1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21 /* 22 * Copyright 2009 Sun Microsystems, Inc. All rights reserved. 23 * Use is subject to license terms. 24 */ 25 26 /* 27 * VM - Hardware Address Translation management for Spitfire MMU. 28 * 29 * This file implements the machine specific hardware translation 30 * needed by the VM system. The machine independent interface is 31 * described in <vm/hat.h> while the machine dependent interface 32 * and data structures are described in <vm/hat_sfmmu.h>. 33 * 34 * The hat layer manages the address translation hardware as a cache 35 * driven by calls from the higher levels in the VM system. 36 */ 37 38 #include <sys/types.h> 39 #include <sys/kstat.h> 40 #include <vm/hat.h> 41 #include <vm/hat_sfmmu.h> 42 #include <vm/page.h> 43 #include <sys/pte.h> 44 #include <sys/systm.h> 45 #include <sys/mman.h> 46 #include <sys/sysmacros.h> 47 #include <sys/machparam.h> 48 #include <sys/vtrace.h> 49 #include <sys/kmem.h> 50 #include <sys/mmu.h> 51 #include <sys/cmn_err.h> 52 #include <sys/cpu.h> 53 #include <sys/cpuvar.h> 54 #include <sys/debug.h> 55 #include <sys/lgrp.h> 56 #include <sys/archsystm.h> 57 #include <sys/machsystm.h> 58 #include <sys/vmsystm.h> 59 #include <vm/as.h> 60 #include <vm/seg.h> 61 #include <vm/seg_kp.h> 62 #include <vm/seg_kmem.h> 63 #include <vm/seg_kpm.h> 64 #include <vm/rm.h> 65 #include <sys/t_lock.h> 66 #include <sys/obpdefs.h> 67 #include <sys/vm_machparam.h> 68 #include <sys/var.h> 69 #include <sys/trap.h> 70 #include <sys/machtrap.h> 71 #include <sys/scb.h> 72 #include <sys/bitmap.h> 73 #include <sys/machlock.h> 74 #include <sys/membar.h> 75 #include <sys/atomic.h> 76 #include <sys/cpu_module.h> 77 #include <sys/prom_debug.h> 78 #include <sys/ksynch.h> 79 #include <sys/mem_config.h> 80 #include <sys/mem_cage.h> 81 #include <vm/vm_dep.h> 82 #include <vm/xhat_sfmmu.h> 83 #include <sys/fpu/fpusystm.h> 84 #include <vm/mach_kpm.h> 85 #include <sys/callb.h> 86 87 #ifdef DEBUG 88 #define SFMMU_VALIDATE_HMERID(hat, rid, saddr, len) \ 89 if (SFMMU_IS_SHMERID_VALID(rid)) { \ 90 caddr_t _eaddr = (saddr) + (len); \ 91 sf_srd_t *_srdp; \ 92 sf_region_t *_rgnp; \ 93 ASSERT((rid) < SFMMU_MAX_HME_REGIONS); \ 94 ASSERT(SF_RGNMAP_TEST(hat->sfmmu_hmeregion_map, rid)); \ 95 ASSERT((hat) != ksfmmup); \ 96 _srdp = (hat)->sfmmu_srdp; \ 97 ASSERT(_srdp != NULL); \ 98 ASSERT(_srdp->srd_refcnt != 0); \ 99 _rgnp = _srdp->srd_hmergnp[(rid)]; \ 100 ASSERT(_rgnp != NULL && _rgnp->rgn_id == rid); \ 101 ASSERT(_rgnp->rgn_refcnt != 0); \ 102 ASSERT(!(_rgnp->rgn_flags & SFMMU_REGION_FREE)); \ 103 ASSERT((_rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK) == \ 104 SFMMU_REGION_HME); \ 105 ASSERT((saddr) >= _rgnp->rgn_saddr); \ 106 ASSERT((saddr) < _rgnp->rgn_saddr + _rgnp->rgn_size); \ 107 ASSERT(_eaddr > _rgnp->rgn_saddr); \ 108 ASSERT(_eaddr <= _rgnp->rgn_saddr + _rgnp->rgn_size); \ 109 } 110 111 #define SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid) \ 112 { \ 113 caddr_t _hsva; \ 114 caddr_t _heva; \ 115 caddr_t _rsva; \ 116 caddr_t _reva; \ 117 int _ttesz = get_hblk_ttesz(hmeblkp); \ 118 int _flagtte; \ 119 ASSERT((srdp)->srd_refcnt != 0); \ 120 ASSERT((rid) < SFMMU_MAX_HME_REGIONS); \ 121 ASSERT((rgnp)->rgn_id == rid); \ 122 ASSERT(!((rgnp)->rgn_flags & SFMMU_REGION_FREE)); \ 123 ASSERT(((rgnp)->rgn_flags & SFMMU_REGION_TYPE_MASK) == \ 124 SFMMU_REGION_HME); \ 125 ASSERT(_ttesz <= (rgnp)->rgn_pgszc); \ 126 _hsva = (caddr_t)get_hblk_base(hmeblkp); \ 127 _heva = get_hblk_endaddr(hmeblkp); \ 128 _rsva = (caddr_t)P2ALIGN( \ 129 (uintptr_t)(rgnp)->rgn_saddr, HBLK_MIN_BYTES); \ 130 _reva = (caddr_t)P2ROUNDUP( \ 131 (uintptr_t)((rgnp)->rgn_saddr + (rgnp)->rgn_size), \ 132 HBLK_MIN_BYTES); \ 133 ASSERT(_hsva >= _rsva); \ 134 ASSERT(_hsva < _reva); \ 135 ASSERT(_heva > _rsva); \ 136 ASSERT(_heva <= _reva); \ 137 _flagtte = (_ttesz < HBLK_MIN_TTESZ) ? HBLK_MIN_TTESZ : \ 138 _ttesz; \ 139 ASSERT(rgnp->rgn_hmeflags & (0x1 << _flagtte)); \ 140 } 141 142 #else /* DEBUG */ 143 #define SFMMU_VALIDATE_HMERID(hat, rid, addr, len) 144 #define SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid) 145 #endif /* DEBUG */ 146 147 #if defined(SF_ERRATA_57) 148 extern caddr_t errata57_limit; 149 #endif 150 151 #define HME8BLK_SZ_RND ((roundup(HME8BLK_SZ, sizeof (int64_t))) / \ 152 (sizeof (int64_t))) 153 #define HBLK_RESERVE ((struct hme_blk *)hblk_reserve) 154 155 #define HBLK_RESERVE_CNT 128 156 #define HBLK_RESERVE_MIN 20 157 158 static struct hme_blk *freehblkp; 159 static kmutex_t freehblkp_lock; 160 static int freehblkcnt; 161 162 static int64_t hblk_reserve[HME8BLK_SZ_RND]; 163 static kmutex_t hblk_reserve_lock; 164 static kthread_t *hblk_reserve_thread; 165 166 static nucleus_hblk8_info_t nucleus_hblk8; 167 static nucleus_hblk1_info_t nucleus_hblk1; 168 169 /* 170 * Data to manage per-cpu hmeblk pending queues, hmeblks are queued here 171 * after the initial phase of removing an hmeblk from the hash chain, see 172 * the detailed comment in sfmmu_hblk_hash_rm() for further details. 173 */ 174 static cpu_hme_pend_t *cpu_hme_pend; 175 static uint_t cpu_hme_pend_thresh; 176 /* 177 * SFMMU specific hat functions 178 */ 179 void hat_pagecachectl(struct page *, int); 180 181 /* flags for hat_pagecachectl */ 182 #define HAT_CACHE 0x1 183 #define HAT_UNCACHE 0x2 184 #define HAT_TMPNC 0x4 185 186 /* 187 * Flag to allow the creation of non-cacheable translations 188 * to system memory. It is off by default. At the moment this 189 * flag is used by the ecache error injector. The error injector 190 * will turn it on when creating such a translation then shut it 191 * off when it's finished. 192 */ 193 194 int sfmmu_allow_nc_trans = 0; 195 196 /* 197 * Flag to disable large page support. 198 * value of 1 => disable all large pages. 199 * bits 1, 2, and 3 are to disable 64K, 512K and 4M pages respectively. 200 * 201 * For example, use the value 0x4 to disable 512K pages. 202 * 203 */ 204 #define LARGE_PAGES_OFF 0x1 205 206 /* 207 * The disable_large_pages and disable_ism_large_pages variables control 208 * hat_memload_array and the page sizes to be used by ISM and the kernel. 209 * 210 * The disable_auto_data_large_pages and disable_auto_text_large_pages variables 211 * are only used to control which OOB pages to use at upper VM segment creation 212 * time, and are set in hat_init_pagesizes and used in the map_pgsz* routines. 213 * Their values may come from platform or CPU specific code to disable page 214 * sizes that should not be used. 215 * 216 * WARNING: 512K pages are currently not supported for ISM/DISM. 217 */ 218 uint_t disable_large_pages = 0; 219 uint_t disable_ism_large_pages = (1 << TTE512K); 220 uint_t disable_auto_data_large_pages = 0; 221 uint_t disable_auto_text_large_pages = 0; 222 223 /* 224 * Private sfmmu data structures for hat management 225 */ 226 static struct kmem_cache *sfmmuid_cache; 227 static struct kmem_cache *mmuctxdom_cache; 228 229 /* 230 * Private sfmmu data structures for tsb management 231 */ 232 static struct kmem_cache *sfmmu_tsbinfo_cache; 233 static struct kmem_cache *sfmmu_tsb8k_cache; 234 static struct kmem_cache *sfmmu_tsb_cache[NLGRPS_MAX]; 235 static vmem_t *kmem_bigtsb_arena; 236 static vmem_t *kmem_tsb_arena; 237 238 /* 239 * sfmmu static variables for hmeblk resource management. 240 */ 241 static vmem_t *hat_memload1_arena; /* HAT translation arena for sfmmu1_cache */ 242 static struct kmem_cache *sfmmu8_cache; 243 static struct kmem_cache *sfmmu1_cache; 244 static struct kmem_cache *pa_hment_cache; 245 246 static kmutex_t ism_mlist_lock; /* mutex for ism mapping list */ 247 /* 248 * private data for ism 249 */ 250 static struct kmem_cache *ism_blk_cache; 251 static struct kmem_cache *ism_ment_cache; 252 #define ISMID_STARTADDR NULL 253 254 /* 255 * Region management data structures and function declarations. 256 */ 257 258 static void sfmmu_leave_srd(sfmmu_t *); 259 static int sfmmu_srdcache_constructor(void *, void *, int); 260 static void sfmmu_srdcache_destructor(void *, void *); 261 static int sfmmu_rgncache_constructor(void *, void *, int); 262 static void sfmmu_rgncache_destructor(void *, void *); 263 static int sfrgnmap_isnull(sf_region_map_t *); 264 static int sfhmergnmap_isnull(sf_hmeregion_map_t *); 265 static int sfmmu_scdcache_constructor(void *, void *, int); 266 static void sfmmu_scdcache_destructor(void *, void *); 267 static void sfmmu_rgn_cb_noop(caddr_t, caddr_t, caddr_t, 268 size_t, void *, u_offset_t); 269 270 static uint_t srd_hashmask = SFMMU_MAX_SRD_BUCKETS - 1; 271 static sf_srd_bucket_t *srd_buckets; 272 static struct kmem_cache *srd_cache; 273 static uint_t srd_rgn_hashmask = SFMMU_MAX_REGION_BUCKETS - 1; 274 static struct kmem_cache *region_cache; 275 static struct kmem_cache *scd_cache; 276 277 #ifdef sun4v 278 int use_bigtsb_arena = 1; 279 #else 280 int use_bigtsb_arena = 0; 281 #endif 282 283 /* External /etc/system tunable, for turning on&off the shctx support */ 284 int disable_shctx = 0; 285 /* Internal variable, set by MD if the HW supports shctx feature */ 286 int shctx_on = 0; 287 288 #ifdef DEBUG 289 static void check_scd_sfmmu_list(sfmmu_t **, sfmmu_t *, int); 290 #endif 291 static void sfmmu_to_scd_list(sfmmu_t **, sfmmu_t *); 292 static void sfmmu_from_scd_list(sfmmu_t **, sfmmu_t *); 293 294 static sf_scd_t *sfmmu_alloc_scd(sf_srd_t *, sf_region_map_t *); 295 static void sfmmu_find_scd(sfmmu_t *); 296 static void sfmmu_join_scd(sf_scd_t *, sfmmu_t *); 297 static void sfmmu_finish_join_scd(sfmmu_t *); 298 static void sfmmu_leave_scd(sfmmu_t *, uchar_t); 299 static void sfmmu_destroy_scd(sf_srd_t *, sf_scd_t *, sf_region_map_t *); 300 static int sfmmu_alloc_scd_tsbs(sf_srd_t *, sf_scd_t *); 301 static void sfmmu_free_scd_tsbs(sfmmu_t *); 302 static void sfmmu_tsb_inv_ctx(sfmmu_t *); 303 static int find_ism_rid(sfmmu_t *, sfmmu_t *, caddr_t, uint_t *); 304 static void sfmmu_ism_hatflags(sfmmu_t *, int); 305 static int sfmmu_srd_lock_held(sf_srd_t *); 306 static void sfmmu_remove_scd(sf_scd_t **, sf_scd_t *); 307 static void sfmmu_add_scd(sf_scd_t **headp, sf_scd_t *); 308 static void sfmmu_link_scd_to_regions(sf_srd_t *, sf_scd_t *); 309 static void sfmmu_unlink_scd_from_regions(sf_srd_t *, sf_scd_t *); 310 static void sfmmu_link_to_hmeregion(sfmmu_t *, sf_region_t *); 311 static void sfmmu_unlink_from_hmeregion(sfmmu_t *, sf_region_t *); 312 313 /* 314 * ``hat_lock'' is a hashed mutex lock for protecting sfmmu TSB lists, 315 * HAT flags, synchronizing TLB/TSB coherency, and context management. 316 * The lock is hashed on the sfmmup since the case where we need to lock 317 * all processes is rare but does occur (e.g. we need to unload a shared 318 * mapping from all processes using the mapping). We have a lot of buckets, 319 * and each slab of sfmmu_t's can use about a quarter of them, giving us 320 * a fairly good distribution without wasting too much space and overhead 321 * when we have to grab them all. 322 */ 323 #define SFMMU_NUM_LOCK 128 /* must be power of two */ 324 hatlock_t hat_lock[SFMMU_NUM_LOCK]; 325 326 /* 327 * Hash algorithm optimized for a small number of slabs. 328 * 7 is (highbit((sizeof sfmmu_t)) - 1) 329 * This hash algorithm is based upon the knowledge that sfmmu_t's come from a 330 * kmem_cache, and thus they will be sequential within that cache. In 331 * addition, each new slab will have a different "color" up to cache_maxcolor 332 * which will skew the hashing for each successive slab which is allocated. 333 * If the size of sfmmu_t changed to a larger size, this algorithm may need 334 * to be revisited. 335 */ 336 #define TSB_HASH_SHIFT_BITS (7) 337 #define PTR_HASH(x) ((uintptr_t)x >> TSB_HASH_SHIFT_BITS) 338 339 #ifdef DEBUG 340 int tsb_hash_debug = 0; 341 #define TSB_HASH(sfmmup) \ 342 (tsb_hash_debug ? &hat_lock[0] : \ 343 &hat_lock[PTR_HASH(sfmmup) & (SFMMU_NUM_LOCK-1)]) 344 #else /* DEBUG */ 345 #define TSB_HASH(sfmmup) &hat_lock[PTR_HASH(sfmmup) & (SFMMU_NUM_LOCK-1)] 346 #endif /* DEBUG */ 347 348 349 /* sfmmu_replace_tsb() return codes. */ 350 typedef enum tsb_replace_rc { 351 TSB_SUCCESS, 352 TSB_ALLOCFAIL, 353 TSB_LOSTRACE, 354 TSB_ALREADY_SWAPPED, 355 TSB_CANTGROW 356 } tsb_replace_rc_t; 357 358 /* 359 * Flags for TSB allocation routines. 360 */ 361 #define TSB_ALLOC 0x01 362 #define TSB_FORCEALLOC 0x02 363 #define TSB_GROW 0x04 364 #define TSB_SHRINK 0x08 365 #define TSB_SWAPIN 0x10 366 367 /* 368 * Support for HAT callbacks. 369 */ 370 #define SFMMU_MAX_RELOC_CALLBACKS 10 371 int sfmmu_max_cb_id = SFMMU_MAX_RELOC_CALLBACKS; 372 static id_t sfmmu_cb_nextid = 0; 373 static id_t sfmmu_tsb_cb_id; 374 struct sfmmu_callback *sfmmu_cb_table; 375 376 /* 377 * Kernel page relocation is enabled by default for non-caged 378 * kernel pages. This has little effect unless segkmem_reloc is 379 * set, since by default kernel memory comes from inside the 380 * kernel cage. 381 */ 382 int hat_kpr_enabled = 1; 383 384 kmutex_t kpr_mutex; 385 kmutex_t kpr_suspendlock; 386 kthread_t *kreloc_thread; 387 388 /* 389 * Enable VA->PA translation sanity checking on DEBUG kernels. 390 * Disabled by default. This is incompatible with some 391 * drivers (error injector, RSM) so if it breaks you get 392 * to keep both pieces. 393 */ 394 int hat_check_vtop = 0; 395 396 /* 397 * Private sfmmu routines (prototypes) 398 */ 399 static struct hme_blk *sfmmu_shadow_hcreate(sfmmu_t *, caddr_t, int, uint_t); 400 static struct hme_blk *sfmmu_hblk_alloc(sfmmu_t *, caddr_t, 401 struct hmehash_bucket *, uint_t, hmeblk_tag, uint_t, 402 uint_t); 403 static caddr_t sfmmu_hblk_unload(struct hat *, struct hme_blk *, caddr_t, 404 caddr_t, demap_range_t *, uint_t); 405 static caddr_t sfmmu_hblk_sync(struct hat *, struct hme_blk *, caddr_t, 406 caddr_t, int); 407 static void sfmmu_hblk_free(struct hme_blk **); 408 static void sfmmu_hblks_list_purge(struct hme_blk **, int); 409 static uint_t sfmmu_get_free_hblk(struct hme_blk **, uint_t); 410 static uint_t sfmmu_put_free_hblk(struct hme_blk *, uint_t); 411 static struct hme_blk *sfmmu_hblk_steal(int); 412 static int sfmmu_steal_this_hblk(struct hmehash_bucket *, 413 struct hme_blk *, uint64_t, struct hme_blk *); 414 static caddr_t sfmmu_hblk_unlock(struct hme_blk *, caddr_t, caddr_t); 415 416 static void hat_do_memload_array(struct hat *, caddr_t, size_t, 417 struct page **, uint_t, uint_t, uint_t); 418 static void hat_do_memload(struct hat *, caddr_t, struct page *, 419 uint_t, uint_t, uint_t); 420 static void sfmmu_memload_batchsmall(struct hat *, caddr_t, page_t **, 421 uint_t, uint_t, pgcnt_t, uint_t); 422 void sfmmu_tteload(struct hat *, tte_t *, caddr_t, page_t *, 423 uint_t); 424 static int sfmmu_tteload_array(sfmmu_t *, tte_t *, caddr_t, page_t **, 425 uint_t, uint_t); 426 static struct hmehash_bucket *sfmmu_tteload_acquire_hashbucket(sfmmu_t *, 427 caddr_t, int, uint_t); 428 static struct hme_blk *sfmmu_tteload_find_hmeblk(sfmmu_t *, 429 struct hmehash_bucket *, caddr_t, uint_t, uint_t, 430 uint_t); 431 static int sfmmu_tteload_addentry(sfmmu_t *, struct hme_blk *, tte_t *, 432 caddr_t, page_t **, uint_t, uint_t); 433 static void sfmmu_tteload_release_hashbucket(struct hmehash_bucket *); 434 435 static int sfmmu_pagearray_setup(caddr_t, page_t **, tte_t *, int); 436 static pfn_t sfmmu_uvatopfn(caddr_t, sfmmu_t *, tte_t *); 437 void sfmmu_memtte(tte_t *, pfn_t, uint_t, int); 438 #ifdef VAC 439 static void sfmmu_vac_conflict(struct hat *, caddr_t, page_t *); 440 static int sfmmu_vacconflict_array(caddr_t, page_t *, int *); 441 int tst_tnc(page_t *pp, pgcnt_t); 442 void conv_tnc(page_t *pp, int); 443 #endif 444 445 static void sfmmu_get_ctx(sfmmu_t *); 446 static void sfmmu_free_sfmmu(sfmmu_t *); 447 448 static void sfmmu_ttesync(struct hat *, caddr_t, tte_t *, page_t *); 449 static void sfmmu_chgattr(struct hat *, caddr_t, size_t, uint_t, int); 450 451 cpuset_t sfmmu_pageunload(page_t *, struct sf_hment *, int); 452 static void hat_pagereload(struct page *, struct page *); 453 static cpuset_t sfmmu_pagesync(page_t *, struct sf_hment *, uint_t); 454 #ifdef VAC 455 void sfmmu_page_cache_array(page_t *, int, int, pgcnt_t); 456 static void sfmmu_page_cache(page_t *, int, int, int); 457 #endif 458 459 cpuset_t sfmmu_rgntlb_demap(caddr_t, sf_region_t *, 460 struct hme_blk *, int); 461 static void sfmmu_tlbcache_demap(caddr_t, sfmmu_t *, struct hme_blk *, 462 pfn_t, int, int, int, int); 463 static void sfmmu_ismtlbcache_demap(caddr_t, sfmmu_t *, struct hme_blk *, 464 pfn_t, int); 465 static void sfmmu_tlb_demap(caddr_t, sfmmu_t *, struct hme_blk *, int, int); 466 static void sfmmu_tlb_range_demap(demap_range_t *); 467 static void sfmmu_invalidate_ctx(sfmmu_t *); 468 static void sfmmu_sync_mmustate(sfmmu_t *); 469 470 static void sfmmu_tsbinfo_setup_phys(struct tsb_info *, pfn_t); 471 static int sfmmu_tsbinfo_alloc(struct tsb_info **, int, int, uint_t, 472 sfmmu_t *); 473 static void sfmmu_tsb_free(struct tsb_info *); 474 static void sfmmu_tsbinfo_free(struct tsb_info *); 475 static int sfmmu_init_tsbinfo(struct tsb_info *, int, int, uint_t, 476 sfmmu_t *); 477 static void sfmmu_tsb_chk_reloc(sfmmu_t *, hatlock_t *); 478 static void sfmmu_tsb_swapin(sfmmu_t *, hatlock_t *); 479 static int sfmmu_select_tsb_szc(pgcnt_t); 480 static void sfmmu_mod_tsb(sfmmu_t *, caddr_t, tte_t *, int); 481 #define sfmmu_load_tsb(sfmmup, vaddr, tte, szc) \ 482 sfmmu_mod_tsb(sfmmup, vaddr, tte, szc) 483 #define sfmmu_unload_tsb(sfmmup, vaddr, szc) \ 484 sfmmu_mod_tsb(sfmmup, vaddr, NULL, szc) 485 static void sfmmu_copy_tsb(struct tsb_info *, struct tsb_info *); 486 static tsb_replace_rc_t sfmmu_replace_tsb(sfmmu_t *, struct tsb_info *, uint_t, 487 hatlock_t *, uint_t); 488 static void sfmmu_size_tsb(sfmmu_t *, int, uint64_t, uint64_t, int); 489 490 #ifdef VAC 491 void sfmmu_cache_flush(pfn_t, int); 492 void sfmmu_cache_flushcolor(int, pfn_t); 493 #endif 494 static caddr_t sfmmu_hblk_chgattr(sfmmu_t *, struct hme_blk *, caddr_t, 495 caddr_t, demap_range_t *, uint_t, int); 496 497 static uint64_t sfmmu_vtop_attr(uint_t, int mode, tte_t *); 498 static uint_t sfmmu_ptov_attr(tte_t *); 499 static caddr_t sfmmu_hblk_chgprot(sfmmu_t *, struct hme_blk *, caddr_t, 500 caddr_t, demap_range_t *, uint_t); 501 static uint_t sfmmu_vtop_prot(uint_t, uint_t *); 502 static int sfmmu_idcache_constructor(void *, void *, int); 503 static void sfmmu_idcache_destructor(void *, void *); 504 static int sfmmu_hblkcache_constructor(void *, void *, int); 505 static void sfmmu_hblkcache_destructor(void *, void *); 506 static void sfmmu_hblkcache_reclaim(void *); 507 static void sfmmu_shadow_hcleanup(sfmmu_t *, struct hme_blk *, 508 struct hmehash_bucket *); 509 static void sfmmu_hblk_hash_rm(struct hmehash_bucket *, struct hme_blk *, 510 struct hme_blk *, struct hme_blk **, int); 511 static void sfmmu_hblk_hash_add(struct hmehash_bucket *, struct hme_blk *, 512 uint64_t); 513 static struct hme_blk *sfmmu_check_pending_hblks(int); 514 static void sfmmu_free_hblks(sfmmu_t *, caddr_t, caddr_t, int); 515 static void sfmmu_cleanup_rhblk(sf_srd_t *, caddr_t, uint_t, int); 516 static void sfmmu_unload_hmeregion_va(sf_srd_t *, uint_t, caddr_t, caddr_t, 517 int, caddr_t *); 518 static void sfmmu_unload_hmeregion(sf_srd_t *, sf_region_t *); 519 520 static void sfmmu_rm_large_mappings(page_t *, int); 521 522 static void hat_lock_init(void); 523 static void hat_kstat_init(void); 524 static int sfmmu_kstat_percpu_update(kstat_t *ksp, int rw); 525 static void sfmmu_set_scd_rttecnt(sf_srd_t *, sf_scd_t *); 526 static int sfmmu_is_rgnva(sf_srd_t *, caddr_t, ulong_t, ulong_t); 527 static void sfmmu_check_page_sizes(sfmmu_t *, int); 528 int fnd_mapping_sz(page_t *); 529 static void iment_add(struct ism_ment *, struct hat *); 530 static void iment_sub(struct ism_ment *, struct hat *); 531 static pgcnt_t ism_tsb_entries(sfmmu_t *, int szc); 532 extern void sfmmu_setup_tsbinfo(sfmmu_t *); 533 extern void sfmmu_clear_utsbinfo(void); 534 535 static void sfmmu_ctx_wrap_around(mmu_ctx_t *); 536 537 extern int vpm_enable; 538 539 /* kpm globals */ 540 #ifdef DEBUG 541 /* 542 * Enable trap level tsbmiss handling 543 */ 544 int kpm_tsbmtl = 1; 545 546 /* 547 * Flush the TLB on kpm mapout. Note: Xcalls are used (again) for the 548 * required TLB shootdowns in this case, so handle w/ care. Off by default. 549 */ 550 int kpm_tlb_flush; 551 #endif /* DEBUG */ 552 553 static void *sfmmu_vmem_xalloc_aligned_wrapper(vmem_t *, size_t, int); 554 555 #ifdef DEBUG 556 static void sfmmu_check_hblk_flist(); 557 #endif 558 559 /* 560 * Semi-private sfmmu data structures. Some of them are initialize in 561 * startup or in hat_init. Some of them are private but accessed by 562 * assembly code or mach_sfmmu.c 563 */ 564 struct hmehash_bucket *uhme_hash; /* user hmeblk hash table */ 565 struct hmehash_bucket *khme_hash; /* kernel hmeblk hash table */ 566 uint64_t uhme_hash_pa; /* PA of uhme_hash */ 567 uint64_t khme_hash_pa; /* PA of khme_hash */ 568 int uhmehash_num; /* # of buckets in user hash table */ 569 int khmehash_num; /* # of buckets in kernel hash table */ 570 571 uint_t max_mmu_ctxdoms = 0; /* max context domains in the system */ 572 mmu_ctx_t **mmu_ctxs_tbl; /* global array of context domains */ 573 uint64_t mmu_saved_gnum = 0; /* to init incoming MMUs' gnums */ 574 575 #define DEFAULT_NUM_CTXS_PER_MMU 8192 576 static uint_t nctxs = DEFAULT_NUM_CTXS_PER_MMU; 577 578 int cache; /* describes system cache */ 579 580 caddr_t ktsb_base; /* kernel 8k-indexed tsb base address */ 581 uint64_t ktsb_pbase; /* kernel 8k-indexed tsb phys address */ 582 int ktsb_szcode; /* kernel 8k-indexed tsb size code */ 583 int ktsb_sz; /* kernel 8k-indexed tsb size */ 584 585 caddr_t ktsb4m_base; /* kernel 4m-indexed tsb base address */ 586 uint64_t ktsb4m_pbase; /* kernel 4m-indexed tsb phys address */ 587 int ktsb4m_szcode; /* kernel 4m-indexed tsb size code */ 588 int ktsb4m_sz; /* kernel 4m-indexed tsb size */ 589 590 uint64_t kpm_tsbbase; /* kernel seg_kpm 4M TSB base address */ 591 int kpm_tsbsz; /* kernel seg_kpm 4M TSB size code */ 592 uint64_t kpmsm_tsbbase; /* kernel seg_kpm 8K TSB base address */ 593 int kpmsm_tsbsz; /* kernel seg_kpm 8K TSB size code */ 594 595 #ifndef sun4v 596 int utsb_dtlb_ttenum = -1; /* index in TLB for utsb locked TTE */ 597 int utsb4m_dtlb_ttenum = -1; /* index in TLB for 4M TSB TTE */ 598 int dtlb_resv_ttenum; /* index in TLB of first reserved TTE */ 599 caddr_t utsb_vabase; /* reserved kernel virtual memory */ 600 caddr_t utsb4m_vabase; /* for trap handler TSB accesses */ 601 #endif /* sun4v */ 602 uint64_t tsb_alloc_bytes = 0; /* bytes allocated to TSBs */ 603 vmem_t *kmem_tsb_default_arena[NLGRPS_MAX]; /* For dynamic TSBs */ 604 vmem_t *kmem_bigtsb_default_arena[NLGRPS_MAX]; /* dynamic 256M TSBs */ 605 606 /* 607 * Size to use for TSB slabs. Future platforms that support page sizes 608 * larger than 4M may wish to change these values, and provide their own 609 * assembly macros for building and decoding the TSB base register contents. 610 * Note disable_large_pages will override the value set here. 611 */ 612 static uint_t tsb_slab_ttesz = TTE4M; 613 size_t tsb_slab_size = MMU_PAGESIZE4M; 614 uint_t tsb_slab_shift = MMU_PAGESHIFT4M; 615 /* PFN mask for TTE */ 616 size_t tsb_slab_mask = MMU_PAGEOFFSET4M >> MMU_PAGESHIFT; 617 618 /* 619 * Size to use for TSB slabs. These are used only when 256M tsb arenas 620 * exist. 621 */ 622 static uint_t bigtsb_slab_ttesz = TTE256M; 623 static size_t bigtsb_slab_size = MMU_PAGESIZE256M; 624 static uint_t bigtsb_slab_shift = MMU_PAGESHIFT256M; 625 /* 256M page alignment for 8K pfn */ 626 static size_t bigtsb_slab_mask = MMU_PAGEOFFSET256M >> MMU_PAGESHIFT; 627 628 /* largest TSB size to grow to, will be smaller on smaller memory systems */ 629 static int tsb_max_growsize = 0; 630 631 /* 632 * Tunable parameters dealing with TSB policies. 633 */ 634 635 /* 636 * This undocumented tunable forces all 8K TSBs to be allocated from 637 * the kernel heap rather than from the kmem_tsb_default_arena arenas. 638 */ 639 #ifdef DEBUG 640 int tsb_forceheap = 0; 641 #endif /* DEBUG */ 642 643 /* 644 * Decide whether to use per-lgroup arenas, or one global set of 645 * TSB arenas. The default is not to break up per-lgroup, since 646 * most platforms don't recognize any tangible benefit from it. 647 */ 648 int tsb_lgrp_affinity = 0; 649 650 /* 651 * Used for growing the TSB based on the process RSS. 652 * tsb_rss_factor is based on the smallest TSB, and is 653 * shifted by the TSB size to determine if we need to grow. 654 * The default will grow the TSB if the number of TTEs for 655 * this page size exceeds 75% of the number of TSB entries, 656 * which should _almost_ eliminate all conflict misses 657 * (at the expense of using up lots and lots of memory). 658 */ 659 #define TSB_RSS_FACTOR (TSB_ENTRIES(TSB_MIN_SZCODE) * 0.75) 660 #define SFMMU_RSS_TSBSIZE(tsbszc) (tsb_rss_factor << tsbszc) 661 #define SELECT_TSB_SIZECODE(pgcnt) ( \ 662 (enable_tsb_rss_sizing)? sfmmu_select_tsb_szc(pgcnt) : \ 663 default_tsb_size) 664 #define TSB_OK_SHRINK() \ 665 (tsb_alloc_bytes > tsb_alloc_hiwater || freemem < desfree) 666 #define TSB_OK_GROW() \ 667 (tsb_alloc_bytes < tsb_alloc_hiwater && freemem > desfree) 668 669 int enable_tsb_rss_sizing = 1; 670 int tsb_rss_factor = (int)TSB_RSS_FACTOR; 671 672 /* which TSB size code to use for new address spaces or if rss sizing off */ 673 int default_tsb_size = TSB_8K_SZCODE; 674 675 static uint64_t tsb_alloc_hiwater; /* limit TSB reserved memory */ 676 uint64_t tsb_alloc_hiwater_factor; /* tsb_alloc_hiwater = physmem / this */ 677 #define TSB_ALLOC_HIWATER_FACTOR_DEFAULT 32 678 679 #ifdef DEBUG 680 static int tsb_random_size = 0; /* set to 1 to test random tsb sizes on alloc */ 681 static int tsb_grow_stress = 0; /* if set to 1, keep replacing TSB w/ random */ 682 static int tsb_alloc_mtbf = 0; /* fail allocation every n attempts */ 683 static int tsb_alloc_fail_mtbf = 0; 684 static int tsb_alloc_count = 0; 685 #endif /* DEBUG */ 686 687 /* if set to 1, will remap valid TTEs when growing TSB. */ 688 int tsb_remap_ttes = 1; 689 690 /* 691 * If we have more than this many mappings, allocate a second TSB. 692 * This default is chosen because the I/D fully associative TLBs are 693 * assumed to have at least 8 available entries. Platforms with a 694 * larger fully-associative TLB could probably override the default. 695 */ 696 697 #ifdef sun4v 698 int tsb_sectsb_threshold = 0; 699 #else 700 int tsb_sectsb_threshold = 8; 701 #endif 702 703 /* 704 * kstat data 705 */ 706 struct sfmmu_global_stat sfmmu_global_stat; 707 struct sfmmu_tsbsize_stat sfmmu_tsbsize_stat; 708 709 /* 710 * Global data 711 */ 712 sfmmu_t *ksfmmup; /* kernel's hat id */ 713 714 #ifdef DEBUG 715 static void chk_tte(tte_t *, tte_t *, tte_t *, struct hme_blk *); 716 #endif 717 718 /* sfmmu locking operations */ 719 static kmutex_t *sfmmu_mlspl_enter(struct page *, int); 720 static int sfmmu_mlspl_held(struct page *, int); 721 722 kmutex_t *sfmmu_page_enter(page_t *); 723 void sfmmu_page_exit(kmutex_t *); 724 int sfmmu_page_spl_held(struct page *); 725 726 /* sfmmu internal locking operations - accessed directly */ 727 static void sfmmu_mlist_reloc_enter(page_t *, page_t *, 728 kmutex_t **, kmutex_t **); 729 static void sfmmu_mlist_reloc_exit(kmutex_t *, kmutex_t *); 730 static hatlock_t * 731 sfmmu_hat_enter(sfmmu_t *); 732 static hatlock_t * 733 sfmmu_hat_tryenter(sfmmu_t *); 734 static void sfmmu_hat_exit(hatlock_t *); 735 static void sfmmu_hat_lock_all(void); 736 static void sfmmu_hat_unlock_all(void); 737 static void sfmmu_ismhat_enter(sfmmu_t *, int); 738 static void sfmmu_ismhat_exit(sfmmu_t *, int); 739 740 /* 741 * Array of mutexes protecting a page's mapping list and p_nrm field. 742 * 743 * The hash function looks complicated, but is made up so that: 744 * 745 * "pp" not shifted, so adjacent pp values will hash to different cache lines 746 * (8 byte alignment * 8 bytes/mutes == 64 byte coherency subblock) 747 * 748 * "pp" >> mml_shift, incorporates more source bits into the hash result 749 * 750 * "& (mml_table_size - 1), should be faster than using remainder "%" 751 * 752 * Hopefully, mml_table, mml_table_size and mml_shift are all in the same 753 * cacheline, since they get declared next to each other below. We'll trust 754 * ld not to do something random. 755 */ 756 #ifdef DEBUG 757 int mlist_hash_debug = 0; 758 #define MLIST_HASH(pp) (mlist_hash_debug ? &mml_table[0] : \ 759 &mml_table[((uintptr_t)(pp) + \ 760 ((uintptr_t)(pp) >> mml_shift)) & (mml_table_sz - 1)]) 761 #else /* !DEBUG */ 762 #define MLIST_HASH(pp) &mml_table[ \ 763 ((uintptr_t)(pp) + ((uintptr_t)(pp) >> mml_shift)) & (mml_table_sz - 1)] 764 #endif /* !DEBUG */ 765 766 kmutex_t *mml_table; 767 uint_t mml_table_sz; /* must be a power of 2 */ 768 uint_t mml_shift; /* log2(mml_table_sz) + 3 for align */ 769 770 kpm_hlk_t *kpmp_table; 771 uint_t kpmp_table_sz; /* must be a power of 2 */ 772 uchar_t kpmp_shift; 773 774 kpm_shlk_t *kpmp_stable; 775 uint_t kpmp_stable_sz; /* must be a power of 2 */ 776 777 /* 778 * SPL_HASH was improved to avoid false cache line sharing 779 */ 780 #define SPL_TABLE_SIZE 128 781 #define SPL_MASK (SPL_TABLE_SIZE - 1) 782 #define SPL_SHIFT 7 /* log2(SPL_TABLE_SIZE) */ 783 784 #define SPL_INDEX(pp) \ 785 ((((uintptr_t)(pp) >> SPL_SHIFT) ^ \ 786 ((uintptr_t)(pp) >> (SPL_SHIFT << 1))) & \ 787 (SPL_TABLE_SIZE - 1)) 788 789 #define SPL_HASH(pp) \ 790 (&sfmmu_page_lock[SPL_INDEX(pp) & SPL_MASK].pad_mutex) 791 792 static pad_mutex_t sfmmu_page_lock[SPL_TABLE_SIZE]; 793 794 795 /* 796 * hat_unload_callback() will group together callbacks in order 797 * to avoid xt_sync() calls. This is the maximum size of the group. 798 */ 799 #define MAX_CB_ADDR 32 800 801 tte_t hw_tte; 802 static ulong_t sfmmu_dmr_maxbit = DMR_MAXBIT; 803 804 static char *mmu_ctx_kstat_names[] = { 805 "mmu_ctx_tsb_exceptions", 806 "mmu_ctx_tsb_raise_exception", 807 "mmu_ctx_wrap_around", 808 }; 809 810 /* 811 * Wrapper for vmem_xalloc since vmem_create only allows limited 812 * parameters for vm_source_alloc functions. This function allows us 813 * to specify alignment consistent with the size of the object being 814 * allocated. 815 */ 816 static void * 817 sfmmu_vmem_xalloc_aligned_wrapper(vmem_t *vmp, size_t size, int vmflag) 818 { 819 return (vmem_xalloc(vmp, size, size, 0, 0, NULL, NULL, vmflag)); 820 } 821 822 /* Common code for setting tsb_alloc_hiwater. */ 823 #define SFMMU_SET_TSB_ALLOC_HIWATER(pages) tsb_alloc_hiwater = \ 824 ptob(pages) / tsb_alloc_hiwater_factor 825 826 /* 827 * Set tsb_max_growsize to allow at most all of physical memory to be mapped by 828 * a single TSB. physmem is the number of physical pages so we need physmem 8K 829 * TTEs to represent all those physical pages. We round this up by using 830 * 1<<highbit(). To figure out which size code to use, remember that the size 831 * code is just an amount to shift the smallest TSB size to get the size of 832 * this TSB. So we subtract that size, TSB_START_SIZE, from highbit() (or 833 * highbit() - 1) to get the size code for the smallest TSB that can represent 834 * all of physical memory, while erring on the side of too much. 835 * 836 * Restrict tsb_max_growsize to make sure that: 837 * 1) TSBs can't grow larger than the TSB slab size 838 * 2) TSBs can't grow larger than UTSB_MAX_SZCODE. 839 */ 840 #define SFMMU_SET_TSB_MAX_GROWSIZE(pages) { \ 841 int _i, _szc, _slabszc, _tsbszc; \ 842 \ 843 _i = highbit(pages); \ 844 if ((1 << (_i - 1)) == (pages)) \ 845 _i--; /* 2^n case, round down */ \ 846 _szc = _i - TSB_START_SIZE; \ 847 _slabszc = bigtsb_slab_shift - (TSB_START_SIZE + TSB_ENTRY_SHIFT); \ 848 _tsbszc = MIN(_szc, _slabszc); \ 849 tsb_max_growsize = MIN(_tsbszc, UTSB_MAX_SZCODE); \ 850 } 851 852 /* 853 * Given a pointer to an sfmmu and a TTE size code, return a pointer to the 854 * tsb_info which handles that TTE size. 855 */ 856 #define SFMMU_GET_TSBINFO(tsbinfop, sfmmup, tte_szc) { \ 857 (tsbinfop) = (sfmmup)->sfmmu_tsb; \ 858 ASSERT(((tsbinfop)->tsb_flags & TSB_SHAREDCTX) || \ 859 sfmmu_hat_lock_held(sfmmup)); \ 860 if ((tte_szc) >= TTE4M) { \ 861 ASSERT((tsbinfop) != NULL); \ 862 (tsbinfop) = (tsbinfop)->tsb_next; \ 863 } \ 864 } 865 866 /* 867 * Macro to use to unload entries from the TSB. 868 * It has knowledge of which page sizes get replicated in the TSB 869 * and will call the appropriate unload routine for the appropriate size. 870 */ 871 #define SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp, ismhat) \ 872 { \ 873 int ttesz = get_hblk_ttesz(hmeblkp); \ 874 if (ttesz == TTE8K || ttesz == TTE4M) { \ 875 sfmmu_unload_tsb(sfmmup, addr, ttesz); \ 876 } else { \ 877 caddr_t sva = ismhat ? addr : \ 878 (caddr_t)get_hblk_base(hmeblkp); \ 879 caddr_t eva = sva + get_hblk_span(hmeblkp); \ 880 ASSERT(addr >= sva && addr < eva); \ 881 sfmmu_unload_tsb_range(sfmmup, sva, eva, ttesz); \ 882 } \ 883 } 884 885 886 /* Update tsb_alloc_hiwater after memory is configured. */ 887 /*ARGSUSED*/ 888 static void 889 sfmmu_update_post_add(void *arg, pgcnt_t delta_pages) 890 { 891 /* Assumes physmem has already been updated. */ 892 SFMMU_SET_TSB_ALLOC_HIWATER(physmem); 893 SFMMU_SET_TSB_MAX_GROWSIZE(physmem); 894 } 895 896 /* 897 * Update tsb_alloc_hiwater before memory is deleted. We'll do nothing here 898 * and update tsb_alloc_hiwater and tsb_max_growsize after the memory is 899 * deleted. 900 */ 901 /*ARGSUSED*/ 902 static int 903 sfmmu_update_pre_del(void *arg, pgcnt_t delta_pages) 904 { 905 return (0); 906 } 907 908 /* Update tsb_alloc_hiwater after memory fails to be unconfigured. */ 909 /*ARGSUSED*/ 910 static void 911 sfmmu_update_post_del(void *arg, pgcnt_t delta_pages, int cancelled) 912 { 913 /* 914 * Whether the delete was cancelled or not, just go ahead and update 915 * tsb_alloc_hiwater and tsb_max_growsize. 916 */ 917 SFMMU_SET_TSB_ALLOC_HIWATER(physmem); 918 SFMMU_SET_TSB_MAX_GROWSIZE(physmem); 919 } 920 921 static kphysm_setup_vector_t sfmmu_update_vec = { 922 KPHYSM_SETUP_VECTOR_VERSION, /* version */ 923 sfmmu_update_post_add, /* post_add */ 924 sfmmu_update_pre_del, /* pre_del */ 925 sfmmu_update_post_del /* post_del */ 926 }; 927 928 929 /* 930 * HME_BLK HASH PRIMITIVES 931 */ 932 933 /* 934 * Enter a hme on the mapping list for page pp. 935 * When large pages are more prevalent in the system we might want to 936 * keep the mapping list in ascending order by the hment size. For now, 937 * small pages are more frequent, so don't slow it down. 938 */ 939 #define HME_ADD(hme, pp) \ 940 { \ 941 ASSERT(sfmmu_mlist_held(pp)); \ 942 \ 943 hme->hme_prev = NULL; \ 944 hme->hme_next = pp->p_mapping; \ 945 hme->hme_page = pp; \ 946 if (pp->p_mapping) { \ 947 ((struct sf_hment *)(pp->p_mapping))->hme_prev = hme;\ 948 ASSERT(pp->p_share > 0); \ 949 } else { \ 950 /* EMPTY */ \ 951 ASSERT(pp->p_share == 0); \ 952 } \ 953 pp->p_mapping = hme; \ 954 pp->p_share++; \ 955 } 956 957 /* 958 * Enter a hme on the mapping list for page pp. 959 * If we are unmapping a large translation, we need to make sure that the 960 * change is reflect in the corresponding bit of the p_index field. 961 */ 962 #define HME_SUB(hme, pp) \ 963 { \ 964 ASSERT(sfmmu_mlist_held(pp)); \ 965 ASSERT(hme->hme_page == pp || IS_PAHME(hme)); \ 966 \ 967 if (pp->p_mapping == NULL) { \ 968 panic("hme_remove - no mappings"); \ 969 } \ 970 \ 971 membar_stst(); /* ensure previous stores finish */ \ 972 \ 973 ASSERT(pp->p_share > 0); \ 974 pp->p_share--; \ 975 \ 976 if (hme->hme_prev) { \ 977 ASSERT(pp->p_mapping != hme); \ 978 ASSERT(hme->hme_prev->hme_page == pp || \ 979 IS_PAHME(hme->hme_prev)); \ 980 hme->hme_prev->hme_next = hme->hme_next; \ 981 } else { \ 982 ASSERT(pp->p_mapping == hme); \ 983 pp->p_mapping = hme->hme_next; \ 984 ASSERT((pp->p_mapping == NULL) ? \ 985 (pp->p_share == 0) : 1); \ 986 } \ 987 \ 988 if (hme->hme_next) { \ 989 ASSERT(hme->hme_next->hme_page == pp || \ 990 IS_PAHME(hme->hme_next)); \ 991 hme->hme_next->hme_prev = hme->hme_prev; \ 992 } \ 993 \ 994 /* zero out the entry */ \ 995 hme->hme_next = NULL; \ 996 hme->hme_prev = NULL; \ 997 hme->hme_page = NULL; \ 998 \ 999 if (hme_size(hme) > TTE8K) { \ 1000 /* remove mappings for remainder of large pg */ \ 1001 sfmmu_rm_large_mappings(pp, hme_size(hme)); \ 1002 } \ 1003 } 1004 1005 /* 1006 * This function returns the hment given the hme_blk and a vaddr. 1007 * It assumes addr has already been checked to belong to hme_blk's 1008 * range. 1009 */ 1010 #define HBLKTOHME(hment, hmeblkp, addr) \ 1011 { \ 1012 int index; \ 1013 HBLKTOHME_IDX(hment, hmeblkp, addr, index) \ 1014 } 1015 1016 /* 1017 * Version of HBLKTOHME that also returns the index in hmeblkp 1018 * of the hment. 1019 */ 1020 #define HBLKTOHME_IDX(hment, hmeblkp, addr, idx) \ 1021 { \ 1022 ASSERT(in_hblk_range((hmeblkp), (addr))); \ 1023 \ 1024 if (get_hblk_ttesz(hmeblkp) == TTE8K) { \ 1025 idx = (((uintptr_t)(addr) >> MMU_PAGESHIFT) & (NHMENTS-1)); \ 1026 } else \ 1027 idx = 0; \ 1028 \ 1029 (hment) = &(hmeblkp)->hblk_hme[idx]; \ 1030 } 1031 1032 /* 1033 * Disable any page sizes not supported by the CPU 1034 */ 1035 void 1036 hat_init_pagesizes() 1037 { 1038 int i; 1039 1040 mmu_exported_page_sizes = 0; 1041 for (i = TTE8K; i < max_mmu_page_sizes; i++) { 1042 1043 szc_2_userszc[i] = (uint_t)-1; 1044 userszc_2_szc[i] = (uint_t)-1; 1045 1046 if ((mmu_exported_pagesize_mask & (1 << i)) == 0) { 1047 disable_large_pages |= (1 << i); 1048 } else { 1049 szc_2_userszc[i] = mmu_exported_page_sizes; 1050 userszc_2_szc[mmu_exported_page_sizes] = i; 1051 mmu_exported_page_sizes++; 1052 } 1053 } 1054 1055 disable_ism_large_pages |= disable_large_pages; 1056 disable_auto_data_large_pages = disable_large_pages; 1057 disable_auto_text_large_pages = disable_large_pages; 1058 1059 /* 1060 * Initialize mmu-specific large page sizes. 1061 */ 1062 if (&mmu_large_pages_disabled) { 1063 disable_large_pages |= mmu_large_pages_disabled(HAT_LOAD); 1064 disable_ism_large_pages |= 1065 mmu_large_pages_disabled(HAT_LOAD_SHARE); 1066 disable_auto_data_large_pages |= 1067 mmu_large_pages_disabled(HAT_AUTO_DATA); 1068 disable_auto_text_large_pages |= 1069 mmu_large_pages_disabled(HAT_AUTO_TEXT); 1070 } 1071 } 1072 1073 /* 1074 * Initialize the hardware address translation structures. 1075 */ 1076 void 1077 hat_init(void) 1078 { 1079 int i; 1080 uint_t sz; 1081 size_t size; 1082 1083 hat_lock_init(); 1084 hat_kstat_init(); 1085 1086 /* 1087 * Hardware-only bits in a TTE 1088 */ 1089 MAKE_TTE_MASK(&hw_tte); 1090 1091 hat_init_pagesizes(); 1092 1093 /* Initialize the hash locks */ 1094 for (i = 0; i < khmehash_num; i++) { 1095 mutex_init(&khme_hash[i].hmehash_mutex, NULL, 1096 MUTEX_DEFAULT, NULL); 1097 khme_hash[i].hmeh_nextpa = HMEBLK_ENDPA; 1098 } 1099 for (i = 0; i < uhmehash_num; i++) { 1100 mutex_init(&uhme_hash[i].hmehash_mutex, NULL, 1101 MUTEX_DEFAULT, NULL); 1102 uhme_hash[i].hmeh_nextpa = HMEBLK_ENDPA; 1103 } 1104 khmehash_num--; /* make sure counter starts from 0 */ 1105 uhmehash_num--; /* make sure counter starts from 0 */ 1106 1107 /* 1108 * Allocate context domain structures. 1109 * 1110 * A platform may choose to modify max_mmu_ctxdoms in 1111 * set_platform_defaults(). If a platform does not define 1112 * a set_platform_defaults() or does not choose to modify 1113 * max_mmu_ctxdoms, it gets one MMU context domain for every CPU. 1114 * 1115 * For sun4v, there will be one global context domain, this is to 1116 * avoid the ldom cpu substitution problem. 1117 * 1118 * For all platforms that have CPUs sharing MMUs, this 1119 * value must be defined. 1120 */ 1121 if (max_mmu_ctxdoms == 0) { 1122 #ifndef sun4v 1123 max_mmu_ctxdoms = max_ncpus; 1124 #else /* sun4v */ 1125 max_mmu_ctxdoms = 1; 1126 #endif /* sun4v */ 1127 } 1128 1129 size = max_mmu_ctxdoms * sizeof (mmu_ctx_t *); 1130 mmu_ctxs_tbl = kmem_zalloc(size, KM_SLEEP); 1131 1132 /* mmu_ctx_t is 64 bytes aligned */ 1133 mmuctxdom_cache = kmem_cache_create("mmuctxdom_cache", 1134 sizeof (mmu_ctx_t), 64, NULL, NULL, NULL, NULL, NULL, 0); 1135 /* 1136 * MMU context domain initialization for the Boot CPU. 1137 * This needs the context domains array allocated above. 1138 */ 1139 mutex_enter(&cpu_lock); 1140 sfmmu_cpu_init(CPU); 1141 mutex_exit(&cpu_lock); 1142 1143 /* 1144 * Intialize ism mapping list lock. 1145 */ 1146 1147 mutex_init(&ism_mlist_lock, NULL, MUTEX_DEFAULT, NULL); 1148 1149 /* 1150 * Each sfmmu structure carries an array of MMU context info 1151 * structures, one per context domain. The size of this array depends 1152 * on the maximum number of context domains. So, the size of the 1153 * sfmmu structure varies per platform. 1154 * 1155 * sfmmu is allocated from static arena, because trap 1156 * handler at TL > 0 is not allowed to touch kernel relocatable 1157 * memory. sfmmu's alignment is changed to 64 bytes from 1158 * default 8 bytes, as the lower 6 bits will be used to pass 1159 * pgcnt to vtag_flush_pgcnt_tl1. 1160 */ 1161 size = sizeof (sfmmu_t) + sizeof (sfmmu_ctx_t) * (max_mmu_ctxdoms - 1); 1162 1163 sfmmuid_cache = kmem_cache_create("sfmmuid_cache", size, 1164 64, sfmmu_idcache_constructor, sfmmu_idcache_destructor, 1165 NULL, NULL, static_arena, 0); 1166 1167 sfmmu_tsbinfo_cache = kmem_cache_create("sfmmu_tsbinfo_cache", 1168 sizeof (struct tsb_info), 0, NULL, NULL, NULL, NULL, NULL, 0); 1169 1170 /* 1171 * Since we only use the tsb8k cache to "borrow" pages for TSBs 1172 * from the heap when low on memory or when TSB_FORCEALLOC is 1173 * specified, don't use magazines to cache them--we want to return 1174 * them to the system as quickly as possible. 1175 */ 1176 sfmmu_tsb8k_cache = kmem_cache_create("sfmmu_tsb8k_cache", 1177 MMU_PAGESIZE, MMU_PAGESIZE, NULL, NULL, NULL, NULL, 1178 static_arena, KMC_NOMAGAZINE); 1179 1180 /* 1181 * Set tsb_alloc_hiwater to 1/tsb_alloc_hiwater_factor of physical 1182 * memory, which corresponds to the old static reserve for TSBs. 1183 * tsb_alloc_hiwater_factor defaults to 32. This caps the amount of 1184 * memory we'll allocate for TSB slabs; beyond this point TSB 1185 * allocations will be taken from the kernel heap (via 1186 * sfmmu_tsb8k_cache) and will be throttled as would any other kmem 1187 * consumer. 1188 */ 1189 if (tsb_alloc_hiwater_factor == 0) { 1190 tsb_alloc_hiwater_factor = TSB_ALLOC_HIWATER_FACTOR_DEFAULT; 1191 } 1192 SFMMU_SET_TSB_ALLOC_HIWATER(physmem); 1193 1194 for (sz = tsb_slab_ttesz; sz > 0; sz--) { 1195 if (!(disable_large_pages & (1 << sz))) 1196 break; 1197 } 1198 1199 if (sz < tsb_slab_ttesz) { 1200 tsb_slab_ttesz = sz; 1201 tsb_slab_shift = MMU_PAGESHIFT + (sz << 1) + sz; 1202 tsb_slab_size = 1 << tsb_slab_shift; 1203 tsb_slab_mask = (1 << (tsb_slab_shift - MMU_PAGESHIFT)) - 1; 1204 use_bigtsb_arena = 0; 1205 } else if (use_bigtsb_arena && 1206 (disable_large_pages & (1 << bigtsb_slab_ttesz))) { 1207 use_bigtsb_arena = 0; 1208 } 1209 1210 if (!use_bigtsb_arena) { 1211 bigtsb_slab_shift = tsb_slab_shift; 1212 } 1213 SFMMU_SET_TSB_MAX_GROWSIZE(physmem); 1214 1215 /* 1216 * On smaller memory systems, allocate TSB memory in smaller chunks 1217 * than the default 4M slab size. We also honor disable_large_pages 1218 * here. 1219 * 1220 * The trap handlers need to be patched with the final slab shift, 1221 * since they need to be able to construct the TSB pointer at runtime. 1222 */ 1223 if ((tsb_max_growsize <= TSB_512K_SZCODE) && 1224 !(disable_large_pages & (1 << TTE512K))) { 1225 tsb_slab_ttesz = TTE512K; 1226 tsb_slab_shift = MMU_PAGESHIFT512K; 1227 tsb_slab_size = MMU_PAGESIZE512K; 1228 tsb_slab_mask = MMU_PAGEOFFSET512K >> MMU_PAGESHIFT; 1229 use_bigtsb_arena = 0; 1230 } 1231 1232 if (!use_bigtsb_arena) { 1233 bigtsb_slab_ttesz = tsb_slab_ttesz; 1234 bigtsb_slab_shift = tsb_slab_shift; 1235 bigtsb_slab_size = tsb_slab_size; 1236 bigtsb_slab_mask = tsb_slab_mask; 1237 } 1238 1239 1240 /* 1241 * Set up memory callback to update tsb_alloc_hiwater and 1242 * tsb_max_growsize. 1243 */ 1244 i = kphysm_setup_func_register(&sfmmu_update_vec, (void *) 0); 1245 ASSERT(i == 0); 1246 1247 /* 1248 * kmem_tsb_arena is the source from which large TSB slabs are 1249 * drawn. The quantum of this arena corresponds to the largest 1250 * TSB size we can dynamically allocate for user processes. 1251 * Currently it must also be a supported page size since we 1252 * use exactly one translation entry to map each slab page. 1253 * 1254 * The per-lgroup kmem_tsb_default_arena arenas are the arenas from 1255 * which most TSBs are allocated. Since most TSB allocations are 1256 * typically 8K we have a kmem cache we stack on top of each 1257 * kmem_tsb_default_arena to speed up those allocations. 1258 * 1259 * Note the two-level scheme of arenas is required only 1260 * because vmem_create doesn't allow us to specify alignment 1261 * requirements. If this ever changes the code could be 1262 * simplified to use only one level of arenas. 1263 * 1264 * If 256M page support exists on sun4v, 256MB kmem_bigtsb_arena 1265 * will be provided in addition to the 4M kmem_tsb_arena. 1266 */ 1267 if (use_bigtsb_arena) { 1268 kmem_bigtsb_arena = vmem_create("kmem_bigtsb", NULL, 0, 1269 bigtsb_slab_size, sfmmu_vmem_xalloc_aligned_wrapper, 1270 vmem_xfree, heap_arena, 0, VM_SLEEP); 1271 } 1272 1273 kmem_tsb_arena = vmem_create("kmem_tsb", NULL, 0, tsb_slab_size, 1274 sfmmu_vmem_xalloc_aligned_wrapper, 1275 vmem_xfree, heap_arena, 0, VM_SLEEP); 1276 1277 if (tsb_lgrp_affinity) { 1278 char s[50]; 1279 for (i = 0; i < NLGRPS_MAX; i++) { 1280 if (use_bigtsb_arena) { 1281 (void) sprintf(s, "kmem_bigtsb_lgrp%d", i); 1282 kmem_bigtsb_default_arena[i] = vmem_create(s, 1283 NULL, 0, 2 * tsb_slab_size, 1284 sfmmu_tsb_segkmem_alloc, 1285 sfmmu_tsb_segkmem_free, kmem_bigtsb_arena, 1286 0, VM_SLEEP | VM_BESTFIT); 1287 } 1288 1289 (void) sprintf(s, "kmem_tsb_lgrp%d", i); 1290 kmem_tsb_default_arena[i] = vmem_create(s, 1291 NULL, 0, PAGESIZE, sfmmu_tsb_segkmem_alloc, 1292 sfmmu_tsb_segkmem_free, kmem_tsb_arena, 0, 1293 VM_SLEEP | VM_BESTFIT); 1294 1295 (void) sprintf(s, "sfmmu_tsb_lgrp%d_cache", i); 1296 sfmmu_tsb_cache[i] = kmem_cache_create(s, 1297 PAGESIZE, PAGESIZE, NULL, NULL, NULL, NULL, 1298 kmem_tsb_default_arena[i], 0); 1299 } 1300 } else { 1301 if (use_bigtsb_arena) { 1302 kmem_bigtsb_default_arena[0] = 1303 vmem_create("kmem_bigtsb_default", NULL, 0, 1304 2 * tsb_slab_size, sfmmu_tsb_segkmem_alloc, 1305 sfmmu_tsb_segkmem_free, kmem_bigtsb_arena, 0, 1306 VM_SLEEP | VM_BESTFIT); 1307 } 1308 1309 kmem_tsb_default_arena[0] = vmem_create("kmem_tsb_default", 1310 NULL, 0, PAGESIZE, sfmmu_tsb_segkmem_alloc, 1311 sfmmu_tsb_segkmem_free, kmem_tsb_arena, 0, 1312 VM_SLEEP | VM_BESTFIT); 1313 sfmmu_tsb_cache[0] = kmem_cache_create("sfmmu_tsb_cache", 1314 PAGESIZE, PAGESIZE, NULL, NULL, NULL, NULL, 1315 kmem_tsb_default_arena[0], 0); 1316 } 1317 1318 sfmmu8_cache = kmem_cache_create("sfmmu8_cache", HME8BLK_SZ, 1319 HMEBLK_ALIGN, sfmmu_hblkcache_constructor, 1320 sfmmu_hblkcache_destructor, 1321 sfmmu_hblkcache_reclaim, (void *)HME8BLK_SZ, 1322 hat_memload_arena, KMC_NOHASH); 1323 1324 hat_memload1_arena = vmem_create("hat_memload1", NULL, 0, PAGESIZE, 1325 segkmem_alloc_permanent, segkmem_free, heap_arena, 0, VM_SLEEP); 1326 1327 sfmmu1_cache = kmem_cache_create("sfmmu1_cache", HME1BLK_SZ, 1328 HMEBLK_ALIGN, sfmmu_hblkcache_constructor, 1329 sfmmu_hblkcache_destructor, 1330 NULL, (void *)HME1BLK_SZ, 1331 hat_memload1_arena, KMC_NOHASH); 1332 1333 pa_hment_cache = kmem_cache_create("pa_hment_cache", PAHME_SZ, 1334 0, NULL, NULL, NULL, NULL, static_arena, KMC_NOHASH); 1335 1336 ism_blk_cache = kmem_cache_create("ism_blk_cache", 1337 sizeof (ism_blk_t), ecache_alignsize, NULL, NULL, 1338 NULL, NULL, static_arena, KMC_NOHASH); 1339 1340 ism_ment_cache = kmem_cache_create("ism_ment_cache", 1341 sizeof (ism_ment_t), 0, NULL, NULL, 1342 NULL, NULL, NULL, 0); 1343 1344 /* 1345 * We grab the first hat for the kernel, 1346 */ 1347 AS_LOCK_ENTER(&kas, &kas.a_lock, RW_WRITER); 1348 kas.a_hat = hat_alloc(&kas); 1349 AS_LOCK_EXIT(&kas, &kas.a_lock); 1350 1351 /* 1352 * Initialize hblk_reserve. 1353 */ 1354 ((struct hme_blk *)hblk_reserve)->hblk_nextpa = 1355 va_to_pa((caddr_t)hblk_reserve); 1356 1357 #ifndef UTSB_PHYS 1358 /* 1359 * Reserve some kernel virtual address space for the locked TTEs 1360 * that allow us to probe the TSB from TL>0. 1361 */ 1362 utsb_vabase = vmem_xalloc(heap_arena, tsb_slab_size, tsb_slab_size, 1363 0, 0, NULL, NULL, VM_SLEEP); 1364 utsb4m_vabase = vmem_xalloc(heap_arena, tsb_slab_size, tsb_slab_size, 1365 0, 0, NULL, NULL, VM_SLEEP); 1366 #endif 1367 1368 #ifdef VAC 1369 /* 1370 * The big page VAC handling code assumes VAC 1371 * will not be bigger than the smallest big 1372 * page- which is 64K. 1373 */ 1374 if (TTEPAGES(TTE64K) < CACHE_NUM_COLOR) { 1375 cmn_err(CE_PANIC, "VAC too big!"); 1376 } 1377 #endif 1378 1379 (void) xhat_init(); 1380 1381 uhme_hash_pa = va_to_pa(uhme_hash); 1382 khme_hash_pa = va_to_pa(khme_hash); 1383 1384 /* 1385 * Initialize relocation locks. kpr_suspendlock is held 1386 * at PIL_MAX to prevent interrupts from pinning the holder 1387 * of a suspended TTE which may access it leading to a 1388 * deadlock condition. 1389 */ 1390 mutex_init(&kpr_mutex, NULL, MUTEX_DEFAULT, NULL); 1391 mutex_init(&kpr_suspendlock, NULL, MUTEX_SPIN, (void *)PIL_MAX); 1392 1393 /* 1394 * If Shared context support is disabled via /etc/system 1395 * set shctx_on to 0 here if it was set to 1 earlier in boot 1396 * sequence by cpu module initialization code. 1397 */ 1398 if (shctx_on && disable_shctx) { 1399 shctx_on = 0; 1400 } 1401 1402 if (shctx_on) { 1403 srd_buckets = kmem_zalloc(SFMMU_MAX_SRD_BUCKETS * 1404 sizeof (srd_buckets[0]), KM_SLEEP); 1405 for (i = 0; i < SFMMU_MAX_SRD_BUCKETS; i++) { 1406 mutex_init(&srd_buckets[i].srdb_lock, NULL, 1407 MUTEX_DEFAULT, NULL); 1408 } 1409 1410 srd_cache = kmem_cache_create("srd_cache", sizeof (sf_srd_t), 1411 0, sfmmu_srdcache_constructor, sfmmu_srdcache_destructor, 1412 NULL, NULL, NULL, 0); 1413 region_cache = kmem_cache_create("region_cache", 1414 sizeof (sf_region_t), 0, sfmmu_rgncache_constructor, 1415 sfmmu_rgncache_destructor, NULL, NULL, NULL, 0); 1416 scd_cache = kmem_cache_create("scd_cache", sizeof (sf_scd_t), 1417 0, sfmmu_scdcache_constructor, sfmmu_scdcache_destructor, 1418 NULL, NULL, NULL, 0); 1419 } 1420 1421 /* 1422 * Pre-allocate hrm_hashtab before enabling the collection of 1423 * refmod statistics. Allocating on the fly would mean us 1424 * running the risk of suffering recursive mutex enters or 1425 * deadlocks. 1426 */ 1427 hrm_hashtab = kmem_zalloc(HRM_HASHSIZE * sizeof (struct hrmstat *), 1428 KM_SLEEP); 1429 1430 /* Allocate per-cpu pending freelist of hmeblks */ 1431 cpu_hme_pend = kmem_zalloc((NCPU * sizeof (cpu_hme_pend_t)) + 64, 1432 KM_SLEEP); 1433 cpu_hme_pend = (cpu_hme_pend_t *)P2ROUNDUP( 1434 (uintptr_t)cpu_hme_pend, 64); 1435 1436 for (i = 0; i < NCPU; i++) { 1437 mutex_init(&cpu_hme_pend[i].chp_mutex, NULL, MUTEX_DEFAULT, 1438 NULL); 1439 } 1440 1441 if (cpu_hme_pend_thresh == 0) { 1442 cpu_hme_pend_thresh = CPU_HME_PEND_THRESH; 1443 } 1444 } 1445 1446 /* 1447 * Initialize locking for the hat layer, called early during boot. 1448 */ 1449 static void 1450 hat_lock_init() 1451 { 1452 int i; 1453 1454 /* 1455 * initialize the array of mutexes protecting a page's mapping 1456 * list and p_nrm field. 1457 */ 1458 for (i = 0; i < mml_table_sz; i++) 1459 mutex_init(&mml_table[i], NULL, MUTEX_DEFAULT, NULL); 1460 1461 if (kpm_enable) { 1462 for (i = 0; i < kpmp_table_sz; i++) { 1463 mutex_init(&kpmp_table[i].khl_mutex, NULL, 1464 MUTEX_DEFAULT, NULL); 1465 } 1466 } 1467 1468 /* 1469 * Initialize array of mutex locks that protects sfmmu fields and 1470 * TSB lists. 1471 */ 1472 for (i = 0; i < SFMMU_NUM_LOCK; i++) 1473 mutex_init(HATLOCK_MUTEXP(&hat_lock[i]), NULL, MUTEX_DEFAULT, 1474 NULL); 1475 } 1476 1477 #define SFMMU_KERNEL_MAXVA \ 1478 (kmem64_base ? (uintptr_t)kmem64_end : (SYSLIMIT)) 1479 1480 /* 1481 * Allocate a hat structure. 1482 * Called when an address space first uses a hat. 1483 */ 1484 struct hat * 1485 hat_alloc(struct as *as) 1486 { 1487 sfmmu_t *sfmmup; 1488 int i; 1489 uint64_t cnum; 1490 extern uint_t get_color_start(struct as *); 1491 1492 ASSERT(AS_WRITE_HELD(as, &as->a_lock)); 1493 sfmmup = kmem_cache_alloc(sfmmuid_cache, KM_SLEEP); 1494 sfmmup->sfmmu_as = as; 1495 sfmmup->sfmmu_flags = 0; 1496 sfmmup->sfmmu_tteflags = 0; 1497 sfmmup->sfmmu_rtteflags = 0; 1498 LOCK_INIT_CLEAR(&sfmmup->sfmmu_ctx_lock); 1499 1500 if (as == &kas) { 1501 ksfmmup = sfmmup; 1502 sfmmup->sfmmu_cext = 0; 1503 cnum = KCONTEXT; 1504 1505 sfmmup->sfmmu_clrstart = 0; 1506 sfmmup->sfmmu_tsb = NULL; 1507 /* 1508 * hat_kern_setup() will call sfmmu_init_ktsbinfo() 1509 * to setup tsb_info for ksfmmup. 1510 */ 1511 } else { 1512 1513 /* 1514 * Just set to invalid ctx. When it faults, it will 1515 * get a valid ctx. This would avoid the situation 1516 * where we get a ctx, but it gets stolen and then 1517 * we fault when we try to run and so have to get 1518 * another ctx. 1519 */ 1520 sfmmup->sfmmu_cext = 0; 1521 cnum = INVALID_CONTEXT; 1522 1523 /* initialize original physical page coloring bin */ 1524 sfmmup->sfmmu_clrstart = get_color_start(as); 1525 #ifdef DEBUG 1526 if (tsb_random_size) { 1527 uint32_t randval = (uint32_t)gettick() >> 4; 1528 int size = randval % (tsb_max_growsize + 1); 1529 1530 /* chose a random tsb size for stress testing */ 1531 (void) sfmmu_tsbinfo_alloc(&sfmmup->sfmmu_tsb, size, 1532 TSB8K|TSB64K|TSB512K, 0, sfmmup); 1533 } else 1534 #endif /* DEBUG */ 1535 (void) sfmmu_tsbinfo_alloc(&sfmmup->sfmmu_tsb, 1536 default_tsb_size, 1537 TSB8K|TSB64K|TSB512K, 0, sfmmup); 1538 sfmmup->sfmmu_flags = HAT_SWAPPED | HAT_ALLCTX_INVALID; 1539 ASSERT(sfmmup->sfmmu_tsb != NULL); 1540 } 1541 1542 ASSERT(max_mmu_ctxdoms > 0); 1543 for (i = 0; i < max_mmu_ctxdoms; i++) { 1544 sfmmup->sfmmu_ctxs[i].cnum = cnum; 1545 sfmmup->sfmmu_ctxs[i].gnum = 0; 1546 } 1547 1548 for (i = 0; i < max_mmu_page_sizes; i++) { 1549 sfmmup->sfmmu_ttecnt[i] = 0; 1550 sfmmup->sfmmu_scdrttecnt[i] = 0; 1551 sfmmup->sfmmu_ismttecnt[i] = 0; 1552 sfmmup->sfmmu_scdismttecnt[i] = 0; 1553 sfmmup->sfmmu_pgsz[i] = TTE8K; 1554 } 1555 sfmmup->sfmmu_tsb0_4minflcnt = 0; 1556 sfmmup->sfmmu_iblk = NULL; 1557 sfmmup->sfmmu_ismhat = 0; 1558 sfmmup->sfmmu_scdhat = 0; 1559 sfmmup->sfmmu_ismblkpa = (uint64_t)-1; 1560 if (sfmmup == ksfmmup) { 1561 CPUSET_ALL(sfmmup->sfmmu_cpusran); 1562 } else { 1563 CPUSET_ZERO(sfmmup->sfmmu_cpusran); 1564 } 1565 sfmmup->sfmmu_free = 0; 1566 sfmmup->sfmmu_rmstat = 0; 1567 sfmmup->sfmmu_clrbin = sfmmup->sfmmu_clrstart; 1568 sfmmup->sfmmu_xhat_provider = NULL; 1569 cv_init(&sfmmup->sfmmu_tsb_cv, NULL, CV_DEFAULT, NULL); 1570 sfmmup->sfmmu_srdp = NULL; 1571 SF_RGNMAP_ZERO(sfmmup->sfmmu_region_map); 1572 bzero(sfmmup->sfmmu_hmeregion_links, SFMMU_L1_HMERLINKS_SIZE); 1573 sfmmup->sfmmu_scdp = NULL; 1574 sfmmup->sfmmu_scd_link.next = NULL; 1575 sfmmup->sfmmu_scd_link.prev = NULL; 1576 return (sfmmup); 1577 } 1578 1579 /* 1580 * Create per-MMU context domain kstats for a given MMU ctx. 1581 */ 1582 static void 1583 sfmmu_mmu_kstat_create(mmu_ctx_t *mmu_ctxp) 1584 { 1585 mmu_ctx_stat_t stat; 1586 kstat_t *mmu_kstat; 1587 1588 ASSERT(MUTEX_HELD(&cpu_lock)); 1589 ASSERT(mmu_ctxp->mmu_kstat == NULL); 1590 1591 mmu_kstat = kstat_create("unix", mmu_ctxp->mmu_idx, "mmu_ctx", 1592 "hat", KSTAT_TYPE_NAMED, MMU_CTX_NUM_STATS, KSTAT_FLAG_VIRTUAL); 1593 1594 if (mmu_kstat == NULL) { 1595 cmn_err(CE_WARN, "kstat_create for MMU %d failed", 1596 mmu_ctxp->mmu_idx); 1597 } else { 1598 mmu_kstat->ks_data = mmu_ctxp->mmu_kstat_data; 1599 for (stat = 0; stat < MMU_CTX_NUM_STATS; stat++) 1600 kstat_named_init(&mmu_ctxp->mmu_kstat_data[stat], 1601 mmu_ctx_kstat_names[stat], KSTAT_DATA_INT64); 1602 mmu_ctxp->mmu_kstat = mmu_kstat; 1603 kstat_install(mmu_kstat); 1604 } 1605 } 1606 1607 /* 1608 * plat_cpuid_to_mmu_ctx_info() is a platform interface that returns MMU 1609 * context domain information for a given CPU. If a platform does not 1610 * specify that interface, then the function below is used instead to return 1611 * default information. The defaults are as follows: 1612 * 1613 * - For sun4u systems there's one MMU context domain per CPU. 1614 * This default is used by all sun4u systems except OPL. OPL systems 1615 * provide platform specific interface to map CPU ids to MMU ids 1616 * because on OPL more than 1 CPU shares a single MMU. 1617 * Note that on sun4v, there is one global context domain for 1618 * the entire system. This is to avoid running into potential problem 1619 * with ldom physical cpu substitution feature. 1620 * - The number of MMU context IDs supported on any CPU in the 1621 * system is 8K. 1622 */ 1623 /*ARGSUSED*/ 1624 static void 1625 sfmmu_cpuid_to_mmu_ctx_info(processorid_t cpuid, mmu_ctx_info_t *infop) 1626 { 1627 infop->mmu_nctxs = nctxs; 1628 #ifndef sun4v 1629 infop->mmu_idx = cpu[cpuid]->cpu_seqid; 1630 #else /* sun4v */ 1631 infop->mmu_idx = 0; 1632 #endif /* sun4v */ 1633 } 1634 1635 /* 1636 * Called during CPU initialization to set the MMU context-related information 1637 * for a CPU. 1638 * 1639 * cpu_lock serializes accesses to mmu_ctxs and mmu_saved_gnum. 1640 */ 1641 void 1642 sfmmu_cpu_init(cpu_t *cp) 1643 { 1644 mmu_ctx_info_t info; 1645 mmu_ctx_t *mmu_ctxp; 1646 1647 ASSERT(MUTEX_HELD(&cpu_lock)); 1648 1649 if (&plat_cpuid_to_mmu_ctx_info == NULL) 1650 sfmmu_cpuid_to_mmu_ctx_info(cp->cpu_id, &info); 1651 else 1652 plat_cpuid_to_mmu_ctx_info(cp->cpu_id, &info); 1653 1654 ASSERT(info.mmu_idx < max_mmu_ctxdoms); 1655 1656 if ((mmu_ctxp = mmu_ctxs_tbl[info.mmu_idx]) == NULL) { 1657 /* Each mmu_ctx is cacheline aligned. */ 1658 mmu_ctxp = kmem_cache_alloc(mmuctxdom_cache, KM_SLEEP); 1659 bzero(mmu_ctxp, sizeof (mmu_ctx_t)); 1660 1661 mutex_init(&mmu_ctxp->mmu_lock, NULL, MUTEX_SPIN, 1662 (void *)ipltospl(DISP_LEVEL)); 1663 mmu_ctxp->mmu_idx = info.mmu_idx; 1664 mmu_ctxp->mmu_nctxs = info.mmu_nctxs; 1665 /* 1666 * Globally for lifetime of a system, 1667 * gnum must always increase. 1668 * mmu_saved_gnum is protected by the cpu_lock. 1669 */ 1670 mmu_ctxp->mmu_gnum = mmu_saved_gnum + 1; 1671 mmu_ctxp->mmu_cnum = NUM_LOCKED_CTXS; 1672 1673 sfmmu_mmu_kstat_create(mmu_ctxp); 1674 1675 mmu_ctxs_tbl[info.mmu_idx] = mmu_ctxp; 1676 } else { 1677 ASSERT(mmu_ctxp->mmu_idx == info.mmu_idx); 1678 } 1679 1680 /* 1681 * The mmu_lock is acquired here to prevent races with 1682 * the wrap-around code. 1683 */ 1684 mutex_enter(&mmu_ctxp->mmu_lock); 1685 1686 1687 mmu_ctxp->mmu_ncpus++; 1688 CPUSET_ADD(mmu_ctxp->mmu_cpuset, cp->cpu_id); 1689 CPU_MMU_IDX(cp) = info.mmu_idx; 1690 CPU_MMU_CTXP(cp) = mmu_ctxp; 1691 1692 mutex_exit(&mmu_ctxp->mmu_lock); 1693 } 1694 1695 /* 1696 * Called to perform MMU context-related cleanup for a CPU. 1697 */ 1698 void 1699 sfmmu_cpu_cleanup(cpu_t *cp) 1700 { 1701 mmu_ctx_t *mmu_ctxp; 1702 1703 ASSERT(MUTEX_HELD(&cpu_lock)); 1704 1705 mmu_ctxp = CPU_MMU_CTXP(cp); 1706 ASSERT(mmu_ctxp != NULL); 1707 1708 /* 1709 * The mmu_lock is acquired here to prevent races with 1710 * the wrap-around code. 1711 */ 1712 mutex_enter(&mmu_ctxp->mmu_lock); 1713 1714 CPU_MMU_CTXP(cp) = NULL; 1715 1716 CPUSET_DEL(mmu_ctxp->mmu_cpuset, cp->cpu_id); 1717 if (--mmu_ctxp->mmu_ncpus == 0) { 1718 mmu_ctxs_tbl[mmu_ctxp->mmu_idx] = NULL; 1719 mutex_exit(&mmu_ctxp->mmu_lock); 1720 mutex_destroy(&mmu_ctxp->mmu_lock); 1721 1722 if (mmu_ctxp->mmu_kstat) 1723 kstat_delete(mmu_ctxp->mmu_kstat); 1724 1725 /* mmu_saved_gnum is protected by the cpu_lock. */ 1726 if (mmu_saved_gnum < mmu_ctxp->mmu_gnum) 1727 mmu_saved_gnum = mmu_ctxp->mmu_gnum; 1728 1729 kmem_cache_free(mmuctxdom_cache, mmu_ctxp); 1730 1731 return; 1732 } 1733 1734 mutex_exit(&mmu_ctxp->mmu_lock); 1735 } 1736 1737 /* 1738 * Hat_setup, makes an address space context the current active one. 1739 * In sfmmu this translates to setting the secondary context with the 1740 * corresponding context. 