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 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN, &boot_pages, 0, 138 "number of pages allocated for bootstrapping the VM system"); 139 140 static int pa_tryrelock_restart; 141 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD, 142 &pa_tryrelock_restart, 0, "Number of tryrelock restarts"); 143 144 static uma_zone_t fakepg_zone; 145 146 static struct vnode *vm_page_alloc_init(vm_page_t m); 147 static void vm_page_cache_turn_free(vm_page_t m); 148 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits); 149 static void vm_page_enqueue(uint8_t queue, vm_page_t m); 150 static void vm_page_init_fakepg(void *dummy); 151 static int vm_page_insert_after(vm_page_t m, vm_object_t object, 152 vm_pindex_t pindex, vm_page_t mpred); 153 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object, 154 vm_page_t mpred); 155 156 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init_fakepg, NULL); 157 158 static void 159 vm_page_init_fakepg(void *dummy) 160 { 161 162 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL, 163 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM); 164 } 165 166 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */ 167 #if PAGE_SIZE == 32768 168 #ifdef CTASSERT 169 CTASSERT(sizeof(u_long) >= 8); 170 #endif 171 #endif 172 173 /* 174 * Try to acquire a physical address lock while a pmap is locked. If we 175 * fail to trylock we unlock and lock the pmap directly and cache the 176 * locked pa in *locked. The caller should then restart their loop in case 177 * the virtual to physical mapping has changed. 178 */ 179 int 180 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked) 181 { 182 vm_paddr_t lockpa; 183 184 lockpa = *locked; 185 *locked = pa; 186 if (lockpa) { 187 PA_LOCK_ASSERT(lockpa, MA_OWNED); 188 if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa)) 189 return (0); 190 PA_UNLOCK(lockpa); 191 } 192 if (PA_TRYLOCK(pa)) 193 return (0); 194 PMAP_UNLOCK(pmap); 195 atomic_add_int(&pa_tryrelock_restart, 1); 196 PA_LOCK(pa); 197 PMAP_LOCK(pmap); 198 return (EAGAIN); 199 } 200 201 /* 202 * vm_set_page_size: 203 * 204 * Sets the page size, perhaps based upon the memory 205 * size. Must be called before any use of page-size 206 * dependent functions. 207 */ 208 void 209 vm_set_page_size(void) 210 { 211 if (vm_cnt.v_page_size == 0) 212 vm_cnt.v_page_size = PAGE_SIZE; 213 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0) 214 panic("vm_set_page_size: page size not a power of two"); 215 } 216 217 /* 218 * vm_page_blacklist_lookup: 219 * 220 * See if a physical address in this page has been listed 221 * in the blacklist tunable. Entries in the tunable are 222 * separated by spaces or commas. If an invalid integer is 223 * encountered then the rest of the string is skipped. 224 */ 225 static int 226 vm_page_blacklist_lookup(char *list, vm_paddr_t pa) 227 { 228 vm_paddr_t bad; 229 char *cp, *pos; 230 231 for (pos = list; *pos != '\0'; pos = cp) { 232 bad = strtoq(pos, &cp, 0); 233 if (*cp != '\0') { 234 if (*cp == ' ' || *cp == ',') { 235 cp++; 236 if (cp == pos) 237 continue; 238 } else 239 break; 240 } 241 if (pa == trunc_page(bad)) 242 return (1); 243 } 244 return (0); 245 } 246 247 static void 248 vm_page_domain_init(struct vm_domain *vmd) 249 { 250 struct vm_pagequeue *pq; 251 int i; 252 253 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) = 254 "vm inactive pagequeue"; 255 *__DECONST(int **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_vcnt) = 256 &vm_cnt.v_inactive_count; 257 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) = 258 "vm active pagequeue"; 259 *__DECONST(int **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_vcnt) = 260 &vm_cnt.v_active_count; 261 vmd->vmd_page_count = 0; 262 vmd->vmd_free_count = 0; 263 vmd->vmd_segs = 0; 264 vmd->vmd_oom = FALSE; 265 vmd->vmd_pass = 0; 266 for (i = 0; i < PQ_COUNT; i++) { 267 pq = &vmd->vmd_pagequeues[i]; 268 TAILQ_INIT(&pq->pq_pl); 269 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue", 270 MTX_DEF | MTX_DUPOK); 271 } 272 } 273 274 /* 275 * vm_page_startup: 276 * 277 * Initializes the resident memory module. 278 * 279 * Allocates memory for the page cells, and 280 * for the object/offset-to-page hash table headers. 281 * Each page cell is initialized and placed on the free list. 282 */ 283 vm_offset_t 284 vm_page_startup(vm_offset_t vaddr) 285 { 286 vm_offset_t mapped; 287 vm_paddr_t page_range; 288 vm_paddr_t new_end; 289 int i; 290 vm_paddr_t pa; 291 vm_paddr_t last_pa; 292 char *list; 293 294 /* the biggest memory array is the second group of pages */ 295 vm_paddr_t end; 296 vm_paddr_t biggestsize; 297 vm_paddr_t low_water, high_water; 298 int biggestone; 299 300 biggestsize = 0; 301 biggestone = 0; 302 vaddr = round_page(vaddr); 303 304 for (i = 0; phys_avail[i + 1]; i += 2) { 305 phys_avail[i] = round_page(phys_avail[i]); 306 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 307 } 308 309 low_water = phys_avail[0]; 310 high_water = phys_avail[1]; 311 312 for (i = 0; phys_avail[i + 1]; i += 2) { 313 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i]; 314 315 if (size > biggestsize) { 316 biggestone = i; 317 biggestsize = size; 318 } 319 if (phys_avail[i] < low_water) 320 low_water = phys_avail[i]; 321 if (phys_avail[i + 1] > high_water) 322 high_water = phys_avail[i + 1]; 323 } 324 325 #ifdef XEN 326 low_water = 0; 327 #endif 328 329 end = phys_avail[biggestone+1]; 330 331 /* 332 * Initialize the page and queue locks. 333 */ 334 mtx_init(&vm_page_queue_free_mtx, "vm page free queue", NULL, MTX_DEF); 335 for (i = 0; i < PA_LOCK_COUNT; i++) 336 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF); 337 for (i = 0; i < vm_ndomains; i++) 338 vm_page_domain_init(&vm_dom[i]); 339 340 /* 341 * Allocate memory for use when boot strapping the kernel memory 342 * allocator. 343 */ 344 new_end = end - (boot_pages * UMA_SLAB_SIZE); 345 new_end = trunc_page(new_end); 346 mapped = pmap_map(&vaddr, new_end, end, 347 VM_PROT_READ | VM_PROT_WRITE); 348 bzero((void *)mapped, end - new_end); 349 uma_startup((void *)mapped, boot_pages); 350 351 #if defined(__amd64__) || defined(__i386__) || defined(__arm__) || \ 352 defined(__mips__) 353 /* 354 * Allocate a bitmap to indicate that a random physical page 355 * needs to be included in a minidump. 356 * 357 * The amd64 port needs this to indicate which direct map pages 358 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 359 * 360 * However, i386 still needs this workspace internally within the 361 * minidump code. In theory, they are not needed on i386, but are 362 * included should the sf_buf code decide to use them. 363 */ 364 last_pa = 0; 365 for (i = 0; dump_avail[i + 1] != 0; i += 2) 366 if (dump_avail[i + 1] > last_pa) 367 last_pa = dump_avail[i + 1]; 368 page_range = last_pa / PAGE_SIZE; 369 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 370 new_end -= vm_page_dump_size; 371 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, 372 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); 373 bzero((void *)vm_page_dump, vm_page_dump_size); 374 #endif 375 #ifdef __amd64__ 376 /* 377 * Request that the physical pages underlying the message buffer be 378 * included in a crash dump. Since the message buffer is accessed 379 * through the direct map, they are not automatically included. 380 */ 381 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); 382 last_pa = pa + round_page(msgbufsize); 383 while (pa < last_pa) { 384 dump_add_page(pa); 385 pa += PAGE_SIZE; 386 } 387 #endif 388 /* 389 * Compute the number of pages of memory that will be available for 390 * use (taking into account the overhead of a page structure per 391 * page). 392 */ 393 first_page = low_water / PAGE_SIZE; 394 #ifdef VM_PHYSSEG_SPARSE 395 page_range = 0; 396 for (i = 0; phys_avail[i + 1] != 0; i += 2) 397 page_range += atop(phys_avail[i + 1] - phys_avail[i]); 398 #elif defined(VM_PHYSSEG_DENSE) 399 page_range = high_water / PAGE_SIZE - first_page; 400 #else 401 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 402 #endif 403 end = new_end; 404 405 /* 406 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 407 */ 408 vaddr += PAGE_SIZE; 409 410 /* 411 * Initialize the mem entry structures now, and put them in the free 412 * queue. 