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) 1984, 1986, 1987, 1988, 1989 AT&T */ 27 /* All Rights Reserved */ 28 29 /* 30 * Portions of this source code were derived from Berkeley 4.3 BSD 31 * under license from the Regents of the University of California. 32 */ 33 34 #pragma ident "%Z%%M% %I% %E% SMI" 35 36 /* 37 * UNIX machine dependent virtual memory support. 38 */ 39 40 #include <sys/types.h> 41 #include <sys/param.h> 42 #include <sys/systm.h> 43 #include <sys/user.h> 44 #include <sys/proc.h> 45 #include <sys/kmem.h> 46 #include <sys/vmem.h> 47 #include <sys/buf.h> 48 #include <sys/cpuvar.h> 49 #include <sys/lgrp.h> 50 #include <sys/disp.h> 51 #include <sys/vm.h> 52 #include <sys/mman.h> 53 #include <sys/vnode.h> 54 #include <sys/cred.h> 55 #include <sys/exec.h> 56 #include <sys/exechdr.h> 57 #include <sys/debug.h> 58 #include <sys/vmsystm.h> 59 60 #include <vm/hat.h> 61 #include <vm/as.h> 62 #include <vm/seg.h> 63 #include <vm/seg_kp.h> 64 #include <vm/seg_vn.h> 65 #include <vm/page.h> 66 #include <vm/seg_kmem.h> 67 #include <vm/seg_kpm.h> 68 #include <vm/vm_dep.h> 69 70 #include <sys/cpu.h> 71 #include <sys/vm_machparam.h> 72 #include <sys/memlist.h> 73 #include <sys/bootconf.h> /* XXX the memlist stuff belongs in memlist_plat.h */ 74 #include <vm/hat_i86.h> 75 #include <sys/x86_archext.h> 76 #include <sys/elf_386.h> 77 #include <sys/cmn_err.h> 78 #include <sys/archsystm.h> 79 #include <sys/machsystm.h> 80 81 #include <sys/vtrace.h> 82 #include <sys/ddidmareq.h> 83 #include <sys/promif.h> 84 #include <sys/memnode.h> 85 #include <sys/stack.h> 86 #include <util/qsort.h> 87 #include <sys/taskq.h> 88 89 #ifdef __xpv 90 91 #include <sys/hypervisor.h> 92 #include <sys/xen_mmu.h> 93 #include <sys/balloon_impl.h> 94 95 /* 96 * domain 0 pages usable for DMA are kept pre-allocated and kept in 97 * distinct lists, ordered by increasing mfn. 98 */ 99 static kmutex_t io_pool_lock; 100 static kmutex_t contig_list_lock; 101 static page_t *io_pool_4g; /* pool for 32 bit dma limited devices */ 102 static page_t *io_pool_16m; /* pool for 24 bit dma limited legacy devices */ 103 static long io_pool_cnt; 104 static long io_pool_cnt_max = 0; 105 #define DEFAULT_IO_POOL_MIN 128 106 static long io_pool_cnt_min = DEFAULT_IO_POOL_MIN; 107 static long io_pool_cnt_lowater = 0; 108 static long io_pool_shrink_attempts; /* how many times did we try to shrink */ 109 static long io_pool_shrinks; /* how many times did we really shrink */ 110 static long io_pool_grows; /* how many times did we grow */ 111 static mfn_t start_mfn = 1; 112 static caddr_t io_pool_kva; /* use to alloc pages when needed */ 113 114 static int create_contig_pfnlist(uint_t); 115 116 /* 117 * percentage of phys mem to hold in the i/o pool 118 */ 119 #define DEFAULT_IO_POOL_PCT 2 120 static long io_pool_physmem_pct = DEFAULT_IO_POOL_PCT; 121 static void page_io_pool_sub(page_t **, page_t *, page_t *); 122 int ioalloc_dbg = 0; 123 124 #endif /* __xpv */ 125 126 uint_t vac_colors = 1; 127 128 int largepagesupport = 0; 129 extern uint_t page_create_new; 130 extern uint_t page_create_exists; 131 extern uint_t page_create_putbacks; 132 extern uint_t page_create_putbacks; 133 /* 134 * Allow users to disable the kernel's use of SSE. 135 */ 136 extern int use_sse_pagecopy, use_sse_pagezero; 137 138 /* 139 * combined memory ranges from mnode and memranges[] to manage single 140 * mnode/mtype dimension in the page lists. 141 */ 142 typedef struct { 143 pfn_t mnr_pfnlo; 144 pfn_t mnr_pfnhi; 145 int mnr_mnode; 146 int mnr_memrange; /* index into memranges[] */ 147 /* maintain page list stats */ 148 pgcnt_t mnr_mt_clpgcnt; /* cache list cnt */ 149 pgcnt_t mnr_mt_flpgcnt[MMU_PAGE_SIZES]; /* free list cnt per szc */ 150 pgcnt_t mnr_mt_totcnt; /* sum of cache and free lists */ 151 #ifdef DEBUG 152 struct mnr_mts { /* mnode/mtype szc stats */ 153 pgcnt_t mnr_mts_pgcnt; 154 int mnr_mts_colors; 155 pgcnt_t *mnr_mtsc_pgcnt; 156 } *mnr_mts; 157 #endif 158 } mnoderange_t; 159 160 #define MEMRANGEHI(mtype) \ 161 ((mtype > 0) ? memranges[mtype - 1] - 1: physmax) 162 #define MEMRANGELO(mtype) (memranges[mtype]) 163 164 #define MTYPE_FREEMEM(mt) (mnoderanges[mt].mnr_mt_totcnt) 165 166 /* 167 * As the PC architecture evolved memory up was clumped into several 168 * ranges for various historical I/O devices to do DMA. 169 * < 16Meg - ISA bus 170 * < 2Gig - ??? 171 * < 4Gig - PCI bus or drivers that don't understand PAE mode 172 * 173 * These are listed in reverse order, so that we can skip over unused 174 * ranges on machines with small memories. 175 * 176 * For now under the Hypervisor, we'll only ever have one memrange. 177 */ 178 #define PFN_4GIG 0x100000 179 #define PFN_16MEG 0x1000 180 static pfn_t arch_memranges[NUM_MEM_RANGES] = { 181 PFN_4GIG, /* pfn range for 4G and above */ 182 0x80000, /* pfn range for 2G-4G */ 183 PFN_16MEG, /* pfn range for 16M-2G */ 184 0x00000, /* pfn range for 0-16M */ 185 }; 186 pfn_t *memranges = &arch_memranges[0]; 187 int nranges = NUM_MEM_RANGES; 188 189 /* 190 * This combines mem_node_config and memranges into one data 191 * structure to be used for page list management. 192 */ 193 mnoderange_t *mnoderanges; 194 int mnoderangecnt; 195 int mtype4g; 196 197 /* 198 * 4g memory management variables for systems with more than 4g of memory: 199 * 200 * physical memory below 4g is required for 32bit dma devices and, currently, 201 * for kmem memory. On systems with more than 4g of memory, the pool of memory 202 * below 4g can be depleted without any paging activity given that there is 203 * likely to be sufficient memory above 4g. 204 * 205 * physmax4g is set true if the largest pfn is over 4g. The rest of the 206 * 4g memory management code is enabled only when physmax4g is true. 207 * 208 * maxmem4g is the count of the maximum number of pages on the page lists 209 * with physical addresses below 4g. It can be a lot less then 4g given that 210 * BIOS may reserve large chunks of space below 4g for hot plug pci devices, 211 * agp aperture etc. 212 * 213 * freemem4g maintains the count of the number of available pages on the 214 * page lists with physical addresses below 4g. 215 * 216 * DESFREE4G specifies the desired amount of below 4g memory. It defaults to 217 * 6% (desfree4gshift = 4) of maxmem4g. 218 * 219 * RESTRICT4G_ALLOC returns true if freemem4g falls below DESFREE4G 220 * and the amount of physical memory above 4g is greater than freemem4g. 221 * In this case, page_get_* routines will restrict below 4g allocations 222 * for requests that don't specifically require it. 223 */ 224 225 #define LOTSFREE4G (maxmem4g >> lotsfree4gshift) 226 #define DESFREE4G (maxmem4g >> desfree4gshift) 227 228 #define RESTRICT4G_ALLOC \ 229 (physmax4g && (freemem4g < DESFREE4G) && ((freemem4g << 1) < freemem)) 230 231 static pgcnt_t maxmem4g; 232 static pgcnt_t freemem4g; 233 static int physmax4g; 234 static int desfree4gshift = 4; /* maxmem4g shift to derive DESFREE4G */ 235 static int lotsfree4gshift = 3; 236 237 /* 238 * 16m memory management: 239 * 240 * reserve some amount of physical memory below 16m for legacy devices. 241 * 242 * RESTRICT16M_ALLOC returns true if an there are sufficient free pages above 243 * 16m or if the 16m pool drops below DESFREE16M. 244 * 245 * In this case, general page allocations via page_get_{free,cache}list 246 * routines will be restricted from allocating from the 16m pool. Allocations 247 * that require specific pfn ranges (page_get_anylist) and PG_PANIC allocations 248 * are not restricted. 249 */ 250 251 #define FREEMEM16M MTYPE_FREEMEM(0) 252 #define DESFREE16M desfree16m 253 #define RESTRICT16M_ALLOC(freemem, pgcnt, flags) \ 254 ((freemem != 0) && ((flags & PG_PANIC) == 0) && \ 255 ((freemem >= (FREEMEM16M)) || \ 256 (FREEMEM16M < (DESFREE16M + pgcnt)))) 257 258 static pgcnt_t desfree16m = 0x380; 259 260 /* 261 * This can be patched via /etc/system to allow old non-PAE aware device 262 * drivers to use kmem_alloc'd memory on 32 bit systems with > 4Gig RAM. 263 */ 264 int restricted_kmemalloc = 0; 265 266 #ifdef VM_STATS 267 struct { 268 ulong_t pga_alloc; 269 ulong_t pga_notfullrange; 270 ulong_t pga_nulldmaattr; 271 ulong_t pga_allocok; 272 ulong_t pga_allocfailed; 273 ulong_t pgma_alloc; 274 ulong_t pgma_allocok; 275 ulong_t pgma_allocfailed; 276 ulong_t pgma_allocempty; 277 } pga_vmstats; 278 #endif 279 280 uint_t mmu_page_sizes; 281 282 /* How many page sizes the users can see */ 283 uint_t mmu_exported_page_sizes; 284 285 /* page sizes that legacy applications can see */ 286 uint_t mmu_legacy_page_sizes; 287 288 /* 289 * Number of pages in 1 GB. Don't enable automatic large pages if we have 290 * fewer than this many pages. 291 */ 292 pgcnt_t shm_lpg_min_physmem = 1 << (30 - MMU_PAGESHIFT); 293 pgcnt_t privm_lpg_min_physmem = 1 << (30 - MMU_PAGESHIFT); 294 295 /* 296 * Maximum and default segment size tunables for user private 297 * and shared anon memory, and user text and initialized data. 298 * These can be patched via /etc/system to allow large pages 299 * to be used for mapping application private and shared anon memory. 300 */ 301 size_t mcntl0_lpsize = MMU_PAGESIZE; 302 size_t max_uheap_lpsize = MMU_PAGESIZE; 303 size_t default_uheap_lpsize = MMU_PAGESIZE; 304 size_t max_ustack_lpsize = MMU_PAGESIZE; 305 size_t default_ustack_lpsize = MMU_PAGESIZE; 306 size_t max_privmap_lpsize = MMU_PAGESIZE; 307 size_t max_uidata_lpsize = MMU_PAGESIZE; 308 size_t max_utext_lpsize = MMU_PAGESIZE; 309 size_t max_shm_lpsize = MMU_PAGESIZE; 310 311 312 /* 313 * initialized by page_coloring_init(). 314 */ 315 uint_t page_colors; 316 uint_t page_colors_mask; 317 uint_t page_coloring_shift; 318 int cpu_page_colors; 319 static uint_t l2_colors; 320 321 /* 322 * Page freelists and cachelists are dynamically allocated once mnoderangecnt 323 * and page_colors are calculated from the l2 cache n-way set size. Within a 324 * mnode range, the page freelist and cachelist are hashed into bins based on 325 * color. This makes it easier to search for a page within a specific memory 326 * range. 327 */ 328 #define PAGE_COLORS_MIN 16 329 330 page_t ****page_freelists; 331 page_t ***page_cachelists; 332 333 334 /* 335 * Used by page layer to know about page sizes 336 */ 337 hw_pagesize_t hw_page_array[MAX_NUM_LEVEL + 1]; 338 339 kmutex_t *fpc_mutex[NPC_MUTEX]; 340 kmutex_t *cpc_mutex[NPC_MUTEX]; 341 342 /* 343 * Only let one thread at a time try to coalesce large pages, to 344 * prevent them from working against each other. 345 */ 346 static kmutex_t contig_lock; 347 #define CONTIG_LOCK() mutex_enter(&contig_lock); 348 #define CONTIG_UNLOCK() mutex_exit(&contig_lock); 349 350 #define PFN_16M (mmu_btop((uint64_t)0x1000000)) 351 352 /* 353 * Return the optimum page size for a given mapping 354 */ 355 /*ARGSUSED*/ 356 size_t 357 map_pgsz(int maptype, struct proc *p, caddr_t addr, size_t len, int memcntl) 358 { 359 level_t l = 0; 360 size_t pgsz = MMU_PAGESIZE; 361 size_t max_lpsize; 362 uint_t mszc; 363 364 ASSERT(maptype != MAPPGSZ_VA); 365 366 if (maptype != MAPPGSZ_ISM && physmem < privm_lpg_min_physmem) { 367 return (MMU_PAGESIZE); 368 } 369 370 switch (maptype) { 371 case MAPPGSZ_HEAP: 372 case MAPPGSZ_STK: 373 max_lpsize = memcntl ? mcntl0_lpsize : (maptype == 374 MAPPGSZ_HEAP ? max_uheap_lpsize : max_ustack_lpsize); 375 if (max_lpsize == MMU_PAGESIZE) { 376 return (MMU_PAGESIZE); 377 } 378 if (len == 0) { 379 len = (maptype == MAPPGSZ_HEAP) ? p->p_brkbase + 380 p->p_brksize - p->p_bssbase : p->p_stksize; 381 } 382 len = (maptype == MAPPGSZ_HEAP) ? MAX(len, 383 default_uheap_lpsize) : MAX(len, default_ustack_lpsize); 384 385 /* 386 * use the pages size that best fits len 387 */ 388 for (l = mmu.umax_page_level; l > 0; --l) { 389 if (LEVEL_SIZE(l) > max_lpsize || len < LEVEL_SIZE(l)) { 390 continue; 391 } else { 392 pgsz = LEVEL_SIZE(l); 393 } 394 break; 395 } 396 397 mszc = (maptype == MAPPGSZ_HEAP ? p->p_brkpageszc : 398 p->p_stkpageszc); 399 if (addr == 0 && (pgsz < hw_page_array[mszc].hp_size)) { 400 pgsz = hw_page_array[mszc].hp_size; 401 } 402 return (pgsz); 403 404 case MAPPGSZ_ISM: 405 for (l = mmu.umax_page_level; l > 0; --l) { 406 if (len >= LEVEL_SIZE(l)) 407 return (LEVEL_SIZE(l)); 408 } 409 return (LEVEL_SIZE(0)); 410 } 411 return (pgsz); 412 } 413 414 static uint_t 415 map_szcvec(caddr_t addr, size_t size, uintptr_t off, size_t max_lpsize, 416 size_t min_physmem) 417 { 418 caddr_t eaddr = addr + size; 419 uint_t szcvec = 0; 420 caddr_t raddr; 421 caddr_t readdr; 422 size_t pgsz; 423 int i; 424 425 if (physmem < min_physmem || max_lpsize <= MMU_PAGESIZE) { 426 return (0); 427 } 428 429 for (i = mmu_exported_page_sizes - 1; i > 0; i--) { 430 pgsz = page_get_pagesize(i); 431 if (pgsz > max_lpsize) { 432 continue; 433 } 434 raddr = (caddr_t)P2ROUNDUP((uintptr_t)addr, pgsz); 435 readdr = (caddr_t)P2ALIGN((uintptr_t)eaddr, pgsz); 436 if (raddr < addr || raddr >= readdr) { 437 continue; 438 } 439 if (P2PHASE((uintptr_t)addr ^ off, pgsz)) { 440 continue; 441 } 442 /* 443 * Set szcvec to the remaining page sizes. 