1741 */ 1742 void 1743 hat_setup(struct hat *sfmmup, int allocflag) 1744 { 1745 hatlock_t *hatlockp; 1746 1747 /* Init needs some special treatment. */ 1748 if (allocflag == HAT_INIT) { 1749 /* 1750 * Make sure that we have 1751 * 1. a TSB 1752 * 2. a valid ctx that doesn't get stolen after this point. 1753 */ 1754 hatlockp = sfmmu_hat_enter(sfmmup); 1755 1756 /* 1757 * Swap in the TSB. hat_init() allocates tsbinfos without 1758 * TSBs, but we need one for init, since the kernel does some 1759 * special things to set up its stack and needs the TSB to 1760 * resolve page faults. 1761 */ 1762 sfmmu_tsb_swapin(sfmmup, hatlockp); 1763 1764 sfmmu_get_ctx(sfmmup); 1765 1766 sfmmu_hat_exit(hatlockp); 1767 } else { 1768 ASSERT(allocflag == HAT_ALLOC); 1769 1770 hatlockp = sfmmu_hat_enter(sfmmup); 1771 kpreempt_disable(); 1772 1773 CPUSET_ADD(sfmmup->sfmmu_cpusran, CPU->cpu_id); 1774 /* 1775 * sfmmu_setctx_sec takes <pgsz|cnum> as a parameter, 1776 * pagesize bits don't matter in this case since we are passing 1777 * INVALID_CONTEXT to it. 1778 * Compatibility Note: hw takes care of MMU_SCONTEXT1 1779 */ 1780 sfmmu_setctx_sec(INVALID_CONTEXT); 1781 sfmmu_clear_utsbinfo(); 1782 1783 kpreempt_enable(); 1784 sfmmu_hat_exit(hatlockp); 1785 } 1786 } 1787 1788 /* 1789 * Free all the translation resources for the specified address space. 1790 * Called from as_free when an address space is being destroyed. 1791 */ 1792 void 1793 hat_free_start(struct hat *sfmmup) 1794 { 1795 ASSERT(AS_WRITE_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock)); 1796 ASSERT(sfmmup != ksfmmup); 1797 ASSERT(sfmmup->sfmmu_xhat_provider == NULL); 1798 1799 sfmmup->sfmmu_free = 1; 1800 if (sfmmup->sfmmu_scdp != NULL) { 1801 sfmmu_leave_scd(sfmmup, 0); 1802 } 1803 1804 ASSERT(sfmmup->sfmmu_scdp == NULL); 1805 } 1806 1807 void 1808 hat_free_end(struct hat *sfmmup) 1809 { 1810 int i; 1811 1812 ASSERT(sfmmup->sfmmu_xhat_provider == NULL); 1813 ASSERT(sfmmup->sfmmu_free == 1); 1814 ASSERT(sfmmup->sfmmu_ttecnt[TTE8K] == 0); 1815 ASSERT(sfmmup->sfmmu_ttecnt[TTE64K] == 0); 1816 ASSERT(sfmmup->sfmmu_ttecnt[TTE512K] == 0); 1817 ASSERT(sfmmup->sfmmu_ttecnt[TTE4M] == 0); 1818 ASSERT(sfmmup->sfmmu_ttecnt[TTE32M] == 0); 1819 ASSERT(sfmmup->sfmmu_ttecnt[TTE256M] == 0); 1820 1821 if (sfmmup->sfmmu_rmstat) { 1822 hat_freestat(sfmmup->sfmmu_as, NULL); 1823 } 1824 1825 while (sfmmup->sfmmu_tsb != NULL) { 1826 struct tsb_info *next = sfmmup->sfmmu_tsb->tsb_next; 1827 sfmmu_tsbinfo_free(sfmmup->sfmmu_tsb); 1828 sfmmup->sfmmu_tsb = next; 1829 } 1830 1831 if (sfmmup->sfmmu_srdp != NULL) { 1832 sfmmu_leave_srd(sfmmup); 1833 ASSERT(sfmmup->sfmmu_srdp == NULL); 1834 for (i = 0; i < SFMMU_L1_HMERLINKS; i++) { 1835 if (sfmmup->sfmmu_hmeregion_links[i] != NULL) { 1836 kmem_free(sfmmup->sfmmu_hmeregion_links[i], 1837 SFMMU_L2_HMERLINKS_SIZE); 1838 sfmmup->sfmmu_hmeregion_links[i] = NULL; 1839 } 1840 } 1841 } 1842 sfmmu_free_sfmmu(sfmmup); 1843 1844 #ifdef DEBUG 1845 for (i = 0; i < SFMMU_L1_HMERLINKS; i++) { 1846 ASSERT(sfmmup->sfmmu_hmeregion_links[i] == NULL); 1847 } 1848 #endif 1849 1850 kmem_cache_free(sfmmuid_cache, sfmmup); 1851 } 1852 1853 /* 1854 * Set up any translation structures, for the specified address space, 1855 * that are needed or preferred when the process is being swapped in. 1856 */ 1857 /* ARGSUSED */ 1858 void 1859 hat_swapin(struct hat *hat) 1860 { 1861 ASSERT(hat->sfmmu_xhat_provider == NULL); 1862 } 1863 1864 /* 1865 * Free all of the translation resources, for the specified address space, 1866 * that can be freed while the process is swapped out. Called from as_swapout. 1867 * Also, free up the ctx that this process was using. 1868 */ 1869 void 1870 hat_swapout(struct hat *sfmmup) 1871 { 1872 struct hmehash_bucket *hmebp; 1873 struct hme_blk *hmeblkp; 1874 struct hme_blk *pr_hblk = NULL; 1875 struct hme_blk *nx_hblk; 1876 int i; 1877 struct hme_blk *list = NULL; 1878 hatlock_t *hatlockp; 1879 struct tsb_info *tsbinfop; 1880 struct free_tsb { 1881 struct free_tsb *next; 1882 struct tsb_info *tsbinfop; 1883 }; /* free list of TSBs */ 1884 struct free_tsb *freelist, *last, *next; 1885 1886 ASSERT(sfmmup->sfmmu_xhat_provider == NULL); 1887 SFMMU_STAT(sf_swapout); 1888 1889 /* 1890 * There is no way to go from an as to all its translations in sfmmu. 1891 * Here is one of the times when we take the big hit and traverse 1892 * the hash looking for hme_blks to free up. Not only do we free up 1893 * this as hme_blks but all those that are free. We are obviously 1894 * swapping because we need memory so let's free up as much 1895 * as we can. 1896 * 1897 * Note that we don't flush TLB/TSB here -- it's not necessary 1898 * because: 1899 * 1) we free the ctx we're using and throw away the TSB(s); 1900 * 2) processes aren't runnable while being swapped out. 1901 */ 1902 ASSERT(sfmmup != KHATID); 1903 for (i = 0; i <= UHMEHASH_SZ; i++) { 1904 hmebp = &uhme_hash[i]; 1905 SFMMU_HASH_LOCK(hmebp); 1906 hmeblkp = hmebp->hmeblkp; 1907 pr_hblk = NULL; 1908 while (hmeblkp) { 1909 1910 ASSERT(!hmeblkp->hblk_xhat_bit); 1911 1912 if ((hmeblkp->hblk_tag.htag_id == sfmmup) && 1913 !hmeblkp->hblk_shw_bit && !hmeblkp->hblk_lckcnt) { 1914 ASSERT(!hmeblkp->hblk_shared); 1915 (void) sfmmu_hblk_unload(sfmmup, hmeblkp, 1916 (caddr_t)get_hblk_base(hmeblkp), 1917 get_hblk_endaddr(hmeblkp), 1918 NULL, HAT_UNLOAD); 1919 } 1920 nx_hblk = hmeblkp->hblk_next; 1921 if (!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt) { 1922 ASSERT(!hmeblkp->hblk_lckcnt); 1923 sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk, 1924 &list, 0); 1925 } else { 1926 pr_hblk = hmeblkp; 1927 } 1928 hmeblkp = nx_hblk; 1929 } 1930 SFMMU_HASH_UNLOCK(hmebp); 1931 } 1932 1933 sfmmu_hblks_list_purge(&list, 0); 1934 1935 /* 1936 * Now free up the ctx so that others can reuse it. 1937 */ 1938 hatlockp = sfmmu_hat_enter(sfmmup); 1939 1940 sfmmu_invalidate_ctx(sfmmup); 1941 1942 /* 1943 * Free TSBs, but not tsbinfos, and set SWAPPED flag. 1944 * If TSBs were never swapped in, just return. 1945 * This implies that we don't support partial swapping 1946 * of TSBs -- either all are swapped out, or none are. 1947 * 1948 * We must hold the HAT lock here to prevent racing with another 1949 * thread trying to unmap TTEs from the TSB or running the post- 1950 * relocator after relocating the TSB's memory. Unfortunately, we 1951 * can't free memory while holding the HAT lock or we could 1952 * deadlock, so we build a list of TSBs to be freed after marking 1953 * the tsbinfos as swapped out and free them after dropping the 1954 * lock. 1955 */ 1956 if (SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) { 1957 sfmmu_hat_exit(hatlockp); 1958 return; 1959 } 1960 1961 SFMMU_FLAGS_SET(sfmmup, HAT_SWAPPED); 1962 last = freelist = NULL; 1963 for (tsbinfop = sfmmup->sfmmu_tsb; tsbinfop != NULL; 1964 tsbinfop = tsbinfop->tsb_next) { 1965 ASSERT((tsbinfop->tsb_flags & TSB_SWAPPED) == 0); 1966 1967 /* 1968 * Cast the TSB into a struct free_tsb and put it on the free 1969 * list. 1970 */ 1971 if (freelist == NULL) { 1972 last = freelist = (struct free_tsb *)tsbinfop->tsb_va; 1973 } else { 1974 last->next = (struct free_tsb *)tsbinfop->tsb_va; 1975 last = last->next; 1976 } 1977 last->next = NULL; 1978 last->tsbinfop = tsbinfop; 1979 tsbinfop->tsb_flags |= TSB_SWAPPED; 1980 /* 1981 * Zero out the TTE to clear the valid bit. 1982 * Note we can't use a value like 0xbad because we want to 1983 * ensure diagnostic bits are NEVER set on TTEs that might 1984 * be loaded. The intent is to catch any invalid access 1985 * to the swapped TSB, such as a thread running with a valid 1986 * context without first calling sfmmu_tsb_swapin() to 1987 * allocate TSB memory. 1988 */ 1989 tsbinfop->tsb_tte.ll = 0; 1990 } 1991 1992 /* Now we can drop the lock and free the TSB memory. */ 1993 sfmmu_hat_exit(hatlockp); 1994 for (; freelist != NULL; freelist = next) { 1995 next = freelist->next; 1996 sfmmu_tsb_free(freelist->tsbinfop); 1997 } 1998 } 1999 2000 /* 2001 * Duplicate the translations of an as into another newas 2002 */ 2003 /* ARGSUSED */ 2004 int 2005 hat_dup(struct hat *hat, struct hat *newhat, caddr_t addr, size_t len, 2006 uint_t flag) 2007 { 2008 sf_srd_t *srdp; 2009 sf_scd_t *scdp; 2010 int i; 2011 extern uint_t get_color_start(struct as *); 2012 2013 ASSERT(hat->sfmmu_xhat_provider == NULL); 2014 ASSERT((flag == 0) || (flag == HAT_DUP_ALL) || (flag == HAT_DUP_COW) || 2015 (flag == HAT_DUP_SRD)); 2016 ASSERT(hat != ksfmmup); 2017 ASSERT(newhat != ksfmmup); 2018 ASSERT(flag != HAT_DUP_ALL || hat->sfmmu_srdp == newhat->sfmmu_srdp); 2019 2020 if (flag == HAT_DUP_COW) { 2021 panic("hat_dup: HAT_DUP_COW not supported"); 2022 } 2023 2024 if (flag == HAT_DUP_SRD && ((srdp = hat->sfmmu_srdp) != NULL)) { 2025 ASSERT(srdp->srd_evp != NULL); 2026 VN_HOLD(srdp->srd_evp); 2027 ASSERT(srdp->srd_refcnt > 0); 2028 newhat->sfmmu_srdp = srdp; 2029 atomic_add_32((volatile uint_t *)&srdp->srd_refcnt, 1); 2030 } 2031 2032 /* 2033 * HAT_DUP_ALL flag is used after as duplication is done. 2034 */ 2035 if (flag == HAT_DUP_ALL && ((srdp = newhat->sfmmu_srdp) != NULL)) { 2036 ASSERT(newhat->sfmmu_srdp->srd_refcnt >= 2); 2037 newhat->sfmmu_rtteflags = hat->sfmmu_rtteflags; 2038 if (hat->sfmmu_flags & HAT_4MTEXT_FLAG) { 2039 newhat->sfmmu_flags |= HAT_4MTEXT_FLAG; 2040 } 2041 2042 /* check if need to join scd */ 2043 if ((scdp = hat->sfmmu_scdp) != NULL && 2044 newhat->sfmmu_scdp != scdp) { 2045 int ret; 2046 SF_RGNMAP_IS_SUBSET(&newhat->sfmmu_region_map, 2047 &scdp->scd_region_map, ret); 2048 ASSERT(ret); 2049 sfmmu_join_scd(scdp, newhat); 2050 ASSERT(newhat->sfmmu_scdp == scdp && 2051 scdp->scd_refcnt >= 2); 2052 for (i = 0; i < max_mmu_page_sizes; i++) { 2053 newhat->sfmmu_ismttecnt[i] = 2054 hat->sfmmu_ismttecnt[i]; 2055 newhat->sfmmu_scdismttecnt[i] = 2056 hat->sfmmu_scdismttecnt[i]; 2057 } 2058 } 2059 2060 sfmmu_check_page_sizes(newhat, 1); 2061 } 2062 2063 if (flag == HAT_DUP_ALL && consistent_coloring == 0 && 2064 update_proc_pgcolorbase_after_fork != 0) { 2065 hat->sfmmu_clrbin = get_color_start(hat->sfmmu_as); 2066 } 2067 return (0); 2068 } 2069 2070 void 2071 hat_memload(struct hat *hat, caddr_t addr, struct page *pp, 2072 uint_t attr, uint_t flags) 2073 { 2074 hat_do_memload(hat, addr, pp, attr, flags, 2075 SFMMU_INVALID_SHMERID); 2076 } 2077 2078 void 2079 hat_memload_region(struct hat *hat, caddr_t addr, struct page *pp, 2080 uint_t attr, uint_t flags, hat_region_cookie_t rcookie) 2081 { 2082 uint_t rid; 2083 if (rcookie == HAT_INVALID_REGION_COOKIE || 2084 hat->sfmmu_xhat_provider != NULL) { 2085 hat_do_memload(hat, addr, pp, attr, flags, 2086 SFMMU_INVALID_SHMERID); 2087 return; 2088 } 2089 rid = (uint_t)((uint64_t)rcookie); 2090 ASSERT(rid < SFMMU_MAX_HME_REGIONS); 2091 hat_do_memload(hat, addr, pp, attr, flags, rid); 2092 } 2093 2094 /* 2095 * Set up addr to map to page pp with protection prot. 2096 * As an optimization we also load the TSB with the 2097 * corresponding tte but it is no big deal if the tte gets kicked out. 2098 */ 2099 static void 2100 hat_do_memload(struct hat *hat, caddr_t addr, struct page *pp, 2101 uint_t attr, uint_t flags, uint_t rid) 2102 { 2103 tte_t tte; 2104 2105 2106 ASSERT(hat != NULL); 2107 ASSERT(PAGE_LOCKED(pp)); 2108 ASSERT(!((uintptr_t)addr & MMU_PAGEOFFSET)); 2109 ASSERT(!(flags & ~SFMMU_LOAD_ALLFLAG)); 2110 ASSERT(!(attr & ~SFMMU_LOAD_ALLATTR)); 2111 SFMMU_VALIDATE_HMERID(hat, rid, addr, MMU_PAGESIZE); 2112 2113 if (PP_ISFREE(pp)) { 2114 panic("hat_memload: loading a mapping to free page %p", 2115 (void *)pp); 2116 } 2117 2118 if (hat->sfmmu_xhat_provider) { 2119 /* no regions for xhats */ 2120 ASSERT(!SFMMU_IS_SHMERID_VALID(rid)); 2121 XHAT_MEMLOAD(hat, addr, pp, attr, flags); 2122 return; 2123 } 2124 2125 ASSERT((hat == ksfmmup) || 2126 AS_LOCK_HELD(hat->sfmmu_as, &hat->sfmmu_as->a_lock)); 2127 2128 if (flags & ~SFMMU_LOAD_ALLFLAG) 2129 cmn_err(CE_NOTE, "hat_memload: unsupported flags %d", 2130 flags & ~SFMMU_LOAD_ALLFLAG); 2131 2132 if (hat->sfmmu_rmstat) 2133 hat_resvstat(MMU_PAGESIZE, hat->sfmmu_as, addr); 2134 2135 #if defined(SF_ERRATA_57) 2136 if ((hat != ksfmmup) && AS_TYPE_64BIT(hat->sfmmu_as) && 2137 (addr < errata57_limit) && (attr & PROT_EXEC) && 2138 !(flags & HAT_LOAD_SHARE)) { 2139 cmn_err(CE_WARN, "hat_memload: illegal attempt to make user " 2140 " page executable"); 2141 attr &= ~PROT_EXEC; 2142 } 2143 #endif 2144 2145 sfmmu_memtte(&tte, pp->p_pagenum, attr, TTE8K); 2146 (void) sfmmu_tteload_array(hat, &tte, addr, &pp, flags, rid); 2147 2148 /* 2149 * Check TSB and TLB page sizes. 2150 */ 2151 if ((flags & HAT_LOAD_SHARE) == 0) { 2152 sfmmu_check_page_sizes(hat, 1); 2153 } 2154 } 2155 2156 /* 2157 * hat_devload can be called to map real memory (e.g. 2158 * /dev/kmem) and even though hat_devload will determine pf is 2159 * for memory, it will be unable to get a shared lock on the 2160 * page (because someone else has it exclusively) and will 2161 * pass dp = NULL. If tteload doesn't get a non-NULL 2162 * page pointer it can't cache memory. 2163 */ 2164 void 2165 hat_devload(struct hat *hat, caddr_t addr, size_t len, pfn_t pfn, 2166 uint_t attr, int flags) 2167 { 2168 tte_t tte; 2169 struct page *pp = NULL; 2170 int use_lgpg = 0; 2171 2172 ASSERT(hat != NULL); 2173 2174 if (hat->sfmmu_xhat_provider) { 2175 XHAT_DEVLOAD(hat, addr, len, pfn, attr, flags); 2176 return; 2177 } 2178 2179 ASSERT(!(flags & ~SFMMU_LOAD_ALLFLAG)); 2180 ASSERT(!(attr & ~SFMMU_LOAD_ALLATTR)); 2181 ASSERT((hat == ksfmmup) || 2182 AS_LOCK_HELD(hat->sfmmu_as, &hat->sfmmu_as->a_lock)); 2183 if (len == 0) 2184 panic("hat_devload: zero len"); 2185 if (flags & ~SFMMU_LOAD_ALLFLAG) 2186 cmn_err(CE_NOTE, "hat_devload: unsupported flags %d", 2187 flags & ~SFMMU_LOAD_ALLFLAG); 2188 2189 #if defined(SF_ERRATA_57) 2190 if ((hat != ksfmmup) && AS_TYPE_64BIT(hat->sfmmu_as) && 2191 (addr < errata57_limit) && (attr & PROT_EXEC) && 2192 !(flags & HAT_LOAD_SHARE)) { 2193 cmn_err(CE_WARN, "hat_devload: illegal attempt to make user " 2194 " page executable"); 2195 attr &= ~PROT_EXEC; 2196 } 2197 #endif 2198 2199 /* 2200 * If it's a memory page find its pp 2201 */ 2202 if (!(flags & HAT_LOAD_NOCONSIST) && pf_is_memory(pfn)) { 2203 pp = page_numtopp_nolock(pfn); 2204 if (pp == NULL) { 2205 flags |= HAT_LOAD_NOCONSIST; 2206 } else { 2207 if (PP_ISFREE(pp)) { 2208 panic("hat_memload: loading " 2209 "a mapping to free page %p", 2210 (void *)pp); 2211 } 2212 if (!PAGE_LOCKED(pp) && !PP_ISNORELOC(pp)) { 2213 panic("hat_memload: loading a mapping " 2214 "to unlocked relocatable page %p", 2215 (void *)pp); 2216 } 2217 ASSERT(len == MMU_PAGESIZE); 2218 } 2219 } 2220 2221 if (hat->sfmmu_rmstat) 2222 hat_resvstat(len, hat->sfmmu_as, addr); 2223 2224 if (flags & HAT_LOAD_NOCONSIST) { 2225 attr |= SFMMU_UNCACHEVTTE; 2226 use_lgpg = 1; 2227 } 2228 if (!pf_is_memory(pfn)) { 2229 attr |= SFMMU_UNCACHEPTTE | HAT_NOSYNC; 2230 use_lgpg = 1; 2231 switch (attr & HAT_ORDER_MASK) { 2232 case HAT_STRICTORDER: 2233 case HAT_UNORDERED_OK: 2234 /* 2235 * we set the side effect bit for all non 2236 * memory mappings unless merging is ok 2237 */ 2238 attr |= SFMMU_SIDEFFECT; 2239 break; 2240 case HAT_MERGING_OK: 2241 case HAT_LOADCACHING_OK: 2242 case HAT_STORECACHING_OK: 2243 break; 2244 default: 2245 panic("hat_devload: bad attr"); 2246 break; 2247 } 2248 } 2249 while (len) { 2250 if (!use_lgpg) { 2251 sfmmu_memtte(&tte, pfn, attr, TTE8K); 2252 (void) sfmmu_tteload_array(hat, &tte, addr, &pp, 2253 flags, SFMMU_INVALID_SHMERID); 2254 len -= MMU_PAGESIZE; 2255 addr += MMU_PAGESIZE; 2256 pfn++; 2257 continue; 2258 } 2259 /* 2260 * try to use large pages, check va/pa alignments 2261 * Note that 32M/256M page sizes are not (yet) supported. 2262 */ 2263 if ((len >= MMU_PAGESIZE4M) && 2264 !((uintptr_t)addr & MMU_PAGEOFFSET4M) && 2265 !(disable_large_pages & (1 << TTE4M)) && 2266 !(mmu_ptob(pfn) & MMU_PAGEOFFSET4M)) { 2267 sfmmu_memtte(&tte, pfn, attr, TTE4M); 2268 (void) sfmmu_tteload_array(hat, &tte, addr, &pp, 2269 flags, SFMMU_INVALID_SHMERID); 2270 len -= MMU_PAGESIZE4M; 2271 addr += MMU_PAGESIZE4M; 2272 pfn += MMU_PAGESIZE4M / MMU_PAGESIZE; 2273 } else if ((len >= MMU_PAGESIZE512K) && 2274 !((uintptr_t)addr & MMU_PAGEOFFSET512K) && 2275 !(disable_large_pages & (1 << TTE512K)) && 2276 !(mmu_ptob(pfn) & MMU_PAGEOFFSET512K)) { 2277 sfmmu_memtte(&tte, pfn, attr, TTE512K); 2278 (void) sfmmu_tteload_array(hat, &tte, addr, &pp, 2279 flags, SFMMU_INVALID_SHMERID); 2280 len -= MMU_PAGESIZE512K; 2281 addr += MMU_PAGESIZE512K; 2282 pfn += MMU_PAGESIZE512K / MMU_PAGESIZE; 2283 } else if ((len >= MMU_PAGESIZE64K) && 2284 !((uintptr_t)addr & MMU_PAGEOFFSET64K) && 2285 !(disable_large_pages & (1 << TTE64K)) && 2286 !(mmu_ptob(pfn) & MMU_PAGEOFFSET64K)) { 2287 sfmmu_memtte(&tte, pfn, attr, TTE64K); 2288 (void) sfmmu_tteload_array(hat, &tte, addr, &pp, 2289 flags, SFMMU_INVALID_SHMERID); 2290 len -= MMU_PAGESIZE64K; 2291 addr += MMU_PAGESIZE64K; 2292 pfn += MMU_PAGESIZE64K / MMU_PAGESIZE; 2293 } else { 2294 sfmmu_memtte(&tte, pfn, attr, TTE8K); 2295 (void) sfmmu_tteload_array(hat, &tte, addr, &pp, 2296 flags, SFMMU_INVALID_SHMERID); 2297 len -= MMU_PAGESIZE; 2298 addr += MMU_PAGESIZE; 2299 pfn++; 2300 } 2301 } 2302 2303 /* 2304 * Check TSB and TLB page sizes. 2305 */ 2306 if ((flags & HAT_LOAD_SHARE) == 0) { 2307 sfmmu_check_page_sizes(hat, 1); 2308 } 2309 } 2310 2311 void 2312 hat_memload_array(struct hat *hat, caddr_t addr, size_t len, 2313 struct page **pps, uint_t attr, uint_t flags) 2314 { 2315 hat_do_memload_array(hat, addr, len, pps, attr, flags, 2316 SFMMU_INVALID_SHMERID); 2317 } 2318 2319 void 2320 hat_memload_array_region(struct hat *hat, caddr_t addr, size_t len, 2321 struct page **pps, uint_t attr, uint_t flags, 2322 hat_region_cookie_t rcookie) 2323 { 2324 uint_t rid; 2325 if (rcookie == HAT_INVALID_REGION_COOKIE || 2326 hat->sfmmu_xhat_provider != NULL) { 2327 hat_do_memload_array(hat, addr, len, pps, attr, flags, 2328 SFMMU_INVALID_SHMERID); 2329 return; 2330 } 2331 rid = (uint_t)((uint64_t)rcookie); 2332 ASSERT(rid < SFMMU_MAX_HME_REGIONS); 2333 hat_do_memload_array(hat, addr, len, pps, attr, flags, rid); 2334 } 2335 2336 /* 2337 * Map the largest extend possible out of the page array. The array may NOT 2338 * be in order. The largest possible mapping a page can have 2339 * is specified in the p_szc field. The p_szc field 2340 * cannot change as long as there any mappings (large or small) 2341 * to any of the pages that make up the large page. (ie. any 2342 * promotion/demotion of page size is not up to the hat but up to 2343 * the page free list manager). The array 2344 * should consist of properly aligned contigous pages that are 2345 * part of a big page for a large mapping to be created. 2346 */ 2347 static void 2348 hat_do_memload_array(struct hat *hat, caddr_t addr, size_t len, 2349 struct page **pps, uint_t attr, uint_t flags, uint_t rid) 2350 { 2351 int ttesz; 2352 size_t mapsz; 2353 pgcnt_t numpg, npgs; 2354 tte_t tte; 2355 page_t *pp; 2356 uint_t large_pages_disable; 2357 2358 ASSERT(!((uintptr_t)addr & MMU_PAGEOFFSET)); 2359 SFMMU_VALIDATE_HMERID(hat, rid, addr, len); 2360 2361 if (hat->sfmmu_xhat_provider) { 2362 ASSERT(!SFMMU_IS_SHMERID_VALID(rid)); 2363 XHAT_MEMLOAD_ARRAY(hat, addr, len, pps, attr, flags); 2364 return; 2365 } 2366 2367 if (hat->sfmmu_rmstat) 2368 hat_resvstat(len, hat->sfmmu_as, addr); 2369 2370 #if defined(SF_ERRATA_57) 2371 if ((hat != ksfmmup) && AS_TYPE_64BIT(hat->sfmmu_as) && 2372 (addr < errata57_limit) && (attr & PROT_EXEC) && 2373 !(flags & HAT_LOAD_SHARE)) { 2374 cmn_err(CE_WARN, "hat_memload_array: illegal attempt to make " 2375 "user page executable"); 2376 attr &= ~PROT_EXEC; 2377 } 2378 #endif 2379 2380 /* Get number of pages */ 2381 npgs = len >> MMU_PAGESHIFT; 2382 2383 if (flags & HAT_LOAD_SHARE) { 2384 large_pages_disable = disable_ism_large_pages; 2385 } else { 2386 large_pages_disable = disable_large_pages; 2387 } 2388 2389 if (npgs < NHMENTS || large_pages_disable == LARGE_PAGES_OFF) { 2390 sfmmu_memload_batchsmall(hat, addr, pps, attr, flags, npgs, 2391 rid); 2392 return; 2393 } 2394 2395 while (npgs >= NHMENTS) { 2396 pp = *pps; 2397 for (ttesz = pp->p_szc; ttesz != TTE8K; ttesz--) { 2398 /* 2399 * Check if this page size is disabled. 2400 */ 2401 if (large_pages_disable & (1 << ttesz)) 2402 continue; 2403 2404 numpg = TTEPAGES(ttesz); 2405 mapsz = numpg << MMU_PAGESHIFT; 2406 if ((npgs >= numpg) && 2407 IS_P2ALIGNED(addr, mapsz) && 2408 IS_P2ALIGNED(pp->p_pagenum, numpg)) { 2409 /* 2410 * At this point we have enough pages and 2411 * we know the virtual address and the pfn 2412 * are properly aligned. We still need 2413 * to check for physical contiguity but since 2414 * it is very likely that this is the case 2415 * we will assume they are so and undo 2416 * the request if necessary. It would 2417 * be great if we could get a hint flag 2418 * like HAT_CONTIG which would tell us 2419 * the pages are contigous for sure. 2420 */ 2421 sfmmu_memtte(&tte, (*pps)->p_pagenum, 2422 attr, ttesz); 2423 if (!sfmmu_tteload_array(hat, &tte, addr, 2424 pps, flags, rid)) { 2425 break; 2426 } 2427 } 2428 } 2429 if (ttesz == TTE8K) { 2430 /* 2431 * We were not able to map array using a large page 2432 * batch a hmeblk or fraction at a time. 2433 */ 2434 numpg = ((uintptr_t)addr >> MMU_PAGESHIFT) 2435 & (NHMENTS-1); 2436 numpg = NHMENTS - numpg; 2437 ASSERT(numpg <= npgs); 2438 mapsz = numpg * MMU_PAGESIZE; 2439 sfmmu_memload_batchsmall(hat, addr, pps, attr, flags, 2440 numpg, rid); 2441 } 2442 addr += mapsz; 2443 npgs -= numpg; 2444 pps += numpg; 2445 } 2446 2447 if (npgs) { 2448 sfmmu_memload_batchsmall(hat, addr, pps, attr, flags, npgs, 2449 rid); 2450 } 2451 2452 /* 2453 * Check TSB and TLB page sizes. 2454 */ 2455 if ((flags & HAT_LOAD_SHARE) == 0) { 2456 sfmmu_check_page_sizes(hat, 1); 2457 } 2458 } 2459 2460 /* 2461 * Function tries to batch 8K pages into the same hme blk. 2462 */ 2463 static void 2464 sfmmu_memload_batchsmall(struct hat *hat, caddr_t vaddr, page_t **pps, 2465 uint_t attr, uint_t flags, pgcnt_t npgs, uint_t rid) 2466 { 2467 tte_t tte; 2468 page_t *pp; 2469 struct hmehash_bucket *hmebp; 2470 struct hme_blk *hmeblkp; 2471 int index; 2472 2473 while (npgs) { 2474 /* 2475 * Acquire the hash bucket. 2476 */ 2477 hmebp = sfmmu_tteload_acquire_hashbucket(hat, vaddr, TTE8K, 2478 rid); 2479 ASSERT(hmebp); 2480 2481 /* 2482 * Find the hment block. 2483 */ 2484 hmeblkp = sfmmu_tteload_find_hmeblk(hat, hmebp, vaddr, 2485 TTE8K, flags, rid); 2486 ASSERT(hmeblkp); 2487 2488 do { 2489 /* 2490 * Make the tte. 2491 */ 2492 pp = *pps; 2493 sfmmu_memtte(&tte, pp->p_pagenum, attr, TTE8K); 2494 2495 /* 2496 * Add the translation. 2497 */ 2498 (void) sfmmu_tteload_addentry(hat, hmeblkp, &tte, 2499 vaddr, pps, flags, rid); 2500 2501 /* 2502 * Goto next page. 2503 */ 2504 pps++; 2505 npgs--; 2506 2507 /* 2508 * Goto next address. 2509 */ 2510 vaddr += MMU_PAGESIZE; 2511 2512 /* 2513 * Don't crossover into a different hmentblk. 2514 */ 2515 index = (int)(((uintptr_t)vaddr >> MMU_PAGESHIFT) & 2516 (NHMENTS-1)); 2517 2518 } while (index != 0 && npgs != 0); 2519 2520 /* 2521 * Release the hash bucket. 2522 */ 2523 2524 sfmmu_tteload_release_hashbucket(hmebp); 2525 } 2526 } 2527 2528 /* 2529 * Construct a tte for a page: 2530 * 2531 * tte_valid = 1 2532 * tte_size2 = size & TTE_SZ2_BITS (Panther and Olympus-C only) 2533 * tte_size = size 2534 * tte_nfo = attr & HAT_NOFAULT 2535 * tte_ie = attr & HAT_STRUCTURE_LE 2536 * tte_hmenum = hmenum 2537 * tte_pahi = pp->p_pagenum >> TTE_PASHIFT; 2538 * tte_palo = pp->p_pagenum & TTE_PALOMASK; 2539 * tte_ref = 1 (optimization) 2540 * tte_wr_perm = attr & PROT_WRITE; 2541 * tte_no_sync = attr & HAT_NOSYNC 2542 * tte_lock = attr & SFMMU_LOCKTTE 2543 * tte_cp = !(attr & SFMMU_UNCACHEPTTE) 2544 * tte_cv = !(attr & SFMMU_UNCACHEVTTE) 2545 * tte_e = attr & SFMMU_SIDEFFECT 2546 * tte_priv = !(attr & PROT_USER) 2547 * tte_hwwr = if nosync is set and it is writable we set the mod bit (opt) 2548 * tte_glb = 0 2549 */ 2550 void 2551 sfmmu_memtte(tte_t *ttep, pfn_t pfn, uint_t attr, int tte_sz) 2552 { 2553 ASSERT(!(attr & ~SFMMU_LOAD_ALLATTR)); 2554 2555 ttep->tte_inthi = MAKE_TTE_INTHI(pfn, attr, tte_sz, 0 /* hmenum */); 2556 ttep->tte_intlo = MAKE_TTE_INTLO(pfn, attr, tte_sz, 0 /* hmenum */); 2557 2558 if (TTE_IS_NOSYNC(ttep)) { 2559 TTE_SET_REF(ttep); 2560 if (TTE_IS_WRITABLE(ttep)) { 2561 TTE_SET_MOD(ttep); 2562 } 2563 } 2564 if (TTE_IS_NFO(ttep) && TTE_IS_EXECUTABLE(ttep)) { 2565 panic("sfmmu_memtte: can't set both NFO and EXEC bits"); 2566 } 2567 } 2568 2569 /* 2570 * This function will add a translation to the hme_blk and allocate the 2571 * hme_blk if one does not exist. 2572 * If a page structure is specified then it will add the 2573 * corresponding hment to the mapping list. 2574 * It will also update the hmenum field for the tte. 2575 * 2576 * Currently this function is only used for kernel mappings. 2577 * So pass invalid region to sfmmu_tteload_array(). 2578 */ 2579 void 2580 sfmmu_tteload(struct hat *sfmmup, tte_t *ttep, caddr_t vaddr, page_t *pp, 2581 uint_t flags) 2582 { 2583 ASSERT(sfmmup == ksfmmup); 2584 (void) sfmmu_tteload_array(sfmmup, ttep, vaddr, &pp, flags, 2585 SFMMU_INVALID_SHMERID); 2586 } 2587 2588 /* 2589 * Load (ttep != NULL) or unload (ttep == NULL) one entry in the TSB. 2590 * Assumes that a particular page size may only be resident in one TSB. 2591 */ 2592 static void 2593 sfmmu_mod_tsb(sfmmu_t *sfmmup, caddr_t vaddr, tte_t *ttep, int ttesz) 2594 { 2595 struct tsb_info *tsbinfop = NULL; 2596 uint64_t tag; 2597 struct tsbe *tsbe_addr; 2598 uint64_t tsb_base; 2599 uint_t tsb_size; 2600 int vpshift = MMU_PAGESHIFT; 2601 int phys = 0; 2602 2603 if (sfmmup == ksfmmup) { /* No support for 32/256M ksfmmu pages */ 2604 phys = ktsb_phys; 2605 if (ttesz >= TTE4M) { 2606 #ifndef sun4v 2607 ASSERT((ttesz != TTE32M) && (ttesz != TTE256M)); 2608 #endif 2609 tsb_base = (phys)? ktsb4m_pbase : (uint64_t)ktsb4m_base; 2610 tsb_size = ktsb4m_szcode; 2611 } else { 2612 tsb_base = (phys)? ktsb_pbase : (uint64_t)ktsb_base; 2613 tsb_size = ktsb_szcode; 2614 } 2615 } else { 2616 SFMMU_GET_TSBINFO(tsbinfop, sfmmup, ttesz); 2617 2618 /* 2619 * If there isn't a TSB for this page size, or the TSB is 2620 * swapped out, there is nothing to do. Note that the latter 2621 * case seems impossible but can occur if hat_pageunload() 2622 * is called on an ISM mapping while the process is swapped 2623 * out. 2624 */ 2625 if (tsbinfop == NULL || (tsbinfop->tsb_flags & TSB_SWAPPED)) 2626 return; 2627 2628 /* 2629 * If another thread is in the middle of relocating a TSB 2630 * we can't unload the entry so set a flag so that the 2631 * TSB will be flushed before it can be accessed by the 2632 * process. 2633 */ 2634 if ((tsbinfop->tsb_flags & TSB_RELOC_FLAG) != 0) { 2635 if (ttep == NULL) 2636 tsbinfop->tsb_flags |= TSB_FLUSH_NEEDED; 2637 return; 2638 } 2639 #if defined(UTSB_PHYS) 2640 phys = 1; 2641 tsb_base = (uint64_t)tsbinfop->tsb_pa; 2642 #else 2643 tsb_base = (uint64_t)tsbinfop->tsb_va; 2644 #endif 2645 tsb_size = tsbinfop->tsb_szc; 2646 } 2647 if (ttesz >= TTE4M) 2648 vpshift = MMU_PAGESHIFT4M; 2649 2650 tsbe_addr = sfmmu_get_tsbe(tsb_base, vaddr, vpshift, tsb_size); 2651 tag = sfmmu_make_tsbtag(vaddr); 2652 2653 if (ttep == NULL) { 2654 sfmmu_unload_tsbe(tsbe_addr, tag, phys); 2655 } else { 2656 if (ttesz >= TTE4M) { 2657 SFMMU_STAT(sf_tsb_load4m); 2658 } else { 2659 SFMMU_STAT(sf_tsb_load8k); 2660 } 2661 2662 sfmmu_load_tsbe(tsbe_addr, tag, ttep, phys); 2663 } 2664 } 2665 2666 /* 2667 * Unmap all entries from [start, end) matching the given page size. 2668 * 2669 * This function is used primarily to unmap replicated 64K or 512K entries 2670 * from the TSB that are inserted using the base page size TSB pointer, but 2671 * it may also be called to unmap a range of addresses from the TSB. 2672 */ 2673 void 2674 sfmmu_unload_tsb_range(sfmmu_t *sfmmup, caddr_t start, caddr_t end, int ttesz) 2675 { 2676 struct tsb_info *tsbinfop; 2677 uint64_t tag; 2678 struct tsbe *tsbe_addr; 2679 caddr_t vaddr; 2680 uint64_t tsb_base; 2681 int vpshift, vpgsz; 2682 uint_t tsb_size; 2683 int phys = 0; 2684 2685 /* 2686 * Assumptions: 2687 * If ttesz == 8K, 64K or 512K, we walk through the range 8K 2688 * at a time shooting down any valid entries we encounter. 2689 * 2690 * If ttesz >= 4M we walk the range 4M at a time shooting 2691 * down any valid mappings we find. 2692 */ 2693 if (sfmmup == ksfmmup) { 2694 phys = ktsb_phys; 2695 if (ttesz >= TTE4M) { 2696 #ifndef sun4v 2697 ASSERT((ttesz != TTE32M) && (ttesz != TTE256M)); 2698 #endif 2699 tsb_base = (phys)? ktsb4m_pbase : (uint64_t)ktsb4m_base; 2700 tsb_size = ktsb4m_szcode; 2701 } else { 2702 tsb_base = (phys)? ktsb_pbase : (uint64_t)ktsb_base; 2703 tsb_size = ktsb_szcode; 2704 } 2705 } else { 2706 SFMMU_GET_TSBINFO(tsbinfop, sfmmup, ttesz); 2707 2708 /* 2709 * If there isn't a TSB for this page size, or the TSB is 2710 * swapped out, there is nothing to do. Note that the latter 2711 * case seems impossible but can occur if hat_pageunload() 2712 * is called on an ISM mapping while the process is swapped 2713 * out. 2714 */ 2715 if (tsbinfop == NULL || (tsbinfop->tsb_flags & TSB_SWAPPED)) 2716 return; 2717 2718 /* 2719 * If another thread is in the middle of relocating a TSB 2720 * we can't unload the entry so set a flag so that the 2721 * TSB will be flushed before it can be accessed by the 2722 * process. 2723 */ 2724 if ((tsbinfop->tsb_flags & TSB_RELOC_FLAG) != 0) { 2725 tsbinfop->tsb_flags |= TSB_FLUSH_NEEDED; 2726 return; 2727 } 2728 #if defined(UTSB_PHYS) 2729 phys = 1; 2730 tsb_base = (uint64_t)tsbinfop->tsb_pa; 2731 #else 2732 tsb_base = (uint64_t)tsbinfop->tsb_va; 2733 #endif 2734 tsb_size = tsbinfop->tsb_szc; 2735 } 2736 if (ttesz >= TTE4M) { 2737 vpshift = MMU_PAGESHIFT4M; 2738 vpgsz = MMU_PAGESIZE4M; 2739 } else { 2740 vpshift = MMU_PAGESHIFT; 2741 vpgsz = MMU_PAGESIZE; 2742 } 2743 2744 for (vaddr = start; vaddr < end; vaddr += vpgsz) { 2745 tag = sfmmu_make_tsbtag(vaddr); 2746 tsbe_addr = sfmmu_get_tsbe(tsb_base, vaddr, vpshift, tsb_size); 2747 sfmmu_unload_tsbe(tsbe_addr, tag, phys); 2748 } 2749 } 2750 2751 /* 2752 * Select the optimum TSB size given the number of mappings 2753 * that need to be cached. 2754 */ 2755 static int 2756 sfmmu_select_tsb_szc(pgcnt_t pgcnt) 2757 { 2758 int szc = 0; 2759 2760 #ifdef DEBUG 2761 if (tsb_grow_stress) { 2762 uint32_t randval = (uint32_t)gettick() >> 4; 2763 return (randval % (tsb_max_growsize + 1)); 2764 } 2765 #endif /* DEBUG */ 2766 2767 while ((szc < tsb_max_growsize) && (pgcnt > SFMMU_RSS_TSBSIZE(szc))) 2768 szc++; 2769 return (szc); 2770 } 2771 2772 /* 2773 * This function will add a translation to the hme_blk and allocate the 2774 * hme_blk if one does not exist. 2775 * If a page structure is specified then it will add the 2776 * corresponding hment to the mapping list. 2777 * It will also update the hmenum field for the tte. 2778 * Furthermore, it attempts to create a large page translation 2779 * for <addr,hat> at page array pps. It assumes addr and first 2780 * pp is correctly aligned. It returns 0 if successful and 1 otherwise. 2781 */ 2782 static int 2783 sfmmu_tteload_array(sfmmu_t *sfmmup, tte_t *ttep, caddr_t vaddr, 2784 page_t **pps, uint_t flags, uint_t rid) 2785 { 2786 struct hmehash_bucket *hmebp; 2787 struct hme_blk *hmeblkp; 2788 int ret; 2789 uint_t size; 2790 2791 /* 2792 * Get mapping size. 2793 */ 2794 size = TTE_CSZ(ttep); 2795 ASSERT(!((uintptr_t)vaddr & TTE_PAGE_OFFSET(size))); 2796 2797 /* 2798 * Acquire the hash bucket. 2799 */ 2800 hmebp = sfmmu_tteload_acquire_hashbucket(sfmmup, vaddr, size, rid); 2801 ASSERT(hmebp); 2802 2803 /* 2804 * Find the hment block. 2805 */ 2806 hmeblkp = sfmmu_tteload_find_hmeblk(sfmmup, hmebp, vaddr, size, flags, 2807 rid); 2808 ASSERT(hmeblkp); 2809 2810 /* 2811 * Add the translation. 2812 */ 2813 ret = sfmmu_tteload_addentry(sfmmup, hmeblkp, ttep, vaddr, pps, flags, 2814 rid); 2815 2816 /* 2817 * Release the hash bucket. 2818 */ 2819 sfmmu_tteload_release_hashbucket(hmebp); 2820 2821 return (ret); 2822 } 2823 2824 /* 2825 * Function locks and returns a pointer to the hash bucket for vaddr and size. 2826 */ 2827 static struct hmehash_bucket * 2828 sfmmu_tteload_acquire_hashbucket(sfmmu_t *sfmmup, caddr_t vaddr, int size, 2829 uint_t rid) 2830 { 2831 struct hmehash_bucket *hmebp; 2832 int hmeshift; 2833 void *htagid = sfmmutohtagid(sfmmup, rid); 2834 2835 ASSERT(htagid != NULL); 2836 2837 hmeshift = HME_HASH_SHIFT(size); 2838 2839 hmebp = HME_HASH_FUNCTION(htagid, vaddr, hmeshift); 2840 2841 SFMMU_HASH_LOCK(hmebp); 2842 2843 return (hmebp); 2844 } 2845 2846 /* 2847 * Function returns a pointer to an hmeblk in the hash bucket, hmebp. If the 2848 * hmeblk doesn't exists for the [sfmmup, vaddr & size] signature, a hmeblk is 2849 * allocated. 2850 */ 2851 static struct hme_blk * 2852 sfmmu_tteload_find_hmeblk(sfmmu_t *sfmmup, struct hmehash_bucket *hmebp, 2853 caddr_t vaddr, uint_t size, uint_t flags, uint_t rid) 2854 { 2855 hmeblk_tag hblktag; 2856 int hmeshift; 2857 struct hme_blk *hmeblkp, *pr_hblk, *list = NULL; 2858 2859 SFMMU_VALIDATE_HMERID(sfmmup, rid, vaddr, TTEBYTES(size)); 2860 2861 hblktag.htag_id = sfmmutohtagid(sfmmup, rid); 2862 ASSERT(hblktag.htag_id != NULL); 2863 hmeshift = HME_HASH_SHIFT(size); 2864 hblktag.htag_bspage = HME_HASH_BSPAGE(vaddr, hmeshift); 2865 hblktag.htag_rehash = HME_HASH_REHASH(size); 2866 hblktag.htag_rid = rid; 2867 2868 ttearray_realloc: 2869 2870 HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, pr_hblk, &list); 2871 2872 /* 2873 * We block until hblk_reserve_lock is released; it's held by 2874 * the thread, temporarily using hblk_reserve, until hblk_reserve is 2875 * replaced by a hblk from sfmmu8_cache. 2876 */ 2877 if (hmeblkp == (struct hme_blk *)hblk_reserve && 2878 hblk_reserve_thread != curthread) { 2879 SFMMU_HASH_UNLOCK(hmebp); 2880 mutex_enter(&hblk_reserve_lock); 2881 mutex_exit(&hblk_reserve_lock); 2882 SFMMU_STAT(sf_hblk_reserve_hit); 2883 SFMMU_HASH_LOCK(hmebp); 2884 goto ttearray_realloc; 2885 } 2886 2887 if (hmeblkp == NULL) { 2888 hmeblkp = sfmmu_hblk_alloc(sfmmup, vaddr, hmebp, size, 2889 hblktag, flags, rid); 2890 ASSERT(!SFMMU_IS_SHMERID_VALID(rid) || hmeblkp->hblk_shared); 2891 ASSERT(SFMMU_IS_SHMERID_VALID(rid) || !hmeblkp->hblk_shared); 2892 } else { 2893 /* 2894 * It is possible for 8k and 64k hblks to collide since they 2895 * have the same rehash value. This is because we 2896 * lazily free hblks and 8K/64K blks could be lingering. 2897 * If we find size mismatch we free the block and & try again. 2898 */ 2899 if (get_hblk_ttesz(hmeblkp) != size) { 2900 ASSERT(!hmeblkp->hblk_vcnt); 2901 ASSERT(!hmeblkp->hblk_hmecnt); 2902 sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk, 2903 &list, 0); 2904 goto ttearray_realloc; 2905 } 2906 if (hmeblkp->hblk_shw_bit) { 2907 /* 2908 * if the hblk was previously used as a shadow hblk then 2909 * we will change it to a normal hblk 2910 */ 2911 ASSERT(!hmeblkp->hblk_shared); 2912 if (hmeblkp->hblk_shw_mask) { 2913 sfmmu_shadow_hcleanup(sfmmup, hmeblkp, hmebp); 2914 ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp)); 2915 goto ttearray_realloc; 2916 } else { 2917 hmeblkp->hblk_shw_bit = 0; 2918 } 2919 } 2920 SFMMU_STAT(sf_hblk_hit); 2921 } 2922 2923 /* 2924 * hat_memload() should never call kmem_cache_free() for kernel hmeblks; 2925 * see block comment showing the stacktrace in sfmmu_hblk_alloc(); 2926 * set the flag parameter to 1 so that sfmmu_hblks_list_purge() will 2927 * just add these hmeblks to the per-cpu pending queue. 2928 */ 2929 sfmmu_hblks_list_purge(&list, 1); 2930 2931 ASSERT(get_hblk_ttesz(hmeblkp) == size); 2932 ASSERT(!hmeblkp->hblk_shw_bit); 2933 ASSERT(!SFMMU_IS_SHMERID_VALID(rid) || hmeblkp->hblk_shared); 2934 ASSERT(SFMMU_IS_SHMERID_VALID(rid) || !hmeblkp->hblk_shared); 2935 ASSERT(hmeblkp->hblk_tag.htag_rid == rid); 2936 2937 return (hmeblkp); 2938 } 2939 2940 /* 2941 * Function adds a tte entry into the hmeblk. It returns 0 if successful and 1 2942 * otherwise. 2943 */ 2944 static int 2945 sfmmu_tteload_addentry(sfmmu_t *sfmmup, struct hme_blk *hmeblkp, tte_t *ttep, 2946 caddr_t vaddr, page_t **pps, uint_t flags, uint_t rid) 2947 { 2948 page_t *pp = *pps; 2949 int hmenum, size, remap; 2950 tte_t tteold, flush_tte; 2951 #ifdef DEBUG 2952 tte_t orig_old; 2953 #endif /* DEBUG */ 2954 struct sf_hment *sfhme; 2955 kmutex_t *pml, *pmtx; 2956 hatlock_t *hatlockp; 2957 int myflt; 2958 2959 /* 2960 * remove this panic when we decide to let user virtual address 2961 * space be >= USERLIMIT. 2962 */ 2963 if (!TTE_IS_PRIVILEGED(ttep) && vaddr >= (caddr_t)USERLIMIT) 2964 panic("user addr %p in kernel space", (void *)vaddr); 2965 #if defined(TTE_IS_GLOBAL) 2966 if (TTE_IS_GLOBAL(ttep)) 2967 panic("sfmmu_tteload: creating global tte"); 2968 #endif 2969 2970 #ifdef DEBUG 2971 if (pf_is_memory(sfmmu_ttetopfn(ttep, vaddr)) && 2972 !TTE_IS_PCACHEABLE(ttep) && !sfmmu_allow_nc_trans) 2973 panic("sfmmu_tteload: non cacheable memory tte"); 2974 #endif /* DEBUG */ 2975 2976 /* don't simulate dirty bit for writeable ISM/DISM mappings */ 2977 if ((flags & HAT_LOAD_SHARE) && TTE_IS_WRITABLE(ttep)) { 2978 TTE_SET_REF(ttep); 2979 TTE_SET_MOD(ttep); 2980 } 2981 2982 if ((flags & HAT_LOAD_SHARE) || !TTE_IS_REF(ttep) || 2983 !TTE_IS_MOD(ttep)) { 2984 /* 2985 * Don't load TSB for dummy as in ISM. Also don't preload 2986 * the TSB if the TTE isn't writable since we're likely to 2987 * fault on it again -- preloading can be fairly expensive. 2988 */ 2989 flags |= SFMMU_NO_TSBLOAD; 2990 } 2991 2992 size = TTE_CSZ(ttep); 2993 switch (size) { 2994 case TTE8K: 2995 SFMMU_STAT(sf_tteload8k); 2996 break; 2997 case TTE64K: 2998 SFMMU_STAT(sf_tteload64k); 2999 break; 3000 case TTE512K: 3001 SFMMU_STAT(sf_tteload512k); 3002 break; 3003 case TTE4M: 3004 SFMMU_STAT(sf_tteload4m); 3005 break; 3006 case (TTE32M): 3007 SFMMU_STAT(sf_tteload32m); 3008 ASSERT(mmu_page_sizes == max_mmu_page_sizes); 3009 break; 3010 case (TTE256M): 3011 SFMMU_STAT(sf_tteload256m); 3012 ASSERT(mmu_page_sizes == max_mmu_page_sizes); 3013 break; 3014 } 3015 3016 ASSERT(!((uintptr_t)vaddr & TTE_PAGE_OFFSET(size))); 3017 SFMMU_VALIDATE_HMERID(sfmmup, rid, vaddr, TTEBYTES(size)); 3018 ASSERT(!SFMMU_IS_SHMERID_VALID(rid) || hmeblkp->hblk_shared); 3019 ASSERT(SFMMU_IS_SHMERID_VALID(rid) || !hmeblkp->hblk_shared); 3020 3021 HBLKTOHME_IDX(sfhme, hmeblkp, vaddr, hmenum); 3022 3023 /* 3024 * Need to grab mlist lock here so that pageunload 3025 * will not change tte behind us. 3026 */ 3027 if (pp) { 3028 pml = sfmmu_mlist_enter(pp); 3029 } 3030 3031 sfmmu_copytte(&sfhme->hme_tte, &tteold); 3032 /* 3033 * Look for corresponding hment and if valid verify 3034 * pfns are equal. 3035 */ 3036 remap = TTE_IS_VALID(&tteold); 3037 if (remap) { 3038 pfn_t new_pfn, old_pfn; 3039 3040 old_pfn = TTE_TO_PFN(vaddr, &tteold); 3041 new_pfn = TTE_TO_PFN(vaddr, ttep); 3042 3043 if (flags & HAT_LOAD_REMAP) { 3044 /* make sure we are remapping same type of pages */ 3045 if (pf_is_memory(old_pfn) != pf_is_memory(new_pfn)) { 3046 panic("sfmmu_tteload - tte remap io<->memory"); 3047 } 3048 if (old_pfn != new_pfn && 3049 (pp != NULL || sfhme->hme_page != NULL)) { 3050 panic("sfmmu_tteload - tte remap pp != NULL"); 3051 } 3052 } else if (old_pfn != new_pfn) { 3053 panic("sfmmu_tteload - tte remap, hmeblkp 0x%p", 3054 (void *)hmeblkp); 3055 } 3056 ASSERT(TTE_CSZ(&tteold) == TTE_CSZ(ttep)); 3057 } 3058 3059 if (pp) { 3060 if (size == TTE8K) { 3061 #ifdef VAC 3062 /* 3063 * Handle VAC consistency 3064 */ 3065 if (!remap && (cache & CACHE_VAC) && !PP_ISNC(pp)) { 3066 sfmmu_vac_conflict(sfmmup, vaddr, pp); 3067 } 3068 #endif 3069 3070 if (TTE_IS_WRITABLE(ttep) && PP_ISRO(pp)) { 3071 pmtx = sfmmu_page_enter(pp); 3072 PP_CLRRO(pp); 3073 sfmmu_page_exit(pmtx); 3074 } else if (!PP_ISMAPPED(pp) && 3075 (!TTE_IS_WRITABLE(ttep)) && !(PP_ISMOD(pp))) { 3076 pmtx = sfmmu_page_enter(pp); 3077 if (!(PP_ISMOD(pp))) { 3078 PP_SETRO(pp); 3079 } 3080 sfmmu_page_exit(pmtx); 3081 } 3082 3083 } else if (sfmmu_pagearray_setup(vaddr, pps, ttep, remap)) { 3084 /* 3085 * sfmmu_pagearray_setup failed so return 3086 */ 3087 sfmmu_mlist_exit(pml); 3088 return (1); 3089 } 3090 } 3091 3092 /* 3093 * Make sure hment is not on a mapping list. 3094 */ 3095 ASSERT(remap || (sfhme->hme_page == NULL)); 3096 3097 /* if it is not a remap then hme->next better be NULL */ 3098 ASSERT((!remap) ? sfhme->hme_next == NULL : 1); 3099 3100 if (flags & HAT_LOAD_LOCK) { 3101 if ((hmeblkp->hblk_lckcnt + 1) >= MAX_HBLK_LCKCNT) { 3102 panic("too high lckcnt-hmeblk %p", 3103 (void *)hmeblkp); 3104 } 3105 atomic_add_32(&hmeblkp->hblk_lckcnt, 1); 3106 3107 HBLK_STACK_TRACE(hmeblkp, HBLK_LOCK); 3108 } 3109 3110 #ifdef VAC 3111 if (pp && PP_ISNC(pp)) { 3112 /* 3113 * If the physical page is marked to be uncacheable, like 3114 * by a vac conflict, make sure the new mapping is also 3115 * uncacheable. 3116 */ 3117 TTE_CLR_VCACHEABLE(ttep); 3118 ASSERT(PP_GET_VCOLOR(pp) == NO_VCOLOR); 3119 } 3120 #endif 3121 ttep->tte_hmenum = hmenum; 3122 3123 #ifdef DEBUG 3124 orig_old = tteold; 3125 #endif /* DEBUG */ 3126 3127 while (sfmmu_modifytte_try(&tteold, ttep, &sfhme->hme_tte) < 0) { 3128 if ((sfmmup == KHATID) && 3129 (flags & (HAT_LOAD_LOCK | HAT_LOAD_REMAP))) { 3130 sfmmu_copytte(&sfhme->hme_tte, &tteold); 3131 } 3132 #ifdef DEBUG 3133 chk_tte(&orig_old, &tteold, ttep, hmeblkp); 3134 #endif /* DEBUG */ 3135 } 3136 ASSERT(TTE_IS_VALID(&sfhme->hme_tte)); 3137 3138 if (!TTE_IS_VALID(&tteold)) { 3139 3140 atomic_add_16(&hmeblkp->hblk_vcnt, 1); 3141 if (rid == SFMMU_INVALID_SHMERID) { 3142 atomic_add_long(&sfmmup->sfmmu_ttecnt[size], 1); 3143 } else { 3144 sf_srd_t *srdp = sfmmup->sfmmu_srdp; 3145 sf_region_t *rgnp = srdp->srd_hmergnp[rid]; 3146 /* 3147 * We already accounted for region ttecnt's in sfmmu 3148 * during hat_join_region() processing. Here we 3149 * only update ttecnt's in region struture. 3150 */ 3151 atomic_add_long(&rgnp->rgn_ttecnt[size], 1); 3152 } 3153 } 3154 3155 myflt = (astosfmmu(curthread->t_procp->p_as) == sfmmup); 3156 if (size > TTE8K && (flags & HAT_LOAD_SHARE) == 0 && 3157 sfmmup != ksfmmup) { 3158 uchar_t tteflag = 1 << size; 3159 if (rid == SFMMU_INVALID_SHMERID) { 3160 if (!(sfmmup->sfmmu_tteflags & tteflag)) { 3161 hatlockp = sfmmu_hat_enter(sfmmup); 3162 sfmmup->sfmmu_tteflags |= tteflag; 3163 sfmmu_hat_exit(hatlockp); 3164 } 3165 } else if (!(sfmmup->sfmmu_rtteflags & tteflag)) { 3166 hatlockp = sfmmu_hat_enter(sfmmup); 3167 sfmmup->sfmmu_rtteflags |= tteflag; 3168 sfmmu_hat_exit(hatlockp); 3169 } 3170 /* 3171 * Update the current CPU tsbmiss area, so the current thread 3172 * won't need to take the tsbmiss for the new pagesize. 3173 * The other threads in the process will update their tsb 3174 * miss area lazily in sfmmu_tsbmiss_exception() when they 3175 * fail to find the translation for a newly added pagesize. 3176 */ 3177 if (size > TTE64K && myflt) { 3178 struct tsbmiss *tsbmp; 3179 kpreempt_disable(); 3180 tsbmp = &tsbmiss_area[CPU->cpu_id]; 3181 if (rid == SFMMU_INVALID_SHMERID) { 3182 if (!(tsbmp->uhat_tteflags & tteflag)) { 3183 tsbmp->uhat_tteflags |= tteflag; 3184 } 3185 } else { 3186 if (!(tsbmp->uhat_rtteflags & tteflag)) { 3187 tsbmp->uhat_rtteflags |= tteflag; 3188 } 3189 } 3190 kpreempt_enable(); 3191 } 3192 } 3193 3194 if (size >= TTE4M && (flags & HAT_LOAD_TEXT) && 3195 !SFMMU_FLAGS_ISSET(sfmmup, HAT_4MTEXT_FLAG)) { 3196 hatlockp = sfmmu_hat_enter(sfmmup); 3197 SFMMU_FLAGS_SET(sfmmup, HAT_4MTEXT_FLAG); 3198 sfmmu_hat_exit(hatlockp); 3199 } 3200 3201 flush_tte.tte_intlo = (tteold.tte_intlo ^ ttep->tte_intlo) & 3202 hw_tte.tte_intlo; 3203 flush_tte.tte_inthi = (tteold.tte_inthi ^ ttep->tte_inthi) & 3204 hw_tte.tte_inthi; 3205 3206 if (remap && (flush_tte.tte_inthi || flush_tte.tte_intlo)) { 3207 /* 3208 * If remap and new tte differs from old tte we need 3209 * to sync the mod bit and flush TLB/TSB. We don't 3210 * need to sync ref bit because we currently always set 3211 * ref bit in tteload. 3212 */ 3213 ASSERT(TTE_IS_REF(ttep)); 3214 if (TTE_IS_MOD(&tteold)) { 3215 sfmmu_ttesync(sfmmup, vaddr, &tteold, pp); 3216 } 3217 /* 3218 * hwtte bits shouldn't change for SRD hmeblks as long as SRD 3219 * hmes are only used for read only text. Adding this code for 3220 * completeness and future use of shared hmeblks with writable 3221 * mappings of VMODSORT vnodes. 3222 */ 3223 if (hmeblkp->hblk_shared) { 3224 cpuset_t cpuset = sfmmu_rgntlb_demap(vaddr, 3225 sfmmup->sfmmu_srdp->srd_hmergnp[rid], hmeblkp, 1); 3226 xt_sync(cpuset); 3227 SFMMU_STAT_ADD(sf_region_remap_demap, 1); 3228 } else { 3229 sfmmu_tlb_demap(vaddr, sfmmup, hmeblkp, 0, 0); 3230 xt_sync(sfmmup->sfmmu_cpusran); 3231 } 3232 } 3233 3234 if ((flags & SFMMU_NO_TSBLOAD) == 0) { 3235 /* 3236 * We only preload 8K and 4M mappings into the TSB, since 3237 * 64K and 512K mappings are replicated and hence don't 3238 * have a single, unique TSB entry. Ditto for 32M/256M. 3239 */ 3240 if (size == TTE8K || size == TTE4M) { 3241 sf_scd_t *scdp; 3242 hatlockp = sfmmu_hat_enter(sfmmup); 3243 /* 3244 * Don't preload private TSB if the mapping is used 3245 * by the shctx in the SCD. 3246 */ 3247 scdp = sfmmup->sfmmu_scdp; 3248 if (rid == SFMMU_INVALID_SHMERID || scdp == NULL || 3249 !SF_RGNMAP_TEST(scdp->scd_hmeregion_map, rid)) { 3250 sfmmu_load_tsb(sfmmup, vaddr, &sfhme->hme_tte, 3251 size); 3252 } 3253 sfmmu_hat_exit(hatlockp); 3254 } 3255 } 3256 if (pp) { 3257 if (!remap) { 3258 HME_ADD(sfhme, pp); 3259 atomic_add_16(&hmeblkp->hblk_hmecnt, 1); 3260 ASSERT(hmeblkp->hblk_hmecnt > 0); 3261 3262 /* 3263 * Cannot ASSERT(hmeblkp->hblk_hmecnt <= NHMENTS) 3264 * see pageunload() for comment. 3265 */ 3266 } 3267 sfmmu_mlist_exit(pml); 3268 } 3269 3270 return (0); 3271 } 3272 /* 3273 * Function unlocks hash bucket. 3274 */ 3275 static void 3276 sfmmu_tteload_release_hashbucket(struct hmehash_bucket *hmebp) 3277 { 3278 ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp)); 3279 SFMMU_HASH_UNLOCK(hmebp); 3280 } 3281 3282 /* 3283 * function which checks and sets up page array for a large 3284 * translation. Will set p_vcolor, p_index, p_ro fields. 3285 * Assumes addr and pfnum of first page are properly aligned. 3286 * Will check for physical contiguity. If check fails it return 3287 * non null. 3288 */ 3289 static int 3290 sfmmu_pagearray_setup(caddr_t addr, page_t **pps, tte_t *ttep, int remap) 3291 { 3292 int i, index, ttesz; 3293 pfn_t pfnum; 3294 pgcnt_t npgs; 3295 page_t *pp, *pp1; 3296 kmutex_t *pmtx; 3297 #ifdef VAC 3298 int osz; 3299 int cflags = 0; 3300 int vac_err = 0; 3301 #endif 3302 int newidx = 0; 3303 3304 ttesz = TTE_CSZ(ttep); 3305 3306 ASSERT(ttesz > TTE8K); 3307 3308 npgs = TTEPAGES(ttesz); 3309 index = PAGESZ_TO_INDEX(ttesz); 3310 3311 pfnum = (*pps)->p_pagenum; 3312 ASSERT(IS_P2ALIGNED(pfnum, npgs)); 3313 3314 /* 3315 * Save the first pp so we can do HAT_TMPNC at the end. 3316 */ 3317 pp1 = *pps; 3318 #ifdef VAC 3319 osz = fnd_mapping_sz(pp1); 3320 #endif 3321 3322 for (i = 0; i < npgs; i++, pps++) { 3323 pp = *pps; 3324 ASSERT(PAGE_LOCKED(pp)); 3325 ASSERT(pp->p_szc >= ttesz); 3326 ASSERT(pp->p_szc == pp1->p_szc); 3327 ASSERT(sfmmu_mlist_held(pp)); 3328 3329 /* 3330 * XXX is it possible to maintain P_RO on the root only? 3331 */ 3332 if (TTE_IS_WRITABLE(ttep) && PP_ISRO(pp)) { 3333 pmtx = sfmmu_page_enter(pp); 3334 PP_CLRRO(pp); 3335 sfmmu_page_exit(pmtx); 3336 } else if (!PP_ISMAPPED(pp) && !TTE_IS_WRITABLE(ttep) && 3337 !PP_ISMOD(pp)) { 3338 pmtx = sfmmu_page_enter(pp); 3339 if (!(PP_ISMOD(pp))) { 3340 PP_SETRO(pp); 3341 } 3342 sfmmu_page_exit(pmtx); 3343 } 3344 3345 /* 3346 * If this is a remap we skip vac & contiguity checks. 3347 */ 3348 if (remap) 3349 continue; 3350 3351 /* 3352 * set p_vcolor and detect any vac conflicts. 3353 */ 3354 #ifdef VAC 3355 if (vac_err == 0) { 3356 vac_err = sfmmu_vacconflict_array(addr, pp, &cflags); 3357 3358 } 3359 #endif 3360 3361 /* 3362 * Save current index in case we need to undo it. 3363 * Note: "PAGESZ_TO_INDEX(sz) (1 << (sz))" 3364 * "SFMMU_INDEX_SHIFT 6" 3365 * "SFMMU_INDEX_MASK ((1 << SFMMU_INDEX_SHIFT) - 1)" 3366 * "PP_MAPINDEX(p_index) (p_index & SFMMU_INDEX_MASK)" 3367 * 3368 * So: index = PAGESZ_TO_INDEX(ttesz); 3369 * if ttesz == 1 then index = 0x2 3370 * 2 then index = 0x4 3371 * 3 then index = 0x8 3372 * 4 then index = 0x10 3373 * 5 then index = 0x20 3374 * The code below checks if it's a new pagesize (ie, newidx) 3375 * in case we need to take it back out of p_index, 3376 * and then or's the new index into the existing index. 3377 */ 3378 if ((PP_MAPINDEX(pp) & index) == 0) 3379 newidx = 1; 3380 pp->p_index = (PP_MAPINDEX(pp) | index); 3381 3382 /* 3383 * contiguity check 3384 */ 3385 if (pp->p_pagenum != pfnum) { 3386 /* 3387 * If we fail the contiguity test then 3388 * the only thing we need to fix is the p_index field. 3389 * We might get a few extra flushes but since this 3390 * path is rare that is ok. The p_ro field will 3391 * get automatically fixed on the next tteload to 3392 * the page. NO TNC bit is set yet. 3393 */ 3394 while (i >= 0) { 3395 pp = *pps; 3396 if (newidx) 3397 pp->p_index = (PP_MAPINDEX(pp) & 3398 ~index); 3399 pps--; 3400 i--; 3401 } 3402 return (1); 3403 } 3404 pfnum++; 3405 addr += MMU_PAGESIZE; 3406 } 3407 3408 #ifdef VAC 3409 if (vac_err) { 3410 if (ttesz > osz) { 3411 /* 3412 * There are some smaller mappings that causes vac 3413 * conflicts. Convert all existing small mappings to 3414 * TNC. 3415 */ 3416 SFMMU_STAT_ADD(sf_uncache_conflict, npgs); 3417 sfmmu_page_cache_array(pp1, HAT_TMPNC, CACHE_FLUSH, 3418 npgs); 3419 } else { 3420 /* EMPTY */ 3421 /* 3422 * If there exists an big page mapping, 3423 * that means the whole existing big page 3424 * has TNC setting already. No need to covert to 3425 * TNC again. 3426 */ 3427 ASSERT(PP_ISTNC(pp1)); 3428 } 3429 } 3430 #endif /* VAC */ 3431 3432 return (0); 3433 } 3434 3435 #ifdef VAC 3436 /* 3437 * Routine that detects vac consistency for a large page. It also 3438 * sets virtual color for all pp's for this big mapping. 3439 */ 3440 static int 3441 sfmmu_vacconflict_array(caddr_t addr, page_t *pp, int *cflags) 3442 { 3443 int vcolor, ocolor; 3444 3445 ASSERT(sfmmu_mlist_held(pp)); 3446 3447 if (PP_ISNC(pp)) { 3448 return (HAT_TMPNC); 3449 } 3450 3451 vcolor = addr_to_vcolor(addr); 3452 if (PP_NEWPAGE(pp)) { 3453 PP_SET_VCOLOR(pp, vcolor); 3454 return (0); 3455 } 3456 3457 ocolor = PP_GET_VCOLOR(pp); 3458 if (ocolor == vcolor) { 3459 return (0); 3460 } 3461 3462 if (!PP_ISMAPPED(pp) && !PP_ISMAPPED_KPM(pp)) { 3463 /* 3464 * Previous user of page had a differnet color 3465 * but since there are no current users 3466 * we just flush the cache and change the color. 3467 * As an optimization for large pages we flush the 3468 * entire cache of that color and set a flag. 3469 */ 3470 SFMMU_STAT(sf_pgcolor_conflict); 3471 if (!CacheColor_IsFlushed(*cflags, ocolor)) { 3472 CacheColor_SetFlushed(*cflags, ocolor); 3473 sfmmu_cache_flushcolor(ocolor, pp->p_pagenum); 3474 } 3475 PP_SET_VCOLOR(pp, vcolor); 3476 return (0); 3477 } 3478 3479 /* 3480 * We got a real conflict with a current mapping. 3481 * set flags to start unencaching all mappings 3482 * and return failure so we restart looping 3483 * the pp array from the beginning. 3484 */ 3485 return (HAT_TMPNC); 3486 } 3487 #endif /* VAC */ 3488 3489 /* 3490 * creates a large page shadow hmeblk for a tte. 3491 * The purpose of this routine is to allow us to do quick unloads because 3492 * the vm layer can easily pass a very large but sparsely populated range. 3493 */ 3494 static struct hme_blk * 3495 sfmmu_shadow_hcreate(sfmmu_t *sfmmup, caddr_t vaddr, int ttesz, uint_t flags) 3496 { 3497 struct hmehash_bucket *hmebp; 3498 hmeblk_tag hblktag; 3499 int hmeshift, size, vshift; 3500 uint_t shw_mask, newshw_mask; 3501 struct hme_blk *hmeblkp; 3502 3503 ASSERT(sfmmup != KHATID); 3504 if (mmu_page_sizes == max_mmu_page_sizes) { 3505 ASSERT(ttesz < TTE256M); 3506 } else { 3507 ASSERT(ttesz < TTE4M); 3508 ASSERT(sfmmup->sfmmu_ttecnt[TTE32M] == 0); 3509 ASSERT(sfmmup->sfmmu_ttecnt[TTE256M] == 0); 3510 } 3511 3512 if (ttesz == TTE8K) { 3513 size = TTE512K; 3514 } else { 3515 size = ++ttesz; 3516 } 3517 3518 hblktag.htag_id = sfmmup; 3519 hmeshift = HME_HASH_SHIFT(size); 3520 hblktag.htag_bspage = HME_HASH_BSPAGE(vaddr, hmeshift); 3521 hblktag.htag_rehash = HME_HASH_REHASH(size); 3522 hblktag.htag_rid = SFMMU_INVALID_SHMERID; 3523 hmebp = HME_HASH_FUNCTION(sfmmup, vaddr, hmeshift); 3524 3525 SFMMU_HASH_LOCK(hmebp); 3526 3527 HME_HASH_FAST_SEARCH(hmebp, hblktag, hmeblkp); 3528 ASSERT(hmeblkp != (struct hme_blk *)hblk_reserve); 3529 if (hmeblkp == NULL) { 3530 hmeblkp = sfmmu_hblk_alloc(sfmmup, vaddr, hmebp, size, 3531 hblktag, flags, SFMMU_INVALID_SHMERID); 3532 } 3533 ASSERT(hmeblkp); 3534 if (!hmeblkp->hblk_shw_mask) { 3535 /* 3536 * if this is a unused hblk it was just allocated or could 3537 * potentially be a previous large page hblk so we need to 3538 * set the shadow bit. 3539 */ 3540 ASSERT(!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt); 3541 hmeblkp->hblk_shw_bit = 1; 3542 } else if (hmeblkp->hblk_shw_bit == 0) { 3543 panic("sfmmu_shadow_hcreate: shw bit not set in hmeblkp 0x%p", 3544 (void *)hmeblkp); 3545 } 3546 ASSERT(hmeblkp->hblk_shw_bit == 1); 3547 ASSERT(!hmeblkp->hblk_shared); 3548 vshift = vaddr_to_vshift(hblktag, vaddr, size); 3549 ASSERT(vshift < 8); 3550 /* 3551 * Atomically set shw mask bit 3552 */ 3553 do { 3554 shw_mask = hmeblkp->hblk_shw_mask; 3555 newshw_mask = shw_mask | (1 << vshift); 3556 newshw_mask = cas32(&hmeblkp->hblk_shw_mask, shw_mask, 3557 newshw_mask); 3558 } while (newshw_mask != shw_mask); 3559 3560 SFMMU_HASH_UNLOCK(hmebp); 3561 3562 return (hmeblkp); 3563 } 3564 3565 /* 3566 * This routine cleanup a previous shadow hmeblk and changes it to 3567 * a regular hblk. This happens rarely but it is possible 3568 * when a process wants to use large pages and there are hblks still 3569 * lying around from the previous as that used these hmeblks. 3570 * The alternative was to cleanup the shadow hblks at unload time 3571 * but since so few user processes actually use large pages, it is 3572 * better to be lazy and cleanup at this time. 3573 */ 3574 static void 3575 sfmmu_shadow_hcleanup(sfmmu_t *sfmmup, struct hme_blk *hmeblkp, 3576 struct hmehash_bucket *hmebp) 3577 { 3578 caddr_t addr, endaddr; 3579 int hashno, size; 3580 3581 ASSERT(hmeblkp->hblk_shw_bit); 3582 ASSERT(!hmeblkp->hblk_shared); 3583 3584 ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp)); 3585 3586 if (!hmeblkp->hblk_shw_mask) { 3587 hmeblkp->hblk_shw_bit = 0; 3588 return; 3589 } 3590 addr = (caddr_t)get_hblk_base(hmeblkp); 3591 endaddr = get_hblk_endaddr(hmeblkp); 3592 size = get_hblk_ttesz(hmeblkp); 3593 hashno = size - 1; 3594 ASSERT(hashno > 0); 3595 SFMMU_HASH_UNLOCK(hmebp); 3596 3597 sfmmu_free_hblks(sfmmup, addr, endaddr, hashno); 3598 3599 SFMMU_HASH_LOCK(hmebp); 3600 } 3601 3602 static void 3603 sfmmu_free_hblks(sfmmu_t *sfmmup, caddr_t addr, caddr_t endaddr, 3604 int hashno) 3605 { 3606 int hmeshift, shadow = 0; 3607 hmeblk_tag hblktag; 3608 struct hmehash_bucket *hmebp; 3609 struct hme_blk *hmeblkp; 3610 struct hme_blk *nx_hblk, *pr_hblk, *list = NULL; 3611 3612 ASSERT(hashno > 0); 3613 hblktag.htag_id = sfmmup; 3614 hblktag.htag_rehash = hashno; 3615 hblktag.htag_rid = SFMMU_INVALID_SHMERID; 3616 3617 hmeshift = HME_HASH_SHIFT(hashno); 3618 3619 while (addr < endaddr) { 3620 hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift); 3621 hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift); 3622 SFMMU_HASH_LOCK(hmebp); 3623 /* inline HME_HASH_SEARCH */ 3624 hmeblkp = hmebp->hmeblkp; 3625 pr_hblk = NULL; 3626 while (hmeblkp) { 3627 if (HTAGS_EQ(hmeblkp->hblk_tag, hblktag)) { 3628 /* found hme_blk */ 3629 ASSERT(!hmeblkp->hblk_shared); 3630 if (hmeblkp->hblk_shw_bit) { 3631 if (hmeblkp->hblk_shw_mask) { 3632 shadow = 1; 3633 sfmmu_shadow_hcleanup(sfmmup, 3634 hmeblkp, hmebp); 3635 break; 3636 } else { 3637 hmeblkp->hblk_shw_bit = 0; 3638 } 3639 } 3640 3641 /* 3642 * Hblk_hmecnt and hblk_vcnt could be non zero 3643 * since hblk_unload() does not gurantee that. 3644 * 3645 * XXX - this could cause tteload() to spin 3646 * where sfmmu_shadow_hcleanup() is called. 3647 */ 3648 } 3649 3650 nx_hblk = hmeblkp->hblk_next; 3651 if (!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt) { 3652 sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk, 3653 &list, 0); 3654 } else { 3655 pr_hblk = hmeblkp; 3656 } 3657 hmeblkp = nx_hblk; 3658 } 3659 3660 SFMMU_HASH_UNLOCK(hmebp); 3661 3662 if (shadow) { 3663 /* 3664 * We found another shadow hblk so cleaned its 3665 * children. We need to go back and cleanup 3666 * the original hblk so we don't change the 3667 * addr. 3668 */ 3669 shadow = 0; 3670 } else { 3671 addr = (caddr_t)roundup((uintptr_t)addr + 1, 3672 (1 << hmeshift)); 3673 } 3674 } 3675 sfmmu_hblks_list_purge(&list, 0); 3676 } 3677 3678 /* 3679 * This routine's job is to delete stale invalid shared hmeregions hmeblks that 3680 * may still linger on after pageunload. 3681 */ 3682 static void 3683 sfmmu_cleanup_rhblk(sf_srd_t *srdp, caddr_t addr, uint_t rid, int ttesz) 3684 { 3685 int hmeshift; 3686 hmeblk_tag hblktag; 3687 struct hmehash_bucket *hmebp; 3688 struct hme_blk *hmeblkp; 3689 struct hme_blk *pr_hblk; 3690 struct hme_blk *list = NULL; 3691 3692 ASSERT(SFMMU_IS_SHMERID_VALID(rid)); 3693 ASSERT(rid < SFMMU_MAX_HME_REGIONS); 3694 3695 hmeshift = HME_HASH_SHIFT(ttesz); 3696 hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift); 3697 hblktag.htag_rehash = ttesz; 3698 hblktag.htag_rid = rid; 3699 hblktag.htag_id = srdp; 3700 hmebp = HME_HASH_FUNCTION(srdp, addr, hmeshift); 3701 3702 SFMMU_HASH_LOCK(hmebp); 3703 HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, pr_hblk, &list); 3704 if (hmeblkp != NULL) { 3705 ASSERT(hmeblkp->hblk_shared); 3706 ASSERT(!hmeblkp->hblk_shw_bit); 3707 if (hmeblkp->hblk_vcnt || hmeblkp->hblk_hmecnt) { 3708 panic("sfmmu_cleanup_rhblk: valid hmeblk"); 3709 } 3710 ASSERT(!hmeblkp->hblk_lckcnt); 3711 sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk, 3712 &list, 0); 3713 } 3714 SFMMU_HASH_UNLOCK(hmebp); 3715 sfmmu_hblks_list_purge(&list, 0); 3716 } 3717 3718 /* ARGSUSED */ 3719 static void 3720 sfmmu_rgn_cb_noop(caddr_t saddr, caddr_t eaddr, caddr_t r_saddr, 3721 size_t r_size, void *r_obj, u_offset_t r_objoff) 3722 { 3723 } 3724 3725 /* 3726 * Searches for an hmeblk which maps addr, then unloads this mapping 3727 * and updates *eaddrp, if the hmeblk is found. 