1/*- 2 * Copyright (c) 1991 Regents of the University of California. 3 * All rights reserved. 4 * Copyright (c) 1998 Matthew Dillon. All Rights Reserved. 5 * 6 * This code is derived from software contributed to Berkeley by 7 * The Mach Operating System project at Carnegie-Mellon University. 8 * 9 * Redistribution and use in source and binary forms, with or without 10 * modification, are permitted provided that the following conditions 11 * are met: 12 * 1. Redistributions of source code must retain the above copyright 13 * notice, this list of conditions and the following disclaimer. 14 * 2. Redistributions in binary form must reproduce the above copyright 15 * notice, this list of conditions and the following disclaimer in the 16 * documentation and/or other materials provided with the distribution. 17 * 4. Neither the name of the University nor the names of its contributors 18 * may be used to endorse or promote products derived from this software 19 * without specific prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 24 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 31 * SUCH DAMAGE. 32 * 33 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91 34 */ 35 36/*- 37 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 38 * All rights reserved. 39 * 40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 41 * 42 * Permission to use, copy, modify and distribute this software and 43 * its documentation is hereby granted, provided that both the copyright 44 * notice and this permission notice appear in all copies of the 45 * software, derivative works or modified versions, and any portions 46 * thereof, and that both notices appear in supporting documentation. 47 * 48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 51 * 52 * Carnegie Mellon requests users of this software to return to 53 * 54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 55 * School of Computer Science 56 * Carnegie Mellon University 57 * Pittsburgh PA 15213-3890 58 * 59 * any improvements or extensions that they make and grant Carnegie the 60 * rights to redistribute these changes. 61 */ 62 63/* 64 * GENERAL RULES ON VM_PAGE MANIPULATION 65 * 66 * - A page queue lock is required when adding or removing a page from a 67 * page queue regardless of other locks or the busy state of a page. 68 * 69 * * In general, no thread besides the page daemon can acquire or 70 * hold more than one page queue lock at a time. 71 * 72 * * The page daemon can acquire and hold any pair of page queue 73 * locks in any order. 74 * 75 * - The object lock is required when inserting or removing 76 * pages from an object (vm_page_insert() or vm_page_remove()). 77 * 78 */ 79 80/* 81 * Resident memory management module. 82 */ 83 84#include <sys/cdefs.h> 85__FBSDID("$FreeBSD: stable/10/sys/vm/vm_page.c 320190 2017-06-21 14:39:31Z jhb $"); 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/malloc.h> 95#include <sys/mman.h> 96#include <sys/msgbuf.h> 97#include <sys/mutex.h> 98#include <sys/proc.h> 99#include <sys/rwlock.h> 100#include <sys/sysctl.h> 101#include <sys/vmmeter.h> 102#include <sys/vnode.h> 103 104#include <vm/vm.h> 105#include <vm/pmap.h> 106#include <vm/vm_param.h> 107#include <vm/vm_kern.h> 108#include <vm/vm_object.h> 109#include <vm/vm_page.h> 110#include <vm/vm_pageout.h> 111#include <vm/vm_pager.h> 112#include <vm/vm_phys.h> 113#include <vm/vm_radix.h> 114#include <vm/vm_reserv.h> 115#include <vm/vm_extern.h> 116#include <vm/uma.h> 117#include <vm/uma_int.h> 118 119#include <machine/md_var.h> 120 121/* 122 * Associated with page of user-allocatable memory is a 123 * page structure. 124 */ 125 126struct vm_domain vm_dom[MAXMEMDOM]; 127struct mtx_padalign vm_page_queue_free_mtx; 128 129struct mtx_padalign pa_lock[PA_LOCK_COUNT]; 130 131vm_page_t vm_page_array; 132long vm_page_array_size; 133long first_page; 134int vm_page_zero_count; 135 136static int boot_pages = UMA_BOOT_PAGES; 137TUNABLE_INT("vm.boot_pages", &boot_pages); 138SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RD, &boot_pages, 0, 139 "number of pages allocated for bootstrapping the VM system"); 140 141static int pa_tryrelock_restart; 142SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD, 143 &pa_tryrelock_restart, 0, "Number of tryrelock restarts"); 144 145static uma_zone_t fakepg_zone; 146 147static struct vnode *vm_page_alloc_init(vm_page_t m); 148static void vm_page_cache_turn_free(vm_page_t m); 149static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits); 150static void vm_page_enqueue(int queue, vm_page_t m); 151static void vm_page_init_fakepg(void *dummy); 152static int vm_page_insert_after(vm_page_t m, vm_object_t object, 153 vm_pindex_t pindex, vm_page_t mpred); 154static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object, 155 vm_page_t mpred); 156 157SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init_fakepg, NULL); 158 159static void 160vm_page_init_fakepg(void *dummy) 161{ 162 163 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL, 164 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM); 165} 166 167/* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */ 168#if PAGE_SIZE == 32768 169#ifdef CTASSERT 170CTASSERT(sizeof(u_long) >= 8); 171#endif 172#endif 173 174/* 175 * Try to acquire a physical address lock while a pmap is locked. If we 176 * fail to trylock we unlock and lock the pmap directly and cache the 177 * locked pa in *locked. The caller should then restart their loop in case 178 * the virtual to physical mapping has changed. 179 */ 180int 181vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked) 182{ 183 vm_paddr_t lockpa; 184 185 lockpa = *locked; 186 *locked = pa; 187 if (lockpa) { 188 PA_LOCK_ASSERT(lockpa, MA_OWNED); 189 if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa)) 190 return (0); 191 PA_UNLOCK(lockpa); 192 } 193 if (PA_TRYLOCK(pa)) 194 return (0); 195 PMAP_UNLOCK(pmap); 196 atomic_add_int(&pa_tryrelock_restart, 1); 197 PA_LOCK(pa); 198 PMAP_LOCK(pmap); 199 return (EAGAIN); 200} 201 202/* 203 * vm_set_page_size: 204 * 205 * Sets the page size, perhaps based upon the memory 206 * size. Must be called before any use of page-size 207 * dependent functions. 208 */ 209void 210vm_set_page_size(void) 211{ 212 if (cnt.v_page_size == 0) 213 cnt.v_page_size = PAGE_SIZE; 214 if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0) 215 panic("vm_set_page_size: page size not a power of two"); 216} 217 218/* 219 * vm_page_blacklist_lookup: 220 * 221 * See if a physical address in this page has been listed 222 * in the blacklist tunable. Entries in the tunable are 223 * separated by spaces or commas. If an invalid integer is 224 * encountered then the rest of the string is skipped. 225 */ 226static int 227vm_page_blacklist_lookup(char *list, vm_paddr_t pa) 228{ 229 vm_paddr_t bad; 230 char *cp, *pos; 231 232 for (pos = list; *pos != '\0'; pos = cp) { 233 bad = strtoq(pos, &cp, 0); 234 if (*cp != '\0') { 235 if (*cp == ' ' || *cp == ',') { 236 cp++; 237 if (cp == pos) 238 continue; 239 } else 240 break; 241 } 242 if (pa == trunc_page(bad)) 243 return (1); 244 } 245 return (0); 246} 247 248static void 249vm_page_domain_init(struct vm_domain *vmd) 250{ 251 struct vm_pagequeue *pq; 252 int i; 253 254 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) = 255 "vm inactive pagequeue"; 256 *__DECONST(u_int **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_vcnt) = 257 &cnt.v_inactive_count; 258 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) = 259 "vm active pagequeue"; 260 *__DECONST(u_int **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_vcnt) = 261 &cnt.v_active_count; 262 vmd->vmd_page_count = 0; 263 vmd->vmd_free_count = 0; 264 vmd->vmd_segs = 0; 265 vmd->vmd_oom = FALSE; 266 vmd->vmd_pass = 0; 267 for (i = 0; i < PQ_COUNT; i++) { 268 pq = &vmd->vmd_pagequeues[i]; 269 TAILQ_INIT(&pq->pq_pl); 270 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue", 271 MTX_DEF | MTX_DUPOK); 272 } 273} 274 275/* 276 * vm_page_startup: 277 * 278 * Initializes the resident memory module. Allocates physical memory for 279 * bootstrapping UMA and some data structures that are used to manage 280 * physical pages. Initializes these structures, and populates the free 281 * page queues. 282 */ 283vm_offset_t 284vm_page_startup(vm_offset_t vaddr) 285{ 286 vm_offset_t mapped; 287 vm_paddr_t high_avail, low_avail, page_range, size; 288 vm_paddr_t new_end; 289 int i; 290 vm_paddr_t pa; 291 vm_paddr_t last_pa; 292 char *list; 293 294 /* the biggest memory array is the second group of pages */ 295 vm_paddr_t end; 296 vm_paddr_t biggestsize; 297 int biggestone; 298 299 biggestsize = 0; 300 biggestone = 0; 301 vaddr = round_page(vaddr); 302 303 for (i = 0; phys_avail[i + 1]; i += 2) { 304 phys_avail[i] = round_page(phys_avail[i]); 305 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 306 } 307 308#ifdef XEN 309 /* 310 * There is no obvious reason why i386 PV Xen needs vm_page structs 311 * created for these pseudo-physical addresses. XXX 312 */ 313 vm_phys_add_seg(0, phys_avail[0]); 314#endif 315 316 for (i = 0; phys_avail[i + 1]; i += 2) { 317 size = phys_avail[i + 1] - phys_avail[i]; 318 if (size > biggestsize) { 319 biggestone = i; 320 biggestsize = size; 321 } 322 } 323 324 end = phys_avail[biggestone+1]; 325 326 /* 327 * Initialize the page and queue locks. 328 */ 329 mtx_init(&vm_page_queue_free_mtx, "vm page free queue", NULL, MTX_DEF); 330 for (i = 0; i < PA_LOCK_COUNT; i++) 331 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF); 332 for (i = 0; i < vm_ndomains; i++) 333 vm_page_domain_init(&vm_dom[i]); 334 335 /* 336 * Allocate memory for use when boot strapping the kernel memory 337 * allocator. 338 */ 339 new_end = end - (boot_pages * UMA_SLAB_SIZE); 340 new_end = trunc_page(new_end); 341 mapped = pmap_map(&vaddr, new_end, end, 342 VM_PROT_READ | VM_PROT_WRITE); 343 bzero((void *)mapped, end - new_end); 344 uma_startup((void *)mapped, boot_pages); 345 346#if defined(__amd64__) || defined(__i386__) || defined(__arm__) || \ 347 defined(__mips__) 348 /* 349 * Allocate a bitmap to indicate that a random physical page 350 * needs to be included in a minidump. 351 * 352 * The amd64 port needs this to indicate which direct map pages 353 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 354 * 355 * However, i386 still needs this workspace internally within the 356 * minidump code. In theory, they are not needed on i386, but are 357 * included should the sf_buf code decide to use them. 358 */ 359 last_pa = 0; 360 for (i = 0; dump_avail[i + 1] != 0; i += 2) 361 if (dump_avail[i + 1] > last_pa) 362 last_pa = dump_avail[i + 1]; 363 page_range = last_pa / PAGE_SIZE; 364 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 365 new_end -= vm_page_dump_size; 366 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, 367 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); 368 bzero((void *)vm_page_dump, vm_page_dump_size); 369#endif 370#if defined(__amd64__) || defined(__mips__) 371 /* 372 * Include the UMA bootstrap pages and vm_page_dump in a crash dump. 373 * When pmap_map() uses the direct map, they are not automatically 374 * included. 375 */ 376 for (pa = new_end; pa < end; pa += PAGE_SIZE) 377 dump_add_page(pa); 378#endif 379 phys_avail[biggestone + 1] = new_end; 380#ifdef __amd64__ 381 /* 382 * Request that the physical pages underlying the message buffer be 383 * included in a crash dump. Since the message buffer is accessed 384 * through the direct map, they are not automatically included. 385 */ 386 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); 387 last_pa = pa + round_page(msgbufsize); 388 while (pa < last_pa) { 389 dump_add_page(pa); 390 pa += PAGE_SIZE; 391 } 392#endif 393 /* 394 * Compute the number of pages of memory that will be available for 395 * use, taking into account the overhead of a page structure per page. 396 * In other words, solve 397 * "available physical memory" - round_page(page_range * 398 * sizeof(struct vm_page)) = page_range * PAGE_SIZE 399 * for page_range. 