1/*- 2 * Copyright (c) 1991 Regents of the University of California. 3 * All rights reserved. 4 * Copyright (c) 1998 Matthew Dillon. All Rights Reserved. 5 * 6 * This code is derived from software contributed to Berkeley by 7 * The Mach Operating System project at Carnegie-Mellon University. 8 * 9 * Redistribution and use in source and binary forms, with or without 10 * modification, are permitted provided that the following conditions 11 * are met: 12 * 1. Redistributions of source code must retain the above copyright 13 * notice, this list of conditions and the following disclaimer. 14 * 2. Redistributions in binary form must reproduce the above copyright 15 * notice, this list of conditions and the following disclaimer in the 16 * documentation and/or other materials provided with the distribution. 17 * 4. Neither the name of the University nor the names of its contributors 18 * may be used to endorse or promote products derived from this software 19 * without specific prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 24 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 31 * SUCH DAMAGE. 32 * 33 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91 34 */ 35 36/*- 37 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 38 * All rights reserved. 39 * 40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 41 * 42 * Permission to use, copy, modify and distribute this software and 43 * its documentation is hereby granted, provided that both the copyright 44 * notice and this permission notice appear in all copies of the 45 * software, derivative works or modified versions, and any portions 46 * thereof, and that both notices appear in supporting documentation. 47 * 48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 51 * 52 * Carnegie Mellon requests users of this software to return to 53 * 54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 55 * School of Computer Science 56 * Carnegie Mellon University 57 * Pittsburgh PA 15213-3890 58 * 59 * any improvements or extensions that they make and grant Carnegie the 60 * rights to redistribute these changes. 61 */ 62 63/* 64 * GENERAL RULES ON VM_PAGE MANIPULATION 65 * 66 * - A page queue lock is required when adding or removing a page from a 67 * page queue regardless of other locks or the busy state of a page. 68 * 69 * * In general, no thread besides the page daemon can acquire or 70 * hold more than one page queue lock at a time. 71 * 72 * * The page daemon can acquire and hold any pair of page queue 73 * locks in any order. 74 * 75 * - The object lock is required when inserting or removing 76 * pages from an object (vm_page_insert() or vm_page_remove()). 77 * 78 */ 79 80/* 81 * Resident memory management module. 82 */ 83 84#include <sys/cdefs.h> 85__FBSDID("$FreeBSD: stable/11/sys/vm/vm_page.c 342797 2019-01-06 00:38:28Z kib $"); 86 87#include "opt_vm.h" 88 89#include <sys/param.h> 90#include <sys/systm.h> 91#include <sys/lock.h> 92#include <sys/kernel.h> 93#include <sys/limits.h> 94#include <sys/linker.h> 95#include <sys/malloc.h> 96#include <sys/mman.h> 97#include <sys/msgbuf.h> 98#include <sys/mutex.h> 99#include <sys/proc.h> 100#include <sys/rwlock.h> 101#include <sys/sbuf.h> 102#include <sys/smp.h> 103#include <sys/sysctl.h> 104#include <sys/vmmeter.h> 105#include <sys/vnode.h> 106 107#include <vm/vm.h> 108#include <vm/pmap.h> 109#include <vm/vm_param.h> 110#include <vm/vm_kern.h> 111#include <vm/vm_object.h> 112#include <vm/vm_page.h> 113#include <vm/vm_pageout.h> 114#include <vm/vm_pager.h> 115#include <vm/vm_phys.h> 116#include <vm/vm_radix.h> 117#include <vm/vm_reserv.h> 118#include <vm/vm_extern.h> 119#include <vm/uma.h> 120#include <vm/uma_int.h> 121 122#include <machine/md_var.h> 123 124/* 125 * Associated with page of user-allocatable memory is a 126 * page structure. 127 */ 128 129struct vm_domain vm_dom[MAXMEMDOM]; 130struct mtx_padalign __exclusive_cache_line vm_page_queue_free_mtx; 131 132struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT]; 133 134vm_page_t vm_page_array; 135long vm_page_array_size; 136long first_page; 137int vm_page_zero_count; 138 139static int boot_pages = UMA_BOOT_PAGES; 140SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, 141 &boot_pages, 0, 142 "number of pages allocated for bootstrapping the VM system"); 143 144static int pa_tryrelock_restart; 145SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD, 146 &pa_tryrelock_restart, 0, "Number of tryrelock restarts"); 147 148static TAILQ_HEAD(, vm_page) blacklist_head; 149static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS); 150SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD | 151 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages"); 152 153/* Is the page daemon waiting for free pages? */ 154static int vm_pageout_pages_needed; 155 156static uma_zone_t fakepg_zone; 157 158static void vm_page_alloc_check(vm_page_t m); 159static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits); 160static void vm_page_enqueue(uint8_t queue, vm_page_t m); 161static void vm_page_free_phys(vm_page_t m); 162static void vm_page_free_wakeup(void); 163static void vm_page_init_fakepg(void *dummy); 164static int vm_page_insert_after(vm_page_t m, vm_object_t object, 165 vm_pindex_t pindex, vm_page_t mpred); 166static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object, 167 vm_page_t mpred); 168static int vm_page_reclaim_run(int req_class, u_long npages, vm_page_t m_run, 169 vm_paddr_t high); 170static int vm_page_alloc_fail(vm_object_t object, int req); 171 172SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init_fakepg, NULL); 173 174static void 175vm_page_init_fakepg(void *dummy) 176{ 177 178 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL, 179 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM); 180} 181 182/* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */ 183#if PAGE_SIZE == 32768 184#ifdef CTASSERT 185CTASSERT(sizeof(u_long) >= 8); 186#endif 187#endif 188 189/* 190 * Try to acquire a physical address lock while a pmap is locked. If we 191 * fail to trylock we unlock and lock the pmap directly and cache the 192 * locked pa in *locked. The caller should then restart their loop in case 193 * the virtual to physical mapping has changed. 194 */ 195int 196vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked) 197{ 198 vm_paddr_t lockpa; 199 200 lockpa = *locked; 201 *locked = pa; 202 if (lockpa) { 203 PA_LOCK_ASSERT(lockpa, MA_OWNED); 204 if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa)) 205 return (0); 206 PA_UNLOCK(lockpa); 207 } 208 if (PA_TRYLOCK(pa)) 209 return (0); 210 PMAP_UNLOCK(pmap); 211 atomic_add_int(&pa_tryrelock_restart, 1); 212 PA_LOCK(pa); 213 PMAP_LOCK(pmap); 214 return (EAGAIN); 215} 216 217/* 218 * vm_set_page_size: 219 * 220 * Sets the page size, perhaps based upon the memory 221 * size. Must be called before any use of page-size 222 * dependent functions. 223 */ 224void 225vm_set_page_size(void) 226{ 227 if (vm_cnt.v_page_size == 0) 228 vm_cnt.v_page_size = PAGE_SIZE; 229 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0) 230 panic("vm_set_page_size: page size not a power of two"); 231} 232 233/* 234 * vm_page_blacklist_next: 235 * 236 * Find the next entry in the provided string of blacklist 237 * addresses. Entries are separated by space, comma, or newline. 238 * If an invalid integer is encountered then the rest of the 239 * string is skipped. Updates the list pointer to the next 240 * character, or NULL if the string is exhausted or invalid. 241 */ 242static vm_paddr_t 243vm_page_blacklist_next(char **list, char *end) 244{ 245 vm_paddr_t bad; 246 char *cp, *pos; 247 248 if (list == NULL || *list == NULL) 249 return (0); 250 if (**list =='\0') { 251 *list = NULL; 252 return (0); 253 } 254 255 /* 256 * If there's no end pointer then the buffer is coming from 257 * the kenv and we know it's null-terminated. 258 */ 259 if (end == NULL) 260 end = *list + strlen(*list); 261 262 /* Ensure that strtoq() won't walk off the end */ 263 if (*end != '\0') { 264 if (*end == '\n' || *end == ' ' || *end == ',') 265 *end = '\0'; 266 else { 267 printf("Blacklist not terminated, skipping\n"); 268 *list = NULL; 269 return (0); 270 } 271 } 272 273 for (pos = *list; *pos != '\0'; pos = cp) { 274 bad = strtoq(pos, &cp, 0); 275 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') { 276 if (bad == 0) { 277 if (++cp < end) 278 continue; 279 else 280 break; 281 } 282 } else 283 break; 284 if (*cp == '\0' || ++cp >= end) 285 *list = NULL; 286 else 287 *list = cp; 288 return (trunc_page(bad)); 289 } 290 printf("Garbage in RAM blacklist, skipping\n"); 291 *list = NULL; 292 return (0); 293} 294 295bool 296vm_page_blacklist_add(vm_paddr_t pa, bool verbose) 297{ 298 vm_page_t m; 299 int ret; 300 301 m = vm_phys_paddr_to_vm_page(pa); 302 if (m == NULL) 303 return (true); /* page does not exist, no failure */ 304 305 mtx_lock(&vm_page_queue_free_mtx); 306 ret = vm_phys_unfree_page(m); 307 if (ret != 0) 308 vm_phys_freecnt_adj(m, -1); 309 mtx_unlock(&vm_page_queue_free_mtx); 310 if (ret != 0) { 311 TAILQ_INSERT_TAIL(&blacklist_head, m, listq); 312 if (verbose) 313 printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa); 314 } 315 return (ret); 316} 317 318/* 319 * vm_page_blacklist_check: 320 * 321 * Iterate through the provided string of blacklist addresses, pulling 322 * each entry out of the physical allocator free list and putting it 323 * onto a list for reporting via the vm.page_blacklist sysctl. 324 */ 325static void 326vm_page_blacklist_check(char *list, char *end) 327{ 328 vm_paddr_t pa; 329 char *next; 330 331 next = list; 332 while (next != NULL) { 333 if ((pa = vm_page_blacklist_next(&next, end)) == 0) 334 continue; 335 vm_page_blacklist_add(pa, bootverbose); 336 } 337} 338 339/* 340 * vm_page_blacklist_load: 341 * 342 * Search for a special module named "ram_blacklist". It'll be a 343 * plain text file provided by the user via the loader directive 344 * of the same name. 345 */ 346static void 347vm_page_blacklist_load(char **list, char **end) 348{ 349 void *mod; 350 u_char *ptr; 351 u_int len; 352 353 mod = NULL; 354 ptr = NULL; 355 356 mod = preload_search_by_type("ram_blacklist"); 357 if (mod != NULL) { 358 ptr = preload_fetch_addr(mod); 359 len = preload_fetch_size(mod); 360 } 361 *list = ptr; 362 if (ptr != NULL) 363 *end = ptr + len; 364 else 365 *end = NULL; 366 return; 367} 368 369static int 370sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS) 371{ 372 vm_page_t m; 373 struct sbuf sbuf; 374 int error, first; 375 376 first = 1; 377 error = sysctl_wire_old_buffer(req, 0); 378 if (error != 0) 379 return (error); 380 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 381 TAILQ_FOREACH(m, &blacklist_head, listq) { 382 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",", 383 (uintmax_t)m->phys_addr); 384 first = 0; 385 } 386 error = sbuf_finish(&sbuf); 387 sbuf_delete(&sbuf); 388 return (error); 389} 390 391static void 392vm_page_domain_init(struct vm_domain *vmd) 393{ 394 struct vm_pagequeue *pq; 395 int i; 396 397 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) = 398 "vm inactive pagequeue"; 399 *__DECONST(u_int **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_vcnt) = 400 &vm_cnt.v_inactive_count; 401 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) = 402 "vm active pagequeue"; 403 *__DECONST(u_int **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_vcnt) = 404 &vm_cnt.v_active_count; 405 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) = 406 "vm laundry pagequeue"; 407 *__DECONST(int **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_vcnt) = 408 &vm_cnt.v_laundry_count; 409 vmd->vmd_page_count = 0; 410 vmd->vmd_free_count = 0; 411 vmd->vmd_segs = 0; 412 vmd->vmd_oom = FALSE; 413 for (i = 0; i < PQ_COUNT; i++) { 414 pq = &vmd->vmd_pagequeues[i]; 415 TAILQ_INIT(&pq->pq_pl); 416 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue", 417 MTX_DEF | MTX_DUPOK); 418 } 419} 420 421/* 422 * Initialize a physical page in preparation for adding it to the free 423 * lists. 424 */ 425static void 426vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind) 427{ 428 429 m->object = NULL; 430 m->wire_count = 0; 431 m->busy_lock = VPB_UNBUSIED; 432 m->hold_count = 0; 433 m->flags = 0; 434 m->phys_addr = pa; 435 m->queue = PQ_NONE; 436 m->psind = 0; 437 m->segind = segind; 438 m->order = VM_NFREEORDER; 439 m->pool = VM_FREEPOOL_DEFAULT; 440 m->valid = m->dirty = 0; 441 pmap_page_init(m); 442} 443 444/* 445 * vm_page_startup: 446 * 447 * Initializes the resident memory module. Allocates physical memory for 448 * bootstrapping UMA and some data structures that are used to manage 449 * physical pages. Initializes these structures, and populates the free 450 * page queues. 451 */ 452vm_offset_t 453vm_page_startup(vm_offset_t vaddr) 454{ 455 struct vm_domain *vmd; 456 struct vm_phys_seg *seg; 457 vm_page_t m; 458 char *list, *listend; 459 vm_offset_t mapped; 460 vm_paddr_t end, high_avail, low_avail, new_end, page_range, size; 461 vm_paddr_t biggestsize, last_pa, pa; 462 u_long pagecount; 463 int biggestone, i, pages_per_zone, segind; 464 465 biggestsize = 0; 466 biggestone = 0; 467 vaddr = round_page(vaddr); 468 469 for (i = 0; phys_avail[i + 1]; i += 2) { 470 phys_avail[i] = round_page(phys_avail[i]); 471 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 472 } 473 for (i = 0; phys_avail[i + 1]; i += 2) { 474 size = phys_avail[i + 1] - phys_avail[i]; 475 if (size > biggestsize) { 476 biggestone = i; 477 biggestsize = size; 478 } 479 } 480 481 end = phys_avail[biggestone+1]; 482 483 /* 484 * Initialize the page and queue locks. 485 */ 486 mtx_init(&vm_page_queue_free_mtx, "vm page free queue", NULL, MTX_DEF); 487 for (i = 0; i < PA_LOCK_COUNT; i++) 488 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF); 489 for (i = 0; i < vm_ndomains; i++) 490 vm_page_domain_init(&vm_dom[i]); 491 492 /* 493 * Almost all of the pages needed for bootstrapping UMA are used 494 * for zone structures, so if the number of CPUs results in those 495 * structures taking more than one page each, we set aside more pages 496 * in proportion to the zone structure size. 497 */ 498 pages_per_zone = howmany(sizeof(struct uma_zone) + 499 sizeof(struct uma_cache) * (mp_maxid + 1) + 500 roundup2(sizeof(struct uma_slab), sizeof(void *)), UMA_SLAB_SIZE); 501 if (pages_per_zone > 1) { 502 /* Reserve more pages so that we don't run out. */ 503 boot_pages = UMA_BOOT_PAGES_ZONES * pages_per_zone; 504 } 505 506 /* 507 * Allocate memory for use when boot strapping the kernel memory 508 * allocator. 