3728 */ 3729 static void 3730 sfmmu_unload_hmeregion_va(sf_srd_t *srdp, uint_t rid, caddr_t addr, 3731 caddr_t eaddr, int ttesz, caddr_t *eaddrp) 3732 { 3733 int hmeshift; 3734 hmeblk_tag hblktag; 3735 struct hmehash_bucket *hmebp; 3736 struct hme_blk *hmeblkp; 3737 struct hme_blk *pr_hblk; 3738 struct hme_blk *list = NULL; 3739 3740 ASSERT(SFMMU_IS_SHMERID_VALID(rid)); 3741 ASSERT(rid < SFMMU_MAX_HME_REGIONS); 3742 ASSERT(ttesz >= HBLK_MIN_TTESZ); 3743 3744 hmeshift = HME_HASH_SHIFT(ttesz); 3745 hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift); 3746 hblktag.htag_rehash = ttesz; 3747 hblktag.htag_rid = rid; 3748 hblktag.htag_id = srdp; 3749 hmebp = HME_HASH_FUNCTION(srdp, addr, hmeshift); 3750 3751 SFMMU_HASH_LOCK(hmebp); 3752 HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, pr_hblk, &list); 3753 if (hmeblkp != NULL) { 3754 ASSERT(hmeblkp->hblk_shared); 3755 ASSERT(!hmeblkp->hblk_lckcnt); 3756 if (hmeblkp->hblk_vcnt || hmeblkp->hblk_hmecnt) { 3757 *eaddrp = sfmmu_hblk_unload(NULL, hmeblkp, addr, 3758 eaddr, NULL, HAT_UNLOAD); 3759 ASSERT(*eaddrp > addr); 3760 } 3761 ASSERT(!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt); 3762 sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk, 3763 &list, 0); 3764 } 3765 SFMMU_HASH_UNLOCK(hmebp); 3766 sfmmu_hblks_list_purge(&list, 0); 3767 } 3768 3769 static void 3770 sfmmu_unload_hmeregion(sf_srd_t *srdp, sf_region_t *rgnp) 3771 { 3772 int ttesz = rgnp->rgn_pgszc; 3773 size_t rsz = rgnp->rgn_size; 3774 caddr_t rsaddr = rgnp->rgn_saddr; 3775 caddr_t readdr = rsaddr + rsz; 3776 caddr_t rhsaddr; 3777 caddr_t va; 3778 uint_t rid = rgnp->rgn_id; 3779 caddr_t cbsaddr; 3780 caddr_t cbeaddr; 3781 hat_rgn_cb_func_t rcbfunc; 3782 ulong_t cnt; 3783 3784 ASSERT(SFMMU_IS_SHMERID_VALID(rid)); 3785 ASSERT(rid < SFMMU_MAX_HME_REGIONS); 3786 3787 ASSERT(IS_P2ALIGNED(rsaddr, TTEBYTES(ttesz))); 3788 ASSERT(IS_P2ALIGNED(rsz, TTEBYTES(ttesz))); 3789 if (ttesz < HBLK_MIN_TTESZ) { 3790 ttesz = HBLK_MIN_TTESZ; 3791 rhsaddr = (caddr_t)P2ALIGN((uintptr_t)rsaddr, HBLK_MIN_BYTES); 3792 } else { 3793 rhsaddr = rsaddr; 3794 } 3795 3796 if ((rcbfunc = rgnp->rgn_cb_function) == NULL) { 3797 rcbfunc = sfmmu_rgn_cb_noop; 3798 } 3799 3800 while (ttesz >= HBLK_MIN_TTESZ) { 3801 cbsaddr = rsaddr; 3802 cbeaddr = rsaddr; 3803 if (!(rgnp->rgn_hmeflags & (1 << ttesz))) { 3804 ttesz--; 3805 continue; 3806 } 3807 cnt = 0; 3808 va = rsaddr; 3809 while (va < readdr) { 3810 ASSERT(va >= rhsaddr); 3811 if (va != cbeaddr) { 3812 if (cbeaddr != cbsaddr) { 3813 ASSERT(cbeaddr > cbsaddr); 3814 (*rcbfunc)(cbsaddr, cbeaddr, 3815 rsaddr, rsz, rgnp->rgn_obj, 3816 rgnp->rgn_objoff); 3817 } 3818 cbsaddr = va; 3819 cbeaddr = va; 3820 } 3821 sfmmu_unload_hmeregion_va(srdp, rid, va, readdr, 3822 ttesz, &cbeaddr); 3823 cnt++; 3824 va = rhsaddr + (cnt << TTE_PAGE_SHIFT(ttesz)); 3825 } 3826 if (cbeaddr != cbsaddr) { 3827 ASSERT(cbeaddr > cbsaddr); 3828 (*rcbfunc)(cbsaddr, cbeaddr, rsaddr, 3829 rsz, rgnp->rgn_obj, 3830 rgnp->rgn_objoff); 3831 } 3832 ttesz--; 3833 } 3834 } 3835 3836 /* 3837 * Release one hardware address translation lock on the given address range. 3838 */ 3839 void 3840 hat_unlock(struct hat *sfmmup, caddr_t addr, size_t len) 3841 { 3842 struct hmehash_bucket *hmebp; 3843 hmeblk_tag hblktag; 3844 int hmeshift, hashno = 1; 3845 struct hme_blk *hmeblkp, *list = NULL; 3846 caddr_t endaddr; 3847 3848 ASSERT(sfmmup != NULL); 3849 ASSERT(sfmmup->sfmmu_xhat_provider == NULL); 3850 3851 ASSERT((sfmmup == ksfmmup) || 3852 AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock)); 3853 ASSERT((len & MMU_PAGEOFFSET) == 0); 3854 endaddr = addr + len; 3855 hblktag.htag_id = sfmmup; 3856 hblktag.htag_rid = SFMMU_INVALID_SHMERID; 3857 3858 /* 3859 * Spitfire supports 4 page sizes. 3860 * Most pages are expected to be of the smallest page size (8K) and 3861 * these will not need to be rehashed. 64K pages also don't need to be 3862 * rehashed because an hmeblk spans 64K of address space. 512K pages 3863 * might need 1 rehash and and 4M pages might need 2 rehashes. 3864 */ 3865 while (addr < endaddr) { 3866 hmeshift = HME_HASH_SHIFT(hashno); 3867 hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift); 3868 hblktag.htag_rehash = hashno; 3869 hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift); 3870 3871 SFMMU_HASH_LOCK(hmebp); 3872 3873 HME_HASH_SEARCH(hmebp, hblktag, hmeblkp, &list); 3874 if (hmeblkp != NULL) { 3875 ASSERT(!hmeblkp->hblk_shared); 3876 /* 3877 * If we encounter a shadow hmeblk then 3878 * we know there are no valid hmeblks mapping 3879 * this address at this size or larger. 3880 * Just increment address by the smallest 3881 * page size. 3882 */ 3883 if (hmeblkp->hblk_shw_bit) { 3884 addr += MMU_PAGESIZE; 3885 } else { 3886 addr = sfmmu_hblk_unlock(hmeblkp, addr, 3887 endaddr); 3888 } 3889 SFMMU_HASH_UNLOCK(hmebp); 3890 hashno = 1; 3891 continue; 3892 } 3893 SFMMU_HASH_UNLOCK(hmebp); 3894 3895 if (!HME_REHASH(sfmmup) || (hashno >= mmu_hashcnt)) { 3896 /* 3897 * We have traversed the whole list and rehashed 3898 * if necessary without finding the address to unlock 3899 * which should never happen. 3900 */ 3901 panic("sfmmu_unlock: addr not found. " 3902 "addr %p hat %p", (void *)addr, (void *)sfmmup); 3903 } else { 3904 hashno++; 3905 } 3906 } 3907 3908 sfmmu_hblks_list_purge(&list, 0); 3909 } 3910 3911 void 3912 hat_unlock_region(struct hat *sfmmup, caddr_t addr, size_t len, 3913 hat_region_cookie_t rcookie) 3914 { 3915 sf_srd_t *srdp; 3916 sf_region_t *rgnp; 3917 int ttesz; 3918 uint_t rid; 3919 caddr_t eaddr; 3920 caddr_t va; 3921 int hmeshift; 3922 hmeblk_tag hblktag; 3923 struct hmehash_bucket *hmebp; 3924 struct hme_blk *hmeblkp; 3925 struct hme_blk *pr_hblk; 3926 struct hme_blk *list; 3927 3928 if (rcookie == HAT_INVALID_REGION_COOKIE) { 3929 hat_unlock(sfmmup, addr, len); 3930 return; 3931 } 3932 3933 ASSERT(sfmmup != NULL); 3934 ASSERT(sfmmup->sfmmu_xhat_provider == NULL); 3935 ASSERT(sfmmup != ksfmmup); 3936 3937 srdp = sfmmup->sfmmu_srdp; 3938 rid = (uint_t)((uint64_t)rcookie); 3939 ASSERT(rid < SFMMU_MAX_HME_REGIONS); 3940 eaddr = addr + len; 3941 va = addr; 3942 list = NULL; 3943 rgnp = srdp->srd_hmergnp[rid]; 3944 SFMMU_VALIDATE_HMERID(sfmmup, rid, addr, len); 3945 3946 ASSERT(IS_P2ALIGNED(addr, TTEBYTES(rgnp->rgn_pgszc))); 3947 ASSERT(IS_P2ALIGNED(len, TTEBYTES(rgnp->rgn_pgszc))); 3948 if (rgnp->rgn_pgszc < HBLK_MIN_TTESZ) { 3949 ttesz = HBLK_MIN_TTESZ; 3950 } else { 3951 ttesz = rgnp->rgn_pgszc; 3952 } 3953 while (va < eaddr) { 3954 while (ttesz < rgnp->rgn_pgszc && 3955 IS_P2ALIGNED(va, TTEBYTES(ttesz + 1))) { 3956 ttesz++; 3957 } 3958 while (ttesz >= HBLK_MIN_TTESZ) { 3959 if (!(rgnp->rgn_hmeflags & (1 << ttesz))) { 3960 ttesz--; 3961 continue; 3962 } 3963 hmeshift = HME_HASH_SHIFT(ttesz); 3964 hblktag.htag_bspage = HME_HASH_BSPAGE(va, hmeshift); 3965 hblktag.htag_rehash = ttesz; 3966 hblktag.htag_rid = rid; 3967 hblktag.htag_id = srdp; 3968 hmebp = HME_HASH_FUNCTION(srdp, va, hmeshift); 3969 SFMMU_HASH_LOCK(hmebp); 3970 HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, pr_hblk, 3971 &list); 3972 if (hmeblkp == NULL) { 3973 SFMMU_HASH_UNLOCK(hmebp); 3974 ttesz--; 3975 continue; 3976 } 3977 ASSERT(hmeblkp->hblk_shared); 3978 va = sfmmu_hblk_unlock(hmeblkp, va, eaddr); 3979 ASSERT(va >= eaddr || 3980 IS_P2ALIGNED((uintptr_t)va, TTEBYTES(ttesz))); 3981 SFMMU_HASH_UNLOCK(hmebp); 3982 break; 3983 } 3984 if (ttesz < HBLK_MIN_TTESZ) { 3985 panic("hat_unlock_region: addr not found " 3986 "addr %p hat %p", (void *)va, (void *)sfmmup); 3987 } 3988 } 3989 sfmmu_hblks_list_purge(&list, 0); 3990 } 3991 3992 /* 3993 * Function to unlock a range of addresses in an hmeblk. It returns the 3994 * next address that needs to be unlocked. 3995 * Should be called with the hash lock held. 3996 */ 3997 static caddr_t 3998 sfmmu_hblk_unlock(struct hme_blk *hmeblkp, caddr_t addr, caddr_t endaddr) 3999 { 4000 struct sf_hment *sfhme; 4001 tte_t tteold, ttemod; 4002 int ttesz, ret; 4003 4004 ASSERT(in_hblk_range(hmeblkp, addr)); 4005 ASSERT(hmeblkp->hblk_shw_bit == 0); 4006 4007 endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp)); 4008 ttesz = get_hblk_ttesz(hmeblkp); 4009 4010 HBLKTOHME(sfhme, hmeblkp, addr); 4011 while (addr < endaddr) { 4012 readtte: 4013 sfmmu_copytte(&sfhme->hme_tte, &tteold); 4014 if (TTE_IS_VALID(&tteold)) { 4015 4016 ttemod = tteold; 4017 4018 ret = sfmmu_modifytte_try(&tteold, &ttemod, 4019 &sfhme->hme_tte); 4020 4021 if (ret < 0) 4022 goto readtte; 4023 4024 if (hmeblkp->hblk_lckcnt == 0) 4025 panic("zero hblk lckcnt"); 4026 4027 if (((uintptr_t)addr + TTEBYTES(ttesz)) > 4028 (uintptr_t)endaddr) 4029 panic("can't unlock large tte"); 4030 4031 ASSERT(hmeblkp->hblk_lckcnt > 0); 4032 atomic_add_32(&hmeblkp->hblk_lckcnt, -1); 4033 HBLK_STACK_TRACE(hmeblkp, HBLK_UNLOCK); 4034 } else { 4035 panic("sfmmu_hblk_unlock: invalid tte"); 4036 } 4037 addr += TTEBYTES(ttesz); 4038 sfhme++; 4039 } 4040 return (addr); 4041 } 4042 4043 /* 4044 * Physical Address Mapping Framework 4045 * 4046 * General rules: 4047 * 4048 * (1) Applies only to seg_kmem memory pages. To make things easier, 4049 * seg_kpm addresses are also accepted by the routines, but nothing 4050 * is done with them since by definition their PA mappings are static. 4051 * (2) hat_add_callback() may only be called while holding the page lock 4052 * SE_SHARED or SE_EXCL of the underlying page (e.g., as_pagelock()), 4053 * or passing HAC_PAGELOCK flag. 4054 * (3) prehandler() and posthandler() may not call hat_add_callback() or 4055 * hat_delete_callback(), nor should they allocate memory. Post quiesce 4056 * callbacks may not sleep or acquire adaptive mutex locks. 4057 * (4) Either prehandler() or posthandler() (but not both) may be specified 4058 * as being NULL. Specifying an errhandler() is optional. 4059 * 4060 * Details of using the framework: 4061 * 4062 * registering a callback (hat_register_callback()) 4063 * 4064 * Pass prehandler, posthandler, errhandler addresses 4065 * as described below. If capture_cpus argument is nonzero, 4066 * suspend callback to the prehandler will occur with CPUs 4067 * captured and executing xc_loop() and CPUs will remain 4068 * captured until after the posthandler suspend callback 4069 * occurs. 4070 * 4071 * adding a callback (hat_add_callback()) 4072 * 4073 * as_pagelock(); 4074 * hat_add_callback(); 4075 * save returned pfn in private data structures or program registers; 4076 * as_pageunlock(); 4077 * 4078 * prehandler() 4079 * 4080 * Stop all accesses by physical address to this memory page. 4081 * Called twice: the first, PRESUSPEND, is a context safe to acquire 4082 * adaptive locks. The second, SUSPEND, is called at high PIL with 4083 * CPUs captured so adaptive locks may NOT be acquired (and all spin 4084 * locks must be XCALL_PIL or higher locks). 4085 * 4086 * May return the following errors: 4087 * EIO: A fatal error has occurred. This will result in panic. 4088 * EAGAIN: The page cannot be suspended. This will fail the 4089 * relocation. 4090 * 0: Success. 4091 * 4092 * posthandler() 4093 * 4094 * Save new pfn in private data structures or program registers; 4095 * not allowed to fail (non-zero return values will result in panic). 4096 * 4097 * errhandler() 4098 * 4099 * called when an error occurs related to the callback. Currently 4100 * the only such error is HAT_CB_ERR_LEAKED which indicates that 4101 * a page is being freed, but there are still outstanding callback(s) 4102 * registered on the page. 4103 * 4104 * removing a callback (hat_delete_callback(); e.g., prior to freeing memory) 4105 * 4106 * stop using physical address 4107 * hat_delete_callback(); 4108 * 4109 */ 4110 4111 /* 4112 * Register a callback class. Each subsystem should do this once and 4113 * cache the id_t returned for use in setting up and tearing down callbacks. 4114 * 4115 * There is no facility for removing callback IDs once they are created; 4116 * the "key" should be unique for each module, so in case a module is unloaded 4117 * and subsequently re-loaded, we can recycle the module's previous entry. 4118 */ 4119 id_t 4120 hat_register_callback(int key, 4121 int (*prehandler)(caddr_t, uint_t, uint_t, void *), 4122 int (*posthandler)(caddr_t, uint_t, uint_t, void *, pfn_t), 4123 int (*errhandler)(caddr_t, uint_t, uint_t, void *), 4124 int capture_cpus) 4125 { 4126 id_t id; 4127 4128 /* 4129 * Search the table for a pre-existing callback associated with 4130 * the identifier "key". If one exists, we re-use that entry in 4131 * the table for this instance, otherwise we assign the next 4132 * available table slot. 4133 */ 4134 for (id = 0; id < sfmmu_max_cb_id; id++) { 4135 if (sfmmu_cb_table[id].key == key) 4136 break; 4137 } 4138 4139 if (id == sfmmu_max_cb_id) { 4140 id = sfmmu_cb_nextid++; 4141 if (id >= sfmmu_max_cb_id) 4142 panic("hat_register_callback: out of callback IDs"); 4143 } 4144 4145 ASSERT(prehandler != NULL || posthandler != NULL); 4146 4147 sfmmu_cb_table[id].key = key; 4148 sfmmu_cb_table[id].prehandler = prehandler; 4149 sfmmu_cb_table[id].posthandler = posthandler; 4150 sfmmu_cb_table[id].errhandler = errhandler; 4151 sfmmu_cb_table[id].capture_cpus = capture_cpus; 4152 4153 return (id); 4154 } 4155 4156 #define HAC_COOKIE_NONE (void *)-1 4157 4158 /* 4159 * Add relocation callbacks to the specified addr/len which will be called 4160 * when relocating the associated page. See the description of pre and 4161 * posthandler above for more details. 4162 * 4163 * If HAC_PAGELOCK is included in flags, the underlying memory page is 4164 * locked internally so the caller must be able to deal with the callback 4165 * running even before this function has returned. If HAC_PAGELOCK is not 4166 * set, it is assumed that the underlying memory pages are locked. 4167 * 4168 * Since the caller must track the individual page boundaries anyway, 4169 * we only allow a callback to be added to a single page (large 4170 * or small). Thus [addr, addr + len) MUST be contained within a single 4171 * page. 4172 * 4173 * Registering multiple callbacks on the same [addr, addr+len) is supported, 4174 * _provided_that_ a unique parameter is specified for each callback. 4175 * If multiple callbacks are registered on the same range the callback will 4176 * be invoked with each unique parameter. Registering the same callback with 4177 * the same argument more than once will result in corrupted kernel state. 4178 * 4179 * Returns the pfn of the underlying kernel page in *rpfn 4180 * on success, or PFN_INVALID on failure. 4181 * 4182 * cookiep (if passed) provides storage space for an opaque cookie 4183 * to return later to hat_delete_callback(). This cookie makes the callback 4184 * deletion significantly quicker by avoiding a potentially lengthy hash 4185 * search. 4186 * 4187 * Returns values: 4188 * 0: success 4189 * ENOMEM: memory allocation failure (e.g. flags was passed as HAC_NOSLEEP) 4190 * EINVAL: callback ID is not valid 4191 * ENXIO: ["vaddr", "vaddr" + len) is not mapped in the kernel's address 4192 * space 4193 * ERANGE: ["vaddr", "vaddr" + len) crosses a page boundary 4194 */ 4195 int 4196 hat_add_callback(id_t callback_id, caddr_t vaddr, uint_t len, uint_t flags, 4197 void *pvt, pfn_t *rpfn, void **cookiep) 4198 { 4199 struct hmehash_bucket *hmebp; 4200 hmeblk_tag hblktag; 4201 struct hme_blk *hmeblkp; 4202 int hmeshift, hashno; 4203 caddr_t saddr, eaddr, baseaddr; 4204 struct pa_hment *pahmep; 4205 struct sf_hment *sfhmep, *osfhmep; 4206 kmutex_t *pml; 4207 tte_t tte; 4208 page_t *pp; 4209 vnode_t *vp; 4210 u_offset_t off; 4211 pfn_t pfn; 4212 int kmflags = (flags & HAC_SLEEP)? KM_SLEEP : KM_NOSLEEP; 4213 int locked = 0; 4214 4215 /* 4216 * For KPM mappings, just return the physical address since we 4217 * don't need to register any callbacks. 4218 */ 4219 if (IS_KPM_ADDR(vaddr)) { 4220 uint64_t paddr; 4221 SFMMU_KPM_VTOP(vaddr, paddr); 4222 *rpfn = btop(paddr); 4223 if (cookiep != NULL) 4224 *cookiep = HAC_COOKIE_NONE; 4225 return (0); 4226 } 4227 4228 if (callback_id < (id_t)0 || callback_id >= sfmmu_cb_nextid) { 4229 *rpfn = PFN_INVALID; 4230 return (EINVAL); 4231 } 4232 4233 if ((pahmep = kmem_cache_alloc(pa_hment_cache, kmflags)) == NULL) { 4234 *rpfn = PFN_INVALID; 4235 return (ENOMEM); 4236 } 4237 4238 sfhmep = &pahmep->sfment; 4239 4240 saddr = (caddr_t)((uintptr_t)vaddr & MMU_PAGEMASK); 4241 eaddr = saddr + len; 4242 4243 rehash: 4244 /* Find the mapping(s) for this page */ 4245 for (hashno = TTE64K, hmeblkp = NULL; 4246 hmeblkp == NULL && hashno <= mmu_hashcnt; 4247 hashno++) { 4248 hmeshift = HME_HASH_SHIFT(hashno); 4249 hblktag.htag_id = ksfmmup; 4250 hblktag.htag_rid = SFMMU_INVALID_SHMERID; 4251 hblktag.htag_bspage = HME_HASH_BSPAGE(saddr, hmeshift); 4252 hblktag.htag_rehash = hashno; 4253 hmebp = HME_HASH_FUNCTION(ksfmmup, saddr, hmeshift); 4254 4255 SFMMU_HASH_LOCK(hmebp); 4256 4257 HME_HASH_FAST_SEARCH(hmebp, hblktag, hmeblkp); 4258 4259 if (hmeblkp == NULL) 4260 SFMMU_HASH_UNLOCK(hmebp); 4261 } 4262 4263 if (hmeblkp == NULL) { 4264 kmem_cache_free(pa_hment_cache, pahmep); 4265 *rpfn = PFN_INVALID; 4266 return (ENXIO); 4267 } 4268 4269 ASSERT(!hmeblkp->hblk_shared); 4270 4271 HBLKTOHME(osfhmep, hmeblkp, saddr); 4272 sfmmu_copytte(&osfhmep->hme_tte, &tte); 4273 4274 if (!TTE_IS_VALID(&tte)) { 4275 SFMMU_HASH_UNLOCK(hmebp); 4276 kmem_cache_free(pa_hment_cache, pahmep); 4277 *rpfn = PFN_INVALID; 4278 return (ENXIO); 4279 } 4280 4281 /* 4282 * Make sure the boundaries for the callback fall within this 4283 * single mapping. 4284 */ 4285 baseaddr = (caddr_t)get_hblk_base(hmeblkp); 4286 ASSERT(saddr >= baseaddr); 4287 if (eaddr > saddr + TTEBYTES(TTE_CSZ(&tte))) { 4288 SFMMU_HASH_UNLOCK(hmebp); 4289 kmem_cache_free(pa_hment_cache, pahmep); 4290 *rpfn = PFN_INVALID; 4291 return (ERANGE); 4292 } 4293 4294 pfn = sfmmu_ttetopfn(&tte, vaddr); 4295 4296 /* 4297 * The pfn may not have a page_t underneath in which case we 4298 * just return it. This can happen if we are doing I/O to a 4299 * static portion of the kernel's address space, for instance. 4300 */ 4301 pp = osfhmep->hme_page; 4302 if (pp == NULL) { 4303 SFMMU_HASH_UNLOCK(hmebp); 4304 kmem_cache_free(pa_hment_cache, pahmep); 4305 *rpfn = pfn; 4306 if (cookiep) 4307 *cookiep = HAC_COOKIE_NONE; 4308 return (0); 4309 } 4310 ASSERT(pp == PP_PAGEROOT(pp)); 4311 4312 vp = pp->p_vnode; 4313 off = pp->p_offset; 4314 4315 pml = sfmmu_mlist_enter(pp); 4316 4317 if (flags & HAC_PAGELOCK) { 4318 if (!page_trylock(pp, SE_SHARED)) { 4319 /* 4320 * Somebody is holding SE_EXCL lock. Might 4321 * even be hat_page_relocate(). Drop all 4322 * our locks, lookup the page in &kvp, and 4323 * retry. If it doesn't exist in &kvp and &zvp, 4324 * then we must be dealing with a kernel mapped 4325 * page which doesn't actually belong to 4326 * segkmem so we punt. 4327 */ 4328 sfmmu_mlist_exit(pml); 4329 SFMMU_HASH_UNLOCK(hmebp); 4330 pp = page_lookup(&kvp, (u_offset_t)saddr, SE_SHARED); 4331 4332 /* check zvp before giving up */ 4333 if (pp == NULL) 4334 pp = page_lookup(&zvp, (u_offset_t)saddr, 4335 SE_SHARED); 4336 4337 /* Okay, we didn't find it, give up */ 4338 if (pp == NULL) { 4339 kmem_cache_free(pa_hment_cache, pahmep); 4340 *rpfn = pfn; 4341 if (cookiep) 4342 *cookiep = HAC_COOKIE_NONE; 4343 return (0); 4344 } 4345 page_unlock(pp); 4346 goto rehash; 4347 } 4348 locked = 1; 4349 } 4350 4351 if (!PAGE_LOCKED(pp) && !panicstr) 4352 panic("hat_add_callback: page 0x%p not locked", (void *)pp); 4353 4354 if (osfhmep->hme_page != pp || pp->p_vnode != vp || 4355 pp->p_offset != off) { 4356 /* 4357 * The page moved before we got our hands on it. Drop 4358 * all the locks and try again. 4359 */ 4360 ASSERT((flags & HAC_PAGELOCK) != 0); 4361 sfmmu_mlist_exit(pml); 4362 SFMMU_HASH_UNLOCK(hmebp); 4363 page_unlock(pp); 4364 locked = 0; 4365 goto rehash; 4366 } 4367 4368 if (!VN_ISKAS(vp)) { 4369 /* 4370 * This is not a segkmem page but another page which 4371 * has been kernel mapped. It had better have at least 4372 * a share lock on it. Return the pfn. 4373 */ 4374 sfmmu_mlist_exit(pml); 4375 SFMMU_HASH_UNLOCK(hmebp); 4376 if (locked) 4377 page_unlock(pp); 4378 kmem_cache_free(pa_hment_cache, pahmep); 4379 ASSERT(PAGE_LOCKED(pp)); 4380 *rpfn = pfn; 4381 if (cookiep) 4382 *cookiep = HAC_COOKIE_NONE; 4383 return (0); 4384 } 4385 4386 /* 4387 * Setup this pa_hment and link its embedded dummy sf_hment into 4388 * the mapping list. 4389 */ 4390 pp->p_share++; 4391 pahmep->cb_id = callback_id; 4392 pahmep->addr = vaddr; 4393 pahmep->len = len; 4394 pahmep->refcnt = 1; 4395 pahmep->flags = 0; 4396 pahmep->pvt = pvt; 4397 4398 sfhmep->hme_tte.ll = 0; 4399 sfhmep->hme_data = pahmep; 4400 sfhmep->hme_prev = osfhmep; 4401 sfhmep->hme_next = osfhmep->hme_next; 4402 4403 if (osfhmep->hme_next) 4404 osfhmep->hme_next->hme_prev = sfhmep; 4405 4406 osfhmep->hme_next = sfhmep; 4407 4408 sfmmu_mlist_exit(pml); 4409 SFMMU_HASH_UNLOCK(hmebp); 4410 4411 if (locked) 4412 page_unlock(pp); 4413 4414 *rpfn = pfn; 4415 if (cookiep) 4416 *cookiep = (void *)pahmep; 4417 4418 return (0); 4419 } 4420 4421 /* 4422 * Remove the relocation callbacks from the specified addr/len. 4423 */ 4424 void 4425 hat_delete_callback(caddr_t vaddr, uint_t len, void *pvt, uint_t flags, 4426 void *cookie) 4427 { 4428 struct hmehash_bucket *hmebp; 4429 hmeblk_tag hblktag; 4430 struct hme_blk *hmeblkp; 4431 int hmeshift, hashno; 4432 caddr_t saddr; 4433 struct pa_hment *pahmep; 4434 struct sf_hment *sfhmep, *osfhmep; 4435 kmutex_t *pml; 4436 tte_t tte; 4437 page_t *pp; 4438 vnode_t *vp; 4439 u_offset_t off; 4440 int locked = 0; 4441 4442 /* 4443 * If the cookie is HAC_COOKIE_NONE then there is no pa_hment to 4444 * remove so just return. 4445 */ 4446 if (cookie == HAC_COOKIE_NONE || IS_KPM_ADDR(vaddr)) 4447 return; 4448 4449 saddr = (caddr_t)((uintptr_t)vaddr & MMU_PAGEMASK); 4450 4451 rehash: 4452 /* Find the mapping(s) for this page */ 4453 for (hashno = TTE64K, hmeblkp = NULL; 4454 hmeblkp == NULL && hashno <= mmu_hashcnt; 4455 hashno++) { 4456 hmeshift = HME_HASH_SHIFT(hashno); 4457 hblktag.htag_id = ksfmmup; 4458 hblktag.htag_rid = SFMMU_INVALID_SHMERID; 4459 hblktag.htag_bspage = HME_HASH_BSPAGE(saddr, hmeshift); 4460 hblktag.htag_rehash = hashno; 4461 hmebp = HME_HASH_FUNCTION(ksfmmup, saddr, hmeshift); 4462 4463 SFMMU_HASH_LOCK(hmebp); 4464 4465 HME_HASH_FAST_SEARCH(hmebp, hblktag, hmeblkp); 4466 4467 if (hmeblkp == NULL) 4468 SFMMU_HASH_UNLOCK(hmebp); 4469 } 4470 4471 if (hmeblkp == NULL) 4472 return; 4473 4474 ASSERT(!hmeblkp->hblk_shared); 4475 4476 HBLKTOHME(osfhmep, hmeblkp, saddr); 4477 4478 sfmmu_copytte(&osfhmep->hme_tte, &tte); 4479 if (!TTE_IS_VALID(&tte)) { 4480 SFMMU_HASH_UNLOCK(hmebp); 4481 return; 4482 } 4483 4484 pp = osfhmep->hme_page; 4485 if (pp == NULL) { 4486 SFMMU_HASH_UNLOCK(hmebp); 4487 ASSERT(cookie == NULL); 4488 return; 4489 } 4490 4491 vp = pp->p_vnode; 4492 off = pp->p_offset; 4493 4494 pml = sfmmu_mlist_enter(pp); 4495 4496 if (flags & HAC_PAGELOCK) { 4497 if (!page_trylock(pp, SE_SHARED)) { 4498 /* 4499 * Somebody is holding SE_EXCL lock. Might 4500 * even be hat_page_relocate(). Drop all 4501 * our locks, lookup the page in &kvp, and 4502 * retry. If it doesn't exist in &kvp and &zvp, 4503 * then we must be dealing with a kernel mapped 4504 * page which doesn't actually belong to 4505 * segkmem so we punt. 4506 */ 4507 sfmmu_mlist_exit(pml); 4508 SFMMU_HASH_UNLOCK(hmebp); 4509 pp = page_lookup(&kvp, (u_offset_t)saddr, SE_SHARED); 4510 /* check zvp before giving up */ 4511 if (pp == NULL) 4512 pp = page_lookup(&zvp, (u_offset_t)saddr, 4513 SE_SHARED); 4514 4515 if (pp == NULL) { 4516 ASSERT(cookie == NULL); 4517 return; 4518 } 4519 page_unlock(pp); 4520 goto rehash; 4521 } 4522 locked = 1; 4523 } 4524 4525 ASSERT(PAGE_LOCKED(pp)); 4526 4527 if (osfhmep->hme_page != pp || pp->p_vnode != vp || 4528 pp->p_offset != off) { 4529 /* 4530 * The page moved before we got our hands on it. Drop 4531 * all the locks and try again. 4532 */ 4533 ASSERT((flags & HAC_PAGELOCK) != 0); 4534 sfmmu_mlist_exit(pml); 4535 SFMMU_HASH_UNLOCK(hmebp); 4536 page_unlock(pp); 4537 locked = 0; 4538 goto rehash; 4539 } 4540 4541 if (!VN_ISKAS(vp)) { 4542 /* 4543 * This is not a segkmem page but another page which 4544 * has been kernel mapped. 4545 */ 4546 sfmmu_mlist_exit(pml); 4547 SFMMU_HASH_UNLOCK(hmebp); 4548 if (locked) 4549 page_unlock(pp); 4550 ASSERT(cookie == NULL); 4551 return; 4552 } 4553 4554 if (cookie != NULL) { 4555 pahmep = (struct pa_hment *)cookie; 4556 sfhmep = &pahmep->sfment; 4557 } else { 4558 for (sfhmep = pp->p_mapping; sfhmep != NULL; 4559 sfhmep = sfhmep->hme_next) { 4560 4561 /* 4562 * skip va<->pa mappings 4563 */ 4564 if (!IS_PAHME(sfhmep)) 4565 continue; 4566 4567 pahmep = sfhmep->hme_data; 4568 ASSERT(pahmep != NULL); 4569 4570 /* 4571 * if pa_hment matches, remove it 4572 */ 4573 if ((pahmep->pvt == pvt) && 4574 (pahmep->addr == vaddr) && 4575 (pahmep->len == len)) { 4576 break; 4577 } 4578 } 4579 } 4580 4581 if (sfhmep == NULL) { 4582 if (!panicstr) { 4583 panic("hat_delete_callback: pa_hment not found, pp %p", 4584 (void *)pp); 4585 } 4586 return; 4587 } 4588 4589 /* 4590 * Note: at this point a valid kernel mapping must still be 4591 * present on this page. 4592 */ 4593 pp->p_share--; 4594 if (pp->p_share <= 0) 4595 panic("hat_delete_callback: zero p_share"); 4596 4597 if (--pahmep->refcnt == 0) { 4598 if (pahmep->flags != 0) 4599 panic("hat_delete_callback: pa_hment is busy"); 4600 4601 /* 4602 * Remove sfhmep from the mapping list for the page. 4603 */ 4604 if (sfhmep->hme_prev) { 4605 sfhmep->hme_prev->hme_next = sfhmep->hme_next; 4606 } else { 4607 pp->p_mapping = sfhmep->hme_next; 4608 } 4609 4610 if (sfhmep->hme_next) 4611 sfhmep->hme_next->hme_prev = sfhmep->hme_prev; 4612 4613 sfmmu_mlist_exit(pml); 4614 SFMMU_HASH_UNLOCK(hmebp); 4615 4616 if (locked) 4617 page_unlock(pp); 4618 4619 kmem_cache_free(pa_hment_cache, pahmep); 4620 return; 4621 } 4622 4623 sfmmu_mlist_exit(pml); 4624 SFMMU_HASH_UNLOCK(hmebp); 4625 if (locked) 4626 page_unlock(pp); 4627 } 4628 4629 /* 4630 * hat_probe returns 1 if the translation for the address 'addr' is 4631 * loaded, zero otherwise. 4632 * 4633 * hat_probe should be used only for advisorary purposes because it may 4634 * occasionally return the wrong value. The implementation must guarantee that 4635 * returning the wrong value is a very rare event. hat_probe is used 4636 * to implement optimizations in the segment drivers. 4637 * 4638 */ 4639 int 4640 hat_probe(struct hat *sfmmup, caddr_t addr) 4641 { 4642 pfn_t pfn; 4643 tte_t tte; 4644 4645 ASSERT(sfmmup != NULL); 4646 ASSERT(sfmmup->sfmmu_xhat_provider == NULL); 4647 4648 ASSERT((sfmmup == ksfmmup) || 4649 AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock)); 4650 4651 if (sfmmup == ksfmmup) { 4652 while ((pfn = sfmmu_vatopfn(addr, sfmmup, &tte)) 4653 == PFN_SUSPENDED) { 4654 sfmmu_vatopfn_suspended(addr, sfmmup, &tte); 4655 } 4656 } else { 4657 pfn = sfmmu_uvatopfn(addr, sfmmup, NULL); 4658 } 4659 4660 if (pfn != PFN_INVALID) 4661 return (1); 4662 else 4663 return (0); 4664 } 4665 4666 ssize_t 4667 hat_getpagesize(struct hat *sfmmup, caddr_t addr) 4668 { 4669 tte_t tte; 4670 4671 ASSERT(sfmmup->sfmmu_xhat_provider == NULL); 4672 4673 if (sfmmup == ksfmmup) { 4674 if (sfmmu_vatopfn(addr, sfmmup, &tte) == PFN_INVALID) { 4675 return (-1); 4676 } 4677 } else { 4678 if (sfmmu_uvatopfn(addr, sfmmup, &tte) == PFN_INVALID) { 4679 return (-1); 4680 } 4681 } 4682 4683 ASSERT(TTE_IS_VALID(&tte)); 4684 return (TTEBYTES(TTE_CSZ(&tte))); 4685 } 4686 4687 uint_t 4688 hat_getattr(struct hat *sfmmup, caddr_t addr, uint_t *attr) 4689 { 4690 tte_t tte; 4691 4692 ASSERT(sfmmup->sfmmu_xhat_provider == NULL); 4693 4694 if (sfmmup == ksfmmup) { 4695 if (sfmmu_vatopfn(addr, sfmmup, &tte) == PFN_INVALID) { 4696 tte.ll = 0; 4697 } 4698 } else { 4699 if (sfmmu_uvatopfn(addr, sfmmup, &tte) == PFN_INVALID) { 4700 tte.ll = 0; 4701 } 4702 } 4703 if (TTE_IS_VALID(&tte)) { 4704 *attr = sfmmu_ptov_attr(&tte); 4705 return (0); 4706 } 4707 *attr = 0; 4708 return ((uint_t)0xffffffff); 4709 } 4710 4711 /* 4712 * Enables more attributes on specified address range (ie. logical OR) 4713 */ 4714 void 4715 hat_setattr(struct hat *hat, caddr_t addr, size_t len, uint_t attr) 4716 { 4717 if (hat->sfmmu_xhat_provider) { 4718 XHAT_SETATTR(hat, addr, len, attr); 4719 return; 4720 } else { 4721 /* 4722 * This must be a CPU HAT. If the address space has 4723 * XHATs attached, change attributes for all of them, 4724 * just in case 4725 */ 4726 ASSERT(hat->sfmmu_as != NULL); 4727 if (hat->sfmmu_as->a_xhat != NULL) 4728 xhat_setattr_all(hat->sfmmu_as, addr, len, attr); 4729 } 4730 4731 sfmmu_chgattr(hat, addr, len, attr, SFMMU_SETATTR); 4732 } 4733 4734 /* 4735 * Assigns attributes to the specified address range. All the attributes 4736 * are specified. 4737 */ 4738 void 4739 hat_chgattr(struct hat *hat, caddr_t addr, size_t len, uint_t attr) 4740 { 4741 if (hat->sfmmu_xhat_provider) { 4742 XHAT_CHGATTR(hat, addr, len, attr); 4743 return; 4744 } else { 4745 /* 4746 * This must be a CPU HAT. If the address space has 4747 * XHATs attached, change attributes for all of them, 4748 * just in case 4749 */ 4750 ASSERT(hat->sfmmu_as != NULL); 4751 if (hat->sfmmu_as->a_xhat != NULL) 4752 xhat_chgattr_all(hat->sfmmu_as, addr, len, attr); 4753 } 4754 4755 sfmmu_chgattr(hat, addr, len, attr, SFMMU_CHGATTR); 4756 } 4757 4758 /* 4759 * Remove attributes on the specified address range (ie. loginal NAND) 4760 */ 4761 void 4762 hat_clrattr(struct hat *hat, caddr_t addr, size_t len, uint_t attr) 4763 { 4764 if (hat->sfmmu_xhat_provider) { 4765 XHAT_CLRATTR(hat, addr, len, attr); 4766 return; 4767 } else { 4768 /* 4769 * This must be a CPU HAT. If the address space has 4770 * XHATs attached, change attributes for all of them, 4771 * just in case 4772 */ 4773 ASSERT(hat->sfmmu_as != NULL); 4774 if (hat->sfmmu_as->a_xhat != NULL) 4775 xhat_clrattr_all(hat->sfmmu_as, addr, len, attr); 4776 } 4777 4778 sfmmu_chgattr(hat, addr, len, attr, SFMMU_CLRATTR); 4779 } 4780 4781 /* 4782 * Change attributes on an address range to that specified by attr and mode. 4783 */ 4784 static void 4785 sfmmu_chgattr(struct hat *sfmmup, caddr_t addr, size_t len, uint_t attr, 4786 int mode) 4787 { 4788 struct hmehash_bucket *hmebp; 4789 hmeblk_tag hblktag; 4790 int hmeshift, hashno = 1; 4791 struct hme_blk *hmeblkp, *list = NULL; 4792 caddr_t endaddr; 4793 cpuset_t cpuset; 4794 demap_range_t dmr; 4795 4796 CPUSET_ZERO(cpuset); 4797 4798 ASSERT((sfmmup == ksfmmup) || 4799 AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock)); 4800 ASSERT((len & MMU_PAGEOFFSET) == 0); 4801 ASSERT(((uintptr_t)addr & MMU_PAGEOFFSET) == 0); 4802 4803 if ((attr & PROT_USER) && (mode != SFMMU_CLRATTR) && 4804 ((addr + len) > (caddr_t)USERLIMIT)) { 4805 panic("user addr %p in kernel space", 4806 (void *)addr); 4807 } 4808 4809 endaddr = addr + len; 4810 hblktag.htag_id = sfmmup; 4811 hblktag.htag_rid = SFMMU_INVALID_SHMERID; 4812 DEMAP_RANGE_INIT(sfmmup, &dmr); 4813 4814 while (addr < endaddr) { 4815 hmeshift = HME_HASH_SHIFT(hashno); 4816 hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift); 4817 hblktag.htag_rehash = hashno; 4818 hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift); 4819 4820 SFMMU_HASH_LOCK(hmebp); 4821 4822 HME_HASH_SEARCH(hmebp, hblktag, hmeblkp, &list); 4823 if (hmeblkp != NULL) { 4824 ASSERT(!hmeblkp->hblk_shared); 4825 /* 4826 * We've encountered a shadow hmeblk so skip the range 4827 * of the next smaller mapping size. 4828 */ 4829 if (hmeblkp->hblk_shw_bit) { 4830 ASSERT(sfmmup != ksfmmup); 4831 ASSERT(hashno > 1); 4832 addr = (caddr_t)P2END((uintptr_t)addr, 4833 TTEBYTES(hashno - 1)); 4834 } else { 4835 addr = sfmmu_hblk_chgattr(sfmmup, 4836 hmeblkp, addr, endaddr, &dmr, attr, mode); 4837 } 4838 SFMMU_HASH_UNLOCK(hmebp); 4839 hashno = 1; 4840 continue; 4841 } 4842 SFMMU_HASH_UNLOCK(hmebp); 4843 4844 if (!HME_REHASH(sfmmup) || (hashno >= mmu_hashcnt)) { 4845 /* 4846 * We have traversed the whole list and rehashed 4847 * if necessary without finding the address to chgattr. 4848 * This is ok, so we increment the address by the 4849 * smallest hmeblk range for kernel mappings or for 4850 * user mappings with no large pages, and the largest 4851 * hmeblk range, to account for shadow hmeblks, for 4852 * user mappings with large pages and continue. 4853 */ 4854 if (sfmmup == ksfmmup) 4855 addr = (caddr_t)P2END((uintptr_t)addr, 4856 TTEBYTES(1)); 4857 else 4858 addr = (caddr_t)P2END((uintptr_t)addr, 4859 TTEBYTES(hashno)); 4860 hashno = 1; 4861 } else { 4862 hashno++; 4863 } 4864 } 4865 4866 sfmmu_hblks_list_purge(&list, 0); 4867 DEMAP_RANGE_FLUSH(&dmr); 4868 cpuset = sfmmup->sfmmu_cpusran; 4869 xt_sync(cpuset); 4870 } 4871 4872 /* 4873 * This function chgattr on a range of addresses in an hmeblk. It returns the 4874 * next addres that needs to be chgattr. 4875 * It should be called with the hash lock held. 4876 * XXX It should be possible to optimize chgattr by not flushing every time but 4877 * on the other hand: 4878 * 1. do one flush crosscall. 4879 * 2. only flush if we are increasing permissions (make sure this will work) 4880 */ 4881 static caddr_t 4882 sfmmu_hblk_chgattr(struct hat *sfmmup, struct hme_blk *hmeblkp, caddr_t addr, 4883 caddr_t endaddr, demap_range_t *dmrp, uint_t attr, int mode) 4884 { 4885 tte_t tte, tteattr, tteflags, ttemod; 4886 struct sf_hment *sfhmep; 4887 int ttesz; 4888 struct page *pp = NULL; 4889 kmutex_t *pml, *pmtx; 4890 int ret; 4891 int use_demap_range; 4892 #if defined(SF_ERRATA_57) 4893 int check_exec; 4894 #endif 4895 4896 ASSERT(in_hblk_range(hmeblkp, addr)); 4897 ASSERT(hmeblkp->hblk_shw_bit == 0); 4898 ASSERT(!hmeblkp->hblk_shared); 4899 4900 endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp)); 4901 ttesz = get_hblk_ttesz(hmeblkp); 4902 4903 /* 4904 * Flush the current demap region if addresses have been 4905 * skipped or the page size doesn't match. 4906 */ 4907 use_demap_range = (TTEBYTES(ttesz) == DEMAP_RANGE_PGSZ(dmrp)); 4908 if (use_demap_range) { 4909 DEMAP_RANGE_CONTINUE(dmrp, addr, endaddr); 4910 } else { 4911 DEMAP_RANGE_FLUSH(dmrp); 4912 } 4913 4914 tteattr.ll = sfmmu_vtop_attr(attr, mode, &tteflags); 4915 #if defined(SF_ERRATA_57) 4916 check_exec = (sfmmup != ksfmmup) && 4917 AS_TYPE_64BIT(sfmmup->sfmmu_as) && 4918 TTE_IS_EXECUTABLE(&tteattr); 4919 #endif 4920 HBLKTOHME(sfhmep, hmeblkp, addr); 4921 while (addr < endaddr) { 4922 sfmmu_copytte(&sfhmep->hme_tte, &tte); 4923 if (TTE_IS_VALID(&tte)) { 4924 if ((tte.ll & tteflags.ll) == tteattr.ll) { 4925 /* 4926 * if the new attr is the same as old 4927 * continue 4928 */ 4929 goto next_addr; 4930 } 4931 if (!TTE_IS_WRITABLE(&tteattr)) { 4932 /* 4933 * make sure we clear hw modify bit if we 4934 * removing write protections 4935 */ 4936 tteflags.tte_intlo |= TTE_HWWR_INT; 4937 } 4938 4939 pml = NULL; 4940 pp = sfhmep->hme_page; 4941 if (pp) { 4942 pml = sfmmu_mlist_enter(pp); 4943 } 4944 4945 if (pp != sfhmep->hme_page) { 4946 /* 4947 * tte must have been unloaded. 4948 */ 4949 ASSERT(pml); 4950 sfmmu_mlist_exit(pml); 4951 continue; 4952 } 4953 4954 ASSERT(pp == NULL || sfmmu_mlist_held(pp)); 4955 4956 ttemod = tte; 4957 ttemod.ll = (ttemod.ll & ~tteflags.ll) | tteattr.ll; 4958 ASSERT(TTE_TO_TTEPFN(&ttemod) == TTE_TO_TTEPFN(&tte)); 4959 4960 #if defined(SF_ERRATA_57) 4961 if (check_exec && addr < errata57_limit) 4962 ttemod.tte_exec_perm = 0; 4963 #endif 4964 ret = sfmmu_modifytte_try(&tte, &ttemod, 4965 &sfhmep->hme_tte); 4966 4967 if (ret < 0) { 4968 /* tte changed underneath us */ 4969 if (pml) { 4970 sfmmu_mlist_exit(pml); 4971 } 4972 continue; 4973 } 4974 4975 if (tteflags.tte_intlo & TTE_HWWR_INT) { 4976 /* 4977 * need to sync if we are clearing modify bit. 4978 */ 4979 sfmmu_ttesync(sfmmup, addr, &tte, pp); 4980 } 4981 4982 if (pp && PP_ISRO(pp)) { 4983 if (tteattr.tte_intlo & TTE_WRPRM_INT) { 4984 pmtx = sfmmu_page_enter(pp); 4985 PP_CLRRO(pp); 4986 sfmmu_page_exit(pmtx); 4987 } 4988 } 4989 4990 if (ret > 0 && use_demap_range) { 4991 DEMAP_RANGE_MARKPG(dmrp, addr); 4992 } else if (ret > 0) { 4993 sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 0, 0); 4994 } 4995 4996 if (pml) { 4997 sfmmu_mlist_exit(pml); 4998 } 4999 } 5000 next_addr: 5001 addr += TTEBYTES(ttesz); 5002 sfhmep++; 5003 DEMAP_RANGE_NEXTPG(dmrp); 5004 } 5005 return (addr); 5006 } 5007 5008 /* 5009 * This routine converts virtual attributes to physical ones. It will 5010 * update the tteflags field with the tte mask corresponding to the attributes 5011 * affected and it returns the new attributes. It will also clear the modify 5012 * bit if we are taking away write permission. This is necessary since the 5013 * modify bit is the hardware permission bit and we need to clear it in order 5014 * to detect write faults. 5015 */ 5016 static uint64_t 5017 sfmmu_vtop_attr(uint_t attr, int mode, tte_t *ttemaskp) 5018 { 5019 tte_t ttevalue; 5020 5021 ASSERT(!(attr & ~SFMMU_LOAD_ALLATTR)); 5022 5023 switch (mode) { 5024 case SFMMU_CHGATTR: 5025 /* all attributes specified */ 5026 ttevalue.tte_inthi = MAKE_TTEATTR_INTHI(attr); 5027 ttevalue.tte_intlo = MAKE_TTEATTR_INTLO(attr); 5028 ttemaskp->tte_inthi = TTEINTHI_ATTR; 5029 ttemaskp->tte_intlo = TTEINTLO_ATTR; 5030 break; 5031 case SFMMU_SETATTR: 5032 ASSERT(!(attr & ~HAT_PROT_MASK)); 5033 ttemaskp->ll = 0; 5034 ttevalue.ll = 0; 5035 /* 5036 * a valid tte implies exec and read for sfmmu 5037 * so no need to do anything about them. 5038 * since priviledged access implies user access 5039 * PROT_USER doesn't make sense either. 5040 */ 5041 if (attr & PROT_WRITE) { 5042 ttemaskp->tte_intlo |= TTE_WRPRM_INT; 5043 ttevalue.tte_intlo |= TTE_WRPRM_INT; 5044 } 5045 break; 5046 case SFMMU_CLRATTR: 5047 /* attributes will be nand with current ones */ 5048 if (attr & ~(PROT_WRITE | PROT_USER)) { 5049 panic("sfmmu: attr %x not supported", attr); 5050 } 5051 ttemaskp->ll = 0; 5052 ttevalue.ll = 0; 5053 if (attr & PROT_WRITE) { 5054 /* clear both writable and modify bit */ 5055 ttemaskp->tte_intlo |= TTE_WRPRM_INT | TTE_HWWR_INT; 5056 } 5057 if (attr & PROT_USER) { 5058 ttemaskp->tte_intlo |= TTE_PRIV_INT; 5059 ttevalue.tte_intlo |= TTE_PRIV_INT; 5060 } 5061 break; 5062 default: 5063 panic("sfmmu_vtop_attr: bad mode %x", mode); 5064 } 5065 ASSERT(TTE_TO_TTEPFN(&ttevalue) == 0); 5066 return (ttevalue.ll); 5067 } 5068 5069 static uint_t 5070 sfmmu_ptov_attr(tte_t *ttep) 5071 { 5072 uint_t attr; 5073 5074 ASSERT(TTE_IS_VALID(ttep)); 5075 5076 attr = PROT_READ; 5077 5078 if (TTE_IS_WRITABLE(ttep)) { 5079 attr |= PROT_WRITE; 5080 } 5081 if (TTE_IS_EXECUTABLE(ttep)) { 5082 attr |= PROT_EXEC; 5083 } 5084 if (!TTE_IS_PRIVILEGED(ttep)) { 5085 attr |= PROT_USER; 5086 } 5087 if (TTE_IS_NFO(ttep)) { 5088 attr |= HAT_NOFAULT; 5089 } 5090 if (TTE_IS_NOSYNC(ttep)) { 5091 attr |= HAT_NOSYNC; 5092 } 5093 if (TTE_IS_SIDEFFECT(ttep)) { 5094 attr |= SFMMU_SIDEFFECT; 5095 } 5096 if (!TTE_IS_VCACHEABLE(ttep)) { 5097 attr |= SFMMU_UNCACHEVTTE; 5098 } 5099 if (!TTE_IS_PCACHEABLE(ttep)) { 5100 attr |= SFMMU_UNCACHEPTTE; 5101 } 5102 return (attr); 5103 } 5104 5105 /* 5106 * hat_chgprot is a deprecated hat call. New segment drivers 5107 * should store all attributes and use hat_*attr calls. 5108 * 5109 * Change the protections in the virtual address range 5110 * given to the specified virtual protection. If vprot is ~PROT_WRITE, 5111 * then remove write permission, leaving the other 5112 * permissions unchanged. If vprot is ~PROT_USER, remove user permissions. 5113 * 5114 */ 5115 void 5116 hat_chgprot(struct hat *sfmmup, caddr_t addr, size_t len, uint_t vprot) 5117 { 5118 struct hmehash_bucket *hmebp; 5119 hmeblk_tag hblktag; 5120 int hmeshift, hashno = 1; 5121 struct hme_blk *hmeblkp, *list = NULL; 5122 caddr_t endaddr; 5123 cpuset_t cpuset; 5124 demap_range_t dmr; 5125 5126 ASSERT((len & MMU_PAGEOFFSET) == 0); 5127 ASSERT(((uintptr_t)addr & MMU_PAGEOFFSET) == 0); 5128 5129 if (sfmmup->sfmmu_xhat_provider) { 5130 XHAT_CHGPROT(sfmmup, addr, len, vprot); 5131 return; 5132 } else { 5133 /* 5134 * This must be a CPU HAT. If the address space has 5135 * XHATs attached, change attributes for all of them, 5136 * just in case 5137 */ 5138 ASSERT(sfmmup->sfmmu_as != NULL); 5139 if (sfmmup->sfmmu_as->a_xhat != NULL) 5140 xhat_chgprot_all(sfmmup->sfmmu_as, addr, len, vprot); 5141 } 5142 5143 CPUSET_ZERO(cpuset); 5144 5145 if ((vprot != (uint_t)~PROT_WRITE) && (vprot & PROT_USER) && 5146 ((addr + len) > (caddr_t)USERLIMIT)) { 5147 panic("user addr %p vprot %x in kernel space", 5148 (void *)addr, vprot); 5149 } 5150 endaddr = addr + len; 5151 hblktag.htag_id = sfmmup; 5152 hblktag.htag_rid = SFMMU_INVALID_SHMERID; 5153 DEMAP_RANGE_INIT(sfmmup, &dmr); 5154 5155 while (addr < endaddr) { 5156 hmeshift = HME_HASH_SHIFT(hashno); 5157 hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift); 5158 hblktag.htag_rehash = hashno; 5159 hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift); 5160 5161 SFMMU_HASH_LOCK(hmebp); 5162 5163 HME_HASH_SEARCH(hmebp, hblktag, hmeblkp, &list); 5164 if (hmeblkp != NULL) { 5165 ASSERT(!hmeblkp->hblk_shared); 5166 /* 5167 * We've encountered a shadow hmeblk so skip the range 5168 * of the next smaller mapping size. 5169 */ 5170 if (hmeblkp->hblk_shw_bit) { 5171 ASSERT(sfmmup != ksfmmup); 5172 ASSERT(hashno > 1); 5173 addr = (caddr_t)P2END((uintptr_t)addr, 5174 TTEBYTES(hashno - 1)); 5175 } else { 5176 addr = sfmmu_hblk_chgprot(sfmmup, hmeblkp, 5177 addr, endaddr, &dmr, vprot); 5178 } 5179 SFMMU_HASH_UNLOCK(hmebp); 5180 hashno = 1; 5181 continue; 5182 } 5183 SFMMU_HASH_UNLOCK(hmebp); 5184 5185 if (!HME_REHASH(sfmmup) || (hashno >= mmu_hashcnt)) { 5186 /* 5187 * We have traversed the whole list and rehashed 5188 * if necessary without finding the address to chgprot. 5189 * This is ok so we increment the address by the 5190 * smallest hmeblk range for kernel mappings and the 5191 * largest hmeblk range, to account for shadow hmeblks, 5192 * for user mappings and continue. 5193 */ 5194 if (sfmmup == ksfmmup) 5195 addr = (caddr_t)P2END((uintptr_t)addr, 5196 TTEBYTES(1)); 5197 else 5198 addr = (caddr_t)P2END((uintptr_t)addr, 5199 TTEBYTES(hashno)); 5200 hashno = 1; 5201 } else { 5202 hashno++; 5203 } 5204 } 5205 5206 sfmmu_hblks_list_purge(&list, 0); 5207 DEMAP_RANGE_FLUSH(&dmr); 5208 cpuset = sfmmup->sfmmu_cpusran; 5209 xt_sync(cpuset); 5210 } 5211 5212 /* 5213 * This function chgprots a range of addresses in an hmeblk. It returns the 5214 * next addres that needs to be chgprot. 5215 * It should be called with the hash lock held. 5216 * XXX It shold be possible to optimize chgprot by not flushing every time but 5217 * on the other hand: 5218 * 1. do one flush crosscall. 5219 * 2. only flush if we are increasing permissions (make sure this will work) 5220 */ 5221 static caddr_t 5222 sfmmu_hblk_chgprot(sfmmu_t *sfmmup, struct hme_blk *hmeblkp, caddr_t addr, 5223 caddr_t endaddr, demap_range_t *dmrp, uint_t vprot) 5224 { 5225 uint_t pprot; 5226 tte_t tte, ttemod; 5227 struct sf_hment *sfhmep; 5228 uint_t tteflags; 5229 int ttesz; 5230 struct page *pp = NULL; 5231 kmutex_t *pml, *pmtx; 5232 int ret; 5233 int use_demap_range; 5234 #if defined(SF_ERRATA_57) 5235 int check_exec; 5236 #endif 5237 5238 ASSERT(in_hblk_range(hmeblkp, addr)); 5239 ASSERT(hmeblkp->hblk_shw_bit == 0); 5240 ASSERT(!hmeblkp->hblk_shared); 5241 5242 #ifdef DEBUG 5243 if (get_hblk_ttesz(hmeblkp) != TTE8K && 5244 (endaddr < get_hblk_endaddr(hmeblkp))) { 5245 panic("sfmmu_hblk_chgprot: partial chgprot of large page"); 5246 } 5247 #endif /* DEBUG */ 5248 5249 endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp)); 5250 ttesz = get_hblk_ttesz(hmeblkp); 5251 5252 pprot = sfmmu_vtop_prot(vprot, &tteflags); 5253 #if defined(SF_ERRATA_57) 5254 check_exec = (sfmmup != ksfmmup) && 5255 AS_TYPE_64BIT(sfmmup->sfmmu_as) && 5256 ((vprot & PROT_EXEC) == PROT_EXEC); 5257 #endif 5258 HBLKTOHME(sfhmep, hmeblkp, addr); 5259 5260 /* 5261 * Flush the current demap region if addresses have been 5262 * skipped or the page size doesn't match. 5263 */ 5264 use_demap_range = (TTEBYTES(ttesz) == MMU_PAGESIZE); 5265 if (use_demap_range) { 5266 DEMAP_RANGE_CONTINUE(dmrp, addr, endaddr); 5267 } else { 5268 DEMAP_RANGE_FLUSH(dmrp); 5269 } 5270 5271 while (addr < endaddr) { 5272 sfmmu_copytte(&sfhmep->hme_tte, &tte); 5273 if (TTE_IS_VALID(&tte)) { 5274 if (TTE_GET_LOFLAGS(&tte, tteflags) == pprot) { 5275 /* 5276 * if the new protection is the same as old 5277 * continue 5278 */ 5279 goto next_addr; 5280 } 5281 pml = NULL; 5282 pp = sfhmep->hme_page; 5283 if (pp) { 5284 pml = sfmmu_mlist_enter(pp); 5285 } 5286 if (pp != sfhmep->hme_page) { 5287 /* 5288 * tte most have been unloaded 5289 * underneath us. Recheck 5290 */ 5291 ASSERT(pml); 5292 sfmmu_mlist_exit(pml); 5293 continue; 5294 } 5295 5296 ASSERT(pp == NULL || sfmmu_mlist_held(pp)); 5297 5298 ttemod = tte; 5299 TTE_SET_LOFLAGS(&ttemod, tteflags, pprot); 5300 #if defined(SF_ERRATA_57) 5301 if (check_exec && addr < errata57_limit) 5302 ttemod.tte_exec_perm = 0; 5303 #endif 5304 ret = sfmmu_modifytte_try(&tte, &ttemod, 5305 &sfhmep->hme_tte); 5306 5307 if (ret < 0) { 5308 /* tte changed underneath us */ 5309 if (pml) { 5310 sfmmu_mlist_exit(pml); 5311 } 5312 continue; 5313 } 5314 5315 if (tteflags & TTE_HWWR_INT) { 5316 /* 5317 * need to sync if we are clearing modify bit. 5318 */ 5319 sfmmu_ttesync(sfmmup, addr, &tte, pp); 5320 } 5321 5322 if (pp && PP_ISRO(pp)) { 5323 if (pprot & TTE_WRPRM_INT) { 5324 pmtx = sfmmu_page_enter(pp); 5325 PP_CLRRO(pp); 5326 sfmmu_page_exit(pmtx); 5327 } 5328 } 5329 5330 if (ret > 0 && use_demap_range) { 5331 DEMAP_RANGE_MARKPG(dmrp, addr); 5332 } else if (ret > 0) { 5333 sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 0, 0); 5334 } 5335 5336 if (pml) { 5337 sfmmu_mlist_exit(pml); 5338 } 5339 } 5340 next_addr: 5341 addr += TTEBYTES(ttesz); 5342 sfhmep++; 5343 DEMAP_RANGE_NEXTPG(dmrp); 5344 } 5345 return (addr); 5346 } 5347 5348 /* 5349 * This routine is deprecated and should only be used by hat_chgprot. 5350 * The correct routine is sfmmu_vtop_attr. 5351 * This routine converts virtual page protections to physical ones. It will 5352 * update the tteflags field with the tte mask corresponding to the protections 5353 * affected and it returns the new protections. It will also clear the modify 5354 * bit if we are taking away write permission. This is necessary since the 5355 * modify bit is the hardware permission bit and we need to clear it in order 5356 * to detect write faults. 5357 * It accepts the following special protections: 5358 * ~PROT_WRITE = remove write permissions. 5359 * ~PROT_USER = remove user permissions. 5360 */ 5361 static uint_t 5362 sfmmu_vtop_prot(uint_t vprot, uint_t *tteflagsp) 5363 { 5364 if (vprot == (uint_t)~PROT_WRITE) { 5365 *tteflagsp = TTE_WRPRM_INT | TTE_HWWR_INT; 5366 return (0); /* will cause wrprm to be cleared */ 5367 } 5368 if (vprot == (uint_t)~PROT_USER) { 5369 *tteflagsp = TTE_PRIV_INT; 5370 return (0); /* will cause privprm to be cleared */ 5371 } 5372 if ((vprot == 0) || (vprot == PROT_USER) || 5373 ((vprot & PROT_ALL) != vprot)) { 5374 panic("sfmmu_vtop_prot -- bad prot %x", vprot); 5375 } 5376 5377 switch (vprot) { 5378 case (PROT_READ): 5379 case (PROT_EXEC): 5380 case (PROT_EXEC | PROT_READ): 5381 *tteflagsp = TTE_PRIV_INT | TTE_WRPRM_INT | TTE_HWWR_INT; 5382 return (TTE_PRIV_INT); /* set prv and clr wrt */ 5383 case (PROT_WRITE): 5384 case (PROT_WRITE | PROT_READ): 5385 case (PROT_EXEC | PROT_WRITE): 5386 case (PROT_EXEC | PROT_WRITE | PROT_READ): 5387 *tteflagsp = TTE_PRIV_INT | TTE_WRPRM_INT; 5388 return (TTE_PRIV_INT | TTE_WRPRM_INT); /* set prv and wrt */ 5389 case (PROT_USER | PROT_READ): 5390 case (PROT_USER | PROT_EXEC): 5391 case (PROT_USER | PROT_EXEC | PROT_READ): 5392 *tteflagsp = TTE_PRIV_INT | TTE_WRPRM_INT | TTE_HWWR_INT; 5393 return (0); /* clr prv and wrt */ 5394 case (PROT_USER | PROT_WRITE): 5395 case (PROT_USER | PROT_WRITE | PROT_READ): 5396 case (PROT_USER | PROT_EXEC | PROT_WRITE): 5397 case (PROT_USER | PROT_EXEC | PROT_WRITE | PROT_READ): 5398 *tteflagsp = TTE_PRIV_INT | TTE_WRPRM_INT; 5399 return (TTE_WRPRM_INT); /* clr prv and set wrt */ 5400 default: 5401 panic("sfmmu_vtop_prot -- bad prot %x", vprot); 5402 } 5403 return (0); 5404 } 5405 5406 /* 5407 * Alternate unload for very large virtual ranges. With a true 64 bit VA, 5408 * the normal algorithm would take too long for a very large VA range with 5409 * few real mappings. This routine just walks thru all HMEs in the global 5410 * hash table to find and remove mappings. 5411 */ 5412 static void 5413 hat_unload_large_virtual( 5414 struct hat *sfmmup, 5415 caddr_t startaddr, 5416 size_t len, 5417 uint_t flags, 5418 hat_callback_t *callback) 5419 { 5420 struct hmehash_bucket *hmebp; 5421 struct hme_blk *hmeblkp; 5422 struct hme_blk *pr_hblk = NULL; 5423 struct hme_blk *nx_hblk; 5424 struct hme_blk *list = NULL; 5425 int i; 5426 demap_range_t dmr, *dmrp; 5427 cpuset_t cpuset; 5428 caddr_t endaddr = startaddr + len; 5429 caddr_t sa; 5430 caddr_t ea; 5431 caddr_t cb_sa[MAX_CB_ADDR]; 5432 caddr_t cb_ea[MAX_CB_ADDR]; 5433 int addr_cnt = 0; 5434 int a = 0; 5435 5436 if (sfmmup->sfmmu_free) { 5437 dmrp = NULL; 5438 } else { 5439 dmrp = &dmr; 5440 DEMAP_RANGE_INIT(sfmmup, dmrp); 5441 } 5442 5443 /* 5444 * Loop through all the hash buckets of HME blocks looking for matches. 5445 */ 5446 for (i = 0; i <= UHMEHASH_SZ; i++) { 5447 hmebp = &uhme_hash[i]; 5448 SFMMU_HASH_LOCK(hmebp); 5449 hmeblkp = hmebp->hmeblkp; 5450 pr_hblk = NULL; 5451 while (hmeblkp) { 5452 nx_hblk = hmeblkp->hblk_next; 5453 5454 /* 5455 * skip if not this context, if a shadow block or 5456 * if the mapping is not in the requested range 5457 */ 5458 if (hmeblkp->hblk_tag.htag_id != sfmmup || 5459 hmeblkp->hblk_shw_bit || 5460 (sa = (caddr_t)get_hblk_base(hmeblkp)) >= endaddr || 5461 (ea = get_hblk_endaddr(hmeblkp)) <= startaddr) { 5462 pr_hblk = hmeblkp; 5463 goto next_block; 5464 } 5465 5466 ASSERT(!hmeblkp->hblk_shared); 5467 /* 5468 * unload if there are any current valid mappings 5469 */ 5470 if (hmeblkp->hblk_vcnt != 0 || 5471 hmeblkp->hblk_hmecnt != 0) 5472 (void) sfmmu_hblk_unload(sfmmup, hmeblkp, 5473 sa, ea, dmrp, flags); 5474 5475 /* 5476 * on unmap we also release the HME block itself, once 5477 * all mappings are gone. 5478 */ 5479 if ((flags & HAT_UNLOAD_UNMAP) != 0 && 5480 !hmeblkp->hblk_vcnt && 5481 !hmeblkp->hblk_hmecnt) { 5482 ASSERT(!hmeblkp->hblk_lckcnt); 5483 sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk, 5484 &list, 0); 5485 } else { 5486 pr_hblk = hmeblkp; 5487 } 5488 5489 if (callback == NULL) 5490 goto next_block; 5491 5492 /* 5493 * HME blocks may span more than one page, but we may be 5494 * unmapping only one page, so check for a smaller range 5495 * for the callback 5496 */ 5497 if (sa < startaddr) 5498 sa = startaddr; 5499 if (--ea > endaddr) 5500 ea = endaddr - 1; 5501 5502 cb_sa[addr_cnt] = sa; 5503 cb_ea[addr_cnt] = ea; 5504 if (++addr_cnt == MAX_CB_ADDR) { 5505 if (dmrp != NULL) { 5506 DEMAP_RANGE_FLUSH(dmrp); 5507 cpuset = sfmmup->sfmmu_cpusran; 5508 xt_sync(cpuset); 5509 } 5510 5511 for (a = 0; a < MAX_CB_ADDR; ++a) { 5512 callback->hcb_start_addr = cb_sa[a]; 5513 callback->hcb_end_addr = cb_ea[a]; 5514 callback->hcb_function(callback); 5515 } 5516 addr_cnt = 0; 5517 } 5518 5519 next_block: 5520 hmeblkp = nx_hblk; 5521 } 5522 SFMMU_HASH_UNLOCK(hmebp); 5523 } 5524 5525 sfmmu_hblks_list_purge(&list, 0); 5526 if (dmrp != NULL) { 5527 DEMAP_RANGE_FLUSH(dmrp); 5528 cpuset = sfmmup->sfmmu_cpusran; 5529 xt_sync(cpuset); 5530 } 5531 5532 for (a = 0; a < addr_cnt; ++a) { 5533 callback->hcb_start_addr = cb_sa[a]; 5534 callback->hcb_end_addr = cb_ea[a]; 5535 callback->hcb_function(callback); 5536 } 5537 5538 /* 5539 * Check TSB and TLB page sizes if the process isn't exiting. 5540 */ 5541 if (!sfmmup->sfmmu_free) 5542 sfmmu_check_page_sizes(sfmmup, 0); 5543 } 5544 5545 /* 5546 * Unload all the mappings in the range [addr..addr+len). addr and len must 5547 * be MMU_PAGESIZE aligned. 5548 */ 5549 5550 extern struct seg *segkmap; 5551 #define ISSEGKMAP(sfmmup, addr) (sfmmup == ksfmmup && \ 5552 segkmap->s_base <= (addr) && (addr) < (segkmap->s_base + segkmap->s_size)) 5553 5554 5555 void 5556 hat_unload_callback( 5557 struct hat *sfmmup, 5558 caddr_t addr, 5559 size_t len, 5560 uint_t flags, 5561 hat_callback_t *callback) 5562 { 5563 struct hmehash_bucket *hmebp; 5564 hmeblk_tag hblktag; 5565 int hmeshift, hashno, iskernel; 5566 struct hme_blk *hmeblkp, *pr_hblk, *list = NULL; 5567 caddr_t endaddr; 5568 cpuset_t cpuset; 5569 int addr_count = 0; 5570 int a; 5571 caddr_t cb_start_addr[MAX_CB_ADDR]; 5572 caddr_t cb_end_addr[MAX_CB_ADDR]; 5573 int issegkmap = ISSEGKMAP(sfmmup, addr); 5574 demap_range_t dmr, *dmrp; 5575 5576 if (sfmmup->sfmmu_xhat_provider) { 5577 XHAT_UNLOAD_CALLBACK(sfmmup, addr, len, flags, callback); 5578 return; 5579 } else { 5580 /* 5581 * This must be a CPU HAT. If the address space has 5582 * XHATs attached, unload the mappings for all of them, 5583 * just in case 5584 */ 5585 ASSERT(sfmmup->sfmmu_as != NULL); 5586 if (sfmmup->sfmmu_as->a_xhat != NULL) 5587 xhat_unload_callback_all(sfmmup->sfmmu_as, addr, 5588 len, flags, callback); 5589 } 5590 5591 ASSERT((sfmmup == ksfmmup) || (flags & HAT_UNLOAD_OTHER) || \ 5592 AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock)); 5593 5594 ASSERT(sfmmup != NULL); 5595 ASSERT((len & MMU_PAGEOFFSET) == 0); 5596 ASSERT(!((uintptr_t)addr & MMU_PAGEOFFSET)); 5597 5598 /* 5599 * Probing through a large VA range (say 63 bits) will be slow, even 5600 * at 4 Meg steps between the probes. So, when the virtual address range 5601 * is very large, search the HME entries for what to unload. 5602 * 5603 * len >> TTE_PAGE_SHIFT(TTE4M) is the # of 4Meg probes we'd need 5604 * 5605 * UHMEHASH_SZ is number of hash buckets to examine 5606 * 5607 */ 5608 if (sfmmup != KHATID && (len >> TTE_PAGE_SHIFT(TTE4M)) > UHMEHASH_SZ) { 5609 hat_unload_large_virtual(sfmmup, addr, len, flags, callback); 5610 return; 5611 } 5612 5613 CPUSET_ZERO(cpuset); 5614 5615 /* 5616 * If the process is exiting, we can save a lot of fuss since 5617 * we'll flush the TLB when we free the ctx anyway. 5618 */ 5619 if (sfmmup->sfmmu_free) 5620 dmrp = NULL; 5621 else 5622 dmrp = &dmr; 5623 5624 DEMAP_RANGE_INIT(sfmmup, dmrp); 5625 endaddr = addr + len; 5626 hblktag.htag_id = sfmmup; 5627 hblktag.htag_rid = SFMMU_INVALID_SHMERID; 5628 5629 /* 5630 * It is likely for the vm to call unload over a wide range of 5631 * addresses that are actually very sparsely populated by 5632 * translations. In order to speed this up the sfmmu hat supports 5633 * the concept of shadow hmeblks. Dummy large page hmeblks that 5634 * correspond to actual small translations are allocated at tteload 5635 * time and are referred to as shadow hmeblks. Now, during unload 5636 * time, we first check if we have a shadow hmeblk for that 5637 * translation. The absence of one means the corresponding address 5638 * range is empty and can be skipped. 5639 * 5640 * The kernel is an exception to above statement and that is why 5641 * we don't use shadow hmeblks and hash starting from the smallest 5642 * page size. 5643 */ 5644 if (sfmmup == KHATID) { 5645 iskernel = 1; 5646 hashno = TTE64K; 5647 } else { 5648 iskernel = 0; 5649 if (mmu_page_sizes == max_mmu_page_sizes) { 5650 hashno = TTE256M; 5651 } else { 5652 hashno = TTE4M; 5653 } 5654 } 5655 while (addr < endaddr) { 5656 hmeshift = HME_HASH_SHIFT(hashno); 5657 hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift); 5658 hblktag.htag_rehash = hashno; 5659 hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift); 5660 5661 SFMMU_HASH_LOCK(hmebp); 5662 5663 HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, pr_hblk, &list); 5664 if (hmeblkp == NULL) { 5665 /* 5666 * didn't find an hmeblk. skip the appropiate 5667 * address range. 5668 */ 5669 SFMMU_HASH_UNLOCK(hmebp); 5670 if (iskernel) { 5671 if (hashno < mmu_hashcnt) { 5672 hashno++; 5673 continue; 5674 } else { 5675 hashno = TTE64K; 5676 addr = (caddr_t)roundup((uintptr_t)addr 5677 + 1, MMU_PAGESIZE64K); 5678 continue; 5679 } 5680 } 5681 addr = (caddr_t)roundup((uintptr_t)addr + 1, 5682 (1 << hmeshift)); 5683 if ((uintptr_t)addr & MMU_PAGEOFFSET512K) { 5684 ASSERT(hashno == TTE64K); 5685 continue; 5686 } 5687 if ((uintptr_t)addr & MMU_PAGEOFFSET4M) { 5688 hashno = TTE512K; 5689 continue; 5690 } 5691 if (mmu_page_sizes == max_mmu_page_sizes) { 5692 if ((uintptr_t)addr & MMU_PAGEOFFSET32M) { 5693 hashno = TTE4M; 5694 continue; 5695 } 5696 if ((uintptr_t)addr & MMU_PAGEOFFSET256M) { 5697 hashno = TTE32M; 5698 continue; 5699 } 5700 hashno = TTE256M; 5701 continue; 5702 } else { 5703 hashno = TTE4M; 5704 continue; 5705 } 5706 } 5707 ASSERT(hmeblkp); 5708 ASSERT(!hmeblkp->hblk_shared); 5709 if (!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt) { 5710 /* 5711 * If the valid count is zero we can skip the range 5712 * mapped by this hmeblk. 5713 * We free hblks in the case of HAT_UNMAP. HAT_UNMAP 5714 * is used by segment drivers as a hint 5715 * that the mapping resource won't be used any longer. 5716 * The best example of this is during exit(). 5717 */ 5718 addr = (caddr_t)roundup((uintptr_t)addr + 1, 5719 get_hblk_span(hmeblkp)); 5720 if ((flags & HAT_UNLOAD_UNMAP) || 5721 (iskernel && !issegkmap)) { 5722 sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk, 5723 &list, 0); 5724 } 5725 SFMMU_HASH_UNLOCK(hmebp); 5726 5727 if (iskernel) { 5728 hashno = TTE64K; 5729 continue; 5730 } 5731 if ((uintptr_t)addr & MMU_PAGEOFFSET512K) { 5732 ASSERT(hashno == TTE64K); 5733 continue; 5734 } 5735 if ((uintptr_t)addr & MMU_PAGEOFFSET4M) { 5736 hashno = TTE512K; 5737 continue; 5738 } 5739 if (mmu_page_sizes == max_mmu_page_sizes) { 5740 if ((uintptr_t)addr & MMU_PAGEOFFSET32M) { 5741 hashno = TTE4M; 5742 continue; 5743 } 5744 if ((uintptr_t)addr & MMU_PAGEOFFSET256M) { 5745 hashno = TTE32M; 5746 continue; 5747 } 5748 hashno = TTE256M; 5749 continue; 5750 } else { 5751 hashno = TTE4M; 5752 continue; 5753 } 5754 } 5755 if (hmeblkp->hblk_shw_bit) { 5756 /* 5757 * If we encounter a shadow hmeblk we know there is 5758 * smaller sized hmeblks mapping the same address space. 5759 * Decrement the hash size and rehash. 5760 */ 5761 ASSERT(sfmmup != KHATID); 5762 hashno--; 5763 SFMMU_HASH_UNLOCK(hmebp); 5764 continue; 5765 } 5766 5767 /* 5768 * track callback address ranges. 5769 * only start a new range when it's not contiguous 5770 */ 5771 if (callback != NULL) { 5772 if (addr_count > 0 && 5773 addr == cb_end_addr[addr_count - 1]) 5774 --addr_count; 5775 else 5776 cb_start_addr[addr_count] = addr; 5777 } 5778 5779 addr = sfmmu_hblk_unload(sfmmup, hmeblkp, addr, endaddr, 5780 dmrp, flags); 5781 5782 if (callback != NULL) 5783 cb_end_addr[addr_count++] = addr; 5784 5785 if (((flags & HAT_UNLOAD_UNMAP) || (iskernel && !issegkmap)) && 5786 !hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt) { 5787 sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk, &list, 0); 5788 } 5789 SFMMU_HASH_UNLOCK(hmebp); 5790 5791 /* 5792 * Notify our caller as to exactly which pages 5793 * have been unloaded. We do these in clumps, 5794 * to minimize the number of xt_sync()s that need to occur. 