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("$FreeBSD$"); 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 */ 125 126 struct vm_domain vm_dom[MAXMEMDOM]; 127 struct mtx_padalign vm_page_queue_free_mtx; 128 129 struct mtx_padalign pa_lock[PA_LOCK_COUNT]; 130 131 vm_page_t vm_page_array; 132 long vm_page_array_size; 133 long first_page; 134 int vm_page_zero_count; 135 136 static int boot_pages = UMA_BOOT_PAGES; 137 TUNABLE_INT("vm.boot_pages", &boot_pages); 138 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RD, &boot_pages, 0, 139 "number of pages allocated for bootstrapping the VM system"); 140 141 static int pa_tryrelock_restart; 142 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD, 143 &pa_tryrelock_restart, 0, "Number of tryrelock restarts"); 144 145 static uma_zone_t fakepg_zone; 146 147 static struct vnode *vm_page_alloc_init(vm_page_t m); 148 static void vm_page_cache_turn_free(vm_page_t m); 149 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits); 150 static void vm_page_enqueue(int queue, vm_page_t m); 151 static void vm_page_init_fakepg(void *dummy); 152 static int vm_page_insert_after(vm_page_t m, vm_object_t object, 153 vm_pindex_t pindex, vm_page_t mpred); 154 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object, 155 vm_page_t mpred); 156 157 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init_fakepg, NULL); 158 159 static void 160 vm_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 170 CTASSERT(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 */ 180 int 181 vm_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 */ 209 void 210 vm_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 */ 226 static int 227 vm_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 } 247 248 static void 249 vm_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(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(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. 279 * 280 * Allocates memory for the page cells, and 281 * for the object/offset-to-page hash table headers. 282 * Each page cell is initialized and placed on the free list. 283 */ 284 vm_offset_t 285 vm_page_startup(vm_offset_t vaddr) 286 { 287 vm_offset_t mapped; 288 vm_paddr_t page_range; 289 vm_paddr_t new_end; 290 int i; 291 vm_paddr_t pa; 292 vm_paddr_t last_pa; 293 char *list; 294 295 /* the biggest memory array is the second group of pages */ 296 vm_paddr_t end; 297 vm_paddr_t biggestsize; 298 vm_paddr_t low_water, high_water; 299 int biggestone; 300 301 biggestsize = 0; 302 biggestone = 0; 303 vaddr = round_page(vaddr); 304 305 for (i = 0; phys_avail[i + 1]; i += 2) { 306 phys_avail[i] = round_page(phys_avail[i]); 307 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 308 } 309 310 low_water = phys_avail[0]; 311 high_water = phys_avail[1]; 312 313 for (i = 0; phys_avail[i + 1]; i += 2) { 314 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i]; 315 316 if (size > biggestsize) { 317 biggestone = i; 318 biggestsize = size; 319 } 320 if (phys_avail[i] < low_water) 321 low_water = phys_avail[i]; 322 if (phys_avail[i + 1] > high_water) 323 high_water = phys_avail[i + 1]; 324 } 325 326 #ifdef XEN 327 low_water = 0; 328 #endif 329 330 end = phys_avail[biggestone+1]; 331 332 /* 333 * Initialize the page and queue locks. 334 */ 335 mtx_init(&vm_page_queue_free_mtx, "vm page free queue", NULL, MTX_DEF); 336 for (i = 0; i < PA_LOCK_COUNT; i++) 337 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF); 338 for (i = 0; i < vm_ndomains; i++) 339 vm_page_domain_init(&vm_dom[i]); 340 341 /* 342 * Allocate memory for use when boot strapping the kernel memory 343 * allocator. 344 */ 345 new_end = end - (boot_pages * UMA_SLAB_SIZE); 346 new_end = trunc_page(new_end); 347 mapped = pmap_map(&vaddr, new_end, end, 348 VM_PROT_READ | VM_PROT_WRITE); 349 bzero((void *)mapped, end - new_end); 350 uma_startup((void *)mapped, boot_pages); 351 352 #if defined(__amd64__) || defined(__i386__) || defined(__arm__) || \ 353 defined(__mips__) 354 /* 355 * Allocate a bitmap to indicate that a random physical page 356 * needs to be included in a minidump. 357 * 358 * The amd64 port needs this to indicate which direct map pages 359 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 360 * 361 * However, i386 still needs this workspace internally within the 362 * minidump code. In theory, they are not needed on i386, but are 363 * included should the sf_buf code decide to use them. 364 */ 365 last_pa = 0; 366 for (i = 0; dump_avail[i + 1] != 0; i += 2) 367 if (dump_avail[i + 1] > last_pa) 368 last_pa = dump_avail[i + 1]; 369 page_range = last_pa / PAGE_SIZE; 370 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 371 new_end -= vm_page_dump_size; 372 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, 373 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); 374 bzero((void *)vm_page_dump, vm_page_dump_size); 375 #endif 376 #ifdef __amd64__ 377 /* 378 * Request that the physical pages underlying the message buffer be 379 * included in a crash dump. Since the message buffer is accessed 380 * through the direct map, they are not automatically included. 381 */ 382 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); 383 last_pa = pa + round_page(msgbufsize); 384 while (pa < last_pa) { 385 dump_add_page(pa); 386 pa += PAGE_SIZE; 387 } 388 #endif 389 /* 390 * Compute the number of pages of memory that will be available for 391 * use (taking into account the overhead of a page structure per 392 * page). 393 */ 394 first_page = low_water / PAGE_SIZE; 395 #ifdef VM_PHYSSEG_SPARSE 396 page_range = 0; 397 for (i = 0; phys_avail[i + 1] != 0; i += 2) 398 page_range += atop(phys_avail[i + 1] - phys_avail[i]); 399 #elif defined(VM_PHYSSEG_DENSE) 400 page_range = high_water / PAGE_SIZE - first_page; 401 #else 402 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 403 #endif 404 end = new_end; 405 406 /* 407 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 408 */ 409 vaddr += PAGE_SIZE; 410 411 /* 412 * Initialize the mem entry structures now, and put them in the free 413 * queue. 414 */ 415 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 416 mapped = pmap_map(&vaddr, new_end, end, 417 VM_PROT_READ | VM_PROT_WRITE); 418 vm_page_array = (vm_page_t) mapped; 419 #if VM_NRESERVLEVEL > 0 420 /* 421 * Allocate memory for the reservation management system's data 422 * structures. 423 */ 424 new_end = vm_reserv_startup(&vaddr, new_end, high_water); 425 #endif 426 #if defined(__amd64__) || defined(__mips__) 427 /* 428 * pmap_map on amd64 and mips can come out of the direct-map, not kvm 429 * like i386, so the pages must be tracked for a crashdump to include 430 * this data. This includes the vm_page_array and the early UMA 431 * bootstrap pages. 432 */ 433 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE) 434 dump_add_page(pa); 435 #endif 436 phys_avail[biggestone + 1] = new_end; 437 438 /* 439 * Clear all of the page structures 440 */ 441 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 442 for (i = 0; i < page_range; i++) 443 vm_page_array[i].order = VM_NFREEORDER; 444 vm_page_array_size = page_range; 445 446 /* 447 * Initialize the physical memory allocator. 448 */ 449 vm_phys_init(); 450 451 /* 452 * Add every available physical page that is not blacklisted to 453 * the free lists. 454 */ 455 cnt.v_page_count = 0; 456 cnt.v_free_count = 0; 457 list = getenv("vm.blacklist"); 458 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 459 pa = phys_avail[i]; 460 last_pa = phys_avail[i + 1]; 461 while (pa < last_pa) { 462 if (list != NULL && 463 vm_page_blacklist_lookup(list, pa)) 464 printf("Skipping page with pa 0x%jx\n", 465 (uintmax_t)pa); 466 else 467 vm_phys_add_page(pa); 468 pa += PAGE_SIZE; 469 } 470 } 471 freeenv(list); 472 #if VM_NRESERVLEVEL > 0 473 /* 474 * Initialize the reservation management system. 475 */ 476 vm_reserv_init(); 477 #endif 478 return (vaddr); 479 } 480 481 void 482 vm_page_reference(vm_page_t m) 483 { 484 485 vm_page_aflag_set(m, PGA_REFERENCED); 486 } 487 488 /* 489 * vm_page_busy_downgrade: 490 * 491 * Downgrade an exclusive busy page into a single shared busy page. 492 */ 493 void 494 vm_page_busy_downgrade(vm_page_t m) 495 { 496 u_int x; 497 498 vm_page_assert_xbusied(m); 499 500 for (;;) { 501 x = m->busy_lock; 502 x &= VPB_BIT_WAITERS; 503 if (atomic_cmpset_rel_int(&m->busy_lock, 504 VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1) | x)) 505 break; 506 } 507 } 508 509 /* 510 * vm_page_sbusied: 511 * 512 * Return a positive value if the page is shared busied, 0 otherwise. 513 */ 514 int 515 vm_page_sbusied(vm_page_t m) 516 { 517 u_int x; 518 519 x = m->busy_lock; 520 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED); 521 } 522 523 /* 524 * vm_page_sunbusy: 525 * 526 * Shared unbusy a page. 527 */ 528 void 529 vm_page_sunbusy(vm_page_t m) 530 { 531 u_int x; 532 533 vm_page_assert_sbusied(m); 534 535 for (;;) { 536 x = m->busy_lock; 537 if (VPB_SHARERS(x) > 1) { 538 if (atomic_cmpset_int(&m->busy_lock, x, 539 x - VPB_ONE_SHARER)) 540 break; 541 continue; 542 } 543 if ((x & VPB_BIT_WAITERS) == 0) { 544 KASSERT(x == VPB_SHARERS_WORD(1), 545 ("vm_page_sunbusy: invalid lock state")); 546 if (atomic_cmpset_int(&m->busy_lock, 547 VPB_SHARERS_WORD(1), VPB_UNBUSIED)) 548 break; 549 continue; 550 } 551 KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS), 552 ("vm_page_sunbusy: invalid lock state for waiters")); 553 554 vm_page_lock(m); 555 if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) { 556 vm_page_unlock(m); 557 continue; 558 } 559 wakeup(m); 560 vm_page_unlock(m); 561 break; 562 } 563 } 564 565 /* 566 * vm_page_busy_sleep: 567 * 568 * Sleep and release the page lock, using the page pointer as wchan. 569 * This is used to implement the hard-path of busying mechanism. 570 * 571 * The given page must be locked. 572 */ 573 void 574 vm_page_busy_sleep(vm_page_t m, const char *wmesg) 575 { 576 u_int x; 577 578 vm_page_lock_assert(m, MA_OWNED); 579 580 x = m->busy_lock; 581 if (x == VPB_UNBUSIED) { 582 vm_page_unlock(m); 583 return; 584 } 585 if ((x & VPB_BIT_WAITERS) == 0 && 586 !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS)) { 587 vm_page_unlock(m); 588 return; 589 } 590 msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0); 591 } 592 593 /* 594 * vm_page_trysbusy: 595 * 596 * Try to shared busy a page. 597 * If the operation succeeds 1 is returned otherwise 0. 598 * The operation never sleeps. 599 */ 600 int 601 vm_page_trysbusy(vm_page_t m) 602 { 603 u_int x; 604 605 for (;;) { 606 x = m->busy_lock; 607 if ((x & VPB_BIT_SHARED) == 0) 608 return (0); 609 if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER)) 610 return (1); 611 } 612 } 613 614 /* 615 * vm_page_xunbusy_hard: 616 * 617 * Called after the first try the exclusive unbusy of a page failed. 