509 * 510 * CTFLAG_RDTUN doesn't work during the early boot process, so we must 511 * manually fetch the value. 512 */ 513 TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages); 514 new_end = end - (boot_pages * UMA_SLAB_SIZE); 515 new_end = trunc_page(new_end); 516 mapped = pmap_map(&vaddr, new_end, end, 517 VM_PROT_READ | VM_PROT_WRITE); 518 bzero((void *)mapped, end - new_end); 519 uma_startup((void *)mapped, boot_pages); 520 521#if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \ 522 defined(__i386__) || defined(__mips__) 523 /* 524 * Allocate a bitmap to indicate that a random physical page 525 * needs to be included in a minidump. 526 * 527 * The amd64 port needs this to indicate which direct map pages 528 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 529 * 530 * However, i386 still needs this workspace internally within the 531 * minidump code. In theory, they are not needed on i386, but are 532 * included should the sf_buf code decide to use them. 533 */ 534 last_pa = 0; 535 for (i = 0; dump_avail[i + 1] != 0; i += 2) 536 if (dump_avail[i + 1] > last_pa) 537 last_pa = dump_avail[i + 1]; 538 page_range = last_pa / PAGE_SIZE; 539 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 540 new_end -= vm_page_dump_size; 541 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, 542 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); 543 bzero((void *)vm_page_dump, vm_page_dump_size); 544#else 545 (void)last_pa; 546#endif 547#if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) 548 /* 549 * Include the UMA bootstrap pages and vm_page_dump in a crash dump. 550 * When pmap_map() uses the direct map, they are not automatically 551 * included. 552 */ 553 for (pa = new_end; pa < end; pa += PAGE_SIZE) 554 dump_add_page(pa); 555#endif 556 phys_avail[biggestone + 1] = new_end; 557#ifdef __amd64__ 558 /* 559 * Request that the physical pages underlying the message buffer be 560 * included in a crash dump. Since the message buffer is accessed 561 * through the direct map, they are not automatically included. 562 */ 563 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); 564 last_pa = pa + round_page(msgbufsize); 565 while (pa < last_pa) { 566 dump_add_page(pa); 567 pa += PAGE_SIZE; 568 } 569#endif 570 /* 571 * Compute the number of pages of memory that will be available for 572 * use, taking into account the overhead of a page structure per page. 573 * In other words, solve 574 * "available physical memory" - round_page(page_range * 575 * sizeof(struct vm_page)) = page_range * PAGE_SIZE 576 * for page_range. 577 */ 578 low_avail = phys_avail[0]; 579 high_avail = phys_avail[1]; 580 for (i = 0; i < vm_phys_nsegs; i++) { 581 if (vm_phys_segs[i].start < low_avail) 582 low_avail = vm_phys_segs[i].start; 583 if (vm_phys_segs[i].end > high_avail) 584 high_avail = vm_phys_segs[i].end; 585 } 586 /* Skip the first chunk. It is already accounted for. */ 587 for (i = 2; phys_avail[i + 1] != 0; i += 2) { 588 if (phys_avail[i] < low_avail) 589 low_avail = phys_avail[i]; 590 if (phys_avail[i + 1] > high_avail) 591 high_avail = phys_avail[i + 1]; 592 } 593 first_page = low_avail / PAGE_SIZE; 594#ifdef VM_PHYSSEG_SPARSE 595 size = 0; 596 for (i = 0; i < vm_phys_nsegs; i++) 597 size += vm_phys_segs[i].end - vm_phys_segs[i].start; 598 for (i = 0; phys_avail[i + 1] != 0; i += 2) 599 size += phys_avail[i + 1] - phys_avail[i]; 600#elif defined(VM_PHYSSEG_DENSE) 601 size = high_avail - low_avail; 602#else 603#error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 604#endif 605 606#ifdef VM_PHYSSEG_DENSE 607 /* 608 * In the VM_PHYSSEG_DENSE case, the number of pages can account for 609 * the overhead of a page structure per page only if vm_page_array is 610 * allocated from the last physical memory chunk. Otherwise, we must 611 * allocate page structures representing the physical memory 612 * underlying vm_page_array, even though they will not be used. 613 */ 614 if (new_end != high_avail) 615 page_range = size / PAGE_SIZE; 616 else 617#endif 618 { 619 page_range = size / (PAGE_SIZE + sizeof(struct vm_page)); 620 621 /* 622 * If the partial bytes remaining are large enough for 623 * a page (PAGE_SIZE) without a corresponding 624 * 'struct vm_page', then new_end will contain an 625 * extra page after subtracting the length of the VM 626 * page array. Compensate by subtracting an extra 627 * page from new_end. 628 */ 629 if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) { 630 if (new_end == high_avail) 631 high_avail -= PAGE_SIZE; 632 new_end -= PAGE_SIZE; 633 } 634 } 635 end = new_end; 636 637 /* 638 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 639 * However, because this page is allocated from KVM, out-of-bounds 640 * accesses using the direct map will not be trapped. 641 */ 642 vaddr += PAGE_SIZE; 643 644 /* 645 * Allocate physical memory for the page structures, and map it. 646 */ 647 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 648 mapped = pmap_map(&vaddr, new_end, end, 649 VM_PROT_READ | VM_PROT_WRITE); 650 vm_page_array = (vm_page_t)mapped; 651 vm_page_array_size = page_range; 652 653#if VM_NRESERVLEVEL > 0 654 /* 655 * Allocate physical memory for the reservation management system's 656 * data structures, and map it. 657 */ 658 if (high_avail == end) 659 high_avail = new_end; 660 new_end = vm_reserv_startup(&vaddr, new_end, high_avail); 661#endif 662#if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) 663 /* 664 * Include vm_page_array and vm_reserv_array in a crash dump. 665 */ 666 for (pa = new_end; pa < end; pa += PAGE_SIZE) 667 dump_add_page(pa); 668#endif 669 phys_avail[biggestone + 1] = new_end; 670 671 /* 672 * Add physical memory segments corresponding to the available 673 * physical pages. 674 */ 675 for (i = 0; phys_avail[i + 1] != 0; i += 2) 676 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]); 677 678 /* 679 * Initialize the physical memory allocator. 680 */ 681 vm_phys_init(); 682 683 /* 684 * Initialize the page structures and add every available page to the 685 * physical memory allocator's free lists. 686 */ 687 vm_cnt.v_page_count = 0; 688 vm_cnt.v_free_count = 0; 689 for (segind = 0; segind < vm_phys_nsegs; segind++) { 690 seg = &vm_phys_segs[segind]; 691 for (m = seg->first_page, pa = seg->start; pa < seg->end; 692 m++, pa += PAGE_SIZE) 693 vm_page_init_page(m, pa, segind); 694 695 /* 696 * Add the segment to the free lists only if it is covered by 697 * one of the ranges in phys_avail. Because we've added the 698 * ranges to the vm_phys_segs array, we can assume that each 699 * segment is either entirely contained in one of the ranges, 700 * or doesn't overlap any of them. 701 */ 702 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 703 if (seg->start < phys_avail[i] || 704 seg->end > phys_avail[i + 1]) 705 continue; 706 707 m = seg->first_page; 708 pagecount = (u_long)atop(seg->end - seg->start); 709 710 mtx_lock(&vm_page_queue_free_mtx); 711 vm_phys_free_contig(m, pagecount); 712 vm_phys_freecnt_adj(m, (int)pagecount); 713 mtx_unlock(&vm_page_queue_free_mtx); 714 vm_cnt.v_page_count += (u_int)pagecount; 715 716 vmd = &vm_dom[seg->domain]; 717 vmd->vmd_page_count += (u_int)pagecount; 718 vmd->vmd_segs |= 1UL << m->segind; 719 break; 720 } 721 } 722 723 /* 724 * Remove blacklisted pages from the physical memory allocator. 725 */ 726 TAILQ_INIT(&blacklist_head); 727 vm_page_blacklist_load(&list, &listend); 728 vm_page_blacklist_check(list, listend); 729 730 list = kern_getenv("vm.blacklist"); 731 vm_page_blacklist_check(list, NULL); 732 733 freeenv(list); 734#if VM_NRESERVLEVEL > 0 735 /* 736 * Initialize the reservation management system. 737 */ 738 vm_reserv_init(); 739#endif 740 return (vaddr); 741} 742 743void 744vm_page_reference(vm_page_t m) 745{ 746 747 vm_page_aflag_set(m, PGA_REFERENCED); 748} 749 750/* 751 * vm_page_busy_downgrade: 752 * 753 * Downgrade an exclusive busy page into a single shared busy page. 754 */ 755void 756vm_page_busy_downgrade(vm_page_t m) 757{ 758 u_int x; 759 bool locked; 760 761 vm_page_assert_xbusied(m); 762 locked = mtx_owned(vm_page_lockptr(m)); 763 764 for (;;) { 765 x = m->busy_lock; 766 x &= VPB_BIT_WAITERS; 767 if (x != 0 && !locked) 768 vm_page_lock(m); 769 if (atomic_cmpset_rel_int(&m->busy_lock, 770 VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1))) 771 break; 772 if (x != 0 && !locked) 773 vm_page_unlock(m); 774 } 775 if (x != 0) { 776 wakeup(m); 777 if (!locked) 778 vm_page_unlock(m); 779 } 780} 781 782/* 783 * vm_page_sbusied: 784 * 785 * Return a positive value if the page is shared busied, 0 otherwise. 786 */ 787int 788vm_page_sbusied(vm_page_t m) 789{ 790 u_int x; 791 792 x = m->busy_lock; 793 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED); 794} 795 796/* 797 * vm_page_sunbusy: 798 * 799 * Shared unbusy a page. 800 */ 801void 802vm_page_sunbusy(vm_page_t m) 803{ 804 u_int x; 805 806 vm_page_lock_assert(m, MA_NOTOWNED); 807 vm_page_assert_sbusied(m); 808 809 for (;;) { 810 x = m->busy_lock; 811 if (VPB_SHARERS(x) > 1) { 812 if (atomic_cmpset_int(&m->busy_lock, x, 813 x - VPB_ONE_SHARER)) 814 break; 815 continue; 816 } 817 if ((x & VPB_BIT_WAITERS) == 0) { 818 KASSERT(x == VPB_SHARERS_WORD(1), 819 ("vm_page_sunbusy: invalid lock state")); 820 if (atomic_cmpset_int(&m->busy_lock, 821 VPB_SHARERS_WORD(1), VPB_UNBUSIED)) 822 break; 823 continue; 824 } 825 KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS), 826 ("vm_page_sunbusy: invalid lock state for waiters")); 827 828 vm_page_lock(m); 829 if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) { 830 vm_page_unlock(m); 831 continue; 832 } 833 wakeup(m); 834 vm_page_unlock(m); 835 break; 836 } 837} 838 839/* 840 * vm_page_busy_sleep: 841 * 842 * Sleep and release the page lock, using the page pointer as wchan. 843 * This is used to implement the hard-path of busying mechanism. 844 * 845 * The given page must be locked. 846 * 847 * If nonshared is true, sleep only if the page is xbusy. 848 */ 849void 850vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared) 851{ 852 u_int x; 853 854 vm_page_assert_locked(m); 855 856 x = m->busy_lock; 857 if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) || 858 ((x & VPB_BIT_WAITERS) == 0 && 859 !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) { 860 vm_page_unlock(m); 861 return; 862 } 863 msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0); 864} 865 866/* 867 * vm_page_trysbusy: 868 * 869 * Try to shared busy a page. 870 * If the operation succeeds 1 is returned otherwise 0. 871 * The operation never sleeps. 872 */ 873int 874vm_page_trysbusy(vm_page_t m) 875{ 876 u_int x; 877 878 for (;;) { 879 x = m->busy_lock; 880 if ((x & VPB_BIT_SHARED) == 0) 881 return (0); 882 if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER)) 883 return (1); 884 } 885} 886 887static void 888vm_page_xunbusy_locked(vm_page_t m) 889{ 890 891 vm_page_assert_xbusied(m); 892 vm_page_assert_locked(m); 893 894 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); 895 /* There is a waiter, do wakeup() instead of vm_page_flash(). */ 896 wakeup(m); 897} 898 899void 900vm_page_xunbusy_maybelocked(vm_page_t m) 901{ 902 bool lockacq; 903 904 vm_page_assert_xbusied(m); 905 906 /* 907 * Fast path for unbusy. If it succeeds, we know that there 908 * are no waiters, so we do not need a wakeup. 909 */ 910 if (atomic_cmpset_rel_int(&m->busy_lock, VPB_SINGLE_EXCLUSIVER, 911 VPB_UNBUSIED)) 912 return; 913 914 lockacq = !mtx_owned(vm_page_lockptr(m)); 915 if (lockacq) 916 vm_page_lock(m); 917 vm_page_xunbusy_locked(m); 918 if (lockacq) 919 vm_page_unlock(m); 920} 921 922/* 923 * vm_page_xunbusy_hard: 924 * 925 * Called after the first try the exclusive unbusy of a page failed. 926 * It is assumed that the waiters bit is on. 927 */ 928void 929vm_page_xunbusy_hard(vm_page_t m) 930{ 931 932 vm_page_assert_xbusied(m); 933 934 vm_page_lock(m); 935 vm_page_xunbusy_locked(m); 936 vm_page_unlock(m); 937} 938 939/* 940 * vm_page_flash: 941 * 942 * Wakeup anyone waiting for the page. 943 * The ownership bits do not change. 944 * 945 * The given page must be locked. 946 */ 947void 948vm_page_flash(vm_page_t m) 949{ 950 u_int x; 951 952 vm_page_lock_assert(m, MA_OWNED); 953 954 for (;;) { 955 x = m->busy_lock; 956 if ((x & VPB_BIT_WAITERS) == 0) 957 return; 958 if (atomic_cmpset_int(&m->busy_lock, x, 959 x & (~VPB_BIT_WAITERS))) 960 break; 961 } 962 wakeup(m); 963} 964 965/* 966 * Avoid releasing and reacquiring the same page lock. 967 */ 968void 969vm_page_change_lock(vm_page_t m, struct mtx **mtx) 970{ 971 struct mtx *mtx1; 972 973 mtx1 = vm_page_lockptr(m); 974 if (*mtx == mtx1) 975 return; 976 if (*mtx != NULL) 977 mtx_unlock(*mtx); 978 *mtx = mtx1; 979 mtx_lock(mtx1); 980} 981 982/* 983 * Keep page from being freed by the page daemon 984 * much of the same effect as wiring, except much lower 985 * overhead and should be used only for *very* temporary 986 * holding ("wiring"). 987 */ 988void 989vm_page_hold(vm_page_t mem) 990{ 991 992 vm_page_lock_assert(mem, MA_OWNED); 993 mem->hold_count++; 994} 995 996void 997vm_page_unhold(vm_page_t mem) 998{ 999 1000 vm_page_lock_assert(mem, MA_OWNED); 1001 KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!")); 1002 --mem->hold_count; 1003 if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0) 1004 vm_page_free_toq(mem); 1005} 1006 1007/* 1008 * vm_page_unhold_pages: 1009 * 1010 * Unhold each of the pages that is referenced by the given array. 1011 */ 1012void 1013vm_page_unhold_pages(vm_page_t *ma, int count) 1014{ 1015 struct mtx *mtx; 1016 1017 mtx = NULL; 1018 for (; count != 0; count--) { 1019 vm_page_change_lock(*ma, &mtx); 1020 vm_page_unhold(*ma); 1021 ma++; 1022 } 1023 if (mtx != NULL) 1024 mtx_unlock(mtx); 1025} 1026 1027vm_page_t 1028PHYS_TO_VM_PAGE(vm_paddr_t pa) 1029{ 1030 vm_page_t m; 1031 1032#ifdef VM_PHYSSEG_SPARSE 1033 m = vm_phys_paddr_to_vm_page(pa); 1034 if (m == NULL) 1035 m = vm_phys_fictitious_to_vm_page(pa); 1036 return (m); 1037#elif defined(VM_PHYSSEG_DENSE) 1038 long pi; 1039 1040 pi = atop(pa); 1041 if (pi >= first_page && (pi - first_page) < vm_page_array_size) { 1042 m = &vm_page_array[pi - first_page]; 1043 return (m); 1044 } 1045 return (vm_phys_fictitious_to_vm_page(pa)); 1046#else 1047#error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 1048#endif 1049} 1050 1051/* 1052 * vm_page_getfake: 1053 * 1054 * Create a fictitious page with the specified physical address and 1055 * memory attribute. The memory attribute is the only the machine- 1056 * dependent aspect of a fictitious page that must be initialized. 