5795 */ 5796 if (callback != NULL && addr_count == MAX_CB_ADDR) { 5797 DEMAP_RANGE_FLUSH(dmrp); 5798 if (dmrp != NULL) { 5799 cpuset = sfmmup->sfmmu_cpusran; 5800 xt_sync(cpuset); 5801 } 5802 5803 for (a = 0; a < MAX_CB_ADDR; ++a) { 5804 callback->hcb_start_addr = cb_start_addr[a]; 5805 callback->hcb_end_addr = cb_end_addr[a]; 5806 callback->hcb_function(callback); 5807 } 5808 addr_count = 0; 5809 } 5810 if (iskernel) { 5811 hashno = TTE64K; 5812 continue; 5813 } 5814 if ((uintptr_t)addr & MMU_PAGEOFFSET512K) { 5815 ASSERT(hashno == TTE64K); 5816 continue; 5817 } 5818 if ((uintptr_t)addr & MMU_PAGEOFFSET4M) { 5819 hashno = TTE512K; 5820 continue; 5821 } 5822 if (mmu_page_sizes == max_mmu_page_sizes) { 5823 if ((uintptr_t)addr & MMU_PAGEOFFSET32M) { 5824 hashno = TTE4M; 5825 continue; 5826 } 5827 if ((uintptr_t)addr & MMU_PAGEOFFSET256M) { 5828 hashno = TTE32M; 5829 continue; 5830 } 5831 hashno = TTE256M; 5832 } else { 5833 hashno = TTE4M; 5834 } 5835 } 5836 5837 sfmmu_hblks_list_purge(&list, 0); 5838 DEMAP_RANGE_FLUSH(dmrp); 5839 if (dmrp != NULL) { 5840 cpuset = sfmmup->sfmmu_cpusran; 5841 xt_sync(cpuset); 5842 } 5843 if (callback && addr_count != 0) { 5844 for (a = 0; a < addr_count; ++a) { 5845 callback->hcb_start_addr = cb_start_addr[a]; 5846 callback->hcb_end_addr = cb_end_addr[a]; 5847 callback->hcb_function(callback); 5848 } 5849 } 5850 5851 /* 5852 * Check TSB and TLB page sizes if the process isn't exiting. 5853 */ 5854 if (!sfmmup->sfmmu_free) 5855 sfmmu_check_page_sizes(sfmmup, 0); 5856 } 5857 5858 /* 5859 * Unload all the mappings in the range [addr..addr+len). addr and len must 5860 * be MMU_PAGESIZE aligned. 5861 */ 5862 void 5863 hat_unload(struct hat *sfmmup, caddr_t addr, size_t len, uint_t flags) 5864 { 5865 if (sfmmup->sfmmu_xhat_provider) { 5866 XHAT_UNLOAD(sfmmup, addr, len, flags); 5867 return; 5868 } 5869 hat_unload_callback(sfmmup, addr, len, flags, NULL); 5870 } 5871 5872 5873 /* 5874 * Find the largest mapping size for this page. 5875 */ 5876 int 5877 fnd_mapping_sz(page_t *pp) 5878 { 5879 int sz; 5880 int p_index; 5881 5882 p_index = PP_MAPINDEX(pp); 5883 5884 sz = 0; 5885 p_index >>= 1; /* don't care about 8K bit */ 5886 for (; p_index; p_index >>= 1) { 5887 sz++; 5888 } 5889 5890 return (sz); 5891 } 5892 5893 /* 5894 * This function unloads a range of addresses for an hmeblk. 5895 * It returns the next address to be unloaded. 5896 * It should be called with the hash lock held. 5897 */ 5898 static caddr_t 5899 sfmmu_hblk_unload(struct hat *sfmmup, struct hme_blk *hmeblkp, caddr_t addr, 5900 caddr_t endaddr, demap_range_t *dmrp, uint_t flags) 5901 { 5902 tte_t tte, ttemod; 5903 struct sf_hment *sfhmep; 5904 int ttesz; 5905 long ttecnt; 5906 page_t *pp; 5907 kmutex_t *pml; 5908 int ret; 5909 int use_demap_range; 5910 5911 ASSERT(in_hblk_range(hmeblkp, addr)); 5912 ASSERT(!hmeblkp->hblk_shw_bit); 5913 ASSERT(sfmmup != NULL || hmeblkp->hblk_shared); 5914 ASSERT(sfmmup == NULL || !hmeblkp->hblk_shared); 5915 ASSERT(dmrp == NULL || !hmeblkp->hblk_shared); 5916 5917 #ifdef DEBUG 5918 if (get_hblk_ttesz(hmeblkp) != TTE8K && 5919 (endaddr < get_hblk_endaddr(hmeblkp))) { 5920 panic("sfmmu_hblk_unload: partial unload of large page"); 5921 } 5922 #endif /* DEBUG */ 5923 5924 endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp)); 5925 ttesz = get_hblk_ttesz(hmeblkp); 5926 5927 use_demap_range = ((dmrp == NULL) || 5928 (TTEBYTES(ttesz) == DEMAP_RANGE_PGSZ(dmrp))); 5929 5930 if (use_demap_range) { 5931 DEMAP_RANGE_CONTINUE(dmrp, addr, endaddr); 5932 } else { 5933 DEMAP_RANGE_FLUSH(dmrp); 5934 } 5935 ttecnt = 0; 5936 HBLKTOHME(sfhmep, hmeblkp, addr); 5937 5938 while (addr < endaddr) { 5939 pml = NULL; 5940 sfmmu_copytte(&sfhmep->hme_tte, &tte); 5941 if (TTE_IS_VALID(&tte)) { 5942 pp = sfhmep->hme_page; 5943 if (pp != NULL) { 5944 pml = sfmmu_mlist_enter(pp); 5945 } 5946 5947 /* 5948 * Verify if hme still points to 'pp' now that 5949 * we have p_mapping lock. 5950 */ 5951 if (sfhmep->hme_page != pp) { 5952 if (pp != NULL && sfhmep->hme_page != NULL) { 5953 ASSERT(pml != NULL); 5954 sfmmu_mlist_exit(pml); 5955 /* Re-start this iteration. */ 5956 continue; 5957 } 5958 ASSERT((pp != NULL) && 5959 (sfhmep->hme_page == NULL)); 5960 goto tte_unloaded; 5961 } 5962 5963 /* 5964 * This point on we have both HASH and p_mapping 5965 * lock. 5966 */ 5967 ASSERT(pp == sfhmep->hme_page); 5968 ASSERT(pp == NULL || sfmmu_mlist_held(pp)); 5969 5970 /* 5971 * We need to loop on modify tte because it is 5972 * possible for pagesync to come along and 5973 * change the software bits beneath us. 5974 * 5975 * Page_unload can also invalidate the tte after 5976 * we read tte outside of p_mapping lock. 5977 */ 5978 again: 5979 ttemod = tte; 5980 5981 TTE_SET_INVALID(&ttemod); 5982 ret = sfmmu_modifytte_try(&tte, &ttemod, 5983 &sfhmep->hme_tte); 5984 5985 if (ret <= 0) { 5986 if (TTE_IS_VALID(&tte)) { 5987 ASSERT(ret < 0); 5988 goto again; 5989 } 5990 if (pp != NULL) { 5991 panic("sfmmu_hblk_unload: pp = 0x%p " 5992 "tte became invalid under mlist" 5993 " lock = 0x%p", (void *)pp, 5994 (void *)pml); 5995 } 5996 continue; 5997 } 5998 5999 if (!(flags & HAT_UNLOAD_NOSYNC)) { 6000 sfmmu_ttesync(sfmmup, addr, &tte, pp); 6001 } 6002 6003 /* 6004 * Ok- we invalidated the tte. Do the rest of the job. 6005 */ 6006 ttecnt++; 6007 6008 if (flags & HAT_UNLOAD_UNLOCK) { 6009 ASSERT(hmeblkp->hblk_lckcnt > 0); 6010 atomic_add_32(&hmeblkp->hblk_lckcnt, -1); 6011 HBLK_STACK_TRACE(hmeblkp, HBLK_UNLOCK); 6012 } 6013 6014 /* 6015 * Normally we would need to flush the page 6016 * from the virtual cache at this point in 6017 * order to prevent a potential cache alias 6018 * inconsistency. 6019 * The particular scenario we need to worry 6020 * about is: 6021 * Given: va1 and va2 are two virtual address 6022 * that alias and map the same physical 6023 * address. 6024 * 1. mapping exists from va1 to pa and data 6025 * has been read into the cache. 6026 * 2. unload va1. 6027 * 3. load va2 and modify data using va2. 6028 * 4 unload va2. 6029 * 5. load va1 and reference data. Unless we 6030 * flush the data cache when we unload we will 6031 * get stale data. 6032 * Fortunately, page coloring eliminates the 6033 * above scenario by remembering the color a 6034 * physical page was last or is currently 6035 * mapped to. Now, we delay the flush until 6036 * the loading of translations. Only when the 6037 * new translation is of a different color 6038 * are we forced to flush. 6039 */ 6040 if (use_demap_range) { 6041 /* 6042 * Mark this page as needing a demap. 6043 */ 6044 DEMAP_RANGE_MARKPG(dmrp, addr); 6045 } else { 6046 ASSERT(sfmmup != NULL); 6047 ASSERT(!hmeblkp->hblk_shared); 6048 sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 6049 sfmmup->sfmmu_free, 0); 6050 } 6051 6052 if (pp) { 6053 /* 6054 * Remove the hment from the mapping list 6055 */ 6056 ASSERT(hmeblkp->hblk_hmecnt > 0); 6057 6058 /* 6059 * Again, we cannot 6060 * ASSERT(hmeblkp->hblk_hmecnt <= NHMENTS); 6061 */ 6062 HME_SUB(sfhmep, pp); 6063 membar_stst(); 6064 atomic_add_16(&hmeblkp->hblk_hmecnt, -1); 6065 } 6066 6067 ASSERT(hmeblkp->hblk_vcnt > 0); 6068 atomic_add_16(&hmeblkp->hblk_vcnt, -1); 6069 6070 ASSERT(hmeblkp->hblk_hmecnt || hmeblkp->hblk_vcnt || 6071 !hmeblkp->hblk_lckcnt); 6072 6073 #ifdef VAC 6074 if (pp && (pp->p_nrm & (P_KPMC | P_KPMS | P_TNC))) { 6075 if (PP_ISTNC(pp)) { 6076 /* 6077 * If page was temporary 6078 * uncached, try to recache 6079 * it. Note that HME_SUB() was 6080 * called above so p_index and 6081 * mlist had been updated. 6082 */ 6083 conv_tnc(pp, ttesz); 6084 } else if (pp->p_mapping == NULL) { 6085 ASSERT(kpm_enable); 6086 /* 6087 * Page is marked to be in VAC conflict 6088 * to an existing kpm mapping and/or is 6089 * kpm mapped using only the regular 6090 * pagesize. 6091 */ 6092 sfmmu_kpm_hme_unload(pp); 6093 } 6094 } 6095 #endif /* VAC */ 6096 } else if ((pp = sfhmep->hme_page) != NULL) { 6097 /* 6098 * TTE is invalid but the hme 6099 * still exists. let pageunload 6100 * complete its job. 6101 */ 6102 ASSERT(pml == NULL); 6103 pml = sfmmu_mlist_enter(pp); 6104 if (sfhmep->hme_page != NULL) { 6105 sfmmu_mlist_exit(pml); 6106 continue; 6107 } 6108 ASSERT(sfhmep->hme_page == NULL); 6109 } else if (hmeblkp->hblk_hmecnt != 0) { 6110 /* 6111 * pageunload may have not finished decrementing 6112 * hblk_vcnt and hblk_hmecnt. Find page_t if any and 6113 * wait for pageunload to finish. Rely on pageunload 6114 * to decrement hblk_hmecnt after hblk_vcnt. 6115 */ 6116 pfn_t pfn = TTE_TO_TTEPFN(&tte); 6117 ASSERT(pml == NULL); 6118 if (pf_is_memory(pfn)) { 6119 pp = page_numtopp_nolock(pfn); 6120 if (pp != NULL) { 6121 pml = sfmmu_mlist_enter(pp); 6122 sfmmu_mlist_exit(pml); 6123 pml = NULL; 6124 } 6125 } 6126 } 6127 6128 tte_unloaded: 6129 /* 6130 * At this point, the tte we are looking at 6131 * should be unloaded, and hme has been unlinked 6132 * from page too. This is important because in 6133 * pageunload, it does ttesync() then HME_SUB. 6134 * We need to make sure HME_SUB has been completed 6135 * so we know ttesync() has been completed. Otherwise, 6136 * at exit time, after return from hat layer, VM will 6137 * release as structure which hat_setstat() (called 6138 * by ttesync()) needs. 6139 */ 6140 #ifdef DEBUG 6141 { 6142 tte_t dtte; 6143 6144 ASSERT(sfhmep->hme_page == NULL); 6145 6146 sfmmu_copytte(&sfhmep->hme_tte, &dtte); 6147 ASSERT(!TTE_IS_VALID(&dtte)); 6148 } 6149 #endif 6150 6151 if (pml) { 6152 sfmmu_mlist_exit(pml); 6153 } 6154 6155 addr += TTEBYTES(ttesz); 6156 sfhmep++; 6157 DEMAP_RANGE_NEXTPG(dmrp); 6158 } 6159 /* 6160 * For shared hmeblks this routine is only called when region is freed 6161 * and no longer referenced. So no need to decrement ttecnt 6162 * in the region structure here. 6163 */ 6164 if (ttecnt > 0 && sfmmup != NULL) { 6165 atomic_add_long(&sfmmup->sfmmu_ttecnt[ttesz], -ttecnt); 6166 } 6167 return (addr); 6168 } 6169 6170 /* 6171 * Synchronize all the mappings in the range [addr..addr+len). 6172 * Can be called with clearflag having two states: 6173 * HAT_SYNC_DONTZERO means just return the rm stats 6174 * HAT_SYNC_ZERORM means zero rm bits in the tte and return the stats 6175 */ 6176 void 6177 hat_sync(struct hat *sfmmup, caddr_t addr, size_t len, uint_t clearflag) 6178 { 6179 struct hmehash_bucket *hmebp; 6180 hmeblk_tag hblktag; 6181 int hmeshift, hashno = 1; 6182 struct hme_blk *hmeblkp, *list = NULL; 6183 caddr_t endaddr; 6184 cpuset_t cpuset; 6185 6186 ASSERT(sfmmup->sfmmu_xhat_provider == NULL); 6187 ASSERT((sfmmup == ksfmmup) || 6188 AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock)); 6189 ASSERT((len & MMU_PAGEOFFSET) == 0); 6190 ASSERT((clearflag == HAT_SYNC_DONTZERO) || 6191 (clearflag == HAT_SYNC_ZERORM)); 6192 6193 CPUSET_ZERO(cpuset); 6194 6195 endaddr = addr + len; 6196 hblktag.htag_id = sfmmup; 6197 hblktag.htag_rid = SFMMU_INVALID_SHMERID; 6198 6199 /* 6200 * Spitfire supports 4 page sizes. 6201 * Most pages are expected to be of the smallest page 6202 * size (8K) and these will not need to be rehashed. 64K 6203 * pages also don't need to be rehashed because the an hmeblk 6204 * spans 64K of address space. 512K pages might need 1 rehash and 6205 * and 4M pages 2 rehashes. 6206 */ 6207 while (addr < endaddr) { 6208 hmeshift = HME_HASH_SHIFT(hashno); 6209 hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift); 6210 hblktag.htag_rehash = hashno; 6211 hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift); 6212 6213 SFMMU_HASH_LOCK(hmebp); 6214 6215 HME_HASH_SEARCH(hmebp, hblktag, hmeblkp, &list); 6216 if (hmeblkp != NULL) { 6217 ASSERT(!hmeblkp->hblk_shared); 6218 /* 6219 * We've encountered a shadow hmeblk so skip the range 6220 * of the next smaller mapping size. 6221 */ 6222 if (hmeblkp->hblk_shw_bit) { 6223 ASSERT(sfmmup != ksfmmup); 6224 ASSERT(hashno > 1); 6225 addr = (caddr_t)P2END((uintptr_t)addr, 6226 TTEBYTES(hashno - 1)); 6227 } else { 6228 addr = sfmmu_hblk_sync(sfmmup, hmeblkp, 6229 addr, endaddr, clearflag); 6230 } 6231 SFMMU_HASH_UNLOCK(hmebp); 6232 hashno = 1; 6233 continue; 6234 } 6235 SFMMU_HASH_UNLOCK(hmebp); 6236 6237 if (!HME_REHASH(sfmmup) || (hashno >= mmu_hashcnt)) { 6238 /* 6239 * We have traversed the whole list and rehashed 6240 * if necessary without finding the address to sync. 6241 * This is ok so we increment the address by the 6242 * smallest hmeblk range for kernel mappings and the 6243 * largest hmeblk range, to account for shadow hmeblks, 6244 * for user mappings and continue. 6245 */ 6246 if (sfmmup == ksfmmup) 6247 addr = (caddr_t)P2END((uintptr_t)addr, 6248 TTEBYTES(1)); 6249 else 6250 addr = (caddr_t)P2END((uintptr_t)addr, 6251 TTEBYTES(hashno)); 6252 hashno = 1; 6253 } else { 6254 hashno++; 6255 } 6256 } 6257 sfmmu_hblks_list_purge(&list, 0); 6258 cpuset = sfmmup->sfmmu_cpusran; 6259 xt_sync(cpuset); 6260 } 6261 6262 static caddr_t 6263 sfmmu_hblk_sync(struct hat *sfmmup, struct hme_blk *hmeblkp, caddr_t addr, 6264 caddr_t endaddr, int clearflag) 6265 { 6266 tte_t tte, ttemod; 6267 struct sf_hment *sfhmep; 6268 int ttesz; 6269 struct page *pp; 6270 kmutex_t *pml; 6271 int ret; 6272 6273 ASSERT(hmeblkp->hblk_shw_bit == 0); 6274 ASSERT(!hmeblkp->hblk_shared); 6275 6276 endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp)); 6277 6278 ttesz = get_hblk_ttesz(hmeblkp); 6279 HBLKTOHME(sfhmep, hmeblkp, addr); 6280 6281 while (addr < endaddr) { 6282 sfmmu_copytte(&sfhmep->hme_tte, &tte); 6283 if (TTE_IS_VALID(&tte)) { 6284 pml = NULL; 6285 pp = sfhmep->hme_page; 6286 if (pp) { 6287 pml = sfmmu_mlist_enter(pp); 6288 } 6289 if (pp != sfhmep->hme_page) { 6290 /* 6291 * tte most have been unloaded 6292 * underneath us. Recheck 6293 */ 6294 ASSERT(pml); 6295 sfmmu_mlist_exit(pml); 6296 continue; 6297 } 6298 6299 ASSERT(pp == NULL || sfmmu_mlist_held(pp)); 6300 6301 if (clearflag == HAT_SYNC_ZERORM) { 6302 ttemod = tte; 6303 TTE_CLR_RM(&ttemod); 6304 ret = sfmmu_modifytte_try(&tte, &ttemod, 6305 &sfhmep->hme_tte); 6306 if (ret < 0) { 6307 if (pml) { 6308 sfmmu_mlist_exit(pml); 6309 } 6310 continue; 6311 } 6312 6313 if (ret > 0) { 6314 sfmmu_tlb_demap(addr, sfmmup, 6315 hmeblkp, 0, 0); 6316 } 6317 } 6318 sfmmu_ttesync(sfmmup, addr, &tte, pp); 6319 if (pml) { 6320 sfmmu_mlist_exit(pml); 6321 } 6322 } 6323 addr += TTEBYTES(ttesz); 6324 sfhmep++; 6325 } 6326 return (addr); 6327 } 6328 6329 /* 6330 * This function will sync a tte to the page struct and it will 6331 * update the hat stats. Currently it allows us to pass a NULL pp 6332 * and we will simply update the stats. We may want to change this 6333 * so we only keep stats for pages backed by pp's. 6334 */ 6335 static void 6336 sfmmu_ttesync(struct hat *sfmmup, caddr_t addr, tte_t *ttep, page_t *pp) 6337 { 6338 uint_t rm = 0; 6339 int sz; 6340 pgcnt_t npgs; 6341 6342 ASSERT(TTE_IS_VALID(ttep)); 6343 6344 if (TTE_IS_NOSYNC(ttep)) { 6345 return; 6346 } 6347 6348 if (TTE_IS_REF(ttep)) { 6349 rm = P_REF; 6350 } 6351 if (TTE_IS_MOD(ttep)) { 6352 rm |= P_MOD; 6353 } 6354 6355 if (rm == 0) { 6356 return; 6357 } 6358 6359 sz = TTE_CSZ(ttep); 6360 if (sfmmup != NULL && sfmmup->sfmmu_rmstat) { 6361 int i; 6362 caddr_t vaddr = addr; 6363 6364 for (i = 0; i < TTEPAGES(sz); i++, vaddr += MMU_PAGESIZE) { 6365 hat_setstat(sfmmup->sfmmu_as, vaddr, MMU_PAGESIZE, rm); 6366 } 6367 6368 } 6369 6370 /* 6371 * XXX I want to use cas to update nrm bits but they 6372 * currently belong in common/vm and not in hat where 6373 * they should be. 6374 * The nrm bits are protected by the same mutex as 6375 * the one that protects the page's mapping list. 6376 */ 6377 if (!pp) 6378 return; 6379 ASSERT(sfmmu_mlist_held(pp)); 6380 /* 6381 * If the tte is for a large page, we need to sync all the 6382 * pages covered by the tte. 6383 */ 6384 if (sz != TTE8K) { 6385 ASSERT(pp->p_szc != 0); 6386 pp = PP_GROUPLEADER(pp, sz); 6387 ASSERT(sfmmu_mlist_held(pp)); 6388 } 6389 6390 /* Get number of pages from tte size. */ 6391 npgs = TTEPAGES(sz); 6392 6393 do { 6394 ASSERT(pp); 6395 ASSERT(sfmmu_mlist_held(pp)); 6396 if (((rm & P_REF) != 0 && !PP_ISREF(pp)) || 6397 ((rm & P_MOD) != 0 && !PP_ISMOD(pp))) 6398 hat_page_setattr(pp, rm); 6399 6400 /* 6401 * Are we done? If not, we must have a large mapping. 6402 * For large mappings we need to sync the rest of the pages 6403 * covered by this tte; goto the next page. 6404 */ 6405 } while (--npgs > 0 && (pp = PP_PAGENEXT(pp))); 6406 } 6407 6408 /* 6409 * Execute pre-callback handler of each pa_hment linked to pp 6410 * 6411 * Inputs: 6412 * flag: either HAT_PRESUSPEND or HAT_SUSPEND. 6413 * capture_cpus: pointer to return value (below) 6414 * 6415 * Returns: 6416 * Propagates the subsystem callback return values back to the caller; 6417 * returns 0 on success. If capture_cpus is non-NULL, the value returned 6418 * is zero if all of the pa_hments are of a type that do not require 6419 * capturing CPUs prior to suspending the mapping, else it is 1. 6420 */ 6421 static int 6422 hat_pageprocess_precallbacks(struct page *pp, uint_t flag, int *capture_cpus) 6423 { 6424 struct sf_hment *sfhmep; 6425 struct pa_hment *pahmep; 6426 int (*f)(caddr_t, uint_t, uint_t, void *); 6427 int ret; 6428 id_t id; 6429 int locked = 0; 6430 kmutex_t *pml; 6431 6432 ASSERT(PAGE_EXCL(pp)); 6433 if (!sfmmu_mlist_held(pp)) { 6434 pml = sfmmu_mlist_enter(pp); 6435 locked = 1; 6436 } 6437 6438 if (capture_cpus) 6439 *capture_cpus = 0; 6440 6441 top: 6442 for (sfhmep = pp->p_mapping; sfhmep; sfhmep = sfhmep->hme_next) { 6443 /* 6444 * skip sf_hments corresponding to VA<->PA mappings; 6445 * for pa_hment's, hme_tte.ll is zero 6446 */ 6447 if (!IS_PAHME(sfhmep)) 6448 continue; 6449 6450 pahmep = sfhmep->hme_data; 6451 ASSERT(pahmep != NULL); 6452 6453 /* 6454 * skip if pre-handler has been called earlier in this loop 6455 */ 6456 if (pahmep->flags & flag) 6457 continue; 6458 6459 id = pahmep->cb_id; 6460 ASSERT(id >= (id_t)0 && id < sfmmu_cb_nextid); 6461 if (capture_cpus && sfmmu_cb_table[id].capture_cpus != 0) 6462 *capture_cpus = 1; 6463 if ((f = sfmmu_cb_table[id].prehandler) == NULL) { 6464 pahmep->flags |= flag; 6465 continue; 6466 } 6467 6468 /* 6469 * Drop the mapping list lock to avoid locking order issues. 6470 */ 6471 if (locked) 6472 sfmmu_mlist_exit(pml); 6473 6474 ret = f(pahmep->addr, pahmep->len, flag, pahmep->pvt); 6475 if (ret != 0) 6476 return (ret); /* caller must do the cleanup */ 6477 6478 if (locked) { 6479 pml = sfmmu_mlist_enter(pp); 6480 pahmep->flags |= flag; 6481 goto top; 6482 } 6483 6484 pahmep->flags |= flag; 6485 } 6486 6487 if (locked) 6488 sfmmu_mlist_exit(pml); 6489 6490 return (0); 6491 } 6492 6493 /* 6494 * Execute post-callback handler of each pa_hment linked to pp 6495 * 6496 * Same overall assumptions and restrictions apply as for 6497 * hat_pageprocess_precallbacks(). 6498 */ 6499 static void 6500 hat_pageprocess_postcallbacks(struct page *pp, uint_t flag) 6501 { 6502 pfn_t pgpfn = pp->p_pagenum; 6503 pfn_t pgmask = btop(page_get_pagesize(pp->p_szc)) - 1; 6504 pfn_t newpfn; 6505 struct sf_hment *sfhmep; 6506 struct pa_hment *pahmep; 6507 int (*f)(caddr_t, uint_t, uint_t, void *, pfn_t); 6508 id_t id; 6509 int locked = 0; 6510 kmutex_t *pml; 6511 6512 ASSERT(PAGE_EXCL(pp)); 6513 if (!sfmmu_mlist_held(pp)) { 6514 pml = sfmmu_mlist_enter(pp); 6515 locked = 1; 6516 } 6517 6518 top: 6519 for (sfhmep = pp->p_mapping; sfhmep; sfhmep = sfhmep->hme_next) { 6520 /* 6521 * skip sf_hments corresponding to VA<->PA mappings; 6522 * for pa_hment's, hme_tte.ll is zero 6523 */ 6524 if (!IS_PAHME(sfhmep)) 6525 continue; 6526 6527 pahmep = sfhmep->hme_data; 6528 ASSERT(pahmep != NULL); 6529 6530 if ((pahmep->flags & flag) == 0) 6531 continue; 6532 6533 pahmep->flags &= ~flag; 6534 6535 id = pahmep->cb_id; 6536 ASSERT(id >= (id_t)0 && id < sfmmu_cb_nextid); 6537 if ((f = sfmmu_cb_table[id].posthandler) == NULL) 6538 continue; 6539 6540 /* 6541 * Convert the base page PFN into the constituent PFN 6542 * which is needed by the callback handler. 6543 */ 6544 newpfn = pgpfn | (btop((uintptr_t)pahmep->addr) & pgmask); 6545 6546 /* 6547 * Drop the mapping list lock to avoid locking order issues. 6548 */ 6549 if (locked) 6550 sfmmu_mlist_exit(pml); 6551 6552 if (f(pahmep->addr, pahmep->len, flag, pahmep->pvt, newpfn) 6553 != 0) 6554 panic("sfmmu: posthandler failed"); 6555 6556 if (locked) { 6557 pml = sfmmu_mlist_enter(pp); 6558 goto top; 6559 } 6560 } 6561 6562 if (locked) 6563 sfmmu_mlist_exit(pml); 6564 } 6565 6566 /* 6567 * Suspend locked kernel mapping 6568 */ 6569 void 6570 hat_pagesuspend(struct page *pp) 6571 { 6572 struct sf_hment *sfhmep; 6573 sfmmu_t *sfmmup; 6574 tte_t tte, ttemod; 6575 struct hme_blk *hmeblkp; 6576 caddr_t addr; 6577 int index, cons; 6578 cpuset_t cpuset; 6579 6580 ASSERT(PAGE_EXCL(pp)); 6581 ASSERT(sfmmu_mlist_held(pp)); 6582 6583 mutex_enter(&kpr_suspendlock); 6584 6585 /* 6586 * We're about to suspend a kernel mapping so mark this thread as 6587 * non-traceable by DTrace. This prevents us from running into issues 6588 * with probe context trying to touch a suspended page 6589 * in the relocation codepath itself. 6590 */ 6591 curthread->t_flag |= T_DONTDTRACE; 6592 6593 index = PP_MAPINDEX(pp); 6594 cons = TTE8K; 6595 6596 retry: 6597 for (sfhmep = pp->p_mapping; sfhmep; sfhmep = sfhmep->hme_next) { 6598 6599 if (IS_PAHME(sfhmep)) 6600 continue; 6601 6602 if (get_hblk_ttesz(sfmmu_hmetohblk(sfhmep)) != cons) 6603 continue; 6604 6605 /* 6606 * Loop until we successfully set the suspend bit in 6607 * the TTE. 6608 */ 6609 again: 6610 sfmmu_copytte(&sfhmep->hme_tte, &tte); 6611 ASSERT(TTE_IS_VALID(&tte)); 6612 6613 ttemod = tte; 6614 TTE_SET_SUSPEND(&ttemod); 6615 if (sfmmu_modifytte_try(&tte, &ttemod, 6616 &sfhmep->hme_tte) < 0) 6617 goto again; 6618 6619 /* 6620 * Invalidate TSB entry 6621 */ 6622 hmeblkp = sfmmu_hmetohblk(sfhmep); 6623 6624 sfmmup = hblktosfmmu(hmeblkp); 6625 ASSERT(sfmmup == ksfmmup); 6626 ASSERT(!hmeblkp->hblk_shared); 6627 6628 addr = tte_to_vaddr(hmeblkp, tte); 6629 6630 /* 6631 * No need to make sure that the TSB for this sfmmu is 6632 * not being relocated since it is ksfmmup and thus it 6633 * will never be relocated. 6634 */ 6635 SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp, 0); 6636 6637 /* 6638 * Update xcall stats 6639 */ 6640 cpuset = cpu_ready_set; 6641 CPUSET_DEL(cpuset, CPU->cpu_id); 6642 6643 /* LINTED: constant in conditional context */ 6644 SFMMU_XCALL_STATS(ksfmmup); 6645 6646 /* 6647 * Flush TLB entry on remote CPU's 6648 */ 6649 xt_some(cpuset, vtag_flushpage_tl1, (uint64_t)addr, 6650 (uint64_t)ksfmmup); 6651 xt_sync(cpuset); 6652 6653 /* 6654 * Flush TLB entry on local CPU 6655 */ 6656 vtag_flushpage(addr, (uint64_t)ksfmmup); 6657 } 6658 6659 while (index != 0) { 6660 index = index >> 1; 6661 if (index != 0) 6662 cons++; 6663 if (index & 0x1) { 6664 pp = PP_GROUPLEADER(pp, cons); 6665 goto retry; 6666 } 6667 } 6668 } 6669 6670 #ifdef DEBUG 6671 6672 #define N_PRLE 1024 6673 struct prle { 6674 page_t *targ; 6675 page_t *repl; 6676 int status; 6677 int pausecpus; 6678 hrtime_t whence; 6679 }; 6680 6681 static struct prle page_relocate_log[N_PRLE]; 6682 static int prl_entry; 6683 static kmutex_t prl_mutex; 6684 6685 #define PAGE_RELOCATE_LOG(t, r, s, p) \ 6686 mutex_enter(&prl_mutex); \ 6687 page_relocate_log[prl_entry].targ = *(t); \ 6688 page_relocate_log[prl_entry].repl = *(r); \ 6689 page_relocate_log[prl_entry].status = (s); \ 6690 page_relocate_log[prl_entry].pausecpus = (p); \ 6691 page_relocate_log[prl_entry].whence = gethrtime(); \ 6692 prl_entry = (prl_entry == (N_PRLE - 1))? 0 : prl_entry + 1; \ 6693 mutex_exit(&prl_mutex); 6694 6695 #else /* !DEBUG */ 6696 #define PAGE_RELOCATE_LOG(t, r, s, p) 6697 #endif 6698 6699 /* 6700 * Core Kernel Page Relocation Algorithm 6701 * 6702 * Input: 6703 * 6704 * target : constituent pages are SE_EXCL locked. 6705 * replacement: constituent pages are SE_EXCL locked. 6706 * 6707 * Output: 6708 * 6709 * nrelocp: number of pages relocated 6710 */ 6711 int 6712 hat_page_relocate(page_t **target, page_t **replacement, spgcnt_t *nrelocp) 6713 { 6714 page_t *targ, *repl; 6715 page_t *tpp, *rpp; 6716 kmutex_t *low, *high; 6717 spgcnt_t npages, i; 6718 page_t *pl = NULL; 6719 int old_pil; 6720 cpuset_t cpuset; 6721 int cap_cpus; 6722 int ret; 6723 #ifdef VAC 6724 int cflags = 0; 6725 #endif 6726 6727 if (hat_kpr_enabled == 0 || !kcage_on || PP_ISNORELOC(*target)) { 6728 PAGE_RELOCATE_LOG(target, replacement, EAGAIN, -1); 6729 return (EAGAIN); 6730 } 6731 6732 mutex_enter(&kpr_mutex); 6733 kreloc_thread = curthread; 6734 6735 targ = *target; 6736 repl = *replacement; 6737 ASSERT(repl != NULL); 6738 ASSERT(targ->p_szc == repl->p_szc); 6739 6740 npages = page_get_pagecnt(targ->p_szc); 6741 6742 /* 6743 * unload VA<->PA mappings that are not locked 6744 */ 6745 tpp = targ; 6746 for (i = 0; i < npages; i++) { 6747 (void) hat_pageunload(tpp, SFMMU_KERNEL_RELOC); 6748 tpp++; 6749 } 6750 6751 /* 6752 * Do "presuspend" callbacks, in a context from which we can still 6753 * block as needed. Note that we don't hold the mapping list lock 6754 * of "targ" at this point due to potential locking order issues; 6755 * we assume that between the hat_pageunload() above and holding 6756 * the SE_EXCL lock that the mapping list *cannot* change at this 6757 * point. 6758 */ 6759 ret = hat_pageprocess_precallbacks(targ, HAT_PRESUSPEND, &cap_cpus); 6760 if (ret != 0) { 6761 /* 6762 * EIO translates to fatal error, for all others cleanup 6763 * and return EAGAIN. 6764 */ 6765 ASSERT(ret != EIO); 6766 hat_pageprocess_postcallbacks(targ, HAT_POSTUNSUSPEND); 6767 PAGE_RELOCATE_LOG(target, replacement, ret, -1); 6768 kreloc_thread = NULL; 6769 mutex_exit(&kpr_mutex); 6770 return (EAGAIN); 6771 } 6772 6773 /* 6774 * acquire p_mapping list lock for both the target and replacement 6775 * root pages. 6776 * 6777 * low and high refer to the need to grab the mlist locks in a 6778 * specific order in order to prevent race conditions. Thus the 6779 * lower lock must be grabbed before the higher lock. 6780 * 6781 * This will block hat_unload's accessing p_mapping list. Since 6782 * we have SE_EXCL lock, hat_memload and hat_pageunload will be 6783 * blocked. Thus, no one else will be accessing the p_mapping list 6784 * while we suspend and reload the locked mapping below. 6785 */ 6786 tpp = targ; 6787 rpp = repl; 6788 sfmmu_mlist_reloc_enter(tpp, rpp, &low, &high); 6789 6790 kpreempt_disable(); 6791 6792 /* 6793 * We raise our PIL to 13 so that we don't get captured by 6794 * another CPU or pinned by an interrupt thread. We can't go to 6795 * PIL 14 since the nexus driver(s) may need to interrupt at 6796 * that level in the case of IOMMU pseudo mappings. 6797 */ 6798 cpuset = cpu_ready_set; 6799 CPUSET_DEL(cpuset, CPU->cpu_id); 6800 if (!cap_cpus || CPUSET_ISNULL(cpuset)) { 6801 old_pil = splr(XCALL_PIL); 6802 } else { 6803 old_pil = -1; 6804 xc_attention(cpuset); 6805 } 6806 ASSERT(getpil() == XCALL_PIL); 6807 6808 /* 6809 * Now do suspend callbacks. In the case of an IOMMU mapping 6810 * this will suspend all DMA activity to the page while it is 6811 * being relocated. Since we are well above LOCK_LEVEL and CPUs 6812 * may be captured at this point we should have acquired any needed 6813 * locks in the presuspend callback. 6814 */ 6815 ret = hat_pageprocess_precallbacks(targ, HAT_SUSPEND, NULL); 6816 if (ret != 0) { 6817 repl = targ; 6818 goto suspend_fail; 6819 } 6820 6821 /* 6822 * Raise the PIL yet again, this time to block all high-level 6823 * interrupts on this CPU. This is necessary to prevent an 6824 * interrupt routine from pinning the thread which holds the 6825 * mapping suspended and then touching the suspended page. 6826 * 6827 * Once the page is suspended we also need to be careful to 6828 * avoid calling any functions which touch any seg_kmem memory 6829 * since that memory may be backed by the very page we are 6830 * relocating in here! 6831 */ 6832 hat_pagesuspend(targ); 6833 6834 /* 6835 * Now that we are confident everybody has stopped using this page, 6836 * copy the page contents. Note we use a physical copy to prevent 6837 * locking issues and to avoid fpRAS because we can't handle it in 6838 * this context. 6839 */ 6840 for (i = 0; i < npages; i++, tpp++, rpp++) { 6841 #ifdef VAC 6842 /* 6843 * If the replacement has a different vcolor than 6844 * the one being replacd, we need to handle VAC 6845 * consistency for it just as we were setting up 6846 * a new mapping to it. 6847 */ 6848 if ((PP_GET_VCOLOR(rpp) != NO_VCOLOR) && 6849 (tpp->p_vcolor != rpp->p_vcolor) && 6850 !CacheColor_IsFlushed(cflags, PP_GET_VCOLOR(rpp))) { 6851 CacheColor_SetFlushed(cflags, PP_GET_VCOLOR(rpp)); 6852 sfmmu_cache_flushcolor(PP_GET_VCOLOR(rpp), 6853 rpp->p_pagenum); 6854 } 6855 #endif 6856 /* 6857 * Copy the contents of the page. 6858 */ 6859 ppcopy_kernel(tpp, rpp); 6860 } 6861 6862 tpp = targ; 6863 rpp = repl; 6864 for (i = 0; i < npages; i++, tpp++, rpp++) { 6865 /* 6866 * Copy attributes. VAC consistency was handled above, 6867 * if required. 6868 */ 6869 rpp->p_nrm = tpp->p_nrm; 6870 tpp->p_nrm = 0; 6871 rpp->p_index = tpp->p_index; 6872 tpp->p_index = 0; 6873 #ifdef VAC 6874 rpp->p_vcolor = tpp->p_vcolor; 6875 #endif 6876 } 6877 6878 /* 6879 * First, unsuspend the page, if we set the suspend bit, and transfer 6880 * the mapping list from the target page to the replacement page. 6881 * Next process postcallbacks; since pa_hment's are linked only to the 6882 * p_mapping list of root page, we don't iterate over the constituent 6883 * pages. 6884 */ 6885 hat_pagereload(targ, repl); 6886 6887 suspend_fail: 6888 hat_pageprocess_postcallbacks(repl, HAT_UNSUSPEND); 6889 6890 /* 6891 * Now lower our PIL and release any captured CPUs since we 6892 * are out of the "danger zone". After this it will again be 6893 * safe to acquire adaptive mutex locks, or to drop them... 6894 */ 6895 if (old_pil != -1) { 6896 splx(old_pil); 6897 } else { 6898 xc_dismissed(cpuset); 6899 } 6900 6901 kpreempt_enable(); 6902 6903 sfmmu_mlist_reloc_exit(low, high); 6904 6905 /* 6906 * Postsuspend callbacks should drop any locks held across 6907 * the suspend callbacks. As before, we don't hold the mapping 6908 * list lock at this point.. our assumption is that the mapping 6909 * list still can't change due to our holding SE_EXCL lock and 6910 * there being no unlocked mappings left. Hence the restriction 6911 * on calling context to hat_delete_callback() 6912 */ 6913 hat_pageprocess_postcallbacks(repl, HAT_POSTUNSUSPEND); 6914 if (ret != 0) { 6915 /* 6916 * The second presuspend call failed: we got here through 6917 * the suspend_fail label above. 6918 */ 6919 ASSERT(ret != EIO); 6920 PAGE_RELOCATE_LOG(target, replacement, ret, cap_cpus); 6921 kreloc_thread = NULL; 6922 mutex_exit(&kpr_mutex); 6923 return (EAGAIN); 6924 } 6925 6926 /* 6927 * Now that we're out of the performance critical section we can 6928 * take care of updating the hash table, since we still 6929 * hold all the pages locked SE_EXCL at this point we 6930 * needn't worry about things changing out from under us. 6931 */ 6932 tpp = targ; 6933 rpp = repl; 6934 for (i = 0; i < npages; i++, tpp++, rpp++) { 6935 6936 /* 6937 * replace targ with replacement in page_hash table 6938 */ 6939 targ = tpp; 6940 page_relocate_hash(rpp, targ); 6941 6942 /* 6943 * concatenate target; caller of platform_page_relocate() 6944 * expects target to be concatenated after returning. 6945 */ 6946 ASSERT(targ->p_next == targ); 6947 ASSERT(targ->p_prev == targ); 6948 page_list_concat(&pl, &targ); 6949 } 6950 6951 ASSERT(*target == pl); 6952 *nrelocp = npages; 6953 PAGE_RELOCATE_LOG(target, replacement, 0, cap_cpus); 6954 kreloc_thread = NULL; 6955 mutex_exit(&kpr_mutex); 6956 return (0); 6957 } 6958 6959 /* 6960 * Called when stray pa_hments are found attached to a page which is 6961 * being freed. Notify the subsystem which attached the pa_hment of 6962 * the error if it registered a suitable handler, else panic. 6963 */ 6964 static void 6965 sfmmu_pahment_leaked(struct pa_hment *pahmep) 6966 { 6967 id_t cb_id = pahmep->cb_id; 6968 6969 ASSERT(cb_id >= (id_t)0 && cb_id < sfmmu_cb_nextid); 6970 if (sfmmu_cb_table[cb_id].errhandler != NULL) { 6971 if (sfmmu_cb_table[cb_id].errhandler(pahmep->addr, pahmep->len, 6972 HAT_CB_ERR_LEAKED, pahmep->pvt) == 0) 6973 return; /* non-fatal */ 6974 } 6975 panic("pa_hment leaked: 0x%p", (void *)pahmep); 6976 } 6977 6978 /* 6979 * Remove all mappings to page 'pp'. 6980 */ 6981 int 6982 hat_pageunload(struct page *pp, uint_t forceflag) 6983 { 6984 struct page *origpp = pp; 6985 struct sf_hment *sfhme, *tmphme; 6986 struct hme_blk *hmeblkp; 6987 kmutex_t *pml; 6988 #ifdef VAC 6989 kmutex_t *pmtx; 6990 #endif 6991 cpuset_t cpuset, tset; 6992 int index, cons; 6993 int xhme_blks; 6994 int pa_hments; 6995 6996 ASSERT(PAGE_EXCL(pp)); 6997 6998 retry_xhat: 6999 tmphme = NULL; 7000 xhme_blks = 0; 7001 pa_hments = 0; 7002 CPUSET_ZERO(cpuset); 7003 7004 pml = sfmmu_mlist_enter(pp); 7005 7006 #ifdef VAC 7007 if (pp->p_kpmref) 7008 sfmmu_kpm_pageunload(pp); 7009 ASSERT(!PP_ISMAPPED_KPM(pp)); 7010 #endif 7011 /* 7012 * Clear vpm reference. Since the page is exclusively locked 7013 * vpm cannot be referencing it. 7014 */ 7015 if (vpm_enable) { 7016 pp->p_vpmref = 0; 7017 } 7018 7019 index = PP_MAPINDEX(pp); 7020 cons = TTE8K; 7021 retry: 7022 for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) { 7023 tmphme = sfhme->hme_next; 7024 7025 if (IS_PAHME(sfhme)) { 7026 ASSERT(sfhme->hme_data != NULL); 7027 pa_hments++; 7028 continue; 7029 } 7030 7031 hmeblkp = sfmmu_hmetohblk(sfhme); 7032 if (hmeblkp->hblk_xhat_bit) { 7033 struct xhat_hme_blk *xblk = 7034 (struct xhat_hme_blk *)hmeblkp; 7035 7036 (void) XHAT_PAGEUNLOAD(xblk->xhat_hme_blk_hat, 7037 pp, forceflag, XBLK2PROVBLK(xblk)); 7038 7039 xhme_blks = 1; 7040 continue; 7041 } 7042 7043 /* 7044 * If there are kernel mappings don't unload them, they will 7045 * be suspended. 7046 */ 7047 if (forceflag == SFMMU_KERNEL_RELOC && hmeblkp->hblk_lckcnt && 7048 hmeblkp->hblk_tag.htag_id == ksfmmup) 7049 continue; 7050 7051 tset = sfmmu_pageunload(pp, sfhme, cons); 7052 CPUSET_OR(cpuset, tset); 7053 } 7054 7055 while (index != 0) { 7056 index = index >> 1; 7057 if (index != 0) 7058 cons++; 7059 if (index & 0x1) { 7060 /* Go to leading page */ 7061 pp = PP_GROUPLEADER(pp, cons); 7062 ASSERT(sfmmu_mlist_held(pp)); 7063 goto retry; 7064 } 7065 } 7066 7067 /* 7068 * cpuset may be empty if the page was only mapped by segkpm, 7069 * in which case we won't actually cross-trap. 7070 */ 7071 xt_sync(cpuset); 7072 7073 /* 7074 * The page should have no mappings at this point, unless 7075 * we were called from hat_page_relocate() in which case we 7076 * leave the locked mappings which will be suspended later. 7077 */ 7078 ASSERT(!PP_ISMAPPED(origpp) || xhme_blks || pa_hments || 7079 (forceflag == SFMMU_KERNEL_RELOC)); 7080 7081 #ifdef VAC 7082 if (PP_ISTNC(pp)) { 7083 if (cons == TTE8K) { 7084 pmtx = sfmmu_page_enter(pp); 7085 PP_CLRTNC(pp); 7086 sfmmu_page_exit(pmtx); 7087 } else { 7088 conv_tnc(pp, cons); 7089 } 7090 } 7091 #endif /* VAC */ 7092 7093 if (pa_hments && forceflag != SFMMU_KERNEL_RELOC) { 7094 /* 7095 * Unlink any pa_hments and free them, calling back 7096 * the responsible subsystem to notify it of the error. 7097 * This can occur in situations such as drivers leaking 7098 * DMA handles: naughty, but common enough that we'd like 7099 * to keep the system running rather than bringing it 7100 * down with an obscure error like "pa_hment leaked" 7101 * which doesn't aid the user in debugging their driver. 7102 */ 7103 for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) { 7104 tmphme = sfhme->hme_next; 7105 if (IS_PAHME(sfhme)) { 7106 struct pa_hment *pahmep = sfhme->hme_data; 7107 sfmmu_pahment_leaked(pahmep); 7108 HME_SUB(sfhme, pp); 7109 kmem_cache_free(pa_hment_cache, pahmep); 7110 } 7111 } 7112 7113 ASSERT(!PP_ISMAPPED(origpp) || xhme_blks); 7114 } 7115 7116 sfmmu_mlist_exit(pml); 7117 7118 /* 7119 * XHAT may not have finished unloading pages 7120 * because some other thread was waiting for 7121 * mlist lock and XHAT_PAGEUNLOAD let it do 7122 * the job. 7123 */ 7124 if (xhme_blks) { 7125 pp = origpp; 7126 goto retry_xhat; 7127 } 7128 7129 return (0); 7130 } 7131 7132 cpuset_t 7133 sfmmu_pageunload(page_t *pp, struct sf_hment *sfhme, int cons) 7134 { 7135 struct hme_blk *hmeblkp; 7136 sfmmu_t *sfmmup; 7137 tte_t tte, ttemod; 7138 #ifdef DEBUG 7139 tte_t orig_old; 7140 #endif /* DEBUG */ 7141 caddr_t addr; 7142 int ttesz; 7143 int ret; 7144 cpuset_t cpuset; 7145 7146 ASSERT(pp != NULL); 7147 ASSERT(sfmmu_mlist_held(pp)); 7148 ASSERT(!PP_ISKAS(pp)); 7149 7150 CPUSET_ZERO(cpuset); 7151 7152 hmeblkp = sfmmu_hmetohblk(sfhme); 7153 7154 readtte: 7155 sfmmu_copytte(&sfhme->hme_tte, &tte); 7156 if (TTE_IS_VALID(&tte)) { 7157 sfmmup = hblktosfmmu(hmeblkp); 7158 ttesz = get_hblk_ttesz(hmeblkp); 7159 /* 7160 * Only unload mappings of 'cons' size. 7161 */ 7162 if (ttesz != cons) 7163 return (cpuset); 7164 7165 /* 7166 * Note that we have p_mapping lock, but no hash lock here. 7167 * hblk_unload() has to have both hash lock AND p_mapping 7168 * lock before it tries to modify tte. So, the tte could 7169 * not become invalid in the sfmmu_modifytte_try() below. 7170 */ 7171 ttemod = tte; 7172 #ifdef DEBUG 7173 orig_old = tte; 7174 #endif /* DEBUG */ 7175 7176 TTE_SET_INVALID(&ttemod); 7177 ret = sfmmu_modifytte_try(&tte, &ttemod, &sfhme->hme_tte); 7178 if (ret < 0) { 7179 #ifdef DEBUG 7180 /* only R/M bits can change. */ 7181 chk_tte(&orig_old, &tte, &ttemod, hmeblkp); 7182 #endif /* DEBUG */ 7183 goto readtte; 7184 } 7185 7186 if (ret == 0) { 7187 panic("pageunload: cas failed?"); 7188 } 7189 7190 addr = tte_to_vaddr(hmeblkp, tte); 7191 7192 if (hmeblkp->hblk_shared) { 7193 sf_srd_t *srdp = (sf_srd_t *)sfmmup; 7194 uint_t rid = hmeblkp->hblk_tag.htag_rid; 7195 sf_region_t *rgnp; 7196 ASSERT(SFMMU_IS_SHMERID_VALID(rid)); 7197 ASSERT(rid < SFMMU_MAX_HME_REGIONS); 7198 ASSERT(srdp != NULL); 7199 rgnp = srdp->srd_hmergnp[rid]; 7200 SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid); 7201 cpuset = sfmmu_rgntlb_demap(addr, rgnp, hmeblkp, 1); 7202 sfmmu_ttesync(NULL, addr, &tte, pp); 7203 ASSERT(rgnp->rgn_ttecnt[ttesz] > 0); 7204 atomic_add_long(&rgnp->rgn_ttecnt[ttesz], -1); 7205 } else { 7206 sfmmu_ttesync(sfmmup, addr, &tte, pp); 7207 atomic_add_long(&sfmmup->sfmmu_ttecnt[ttesz], -1); 7208 7209 /* 7210 * We need to flush the page from the virtual cache 7211 * in order to prevent a virtual cache alias 7212 * inconsistency. The particular scenario we need 7213 * to worry about is: 7214 * Given: va1 and va2 are two virtual address that 7215 * alias and will map the same physical address. 7216 * 1. mapping exists from va1 to pa and data has 7217 * been read into the cache. 7218 * 2. unload va1. 7219 * 3. load va2 and modify data using va2. 7220 * 4 unload va2. 7221 * 5. load va1 and reference data. Unless we flush 7222 * the data cache when we unload we will get 7223 * stale data. 7224 * This scenario is taken care of by using virtual 7225 * page coloring. 7226 */ 7227 if (sfmmup->sfmmu_ismhat) { 7228 /* 7229 * Flush TSBs, TLBs and caches 7230 * of every process 7231 * sharing this ism segment. 7232 */ 7233 sfmmu_hat_lock_all(); 7234 mutex_enter(&ism_mlist_lock); 7235 kpreempt_disable(); 7236 sfmmu_ismtlbcache_demap(addr, sfmmup, hmeblkp, 7237 pp->p_pagenum, CACHE_NO_FLUSH); 7238 kpreempt_enable(); 7239 mutex_exit(&ism_mlist_lock); 7240 sfmmu_hat_unlock_all(); 7241 cpuset = cpu_ready_set; 7242 } else { 7243 sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 0, 0); 7244 cpuset = sfmmup->sfmmu_cpusran; 7245 } 7246 } 7247 7248 /* 7249 * Hme_sub has to run after ttesync() and a_rss update. 7250 * See hblk_unload(). 7251 */ 7252 HME_SUB(sfhme, pp); 7253 membar_stst(); 7254 7255 /* 7256 * We can not make ASSERT(hmeblkp->hblk_hmecnt <= NHMENTS) 7257 * since pteload may have done a HME_ADD() right after 7258 * we did the HME_SUB() above. Hmecnt is now maintained 7259 * by cas only. no lock guranteed its value. The only 7260 * gurantee we have is the hmecnt should not be less than 7261 * what it should be so the hblk will not be taken away. 7262 * It's also important that we decremented the hmecnt after 7263 * we are done with hmeblkp so that this hmeblk won't be 7264 * stolen. 7265 */ 7266 ASSERT(hmeblkp->hblk_hmecnt > 0); 7267 ASSERT(hmeblkp->hblk_vcnt > 0); 7268 atomic_add_16(&hmeblkp->hblk_vcnt, -1); 7269 atomic_add_16(&hmeblkp->hblk_hmecnt, -1); 7270 /* 7271 * This is bug 4063182. 7272 * XXX: fixme 7273 * ASSERT(hmeblkp->hblk_hmecnt || hmeblkp->hblk_vcnt || 7274 * !hmeblkp->hblk_lckcnt); 7275 */ 7276 } else { 7277 panic("invalid tte? pp %p &tte %p", 7278 (void *)pp, (void *)&tte); 7279 } 7280 7281 return (cpuset); 7282 } 7283 7284 /* 7285 * While relocating a kernel page, this function will move the mappings 7286 * from tpp to dpp and modify any associated data with these mappings. 7287 * It also unsuspends the suspended kernel mapping. 7288 */ 7289 static void 7290 hat_pagereload(struct page *tpp, struct page *dpp) 7291 { 7292 struct sf_hment *sfhme; 7293 tte_t tte, ttemod; 7294 int index, cons; 7295 7296 ASSERT(getpil() == PIL_MAX); 7297 ASSERT(sfmmu_mlist_held(tpp)); 7298 ASSERT(sfmmu_mlist_held(dpp)); 7299 7300 index = PP_MAPINDEX(tpp); 7301 cons = TTE8K; 7302 7303 /* Update real mappings to the page */ 7304 retry: 7305 for (sfhme = tpp->p_mapping; sfhme != NULL; sfhme = sfhme->hme_next) { 7306 if (IS_PAHME(sfhme)) 7307 continue; 7308 sfmmu_copytte(&sfhme->hme_tte, &tte); 7309 ttemod = tte; 7310 7311 /* 7312 * replace old pfn with new pfn in TTE 7313 */ 7314 PFN_TO_TTE(ttemod, dpp->p_pagenum); 7315 7316 /* 7317 * clear suspend bit 7318 */ 7319 ASSERT(TTE_IS_SUSPEND(&ttemod)); 7320 TTE_CLR_SUSPEND(&ttemod); 7321 7322 if (sfmmu_modifytte_try(&tte, &ttemod, &sfhme->hme_tte) < 0) 7323 panic("hat_pagereload(): sfmmu_modifytte_try() failed"); 7324 7325 /* 7326 * set hme_page point to new page 7327 */ 7328 sfhme->hme_page = dpp; 7329 } 7330 7331 /* 7332 * move p_mapping list from old page to new page 7333 */ 7334 dpp->p_mapping = tpp->p_mapping; 7335 tpp->p_mapping = NULL; 7336 dpp->p_share = tpp->p_share; 7337 tpp->p_share = 0; 7338 7339 while (index != 0) { 7340 index = index >> 1; 7341 if (index != 0) 7342 cons++; 7343 if (index & 0x1) { 7344 tpp = PP_GROUPLEADER(tpp, cons); 7345 dpp = PP_GROUPLEADER(dpp, cons); 7346 goto retry; 7347 } 7348 } 7349 7350 curthread->t_flag &= ~T_DONTDTRACE; 7351 mutex_exit(&kpr_suspendlock); 7352 } 7353 7354 uint_t 7355 hat_pagesync(struct page *pp, uint_t clearflag) 7356 { 7357 struct sf_hment *sfhme, *tmphme = NULL; 7358 struct hme_blk *hmeblkp; 7359 kmutex_t *pml; 7360 cpuset_t cpuset, tset; 7361 int index, cons; 7362 extern ulong_t po_share; 7363 page_t *save_pp = pp; 7364 int stop_on_sh = 0; 7365 uint_t shcnt; 7366 7367 CPUSET_ZERO(cpuset); 7368 7369 if (PP_ISRO(pp) && (clearflag & HAT_SYNC_STOPON_MOD)) { 7370 return (PP_GENERIC_ATTR(pp)); 7371 } 7372 7373 if ((clearflag & HAT_SYNC_ZERORM) == 0) { 7374 if ((clearflag & HAT_SYNC_STOPON_REF) && PP_ISREF(pp)) { 7375 return (PP_GENERIC_ATTR(pp)); 7376 } 7377 if ((clearflag & HAT_SYNC_STOPON_MOD) && PP_ISMOD(pp)) { 7378 return (PP_GENERIC_ATTR(pp)); 7379 } 7380 if (clearflag & HAT_SYNC_STOPON_SHARED) { 7381 if (pp->p_share > po_share) { 7382 hat_page_setattr(pp, P_REF); 7383 return (PP_GENERIC_ATTR(pp)); 7384 } 7385 stop_on_sh = 1; 7386 shcnt = 0; 7387 } 7388 } 7389 7390 clearflag &= ~HAT_SYNC_STOPON_SHARED; 7391 pml = sfmmu_mlist_enter(pp); 7392 index = PP_MAPINDEX(pp); 7393 cons = TTE8K; 7394 retry: 7395 for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) { 7396 /* 7397 * We need to save the next hment on the list since 7398 * it is possible for pagesync to remove an invalid hment 7399 * from the list. 7400 */ 7401 tmphme = sfhme->hme_next; 7402 if (IS_PAHME(sfhme)) 7403 continue; 7404 /* 7405 * If we are looking for large mappings and this hme doesn't 7406 * reach the range we are seeking, just ignore it. 7407 */ 7408 hmeblkp = sfmmu_hmetohblk(sfhme); 7409 if (hmeblkp->hblk_xhat_bit) 7410 continue; 7411 7412 if (hme_size(sfhme) < cons) 7413 continue; 7414 7415 if (stop_on_sh) { 7416 if (hmeblkp->hblk_shared) { 7417 sf_srd_t *srdp = hblktosrd(hmeblkp); 7418 uint_t rid = hmeblkp->hblk_tag.htag_rid; 7419 sf_region_t *rgnp; 7420 ASSERT(SFMMU_IS_SHMERID_VALID(rid)); 7421 ASSERT(rid < SFMMU_MAX_HME_REGIONS); 7422 ASSERT(srdp != NULL); 7423 rgnp = srdp->srd_hmergnp[rid]; 7424 SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, 7425 rgnp, rid); 7426 shcnt += rgnp->rgn_refcnt; 7427 } else { 7428 shcnt++; 7429 } 7430 if (shcnt > po_share) { 7431 /* 7432 * tell the pager to spare the page this time 7433 * around. 7434 */ 7435 hat_page_setattr(save_pp, P_REF); 7436 index = 0; 7437 break; 7438 } 7439 } 7440 tset = sfmmu_pagesync(pp, sfhme, 7441 clearflag & ~HAT_SYNC_STOPON_RM); 7442 CPUSET_OR(cpuset, tset); 7443 7444 /* 7445 * If clearflag is HAT_SYNC_DONTZERO, break out as soon 7446 * as the "ref" or "mod" is set or share cnt exceeds po_share. 7447 */ 7448 if ((clearflag & ~HAT_SYNC_STOPON_RM) == HAT_SYNC_DONTZERO && 7449 (((clearflag & HAT_SYNC_STOPON_MOD) && PP_ISMOD(save_pp)) || 7450 ((clearflag & HAT_SYNC_STOPON_REF) && PP_ISREF(save_pp)))) { 7451 index = 0; 7452 break; 7453 } 7454 } 7455 7456 while (index) { 7457 index = index >> 1; 7458 cons++; 7459 if (index & 0x1) { 7460 /* Go to leading page */ 7461 pp = PP_GROUPLEADER(pp, cons); 7462 goto retry; 7463 } 7464 } 7465 7466 xt_sync(cpuset); 7467 sfmmu_mlist_exit(pml); 7468 return (PP_GENERIC_ATTR(save_pp)); 7469 } 7470 7471 /* 7472 * Get all the hardware dependent attributes for a page struct 7473 */ 7474 static cpuset_t 7475 sfmmu_pagesync(struct page *pp, struct sf_hment *sfhme, 7476 uint_t clearflag) 7477 { 7478 caddr_t addr; 7479 tte_t tte, ttemod; 7480 struct hme_blk *hmeblkp; 7481 int ret; 7482 sfmmu_t *sfmmup; 7483 cpuset_t cpuset; 7484 7485 ASSERT(pp != NULL); 7486 ASSERT(sfmmu_mlist_held(pp)); 7487 ASSERT((clearflag == HAT_SYNC_DONTZERO) || 7488 (clearflag == HAT_SYNC_ZERORM)); 7489 7490 SFMMU_STAT(sf_pagesync); 7491 7492 CPUSET_ZERO(cpuset); 7493 7494 sfmmu_pagesync_retry: 7495 7496 sfmmu_copytte(&sfhme->hme_tte, &tte); 7497 if (TTE_IS_VALID(&tte)) { 7498 hmeblkp = sfmmu_hmetohblk(sfhme); 7499 sfmmup = hblktosfmmu(hmeblkp); 7500 addr = tte_to_vaddr(hmeblkp, tte); 7501 if (clearflag == HAT_SYNC_ZERORM) { 7502 ttemod = tte; 7503 TTE_CLR_RM(&ttemod); 7504 ret = sfmmu_modifytte_try(&tte, &ttemod, 7505 &sfhme->hme_tte); 7506 if (ret < 0) { 7507 /* 7508 * cas failed and the new value is not what 7509 * we want. 7510 */ 7511 goto sfmmu_pagesync_retry; 7512 } 7513 7514 if (ret > 0) { 7515 /* we win the cas */ 7516 if (hmeblkp->hblk_shared) { 7517 sf_srd_t *srdp = (sf_srd_t *)sfmmup; 7518 uint_t rid = 7519 hmeblkp->hblk_tag.htag_rid; 7520 sf_region_t *rgnp; 7521 ASSERT(SFMMU_IS_SHMERID_VALID(rid)); 7522 ASSERT(rid < SFMMU_MAX_HME_REGIONS); 7523 ASSERT(srdp != NULL); 7524 rgnp = srdp->srd_hmergnp[rid]; 7525 SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, 7526 srdp, rgnp, rid); 7527 cpuset = sfmmu_rgntlb_demap(addr, 7528 rgnp, hmeblkp, 1); 7529 } else { 7530 sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 7531 0, 0); 7532 cpuset = sfmmup->sfmmu_cpusran; 7533 } 7534 } 7535 } 7536 sfmmu_ttesync(hmeblkp->hblk_shared ? NULL : sfmmup, addr, 7537 &tte, pp); 7538 } 7539 return (cpuset); 7540 } 7541 7542 /* 7543 * Remove write permission from a mappings to a page, so that 7544 * we can detect the next modification of it. This requires modifying 7545 * the TTE then invalidating (demap) any TLB entry using that TTE. 7546 * This code is similar to sfmmu_pagesync(). 7547 */ 7548 static cpuset_t 7549 sfmmu_pageclrwrt(struct page *pp, struct sf_hment *sfhme) 7550 { 7551 caddr_t addr; 7552 tte_t tte; 7553 tte_t ttemod; 7554 struct hme_blk *hmeblkp; 7555 int ret; 7556 sfmmu_t *sfmmup; 7557 cpuset_t cpuset; 7558 7559 ASSERT(pp != NULL); 7560 ASSERT(sfmmu_mlist_held(pp)); 7561 7562 CPUSET_ZERO(cpuset); 7563 SFMMU_STAT(sf_clrwrt); 7564 7565 retry: 7566 7567 sfmmu_copytte(&sfhme->hme_tte, &tte); 7568 if (TTE_IS_VALID(&tte) && TTE_IS_WRITABLE(&tte)) { 7569 hmeblkp = sfmmu_hmetohblk(sfhme); 7570 7571 /* 7572 * xhat mappings should never be to a VMODSORT page. 7573 */ 7574 ASSERT(hmeblkp->hblk_xhat_bit == 0); 7575 7576 sfmmup = hblktosfmmu(hmeblkp); 7577 addr = tte_to_vaddr(hmeblkp, tte); 7578 7579 ttemod = tte; 7580 TTE_CLR_WRT(&ttemod); 7581 TTE_CLR_MOD(&ttemod); 7582 ret = sfmmu_modifytte_try(&tte, &ttemod, &sfhme->hme_tte); 7583 7584 /* 7585 * if cas failed and the new value is not what 7586 * we want retry 7587 */ 7588 if (ret < 0) 7589 goto retry; 7590 7591 /* we win the cas */ 7592 if (ret > 0) { 7593 if (hmeblkp->hblk_shared) { 7594 sf_srd_t *srdp = (sf_srd_t *)sfmmup; 7595 uint_t rid = hmeblkp->hblk_tag.htag_rid; 7596 sf_region_t *rgnp; 7597 ASSERT(SFMMU_IS_SHMERID_VALID(rid)); 7598 ASSERT(rid < SFMMU_MAX_HME_REGIONS); 7599 ASSERT(srdp != NULL); 7600 rgnp = srdp->srd_hmergnp[rid]; 7601 SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, 7602 srdp, rgnp, rid); 7603 cpuset = sfmmu_rgntlb_demap(addr, 7604 rgnp, hmeblkp, 1); 7605 } else { 7606 sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 0, 0); 7607 cpuset = sfmmup->sfmmu_cpusran; 7608 } 7609 } 7610 } 7611 7612 return (cpuset); 7613 } 7614 7615 /* 7616 * Walk all mappings of a page, removing write permission and clearing the 7617 * ref/mod bits. This code is similar to hat_pagesync() 7618 */ 7619 static void 7620 hat_page_clrwrt(page_t *pp) 7621 { 7622 struct sf_hment *sfhme; 7623 struct sf_hment *tmphme = NULL; 7624 kmutex_t *pml; 7625 cpuset_t cpuset; 7626 cpuset_t tset; 7627 int index; 7628 int cons; 7629 7630 CPUSET_ZERO(cpuset); 7631 7632 pml = sfmmu_mlist_enter(pp); 7633 index = PP_MAPINDEX(pp); 7634 cons = TTE8K; 7635 retry: 7636 for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) { 7637 tmphme = sfhme->hme_next; 7638 7639 /* 7640 * If we are looking for large mappings and this hme doesn't 7641 * reach the range we are seeking, just ignore its. 7642 */ 7643 7644 if (hme_size(sfhme) < cons) 7645 continue; 7646 7647 tset = sfmmu_pageclrwrt(pp, sfhme); 7648 CPUSET_OR(cpuset, tset); 7649 } 7650 7651 while (index) { 7652 index = index >> 1; 7653 cons++; 7654 if (index & 0x1) { 7655 /* Go to leading page */ 7656 pp = PP_GROUPLEADER(pp, cons); 7657 goto retry; 7658 } 7659 } 7660 7661 xt_sync(cpuset); 7662 sfmmu_mlist_exit(pml); 7663 } 7664 7665 /* 7666 * Set the given REF/MOD/RO bits for the given page. 7667 * For a vnode with a sorted v_pages list, we need to change 7668 * the attributes and the v_pages list together under page_vnode_mutex. 7669 */ 7670 void 7671 hat_page_setattr(page_t *pp, uint_t flag) 7672 { 7673 vnode_t *vp = pp->p_vnode; 7674 page_t **listp; 7675 kmutex_t *pmtx; 7676 kmutex_t *vphm = NULL; 7677 int noshuffle; 7678 7679 noshuffle = flag & P_NSH; 7680 flag &= ~P_NSH; 7681 7682 ASSERT(!(flag & ~(P_MOD | P_REF | P_RO))); 7683 7684 /* 7685 * nothing to do if attribute already set 7686 */ 7687 if ((pp->p_nrm & flag) == flag) 7688 return; 7689 7690 if ((flag & P_MOD) != 0 && vp != NULL && IS_VMODSORT(vp) && 7691 !noshuffle) { 7692 vphm = page_vnode_mutex(vp); 7693 mutex_enter(vphm); 7694 } 7695 7696 pmtx = sfmmu_page_enter(pp); 7697 pp->p_nrm |= flag; 7698 sfmmu_page_exit(pmtx); 7699 7700 if (vphm != NULL) { 7701 /* 7702 * Some File Systems examine v_pages for NULL w/o 7703 * grabbing the vphm mutex. Must not let it become NULL when 7704 * pp is the only page on the list. 7705 */ 7706 if (pp->p_vpnext != pp) { 7707 page_vpsub(&vp->v_pages, pp); 7708 if (vp->v_pages != NULL) 7709 listp = &vp->v_pages->p_vpprev->p_vpnext; 7710 else 7711 listp = &vp->v_pages; 7712 page_vpadd(listp, pp); 7713 } 7714 mutex_exit(vphm); 7715 } 7716 } 7717 7718 void 7719 hat_page_clrattr(page_t *pp, uint_t flag) 7720 { 7721 vnode_t *vp = pp->p_vnode; 7722 kmutex_t *pmtx; 7723 7724 ASSERT(!(flag & ~(P_MOD | P_REF | P_RO))); 7725 7726 pmtx = sfmmu_page_enter(pp); 7727 7728 /* 7729 * Caller is expected to hold page's io lock for VMODSORT to work 7730 * correctly with pvn_vplist_dirty() and pvn_getdirty() when mod 7731 * bit is cleared. 7732 * We don't have assert to avoid tripping some existing third party 7733 * code. The dirty page is moved back to top of the v_page list 7734 * after IO is done in pvn_write_done(). 7735 */ 7736 pp->p_nrm &= ~flag; 7737 sfmmu_page_exit(pmtx); 7738 7739 if ((flag & P_MOD) != 0 && vp != NULL && IS_VMODSORT(vp)) { 7740 7741 /* 7742 * VMODSORT works by removing write permissions and getting 7743 * a fault when a page is made dirty. At this point 7744 * we need to remove write permission from all mappings 7745 * to this page. 7746 */ 7747 hat_page_clrwrt(pp); 7748 } 7749 } 7750 7751 uint_t 7752 hat_page_getattr(page_t *pp, uint_t flag) 7753 { 7754 ASSERT(!(flag & ~(P_MOD | P_REF | P_RO))); 7755 return ((uint_t)(pp->p_nrm & flag)); 7756 } 7757 7758 /* 7759 * DEBUG kernels: verify that a kernel va<->pa translation 7760 * is safe by checking the underlying page_t is in a page 7761 * relocation-safe state. 7762 */ 7763 #ifdef DEBUG 7764 void 7765 sfmmu_check_kpfn(pfn_t pfn) 7766 { 7767 page_t *pp; 7768 int index, cons; 7769 7770 if (hat_check_vtop == 0) 7771 return; 7772 7773 if (hat_kpr_enabled == 0 || kvseg.s_base == NULL || panicstr) 7774 return; 7775 7776 pp = page_numtopp_nolock(pfn); 7777 if (!pp) 7778 return; 7779 7780 if (PAGE_LOCKED(pp) || PP_ISNORELOC(pp)) 7781 return; 7782 7783 /* 7784 * Handed a large kernel page, we dig up the root page since we 7785 * know the root page might have the lock also. 7786 */ 7787 if (pp->p_szc != 0) { 7788 index = PP_MAPINDEX(pp); 7789 cons = TTE8K; 7790 again: 7791 while (index != 0) { 7792 index >>= 1; 7793 if (index != 0) 7794 cons++; 7795 if (index & 0x1) { 7796 pp = PP_GROUPLEADER(pp, cons); 7797 goto again; 7798 } 7799 } 7800 } 7801 7802 if (PAGE_LOCKED(pp) || PP_ISNORELOC(pp)) 7803 return; 7804 7805 /* 7806 * Pages need to be locked or allocated "permanent" (either from 7807 * static_arena arena or explicitly setting PG_NORELOC when calling 7808 * page_create_va()) for VA->PA translations to be valid. 7809 */ 7810 if (!PP_ISNORELOC(pp)) 7811 panic("Illegal VA->PA translation, pp 0x%p not permanent", 7812 (void *)pp); 7813 else 7814 panic("Illegal VA->PA translation, pp 0x%p not locked", 7815 (void *)pp); 7816 } 7817 #endif /* DEBUG */ 7818 7819 /* 7820 * Returns a page frame number for a given virtual address. 7821 * Returns PFN_INVALID to indicate an invalid mapping 7822 */ 7823 pfn_t 7824 hat_getpfnum(struct hat *hat, caddr_t addr) 7825 { 7826 pfn_t pfn; 7827 tte_t tte; 7828 7829 /* 7830 * We would like to 7831 * ASSERT(AS_LOCK_HELD(as, &as->a_lock)); 7832 * but we can't because the iommu driver will call this 7833 * routine at interrupt time and it can't grab the as lock 7834 * or it will deadlock: A thread could have the as lock 7835 * and be waiting for io. The io can't complete 7836 * because the interrupt thread is blocked trying to grab 7837 * the as lock. 7838 */ 7839 7840 ASSERT(hat->sfmmu_xhat_provider == NULL); 7841 7842 if (hat == ksfmmup) { 7843 if (IS_KMEM_VA_LARGEPAGE(addr)) { 7844 ASSERT(segkmem_lpszc > 0); 7845 pfn = sfmmu_kvaszc2pfn(addr, segkmem_lpszc); 7846 if (pfn != PFN_INVALID) { 7847 sfmmu_check_kpfn(pfn); 7848 return (pfn); 7849 } 7850 } else if (segkpm && IS_KPM_ADDR(addr)) { 7851 return (sfmmu_kpm_vatopfn(addr)); 7852 } 7853 while ((pfn = sfmmu_vatopfn(addr, ksfmmup, &tte)) 7854 == PFN_SUSPENDED) { 7855 sfmmu_vatopfn_suspended(addr, ksfmmup, &tte); 7856 } 7857 sfmmu_check_kpfn(pfn); 7858 return (pfn); 7859 } else { 7860 return (sfmmu_uvatopfn(addr, hat, NULL)); 7861 } 7862 } 7863 7864 /* 7865 * hat_getkpfnum() is an obsolete DDI routine, and its use is discouraged. 7866 * Use hat_getpfnum(kas.a_hat, ...) instead. 7867 * 7868 * We'd like to return PFN_INVALID if the mappings have underlying page_t's 7869 * but can't right now due to the fact that some software has grown to use 7870 * this interface incorrectly. So for now when the interface is misused, 7871 * return a warning to the user that in the future it won't work in the 7872 * way they're abusing it, and carry on (after disabling page relocation). 7873 */ 7874 pfn_t 7875 hat_getkpfnum(caddr_t addr) 7876 { 7877 pfn_t pfn; 7878 tte_t tte; 7879 int badcaller = 0; 7880 extern int segkmem_reloc; 7881 7882 if (segkpm && IS_KPM_ADDR(addr)) { 7883 badcaller = 1; 7884 pfn = sfmmu_kpm_vatopfn(addr); 7885 } else { 7886 while ((pfn = sfmmu_vatopfn(addr, ksfmmup, &tte)) 7887 == PFN_SUSPENDED) { 7888 sfmmu_vatopfn_suspended(addr, ksfmmup, &tte); 7889 } 7890 badcaller = pf_is_memory(pfn); 7891 } 7892 7893 if (badcaller) { 7894 /* 7895 * We can't return PFN_INVALID or the caller may panic 7896 * or corrupt the system. The only alternative is to 7897 * disable page relocation at this point for all kernel 7898 * memory. This will impact any callers of page_relocate() 7899 * such as FMA or DR. 7900 * 7901 * RFE: Add junk here to spit out an ereport so the sysadmin 7902 * can be advised that he should upgrade his device driver 7903 * so that this doesn't happen. 7904 */ 7905 hat_getkpfnum_badcall(caller()); 7906 if (hat_kpr_enabled && segkmem_reloc) { 7907 hat_kpr_enabled = 0; 7908 segkmem_reloc = 0; 7909 cmn_err(CE_WARN, "Kernel Page Relocation is DISABLED"); 7910 } 7911 } 7912 return (pfn); 7913 } 7914 7915 /* 7916 * This routine will return both pfn and tte for the vaddr. 7917 */ 7918 static pfn_t 7919 sfmmu_uvatopfn(caddr_t vaddr, struct hat *sfmmup, tte_t *ttep) 7920 { 7921 struct hmehash_bucket *hmebp; 7922 hmeblk_tag hblktag; 7923 int hmeshift, hashno = 1; 7924 struct hme_blk *hmeblkp = NULL; 7925 tte_t tte; 7926 7927 struct sf_hment *sfhmep; 7928 pfn_t pfn; 7929 7930 /* support for ISM */ 7931 ism_map_t *ism_map; 7932 ism_blk_t *ism_blkp; 7933 int i; 7934 sfmmu_t *ism_hatid = NULL; 7935 sfmmu_t *locked_hatid = NULL; 7936 sfmmu_t *sv_sfmmup = sfmmup; 7937 caddr_t sv_vaddr = vaddr; 7938 sf_srd_t *srdp; 7939 7940 if (ttep == NULL) { 7941 ttep = &tte; 7942 } else { 7943 ttep->ll = 0; 7944 } 7945 7946 ASSERT(sfmmup != ksfmmup); 7947 SFMMU_STAT(sf_user_vtop); 7948 /* 7949 * Set ism_hatid if vaddr falls in a ISM segment. 