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 2008 Sun Microsystems, Inc. All rights reserved. 23 * Use is subject to license terms. 24 */ 25 26 /* Copyright (c) 1983, 1984, 1985, 1986, 1987, 1988, 1989 AT&T */ 27 /* All Rights Reserved */ 28 29 /* 30 * University Copyright- Copyright (c) 1982, 1986, 1988 31 * The Regents of the University of California 32 * All Rights Reserved 33 * 34 * University Acknowledgment- Portions of this document are derived from 35 * software developed by the University of California, Berkeley, and its 36 * contributors. 37 */ 38 39 /* 40 * VM - physical page management. 41 */ 42 43 #include <sys/types.h> 44 #include <sys/t_lock.h> 45 #include <sys/param.h> 46 #include <sys/systm.h> 47 #include <sys/errno.h> 48 #include <sys/time.h> 49 #include <sys/vnode.h> 50 #include <sys/vm.h> 51 #include <sys/vtrace.h> 52 #include <sys/swap.h> 53 #include <sys/cmn_err.h> 54 #include <sys/tuneable.h> 55 #include <sys/sysmacros.h> 56 #include <sys/cpuvar.h> 57 #include <sys/callb.h> 58 #include <sys/debug.h> 59 #include <sys/tnf_probe.h> 60 #include <sys/condvar_impl.h> 61 #include <sys/mem_config.h> 62 #include <sys/mem_cage.h> 63 #include <sys/kmem.h> 64 #include <sys/atomic.h> 65 #include <sys/strlog.h> 66 #include <sys/mman.h> 67 #include <sys/ontrap.h> 68 #include <sys/lgrp.h> 69 #include <sys/vfs.h> 70 71 #include <vm/hat.h> 72 #include <vm/anon.h> 73 #include <vm/page.h> 74 #include <vm/seg.h> 75 #include <vm/pvn.h> 76 #include <vm/seg_kmem.h> 77 #include <vm/vm_dep.h> 78 #include <sys/vm_usage.h> 79 #include <fs/fs_subr.h> 80 #include <sys/ddi.h> 81 #include <sys/modctl.h> 82 83 static int nopageage = 0; 84 85 static pgcnt_t max_page_get; /* max page_get request size in pages */ 86 pgcnt_t total_pages = 0; /* total number of pages (used by /proc) */ 87 88 /* 89 * freemem_lock protects all freemem variables: 90 * availrmem. Also this lock protects the globals which track the 91 * availrmem changes for accurate kernel footprint calculation. 92 * See below for an explanation of these 93 * globals. 94 */ 95 kmutex_t freemem_lock; 96 pgcnt_t availrmem; 97 pgcnt_t availrmem_initial; 98 99 /* 100 * These globals track availrmem changes to get a more accurate 101 * estimate of tke kernel size. Historically pp_kernel is used for 102 * kernel size and is based on availrmem. But availrmem is adjusted for 103 * locked pages in the system not just for kernel locked pages. 104 * These new counters will track the pages locked through segvn and 105 * by explicit user locking. 106 * 107 * pages_locked : How many pages are locked because of user specified 108 * locking through mlock or plock. 109 * 110 * pages_useclaim,pages_claimed : These two variables track the 111 * claim adjustments because of the protection changes on a segvn segment. 112 * 113 * All these globals are protected by the same lock which protects availrmem. 114 */ 115 pgcnt_t pages_locked = 0; 116 pgcnt_t pages_useclaim = 0; 117 pgcnt_t pages_claimed = 0; 118 119 120 /* 121 * new_freemem_lock protects freemem, freemem_wait & freemem_cv. 122 */ 123 static kmutex_t new_freemem_lock; 124 static uint_t freemem_wait; /* someone waiting for freemem */ 125 static kcondvar_t freemem_cv; 126 127 /* 128 * The logical page free list is maintained as two lists, the 'free' 129 * and the 'cache' lists. 130 * The free list contains those pages that should be reused first. 131 * 132 * The implementation of the lists is machine dependent. 133 * page_get_freelist(), page_get_cachelist(), 134 * page_list_sub(), and page_list_add() 135 * form the interface to the machine dependent implementation. 136 * 137 * Pages with p_free set are on the cache list. 138 * Pages with p_free and p_age set are on the free list, 139 * 140 * A page may be locked while on either list. 141 */ 142 143 /* 144 * free list accounting stuff. 145 * 146 * 147 * Spread out the value for the number of pages on the 148 * page free and page cache lists. If there is just one 149 * value, then it must be under just one lock. 150 * The lock contention and cache traffic are a real bother. 151 * 152 * When we acquire and then drop a single pcf lock 153 * we can start in the middle of the array of pcf structures. 154 * If we acquire more than one pcf lock at a time, we need to 155 * start at the front to avoid deadlocking. 156 * 157 * pcf_count holds the number of pages in each pool. 158 * 159 * pcf_block is set when page_create_get_something() has asked the 160 * PSM page freelist and page cachelist routines without specifying 161 * a color and nothing came back. This is used to block anything 162 * else from moving pages from one list to the other while the 163 * lists are searched again. If a page is freeed while pcf_block is 164 * set, then pcf_reserve is incremented. pcgs_unblock() takes care 165 * of clearning pcf_block, doing the wakeups, etc. 166 */ 167 168 #define MAX_PCF_FANOUT NCPU 169 static uint_t pcf_fanout = 1; /* Will get changed at boot time */ 170 static uint_t pcf_fanout_mask = 0; 171 172 struct pcf { 173 kmutex_t pcf_lock; /* protects the structure */ 174 uint_t pcf_count; /* page count */ 175 uint_t pcf_wait; /* number of waiters */ 176 uint_t pcf_block; /* pcgs flag to page_free() */ 177 uint_t pcf_reserve; /* pages freed after pcf_block set */ 178 uint_t pcf_fill[10]; /* to line up on the caches */ 179 }; 180 181 /* 182 * PCF_INDEX hash needs to be dynamic (every so often the hash changes where 183 * it will hash the cpu to). This is done to prevent a drain condition 184 * from happening. This drain condition will occur when pcf_count decrement 185 * occurs on cpu A and the increment of pcf_count always occurs on cpu B. An 186 * example of this shows up with device interrupts. The dma buffer is allocated 187 * by the cpu requesting the IO thus the pcf_count is decremented based on that. 188 * When the memory is returned by the interrupt thread, the pcf_count will be 189 * incremented based on the cpu servicing the interrupt. 190 */ 191 static struct pcf pcf[MAX_PCF_FANOUT]; 192 #define PCF_INDEX() ((int)(((long)CPU->cpu_seqid) + \ 193 (randtick() >> 24)) & (pcf_fanout_mask)) 194 195 static int pcf_decrement_bucket(pgcnt_t); 196 static int pcf_decrement_multiple(pgcnt_t *, pgcnt_t, int); 197 198 kmutex_t pcgs_lock; /* serializes page_create_get_ */ 199 kmutex_t pcgs_cagelock; /* serializes NOSLEEP cage allocs */ 200 kmutex_t pcgs_wait_lock; /* used for delay in pcgs */ 201 static kcondvar_t pcgs_cv; /* cv for delay in pcgs */ 202 203 #ifdef VM_STATS 204 205 /* 206 * No locks, but so what, they are only statistics. 207 */ 208 209 static struct page_tcnt { 210 int pc_free_cache; /* free's into cache list */ 211 int pc_free_dontneed; /* free's with dontneed */ 212 int pc_free_pageout; /* free's from pageout */ 213 int pc_free_free; /* free's into free list */ 214 int pc_free_pages; /* free's into large page free list */ 215 int pc_destroy_pages; /* large page destroy's */ 216 int pc_get_cache; /* get's from cache list */ 217 int pc_get_free; /* get's from free list */ 218 int pc_reclaim; /* reclaim's */ 219 int pc_abortfree; /* abort's of free pages */ 220 int pc_find_hit; /* find's that find page */ 221 int pc_find_miss; /* find's that don't find page */ 222 int pc_destroy_free; /* # of free pages destroyed */ 223 #define PC_HASH_CNT (4*PAGE_HASHAVELEN) 224 int pc_find_hashlen[PC_HASH_CNT+1]; 225 int pc_addclaim_pages; 226 int pc_subclaim_pages; 227 int pc_free_replacement_page[2]; 228 int pc_try_demote_pages[6]; 229 int pc_demote_pages[2]; 230 } pagecnt; 231 232 uint_t hashin_count; 233 uint_t hashin_not_held; 234 uint_t hashin_already; 235 236 uint_t hashout_count; 237 uint_t hashout_not_held; 238 239 uint_t page_create_count; 240 uint_t page_create_not_enough; 241 uint_t page_create_not_enough_again; 242 uint_t page_create_zero; 243 uint_t page_create_hashout; 244 uint_t page_create_page_lock_failed; 245 uint_t page_create_trylock_failed; 246 uint_t page_create_found_one; 247 uint_t page_create_hashin_failed; 248 uint_t page_create_dropped_phm; 249 250 uint_t page_create_new; 251 uint_t page_create_exists; 252 uint_t page_create_putbacks; 253 uint_t page_create_overshoot; 254 255 uint_t page_reclaim_zero; 256 uint_t page_reclaim_zero_locked; 257 258 uint_t page_rename_exists; 259 uint_t page_rename_count; 260 261 uint_t page_lookup_cnt[20]; 262 uint_t page_lookup_nowait_cnt[10]; 263 uint_t page_find_cnt; 264 uint_t page_exists_cnt; 265 uint_t page_exists_forreal_cnt; 266 uint_t page_lookup_dev_cnt; 267 uint_t get_cachelist_cnt; 268 uint_t page_create_cnt[10]; 269 uint_t alloc_pages[9]; 270 uint_t page_exphcontg[19]; 271 uint_t page_create_large_cnt[10]; 272 273 /* 274 * Collects statistics. 275 */ 276 #define PAGE_HASH_SEARCH(index, pp, vp, off) { \ 277 uint_t mylen = 0; \ 278 \ 279 for ((pp) = page_hash[(index)]; (pp); (pp) = (pp)->p_hash, mylen++) { \ 280 if ((pp)->p_vnode == (vp) && (pp)->p_offset == (off)) \ 281 break; \ 282 } \ 283 if ((pp) != NULL) \ 284 pagecnt.pc_find_hit++; \ 285 else \ 286 pagecnt.pc_find_miss++; \ 287 if (mylen > PC_HASH_CNT) \ 288 mylen = PC_HASH_CNT; \ 289 pagecnt.pc_find_hashlen[mylen]++; \ 290 } 291 292 #else /* VM_STATS */ 293 294 /* 295 * Don't collect statistics 296 */ 297 #define PAGE_HASH_SEARCH(index, pp, vp, off) { \ 298 for ((pp) = page_hash[(index)]; (pp); (pp) = (pp)->p_hash) { \ 299 if ((pp)->p_vnode == (vp) && (pp)->p_offset == (off)) \ 300 break; \ 301 } \ 302 } 303 304 #endif /* VM_STATS */ 305 306 307 308 #ifdef DEBUG 309 #define MEMSEG_SEARCH_STATS 310 #endif 311 312 #ifdef MEMSEG_SEARCH_STATS 313 struct memseg_stats { 314 uint_t nsearch; 315 uint_t nlastwon; 316 uint_t nhashwon; 317 uint_t nnotfound; 318 } memseg_stats; 319 320 #define MEMSEG_STAT_INCR(v) \ 321 atomic_add_32(&memseg_stats.v, 1) 322 #else 323 #define MEMSEG_STAT_INCR(x) 324 #endif 325 326 struct memseg *memsegs; /* list of memory segments */ 327 328 /* 329 * /etc/system tunable to control large page allocation hueristic. 330 * 331 * Setting to LPAP_LOCAL will heavily prefer the local lgroup over remote lgroup 332 * for large page allocation requests. If a large page is not readily 333 * avaliable on the local freelists we will go through additional effort 334 * to create a large page, potentially moving smaller pages around to coalesce 335 * larger pages in the local lgroup. 336 * Default value of LPAP_DEFAULT will go to remote freelists if large pages 337 * are not readily available in the local lgroup. 338 */ 339 enum lpap { 340 LPAP_DEFAULT, /* default large page allocation policy */ 341 LPAP_LOCAL /* local large page allocation policy */ 342 }; 343 344 enum lpap lpg_alloc_prefer = LPAP_DEFAULT; 345 346 static void page_init_mem_config(void); 347 static int page_do_hashin(page_t *, vnode_t *, u_offset_t); 348 static void page_do_hashout(page_t *); 349 static void page_capture_init(); 350 int page_capture_take_action(page_t *, uint_t, void *); 351 352 static void page_demote_vp_pages(page_t *); 353 354 355 void 356 pcf_init(void) 357 358 { 359 if (boot_ncpus != -1) { 360 pcf_fanout = boot_ncpus; 361 } else { 362 pcf_fanout = max_ncpus; 363 } 364 #ifdef sun4v 365 /* 366 * Force at least 4 buckets if possible for sun4v. 367 */ 368 pcf_fanout = MAX(pcf_fanout, 4); 369 #endif /* sun4v */ 370 371 /* 372 * Round up to the nearest power of 2. 373 */ 374 pcf_fanout = MIN(pcf_fanout, MAX_PCF_FANOUT); 375 if (!ISP2(pcf_fanout)) { 376 pcf_fanout = 1 << highbit(pcf_fanout); 377 378 if (pcf_fanout > MAX_PCF_FANOUT) { 379 pcf_fanout = 1 << (highbit(MAX_PCF_FANOUT) - 1); 380 } 381 } 382 pcf_fanout_mask = pcf_fanout - 1; 383 } 384 385 /* 386 * vm subsystem related initialization 387 */ 388 void 389 vm_init(void) 390 { 391 boolean_t callb_vm_cpr(void *, int); 392 393 (void) callb_add(callb_vm_cpr, 0, CB_CL_CPR_VM, "vm"); 394 page_init_mem_config(); 395 page_retire_init(); 396 vm_usage_init(); 397 page_capture_init(); 398 } 399 400 /* 401 * This function is called at startup and when memory is added or deleted. 402 */ 403 void 404 init_pages_pp_maximum() 405 { 406 static pgcnt_t p_min; 407 static pgcnt_t pages_pp_maximum_startup; 408 static pgcnt_t avrmem_delta; 409 static int init_done; 410 static int user_set; /* true if set in /etc/system */ 411 412 if (init_done == 0) { 413 414 /* If the user specified a value, save it */ 415 if (pages_pp_maximum != 0) { 416 user_set = 1; 417 pages_pp_maximum_startup = pages_pp_maximum; 418 } 419 420 /* 421 * Setting of pages_pp_maximum is based first time 422 * on the value of availrmem just after the start-up 423 * allocations. To preserve this relationship at run 424 * time, use a delta from availrmem_initial. 425 */ 426 ASSERT(availrmem_initial >= availrmem); 427 avrmem_delta = availrmem_initial - availrmem; 428 429 /* The allowable floor of pages_pp_maximum */ 430 p_min = tune.t_minarmem + 100; 431 432 /* Make sure we don't come through here again. */ 433 init_done = 1; 434 } 435 /* 436 * Determine pages_pp_maximum, the number of currently available 437 * pages (availrmem) that can't be `locked'. If not set by 438 * the user, we set it to 4% of the currently available memory 439 * plus 4MB. 440 * But we also insist that it be greater than tune.t_minarmem; 441 * otherwise a process could lock down a lot of memory, get swapped 442 * out, and never have enough to get swapped back in. 443 */ 444 if (user_set) 445 pages_pp_maximum = pages_pp_maximum_startup; 446 else 447 pages_pp_maximum = ((availrmem_initial - avrmem_delta) / 25) 448 + btop(4 * 1024 * 1024); 449 450 if (pages_pp_maximum <= p_min) { 451 pages_pp_maximum = p_min; 452 } 453 } 454 455 void 456 set_max_page_get(pgcnt_t target_total_pages) 457 { 458 max_page_get = target_total_pages / 2; 459 } 460 461 static pgcnt_t pending_delete; 462 463 /*ARGSUSED*/ 464 static void 465 page_mem_config_post_add( 466 void *arg, 467 pgcnt_t delta_pages) 468 { 469 set_max_page_get(total_pages - pending_delete); 470 init_pages_pp_maximum(); 471 } 472 473 /*ARGSUSED*/ 474 static int 475 page_mem_config_pre_del( 476 void *arg, 477 pgcnt_t delta_pages) 478 { 479 pgcnt_t nv; 480 481 nv = atomic_add_long_nv(&pending_delete, (spgcnt_t)delta_pages); 482 set_max_page_get(total_pages - nv); 483 return (0); 484 } 485 486 /*ARGSUSED*/ 487 static void 488 page_mem_config_post_del( 489 void *arg, 490 pgcnt_t delta_pages, 491 int cancelled) 492 { 493 pgcnt_t nv; 494 495 nv = atomic_add_long_nv(&pending_delete, -(spgcnt_t)delta_pages); 496 set_max_page_get(total_pages - nv); 497 if (!cancelled) 498 init_pages_pp_maximum(); 499 } 500 501 static kphysm_setup_vector_t page_mem_config_vec = { 502 KPHYSM_SETUP_VECTOR_VERSION, 503 page_mem_config_post_add, 504 page_mem_config_pre_del, 505 page_mem_config_post_del, 506 }; 507 508 static void 509 page_init_mem_config(void) 510 { 511 int ret; 512 513 ret = kphysm_setup_func_register(&page_mem_config_vec, (void *)NULL); 514 ASSERT(ret == 0); 515 } 516 517 /* 518 * Evenly spread out the PCF counters for large free pages 519 */ 520 static void 521 page_free_large_ctr(pgcnt_t npages) 522 { 523 static struct pcf *p = pcf; 524 pgcnt_t lump; 525 526 freemem += npages; 527 528 lump = roundup(npages, pcf_fanout) / pcf_fanout; 529 530 while (npages > 0) { 531 532 ASSERT(!p->pcf_block); 533 534 if (lump < npages) { 535 p->pcf_count += (uint_t)lump; 536 npages -= lump; 537 } else { 538 p->pcf_count += (uint_t)npages; 539 npages = 0; 540 } 541 542 ASSERT(!p->pcf_wait); 543 544 if (++p > &pcf[pcf_fanout - 1]) 545 p = pcf; 546 } 547 548 ASSERT(npages == 0); 549 } 550 551 /* 552 * Add a physical chunk of memory to the system free lists during startup. 553 * Platform specific startup() allocates the memory for the page structs. 554 * 555 * num - number of page structures 556 * base - page number (pfn) to be associated with the first page. 557 * 558 * Since we are doing this during startup (ie. single threaded), we will 559 * use shortcut routines to avoid any locking overhead while putting all 560 * these pages on the freelists. 561 * 562 * NOTE: Any changes performed to page_free(), must also be performed to 563 * add_physmem() since this is how we initialize all page_t's at 564 * boot time. 565 */ 566 void 567 add_physmem( 568 page_t *pp, 569 pgcnt_t num, 570 pfn_t pnum) 571 { 572 page_t *root = NULL; 573 uint_t szc = page_num_pagesizes() - 1; 574 pgcnt_t large = page_get_pagecnt(szc); 575 pgcnt_t cnt = 0; 576 577 TRACE_2(TR_FAC_VM, TR_PAGE_INIT, 578 "add_physmem:pp %p num %lu", pp, num); 579 580 /* 581 * Arbitrarily limit the max page_get request 582 * to 1/2 of the page structs we have. 583 */ 584 total_pages += num; 585 set_max_page_get(total_pages); 586 587 PLCNT_MODIFY_MAX(pnum, (long)num); 588 589 /* 590 * The physical space for the pages array 591 * representing ram pages has already been 592 * allocated. Here we initialize each lock 593 * in the page structure, and put each on 594 * the free list 595 */ 596 for (; num; pp++, pnum++, num--) { 597 598 /* 599 * this needs to fill in the page number 600 * and do any other arch specific initialization 601 */ 602 add_physmem_cb(pp, pnum); 603 604 pp->p_lckcnt = 0; 605 pp->p_cowcnt = 0; 606 pp->p_slckcnt = 0; 607 608 /* 609 * Initialize the page lock as unlocked, since nobody 610 * can see or access this page yet. 611 */ 612 pp->p_selock = 0; 613 614 /* 615 * Initialize IO lock 616 */ 617 page_iolock_init(pp); 618 619 /* 620 * initialize other fields in the page_t 621 */ 622 PP_SETFREE(pp); 623 page_clr_all_props(pp, 0); 624 PP_SETAGED(pp); 625 pp->p_offset = (u_offset_t)-1; 626 pp->p_next = pp; 627 pp->p_prev = pp; 628 629 /* 630 * Simple case: System doesn't support large pages. 631 */ 632 if (szc == 0) { 633 pp->p_szc = 0; 634 page_free_at_startup(pp); 635 continue; 636 } 637 638 /* 639 * Handle unaligned pages, we collect them up onto 640 * the root page until we have a full large page. 641 */ 642 if (!IS_P2ALIGNED(pnum, large)) { 643 644 /* 645 * If not in a large page, 646 * just free as small page. 647 */ 648 if (root == NULL) { 649 pp->p_szc = 0; 650 page_free_at_startup(pp); 651 continue; 652 } 653 654 /* 655 * Link a constituent page into the large page. 656 */ 657 pp->p_szc = szc; 658 page_list_concat(&root, &pp); 659 660 /* 661 * When large page is fully formed, free it. 662 */ 663 if (++cnt == large) { 664 page_free_large_ctr(cnt); 665 page_list_add_pages(root, PG_LIST_ISINIT); 666 root = NULL; 667 cnt = 0; 668 } 669 continue; 670 } 671 672 /* 673 * At this point we have a page number which 674 * is aligned. We assert that we aren't already 675 * in a different large page. 676 */ 677 ASSERT(IS_P2ALIGNED(pnum, large)); 678 ASSERT(root == NULL && cnt == 0); 679 680 /* 681 * If insufficient number of pages left to form 682 * a large page, just free the small page. 683 */ 684 if (num < large) { 685 pp->p_szc = 0; 686 page_free_at_startup(pp); 687 continue; 688 } 689 690 /* 691 * Otherwise start a new large page. 692 */ 693 pp->p_szc = szc; 694 cnt++; 695 root = pp; 696 } 697 ASSERT(root == NULL && cnt == 0); 698 } 699 700 /* 701 * Find a page representing the specified [vp, offset]. 702 * If we find the page but it is intransit coming in, 703 * it will have an "exclusive" lock and we wait for 704 * the i/o to complete. A page found on the free list 705 * is always reclaimed and then locked. On success, the page 706 * is locked, its data is valid and it isn't on the free 707 * list, while a NULL is returned if the page doesn't exist. 708 */ 709 page_t * 710 page_lookup(vnode_t *vp, u_offset_t off, se_t se) 711 { 712 return (page_lookup_create(vp, off, se, NULL, NULL, 0)); 713 } 714 715 /* 716 * Find a page representing the specified [vp, offset]. 717 * We either return the one we found or, if passed in, 718 * create one with identity of [vp, offset] of the 719 * pre-allocated page. If we find existing page but it is 720 * intransit coming in, it will have an "exclusive" lock 721 * and we wait for the i/o to complete. A page found on 722 * the free list is always reclaimed and then locked. 723 * On success, the page is locked, its data is valid and 724 * it isn't on the free list, while a NULL is returned 725 * if the page doesn't exist and newpp is NULL; 726 */ 727 page_t * 728 page_lookup_create( 729 vnode_t *vp, 730 u_offset_t off, 731 se_t se, 732 page_t *newpp, 733 spgcnt_t *nrelocp, 734 int flags) 735 { 736 page_t *pp; 737 kmutex_t *phm; 738 ulong_t index; 739 uint_t hash_locked; 740 uint_t es; 741 742 ASSERT(MUTEX_NOT_HELD(page_vnode_mutex(vp))); 743 VM_STAT_ADD(page_lookup_cnt[0]); 744 ASSERT(newpp ? PAGE_EXCL(newpp) : 1); 745 746 /* 747 * Acquire the appropriate page hash lock since 748 * we have to search the hash list. Pages that 749 * hash to this list can't change identity while 750 * this lock is held. 751 */ 752 hash_locked = 0; 753 index = PAGE_HASH_FUNC(vp, off); 754 phm = NULL; 755 top: 756 PAGE_HASH_SEARCH(index, pp, vp, off); 757 if (pp != NULL) { 758 VM_STAT_ADD(page_lookup_cnt[1]); 759 es = (newpp != NULL) ? 1 : 0; 760 es |= flags; 761 if (!hash_locked) { 762 VM_STAT_ADD(page_lookup_cnt[2]); 763 if (!page_try_reclaim_lock(pp, se, es)) { 764 /* 765 * On a miss, acquire the phm. Then 766 * next time, page_lock() will be called, 767 * causing a wait if the page is busy. 768 * just looping with page_trylock() would 769 * get pretty boring. 770 */ 771 VM_STAT_ADD(page_lookup_cnt[3]); 772 phm = PAGE_HASH_MUTEX(index); 773 mutex_enter(phm); 774 hash_locked = 1; 775 goto top; 776 } 777 } else { 778 VM_STAT_ADD(page_lookup_cnt[4]); 779 if (!page_lock_es(pp, se, phm, P_RECLAIM, es)) { 780 VM_STAT_ADD(page_lookup_cnt[5]); 781 goto top; 782 } 783 } 784 785 /* 786 * Since `pp' is locked it can not change identity now. 787 * Reconfirm we locked the correct page. 788 * 789 * Both the p_vnode and p_offset *must* be cast volatile 790 * to force a reload of their values: The PAGE_HASH_SEARCH 791 * macro will have stuffed p_vnode and p_offset into 792 * registers before calling page_trylock(); another thread, 793 * actually holding the hash lock, could have changed the 794 * page's identity in memory, but our registers would not 795 * be changed, fooling the reconfirmation. If the hash 796 * lock was held during the search, the casting would 797 * not be needed. 798 */ 799 VM_STAT_ADD(page_lookup_cnt[6]); 800 if (((volatile struct vnode *)(pp->p_vnode) != vp) || 801 ((volatile u_offset_t)(pp->p_offset) != off)) { 802 VM_STAT_ADD(page_lookup_cnt[7]); 803 if (hash_locked) { 804 panic("page_lookup_create: lost page %p", 805 (void *)pp); 806 /*NOTREACHED*/ 807 } 808 page_unlock(pp); 809 phm = PAGE_HASH_MUTEX(index); 810 mutex_enter(phm); 811 hash_locked = 1; 812 goto top; 813 } 814 815 /* 816 * If page_trylock() was called, then pp may still be on 817 * the cachelist (can't be on the free list, it would not 818 * have been found in the search). If it is on the 819 * cachelist it must be pulled now. To pull the page from 820 * the cachelist, it must be exclusively locked. 821 * 822 * The other big difference between page_trylock() and 823 * page_lock(), is that page_lock() will pull the 824 * page from whatever free list (the cache list in this 825 * case) the page is on. If page_trylock() was used 826 * above, then we have to do the reclaim ourselves. 827 */ 828 if ((!hash_locked) && (PP_ISFREE(pp))) { 829 ASSERT(PP_ISAGED(pp) == 0); 830 VM_STAT_ADD(page_lookup_cnt[8]); 831 832 /* 833 * page_relcaim will insure that we 834 * have this page exclusively 835 */ 836 837 if (!page_reclaim(pp, NULL)) { 838 /* 839 * Page_reclaim dropped whatever lock 840 * we held. 841 */ 842 VM_STAT_ADD(page_lookup_cnt[9]); 843 phm = PAGE_HASH_MUTEX(index); 844 mutex_enter(phm); 845 hash_locked = 1; 846 goto top; 847 } else if (se == SE_SHARED && newpp == NULL) { 848 VM_STAT_ADD(page_lookup_cnt[10]); 849 page_downgrade(pp); 850 } 851 } 852 853 if (hash_locked) { 854 mutex_exit(phm); 855 } 856 857 if (newpp != NULL && pp->p_szc < newpp->p_szc && 858 PAGE_EXCL(pp) && nrelocp != NULL) { 859 ASSERT(nrelocp != NULL); 860 (void) page_relocate(&pp, &newpp, 1, 1, nrelocp, 861 NULL); 862 if (*nrelocp > 0) { 863 VM_STAT_COND_ADD(*nrelocp == 1, 864 page_lookup_cnt[11]); 865 VM_STAT_COND_ADD(*nrelocp > 1, 866 page_lookup_cnt[12]); 867 pp = newpp; 868 se = SE_EXCL; 869 } else { 870 if (se == SE_SHARED) { 871 page_downgrade(pp); 872 } 873 VM_STAT_ADD(page_lookup_cnt[13]); 874 } 875 } else if (newpp != NULL && nrelocp != NULL) { 876 if (PAGE_EXCL(pp) && se == SE_SHARED) { 877 page_downgrade(pp); 878 } 879 VM_STAT_COND_ADD(pp->p_szc < newpp->p_szc, 880 page_lookup_cnt[14]); 881 VM_STAT_COND_ADD(pp->p_szc == newpp->p_szc, 882 page_lookup_cnt[15]); 883 VM_STAT_COND_ADD(pp->p_szc > newpp->p_szc, 884 page_lookup_cnt[16]); 885 } else if (newpp != NULL && PAGE_EXCL(pp)) { 886 se = SE_EXCL; 887 } 888 } else if (!hash_locked) { 889 VM_STAT_ADD(page_lookup_cnt[17]); 890 phm = PAGE_HASH_MUTEX(index); 891 mutex_enter(phm); 892 hash_locked = 1; 893 goto top; 894 } else if (newpp != NULL) { 895 /* 896 * If we have a preallocated page then 897 * insert it now and basically behave like 898 * page_create. 899 */ 900 VM_STAT_ADD(page_lookup_cnt[18]); 901 /* 902 * Since we hold the page hash mutex and 903 * just searched for this page, page_hashin 904 * had better not fail. If it does, that 905 * means some thread did not follow the 906 * page hash mutex rules. Panic now and 907 * get it over with. As usual, go down 908 * holding all the locks. 909 */ 910 ASSERT(MUTEX_HELD(phm)); 911 if (!page_hashin(newpp, vp, off, phm)) { 912 ASSERT(MUTEX_HELD(phm)); 913 panic("page_lookup_create: hashin failed %p %p %llx %p", 914 (void *)newpp, (void *)vp, off, (void *)phm); 915 /*NOTREACHED*/ 916 } 917 ASSERT(MUTEX_HELD(phm)); 918 mutex_exit(phm); 919 phm = NULL; 920 page_set_props(newpp, P_REF); 921 page_io_lock(newpp); 922 pp = newpp; 923 se = SE_EXCL; 924 } else { 925 VM_STAT_ADD(page_lookup_cnt[19]); 926 mutex_exit(phm); 927 } 928 929 ASSERT(pp ? PAGE_LOCKED_SE(pp, se) : 1); 930 931 ASSERT(pp ? ((PP_ISFREE(pp) == 0) && (PP_ISAGED(pp) == 0)) : 1); 932 933 return (pp); 934 } 935 936 /* 937 * Search the hash list for the page representing the 938 * specified [vp, offset] and return it locked. Skip 939 * free pages and pages that cannot be locked as requested. 