1057 */ 1058vm_page_t 1059vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr) 1060{ 1061 vm_page_t m; 1062 1063 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO); 1064 vm_page_initfake(m, paddr, memattr); 1065 return (m); 1066} 1067 1068void 1069vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1070{ 1071 1072 if ((m->flags & PG_FICTITIOUS) != 0) { 1073 /* 1074 * The page's memattr might have changed since the 1075 * previous initialization. Update the pmap to the 1076 * new memattr. 1077 */ 1078 goto memattr; 1079 } 1080 m->phys_addr = paddr; 1081 m->queue = PQ_NONE; 1082 /* Fictitious pages don't use "segind". */ 1083 m->flags = PG_FICTITIOUS; 1084 /* Fictitious pages don't use "order" or "pool". */ 1085 m->oflags = VPO_UNMANAGED; 1086 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 1087 m->wire_count = 1; 1088 pmap_page_init(m); 1089memattr: 1090 pmap_page_set_memattr(m, memattr); 1091} 1092 1093/* 1094 * vm_page_putfake: 1095 * 1096 * Release a fictitious page. 1097 */ 1098void 1099vm_page_putfake(vm_page_t m) 1100{ 1101 1102 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m)); 1103 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1104 ("vm_page_putfake: bad page %p", m)); 1105 uma_zfree(fakepg_zone, m); 1106} 1107 1108/* 1109 * vm_page_updatefake: 1110 * 1111 * Update the given fictitious page to the specified physical address and 1112 * memory attribute. 1113 */ 1114void 1115vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1116{ 1117 1118 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1119 ("vm_page_updatefake: bad page %p", m)); 1120 m->phys_addr = paddr; 1121 pmap_page_set_memattr(m, memattr); 1122} 1123 1124/* 1125 * vm_page_free: 1126 * 1127 * Free a page. 1128 */ 1129void 1130vm_page_free(vm_page_t m) 1131{ 1132 1133 m->flags &= ~PG_ZERO; 1134 vm_page_free_toq(m); 1135} 1136 1137/* 1138 * vm_page_free_zero: 1139 * 1140 * Free a page to the zerod-pages queue 1141 */ 1142void 1143vm_page_free_zero(vm_page_t m) 1144{ 1145 1146 m->flags |= PG_ZERO; 1147 vm_page_free_toq(m); 1148} 1149 1150/* 1151 * Unbusy and handle the page queueing for a page from a getpages request that 1152 * was optionally read ahead or behind. 1153 */ 1154void 1155vm_page_readahead_finish(vm_page_t m) 1156{ 1157 1158 /* We shouldn't put invalid pages on queues. */ 1159 KASSERT(m->valid != 0, ("%s: %p is invalid", __func__, m)); 1160 1161 /* 1162 * Since the page is not the actually needed one, whether it should 1163 * be activated or deactivated is not obvious. Empirical results 1164 * have shown that deactivating the page is usually the best choice, 1165 * unless the page is wanted by another thread. 1166 */ 1167 vm_page_lock(m); 1168 if ((m->busy_lock & VPB_BIT_WAITERS) != 0) 1169 vm_page_activate(m); 1170 else 1171 vm_page_deactivate(m); 1172 vm_page_unlock(m); 1173 vm_page_xunbusy(m); 1174} 1175 1176/* 1177 * vm_page_sleep_if_busy: 1178 * 1179 * Sleep and release the page queues lock if the page is busied. 1180 * Returns TRUE if the thread slept. 1181 * 1182 * The given page must be unlocked and object containing it must 1183 * be locked. 1184 */ 1185int 1186vm_page_sleep_if_busy(vm_page_t m, const char *msg) 1187{ 1188 vm_object_t obj; 1189 1190 vm_page_lock_assert(m, MA_NOTOWNED); 1191 VM_OBJECT_ASSERT_WLOCKED(m->object); 1192 1193 if (vm_page_busied(m)) { 1194 /* 1195 * The page-specific object must be cached because page 1196 * identity can change during the sleep, causing the 1197 * re-lock of a different object. 1198 * It is assumed that a reference to the object is already 1199 * held by the callers. 1200 */ 1201 obj = m->object; 1202 vm_page_lock(m); 1203 VM_OBJECT_WUNLOCK(obj); 1204 vm_page_busy_sleep(m, msg, false); 1205 VM_OBJECT_WLOCK(obj); 1206 return (TRUE); 1207 } 1208 return (FALSE); 1209} 1210 1211/* 1212 * vm_page_dirty_KBI: [ internal use only ] 1213 * 1214 * Set all bits in the page's dirty field. 1215 * 1216 * The object containing the specified page must be locked if the 1217 * call is made from the machine-independent layer. 1218 * 1219 * See vm_page_clear_dirty_mask(). 1220 * 1221 * This function should only be called by vm_page_dirty(). 1222 */ 1223void 1224vm_page_dirty_KBI(vm_page_t m) 1225{ 1226 1227 /* Refer to this operation by its public name. */ 1228 KASSERT(m->valid == VM_PAGE_BITS_ALL, 1229 ("vm_page_dirty: page is invalid!")); 1230 m->dirty = VM_PAGE_BITS_ALL; 1231} 1232 1233/* 1234 * vm_page_insert: [ internal use only ] 1235 * 1236 * Inserts the given mem entry into the object and object list. 1237 * 1238 * The object must be locked. 1239 */ 1240int 1241vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 1242{ 1243 vm_page_t mpred; 1244 1245 VM_OBJECT_ASSERT_WLOCKED(object); 1246 mpred = vm_radix_lookup_le(&object->rtree, pindex); 1247 return (vm_page_insert_after(m, object, pindex, mpred)); 1248} 1249 1250/* 1251 * vm_page_insert_after: 1252 * 1253 * Inserts the page "m" into the specified object at offset "pindex". 1254 * 1255 * The page "mpred" must immediately precede the offset "pindex" within 1256 * the specified object. 1257 * 1258 * The object must be locked. 1259 */ 1260static int 1261vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, 1262 vm_page_t mpred) 1263{ 1264 vm_page_t msucc; 1265 1266 VM_OBJECT_ASSERT_WLOCKED(object); 1267 KASSERT(m->object == NULL, 1268 ("vm_page_insert_after: page already inserted")); 1269 if (mpred != NULL) { 1270 KASSERT(mpred->object == object, 1271 ("vm_page_insert_after: object doesn't contain mpred")); 1272 KASSERT(mpred->pindex < pindex, 1273 ("vm_page_insert_after: mpred doesn't precede pindex")); 1274 msucc = TAILQ_NEXT(mpred, listq); 1275 } else 1276 msucc = TAILQ_FIRST(&object->memq); 1277 if (msucc != NULL) 1278 KASSERT(msucc->pindex > pindex, 1279 ("vm_page_insert_after: msucc doesn't succeed pindex")); 1280 1281 /* 1282 * Record the object/offset pair in this page 1283 */ 1284 m->object = object; 1285 m->pindex = pindex; 1286 1287 /* 1288 * Now link into the object's ordered list of backed pages. 1289 */ 1290 if (vm_radix_insert(&object->rtree, m)) { 1291 m->object = NULL; 1292 m->pindex = 0; 1293 return (1); 1294 } 1295 vm_page_insert_radixdone(m, object, mpred); 1296 return (0); 1297} 1298 1299/* 1300 * vm_page_insert_radixdone: 1301 * 1302 * Complete page "m" insertion into the specified object after the 1303 * radix trie hooking. 1304 * 1305 * The page "mpred" must precede the offset "m->pindex" within the 1306 * specified object. 1307 * 1308 * The object must be locked. 1309 */ 1310static void 1311vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred) 1312{ 1313 1314 VM_OBJECT_ASSERT_WLOCKED(object); 1315 KASSERT(object != NULL && m->object == object, 1316 ("vm_page_insert_radixdone: page %p has inconsistent object", m)); 1317 if (mpred != NULL) { 1318 KASSERT(mpred->object == object, 1319 ("vm_page_insert_after: object doesn't contain mpred")); 1320 KASSERT(mpred->pindex < m->pindex, 1321 ("vm_page_insert_after: mpred doesn't precede pindex")); 1322 } 1323 1324 if (mpred != NULL) 1325 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq); 1326 else 1327 TAILQ_INSERT_HEAD(&object->memq, m, listq); 1328 1329 /* 1330 * Show that the object has one more resident page. 1331 */ 1332 object->resident_page_count++; 1333 1334 /* 1335 * Hold the vnode until the last page is released. 1336 */ 1337 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 1338 vhold(object->handle); 1339 1340 /* 1341 * Since we are inserting a new and possibly dirty page, 1342 * update the object's OBJ_MIGHTBEDIRTY flag. 1343 */ 1344 if (pmap_page_is_write_mapped(m)) 1345 vm_object_set_writeable_dirty(object); 1346} 1347 1348/* 1349 * vm_page_remove: 1350 * 1351 * Removes the specified page from its containing object, but does not 1352 * invalidate any backing storage. 1353 * 1354 * The object must be locked. The page must be locked if it is managed. 1355 */ 1356void 1357vm_page_remove(vm_page_t m) 1358{ 1359 vm_object_t object; 1360 vm_page_t mrem; 1361 1362 if ((m->oflags & VPO_UNMANAGED) == 0) 1363 vm_page_assert_locked(m); 1364 if ((object = m->object) == NULL) 1365 return; 1366 VM_OBJECT_ASSERT_WLOCKED(object); 1367 if (vm_page_xbusied(m)) 1368 vm_page_xunbusy_maybelocked(m); 1369 mrem = vm_radix_remove(&object->rtree, m->pindex); 1370 KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m)); 1371 1372 /* 1373 * Now remove from the object's list of backed pages. 1374 */ 1375 TAILQ_REMOVE(&object->memq, m, listq); 1376 1377 /* 1378 * And show that the object has one fewer resident page. 1379 */ 1380 object->resident_page_count--; 1381 1382 /* 1383 * The vnode may now be recycled. 1384 */ 1385 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 1386 vdrop(object->handle); 1387 1388 m->object = NULL; 1389} 1390 1391/* 1392 * vm_page_lookup: 1393 * 1394 * Returns the page associated with the object/offset 1395 * pair specified; if none is found, NULL is returned. 1396 * 1397 * The object must be locked. 1398 */ 1399vm_page_t 1400vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1401{ 1402 1403 VM_OBJECT_ASSERT_LOCKED(object); 1404 return (vm_radix_lookup(&object->rtree, pindex)); 1405} 1406 1407/* 1408 * vm_page_find_least: 1409 * 1410 * Returns the page associated with the object with least pindex 1411 * greater than or equal to the parameter pindex, or NULL. 1412 * 1413 * The object must be locked. 1414 */ 1415vm_page_t 1416vm_page_find_least(vm_object_t object, vm_pindex_t pindex) 1417{ 1418 vm_page_t m; 1419 1420 VM_OBJECT_ASSERT_LOCKED(object); 1421 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex) 1422 m = vm_radix_lookup_ge(&object->rtree, pindex); 1423 return (m); 1424} 1425 1426/* 1427 * Returns the given page's successor (by pindex) within the object if it is 1428 * resident; if none is found, NULL is returned. 1429 * 1430 * The object must be locked. 1431 */ 1432vm_page_t 1433vm_page_next(vm_page_t m) 1434{ 1435 vm_page_t next; 1436 1437 VM_OBJECT_ASSERT_LOCKED(m->object); 1438 if ((next = TAILQ_NEXT(m, listq)) != NULL) { 1439 MPASS(next->object == m->object); 1440 if (next->pindex != m->pindex + 1) 1441 next = NULL; 1442 } 1443 return (next); 1444} 1445 1446/* 1447 * Returns the given page's predecessor (by pindex) within the object if it is 1448 * resident; if none is found, NULL is returned. 1449 * 1450 * The object must be locked. 1451 */ 1452vm_page_t 1453vm_page_prev(vm_page_t m) 1454{ 1455 vm_page_t prev; 1456 1457 VM_OBJECT_ASSERT_LOCKED(m->object); 1458 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) { 1459 MPASS(prev->object == m->object); 1460 if (prev->pindex != m->pindex - 1) 1461 prev = NULL; 1462 } 1463 return (prev); 1464} 1465 1466/* 1467 * Uses the page mnew as a replacement for an existing page at index 1468 * pindex which must be already present in the object. 1469 * 1470 * The existing page must not be on a paging queue. 1471 */ 1472vm_page_t 1473vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex) 1474{ 1475 vm_page_t mold; 1476 1477 VM_OBJECT_ASSERT_WLOCKED(object); 1478 KASSERT(mnew->object == NULL, 1479 ("vm_page_replace: page %p already in object", mnew)); 1480 KASSERT(mnew->queue == PQ_NONE, 1481 ("vm_page_replace: new page %p is on a paging queue", mnew)); 1482 1483 /* 1484 * This function mostly follows vm_page_insert() and 1485 * vm_page_remove() without the radix, object count and vnode 1486 * dance. Double check such functions for more comments. 1487 */ 1488 1489 mnew->object = object; 1490 mnew->pindex = pindex; 1491 mold = vm_radix_replace(&object->rtree, mnew); 1492 KASSERT(mold->queue == PQ_NONE, 1493 ("vm_page_replace: old page %p is on a paging queue", mold)); 1494 1495 /* Keep the resident page list in sorted order. */ 1496 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq); 1497 TAILQ_REMOVE(&object->memq, mold, listq); 1498 1499 mold->object = NULL; 1500 vm_page_xunbusy_maybelocked(mold); 1501 1502 /* 1503 * The object's resident_page_count does not change because we have 1504 * swapped one page for another, but OBJ_MIGHTBEDIRTY. 1505 */ 1506 if (pmap_page_is_write_mapped(mnew)) 1507 vm_object_set_writeable_dirty(object); 1508 return (mold); 1509} 1510 1511/* 1512 * vm_page_rename: 1513 * 1514 * Move the given memory entry from its 1515 * current object to the specified target object/offset. 1516 * 1517 * Note: swap associated with the page must be invalidated by the move. We 1518 * have to do this for several reasons: (1) we aren't freeing the 1519 * page, (2) we are dirtying the page, (3) the VM system is probably 1520 * moving the page from object A to B, and will then later move 1521 * the backing store from A to B and we can't have a conflict. 1522 * 1523 * Note: we *always* dirty the page. It is necessary both for the 1524 * fact that we moved it, and because we may be invalidating 1525 * swap. 1526 * 1527 * The objects must be locked. 1528 */ 1529int 1530vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1531{ 1532 vm_page_t mpred; 1533 vm_pindex_t opidx; 1534 1535 VM_OBJECT_ASSERT_WLOCKED(new_object); 1536 1537 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex); 1538 KASSERT(mpred == NULL || mpred->pindex != new_pindex, 1539 ("vm_page_rename: pindex already renamed")); 1540 1541 /* 1542 * Create a custom version of vm_page_insert() which does not depend 1543 * by m_prev and can cheat on the implementation aspects of the 1544 * function. 1545 */ 1546 opidx = m->pindex; 1547 m->pindex = new_pindex; 1548 if (vm_radix_insert(&new_object->rtree, m)) { 1549 m->pindex = opidx; 1550 return (1); 1551 } 1552 1553 /* 1554 * The operation cannot fail anymore. The removal must happen before 1555 * the listq iterator is tainted. 