vm_page.c revision 301210
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: head/sys/vm/vm_page.c 301210 2016-06-02 16:58:47Z markj $"); 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/sysctl.h> 103#include <sys/vmmeter.h> 104#include <sys/vnode.h> 105 106#include <vm/vm.h> 107#include <vm/pmap.h> 108#include <vm/vm_param.h> 109#include <vm/vm_kern.h> 110#include <vm/vm_object.h> 111#include <vm/vm_page.h> 112#include <vm/vm_pageout.h> 113#include <vm/vm_pager.h> 114#include <vm/vm_phys.h> 115#include <vm/vm_radix.h> 116#include <vm/vm_reserv.h> 117#include <vm/vm_extern.h> 118#include <vm/uma.h> 119#include <vm/uma_int.h> 120 121#include <machine/md_var.h> 122 123/* 124 * Associated with page of user-allocatable memory is a 125 * page structure. 126 */ 127 128struct vm_domain vm_dom[MAXMEMDOM]; 129struct mtx_padalign vm_page_queue_free_mtx; 130 131struct mtx_padalign pa_lock[PA_LOCK_COUNT]; 132 133vm_page_t vm_page_array; 134long vm_page_array_size; 135long first_page; 136int vm_page_zero_count; 137 138static int boot_pages = UMA_BOOT_PAGES; 139SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, 140 &boot_pages, 0, 141 "number of pages allocated for bootstrapping the VM system"); 142 143static int pa_tryrelock_restart; 144SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD, 145 &pa_tryrelock_restart, 0, "Number of tryrelock restarts"); 146 147static TAILQ_HEAD(, vm_page) blacklist_head; 148static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS); 149SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD | 150 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages"); 151 152/* Is the page daemon waiting for free pages? */ 153static int vm_pageout_pages_needed; 154 155static uma_zone_t fakepg_zone; 156 157static struct vnode *vm_page_alloc_init(vm_page_t m); 158static void vm_page_cache_turn_free(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_wakeup(void); 162static void vm_page_init_fakepg(void *dummy); 163static int vm_page_insert_after(vm_page_t m, vm_object_t object, 164 vm_pindex_t pindex, vm_page_t mpred); 165static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object, 166 vm_page_t mpred); 167static int vm_page_reclaim_run(int req_class, u_long npages, vm_page_t m_run, 168 vm_paddr_t high); 169 170SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init_fakepg, NULL); 171 172static void 173vm_page_init_fakepg(void *dummy) 174{ 175 176 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL, 177 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM); 178} 179 180/* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */ 181#if PAGE_SIZE == 32768 182#ifdef CTASSERT 183CTASSERT(sizeof(u_long) >= 8); 184#endif 185#endif 186 187/* 188 * Try to acquire a physical address lock while a pmap is locked. If we 189 * fail to trylock we unlock and lock the pmap directly and cache the 190 * locked pa in *locked. The caller should then restart their loop in case 191 * the virtual to physical mapping has changed. 192 */ 193int 194vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked) 195{ 196 vm_paddr_t lockpa; 197 198 lockpa = *locked; 199 *locked = pa; 200 if (lockpa) { 201 PA_LOCK_ASSERT(lockpa, MA_OWNED); 202 if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa)) 203 return (0); 204 PA_UNLOCK(lockpa); 205 } 206 if (PA_TRYLOCK(pa)) 207 return (0); 208 PMAP_UNLOCK(pmap); 209 atomic_add_int(&pa_tryrelock_restart, 1); 210 PA_LOCK(pa); 211 PMAP_LOCK(pmap); 212 return (EAGAIN); 213} 214 215/* 216 * vm_set_page_size: 217 * 218 * Sets the page size, perhaps based upon the memory 219 * size. Must be called before any use of page-size 220 * dependent functions. 221 */ 222void 223vm_set_page_size(void) 224{ 225 if (vm_cnt.v_page_size == 0) 226 vm_cnt.v_page_size = PAGE_SIZE; 227 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0) 228 panic("vm_set_page_size: page size not a power of two"); 229} 230 231/* 232 * vm_page_blacklist_next: 233 * 234 * Find the next entry in the provided string of blacklist 235 * addresses. Entries are separated by space, comma, or newline. 236 * If an invalid integer is encountered then the rest of the 237 * string is skipped. Updates the list pointer to the next 238 * character, or NULL if the string is exhausted or invalid. 239 */ 240static vm_paddr_t 241vm_page_blacklist_next(char **list, char *end) 242{ 243 vm_paddr_t bad; 244 char *cp, *pos; 245 246 if (list == NULL || *list == NULL) 247 return (0); 248 if (**list =='\0') { 249 *list = NULL; 250 return (0); 251 } 252 253 /* 254 * If there's no end pointer then the buffer is coming from 255 * the kenv and we know it's null-terminated. 256 */ 257 if (end == NULL) 258 end = *list + strlen(*list); 259 260 /* Ensure that strtoq() won't walk off the end */ 261 if (*end != '\0') { 262 if (*end == '\n' || *end == ' ' || *end == ',') 263 *end = '\0'; 264 else { 265 printf("Blacklist not terminated, skipping\n"); 266 *list = NULL; 267 return (0); 268 } 269 } 270 271 for (pos = *list; *pos != '\0'; pos = cp) { 272 bad = strtoq(pos, &cp, 0); 273 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') { 274 if (bad == 0) { 275 if (++cp < end) 276 continue; 277 else 278 break; 279 } 280 } else 281 break; 282 if (*cp == '\0' || ++cp >= end) 283 *list = NULL; 284 else 285 *list = cp; 286 return (trunc_page(bad)); 287 } 288 printf("Garbage in RAM blacklist, skipping\n"); 289 *list = NULL; 290 return (0); 291} 292 293/* 294 * vm_page_blacklist_check: 295 * 296 * Iterate through the provided string of blacklist addresses, pulling 297 * each entry out of the physical allocator free list and putting it 298 * onto a list for reporting via the vm.page_blacklist sysctl. 299 */ 300static void 301vm_page_blacklist_check(char *list, char *end) 302{ 303 vm_paddr_t pa; 304 vm_page_t m; 305 char *next; 306 int ret; 307 308 next = list; 309 while (next != NULL) { 310 if ((pa = vm_page_blacklist_next(&next, end)) == 0) 311 continue; 312 m = vm_phys_paddr_to_vm_page(pa); 313 if (m == NULL) 314 continue; 315 mtx_lock(&vm_page_queue_free_mtx); 316 ret = vm_phys_unfree_page(m); 317 mtx_unlock(&vm_page_queue_free_mtx); 318 if (ret == TRUE) { 319 TAILQ_INSERT_TAIL(&blacklist_head, m, listq); 320 if (bootverbose) 321 printf("Skipping page with pa 0x%jx\n", 322 (uintmax_t)pa); 323 } 324 } 325} 326 327/* 328 * vm_page_blacklist_load: 329 * 330 * Search for a special module named "ram_blacklist". It'll be a 331 * plain text file provided by the user via the loader directive 332 * of the same name. 333 */ 334static void 335vm_page_blacklist_load(char **list, char **end) 336{ 337 void *mod; 338 u_char *ptr; 339 u_int len; 340 341 mod = NULL; 342 ptr = NULL; 343 344 mod = preload_search_by_type("ram_blacklist"); 345 if (mod != NULL) { 346 ptr = preload_fetch_addr(mod); 347 len = preload_fetch_size(mod); 348 } 349 *list = ptr; 350 if (ptr != NULL) 351 *end = ptr + len; 352 else 353 *end = NULL; 354 return; 355} 356 357static int 358sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS) 359{ 360 vm_page_t m; 361 struct sbuf sbuf; 362 int error, first; 363 364 first = 1; 365 error = sysctl_wire_old_buffer(req, 0); 366 if (error != 0) 367 return (error); 368 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 369 TAILQ_FOREACH(m, &blacklist_head, listq) { 370 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",", 371 (uintmax_t)m->phys_addr); 372 first = 0; 373 } 374 error = sbuf_finish(&sbuf); 375 sbuf_delete(&sbuf); 376 return (error); 377} 378 379static void 380vm_page_domain_init(struct vm_domain *vmd) 381{ 382 struct vm_pagequeue *pq; 383 int i; 384 385 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) = 386 "vm inactive pagequeue"; 387 *__DECONST(u_int **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_vcnt) = 388 &vm_cnt.v_inactive_count; 389 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) = 390 "vm active pagequeue"; 391 *__DECONST(u_int **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_vcnt) = 392 &vm_cnt.v_active_count; 393 vmd->vmd_page_count = 0; 394 vmd->vmd_free_count = 0; 395 vmd->vmd_segs = 0; 396 vmd->vmd_oom = FALSE; 397 vmd->vmd_pass = 0; 398 for (i = 0; i < PQ_COUNT; i++) { 399 pq = &vmd->vmd_pagequeues[i]; 400 TAILQ_INIT(&pq->pq_pl); 401 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue", 402 MTX_DEF | MTX_DUPOK); 403 } 404} 405 406/* 407 * vm_page_startup: 408 * 409 * Initializes the resident memory module. 410 * 411 * Allocates memory for the page cells, and 412 * for the object/offset-to-page hash table headers. 413 * Each page cell is initialized and placed on the free list. 414 */ 415vm_offset_t 416vm_page_startup(vm_offset_t vaddr) 417{ 418 vm_offset_t mapped; 419 vm_paddr_t page_range; 420 vm_paddr_t new_end; 421 int i; 422 vm_paddr_t pa; 423 vm_paddr_t last_pa; 424 char *list, *listend; 425 vm_paddr_t end; 426 vm_paddr_t biggestsize; 427 vm_paddr_t low_water, high_water; 428 int biggestone; 429 430 biggestsize = 0; 431 biggestone = 0; 432 vaddr = round_page(vaddr); 433 434 for (i = 0; phys_avail[i + 1]; i += 2) { 435 phys_avail[i] = round_page(phys_avail[i]); 436 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 437 } 438 439 low_water = phys_avail[0]; 440 high_water = phys_avail[1]; 441 442 for (i = 0; i < vm_phys_nsegs; i++) { 443 if (vm_phys_segs[i].start < low_water) 444 low_water = vm_phys_segs[i].start; 445 if (vm_phys_segs[i].end > high_water) 446 high_water = vm_phys_segs[i].end; 447 } 448 for (i = 0; phys_avail[i + 1]; i += 2) { 449 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i]; 450 451 if (size > biggestsize) { 452 biggestone = i; 453 biggestsize = size; 454 } 455 if (phys_avail[i] < low_water) 456 low_water = phys_avail[i]; 457 if (phys_avail[i + 1] > high_water) 458 high_water = phys_avail[i + 1]; 459 } 460 461 end = phys_avail[biggestone+1]; 462 463 /* 464 * Initialize the page and queue locks. 465 */ 466 mtx_init(&vm_page_queue_free_mtx, "vm page free queue", NULL, MTX_DEF); 467 for (i = 0; i < PA_LOCK_COUNT; i++) 468 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF); 469 for (i = 0; i < vm_ndomains; i++) 470 vm_page_domain_init(&vm_dom[i]); 471 472 /* 473 * Allocate memory for use when boot strapping the kernel memory 474 * allocator. 475 * 476 * CTFLAG_RDTUN doesn't work during the early boot process, so we must 477 * manually fetch the value. 478 */ 479 TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages); 480 new_end = end - (boot_pages * UMA_SLAB_SIZE); 481 new_end = trunc_page(new_end); 482 mapped = pmap_map(&vaddr, new_end, end, 483 VM_PROT_READ | VM_PROT_WRITE); 484 bzero((void *)mapped, end - new_end); 485 uma_startup((void *)mapped, boot_pages); 486 487#if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \ 488 defined(__i386__) || defined(__mips__) 489 /* 490 * Allocate a bitmap to indicate that a random physical page 491 * needs to be included in a minidump. 492 * 493 * The amd64 port needs this to indicate which direct map pages 494 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 495 * 496 * However, i386 still needs this workspace internally within the 497 * minidump code. In theory, they are not needed on i386, but are 498 * included should the sf_buf code decide to use them. 499 */ 500 last_pa = 0; 501 for (i = 0; dump_avail[i + 1] != 0; i += 2) 502 if (dump_avail[i + 1] > last_pa) 503 last_pa = dump_avail[i + 1]; 504 page_range = last_pa / PAGE_SIZE; 505 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 506 new_end -= vm_page_dump_size; 507 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, 508 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); 509 bzero((void *)vm_page_dump, vm_page_dump_size); 510#endif 511#ifdef __amd64__ 512 /* 513 * Request that the physical pages underlying the message buffer be 514 * included in a crash dump. Since the message buffer is accessed 515 * through the direct map, they are not automatically included. 516 */ 517 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); 518 last_pa = pa + round_page(msgbufsize); 519 while (pa < last_pa) { 520 dump_add_page(pa); 521 pa += PAGE_SIZE; 522 } 523#endif 524 /* 525 * Compute the number of pages of memory that will be available for 526 * use (taking into account the overhead of a page structure per 527 * page). 528 */ 529 first_page = low_water / PAGE_SIZE; 530#ifdef VM_PHYSSEG_SPARSE 531 page_range = 0; 532 for (i = 0; i < vm_phys_nsegs; i++) { 533 page_range += atop(vm_phys_segs[i].end - 534 vm_phys_segs[i].start); 535 } 536 for (i = 0; phys_avail[i + 1] != 0; i += 2) 537 page_range += atop(phys_avail[i + 1] - phys_avail[i]); 538#elif defined(VM_PHYSSEG_DENSE) 539 page_range = high_water / PAGE_SIZE - first_page; 540#else 541#error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 542#endif 543 end = new_end; 544 545 /* 546 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 547 */ 548 vaddr += PAGE_SIZE; 549 550 /* 551 * Initialize the mem entry structures now, and put them in the free 552 * queue. 553 */ 554 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 555 mapped = pmap_map(&vaddr, new_end, end, 556 VM_PROT_READ | VM_PROT_WRITE); 557 vm_page_array = (vm_page_t) mapped; 558#if VM_NRESERVLEVEL > 0 559 /* 560 * Allocate memory for the reservation management system's data 561 * structures. 562 */ 563 new_end = vm_reserv_startup(&vaddr, new_end, high_water); 564#endif 565#if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) 566 /* 567 * pmap_map on arm64, amd64, and mips can come out of the direct-map, 568 * not kvm like i386, so the pages must be tracked for a crashdump to 569 * include this data. This includes the vm_page_array and the early 570 * UMA bootstrap pages. 571 */ 572 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE) 573 dump_add_page(pa); 574#endif 575 phys_avail[biggestone + 1] = new_end; 576 577 /* 578 * Add physical memory segments corresponding to the available 579 * physical pages. 