940 * Used while attempting to kluster pages. 941 */ 942 page_t * 943 page_lookup_nowait(vnode_t *vp, u_offset_t off, se_t se) 944 { 945 page_t *pp; 946 kmutex_t *phm; 947 ulong_t index; 948 uint_t locked; 949 950 ASSERT(MUTEX_NOT_HELD(page_vnode_mutex(vp))); 951 VM_STAT_ADD(page_lookup_nowait_cnt[0]); 952 953 index = PAGE_HASH_FUNC(vp, off); 954 PAGE_HASH_SEARCH(index, pp, vp, off); 955 locked = 0; 956 if (pp == NULL) { 957 top: 958 VM_STAT_ADD(page_lookup_nowait_cnt[1]); 959 locked = 1; 960 phm = PAGE_HASH_MUTEX(index); 961 mutex_enter(phm); 962 PAGE_HASH_SEARCH(index, pp, vp, off); 963 } 964 965 if (pp == NULL || PP_ISFREE(pp)) { 966 VM_STAT_ADD(page_lookup_nowait_cnt[2]); 967 pp = NULL; 968 } else { 969 if (!page_trylock(pp, se)) { 970 VM_STAT_ADD(page_lookup_nowait_cnt[3]); 971 pp = NULL; 972 } else { 973 VM_STAT_ADD(page_lookup_nowait_cnt[4]); 974 /* 975 * See the comment in page_lookup() 976 */ 977 if (((volatile struct vnode *)(pp->p_vnode) != vp) || 978 ((u_offset_t)(pp->p_offset) != off)) { 979 VM_STAT_ADD(page_lookup_nowait_cnt[5]); 980 if (locked) { 981 panic("page_lookup_nowait %p", 982 (void *)pp); 983 /*NOTREACHED*/ 984 } 985 page_unlock(pp); 986 goto top; 987 } 988 if (PP_ISFREE(pp)) { 989 VM_STAT_ADD(page_lookup_nowait_cnt[6]); 990 page_unlock(pp); 991 pp = NULL; 992 } 993 } 994 } 995 if (locked) { 996 VM_STAT_ADD(page_lookup_nowait_cnt[7]); 997 mutex_exit(phm); 998 } 999 1000 ASSERT(pp ? PAGE_LOCKED_SE(pp, se) : 1); 1001 1002 return (pp); 1003 } 1004 1005 /* 1006 * Search the hash list for a page with the specified [vp, off] 1007 * that is known to exist and is already locked. This routine 1008 * is typically used by segment SOFTUNLOCK routines. 1009 */ 1010 page_t * 1011 page_find(vnode_t *vp, u_offset_t off) 1012 { 1013 page_t *pp; 1014 kmutex_t *phm; 1015 ulong_t index; 1016 1017 ASSERT(MUTEX_NOT_HELD(page_vnode_mutex(vp))); 1018 VM_STAT_ADD(page_find_cnt); 1019 1020 index = PAGE_HASH_FUNC(vp, off); 1021 phm = PAGE_HASH_MUTEX(index); 1022 1023 mutex_enter(phm); 1024 PAGE_HASH_SEARCH(index, pp, vp, off); 1025 mutex_exit(phm); 1026 1027 ASSERT(pp == NULL || PAGE_LOCKED(pp) || panicstr); 1028 return (pp); 1029 } 1030 1031 /* 1032 * Determine whether a page with the specified [vp, off] 1033 * currently exists in the system. Obviously this should 1034 * only be considered as a hint since nothing prevents the 1035 * page from disappearing or appearing immediately after 1036 * the return from this routine. Subsequently, we don't 1037 * even bother to lock the list. 1038 */ 1039 page_t * 1040 page_exists(vnode_t *vp, u_offset_t off) 1041 { 1042 page_t *pp; 1043 ulong_t index; 1044 1045 ASSERT(MUTEX_NOT_HELD(page_vnode_mutex(vp))); 1046 VM_STAT_ADD(page_exists_cnt); 1047 1048 index = PAGE_HASH_FUNC(vp, off); 1049 PAGE_HASH_SEARCH(index, pp, vp, off); 1050 1051 return (pp); 1052 } 1053 1054 /* 1055 * Determine if physically contiguous pages exist for [vp, off] - [vp, off + 1056 * page_size(szc)) range. if they exist and ppa is not NULL fill ppa array 1057 * with these pages locked SHARED. If necessary reclaim pages from 1058 * freelist. Return 1 if contiguous pages exist and 0 otherwise. 1059 * 1060 * If we fail to lock pages still return 1 if pages exist and contiguous. 1061 * But in this case return value is just a hint. ppa array won't be filled. 1062 * Caller should initialize ppa[0] as NULL to distinguish return value. 1063 * 1064 * Returns 0 if pages don't exist or not physically contiguous. 1065 * 1066 * This routine doesn't work for anonymous(swapfs) pages. 1067 */ 1068 int 1069 page_exists_physcontig(vnode_t *vp, u_offset_t off, uint_t szc, page_t *ppa[]) 1070 { 1071 pgcnt_t pages; 1072 pfn_t pfn; 1073 page_t *rootpp; 1074 pgcnt_t i; 1075 pgcnt_t j; 1076 u_offset_t save_off = off; 1077 ulong_t index; 1078 kmutex_t *phm; 1079 page_t *pp; 1080 uint_t pszc; 1081 int loopcnt = 0; 1082 1083 ASSERT(szc != 0); 1084 ASSERT(vp != NULL); 1085 ASSERT(!IS_SWAPFSVP(vp)); 1086 ASSERT(!VN_ISKAS(vp)); 1087 1088 again: 1089 if (++loopcnt > 3) { 1090 VM_STAT_ADD(page_exphcontg[0]); 1091 return (0); 1092 } 1093 1094 index = PAGE_HASH_FUNC(vp, off); 1095 phm = PAGE_HASH_MUTEX(index); 1096 1097 mutex_enter(phm); 1098 PAGE_HASH_SEARCH(index, pp, vp, off); 1099 mutex_exit(phm); 1100 1101 VM_STAT_ADD(page_exphcontg[1]); 1102 1103 if (pp == NULL) { 1104 VM_STAT_ADD(page_exphcontg[2]); 1105 return (0); 1106 } 1107 1108 pages = page_get_pagecnt(szc); 1109 rootpp = pp; 1110 pfn = rootpp->p_pagenum; 1111 1112 if ((pszc = pp->p_szc) >= szc && ppa != NULL) { 1113 VM_STAT_ADD(page_exphcontg[3]); 1114 if (!page_trylock(pp, SE_SHARED)) { 1115 VM_STAT_ADD(page_exphcontg[4]); 1116 return (1); 1117 } 1118 if (pp->p_szc != pszc || pp->p_vnode != vp || 1119 pp->p_offset != off) { 1120 VM_STAT_ADD(page_exphcontg[5]); 1121 page_unlock(pp); 1122 off = save_off; 1123 goto again; 1124 } 1125 /* 1126 * szc was non zero and vnode and offset matched after we 1127 * locked the page it means it can't become free on us. 1128 */ 1129 ASSERT(!PP_ISFREE(pp)); 1130 if (!IS_P2ALIGNED(pfn, pages)) { 1131 page_unlock(pp); 1132 return (0); 1133 } 1134 ppa[0] = pp; 1135 pp++; 1136 off += PAGESIZE; 1137 pfn++; 1138 for (i = 1; i < pages; i++, pp++, off += PAGESIZE, pfn++) { 1139 if (!page_trylock(pp, SE_SHARED)) { 1140 VM_STAT_ADD(page_exphcontg[6]); 1141 pp--; 1142 while (i-- > 0) { 1143 page_unlock(pp); 1144 pp--; 1145 } 1146 ppa[0] = NULL; 1147 return (1); 1148 } 1149 if (pp->p_szc != pszc) { 1150 VM_STAT_ADD(page_exphcontg[7]); 1151 page_unlock(pp); 1152 pp--; 1153 while (i-- > 0) { 1154 page_unlock(pp); 1155 pp--; 1156 } 1157 ppa[0] = NULL; 1158 off = save_off; 1159 goto again; 1160 } 1161 /* 1162 * szc the same as for previous already locked pages 1163 * with right identity. Since this page had correct 1164 * szc after we locked it can't get freed or destroyed 1165 * and therefore must have the expected identity. 1166 */ 1167 ASSERT(!PP_ISFREE(pp)); 1168 if (pp->p_vnode != vp || 1169 pp->p_offset != off) { 1170 panic("page_exists_physcontig: " 1171 "large page identity doesn't match"); 1172 } 1173 ppa[i] = pp; 1174 ASSERT(pp->p_pagenum == pfn); 1175 } 1176 VM_STAT_ADD(page_exphcontg[8]); 1177 ppa[pages] = NULL; 1178 return (1); 1179 } else if (pszc >= szc) { 1180 VM_STAT_ADD(page_exphcontg[9]); 1181 if (!IS_P2ALIGNED(pfn, pages)) { 1182 return (0); 1183 } 1184 return (1); 1185 } 1186 1187 if (!IS_P2ALIGNED(pfn, pages)) { 1188 VM_STAT_ADD(page_exphcontg[10]); 1189 return (0); 1190 } 1191 1192 if (page_numtomemseg_nolock(pfn) != 1193 page_numtomemseg_nolock(pfn + pages - 1)) { 1194 VM_STAT_ADD(page_exphcontg[11]); 1195 return (0); 1196 } 1197 1198 /* 1199 * We loop up 4 times across pages to promote page size. 1200 * We're extra cautious to promote page size atomically with respect 1201 * to everybody else. But we can probably optimize into 1 loop if 1202 * this becomes an issue. 1203 */ 1204 1205 for (i = 0; i < pages; i++, pp++, off += PAGESIZE, pfn++) { 1206 ASSERT(pp->p_pagenum == pfn); 1207 if (!page_trylock(pp, SE_EXCL)) { 1208 VM_STAT_ADD(page_exphcontg[12]); 1209 break; 1210 } 1211 if (pp->p_vnode != vp || 1212 pp->p_offset != off) { 1213 VM_STAT_ADD(page_exphcontg[13]); 1214 page_unlock(pp); 1215 break; 1216 } 1217 if (pp->p_szc >= szc) { 1218 ASSERT(i == 0); 1219 page_unlock(pp); 1220 off = save_off; 1221 goto again; 1222 } 1223 } 1224 1225 if (i != pages) { 1226 VM_STAT_ADD(page_exphcontg[14]); 1227 --pp; 1228 while (i-- > 0) { 1229 page_unlock(pp); 1230 --pp; 1231 } 1232 return (0); 1233 } 1234 1235 pp = rootpp; 1236 for (i = 0; i < pages; i++, pp++) { 1237 if (PP_ISFREE(pp)) { 1238 VM_STAT_ADD(page_exphcontg[15]); 1239 ASSERT(!PP_ISAGED(pp)); 1240 ASSERT(pp->p_szc == 0); 1241 if (!page_reclaim(pp, NULL)) { 1242 break; 1243 } 1244 } else { 1245 ASSERT(pp->p_szc < szc); 1246 VM_STAT_ADD(page_exphcontg[16]); 1247 (void) hat_pageunload(pp, HAT_FORCE_PGUNLOAD); 1248 } 1249 } 1250 if (i < pages) { 1251 VM_STAT_ADD(page_exphcontg[17]); 1252 /* 1253 * page_reclaim failed because we were out of memory. 1254 * drop the rest of the locks and return because this page 1255 * must be already reallocated anyway. 1256 */ 1257 pp = rootpp; 1258 for (j = 0; j < pages; j++, pp++) { 1259 if (j != i) { 1260 page_unlock(pp); 1261 } 1262 } 1263 return (0); 1264 } 1265 1266 off = save_off; 1267 pp = rootpp; 1268 for (i = 0; i < pages; i++, pp++, off += PAGESIZE) { 1269 ASSERT(PAGE_EXCL(pp)); 1270 ASSERT(!PP_ISFREE(pp)); 1271 ASSERT(!hat_page_is_mapped(pp)); 1272 ASSERT(pp->p_vnode == vp); 1273 ASSERT(pp->p_offset == off); 1274 pp->p_szc = szc; 1275 } 1276 pp = rootpp; 1277 for (i = 0; i < pages; i++, pp++) { 1278 if (ppa == NULL) { 1279 page_unlock(pp); 1280 } else { 1281 ppa[i] = pp; 1282 page_downgrade(ppa[i]); 1283 } 1284 } 1285 if (ppa != NULL) { 1286 ppa[pages] = NULL; 1287 } 1288 VM_STAT_ADD(page_exphcontg[18]); 1289 ASSERT(vp->v_pages != NULL); 1290 return (1); 1291 } 1292 1293 /* 1294 * Determine whether a page with the specified [vp, off] 1295 * currently exists in the system and if so return its 1296 * size code. Obviously this should only be considered as 1297 * a hint since nothing prevents the page from disappearing 1298 * or appearing immediately after the return from this routine. 1299 */ 1300 int 1301 page_exists_forreal(vnode_t *vp, u_offset_t off, uint_t *szc) 1302 { 1303 page_t *pp; 1304 kmutex_t *phm; 1305 ulong_t index; 1306 int rc = 0; 1307 1308 ASSERT(MUTEX_NOT_HELD(page_vnode_mutex(vp))); 1309 ASSERT(szc != NULL); 1310 VM_STAT_ADD(page_exists_forreal_cnt); 1311 1312 index = PAGE_HASH_FUNC(vp, off); 1313 phm = PAGE_HASH_MUTEX(index); 1314 1315 mutex_enter(phm); 1316 PAGE_HASH_SEARCH(index, pp, vp, off); 1317 if (pp != NULL) { 1318 *szc = pp->p_szc; 1319 rc = 1; 1320 } 1321 mutex_exit(phm); 1322 return (rc); 1323 } 1324 1325 /* wakeup threads waiting for pages in page_create_get_something() */ 1326 void 1327 wakeup_pcgs(void) 1328 { 1329 if (!CV_HAS_WAITERS(&pcgs_cv)) 1330 return; 1331 cv_broadcast(&pcgs_cv); 1332 } 1333 1334 /* 1335 * 'freemem' is used all over the kernel as an indication of how many 1336 * pages are free (either on the cache list or on the free page list) 1337 * in the system. In very few places is a really accurate 'freemem' 1338 * needed. To avoid contention of the lock protecting a the 1339 * single freemem, it was spread out into NCPU buckets. Set_freemem 1340 * sets freemem to the total of all NCPU buckets. It is called from 1341 * clock() on each TICK. 1342 */ 1343 void 1344 set_freemem() 1345 { 1346 struct pcf *p; 1347 ulong_t t; 1348 uint_t i; 1349 1350 t = 0; 1351 p = pcf; 1352 for (i = 0; i < pcf_fanout; i++) { 1353 t += p->pcf_count; 1354 p++; 1355 } 1356 freemem = t; 1357 1358 /* 1359 * Don't worry about grabbing mutex. It's not that 1360 * critical if we miss a tick or two. This is 1361 * where we wakeup possible delayers in 1362 * page_create_get_something(). 1363 */ 1364 wakeup_pcgs(); 1365 } 1366 1367 ulong_t 1368 get_freemem() 1369 { 1370 struct pcf *p; 1371 ulong_t t; 1372 uint_t i; 1373 1374 t = 0; 1375 p = pcf; 1376 for (i = 0; i < pcf_fanout; i++) { 1377 t += p->pcf_count; 1378 p++; 1379 } 1380 /* 1381 * We just calculated it, might as well set it. 1382 */ 1383 freemem = t; 1384 return (t); 1385 } 1386 1387 /* 1388 * Acquire all of the page cache & free (pcf) locks. 1389 */ 1390 void 1391 pcf_acquire_all() 1392 { 1393 struct pcf *p; 1394 uint_t i; 1395 1396 p = pcf; 1397 for (i = 0; i < pcf_fanout; i++) { 1398 mutex_enter(&p->pcf_lock); 1399 p++; 1400 } 1401 } 1402 1403 /* 1404 * Release all the pcf_locks. 1405 */ 1406 void 1407 pcf_release_all() 1408 { 1409 struct pcf *p; 1410 uint_t i; 1411 1412 p = pcf; 1413 for (i = 0; i < pcf_fanout; i++) { 1414 mutex_exit(&p->pcf_lock); 1415 p++; 1416 } 1417 } 1418 1419 /* 1420 * Inform the VM system that we need some pages freed up. 1421 * Calls must be symmetric, e.g.: 1422 * 1423 * page_needfree(100); 1424 * wait a bit; 1425 * page_needfree(-100); 1426 */ 1427 void 1428 page_needfree(spgcnt_t npages) 1429 { 1430 mutex_enter(&new_freemem_lock); 1431 needfree += npages; 1432 mutex_exit(&new_freemem_lock); 1433 } 1434 1435 /* 1436 * Throttle for page_create(): try to prevent freemem from dropping 1437 * below throttlefree. We can't provide a 100% guarantee because 1438 * KM_NOSLEEP allocations, page_reclaim(), and various other things 1439 * nibble away at the freelist. However, we can block all PG_WAIT 1440 * allocations until memory becomes available. The motivation is 1441 * that several things can fall apart when there's no free memory: 1442 * 1443 * (1) If pageout() needs memory to push a page, the system deadlocks. 1444 * 1445 * (2) By (broken) specification, timeout(9F) can neither fail nor 1446 * block, so it has no choice but to panic the system if it 1447 * cannot allocate a callout structure. 1448 * 1449 * (3) Like timeout(), ddi_set_callback() cannot fail and cannot block; 1450 * it panics if it cannot allocate a callback structure. 1451 * 1452 * (4) Untold numbers of third-party drivers have not yet been hardened 1453 * against KM_NOSLEEP and/or allocb() failures; they simply assume 1454 * success and panic the system with a data fault on failure. 1455 * (The long-term solution to this particular problem is to ship 1456 * hostile fault-injecting DEBUG kernels with the DDK.) 1457 * 1458 * It is theoretically impossible to guarantee success of non-blocking 1459 * allocations, but in practice, this throttle is very hard to break. 1460 */ 1461 static int 1462 page_create_throttle(pgcnt_t npages, int flags) 1463 { 1464 ulong_t fm; 1465 uint_t i; 1466 pgcnt_t tf; /* effective value of throttlefree */ 1467 1468 /* 1469 * Never deny pages when: 1470 * - it's a thread that cannot block [NOMEMWAIT()] 1471 * - the allocation cannot block and must not fail 1472 * - the allocation cannot block and is pageout dispensated 1473 */ 1474 if (NOMEMWAIT() || 1475 ((flags & (PG_WAIT | PG_PANIC)) == PG_PANIC) || 1476 ((flags & (PG_WAIT | PG_PUSHPAGE)) == PG_PUSHPAGE)) 1477 return (1); 1478 1479 /* 1480 * If the allocation can't block, we look favorably upon it 1481 * unless we're below pageout_reserve. In that case we fail 1482 * the allocation because we want to make sure there are a few 1483 * pages available for pageout. 1484 */ 1485 if ((flags & PG_WAIT) == 0) 1486 return (freemem >= npages + pageout_reserve); 1487 1488 /* Calculate the effective throttlefree value */ 1489 tf = throttlefree - 1490 ((flags & PG_PUSHPAGE) ? pageout_reserve : 0); 1491 1492 cv_signal(&proc_pageout->p_cv); 1493 1494 for (;;) { 1495 fm = 0; 1496 pcf_acquire_all(); 1497 mutex_enter(&new_freemem_lock); 1498 for (i = 0; i < pcf_fanout; i++) { 1499 fm += pcf[i].pcf_count; 1500 pcf[i].pcf_wait++; 1501 mutex_exit(&pcf[i].pcf_lock); 1502 } 1503 freemem = fm; 1504 if (freemem >= npages + tf) { 1505 mutex_exit(&new_freemem_lock); 1506 break; 1507 } 1508 needfree += npages; 1509 freemem_wait++; 1510 cv_wait(&freemem_cv, &new_freemem_lock); 1511 freemem_wait--; 1512 needfree -= npages; 1513 mutex_exit(&new_freemem_lock); 1514 } 1515 return (1); 1516 } 1517 1518 /* 1519 * page_create_wait() is called to either coalesce pages from the 1520 * different pcf buckets or to wait because there simply are not 1521 * enough pages to satisfy the caller's request. 1522 * 1523 * Sadly, this is called from platform/vm/vm_machdep.c 1524 */ 1525 int 1526 page_create_wait(pgcnt_t npages, uint_t flags) 1527 { 1528 pgcnt_t total; 1529 uint_t i; 1530 struct pcf *p; 1531 1532 /* 1533 * Wait until there are enough free pages to satisfy our 1534 * entire request. 1535 * We set needfree += npages before prodding pageout, to make sure 1536 * it does real work when npages > lotsfree > freemem. 1537 */ 1538 VM_STAT_ADD(page_create_not_enough); 1539 1540 ASSERT(!kcage_on ? !(flags & PG_NORELOC) : 1); 1541 checkagain: 1542 if ((flags & PG_NORELOC) && 1543 kcage_freemem < kcage_throttlefree + npages) 1544 (void) kcage_create_throttle(npages, flags); 1545 1546 if (freemem < npages + throttlefree) 1547 if (!page_create_throttle(npages, flags)) 1548 return (0); 1549 1550 if (pcf_decrement_bucket(npages) || 1551 pcf_decrement_multiple(&total, npages, 0)) 1552 return (1); 1553 1554 /* 1555 * All of the pcf locks are held, there are not enough pages 1556 * to satisfy the request (npages < total). 1557 * Be sure to acquire the new_freemem_lock before dropping 1558 * the pcf locks. This prevents dropping wakeups in page_free(). 1559 * The order is always pcf_lock then new_freemem_lock. 1560 * 1561 * Since we hold all the pcf locks, it is a good time to set freemem. 1562 * 1563 * If the caller does not want to wait, return now. 1564 * Else turn the pageout daemon loose to find something 1565 * and wait till it does. 1566 * 1567 */ 1568 freemem = total; 1569 1570 if ((flags & PG_WAIT) == 0) { 1571 pcf_release_all(); 1572 1573 TRACE_2(TR_FAC_VM, TR_PAGE_CREATE_NOMEM, 1574 "page_create_nomem:npages %ld freemem %ld", npages, freemem); 1575 return (0); 1576 } 1577 1578 ASSERT(proc_pageout != NULL); 1579 cv_signal(&proc_pageout->p_cv); 1580 1581 TRACE_2(TR_FAC_VM, TR_PAGE_CREATE_SLEEP_START, 1582 "page_create_sleep_start: freemem %ld needfree %ld", 1583 freemem, needfree); 1584 1585 /* 1586 * We are going to wait. 1587 * We currently hold all of the pcf_locks, 1588 * get the new_freemem_lock (it protects freemem_wait), 1589 * before dropping the pcf_locks. 1590 */ 1591 mutex_enter(&new_freemem_lock); 1592 1593 p = pcf; 1594 for (i = 0; i < pcf_fanout; i++) { 1595 p->pcf_wait++; 1596 mutex_exit(&p->pcf_lock); 1597 p++; 1598 } 1599 1600 needfree += npages; 1601 freemem_wait++; 1602 1603 cv_wait(&freemem_cv, &new_freemem_lock); 1604 1605 freemem_wait--; 1606 needfree -= npages; 1607 1608 mutex_exit(&new_freemem_lock); 1609 1610 TRACE_2(TR_FAC_VM, TR_PAGE_CREATE_SLEEP_END, 1611 "page_create_sleep_end: freemem %ld needfree %ld", 1612 freemem, needfree); 1613 1614 VM_STAT_ADD(page_create_not_enough_again); 1615 goto checkagain; 1616 } 1617 /* 1618 * A routine to do the opposite of page_create_wait(). 1619 */ 1620 void 1621 page_create_putback(spgcnt_t npages) 1622 { 1623 struct pcf *p; 1624 pgcnt_t lump; 1625 uint_t *which; 1626 1627 /* 1628 * When a contiguous lump is broken up, we have to 1629 * deal with lots of pages (min 64) so lets spread 1630 * the wealth around. 1631 */ 1632 lump = roundup(npages, pcf_fanout) / pcf_fanout; 1633 freemem += npages; 1634 1635 for (p = pcf; (npages > 0) && (p < &pcf[pcf_fanout]); p++) { 1636 which = &p->pcf_count; 1637 1638 mutex_enter(&p->pcf_lock); 1639 1640 if (p->pcf_block) { 1641 which = &p->pcf_reserve; 1642 } 1643 1644 if (lump < npages) { 1645 *which += (uint_t)lump; 1646 npages -= lump; 1647 } else { 1648 *which += (uint_t)npages; 1649 npages = 0; 1650 } 1651 1652 if (p->pcf_wait) { 1653 mutex_enter(&new_freemem_lock); 1654 /* 1655 * Check to see if some other thread 1656 * is actually waiting. Another bucket 1657 * may have woken it up by now. If there 1658 * are no waiters, then set our pcf_wait 1659 * count to zero to avoid coming in here 1660 * next time. 1661 */ 1662 if (freemem_wait) { 1663 if (npages > 1) { 1664 cv_broadcast(&freemem_cv); 1665 } else { 1666 cv_signal(&freemem_cv); 1667 } 1668 p->pcf_wait--; 1669 } else { 1670 p->pcf_wait = 0; 1671 } 1672 mutex_exit(&new_freemem_lock); 1673 } 1674 mutex_exit(&p->pcf_lock); 1675 } 1676 ASSERT(npages == 0); 1677 } 1678 1679 /* 1680 * A helper routine for page_create_get_something. 1681 * The indenting got to deep down there. 1682 * Unblock the pcf counters. Any pages freed after 1683 * pcf_block got set are moved to pcf_count and 1684 * wakeups (cv_broadcast() or cv_signal()) are done as needed. 1685 */ 1686 static void 1687 pcgs_unblock(void) 1688 { 1689 int i; 1690 struct pcf *p; 1691 1692 /* Update freemem while we're here. */ 1693 freemem = 0; 1694 p = pcf; 1695 for (i = 0; i < pcf_fanout; i++) { 1696 mutex_enter(&p->pcf_lock); 1697 ASSERT(p->pcf_count == 0); 1698 p->pcf_count = p->pcf_reserve; 1699 p->pcf_block = 0; 1700 freemem += p->pcf_count; 1701 if (p->pcf_wait) { 1702 mutex_enter(&new_freemem_lock); 1703 if (freemem_wait) { 1704 if (p->pcf_reserve > 1) { 1705 cv_broadcast(&freemem_cv); 1706 p->pcf_wait = 0; 1707 } else { 1708 cv_signal(&freemem_cv); 1709 p->pcf_wait--; 1710 } 1711 } else { 1712 p->pcf_wait = 0; 1713 } 1714 mutex_exit(&new_freemem_lock); 1715 } 1716 p->pcf_reserve = 0; 1717 mutex_exit(&p->pcf_lock); 1718 p++; 1719 } 1720 } 1721 1722 /* 1723 * Called from page_create_va() when both the cache and free lists 1724 * have been checked once. 1725 * 1726 * Either returns a page or panics since the accounting was done 1727 * way before we got here. 1728 * 1729 * We don't come here often, so leave the accounting on permanently. 1730 */ 1731 1732 #define MAX_PCGS 100 1733 1734 #ifdef DEBUG 1735 #define PCGS_TRIES 100 1736 #else /* DEBUG */ 1737 #define PCGS_TRIES 10 1738 #endif /* DEBUG */ 1739 1740 #ifdef VM_STATS 1741 uint_t pcgs_counts[PCGS_TRIES]; 1742 uint_t pcgs_too_many; 1743 uint_t pcgs_entered; 1744 uint_t pcgs_entered_noreloc; 1745 uint_t pcgs_locked; 1746 uint_t pcgs_cagelocked; 1747 #endif /* VM_STATS */ 1748 1749 static page_t * 1750 page_create_get_something(vnode_t *vp, u_offset_t off, struct seg *seg, 1751 caddr_t vaddr, uint_t flags) 1752 { 1753 uint_t count; 1754 page_t *pp; 1755 uint_t locked, i; 1756 struct pcf *p; 1757 lgrp_t *lgrp; 1758 int cagelocked = 0; 1759 1760 VM_STAT_ADD(pcgs_entered); 1761 1762 /* 1763 * Tap any reserve freelists: if we fail now, we'll die 1764 * since the page(s) we're looking for have already been 1765 * accounted for. 1766 */ 1767 flags |= PG_PANIC; 1768 1769 if ((flags & PG_NORELOC) != 0) { 1770 VM_STAT_ADD(pcgs_entered_noreloc); 1771 /* 1772 * Requests for free pages from critical threads 1773 * such as pageout still won't throttle here, but 1774 * we must try again, to give the cageout thread 1775 * another chance to catch up. Since we already 1776 * accounted for the pages, we had better get them 1777 * this time. 1778 * 1779 * N.B. All non-critical threads acquire the pcgs_cagelock 1780 * to serialize access to the freelists. This implements a 1781 * turnstile-type synchornization to avoid starvation of 1782 * critical requests for PG_NORELOC memory by non-critical 1783 * threads: all non-critical threads must acquire a 'ticket' 1784 * before passing through, which entails making sure 1785 * kcage_freemem won't fall below minfree prior to grabbing 1786 * pages from the freelists. 1787 */ 1788 if (kcage_create_throttle(1, flags) == KCT_NONCRIT) { 1789 mutex_enter(&pcgs_cagelock); 1790 cagelocked = 1; 1791 VM_STAT_ADD(pcgs_cagelocked); 1792 } 1793 } 1794 1795 /* 1796 * Time to get serious. 1797 * We failed to get a `correctly colored' page from both the 1798 * free and cache lists. 1799 * We escalate in stage. 1800 * 1801 * First try both lists without worring about color. 1802 * 1803 * Then, grab all page accounting locks (ie. pcf[]) and 1804 * steal any pages that they have and set the pcf_block flag to 1805 * stop deletions from the lists. This will help because 1806 * a page can get added to the free list while we are looking 1807 * at the cache list, then another page could be added to the cache 1808 * list allowing the page on the free list to be removed as we 1809 * move from looking at the cache list to the free list. This 1810 * could happen over and over. We would never find the page 1811 * we have accounted for. 1812 * 1813 * Noreloc pages are a subset of the global (relocatable) page pool. 1814 * They are not tracked separately in the pcf bins, so it is 1815 * impossible to know when doing pcf accounting if the available 1816 * page(s) are noreloc pages or not. When looking for a noreloc page 1817 * it is quite easy to end up here even if the global (relocatable) 1818 * page pool has plenty of free pages but the noreloc pool is empty. 1819 * 1820 * When the noreloc pool is empty (or low), additional noreloc pages 1821 * are created by converting pages from the global page pool. This 1822 * process will stall during pcf accounting if the pcf bins are 1823 * already locked. Such is the case when a noreloc allocation is 1824 * looping here in page_create_get_something waiting for more noreloc 1825 * pages to appear. 1826 * 1827 * Short of adding a new field to the pcf bins to accurately track 1828 * the number of free noreloc pages, we instead do not grab the 1829 * pcgs_lock, do not set the pcf blocks and do not timeout when 1830 * allocating a noreloc page. This allows noreloc allocations to 1831 * loop without blocking global page pool allocations. 1832 * 1833 * NOTE: the behaviour of page_create_get_something has not changed 1834 * for the case of global page pool allocations. 1835 */ 1836 1837 flags &= ~PG_MATCH_COLOR; 1838 locked = 0; 1839 #if defined(__i386) || defined(__amd64) 1840 flags = page_create_update_flags_x86(flags); 1841 #endif 1842 1843 lgrp = lgrp_mem_choose(seg, vaddr, PAGESIZE); 1844 1845 for (count = 0; kcage_on || count < MAX_PCGS; count++) { 1846 pp = page_get_freelist(vp, off, seg, vaddr, PAGESIZE, 1847 flags, lgrp); 1848 if (pp == NULL) { 1849 pp = page_get_cachelist(vp, off, seg, vaddr, 1850 flags, lgrp); 1851 } 1852 if (pp == NULL) { 1853 /* 1854 * Serialize. Don't fight with other pcgs(). 1855 */ 1856 if (!locked && (!kcage_on || !