618 * It is assumed that the waiters bit is on. 619 */ 620 void 621 vm_page_xunbusy_hard(vm_page_t m) 622 { 623 624 vm_page_assert_xbusied(m); 625 626 vm_page_lock(m); 627 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); 628 wakeup(m); 629 vm_page_unlock(m); 630 } 631 632 /* 633 * vm_page_flash: 634 * 635 * Wakeup anyone waiting for the page. 636 * The ownership bits do not change. 637 * 638 * The given page must be locked. 639 */ 640 void 641 vm_page_flash(vm_page_t m) 642 { 643 u_int x; 644 645 vm_page_lock_assert(m, MA_OWNED); 646 647 for (;;) { 648 x = m->busy_lock; 649 if ((x & VPB_BIT_WAITERS) == 0) 650 return; 651 if (atomic_cmpset_int(&m->busy_lock, x, 652 x & (~VPB_BIT_WAITERS))) 653 break; 654 } 655 wakeup(m); 656 } 657 658 /* 659 * Keep page from being freed by the page daemon 660 * much of the same effect as wiring, except much lower 661 * overhead and should be used only for *very* temporary 662 * holding ("wiring"). 663 */ 664 void 665 vm_page_hold(vm_page_t mem) 666 { 667 668 vm_page_lock_assert(mem, MA_OWNED); 669 mem->hold_count++; 670 } 671 672 void 673 vm_page_unhold(vm_page_t mem) 674 { 675 676 vm_page_lock_assert(mem, MA_OWNED); 677 KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!")); 678 --mem->hold_count; 679 if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0) 680 vm_page_free_toq(mem); 681 } 682 683 /* 684 * vm_page_unhold_pages: 685 * 686 * Unhold each of the pages that is referenced by the given array. 687 */ 688 void 689 vm_page_unhold_pages(vm_page_t *ma, int count) 690 { 691 struct mtx *mtx, *new_mtx; 692 693 mtx = NULL; 694 for (; count != 0; count--) { 695 /* 696 * Avoid releasing and reacquiring the same page lock. 697 */ 698 new_mtx = vm_page_lockptr(*ma); 699 if (mtx != new_mtx) { 700 if (mtx != NULL) 701 mtx_unlock(mtx); 702 mtx = new_mtx; 703 mtx_lock(mtx); 704 } 705 vm_page_unhold(*ma); 706 ma++; 707 } 708 if (mtx != NULL) 709 mtx_unlock(mtx); 710 } 711 712 vm_page_t 713 PHYS_TO_VM_PAGE(vm_paddr_t pa) 714 { 715 vm_page_t m; 716 717 #ifdef VM_PHYSSEG_SPARSE 718 m = vm_phys_paddr_to_vm_page(pa); 719 if (m == NULL) 720 m = vm_phys_fictitious_to_vm_page(pa); 721 return (m); 722 #elif defined(VM_PHYSSEG_DENSE) 723 long pi; 724 725 pi = atop(pa); 726 if (pi >= first_page && (pi - first_page) < vm_page_array_size) { 727 m = &vm_page_array[pi - first_page]; 728 return (m); 729 } 730 return (vm_phys_fictitious_to_vm_page(pa)); 731 #else 732 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 733 #endif 734 } 735 736 /* 737 * vm_page_getfake: 738 * 739 * Create a fictitious page with the specified physical address and 740 * memory attribute. The memory attribute is the only the machine- 741 * dependent aspect of a fictitious page that must be initialized. 742 */ 743 vm_page_t 744 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr) 745 { 746 vm_page_t m; 747 748 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO); 749 vm_page_initfake(m, paddr, memattr); 750 return (m); 751 } 752 753 void 754 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 755 { 756 757 if ((m->flags & PG_FICTITIOUS) != 0) { 758 /* 759 * The page's memattr might have changed since the 760 * previous initialization. Update the pmap to the 761 * new memattr. 762 */ 763 goto memattr; 764 } 765 m->phys_addr = paddr; 766 m->queue = PQ_NONE; 767 /* Fictitious pages don't use "segind". */ 768 m->flags = PG_FICTITIOUS; 769 /* Fictitious pages don't use "order" or "pool". */ 770 m->oflags = VPO_UNMANAGED; 771 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 772 m->wire_count = 1; 773 pmap_page_init(m); 774 memattr: 775 pmap_page_set_memattr(m, memattr); 776 } 777 778 /* 779 * vm_page_putfake: 780 * 781 * Release a fictitious page. 782 */ 783 void 784 vm_page_putfake(vm_page_t m) 785 { 786 787 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m)); 788 KASSERT((m->flags & PG_FICTITIOUS) != 0, 789 ("vm_page_putfake: bad page %p", m)); 790 uma_zfree(fakepg_zone, m); 791 } 792 793 /* 794 * vm_page_updatefake: 795 * 796 * Update the given fictitious page to the specified physical address and 797 * memory attribute. 798 */ 799 void 800 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 801 { 802 803 KASSERT((m->flags & PG_FICTITIOUS) != 0, 804 ("vm_page_updatefake: bad page %p", m)); 805 m->phys_addr = paddr; 806 pmap_page_set_memattr(m, memattr); 807 } 808 809 /* 810 * vm_page_free: 811 * 812 * Free a page. 813 */ 814 void 815 vm_page_free(vm_page_t m) 816 { 817 818 m->flags &= ~PG_ZERO; 819 vm_page_free_toq(m); 820 } 821 822 /* 823 * vm_page_free_zero: 824 * 825 * Free a page to the zerod-pages queue 826 */ 827 void 828 vm_page_free_zero(vm_page_t m) 829 { 830 831 m->flags |= PG_ZERO; 832 vm_page_free_toq(m); 833 } 834 835 /* 836 * Unbusy and handle the page queueing for a page from the VOP_GETPAGES() 837 * array which is not the request page. 838 */ 839 void 840 vm_page_readahead_finish(vm_page_t m) 841 { 842 843 if (m->valid != 0) { 844 /* 845 * Since the page is not the requested page, whether 846 * it should be activated or deactivated is not 847 * obvious. Empirical results have shown that 848 * deactivating the page is usually the best choice, 849 * unless the page is wanted by another thread. 850 */ 851 vm_page_lock(m); 852 if ((m->busy_lock & VPB_BIT_WAITERS) != 0) 853 vm_page_activate(m); 854 else 855 vm_page_deactivate(m); 856 vm_page_unlock(m); 857 vm_page_xunbusy(m); 858 } else { 859 /* 860 * Free the completely invalid page. Such page state 861 * occurs due to the short read operation which did 862 * not covered our page at all, or in case when a read 863 * error happens. 864 */ 865 vm_page_lock(m); 866 vm_page_free(m); 867 vm_page_unlock(m); 868 } 869 } 870 871 /* 872 * vm_page_sleep_if_busy: 873 * 874 * Sleep and release the page queues lock if the page is busied. 875 * Returns TRUE if the thread slept. 876 * 877 * The given page must be unlocked and object containing it must 878 * be locked. 879 */ 880 int 881 vm_page_sleep_if_busy(vm_page_t m, const char *msg) 882 { 883 vm_object_t obj; 884 885 vm_page_lock_assert(m, MA_NOTOWNED); 886 VM_OBJECT_ASSERT_WLOCKED(m->object); 887 888 if (vm_page_busied(m)) { 889 /* 890 * The page-specific object must be cached because page 891 * identity can change during the sleep, causing the 892 * re-lock of a different object. 893 * It is assumed that a reference to the object is already 894 * held by the callers. 895 */ 896 obj = m->object; 897 vm_page_lock(m); 898 VM_OBJECT_WUNLOCK(obj); 899 vm_page_busy_sleep(m, msg); 900 VM_OBJECT_WLOCK(obj); 901 return (TRUE); 902 } 903 return (FALSE); 904 } 905 906 /* 907 * vm_page_dirty_KBI: [ internal use only ] 908 * 909 * Set all bits in the page's dirty field. 910 * 911 * The object containing the specified page must be locked if the 912 * call is made from the machine-independent layer. 913 * 914 * See vm_page_clear_dirty_mask(). 915 * 916 * This function should only be called by vm_page_dirty(). 917 */ 918 void 919 vm_page_dirty_KBI(vm_page_t m) 920 { 921 922 /* These assertions refer to this operation by its public name. */ 923 KASSERT((m->flags & PG_CACHED) == 0, 924 ("vm_page_dirty: page in cache!")); 925 KASSERT(!VM_PAGE_IS_FREE(m), 926 ("vm_page_dirty: page is free!")); 927 KASSERT(m->valid == VM_PAGE_BITS_ALL, 928 ("vm_page_dirty: page is invalid!")); 929 m->dirty = VM_PAGE_BITS_ALL; 930 } 931 932 /* 933 * vm_page_insert: [ internal use only ] 934 * 935 * Inserts the given mem entry into the object and object list. 936 * 937 * The object must be locked. 938 */ 939 int 940 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 941 { 942 vm_page_t mpred; 943 944 VM_OBJECT_ASSERT_WLOCKED(object); 945 mpred = vm_radix_lookup_le(&object->rtree, pindex); 946 return (vm_page_insert_after(m, object, pindex, mpred)); 947 } 948 949 /* 950 * vm_page_insert_after: 951 * 952 * Inserts the page "m" into the specified object at offset "pindex". 953 * 954 * The page "mpred" must immediately precede the offset "pindex" within 955 * the specified object. 956 * 957 * The object must be locked. 958 */ 959 static int 960 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, 961 vm_page_t mpred) 962 { 963 vm_pindex_t sidx; 964 vm_object_t sobj; 965 vm_page_t msucc; 966 967 VM_OBJECT_ASSERT_WLOCKED(object); 968 KASSERT(m->object == NULL, 969 ("vm_page_insert_after: page already inserted")); 970 if (mpred != NULL) { 971 KASSERT(mpred->object == object, 972 ("vm_page_insert_after: object doesn't contain mpred")); 973 KASSERT(mpred->pindex < pindex, 974 ("vm_page_insert_after: mpred doesn't precede pindex")); 975 msucc = TAILQ_NEXT(mpred, listq); 976 } else 977 msucc = TAILQ_FIRST(&object->memq); 978 if (msucc != NULL) 979 KASSERT(msucc->pindex > pindex, 980 ("vm_page_insert_after: msucc doesn't succeed pindex")); 981 982 /* 983 * Record the object/offset pair in this page 984 */ 985 sobj = m->object; 986 sidx = m->pindex; 987 m->object = object; 988 m->pindex = pindex; 989 990 /* 991 * Now link into the object's ordered list of backed pages. 992 */ 993 if (vm_radix_insert(&object->rtree, m)) { 994 m->object = sobj; 995 m->pindex = sidx; 996 return (1); 997 } 998 vm_page_insert_radixdone(m, object, mpred); 999 return (0); 1000 } 1001 1002 /* 1003 * vm_page_insert_radixdone: 1004 * 1005 * Complete page "m" insertion into the specified object after the 1006 * radix trie hooking. 1007 * 1008 * The page "mpred" must precede the offset "m->pindex" within the 1009 * specified object. 1010 * 1011 * The object must be locked. 1012 */ 1013 static void 1014 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred) 1015 { 1016 1017 VM_OBJECT_ASSERT_WLOCKED(object); 1018 KASSERT(object != NULL && m->object == object, 1019 ("vm_page_insert_radixdone: page %p has inconsistent object", m)); 1020 if (mpred != NULL) { 1021 KASSERT(mpred->object == object, 1022 ("vm_page_insert_after: object doesn't contain mpred")); 1023 KASSERT(mpred->pindex < m->pindex, 1024 ("vm_page_insert_after: mpred doesn't precede pindex")); 1025 } 1026 1027 if (mpred != NULL) 1028 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq); 1029 else 1030 TAILQ_INSERT_HEAD(&object->memq, m, listq); 1031 1032 /* 1033 * Show that the object has one more resident page. 1034 */ 1035 object->resident_page_count++; 1036 1037 /* 1038 * Hold the vnode until the last page is released. 1039 */ 1040 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 1041 vhold(object->handle); 1042 1043 /* 1044 * Since we are inserting a new and possibly dirty page, 1045 * update the object's OBJ_MIGHTBEDIRTY flag. 1046 */ 1047 if (pmap_page_is_write_mapped(m)) 1048 vm_object_set_writeable_dirty(object); 1049 } 1050 1051 /* 1052 * vm_page_remove: 1053 * 1054 * Removes the given mem entry from the object/offset-page 1055 * table and the object page list, but do not invalidate/terminate 1056 * the backing store. 1057 * 1058 * The object must be locked. The page must be locked if it is managed. 