580 */ 581 for (i = 0; phys_avail[i + 1] != 0; i += 2) 582 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]); 583 584 /* 585 * Clear all of the page structures 586 */ 587 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 588 for (i = 0; i < page_range; i++) 589 vm_page_array[i].order = VM_NFREEORDER; 590 vm_page_array_size = page_range; 591 592 /* 593 * Initialize the physical memory allocator. 594 */ 595 vm_phys_init(); 596 597 /* 598 * Add every available physical page that is not blacklisted to 599 * the free lists. 600 */ 601 vm_cnt.v_page_count = 0; 602 vm_cnt.v_free_count = 0; 603 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 604 pa = phys_avail[i]; 605 last_pa = phys_avail[i + 1]; 606 while (pa < last_pa) { 607 vm_phys_add_page(pa); 608 pa += PAGE_SIZE; 609 } 610 } 611 612 TAILQ_INIT(&blacklist_head); 613 vm_page_blacklist_load(&list, &listend); 614 vm_page_blacklist_check(list, listend); 615 616 list = kern_getenv("vm.blacklist"); 617 vm_page_blacklist_check(list, NULL); 618 619 freeenv(list); 620#if VM_NRESERVLEVEL > 0 621 /* 622 * Initialize the reservation management system. 623 */ 624 vm_reserv_init(); 625#endif 626 return (vaddr); 627} 628 629void 630vm_page_reference(vm_page_t m) 631{ 632 633 vm_page_aflag_set(m, PGA_REFERENCED); 634} 635 636/* 637 * vm_page_busy_downgrade: 638 * 639 * Downgrade an exclusive busy page into a single shared busy page. 640 */ 641void 642vm_page_busy_downgrade(vm_page_t m) 643{ 644 u_int x; 645 646 vm_page_assert_xbusied(m); 647 648 for (;;) { 649 x = m->busy_lock; 650 x &= VPB_BIT_WAITERS; 651 if (atomic_cmpset_rel_int(&m->busy_lock, 652 VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1) | x)) 653 break; 654 } 655} 656 657/* 658 * vm_page_sbusied: 659 * 660 * Return a positive value if the page is shared busied, 0 otherwise. 661 */ 662int 663vm_page_sbusied(vm_page_t m) 664{ 665 u_int x; 666 667 x = m->busy_lock; 668 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED); 669} 670 671/* 672 * vm_page_sunbusy: 673 * 674 * Shared unbusy a page. 675 */ 676void 677vm_page_sunbusy(vm_page_t m) 678{ 679 u_int x; 680 681 vm_page_assert_sbusied(m); 682 683 for (;;) { 684 x = m->busy_lock; 685 if (VPB_SHARERS(x) > 1) { 686 if (atomic_cmpset_int(&m->busy_lock, x, 687 x - VPB_ONE_SHARER)) 688 break; 689 continue; 690 } 691 if ((x & VPB_BIT_WAITERS) == 0) { 692 KASSERT(x == VPB_SHARERS_WORD(1), 693 ("vm_page_sunbusy: invalid lock state")); 694 if (atomic_cmpset_int(&m->busy_lock, 695 VPB_SHARERS_WORD(1), VPB_UNBUSIED)) 696 break; 697 continue; 698 } 699 KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS), 700 ("vm_page_sunbusy: invalid lock state for waiters")); 701 702 vm_page_lock(m); 703 if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) { 704 vm_page_unlock(m); 705 continue; 706 } 707 wakeup(m); 708 vm_page_unlock(m); 709 break; 710 } 711} 712 713/* 714 * vm_page_busy_sleep: 715 * 716 * Sleep and release the page lock, using the page pointer as wchan. 717 * This is used to implement the hard-path of busying mechanism. 718 * 719 * The given page must be locked. 720 */ 721void 722vm_page_busy_sleep(vm_page_t m, const char *wmesg) 723{ 724 u_int x; 725 726 vm_page_lock_assert(m, MA_OWNED); 727 728 x = m->busy_lock; 729 if (x == VPB_UNBUSIED) { 730 vm_page_unlock(m); 731 return; 732 } 733 if ((x & VPB_BIT_WAITERS) == 0 && 734 !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS)) { 735 vm_page_unlock(m); 736 return; 737 } 738 msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0); 739} 740 741/* 742 * vm_page_trysbusy: 743 * 744 * Try to shared busy a page. 745 * If the operation succeeds 1 is returned otherwise 0. 746 * The operation never sleeps. 747 */ 748int 749vm_page_trysbusy(vm_page_t m) 750{ 751 u_int x; 752 753 for (;;) { 754 x = m->busy_lock; 755 if ((x & VPB_BIT_SHARED) == 0) 756 return (0); 757 if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER)) 758 return (1); 759 } 760} 761 762static void 763vm_page_xunbusy_maybelocked(vm_page_t m) 764{ 765 bool lockacq; 766 767 vm_page_assert_xbusied(m); 768 769 lockacq = !mtx_owned(vm_page_lockptr(m)); 770 if (lockacq) 771 vm_page_lock(m); 772 vm_page_flash(m); 773 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); 774 if (lockacq) 775 vm_page_unlock(m); 776} 777 778/* 779 * vm_page_xunbusy_hard: 780 * 781 * Called after the first try the exclusive unbusy of a page failed. 782 * It is assumed that the waiters bit is on. 783 */ 784void 785vm_page_xunbusy_hard(vm_page_t m) 786{ 787 788 vm_page_assert_xbusied(m); 789 790 vm_page_lock(m); 791 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); 792 wakeup(m); 793 vm_page_unlock(m); 794} 795 796/* 797 * vm_page_flash: 798 * 799 * Wakeup anyone waiting for the page. 800 * The ownership bits do not change. 801 * 802 * The given page must be locked. 803 */ 804void 805vm_page_flash(vm_page_t m) 806{ 807 u_int x; 808 809 vm_page_lock_assert(m, MA_OWNED); 810 811 for (;;) { 812 x = m->busy_lock; 813 if ((x & VPB_BIT_WAITERS) == 0) 814 return; 815 if (atomic_cmpset_int(&m->busy_lock, x, 816 x & (~VPB_BIT_WAITERS))) 817 break; 818 } 819 wakeup(m); 820} 821 822/* 823 * Keep page from being freed by the page daemon 824 * much of the same effect as wiring, except much lower 825 * overhead and should be used only for *very* temporary 826 * holding ("wiring"). 827 */ 828void 829vm_page_hold(vm_page_t mem) 830{ 831 832 vm_page_lock_assert(mem, MA_OWNED); 833 mem->hold_count++; 834} 835 836void 837vm_page_unhold(vm_page_t mem) 838{ 839 840 vm_page_lock_assert(mem, MA_OWNED); 841 KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!")); 842 --mem->hold_count; 843 if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0) 844 vm_page_free_toq(mem); 845} 846 847/* 848 * vm_page_unhold_pages: 849 * 850 * Unhold each of the pages that is referenced by the given array. 851 */ 852void 853vm_page_unhold_pages(vm_page_t *ma, int count) 854{ 855 struct mtx *mtx, *new_mtx; 856 857 mtx = NULL; 858 for (; count != 0; count--) { 859 /* 860 * Avoid releasing and reacquiring the same page lock. 861 */ 862 new_mtx = vm_page_lockptr(*ma); 863 if (mtx != new_mtx) { 864 if (mtx != NULL) 865 mtx_unlock(mtx); 866 mtx = new_mtx; 867 mtx_lock(mtx); 868 } 869 vm_page_unhold(*ma); 870 ma++; 871 } 872 if (mtx != NULL) 873 mtx_unlock(mtx); 874} 875 876vm_page_t 877PHYS_TO_VM_PAGE(vm_paddr_t pa) 878{ 879 vm_page_t m; 880 881#ifdef VM_PHYSSEG_SPARSE 882 m = vm_phys_paddr_to_vm_page(pa); 883 if (m == NULL) 884 m = vm_phys_fictitious_to_vm_page(pa); 885 return (m); 886#elif defined(VM_PHYSSEG_DENSE) 887 long pi; 888 889 pi = atop(pa); 890 if (pi >= first_page && (pi - first_page) < vm_page_array_size) { 891 m = &vm_page_array[pi - first_page]; 892 return (m); 893 } 894 return (vm_phys_fictitious_to_vm_page(pa)); 895#else 896#error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 897#endif 898} 899 900/* 901 * vm_page_getfake: 902 * 903 * Create a fictitious page with the specified physical address and 904 * memory attribute. The memory attribute is the only the machine- 905 * dependent aspect of a fictitious page that must be initialized. 906 */ 907vm_page_t 908vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr) 909{ 910 vm_page_t m; 911 912 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO); 913 vm_page_initfake(m, paddr, memattr); 914 return (m); 915} 916 917void 918vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 919{ 920 921 if ((m->flags & PG_FICTITIOUS) != 0) { 922 /* 923 * The page's memattr might have changed since the 924 * previous initialization. Update the pmap to the 925 * new memattr. 926 */ 927 goto memattr; 928 } 929 m->phys_addr = paddr; 930 m->queue = PQ_NONE; 931 /* Fictitious pages don't use "segind". */ 932 m->flags = PG_FICTITIOUS; 933 /* Fictitious pages don't use "order" or "pool". */ 934 m->oflags = VPO_UNMANAGED; 935 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 936 m->wire_count = 1; 937 pmap_page_init(m); 938memattr: 939 pmap_page_set_memattr(m, memattr); 940} 941 942/* 943 * vm_page_putfake: 944 * 945 * Release a fictitious page. 946 */ 947void 948vm_page_putfake(vm_page_t m) 949{ 950 951 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m)); 952 KASSERT((m->flags & PG_FICTITIOUS) != 0, 953 ("vm_page_putfake: bad page %p", m)); 954 uma_zfree(fakepg_zone, m); 955} 956 957/* 958 * vm_page_updatefake: 959 * 960 * Update the given fictitious page to the specified physical address and 961 * memory attribute. 962 */ 963void 964vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 965{ 966 967 KASSERT((m->flags & PG_FICTITIOUS) != 0, 968 ("vm_page_updatefake: bad page %p", m)); 969 m->phys_addr = paddr; 970 pmap_page_set_memattr(m, memattr); 971} 972 973/* 974 * vm_page_free: 975 * 976 * Free a page. 977 */ 978void 979vm_page_free(vm_page_t m) 980{ 981 982 m->flags &= ~PG_ZERO; 983 vm_page_free_toq(m); 984} 985 986/* 987 * vm_page_free_zero: 988 * 989 * Free a page to the zerod-pages queue 990 */ 991void 992vm_page_free_zero(vm_page_t m) 993{ 994 995 m->flags |= PG_ZERO; 996 vm_page_free_toq(m); 997} 998 999/* 1000 * Unbusy and handle the page queueing for a page from the VOP_GETPAGES() 1001 * array which was optionally read ahead or behind. 1002 */ 1003void 1004vm_page_readahead_finish(vm_page_t m) 1005{ 1006 1007 /* We shouldn't put invalid pages on queues. */ 1008 KASSERT(m->valid != 0, ("%s: %p is invalid", __func__, m)); 1009 1010 /* 1011 * Since the page is not the actually needed one, whether it should 1012 * be activated or deactivated is not obvious. Empirical results 1013 * have shown that deactivating the page is usually the best choice, 1014 * unless the page is wanted by another thread. 1015 */ 1016 vm_page_lock(m); 1017 if ((m->busy_lock & VPB_BIT_WAITERS) != 0) 1018 vm_page_activate(m); 1019 else 1020 vm_page_deactivate(m); 1021 vm_page_unlock(m); 1022 vm_page_xunbusy(m); 1023} 1024 1025/* 1026 * vm_page_sleep_if_busy: 1027 * 1028 * Sleep and release the page queues lock if the page is busied. 1029 * Returns TRUE if the thread slept. 1030 * 1031 * The given page must be unlocked and object containing it must 1032 * be locked. 1033 */ 1034int 1035vm_page_sleep_if_busy(vm_page_t m, const char *msg) 1036{ 1037 vm_object_t obj; 1038 1039 vm_page_lock_assert(m, MA_NOTOWNED); 1040 VM_OBJECT_ASSERT_WLOCKED(m->object); 1041 1042 if (vm_page_busied(m)) { 1043 /* 1044 * The page-specific object must be cached because page 1045 * identity can change during the sleep, causing the 1046 * re-lock of a different object. 1047 * It is assumed that a reference to the object is already 1048 * held by the callers. 1049 */ 1050 obj = m->object; 1051 vm_page_lock(m); 1052 VM_OBJECT_WUNLOCK(obj); 1053 vm_page_busy_sleep(m, msg); 1054 VM_OBJECT_WLOCK(obj); 1055 return (TRUE); 1056 } 1057 return (FALSE); 1058} 1059 1060/* 1061 * vm_page_dirty_KBI: [ internal use only ] 1062 * 1063 * Set all bits in the page's dirty field. 1064 * 1065 * The object containing the specified page must be locked if the 1066 * call is made from the machine-independent layer. 1067 * 1068 * See vm_page_clear_dirty_mask(). 1069 * 1070 * This function should only be called by vm_page_dirty(). 1071 */ 1072void 1073vm_page_dirty_KBI(vm_page_t m) 1074{ 1075 1076 /* These assertions refer to this operation by its public name. */ 1077 KASSERT((m->flags & PG_CACHED) == 0, 1078 ("vm_page_dirty: page in cache!")); 1079 KASSERT(m->valid == VM_PAGE_BITS_ALL, 1080 ("vm_page_dirty: page is invalid!")); 1081 m->dirty = VM_PAGE_BITS_ALL; 1082} 1083 1084/* 1085 * vm_page_insert: [ internal use only ] 1086 * 1087 * Inserts the given mem entry into the object and object list. 1088 * 1089 * The object must be locked. 1090 */ 1091int 1092vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 1093{ 1094 vm_page_t mpred; 1095 1096 VM_OBJECT_ASSERT_WLOCKED(object); 1097 mpred = vm_radix_lookup_le(&object->rtree, pindex); 1098 return (vm_page_insert_after(m, object, pindex, mpred)); 1099} 1100 1101/* 1102 * vm_page_insert_after: 1103 * 1104 * Inserts the page "m" into the specified object at offset "pindex". 1105 * 1106 * The page "mpred" must immediately precede the offset "pindex" within 1107 * the specified object. 1108 * 1109 * The object must be locked. 1110 */ 1111static int 1112vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, 1113 vm_page_t mpred) 1114{ 1115 vm_page_t msucc; 1116 1117 VM_OBJECT_ASSERT_WLOCKED(object); 1118 KASSERT(m->object == NULL, 1119 ("vm_page_insert_after: page already inserted")); 1120 if (mpred != NULL) { 1121 KASSERT(mpred->object == object, 1122 ("vm_page_insert_after: object doesn't contain mpred")); 1123 KASSERT(mpred->pindex < pindex, 1124 ("vm_page_insert_after: mpred doesn't precede pindex")); 1125 msucc = TAILQ_NEXT(mpred, listq); 1126 } else 1127 msucc = TAILQ_FIRST(&object->memq); 1128 if (msucc != NULL) 1129 KASSERT(msucc->pindex > pindex, 1130 ("vm_page_insert_after: msucc doesn't succeed pindex")); 1131 1132 /* 1133 * Record the object/offset pair in this page 1134 */ 1135 m->object = object; 1136 m->pindex = pindex; 1137 1138 /* 1139 * Now link into the object's ordered list of backed pages. 1140 */ 1141 if (vm_radix_insert(&object->rtree, m)) { 1142 m->object = NULL; 1143 m->pindex = 0; 1144 return (1); 1145 } 1146 vm_page_insert_radixdone(m, object, mpred); 1147 return (0); 1148} 1149 1150/* 1151 * vm_page_insert_radixdone: 1152 * 1153 * Complete page "m" insertion into the specified object after the 1154 * radix trie hooking. 1155 * 1156 * The page "mpred" must precede the offset "m->pindex" within the 1157 * specified object. 1158 * 1159 * The object must be locked. 1160 */ 1161static void 1162vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred) 1163{ 1164 1165 VM_OBJECT_ASSERT_WLOCKED(object); 1166 KASSERT(object != NULL && m->object == object, 1167 ("vm_page_insert_radixdone: page %p has inconsistent object", m)); 1168 if (mpred != NULL) { 1169 KASSERT(mpred->object == object, 1170 ("vm_page_insert_after: object doesn't contain mpred")); 1171 KASSERT(mpred->pindex < m->pindex, 1172 ("vm_page_insert_after: mpred doesn't precede pindex")); 1173 } 1174 1175 if (mpred != NULL) 1176 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq); 1177 else 1178 TAILQ_INSERT_HEAD(&object->memq, m, listq); 1179 1180 /* 1181 * Show that the object has one more resident page. 1182 */ 1183 object->resident_page_count++; 1184 1185 /* 1186 * Hold the vnode until the last page is released. 