(flags & PG_NORELOC))) { 1857 mutex_enter(&pcgs_lock); 1858 VM_STAT_ADD(pcgs_locked); 1859 locked = 1; 1860 p = pcf; 1861 for (i = 0; i < pcf_fanout; i++) { 1862 mutex_enter(&p->pcf_lock); 1863 ASSERT(p->pcf_block == 0); 1864 p->pcf_block = 1; 1865 p->pcf_reserve = p->pcf_count; 1866 p->pcf_count = 0; 1867 mutex_exit(&p->pcf_lock); 1868 p++; 1869 } 1870 freemem = 0; 1871 } 1872 1873 if (count) { 1874 /* 1875 * Since page_free() puts pages on 1876 * a list then accounts for it, we 1877 * just have to wait for page_free() 1878 * to unlock any page it was working 1879 * with. The page_lock()-page_reclaim() 1880 * path falls in the same boat. 1881 * 1882 * We don't need to check on the 1883 * PG_WAIT flag, we have already 1884 * accounted for the page we are 1885 * looking for in page_create_va(). 1886 * 1887 * We just wait a moment to let any 1888 * locked pages on the lists free up, 1889 * then continue around and try again. 1890 * 1891 * Will be awakened by set_freemem(). 1892 */ 1893 mutex_enter(&pcgs_wait_lock); 1894 cv_wait(&pcgs_cv, &pcgs_wait_lock); 1895 mutex_exit(&pcgs_wait_lock); 1896 } 1897 } else { 1898 #ifdef VM_STATS 1899 if (count >= PCGS_TRIES) { 1900 VM_STAT_ADD(pcgs_too_many); 1901 } else { 1902 VM_STAT_ADD(pcgs_counts[count]); 1903 } 1904 #endif 1905 if (locked) { 1906 pcgs_unblock(); 1907 mutex_exit(&pcgs_lock); 1908 } 1909 if (cagelocked) 1910 mutex_exit(&pcgs_cagelock); 1911 return (pp); 1912 } 1913 } 1914 /* 1915 * we go down holding the pcf locks. 1916 */ 1917 panic("no %spage found %d", 1918 ((flags & PG_NORELOC) ? "non-reloc " : ""), count); 1919 /*NOTREACHED*/ 1920 } 1921 1922 /* 1923 * Create enough pages for "bytes" worth of data starting at 1924 * "off" in "vp". 1925 * 1926 * Where flag must be one of: 1927 * 1928 * PG_EXCL: Exclusive create (fail if any page already 1929 * exists in the page cache) which does not 1930 * wait for memory to become available. 1931 * 1932 * PG_WAIT: Non-exclusive create which can wait for 1933 * memory to become available. 1934 * 1935 * PG_PHYSCONTIG: Allocate physically contiguous pages. 1936 * (Not Supported) 1937 * 1938 * A doubly linked list of pages is returned to the caller. Each page 1939 * on the list has the "exclusive" (p_selock) lock and "iolock" (p_iolock) 1940 * lock. 1941 * 1942 * Unable to change the parameters to page_create() in a minor release, 1943 * we renamed page_create() to page_create_va(), changed all known calls 1944 * from page_create() to page_create_va(), and created this wrapper. 1945 * 1946 * Upon a major release, we should break compatibility by deleting this 1947 * wrapper, and replacing all the strings "page_create_va", with "page_create". 1948 * 1949 * NOTE: There is a copy of this interface as page_create_io() in 1950 * i86/vm/vm_machdep.c. Any bugs fixed here should be applied 1951 * there. 1952 */ 1953 page_t * 1954 page_create(vnode_t *vp, u_offset_t off, size_t bytes, uint_t flags) 1955 { 1956 caddr_t random_vaddr; 1957 struct seg kseg; 1958 1959 #ifdef DEBUG 1960 cmn_err(CE_WARN, "Using deprecated interface page_create: caller %p", 1961 (void *)caller()); 1962 #endif 1963 1964 random_vaddr = (caddr_t)(((uintptr_t)vp >> 7) ^ 1965 (uintptr_t)(off >> PAGESHIFT)); 1966 kseg.s_as = &kas; 1967 1968 return (page_create_va(vp, off, bytes, flags, &kseg, random_vaddr)); 1969 } 1970 1971 #ifdef DEBUG 1972 uint32_t pg_alloc_pgs_mtbf = 0; 1973 #endif 1974 1975 /* 1976 * Used for large page support. It will attempt to allocate 1977 * a large page(s) off the freelist. 1978 * 1979 * Returns non zero on failure. 1980 */ 1981 int 1982 page_alloc_pages(struct vnode *vp, struct seg *seg, caddr_t addr, 1983 page_t **basepp, page_t *ppa[], uint_t szc, int anypgsz, int pgflags) 1984 { 1985 pgcnt_t npgs, curnpgs, totpgs; 1986 size_t pgsz; 1987 page_t *pplist = NULL, *pp; 1988 int err = 0; 1989 lgrp_t *lgrp; 1990 1991 ASSERT(szc != 0 && szc <= (page_num_pagesizes() - 1)); 1992 ASSERT(pgflags == 0 || pgflags == PG_LOCAL); 1993 1994 /* 1995 * Check if system heavily prefers local large pages over remote 1996 * on systems with multiple lgroups. 1997 */ 1998 if (lpg_alloc_prefer == LPAP_LOCAL && nlgrps > 1) { 1999 pgflags = PG_LOCAL; 2000 } 2001 2002 VM_STAT_ADD(alloc_pages[0]); 2003 2004 #ifdef DEBUG 2005 if (pg_alloc_pgs_mtbf && !(gethrtime() % pg_alloc_pgs_mtbf)) { 2006 return (ENOMEM); 2007 } 2008 #endif 2009 2010 /* 2011 * One must be NULL but not both. 2012 * And one must be non NULL but not both. 2013 */ 2014 ASSERT(basepp != NULL || ppa != NULL); 2015 ASSERT(basepp == NULL || ppa == NULL); 2016 2017 #if defined(__i386) || defined(__amd64) 2018 while (page_chk_freelist(szc) == 0) { 2019 VM_STAT_ADD(alloc_pages[8]); 2020 if (anypgsz == 0 || --szc == 0) 2021 return (ENOMEM); 2022 } 2023 #endif 2024 2025 pgsz = page_get_pagesize(szc); 2026 totpgs = curnpgs = npgs = pgsz >> PAGESHIFT; 2027 2028 ASSERT(((uintptr_t)addr & (pgsz - 1)) == 0); 2029 2030 (void) page_create_wait(npgs, PG_WAIT); 2031 2032 while (npgs && szc) { 2033 lgrp = lgrp_mem_choose(seg, addr, pgsz); 2034 if (pgflags == PG_LOCAL) { 2035 pp = page_get_freelist(vp, 0, seg, addr, pgsz, 2036 pgflags, lgrp); 2037 if (pp == NULL) { 2038 pp = page_get_freelist(vp, 0, seg, addr, pgsz, 2039 0, lgrp); 2040 } 2041 } else { 2042 pp = page_get_freelist(vp, 0, seg, addr, pgsz, 2043 0, lgrp); 2044 } 2045 if (pp != NULL) { 2046 VM_STAT_ADD(alloc_pages[1]); 2047 page_list_concat(&pplist, &pp); 2048 ASSERT(npgs >= curnpgs); 2049 npgs -= curnpgs; 2050 } else if (anypgsz) { 2051 VM_STAT_ADD(alloc_pages[2]); 2052 szc--; 2053 pgsz = page_get_pagesize(szc); 2054 curnpgs = pgsz >> PAGESHIFT; 2055 } else { 2056 VM_STAT_ADD(alloc_pages[3]); 2057 ASSERT(npgs == totpgs); 2058 page_create_putback(npgs); 2059 return (ENOMEM); 2060 } 2061 } 2062 if (szc == 0) { 2063 VM_STAT_ADD(alloc_pages[4]); 2064 ASSERT(npgs != 0); 2065 page_create_putback(npgs); 2066 err = ENOMEM; 2067 } else if (basepp != NULL) { 2068 ASSERT(npgs == 0); 2069 ASSERT(ppa == NULL); 2070 *basepp = pplist; 2071 } 2072 2073 npgs = totpgs - npgs; 2074 pp = pplist; 2075 2076 /* 2077 * Clear the free and age bits. Also if we were passed in a ppa then 2078 * fill it in with all the constituent pages from the large page. But 2079 * if we failed to allocate all the pages just free what we got. 2080 */ 2081 while (npgs != 0) { 2082 ASSERT(PP_ISFREE(pp)); 2083 ASSERT(PP_ISAGED(pp)); 2084 if (ppa != NULL || err != 0) { 2085 if (err == 0) { 2086 VM_STAT_ADD(alloc_pages[5]); 2087 PP_CLRFREE(pp); 2088 PP_CLRAGED(pp); 2089 page_sub(&pplist, pp); 2090 *ppa++ = pp; 2091 npgs--; 2092 } else { 2093 VM_STAT_ADD(alloc_pages[6]); 2094 ASSERT(pp->p_szc != 0); 2095 curnpgs = page_get_pagecnt(pp->p_szc); 2096 page_list_break(&pp, &pplist, curnpgs); 2097 page_list_add_pages(pp, 0); 2098 page_create_putback(curnpgs); 2099 ASSERT(npgs >= curnpgs); 2100 npgs -= curnpgs; 2101 } 2102 pp = pplist; 2103 } else { 2104 VM_STAT_ADD(alloc_pages[7]); 2105 PP_CLRFREE(pp); 2106 PP_CLRAGED(pp); 2107 pp = pp->p_next; 2108 npgs--; 2109 } 2110 } 2111 return (err); 2112 } 2113 2114 /* 2115 * Get a single large page off of the freelists, and set it up for use. 2116 * Number of bytes requested must be a supported page size. 2117 * 2118 * Note that this call may fail even if there is sufficient 2119 * memory available or PG_WAIT is set, so the caller must 2120 * be willing to fallback on page_create_va(), block and retry, 2121 * or fail the requester. 2122 */ 2123 page_t * 2124 page_create_va_large(vnode_t *vp, u_offset_t off, size_t bytes, uint_t flags, 2125 struct seg *seg, caddr_t vaddr, void *arg) 2126 { 2127 pgcnt_t npages; 2128 page_t *pp; 2129 page_t *rootpp; 2130 lgrp_t *lgrp; 2131 lgrp_id_t *lgrpid = (lgrp_id_t *)arg; 2132 2133 ASSERT(vp != NULL); 2134 2135 ASSERT((flags & ~(PG_EXCL | PG_WAIT | 2136 PG_NORELOC | PG_PANIC | PG_PUSHPAGE)) == 0); 2137 /* but no others */ 2138 2139 ASSERT((flags & PG_EXCL) == PG_EXCL); 2140 2141 npages = btop(bytes); 2142 2143 if (!kcage_on || panicstr) { 2144 /* 2145 * Cage is OFF, or we are single threaded in 2146 * panic, so make everything a RELOC request. 2147 */ 2148 flags &= ~PG_NORELOC; 2149 } 2150 2151 /* 2152 * Make sure there's adequate physical memory available. 2153 * Note: PG_WAIT is ignored here. 2154 */ 2155 if (freemem <= throttlefree + npages) { 2156 VM_STAT_ADD(page_create_large_cnt[1]); 2157 return (NULL); 2158 } 2159 2160 /* 2161 * If cage is on, dampen draw from cage when available 2162 * cage space is low. 2163 */ 2164 if ((flags & (PG_NORELOC | PG_WAIT)) == (PG_NORELOC | PG_WAIT) && 2165 kcage_freemem < kcage_throttlefree + npages) { 2166 2167 /* 2168 * The cage is on, the caller wants PG_NORELOC 2169 * pages and available cage memory is very low. 2170 * Call kcage_create_throttle() to attempt to 2171 * control demand on the cage. 2172 */ 2173 if (kcage_create_throttle(npages, flags) == KCT_FAILURE) { 2174 VM_STAT_ADD(page_create_large_cnt[2]); 2175 return (NULL); 2176 } 2177 } 2178 2179 if (!pcf_decrement_bucket(npages) && 2180 !pcf_decrement_multiple(NULL, npages, 1)) { 2181 VM_STAT_ADD(page_create_large_cnt[4]); 2182 return (NULL); 2183 } 2184 2185 /* 2186 * This is where this function behaves fundamentally differently 2187 * than page_create_va(); since we're intending to map the page 2188 * with a single TTE, we have to get it as a physically contiguous 2189 * hardware pagesize chunk. If we can't, we fail. 2190 */ 2191 if (lgrpid != NULL && *lgrpid >= 0 && *lgrpid <= lgrp_alloc_max && 2192 LGRP_EXISTS(lgrp_table[*lgrpid])) 2193 lgrp = lgrp_table[*lgrpid]; 2194 else 2195 lgrp = lgrp_mem_choose(seg, vaddr, bytes); 2196 2197 if ((rootpp = page_get_freelist(&kvp, off, seg, vaddr, 2198 bytes, flags & ~PG_MATCH_COLOR, lgrp)) == NULL) { 2199 page_create_putback(npages); 2200 VM_STAT_ADD(page_create_large_cnt[5]); 2201 return (NULL); 2202 } 2203 2204 /* 2205 * if we got the page with the wrong mtype give it back this is a 2206 * workaround for CR 6249718. When CR 6249718 is fixed we never get 2207 * inside "if" and the workaround becomes just a nop 2208 */ 2209 if (kcage_on && (flags & PG_NORELOC) && !PP_ISNORELOC(rootpp)) { 2210 page_list_add_pages(rootpp, 0); 2211 page_create_putback(npages); 2212 VM_STAT_ADD(page_create_large_cnt[6]); 2213 return (NULL); 2214 } 2215 2216 /* 2217 * If satisfying this request has left us with too little 2218 * memory, start the wheels turning to get some back. The 2219 * first clause of the test prevents waking up the pageout 2220 * daemon in situations where it would decide that there's 2221 * nothing to do. 2222 */ 2223 if (nscan < desscan && freemem < minfree) { 2224 TRACE_1(TR_FAC_VM, TR_PAGEOUT_CV_SIGNAL, 2225 "pageout_cv_signal:freemem %ld", freemem); 2226 cv_signal(&proc_pageout->p_cv); 2227 } 2228 2229 pp = rootpp; 2230 while (npages--) { 2231 ASSERT(PAGE_EXCL(pp)); 2232 ASSERT(pp->p_vnode == NULL); 2233 ASSERT(!hat_page_is_mapped(pp)); 2234 PP_CLRFREE(pp); 2235 PP_CLRAGED(pp); 2236 if (!page_hashin(pp, vp, off, NULL)) 2237 panic("page_create_large: hashin failed: page %p", 2238 (void *)pp); 2239 page_io_lock(pp); 2240 off += PAGESIZE; 2241 pp = pp->p_next; 2242 } 2243 2244 VM_STAT_ADD(page_create_large_cnt[0]); 2245 return (rootpp); 2246 } 2247 2248 page_t * 2249 page_create_va(vnode_t *vp, u_offset_t off, size_t bytes, uint_t flags, 2250 struct seg *seg, caddr_t vaddr) 2251 { 2252 page_t *plist = NULL; 2253 pgcnt_t npages; 2254 pgcnt_t found_on_free = 0; 2255 pgcnt_t pages_req; 2256 page_t *npp = NULL; 2257 struct pcf *p; 2258 lgrp_t *lgrp; 2259 2260 TRACE_4(TR_FAC_VM, TR_PAGE_CREATE_START, 2261 "page_create_start:vp %p off %llx bytes %lu flags %x", 2262 vp, off, bytes, flags); 2263 2264 ASSERT(bytes != 0 && vp != NULL); 2265 2266 if ((flags & PG_EXCL) == 0 && (flags & PG_WAIT) == 0) { 2267 panic("page_create: invalid flags"); 2268 /*NOTREACHED*/ 2269 } 2270 ASSERT((flags & ~(PG_EXCL | PG_WAIT | 2271 PG_NORELOC | PG_PANIC | PG_PUSHPAGE)) == 0); 2272 /* but no others */ 2273 2274 pages_req = npages = btopr(bytes); 2275 /* 2276 * Try to see whether request is too large to *ever* be 2277 * satisfied, in order to prevent deadlock. We arbitrarily 2278 * decide to limit maximum size requests to max_page_get. 2279 */ 2280 if (npages >= max_page_get) { 2281 if ((flags & PG_WAIT) == 0) { 2282 TRACE_4(TR_FAC_VM, TR_PAGE_CREATE_TOOBIG, 2283 "page_create_toobig:vp %p off %llx npages " 2284 "%lu max_page_get %lu", 2285 vp, off, npages, max_page_get); 2286 return (NULL); 2287 } else { 2288 cmn_err(CE_WARN, 2289 "Request for too much kernel memory " 2290 "(%lu bytes), will hang forever", bytes); 2291 for (;;) 2292 delay(1000000000); 2293 } 2294 } 2295 2296 if (!kcage_on || panicstr) { 2297 /* 2298 * Cage is OFF, or we are single threaded in 2299 * panic, so make everything a RELOC request. 2300 */ 2301 flags &= ~PG_NORELOC; 2302 } 2303 2304 if (freemem <= throttlefree + npages) 2305 if (!page_create_throttle(npages, flags)) 2306 return (NULL); 2307 2308 /* 2309 * If cage is on, dampen draw from cage when available 2310 * cage space is low. 2311 */ 2312 if ((flags & PG_NORELOC) && 2313 kcage_freemem < kcage_throttlefree + npages) { 2314 2315 /* 2316 * The cage is on, the caller wants PG_NORELOC 2317 * pages and available cage memory is very low. 2318 * Call kcage_create_throttle() to attempt to 2319 * control demand on the cage. 2320 */ 2321 if (kcage_create_throttle(npages, flags) == KCT_FAILURE) 2322 return (NULL); 2323 } 2324 2325 VM_STAT_ADD(page_create_cnt[0]); 2326 2327 if (!pcf_decrement_bucket(npages)) { 2328 /* 2329 * Have to look harder. If npages is greater than 2330 * one, then we might have to coalesce the counters. 2331 * 2332 * Go wait. We come back having accounted 2333 * for the memory. 2334 */ 2335 VM_STAT_ADD(page_create_cnt[1]); 2336 if (!page_create_wait(npages, flags)) { 2337 VM_STAT_ADD(page_create_cnt[2]); 2338 return (NULL); 2339 } 2340 } 2341 2342 TRACE_2(TR_FAC_VM, TR_PAGE_CREATE_SUCCESS, 2343 "page_create_success:vp %p off %llx", vp, off); 2344 2345 /* 2346 * If satisfying this request has left us with too little 2347 * memory, start the wheels turning to get some back. The 2348 * first clause of the test prevents waking up the pageout 2349 * daemon in situations where it would decide that there's 2350 * nothing to do. 2351 */ 2352 if (nscan < desscan && freemem < minfree) { 2353 TRACE_1(TR_FAC_VM, TR_PAGEOUT_CV_SIGNAL, 2354 "pageout_cv_signal:freemem %ld", freemem); 2355 cv_signal(&proc_pageout->p_cv); 2356 } 2357 2358 /* 2359 * Loop around collecting the requested number of pages. 2360 * Most of the time, we have to `create' a new page. With 2361 * this in mind, pull the page off the free list before 2362 * getting the hash lock. This will minimize the hash 2363 * lock hold time, nesting, and the like. If it turns 2364 * out we don't need the page, we put it back at the end. 2365 */ 2366 while (npages--) { 2367 page_t *pp; 2368 kmutex_t *phm = NULL; 2369 ulong_t index; 2370 2371 index = PAGE_HASH_FUNC(vp, off); 2372 top: 2373 ASSERT(phm == NULL); 2374 ASSERT(index == PAGE_HASH_FUNC(vp, off)); 2375 ASSERT(MUTEX_NOT_HELD(page_vnode_mutex(vp))); 2376 2377 if (npp == NULL) { 2378 /* 2379 * Try to get a page from the freelist (ie, 2380 * a page with no [vp, off] tag). If that 2381 * fails, use the cachelist. 2382 * 2383 * During the first attempt at both the free 2384 * and cache lists we try for the correct color. 2385 */ 2386 /* 2387 * XXXX-how do we deal with virtual indexed 2388 * caches and and colors? 2389 */ 2390 VM_STAT_ADD(page_create_cnt[4]); 2391 /* 2392 * Get lgroup to allocate next page of shared memory 2393 * from and use it to specify where to allocate 2394 * the physical memory 2395 */ 2396 lgrp = lgrp_mem_choose(seg, vaddr, PAGESIZE); 2397 npp = page_get_freelist(vp, off, seg, vaddr, PAGESIZE, 2398 flags | PG_MATCH_COLOR, lgrp); 2399 if (npp == NULL) { 2400 npp = page_get_cachelist(vp, off, seg, 2401 vaddr, flags | PG_MATCH_COLOR, lgrp); 2402 if (npp == NULL) { 2403 npp = page_create_get_something(vp, 2404 off, seg, vaddr, 2405 flags & ~PG_MATCH_COLOR); 2406 } 2407 2408 if (PP_ISAGED(npp) == 0) { 2409 /* 2410 * Since this page came from the 2411 * cachelist, we must destroy the 2412 * old vnode association. 2413 */ 2414 page_hashout(npp, NULL); 2415 } 2416 } 2417 } 2418 2419 /* 2420 * We own this page! 2421 */ 2422 ASSERT(PAGE_EXCL(npp)); 2423 ASSERT(npp->p_vnode == NULL); 2424 ASSERT(!hat_page_is_mapped(npp)); 2425 PP_CLRFREE(npp); 2426 PP_CLRAGED(npp); 2427 2428 /* 2429 * Here we have a page in our hot little mits and are 2430 * just waiting to stuff it on the appropriate lists. 2431 * Get the mutex and check to see if it really does 2432 * not exist. 2433 */ 2434 phm = PAGE_HASH_MUTEX(index); 2435 mutex_enter(phm); 2436 PAGE_HASH_SEARCH(index, pp, vp, off); 2437 if (pp == NULL) { 2438 VM_STAT_ADD(page_create_new); 2439 pp = npp; 2440 npp = NULL; 2441 if (!page_hashin(pp, vp, off, phm)) { 2442 /* 2443 * Since we hold the page hash mutex and 2444 * just searched for this page, page_hashin 2445 * had better not fail. If it does, that 2446 * means somethread did not follow the 2447 * page hash mutex rules. Panic now and 2448 * get it over with. As usual, go down 2449 * holding all the locks. 2450 */ 2451 ASSERT(MUTEX_HELD(phm)); 2452 panic("page_create: " 2453 "hashin failed %p %p %llx %p", 2454 (void *)pp, (void *)vp, off, (void *)phm); 2455 /*NOTREACHED*/ 2456 } 2457 ASSERT(MUTEX_HELD(phm)); 2458 mutex_exit(phm); 2459 phm = NULL; 2460 2461 /* 2462 * Hat layer locking need not be done to set 2463 * the following bits since the page is not hashed 2464 * and was on the free list (i.e., had no mappings). 2465 * 2466 * Set the reference bit to protect 2467 * against immediate pageout 2468 * 2469 * XXXmh modify freelist code to set reference 2470 * bit so we don't have to do it here. 2471 */ 2472 page_set_props(pp, P_REF); 2473 found_on_free++; 2474 } else { 2475 VM_STAT_ADD(page_create_exists); 2476 if (flags & PG_EXCL) { 2477 /* 2478 * Found an existing page, and the caller 2479 * wanted all new pages. Undo all of the work 2480 * we have done. 2481 */ 2482 mutex_exit(phm); 2483 phm = NULL; 2484 while (plist != NULL) { 2485 pp = plist; 2486 page_sub(&plist, pp); 2487 page_io_unlock(pp); 2488 /* large pages should not end up here */ 2489 ASSERT(pp->p_szc == 0); 2490 /*LINTED: constant in conditional ctx*/ 2491 VN_DISPOSE(pp, B_INVAL, 0, kcred); 2492 } 2493 VM_STAT_ADD(page_create_found_one); 2494 goto fail; 2495 } 2496 ASSERT(flags & PG_WAIT); 2497 if (!page_lock(pp, SE_EXCL, phm, P_NO_RECLAIM)) { 2498 /* 2499 * Start all over again if we blocked trying 2500 * to lock the page. 2501 */ 2502 mutex_exit(phm); 2503 VM_STAT_ADD(page_create_page_lock_failed); 2504 phm = NULL; 2505 goto top; 2506 } 2507 mutex_exit(phm); 2508 phm = NULL; 2509 2510 if (PP_ISFREE(pp)) { 2511 ASSERT(PP_ISAGED(pp) == 0); 2512 VM_STAT_ADD(pagecnt.pc_get_cache); 2513 page_list_sub(pp, PG_CACHE_LIST); 2514 PP_CLRFREE(pp); 2515 found_on_free++; 2516 } 2517 } 2518 2519 /* 2520 * Got a page! It is locked. Acquire the i/o 2521 * lock since we are going to use the p_next and 2522 * p_prev fields to link the requested pages together. 2523 */ 2524 page_io_lock(pp); 2525 page_add(&plist, pp); 2526 plist = plist->p_next; 2527 off += PAGESIZE; 2528 vaddr += PAGESIZE; 2529 } 2530 2531 ASSERT((flags & PG_EXCL) ? (found_on_free == pages_req) : 1); 2532 fail: 2533 if (npp != NULL) { 2534 /* 2535 * Did not need this page after all. 2536 * Put it back on the free list. 2537 */ 2538 VM_STAT_ADD(page_create_putbacks); 2539 PP_SETFREE(npp); 2540 PP_SETAGED(npp); 2541 npp->p_offset = (u_offset_t)-1; 2542 page_list_add(npp, PG_FREE_LIST | PG_LIST_TAIL); 2543 page_unlock(npp); 2544 2545 } 2546 2547 ASSERT(pages_req >= found_on_free); 2548 2549 { 2550 uint_t overshoot = (uint_t)(pages_req - found_on_free); 2551 2552 if (overshoot) { 2553 VM_STAT_ADD(page_create_overshoot); 2554 p = &pcf[PCF_INDEX()]; 2555 mutex_enter(&p->pcf_lock); 2556 if (p->pcf_block) { 2557 p->pcf_reserve += overshoot; 2558 } else { 2559 p->pcf_count += overshoot; 2560 if (p->pcf_wait) { 2561 mutex_enter(&new_freemem_lock); 2562 if (freemem_wait) { 2563 cv_signal(&freemem_cv); 2564 p->pcf_wait--; 2565 } else { 2566 p->pcf_wait = 0; 2567 } 2568 mutex_exit(&new_freemem_lock); 2569 } 2570 } 2571 mutex_exit(&p->pcf_lock); 2572 /* freemem is approximate, so this test OK */ 2573 if (!p->pcf_block) 2574 freemem += overshoot; 2575 } 2576 } 2577 2578 return (plist); 2579 } 2580 2581 /* 2582 * One or more constituent pages of this large page has been marked 2583 * toxic. Simply demote the large page to PAGESIZE pages and let 2584 * page_free() handle it. This routine should only be called by 2585 * large page free routines (page_free_pages() and page_destroy_pages(). 2586 * All pages are locked SE_EXCL and have already been marked free. 2587 */ 2588 static void 2589 page_free_toxic_pages(page_t *rootpp) 2590 { 2591 page_t *tpp; 2592 pgcnt_t i, pgcnt = page_get_pagecnt(rootpp->p_szc); 2593 uint_t szc = rootpp->p_szc; 2594 2595 for (i = 0, tpp = rootpp; i < pgcnt; i++, tpp = tpp->p_next) { 2596 ASSERT(tpp->p_szc == szc); 2597 ASSERT((PAGE_EXCL(tpp) && 2598 !page_iolock_assert(tpp)) || panicstr); 2599 tpp->p_szc = 0; 2600 } 2601 2602 while (rootpp != NULL) { 2603 tpp = rootpp; 2604 page_sub(&rootpp, tpp); 2605 ASSERT(PP_ISFREE(tpp)); 2606 PP_CLRFREE(tpp); 2607 page_free(tpp, 1); 2608 } 2609 } 2610 2611 /* 2612 * Put page on the "free" list. 2613 * The free list is really two lists maintained by 2614 * the PSM of whatever machine we happen to be on. 2615 */ 2616 void 2617 page_free(page_t *pp, int dontneed) 2618 { 2619 struct pcf *p; 2620 uint_t pcf_index; 2621 2622 ASSERT((PAGE_EXCL(pp) && 2623 !page_iolock_assert(pp)) || panicstr); 2624 2625 if (PP_ISFREE(pp)) { 2626 panic("page_free: page %p is free", (void *)pp); 2627 } 2628 2629 if (pp->p_szc != 0) { 2630 if (pp->p_vnode == NULL || IS_SWAPFSVP(pp->p_vnode) || 2631 PP_ISKAS(pp)) { 2632 panic("page_free: anon or kernel " 2633 "or no vnode large page %p", (void *)pp); 2634 } 2635 page_demote_vp_pages(pp); 2636 ASSERT(pp->p_szc == 0); 2637 } 2638 2639 /* 2640 * The page_struct_lock need not be acquired to examine these 2641 * fields since the page has an "exclusive" lock. 2642 */ 2643 if (hat_page_is_mapped(pp) || pp->p_lckcnt != 0 || pp->p_cowcnt != 0 || 2644 pp->p_slckcnt != 0) { 2645 panic("page_free pp=%p, pfn=%lx, lckcnt=%d, cowcnt=%d " 2646 "slckcnt = %d", (void *)pp, page_pptonum(pp), pp->p_lckcnt, 2647 pp->p_cowcnt, pp->p_slckcnt); 2648 /*NOTREACHED*/ 2649 } 2650 2651 ASSERT(!hat_page_getshare(pp)); 2652 2653 PP_SETFREE(pp); 2654 ASSERT(pp->p_vnode == NULL || !IS_VMODSORT(pp->p_vnode) || 2655 !hat_ismod(pp)); 2656 page_clr_all_props(pp, 0); 2657 ASSERT(!hat_page_getshare(pp)); 2658 2659 /* 2660 * Now we add the page to the head of the free list. 2661 * But if this page is associated with a paged vnode 2662 * then we adjust the head forward so that the page is 2663 * effectively at the end of the list. 2664 */ 2665 if (pp->p_vnode == NULL) { 2666 /* 2667 * Page has no identity, put it on the free list. 2668 */ 2669 PP_SETAGED(pp); 2670 pp->p_offset = (u_offset_t)-1; 2671 page_list_add(pp, PG_FREE_LIST | PG_LIST_TAIL); 2672 VM_STAT_ADD(pagecnt.pc_free_free); 2673 TRACE_1(TR_FAC_VM, TR_PAGE_FREE_FREE, 2674 "page_free_free:pp %p", pp); 2675 } else { 2676 PP_CLRAGED(pp); 2677 2678 if (!dontneed || nopageage) { 2679 /* move it to the tail of the list */ 2680 page_list_add(pp, PG_CACHE_LIST | PG_LIST_TAIL); 2681 2682 VM_STAT_ADD(pagecnt.pc_free_cache); 2683 TRACE_1(TR_FAC_VM, TR_PAGE_FREE_CACHE_TAIL, 2684 "page_free_cache_tail:pp %p", pp); 2685 } else { 2686 page_list_add(pp, PG_CACHE_LIST | PG_LIST_HEAD); 2687 2688 VM_STAT_ADD(pagecnt.pc_free_dontneed); 2689 TRACE_1(TR_FAC_VM, TR_PAGE_FREE_CACHE_HEAD, 2690 "page_free_cache_head:pp %p", pp); 2691 } 2692 } 2693 page_unlock(pp); 2694 2695 /* 2696 * Now do the `freemem' accounting. 2697 */ 2698 pcf_index = PCF_INDEX(); 2699 p = &pcf[pcf_index]; 2700 2701 mutex_enter(&p->pcf_lock); 2702 if (p->pcf_block) { 2703 p->pcf_reserve += 1; 2704 } else { 2705 p->pcf_count += 1; 2706 if (p->pcf_wait) { 2707 mutex_enter(&new_freemem_lock); 2708 /* 2709 * Check to see if some other thread 2710 * is actually waiting. Another bucket 2711 * may have woken it up by now. If there 2712 * are no waiters, then set our pcf_wait 2713 * count to zero to avoid coming in here 2714 * next time. Also, since only one page 2715 * was put on the free list, just wake 2716 * up one waiter. 2717 */ 2718 if (freemem_wait) { 2719 cv_signal(&freemem_cv); 2720 p->pcf_wait--; 2721 } else { 2722 p->pcf_wait = 0; 2723 } 2724 mutex_exit(&new_freemem_lock); 2725 } 2726 } 2727 mutex_exit(&p->pcf_lock); 2728 2729 /* freemem is approximate, so this test OK */ 2730 if (!p->pcf_block) 2731 freemem += 1; 2732 } 2733 2734 /* 2735 * Put page on the "free" list during intial startup. 2736 * This happens during initial single threaded execution. 2737 */ 2738 void 2739 page_free_at_startup(page_t *pp) 2740 { 2741 struct pcf *p; 2742 uint_t pcf_index; 2743 2744 page_list_add(pp, PG_FREE_LIST | PG_LIST_HEAD | PG_LIST_ISINIT); 2745 VM_STAT_ADD(pagecnt.pc_free_free); 2746 2747 /* 2748 * Now do the `freemem' accounting. 2749 */ 2750 pcf_index = PCF_INDEX(); 2751 p = &pcf[pcf_index]; 2752 2753 ASSERT(p->pcf_block == 0); 2754 ASSERT(p->pcf_wait == 0); 2755 p->pcf_count += 1; 2756 2757 /* freemem is approximate, so this is OK */ 2758 freemem += 1; 2759 } 2760 2761 void 2762 page_free_pages(page_t *pp) 2763 { 2764 page_t *tpp, *rootpp = NULL; 2765 pgcnt_t pgcnt = page_get_pagecnt(pp->p_szc); 2766 pgcnt_t i; 2767 uint_t szc = pp->p_szc; 2768 2769 VM_STAT_ADD(pagecnt.pc_free_pages); 2770 TRACE_1(TR_FAC_VM, TR_PAGE_FREE_FREE, 2771 "page_free_free:pp %p", pp); 2772 2773 ASSERT(pp->p_szc != 0 && pp->p_szc < page_num_pagesizes()); 2774 if ((page_pptonum(pp) & (pgcnt - 1)) != 0) { 2775 panic("page_free_pages: not root page %p", (void *)pp); 2776 /*NOTREACHED*/ 2777 } 2778 2779 for (i = 0, tpp = pp; i < pgcnt; i++, tpp++) { 2780 ASSERT((PAGE_EXCL(tpp) && 2781 !page_iolock_assert(tpp)) || panicstr); 2782 if (PP_ISFREE(tpp)) { 2783 panic("page_free_pages: page %p is free", (void *)tpp); 2784 /*NOTREACHED*/ 2785 } 2786 if (hat_page_is_mapped(tpp) || tpp->p_lckcnt != 0 || 2787 tpp->p_cowcnt != 0 || tpp->p_slckcnt != 0) { 2788 panic("page_free_pages %p", (void *)tpp); 2789 /*NOTREACHED*/ 2790 } 2791 2792 ASSERT(!hat_page_getshare(tpp)); 2793 ASSERT(tpp->p_vnode == NULL); 2794 ASSERT(tpp->p_szc == szc); 2795 2796 PP_SETFREE(tpp); 2797 page_clr_all_props(tpp, 0); 2798 PP_SETAGED(tpp); 2799 tpp->p_offset = (u_offset_t)-1; 2800 ASSERT(tpp->p_next == tpp); 2801 ASSERT(tpp->p_prev == tpp); 2802 page_list_concat(&rootpp, &tpp); 2803 } 2804 ASSERT(rootpp == pp); 2805 2806 page_list_add_pages(rootpp, 0); 2807 page_create_putback(pgcnt); 2808 } 2809 2810 int free_pages = 1; 2811 2812 /* 2813 * This routine attempts to return pages to the cachelist via page_release(). 