413 */ 414 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 415 mapped = pmap_map(&vaddr, new_end, end, 416 VM_PROT_READ | VM_PROT_WRITE); 417 vm_page_array = (vm_page_t) mapped; 418 #if VM_NRESERVLEVEL > 0 419 /* 420 * Allocate memory for the reservation management system's data 421 * structures. 422 */ 423 new_end = vm_reserv_startup(&vaddr, new_end, high_water); 424 #endif 425 #if defined(__amd64__) || defined(__mips__) 426 /* 427 * pmap_map on amd64 and mips can come out of the direct-map, not kvm 428 * like i386, so the pages must be tracked for a crashdump to include 429 * this data. This includes the vm_page_array and the early UMA 430 * bootstrap pages. 431 */ 432 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE) 433 dump_add_page(pa); 434 #endif 435 phys_avail[biggestone + 1] = new_end; 436 437 /* 438 * Clear all of the page structures 439 */ 440 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 441 for (i = 0; i < page_range; i++) 442 vm_page_array[i].order = VM_NFREEORDER; 443 vm_page_array_size = page_range; 444 445 /* 446 * Initialize the physical memory allocator. 447 */ 448 vm_phys_init(); 449 450 /* 451 * Add every available physical page that is not blacklisted to 452 * the free lists. 453 */ 454 vm_cnt.v_page_count = 0; 455 vm_cnt.v_free_count = 0; 456 list = getenv("vm.blacklist"); 457 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 458 pa = phys_avail[i]; 459 last_pa = phys_avail[i + 1]; 460 while (pa < last_pa) { 461 if (list != NULL && 462 vm_page_blacklist_lookup(list, pa)) 463 printf("Skipping page with pa 0x%jx\n", 464 (uintmax_t)pa); 465 else 466 vm_phys_add_page(pa); 467 pa += PAGE_SIZE; 468 } 469 } 470 freeenv(list); 471 #if VM_NRESERVLEVEL > 0 472 /* 473 * Initialize the reservation management system. 474 */ 475 vm_reserv_init(); 476 #endif 477 return (vaddr); 478 } 479 480 void 481 vm_page_reference(vm_page_t m) 482 { 483 484 vm_page_aflag_set(m, PGA_REFERENCED); 485 } 486 487 /* 488 * vm_page_busy_downgrade: 489 * 490 * Downgrade an exclusive busy page into a single shared busy page. 491 */ 492 void 493 vm_page_busy_downgrade(vm_page_t m) 494 { 495 u_int x; 496 497 vm_page_assert_xbusied(m); 498 499 for (;;) { 500 x = m->busy_lock; 501 x &= VPB_BIT_WAITERS; 502 if (atomic_cmpset_rel_int(&m->busy_lock, 503 VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1) | x)) 504 break; 505 } 506 } 507 508 /* 509 * vm_page_sbusied: 510 * 511 * Return a positive value if the page is shared busied, 0 otherwise. 512 */ 513 int 514 vm_page_sbusied(vm_page_t m) 515 { 516 u_int x; 517 518 x = m->busy_lock; 519 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED); 520 } 521 522 /* 523 * vm_page_sunbusy: 524 * 525 * Shared unbusy a page. 526 */ 527 void 528 vm_page_sunbusy(vm_page_t m) 529 { 530 u_int x; 531 532 vm_page_assert_sbusied(m); 533 534 for (;;) { 535 x = m->busy_lock; 536 if (VPB_SHARERS(x) > 1) { 537 if (atomic_cmpset_int(&m->busy_lock, x, 538 x - VPB_ONE_SHARER)) 539 break; 540 continue; 541 } 542 if ((x & VPB_BIT_WAITERS) == 0) { 543 KASSERT(x == VPB_SHARERS_WORD(1), 544 ("vm_page_sunbusy: invalid lock state")); 545 if (atomic_cmpset_int(&m->busy_lock, 546 VPB_SHARERS_WORD(1), VPB_UNBUSIED)) 547 break; 548 continue; 549 } 550 KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS), 551 ("vm_page_sunbusy: invalid lock state for waiters")); 552 553 vm_page_lock(m); 554 if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) { 555 vm_page_unlock(m); 556 continue; 557 } 558 wakeup(m); 559 vm_page_unlock(m); 560 break; 561 } 562 } 563 564 /* 565 * vm_page_busy_sleep: 566 * 567 * Sleep and release the page lock, using the page pointer as wchan. 568 * This is used to implement the hard-path of busying mechanism. 569 * 570 * The given page must be locked. 571 */ 572 void 573 vm_page_busy_sleep(vm_page_t m, const char *wmesg) 574 { 575 u_int x; 576 577 vm_page_lock_assert(m, MA_OWNED); 578 579 x = m->busy_lock; 580 if (x == VPB_UNBUSIED) { 581 vm_page_unlock(m); 582 return; 583 } 584 if ((x & VPB_BIT_WAITERS) == 0 && 585 !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS)) { 586 vm_page_unlock(m); 587 return; 588 } 589 msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0); 590 } 591 592 /* 593 * vm_page_trysbusy: 594 * 595 * Try to shared busy a page. 596 * If the operation succeeds 1 is returned otherwise 0. 597 * The operation never sleeps. 598 */ 599 int 600 vm_page_trysbusy(vm_page_t m) 601 { 602 u_int x; 603 604 for (;;) { 605 x = m->busy_lock; 606 if ((x & VPB_BIT_SHARED) == 0) 607 return (0); 608 if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER)) 609 return (1); 610 } 611 } 612 613 /* 614 * vm_page_xunbusy_hard: 615 * 616 * Called after the first try the exclusive unbusy of a page failed. 617 * It is assumed that the waiters bit is on. 618 */ 619 void 620 vm_page_xunbusy_hard(vm_page_t m) 621 { 622 623 vm_page_assert_xbusied(m); 624 625 vm_page_lock(m); 626 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); 627 wakeup(m); 628 vm_page_unlock(m); 629 } 630 631 /* 632 * vm_page_flash: 633 * 634 * Wakeup anyone waiting for the page. 635 * The ownership bits do not change. 636 * 637 * The given page must be locked. 638 */ 639 void 640 vm_page_flash(vm_page_t m) 641 { 642 u_int x; 643 644 vm_page_lock_assert(m, MA_OWNED); 645 646 for (;;) { 647 x = m->busy_lock; 648 if ((x & VPB_BIT_WAITERS) == 0) 649 return; 650 if (atomic_cmpset_int(&m->busy_lock, x, 651 x & (~VPB_BIT_WAITERS))) 652 break; 653 } 654 wakeup(m); 655 } 656 657 /* 658 * Keep page from being freed by the page daemon 659 * much of the same effect as wiring, except much lower 660 * overhead and should be used only for *very* temporary 661 * holding ("wiring"). 662 */ 663 void 664 vm_page_hold(vm_page_t mem) 665 { 666 667 vm_page_lock_assert(mem, MA_OWNED); 668 mem->hold_count++; 669 } 670 671 void 672 vm_page_unhold(vm_page_t mem) 673 { 674 675 vm_page_lock_assert(mem, MA_OWNED); 676 KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!")); 677 --mem->hold_count; 678 if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0) 679 vm_page_free_toq(mem); 680 } 681 682 /* 683 * vm_page_unhold_pages: 684 * 685 * Unhold each of the pages that is referenced by the given array. 686 */ 687 void 688 vm_page_unhold_pages(vm_page_t *ma, int count) 689 { 690 struct mtx *mtx, *new_mtx; 691 692 mtx = NULL; 693 for (; count != 0; count--) { 694 /* 695 * Avoid releasing and reacquiring the same page lock. 696 */ 697 new_mtx = vm_page_lockptr(*ma); 698 if (mtx != new_mtx) { 699 if (mtx != NULL) 700 mtx_unlock(mtx); 701 mtx = new_mtx; 702 mtx_lock(mtx); 703 } 704 vm_page_unhold(*ma); 705 ma++; 706 } 707 if (mtx != NULL) 708 mtx_unlock(mtx); 709 } 710 711 vm_page_t 712 PHYS_TO_VM_PAGE(vm_paddr_t pa) 713 { 714 vm_page_t m; 715 716 #ifdef VM_PHYSSEG_SPARSE 717 m = vm_phys_paddr_to_vm_page(pa); 718 if (m == NULL) 719 m = vm_phys_fictitious_to_vm_page(pa); 720 return (m); 721 #elif defined(VM_PHYSSEG_DENSE) 722 long pi; 723 724 pi = atop(pa); 725 if (pi >= first_page && (pi - first_page) < vm_page_array_size) { 726 m = &vm_page_array[pi - first_page]; 727 return (m); 728 } 729 return (vm_phys_fictitious_to_vm_page(pa)); 730 #else 731 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 732 #endif 733 } 734 735 /* 736 * vm_page_getfake: 737 * 738 * Create a fictitious page with the specified physical address and 739 * memory attribute. The memory attribute is the only the machine- 740 * dependent aspect of a fictitious page that must be initialized. 741 */ 742 vm_page_t 743 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr) 744 { 745 vm_page_t m; 746 747 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO); 748 vm_page_initfake(m, paddr, memattr); 749 return (m); 750 } 751 752 void 753 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 754 { 755 756 if ((m->flags & PG_FICTITIOUS) != 0) { 757 /* 758 * The page's memattr might have changed since the 759 * previous initialization. Update the pmap to the 760 * new memattr. 