1059 */ 1060 void 1061 vm_page_remove(vm_page_t m) 1062 { 1063 vm_object_t object; 1064 boolean_t lockacq; 1065 1066 if ((m->oflags & VPO_UNMANAGED) == 0) 1067 vm_page_lock_assert(m, MA_OWNED); 1068 if ((object = m->object) == NULL) 1069 return; 1070 VM_OBJECT_ASSERT_WLOCKED(object); 1071 if (vm_page_xbusied(m)) { 1072 lockacq = FALSE; 1073 if ((m->oflags & VPO_UNMANAGED) != 0 && 1074 !mtx_owned(vm_page_lockptr(m))) { 1075 lockacq = TRUE; 1076 vm_page_lock(m); 1077 } 1078 vm_page_flash(m); 1079 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); 1080 if (lockacq) 1081 vm_page_unlock(m); 1082 } 1083 1084 /* 1085 * Now remove from the object's list of backed pages. 1086 */ 1087 vm_radix_remove(&object->rtree, m->pindex); 1088 TAILQ_REMOVE(&object->memq, m, listq); 1089 1090 /* 1091 * And show that the object has one fewer resident page. 1092 */ 1093 object->resident_page_count--; 1094 1095 /* 1096 * The vnode may now be recycled. 1097 */ 1098 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 1099 vdrop(object->handle); 1100 1101 m->object = NULL; 1102 } 1103 1104 /* 1105 * vm_page_lookup: 1106 * 1107 * Returns the page associated with the object/offset 1108 * pair specified; if none is found, NULL is returned. 1109 * 1110 * The object must be locked. 1111 */ 1112 vm_page_t 1113 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1114 { 1115 1116 VM_OBJECT_ASSERT_LOCKED(object); 1117 return (vm_radix_lookup(&object->rtree, pindex)); 1118 } 1119 1120 /* 1121 * vm_page_find_least: 1122 * 1123 * Returns the page associated with the object with least pindex 1124 * greater than or equal to the parameter pindex, or NULL. 1125 * 1126 * The object must be locked. 1127 */ 1128 vm_page_t 1129 vm_page_find_least(vm_object_t object, vm_pindex_t pindex) 1130 { 1131 vm_page_t m; 1132 1133 VM_OBJECT_ASSERT_LOCKED(object); 1134 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex) 1135 m = vm_radix_lookup_ge(&object->rtree, pindex); 1136 return (m); 1137 } 1138 1139 /* 1140 * Returns the given page's successor (by pindex) within the object if it is 1141 * resident; if none is found, NULL is returned. 1142 * 1143 * The object must be locked. 1144 */ 1145 vm_page_t 1146 vm_page_next(vm_page_t m) 1147 { 1148 vm_page_t next; 1149 1150 VM_OBJECT_ASSERT_WLOCKED(m->object); 1151 if ((next = TAILQ_NEXT(m, listq)) != NULL && 1152 next->pindex != m->pindex + 1) 1153 next = NULL; 1154 return (next); 1155 } 1156 1157 /* 1158 * Returns the given page's predecessor (by pindex) within the object if it is 1159 * resident; if none is found, NULL is returned. 1160 * 1161 * The object must be locked. 1162 */ 1163 vm_page_t 1164 vm_page_prev(vm_page_t m) 1165 { 1166 vm_page_t prev; 1167 1168 VM_OBJECT_ASSERT_WLOCKED(m->object); 1169 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL && 1170 prev->pindex != m->pindex - 1) 1171 prev = NULL; 1172 return (prev); 1173 } 1174 1175 /* 1176 * Uses the page mnew as a replacement for an existing page at index 1177 * pindex which must be already present in the object. 1178 * 1179 * The existing page must not be on a paging queue. 1180 */ 1181 vm_page_t 1182 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex) 1183 { 1184 vm_page_t mold, mpred; 1185 1186 VM_OBJECT_ASSERT_WLOCKED(object); 1187 1188 /* 1189 * This function mostly follows vm_page_insert() and 1190 * vm_page_remove() without the radix, object count and vnode 1191 * dance. Double check such functions for more comments. 1192 */ 1193 mpred = vm_radix_lookup(&object->rtree, pindex); 1194 KASSERT(mpred != NULL, 1195 ("vm_page_replace: replacing page not present with pindex")); 1196 mpred = TAILQ_PREV(mpred, respgs, listq); 1197 if (mpred != NULL) 1198 KASSERT(mpred->pindex < pindex, 1199 ("vm_page_insert_after: mpred doesn't precede pindex")); 1200 1201 mnew->object = object; 1202 mnew->pindex = pindex; 1203 mold = vm_radix_replace(&object->rtree, mnew, pindex); 1204 KASSERT(mold->queue == PQ_NONE, 1205 ("vm_page_replace: mold is on a paging queue")); 1206 1207 /* Detach the old page from the resident tailq. */ 1208 TAILQ_REMOVE(&object->memq, mold, listq); 1209 1210 mold->object = NULL; 1211 vm_page_xunbusy(mold); 1212 1213 /* Insert the new page in the resident tailq. */ 1214 if (mpred != NULL) 1215 TAILQ_INSERT_AFTER(&object->memq, mpred, mnew, listq); 1216 else 1217 TAILQ_INSERT_HEAD(&object->memq, mnew, listq); 1218 if (pmap_page_is_write_mapped(mnew)) 1219 vm_object_set_writeable_dirty(object); 1220 return (mold); 1221 } 1222 1223 /* 1224 * vm_page_rename: 1225 * 1226 * Move the given memory entry from its 1227 * current object to the specified target object/offset. 1228 * 1229 * Note: swap associated with the page must be invalidated by the move. We 1230 * have to do this for several reasons: (1) we aren't freeing the 1231 * page, (2) we are dirtying the page, (3) the VM system is probably 1232 * moving the page from object A to B, and will then later move 1233 * the backing store from A to B and we can't have a conflict. 1234 * 1235 * Note: we *always* dirty the page. It is necessary both for the 1236 * fact that we moved it, and because we may be invalidating 1237 * swap. If the page is on the cache, we have to deactivate it 1238 * or vm_page_dirty() will panic. Dirty pages are not allowed 1239 * on the cache. 1240 * 1241 * The objects must be locked. 1242 */ 1243 int 1244 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1245 { 1246 vm_page_t mpred; 1247 vm_pindex_t opidx; 1248 1249 VM_OBJECT_ASSERT_WLOCKED(new_object); 1250 1251 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex); 1252 KASSERT(mpred == NULL || mpred->pindex != new_pindex, 1253 ("vm_page_rename: pindex already renamed")); 1254 1255 /* 1256 * Create a custom version of vm_page_insert() which does not depend 1257 * by m_prev and can cheat on the implementation aspects of the 1258 * function. 1259 */ 1260 opidx = m->pindex; 1261 m->pindex = new_pindex; 1262 if (vm_radix_insert(&new_object->rtree, m)) { 1263 m->pindex = opidx; 1264 return (1); 1265 } 1266 1267 /* 1268 * The operation cannot fail anymore. The removal must happen before 1269 * the listq iterator is tainted. 1270 */ 1271 m->pindex = opidx; 1272 vm_page_lock(m); 1273 vm_page_remove(m); 1274 1275 /* Return back to the new pindex to complete vm_page_insert(). */ 1276 m->pindex = new_pindex; 1277 m->object = new_object; 1278 vm_page_unlock(m); 1279 vm_page_insert_radixdone(m, new_object, mpred); 1280 vm_page_dirty(m); 1281 return (0); 1282 } 1283 1284 /* 1285 * Convert all of the given object's cached pages that have a 1286 * pindex within the given range into free pages. If the value 1287 * zero is given for "end", then the range's upper bound is 1288 * infinity. If the given object is backed by a vnode and it 1289 * transitions from having one or more cached pages to none, the 1290 * vnode's hold count is reduced. 1291 */ 1292 void 1293 vm_page_cache_free(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 1294 { 1295 vm_page_t m; 1296 boolean_t empty; 1297 1298 mtx_lock(&vm_page_queue_free_mtx); 1299 if (__predict_false(vm_radix_is_empty(&object->cache))) { 1300 mtx_unlock(&vm_page_queue_free_mtx); 1301 return; 1302 } 1303 while ((m = vm_radix_lookup_ge(&object->cache, start)) != NULL) { 1304 if (end != 0 && m->pindex >= end) 1305 break; 1306 vm_radix_remove(&object->cache, m->pindex); 1307 vm_page_cache_turn_free(m); 1308 } 1309 empty = vm_radix_is_empty(&object->cache); 1310 mtx_unlock(&vm_page_queue_free_mtx); 1311 if (object->type == OBJT_VNODE && empty) 1312 vdrop(object->handle); 1313 } 1314 1315 /* 1316 * Returns the cached page that is associated with the given 1317 * object and offset. If, however, none exists, returns NULL. 1318 * 1319 * The free page queue must be locked. 1320 */ 1321 static inline vm_page_t 1322 vm_page_cache_lookup(vm_object_t object, vm_pindex_t pindex) 1323 { 1324 1325 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1326 return (vm_radix_lookup(&object->cache, pindex)); 1327 } 1328 1329 /* 1330 * Remove the given cached page from its containing object's 1331 * collection of cached pages. 1332 * 1333 * The free page queue must be locked. 1334 */ 1335 static void 1336 vm_page_cache_remove(vm_page_t m) 1337 { 1338 1339 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1340 KASSERT((m->flags & PG_CACHED) != 0, 1341 ("vm_page_cache_remove: page %p is not cached", m)); 1342 vm_radix_remove(&m->object->cache, m->pindex); 1343 m->object = NULL; 1344 cnt.v_cache_count--; 1345 } 1346 1347 /* 1348 * Transfer all of the cached pages with offset greater than or 1349 * equal to 'offidxstart' from the original object's cache to the 1350 * new object's cache. However, any cached pages with offset 1351 * greater than or equal to the new object's size are kept in the 1352 * original object. Initially, the new object's cache must be 1353 * empty. Offset 'offidxstart' in the original object must 1354 * correspond to offset zero in the new object. 1355 * 1356 * The new object must be locked. 1357 */ 1358 void 1359 vm_page_cache_transfer(vm_object_t orig_object, vm_pindex_t offidxstart, 1360 vm_object_t new_object) 1361 { 1362 vm_page_t m; 1363 1364 /* 1365 * Insertion into an object's collection of cached pages 1366 * requires the object to be locked. In contrast, removal does 1367 * not. 1368 */ 1369 VM_OBJECT_ASSERT_WLOCKED(new_object); 1370 KASSERT(vm_radix_is_empty(&new_object->cache), 1371 ("vm_page_cache_transfer: object %p has cached pages", 1372 new_object)); 1373 mtx_lock(&vm_page_queue_free_mtx); 1374 while ((m = vm_radix_lookup_ge(&orig_object->cache, 1375 offidxstart)) != NULL) { 1376 /* 1377 * Transfer all of the pages with offset greater than or 1378 * equal to 'offidxstart' from the original object's 1379 * cache to the new object's cache. 1380 */ 1381 if ((m->pindex - offidxstart) >= new_object->size) 1382 break; 1383 vm_radix_remove(&orig_object->cache, m->pindex); 1384 /* Update the page's object and offset. */ 1385 m->object = new_object; 1386 m->pindex -= offidxstart; 1387 if (vm_radix_insert(&new_object->cache, m)) 1388 vm_page_cache_turn_free(m); 1389 } 1390 mtx_unlock(&vm_page_queue_free_mtx); 1391 } 1392 1393 /* 1394 * Returns TRUE if a cached page is associated with the given object and 1395 * offset, and FALSE otherwise. 1396 * 1397 * The object must be locked. 1398 */ 1399 boolean_t 1400 vm_page_is_cached(vm_object_t object, vm_pindex_t pindex) 1401 { 1402 vm_page_t m; 1403 1404 /* 1405 * Insertion into an object's collection of cached pages requires the 1406 * object to be locked. Therefore, if the object is locked and the 1407 * object's collection is empty, there is no need to acquire the free 1408 * page queues lock in order to prove that the specified page doesn't 1409 * exist. 