400 */ 401 low_avail = phys_avail[0]; 402 high_avail = phys_avail[1]; 403 for (i = 0; i < vm_phys_nsegs; i++) { 404 if (vm_phys_segs[i].start < low_avail) 405 low_avail = vm_phys_segs[i].start; 406 if (vm_phys_segs[i].end > high_avail) 407 high_avail = vm_phys_segs[i].end; 408 } 409 /* Skip the first chunk. It is already accounted for. */ 410 for (i = 2; phys_avail[i + 1] != 0; i += 2) { 411 if (phys_avail[i] < low_avail) 412 low_avail = phys_avail[i]; 413 if (phys_avail[i + 1] > high_avail) 414 high_avail = phys_avail[i + 1]; 415 } 416 first_page = low_avail / PAGE_SIZE; 417#ifdef VM_PHYSSEG_SPARSE 418 size = 0; 419 for (i = 0; i < vm_phys_nsegs; i++) 420 size += vm_phys_segs[i].end - vm_phys_segs[i].start; 421 for (i = 0; phys_avail[i + 1] != 0; i += 2) 422 size += phys_avail[i + 1] - phys_avail[i]; 423#elif defined(VM_PHYSSEG_DENSE) 424 size = high_avail - low_avail; 425#else 426#error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 427#endif 428 429#ifdef VM_PHYSSEG_DENSE 430 /* 431 * In the VM_PHYSSEG_DENSE case, the number of pages can account for 432 * the overhead of a page structure per page only if vm_page_array is 433 * allocated from the last physical memory chunk. Otherwise, we must 434 * allocate page structures representing the physical memory 435 * underlying vm_page_array, even though they will not be used. 436 */ 437 if (new_end != high_avail) 438 page_range = size / PAGE_SIZE; 439 else 440#endif 441 { 442 page_range = size / (PAGE_SIZE + sizeof(struct vm_page)); 443 444 /* 445 * If the partial bytes remaining are large enough for 446 * a page (PAGE_SIZE) without a corresponding 447 * 'struct vm_page', then new_end will contain an 448 * extra page after subtracting the length of the VM 449 * page array. Compensate by subtracting an extra 450 * page from new_end. 451 */ 452 if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) { 453 if (new_end == high_avail) 454 high_avail -= PAGE_SIZE; 455 new_end -= PAGE_SIZE; 456 } 457 } 458 end = new_end; 459 460 /* 461 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 462 * However, because this page is allocated from KVM, out-of-bounds 463 * accesses using the direct map will not be trapped. 464 */ 465 vaddr += PAGE_SIZE; 466 467 /* 468 * Allocate physical memory for the page structures, and map it. 469 */ 470 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 471 mapped = pmap_map(&vaddr, new_end, end, 472 VM_PROT_READ | VM_PROT_WRITE); 473 vm_page_array = (vm_page_t) mapped; 474#if VM_NRESERVLEVEL > 0 475 /* 476 * Allocate physical memory for the reservation management system's 477 * data structures, and map it. 478 */ 479 if (high_avail == end) 480 high_avail = new_end; 481 new_end = vm_reserv_startup(&vaddr, new_end, high_avail); 482#endif 483#if defined(__amd64__) || defined(__mips__) 484 /* 485 * Include vm_page_array and vm_reserv_array in a crash dump. 486 */ 487 for (pa = new_end; pa < end; pa += PAGE_SIZE) 488 dump_add_page(pa); 489#endif 490 phys_avail[biggestone + 1] = new_end; 491 492 /* 493 * Add physical memory segments corresponding to the available 494 * physical pages. 495 */ 496 for (i = 0; phys_avail[i + 1] != 0; i += 2) 497 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]); 498 499 /* 500 * Clear all of the page structures 501 */ 502 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 503 for (i = 0; i < page_range; i++) 504 vm_page_array[i].order = VM_NFREEORDER; 505 vm_page_array_size = page_range; 506 507 /* 508 * Initialize the physical memory allocator. 509 */ 510 vm_phys_init(); 511 512 /* 513 * Add every available physical page that is not blacklisted to 514 * the free lists. 515 */ 516 cnt.v_page_count = 0; 517 cnt.v_free_count = 0; 518 list = getenv("vm.blacklist"); 519 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 520 pa = phys_avail[i]; 521 last_pa = phys_avail[i + 1]; 522 while (pa < last_pa) { 523 if (list != NULL && 524 vm_page_blacklist_lookup(list, pa)) 525 printf("Skipping page with pa 0x%jx\n", 526 (uintmax_t)pa); 527 else 528 vm_phys_add_page(pa); 529 pa += PAGE_SIZE; 530 } 531 } 532 freeenv(list); 533#if VM_NRESERVLEVEL > 0 534 /* 535 * Initialize the reservation management system. 536 */ 537 vm_reserv_init(); 538#endif 539 return (vaddr); 540} 541 542void 543vm_page_reference(vm_page_t m) 544{ 545 546 vm_page_aflag_set(m, PGA_REFERENCED); 547} 548 549/* 550 * vm_page_busy_downgrade: 551 * 552 * Downgrade an exclusive busy page into a single shared busy page. 553 */ 554void 555vm_page_busy_downgrade(vm_page_t m) 556{ 557 u_int x; 558 bool locked; 559 560 vm_page_assert_xbusied(m); 561 locked = mtx_owned(vm_page_lockptr(m)); 562 563 for (;;) { 564 x = m->busy_lock; 565 x &= VPB_BIT_WAITERS; 566 if (x != 0 && !locked) 567 vm_page_lock(m); 568 if (atomic_cmpset_rel_int(&m->busy_lock, 569 VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1))) 570 break; 571 if (x != 0 && !locked) 572 vm_page_unlock(m); 573 } 574 if (x != 0) { 575 wakeup(m); 576 if (!locked) 577 vm_page_unlock(m); 578 } 579} 580 581/* 582 * vm_page_sbusied: 583 * 584 * Return a positive value if the page is shared busied, 0 otherwise. 585 */ 586int 587vm_page_sbusied(vm_page_t m) 588{ 589 u_int x; 590 591 x = m->busy_lock; 592 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED); 593} 594 595/* 596 * vm_page_sunbusy: 597 * 598 * Shared unbusy a page. 599 */ 600void 601vm_page_sunbusy(vm_page_t m) 602{ 603 u_int x; 604 605 vm_page_assert_sbusied(m); 606 607 for (;;) { 608 x = m->busy_lock; 609 if (VPB_SHARERS(x) > 1) { 610 if (atomic_cmpset_int(&m->busy_lock, x, 611 x - VPB_ONE_SHARER)) 612 break; 613 continue; 614 } 615 if ((x & VPB_BIT_WAITERS) == 0) { 616 KASSERT(x == VPB_SHARERS_WORD(1), 617 ("vm_page_sunbusy: invalid lock state")); 618 if (atomic_cmpset_int(&m->busy_lock, 619 VPB_SHARERS_WORD(1), VPB_UNBUSIED)) 620 break; 621 continue; 622 } 623 KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS), 624 ("vm_page_sunbusy: invalid lock state for waiters")); 625 626 vm_page_lock(m); 627 if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) { 628 vm_page_unlock(m); 629 continue; 630 } 631 wakeup(m); 632 vm_page_unlock(m); 633 break; 634 } 635} 636 637/* 638 * vm_page_busy_sleep: 639 * 640 * Sleep and release the page lock, using the page pointer as wchan. 641 * This is used to implement the hard-path of busying mechanism. 642 * 643 * The given page must be locked. 644 * 645 * If nonshared is true, sleep only if the page is xbusy. 646 */ 647void 648vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared) 649{ 650 u_int x; 651 652 vm_page_assert_locked(m); 653 654 x = m->busy_lock; 655 if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) || 656 ((x & VPB_BIT_WAITERS) == 0 && 657 !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) { 658 vm_page_unlock(m); 659 return; 660 } 661 msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0); 662} 663 664/* 665 * vm_page_trysbusy: 666 * 667 * Try to shared busy a page. 668 * If the operation succeeds 1 is returned otherwise 0. 669 * The operation never sleeps. 670 */ 671int 672vm_page_trysbusy(vm_page_t m) 673{ 674 u_int x; 675 676 for (;;) { 677 x = m->busy_lock; 678 if ((x & VPB_BIT_SHARED) == 0) 679 return (0); 680 if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER)) 681 return (1); 682 } 683} 684 685/* 686 * vm_page_xunbusy_hard: 687 * 688 * Called after the first try the exclusive unbusy of a page failed. 689 * It is assumed that the waiters bit is on. 690 */ 691void 692vm_page_xunbusy_hard(vm_page_t m) 693{ 694 695 vm_page_assert_xbusied(m); 696 697 vm_page_lock(m); 698 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); 699 wakeup(m); 700 vm_page_unlock(m); 701} 702 703/* 704 * vm_page_flash: 705 * 706 * Wakeup anyone waiting for the page. 707 * The ownership bits do not change. 708 * 709 * The given page must be locked. 710 */ 711void 712vm_page_flash(vm_page_t m) 713{ 714 u_int x; 715 716 vm_page_lock_assert(m, MA_OWNED); 717 718 for (;;) { 719 x = m->busy_lock; 720 if ((x & VPB_BIT_WAITERS) == 0) 721 return; 722 if (atomic_cmpset_int(&m->busy_lock, x, 723 x & (~VPB_BIT_WAITERS))) 724 break; 725 } 726 wakeup(m); 727} 728 729/* 730 * Keep page from being freed by the page daemon 731 * much of the same effect as wiring, except much lower 732 * overhead and should be used only for *very* temporary 733 * holding ("wiring"). 734 */ 735void 736vm_page_hold(vm_page_t mem) 737{ 738 739 vm_page_lock_assert(mem, MA_OWNED); 740 mem->hold_count++; 741} 742 743void 744vm_page_unhold(vm_page_t mem) 745{ 746 747 vm_page_lock_assert(mem, MA_OWNED); 748 KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!")); 749 --mem->hold_count; 750 if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0) 751 vm_page_free_toq(mem); 752} 753 754/* 755 * vm_page_unhold_pages: 756 * 757 * Unhold each of the pages that is referenced by the given array. 758 */ 759void 760vm_page_unhold_pages(vm_page_t *ma, int count) 761{ 762 struct mtx *mtx, *new_mtx; 763 764 mtx = NULL; 765 for (; count != 0; count--) { 766 /* 767 * Avoid releasing and reacquiring the same page lock. 768 */ 769 new_mtx = vm_page_lockptr(*ma); 770 if (mtx != new_mtx) { 771 if (mtx != NULL) 772 mtx_unlock(mtx); 773 mtx = new_mtx; 774 mtx_lock(mtx); 775 } 776 vm_page_unhold(*ma); 777 ma++; 778 } 779 if (mtx != NULL) 780 mtx_unlock(mtx); 781} 782 783vm_page_t 784PHYS_TO_VM_PAGE(vm_paddr_t pa) 785{ 786 vm_page_t m; 787 788#ifdef VM_PHYSSEG_SPARSE 789 m = vm_phys_paddr_to_vm_page(pa); 790 if (m == NULL) 791 m = vm_phys_fictitious_to_vm_page(pa); 792 return (m); 793#elif defined(VM_PHYSSEG_DENSE) 794 long pi; 795 796 pi = atop(pa); 797 if (pi >= first_page && (pi - first_page) < vm_page_array_size) { 798 m = &vm_page_array[pi - first_page]; 799 return (m); 800 } 801 return (vm_phys_fictitious_to_vm_page(pa)); 802#else 803#error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 804#endif 805} 806 807/* 808 * vm_page_getfake: 809 * 810 * Create a fictitious page with the specified physical address and 811 * memory attribute. The memory attribute is the only the machine- 812 * dependent aspect of a fictitious page that must be initialized. 813 */ 814vm_page_t 815vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr) 816{ 817 vm_page_t m; 818 819 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO); 820 vm_page_initfake(m, paddr, memattr); 821 return (m); 822} 823 824void 825vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 826{ 827 828 if ((m->flags & PG_FICTITIOUS) != 0) { 829 /* 830 * The page's memattr might have changed since the 831 * previous initialization. Update the pmap to the 832 * new memattr. 833 */ 834 goto memattr; 835 } 836 m->phys_addr = paddr; 837 m->queue = PQ_NONE; 838 /* Fictitious pages don't use "segind". */ 839 m->flags = PG_FICTITIOUS; 840 /* Fictitious pages don't use "order" or "pool". */ 841 m->oflags = VPO_UNMANAGED; 842 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 843 m->wire_count = 1; 844 pmap_page_init(m); 845memattr: 846 pmap_page_set_memattr(m, memattr); 847} 848 849/* 850 * vm_page_putfake: 851 * 852 * Release a fictitious page. 853 */ 854void 855vm_page_putfake(vm_page_t m) 856{ 857 858 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m)); 859 KASSERT((m->flags & PG_FICTITIOUS) != 0, 860 ("vm_page_putfake: bad page %p", m)); 861 uma_zfree(fakepg_zone, m); 862} 863 864/* 865 * vm_page_updatefake: 866 * 867 * Update the given fictitious page to the specified physical address and 868 * memory attribute. 