761 */ 762 goto memattr; 763 } 764 m->phys_addr = paddr; 765 m->queue = PQ_NONE; 766 /* Fictitious pages don't use "segind". */ 767 m->flags = PG_FICTITIOUS; 768 /* Fictitious pages don't use "order" or "pool". */ 769 m->oflags = VPO_UNMANAGED; 770 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 771 m->wire_count = 1; 772 pmap_page_init(m); 773 memattr: 774 pmap_page_set_memattr(m, memattr); 775 } 776 777 /* 778 * vm_page_putfake: 779 * 780 * Release a fictitious page. 781 */ 782 void 783 vm_page_putfake(vm_page_t m) 784 { 785 786 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m)); 787 KASSERT((m->flags & PG_FICTITIOUS) != 0, 788 ("vm_page_putfake: bad page %p", m)); 789 uma_zfree(fakepg_zone, m); 790 } 791 792 /* 793 * vm_page_updatefake: 794 * 795 * Update the given fictitious page to the specified physical address and 796 * memory attribute. 797 */ 798 void 799 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 800 { 801 802 KASSERT((m->flags & PG_FICTITIOUS) != 0, 803 ("vm_page_updatefake: bad page %p", m)); 804 m->phys_addr = paddr; 805 pmap_page_set_memattr(m, memattr); 806 } 807 808 /* 809 * vm_page_free: 810 * 811 * Free a page. 812 */ 813 void 814 vm_page_free(vm_page_t m) 815 { 816 817 m->flags &= ~PG_ZERO; 818 vm_page_free_toq(m); 819 } 820 821 /* 822 * vm_page_free_zero: 823 * 824 * Free a page to the zerod-pages queue 825 */ 826 void 827 vm_page_free_zero(vm_page_t m) 828 { 829 830 m->flags |= PG_ZERO; 831 vm_page_free_toq(m); 832 } 833 834 /* 835 * Unbusy and handle the page queueing for a page from the VOP_GETPAGES() 836 * array which is not the request page. 837 */ 838 void 839 vm_page_readahead_finish(vm_page_t m) 840 { 841 842 if (m->valid != 0) { 843 /* 844 * Since the page is not the requested page, whether 845 * it should be activated or deactivated is not 846 * obvious. Empirical results have shown that 847 * deactivating the page is usually the best choice, 848 * unless the page is wanted by another thread. 849 */ 850 vm_page_lock(m); 851 if ((m->busy_lock & VPB_BIT_WAITERS) != 0) 852 vm_page_activate(m); 853 else 854 vm_page_deactivate(m); 855 vm_page_unlock(m); 856 vm_page_xunbusy(m); 857 } else { 858 /* 859 * Free the completely invalid page. Such page state 860 * occurs due to the short read operation which did 861 * not covered our page at all, or in case when a read 862 * error happens. 863 */ 864 vm_page_lock(m); 865 vm_page_free(m); 866 vm_page_unlock(m); 867 } 868 } 869 870 /* 871 * vm_page_sleep_if_busy: 872 * 873 * Sleep and release the page queues lock if the page is busied. 874 * Returns TRUE if the thread slept. 875 * 876 * The given page must be unlocked and object containing it must 877 * be locked. 878 */ 879 int 880 vm_page_sleep_if_busy(vm_page_t m, const char *msg) 881 { 882 vm_object_t obj; 883 884 vm_page_lock_assert(m, MA_NOTOWNED); 885 VM_OBJECT_ASSERT_WLOCKED(m->object); 886 887 if (vm_page_busied(m)) { 888 /* 889 * The page-specific object must be cached because page 890 * identity can change during the sleep, causing the 891 * re-lock of a different object. 892 * It is assumed that a reference to the object is already 893 * held by the callers. 894 */ 895 obj = m->object; 896 vm_page_lock(m); 897 VM_OBJECT_WUNLOCK(obj); 898 vm_page_busy_sleep(m, msg); 899 VM_OBJECT_WLOCK(obj); 900 return (TRUE); 901 } 902 return (FALSE); 903 } 904 905 /* 906 * vm_page_dirty_KBI: [ internal use only ] 907 * 908 * Set all bits in the page's dirty field. 909 * 910 * The object containing the specified page must be locked if the 911 * call is made from the machine-independent layer. 912 * 913 * See vm_page_clear_dirty_mask(). 914 * 915 * This function should only be called by vm_page_dirty(). 916 */ 917 void 918 vm_page_dirty_KBI(vm_page_t m) 919 { 920 921 /* These assertions refer to this operation by its public name. */ 922 KASSERT((m->flags & PG_CACHED) == 0, 923 ("vm_page_dirty: page in cache!")); 924 KASSERT(m->valid == VM_PAGE_BITS_ALL, 925 ("vm_page_dirty: page is invalid!")); 926 m->dirty = VM_PAGE_BITS_ALL; 927 } 928 929 /* 930 * vm_page_insert: [ internal use only ] 931 * 932 * Inserts the given mem entry into the object and object list. 933 * 934 * The object must be locked. 935 */ 936 int 937 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 938 { 939 vm_page_t mpred; 940 941 VM_OBJECT_ASSERT_WLOCKED(object); 942 mpred = vm_radix_lookup_le(&object->rtree, pindex); 943 return (vm_page_insert_after(m, object, pindex, mpred)); 944 } 945 946 /* 947 * vm_page_insert_after: 948 * 949 * Inserts the page "m" into the specified object at offset "pindex". 950 * 951 * The page "mpred" must immediately precede the offset "pindex" within 952 * the specified object. 953 * 954 * The object must be locked. 955 */ 956 static int 957 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, 958 vm_page_t mpred) 959 { 960 vm_pindex_t sidx; 961 vm_object_t sobj; 962 vm_page_t msucc; 963 964 VM_OBJECT_ASSERT_WLOCKED(object); 965 KASSERT(m->object == NULL, 966 ("vm_page_insert_after: page already inserted")); 967 if (mpred != NULL) { 968 KASSERT(mpred->object == object, 969 ("vm_page_insert_after: object doesn't contain mpred")); 970 KASSERT(mpred->pindex < pindex, 971 ("vm_page_insert_after: mpred doesn't precede pindex")); 972 msucc = TAILQ_NEXT(mpred, listq); 973 } else 974 msucc = TAILQ_FIRST(&object->memq); 975 if (msucc != NULL) 976 KASSERT(msucc->pindex > pindex, 977 ("vm_page_insert_after: msucc doesn't succeed pindex")); 978 979 /* 980 * Record the object/offset pair in this page 981 */ 982 sobj = m->object; 983 sidx = m->pindex; 984 m->object = object; 985 m->pindex = pindex; 986 987 /* 988 * Now link into the object's ordered list of backed pages. 989 */ 990 if (vm_radix_insert(&object->rtree, m)) { 991 m->object = sobj; 992 m->pindex = sidx; 993 return (1); 994 } 995 vm_page_insert_radixdone(m, object, mpred); 996 return (0); 997 } 998 999 /* 1000 * vm_page_insert_radixdone: 1001 * 1002 * Complete page "m" insertion into the specified object after the 1003 * radix trie hooking. 1004 * 1005 * The page "mpred" must precede the offset "m->pindex" within the 1006 * specified object. 1007 * 1008 * The object must be locked. 1009 */ 1010 static void 1011 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred) 1012 { 1013 1014 VM_OBJECT_ASSERT_WLOCKED(object); 1015 KASSERT(object != NULL && m->object == object, 1016 ("vm_page_insert_radixdone: page %p has inconsistent object", m)); 1017 if (mpred != NULL) { 1018 KASSERT(mpred->object == object, 1019 ("vm_page_insert_after: object doesn't contain mpred")); 1020 KASSERT(mpred->pindex < m->pindex, 1021 ("vm_page_insert_after: mpred doesn't precede pindex")); 1022 } 1023 1024 if (mpred != NULL) 1025 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq); 1026 else 1027 TAILQ_INSERT_HEAD(&object->memq, m, listq); 1028 1029 /* 1030 * Show that the object has one more resident page. 1031 */ 1032 object->resident_page_count++; 1033 1034 /* 1035 * Hold the vnode until the last page is released. 1036 */ 1037 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 1038 vhold(object->handle); 1039 1040 /* 1041 * Since we are inserting a new and possibly dirty page, 1042 * update the object's OBJ_MIGHTBEDIRTY flag. 1043 */ 1044 if (pmap_page_is_write_mapped(m)) 1045 vm_object_set_writeable_dirty(object); 1046 } 1047 1048 /* 1049 * vm_page_remove: 1050 * 1051 * Removes the given mem entry from the object/offset-page 1052 * table and the object page list, but do not invalidate/terminate 1053 * the backing store. 1054 * 1055 * The object must be locked. The page must be locked if it is managed. 1056 */ 1057 void 1058 vm_page_remove(vm_page_t m) 1059 { 1060 vm_object_t object; 1061 boolean_t lockacq; 1062 1063 if ((m->oflags & VPO_UNMANAGED) == 0) 1064 vm_page_lock_assert(m, MA_OWNED); 1065 if ((object = m->object) == NULL) 1066 return; 1067 VM_OBJECT_ASSERT_WLOCKED(object); 1068 if (vm_page_xbusied(m)) { 1069 lockacq = FALSE; 1070 if ((m->oflags & VPO_UNMANAGED) != 0 && 1071 !mtx_owned(vm_page_lockptr(m))) { 1072 lockacq = TRUE; 1073 vm_page_lock(m); 1074 } 1075 vm_page_flash(m); 1076 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); 1077 if (lockacq) 1078 vm_page_unlock(m); 1079 } 1080 1081 /* 1082 * Now remove from the object's list of backed pages. 1083 */ 1084 vm_radix_remove(&object->rtree, m->pindex); 1085 TAILQ_REMOVE(&object->memq, m, listq); 1086 1087 /* 1088 * And show that the object has one fewer resident page. 1089 */ 1090 object->resident_page_count--; 1091 1092 /* 1093 * The vnode may now be recycled. 