1410 */ 1411 VM_OBJECT_ASSERT_WLOCKED(object); 1412 if (__predict_true(vm_object_cache_is_empty(object))) 1413 return (FALSE); 1414 mtx_lock(&vm_page_queue_free_mtx); 1415 m = vm_page_cache_lookup(object, pindex); 1416 mtx_unlock(&vm_page_queue_free_mtx); 1417 return (m != NULL); 1418 } 1419 1420 /* 1421 * vm_page_alloc: 1422 * 1423 * Allocate and return a page that is associated with the specified 1424 * object and offset pair. By default, this page is exclusive busied. 1425 * 1426 * The caller must always specify an allocation class. 1427 * 1428 * allocation classes: 1429 * VM_ALLOC_NORMAL normal process request 1430 * VM_ALLOC_SYSTEM system *really* needs a page 1431 * VM_ALLOC_INTERRUPT interrupt time request 1432 * 1433 * optional allocation flags: 1434 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1435 * intends to allocate 1436 * VM_ALLOC_IFCACHED return page only if it is cached 1437 * VM_ALLOC_IFNOTCACHED return NULL, do not reactivate if the page 1438 * is cached 1439 * VM_ALLOC_NOBUSY do not exclusive busy the page 1440 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1441 * VM_ALLOC_NOOBJ page is not associated with an object and 1442 * should not be exclusive busy 1443 * VM_ALLOC_SBUSY shared busy the allocated page 1444 * VM_ALLOC_WIRED wire the allocated page 1445 * VM_ALLOC_ZERO prefer a zeroed page 1446 * 1447 * This routine may not sleep. 1448 */ 1449 vm_page_t 1450 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 1451 { 1452 struct vnode *vp = NULL; 1453 vm_object_t m_object; 1454 vm_page_t m, mpred; 1455 int flags, req_class; 1456 1457 mpred = 0; /* XXX: pacify gcc */ 1458 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 1459 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 1460 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 1461 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 1462 ("vm_page_alloc: inconsistent object(%p)/req(%x)", (void *)object, 1463 req)); 1464 if (object != NULL) 1465 VM_OBJECT_ASSERT_WLOCKED(object); 1466 1467 req_class = req & VM_ALLOC_CLASS_MASK; 1468 1469 /* 1470 * The page daemon is allowed to dig deeper into the free page list. 1471 */ 1472 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1473 req_class = VM_ALLOC_SYSTEM; 1474 1475 if (object != NULL) { 1476 mpred = vm_radix_lookup_le(&object->rtree, pindex); 1477 KASSERT(mpred == NULL || mpred->pindex != pindex, 1478 ("vm_page_alloc: pindex already allocated")); 1479 } 1480 1481 /* 1482 * The page allocation request can came from consumers which already 1483 * hold the free page queue mutex, like vm_page_insert() in 1484 * vm_page_cache(). 1485 */ 1486 mtx_lock_flags(&vm_page_queue_free_mtx, MTX_RECURSE); 1487 if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved || 1488 (req_class == VM_ALLOC_SYSTEM && 1489 cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) || 1490 (req_class == VM_ALLOC_INTERRUPT && 1491 cnt.v_free_count + cnt.v_cache_count > 0)) { 1492 /* 1493 * Allocate from the free queue if the number of free pages 1494 * exceeds the minimum for the request class. 1495 */ 1496 if (object != NULL && 1497 (m = vm_page_cache_lookup(object, pindex)) != NULL) { 1498 if ((req & VM_ALLOC_IFNOTCACHED) != 0) { 1499 mtx_unlock(&vm_page_queue_free_mtx); 1500 return (NULL); 1501 } 1502 if (vm_phys_unfree_page(m)) 1503 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0); 1504 #if VM_NRESERVLEVEL > 0 1505 else if (!vm_reserv_reactivate_page(m)) 1506 #else 1507 else 1508 #endif 1509 panic("vm_page_alloc: cache page %p is missing" 1510 " from the free queue", m); 1511 } else if ((req & VM_ALLOC_IFCACHED) != 0) { 1512 mtx_unlock(&vm_page_queue_free_mtx); 1513 return (NULL); 1514 #if VM_NRESERVLEVEL > 0 1515 } else if (object == NULL || (object->flags & (OBJ_COLORED | 1516 OBJ_FICTITIOUS)) != OBJ_COLORED || (m = 1517 vm_reserv_alloc_page(object, pindex, mpred)) == NULL) { 1518 #else 1519 } else { 1520 #endif 1521 m = vm_phys_alloc_pages(object != NULL ? 1522 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0); 1523 #if VM_NRESERVLEVEL > 0 1524 if (m == NULL && vm_reserv_reclaim_inactive()) { 1525 m = vm_phys_alloc_pages(object != NULL ? 1526 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 1527 0); 1528 } 1529 #endif 1530 } 1531 } else { 1532 /* 1533 * Not allocatable, give up. 1534 */ 1535 mtx_unlock(&vm_page_queue_free_mtx); 1536 atomic_add_int(&vm_pageout_deficit, 1537 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 1538 pagedaemon_wakeup(); 1539 return (NULL); 1540 } 1541 1542 /* 1543 * At this point we had better have found a good page. 1544 */ 1545 KASSERT(m != NULL, ("vm_page_alloc: missing page")); 1546 KASSERT(m->queue == PQ_NONE, 1547 ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue)); 1548 KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m)); 1549 KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m)); 1550 KASSERT(!vm_page_sbusied(m), 1551 ("vm_page_alloc: page %p is busy", m)); 1552 KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m)); 1553 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 1554 ("vm_page_alloc: page %p has unexpected memattr %d", m, 1555 pmap_page_get_memattr(m))); 1556 if ((m->flags & PG_CACHED) != 0) { 1557 KASSERT((m->flags & PG_ZERO) == 0, 1558 ("vm_page_alloc: cached page %p is PG_ZERO", m)); 1559 KASSERT(m->valid != 0, 1560 ("vm_page_alloc: cached page %p is invalid", m)); 1561 if (m->object == object && m->pindex == pindex) 1562 cnt.v_reactivated++; 1563 else 1564 m->valid = 0; 1565 m_object = m->object; 1566 vm_page_cache_remove(m); 1567 if (m_object->type == OBJT_VNODE && 1568 vm_object_cache_is_empty(m_object)) 1569 vp = m_object->handle; 1570 } else { 1571 KASSERT(VM_PAGE_IS_FREE(m), 1572 ("vm_page_alloc: page %p is not free", m)); 1573 KASSERT(m->valid == 0, 1574 ("vm_page_alloc: free page %p is valid", m)); 1575 vm_phys_freecnt_adj(m, -1); 1576 } 1577 1578 /* 1579 * Only the PG_ZERO flag is inherited. The PG_CACHED or PG_FREE flag 1580 * must be cleared before the free page queues lock is released. 1581 */ 1582 flags = 0; 1583 if (m->flags & PG_ZERO) { 1584 vm_page_zero_count--; 1585 if (req & VM_ALLOC_ZERO) 1586 flags = PG_ZERO; 1587 } 1588 if (req & VM_ALLOC_NODUMP) 1589 flags |= PG_NODUMP; 1590 m->flags = flags; 1591 mtx_unlock(&vm_page_queue_free_mtx); 1592 m->aflags = 0; 1593 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? 1594 VPO_UNMANAGED : 0; 1595 m->busy_lock = VPB_UNBUSIED; 1596 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) 1597 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 1598 if ((req & VM_ALLOC_SBUSY) != 0) 1599 m->busy_lock = VPB_SHARERS_WORD(1); 1600 if (req & VM_ALLOC_WIRED) { 1601 /* 1602 * The page lock is not required for wiring a page until that 1603 * page is inserted into the object. 1604 */ 1605 atomic_add_int(&cnt.v_wire_count, 1); 1606 m->wire_count = 1; 1607 } 1608 m->act_count = 0; 1609 1610 if (object != NULL) { 1611 if (vm_page_insert_after(m, object, pindex, mpred)) { 1612 /* See the comment below about hold count. */ 1613 if (vp != NULL) 1614 vdrop(vp); 1615 pagedaemon_wakeup(); 1616 if (req & VM_ALLOC_WIRED) { 1617 atomic_subtract_int(&cnt.v_wire_count, 1); 1618 m->wire_count = 0; 1619 } 1620 m->object = NULL; 1621 vm_page_free(m); 1622 return (NULL); 1623 } 1624 1625 /* Ignore device objects; the pager sets "memattr" for them. */ 1626 if (object->memattr != VM_MEMATTR_DEFAULT && 1627 (object->flags & OBJ_FICTITIOUS) == 0) 1628 pmap_page_set_memattr(m, object->memattr); 1629 } else 1630 m->pindex = pindex; 1631 1632 /* 1633 * The following call to vdrop() must come after the above call 1634 * to vm_page_insert() in case both affect the same object and 1635 * vnode. Otherwise, the affected vnode's hold count could 1636 * temporarily become zero. 1637 */ 1638 if (vp != NULL) 1639 vdrop(vp); 1640 1641 /* 1642 * Don't wakeup too often - wakeup the pageout daemon when 1643 * we would be nearly out of memory. 1644 */ 1645 if (vm_paging_needed()) 1646 pagedaemon_wakeup(); 1647 1648 return (m); 1649 } 1650 1651 static void 1652 vm_page_alloc_contig_vdrop(struct spglist *lst) 1653 { 1654 1655 while (!SLIST_EMPTY(lst)) { 1656 vdrop((struct vnode *)SLIST_FIRST(lst)-> plinks.s.pv); 1657 SLIST_REMOVE_HEAD(lst, plinks.s.ss); 1658 } 1659 } 1660 1661 /* 1662 * vm_page_alloc_contig: 1663 * 1664 * Allocate a contiguous set of physical pages of the given size "npages" 1665 * from the free lists. All of the physical pages must be at or above 1666 * the given physical address "low" and below the given physical address 1667 * "high". The given value "alignment" determines the alignment of the 1668 * first physical page in the set. If the given value "boundary" is 1669 * non-zero, then the set of physical pages cannot cross any physical 1670 * address boundary that is a multiple of that value. Both "alignment" 1671 * and "boundary" must be a power of two. 1672 * 1673 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, 1674 * then the memory attribute setting for the physical pages is configured 1675 * to the object's memory attribute setting. Otherwise, the memory 1676 * attribute setting for the physical pages is configured to "memattr", 1677 * overriding the object's memory attribute setting. However, if the 1678 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the 1679 * memory attribute setting for the physical pages cannot be configured 1680 * to VM_MEMATTR_DEFAULT. 1681 * 1682 * The caller must always specify an allocation class. 1683 * 1684 * allocation classes: 1685 * VM_ALLOC_NORMAL normal process request 1686 * VM_ALLOC_SYSTEM system *really* needs a page 1687 * VM_ALLOC_INTERRUPT interrupt time request 1688 * 1689 * optional allocation flags: 1690 * VM_ALLOC_NOBUSY do not exclusive busy the page 1691 * VM_ALLOC_NOOBJ page is not associated with an object and 1692 * should not be exclusive busy 1693 * VM_ALLOC_SBUSY shared busy the allocated page 1694 * VM_ALLOC_WIRED wire the allocated page 1695 * VM_ALLOC_ZERO prefer a zeroed page 1696 * 1697 * This routine may not sleep. 