1556 */ 1557 m->pindex = opidx; 1558 vm_page_lock(m); 1559 vm_page_remove(m); 1560 1561 /* Return back to the new pindex to complete vm_page_insert(). */ 1562 m->pindex = new_pindex; 1563 m->object = new_object; 1564 vm_page_unlock(m); 1565 vm_page_insert_radixdone(m, new_object, mpred); 1566 vm_page_dirty(m); 1567 return (0); 1568} 1569 1570/* 1571 * vm_page_alloc: 1572 * 1573 * Allocate and return a page that is associated with the specified 1574 * object and offset pair. By default, this page is exclusive busied. 1575 * 1576 * The caller must always specify an allocation class. 1577 * 1578 * allocation classes: 1579 * VM_ALLOC_NORMAL normal process request 1580 * VM_ALLOC_SYSTEM system *really* needs a page 1581 * VM_ALLOC_INTERRUPT interrupt time request 1582 * 1583 * optional allocation flags: 1584 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1585 * intends to allocate 1586 * VM_ALLOC_NOBUSY do not exclusive busy the page 1587 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1588 * VM_ALLOC_NOOBJ page is not associated with an object and 1589 * should not be exclusive busy 1590 * VM_ALLOC_SBUSY shared busy the allocated page 1591 * VM_ALLOC_WIRED wire the allocated page 1592 * VM_ALLOC_ZERO prefer a zeroed page 1593 * 1594 * This routine may not sleep. 1595 */ 1596vm_page_t 1597vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 1598{ 1599 1600 return (vm_page_alloc_after(object, pindex, req, object != NULL ? 1601 vm_radix_lookup_le(&object->rtree, pindex) : NULL)); 1602} 1603 1604/* 1605 * Allocate a page in the specified object with the given page index. To 1606 * optimize insertion of the page into the object, the caller must also specifiy 1607 * the resident page in the object with largest index smaller than the given 1608 * page index, or NULL if no such page exists. 1609 */ 1610vm_page_t 1611vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex, int req, 1612 vm_page_t mpred) 1613{ 1614 vm_page_t m; 1615 int flags, req_class; 1616 u_int free_count; 1617 1618 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 1619 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 1620 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 1621 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 1622 ("inconsistent object(%p)/req(%x)", object, req)); 1623 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0, 1624 ("Can't sleep and retry object insertion.")); 1625 KASSERT(mpred == NULL || mpred->pindex < pindex, 1626 ("mpred %p doesn't precede pindex 0x%jx", mpred, 1627 (uintmax_t)pindex)); 1628 if (object != NULL) 1629 VM_OBJECT_ASSERT_WLOCKED(object); 1630 1631 if (__predict_false((req & VM_ALLOC_IFCACHED) != 0)) 1632 return (NULL); 1633 1634 req_class = req & VM_ALLOC_CLASS_MASK; 1635 1636 /* 1637 * The page daemon is allowed to dig deeper into the free page list. 1638 */ 1639 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1640 req_class = VM_ALLOC_SYSTEM; 1641 1642 /* 1643 * Allocate a page if the number of free pages exceeds the minimum 1644 * for the request class. 1645 */ 1646again: 1647 mtx_lock(&vm_page_queue_free_mtx); 1648 if (vm_cnt.v_free_count > vm_cnt.v_free_reserved || 1649 (req_class == VM_ALLOC_SYSTEM && 1650 vm_cnt.v_free_count > vm_cnt.v_interrupt_free_min) || 1651 (req_class == VM_ALLOC_INTERRUPT && 1652 vm_cnt.v_free_count > 0)) { 1653 /* 1654 * Can we allocate the page from a reservation? 1655 */ 1656#if VM_NRESERVLEVEL > 0 1657 if (object == NULL || (object->flags & (OBJ_COLORED | 1658 OBJ_FICTITIOUS)) != OBJ_COLORED || (m = 1659 vm_reserv_alloc_page(object, pindex, mpred)) == NULL) 1660#endif 1661 { 1662 /* 1663 * If not, allocate it from the free page queues. 1664 */ 1665 m = vm_phys_alloc_pages(object != NULL ? 1666 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0); 1667#if VM_NRESERVLEVEL > 0 1668 if (m == NULL && vm_reserv_reclaim_inactive()) { 1669 m = vm_phys_alloc_pages(object != NULL ? 1670 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 1671 0); 1672 } 1673#endif 1674 } 1675 } else { 1676 /* 1677 * Not allocatable, give up. 1678 */ 1679 if (vm_page_alloc_fail(object, req)) 1680 goto again; 1681 return (NULL); 1682 } 1683 1684 /* 1685 * At this point we had better have found a good page. 1686 */ 1687 KASSERT(m != NULL, ("missing page")); 1688 free_count = vm_phys_freecnt_adj(m, -1); 1689 if ((m->flags & PG_ZERO) != 0) 1690 vm_page_zero_count--; 1691 mtx_unlock(&vm_page_queue_free_mtx); 1692 vm_page_alloc_check(m); 1693 1694 /* 1695 * Initialize the page. Only the PG_ZERO flag is inherited. 1696 */ 1697 flags = 0; 1698 if ((req & VM_ALLOC_ZERO) != 0) 1699 flags = PG_ZERO; 1700 flags &= m->flags; 1701 if ((req & VM_ALLOC_NODUMP) != 0) 1702 flags |= PG_NODUMP; 1703 m->flags = flags; 1704 m->aflags = 0; 1705 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? 1706 VPO_UNMANAGED : 0; 1707 m->busy_lock = VPB_UNBUSIED; 1708 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) 1709 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 1710 if ((req & VM_ALLOC_SBUSY) != 0) 1711 m->busy_lock = VPB_SHARERS_WORD(1); 1712 if (req & VM_ALLOC_WIRED) { 1713 /* 1714 * The page lock is not required for wiring a page until that 1715 * page is inserted into the object. 1716 */ 1717 atomic_add_int(&vm_cnt.v_wire_count, 1); 1718 m->wire_count = 1; 1719 } 1720 m->act_count = 0; 1721 1722 if (object != NULL) { 1723 if (vm_page_insert_after(m, object, pindex, mpred)) { 1724 pagedaemon_wakeup(); 1725 if (req & VM_ALLOC_WIRED) { 1726 atomic_subtract_int(&vm_cnt.v_wire_count, 1); 1727 m->wire_count = 0; 1728 } 1729 KASSERT(m->object == NULL, ("page %p has object", m)); 1730 m->oflags = VPO_UNMANAGED; 1731 m->busy_lock = VPB_UNBUSIED; 1732 /* Don't change PG_ZERO. */ 1733 vm_page_free_toq(m); 1734 if (req & VM_ALLOC_WAITFAIL) { 1735 VM_OBJECT_WUNLOCK(object); 1736 vm_radix_wait(); 1737 VM_OBJECT_WLOCK(object); 1738 } 1739 return (NULL); 1740 } 1741 1742 /* Ignore device objects; the pager sets "memattr" for them. */ 1743 if (object->memattr != VM_MEMATTR_DEFAULT && 1744 (object->flags & OBJ_FICTITIOUS) == 0) 1745 pmap_page_set_memattr(m, object->memattr); 1746 } else 1747 m->pindex = pindex; 1748 1749 /* 1750 * Don't wakeup too often - wakeup the pageout daemon when 1751 * we would be nearly out of memory. 1752 */ 1753 if (vm_paging_needed(free_count)) 1754 pagedaemon_wakeup(); 1755 1756 return (m); 1757} 1758 1759/* 1760 * vm_page_alloc_contig: 1761 * 1762 * Allocate a contiguous set of physical pages of the given size "npages" 1763 * from the free lists. All of the physical pages must be at or above 1764 * the given physical address "low" and below the given physical address 1765 * "high". The given value "alignment" determines the alignment of the 1766 * first physical page in the set. If the given value "boundary" is 1767 * non-zero, then the set of physical pages cannot cross any physical 1768 * address boundary that is a multiple of that value. Both "alignment" 1769 * and "boundary" must be a power of two. 1770 * 1771 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, 1772 * then the memory attribute setting for the physical pages is configured 1773 * to the object's memory attribute setting. Otherwise, the memory 1774 * attribute setting for the physical pages is configured to "memattr", 1775 * overriding the object's memory attribute setting. However, if the 1776 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the 1777 * memory attribute setting for the physical pages cannot be configured 1778 * to VM_MEMATTR_DEFAULT. 1779 * 1780 * The specified object may not contain fictitious pages. 1781 * 1782 * The caller must always specify an allocation class. 1783 * 1784 * allocation classes: 1785 * VM_ALLOC_NORMAL normal process request 1786 * VM_ALLOC_SYSTEM system *really* needs a page 1787 * VM_ALLOC_INTERRUPT interrupt time request 1788 * 1789 * optional allocation flags: 1790 * VM_ALLOC_NOBUSY do not exclusive busy the page 1791 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1792 * VM_ALLOC_NOOBJ page is not associated with an object and 1793 * should not be exclusive busy 1794 * VM_ALLOC_SBUSY shared busy the allocated page 1795 * VM_ALLOC_WIRED wire the allocated page 1796 * VM_ALLOC_ZERO prefer a zeroed page 1797 * 1798 * This routine may not sleep. 1799 */ 1800vm_page_t 1801vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, 1802 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 1803 vm_paddr_t boundary, vm_memattr_t memattr) 1804{ 1805 vm_page_t m, m_ret, mpred; 1806 u_int busy_lock, flags, oflags; 1807 int req_class; 1808 1809 mpred = NULL; /* XXX: pacify gcc */ 1810 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 1811 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 1812 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 1813 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 1814 ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object, 1815 req)); 1816 KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0, 1817 ("Can't sleep and retry object insertion.")); 1818 if (object != NULL) { 1819 VM_OBJECT_ASSERT_WLOCKED(object); 1820 KASSERT((object->flags & OBJ_FICTITIOUS) == 0, 1821 ("vm_page_alloc_contig: object %p has fictitious pages", 1822 object)); 1823 } 1824 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 1825 req_class = req & VM_ALLOC_CLASS_MASK; 1826 1827 /* 1828 * The page daemon is allowed to dig deeper into the free page list. 1829 */ 1830 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1831 req_class = VM_ALLOC_SYSTEM; 1832 1833 if (object != NULL) { 1834 mpred = vm_radix_lookup_le(&object->rtree, pindex); 1835 KASSERT(mpred == NULL || mpred->pindex != pindex, 1836 ("vm_page_alloc_contig: pindex already allocated")); 1837 } 1838 1839 /* 1840 * Can we allocate the pages without the number of free pages falling 1841 * below the lower bound for the allocation class? 1842 */ 1843again: 1844 mtx_lock(&vm_page_queue_free_mtx); 1845 if (vm_cnt.v_free_count >= npages + vm_cnt.v_free_reserved || 1846 (req_class == VM_ALLOC_SYSTEM && 1847 vm_cnt.v_free_count >= npages + vm_cnt.v_interrupt_free_min) || 1848 (req_class == VM_ALLOC_INTERRUPT && 1849 vm_cnt.v_free_count >= npages)) { 1850 /* 1851 * Can we allocate the pages from a reservation? 1852 */ 1853#if VM_NRESERVLEVEL > 0 1854retry: 1855 if (object == NULL || (object->flags & OBJ_COLORED) == 0 || 1856 (m_ret = vm_reserv_alloc_contig(object, pindex, npages, 1857 low, high, alignment, boundary, mpred)) == NULL) 1858#endif 1859 /* 1860 * If not, allocate them from the free page queues. 1861 */ 1862 m_ret = vm_phys_alloc_contig(npages, low, high, 1863 alignment, boundary); 1864 } else { 1865 if (vm_page_alloc_fail(object, req)) 1866 goto again; 1867 return (NULL); 1868 } 1869 if (m_ret != NULL) { 1870 vm_phys_freecnt_adj(m_ret, -npages); 1871 for (m = m_ret; m < &m_ret[npages]; m++) 1872 if ((m->flags & PG_ZERO) != 0) 1873 vm_page_zero_count--; 1874 } else { 1875#if VM_NRESERVLEVEL > 0 1876 if (vm_reserv_reclaim_contig(npages, low, high, alignment, 1877 boundary)) 1878 goto retry; 1879#endif 1880 } 1881 mtx_unlock(&vm_page_queue_free_mtx); 1882 if (m_ret == NULL) 1883 return (NULL); 1884 for (m = m_ret; m < &m_ret[npages]; m++) 1885 vm_page_alloc_check(m); 1886 1887 /* 1888 * Initialize the pages. Only the PG_ZERO flag is inherited. 1889 */ 1890 flags = 0; 1891 if ((req & VM_ALLOC_ZERO) != 0) 1892 flags = PG_ZERO; 1893 if ((req & VM_ALLOC_NODUMP) != 0) 1894 flags |= PG_NODUMP; 1895 oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? 1896 VPO_UNMANAGED : 0; 1897 busy_lock = VPB_UNBUSIED; 1898 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) 1899 busy_lock = VPB_SINGLE_EXCLUSIVER; 1900 if ((req & VM_ALLOC_SBUSY) != 0) 1901 busy_lock = VPB_SHARERS_WORD(1); 1902 if ((req & VM_ALLOC_WIRED) != 0) 1903 atomic_add_int(&vm_cnt.v_wire_count, npages); 1904 if (object != NULL) { 1905 if (object->memattr != VM_MEMATTR_DEFAULT && 1906 memattr == VM_MEMATTR_DEFAULT) 1907 memattr = object->memattr; 1908 } 1909 for (m = m_ret; m < &m_ret[npages]; m++) { 1910 m->aflags = 0; 1911 m->flags = (m->flags | PG_NODUMP) & flags; 1912 m->busy_lock = busy_lock; 1913 if ((req & VM_ALLOC_WIRED) != 0) 1914 m->wire_count = 1; 1915 m->act_count = 0; 1916 m->oflags = oflags; 1917 if (object != NULL) { 1918 if (vm_page_insert_after(m, object, pindex, mpred)) { 1919 pagedaemon_wakeup(); 1920 if ((req & VM_ALLOC_WIRED) != 0) 1921 atomic_subtract_int( 1922 &vm_cnt.v_wire_count, npages); 1923 KASSERT(m->object == NULL, 1924 ("page %p has object", m)); 1925 mpred = m; 1926 for (m = m_ret; m < &m_ret[npages]; m++) { 1927 if (m <= mpred && 1928 (req & VM_ALLOC_WIRED) != 0) 1929 m->wire_count = 0; 1930 m->oflags = VPO_UNMANAGED; 1931 m->busy_lock = VPB_UNBUSIED; 1932 /* Don't change PG_ZERO. */ 1933 vm_page_free_toq(m); 1934 } 1935 if (req & VM_ALLOC_WAITFAIL) { 1936 VM_OBJECT_WUNLOCK(object); 1937 vm_radix_wait(); 1938 VM_OBJECT_WLOCK(object); 1939 } 1940 return (NULL); 1941 } 1942 mpred = m; 1943 } else 1944 m->pindex = pindex; 1945 if (memattr != VM_MEMATTR_DEFAULT) 1946 pmap_page_set_memattr(m, memattr); 1947 pindex++; 1948 } 1949 if (vm_paging_needed(vm_cnt.v_free_count)) 1950 pagedaemon_wakeup(); 1951 return (m_ret); 1952} 1953 1954/* 1955 * Check a page that has been freshly dequeued from a freelist. 1956 */ 1957static void 1958vm_page_alloc_check(vm_page_t m) 1959{ 1960 1961 KASSERT(m->object == NULL, ("page %p has object", m)); 1962 KASSERT(m->queue == PQ_NONE, 1963 ("page %p has unexpected queue %d", m, m->queue)); 1964 KASSERT(m->wire_count == 0, ("page %p is wired", m)); 1965 KASSERT(m->hold_count == 0, ("page %p is held", m)); 1966 KASSERT(!vm_page_busied(m), ("page %p is busy", m)); 1967 KASSERT(m->dirty == 0, ("page %p is dirty", m)); 1968 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 1969 ("page %p has unexpected memattr %d", 1970 m, pmap_page_get_memattr(m))); 1971 KASSERT(m->valid == 0, ("free page %p is valid", m)); 1972} 1973 1974/* 1975 * vm_page_alloc_freelist: 1976 * 1977 * Allocate a physical page from the specified free page list. 1978 * 1979 * The caller must always specify an allocation class. 1980 * 1981 * allocation classes: 1982 * VM_ALLOC_NORMAL normal process request 1983 * VM_ALLOC_SYSTEM system *really* needs a page 1984 * VM_ALLOC_INTERRUPT interrupt time request 1985 * 1986 * optional allocation flags: 1987 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1988 * intends to allocate 1989 * VM_ALLOC_WIRED wire the allocated page 1990 * VM_ALLOC_ZERO prefer a zeroed page 1991 * 1992 * This routine may not sleep. 