444 */ 445 szcvec = ((1 << (i + 1)) - 1) & ~1; 446 break; 447 } 448 return (szcvec); 449 } 450 451 /* 452 * Return a bit vector of large page size codes that 453 * can be used to map [addr, addr + len) region. 454 */ 455 /*ARGSUSED*/ 456 uint_t 457 map_pgszcvec(caddr_t addr, size_t size, uintptr_t off, int flags, int type, 458 int memcntl) 459 { 460 size_t max_lpsize = mcntl0_lpsize; 461 462 if (mmu.max_page_level == 0) 463 return (0); 464 465 if (flags & MAP_TEXT) { 466 if (!memcntl) 467 max_lpsize = max_utext_lpsize; 468 return (map_szcvec(addr, size, off, max_lpsize, 469 shm_lpg_min_physmem)); 470 471 } else if (flags & MAP_INITDATA) { 472 if (!memcntl) 473 max_lpsize = max_uidata_lpsize; 474 return (map_szcvec(addr, size, off, max_lpsize, 475 privm_lpg_min_physmem)); 476 477 } else if (type == MAPPGSZC_SHM) { 478 if (!memcntl) 479 max_lpsize = max_shm_lpsize; 480 return (map_szcvec(addr, size, off, max_lpsize, 481 shm_lpg_min_physmem)); 482 483 } else if (type == MAPPGSZC_HEAP) { 484 if (!memcntl) 485 max_lpsize = max_uheap_lpsize; 486 return (map_szcvec(addr, size, off, max_lpsize, 487 privm_lpg_min_physmem)); 488 489 } else if (type == MAPPGSZC_STACK) { 490 if (!memcntl) 491 max_lpsize = max_ustack_lpsize; 492 return (map_szcvec(addr, size, off, max_lpsize, 493 privm_lpg_min_physmem)); 494 495 } else { 496 if (!memcntl) 497 max_lpsize = max_privmap_lpsize; 498 return (map_szcvec(addr, size, off, max_lpsize, 499 privm_lpg_min_physmem)); 500 } 501 } 502 503 /* 504 * Handle a pagefault. 505 */ 506 faultcode_t 507 pagefault( 508 caddr_t addr, 509 enum fault_type type, 510 enum seg_rw rw, 511 int iskernel) 512 { 513 struct as *as; 514 struct hat *hat; 515 struct proc *p; 516 kthread_t *t; 517 faultcode_t res; 518 caddr_t base; 519 size_t len; 520 int err; 521 int mapped_red; 522 uintptr_t ea; 523 524 ASSERT_STACK_ALIGNED(); 525 526 if (INVALID_VADDR(addr)) 527 return (FC_NOMAP); 528 529 mapped_red = segkp_map_red(); 530 531 if (iskernel) { 532 as = &kas; 533 hat = as->a_hat; 534 } else { 535 t = curthread; 536 p = ttoproc(t); 537 as = p->p_as; 538 hat = as->a_hat; 539 } 540 541 /* 542 * Dispatch pagefault. 543 */ 544 res = as_fault(hat, as, addr, 1, type, rw); 545 546 /* 547 * If this isn't a potential unmapped hole in the user's 548 * UNIX data or stack segments, just return status info. 549 */ 550 if (res != FC_NOMAP || iskernel) 551 goto out; 552 553 /* 554 * Check to see if we happened to faulted on a currently unmapped 555 * part of the UNIX data or stack segments. If so, create a zfod 556 * mapping there and then try calling the fault routine again. 557 */ 558 base = p->p_brkbase; 559 len = p->p_brksize; 560 561 if (addr < base || addr >= base + len) { /* data seg? */ 562 base = (caddr_t)p->p_usrstack - p->p_stksize; 563 len = p->p_stksize; 564 if (addr < base || addr >= p->p_usrstack) { /* stack seg? */ 565 /* not in either UNIX data or stack segments */ 566 res = FC_NOMAP; 567 goto out; 568 } 569 } 570 571 /* 572 * the rest of this function implements a 3.X 4.X 5.X compatibility 573 * This code is probably not needed anymore 574 */ 575 if (p->p_model == DATAMODEL_ILP32) { 576 577 /* expand the gap to the page boundaries on each side */ 578 ea = P2ROUNDUP((uintptr_t)base + len, MMU_PAGESIZE); 579 base = (caddr_t)P2ALIGN((uintptr_t)base, MMU_PAGESIZE); 580 len = ea - (uintptr_t)base; 581 582 as_rangelock(as); 583 if (as_gap(as, MMU_PAGESIZE, &base, &len, AH_CONTAIN, addr) == 584 0) { 585 err = as_map(as, base, len, segvn_create, zfod_argsp); 586 as_rangeunlock(as); 587 if (err) { 588 res = FC_MAKE_ERR(err); 589 goto out; 590 } 591 } else { 592 /* 593 * This page is already mapped by another thread after 594 * we returned from as_fault() above. We just fall 595 * through as_fault() below. 596 */ 597 as_rangeunlock(as); 598 } 599 600 res = as_fault(hat, as, addr, 1, F_INVAL, rw); 601 } 602 603 out: 604 if (mapped_red) 605 segkp_unmap_red(); 606 607 return (res); 608 } 609 610 void 611 map_addr(caddr_t *addrp, size_t len, offset_t off, int vacalign, uint_t flags) 612 { 613 struct proc *p = curproc; 614 caddr_t userlimit = (flags & _MAP_LOW32) ? 615 (caddr_t)_userlimit32 : p->p_as->a_userlimit; 616 617 map_addr_proc(addrp, len, off, vacalign, userlimit, curproc, flags); 618 } 619 620 /*ARGSUSED*/ 621 int 622 map_addr_vacalign_check(caddr_t addr, u_offset_t off) 623 { 624 return (0); 625 } 626 627 /* 628 * map_addr_proc() is the routine called when the system is to 629 * choose an address for the user. We will pick an address 630 * range which is the highest available below userlimit. 631 * 632 * Every mapping will have a redzone of a single page on either side of 633 * the request. This is done to leave one page unmapped between segments. 634 * This is not required, but it's useful for the user because if their 635 * program strays across a segment boundary, it will catch a fault 636 * immediately making debugging a little easier. Currently the redzone 637 * is mandatory. 638 * 639 * addrp is a value/result parameter. 640 * On input it is a hint from the user to be used in a completely 641 * machine dependent fashion. We decide to completely ignore this hint. 642 * If MAP_ALIGN was specified, addrp contains the minimal alignment, which 643 * must be some "power of two" multiple of pagesize. 644 * 645 * On output it is NULL if no address can be found in the current 646 * processes address space or else an address that is currently 647 * not mapped for len bytes with a page of red zone on either side. 648 * 649 * vacalign is not needed on x86 (it's for viturally addressed caches) 650 */ 651 /*ARGSUSED*/ 652 void 653 map_addr_proc( 654 caddr_t *addrp, 655 size_t len, 656 offset_t off, 657 int vacalign, 658 caddr_t userlimit, 659 struct proc *p, 660 uint_t flags) 661 { 662 struct as *as = p->p_as; 663 caddr_t addr; 664 caddr_t base; 665 size_t slen; 666 size_t align_amount; 667 668 ASSERT32(userlimit == as->a_userlimit); 669 670 base = p->p_brkbase; 671 #if defined(__amd64) 672 /* 673 * XX64 Yes, this needs more work. 674 */ 675 if (p->p_model == DATAMODEL_NATIVE) { 676 if (userlimit < as->a_userlimit) { 677 /* 678 * This happens when a program wants to map 679 * something in a range that's accessible to a 680 * program in a smaller address space. For example, 681 * a 64-bit program calling mmap32(2) to guarantee 682 * that the returned address is below 4Gbytes. 683 */ 684 ASSERT((uintptr_t)userlimit < ADDRESS_C(0xffffffff)); 685 686 if (userlimit > base) 687 slen = userlimit - base; 688 else { 689 *addrp = NULL; 690 return; 691 } 692 } else { 693 /* 694 * XX64 This layout is probably wrong .. but in 695 * the event we make the amd64 address space look 696 * like sparcv9 i.e. with the stack -above- the 697 * heap, this bit of code might even be correct. 698 */ 699 slen = p->p_usrstack - base - 700 (((size_t)rctl_enforced_value( 701 rctlproc_legacy[RLIMIT_STACK], 702 p->p_rctls, p) + PAGEOFFSET) & PAGEMASK); 703 } 704 } else 705 #endif 706 slen = userlimit - base; 707 708 /* Make len be a multiple of PAGESIZE */ 709 len = (len + PAGEOFFSET) & PAGEMASK; 710 711 /* 712 * figure out what the alignment should be 713 * 714 * XX64 -- is there an ELF_AMD64_MAXPGSZ or is it the same???? 715 */ 716 if (len <= ELF_386_MAXPGSZ) { 717 /* 718 * Align virtual addresses to ensure that ELF shared libraries 719 * are mapped with the appropriate alignment constraints by 720 * the run-time linker. 721 */ 722 align_amount = ELF_386_MAXPGSZ; 723 } else { 724 int l = mmu.umax_page_level; 725 726 while (l && len < LEVEL_SIZE(l)) 727 --l; 728 729 align_amount = LEVEL_SIZE(l); 730 } 731 732 if ((flags & MAP_ALIGN) && ((uintptr_t)*addrp > align_amount)) 733 align_amount = (uintptr_t)*addrp; 734 735 ASSERT(ISP2(align_amount)); 736 ASSERT(align_amount == 0 || align_amount >= PAGESIZE); 737 738 off = off & (align_amount - 1); 739 /* 740 * Look for a large enough hole starting below userlimit. 741 * After finding it, use the upper part. 742 */ 743 if (as_gap_aligned(as, len, &base, &slen, AH_HI, NULL, align_amount, 744 PAGESIZE, off) == 0) { 745 caddr_t as_addr; 746 747 /* 748 * addr is the highest possible address to use since we have 749 * a PAGESIZE redzone at the beginning and end. 750 */ 751 addr = base + slen - (PAGESIZE + len); 752 as_addr = addr; 753 /* 754 * Round address DOWN to the alignment amount and 755 * add the offset in. 756 * If addr is greater than as_addr, len would not be large 757 * enough to include the redzone, so we must adjust down 758 * by the alignment amount. 759 */ 760 addr = (caddr_t)((uintptr_t)addr & (~(align_amount - 1))); 761 addr += (uintptr_t)off; 762 if (addr > as_addr) { 763 addr -= align_amount; 764 } 765 766 ASSERT(addr > base); 767 ASSERT(addr + len < base + slen); 768 ASSERT(((uintptr_t)addr & (align_amount - 1)) == 769 ((uintptr_t)(off))); 770 *addrp = addr; 771 } else { 772 *addrp = NULL; /* no more virtual space */ 773 } 774 } 775 776 int valid_va_range_aligned_wraparound; 777 778 /* 779 * Determine whether [*basep, *basep + *lenp) contains a mappable range of 780 * addresses at least "minlen" long, where the base of the range is at "off" 781 * phase from an "align" boundary and there is space for a "redzone"-sized 782 * redzone on either side of the range. On success, 1 is returned and *basep 783 * and *lenp are adjusted to describe the acceptable range (including 784 * the redzone). On failure, 0 is returned. 785 */ 786 /*ARGSUSED3*/ 787 int 788 valid_va_range_aligned(caddr_t *basep, size_t *lenp, size_t minlen, int dir, 789 size_t align, size_t redzone, size_t off) 790 { 791 uintptr_t hi, lo; 792 size_t tot_len; 793 794 ASSERT(align == 0 ? off == 0 : off < align); 795 ASSERT(ISP2(align)); 796 ASSERT(align == 0 || align >= PAGESIZE); 797 798 lo = (uintptr_t)*basep; 799 hi = lo + *lenp; 800 tot_len = minlen + 2 * redzone; /* need at least this much space */ 801 802 /* 803 * If hi rolled over the top, try cutting back. 804 */ 805 if (hi < lo) { 806 *lenp = 0UL - lo - 1UL; 807 /* See if this really happens. If so, then we figure out why */ 808 valid_va_range_aligned_wraparound++; 809 hi = lo + *lenp; 810 } 811 if (*lenp < tot_len) { 812 return (0); 813 } 814 815 #if defined(__amd64) 816 /* 817 * Deal with a possible hole in the address range between 818 * hole_start and hole_end that should never be mapped. 819 */ 820 if (lo < hole_start) { 821 if (hi > hole_start) { 822 if (hi < hole_end) { 823 hi = hole_start; 824 } else { 825 /* lo < hole_start && hi >= hole_end */ 826 if (dir == AH_LO) { 827 /* 828 * prefer lowest range 829 */ 830 if (hole_start - lo >= tot_len) 831 hi = hole_start; 832 else if (hi - hole_end >= tot_len) 833 lo = hole_end; 834 else 835 return (0); 836 } else { 837 /* 838 * prefer highest range 839 */ 840 if (hi - hole_end >= tot_len) 841 lo = hole_end; 842 else if (hole_start - lo >= tot_len) 843 hi = hole_start; 844 else 845 return (0); 846 } 847 } 848 } 849 } else { 850 /* lo >= hole_start */ 851 if (hi < hole_end) 852 return (0); 853 if (lo < hole_end) 854 lo = hole_end; 855 } 856 #endif 857 858 if (hi - lo < tot_len) 859 return (0); 860 861 if (align > 1) { 862 uintptr_t tlo = lo + redzone; 863 uintptr_t thi = hi - redzone; 864 tlo = (uintptr_t)P2PHASEUP(tlo, align, off); 865 if (tlo < lo + redzone) { 866 return (0); 867 } 868 if (thi < tlo || thi - tlo < minlen) { 869 return (0); 870 } 871 } 872 873 *basep = (caddr_t)lo; 874 *lenp = hi - lo; 875 return (1); 876 } 877 878 /* 879 * Determine whether [*basep, *basep + *lenp) contains a mappable range of 880 * addresses at least "minlen" long. On success, 1 is returned and *basep 881 * and *lenp are adjusted to describe the acceptable range. On failure, 0 882 * is returned. 883 */ 884 int 885 valid_va_range(caddr_t *basep, size_t *lenp, size_t minlen, int dir) 886 { 887 return (valid_va_range_aligned(basep, lenp, minlen, dir, 0, 0, 0)); 888 } 889 890 /* 891 * Determine whether [addr, addr+len] are valid user addresses. 892 */ 893 /*ARGSUSED*/ 894 int 895 valid_usr_range(caddr_t addr, size_t len, uint_t prot, struct as *as, 896 caddr_t userlimit) 897 { 898 caddr_t eaddr = addr + len; 899 900 if (eaddr <= addr || addr >= userlimit || eaddr > userlimit) 901 return (RANGE_BADADDR); 902 903 #if defined(__amd64) 904 /* 905 * Check for the VA hole 906 */ 907 if (eaddr > (caddr_t)hole_start && addr < (caddr_t)hole_end) 908 return (RANGE_BADADDR); 909 #endif 910 911 return (RANGE_OKAY); 912 } 913 914 /* 915 * Return 1 if the page frame is onboard memory, else 0. 