2814 * It does not *have* to be successful in all cases, since the pageout scanner 2815 * will catch any pages it misses. It does need to be fast and not introduce 2816 * too much overhead. 2817 * 2818 * If a page isn't found on the unlocked sweep of the page_hash bucket, we 2819 * don't lock and retry. This is ok, since the page scanner will eventually 2820 * find any page we miss in free_vp_pages(). 2821 */ 2822 void 2823 free_vp_pages(vnode_t *vp, u_offset_t off, size_t len) 2824 { 2825 page_t *pp; 2826 u_offset_t eoff; 2827 extern int swap_in_range(vnode_t *, u_offset_t, size_t); 2828 2829 eoff = off + len; 2830 2831 if (free_pages == 0) 2832 return; 2833 if (swap_in_range(vp, off, len)) 2834 return; 2835 2836 for (; off < eoff; off += PAGESIZE) { 2837 2838 /* 2839 * find the page using a fast, but inexact search. It'll be OK 2840 * if a few pages slip through the cracks here. 2841 */ 2842 pp = page_exists(vp, off); 2843 2844 /* 2845 * If we didn't find the page (it may not exist), the page 2846 * is free, looks still in use (shared), or we can't lock it, 2847 * just give up. 2848 */ 2849 if (pp == NULL || 2850 PP_ISFREE(pp) || 2851 page_share_cnt(pp) > 0 || 2852 !page_trylock(pp, SE_EXCL)) 2853 continue; 2854 2855 /* 2856 * Once we have locked pp, verify that it's still the 2857 * correct page and not already free 2858 */ 2859 ASSERT(PAGE_LOCKED_SE(pp, SE_EXCL)); 2860 if (pp->p_vnode != vp || pp->p_offset != off || PP_ISFREE(pp)) { 2861 page_unlock(pp); 2862 continue; 2863 } 2864 2865 /* 2866 * try to release the page... 2867 */ 2868 (void) page_release(pp, 1); 2869 } 2870 } 2871 2872 /* 2873 * Reclaim the given page from the free list. 2874 * If pp is part of a large pages, only the given constituent page is reclaimed 2875 * and the large page it belonged to will be demoted. This can only happen 2876 * if the page is not on the cachelist. 2877 * 2878 * Returns 1 on success or 0 on failure. 2879 * 2880 * The page is unlocked if it can't be reclaimed (when freemem == 0). 2881 * If `lock' is non-null, it will be dropped and re-acquired if 2882 * the routine must wait while freemem is 0. 2883 * 2884 * As it turns out, boot_getpages() does this. It picks a page, 2885 * based on where OBP mapped in some address, gets its pfn, searches 2886 * the memsegs, locks the page, then pulls it off the free list! 2887 */ 2888 int 2889 page_reclaim(page_t *pp, kmutex_t *lock) 2890 { 2891 struct pcf *p; 2892 struct cpu *cpup; 2893 int enough; 2894 uint_t i; 2895 2896 ASSERT(lock != NULL ? MUTEX_HELD(lock) : 1); 2897 ASSERT(PAGE_EXCL(pp) && PP_ISFREE(pp)); 2898 2899 /* 2900 * If `freemem' is 0, we cannot reclaim this page from the 2901 * freelist, so release every lock we might hold: the page, 2902 * and the `lock' before blocking. 2903 * 2904 * The only way `freemem' can become 0 while there are pages 2905 * marked free (have their p->p_free bit set) is when the 2906 * system is low on memory and doing a page_create(). In 2907 * order to guarantee that once page_create() starts acquiring 2908 * pages it will be able to get all that it needs since `freemem' 2909 * was decreased by the requested amount. So, we need to release 2910 * this page, and let page_create() have it. 2911 * 2912 * Since `freemem' being zero is not supposed to happen, just 2913 * use the usual hash stuff as a starting point. If that bucket 2914 * is empty, then assume the worst, and start at the beginning 2915 * of the pcf array. If we always start at the beginning 2916 * when acquiring more than one pcf lock, there won't be any 2917 * deadlock problems. 2918 */ 2919 2920 /* TODO: Do we need to test kcage_freemem if PG_NORELOC(pp)? */ 2921 2922 if (freemem <= throttlefree && !page_create_throttle(1l, 0)) { 2923 pcf_acquire_all(); 2924 goto page_reclaim_nomem; 2925 } 2926 2927 enough = pcf_decrement_bucket(1); 2928 2929 if (!enough) { 2930 VM_STAT_ADD(page_reclaim_zero); 2931 /* 2932 * Check again. Its possible that some other thread 2933 * could have been right behind us, and added one 2934 * to a list somewhere. Acquire each of the pcf locks 2935 * until we find a page. 2936 */ 2937 p = pcf; 2938 for (i = 0; i < pcf_fanout; i++) { 2939 mutex_enter(&p->pcf_lock); 2940 if (p->pcf_count >= 1) { 2941 p->pcf_count -= 1; 2942 enough = 1; 2943 break; 2944 } 2945 p++; 2946 } 2947 2948 if (!enough) { 2949 page_reclaim_nomem: 2950 /* 2951 * We really can't have page `pp'. 2952 * Time for the no-memory dance with 2953 * page_free(). This is just like 2954 * page_create_wait(). Plus the added 2955 * attraction of releasing whatever mutex 2956 * we held when we were called with in `lock'. 2957 * Page_unlock() will wakeup any thread 2958 * waiting around for this page. 2959 */ 2960 if (lock) { 2961 VM_STAT_ADD(page_reclaim_zero_locked); 2962 mutex_exit(lock); 2963 } 2964 page_unlock(pp); 2965 2966 /* 2967 * get this before we drop all the pcf locks. 2968 */ 2969 mutex_enter(&new_freemem_lock); 2970 2971 p = pcf; 2972 for (i = 0; i < pcf_fanout; i++) { 2973 p->pcf_wait++; 2974 mutex_exit(&p->pcf_lock); 2975 p++; 2976 } 2977 2978 freemem_wait++; 2979 cv_wait(&freemem_cv, &new_freemem_lock); 2980 freemem_wait--; 2981 2982 mutex_exit(&new_freemem_lock); 2983 2984 if (lock) { 2985 mutex_enter(lock); 2986 } 2987 return (0); 2988 } 2989 2990 /* 2991 * The pcf accounting has been done, 2992 * though none of the pcf_wait flags have been set, 2993 * drop the locks and continue on. 2994 */ 2995 while (p >= pcf) { 2996 mutex_exit(&p->pcf_lock); 2997 p--; 2998 } 2999 } 3000 3001 /* 3002 * freemem is not protected by any lock. Thus, we cannot 3003 * have any assertion containing freemem here. 3004 */ 3005 freemem -= 1; 3006 3007 VM_STAT_ADD(pagecnt.pc_reclaim); 3008 3009 /* 3010 * page_list_sub will handle the case where pp is a large page. 3011 * It's possible that the page was promoted while on the freelist 3012 */ 3013 if (PP_ISAGED(pp)) { 3014 page_list_sub(pp, PG_FREE_LIST); 3015 TRACE_1(TR_FAC_VM, TR_PAGE_UNFREE_FREE, 3016 "page_reclaim_free:pp %p", pp); 3017 } else { 3018 page_list_sub(pp, PG_CACHE_LIST); 3019 TRACE_1(TR_FAC_VM, TR_PAGE_UNFREE_CACHE, 3020 "page_reclaim_cache:pp %p", pp); 3021 } 3022 3023 /* 3024 * clear the p_free & p_age bits since this page is no longer 3025 * on the free list. Notice that there was a brief time where 3026 * a page is marked as free, but is not on the list. 3027 * 3028 * Set the reference bit to protect against immediate pageout. 3029 */ 3030 PP_CLRFREE(pp); 3031 PP_CLRAGED(pp); 3032 page_set_props(pp, P_REF); 3033 3034 CPU_STATS_ENTER_K(); 3035 cpup = CPU; /* get cpup now that CPU cannot change */ 3036 CPU_STATS_ADDQ(cpup, vm, pgrec, 1); 3037 CPU_STATS_ADDQ(cpup, vm, pgfrec, 1); 3038 CPU_STATS_EXIT_K(); 3039 ASSERT(pp->p_szc == 0); 3040 3041 return (1); 3042 } 3043 3044 /* 3045 * Destroy identity of the page and put it back on 3046 * the page free list. Assumes that the caller has 3047 * acquired the "exclusive" lock on the page. 3048 */ 3049 void 3050 page_destroy(page_t *pp, int dontfree) 3051 { 3052 ASSERT((PAGE_EXCL(pp) && 3053 !page_iolock_assert(pp)) || panicstr); 3054 ASSERT(pp->p_slckcnt == 0 || panicstr); 3055 3056 if (pp->p_szc != 0) { 3057 if (pp->p_vnode == NULL || IS_SWAPFSVP(pp->p_vnode) || 3058 PP_ISKAS(pp)) { 3059 panic("page_destroy: anon or kernel or no vnode " 3060 "large page %p", (void *)pp); 3061 } 3062 page_demote_vp_pages(pp); 3063 ASSERT(pp->p_szc == 0); 3064 } 3065 3066 TRACE_1(TR_FAC_VM, TR_PAGE_DESTROY, "page_destroy:pp %p", pp); 3067 3068 /* 3069 * Unload translations, if any, then hash out the 3070 * page to erase its identity. 3071 */ 3072 (void) hat_pageunload(pp, HAT_FORCE_PGUNLOAD); 3073 page_hashout(pp, NULL); 3074 3075 if (!dontfree) { 3076 /* 3077 * Acquire the "freemem_lock" for availrmem. 3078 * The page_struct_lock need not be acquired for lckcnt 3079 * and cowcnt since the page has an "exclusive" lock. 3080 */ 3081 if ((pp->p_lckcnt != 0) || (pp->p_cowcnt != 0)) { 3082 mutex_enter(&freemem_lock); 3083 if (pp->p_lckcnt != 0) { 3084 availrmem++; 3085 pp->p_lckcnt = 0; 3086 } 3087 if (pp->p_cowcnt != 0) { 3088 availrmem += pp->p_cowcnt; 3089 pp->p_cowcnt = 0; 3090 } 3091 mutex_exit(&freemem_lock); 3092 } 3093 /* 3094 * Put the page on the "free" list. 3095 */ 3096 page_free(pp, 0); 3097 } 3098 } 3099 3100 void 3101 page_destroy_pages(page_t *pp) 3102 { 3103 3104 page_t *tpp, *rootpp = NULL; 3105 pgcnt_t pgcnt = page_get_pagecnt(pp->p_szc); 3106 pgcnt_t i, pglcks = 0; 3107 uint_t szc = pp->p_szc; 3108 3109 ASSERT(pp->p_szc != 0 && pp->p_szc < page_num_pagesizes()); 3110 3111 VM_STAT_ADD(pagecnt.pc_destroy_pages); 3112 3113 TRACE_1(TR_FAC_VM, TR_PAGE_DESTROY, "page_destroy_pages:pp %p", pp); 3114 3115 if ((page_pptonum(pp) & (pgcnt - 1)) != 0) { 3116 panic("page_destroy_pages: not root page %p", (void *)pp); 3117 /*NOTREACHED*/ 3118 } 3119 3120 for (i = 0, tpp = pp; i < pgcnt; i++, tpp++) { 3121 ASSERT((PAGE_EXCL(tpp) && 3122 !page_iolock_assert(tpp)) || panicstr); 3123 ASSERT(tpp->p_slckcnt == 0 || panicstr); 3124 (void) hat_pageunload(tpp, HAT_FORCE_PGUNLOAD); 3125 page_hashout(tpp, NULL); 3126 ASSERT(tpp->p_offset == (u_offset_t)-1); 3127 if (tpp->p_lckcnt != 0) { 3128 pglcks++; 3129 tpp->p_lckcnt = 0; 3130 } else if (tpp->p_cowcnt != 0) { 3131 pglcks += tpp->p_cowcnt; 3132 tpp->p_cowcnt = 0; 3133 } 3134 ASSERT(!hat_page_getshare(tpp)); 3135 ASSERT(tpp->p_vnode == NULL); 3136 ASSERT(tpp->p_szc == szc); 3137 3138 PP_SETFREE(tpp); 3139 page_clr_all_props(tpp, 0); 3140 PP_SETAGED(tpp); 3141 ASSERT(tpp->p_next == tpp); 3142 ASSERT(tpp->p_prev == tpp); 3143 page_list_concat(&rootpp, &tpp); 3144 } 3145 3146 ASSERT(rootpp == pp); 3147 if (pglcks != 0) { 3148 mutex_enter(&freemem_lock); 3149 availrmem += pglcks; 3150 mutex_exit(&freemem_lock); 3151 } 3152 3153 page_list_add_pages(rootpp, 0); 3154 page_create_putback(pgcnt); 3155 } 3156 3157 /* 3158 * Similar to page_destroy(), but destroys pages which are 3159 * locked and known to be on the page free list. Since 3160 * the page is known to be free and locked, no one can access 3161 * it. 3162 * 3163 * Also, the number of free pages does not change. 3164 */ 3165 void 3166 page_destroy_free(page_t *pp) 3167 { 3168 ASSERT(PAGE_EXCL(pp)); 3169 ASSERT(PP_ISFREE(pp)); 3170 ASSERT(pp->p_vnode); 3171 ASSERT(hat_page_getattr(pp, P_MOD | P_REF | P_RO) == 0); 3172 ASSERT(!hat_page_is_mapped(pp)); 3173 ASSERT(PP_ISAGED(pp) == 0); 3174 ASSERT(pp->p_szc == 0); 3175 3176 VM_STAT_ADD(pagecnt.pc_destroy_free); 3177 page_list_sub(pp, PG_CACHE_LIST); 3178 3179 page_hashout(pp, NULL); 3180 ASSERT(pp->p_vnode == NULL); 3181 ASSERT(pp->p_offset == (u_offset_t)-1); 3182 ASSERT(pp->p_hash == NULL); 3183 3184 PP_SETAGED(pp); 3185 page_list_add(pp, PG_FREE_LIST | PG_LIST_TAIL); 3186 page_unlock(pp); 3187 3188 mutex_enter(&new_freemem_lock); 3189 if (freemem_wait) { 3190 cv_signal(&freemem_cv); 3191 } 3192 mutex_exit(&new_freemem_lock); 3193 } 3194 3195 /* 3196 * Rename the page "opp" to have an identity specified 3197 * by [vp, off]. If a page already exists with this name 3198 * it is locked and destroyed. Note that the page's 3199 * translations are not unloaded during the rename. 3200 * 3201 * This routine is used by the anon layer to "steal" the 3202 * original page and is not unlike destroying a page and 3203 * creating a new page using the same page frame. 3204 * 3205 * XXX -- Could deadlock if caller 1 tries to rename A to B while 3206 * caller 2 tries to rename B to A. 3207 */ 3208 void 3209 page_rename(page_t *opp, vnode_t *vp, u_offset_t off) 3210 { 3211 page_t *pp; 3212 int olckcnt = 0; 3213 int ocowcnt = 0; 3214 kmutex_t *phm; 3215 ulong_t index; 3216 3217 ASSERT(PAGE_EXCL(opp) && !page_iolock_assert(opp)); 3218 ASSERT(MUTEX_NOT_HELD(page_vnode_mutex(vp))); 3219 ASSERT(PP_ISFREE(opp) == 0); 3220 3221 VM_STAT_ADD(page_rename_count); 3222 3223 TRACE_3(TR_FAC_VM, TR_PAGE_RENAME, 3224 "page rename:pp %p vp %p off %llx", opp, vp, off); 3225 3226 /* 3227 * CacheFS may call page_rename for a large NFS page 3228 * when both CacheFS and NFS mount points are used 3229 * by applications. Demote this large page before 3230 * renaming it, to ensure that there are no "partial" 3231 * large pages left lying around. 3232 */ 3233 if (opp->p_szc != 0) { 3234 vnode_t *ovp = opp->p_vnode; 3235 ASSERT(ovp != NULL); 3236 ASSERT(!IS_SWAPFSVP(ovp)); 3237 ASSERT(!VN_ISKAS(ovp)); 3238 page_demote_vp_pages(opp); 3239 ASSERT(opp->p_szc == 0); 3240 } 3241 3242 page_hashout(opp, NULL); 3243 PP_CLRAGED(opp); 3244 3245 /* 3246 * Acquire the appropriate page hash lock, since 3247 * we're going to rename the page. 3248 */ 3249 index = PAGE_HASH_FUNC(vp, off); 3250 phm = PAGE_HASH_MUTEX(index); 3251 mutex_enter(phm); 3252 top: 3253 /* 3254 * Look for an existing page with this name and destroy it if found. 3255 * By holding the page hash lock all the way to the page_hashin() 3256 * call, we are assured that no page can be created with this 3257 * identity. In the case when the phm lock is dropped to undo any 3258 * hat layer mappings, the existing page is held with an "exclusive" 3259 * lock, again preventing another page from being created with 3260 * this identity. 3261 */ 3262 PAGE_HASH_SEARCH(index, pp, vp, off); 3263 if (pp != NULL) { 3264 VM_STAT_ADD(page_rename_exists); 3265 3266 /* 3267 * As it turns out, this is one of only two places where 3268 * page_lock() needs to hold the passed in lock in the 3269 * successful case. In all of the others, the lock could 3270 * be dropped as soon as the attempt is made to lock 3271 * the page. It is tempting to add yet another arguement, 3272 * PL_KEEP or PL_DROP, to let page_lock know what to do. 3273 */ 3274 if (!page_lock(pp, SE_EXCL, phm, P_RECLAIM)) { 3275 /* 3276 * Went to sleep because the page could not 3277 * be locked. We were woken up when the page 3278 * was unlocked, or when the page was destroyed. 3279 * In either case, `phm' was dropped while we 3280 * slept. Hence we should not just roar through 3281 * this loop. 3282 */ 3283 goto top; 3284 } 3285 3286 /* 3287 * If an existing page is a large page, then demote 3288 * it to ensure that no "partial" large pages are 3289 * "created" after page_rename. An existing page 3290 * can be a CacheFS page, and can't belong to swapfs. 3291 */ 3292 if (hat_page_is_mapped(pp)) { 3293 /* 3294 * Unload translations. Since we hold the 3295 * exclusive lock on this page, the page 3296 * can not be changed while we drop phm. 3297 * This is also not a lock protocol violation, 3298 * but rather the proper way to do things. 3299 */ 3300 mutex_exit(phm); 3301 (void) hat_pageunload(pp, HAT_FORCE_PGUNLOAD); 3302 if (pp->p_szc != 0) { 3303 ASSERT(!IS_SWAPFSVP(vp)); 3304 ASSERT(!VN_ISKAS(vp)); 3305 page_demote_vp_pages(pp); 3306 ASSERT(pp->p_szc == 0); 3307 } 3308 mutex_enter(phm); 3309 } else if (pp->p_szc != 0) { 3310 ASSERT(!IS_SWAPFSVP(vp)); 3311 ASSERT(!VN_ISKAS(vp)); 3312 mutex_exit(phm); 3313 page_demote_vp_pages(pp); 3314 ASSERT(pp->p_szc == 0); 3315 mutex_enter(phm); 3316 } 3317 page_hashout(pp, phm); 3318 } 3319 /* 3320 * Hash in the page with the new identity. 3321 */ 3322 if (!page_hashin(opp, vp, off, phm)) { 3323 /* 3324 * We were holding phm while we searched for [vp, off] 3325 * and only dropped phm if we found and locked a page. 3326 * If we can't create this page now, then some thing 3327 * is really broken. 3328 */ 3329 panic("page_rename: Can't hash in page: %p", (void *)pp); 3330 /*NOTREACHED*/ 3331 } 3332 3333 ASSERT(MUTEX_HELD(phm)); 3334 mutex_exit(phm); 3335 3336 /* 3337 * Now that we have dropped phm, lets get around to finishing up 3338 * with pp. 3339 */ 3340 if (pp != NULL) { 3341 ASSERT(!hat_page_is_mapped(pp)); 3342 /* for now large pages should not end up here */ 3343 ASSERT(pp->p_szc == 0); 3344 /* 3345 * Save the locks for transfer to the new page and then 3346 * clear them so page_free doesn't think they're important. 3347 * The page_struct_lock need not be acquired for lckcnt and 3348 * cowcnt since the page has an "exclusive" lock. 3349 */ 3350 olckcnt = pp->p_lckcnt; 3351 ocowcnt = pp->p_cowcnt; 3352 pp->p_lckcnt = pp->p_cowcnt = 0; 3353 3354 /* 3355 * Put the page on the "free" list after we drop 3356 * the lock. The less work under the lock the better. 3357 */ 3358 /*LINTED: constant in conditional context*/ 3359 VN_DISPOSE(pp, B_FREE, 0, kcred); 3360 } 3361 3362 /* 3363 * Transfer the lock count from the old page (if any). 3364 * The page_struct_lock need not be acquired for lckcnt and 3365 * cowcnt since the page has an "exclusive" lock. 3366 */ 3367 opp->p_lckcnt += olckcnt; 3368 opp->p_cowcnt += ocowcnt; 3369 } 3370 3371 /* 3372 * low level routine to add page `pp' to the hash and vp chains for [vp, offset] 3373 * 3374 * Pages are normally inserted at the start of a vnode's v_pages list. 3375 * If the vnode is VMODSORT and the page is modified, it goes at the end. 3376 * This can happen when a modified page is relocated for DR. 3377 * 3378 * Returns 1 on success and 0 on failure. 3379 */ 3380 static int 3381 page_do_hashin(page_t *pp, vnode_t *vp, u_offset_t offset) 3382 { 3383 page_t **listp; 3384 page_t *tp; 3385 ulong_t index; 3386 3387 ASSERT(PAGE_EXCL(pp)); 3388 ASSERT(vp != NULL); 3389 ASSERT(MUTEX_HELD(page_vnode_mutex(vp))); 3390 3391 /* 3392 * Be sure to set these up before the page is inserted on the hash 3393 * list. As soon as the page is placed on the list some other 3394 * thread might get confused and wonder how this page could 3395 * possibly hash to this list. 3396 */ 3397 pp->p_vnode = vp; 3398 pp->p_offset = offset; 3399 3400 /* 3401 * record if this page is on a swap vnode 3402 */ 3403 if ((vp->v_flag & VISSWAP) != 0) 3404 PP_SETSWAP(pp); 3405 3406 index = PAGE_HASH_FUNC(vp, offset); 3407 ASSERT(MUTEX_HELD(PAGE_HASH_MUTEX(index))); 3408 listp = &page_hash[index]; 3409 3410 /* 3411 * If this page is already hashed in, fail this attempt to add it. 3412 */ 3413 for (tp = *listp; tp != NULL; tp = tp->p_hash) { 3414 if (tp->p_vnode == vp && tp->p_offset == offset) { 3415 pp->p_vnode = NULL; 3416 pp->p_offset = (u_offset_t)(-1); 3417 return (0); 3418 } 3419 } 3420 pp->p_hash = *listp; 3421 *listp = pp; 3422 3423 /* 3424 * Add the page to the vnode's list of pages 3425 */ 3426 if (vp->v_pages != NULL && IS_VMODSORT(vp) && hat_ismod(pp)) 3427 listp = &vp->v_pages->p_vpprev->p_vpnext; 3428 else 3429 listp = &vp->v_pages; 3430 3431 page_vpadd(listp, pp); 3432 3433 return (1); 3434 } 3435 3436 /* 3437 * Add page `pp' to both the hash and vp chains for [vp, offset]. 3438 * 3439 * Returns 1 on success and 0 on failure. 3440 * If hold is passed in, it is not dropped. 3441 */ 3442 int 3443 page_hashin(page_t *pp, vnode_t *vp, u_offset_t offset, kmutex_t *hold) 3444 { 3445 kmutex_t *phm = NULL; 3446 kmutex_t *vphm; 3447 int rc; 3448 3449 ASSERT(MUTEX_NOT_HELD(page_vnode_mutex(vp))); 3450 3451 TRACE_3(TR_FAC_VM, TR_PAGE_HASHIN, 3452 "page_hashin:pp %p vp %p offset %llx", 3453 pp, vp, offset); 3454 3455 VM_STAT_ADD(hashin_count); 3456 3457 if (hold != NULL) 3458 phm = hold; 3459 else { 3460 VM_STAT_ADD(hashin_not_held); 3461 phm = PAGE_HASH_MUTEX(PAGE_HASH_FUNC(vp, offset)); 3462 mutex_enter(phm); 3463 } 3464 3465 vphm = page_vnode_mutex(vp); 3466 mutex_enter(vphm); 3467 rc = page_do_hashin(pp, vp, offset); 3468 mutex_exit(vphm); 3469 if (hold == NULL) 3470 mutex_exit(phm); 3471 if (rc == 0) 3472 VM_STAT_ADD(hashin_already); 3473 return (rc); 3474 } 3475 3476 /* 3477 * Remove page ``pp'' from the hash and vp chains and remove vp association. 3478 * All mutexes must be held 3479 */ 3480 static void 3481 page_do_hashout(page_t *pp) 3482 { 3483 page_t **hpp; 3484 page_t *hp; 3485 vnode_t *vp = pp->p_vnode; 3486 3487 ASSERT(vp != NULL); 3488 ASSERT(MUTEX_HELD(page_vnode_mutex(vp))); 3489 3490 /* 3491 * First, take pp off of its hash chain. 3492 */ 3493 hpp = &page_hash[PAGE_HASH_FUNC(vp, pp->p_offset)]; 3494 3495 for (;;) { 3496 hp = *hpp; 3497 if (hp == pp) 3498 break; 3499 if (hp == NULL) { 3500 panic("page_do_hashout"); 3501 /*NOTREACHED*/ 3502 } 3503 hpp = &hp->p_hash; 3504 } 3505 *hpp = pp->p_hash; 3506 3507 /* 3508 * Now remove it from its associated vnode. 3509 */ 3510 if (vp->v_pages) 3511 page_vpsub(&vp->v_pages, pp); 3512 3513 pp->p_hash = NULL; 3514 page_clr_all_props(pp, 1); 3515 PP_CLRSWAP(pp); 3516 pp->p_vnode = NULL; 3517 pp->p_offset = (u_offset_t)-1; 3518 } 3519 3520 /* 3521 * Remove page ``pp'' from the hash and vp chains and remove vp association. 3522 * 3523 * When `phm' is non-NULL it contains the address of the mutex protecting the 3524 * hash list pp is on. It is not dropped. 3525 */ 3526 void 3527 page_hashout(page_t *pp, kmutex_t *phm) 3528 { 3529 vnode_t *vp; 3530 ulong_t index; 3531 kmutex_t *nphm; 3532 kmutex_t *vphm; 3533 kmutex_t *sep; 3534 3535 ASSERT(phm != NULL ? MUTEX_HELD(phm) : 1); 3536 ASSERT(pp->p_vnode != NULL); 3537 ASSERT((PAGE_EXCL(pp) && !page_iolock_assert(pp)) || panicstr); 3538 ASSERT(MUTEX_NOT_HELD(page_vnode_mutex(pp->p_vnode))); 3539 3540 vp = pp->p_vnode; 3541 3542 TRACE_2(TR_FAC_VM, TR_PAGE_HASHOUT, 3543 "page_hashout:pp %p vp %p", pp, vp); 3544 3545 /* Kernel probe */ 3546 TNF_PROBE_2(page_unmap, "vm pagefault", /* CSTYLED */, 3547 tnf_opaque, vnode, vp, 3548 tnf_offset, offset, pp->p_offset); 3549 3550 /* 3551 * 3552 */ 3553 VM_STAT_ADD(hashout_count); 3554 index = PAGE_HASH_FUNC(vp, pp->p_offset); 3555 if (phm == NULL) { 3556 VM_STAT_ADD(hashout_not_held); 3557 nphm = PAGE_HASH_MUTEX(index); 3558 mutex_enter(nphm); 3559 } 3560 ASSERT(phm ? phm == PAGE_HASH_MUTEX(index) : 1); 3561 3562 3563 /* 3564 * grab page vnode mutex and remove it... 3565 */ 3566 vphm = page_vnode_mutex(vp); 3567 mutex_enter(vphm); 3568 3569 page_do_hashout(pp); 3570 3571 mutex_exit(vphm); 3572 if (phm == NULL) 3573 mutex_exit(nphm); 3574 3575 /* 3576 * Wake up processes waiting for this page. The page's 3577 * identity has been changed, and is probably not the 3578 * desired page any longer. 3579 */ 3580 sep = page_se_mutex(pp); 3581 mutex_enter(sep); 3582 pp->p_selock &= ~SE_EWANTED; 3583 if (CV_HAS_WAITERS(&pp->p_cv)) 3584 cv_broadcast(&pp->p_cv); 3585 mutex_exit(sep); 3586 } 3587 3588 /* 3589 * Add the page to the front of a linked list of pages 3590 * using the p_next & p_prev pointers for the list. 3591 * The caller is responsible for protecting the list pointers. 3592 */ 3593 void 3594 page_add(page_t **ppp, page_t *pp) 3595 { 3596 ASSERT(PAGE_EXCL(pp) || (PAGE_SHARED(pp) && page_iolock_assert(pp))); 3597 3598 page_add_common(ppp, pp); 3599 } 3600 3601 3602 3603 /* 3604 * Common code for page_add() and mach_page_add() 3605 */ 3606 void 3607 page_add_common(page_t **ppp, page_t *pp) 3608 { 3609 if (*ppp == NULL) { 3610 pp->p_next = pp->p_prev = pp; 3611 } else { 3612 pp->p_next = *ppp; 3613 pp->p_prev = (*ppp)->p_prev; 3614 (*ppp)->p_prev = pp; 3615 pp->p_prev->p_next = pp; 3616 } 3617 *ppp = pp; 3618 } 3619 3620 3621 /* 3622 * Remove this page from a linked list of pages 3623 * using the p_next & p_prev pointers for the list. 3624 * 3625 * The caller is responsible for protecting the list pointers. 3626 */ 3627 void 3628 page_sub(page_t **ppp, page_t *pp) 3629 { 3630 ASSERT((PP_ISFREE(pp)) ? 1 : 3631 (PAGE_EXCL(pp)) || (PAGE_SHARED(pp) && page_iolock_assert(pp))); 3632 3633 if (*ppp == NULL || pp == NULL) { 3634 panic("page_sub: bad arg(s): pp %p, *ppp %p", 3635 (void *)pp, (void *)(*ppp)); 3636 /*NOTREACHED*/ 3637 } 3638 3639 page_sub_common(ppp, pp); 3640 } 3641 3642 3643 /* 3644 * Common code for page_sub() and mach_page_sub() 3645 */ 3646 void 3647 page_sub_common(page_t **ppp, page_t *pp) 3648 { 3649 if (*ppp == pp) 3650 *ppp = pp->p_next; /* go to next page */ 3651 3652 if (*ppp == pp) 3653 *ppp = NULL; /* page list is gone */ 3654 else { 3655 pp->p_prev->p_next = pp->p_next; 3656 pp->p_next->p_prev = pp->p_prev; 3657 } 3658 pp->p_prev = pp->p_next = pp; /* make pp a list of one */ 3659 } 3660 3661 3662 /* 3663 * Break page list cppp into two lists with npages in the first list. 3664 * The tail is returned in nppp. 3665 */ 3666 void 3667 page_list_break(page_t **oppp, page_t **nppp, pgcnt_t npages) 3668 { 3669 page_t *s1pp = *oppp; 3670 page_t *s2pp; 3671 page_t *e1pp, *e2pp; 3672 long n = 0; 3673 3674 if (s1pp == NULL) { 3675 *nppp = NULL; 3676 return; 3677 } 3678 if (npages == 0) { 3679 *nppp = s1pp; 3680 *oppp = NULL; 3681 return; 3682 } 3683 for (n = 0, s2pp = *oppp; n < npages; n++) { 3684 s2pp = s2pp->p_next; 3685 } 3686 /* Fix head and tail of new lists */ 3687 e1pp = s2pp->p_prev; 3688 e2pp = s1pp->p_prev; 3689 s1pp->p_prev = e1pp; 3690 e1pp->p_next = s1pp; 3691 s2pp->p_prev = e2pp; 3692 e2pp->p_next = s2pp; 3693 3694 /* second list empty */ 3695 if (s2pp == s1pp) { 3696 *oppp = s1pp; 3697 *nppp = NULL; 3698 } else { 3699 *oppp = s1pp; 3700 *nppp = s2pp; 3701 } 3702 } 3703 3704 /* 3705 * Concatenate page list nppp onto the end of list ppp. 3706 */ 3707 void 3708 page_list_concat(page_t **ppp, page_t **nppp) 3709 { 3710 page_t *s1pp, *s2pp, *e1pp, *e2pp; 3711 3712 if (*nppp == NULL) { 3713 return; 3714 } 3715 if (*ppp == NULL) { 3716 *ppp = *nppp; 3717 return; 3718 } 3719 s1pp = *ppp; 3720 e1pp = s1pp->p_prev; 3721 s2pp = *nppp; 3722 e2pp = s2pp->p_prev; 3723 s1pp->p_prev = e2pp; 3724 e2pp->p_next = s1pp; 3725 e1pp->p_next = s2pp; 3726 s2pp->p_prev = e1pp; 3727 } 3728 3729 /* 3730 * return the next page in the page list 3731 */ 3732 page_t * 3733 page_list_next(page_t *pp) 3734 { 3735 return (pp->p_next); 3736 } 3737 3738 3739 /* 3740 * Add the page to the front of the linked list of pages 3741 * using p_vpnext/p_vpprev pointers for the list. 3742 * 3743 * The caller is responsible for protecting the lists. 