1993 */ 1994vm_page_t 1995vm_page_alloc_freelist(int flind, int req) 1996{ 1997 vm_page_t m; 1998 u_int flags, free_count; 1999 int req_class; 2000 2001 req_class = req & VM_ALLOC_CLASS_MASK; 2002 2003 /* 2004 * The page daemon is allowed to dig deeper into the free page list. 2005 */ 2006 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 2007 req_class = VM_ALLOC_SYSTEM; 2008 2009 /* 2010 * Do not allocate reserved pages unless the req has asked for it. 2011 */ 2012again: 2013 mtx_lock(&vm_page_queue_free_mtx); 2014 if (vm_cnt.v_free_count > vm_cnt.v_free_reserved || 2015 (req_class == VM_ALLOC_SYSTEM && 2016 vm_cnt.v_free_count > vm_cnt.v_interrupt_free_min) || 2017 (req_class == VM_ALLOC_INTERRUPT && 2018 vm_cnt.v_free_count > 0)) { 2019 m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0); 2020 } else { 2021 if (vm_page_alloc_fail(NULL, req)) 2022 goto again; 2023 return (NULL); 2024 } 2025 if (m == NULL) { 2026 mtx_unlock(&vm_page_queue_free_mtx); 2027 return (NULL); 2028 } 2029 free_count = vm_phys_freecnt_adj(m, -1); 2030 if ((m->flags & PG_ZERO) != 0) 2031 vm_page_zero_count--; 2032 mtx_unlock(&vm_page_queue_free_mtx); 2033 vm_page_alloc_check(m); 2034 2035 /* 2036 * Initialize the page. Only the PG_ZERO flag is inherited. 2037 */ 2038 m->aflags = 0; 2039 flags = 0; 2040 if ((req & VM_ALLOC_ZERO) != 0) 2041 flags = PG_ZERO; 2042 m->flags &= flags; 2043 if ((req & VM_ALLOC_WIRED) != 0) { 2044 /* 2045 * The page lock is not required for wiring a page that does 2046 * not belong to an object. 2047 */ 2048 atomic_add_int(&vm_cnt.v_wire_count, 1); 2049 m->wire_count = 1; 2050 } 2051 /* Unmanaged pages don't use "act_count". */ 2052 m->oflags = VPO_UNMANAGED; 2053 if (vm_paging_needed(free_count)) 2054 pagedaemon_wakeup(); 2055 return (m); 2056} 2057 2058#define VPSC_ANY 0 /* No restrictions. */ 2059#define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */ 2060#define VPSC_NOSUPER 2 /* Skip superpages. */ 2061 2062/* 2063 * vm_page_scan_contig: 2064 * 2065 * Scan vm_page_array[] between the specified entries "m_start" and 2066 * "m_end" for a run of contiguous physical pages that satisfy the 2067 * specified conditions, and return the lowest page in the run. The 2068 * specified "alignment" determines the alignment of the lowest physical 2069 * page in the run. If the specified "boundary" is non-zero, then the 2070 * run of physical pages cannot span a physical address that is a 2071 * multiple of "boundary". 2072 * 2073 * "m_end" is never dereferenced, so it need not point to a vm_page 2074 * structure within vm_page_array[]. 2075 * 2076 * "npages" must be greater than zero. "m_start" and "m_end" must not 2077 * span a hole (or discontiguity) in the physical address space. Both 2078 * "alignment" and "boundary" must be a power of two. 2079 */ 2080vm_page_t 2081vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end, 2082 u_long alignment, vm_paddr_t boundary, int options) 2083{ 2084 struct mtx *m_mtx; 2085 vm_object_t object; 2086 vm_paddr_t pa; 2087 vm_page_t m, m_run; 2088#if VM_NRESERVLEVEL > 0 2089 int level; 2090#endif 2091 int m_inc, order, run_ext, run_len; 2092 2093 KASSERT(npages > 0, ("npages is 0")); 2094 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2095 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2096 m_run = NULL; 2097 run_len = 0; 2098 m_mtx = NULL; 2099 for (m = m_start; m < m_end && run_len < npages; m += m_inc) { 2100 KASSERT((m->flags & PG_MARKER) == 0, 2101 ("page %p is PG_MARKER", m)); 2102 KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->wire_count == 1, 2103 ("fictitious page %p has invalid wire count", m)); 2104 2105 /* 2106 * If the current page would be the start of a run, check its 2107 * physical address against the end, alignment, and boundary 2108 * conditions. If it doesn't satisfy these conditions, either 2109 * terminate the scan or advance to the next page that 2110 * satisfies the failed condition. 2111 */ 2112 if (run_len == 0) { 2113 KASSERT(m_run == NULL, ("m_run != NULL")); 2114 if (m + npages > m_end) 2115 break; 2116 pa = VM_PAGE_TO_PHYS(m); 2117 if ((pa & (alignment - 1)) != 0) { 2118 m_inc = atop(roundup2(pa, alignment) - pa); 2119 continue; 2120 } 2121 if (rounddown2(pa ^ (pa + ptoa(npages) - 1), 2122 boundary) != 0) { 2123 m_inc = atop(roundup2(pa, boundary) - pa); 2124 continue; 2125 } 2126 } else 2127 KASSERT(m_run != NULL, ("m_run == NULL")); 2128 2129 vm_page_change_lock(m, &m_mtx); 2130 m_inc = 1; 2131retry: 2132 if (m->wire_count != 0 || m->hold_count != 0) 2133 run_ext = 0; 2134#if VM_NRESERVLEVEL > 0 2135 else if ((level = vm_reserv_level(m)) >= 0 && 2136 (options & VPSC_NORESERV) != 0) { 2137 run_ext = 0; 2138 /* Advance to the end of the reservation. */ 2139 pa = VM_PAGE_TO_PHYS(m); 2140 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - 2141 pa); 2142 } 2143#endif 2144 else if ((object = m->object) != NULL) { 2145 /* 2146 * The page is considered eligible for relocation if 2147 * and only if it could be laundered or reclaimed by 2148 * the page daemon. 2149 */ 2150 if (!VM_OBJECT_TRYRLOCK(object)) { 2151 mtx_unlock(m_mtx); 2152 VM_OBJECT_RLOCK(object); 2153 mtx_lock(m_mtx); 2154 if (m->object != object) { 2155 /* 2156 * The page may have been freed. 2157 */ 2158 VM_OBJECT_RUNLOCK(object); 2159 goto retry; 2160 } else if (m->wire_count != 0 || 2161 m->hold_count != 0) { 2162 run_ext = 0; 2163 goto unlock; 2164 } 2165 } 2166 KASSERT((m->flags & PG_UNHOLDFREE) == 0, 2167 ("page %p is PG_UNHOLDFREE", m)); 2168 /* Don't care: PG_NODUMP, PG_ZERO. */ 2169 if (object->type != OBJT_DEFAULT && 2170 object->type != OBJT_SWAP && 2171 object->type != OBJT_VNODE) { 2172 run_ext = 0; 2173#if VM_NRESERVLEVEL > 0 2174 } else if ((options & VPSC_NOSUPER) != 0 && 2175 (level = vm_reserv_level_iffullpop(m)) >= 0) { 2176 run_ext = 0; 2177 /* Advance to the end of the superpage. */ 2178 pa = VM_PAGE_TO_PHYS(m); 2179 m_inc = atop(roundup2(pa + 1, 2180 vm_reserv_size(level)) - pa); 2181#endif 2182 } else if (object->memattr == VM_MEMATTR_DEFAULT && 2183 m->queue != PQ_NONE && !vm_page_busied(m)) { 2184 /* 2185 * The page is allocated but eligible for 2186 * relocation. Extend the current run by one 2187 * page. 2188 */ 2189 KASSERT(pmap_page_get_memattr(m) == 2190 VM_MEMATTR_DEFAULT, 2191 ("page %p has an unexpected memattr", m)); 2192 KASSERT((m->oflags & (VPO_SWAPINPROG | 2193 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, 2194 ("page %p has unexpected oflags", m)); 2195 /* Don't care: VPO_NOSYNC. */ 2196 run_ext = 1; 2197 } else 2198 run_ext = 0; 2199unlock: 2200 VM_OBJECT_RUNLOCK(object); 2201#if VM_NRESERVLEVEL > 0 2202 } else if (level >= 0) { 2203 /* 2204 * The page is reserved but not yet allocated. In 2205 * other words, it is still free. Extend the current 2206 * run by one page. 2207 */ 2208 run_ext = 1; 2209#endif 2210 } else if ((order = m->order) < VM_NFREEORDER) { 2211 /* 2212 * The page is enqueued in the physical memory 2213 * allocator's free page queues. Moreover, it is the 2214 * first page in a power-of-two-sized run of 2215 * contiguous free pages. Add these pages to the end 2216 * of the current run, and jump ahead. 2217 */ 2218 run_ext = 1 << order; 2219 m_inc = 1 << order; 2220 } else { 2221 /* 2222 * Skip the page for one of the following reasons: (1) 2223 * It is enqueued in the physical memory allocator's 2224 * free page queues. However, it is not the first 2225 * page in a run of contiguous free pages. (This case 2226 * rarely occurs because the scan is performed in 2227 * ascending order.) (2) It is not reserved, and it is 2228 * transitioning from free to allocated. (Conversely, 2229 * the transition from allocated to free for managed 2230 * pages is blocked by the page lock.) (3) It is 2231 * allocated but not contained by an object and not 2232 * wired, e.g., allocated by Xen's balloon driver. 2233 */ 2234 run_ext = 0; 2235 } 2236 2237 /* 2238 * Extend or reset the current run of pages. 2239 */ 2240 if (run_ext > 0) { 2241 if (run_len == 0) 2242 m_run = m; 2243 run_len += run_ext; 2244 } else { 2245 if (run_len > 0) { 2246 m_run = NULL; 2247 run_len = 0; 2248 } 2249 } 2250 } 2251 if (m_mtx != NULL) 2252 mtx_unlock(m_mtx); 2253 if (run_len >= npages) 2254 return (m_run); 2255 return (NULL); 2256} 2257 2258/* 2259 * vm_page_reclaim_run: 2260 * 2261 * Try to relocate each of the allocated virtual pages within the 2262 * specified run of physical pages to a new physical address. Free the 2263 * physical pages underlying the relocated virtual pages. A virtual page 2264 * is relocatable if and only if it could be laundered or reclaimed by 2265 * the page daemon. Whenever possible, a virtual page is relocated to a 2266 * physical address above "high". 2267 * 2268 * Returns 0 if every physical page within the run was already free or 2269 * just freed by a successful relocation. Otherwise, returns a non-zero 2270 * value indicating why the last attempt to relocate a virtual page was 2271 * unsuccessful. 2272 * 2273 * "req_class" must be an allocation class. 2274 */ 2275static int 2276vm_page_reclaim_run(int req_class, u_long npages, vm_page_t m_run, 2277 vm_paddr_t high) 2278{ 2279 struct mtx *m_mtx; 2280 struct spglist free; 2281 vm_object_t object; 2282 vm_paddr_t pa; 2283 vm_page_t m, m_end, m_new; 2284 int error, order, req; 2285 2286 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class, 2287 ("req_class is not an allocation class")); 2288 SLIST_INIT(&free); 2289 error = 0; 2290 m = m_run; 2291 m_end = m_run + npages; 2292 m_mtx = NULL; 2293 for (; error == 0 && m < m_end; m++) { 2294 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0, 2295 ("page %p is PG_FICTITIOUS or PG_MARKER", m)); 2296 2297 /* 2298 * Avoid releasing and reacquiring the same page lock. 2299 */ 2300 vm_page_change_lock(m, &m_mtx); 2301retry: 2302 if (m->wire_count != 0 || m->hold_count != 0) 2303 error = EBUSY; 2304 else if ((object = m->object) != NULL) { 2305 /* 2306 * The page is relocated if and only if it could be 2307 * laundered or reclaimed by the page daemon. 2308 */ 2309 if (!VM_OBJECT_TRYWLOCK(object)) { 2310 mtx_unlock(m_mtx); 2311 VM_OBJECT_WLOCK(object); 2312 mtx_lock(m_mtx); 2313 if (m->object != object) { 2314 /* 2315 * The page may have been freed. 2316 */ 2317 VM_OBJECT_WUNLOCK(object); 2318 goto retry; 2319 } else if (m->wire_count != 0 || 2320 m->hold_count != 0) { 2321 error = EBUSY; 2322 goto unlock; 2323 } 2324 } 2325 KASSERT((m->flags & PG_UNHOLDFREE) == 0, 2326 ("page %p is PG_UNHOLDFREE", m)); 2327 /* Don't care: PG_NODUMP, PG_ZERO. */ 2328 if (object->type != OBJT_DEFAULT && 2329 object->type != OBJT_SWAP && 2330 object->type != OBJT_VNODE) 2331 error = EINVAL; 2332 else if (object->memattr != VM_MEMATTR_DEFAULT) 2333 error = EINVAL; 2334 else if (m->queue != PQ_NONE && !vm_page_busied(m)) { 2335 KASSERT(pmap_page_get_memattr(m) == 2336 VM_MEMATTR_DEFAULT, 2337 ("page %p has an unexpected memattr", m)); 2338 KASSERT((m->oflags & (VPO_SWAPINPROG | 2339 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, 2340 ("page %p has unexpected oflags", m)); 2341 /* Don't care: VPO_NOSYNC. */ 2342 if (m->valid != 0) { 2343 /* 2344 * First, try to allocate a new page 2345 * that is above "high". Failing 2346 * that, try to allocate a new page 2347 * that is below "m_run". Allocate 2348 * the new page between the end of 2349 * "m_run" and "high" only as a last 2350 * resort. 2351 */ 2352 req = req_class | VM_ALLOC_NOOBJ; 2353 if ((m->flags & PG_NODUMP) != 0) 2354 req |= VM_ALLOC_NODUMP; 2355 if (trunc_page(high) != 2356 ~(vm_paddr_t)PAGE_MASK) { 2357 m_new = vm_page_alloc_contig( 2358 NULL, 0, req, 1, 2359 round_page(high), 2360 ~(vm_paddr_t)0, 2361 PAGE_SIZE, 0, 2362 VM_MEMATTR_DEFAULT); 2363 } else 2364 m_new = NULL; 2365 if (m_new == NULL) { 2366 pa = VM_PAGE_TO_PHYS(m_run); 2367 m_new = vm_page_alloc_contig( 2368 NULL, 0, req, 1, 2369 0, pa - 1, PAGE_SIZE, 0, 2370 VM_MEMATTR_DEFAULT); 2371 } 2372 if (m_new == NULL) { 2373 pa += ptoa(npages); 2374 m_new = vm_page_alloc_contig( 2375 NULL, 0, req, 1, 2376 pa, high, PAGE_SIZE, 0, 2377 VM_MEMATTR_DEFAULT); 2378 } 2379 if (m_new == NULL) { 2380 error = ENOMEM; 2381 goto unlock; 2382 } 2383 KASSERT(m_new->wire_count == 0, 2384 ("page %p is wired", m_new)); 2385 2386 /* 2387 * Replace "m" with the new page. For 2388 * vm_page_replace(), "m" must be busy 2389 * and dequeued. Finally, change "m" 2390 * as if vm_page_free() was called. 2391 */ 2392 if (object->ref_count != 0) 2393 pmap_remove_all(m); 2394 m_new->aflags = m->aflags; 2395 KASSERT(m_new->oflags == VPO_UNMANAGED, 2396 ("page %p is managed", m_new)); 2397 m_new->oflags = m->oflags & VPO_NOSYNC; 2398 pmap_copy_page(m, m_new); 2399 m_new->valid = m->valid; 2400 m_new->dirty = m->dirty; 2401 m->flags &= ~PG_ZERO; 2402 vm_page_xbusy(m); 2403 vm_page_remque(m); 2404 vm_page_replace_checked(m_new, object, 2405 m->pindex, m); 2406 m->valid = 0; 2407 vm_page_undirty(m); 2408 2409 /* 2410 * The new page must be deactivated 2411 * before the object is unlocked. 2412 */ 2413 vm_page_change_lock(m_new, &m_mtx); 2414 vm_page_deactivate(m_new); 2415 } else { 2416 m->flags &= ~PG_ZERO; 2417 vm_page_remque(m); 2418 vm_page_remove(m); 2419 KASSERT(m->dirty == 0, 2420 ("page %p is dirty", m)); 2421 } 2422 SLIST_INSERT_HEAD(&free, m, plinks.s.ss); 2423 } else 2424 error = EBUSY; 2425unlock: 2426 VM_OBJECT_WUNLOCK(object); 2427 } else { 2428 mtx_lock(&vm_page_queue_free_mtx); 2429 order = m->order; 2430 if (order < VM_NFREEORDER) { 2431 /* 2432 * The page is enqueued in the physical memory 2433 * allocator's free page queues. Moreover, it 2434 * is the first page in a power-of-two-sized 2435 * run of contiguous free pages. Jump ahead 2436 * to the last page within that run, and 2437 * continue from there. 2438 */ 2439 m += (1 << order) - 1; 2440 } 2441#if VM_NRESERVLEVEL > 0 2442 else if (vm_reserv_is_page_free(m)) 2443 order = 0; 2444#endif 2445 mtx_unlock(&vm_page_queue_free_mtx); 2446 if (order == VM_NFREEORDER) 2447 error = EINVAL; 2448 } 2449 } 2450 if (m_mtx != NULL) 2451 mtx_unlock(m_mtx); 2452 if ((m = SLIST_FIRST(&free)) != NULL) { 2453 mtx_lock(&vm_page_queue_free_mtx); 2454 do { 2455 SLIST_REMOVE_HEAD(&free, plinks.s.ss); 2456 vm_page_free_phys(m); 2457 } while ((m = SLIST_FIRST(&free)) != NULL); 2458 vm_page_zero_idle_wakeup(); 2459 vm_page_free_wakeup(); 2460 mtx_unlock(&vm_page_queue_free_mtx); 2461 } 2462 return (error); 2463} 2464 2465#define NRUNS 16 2466 2467CTASSERT(powerof2(NRUNS)); 2468 2469#define RUN_INDEX(count) ((count) & (NRUNS - 1)) 2470 2471#define MIN_RECLAIM 8 2472 2473/* 2474 * vm_page_reclaim_contig: 2475 * 2476 * Reclaim allocated, contiguous physical memory satisfying the specified 2477 * conditions by relocating the virtual pages using that physical memory. 