916 */ 917 int 918 pf_is_memory(pfn_t pf) 919 { 920 if (pfn_is_foreign(pf)) 921 return (0); 922 return (address_in_memlist(phys_install, pfn_to_pa(pf), 1)); 923 } 924 925 /* 926 * return the memrange containing pfn 927 */ 928 int 929 memrange_num(pfn_t pfn) 930 { 931 int n; 932 933 for (n = 0; n < nranges - 1; ++n) { 934 if (pfn >= memranges[n]) 935 break; 936 } 937 return (n); 938 } 939 940 /* 941 * return the mnoderange containing pfn 942 */ 943 /*ARGSUSED*/ 944 int 945 pfn_2_mtype(pfn_t pfn) 946 { 947 #if defined(__xpv) 948 return (0); 949 #else 950 int n; 951 952 for (n = mnoderangecnt - 1; n >= 0; n--) { 953 if (pfn >= mnoderanges[n].mnr_pfnlo) { 954 break; 955 } 956 } 957 return (n); 958 #endif 959 } 960 961 #if !defined(__xpv) 962 /* 963 * is_contigpage_free: 964 * returns a page list of contiguous pages. It minimally has to return 965 * minctg pages. Caller determines minctg based on the scatter-gather 966 * list length. 967 * 968 * pfnp is set to the next page frame to search on return. 969 */ 970 static page_t * 971 is_contigpage_free( 972 pfn_t *pfnp, 973 pgcnt_t *pgcnt, 974 pgcnt_t minctg, 975 uint64_t pfnseg, 976 int iolock) 977 { 978 int i = 0; 979 pfn_t pfn = *pfnp; 980 page_t *pp; 981 page_t *plist = NULL; 982 983 /* 984 * fail if pfn + minctg crosses a segment boundary. 985 * Adjust for next starting pfn to begin at segment boundary. 986 */ 987 988 if (((*pfnp + minctg - 1) & pfnseg) < (*pfnp & pfnseg)) { 989 *pfnp = roundup(*pfnp, pfnseg + 1); 990 return (NULL); 991 } 992 993 do { 994 retry: 995 pp = page_numtopp_nolock(pfn + i); 996 if ((pp == NULL) || 997 (page_trylock(pp, SE_EXCL) == 0)) { 998 (*pfnp)++; 999 break; 1000 } 1001 if (page_pptonum(pp) != pfn + i) { 1002 page_unlock(pp); 1003 goto retry; 1004 } 1005 1006 if (!(PP_ISFREE(pp))) { 1007 page_unlock(pp); 1008 (*pfnp)++; 1009 break; 1010 } 1011 1012 if (!PP_ISAGED(pp)) { 1013 page_list_sub(pp, PG_CACHE_LIST); 1014 page_hashout(pp, (kmutex_t *)NULL); 1015 } else { 1016 page_list_sub(pp, PG_FREE_LIST); 1017 } 1018 1019 if (iolock) 1020 page_io_lock(pp); 1021 page_list_concat(&plist, &pp); 1022 1023 /* 1024 * exit loop when pgcnt satisfied or segment boundary reached. 1025 */ 1026 1027 } while ((++i < *pgcnt) && ((pfn + i) & pfnseg)); 1028 1029 *pfnp += i; /* set to next pfn to search */ 1030 1031 if (i >= minctg) { 1032 *pgcnt -= i; 1033 return (plist); 1034 } 1035 1036 /* 1037 * failure: minctg not satisfied. 1038 * 1039 * if next request crosses segment boundary, set next pfn 1040 * to search from the segment boundary. 1041 */ 1042 if (((*pfnp + minctg - 1) & pfnseg) < (*pfnp & pfnseg)) 1043 *pfnp = roundup(*pfnp, pfnseg + 1); 1044 1045 /* clean up any pages already allocated */ 1046 1047 while (plist) { 1048 pp = plist; 1049 page_sub(&plist, pp); 1050 page_list_add(pp, PG_FREE_LIST | PG_LIST_TAIL); 1051 if (iolock) 1052 page_io_unlock(pp); 1053 page_unlock(pp); 1054 } 1055 1056 return (NULL); 1057 } 1058 #endif /* !__xpv */ 1059 1060 /* 1061 * verify that pages being returned from allocator have correct DMA attribute 1062 */ 1063 #ifndef DEBUG 1064 #define check_dma(a, b, c) (0) 1065 #else 1066 static void 1067 check_dma(ddi_dma_attr_t *dma_attr, page_t *pp, int cnt) 1068 { 1069 if (dma_attr == NULL) 1070 return; 1071 1072 while (cnt-- > 0) { 1073 if (pa_to_ma(pfn_to_pa(pp->p_pagenum)) < 1074 dma_attr->dma_attr_addr_lo) 1075 panic("PFN (pp=%p) below dma_attr_addr_lo", pp); 1076 if (pa_to_ma(pfn_to_pa(pp->p_pagenum)) >= 1077 dma_attr->dma_attr_addr_hi) 1078 panic("PFN (pp=%p) above dma_attr_addr_hi", pp); 1079 pp = pp->p_next; 1080 } 1081 } 1082 #endif 1083 1084 #if !defined(__xpv) 1085 static page_t * 1086 page_get_contigpage(pgcnt_t *pgcnt, ddi_dma_attr_t *mattr, int iolock) 1087 { 1088 pfn_t pfn; 1089 int sgllen; 1090 uint64_t pfnseg; 1091 pgcnt_t minctg; 1092 page_t *pplist = NULL, *plist; 1093 uint64_t lo, hi; 1094 pgcnt_t pfnalign = 0; 1095 static pfn_t startpfn; 1096 static pgcnt_t lastctgcnt; 1097 uintptr_t align; 1098 1099 CONTIG_LOCK(); 1100 1101 if (mattr) { 1102 lo = mmu_btop((mattr->dma_attr_addr_lo + MMU_PAGEOFFSET)); 1103 hi = mmu_btop(mattr->dma_attr_addr_hi); 1104 if (hi >= physmax) 1105 hi = physmax - 1; 1106 sgllen = mattr->dma_attr_sgllen; 1107 pfnseg = mmu_btop(mattr->dma_attr_seg); 1108 1109 align = maxbit(mattr->dma_attr_align, mattr->dma_attr_minxfer); 1110 if (align > MMU_PAGESIZE) 1111 pfnalign = mmu_btop(align); 1112 1113 /* 1114 * in order to satisfy the request, must minimally 1115 * acquire minctg contiguous pages 1116 */ 1117 minctg = howmany(*pgcnt, sgllen); 1118 1119 ASSERT(hi >= lo); 1120 1121 /* 1122 * start from where last searched if the minctg >= lastctgcnt 1123 */ 1124 if (minctg < lastctgcnt || startpfn < lo || startpfn > hi) 1125 startpfn = lo; 1126 } else { 1127 hi = physmax - 1; 1128 lo = 0; 1129 sgllen = 1; 1130 pfnseg = mmu.highest_pfn; 1131 minctg = *pgcnt; 1132 1133 if (minctg < lastctgcnt) 1134 startpfn = lo; 1135 } 1136 lastctgcnt = minctg; 1137 1138 ASSERT(pfnseg + 1 >= (uint64_t)minctg); 1139 1140 /* conserve 16m memory - start search above 16m when possible */ 1141 if (hi > PFN_16M && startpfn < PFN_16M) 1142 startpfn = PFN_16M; 1143 1144 pfn = startpfn; 1145 if (pfnalign) 1146 pfn = P2ROUNDUP(pfn, pfnalign); 1147 1148 while (pfn + minctg - 1 <= hi) { 1149 1150 plist = is_contigpage_free(&pfn, pgcnt, minctg, pfnseg, iolock); 1151 if (plist) { 1152 page_list_concat(&pplist, &plist); 1153 sgllen--; 1154 /* 1155 * return when contig pages no longer needed 1156 */ 1157 if (!*pgcnt || ((*pgcnt <= sgllen) && !pfnalign)) { 1158 startpfn = pfn; 1159 CONTIG_UNLOCK(); 1160 check_dma(mattr, pplist, *pgcnt); 1161 return (pplist); 1162 } 1163 minctg = howmany(*pgcnt, sgllen); 1164 } 1165 if (pfnalign) 1166 pfn = P2ROUNDUP(pfn, pfnalign); 1167 } 1168 1169 /* cannot find contig pages in specified range */ 1170 if (startpfn == lo) { 1171 CONTIG_UNLOCK(); 1172 return (NULL); 1173 } 1174 1175 /* did not start with lo previously */ 1176 pfn = lo; 1177 if (pfnalign) 1178 pfn = P2ROUNDUP(pfn, pfnalign); 1179 1180 /* allow search to go above startpfn */ 1181 while (pfn < startpfn) { 1182 1183 plist = is_contigpage_free(&pfn, pgcnt, minctg, pfnseg, iolock); 1184 if (plist != NULL) { 1185 1186 page_list_concat(&pplist, &plist); 1187 sgllen--; 1188 1189 /* 1190 * return when contig pages no longer needed 1191 */ 1192 if (!*pgcnt || ((*pgcnt <= sgllen) && !pfnalign)) { 1193 startpfn = pfn; 1194 CONTIG_UNLOCK(); 1195 check_dma(mattr, pplist, *pgcnt); 1196 return (pplist); 1197 } 1198 minctg = howmany(*pgcnt, sgllen); 1199 } 1200 if (pfnalign) 1201 pfn = P2ROUNDUP(pfn, pfnalign); 1202 } 1203 CONTIG_UNLOCK(); 1204 return (NULL); 1205 } 1206 #endif /* !__xpv */ 1207 1208 /* 1209 * mnode_range_cnt() calculates the number of memory ranges for mnode and 1210 * memranges[]. Used to determine the size of page lists and mnoderanges. 1211 */ 1212 int 1213 mnode_range_cnt(int mnode) 1214 { 1215 #if defined(__xpv) 1216 ASSERT(mnode == 0); 1217 return (1); 1218 #else /* __xpv */ 1219 int mri; 1220 int mnrcnt = 0; 1221 1222 if (mem_node_config[mnode].exists != 0) { 1223 mri = nranges - 1; 1224 1225 /* find the memranges index below contained in mnode range */ 1226 1227 while (MEMRANGEHI(mri) < mem_node_config[mnode].physbase) 1228 mri--; 1229 1230 /* 1231 * increment mnode range counter when memranges or mnode 1232 * boundary is reached. 1233 */ 1234 while (mri >= 0 && 1235 mem_node_config[mnode].physmax >= MEMRANGELO(mri)) { 1236 mnrcnt++; 1237 if (mem_node_config[mnode].physmax > MEMRANGEHI(mri)) 1238 mri--; 1239 else 1240 break; 1241 } 1242 } 1243 ASSERT(mnrcnt <= MAX_MNODE_MRANGES); 1244 return (mnrcnt); 1245 #endif /* __xpv */ 1246 } 1247 1248 /* 1249 * mnode_range_setup() initializes mnoderanges. 1250 */ 1251 void 1252 mnode_range_setup(mnoderange_t *mnoderanges) 1253 { 1254 int mnode, mri; 1255 1256 for (mnode = 0; mnode < max_mem_nodes; mnode++) { 1257 if (mem_node_config[mnode].exists == 0) 1258 continue; 1259 1260 mri = nranges - 1; 1261 1262 while (MEMRANGEHI(mri) < mem_node_config[mnode].physbase) 1263 mri--; 1264 1265 while (mri >= 0 && mem_node_config[mnode].physmax >= 1266 MEMRANGELO(mri)) { 1267 mnoderanges->mnr_pfnlo = MAX(MEMRANGELO(mri), 1268 mem_node_config[mnode].physbase); 1269 mnoderanges->mnr_pfnhi = MIN(MEMRANGEHI(mri), 1270 mem_node_config[mnode].physmax); 1271 mnoderanges->mnr_mnode = mnode; 1272 mnoderanges->mnr_memrange = mri; 1273 mnoderanges++; 1274 if (mem_node_config[mnode].physmax > MEMRANGEHI(mri)) 1275 mri--; 1276 else 1277 break; 1278 } 1279 } 1280 } 1281 1282 /*ARGSUSED*/ 1283 int 1284 mtype_init(vnode_t *vp, caddr_t vaddr, uint_t *flags, size_t pgsz) 1285 { 1286 int mtype = mnoderangecnt - 1; 1287 1288 #if !defined(__xpv) 1289 #if defined(__i386) 1290 /* 1291 * set the mtype range 1292 * - kmem requests needs to be below 4g if restricted_kmemalloc is set. 1293 * - for non kmem requests, set range to above 4g if memory below 4g 1294 * runs low. 1295 */ 1296 if (restricted_kmemalloc && VN_ISKAS(vp) && 1297 (caddr_t)(vaddr) >= kernelheap && 1298 (caddr_t)(vaddr) < ekernelheap) { 1299 ASSERT(physmax4g); 1300 mtype = mtype4g; 1301 if (RESTRICT16M_ALLOC(freemem4g - btop(pgsz), 1302 btop(pgsz), *flags)) { 1303 *flags |= PGI_MT_RANGE16M; 1304 } else { 1305 VM_STAT_ADD(vmm_vmstats.unrestrict16mcnt); 1306 VM_STAT_COND_ADD((*flags & PG_PANIC), 1307 vmm_vmstats.pgpanicalloc); 1308 *flags |= PGI_MT_RANGE0; 1309 } 1310 return (mtype); 1311 } 1312 #endif /* __i386 */ 1313 1314 if (RESTRICT4G_ALLOC) { 1315 VM_STAT_ADD(vmm_vmstats.restrict4gcnt); 1316 /* here only for > 4g systems */ 1317 *flags |= PGI_MT_RANGE4G; 1318 } else if (RESTRICT16M_ALLOC(freemem, btop(pgsz), *flags)) { 1319 *flags |= PGI_MT_RANGE16M; 1320 } else { 1321 VM_STAT_ADD(vmm_vmstats.unrestrict16mcnt); 1322 VM_STAT_COND_ADD((*flags & PG_PANIC), vmm_vmstats.pgpanicalloc); 1323 *flags |= PGI_MT_RANGE0; 1324 } 1325 #endif /* !__xpv */ 1326 return (mtype); 1327 } 1328 1329 1330 /* mtype init for page_get_replacement_page */ 1331 /*ARGSUSED*/ 1332 int 1333 mtype_pgr_init(int *flags, page_t *pp, int mnode, pgcnt_t pgcnt) 1334 { 1335 int mtype = mnoderangecnt - 1; 1336 #if !defined(__ixpv) 1337 if (RESTRICT16M_ALLOC(freemem, pgcnt, *flags)) { 1338 *flags |= PGI_MT_RANGE16M; 1339 } else { 1340 VM_STAT_ADD(vmm_vmstats.unrestrict16mcnt); 1341 *flags |= PGI_MT_RANGE0; 1342 } 1343 #endif 1344 return (mtype); 1345 } 1346 1347 /* 1348 * Determine if the mnode range specified in mtype contains memory belonging 1349 * to memory node mnode. If flags & PGI_MT_RANGE is set then mtype contains 1350 * the range of indices from high pfn to 0, 16m or 4g. 1351 * 1352 * Return first mnode range type index found otherwise return -1 if none found. 1353 */ 1354 int 1355 mtype_func(int mnode, int mtype, uint_t flags) 1356 { 1357 if (flags & PGI_MT_RANGE) { 1358 int mtlim = 0; 1359 1360 if (flags & PGI_MT_NEXT) 1361 mtype--; 1362 if (flags & PGI_MT_RANGE4G) 1363 mtlim = mtype4g + 1; /* exclude 0-4g range */ 1364 else if (flags & PGI_MT_RANGE16M) 1365 mtlim = 1; /* exclude 0-16m range */ 1366 while (mtype >= mtlim) { 1367 if (mnoderanges[mtype].mnr_mnode == mnode) 1368 return (mtype); 1369 mtype--; 1370 } 1371 } else if (mnoderanges[mtype].mnr_mnode == mnode) { 1372 return (mtype); 1373 } 1374 return (-1); 1375 } 1376 1377 /* 1378 * Update the page list max counts with the pfn range specified by the 1379 * input parameters. Called from add_physmem() when physical memory with 1380 * page_t's are initially added to the page lists. 