3744 */ 3745 void 3746 page_vpadd(page_t **ppp, page_t *pp) 3747 { 3748 if (*ppp == NULL) { 3749 pp->p_vpnext = pp->p_vpprev = pp; 3750 } else { 3751 pp->p_vpnext = *ppp; 3752 pp->p_vpprev = (*ppp)->p_vpprev; 3753 (*ppp)->p_vpprev = pp; 3754 pp->p_vpprev->p_vpnext = pp; 3755 } 3756 *ppp = pp; 3757 } 3758 3759 /* 3760 * Remove this page from the linked list of pages 3761 * using p_vpnext/p_vpprev pointers for the list. 3762 * 3763 * The caller is responsible for protecting the lists. 3764 */ 3765 void 3766 page_vpsub(page_t **ppp, page_t *pp) 3767 { 3768 if (*ppp == NULL || pp == NULL) { 3769 panic("page_vpsub: bad arg(s): pp %p, *ppp %p", 3770 (void *)pp, (void *)(*ppp)); 3771 /*NOTREACHED*/ 3772 } 3773 3774 if (*ppp == pp) 3775 *ppp = pp->p_vpnext; /* go to next page */ 3776 3777 if (*ppp == pp) 3778 *ppp = NULL; /* page list is gone */ 3779 else { 3780 pp->p_vpprev->p_vpnext = pp->p_vpnext; 3781 pp->p_vpnext->p_vpprev = pp->p_vpprev; 3782 } 3783 pp->p_vpprev = pp->p_vpnext = pp; /* make pp a list of one */ 3784 } 3785 3786 /* 3787 * Lock a physical page into memory "long term". Used to support "lock 3788 * in memory" functions. Accepts the page to be locked, and a cow variable 3789 * to indicate whether a the lock will travel to the new page during 3790 * a potential copy-on-write. 3791 */ 3792 int 3793 page_pp_lock( 3794 page_t *pp, /* page to be locked */ 3795 int cow, /* cow lock */ 3796 int kernel) /* must succeed -- ignore checking */ 3797 { 3798 int r = 0; /* result -- assume failure */ 3799 3800 ASSERT(PAGE_LOCKED(pp)); 3801 3802 page_struct_lock(pp); 3803 /* 3804 * Acquire the "freemem_lock" for availrmem. 3805 */ 3806 if (cow) { 3807 mutex_enter(&freemem_lock); 3808 if ((availrmem > pages_pp_maximum) && 3809 (pp->p_cowcnt < (ushort_t)PAGE_LOCK_MAXIMUM)) { 3810 availrmem--; 3811 pages_locked++; 3812 mutex_exit(&freemem_lock); 3813 r = 1; 3814 if (++pp->p_cowcnt == (ushort_t)PAGE_LOCK_MAXIMUM) { 3815 cmn_err(CE_WARN, 3816 "COW lock limit reached on pfn 0x%lx", 3817 page_pptonum(pp)); 3818 } 3819 } else 3820 mutex_exit(&freemem_lock); 3821 } else { 3822 if (pp->p_lckcnt) { 3823 if (pp->p_lckcnt < (ushort_t)PAGE_LOCK_MAXIMUM) { 3824 r = 1; 3825 if (++pp->p_lckcnt == 3826 (ushort_t)PAGE_LOCK_MAXIMUM) { 3827 cmn_err(CE_WARN, "Page lock limit " 3828 "reached on pfn 0x%lx", 3829 page_pptonum(pp)); 3830 } 3831 } 3832 } else { 3833 if (kernel) { 3834 /* availrmem accounting done by caller */ 3835 ++pp->p_lckcnt; 3836 r = 1; 3837 } else { 3838 mutex_enter(&freemem_lock); 3839 if (availrmem > pages_pp_maximum) { 3840 availrmem--; 3841 pages_locked++; 3842 ++pp->p_lckcnt; 3843 r = 1; 3844 } 3845 mutex_exit(&freemem_lock); 3846 } 3847 } 3848 } 3849 page_struct_unlock(pp); 3850 return (r); 3851 } 3852 3853 /* 3854 * Decommit a lock on a physical page frame. Account for cow locks if 3855 * appropriate. 3856 */ 3857 void 3858 page_pp_unlock( 3859 page_t *pp, /* page to be unlocked */ 3860 int cow, /* expect cow lock */ 3861 int kernel) /* this was a kernel lock */ 3862 { 3863 ASSERT(PAGE_LOCKED(pp)); 3864 3865 page_struct_lock(pp); 3866 /* 3867 * Acquire the "freemem_lock" for availrmem. 3868 * If cowcnt or lcknt is already 0 do nothing; i.e., we 3869 * could be called to unlock even if nothing is locked. This could 3870 * happen if locked file pages were truncated (removing the lock) 3871 * and the file was grown again and new pages faulted in; the new 3872 * pages are unlocked but the segment still thinks they're locked. 3873 */ 3874 if (cow) { 3875 if (pp->p_cowcnt) { 3876 mutex_enter(&freemem_lock); 3877 pp->p_cowcnt--; 3878 availrmem++; 3879 pages_locked--; 3880 mutex_exit(&freemem_lock); 3881 } 3882 } else { 3883 if (pp->p_lckcnt && --pp->p_lckcnt == 0) { 3884 if (!kernel) { 3885 mutex_enter(&freemem_lock); 3886 availrmem++; 3887 pages_locked--; 3888 mutex_exit(&freemem_lock); 3889 } 3890 } 3891 } 3892 page_struct_unlock(pp); 3893 } 3894 3895 /* 3896 * This routine reserves availrmem for npages; 3897 * flags: KM_NOSLEEP or KM_SLEEP 3898 * returns 1 on success or 0 on failure 3899 */ 3900 int 3901 page_resv(pgcnt_t npages, uint_t flags) 3902 { 3903 mutex_enter(&freemem_lock); 3904 while (availrmem < tune.t_minarmem + npages) { 3905 if (flags & KM_NOSLEEP) { 3906 mutex_exit(&freemem_lock); 3907 return (0); 3908 } 3909 mutex_exit(&freemem_lock); 3910 page_needfree(npages); 3911 kmem_reap(); 3912 delay(hz >> 2); 3913 page_needfree(-(spgcnt_t)npages); 3914 mutex_enter(&freemem_lock); 3915 } 3916 availrmem -= npages; 3917 mutex_exit(&freemem_lock); 3918 return (1); 3919 } 3920 3921 /* 3922 * This routine unreserves availrmem for npages; 3923 */ 3924 void 3925 page_unresv(pgcnt_t npages) 3926 { 3927 mutex_enter(&freemem_lock); 3928 availrmem += npages; 3929 mutex_exit(&freemem_lock); 3930 } 3931 3932 /* 3933 * See Statement at the beginning of segvn_lockop() regarding 3934 * the way we handle cowcnts and lckcnts. 3935 * 3936 * Transfer cowcnt on 'opp' to cowcnt on 'npp' if the vpage 3937 * that breaks COW has PROT_WRITE. 3938 * 3939 * Note that, we may also break COW in case we are softlocking 3940 * on read access during physio; 3941 * in this softlock case, the vpage may not have PROT_WRITE. 3942 * So, we need to transfer lckcnt on 'opp' to lckcnt on 'npp' 3943 * if the vpage doesn't have PROT_WRITE. 3944 * 3945 * This routine is never called if we are stealing a page 3946 * in anon_private. 3947 * 3948 * The caller subtracted from availrmem for read only mapping. 3949 * if lckcnt is 1 increment availrmem. 3950 */ 3951 void 3952 page_pp_useclaim( 3953 page_t *opp, /* original page frame losing lock */ 3954 page_t *npp, /* new page frame gaining lock */ 3955 uint_t write_perm) /* set if vpage has PROT_WRITE */ 3956 { 3957 int payback = 0; 3958 3959 ASSERT(PAGE_LOCKED(opp)); 3960 ASSERT(PAGE_LOCKED(npp)); 3961 3962 page_struct_lock(opp); 3963 3964 ASSERT(npp->p_cowcnt == 0); 3965 ASSERT(npp->p_lckcnt == 0); 3966 3967 /* Don't use claim if nothing is locked (see page_pp_unlock above) */ 3968 if ((write_perm && opp->p_cowcnt != 0) || 3969 (!write_perm && opp->p_lckcnt != 0)) { 3970 3971 if (write_perm) { 3972 npp->p_cowcnt++; 3973 ASSERT(opp->p_cowcnt != 0); 3974 opp->p_cowcnt--; 3975 } else { 3976 3977 ASSERT(opp->p_lckcnt != 0); 3978 3979 /* 3980 * We didn't need availrmem decremented if p_lckcnt on 3981 * original page is 1. Here, we are unlocking 3982 * read-only copy belonging to original page and 3983 * are locking a copy belonging to new page. 3984 */ 3985 if (opp->p_lckcnt == 1) 3986 payback = 1; 3987 3988 npp->p_lckcnt++; 3989 opp->p_lckcnt--; 3990 } 3991 } 3992 if (payback) { 3993 mutex_enter(&freemem_lock); 3994 availrmem++; 3995 pages_useclaim--; 3996 mutex_exit(&freemem_lock); 3997 } 3998 page_struct_unlock(opp); 3999 } 4000 4001 /* 4002 * Simple claim adjust functions -- used to support changes in 4003 * claims due to changes in access permissions. Used by segvn_setprot(). 4004 */ 4005 int 4006 page_addclaim(page_t *pp) 4007 { 4008 int r = 0; /* result */ 4009 4010 ASSERT(PAGE_LOCKED(pp)); 4011 4012 page_struct_lock(pp); 4013 ASSERT(pp->p_lckcnt != 0); 4014 4015 if (pp->p_lckcnt == 1) { 4016 if (pp->p_cowcnt < (ushort_t)PAGE_LOCK_MAXIMUM) { 4017 --pp->p_lckcnt; 4018 r = 1; 4019 if (++pp->p_cowcnt == (ushort_t)PAGE_LOCK_MAXIMUM) { 4020 cmn_err(CE_WARN, 4021 "COW lock limit reached on pfn 0x%lx", 4022 page_pptonum(pp)); 4023 } 4024 } 4025 } else { 4026 mutex_enter(&freemem_lock); 4027 if ((availrmem > pages_pp_maximum) && 4028 (pp->p_cowcnt < (ushort_t)PAGE_LOCK_MAXIMUM)) { 4029 --availrmem; 4030 ++pages_claimed; 4031 mutex_exit(&freemem_lock); 4032 --pp->p_lckcnt; 4033 r = 1; 4034 if (++pp->p_cowcnt == (ushort_t)PAGE_LOCK_MAXIMUM) { 4035 cmn_err(CE_WARN, 4036 "COW lock limit reached on pfn 0x%lx", 4037 page_pptonum(pp)); 4038 } 4039 } else 4040 mutex_exit(&freemem_lock); 4041 } 4042 page_struct_unlock(pp); 4043 return (r); 4044 } 4045 4046 int 4047 page_subclaim(page_t *pp) 4048 { 4049 int r = 0; 4050 4051 ASSERT(PAGE_LOCKED(pp)); 4052 4053 page_struct_lock(pp); 4054 ASSERT(pp->p_cowcnt != 0); 4055 4056 if (pp->p_lckcnt) { 4057 if (pp->p_lckcnt < (ushort_t)PAGE_LOCK_MAXIMUM) { 4058 r = 1; 4059 /* 4060 * for availrmem 4061 */ 4062 mutex_enter(&freemem_lock); 4063 availrmem++; 4064 pages_claimed--; 4065 mutex_exit(&freemem_lock); 4066 4067 pp->p_cowcnt--; 4068 4069 if (++pp->p_lckcnt == (ushort_t)PAGE_LOCK_MAXIMUM) { 4070 cmn_err(CE_WARN, 4071 "Page lock limit reached on pfn 0x%lx", 4072 page_pptonum(pp)); 4073 } 4074 } 4075 } else { 4076 r = 1; 4077 pp->p_cowcnt--; 4078 pp->p_lckcnt++; 4079 } 4080 page_struct_unlock(pp); 4081 return (r); 4082 } 4083 4084 int 4085 page_addclaim_pages(page_t **ppa) 4086 { 4087 4088 pgcnt_t lckpgs = 0, pg_idx; 4089 4090 VM_STAT_ADD(pagecnt.pc_addclaim_pages); 4091 4092 mutex_enter(&page_llock); 4093 for (pg_idx = 0; ppa[pg_idx] != NULL; pg_idx++) { 4094 4095 ASSERT(PAGE_LOCKED(ppa[pg_idx])); 4096 ASSERT(ppa[pg_idx]->p_lckcnt != 0); 4097 if (ppa[pg_idx]->p_cowcnt == (ushort_t)PAGE_LOCK_MAXIMUM) { 4098 mutex_exit(&page_llock); 4099 return (0); 4100 } 4101 if (ppa[pg_idx]->p_lckcnt > 1) 4102 lckpgs++; 4103 } 4104 4105 if (lckpgs != 0) { 4106 mutex_enter(&freemem_lock); 4107 if (availrmem >= pages_pp_maximum + lckpgs) { 4108 availrmem -= lckpgs; 4109 pages_claimed += lckpgs; 4110 } else { 4111 mutex_exit(&freemem_lock); 4112 mutex_exit(&page_llock); 4113 return (0); 4114 } 4115 mutex_exit(&freemem_lock); 4116 } 4117 4118 for (pg_idx = 0; ppa[pg_idx] != NULL; pg_idx++) { 4119 ppa[pg_idx]->p_lckcnt--; 4120 ppa[pg_idx]->p_cowcnt++; 4121 } 4122 mutex_exit(&page_llock); 4123 return (1); 4124 } 4125 4126 int 4127 page_subclaim_pages(page_t **ppa) 4128 { 4129 pgcnt_t ulckpgs = 0, pg_idx; 4130 4131 VM_STAT_ADD(pagecnt.pc_subclaim_pages); 4132 4133 mutex_enter(&page_llock); 4134 for (pg_idx = 0; ppa[pg_idx] != NULL; pg_idx++) { 4135 4136 ASSERT(PAGE_LOCKED(ppa[pg_idx])); 4137 ASSERT(ppa[pg_idx]->p_cowcnt != 0); 4138 if (ppa[pg_idx]->p_lckcnt == (ushort_t)PAGE_LOCK_MAXIMUM) { 4139 mutex_exit(&page_llock); 4140 return (0); 4141 } 4142 if (ppa[pg_idx]->p_lckcnt != 0) 4143 ulckpgs++; 4144 } 4145 4146 if (ulckpgs != 0) { 4147 mutex_enter(&freemem_lock); 4148 availrmem += ulckpgs; 4149 pages_claimed -= ulckpgs; 4150 mutex_exit(&freemem_lock); 4151 } 4152 4153 for (pg_idx = 0; ppa[pg_idx] != NULL; pg_idx++) { 4154 ppa[pg_idx]->p_cowcnt--; 4155 ppa[pg_idx]->p_lckcnt++; 4156 4157 } 4158 mutex_exit(&page_llock); 4159 return (1); 4160 } 4161 4162 page_t * 4163 page_numtopp(pfn_t pfnum, se_t se) 4164 { 4165 page_t *pp; 4166 4167 retry: 4168 pp = page_numtopp_nolock(pfnum); 4169 if (pp == NULL) { 4170 return ((page_t *)NULL); 4171 } 4172 4173 /* 4174 * Acquire the appropriate lock on the page. 4175 */ 4176 while (!page_lock(pp, se, (kmutex_t *)NULL, P_RECLAIM)) { 4177 if (page_pptonum(pp) != pfnum) 4178 goto retry; 4179 continue; 4180 } 4181 4182 if (page_pptonum(pp) != pfnum) { 4183 page_unlock(pp); 4184 goto retry; 4185 } 4186 4187 return (pp); 4188 } 4189 4190 page_t * 4191 page_numtopp_noreclaim(pfn_t pfnum, se_t se) 4192 { 4193 page_t *pp; 4194 4195 retry: 4196 pp = page_numtopp_nolock(pfnum); 4197 if (pp == NULL) { 4198 return ((page_t *)NULL); 4199 } 4200 4201 /* 4202 * Acquire the appropriate lock on the page. 4203 */ 4204 while (!page_lock(pp, se, (kmutex_t *)NULL, P_NO_RECLAIM)) { 4205 if (page_pptonum(pp) != pfnum) 4206 goto retry; 4207 continue; 4208 } 4209 4210 if (page_pptonum(pp) != pfnum) { 4211 page_unlock(pp); 4212 goto retry; 4213 } 4214 4215 return (pp); 4216 } 4217 4218 /* 4219 * This routine is like page_numtopp, but will only return page structs 4220 * for pages which are ok for loading into hardware using the page struct. 4221 */ 4222 page_t * 4223 page_numtopp_nowait(pfn_t pfnum, se_t se) 4224 { 4225 page_t *pp; 4226 4227 retry: 4228 pp = page_numtopp_nolock(pfnum); 4229 if (pp == NULL) { 4230 return ((page_t *)NULL); 4231 } 4232 4233 /* 4234 * Try to acquire the appropriate lock on the page. 4235 */ 4236 if (PP_ISFREE(pp)) 4237 pp = NULL; 4238 else { 4239 if (!page_trylock(pp, se)) 4240 pp = NULL; 4241 else { 4242 if (page_pptonum(pp) != pfnum) { 4243 page_unlock(pp); 4244 goto retry; 4245 } 4246 if (PP_ISFREE(pp)) { 4247 page_unlock(pp); 4248 pp = NULL; 4249 } 4250 } 4251 } 4252 return (pp); 4253 } 4254 4255 /* 4256 * Returns a count of dirty pages that are in the process 4257 * of being written out. If 'cleanit' is set, try to push the page. 4258 */ 4259 pgcnt_t 4260 page_busy(int cleanit) 4261 { 4262 page_t *page0 = page_first(); 4263 page_t *pp = page0; 4264 pgcnt_t nppbusy = 0; 4265 u_offset_t off; 4266 4267 do { 4268 vnode_t *vp = pp->p_vnode; 4269 4270 /* 4271 * A page is a candidate for syncing if it is: 4272 * 4273 * (a) On neither the freelist nor the cachelist 4274 * (b) Hashed onto a vnode 4275 * (c) Not a kernel page 4276 * (d) Dirty 4277 * (e) Not part of a swapfile 4278 * (f) a page which belongs to a real vnode; eg has a non-null 4279 * v_vfsp pointer. 4280 * (g) Backed by a filesystem which doesn't have a 4281 * stubbed-out sync operation 4282 */ 4283 if (!PP_ISFREE(pp) && vp != NULL && !VN_ISKAS(vp) && 4284 hat_ismod(pp) && !IS_SWAPVP(vp) && vp->v_vfsp != NULL && 4285 vfs_can_sync(vp->v_vfsp)) { 4286 nppbusy++; 4287 vfs_syncprogress(); 4288 4289 if (!cleanit) 4290 continue; 4291 if (!page_trylock(pp, SE_EXCL)) 4292 continue; 4293 4294 if (PP_ISFREE(pp) || vp == NULL || IS_SWAPVP(vp) || 4295 pp->p_lckcnt != 0 || pp->p_cowcnt != 0 || 4296 !(hat_pagesync(pp, 4297 HAT_SYNC_DONTZERO | HAT_SYNC_STOPON_MOD) & P_MOD)) { 4298 page_unlock(pp); 4299 continue; 4300 } 4301 off = pp->p_offset; 4302 VN_HOLD(vp); 4303 page_unlock(pp); 4304 (void) VOP_PUTPAGE(vp, off, PAGESIZE, 4305 B_ASYNC | B_FREE, kcred, NULL); 4306 VN_RELE(vp); 4307 } 4308 } while ((pp = page_next(pp)) != page0); 4309 4310 return (nppbusy); 4311 } 4312 4313 void page_invalidate_pages(void); 4314 4315 /* 4316 * callback handler to vm sub-system 4317 * 4318 * callers make sure no recursive entries to this func. 4319 */ 4320 /*ARGSUSED*/ 4321 boolean_t 4322 callb_vm_cpr(void *arg, int code) 4323 { 4324 if (code == CB_CODE_CPR_CHKPT) 4325 page_invalidate_pages(); 4326 return (B_TRUE); 4327 } 4328 4329 /* 4330 * Invalidate all pages of the system. 4331 * It shouldn't be called until all user page activities are all stopped. 4332 */ 4333 void 4334 page_invalidate_pages() 4335 { 4336 page_t *pp; 4337 page_t *page0; 4338 pgcnt_t nbusypages; 4339 int retry = 0; 4340 const int MAXRETRIES = 4; 4341 #if defined(__sparc) 4342 extern struct vnode prom_ppages; 4343 #endif /* __sparc */ 4344 4345 top: 4346 /* 4347 * Flush dirty pages and destroy the clean ones. 4348 */ 4349 nbusypages = 0; 4350 4351 pp = page0 = page_first(); 4352 do { 4353 struct vnode *vp; 4354 u_offset_t offset; 4355 int mod; 4356 4357 /* 4358 * skip the page if it has no vnode or the page associated 4359 * with the kernel vnode or prom allocated kernel mem. 4360 */ 4361 #if defined(__sparc) 4362 if ((vp = pp->p_vnode) == NULL || VN_ISKAS(vp) || 4363 vp == &prom_ppages) 4364 #else /* x86 doesn't have prom or prom_ppage */ 4365 if ((vp = pp->p_vnode) == NULL || VN_ISKAS(vp)) 4366 #endif /* __sparc */ 4367 continue; 4368 4369 /* 4370 * skip the page which is already free invalidated. 4371 */ 4372 if (PP_ISFREE(pp) && PP_ISAGED(pp)) 4373 continue; 4374 4375 /* 4376 * skip pages that are already locked or can't be "exclusively" 4377 * locked or are already free. After we lock the page, check 4378 * the free and age bits again to be sure it's not destroied 4379 * yet. 4380 * To achieve max. parallelization, we use page_trylock instead 4381 * of page_lock so that we don't get block on individual pages 4382 * while we have thousands of other pages to process. 4383 */ 4384 if (!page_trylock(pp, SE_EXCL)) { 4385 nbusypages++; 4386 continue; 4387 } else if (PP_ISFREE(pp)) { 4388 if (!PP_ISAGED(pp)) { 4389 page_destroy_free(pp); 4390 } else { 4391 page_unlock(pp); 4392 } 4393 continue; 4394 } 4395 /* 4396 * Is this page involved in some I/O? shared? 4397 * 4398 * The page_struct_lock need not be acquired to 4399 * examine these fields since the page has an 4400 * "exclusive" lock. 4401 */ 4402 if (pp->p_lckcnt != 0 || pp->p_cowcnt != 0) { 4403 page_unlock(pp); 4404 continue; 4405 } 4406 4407 if (vp->v_type == VCHR) { 4408 panic("vp->v_type == VCHR"); 4409 /*NOTREACHED*/ 4410 } 4411 4412 if (!page_try_demote_pages(pp)) { 4413 page_unlock(pp); 4414 continue; 4415 } 4416 4417 /* 4418 * Check the modified bit. Leave the bits alone in hardware 4419 * (they will be modified if we do the putpage). 4420 */ 4421 mod = (hat_pagesync(pp, HAT_SYNC_DONTZERO | HAT_SYNC_STOPON_MOD) 4422 & P_MOD); 4423 if (mod) { 4424 offset = pp->p_offset; 4425 /* 4426 * Hold the vnode before releasing the page lock 4427 * to prevent it from being freed and re-used by 4428 * some other thread. 4429 */ 4430 VN_HOLD(vp); 4431 page_unlock(pp); 4432 /* 4433 * No error return is checked here. Callers such as 4434 * cpr deals with the dirty pages at the dump time 4435 * if this putpage fails. 4436 */ 4437 (void) VOP_PUTPAGE(vp, offset, PAGESIZE, B_INVAL, 4438 kcred, NULL); 4439 VN_RELE(vp); 4440 } else { 4441 page_destroy(pp, 0); 4442 } 4443 } while ((pp = page_next(pp)) != page0); 4444 if (nbusypages && retry++ < MAXRETRIES) { 4445 delay(1); 4446 goto top; 4447 } 4448 } 4449 4450 /* 4451 * Replace the page "old" with the page "new" on the page hash and vnode lists 4452 * 4453 * the replacement must be done in place, ie the equivalent sequence: 4454 * 4455 * vp = old->p_vnode; 4456 * off = old->p_offset; 4457 * page_do_hashout(old) 4458 * page_do_hashin(new, vp, off) 4459 * 4460 * doesn't work, since 4461 * 1) if old is the only page on the vnode, the v_pages list has a window 4462 * where it looks empty. This will break file system assumptions. 4463 * and 4464 * 2) pvn_vplist_dirty() can't deal with pages moving on the v_pages list. 4465 */ 4466 static void 4467 page_do_relocate_hash(page_t *new, page_t *old) 4468 { 4469 page_t **hash_list; 4470 vnode_t *vp = old->p_vnode; 4471 kmutex_t *sep; 4472 4473 ASSERT(PAGE_EXCL(old)); 4474 ASSERT(PAGE_EXCL(new)); 4475 ASSERT(vp != NULL); 4476 ASSERT(MUTEX_HELD(page_vnode_mutex(vp))); 4477 ASSERT(MUTEX_HELD(PAGE_HASH_MUTEX(PAGE_HASH_FUNC(vp, old->p_offset)))); 4478 4479 /* 4480 * First find old page on the page hash list 4481 */ 4482 hash_list = &page_hash[PAGE_HASH_FUNC(vp, old->p_offset)]; 4483 4484 for (;;) { 4485 if (*hash_list == old) 4486 break; 4487 if (*hash_list == NULL) { 4488 panic("page_do_hashout"); 4489 /*NOTREACHED*/ 4490 } 4491 hash_list = &(*hash_list)->p_hash; 4492 } 4493 4494 /* 4495 * update new and replace old with new on the page hash list 4496 */ 4497 new->p_vnode = old->p_vnode; 4498 new->p_offset = old->p_offset; 4499 new->p_hash = old->p_hash; 4500 *hash_list = new; 4501 4502 if ((new->p_vnode->v_flag & VISSWAP) != 0) 4503 PP_SETSWAP(new); 4504 4505 /* 4506 * replace old with new on the vnode's page list 4507 */ 4508 if (old->p_vpnext == old) { 4509 new->p_vpnext = new; 4510 new->p_vpprev = new; 4511 } else { 4512 new->p_vpnext = old->p_vpnext; 4513 new->p_vpprev = old->p_vpprev; 4514 new->p_vpnext->p_vpprev = new; 4515 new->p_vpprev->p_vpnext = new; 4516 } 4517 if (vp->v_pages == old) 4518 vp->v_pages = new; 4519 4520 /* 4521 * clear out the old page 4522 */ 4523 old->p_hash = NULL; 4524 old->p_vpnext = NULL; 4525 old->p_vpprev = NULL; 4526 old->p_vnode = NULL; 4527 PP_CLRSWAP(old); 4528 old->p_offset = (u_offset_t)-1; 4529 page_clr_all_props(old, 1); 4530 4531 /* 4532 * Wake up processes waiting for this page. The page's 4533 * identity has been changed, and is probably not the 4534 * desired page any longer. 4535 */ 4536 sep = page_se_mutex(old); 4537 mutex_enter(sep); 4538 old->p_selock &= ~SE_EWANTED; 4539 if (CV_HAS_WAITERS(&old->p_cv)) 4540 cv_broadcast(&old->p_cv); 4541 mutex_exit(sep); 4542 } 4543 4544 /* 4545 * This function moves the identity of page "pp_old" to page "pp_new". 4546 * Both pages must be locked on entry. "pp_new" is free, has no identity, 4547 * and need not be hashed out from anywhere. 4548 */ 4549 void 4550 page_relocate_hash(page_t *pp_new, page_t *pp_old) 4551 { 4552 vnode_t *vp = pp_old->p_vnode; 4553 u_offset_t off = pp_old->p_offset; 4554 kmutex_t *phm, *vphm; 4555 4556 /* 4557 * Rehash two pages 4558 */ 4559 ASSERT(PAGE_EXCL(pp_old)); 4560 ASSERT(PAGE_EXCL(pp_new)); 4561 ASSERT(vp != NULL); 4562 ASSERT(pp_new->p_vnode == NULL); 4563 4564 /* 4565 * hashout then hashin while holding the mutexes 4566 */ 4567 phm = PAGE_HASH_MUTEX(PAGE_HASH_FUNC(vp, off)); 4568 mutex_enter(phm); 4569 vphm = page_vnode_mutex(vp); 4570 mutex_enter(vphm); 4571 4572 page_do_relocate_hash(pp_new, pp_old); 4573 4574 mutex_exit(vphm); 4575 mutex_exit(phm); 4576 4577 /* 4578 * The page_struct_lock need not be acquired for lckcnt and 4579 * cowcnt since the page has an "exclusive" lock. 4580 */ 4581 ASSERT(pp_new->p_lckcnt == 0); 4582 ASSERT(pp_new->p_cowcnt == 0); 4583 pp_new->p_lckcnt = pp_old->p_lckcnt; 4584 pp_new->p_cowcnt = pp_old->p_cowcnt; 4585 pp_old->p_lckcnt = pp_old->p_cowcnt = 0; 4586 4587 /* The following comment preserved from page_flip(). */ 4588 /* XXX - Do we need to protect fsdata? */ 4589 pp_new->p_fsdata = pp_old->p_fsdata; 4590 } 4591 4592 /* 4593 * Helper routine used to lock all remaining members of a 4594 * large page. The caller is responsible for passing in a locked 4595 * pp. If pp is a large page, then it succeeds in locking all the 4596 * remaining constituent pages or it returns with only the 4597 * original page locked. 4598 * 4599 * Returns 1 on success, 0 on failure. 4600 * 4601 * If success is returned this routine guarantees p_szc for all constituent 4602 * pages of a large page pp belongs to can't change. To achieve this we 4603 * recheck szc of pp after locking all constituent pages and retry if szc 4604 * changed (it could only decrease). Since hat_page_demote() needs an EXCL 4605 * lock on one of constituent pages it can't be running after all constituent 4606 * pages are locked. hat_page_demote() with a lock on a constituent page 4607 * outside of this large page (i.e. pp belonged to a larger large page) is 4608 * already done with all constituent pages of pp since the root's p_szc is 4609 * changed last. Therefore no need to synchronize with hat_page_demote() that 4610 * locked a constituent page outside of pp's current large page. 4611 */ 4612 #ifdef DEBUG 4613 uint32_t gpg_trylock_mtbf = 0; 4614 #endif 4615 4616 int 4617 group_page_trylock(page_t *pp, se_t se) 4618 { 4619 page_t *tpp; 4620 pgcnt_t npgs, i, j; 4621 uint_t pszc = pp->p_szc; 4622 4623 #ifdef DEBUG 4624 if (gpg_trylock_mtbf && !(gethrtime() % gpg_trylock_mtbf)) { 4625 return (0); 4626 } 4627 #endif 4628 4629 if (pp != PP_GROUPLEADER(pp, pszc)) { 4630 return (0); 4631 } 4632 4633 retry: 4634 ASSERT(PAGE_LOCKED_SE(pp, se)); 4635 ASSERT(!PP_ISFREE(pp)); 4636 if (pszc == 0) { 4637 return (1); 4638 } 4639 npgs = page_get_pagecnt(pszc); 4640 tpp = pp + 1; 4641 for (i = 1; i < npgs; i++, tpp++) { 4642 if (!page_trylock(tpp, se)) { 4643 tpp = pp + 1; 4644 for (j = 1; j < i; j++, tpp++) { 4645 page_unlock(tpp); 4646 } 4647 return (0); 4648 } 4649 } 4650 if (pp->p_szc != pszc) { 4651 ASSERT(pp->p_szc < pszc); 4652 ASSERT(pp->p_vnode != NULL && !PP_ISKAS(pp) && 4653 !IS_SWAPFSVP(pp->p_vnode)); 4654 tpp = pp + 1; 4655 for (i = 1; i < npgs; i++, tpp++) { 4656 page_unlock(tpp); 4657 } 4658 pszc = pp->p_szc; 4659 goto retry; 4660 } 4661 return (1); 4662 } 4663 4664 void 4665 group_page_unlock(page_t *pp) 4666 { 4667 page_t *tpp; 4668 pgcnt_t npgs, i; 4669 4670 ASSERT(PAGE_LOCKED(pp)); 4671 ASSERT(!PP_ISFREE(pp)); 4672 ASSERT(pp == PP_PAGEROOT(pp)); 4673 npgs = page_get_pagecnt(pp->p_szc); 4674 for (i = 1, tpp = pp + 1; i < npgs; i++, tpp++) { 4675 page_unlock(tpp); 4676 } 4677 } 4678 4679 /* 4680 * returns 4681 * 0 : on success and *nrelocp is number of relocated PAGESIZE pages 4682 * ERANGE : this is not a base page 4683 * EBUSY : failure to get locks on the page/pages 4684 * ENOMEM : failure to obtain replacement pages 4685 * EAGAIN : OBP has not yet completed its boot-time handoff to the kernel 4686 * EIO : An error occurred while trying to copy the page data 4687 * 4688 * Return with all constituent members of target and replacement 4689 * SE_EXCL locked. It is the callers responsibility to drop the 4690 * locks. 4691 */ 4692 int 4693 do_page_relocate( 4694 page_t **target, 4695 page_t **replacement, 4696 int grouplock, 4697 spgcnt_t *nrelocp, 4698 lgrp_t *lgrp) 4699 { 4700 page_t *first_repl; 4701 page_t *repl; 4702 page_t *targ; 4703 page_t *pl = NULL; 4704 uint_t ppattr; 4705 pfn_t pfn, repl_pfn; 4706 uint_t szc; 4707 spgcnt_t npgs, i; 4708 int repl_contig = 0; 4709 uint_t flags = 0; 4710 spgcnt_t dofree = 0; 4711 4712 *nrelocp = 0; 4713 4714 #if defined(__sparc) 4715 /* 4716 * We need to wait till OBP has completed 4717 * its boot-time handoff of its resources to the kernel 4718 * before we allow page relocation 4719 */ 4720 if (page_relocate_ready == 0) { 4721 return (EAGAIN); 4722 } 4723 #endif 4724 4725 /* 4726 * If this is not a base page, 4727 * just return with 0x0 pages relocated. 