2478 * Returns true if reclamation is successful and false otherwise. Since 2479 * relocation requires the allocation of physical pages, reclamation may 2480 * fail due to a shortage of free pages. When reclamation fails, callers 2481 * are expected to perform VM_WAIT before retrying a failed allocation 2482 * operation, e.g., vm_page_alloc_contig(). 2483 * 2484 * The caller must always specify an allocation class through "req". 2485 * 2486 * allocation classes: 2487 * VM_ALLOC_NORMAL normal process request 2488 * VM_ALLOC_SYSTEM system *really* needs a page 2489 * VM_ALLOC_INTERRUPT interrupt time request 2490 * 2491 * The optional allocation flags are ignored. 2492 * 2493 * "npages" must be greater than zero. Both "alignment" and "boundary" 2494 * must be a power of two. 2495 */ 2496bool 2497vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, 2498 u_long alignment, vm_paddr_t boundary) 2499{ 2500 vm_paddr_t curr_low; 2501 vm_page_t m_run, m_runs[NRUNS]; 2502 u_long count, reclaimed; 2503 int error, i, options, req_class; 2504 2505 KASSERT(npages > 0, ("npages is 0")); 2506 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2507 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2508 req_class = req & VM_ALLOC_CLASS_MASK; 2509 2510 /* 2511 * The page daemon is allowed to dig deeper into the free page list. 2512 */ 2513 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 2514 req_class = VM_ALLOC_SYSTEM; 2515 2516 /* 2517 * Return if the number of free pages cannot satisfy the requested 2518 * allocation. 2519 */ 2520 count = vm_cnt.v_free_count; 2521 if (count < npages + vm_cnt.v_free_reserved || (count < npages + 2522 vm_cnt.v_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) || 2523 (count < npages && req_class == VM_ALLOC_INTERRUPT)) 2524 return (false); 2525 2526 /* 2527 * Scan up to three times, relaxing the restrictions ("options") on 2528 * the reclamation of reservations and superpages each time. 2529 */ 2530 for (options = VPSC_NORESERV;;) { 2531 /* 2532 * Find the highest runs that satisfy the given constraints 2533 * and restrictions, and record them in "m_runs". 2534 */ 2535 curr_low = low; 2536 count = 0; 2537 for (;;) { 2538 m_run = vm_phys_scan_contig(npages, curr_low, high, 2539 alignment, boundary, options); 2540 if (m_run == NULL) 2541 break; 2542 curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages); 2543 m_runs[RUN_INDEX(count)] = m_run; 2544 count++; 2545 } 2546 2547 /* 2548 * Reclaim the highest runs in LIFO (descending) order until 2549 * the number of reclaimed pages, "reclaimed", is at least 2550 * MIN_RECLAIM. Reset "reclaimed" each time because each 2551 * reclamation is idempotent, and runs will (likely) recur 2552 * from one scan to the next as restrictions are relaxed. 2553 */ 2554 reclaimed = 0; 2555 for (i = 0; count > 0 && i < NRUNS; i++) { 2556 count--; 2557 m_run = m_runs[RUN_INDEX(count)]; 2558 error = vm_page_reclaim_run(req_class, npages, m_run, 2559 high); 2560 if (error == 0) { 2561 reclaimed += npages; 2562 if (reclaimed >= MIN_RECLAIM) 2563 return (true); 2564 } 2565 } 2566 2567 /* 2568 * Either relax the restrictions on the next scan or return if 2569 * the last scan had no restrictions. 2570 */ 2571 if (options == VPSC_NORESERV) 2572 options = VPSC_NOSUPER; 2573 else if (options == VPSC_NOSUPER) 2574 options = VPSC_ANY; 2575 else if (options == VPSC_ANY) 2576 return (reclaimed != 0); 2577 } 2578} 2579 2580/* 2581 * vm_wait: (also see VM_WAIT macro) 2582 * 2583 * Sleep until free pages are available for allocation. 2584 * - Called in various places before memory allocations. 2585 */ 2586static void 2587_vm_wait(void) 2588{ 2589 2590 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 2591 if (curproc == pageproc) { 2592 vm_pageout_pages_needed = 1; 2593 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx, 2594 PDROP | PSWP, "VMWait", 0); 2595 } else { 2596 if (pageproc == NULL) 2597 panic("vm_wait in early boot"); 2598 pagedaemon_wait(PVM, "vmwait"); 2599 } 2600} 2601 2602void 2603vm_wait(void) 2604{ 2605 2606 mtx_lock(&vm_page_queue_free_mtx); 2607 _vm_wait(); 2608} 2609 2610/* 2611 * vm_page_alloc_fail: 2612 * 2613 * Called when a page allocation function fails. Informs the 2614 * pagedaemon and performs the requested wait. Requires the 2615 * page_queue_free and object lock on entry. Returns with the 2616 * object lock held and free lock released. Returns an error when 2617 * retry is necessary. 2618 * 2619 */ 2620static int 2621vm_page_alloc_fail(vm_object_t object, int req) 2622{ 2623 2624 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 2625 2626 atomic_add_int(&vm_pageout_deficit, 2627 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 2628 if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) { 2629 if (object != NULL) 2630 VM_OBJECT_WUNLOCK(object); 2631 _vm_wait(); 2632 if (object != NULL) 2633 VM_OBJECT_WLOCK(object); 2634 if (req & VM_ALLOC_WAITOK) 2635 return (EAGAIN); 2636 } else { 2637 mtx_unlock(&vm_page_queue_free_mtx); 2638 pagedaemon_wakeup(); 2639 } 2640 return (0); 2641} 2642 2643/* 2644 * vm_waitpfault: (also see VM_WAITPFAULT macro) 2645 * 2646 * Sleep until free pages are available for allocation. 2647 * - Called only in vm_fault so that processes page faulting 2648 * can be easily tracked. 2649 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 2650 * processes will be able to grab memory first. Do not change 2651 * this balance without careful testing first. 2652 */ 2653void 2654vm_waitpfault(void) 2655{ 2656 2657 mtx_lock(&vm_page_queue_free_mtx); 2658 pagedaemon_wait(PUSER, "pfault"); 2659} 2660 2661struct vm_pagequeue * 2662vm_page_pagequeue(vm_page_t m) 2663{ 2664 2665 if (vm_page_in_laundry(m)) 2666 return (&vm_dom[0].vmd_pagequeues[m->queue]); 2667 else 2668 return (&vm_phys_domain(m)->vmd_pagequeues[m->queue]); 2669} 2670 2671/* 2672 * vm_page_dequeue: 2673 * 2674 * Remove the given page from its current page queue. 2675 * 2676 * The page must be locked. 2677 */ 2678void 2679vm_page_dequeue(vm_page_t m) 2680{ 2681 struct vm_pagequeue *pq; 2682 2683 vm_page_assert_locked(m); 2684 KASSERT(m->queue < PQ_COUNT, ("vm_page_dequeue: page %p is not queued", 2685 m)); 2686 pq = vm_page_pagequeue(m); 2687 vm_pagequeue_lock(pq); 2688 m->queue = PQ_NONE; 2689 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2690 vm_pagequeue_cnt_dec(pq); 2691 vm_pagequeue_unlock(pq); 2692} 2693 2694/* 2695 * vm_page_dequeue_locked: 2696 * 2697 * Remove the given page from its current page queue. 2698 * 2699 * The page and page queue must be locked. 2700 */ 2701void 2702vm_page_dequeue_locked(vm_page_t m) 2703{ 2704 struct vm_pagequeue *pq; 2705 2706 vm_page_lock_assert(m, MA_OWNED); 2707 pq = vm_page_pagequeue(m); 2708 vm_pagequeue_assert_locked(pq); 2709 m->queue = PQ_NONE; 2710 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2711 vm_pagequeue_cnt_dec(pq); 2712} 2713 2714/* 2715 * vm_page_enqueue: 2716 * 2717 * Add the given page to the specified page queue. 2718 * 2719 * The page must be locked. 2720 */ 2721static void 2722vm_page_enqueue(uint8_t queue, vm_page_t m) 2723{ 2724 struct vm_pagequeue *pq; 2725 2726 vm_page_lock_assert(m, MA_OWNED); 2727 KASSERT(queue < PQ_COUNT, 2728 ("vm_page_enqueue: invalid queue %u request for page %p", 2729 queue, m)); 2730 if (queue == PQ_LAUNDRY) 2731 pq = &vm_dom[0].vmd_pagequeues[queue]; 2732 else 2733 pq = &vm_phys_domain(m)->vmd_pagequeues[queue]; 2734 vm_pagequeue_lock(pq); 2735 m->queue = queue; 2736 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2737 vm_pagequeue_cnt_inc(pq); 2738 vm_pagequeue_unlock(pq); 2739} 2740 2741/* 2742 * vm_page_requeue: 2743 * 2744 * Move the given page to the tail of its current page queue. 2745 * 2746 * The page must be locked. 2747 */ 2748void 2749vm_page_requeue(vm_page_t m) 2750{ 2751 struct vm_pagequeue *pq; 2752 2753 vm_page_lock_assert(m, MA_OWNED); 2754 KASSERT(m->queue != PQ_NONE, 2755 ("vm_page_requeue: page %p is not queued", m)); 2756 pq = vm_page_pagequeue(m); 2757 vm_pagequeue_lock(pq); 2758 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2759 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2760 vm_pagequeue_unlock(pq); 2761} 2762 2763/* 2764 * vm_page_requeue_locked: 2765 * 2766 * Move the given page to the tail of its current page queue. 2767 * 2768 * The page queue must be locked. 2769 */ 2770void 2771vm_page_requeue_locked(vm_page_t m) 2772{ 2773 struct vm_pagequeue *pq; 2774 2775 KASSERT(m->queue != PQ_NONE, 2776 ("vm_page_requeue_locked: page %p is not queued", m)); 2777 pq = vm_page_pagequeue(m); 2778 vm_pagequeue_assert_locked(pq); 2779 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2780 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2781} 2782 2783/* 2784 * vm_page_activate: 2785 * 2786 * Put the specified page on the active list (if appropriate). 2787 * Ensure that act_count is at least ACT_INIT but do not otherwise 2788 * mess with it. 2789 * 2790 * The page must be locked. 2791 */ 2792void 2793vm_page_activate(vm_page_t m) 2794{ 2795 int queue; 2796 2797 vm_page_lock_assert(m, MA_OWNED); 2798 if ((queue = m->queue) != PQ_ACTIVE) { 2799 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 2800 if (m->act_count < ACT_INIT) 2801 m->act_count = ACT_INIT; 2802 if (queue != PQ_NONE) 2803 vm_page_dequeue(m); 2804 vm_page_enqueue(PQ_ACTIVE, m); 2805 } else 2806 KASSERT(queue == PQ_NONE, 2807 ("vm_page_activate: wired page %p is queued", m)); 2808 } else { 2809 if (m->act_count < ACT_INIT) 2810 m->act_count = ACT_INIT; 2811 } 2812} 2813 2814/* 2815 * vm_page_free_wakeup: 2816 * 2817 * Helper routine for vm_page_free_toq(). This routine is called 2818 * when a page is added to the free queues. 2819 * 2820 * The page queues must be locked. 2821 */ 2822static void 2823vm_page_free_wakeup(void) 2824{ 2825 2826 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 2827 /* 2828 * if pageout daemon needs pages, then tell it that there are 2829 * some free. 2830 */ 2831 if (vm_pageout_pages_needed && 2832 vm_cnt.v_free_count >= vm_cnt.v_pageout_free_min) { 2833 wakeup(&vm_pageout_pages_needed); 2834 vm_pageout_pages_needed = 0; 2835 } 2836 /* 2837 * wakeup processes that are waiting on memory if we hit a 2838 * high water mark. And wakeup scheduler process if we have 2839 * lots of memory. this process will swapin processes. 2840 */ 2841 if (vm_pages_needed && !vm_page_count_min()) { 2842 vm_pages_needed = false; 2843 wakeup(&vm_cnt.v_free_count); 2844 } 2845} 2846 2847/* 2848 * vm_page_free_prep: 2849 * 2850 * Prepares the given page to be put on the free list, 2851 * disassociating it from any VM object. The caller may return 2852 * the page to the free list only if this function returns true. 2853 * 2854 * The object must be locked. The page must be locked if it is 2855 * managed. For a queued managed page, the pagequeue_locked 2856 * argument specifies whether the page queue is already locked. 2857 */ 2858bool 2859vm_page_free_prep(vm_page_t m, bool pagequeue_locked) 2860{ 2861 2862#if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP) 2863 if ((m->flags & PG_ZERO) != 0) { 2864 uint64_t *p; 2865 int i; 2866 p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); 2867 for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++) 2868 KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx", 2869 m, i, (uintmax_t)*p)); 2870 } 2871#endif 2872 if ((m->oflags & VPO_UNMANAGED) == 0) { 2873 vm_page_lock_assert(m, MA_OWNED); 2874 KASSERT(!pmap_page_is_mapped(m), 2875 ("vm_page_free_toq: freeing mapped page %p", m)); 2876 } else 2877 KASSERT(m->queue == PQ_NONE, 2878 ("vm_page_free_toq: unmanaged page %p is queued", m)); 2879 PCPU_INC(cnt.v_tfree); 2880 2881 if (vm_page_sbusied(m)) 2882 panic("vm_page_free: freeing busy page %p", m); 2883 2884 /* 2885 * Unqueue, then remove page. Note that we cannot destroy 2886 * the page here because we do not want to call the pager's 2887 * callback routine until after we've put the page on the 2888 * appropriate free queue. 2889 */ 2890 if (m->queue != PQ_NONE) { 2891 if (pagequeue_locked) 2892 vm_page_dequeue_locked(m); 2893 else 2894 vm_page_dequeue(m); 2895 } 2896 vm_page_remove(m); 2897 2898 /* 2899 * If fictitious remove object association and 2900 * return, otherwise delay object association removal. 2901 */ 2902 if ((m->flags & PG_FICTITIOUS) != 0) 2903 return (false); 2904 2905 m->valid = 0; 2906 vm_page_undirty(m); 2907 2908 if (m->wire_count != 0) 2909 panic("vm_page_free: freeing wired page %p", m); 2910 if (m->hold_count != 0) { 2911 m->flags &= ~PG_ZERO; 2912 KASSERT((m->flags & PG_UNHOLDFREE) == 0, 2913 ("vm_page_free: freeing PG_UNHOLDFREE page %p", m)); 2914 m->flags |= PG_UNHOLDFREE; 2915 return (false); 2916 } 2917 2918 /* 2919 * Restore the default memory attribute to the page. 2920 */ 2921 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 2922 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 2923 2924 return (true); 2925} 2926 2927/* 2928 * Insert the page into the physical memory allocator's free page 2929 * queues. This is the last step to free a page. 2930 */ 2931static void 2932vm_page_free_phys(vm_page_t m) 2933{ 2934 2935 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 2936 2937 vm_phys_freecnt_adj(m, 1); 2938#if VM_NRESERVLEVEL > 0 2939 if (!vm_reserv_free_page(m)) 2940#endif 2941 vm_phys_free_pages(m, 0); 2942 if ((m->flags & PG_ZERO) != 0) 2943 ++vm_page_zero_count; 2944 else 2945 vm_page_zero_idle_wakeup(); 2946} 2947 2948void 2949vm_page_free_phys_pglist(struct pglist *tq) 2950{ 2951 vm_page_t m; 2952 2953 if (TAILQ_EMPTY(tq)) 2954 return; 2955 mtx_lock(&vm_page_queue_free_mtx); 2956 TAILQ_FOREACH(m, tq, listq) 2957 vm_page_free_phys(m); 2958 vm_page_free_wakeup(); 2959 mtx_unlock(&vm_page_queue_free_mtx); 2960} 2961 2962/* 2963 * vm_page_free_toq: 2964 * 2965 * Returns the given page to the free list, disassociating it 2966 * from any VM object. 