1094 */ 1095 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 1096 vdrop(object->handle); 1097 1098 m->object = NULL; 1099 } 1100 1101 /* 1102 * vm_page_lookup: 1103 * 1104 * Returns the page associated with the object/offset 1105 * pair specified; if none is found, NULL is returned. 1106 * 1107 * The object must be locked. 1108 */ 1109 vm_page_t 1110 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1111 { 1112 1113 VM_OBJECT_ASSERT_LOCKED(object); 1114 return (vm_radix_lookup(&object->rtree, pindex)); 1115 } 1116 1117 /* 1118 * vm_page_find_least: 1119 * 1120 * Returns the page associated with the object with least pindex 1121 * greater than or equal to the parameter pindex, or NULL. 1122 * 1123 * The object must be locked. 1124 */ 1125 vm_page_t 1126 vm_page_find_least(vm_object_t object, vm_pindex_t pindex) 1127 { 1128 vm_page_t m; 1129 1130 VM_OBJECT_ASSERT_LOCKED(object); 1131 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex) 1132 m = vm_radix_lookup_ge(&object->rtree, pindex); 1133 return (m); 1134 } 1135 1136 /* 1137 * Returns the given page's successor (by pindex) within the object if it is 1138 * resident; if none is found, NULL is returned. 1139 * 1140 * The object must be locked. 1141 */ 1142 vm_page_t 1143 vm_page_next(vm_page_t m) 1144 { 1145 vm_page_t next; 1146 1147 VM_OBJECT_ASSERT_WLOCKED(m->object); 1148 if ((next = TAILQ_NEXT(m, listq)) != NULL && 1149 next->pindex != m->pindex + 1) 1150 next = NULL; 1151 return (next); 1152 } 1153 1154 /* 1155 * Returns the given page's predecessor (by pindex) within the object if it is 1156 * resident; if none is found, NULL is returned. 1157 * 1158 * The object must be locked. 1159 */ 1160 vm_page_t 1161 vm_page_prev(vm_page_t m) 1162 { 1163 vm_page_t prev; 1164 1165 VM_OBJECT_ASSERT_WLOCKED(m->object); 1166 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL && 1167 prev->pindex != m->pindex - 1) 1168 prev = NULL; 1169 return (prev); 1170 } 1171 1172 /* 1173 * Uses the page mnew as a replacement for an existing page at index 1174 * pindex which must be already present in the object. 1175 * 1176 * The existing page must not be on a paging queue. 1177 */ 1178 vm_page_t 1179 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex) 1180 { 1181 vm_page_t mold, mpred; 1182 1183 VM_OBJECT_ASSERT_WLOCKED(object); 1184 1185 /* 1186 * This function mostly follows vm_page_insert() and 1187 * vm_page_remove() without the radix, object count and vnode 1188 * dance. Double check such functions for more comments. 1189 */ 1190 mpred = vm_radix_lookup(&object->rtree, pindex); 1191 KASSERT(mpred != NULL, 1192 ("vm_page_replace: replacing page not present with pindex")); 1193 mpred = TAILQ_PREV(mpred, respgs, listq); 1194 if (mpred != NULL) 1195 KASSERT(mpred->pindex < pindex, 1196 ("vm_page_insert_after: mpred doesn't precede pindex")); 1197 1198 mnew->object = object; 1199 mnew->pindex = pindex; 1200 mold = vm_radix_replace(&object->rtree, mnew); 1201 KASSERT(mold->queue == PQ_NONE, 1202 ("vm_page_replace: mold is on a paging queue")); 1203 1204 /* Detach the old page from the resident tailq. */ 1205 TAILQ_REMOVE(&object->memq, mold, listq); 1206 1207 mold->object = NULL; 1208 vm_page_xunbusy(mold); 1209 1210 /* Insert the new page in the resident tailq. */ 1211 if (mpred != NULL) 1212 TAILQ_INSERT_AFTER(&object->memq, mpred, mnew, listq); 1213 else 1214 TAILQ_INSERT_HEAD(&object->memq, mnew, listq); 1215 if (pmap_page_is_write_mapped(mnew)) 1216 vm_object_set_writeable_dirty(object); 1217 return (mold); 1218 } 1219 1220 /* 1221 * vm_page_rename: 1222 * 1223 * Move the given memory entry from its 1224 * current object to the specified target object/offset. 1225 * 1226 * Note: swap associated with the page must be invalidated by the move. We 1227 * have to do this for several reasons: (1) we aren't freeing the 1228 * page, (2) we are dirtying the page, (3) the VM system is probably 1229 * moving the page from object A to B, and will then later move 1230 * the backing store from A to B and we can't have a conflict. 1231 * 1232 * Note: we *always* dirty the page. It is necessary both for the 1233 * fact that we moved it, and because we may be invalidating 1234 * swap. If the page is on the cache, we have to deactivate it 1235 * or vm_page_dirty() will panic. Dirty pages are not allowed 1236 * on the cache. 1237 * 1238 * The objects must be locked. 1239 */ 1240 int 1241 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1242 { 1243 vm_page_t mpred; 1244 vm_pindex_t opidx; 1245 1246 VM_OBJECT_ASSERT_WLOCKED(new_object); 1247 1248 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex); 1249 KASSERT(mpred == NULL || mpred->pindex != new_pindex, 1250 ("vm_page_rename: pindex already renamed")); 1251 1252 /* 1253 * Create a custom version of vm_page_insert() which does not depend 1254 * by m_prev and can cheat on the implementation aspects of the 1255 * function. 1256 */ 1257 opidx = m->pindex; 1258 m->pindex = new_pindex; 1259 if (vm_radix_insert(&new_object->rtree, m)) { 1260 m->pindex = opidx; 1261 return (1); 1262 } 1263 1264 /* 1265 * The operation cannot fail anymore. The removal must happen before 1266 * the listq iterator is tainted. 1267 */ 1268 m->pindex = opidx; 1269 vm_page_lock(m); 1270 vm_page_remove(m); 1271 1272 /* Return back to the new pindex to complete vm_page_insert(). */ 1273 m->pindex = new_pindex; 1274 m->object = new_object; 1275 vm_page_unlock(m); 1276 vm_page_insert_radixdone(m, new_object, mpred); 1277 vm_page_dirty(m); 1278 return (0); 1279 } 1280 1281 /* 1282 * Convert all of the given object's cached pages that have a 1283 * pindex within the given range into free pages. If the value 1284 * zero is given for "end", then the range's upper bound is 1285 * infinity. If the given object is backed by a vnode and it 1286 * transitions from having one or more cached pages to none, the 1287 * vnode's hold count is reduced. 1288 */ 1289 void 1290 vm_page_cache_free(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 1291 { 1292 vm_page_t m; 1293 boolean_t empty; 1294 1295 mtx_lock(&vm_page_queue_free_mtx); 1296 if (__predict_false(vm_radix_is_empty(&object->cache))) { 1297 mtx_unlock(&vm_page_queue_free_mtx); 1298 return; 1299 } 1300 while ((m = vm_radix_lookup_ge(&object->cache, start)) != NULL) { 1301 if (end != 0 && m->pindex >= end) 1302 break; 1303 vm_radix_remove(&object->cache, m->pindex); 1304 vm_page_cache_turn_free(m); 1305 } 1306 empty = vm_radix_is_empty(&object->cache); 1307 mtx_unlock(&vm_page_queue_free_mtx); 1308 if (object->type == OBJT_VNODE && empty) 1309 vdrop(object->handle); 1310 } 1311 1312 /* 1313 * Returns the cached page that is associated with the given 1314 * object and offset. If, however, none exists, returns NULL. 1315 * 1316 * The free page queue must be locked. 1317 */ 1318 static inline vm_page_t 1319 vm_page_cache_lookup(vm_object_t object, vm_pindex_t pindex) 1320 { 1321 1322 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1323 return (vm_radix_lookup(&object->cache, pindex)); 1324 } 1325 1326 /* 1327 * Remove the given cached page from its containing object's 1328 * collection of cached pages. 1329 * 1330 * The free page queue must be locked. 1331 */ 1332 static void 1333 vm_page_cache_remove(vm_page_t m) 1334 { 1335 1336 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1337 KASSERT((m->flags & PG_CACHED) != 0, 1338 ("vm_page_cache_remove: page %p is not cached", m)); 1339 vm_radix_remove(&m->object->cache, m->pindex); 1340 m->object = NULL; 1341 vm_cnt.v_cache_count--; 1342 } 1343 1344 /* 1345 * Transfer all of the cached pages with offset greater than or 1346 * equal to 'offidxstart' from the original object's cache to the 1347 * new object's cache. However, any cached pages with offset 1348 * greater than or equal to the new object's size are kept in the 1349 * original object. Initially, the new object's cache must be 1350 * empty. Offset 'offidxstart' in the original object must 1351 * correspond to offset zero in the new object. 1352 * 1353 * The new object must be locked. 