1698 */ 1699 vm_page_t 1700 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, 1701 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 1702 vm_paddr_t boundary, vm_memattr_t memattr) 1703 { 1704 struct vnode *drop; 1705 struct spglist deferred_vdrop_list; 1706 vm_page_t m, m_tmp, m_ret; 1707 u_int flags, oflags; 1708 int req_class; 1709 1710 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 1711 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 1712 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 1713 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 1714 ("vm_page_alloc: inconsistent object(%p)/req(%x)", (void *)object, 1715 req)); 1716 if (object != NULL) { 1717 VM_OBJECT_ASSERT_WLOCKED(object); 1718 KASSERT(object->type == OBJT_PHYS, 1719 ("vm_page_alloc_contig: object %p isn't OBJT_PHYS", 1720 object)); 1721 } 1722 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 1723 req_class = req & VM_ALLOC_CLASS_MASK; 1724 1725 /* 1726 * The page daemon is allowed to dig deeper into the free page list. 1727 */ 1728 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1729 req_class = VM_ALLOC_SYSTEM; 1730 1731 SLIST_INIT(&deferred_vdrop_list); 1732 mtx_lock(&vm_page_queue_free_mtx); 1733 if (cnt.v_free_count + cnt.v_cache_count >= npages + 1734 cnt.v_free_reserved || (req_class == VM_ALLOC_SYSTEM && 1735 cnt.v_free_count + cnt.v_cache_count >= npages + 1736 cnt.v_interrupt_free_min) || (req_class == VM_ALLOC_INTERRUPT && 1737 cnt.v_free_count + cnt.v_cache_count >= npages)) { 1738 #if VM_NRESERVLEVEL > 0 1739 retry: 1740 if (object == NULL || (object->flags & OBJ_COLORED) == 0 || 1741 (m_ret = vm_reserv_alloc_contig(object, pindex, npages, 1742 low, high, alignment, boundary)) == NULL) 1743 #endif 1744 m_ret = vm_phys_alloc_contig(npages, low, high, 1745 alignment, boundary); 1746 } else { 1747 mtx_unlock(&vm_page_queue_free_mtx); 1748 atomic_add_int(&vm_pageout_deficit, npages); 1749 pagedaemon_wakeup(); 1750 return (NULL); 1751 } 1752 if (m_ret != NULL) 1753 for (m = m_ret; m < &m_ret[npages]; m++) { 1754 drop = vm_page_alloc_init(m); 1755 if (drop != NULL) { 1756 /* 1757 * Enqueue the vnode for deferred vdrop(). 1758 */ 1759 m->plinks.s.pv = drop; 1760 SLIST_INSERT_HEAD(&deferred_vdrop_list, m, 1761 plinks.s.ss); 1762 } 1763 } 1764 else { 1765 #if VM_NRESERVLEVEL > 0 1766 if (vm_reserv_reclaim_contig(npages, low, high, alignment, 1767 boundary)) 1768 goto retry; 1769 #endif 1770 } 1771 mtx_unlock(&vm_page_queue_free_mtx); 1772 if (m_ret == NULL) 1773 return (NULL); 1774 1775 /* 1776 * Initialize the pages. Only the PG_ZERO flag is inherited. 1777 */ 1778 flags = 0; 1779 if ((req & VM_ALLOC_ZERO) != 0) 1780 flags = PG_ZERO; 1781 if ((req & VM_ALLOC_NODUMP) != 0) 1782 flags |= PG_NODUMP; 1783 if ((req & VM_ALLOC_WIRED) != 0) 1784 atomic_add_int(&cnt.v_wire_count, npages); 1785 oflags = VPO_UNMANAGED; 1786 if (object != NULL) { 1787 if (object->memattr != VM_MEMATTR_DEFAULT && 1788 memattr == VM_MEMATTR_DEFAULT) 1789 memattr = object->memattr; 1790 } 1791 for (m = m_ret; m < &m_ret[npages]; m++) { 1792 m->aflags = 0; 1793 m->flags = (m->flags | PG_NODUMP) & flags; 1794 m->busy_lock = VPB_UNBUSIED; 1795 if (object != NULL) { 1796 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 1797 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 1798 if ((req & VM_ALLOC_SBUSY) != 0) 1799 m->busy_lock = VPB_SHARERS_WORD(1); 1800 } 1801 if ((req & VM_ALLOC_WIRED) != 0) 1802 m->wire_count = 1; 1803 /* Unmanaged pages don't use "act_count". */ 1804 m->oflags = oflags; 1805 if (object != NULL) { 1806 if (vm_page_insert(m, object, pindex)) { 1807 vm_page_alloc_contig_vdrop( 1808 &deferred_vdrop_list); 1809 if (vm_paging_needed()) 1810 pagedaemon_wakeup(); 1811 if ((req & VM_ALLOC_WIRED) != 0) 1812 atomic_subtract_int(&cnt.v_wire_count, 1813 npages); 1814 for (m_tmp = m, m = m_ret; 1815 m < &m_ret[npages]; m++) { 1816 if ((req & VM_ALLOC_WIRED) != 0) 1817 m->wire_count = 0; 1818 if (m >= m_tmp) 1819 m->object = NULL; 1820 vm_page_free(m); 1821 } 1822 return (NULL); 1823 } 1824 } else 1825 m->pindex = pindex; 1826 if (memattr != VM_MEMATTR_DEFAULT) 1827 pmap_page_set_memattr(m, memattr); 1828 pindex++; 1829 } 1830 vm_page_alloc_contig_vdrop(&deferred_vdrop_list); 1831 if (vm_paging_needed()) 1832 pagedaemon_wakeup(); 1833 return (m_ret); 1834 } 1835 1836 /* 1837 * Initialize a page that has been freshly dequeued from a freelist. 1838 * The caller has to drop the vnode returned, if it is not NULL. 1839 * 1840 * This function may only be used to initialize unmanaged pages. 1841 * 1842 * To be called with vm_page_queue_free_mtx held. 1843 */ 1844 static struct vnode * 1845 vm_page_alloc_init(vm_page_t m) 1846 { 1847 struct vnode *drop; 1848 vm_object_t m_object; 1849 1850 KASSERT(m->queue == PQ_NONE, 1851 ("vm_page_alloc_init: page %p has unexpected queue %d", 1852 m, m->queue)); 1853 KASSERT(m->wire_count == 0, 1854 ("vm_page_alloc_init: page %p is wired", m)); 1855 KASSERT(m->hold_count == 0, 1856 ("vm_page_alloc_init: page %p is held", m)); 1857 KASSERT(!vm_page_sbusied(m), 1858 ("vm_page_alloc_init: page %p is busy", m)); 1859 KASSERT(m->dirty == 0, 1860 ("vm_page_alloc_init: page %p is dirty", m)); 1861 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 1862 ("vm_page_alloc_init: page %p has unexpected memattr %d", 1863 m, pmap_page_get_memattr(m))); 1864 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1865 drop = NULL; 1866 if ((m->flags & PG_CACHED) != 0) { 1867 KASSERT((m->flags & PG_ZERO) == 0, 1868 ("vm_page_alloc_init: cached page %p is PG_ZERO", m)); 1869 m->valid = 0; 1870 m_object = m->object; 1871 vm_page_cache_remove(m); 1872 if (m_object->type == OBJT_VNODE && 1873 vm_object_cache_is_empty(m_object)) 1874 drop = m_object->handle; 1875 } else { 1876 KASSERT(VM_PAGE_IS_FREE(m), 1877 ("vm_page_alloc_init: page %p is not free", m)); 1878 KASSERT(m->valid == 0, 1879 ("vm_page_alloc_init: free page %p is valid", m)); 1880 vm_phys_freecnt_adj(m, -1); 1881 if ((m->flags & PG_ZERO) != 0) 1882 vm_page_zero_count--; 1883 } 1884 /* Don't clear the PG_ZERO flag; we'll need it later. */ 1885 m->flags &= PG_ZERO; 1886 return (drop); 1887 } 1888 1889 /* 1890 * vm_page_alloc_freelist: 1891 * 1892 * Allocate a physical page from the specified free page list. 1893 * 1894 * The caller must always specify an allocation class. 1895 * 1896 * allocation classes: 1897 * VM_ALLOC_NORMAL normal process request 1898 * VM_ALLOC_SYSTEM system *really* needs a page 1899 * VM_ALLOC_INTERRUPT interrupt time request 1900 * 1901 * optional allocation flags: 1902 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1903 * intends to allocate 1904 * VM_ALLOC_WIRED wire the allocated page 1905 * VM_ALLOC_ZERO prefer a zeroed page 1906 * 1907 * This routine may not sleep. 1908 */ 1909 vm_page_t 1910 vm_page_alloc_freelist(int flind, int req) 1911 { 1912 struct vnode *drop; 1913 vm_page_t m; 1914 u_int flags; 1915 int req_class; 1916 1917 req_class = req & VM_ALLOC_CLASS_MASK; 1918 1919 /* 1920 * The page daemon is allowed to dig deeper into the free page list. 1921 */ 1922 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1923 req_class = VM_ALLOC_SYSTEM; 1924 1925 /* 1926 * Do not allocate reserved pages unless the req has asked for it. 1927 */ 1928 mtx_lock_flags(&vm_page_queue_free_mtx, MTX_RECURSE); 1929 if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved || 1930 (req_class == VM_ALLOC_SYSTEM && 1931 cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) || 1932 (req_class == VM_ALLOC_INTERRUPT && 1933 cnt.v_free_count + cnt.v_cache_count > 0)) 1934 m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0); 1935 else { 1936 mtx_unlock(&vm_page_queue_free_mtx); 1937 atomic_add_int(&vm_pageout_deficit, 1938 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 1939 pagedaemon_wakeup(); 1940 return (NULL); 1941 } 1942 if (m == NULL) { 1943 mtx_unlock(&vm_page_queue_free_mtx); 1944 return (NULL); 1945 } 1946 drop = vm_page_alloc_init(m); 1947 mtx_unlock(&vm_page_queue_free_mtx); 1948 1949 /* 1950 * Initialize the page. Only the PG_ZERO flag is inherited. 1951 */ 1952 m->aflags = 0; 1953 flags = 0; 1954 if ((req & VM_ALLOC_ZERO) != 0) 1955 flags = PG_ZERO; 1956 m->flags &= flags; 1957 if ((req & VM_ALLOC_WIRED) != 0) { 1958 /* 1959 * The page lock is not required for wiring a page that does 1960 * not belong to an object. 1961 */ 1962 atomic_add_int(&cnt.v_wire_count, 1); 1963 m->wire_count = 1; 1964 } 1965 /* Unmanaged pages don't use "act_count". */ 1966 m->oflags = VPO_UNMANAGED; 1967 if (drop != NULL) 1968 vdrop(drop); 1969 if (vm_paging_needed()) 1970 pagedaemon_wakeup(); 1971 return (m); 1972 } 1973 1974 /* 1975 * vm_wait: (also see VM_WAIT macro) 1976 * 1977 * Sleep until free pages are available for allocation. 1978 * - Called in various places before memory allocations. 1979 */ 1980 void 1981 vm_wait(void) 1982 { 1983 1984 mtx_lock(&vm_page_queue_free_mtx); 1985 if (curproc == pageproc) { 1986 vm_pageout_pages_needed = 1; 1987 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx, 1988 PDROP | PSWP, "VMWait", 0); 1989 } else { 1990 if (!vm_pages_needed) { 1991 vm_pages_needed = 1; 1992 wakeup(&vm_pages_needed); 1993 } 1994 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM, 1995 "vmwait", 0); 1996 } 1997 } 1998 1999 /* 2000 * vm_waitpfault: (also see VM_WAITPFAULT macro) 2001 * 2002 * Sleep until free pages are available for allocation. 2003 * - Called only in vm_fault so that processes page faulting 2004 * can be easily tracked. 2005 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 2006 * processes will be able to grab memory first. Do not change 2007 * this balance without careful testing first. 2008 */ 2009 void 2010 vm_waitpfault(void) 2011 { 2012 2013 mtx_lock(&vm_page_queue_free_mtx); 2014 if (!vm_pages_needed) { 2015 vm_pages_needed = 1; 2016 wakeup(&vm_pages_needed); 2017 } 2018 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER, 2019 "pfault", 0); 2020 } 2021 2022 struct vm_pagequeue * 2023 vm_page_pagequeue(vm_page_t m) 2024 { 2025 2026 return (&vm_phys_domain(m)->vmd_pagequeues[m->queue]); 2027 } 2028 2029 /* 2030 * vm_page_dequeue: 2031 * 2032 * Remove the given page from its current page queue. 2033 * 2034 * The page must be locked. 2035 */ 2036 void 2037 vm_page_dequeue(vm_page_t m) 2038 { 2039 struct vm_pagequeue *pq; 2040 2041 vm_page_lock_assert(m, MA_OWNED); 2042 KASSERT(m->queue != PQ_NONE, 2043 ("vm_page_dequeue: page %p is not queued", m)); 2044 pq = vm_page_pagequeue(m); 2045 vm_pagequeue_lock(pq); 2046 m->queue = PQ_NONE; 2047 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2048 vm_pagequeue_cnt_dec(pq); 2049 vm_pagequeue_unlock(pq); 2050 } 2051 2052 /* 2053 * vm_page_dequeue_locked: 2054 * 2055 * Remove the given page from its current page queue. 2056 * 2057 * The page and page queue must be locked. 2058 */ 2059 void 2060 vm_page_dequeue_locked(vm_page_t m) 2061 { 2062 struct vm_pagequeue *pq; 2063 2064 vm_page_lock_assert(m, MA_OWNED); 2065 pq = vm_page_pagequeue(m); 2066 vm_pagequeue_assert_locked(pq); 2067 m->queue = PQ_NONE; 2068 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2069 vm_pagequeue_cnt_dec(pq); 2070 } 2071 2072 /* 2073 * vm_page_enqueue: 2074 * 2075 * Add the given page to the specified page queue. 2076 * 2077 * The page must be locked. 2078 */ 2079 static void 2080 vm_page_enqueue(int queue, vm_page_t m) 2081 { 2082 struct vm_pagequeue *pq; 2083 2084 vm_page_lock_assert(m, MA_OWNED); 2085 pq = &vm_phys_domain(m)->vmd_pagequeues[queue]; 2086 vm_pagequeue_lock(pq); 2087 m->queue = queue; 2088 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2089 vm_pagequeue_cnt_inc(pq); 2090 vm_pagequeue_unlock(pq); 2091 } 2092 2093 /* 2094 * vm_page_requeue: 2095 * 2096 * Move the given page to the tail of its current page queue. 2097 * 2098 * The page must be locked. 