869 */ 870void 871vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 872{ 873 874 KASSERT((m->flags & PG_FICTITIOUS) != 0, 875 ("vm_page_updatefake: bad page %p", m)); 876 m->phys_addr = paddr; 877 pmap_page_set_memattr(m, memattr); 878} 879 880/* 881 * vm_page_free: 882 * 883 * Free a page. 884 */ 885void 886vm_page_free(vm_page_t m) 887{ 888 889 m->flags &= ~PG_ZERO; 890 vm_page_free_toq(m); 891} 892 893/* 894 * vm_page_free_zero: 895 * 896 * Free a page to the zerod-pages queue 897 */ 898void 899vm_page_free_zero(vm_page_t m) 900{ 901 902 m->flags |= PG_ZERO; 903 vm_page_free_toq(m); 904} 905 906/* 907 * Unbusy and handle the page queueing for a page from the VOP_GETPAGES() 908 * array which is not the request page. 909 */ 910void 911vm_page_readahead_finish(vm_page_t m) 912{ 913 914 if (m->valid != 0) { 915 /* 916 * Since the page is not the requested page, whether 917 * it should be activated or deactivated is not 918 * obvious. Empirical results have shown that 919 * deactivating the page is usually the best choice, 920 * unless the page is wanted by another thread. 921 */ 922 vm_page_lock(m); 923 if ((m->busy_lock & VPB_BIT_WAITERS) != 0) 924 vm_page_activate(m); 925 else 926 vm_page_deactivate(m); 927 vm_page_unlock(m); 928 vm_page_xunbusy(m); 929 } else { 930 /* 931 * Free the completely invalid page. Such page state 932 * occurs due to the short read operation which did 933 * not covered our page at all, or in case when a read 934 * error happens. 935 */ 936 vm_page_lock(m); 937 vm_page_free(m); 938 vm_page_unlock(m); 939 } 940} 941 942/* 943 * vm_page_sleep_if_busy: 944 * 945 * Sleep and release the page queues lock if the page is busied. 946 * Returns TRUE if the thread slept. 947 * 948 * The given page must be unlocked and object containing it must 949 * be locked. 950 */ 951int 952vm_page_sleep_if_busy(vm_page_t m, const char *msg) 953{ 954 vm_object_t obj; 955 956 vm_page_lock_assert(m, MA_NOTOWNED); 957 VM_OBJECT_ASSERT_WLOCKED(m->object); 958 959 if (vm_page_busied(m)) { 960 /* 961 * The page-specific object must be cached because page 962 * identity can change during the sleep, causing the 963 * re-lock of a different object. 964 * It is assumed that a reference to the object is already 965 * held by the callers. 966 */ 967 obj = m->object; 968 vm_page_lock(m); 969 VM_OBJECT_WUNLOCK(obj); 970 vm_page_busy_sleep(m, msg, false); 971 VM_OBJECT_WLOCK(obj); 972 return (TRUE); 973 } 974 return (FALSE); 975} 976 977/* 978 * vm_page_dirty_KBI: [ internal use only ] 979 * 980 * Set all bits in the page's dirty field. 981 * 982 * The object containing the specified page must be locked if the 983 * call is made from the machine-independent layer. 984 * 985 * See vm_page_clear_dirty_mask(). 986 * 987 * This function should only be called by vm_page_dirty(). 988 */ 989void 990vm_page_dirty_KBI(vm_page_t m) 991{ 992 993 /* These assertions refer to this operation by its public name. */ 994 KASSERT((m->flags & PG_CACHED) == 0, 995 ("vm_page_dirty: page in cache!")); 996 KASSERT(!VM_PAGE_IS_FREE(m), 997 ("vm_page_dirty: page is free!")); 998 KASSERT(m->valid == VM_PAGE_BITS_ALL, 999 ("vm_page_dirty: page is invalid!")); 1000 m->dirty = VM_PAGE_BITS_ALL; 1001} 1002 1003/* 1004 * vm_page_insert: [ internal use only ] 1005 * 1006 * Inserts the given mem entry into the object and object list. 1007 * 1008 * The object must be locked. 1009 */ 1010int 1011vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 1012{ 1013 vm_page_t mpred; 1014 1015 VM_OBJECT_ASSERT_WLOCKED(object); 1016 mpred = vm_radix_lookup_le(&object->rtree, pindex); 1017 return (vm_page_insert_after(m, object, pindex, mpred)); 1018} 1019 1020/* 1021 * vm_page_insert_after: 1022 * 1023 * Inserts the page "m" into the specified object at offset "pindex". 1024 * 1025 * The page "mpred" must immediately precede the offset "pindex" within 1026 * the specified object. 1027 * 1028 * The object must be locked. 1029 */ 1030static int 1031vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, 1032 vm_page_t mpred) 1033{ 1034 vm_page_t msucc; 1035 1036 VM_OBJECT_ASSERT_WLOCKED(object); 1037 KASSERT(m->object == NULL, 1038 ("vm_page_insert_after: page already inserted")); 1039 if (mpred != NULL) { 1040 KASSERT(mpred->object == object, 1041 ("vm_page_insert_after: object doesn't contain mpred")); 1042 KASSERT(mpred->pindex < pindex, 1043 ("vm_page_insert_after: mpred doesn't precede pindex")); 1044 msucc = TAILQ_NEXT(mpred, listq); 1045 } else 1046 msucc = TAILQ_FIRST(&object->memq); 1047 if (msucc != NULL) 1048 KASSERT(msucc->pindex > pindex, 1049 ("vm_page_insert_after: msucc doesn't succeed pindex")); 1050 1051 /* 1052 * Record the object/offset pair in this page 1053 */ 1054 m->object = object; 1055 m->pindex = pindex; 1056 1057 /* 1058 * Now link into the object's ordered list of backed pages. 1059 */ 1060 if (vm_radix_insert(&object->rtree, m)) { 1061 m->object = NULL; 1062 m->pindex = 0; 1063 return (1); 1064 } 1065 vm_page_insert_radixdone(m, object, mpred); 1066 return (0); 1067} 1068 1069/* 1070 * vm_page_insert_radixdone: 1071 * 1072 * Complete page "m" insertion into the specified object after the 1073 * radix trie hooking. 1074 * 1075 * The page "mpred" must precede the offset "m->pindex" within the 1076 * specified object. 1077 * 1078 * The object must be locked. 1079 */ 1080static void 1081vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred) 1082{ 1083 1084 VM_OBJECT_ASSERT_WLOCKED(object); 1085 KASSERT(object != NULL && m->object == object, 1086 ("vm_page_insert_radixdone: page %p has inconsistent object", m)); 1087 if (mpred != NULL) { 1088 KASSERT(mpred->object == object, 1089 ("vm_page_insert_after: object doesn't contain mpred")); 1090 KASSERT(mpred->pindex < m->pindex, 1091 ("vm_page_insert_after: mpred doesn't precede pindex")); 1092 } 1093 1094 if (mpred != NULL) 1095 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq); 1096 else 1097 TAILQ_INSERT_HEAD(&object->memq, m, listq); 1098 1099 /* 1100 * Show that the object has one more resident page. 1101 */ 1102 object->resident_page_count++; 1103 1104 /* 1105 * Hold the vnode until the last page is released. 1106 */ 1107 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 1108 vhold(object->handle); 1109 1110 /* 1111 * Since we are inserting a new and possibly dirty page, 1112 * update the object's OBJ_MIGHTBEDIRTY flag. 1113 */ 1114 if (pmap_page_is_write_mapped(m)) 1115 vm_object_set_writeable_dirty(object); 1116} 1117 1118/* 1119 * vm_page_remove: 1120 * 1121 * Removes the given mem entry from the object/offset-page 1122 * table and the object page list, but do not invalidate/terminate 1123 * the backing store. 1124 * 1125 * The object must be locked. The page must be locked if it is managed. 1126 */ 1127void 1128vm_page_remove(vm_page_t m) 1129{ 1130 vm_object_t object; 1131 boolean_t lockacq; 1132 1133 if ((m->oflags & VPO_UNMANAGED) == 0) 1134 vm_page_lock_assert(m, MA_OWNED); 1135 if ((object = m->object) == NULL) 1136 return; 1137 VM_OBJECT_ASSERT_WLOCKED(object); 1138 if (vm_page_xbusied(m)) { 1139 lockacq = FALSE; 1140 if ((m->oflags & VPO_UNMANAGED) != 0 && 1141 !mtx_owned(vm_page_lockptr(m))) { 1142 lockacq = TRUE; 1143 vm_page_lock(m); 1144 } 1145 vm_page_flash(m); 1146 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); 1147 if (lockacq) 1148 vm_page_unlock(m); 1149 } 1150 1151 /* 1152 * Now remove from the object's list of backed pages. 1153 */ 1154 vm_radix_remove(&object->rtree, m->pindex); 1155 TAILQ_REMOVE(&object->memq, m, listq); 1156 1157 /* 1158 * And show that the object has one fewer resident page. 1159 */ 1160 object->resident_page_count--; 1161 1162 /* 1163 * The vnode may now be recycled. 1164 */ 1165 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 1166 vdrop(object->handle); 1167 1168 m->object = NULL; 1169} 1170 1171/* 1172 * vm_page_lookup: 1173 * 1174 * Returns the page associated with the object/offset 1175 * pair specified; if none is found, NULL is returned. 1176 * 1177 * The object must be locked. 1178 */ 1179vm_page_t 1180vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1181{ 1182 1183 VM_OBJECT_ASSERT_LOCKED(object); 1184 return (vm_radix_lookup(&object->rtree, pindex)); 1185} 1186 1187/* 1188 * vm_page_find_least: 1189 * 1190 * Returns the page associated with the object with least pindex 1191 * greater than or equal to the parameter pindex, or NULL. 1192 * 1193 * The object must be locked. 1194 */ 1195vm_page_t 1196vm_page_find_least(vm_object_t object, vm_pindex_t pindex) 1197{ 1198 vm_page_t m; 1199 1200 VM_OBJECT_ASSERT_LOCKED(object); 1201 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex) 1202 m = vm_radix_lookup_ge(&object->rtree, pindex); 1203 return (m); 1204} 1205 1206/* 1207 * Returns the given page's successor (by pindex) within the object if it is 1208 * resident; if none is found, NULL is returned. 1209 * 1210 * The object must be locked. 1211 */ 1212vm_page_t 1213vm_page_next(vm_page_t m) 1214{ 1215 vm_page_t next; 1216 1217 VM_OBJECT_ASSERT_WLOCKED(m->object); 1218 if ((next = TAILQ_NEXT(m, listq)) != NULL) { 1219 MPASS(next->object == m->object); 1220 if (next->pindex != m->pindex + 1) 1221 next = NULL; 1222 } 1223 return (next); 1224} 1225 1226/* 1227 * Returns the given page's predecessor (by pindex) within the object if it is 1228 * resident; if none is found, NULL is returned. 1229 * 1230 * The object must be locked. 1231 */ 1232vm_page_t 1233vm_page_prev(vm_page_t m) 1234{ 1235 vm_page_t prev; 1236 1237 VM_OBJECT_ASSERT_WLOCKED(m->object); 1238 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) { 1239 MPASS(prev->object == m->object); 1240 if (prev->pindex != m->pindex - 1) 1241 prev = NULL; 1242 } 1243 return (prev); 1244} 1245 1246/* 1247 * Uses the page mnew as a replacement for an existing page at index 1248 * pindex which must be already present in the object. 1249 * 1250 * The existing page must not be on a paging queue. 1251 */ 1252vm_page_t 1253vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex) 1254{ 1255 vm_page_t mold, mpred; 1256 1257 VM_OBJECT_ASSERT_WLOCKED(object); 1258 1259 /* 1260 * This function mostly follows vm_page_insert() and 1261 * vm_page_remove() without the radix, object count and vnode 1262 * dance. Double check such functions for more comments. 1263 */ 1264 mpred = vm_radix_lookup(&object->rtree, pindex); 1265 KASSERT(mpred != NULL, 1266 ("vm_page_replace: replacing page not present with pindex")); 1267 mpred = TAILQ_PREV(mpred, respgs, listq); 1268 if (mpred != NULL) 1269 KASSERT(mpred->pindex < pindex, 1270 ("vm_page_insert_after: mpred doesn't precede pindex")); 1271 1272 mnew->object = object; 1273 mnew->pindex = pindex; 1274 mold = vm_radix_replace(&object->rtree, mnew); 1275 KASSERT(mold->queue == PQ_NONE, 1276 ("vm_page_replace: mold is on a paging queue")); 1277 1278 /* Detach the old page from the resident tailq. */ 1279 TAILQ_REMOVE(&object->memq, mold, listq); 1280 1281 mold->object = NULL; 1282 vm_page_xunbusy(mold); 1283 1284 /* Insert the new page in the resident tailq. */ 1285 if (mpred != NULL) 1286 TAILQ_INSERT_AFTER(&object->memq, mpred, mnew, listq); 1287 else 1288 TAILQ_INSERT_HEAD(&object->memq, mnew, listq); 1289 if (pmap_page_is_write_mapped(mnew)) 1290 vm_object_set_writeable_dirty(object); 1291 return (mold); 1292} 1293 1294/* 1295 * vm_page_rename: 1296 * 1297 * Move the given memory entry from its 1298 * current object to the specified target object/offset. 1299 * 1300 * Note: swap associated with the page must be invalidated by the move. We 1301 * have to do this for several reasons: (1) we aren't freeing the 1302 * page, (2) we are dirtying the page, (3) the VM system is probably 1303 * moving the page from object A to B, and will then later move 1304 * the backing store from A to B and we can't have a conflict. 1305 * 1306 * Note: we *always* dirty the page. It is necessary both for the 1307 * fact that we moved it, and because we may be invalidating 1308 * swap. If the page is on the cache, we have to deactivate it 1309 * or vm_page_dirty() will panic. Dirty pages are not allowed 1310 * on the cache. 1311 * 1312 * The objects must be locked. 