1354 */ 1355 void 1356 vm_page_cache_transfer(vm_object_t orig_object, vm_pindex_t offidxstart, 1357 vm_object_t new_object) 1358 { 1359 vm_page_t m; 1360 1361 /* 1362 * Insertion into an object's collection of cached pages 1363 * requires the object to be locked. In contrast, removal does 1364 * not. 1365 */ 1366 VM_OBJECT_ASSERT_WLOCKED(new_object); 1367 KASSERT(vm_radix_is_empty(&new_object->cache), 1368 ("vm_page_cache_transfer: object %p has cached pages", 1369 new_object)); 1370 mtx_lock(&vm_page_queue_free_mtx); 1371 while ((m = vm_radix_lookup_ge(&orig_object->cache, 1372 offidxstart)) != NULL) { 1373 /* 1374 * Transfer all of the pages with offset greater than or 1375 * equal to 'offidxstart' from the original object's 1376 * cache to the new object's cache. 1377 */ 1378 if ((m->pindex - offidxstart) >= new_object->size) 1379 break; 1380 vm_radix_remove(&orig_object->cache, m->pindex); 1381 /* Update the page's object and offset. */ 1382 m->object = new_object; 1383 m->pindex -= offidxstart; 1384 if (vm_radix_insert(&new_object->cache, m)) 1385 vm_page_cache_turn_free(m); 1386 } 1387 mtx_unlock(&vm_page_queue_free_mtx); 1388 } 1389 1390 /* 1391 * Returns TRUE if a cached page is associated with the given object and 1392 * offset, and FALSE otherwise. 1393 * 1394 * The object must be locked. 1395 */ 1396 boolean_t 1397 vm_page_is_cached(vm_object_t object, vm_pindex_t pindex) 1398 { 1399 vm_page_t m; 1400 1401 /* 1402 * Insertion into an object's collection of cached pages requires the 1403 * object to be locked. Therefore, if the object is locked and the 1404 * object's collection is empty, there is no need to acquire the free 1405 * page queues lock in order to prove that the specified page doesn't 1406 * exist. 1407 */ 1408 VM_OBJECT_ASSERT_WLOCKED(object); 1409 if (__predict_true(vm_object_cache_is_empty(object))) 1410 return (FALSE); 1411 mtx_lock(&vm_page_queue_free_mtx); 1412 m = vm_page_cache_lookup(object, pindex); 1413 mtx_unlock(&vm_page_queue_free_mtx); 1414 return (m != NULL); 1415 } 1416 1417 /* 1418 * vm_page_alloc: 1419 * 1420 * Allocate and return a page that is associated with the specified 1421 * object and offset pair. By default, this page is exclusive busied. 1422 * 1423 * The caller must always specify an allocation class. 1424 * 1425 * allocation classes: 1426 * VM_ALLOC_NORMAL normal process request 1427 * VM_ALLOC_SYSTEM system *really* needs a page 1428 * VM_ALLOC_INTERRUPT interrupt time request 1429 * 1430 * optional allocation flags: 1431 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1432 * intends to allocate 1433 * VM_ALLOC_IFCACHED return page only if it is cached 1434 * VM_ALLOC_IFNOTCACHED return NULL, do not reactivate if the page 1435 * is cached 1436 * VM_ALLOC_NOBUSY do not exclusive busy the page 1437 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1438 * VM_ALLOC_NOOBJ page is not associated with an object and 1439 * should not be exclusive busy 1440 * VM_ALLOC_SBUSY shared busy the allocated page 1441 * VM_ALLOC_WIRED wire the allocated page 1442 * VM_ALLOC_ZERO prefer a zeroed page 1443 * 1444 * This routine may not sleep. 1445 */ 1446 vm_page_t 1447 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 1448 { 1449 struct vnode *vp = NULL; 1450 vm_object_t m_object; 1451 vm_page_t m, mpred; 1452 int flags, req_class; 1453 1454 mpred = 0; /* XXX: pacify gcc */ 1455 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 1456 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 1457 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 1458 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 1459 ("vm_page_alloc: inconsistent object(%p)/req(%x)", (void *)object, 1460 req)); 1461 if (object != NULL) 1462 VM_OBJECT_ASSERT_WLOCKED(object); 1463 1464 req_class = req & VM_ALLOC_CLASS_MASK; 1465 1466 /* 1467 * The page daemon is allowed to dig deeper into the free page list. 1468 */ 1469 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1470 req_class = VM_ALLOC_SYSTEM; 1471 1472 if (object != NULL) { 1473 mpred = vm_radix_lookup_le(&object->rtree, pindex); 1474 KASSERT(mpred == NULL || mpred->pindex != pindex, 1475 ("vm_page_alloc: pindex already allocated")); 1476 } 1477 1478 /* 1479 * The page allocation request can came from consumers which already 1480 * hold the free page queue mutex, like vm_page_insert() in 1481 * vm_page_cache(). 1482 */ 1483 mtx_lock_flags(&vm_page_queue_free_mtx, MTX_RECURSE); 1484 if (vm_cnt.v_free_count + vm_cnt.v_cache_count > vm_cnt.v_free_reserved || 1485 (req_class == VM_ALLOC_SYSTEM && 1486 vm_cnt.v_free_count + vm_cnt.v_cache_count > vm_cnt.v_interrupt_free_min) || 1487 (req_class == VM_ALLOC_INTERRUPT && 1488 vm_cnt.v_free_count + vm_cnt.v_cache_count > 0)) { 1489 /* 1490 * Allocate from the free queue if the number of free pages 1491 * exceeds the minimum for the request class. 1492 */ 1493 if (object != NULL && 1494 (m = vm_page_cache_lookup(object, pindex)) != NULL) { 1495 if ((req & VM_ALLOC_IFNOTCACHED) != 0) { 1496 mtx_unlock(&vm_page_queue_free_mtx); 1497 return (NULL); 1498 } 1499 if (vm_phys_unfree_page(m)) 1500 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0); 1501 #if VM_NRESERVLEVEL > 0 1502 else if (!vm_reserv_reactivate_page(m)) 1503 #else 1504 else 1505 #endif 1506 panic("vm_page_alloc: cache page %p is missing" 1507 " from the free queue", m); 1508 } else if ((req & VM_ALLOC_IFCACHED) != 0) { 1509 mtx_unlock(&vm_page_queue_free_mtx); 1510 return (NULL); 1511 #if VM_NRESERVLEVEL > 0 1512 } else if (object == NULL || (object->flags & (OBJ_COLORED | 1513 OBJ_FICTITIOUS)) != OBJ_COLORED || (m = 1514 vm_reserv_alloc_page(object, pindex, mpred)) == NULL) { 1515 #else 1516 } else { 1517 #endif 1518 m = vm_phys_alloc_pages(object != NULL ? 1519 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0); 1520 #if VM_NRESERVLEVEL > 0 1521 if (m == NULL && vm_reserv_reclaim_inactive()) { 1522 m = vm_phys_alloc_pages(object != NULL ? 1523 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 1524 0); 1525 } 1526 #endif 1527 } 1528 } else { 1529 /* 1530 * Not allocatable, give up. 1531 */ 1532 mtx_unlock(&vm_page_queue_free_mtx); 1533 atomic_add_int(&vm_pageout_deficit, 1534 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 1535 pagedaemon_wakeup(); 1536 return (NULL); 1537 } 1538 1539 /* 1540 * At this point we had better have found a good page. 1541 */ 1542 KASSERT(m != NULL, ("vm_page_alloc: missing page")); 1543 KASSERT(m->queue == PQ_NONE, 1544 ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue)); 1545 KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m)); 1546 KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m)); 1547 KASSERT(!vm_page_sbusied(m), 1548 ("vm_page_alloc: page %p is busy", m)); 1549 KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m)); 1550 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 1551 ("vm_page_alloc: page %p has unexpected memattr %d", m, 1552 pmap_page_get_memattr(m))); 1553 if ((m->flags & PG_CACHED) != 0) { 1554 KASSERT((m->flags & PG_ZERO) == 0, 1555 ("vm_page_alloc: cached page %p is PG_ZERO", m)); 1556 KASSERT(m->valid != 0, 1557 ("vm_page_alloc: cached page %p is invalid", m)); 1558 if (m->object == object && m->pindex == pindex) 1559 vm_cnt.v_reactivated++; 1560 else 1561 m->valid = 0; 1562 m_object = m->object; 1563 vm_page_cache_remove(m); 1564 if (m_object->type == OBJT_VNODE && 1565 vm_object_cache_is_empty(m_object)) 1566 vp = m_object->handle; 1567 } else { 1568 KASSERT(m->valid == 0, 1569 ("vm_page_alloc: free page %p is valid", m)); 1570 vm_phys_freecnt_adj(m, -1); 1571 if ((m->flags & PG_ZERO) != 0) 1572 vm_page_zero_count--; 1573 } 1574 mtx_unlock(&vm_page_queue_free_mtx); 1575 1576 /* 1577 * Initialize the page. Only the PG_ZERO flag is inherited. 1578 */ 1579 flags = 0; 1580 if ((req & VM_ALLOC_ZERO) != 0) 1581 flags = PG_ZERO; 1582 flags &= m->flags; 1583 if ((req & VM_ALLOC_NODUMP) != 0) 1584 flags |= PG_NODUMP; 1585 m->flags = flags; 1586 m->aflags = 0; 1587 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? 1588 VPO_UNMANAGED : 0; 1589 m->busy_lock = VPB_UNBUSIED; 1590 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) 1591 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 1592 if ((req & VM_ALLOC_SBUSY) != 0) 1593 m->busy_lock = VPB_SHARERS_WORD(1); 1594 if (req & VM_ALLOC_WIRED) { 1595 /* 1596 * The page lock is not required for wiring a page until that 1597 * page is inserted into the object. 