2099 */ 2100 void 2101 vm_page_requeue(vm_page_t m) 2102 { 2103 struct vm_pagequeue *pq; 2104 2105 vm_page_lock_assert(m, MA_OWNED); 2106 KASSERT(m->queue != PQ_NONE, 2107 ("vm_page_requeue: page %p is not queued", m)); 2108 pq = vm_page_pagequeue(m); 2109 vm_pagequeue_lock(pq); 2110 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2111 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2112 vm_pagequeue_unlock(pq); 2113 } 2114 2115 /* 2116 * vm_page_requeue_locked: 2117 * 2118 * Move the given page to the tail of its current page queue. 2119 * 2120 * The page queue must be locked. 2121 */ 2122 void 2123 vm_page_requeue_locked(vm_page_t m) 2124 { 2125 struct vm_pagequeue *pq; 2126 2127 KASSERT(m->queue != PQ_NONE, 2128 ("vm_page_requeue_locked: page %p is not queued", m)); 2129 pq = vm_page_pagequeue(m); 2130 vm_pagequeue_assert_locked(pq); 2131 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2132 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2133 } 2134 2135 /* 2136 * vm_page_activate: 2137 * 2138 * Put the specified page on the active list (if appropriate). 2139 * Ensure that act_count is at least ACT_INIT but do not otherwise 2140 * mess with it. 2141 * 2142 * The page must be locked. 2143 */ 2144 void 2145 vm_page_activate(vm_page_t m) 2146 { 2147 int queue; 2148 2149 vm_page_lock_assert(m, MA_OWNED); 2150 if ((queue = m->queue) != PQ_ACTIVE) { 2151 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 2152 if (m->act_count < ACT_INIT) 2153 m->act_count = ACT_INIT; 2154 if (queue != PQ_NONE) 2155 vm_page_dequeue(m); 2156 vm_page_enqueue(PQ_ACTIVE, m); 2157 } else 2158 KASSERT(queue == PQ_NONE, 2159 ("vm_page_activate: wired page %p is queued", m)); 2160 } else { 2161 if (m->act_count < ACT_INIT) 2162 m->act_count = ACT_INIT; 2163 } 2164 } 2165 2166 /* 2167 * vm_page_free_wakeup: 2168 * 2169 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 2170 * routine is called when a page has been added to the cache or free 2171 * queues. 2172 * 2173 * The page queues must be locked. 2174 */ 2175 static inline void 2176 vm_page_free_wakeup(void) 2177 { 2178 2179 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 2180 /* 2181 * if pageout daemon needs pages, then tell it that there are 2182 * some free. 2183 */ 2184 if (vm_pageout_pages_needed && 2185 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 2186 wakeup(&vm_pageout_pages_needed); 2187 vm_pageout_pages_needed = 0; 2188 } 2189 /* 2190 * wakeup processes that are waiting on memory if we hit a 2191 * high water mark. And wakeup scheduler process if we have 2192 * lots of memory. this process will swapin processes. 2193 */ 2194 if (vm_pages_needed && !vm_page_count_min()) { 2195 vm_pages_needed = 0; 2196 wakeup(&cnt.v_free_count); 2197 } 2198 } 2199 2200 /* 2201 * Turn a cached page into a free page, by changing its attributes. 2202 * Keep the statistics up-to-date. 2203 * 2204 * The free page queue must be locked. 2205 */ 2206 static void 2207 vm_page_cache_turn_free(vm_page_t m) 2208 { 2209 2210 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 2211 2212 m->object = NULL; 2213 m->valid = 0; 2214 /* Clear PG_CACHED and set PG_FREE. */ 2215 m->flags ^= PG_CACHED | PG_FREE; 2216 KASSERT((m->flags & (PG_CACHED | PG_FREE)) == PG_FREE, 2217 ("vm_page_cache_free: page %p has inconsistent flags", m)); 2218 cnt.v_cache_count--; 2219 vm_phys_freecnt_adj(m, 1); 2220 } 2221 2222 /* 2223 * vm_page_free_toq: 2224 * 2225 * Returns the given page to the free list, 2226 * disassociating it with any VM object. 2227 * 2228 * The object must be locked. The page must be locked if it is managed. 2229 */ 2230 void 2231 vm_page_free_toq(vm_page_t m) 2232 { 2233 2234 if ((m->oflags & VPO_UNMANAGED) == 0) { 2235 vm_page_lock_assert(m, MA_OWNED); 2236 KASSERT(!pmap_page_is_mapped(m), 2237 ("vm_page_free_toq: freeing mapped page %p", m)); 2238 } else 2239 KASSERT(m->queue == PQ_NONE, 2240 ("vm_page_free_toq: unmanaged page %p is queued", m)); 2241 PCPU_INC(cnt.v_tfree); 2242 2243 if (VM_PAGE_IS_FREE(m)) 2244 panic("vm_page_free: freeing free page %p", m); 2245 else if (vm_page_sbusied(m)) 2246 panic("vm_page_free: freeing busy page %p", m); 2247 2248 /* 2249 * Unqueue, then remove page. Note that we cannot destroy 2250 * the page here because we do not want to call the pager's 2251 * callback routine until after we've put the page on the 2252 * appropriate free queue. 2253 */ 2254 vm_page_remque(m); 2255 vm_page_remove(m); 2256 2257 /* 2258 * If fictitious remove object association and 2259 * return, otherwise delay object association removal. 2260 */ 2261 if ((m->flags & PG_FICTITIOUS) != 0) { 2262 return; 2263 } 2264 2265 m->valid = 0; 2266 vm_page_undirty(m); 2267 2268 if (m->wire_count != 0) 2269 panic("vm_page_free: freeing wired page %p", m); 2270 if (m->hold_count != 0) { 2271 m->flags &= ~PG_ZERO; 2272 KASSERT((m->flags & PG_UNHOLDFREE) == 0, 2273 ("vm_page_free: freeing PG_UNHOLDFREE page %p", m)); 2274 m->flags |= PG_UNHOLDFREE; 2275 } else { 2276 /* 2277 * Restore the default memory attribute to the page. 2278 */ 2279 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 2280 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 2281 2282 /* 2283 * Insert the page into the physical memory allocator's 2284 * cache/free page queues. 2285 */ 2286 mtx_lock(&vm_page_queue_free_mtx); 2287 m->flags |= PG_FREE; 2288 vm_phys_freecnt_adj(m, 1); 2289 #if VM_NRESERVLEVEL > 0 2290 if (!vm_reserv_free_page(m)) 2291 #else 2292 if (TRUE) 2293 #endif 2294 vm_phys_free_pages(m, 0); 2295 if ((m->flags & PG_ZERO) != 0) 2296 ++vm_page_zero_count; 2297 else 2298 vm_page_zero_idle_wakeup(); 2299 vm_page_free_wakeup(); 2300 mtx_unlock(&vm_page_queue_free_mtx); 2301 } 2302 } 2303 2304 /* 2305 * vm_page_wire: 2306 * 2307 * Mark this page as wired down by yet 2308 * another map, removing it from paging queues 2309 * as necessary. 2310 * 2311 * If the page is fictitious, then its wire count must remain one. 2312 * 2313 * The page must be locked. 2314 */ 2315 void 2316 vm_page_wire(vm_page_t m) 2317 { 2318 2319 /* 2320 * Only bump the wire statistics if the page is not already wired, 2321 * and only unqueue the page if it is on some queue (if it is unmanaged 2322 * it is already off the queues). 2323 */ 2324 vm_page_lock_assert(m, MA_OWNED); 2325 if ((m->flags & PG_FICTITIOUS) != 0) { 2326 KASSERT(m->wire_count == 1, 2327 ("vm_page_wire: fictitious page %p's wire count isn't one", 2328 m)); 2329 return; 2330 } 2331 if (m->wire_count == 0) { 2332 KASSERT((m->oflags & VPO_UNMANAGED) == 0 || 2333 m->queue == PQ_NONE, 2334 ("vm_page_wire: unmanaged page %p is queued", m)); 2335 vm_page_remque(m); 2336 atomic_add_int(&cnt.v_wire_count, 1); 2337 } 2338 m->wire_count++; 2339 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 2340 } 2341 2342 /* 2343 * vm_page_unwire: 2344 * 2345 * Release one wiring of the specified page, potentially enabling it to be 2346 * paged again. If paging is enabled, then the value of the parameter 2347 * "activate" determines to which queue the page is added. If "activate" is 2348 * non-zero, then the page is added to the active queue. Otherwise, it is 2349 * added to the inactive queue. 2350 * 2351 * However, unless the page belongs to an object, it is not enqueued because 2352 * it cannot be paged out. 2353 * 2354 * If a page is fictitious, then its wire count must always be one. 2355 * 2356 * A managed page must be locked. 2357 */ 2358 void 2359 vm_page_unwire(vm_page_t m, int activate) 2360 { 2361 2362 if ((m->oflags & VPO_UNMANAGED) == 0) 2363 vm_page_lock_assert(m, MA_OWNED); 2364 if ((m->flags & PG_FICTITIOUS) != 0) { 2365 KASSERT(m->wire_count == 1, 2366 ("vm_page_unwire: fictitious page %p's wire count isn't one", m)); 2367 return; 2368 } 2369 if (m->wire_count > 0) { 2370 m->wire_count--; 2371 if (m->wire_count == 0) { 2372 atomic_subtract_int(&cnt.v_wire_count, 1); 2373 if ((m->oflags & VPO_UNMANAGED) != 0 || 2374 m->object == NULL) 2375 return; 2376 if (!activate) 2377 m->flags &= ~PG_WINATCFLS; 2378 vm_page_enqueue(activate ? PQ_ACTIVE : PQ_INACTIVE, m); 2379 } 2380 } else 2381 panic("vm_page_unwire: page %p's wire count is zero", m); 2382 } 2383 2384 /* 2385 * Move the specified page to the inactive queue. 2386 * 2387 * Many pages placed on the inactive queue should actually go 2388 * into the cache, but it is difficult to figure out which. What 2389 * we do instead, if the inactive target is well met, is to put 2390 * clean pages at the head of the inactive queue instead of the tail. 2391 * This will cause them to be moved to the cache more quickly and 2392 * if not actively re-referenced, reclaimed more quickly. If we just 2393 * stick these pages at the end of the inactive queue, heavy filesystem 2394 * meta-data accesses can cause an unnecessary paging load on memory bound 2395 * processes. This optimization causes one-time-use metadata to be 2396 * reused more quickly. 2397 * 2398 * Normally athead is 0 resulting in LRU operation. athead is set 2399 * to 1 if we want this page to be 'as if it were placed in the cache', 2400 * except without unmapping it from the process address space. 2401 * 2402 * The page must be locked. 2403 */ 2404 static inline void 2405 _vm_page_deactivate(vm_page_t m, int athead) 2406 { 2407 struct vm_pagequeue *pq; 2408 int queue; 2409 2410 vm_page_lock_assert(m, MA_OWNED); 2411 2412 /* 2413 * Ignore if already inactive. 2414 */ 2415 if ((queue = m->queue) == PQ_INACTIVE) 2416 return; 2417 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 2418 if (queue != PQ_NONE) 2419 vm_page_dequeue(m); 2420 m->flags &= ~PG_WINATCFLS; 2421 pq = &vm_phys_domain(m)->vmd_pagequeues[PQ_INACTIVE]; 2422 vm_pagequeue_lock(pq); 2423 m->queue = PQ_INACTIVE; 2424 if (athead) 2425 TAILQ_INSERT_HEAD(&pq->pq_pl, m, plinks.q); 2426 else 2427 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2428 vm_pagequeue_cnt_inc(pq); 2429 vm_pagequeue_unlock(pq); 2430 } 2431 } 2432 2433 /* 2434 * Move the specified page to the inactive queue. 2435 * 2436 * The page must be locked. 2437 */ 2438 void 2439 vm_page_deactivate(vm_page_t m) 2440 { 2441 2442 _vm_page_deactivate(m, 0); 2443 } 2444 2445 /* 2446 * vm_page_try_to_cache: 2447 * 2448 * Returns 0 on failure, 1 on success 2449 */ 2450 int 2451 vm_page_try_to_cache(vm_page_t m) 2452 { 2453 2454 vm_page_lock_assert(m, MA_OWNED); 2455 VM_OBJECT_ASSERT_WLOCKED(m->object); 2456 if (m->dirty || m->hold_count || m->wire_count || 2457 (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m)) 2458 return (0); 2459 pmap_remove_all(m); 2460 if (m->dirty) 2461 return (0); 2462 vm_page_cache(m); 2463 return (1); 2464 } 2465 2466 /* 2467 * vm_page_try_to_free() 2468 * 2469 * Attempt to free the page. If we cannot free it, we do nothing. 2470 * 1 is returned on success, 0 on failure. 