1381 */ 1382 void 1383 mtype_modify_max(pfn_t startpfn, long cnt) 1384 { 1385 int mtype = 0; 1386 pfn_t endpfn = startpfn + cnt, pfn; 1387 pgcnt_t inc; 1388 1389 ASSERT(cnt > 0); 1390 1391 if (!physmax4g) 1392 return; 1393 1394 for (pfn = startpfn; pfn < endpfn; ) { 1395 if (pfn <= mnoderanges[mtype].mnr_pfnhi) { 1396 if (endpfn < mnoderanges[mtype].mnr_pfnhi) { 1397 inc = endpfn - pfn; 1398 } else { 1399 inc = mnoderanges[mtype].mnr_pfnhi - pfn + 1; 1400 } 1401 if (mtype <= mtype4g) 1402 maxmem4g += inc; 1403 pfn += inc; 1404 } 1405 mtype++; 1406 ASSERT(mtype < mnoderangecnt || pfn >= endpfn); 1407 } 1408 } 1409 1410 int 1411 mtype_2_mrange(int mtype) 1412 { 1413 return (mnoderanges[mtype].mnr_memrange); 1414 } 1415 1416 void 1417 mnodetype_2_pfn(int mnode, int mtype, pfn_t *pfnlo, pfn_t *pfnhi) 1418 { 1419 ASSERT(mnoderanges[mtype].mnr_mnode == mnode); 1420 *pfnlo = mnoderanges[mtype].mnr_pfnlo; 1421 *pfnhi = mnoderanges[mtype].mnr_pfnhi; 1422 } 1423 1424 size_t 1425 plcnt_sz(size_t ctrs_sz) 1426 { 1427 #ifdef DEBUG 1428 int szc, colors; 1429 1430 ctrs_sz += mnoderangecnt * sizeof (struct mnr_mts) * mmu_page_sizes; 1431 for (szc = 0; szc < mmu_page_sizes; szc++) { 1432 colors = page_get_pagecolors(szc); 1433 ctrs_sz += mnoderangecnt * sizeof (pgcnt_t) * colors; 1434 } 1435 #endif 1436 return (ctrs_sz); 1437 } 1438 1439 caddr_t 1440 plcnt_init(caddr_t addr) 1441 { 1442 #ifdef DEBUG 1443 int mt, szc, colors; 1444 1445 for (mt = 0; mt < mnoderangecnt; mt++) { 1446 mnoderanges[mt].mnr_mts = (struct mnr_mts *)addr; 1447 addr += (sizeof (struct mnr_mts) * mmu_page_sizes); 1448 for (szc = 0; szc < mmu_page_sizes; szc++) { 1449 colors = page_get_pagecolors(szc); 1450 mnoderanges[mt].mnr_mts[szc].mnr_mts_colors = colors; 1451 mnoderanges[mt].mnr_mts[szc].mnr_mtsc_pgcnt = 1452 (pgcnt_t *)addr; 1453 addr += (sizeof (pgcnt_t) * colors); 1454 } 1455 } 1456 #endif 1457 return (addr); 1458 } 1459 1460 void 1461 plcnt_inc_dec(page_t *pp, int mtype, int szc, long cnt, int flags) 1462 { 1463 #ifdef DEBUG 1464 int bin = PP_2_BIN(pp); 1465 1466 atomic_add_long(&mnoderanges[mtype].mnr_mts[szc].mnr_mts_pgcnt, cnt); 1467 atomic_add_long(&mnoderanges[mtype].mnr_mts[szc].mnr_mtsc_pgcnt[bin], 1468 cnt); 1469 #endif 1470 ASSERT(mtype == PP_2_MTYPE(pp)); 1471 if (physmax4g && mtype <= mtype4g) 1472 atomic_add_long(&freemem4g, cnt); 1473 if (flags & PG_CACHE_LIST) 1474 atomic_add_long(&mnoderanges[mtype].mnr_mt_clpgcnt, cnt); 1475 else 1476 atomic_add_long(&mnoderanges[mtype].mnr_mt_flpgcnt[szc], cnt); 1477 atomic_add_long(&mnoderanges[mtype].mnr_mt_totcnt, cnt); 1478 } 1479 1480 /* 1481 * Returns the free page count for mnode 1482 */ 1483 int 1484 mnode_pgcnt(int mnode) 1485 { 1486 int mtype = mnoderangecnt - 1; 1487 int flags = PGI_MT_RANGE0; 1488 pgcnt_t pgcnt = 0; 1489 1490 mtype = mtype_func(mnode, mtype, flags); 1491 1492 while (mtype != -1) { 1493 pgcnt += MTYPE_FREEMEM(mtype); 1494 mtype = mtype_func(mnode, mtype, flags | PGI_MT_NEXT); 1495 } 1496 return (pgcnt); 1497 } 1498 1499 /* 1500 * Initialize page coloring variables based on the l2 cache parameters. 1501 * Calculate and return memory needed for page coloring data structures. 1502 */ 1503 size_t 1504 page_coloring_init(uint_t l2_sz, int l2_linesz, int l2_assoc) 1505 { 1506 size_t colorsz = 0; 1507 int i; 1508 int colors; 1509 1510 #if defined(__xpv) 1511 /* 1512 * Hypervisor domains currently don't have any concept of NUMA. 1513 * Hence we'll act like there is only 1 memrange. 1514 */ 1515 i = memrange_num(1); 1516 #else /* !__xpv */ 1517 /* 1518 * Reduce the memory ranges lists if we don't have large amounts 1519 * of memory. This avoids searching known empty free lists. 1520 */ 1521 i = memrange_num(physmax); 1522 #if defined(__i386) 1523 if (i > 0) 1524 restricted_kmemalloc = 0; 1525 #endif 1526 /* physmax greater than 4g */ 1527 if (i == 0) 1528 physmax4g = 1; 1529 #endif /* !__xpv */ 1530 memranges += i; 1531 nranges -= i; 1532 1533 ASSERT(mmu_page_sizes <= MMU_PAGE_SIZES); 1534 1535 ASSERT(ISP2(l2_sz)); 1536 ASSERT(ISP2(l2_linesz)); 1537 ASSERT(l2_sz > MMU_PAGESIZE); 1538 1539 /* l2_assoc is 0 for fully associative l2 cache */ 1540 if (l2_assoc) 1541 l2_colors = MAX(1, l2_sz / (l2_assoc * MMU_PAGESIZE)); 1542 else 1543 l2_colors = 1; 1544 1545 /* for scalability, configure at least PAGE_COLORS_MIN color bins */ 1546 page_colors = MAX(l2_colors, PAGE_COLORS_MIN); 1547 1548 /* 1549 * cpu_page_colors is non-zero when a page color may be spread across 1550 * multiple bins. 1551 */ 1552 if (l2_colors < page_colors) 1553 cpu_page_colors = l2_colors; 1554 1555 ASSERT(ISP2(page_colors)); 1556 1557 page_colors_mask = page_colors - 1; 1558 1559 ASSERT(ISP2(CPUSETSIZE())); 1560 page_coloring_shift = lowbit(CPUSETSIZE()); 1561 1562 /* initialize number of colors per page size */ 1563 for (i = 0; i <= mmu.max_page_level; i++) { 1564 hw_page_array[i].hp_size = LEVEL_SIZE(i); 1565 hw_page_array[i].hp_shift = LEVEL_SHIFT(i); 1566 hw_page_array[i].hp_pgcnt = LEVEL_SIZE(i) >> LEVEL_SHIFT(0); 1567 hw_page_array[i].hp_colors = (page_colors_mask >> 1568 (hw_page_array[i].hp_shift - hw_page_array[0].hp_shift)) 1569 + 1; 1570 colorequivszc[i] = 0; 1571 } 1572 1573 /* 1574 * The value of cpu_page_colors determines if additional color bins 1575 * need to be checked for a particular color in the page_get routines. 1576 */ 1577 if (cpu_page_colors != 0) { 1578 1579 int a = lowbit(page_colors) - lowbit(cpu_page_colors); 1580 ASSERT(a > 0); 1581 ASSERT(a < 16); 1582 1583 for (i = 0; i <= mmu.max_page_level; i++) { 1584 if ((colors = hw_page_array[i].hp_colors) <= 1) { 1585 colorequivszc[i] = 0; 1586 continue; 1587 } 1588 while ((colors >> a) == 0) 1589 a--; 1590 ASSERT(a >= 0); 1591 1592 /* higher 4 bits encodes color equiv mask */ 1593 colorequivszc[i] = (a << 4); 1594 } 1595 } 1596 1597 /* factor in colorequiv to check additional 'equivalent' bins. */ 1598 if (colorequiv > 1) { 1599 1600 int a = lowbit(colorequiv) - 1; 1601 if (a > 15) 1602 a = 15; 1603 1604 for (i = 0; i <= mmu.max_page_level; i++) { 1605 if ((colors = hw_page_array[i].hp_colors) <= 1) { 1606 continue; 1607 } 1608 while ((colors >> a) == 0) 1609 a--; 1610 if ((a << 4) > colorequivszc[i]) { 1611 colorequivszc[i] = (a << 4); 1612 } 1613 } 1614 } 1615 1616 /* size for mnoderanges */ 1617 for (mnoderangecnt = 0, i = 0; i < max_mem_nodes; i++) 1618 mnoderangecnt += mnode_range_cnt(i); 1619 colorsz = mnoderangecnt * sizeof (mnoderange_t); 1620 1621 /* size for fpc_mutex and cpc_mutex */ 1622 colorsz += (2 * max_mem_nodes * sizeof (kmutex_t) * NPC_MUTEX); 1623 1624 /* size of page_freelists */ 1625 colorsz += mnoderangecnt * sizeof (page_t ***); 1626 colorsz += mnoderangecnt * mmu_page_sizes * sizeof (page_t **); 1627 1628 for (i = 0; i < mmu_page_sizes; i++) { 1629 colors = page_get_pagecolors(i); 1630 colorsz += mnoderangecnt * colors * sizeof (page_t *); 1631 } 1632 1633 /* size of page_cachelists */ 1634 colorsz += mnoderangecnt * sizeof (page_t **); 1635 colorsz += mnoderangecnt * page_colors * sizeof (page_t *); 1636 1637 return (colorsz); 1638 } 1639 1640 /* 1641 * Called once at startup to configure page_coloring data structures and 1642 * does the 1st page_free()/page_freelist_add(). 1643 */ 1644 void 1645 page_coloring_setup(caddr_t pcmemaddr) 1646 { 1647 int i; 1648 int j; 1649 int k; 1650 caddr_t addr; 1651 int colors; 1652 1653 /* 1654 * do page coloring setup 1655 */ 1656 addr = pcmemaddr; 1657 1658 mnoderanges = (mnoderange_t *)addr; 1659 addr += (mnoderangecnt * sizeof (mnoderange_t)); 1660 1661 mnode_range_setup(mnoderanges); 1662 1663 if (physmax4g) 1664 mtype4g = pfn_2_mtype(0xfffff); 1665 1666 for (k = 0; k < NPC_MUTEX; k++) { 1667 fpc_mutex[k] = (kmutex_t *)addr; 1668 addr += (max_mem_nodes * sizeof (kmutex_t)); 1669 } 1670 for (k = 0; k < NPC_MUTEX; k++) { 1671 cpc_mutex[k] = (kmutex_t *)addr; 1672 addr += (max_mem_nodes * sizeof (kmutex_t)); 1673 } 1674 page_freelists = (page_t ****)addr; 1675 addr += (mnoderangecnt * sizeof (page_t ***)); 1676 1677 page_cachelists = (page_t ***)addr; 1678 addr += (mnoderangecnt * sizeof (page_t **)); 1679 1680 for (i = 0; i < mnoderangecnt; i++) { 1681 page_freelists[i] = (page_t ***)addr; 1682 addr += (mmu_page_sizes * sizeof (page_t **)); 1683 1684 for (j = 0; j < mmu_page_sizes; j++) { 1685 colors = page_get_pagecolors(j); 1686 page_freelists[i][j] = (page_t **)addr; 1687 addr += (colors * sizeof (page_t *)); 1688 } 1689 page_cachelists[i] = (page_t **)addr; 1690 addr += (page_colors * sizeof (page_t *)); 1691 } 1692 } 1693 1694 #if defined(__xpv) 1695 /* 1696 * Give back 10% of the io_pool pages to the free list. 1697 * Don't shrink the pool below some absolute minimum. 1698 */ 1699 static void 1700 page_io_pool_shrink() 1701 { 1702 int retcnt; 1703 page_t *pp, *pp_first, *pp_last, **curpool; 1704 mfn_t mfn; 1705 int bothpools = 0; 1706 1707 mutex_enter(&io_pool_lock); 1708 io_pool_shrink_attempts++; /* should be a kstat? */ 1709 retcnt = io_pool_cnt / 10; 1710 if (io_pool_cnt - retcnt < io_pool_cnt_min) 1711 retcnt = io_pool_cnt - io_pool_cnt_min; 1712 if (retcnt <= 0) 1713 goto done; 1714 io_pool_shrinks++; /* should be a kstat? */ 1715 curpool = &io_pool_4g; 1716 domore: 1717 /* 1718 * Loop through taking pages from the end of the list 1719 * (highest mfns) till amount to return reached. 1720 */ 1721 for (pp = *curpool; pp && retcnt > 0; ) { 1722 pp_first = pp_last = pp->p_prev; 1723 if (pp_first == *curpool) 1724 break; 1725 retcnt--; 1726 io_pool_cnt--; 1727 page_io_pool_sub(curpool, pp_first, pp_last); 1728 if ((mfn = pfn_to_mfn(pp->p_pagenum)) < start_mfn) 1729 start_mfn = mfn; 1730 page_free(pp_first, 1); 1731 pp = *curpool; 1732 } 1733 if (retcnt != 0 && !bothpools) { 1734 /* 1735 * If not enough found in less constrained pool try the 1736 * more constrained one. 1737 */ 1738 curpool = &io_pool_16m; 1739 bothpools = 1; 1740 goto domore; 1741 } 1742 done: 1743 mutex_exit(&io_pool_lock); 1744 } 1745 1746 #endif /* __xpv */ 1747 1748 uint_t 1749 page_create_update_flags_x86(uint_t flags) 1750 { 1751 #if defined(__xpv) 1752 /* 1753 * Check this is an urgent allocation and free pages are depleted. 1754 */ 1755 if (!(flags & PG_WAIT) && freemem < desfree) 1756 page_io_pool_shrink(); 1757 #else /* !__xpv */ 1758 /* 1759 * page_create_get_something may call this because 4g memory may be 1760 * depleted. Set flags to allow for relocation of base page below 1761 * 4g if necessary. 1762 */ 1763 if (physmax4g) 1764 flags |= (PGI_PGCPSZC0 | PGI_PGCPHIPRI); 1765 #endif /* __xpv */ 1766 return (flags); 1767 } 1768 1769 /*ARGSUSED*/ 1770 int 1771 bp_color(struct buf *bp) 1772 { 1773 return (0); 1774 } 1775 1776 #if defined(__xpv) 1777 1778 /* 1779 * Take pages out of an io_pool 1780 */ 1781 static void 1782 page_io_pool_sub(page_t **poolp, page_t *pp_first, page_t *pp_last) 1783 { 1784 if (*poolp == pp_first) { 1785 *poolp = pp_last->p_next; 1786 if (*poolp == pp_first) 1787 *poolp = NULL; 1788 } 1789 pp_first->p_prev->p_next = pp_last->p_next; 1790 pp_last->p_next->p_prev = pp_first->p_prev; 1791 pp_first->p_prev = pp_last; 1792 pp_last->p_next = pp_first; 1793 } 1794 1795 /* 1796 * Put a page on the io_pool list. The list is ordered by increasing MFN. 1797 */ 1798 static void 1799 page_io_pool_add(page_t **poolp, page_t *pp) 1800 { 1801 page_t *look; 1802 mfn_t mfn = mfn_list[pp->p_pagenum]; 1803 1804 if (*poolp == NULL) { 1805 *poolp = pp; 1806 pp->p_next = pp; 1807 pp->p_prev = pp; 1808 return; 1809 } 1810 1811 /* 1812 * Since we try to take pages from the high end of the pool 1813 * chances are good that the pages to be put on the list will 1814 * go at or near the end of the list. so start at the end and 1815 * work backwards. 1816 */ 1817 look = (*poolp)->p_prev; 1818 while (mfn < mfn_list[look->p_pagenum]) { 1819 look = look->p_prev; 1820 if (look == (*poolp)->p_prev) 1821 break; /* backed all the way to front of list */ 1822 } 1823 1824 /* insert after look */ 1825 pp->p_prev = look; 1826 pp->p_next = look->p_next; 1827 pp->p_next->p_prev = pp; 1828 look->p_next = pp; 1829 if (mfn < mfn_list[(*poolp)->p_pagenum]) { 1830 /* 1831 * we inserted a new first list element 1832 * adjust pool pointer to newly inserted element 1833 */ 1834 *poolp = pp; 1835 } 1836 } 1837 1838 /* 1839 * Add a page to the io_pool. Setting the force flag will force the page 1840 * into the io_pool no matter what. 1841 */ 1842 static void 1843 add_page_to_pool(page_t *pp, int force) 1844 { 1845 page_t *highest; 1846 page_t *freep = NULL; 1847 1848 mutex_enter(&io_pool_lock); 1849 /* 1850 * Always keep the scarce low memory pages 1851 */ 1852 if (mfn_list[pp->p_pagenum] < PFN_16MEG) { 1853 ++io_pool_cnt; 1854 page_io_pool_add(&io_pool_16m, pp); 1855 goto done; 1856 } 1857 if (io_pool_cnt < io_pool_cnt_max || force || io_pool_4g == NULL) { 1858 ++io_pool_cnt; 1859 page_io_pool_add(&io_pool_4g, pp); 1860 } else { 1861 highest = io_pool_4g->p_prev; 1862 if (mfn_list[pp->p_pagenum] < mfn_list[highest->p_pagenum]) { 1863 page_io_pool_sub(&io_pool_4g, highest, highest); 1864 page_io_pool_add(&io_pool_4g, pp); 1865 freep = highest; 1866 } else { 1867 freep = pp; 1868 } 1869 } 1870 done: 1871 mutex_exit(&io_pool_lock); 1872 if (freep) 1873 page_free(freep, 1); 1874 } 1875 1876 1877 int contig_pfn_cnt; /* no of pfns in the contig pfn list */ 1878 int contig_pfn_max; /* capacity of the contig pfn list */ 1879 int next_alloc_pfn; /* next position in list to start a contig search */ 1880 int contig_pfnlist_updates; /* pfn list update count */ 1881 int contig_pfnlist_builds; /* how many times have we (re)built list */ 1882 int contig_pfnlist_buildfailed; /* how many times has list build failed */ 1883 int create_contig_pending; /* nonzero means taskq creating contig list */ 1884 pfn_t *contig_pfn_list = NULL; /* list of contig pfns in ascending mfn order */ 1885 1886 /* 1887 * Function to use in sorting a list of pfns by their underlying mfns. 1888 */ 1889 static int 1890 mfn_compare(const void *pfnp1, const void *pfnp2) 1891 { 1892 mfn_t mfn1 = mfn_list[*(pfn_t *)pfnp1]; 1893 mfn_t mfn2 = mfn_list[*(pfn_t *)pfnp2]; 1894 1895 if (mfn1 > mfn2) 1896 return (1); 1897 if (mfn1 < mfn2) 1898 return (-1); 1899 return (0); 1900 } 1901 1902 /* 1903 * Compact the contig_pfn_list by tossing all the non-contiguous 1904 * elements from the list. 1905 */ 1906 static void 1907 compact_contig_pfn_list(void) 1908 { 1909 pfn_t pfn, lapfn, prev_lapfn; 1910 mfn_t mfn; 1911 int i, newcnt = 0; 1912 1913 prev_lapfn = 0; 1914 for (i = 0; i < contig_pfn_cnt - 1; i++) { 1915 pfn = contig_pfn_list[i]; 1916 lapfn = contig_pfn_list[i + 1]; 1917 mfn = mfn_list[pfn]; 1918 /* 1919 * See if next pfn is for a contig mfn 1920 */ 1921 if (mfn_list[lapfn] != mfn + 1) 1922 continue; 1923 /* 1924 * pfn and lookahead are both put in list 1925 * unless pfn is the previous lookahead. 