1187 */ 1188 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 1189 vhold(object->handle); 1190 1191 /* 1192 * Since we are inserting a new and possibly dirty page, 1193 * update the object's OBJ_MIGHTBEDIRTY flag. 1194 */ 1195 if (pmap_page_is_write_mapped(m)) 1196 vm_object_set_writeable_dirty(object); 1197} 1198 1199/* 1200 * vm_page_remove: 1201 * 1202 * Removes the given mem entry from the object/offset-page 1203 * table and the object page list, but do not invalidate/terminate 1204 * the backing store. 1205 * 1206 * The object must be locked. The page must be locked if it is managed. 1207 */ 1208void 1209vm_page_remove(vm_page_t m) 1210{ 1211 vm_object_t object; 1212 1213 if ((m->oflags & VPO_UNMANAGED) == 0) 1214 vm_page_assert_locked(m); 1215 if ((object = m->object) == NULL) 1216 return; 1217 VM_OBJECT_ASSERT_WLOCKED(object); 1218 if (vm_page_xbusied(m)) 1219 vm_page_xunbusy_maybelocked(m); 1220 1221 /* 1222 * Now remove from the object's list of backed pages. 1223 */ 1224 vm_radix_remove(&object->rtree, m->pindex); 1225 TAILQ_REMOVE(&object->memq, m, listq); 1226 1227 /* 1228 * And show that the object has one fewer resident page. 1229 */ 1230 object->resident_page_count--; 1231 1232 /* 1233 * The vnode may now be recycled. 1234 */ 1235 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 1236 vdrop(object->handle); 1237 1238 m->object = NULL; 1239} 1240 1241/* 1242 * vm_page_lookup: 1243 * 1244 * Returns the page associated with the object/offset 1245 * pair specified; if none is found, NULL is returned. 1246 * 1247 * The object must be locked. 1248 */ 1249vm_page_t 1250vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1251{ 1252 1253 VM_OBJECT_ASSERT_LOCKED(object); 1254 return (vm_radix_lookup(&object->rtree, pindex)); 1255} 1256 1257/* 1258 * vm_page_find_least: 1259 * 1260 * Returns the page associated with the object with least pindex 1261 * greater than or equal to the parameter pindex, or NULL. 1262 * 1263 * The object must be locked. 1264 */ 1265vm_page_t 1266vm_page_find_least(vm_object_t object, vm_pindex_t pindex) 1267{ 1268 vm_page_t m; 1269 1270 VM_OBJECT_ASSERT_LOCKED(object); 1271 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex) 1272 m = vm_radix_lookup_ge(&object->rtree, pindex); 1273 return (m); 1274} 1275 1276/* 1277 * Returns the given page's successor (by pindex) within the object if it is 1278 * resident; if none is found, NULL is returned. 1279 * 1280 * The object must be locked. 1281 */ 1282vm_page_t 1283vm_page_next(vm_page_t m) 1284{ 1285 vm_page_t next; 1286 1287 VM_OBJECT_ASSERT_LOCKED(m->object); 1288 if ((next = TAILQ_NEXT(m, listq)) != NULL && 1289 next->pindex != m->pindex + 1) 1290 next = NULL; 1291 return (next); 1292} 1293 1294/* 1295 * Returns the given page's predecessor (by pindex) within the object if it is 1296 * resident; if none is found, NULL is returned. 1297 * 1298 * The object must be locked. 1299 */ 1300vm_page_t 1301vm_page_prev(vm_page_t m) 1302{ 1303 vm_page_t prev; 1304 1305 VM_OBJECT_ASSERT_LOCKED(m->object); 1306 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL && 1307 prev->pindex != m->pindex - 1) 1308 prev = NULL; 1309 return (prev); 1310} 1311 1312/* 1313 * Uses the page mnew as a replacement for an existing page at index 1314 * pindex which must be already present in the object. 1315 * 1316 * The existing page must not be on a paging queue. 1317 */ 1318vm_page_t 1319vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex) 1320{ 1321 vm_page_t mold; 1322 1323 VM_OBJECT_ASSERT_WLOCKED(object); 1324 KASSERT(mnew->object == NULL, 1325 ("vm_page_replace: page already in object")); 1326 1327 /* 1328 * This function mostly follows vm_page_insert() and 1329 * vm_page_remove() without the radix, object count and vnode 1330 * dance. Double check such functions for more comments. 1331 */ 1332 1333 mnew->object = object; 1334 mnew->pindex = pindex; 1335 mold = vm_radix_replace(&object->rtree, mnew); 1336 KASSERT(mold->queue == PQ_NONE, 1337 ("vm_page_replace: mold is on a paging queue")); 1338 1339 /* Keep the resident page list in sorted order. */ 1340 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq); 1341 TAILQ_REMOVE(&object->memq, mold, listq); 1342 1343 mold->object = NULL; 1344 vm_page_xunbusy_maybelocked(mold); 1345 1346 /* 1347 * The object's resident_page_count does not change because we have 1348 * swapped one page for another, but OBJ_MIGHTBEDIRTY. 1349 */ 1350 if (pmap_page_is_write_mapped(mnew)) 1351 vm_object_set_writeable_dirty(object); 1352 return (mold); 1353} 1354 1355/* 1356 * vm_page_rename: 1357 * 1358 * Move the given memory entry from its 1359 * current object to the specified target object/offset. 1360 * 1361 * Note: swap associated with the page must be invalidated by the move. We 1362 * have to do this for several reasons: (1) we aren't freeing the 1363 * page, (2) we are dirtying the page, (3) the VM system is probably 1364 * moving the page from object A to B, and will then later move 1365 * the backing store from A to B and we can't have a conflict. 1366 * 1367 * Note: we *always* dirty the page. It is necessary both for the 1368 * fact that we moved it, and because we may be invalidating 1369 * swap. If the page is on the cache, we have to deactivate it 1370 * or vm_page_dirty() will panic. Dirty pages are not allowed 1371 * on the cache. 1372 * 1373 * The objects must be locked. 1374 */ 1375int 1376vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1377{ 1378 vm_page_t mpred; 1379 vm_pindex_t opidx; 1380 1381 VM_OBJECT_ASSERT_WLOCKED(new_object); 1382 1383 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex); 1384 KASSERT(mpred == NULL || mpred->pindex != new_pindex, 1385 ("vm_page_rename: pindex already renamed")); 1386 1387 /* 1388 * Create a custom version of vm_page_insert() which does not depend 1389 * by m_prev and can cheat on the implementation aspects of the 1390 * function. 1391 */ 1392 opidx = m->pindex; 1393 m->pindex = new_pindex; 1394 if (vm_radix_insert(&new_object->rtree, m)) { 1395 m->pindex = opidx; 1396 return (1); 1397 } 1398 1399 /* 1400 * The operation cannot fail anymore. The removal must happen before 1401 * the listq iterator is tainted. 1402 */ 1403 m->pindex = opidx; 1404 vm_page_lock(m); 1405 vm_page_remove(m); 1406 1407 /* Return back to the new pindex to complete vm_page_insert(). */ 1408 m->pindex = new_pindex; 1409 m->object = new_object; 1410 vm_page_unlock(m); 1411 vm_page_insert_radixdone(m, new_object, mpred); 1412 vm_page_dirty(m); 1413 return (0); 1414} 1415 1416/* 1417 * Convert all of the given object's cached pages that have a 1418 * pindex within the given range into free pages. If the value 1419 * zero is given for "end", then the range's upper bound is 1420 * infinity. If the given object is backed by a vnode and it 1421 * transitions from having one or more cached pages to none, the 1422 * vnode's hold count is reduced. 1423 */ 1424void 1425vm_page_cache_free(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 1426{ 1427 vm_page_t m; 1428 boolean_t empty; 1429 1430 mtx_lock(&vm_page_queue_free_mtx); 1431 if (__predict_false(vm_radix_is_empty(&object->cache))) { 1432 mtx_unlock(&vm_page_queue_free_mtx); 1433 return; 1434 } 1435 while ((m = vm_radix_lookup_ge(&object->cache, start)) != NULL) { 1436 if (end != 0 && m->pindex >= end) 1437 break; 1438 vm_radix_remove(&object->cache, m->pindex); 1439 vm_page_cache_turn_free(m); 1440 } 1441 empty = vm_radix_is_empty(&object->cache); 1442 mtx_unlock(&vm_page_queue_free_mtx); 1443 if (object->type == OBJT_VNODE && empty) 1444 vdrop(object->handle); 1445} 1446 1447/* 1448 * Returns the cached page that is associated with the given 1449 * object and offset. If, however, none exists, returns NULL. 1450 * 1451 * The free page queue must be locked. 1452 */ 1453static inline vm_page_t 1454vm_page_cache_lookup(vm_object_t object, vm_pindex_t pindex) 1455{ 1456 1457 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1458 return (vm_radix_lookup(&object->cache, pindex)); 1459} 1460 1461/* 1462 * Remove the given cached page from its containing object's 1463 * collection of cached pages. 1464 * 1465 * The free page queue must be locked. 1466 */ 1467static void 1468vm_page_cache_remove(vm_page_t m) 1469{ 1470 1471 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1472 KASSERT((m->flags & PG_CACHED) != 0, 1473 ("vm_page_cache_remove: page %p is not cached", m)); 1474 vm_radix_remove(&m->object->cache, m->pindex); 1475 m->object = NULL; 1476 vm_cnt.v_cache_count--; 1477} 1478 1479/* 1480 * Transfer all of the cached pages with offset greater than or 1481 * equal to 'offidxstart' from the original object's cache to the 1482 * new object's cache. However, any cached pages with offset 1483 * greater than or equal to the new object's size are kept in the 1484 * original object. Initially, the new object's cache must be 1485 * empty. Offset 'offidxstart' in the original object must 1486 * correspond to offset zero in the new object. 1487 * 1488 * The new object must be locked. 1489 */ 1490void 1491vm_page_cache_transfer(vm_object_t orig_object, vm_pindex_t offidxstart, 1492 vm_object_t new_object) 1493{ 1494 vm_page_t m; 1495 1496 /* 1497 * Insertion into an object's collection of cached pages 1498 * requires the object to be locked. In contrast, removal does 1499 * not. 1500 */ 1501 VM_OBJECT_ASSERT_WLOCKED(new_object); 1502 KASSERT(vm_radix_is_empty(&new_object->cache), 1503 ("vm_page_cache_transfer: object %p has cached pages", 1504 new_object)); 1505 mtx_lock(&vm_page_queue_free_mtx); 1506 while ((m = vm_radix_lookup_ge(&orig_object->cache, 1507 offidxstart)) != NULL) { 1508 /* 1509 * Transfer all of the pages with offset greater than or 1510 * equal to 'offidxstart' from the original object's 1511 * cache to the new object's cache. 1512 */ 1513 if ((m->pindex - offidxstart) >= new_object->size) 1514 break; 1515 vm_radix_remove(&orig_object->cache, m->pindex); 1516 /* Update the page's object and offset. */ 1517 m->object = new_object; 1518 m->pindex -= offidxstart; 1519 if (vm_radix_insert(&new_object->cache, m)) 1520 vm_page_cache_turn_free(m); 1521 } 1522 mtx_unlock(&vm_page_queue_free_mtx); 1523} 1524 1525/* 1526 * Returns TRUE if a cached page is associated with the given object and 1527 * offset, and FALSE otherwise. 1528 * 1529 * The object must be locked. 1530 */ 1531boolean_t 1532vm_page_is_cached(vm_object_t object, vm_pindex_t pindex) 1533{ 1534 vm_page_t m; 1535 1536 /* 1537 * Insertion into an object's collection of cached pages requires the 1538 * object to be locked. Therefore, if the object is locked and the 1539 * object's collection is empty, there is no need to acquire the free 1540 * page queues lock in order to prove that the specified page doesn't 1541 * exist. 1542 */ 1543 VM_OBJECT_ASSERT_WLOCKED(object); 1544 if (__predict_true(vm_object_cache_is_empty(object))) 1545 return (FALSE); 1546 mtx_lock(&vm_page_queue_free_mtx); 1547 m = vm_page_cache_lookup(object, pindex); 1548 mtx_unlock(&vm_page_queue_free_mtx); 1549 return (m != NULL); 1550} 1551 1552/* 1553 * vm_page_alloc: 1554 * 1555 * Allocate and return a page that is associated with the specified 1556 * object and offset pair. By default, this page is exclusive busied. 1557 * 1558 * The caller must always specify an allocation class. 1559 * 1560 * allocation classes: 1561 * VM_ALLOC_NORMAL normal process request 1562 * VM_ALLOC_SYSTEM system *really* needs a page 1563 * VM_ALLOC_INTERRUPT interrupt time request 1564 * 1565 * optional allocation flags: 1566 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1567 * intends to allocate 1568 * VM_ALLOC_IFCACHED return page only if it is cached 1569 * VM_ALLOC_IFNOTCACHED return NULL, do not reactivate if the page 1570 * is cached 1571 * VM_ALLOC_NOBUSY do not exclusive busy the page 1572 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1573 * VM_ALLOC_NOOBJ page is not associated with an object and 1574 * should not be exclusive busy 1575 * VM_ALLOC_SBUSY shared busy the allocated page 1576 * VM_ALLOC_WIRED wire the allocated page 1577 * VM_ALLOC_ZERO prefer a zeroed page 1578 * 1579 * This routine may not sleep. 1580 */ 1581vm_page_t 1582vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 1583{ 1584 struct vnode *vp = NULL; 1585 vm_object_t m_object; 1586 vm_page_t m, mpred; 1587 int flags, req_class; 1588 1589 mpred = 0; /* XXX: pacify gcc */ 1590 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 1591 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 1592 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 1593 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 1594 ("vm_page_alloc: inconsistent object(%p)/req(%x)", (void *)object, 1595 req)); 1596 if (object != NULL) 1597 VM_OBJECT_ASSERT_WLOCKED(object); 1598 1599 req_class = req & VM_ALLOC_CLASS_MASK; 1600 1601 /* 1602 * The page daemon is allowed to dig deeper into the free page list. 1603 */ 1604 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1605 req_class = VM_ALLOC_SYSTEM; 1606 1607 if (object != NULL) { 1608 mpred = vm_radix_lookup_le(&object->rtree, pindex); 1609 KASSERT(mpred == NULL || mpred->pindex != pindex, 1610 ("vm_page_alloc: pindex already allocated")); 1611 } 1612 1613 /* 1614 * The page allocation request can came from consumers which already 1615 * hold the free page queue mutex, like vm_page_insert() in 1616 * vm_page_cache(). 1617 */ 1618 mtx_lock_flags(&vm_page_queue_free_mtx, MTX_RECURSE); 1619 if (vm_cnt.v_free_count + vm_cnt.v_cache_count > vm_cnt.v_free_reserved || 1620 (req_class == VM_ALLOC_SYSTEM && 1621 vm_cnt.v_free_count + vm_cnt.v_cache_count > vm_cnt.v_interrupt_free_min) || 1622 (req_class == VM_ALLOC_INTERRUPT && 1623 vm_cnt.v_free_count + vm_cnt.v_cache_count > 0)) { 1624 /* 1625 * Allocate from the free queue if the number of free pages 1626 * exceeds the minimum for the request class. 