2471 */ 2472 int 2473 vm_page_try_to_free(vm_page_t m) 2474 { 2475 2476 vm_page_lock_assert(m, MA_OWNED); 2477 if (m->object != NULL) 2478 VM_OBJECT_ASSERT_WLOCKED(m->object); 2479 if (m->dirty || m->hold_count || m->wire_count || 2480 (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m)) 2481 return (0); 2482 pmap_remove_all(m); 2483 if (m->dirty) 2484 return (0); 2485 vm_page_free(m); 2486 return (1); 2487 } 2488 2489 /* 2490 * vm_page_cache 2491 * 2492 * Put the specified page onto the page cache queue (if appropriate). 2493 * 2494 * The object and page must be locked. 2495 */ 2496 void 2497 vm_page_cache(vm_page_t m) 2498 { 2499 vm_object_t object; 2500 boolean_t cache_was_empty; 2501 2502 vm_page_lock_assert(m, MA_OWNED); 2503 object = m->object; 2504 VM_OBJECT_ASSERT_WLOCKED(object); 2505 if (vm_page_busied(m) || (m->oflags & VPO_UNMANAGED) || 2506 m->hold_count || m->wire_count) 2507 panic("vm_page_cache: attempting to cache busy page"); 2508 KASSERT(!pmap_page_is_mapped(m), 2509 ("vm_page_cache: page %p is mapped", m)); 2510 KASSERT(m->dirty == 0, ("vm_page_cache: page %p is dirty", m)); 2511 if (m->valid == 0 || object->type == OBJT_DEFAULT || 2512 (object->type == OBJT_SWAP && 2513 !vm_pager_has_page(object, m->pindex, NULL, NULL))) { 2514 /* 2515 * Hypothesis: A cache-elgible page belonging to a 2516 * default object or swap object but without a backing 2517 * store must be zero filled. 2518 */ 2519 vm_page_free(m); 2520 return; 2521 } 2522 KASSERT((m->flags & PG_CACHED) == 0, 2523 ("vm_page_cache: page %p is already cached", m)); 2524 2525 /* 2526 * Remove the page from the paging queues. 2527 */ 2528 vm_page_remque(m); 2529 2530 /* 2531 * Remove the page from the object's collection of resident 2532 * pages. 2533 */ 2534 vm_radix_remove(&object->rtree, m->pindex); 2535 TAILQ_REMOVE(&object->memq, m, listq); 2536 object->resident_page_count--; 2537 2538 /* 2539 * Restore the default memory attribute to the page. 2540 */ 2541 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 2542 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 2543 2544 /* 2545 * Insert the page into the object's collection of cached pages 2546 * and the physical memory allocator's cache/free page queues. 2547 */ 2548 m->flags &= ~PG_ZERO; 2549 mtx_lock(&vm_page_queue_free_mtx); 2550 cache_was_empty = vm_radix_is_empty(&object->cache); 2551 if (vm_radix_insert(&object->cache, m)) { 2552 mtx_unlock(&vm_page_queue_free_mtx); 2553 if (object->resident_page_count == 0) 2554 vdrop(object->handle); 2555 m->object = NULL; 2556 vm_page_free(m); 2557 return; 2558 } 2559 2560 /* 2561 * The above call to vm_radix_insert() could reclaim the one pre- 2562 * existing cached page from this object, resulting in a call to 2563 * vdrop(). 2564 */ 2565 if (!cache_was_empty) 2566 cache_was_empty = vm_radix_is_singleton(&object->cache); 2567 2568 m->flags |= PG_CACHED; 2569 cnt.v_cache_count++; 2570 PCPU_INC(cnt.v_tcached); 2571 #if VM_NRESERVLEVEL > 0 2572 if (!vm_reserv_free_page(m)) { 2573 #else 2574 if (TRUE) { 2575 #endif 2576 vm_phys_set_pool(VM_FREEPOOL_CACHE, m, 0); 2577 vm_phys_free_pages(m, 0); 2578 } 2579 vm_page_free_wakeup(); 2580 mtx_unlock(&vm_page_queue_free_mtx); 2581 2582 /* 2583 * Increment the vnode's hold count if this is the object's only 2584 * cached page. Decrement the vnode's hold count if this was 2585 * the object's only resident page. 2586 */ 2587 if (object->type == OBJT_VNODE) { 2588 if (cache_was_empty && object->resident_page_count != 0) 2589 vhold(object->handle); 2590 else if (!cache_was_empty && object->resident_page_count == 0) 2591 vdrop(object->handle); 2592 } 2593 } 2594 2595 /* 2596 * vm_page_advise 2597 * 2598 * Cache, deactivate, or do nothing as appropriate. This routine 2599 * is used by madvise(). 2600 * 2601 * Generally speaking we want to move the page into the cache so 2602 * it gets reused quickly. However, this can result in a silly syndrome 2603 * due to the page recycling too quickly. Small objects will not be 2604 * fully cached. On the other hand, if we move the page to the inactive 2605 * queue we wind up with a problem whereby very large objects 2606 * unnecessarily blow away our inactive and cache queues. 2607 * 2608 * The solution is to move the pages based on a fixed weighting. We 2609 * either leave them alone, deactivate them, or move them to the cache, 2610 * where moving them to the cache has the highest weighting. 2611 * By forcing some pages into other queues we eventually force the 2612 * system to balance the queues, potentially recovering other unrelated 2613 * space from active. The idea is to not force this to happen too 2614 * often. 2615 * 2616 * The object and page must be locked. 2617 */ 2618 void 2619 vm_page_advise(vm_page_t m, int advice) 2620 { 2621 int dnw, head; 2622 2623 vm_page_assert_locked(m); 2624 VM_OBJECT_ASSERT_WLOCKED(m->object); 2625 if (advice == MADV_FREE) { 2626 /* 2627 * Mark the page clean. This will allow the page to be freed 2628 * up by the system. However, such pages are often reused 2629 * quickly by malloc() so we do not do anything that would 2630 * cause a page fault if we can help it. 2631 * 2632 * Specifically, we do not try to actually free the page now 2633 * nor do we try to put it in the cache (which would cause a 2634 * page fault on reuse). 2635 * 2636 * But we do make the page is freeable as we can without 2637 * actually taking the step of unmapping it. 2638 */ 2639 m->dirty = 0; 2640 m->act_count = 0; 2641 } else if (advice != MADV_DONTNEED) 2642 return; 2643 dnw = PCPU_GET(dnweight); 2644 PCPU_INC(dnweight); 2645 2646 /* 2647 * Occasionally leave the page alone. 2648 */ 2649 if ((dnw & 0x01F0) == 0 || m->queue == PQ_INACTIVE) { 2650 if (m->act_count >= ACT_INIT) 2651 --m->act_count; 2652 return; 2653 } 2654 2655 /* 2656 * Clear any references to the page. Otherwise, the page daemon will 2657 * immediately reactivate the page. 2658 */ 2659 vm_page_aflag_clear(m, PGA_REFERENCED); 2660 2661 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) 2662 vm_page_dirty(m); 2663 2664 if (m->dirty || (dnw & 0x0070) == 0) { 2665 /* 2666 * Deactivate the page 3 times out of 32. 2667 */ 2668 head = 0; 2669 } else { 2670 /* 2671 * Cache the page 28 times out of every 32. Note that 2672 * the page is deactivated instead of cached, but placed 2673 * at the head of the queue instead of the tail. 2674 */ 2675 head = 1; 2676 } 2677 _vm_page_deactivate(m, head); 2678 } 2679 2680 /* 2681 * Grab a page, waiting until we are waken up due to the page 2682 * changing state. We keep on waiting, if the page continues 2683 * to be in the object. If the page doesn't exist, first allocate it 2684 * and then conditionally zero it. 2685 * 2686 * This routine may sleep. 2687 * 2688 * The object must be locked on entry. The lock will, however, be released 2689 * and reacquired if the routine sleeps. 2690 */ 2691 vm_page_t 2692 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 2693 { 2694 vm_page_t m; 2695 int sleep; 2696 2697 VM_OBJECT_ASSERT_WLOCKED(object); 2698 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 2699 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 2700 ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 2701 retrylookup: 2702 if ((m = vm_page_lookup(object, pindex)) != NULL) { 2703 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ? 2704 vm_page_xbusied(m) : vm_page_busied(m); 2705 if (sleep) { 2706 /* 2707 * Reference the page before unlocking and 2708 * sleeping so that the page daemon is less 2709 * likely to reclaim it. 2710 */ 2711 vm_page_aflag_set(m, PGA_REFERENCED); 2712 vm_page_lock(m); 2713 VM_OBJECT_WUNLOCK(object); 2714 vm_page_busy_sleep(m, "pgrbwt"); 2715 VM_OBJECT_WLOCK(object); 2716 goto retrylookup; 2717 } else { 2718 if ((allocflags & VM_ALLOC_WIRED) != 0) { 2719 vm_page_lock(m); 2720 vm_page_wire(m); 2721 vm_page_unlock(m); 2722 } 2723 if ((allocflags & 2724 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 2725 vm_page_xbusy(m); 2726 if ((allocflags & VM_ALLOC_SBUSY) != 0) 2727 vm_page_sbusy(m); 2728 return (m); 2729 } 2730 } 2731 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_IGN_SBUSY); 2732 if (m == NULL) { 2733 VM_OBJECT_WUNLOCK(object); 2734 VM_WAIT; 2735 VM_OBJECT_WLOCK(object); 2736 goto retrylookup; 2737 } else if (m->valid != 0) 2738 return (m); 2739 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 2740 pmap_zero_page(m); 2741 return (m); 2742 } 2743 2744 /* 2745 * Mapping function for valid or dirty bits in a page. 2746 * 2747 * Inputs are required to range within a page. 2748 */ 2749 vm_page_bits_t 2750 vm_page_bits(int base, int size) 2751 { 2752 int first_bit; 2753 int last_bit; 2754 2755 KASSERT( 2756 base + size <= PAGE_SIZE, 2757 ("vm_page_bits: illegal base/size %d/%d", base, size) 2758 ); 2759 2760 if (size == 0) /* handle degenerate case */ 2761 return (0); 2762 2763 first_bit = base >> DEV_BSHIFT; 2764 last_bit = (base + size - 1) >> DEV_BSHIFT; 2765 2766 return (((vm_page_bits_t)2 << last_bit) - 2767 ((vm_page_bits_t)1 << first_bit)); 2768 } 2769 2770 /* 2771 * vm_page_set_valid_range: 2772 * 2773 * Sets portions of a page valid. The arguments are expected 2774 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2775 * of any partial chunks touched by the range. The invalid portion of 2776 * such chunks will be zeroed. 2777 * 2778 * (base + size) must be less then or equal to PAGE_SIZE. 2779 */ 2780 void 2781 vm_page_set_valid_range(vm_page_t m, int base, int size) 2782 { 2783 int endoff, frag; 2784 2785 VM_OBJECT_ASSERT_WLOCKED(m->object); 2786 if (size == 0) /* handle degenerate case */ 2787 return; 2788 2789 /* 2790 * If the base is not DEV_BSIZE aligned and the valid 2791 * bit is clear, we have to zero out a portion of the 2792 * first block. 2793 */ 2794 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 2795 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 2796 pmap_zero_page_area(m, frag, base - frag); 2797 2798 /* 2799 * If the ending offset is not DEV_BSIZE aligned and the 2800 * valid bit is clear, we have to zero out a portion of 2801 * the last block. 2802 */ 2803 endoff = base + size; 2804 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 2805 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 2806 pmap_zero_page_area(m, endoff, 2807 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 2808 2809 /* 2810 * Assert that no previously invalid block that is now being validated 2811 * is already dirty. 2812 */ 2813 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 2814 ("vm_page_set_valid_range: page %p is dirty", m)); 2815 2816 /* 2817 * Set valid bits inclusive of any overlap. 2818 */ 2819 m->valid |= vm_page_bits(base, size); 2820 } 2821 2822 /* 2823 * Clear the given bits from the specified page's dirty field. 2824 */ 2825 static __inline void 2826 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) 2827 { 2828 uintptr_t addr; 2829 #if PAGE_SIZE < 16384 2830 int shift; 2831 #endif 2832 2833 /* 2834 * If the object is locked and the page is neither exclusive busy nor 2835 * write mapped, then the page's dirty field cannot possibly be 2836 * set by a concurrent pmap operation. 