1313 */ 1314int 1315vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1316{ 1317 vm_page_t mpred; 1318 vm_pindex_t opidx; 1319 1320 VM_OBJECT_ASSERT_WLOCKED(new_object); 1321 1322 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex); 1323 KASSERT(mpred == NULL || mpred->pindex != new_pindex, 1324 ("vm_page_rename: pindex already renamed")); 1325 1326 /* 1327 * Create a custom version of vm_page_insert() which does not depend 1328 * by m_prev and can cheat on the implementation aspects of the 1329 * function. 1330 */ 1331 opidx = m->pindex; 1332 m->pindex = new_pindex; 1333 if (vm_radix_insert(&new_object->rtree, m)) { 1334 m->pindex = opidx; 1335 return (1); 1336 } 1337 1338 /* 1339 * The operation cannot fail anymore. The removal must happen before 1340 * the listq iterator is tainted. 1341 */ 1342 m->pindex = opidx; 1343 vm_page_lock(m); 1344 vm_page_remove(m); 1345 1346 /* Return back to the new pindex to complete vm_page_insert(). */ 1347 m->pindex = new_pindex; 1348 m->object = new_object; 1349 vm_page_unlock(m); 1350 vm_page_insert_radixdone(m, new_object, mpred); 1351 vm_page_dirty(m); 1352 return (0); 1353} 1354 1355/* 1356 * Convert all of the given object's cached pages that have a 1357 * pindex within the given range into free pages. If the value 1358 * zero is given for "end", then the range's upper bound is 1359 * infinity. If the given object is backed by a vnode and it 1360 * transitions from having one or more cached pages to none, the 1361 * vnode's hold count is reduced. 1362 */ 1363void 1364vm_page_cache_free(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 1365{ 1366 vm_page_t m; 1367 boolean_t empty; 1368 1369 mtx_lock(&vm_page_queue_free_mtx); 1370 if (__predict_false(vm_radix_is_empty(&object->cache))) { 1371 mtx_unlock(&vm_page_queue_free_mtx); 1372 return; 1373 } 1374 while ((m = vm_radix_lookup_ge(&object->cache, start)) != NULL) { 1375 if (end != 0 && m->pindex >= end) 1376 break; 1377 vm_radix_remove(&object->cache, m->pindex); 1378 vm_page_cache_turn_free(m); 1379 } 1380 empty = vm_radix_is_empty(&object->cache); 1381 mtx_unlock(&vm_page_queue_free_mtx); 1382 if (object->type == OBJT_VNODE && empty) 1383 vdrop(object->handle); 1384} 1385 1386/* 1387 * Returns the cached page that is associated with the given 1388 * object and offset. If, however, none exists, returns NULL. 1389 * 1390 * The free page queue must be locked. 1391 */ 1392static inline vm_page_t 1393vm_page_cache_lookup(vm_object_t object, vm_pindex_t pindex) 1394{ 1395 1396 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1397 return (vm_radix_lookup(&object->cache, pindex)); 1398} 1399 1400/* 1401 * Remove the given cached page from its containing object's 1402 * collection of cached pages. 1403 * 1404 * The free page queue must be locked. 1405 */ 1406static void 1407vm_page_cache_remove(vm_page_t m) 1408{ 1409 1410 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1411 KASSERT((m->flags & PG_CACHED) != 0, 1412 ("vm_page_cache_remove: page %p is not cached", m)); 1413 vm_radix_remove(&m->object->cache, m->pindex); 1414 m->object = NULL; 1415 cnt.v_cache_count--; 1416} 1417 1418/* 1419 * Transfer all of the cached pages with offset greater than or 1420 * equal to 'offidxstart' from the original object's cache to the 1421 * new object's cache. However, any cached pages with offset 1422 * greater than or equal to the new object's size are kept in the 1423 * original object. Initially, the new object's cache must be 1424 * empty. Offset 'offidxstart' in the original object must 1425 * correspond to offset zero in the new object. 1426 * 1427 * The new object must be locked. 1428 */ 1429void 1430vm_page_cache_transfer(vm_object_t orig_object, vm_pindex_t offidxstart, 1431 vm_object_t new_object) 1432{ 1433 vm_page_t m; 1434 1435 /* 1436 * Insertion into an object's collection of cached pages 1437 * requires the object to be locked. In contrast, removal does 1438 * not. 1439 */ 1440 VM_OBJECT_ASSERT_WLOCKED(new_object); 1441 KASSERT(vm_radix_is_empty(&new_object->cache), 1442 ("vm_page_cache_transfer: object %p has cached pages", 1443 new_object)); 1444 mtx_lock(&vm_page_queue_free_mtx); 1445 while ((m = vm_radix_lookup_ge(&orig_object->cache, 1446 offidxstart)) != NULL) { 1447 /* 1448 * Transfer all of the pages with offset greater than or 1449 * equal to 'offidxstart' from the original object's 1450 * cache to the new object's cache. 1451 */ 1452 if ((m->pindex - offidxstart) >= new_object->size) 1453 break; 1454 vm_radix_remove(&orig_object->cache, m->pindex); 1455 /* Update the page's object and offset. */ 1456 m->object = new_object; 1457 m->pindex -= offidxstart; 1458 if (vm_radix_insert(&new_object->cache, m)) 1459 vm_page_cache_turn_free(m); 1460 } 1461 mtx_unlock(&vm_page_queue_free_mtx); 1462} 1463 1464/* 1465 * Returns TRUE if a cached page is associated with the given object and 1466 * offset, and FALSE otherwise. 1467 * 1468 * The object must be locked. 1469 */ 1470boolean_t 1471vm_page_is_cached(vm_object_t object, vm_pindex_t pindex) 1472{ 1473 vm_page_t m; 1474 1475 /* 1476 * Insertion into an object's collection of cached pages requires the 1477 * object to be locked. Therefore, if the object is locked and the 1478 * object's collection is empty, there is no need to acquire the free 1479 * page queues lock in order to prove that the specified page doesn't 1480 * exist. 1481 */ 1482 VM_OBJECT_ASSERT_WLOCKED(object); 1483 if (__predict_true(vm_object_cache_is_empty(object))) 1484 return (FALSE); 1485 mtx_lock(&vm_page_queue_free_mtx); 1486 m = vm_page_cache_lookup(object, pindex); 1487 mtx_unlock(&vm_page_queue_free_mtx); 1488 return (m != NULL); 1489} 1490 1491/* 1492 * vm_page_alloc: 1493 * 1494 * Allocate and return a page that is associated with the specified 1495 * object and offset pair. By default, this page is exclusive busied. 1496 * 1497 * The caller must always specify an allocation class. 1498 * 1499 * allocation classes: 1500 * VM_ALLOC_NORMAL normal process request 1501 * VM_ALLOC_SYSTEM system *really* needs a page 1502 * VM_ALLOC_INTERRUPT interrupt time request 1503 * 1504 * optional allocation flags: 1505 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1506 * intends to allocate 1507 * VM_ALLOC_IFCACHED return page only if it is cached 1508 * VM_ALLOC_IFNOTCACHED return NULL, do not reactivate if the page 1509 * is cached 1510 * VM_ALLOC_NOBUSY do not exclusive busy the page 1511 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1512 * VM_ALLOC_NOOBJ page is not associated with an object and 1513 * should not be exclusive busy 1514 * VM_ALLOC_SBUSY shared busy the allocated page 1515 * VM_ALLOC_WIRED wire the allocated page 1516 * VM_ALLOC_ZERO prefer a zeroed page 1517 * 1518 * This routine may not sleep. 1519 */ 1520vm_page_t 1521vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 1522{ 1523 struct vnode *vp = NULL; 1524 vm_object_t m_object; 1525 vm_page_t m, mpred; 1526 int flags, req_class; 1527 1528 mpred = 0; /* XXX: pacify gcc */ 1529 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 1530 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 1531 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 1532 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 1533 ("vm_page_alloc: inconsistent object(%p)/req(%x)", (void *)object, 1534 req)); 1535 if (object != NULL) 1536 VM_OBJECT_ASSERT_WLOCKED(object); 1537 1538 req_class = req & VM_ALLOC_CLASS_MASK; 1539 1540 /* 1541 * The page daemon is allowed to dig deeper into the free page list. 1542 */ 1543 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1544 req_class = VM_ALLOC_SYSTEM; 1545 1546 if (object != NULL) { 1547 mpred = vm_radix_lookup_le(&object->rtree, pindex); 1548 KASSERT(mpred == NULL || mpred->pindex != pindex, 1549 ("vm_page_alloc: pindex already allocated")); 1550 } 1551 1552 /* 1553 * The page allocation request can came from consumers which already 1554 * hold the free page queue mutex, like vm_page_insert() in 1555 * vm_page_cache(). 1556 */ 1557 mtx_lock_flags(&vm_page_queue_free_mtx, MTX_RECURSE); 1558 if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved || 1559 (req_class == VM_ALLOC_SYSTEM && 1560 cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) || 1561 (req_class == VM_ALLOC_INTERRUPT && 1562 cnt.v_free_count + cnt.v_cache_count > 0)) { 1563 /* 1564 * Allocate from the free queue if the number of free pages 1565 * exceeds the minimum for the request class. 1566 */ 1567 if (object != NULL && 1568 (m = vm_page_cache_lookup(object, pindex)) != NULL) { 1569 if ((req & VM_ALLOC_IFNOTCACHED) != 0) { 1570 mtx_unlock(&vm_page_queue_free_mtx); 1571 return (NULL); 1572 } 1573 if (vm_phys_unfree_page(m)) 1574 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0); 1575#if VM_NRESERVLEVEL > 0 1576 else if (!vm_reserv_reactivate_page(m)) 1577#else 1578 else 1579#endif 1580 panic("vm_page_alloc: cache page %p is missing" 1581 " from the free queue", m); 1582 } else if ((req & VM_ALLOC_IFCACHED) != 0) { 1583 mtx_unlock(&vm_page_queue_free_mtx); 1584 return (NULL); 1585#if VM_NRESERVLEVEL > 0 1586 } else if (object == NULL || (object->flags & (OBJ_COLORED | 1587 OBJ_FICTITIOUS)) != OBJ_COLORED || (m = 1588 vm_reserv_alloc_page(object, pindex, mpred)) == NULL) { 1589#else 1590 } else { 1591#endif 1592 m = vm_phys_alloc_pages(object != NULL ? 1593 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0); 1594#if VM_NRESERVLEVEL > 0 1595 if (m == NULL && vm_reserv_reclaim_inactive()) { 1596 m = vm_phys_alloc_pages(object != NULL ? 1597 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 1598 0); 1599 } 1600#endif 1601 } 1602 } else { 1603 /* 1604 * Not allocatable, give up. 1605 */ 1606 mtx_unlock(&vm_page_queue_free_mtx); 1607 atomic_add_int(&vm_pageout_deficit, 1608 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 1609 pagedaemon_wakeup(); 1610 return (NULL); 1611 } 1612 1613 /* 1614 * At this point we had better have found a good page. 1615 */ 1616 KASSERT(m != NULL, ("vm_page_alloc: missing page")); 1617 KASSERT(m->queue == PQ_NONE, 1618 ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue)); 1619 KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m)); 1620 KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m)); 1621 KASSERT(!vm_page_busied(m), ("vm_page_alloc: page %p is busy", m)); 1622 KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m)); 1623 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 1624 ("vm_page_alloc: page %p has unexpected memattr %d", m, 1625 pmap_page_get_memattr(m))); 1626 if ((m->flags & PG_CACHED) != 0) { 1627 KASSERT((m->flags & PG_ZERO) == 0, 1628 ("vm_page_alloc: cached page %p is PG_ZERO", m)); 1629 KASSERT(m->valid != 0, 1630 ("vm_page_alloc: cached page %p is invalid", m)); 1631 if (m->object == object && m->pindex == pindex) 1632 cnt.v_reactivated++; 1633 else 1634 m->valid = 0; 1635 m_object = m->object; 1636 vm_page_cache_remove(m); 1637 if (m_object->type == OBJT_VNODE && 1638 vm_object_cache_is_empty(m_object)) 1639 vp = m_object->handle; 1640 } else { 1641 KASSERT(VM_PAGE_IS_FREE(m), 1642 ("vm_page_alloc: page %p is not free", m)); 1643 KASSERT(m->valid == 0, 1644 ("vm_page_alloc: free page %p is valid", m)); 1645 vm_phys_freecnt_adj(m, -1); 1646 } 1647 1648 /* 1649 * Only the PG_ZERO flag is inherited. The PG_CACHED or PG_FREE flag 1650 * must be cleared before the free page queues lock is released. 1651 */ 1652 flags = 0; 1653 if (m->flags & PG_ZERO) { 1654 vm_page_zero_count--; 1655 if (req & VM_ALLOC_ZERO) 1656 flags = PG_ZERO; 1657 } 1658 if (req & VM_ALLOC_NODUMP) 1659 flags |= PG_NODUMP; 1660 m->flags = flags; 1661 mtx_unlock(&vm_page_queue_free_mtx); 1662 m->aflags = 0; 1663 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? 1664 VPO_UNMANAGED : 0; 1665 m->busy_lock = VPB_UNBUSIED; 1666 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) 1667 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 1668 if ((req & VM_ALLOC_SBUSY) != 0) 1669 m->busy_lock = VPB_SHARERS_WORD(1); 1670 if (req & VM_ALLOC_WIRED) { 1671 /* 1672 * The page lock is not required for wiring a page until that 1673 * page is inserted into the object. 