1627 */ 1628 if (object != NULL && 1629 (m = vm_page_cache_lookup(object, pindex)) != NULL) { 1630 if ((req & VM_ALLOC_IFNOTCACHED) != 0) { 1631 mtx_unlock(&vm_page_queue_free_mtx); 1632 return (NULL); 1633 } 1634 if (vm_phys_unfree_page(m)) 1635 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0); 1636#if VM_NRESERVLEVEL > 0 1637 else if (!vm_reserv_reactivate_page(m)) 1638#else 1639 else 1640#endif 1641 panic("vm_page_alloc: cache page %p is missing" 1642 " from the free queue", m); 1643 } else if ((req & VM_ALLOC_IFCACHED) != 0) { 1644 mtx_unlock(&vm_page_queue_free_mtx); 1645 return (NULL); 1646#if VM_NRESERVLEVEL > 0 1647 } else if (object == NULL || (object->flags & (OBJ_COLORED | 1648 OBJ_FICTITIOUS)) != OBJ_COLORED || (m = 1649 vm_reserv_alloc_page(object, pindex, mpred)) == NULL) { 1650#else 1651 } else { 1652#endif 1653 m = vm_phys_alloc_pages(object != NULL ? 1654 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0); 1655#if VM_NRESERVLEVEL > 0 1656 if (m == NULL && vm_reserv_reclaim_inactive()) { 1657 m = vm_phys_alloc_pages(object != NULL ? 1658 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 1659 0); 1660 } 1661#endif 1662 } 1663 } else { 1664 /* 1665 * Not allocatable, give up. 1666 */ 1667 mtx_unlock(&vm_page_queue_free_mtx); 1668 atomic_add_int(&vm_pageout_deficit, 1669 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 1670 pagedaemon_wakeup(); 1671 return (NULL); 1672 } 1673 1674 /* 1675 * At this point we had better have found a good page. 1676 */ 1677 KASSERT(m != NULL, ("vm_page_alloc: missing page")); 1678 KASSERT(m->queue == PQ_NONE, 1679 ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue)); 1680 KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m)); 1681 KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m)); 1682 KASSERT(!vm_page_sbusied(m), 1683 ("vm_page_alloc: page %p is busy", m)); 1684 KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m)); 1685 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 1686 ("vm_page_alloc: page %p has unexpected memattr %d", m, 1687 pmap_page_get_memattr(m))); 1688 if ((m->flags & PG_CACHED) != 0) { 1689 KASSERT((m->flags & PG_ZERO) == 0, 1690 ("vm_page_alloc: cached page %p is PG_ZERO", m)); 1691 KASSERT(m->valid != 0, 1692 ("vm_page_alloc: cached page %p is invalid", m)); 1693 if (m->object == object && m->pindex == pindex) 1694 vm_cnt.v_reactivated++; 1695 else 1696 m->valid = 0; 1697 m_object = m->object; 1698 vm_page_cache_remove(m); 1699 if (m_object->type == OBJT_VNODE && 1700 vm_object_cache_is_empty(m_object)) 1701 vp = m_object->handle; 1702 } else { 1703 KASSERT(m->valid == 0, 1704 ("vm_page_alloc: free page %p is valid", m)); 1705 vm_phys_freecnt_adj(m, -1); 1706 if ((m->flags & PG_ZERO) != 0) 1707 vm_page_zero_count--; 1708 } 1709 mtx_unlock(&vm_page_queue_free_mtx); 1710 1711 /* 1712 * Initialize the page. Only the PG_ZERO flag is inherited. 1713 */ 1714 flags = 0; 1715 if ((req & VM_ALLOC_ZERO) != 0) 1716 flags = PG_ZERO; 1717 flags &= m->flags; 1718 if ((req & VM_ALLOC_NODUMP) != 0) 1719 flags |= PG_NODUMP; 1720 m->flags = flags; 1721 m->aflags = 0; 1722 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? 1723 VPO_UNMANAGED : 0; 1724 m->busy_lock = VPB_UNBUSIED; 1725 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) 1726 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 1727 if ((req & VM_ALLOC_SBUSY) != 0) 1728 m->busy_lock = VPB_SHARERS_WORD(1); 1729 if (req & VM_ALLOC_WIRED) { 1730 /* 1731 * The page lock is not required for wiring a page until that 1732 * page is inserted into the object. 1733 */ 1734 atomic_add_int(&vm_cnt.v_wire_count, 1); 1735 m->wire_count = 1; 1736 } 1737 m->act_count = 0; 1738 1739 if (object != NULL) { 1740 if (vm_page_insert_after(m, object, pindex, mpred)) { 1741 /* See the comment below about hold count. */ 1742 if (vp != NULL) 1743 vdrop(vp); 1744 pagedaemon_wakeup(); 1745 if (req & VM_ALLOC_WIRED) { 1746 atomic_subtract_int(&vm_cnt.v_wire_count, 1); 1747 m->wire_count = 0; 1748 } 1749 m->object = NULL; 1750 m->oflags = VPO_UNMANAGED; 1751 vm_page_free(m); 1752 return (NULL); 1753 } 1754 1755 /* Ignore device objects; the pager sets "memattr" for them. */ 1756 if (object->memattr != VM_MEMATTR_DEFAULT && 1757 (object->flags & OBJ_FICTITIOUS) == 0) 1758 pmap_page_set_memattr(m, object->memattr); 1759 } else 1760 m->pindex = pindex; 1761 1762 /* 1763 * The following call to vdrop() must come after the above call 1764 * to vm_page_insert() in case both affect the same object and 1765 * vnode. Otherwise, the affected vnode's hold count could 1766 * temporarily become zero. 1767 */ 1768 if (vp != NULL) 1769 vdrop(vp); 1770 1771 /* 1772 * Don't wakeup too often - wakeup the pageout daemon when 1773 * we would be nearly out of memory. 1774 */ 1775 if (vm_paging_needed()) 1776 pagedaemon_wakeup(); 1777 1778 return (m); 1779} 1780 1781static void 1782vm_page_alloc_contig_vdrop(struct spglist *lst) 1783{ 1784 1785 while (!SLIST_EMPTY(lst)) { 1786 vdrop((struct vnode *)SLIST_FIRST(lst)-> plinks.s.pv); 1787 SLIST_REMOVE_HEAD(lst, plinks.s.ss); 1788 } 1789} 1790 1791/* 1792 * vm_page_alloc_contig: 1793 * 1794 * Allocate a contiguous set of physical pages of the given size "npages" 1795 * from the free lists. All of the physical pages must be at or above 1796 * the given physical address "low" and below the given physical address 1797 * "high". The given value "alignment" determines the alignment of the 1798 * first physical page in the set. If the given value "boundary" is 1799 * non-zero, then the set of physical pages cannot cross any physical 1800 * address boundary that is a multiple of that value. Both "alignment" 1801 * and "boundary" must be a power of two. 1802 * 1803 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, 1804 * then the memory attribute setting for the physical pages is configured 1805 * to the object's memory attribute setting. Otherwise, the memory 1806 * attribute setting for the physical pages is configured to "memattr", 1807 * overriding the object's memory attribute setting. However, if the 1808 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the 1809 * memory attribute setting for the physical pages cannot be configured 1810 * to VM_MEMATTR_DEFAULT. 1811 * 1812 * The caller must always specify an allocation class. 1813 * 1814 * allocation classes: 1815 * VM_ALLOC_NORMAL normal process request 1816 * VM_ALLOC_SYSTEM system *really* needs a page 1817 * VM_ALLOC_INTERRUPT interrupt time request 1818 * 1819 * optional allocation flags: 1820 * VM_ALLOC_NOBUSY do not exclusive busy the page 1821 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1822 * VM_ALLOC_NOOBJ page is not associated with an object and 1823 * should not be exclusive busy 1824 * VM_ALLOC_SBUSY shared busy the allocated page 1825 * VM_ALLOC_WIRED wire the allocated page 1826 * VM_ALLOC_ZERO prefer a zeroed page 1827 * 1828 * This routine may not sleep. 1829 */ 1830vm_page_t 1831vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, 1832 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 1833 vm_paddr_t boundary, vm_memattr_t memattr) 1834{ 1835 struct vnode *drop; 1836 struct spglist deferred_vdrop_list; 1837 vm_page_t m, m_tmp, m_ret; 1838 u_int flags; 1839 int req_class; 1840 1841 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 1842 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 1843 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 1844 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 1845 ("vm_page_alloc: inconsistent object(%p)/req(%x)", (void *)object, 1846 req)); 1847 if (object != NULL) { 1848 VM_OBJECT_ASSERT_WLOCKED(object); 1849 KASSERT(object->type == OBJT_PHYS, 1850 ("vm_page_alloc_contig: object %p isn't OBJT_PHYS", 1851 object)); 1852 } 1853 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 1854 req_class = req & VM_ALLOC_CLASS_MASK; 1855 1856 /* 1857 * The page daemon is allowed to dig deeper into the free page list. 1858 */ 1859 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1860 req_class = VM_ALLOC_SYSTEM; 1861 1862 SLIST_INIT(&deferred_vdrop_list); 1863 mtx_lock(&vm_page_queue_free_mtx); 1864 if (vm_cnt.v_free_count + vm_cnt.v_cache_count >= npages + 1865 vm_cnt.v_free_reserved || (req_class == VM_ALLOC_SYSTEM && 1866 vm_cnt.v_free_count + vm_cnt.v_cache_count >= npages + 1867 vm_cnt.v_interrupt_free_min) || (req_class == VM_ALLOC_INTERRUPT && 1868 vm_cnt.v_free_count + vm_cnt.v_cache_count >= npages)) { 1869#if VM_NRESERVLEVEL > 0 1870retry: 1871 if (object == NULL || (object->flags & OBJ_COLORED) == 0 || 1872 (m_ret = vm_reserv_alloc_contig(object, pindex, npages, 1873 low, high, alignment, boundary)) == NULL) 1874#endif 1875 m_ret = vm_phys_alloc_contig(npages, low, high, 1876 alignment, boundary); 1877 } else { 1878 mtx_unlock(&vm_page_queue_free_mtx); 1879 atomic_add_int(&vm_pageout_deficit, npages); 1880 pagedaemon_wakeup(); 1881 return (NULL); 1882 } 1883 if (m_ret != NULL) 1884 for (m = m_ret; m < &m_ret[npages]; m++) { 1885 drop = vm_page_alloc_init(m); 1886 if (drop != NULL) { 1887 /* 1888 * Enqueue the vnode for deferred vdrop(). 1889 */ 1890 m->plinks.s.pv = drop; 1891 SLIST_INSERT_HEAD(&deferred_vdrop_list, m, 1892 plinks.s.ss); 1893 } 1894 } 1895 else { 1896#if VM_NRESERVLEVEL > 0 1897 if (vm_reserv_reclaim_contig(npages, low, high, alignment, 1898 boundary)) 1899 goto retry; 1900#endif 1901 } 1902 mtx_unlock(&vm_page_queue_free_mtx); 1903 if (m_ret == NULL) 1904 return (NULL); 1905 1906 /* 1907 * Initialize the pages. Only the PG_ZERO flag is inherited. 1908 */ 1909 flags = 0; 1910 if ((req & VM_ALLOC_ZERO) != 0) 1911 flags = PG_ZERO; 1912 if ((req & VM_ALLOC_NODUMP) != 0) 1913 flags |= PG_NODUMP; 1914 if ((req & VM_ALLOC_WIRED) != 0) 1915 atomic_add_int(&vm_cnt.v_wire_count, npages); 1916 if (object != NULL) { 1917 if (object->memattr != VM_MEMATTR_DEFAULT && 1918 memattr == VM_MEMATTR_DEFAULT) 1919 memattr = object->memattr; 1920 } 1921 for (m = m_ret; m < &m_ret[npages]; m++) { 1922 m->aflags = 0; 1923 m->flags = (m->flags | PG_NODUMP) & flags; 1924 m->busy_lock = VPB_UNBUSIED; 1925 if (object != NULL) { 1926 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 1927 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 1928 if ((req & VM_ALLOC_SBUSY) != 0) 1929 m->busy_lock = VPB_SHARERS_WORD(1); 1930 } 1931 if ((req & VM_ALLOC_WIRED) != 0) 1932 m->wire_count = 1; 1933 /* Unmanaged pages don't use "act_count". */ 1934 m->oflags = VPO_UNMANAGED; 1935 if (object != NULL) { 1936 if (vm_page_insert(m, object, pindex)) { 1937 vm_page_alloc_contig_vdrop( 1938 &deferred_vdrop_list); 1939 if (vm_paging_needed()) 1940 pagedaemon_wakeup(); 1941 if ((req & VM_ALLOC_WIRED) != 0) 1942 atomic_subtract_int(&vm_cnt.v_wire_count, 1943 npages); 1944 for (m_tmp = m, m = m_ret; 1945 m < &m_ret[npages]; m++) { 1946 if ((req & VM_ALLOC_WIRED) != 0) 1947 m->wire_count = 0; 1948 if (m >= m_tmp) { 1949 m->object = NULL; 1950 m->oflags |= VPO_UNMANAGED; 1951 } 1952 vm_page_free(m); 1953 } 1954 return (NULL); 1955 } 1956 } else 1957 m->pindex = pindex; 1958 if (memattr != VM_MEMATTR_DEFAULT) 1959 pmap_page_set_memattr(m, memattr); 1960 pindex++; 1961 } 1962 vm_page_alloc_contig_vdrop(&deferred_vdrop_list); 1963 if (vm_paging_needed()) 1964 pagedaemon_wakeup(); 1965 return (m_ret); 1966} 1967 1968/* 1969 * Initialize a page that has been freshly dequeued from a freelist. 1970 * The caller has to drop the vnode returned, if it is not NULL. 1971 * 1972 * This function may only be used to initialize unmanaged pages. 1973 * 1974 * To be called with vm_page_queue_free_mtx held. 1975 */ 1976static struct vnode * 1977vm_page_alloc_init(vm_page_t m) 1978{ 1979 struct vnode *drop; 1980 vm_object_t m_object; 1981 1982 KASSERT(m->queue == PQ_NONE, 1983 ("vm_page_alloc_init: page %p has unexpected queue %d", 1984 m, m->queue)); 1985 KASSERT(m->wire_count == 0, 1986 ("vm_page_alloc_init: page %p is wired", m)); 1987 KASSERT(m->hold_count == 0, 1988 ("vm_page_alloc_init: page %p is held", m)); 1989 KASSERT(!vm_page_sbusied(m), 1990 ("vm_page_alloc_init: page %p is busy", m)); 1991 KASSERT(m->dirty == 0, 1992 ("vm_page_alloc_init: page %p is dirty", m)); 1993 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 1994 ("vm_page_alloc_init: page %p has unexpected memattr %d", 1995 m, pmap_page_get_memattr(m))); 1996 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1997 drop = NULL; 1998 if ((m->flags & PG_CACHED) != 0) { 1999 KASSERT((m->flags & PG_ZERO) == 0, 2000 ("vm_page_alloc_init: cached page %p is PG_ZERO", m)); 2001 m->valid = 0; 2002 m_object = m->object; 2003 vm_page_cache_remove(m); 2004 if (m_object->type == OBJT_VNODE && 2005 vm_object_cache_is_empty(m_object)) 2006 drop = m_object->handle; 2007 } else { 2008 KASSERT(m->valid == 0, 2009 ("vm_page_alloc_init: free page %p is valid", m)); 2010 vm_phys_freecnt_adj(m, -1); 2011 if ((m->flags & PG_ZERO) != 0) 2012 vm_page_zero_count--; 2013 } 2014 return (drop); 2015} 2016 2017/* 2018 * vm_page_alloc_freelist: 2019 * 2020 * Allocate a physical page from the specified free page list. 2021 * 2022 * The caller must always specify an allocation class. 2023 * 2024 * allocation classes: 2025 * VM_ALLOC_NORMAL normal process request 2026 * VM_ALLOC_SYSTEM system *really* needs a page 2027 * VM_ALLOC_INTERRUPT interrupt time request 2028 * 2029 * optional allocation flags: 2030 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 2031 * intends to allocate 2032 * VM_ALLOC_WIRED wire the allocated page 2033 * VM_ALLOC_ZERO prefer a zeroed page 2034 * 2035 * This routine may not sleep. 