2837 */ 2838 VM_OBJECT_ASSERT_WLOCKED(m->object); 2839 if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) 2840 m->dirty &= ~pagebits; 2841 else { 2842 /* 2843 * The pmap layer can call vm_page_dirty() without 2844 * holding a distinguished lock. The combination of 2845 * the object's lock and an atomic operation suffice 2846 * to guarantee consistency of the page dirty field. 2847 * 2848 * For PAGE_SIZE == 32768 case, compiler already 2849 * properly aligns the dirty field, so no forcible 2850 * alignment is needed. Only require existence of 2851 * atomic_clear_64 when page size is 32768. 2852 */ 2853 addr = (uintptr_t)&m->dirty; 2854 #if PAGE_SIZE == 32768 2855 atomic_clear_64((uint64_t *)addr, pagebits); 2856 #elif PAGE_SIZE == 16384 2857 atomic_clear_32((uint32_t *)addr, pagebits); 2858 #else /* PAGE_SIZE <= 8192 */ 2859 /* 2860 * Use a trick to perform a 32-bit atomic on the 2861 * containing aligned word, to not depend on the existence 2862 * of atomic_clear_{8, 16}. 2863 */ 2864 shift = addr & (sizeof(uint32_t) - 1); 2865 #if BYTE_ORDER == BIG_ENDIAN 2866 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY; 2867 #else 2868 shift *= NBBY; 2869 #endif 2870 addr &= ~(sizeof(uint32_t) - 1); 2871 atomic_clear_32((uint32_t *)addr, pagebits << shift); 2872 #endif /* PAGE_SIZE */ 2873 } 2874 } 2875 2876 /* 2877 * vm_page_set_validclean: 2878 * 2879 * Sets portions of a page valid and clean. The arguments are expected 2880 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2881 * of any partial chunks touched by the range. The invalid portion of 2882 * such chunks will be zero'd. 2883 * 2884 * (base + size) must be less then or equal to PAGE_SIZE. 2885 */ 2886 void 2887 vm_page_set_validclean(vm_page_t m, int base, int size) 2888 { 2889 vm_page_bits_t oldvalid, pagebits; 2890 int endoff, frag; 2891 2892 VM_OBJECT_ASSERT_WLOCKED(m->object); 2893 if (size == 0) /* handle degenerate case */ 2894 return; 2895 2896 /* 2897 * If the base is not DEV_BSIZE aligned and the valid 2898 * bit is clear, we have to zero out a portion of the 2899 * first block. 2900 */ 2901 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 2902 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) 2903 pmap_zero_page_area(m, frag, base - frag); 2904 2905 /* 2906 * If the ending offset is not DEV_BSIZE aligned and the 2907 * valid bit is clear, we have to zero out a portion of 2908 * the last block. 2909 */ 2910 endoff = base + size; 2911 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 2912 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) 2913 pmap_zero_page_area(m, endoff, 2914 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 2915 2916 /* 2917 * Set valid, clear dirty bits. If validating the entire 2918 * page we can safely clear the pmap modify bit. We also 2919 * use this opportunity to clear the VPO_NOSYNC flag. If a process 2920 * takes a write fault on a MAP_NOSYNC memory area the flag will 2921 * be set again. 2922 * 2923 * We set valid bits inclusive of any overlap, but we can only 2924 * clear dirty bits for DEV_BSIZE chunks that are fully within 2925 * the range. 2926 */ 2927 oldvalid = m->valid; 2928 pagebits = vm_page_bits(base, size); 2929 m->valid |= pagebits; 2930 #if 0 /* NOT YET */ 2931 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 2932 frag = DEV_BSIZE - frag; 2933 base += frag; 2934 size -= frag; 2935 if (size < 0) 2936 size = 0; 2937 } 2938 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 2939 #endif 2940 if (base == 0 && size == PAGE_SIZE) { 2941 /* 2942 * The page can only be modified within the pmap if it is 2943 * mapped, and it can only be mapped if it was previously 2944 * fully valid. 2945 */ 2946 if (oldvalid == VM_PAGE_BITS_ALL) 2947 /* 2948 * Perform the pmap_clear_modify() first. Otherwise, 2949 * a concurrent pmap operation, such as 2950 * pmap_protect(), could clear a modification in the 2951 * pmap and set the dirty field on the page before 2952 * pmap_clear_modify() had begun and after the dirty 2953 * field was cleared here. 2954 */ 2955 pmap_clear_modify(m); 2956 m->dirty = 0; 2957 m->oflags &= ~VPO_NOSYNC; 2958 } else if (oldvalid != VM_PAGE_BITS_ALL) 2959 m->dirty &= ~pagebits; 2960 else 2961 vm_page_clear_dirty_mask(m, pagebits); 2962 } 2963 2964 void 2965 vm_page_clear_dirty(vm_page_t m, int base, int size) 2966 { 2967 2968 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 2969 } 2970 2971 /* 2972 * vm_page_set_invalid: 2973 * 2974 * Invalidates DEV_BSIZE'd chunks within a page. Both the 2975 * valid and dirty bits for the effected areas are cleared. 2976 */ 2977 void 2978 vm_page_set_invalid(vm_page_t m, int base, int size) 2979 { 2980 vm_page_bits_t bits; 2981 vm_object_t object; 2982 2983 object = m->object; 2984 VM_OBJECT_ASSERT_WLOCKED(object); 2985 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + 2986 size >= object->un_pager.vnp.vnp_size) 2987 bits = VM_PAGE_BITS_ALL; 2988 else 2989 bits = vm_page_bits(base, size); 2990 if (m->valid == VM_PAGE_BITS_ALL && bits != 0) 2991 pmap_remove_all(m); 2992 KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) || 2993 !pmap_page_is_mapped(m), 2994 ("vm_page_set_invalid: page %p is mapped", m)); 2995 m->valid &= ~bits; 2996 m->dirty &= ~bits; 2997 } 2998 2999 /* 3000 * vm_page_zero_invalid() 3001 * 3002 * The kernel assumes that the invalid portions of a page contain 3003 * garbage, but such pages can be mapped into memory by user code. 3004 * When this occurs, we must zero out the non-valid portions of the 3005 * page so user code sees what it expects. 3006 * 3007 * Pages are most often semi-valid when the end of a file is mapped 3008 * into memory and the file's size is not page aligned. 3009 */ 3010 void 3011 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 3012 { 3013 int b; 3014 int i; 3015 3016 VM_OBJECT_ASSERT_WLOCKED(m->object); 3017 /* 3018 * Scan the valid bits looking for invalid sections that 3019 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 3020 * valid bit may be set ) have already been zerod by 3021 * vm_page_set_validclean(). 3022 */ 3023 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 3024 if (i == (PAGE_SIZE / DEV_BSIZE) || 3025 (m->valid & ((vm_page_bits_t)1 << i))) { 3026 if (i > b) { 3027 pmap_zero_page_area(m, 3028 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 3029 } 3030 b = i + 1; 3031 } 3032 } 3033 3034 /* 3035 * setvalid is TRUE when we can safely set the zero'd areas 3036 * as being valid. We can do this if there are no cache consistancy 3037 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 3038 */ 3039 if (setvalid) 3040 m->valid = VM_PAGE_BITS_ALL; 3041 } 3042 3043 /* 3044 * vm_page_is_valid: 3045 * 3046 * Is (partial) page valid? Note that the case where size == 0 3047 * will return FALSE in the degenerate case where the page is 3048 * entirely invalid, and TRUE otherwise. 3049 */ 3050 int 3051 vm_page_is_valid(vm_page_t m, int base, int size) 3052 { 3053 vm_page_bits_t bits; 3054 3055 VM_OBJECT_ASSERT_LOCKED(m->object); 3056 bits = vm_page_bits(base, size); 3057 return (m->valid != 0 && (m->valid & bits) == bits); 3058 } 3059 3060 /* 3061 * Set the page's dirty bits if the page is modified. 3062 */ 3063 void 3064 vm_page_test_dirty(vm_page_t m) 3065 { 3066 3067 VM_OBJECT_ASSERT_WLOCKED(m->object); 3068 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 3069 vm_page_dirty(m); 3070 } 3071 3072 void 3073 vm_page_lock_KBI(vm_page_t m, const char *file, int line) 3074 { 3075 3076 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); 3077 } 3078 3079 void 3080 vm_page_unlock_KBI(vm_page_t m, const char *file, int line) 3081 { 3082 3083 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); 3084 } 3085 3086 int 3087 vm_page_trylock_KBI(vm_page_t m, const char *file, int line) 3088 { 3089 3090 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); 3091 } 3092 3093 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) 3094 void 3095 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line) 3096 { 3097 3098 vm_page_lock_assert_KBI(m, MA_OWNED, file, line); 3099 } 3100 3101 void 3102 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) 3103 { 3104 3105 mtx_assert_(vm_page_lockptr(m), a, file, line); 3106 } 3107 #endif 3108 3109 #ifdef INVARIANTS 3110 void 3111 vm_page_object_lock_assert(vm_page_t m) 3112 { 3113 3114 /* 3115 * Certain of the page's fields may only be modified by the 3116 * holder of the containing object's lock or the exclusive busy. 3117 * holder. Unfortunately, the holder of the write busy is 3118 * not recorded, and thus cannot be checked here. 3119 */ 3120 if (m->object != NULL && !vm_page_xbusied(m)) 3121 VM_OBJECT_ASSERT_WLOCKED(m->object); 3122 } 3123 #endif 3124 3125 #include "opt_ddb.h" 3126 #ifdef DDB 3127 #include <sys/kernel.h> 3128 3129 #include <ddb/ddb.h> 3130 3131 DB_SHOW_COMMAND(page, vm_page_print_page_info) 3132 { 3133 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 3134 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 3135 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 3136 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 3137 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 3138 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 3139 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 3140 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 3141 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 3142 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 3143 } 3144 3145 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 3146 { 3147 int dom; 3148 3149 db_printf("pq_free %d pq_cache %d\n", 3150 cnt.v_free_count, cnt.v_cache_count); 3151 for (dom = 0; dom < vm_ndomains; dom++) { 3152 db_printf( 3153 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pass %d\n", 3154 dom, 3155 vm_dom[dom].vmd_page_count, 3156 vm_dom[dom].vmd_free_count, 3157 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, 3158 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, 3159 vm_dom[dom].vmd_pass); 3160 } 3161 } 3162 3163 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) 3164 { 3165 vm_page_t m; 3166 boolean_t phys; 3167 3168 if (!have_addr) { 3169 db_printf("show pginfo addr\n"); 3170 return; 3171 } 3172 3173 phys = strchr(modif, 'p') != NULL; 3174 if (phys) 3175 m = PHYS_TO_VM_PAGE(addr); 3176 else 3177 m = (vm_page_t)addr; 3178 db_printf( 3179 "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n" 3180 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", 3181 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, 3182 m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags, 3183 m->flags, m->act_count, m->busy_lock, m->valid, m->dirty); 3184 } 3185 #endif /* DDB */ 3186