7950 */ 7951 ism_blkp = sfmmup->sfmmu_iblk; 7952 if (ism_blkp != NULL) { 7953 sfmmu_ismhat_enter(sfmmup, 0); 7954 locked_hatid = sfmmup; 7955 } 7956 while (ism_blkp != NULL && ism_hatid == NULL) { 7957 ism_map = ism_blkp->iblk_maps; 7958 for (i = 0; ism_map[i].imap_ismhat && i < ISM_MAP_SLOTS; i++) { 7959 if (vaddr >= ism_start(ism_map[i]) && 7960 vaddr < ism_end(ism_map[i])) { 7961 sfmmup = ism_hatid = ism_map[i].imap_ismhat; 7962 vaddr = (caddr_t)(vaddr - 7963 ism_start(ism_map[i])); 7964 break; 7965 } 7966 } 7967 ism_blkp = ism_blkp->iblk_next; 7968 } 7969 if (locked_hatid) { 7970 sfmmu_ismhat_exit(locked_hatid, 0); 7971 } 7972 7973 hblktag.htag_id = sfmmup; 7974 hblktag.htag_rid = SFMMU_INVALID_SHMERID; 7975 do { 7976 hmeshift = HME_HASH_SHIFT(hashno); 7977 hblktag.htag_bspage = HME_HASH_BSPAGE(vaddr, hmeshift); 7978 hblktag.htag_rehash = hashno; 7979 hmebp = HME_HASH_FUNCTION(sfmmup, vaddr, hmeshift); 7980 7981 SFMMU_HASH_LOCK(hmebp); 7982 7983 HME_HASH_FAST_SEARCH(hmebp, hblktag, hmeblkp); 7984 if (hmeblkp != NULL) { 7985 ASSERT(!hmeblkp->hblk_shared); 7986 HBLKTOHME(sfhmep, hmeblkp, vaddr); 7987 sfmmu_copytte(&sfhmep->hme_tte, ttep); 7988 SFMMU_HASH_UNLOCK(hmebp); 7989 if (TTE_IS_VALID(ttep)) { 7990 pfn = TTE_TO_PFN(vaddr, ttep); 7991 return (pfn); 7992 } 7993 break; 7994 } 7995 SFMMU_HASH_UNLOCK(hmebp); 7996 hashno++; 7997 } while (HME_REHASH(sfmmup) && (hashno <= mmu_hashcnt)); 7998 7999 if (SF_HMERGNMAP_ISNULL(sv_sfmmup)) { 8000 return (PFN_INVALID); 8001 } 8002 srdp = sv_sfmmup->sfmmu_srdp; 8003 ASSERT(srdp != NULL); 8004 ASSERT(srdp->srd_refcnt != 0); 8005 hblktag.htag_id = srdp; 8006 hashno = 1; 8007 do { 8008 hmeshift = HME_HASH_SHIFT(hashno); 8009 hblktag.htag_bspage = HME_HASH_BSPAGE(sv_vaddr, hmeshift); 8010 hblktag.htag_rehash = hashno; 8011 hmebp = HME_HASH_FUNCTION(srdp, sv_vaddr, hmeshift); 8012 8013 SFMMU_HASH_LOCK(hmebp); 8014 for (hmeblkp = hmebp->hmeblkp; hmeblkp != NULL; 8015 hmeblkp = hmeblkp->hblk_next) { 8016 uint_t rid; 8017 sf_region_t *rgnp; 8018 caddr_t rsaddr; 8019 caddr_t readdr; 8020 8021 if (!HTAGS_EQ_SHME(hmeblkp->hblk_tag, hblktag, 8022 sv_sfmmup->sfmmu_hmeregion_map)) { 8023 continue; 8024 } 8025 ASSERT(hmeblkp->hblk_shared); 8026 rid = hmeblkp->hblk_tag.htag_rid; 8027 ASSERT(SFMMU_IS_SHMERID_VALID(rid)); 8028 ASSERT(rid < SFMMU_MAX_HME_REGIONS); 8029 rgnp = srdp->srd_hmergnp[rid]; 8030 SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid); 8031 HBLKTOHME(sfhmep, hmeblkp, sv_vaddr); 8032 sfmmu_copytte(&sfhmep->hme_tte, ttep); 8033 rsaddr = rgnp->rgn_saddr; 8034 readdr = rsaddr + rgnp->rgn_size; 8035 #ifdef DEBUG 8036 if (TTE_IS_VALID(ttep) || 8037 get_hblk_ttesz(hmeblkp) > TTE8K) { 8038 caddr_t eva = tte_to_evaddr(hmeblkp, ttep); 8039 ASSERT(eva > sv_vaddr); 8040 ASSERT(sv_vaddr >= rsaddr); 8041 ASSERT(sv_vaddr < readdr); 8042 ASSERT(eva <= readdr); 8043 } 8044 #endif /* DEBUG */ 8045 /* 8046 * Continue the search if we 8047 * found an invalid 8K tte outside of the area 8048 * covered by this hmeblk's region. 8049 */ 8050 if (TTE_IS_VALID(ttep)) { 8051 SFMMU_HASH_UNLOCK(hmebp); 8052 pfn = TTE_TO_PFN(sv_vaddr, ttep); 8053 return (pfn); 8054 } else if (get_hblk_ttesz(hmeblkp) > TTE8K || 8055 (sv_vaddr >= rsaddr && sv_vaddr < readdr)) { 8056 SFMMU_HASH_UNLOCK(hmebp); 8057 pfn = PFN_INVALID; 8058 return (pfn); 8059 } 8060 } 8061 SFMMU_HASH_UNLOCK(hmebp); 8062 hashno++; 8063 } while (hashno <= mmu_hashcnt); 8064 return (PFN_INVALID); 8065 } 8066 8067 8068 /* 8069 * For compatability with AT&T and later optimizations 8070 */ 8071 /* ARGSUSED */ 8072 void 8073 hat_map(struct hat *hat, caddr_t addr, size_t len, uint_t flags) 8074 { 8075 ASSERT(hat != NULL); 8076 ASSERT(hat->sfmmu_xhat_provider == NULL); 8077 } 8078 8079 /* 8080 * Return the number of mappings to a particular page. This number is an 8081 * approximation of the number of people sharing the page. 8082 * 8083 * shared hmeblks or ism hmeblks are counted as 1 mapping here. 8084 * hat_page_checkshare() can be used to compare threshold to share 8085 * count that reflects the number of region sharers albeit at higher cost. 8086 */ 8087 ulong_t 8088 hat_page_getshare(page_t *pp) 8089 { 8090 page_t *spp = pp; /* start page */ 8091 kmutex_t *pml; 8092 ulong_t cnt; 8093 int index, sz = TTE64K; 8094 8095 /* 8096 * We need to grab the mlist lock to make sure any outstanding 8097 * load/unloads complete. Otherwise we could return zero 8098 * even though the unload(s) hasn't finished yet. 8099 */ 8100 pml = sfmmu_mlist_enter(spp); 8101 cnt = spp->p_share; 8102 8103 #ifdef VAC 8104 if (kpm_enable) 8105 cnt += spp->p_kpmref; 8106 #endif 8107 if (vpm_enable && pp->p_vpmref) { 8108 cnt += 1; 8109 } 8110 8111 /* 8112 * If we have any large mappings, we count the number of 8113 * mappings that this large page is part of. 8114 */ 8115 index = PP_MAPINDEX(spp); 8116 index >>= 1; 8117 while (index) { 8118 pp = PP_GROUPLEADER(spp, sz); 8119 if ((index & 0x1) && pp != spp) { 8120 cnt += pp->p_share; 8121 spp = pp; 8122 } 8123 index >>= 1; 8124 sz++; 8125 } 8126 sfmmu_mlist_exit(pml); 8127 return (cnt); 8128 } 8129 8130 /* 8131 * Return 1 if the number of mappings exceeds sh_thresh. Return 0 8132 * otherwise. Count shared hmeblks by region's refcnt. 8133 */ 8134 int 8135 hat_page_checkshare(page_t *pp, ulong_t sh_thresh) 8136 { 8137 kmutex_t *pml; 8138 ulong_t cnt = 0; 8139 int index, sz = TTE8K; 8140 struct sf_hment *sfhme, *tmphme = NULL; 8141 struct hme_blk *hmeblkp; 8142 8143 pml = sfmmu_mlist_enter(pp); 8144 8145 #ifdef VAC 8146 if (kpm_enable) 8147 cnt = pp->p_kpmref; 8148 #endif 8149 8150 if (vpm_enable && pp->p_vpmref) { 8151 cnt += 1; 8152 } 8153 8154 if (pp->p_share + cnt > sh_thresh) { 8155 sfmmu_mlist_exit(pml); 8156 return (1); 8157 } 8158 8159 index = PP_MAPINDEX(pp); 8160 8161 again: 8162 for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) { 8163 tmphme = sfhme->hme_next; 8164 if (IS_PAHME(sfhme)) { 8165 continue; 8166 } 8167 8168 hmeblkp = sfmmu_hmetohblk(sfhme); 8169 if (hmeblkp->hblk_xhat_bit) { 8170 cnt++; 8171 if (cnt > sh_thresh) { 8172 sfmmu_mlist_exit(pml); 8173 return (1); 8174 } 8175 continue; 8176 } 8177 if (hme_size(sfhme) != sz) { 8178 continue; 8179 } 8180 8181 if (hmeblkp->hblk_shared) { 8182 sf_srd_t *srdp = hblktosrd(hmeblkp); 8183 uint_t rid = hmeblkp->hblk_tag.htag_rid; 8184 sf_region_t *rgnp; 8185 ASSERT(SFMMU_IS_SHMERID_VALID(rid)); 8186 ASSERT(rid < SFMMU_MAX_HME_REGIONS); 8187 ASSERT(srdp != NULL); 8188 rgnp = srdp->srd_hmergnp[rid]; 8189 SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, 8190 rgnp, rid); 8191 cnt += rgnp->rgn_refcnt; 8192 } else { 8193 cnt++; 8194 } 8195 if (cnt > sh_thresh) { 8196 sfmmu_mlist_exit(pml); 8197 return (1); 8198 } 8199 } 8200 8201 index >>= 1; 8202 sz++; 8203 while (index) { 8204 pp = PP_GROUPLEADER(pp, sz); 8205 ASSERT(sfmmu_mlist_held(pp)); 8206 if (index & 0x1) { 8207 goto again; 8208 } 8209 index >>= 1; 8210 sz++; 8211 } 8212 sfmmu_mlist_exit(pml); 8213 return (0); 8214 } 8215 8216 /* 8217 * Unload all large mappings to the pp and reset the p_szc field of every 8218 * constituent page according to the remaining mappings. 8219 * 8220 * pp must be locked SE_EXCL. Even though no other constituent pages are 8221 * locked it's legal to unload the large mappings to the pp because all 8222 * constituent pages of large locked mappings have to be locked SE_SHARED. 8223 * This means if we have SE_EXCL lock on one of constituent pages none of the 8224 * large mappings to pp are locked. 8225 * 8226 * Decrease p_szc field starting from the last constituent page and ending 8227 * with the root page. This method is used because other threads rely on the 8228 * root's p_szc to find the lock to syncronize on. After a root page_t's p_szc 8229 * is demoted then other threads will succeed in sfmmu_mlspl_enter(). This 8230 * ensures that p_szc changes of the constituent pages appears atomic for all 8231 * threads that use sfmmu_mlspl_enter() to examine p_szc field. 8232 * 8233 * This mechanism is only used for file system pages where it's not always 8234 * possible to get SE_EXCL locks on all constituent pages to demote the size 8235 * code (as is done for anonymous or kernel large pages). 8236 * 8237 * See more comments in front of sfmmu_mlspl_enter(). 8238 */ 8239 void 8240 hat_page_demote(page_t *pp) 8241 { 8242 int index; 8243 int sz; 8244 cpuset_t cpuset; 8245 int sync = 0; 8246 page_t *rootpp; 8247 struct sf_hment *sfhme; 8248 struct sf_hment *tmphme = NULL; 8249 struct hme_blk *hmeblkp; 8250 uint_t pszc; 8251 page_t *lastpp; 8252 cpuset_t tset; 8253 pgcnt_t npgs; 8254 kmutex_t *pml; 8255 kmutex_t *pmtx = NULL; 8256 8257 ASSERT(PAGE_EXCL(pp)); 8258 ASSERT(!PP_ISFREE(pp)); 8259 ASSERT(!PP_ISKAS(pp)); 8260 ASSERT(page_szc_lock_assert(pp)); 8261 pml = sfmmu_mlist_enter(pp); 8262 8263 pszc = pp->p_szc; 8264 if (pszc == 0) { 8265 goto out; 8266 } 8267 8268 index = PP_MAPINDEX(pp) >> 1; 8269 8270 if (index) { 8271 CPUSET_ZERO(cpuset); 8272 sz = TTE64K; 8273 sync = 1; 8274 } 8275 8276 while (index) { 8277 if (!(index & 0x1)) { 8278 index >>= 1; 8279 sz++; 8280 continue; 8281 } 8282 ASSERT(sz <= pszc); 8283 rootpp = PP_GROUPLEADER(pp, sz); 8284 for (sfhme = rootpp->p_mapping; sfhme; sfhme = tmphme) { 8285 tmphme = sfhme->hme_next; 8286 ASSERT(!IS_PAHME(sfhme)); 8287 hmeblkp = sfmmu_hmetohblk(sfhme); 8288 if (hme_size(sfhme) != sz) { 8289 continue; 8290 } 8291 if (hmeblkp->hblk_xhat_bit) { 8292 cmn_err(CE_PANIC, 8293 "hat_page_demote: xhat hmeblk"); 8294 } 8295 tset = sfmmu_pageunload(rootpp, sfhme, sz); 8296 CPUSET_OR(cpuset, tset); 8297 } 8298 if (index >>= 1) { 8299 sz++; 8300 } 8301 } 8302 8303 ASSERT(!PP_ISMAPPED_LARGE(pp)); 8304 8305 if (sync) { 8306 xt_sync(cpuset); 8307 #ifdef VAC 8308 if (PP_ISTNC(pp)) { 8309 conv_tnc(rootpp, sz); 8310 } 8311 #endif /* VAC */ 8312 } 8313 8314 pmtx = sfmmu_page_enter(pp); 8315 8316 ASSERT(pp->p_szc == pszc); 8317 rootpp = PP_PAGEROOT(pp); 8318 ASSERT(rootpp->p_szc == pszc); 8319 lastpp = PP_PAGENEXT_N(rootpp, TTEPAGES(pszc) - 1); 8320 8321 while (lastpp != rootpp) { 8322 sz = PP_MAPINDEX(lastpp) ? fnd_mapping_sz(lastpp) : 0; 8323 ASSERT(sz < pszc); 8324 npgs = (sz == 0) ? 1 : TTEPAGES(sz); 8325 ASSERT(P2PHASE(lastpp->p_pagenum, npgs) == npgs - 1); 8326 while (--npgs > 0) { 8327 lastpp->p_szc = (uchar_t)sz; 8328 lastpp = PP_PAGEPREV(lastpp); 8329 } 8330 if (sz) { 8331 /* 8332 * make sure before current root's pszc 8333 * is updated all updates to constituent pages pszc 8334 * fields are globally visible. 8335 */ 8336 membar_producer(); 8337 } 8338 lastpp->p_szc = sz; 8339 ASSERT(IS_P2ALIGNED(lastpp->p_pagenum, TTEPAGES(sz))); 8340 if (lastpp != rootpp) { 8341 lastpp = PP_PAGEPREV(lastpp); 8342 } 8343 } 8344 if (sz == 0) { 8345 /* the loop above doesn't cover this case */ 8346 rootpp->p_szc = 0; 8347 } 8348 out: 8349 ASSERT(pp->p_szc == 0); 8350 if (pmtx != NULL) { 8351 sfmmu_page_exit(pmtx); 8352 } 8353 sfmmu_mlist_exit(pml); 8354 } 8355 8356 /* 8357 * Refresh the HAT ismttecnt[] element for size szc. 8358 * Caller must have set ISM busy flag to prevent mapping 8359 * lists from changing while we're traversing them. 8360 */ 8361 pgcnt_t 8362 ism_tsb_entries(sfmmu_t *sfmmup, int szc) 8363 { 8364 ism_blk_t *ism_blkp = sfmmup->sfmmu_iblk; 8365 ism_map_t *ism_map; 8366 pgcnt_t npgs = 0; 8367 pgcnt_t npgs_scd = 0; 8368 int j; 8369 sf_scd_t *scdp; 8370 uchar_t rid; 8371 8372 ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY)); 8373 scdp = sfmmup->sfmmu_scdp; 8374 8375 for (; ism_blkp != NULL; ism_blkp = ism_blkp->iblk_next) { 8376 ism_map = ism_blkp->iblk_maps; 8377 for (j = 0; ism_map[j].imap_ismhat && j < ISM_MAP_SLOTS; j++) { 8378 rid = ism_map[j].imap_rid; 8379 ASSERT(rid == SFMMU_INVALID_ISMRID || 8380 rid < sfmmup->sfmmu_srdp->srd_next_ismrid); 8381 8382 if (scdp != NULL && rid != SFMMU_INVALID_ISMRID && 8383 SF_RGNMAP_TEST(scdp->scd_ismregion_map, rid)) { 8384 /* ISM is in sfmmup's SCD */ 8385 npgs_scd += 8386 ism_map[j].imap_ismhat->sfmmu_ttecnt[szc]; 8387 } else { 8388 /* ISMs is not in SCD */ 8389 npgs += 8390 ism_map[j].imap_ismhat->sfmmu_ttecnt[szc]; 8391 } 8392 } 8393 } 8394 sfmmup->sfmmu_ismttecnt[szc] = npgs; 8395 sfmmup->sfmmu_scdismttecnt[szc] = npgs_scd; 8396 return (npgs); 8397 } 8398 8399 /* 8400 * Yield the memory claim requirement for an address space. 8401 * 8402 * This is currently implemented as the number of bytes that have active 8403 * hardware translations that have page structures. Therefore, it can 8404 * underestimate the traditional resident set size, eg, if the 8405 * physical page is present and the hardware translation is missing; 8406 * and it can overestimate the rss, eg, if there are active 8407 * translations to a frame buffer with page structs. 8408 * Also, it does not take sharing into account. 8409 * 8410 * Note that we don't acquire locks here since this function is most often 8411 * called from the clock thread. 8412 */ 8413 size_t 8414 hat_get_mapped_size(struct hat *hat) 8415 { 8416 size_t assize = 0; 8417 int i; 8418 8419 if (hat == NULL) 8420 return (0); 8421 8422 ASSERT(hat->sfmmu_xhat_provider == NULL); 8423 8424 for (i = 0; i < mmu_page_sizes; i++) 8425 assize += ((pgcnt_t)hat->sfmmu_ttecnt[i] + 8426 (pgcnt_t)hat->sfmmu_scdrttecnt[i]) * TTEBYTES(i); 8427 8428 if (hat->sfmmu_iblk == NULL) 8429 return (assize); 8430 8431 for (i = 0; i < mmu_page_sizes; i++) 8432 assize += ((pgcnt_t)hat->sfmmu_ismttecnt[i] + 8433 (pgcnt_t)hat->sfmmu_scdismttecnt[i]) * TTEBYTES(i); 8434 8435 return (assize); 8436 } 8437 8438 int 8439 hat_stats_enable(struct hat *hat) 8440 { 8441 hatlock_t *hatlockp; 8442 8443 ASSERT(hat->sfmmu_xhat_provider == NULL); 8444 8445 hatlockp = sfmmu_hat_enter(hat); 8446 hat->sfmmu_rmstat++; 8447 sfmmu_hat_exit(hatlockp); 8448 return (1); 8449 } 8450 8451 void 8452 hat_stats_disable(struct hat *hat) 8453 { 8454 hatlock_t *hatlockp; 8455 8456 ASSERT(hat->sfmmu_xhat_provider == NULL); 8457 8458 hatlockp = sfmmu_hat_enter(hat); 8459 hat->sfmmu_rmstat--; 8460 sfmmu_hat_exit(hatlockp); 8461 } 8462 8463 /* 8464 * Routines for entering or removing ourselves from the 8465 * ism_hat's mapping list. This is used for both private and 8466 * SCD hats. 8467 */ 8468 static void 8469 iment_add(struct ism_ment *iment, struct hat *ism_hat) 8470 { 8471 ASSERT(MUTEX_HELD(&ism_mlist_lock)); 8472 8473 iment->iment_prev = NULL; 8474 iment->iment_next = ism_hat->sfmmu_iment; 8475 if (ism_hat->sfmmu_iment) { 8476 ism_hat->sfmmu_iment->iment_prev = iment; 8477 } 8478 ism_hat->sfmmu_iment = iment; 8479 } 8480 8481 static void 8482 iment_sub(struct ism_ment *iment, struct hat *ism_hat) 8483 { 8484 ASSERT(MUTEX_HELD(&ism_mlist_lock)); 8485 8486 if (ism_hat->sfmmu_iment == NULL) { 8487 panic("ism map entry remove - no entries"); 8488 } 8489 8490 if (iment->iment_prev) { 8491 ASSERT(ism_hat->sfmmu_iment != iment); 8492 iment->iment_prev->iment_next = iment->iment_next; 8493 } else { 8494 ASSERT(ism_hat->sfmmu_iment == iment); 8495 ism_hat->sfmmu_iment = iment->iment_next; 8496 } 8497 8498 if (iment->iment_next) { 8499 iment->iment_next->iment_prev = iment->iment_prev; 8500 } 8501 8502 /* 8503 * zero out the entry 8504 */ 8505 iment->iment_next = NULL; 8506 iment->iment_prev = NULL; 8507 iment->iment_hat = NULL; 8508 iment->iment_base_va = 0; 8509 } 8510 8511 /* 8512 * Hat_share()/unshare() return an (non-zero) error 8513 * when saddr and daddr are not properly aligned. 8514 * 8515 * The top level mapping element determines the alignment 8516 * requirement for saddr and daddr, depending on different 8517 * architectures. 8518 * 8519 * When hat_share()/unshare() are not supported, 8520 * HATOP_SHARE()/UNSHARE() return 0 8521 */ 8522 int 8523 hat_share(struct hat *sfmmup, caddr_t addr, 8524 struct hat *ism_hatid, caddr_t sptaddr, size_t len, uint_t ismszc) 8525 { 8526 ism_blk_t *ism_blkp; 8527 ism_blk_t *new_iblk; 8528 ism_map_t *ism_map; 8529 ism_ment_t *ism_ment; 8530 int i, added; 8531 hatlock_t *hatlockp; 8532 int reload_mmu = 0; 8533 uint_t ismshift = page_get_shift(ismszc); 8534 size_t ismpgsz = page_get_pagesize(ismszc); 8535 uint_t ismmask = (uint_t)ismpgsz - 1; 8536 size_t sh_size = ISM_SHIFT(ismshift, len); 8537 ushort_t ismhatflag; 8538 hat_region_cookie_t rcookie; 8539 sf_scd_t *old_scdp; 8540 8541 #ifdef DEBUG 8542 caddr_t eaddr = addr + len; 8543 #endif /* DEBUG */ 8544 8545 ASSERT(ism_hatid != NULL && sfmmup != NULL); 8546 ASSERT(sptaddr == ISMID_STARTADDR); 8547 /* 8548 * Check the alignment. 8549 */ 8550 if (!ISM_ALIGNED(ismshift, addr) || !ISM_ALIGNED(ismshift, sptaddr)) 8551 return (EINVAL); 8552 8553 /* 8554 * Check size alignment. 8555 */ 8556 if (!ISM_ALIGNED(ismshift, len)) 8557 return (EINVAL); 8558 8559 ASSERT(sfmmup->sfmmu_xhat_provider == NULL); 8560 8561 /* 8562 * Allocate ism_ment for the ism_hat's mapping list, and an 8563 * ism map blk in case we need one. We must do our 8564 * allocations before acquiring locks to prevent a deadlock 8565 * in the kmem allocator on the mapping list lock. 8566 */ 8567 new_iblk = kmem_cache_alloc(ism_blk_cache, KM_SLEEP); 8568 ism_ment = kmem_cache_alloc(ism_ment_cache, KM_SLEEP); 8569 8570 /* 8571 * Serialize ISM mappings with the ISM busy flag, and also the 8572 * trap handlers. 8573 */ 8574 sfmmu_ismhat_enter(sfmmup, 0); 8575 8576 /* 8577 * Allocate an ism map blk if necessary. 8578 */ 8579 if (sfmmup->sfmmu_iblk == NULL) { 8580 sfmmup->sfmmu_iblk = new_iblk; 8581 bzero(new_iblk, sizeof (*new_iblk)); 8582 new_iblk->iblk_nextpa = (uint64_t)-1; 8583 membar_stst(); /* make sure next ptr visible to all CPUs */ 8584 sfmmup->sfmmu_ismblkpa = va_to_pa((caddr_t)new_iblk); 8585 reload_mmu = 1; 8586 new_iblk = NULL; 8587 } 8588 8589 #ifdef DEBUG 8590 /* 8591 * Make sure mapping does not already exist. 8592 */ 8593 ism_blkp = sfmmup->sfmmu_iblk; 8594 while (ism_blkp != NULL) { 8595 ism_map = ism_blkp->iblk_maps; 8596 for (i = 0; i < ISM_MAP_SLOTS && ism_map[i].imap_ismhat; i++) { 8597 if ((addr >= ism_start(ism_map[i]) && 8598 addr < ism_end(ism_map[i])) || 8599 eaddr > ism_start(ism_map[i]) && 8600 eaddr <= ism_end(ism_map[i])) { 8601 panic("sfmmu_share: Already mapped!"); 8602 } 8603 } 8604 ism_blkp = ism_blkp->iblk_next; 8605 } 8606 #endif /* DEBUG */ 8607 8608 ASSERT(ismszc >= TTE4M); 8609 if (ismszc == TTE4M) { 8610 ismhatflag = HAT_4M_FLAG; 8611 } else if (ismszc == TTE32M) { 8612 ismhatflag = HAT_32M_FLAG; 8613 } else if (ismszc == TTE256M) { 8614 ismhatflag = HAT_256M_FLAG; 8615 } 8616 /* 8617 * Add mapping to first available mapping slot. 8618 */ 8619 ism_blkp = sfmmup->sfmmu_iblk; 8620 added = 0; 8621 while (!added) { 8622 ism_map = ism_blkp->iblk_maps; 8623 for (i = 0; i < ISM_MAP_SLOTS; i++) { 8624 if (ism_map[i].imap_ismhat == NULL) { 8625 8626 ism_map[i].imap_ismhat = ism_hatid; 8627 ism_map[i].imap_vb_shift = (uchar_t)ismshift; 8628 ism_map[i].imap_rid = SFMMU_INVALID_ISMRID; 8629 ism_map[i].imap_hatflags = ismhatflag; 8630 ism_map[i].imap_sz_mask = ismmask; 8631 /* 8632 * imap_seg is checked in ISM_CHECK to see if 8633 * non-NULL, then other info assumed valid. 8634 */ 8635 membar_stst(); 8636 ism_map[i].imap_seg = (uintptr_t)addr | sh_size; 8637 ism_map[i].imap_ment = ism_ment; 8638 8639 /* 8640 * Now add ourselves to the ism_hat's 8641 * mapping list. 8642 */ 8643 ism_ment->iment_hat = sfmmup; 8644 ism_ment->iment_base_va = addr; 8645 ism_hatid->sfmmu_ismhat = 1; 8646 mutex_enter(&ism_mlist_lock); 8647 iment_add(ism_ment, ism_hatid); 8648 mutex_exit(&ism_mlist_lock); 8649 added = 1; 8650 break; 8651 } 8652 } 8653 if (!added && ism_blkp->iblk_next == NULL) { 8654 ism_blkp->iblk_next = new_iblk; 8655 new_iblk = NULL; 8656 bzero(ism_blkp->iblk_next, 8657 sizeof (*ism_blkp->iblk_next)); 8658 ism_blkp->iblk_next->iblk_nextpa = (uint64_t)-1; 8659 membar_stst(); 8660 ism_blkp->iblk_nextpa = 8661 va_to_pa((caddr_t)ism_blkp->iblk_next); 8662 } 8663 ism_blkp = ism_blkp->iblk_next; 8664 } 8665 8666 /* 8667 * After calling hat_join_region, sfmmup may join a new SCD or 8668 * move from the old scd to a new scd, in which case, we want to 8669 * shrink the sfmmup's private tsb size, i.e., pass shrink to 8670 * sfmmu_check_page_sizes at the end of this routine. 8671 */ 8672 old_scdp = sfmmup->sfmmu_scdp; 8673 8674 rcookie = hat_join_region(sfmmup, addr, len, (void *)ism_hatid, 0, 8675 PROT_ALL, ismszc, NULL, HAT_REGION_ISM); 8676 if (rcookie != HAT_INVALID_REGION_COOKIE) { 8677 ism_map[i].imap_rid = (uchar_t)((uint64_t)rcookie); 8678 } 8679 /* 8680 * Update our counters for this sfmmup's ism mappings. 8681 */ 8682 for (i = 0; i <= ismszc; i++) { 8683 if (!(disable_ism_large_pages & (1 << i))) 8684 (void) ism_tsb_entries(sfmmup, i); 8685 } 8686 8687 /* 8688 * For ISM and DISM we do not support 512K pages, so we only only 8689 * search the 4M and 8K/64K hashes for 4 pagesize cpus, and search the 8690 * 256M or 32M, and 4M and 8K/64K hashes for 6 pagesize cpus. 8691 * 8692 * Need to set 32M/256M ISM flags to make sure 8693 * sfmmu_check_page_sizes() enables them on Panther. 8694 */ 8695 ASSERT((disable_ism_large_pages & (1 << TTE512K)) != 0); 8696 8697 switch (ismszc) { 8698 case TTE256M: 8699 if (!SFMMU_FLAGS_ISSET(sfmmup, HAT_256M_ISM)) { 8700 hatlockp = sfmmu_hat_enter(sfmmup); 8701 SFMMU_FLAGS_SET(sfmmup, HAT_256M_ISM); 8702 sfmmu_hat_exit(hatlockp); 8703 } 8704 break; 8705 case TTE32M: 8706 if (!SFMMU_FLAGS_ISSET(sfmmup, HAT_32M_ISM)) { 8707 hatlockp = sfmmu_hat_enter(sfmmup); 8708 SFMMU_FLAGS_SET(sfmmup, HAT_32M_ISM); 8709 sfmmu_hat_exit(hatlockp); 8710 } 8711 break; 8712 default: 8713 break; 8714 } 8715 8716 /* 8717 * If we updated the ismblkpa for this HAT we must make 8718 * sure all CPUs running this process reload their tsbmiss area. 8719 * Otherwise they will fail to load the mappings in the tsbmiss 8720 * handler and will loop calling pagefault(). 8721 */ 8722 if (reload_mmu) { 8723 hatlockp = sfmmu_hat_enter(sfmmup); 8724 sfmmu_sync_mmustate(sfmmup); 8725 sfmmu_hat_exit(hatlockp); 8726 } 8727 8728 sfmmu_ismhat_exit(sfmmup, 0); 8729 8730 /* 8731 * Free up ismblk if we didn't use it. 8732 */ 8733 if (new_iblk != NULL) 8734 kmem_cache_free(ism_blk_cache, new_iblk); 8735 8736 /* 8737 * Check TSB and TLB page sizes. 8738 */ 8739 if (sfmmup->sfmmu_scdp != NULL && old_scdp != sfmmup->sfmmu_scdp) { 8740 sfmmu_check_page_sizes(sfmmup, 0); 8741 } else { 8742 sfmmu_check_page_sizes(sfmmup, 1); 8743 } 8744 return (0); 8745 } 8746 8747 /* 8748 * hat_unshare removes exactly one ism_map from 8749 * this process's as. It expects multiple calls 8750 * to hat_unshare for multiple shm segments. 8751 */ 8752 void 8753 hat_unshare(struct hat *sfmmup, caddr_t addr, size_t len, uint_t ismszc) 8754 { 8755 ism_map_t *ism_map; 8756 ism_ment_t *free_ment = NULL; 8757 ism_blk_t *ism_blkp; 8758 struct hat *ism_hatid; 8759 int found, i; 8760 hatlock_t *hatlockp; 8761 struct tsb_info *tsbinfo; 8762 uint_t ismshift = page_get_shift(ismszc); 8763 size_t sh_size = ISM_SHIFT(ismshift, len); 8764 uchar_t ism_rid; 8765 sf_scd_t *old_scdp; 8766 8767 ASSERT(ISM_ALIGNED(ismshift, addr)); 8768 ASSERT(ISM_ALIGNED(ismshift, len)); 8769 ASSERT(sfmmup != NULL); 8770 ASSERT(sfmmup != ksfmmup); 8771 8772 if (sfmmup->sfmmu_xhat_provider) { 8773 XHAT_UNSHARE(sfmmup, addr, len); 8774 return; 8775 } else { 8776 /* 8777 * This must be a CPU HAT. If the address space has 8778 * XHATs attached, inform all XHATs that ISM segment 8779 * is going away 8780 */ 8781 ASSERT(sfmmup->sfmmu_as != NULL); 8782 if (sfmmup->sfmmu_as->a_xhat != NULL) 8783 xhat_unshare_all(sfmmup->sfmmu_as, addr, len); 8784 } 8785 8786 /* 8787 * Make sure that during the entire time ISM mappings are removed, 8788 * the trap handlers serialize behind us, and that no one else 8789 * can be mucking with ISM mappings. This also lets us get away 8790 * with not doing expensive cross calls to flush the TLB -- we 8791 * just discard the context, flush the entire TSB, and call it 8792 * a day. 8793 */ 8794 sfmmu_ismhat_enter(sfmmup, 0); 8795 8796 /* 8797 * Remove the mapping. 8798 * 8799 * We can't have any holes in the ism map. 8800 * The tsb miss code while searching the ism map will 8801 * stop on an empty map slot. So we must move 8802 * everyone past the hole up 1 if any. 8803 * 8804 * Also empty ism map blks are not freed until the 8805 * process exits. This is to prevent a MT race condition 8806 * between sfmmu_unshare() and sfmmu_tsbmiss_exception(). 8807 */ 8808 found = 0; 8809 ism_blkp = sfmmup->sfmmu_iblk; 8810 while (!found && ism_blkp != NULL) { 8811 ism_map = ism_blkp->iblk_maps; 8812 for (i = 0; i < ISM_MAP_SLOTS; i++) { 8813 if (addr == ism_start(ism_map[i]) && 8814 sh_size == (size_t)(ism_size(ism_map[i]))) { 8815 found = 1; 8816 break; 8817 } 8818 } 8819 if (!found) 8820 ism_blkp = ism_blkp->iblk_next; 8821 } 8822 8823 if (found) { 8824 ism_hatid = ism_map[i].imap_ismhat; 8825 ism_rid = ism_map[i].imap_rid; 8826 ASSERT(ism_hatid != NULL); 8827 ASSERT(ism_hatid->sfmmu_ismhat == 1); 8828 8829 /* 8830 * After hat_leave_region, the sfmmup may leave SCD, 8831 * in which case, we want to grow the private tsb size when 8832 * calling sfmmu_check_page_sizes at the end of the routine. 8833 */ 8834 old_scdp = sfmmup->sfmmu_scdp; 8835 /* 8836 * Then remove ourselves from the region. 8837 */ 8838 if (ism_rid != SFMMU_INVALID_ISMRID) { 8839 hat_leave_region(sfmmup, (void *)((uint64_t)ism_rid), 8840 HAT_REGION_ISM); 8841 } 8842 8843 /* 8844 * And now guarantee that any other cpu 8845 * that tries to process an ISM miss 8846 * will go to tl=0. 8847 */ 8848 hatlockp = sfmmu_hat_enter(sfmmup); 8849 sfmmu_invalidate_ctx(sfmmup); 8850 sfmmu_hat_exit(hatlockp); 8851 8852 /* 8853 * Remove ourselves from the ism mapping list. 8854 */ 8855 mutex_enter(&ism_mlist_lock); 8856 iment_sub(ism_map[i].imap_ment, ism_hatid); 8857 mutex_exit(&ism_mlist_lock); 8858 free_ment = ism_map[i].imap_ment; 8859 8860 /* 8861 * We delete the ism map by copying 8862 * the next map over the current one. 8863 * We will take the next one in the maps 8864 * array or from the next ism_blk. 8865 */ 8866 while (ism_blkp != NULL) { 8867 ism_map = ism_blkp->iblk_maps; 8868 while (i < (ISM_MAP_SLOTS - 1)) { 8869 ism_map[i] = ism_map[i + 1]; 8870 i++; 8871 } 8872 /* i == (ISM_MAP_SLOTS - 1) */ 8873 ism_blkp = ism_blkp->iblk_next; 8874 if (ism_blkp != NULL) { 8875 ism_map[i] = ism_blkp->iblk_maps[0]; 8876 i = 0; 8877 } else { 8878 ism_map[i].imap_seg = 0; 8879 ism_map[i].imap_vb_shift = 0; 8880 ism_map[i].imap_rid = SFMMU_INVALID_ISMRID; 8881 ism_map[i].imap_hatflags = 0; 8882 ism_map[i].imap_sz_mask = 0; 8883 ism_map[i].imap_ismhat = NULL; 8884 ism_map[i].imap_ment = NULL; 8885 } 8886 } 8887 8888 /* 8889 * Now flush entire TSB for the process, since 8890 * demapping page by page can be too expensive. 8891 * We don't have to flush the TLB here anymore 8892 * since we switch to a new TLB ctx instead. 8893 * Also, there is no need to flush if the process 8894 * is exiting since the TSB will be freed later. 8895 */ 8896 if (!sfmmup->sfmmu_free) { 8897 hatlockp = sfmmu_hat_enter(sfmmup); 8898 for (tsbinfo = sfmmup->sfmmu_tsb; tsbinfo != NULL; 8899 tsbinfo = tsbinfo->tsb_next) { 8900 if (tsbinfo->tsb_flags & TSB_SWAPPED) 8901 continue; 8902 if (tsbinfo->tsb_flags & TSB_RELOC_FLAG) { 8903 tsbinfo->tsb_flags |= 8904 TSB_FLUSH_NEEDED; 8905 continue; 8906 } 8907 8908 sfmmu_inv_tsb(tsbinfo->tsb_va, 8909 TSB_BYTES(tsbinfo->tsb_szc)); 8910 } 8911 sfmmu_hat_exit(hatlockp); 8912 } 8913 } 8914 8915 /* 8916 * Update our counters for this sfmmup's ism mappings. 8917 */ 8918 for (i = 0; i <= ismszc; i++) { 8919 if (!(disable_ism_large_pages & (1 << i))) 8920 (void) ism_tsb_entries(sfmmup, i); 8921 } 8922 8923 sfmmu_ismhat_exit(sfmmup, 0); 8924 8925 /* 8926 * We must do our freeing here after dropping locks 8927 * to prevent a deadlock in the kmem allocator on the 8928 * mapping list lock. 8929 */ 8930 if (free_ment != NULL) 8931 kmem_cache_free(ism_ment_cache, free_ment); 8932 8933 /* 8934 * Check TSB and TLB page sizes if the process isn't exiting. 8935 */ 8936 if (!sfmmup->sfmmu_free) { 8937 if (found && old_scdp != NULL && sfmmup->sfmmu_scdp == NULL) { 8938 sfmmu_check_page_sizes(sfmmup, 1); 8939 } else { 8940 sfmmu_check_page_sizes(sfmmup, 0); 8941 } 8942 } 8943 } 8944 8945 /* ARGSUSED */ 8946 static int 8947 sfmmu_idcache_constructor(void *buf, void *cdrarg, int kmflags) 8948 { 8949 /* void *buf is sfmmu_t pointer */ 8950 bzero(buf, sizeof (sfmmu_t)); 8951 8952 return (0); 8953 } 8954 8955 /* ARGSUSED */ 8956 static void 8957 sfmmu_idcache_destructor(void *buf, void *cdrarg) 8958 { 8959 /* void *buf is sfmmu_t pointer */ 8960 } 8961 8962 /* 8963 * setup kmem hmeblks by bzeroing all members and initializing the nextpa 8964 * field to be the pa of this hmeblk 8965 */ 8966 /* ARGSUSED */ 8967 static int 8968 sfmmu_hblkcache_constructor(void *buf, void *cdrarg, int kmflags) 8969 { 8970 struct hme_blk *hmeblkp; 8971 8972 bzero(buf, (size_t)cdrarg); 8973 hmeblkp = (struct hme_blk *)buf; 8974 hmeblkp->hblk_nextpa = va_to_pa((caddr_t)hmeblkp); 8975 8976 #ifdef HBLK_TRACE 8977 mutex_init(&hmeblkp->hblk_audit_lock, NULL, MUTEX_DEFAULT, NULL); 8978 #endif /* HBLK_TRACE */ 8979 8980 return (0); 8981 } 8982 8983 /* ARGSUSED */ 8984 static void 8985 sfmmu_hblkcache_destructor(void *buf, void *cdrarg) 8986 { 8987 8988 #ifdef HBLK_TRACE 8989 8990 struct hme_blk *hmeblkp; 8991 8992 hmeblkp = (struct hme_blk *)buf; 8993 mutex_destroy(&hmeblkp->hblk_audit_lock); 8994 8995 #endif /* HBLK_TRACE */ 8996 } 8997 8998 #define SFMMU_CACHE_RECLAIM_SCAN_RATIO 8 8999 static int sfmmu_cache_reclaim_scan_ratio = SFMMU_CACHE_RECLAIM_SCAN_RATIO; 9000 /* 9001 * The kmem allocator will callback into our reclaim routine when the system 9002 * is running low in memory. We traverse the hash and free up all unused but 9003 * still cached hme_blks. We also traverse the free list and free them up 9004 * as well. 9005 */ 9006 /*ARGSUSED*/ 9007 static void 9008 sfmmu_hblkcache_reclaim(void *cdrarg) 9009 { 9010 int i; 9011 struct hmehash_bucket *hmebp; 9012 struct hme_blk *hmeblkp, *nx_hblk, *pr_hblk = NULL; 9013 static struct hmehash_bucket *uhmehash_reclaim_hand; 9014 static struct hmehash_bucket *khmehash_reclaim_hand; 9015 struct hme_blk *list = NULL, *last_hmeblkp; 9016 cpuset_t cpuset = cpu_ready_set; 9017 cpu_hme_pend_t *cpuhp; 9018 9019 /* Free up hmeblks on the cpu pending lists */ 9020 for (i = 0; i < NCPU; i++) { 9021 cpuhp = &cpu_hme_pend[i]; 9022 if (cpuhp->chp_listp != NULL) { 9023 mutex_enter(&cpuhp->chp_mutex); 9024 if (cpuhp->chp_listp == NULL) { 9025 mutex_exit(&cpuhp->chp_mutex); 9026 continue; 9027 } 9028 for (last_hmeblkp = cpuhp->chp_listp; 9029 last_hmeblkp->hblk_next != NULL; 9030 last_hmeblkp = last_hmeblkp->hblk_next) 9031 ; 9032 last_hmeblkp->hblk_next = list; 9033 list = cpuhp->chp_listp; 9034 cpuhp->chp_listp = NULL; 9035 cpuhp->chp_count = 0; 9036 mutex_exit(&cpuhp->chp_mutex); 9037 } 9038 9039 } 9040 9041 if (list != NULL) { 9042 kpreempt_disable(); 9043 CPUSET_DEL(cpuset, CPU->cpu_id); 9044 xt_sync(cpuset); 9045 xt_sync(cpuset); 9046 kpreempt_enable(); 9047 sfmmu_hblk_free(&list); 9048 list = NULL; 9049 } 9050 9051 hmebp = uhmehash_reclaim_hand; 9052 if (hmebp == NULL || hmebp > &uhme_hash[UHMEHASH_SZ]) 9053 uhmehash_reclaim_hand = hmebp = uhme_hash; 9054 uhmehash_reclaim_hand += UHMEHASH_SZ / sfmmu_cache_reclaim_scan_ratio; 9055 9056 for (i = UHMEHASH_SZ / sfmmu_cache_reclaim_scan_ratio; i; i--) { 9057 if (SFMMU_HASH_LOCK_TRYENTER(hmebp) != 0) { 9058 hmeblkp = hmebp->hmeblkp; 9059 pr_hblk = NULL; 9060 while (hmeblkp) { 9061 nx_hblk = hmeblkp->hblk_next; 9062 if (!hmeblkp->hblk_vcnt && 9063 !hmeblkp->hblk_hmecnt) { 9064 sfmmu_hblk_hash_rm(hmebp, hmeblkp, 9065 pr_hblk, &list, 0); 9066 } else { 9067 pr_hblk = hmeblkp; 9068 } 9069 hmeblkp = nx_hblk; 9070 } 9071 SFMMU_HASH_UNLOCK(hmebp); 9072 } 9073 if (hmebp++ == &uhme_hash[UHMEHASH_SZ]) 9074 hmebp = uhme_hash; 9075 } 9076 9077 hmebp = khmehash_reclaim_hand; 9078 if (hmebp == NULL || hmebp > &khme_hash[KHMEHASH_SZ]) 9079 khmehash_reclaim_hand = hmebp = khme_hash; 9080 khmehash_reclaim_hand += KHMEHASH_SZ / sfmmu_cache_reclaim_scan_ratio; 9081 9082 for (i = KHMEHASH_SZ / sfmmu_cache_reclaim_scan_ratio; i; i--) { 9083 if (SFMMU_HASH_LOCK_TRYENTER(hmebp) != 0) { 9084 hmeblkp = hmebp->hmeblkp; 9085 pr_hblk = NULL; 9086 while (hmeblkp) { 9087 nx_hblk = hmeblkp->hblk_next; 9088 if (!hmeblkp->hblk_vcnt && 9089 !hmeblkp->hblk_hmecnt) { 9090 sfmmu_hblk_hash_rm(hmebp, hmeblkp, 9091 pr_hblk, &list, 0); 9092 } else { 9093 pr_hblk = hmeblkp; 9094 } 9095 hmeblkp = nx_hblk; 9096 } 9097 SFMMU_HASH_UNLOCK(hmebp); 9098 } 9099 if (hmebp++ == &khme_hash[KHMEHASH_SZ]) 9100 hmebp = khme_hash; 9101 } 9102 sfmmu_hblks_list_purge(&list, 0); 9103 } 9104 9105 /* 9106 * sfmmu_get_ppvcolor should become a vm_machdep or hatop interface. 9107 * same goes for sfmmu_get_addrvcolor(). 9108 * 9109 * This function will return the virtual color for the specified page. The 9110 * virtual color corresponds to this page current mapping or its last mapping. 9111 * It is used by memory allocators to choose addresses with the correct 9112 * alignment so vac consistency is automatically maintained. If the page 9113 * has no color it returns -1. 9114 */ 9115 /*ARGSUSED*/ 9116 int 9117 sfmmu_get_ppvcolor(struct page *pp) 9118 { 9119 #ifdef VAC 9120 int color; 9121 9122 if (!(cache & CACHE_VAC) || PP_NEWPAGE(pp)) { 9123 return (-1); 9124 } 9125 color = PP_GET_VCOLOR(pp); 9126 ASSERT(color < mmu_btop(shm_alignment)); 9127 return (color); 9128 #else 9129 return (-1); 9130 #endif /* VAC */ 9131 } 9132 9133 /* 9134 * This function will return the desired alignment for vac consistency 9135 * (vac color) given a virtual address. If no vac is present it returns -1. 9136 */ 9137 /*ARGSUSED*/ 9138 int 9139 sfmmu_get_addrvcolor(caddr_t vaddr) 9140 { 9141 #ifdef VAC 9142 if (cache & CACHE_VAC) { 9143 return (addr_to_vcolor(vaddr)); 9144 } else { 9145 return (-1); 9146 } 9147 #else 9148 return (-1); 9149 #endif /* VAC */ 9150 } 9151 9152 #ifdef VAC 9153 /* 9154 * Check for conflicts. 9155 * A conflict exists if the new and existent mappings do not match in 9156 * their "shm_alignment fields. If conflicts exist, the existant mappings 9157 * are flushed unless one of them is locked. If one of them is locked, then 9158 * the mappings are flushed and converted to non-cacheable mappings. 9159 */ 9160 static void 9161 sfmmu_vac_conflict(struct hat *hat, caddr_t addr, page_t *pp) 9162 { 9163 struct hat *tmphat; 9164 struct sf_hment *sfhmep, *tmphme = NULL; 9165 struct hme_blk *hmeblkp; 9166 int vcolor; 9167 tte_t tte; 9168 9169 ASSERT(sfmmu_mlist_held(pp)); 9170 ASSERT(!PP_ISNC(pp)); /* page better be cacheable */ 9171 9172 vcolor = addr_to_vcolor(addr); 9173 if (PP_NEWPAGE(pp)) { 9174 PP_SET_VCOLOR(pp, vcolor); 9175 return; 9176 } 9177 9178 if (PP_GET_VCOLOR(pp) == vcolor) { 9179 return; 9180 } 9181 9182 if (!PP_ISMAPPED(pp) && !PP_ISMAPPED_KPM(pp)) { 9183 /* 9184 * Previous user of page had a different color 9185 * but since there are no current users 9186 * we just flush the cache and change the color. 9187 */ 9188 SFMMU_STAT(sf_pgcolor_conflict); 9189 sfmmu_cache_flush(pp->p_pagenum, PP_GET_VCOLOR(pp)); 9190 PP_SET_VCOLOR(pp, vcolor); 9191 return; 9192 } 9193 9194 /* 9195 * If we get here we have a vac conflict with a current 9196 * mapping. VAC conflict policy is as follows. 9197 * - The default is to unload the other mappings unless: 9198 * - If we have a large mapping we uncache the page. 9199 * We need to uncache the rest of the large page too. 9200 * - If any of the mappings are locked we uncache the page. 9201 * - If the requested mapping is inconsistent 9202 * with another mapping and that mapping 9203 * is in the same address space we have to 9204 * make it non-cached. The default thing 9205 * to do is unload the inconsistent mapping 9206 * but if they are in the same address space 9207 * we run the risk of unmapping the pc or the 9208 * stack which we will use as we return to the user, 9209 * in which case we can then fault on the thing 9210 * we just unloaded and get into an infinite loop. 9211 */ 9212 if (PP_ISMAPPED_LARGE(pp)) { 9213 int sz; 9214 9215 /* 9216 * Existing mapping is for big pages. We don't unload 9217 * existing big mappings to satisfy new mappings. 9218 * Always convert all mappings to TNC. 9219 */ 9220 sz = fnd_mapping_sz(pp); 9221 pp = PP_GROUPLEADER(pp, sz); 9222 SFMMU_STAT_ADD(sf_uncache_conflict, TTEPAGES(sz)); 9223 sfmmu_page_cache_array(pp, HAT_TMPNC, CACHE_FLUSH, 9224 TTEPAGES(sz)); 9225 9226 return; 9227 } 9228 9229 /* 9230 * check if any mapping is in same as or if it is locked 9231 * since in that case we need to uncache. 9232 */ 9233 for (sfhmep = pp->p_mapping; sfhmep; sfhmep = tmphme) { 9234 tmphme = sfhmep->hme_next; 9235 if (IS_PAHME(sfhmep)) 9236 continue; 9237 hmeblkp = sfmmu_hmetohblk(sfhmep); 9238 if (hmeblkp->hblk_xhat_bit) 9239 continue; 9240 tmphat = hblktosfmmu(hmeblkp); 9241 sfmmu_copytte(&sfhmep->hme_tte, &tte); 9242 ASSERT(TTE_IS_VALID(&tte)); 9243 if (hmeblkp->hblk_shared || tmphat == hat || 9244 hmeblkp->hblk_lckcnt) { 9245 /* 9246 * We have an uncache conflict 9247 */ 9248 SFMMU_STAT(sf_uncache_conflict); 9249 sfmmu_page_cache_array(pp, HAT_TMPNC, CACHE_FLUSH, 1); 9250 return; 9251 } 9252 } 9253 9254 /* 9255 * We have an unload conflict 9256 * We have already checked for LARGE mappings, therefore 9257 * the remaining mapping(s) must be TTE8K. 9258 */ 9259 SFMMU_STAT(sf_unload_conflict); 9260 9261 for (sfhmep = pp->p_mapping; sfhmep; sfhmep = tmphme) { 9262 tmphme = sfhmep->hme_next; 9263 if (IS_PAHME(sfhmep)) 9264 continue; 9265 hmeblkp = sfmmu_hmetohblk(sfhmep); 9266 if (hmeblkp->hblk_xhat_bit) 9267 continue; 9268 ASSERT(!hmeblkp->hblk_shared); 9269 (void) sfmmu_pageunload(pp, sfhmep, TTE8K); 9270 } 9271 9272 if (PP_ISMAPPED_KPM(pp)) 9273 sfmmu_kpm_vac_unload(pp, addr); 9274 9275 /* 9276 * Unloads only do TLB flushes so we need to flush the 9277 * cache here. 9278 */ 9279 sfmmu_cache_flush(pp->p_pagenum, PP_GET_VCOLOR(pp)); 9280 PP_SET_VCOLOR(pp, vcolor); 9281 } 9282 9283 /* 9284 * Whenever a mapping is unloaded and the page is in TNC state, 9285 * we see if the page can be made cacheable again. 'pp' is 9286 * the page that we just unloaded a mapping from, the size 9287 * of mapping that was unloaded is 'ottesz'. 9288 * Remark: 9289 * The recache policy for mpss pages can leave a performance problem 9290 * under the following circumstances: 9291 * . A large page in uncached mode has just been unmapped. 9292 * . All constituent pages are TNC due to a conflicting small mapping. 9293 * . There are many other, non conflicting, small mappings around for 9294 * a lot of the constituent pages. 9295 * . We're called w/ the "old" groupleader page and the old ottesz, 9296 * but this is irrelevant, since we're no more "PP_ISMAPPED_LARGE", so 9297 * we end up w/ TTE8K or npages == 1. 9298 * . We call tst_tnc w/ the old groupleader only, and if there is no 9299 * conflict, we re-cache only this page. 9300 * . All other small mappings are not checked and will be left in TNC mode. 9301 * The problem is not very serious because: 9302 * . mpss is actually only defined for heap and stack, so the probability 9303 * is not very high that a large page mapping exists in parallel to a small 9304 * one (this is possible, but seems to be bad programming style in the 9305 * appl). 9306 * . The problem gets a little bit more serious, when those TNC pages 9307 * have to be mapped into kernel space, e.g. for networking. 9308 * . When VAC alias conflicts occur in applications, this is regarded 9309 * as an application bug. So if kstat's show them, the appl should 9310 * be changed anyway. 9311 */ 9312 void 9313 conv_tnc(page_t *pp, int ottesz) 9314 { 9315 int cursz, dosz; 9316 pgcnt_t curnpgs, dopgs; 9317 pgcnt_t pg64k; 9318 page_t *pp2; 9319 9320 /* 9321 * Determine how big a range we check for TNC and find 9322 * leader page. cursz is the size of the biggest 9323 * mapping that still exist on 'pp'. 9324 */ 9325 if (PP_ISMAPPED_LARGE(pp)) { 9326 cursz = fnd_mapping_sz(pp); 9327 } else { 9328 cursz = TTE8K; 9329 } 9330 9331 if (ottesz >= cursz) { 9332 dosz = ottesz; 9333 pp2 = pp; 9334 } else { 9335 dosz = cursz; 9336 pp2 = PP_GROUPLEADER(pp, dosz); 9337 } 9338 9339 pg64k = TTEPAGES(TTE64K); 9340 dopgs = TTEPAGES(dosz); 9341 9342 ASSERT(dopgs == 1 || ((dopgs & (pg64k - 1)) == 0)); 9343 9344 while (dopgs != 0) { 9345 curnpgs = TTEPAGES(cursz); 9346 if (tst_tnc(pp2, curnpgs)) { 9347 SFMMU_STAT_ADD(sf_recache, curnpgs); 9348 sfmmu_page_cache_array(pp2, HAT_CACHE, CACHE_NO_FLUSH, 9349 curnpgs); 9350 } 9351 9352 ASSERT(dopgs >= curnpgs); 9353 dopgs -= curnpgs; 9354 9355 if (dopgs == 0) { 9356 break; 9357 } 9358 9359 pp2 = PP_PAGENEXT_N(pp2, curnpgs); 9360 if (((dopgs & (pg64k - 1)) == 0) && PP_ISMAPPED_LARGE(pp2)) { 9361 cursz = fnd_mapping_sz(pp2); 9362 } else { 9363 cursz = TTE8K; 9364 } 9365 } 9366 } 9367 9368 /* 9369 * Returns 1 if page(s) can be converted from TNC to cacheable setting, 9370 * returns 0 otherwise. Note that oaddr argument is valid for only 9371 * 8k pages. 9372 */ 9373 int 9374 tst_tnc(page_t *pp, pgcnt_t npages) 9375 { 9376 struct sf_hment *sfhme; 9377 struct hme_blk *hmeblkp; 9378 tte_t tte; 9379 caddr_t vaddr; 9380 int clr_valid = 0; 9381 int color, color1, bcolor; 9382 int i, ncolors; 9383 9384 ASSERT(pp != NULL); 9385 ASSERT(!(cache & CACHE_WRITEBACK)); 9386 9387 if (npages > 1) { 9388 ncolors = CACHE_NUM_COLOR; 9389 } 9390 9391 for (i = 0; i < npages; i++) { 9392 ASSERT(sfmmu_mlist_held(pp)); 9393 ASSERT(PP_ISTNC(pp)); 9394 ASSERT(PP_GET_VCOLOR(pp) == NO_VCOLOR); 9395 9396 if (PP_ISPNC(pp)) { 9397 return (0); 9398 } 9399 9400 clr_valid = 0; 9401 if (PP_ISMAPPED_KPM(pp)) { 9402 caddr_t kpmvaddr; 9403 9404 ASSERT(kpm_enable); 9405 kpmvaddr = hat_kpm_page2va(pp, 1); 9406 ASSERT(!(npages > 1 && IS_KPM_ALIAS_RANGE(kpmvaddr))); 9407 color1 = addr_to_vcolor(kpmvaddr); 9408 clr_valid = 1; 9409 } 9410 9411 for (sfhme = pp->p_mapping; sfhme; sfhme = sfhme->hme_next) { 9412 if (IS_PAHME(sfhme)) 9413 continue; 9414 hmeblkp = sfmmu_hmetohblk(sfhme); 9415 if (hmeblkp->hblk_xhat_bit) 9416 continue; 9417 9418 sfmmu_copytte(&sfhme->hme_tte, &tte); 9419 ASSERT(TTE_IS_VALID(&tte)); 9420 9421 vaddr = tte_to_vaddr(hmeblkp, tte); 9422 color = addr_to_vcolor(vaddr); 9423 9424 if (npages > 1) { 9425 /* 9426 * If there is a big mapping, make sure 9427 * 8K mapping is consistent with the big 9428 * mapping. 9429 */ 9430 bcolor = i % ncolors; 9431 if (color != bcolor) { 9432 return (0); 9433 } 9434 } 9435 if (!clr_valid) { 9436 clr_valid = 1; 9437 color1 = color; 9438 } 9439 9440 if (color1 != color) { 9441 return (0); 9442 } 9443 } 9444 9445 pp = PP_PAGENEXT(pp); 9446 } 9447 9448 return (1); 9449 } 9450 9451 void 9452 sfmmu_page_cache_array(page_t *pp, int flags, int cache_flush_flag, 9453 pgcnt_t npages) 9454 { 9455 kmutex_t *pmtx; 9456 int i, ncolors, bcolor; 9457 kpm_hlk_t *kpmp; 9458 cpuset_t cpuset; 9459 9460 ASSERT(pp != NULL); 9461 ASSERT(!(cache & CACHE_WRITEBACK)); 9462 9463 kpmp = sfmmu_kpm_kpmp_enter(pp, npages); 9464 pmtx = sfmmu_page_enter(pp); 9465 9466 /* 9467 * Fast path caching single unmapped page 9468 */ 9469 if (npages == 1 && !PP_ISMAPPED(pp) && !PP_ISMAPPED_KPM(pp) && 9470 flags == HAT_CACHE) { 9471 PP_CLRTNC(pp); 9472 PP_CLRPNC(pp); 9473 sfmmu_page_exit(pmtx); 9474 sfmmu_kpm_kpmp_exit(kpmp); 9475 return; 9476 } 9477 9478 /* 9479 * We need to capture all cpus in order to change cacheability 9480 * because we can't allow one cpu to access the same physical 9481 * page using a cacheable and a non-cachebale mapping at the same 9482 * time. Since we may end up walking the ism mapping list 9483 * have to grab it's lock now since we can't after all the 9484 * cpus have been captured. 9485 */ 9486 sfmmu_hat_lock_all(); 9487 mutex_enter(&ism_mlist_lock); 9488 kpreempt_disable(); 9489 cpuset = cpu_ready_set; 9490 xc_attention(cpuset); 9491 9492 if (npages > 1) { 9493 /* 9494 * Make sure all colors are flushed since the 9495 * sfmmu_page_cache() only flushes one color- 9496 * it does not know big pages. 9497 */ 9498 ncolors = CACHE_NUM_COLOR; 9499 if (flags & HAT_TMPNC) { 9500 for (i = 0; i < ncolors; i++) { 9501 sfmmu_cache_flushcolor(i, pp->p_pagenum); 9502 } 9503 cache_flush_flag = CACHE_NO_FLUSH; 9504 } 9505 } 9506 9507 for (i = 0; i < npages; i++) { 9508 9509 ASSERT(sfmmu_mlist_held(pp)); 9510 9511 if (!(flags == HAT_TMPNC && PP_ISTNC(pp))) { 9512 9513 if (npages > 1) { 9514 bcolor = i % ncolors; 9515 } else { 9516 bcolor = NO_VCOLOR; 9517 } 9518 9519 sfmmu_page_cache(pp, flags, cache_flush_flag, 9520 bcolor); 9521 } 9522 9523 pp = PP_PAGENEXT(pp); 9524 } 9525 9526 xt_sync(cpuset); 9527 xc_dismissed(cpuset); 9528 mutex_exit(&ism_mlist_lock); 9529 sfmmu_hat_unlock_all(); 9530 sfmmu_page_exit(pmtx); 9531 sfmmu_kpm_kpmp_exit(kpmp); 9532 kpreempt_enable(); 9533 } 9534 9535 /* 9536 * This function changes the virtual cacheability of all mappings to a 9537 * particular page. When changing from uncache to cacheable the mappings will 9538 * only be changed if all of them have the same virtual color. 9539 * We need to flush the cache in all cpus. It is possible that 9540 * a process referenced a page as cacheable but has sinced exited 9541 * and cleared the mapping list. We still to flush it but have no 9542 * state so all cpus is the only alternative. 9543 */ 9544 static void 9545 sfmmu_page_cache(page_t *pp, int flags, int cache_flush_flag, int bcolor) 9546 { 9547 struct sf_hment *sfhme; 9548 struct hme_blk *hmeblkp; 9549 sfmmu_t *sfmmup; 9550 tte_t tte, ttemod; 9551 caddr_t vaddr; 9552 int ret, color; 9553 pfn_t pfn; 9554 9555 color = bcolor; 9556 pfn = pp->p_pagenum; 9557 9558 for (sfhme = pp->p_mapping; sfhme; sfhme = sfhme->hme_next) { 9559 9560 if (IS_PAHME(sfhme)) 9561 continue; 9562 hmeblkp = sfmmu_hmetohblk(sfhme); 9563 9564 if (hmeblkp->hblk_xhat_bit) 9565 continue; 9566 9567 sfmmu_copytte(&sfhme->hme_tte, &tte); 9568 ASSERT(TTE_IS_VALID(&tte)); 9569 vaddr = tte_to_vaddr(hmeblkp, tte); 9570 color = addr_to_vcolor(vaddr); 9571 9572 #ifdef DEBUG 9573 if ((flags & HAT_CACHE) && bcolor != NO_VCOLOR) { 9574 ASSERT(color == bcolor); 9575 } 9576 #endif 9577 9578 ASSERT(flags != HAT_TMPNC || color == PP_GET_VCOLOR(pp)); 9579 9580 ttemod = tte; 9581 if (flags & (HAT_UNCACHE | HAT_TMPNC)) { 9582 TTE_CLR_VCACHEABLE(&ttemod); 9583 } else { /* flags & HAT_CACHE */ 9584 TTE_SET_VCACHEABLE(&ttemod); 9585 } 9586 ret = sfmmu_modifytte_try(&tte, &ttemod, &sfhme->hme_tte); 9587 if (ret < 0) { 9588 /* 9589 * Since all cpus are captured modifytte should not 9590 * fail. 9591 */ 9592 panic("sfmmu_page_cache: write to tte failed"); 9593 } 9594 9595 sfmmup = hblktosfmmu(hmeblkp); 9596 if (cache_flush_flag == CACHE_FLUSH) { 9597 /* 9598 * Flush TSBs, TLBs and caches 9599 */ 9600 if (hmeblkp->hblk_shared) { 9601 sf_srd_t *srdp = (sf_srd_t *)sfmmup; 9602 uint_t rid = hmeblkp->hblk_tag.htag_rid; 9603 sf_region_t *rgnp; 9604 ASSERT(SFMMU_IS_SHMERID_VALID(rid)); 9605 ASSERT(rid < SFMMU_MAX_HME_REGIONS); 9606 ASSERT(srdp != NULL); 9607 rgnp = srdp->srd_hmergnp[rid]; 9608 SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, 9609 srdp, rgnp, rid); 9610 (void) sfmmu_rgntlb_demap(vaddr, rgnp, 9611 hmeblkp, 0); 9612 sfmmu_cache_flush(pfn, addr_to_vcolor(vaddr)); 9613 } else if (sfmmup->sfmmu_ismhat) { 9614 if (flags & HAT_CACHE) { 9615 SFMMU_STAT(sf_ism_recache); 9616 } else { 9617 SFMMU_STAT(sf_ism_uncache); 9618 } 9619 sfmmu_ismtlbcache_demap(vaddr, sfmmup, hmeblkp, 9620 pfn, CACHE_FLUSH); 9621 } else { 9622 sfmmu_tlbcache_demap(vaddr, sfmmup, hmeblkp, 9623 pfn, 0, FLUSH_ALL_CPUS, CACHE_FLUSH, 1); 9624 } 9625 9626 /* 9627 * all cache entries belonging to this pfn are 9628 * now flushed. 9629 */ 9630 cache_flush_flag = CACHE_NO_FLUSH; 9631 } else { 9632 /* 9633 * Flush only TSBs and TLBs. 9634 */ 9635 if (hmeblkp->hblk_shared) { 9636 sf_srd_t *srdp = (sf_srd_t *)sfmmup; 9637 uint_t rid = hmeblkp->hblk_tag.htag_rid; 9638 sf_region_t *rgnp; 9639 ASSERT(SFMMU_IS_SHMERID_VALID(rid)); 9640 ASSERT(rid < SFMMU_MAX_HME_REGIONS); 9641 ASSERT(srdp != NULL); 9642 rgnp = srdp->srd_hmergnp[rid]; 9643 SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, 9644 srdp, rgnp, rid); 9645 (void) sfmmu_rgntlb_demap(vaddr, rgnp, 9646 hmeblkp, 0); 9647 } else if (sfmmup->sfmmu_ismhat) { 9648 if (flags & HAT_CACHE) { 9649 SFMMU_STAT(sf_ism_recache); 9650 } else { 9651 SFMMU_STAT(sf_ism_uncache); 9652 } 9653 sfmmu_ismtlbcache_demap(vaddr, sfmmup, hmeblkp, 9654 pfn, CACHE_NO_FLUSH); 9655 } else { 9656 sfmmu_tlb_demap(vaddr, sfmmup, hmeblkp, 0, 1); 9657 } 9658 } 9659 } 9660 9661 if (PP_ISMAPPED_KPM(pp)) 9662 sfmmu_kpm_page_cache(pp, flags, cache_flush_flag); 9663 9664 switch (flags) { 9665 9666 default: 9667 panic("sfmmu_pagecache: unknown flags"); 9668 break; 9669 9670 case HAT_CACHE: 9671 PP_CLRTNC(pp); 9672 PP_CLRPNC(pp); 9673 PP_SET_VCOLOR(pp, color); 9674 break; 9675 9676 case HAT_TMPNC: 9677 PP_SETTNC(pp); 9678 PP_SET_VCOLOR(pp, NO_VCOLOR); 9679 break; 9680 9681 case HAT_UNCACHE: 9682 PP_SETPNC(pp); 9683 PP_CLRTNC(pp); 9684 PP_SET_VCOLOR(pp, NO_VCOLOR); 9685 break; 9686 } 9687 } 9688 #endif /* VAC */ 9689 9690 9691 /* 9692 * Wrapper routine used to return a context. 9693 * 9694 * It's the responsibility of the caller to guarantee that the 9695 * process serializes on calls here by taking the HAT lock for 9696 * the hat. 9697 * 9698 */ 9699 static void 9700 sfmmu_get_ctx(sfmmu_t *sfmmup) 9701 { 9702 mmu_ctx_t *mmu_ctxp; 9703 uint_t pstate_save; 9704 int ret; 9705 9706 ASSERT(sfmmu_hat_lock_held(sfmmup)); 9707 ASSERT(sfmmup != ksfmmup); 9708 9709 if (SFMMU_FLAGS_ISSET(sfmmup, HAT_ALLCTX_INVALID)) { 9710 sfmmu_setup_tsbinfo(sfmmup); 9711 SFMMU_FLAGS_CLEAR(sfmmup, HAT_ALLCTX_INVALID); 9712 } 9713 9714 kpreempt_disable(); 9715 9716 mmu_ctxp = CPU_MMU_CTXP(CPU); 9717 ASSERT(mmu_ctxp); 9718 ASSERT(mmu_ctxp->mmu_idx < max_mmu_ctxdoms); 9719 ASSERT(mmu_ctxp == mmu_ctxs_tbl[mmu_ctxp->mmu_idx]); 9720 9721 /* 9722 * Do a wrap-around if cnum reaches the max # cnum supported by a MMU. 9723 */ 9724 if (mmu_ctxp->mmu_cnum == mmu_ctxp->mmu_nctxs) 9725 sfmmu_ctx_wrap_around(mmu_ctxp); 9726 9727 /* 9728 * Let the MMU set up the page sizes to use for 9729 * this context in the TLB. Don't program 2nd dtlb for ism hat. 9730 */ 9731 if ((&mmu_set_ctx_page_sizes) && (sfmmup->sfmmu_ismhat == 0)) { 9732 mmu_set_ctx_page_sizes(sfmmup); 9733 } 9734 9735 /* 9736 * sfmmu_alloc_ctx and sfmmu_load_mmustate will be performed with 9737 * interrupts disabled to prevent race condition with wrap-around 9738 * ctx invalidatation. In sun4v, ctx invalidation also involves 9739 * a HV call to set the number of TSBs to 0. If interrupts are not 9740 * disabled until after sfmmu_load_mmustate is complete TSBs may 9741 * become assigned to INVALID_CONTEXT. This is not allowed. 9742 */ 9743 pstate_save = sfmmu_disable_intrs(); 9744 9745 if (sfmmu_alloc_ctx(sfmmup, 1, CPU, SFMMU_PRIVATE) && 9746 sfmmup->sfmmu_scdp != NULL) { 9747 sf_scd_t *scdp = sfmmup->sfmmu_scdp; 9748 sfmmu_t *scsfmmup = scdp->scd_sfmmup; 9749 ret = sfmmu_alloc_ctx(scsfmmup, 1, CPU, SFMMU_SHARED); 9750 /* debug purpose only */ 9751 ASSERT(!ret || scsfmmup->sfmmu_ctxs[CPU_MMU_IDX(CPU)].cnum 9752 != INVALID_CONTEXT); 9753 } 9754 sfmmu_load_mmustate(sfmmup); 9755 9756 sfmmu_enable_intrs(pstate_save); 9757 9758 kpreempt_enable(); 9759 } 9760 9761 /* 9762 * When all cnums are used up in a MMU, cnum will wrap around to the 9763 * next generation and start from 2. 9764 */ 9765 static void 9766 sfmmu_ctx_wrap_around(mmu_ctx_t *mmu_ctxp) 9767 { 9768 9769 /* caller must have disabled the preemption */ 9770 ASSERT(curthread->t_preempt >= 1); 9771 ASSERT(mmu_ctxp != NULL); 9772 9773 /* acquire Per-MMU (PM) spin lock */ 9774 mutex_enter(&mmu_ctxp->mmu_lock); 9775 9776 /* re-check to see if wrap-around is needed */ 9777 if (mmu_ctxp->mmu_cnum < mmu_ctxp->mmu_nctxs) 9778 goto done; 9779 9780 SFMMU_MMU_STAT(mmu_wrap_around); 9781 9782 /* update gnum */ 9783 ASSERT(mmu_ctxp->mmu_gnum != 0); 9784 mmu_ctxp->mmu_gnum++; 9785 if (mmu_ctxp->mmu_gnum == 0 || 9786 mmu_ctxp->mmu_gnum > MAX_SFMMU_GNUM_VAL) { 9787 cmn_err(CE_PANIC, "mmu_gnum of mmu_ctx 0x%p is out of bound.", 9788 (void *)mmu_ctxp); 9789 } 9790 9791 if (mmu_ctxp->mmu_ncpus > 1) { 9792 cpuset_t cpuset; 9793 9794 membar_enter(); /* make sure updated gnum visible */ 9795 9796 SFMMU_XCALL_STATS(NULL); 9797 9798 /* xcall to others on the same MMU to invalidate ctx */ 9799 cpuset = mmu_ctxp->mmu_cpuset; 9800 ASSERT(CPU_IN_SET(cpuset, CPU->cpu_id)); 9801 CPUSET_DEL(cpuset, CPU->cpu_id); 9802 CPUSET_AND(cpuset, cpu_ready_set); 9803 9804 /* 9805 * Pass in INVALID_CONTEXT as the first parameter to 9806 * sfmmu_raise_tsb_exception, which invalidates the context 9807 * of any process running on the CPUs in the MMU. 9808 */ 9809 xt_some(cpuset, sfmmu_raise_tsb_exception, 9810 INVALID_CONTEXT, INVALID_CONTEXT); 9811 xt_sync(cpuset); 9812 9813 SFMMU_MMU_STAT(mmu_tsb_raise_exception); 9814 } 9815 9816 if (sfmmu_getctx_sec() != INVALID_CONTEXT) { 9817 sfmmu_setctx_sec(INVALID_CONTEXT); 9818 sfmmu_clear_utsbinfo(); 9819 } 9820 9821 /* 9822 * No xcall is needed here. For sun4u systems all CPUs in context 9823 * domain share a single physical MMU therefore it's enough to flush 9824 * TLB on local CPU. On sun4v systems we use 1 global context 9825 * domain and flush all remote TLBs in sfmmu_raise_tsb_exception 9826 * handler. Note that vtag_flushall_uctxs() is called 9827 * for Ultra II machine, where the equivalent flushall functionality 9828 * is implemented in SW, and only user ctx TLB entries are flushed. 9829 */ 9830 if (&vtag_flushall_uctxs != NULL) { 9831 vtag_flushall_uctxs(); 9832 } else { 9833 vtag_flushall(); 9834 } 9835 9836 /* reset mmu cnum, skips cnum 0 and 1 */ 9837 mmu_ctxp->mmu_cnum = NUM_LOCKED_CTXS; 9838 9839 done: 9840 mutex_exit(&mmu_ctxp->mmu_lock); 9841 } 9842 9843 9844 /* 9845 * For multi-threaded process, set the process context to INVALID_CONTEXT 9846 * so that it faults and reloads the MMU state from TL=0. For single-threaded 9847 * process, we can just load the MMU state directly without having to 9848 * set context invalid. Caller must hold the hat lock since we don't 9849 * acquire it here. 9850 */ 9851 static void 9852 sfmmu_sync_mmustate(sfmmu_t *sfmmup) 9853 { 9854 uint_t cnum; 9855 uint_t pstate_save; 9856 9857 ASSERT(sfmmup != ksfmmup); 9858 ASSERT(sfmmu_hat_lock_held(sfmmup)); 9859 9860 kpreempt_disable(); 9861 9862 /* 9863 * We check whether the pass'ed-in sfmmup is the same as the 9864 * current running proc. This is to makes sure the current proc 9865 * stays single-threaded if it already is. 9866 */ 9867 if ((sfmmup == curthread->t_procp->p_as->a_hat) && 9868 (curthread->t_procp->p_lwpcnt == 1)) { 9869 /* single-thread */ 9870 cnum = sfmmup->sfmmu_ctxs[CPU_MMU_IDX(CPU)].cnum; 9871 if (cnum != INVALID_CONTEXT) { 9872 uint_t curcnum; 9873 /* 9874 * Disable interrupts to prevent race condition 9875 * with sfmmu_ctx_wrap_around ctx invalidation. 9876 * In sun4v, ctx invalidation involves setting 9877 * TSB to NULL, hence, interrupts should be disabled 9878 * untill after sfmmu_load_mmustate is completed. 9879 */ 9880 pstate_save = sfmmu_disable_intrs(); 9881 curcnum = sfmmu_getctx_sec(); 9882 if (curcnum == cnum) 9883 sfmmu_load_mmustate(sfmmup); 9884 sfmmu_enable_intrs(pstate_save); 9885 ASSERT(curcnum == cnum || curcnum == INVALID_CONTEXT); 9886 } 9887 } else { 9888 /* 9889 * multi-thread 9890 * or when sfmmup is not the same as the curproc. 9891 */ 9892 sfmmu_invalidate_ctx(sfmmup); 9893 } 9894 9895 kpreempt_enable(); 9896 } 9897 9898 9899 /* 9900 * Replace the specified TSB with a new TSB. This function gets called when 9901 * we grow, shrink or swapin a TSB. When swapping in a TSB (TSB_SWAPIN), the 9902 * TSB_FORCEALLOC flag may be used to force allocation of a minimum-sized TSB 9903 * (8K). 9904 * 9905 * Caller must hold the HAT lock, but should assume any tsb_info 9906 * pointers it has are no longer valid after calling this function. 9907 * 9908 * Return values: 9909 * TSB_ALLOCFAIL Failed to allocate a TSB, due to memory constraints 9910 * TSB_LOSTRACE HAT is busy, i.e. another thread is already doing 9911 * something to this tsbinfo/TSB 9912 * TSB_SUCCESS Operation succeeded 9913 */ 9914 static tsb_replace_rc_t 9915 sfmmu_replace_tsb(sfmmu_t *sfmmup, struct tsb_info *old_tsbinfo, uint_t szc, 9916 hatlock_t *hatlockp, uint_t flags) 9917 { 9918 struct tsb_info *new_tsbinfo = NULL; 9919 struct tsb_info *curtsb, *prevtsb; 9920 uint_t tte_sz_mask; 9921 int i; 9922 9923 ASSERT(sfmmup != ksfmmup); 9924 ASSERT(sfmmup->sfmmu_ismhat == 0); 9925 ASSERT(sfmmu_hat_lock_held(sfmmup)); 9926 ASSERT(szc <= tsb_max_growsize); 9927 9928 if (SFMMU_FLAGS_ISSET(sfmmup, HAT_BUSY)) 9929 return (TSB_LOSTRACE); 9930 9931 /* 9932 * Find the tsb_info ahead of this one in the list, and 9933 * also make sure that the tsb_info passed in really 9934 * exists! 9935 */ 9936 for (prevtsb = NULL, curtsb = sfmmup->sfmmu_tsb; 9937 curtsb != old_tsbinfo && curtsb != NULL; 9938 prevtsb = curtsb, curtsb = curtsb->tsb_next) 9939 ; 9940 ASSERT(curtsb != NULL); 9941 9942 if (!(flags & TSB_SWAPIN) && SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) { 9943 /* 9944 * The process is swapped out, so just set the new size 9945 * code. When it swaps back in, we'll allocate a new one 9946 * of the new chosen size. 9947 */ 9948 curtsb->tsb_szc = szc; 9949 return (TSB_SUCCESS); 9950 } 9951 SFMMU_FLAGS_SET(sfmmup, HAT_BUSY); 9952 9953 tte_sz_mask = old_tsbinfo->tsb_ttesz_mask; 9954 9955 /* 9956 * All initialization is done inside of sfmmu_tsbinfo_alloc(). 9957 * If we fail to allocate a TSB, exit. 9958 * 9959 * If tsb grows with new tsb size > 4M and old tsb size < 4M, 9960 * then try 4M slab after the initial alloc fails. 9961 * 9962 * If tsb swapin with tsb size > 4M, then try 4M after the 9963 * initial alloc fails. 9964 */ 9965 sfmmu_hat_exit(hatlockp); 9966 if (sfmmu_tsbinfo_alloc(&new_tsbinfo, szc, 9967 tte_sz_mask, flags, sfmmup) && 9968 (!(flags & (TSB_GROW | TSB_SWAPIN)) || (szc <= TSB_4M_SZCODE) || 9969 (!(flags & TSB_SWAPIN) && 9970 (old_tsbinfo->tsb_szc >= TSB_4M_SZCODE)) || 9971 sfmmu_tsbinfo_alloc(&new_tsbinfo, TSB_4M_SZCODE, 9972 tte_sz_mask, flags, sfmmup))) { 9973 (void) sfmmu_hat_enter(sfmmup); 9974 if (!(flags & TSB_SWAPIN)) 9975 SFMMU_STAT(sf_tsb_resize_failures); 9976 SFMMU_FLAGS_CLEAR(sfmmup, HAT_BUSY); 9977 return (TSB_ALLOCFAIL); 9978 } 9979 (void) sfmmu_hat_enter(sfmmup); 9980 9981 /* 9982 * Re-check to make sure somebody else didn't muck with us while we 9983 * didn't hold the HAT lock. If the process swapped out, fine, just 9984 * exit; this can happen if we try to shrink the TSB from the context 9985 * of another process (such as on an ISM unmap), though it is rare. 9986 */ 9987 if (!(flags & TSB_SWAPIN) && SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) { 9988 SFMMU_STAT(sf_tsb_resize_failures); 9989 SFMMU_FLAGS_CLEAR(sfmmup, HAT_BUSY); 9990 sfmmu_hat_exit(hatlockp); 9991 sfmmu_tsbinfo_free(new_tsbinfo); 9992 (void) sfmmu_hat_enter(sfmmup); 9993 return (TSB_LOSTRACE); 9994 } 9995 9996 #ifdef DEBUG 9997 /* Reverify that the tsb_info still exists.. for debugging only */ 9998 for (prevtsb = NULL, curtsb = sfmmup->sfmmu_tsb; 9999 curtsb != old_tsbinfo && curtsb != NULL; 10000 prevtsb = curtsb, curtsb = curtsb->tsb_next) 10001 ; 10002 ASSERT(curtsb != NULL); 10003 #endif /* DEBUG */ 10004 10005 /* 10006 * Quiesce any CPUs running this process on their next TLB miss 10007 * so they atomically see the new tsb_info. We temporarily set the 10008 * context to invalid context so new threads that come on processor 10009 * after we do the xcall to cpusran will also serialize behind the 10010 * HAT lock on TLB miss and will see the new TSB. Since this short 10011 * race with a new thread coming on processor is relatively rare, 10012 * this synchronization mechanism should be cheaper than always 10013 * pausing all CPUs for the duration of the setup, which is what 10014 * the old implementation did. This is particuarly true if we are 10015 * copying a huge chunk of memory around during that window. 10016 * 10017 * The memory barriers are to make sure things stay consistent 10018 * with resume() since it does not hold the HAT lock while 10019 * walking the list of tsb_info structures. 10020 */ 10021 if ((flags & TSB_SWAPIN) != TSB_SWAPIN) { 10022 /* The TSB is either growing or shrinking. */ 10023 sfmmu_invalidate_ctx(sfmmup); 10024 } else { 10025 /* 10026 * It is illegal to swap in TSBs from a process other 10027 * than a process being swapped in. This in turn 10028 * implies we do not have a valid MMU context here 10029 * since a process needs one to resolve translation 10030 * misses. 