1926 */ 1927 if (pfn != prev_lapfn) 1928 contig_pfn_list[newcnt++] = pfn; 1929 contig_pfn_list[newcnt++] = lapfn; 1930 prev_lapfn = lapfn; 1931 } 1932 for (i = newcnt; i < contig_pfn_cnt; i++) 1933 contig_pfn_list[i] = 0; 1934 contig_pfn_cnt = newcnt; 1935 } 1936 1937 /*ARGSUSED*/ 1938 static void 1939 call_create_contiglist(void *arg) 1940 { 1941 (void) create_contig_pfnlist(PG_WAIT); 1942 } 1943 1944 /* 1945 * Create list of freelist pfns that have underlying 1946 * contiguous mfns. The list is kept in ascending mfn order. 1947 * returns 1 if list created else 0. 1948 */ 1949 static int 1950 create_contig_pfnlist(uint_t flags) 1951 { 1952 pfn_t pfn; 1953 page_t *pp; 1954 int ret = 1; 1955 1956 mutex_enter(&contig_list_lock); 1957 if (contig_pfn_list != NULL) 1958 goto out; 1959 contig_pfn_max = freemem + (freemem / 10); 1960 contig_pfn_list = kmem_zalloc(contig_pfn_max * sizeof (pfn_t), 1961 (flags & PG_WAIT) ? KM_SLEEP : KM_NOSLEEP); 1962 if (contig_pfn_list == NULL) { 1963 /* 1964 * If we could not create the contig list (because 1965 * we could not sleep for memory). Dispatch a taskq that can 1966 * sleep to get the memory. 1967 */ 1968 if (!create_contig_pending) { 1969 if (taskq_dispatch(system_taskq, call_create_contiglist, 1970 NULL, TQ_NOSLEEP) != NULL) 1971 create_contig_pending = 1; 1972 } 1973 contig_pfnlist_buildfailed++; /* count list build failures */ 1974 ret = 0; 1975 goto out; 1976 } 1977 create_contig_pending = 0; 1978 ASSERT(contig_pfn_cnt == 0); 1979 for (pfn = 0; pfn < mfn_count; pfn++) { 1980 pp = page_numtopp_nolock(pfn); 1981 if (pp == NULL || !PP_ISFREE(pp)) 1982 continue; 1983 contig_pfn_list[contig_pfn_cnt] = pfn; 1984 if (++contig_pfn_cnt == contig_pfn_max) 1985 break; 1986 } 1987 qsort(contig_pfn_list, contig_pfn_cnt, sizeof (pfn_t), mfn_compare); 1988 compact_contig_pfn_list(); 1989 /* 1990 * Make sure next search of the newly created contiguous pfn 1991 * list starts at the beginning of the list. 1992 */ 1993 next_alloc_pfn = 0; 1994 contig_pfnlist_builds++; /* count list builds */ 1995 out: 1996 mutex_exit(&contig_list_lock); 1997 return (ret); 1998 } 1999 2000 2001 /* 2002 * Toss the current contig pfnlist. Someone is about to do a massive 2003 * update to pfn<->mfn mappings. So we have them destroy the list and lock 2004 * it till they are done with their update. 2005 */ 2006 void 2007 clear_and_lock_contig_pfnlist() 2008 { 2009 pfn_t *listp = NULL; 2010 size_t listsize; 2011 2012 mutex_enter(&contig_list_lock); 2013 if (contig_pfn_list != NULL) { 2014 listp = contig_pfn_list; 2015 listsize = contig_pfn_max * sizeof (pfn_t); 2016 contig_pfn_list = NULL; 2017 contig_pfn_max = contig_pfn_cnt = 0; 2018 } 2019 if (listp != NULL) 2020 kmem_free(listp, listsize); 2021 } 2022 2023 /* 2024 * Unlock the contig_pfn_list. The next attempted use of it will cause 2025 * it to be re-created. 2026 */ 2027 void 2028 unlock_contig_pfnlist() 2029 { 2030 mutex_exit(&contig_list_lock); 2031 } 2032 2033 /* 2034 * Update the contiguous pfn list in response to a pfn <-> mfn reassignment 2035 */ 2036 void 2037 update_contig_pfnlist(pfn_t pfn, mfn_t oldmfn, mfn_t newmfn) 2038 { 2039 int probe_hi, probe_lo, probe_pos, insert_after, insert_point; 2040 pfn_t probe_pfn; 2041 mfn_t probe_mfn; 2042 int drop_lock = 0; 2043 2044 if (mutex_owner(&contig_list_lock) != curthread) { 2045 drop_lock = 1; 2046 mutex_enter(&contig_list_lock); 2047 } 2048 if (contig_pfn_list == NULL) 2049 goto done; 2050 contig_pfnlist_updates++; 2051 /* 2052 * Find the pfn in the current list. Use a binary chop to locate it. 2053 */ 2054 probe_hi = contig_pfn_cnt - 1; 2055 probe_lo = 0; 2056 probe_pos = (probe_hi + probe_lo) / 2; 2057 while ((probe_pfn = contig_pfn_list[probe_pos]) != pfn) { 2058 if (probe_pos == probe_lo) { /* pfn not in list */ 2059 probe_pos = -1; 2060 break; 2061 } 2062 if (pfn_to_mfn(probe_pfn) <= oldmfn) 2063 probe_lo = probe_pos; 2064 else 2065 probe_hi = probe_pos; 2066 probe_pos = (probe_hi + probe_lo) / 2; 2067 } 2068 if (probe_pos >= 0) { /* remove pfn fom list */ 2069 contig_pfn_cnt--; 2070 ovbcopy(&contig_pfn_list[probe_pos + 1], 2071 &contig_pfn_list[probe_pos], 2072 (contig_pfn_cnt - probe_pos) * sizeof (pfn_t)); 2073 } 2074 if (newmfn == MFN_INVALID) 2075 goto done; 2076 /* 2077 * Check if new mfn has adjacent mfns in the list 2078 */ 2079 probe_hi = contig_pfn_cnt - 1; 2080 probe_lo = 0; 2081 insert_after = -2; 2082 do { 2083 probe_pos = (probe_hi + probe_lo) / 2; 2084 probe_mfn = pfn_to_mfn(contig_pfn_list[probe_pos]); 2085 if (newmfn == probe_mfn + 1) 2086 insert_after = probe_pos; 2087 else if (newmfn == probe_mfn - 1) 2088 insert_after = probe_pos - 1; 2089 if (probe_pos == probe_lo) 2090 break; 2091 if (probe_mfn <= newmfn) 2092 probe_lo = probe_pos; 2093 else 2094 probe_hi = probe_pos; 2095 } while (insert_after == -2); 2096 /* 2097 * If there is space in the list and there are adjacent mfns 2098 * insert the pfn in to its proper place in the list. 2099 */ 2100 if (insert_after != -2 && contig_pfn_cnt + 1 <= contig_pfn_max) { 2101 insert_point = insert_after + 1; 2102 ovbcopy(&contig_pfn_list[insert_point], 2103 &contig_pfn_list[insert_point + 1], 2104 (contig_pfn_cnt - insert_point) * sizeof (pfn_t)); 2105 contig_pfn_list[insert_point] = pfn; 2106 contig_pfn_cnt++; 2107 } 2108 done: 2109 if (drop_lock) 2110 mutex_exit(&contig_list_lock); 2111 } 2112 2113 /* 2114 * Called to (re-)populate the io_pool from the free page lists. 2115 */ 2116 long 2117 populate_io_pool(void) 2118 { 2119 pfn_t pfn; 2120 mfn_t mfn, max_mfn; 2121 page_t *pp; 2122 2123 /* 2124 * Figure out the bounds of the pool on first invocation. 2125 * We use a percentage of memory for the io pool size. 2126 * we allow that to shrink, but not to less than a fixed minimum 2127 */ 2128 if (io_pool_cnt_max == 0) { 2129 io_pool_cnt_max = physmem / (100 / io_pool_physmem_pct); 2130 io_pool_cnt_lowater = io_pool_cnt_max; 2131 /* 2132 * This is the first time in populate_io_pool, grab a va to use 2133 * when we need to allocate pages. 2134 */ 2135 io_pool_kva = vmem_alloc(heap_arena, PAGESIZE, VM_SLEEP); 2136 } 2137 /* 2138 * If we are out of pages in the pool, then grow the size of the pool 2139 */ 2140 if (io_pool_cnt == 0) { 2141 /* 2142 * Grow the max size of the io pool by 5%, but never more than 2143 * 25% of physical memory. 2144 */ 2145 if (io_pool_cnt_max < physmem / 4) 2146 io_pool_cnt_max += io_pool_cnt_max / 20; 2147 } 2148 io_pool_grows++; /* should be a kstat? */ 2149 2150 /* 2151 * Get highest mfn on this platform, but limit to the 32 bit DMA max. 2152 */ 2153 (void) mfn_to_pfn(start_mfn); 2154 max_mfn = MIN(cached_max_mfn, PFN_4GIG); 2155 for (mfn = start_mfn; mfn < max_mfn; start_mfn = ++mfn) { 2156 pfn = mfn_to_pfn(mfn); 2157 if (pfn & PFN_IS_FOREIGN_MFN) 2158 continue; 2159 /* 2160 * try to allocate it from free pages 2161 */ 2162 pp = page_numtopp_alloc(pfn); 2163 if (pp == NULL) 2164 continue; 2165 PP_CLRFREE(pp); 2166 add_page_to_pool(pp, 1); 2167 if (io_pool_cnt >= io_pool_cnt_max) 2168 break; 2169 } 2170 2171 return (io_pool_cnt); 2172 } 2173 2174 /* 2175 * Destroy a page that was being used for DMA I/O. It may or 2176 * may not actually go back to the io_pool. 2177 */ 2178 void 2179 page_destroy_io(page_t *pp) 2180 { 2181 mfn_t mfn = mfn_list[pp->p_pagenum]; 2182 2183 /* 2184 * When the page was alloc'd a reservation was made, release it now 2185 */ 2186 page_unresv(1); 2187 /* 2188 * Unload translations, if any, then hash out the 2189 * page to erase its identity. 2190 */ 2191 (void) hat_pageunload(pp, HAT_FORCE_PGUNLOAD); 2192 page_hashout(pp, NULL); 2193 2194 /* 2195 * If the page came from the free lists, just put it back to them. 2196 * DomU pages always go on the free lists as well. 2197 */ 2198 if (!DOMAIN_IS_INITDOMAIN(xen_info) || mfn >= PFN_4GIG) { 2199 page_free(pp, 1); 2200 return; 2201 } 2202 2203 add_page_to_pool(pp, 0); 2204 } 2205 2206 2207 long contig_searches; /* count of times contig pages requested */ 2208 long contig_search_restarts; /* count of contig ranges tried */ 2209 long contig_search_failed; /* count of contig alloc failures */ 2210 2211 /* 2212 * Look thru the contiguous pfns that are not part of the io_pool for 2213 * contiguous free pages. Return a list of the found pages or NULL. 2214 */ 2215 page_t * 2216 find_contig_free(uint_t npages, uint_t flags, uint64_t pfnseg) 2217 { 2218 page_t *pp, *plist = NULL; 2219 mfn_t mfn, prev_mfn, start_mfn; 2220 pfn_t pfn; 2221 int pages_needed, pages_requested; 2222 int search_start; 2223 2224 /* 2225 * create the contig pfn list if not already done 2226 */ 2227 retry: 2228 mutex_enter(&contig_list_lock); 2229 if (contig_pfn_list == NULL) { 2230 mutex_exit(&contig_list_lock); 2231 if (!create_contig_pfnlist(flags)) { 2232 return (NULL); 2233 } 2234 goto retry; 2235 } 2236 contig_searches++; 2237 /* 2238 * Search contiguous pfn list for physically contiguous pages not in 2239 * the io_pool. Start the search where the last search left off. 2240 */ 2241 pages_requested = pages_needed = npages; 2242 search_start = next_alloc_pfn; 2243 start_mfn = prev_mfn = 0; 2244 while (pages_needed) { 2245 pfn = contig_pfn_list[next_alloc_pfn]; 2246 mfn = pfn_to_mfn(pfn); 2247 /* 2248 * Check if mfn is first one or contig to previous one and 2249 * if page corresponding to mfn is free and that mfn 2250 * range is not crossing a segment boundary. 2251 */ 2252 if ((prev_mfn == 0 || mfn == prev_mfn + 1) && 2253 (pp = page_numtopp_alloc(pfn)) != NULL && 2254 !((mfn & pfnseg) < (start_mfn & pfnseg))) { 2255 PP_CLRFREE(pp); 2256 page_io_pool_add(&plist, pp); 2257 pages_needed--; 2258 if (prev_mfn == 0) 2259 start_mfn = mfn; 2260 prev_mfn = mfn; 2261 } else { 2262 contig_search_restarts++; 2263 /* 2264 * free partial page list 2265 */ 2266 while (plist != NULL) { 2267 pp = plist; 2268 page_io_pool_sub(&plist, pp, pp); 2269 page_free(pp, 1); 2270 } 2271 pages_needed = pages_requested; 2272 start_mfn = prev_mfn = 0; 2273 } 2274 if (++next_alloc_pfn == contig_pfn_cnt) 2275 next_alloc_pfn = 0; 2276 if (next_alloc_pfn == search_start) 2277 break; /* all pfns searched */ 2278 } 2279 mutex_exit(&contig_list_lock); 2280 if (pages_needed) { 2281 contig_search_failed++; 2282 /* 2283 * Failed to find enough contig pages. 2284 * free partial page list 2285 */ 2286 while (plist != NULL) { 2287 pp = plist; 2288 page_io_pool_sub(&plist, pp, pp); 2289 page_free(pp, 1); 2290 } 2291 } 2292 return (plist); 2293 } 2294 2295 /* 2296 * Search the reserved io pool pages for a page range with the 2297 * desired characteristics. 2298 */ 2299 page_t * 2300 page_io_pool_alloc(ddi_dma_attr_t *mattr, int contig, pgcnt_t minctg) 2301 { 2302 page_t *pp_first, *pp_last; 2303 page_t *pp, **poolp; 2304 pgcnt_t nwanted, pfnalign; 2305 uint64_t pfnseg; 2306 mfn_t mfn, tmfn, hi_mfn, lo_mfn; 2307 int align, attempt = 0; 2308 2309 if (minctg == 1) 2310 contig = 0; 2311 lo_mfn = mmu_btop(mattr->dma_attr_addr_lo); 2312 hi_mfn = mmu_btop(mattr->dma_attr_addr_hi); 2313 pfnseg = mmu_btop(mattr->dma_attr_seg); 2314 align = maxbit(mattr->dma_attr_align, mattr->dma_attr_minxfer); 2315 if (align > MMU_PAGESIZE) 2316 pfnalign = mmu_btop(align); 2317 else 2318 pfnalign = 0; 2319 2320 try_again: 2321 /* 2322 * See if we want pages for a legacy device 2323 */ 2324 if (hi_mfn < PFN_16MEG) 2325 poolp = &io_pool_16m; 2326 else 2327 poolp = &io_pool_4g; 2328 try_smaller: 2329 /* 2330 * Take pages from I/O pool. We'll use pages from the highest 2331 * MFN range possible. 