4728 */ 4729 targ = *target; 4730 ASSERT(PAGE_EXCL(targ)); 4731 ASSERT(!PP_ISFREE(targ)); 4732 szc = targ->p_szc; 4733 ASSERT(szc < mmu_page_sizes); 4734 VM_STAT_ADD(vmm_vmstats.ppr_reloc[szc]); 4735 pfn = targ->p_pagenum; 4736 if (pfn != PFN_BASE(pfn, szc)) { 4737 VM_STAT_ADD(vmm_vmstats.ppr_relocnoroot[szc]); 4738 return (ERANGE); 4739 } 4740 4741 if ((repl = *replacement) != NULL && repl->p_szc >= szc) { 4742 repl_pfn = repl->p_pagenum; 4743 if (repl_pfn != PFN_BASE(repl_pfn, szc)) { 4744 VM_STAT_ADD(vmm_vmstats.ppr_reloc_replnoroot[szc]); 4745 return (ERANGE); 4746 } 4747 repl_contig = 1; 4748 } 4749 4750 /* 4751 * We must lock all members of this large page or we cannot 4752 * relocate any part of it. 4753 */ 4754 if (grouplock != 0 && !group_page_trylock(targ, SE_EXCL)) { 4755 VM_STAT_ADD(vmm_vmstats.ppr_relocnolock[targ->p_szc]); 4756 return (EBUSY); 4757 } 4758 4759 /* 4760 * reread szc it could have been decreased before 4761 * group_page_trylock() was done. 4762 */ 4763 szc = targ->p_szc; 4764 ASSERT(szc < mmu_page_sizes); 4765 VM_STAT_ADD(vmm_vmstats.ppr_reloc[szc]); 4766 ASSERT(pfn == PFN_BASE(pfn, szc)); 4767 4768 npgs = page_get_pagecnt(targ->p_szc); 4769 4770 if (repl == NULL) { 4771 dofree = npgs; /* Size of target page in MMU pages */ 4772 if (!page_create_wait(dofree, 0)) { 4773 if (grouplock != 0) { 4774 group_page_unlock(targ); 4775 } 4776 VM_STAT_ADD(vmm_vmstats.ppr_relocnomem[szc]); 4777 return (ENOMEM); 4778 } 4779 4780 /* 4781 * seg kmem pages require that the target and replacement 4782 * page be the same pagesize. 4783 */ 4784 flags = (VN_ISKAS(targ->p_vnode)) ? PGR_SAMESZC : 0; 4785 repl = page_get_replacement_page(targ, lgrp, flags); 4786 if (repl == NULL) { 4787 if (grouplock != 0) { 4788 group_page_unlock(targ); 4789 } 4790 page_create_putback(dofree); 4791 VM_STAT_ADD(vmm_vmstats.ppr_relocnomem[szc]); 4792 return (ENOMEM); 4793 } 4794 } 4795 #ifdef DEBUG 4796 else { 4797 ASSERT(PAGE_LOCKED(repl)); 4798 } 4799 #endif /* DEBUG */ 4800 4801 #if defined(__sparc) 4802 /* 4803 * Let hat_page_relocate() complete the relocation if it's kernel page 4804 */ 4805 if (VN_ISKAS(targ->p_vnode)) { 4806 *replacement = repl; 4807 if (hat_page_relocate(target, replacement, nrelocp) != 0) { 4808 if (grouplock != 0) { 4809 group_page_unlock(targ); 4810 } 4811 if (dofree) { 4812 *replacement = NULL; 4813 page_free_replacement_page(repl); 4814 page_create_putback(dofree); 4815 } 4816 VM_STAT_ADD(vmm_vmstats.ppr_krelocfail[szc]); 4817 return (EAGAIN); 4818 } 4819 VM_STAT_ADD(vmm_vmstats.ppr_relocok[szc]); 4820 return (0); 4821 } 4822 #else 4823 #if defined(lint) 4824 dofree = dofree; 4825 #endif 4826 #endif 4827 4828 first_repl = repl; 4829 4830 for (i = 0; i < npgs; i++) { 4831 ASSERT(PAGE_EXCL(targ)); 4832 ASSERT(targ->p_slckcnt == 0); 4833 ASSERT(repl->p_slckcnt == 0); 4834 4835 (void) hat_pageunload(targ, HAT_FORCE_PGUNLOAD); 4836 4837 ASSERT(hat_page_getshare(targ) == 0); 4838 ASSERT(!PP_ISFREE(targ)); 4839 ASSERT(targ->p_pagenum == (pfn + i)); 4840 ASSERT(repl_contig == 0 || 4841 repl->p_pagenum == (repl_pfn + i)); 4842 4843 /* 4844 * Copy the page contents and attributes then 4845 * relocate the page in the page hash. 4846 */ 4847 if (ppcopy(targ, repl) == 0) { 4848 targ = *target; 4849 repl = first_repl; 4850 VM_STAT_ADD(vmm_vmstats.ppr_copyfail); 4851 if (grouplock != 0) { 4852 group_page_unlock(targ); 4853 } 4854 if (dofree) { 4855 *replacement = NULL; 4856 page_free_replacement_page(repl); 4857 page_create_putback(dofree); 4858 } 4859 return (EIO); 4860 } 4861 4862 targ++; 4863 if (repl_contig != 0) { 4864 repl++; 4865 } else { 4866 repl = repl->p_next; 4867 } 4868 } 4869 4870 repl = first_repl; 4871 targ = *target; 4872 4873 for (i = 0; i < npgs; i++) { 4874 ppattr = hat_page_getattr(targ, (P_MOD | P_REF | P_RO)); 4875 page_clr_all_props(repl, 0); 4876 page_set_props(repl, ppattr); 4877 page_relocate_hash(repl, targ); 4878 4879 ASSERT(hat_page_getshare(targ) == 0); 4880 ASSERT(hat_page_getshare(repl) == 0); 4881 /* 4882 * Now clear the props on targ, after the 4883 * page_relocate_hash(), they no longer 4884 * have any meaning. 4885 */ 4886 page_clr_all_props(targ, 0); 4887 ASSERT(targ->p_next == targ); 4888 ASSERT(targ->p_prev == targ); 4889 page_list_concat(&pl, &targ); 4890 4891 targ++; 4892 if (repl_contig != 0) { 4893 repl++; 4894 } else { 4895 repl = repl->p_next; 4896 } 4897 } 4898 /* assert that we have come full circle with repl */ 4899 ASSERT(repl_contig == 1 || first_repl == repl); 4900 4901 *target = pl; 4902 if (*replacement == NULL) { 4903 ASSERT(first_repl == repl); 4904 *replacement = repl; 4905 } 4906 VM_STAT_ADD(vmm_vmstats.ppr_relocok[szc]); 4907 *nrelocp = npgs; 4908 return (0); 4909 } 4910 /* 4911 * On success returns 0 and *nrelocp the number of PAGESIZE pages relocated. 4912 */ 4913 int 4914 page_relocate( 4915 page_t **target, 4916 page_t **replacement, 4917 int grouplock, 4918 int freetarget, 4919 spgcnt_t *nrelocp, 4920 lgrp_t *lgrp) 4921 { 4922 spgcnt_t ret; 4923 4924 /* do_page_relocate returns 0 on success or errno value */ 4925 ret = do_page_relocate(target, replacement, grouplock, nrelocp, lgrp); 4926 4927 if (ret != 0 || freetarget == 0) { 4928 return (ret); 4929 } 4930 if (*nrelocp == 1) { 4931 ASSERT(*target != NULL); 4932 page_free(*target, 1); 4933 } else { 4934 page_t *tpp = *target; 4935 uint_t szc = tpp->p_szc; 4936 pgcnt_t npgs = page_get_pagecnt(szc); 4937 ASSERT(npgs > 1); 4938 ASSERT(szc != 0); 4939 do { 4940 ASSERT(PAGE_EXCL(tpp)); 4941 ASSERT(!hat_page_is_mapped(tpp)); 4942 ASSERT(tpp->p_szc == szc); 4943 PP_SETFREE(tpp); 4944 PP_SETAGED(tpp); 4945 npgs--; 4946 } while ((tpp = tpp->p_next) != *target); 4947 ASSERT(npgs == 0); 4948 page_list_add_pages(*target, 0); 4949 npgs = page_get_pagecnt(szc); 4950 page_create_putback(npgs); 4951 } 4952 return (ret); 4953 } 4954 4955 /* 4956 * it is up to the caller to deal with pcf accounting. 4957 */ 4958 void 4959 page_free_replacement_page(page_t *pplist) 4960 { 4961 page_t *pp; 4962 4963 while (pplist != NULL) { 4964 /* 4965 * pp_targ is a linked list. 4966 */ 4967 pp = pplist; 4968 if (pp->p_szc == 0) { 4969 page_sub(&pplist, pp); 4970 page_clr_all_props(pp, 0); 4971 PP_SETFREE(pp); 4972 PP_SETAGED(pp); 4973 page_list_add(pp, PG_FREE_LIST | PG_LIST_TAIL); 4974 page_unlock(pp); 4975 VM_STAT_ADD(pagecnt.pc_free_replacement_page[0]); 4976 } else { 4977 spgcnt_t curnpgs = page_get_pagecnt(pp->p_szc); 4978 page_t *tpp; 4979 page_list_break(&pp, &pplist, curnpgs); 4980 tpp = pp; 4981 do { 4982 ASSERT(PAGE_EXCL(tpp)); 4983 ASSERT(!hat_page_is_mapped(tpp)); 4984 page_clr_all_props(tpp, 0); 4985 PP_SETFREE(tpp); 4986 PP_SETAGED(tpp); 4987 } while ((tpp = tpp->p_next) != pp); 4988 page_list_add_pages(pp, 0); 4989 VM_STAT_ADD(pagecnt.pc_free_replacement_page[1]); 4990 } 4991 } 4992 } 4993 4994 /* 4995 * Relocate target to non-relocatable replacement page. 4996 */ 4997 int 4998 page_relocate_cage(page_t **target, page_t **replacement) 4999 { 5000 page_t *tpp, *rpp; 5001 spgcnt_t pgcnt, npgs; 5002 int result; 5003 5004 tpp = *target; 5005 5006 ASSERT(PAGE_EXCL(tpp)); 5007 ASSERT(tpp->p_szc == 0); 5008 5009 pgcnt = btop(page_get_pagesize(tpp->p_szc)); 5010 5011 do { 5012 (void) page_create_wait(pgcnt, PG_WAIT | PG_NORELOC); 5013 rpp = page_get_replacement_page(tpp, NULL, PGR_NORELOC); 5014 if (rpp == NULL) { 5015 page_create_putback(pgcnt); 5016 kcage_cageout_wakeup(); 5017 } 5018 } while (rpp == NULL); 5019 5020 ASSERT(PP_ISNORELOC(rpp)); 5021 5022 result = page_relocate(&tpp, &rpp, 0, 1, &npgs, NULL); 5023 5024 if (result == 0) { 5025 *replacement = rpp; 5026 if (pgcnt != npgs) 5027 panic("page_relocate_cage: partial relocation"); 5028 } 5029 5030 return (result); 5031 } 5032 5033 /* 5034 * Release the page lock on a page, place on cachelist 5035 * tail if no longer mapped. Caller can let us know if 5036 * the page is known to be clean. 5037 */ 5038 int 5039 page_release(page_t *pp, int checkmod) 5040 { 5041 int status; 5042 5043 ASSERT(PAGE_LOCKED(pp) && !PP_ISFREE(pp) && 5044 (pp->p_vnode != NULL)); 5045 5046 if (!hat_page_is_mapped(pp) && !IS_SWAPVP(pp->p_vnode) && 5047 ((PAGE_SHARED(pp) && page_tryupgrade(pp)) || PAGE_EXCL(pp)) && 5048 pp->p_lckcnt == 0 && pp->p_cowcnt == 0 && 5049 !hat_page_is_mapped(pp)) { 5050 5051 /* 5052 * If page is modified, unlock it 5053 * 5054 * (p_nrm & P_MOD) bit has the latest stuff because: 5055 * (1) We found that this page doesn't have any mappings 5056 * _after_ holding SE_EXCL and 5057 * (2) We didn't drop SE_EXCL lock after the check in (1) 5058 */ 5059 if (checkmod && hat_ismod(pp)) { 5060 page_unlock(pp); 5061 status = PGREL_MOD; 5062 } else { 5063 /*LINTED: constant in conditional context*/ 5064 VN_DISPOSE(pp, B_FREE, 0, kcred); 5065 status = PGREL_CLEAN; 5066 } 5067 } else { 5068 page_unlock(pp); 5069 status = PGREL_NOTREL; 5070 } 5071 return (status); 5072 } 5073 5074 /* 5075 * Given a constituent page, try to demote the large page on the freelist. 5076 * 5077 * Returns nonzero if the page could be demoted successfully. Returns with 5078 * the constituent page still locked. 5079 */ 5080 int 5081 page_try_demote_free_pages(page_t *pp) 5082 { 5083 page_t *rootpp = pp; 5084 pfn_t pfn = page_pptonum(pp); 5085 spgcnt_t npgs; 5086 uint_t szc = pp->p_szc; 5087 5088 ASSERT(PP_ISFREE(pp)); 5089 ASSERT(PAGE_EXCL(pp)); 5090 5091 /* 5092 * Adjust rootpp and lock it, if `pp' is not the base 5093 * constituent page. 5094 */ 5095 npgs = page_get_pagecnt(pp->p_szc); 5096 if (npgs == 1) { 5097 return (0); 5098 } 5099 5100 if (!IS_P2ALIGNED(pfn, npgs)) { 5101 pfn = P2ALIGN(pfn, npgs); 5102 rootpp = page_numtopp_nolock(pfn); 5103 } 5104 5105 if (pp != rootpp && !page_trylock(rootpp, SE_EXCL)) { 5106 return (0); 5107 } 5108 5109 if (rootpp->p_szc != szc) { 5110 if (pp != rootpp) 5111 page_unlock(rootpp); 5112 return (0); 5113 } 5114 5115 page_demote_free_pages(rootpp); 5116 5117 if (pp != rootpp) 5118 page_unlock(rootpp); 5119 5120 ASSERT(PP_ISFREE(pp)); 5121 ASSERT(PAGE_EXCL(pp)); 5122 return (1); 5123 } 5124 5125 /* 5126 * Given a constituent page, try to demote the large page. 5127 * 5128 * Returns nonzero if the page could be demoted successfully. Returns with 5129 * the constituent page still locked. 5130 */ 5131 int 5132 page_try_demote_pages(page_t *pp) 5133 { 5134 page_t *tpp, *rootpp = pp; 5135 pfn_t pfn = page_pptonum(pp); 5136 spgcnt_t i, npgs; 5137 uint_t szc = pp->p_szc; 5138 vnode_t *vp = pp->p_vnode; 5139 5140 ASSERT(PAGE_EXCL(pp)); 5141 5142 VM_STAT_ADD(pagecnt.pc_try_demote_pages[0]); 5143 5144 if (pp->p_szc == 0) { 5145 VM_STAT_ADD(pagecnt.pc_try_demote_pages[1]); 5146 return (1); 5147 } 5148 5149 if (vp != NULL && !IS_SWAPFSVP(vp) && !VN_ISKAS(vp)) { 5150 VM_STAT_ADD(pagecnt.pc_try_demote_pages[2]); 5151 page_demote_vp_pages(pp); 5152 ASSERT(pp->p_szc == 0); 5153 return (1); 5154 } 5155 5156 /* 5157 * Adjust rootpp if passed in is not the base 5158 * constituent page. 5159 */ 5160 npgs = page_get_pagecnt(pp->p_szc); 5161 ASSERT(npgs > 1); 5162 if (!IS_P2ALIGNED(pfn, npgs)) { 5163 pfn = P2ALIGN(pfn, npgs); 5164 rootpp = page_numtopp_nolock(pfn); 5165 VM_STAT_ADD(pagecnt.pc_try_demote_pages[3]); 5166 ASSERT(rootpp->p_vnode != NULL); 5167 ASSERT(rootpp->p_szc == szc); 5168 } 5169 5170 /* 5171 * We can't demote kernel pages since we can't hat_unload() 5172 * the mappings. 5173 */ 5174 if (VN_ISKAS(rootpp->p_vnode)) 5175 return (0); 5176 5177 /* 5178 * Attempt to lock all constituent pages except the page passed 5179 * in since it's already locked. 5180 */ 5181 for (tpp = rootpp, i = 0; i < npgs; i++, tpp++) { 5182 ASSERT(!PP_ISFREE(tpp)); 5183 ASSERT(tpp->p_vnode != NULL); 5184 5185 if (tpp != pp && !page_trylock(tpp, SE_EXCL)) 5186 break; 5187 ASSERT(tpp->p_szc == rootpp->p_szc); 5188 ASSERT(page_pptonum(tpp) == page_pptonum(rootpp) + i); 5189 } 5190 5191 /* 5192 * If we failed to lock them all then unlock what we have 5193 * locked so far and bail. 5194 */ 5195 if (i < npgs) { 5196 tpp = rootpp; 5197 while (i-- > 0) { 5198 if (tpp != pp) 5199 page_unlock(tpp); 5200 tpp++; 5201 } 5202 VM_STAT_ADD(pagecnt.pc_try_demote_pages[4]); 5203 return (0); 5204 } 5205 5206 for (tpp = rootpp, i = 0; i < npgs; i++, tpp++) { 5207 ASSERT(PAGE_EXCL(tpp)); 5208 ASSERT(tpp->p_slckcnt == 0); 5209 (void) hat_pageunload(tpp, HAT_FORCE_PGUNLOAD); 5210 tpp->p_szc = 0; 5211 } 5212 5213 /* 5214 * Unlock all pages except the page passed in. 5215 */ 5216 for (tpp = rootpp, i = 0; i < npgs; i++, tpp++) { 5217 ASSERT(!hat_page_is_mapped(tpp)); 5218 if (tpp != pp) 5219 page_unlock(tpp); 5220 } 5221 5222 VM_STAT_ADD(pagecnt.pc_try_demote_pages[5]); 5223 return (1); 5224 } 5225 5226 /* 5227 * Called by page_free() and page_destroy() to demote the page size code 5228 * (p_szc) to 0 (since we can't just put a single PAGESIZE page with non zero 5229 * p_szc on free list, neither can we just clear p_szc of a single page_t 5230 * within a large page since it will break other code that relies on p_szc 5231 * being the same for all page_t's of a large page). Anonymous pages should 5232 * never end up here because anon_map_getpages() cannot deal with p_szc 5233 * changes after a single constituent page is locked. While anonymous or 5234 * kernel large pages are demoted or freed the entire large page at a time 5235 * with all constituent pages locked EXCL for the file system pages we 5236 * have to be able to demote a large page (i.e. decrease all constituent pages 5237 * p_szc) with only just an EXCL lock on one of constituent pages. The reason 5238 * we can easily deal with anonymous page demotion the entire large page at a 5239 * time is that those operation originate at address space level and concern 5240 * the entire large page region with actual demotion only done when pages are 5241 * not shared with any other processes (therefore we can always get EXCL lock 5242 * on all anonymous constituent pages after clearing segment page 5243 * cache). However file system pages can be truncated or invalidated at a 5244 * PAGESIZE level from the file system side and end up in page_free() or 5245 * page_destroy() (we also allow only part of the large page to be SOFTLOCKed 5246 * and therefore pageout should be able to demote a large page by EXCL locking 5247 * any constituent page that is not under SOFTLOCK). In those cases we cannot 5248 * rely on being able to lock EXCL all constituent pages. 5249 * 5250 * To prevent szc changes on file system pages one has to lock all constituent 5251 * pages at least SHARED (or call page_szc_lock()). The only subsystem that 5252 * doesn't rely on locking all constituent pages (or using page_szc_lock()) to 5253 * prevent szc changes is hat layer that uses its own page level mlist 5254 * locks. hat assumes that szc doesn't change after mlist lock for a page is 5255 * taken. Therefore we need to change szc under hat level locks if we only 5256 * have an EXCL lock on a single constituent page and hat still references any 5257 * of constituent pages. (Note we can't "ignore" hat layer by simply 5258 * hat_pageunload() all constituent pages without having EXCL locks on all of 5259 * constituent pages). We use hat_page_demote() call to safely demote szc of 5260 * all constituent pages under hat locks when we only have an EXCL lock on one 5261 * of constituent pages. 5262 * 5263 * This routine calls page_szc_lock() before calling hat_page_demote() to 5264 * allow segvn in one special case not to lock all constituent pages SHARED 5265 * before calling hat_memload_array() that relies on p_szc not changing even 5266 * before hat level mlist lock is taken. In that case segvn uses 5267 * page_szc_lock() to prevent hat_page_demote() changing p_szc values. 5268 * 5269 * Anonymous or kernel page demotion still has to lock all pages exclusively 5270 * and do hat_pageunload() on all constituent pages before demoting the page 5271 * therefore there's no need for anonymous or kernel page demotion to use 5272 * hat_page_demote() mechanism. 5273 * 5274 * hat_page_demote() removes all large mappings that map pp and then decreases 5275 * p_szc starting from the last constituent page of the large page. By working 5276 * from the tail of a large page in pfn decreasing order allows one looking at 5277 * the root page to know that hat_page_demote() is done for root's szc area. 5278 * e.g. if a root page has szc 1 one knows it only has to lock all constituent 5279 * pages within szc 1 area to prevent szc changes because hat_page_demote() 5280 * that started on this page when it had szc > 1 is done for this szc 1 area. 5281 * 5282 * We are guaranteed that all constituent pages of pp's large page belong to 5283 * the same vnode with the consecutive offsets increasing in the direction of 5284 * the pfn i.e. the identity of constituent pages can't change until their 5285 * p_szc is decreased. Therefore it's safe for hat_page_demote() to remove 5286 * large mappings to pp even though we don't lock any constituent page except 5287 * pp (i.e. we won't unload e.g. kernel locked page). 5288 */ 5289 static void 5290 page_demote_vp_pages(page_t *pp) 5291 { 5292 kmutex_t *mtx; 5293 5294 ASSERT(PAGE_EXCL(pp)); 5295 ASSERT(!PP_ISFREE(pp)); 5296 ASSERT(pp->p_vnode != NULL); 5297 ASSERT(!IS_SWAPFSVP(pp->p_vnode)); 5298 ASSERT(!PP_ISKAS(pp)); 5299 5300 VM_STAT_ADD(pagecnt.pc_demote_pages[0]); 5301 5302 mtx = page_szc_lock(pp); 5303 if (mtx != NULL) { 5304 hat_page_demote(pp); 5305 mutex_exit(mtx); 5306 } 5307 ASSERT(pp->p_szc == 0); 5308 } 5309 5310 /* 5311 * Mark any existing pages for migration in the given range 5312 */ 5313 void 5314 page_mark_migrate(struct seg *seg, caddr_t addr, size_t len, 5315 struct anon_map *amp, ulong_t anon_index, vnode_t *vp, 5316 u_offset_t vnoff, int rflag) 5317 { 5318 struct anon *ap; 5319 vnode_t *curvp; 5320 lgrp_t *from; 5321 pgcnt_t i; 5322 pgcnt_t nlocked; 5323 u_offset_t off; 5324 pfn_t pfn; 5325 size_t pgsz; 5326 size_t segpgsz; 5327 pgcnt_t pages; 5328 uint_t pszc; 5329 page_t **ppa; 5330 pgcnt_t ppa_nentries; 5331 page_t *pp; 5332 caddr_t va; 5333 ulong_t an_idx; 5334 anon_sync_obj_t cookie; 5335 5336 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock)); 5337 5338 /* 5339 * Don't do anything if don't need to do lgroup optimizations 5340 * on this system 5341 */ 5342 if (!lgrp_optimizations()) 5343 return; 5344 5345 /* 5346 * Align address and length to (potentially large) page boundary 5347 */ 5348 segpgsz = page_get_pagesize(seg->s_szc); 5349 addr = (caddr_t)P2ALIGN((uintptr_t)addr, segpgsz); 5350 if (rflag) 5351 len = P2ROUNDUP(len, segpgsz); 5352 5353 /* 5354 * Allocate page array to accommodate largest page size 5355 */ 5356 pgsz = page_get_pagesize(page_num_pagesizes() - 1); 5357 ppa_nentries = btop(pgsz); 5358 ppa = kmem_zalloc(ppa_nentries * sizeof (page_t *), KM_SLEEP); 5359 5360 /* 5361 * Do one (large) page at a time 5362 */ 5363 va = addr; 5364 while (va < addr + len) { 5365 /* 5366 * Lookup (root) page for vnode and offset corresponding to 5367 * this virtual address 5368 * Try anonmap first since there may be copy-on-write 5369 * pages, but initialize vnode pointer and offset using 5370 * vnode arguments just in case there isn't an amp. 5371 */ 5372 curvp = vp; 5373 off = vnoff + va - seg->s_base; 5374 if (amp) { 5375 ANON_LOCK_ENTER(&->a_rwlock, RW_READER); 5376 an_idx = anon_index + seg_page(seg, va); 5377 anon_array_enter(amp, an_idx, &cookie); 5378 ap = anon_get_ptr(amp->ahp, an_idx); 5379 if (ap) 5380 swap_xlate(ap, &curvp, &off); 5381 anon_array_exit(&cookie); 5382 ANON_LOCK_EXIT(&->a_rwlock); 5383 } 5384 5385 pp = NULL; 5386 if (curvp) 5387 pp = page_lookup(curvp, off, SE_SHARED); 5388 5389 /* 5390 * If there isn't a page at this virtual address, 5391 * skip to next page 5392 */ 5393 if (pp == NULL) { 5394 va += PAGESIZE; 5395 continue; 5396 } 5397 5398 /* 5399 * Figure out which lgroup this page is in for kstats 5400 */ 5401 pfn = page_pptonum(pp); 5402 from = lgrp_pfn_to_lgrp(pfn); 5403 5404 /* 5405 * Get page size, and round up and skip to next page boundary 5406 * if unaligned address 5407 */ 5408 pszc = pp->p_szc; 5409 pgsz = page_get_pagesize(pszc); 5410 pages = btop(pgsz); 5411 if (!IS_P2ALIGNED(va, pgsz) || 5412 !IS_P2ALIGNED(pfn, pages) || 5413 pgsz > segpgsz) { 5414 pgsz = MIN(pgsz, segpgsz); 5415 page_unlock(pp); 5416 i = btop(P2END((uintptr_t)va, pgsz) - 5417 (uintptr_t)va); 5418 va = (caddr_t)P2END((uintptr_t)va, pgsz); 5419 lgrp_stat_add(from->lgrp_id, LGRP_PMM_FAIL_PGS, i); 5420 continue; 5421 } 5422 5423 /* 5424 * Upgrade to exclusive lock on page 5425 */ 5426 if (!page_tryupgrade(pp)) { 5427 page_unlock(pp); 5428 va += pgsz; 5429 lgrp_stat_add(from->lgrp_id, LGRP_PMM_FAIL_PGS, 5430 btop(pgsz)); 5431 continue; 5432 } 5433 5434 /* 5435 * Remember pages locked exclusively and how many 5436 */ 5437 ppa[0] = pp; 5438 nlocked = 1; 5439 5440 /* 5441 * Lock constituent pages if this is large page 5442 */ 5443 if (pages > 1) { 5444 /* 5445 * Lock all constituents except root page, since it 5446 * should be locked already. 5447 */ 5448 for (i = 1; i < pages; i++) { 5449 pp++; 5450 if (!page_trylock(pp, SE_EXCL)) { 5451 break; 5452 } 5453 if (PP_ISFREE(pp) || 5454 pp->p_szc != pszc) { 5455 /* 5456 * hat_page_demote() raced in with us. 5457 */ 5458 ASSERT(!IS_SWAPFSVP(curvp)); 5459 page_unlock(pp); 5460 break; 5461 } 5462 ppa[nlocked] = pp; 5463 nlocked++; 5464 } 5465 } 5466 5467 /* 5468 * If all constituent pages couldn't be locked, 5469 * unlock pages locked so far and skip to next page. 5470 */ 5471 if (nlocked != pages) { 5472 for (i = 0; i < nlocked; i++) 5473 page_unlock(ppa[i]); 5474 va += pgsz; 5475 lgrp_stat_add(from->lgrp_id, LGRP_PMM_FAIL_PGS, 5476 btop(pgsz)); 5477 continue; 5478 } 5479 5480 /* 5481 * hat_page_demote() can no longer happen 5482 * since last cons page had the right p_szc after 5483 * all cons pages were locked. all cons pages 5484 * should now have the same p_szc. 5485 */ 5486 5487 /* 5488 * All constituent pages locked successfully, so mark 5489 * large page for migration and unload the mappings of 5490 * constituent pages, so a fault will occur on any part of the 5491 * large page 5492 */ 5493 PP_SETMIGRATE(ppa[0]); 5494 for (i = 0; i < nlocked; i++) { 5495 pp = ppa[i]; 5496 (void) hat_pageunload(pp, HAT_FORCE_PGUNLOAD); 5497 ASSERT(hat_page_getshare(pp) == 0); 5498 page_unlock(pp); 5499 } 5500 lgrp_stat_add(from->lgrp_id, LGRP_PMM_PGS, nlocked); 5501 5502 va += pgsz; 5503 } 5504 kmem_free(ppa, ppa_nentries * sizeof (page_t *)); 5505 } 5506 5507 /* 5508 * Migrate any pages that have been marked for migration in the given range 5509 */ 5510 void 5511 page_migrate( 5512 struct seg *seg, 5513 caddr_t addr, 5514 page_t **ppa, 5515 pgcnt_t npages) 5516 { 5517 lgrp_t *from; 5518 lgrp_t *to; 5519 page_t *newpp; 5520 page_t *pp; 5521 pfn_t pfn; 5522 size_t pgsz; 5523 spgcnt_t page_cnt; 5524 spgcnt_t i; 5525 uint_t pszc; 5526 5527 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock)); 5528 5529 while (npages > 0) { 5530 pp = *ppa; 5531 pszc = pp->p_szc; 5532 pgsz = page_get_pagesize(pszc); 5533 page_cnt = btop(pgsz); 5534 5535 /* 5536 * Check to see whether this page is marked for migration 5537 * 5538 * Assume that root page of large page is marked for 5539 * migration and none of the other constituent pages 5540 * are marked. This really simplifies clearing the 5541 * migrate bit by not having to clear it from each 5542 * constituent page. 5543 * 5544 * note we don't want to relocate an entire large page if 5545 * someone is only using one subpage. 5546 */ 5547 if (npages < page_cnt) 5548 break; 5549 5550 /* 5551 * Is it marked for migration? 5552 */ 5553 if (!PP_ISMIGRATE(pp)) 5554 goto next; 5555 5556 /* 5557 * Determine lgroups that page is being migrated between 5558 */ 5559 pfn = page_pptonum(pp); 5560 if (!IS_P2ALIGNED(pfn, page_cnt)) { 5561 break; 5562 } 5563 from = lgrp_pfn_to_lgrp(pfn); 5564 to = lgrp_mem_choose(seg, addr, pgsz); 5565 5566 /* 5567 * Need to get exclusive lock's to migrate 5568 */ 5569 for (i = 0; i < page_cnt; i++) { 5570 ASSERT(PAGE_LOCKED(ppa[i])); 5571 if (page_pptonum(ppa[i]) != pfn + i || 5572 ppa[i]->p_szc != pszc) { 5573 break; 5574 } 5575 if (!page_tryupgrade(ppa[i])) { 5576 lgrp_stat_add(from->lgrp_id, 5577 LGRP_PM_FAIL_LOCK_PGS, 5578 page_cnt); 5579 break; 5580 } 5581 5582 /* 5583 * Check to see whether we are trying to migrate 5584 * page to lgroup where it is allocated already. 5585 * If so, clear the migrate bit and skip to next 5586 * page. 5587 */ 5588 if (i == 0 && to == from) { 5589 PP_CLRMIGRATE(ppa[0]); 5590 page_downgrade(ppa[0]); 5591 goto next; 5592 } 5593 } 5594 5595 /* 5596 * If all constituent pages couldn't be locked, 5597 * unlock pages locked so far and skip to next page. 5598 */ 5599 if (i != page_cnt) { 5600 while (--i != -1) { 5601 page_downgrade(ppa[i]); 5602 } 5603 goto next; 5604 } 5605 5606 (void) page_create_wait(page_cnt, PG_WAIT); 5607 newpp = page_get_replacement_page(pp, to, PGR_SAMESZC); 5608 if (newpp == NULL) { 5609 page_create_putback(page_cnt); 5610 for (i = 0; i < page_cnt; i++) { 5611 page_downgrade(ppa[i]); 5612 } 5613 lgrp_stat_add(to->lgrp_id, LGRP_PM_FAIL_ALLOC_PGS, 5614 page_cnt); 5615 goto next; 5616 } 5617 ASSERT(newpp->p_szc == pszc); 5618 /* 5619 * Clear migrate bit and relocate page 5620 */ 5621 PP_CLRMIGRATE(pp); 5622 if (page_relocate(&pp, &newpp, 0, 1, &page_cnt, to)) { 5623 panic("page_migrate: page_relocate failed"); 5624 } 5625 ASSERT(page_cnt * PAGESIZE == pgsz); 5626 5627 /* 5628 * Keep stats for number of pages migrated from and to 5629 * each lgroup 5630 */ 5631 lgrp_stat_add(from->lgrp_id, LGRP_PM_SRC_PGS, page_cnt); 5632 lgrp_stat_add(to->lgrp_id, LGRP_PM_DEST_PGS, page_cnt); 5633 /* 5634 * update the page_t array we were passed in and 5635 * unlink constituent pages of a large page. 5636 */ 5637 for (i = 0; i < page_cnt; ++i, ++pp) { 5638 ASSERT(PAGE_EXCL(newpp)); 5639 ASSERT(newpp->p_szc == pszc); 5640 ppa[i] = newpp; 5641 pp = newpp; 5642 page_sub(&newpp, pp); 5643 page_downgrade(pp); 5644 } 5645 ASSERT(newpp == NULL); 5646 next: 5647 addr += pgsz; 5648 ppa += page_cnt; 5649 npages -= page_cnt; 5650 } 5651 } 5652 5653 ulong_t mem_waiters = 0; 5654 ulong_t max_count = 20; 5655 #define MAX_DELAY 0x1ff 5656 5657 /* 5658 * Check if enough memory is available to proceed. 5659 * Depending on system configuration and how much memory is 5660 * reserved for swap we need to check against two variables. 