1674 */ 1675 atomic_add_int(&cnt.v_wire_count, 1); 1676 m->wire_count = 1; 1677 } 1678 m->act_count = 0; 1679 1680 if (object != NULL) { 1681 if (vm_page_insert_after(m, object, pindex, mpred)) { 1682 /* See the comment below about hold count. */ 1683 if (vp != NULL) 1684 vdrop(vp); 1685 pagedaemon_wakeup(); 1686 if (req & VM_ALLOC_WIRED) { 1687 atomic_subtract_int(&cnt.v_wire_count, 1); 1688 m->wire_count = 0; 1689 } 1690 m->object = NULL; 1691 m->oflags = VPO_UNMANAGED; 1692 m->busy_lock = VPB_UNBUSIED; 1693 vm_page_free(m); 1694 return (NULL); 1695 } 1696 1697 /* Ignore device objects; the pager sets "memattr" for them. */ 1698 if (object->memattr != VM_MEMATTR_DEFAULT && 1699 (object->flags & OBJ_FICTITIOUS) == 0) 1700 pmap_page_set_memattr(m, object->memattr); 1701 } else 1702 m->pindex = pindex; 1703 1704 /* 1705 * The following call to vdrop() must come after the above call 1706 * to vm_page_insert() in case both affect the same object and 1707 * vnode. Otherwise, the affected vnode's hold count could 1708 * temporarily become zero. 1709 */ 1710 if (vp != NULL) 1711 vdrop(vp); 1712 1713 /* 1714 * Don't wakeup too often - wakeup the pageout daemon when 1715 * we would be nearly out of memory. 1716 */ 1717 if (vm_paging_needed()) 1718 pagedaemon_wakeup(); 1719 1720 return (m); 1721} 1722 1723static void 1724vm_page_alloc_contig_vdrop(struct spglist *lst) 1725{ 1726 1727 while (!SLIST_EMPTY(lst)) { 1728 vdrop((struct vnode *)SLIST_FIRST(lst)-> plinks.s.pv); 1729 SLIST_REMOVE_HEAD(lst, plinks.s.ss); 1730 } 1731} 1732 1733/* 1734 * vm_page_alloc_contig: 1735 * 1736 * Allocate a contiguous set of physical pages of the given size "npages" 1737 * from the free lists. All of the physical pages must be at or above 1738 * the given physical address "low" and below the given physical address 1739 * "high". The given value "alignment" determines the alignment of the 1740 * first physical page in the set. If the given value "boundary" is 1741 * non-zero, then the set of physical pages cannot cross any physical 1742 * address boundary that is a multiple of that value. Both "alignment" 1743 * and "boundary" must be a power of two. 1744 * 1745 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, 1746 * then the memory attribute setting for the physical pages is configured 1747 * to the object's memory attribute setting. Otherwise, the memory 1748 * attribute setting for the physical pages is configured to "memattr", 1749 * overriding the object's memory attribute setting. However, if the 1750 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the 1751 * memory attribute setting for the physical pages cannot be configured 1752 * to VM_MEMATTR_DEFAULT. 1753 * 1754 * The caller must always specify an allocation class. 1755 * 1756 * allocation classes: 1757 * VM_ALLOC_NORMAL normal process request 1758 * VM_ALLOC_SYSTEM system *really* needs a page 1759 * VM_ALLOC_INTERRUPT interrupt time request 1760 * 1761 * optional allocation flags: 1762 * VM_ALLOC_NOBUSY do not exclusive busy the page 1763 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1764 * VM_ALLOC_NOOBJ page is not associated with an object and 1765 * should not be exclusive busy 1766 * VM_ALLOC_SBUSY shared busy the allocated page 1767 * VM_ALLOC_WIRED wire the allocated page 1768 * VM_ALLOC_ZERO prefer a zeroed page 1769 * 1770 * This routine may not sleep. 1771 */ 1772vm_page_t 1773vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, 1774 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 1775 vm_paddr_t boundary, vm_memattr_t memattr) 1776{ 1777 struct vnode *drop; 1778 struct spglist deferred_vdrop_list; 1779 vm_page_t m, m_tmp, m_ret; 1780 u_int flags, oflags; 1781 int req_class; 1782 1783 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 1784 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 1785 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 1786 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 1787 ("vm_page_alloc: inconsistent object(%p)/req(%x)", (void *)object, 1788 req)); 1789 if (object != NULL) { 1790 VM_OBJECT_ASSERT_WLOCKED(object); 1791 KASSERT(object->type == OBJT_PHYS, 1792 ("vm_page_alloc_contig: object %p isn't OBJT_PHYS", 1793 object)); 1794 } 1795 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 1796 req_class = req & VM_ALLOC_CLASS_MASK; 1797 1798 /* 1799 * The page daemon is allowed to dig deeper into the free page list. 1800 */ 1801 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1802 req_class = VM_ALLOC_SYSTEM; 1803 1804 SLIST_INIT(&deferred_vdrop_list); 1805 mtx_lock(&vm_page_queue_free_mtx); 1806 if (cnt.v_free_count + cnt.v_cache_count >= npages + 1807 cnt.v_free_reserved || (req_class == VM_ALLOC_SYSTEM && 1808 cnt.v_free_count + cnt.v_cache_count >= npages + 1809 cnt.v_interrupt_free_min) || (req_class == VM_ALLOC_INTERRUPT && 1810 cnt.v_free_count + cnt.v_cache_count >= npages)) { 1811#if VM_NRESERVLEVEL > 0 1812retry: 1813 if (object == NULL || (object->flags & OBJ_COLORED) == 0 || 1814 (m_ret = vm_reserv_alloc_contig(object, pindex, npages, 1815 low, high, alignment, boundary)) == NULL) 1816#endif 1817 m_ret = vm_phys_alloc_contig(npages, low, high, 1818 alignment, boundary); 1819 } else { 1820 mtx_unlock(&vm_page_queue_free_mtx); 1821 atomic_add_int(&vm_pageout_deficit, npages); 1822 pagedaemon_wakeup(); 1823 return (NULL); 1824 } 1825 if (m_ret != NULL) 1826 for (m = m_ret; m < &m_ret[npages]; m++) { 1827 drop = vm_page_alloc_init(m); 1828 if (drop != NULL) { 1829 /* 1830 * Enqueue the vnode for deferred vdrop(). 1831 */ 1832 m->plinks.s.pv = drop; 1833 SLIST_INSERT_HEAD(&deferred_vdrop_list, m, 1834 plinks.s.ss); 1835 } 1836 } 1837 else { 1838#if VM_NRESERVLEVEL > 0 1839 if (vm_reserv_reclaim_contig(npages, low, high, alignment, 1840 boundary)) 1841 goto retry; 1842#endif 1843 } 1844 mtx_unlock(&vm_page_queue_free_mtx); 1845 if (m_ret == NULL) 1846 return (NULL); 1847 1848 /* 1849 * Initialize the pages. Only the PG_ZERO flag is inherited. 1850 */ 1851 flags = 0; 1852 if ((req & VM_ALLOC_ZERO) != 0) 1853 flags = PG_ZERO; 1854 if ((req & VM_ALLOC_NODUMP) != 0) 1855 flags |= PG_NODUMP; 1856 if ((req & VM_ALLOC_WIRED) != 0) 1857 atomic_add_int(&cnt.v_wire_count, npages); 1858 oflags = VPO_UNMANAGED; 1859 if (object != NULL) { 1860 if (object->memattr != VM_MEMATTR_DEFAULT && 1861 memattr == VM_MEMATTR_DEFAULT) 1862 memattr = object->memattr; 1863 } 1864 for (m = m_ret; m < &m_ret[npages]; m++) { 1865 m->aflags = 0; 1866 m->flags = (m->flags | PG_NODUMP) & flags; 1867 m->busy_lock = VPB_UNBUSIED; 1868 if (object != NULL) { 1869 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 1870 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 1871 if ((req & VM_ALLOC_SBUSY) != 0) 1872 m->busy_lock = VPB_SHARERS_WORD(1); 1873 } 1874 if ((req & VM_ALLOC_WIRED) != 0) 1875 m->wire_count = 1; 1876 /* Unmanaged pages don't use "act_count". */ 1877 m->oflags = oflags; 1878 if (object != NULL) { 1879 if (vm_page_insert(m, object, pindex)) { 1880 vm_page_alloc_contig_vdrop( 1881 &deferred_vdrop_list); 1882 if (vm_paging_needed()) 1883 pagedaemon_wakeup(); 1884 if ((req & VM_ALLOC_WIRED) != 0) 1885 atomic_subtract_int(&cnt.v_wire_count, 1886 npages); 1887 for (m_tmp = m, m = m_ret; 1888 m < &m_ret[npages]; m++) { 1889 if ((req & VM_ALLOC_WIRED) != 0) 1890 m->wire_count = 0; 1891 if (m >= m_tmp) { 1892 m->object = NULL; 1893 m->oflags |= VPO_UNMANAGED; 1894 } 1895 m->busy_lock = VPB_UNBUSIED; 1896 vm_page_free(m); 1897 } 1898 return (NULL); 1899 } 1900 } else 1901 m->pindex = pindex; 1902 if (memattr != VM_MEMATTR_DEFAULT) 1903 pmap_page_set_memattr(m, memattr); 1904 pindex++; 1905 } 1906 vm_page_alloc_contig_vdrop(&deferred_vdrop_list); 1907 if (vm_paging_needed()) 1908 pagedaemon_wakeup(); 1909 return (m_ret); 1910} 1911 1912/* 1913 * Initialize a page that has been freshly dequeued from a freelist. 1914 * The caller has to drop the vnode returned, if it is not NULL. 1915 * 1916 * This function may only be used to initialize unmanaged pages. 1917 * 1918 * To be called with vm_page_queue_free_mtx held. 1919 */ 1920static struct vnode * 1921vm_page_alloc_init(vm_page_t m) 1922{ 1923 struct vnode *drop; 1924 vm_object_t m_object; 1925 1926 KASSERT(m->queue == PQ_NONE, 1927 ("vm_page_alloc_init: page %p has unexpected queue %d", 1928 m, m->queue)); 1929 KASSERT(m->wire_count == 0, 1930 ("vm_page_alloc_init: page %p is wired", m)); 1931 KASSERT(m->hold_count == 0, 1932 ("vm_page_alloc_init: page %p is held", m)); 1933 KASSERT(!vm_page_busied(m), 1934 ("vm_page_alloc_init: page %p is busy", m)); 1935 KASSERT(m->dirty == 0, 1936 ("vm_page_alloc_init: page %p is dirty", m)); 1937 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 1938 ("vm_page_alloc_init: page %p has unexpected memattr %d", 1939 m, pmap_page_get_memattr(m))); 1940 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1941 drop = NULL; 1942 if ((m->flags & PG_CACHED) != 0) { 1943 KASSERT((m->flags & PG_ZERO) == 0, 1944 ("vm_page_alloc_init: cached page %p is PG_ZERO", m)); 1945 m->valid = 0; 1946 m_object = m->object; 1947 vm_page_cache_remove(m); 1948 if (m_object->type == OBJT_VNODE && 1949 vm_object_cache_is_empty(m_object)) 1950 drop = m_object->handle; 1951 } else { 1952 KASSERT(VM_PAGE_IS_FREE(m), 1953 ("vm_page_alloc_init: page %p is not free", m)); 1954 KASSERT(m->valid == 0, 1955 ("vm_page_alloc_init: free page %p is valid", m)); 1956 vm_phys_freecnt_adj(m, -1); 1957 if ((m->flags & PG_ZERO) != 0) 1958 vm_page_zero_count--; 1959 } 1960 /* Don't clear the PG_ZERO flag; we'll need it later. */ 1961 m->flags &= PG_ZERO; 1962 return (drop); 1963} 1964 1965/* 1966 * vm_page_alloc_freelist: 1967 * 1968 * Allocate a physical page from the specified free page list. 1969 * 1970 * The caller must always specify an allocation class. 1971 * 1972 * allocation classes: 1973 * VM_ALLOC_NORMAL normal process request 1974 * VM_ALLOC_SYSTEM system *really* needs a page 1975 * VM_ALLOC_INTERRUPT interrupt time request 1976 * 1977 * optional allocation flags: 1978 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1979 * intends to allocate 1980 * VM_ALLOC_WIRED wire the allocated page 1981 * VM_ALLOC_ZERO prefer a zeroed page 1982 * 1983 * This routine may not sleep. 1984 */ 1985vm_page_t 1986vm_page_alloc_freelist(int flind, int req) 1987{ 1988 struct vnode *drop; 1989 vm_page_t m; 1990 u_int flags; 1991 int req_class; 1992 1993 req_class = req & VM_ALLOC_CLASS_MASK; 1994 1995 /* 1996 * The page daemon is allowed to dig deeper into the free page list. 1997 */ 1998 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1999 req_class = VM_ALLOC_SYSTEM; 2000 2001 /* 2002 * Do not allocate reserved pages unless the req has asked for it. 2003 */ 2004 mtx_lock_flags(&vm_page_queue_free_mtx, MTX_RECURSE); 2005 if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved || 2006 (req_class == VM_ALLOC_SYSTEM && 2007 cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) || 2008 (req_class == VM_ALLOC_INTERRUPT && 2009 cnt.v_free_count + cnt.v_cache_count > 0)) 2010 m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0); 2011 else { 2012 mtx_unlock(&vm_page_queue_free_mtx); 2013 atomic_add_int(&vm_pageout_deficit, 2014 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 2015 pagedaemon_wakeup(); 2016 return (NULL); 2017 } 2018 if (m == NULL) { 2019 mtx_unlock(&vm_page_queue_free_mtx); 2020 return (NULL); 2021 } 2022 drop = vm_page_alloc_init(m); 2023 mtx_unlock(&vm_page_queue_free_mtx); 2024 2025 /* 2026 * Initialize the page. Only the PG_ZERO flag is inherited. 2027 */ 2028 m->aflags = 0; 2029 flags = 0; 2030 if ((req & VM_ALLOC_ZERO) != 0) 2031 flags = PG_ZERO; 2032 m->flags &= flags; 2033 if ((req & VM_ALLOC_WIRED) != 0) { 2034 /* 2035 * The page lock is not required for wiring a page that does 2036 * not belong to an object. 2037 */ 2038 atomic_add_int(&cnt.v_wire_count, 1); 2039 m->wire_count = 1; 2040 } 2041 /* Unmanaged pages don't use "act_count". */ 2042 m->oflags = VPO_UNMANAGED; 2043 if (drop != NULL) 2044 vdrop(drop); 2045 if (vm_paging_needed()) 2046 pagedaemon_wakeup(); 2047 return (m); 2048} 2049 2050/* 2051 * vm_wait: (also see VM_WAIT macro) 2052 * 2053 * Sleep until free pages are available for allocation. 