2036 */ 2037vm_page_t 2038vm_page_alloc_freelist(int flind, int req) 2039{ 2040 struct vnode *drop; 2041 vm_page_t m; 2042 u_int flags; 2043 int req_class; 2044 2045 req_class = req & VM_ALLOC_CLASS_MASK; 2046 2047 /* 2048 * The page daemon is allowed to dig deeper into the free page list. 2049 */ 2050 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 2051 req_class = VM_ALLOC_SYSTEM; 2052 2053 /* 2054 * Do not allocate reserved pages unless the req has asked for it. 2055 */ 2056 mtx_lock_flags(&vm_page_queue_free_mtx, MTX_RECURSE); 2057 if (vm_cnt.v_free_count + vm_cnt.v_cache_count > vm_cnt.v_free_reserved || 2058 (req_class == VM_ALLOC_SYSTEM && 2059 vm_cnt.v_free_count + vm_cnt.v_cache_count > vm_cnt.v_interrupt_free_min) || 2060 (req_class == VM_ALLOC_INTERRUPT && 2061 vm_cnt.v_free_count + vm_cnt.v_cache_count > 0)) 2062 m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0); 2063 else { 2064 mtx_unlock(&vm_page_queue_free_mtx); 2065 atomic_add_int(&vm_pageout_deficit, 2066 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 2067 pagedaemon_wakeup(); 2068 return (NULL); 2069 } 2070 if (m == NULL) { 2071 mtx_unlock(&vm_page_queue_free_mtx); 2072 return (NULL); 2073 } 2074 drop = vm_page_alloc_init(m); 2075 mtx_unlock(&vm_page_queue_free_mtx); 2076 2077 /* 2078 * Initialize the page. Only the PG_ZERO flag is inherited. 2079 */ 2080 m->aflags = 0; 2081 flags = 0; 2082 if ((req & VM_ALLOC_ZERO) != 0) 2083 flags = PG_ZERO; 2084 m->flags &= flags; 2085 if ((req & VM_ALLOC_WIRED) != 0) { 2086 /* 2087 * The page lock is not required for wiring a page that does 2088 * not belong to an object. 2089 */ 2090 atomic_add_int(&vm_cnt.v_wire_count, 1); 2091 m->wire_count = 1; 2092 } 2093 /* Unmanaged pages don't use "act_count". */ 2094 m->oflags = VPO_UNMANAGED; 2095 if (drop != NULL) 2096 vdrop(drop); 2097 if (vm_paging_needed()) 2098 pagedaemon_wakeup(); 2099 return (m); 2100} 2101 2102#define VPSC_ANY 0 /* No restrictions. */ 2103#define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */ 2104#define VPSC_NOSUPER 2 /* Skip superpages. */ 2105 2106/* 2107 * vm_page_scan_contig: 2108 * 2109 * Scan vm_page_array[] between the specified entries "m_start" and 2110 * "m_end" for a run of contiguous physical pages that satisfy the 2111 * specified conditions, and return the lowest page in the run. The 2112 * specified "alignment" determines the alignment of the lowest physical 2113 * page in the run. If the specified "boundary" is non-zero, then the 2114 * run of physical pages cannot span a physical address that is a 2115 * multiple of "boundary". 2116 * 2117 * "m_end" is never dereferenced, so it need not point to a vm_page 2118 * structure within vm_page_array[]. 2119 * 2120 * "npages" must be greater than zero. "m_start" and "m_end" must not 2121 * span a hole (or discontiguity) in the physical address space. Both 2122 * "alignment" and "boundary" must be a power of two. 2123 */ 2124vm_page_t 2125vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end, 2126 u_long alignment, vm_paddr_t boundary, int options) 2127{ 2128 struct mtx *m_mtx, *new_mtx; 2129 vm_object_t object; 2130 vm_paddr_t pa; 2131 vm_page_t m, m_run; 2132#if VM_NRESERVLEVEL > 0 2133 int level; 2134#endif 2135 int m_inc, order, run_ext, run_len; 2136 2137 KASSERT(npages > 0, ("npages is 0")); 2138 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2139 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2140 m_run = NULL; 2141 run_len = 0; 2142 m_mtx = NULL; 2143 for (m = m_start; m < m_end && run_len < npages; m += m_inc) { 2144 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0, 2145 ("page %p is PG_FICTITIOUS or PG_MARKER", m)); 2146 2147 /* 2148 * If the current page would be the start of a run, check its 2149 * physical address against the end, alignment, and boundary 2150 * conditions. If it doesn't satisfy these conditions, either 2151 * terminate the scan or advance to the next page that 2152 * satisfies the failed condition. 2153 */ 2154 if (run_len == 0) { 2155 KASSERT(m_run == NULL, ("m_run != NULL")); 2156 if (m + npages > m_end) 2157 break; 2158 pa = VM_PAGE_TO_PHYS(m); 2159 if ((pa & (alignment - 1)) != 0) { 2160 m_inc = atop(roundup2(pa, alignment) - pa); 2161 continue; 2162 } 2163 if (rounddown2(pa ^ (pa + ptoa(npages) - 1), 2164 boundary) != 0) { 2165 m_inc = atop(roundup2(pa, boundary) - pa); 2166 continue; 2167 } 2168 } else 2169 KASSERT(m_run != NULL, ("m_run == NULL")); 2170 2171 /* 2172 * Avoid releasing and reacquiring the same page lock. 2173 */ 2174 new_mtx = vm_page_lockptr(m); 2175 if (m_mtx != new_mtx) { 2176 if (m_mtx != NULL) 2177 mtx_unlock(m_mtx); 2178 m_mtx = new_mtx; 2179 mtx_lock(m_mtx); 2180 } 2181 m_inc = 1; 2182retry: 2183 if (m->wire_count != 0 || m->hold_count != 0) 2184 run_ext = 0; 2185#if VM_NRESERVLEVEL > 0 2186 else if ((level = vm_reserv_level(m)) >= 0 && 2187 (options & VPSC_NORESERV) != 0) { 2188 run_ext = 0; 2189 /* Advance to the end of the reservation. */ 2190 pa = VM_PAGE_TO_PHYS(m); 2191 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - 2192 pa); 2193 } 2194#endif 2195 else if ((object = m->object) != NULL) { 2196 /* 2197 * The page is considered eligible for relocation if 2198 * and only if it could be laundered or reclaimed by 2199 * the page daemon. 2200 */ 2201 if (!VM_OBJECT_TRYRLOCK(object)) { 2202 mtx_unlock(m_mtx); 2203 VM_OBJECT_RLOCK(object); 2204 mtx_lock(m_mtx); 2205 if (m->object != object) { 2206 /* 2207 * The page may have been freed. 2208 */ 2209 VM_OBJECT_RUNLOCK(object); 2210 goto retry; 2211 } else if (m->wire_count != 0 || 2212 m->hold_count != 0) { 2213 run_ext = 0; 2214 goto unlock; 2215 } 2216 } 2217 KASSERT((m->flags & PG_UNHOLDFREE) == 0, 2218 ("page %p is PG_UNHOLDFREE", m)); 2219 /* Don't care: PG_NODUMP, PG_WINATCFLS, PG_ZERO. */ 2220 if (object->type != OBJT_DEFAULT && 2221 object->type != OBJT_SWAP && 2222 object->type != OBJT_VNODE) 2223 run_ext = 0; 2224 else if ((m->flags & PG_CACHED) != 0 || 2225 m != vm_page_lookup(object, m->pindex)) { 2226 /* 2227 * The page is cached or recently converted 2228 * from cached to free. 2229 */ 2230#if VM_NRESERVLEVEL > 0 2231 if (level >= 0) { 2232 /* 2233 * The page is reserved. Extend the 2234 * current run by one page. 2235 */ 2236 run_ext = 1; 2237 } else 2238#endif 2239 if ((order = m->order) < VM_NFREEORDER) { 2240 /* 2241 * The page is enqueued in the 2242 * physical memory allocator's cache/ 2243 * free page queues. Moreover, it is 2244 * the first page in a power-of-two- 2245 * sized run of contiguous cache/free 2246 * pages. Add these pages to the end 2247 * of the current run, and jump 2248 * ahead. 2249 */ 2250 run_ext = 1 << order; 2251 m_inc = 1 << order; 2252 } else 2253 run_ext = 0; 2254#if VM_NRESERVLEVEL > 0 2255 } else if ((options & VPSC_NOSUPER) != 0 && 2256 (level = vm_reserv_level_iffullpop(m)) >= 0) { 2257 run_ext = 0; 2258 /* Advance to the end of the superpage. */ 2259 pa = VM_PAGE_TO_PHYS(m); 2260 m_inc = atop(roundup2(pa + 1, 2261 vm_reserv_size(level)) - pa); 2262#endif 2263 } else if (object->memattr == VM_MEMATTR_DEFAULT && 2264 m->queue != PQ_NONE && !vm_page_busied(m)) { 2265 /* 2266 * The page is allocated but eligible for 2267 * relocation. Extend the current run by one 2268 * page. 2269 */ 2270 KASSERT(pmap_page_get_memattr(m) == 2271 VM_MEMATTR_DEFAULT, 2272 ("page %p has an unexpected memattr", m)); 2273 KASSERT((m->oflags & (VPO_SWAPINPROG | 2274 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, 2275 ("page %p has unexpected oflags", m)); 2276 /* Don't care: VPO_NOSYNC. */ 2277 run_ext = 1; 2278 } else 2279 run_ext = 0; 2280unlock: 2281 VM_OBJECT_RUNLOCK(object); 2282#if VM_NRESERVLEVEL > 0 2283 } else if (level >= 0) { 2284 /* 2285 * The page is reserved but not yet allocated. In 2286 * other words, it is still cached or free. Extend 2287 * the current run by one page. 2288 */ 2289 run_ext = 1; 2290#endif 2291 } else if ((order = m->order) < VM_NFREEORDER) { 2292 /* 2293 * The page is enqueued in the physical memory 2294 * allocator's cache/free page queues. Moreover, it 2295 * is the first page in a power-of-two-sized run of 2296 * contiguous cache/free pages. Add these pages to 2297 * the end of the current run, and jump ahead. 2298 */ 2299 run_ext = 1 << order; 2300 m_inc = 1 << order; 2301 } else { 2302 /* 2303 * Skip the page for one of the following reasons: (1) 2304 * It is enqueued in the physical memory allocator's 2305 * cache/free page queues. However, it is not the 2306 * first page in a run of contiguous cache/free pages. 2307 * (This case rarely occurs because the scan is 2308 * performed in ascending order.) (2) It is not 2309 * reserved, and it is transitioning from free to 2310 * allocated. (Conversely, the transition from 2311 * allocated to free for managed pages is blocked by 2312 * the page lock.) (3) It is allocated but not 2313 * contained by an object and not wired, e.g., 2314 * allocated by Xen's balloon driver. 2315 */ 2316 run_ext = 0; 2317 } 2318 2319 /* 2320 * Extend or reset the current run of pages. 2321 */ 2322 if (run_ext > 0) { 2323 if (run_len == 0) 2324 m_run = m; 2325 run_len += run_ext; 2326 } else { 2327 if (run_len > 0) { 2328 m_run = NULL; 2329 run_len = 0; 2330 } 2331 } 2332 } 2333 if (m_mtx != NULL) 2334 mtx_unlock(m_mtx); 2335 if (run_len >= npages) 2336 return (m_run); 2337 return (NULL); 2338} 2339 2340/* 2341 * vm_page_reclaim_run: 2342 * 2343 * Try to relocate each of the allocated virtual pages within the 2344 * specified run of physical pages to a new physical address. Free the 2345 * physical pages underlying the relocated virtual pages. A virtual page 2346 * is relocatable if and only if it could be laundered or reclaimed by 2347 * the page daemon. Whenever possible, a virtual page is relocated to a 2348 * physical address above "high". 2349 * 2350 * Returns 0 if every physical page within the run was already free or 2351 * just freed by a successful relocation. Otherwise, returns a non-zero 2352 * value indicating why the last attempt to relocate a virtual page was 2353 * unsuccessful. 2354 * 2355 * "req_class" must be an allocation class. 2356 */ 2357static int 2358vm_page_reclaim_run(int req_class, u_long npages, vm_page_t m_run, 2359 vm_paddr_t high) 2360{ 2361 struct mtx *m_mtx, *new_mtx; 2362 struct spglist free; 2363 vm_object_t object; 2364 vm_paddr_t pa; 2365 vm_page_t m, m_end, m_new; 2366 int error, order, req; 2367 2368 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class, 2369 ("req_class is not an allocation class")); 2370 SLIST_INIT(&free); 2371 error = 0; 2372 m = m_run; 2373 m_end = m_run + npages; 2374 m_mtx = NULL; 2375 for (; error == 0 && m < m_end; m++) { 2376 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0, 2377 ("page %p is PG_FICTITIOUS or PG_MARKER", m)); 2378 2379 /* 2380 * Avoid releasing and reacquiring the same page lock. 2381 */ 2382 new_mtx = vm_page_lockptr(m); 2383 if (m_mtx != new_mtx) { 2384 if (m_mtx != NULL) 2385 mtx_unlock(m_mtx); 2386 m_mtx = new_mtx; 2387 mtx_lock(m_mtx); 2388 } 2389retry: 2390 if (m->wire_count != 0 || m->hold_count != 0) 2391 error = EBUSY; 2392 else if ((object = m->object) != NULL) { 2393 /* 2394 * The page is relocated if and only if it could be 2395 * laundered or reclaimed by the page daemon. 2396 */ 2397 if (!VM_OBJECT_TRYWLOCK(object)) { 2398 mtx_unlock(m_mtx); 2399 VM_OBJECT_WLOCK(object); 2400 mtx_lock(m_mtx); 2401 if (m->object != object) { 2402 /* 2403 * The page may have been freed. 2404 */ 2405 VM_OBJECT_WUNLOCK(object); 2406 goto retry; 2407 } else if (m->wire_count != 0 || 2408 m->hold_count != 0) { 2409 error = EBUSY; 2410 goto unlock; 2411 } 2412 } 2413 KASSERT((m->flags & PG_UNHOLDFREE) == 0, 2414 ("page %p is PG_UNHOLDFREE", m)); 2415 /* Don't care: PG_NODUMP, PG_WINATCFLS, PG_ZERO. */ 2416 if (object->type != OBJT_DEFAULT && 2417 object->type != OBJT_SWAP && 2418 object->type != OBJT_VNODE) 2419 error = EINVAL; 2420 else if ((m->flags & PG_CACHED) != 0 || 2421 m != vm_page_lookup(object, m->pindex)) { 2422 /* 2423 * The page is cached or recently converted 2424 * from cached to free. 2425 */ 2426 VM_OBJECT_WUNLOCK(object); 2427 goto cached; 2428 } else if (object->memattr != VM_MEMATTR_DEFAULT) 2429 error = EINVAL; 2430 else if (m->queue != PQ_NONE && !vm_page_busied(m)) { 2431 KASSERT(pmap_page_get_memattr(m) == 2432 VM_MEMATTR_DEFAULT, 2433 ("page %p has an unexpected memattr", m)); 2434 KASSERT((m->oflags & (VPO_SWAPINPROG | 2435 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, 2436 ("page %p has unexpected oflags", m)); 2437 /* Don't care: VPO_NOSYNC. */ 2438 if (m->valid != 0) { 2439 /* 2440 * First, try to allocate a new page 2441 * that is above "high". Failing 2442 * that, try to allocate a new page 2443 * that is below "m_run". Allocate 2444 * the new page between the end of 2445 * "m_run" and "high" only as a last 2446 * resort. 2447 */ 2448 req = req_class | VM_ALLOC_NOOBJ; 2449 if ((m->flags & PG_NODUMP) != 0) 2450 req |= VM_ALLOC_NODUMP; 2451 if (trunc_page(high) != 2452 ~(vm_paddr_t)PAGE_MASK) { 2453 m_new = vm_page_alloc_contig( 2454 NULL, 0, req, 1, 2455 round_page(high), 2456 ~(vm_paddr_t)0, 2457 PAGE_SIZE, 0, 2458 VM_MEMATTR_DEFAULT); 2459 } else 2460 m_new = NULL; 2461 if (m_new == NULL) { 2462 pa = VM_PAGE_TO_PHYS(m_run); 2463 m_new = vm_page_alloc_contig( 2464 NULL, 0, req, 1, 2465 0, pa - 1, PAGE_SIZE, 0, 2466 VM_MEMATTR_DEFAULT); 2467 } 2468 if (m_new == NULL) { 2469 pa += ptoa(npages); 2470 m_new = vm_page_alloc_contig( 2471 NULL, 0, req, 1, 2472 pa, high, PAGE_SIZE, 0, 2473 VM_MEMATTR_DEFAULT); 2474 } 2475 if (m_new == NULL) { 2476 error = ENOMEM; 2477 goto unlock; 2478 } 2479 KASSERT(m_new->wire_count == 0, 2480 ("page %p is wired", m)); 2481 2482 /* 2483 * Replace "m" with the new page. For 2484 * vm_page_replace(), "m" must be busy 2485 * and dequeued. Finally, change "m" 2486 * as if vm_page_free() was called. 2487 */ 2488 if (object->ref_count != 0) 2489 pmap_remove_all(m); 2490 m_new->aflags = m->aflags; 2491 KASSERT(m_new->oflags == VPO_UNMANAGED, 2492 ("page %p is managed", m)); 2493 m_new->oflags = m->oflags & VPO_NOSYNC; 2494 pmap_copy_page(m, m_new); 2495 m_new->valid = m->valid; 2496 m_new->dirty = m->dirty; 2497 m->flags &= ~PG_ZERO; 2498 vm_page_xbusy(m); 2499 vm_page_remque(m); 2500 vm_page_replace_checked(m_new, object, 2501 m->pindex, m); 2502 m->valid = 0; 2503 vm_page_undirty(m); 2504 2505 /* 2506 * The new page must be deactivated 2507 * before the object is unlocked. 