2332 */ 2333 pp_first = pp_last = NULL; 2334 mutex_enter(&io_pool_lock); 2335 nwanted = minctg; 2336 for (pp = *poolp; pp && nwanted > 0; ) { 2337 pp = pp->p_prev; 2338 2339 /* 2340 * skip pages above allowable range 2341 */ 2342 mfn = mfn_list[pp->p_pagenum]; 2343 if (hi_mfn < mfn) 2344 goto skip; 2345 2346 /* 2347 * stop at pages below allowable range 2348 */ 2349 if (lo_mfn > mfn) 2350 break; 2351 restart: 2352 if (pp_last == NULL) { 2353 /* 2354 * Check alignment 2355 */ 2356 tmfn = mfn - (minctg - 1); 2357 if (pfnalign && tmfn != P2ROUNDUP(tmfn, pfnalign)) 2358 goto skip; /* not properly aligned */ 2359 /* 2360 * Check segment 2361 */ 2362 if ((mfn & pfnseg) < (tmfn & pfnseg)) 2363 goto skip; /* crosses seg boundary */ 2364 /* 2365 * Start building page list 2366 */ 2367 pp_first = pp_last = pp; 2368 nwanted--; 2369 } else { 2370 /* 2371 * check physical contiguity if required 2372 */ 2373 if (contig && 2374 mfn_list[pp_first->p_pagenum] != mfn + 1) { 2375 /* 2376 * not a contiguous page, restart list. 2377 */ 2378 pp_last = NULL; 2379 nwanted = minctg; 2380 goto restart; 2381 } else { /* add page to list */ 2382 pp_first = pp; 2383 nwanted--; 2384 } 2385 } 2386 skip: 2387 if (pp == *poolp) 2388 break; 2389 } 2390 2391 /* 2392 * If we didn't find memory. Try the more constrained pool, then 2393 * sweep free pages into the DMA pool and try again. 2394 */ 2395 if (nwanted != 0) { 2396 mutex_exit(&io_pool_lock); 2397 /* 2398 * If we were looking in the less constrained pool and 2399 * didn't find pages, try the more constrained pool. 2400 */ 2401 if (poolp == &io_pool_4g) { 2402 poolp = &io_pool_16m; 2403 goto try_smaller; 2404 } 2405 kmem_reap(); 2406 if (++attempt < 4) { 2407 /* 2408 * Grab some more io_pool pages 2409 */ 2410 (void) populate_io_pool(); 2411 goto try_again; /* go around and retry */ 2412 } 2413 return (NULL); 2414 } 2415 /* 2416 * Found the pages, now snip them from the list 2417 */ 2418 page_io_pool_sub(poolp, pp_first, pp_last); 2419 io_pool_cnt -= minctg; 2420 /* 2421 * reset low water mark 2422 */ 2423 if (io_pool_cnt < io_pool_cnt_lowater) 2424 io_pool_cnt_lowater = io_pool_cnt; 2425 mutex_exit(&io_pool_lock); 2426 return (pp_first); 2427 } 2428 2429 page_t * 2430 page_swap_with_hypervisor(struct vnode *vp, u_offset_t off, caddr_t vaddr, 2431 ddi_dma_attr_t *mattr, uint_t flags, pgcnt_t minctg) 2432 { 2433 uint_t kflags; 2434 int order, extra, extpages, i, contig, nbits, extents; 2435 page_t *pp, *expp, *pp_first, **pplist = NULL; 2436 mfn_t *mfnlist = NULL; 2437 2438 contig = flags & PG_PHYSCONTIG; 2439 if (minctg == 1) 2440 contig = 0; 2441 flags &= ~PG_PHYSCONTIG; 2442 kflags = flags & PG_WAIT ? KM_SLEEP : KM_NOSLEEP; 2443 /* 2444 * Hypervisor will allocate extents, if we want contig 2445 * pages extent must be >= minctg 2446 */ 2447 if (contig) { 2448 order = highbit(minctg) - 1; 2449 if (minctg & ((1 << order) - 1)) 2450 order++; 2451 extpages = 1 << order; 2452 } else { 2453 order = 0; 2454 extpages = minctg; 2455 } 2456 if (extpages > minctg) { 2457 extra = extpages - minctg; 2458 if (!page_resv(extra, kflags)) 2459 return (NULL); 2460 } 2461 pp_first = NULL; 2462 pplist = kmem_alloc(extpages * sizeof (page_t *), kflags); 2463 if (pplist == NULL) 2464 goto balloon_fail; 2465 mfnlist = kmem_alloc(extpages * sizeof (mfn_t), kflags); 2466 if (mfnlist == NULL) 2467 goto balloon_fail; 2468 pp = page_create_va(vp, off, minctg * PAGESIZE, flags, &kvseg, vaddr); 2469 if (pp == NULL) 2470 goto balloon_fail; 2471 pp_first = pp; 2472 if (extpages > minctg) { 2473 /* 2474 * fill out the rest of extent pages to swap 2475 * with the hypervisor 2476 */ 2477 for (i = 0; i < extra; i++) { 2478 expp = page_create_va(vp, 2479 (u_offset_t)(uintptr_t)io_pool_kva, 2480 PAGESIZE, flags, &kvseg, io_pool_kva); 2481 if (expp == NULL) 2482 goto balloon_fail; 2483 (void) hat_pageunload(expp, HAT_FORCE_PGUNLOAD); 2484 page_io_unlock(expp); 2485 page_hashout(expp, NULL); 2486 page_io_lock(expp); 2487 /* 2488 * add page to end of list 2489 */ 2490 expp->p_prev = pp_first->p_prev; 2491 expp->p_next = pp_first; 2492 expp->p_prev->p_next = expp; 2493 pp_first->p_prev = expp; 2494 } 2495 2496 } 2497 for (i = 0; i < extpages; i++) { 2498 pplist[i] = pp; 2499 pp = pp->p_next; 2500 } 2501 nbits = highbit(mattr->dma_attr_addr_hi); 2502 extents = contig ? 1 : minctg; 2503 if (balloon_replace_pages(extents, pplist, nbits, order, 2504 mfnlist) != extents) { 2505 if (ioalloc_dbg) 2506 cmn_err(CE_NOTE, "request to hypervisor" 2507 " for %d pages, maxaddr %" PRIx64 " failed", 2508 extpages, mattr->dma_attr_addr_hi); 2509 goto balloon_fail; 2510 } 2511 2512 kmem_free(pplist, extpages * sizeof (page_t *)); 2513 kmem_free(mfnlist, extpages * sizeof (mfn_t)); 2514 /* 2515 * Return any excess pages to free list 2516 */ 2517 if (extpages > minctg) { 2518 for (i = 0; i < extra; i++) { 2519 pp = pp_first->p_prev; 2520 page_sub(&pp_first, pp); 2521 page_io_unlock(pp); 2522 page_unresv(1); 2523 page_free(pp, 1); 2524 } 2525 } 2526 return (pp_first); 2527 balloon_fail: 2528 /* 2529 * Return pages to free list and return failure 2530 */ 2531 while (pp_first != NULL) { 2532 pp = pp_first; 2533 page_sub(&pp_first, pp); 2534 page_io_unlock(pp); 2535 if (pp->p_vnode != NULL) 2536 page_hashout(pp, NULL); 2537 page_free(pp, 1); 2538 } 2539 if (pplist) 2540 kmem_free(pplist, extpages * sizeof (page_t *)); 2541 if (mfnlist) 2542 kmem_free(mfnlist, extpages * sizeof (mfn_t)); 2543 page_unresv(extpages - minctg); 2544 return (NULL); 2545 } 2546 2547 static void 2548 return_partial_alloc(page_t *plist) 2549 { 2550 page_t *pp; 2551 2552 while (plist != NULL) { 2553 pp = plist; 2554 page_sub(&plist, pp); 2555 page_destroy_io(pp); 2556 } 2557 } 2558 2559 static page_t * 2560 page_get_contigpages( 2561 struct vnode *vp, 2562 u_offset_t off, 2563 int *npagesp, 2564 uint_t flags, 2565 caddr_t vaddr, 2566 ddi_dma_attr_t *mattr) 2567 { 2568 mfn_t max_mfn = HYPERVISOR_memory_op(XENMEM_maximum_ram_page, NULL); 2569 page_t *plist; /* list to return */ 2570 page_t *pp, *mcpl; 2571 int contig, anyaddr, npages, getone = 0; 2572 mfn_t lo_mfn; 2573 mfn_t hi_mfn; 2574 pgcnt_t pfnalign = 0; 2575 int align, sgllen; 2576 uint64_t pfnseg; 2577 pgcnt_t minctg; 2578 2579 npages = *npagesp; 2580 ASSERT(mattr != NULL); 2581 lo_mfn = mmu_btop(mattr->dma_attr_addr_lo); 2582 hi_mfn = mmu_btop(mattr->dma_attr_addr_hi); 2583 sgllen = mattr->dma_attr_sgllen; 2584 pfnseg = mmu_btop(mattr->dma_attr_seg); 2585 align = maxbit(mattr->dma_attr_align, mattr->dma_attr_minxfer); 2586 if (align > MMU_PAGESIZE) 2587 pfnalign = mmu_btop(align); 2588 2589 /* 2590 * Clear the contig flag if only one page is needed. 2591 */ 2592 contig = flags & PG_PHYSCONTIG; 2593 if (npages == 1) { 2594 getone = 1; 2595 contig = 0; 2596 } 2597 2598 /* 2599 * Check if any page in the system is fine. 2600 */ 2601 anyaddr = lo_mfn == 0 && hi_mfn >= max_mfn && !pfnalign; 2602 if (!contig && anyaddr) { 2603 flags &= ~PG_PHYSCONTIG; 2604 plist = page_create_va(vp, off, npages * MMU_PAGESIZE, 2605 flags, &kvseg, vaddr); 2606 if (plist != NULL) { 2607 *npagesp = 0; 2608 return (plist); 2609 } 2610 } 2611 plist = NULL; 2612 minctg = howmany(npages, sgllen); 2613 while (npages > sgllen || getone) { 2614 if (minctg > npages) 2615 minctg = npages; 2616 mcpl = NULL; 2617 /* 2618 * We could just want unconstrained but contig pages. 2619 */ 2620 if (anyaddr && contig) { 2621 /* 2622 * Look for free contig pages to satisfy the request. 2623 */ 2624 mcpl = find_contig_free(minctg, flags, pfnseg); 2625 } 2626 /* 2627 * Try the reserved io pools next 2628 */ 2629 if (mcpl == NULL) 2630 mcpl = page_io_pool_alloc(mattr, contig, minctg); 2631 if (mcpl != NULL) { 2632 pp = mcpl; 2633 do { 2634 if (!page_hashin(pp, vp, off, NULL)) { 2635 panic("page_get_contigpages:" 2636 " hashin failed" 2637 " pp %p, vp %p, off %llx", 2638 (void *)pp, (void *)vp, off); 2639 } 2640 off += MMU_PAGESIZE; 2641 PP_CLRFREE(pp); 2642 PP_CLRAGED(pp); 2643 page_set_props(pp, P_REF); 2644 page_io_lock(pp); 2645 pp = pp->p_next; 2646 } while (pp != mcpl); 2647 } else { 2648 /* 2649 * Hypervisor exchange doesn't handle segment or 2650 * alignment constraints 2651 */ 2652 if (mattr->dma_attr_seg < mattr->dma_attr_addr_hi || 2653 pfnalign) 2654 goto fail; 2655 /* 2656 * Try exchanging pages with the hypervisor 2657 */ 2658 mcpl = page_swap_with_hypervisor(vp, off, vaddr, mattr, 2659 flags, minctg); 2660 if (mcpl == NULL) 2661 goto fail; 2662 off += minctg * MMU_PAGESIZE; 2663 } 2664 check_dma(mattr, mcpl, minctg); 2665 /* 2666 * Here with a minctg run of contiguous pages, add them to the 2667 * list we will return for this request. 2668 */ 2669 page_list_concat(&plist, &mcpl); 2670 npages -= minctg; 2671 *npagesp = npages; 2672 sgllen--; 2673 if (getone) 2674 break; 2675 } 2676 return (plist); 2677 fail: 2678 return_partial_alloc(plist); 2679 return (NULL); 2680 } 2681 2682 /* 2683 * Allocator for domain 0 I/O pages. We match the required 2684 * DMA attributes and contiguity constraints. 2685 */ 2686 /*ARGSUSED*/ 2687 page_t * 2688 page_create_io( 2689 struct vnode *vp, 2690 u_offset_t off, 2691 uint_t bytes, 2692 uint_t flags, 2693 struct as *as, 2694 caddr_t vaddr, 2695 ddi_dma_attr_t *mattr) 2696 { 2697 page_t *plist = NULL, *pp; 2698 int npages = 0, contig, anyaddr, pages_req; 2699 mfn_t lo_mfn; 2700 mfn_t hi_mfn; 2701 pgcnt_t pfnalign = 0; 2702 int align; 2703 int is_domu = 0; 2704 int dummy, bytes_got; 2705 mfn_t max_mfn = HYPERVISOR_memory_op(XENMEM_maximum_ram_page, NULL); 2706 2707 ASSERT(mattr != NULL); 2708 lo_mfn = mmu_btop(mattr->dma_attr_addr_lo); 2709 hi_mfn = mmu_btop(mattr->dma_attr_addr_hi); 2710 align = maxbit(mattr->dma_attr_align, mattr->dma_attr_minxfer); 2711 if (align > MMU_PAGESIZE) 2712 pfnalign = mmu_btop(align); 2713 2714 /* 2715 * Clear the contig flag if only one page is needed or the scatter 2716 * gather list length is >= npages. 2717 */ 2718 pages_req = npages = mmu_btopr(bytes); 2719 contig = (flags & PG_PHYSCONTIG); 2720 bytes = P2ROUNDUP(bytes, MMU_PAGESIZE); 2721 if (bytes == MMU_PAGESIZE || mattr->dma_attr_sgllen >= npages) 2722 contig = 0; 2723 2724 /* 2725 * Check if any old page in the system is fine. 2726 * DomU should always go down this path. 2727 */ 2728 is_domu = !DOMAIN_IS_INITDOMAIN(xen_info); 2729 anyaddr = lo_mfn == 0 && hi_mfn >= max_mfn && !pfnalign; 2730 if ((!contig && anyaddr) || is_domu) { 2731 flags &= ~PG_PHYSCONTIG; 2732 plist = page_create_va(vp, off, bytes, flags, &kvseg, vaddr); 2733 if (plist != NULL) 2734 return (plist); 2735 else if (is_domu) 2736 return (NULL); /* no memory available */ 2737 } 2738 /* 2739 * DomU should never reach here 2740 */ 2741 if (contig) { 2742 plist = page_get_contigpages(vp, off, &npages, flags, vaddr, 2743 mattr); 2744 if (plist == NULL) 2745 goto fail; 2746 bytes_got = (pages_req - npages) << MMU_PAGESHIFT; 2747 vaddr += bytes_got; 2748 off += bytes_got; 2749 /* 2750 * We now have all the contiguous pages we need, but 2751 * we may still need additional non-contiguous pages. 2752 */ 2753 } 2754 /* 2755 * now loop collecting the requested number of pages, these do 2756 * not have to be contiguous pages but we will use the contig 2757 * page alloc code to get the pages since it will honor any 2758 * other constraints the pages may have. 2759 */ 2760 while (npages--) { 2761 dummy = 1; 2762 pp = page_get_contigpages(vp, off, &dummy, flags, vaddr, mattr); 2763 if (pp == NULL) 2764 goto fail; 2765 page_add(&plist, pp); 2766 vaddr += MMU_PAGESIZE; 2767 off += MMU_PAGESIZE; 2768 } 2769 return (plist); 2770 fail: 2771 /* 2772 * Failed to get enough pages, return ones we did get 2773 */ 2774 return_partial_alloc(plist); 2775 return (NULL); 2776 } 2777 2778 /* 2779 * Lock and return the page with the highest mfn that we can find. last_mfn 2780 * holds the last one found, so the next search can start from there. We 2781 * also keep a counter so that we don't loop forever if the machine has no 2782 * free pages. 2783 * 2784 * This is called from the balloon thread to find pages to give away. new_high 2785 * is used when new mfn's have been added to the system - we will reset our 2786 * search if the new mfn's are higher than our current search position. 2787 */ 2788 page_t * 2789 page_get_high_mfn(mfn_t new_high) 2790 { 2791 static mfn_t last_mfn = 0; 2792 pfn_t pfn; 2793 page_t *pp; 2794 ulong_t loop_count = 0; 2795 2796 if (new_high > last_mfn) 2797 last_mfn = new_high; 2798 2799 for (; loop_count < mfn_count; loop_count++, last_mfn--) { 2800 if (last_mfn == 0) { 2801 last_mfn = cached_max_mfn; 2802 } 2803 2804 pfn = mfn_to_pfn(last_mfn); 2805 if (pfn & PFN_IS_FOREIGN_MFN) 2806 continue; 2807 2808 /* See if the page is free. If so, lock it. */ 2809 pp = page_numtopp_alloc(pfn); 2810 if (pp == NULL) 2811 continue; 2812 PP_CLRFREE(pp); 2813 2814 ASSERT(PAGE_EXCL(pp)); 2815 ASSERT(pp->p_vnode == NULL); 2816 ASSERT(!hat_page_is_mapped(pp)); 2817 last_mfn--; 2818 return (pp); 2819 } 2820 return (NULL); 2821 } 2822 2823 #else /* !