2967 * 2968 * The object must be locked. The page must be locked if it is 2969 * managed. 2970 */ 2971void 2972vm_page_free_toq(vm_page_t m) 2973{ 2974 2975 if (!vm_page_free_prep(m, false)) 2976 return; 2977 mtx_lock(&vm_page_queue_free_mtx); 2978 vm_page_free_phys(m); 2979 vm_page_free_wakeup(); 2980 mtx_unlock(&vm_page_queue_free_mtx); 2981} 2982 2983/* 2984 * vm_page_wire: 2985 * 2986 * Mark this page as wired down by yet 2987 * another map, removing it from paging queues 2988 * as necessary. 2989 * 2990 * If the page is fictitious, then its wire count must remain one. 2991 * 2992 * The page must be locked. 2993 */ 2994void 2995vm_page_wire(vm_page_t m) 2996{ 2997 2998 /* 2999 * Only bump the wire statistics if the page is not already wired, 3000 * and only unqueue the page if it is on some queue (if it is unmanaged 3001 * it is already off the queues). 3002 */ 3003 vm_page_lock_assert(m, MA_OWNED); 3004 if ((m->flags & PG_FICTITIOUS) != 0) { 3005 KASSERT(m->wire_count == 1, 3006 ("vm_page_wire: fictitious page %p's wire count isn't one", 3007 m)); 3008 return; 3009 } 3010 if (m->wire_count == 0) { 3011 KASSERT((m->oflags & VPO_UNMANAGED) == 0 || 3012 m->queue == PQ_NONE, 3013 ("vm_page_wire: unmanaged page %p is queued", m)); 3014 vm_page_remque(m); 3015 atomic_add_int(&vm_cnt.v_wire_count, 1); 3016 } 3017 m->wire_count++; 3018 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 3019} 3020 3021/* 3022 * vm_page_unwire: 3023 * 3024 * Release one wiring of the specified page, potentially allowing it to be 3025 * paged out. Returns TRUE if the number of wirings transitions to zero and 3026 * FALSE otherwise. 3027 * 3028 * Only managed pages belonging to an object can be paged out. If the number 3029 * of wirings transitions to zero and the page is eligible for page out, then 3030 * the page is added to the specified paging queue (unless PQ_NONE is 3031 * specified). 3032 * 3033 * If a page is fictitious, then its wire count must always be one. 3034 * 3035 * A managed page must be locked. 3036 */ 3037boolean_t 3038vm_page_unwire(vm_page_t m, uint8_t queue) 3039{ 3040 3041 KASSERT(queue < PQ_COUNT || queue == PQ_NONE, 3042 ("vm_page_unwire: invalid queue %u request for page %p", 3043 queue, m)); 3044 if ((m->oflags & VPO_UNMANAGED) == 0) 3045 vm_page_assert_locked(m); 3046 if ((m->flags & PG_FICTITIOUS) != 0) { 3047 KASSERT(m->wire_count == 1, 3048 ("vm_page_unwire: fictitious page %p's wire count isn't one", m)); 3049 return (FALSE); 3050 } 3051 if (m->wire_count > 0) { 3052 m->wire_count--; 3053 if (m->wire_count == 0) { 3054 atomic_subtract_int(&vm_cnt.v_wire_count, 1); 3055 if ((m->oflags & VPO_UNMANAGED) == 0 && 3056 m->object != NULL && queue != PQ_NONE) 3057 vm_page_enqueue(queue, m); 3058 return (TRUE); 3059 } else 3060 return (FALSE); 3061 } else 3062 panic("vm_page_unwire: page %p's wire count is zero", m); 3063} 3064 3065/* 3066 * Move the specified page to the inactive queue. 3067 * 3068 * Normally, "noreuse" is FALSE, resulting in LRU ordering of the inactive 3069 * queue. However, setting "noreuse" to TRUE will accelerate the specified 3070 * page's reclamation, but it will not unmap the page from any address space. 3071 * This is implemented by inserting the page near the head of the inactive 3072 * queue, using a marker page to guide FIFO insertion ordering. 3073 * 3074 * The page must be locked. 3075 */ 3076static inline void 3077_vm_page_deactivate(vm_page_t m, boolean_t noreuse) 3078{ 3079 struct vm_pagequeue *pq; 3080 int queue; 3081 3082 vm_page_assert_locked(m); 3083 3084 /* 3085 * Ignore if the page is already inactive, unless it is unlikely to be 3086 * reactivated. 3087 */ 3088 if ((queue = m->queue) == PQ_INACTIVE && !noreuse) 3089 return; 3090 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 3091 pq = &vm_phys_domain(m)->vmd_pagequeues[PQ_INACTIVE]; 3092 /* Avoid multiple acquisitions of the inactive queue lock. */ 3093 if (queue == PQ_INACTIVE) { 3094 vm_pagequeue_lock(pq); 3095 vm_page_dequeue_locked(m); 3096 } else { 3097 if (queue != PQ_NONE) 3098 vm_page_dequeue(m); 3099 vm_pagequeue_lock(pq); 3100 } 3101 m->queue = PQ_INACTIVE; 3102 if (noreuse) 3103 TAILQ_INSERT_BEFORE(&vm_phys_domain(m)->vmd_inacthead, 3104 m, plinks.q); 3105 else 3106 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3107 vm_pagequeue_cnt_inc(pq); 3108 vm_pagequeue_unlock(pq); 3109 } 3110} 3111 3112/* 3113 * Move the specified page to the inactive queue. 3114 * 3115 * The page must be locked. 3116 */ 3117void 3118vm_page_deactivate(vm_page_t m) 3119{ 3120 3121 _vm_page_deactivate(m, FALSE); 3122} 3123 3124/* 3125 * Move the specified page to the inactive queue with the expectation 3126 * that it is unlikely to be reused. 3127 * 3128 * The page must be locked. 3129 */ 3130void 3131vm_page_deactivate_noreuse(vm_page_t m) 3132{ 3133 3134 _vm_page_deactivate(m, TRUE); 3135} 3136 3137/* 3138 * vm_page_launder 3139 * 3140 * Put a page in the laundry. 3141 */ 3142void 3143vm_page_launder(vm_page_t m) 3144{ 3145 int queue; 3146 3147 vm_page_assert_locked(m); 3148 if ((queue = m->queue) != PQ_LAUNDRY) { 3149 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 3150 if (queue != PQ_NONE) 3151 vm_page_dequeue(m); 3152 vm_page_enqueue(PQ_LAUNDRY, m); 3153 } else 3154 KASSERT(queue == PQ_NONE, 3155 ("wired page %p is queued", m)); 3156 } 3157} 3158 3159/* 3160 * vm_page_try_to_free() 3161 * 3162 * Attempt to free the page. If we cannot free it, we do nothing. 3163 * true is returned on success, false on failure. 3164 */ 3165bool 3166vm_page_try_to_free(vm_page_t m) 3167{ 3168 3169 vm_page_assert_locked(m); 3170 if (m->object != NULL) 3171 VM_OBJECT_ASSERT_WLOCKED(m->object); 3172 if (m->dirty != 0 || m->hold_count != 0 || m->wire_count != 0 || 3173 (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m)) 3174 return (false); 3175 if (m->object != NULL && m->object->ref_count != 0) { 3176 pmap_remove_all(m); 3177 if (m->dirty != 0) 3178 return (false); 3179 } 3180 vm_page_free(m); 3181 return (true); 3182} 3183 3184/* 3185 * vm_page_advise 3186 * 3187 * Apply the specified advice to the given page. 3188 * 3189 * The object and page must be locked. 3190 */ 3191void 3192vm_page_advise(vm_page_t m, int advice) 3193{ 3194 3195 vm_page_assert_locked(m); 3196 VM_OBJECT_ASSERT_WLOCKED(m->object); 3197 if (advice == MADV_FREE) 3198 /* 3199 * Mark the page clean. This will allow the page to be freed 3200 * without first paging it out. MADV_FREE pages are often 3201 * quickly reused by malloc(3), so we do not do anything that 3202 * would result in a page fault on a later access. 3203 */ 3204 vm_page_undirty(m); 3205 else if (advice != MADV_DONTNEED) { 3206 if (advice == MADV_WILLNEED) 3207 vm_page_activate(m); 3208 return; 3209 } 3210 3211 /* 3212 * Clear any references to the page. Otherwise, the page daemon will 3213 * immediately reactivate the page. 3214 */ 3215 vm_page_aflag_clear(m, PGA_REFERENCED); 3216 3217 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) 3218 vm_page_dirty(m); 3219 3220 /* 3221 * Place clean pages near the head of the inactive queue rather than 3222 * the tail, thus defeating the queue's LRU operation and ensuring that 3223 * the page will be reused quickly. Dirty pages not already in the 3224 * laundry are moved there. 3225 */ 3226 if (m->dirty == 0) 3227 vm_page_deactivate_noreuse(m); 3228 else 3229 vm_page_launder(m); 3230} 3231 3232/* 3233 * Grab a page, waiting until we are waken up due to the page 3234 * changing state. We keep on waiting, if the page continues 3235 * to be in the object. If the page doesn't exist, first allocate it 3236 * and then conditionally zero it. 3237 * 3238 * This routine may sleep. 3239 * 3240 * The object must be locked on entry. The lock will, however, be released 3241 * and reacquired if the routine sleeps. 3242 */ 3243vm_page_t 3244vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 3245{ 3246 vm_page_t m; 3247 int sleep; 3248 int pflags; 3249 3250 VM_OBJECT_ASSERT_WLOCKED(object); 3251 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 3252 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 3253 ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 3254 pflags = allocflags & 3255 ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL); 3256 if ((allocflags & VM_ALLOC_NOWAIT) == 0) 3257 pflags |= VM_ALLOC_WAITFAIL; 3258retrylookup: 3259 if ((m = vm_page_lookup(object, pindex)) != NULL) { 3260 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ? 3261 vm_page_xbusied(m) : vm_page_busied(m); 3262 if (sleep) { 3263 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 3264 return (NULL); 3265 /* 3266 * Reference the page before unlocking and 3267 * sleeping so that the page daemon is less 3268 * likely to reclaim it. 3269 */ 3270 vm_page_aflag_set(m, PGA_REFERENCED); 3271 vm_page_lock(m); 3272 VM_OBJECT_WUNLOCK(object); 3273 vm_page_busy_sleep(m, "pgrbwt", (allocflags & 3274 VM_ALLOC_IGN_SBUSY) != 0); 3275 VM_OBJECT_WLOCK(object); 3276 goto retrylookup; 3277 } else { 3278 if ((allocflags & VM_ALLOC_WIRED) != 0) { 3279 vm_page_lock(m); 3280 vm_page_wire(m); 3281 vm_page_unlock(m); 3282 } 3283 if ((allocflags & 3284 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 3285 vm_page_xbusy(m); 3286 if ((allocflags & VM_ALLOC_SBUSY) != 0) 3287 vm_page_sbusy(m); 3288 return (m); 3289 } 3290 } 3291 m = vm_page_alloc(object, pindex, pflags); 3292 if (m == NULL) { 3293 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 3294 return (NULL); 3295 goto retrylookup; 3296 } 3297 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 3298 pmap_zero_page(m); 3299 return (m); 3300} 3301 3302/* 3303 * Return the specified range of pages from the given object. For each 3304 * page offset within the range, if a page already exists within the object 3305 * at that offset and it is busy, then wait for it to change state. If, 3306 * instead, the page doesn't exist, then allocate it. 3307 * 3308 * The caller must always specify an allocation class. 3309 * 3310 * allocation classes: 3311 * VM_ALLOC_NORMAL normal process request 3312 * VM_ALLOC_SYSTEM system *really* needs the pages 3313 * 3314 * The caller must always specify that the pages are to be busied and/or 3315 * wired. 3316 * 3317 * optional allocation flags: 3318 * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages 3319 * VM_ALLOC_NOBUSY do not exclusive busy the page 3320 * VM_ALLOC_NOWAIT do not sleep 3321 * VM_ALLOC_SBUSY set page to sbusy state 3322 * VM_ALLOC_WIRED wire the pages 3323 * VM_ALLOC_ZERO zero and validate any invalid pages 3324 * 3325 * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it 3326 * may return a partial prefix of the requested range. 3327 */ 3328int 3329vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags, 3330 vm_page_t *ma, int count) 3331{ 3332 vm_page_t m, mpred; 3333 int pflags; 3334 int i; 3335 bool sleep; 3336 3337 VM_OBJECT_ASSERT_WLOCKED(object); 3338 KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0, 3339 ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed")); 3340 KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 || 3341 (allocflags & VM_ALLOC_WIRED) != 0, 3342 ("vm_page_grab_pages: the pages must be busied or wired")); 3343 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 3344 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 3345 ("vm_page_grab_pages: VM_ALLOC_SBUSY/IGN_SBUSY mismatch")); 3346 if (count == 0) 3347 return (0); 3348 pflags = allocflags & ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | 3349 VM_ALLOC_WAITFAIL | VM_ALLOC_IGN_SBUSY); 3350 if ((allocflags & VM_ALLOC_NOWAIT) == 0) 3351 pflags |= VM_ALLOC_WAITFAIL; 3352 i = 0; 3353retrylookup: 3354 m = vm_radix_lookup_le(&object->rtree, pindex + i); 3355 if (m == NULL || m->pindex != pindex + i) { 3356 mpred = m; 3357 m = NULL; 3358 } else 3359 mpred = TAILQ_PREV(m, pglist, listq); 3360 for (; i < count; i++) { 3361 if (m != NULL) { 3362 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ? 3363 vm_page_xbusied(m) : vm_page_busied(m); 3364 if (sleep) { 3365 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 3366 break; 3367 /* 3368 * Reference the page before unlocking and 3369 * sleeping so that the page daemon is less 3370 * likely to reclaim it. 3371 */ 3372 vm_page_aflag_set(m, PGA_REFERENCED); 3373 vm_page_lock(m); 3374 VM_OBJECT_WUNLOCK(object); 3375 vm_page_busy_sleep(m, "grbmaw", (allocflags & 3376 VM_ALLOC_IGN_SBUSY) != 0); 3377 VM_OBJECT_WLOCK(object); 3378 goto retrylookup; 3379 } 3380 if ((allocflags & VM_ALLOC_WIRED) != 0) { 3381 vm_page_lock(m); 3382 vm_page_wire(m); 3383 vm_page_unlock(m); 3384 } 3385 if ((allocflags & (VM_ALLOC_NOBUSY | 3386 VM_ALLOC_SBUSY)) == 0) 3387 vm_page_xbusy(m); 3388 if ((allocflags & VM_ALLOC_SBUSY) != 0) 3389 vm_page_sbusy(m); 3390 } else { 3391 m = vm_page_alloc_after(object, pindex + i, 3392 pflags | VM_ALLOC_COUNT(count - i), mpred); 3393 if (m == NULL) { 3394 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 3395 break; 3396 goto retrylookup; 3397 } 3398 } 3399 if (m->valid == 0 && (allocflags & VM_ALLOC_ZERO) != 0) { 3400 if ((m->flags & PG_ZERO) == 0) 3401 pmap_zero_page(m); 3402 m->valid = VM_PAGE_BITS_ALL; 3403 } 3404 ma[i] = mpred = m; 3405 m = vm_page_next(m); 3406 } 3407 return (i); 3408} 3409 3410/* 3411 * Mapping function for valid or dirty bits in a page. 3412 * 3413 * Inputs are required to range within a page. 3414 */ 3415vm_page_bits_t 3416vm_page_bits(int base, int size) 3417{ 3418 int first_bit; 3419 int last_bit; 3420 3421 KASSERT( 3422 base + size <= PAGE_SIZE, 3423 ("vm_page_bits: illegal base/size %d/%d", base, size) 3424 ); 3425 3426 if (size == 0) /* handle degenerate case */ 3427 return (0); 3428 3429 first_bit = base >> DEV_BSHIFT; 3430 last_bit = (base + size - 1) >> DEV_BSHIFT; 3431 3432 return (((vm_page_bits_t)2 << last_bit) - 3433 ((vm_page_bits_t)1 << first_bit)); 3434} 3435 3436/* 3437 * vm_page_set_valid_range: 3438 * 3439 * Sets portions of a page valid. The arguments are expected 3440 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 3441 * of any partial chunks touched by the range. The invalid portion of 3442 * such chunks will be zeroed. 