1598 */ 1599 atomic_add_int(&vm_cnt.v_wire_count, 1); 1600 m->wire_count = 1; 1601 } 1602 m->act_count = 0; 1603 1604 if (object != NULL) { 1605 if (vm_page_insert_after(m, object, pindex, mpred)) { 1606 /* See the comment below about hold count. */ 1607 if (vp != NULL) 1608 vdrop(vp); 1609 pagedaemon_wakeup(); 1610 if (req & VM_ALLOC_WIRED) { 1611 atomic_subtract_int(&vm_cnt.v_wire_count, 1); 1612 m->wire_count = 0; 1613 } 1614 m->object = NULL; 1615 vm_page_free(m); 1616 return (NULL); 1617 } 1618 1619 /* Ignore device objects; the pager sets "memattr" for them. */ 1620 if (object->memattr != VM_MEMATTR_DEFAULT && 1621 (object->flags & OBJ_FICTITIOUS) == 0) 1622 pmap_page_set_memattr(m, object->memattr); 1623 } else 1624 m->pindex = pindex; 1625 1626 /* 1627 * The following call to vdrop() must come after the above call 1628 * to vm_page_insert() in case both affect the same object and 1629 * vnode. Otherwise, the affected vnode's hold count could 1630 * temporarily become zero. 1631 */ 1632 if (vp != NULL) 1633 vdrop(vp); 1634 1635 /* 1636 * Don't wakeup too often - wakeup the pageout daemon when 1637 * we would be nearly out of memory. 1638 */ 1639 if (vm_paging_needed()) 1640 pagedaemon_wakeup(); 1641 1642 return (m); 1643 } 1644 1645 static void 1646 vm_page_alloc_contig_vdrop(struct spglist *lst) 1647 { 1648 1649 while (!SLIST_EMPTY(lst)) { 1650 vdrop((struct vnode *)SLIST_FIRST(lst)-> plinks.s.pv); 1651 SLIST_REMOVE_HEAD(lst, plinks.s.ss); 1652 } 1653 } 1654 1655 /* 1656 * vm_page_alloc_contig: 1657 * 1658 * Allocate a contiguous set of physical pages of the given size "npages" 1659 * from the free lists. All of the physical pages must be at or above 1660 * the given physical address "low" and below the given physical address 1661 * "high". The given value "alignment" determines the alignment of the 1662 * first physical page in the set. If the given value "boundary" is 1663 * non-zero, then the set of physical pages cannot cross any physical 1664 * address boundary that is a multiple of that value. Both "alignment" 1665 * and "boundary" must be a power of two. 1666 * 1667 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, 1668 * then the memory attribute setting for the physical pages is configured 1669 * to the object's memory attribute setting. Otherwise, the memory 1670 * attribute setting for the physical pages is configured to "memattr", 1671 * overriding the object's memory attribute setting. However, if the 1672 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the 1673 * memory attribute setting for the physical pages cannot be configured 1674 * to VM_MEMATTR_DEFAULT. 1675 * 1676 * The caller must always specify an allocation class. 1677 * 1678 * allocation classes: 1679 * VM_ALLOC_NORMAL normal process request 1680 * VM_ALLOC_SYSTEM system *really* needs a page 1681 * VM_ALLOC_INTERRUPT interrupt time request 1682 * 1683 * optional allocation flags: 1684 * VM_ALLOC_NOBUSY do not exclusive busy the page 1685 * VM_ALLOC_NOOBJ page is not associated with an object and 1686 * should not be exclusive busy 1687 * VM_ALLOC_SBUSY shared busy the allocated page 1688 * VM_ALLOC_WIRED wire the allocated page 1689 * VM_ALLOC_ZERO prefer a zeroed page 1690 * 1691 * This routine may not sleep. 1692 */ 1693 vm_page_t 1694 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, 1695 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 1696 vm_paddr_t boundary, vm_memattr_t memattr) 1697 { 1698 struct vnode *drop; 1699 struct spglist deferred_vdrop_list; 1700 vm_page_t m, m_tmp, m_ret; 1701 u_int flags; 1702 int req_class; 1703 1704 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 1705 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 1706 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 1707 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 1708 ("vm_page_alloc: inconsistent object(%p)/req(%x)", (void *)object, 1709 req)); 1710 if (object != NULL) { 1711 VM_OBJECT_ASSERT_WLOCKED(object); 1712 KASSERT(object->type == OBJT_PHYS, 1713 ("vm_page_alloc_contig: object %p isn't OBJT_PHYS", 1714 object)); 1715 } 1716 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 1717 req_class = req & VM_ALLOC_CLASS_MASK; 1718 1719 /* 1720 * The page daemon is allowed to dig deeper into the free page list. 1721 */ 1722 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1723 req_class = VM_ALLOC_SYSTEM; 1724 1725 SLIST_INIT(&deferred_vdrop_list); 1726 mtx_lock(&vm_page_queue_free_mtx); 1727 if (vm_cnt.v_free_count + vm_cnt.v_cache_count >= npages + 1728 vm_cnt.v_free_reserved || (req_class == VM_ALLOC_SYSTEM && 1729 vm_cnt.v_free_count + vm_cnt.v_cache_count >= npages + 1730 vm_cnt.v_interrupt_free_min) || (req_class == VM_ALLOC_INTERRUPT && 1731 vm_cnt.v_free_count + vm_cnt.v_cache_count >= npages)) { 1732 #if VM_NRESERVLEVEL > 0 1733 retry: 1734 if (object == NULL || (object->flags & OBJ_COLORED) == 0 || 1735 (m_ret = vm_reserv_alloc_contig(object, pindex, npages, 1736 low, high, alignment, boundary)) == NULL) 1737 #endif 1738 m_ret = vm_phys_alloc_contig(npages, low, high, 1739 alignment, boundary); 1740 } else { 1741 mtx_unlock(&vm_page_queue_free_mtx); 1742 atomic_add_int(&vm_pageout_deficit, npages); 1743 pagedaemon_wakeup(); 1744 return (NULL); 1745 } 1746 if (m_ret != NULL) 1747 for (m = m_ret; m < &m_ret[npages]; m++) { 1748 drop = vm_page_alloc_init(m); 1749 if (drop != NULL) { 1750 /* 1751 * Enqueue the vnode for deferred vdrop(). 1752 */ 1753 m->plinks.s.pv = drop; 1754 SLIST_INSERT_HEAD(&deferred_vdrop_list, m, 1755 plinks.s.ss); 1756 } 1757 } 1758 else { 1759 #if VM_NRESERVLEVEL > 0 1760 if (vm_reserv_reclaim_contig(npages, low, high, alignment, 1761 boundary)) 1762 goto retry; 1763 #endif 1764 } 1765 mtx_unlock(&vm_page_queue_free_mtx); 1766 if (m_ret == NULL) 1767 return (NULL); 1768 1769 /* 1770 * Initialize the pages. Only the PG_ZERO flag is inherited. 1771 */ 1772 flags = 0; 1773 if ((req & VM_ALLOC_ZERO) != 0) 1774 flags = PG_ZERO; 1775 if ((req & VM_ALLOC_NODUMP) != 0) 1776 flags |= PG_NODUMP; 1777 if ((req & VM_ALLOC_WIRED) != 0) 1778 atomic_add_int(&vm_cnt.v_wire_count, npages); 1779 if (object != NULL) { 1780 if (object->memattr != VM_MEMATTR_DEFAULT && 1781 memattr == VM_MEMATTR_DEFAULT) 1782 memattr = object->memattr; 1783 } 1784 for (m = m_ret; m < &m_ret[npages]; m++) { 1785 m->aflags = 0; 1786 m->flags = (m->flags | PG_NODUMP) & flags; 1787 m->busy_lock = VPB_UNBUSIED; 1788 if (object != NULL) { 1789 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 1790 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 1791 if ((req & VM_ALLOC_SBUSY) != 0) 1792 m->busy_lock = VPB_SHARERS_WORD(1); 1793 } 1794 if ((req & VM_ALLOC_WIRED) != 0) 1795 m->wire_count = 1; 1796 /* Unmanaged pages don't use "act_count". */ 1797 m->oflags = VPO_UNMANAGED; 1798 if (object != NULL) { 1799 if (vm_page_insert(m, object, pindex)) { 1800 vm_page_alloc_contig_vdrop( 1801 &deferred_vdrop_list); 1802 if (vm_paging_needed()) 1803 pagedaemon_wakeup(); 1804 if ((req & VM_ALLOC_WIRED) != 0) 1805 atomic_subtract_int(&vm_cnt.v_wire_count, 1806 npages); 1807 for (m_tmp = m, m = m_ret; 1808 m < &m_ret[npages]; m++) { 1809 if ((req & VM_ALLOC_WIRED) != 0) 1810 m->wire_count = 0; 1811 if (m >= m_tmp) 1812 m->object = NULL; 1813 vm_page_free(m); 1814 } 1815 return (NULL); 1816 } 1817 } else 1818 m->pindex = pindex; 1819 if (memattr != VM_MEMATTR_DEFAULT) 1820 pmap_page_set_memattr(m, memattr); 1821 pindex++; 1822 } 1823 vm_page_alloc_contig_vdrop(&deferred_vdrop_list); 1824 if (vm_paging_needed()) 1825 pagedaemon_wakeup(); 1826 return (m_ret); 1827 } 1828 1829 /* 1830 * Initialize a page that has been freshly dequeued from a freelist. 1831 * The caller has to drop the vnode returned, if it is not NULL. 1832 * 1833 * This function may only be used to initialize unmanaged pages. 