2508 */ 2509 new_mtx = vm_page_lockptr(m_new); 2510 if (m_mtx != new_mtx) { 2511 mtx_unlock(m_mtx); 2512 m_mtx = new_mtx; 2513 mtx_lock(m_mtx); 2514 } 2515 vm_page_deactivate(m_new); 2516 } else { 2517 m->flags &= ~PG_ZERO; 2518 vm_page_remque(m); 2519 vm_page_remove(m); 2520 KASSERT(m->dirty == 0, 2521 ("page %p is dirty", m)); 2522 } 2523 SLIST_INSERT_HEAD(&free, m, plinks.s.ss); 2524 } else 2525 error = EBUSY; 2526unlock: 2527 VM_OBJECT_WUNLOCK(object); 2528 } else { 2529cached: 2530 mtx_lock(&vm_page_queue_free_mtx); 2531 order = m->order; 2532 if (order < VM_NFREEORDER) { 2533 /* 2534 * The page is enqueued in the physical memory 2535 * allocator's cache/free page queues. 2536 * Moreover, it is the first page in a power- 2537 * of-two-sized run of contiguous cache/free 2538 * pages. Jump ahead to the last page within 2539 * that run, and continue from there. 2540 */ 2541 m += (1 << order) - 1; 2542 } 2543#if VM_NRESERVLEVEL > 0 2544 else if (vm_reserv_is_page_free(m)) 2545 order = 0; 2546#endif 2547 mtx_unlock(&vm_page_queue_free_mtx); 2548 if (order == VM_NFREEORDER) 2549 error = EINVAL; 2550 } 2551 } 2552 if (m_mtx != NULL) 2553 mtx_unlock(m_mtx); 2554 if ((m = SLIST_FIRST(&free)) != NULL) { 2555 mtx_lock(&vm_page_queue_free_mtx); 2556 do { 2557 SLIST_REMOVE_HEAD(&free, plinks.s.ss); 2558 vm_phys_freecnt_adj(m, 1); 2559#if VM_NRESERVLEVEL > 0 2560 if (!vm_reserv_free_page(m)) 2561#else 2562 if (true) 2563#endif 2564 vm_phys_free_pages(m, 0); 2565 } while ((m = SLIST_FIRST(&free)) != NULL); 2566 vm_page_zero_idle_wakeup(); 2567 vm_page_free_wakeup(); 2568 mtx_unlock(&vm_page_queue_free_mtx); 2569 } 2570 return (error); 2571} 2572 2573#define NRUNS 16 2574 2575CTASSERT(powerof2(NRUNS)); 2576 2577#define RUN_INDEX(count) ((count) & (NRUNS - 1)) 2578 2579#define MIN_RECLAIM 8 2580 2581/* 2582 * vm_page_reclaim_contig: 2583 * 2584 * Reclaim allocated, contiguous physical memory satisfying the specified 2585 * conditions by relocating the virtual pages using that physical memory. 2586 * Returns true if reclamation is successful and false otherwise. Since 2587 * relocation requires the allocation of physical pages, reclamation may 2588 * fail due to a shortage of cache/free pages. When reclamation fails, 2589 * callers are expected to perform VM_WAIT before retrying a failed 2590 * allocation operation, e.g., vm_page_alloc_contig(). 2591 * 2592 * The caller must always specify an allocation class through "req". 2593 * 2594 * allocation classes: 2595 * VM_ALLOC_NORMAL normal process request 2596 * VM_ALLOC_SYSTEM system *really* needs a page 2597 * VM_ALLOC_INTERRUPT interrupt time request 2598 * 2599 * The optional allocation flags are ignored. 2600 * 2601 * "npages" must be greater than zero. Both "alignment" and "boundary" 2602 * must be a power of two. 2603 */ 2604bool 2605vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, 2606 u_long alignment, vm_paddr_t boundary) 2607{ 2608 vm_paddr_t curr_low; 2609 vm_page_t m_run, m_runs[NRUNS]; 2610 u_long count, reclaimed; 2611 int error, i, options, req_class; 2612 2613 KASSERT(npages > 0, ("npages is 0")); 2614 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2615 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2616 req_class = req & VM_ALLOC_CLASS_MASK; 2617 2618 /* 2619 * The page daemon is allowed to dig deeper into the free page list. 2620 */ 2621 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 2622 req_class = VM_ALLOC_SYSTEM; 2623 2624 /* 2625 * Return if the number of cached and free pages cannot satisfy the 2626 * requested allocation. 2627 */ 2628 count = vm_cnt.v_free_count + vm_cnt.v_cache_count; 2629 if (count < npages + vm_cnt.v_free_reserved || (count < npages + 2630 vm_cnt.v_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) || 2631 (count < npages && req_class == VM_ALLOC_INTERRUPT)) 2632 return (false); 2633 2634 /* 2635 * Scan up to three times, relaxing the restrictions ("options") on 2636 * the reclamation of reservations and superpages each time. 2637 */ 2638 for (options = VPSC_NORESERV;;) { 2639 /* 2640 * Find the highest runs that satisfy the given constraints 2641 * and restrictions, and record them in "m_runs". 2642 */ 2643 curr_low = low; 2644 count = 0; 2645 for (;;) { 2646 m_run = vm_phys_scan_contig(npages, curr_low, high, 2647 alignment, boundary, options); 2648 if (m_run == NULL) 2649 break; 2650 curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages); 2651 m_runs[RUN_INDEX(count)] = m_run; 2652 count++; 2653 } 2654 2655 /* 2656 * Reclaim the highest runs in LIFO (descending) order until 2657 * the number of reclaimed pages, "reclaimed", is at least 2658 * MIN_RECLAIM. Reset "reclaimed" each time because each 2659 * reclamation is idempotent, and runs will (likely) recur 2660 * from one scan to the next as restrictions are relaxed. 2661 */ 2662 reclaimed = 0; 2663 for (i = 0; count > 0 && i < NRUNS; i++) { 2664 count--; 2665 m_run = m_runs[RUN_INDEX(count)]; 2666 error = vm_page_reclaim_run(req_class, npages, m_run, 2667 high); 2668 if (error == 0) { 2669 reclaimed += npages; 2670 if (reclaimed >= MIN_RECLAIM) 2671 return (true); 2672 } 2673 } 2674 2675 /* 2676 * Either relax the restrictions on the next scan or return if 2677 * the last scan had no restrictions. 2678 */ 2679 if (options == VPSC_NORESERV) 2680 options = VPSC_NOSUPER; 2681 else if (options == VPSC_NOSUPER) 2682 options = VPSC_ANY; 2683 else if (options == VPSC_ANY) 2684 return (reclaimed != 0); 2685 } 2686} 2687 2688/* 2689 * vm_wait: (also see VM_WAIT macro) 2690 * 2691 * Sleep until free pages are available for allocation. 2692 * - Called in various places before memory allocations. 2693 */ 2694void 2695vm_wait(void) 2696{ 2697 2698 mtx_lock(&vm_page_queue_free_mtx); 2699 if (curproc == pageproc) { 2700 vm_pageout_pages_needed = 1; 2701 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx, 2702 PDROP | PSWP, "VMWait", 0); 2703 } else { 2704 if (!vm_pageout_wanted) { 2705 vm_pageout_wanted = true; 2706 wakeup(&vm_pageout_wanted); 2707 } 2708 vm_pages_needed = true; 2709 msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM, 2710 "vmwait", 0); 2711 } 2712} 2713 2714/* 2715 * vm_waitpfault: (also see VM_WAITPFAULT macro) 2716 * 2717 * Sleep until free pages are available for allocation. 2718 * - Called only in vm_fault so that processes page faulting 2719 * can be easily tracked. 2720 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 2721 * processes will be able to grab memory first. Do not change 2722 * this balance without careful testing first. 2723 */ 2724void 2725vm_waitpfault(void) 2726{ 2727 2728 mtx_lock(&vm_page_queue_free_mtx); 2729 if (!vm_pageout_wanted) { 2730 vm_pageout_wanted = true; 2731 wakeup(&vm_pageout_wanted); 2732 } 2733 vm_pages_needed = true; 2734 msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER, 2735 "pfault", 0); 2736} 2737 2738struct vm_pagequeue * 2739vm_page_pagequeue(vm_page_t m) 2740{ 2741 2742 return (&vm_phys_domain(m)->vmd_pagequeues[m->queue]); 2743} 2744 2745/* 2746 * vm_page_dequeue: 2747 * 2748 * Remove the given page from its current page queue. 2749 * 2750 * The page must be locked. 2751 */ 2752void 2753vm_page_dequeue(vm_page_t m) 2754{ 2755 struct vm_pagequeue *pq; 2756 2757 vm_page_assert_locked(m); 2758 KASSERT(m->queue < PQ_COUNT, ("vm_page_dequeue: page %p is not queued", 2759 m)); 2760 pq = vm_page_pagequeue(m); 2761 vm_pagequeue_lock(pq); 2762 m->queue = PQ_NONE; 2763 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2764 vm_pagequeue_cnt_dec(pq); 2765 vm_pagequeue_unlock(pq); 2766} 2767 2768/* 2769 * vm_page_dequeue_locked: 2770 * 2771 * Remove the given page from its current page queue. 2772 * 2773 * The page and page queue must be locked. 2774 */ 2775void 2776vm_page_dequeue_locked(vm_page_t m) 2777{ 2778 struct vm_pagequeue *pq; 2779 2780 vm_page_lock_assert(m, MA_OWNED); 2781 pq = vm_page_pagequeue(m); 2782 vm_pagequeue_assert_locked(pq); 2783 m->queue = PQ_NONE; 2784 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2785 vm_pagequeue_cnt_dec(pq); 2786} 2787 2788/* 2789 * vm_page_enqueue: 2790 * 2791 * Add the given page to the specified page queue. 2792 * 2793 * The page must be locked. 2794 */ 2795static void 2796vm_page_enqueue(uint8_t queue, vm_page_t m) 2797{ 2798 struct vm_pagequeue *pq; 2799 2800 vm_page_lock_assert(m, MA_OWNED); 2801 KASSERT(queue < PQ_COUNT, 2802 ("vm_page_enqueue: invalid queue %u request for page %p", 2803 queue, m)); 2804 pq = &vm_phys_domain(m)->vmd_pagequeues[queue]; 2805 vm_pagequeue_lock(pq); 2806 m->queue = queue; 2807 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2808 vm_pagequeue_cnt_inc(pq); 2809 vm_pagequeue_unlock(pq); 2810} 2811 2812/* 2813 * vm_page_requeue: 2814 * 2815 * Move the given page to the tail of its current page queue. 2816 * 2817 * The page must be locked. 2818 */ 2819void 2820vm_page_requeue(vm_page_t m) 2821{ 2822 struct vm_pagequeue *pq; 2823 2824 vm_page_lock_assert(m, MA_OWNED); 2825 KASSERT(m->queue != PQ_NONE, 2826 ("vm_page_requeue: page %p is not queued", m)); 2827 pq = vm_page_pagequeue(m); 2828 vm_pagequeue_lock(pq); 2829 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2830 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2831 vm_pagequeue_unlock(pq); 2832} 2833 2834/* 2835 * vm_page_requeue_locked: 2836 * 2837 * Move the given page to the tail of its current page queue. 2838 * 2839 * The page queue must be locked. 2840 */ 2841void 2842vm_page_requeue_locked(vm_page_t m) 2843{ 2844 struct vm_pagequeue *pq; 2845 2846 KASSERT(m->queue != PQ_NONE, 2847 ("vm_page_requeue_locked: page %p is not queued", m)); 2848 pq = vm_page_pagequeue(m); 2849 vm_pagequeue_assert_locked(pq); 2850 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2851 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2852} 2853 2854/* 2855 * vm_page_activate: 2856 * 2857 * Put the specified page on the active list (if appropriate). 2858 * Ensure that act_count is at least ACT_INIT but do not otherwise 2859 * mess with it. 2860 * 2861 * The page must be locked. 2862 */ 2863void 2864vm_page_activate(vm_page_t m) 2865{ 2866 int queue; 2867 2868 vm_page_lock_assert(m, MA_OWNED); 2869 if ((queue = m->queue) != PQ_ACTIVE) { 2870 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 2871 if (m->act_count < ACT_INIT) 2872 m->act_count = ACT_INIT; 2873 if (queue != PQ_NONE) 2874 vm_page_dequeue(m); 2875 vm_page_enqueue(PQ_ACTIVE, m); 2876 } else 2877 KASSERT(queue == PQ_NONE, 2878 ("vm_page_activate: wired page %p is queued", m)); 2879 } else { 2880 if (m->act_count < ACT_INIT) 2881 m->act_count = ACT_INIT; 2882 } 2883} 2884 2885/* 2886 * vm_page_free_wakeup: 2887 * 2888 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 2889 * routine is called when a page has been added to the cache or free 2890 * queues. 2891 * 2892 * The page queues must be locked. 2893 */ 2894static inline void 2895vm_page_free_wakeup(void) 2896{ 2897 2898 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 2899 /* 2900 * if pageout daemon needs pages, then tell it that there are 2901 * some free. 2902 */ 2903 if (vm_pageout_pages_needed && 2904 vm_cnt.v_cache_count + vm_cnt.v_free_count >= vm_cnt.v_pageout_free_min) { 2905 wakeup(&vm_pageout_pages_needed); 2906 vm_pageout_pages_needed = 0; 2907 } 2908 /* 2909 * wakeup processes that are waiting on memory if we hit a 2910 * high water mark. And wakeup scheduler process if we have 2911 * lots of memory. this process will swapin processes. 2912 */ 2913 if (vm_pages_needed && !vm_page_count_min()) { 2914 vm_pages_needed = false; 2915 wakeup(&vm_cnt.v_free_count); 2916 } 2917} 2918 2919/* 2920 * Turn a cached page into a free page, by changing its attributes. 2921 * Keep the statistics up-to-date. 2922 * 2923 * The free page queue must be locked. 2924 */ 2925static void 2926vm_page_cache_turn_free(vm_page_t m) 2927{ 2928 2929 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 2930 2931 m->object = NULL; 2932 m->valid = 0; 2933 KASSERT((m->flags & PG_CACHED) != 0, 2934 ("vm_page_cache_turn_free: page %p is not cached", m)); 2935 m->flags &= ~PG_CACHED; 2936 vm_cnt.v_cache_count--; 2937 vm_phys_freecnt_adj(m, 1); 2938} 2939 2940/* 2941 * vm_page_free_toq: 2942 * 2943 * Returns the given page to the free list, 2944 * disassociating it with any VM object. 2945 * 2946 * The object must be locked. The page must be locked if it is managed. 2947 */ 2948void 2949vm_page_free_toq(vm_page_t m) 2950{ 2951 2952 if ((m->oflags & VPO_UNMANAGED) == 0) { 2953 vm_page_lock_assert(m, MA_OWNED); 2954 KASSERT(!pmap_page_is_mapped(m), 2955 ("vm_page_free_toq: freeing mapped page %p", m)); 2956 } else 2957 KASSERT(m->queue == PQ_NONE, 2958 ("vm_page_free_toq: unmanaged page %p is queued", m)); 2959 PCPU_INC(cnt.v_tfree); 2960 2961 if (vm_page_sbusied(m)) 2962 panic("vm_page_free: freeing busy page %p", m); 2963 2964 /* 2965 * Unqueue, then remove page. Note that we cannot destroy 2966 * the page here because we do not want to call the pager's 2967 * callback routine until after we've put the page on the 2968 * appropriate free queue. 2969 */ 2970 vm_page_remque(m); 2971 vm_page_remove(m); 2972 2973 /* 2974 * If fictitious remove object association and 2975 * return, otherwise delay object association removal. 2976 */ 2977 if ((m->flags & PG_FICTITIOUS) != 0) { 2978 return; 2979 } 2980 2981 m->valid = 0; 2982 vm_page_undirty(m); 2983 2984 if (m->wire_count != 0) 2985 panic("vm_page_free: freeing wired page %p", m); 2986 if (m->hold_count != 0) { 2987 m->flags &= ~PG_ZERO; 2988 KASSERT((m->flags & PG_UNHOLDFREE) == 0, 2989 ("vm_page_free: freeing PG_UNHOLDFREE page %p", m)); 2990 m->flags |= PG_UNHOLDFREE; 2991 } else { 2992 /* 2993 * Restore the default memory attribute to the page. 2994 */ 2995 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 2996 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 2997 2998 /* 2999 * Insert the page into the physical memory allocator's 3000 * cache/free page queues. 