5661 * e.g. on systems with little physical swap availrmem can be 5662 * more reliable indicator of how much memory is available. 5663 * On systems with large phys swap freemem can be better indicator. 5664 * If freemem drops below threshold level don't return an error 5665 * immediately but wake up pageout to free memory and block. 5666 * This is done number of times. If pageout is not able to free 5667 * memory within certain time return an error. 5668 * The same applies for availrmem but kmem_reap is used to 5669 * free memory. 5670 */ 5671 int 5672 page_mem_avail(pgcnt_t npages) 5673 { 5674 ulong_t count; 5675 5676 #if defined(__i386) 5677 if (freemem > desfree + npages && 5678 availrmem > swapfs_reserve + npages && 5679 btop(vmem_size(heap_arena, VMEM_FREE)) > tune.t_minarmem + 5680 npages) 5681 return (1); 5682 #else 5683 if (freemem > desfree + npages && 5684 availrmem > swapfs_reserve + npages) 5685 return (1); 5686 #endif 5687 5688 count = max_count; 5689 atomic_add_long(&mem_waiters, 1); 5690 5691 while (freemem < desfree + npages && --count) { 5692 cv_signal(&proc_pageout->p_cv); 5693 if (delay_sig(hz + (mem_waiters & MAX_DELAY))) { 5694 atomic_add_long(&mem_waiters, -1); 5695 return (0); 5696 } 5697 } 5698 if (count == 0) { 5699 atomic_add_long(&mem_waiters, -1); 5700 return (0); 5701 } 5702 5703 count = max_count; 5704 while (availrmem < swapfs_reserve + npages && --count) { 5705 kmem_reap(); 5706 if (delay_sig(hz + (mem_waiters & MAX_DELAY))) { 5707 atomic_add_long(&mem_waiters, -1); 5708 return (0); 5709 } 5710 } 5711 atomic_add_long(&mem_waiters, -1); 5712 if (count == 0) 5713 return (0); 5714 5715 #if defined(__i386) 5716 if (btop(vmem_size(heap_arena, VMEM_FREE)) < 5717 tune.t_minarmem + npages) 5718 return (0); 5719 #endif 5720 return (1); 5721 } 5722 5723 #define MAX_CNT 60 /* max num of iterations */ 5724 /* 5725 * Reclaim/reserve availrmem for npages. 5726 * If there is not enough memory start reaping seg, kmem caches. 5727 * Start pageout scanner (via page_needfree()). 5728 * Exit after ~ MAX_CNT s regardless of how much memory has been released. 5729 * Note: There is no guarantee that any availrmem will be freed as 5730 * this memory typically is locked (kernel heap) or reserved for swap. 5731 * Also due to memory fragmentation kmem allocator may not be able 5732 * to free any memory (single user allocated buffer will prevent 5733 * freeing slab or a page). 5734 */ 5735 int 5736 page_reclaim_mem(pgcnt_t npages, pgcnt_t epages, int adjust) 5737 { 5738 int i = 0; 5739 int ret = 0; 5740 pgcnt_t deficit; 5741 pgcnt_t old_availrmem; 5742 5743 mutex_enter(&freemem_lock); 5744 old_availrmem = availrmem - 1; 5745 while ((availrmem < tune.t_minarmem + npages + epages) && 5746 (old_availrmem < availrmem) && (i++ < MAX_CNT)) { 5747 old_availrmem = availrmem; 5748 deficit = tune.t_minarmem + npages + epages - availrmem; 5749 mutex_exit(&freemem_lock); 5750 page_needfree(deficit); 5751 kmem_reap(); 5752 delay(hz); 5753 page_needfree(-(spgcnt_t)deficit); 5754 mutex_enter(&freemem_lock); 5755 } 5756 5757 if (adjust && (availrmem >= tune.t_minarmem + npages + epages)) { 5758 availrmem -= npages; 5759 ret = 1; 5760 } 5761 5762 mutex_exit(&freemem_lock); 5763 5764 return (ret); 5765 } 5766 5767 /* 5768 * Search the memory segments to locate the desired page. Within a 5769 * segment, pages increase linearly with one page structure per 5770 * physical page frame (size PAGESIZE). The search begins 5771 * with the segment that was accessed last, to take advantage of locality. 5772 * If the hint misses, we start from the beginning of the sorted memseg list 5773 */ 5774 5775 5776 /* 5777 * Some data structures for pfn to pp lookup. 5778 */ 5779 ulong_t mhash_per_slot; 5780 struct memseg *memseg_hash[N_MEM_SLOTS]; 5781 5782 page_t * 5783 page_numtopp_nolock(pfn_t pfnum) 5784 { 5785 struct memseg *seg; 5786 page_t *pp; 5787 vm_cpu_data_t *vc = CPU->cpu_vm_data; 5788 5789 ASSERT(vc != NULL); 5790 5791 MEMSEG_STAT_INCR(nsearch); 5792 5793 /* Try last winner first */ 5794 if (((seg = vc->vc_pnum_memseg) != NULL) && 5795 (pfnum >= seg->pages_base) && (pfnum < seg->pages_end)) { 5796 MEMSEG_STAT_INCR(nlastwon); 5797 pp = seg->pages + (pfnum - seg->pages_base); 5798 if (pp->p_pagenum == pfnum) 5799 return ((page_t *)pp); 5800 } 5801 5802 /* Else Try hash */ 5803 if (((seg = memseg_hash[MEMSEG_PFN_HASH(pfnum)]) != NULL) && 5804 (pfnum >= seg->pages_base) && (pfnum < seg->pages_end)) { 5805 MEMSEG_STAT_INCR(nhashwon); 5806 vc->vc_pnum_memseg = seg; 5807 pp = seg->pages + (pfnum - seg->pages_base); 5808 if (pp->p_pagenum == pfnum) 5809 return ((page_t *)pp); 5810 } 5811 5812 /* Else Brute force */ 5813 for (seg = memsegs; seg != NULL; seg = seg->next) { 5814 if (pfnum >= seg->pages_base && pfnum < seg->pages_end) { 5815 vc->vc_pnum_memseg = seg; 5816 pp = seg->pages + (pfnum - seg->pages_base); 5817 return ((page_t *)pp); 5818 } 5819 } 5820 vc->vc_pnum_memseg = NULL; 5821 MEMSEG_STAT_INCR(nnotfound); 5822 return ((page_t *)NULL); 5823 5824 } 5825 5826 struct memseg * 5827 page_numtomemseg_nolock(pfn_t pfnum) 5828 { 5829 struct memseg *seg; 5830 page_t *pp; 5831 5832 /* Try hash */ 5833 if (((seg = memseg_hash[MEMSEG_PFN_HASH(pfnum)]) != NULL) && 5834 (pfnum >= seg->pages_base) && (pfnum < seg->pages_end)) { 5835 pp = seg->pages + (pfnum - seg->pages_base); 5836 if (pp->p_pagenum == pfnum) 5837 return (seg); 5838 } 5839 5840 /* Else Brute force */ 5841 for (seg = memsegs; seg != NULL; seg = seg->next) { 5842 if (pfnum >= seg->pages_base && pfnum < seg->pages_end) { 5843 return (seg); 5844 } 5845 } 5846 return ((struct memseg *)NULL); 5847 } 5848 5849 /* 5850 * Given a page and a count return the page struct that is 5851 * n structs away from the current one in the global page 5852 * list. 5853 * 5854 * This function wraps to the first page upon 5855 * reaching the end of the memseg list. 5856 */ 5857 page_t * 5858 page_nextn(page_t *pp, ulong_t n) 5859 { 5860 struct memseg *seg; 5861 page_t *ppn; 5862 vm_cpu_data_t *vc = (vm_cpu_data_t *)CPU->cpu_vm_data; 5863 5864 ASSERT(vc != NULL); 5865 5866 if (((seg = vc->vc_pnext_memseg) == NULL) || 5867 (seg->pages_base == seg->pages_end) || 5868 !(pp >= seg->pages && pp < seg->epages)) { 5869 5870 for (seg = memsegs; seg; seg = seg->next) { 5871 if (pp >= seg->pages && pp < seg->epages) 5872 break; 5873 } 5874 5875 if (seg == NULL) { 5876 /* Memory delete got in, return something valid. */ 5877 /* TODO: fix me. */ 5878 seg = memsegs; 5879 pp = seg->pages; 5880 } 5881 } 5882 5883 /* check for wraparound - possible if n is large */ 5884 while ((ppn = (pp + n)) >= seg->epages || ppn < pp) { 5885 n -= seg->epages - pp; 5886 seg = seg->next; 5887 if (seg == NULL) 5888 seg = memsegs; 5889 pp = seg->pages; 5890 } 5891 vc->vc_pnext_memseg = seg; 5892 return (ppn); 5893 } 5894 5895 /* 5896 * Initialize for a loop using page_next_scan_large(). 5897 */ 5898 page_t * 5899 page_next_scan_init(void **cookie) 5900 { 5901 ASSERT(cookie != NULL); 5902 *cookie = (void *)memsegs; 5903 return ((page_t *)memsegs->pages); 5904 } 5905 5906 /* 5907 * Return the next page in a scan of page_t's, assuming we want 5908 * to skip over sub-pages within larger page sizes. 5909 * 5910 * The cookie is used to keep track of the current memseg. 5911 */ 5912 page_t * 5913 page_next_scan_large( 5914 page_t *pp, 5915 ulong_t *n, 5916 void **cookie) 5917 { 5918 struct memseg *seg = (struct memseg *)*cookie; 5919 page_t *new_pp; 5920 ulong_t cnt; 5921 pfn_t pfn; 5922 5923 5924 /* 5925 * get the count of page_t's to skip based on the page size 5926 */ 5927 ASSERT(pp != NULL); 5928 if (pp->p_szc == 0) { 5929 cnt = 1; 5930 } else { 5931 pfn = page_pptonum(pp); 5932 cnt = page_get_pagecnt(pp->p_szc); 5933 cnt -= pfn & (cnt - 1); 5934 } 5935 *n += cnt; 5936 new_pp = pp + cnt; 5937 5938 /* 5939 * Catch if we went past the end of the current memory segment. If so, 5940 * just move to the next segment with pages. 5941 */ 5942 if (new_pp >= seg->epages) { 5943 do { 5944 seg = seg->next; 5945 if (seg == NULL) 5946 seg = memsegs; 5947 } while (seg->pages == seg->epages); 5948 new_pp = seg->pages; 5949 *cookie = (void *)seg; 5950 } 5951 5952 return (new_pp); 5953 } 5954 5955 5956 /* 5957 * Returns next page in list. Note: this function wraps 5958 * to the first page in the list upon reaching the end 5959 * of the list. Callers should be aware of this fact. 5960 */ 5961 5962 /* We should change this be a #define */ 5963 5964 page_t * 5965 page_next(page_t *pp) 5966 { 5967 return (page_nextn(pp, 1)); 5968 } 5969 5970 page_t * 5971 page_first() 5972 { 5973 return ((page_t *)memsegs->pages); 5974 } 5975 5976 5977 /* 5978 * This routine is called at boot with the initial memory configuration 5979 * and when memory is added or removed. 5980 */ 5981 void 5982 build_pfn_hash() 5983 { 5984 pfn_t cur; 5985 pgcnt_t index; 5986 struct memseg *pseg; 5987 int i; 5988 5989 /* 5990 * Clear memseg_hash array. 5991 * Since memory add/delete is designed to operate concurrently 5992 * with normal operation, the hash rebuild must be able to run 5993 * concurrently with page_numtopp_nolock(). To support this 5994 * functionality, assignments to memseg_hash array members must 5995 * be done atomically. 5996 * 5997 * NOTE: bzero() does not currently guarantee this for kernel 5998 * threads, and cannot be used here. 5999 */ 6000 for (i = 0; i < N_MEM_SLOTS; i++) 6001 memseg_hash[i] = NULL; 6002 6003 hat_kpm_mseghash_clear(N_MEM_SLOTS); 6004 6005 /* 6006 * Physmax is the last valid pfn. 6007 */ 6008 mhash_per_slot = (physmax + 1) >> MEM_HASH_SHIFT; 6009 for (pseg = memsegs; pseg != NULL; pseg = pseg->next) { 6010 index = MEMSEG_PFN_HASH(pseg->pages_base); 6011 cur = pseg->pages_base; 6012 do { 6013 if (index >= N_MEM_SLOTS) 6014 index = MEMSEG_PFN_HASH(cur); 6015 6016 if (memseg_hash[index] == NULL || 6017 memseg_hash[index]->pages_base > pseg->pages_base) { 6018 memseg_hash[index] = pseg; 6019 hat_kpm_mseghash_update(index, pseg); 6020 } 6021 cur += mhash_per_slot; 6022 index++; 6023 } while (cur < pseg->pages_end); 6024 } 6025 } 6026 6027 /* 6028 * Return the pagenum for the pp 6029 */ 6030 pfn_t 6031 page_pptonum(page_t *pp) 6032 { 6033 return (pp->p_pagenum); 6034 } 6035 6036 /* 6037 * interface to the referenced and modified etc bits 6038 * in the PSM part of the page struct 6039 * when no locking is desired. 6040 */ 6041 void 6042 page_set_props(page_t *pp, uint_t flags) 6043 { 6044 ASSERT((flags & ~(P_MOD | P_REF | P_RO)) == 0); 6045 pp->p_nrm |= (uchar_t)flags; 6046 } 6047 6048 extern void mach_sync_icache_pp(page_t *); 6049 #pragma weak mach_sync_icache_pp 6050 6051 /* 6052 * Flush I-cache if the page is being reassigned. The hashout flag is 6053 * set when a page has been removed from a hash chain (i.e. vnode 6054 * pages). If the page stays on the hash chain there is a chance it 6055 * will be re-used, therefore there is no need to flush the 6056 * I-cache. However, if the page is being removed from a hash chain 6057 * then the page can be used for any new purpose, and the I-cache must 6058 * be flushed. 6059 */ 6060 /* ARGSUSED */ 6061 void 6062 page_clr_all_props(page_t *pp, int hashout) 6063 { 6064 if (&mach_sync_icache_pp != NULL && hashout) { 6065 mach_sync_icache_pp(pp); 6066 } 6067 pp->p_nrm = 0; 6068 } 6069 6070 /* 6071 * Clear p_lckcnt and p_cowcnt, adjusting freemem if required. 6072 */ 6073 int 6074 page_clear_lck_cow(page_t *pp, int adjust) 6075 { 6076 int f_amount; 6077 6078 ASSERT(PAGE_EXCL(pp)); 6079 6080 /* 6081 * The page_struct_lock need not be acquired here since 6082 * we require the caller hold the page exclusively locked. 6083 */ 6084 f_amount = 0; 6085 if (pp->p_lckcnt) { 6086 f_amount = 1; 6087 pp->p_lckcnt = 0; 6088 } 6089 if (pp->p_cowcnt) { 6090 f_amount += pp->p_cowcnt; 6091 pp->p_cowcnt = 0; 6092 } 6093 6094 if (adjust && f_amount) { 6095 mutex_enter(&freemem_lock); 6096 availrmem += f_amount; 6097 mutex_exit(&freemem_lock); 6098 } 6099 6100 return (f_amount); 6101 } 6102 6103 /* 6104 * The following functions is called from free_vp_pages() 6105 * for an inexact estimate of a newly free'd page... 6106 */ 6107 ulong_t 6108 page_share_cnt(page_t *pp) 6109 { 6110 return (hat_page_getshare(pp)); 6111 } 6112 6113 int 6114 page_isshared(page_t *pp) 6115 { 6116 return (hat_page_checkshare(pp, 1)); 6117 } 6118 6119 int 6120 page_isfree(page_t *pp) 6121 { 6122 return (PP_ISFREE(pp)); 6123 } 6124 6125 int 6126 page_isref(page_t *pp) 6127 { 6128 return (hat_page_getattr(pp, P_REF)); 6129 } 6130 6131 int 6132 page_ismod(page_t *pp) 6133 { 6134 return (hat_page_getattr(pp, P_MOD)); 6135 } 6136 6137 /* 6138 * The following code all currently relates to the page capture logic: 6139 * 6140 * This logic is used for cases where there is a desire to claim a certain 6141 * physical page in the system for the caller. As it may not be possible 6142 * to capture the page immediately, the p_toxic bits are used in the page 6143 * structure to indicate that someone wants to capture this page. When the 6144 * page gets unlocked, the toxic flag will be noted and an attempt to capture 6145 * the page will be made. If it is successful, the original callers callback 6146 * will be called with the page to do with it what they please. 6147 * 6148 * There is also an async thread which wakes up to attempt to capture 6149 * pages occasionally which have the capture bit set. All of the pages which 6150 * need to be captured asynchronously have been inserted into the 6151 * page_capture_hash and thus this thread walks that hash list. Items in the 6152 * hash have an expiration time so this thread handles that as well by removing 6153 * the item from the hash if it has expired. 6154 * 6155 * Some important things to note are: 6156 * - if the PR_CAPTURE bit is set on a page, then the page is in the 6157 * page_capture_hash. The page_capture_hash_head.pchh_mutex is needed 6158 * to set and clear this bit, and while the lock is held is the only time 6159 * you can add or remove an entry from the hash. 6160 * - the PR_CAPTURE bit can only be set and cleared while holding the 6161 * page_capture_hash_head.pchh_mutex 6162 * - the t_flag field of the thread struct is used with the T_CAPTURING 6163 * flag to prevent recursion while dealing with large pages. 6164 * - pages which need to be retired never expire on the page_capture_hash. 6165 */ 6166 6167 static void page_capture_thread(void); 6168 static kthread_t *pc_thread_id; 6169 kcondvar_t pc_cv; 6170 static kmutex_t pc_thread_mutex; 6171 static clock_t pc_thread_shortwait; 6172 static clock_t pc_thread_longwait; 6173 static int pc_thread_retry; 6174 6175 struct page_capture_callback pc_cb[PC_NUM_CALLBACKS]; 6176 6177 /* Note that this is a circular linked list */ 6178 typedef struct page_capture_hash_bucket { 6179 page_t *pp; 6180 uint_t szc; 6181 uint_t flags; 6182 clock_t expires; /* lbolt at which this request expires. */ 6183 void *datap; /* Cached data passed in for callback */ 6184 struct page_capture_hash_bucket *next; 6185 struct page_capture_hash_bucket *prev; 6186 } page_capture_hash_bucket_t; 6187 6188 /* 6189 * Each hash bucket will have it's own mutex and two lists which are: 6190 * active (0): represents requests which have not been processed by 6191 * the page_capture async thread yet. 6192 * walked (1): represents requests which have been processed by the 6193 * page_capture async thread within it's given walk of this bucket. 6194 * 6195 * These are all needed so that we can synchronize all async page_capture 6196 * events. When the async thread moves to a new bucket, it will append the 6197 * walked list to the active list and walk each item one at a time, moving it 6198 * from the active list to the walked list. Thus if there is an async request 6199 * outstanding for a given page, it will always be in one of the two lists. 6200 * New requests will always be added to the active list. 6201 * If we were not able to capture a page before the request expired, we'd free 6202 * up the request structure which would indicate to page_capture that there is 6203 * no longer a need for the given page, and clear the PR_CAPTURE flag if 6204 * possible. 6205 */ 6206 typedef struct page_capture_hash_head { 6207 kmutex_t pchh_mutex; 6208 uint_t num_pages; 6209 page_capture_hash_bucket_t lists[2]; /* sentinel nodes */ 6210 } page_capture_hash_head_t; 6211 6212 #ifdef DEBUG 6213 #define NUM_PAGE_CAPTURE_BUCKETS 4 6214 #else 6215 #define NUM_PAGE_CAPTURE_BUCKETS 64 6216 #endif 6217 6218 page_capture_hash_head_t page_capture_hash[NUM_PAGE_CAPTURE_BUCKETS]; 6219 6220 /* for now use a very simple hash based upon the size of a page struct */ 6221 #define PAGE_CAPTURE_HASH(pp) \ 6222 ((int)(((uintptr_t)pp >> 7) & (NUM_PAGE_CAPTURE_BUCKETS - 1))) 6223 6224 extern pgcnt_t swapfs_minfree; 6225 6226 int page_trycapture(page_t *pp, uint_t szc, uint_t flags, void *datap); 6227 6228 /* 6229 * a callback function is required for page capture requests. 6230 */ 6231 void 6232 page_capture_register_callback(uint_t index, clock_t duration, 6233 int (*cb_func)(page_t *, void *, uint_t)) 6234 { 6235 ASSERT(pc_cb[index].cb_active == 0); 6236 ASSERT(cb_func != NULL); 6237 rw_enter(&pc_cb[index].cb_rwlock, RW_WRITER); 6238 pc_cb[index].duration = duration; 6239 pc_cb[index].cb_func = cb_func; 6240 pc_cb[index].cb_active = 1; 6241 rw_exit(&pc_cb[index].cb_rwlock); 6242 } 6243 6244 void 6245 page_capture_unregister_callback(uint_t index) 6246 { 6247 int i, j; 6248 struct page_capture_hash_bucket *bp1; 6249 struct page_capture_hash_bucket *bp2; 6250 struct page_capture_hash_bucket *head = NULL; 6251 uint_t flags = (1 << index); 6252 6253 rw_enter(&pc_cb[index].cb_rwlock, RW_WRITER); 6254 ASSERT(pc_cb[index].cb_active == 1); 6255 pc_cb[index].duration = 0; /* Paranoia */ 6256 pc_cb[index].cb_func = NULL; /* Paranoia */ 6257 pc_cb[index].cb_active = 0; 6258 rw_exit(&pc_cb[index].cb_rwlock); 6259 6260 /* 6261 * Just move all the entries to a private list which we can walk 6262 * through without the need to hold any locks. 6263 * No more requests can get added to the hash lists for this consumer 6264 * as the cb_active field for the callback has been cleared. 6265 */ 6266 for (i = 0; i < NUM_PAGE_CAPTURE_BUCKETS; i++) { 6267 mutex_enter(&page_capture_hash[i].pchh_mutex); 6268 for (j = 0; j < 2; j++) { 6269 bp1 = page_capture_hash[i].lists[j].next; 6270 /* walk through all but first (sentinel) element */ 6271 while (bp1 != &page_capture_hash[i].lists[j]) { 6272 bp2 = bp1; 6273 if (bp2->flags & flags) { 6274 bp1 = bp2->next; 6275 bp1->prev = bp2->prev; 6276 bp2->prev->next = bp1; 6277 bp2->next = head; 6278 head = bp2; 6279 /* 6280 * Clear the PR_CAPTURE bit as we 6281 * hold appropriate locks here. 6282 */ 6283 page_clrtoxic(head->pp, PR_CAPTURE); 6284 page_capture_hash[i].num_pages--; 6285 continue; 6286 } 6287 bp1 = bp1->next; 6288 } 6289 } 6290 mutex_exit(&page_capture_hash[i].pchh_mutex); 6291 } 6292 6293 while (head != NULL) { 6294 bp1 = head; 6295 head = head->next; 6296 kmem_free(bp1, sizeof (*bp1)); 6297 } 6298 } 6299 6300 6301 /* 6302 * Find pp in the active list and move it to the walked list if it 6303 * exists. 6304 * Note that most often pp should be at the front of the active list 6305 * as it is currently used and thus there is no other sort of optimization 6306 * being done here as this is a linked list data structure. 6307 * Returns 1 on successful move or 0 if page could not be found. 6308 */ 6309 static int 6310 page_capture_move_to_walked(page_t *pp) 6311 { 6312 page_capture_hash_bucket_t *bp; 6313 int index; 6314 6315 index = PAGE_CAPTURE_HASH(pp); 6316 6317 mutex_enter(&page_capture_hash[index].pchh_mutex); 6318 bp = page_capture_hash[index].lists[0].next; 6319 while (bp != &page_capture_hash[index].lists[0]) { 6320 if (bp->pp == pp) { 6321 /* Remove from old list */ 6322 bp->next->prev = bp->prev; 6323 bp->prev->next = bp->next; 6324 6325 /* Add to new list */ 6326 bp->next = page_capture_hash[index].lists[1].next; 6327 bp->prev = &page_capture_hash[index].lists[1]; 6328 page_capture_hash[index].lists[1].next = bp; 6329 bp->next->prev = bp; 6330 mutex_exit(&page_capture_hash[index].pchh_mutex); 6331 6332 return (1); 6333 } 6334 bp = bp->next; 6335 } 6336 mutex_exit(&page_capture_hash[index].pchh_mutex); 6337 return (0); 6338 } 6339 6340 /* 6341 * Add a new entry to the page capture hash. The only case where a new 6342 * entry is not added is when the page capture consumer is no longer registered. 6343 * In this case, we'll silently not add the page to the hash. We know that 6344 * page retire will always be registered for the case where we are currently 6345 * unretiring a page and thus there are no conflicts. 6346 */ 6347 static void 6348 page_capture_add_hash(page_t *pp, uint_t szc, uint_t flags, void *datap) 6349 { 6350 page_capture_hash_bucket_t *bp1; 6351 page_capture_hash_bucket_t *bp2; 6352 int index; 6353 int cb_index; 6354 int i; 6355 #ifdef DEBUG 6356 page_capture_hash_bucket_t *tp1; 6357 int l; 6358 #endif 6359 6360 ASSERT(!(flags & CAPTURE_ASYNC)); 6361 6362 bp1 = kmem_alloc(sizeof (struct page_capture_hash_bucket), KM_SLEEP); 6363 6364 bp1->pp = pp; 6365 bp1->szc = szc; 6366 bp1->flags = flags; 6367 bp1->datap = datap; 6368 6369 for (cb_index = 0; cb_index < PC_NUM_CALLBACKS; cb_index++) { 6370 if ((flags >> cb_index) & 1) { 6371 break; 6372 } 6373 } 6374 6375 ASSERT(cb_index != PC_NUM_CALLBACKS); 6376 6377 rw_enter(&pc_cb[cb_index].cb_rwlock, RW_READER); 6378 if (pc_cb[cb_index].cb_active) { 6379 if (pc_cb[cb_index].duration == -1) { 6380 bp1->expires = (clock_t)-1; 6381 } else { 6382 bp1->expires = lbolt + pc_cb[cb_index].duration; 6383 } 6384 } else { 6385 /* There's no callback registered so don't add to the hash */ 6386 rw_exit(&pc_cb[cb_index].cb_rwlock); 6387 kmem_free(bp1, sizeof (*bp1)); 6388 return; 6389 } 6390 6391 index = PAGE_CAPTURE_HASH(pp); 6392 6393 /* 6394 * Only allow capture flag to be modified under this mutex. 6395 * Prevents multiple entries for same page getting added. 6396 */ 6397 mutex_enter(&page_capture_hash[index].pchh_mutex); 6398 6399 /* 6400 * if not already on the hash, set capture bit and add to the hash 6401 */ 6402 if (!(pp->p_toxic & PR_CAPTURE)) { 6403 #ifdef DEBUG 6404 /* Check for duplicate entries */ 6405 for (l = 0; l < 2; l++) { 6406 tp1 = page_capture_hash[index].lists[l].next; 6407 while (tp1 != &page_capture_hash[index].lists[l]) { 6408 if (tp1->pp == pp) { 6409 panic("page pp 0x%p already on hash " 6410 "at 0x%p\n", 6411 (void *)pp, (void *)tp1); 6412 } 6413 tp1 = tp1->next; 6414 } 6415 } 6416 6417 #endif 6418 page_settoxic(pp, PR_CAPTURE); 6419 bp1->next = page_capture_hash[index].lists[0].next; 6420 bp1->prev = &page_capture_hash[index].lists[0]; 6421 bp1->next->prev = bp1; 6422 page_capture_hash[index].lists[0].next = bp1; 6423 page_capture_hash[index].num_pages++; 6424 if (flags & CAPTURE_RETIRE) { 6425 page_retire_incr_pend_count(); 6426 } 6427 mutex_exit(&page_capture_hash[index].pchh_mutex); 6428 rw_exit(&pc_cb[cb_index].cb_rwlock); 6429 cv_signal(&pc_cv); 6430 return; 6431 } 6432 6433 /* 6434 * A page retire request will replace any other request. 6435 * A second physmem request which is for a different process than 6436 * the currently registered one will be dropped as there is 6437 * no way to hold the private data for both calls. 6438 * In the future, once there are more callers, this will have to 6439 * be worked out better as there needs to be private storage for 6440 * at least each type of caller (maybe have datap be an array of 6441 * *void's so that we can index based upon callers index). 6442 */ 6443 6444 /* walk hash list to update expire time */ 6445 for (i = 0; i < 2; i++) { 6446 bp2 = page_capture_hash[index].lists[i].next; 6447 while (bp2 != &page_capture_hash[index].lists[i]) { 6448 if (bp2->pp == pp) { 6449 if (flags & CAPTURE_RETIRE) { 6450 if (!(bp2->flags & CAPTURE_RETIRE)) { 6451 page_retire_incr_pend_count(); 6452 bp2->flags = flags; 6453 bp2->expires = bp1->expires; 6454 bp2->datap = datap; 6455 } 6456 } else { 6457 ASSERT(flags & CAPTURE_PHYSMEM); 6458 if (!(bp2->flags & CAPTURE_RETIRE) && 6459 (datap == bp2->datap)) { 6460 bp2->expires = bp1->expires; 6461 } 6462 } 6463 mutex_exit(&page_capture_hash[index]. 6464 pchh_mutex); 6465 rw_exit(&pc_cb[cb_index].cb_rwlock); 6466 kmem_free(bp1, sizeof (*bp1)); 6467 return; 6468 } 6469 bp2 = bp2->next; 6470 } 6471 } 6472 6473 /* 6474 * the PR_CAPTURE flag is protected by the page_capture_hash mutexes 6475 * and thus it either has to be set or not set and can't change 6476 * while holding the mutex above. 6477 */ 6478 panic("page_capture_add_hash, PR_CAPTURE flag set on pp %p\n", 6479 (void *)pp); 6480 } 6481 6482 /* 6483 * We have a page in our hands, lets try and make it ours by turning 6484 * it into a clean page like it had just come off the freelists. 6485 * 6486 * Returns 0 on success, with the page still EXCL locked. 6487 * On failure, the page will be unlocked, and returns EAGAIN 6488 */ 6489 static int 6490 page_capture_clean_page(page_t *pp) 6491 { 6492 page_t *newpp; 6493 int skip_unlock = 0; 6494 spgcnt_t count; 6495 page_t *tpp; 6496 int ret = 0; 6497 int extra; 6498 6499 ASSERT(PAGE_EXCL(pp)); 6500 ASSERT(!PP_RETIRED(pp)); 6501 ASSERT(curthread->t_flag & T_CAPTURING); 6502 6503 if (PP_ISFREE(pp)) { 6504 if (!page_reclaim(pp, NULL)) { 6505 skip_unlock = 1; 6506 ret = EAGAIN; 6507 goto cleanup; 6508 } 6509 ASSERT(pp->p_szc == 0); 6510 if (pp->p_vnode != NULL) { 6511 /* 6512 * Since this page came from the 6513 * cachelist, we must destroy the 6514 * old vnode association. 