2054 * - Called in various places before memory allocations. 2055 */ 2056void 2057vm_wait(void) 2058{ 2059 2060 mtx_lock(&vm_page_queue_free_mtx); 2061 if (curproc == pageproc) { 2062 vm_pageout_pages_needed = 1; 2063 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx, 2064 PDROP | PSWP, "VMWait", 0); 2065 } else { 2066 if (!vm_pages_needed) { 2067 vm_pages_needed = 1; 2068 wakeup(&vm_pages_needed); 2069 } 2070 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM, 2071 "vmwait", 0); 2072 } 2073} 2074 2075/* 2076 * vm_waitpfault: (also see VM_WAITPFAULT macro) 2077 * 2078 * Sleep until free pages are available for allocation. 2079 * - Called only in vm_fault so that processes page faulting 2080 * can be easily tracked. 2081 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 2082 * processes will be able to grab memory first. Do not change 2083 * this balance without careful testing first. 2084 */ 2085void 2086vm_waitpfault(void) 2087{ 2088 2089 mtx_lock(&vm_page_queue_free_mtx); 2090 if (!vm_pages_needed) { 2091 vm_pages_needed = 1; 2092 wakeup(&vm_pages_needed); 2093 } 2094 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER, 2095 "pfault", 0); 2096} 2097 2098struct vm_pagequeue * 2099vm_page_pagequeue(vm_page_t m) 2100{ 2101 2102 return (&vm_phys_domain(m)->vmd_pagequeues[m->queue]); 2103} 2104 2105/* 2106 * vm_page_dequeue: 2107 * 2108 * Remove the given page from its current page queue. 2109 * 2110 * The page must be locked. 2111 */ 2112void 2113vm_page_dequeue(vm_page_t m) 2114{ 2115 struct vm_pagequeue *pq; 2116 2117 vm_page_lock_assert(m, MA_OWNED); 2118 KASSERT(m->queue != PQ_NONE, 2119 ("vm_page_dequeue: page %p is not queued", m)); 2120 pq = vm_page_pagequeue(m); 2121 vm_pagequeue_lock(pq); 2122 m->queue = PQ_NONE; 2123 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2124 vm_pagequeue_cnt_dec(pq); 2125 vm_pagequeue_unlock(pq); 2126} 2127 2128/* 2129 * vm_page_dequeue_locked: 2130 * 2131 * Remove the given page from its current page queue. 2132 * 2133 * The page and page queue must be locked. 2134 */ 2135void 2136vm_page_dequeue_locked(vm_page_t m) 2137{ 2138 struct vm_pagequeue *pq; 2139 2140 vm_page_lock_assert(m, MA_OWNED); 2141 pq = vm_page_pagequeue(m); 2142 vm_pagequeue_assert_locked(pq); 2143 m->queue = PQ_NONE; 2144 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2145 vm_pagequeue_cnt_dec(pq); 2146} 2147 2148/* 2149 * vm_page_enqueue: 2150 * 2151 * Add the given page to the specified page queue. 2152 * 2153 * The page must be locked. 2154 */ 2155static void 2156vm_page_enqueue(int queue, vm_page_t m) 2157{ 2158 struct vm_pagequeue *pq; 2159 2160 vm_page_lock_assert(m, MA_OWNED); 2161 pq = &vm_phys_domain(m)->vmd_pagequeues[queue]; 2162 vm_pagequeue_lock(pq); 2163 m->queue = queue; 2164 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2165 vm_pagequeue_cnt_inc(pq); 2166 vm_pagequeue_unlock(pq); 2167} 2168 2169/* 2170 * vm_page_requeue: 2171 * 2172 * Move the given page to the tail of its current page queue. 2173 * 2174 * The page must be locked. 2175 */ 2176void 2177vm_page_requeue(vm_page_t m) 2178{ 2179 struct vm_pagequeue *pq; 2180 2181 vm_page_lock_assert(m, MA_OWNED); 2182 KASSERT(m->queue != PQ_NONE, 2183 ("vm_page_requeue: page %p is not queued", m)); 2184 pq = vm_page_pagequeue(m); 2185 vm_pagequeue_lock(pq); 2186 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2187 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2188 vm_pagequeue_unlock(pq); 2189} 2190 2191/* 2192 * vm_page_requeue_locked: 2193 * 2194 * Move the given page to the tail of its current page queue. 2195 * 2196 * The page queue must be locked. 2197 */ 2198void 2199vm_page_requeue_locked(vm_page_t m) 2200{ 2201 struct vm_pagequeue *pq; 2202 2203 KASSERT(m->queue != PQ_NONE, 2204 ("vm_page_requeue_locked: page %p is not queued", m)); 2205 pq = vm_page_pagequeue(m); 2206 vm_pagequeue_assert_locked(pq); 2207 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2208 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2209} 2210 2211/* 2212 * vm_page_activate: 2213 * 2214 * Put the specified page on the active list (if appropriate). 2215 * Ensure that act_count is at least ACT_INIT but do not otherwise 2216 * mess with it. 2217 * 2218 * The page must be locked. 2219 */ 2220void 2221vm_page_activate(vm_page_t m) 2222{ 2223 int queue; 2224 2225 vm_page_lock_assert(m, MA_OWNED); 2226 if ((queue = m->queue) != PQ_ACTIVE) { 2227 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 2228 if (m->act_count < ACT_INIT) 2229 m->act_count = ACT_INIT; 2230 if (queue != PQ_NONE) 2231 vm_page_dequeue(m); 2232 vm_page_enqueue(PQ_ACTIVE, m); 2233 } else 2234 KASSERT(queue == PQ_NONE, 2235 ("vm_page_activate: wired page %p is queued", m)); 2236 } else { 2237 if (m->act_count < ACT_INIT) 2238 m->act_count = ACT_INIT; 2239 } 2240} 2241 2242/* 2243 * vm_page_free_wakeup: 2244 * 2245 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 2246 * routine is called when a page has been added to the cache or free 2247 * queues. 2248 * 2249 * The page queues must be locked. 2250 */ 2251static inline void 2252vm_page_free_wakeup(void) 2253{ 2254 2255 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 2256 /* 2257 * if pageout daemon needs pages, then tell it that there are 2258 * some free. 2259 */ 2260 if (vm_pageout_pages_needed && 2261 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 2262 wakeup(&vm_pageout_pages_needed); 2263 vm_pageout_pages_needed = 0; 2264 } 2265 /* 2266 * wakeup processes that are waiting on memory if we hit a 2267 * high water mark. And wakeup scheduler process if we have 2268 * lots of memory. this process will swapin processes. 2269 */ 2270 if (vm_pages_needed && !vm_page_count_min()) { 2271 vm_pages_needed = 0; 2272 wakeup(&cnt.v_free_count); 2273 } 2274} 2275 2276/* 2277 * Turn a cached page into a free page, by changing its attributes. 2278 * Keep the statistics up-to-date. 2279 * 2280 * The free page queue must be locked. 2281 */ 2282static void 2283vm_page_cache_turn_free(vm_page_t m) 2284{ 2285 2286 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 2287 2288 m->object = NULL; 2289 m->valid = 0; 2290 /* Clear PG_CACHED and set PG_FREE. */ 2291 m->flags ^= PG_CACHED | PG_FREE; 2292 KASSERT((m->flags & (PG_CACHED | PG_FREE)) == PG_FREE, 2293 ("vm_page_cache_free: page %p has inconsistent flags", m)); 2294 cnt.v_cache_count--; 2295 vm_phys_freecnt_adj(m, 1); 2296} 2297 2298/* 2299 * vm_page_free_toq: 2300 * 2301 * Returns the given page to the free list, 2302 * disassociating it with any VM object. 2303 * 2304 * The object must be locked. The page must be locked if it is managed. 2305 */ 2306void 2307vm_page_free_toq(vm_page_t m) 2308{ 2309 2310 if ((m->oflags & VPO_UNMANAGED) == 0) { 2311 vm_page_lock_assert(m, MA_OWNED); 2312 KASSERT(!pmap_page_is_mapped(m), 2313 ("vm_page_free_toq: freeing mapped page %p", m)); 2314 } else 2315 KASSERT(m->queue == PQ_NONE, 2316 ("vm_page_free_toq: unmanaged page %p is queued", m)); 2317 PCPU_INC(cnt.v_tfree); 2318 2319 if (VM_PAGE_IS_FREE(m)) 2320 panic("vm_page_free: freeing free page %p", m); 2321 else if (vm_page_sbusied(m)) 2322 panic("vm_page_free: freeing busy page %p", m); 2323 2324 /* 2325 * Unqueue, then remove page. Note that we cannot destroy 2326 * the page here because we do not want to call the pager's 2327 * callback routine until after we've put the page on the 2328 * appropriate free queue. 2329 */ 2330 vm_page_remque(m); 2331 vm_page_remove(m); 2332 2333 /* 2334 * If fictitious remove object association and 2335 * return, otherwise delay object association removal. 2336 */ 2337 if ((m->flags & PG_FICTITIOUS) != 0) { 2338 return; 2339 } 2340 2341 m->valid = 0; 2342 vm_page_undirty(m); 2343 2344 if (m->wire_count != 0) 2345 panic("vm_page_free: freeing wired page %p", m); 2346 if (m->hold_count != 0) { 2347 m->flags &= ~PG_ZERO; 2348 KASSERT((m->flags & PG_UNHOLDFREE) == 0, 2349 ("vm_page_free: freeing PG_UNHOLDFREE page %p", m)); 2350 m->flags |= PG_UNHOLDFREE; 2351 } else { 2352 /* 2353 * Restore the default memory attribute to the page. 2354 */ 2355 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 2356 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 2357 2358 /* 2359 * Insert the page into the physical memory allocator's 2360 * cache/free page queues. 2361 */ 2362 mtx_lock(&vm_page_queue_free_mtx); 2363 m->flags |= PG_FREE; 2364 vm_phys_freecnt_adj(m, 1); 2365#if VM_NRESERVLEVEL > 0 2366 if (!vm_reserv_free_page(m)) 2367#else 2368 if (TRUE) 2369#endif 2370 vm_phys_free_pages(m, 0); 2371 if ((m->flags & PG_ZERO) != 0) 2372 ++vm_page_zero_count; 2373 else 2374 vm_page_zero_idle_wakeup(); 2375 vm_page_free_wakeup(); 2376 mtx_unlock(&vm_page_queue_free_mtx); 2377 } 2378} 2379 2380/* 2381 * vm_page_wire: 2382 * 2383 * Mark this page as wired down by yet 2384 * another map, removing it from paging queues 2385 * as necessary. 2386 * 2387 * If the page is fictitious, then its wire count must remain one. 2388 * 2389 * The page must be locked. 2390 */ 2391void 2392vm_page_wire(vm_page_t m) 2393{ 2394 2395 /* 2396 * Only bump the wire statistics if the page is not already wired, 2397 * and only unqueue the page if it is on some queue (if it is unmanaged 2398 * it is already off the queues). 2399 */ 2400 vm_page_lock_assert(m, MA_OWNED); 2401 if ((m->flags & PG_FICTITIOUS) != 0) { 2402 KASSERT(m->wire_count == 1, 2403 ("vm_page_wire: fictitious page %p's wire count isn't one", 2404 m)); 2405 return; 2406 } 2407 if (m->wire_count == 0) { 2408 KASSERT((m->oflags & VPO_UNMANAGED) == 0 || 2409 m->queue == PQ_NONE, 2410 ("vm_page_wire: unmanaged page %p is queued", m)); 2411 vm_page_remque(m); 2412 atomic_add_int(&cnt.v_wire_count, 1); 2413 } 2414 m->wire_count++; 2415 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 2416} 2417 2418/* 2419 * vm_page_unwire: 2420 * 2421 * Release one wiring of the specified page, potentially enabling it to be 2422 * paged again. If paging is enabled, then the value of the parameter 2423 * "activate" determines to which queue the page is added. If "activate" is 2424 * non-zero, then the page is added to the active queue. Otherwise, it is 2425 * added to the inactive queue. 2426 * 2427 * However, unless the page belongs to an object, it is not enqueued because 2428 * it cannot be paged out. 2429 * 2430 * If a page is fictitious, then its wire count must always be one. 2431 * 2432 * A managed page must be locked. 2433 */ 2434void 2435vm_page_unwire(vm_page_t m, int activate) 2436{ 2437 2438 if ((m->oflags & VPO_UNMANAGED) == 0) 2439 vm_page_lock_assert(m, MA_OWNED); 2440 if ((m->flags & PG_FICTITIOUS) != 0) { 2441 KASSERT(m->wire_count == 1, 2442 ("vm_page_unwire: fictitious page %p's wire count isn't one", m)); 2443 return; 2444 } 2445 if (m->wire_count > 0) { 2446 m->wire_count--; 2447 if (m->wire_count == 0) { 2448 atomic_subtract_int(&cnt.v_wire_count, 1); 2449 if ((m->oflags & VPO_UNMANAGED) != 0 || 2450 m->object == NULL) 2451 return; 2452 if (!activate) 2453 m->flags &= ~PG_WINATCFLS; 2454 vm_page_enqueue(activate ? PQ_ACTIVE : PQ_INACTIVE, m); 2455 } 2456 } else 2457 panic("vm_page_unwire: page %p's wire count is zero", m); 2458} 2459 2460/* 2461 * Move the specified page to the inactive queue. 2462 * 2463 * Many pages placed on the inactive queue should actually go 2464 * into the cache, but it is difficult to figure out which. What 2465 * we do instead, if the inactive target is well met, is to put 2466 * clean pages at the head of the inactive queue instead of the tail. 2467 * This will cause them to be moved to the cache more quickly and 2468 * if not actively re-referenced, reclaimed more quickly. If we just 2469 * stick these pages at the end of the inactive queue, heavy filesystem 2470 * meta-data accesses can cause an unnecessary paging load on memory bound 2471 * processes. This optimization causes one-time-use metadata to be 2472 * reused more quickly. 