10031 */ 10032 ASSERT(curthread->t_procp->p_as->a_hat == sfmmup); 10033 } 10034 10035 #ifdef DEBUG 10036 ASSERT(max_mmu_ctxdoms > 0); 10037 10038 /* 10039 * Process should have INVALID_CONTEXT on all MMUs 10040 */ 10041 for (i = 0; i < max_mmu_ctxdoms; i++) { 10042 10043 ASSERT(sfmmup->sfmmu_ctxs[i].cnum == INVALID_CONTEXT); 10044 } 10045 #endif 10046 10047 new_tsbinfo->tsb_next = old_tsbinfo->tsb_next; 10048 membar_stst(); /* strict ordering required */ 10049 if (prevtsb) 10050 prevtsb->tsb_next = new_tsbinfo; 10051 else 10052 sfmmup->sfmmu_tsb = new_tsbinfo; 10053 membar_enter(); /* make sure new TSB globally visible */ 10054 10055 /* 10056 * We need to migrate TSB entries from the old TSB to the new TSB 10057 * if tsb_remap_ttes is set and the TSB is growing. 10058 */ 10059 if (tsb_remap_ttes && ((flags & TSB_GROW) == TSB_GROW)) 10060 sfmmu_copy_tsb(old_tsbinfo, new_tsbinfo); 10061 10062 SFMMU_FLAGS_CLEAR(sfmmup, HAT_BUSY); 10063 10064 /* 10065 * Drop the HAT lock to free our old tsb_info. 10066 */ 10067 sfmmu_hat_exit(hatlockp); 10068 10069 if ((flags & TSB_GROW) == TSB_GROW) { 10070 SFMMU_STAT(sf_tsb_grow); 10071 } else if ((flags & TSB_SHRINK) == TSB_SHRINK) { 10072 SFMMU_STAT(sf_tsb_shrink); 10073 } 10074 10075 sfmmu_tsbinfo_free(old_tsbinfo); 10076 10077 (void) sfmmu_hat_enter(sfmmup); 10078 return (TSB_SUCCESS); 10079 } 10080 10081 /* 10082 * This function will re-program hat pgsz array, and invalidate the 10083 * process' context, forcing the process to switch to another 10084 * context on the next TLB miss, and therefore start using the 10085 * TLB that is reprogrammed for the new page sizes. 10086 */ 10087 void 10088 sfmmu_reprog_pgsz_arr(sfmmu_t *sfmmup, uint8_t *tmp_pgsz) 10089 { 10090 int i; 10091 hatlock_t *hatlockp = NULL; 10092 10093 hatlockp = sfmmu_hat_enter(sfmmup); 10094 /* USIII+-IV+ optimization, requires hat lock */ 10095 if (tmp_pgsz) { 10096 for (i = 0; i < mmu_page_sizes; i++) 10097 sfmmup->sfmmu_pgsz[i] = tmp_pgsz[i]; 10098 } 10099 SFMMU_STAT(sf_tlb_reprog_pgsz); 10100 10101 sfmmu_invalidate_ctx(sfmmup); 10102 10103 sfmmu_hat_exit(hatlockp); 10104 } 10105 10106 /* 10107 * The scd_rttecnt field in the SCD must be updated to take account of the 10108 * regions which it contains. 10109 */ 10110 static void 10111 sfmmu_set_scd_rttecnt(sf_srd_t *srdp, sf_scd_t *scdp) 10112 { 10113 uint_t rid; 10114 uint_t i, j; 10115 ulong_t w; 10116 sf_region_t *rgnp; 10117 10118 ASSERT(srdp != NULL); 10119 10120 for (i = 0; i < SFMMU_HMERGNMAP_WORDS; i++) { 10121 if ((w = scdp->scd_region_map.bitmap[i]) == 0) { 10122 continue; 10123 } 10124 10125 j = 0; 10126 while (w) { 10127 if (!(w & 0x1)) { 10128 j++; 10129 w >>= 1; 10130 continue; 10131 } 10132 rid = (i << BT_ULSHIFT) | j; 10133 j++; 10134 w >>= 1; 10135 10136 ASSERT(SFMMU_IS_SHMERID_VALID(rid)); 10137 ASSERT(rid < SFMMU_MAX_HME_REGIONS); 10138 rgnp = srdp->srd_hmergnp[rid]; 10139 ASSERT(rgnp->rgn_refcnt > 0); 10140 ASSERT(rgnp->rgn_id == rid); 10141 10142 scdp->scd_rttecnt[rgnp->rgn_pgszc] += 10143 rgnp->rgn_size >> TTE_PAGE_SHIFT(rgnp->rgn_pgszc); 10144 10145 /* 10146 * Maintain the tsb0 inflation cnt for the regions 10147 * in the SCD. 10148 */ 10149 if (rgnp->rgn_pgszc >= TTE4M) { 10150 scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt += 10151 rgnp->rgn_size >> 10152 (TTE_PAGE_SHIFT(TTE8K) + 2); 10153 } 10154 } 10155 } 10156 } 10157 10158 /* 10159 * This function assumes that there are either four or six supported page 10160 * sizes and at most two programmable TLBs, so we need to decide which 10161 * page sizes are most important and then tell the MMU layer so it 10162 * can adjust the TLB page sizes accordingly (if supported). 10163 * 10164 * If these assumptions change, this function will need to be 10165 * updated to support whatever the new limits are. 10166 * 10167 * The growing flag is nonzero if we are growing the address space, 10168 * and zero if it is shrinking. This allows us to decide whether 10169 * to grow or shrink our TSB, depending upon available memory 10170 * conditions. 10171 */ 10172 static void 10173 sfmmu_check_page_sizes(sfmmu_t *sfmmup, int growing) 10174 { 10175 uint64_t ttecnt[MMU_PAGE_SIZES]; 10176 uint64_t tte8k_cnt, tte4m_cnt; 10177 uint8_t i; 10178 int sectsb_thresh; 10179 10180 /* 10181 * Kernel threads, processes with small address spaces not using 10182 * large pages, and dummy ISM HATs need not apply. 10183 */ 10184 if (sfmmup == ksfmmup || sfmmup->sfmmu_ismhat != NULL) 10185 return; 10186 10187 if (!SFMMU_LGPGS_INUSE(sfmmup) && 10188 sfmmup->sfmmu_ttecnt[TTE8K] <= tsb_rss_factor) 10189 return; 10190 10191 for (i = 0; i < mmu_page_sizes; i++) { 10192 ttecnt[i] = sfmmup->sfmmu_ttecnt[i] + 10193 sfmmup->sfmmu_ismttecnt[i]; 10194 } 10195 10196 /* Check pagesizes in use, and possibly reprogram DTLB. */ 10197 if (&mmu_check_page_sizes) 10198 mmu_check_page_sizes(sfmmup, ttecnt); 10199 10200 /* 10201 * Calculate the number of 8k ttes to represent the span of these 10202 * pages. 10203 */ 10204 tte8k_cnt = ttecnt[TTE8K] + 10205 (ttecnt[TTE64K] << (MMU_PAGESHIFT64K - MMU_PAGESHIFT)) + 10206 (ttecnt[TTE512K] << (MMU_PAGESHIFT512K - MMU_PAGESHIFT)); 10207 if (mmu_page_sizes == max_mmu_page_sizes) { 10208 tte4m_cnt = ttecnt[TTE4M] + 10209 (ttecnt[TTE32M] << (MMU_PAGESHIFT32M - MMU_PAGESHIFT4M)) + 10210 (ttecnt[TTE256M] << (MMU_PAGESHIFT256M - MMU_PAGESHIFT4M)); 10211 } else { 10212 tte4m_cnt = ttecnt[TTE4M]; 10213 } 10214 10215 /* 10216 * Inflate tte8k_cnt to allow for region large page allocation failure. 10217 */ 10218 tte8k_cnt += sfmmup->sfmmu_tsb0_4minflcnt; 10219 10220 /* 10221 * Inflate TSB sizes by a factor of 2 if this process 10222 * uses 4M text pages to minimize extra conflict misses 10223 * in the first TSB since without counting text pages 10224 * 8K TSB may become too small. 10225 * 10226 * Also double the size of the second TSB to minimize 10227 * extra conflict misses due to competition between 4M text pages 10228 * and data pages. 10229 * 10230 * We need to adjust the second TSB allocation threshold by the 10231 * inflation factor, since there is no point in creating a second 10232 * TSB when we know all the mappings can fit in the I/D TLBs. 10233 */ 10234 sectsb_thresh = tsb_sectsb_threshold; 10235 if (sfmmup->sfmmu_flags & HAT_4MTEXT_FLAG) { 10236 tte8k_cnt <<= 1; 10237 tte4m_cnt <<= 1; 10238 sectsb_thresh <<= 1; 10239 } 10240 10241 /* 10242 * Check to see if our TSB is the right size; we may need to 10243 * grow or shrink it. If the process is small, our work is 10244 * finished at this point. 10245 */ 10246 if (tte8k_cnt <= tsb_rss_factor && tte4m_cnt <= sectsb_thresh) { 10247 return; 10248 } 10249 sfmmu_size_tsb(sfmmup, growing, tte8k_cnt, tte4m_cnt, sectsb_thresh); 10250 } 10251 10252 static void 10253 sfmmu_size_tsb(sfmmu_t *sfmmup, int growing, uint64_t tte8k_cnt, 10254 uint64_t tte4m_cnt, int sectsb_thresh) 10255 { 10256 int tsb_bits; 10257 uint_t tsb_szc; 10258 struct tsb_info *tsbinfop; 10259 hatlock_t *hatlockp = NULL; 10260 10261 hatlockp = sfmmu_hat_enter(sfmmup); 10262 ASSERT(hatlockp != NULL); 10263 tsbinfop = sfmmup->sfmmu_tsb; 10264 ASSERT(tsbinfop != NULL); 10265 10266 /* 10267 * If we're growing, select the size based on RSS. If we're 10268 * shrinking, leave some room so we don't have to turn around and 10269 * grow again immediately. 10270 */ 10271 if (growing) 10272 tsb_szc = SELECT_TSB_SIZECODE(tte8k_cnt); 10273 else 10274 tsb_szc = SELECT_TSB_SIZECODE(tte8k_cnt << 1); 10275 10276 if (!growing && (tsb_szc < tsbinfop->tsb_szc) && 10277 (tsb_szc >= default_tsb_size) && TSB_OK_SHRINK()) { 10278 (void) sfmmu_replace_tsb(sfmmup, tsbinfop, tsb_szc, 10279 hatlockp, TSB_SHRINK); 10280 } else if (growing && tsb_szc > tsbinfop->tsb_szc && TSB_OK_GROW()) { 10281 (void) sfmmu_replace_tsb(sfmmup, tsbinfop, tsb_szc, 10282 hatlockp, TSB_GROW); 10283 } 10284 tsbinfop = sfmmup->sfmmu_tsb; 10285 10286 /* 10287 * With the TLB and first TSB out of the way, we need to see if 10288 * we need a second TSB for 4M pages. If we managed to reprogram 10289 * the TLB page sizes above, the process will start using this new 10290 * TSB right away; otherwise, it will start using it on the next 10291 * context switch. Either way, it's no big deal so there's no 10292 * synchronization with the trap handlers here unless we grow the 10293 * TSB (in which case it's required to prevent using the old one 10294 * after it's freed). Note: second tsb is required for 32M/256M 10295 * page sizes. 10296 */ 10297 if (tte4m_cnt > sectsb_thresh) { 10298 /* 10299 * If we're growing, select the size based on RSS. If we're 10300 * shrinking, leave some room so we don't have to turn 10301 * around and grow again immediately. 10302 */ 10303 if (growing) 10304 tsb_szc = SELECT_TSB_SIZECODE(tte4m_cnt); 10305 else 10306 tsb_szc = SELECT_TSB_SIZECODE(tte4m_cnt << 1); 10307 if (tsbinfop->tsb_next == NULL) { 10308 struct tsb_info *newtsb; 10309 int allocflags = SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)? 10310 0 : TSB_ALLOC; 10311 10312 sfmmu_hat_exit(hatlockp); 10313 10314 /* 10315 * Try to allocate a TSB for 4[32|256]M pages. If we 10316 * can't get the size we want, retry w/a minimum sized 10317 * TSB. If that still didn't work, give up; we can 10318 * still run without one. 10319 */ 10320 tsb_bits = (mmu_page_sizes == max_mmu_page_sizes)? 10321 TSB4M|TSB32M|TSB256M:TSB4M; 10322 if ((sfmmu_tsbinfo_alloc(&newtsb, tsb_szc, tsb_bits, 10323 allocflags, sfmmup)) && 10324 (tsb_szc <= TSB_4M_SZCODE || 10325 sfmmu_tsbinfo_alloc(&newtsb, TSB_4M_SZCODE, 10326 tsb_bits, allocflags, sfmmup)) && 10327 sfmmu_tsbinfo_alloc(&newtsb, TSB_MIN_SZCODE, 10328 tsb_bits, allocflags, sfmmup)) { 10329 return; 10330 } 10331 10332 hatlockp = sfmmu_hat_enter(sfmmup); 10333 10334 sfmmu_invalidate_ctx(sfmmup); 10335 10336 if (sfmmup->sfmmu_tsb->tsb_next == NULL) { 10337 sfmmup->sfmmu_tsb->tsb_next = newtsb; 10338 SFMMU_STAT(sf_tsb_sectsb_create); 10339 sfmmu_hat_exit(hatlockp); 10340 return; 10341 } else { 10342 /* 10343 * It's annoying, but possible for us 10344 * to get here.. we dropped the HAT lock 10345 * because of locking order in the kmem 10346 * allocator, and while we were off getting 10347 * our memory, some other thread decided to 10348 * do us a favor and won the race to get a 10349 * second TSB for this process. Sigh. 10350 */ 10351 sfmmu_hat_exit(hatlockp); 10352 sfmmu_tsbinfo_free(newtsb); 10353 return; 10354 } 10355 } 10356 10357 /* 10358 * We have a second TSB, see if it's big enough. 10359 */ 10360 tsbinfop = tsbinfop->tsb_next; 10361 10362 /* 10363 * Check to see if our second TSB is the right size; 10364 * we may need to grow or shrink it. 10365 * To prevent thrashing (e.g. growing the TSB on a 10366 * subsequent map operation), only try to shrink if 10367 * the TSB reach exceeds twice the virtual address 10368 * space size. 10369 */ 10370 if (!growing && (tsb_szc < tsbinfop->tsb_szc) && 10371 (tsb_szc >= default_tsb_size) && TSB_OK_SHRINK()) { 10372 (void) sfmmu_replace_tsb(sfmmup, tsbinfop, 10373 tsb_szc, hatlockp, TSB_SHRINK); 10374 } else if (growing && tsb_szc > tsbinfop->tsb_szc && 10375 TSB_OK_GROW()) { 10376 (void) sfmmu_replace_tsb(sfmmup, tsbinfop, 10377 tsb_szc, hatlockp, TSB_GROW); 10378 } 10379 } 10380 10381 sfmmu_hat_exit(hatlockp); 10382 } 10383 10384 /* 10385 * Free up a sfmmu 10386 * Since the sfmmu is currently embedded in the hat struct we simply zero 10387 * out our fields and free up the ism map blk list if any. 10388 */ 10389 static void 10390 sfmmu_free_sfmmu(sfmmu_t *sfmmup) 10391 { 10392 ism_blk_t *blkp, *nx_blkp; 10393 #ifdef DEBUG 10394 ism_map_t *map; 10395 int i; 10396 #endif 10397 10398 ASSERT(sfmmup->sfmmu_ttecnt[TTE8K] == 0); 10399 ASSERT(sfmmup->sfmmu_ttecnt[TTE64K] == 0); 10400 ASSERT(sfmmup->sfmmu_ttecnt[TTE512K] == 0); 10401 ASSERT(sfmmup->sfmmu_ttecnt[TTE4M] == 0); 10402 ASSERT(sfmmup->sfmmu_ttecnt[TTE32M] == 0); 10403 ASSERT(sfmmup->sfmmu_ttecnt[TTE256M] == 0); 10404 ASSERT(SF_RGNMAP_ISNULL(sfmmup)); 10405 10406 sfmmup->sfmmu_free = 0; 10407 sfmmup->sfmmu_ismhat = 0; 10408 10409 blkp = sfmmup->sfmmu_iblk; 10410 sfmmup->sfmmu_iblk = NULL; 10411 10412 while (blkp) { 10413 #ifdef DEBUG 10414 map = blkp->iblk_maps; 10415 for (i = 0; i < ISM_MAP_SLOTS; i++) { 10416 ASSERT(map[i].imap_seg == 0); 10417 ASSERT(map[i].imap_ismhat == NULL); 10418 ASSERT(map[i].imap_ment == NULL); 10419 } 10420 #endif 10421 nx_blkp = blkp->iblk_next; 10422 blkp->iblk_next = NULL; 10423 blkp->iblk_nextpa = (uint64_t)-1; 10424 kmem_cache_free(ism_blk_cache, blkp); 10425 blkp = nx_blkp; 10426 } 10427 } 10428 10429 /* 10430 * Locking primitves accessed by HATLOCK macros 10431 */ 10432 10433 #define SFMMU_SPL_MTX (0x0) 10434 #define SFMMU_ML_MTX (0x1) 10435 10436 #define SFMMU_MLSPL_MTX(type, pg) (((type) == SFMMU_SPL_MTX) ? \ 10437 SPL_HASH(pg) : MLIST_HASH(pg)) 10438 10439 kmutex_t * 10440 sfmmu_page_enter(struct page *pp) 10441 { 10442 return (sfmmu_mlspl_enter(pp, SFMMU_SPL_MTX)); 10443 } 10444 10445 void 10446 sfmmu_page_exit(kmutex_t *spl) 10447 { 10448 mutex_exit(spl); 10449 } 10450 10451 int 10452 sfmmu_page_spl_held(struct page *pp) 10453 { 10454 return (sfmmu_mlspl_held(pp, SFMMU_SPL_MTX)); 10455 } 10456 10457 kmutex_t * 10458 sfmmu_mlist_enter(struct page *pp) 10459 { 10460 return (sfmmu_mlspl_enter(pp, SFMMU_ML_MTX)); 10461 } 10462 10463 void 10464 sfmmu_mlist_exit(kmutex_t *mml) 10465 { 10466 mutex_exit(mml); 10467 } 10468 10469 int 10470 sfmmu_mlist_held(struct page *pp) 10471 { 10472 10473 return (sfmmu_mlspl_held(pp, SFMMU_ML_MTX)); 10474 } 10475 10476 /* 10477 * Common code for sfmmu_mlist_enter() and sfmmu_page_enter(). For 10478 * sfmmu_mlist_enter() case mml_table lock array is used and for 10479 * sfmmu_page_enter() sfmmu_page_lock lock array is used. 10480 * 10481 * The lock is taken on a root page so that it protects an operation on all 10482 * constituent pages of a large page pp belongs to. 10483 * 10484 * The routine takes a lock from the appropriate array. The lock is determined 10485 * by hashing the root page. After taking the lock this routine checks if the 10486 * root page has the same size code that was used to determine the root (i.e 10487 * that root hasn't changed). If root page has the expected p_szc field we 10488 * have the right lock and it's returned to the caller. If root's p_szc 10489 * decreased we release the lock and retry from the beginning. This case can 10490 * happen due to hat_page_demote() decreasing p_szc between our load of p_szc 10491 * value and taking the lock. The number of retries due to p_szc decrease is 10492 * limited by the maximum p_szc value. If p_szc is 0 we return the lock 10493 * determined by hashing pp itself. 10494 * 10495 * If our caller doesn't hold a SE_SHARED or SE_EXCL lock on pp it's also 10496 * possible that p_szc can increase. To increase p_szc a thread has to lock 10497 * all constituent pages EXCL and do hat_pageunload() on all of them. All the 10498 * callers that don't hold a page locked recheck if hmeblk through which pp 10499 * was found still maps this pp. If it doesn't map it anymore returned lock 10500 * is immediately dropped. Therefore if sfmmu_mlspl_enter() hits the case of 10501 * p_szc increase after taking the lock it returns this lock without further 10502 * retries because in this case the caller doesn't care about which lock was 10503 * taken. The caller will drop it right away. 10504 * 10505 * After the routine returns it's guaranteed that hat_page_demote() can't 10506 * change p_szc field of any of constituent pages of a large page pp belongs 10507 * to as long as pp was either locked at least SHARED prior to this call or 10508 * the caller finds that hment that pointed to this pp still references this 10509 * pp (this also assumes that the caller holds hme hash bucket lock so that 10510 * the same pp can't be remapped into the same hmeblk after it was unmapped by 10511 * hat_pageunload()). 10512 */ 10513 static kmutex_t * 10514 sfmmu_mlspl_enter(struct page *pp, int type) 10515 { 10516 kmutex_t *mtx; 10517 uint_t prev_rszc = UINT_MAX; 10518 page_t *rootpp; 10519 uint_t szc; 10520 uint_t rszc; 10521 uint_t pszc = pp->p_szc; 10522 10523 ASSERT(pp != NULL); 10524 10525 again: 10526 if (pszc == 0) { 10527 mtx = SFMMU_MLSPL_MTX(type, pp); 10528 mutex_enter(mtx); 10529 return (mtx); 10530 } 10531 10532 /* The lock lives in the root page */ 10533 rootpp = PP_GROUPLEADER(pp, pszc); 10534 mtx = SFMMU_MLSPL_MTX(type, rootpp); 10535 mutex_enter(mtx); 10536 10537 /* 10538 * Return mml in the following 3 cases: 10539 * 10540 * 1) If pp itself is root since if its p_szc decreased before we took 10541 * the lock pp is still the root of smaller szc page. And if its p_szc 10542 * increased it doesn't matter what lock we return (see comment in 10543 * front of this routine). 10544 * 10545 * 2) If pp's not root but rootpp is the root of a rootpp->p_szc size 10546 * large page we have the right lock since any previous potential 10547 * hat_page_demote() is done demoting from greater than current root's 10548 * p_szc because hat_page_demote() changes root's p_szc last. No 10549 * further hat_page_demote() can start or be in progress since it 10550 * would need the same lock we currently hold. 10551 * 10552 * 3) If rootpp's p_szc increased since previous iteration it doesn't 10553 * matter what lock we return (see comment in front of this routine). 10554 */ 10555 if (pp == rootpp || (rszc = rootpp->p_szc) == pszc || 10556 rszc >= prev_rszc) { 10557 return (mtx); 10558 } 10559 10560 /* 10561 * hat_page_demote() could have decreased root's p_szc. 10562 * In this case pp's p_szc must also be smaller than pszc. 10563 * Retry. 10564 */ 10565 if (rszc < pszc) { 10566 szc = pp->p_szc; 10567 if (szc < pszc) { 10568 mutex_exit(mtx); 10569 pszc = szc; 10570 goto again; 10571 } 10572 /* 10573 * pp's p_szc increased after it was decreased. 10574 * page cannot be mapped. Return current lock. The caller 10575 * will drop it right away. 10576 */ 10577 return (mtx); 10578 } 10579 10580 /* 10581 * root's p_szc is greater than pp's p_szc. 10582 * hat_page_demote() is not done with all pages 10583 * yet. Wait for it to complete. 10584 */ 10585 mutex_exit(mtx); 10586 rootpp = PP_GROUPLEADER(rootpp, rszc); 10587 mtx = SFMMU_MLSPL_MTX(type, rootpp); 10588 mutex_enter(mtx); 10589 mutex_exit(mtx); 10590 prev_rszc = rszc; 10591 goto again; 10592 } 10593 10594 static int 10595 sfmmu_mlspl_held(struct page *pp, int type) 10596 { 10597 kmutex_t *mtx; 10598 10599 ASSERT(pp != NULL); 10600 /* The lock lives in the root page */ 10601 pp = PP_PAGEROOT(pp); 10602 ASSERT(pp != NULL); 10603 10604 mtx = SFMMU_MLSPL_MTX(type, pp); 10605 return (MUTEX_HELD(mtx)); 10606 } 10607 10608 static uint_t 10609 sfmmu_get_free_hblk(struct hme_blk **hmeblkpp, uint_t critical) 10610 { 10611 struct hme_blk *hblkp; 10612 10613 10614 if (freehblkp != NULL) { 10615 mutex_enter(&freehblkp_lock); 10616 if (freehblkp != NULL) { 10617 /* 10618 * If the current thread is owning hblk_reserve OR 10619 * critical request from sfmmu_hblk_steal() 10620 * let it succeed even if freehblkcnt is really low. 10621 */ 10622 if (freehblkcnt <= HBLK_RESERVE_MIN && !critical) { 10623 SFMMU_STAT(sf_get_free_throttle); 10624 mutex_exit(&freehblkp_lock); 10625 return (0); 10626 } 10627 freehblkcnt--; 10628 *hmeblkpp = freehblkp; 10629 hblkp = *hmeblkpp; 10630 freehblkp = hblkp->hblk_next; 10631 mutex_exit(&freehblkp_lock); 10632 hblkp->hblk_next = NULL; 10633 SFMMU_STAT(sf_get_free_success); 10634 10635 ASSERT(hblkp->hblk_hmecnt == 0); 10636 ASSERT(hblkp->hblk_vcnt == 0); 10637 ASSERT(hblkp->hblk_nextpa == va_to_pa((caddr_t)hblkp)); 10638 10639 return (1); 10640 } 10641 mutex_exit(&freehblkp_lock); 10642 } 10643 10644 /* Check cpu hblk pending queues */ 10645 if ((*hmeblkpp = sfmmu_check_pending_hblks(TTE8K)) != NULL) { 10646 hblkp = *hmeblkpp; 10647 hblkp->hblk_next = NULL; 10648 hblkp->hblk_nextpa = va_to_pa((caddr_t)hblkp); 10649 10650 ASSERT(hblkp->hblk_hmecnt == 0); 10651 ASSERT(hblkp->hblk_vcnt == 0); 10652 10653 return (1); 10654 } 10655 10656 SFMMU_STAT(sf_get_free_fail); 10657 return (0); 10658 } 10659 10660 static uint_t 10661 sfmmu_put_free_hblk(struct hme_blk *hmeblkp, uint_t critical) 10662 { 10663 struct hme_blk *hblkp; 10664 10665 ASSERT(hmeblkp->hblk_hmecnt == 0); 10666 ASSERT(hmeblkp->hblk_vcnt == 0); 10667 ASSERT(hmeblkp->hblk_nextpa == va_to_pa((caddr_t)hmeblkp)); 10668 10669 /* 10670 * If the current thread is mapping into kernel space, 10671 * let it succede even if freehblkcnt is max 10672 * so that it will avoid freeing it to kmem. 10673 * This will prevent stack overflow due to 10674 * possible recursion since kmem_cache_free() 10675 * might require creation of a slab which 10676 * in turn needs an hmeblk to map that slab; 10677 * let's break this vicious chain at the first 10678 * opportunity. 10679 */ 10680 if (freehblkcnt < HBLK_RESERVE_CNT || critical) { 10681 mutex_enter(&freehblkp_lock); 10682 if (freehblkcnt < HBLK_RESERVE_CNT || critical) { 10683 SFMMU_STAT(sf_put_free_success); 10684 freehblkcnt++; 10685 hmeblkp->hblk_next = freehblkp; 10686 freehblkp = hmeblkp; 10687 mutex_exit(&freehblkp_lock); 10688 return (1); 10689 } 10690 mutex_exit(&freehblkp_lock); 10691 } 10692 10693 /* 10694 * Bring down freehblkcnt to HBLK_RESERVE_CNT. We are here 10695 * only if freehblkcnt is at least HBLK_RESERVE_CNT *and* 10696 * we are not in the process of mapping into kernel space. 10697 */ 10698 ASSERT(!critical); 10699 while (freehblkcnt > HBLK_RESERVE_CNT) { 10700 mutex_enter(&freehblkp_lock); 10701 if (freehblkcnt > HBLK_RESERVE_CNT) { 10702 freehblkcnt--; 10703 hblkp = freehblkp; 10704 freehblkp = hblkp->hblk_next; 10705 mutex_exit(&freehblkp_lock); 10706 ASSERT(get_hblk_cache(hblkp) == sfmmu8_cache); 10707 kmem_cache_free(sfmmu8_cache, hblkp); 10708 continue; 10709 } 10710 mutex_exit(&freehblkp_lock); 10711 } 10712 SFMMU_STAT(sf_put_free_fail); 10713 return (0); 10714 } 10715 10716 static void 10717 sfmmu_hblk_swap(struct hme_blk *new) 10718 { 10719 struct hme_blk *old, *hblkp, *prev; 10720 uint64_t newpa; 10721 caddr_t base, vaddr, endaddr; 10722 struct hmehash_bucket *hmebp; 10723 struct sf_hment *osfhme, *nsfhme; 10724 page_t *pp; 10725 kmutex_t *pml; 10726 tte_t tte; 10727 struct hme_blk *list = NULL; 10728 10729 #ifdef DEBUG 10730 hmeblk_tag hblktag; 10731 struct hme_blk *found; 10732 #endif 10733 old = HBLK_RESERVE; 10734 ASSERT(!old->hblk_shared); 10735 10736 /* 10737 * save pa before bcopy clobbers it 10738 */ 10739 newpa = new->hblk_nextpa; 10740 10741 base = (caddr_t)get_hblk_base(old); 10742 endaddr = base + get_hblk_span(old); 10743 10744 /* 10745 * acquire hash bucket lock. 10746 */ 10747 hmebp = sfmmu_tteload_acquire_hashbucket(ksfmmup, base, TTE8K, 10748 SFMMU_INVALID_SHMERID); 10749 10750 /* 10751 * copy contents from old to new 10752 */ 10753 bcopy((void *)old, (void *)new, HME8BLK_SZ); 10754 10755 /* 10756 * add new to hash chain 10757 */ 10758 sfmmu_hblk_hash_add(hmebp, new, newpa); 10759 10760 /* 10761 * search hash chain for hblk_reserve; this needs to be performed 10762 * after adding new, otherwise prev won't correspond to the hblk which 10763 * is prior to old in hash chain when we call sfmmu_hblk_hash_rm to 10764 * remove old later. 10765 */ 10766 for (prev = NULL, 10767 hblkp = hmebp->hmeblkp; hblkp != NULL && hblkp != old; 10768 prev = hblkp, hblkp = hblkp->hblk_next) 10769 ; 10770 10771 if (hblkp != old) 10772 panic("sfmmu_hblk_swap: hblk_reserve not found"); 10773 10774 /* 10775 * p_mapping list is still pointing to hments in hblk_reserve; 10776 * fix up p_mapping list so that they point to hments in new. 10777 * 10778 * Since all these mappings are created by hblk_reserve_thread 10779 * on the way and it's using at least one of the buffers from each of 10780 * the newly minted slabs, there is no danger of any of these 10781 * mappings getting unloaded by another thread. 10782 * 10783 * tsbmiss could only modify ref/mod bits of hments in old/new. 10784 * Since all of these hments hold mappings established by segkmem 10785 * and mappings in segkmem are setup with HAT_NOSYNC, ref/mod bits 10786 * have no meaning for the mappings in hblk_reserve. hments in 10787 * old and new are identical except for ref/mod bits. 10788 */ 10789 for (vaddr = base; vaddr < endaddr; vaddr += TTEBYTES(TTE8K)) { 10790 10791 HBLKTOHME(osfhme, old, vaddr); 10792 sfmmu_copytte(&osfhme->hme_tte, &tte); 10793 10794 if (TTE_IS_VALID(&tte)) { 10795 if ((pp = osfhme->hme_page) == NULL) 10796 panic("sfmmu_hblk_swap: page not mapped"); 10797 10798 pml = sfmmu_mlist_enter(pp); 10799 10800 if (pp != osfhme->hme_page) 10801 panic("sfmmu_hblk_swap: mapping changed"); 10802 10803 HBLKTOHME(nsfhme, new, vaddr); 10804 10805 HME_ADD(nsfhme, pp); 10806 HME_SUB(osfhme, pp); 10807 10808 sfmmu_mlist_exit(pml); 10809 } 10810 } 10811 10812 /* 10813 * remove old from hash chain 10814 */ 10815 sfmmu_hblk_hash_rm(hmebp, old, prev, &list, 1); 10816 10817 #ifdef DEBUG 10818 10819 hblktag.htag_id = ksfmmup; 10820 hblktag.htag_rid = SFMMU_INVALID_SHMERID; 10821 hblktag.htag_bspage = HME_HASH_BSPAGE(base, HME_HASH_SHIFT(TTE8K)); 10822 hblktag.htag_rehash = HME_HASH_REHASH(TTE8K); 10823 HME_HASH_FAST_SEARCH(hmebp, hblktag, found); 10824 10825 if (found != new) 10826 panic("sfmmu_hblk_swap: new hblk not found"); 10827 #endif 10828 10829 SFMMU_HASH_UNLOCK(hmebp); 10830 10831 /* 10832 * Reset hblk_reserve 10833 */ 10834 bzero((void *)old, HME8BLK_SZ); 10835 old->hblk_nextpa = va_to_pa((caddr_t)old); 10836 } 10837 10838 /* 10839 * Grab the mlist mutex for both pages passed in. 10840 * 10841 * low and high will be returned as pointers to the mutexes for these pages. 10842 * low refers to the mutex residing in the lower bin of the mlist hash, while 10843 * high refers to the mutex residing in the higher bin of the mlist hash. This 10844 * is due to the locking order restrictions on the same thread grabbing 10845 * multiple mlist mutexes. The low lock must be acquired before the high lock. 10846 * 10847 * If both pages hash to the same mutex, only grab that single mutex, and 10848 * high will be returned as NULL 10849 * If the pages hash to different bins in the hash, grab the lower addressed 10850 * lock first and then the higher addressed lock in order to follow the locking 10851 * rules involved with the same thread grabbing multiple mlist mutexes. 10852 * low and high will both have non-NULL values. 10853 */ 10854 static void 10855 sfmmu_mlist_reloc_enter(struct page *targ, struct page *repl, 10856 kmutex_t **low, kmutex_t **high) 10857 { 10858 kmutex_t *mml_targ, *mml_repl; 10859 10860 /* 10861 * no need to do the dance around szc as in sfmmu_mlist_enter() 10862 * because this routine is only called by hat_page_relocate() and all 10863 * targ and repl pages are already locked EXCL so szc can't change. 10864 */ 10865 10866 mml_targ = MLIST_HASH(PP_PAGEROOT(targ)); 10867 mml_repl = MLIST_HASH(PP_PAGEROOT(repl)); 10868 10869 if (mml_targ == mml_repl) { 10870 *low = mml_targ; 10871 *high = NULL; 10872 } else { 10873 if (mml_targ < mml_repl) { 10874 *low = mml_targ; 10875 *high = mml_repl; 10876 } else { 10877 *low = mml_repl; 10878 *high = mml_targ; 10879 } 10880 } 10881 10882 mutex_enter(*low); 10883 if (*high) 10884 mutex_enter(*high); 10885 } 10886 10887 static void 10888 sfmmu_mlist_reloc_exit(kmutex_t *low, kmutex_t *high) 10889 { 10890 if (high) 10891 mutex_exit(high); 10892 mutex_exit(low); 10893 } 10894 10895 static hatlock_t * 10896 sfmmu_hat_enter(sfmmu_t *sfmmup) 10897 { 10898 hatlock_t *hatlockp; 10899 10900 if (sfmmup != ksfmmup) { 10901 hatlockp = TSB_HASH(sfmmup); 10902 mutex_enter(HATLOCK_MUTEXP(hatlockp)); 10903 return (hatlockp); 10904 } 10905 return (NULL); 10906 } 10907 10908 static hatlock_t * 10909 sfmmu_hat_tryenter(sfmmu_t *sfmmup) 10910 { 10911 hatlock_t *hatlockp; 10912 10913 if (sfmmup != ksfmmup) { 10914 hatlockp = TSB_HASH(sfmmup); 10915 if (mutex_tryenter(HATLOCK_MUTEXP(hatlockp)) == 0) 10916 return (NULL); 10917 return (hatlockp); 10918 } 10919 return (NULL); 10920 } 10921 10922 static void 10923 sfmmu_hat_exit(hatlock_t *hatlockp) 10924 { 10925 if (hatlockp != NULL) 10926 mutex_exit(HATLOCK_MUTEXP(hatlockp)); 10927 } 10928 10929 static void 10930 sfmmu_hat_lock_all(void) 10931 { 10932 int i; 10933 for (i = 0; i < SFMMU_NUM_LOCK; i++) 10934 mutex_enter(HATLOCK_MUTEXP(&hat_lock[i])); 10935 } 10936 10937 static void 10938 sfmmu_hat_unlock_all(void) 10939 { 10940 int i; 10941 for (i = SFMMU_NUM_LOCK - 1; i >= 0; i--) 10942 mutex_exit(HATLOCK_MUTEXP(&hat_lock[i])); 10943 } 10944 10945 int 10946 sfmmu_hat_lock_held(sfmmu_t *sfmmup) 10947 { 10948 ASSERT(sfmmup != ksfmmup); 10949 return (MUTEX_HELD(HATLOCK_MUTEXP(TSB_HASH(sfmmup)))); 10950 } 10951 10952 /* 10953 * Locking primitives to provide consistency between ISM unmap 10954 * and other operations. Since ISM unmap can take a long time, we 10955 * use HAT_ISMBUSY flag (protected by the hatlock) to avoid creating 10956 * contention on the hatlock buckets while ISM segments are being 10957 * unmapped. The tradeoff is that the flags don't prevent priority 10958 * inversion from occurring, so we must request kernel priority in 10959 * case we have to sleep to keep from getting buried while holding 10960 * the HAT_ISMBUSY flag set, which in turn could block other kernel 10961 * threads from running (for example, in sfmmu_uvatopfn()). 10962 */ 10963 static void 10964 sfmmu_ismhat_enter(sfmmu_t *sfmmup, int hatlock_held) 10965 { 10966 hatlock_t *hatlockp; 10967 10968 THREAD_KPRI_REQUEST(); 10969 if (!hatlock_held) 10970 hatlockp = sfmmu_hat_enter(sfmmup); 10971 while (SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY)) 10972 cv_wait(&sfmmup->sfmmu_tsb_cv, HATLOCK_MUTEXP(hatlockp)); 10973 SFMMU_FLAGS_SET(sfmmup, HAT_ISMBUSY); 10974 if (!hatlock_held) 10975 sfmmu_hat_exit(hatlockp); 10976 } 10977 10978 static void 10979 sfmmu_ismhat_exit(sfmmu_t *sfmmup, int hatlock_held) 10980 { 10981 hatlock_t *hatlockp; 10982 10983 if (!hatlock_held) 10984 hatlockp = sfmmu_hat_enter(sfmmup); 10985 ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY)); 10986 SFMMU_FLAGS_CLEAR(sfmmup, HAT_ISMBUSY); 10987 cv_broadcast(&sfmmup->sfmmu_tsb_cv); 10988 if (!hatlock_held) 10989 sfmmu_hat_exit(hatlockp); 10990 THREAD_KPRI_RELEASE(); 10991 } 10992 10993 /* 10994 * 10995 * Algorithm: 10996 * 10997 * (1) if segkmem is not ready, allocate hblk from an array of pre-alloc'ed 10998 * hblks. 10999 * 11000 * (2) if we are allocating an hblk for mapping a slab in sfmmu_cache, 11001 * 11002 * (a) try to return an hblk from reserve pool of free hblks; 11003 * (b) if the reserve pool is empty, acquire hblk_reserve_lock 11004 * and return hblk_reserve. 11005 * 11006 * (3) call kmem_cache_alloc() to allocate hblk; 11007 * 11008 * (a) if hblk_reserve_lock is held by the current thread, 11009 * atomically replace hblk_reserve by the hblk that is 11010 * returned by kmem_cache_alloc; release hblk_reserve_lock 11011 * and call kmem_cache_alloc() again. 11012 * (b) if reserve pool is not full, add the hblk that is 11013 * returned by kmem_cache_alloc to reserve pool and 11014 * call kmem_cache_alloc again. 11015 * 11016 */ 11017 static struct hme_blk * 11018 sfmmu_hblk_alloc(sfmmu_t *sfmmup, caddr_t vaddr, 11019 struct hmehash_bucket *hmebp, uint_t size, hmeblk_tag hblktag, 11020 uint_t flags, uint_t rid) 11021 { 11022 struct hme_blk *hmeblkp = NULL; 11023 struct hme_blk *newhblkp; 11024 struct hme_blk *shw_hblkp = NULL; 11025 struct kmem_cache *sfmmu_cache = NULL; 11026 uint64_t hblkpa; 11027 ulong_t index; 11028 uint_t owner; /* set to 1 if using hblk_reserve */ 11029 uint_t forcefree; 11030 int sleep; 11031 sf_srd_t *srdp; 11032 sf_region_t *rgnp; 11033 11034 ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp)); 11035 ASSERT(hblktag.htag_rid == rid); 11036 SFMMU_VALIDATE_HMERID(sfmmup, rid, vaddr, TTEBYTES(size)); 11037 ASSERT(!SFMMU_IS_SHMERID_VALID(rid) || 11038 IS_P2ALIGNED(vaddr, TTEBYTES(size))); 11039 11040 /* 11041 * If segkmem is not created yet, allocate from static hmeblks 11042 * created at the end of startup_modules(). See the block comment 11043 * in startup_modules() describing how we estimate the number of 11044 * static hmeblks that will be needed during re-map. 11045 */ 11046 if (!hblk_alloc_dynamic) { 11047 11048 ASSERT(!SFMMU_IS_SHMERID_VALID(rid)); 11049 11050 if (size == TTE8K) { 11051 index = nucleus_hblk8.index; 11052 if (index >= nucleus_hblk8.len) { 11053 /* 11054 * If we panic here, see startup_modules() to 11055 * make sure that we are calculating the 11056 * number of hblk8's that we need correctly. 11057 */ 11058 prom_panic("no nucleus hblk8 to allocate"); 11059 } 11060 hmeblkp = 11061 (struct hme_blk *)&nucleus_hblk8.list[index]; 11062 nucleus_hblk8.index++; 11063 SFMMU_STAT(sf_hblk8_nalloc); 11064 } else { 11065 index = nucleus_hblk1.index; 11066 if (nucleus_hblk1.index >= nucleus_hblk1.len) { 11067 /* 11068 * If we panic here, see startup_modules(). 11069 * Most likely you need to update the 11070 * calculation of the number of hblk1 elements 11071 * that the kernel needs to boot. 11072 */ 11073 prom_panic("no nucleus hblk1 to allocate"); 11074 } 11075 hmeblkp = 11076 (struct hme_blk *)&nucleus_hblk1.list[index]; 11077 nucleus_hblk1.index++; 11078 SFMMU_STAT(sf_hblk1_nalloc); 11079 } 11080 11081 goto hblk_init; 11082 } 11083 11084 SFMMU_HASH_UNLOCK(hmebp); 11085 11086 if (sfmmup != KHATID && !SFMMU_IS_SHMERID_VALID(rid)) { 11087 if (mmu_page_sizes == max_mmu_page_sizes) { 11088 if (size < TTE256M) 11089 shw_hblkp = sfmmu_shadow_hcreate(sfmmup, vaddr, 11090 size, flags); 11091 } else { 11092 if (size < TTE4M) 11093 shw_hblkp = sfmmu_shadow_hcreate(sfmmup, vaddr, 11094 size, flags); 11095 } 11096 } else if (SFMMU_IS_SHMERID_VALID(rid)) { 11097 /* 11098 * Shared hmes use per region bitmaps in rgn_hmeflag 11099 * rather than shadow hmeblks to keep track of the 11100 * mapping sizes which have been allocated for the region. 11101 * Here we cleanup old invalid hmeblks with this rid, 11102 * which may be left around by pageunload(). 11103 */ 11104 int ttesz; 11105 caddr_t va; 11106 caddr_t eva = vaddr + TTEBYTES(size); 11107 11108 ASSERT(sfmmup != KHATID); 11109 11110 srdp = sfmmup->sfmmu_srdp; 11111 ASSERT(srdp != NULL && srdp->srd_refcnt != 0); 11112 rgnp = srdp->srd_hmergnp[rid]; 11113 ASSERT(rgnp != NULL && rgnp->rgn_id == rid); 11114 ASSERT(rgnp->rgn_refcnt != 0); 11115 ASSERT(size <= rgnp->rgn_pgszc); 11116 11117 ttesz = HBLK_MIN_TTESZ; 11118 do { 11119 if (!(rgnp->rgn_hmeflags & (0x1 << ttesz))) { 11120 continue; 11121 } 11122 11123 if (ttesz > size && ttesz != HBLK_MIN_TTESZ) { 11124 sfmmu_cleanup_rhblk(srdp, vaddr, rid, ttesz); 11125 } else if (ttesz < size) { 11126 for (va = vaddr; va < eva; 11127 va += TTEBYTES(ttesz)) { 11128 sfmmu_cleanup_rhblk(srdp, va, rid, 11129 ttesz); 11130 } 11131 } 11132 } while (++ttesz <= rgnp->rgn_pgszc); 11133 } 11134 11135 fill_hblk: 11136 owner = (hblk_reserve_thread == curthread) ? 1 : 0; 11137 11138 if (owner && size == TTE8K) { 11139 11140 ASSERT(!SFMMU_IS_SHMERID_VALID(rid)); 11141 /* 11142 * We are really in a tight spot. We already own 11143 * hblk_reserve and we need another hblk. In anticipation 11144 * of this kind of scenario, we specifically set aside 11145 * HBLK_RESERVE_MIN number of hblks to be used exclusively 11146 * by owner of hblk_reserve. 11147 */ 11148 SFMMU_STAT(sf_hblk_recurse_cnt); 11149 11150 if (!sfmmu_get_free_hblk(&hmeblkp, 1)) 11151 panic("sfmmu_hblk_alloc: reserve list is empty"); 11152 11153 goto hblk_verify; 11154 } 11155 11156 ASSERT(!owner); 11157 11158 if ((flags & HAT_NO_KALLOC) == 0) { 11159 11160 sfmmu_cache = ((size == TTE8K) ? sfmmu8_cache : sfmmu1_cache); 11161 sleep = ((sfmmup == KHATID) ? KM_NOSLEEP : KM_SLEEP); 11162 11163 if ((hmeblkp = kmem_cache_alloc(sfmmu_cache, sleep)) == NULL) { 11164 hmeblkp = sfmmu_hblk_steal(size); 11165 } else { 11166 /* 11167 * if we are the owner of hblk_reserve, 11168 * swap hblk_reserve with hmeblkp and 11169 * start a fresh life. Hope things go 11170 * better this time. 11171 */ 11172 if (hblk_reserve_thread == curthread) { 11173 ASSERT(sfmmu_cache == sfmmu8_cache); 11174 sfmmu_hblk_swap(hmeblkp); 11175 hblk_reserve_thread = NULL; 11176 mutex_exit(&hblk_reserve_lock); 11177 goto fill_hblk; 11178 } 11179 /* 11180 * let's donate this hblk to our reserve list if 11181 * we are not mapping kernel range 11182 */ 11183 if (size == TTE8K && sfmmup != KHATID) { 11184 if (sfmmu_put_free_hblk(hmeblkp, 0)) 11185 goto fill_hblk; 11186 } 11187 } 11188 } else { 11189 /* 11190 * We are here to map the slab in sfmmu8_cache; let's 11191 * check if we could tap our reserve list; if successful, 11192 * this will avoid the pain of going thru sfmmu_hblk_swap 11193 */ 11194 SFMMU_STAT(sf_hblk_slab_cnt); 11195 if (!sfmmu_get_free_hblk(&hmeblkp, 0)) { 11196 /* 11197 * let's start hblk_reserve dance 11198 */ 11199 SFMMU_STAT(sf_hblk_reserve_cnt); 11200 owner = 1; 11201 mutex_enter(&hblk_reserve_lock); 11202 hmeblkp = HBLK_RESERVE; 11203 hblk_reserve_thread = curthread; 11204 } 11205 } 11206 11207 hblk_verify: 11208 ASSERT(hmeblkp != NULL); 11209 set_hblk_sz(hmeblkp, size); 11210 ASSERT(hmeblkp->hblk_nextpa == va_to_pa((caddr_t)hmeblkp)); 11211 SFMMU_HASH_LOCK(hmebp); 11212 HME_HASH_FAST_SEARCH(hmebp, hblktag, newhblkp); 11213 if (newhblkp != NULL) { 11214 SFMMU_HASH_UNLOCK(hmebp); 11215 if (hmeblkp != HBLK_RESERVE) { 11216 /* 11217 * This is really tricky! 11218 * 11219 * vmem_alloc(vmem_seg_arena) 11220 * vmem_alloc(vmem_internal_arena) 11221 * segkmem_alloc(heap_arena) 11222 * vmem_alloc(heap_arena) 11223 * page_create() 11224 * hat_memload() 11225 * kmem_cache_free() 11226 * kmem_cache_alloc() 11227 * kmem_slab_create() 11228 * vmem_alloc(kmem_internal_arena) 11229 * segkmem_alloc(heap_arena) 11230 * vmem_alloc(heap_arena) 11231 * page_create() 11232 * hat_memload() 11233 * kmem_cache_free() 11234 * ... 11235 * 11236 * Thus, hat_memload() could call kmem_cache_free 11237 * for enough number of times that we could easily 11238 * hit the bottom of the stack or run out of reserve 11239 * list of vmem_seg structs. So, we must donate 11240 * this hblk to reserve list if it's allocated 11241 * from sfmmu8_cache *and* mapping kernel range. 11242 * We don't need to worry about freeing hmeblk1's 11243 * to kmem since they don't map any kmem slabs. 11244 * 11245 * Note: When segkmem supports largepages, we must 11246 * free hmeblk1's to reserve list as well. 11247 */ 11248 forcefree = (sfmmup == KHATID) ? 1 : 0; 11249 if (size == TTE8K && 11250 sfmmu_put_free_hblk(hmeblkp, forcefree)) { 11251 goto re_verify; 11252 } 11253 ASSERT(sfmmup != KHATID); 11254 kmem_cache_free(get_hblk_cache(hmeblkp), hmeblkp); 11255 } else { 11256 /* 11257 * Hey! we don't need hblk_reserve any more. 11258 */ 11259 ASSERT(owner); 11260 hblk_reserve_thread = NULL; 11261 mutex_exit(&hblk_reserve_lock); 11262 owner = 0; 11263 } 11264 re_verify: 11265 /* 11266 * let's check if the goodies are still present 11267 */ 11268 SFMMU_HASH_LOCK(hmebp); 11269 HME_HASH_FAST_SEARCH(hmebp, hblktag, newhblkp); 11270 if (newhblkp != NULL) { 11271 /* 11272 * return newhblkp if it's not hblk_reserve; 11273 * if newhblkp is hblk_reserve, return it 11274 * _only if_ we are the owner of hblk_reserve. 11275 */ 11276 if (newhblkp != HBLK_RESERVE || owner) { 11277 ASSERT(!SFMMU_IS_SHMERID_VALID(rid) || 11278 newhblkp->hblk_shared); 11279 ASSERT(SFMMU_IS_SHMERID_VALID(rid) || 11280 !newhblkp->hblk_shared); 11281 return (newhblkp); 11282 } else { 11283 /* 11284 * we just hit hblk_reserve in the hash and 11285 * we are not the owner of that; 11286 * 11287 * block until hblk_reserve_thread completes 11288 * swapping hblk_reserve and try the dance 11289 * once again. 11290 */ 11291 SFMMU_HASH_UNLOCK(hmebp); 11292 mutex_enter(&hblk_reserve_lock); 11293 mutex_exit(&hblk_reserve_lock); 11294 SFMMU_STAT(sf_hblk_reserve_hit); 11295 goto fill_hblk; 11296 } 11297 } else { 11298 /* 11299 * it's no more! try the dance once again. 11300 */ 11301 SFMMU_HASH_UNLOCK(hmebp); 11302 goto fill_hblk; 11303 } 11304 } 11305 11306 hblk_init: 11307 if (SFMMU_IS_SHMERID_VALID(rid)) { 11308 uint16_t tteflag = 0x1 << 11309 ((size < HBLK_MIN_TTESZ) ? HBLK_MIN_TTESZ : size); 11310 11311 if (!(rgnp->rgn_hmeflags & tteflag)) { 11312 atomic_or_16(&rgnp->rgn_hmeflags, tteflag); 11313 } 11314 hmeblkp->hblk_shared = 1; 11315 } else { 11316 hmeblkp->hblk_shared = 0; 11317 } 11318 set_hblk_sz(hmeblkp, size); 11319 ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp)); 11320 hmeblkp->hblk_next = (struct hme_blk *)NULL; 11321 hmeblkp->hblk_tag = hblktag; 11322 hmeblkp->hblk_shadow = shw_hblkp; 11323 hblkpa = hmeblkp->hblk_nextpa; 11324 hmeblkp->hblk_nextpa = HMEBLK_ENDPA; 11325 11326 ASSERT(get_hblk_ttesz(hmeblkp) == size); 11327 ASSERT(get_hblk_span(hmeblkp) == HMEBLK_SPAN(size)); 11328 ASSERT(hmeblkp->hblk_hmecnt == 0); 11329 ASSERT(hmeblkp->hblk_vcnt == 0); 11330 ASSERT(hmeblkp->hblk_lckcnt == 0); 11331 ASSERT(hblkpa == va_to_pa((caddr_t)hmeblkp)); 11332 sfmmu_hblk_hash_add(hmebp, hmeblkp, hblkpa); 11333 return (hmeblkp); 11334 } 11335 11336 /* 11337 * This function cleans up the hme_blk and returns it to the free list. 11338 */ 11339 /* ARGSUSED */ 11340 static void 11341 sfmmu_hblk_free(struct hme_blk **listp) 11342 { 11343 struct hme_blk *hmeblkp, *next_hmeblkp; 11344 int size; 11345 uint_t critical; 11346 uint64_t hblkpa; 11347 11348 ASSERT(*listp != NULL); 11349 11350 hmeblkp = *listp; 11351 while (hmeblkp != NULL) { 11352 next_hmeblkp = hmeblkp->hblk_next; 11353 ASSERT(!hmeblkp->hblk_hmecnt); 11354 ASSERT(!hmeblkp->hblk_vcnt); 11355 ASSERT(!hmeblkp->hblk_lckcnt); 11356 ASSERT(hmeblkp != (struct hme_blk *)hblk_reserve); 11357 ASSERT(hmeblkp->hblk_shared == 0); 11358 ASSERT(hmeblkp->hblk_shw_bit == 0); 11359 ASSERT(hmeblkp->hblk_shadow == NULL); 11360 11361 hblkpa = va_to_pa((caddr_t)hmeblkp); 11362 ASSERT(hblkpa != (uint64_t)-1); 11363 critical = (hblktosfmmu(hmeblkp) == KHATID) ? 1 : 0; 11364 11365 size = get_hblk_ttesz(hmeblkp); 11366 hmeblkp->hblk_next = NULL; 11367 hmeblkp->hblk_nextpa = hblkpa; 11368 11369 if (hmeblkp->hblk_nuc_bit == 0) { 11370 11371 if (size != TTE8K || 11372 !sfmmu_put_free_hblk(hmeblkp, critical)) 11373 kmem_cache_free(get_hblk_cache(hmeblkp), 11374 hmeblkp); 11375 } 11376 hmeblkp = next_hmeblkp; 11377 } 11378 } 11379 11380 #define BUCKETS_TO_SEARCH_BEFORE_UNLOAD 30 11381 #define SFMMU_HBLK_STEAL_THRESHOLD 5 11382 11383 static uint_t sfmmu_hblk_steal_twice; 11384 static uint_t sfmmu_hblk_steal_count, sfmmu_hblk_steal_unload_count; 11385 11386 /* 11387 * Steal a hmeblk from user or kernel hme hash lists. 11388 * For 8K tte grab one from reserve pool (freehblkp) before proceeding to 11389 * steal and if we fail to steal after SFMMU_HBLK_STEAL_THRESHOLD attempts 11390 * tap into critical reserve of freehblkp. 11391 * Note: We remain looping in this routine until we find one. 11392 */ 11393 static struct hme_blk * 11394 sfmmu_hblk_steal(int size) 11395 { 11396 static struct hmehash_bucket *uhmehash_steal_hand = NULL; 11397 struct hmehash_bucket *hmebp; 11398 struct hme_blk *hmeblkp = NULL, *pr_hblk; 11399 uint64_t hblkpa; 11400 int i; 11401 uint_t loop_cnt = 0, critical; 11402 11403 for (;;) { 11404 /* Check cpu hblk pending queues */ 11405 if ((hmeblkp = sfmmu_check_pending_hblks(size)) != NULL) { 11406 hmeblkp->hblk_nextpa = va_to_pa((caddr_t)hmeblkp); 11407 ASSERT(hmeblkp->hblk_hmecnt == 0); 11408 ASSERT(hmeblkp->hblk_vcnt == 0); 11409 return (hmeblkp); 11410 } 11411 11412 if (size == TTE8K) { 11413 critical = 11414 (++loop_cnt > SFMMU_HBLK_STEAL_THRESHOLD) ? 1 : 0; 11415 if (sfmmu_get_free_hblk(&hmeblkp, critical)) 11416 return (hmeblkp); 11417 } 11418 11419 hmebp = (uhmehash_steal_hand == NULL) ? uhme_hash : 11420 uhmehash_steal_hand; 11421 ASSERT(hmebp >= uhme_hash && hmebp <= &uhme_hash[UHMEHASH_SZ]); 11422 11423 for (i = 0; hmeblkp == NULL && i <= UHMEHASH_SZ + 11424 BUCKETS_TO_SEARCH_BEFORE_UNLOAD; i++) { 11425 SFMMU_HASH_LOCK(hmebp); 11426 hmeblkp = hmebp->hmeblkp; 11427 hblkpa = hmebp->hmeh_nextpa; 11428 pr_hblk = NULL; 11429 while (hmeblkp) { 11430 /* 11431 * check if it is a hmeblk that is not locked 11432 * and not shared. skip shadow hmeblks with 11433 * shadow_mask set i.e valid count non zero. 11434 */ 11435 if ((get_hblk_ttesz(hmeblkp) == size) && 11436 (hmeblkp->hblk_shw_bit == 0 || 11437 hmeblkp->hblk_vcnt == 0) && 11438 (hmeblkp->hblk_lckcnt == 0)) { 11439 /* 11440 * there is a high probability that we 11441 * will find a free one. search some 11442 * buckets for a free hmeblk initially 11443 * before unloading a valid hmeblk. 11444 */ 11445 if ((hmeblkp->hblk_vcnt == 0 && 11446 hmeblkp->hblk_hmecnt == 0) || (i >= 11447 BUCKETS_TO_SEARCH_BEFORE_UNLOAD)) { 11448 if (sfmmu_steal_this_hblk(hmebp, 11449 hmeblkp, hblkpa, pr_hblk)) { 11450 /* 11451 * Hblk is unloaded 11452 * successfully 11453 */ 11454 break; 11455 } 11456 } 11457 } 11458 pr_hblk = hmeblkp; 11459 hblkpa = hmeblkp->hblk_nextpa; 11460 hmeblkp = hmeblkp->hblk_next; 11461 } 11462 11463 SFMMU_HASH_UNLOCK(hmebp); 11464 if (hmebp++ == &uhme_hash[UHMEHASH_SZ]) 11465 hmebp = uhme_hash; 11466 } 11467 uhmehash_steal_hand = hmebp; 11468 11469 if (hmeblkp != NULL) 11470 break; 11471 11472 /* 11473 * in the worst case, look for a free one in the kernel 11474 * hash table. 11475 */ 11476 for (i = 0, hmebp = khme_hash; i <= KHMEHASH_SZ; i++) { 11477 SFMMU_HASH_LOCK(hmebp); 11478 hmeblkp = hmebp->hmeblkp; 11479 hblkpa = hmebp->hmeh_nextpa; 11480 pr_hblk = NULL; 11481 while (hmeblkp) { 11482 /* 11483 * check if it is free hmeblk 11484 */ 11485 if ((get_hblk_ttesz(hmeblkp) == size) && 11486 (hmeblkp->hblk_lckcnt == 0) && 11487 (hmeblkp->hblk_vcnt == 0) && 11488 (hmeblkp->hblk_hmecnt == 0)) { 11489 if (sfmmu_steal_this_hblk(hmebp, 11490 hmeblkp, hblkpa, pr_hblk)) { 11491 break; 11492 } else { 11493 /* 11494 * Cannot fail since we have 11495 * hash lock. 11496 */ 11497 panic("fail to steal?"); 11498 } 11499 } 11500 11501 pr_hblk = hmeblkp; 11502 hblkpa = hmeblkp->hblk_nextpa; 11503 hmeblkp = hmeblkp->hblk_next; 11504 } 11505 11506 SFMMU_HASH_UNLOCK(hmebp); 11507 if (hmebp++ == &khme_hash[KHMEHASH_SZ]) 11508 hmebp = khme_hash; 11509 } 11510 11511 if (hmeblkp != NULL) 11512 break; 11513 sfmmu_hblk_steal_twice++; 11514 } 11515 return (hmeblkp); 11516 } 11517 11518 /* 11519 * This routine does real work to prepare a hblk to be "stolen" by 11520 * unloading the mappings, updating shadow counts .... 11521 * It returns 1 if the block is ready to be reused (stolen), or 0 11522 * means the block cannot be stolen yet- pageunload is still working 11523 * on this hblk. 11524 */ 11525 static int 11526 sfmmu_steal_this_hblk(struct hmehash_bucket *hmebp, struct hme_blk *hmeblkp, 11527 uint64_t hblkpa, struct hme_blk *pr_hblk) 11528 { 11529 int shw_size, vshift; 11530 struct hme_blk *shw_hblkp; 11531 caddr_t vaddr; 11532 uint_t shw_mask, newshw_mask; 11533 struct hme_blk *list = NULL; 11534 11535 ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp)); 11536 11537 /* 11538 * check if the hmeblk is free, unload if necessary 11539 */ 11540 if (hmeblkp->hblk_vcnt || hmeblkp->hblk_hmecnt) { 11541 sfmmu_t *sfmmup; 11542 demap_range_t dmr; 11543 11544 sfmmup = hblktosfmmu(hmeblkp); 11545 if (hmeblkp->hblk_shared || sfmmup->sfmmu_ismhat) { 11546 return (0); 11547 } 11548 DEMAP_RANGE_INIT(sfmmup, &dmr); 11549 (void) sfmmu_hblk_unload(sfmmup, hmeblkp, 11550 (caddr_t)get_hblk_base(hmeblkp), 11551 get_hblk_endaddr(hmeblkp), &dmr, HAT_UNLOAD); 11552 DEMAP_RANGE_FLUSH(&dmr); 11553 if (hmeblkp->hblk_vcnt || hmeblkp->hblk_hmecnt) { 11554 /* 11555 * Pageunload is working on the same hblk. 11556 */ 11557 return (0); 11558 } 11559 11560 sfmmu_hblk_steal_unload_count++; 11561 } 11562 11563 ASSERT(hmeblkp->hblk_lckcnt == 0); 11564 ASSERT(hmeblkp->hblk_vcnt == 0 && hmeblkp->hblk_hmecnt == 0); 11565 11566 sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk, &list, 1); 11567 hmeblkp->hblk_nextpa = hblkpa; 11568 11569 shw_hblkp = hmeblkp->hblk_shadow; 11570 if (shw_hblkp) { 11571 ASSERT(!hmeblkp->hblk_shared); 11572 shw_size = get_hblk_ttesz(shw_hblkp); 11573 vaddr = (caddr_t)get_hblk_base(hmeblkp); 11574 vshift = vaddr_to_vshift(shw_hblkp->hblk_tag, vaddr, shw_size); 11575 ASSERT(vshift < 8); 11576 /* 11577 * Atomically clear shadow mask bit 11578 */ 11579 do { 11580 shw_mask = shw_hblkp->hblk_shw_mask; 11581 ASSERT(shw_mask & (1 << vshift)); 11582 newshw_mask = shw_mask & ~(1 << vshift); 11583 newshw_mask = cas32(&shw_hblkp->hblk_shw_mask, 11584 shw_mask, newshw_mask); 11585 } while (newshw_mask != shw_mask); 11586 hmeblkp->hblk_shadow = NULL; 11587 } 11588 11589 /* 11590 * remove shadow bit if we are stealing an unused shadow hmeblk. 11591 * sfmmu_hblk_alloc needs it that way, will set shadow bit later if 11592 * we are indeed allocating a shadow hmeblk. 11593 */ 11594 hmeblkp->hblk_shw_bit = 0; 11595 11596 if (hmeblkp->hblk_shared) { 11597 sf_srd_t *srdp; 11598 sf_region_t *rgnp; 11599 uint_t rid; 11600 11601 srdp = hblktosrd(hmeblkp); 11602 ASSERT(srdp != NULL && srdp->srd_refcnt != 0); 11603 rid = hmeblkp->hblk_tag.htag_rid; 11604 ASSERT(SFMMU_IS_SHMERID_VALID(rid)); 11605 ASSERT(rid < SFMMU_MAX_HME_REGIONS); 11606 rgnp = srdp->srd_hmergnp[rid]; 11607 ASSERT(rgnp != NULL); 11608 SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid); 11609 hmeblkp->hblk_shared = 0; 11610 } 11611 11612 sfmmu_hblk_steal_count++; 11613 SFMMU_STAT(sf_steal_count); 11614 11615 return (1); 11616 } 11617 11618 struct hme_blk * 11619 sfmmu_hmetohblk(struct sf_hment *sfhme) 11620 { 11621 struct hme_blk *hmeblkp; 11622 struct sf_hment *sfhme0; 11623 struct hme_blk *hblk_dummy = 0; 11624 11625 /* 11626 * No dummy sf_hments, please. 11627 */ 11628 ASSERT(sfhme->hme_tte.ll != 0); 11629 11630 sfhme0 = sfhme - sfhme->hme_tte.tte_hmenum; 11631 hmeblkp = (struct hme_blk *)((uintptr_t)sfhme0 - 11632 (uintptr_t)&hblk_dummy->hblk_hme[0]); 11633 11634 return (hmeblkp); 11635 } 11636 11637 /* 11638 * On swapin, get appropriately sized TSB(s) and clear the HAT_SWAPPED flag. 11639 * If we can't get appropriately sized TSB(s), try for 8K TSB(s) using 11640 * KM_SLEEP allocation. 11641 * 11642 * Return 0 on success, -1 otherwise. 11643 */ 11644 static void 11645 sfmmu_tsb_swapin(sfmmu_t *sfmmup, hatlock_t *hatlockp) 11646 { 11647 struct tsb_info *tsbinfop, *next; 11648 tsb_replace_rc_t rc; 11649 boolean_t gotfirst = B_FALSE; 11650 11651 ASSERT(sfmmup != ksfmmup); 11652 ASSERT(sfmmu_hat_lock_held(sfmmup)); 11653 11654 while (SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPIN)) { 11655 cv_wait(&sfmmup->sfmmu_tsb_cv, HATLOCK_MUTEXP(hatlockp)); 11656 } 11657 11658 if (SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) { 11659 SFMMU_FLAGS_SET(sfmmup, HAT_SWAPIN); 11660 } else { 11661 return; 11662 } 11663 11664 ASSERT(sfmmup->sfmmu_tsb != NULL); 11665 11666 /* 11667 * Loop over all tsbinfo's replacing them with ones that actually have 11668 * a TSB. If any of the replacements ever fail, bail out of the loop. 11669 */ 11670 for (tsbinfop = sfmmup->sfmmu_tsb; tsbinfop != NULL; tsbinfop = next) { 11671 ASSERT(tsbinfop->tsb_flags & TSB_SWAPPED); 11672 next = tsbinfop->tsb_next; 11673 rc = sfmmu_replace_tsb(sfmmup, tsbinfop, tsbinfop->tsb_szc, 11674 hatlockp, TSB_SWAPIN); 11675 if (rc != TSB_SUCCESS) { 11676 break; 11677 } 11678 gotfirst = B_TRUE; 11679 } 11680 11681 switch (rc) { 11682 case TSB_SUCCESS: 11683 SFMMU_FLAGS_CLEAR(sfmmup, HAT_SWAPPED|HAT_SWAPIN); 11684 cv_broadcast(&sfmmup->sfmmu_tsb_cv); 11685 return; 11686 case TSB_LOSTRACE: 11687 break; 11688 case TSB_ALLOCFAIL: 11689 break; 11690 default: 11691 panic("sfmmu_replace_tsb returned unrecognized failure code " 11692 "%d", rc); 11693 } 11694 11695 /* 11696 * In this case, we failed to get one of our TSBs. If we failed to 11697 * get the first TSB, get one of minimum size (8KB). Walk the list 11698 * and throw away the tsbinfos, starting where the allocation failed; 11699 * we can get by with just one TSB as long as we don't leave the 11700 * SWAPPED tsbinfo structures lying around. 11701 */ 11702 tsbinfop = sfmmup->sfmmu_tsb; 11703 next = tsbinfop->tsb_next; 11704 tsbinfop->tsb_next = NULL; 11705 11706 sfmmu_hat_exit(hatlockp); 11707 for (tsbinfop = next; tsbinfop != NULL; tsbinfop = next) { 11708 next = tsbinfop->tsb_next; 11709 sfmmu_tsbinfo_free(tsbinfop); 11710 } 11711 hatlockp = sfmmu_hat_enter(sfmmup); 11712 11713 /* 11714 * If we don't have any TSBs, get a single 8K TSB for 8K, 64K and 512K 11715 * pages. 11716 */ 11717 if (!gotfirst) { 11718 tsbinfop = sfmmup->sfmmu_tsb; 11719 rc = sfmmu_replace_tsb(sfmmup, tsbinfop, TSB_MIN_SZCODE, 11720 hatlockp, TSB_SWAPIN | TSB_FORCEALLOC); 11721 ASSERT(rc == TSB_SUCCESS); 11722 } 11723 11724 SFMMU_FLAGS_CLEAR(sfmmup, HAT_SWAPPED|HAT_SWAPIN); 11725 cv_broadcast(&sfmmup->sfmmu_tsb_cv); 11726 } 11727 11728 static int 11729 sfmmu_is_rgnva(sf_srd_t *srdp, caddr_t addr, ulong_t w, ulong_t bmw) 11730 { 11731 ulong_t bix = 0; 11732 uint_t rid; 11733 sf_region_t *rgnp; 11734 11735 ASSERT(srdp != NULL); 11736 ASSERT(srdp->srd_refcnt != 0); 11737 11738 w <<= BT_ULSHIFT; 11739 while (bmw) { 11740 if (!(bmw & 0x1)) { 11741 bix++; 11742 bmw >>= 1; 11743 continue; 11744 } 11745 rid = w | bix; 11746 rgnp = srdp->srd_hmergnp[rid]; 11747 ASSERT(rgnp->rgn_refcnt > 0); 11748 ASSERT(rgnp->rgn_id == rid); 11749 if (addr < rgnp->rgn_saddr || 11750 addr >= (rgnp->rgn_saddr + rgnp->rgn_size)) { 11751 bix++; 11752 bmw >>= 1; 11753 } else { 11754 return (1); 11755 } 11756 } 11757 return (0); 11758 } 11759 11760 /* 11761 * Handle exceptions for low level tsb_handler. 11762 * 11763 * There are many scenarios that could land us here: 11764 * 11765 * If the context is invalid we land here. The context can be invalid 11766 * for 3 reasons: 1) we couldn't allocate a new context and now need to 11767 * perform a wrap around operation in order to allocate a new context. 11768 * 2) Context was invalidated to change pagesize programming 3) ISMs or 11769 * TSBs configuration is changeing for this process and we are forced into 11770 * here to do a syncronization operation. If the context is valid we can 11771 * be here from window trap hanlder. In this case just call trap to handle 11772 * the fault. 11773 * 11774 * Note that the process will run in INVALID_CONTEXT before 11775 * faulting into here and subsequently loading the MMU registers 11776 * (including the TSB base register) associated with this process. 11777 * For this reason, the trap handlers must all test for 11778 * INVALID_CONTEXT before attempting to access any registers other 11779 * than the context registers. 11780 */ 11781 void 11782 sfmmu_tsbmiss_exception(struct regs *rp, uintptr_t tagaccess, uint_t traptype) 11783 { 11784 sfmmu_t *sfmmup, *shsfmmup; 11785 uint_t ctxtype; 11786 klwp_id_t lwp; 11787 char lwp_save_state; 11788 hatlock_t *hatlockp, *shatlockp; 11789 struct tsb_info *tsbinfop; 11790 struct tsbmiss *tsbmp; 11791 sf_scd_t *scdp; 11792 11793 SFMMU_STAT(sf_tsb_exceptions); 11794 SFMMU_MMU_STAT(mmu_tsb_exceptions); 11795 sfmmup = astosfmmu(curthread->t_procp->p_as); 11796 /* 11797 * note that in sun4u, tagacces register contains ctxnum 11798 * while sun4v passes ctxtype in the tagaccess register. 11799 */ 11800 ctxtype = tagaccess & TAGACC_CTX_MASK; 11801 11802 ASSERT(sfmmup != ksfmmup && ctxtype != KCONTEXT); 11803 ASSERT(sfmmup->sfmmu_ismhat == 0); 11804 ASSERT(!SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED) || 11805 ctxtype == INVALID_CONTEXT); 11806 11807 if (ctxtype != INVALID_CONTEXT && traptype != T_DATA_PROT) { 11808 /* 11809 * We may land here because shme bitmap and pagesize 11810 * flags are updated lazily in tsbmiss area on other cpus. 11811 * If we detect here that tsbmiss area is out of sync with 11812 * sfmmu update it and retry the trapped instruction. 11813 * Otherwise call trap(). 11814 */ 11815 int ret = 0; 11816 uchar_t tteflag_mask = (1 << TTE64K) | (1 << TTE8K); 11817 caddr_t addr = (caddr_t)(tagaccess & TAGACC_VADDR_MASK); 11818 11819 /* 11820 * Must set lwp state to LWP_SYS before 11821 * trying to acquire any adaptive lock 11822 */ 11823 lwp = ttolwp(curthread); 11824 ASSERT(lwp); 11825 lwp_save_state = lwp->lwp_state; 11826 lwp->lwp_state = LWP_SYS; 11827 11828 hatlockp = sfmmu_hat_enter(sfmmup); 11829 kpreempt_disable(); 11830 tsbmp = &tsbmiss_area[CPU->cpu_id]; 11831 ASSERT(sfmmup == tsbmp->usfmmup); 11832 if (((tsbmp->uhat_tteflags ^ sfmmup->sfmmu_tteflags) & 11833 ~tteflag_mask) || 11834 ((tsbmp->uhat_rtteflags ^ sfmmup->sfmmu_rtteflags) & 11835 ~tteflag_mask)) { 11836 tsbmp->uhat_tteflags = sfmmup->sfmmu_tteflags; 11837 tsbmp->uhat_rtteflags = sfmmup->sfmmu_rtteflags; 11838 ret = 1; 11839 } 11840 if (sfmmup->sfmmu_srdp != NULL) { 11841 ulong_t *sm = sfmmup->sfmmu_hmeregion_map.bitmap; 11842 ulong_t *tm = tsbmp->shmermap; 11843 ulong_t i; 11844 for (i = 0; i < SFMMU_HMERGNMAP_WORDS; i++) { 11845 ulong_t d = tm[i] ^ sm[i]; 11846 if (d) { 11847 if (d & sm[i]) { 11848 if (!ret && sfmmu_is_rgnva( 11849 sfmmup->sfmmu_srdp, 11850 addr, i, d & sm[i])) { 11851 ret = 1; 11852 } 11853 } 11854 tm[i] = sm[i]; 11855 } 11856 } 11857 } 11858 kpreempt_enable(); 11859 sfmmu_hat_exit(hatlockp); 11860 lwp->lwp_state = lwp_save_state; 11861 if (ret) { 11862 return; 11863 } 11864 } else if (ctxtype == INVALID_CONTEXT) { 11865 /* 11866 * First, make sure we come out of here with a valid ctx, 11867 * since if we don't get one we'll simply loop on the 11868 * faulting instruction. 11869 * 11870 * If the ISM mappings are changing, the TSB is relocated, 11871 * the process is swapped, the process is joining SCD or 11872 * leaving SCD or shared regions we serialize behind the 11873 * controlling thread with hat lock, sfmmu_flags and 11874 * sfmmu_tsb_cv condition variable. 11875 */ 11876 11877 /* 11878 * Must set lwp state to LWP_SYS before 11879 * trying to acquire any adaptive lock 11880 */ 11881 lwp = ttolwp(curthread); 11882 ASSERT(lwp); 11883 lwp_save_state = lwp->lwp_state; 11884 lwp->lwp_state = LWP_SYS; 11885 11886 hatlockp = sfmmu_hat_enter(sfmmup); 11887 retry: 11888 if ((scdp = sfmmup->sfmmu_scdp) != NULL) { 11889 shsfmmup = scdp->scd_sfmmup; 11890 ASSERT(shsfmmup != NULL); 11891 11892 for (tsbinfop = shsfmmup->sfmmu_tsb; tsbinfop != NULL; 11893 tsbinfop = tsbinfop->tsb_next) { 11894 if (tsbinfop->tsb_flags & TSB_RELOC_FLAG) { 11895 /* drop the private hat lock */ 11896 sfmmu_hat_exit(hatlockp); 11897 /* acquire the shared hat lock */ 11898 shatlockp = sfmmu_hat_enter(shsfmmup); 11899 /* 11900 * recheck to see if anything changed 11901 * after we drop the private hat lock. 11902 */ 11903 if (sfmmup->sfmmu_scdp == scdp && 11904 shsfmmup == scdp->scd_sfmmup) { 11905 sfmmu_tsb_chk_reloc(shsfmmup, 11906 shatlockp); 11907 } 11908 sfmmu_hat_exit(shatlockp); 11909 hatlockp = sfmmu_hat_enter(sfmmup); 11910 goto retry; 11911 } 11912 } 11913 } 11914 11915 for (tsbinfop = sfmmup->sfmmu_tsb; tsbinfop != NULL; 11916 tsbinfop = tsbinfop->tsb_next) { 11917 if (tsbinfop->tsb_flags & TSB_RELOC_FLAG) { 11918 cv_wait(&sfmmup->sfmmu_tsb_cv, 11919 HATLOCK_MUTEXP(hatlockp)); 11920 goto retry; 11921 } 11922 } 11923 11924 /* 11925 * Wait for ISM maps to be updated. 