__xpv */ 2824 2825 /* 2826 * get a page from any list with the given mnode 2827 */ 2828 static page_t * 2829 page_get_mnode_anylist(ulong_t origbin, uchar_t szc, uint_t flags, 2830 int mnode, int mtype, ddi_dma_attr_t *dma_attr) 2831 { 2832 kmutex_t *pcm; 2833 int i; 2834 page_t *pp; 2835 page_t *first_pp; 2836 uint64_t pgaddr; 2837 ulong_t bin; 2838 int mtypestart; 2839 int plw_initialized; 2840 page_list_walker_t plw; 2841 2842 VM_STAT_ADD(pga_vmstats.pgma_alloc); 2843 2844 ASSERT((flags & PG_MATCH_COLOR) == 0); 2845 ASSERT(szc == 0); 2846 ASSERT(dma_attr != NULL); 2847 2848 MTYPE_START(mnode, mtype, flags); 2849 if (mtype < 0) { 2850 VM_STAT_ADD(pga_vmstats.pgma_allocempty); 2851 return (NULL); 2852 } 2853 2854 mtypestart = mtype; 2855 2856 bin = origbin; 2857 2858 /* 2859 * check up to page_colors + 1 bins - origbin may be checked twice 2860 * because of BIN_STEP skip 2861 */ 2862 do { 2863 plw_initialized = 0; 2864 2865 for (plw.plw_count = 0; 2866 plw.plw_count < page_colors; plw.plw_count++) { 2867 2868 if (PAGE_FREELISTS(mnode, szc, bin, mtype) == NULL) 2869 goto nextfreebin; 2870 2871 pcm = PC_BIN_MUTEX(mnode, bin, PG_FREE_LIST); 2872 mutex_enter(pcm); 2873 pp = PAGE_FREELISTS(mnode, szc, bin, mtype); 2874 first_pp = pp; 2875 while (pp != NULL) { 2876 if (page_trylock(pp, SE_EXCL) == 0) { 2877 pp = pp->p_next; 2878 if (pp == first_pp) { 2879 pp = NULL; 2880 } 2881 continue; 2882 } 2883 2884 ASSERT(PP_ISFREE(pp)); 2885 ASSERT(PP_ISAGED(pp)); 2886 ASSERT(pp->p_vnode == NULL); 2887 ASSERT(pp->p_hash == NULL); 2888 ASSERT(pp->p_offset == (u_offset_t)-1); 2889 ASSERT(pp->p_szc == szc); 2890 ASSERT(PFN_2_MEM_NODE(pp->p_pagenum) == mnode); 2891 /* check if page within DMA attributes */ 2892 pgaddr = pa_to_ma(pfn_to_pa(pp->p_pagenum)); 2893 if ((pgaddr >= dma_attr->dma_attr_addr_lo) && 2894 (pgaddr + MMU_PAGESIZE - 1 <= 2895 dma_attr->dma_attr_addr_hi)) { 2896 break; 2897 } 2898 2899 /* continue looking */ 2900 page_unlock(pp); 2901 pp = pp->p_next; 2902 if (pp == first_pp) 2903 pp = NULL; 2904 2905 } 2906 if (pp != NULL) { 2907 ASSERT(mtype == PP_2_MTYPE(pp)); 2908 ASSERT(pp->p_szc == 0); 2909 2910 /* found a page with specified DMA attributes */ 2911 page_sub(&PAGE_FREELISTS(mnode, szc, bin, 2912 mtype), pp); 2913 page_ctr_sub(mnode, mtype, pp, PG_FREE_LIST); 2914 2915 if ((PP_ISFREE(pp) == 0) || 2916 (PP_ISAGED(pp) == 0)) { 2917 cmn_err(CE_PANIC, "page %p is not free", 2918 (void *)pp); 2919 } 2920 2921 mutex_exit(pcm); 2922 check_dma(dma_attr, pp, 1); 2923 VM_STAT_ADD(pga_vmstats.pgma_allocok); 2924 return (pp); 2925 } 2926 mutex_exit(pcm); 2927 nextfreebin: 2928 if (plw_initialized == 0) { 2929 page_list_walk_init(szc, 0, bin, 1, 0, &plw); 2930 ASSERT(plw.plw_ceq_dif == page_colors); 2931 plw_initialized = 1; 2932 } 2933 2934 if (plw.plw_do_split) { 2935 pp = page_freelist_split(szc, bin, mnode, 2936 mtype, 2937 mmu_btop(dma_attr->dma_attr_addr_hi + 1), 2938 &plw); 2939 if (pp != NULL) 2940 return (pp); 2941 } 2942 2943 bin = page_list_walk_next_bin(szc, bin, &plw); 2944 } 2945 2946 MTYPE_NEXT(mnode, mtype, flags); 2947 } while (mtype >= 0); 2948 2949 /* failed to find a page in the freelist; try it in the cachelist */ 2950 2951 /* reset mtype start for cachelist search */ 2952 mtype = mtypestart; 2953 ASSERT(mtype >= 0); 2954 2955 /* start with the bin of matching color */ 2956 bin = origbin; 2957 2958 do { 2959 for (i = 0; i <= page_colors; i++) { 2960 if (PAGE_CACHELISTS(mnode, bin, mtype) == NULL) 2961 goto nextcachebin; 2962 pcm = PC_BIN_MUTEX(mnode, bin, PG_CACHE_LIST); 2963 mutex_enter(pcm); 2964 pp = PAGE_CACHELISTS(mnode, bin, mtype); 2965 first_pp = pp; 2966 while (pp != NULL) { 2967 if (page_trylock(pp, SE_EXCL) == 0) { 2968 pp = pp->p_next; 2969 if (pp == first_pp) 2970 break; 2971 continue; 2972 } 2973 ASSERT(pp->p_vnode); 2974 ASSERT(PP_ISAGED(pp) == 0); 2975 ASSERT(pp->p_szc == 0); 2976 ASSERT(PFN_2_MEM_NODE(pp->p_pagenum) == mnode); 2977 2978 /* check if page within DMA attributes */ 2979 2980 pgaddr = pa_to_ma(pfn_to_pa(pp->p_pagenum)); 2981 if ((pgaddr >= dma_attr->dma_attr_addr_lo) && 2982 (pgaddr + MMU_PAGESIZE - 1 <= 2983 dma_attr->dma_attr_addr_hi)) { 2984 break; 2985 } 2986 2987 /* continue looking */ 2988 page_unlock(pp); 2989 pp = pp->p_next; 2990 if (pp == first_pp) 2991 pp = NULL; 2992 } 2993 2994 if (pp != NULL) { 2995 ASSERT(mtype == PP_2_MTYPE(pp)); 2996 ASSERT(pp->p_szc == 0); 2997 2998 /* found a page with specified DMA attributes */ 2999 page_sub(&PAGE_CACHELISTS(mnode, bin, 3000 mtype), pp); 3001 page_ctr_sub(mnode, mtype, pp, PG_CACHE_LIST); 3002 3003 mutex_exit(pcm); 3004 ASSERT(pp->p_vnode); 3005 ASSERT(PP_ISAGED(pp) == 0); 3006 check_dma(dma_attr, pp, 1); 3007 VM_STAT_ADD(pga_vmstats.pgma_allocok); 3008 return (pp); 3009 } 3010 mutex_exit(pcm); 3011 nextcachebin: 3012 bin += (i == 0) ? BIN_STEP : 1; 3013 bin &= page_colors_mask; 3014 } 3015 MTYPE_NEXT(mnode, mtype, flags); 3016 } while (mtype >= 0); 3017 3018 VM_STAT_ADD(pga_vmstats.pgma_allocfailed); 3019 return (NULL); 3020 } 3021 3022 /* 3023 * This function is similar to page_get_freelist()/page_get_cachelist() 3024 * but it searches both the lists to find a page with the specified 3025 * color (or no color) and DMA attributes. The search is done in the 3026 * freelist first and then in the cache list within the highest memory 3027 * range (based on DMA attributes) before searching in the lower 3028 * memory ranges. 3029 * 3030 * Note: This function is called only by page_create_io(). 3031 */ 3032 /*ARGSUSED*/ 3033 static page_t * 3034 page_get_anylist(struct vnode *vp, u_offset_t off, struct as *as, caddr_t vaddr, 3035 size_t size, uint_t flags, ddi_dma_attr_t *dma_attr, lgrp_t *lgrp) 3036 { 3037 uint_t bin; 3038 int mtype; 3039 page_t *pp; 3040 int n; 3041 int m; 3042 int szc; 3043 int fullrange; 3044 int mnode; 3045 int local_failed_stat = 0; 3046 lgrp_mnode_cookie_t lgrp_cookie; 3047 3048 VM_STAT_ADD(pga_vmstats.pga_alloc); 3049 3050 /* only base pagesize currently supported */ 3051 if (size != MMU_PAGESIZE) 3052 return (NULL); 3053 3054 /* 3055 * If we're passed a specific lgroup, we use it. Otherwise, 3056 * assume first-touch placement is desired. 3057 */ 3058 if (!LGRP_EXISTS(lgrp)) 3059 lgrp = lgrp_home_lgrp(); 3060 3061 /* LINTED */ 3062 AS_2_BIN(as, seg, vp, vaddr, bin, 0); 3063 3064 /* 3065 * Only hold one freelist or cachelist lock at a time, that way we 3066 * can start anywhere and not have to worry about lock 3067 * ordering. 3068 */ 3069 if (dma_attr == NULL) { 3070 n = 0; 3071 m = mnoderangecnt - 1; 3072 fullrange = 1; 3073 VM_STAT_ADD(pga_vmstats.pga_nulldmaattr); 3074 } else { 3075 pfn_t pfnlo = mmu_btop(dma_attr->dma_attr_addr_lo); 3076 pfn_t pfnhi = mmu_btop(dma_attr->dma_attr_addr_hi); 3077 3078 /* 3079 * We can guarantee alignment only for page boundary. 3080 */ 3081 if (dma_attr->dma_attr_align > MMU_PAGESIZE) 3082 return (NULL); 3083 3084 n = pfn_2_mtype(pfnlo); 3085 m = pfn_2_mtype(pfnhi); 3086 3087 fullrange = ((pfnlo == mnoderanges[n].mnr_pfnlo) && 3088 (pfnhi >= mnoderanges[m].mnr_pfnhi)); 3089 } 3090 VM_STAT_COND_ADD(fullrange == 0, pga_vmstats.pga_notfullrange); 3091 3092 if (n > m) 3093 return (NULL); 3094 3095 szc = 0; 3096 3097 /* cylcing thru mtype handled by RANGE0 if n == 0 */ 3098 if (n == 0) { 3099 flags |= PGI_MT_RANGE0; 3100 n = m; 3101 } 3102 3103 /* 3104 * Try local memory node first, but try remote if we can't 3105 * get a page of the right color. 3106 */ 3107 LGRP_MNODE_COOKIE_INIT(lgrp_cookie, lgrp, LGRP_SRCH_HIER); 3108 while ((mnode = lgrp_memnode_choose(&lgrp_cookie)) >= 0) { 3109 /* 3110 * allocate pages from high pfn to low. 3111 */ 3112 for (mtype = m; mtype >= n; mtype--) { 3113 if (fullrange != 0) { 3114 pp = page_get_mnode_freelist(mnode, 3115 bin, mtype, szc, flags); 3116 if (pp == NULL) { 3117 pp = page_get_mnode_cachelist( 3118 bin, flags, mnode, mtype); 3119 } 3120 } else { 3121 pp = page_get_mnode_anylist(bin, szc, 3122 flags, mnode, mtype, dma_attr); 3123 } 3124 if (pp != NULL) { 3125 VM_STAT_ADD(pga_vmstats.pga_allocok); 3126 check_dma(dma_attr, pp, 1); 3127 return (pp); 3128 } 3129 } 3130 if (!local_failed_stat) { 3131 lgrp_stat_add(lgrp->lgrp_id, LGRP_NUM_ALLOC_FAIL, 1); 3132 local_failed_stat = 1; 3133 } 3134 } 3135 VM_STAT_ADD(pga_vmstats.pga_allocfailed); 3136 3137 return (NULL); 3138 } 3139 3140 /* 3141 * page_create_io() 3142 * 3143 * This function is a copy of page_create_va() with an additional 3144 * argument 'mattr' that specifies DMA memory requirements to 3145 * the page list functions. This function is used by the segkmem 3146 * allocator so it is only to create new pages (i.e PG_EXCL is 3147 * set). 3148 * 3149 * Note: This interface is currently used by x86 PSM only and is 3150 * not fully specified so the commitment level is only for 3151 * private interface specific to x86. This interface uses PSM 3152 * specific page_get_anylist() interface. 3153 */ 3154 3155 #define PAGE_HASH_SEARCH(index, pp, vp, off) { \ 3156 for ((pp) = page_hash[(index)]; (pp); (pp) = (pp)->p_hash) { \ 3157 if ((pp)->p_vnode == (vp) && (pp)->p_offset == (off)) \ 3158 break; \ 3159 } \ 3160 } 3161 3162 3163 page_t * 3164 page_create_io( 3165 struct vnode *vp, 3166 u_offset_t off, 3167 uint_t bytes, 3168 uint_t flags, 3169 struct as *as, 3170 caddr_t vaddr, 3171 ddi_dma_attr_t *mattr) /* DMA memory attributes if any */ 3172 { 3173 page_t *plist = NULL; 3174 uint_t plist_len = 0; 3175 pgcnt_t npages; 3176 page_t *npp = NULL; 3177 uint_t pages_req; 3178 page_t *pp; 3179 kmutex_t *phm = NULL; 3180 uint_t index; 3181 3182 TRACE_4(TR_FAC_VM, TR_PAGE_CREATE_START, 3183 "page_create_start:vp %p off %llx bytes %u flags %x", 3184 vp, off, bytes, flags); 3185 3186 ASSERT((flags & ~(PG_EXCL | PG_WAIT | PG_PHYSCONTIG)) == 0); 3187 3188 pages_req = npages = mmu_btopr(bytes); 3189 3190 /* 3191 * Do the freemem and pcf accounting. 3192 */ 3193 if (!page_create_wait(npages, flags)) { 3194 return (NULL); 3195 } 3196 3197 TRACE_2(TR_FAC_VM, TR_PAGE_CREATE_SUCCESS, 3198 "page_create_success:vp %p off %llx", vp, off); 3199 3200 /* 3201 * If satisfying this request has left us with too little 3202 * memory, start the wheels turning to get some back. The 3203 * first clause of the test prevents waking up the pageout 3204 * daemon in situations where it would decide that there's 3205 * nothing to do. 3206 */ 3207 if (nscan < desscan && freemem < minfree) { 3208 TRACE_1(TR_FAC_VM, TR_PAGEOUT_CV_SIGNAL, 3209 "pageout_cv_signal:freemem %ld", freemem); 3210 cv_signal(&proc_pageout->p_cv); 3211 } 3212 3213 if (flags & PG_PHYSCONTIG) { 3214 3215 plist = page_get_contigpage(&npages, mattr, 1); 3216 if (plist == NULL) { 3217 page_create_putback(npages); 3218 return (NULL); 3219 } 3220 3221 pp = plist; 3222 3223 do { 3224 if (!page_hashin(pp, vp, off, NULL)) { 3225 panic("pg_creat_io: hashin failed %p %p %llx", 3226 (void *)pp, (void *)vp, off); 3227 } 3228 VM_STAT_ADD(page_create_new); 3229 off += MMU_PAGESIZE; 3230 PP_CLRFREE(pp); 3231 PP_CLRAGED(pp); 3232 page_set_props(pp, P_REF); 3233 pp = pp->p_next; 3234 } while (pp != plist); 3235 3236 if (!npages) { 3237 check_dma(mattr, plist, pages_req); 3238 return (plist); 3239 } else { 3240 vaddr += (pages_req - npages) << MMU_PAGESHIFT; 3241 } 3242 3243 /* 3244 * fall-thru: 3245 * 3246 * page_get_contigpage returns when npages <= sgllen. 3247 * Grab the rest of the non-contig pages below from anylist. 3248 */ 3249 } 3250 3251 /* 3252 * Loop around collecting the requested number of pages. 3253 * Most of the time, we have to `create' a new page. With 3254 * this in mind, pull the page off the free list before 3255 * getting the hash lock. This will minimize the hash 3256 * lock hold time, nesting, and the like. If it turns 3257 * out we don't need the page, we put it back at the end. 3258 */ 3259 while (npages--) { 3260 phm = NULL; 3261 3262 index = PAGE_HASH_FUNC(vp, off); 3263 top: 3264 ASSERT(phm == NULL); 3265 ASSERT(index == PAGE_HASH_FUNC(vp, off)); 3266 ASSERT(MUTEX_NOT_HELD(page_vnode_mutex(vp))); 3267 3268 if (npp == NULL) { 3269 /* 3270 * Try to get the page of any color either from 3271 * the freelist or from the cache list. 3272 */ 3273 npp = page_get_anylist(vp, off, as, vaddr, MMU_PAGESIZE, 3274 flags & ~PG_MATCH_COLOR, mattr, NULL); 3275 if (npp == NULL) { 3276 if (mattr == NULL) { 3277 /* 3278 * Not looking for a special page; 3279 * panic! 