2473 * 2474 * Normally athead is 0 resulting in LRU operation. athead is set 2475 * to 1 if we want this page to be 'as if it were placed in the cache', 2476 * except without unmapping it from the process address space. 2477 * 2478 * The page must be locked. 2479 */ 2480static inline void 2481_vm_page_deactivate(vm_page_t m, int athead) 2482{ 2483 struct vm_pagequeue *pq; 2484 int queue; 2485 2486 vm_page_assert_locked(m); 2487 2488 /* 2489 * Ignore if the page is already inactive, unless it is unlikely to be 2490 * reactivated. 2491 */ 2492 if ((queue = m->queue) == PQ_INACTIVE && !athead) 2493 return; 2494 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 2495 pq = &vm_phys_domain(m)->vmd_pagequeues[PQ_INACTIVE]; 2496 /* Avoid multiple acquisitions of the inactive queue lock. */ 2497 if (queue == PQ_INACTIVE) { 2498 vm_pagequeue_lock(pq); 2499 vm_page_dequeue_locked(m); 2500 } else { 2501 if (queue != PQ_NONE) 2502 vm_page_dequeue(m); 2503 m->flags &= ~PG_WINATCFLS; 2504 vm_pagequeue_lock(pq); 2505 } 2506 m->queue = PQ_INACTIVE; 2507 if (athead) 2508 TAILQ_INSERT_HEAD(&pq->pq_pl, m, plinks.q); 2509 else 2510 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2511 vm_pagequeue_cnt_inc(pq); 2512 vm_pagequeue_unlock(pq); 2513 } 2514} 2515 2516/* 2517 * Move the specified page to the inactive queue. 2518 * 2519 * The page must be locked. 2520 */ 2521void 2522vm_page_deactivate(vm_page_t m) 2523{ 2524 2525 _vm_page_deactivate(m, 0); 2526} 2527 2528/* 2529 * vm_page_try_to_cache: 2530 * 2531 * Returns 0 on failure, 1 on success 2532 */ 2533int 2534vm_page_try_to_cache(vm_page_t m) 2535{ 2536 2537 vm_page_lock_assert(m, MA_OWNED); 2538 VM_OBJECT_ASSERT_WLOCKED(m->object); 2539 if (m->dirty || m->hold_count || m->wire_count || 2540 (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m)) 2541 return (0); 2542 pmap_remove_all(m); 2543 if (m->dirty) 2544 return (0); 2545 vm_page_cache(m); 2546 return (1); 2547} 2548 2549/* 2550 * vm_page_try_to_free() 2551 * 2552 * Attempt to free the page. If we cannot free it, we do nothing. 2553 * 1 is returned on success, 0 on failure. 2554 */ 2555int 2556vm_page_try_to_free(vm_page_t m) 2557{ 2558 2559 vm_page_lock_assert(m, MA_OWNED); 2560 if (m->object != NULL) 2561 VM_OBJECT_ASSERT_WLOCKED(m->object); 2562 if (m->dirty || m->hold_count || m->wire_count || 2563 (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m)) 2564 return (0); 2565 pmap_remove_all(m); 2566 if (m->dirty) 2567 return (0); 2568 vm_page_free(m); 2569 return (1); 2570} 2571 2572/* 2573 * vm_page_cache 2574 * 2575 * Put the specified page onto the page cache queue (if appropriate). 2576 * 2577 * The object and page must be locked. 2578 */ 2579void 2580vm_page_cache(vm_page_t m) 2581{ 2582 vm_object_t object; 2583 boolean_t cache_was_empty; 2584 2585 vm_page_lock_assert(m, MA_OWNED); 2586 object = m->object; 2587 VM_OBJECT_ASSERT_WLOCKED(object); 2588 if (vm_page_busied(m) || (m->oflags & VPO_UNMANAGED) || 2589 m->hold_count || m->wire_count) 2590 panic("vm_page_cache: attempting to cache busy page"); 2591 KASSERT(!pmap_page_is_mapped(m), 2592 ("vm_page_cache: page %p is mapped", m)); 2593 KASSERT(m->dirty == 0, ("vm_page_cache: page %p is dirty", m)); 2594 if (m->valid == 0 || object->type == OBJT_DEFAULT || 2595 (object->type == OBJT_SWAP && 2596 !vm_pager_has_page(object, m->pindex, NULL, NULL))) { 2597 /* 2598 * Hypothesis: A cache-elgible page belonging to a 2599 * default object or swap object but without a backing 2600 * store must be zero filled. 2601 */ 2602 vm_page_free(m); 2603 return; 2604 } 2605 KASSERT((m->flags & PG_CACHED) == 0, 2606 ("vm_page_cache: page %p is already cached", m)); 2607 2608 /* 2609 * Remove the page from the paging queues. 2610 */ 2611 vm_page_remque(m); 2612 2613 /* 2614 * Remove the page from the object's collection of resident 2615 * pages. 2616 */ 2617 vm_radix_remove(&object->rtree, m->pindex); 2618 TAILQ_REMOVE(&object->memq, m, listq); 2619 object->resident_page_count--; 2620 2621 /* 2622 * Restore the default memory attribute to the page. 2623 */ 2624 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 2625 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 2626 2627 /* 2628 * Insert the page into the object's collection of cached pages 2629 * and the physical memory allocator's cache/free page queues. 2630 */ 2631 m->flags &= ~PG_ZERO; 2632 mtx_lock(&vm_page_queue_free_mtx); 2633 cache_was_empty = vm_radix_is_empty(&object->cache); 2634 if (vm_radix_insert(&object->cache, m)) { 2635 mtx_unlock(&vm_page_queue_free_mtx); 2636 if (object->type == OBJT_VNODE && 2637 object->resident_page_count == 0) 2638 vdrop(object->handle); 2639 m->object = NULL; 2640 vm_page_free(m); 2641 return; 2642 } 2643 2644 /* 2645 * The above call to vm_radix_insert() could reclaim the one pre- 2646 * existing cached page from this object, resulting in a call to 2647 * vdrop(). 2648 */ 2649 if (!cache_was_empty) 2650 cache_was_empty = vm_radix_is_singleton(&object->cache); 2651 2652 m->flags |= PG_CACHED; 2653 cnt.v_cache_count++; 2654 PCPU_INC(cnt.v_tcached); 2655#if VM_NRESERVLEVEL > 0 2656 if (!vm_reserv_free_page(m)) { 2657#else 2658 if (TRUE) { 2659#endif 2660 vm_phys_set_pool(VM_FREEPOOL_CACHE, m, 0); 2661 vm_phys_free_pages(m, 0); 2662 } 2663 vm_page_free_wakeup(); 2664 mtx_unlock(&vm_page_queue_free_mtx); 2665 2666 /* 2667 * Increment the vnode's hold count if this is the object's only 2668 * cached page. Decrement the vnode's hold count if this was 2669 * the object's only resident page. 2670 */ 2671 if (object->type == OBJT_VNODE) { 2672 if (cache_was_empty && object->resident_page_count != 0) 2673 vhold(object->handle); 2674 else if (!cache_was_empty && object->resident_page_count == 0) 2675 vdrop(object->handle); 2676 } 2677} 2678 2679/* 2680 * vm_page_advise 2681 * 2682 * Deactivate or do nothing, as appropriate. This routine is used 2683 * by madvise() and vop_stdadvise(). 2684 * 2685 * The object and page must be locked. 2686 */ 2687void 2688vm_page_advise(vm_page_t m, int advice) 2689{ 2690 2691 vm_page_assert_locked(m); 2692 VM_OBJECT_ASSERT_WLOCKED(m->object); 2693 if (advice == MADV_FREE) 2694 /* 2695 * Mark the page clean. This will allow the page to be freed 2696 * up by the system. However, such pages are often reused 2697 * quickly by malloc() so we do not do anything that would 2698 * cause a page fault if we can help it. 2699 * 2700 * Specifically, we do not try to actually free the page now 2701 * nor do we try to put it in the cache (which would cause a 2702 * page fault on reuse). 2703 * 2704 * But we do make the page as freeable as we can without 2705 * actually taking the step of unmapping it. 2706 */ 2707 vm_page_undirty(m); 2708 else if (advice != MADV_DONTNEED) 2709 return; 2710 2711 /* 2712 * Clear any references to the page. Otherwise, the page daemon will 2713 * immediately reactivate the page. 2714 */ 2715 vm_page_aflag_clear(m, PGA_REFERENCED); 2716 2717 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) 2718 vm_page_dirty(m); 2719 2720 /* 2721 * Place clean pages at the head of the inactive queue rather than the 2722 * tail, thus defeating the queue's LRU operation and ensuring that the 2723 * page will be reused quickly. 2724 */ 2725 _vm_page_deactivate(m, m->dirty == 0); 2726} 2727 2728/* 2729 * Grab a page, waiting until we are waken up due to the page 2730 * changing state. We keep on waiting, if the page continues 2731 * to be in the object. If the page doesn't exist, first allocate it 2732 * and then conditionally zero it. 2733 * 2734 * This routine may sleep. 2735 * 2736 * The object must be locked on entry. The lock will, however, be released 2737 * and reacquired if the routine sleeps. 2738 */ 2739vm_page_t 2740vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 2741{ 2742 vm_page_t m; 2743 int sleep; 2744 2745 VM_OBJECT_ASSERT_WLOCKED(object); 2746 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 2747 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 2748 ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 2749retrylookup: 2750 if ((m = vm_page_lookup(object, pindex)) != NULL) { 2751 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ? 2752 vm_page_xbusied(m) : vm_page_busied(m); 2753 if (sleep) { 2754 /* 2755 * Reference the page before unlocking and 2756 * sleeping so that the page daemon is less 2757 * likely to reclaim it. 2758 */ 2759 vm_page_aflag_set(m, PGA_REFERENCED); 2760 vm_page_lock(m); 2761 VM_OBJECT_WUNLOCK(object); 2762 vm_page_busy_sleep(m, "pgrbwt", (allocflags & 2763 VM_ALLOC_IGN_SBUSY) != 0); 2764 VM_OBJECT_WLOCK(object); 2765 goto retrylookup; 2766 } else { 2767 if ((allocflags & VM_ALLOC_WIRED) != 0) { 2768 vm_page_lock(m); 2769 vm_page_wire(m); 2770 vm_page_unlock(m); 2771 } 2772 if ((allocflags & 2773 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 2774 vm_page_xbusy(m); 2775 if ((allocflags & VM_ALLOC_SBUSY) != 0) 2776 vm_page_sbusy(m); 2777 return (m); 2778 } 2779 } 2780 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_IGN_SBUSY); 2781 if (m == NULL) { 2782 VM_OBJECT_WUNLOCK(object); 2783 VM_WAIT; 2784 VM_OBJECT_WLOCK(object); 2785 goto retrylookup; 2786 } else if (m->valid != 0) 2787 return (m); 2788 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 2789 pmap_zero_page(m); 2790 return (m); 2791} 2792 2793/* 2794 * Mapping function for valid or dirty bits in a page. 2795 * 2796 * Inputs are required to range within a page. 2797 */ 2798vm_page_bits_t 2799vm_page_bits(int base, int size) 2800{ 2801 int first_bit; 2802 int last_bit; 2803 2804 KASSERT( 2805 base + size <= PAGE_SIZE, 2806 ("vm_page_bits: illegal base/size %d/%d", base, size) 2807 ); 2808 2809 if (size == 0) /* handle degenerate case */ 2810 return (0); 2811 2812 first_bit = base >> DEV_BSHIFT; 2813 last_bit = (base + size - 1) >> DEV_BSHIFT; 2814 2815 return (((vm_page_bits_t)2 << last_bit) - 2816 ((vm_page_bits_t)1 << first_bit)); 2817} 2818 2819/* 2820 * vm_page_set_valid_range: 2821 * 2822 * Sets portions of a page valid. The arguments are expected 2823 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2824 * of any partial chunks touched by the range. The invalid portion of 2825 * such chunks will be zeroed. 2826 * 2827 * (base + size) must be less then or equal to PAGE_SIZE. 2828 */ 2829void 2830vm_page_set_valid_range(vm_page_t m, int base, int size) 2831{ 2832 int endoff, frag; 2833 2834 VM_OBJECT_ASSERT_WLOCKED(m->object); 2835 if (size == 0) /* handle degenerate case */ 2836 return; 2837 2838 /* 2839 * If the base is not DEV_BSIZE aligned and the valid 2840 * bit is clear, we have to zero out a portion of the 2841 * first block. 2842 */ 2843 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 2844 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 2845 pmap_zero_page_area(m, frag, base - frag); 2846 2847 /* 2848 * If the ending offset is not DEV_BSIZE aligned and the 2849 * valid bit is clear, we have to zero out a portion of 2850 * the last block. 2851 */ 2852 endoff = base + size; 2853 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 2854 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 2855 pmap_zero_page_area(m, endoff, 2856 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 2857 2858 /* 2859 * Assert that no previously invalid block that is now being validated 2860 * is already dirty. 2861 */ 2862 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 2863 ("vm_page_set_valid_range: page %p is dirty", m)); 2864 2865 /* 2866 * Set valid bits inclusive of any overlap. 2867 */ 2868 m->valid |= vm_page_bits(base, size); 2869} 2870 2871/* 2872 * Clear the given bits from the specified page's dirty field. 2873 */ 2874static __inline void 2875vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) 2876{ 2877 uintptr_t addr; 2878#if PAGE_SIZE < 16384 2879 int shift; 2880#endif 2881 2882 /* 2883 * If the object is locked and the page is neither exclusive busy nor 2884 * write mapped, then the page's dirty field cannot possibly be 2885 * set by a concurrent pmap operation. 