11926 */ 11927 if (SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY)) { 11928 cv_wait(&sfmmup->sfmmu_tsb_cv, 11929 HATLOCK_MUTEXP(hatlockp)); 11930 goto retry; 11931 } 11932 11933 /* Is this process joining an SCD? */ 11934 if (SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD)) { 11935 /* 11936 * Flush private TSB and setup shared TSB. 11937 * sfmmu_finish_join_scd() does not drop the 11938 * hat lock. 11939 */ 11940 sfmmu_finish_join_scd(sfmmup); 11941 SFMMU_FLAGS_CLEAR(sfmmup, HAT_JOIN_SCD); 11942 } 11943 11944 /* 11945 * If we're swapping in, get TSB(s). Note that we must do 11946 * this before we get a ctx or load the MMU state. Once 11947 * we swap in we have to recheck to make sure the TSB(s) and 11948 * ISM mappings didn't change while we slept. 11949 */ 11950 if (SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) { 11951 sfmmu_tsb_swapin(sfmmup, hatlockp); 11952 goto retry; 11953 } 11954 11955 sfmmu_get_ctx(sfmmup); 11956 11957 sfmmu_hat_exit(hatlockp); 11958 /* 11959 * Must restore lwp_state if not calling 11960 * trap() for further processing. Restore 11961 * it anyway. 11962 */ 11963 lwp->lwp_state = lwp_save_state; 11964 return; 11965 } 11966 trap(rp, (caddr_t)tagaccess, traptype, 0); 11967 } 11968 11969 static void 11970 sfmmu_tsb_chk_reloc(sfmmu_t *sfmmup, hatlock_t *hatlockp) 11971 { 11972 struct tsb_info *tp; 11973 11974 ASSERT(sfmmu_hat_lock_held(sfmmup)); 11975 11976 for (tp = sfmmup->sfmmu_tsb; tp != NULL; tp = tp->tsb_next) { 11977 if (tp->tsb_flags & TSB_RELOC_FLAG) { 11978 cv_wait(&sfmmup->sfmmu_tsb_cv, 11979 HATLOCK_MUTEXP(hatlockp)); 11980 break; 11981 } 11982 } 11983 } 11984 11985 /* 11986 * sfmmu_vatopfn_suspended is called from GET_TTE when TL=0 and 11987 * TTE_SUSPENDED bit set in tte we block on aquiring a page lock 11988 * rather than spinning to avoid send mondo timeouts with 11989 * interrupts enabled. When the lock is acquired it is immediately 11990 * released and we return back to sfmmu_vatopfn just after 11991 * the GET_TTE call. 11992 */ 11993 void 11994 sfmmu_vatopfn_suspended(caddr_t vaddr, sfmmu_t *sfmmu, tte_t *ttep) 11995 { 11996 struct page **pp; 11997 11998 (void) as_pagelock(sfmmu->sfmmu_as, &pp, vaddr, TTE_CSZ(ttep), S_WRITE); 11999 as_pageunlock(sfmmu->sfmmu_as, pp, vaddr, TTE_CSZ(ttep), S_WRITE); 12000 } 12001 12002 /* 12003 * sfmmu_tsbmiss_suspended is called from GET_TTE when TL>0 and 12004 * TTE_SUSPENDED bit set in tte. We do this so that we can handle 12005 * cross traps which cannot be handled while spinning in the 12006 * trap handlers. Simply enter and exit the kpr_suspendlock spin 12007 * mutex, which is held by the holder of the suspend bit, and then 12008 * retry the trapped instruction after unwinding. 12009 */ 12010 /*ARGSUSED*/ 12011 void 12012 sfmmu_tsbmiss_suspended(struct regs *rp, uintptr_t tagacc, uint_t traptype) 12013 { 12014 ASSERT(curthread != kreloc_thread); 12015 mutex_enter(&kpr_suspendlock); 12016 mutex_exit(&kpr_suspendlock); 12017 } 12018 12019 /* 12020 * This routine could be optimized to reduce the number of xcalls by flushing 12021 * the entire TLBs if region reference count is above some threshold but the 12022 * tradeoff will depend on the size of the TLB. So for now flush the specific 12023 * page a context at a time. 12024 * 12025 * If uselocks is 0 then it's called after all cpus were captured and all the 12026 * hat locks were taken. In this case don't take the region lock by relying on 12027 * the order of list region update operations in hat_join_region(), 12028 * hat_leave_region() and hat_dup_region(). The ordering in those routines 12029 * guarantees that list is always forward walkable and reaches active sfmmus 12030 * regardless of where xc_attention() captures a cpu. 12031 */ 12032 cpuset_t 12033 sfmmu_rgntlb_demap(caddr_t addr, sf_region_t *rgnp, 12034 struct hme_blk *hmeblkp, int uselocks) 12035 { 12036 sfmmu_t *sfmmup; 12037 cpuset_t cpuset; 12038 cpuset_t rcpuset; 12039 hatlock_t *hatlockp; 12040 uint_t rid = rgnp->rgn_id; 12041 sf_rgn_link_t *rlink; 12042 sf_scd_t *scdp; 12043 12044 ASSERT(hmeblkp->hblk_shared); 12045 ASSERT(SFMMU_IS_SHMERID_VALID(rid)); 12046 ASSERT(rid < SFMMU_MAX_HME_REGIONS); 12047 12048 CPUSET_ZERO(rcpuset); 12049 if (uselocks) { 12050 mutex_enter(&rgnp->rgn_mutex); 12051 } 12052 sfmmup = rgnp->rgn_sfmmu_head; 12053 while (sfmmup != NULL) { 12054 if (uselocks) { 12055 hatlockp = sfmmu_hat_enter(sfmmup); 12056 } 12057 12058 /* 12059 * When an SCD is created the SCD hat is linked on the sfmmu 12060 * region lists for each hme region which is part of the 12061 * SCD. If we find an SCD hat, when walking these lists, 12062 * then we flush the shared TSBs, if we find a private hat, 12063 * which is part of an SCD, but where the region 12064 * is not part of the SCD then we flush the private TSBs. 12065 */ 12066 if (!sfmmup->sfmmu_scdhat && sfmmup->sfmmu_scdp != NULL && 12067 !SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD)) { 12068 scdp = sfmmup->sfmmu_scdp; 12069 if (SF_RGNMAP_TEST(scdp->scd_hmeregion_map, rid)) { 12070 if (uselocks) { 12071 sfmmu_hat_exit(hatlockp); 12072 } 12073 goto next; 12074 } 12075 } 12076 12077 SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp, 0); 12078 12079 kpreempt_disable(); 12080 cpuset = sfmmup->sfmmu_cpusran; 12081 CPUSET_AND(cpuset, cpu_ready_set); 12082 CPUSET_DEL(cpuset, CPU->cpu_id); 12083 SFMMU_XCALL_STATS(sfmmup); 12084 xt_some(cpuset, vtag_flushpage_tl1, 12085 (uint64_t)addr, (uint64_t)sfmmup); 12086 vtag_flushpage(addr, (uint64_t)sfmmup); 12087 if (uselocks) { 12088 sfmmu_hat_exit(hatlockp); 12089 } 12090 kpreempt_enable(); 12091 CPUSET_OR(rcpuset, cpuset); 12092 12093 next: 12094 /* LINTED: constant in conditional context */ 12095 SFMMU_HMERID2RLINKP(sfmmup, rid, rlink, 0, 0); 12096 ASSERT(rlink != NULL); 12097 sfmmup = rlink->next; 12098 } 12099 if (uselocks) { 12100 mutex_exit(&rgnp->rgn_mutex); 12101 } 12102 return (rcpuset); 12103 } 12104 12105 /* 12106 * This routine takes an sfmmu pointer and the va for an adddress in an 12107 * ISM region as input and returns the corresponding region id in ism_rid. 12108 * The return value of 1 indicates that a region has been found and ism_rid 12109 * is valid, otherwise 0 is returned. 12110 */ 12111 static int 12112 find_ism_rid(sfmmu_t *sfmmup, sfmmu_t *ism_sfmmup, caddr_t va, uint_t *ism_rid) 12113 { 12114 ism_blk_t *ism_blkp; 12115 int i; 12116 ism_map_t *ism_map; 12117 #ifdef DEBUG 12118 struct hat *ism_hatid; 12119 #endif 12120 ASSERT(sfmmu_hat_lock_held(sfmmup)); 12121 12122 ism_blkp = sfmmup->sfmmu_iblk; 12123 while (ism_blkp != NULL) { 12124 ism_map = ism_blkp->iblk_maps; 12125 for (i = 0; i < ISM_MAP_SLOTS && ism_map[i].imap_ismhat; i++) { 12126 if ((va >= ism_start(ism_map[i])) && 12127 (va < ism_end(ism_map[i]))) { 12128 12129 *ism_rid = ism_map[i].imap_rid; 12130 #ifdef DEBUG 12131 ism_hatid = ism_map[i].imap_ismhat; 12132 ASSERT(ism_hatid == ism_sfmmup); 12133 ASSERT(ism_hatid->sfmmu_ismhat); 12134 #endif 12135 return (1); 12136 } 12137 } 12138 ism_blkp = ism_blkp->iblk_next; 12139 } 12140 return (0); 12141 } 12142 12143 /* 12144 * Special routine to flush out ism mappings- TSBs, TLBs and D-caches. 12145 * This routine may be called with all cpu's captured. Therefore, the 12146 * caller is responsible for holding all locks and disabling kernel 12147 * preemption. 12148 */ 12149 /* ARGSUSED */ 12150 static void 12151 sfmmu_ismtlbcache_demap(caddr_t addr, sfmmu_t *ism_sfmmup, 12152 struct hme_blk *hmeblkp, pfn_t pfnum, int cache_flush_flag) 12153 { 12154 cpuset_t cpuset; 12155 caddr_t va; 12156 ism_ment_t *ment; 12157 sfmmu_t *sfmmup; 12158 #ifdef VAC 12159 int vcolor; 12160 #endif 12161 12162 sf_scd_t *scdp; 12163 uint_t ism_rid; 12164 12165 ASSERT(!hmeblkp->hblk_shared); 12166 /* 12167 * Walk the ism_hat's mapping list and flush the page 12168 * from every hat sharing this ism_hat. This routine 12169 * may be called while all cpu's have been captured. 12170 * Therefore we can't attempt to grab any locks. For now 12171 * this means we will protect the ism mapping list under 12172 * a single lock which will be grabbed by the caller. 12173 * If hat_share/unshare scalibility becomes a performance 12174 * problem then we may need to re-think ism mapping list locking. 12175 */ 12176 ASSERT(ism_sfmmup->sfmmu_ismhat); 12177 ASSERT(MUTEX_HELD(&ism_mlist_lock)); 12178 addr = addr - ISMID_STARTADDR; 12179 12180 for (ment = ism_sfmmup->sfmmu_iment; ment; ment = ment->iment_next) { 12181 12182 sfmmup = ment->iment_hat; 12183 12184 va = ment->iment_base_va; 12185 va = (caddr_t)((uintptr_t)va + (uintptr_t)addr); 12186 12187 /* 12188 * When an SCD is created the SCD hat is linked on the ism 12189 * mapping lists for each ISM segment which is part of the 12190 * SCD. If we find an SCD hat, when walking these lists, 12191 * then we flush the shared TSBs, if we find a private hat, 12192 * which is part of an SCD, but where the region 12193 * corresponding to this va is not part of the SCD then we 12194 * flush the private TSBs. 12195 */ 12196 if (!sfmmup->sfmmu_scdhat && sfmmup->sfmmu_scdp != NULL && 12197 !SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD) && 12198 !SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY)) { 12199 if (!find_ism_rid(sfmmup, ism_sfmmup, va, 12200 &ism_rid)) { 12201 cmn_err(CE_PANIC, 12202 "can't find matching ISM rid!"); 12203 } 12204 12205 scdp = sfmmup->sfmmu_scdp; 12206 if (SFMMU_IS_ISMRID_VALID(ism_rid) && 12207 SF_RGNMAP_TEST(scdp->scd_ismregion_map, 12208 ism_rid)) { 12209 continue; 12210 } 12211 } 12212 SFMMU_UNLOAD_TSB(va, sfmmup, hmeblkp, 1); 12213 12214 cpuset = sfmmup->sfmmu_cpusran; 12215 CPUSET_AND(cpuset, cpu_ready_set); 12216 CPUSET_DEL(cpuset, CPU->cpu_id); 12217 SFMMU_XCALL_STATS(sfmmup); 12218 xt_some(cpuset, vtag_flushpage_tl1, (uint64_t)va, 12219 (uint64_t)sfmmup); 12220 vtag_flushpage(va, (uint64_t)sfmmup); 12221 12222 #ifdef VAC 12223 /* 12224 * Flush D$ 12225 * When flushing D$ we must flush all 12226 * cpu's. See sfmmu_cache_flush(). 12227 */ 12228 if (cache_flush_flag == CACHE_FLUSH) { 12229 cpuset = cpu_ready_set; 12230 CPUSET_DEL(cpuset, CPU->cpu_id); 12231 12232 SFMMU_XCALL_STATS(sfmmup); 12233 vcolor = addr_to_vcolor(va); 12234 xt_some(cpuset, vac_flushpage_tl1, pfnum, vcolor); 12235 vac_flushpage(pfnum, vcolor); 12236 } 12237 #endif /* VAC */ 12238 } 12239 } 12240 12241 /* 12242 * Demaps the TSB, CPU caches, and flushes all TLBs on all CPUs of 12243 * a particular virtual address and ctx. If noflush is set we do not 12244 * flush the TLB/TSB. This function may or may not be called with the 12245 * HAT lock held. 12246 */ 12247 static void 12248 sfmmu_tlbcache_demap(caddr_t addr, sfmmu_t *sfmmup, struct hme_blk *hmeblkp, 12249 pfn_t pfnum, int tlb_noflush, int cpu_flag, int cache_flush_flag, 12250 int hat_lock_held) 12251 { 12252 #ifdef VAC 12253 int vcolor; 12254 #endif 12255 cpuset_t cpuset; 12256 hatlock_t *hatlockp; 12257 12258 ASSERT(!hmeblkp->hblk_shared); 12259 12260 #if defined(lint) && !defined(VAC) 12261 pfnum = pfnum; 12262 cpu_flag = cpu_flag; 12263 cache_flush_flag = cache_flush_flag; 12264 #endif 12265 12266 /* 12267 * There is no longer a need to protect against ctx being 12268 * stolen here since we don't store the ctx in the TSB anymore. 12269 */ 12270 #ifdef VAC 12271 vcolor = addr_to_vcolor(addr); 12272 #endif 12273 12274 /* 12275 * We must hold the hat lock during the flush of TLB, 12276 * to avoid a race with sfmmu_invalidate_ctx(), where 12277 * sfmmu_cnum on a MMU could be set to INVALID_CONTEXT, 12278 * causing TLB demap routine to skip flush on that MMU. 12279 * If the context on a MMU has already been set to 12280 * INVALID_CONTEXT, we just get an extra flush on 12281 * that MMU. 12282 */ 12283 if (!hat_lock_held && !tlb_noflush) 12284 hatlockp = sfmmu_hat_enter(sfmmup); 12285 12286 kpreempt_disable(); 12287 if (!tlb_noflush) { 12288 /* 12289 * Flush the TSB and TLB. 12290 */ 12291 SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp, 0); 12292 12293 cpuset = sfmmup->sfmmu_cpusran; 12294 CPUSET_AND(cpuset, cpu_ready_set); 12295 CPUSET_DEL(cpuset, CPU->cpu_id); 12296 12297 SFMMU_XCALL_STATS(sfmmup); 12298 12299 xt_some(cpuset, vtag_flushpage_tl1, (uint64_t)addr, 12300 (uint64_t)sfmmup); 12301 12302 vtag_flushpage(addr, (uint64_t)sfmmup); 12303 } 12304 12305 if (!hat_lock_held && !tlb_noflush) 12306 sfmmu_hat_exit(hatlockp); 12307 12308 #ifdef VAC 12309 /* 12310 * Flush the D$ 12311 * 12312 * Even if the ctx is stolen, we need to flush the 12313 * cache. Our ctx stealer only flushes the TLBs. 12314 */ 12315 if (cache_flush_flag == CACHE_FLUSH) { 12316 if (cpu_flag & FLUSH_ALL_CPUS) { 12317 cpuset = cpu_ready_set; 12318 } else { 12319 cpuset = sfmmup->sfmmu_cpusran; 12320 CPUSET_AND(cpuset, cpu_ready_set); 12321 } 12322 CPUSET_DEL(cpuset, CPU->cpu_id); 12323 SFMMU_XCALL_STATS(sfmmup); 12324 xt_some(cpuset, vac_flushpage_tl1, pfnum, vcolor); 12325 vac_flushpage(pfnum, vcolor); 12326 } 12327 #endif /* VAC */ 12328 kpreempt_enable(); 12329 } 12330 12331 /* 12332 * Demaps the TSB and flushes all TLBs on all cpus for a particular virtual 12333 * address and ctx. If noflush is set we do not currently do anything. 12334 * This function may or may not be called with the HAT lock held. 12335 */ 12336 static void 12337 sfmmu_tlb_demap(caddr_t addr, sfmmu_t *sfmmup, struct hme_blk *hmeblkp, 12338 int tlb_noflush, int hat_lock_held) 12339 { 12340 cpuset_t cpuset; 12341 hatlock_t *hatlockp; 12342 12343 ASSERT(!hmeblkp->hblk_shared); 12344 12345 /* 12346 * If the process is exiting we have nothing to do. 12347 */ 12348 if (tlb_noflush) 12349 return; 12350 12351 /* 12352 * Flush TSB. 12353 */ 12354 if (!hat_lock_held) 12355 hatlockp = sfmmu_hat_enter(sfmmup); 12356 SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp, 0); 12357 12358 kpreempt_disable(); 12359 12360 cpuset = sfmmup->sfmmu_cpusran; 12361 CPUSET_AND(cpuset, cpu_ready_set); 12362 CPUSET_DEL(cpuset, CPU->cpu_id); 12363 12364 SFMMU_XCALL_STATS(sfmmup); 12365 xt_some(cpuset, vtag_flushpage_tl1, (uint64_t)addr, (uint64_t)sfmmup); 12366 12367 vtag_flushpage(addr, (uint64_t)sfmmup); 12368 12369 if (!hat_lock_held) 12370 sfmmu_hat_exit(hatlockp); 12371 12372 kpreempt_enable(); 12373 12374 } 12375 12376 /* 12377 * Special case of sfmmu_tlb_demap for MMU_PAGESIZE hblks. Use the xcall 12378 * call handler that can flush a range of pages to save on xcalls. 12379 */ 12380 static int sfmmu_xcall_save; 12381 12382 /* 12383 * this routine is never used for demaping addresses backed by SRD hmeblks. 12384 */ 12385 static void 12386 sfmmu_tlb_range_demap(demap_range_t *dmrp) 12387 { 12388 sfmmu_t *sfmmup = dmrp->dmr_sfmmup; 12389 hatlock_t *hatlockp; 12390 cpuset_t cpuset; 12391 uint64_t sfmmu_pgcnt; 12392 pgcnt_t pgcnt = 0; 12393 int pgunload = 0; 12394 int dirtypg = 0; 12395 caddr_t addr = dmrp->dmr_addr; 12396 caddr_t eaddr; 12397 uint64_t bitvec = dmrp->dmr_bitvec; 12398 12399 ASSERT(bitvec & 1); 12400 12401 /* 12402 * Flush TSB and calculate number of pages to flush. 12403 */ 12404 while (bitvec != 0) { 12405 dirtypg = 0; 12406 /* 12407 * Find the first page to flush and then count how many 12408 * pages there are after it that also need to be flushed. 12409 * This way the number of TSB flushes is minimized. 12410 */ 12411 while ((bitvec & 1) == 0) { 12412 pgcnt++; 12413 addr += MMU_PAGESIZE; 12414 bitvec >>= 1; 12415 } 12416 while (bitvec & 1) { 12417 dirtypg++; 12418 bitvec >>= 1; 12419 } 12420 eaddr = addr + ptob(dirtypg); 12421 hatlockp = sfmmu_hat_enter(sfmmup); 12422 sfmmu_unload_tsb_range(sfmmup, addr, eaddr, TTE8K); 12423 sfmmu_hat_exit(hatlockp); 12424 pgunload += dirtypg; 12425 addr = eaddr; 12426 pgcnt += dirtypg; 12427 } 12428 12429 ASSERT((pgcnt<<MMU_PAGESHIFT) <= dmrp->dmr_endaddr - dmrp->dmr_addr); 12430 if (sfmmup->sfmmu_free == 0) { 12431 addr = dmrp->dmr_addr; 12432 bitvec = dmrp->dmr_bitvec; 12433 12434 /* 12435 * make sure it has SFMMU_PGCNT_SHIFT bits only, 12436 * as it will be used to pack argument for xt_some 12437 */ 12438 ASSERT((pgcnt > 0) && 12439 (pgcnt <= (1 << SFMMU_PGCNT_SHIFT))); 12440 12441 /* 12442 * Encode pgcnt as (pgcnt -1 ), and pass (pgcnt - 1) in 12443 * the low 6 bits of sfmmup. This is doable since pgcnt 12444 * always >= 1. 12445 */ 12446 ASSERT(!((uint64_t)sfmmup & SFMMU_PGCNT_MASK)); 12447 sfmmu_pgcnt = (uint64_t)sfmmup | 12448 ((pgcnt - 1) & SFMMU_PGCNT_MASK); 12449 12450 /* 12451 * We must hold the hat lock during the flush of TLB, 12452 * to avoid a race with sfmmu_invalidate_ctx(), where 12453 * sfmmu_cnum on a MMU could be set to INVALID_CONTEXT, 12454 * causing TLB demap routine to skip flush on that MMU. 12455 * If the context on a MMU has already been set to 12456 * INVALID_CONTEXT, we just get an extra flush on 12457 * that MMU. 12458 */ 12459 hatlockp = sfmmu_hat_enter(sfmmup); 12460 kpreempt_disable(); 12461 12462 cpuset = sfmmup->sfmmu_cpusran; 12463 CPUSET_AND(cpuset, cpu_ready_set); 12464 CPUSET_DEL(cpuset, CPU->cpu_id); 12465 12466 SFMMU_XCALL_STATS(sfmmup); 12467 xt_some(cpuset, vtag_flush_pgcnt_tl1, (uint64_t)addr, 12468 sfmmu_pgcnt); 12469 12470 for (; bitvec != 0; bitvec >>= 1) { 12471 if (bitvec & 1) 12472 vtag_flushpage(addr, (uint64_t)sfmmup); 12473 addr += MMU_PAGESIZE; 12474 } 12475 kpreempt_enable(); 12476 sfmmu_hat_exit(hatlockp); 12477 12478 sfmmu_xcall_save += (pgunload-1); 12479 } 12480 dmrp->dmr_bitvec = 0; 12481 } 12482 12483 /* 12484 * In cases where we need to synchronize with TLB/TSB miss trap 12485 * handlers, _and_ need to flush the TLB, it's a lot easier to 12486 * throw away the context from the process than to do a 12487 * special song and dance to keep things consistent for the 12488 * handlers. 12489 * 12490 * Since the process suddenly ends up without a context and our caller 12491 * holds the hat lock, threads that fault after this function is called 12492 * will pile up on the lock. We can then do whatever we need to 12493 * atomically from the context of the caller. The first blocked thread 12494 * to resume executing will get the process a new context, and the 12495 * process will resume executing. 12496 * 12497 * One added advantage of this approach is that on MMUs that 12498 * support a "flush all" operation, we will delay the flush until 12499 * cnum wrap-around, and then flush the TLB one time. This 12500 * is rather rare, so it's a lot less expensive than making 8000 12501 * x-calls to flush the TLB 8000 times. 12502 * 12503 * A per-process (PP) lock is used to synchronize ctx allocations in 12504 * resume() and ctx invalidations here. 12505 */ 12506 static void 12507 sfmmu_invalidate_ctx(sfmmu_t *sfmmup) 12508 { 12509 cpuset_t cpuset; 12510 int cnum, currcnum; 12511 mmu_ctx_t *mmu_ctxp; 12512 int i; 12513 uint_t pstate_save; 12514 12515 SFMMU_STAT(sf_ctx_inv); 12516 12517 ASSERT(sfmmu_hat_lock_held(sfmmup)); 12518 ASSERT(sfmmup != ksfmmup); 12519 12520 kpreempt_disable(); 12521 12522 mmu_ctxp = CPU_MMU_CTXP(CPU); 12523 ASSERT(mmu_ctxp); 12524 ASSERT(mmu_ctxp->mmu_idx < max_mmu_ctxdoms); 12525 ASSERT(mmu_ctxp == mmu_ctxs_tbl[mmu_ctxp->mmu_idx]); 12526 12527 currcnum = sfmmup->sfmmu_ctxs[mmu_ctxp->mmu_idx].cnum; 12528 12529 pstate_save = sfmmu_disable_intrs(); 12530 12531 lock_set(&sfmmup->sfmmu_ctx_lock); /* acquire PP lock */ 12532 /* set HAT cnum invalid across all context domains. */ 12533 for (i = 0; i < max_mmu_ctxdoms; i++) { 12534 12535 cnum = sfmmup->sfmmu_ctxs[i].cnum; 12536 if (cnum == INVALID_CONTEXT) { 12537 continue; 12538 } 12539 12540 sfmmup->sfmmu_ctxs[i].cnum = INVALID_CONTEXT; 12541 } 12542 membar_enter(); /* make sure globally visible to all CPUs */ 12543 lock_clear(&sfmmup->sfmmu_ctx_lock); /* release PP lock */ 12544 12545 sfmmu_enable_intrs(pstate_save); 12546 12547 cpuset = sfmmup->sfmmu_cpusran; 12548 CPUSET_DEL(cpuset, CPU->cpu_id); 12549 CPUSET_AND(cpuset, cpu_ready_set); 12550 if (!CPUSET_ISNULL(cpuset)) { 12551 SFMMU_XCALL_STATS(sfmmup); 12552 xt_some(cpuset, sfmmu_raise_tsb_exception, 12553 (uint64_t)sfmmup, INVALID_CONTEXT); 12554 xt_sync(cpuset); 12555 SFMMU_STAT(sf_tsb_raise_exception); 12556 SFMMU_MMU_STAT(mmu_tsb_raise_exception); 12557 } 12558 12559 /* 12560 * If the hat to-be-invalidated is the same as the current 12561 * process on local CPU we need to invalidate 12562 * this CPU context as well. 12563 */ 12564 if ((sfmmu_getctx_sec() == currcnum) && 12565 (currcnum != INVALID_CONTEXT)) { 12566 /* sets shared context to INVALID too */ 12567 sfmmu_setctx_sec(INVALID_CONTEXT); 12568 sfmmu_clear_utsbinfo(); 12569 } 12570 12571 SFMMU_FLAGS_SET(sfmmup, HAT_ALLCTX_INVALID); 12572 12573 kpreempt_enable(); 12574 12575 /* 12576 * we hold the hat lock, so nobody should allocate a context 12577 * for us yet 12578 */ 12579 ASSERT(sfmmup->sfmmu_ctxs[mmu_ctxp->mmu_idx].cnum == INVALID_CONTEXT); 12580 } 12581 12582 #ifdef VAC 12583 /* 12584 * We need to flush the cache in all cpus. It is possible that 12585 * a process referenced a page as cacheable but has sinced exited 12586 * and cleared the mapping list. We still to flush it but have no 12587 * state so all cpus is the only alternative. 12588 */ 12589 void 12590 sfmmu_cache_flush(pfn_t pfnum, int vcolor) 12591 { 12592 cpuset_t cpuset; 12593 12594 kpreempt_disable(); 12595 cpuset = cpu_ready_set; 12596 CPUSET_DEL(cpuset, CPU->cpu_id); 12597 SFMMU_XCALL_STATS(NULL); /* account to any ctx */ 12598 xt_some(cpuset, vac_flushpage_tl1, pfnum, vcolor); 12599 xt_sync(cpuset); 12600 vac_flushpage(pfnum, vcolor); 12601 kpreempt_enable(); 12602 } 12603 12604 void 12605 sfmmu_cache_flushcolor(int vcolor, pfn_t pfnum) 12606 { 12607 cpuset_t cpuset; 12608 12609 ASSERT(vcolor >= 0); 12610 12611 kpreempt_disable(); 12612 cpuset = cpu_ready_set; 12613 CPUSET_DEL(cpuset, CPU->cpu_id); 12614 SFMMU_XCALL_STATS(NULL); /* account to any ctx */ 12615 xt_some(cpuset, vac_flushcolor_tl1, vcolor, pfnum); 12616 xt_sync(cpuset); 12617 vac_flushcolor(vcolor, pfnum); 12618 kpreempt_enable(); 12619 } 12620 #endif /* VAC */ 12621 12622 /* 12623 * We need to prevent processes from accessing the TSB using a cached physical 12624 * address. It's alright if they try to access the TSB via virtual address 12625 * since they will just fault on that virtual address once the mapping has 12626 * been suspended. 12627 */ 12628 #pragma weak sendmondo_in_recover 12629 12630 /* ARGSUSED */ 12631 static int 12632 sfmmu_tsb_pre_relocator(caddr_t va, uint_t tsbsz, uint_t flags, void *tsbinfo) 12633 { 12634 struct tsb_info *tsbinfop = (struct tsb_info *)tsbinfo; 12635 sfmmu_t *sfmmup = tsbinfop->tsb_sfmmu; 12636 hatlock_t *hatlockp; 12637 sf_scd_t *scdp; 12638 12639 if (flags != HAT_PRESUSPEND) 12640 return (0); 12641 12642 /* 12643 * If tsb is a shared TSB with TSB_SHAREDCTX set, sfmmup must 12644 * be a shared hat, then set SCD's tsbinfo's flag. 12645 * If tsb is not shared, sfmmup is a private hat, then set 12646 * its private tsbinfo's flag. 12647 */ 12648 hatlockp = sfmmu_hat_enter(sfmmup); 12649 tsbinfop->tsb_flags |= TSB_RELOC_FLAG; 12650 12651 if (!(tsbinfop->tsb_flags & TSB_SHAREDCTX)) { 12652 sfmmu_tsb_inv_ctx(sfmmup); 12653 sfmmu_hat_exit(hatlockp); 12654 } else { 12655 /* release lock on the shared hat */ 12656 sfmmu_hat_exit(hatlockp); 12657 /* sfmmup is a shared hat */ 12658 ASSERT(sfmmup->sfmmu_scdhat); 12659 scdp = sfmmup->sfmmu_scdp; 12660 ASSERT(scdp != NULL); 12661 /* get private hat from the scd list */ 12662 mutex_enter(&scdp->scd_mutex); 12663 sfmmup = scdp->scd_sf_list; 12664 while (sfmmup != NULL) { 12665 hatlockp = sfmmu_hat_enter(sfmmup); 12666 /* 12667 * We do not call sfmmu_tsb_inv_ctx here because 12668 * sendmondo_in_recover check is only needed for 12669 * sun4u. 12670 */ 12671 sfmmu_invalidate_ctx(sfmmup); 12672 sfmmu_hat_exit(hatlockp); 12673 sfmmup = sfmmup->sfmmu_scd_link.next; 12674 12675 } 12676 mutex_exit(&scdp->scd_mutex); 12677 } 12678 return (0); 12679 } 12680 12681 static void 12682 sfmmu_tsb_inv_ctx(sfmmu_t *sfmmup) 12683 { 12684 extern uint32_t sendmondo_in_recover; 12685 12686 ASSERT(sfmmu_hat_lock_held(sfmmup)); 12687 12688 /* 12689 * For Cheetah+ Erratum 25: 12690 * Wait for any active recovery to finish. We can't risk 12691 * relocating the TSB of the thread running mondo_recover_proc() 12692 * since, if we did that, we would deadlock. The scenario we are 12693 * trying to avoid is as follows: 12694 * 12695 * THIS CPU RECOVER CPU 12696 * -------- ----------- 12697 * Begins recovery, walking through TSB 12698 * hat_pagesuspend() TSB TTE 12699 * TLB miss on TSB TTE, spins at TL1 12700 * xt_sync() 12701 * send_mondo_timeout() 12702 * mondo_recover_proc() 12703 * ((deadlocked)) 12704 * 12705 * The second half of the workaround is that mondo_recover_proc() 12706 * checks to see if the tsb_info has the RELOC flag set, and if it 12707 * does, it skips over that TSB without ever touching tsbinfop->tsb_va 12708 * and hence avoiding the TLB miss that could result in a deadlock. 12709 */ 12710 if (&sendmondo_in_recover) { 12711 membar_enter(); /* make sure RELOC flag visible */ 12712 while (sendmondo_in_recover) { 12713 drv_usecwait(1); 12714 membar_consumer(); 12715 } 12716 } 12717 12718 sfmmu_invalidate_ctx(sfmmup); 12719 } 12720 12721 /* ARGSUSED */ 12722 static int 12723 sfmmu_tsb_post_relocator(caddr_t va, uint_t tsbsz, uint_t flags, 12724 void *tsbinfo, pfn_t newpfn) 12725 { 12726 hatlock_t *hatlockp; 12727 struct tsb_info *tsbinfop = (struct tsb_info *)tsbinfo; 12728 sfmmu_t *sfmmup = tsbinfop->tsb_sfmmu; 12729 12730 if (flags != HAT_POSTUNSUSPEND) 12731 return (0); 12732 12733 hatlockp = sfmmu_hat_enter(sfmmup); 12734 12735 SFMMU_STAT(sf_tsb_reloc); 12736 12737 /* 12738 * The process may have swapped out while we were relocating one 12739 * of its TSBs. If so, don't bother doing the setup since the 12740 * process can't be using the memory anymore. 12741 */ 12742 if ((tsbinfop->tsb_flags & TSB_SWAPPED) == 0) { 12743 ASSERT(va == tsbinfop->tsb_va); 12744 sfmmu_tsbinfo_setup_phys(tsbinfop, newpfn); 12745 12746 if (tsbinfop->tsb_flags & TSB_FLUSH_NEEDED) { 12747 sfmmu_inv_tsb(tsbinfop->tsb_va, 12748 TSB_BYTES(tsbinfop->tsb_szc)); 12749 tsbinfop->tsb_flags &= ~TSB_FLUSH_NEEDED; 12750 } 12751 } 12752 12753 membar_exit(); 12754 tsbinfop->tsb_flags &= ~TSB_RELOC_FLAG; 12755 cv_broadcast(&sfmmup->sfmmu_tsb_cv); 12756 12757 sfmmu_hat_exit(hatlockp); 12758 12759 return (0); 12760 } 12761 12762 /* 12763 * Allocate and initialize a tsb_info structure. Note that we may or may not 12764 * allocate a TSB here, depending on the flags passed in. 12765 */ 12766 static int 12767 sfmmu_tsbinfo_alloc(struct tsb_info **tsbinfopp, int tsb_szc, int tte_sz_mask, 12768 uint_t flags, sfmmu_t *sfmmup) 12769 { 12770 int err; 12771 12772 *tsbinfopp = (struct tsb_info *)kmem_cache_alloc( 12773 sfmmu_tsbinfo_cache, KM_SLEEP); 12774 12775 if ((err = sfmmu_init_tsbinfo(*tsbinfopp, tte_sz_mask, 12776 tsb_szc, flags, sfmmup)) != 0) { 12777 kmem_cache_free(sfmmu_tsbinfo_cache, *tsbinfopp); 12778 SFMMU_STAT(sf_tsb_allocfail); 12779 *tsbinfopp = NULL; 12780 return (err); 12781 } 12782 SFMMU_STAT(sf_tsb_alloc); 12783 12784 /* 12785 * Bump the TSB size counters for this TSB size. 12786 */ 12787 (*(((int *)&sfmmu_tsbsize_stat) + tsb_szc))++; 12788 return (0); 12789 } 12790 12791 static void 12792 sfmmu_tsb_free(struct tsb_info *tsbinfo) 12793 { 12794 caddr_t tsbva = tsbinfo->tsb_va; 12795 uint_t tsb_size = TSB_BYTES(tsbinfo->tsb_szc); 12796 struct kmem_cache *kmem_cachep = tsbinfo->tsb_cache; 12797 vmem_t *vmp = tsbinfo->tsb_vmp; 12798 12799 /* 12800 * If we allocated this TSB from relocatable kernel memory, then we 12801 * need to uninstall the callback handler. 12802 */ 12803 if (tsbinfo->tsb_cache != sfmmu_tsb8k_cache) { 12804 uintptr_t slab_mask; 12805 caddr_t slab_vaddr; 12806 page_t **ppl; 12807 int ret; 12808 12809 ASSERT(tsb_size <= MMU_PAGESIZE4M || use_bigtsb_arena); 12810 if (tsb_size > MMU_PAGESIZE4M) 12811 slab_mask = ~((uintptr_t)bigtsb_slab_mask) << PAGESHIFT; 12812 else 12813 slab_mask = ~((uintptr_t)tsb_slab_mask) << PAGESHIFT; 12814 slab_vaddr = (caddr_t)((uintptr_t)tsbva & slab_mask); 12815 12816 ret = as_pagelock(&kas, &ppl, slab_vaddr, PAGESIZE, S_WRITE); 12817 ASSERT(ret == 0); 12818 hat_delete_callback(tsbva, (uint_t)tsb_size, (void *)tsbinfo, 12819 0, NULL); 12820 as_pageunlock(&kas, ppl, slab_vaddr, PAGESIZE, S_WRITE); 12821 } 12822 12823 if (kmem_cachep != NULL) { 12824 kmem_cache_free(kmem_cachep, tsbva); 12825 } else { 12826 vmem_xfree(vmp, (void *)tsbva, tsb_size); 12827 } 12828 tsbinfo->tsb_va = (caddr_t)0xbad00bad; 12829 atomic_add_64(&tsb_alloc_bytes, -(int64_t)tsb_size); 12830 } 12831 12832 static void 12833 sfmmu_tsbinfo_free(struct tsb_info *tsbinfo) 12834 { 12835 if ((tsbinfo->tsb_flags & TSB_SWAPPED) == 0) { 12836 sfmmu_tsb_free(tsbinfo); 12837 } 12838 kmem_cache_free(sfmmu_tsbinfo_cache, tsbinfo); 12839 12840 } 12841 12842 /* 12843 * Setup all the references to physical memory for this tsbinfo. 12844 * The underlying page(s) must be locked. 12845 */ 12846 static void 12847 sfmmu_tsbinfo_setup_phys(struct tsb_info *tsbinfo, pfn_t pfn) 12848 { 12849 ASSERT(pfn != PFN_INVALID); 12850 ASSERT(pfn == va_to_pfn(tsbinfo->tsb_va)); 12851 12852 #ifndef sun4v 12853 if (tsbinfo->tsb_szc == 0) { 12854 sfmmu_memtte(&tsbinfo->tsb_tte, pfn, 12855 PROT_WRITE|PROT_READ, TTE8K); 12856 } else { 12857 /* 12858 * Round down PA and use a large mapping; the handlers will 12859 * compute the TSB pointer at the correct offset into the 12860 * big virtual page. NOTE: this assumes all TSBs larger 12861 * than 8K must come from physically contiguous slabs of 12862 * size tsb_slab_size. 12863 */ 12864 sfmmu_memtte(&tsbinfo->tsb_tte, pfn & ~tsb_slab_mask, 12865 PROT_WRITE|PROT_READ, tsb_slab_ttesz); 12866 } 12867 tsbinfo->tsb_pa = ptob(pfn); 12868 12869 TTE_SET_LOCKED(&tsbinfo->tsb_tte); /* lock the tte into dtlb */ 12870 TTE_SET_MOD(&tsbinfo->tsb_tte); /* enable writes */ 12871 12872 ASSERT(TTE_IS_PRIVILEGED(&tsbinfo->tsb_tte)); 12873 ASSERT(TTE_IS_LOCKED(&tsbinfo->tsb_tte)); 12874 #else /* sun4v */ 12875 tsbinfo->tsb_pa = ptob(pfn); 12876 #endif /* sun4v */ 12877 } 12878 12879 12880 /* 12881 * Returns zero on success, ENOMEM if over the high water mark, 12882 * or EAGAIN if the caller needs to retry with a smaller TSB 12883 * size (or specify TSB_FORCEALLOC if the allocation can't fail). 12884 * 12885 * This call cannot fail to allocate a TSB if TSB_FORCEALLOC 12886 * is specified and the TSB requested is PAGESIZE, though it 12887 * may sleep waiting for memory if sufficient memory is not 12888 * available. 12889 */ 12890 static int 12891 sfmmu_init_tsbinfo(struct tsb_info *tsbinfo, int tteszmask, 12892 int tsbcode, uint_t flags, sfmmu_t *sfmmup) 12893 { 12894 caddr_t vaddr = NULL; 12895 caddr_t slab_vaddr; 12896 uintptr_t slab_mask; 12897 int tsbbytes = TSB_BYTES(tsbcode); 12898 int lowmem = 0; 12899 struct kmem_cache *kmem_cachep = NULL; 12900 vmem_t *vmp = NULL; 12901 lgrp_id_t lgrpid = LGRP_NONE; 12902 pfn_t pfn; 12903 uint_t cbflags = HAC_SLEEP; 12904 page_t **pplist; 12905 int ret; 12906 12907 ASSERT(tsbbytes <= MMU_PAGESIZE4M || use_bigtsb_arena); 12908 if (tsbbytes > MMU_PAGESIZE4M) 12909 slab_mask = ~((uintptr_t)bigtsb_slab_mask) << PAGESHIFT; 12910 else 12911 slab_mask = ~((uintptr_t)tsb_slab_mask) << PAGESHIFT; 12912 12913 if (flags & (TSB_FORCEALLOC | TSB_SWAPIN | TSB_GROW | TSB_SHRINK)) 12914 flags |= TSB_ALLOC; 12915 12916 ASSERT((flags & TSB_FORCEALLOC) == 0 || tsbcode == TSB_MIN_SZCODE); 12917 12918 tsbinfo->tsb_sfmmu = sfmmup; 12919 12920 /* 12921 * If not allocating a TSB, set up the tsbinfo, set TSB_SWAPPED, and 12922 * return. 12923 */ 12924 if ((flags & TSB_ALLOC) == 0) { 12925 tsbinfo->tsb_szc = tsbcode; 12926 tsbinfo->tsb_ttesz_mask = tteszmask; 12927 tsbinfo->tsb_va = (caddr_t)0xbadbadbeef; 12928 tsbinfo->tsb_pa = -1; 12929 tsbinfo->tsb_tte.ll = 0; 12930 tsbinfo->tsb_next = NULL; 12931 tsbinfo->tsb_flags = TSB_SWAPPED; 12932 tsbinfo->tsb_cache = NULL; 12933 tsbinfo->tsb_vmp = NULL; 12934 return (0); 12935 } 12936 12937 #ifdef DEBUG 12938 /* 12939 * For debugging: 12940 * Randomly force allocation failures every tsb_alloc_mtbf 12941 * tries if TSB_FORCEALLOC is not specified. This will 12942 * return ENOMEM if tsb_alloc_mtbf is odd, or EAGAIN if 12943 * it is even, to allow testing of both failure paths... 12944 */ 12945 if (tsb_alloc_mtbf && ((flags & TSB_FORCEALLOC) == 0) && 12946 (tsb_alloc_count++ == tsb_alloc_mtbf)) { 12947 tsb_alloc_count = 0; 12948 tsb_alloc_fail_mtbf++; 12949 return ((tsb_alloc_mtbf & 1)? ENOMEM : EAGAIN); 12950 } 12951 #endif /* DEBUG */ 12952 12953 /* 12954 * Enforce high water mark if we are not doing a forced allocation 12955 * and are not shrinking a process' TSB. 12956 */ 12957 if ((flags & TSB_SHRINK) == 0 && 12958 (tsbbytes + tsb_alloc_bytes) > tsb_alloc_hiwater) { 12959 if ((flags & TSB_FORCEALLOC) == 0) 12960 return (ENOMEM); 12961 lowmem = 1; 12962 } 12963 12964 /* 12965 * Allocate from the correct location based upon the size of the TSB 12966 * compared to the base page size, and what memory conditions dictate. 12967 * Note we always do nonblocking allocations from the TSB arena since 12968 * we don't want memory fragmentation to cause processes to block 12969 * indefinitely waiting for memory; until the kernel algorithms that 12970 * coalesce large pages are improved this is our best option. 12971 * 12972 * Algorithm: 12973 * If allocating a "large" TSB (>8K), allocate from the 12974 * appropriate kmem_tsb_default_arena vmem arena 12975 * else if low on memory or the TSB_FORCEALLOC flag is set or 12976 * tsb_forceheap is set 12977 * Allocate from kernel heap via sfmmu_tsb8k_cache with 12978 * KM_SLEEP (never fails) 12979 * else 12980 * Allocate from appropriate sfmmu_tsb_cache with 12981 * KM_NOSLEEP 12982 * endif 12983 */ 12984 if (tsb_lgrp_affinity) 12985 lgrpid = lgrp_home_id(curthread); 12986 if (lgrpid == LGRP_NONE) 12987 lgrpid = 0; /* use lgrp of boot CPU */ 12988 12989 if (tsbbytes > MMU_PAGESIZE) { 12990 if (tsbbytes > MMU_PAGESIZE4M) { 12991 vmp = kmem_bigtsb_default_arena[lgrpid]; 12992 vaddr = (caddr_t)vmem_xalloc(vmp, tsbbytes, tsbbytes, 12993 0, 0, NULL, NULL, VM_NOSLEEP); 12994 } else { 12995 vmp = kmem_tsb_default_arena[lgrpid]; 12996 vaddr = (caddr_t)vmem_xalloc(vmp, tsbbytes, tsbbytes, 12997 0, 0, NULL, NULL, VM_NOSLEEP); 12998 } 12999 #ifdef DEBUG 13000 } else if (lowmem || (flags & TSB_FORCEALLOC) || tsb_forceheap) { 13001 #else /* !DEBUG */ 13002 } else if (lowmem || (flags & TSB_FORCEALLOC)) { 13003 #endif /* DEBUG */ 13004 kmem_cachep = sfmmu_tsb8k_cache; 13005 vaddr = (caddr_t)kmem_cache_alloc(kmem_cachep, KM_SLEEP); 13006 ASSERT(vaddr != NULL); 13007 } else { 13008 kmem_cachep = sfmmu_tsb_cache[lgrpid]; 13009 vaddr = (caddr_t)kmem_cache_alloc(kmem_cachep, KM_NOSLEEP); 13010 } 13011 13012 tsbinfo->tsb_cache = kmem_cachep; 13013 tsbinfo->tsb_vmp = vmp; 13014 13015 if (vaddr == NULL) { 13016 return (EAGAIN); 13017 } 13018 13019 atomic_add_64(&tsb_alloc_bytes, (int64_t)tsbbytes); 13020 kmem_cachep = tsbinfo->tsb_cache; 13021 13022 /* 13023 * If we are allocating from outside the cage, then we need to 13024 * register a relocation callback handler. Note that for now 13025 * since pseudo mappings always hang off of the slab's root page, 13026 * we need only lock the first 8K of the TSB slab. This is a bit 13027 * hacky but it is good for performance. 13028 */ 13029 if (kmem_cachep != sfmmu_tsb8k_cache) { 13030 slab_vaddr = (caddr_t)((uintptr_t)vaddr & slab_mask); 13031 ret = as_pagelock(&kas, &pplist, slab_vaddr, PAGESIZE, S_WRITE); 13032 ASSERT(ret == 0); 13033 ret = hat_add_callback(sfmmu_tsb_cb_id, vaddr, (uint_t)tsbbytes, 13034 cbflags, (void *)tsbinfo, &pfn, NULL); 13035 13036 /* 13037 * Need to free up resources if we could not successfully 13038 * add the callback function and return an error condition. 13039 */ 13040 if (ret != 0) { 13041 if (kmem_cachep) { 13042 kmem_cache_free(kmem_cachep, vaddr); 13043 } else { 13044 vmem_xfree(vmp, (void *)vaddr, tsbbytes); 13045 } 13046 as_pageunlock(&kas, pplist, slab_vaddr, PAGESIZE, 13047 S_WRITE); 13048 return (EAGAIN); 13049 } 13050 } else { 13051 /* 13052 * Since allocation of 8K TSBs from heap is rare and occurs 13053 * during memory pressure we allocate them from permanent 13054 * memory rather than using callbacks to get the PFN. 13055 */ 13056 pfn = hat_getpfnum(kas.a_hat, vaddr); 13057 } 13058 13059 tsbinfo->tsb_va = vaddr; 13060 tsbinfo->tsb_szc = tsbcode; 13061 tsbinfo->tsb_ttesz_mask = tteszmask; 13062 tsbinfo->tsb_next = NULL; 13063 tsbinfo->tsb_flags = 0; 13064 13065 sfmmu_tsbinfo_setup_phys(tsbinfo, pfn); 13066 13067 sfmmu_inv_tsb(vaddr, tsbbytes); 13068 13069 if (kmem_cachep != sfmmu_tsb8k_cache) { 13070 as_pageunlock(&kas, pplist, slab_vaddr, PAGESIZE, S_WRITE); 13071 } 13072 13073 return (0); 13074 } 13075 13076 /* 13077 * Initialize per cpu tsb and per cpu tsbmiss_area 13078 */ 13079 void 13080 sfmmu_init_tsbs(void) 13081 { 13082 int i; 13083 struct tsbmiss *tsbmissp; 13084 struct kpmtsbm *kpmtsbmp; 13085 #ifndef sun4v 13086 extern int dcache_line_mask; 13087 #endif /* sun4v */ 13088 extern uint_t vac_colors; 13089 13090 /* 13091 * Init. tsb miss area. 13092 */ 13093 tsbmissp = tsbmiss_area; 13094 13095 for (i = 0; i < NCPU; tsbmissp++, i++) { 13096 /* 13097 * initialize the tsbmiss area. 13098 * Do this for all possible CPUs as some may be added 13099 * while the system is running. There is no cost to this. 13100 */ 13101 tsbmissp->ksfmmup = ksfmmup; 13102 #ifndef sun4v 13103 tsbmissp->dcache_line_mask = (uint16_t)dcache_line_mask; 13104 #endif /* sun4v */ 13105 tsbmissp->khashstart = 13106 (struct hmehash_bucket *)va_to_pa((caddr_t)khme_hash); 13107 tsbmissp->uhashstart = 13108 (struct hmehash_bucket *)va_to_pa((caddr_t)uhme_hash); 13109 tsbmissp->khashsz = khmehash_num; 13110 tsbmissp->uhashsz = uhmehash_num; 13111 } 13112 13113 sfmmu_tsb_cb_id = hat_register_callback('T'<<16 | 'S' << 8 | 'B', 13114 sfmmu_tsb_pre_relocator, sfmmu_tsb_post_relocator, NULL, 0); 13115 13116 if (kpm_enable == 0) 13117 return; 13118 13119 /* -- Begin KPM specific init -- */ 13120 13121 if (kpm_smallpages) { 13122 /* 13123 * If we're using base pagesize pages for seg_kpm 13124 * mappings, we use the kernel TSB since we can't afford 13125 * to allocate a second huge TSB for these mappings. 13126 */ 13127 kpm_tsbbase = ktsb_phys? ktsb_pbase : (uint64_t)ktsb_base; 13128 kpm_tsbsz = ktsb_szcode; 13129 kpmsm_tsbbase = kpm_tsbbase; 13130 kpmsm_tsbsz = kpm_tsbsz; 13131 } else { 13132 /* 13133 * In VAC conflict case, just put the entries in the 13134 * kernel 8K indexed TSB for now so we can find them. 13135 * This could really be changed in the future if we feel 13136 * the need... 13137 */ 13138 kpmsm_tsbbase = ktsb_phys? ktsb_pbase : (uint64_t)ktsb_base; 13139 kpmsm_tsbsz = ktsb_szcode; 13140 kpm_tsbbase = ktsb_phys? ktsb4m_pbase : (uint64_t)ktsb4m_base; 13141 kpm_tsbsz = ktsb4m_szcode; 13142 } 13143 13144 kpmtsbmp = kpmtsbm_area; 13145 for (i = 0; i < NCPU; kpmtsbmp++, i++) { 13146 /* 13147 * Initialize the kpmtsbm area. 13148 * Do this for all possible CPUs as some may be added 13149 * while the system is running. There is no cost to this. 13150 */ 13151 kpmtsbmp->vbase = kpm_vbase; 13152 kpmtsbmp->vend = kpm_vbase + kpm_size * vac_colors; 13153 kpmtsbmp->sz_shift = kpm_size_shift; 13154 kpmtsbmp->kpmp_shift = kpmp_shift; 13155 kpmtsbmp->kpmp2pshft = (uchar_t)kpmp2pshft; 13156 if (kpm_smallpages == 0) { 13157 kpmtsbmp->kpmp_table_sz = kpmp_table_sz; 13158 kpmtsbmp->kpmp_tablepa = va_to_pa(kpmp_table); 13159 } else { 13160 kpmtsbmp->kpmp_table_sz = kpmp_stable_sz; 13161 kpmtsbmp->kpmp_tablepa = va_to_pa(kpmp_stable); 13162 } 13163 kpmtsbmp->msegphashpa = va_to_pa(memseg_phash); 13164 kpmtsbmp->flags = KPMTSBM_ENABLE_FLAG; 13165 #ifdef DEBUG 13166 kpmtsbmp->flags |= (kpm_tsbmtl) ? KPMTSBM_TLTSBM_FLAG : 0; 13167 #endif /* DEBUG */ 13168 if (ktsb_phys) 13169 kpmtsbmp->flags |= KPMTSBM_TSBPHYS_FLAG; 13170 } 13171 13172 /* -- End KPM specific init -- */ 13173 } 13174 13175 /* Avoid using sfmmu_tsbinfo_alloc() to avoid kmem_alloc - no real reason */ 13176 struct tsb_info ktsb_info[2]; 13177 13178 /* 13179 * Called from hat_kern_setup() to setup the tsb_info for ksfmmup. 13180 */ 13181 void 13182 sfmmu_init_ktsbinfo() 13183 { 13184 ASSERT(ksfmmup != NULL); 13185 ASSERT(ksfmmup->sfmmu_tsb == NULL); 13186 /* 13187 * Allocate tsbinfos for kernel and copy in data 13188 * to make debug easier and sun4v setup easier. 13189 */ 13190 ktsb_info[0].tsb_sfmmu = ksfmmup; 13191 ktsb_info[0].tsb_szc = ktsb_szcode; 13192 ktsb_info[0].tsb_ttesz_mask = TSB8K|TSB64K|TSB512K; 13193 ktsb_info[0].tsb_va = ktsb_base; 13194 ktsb_info[0].tsb_pa = ktsb_pbase; 13195 ktsb_info[0].tsb_flags = 0; 13196 ktsb_info[0].tsb_tte.ll = 0; 13197 ktsb_info[0].tsb_cache = NULL; 13198 13199 ktsb_info[1].tsb_sfmmu = ksfmmup; 13200 ktsb_info[1].tsb_szc = ktsb4m_szcode; 13201 ktsb_info[1].tsb_ttesz_mask = TSB4M; 13202 ktsb_info[1].tsb_va = ktsb4m_base; 13203 ktsb_info[1].tsb_pa = ktsb4m_pbase; 13204 ktsb_info[1].tsb_flags = 0; 13205 ktsb_info[1].tsb_tte.ll = 0; 13206 ktsb_info[1].tsb_cache = NULL; 13207 13208 /* Link them into ksfmmup. */ 13209 ktsb_info[0].tsb_next = &ktsb_info[1]; 13210 ktsb_info[1].tsb_next = NULL; 13211 ksfmmup->sfmmu_tsb = &ktsb_info[0]; 13212 13213 sfmmu_setup_tsbinfo(ksfmmup); 13214 } 13215 13216 /* 13217 * Cache the last value returned from va_to_pa(). If the VA specified 13218 * in the current call to cached_va_to_pa() maps to the same Page (as the 13219 * previous call to cached_va_to_pa()), then compute the PA using 13220 * cached info, else call va_to_pa(). 13221 * 13222 * Note: this function is neither MT-safe nor consistent in the presence 13223 * of multiple, interleaved threads. This function was created to enable 13224 * an optimization used during boot (at a point when there's only one thread 13225 * executing on the "boot CPU", and before startup_vm() has been called). 13226 */ 13227 static uint64_t 13228 cached_va_to_pa(void *vaddr) 13229 { 13230 static uint64_t prev_vaddr_base = 0; 13231 static uint64_t prev_pfn = 0; 13232 13233 if ((((uint64_t)vaddr) & MMU_PAGEMASK) == prev_vaddr_base) { 13234 return (prev_pfn | ((uint64_t)vaddr & MMU_PAGEOFFSET)); 13235 } else { 13236 uint64_t pa = va_to_pa(vaddr); 13237 13238 if (pa != ((uint64_t)-1)) { 13239 /* 13240 * Computed physical address is valid. Cache its 13241 * related info for the next cached_va_to_pa() call. 13242 */ 13243 prev_pfn = pa & MMU_PAGEMASK; 13244 prev_vaddr_base = ((uint64_t)vaddr) & MMU_PAGEMASK; 13245 } 13246 13247 return (pa); 13248 } 13249 } 13250 13251 /* 13252 * Carve up our nucleus hblk region. We may allocate more hblks than 13253 * asked due to rounding errors but we are guaranteed to have at least 13254 * enough space to allocate the requested number of hblk8's and hblk1's. 13255 */ 13256 void 13257 sfmmu_init_nucleus_hblks(caddr_t addr, size_t size, int nhblk8, int nhblk1) 13258 { 13259 struct hme_blk *hmeblkp; 13260 size_t hme8blk_sz, hme1blk_sz; 13261 size_t i; 13262 size_t hblk8_bound; 13263 ulong_t j = 0, k = 0; 13264 13265 ASSERT(addr != NULL && size != 0); 13266 13267 /* Need to use proper structure alignment */ 13268 hme8blk_sz = roundup(HME8BLK_SZ, sizeof (int64_t)); 13269 hme1blk_sz = roundup(HME1BLK_SZ, sizeof (int64_t)); 13270 13271 nucleus_hblk8.list = (void *)addr; 13272 nucleus_hblk8.index = 0; 13273 13274 /* 13275 * Use as much memory as possible for hblk8's since we 13276 * expect all bop_alloc'ed memory to be allocated in 8k chunks. 13277 * We need to hold back enough space for the hblk1's which 13278 * we'll allocate next. 13279 */ 13280 hblk8_bound = size - (nhblk1 * hme1blk_sz) - hme8blk_sz; 13281 for (i = 0; i <= hblk8_bound; i += hme8blk_sz, j++) { 13282 hmeblkp = (struct hme_blk *)addr; 13283 addr += hme8blk_sz; 13284 hmeblkp->hblk_nuc_bit = 1; 13285 hmeblkp->hblk_nextpa = cached_va_to_pa((caddr_t)hmeblkp); 13286 } 13287 nucleus_hblk8.len = j; 13288 ASSERT(j >= nhblk8); 13289 SFMMU_STAT_ADD(sf_hblk8_ncreate, j); 13290 13291 nucleus_hblk1.list = (void *)addr; 13292 nucleus_hblk1.index = 0; 13293 for (; i <= (size - hme1blk_sz); i += hme1blk_sz, k++) { 13294 hmeblkp = (struct hme_blk *)addr; 13295 addr += hme1blk_sz; 13296 hmeblkp->hblk_nuc_bit = 1; 13297 hmeblkp->hblk_nextpa = cached_va_to_pa((caddr_t)hmeblkp); 13298 } 13299 ASSERT(k >= nhblk1); 13300 nucleus_hblk1.len = k; 13301 SFMMU_STAT_ADD(sf_hblk1_ncreate, k); 13302 } 13303 13304 /* 13305 * This function is currently not supported on this platform. For what 13306 * it's supposed to do, see hat.c and hat_srmmu.c 13307 */ 13308 /* ARGSUSED */ 13309 faultcode_t 13310 hat_softlock(struct hat *hat, caddr_t addr, size_t *lenp, page_t **ppp, 13311 uint_t flags) 13312 { 13313 ASSERT(hat->sfmmu_xhat_provider == NULL); 13314 return (FC_NOSUPPORT); 13315 } 13316 13317 /* 13318 * Searchs the mapping list of the page for a mapping of the same size. If not 13319 * found the corresponding bit is cleared in the p_index field. When large 13320 * pages are more prevalent in the system, we can maintain the mapping list 13321 * in order and we don't have to traverse the list each time. Just check the 13322 * next and prev entries, and if both are of different size, we clear the bit. 13323 */ 13324 static void 13325 sfmmu_rm_large_mappings(page_t *pp, int ttesz) 13326 { 13327 struct sf_hment *sfhmep; 13328 struct hme_blk *hmeblkp; 13329 int index; 13330 pgcnt_t npgs; 13331 13332 ASSERT(ttesz > TTE8K); 13333 13334 ASSERT(sfmmu_mlist_held(pp)); 13335 13336 ASSERT(PP_ISMAPPED_LARGE(pp)); 13337 13338 /* 13339 * Traverse mapping list looking for another mapping of same size. 13340 * since we only want to clear index field if all mappings of 13341 * that size are gone. 13342 */ 13343 13344 for (sfhmep = pp->p_mapping; sfhmep; sfhmep = sfhmep->hme_next) { 13345 if (IS_PAHME(sfhmep)) 13346 continue; 13347 hmeblkp = sfmmu_hmetohblk(sfhmep); 13348 if (hmeblkp->hblk_xhat_bit) 13349 continue; 13350 if (hme_size(sfhmep) == ttesz) { 13351 /* 13352 * another mapping of the same size. don't clear index. 13353 */ 13354 return; 13355 } 13356 } 13357 13358 /* 13359 * Clear the p_index bit for large page. 13360 */ 13361 index = PAGESZ_TO_INDEX(ttesz); 13362 npgs = TTEPAGES(ttesz); 13363 while (npgs-- > 0) { 13364 ASSERT(pp->p_index & index); 13365 pp->p_index &= ~index; 13366 pp = PP_PAGENEXT(pp); 13367 } 13368 } 13369 13370 /* 13371 * return supported features 13372 */ 13373 /* ARGSUSED */ 13374 int 13375 hat_supported(enum hat_features feature, void *arg) 13376 { 13377 switch (feature) { 13378 case HAT_SHARED_PT: 13379 case HAT_DYNAMIC_ISM_UNMAP: 13380 case HAT_VMODSORT: 13381 return (1); 13382 case HAT_SHARED_REGIONS: 13383 if (shctx_on) 13384 return (1); 13385 else 13386 return (0); 13387 default: 13388 return (0); 13389 } 13390 } 13391 13392 void 13393 hat_enter(struct hat *hat) 13394 { 13395 hatlock_t *hatlockp; 13396 13397 if (hat != ksfmmup) { 13398 hatlockp = TSB_HASH(hat); 13399 mutex_enter(HATLOCK_MUTEXP(hatlockp)); 13400 } 13401 } 13402 13403 void 13404 hat_exit(struct hat *hat) 13405 { 13406 hatlock_t *hatlockp; 13407 13408 if (hat != ksfmmup) { 13409 hatlockp = TSB_HASH(hat); 13410 mutex_exit(HATLOCK_MUTEXP(hatlockp)); 13411 } 13412 } 13413 13414 /*ARGSUSED*/ 13415 void 13416 hat_reserve(struct as *as, caddr_t addr, size_t len) 13417 { 13418 } 13419 13420 static void 13421 hat_kstat_init(void) 13422 { 13423 kstat_t *ksp; 13424 13425 ksp = kstat_create("unix", 0, "sfmmu_global_stat", "hat", 13426 KSTAT_TYPE_RAW, sizeof (struct sfmmu_global_stat), 13427 KSTAT_FLAG_VIRTUAL); 13428 if (ksp) { 13429 ksp->ks_data = (void *) &sfmmu_global_stat; 13430 kstat_install(ksp); 13431 } 13432 ksp = kstat_create("unix", 0, "sfmmu_tsbsize_stat", "hat", 13433 KSTAT_TYPE_RAW, sizeof (struct sfmmu_tsbsize_stat), 13434 KSTAT_FLAG_VIRTUAL); 13435 if (ksp) { 13436 ksp->ks_data = (void *) &sfmmu_tsbsize_stat; 13437 kstat_install(ksp); 13438 } 13439 ksp = kstat_create("unix", 0, "sfmmu_percpu_stat", "hat", 13440 KSTAT_TYPE_RAW, sizeof (struct sfmmu_percpu_stat) * NCPU, 13441 KSTAT_FLAG_WRITABLE); 13442 if (ksp) { 13443 ksp->ks_update = sfmmu_kstat_percpu_update; 13444 kstat_install(ksp); 13445 } 13446 } 13447 13448 /* ARGSUSED */ 13449 static int 13450 sfmmu_kstat_percpu_update(kstat_t *ksp, int rw) 13451 { 13452 struct sfmmu_percpu_stat *cpu_kstat = ksp->ks_data; 13453 struct tsbmiss *tsbm = tsbmiss_area; 13454 struct kpmtsbm *kpmtsbm = kpmtsbm_area; 13455 int i; 13456 13457 ASSERT(cpu_kstat); 13458 if (rw == KSTAT_READ) { 13459 for (i = 0; i < NCPU; cpu_kstat++, tsbm++, kpmtsbm++, i++) { 13460 cpu_kstat->sf_itlb_misses = 0; 13461 cpu_kstat->sf_dtlb_misses = 0; 13462 cpu_kstat->sf_utsb_misses = tsbm->utsb_misses - 13463 tsbm->uprot_traps; 13464 cpu_kstat->sf_ktsb_misses = tsbm->ktsb_misses + 13465 kpmtsbm->kpm_tsb_misses - tsbm->kprot_traps; 13466 cpu_kstat->sf_tsb_hits = 0; 13467 cpu_kstat->sf_umod_faults = tsbm->uprot_traps; 13468 cpu_kstat->sf_kmod_faults = tsbm->kprot_traps; 13469 } 13470 } else { 13471 /* KSTAT_WRITE is used to clear stats */ 13472 for (i = 0; i < NCPU; tsbm++, kpmtsbm++, i++) { 13473 tsbm->utsb_misses = 0; 13474 tsbm->ktsb_misses = 0; 13475 tsbm->uprot_traps = 0; 13476 tsbm->kprot_traps = 0; 13477 kpmtsbm->kpm_dtlb_misses = 0; 13478 kpmtsbm->kpm_tsb_misses = 0; 13479 } 13480 } 13481 return (0); 13482 } 13483 13484 #ifdef DEBUG 13485 13486 tte_t *gorig[NCPU], *gcur[NCPU], *gnew[NCPU]; 13487 13488 /* 13489 * A tte checker. *orig_old is the value we read before cas. 13490 * *cur is the value returned by cas. 13491 * *new is the desired value when we do the cas. 13492 * 13493 * *hmeblkp is currently unused. 13494 */ 13495 13496 /* ARGSUSED */ 13497 void 13498 chk_tte(tte_t *orig_old, tte_t *cur, tte_t *new, struct hme_blk *hmeblkp) 13499 { 13500 pfn_t i, j, k; 13501 int cpuid = CPU->cpu_id; 13502 13503 gorig[cpuid] = orig_old; 13504 gcur[cpuid] = cur; 13505 gnew[cpuid] = new; 13506 13507 #ifdef lint 13508 hmeblkp = hmeblkp; 13509 #endif 13510 13511 if (TTE_IS_VALID(orig_old)) { 13512 if (TTE_IS_VALID(cur)) { 13513 i = TTE_TO_TTEPFN(orig_old); 13514 j = TTE_TO_TTEPFN(cur); 13515 k = TTE_TO_TTEPFN(new); 13516 if (i != j) { 13517 /* remap error? */ 13518 panic("chk_tte: bad pfn, 0x%lx, 0x%lx", i, j); 13519 } 13520 13521 if (i != k) { 13522 /* remap error? */ 13523 panic("chk_tte: bad pfn2, 0x%lx, 0x%lx", i, k); 13524 } 13525 } else { 13526 if (TTE_IS_VALID(new)) { 13527 panic("chk_tte: invalid cur? "); 13528 } 13529 13530 i = TTE_TO_TTEPFN(orig_old); 13531 k = TTE_TO_TTEPFN(new); 13532 if (i != k) { 13533 panic("chk_tte: bad pfn3, 0x%lx, 0x%lx", i, k); 13534 } 13535 } 13536 } else { 13537 if (TTE_IS_VALID(cur)) { 13538 j = TTE_TO_TTEPFN(cur); 13539 if (TTE_IS_VALID(new)) { 13540 k = TTE_TO_TTEPFN(new); 13541 if (j != k) { 13542 panic("chk_tte: bad pfn4, 0x%lx, 0x%lx", 13543 j, k); 13544 } 13545 } else { 13546 panic("chk_tte: why here?"); 13547 } 13548 } else { 13549 if (!TTE_IS_VALID(new)) { 13550 panic("chk_tte: why here2 ?"); 13551 } 13552 } 13553 } 13554 } 13555 13556 #endif /* DEBUG */ 13557 13558 extern void prefetch_tsbe_read(struct tsbe *); 13559 extern void prefetch_tsbe_write(struct tsbe *); 13560 13561 13562 /* 13563 * We want to prefetch 7 cache lines ahead for our read prefetch. This gives 13564 * us optimal performance on Cheetah+. You can only have 8 outstanding 13565 * prefetches at any one time, so we opted for 7 read prefetches and 1 write 13566 * prefetch to make the most utilization of the prefetch capability. 13567 */ 13568 #define TSBE_PREFETCH_STRIDE (7) 13569 13570 void 13571 sfmmu_copy_tsb(struct tsb_info *old_tsbinfo, struct tsb_info *new_tsbinfo) 13572 { 13573 int old_bytes = TSB_BYTES(old_tsbinfo->tsb_szc); 13574 int new_bytes = TSB_BYTES(new_tsbinfo->tsb_szc); 13575 int old_entries = TSB_ENTRIES(old_tsbinfo->tsb_szc); 13576 int new_entries = TSB_ENTRIES(new_tsbinfo->tsb_szc); 13577 struct tsbe *old; 13578 struct tsbe *new; 13579 struct tsbe *new_base = (struct tsbe *)new_tsbinfo->tsb_va; 13580 uint64_t va; 13581 int new_offset; 13582 int i; 13583 int vpshift; 13584 int last_prefetch; 13585 13586 if (old_bytes == new_bytes) { 13587 bcopy(old_tsbinfo->tsb_va, new_tsbinfo->tsb_va, new_bytes); 13588 } else { 13589 13590 /* 13591 * A TSBE is 16 bytes which means there are four TSBE's per 13592 * P$ line (64 bytes), thus every 4 TSBE's we prefetch. 13593 */ 13594 old = (struct tsbe *)old_tsbinfo->tsb_va; 13595 last_prefetch = old_entries - (4*(TSBE_PREFETCH_STRIDE+1)); 13596 for (i = 0; i < old_entries; i++, old++) { 13597 if (((i & (4-1)) == 0) && (i < last_prefetch)) 13598 prefetch_tsbe_read(old); 13599 if (!old->tte_tag.tag_invalid) { 13600 /* 13601 * We have a valid TTE to remap. Check the 13602 * size. We won't remap 64K or 512K TTEs 13603 * because they span more than one TSB entry 13604 * and are indexed using an 8K virt. page. 13605 * Ditto for 32M and 256M TTEs. 13606 */ 13607 if (TTE_CSZ(&old->tte_data) == TTE64K || 13608 TTE_CSZ(&old->tte_data) == TTE512K) 13609 continue; 13610 if (mmu_page_sizes == max_mmu_page_sizes) { 13611 if (TTE_CSZ(&old->tte_data) == TTE32M || 13612 TTE_CSZ(&old->tte_data) == TTE256M) 13613 continue; 13614 } 13615 13616 /* clear the lower 22 bits of the va */ 13617 va = *(uint64_t *)old << 22; 13618 /* turn va into a virtual pfn */ 13619 va >>= 22 - TSB_START_SIZE; 13620 /* 13621 * or in bits from the offset in the tsb 13622 * to get the real virtual pfn. These 13623 * correspond to bits [21:13] in the va 13624 */ 13625 vpshift = 13626 TTE_BSZS_SHIFT(TTE_CSZ(&old->tte_data)) & 13627 0x1ff; 13628 va |= (i << vpshift); 13629 va >>= vpshift; 13630 new_offset = va & (new_entries - 1); 13631 new = new_base + new_offset; 13632 prefetch_tsbe_write(new); 13633 *new = *old; 13634 } 13635 } 13636 } 13637 } 13638 13639 /* 13640 * unused in sfmmu 13641 */ 13642 void 13643 hat_dump(void) 13644 { 13645 } 13646 13647 /* 13648 * Called when a thread is exiting and we have switched to the kernel address 13649 * space. Perform the same VM initialization resume() uses when switching 13650 * processes. 13651 * 13652 * Note that sfmmu_load_mmustate() is currently a no-op for kernel threads, but 13653 * we call it anyway in case the semantics change in the future. 13654 */ 13655 /*ARGSUSED*/ 13656 void 13657 hat_thread_exit(kthread_t *thd) 13658 { 13659 uint_t pgsz_cnum; 13660 uint_t pstate_save; 13661 13662 ASSERT(thd->t_procp->p_as == &kas); 13663 13664 pgsz_cnum = KCONTEXT; 13665 #ifdef sun4u 13666 pgsz_cnum |= (ksfmmup->sfmmu_cext << CTXREG_EXT_SHIFT); 13667 #endif 13668 13669 /* 13670 * Note that sfmmu_load_mmustate() is currently a no-op for 13671 * kernel threads. We need to disable interrupts here, 13672 * simply because otherwise sfmmu_load_mmustate() would panic 13673 * if the caller does not disable interrupts. 13674 */ 13675 pstate_save = sfmmu_disable_intrs(); 13676 13677 /* Compatibility Note: hw takes care of MMU_SCONTEXT1 */ 13678 sfmmu_setctx_sec(pgsz_cnum); 13679 sfmmu_load_mmustate(ksfmmup); 13680 sfmmu_enable_intrs(pstate_save); 13681 } 13682 13683 13684 /* 13685 * SRD support 13686 */ 13687 #define SRD_HASH_FUNCTION(vp) (((((uintptr_t)(vp)) >> 4) ^ \ 13688 (((uintptr_t)(vp)) >> 11)) & \ 13689 srd_hashmask) 13690 13691 /* 13692 * Attach the process to the srd struct associated with the exec vnode 13693 * from which the process is started. 13694 */ 13695 void 13696 hat_join_srd(struct hat *sfmmup, vnode_t *evp) 13697 { 13698 uint_t hash = SRD_HASH_FUNCTION(evp); 13699 sf_srd_t *srdp; 13700 sf_srd_t *newsrdp; 13701 13702 ASSERT(sfmmup != ksfmmup); 13703 ASSERT(sfmmup->sfmmu_srdp == NULL); 13704 13705 if (!shctx_on) { 13706 return; 13707 } 13708 13709 VN_HOLD(evp); 13710 13711 if (srd_buckets[hash].srdb_srdp != NULL) { 13712 mutex_enter(&srd_buckets[hash].srdb_lock); 13713 for (srdp = srd_buckets[hash].srdb_srdp; srdp != NULL; 13714 srdp = srdp->srd_hash) { 13715 if (srdp->srd_evp == evp) { 13716 ASSERT(srdp->srd_refcnt >= 0); 13717 sfmmup->sfmmu_srdp = srdp; 13718 atomic_add_32( 13719 (volatile uint_t *)&srdp->srd_refcnt, 1); 13720 mutex_exit(&srd_buckets[hash].srdb_lock); 13721 return; 13722 } 13723 } 13724 mutex_exit(&srd_buckets[hash].srdb_lock); 13725 } 13726 newsrdp = kmem_cache_alloc(srd_cache, KM_SLEEP); 13727 ASSERT(newsrdp->srd_next_ismrid == 0 && newsrdp->srd_next_hmerid == 0); 13728 13729 newsrdp->srd_evp = evp; 13730 newsrdp->srd_refcnt = 1; 13731 newsrdp->srd_hmergnfree = NULL; 13732 newsrdp->srd_ismrgnfree = NULL; 13733 13734 mutex_enter(&srd_buckets[hash].srdb_lock); 13735 for (srdp = srd_buckets[hash].srdb_srdp; srdp != NULL; 13736 srdp = srdp->srd_hash) { 13737 if (srdp->srd_evp == evp) { 13738 ASSERT(srdp->srd_refcnt >= 0); 13739 sfmmup->sfmmu_srdp = srdp; 13740 atomic_add_32((volatile uint_t *)&srdp->srd_refcnt, 1); 13741 mutex_exit(&srd_buckets[hash].srdb_lock); 13742 kmem_cache_free(srd_cache, newsrdp); 13743 return; 13744 } 13745 } 13746 newsrdp->srd_hash = srd_buckets[hash].srdb_srdp; 13747 srd_buckets[hash].srdb_srdp = newsrdp; 13748 sfmmup->sfmmu_srdp = newsrdp; 13749 13750 mutex_exit(&srd_buckets[hash].srdb_lock); 13751 13752 } 13753 13754 static void 13755 sfmmu_leave_srd(sfmmu_t *sfmmup) 13756 { 13757 vnode_t *evp; 13758 sf_srd_t *srdp = sfmmup->sfmmu_srdp; 13759 uint_t hash; 13760 sf_srd_t **prev_srdpp; 13761 sf_region_t *rgnp; 13762 sf_region_t *nrgnp; 13763 #ifdef DEBUG 13764 int rgns = 0; 13765 #endif 13766 int i; 13767 13768 ASSERT(sfmmup != ksfmmup); 13769 ASSERT(srdp != NULL); 13770 ASSERT(srdp->srd_refcnt > 0); 13771 ASSERT(sfmmup->sfmmu_scdp == NULL); 13772 ASSERT(sfmmup->sfmmu_free == 1); 13773 13774 sfmmup->sfmmu_srdp = NULL; 13775 evp = srdp->srd_evp; 13776 ASSERT(evp != NULL); 13777 if (atomic_add_32_nv( 13778 (volatile uint_t *)&srdp->srd_refcnt, -1)) { 13779 VN_RELE(evp); 13780 return; 13781 } 13782 13783 hash = SRD_HASH_FUNCTION(evp); 13784 mutex_enter(&srd_buckets[hash].srdb_lock); 13785 for (prev_srdpp = &srd_buckets[hash].srdb_srdp; 13786 (srdp = *prev_srdpp) != NULL; prev_srdpp = &srdp->srd_hash) { 13787 if (srdp->srd_evp == evp) { 13788 break; 13789 } 13790 } 13791 if (srdp == NULL || srdp->srd_refcnt) { 13792 mutex_exit(&srd_buckets[hash].srdb_lock); 13793 VN_RELE(evp); 13794 return; 13795 } 13796 *prev_srdpp = srdp->srd_hash; 13797 mutex_exit(&srd_buckets[hash].srdb_lock); 13798 13799 ASSERT(srdp->srd_refcnt == 0); 13800 VN_RELE(evp); 13801 13802 #ifdef DEBUG 13803 for (i = 0; i < SFMMU_MAX_REGION_BUCKETS; i++) { 13804 ASSERT(srdp->srd_rgnhash[i] == NULL); 13805 } 13806 #endif /* DEBUG */ 13807 13808 /* free each hme regions in the srd */ 13809 for (rgnp = srdp->srd_hmergnfree; rgnp != NULL; rgnp = nrgnp) { 13810 nrgnp = rgnp->rgn_next; 13811 ASSERT(rgnp->rgn_id < srdp->srd_next_hmerid); 13812 ASSERT(rgnp->rgn_refcnt == 0); 13813 ASSERT(rgnp->rgn_sfmmu_head == NULL); 13814 ASSERT(rgnp->rgn_flags & SFMMU_REGION_FREE); 13815 ASSERT(rgnp->rgn_hmeflags == 0); 13816 ASSERT(srdp->srd_hmergnp[rgnp->rgn_id] == rgnp); 13817 #ifdef DEBUG 13818 for (i = 0; i < MMU_PAGE_SIZES; i++) { 13819 ASSERT(rgnp->rgn_ttecnt[i] == 0); 13820 } 13821 rgns++; 13822 #endif /* DEBUG */ 13823 kmem_cache_free(region_cache, rgnp); 13824 } 13825 ASSERT(rgns == srdp->srd_next_hmerid); 13826 13827 #ifdef DEBUG 13828 rgns = 0; 13829 #endif 13830 /* free each ism rgns in the srd */ 13831 for (rgnp = srdp->srd_ismrgnfree; rgnp != NULL; rgnp = nrgnp) { 13832 nrgnp = rgnp->rgn_next; 13833 ASSERT(rgnp->rgn_id < srdp->srd_next_ismrid); 13834 ASSERT(rgnp->rgn_refcnt == 0); 13835 ASSERT(rgnp->rgn_sfmmu_head == NULL); 13836 ASSERT(rgnp->rgn_flags & SFMMU_REGION_FREE); 13837 ASSERT(srdp->srd_ismrgnp[rgnp->rgn_id] == rgnp); 13838 #ifdef DEBUG 13839 for (i = 0; i < MMU_PAGE_SIZES; i++) { 13840 ASSERT(rgnp->rgn_ttecnt[i] == 0); 13841 } 13842 rgns++; 13843 #endif /* DEBUG */ 13844 kmem_cache_free(region_cache, rgnp); 13845 } 13846 ASSERT(rgns == srdp->srd_next_ismrid); 13847 ASSERT(srdp->srd_ismbusyrgns == 0); 13848 ASSERT(srdp->srd_hmebusyrgns == 0); 13849 13850 srdp->srd_next_ismrid = 0; 13851 srdp->srd_next_hmerid = 0; 13852 13853 bzero((void *)srdp->srd_ismrgnp, 13854 sizeof (sf_region_t *) * SFMMU_MAX_ISM_REGIONS); 13855 bzero((void *)srdp->srd_hmergnp, 13856 sizeof (sf_region_t *) * SFMMU_MAX_HME_REGIONS); 13857 13858 ASSERT(srdp->srd_scdp == NULL); 13859 kmem_cache_free(srd_cache, srdp); 13860 } 13861 13862 /* ARGSUSED */ 13863 static int 13864 sfmmu_srdcache_constructor(void *buf, void *cdrarg, int kmflags) 13865 { 13866 sf_srd_t *srdp = (sf_srd_t *)buf; 13867 bzero(buf, sizeof (*srdp)); 13868 13869 mutex_init(&srdp->srd_mutex, NULL, MUTEX_DEFAULT, NULL); 13870 mutex_init(&srdp->srd_scd_mutex, NULL, MUTEX_DEFAULT, NULL); 13871 return (0); 13872 } 13873 13874 /* ARGSUSED */ 13875 static void 13876 sfmmu_srdcache_destructor(void *buf, void *cdrarg) 13877 { 13878 sf_srd_t *srdp = (sf_srd_t *)buf; 13879 13880 mutex_destroy(&srdp->srd_mutex); 13881 mutex_destroy(&srdp->srd_scd_mutex); 13882 } 13883 13884 /* 13885 * The caller makes sure hat_join_region()/hat_leave_region() can't be called 13886 * at the same time for the same process and address range. This is ensured by 13887 * the fact that address space is locked as writer when a process joins the 13888 * regions. Therefore there's no need to hold an srd lock during the entire 13889 * execution of hat_join_region()/hat_leave_region(). 