3280 */ 3281 panic("no page found %d", (int)npages); 3282 } 3283 /* 3284 * No page found! This can happen 3285 * if we are looking for a page 3286 * within a specific memory range 3287 * for DMA purposes. If PG_WAIT is 3288 * specified then we wait for a 3289 * while and then try again. The 3290 * wait could be forever if we 3291 * don't get the page(s) we need. 3292 * 3293 * Note: XXX We really need a mechanism 3294 * to wait for pages in the desired 3295 * range. For now, we wait for any 3296 * pages and see if we can use it. 3297 */ 3298 3299 if ((mattr != NULL) && (flags & PG_WAIT)) { 3300 delay(10); 3301 goto top; 3302 } 3303 goto fail; /* undo accounting stuff */ 3304 } 3305 3306 if (PP_ISAGED(npp) == 0) { 3307 /* 3308 * Since this page came from the 3309 * cachelist, we must destroy the 3310 * old vnode association. 3311 */ 3312 page_hashout(npp, (kmutex_t *)NULL); 3313 } 3314 } 3315 3316 /* 3317 * We own this page! 3318 */ 3319 ASSERT(PAGE_EXCL(npp)); 3320 ASSERT(npp->p_vnode == NULL); 3321 ASSERT(!hat_page_is_mapped(npp)); 3322 PP_CLRFREE(npp); 3323 PP_CLRAGED(npp); 3324 3325 /* 3326 * Here we have a page in our hot little mits and are 3327 * just waiting to stuff it on the appropriate lists. 3328 * Get the mutex and check to see if it really does 3329 * not exist. 3330 */ 3331 phm = PAGE_HASH_MUTEX(index); 3332 mutex_enter(phm); 3333 PAGE_HASH_SEARCH(index, pp, vp, off); 3334 if (pp == NULL) { 3335 VM_STAT_ADD(page_create_new); 3336 pp = npp; 3337 npp = NULL; 3338 if (!page_hashin(pp, vp, off, phm)) { 3339 /* 3340 * Since we hold the page hash mutex and 3341 * just searched for this page, page_hashin 3342 * had better not fail. If it does, that 3343 * means somethread did not follow the 3344 * page hash mutex rules. Panic now and 3345 * get it over with. As usual, go down 3346 * holding all the locks. 3347 */ 3348 ASSERT(MUTEX_HELD(phm)); 3349 panic("page_create: hashin fail %p %p %llx %p", 3350 (void *)pp, (void *)vp, off, (void *)phm); 3351 3352 } 3353 ASSERT(MUTEX_HELD(phm)); 3354 mutex_exit(phm); 3355 phm = NULL; 3356 3357 /* 3358 * Hat layer locking need not be done to set 3359 * the following bits since the page is not hashed 3360 * and was on the free list (i.e., had no mappings). 3361 * 3362 * Set the reference bit to protect 3363 * against immediate pageout 3364 * 3365 * XXXmh modify freelist code to set reference 3366 * bit so we don't have to do it here. 3367 */ 3368 page_set_props(pp, P_REF); 3369 } else { 3370 ASSERT(MUTEX_HELD(phm)); 3371 mutex_exit(phm); 3372 phm = NULL; 3373 /* 3374 * NOTE: This should not happen for pages associated 3375 * with kernel vnode 'kvp'. 3376 */ 3377 /* XX64 - to debug why this happens! */ 3378 ASSERT(!VN_ISKAS(vp)); 3379 if (VN_ISKAS(vp)) 3380 cmn_err(CE_NOTE, 3381 "page_create: page not expected " 3382 "in hash list for kernel vnode - pp 0x%p", 3383 (void *)pp); 3384 VM_STAT_ADD(page_create_exists); 3385 goto fail; 3386 } 3387 3388 /* 3389 * Got a page! It is locked. Acquire the i/o 3390 * lock since we are going to use the p_next and 3391 * p_prev fields to link the requested pages together. 3392 */ 3393 page_io_lock(pp); 3394 page_add(&plist, pp); 3395 plist = plist->p_next; 3396 off += MMU_PAGESIZE; 3397 vaddr += MMU_PAGESIZE; 3398 } 3399 3400 check_dma(mattr, plist, pages_req); 3401 return (plist); 3402 3403 fail: 3404 if (npp != NULL) { 3405 /* 3406 * Did not need this page after all. 3407 * Put it back on the free list. 3408 */ 3409 VM_STAT_ADD(page_create_putbacks); 3410 PP_SETFREE(npp); 3411 PP_SETAGED(npp); 3412 npp->p_offset = (u_offset_t)-1; 3413 page_list_add(npp, PG_FREE_LIST | PG_LIST_TAIL); 3414 page_unlock(npp); 3415 } 3416 3417 /* 3418 * Give up the pages we already got. 3419 */ 3420 while (plist != NULL) { 3421 pp = plist; 3422 page_sub(&plist, pp); 3423 page_io_unlock(pp); 3424 plist_len++; 3425 /*LINTED: constant in conditional ctx*/ 3426 VN_DISPOSE(pp, B_INVAL, 0, kcred); 3427 } 3428 3429 /* 3430 * VN_DISPOSE does freemem accounting for the pages in plist 3431 * by calling page_free. So, we need to undo the pcf accounting 3432 * for only the remaining pages. 3433 */ 3434 VM_STAT_ADD(page_create_putbacks); 3435 page_create_putback(pages_req - plist_len); 3436 3437 return (NULL); 3438 } 3439 #endif /* !__xpv */ 3440 3441 3442 /* 3443 * Copy the data from the physical page represented by "frompp" to 3444 * that represented by "topp". ppcopy uses CPU->cpu_caddr1 and 3445 * CPU->cpu_caddr2. It assumes that no one uses either map at interrupt 3446 * level and no one sleeps with an active mapping there. 3447 * 3448 * Note that the ref/mod bits in the page_t's are not affected by 3449 * this operation, hence it is up to the caller to update them appropriately. 3450 */ 3451 int 3452 ppcopy(page_t *frompp, page_t *topp) 3453 { 3454 caddr_t pp_addr1; 3455 caddr_t pp_addr2; 3456 hat_mempte_t pte1; 3457 hat_mempte_t pte2; 3458 kmutex_t *ppaddr_mutex; 3459 label_t ljb; 3460 int ret = 1; 3461 3462 ASSERT_STACK_ALIGNED(); 3463 ASSERT(PAGE_LOCKED(frompp)); 3464 ASSERT(PAGE_LOCKED(topp)); 3465 3466 if (kpm_enable) { 3467 pp_addr1 = hat_kpm_page2va(frompp, 0); 3468 pp_addr2 = hat_kpm_page2va(topp, 0); 3469 kpreempt_disable(); 3470 } else { 3471 /* 3472 * disable pre-emption so that CPU can't change 3473 */ 3474 kpreempt_disable(); 3475 3476 pp_addr1 = CPU->cpu_caddr1; 3477 pp_addr2 = CPU->cpu_caddr2; 3478 pte1 = CPU->cpu_caddr1pte; 3479 pte2 = CPU->cpu_caddr2pte; 3480 3481 ppaddr_mutex = &CPU->cpu_ppaddr_mutex; 3482 mutex_enter(ppaddr_mutex); 3483 3484 hat_mempte_remap(page_pptonum(frompp), pp_addr1, pte1, 3485 PROT_READ | HAT_STORECACHING_OK, HAT_LOAD_NOCONSIST); 3486 hat_mempte_remap(page_pptonum(topp), pp_addr2, pte2, 3487 PROT_READ | PROT_WRITE | HAT_STORECACHING_OK, 3488 HAT_LOAD_NOCONSIST); 3489 } 3490 3491 if (on_fault(&ljb)) { 3492 ret = 0; 3493 goto faulted; 3494 } 3495 if (use_sse_pagecopy) 3496 #ifdef __xpv 3497 page_copy_no_xmm(pp_addr2, pp_addr1); 3498 #else 3499 hwblkpagecopy(pp_addr1, pp_addr2); 3500 #endif 3501 else 3502 bcopy(pp_addr1, pp_addr2, PAGESIZE); 3503 3504 no_fault(); 3505 faulted: 3506 if (!kpm_enable) { 3507 #ifdef __xpv 3508 /* 3509 * We can't leave unused mappings laying about under the 3510 * hypervisor, so blow them away. 3511 */ 3512 if (HYPERVISOR_update_va_mapping((uintptr_t)pp_addr1, 0, 3513 UVMF_INVLPG | UVMF_LOCAL) < 0) 3514 panic("HYPERVISOR_update_va_mapping() failed"); 3515 if (HYPERVISOR_update_va_mapping((uintptr_t)pp_addr2, 0, 3516 UVMF_INVLPG | UVMF_LOCAL) < 0) 3517 panic("HYPERVISOR_update_va_mapping() failed"); 3518 #endif 3519 mutex_exit(ppaddr_mutex); 3520 } 3521 kpreempt_enable(); 3522 return (ret); 3523 } 3524 3525 void 3526 pagezero(page_t *pp, uint_t off, uint_t len) 3527 { 3528 ASSERT(PAGE_LOCKED(pp)); 3529 pfnzero(page_pptonum(pp), off, len); 3530 } 3531 3532 /* 3533 * Zero the physical page from off to off + len given by pfn 3534 * without changing the reference and modified bits of page. 3535 * 3536 * We use this using CPU private page address #2, see ppcopy() for more info. 3537 * pfnzero() must not be called at interrupt level. 3538 */ 3539 void 3540 pfnzero(pfn_t pfn, uint_t off, uint_t len) 3541 { 3542 caddr_t pp_addr2; 3543 hat_mempte_t pte2; 3544 kmutex_t *ppaddr_mutex = NULL; 3545 3546 ASSERT_STACK_ALIGNED(); 3547 ASSERT(len <= MMU_PAGESIZE); 3548 ASSERT(off <= MMU_PAGESIZE); 3549 ASSERT(off + len <= MMU_PAGESIZE); 3550 3551 if (kpm_enable && !pfn_is_foreign(pfn)) { 3552 pp_addr2 = hat_kpm_pfn2va(pfn); 3553 kpreempt_disable(); 3554 } else { 3555 kpreempt_disable(); 3556 3557 pp_addr2 = CPU->cpu_caddr2; 3558 pte2 = CPU->cpu_caddr2pte; 3559 3560 ppaddr_mutex = &CPU->cpu_ppaddr_mutex; 3561 mutex_enter(ppaddr_mutex); 3562 3563 hat_mempte_remap(pfn, pp_addr2, pte2, 3564 PROT_READ | PROT_WRITE | HAT_STORECACHING_OK, 3565 HAT_LOAD_NOCONSIST); 3566 } 3567 3568 if (use_sse_pagezero) { 3569 #ifdef __xpv 3570 uint_t rem; 3571 3572 /* 3573 * zero a byte at a time until properly aligned for 3574 * block_zero_no_xmm(). 3575 */ 3576 while (!P2NPHASE(off, ((uint_t)BLOCKZEROALIGN)) && len-- > 0) 3577 pp_addr2[off++] = 0; 3578 3579 /* 3580 * Now use faster block_zero_no_xmm() for any range 3581 * that is properly aligned and sized. 3582 */ 3583 rem = P2PHASE(len, ((uint_t)BLOCKZEROALIGN)); 3584 len -= rem; 3585 if (len != 0) { 3586 block_zero_no_xmm(pp_addr2 + off, len); 3587 off += len; 3588 } 3589 3590 /* 3591 * zero remainder with byte stores. 3592 */ 3593 while (rem-- > 0) 3594 pp_addr2[off++] = 0; 3595 #else 3596 hwblkclr(pp_addr2 + off, len); 3597 #endif 3598 } else { 3599 bzero(pp_addr2 + off, len); 3600 } 3601 3602 if (!kpm_enable || pfn_is_foreign(pfn)) { 3603 #ifdef __xpv 3604 /* 3605 * On the hypervisor this page might get used for a page 3606 * table before any intervening change to this mapping, 3607 * so blow it away. 3608 */ 3609 if (HYPERVISOR_update_va_mapping((uintptr_t)pp_addr2, 0, 3610 UVMF_INVLPG) < 0) 3611 panic("HYPERVISOR_update_va_mapping() failed"); 3612 #endif 3613 mutex_exit(ppaddr_mutex); 3614 } 3615 3616 kpreempt_enable(); 3617 } 3618 3619 /* 3620 * Platform-dependent page scrub call. 3621 */ 3622 void 3623 pagescrub(page_t *pp, uint_t off, uint_t len) 3624 { 3625 /* 3626 * For now, we rely on the fact that pagezero() will 3627 * always clear UEs. 3628 */ 3629 pagezero(pp, off, len); 3630 } 3631 3632 /* 3633 * set up two private addresses for use on a given CPU for use in ppcopy() 3634 */ 3635 void 3636 setup_vaddr_for_ppcopy(struct cpu *cpup) 3637 { 3638 void *addr; 3639 hat_mempte_t pte_pa; 3640 3641 addr = vmem_alloc(heap_arena, mmu_ptob(1), VM_SLEEP); 3642 pte_pa = hat_mempte_setup(addr); 3643 cpup->cpu_caddr1 = addr; 3644 cpup->cpu_caddr1pte = pte_pa; 3645 3646 addr = vmem_alloc(heap_arena, mmu_ptob(1), VM_SLEEP); 3647 pte_pa = hat_mempte_setup(addr); 3648 cpup->cpu_caddr2 = addr; 3649 cpup->cpu_caddr2pte = pte_pa; 3650 3651 mutex_init(&cpup->cpu_ppaddr_mutex, NULL, MUTEX_DEFAULT, NULL); 3652 } 3653 3654 /* 3655 * Undo setup_vaddr_for_ppcopy 3656 */ 3657 void 3658 teardown_vaddr_for_ppcopy(struct cpu *cpup) 3659 { 3660 mutex_destroy(&cpup->cpu_ppaddr_mutex); 3661 3662 hat_mempte_release(cpup->cpu_caddr2, cpup->cpu_caddr2pte); 3663 cpup->cpu_caddr2pte = 0; 3664 vmem_free(heap_arena, cpup->cpu_caddr2, mmu_ptob(1)); 3665 cpup->cpu_caddr2 = 0; 3666 3667 hat_mempte_release(cpup->cpu_caddr1, cpup->cpu_caddr1pte); 3668 cpup->cpu_caddr1pte = 0; 3669 vmem_free(heap_arena, cpup->cpu_caddr1, mmu_ptob(1)); 3670 cpup->cpu_caddr1 = 0; 3671 } 3672 3673 /* 3674 * Create the pageout scanner thread. The thread has to 3675 * start at procedure with process pp and priority pri. 3676 */ 3677 void 3678 pageout_init(void (*procedure)(), proc_t *pp, pri_t pri) 3679 { 3680 (void) thread_create(NULL, 0, procedure, NULL, 0, pp, TS_RUN, pri); 3681 } 3682 3683 /* 3684 * Function for flushing D-cache when performing module relocations 3685 * to an alternate mapping. Unnecessary on Intel / AMD platforms. 3686 */ 3687 void 3688 dcache_flushall() 3689 {} 3690 3691 size_t 3692 exec_get_spslew(void) 3693 { 3694 return (0); 3695 } 3696 3697 /* 3698 * Allocate a memory page. The argument 'seed' can be any pseudo-random 3699 * number to vary where the pages come from. This is quite a hacked up 3700 * method -- it works for now, but really needs to be fixed up a bit. 3701 * 3702 * We currently use page_create_va() on the kvp with fake offsets, 3703 * segments and virt address. This is pretty bogus, but was copied from the 3704 * old hat_i86.c code. A better approach would be to specify either mnode 3705 * random or mnode local and takes a page from whatever color has the MOST 3706 * available - this would have a minimal impact on page coloring. 3707 */ 3708 page_t * 3709 page_get_physical(uintptr_t seed) 3710 { 3711 page_t *pp; 3712 u_offset_t offset; 3713 static struct seg tmpseg; 3714 static uintptr_t ctr = 0; 3715 3716 /* 3717 * This code is gross, we really need a simpler page allocator. 3718 * 3719 * We need assign an offset for the page to call page_create_va(). 3720 * To avoid conflicts with other pages, we get creative with the offset. 3721 * For 32 bits, we pick an offset > 4Gig 3722 * For 64 bits, pick an offset somewhere in the VA hole. 3723 */ 3724 offset = seed; 3725 if (offset > kernelbase) 3726 offset -= kernelbase; 3727 offset <<= MMU_PAGESHIFT; 3728 #if defined(__amd64) 3729 offset += mmu.hole_start; /* something in VA hole */ 3730 #else 3731 offset += 1ULL << 40; /* something > 4 Gig */ 3732 #endif 3733 3734 if (page_resv(1, KM_NOSLEEP) == 0) 3735 return (NULL); 3736 3737 #ifdef DEBUG 3738 pp = page_exists(&kvp, offset); 3739 if (pp != NULL) 3740 panic("page already exists %p", pp); 3741 #endif 3742 3743 pp = page_create_va(&kvp, offset, MMU_PAGESIZE, PG_EXCL, 3744 &tmpseg, (caddr_t)(ctr += MMU_PAGESIZE)); /* changing VA usage */ 3745 if (pp == NULL) 3746 return (NULL); 3747 page_io_unlock(pp); 3748 page_hashout(pp, NULL); 3749 return (pp); 3750 } 3751