6515 */ 6516 page_hashout(pp, NULL); 6517 } 6518 goto cleanup; 6519 } 6520 6521 /* 6522 * If we know page_relocate will fail, skip it 6523 * It could still fail due to a UE on another page but we 6524 * can't do anything about that. 6525 */ 6526 if (pp->p_toxic & PR_UE) { 6527 goto skip_relocate; 6528 } 6529 6530 /* 6531 * It's possible that pages can not have a vnode as fsflush comes 6532 * through and cleans up these pages. It's ugly but that's how it is. 6533 */ 6534 if (pp->p_vnode == NULL) { 6535 goto skip_relocate; 6536 } 6537 6538 /* 6539 * Page was not free, so lets try to relocate it. 6540 * page_relocate only works with root pages, so if this is not a root 6541 * page, we need to demote it to try and relocate it. 6542 * Unfortunately this is the best we can do right now. 6543 */ 6544 newpp = NULL; 6545 if ((pp->p_szc > 0) && (pp != PP_PAGEROOT(pp))) { 6546 if (page_try_demote_pages(pp) == 0) { 6547 ret = EAGAIN; 6548 goto cleanup; 6549 } 6550 } 6551 ret = page_relocate(&pp, &newpp, 1, 0, &count, NULL); 6552 if (ret == 0) { 6553 page_t *npp; 6554 /* unlock the new page(s) */ 6555 while (count-- > 0) { 6556 ASSERT(newpp != NULL); 6557 npp = newpp; 6558 page_sub(&newpp, npp); 6559 page_unlock(npp); 6560 } 6561 ASSERT(newpp == NULL); 6562 /* 6563 * Check to see if the page we have is too large. 6564 * If so, demote it freeing up the extra pages. 6565 */ 6566 if (pp->p_szc > 0) { 6567 /* For now demote extra pages to szc == 0 */ 6568 extra = page_get_pagecnt(pp->p_szc) - 1; 6569 while (extra > 0) { 6570 tpp = pp->p_next; 6571 page_sub(&pp, tpp); 6572 tpp->p_szc = 0; 6573 page_free(tpp, 1); 6574 extra--; 6575 } 6576 /* Make sure to set our page to szc 0 as well */ 6577 ASSERT(pp->p_next == pp && pp->p_prev == pp); 6578 pp->p_szc = 0; 6579 } 6580 goto cleanup; 6581 } else if (ret == EIO) { 6582 ret = EAGAIN; 6583 goto cleanup; 6584 } else { 6585 /* 6586 * Need to reset return type as we failed to relocate the page 6587 * but that does not mean that some of the next steps will not 6588 * work. 6589 */ 6590 ret = 0; 6591 } 6592 6593 skip_relocate: 6594 6595 if (pp->p_szc > 0) { 6596 if (page_try_demote_pages(pp) == 0) { 6597 ret = EAGAIN; 6598 goto cleanup; 6599 } 6600 } 6601 6602 ASSERT(pp->p_szc == 0); 6603 6604 if (hat_ismod(pp)) { 6605 ret = EAGAIN; 6606 goto cleanup; 6607 } 6608 if (PP_ISKAS(pp)) { 6609 ret = EAGAIN; 6610 goto cleanup; 6611 } 6612 if (pp->p_lckcnt || pp->p_cowcnt) { 6613 ret = EAGAIN; 6614 goto cleanup; 6615 } 6616 6617 (void) hat_pageunload(pp, HAT_FORCE_PGUNLOAD); 6618 ASSERT(!hat_page_is_mapped(pp)); 6619 6620 if (hat_ismod(pp)) { 6621 /* 6622 * This is a semi-odd case as the page is now modified but not 6623 * mapped as we just unloaded the mappings above. 6624 */ 6625 ret = EAGAIN; 6626 goto cleanup; 6627 } 6628 if (pp->p_vnode != NULL) { 6629 page_hashout(pp, NULL); 6630 } 6631 6632 /* 6633 * At this point, the page should be in a clean state and 6634 * we can do whatever we want with it. 6635 */ 6636 6637 cleanup: 6638 if (ret != 0) { 6639 if (!skip_unlock) { 6640 page_unlock(pp); 6641 } 6642 } else { 6643 ASSERT(pp->p_szc == 0); 6644 ASSERT(PAGE_EXCL(pp)); 6645 6646 pp->p_next = pp; 6647 pp->p_prev = pp; 6648 } 6649 return (ret); 6650 } 6651 6652 /* 6653 * Various callers of page_trycapture() can have different restrictions upon 6654 * what memory they have access to. 6655 * Returns 0 on success, with the following error codes on failure: 6656 * EPERM - The requested page is long term locked, and thus repeated 6657 * requests to capture this page will likely fail. 6658 * ENOMEM - There was not enough free memory in the system to safely 6659 * map the requested page. 6660 * ENOENT - The requested page was inside the kernel cage, and the 6661 * PHYSMEM_CAGE flag was not set. 6662 */ 6663 int 6664 page_capture_pre_checks(page_t *pp, uint_t flags) 6665 { 6666 #if defined(__sparc) 6667 extern struct vnode prom_ppages; 6668 #endif /* __sparc */ 6669 6670 ASSERT(pp != NULL); 6671 6672 #if defined(__sparc) 6673 if (pp->p_vnode == &prom_ppages) { 6674 return (EPERM); 6675 } 6676 6677 if (PP_ISNORELOC(pp) && !(flags & CAPTURE_GET_CAGE) && 6678 (flags & CAPTURE_PHYSMEM)) { 6679 return (ENOENT); 6680 } 6681 6682 if (PP_ISNORELOCKERNEL(pp)) { 6683 return (EPERM); 6684 } 6685 #else 6686 if (PP_ISKAS(pp)) { 6687 return (EPERM); 6688 } 6689 #endif /* __sparc */ 6690 6691 /* only physmem currently has the restrictions checked below */ 6692 if (!(flags & CAPTURE_PHYSMEM)) { 6693 return (0); 6694 } 6695 6696 if (availrmem < swapfs_minfree) { 6697 /* 6698 * We won't try to capture this page as we are 6699 * running low on memory. 6700 */ 6701 return (ENOMEM); 6702 } 6703 return (0); 6704 } 6705 6706 /* 6707 * Once we have a page in our mits, go ahead and complete the capture 6708 * operation. 6709 * Returns 1 on failure where page is no longer needed 6710 * Returns 0 on success 6711 * Returns -1 if there was a transient failure. 6712 * Failure cases must release the SE_EXCL lock on pp (usually via page_free). 6713 */ 6714 int 6715 page_capture_take_action(page_t *pp, uint_t flags, void *datap) 6716 { 6717 int cb_index; 6718 int ret = 0; 6719 page_capture_hash_bucket_t *bp1; 6720 page_capture_hash_bucket_t *bp2; 6721 int index; 6722 int found = 0; 6723 int i; 6724 6725 ASSERT(PAGE_EXCL(pp)); 6726 ASSERT(curthread->t_flag & T_CAPTURING); 6727 6728 for (cb_index = 0; cb_index < PC_NUM_CALLBACKS; cb_index++) { 6729 if ((flags >> cb_index) & 1) { 6730 break; 6731 } 6732 } 6733 ASSERT(cb_index < PC_NUM_CALLBACKS); 6734 6735 /* 6736 * Remove the entry from the page_capture hash, but don't free it yet 6737 * as we may need to put it back. 6738 * Since we own the page at this point in time, we should find it 6739 * in the hash if this is an ASYNC call. If we don't it's likely 6740 * that the page_capture_async() thread decided that this request 6741 * had expired, in which case we just continue on. 6742 */ 6743 if (flags & CAPTURE_ASYNC) { 6744 6745 index = PAGE_CAPTURE_HASH(pp); 6746 6747 mutex_enter(&page_capture_hash[index].pchh_mutex); 6748 for (i = 0; i < 2 && !found; i++) { 6749 bp1 = page_capture_hash[index].lists[i].next; 6750 while (bp1 != &page_capture_hash[index].lists[i]) { 6751 if (bp1->pp == pp) { 6752 bp1->next->prev = bp1->prev; 6753 bp1->prev->next = bp1->next; 6754 page_capture_hash[index].num_pages--; 6755 page_clrtoxic(pp, PR_CAPTURE); 6756 found = 1; 6757 break; 6758 } 6759 bp1 = bp1->next; 6760 } 6761 } 6762 mutex_exit(&page_capture_hash[index].pchh_mutex); 6763 } 6764 6765 /* Synchronize with the unregister func. */ 6766 rw_enter(&pc_cb[cb_index].cb_rwlock, RW_READER); 6767 if (!pc_cb[cb_index].cb_active) { 6768 page_free(pp, 1); 6769 rw_exit(&pc_cb[cb_index].cb_rwlock); 6770 if (found) { 6771 kmem_free(bp1, sizeof (*bp1)); 6772 } 6773 return (1); 6774 } 6775 6776 /* 6777 * We need to remove the entry from the page capture hash and turn off 6778 * the PR_CAPTURE bit before calling the callback. We'll need to cache 6779 * the entry here, and then based upon the return value, cleanup 6780 * appropriately or re-add it to the hash, making sure that someone else 6781 * hasn't already done so. 6782 * It should be rare for the callback to fail and thus it's ok for 6783 * the failure path to be a bit complicated as the success path is 6784 * cleaner and the locking rules are easier to follow. 6785 */ 6786 6787 ret = pc_cb[cb_index].cb_func(pp, datap, flags); 6788 6789 rw_exit(&pc_cb[cb_index].cb_rwlock); 6790 6791 /* 6792 * If this was an ASYNC request, we need to cleanup the hash if the 6793 * callback was successful or if the request was no longer valid. 6794 * For non-ASYNC requests, we return failure to map and the caller 6795 * will take care of adding the request to the hash. 6796 * Note also that the callback itself is responsible for the page 6797 * at this point in time in terms of locking ... The most common 6798 * case for the failure path should just be a page_free. 6799 */ 6800 if (ret >= 0) { 6801 if (found) { 6802 if (bp1->flags & CAPTURE_RETIRE) { 6803 page_retire_decr_pend_count(); 6804 } 6805 kmem_free(bp1, sizeof (*bp1)); 6806 } 6807 return (ret); 6808 } 6809 if (!found) { 6810 return (ret); 6811 } 6812 6813 ASSERT(flags & CAPTURE_ASYNC); 6814 6815 /* 6816 * Check for expiration time first as we can just free it up if it's 6817 * expired. 6818 */ 6819 if (lbolt > bp1->expires && bp1->expires != -1) { 6820 kmem_free(bp1, sizeof (*bp1)); 6821 return (ret); 6822 } 6823 6824 /* 6825 * The callback failed and there used to be an entry in the hash for 6826 * this page, so we need to add it back to the hash. 6827 */ 6828 mutex_enter(&page_capture_hash[index].pchh_mutex); 6829 if (!(pp->p_toxic & PR_CAPTURE)) { 6830 /* just add bp1 back to head of walked list */ 6831 page_settoxic(pp, PR_CAPTURE); 6832 bp1->next = page_capture_hash[index].lists[1].next; 6833 bp1->prev = &page_capture_hash[index].lists[1]; 6834 bp1->next->prev = bp1; 6835 page_capture_hash[index].lists[1].next = bp1; 6836 page_capture_hash[index].num_pages++; 6837 mutex_exit(&page_capture_hash[index].pchh_mutex); 6838 return (ret); 6839 } 6840 6841 /* 6842 * Otherwise there was a new capture request added to list 6843 * Need to make sure that our original data is represented if 6844 * appropriate. 6845 */ 6846 for (i = 0; i < 2; i++) { 6847 bp2 = page_capture_hash[index].lists[i].next; 6848 while (bp2 != &page_capture_hash[index].lists[i]) { 6849 if (bp2->pp == pp) { 6850 if (bp1->flags & CAPTURE_RETIRE) { 6851 if (!(bp2->flags & CAPTURE_RETIRE)) { 6852 bp2->szc = bp1->szc; 6853 bp2->flags = bp1->flags; 6854 bp2->expires = bp1->expires; 6855 bp2->datap = bp1->datap; 6856 } 6857 } else { 6858 ASSERT(bp1->flags & CAPTURE_PHYSMEM); 6859 if (!(bp2->flags & CAPTURE_RETIRE)) { 6860 bp2->szc = bp1->szc; 6861 bp2->flags = bp1->flags; 6862 bp2->expires = bp1->expires; 6863 bp2->datap = bp1->datap; 6864 } 6865 } 6866 mutex_exit(&page_capture_hash[index]. 6867 pchh_mutex); 6868 kmem_free(bp1, sizeof (*bp1)); 6869 return (ret); 6870 } 6871 bp2 = bp2->next; 6872 } 6873 } 6874 panic("PR_CAPTURE set but not on hash for pp 0x%p\n", (void *)pp); 6875 /*NOTREACHED*/ 6876 } 6877 6878 /* 6879 * Try to capture the given page for the caller specified in the flags 6880 * parameter. The page will either be captured and handed over to the 6881 * appropriate callback, or will be queued up in the page capture hash 6882 * to be captured asynchronously. 6883 * If the current request is due to an async capture, the page must be 6884 * exclusively locked before calling this function. 6885 * Currently szc must be 0 but in the future this should be expandable to 6886 * other page sizes. 6887 * Returns 0 on success, with the following error codes on failure: 6888 * EPERM - The requested page is long term locked, and thus repeated 6889 * requests to capture this page will likely fail. 6890 * ENOMEM - There was not enough free memory in the system to safely 6891 * map the requested page. 6892 * ENOENT - The requested page was inside the kernel cage, and the 6893 * CAPTURE_GET_CAGE flag was not set. 6894 * EAGAIN - The requested page could not be capturead at this point in 6895 * time but future requests will likely work. 6896 * EBUSY - The requested page is retired and the CAPTURE_GET_RETIRED flag 6897 * was not set. 6898 */ 6899 int 6900 page_itrycapture(page_t *pp, uint_t szc, uint_t flags, void *datap) 6901 { 6902 int ret; 6903 int cb_index; 6904 6905 if (flags & CAPTURE_ASYNC) { 6906 ASSERT(PAGE_EXCL(pp)); 6907 goto async; 6908 } 6909 6910 /* Make sure there's enough availrmem ... */ 6911 ret = page_capture_pre_checks(pp, flags); 6912 if (ret != 0) { 6913 return (ret); 6914 } 6915 6916 if (!page_trylock(pp, SE_EXCL)) { 6917 for (cb_index = 0; cb_index < PC_NUM_CALLBACKS; cb_index++) { 6918 if ((flags >> cb_index) & 1) { 6919 break; 6920 } 6921 } 6922 ASSERT(cb_index < PC_NUM_CALLBACKS); 6923 ret = EAGAIN; 6924 /* Special case for retired pages */ 6925 if (PP_RETIRED(pp)) { 6926 if (flags & CAPTURE_GET_RETIRED) { 6927 if (!page_unretire_pp(pp, PR_UNR_TEMP)) { 6928 /* 6929 * Need to set capture bit and add to 6930 * hash so that the page will be 6931 * retired when freed. 6932 */ 6933 page_capture_add_hash(pp, szc, 6934 CAPTURE_RETIRE, NULL); 6935 ret = 0; 6936 goto own_page; 6937 } 6938 } else { 6939 return (EBUSY); 6940 } 6941 } 6942 page_capture_add_hash(pp, szc, flags, datap); 6943 return (ret); 6944 } 6945 6946 async: 6947 ASSERT(PAGE_EXCL(pp)); 6948 6949 /* Need to check for physmem async requests that availrmem is sane */ 6950 if ((flags & (CAPTURE_ASYNC | CAPTURE_PHYSMEM)) == 6951 (CAPTURE_ASYNC | CAPTURE_PHYSMEM) && 6952 (availrmem < swapfs_minfree)) { 6953 page_unlock(pp); 6954 return (ENOMEM); 6955 } 6956 6957 ret = page_capture_clean_page(pp); 6958 6959 if (ret != 0) { 6960 /* We failed to get the page, so lets add it to the hash */ 6961 if (!(flags & CAPTURE_ASYNC)) { 6962 page_capture_add_hash(pp, szc, flags, datap); 6963 } 6964 return (ret); 6965 } 6966 6967 own_page: 6968 ASSERT(PAGE_EXCL(pp)); 6969 ASSERT(pp->p_szc == 0); 6970 6971 /* Call the callback */ 6972 ret = page_capture_take_action(pp, flags, datap); 6973 6974 if (ret == 0) { 6975 return (0); 6976 } 6977 6978 /* 6979 * Note that in the failure cases from page_capture_take_action, the 6980 * EXCL lock will have already been dropped. 6981 */ 6982 if ((ret == -1) && (!(flags & CAPTURE_ASYNC))) { 6983 page_capture_add_hash(pp, szc, flags, datap); 6984 } 6985 return (EAGAIN); 6986 } 6987 6988 int 6989 page_trycapture(page_t *pp, uint_t szc, uint_t flags, void *datap) 6990 { 6991 int ret; 6992 6993 curthread->t_flag |= T_CAPTURING; 6994 ret = page_itrycapture(pp, szc, flags, datap); 6995 curthread->t_flag &= ~T_CAPTURING; /* xor works as we know its set */ 6996 return (ret); 6997 } 6998 6999 /* 7000 * When unlocking a page which has the PR_CAPTURE bit set, this routine 7001 * gets called to try and capture the page. 7002 */ 7003 void 7004 page_unlock_capture(page_t *pp) 7005 { 7006 page_capture_hash_bucket_t *bp; 7007 int index; 7008 int i; 7009 uint_t szc; 7010 uint_t flags = 0; 7011 void *datap; 7012 kmutex_t *mp; 7013 extern vnode_t retired_pages; 7014 7015 /* 7016 * We need to protect against a possible deadlock here where we own 7017 * the vnode page hash mutex and want to acquire it again as there 7018 * are locations in the code, where we unlock a page while holding 7019 * the mutex which can lead to the page being captured and eventually 7020 * end up here. As we may be hashing out the old page and hashing into 7021 * the retire vnode, we need to make sure we don't own them. 7022 * Other callbacks who do hash operations also need to make sure that 7023 * before they hashin to a vnode that they do not currently own the 7024 * vphm mutex otherwise there will be a panic. 7025 */ 7026 if (mutex_owned(page_vnode_mutex(&retired_pages))) { 7027 page_unlock_nocapture(pp); 7028 return; 7029 } 7030 if (pp->p_vnode != NULL && mutex_owned(page_vnode_mutex(pp->p_vnode))) { 7031 page_unlock_nocapture(pp); 7032 return; 7033 } 7034 7035 index = PAGE_CAPTURE_HASH(pp); 7036 7037 mp = &page_capture_hash[index].pchh_mutex; 7038 mutex_enter(mp); 7039 for (i = 0; i < 2; i++) { 7040 bp = page_capture_hash[index].lists[i].next; 7041 while (bp != &page_capture_hash[index].lists[i]) { 7042 if (bp->pp == pp) { 7043 szc = bp->szc; 7044 flags = bp->flags | CAPTURE_ASYNC; 7045 datap = bp->datap; 7046 mutex_exit(mp); 7047 (void) page_trycapture(pp, szc, flags, datap); 7048 return; 7049 } 7050 bp = bp->next; 7051 } 7052 } 7053 7054 /* Failed to find page in hash so clear flags and unlock it. */ 7055 page_clrtoxic(pp, PR_CAPTURE); 7056 page_unlock(pp); 7057 7058 mutex_exit(mp); 7059 } 7060 7061 void 7062 page_capture_init() 7063 { 7064 int i; 7065 for (i = 0; i < NUM_PAGE_CAPTURE_BUCKETS; i++) { 7066 page_capture_hash[i].lists[0].next = 7067 &page_capture_hash[i].lists[0]; 7068 page_capture_hash[i].lists[0].prev = 7069 &page_capture_hash[i].lists[0]; 7070 page_capture_hash[i].lists[1].next = 7071 &page_capture_hash[i].lists[1]; 7072 page_capture_hash[i].lists[1].prev = 7073 &page_capture_hash[i].lists[1]; 7074 } 7075 7076 pc_thread_shortwait = 23 * hz; 7077 pc_thread_longwait = 1201 * hz; 7078 pc_thread_retry = 3; 7079 mutex_init(&pc_thread_mutex, NULL, MUTEX_DEFAULT, NULL); 7080 cv_init(&pc_cv, NULL, CV_DEFAULT, NULL); 7081 pc_thread_id = thread_create(NULL, 0, page_capture_thread, NULL, 0, &p0, 7082 TS_RUN, minclsyspri); 7083 } 7084 7085 /* 7086 * It is necessary to scrub any failing pages prior to reboot in order to 7087 * prevent a latent error trap from occurring on the next boot. 7088 */ 7089 void 7090 page_retire_mdboot() 7091 { 7092 page_t *pp; 7093 int i, j; 7094 page_capture_hash_bucket_t *bp; 7095 7096 /* walk lists looking for pages to scrub */ 7097 for (i = 0; i < NUM_PAGE_CAPTURE_BUCKETS; i++) { 7098 if (page_capture_hash[i].num_pages == 0) 7099 continue; 7100 7101 mutex_enter(&page_capture_hash[i].pchh_mutex); 7102 7103 for (j = 0; j < 2; j++) { 7104 bp = page_capture_hash[i].lists[j].next; 7105 while (bp != &page_capture_hash[i].lists[j]) { 7106 pp = bp->pp; 7107 if (!PP_ISKAS(pp) && PP_TOXIC(pp)) { 7108 pp->p_selock = -1; /* pacify ASSERTs */ 7109 PP_CLRFREE(pp); 7110 pagescrub(pp, 0, PAGESIZE); 7111 pp->p_selock = 0; 7112 } 7113 bp = bp->next; 7114 } 7115 } 7116 mutex_exit(&page_capture_hash[i].pchh_mutex); 7117 } 7118 } 7119 7120 /* 7121 * Walk the page_capture_hash trying to capture pages and also cleanup old 7122 * entries which have expired. 7123 */ 7124 void 7125 page_capture_async() 7126 { 7127 page_t *pp; 7128 int i; 7129 int ret; 7130 page_capture_hash_bucket_t *bp1, *bp2; 7131 uint_t szc; 7132 uint_t flags; 7133 void *datap; 7134 7135 /* If there are outstanding pages to be captured, get to work */ 7136 for (i = 0; i < NUM_PAGE_CAPTURE_BUCKETS; i++) { 7137 if (page_capture_hash[i].num_pages == 0) 7138 continue; 7139 /* Append list 1 to list 0 and then walk through list 0 */ 7140 mutex_enter(&page_capture_hash[i].pchh_mutex); 7141 bp1 = &page_capture_hash[i].lists[1]; 7142 bp2 = bp1->next; 7143 if (bp1 != bp2) { 7144 bp1->prev->next = page_capture_hash[i].lists[0].next; 7145 bp2->prev = &page_capture_hash[i].lists[0]; 7146 page_capture_hash[i].lists[0].next->prev = bp1->prev; 7147 page_capture_hash[i].lists[0].next = bp2; 7148 bp1->next = bp1; 7149 bp1->prev = bp1; 7150 } 7151 7152 /* list[1] will be empty now */ 7153 7154 bp1 = page_capture_hash[i].lists[0].next; 7155 while (bp1 != &page_capture_hash[i].lists[0]) { 7156 /* Check expiration time */ 7157 if ((lbolt > bp1->expires && bp1->expires != -1) || 7158 page_deleted(bp1->pp)) { 7159 page_capture_hash[i].lists[0].next = bp1->next; 7160 bp1->next->prev = 7161 &page_capture_hash[i].lists[0]; 7162 page_capture_hash[i].num_pages--; 7163 7164 /* 7165 * We can safely remove the PR_CAPTURE bit 7166 * without holding the EXCL lock on the page 7167 * as the PR_CAPTURE bit requres that the 7168 * page_capture_hash[].pchh_mutex be held 7169 * to modify it. 7170 */ 7171 page_clrtoxic(bp1->pp, PR_CAPTURE); 7172 mutex_exit(&page_capture_hash[i].pchh_mutex); 7173 kmem_free(bp1, sizeof (*bp1)); 7174 mutex_enter(&page_capture_hash[i].pchh_mutex); 7175 bp1 = page_capture_hash[i].lists[0].next; 7176 continue; 7177 } 7178 pp = bp1->pp; 7179 szc = bp1->szc; 7180 flags = bp1->flags; 7181 datap = bp1->datap; 7182 mutex_exit(&page_capture_hash[i].pchh_mutex); 7183 if (page_trylock(pp, SE_EXCL)) { 7184 ret = page_trycapture(pp, szc, 7185 flags | CAPTURE_ASYNC, datap); 7186 } else { 7187 ret = 1; /* move to walked hash */ 7188 } 7189 7190 if (ret != 0) { 7191 /* Move to walked hash */ 7192 (void) page_capture_move_to_walked(pp); 7193 } 7194 mutex_enter(&page_capture_hash[i].pchh_mutex); 7195 bp1 = page_capture_hash[i].lists[0].next; 7196 } 7197 7198 mutex_exit(&page_capture_hash[i].pchh_mutex); 7199 } 7200 } 7201 7202 /* 7203 * This function is called by the page_capture_thread, and is needed in 7204 * in order to initiate aio cleanup, so that pages used in aio 7205 * will be unlocked and subsequently retired by page_capture_thread. 7206 */ 7207 static int 7208 do_aio_cleanup(void) 7209 { 7210 proc_t *procp; 7211 int (*aio_cleanup_dr_delete_memory)(proc_t *); 7212 int cleaned = 0; 7213 7214 if (modload("sys", "kaio") == -1) { 7215 cmn_err(CE_WARN, "do_aio_cleanup: cannot load kaio"); 7216 return (0); 7217 } 7218 /* 7219 * We use the aio_cleanup_dr_delete_memory function to 7220 * initiate the actual clean up; this function will wake 7221 * up the per-process aio_cleanup_thread. 7222 */ 7223 aio_cleanup_dr_delete_memory = (int (*)(proc_t *)) 7224 modgetsymvalue("aio_cleanup_dr_delete_memory", 0); 7225 if (aio_cleanup_dr_delete_memory == NULL) { 7226 cmn_err(CE_WARN, 7227 "aio_cleanup_dr_delete_memory not found in kaio"); 7228 return (0); 7229 } 7230 mutex_enter(&pidlock); 7231 for (procp = practive; (procp != NULL); procp = procp->p_next) { 7232 mutex_enter(&procp->p_lock); 7233 if (procp->p_aio != NULL) { 7234 /* cleanup proc's outstanding kaio */ 7235 cleaned += (*aio_cleanup_dr_delete_memory)(procp); 7236 } 7237 mutex_exit(&procp->p_lock); 7238 } 7239 mutex_exit(&pidlock); 7240 return (cleaned); 7241 } 7242 7243 /* 7244 * helper function for page_capture_thread 7245 */ 7246 static void 7247 page_capture_handle_outstanding(void) 7248 { 7249 int ntry; 7250 7251 if (!page_retire_pend_count()) { 7252 /* 7253 * Do we really want to be this aggressive 7254 * for things other than page_retire? 7255 * Maybe have a counter for each callback 7256 * type to guide how aggressive we should 7257 * be here. Thus if there's at least one 7258 * page for page_retire we go ahead and reap 7259 * like this. 7260 */ 7261 kmem_reap(); 7262 seg_preap(); 7263 page_capture_async(); 7264 } else { 7265 /* 7266 * There are pages pending retirement, so 7267 * we reap prior to attempting to capture. 7268 */ 7269 kmem_reap(); 7270 7271 /* disable and purge seg_pcache */ 7272 (void) seg_p_disable(); 7273 for (ntry = 0; ntry < pc_thread_retry; ntry++) { 7274 if (!page_retire_pend_count()) 7275 break; 7276 if (do_aio_cleanup()) { 7277 /* 7278 * allow the apps cleanup threads 7279 * to run 7280 */ 7281 delay(pc_thread_shortwait); 7282 } 7283 page_capture_async(); 7284 } 7285 /* reenable seg_pcache */ 7286 seg_p_enable(); 7287 } 7288 } 7289 7290 /* 7291 * The page_capture_thread loops forever, looking to see if there are 7292 * pages still waiting to be captured. 7293 */ 7294 static void 7295 page_capture_thread(void) 7296 { 7297 callb_cpr_t c; 7298 int outstanding; 7299 int i; 7300 7301 CALLB_CPR_INIT(&c, &pc_thread_mutex, callb_generic_cpr, "page_capture"); 7302 7303 mutex_enter(&pc_thread_mutex); 7304 for (;;) { 7305 outstanding = 0; 7306 for (i = 0; i < NUM_PAGE_CAPTURE_BUCKETS; i++) 7307 outstanding += page_capture_hash[i].num_pages; 7308 if (outstanding) { 7309 page_capture_handle_outstanding(); 7310 CALLB_CPR_SAFE_BEGIN(&c); 7311 (void) cv_timedwait(&pc_cv, &pc_thread_mutex, 7312 lbolt + pc_thread_shortwait); 7313 CALLB_CPR_SAFE_END(&c, &pc_thread_mutex); 7314 } else { 7315 CALLB_CPR_SAFE_BEGIN(&c); 7316 (void) cv_timedwait(&pc_cv, &pc_thread_mutex, 7317 lbolt + pc_thread_longwait); 7318 CALLB_CPR_SAFE_END(&c, &pc_thread_mutex); 7319 } 7320 } 7321 /*NOTREACHED*/ 7322 } 7323 /* 7324 * Attempt to locate a bucket that has enough pages to satisfy the request. 7325 * The initial check is done without the lock to avoid unneeded contention. 7326 * The function returns 1 if enough pages were found, else 0 if it could not 7327 * find enough pages in a bucket. 7328 */ 7329 static int 7330 pcf_decrement_bucket(pgcnt_t npages) 7331 { 7332 struct pcf *p; 7333 struct pcf *q; 7334 int i; 7335 7336 p = &pcf[PCF_INDEX()]; 7337 q = &pcf[pcf_fanout]; 7338 for (i = 0; i < pcf_fanout; i++) { 7339 if (p->pcf_count > npages) { 7340 /* 7341 * a good one to try. 7342 */ 7343 mutex_enter(&p->pcf_lock); 7344 if (p->pcf_count > npages) { 7345 p->pcf_count -= (uint_t)npages; 7346 /* 7347 * freemem is not protected by any lock. 7348 * Thus, we cannot have any assertion 7349 * containing freemem here. 7350 */ 7351 freemem -= npages; 7352 mutex_exit(&p->pcf_lock); 7353 return (1); 7354 } 7355 mutex_exit(&p->pcf_lock); 7356 } 7357 p++; 7358 if (p >= q) { 7359 p = pcf; 7360 } 7361 } 7362 return (0); 7363 } 7364 7365 /* 7366 * Arguments: 7367 * pcftotal_ret: If the value is not NULL and we have walked all the 7368 * buckets but did not find enough pages then it will 7369 * be set to the total number of pages in all the pcf 7370 * buckets. 7371 * npages: Is the number of pages we have been requested to 7372 * find. 7373 * unlock: If set to 0 we will leave the buckets locked if the 7374 * requested number of pages are not found. 7375 * 7376 * Go and try to satisfy the page request from any number of buckets. 7377 * This can be a very expensive operation as we have to lock the buckets 7378 * we are checking (and keep them locked), starting at bucket 0. 7379 * 7380 * The function returns 1 if enough pages were found, else 0 if it could not 7381 * find enough pages in the buckets. 7382 * 7383 */ 7384 static int 7385 pcf_decrement_multiple(pgcnt_t *pcftotal_ret, pgcnt_t npages, int unlock) 7386 { 7387 struct pcf *p; 7388 pgcnt_t pcftotal; 7389 int i; 7390 7391 p = pcf; 7392 /* try to collect pages from several pcf bins */ 7393 for (pcftotal = 0, i = 0; i < pcf_fanout; i++) { 7394 mutex_enter(&p->pcf_lock); 7395 pcftotal += p->pcf_count; 7396 if (pcftotal >= npages) { 7397 /* 7398 * Wow! There are enough pages laying around 7399 * to satisfy the request. Do the accounting, 7400 * drop the locks we acquired, and go back. 7401 * 7402 * freemem is not protected by any lock. So, 7403 * we cannot have any assertion containing 7404 * freemem. 7405 */ 7406 freemem -= npages; 7407 while (p >= pcf) { 7408 if (p->pcf_count <= npages) { 7409 npages -= p->pcf_count; 7410 p->pcf_count = 0; 7411 } else { 7412 p->pcf_count -= (uint_t)npages; 7413 npages = 0; 7414 } 7415 mutex_exit(&p->pcf_lock); 7416 p--; 7417 } 7418 ASSERT(npages == 0); 7419 return (1); 7420 } 7421 p++; 7422 } 7423 if (unlock) { 7424 /* failed to collect pages - release the locks */ 7425 while (--p >= pcf) { 7426 mutex_exit(&p->pcf_lock); 7427 } 7428 } 7429 if (pcftotal_ret != NULL) 7430 *pcftotal_ret = pcftotal; 7431 return (0); 7432 } 7433