2886 */ 2887 VM_OBJECT_ASSERT_WLOCKED(m->object); 2888 if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) 2889 m->dirty &= ~pagebits; 2890 else { 2891 /* 2892 * The pmap layer can call vm_page_dirty() without 2893 * holding a distinguished lock. The combination of 2894 * the object's lock and an atomic operation suffice 2895 * to guarantee consistency of the page dirty field. 2896 * 2897 * For PAGE_SIZE == 32768 case, compiler already 2898 * properly aligns the dirty field, so no forcible 2899 * alignment is needed. Only require existence of 2900 * atomic_clear_64 when page size is 32768. 2901 */ 2902 addr = (uintptr_t)&m->dirty; 2903#if PAGE_SIZE == 32768 2904 atomic_clear_64((uint64_t *)addr, pagebits); 2905#elif PAGE_SIZE == 16384 2906 atomic_clear_32((uint32_t *)addr, pagebits); 2907#else /* PAGE_SIZE <= 8192 */ 2908 /* 2909 * Use a trick to perform a 32-bit atomic on the 2910 * containing aligned word, to not depend on the existence 2911 * of atomic_clear_{8, 16}. 2912 */ 2913 shift = addr & (sizeof(uint32_t) - 1); 2914#if BYTE_ORDER == BIG_ENDIAN 2915 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY; 2916#else 2917 shift *= NBBY; 2918#endif 2919 addr &= ~(sizeof(uint32_t) - 1); 2920 atomic_clear_32((uint32_t *)addr, pagebits << shift); 2921#endif /* PAGE_SIZE */ 2922 } 2923} 2924 2925/* 2926 * vm_page_set_validclean: 2927 * 2928 * Sets portions of a page valid and clean. The arguments are expected 2929 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2930 * of any partial chunks touched by the range. The invalid portion of 2931 * such chunks will be zero'd. 2932 * 2933 * (base + size) must be less then or equal to PAGE_SIZE. 2934 */ 2935void 2936vm_page_set_validclean(vm_page_t m, int base, int size) 2937{ 2938 vm_page_bits_t oldvalid, pagebits; 2939 int endoff, frag; 2940 2941 VM_OBJECT_ASSERT_WLOCKED(m->object); 2942 if (size == 0) /* handle degenerate case */ 2943 return; 2944 2945 /* 2946 * If the base is not DEV_BSIZE aligned and the valid 2947 * bit is clear, we have to zero out a portion of the 2948 * first block. 2949 */ 2950 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 2951 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) 2952 pmap_zero_page_area(m, frag, base - frag); 2953 2954 /* 2955 * If the ending offset is not DEV_BSIZE aligned and the 2956 * valid bit is clear, we have to zero out a portion of 2957 * the last block. 2958 */ 2959 endoff = base + size; 2960 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 2961 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) 2962 pmap_zero_page_area(m, endoff, 2963 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 2964 2965 /* 2966 * Set valid, clear dirty bits. If validating the entire 2967 * page we can safely clear the pmap modify bit. We also 2968 * use this opportunity to clear the VPO_NOSYNC flag. If a process 2969 * takes a write fault on a MAP_NOSYNC memory area the flag will 2970 * be set again. 2971 * 2972 * We set valid bits inclusive of any overlap, but we can only 2973 * clear dirty bits for DEV_BSIZE chunks that are fully within 2974 * the range. 2975 */ 2976 oldvalid = m->valid; 2977 pagebits = vm_page_bits(base, size); 2978 m->valid |= pagebits; 2979#if 0 /* NOT YET */ 2980 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 2981 frag = DEV_BSIZE - frag; 2982 base += frag; 2983 size -= frag; 2984 if (size < 0) 2985 size = 0; 2986 } 2987 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 2988#endif 2989 if (base == 0 && size == PAGE_SIZE) { 2990 /* 2991 * The page can only be modified within the pmap if it is 2992 * mapped, and it can only be mapped if it was previously 2993 * fully valid. 2994 */ 2995 if (oldvalid == VM_PAGE_BITS_ALL) 2996 /* 2997 * Perform the pmap_clear_modify() first. Otherwise, 2998 * a concurrent pmap operation, such as 2999 * pmap_protect(), could clear a modification in the 3000 * pmap and set the dirty field on the page before 3001 * pmap_clear_modify() had begun and after the dirty 3002 * field was cleared here. 3003 */ 3004 pmap_clear_modify(m); 3005 m->dirty = 0; 3006 m->oflags &= ~VPO_NOSYNC; 3007 } else if (oldvalid != VM_PAGE_BITS_ALL) 3008 m->dirty &= ~pagebits; 3009 else 3010 vm_page_clear_dirty_mask(m, pagebits); 3011} 3012 3013void 3014vm_page_clear_dirty(vm_page_t m, int base, int size) 3015{ 3016 3017 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 3018} 3019 3020/* 3021 * vm_page_set_invalid: 3022 * 3023 * Invalidates DEV_BSIZE'd chunks within a page. Both the 3024 * valid and dirty bits for the effected areas are cleared. 3025 */ 3026void 3027vm_page_set_invalid(vm_page_t m, int base, int size) 3028{ 3029 vm_page_bits_t bits; 3030 vm_object_t object; 3031 3032 object = m->object; 3033 VM_OBJECT_ASSERT_WLOCKED(object); 3034 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + 3035 size >= object->un_pager.vnp.vnp_size) 3036 bits = VM_PAGE_BITS_ALL; 3037 else 3038 bits = vm_page_bits(base, size); 3039 if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL && 3040 bits != 0) 3041 pmap_remove_all(m); 3042 KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) || 3043 !pmap_page_is_mapped(m), 3044 ("vm_page_set_invalid: page %p is mapped", m)); 3045 m->valid &= ~bits; 3046 m->dirty &= ~bits; 3047} 3048 3049/* 3050 * vm_page_zero_invalid() 3051 * 3052 * The kernel assumes that the invalid portions of a page contain 3053 * garbage, but such pages can be mapped into memory by user code. 3054 * When this occurs, we must zero out the non-valid portions of the 3055 * page so user code sees what it expects. 3056 * 3057 * Pages are most often semi-valid when the end of a file is mapped 3058 * into memory and the file's size is not page aligned. 3059 */ 3060void 3061vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 3062{ 3063 int b; 3064 int i; 3065 3066 VM_OBJECT_ASSERT_WLOCKED(m->object); 3067 /* 3068 * Scan the valid bits looking for invalid sections that 3069 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the 3070 * valid bit may be set ) have already been zeroed by 3071 * vm_page_set_validclean(). 3072 */ 3073 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 3074 if (i == (PAGE_SIZE / DEV_BSIZE) || 3075 (m->valid & ((vm_page_bits_t)1 << i))) { 3076 if (i > b) { 3077 pmap_zero_page_area(m, 3078 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 3079 } 3080 b = i + 1; 3081 } 3082 } 3083 3084 /* 3085 * setvalid is TRUE when we can safely set the zero'd areas 3086 * as being valid. We can do this if there are no cache consistancy 3087 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 3088 */ 3089 if (setvalid) 3090 m->valid = VM_PAGE_BITS_ALL; 3091} 3092 3093/* 3094 * vm_page_is_valid: 3095 * 3096 * Is (partial) page valid? Note that the case where size == 0 3097 * will return FALSE in the degenerate case where the page is 3098 * entirely invalid, and TRUE otherwise. 3099 */ 3100int 3101vm_page_is_valid(vm_page_t m, int base, int size) 3102{ 3103 vm_page_bits_t bits; 3104 3105 VM_OBJECT_ASSERT_LOCKED(m->object); 3106 bits = vm_page_bits(base, size); 3107 return (m->valid != 0 && (m->valid & bits) == bits); 3108} 3109 3110/* 3111 * vm_page_ps_is_valid: 3112 * 3113 * Returns TRUE if the entire (super)page is valid and FALSE otherwise. 3114 */ 3115boolean_t 3116vm_page_ps_is_valid(vm_page_t m) 3117{ 3118 int i, npages; 3119 3120 VM_OBJECT_ASSERT_LOCKED(m->object); 3121 npages = atop(pagesizes[m->psind]); 3122 3123 /* 3124 * The physically contiguous pages that make up a superpage, i.e., a 3125 * page with a page size index ("psind") greater than zero, will 3126 * occupy adjacent entries in vm_page_array[]. 3127 */ 3128 for (i = 0; i < npages; i++) { 3129 if (m[i].valid != VM_PAGE_BITS_ALL) 3130 return (FALSE); 3131 } 3132 return (TRUE); 3133} 3134 3135/* 3136 * Set the page's dirty bits if the page is modified. 3137 */ 3138void 3139vm_page_test_dirty(vm_page_t m) 3140{ 3141 3142 VM_OBJECT_ASSERT_WLOCKED(m->object); 3143 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 3144 vm_page_dirty(m); 3145} 3146 3147void 3148vm_page_lock_KBI(vm_page_t m, const char *file, int line) 3149{ 3150 3151 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); 3152} 3153 3154void 3155vm_page_unlock_KBI(vm_page_t m, const char *file, int line) 3156{ 3157 3158 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); 3159} 3160 3161int 3162vm_page_trylock_KBI(vm_page_t m, const char *file, int line) 3163{ 3164 3165 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); 3166} 3167 3168#if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) 3169void 3170vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line) 3171{ 3172 3173 vm_page_lock_assert_KBI(m, MA_OWNED, file, line); 3174} 3175 3176void 3177vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) 3178{ 3179 3180 mtx_assert_(vm_page_lockptr(m), a, file, line); 3181} 3182#endif 3183 3184#ifdef INVARIANTS 3185void 3186vm_page_object_lock_assert(vm_page_t m) 3187{ 3188 3189 /* 3190 * Certain of the page's fields may only be modified by the 3191 * holder of the containing object's lock or the exclusive busy. 3192 * holder. Unfortunately, the holder of the write busy is 3193 * not recorded, and thus cannot be checked here. 3194 */ 3195 if (m->object != NULL && !vm_page_xbusied(m)) 3196 VM_OBJECT_ASSERT_WLOCKED(m->object); 3197} 3198 3199void 3200vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits) 3201{ 3202 3203 if ((bits & PGA_WRITEABLE) == 0) 3204 return; 3205 3206 /* 3207 * The PGA_WRITEABLE flag can only be set if the page is 3208 * managed, is exclusively busied or the object is locked. 3209 * Currently, this flag is only set by pmap_enter(). 3210 */ 3211 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 3212 ("PGA_WRITEABLE on unmanaged page")); 3213 if (!vm_page_xbusied(m)) 3214 VM_OBJECT_ASSERT_LOCKED(m->object); 3215} 3216#endif 3217 3218#include "opt_ddb.h" 3219#ifdef DDB 3220#include <sys/kernel.h> 3221 3222#include <ddb/ddb.h> 3223 3224DB_SHOW_COMMAND(page, vm_page_print_page_info) 3225{ 3226 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 3227 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 3228 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 3229 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 3230 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 3231 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 3232 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 3233 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 3234 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 3235 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 3236} 3237 3238DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 3239{ 3240 int dom; 3241 3242 db_printf("pq_free %d pq_cache %d\n", 3243 cnt.v_free_count, cnt.v_cache_count); 3244 for (dom = 0; dom < vm_ndomains; dom++) { 3245 db_printf( 3246 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pass %d\n", 3247 dom, 3248 vm_dom[dom].vmd_page_count, 3249 vm_dom[dom].vmd_free_count, 3250 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, 3251 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, 3252 vm_dom[dom].vmd_pass); 3253 } 3254} 3255 3256DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) 3257{ 3258 vm_page_t m; 3259 boolean_t phys; 3260 3261 if (!have_addr) { 3262 db_printf("show pginfo addr\n"); 3263 return; 3264 } 3265 3266 phys = strchr(modif, 'p') != NULL; 3267 if (phys) 3268 m = PHYS_TO_VM_PAGE(addr); 3269 else 3270 m = (vm_page_t)addr; 3271 db_printf( 3272 "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n" 3273 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", 3274 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, 3275 m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags, 3276 m->flags, m->act_count, m->busy_lock, m->valid, m->dirty); 3277} 3278#endif /* DDB */ 3279