1834 * 1835 * To be called with vm_page_queue_free_mtx held. 1836 */ 1837 static struct vnode * 1838 vm_page_alloc_init(vm_page_t m) 1839 { 1840 struct vnode *drop; 1841 vm_object_t m_object; 1842 1843 KASSERT(m->queue == PQ_NONE, 1844 ("vm_page_alloc_init: page %p has unexpected queue %d", 1845 m, m->queue)); 1846 KASSERT(m->wire_count == 0, 1847 ("vm_page_alloc_init: page %p is wired", m)); 1848 KASSERT(m->hold_count == 0, 1849 ("vm_page_alloc_init: page %p is held", m)); 1850 KASSERT(!vm_page_sbusied(m), 1851 ("vm_page_alloc_init: page %p is busy", m)); 1852 KASSERT(m->dirty == 0, 1853 ("vm_page_alloc_init: page %p is dirty", m)); 1854 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 1855 ("vm_page_alloc_init: page %p has unexpected memattr %d", 1856 m, pmap_page_get_memattr(m))); 1857 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1858 drop = NULL; 1859 if ((m->flags & PG_CACHED) != 0) { 1860 KASSERT((m->flags & PG_ZERO) == 0, 1861 ("vm_page_alloc_init: cached page %p is PG_ZERO", m)); 1862 m->valid = 0; 1863 m_object = m->object; 1864 vm_page_cache_remove(m); 1865 if (m_object->type == OBJT_VNODE && 1866 vm_object_cache_is_empty(m_object)) 1867 drop = m_object->handle; 1868 } else { 1869 KASSERT(m->valid == 0, 1870 ("vm_page_alloc_init: free page %p is valid", m)); 1871 vm_phys_freecnt_adj(m, -1); 1872 if ((m->flags & PG_ZERO) != 0) 1873 vm_page_zero_count--; 1874 } 1875 return (drop); 1876 } 1877 1878 /* 1879 * vm_page_alloc_freelist: 1880 * 1881 * Allocate a physical page from the specified free page list. 1882 * 1883 * The caller must always specify an allocation class. 1884 * 1885 * allocation classes: 1886 * VM_ALLOC_NORMAL normal process request 1887 * VM_ALLOC_SYSTEM system *really* needs a page 1888 * VM_ALLOC_INTERRUPT interrupt time request 1889 * 1890 * optional allocation flags: 1891 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1892 * intends to allocate 1893 * VM_ALLOC_WIRED wire the allocated page 1894 * VM_ALLOC_ZERO prefer a zeroed page 1895 * 1896 * This routine may not sleep. 1897 */ 1898 vm_page_t 1899 vm_page_alloc_freelist(int flind, int req) 1900 { 1901 struct vnode *drop; 1902 vm_page_t m; 1903 u_int flags; 1904 int req_class; 1905 1906 req_class = req & VM_ALLOC_CLASS_MASK; 1907 1908 /* 1909 * The page daemon is allowed to dig deeper into the free page list. 1910 */ 1911 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1912 req_class = VM_ALLOC_SYSTEM; 1913 1914 /* 1915 * Do not allocate reserved pages unless the req has asked for it. 1916 */ 1917 mtx_lock_flags(&vm_page_queue_free_mtx, MTX_RECURSE); 1918 if (vm_cnt.v_free_count + vm_cnt.v_cache_count > vm_cnt.v_free_reserved || 1919 (req_class == VM_ALLOC_SYSTEM && 1920 vm_cnt.v_free_count + vm_cnt.v_cache_count > vm_cnt.v_interrupt_free_min) || 1921 (req_class == VM_ALLOC_INTERRUPT && 1922 vm_cnt.v_free_count + vm_cnt.v_cache_count > 0)) 1923 m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0); 1924 else { 1925 mtx_unlock(&vm_page_queue_free_mtx); 1926 atomic_add_int(&vm_pageout_deficit, 1927 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 1928 pagedaemon_wakeup(); 1929 return (NULL); 1930 } 1931 if (m == NULL) { 1932 mtx_unlock(&vm_page_queue_free_mtx); 1933 return (NULL); 1934 } 1935 drop = vm_page_alloc_init(m); 1936 mtx_unlock(&vm_page_queue_free_mtx); 1937 1938 /* 1939 * Initialize the page. Only the PG_ZERO flag is inherited. 1940 */ 1941 m->aflags = 0; 1942 flags = 0; 1943 if ((req & VM_ALLOC_ZERO) != 0) 1944 flags = PG_ZERO; 1945 m->flags &= flags; 1946 if ((req & VM_ALLOC_WIRED) != 0) { 1947 /* 1948 * The page lock is not required for wiring a page that does 1949 * not belong to an object. 1950 */ 1951 atomic_add_int(&vm_cnt.v_wire_count, 1); 1952 m->wire_count = 1; 1953 } 1954 /* Unmanaged pages don't use "act_count". */ 1955 m->oflags = VPO_UNMANAGED; 1956 if (drop != NULL) 1957 vdrop(drop); 1958 if (vm_paging_needed()) 1959 pagedaemon_wakeup(); 1960 return (m); 1961 } 1962 1963 /* 1964 * vm_wait: (also see VM_WAIT macro) 1965 * 1966 * Sleep until free pages are available for allocation. 1967 * - Called in various places before memory allocations. 1968 */ 1969 void 1970 vm_wait(void) 1971 { 1972 1973 mtx_lock(&vm_page_queue_free_mtx); 1974 if (curproc == pageproc) { 1975 vm_pageout_pages_needed = 1; 1976 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx, 1977 PDROP | PSWP, "VMWait", 0); 1978 } else { 1979 if (!vm_pages_needed) { 1980 vm_pages_needed = 1; 1981 wakeup(&vm_pages_needed); 1982 } 1983 msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM, 1984 "vmwait", 0); 1985 } 1986 } 1987 1988 /* 1989 * vm_waitpfault: (also see VM_WAITPFAULT macro) 1990 * 1991 * Sleep until free pages are available for allocation. 1992 * - Called only in vm_fault so that processes page faulting 1993 * can be easily tracked. 1994 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 1995 * processes will be able to grab memory first. Do not change 1996 * this balance without careful testing first. 1997 */ 1998 void 1999 vm_waitpfault(void) 2000 { 2001 2002 mtx_lock(&vm_page_queue_free_mtx); 2003 if (!vm_pages_needed) { 2004 vm_pages_needed = 1; 2005 wakeup(&vm_pages_needed); 2006 } 2007 msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER, 2008 "pfault", 0); 2009 } 2010 2011 struct vm_pagequeue * 2012 vm_page_pagequeue(vm_page_t m) 2013 { 2014 2015 return (&vm_phys_domain(m)->vmd_pagequeues[m->queue]); 2016 } 2017 2018 /* 2019 * vm_page_dequeue: 2020 * 2021 * Remove the given page from its current page queue. 2022 * 2023 * The page must be locked. 2024 */ 2025 void 2026 vm_page_dequeue(vm_page_t m) 2027 { 2028 struct vm_pagequeue *pq; 2029 2030 vm_page_assert_locked(m); 2031 KASSERT(m->queue < PQ_COUNT, ("vm_page_dequeue: page %p is not queued", 2032 m)); 2033 pq = vm_page_pagequeue(m); 2034 vm_pagequeue_lock(pq); 2035 m->queue = PQ_NONE; 2036 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2037 vm_pagequeue_cnt_dec(pq); 2038 vm_pagequeue_unlock(pq); 2039 } 2040 2041 /* 2042 * vm_page_dequeue_locked: 2043 * 2044 * Remove the given page from its current page queue. 2045 * 2046 * The page and page queue must be locked. 2047 */ 2048 void 2049 vm_page_dequeue_locked(vm_page_t m) 2050 { 2051 struct vm_pagequeue *pq; 2052 2053 vm_page_lock_assert(m, MA_OWNED); 2054 pq = vm_page_pagequeue(m); 2055 vm_pagequeue_assert_locked(pq); 2056 m->queue = PQ_NONE; 2057 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2058 vm_pagequeue_cnt_dec(pq); 2059 } 2060 2061 /* 2062 * vm_page_enqueue: 2063 * 2064 * Add the given page to the specified page queue. 2065 * 2066 * The page must be locked. 2067 */ 2068 static void 2069 vm_page_enqueue(uint8_t queue, vm_page_t m) 2070 { 2071 struct vm_pagequeue *pq; 2072 2073 vm_page_lock_assert(m, MA_OWNED); 2074 KASSERT(queue < PQ_COUNT, 2075 ("vm_page_enqueue: invalid queue %u request for page %p", 2076 queue, m)); 2077 pq = &vm_phys_domain(m)->vmd_pagequeues[queue]; 2078 vm_pagequeue_lock(pq); 2079 m->queue = queue; 2080 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2081 vm_pagequeue_cnt_inc(pq); 2082 vm_pagequeue_unlock(pq); 2083 } 2084 2085 /* 2086 * vm_page_requeue: 2087 * 2088 * Move the given page to the tail of its current page queue. 2089 * 2090 * The page must be locked. 2091 */ 2092 void 2093 vm_page_requeue(vm_page_t m) 2094 { 2095 struct vm_pagequeue *pq; 2096 2097 vm_page_lock_assert(m, MA_OWNED); 2098 KASSERT(m->queue != PQ_NONE, 2099 ("vm_page_requeue: page %p is not queued", m)); 2100 pq = vm_page_pagequeue(m); 2101 vm_pagequeue_lock(pq); 2102 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2103 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2104 vm_pagequeue_unlock(pq); 2105 } 2106 2107 /* 2108 * vm_page_requeue_locked: 2109 * 2110 * Move the given page to the tail of its current page queue. 2111 * 2112 * The page queue must be locked. 2113 */ 2114 void 2115 vm_page_requeue_locked(vm_page_t m) 2116 { 2117 struct vm_pagequeue *pq; 2118 2119 KASSERT(m->queue != PQ_NONE, 2120 ("vm_page_requeue_locked: page %p is not queued", m)); 2121 pq = vm_page_pagequeue(m); 2122 vm_pagequeue_assert_locked(pq); 2123 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2124 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2125 } 2126 2127 /* 2128 * vm_page_activate: 2129 * 2130 * Put the specified page on the active list (if appropriate). 2131 * Ensure that act_count is at least ACT_INIT but do not otherwise 2132 * mess with it. 2133 * 2134 * The page must be locked. 