3001 */ 3002 mtx_lock(&vm_page_queue_free_mtx); 3003 vm_phys_freecnt_adj(m, 1); 3004#if VM_NRESERVLEVEL > 0 3005 if (!vm_reserv_free_page(m)) 3006#else 3007 if (TRUE) 3008#endif 3009 vm_phys_free_pages(m, 0); 3010 if ((m->flags & PG_ZERO) != 0) 3011 ++vm_page_zero_count; 3012 else 3013 vm_page_zero_idle_wakeup(); 3014 vm_page_free_wakeup(); 3015 mtx_unlock(&vm_page_queue_free_mtx); 3016 } 3017} 3018 3019/* 3020 * vm_page_wire: 3021 * 3022 * Mark this page as wired down by yet 3023 * another map, removing it from paging queues 3024 * as necessary. 3025 * 3026 * If the page is fictitious, then its wire count must remain one. 3027 * 3028 * The page must be locked. 3029 */ 3030void 3031vm_page_wire(vm_page_t m) 3032{ 3033 3034 /* 3035 * Only bump the wire statistics if the page is not already wired, 3036 * and only unqueue the page if it is on some queue (if it is unmanaged 3037 * it is already off the queues). 3038 */ 3039 vm_page_lock_assert(m, MA_OWNED); 3040 if ((m->flags & PG_FICTITIOUS) != 0) { 3041 KASSERT(m->wire_count == 1, 3042 ("vm_page_wire: fictitious page %p's wire count isn't one", 3043 m)); 3044 return; 3045 } 3046 if (m->wire_count == 0) { 3047 KASSERT((m->oflags & VPO_UNMANAGED) == 0 || 3048 m->queue == PQ_NONE, 3049 ("vm_page_wire: unmanaged page %p is queued", m)); 3050 vm_page_remque(m); 3051 atomic_add_int(&vm_cnt.v_wire_count, 1); 3052 } 3053 m->wire_count++; 3054 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 3055} 3056 3057/* 3058 * vm_page_unwire: 3059 * 3060 * Release one wiring of the specified page, potentially allowing it to be 3061 * paged out. Returns TRUE if the number of wirings transitions to zero and 3062 * FALSE otherwise. 3063 * 3064 * Only managed pages belonging to an object can be paged out. If the number 3065 * of wirings transitions to zero and the page is eligible for page out, then 3066 * the page is added to the specified paging queue (unless PQ_NONE is 3067 * specified). 3068 * 3069 * If a page is fictitious, then its wire count must always be one. 3070 * 3071 * A managed page must be locked. 3072 */ 3073boolean_t 3074vm_page_unwire(vm_page_t m, uint8_t queue) 3075{ 3076 3077 KASSERT(queue < PQ_COUNT || queue == PQ_NONE, 3078 ("vm_page_unwire: invalid queue %u request for page %p", 3079 queue, m)); 3080 if ((m->oflags & VPO_UNMANAGED) == 0) 3081 vm_page_assert_locked(m); 3082 if ((m->flags & PG_FICTITIOUS) != 0) { 3083 KASSERT(m->wire_count == 1, 3084 ("vm_page_unwire: fictitious page %p's wire count isn't one", m)); 3085 return (FALSE); 3086 } 3087 if (m->wire_count > 0) { 3088 m->wire_count--; 3089 if (m->wire_count == 0) { 3090 atomic_subtract_int(&vm_cnt.v_wire_count, 1); 3091 if ((m->oflags & VPO_UNMANAGED) == 0 && 3092 m->object != NULL && queue != PQ_NONE) { 3093 if (queue == PQ_INACTIVE) 3094 m->flags &= ~PG_WINATCFLS; 3095 vm_page_enqueue(queue, m); 3096 } 3097 return (TRUE); 3098 } else 3099 return (FALSE); 3100 } else 3101 panic("vm_page_unwire: page %p's wire count is zero", m); 3102} 3103 3104/* 3105 * Move the specified page to the inactive queue. 3106 * 3107 * Many pages placed on the inactive queue should actually go 3108 * into the cache, but it is difficult to figure out which. What 3109 * we do instead, if the inactive target is well met, is to put 3110 * clean pages at the head of the inactive queue instead of the tail. 3111 * This will cause them to be moved to the cache more quickly and 3112 * if not actively re-referenced, reclaimed more quickly. If we just 3113 * stick these pages at the end of the inactive queue, heavy filesystem 3114 * meta-data accesses can cause an unnecessary paging load on memory bound 3115 * processes. This optimization causes one-time-use metadata to be 3116 * reused more quickly. 3117 * 3118 * Normally noreuse is FALSE, resulting in LRU operation. noreuse is set 3119 * to TRUE if we want this page to be 'as if it were placed in the cache', 3120 * except without unmapping it from the process address space. In 3121 * practice this is implemented by inserting the page at the head of the 3122 * queue, using a marker page to guide FIFO insertion ordering. 3123 * 3124 * The page must be locked. 3125 */ 3126static inline void 3127_vm_page_deactivate(vm_page_t m, boolean_t noreuse) 3128{ 3129 struct vm_pagequeue *pq; 3130 int queue; 3131 3132 vm_page_assert_locked(m); 3133 3134 /* 3135 * Ignore if the page is already inactive, unless it is unlikely to be 3136 * reactivated. 3137 */ 3138 if ((queue = m->queue) == PQ_INACTIVE && !noreuse) 3139 return; 3140 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 3141 pq = &vm_phys_domain(m)->vmd_pagequeues[PQ_INACTIVE]; 3142 /* Avoid multiple acquisitions of the inactive queue lock. */ 3143 if (queue == PQ_INACTIVE) { 3144 vm_pagequeue_lock(pq); 3145 vm_page_dequeue_locked(m); 3146 } else { 3147 if (queue != PQ_NONE) 3148 vm_page_dequeue(m); 3149 m->flags &= ~PG_WINATCFLS; 3150 vm_pagequeue_lock(pq); 3151 } 3152 m->queue = PQ_INACTIVE; 3153 if (noreuse) 3154 TAILQ_INSERT_BEFORE(&vm_phys_domain(m)->vmd_inacthead, 3155 m, plinks.q); 3156 else 3157 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 3158 vm_pagequeue_cnt_inc(pq); 3159 vm_pagequeue_unlock(pq); 3160 } 3161} 3162 3163/* 3164 * Move the specified page to the inactive queue. 3165 * 3166 * The page must be locked. 3167 */ 3168void 3169vm_page_deactivate(vm_page_t m) 3170{ 3171 3172 _vm_page_deactivate(m, FALSE); 3173} 3174 3175/* 3176 * Move the specified page to the inactive queue with the expectation 3177 * that it is unlikely to be reused. 3178 * 3179 * The page must be locked. 3180 */ 3181void 3182vm_page_deactivate_noreuse(vm_page_t m) 3183{ 3184 3185 _vm_page_deactivate(m, TRUE); 3186} 3187 3188/* 3189 * vm_page_try_to_cache: 3190 * 3191 * Returns 0 on failure, 1 on success 3192 */ 3193int 3194vm_page_try_to_cache(vm_page_t m) 3195{ 3196 3197 vm_page_lock_assert(m, MA_OWNED); 3198 VM_OBJECT_ASSERT_WLOCKED(m->object); 3199 if (m->dirty || m->hold_count || m->wire_count || 3200 (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m)) 3201 return (0); 3202 pmap_remove_all(m); 3203 if (m->dirty) 3204 return (0); 3205 vm_page_cache(m); 3206 return (1); 3207} 3208 3209/* 3210 * vm_page_try_to_free() 3211 * 3212 * Attempt to free the page. If we cannot free it, we do nothing. 3213 * 1 is returned on success, 0 on failure. 3214 */ 3215int 3216vm_page_try_to_free(vm_page_t m) 3217{ 3218 3219 vm_page_lock_assert(m, MA_OWNED); 3220 if (m->object != NULL) 3221 VM_OBJECT_ASSERT_WLOCKED(m->object); 3222 if (m->dirty || m->hold_count || m->wire_count || 3223 (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m)) 3224 return (0); 3225 pmap_remove_all(m); 3226 if (m->dirty) 3227 return (0); 3228 vm_page_free(m); 3229 return (1); 3230} 3231 3232/* 3233 * vm_page_cache 3234 * 3235 * Put the specified page onto the page cache queue (if appropriate). 3236 * 3237 * The object and page must be locked. 3238 */ 3239void 3240vm_page_cache(vm_page_t m) 3241{ 3242 vm_object_t object; 3243 boolean_t cache_was_empty; 3244 3245 vm_page_lock_assert(m, MA_OWNED); 3246 object = m->object; 3247 VM_OBJECT_ASSERT_WLOCKED(object); 3248 if (vm_page_busied(m) || (m->oflags & VPO_UNMANAGED) || 3249 m->hold_count || m->wire_count) 3250 panic("vm_page_cache: attempting to cache busy page"); 3251 KASSERT(!pmap_page_is_mapped(m), 3252 ("vm_page_cache: page %p is mapped", m)); 3253 KASSERT(m->dirty == 0, ("vm_page_cache: page %p is dirty", m)); 3254 if (m->valid == 0 || object->type == OBJT_DEFAULT || 3255 (object->type == OBJT_SWAP && 3256 !vm_pager_has_page(object, m->pindex, NULL, NULL))) { 3257 /* 3258 * Hypothesis: A cache-eligible page belonging to a 3259 * default object or swap object but without a backing 3260 * store must be zero filled. 3261 */ 3262 vm_page_free(m); 3263 return; 3264 } 3265 KASSERT((m->flags & PG_CACHED) == 0, 3266 ("vm_page_cache: page %p is already cached", m)); 3267 3268 /* 3269 * Remove the page from the paging queues. 3270 */ 3271 vm_page_remque(m); 3272 3273 /* 3274 * Remove the page from the object's collection of resident 3275 * pages. 3276 */ 3277 vm_radix_remove(&object->rtree, m->pindex); 3278 TAILQ_REMOVE(&object->memq, m, listq); 3279 object->resident_page_count--; 3280 3281 /* 3282 * Restore the default memory attribute to the page. 3283 */ 3284 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 3285 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 3286 3287 /* 3288 * Insert the page into the object's collection of cached pages 3289 * and the physical memory allocator's cache/free page queues. 3290 */ 3291 m->flags &= ~PG_ZERO; 3292 mtx_lock(&vm_page_queue_free_mtx); 3293 cache_was_empty = vm_radix_is_empty(&object->cache); 3294 if (vm_radix_insert(&object->cache, m)) { 3295 mtx_unlock(&vm_page_queue_free_mtx); 3296 if (object->type == OBJT_VNODE && 3297 object->resident_page_count == 0) 3298 vdrop(object->handle); 3299 m->object = NULL; 3300 vm_page_free(m); 3301 return; 3302 } 3303 3304 /* 3305 * The above call to vm_radix_insert() could reclaim the one pre- 3306 * existing cached page from this object, resulting in a call to 3307 * vdrop(). 3308 */ 3309 if (!cache_was_empty) 3310 cache_was_empty = vm_radix_is_singleton(&object->cache); 3311 3312 m->flags |= PG_CACHED; 3313 vm_cnt.v_cache_count++; 3314 PCPU_INC(cnt.v_tcached); 3315#if VM_NRESERVLEVEL > 0 3316 if (!vm_reserv_free_page(m)) { 3317#else 3318 if (TRUE) { 3319#endif 3320 vm_phys_free_pages(m, 0); 3321 } 3322 vm_page_free_wakeup(); 3323 mtx_unlock(&vm_page_queue_free_mtx); 3324 3325 /* 3326 * Increment the vnode's hold count if this is the object's only 3327 * cached page. Decrement the vnode's hold count if this was 3328 * the object's only resident page. 3329 */ 3330 if (object->type == OBJT_VNODE) { 3331 if (cache_was_empty && object->resident_page_count != 0) 3332 vhold(object->handle); 3333 else if (!cache_was_empty && object->resident_page_count == 0) 3334 vdrop(object->handle); 3335 } 3336} 3337 3338/* 3339 * vm_page_advise 3340 * 3341 * Deactivate or do nothing, as appropriate. 3342 * 3343 * The object and page must be locked. 3344 */ 3345void 3346vm_page_advise(vm_page_t m, int advice) 3347{ 3348 3349 vm_page_assert_locked(m); 3350 VM_OBJECT_ASSERT_WLOCKED(m->object); 3351 if (advice == MADV_FREE) 3352 /* 3353 * Mark the page clean. This will allow the page to be freed 3354 * up by the system. However, such pages are often reused 3355 * quickly by malloc() so we do not do anything that would 3356 * cause a page fault if we can help it. 3357 * 3358 * Specifically, we do not try to actually free the page now 3359 * nor do we try to put it in the cache (which would cause a 3360 * page fault on reuse). 3361 * 3362 * But we do make the page as freeable as we can without 3363 * actually taking the step of unmapping it. 3364 */ 3365 m->dirty = 0; 3366 else if (advice != MADV_DONTNEED) 3367 return; 3368 3369 /* 3370 * Clear any references to the page. Otherwise, the page daemon will 3371 * immediately reactivate the page. 3372 */ 3373 vm_page_aflag_clear(m, PGA_REFERENCED); 3374 3375 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) 3376 vm_page_dirty(m); 3377 3378 /* 3379 * Place clean pages at the head of the inactive queue rather than the 3380 * tail, thus defeating the queue's LRU operation and ensuring that the 3381 * page will be reused quickly. 3382 */ 3383 _vm_page_deactivate(m, m->dirty == 0); 3384} 3385 3386/* 3387 * Grab a page, waiting until we are waken up due to the page 3388 * changing state. We keep on waiting, if the page continues 3389 * to be in the object. If the page doesn't exist, first allocate it 3390 * and then conditionally zero it. 3391 * 3392 * This routine may sleep. 3393 * 3394 * The object must be locked on entry. The lock will, however, be released 3395 * and reacquired if the routine sleeps. 3396 */ 3397vm_page_t 3398vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 3399{ 3400 vm_page_t m; 3401 int sleep; 3402 3403 VM_OBJECT_ASSERT_WLOCKED(object); 3404 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 3405 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 3406 ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 3407retrylookup: 3408 if ((m = vm_page_lookup(object, pindex)) != NULL) { 3409 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ? 3410 vm_page_xbusied(m) : vm_page_busied(m); 3411 if (sleep) { 3412 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 3413 return (NULL); 3414 /* 3415 * Reference the page before unlocking and 3416 * sleeping so that the page daemon is less 3417 * likely to reclaim it. 3418 */ 3419 vm_page_aflag_set(m, PGA_REFERENCED); 3420 vm_page_lock(m); 3421 VM_OBJECT_WUNLOCK(object); 3422 vm_page_busy_sleep(m, "pgrbwt"); 3423 VM_OBJECT_WLOCK(object); 3424 goto retrylookup; 3425 } else { 3426 if ((allocflags & VM_ALLOC_WIRED) != 0) { 3427 vm_page_lock(m); 3428 vm_page_wire(m); 3429 vm_page_unlock(m); 3430 } 3431 if ((allocflags & 3432 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 3433 vm_page_xbusy(m); 3434 if ((allocflags & VM_ALLOC_SBUSY) != 0) 3435 vm_page_sbusy(m); 3436 return (m); 3437 } 3438 } 3439 m = vm_page_alloc(object, pindex, allocflags); 3440 if (m == NULL) { 3441 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 3442 return (NULL); 3443 VM_OBJECT_WUNLOCK(object); 3444 VM_WAIT; 3445 VM_OBJECT_WLOCK(object); 3446 goto retrylookup; 3447 } else if (m->valid != 0) 3448 return (m); 3449 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 3450 pmap_zero_page(m); 3451 return (m); 3452} 3453 3454/* 3455 * Mapping function for valid or dirty bits in a page. 3456 * 3457 * Inputs are required to range within a page. 3458 */ 3459vm_page_bits_t 3460vm_page_bits(int base, int size) 3461{ 3462 int first_bit; 3463 int last_bit; 3464 3465 KASSERT( 3466 base + size <= PAGE_SIZE, 3467 ("vm_page_bits: illegal base/size %d/%d", base, size) 3468 ); 3469 3470 if (size == 0) /* handle degenerate case */ 3471 return (0); 3472 3473 first_bit = base >> DEV_BSHIFT; 3474 last_bit = (base + size - 1) >> DEV_BSHIFT; 3475 3476 return (((vm_page_bits_t)2 << last_bit) - 3477 ((vm_page_bits_t)1 << first_bit)); 3478} 3479 3480/* 3481 * vm_page_set_valid_range: 3482 * 3483 * Sets portions of a page valid. The arguments are expected 3484 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 3485 * of any partial chunks touched by the range. The invalid portion of 3486 * such chunks will be zeroed. 