3443 * 3444 * (base + size) must be less then or equal to PAGE_SIZE. 3445 */ 3446void 3447vm_page_set_valid_range(vm_page_t m, int base, int size) 3448{ 3449 int endoff, frag; 3450 3451 VM_OBJECT_ASSERT_WLOCKED(m->object); 3452 if (size == 0) /* handle degenerate case */ 3453 return; 3454 3455 /* 3456 * If the base is not DEV_BSIZE aligned and the valid 3457 * bit is clear, we have to zero out a portion of the 3458 * first block. 3459 */ 3460 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 3461 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 3462 pmap_zero_page_area(m, frag, base - frag); 3463 3464 /* 3465 * If the ending offset is not DEV_BSIZE aligned and the 3466 * valid bit is clear, we have to zero out a portion of 3467 * the last block. 3468 */ 3469 endoff = base + size; 3470 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 3471 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 3472 pmap_zero_page_area(m, endoff, 3473 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 3474 3475 /* 3476 * Assert that no previously invalid block that is now being validated 3477 * is already dirty. 3478 */ 3479 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 3480 ("vm_page_set_valid_range: page %p is dirty", m)); 3481 3482 /* 3483 * Set valid bits inclusive of any overlap. 3484 */ 3485 m->valid |= vm_page_bits(base, size); 3486} 3487 3488/* 3489 * Clear the given bits from the specified page's dirty field. 3490 */ 3491static __inline void 3492vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) 3493{ 3494 uintptr_t addr; 3495#if PAGE_SIZE < 16384 3496 int shift; 3497#endif 3498 3499 /* 3500 * If the object is locked and the page is neither exclusive busy nor 3501 * write mapped, then the page's dirty field cannot possibly be 3502 * set by a concurrent pmap operation. 3503 */ 3504 VM_OBJECT_ASSERT_WLOCKED(m->object); 3505 if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) 3506 m->dirty &= ~pagebits; 3507 else { 3508 /* 3509 * The pmap layer can call vm_page_dirty() without 3510 * holding a distinguished lock. The combination of 3511 * the object's lock and an atomic operation suffice 3512 * to guarantee consistency of the page dirty field. 3513 * 3514 * For PAGE_SIZE == 32768 case, compiler already 3515 * properly aligns the dirty field, so no forcible 3516 * alignment is needed. Only require existence of 3517 * atomic_clear_64 when page size is 32768. 3518 */ 3519 addr = (uintptr_t)&m->dirty; 3520#if PAGE_SIZE == 32768 3521 atomic_clear_64((uint64_t *)addr, pagebits); 3522#elif PAGE_SIZE == 16384 3523 atomic_clear_32((uint32_t *)addr, pagebits); 3524#else /* PAGE_SIZE <= 8192 */ 3525 /* 3526 * Use a trick to perform a 32-bit atomic on the 3527 * containing aligned word, to not depend on the existence 3528 * of atomic_clear_{8, 16}. 3529 */ 3530 shift = addr & (sizeof(uint32_t) - 1); 3531#if BYTE_ORDER == BIG_ENDIAN 3532 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY; 3533#else 3534 shift *= NBBY; 3535#endif 3536 addr &= ~(sizeof(uint32_t) - 1); 3537 atomic_clear_32((uint32_t *)addr, pagebits << shift); 3538#endif /* PAGE_SIZE */ 3539 } 3540} 3541 3542/* 3543 * vm_page_set_validclean: 3544 * 3545 * Sets portions of a page valid and clean. The arguments are expected 3546 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 3547 * of any partial chunks touched by the range. The invalid portion of 3548 * such chunks will be zero'd. 3549 * 3550 * (base + size) must be less then or equal to PAGE_SIZE. 3551 */ 3552void 3553vm_page_set_validclean(vm_page_t m, int base, int size) 3554{ 3555 vm_page_bits_t oldvalid, pagebits; 3556 int endoff, frag; 3557 3558 VM_OBJECT_ASSERT_WLOCKED(m->object); 3559 if (size == 0) /* handle degenerate case */ 3560 return; 3561 3562 /* 3563 * If the base is not DEV_BSIZE aligned and the valid 3564 * bit is clear, we have to zero out a portion of the 3565 * first block. 3566 */ 3567 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 3568 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) 3569 pmap_zero_page_area(m, frag, base - frag); 3570 3571 /* 3572 * If the ending offset is not DEV_BSIZE aligned and the 3573 * valid bit is clear, we have to zero out a portion of 3574 * the last block. 3575 */ 3576 endoff = base + size; 3577 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 3578 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) 3579 pmap_zero_page_area(m, endoff, 3580 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 3581 3582 /* 3583 * Set valid, clear dirty bits. If validating the entire 3584 * page we can safely clear the pmap modify bit. We also 3585 * use this opportunity to clear the VPO_NOSYNC flag. If a process 3586 * takes a write fault on a MAP_NOSYNC memory area the flag will 3587 * be set again. 3588 * 3589 * We set valid bits inclusive of any overlap, but we can only 3590 * clear dirty bits for DEV_BSIZE chunks that are fully within 3591 * the range. 3592 */ 3593 oldvalid = m->valid; 3594 pagebits = vm_page_bits(base, size); 3595 m->valid |= pagebits; 3596#if 0 /* NOT YET */ 3597 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 3598 frag = DEV_BSIZE - frag; 3599 base += frag; 3600 size -= frag; 3601 if (size < 0) 3602 size = 0; 3603 } 3604 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 3605#endif 3606 if (base == 0 && size == PAGE_SIZE) { 3607 /* 3608 * The page can only be modified within the pmap if it is 3609 * mapped, and it can only be mapped if it was previously 3610 * fully valid. 3611 */ 3612 if (oldvalid == VM_PAGE_BITS_ALL) 3613 /* 3614 * Perform the pmap_clear_modify() first. Otherwise, 3615 * a concurrent pmap operation, such as 3616 * pmap_protect(), could clear a modification in the 3617 * pmap and set the dirty field on the page before 3618 * pmap_clear_modify() had begun and after the dirty 3619 * field was cleared here. 3620 */ 3621 pmap_clear_modify(m); 3622 m->dirty = 0; 3623 m->oflags &= ~VPO_NOSYNC; 3624 } else if (oldvalid != VM_PAGE_BITS_ALL) 3625 m->dirty &= ~pagebits; 3626 else 3627 vm_page_clear_dirty_mask(m, pagebits); 3628} 3629 3630void 3631vm_page_clear_dirty(vm_page_t m, int base, int size) 3632{ 3633 3634 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 3635} 3636 3637/* 3638 * vm_page_set_invalid: 3639 * 3640 * Invalidates DEV_BSIZE'd chunks within a page. Both the 3641 * valid and dirty bits for the effected areas are cleared. 3642 */ 3643void 3644vm_page_set_invalid(vm_page_t m, int base, int size) 3645{ 3646 vm_page_bits_t bits; 3647 vm_object_t object; 3648 3649 object = m->object; 3650 VM_OBJECT_ASSERT_WLOCKED(object); 3651 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + 3652 size >= object->un_pager.vnp.vnp_size) 3653 bits = VM_PAGE_BITS_ALL; 3654 else 3655 bits = vm_page_bits(base, size); 3656 if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL && 3657 bits != 0) 3658 pmap_remove_all(m); 3659 KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) || 3660 !pmap_page_is_mapped(m), 3661 ("vm_page_set_invalid: page %p is mapped", m)); 3662 m->valid &= ~bits; 3663 m->dirty &= ~bits; 3664} 3665 3666/* 3667 * vm_page_zero_invalid() 3668 * 3669 * The kernel assumes that the invalid portions of a page contain 3670 * garbage, but such pages can be mapped into memory by user code. 3671 * When this occurs, we must zero out the non-valid portions of the 3672 * page so user code sees what it expects. 3673 * 3674 * Pages are most often semi-valid when the end of a file is mapped 3675 * into memory and the file's size is not page aligned. 3676 */ 3677void 3678vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 3679{ 3680 int b; 3681 int i; 3682 3683 VM_OBJECT_ASSERT_WLOCKED(m->object); 3684 /* 3685 * Scan the valid bits looking for invalid sections that 3686 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the 3687 * valid bit may be set ) have already been zeroed by 3688 * vm_page_set_validclean(). 3689 */ 3690 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 3691 if (i == (PAGE_SIZE / DEV_BSIZE) || 3692 (m->valid & ((vm_page_bits_t)1 << i))) { 3693 if (i > b) { 3694 pmap_zero_page_area(m, 3695 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 3696 } 3697 b = i + 1; 3698 } 3699 } 3700 3701 /* 3702 * setvalid is TRUE when we can safely set the zero'd areas 3703 * as being valid. We can do this if there are no cache consistancy 3704 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 3705 */ 3706 if (setvalid) 3707 m->valid = VM_PAGE_BITS_ALL; 3708} 3709 3710/* 3711 * vm_page_is_valid: 3712 * 3713 * Is (partial) page valid? Note that the case where size == 0 3714 * will return FALSE in the degenerate case where the page is 3715 * entirely invalid, and TRUE otherwise. 3716 */ 3717int 3718vm_page_is_valid(vm_page_t m, int base, int size) 3719{ 3720 vm_page_bits_t bits; 3721 3722 VM_OBJECT_ASSERT_LOCKED(m->object); 3723 bits = vm_page_bits(base, size); 3724 return (m->valid != 0 && (m->valid & bits) == bits); 3725} 3726 3727/* 3728 * Returns true if all of the specified predicates are true for the entire 3729 * (super)page and false otherwise. 3730 */ 3731bool 3732vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m) 3733{ 3734 vm_object_t object; 3735 int i, npages; 3736 3737 object = m->object; 3738 if (skip_m != NULL && skip_m->object != object) 3739 return (false); 3740 VM_OBJECT_ASSERT_LOCKED(object); 3741 npages = atop(pagesizes[m->psind]); 3742 3743 /* 3744 * The physically contiguous pages that make up a superpage, i.e., a 3745 * page with a page size index ("psind") greater than zero, will 3746 * occupy adjacent entries in vm_page_array[]. 3747 */ 3748 for (i = 0; i < npages; i++) { 3749 /* Always test object consistency, including "skip_m". */ 3750 if (m[i].object != object) 3751 return (false); 3752 if (&m[i] == skip_m) 3753 continue; 3754 if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i])) 3755 return (false); 3756 if ((flags & PS_ALL_DIRTY) != 0) { 3757 /* 3758 * Calling vm_page_test_dirty() or pmap_is_modified() 3759 * might stop this case from spuriously returning 3760 * "false". However, that would require a write lock 3761 * on the object containing "m[i]". 3762 */ 3763 if (m[i].dirty != VM_PAGE_BITS_ALL) 3764 return (false); 3765 } 3766 if ((flags & PS_ALL_VALID) != 0 && 3767 m[i].valid != VM_PAGE_BITS_ALL) 3768 return (false); 3769 } 3770 return (true); 3771} 3772 3773/* 3774 * Set the page's dirty bits if the page is modified. 3775 */ 3776void 3777vm_page_test_dirty(vm_page_t m) 3778{ 3779 3780 VM_OBJECT_ASSERT_WLOCKED(m->object); 3781 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 3782 vm_page_dirty(m); 3783} 3784 3785void 3786vm_page_lock_KBI(vm_page_t m, const char *file, int line) 3787{ 3788 3789 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); 3790} 3791 3792void 3793vm_page_unlock_KBI(vm_page_t m, const char *file, int line) 3794{ 3795 3796 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); 3797} 3798 3799int 3800vm_page_trylock_KBI(vm_page_t m, const char *file, int line) 3801{ 3802 3803 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); 3804} 3805 3806#if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) 3807void 3808vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line) 3809{ 3810 3811 vm_page_lock_assert_KBI(m, MA_OWNED, file, line); 3812} 3813 3814void 3815vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) 3816{ 3817 3818 mtx_assert_(vm_page_lockptr(m), a, file, line); 3819} 3820#endif 3821 3822#ifdef INVARIANTS 3823void 3824vm_page_object_lock_assert(vm_page_t m) 3825{ 3826 3827 /* 3828 * Certain of the page's fields may only be modified by the 3829 * holder of the containing object's lock or the exclusive busy. 3830 * holder. Unfortunately, the holder of the write busy is 3831 * not recorded, and thus cannot be checked here. 3832 */ 3833 if (m->object != NULL && !vm_page_xbusied(m)) 3834 VM_OBJECT_ASSERT_WLOCKED(m->object); 3835} 3836 3837void 3838vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits) 3839{ 3840 3841 if ((bits & PGA_WRITEABLE) == 0) 3842 return; 3843 3844 /* 3845 * The PGA_WRITEABLE flag can only be set if the page is 3846 * managed, is exclusively busied or the object is locked. 3847 * Currently, this flag is only set by pmap_enter(). 3848 */ 3849 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 3850 ("PGA_WRITEABLE on unmanaged page")); 3851 if (!vm_page_xbusied(m)) 3852 VM_OBJECT_ASSERT_LOCKED(m->object); 3853} 3854#endif 3855 3856#include "opt_ddb.h" 3857#ifdef DDB 3858#include <sys/kernel.h> 3859 3860#include <ddb/ddb.h> 3861 3862DB_SHOW_COMMAND(page, vm_page_print_page_info) 3863{ 3864 3865 db_printf("vm_cnt.v_free_count: %d\n", vm_cnt.v_free_count); 3866 db_printf("vm_cnt.v_inactive_count: %d\n", vm_cnt.v_inactive_count); 3867 db_printf("vm_cnt.v_active_count: %d\n", vm_cnt.v_active_count); 3868 db_printf("vm_cnt.v_laundry_count: %d\n", vm_cnt.v_laundry_count); 3869 db_printf("vm_cnt.v_wire_count: %d\n", vm_cnt.v_wire_count); 3870 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved); 3871 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min); 3872 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target); 3873 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target); 3874} 3875 3876DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 3877{ 3878 int dom; 3879 3880 db_printf("pq_free %d\n", vm_cnt.v_free_count); 3881 for (dom = 0; dom < vm_ndomains; dom++) { 3882 db_printf( 3883 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d\n", 3884 dom, 3885 vm_dom[dom].vmd_page_count, 3886 vm_dom[dom].vmd_free_count, 3887 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, 3888 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, 3889 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt); 3890 } 3891} 3892 3893DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) 3894{ 3895 vm_page_t m; 3896 boolean_t phys, virt; 3897 3898 if (!have_addr) { 3899 db_printf("show pginfo addr\n"); 3900 return; 3901 } 3902 3903 phys = strchr(modif, 'p') != NULL; 3904 virt = strchr(modif, 'v') != NULL; 3905 if (virt) 3906 m = PHYS_TO_VM_PAGE(pmap_kextract(addr)); 3907 else if (phys) 3908 m = PHYS_TO_VM_PAGE(addr); 3909 else 3910 m = (vm_page_t)addr; 3911 db_printf( 3912 "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n" 3913 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", 3914 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, 3915 m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags, 3916 m->flags, m->act_count, m->busy_lock, m->valid, m->dirty); 3917} 3918#endif /* DDB */ 3919