13890 */ 13891 13892 #define RGN_HASH_FUNCTION(obj) (((((uintptr_t)(obj)) >> 4) ^ \ 13893 (((uintptr_t)(obj)) >> 11)) & \ 13894 srd_rgn_hashmask) 13895 /* 13896 * This routine implements the shared context functionality required when 13897 * attaching a segment to an address space. It must be called from 13898 * hat_share() for D(ISM) segments and from segvn_create() for segments 13899 * with the MAP_PRIVATE and MAP_TEXT flags set. It returns a region_cookie 13900 * which is saved in the private segment data for hme segments and 13901 * the ism_map structure for ism segments. 13902 */ 13903 hat_region_cookie_t 13904 hat_join_region(struct hat *sfmmup, 13905 caddr_t r_saddr, 13906 size_t r_size, 13907 void *r_obj, 13908 u_offset_t r_objoff, 13909 uchar_t r_perm, 13910 uchar_t r_pgszc, 13911 hat_rgn_cb_func_t r_cb_function, 13912 uint_t flags) 13913 { 13914 sf_srd_t *srdp = sfmmup->sfmmu_srdp; 13915 uint_t rhash; 13916 uint_t rid; 13917 hatlock_t *hatlockp; 13918 sf_region_t *rgnp; 13919 sf_region_t *new_rgnp = NULL; 13920 int i; 13921 uint16_t *nextidp; 13922 sf_region_t **freelistp; 13923 int maxids; 13924 sf_region_t **rarrp; 13925 uint16_t *busyrgnsp; 13926 ulong_t rttecnt; 13927 uchar_t tteflag; 13928 uchar_t r_type = flags & HAT_REGION_TYPE_MASK; 13929 int text = (r_type == HAT_REGION_TEXT); 13930 13931 if (srdp == NULL || r_size == 0) { 13932 return (HAT_INVALID_REGION_COOKIE); 13933 } 13934 13935 ASSERT(sfmmup->sfmmu_xhat_provider == NULL); 13936 ASSERT(sfmmup != ksfmmup); 13937 ASSERT(AS_WRITE_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock)); 13938 ASSERT(srdp->srd_refcnt > 0); 13939 ASSERT(!(flags & ~HAT_REGION_TYPE_MASK)); 13940 ASSERT(flags == HAT_REGION_TEXT || flags == HAT_REGION_ISM); 13941 ASSERT(r_pgszc < mmu_page_sizes); 13942 if (!IS_P2ALIGNED(r_saddr, TTEBYTES(r_pgszc)) || 13943 !IS_P2ALIGNED(r_size, TTEBYTES(r_pgszc))) { 13944 panic("hat_join_region: region addr or size is not aligned\n"); 13945 } 13946 13947 13948 r_type = (r_type == HAT_REGION_ISM) ? SFMMU_REGION_ISM : 13949 SFMMU_REGION_HME; 13950 /* 13951 * Currently only support shared hmes for the read only main text 13952 * region. 13953 */ 13954 if (r_type == SFMMU_REGION_HME && ((r_obj != srdp->srd_evp) || 13955 (r_perm & PROT_WRITE))) { 13956 return (HAT_INVALID_REGION_COOKIE); 13957 } 13958 13959 rhash = RGN_HASH_FUNCTION(r_obj); 13960 13961 if (r_type == SFMMU_REGION_ISM) { 13962 nextidp = &srdp->srd_next_ismrid; 13963 freelistp = &srdp->srd_ismrgnfree; 13964 maxids = SFMMU_MAX_ISM_REGIONS; 13965 rarrp = srdp->srd_ismrgnp; 13966 busyrgnsp = &srdp->srd_ismbusyrgns; 13967 } else { 13968 nextidp = &srdp->srd_next_hmerid; 13969 freelistp = &srdp->srd_hmergnfree; 13970 maxids = SFMMU_MAX_HME_REGIONS; 13971 rarrp = srdp->srd_hmergnp; 13972 busyrgnsp = &srdp->srd_hmebusyrgns; 13973 } 13974 13975 mutex_enter(&srdp->srd_mutex); 13976 13977 for (rgnp = srdp->srd_rgnhash[rhash]; rgnp != NULL; 13978 rgnp = rgnp->rgn_hash) { 13979 if (rgnp->rgn_saddr == r_saddr && rgnp->rgn_size == r_size && 13980 rgnp->rgn_obj == r_obj && rgnp->rgn_objoff == r_objoff && 13981 rgnp->rgn_perm == r_perm && rgnp->rgn_pgszc == r_pgszc) { 13982 break; 13983 } 13984 } 13985 13986 rfound: 13987 if (rgnp != NULL) { 13988 ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK) == r_type); 13989 ASSERT(rgnp->rgn_cb_function == r_cb_function); 13990 ASSERT(rgnp->rgn_refcnt >= 0); 13991 rid = rgnp->rgn_id; 13992 ASSERT(rid < maxids); 13993 ASSERT(rarrp[rid] == rgnp); 13994 ASSERT(rid < *nextidp); 13995 atomic_add_32((volatile uint_t *)&rgnp->rgn_refcnt, 1); 13996 mutex_exit(&srdp->srd_mutex); 13997 if (new_rgnp != NULL) { 13998 kmem_cache_free(region_cache, new_rgnp); 13999 } 14000 if (r_type == SFMMU_REGION_HME) { 14001 int myjoin = 14002 (sfmmup == astosfmmu(curthread->t_procp->p_as)); 14003 14004 sfmmu_link_to_hmeregion(sfmmup, rgnp); 14005 /* 14006 * bitmap should be updated after linking sfmmu on 14007 * region list so that pageunload() doesn't skip 14008 * TSB/TLB flush. As soon as bitmap is updated another 14009 * thread in this process can already start accessing 14010 * this region. 14011 */ 14012 /* 14013 * Normally ttecnt accounting is done as part of 14014 * pagefault handling. But a process may not take any 14015 * pagefaults on shared hmeblks created by some other 14016 * process. To compensate for this assume that the 14017 * entire region will end up faulted in using 14018 * the region's pagesize. 14019 * 14020 */ 14021 if (r_pgszc > TTE8K) { 14022 tteflag = 1 << r_pgszc; 14023 if (disable_large_pages & tteflag) { 14024 tteflag = 0; 14025 } 14026 } else { 14027 tteflag = 0; 14028 } 14029 if (tteflag && !(sfmmup->sfmmu_rtteflags & tteflag)) { 14030 hatlockp = sfmmu_hat_enter(sfmmup); 14031 sfmmup->sfmmu_rtteflags |= tteflag; 14032 sfmmu_hat_exit(hatlockp); 14033 } 14034 hatlockp = sfmmu_hat_enter(sfmmup); 14035 14036 /* 14037 * Preallocate 1/4 of ttecnt's in 8K TSB for >= 4M 14038 * region to allow for large page allocation failure. 14039 */ 14040 if (r_pgszc >= TTE4M) { 14041 sfmmup->sfmmu_tsb0_4minflcnt += 14042 r_size >> (TTE_PAGE_SHIFT(TTE8K) + 2); 14043 } 14044 14045 /* update sfmmu_ttecnt with the shme rgn ttecnt */ 14046 rttecnt = r_size >> TTE_PAGE_SHIFT(r_pgszc); 14047 atomic_add_long(&sfmmup->sfmmu_ttecnt[r_pgszc], 14048 rttecnt); 14049 14050 if (text && r_pgszc >= TTE4M && 14051 (tteflag || ((disable_large_pages >> TTE4M) & 14052 ((1 << (r_pgszc - TTE4M + 1)) - 1))) && 14053 !SFMMU_FLAGS_ISSET(sfmmup, HAT_4MTEXT_FLAG)) { 14054 SFMMU_FLAGS_SET(sfmmup, HAT_4MTEXT_FLAG); 14055 } 14056 14057 sfmmu_hat_exit(hatlockp); 14058 /* 14059 * On Panther we need to make sure TLB is programmed 14060 * to accept 32M/256M pages. Call 14061 * sfmmu_check_page_sizes() now to make sure TLB is 14062 * setup before making hmeregions visible to other 14063 * threads. 14064 */ 14065 sfmmu_check_page_sizes(sfmmup, 1); 14066 hatlockp = sfmmu_hat_enter(sfmmup); 14067 SF_RGNMAP_ADD(sfmmup->sfmmu_hmeregion_map, rid); 14068 14069 /* 14070 * if context is invalid tsb miss exception code will 14071 * call sfmmu_check_page_sizes() and update tsbmiss 14072 * area later. 14073 */ 14074 kpreempt_disable(); 14075 if (myjoin && 14076 (sfmmup->sfmmu_ctxs[CPU_MMU_IDX(CPU)].cnum 14077 != INVALID_CONTEXT)) { 14078 struct tsbmiss *tsbmp; 14079 14080 tsbmp = &tsbmiss_area[CPU->cpu_id]; 14081 ASSERT(sfmmup == tsbmp->usfmmup); 14082 BT_SET(tsbmp->shmermap, rid); 14083 if (r_pgszc > TTE64K) { 14084 tsbmp->uhat_rtteflags |= tteflag; 14085 } 14086 14087 } 14088 kpreempt_enable(); 14089 14090 sfmmu_hat_exit(hatlockp); 14091 ASSERT((hat_region_cookie_t)((uint64_t)rid) != 14092 HAT_INVALID_REGION_COOKIE); 14093 } else { 14094 hatlockp = sfmmu_hat_enter(sfmmup); 14095 SF_RGNMAP_ADD(sfmmup->sfmmu_ismregion_map, rid); 14096 sfmmu_hat_exit(hatlockp); 14097 } 14098 ASSERT(rid < maxids); 14099 14100 if (r_type == SFMMU_REGION_ISM) { 14101 sfmmu_find_scd(sfmmup); 14102 } 14103 return ((hat_region_cookie_t)((uint64_t)rid)); 14104 } 14105 14106 ASSERT(new_rgnp == NULL); 14107 14108 if (*busyrgnsp >= maxids) { 14109 mutex_exit(&srdp->srd_mutex); 14110 return (HAT_INVALID_REGION_COOKIE); 14111 } 14112 14113 ASSERT(MUTEX_HELD(&srdp->srd_mutex)); 14114 if (*freelistp != NULL) { 14115 rgnp = *freelistp; 14116 *freelistp = rgnp->rgn_next; 14117 ASSERT(rgnp->rgn_id < *nextidp); 14118 ASSERT(rgnp->rgn_id < maxids); 14119 ASSERT(rgnp->rgn_flags & SFMMU_REGION_FREE); 14120 ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK) 14121 == r_type); 14122 ASSERT(rarrp[rgnp->rgn_id] == rgnp); 14123 ASSERT(rgnp->rgn_hmeflags == 0); 14124 } else { 14125 /* 14126 * release local locks before memory allocation. 14127 */ 14128 mutex_exit(&srdp->srd_mutex); 14129 14130 new_rgnp = kmem_cache_alloc(region_cache, KM_SLEEP); 14131 14132 mutex_enter(&srdp->srd_mutex); 14133 for (rgnp = srdp->srd_rgnhash[rhash]; rgnp != NULL; 14134 rgnp = rgnp->rgn_hash) { 14135 if (rgnp->rgn_saddr == r_saddr && 14136 rgnp->rgn_size == r_size && 14137 rgnp->rgn_obj == r_obj && 14138 rgnp->rgn_objoff == r_objoff && 14139 rgnp->rgn_perm == r_perm && 14140 rgnp->rgn_pgszc == r_pgszc) { 14141 break; 14142 } 14143 } 14144 if (rgnp != NULL) { 14145 goto rfound; 14146 } 14147 14148 if (*nextidp >= maxids) { 14149 mutex_exit(&srdp->srd_mutex); 14150 goto fail; 14151 } 14152 rgnp = new_rgnp; 14153 new_rgnp = NULL; 14154 rgnp->rgn_id = (*nextidp)++; 14155 ASSERT(rgnp->rgn_id < maxids); 14156 ASSERT(rarrp[rgnp->rgn_id] == NULL); 14157 rarrp[rgnp->rgn_id] = rgnp; 14158 } 14159 14160 ASSERT(rgnp->rgn_sfmmu_head == NULL); 14161 ASSERT(rgnp->rgn_hmeflags == 0); 14162 #ifdef DEBUG 14163 for (i = 0; i < MMU_PAGE_SIZES; i++) { 14164 ASSERT(rgnp->rgn_ttecnt[i] == 0); 14165 } 14166 #endif 14167 rgnp->rgn_saddr = r_saddr; 14168 rgnp->rgn_size = r_size; 14169 rgnp->rgn_obj = r_obj; 14170 rgnp->rgn_objoff = r_objoff; 14171 rgnp->rgn_perm = r_perm; 14172 rgnp->rgn_pgszc = r_pgszc; 14173 rgnp->rgn_flags = r_type; 14174 rgnp->rgn_refcnt = 0; 14175 rgnp->rgn_cb_function = r_cb_function; 14176 rgnp->rgn_hash = srdp->srd_rgnhash[rhash]; 14177 srdp->srd_rgnhash[rhash] = rgnp; 14178 (*busyrgnsp)++; 14179 ASSERT(*busyrgnsp <= maxids); 14180 goto rfound; 14181 14182 fail: 14183 ASSERT(new_rgnp != NULL); 14184 kmem_cache_free(region_cache, new_rgnp); 14185 return (HAT_INVALID_REGION_COOKIE); 14186 } 14187 14188 /* 14189 * This function implements the shared context functionality required 14190 * when detaching a segment from an address space. It must be called 14191 * from hat_unshare() for all D(ISM) segments and from segvn_unmap(), 14192 * for segments with a valid region_cookie. 14193 * It will also be called from all seg_vn routines which change a 14194 * segment's attributes such as segvn_setprot(), segvn_setpagesize(), 14195 * segvn_clrszc() & segvn_advise(), as well as in the case of COW fault 14196 * from segvn_fault(). 14197 */ 14198 void 14199 hat_leave_region(struct hat *sfmmup, hat_region_cookie_t rcookie, uint_t flags) 14200 { 14201 sf_srd_t *srdp = sfmmup->sfmmu_srdp; 14202 sf_scd_t *scdp; 14203 uint_t rhash; 14204 uint_t rid = (uint_t)((uint64_t)rcookie); 14205 hatlock_t *hatlockp = NULL; 14206 sf_region_t *rgnp; 14207 sf_region_t **prev_rgnpp; 14208 sf_region_t *cur_rgnp; 14209 void *r_obj; 14210 int i; 14211 caddr_t r_saddr; 14212 caddr_t r_eaddr; 14213 size_t r_size; 14214 uchar_t r_pgszc; 14215 uchar_t r_type = flags & HAT_REGION_TYPE_MASK; 14216 14217 ASSERT(sfmmup != ksfmmup); 14218 ASSERT(srdp != NULL); 14219 ASSERT(srdp->srd_refcnt > 0); 14220 ASSERT(!(flags & ~HAT_REGION_TYPE_MASK)); 14221 ASSERT(flags == HAT_REGION_TEXT || flags == HAT_REGION_ISM); 14222 ASSERT(!sfmmup->sfmmu_free || sfmmup->sfmmu_scdp == NULL); 14223 14224 r_type = (r_type == HAT_REGION_ISM) ? SFMMU_REGION_ISM : 14225 SFMMU_REGION_HME; 14226 14227 if (r_type == SFMMU_REGION_ISM) { 14228 ASSERT(SFMMU_IS_ISMRID_VALID(rid)); 14229 ASSERT(rid < SFMMU_MAX_ISM_REGIONS); 14230 rgnp = srdp->srd_ismrgnp[rid]; 14231 } else { 14232 ASSERT(SFMMU_IS_SHMERID_VALID(rid)); 14233 ASSERT(rid < SFMMU_MAX_HME_REGIONS); 14234 rgnp = srdp->srd_hmergnp[rid]; 14235 } 14236 ASSERT(rgnp != NULL); 14237 ASSERT(rgnp->rgn_id == rid); 14238 ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK) == r_type); 14239 ASSERT(!(rgnp->rgn_flags & SFMMU_REGION_FREE)); 14240 ASSERT(AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock)); 14241 14242 ASSERT(sfmmup->sfmmu_xhat_provider == NULL); 14243 if (r_type == SFMMU_REGION_HME && sfmmup->sfmmu_as->a_xhat != NULL) { 14244 xhat_unload_callback_all(sfmmup->sfmmu_as, rgnp->rgn_saddr, 14245 rgnp->rgn_size, 0, NULL); 14246 } 14247 14248 if (sfmmup->sfmmu_free) { 14249 ulong_t rttecnt; 14250 r_pgszc = rgnp->rgn_pgszc; 14251 r_size = rgnp->rgn_size; 14252 14253 ASSERT(sfmmup->sfmmu_scdp == NULL); 14254 if (r_type == SFMMU_REGION_ISM) { 14255 SF_RGNMAP_DEL(sfmmup->sfmmu_ismregion_map, rid); 14256 } else { 14257 /* update shme rgns ttecnt in sfmmu_ttecnt */ 14258 rttecnt = r_size >> TTE_PAGE_SHIFT(r_pgszc); 14259 ASSERT(sfmmup->sfmmu_ttecnt[r_pgszc] >= rttecnt); 14260 14261 atomic_add_long(&sfmmup->sfmmu_ttecnt[r_pgszc], 14262 -rttecnt); 14263 14264 SF_RGNMAP_DEL(sfmmup->sfmmu_hmeregion_map, rid); 14265 } 14266 } else if (r_type == SFMMU_REGION_ISM) { 14267 hatlockp = sfmmu_hat_enter(sfmmup); 14268 ASSERT(rid < srdp->srd_next_ismrid); 14269 SF_RGNMAP_DEL(sfmmup->sfmmu_ismregion_map, rid); 14270 scdp = sfmmup->sfmmu_scdp; 14271 if (scdp != NULL && 14272 SF_RGNMAP_TEST(scdp->scd_ismregion_map, rid)) { 14273 sfmmu_leave_scd(sfmmup, r_type); 14274 ASSERT(sfmmu_hat_lock_held(sfmmup)); 14275 } 14276 sfmmu_hat_exit(hatlockp); 14277 } else { 14278 ulong_t rttecnt; 14279 r_pgszc = rgnp->rgn_pgszc; 14280 r_saddr = rgnp->rgn_saddr; 14281 r_size = rgnp->rgn_size; 14282 r_eaddr = r_saddr + r_size; 14283 14284 ASSERT(r_type == SFMMU_REGION_HME); 14285 hatlockp = sfmmu_hat_enter(sfmmup); 14286 ASSERT(rid < srdp->srd_next_hmerid); 14287 SF_RGNMAP_DEL(sfmmup->sfmmu_hmeregion_map, rid); 14288 14289 /* 14290 * If region is part of an SCD call sfmmu_leave_scd(). 14291 * Otherwise if process is not exiting and has valid context 14292 * just drop the context on the floor to lose stale TLB 14293 * entries and force the update of tsb miss area to reflect 14294 * the new region map. After that clean our TSB entries. 14295 */ 14296 scdp = sfmmup->sfmmu_scdp; 14297 if (scdp != NULL && 14298 SF_RGNMAP_TEST(scdp->scd_hmeregion_map, rid)) { 14299 sfmmu_leave_scd(sfmmup, r_type); 14300 ASSERT(sfmmu_hat_lock_held(sfmmup)); 14301 } 14302 sfmmu_invalidate_ctx(sfmmup); 14303 14304 i = TTE8K; 14305 while (i < mmu_page_sizes) { 14306 if (rgnp->rgn_ttecnt[i] != 0) { 14307 sfmmu_unload_tsb_range(sfmmup, r_saddr, 14308 r_eaddr, i); 14309 if (i < TTE4M) { 14310 i = TTE4M; 14311 continue; 14312 } else { 14313 break; 14314 } 14315 } 14316 i++; 14317 } 14318 /* Remove the preallocated 1/4 8k ttecnt for 4M regions. */ 14319 if (r_pgszc >= TTE4M) { 14320 rttecnt = r_size >> (TTE_PAGE_SHIFT(TTE8K) + 2); 14321 ASSERT(sfmmup->sfmmu_tsb0_4minflcnt >= 14322 rttecnt); 14323 sfmmup->sfmmu_tsb0_4minflcnt -= rttecnt; 14324 } 14325 14326 /* update shme rgns ttecnt in sfmmu_ttecnt */ 14327 rttecnt = r_size >> TTE_PAGE_SHIFT(r_pgszc); 14328 ASSERT(sfmmup->sfmmu_ttecnt[r_pgszc] >= rttecnt); 14329 atomic_add_long(&sfmmup->sfmmu_ttecnt[r_pgszc], -rttecnt); 14330 14331 sfmmu_hat_exit(hatlockp); 14332 if (scdp != NULL && sfmmup->sfmmu_scdp == NULL) { 14333 /* sfmmup left the scd, grow private tsb */ 14334 sfmmu_check_page_sizes(sfmmup, 1); 14335 } else { 14336 sfmmu_check_page_sizes(sfmmup, 0); 14337 } 14338 } 14339 14340 if (r_type == SFMMU_REGION_HME) { 14341 sfmmu_unlink_from_hmeregion(sfmmup, rgnp); 14342 } 14343 14344 r_obj = rgnp->rgn_obj; 14345 if (atomic_add_32_nv((volatile uint_t *)&rgnp->rgn_refcnt, -1)) { 14346 return; 14347 } 14348 14349 /* 14350 * looks like nobody uses this region anymore. Free it. 14351 */ 14352 rhash = RGN_HASH_FUNCTION(r_obj); 14353 mutex_enter(&srdp->srd_mutex); 14354 for (prev_rgnpp = &srdp->srd_rgnhash[rhash]; 14355 (cur_rgnp = *prev_rgnpp) != NULL; 14356 prev_rgnpp = &cur_rgnp->rgn_hash) { 14357 if (cur_rgnp == rgnp && cur_rgnp->rgn_refcnt == 0) { 14358 break; 14359 } 14360 } 14361 14362 if (cur_rgnp == NULL) { 14363 mutex_exit(&srdp->srd_mutex); 14364 return; 14365 } 14366 14367 ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK) == r_type); 14368 *prev_rgnpp = rgnp->rgn_hash; 14369 if (r_type == SFMMU_REGION_ISM) { 14370 rgnp->rgn_flags |= SFMMU_REGION_FREE; 14371 ASSERT(rid < srdp->srd_next_ismrid); 14372 rgnp->rgn_next = srdp->srd_ismrgnfree; 14373 srdp->srd_ismrgnfree = rgnp; 14374 ASSERT(srdp->srd_ismbusyrgns > 0); 14375 srdp->srd_ismbusyrgns--; 14376 mutex_exit(&srdp->srd_mutex); 14377 return; 14378 } 14379 mutex_exit(&srdp->srd_mutex); 14380 14381 /* 14382 * Destroy region's hmeblks. 14383 */ 14384 sfmmu_unload_hmeregion(srdp, rgnp); 14385 14386 rgnp->rgn_hmeflags = 0; 14387 14388 ASSERT(rgnp->rgn_sfmmu_head == NULL); 14389 ASSERT(rgnp->rgn_id == rid); 14390 for (i = 0; i < MMU_PAGE_SIZES; i++) { 14391 rgnp->rgn_ttecnt[i] = 0; 14392 } 14393 rgnp->rgn_flags |= SFMMU_REGION_FREE; 14394 mutex_enter(&srdp->srd_mutex); 14395 ASSERT(rid < srdp->srd_next_hmerid); 14396 rgnp->rgn_next = srdp->srd_hmergnfree; 14397 srdp->srd_hmergnfree = rgnp; 14398 ASSERT(srdp->srd_hmebusyrgns > 0); 14399 srdp->srd_hmebusyrgns--; 14400 mutex_exit(&srdp->srd_mutex); 14401 } 14402 14403 /* 14404 * For now only called for hmeblk regions and not for ISM regions. 14405 */ 14406 void 14407 hat_dup_region(struct hat *sfmmup, hat_region_cookie_t rcookie) 14408 { 14409 sf_srd_t *srdp = sfmmup->sfmmu_srdp; 14410 uint_t rid = (uint_t)((uint64_t)rcookie); 14411 sf_region_t *rgnp; 14412 sf_rgn_link_t *rlink; 14413 sf_rgn_link_t *hrlink; 14414 ulong_t rttecnt; 14415 14416 ASSERT(sfmmup != ksfmmup); 14417 ASSERT(srdp != NULL); 14418 ASSERT(srdp->srd_refcnt > 0); 14419 14420 ASSERT(rid < srdp->srd_next_hmerid); 14421 ASSERT(SFMMU_IS_SHMERID_VALID(rid)); 14422 ASSERT(rid < SFMMU_MAX_HME_REGIONS); 14423 14424 rgnp = srdp->srd_hmergnp[rid]; 14425 ASSERT(rgnp->rgn_refcnt > 0); 14426 ASSERT(rgnp->rgn_id == rid); 14427 ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK) == SFMMU_REGION_HME); 14428 ASSERT(!(rgnp->rgn_flags & SFMMU_REGION_FREE)); 14429 14430 atomic_add_32((volatile uint_t *)&rgnp->rgn_refcnt, 1); 14431 14432 /* LINTED: constant in conditional context */ 14433 SFMMU_HMERID2RLINKP(sfmmup, rid, rlink, 1, 0); 14434 ASSERT(rlink != NULL); 14435 mutex_enter(&rgnp->rgn_mutex); 14436 ASSERT(rgnp->rgn_sfmmu_head != NULL); 14437 /* LINTED: constant in conditional context */ 14438 SFMMU_HMERID2RLINKP(rgnp->rgn_sfmmu_head, rid, hrlink, 0, 0); 14439 ASSERT(hrlink != NULL); 14440 ASSERT(hrlink->prev == NULL); 14441 rlink->next = rgnp->rgn_sfmmu_head; 14442 rlink->prev = NULL; 14443 hrlink->prev = sfmmup; 14444 /* 14445 * make sure rlink's next field is correct 14446 * before making this link visible. 14447 */ 14448 membar_stst(); 14449 rgnp->rgn_sfmmu_head = sfmmup; 14450 mutex_exit(&rgnp->rgn_mutex); 14451 14452 /* update sfmmu_ttecnt with the shme rgn ttecnt */ 14453 rttecnt = rgnp->rgn_size >> TTE_PAGE_SHIFT(rgnp->rgn_pgszc); 14454 atomic_add_long(&sfmmup->sfmmu_ttecnt[rgnp->rgn_pgszc], rttecnt); 14455 /* update tsb0 inflation count */ 14456 if (rgnp->rgn_pgszc >= TTE4M) { 14457 sfmmup->sfmmu_tsb0_4minflcnt += 14458 rgnp->rgn_size >> (TTE_PAGE_SHIFT(TTE8K) + 2); 14459 } 14460 /* 14461 * Update regionid bitmask without hat lock since no other thread 14462 * can update this region bitmask right now. 14463 */ 14464 SF_RGNMAP_ADD(sfmmup->sfmmu_hmeregion_map, rid); 14465 } 14466 14467 /* ARGSUSED */ 14468 static int 14469 sfmmu_rgncache_constructor(void *buf, void *cdrarg, int kmflags) 14470 { 14471 sf_region_t *rgnp = (sf_region_t *)buf; 14472 bzero(buf, sizeof (*rgnp)); 14473 14474 mutex_init(&rgnp->rgn_mutex, NULL, MUTEX_DEFAULT, NULL); 14475 14476 return (0); 14477 } 14478 14479 /* ARGSUSED */ 14480 static void 14481 sfmmu_rgncache_destructor(void *buf, void *cdrarg) 14482 { 14483 sf_region_t *rgnp = (sf_region_t *)buf; 14484 mutex_destroy(&rgnp->rgn_mutex); 14485 } 14486 14487 static int 14488 sfrgnmap_isnull(sf_region_map_t *map) 14489 { 14490 int i; 14491 14492 for (i = 0; i < SFMMU_RGNMAP_WORDS; i++) { 14493 if (map->bitmap[i] != 0) { 14494 return (0); 14495 } 14496 } 14497 return (1); 14498 } 14499 14500 static int 14501 sfhmergnmap_isnull(sf_hmeregion_map_t *map) 14502 { 14503 int i; 14504 14505 for (i = 0; i < SFMMU_HMERGNMAP_WORDS; i++) { 14506 if (map->bitmap[i] != 0) { 14507 return (0); 14508 } 14509 } 14510 return (1); 14511 } 14512 14513 #ifdef DEBUG 14514 static void 14515 check_scd_sfmmu_list(sfmmu_t **headp, sfmmu_t *sfmmup, int onlist) 14516 { 14517 sfmmu_t *sp; 14518 sf_srd_t *srdp = sfmmup->sfmmu_srdp; 14519 14520 for (sp = *headp; sp != NULL; sp = sp->sfmmu_scd_link.next) { 14521 ASSERT(srdp == sp->sfmmu_srdp); 14522 if (sp == sfmmup) { 14523 if (onlist) { 14524 return; 14525 } else { 14526 panic("shctx: sfmmu 0x%p found on scd" 14527 "list 0x%p", (void *)sfmmup, 14528 (void *)*headp); 14529 } 14530 } 14531 } 14532 if (onlist) { 14533 panic("shctx: sfmmu 0x%p not found on scd list 0x%p", 14534 (void *)sfmmup, (void *)*headp); 14535 } else { 14536 return; 14537 } 14538 } 14539 #else /* DEBUG */ 14540 #define check_scd_sfmmu_list(headp, sfmmup, onlist) 14541 #endif /* DEBUG */ 14542 14543 /* 14544 * Removes an sfmmu from the SCD sfmmu list. 14545 */ 14546 static void 14547 sfmmu_from_scd_list(sfmmu_t **headp, sfmmu_t *sfmmup) 14548 { 14549 ASSERT(sfmmup->sfmmu_srdp != NULL); 14550 check_scd_sfmmu_list(headp, sfmmup, 1); 14551 if (sfmmup->sfmmu_scd_link.prev != NULL) { 14552 ASSERT(*headp != sfmmup); 14553 sfmmup->sfmmu_scd_link.prev->sfmmu_scd_link.next = 14554 sfmmup->sfmmu_scd_link.next; 14555 } else { 14556 ASSERT(*headp == sfmmup); 14557 *headp = sfmmup->sfmmu_scd_link.next; 14558 } 14559 if (sfmmup->sfmmu_scd_link.next != NULL) { 14560 sfmmup->sfmmu_scd_link.next->sfmmu_scd_link.prev = 14561 sfmmup->sfmmu_scd_link.prev; 14562 } 14563 } 14564 14565 14566 /* 14567 * Adds an sfmmu to the start of the queue. 14568 */ 14569 static void 14570 sfmmu_to_scd_list(sfmmu_t **headp, sfmmu_t *sfmmup) 14571 { 14572 check_scd_sfmmu_list(headp, sfmmup, 0); 14573 sfmmup->sfmmu_scd_link.prev = NULL; 14574 sfmmup->sfmmu_scd_link.next = *headp; 14575 if (*headp != NULL) 14576 (*headp)->sfmmu_scd_link.prev = sfmmup; 14577 *headp = sfmmup; 14578 } 14579 14580 /* 14581 * Remove an scd from the start of the queue. 14582 */ 14583 static void 14584 sfmmu_remove_scd(sf_scd_t **headp, sf_scd_t *scdp) 14585 { 14586 if (scdp->scd_prev != NULL) { 14587 ASSERT(*headp != scdp); 14588 scdp->scd_prev->scd_next = scdp->scd_next; 14589 } else { 14590 ASSERT(*headp == scdp); 14591 *headp = scdp->scd_next; 14592 } 14593 14594 if (scdp->scd_next != NULL) { 14595 scdp->scd_next->scd_prev = scdp->scd_prev; 14596 } 14597 } 14598 14599 /* 14600 * Add an scd to the start of the queue. 14601 */ 14602 static void 14603 sfmmu_add_scd(sf_scd_t **headp, sf_scd_t *scdp) 14604 { 14605 scdp->scd_prev = NULL; 14606 scdp->scd_next = *headp; 14607 if (*headp != NULL) { 14608 (*headp)->scd_prev = scdp; 14609 } 14610 *headp = scdp; 14611 } 14612 14613 static int 14614 sfmmu_alloc_scd_tsbs(sf_srd_t *srdp, sf_scd_t *scdp) 14615 { 14616 uint_t rid; 14617 uint_t i; 14618 uint_t j; 14619 ulong_t w; 14620 sf_region_t *rgnp; 14621 ulong_t tte8k_cnt = 0; 14622 ulong_t tte4m_cnt = 0; 14623 uint_t tsb_szc; 14624 sfmmu_t *scsfmmup = scdp->scd_sfmmup; 14625 sfmmu_t *ism_hatid; 14626 struct tsb_info *newtsb; 14627 int szc; 14628 14629 ASSERT(srdp != NULL); 14630 14631 for (i = 0; i < SFMMU_RGNMAP_WORDS; i++) { 14632 if ((w = scdp->scd_region_map.bitmap[i]) == 0) { 14633 continue; 14634 } 14635 j = 0; 14636 while (w) { 14637 if (!(w & 0x1)) { 14638 j++; 14639 w >>= 1; 14640 continue; 14641 } 14642 rid = (i << BT_ULSHIFT) | j; 14643 j++; 14644 w >>= 1; 14645 14646 if (rid < SFMMU_MAX_HME_REGIONS) { 14647 rgnp = srdp->srd_hmergnp[rid]; 14648 ASSERT(rgnp->rgn_id == rid); 14649 ASSERT(rgnp->rgn_refcnt > 0); 14650 14651 if (rgnp->rgn_pgszc < TTE4M) { 14652 tte8k_cnt += rgnp->rgn_size >> 14653 TTE_PAGE_SHIFT(TTE8K); 14654 } else { 14655 ASSERT(rgnp->rgn_pgszc >= TTE4M); 14656 tte4m_cnt += rgnp->rgn_size >> 14657 TTE_PAGE_SHIFT(TTE4M); 14658 /* 14659 * Inflate SCD tsb0 by preallocating 14660 * 1/4 8k ttecnt for 4M regions to 14661 * allow for lgpg alloc failure. 14662 */ 14663 tte8k_cnt += rgnp->rgn_size >> 14664 (TTE_PAGE_SHIFT(TTE8K) + 2); 14665 } 14666 } else { 14667 rid -= SFMMU_MAX_HME_REGIONS; 14668 rgnp = srdp->srd_ismrgnp[rid]; 14669 ASSERT(rgnp->rgn_id == rid); 14670 ASSERT(rgnp->rgn_refcnt > 0); 14671 14672 ism_hatid = (sfmmu_t *)rgnp->rgn_obj; 14673 ASSERT(ism_hatid->sfmmu_ismhat); 14674 14675 for (szc = 0; szc < TTE4M; szc++) { 14676 tte8k_cnt += 14677 ism_hatid->sfmmu_ttecnt[szc] << 14678 TTE_BSZS_SHIFT(szc); 14679 } 14680 14681 ASSERT(rgnp->rgn_pgszc >= TTE4M); 14682 if (rgnp->rgn_pgszc >= TTE4M) { 14683 tte4m_cnt += rgnp->rgn_size >> 14684 TTE_PAGE_SHIFT(TTE4M); 14685 } 14686 } 14687 } 14688 } 14689 14690 tsb_szc = SELECT_TSB_SIZECODE(tte8k_cnt); 14691 14692 /* Allocate both the SCD TSBs here. */ 14693 if (sfmmu_tsbinfo_alloc(&scsfmmup->sfmmu_tsb, 14694 tsb_szc, TSB8K|TSB64K|TSB512K, TSB_ALLOC, scsfmmup) && 14695 (tsb_szc <= TSB_4M_SZCODE || 14696 sfmmu_tsbinfo_alloc(&scsfmmup->sfmmu_tsb, 14697 TSB_4M_SZCODE, TSB8K|TSB64K|TSB512K, 14698 TSB_ALLOC, scsfmmup))) { 14699 14700 SFMMU_STAT(sf_scd_1sttsb_allocfail); 14701 return (TSB_ALLOCFAIL); 14702 } else { 14703 scsfmmup->sfmmu_tsb->tsb_flags |= TSB_SHAREDCTX; 14704 14705 if (tte4m_cnt) { 14706 tsb_szc = SELECT_TSB_SIZECODE(tte4m_cnt); 14707 if (sfmmu_tsbinfo_alloc(&newtsb, tsb_szc, 14708 TSB4M|TSB32M|TSB256M, TSB_ALLOC, scsfmmup) && 14709 (tsb_szc <= TSB_4M_SZCODE || 14710 sfmmu_tsbinfo_alloc(&newtsb, TSB_4M_SZCODE, 14711 TSB4M|TSB32M|TSB256M, 14712 TSB_ALLOC, scsfmmup))) { 14713 /* 14714 * If we fail to allocate the 2nd shared tsb, 14715 * just free the 1st tsb, return failure. 14716 */ 14717 sfmmu_tsbinfo_free(scsfmmup->sfmmu_tsb); 14718 SFMMU_STAT(sf_scd_2ndtsb_allocfail); 14719 return (TSB_ALLOCFAIL); 14720 } else { 14721 ASSERT(scsfmmup->sfmmu_tsb->tsb_next == NULL); 14722 newtsb->tsb_flags |= TSB_SHAREDCTX; 14723 scsfmmup->sfmmu_tsb->tsb_next = newtsb; 14724 SFMMU_STAT(sf_scd_2ndtsb_alloc); 14725 } 14726 } 14727 SFMMU_STAT(sf_scd_1sttsb_alloc); 14728 } 14729 return (TSB_SUCCESS); 14730 } 14731 14732 static void 14733 sfmmu_free_scd_tsbs(sfmmu_t *scd_sfmmu) 14734 { 14735 while (scd_sfmmu->sfmmu_tsb != NULL) { 14736 struct tsb_info *next = scd_sfmmu->sfmmu_tsb->tsb_next; 14737 sfmmu_tsbinfo_free(scd_sfmmu->sfmmu_tsb); 14738 scd_sfmmu->sfmmu_tsb = next; 14739 } 14740 } 14741 14742 /* 14743 * Link the sfmmu onto the hme region list. 14744 */ 14745 void 14746 sfmmu_link_to_hmeregion(sfmmu_t *sfmmup, sf_region_t *rgnp) 14747 { 14748 uint_t rid; 14749 sf_rgn_link_t *rlink; 14750 sfmmu_t *head; 14751 sf_rgn_link_t *hrlink; 14752 14753 rid = rgnp->rgn_id; 14754 ASSERT(SFMMU_IS_SHMERID_VALID(rid)); 14755 14756 /* LINTED: constant in conditional context */ 14757 SFMMU_HMERID2RLINKP(sfmmup, rid, rlink, 1, 1); 14758 ASSERT(rlink != NULL); 14759 mutex_enter(&rgnp->rgn_mutex); 14760 if ((head = rgnp->rgn_sfmmu_head) == NULL) { 14761 rlink->next = NULL; 14762 rlink->prev = NULL; 14763 /* 14764 * make sure rlink's next field is NULL 14765 * before making this link visible. 14766 */ 14767 membar_stst(); 14768 rgnp->rgn_sfmmu_head = sfmmup; 14769 } else { 14770 /* LINTED: constant in conditional context */ 14771 SFMMU_HMERID2RLINKP(head, rid, hrlink, 0, 0); 14772 ASSERT(hrlink != NULL); 14773 ASSERT(hrlink->prev == NULL); 14774 rlink->next = head; 14775 rlink->prev = NULL; 14776 hrlink->prev = sfmmup; 14777 /* 14778 * make sure rlink's next field is correct 14779 * before making this link visible. 14780 */ 14781 membar_stst(); 14782 rgnp->rgn_sfmmu_head = sfmmup; 14783 } 14784 mutex_exit(&rgnp->rgn_mutex); 14785 } 14786 14787 /* 14788 * Unlink the sfmmu from the hme region list. 14789 */ 14790 void 14791 sfmmu_unlink_from_hmeregion(sfmmu_t *sfmmup, sf_region_t *rgnp) 14792 { 14793 uint_t rid; 14794 sf_rgn_link_t *rlink; 14795 14796 rid = rgnp->rgn_id; 14797 ASSERT(SFMMU_IS_SHMERID_VALID(rid)); 14798 14799 /* LINTED: constant in conditional context */ 14800 SFMMU_HMERID2RLINKP(sfmmup, rid, rlink, 0, 0); 14801 ASSERT(rlink != NULL); 14802 mutex_enter(&rgnp->rgn_mutex); 14803 if (rgnp->rgn_sfmmu_head == sfmmup) { 14804 sfmmu_t *next = rlink->next; 14805 rgnp->rgn_sfmmu_head = next; 14806 /* 14807 * if we are stopped by xc_attention() after this 14808 * point the forward link walking in 14809 * sfmmu_rgntlb_demap() will work correctly since the 14810 * head correctly points to the next element. 14811 */ 14812 membar_stst(); 14813 rlink->next = NULL; 14814 ASSERT(rlink->prev == NULL); 14815 if (next != NULL) { 14816 sf_rgn_link_t *nrlink; 14817 /* LINTED: constant in conditional context */ 14818 SFMMU_HMERID2RLINKP(next, rid, nrlink, 0, 0); 14819 ASSERT(nrlink != NULL); 14820 ASSERT(nrlink->prev == sfmmup); 14821 nrlink->prev = NULL; 14822 } 14823 } else { 14824 sfmmu_t *next = rlink->next; 14825 sfmmu_t *prev = rlink->prev; 14826 sf_rgn_link_t *prlink; 14827 14828 ASSERT(prev != NULL); 14829 /* LINTED: constant in conditional context */ 14830 SFMMU_HMERID2RLINKP(prev, rid, prlink, 0, 0); 14831 ASSERT(prlink != NULL); 14832 ASSERT(prlink->next == sfmmup); 14833 prlink->next = next; 14834 /* 14835 * if we are stopped by xc_attention() 14836 * after this point the forward link walking 14837 * will work correctly since the prev element 14838 * correctly points to the next element. 14839 */ 14840 membar_stst(); 14841 rlink->next = NULL; 14842 rlink->prev = NULL; 14843 if (next != NULL) { 14844 sf_rgn_link_t *nrlink; 14845 /* LINTED: constant in conditional context */ 14846 SFMMU_HMERID2RLINKP(next, rid, nrlink, 0, 0); 14847 ASSERT(nrlink != NULL); 14848 ASSERT(nrlink->prev == sfmmup); 14849 nrlink->prev = prev; 14850 } 14851 } 14852 mutex_exit(&rgnp->rgn_mutex); 14853 } 14854 14855 /* 14856 * Link scd sfmmu onto ism or hme region list for each region in the 14857 * scd region map. 14858 */ 14859 void 14860 sfmmu_link_scd_to_regions(sf_srd_t *srdp, sf_scd_t *scdp) 14861 { 14862 uint_t rid; 14863 uint_t i; 14864 uint_t j; 14865 ulong_t w; 14866 sf_region_t *rgnp; 14867 sfmmu_t *scsfmmup; 14868 14869 scsfmmup = scdp->scd_sfmmup; 14870 ASSERT(scsfmmup->sfmmu_scdhat); 14871 for (i = 0; i < SFMMU_RGNMAP_WORDS; i++) { 14872 if ((w = scdp->scd_region_map.bitmap[i]) == 0) { 14873 continue; 14874 } 14875 j = 0; 14876 while (w) { 14877 if (!(w & 0x1)) { 14878 j++; 14879 w >>= 1; 14880 continue; 14881 } 14882 rid = (i << BT_ULSHIFT) | j; 14883 j++; 14884 w >>= 1; 14885 14886 if (rid < SFMMU_MAX_HME_REGIONS) { 14887 rgnp = srdp->srd_hmergnp[rid]; 14888 ASSERT(rgnp->rgn_id == rid); 14889 ASSERT(rgnp->rgn_refcnt > 0); 14890 sfmmu_link_to_hmeregion(scsfmmup, rgnp); 14891 } else { 14892 sfmmu_t *ism_hatid = NULL; 14893 ism_ment_t *ism_ment; 14894 rid -= SFMMU_MAX_HME_REGIONS; 14895 rgnp = srdp->srd_ismrgnp[rid]; 14896 ASSERT(rgnp->rgn_id == rid); 14897 ASSERT(rgnp->rgn_refcnt > 0); 14898 14899 ism_hatid = (sfmmu_t *)rgnp->rgn_obj; 14900 ASSERT(ism_hatid->sfmmu_ismhat); 14901 ism_ment = &scdp->scd_ism_links[rid]; 14902 ism_ment->iment_hat = scsfmmup; 14903 ism_ment->iment_base_va = rgnp->rgn_saddr; 14904 mutex_enter(&ism_mlist_lock); 14905 iment_add(ism_ment, ism_hatid); 14906 mutex_exit(&ism_mlist_lock); 14907 14908 } 14909 } 14910 } 14911 } 14912 /* 14913 * Unlink scd sfmmu from ism or hme region list for each region in the 14914 * scd region map. 14915 */ 14916 void 14917 sfmmu_unlink_scd_from_regions(sf_srd_t *srdp, sf_scd_t *scdp) 14918 { 14919 uint_t rid; 14920 uint_t i; 14921 uint_t j; 14922 ulong_t w; 14923 sf_region_t *rgnp; 14924 sfmmu_t *scsfmmup; 14925 14926 scsfmmup = scdp->scd_sfmmup; 14927 for (i = 0; i < SFMMU_RGNMAP_WORDS; i++) { 14928 if ((w = scdp->scd_region_map.bitmap[i]) == 0) { 14929 continue; 14930 } 14931 j = 0; 14932 while (w) { 14933 if (!(w & 0x1)) { 14934 j++; 14935 w >>= 1; 14936 continue; 14937 } 14938 rid = (i << BT_ULSHIFT) | j; 14939 j++; 14940 w >>= 1; 14941 14942 if (rid < SFMMU_MAX_HME_REGIONS) { 14943 rgnp = srdp->srd_hmergnp[rid]; 14944 ASSERT(rgnp->rgn_id == rid); 14945 ASSERT(rgnp->rgn_refcnt > 0); 14946 sfmmu_unlink_from_hmeregion(scsfmmup, 14947 rgnp); 14948 14949 } else { 14950 sfmmu_t *ism_hatid = NULL; 14951 ism_ment_t *ism_ment; 14952 rid -= SFMMU_MAX_HME_REGIONS; 14953 rgnp = srdp->srd_ismrgnp[rid]; 14954 ASSERT(rgnp->rgn_id == rid); 14955 ASSERT(rgnp->rgn_refcnt > 0); 14956 14957 ism_hatid = (sfmmu_t *)rgnp->rgn_obj; 14958 ASSERT(ism_hatid->sfmmu_ismhat); 14959 ism_ment = &scdp->scd_ism_links[rid]; 14960 ASSERT(ism_ment->iment_hat == scdp->scd_sfmmup); 14961 ASSERT(ism_ment->iment_base_va == 14962 rgnp->rgn_saddr); 14963 mutex_enter(&ism_mlist_lock); 14964 iment_sub(ism_ment, ism_hatid); 14965 mutex_exit(&ism_mlist_lock); 14966 14967 } 14968 } 14969 } 14970 } 14971 /* 14972 * Allocates and initialises a new SCD structure, this is called with 14973 * the srd_scd_mutex held and returns with the reference count 14974 * initialised to 1. 14975 */ 14976 static sf_scd_t * 14977 sfmmu_alloc_scd(sf_srd_t *srdp, sf_region_map_t *new_map) 14978 { 14979 sf_scd_t *new_scdp; 14980 sfmmu_t *scsfmmup; 14981 int i; 14982 14983 ASSERT(MUTEX_HELD(&srdp->srd_scd_mutex)); 14984 new_scdp = kmem_cache_alloc(scd_cache, KM_SLEEP); 14985 14986 scsfmmup = kmem_cache_alloc(sfmmuid_cache, KM_SLEEP); 14987 new_scdp->scd_sfmmup = scsfmmup; 14988 scsfmmup->sfmmu_srdp = srdp; 14989 scsfmmup->sfmmu_scdp = new_scdp; 14990 scsfmmup->sfmmu_tsb0_4minflcnt = 0; 14991 scsfmmup->sfmmu_scdhat = 1; 14992 CPUSET_ALL(scsfmmup->sfmmu_cpusran); 14993 bzero(scsfmmup->sfmmu_hmeregion_links, SFMMU_L1_HMERLINKS_SIZE); 14994 14995 ASSERT(max_mmu_ctxdoms > 0); 14996 for (i = 0; i < max_mmu_ctxdoms; i++) { 14997 scsfmmup->sfmmu_ctxs[i].cnum = INVALID_CONTEXT; 14998 scsfmmup->sfmmu_ctxs[i].gnum = 0; 14999 } 15000 15001 for (i = 0; i < MMU_PAGE_SIZES; i++) { 15002 new_scdp->scd_rttecnt[i] = 0; 15003 } 15004 15005 new_scdp->scd_region_map = *new_map; 15006 new_scdp->scd_refcnt = 1; 15007 if (sfmmu_alloc_scd_tsbs(srdp, new_scdp) != TSB_SUCCESS) { 15008 kmem_cache_free(scd_cache, new_scdp); 15009 kmem_cache_free(sfmmuid_cache, scsfmmup); 15010 return (NULL); 15011 } 15012 if (&mmu_init_scd) { 15013 mmu_init_scd(new_scdp); 15014 } 15015 return (new_scdp); 15016 } 15017 15018 /* 15019 * The first phase of a process joining an SCD. The hat structure is 15020 * linked to the SCD queue and then the HAT_JOIN_SCD sfmmu flag is set 15021 * and a cross-call with context invalidation is used to cause the 15022 * remaining work to be carried out in the sfmmu_tsbmiss_exception() 15023 * routine. 15024 */ 15025 static void 15026 sfmmu_join_scd(sf_scd_t *scdp, sfmmu_t *sfmmup) 15027 { 15028 hatlock_t *hatlockp; 15029 sf_srd_t *srdp = sfmmup->sfmmu_srdp; 15030 int i; 15031 sf_scd_t *old_scdp; 15032 15033 ASSERT(srdp != NULL); 15034 ASSERT(scdp != NULL); 15035 ASSERT(scdp->scd_refcnt > 0); 15036 ASSERT(AS_WRITE_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock)); 15037 15038 if ((old_scdp = sfmmup->sfmmu_scdp) != NULL) { 15039 ASSERT(old_scdp != scdp); 15040 15041 mutex_enter(&old_scdp->scd_mutex); 15042 sfmmu_from_scd_list(&old_scdp->scd_sf_list, sfmmup); 15043 mutex_exit(&old_scdp->scd_mutex); 15044 /* 15045 * sfmmup leaves the old scd. Update sfmmu_ttecnt to 15046 * include the shme rgn ttecnt for rgns that 15047 * were in the old SCD 15048 */ 15049 for (i = 0; i < mmu_page_sizes; i++) { 15050 ASSERT(sfmmup->sfmmu_scdrttecnt[i] == 15051 old_scdp->scd_rttecnt[i]); 15052 atomic_add_long(&sfmmup->sfmmu_ttecnt[i], 15053 sfmmup->sfmmu_scdrttecnt[i]); 15054 } 15055 } 15056 15057 /* 15058 * Move sfmmu to the scd lists. 15059 */ 15060 mutex_enter(&scdp->scd_mutex); 15061 sfmmu_to_scd_list(&scdp->scd_sf_list, sfmmup); 15062 mutex_exit(&scdp->scd_mutex); 15063 SF_SCD_INCR_REF(scdp); 15064 15065 hatlockp = sfmmu_hat_enter(sfmmup); 15066 /* 15067 * For a multi-thread process, we must stop 15068 * all the other threads before joining the scd. 15069 */ 15070 15071 SFMMU_FLAGS_SET(sfmmup, HAT_JOIN_SCD); 15072 15073 sfmmu_invalidate_ctx(sfmmup); 15074 sfmmup->sfmmu_scdp = scdp; 15075 15076 /* 15077 * Copy scd_rttecnt into sfmmup's sfmmu_scdrttecnt, and update 15078 * sfmmu_ttecnt to not include the rgn ttecnt just joined in SCD. 15079 */ 15080 for (i = 0; i < mmu_page_sizes; i++) { 15081 sfmmup->sfmmu_scdrttecnt[i] = scdp->scd_rttecnt[i]; 15082 ASSERT(sfmmup->sfmmu_ttecnt[i] >= scdp->scd_rttecnt[i]); 15083 atomic_add_long(&sfmmup->sfmmu_ttecnt[i], 15084 -sfmmup->sfmmu_scdrttecnt[i]); 15085 } 15086 /* update tsb0 inflation count */ 15087 if (old_scdp != NULL) { 15088 sfmmup->sfmmu_tsb0_4minflcnt += 15089 old_scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt; 15090 } 15091 ASSERT(sfmmup->sfmmu_tsb0_4minflcnt >= 15092 scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt); 15093 sfmmup->sfmmu_tsb0_4minflcnt -= scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt; 15094 15095 sfmmu_hat_exit(hatlockp); 15096 15097 if (old_scdp != NULL) { 15098 SF_SCD_DECR_REF(srdp, old_scdp); 15099 } 15100 15101 } 15102 15103 /* 15104 * This routine is called by a process to become part of an SCD. It is called 15105 * from sfmmu_tsbmiss_exception() once most of the initial work has been 15106 * done by sfmmu_join_scd(). This routine must not drop the hat lock. 15107 */ 15108 static void 15109 sfmmu_finish_join_scd(sfmmu_t *sfmmup) 15110 { 15111 struct tsb_info *tsbinfop; 15112 15113 ASSERT(sfmmu_hat_lock_held(sfmmup)); 15114 ASSERT(sfmmup->sfmmu_scdp != NULL); 15115 ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD)); 15116 ASSERT(!SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY)); 15117 ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_ALLCTX_INVALID)); 15118 15119 for (tsbinfop = sfmmup->sfmmu_tsb; tsbinfop != NULL; 15120 tsbinfop = tsbinfop->tsb_next) { 15121 if (tsbinfop->tsb_flags & TSB_SWAPPED) { 15122 continue; 15123 } 15124 ASSERT(!(tsbinfop->tsb_flags & TSB_RELOC_FLAG)); 15125 15126 sfmmu_inv_tsb(tsbinfop->tsb_va, 15127 TSB_BYTES(tsbinfop->tsb_szc)); 15128 } 15129 15130 /* Set HAT_CTX1_FLAG for all SCD ISMs */ 15131 sfmmu_ism_hatflags(sfmmup, 1); 15132 15133 SFMMU_STAT(sf_join_scd); 15134 } 15135 15136 /* 15137 * This routine is called in order to check if there is an SCD which matches 15138 * the process's region map if not then a new SCD may be created. 15139 */ 15140 static void 15141 sfmmu_find_scd(sfmmu_t *sfmmup) 15142 { 15143 sf_srd_t *srdp = sfmmup->sfmmu_srdp; 15144 sf_scd_t *scdp, *new_scdp; 15145 int ret; 15146 15147 ASSERT(srdp != NULL); 15148 ASSERT(AS_WRITE_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock)); 15149 15150 mutex_enter(&srdp->srd_scd_mutex); 15151 for (scdp = srdp->srd_scdp; scdp != NULL; 15152 scdp = scdp->scd_next) { 15153 SF_RGNMAP_EQUAL(&scdp->scd_region_map, 15154 &sfmmup->sfmmu_region_map, ret); 15155 if (ret == 1) { 15156 SF_SCD_INCR_REF(scdp); 15157 mutex_exit(&srdp->srd_scd_mutex); 15158 sfmmu_join_scd(scdp, sfmmup); 15159 ASSERT(scdp->scd_refcnt >= 2); 15160 atomic_add_32((volatile uint32_t *) 15161 &scdp->scd_refcnt, -1); 15162 return; 15163 } else { 15164 /* 15165 * If the sfmmu region map is a subset of the scd 15166 * region map, then the assumption is that this process 15167 * will continue attaching to ISM segments until the 15168 * region maps are equal. 15169 */ 15170 SF_RGNMAP_IS_SUBSET(&scdp->scd_region_map, 15171 &sfmmup->sfmmu_region_map, ret); 15172 if (ret == 1) { 15173 mutex_exit(&srdp->srd_scd_mutex); 15174 return; 15175 } 15176 } 15177 } 15178 15179 ASSERT(scdp == NULL); 15180 /* 15181 * No matching SCD has been found, create a new one. 15182 */ 15183 if ((new_scdp = sfmmu_alloc_scd(srdp, &sfmmup->sfmmu_region_map)) == 15184 NULL) { 15185 mutex_exit(&srdp->srd_scd_mutex); 15186 return; 15187 } 15188 15189 /* 15190 * sfmmu_alloc_scd() returns with a ref count of 1 on the scd. 15191 */ 15192 15193 /* Set scd_rttecnt for shme rgns in SCD */ 15194 sfmmu_set_scd_rttecnt(srdp, new_scdp); 15195 15196 /* 15197 * Link scd onto srd_scdp list and scd sfmmu onto region/iment lists. 15198 */ 15199 sfmmu_link_scd_to_regions(srdp, new_scdp); 15200 sfmmu_add_scd(&srdp->srd_scdp, new_scdp); 15201 SFMMU_STAT_ADD(sf_create_scd, 1); 15202 15203 mutex_exit(&srdp->srd_scd_mutex); 15204 sfmmu_join_scd(new_scdp, sfmmup); 15205 ASSERT(new_scdp->scd_refcnt >= 2); 15206 atomic_add_32((volatile uint32_t *)&new_scdp->scd_refcnt, -1); 15207 } 15208 15209 /* 15210 * This routine is called by a process to remove itself from an SCD. It is 15211 * either called when the processes has detached from a segment or from 15212 * hat_free_start() as a result of calling exit. 15213 */ 15214 static void 15215 sfmmu_leave_scd(sfmmu_t *sfmmup, uchar_t r_type) 15216 { 15217 sf_scd_t *scdp = sfmmup->sfmmu_scdp; 15218 sf_srd_t *srdp = sfmmup->sfmmu_srdp; 15219 hatlock_t *hatlockp = TSB_HASH(sfmmup); 15220 int i; 15221 15222 ASSERT(scdp != NULL); 15223 ASSERT(srdp != NULL); 15224 15225 if (sfmmup->sfmmu_free) { 15226 /* 15227 * If the process is part of an SCD the sfmmu is unlinked 15228 * from scd_sf_list. 15229 */ 15230 mutex_enter(&scdp->scd_mutex); 15231 sfmmu_from_scd_list(&scdp->scd_sf_list, sfmmup); 15232 mutex_exit(&scdp->scd_mutex); 15233 /* 15234 * Update sfmmu_ttecnt to include the rgn ttecnt for rgns that 15235 * are about to leave the SCD 15236 */ 15237 for (i = 0; i < mmu_page_sizes; i++) { 15238 ASSERT(sfmmup->sfmmu_scdrttecnt[i] == 15239 scdp->scd_rttecnt[i]); 15240 atomic_add_long(&sfmmup->sfmmu_ttecnt[i], 15241 sfmmup->sfmmu_scdrttecnt[i]); 15242 sfmmup->sfmmu_scdrttecnt[i] = 0; 15243 } 15244 sfmmup->sfmmu_scdp = NULL; 15245 15246 SF_SCD_DECR_REF(srdp, scdp); 15247 return; 15248 } 15249 15250 ASSERT(r_type != SFMMU_REGION_ISM || 15251 SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY)); 15252 ASSERT(scdp->scd_refcnt); 15253 ASSERT(!sfmmup->sfmmu_free); 15254 ASSERT(sfmmu_hat_lock_held(sfmmup)); 15255 ASSERT(AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock)); 15256 15257 /* 15258 * Wait for ISM maps to be updated. 15259 */ 15260 if (r_type != SFMMU_REGION_ISM) { 15261 while (SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY) && 15262 sfmmup->sfmmu_scdp != NULL) { 15263 cv_wait(&sfmmup->sfmmu_tsb_cv, 15264 HATLOCK_MUTEXP(hatlockp)); 15265 } 15266 15267 if (sfmmup->sfmmu_scdp == NULL) { 15268 sfmmu_hat_exit(hatlockp); 15269 return; 15270 } 15271 SFMMU_FLAGS_SET(sfmmup, HAT_ISMBUSY); 15272 } 15273 15274 if (SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD)) { 15275 SFMMU_FLAGS_CLEAR(sfmmup, HAT_JOIN_SCD); 15276 /* 15277 * Since HAT_JOIN_SCD was set our context 15278 * is still invalid. 15279 */ 15280 } else { 15281 /* 15282 * For a multi-thread process, we must stop 15283 * all the other threads before leaving the scd. 15284 */ 15285 15286 sfmmu_invalidate_ctx(sfmmup); 15287 } 15288 15289 /* Clear all the rid's for ISM, delete flags, etc */ 15290 ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY)); 15291 sfmmu_ism_hatflags(sfmmup, 0); 15292 15293 /* 15294 * Update sfmmu_ttecnt to include the rgn ttecnt for rgns that 15295 * are in SCD before this sfmmup leaves the SCD. 15296 */ 15297 for (i = 0; i < mmu_page_sizes; i++) { 15298 ASSERT(sfmmup->sfmmu_scdrttecnt[i] == 15299 scdp->scd_rttecnt[i]); 15300 atomic_add_long(&sfmmup->sfmmu_ttecnt[i], 15301 sfmmup->sfmmu_scdrttecnt[i]); 15302 sfmmup->sfmmu_scdrttecnt[i] = 0; 15303 /* update ismttecnt to include SCD ism before hat leaves SCD */ 15304 sfmmup->sfmmu_ismttecnt[i] += sfmmup->sfmmu_scdismttecnt[i]; 15305 sfmmup->sfmmu_scdismttecnt[i] = 0; 15306 } 15307 /* update tsb0 inflation count */ 15308 sfmmup->sfmmu_tsb0_4minflcnt += scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt; 15309 15310 if (r_type != SFMMU_REGION_ISM) { 15311 SFMMU_FLAGS_CLEAR(sfmmup, HAT_ISMBUSY); 15312 } 15313 sfmmup->sfmmu_scdp = NULL; 15314 15315 sfmmu_hat_exit(hatlockp); 15316 15317 /* 15318 * Unlink sfmmu from scd_sf_list this can be done without holding 15319 * the hat lock as we hold the sfmmu_as lock which prevents 15320 * hat_join_region from adding this thread to the scd again. Other 15321 * threads check if sfmmu_scdp is NULL under hat lock and if it's NULL 15322 * they won't get here, since sfmmu_leave_scd() clears sfmmu_scdp 15323 * while holding the hat lock. 15324 */ 15325 mutex_enter(&scdp->scd_mutex); 15326 sfmmu_from_scd_list(&scdp->scd_sf_list, sfmmup); 15327 mutex_exit(&scdp->scd_mutex); 15328 SFMMU_STAT(sf_leave_scd); 15329 15330 SF_SCD_DECR_REF(srdp, scdp); 15331 hatlockp = sfmmu_hat_enter(sfmmup); 15332 15333 } 15334 15335 /* 15336 * Unlink and free up an SCD structure with a reference count of 0. 15337 */ 15338 static void 15339 sfmmu_destroy_scd(sf_srd_t *srdp, sf_scd_t *scdp, sf_region_map_t *scd_rmap) 15340 { 15341 sfmmu_t *scsfmmup; 15342 sf_scd_t *sp; 15343 hatlock_t *shatlockp; 15344 int i, ret; 15345 15346 mutex_enter(&srdp->srd_scd_mutex); 15347 for (sp = srdp->srd_scdp; sp != NULL; sp = sp->scd_next) { 15348 if (sp == scdp) 15349 break; 15350 } 15351 if (sp == NULL || sp->scd_refcnt) { 15352 mutex_exit(&srdp->srd_scd_mutex); 15353 return; 15354 } 15355 15356 /* 15357 * It is possible that the scd has been freed and reallocated with a 15358 * different region map while we've been waiting for the srd_scd_mutex. 15359 */ 15360 SF_RGNMAP_EQUAL(scd_rmap, &sp->scd_region_map, ret); 15361 if (ret != 1) { 15362 mutex_exit(&srdp->srd_scd_mutex); 15363 return; 15364 } 15365 15366 ASSERT(scdp->scd_sf_list == NULL); 15367 /* 15368 * Unlink scd from srd_scdp list. 15369 */ 15370 sfmmu_remove_scd(&srdp->srd_scdp, scdp); 15371 mutex_exit(&srdp->srd_scd_mutex); 15372 15373 sfmmu_unlink_scd_from_regions(srdp, scdp); 15374 15375 /* Clear shared context tsb and release ctx */ 15376 scsfmmup = scdp->scd_sfmmup; 15377 15378 /* 15379 * create a barrier so that scd will not be destroyed 15380 * if other thread still holds the same shared hat lock. 15381 * E.g., sfmmu_tsbmiss_exception() needs to acquire the 15382 * shared hat lock before checking the shared tsb reloc flag. 15383 */ 15384 shatlockp = sfmmu_hat_enter(scsfmmup); 15385 sfmmu_hat_exit(shatlockp); 15386 15387 sfmmu_free_scd_tsbs(scsfmmup); 15388 15389 for (i = 0; i < SFMMU_L1_HMERLINKS; i++) { 15390 if (scsfmmup->sfmmu_hmeregion_links[i] != NULL) { 15391 kmem_free(scsfmmup->sfmmu_hmeregion_links[i], 15392 SFMMU_L2_HMERLINKS_SIZE); 15393 scsfmmup->sfmmu_hmeregion_links[i] = NULL; 15394 } 15395 } 15396 kmem_cache_free(sfmmuid_cache, scsfmmup); 15397 kmem_cache_free(scd_cache, scdp); 15398 SFMMU_STAT(sf_destroy_scd); 15399 } 15400 15401 /* 15402 * Modifies the HAT_CTX1_FLAG for each of the ISM segments which correspond to 15403 * bits which are set in the ism_region_map parameter. This flag indicates to 15404 * the tsbmiss handler that mapping for these segments should be loaded using 15405 * the shared context. 15406 */ 15407 static void 15408 sfmmu_ism_hatflags(sfmmu_t *sfmmup, int addflag) 15409 { 15410 sf_scd_t *scdp = sfmmup->sfmmu_scdp; 15411 ism_blk_t *ism_blkp; 15412 ism_map_t *ism_map; 15413 int i, rid; 15414 15415 ASSERT(sfmmup->sfmmu_iblk != NULL); 15416 ASSERT(scdp != NULL); 15417 /* 15418 * Note that the caller either set HAT_ISMBUSY flag or checked 15419 * under hat lock that HAT_ISMBUSY was not set by another thread. 15420 */ 15421 ASSERT(sfmmu_hat_lock_held(sfmmup)); 15422 15423 ism_blkp = sfmmup->sfmmu_iblk; 15424 while (ism_blkp != NULL) { 15425 ism_map = ism_blkp->iblk_maps; 15426 for (i = 0; ism_map[i].imap_ismhat && i < ISM_MAP_SLOTS; i++) { 15427 rid = ism_map[i].imap_rid; 15428 if (rid == SFMMU_INVALID_ISMRID) { 15429 continue; 15430 } 15431 ASSERT(rid >= 0 && rid < SFMMU_MAX_ISM_REGIONS); 15432 if (SF_RGNMAP_TEST(scdp->scd_ismregion_map, rid) && 15433 addflag) { 15434 ism_map[i].imap_hatflags |= 15435 HAT_CTX1_FLAG; 15436 } else { 15437 ism_map[i].imap_hatflags &= 15438 ~HAT_CTX1_FLAG; 15439 } 15440 } 15441 ism_blkp = ism_blkp->iblk_next; 15442 } 15443 } 15444 15445 static int 15446 sfmmu_srd_lock_held(sf_srd_t *srdp) 15447 { 15448 return (MUTEX_HELD(&srdp->srd_mutex)); 15449 } 15450 15451 /* ARGSUSED */ 15452 static int 15453 sfmmu_scdcache_constructor(void *buf, void *cdrarg, int kmflags) 15454 { 15455 sf_scd_t *scdp = (sf_scd_t *)buf; 15456 15457 bzero(buf, sizeof (sf_scd_t)); 15458 mutex_init(&scdp->scd_mutex, NULL, MUTEX_DEFAULT, NULL); 15459 return (0); 15460 } 15461 15462 /* ARGSUSED */ 15463 static void 15464 sfmmu_scdcache_destructor(void *buf, void *cdrarg) 15465 { 15466 sf_scd_t *scdp = (sf_scd_t *)buf; 15467 15468 mutex_destroy(&scdp->scd_mutex); 15469 } 15470 15471 /* 15472 * The listp parameter is a pointer to a list of hmeblks which are partially 15473 * freed as result of calling sfmmu_hblk_hash_rm(), the last phase of the 15474 * freeing process is to cross-call all cpus to ensure that there are no 15475 * remaining cached references. 15476 * 15477 * If the local generation number is less than the global then we can free 15478 * hmeblks which are already on the pending queue as another cpu has completed 15479 * the cross-call. 15480 * 15481 * We cross-call to make sure that there are no threads on other cpus accessing 15482 * these hmblks and then complete the process of freeing them under the 15483 * following conditions: 15484 * The total number of pending hmeblks is greater than the threshold 15485 * The reserve list has fewer than HBLK_RESERVE_CNT hmeblks 15486 * It is at least 1 second since the last time we cross-called 15487 * 15488 * Otherwise, we add the hmeblks to the per-cpu pending queue. 15489 */ 15490 static void 15491 sfmmu_hblks_list_purge(struct hme_blk **listp, int dontfree) 15492 { 15493 struct hme_blk *hblkp, *pr_hblkp = NULL; 15494 int count = 0; 15495 cpuset_t cpuset = cpu_ready_set; 15496 cpu_hme_pend_t *cpuhp; 15497 timestruc_t now; 15498 int one_second_expired = 0; 15499 15500 gethrestime_lasttick(&now); 15501 15502 for (hblkp = *listp; hblkp != NULL; hblkp = hblkp->hblk_next) { 15503 ASSERT(hblkp->hblk_shw_bit == 0); 15504 ASSERT(hblkp->hblk_shared == 0); 15505 count++; 15506 pr_hblkp = hblkp; 15507 } 15508 15509 cpuhp = &cpu_hme_pend[CPU->cpu_seqid]; 15510 mutex_enter(&cpuhp->chp_mutex); 15511 15512 if ((cpuhp->chp_count + count) == 0) { 15513 mutex_exit(&cpuhp->chp_mutex); 15514 return; 15515 } 15516 15517 if ((now.tv_sec - cpuhp->chp_timestamp) > 1) { 15518 one_second_expired = 1; 15519 } 15520 15521 if (!dontfree && (freehblkcnt < HBLK_RESERVE_CNT || 15522 (cpuhp->chp_count + count) > cpu_hme_pend_thresh || 15523 one_second_expired)) { 15524 /* Append global list to local */ 15525 if (pr_hblkp == NULL) { 15526 *listp = cpuhp->chp_listp; 15527 } else { 15528 pr_hblkp->hblk_next = cpuhp->chp_listp; 15529 } 15530 cpuhp->chp_listp = NULL; 15531 cpuhp->chp_count = 0; 15532 cpuhp->chp_timestamp = now.tv_sec; 15533 mutex_exit(&cpuhp->chp_mutex); 15534 15535 kpreempt_disable(); 15536 CPUSET_DEL(cpuset, CPU->cpu_id); 15537 xt_sync(cpuset); 15538 xt_sync(cpuset); 15539 kpreempt_enable(); 15540 15541 /* 15542 * At this stage we know that no trap handlers on other 15543 * cpus can have references to hmeblks on the list. 15544 */ 15545 sfmmu_hblk_free(listp); 15546 } else if (*listp != NULL) { 15547 pr_hblkp->hblk_next = cpuhp->chp_listp; 15548 cpuhp->chp_listp = *listp; 15549 cpuhp->chp_count += count; 15550 *listp = NULL; 15551 mutex_exit(&cpuhp->chp_mutex); 15552 } else { 15553 mutex_exit(&cpuhp->chp_mutex); 15554 } 15555 } 15556 15557 /* 15558 * Add an hmeblk to the the hash list. 15559 */ 15560 void 15561 sfmmu_hblk_hash_add(struct hmehash_bucket *hmebp, struct hme_blk *hmeblkp, 15562 uint64_t hblkpa) 15563 { 15564 ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp)); 15565 #ifdef DEBUG 15566 if (hmebp->hmeblkp == NULL) { 15567 ASSERT(hmebp->hmeh_nextpa == HMEBLK_ENDPA); 15568 } 15569 #endif /* DEBUG */ 15570 15571 hmeblkp->hblk_nextpa = hmebp->hmeh_nextpa; 15572 /* 15573 * Since the TSB miss handler now does not lock the hash chain before 15574 * walking it, make sure that the hmeblks nextpa is globally visible 15575 * before we make the hmeblk globally visible by updating the chain root 15576 * pointer in the hash bucket. 15577 */ 15578 membar_producer(); 15579 hmebp->hmeh_nextpa = hblkpa; 15580 hmeblkp->hblk_next = hmebp->hmeblkp; 15581 hmebp->hmeblkp = hmeblkp; 15582 15583 } 15584 15585 /* 15586 * This function is the first part of a 2 part process to remove an hmeblk 15587 * from the hash chain. In this phase we unlink the hmeblk from the hash chain 15588 * but leave the next physical pointer unchanged. The hmeblk is then linked onto 15589 * a per-cpu pending list using the virtual address pointer. 15590 * 15591 * TSB miss trap handlers that start after this phase will no longer see 15592 * this hmeblk. TSB miss handlers that still cache this hmeblk in a register 15593 * can still use it for further chain traversal because we haven't yet modifed 15594 * the next physical pointer or freed it. 15595 * 15596 * In the second phase of hmeblk removal we'll issue a barrier xcall before 15597 * we reuse or free this hmeblk. This will make sure all lingering references to 15598 * the hmeblk after first phase disappear before we finally reclaim it. 15599 * This scheme eliminates the need for TSB miss handlers to lock hmeblk chains 15600 * during their traversal. 15601 * 15602 * The hmehash_mutex must be held when calling this function. 15603 * 15604 * Input: 15605 * hmebp - hme hash bucket pointer 15606 * hmeblkp - address of hmeblk to be removed 15607 * pr_hblk - virtual address of previous hmeblkp 15608 * listp - pointer to list of hmeblks linked by virtual address 15609 * free_now flag - indicates that a complete removal from the hash chains 15610 * is necessary. 15611 * 15612 * It is inefficient to use the free_now flag as a cross-call is required to 15613 * remove a single hmeblk from the hash chain but is necessary when hmeblks are 15614 * in short supply. 15615 */ 15616 void 15617 sfmmu_hblk_hash_rm(struct hmehash_bucket *hmebp, struct hme_blk *hmeblkp, 15618 struct hme_blk *pr_hblk, struct hme_blk **listp, 15619 int free_now) 15620 { 15621 int shw_size, vshift; 15622 struct hme_blk *shw_hblkp; 15623 uint_t shw_mask, newshw_mask; 15624 caddr_t vaddr; 15625 int size; 15626 cpuset_t cpuset = cpu_ready_set; 15627 15628 ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp)); 15629 15630 if (hmebp->hmeblkp == hmeblkp) { 15631 hmebp->hmeh_nextpa = hmeblkp->hblk_nextpa; 15632 hmebp->hmeblkp = hmeblkp->hblk_next; 15633 } else { 15634 pr_hblk->hblk_nextpa = hmeblkp->hblk_nextpa; 15635 pr_hblk->hblk_next = hmeblkp->hblk_next; 15636 } 15637 15638 size = get_hblk_ttesz(hmeblkp); 15639 shw_hblkp = hmeblkp->hblk_shadow; 15640 if (shw_hblkp) { 15641 ASSERT(hblktosfmmu(hmeblkp) != KHATID); 15642 ASSERT(!hmeblkp->hblk_shared); 15643 #ifdef DEBUG 15644 if (mmu_page_sizes == max_mmu_page_sizes) { 15645 ASSERT(size < TTE256M); 15646 } else { 15647 ASSERT(size < TTE4M); 15648 } 15649 #endif /* DEBUG */ 15650 15651 shw_size = get_hblk_ttesz(shw_hblkp); 15652 vaddr = (caddr_t)get_hblk_base(hmeblkp); 15653 vshift = vaddr_to_vshift(shw_hblkp->hblk_tag, vaddr, shw_size); 15654 ASSERT(vshift < 8); 15655 /* 15656 * Atomically clear shadow mask bit 15657 */ 15658 do { 15659 shw_mask = shw_hblkp->hblk_shw_mask; 15660 ASSERT(shw_mask & (1 << vshift)); 15661 newshw_mask = shw_mask & ~(1 << vshift); 15662 newshw_mask = cas32(&shw_hblkp->hblk_shw_mask, 15663 shw_mask, newshw_mask); 15664 } while (newshw_mask != shw_mask); 15665 hmeblkp->hblk_shadow = NULL; 15666 } 15667 hmeblkp->hblk_shw_bit = 0; 15668 15669 if (hmeblkp->hblk_shared) { 15670 #ifdef DEBUG 15671 sf_srd_t *srdp; 15672 sf_region_t *rgnp; 15673 uint_t rid; 15674 15675 srdp = hblktosrd(hmeblkp); 15676 ASSERT(srdp != NULL && srdp->srd_refcnt != 0); 15677 rid = hmeblkp->hblk_tag.htag_rid; 15678 ASSERT(SFMMU_IS_SHMERID_VALID(rid)); 15679 ASSERT(rid < SFMMU_MAX_HME_REGIONS); 15680 rgnp = srdp->srd_hmergnp[rid]; 15681 ASSERT(rgnp != NULL); 15682 SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid); 15683 #endif /* DEBUG */ 15684 hmeblkp->hblk_shared = 0; 15685 } 15686 if (free_now) { 15687 kpreempt_disable(); 15688 CPUSET_DEL(cpuset, CPU->cpu_id); 15689 xt_sync(cpuset); 15690 xt_sync(cpuset); 15691 kpreempt_enable(); 15692 15693 hmeblkp->hblk_nextpa = HMEBLK_ENDPA; 15694 hmeblkp->hblk_next = NULL; 15695 } else { 15696 /* Append hmeblkp to listp for processing later. */ 15697 hmeblkp->hblk_next = *listp; 15698 *listp = hmeblkp; 15699 } 15700 } 15701 15702 /* 15703 * This routine is called when memory is in short supply and returns a free 15704 * hmeblk of the requested size from the cpu pending lists. 15705 */ 15706 static struct hme_blk * 15707 sfmmu_check_pending_hblks(int size) 15708 { 15709 int i; 15710 struct hme_blk *hmeblkp = NULL, *last_hmeblkp; 15711 int found_hmeblk; 15712 cpuset_t cpuset = cpu_ready_set; 15713 cpu_hme_pend_t *cpuhp; 15714 15715 /* Flush cpu hblk pending queues */ 15716 for (i = 0; i < NCPU; i++) { 15717 cpuhp = &cpu_hme_pend[i]; 15718 if (cpuhp->chp_listp != NULL) { 15719 mutex_enter(&cpuhp->chp_mutex); 15720 if (cpuhp->chp_listp == NULL) { 15721 mutex_exit(&cpuhp->chp_mutex); 15722 continue; 15723 } 15724 found_hmeblk = 0; 15725 last_hmeblkp = NULL; 15726 for (hmeblkp = cpuhp->chp_listp; hmeblkp != NULL; 15727 hmeblkp = hmeblkp->hblk_next) { 15728 if (get_hblk_ttesz(hmeblkp) == size) { 15729 if (last_hmeblkp == NULL) { 15730 cpuhp->chp_listp = 15731 hmeblkp->hblk_next; 15732 } else { 15733 last_hmeblkp->hblk_next = 15734 hmeblkp->hblk_next; 15735 } 15736 ASSERT(cpuhp->chp_count > 0); 15737 cpuhp->chp_count--; 15738 found_hmeblk = 1; 15739 break; 15740 } else { 15741 last_hmeblkp = hmeblkp; 15742 } 15743 } 15744 mutex_exit(&cpuhp->chp_mutex); 15745 15746 if (found_hmeblk) { 15747 kpreempt_disable(); 15748 CPUSET_DEL(cpuset, CPU->cpu_id); 15749 xt_sync(cpuset); 15750 xt_sync(cpuset); 15751 kpreempt_enable(); 15752 return (hmeblkp); 15753 } 15754 } 15755 } 15756 return (NULL); 15757 } 15758