3487 * 3488 * (base + size) must be less then or equal to PAGE_SIZE. 3489 */ 3490void 3491vm_page_set_valid_range(vm_page_t m, int base, int size) 3492{ 3493 int endoff, frag; 3494 3495 VM_OBJECT_ASSERT_WLOCKED(m->object); 3496 if (size == 0) /* handle degenerate case */ 3497 return; 3498 3499 /* 3500 * If the base is not DEV_BSIZE aligned and the valid 3501 * bit is clear, we have to zero out a portion of the 3502 * first block. 3503 */ 3504 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 3505 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 3506 pmap_zero_page_area(m, frag, base - frag); 3507 3508 /* 3509 * If the ending offset is not DEV_BSIZE aligned and the 3510 * valid bit is clear, we have to zero out a portion of 3511 * the last block. 3512 */ 3513 endoff = base + size; 3514 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 3515 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 3516 pmap_zero_page_area(m, endoff, 3517 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 3518 3519 /* 3520 * Assert that no previously invalid block that is now being validated 3521 * is already dirty. 3522 */ 3523 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 3524 ("vm_page_set_valid_range: page %p is dirty", m)); 3525 3526 /* 3527 * Set valid bits inclusive of any overlap. 3528 */ 3529 m->valid |= vm_page_bits(base, size); 3530} 3531 3532/* 3533 * Clear the given bits from the specified page's dirty field. 3534 */ 3535static __inline void 3536vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) 3537{ 3538 uintptr_t addr; 3539#if PAGE_SIZE < 16384 3540 int shift; 3541#endif 3542 3543 /* 3544 * If the object is locked and the page is neither exclusive busy nor 3545 * write mapped, then the page's dirty field cannot possibly be 3546 * set by a concurrent pmap operation. 3547 */ 3548 VM_OBJECT_ASSERT_WLOCKED(m->object); 3549 if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) 3550 m->dirty &= ~pagebits; 3551 else { 3552 /* 3553 * The pmap layer can call vm_page_dirty() without 3554 * holding a distinguished lock. The combination of 3555 * the object's lock and an atomic operation suffice 3556 * to guarantee consistency of the page dirty field. 3557 * 3558 * For PAGE_SIZE == 32768 case, compiler already 3559 * properly aligns the dirty field, so no forcible 3560 * alignment is needed. Only require existence of 3561 * atomic_clear_64 when page size is 32768. 3562 */ 3563 addr = (uintptr_t)&m->dirty; 3564#if PAGE_SIZE == 32768 3565 atomic_clear_64((uint64_t *)addr, pagebits); 3566#elif PAGE_SIZE == 16384 3567 atomic_clear_32((uint32_t *)addr, pagebits); 3568#else /* PAGE_SIZE <= 8192 */ 3569 /* 3570 * Use a trick to perform a 32-bit atomic on the 3571 * containing aligned word, to not depend on the existence 3572 * of atomic_clear_{8, 16}. 3573 */ 3574 shift = addr & (sizeof(uint32_t) - 1); 3575#if BYTE_ORDER == BIG_ENDIAN 3576 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY; 3577#else 3578 shift *= NBBY; 3579#endif 3580 addr &= ~(sizeof(uint32_t) - 1); 3581 atomic_clear_32((uint32_t *)addr, pagebits << shift); 3582#endif /* PAGE_SIZE */ 3583 } 3584} 3585 3586/* 3587 * vm_page_set_validclean: 3588 * 3589 * Sets portions of a page valid and clean. The arguments are expected 3590 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 3591 * of any partial chunks touched by the range. The invalid portion of 3592 * such chunks will be zero'd. 3593 * 3594 * (base + size) must be less then or equal to PAGE_SIZE. 3595 */ 3596void 3597vm_page_set_validclean(vm_page_t m, int base, int size) 3598{ 3599 vm_page_bits_t oldvalid, pagebits; 3600 int endoff, frag; 3601 3602 VM_OBJECT_ASSERT_WLOCKED(m->object); 3603 if (size == 0) /* handle degenerate case */ 3604 return; 3605 3606 /* 3607 * If the base is not DEV_BSIZE aligned and the valid 3608 * bit is clear, we have to zero out a portion of the 3609 * first block. 3610 */ 3611 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 3612 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) 3613 pmap_zero_page_area(m, frag, base - frag); 3614 3615 /* 3616 * If the ending offset is not DEV_BSIZE aligned and the 3617 * valid bit is clear, we have to zero out a portion of 3618 * the last block. 3619 */ 3620 endoff = base + size; 3621 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 3622 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) 3623 pmap_zero_page_area(m, endoff, 3624 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 3625 3626 /* 3627 * Set valid, clear dirty bits. If validating the entire 3628 * page we can safely clear the pmap modify bit. We also 3629 * use this opportunity to clear the VPO_NOSYNC flag. If a process 3630 * takes a write fault on a MAP_NOSYNC memory area the flag will 3631 * be set again. 3632 * 3633 * We set valid bits inclusive of any overlap, but we can only 3634 * clear dirty bits for DEV_BSIZE chunks that are fully within 3635 * the range. 3636 */ 3637 oldvalid = m->valid; 3638 pagebits = vm_page_bits(base, size); 3639 m->valid |= pagebits; 3640#if 0 /* NOT YET */ 3641 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 3642 frag = DEV_BSIZE - frag; 3643 base += frag; 3644 size -= frag; 3645 if (size < 0) 3646 size = 0; 3647 } 3648 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 3649#endif 3650 if (base == 0 && size == PAGE_SIZE) { 3651 /* 3652 * The page can only be modified within the pmap if it is 3653 * mapped, and it can only be mapped if it was previously 3654 * fully valid. 3655 */ 3656 if (oldvalid == VM_PAGE_BITS_ALL) 3657 /* 3658 * Perform the pmap_clear_modify() first. Otherwise, 3659 * a concurrent pmap operation, such as 3660 * pmap_protect(), could clear a modification in the 3661 * pmap and set the dirty field on the page before 3662 * pmap_clear_modify() had begun and after the dirty 3663 * field was cleared here. 3664 */ 3665 pmap_clear_modify(m); 3666 m->dirty = 0; 3667 m->oflags &= ~VPO_NOSYNC; 3668 } else if (oldvalid != VM_PAGE_BITS_ALL) 3669 m->dirty &= ~pagebits; 3670 else 3671 vm_page_clear_dirty_mask(m, pagebits); 3672} 3673 3674void 3675vm_page_clear_dirty(vm_page_t m, int base, int size) 3676{ 3677 3678 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 3679} 3680 3681/* 3682 * vm_page_set_invalid: 3683 * 3684 * Invalidates DEV_BSIZE'd chunks within a page. Both the 3685 * valid and dirty bits for the effected areas are cleared. 3686 */ 3687void 3688vm_page_set_invalid(vm_page_t m, int base, int size) 3689{ 3690 vm_page_bits_t bits; 3691 vm_object_t object; 3692 3693 object = m->object; 3694 VM_OBJECT_ASSERT_WLOCKED(object); 3695 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + 3696 size >= object->un_pager.vnp.vnp_size) 3697 bits = VM_PAGE_BITS_ALL; 3698 else 3699 bits = vm_page_bits(base, size); 3700 if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL && 3701 bits != 0) 3702 pmap_remove_all(m); 3703 KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) || 3704 !pmap_page_is_mapped(m), 3705 ("vm_page_set_invalid: page %p is mapped", m)); 3706 m->valid &= ~bits; 3707 m->dirty &= ~bits; 3708} 3709 3710/* 3711 * vm_page_zero_invalid() 3712 * 3713 * The kernel assumes that the invalid portions of a page contain 3714 * garbage, but such pages can be mapped into memory by user code. 3715 * When this occurs, we must zero out the non-valid portions of the 3716 * page so user code sees what it expects. 3717 * 3718 * Pages are most often semi-valid when the end of a file is mapped 3719 * into memory and the file's size is not page aligned. 3720 */ 3721void 3722vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 3723{ 3724 int b; 3725 int i; 3726 3727 VM_OBJECT_ASSERT_WLOCKED(m->object); 3728 /* 3729 * Scan the valid bits looking for invalid sections that 3730 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the 3731 * valid bit may be set ) have already been zeroed by 3732 * vm_page_set_validclean(). 3733 */ 3734 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 3735 if (i == (PAGE_SIZE / DEV_BSIZE) || 3736 (m->valid & ((vm_page_bits_t)1 << i))) { 3737 if (i > b) { 3738 pmap_zero_page_area(m, 3739 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 3740 } 3741 b = i + 1; 3742 } 3743 } 3744 3745 /* 3746 * setvalid is TRUE when we can safely set the zero'd areas 3747 * as being valid. We can do this if there are no cache consistancy 3748 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 3749 */ 3750 if (setvalid) 3751 m->valid = VM_PAGE_BITS_ALL; 3752} 3753 3754/* 3755 * vm_page_is_valid: 3756 * 3757 * Is (partial) page valid? Note that the case where size == 0 3758 * will return FALSE in the degenerate case where the page is 3759 * entirely invalid, and TRUE otherwise. 3760 */ 3761int 3762vm_page_is_valid(vm_page_t m, int base, int size) 3763{ 3764 vm_page_bits_t bits; 3765 3766 VM_OBJECT_ASSERT_LOCKED(m->object); 3767 bits = vm_page_bits(base, size); 3768 return (m->valid != 0 && (m->valid & bits) == bits); 3769} 3770 3771/* 3772 * vm_page_ps_is_valid: 3773 * 3774 * Returns TRUE if the entire (super)page is valid and FALSE otherwise. 3775 */ 3776boolean_t 3777vm_page_ps_is_valid(vm_page_t m) 3778{ 3779 int i, npages; 3780 3781 VM_OBJECT_ASSERT_LOCKED(m->object); 3782 npages = atop(pagesizes[m->psind]); 3783 3784 /* 3785 * The physically contiguous pages that make up a superpage, i.e., a 3786 * page with a page size index ("psind") greater than zero, will 3787 * occupy adjacent entries in vm_page_array[]. 3788 */ 3789 for (i = 0; i < npages; i++) { 3790 if (m[i].valid != VM_PAGE_BITS_ALL) 3791 return (FALSE); 3792 } 3793 return (TRUE); 3794} 3795 3796/* 3797 * Set the page's dirty bits if the page is modified. 3798 */ 3799void 3800vm_page_test_dirty(vm_page_t m) 3801{ 3802 3803 VM_OBJECT_ASSERT_WLOCKED(m->object); 3804 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 3805 vm_page_dirty(m); 3806} 3807 3808void 3809vm_page_lock_KBI(vm_page_t m, const char *file, int line) 3810{ 3811 3812 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); 3813} 3814 3815void 3816vm_page_unlock_KBI(vm_page_t m, const char *file, int line) 3817{ 3818 3819 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); 3820} 3821 3822int 3823vm_page_trylock_KBI(vm_page_t m, const char *file, int line) 3824{ 3825 3826 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); 3827} 3828 3829#if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) 3830void 3831vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line) 3832{ 3833 3834 vm_page_lock_assert_KBI(m, MA_OWNED, file, line); 3835} 3836 3837void 3838vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) 3839{ 3840 3841 mtx_assert_(vm_page_lockptr(m), a, file, line); 3842} 3843#endif 3844 3845#ifdef INVARIANTS 3846void 3847vm_page_object_lock_assert(vm_page_t m) 3848{ 3849 3850 /* 3851 * Certain of the page's fields may only be modified by the 3852 * holder of the containing object's lock or the exclusive busy. 3853 * holder. Unfortunately, the holder of the write busy is 3854 * not recorded, and thus cannot be checked here. 3855 */ 3856 if (m->object != NULL && !vm_page_xbusied(m)) 3857 VM_OBJECT_ASSERT_WLOCKED(m->object); 3858} 3859 3860void 3861vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits) 3862{ 3863 3864 if ((bits & PGA_WRITEABLE) == 0) 3865 return; 3866 3867 /* 3868 * The PGA_WRITEABLE flag can only be set if the page is 3869 * managed, is exclusively busied or the object is locked. 3870 * Currently, this flag is only set by pmap_enter(). 3871 */ 3872 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 3873 ("PGA_WRITEABLE on unmanaged page")); 3874 if (!vm_page_xbusied(m)) 3875 VM_OBJECT_ASSERT_LOCKED(m->object); 3876} 3877#endif 3878 3879#include "opt_ddb.h" 3880#ifdef DDB 3881#include <sys/kernel.h> 3882 3883#include <ddb/ddb.h> 3884 3885DB_SHOW_COMMAND(page, vm_page_print_page_info) 3886{ 3887 db_printf("vm_cnt.v_free_count: %d\n", vm_cnt.v_free_count); 3888 db_printf("vm_cnt.v_cache_count: %d\n", vm_cnt.v_cache_count); 3889 db_printf("vm_cnt.v_inactive_count: %d\n", vm_cnt.v_inactive_count); 3890 db_printf("vm_cnt.v_active_count: %d\n", vm_cnt.v_active_count); 3891 db_printf("vm_cnt.v_wire_count: %d\n", vm_cnt.v_wire_count); 3892 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved); 3893 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min); 3894 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target); 3895 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target); 3896} 3897 3898DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 3899{ 3900 int dom; 3901 3902 db_printf("pq_free %d pq_cache %d\n", 3903 vm_cnt.v_free_count, vm_cnt.v_cache_count); 3904 for (dom = 0; dom < vm_ndomains; dom++) { 3905 db_printf( 3906 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pass %d\n", 3907 dom, 3908 vm_dom[dom].vmd_page_count, 3909 vm_dom[dom].vmd_free_count, 3910 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, 3911 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, 3912 vm_dom[dom].vmd_pass); 3913 } 3914} 3915 3916DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) 3917{ 3918 vm_page_t m; 3919 boolean_t phys; 3920 3921 if (!have_addr) { 3922 db_printf("show pginfo addr\n"); 3923 return; 3924 } 3925 3926 phys = strchr(modif, 'p') != NULL; 3927 if (phys) 3928 m = PHYS_TO_VM_PAGE(addr); 3929 else 3930 m = (vm_page_t)addr; 3931 db_printf( 3932 "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n" 3933 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", 3934 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, 3935 m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags, 3936 m->flags, m->act_count, m->busy_lock, m->valid, m->dirty); 3937} 3938#endif /* DDB */ 3939