2135 */ 2136 void 2137 vm_page_activate(vm_page_t m) 2138 { 2139 int queue; 2140 2141 vm_page_lock_assert(m, MA_OWNED); 2142 if ((queue = m->queue) != PQ_ACTIVE) { 2143 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 2144 if (m->act_count < ACT_INIT) 2145 m->act_count = ACT_INIT; 2146 if (queue != PQ_NONE) 2147 vm_page_dequeue(m); 2148 vm_page_enqueue(PQ_ACTIVE, m); 2149 } else 2150 KASSERT(queue == PQ_NONE, 2151 ("vm_page_activate: wired page %p is queued", m)); 2152 } else { 2153 if (m->act_count < ACT_INIT) 2154 m->act_count = ACT_INIT; 2155 } 2156 } 2157 2158 /* 2159 * vm_page_free_wakeup: 2160 * 2161 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 2162 * routine is called when a page has been added to the cache or free 2163 * queues. 2164 * 2165 * The page queues must be locked. 2166 */ 2167 static inline void 2168 vm_page_free_wakeup(void) 2169 { 2170 2171 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 2172 /* 2173 * if pageout daemon needs pages, then tell it that there are 2174 * some free. 2175 */ 2176 if (vm_pageout_pages_needed && 2177 vm_cnt.v_cache_count + vm_cnt.v_free_count >= vm_cnt.v_pageout_free_min) { 2178 wakeup(&vm_pageout_pages_needed); 2179 vm_pageout_pages_needed = 0; 2180 } 2181 /* 2182 * wakeup processes that are waiting on memory if we hit a 2183 * high water mark. And wakeup scheduler process if we have 2184 * lots of memory. this process will swapin processes. 2185 */ 2186 if (vm_pages_needed && !vm_page_count_min()) { 2187 vm_pages_needed = 0; 2188 wakeup(&vm_cnt.v_free_count); 2189 } 2190 } 2191 2192 /* 2193 * Turn a cached page into a free page, by changing its attributes. 2194 * Keep the statistics up-to-date. 2195 * 2196 * The free page queue must be locked. 2197 */ 2198 static void 2199 vm_page_cache_turn_free(vm_page_t m) 2200 { 2201 2202 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 2203 2204 m->object = NULL; 2205 m->valid = 0; 2206 KASSERT((m->flags & PG_CACHED) != 0, 2207 ("vm_page_cache_turn_free: page %p is not cached", m)); 2208 m->flags &= ~PG_CACHED; 2209 vm_cnt.v_cache_count--; 2210 vm_phys_freecnt_adj(m, 1); 2211 } 2212 2213 /* 2214 * vm_page_free_toq: 2215 * 2216 * Returns the given page to the free list, 2217 * disassociating it with any VM object. 2218 * 2219 * The object must be locked. The page must be locked if it is managed. 2220 */ 2221 void 2222 vm_page_free_toq(vm_page_t m) 2223 { 2224 2225 if ((m->oflags & VPO_UNMANAGED) == 0) { 2226 vm_page_lock_assert(m, MA_OWNED); 2227 KASSERT(!pmap_page_is_mapped(m), 2228 ("vm_page_free_toq: freeing mapped page %p", m)); 2229 } else 2230 KASSERT(m->queue == PQ_NONE, 2231 ("vm_page_free_toq: unmanaged page %p is queued", m)); 2232 PCPU_INC(cnt.v_tfree); 2233 2234 if (vm_page_sbusied(m)) 2235 panic("vm_page_free: freeing busy page %p", m); 2236 2237 /* 2238 * Unqueue, then remove page. Note that we cannot destroy 2239 * the page here because we do not want to call the pager's 2240 * callback routine until after we've put the page on the 2241 * appropriate free queue. 2242 */ 2243 vm_page_remque(m); 2244 vm_page_remove(m); 2245 2246 /* 2247 * If fictitious remove object association and 2248 * return, otherwise delay object association removal. 2249 */ 2250 if ((m->flags & PG_FICTITIOUS) != 0) { 2251 return; 2252 } 2253 2254 m->valid = 0; 2255 vm_page_undirty(m); 2256 2257 if (m->wire_count != 0) 2258 panic("vm_page_free: freeing wired page %p", m); 2259 if (m->hold_count != 0) { 2260 m->flags &= ~PG_ZERO; 2261 KASSERT((m->flags & PG_UNHOLDFREE) == 0, 2262 ("vm_page_free: freeing PG_UNHOLDFREE page %p", m)); 2263 m->flags |= PG_UNHOLDFREE; 2264 } else { 2265 /* 2266 * Restore the default memory attribute to the page. 2267 */ 2268 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 2269 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 2270 2271 /* 2272 * Insert the page into the physical memory allocator's 2273 * cache/free page queues. 2274 */ 2275 mtx_lock(&vm_page_queue_free_mtx); 2276 vm_phys_freecnt_adj(m, 1); 2277 #if VM_NRESERVLEVEL > 0 2278 if (!vm_reserv_free_page(m)) 2279 #else 2280 if (TRUE) 2281 #endif 2282 vm_phys_free_pages(m, 0); 2283 if ((m->flags & PG_ZERO) != 0) 2284 ++vm_page_zero_count; 2285 else 2286 vm_page_zero_idle_wakeup(); 2287 vm_page_free_wakeup(); 2288 mtx_unlock(&vm_page_queue_free_mtx); 2289 } 2290 } 2291 2292 /* 2293 * vm_page_wire: 2294 * 2295 * Mark this page as wired down by yet 2296 * another map, removing it from paging queues 2297 * as necessary. 2298 * 2299 * If the page is fictitious, then its wire count must remain one. 2300 * 2301 * The page must be locked. 2302 */ 2303 void 2304 vm_page_wire(vm_page_t m) 2305 { 2306 2307 /* 2308 * Only bump the wire statistics if the page is not already wired, 2309 * and only unqueue the page if it is on some queue (if it is unmanaged 2310 * it is already off the queues). 2311 */ 2312 vm_page_lock_assert(m, MA_OWNED); 2313 if ((m->flags & PG_FICTITIOUS) != 0) { 2314 KASSERT(m->wire_count == 1, 2315 ("vm_page_wire: fictitious page %p's wire count isn't one", 2316 m)); 2317 return; 2318 } 2319 if (m->wire_count == 0) { 2320 KASSERT((m->oflags & VPO_UNMANAGED) == 0 || 2321 m->queue == PQ_NONE, 2322 ("vm_page_wire: unmanaged page %p is queued", m)); 2323 vm_page_remque(m); 2324 atomic_add_int(&vm_cnt.v_wire_count, 1); 2325 } 2326 m->wire_count++; 2327 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 2328 } 2329 2330 /* 2331 * vm_page_unwire: 2332 * 2333 * Release one wiring of the specified page, potentially enabling it to be 2334 * paged again. If paging is enabled, then the value of the parameter 2335 * "queue" determines the queue to which the page is added. 2336 * 2337 * However, unless the page belongs to an object, it is not enqueued because 2338 * it cannot be paged out. 2339 * 2340 * If a page is fictitious, then its wire count must always be one. 2341 * 2342 * A managed page must be locked. 2343 */ 2344 void 2345 vm_page_unwire(vm_page_t m, uint8_t queue) 2346 { 2347 2348 KASSERT(queue < PQ_COUNT, 2349 ("vm_page_unwire: invalid queue %u request for page %p", 2350 queue, m)); 2351 if ((m->oflags & VPO_UNMANAGED) == 0) 2352 vm_page_lock_assert(m, MA_OWNED); 2353 if ((m->flags & PG_FICTITIOUS) != 0) { 2354 KASSERT(m->wire_count == 1, 2355 ("vm_page_unwire: fictitious page %p's wire count isn't one", m)); 2356 return; 2357 } 2358 if (m->wire_count > 0) { 2359 m->wire_count--; 2360 if (m->wire_count == 0) { 2361 atomic_subtract_int(&vm_cnt.v_wire_count, 1); 2362 if ((m->oflags & VPO_UNMANAGED) != 0 || 2363 m->object == NULL) 2364 return; 2365 if (queue == PQ_INACTIVE) 2366 m->flags &= ~PG_WINATCFLS; 2367 vm_page_enqueue(queue, m); 2368 } 2369 } else 2370 panic("vm_page_unwire: page %p's wire count is zero", m); 2371 } 2372 2373 /* 2374 * Move the specified page to the inactive queue. 2375 * 2376 * Many pages placed on the inactive queue should actually go 2377 * into the cache, but it is difficult to figure out which. What 2378 * we do instead, if the inactive target is well met, is to put 2379 * clean pages at the head of the inactive queue instead of the tail. 2380 * This will cause them to be moved to the cache more quickly and 2381 * if not actively re-referenced, reclaimed more quickly. If we just 2382 * stick these pages at the end of the inactive queue, heavy filesystem 2383 * meta-data accesses can cause an unnecessary paging load on memory bound 2384 * processes. This optimization causes one-time-use metadata to be 2385 * reused more quickly. 2386 * 2387 * Normally athead is 0 resulting in LRU operation. athead is set 2388 * to 1 if we want this page to be 'as if it were placed in the cache', 2389 * except without unmapping it from the process address space. 2390 * 2391 * The page must be locked. 2392 */ 2393 static inline void 2394 _vm_page_deactivate(vm_page_t m, int athead) 2395 { 2396 struct vm_pagequeue *pq; 2397 int queue; 2398 2399 vm_page_lock_assert(m, MA_OWNED); 2400 2401 /* 2402 * Ignore if already inactive. 2403 */ 2404 if ((queue = m->queue) == PQ_INACTIVE) 2405 return; 2406 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 2407 if (queue != PQ_NONE) 2408 vm_page_dequeue(m); 2409 m->flags &= ~PG_WINATCFLS; 2410 pq = &vm_phys_domain(m)->vmd_pagequeues[PQ_INACTIVE]; 2411 vm_pagequeue_lock(pq); 2412 m->queue = PQ_INACTIVE; 2413 if (athead) 2414 TAILQ_INSERT_HEAD(&pq->pq_pl, m, plinks.q); 2415 else 2416 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2417 vm_pagequeue_cnt_inc(pq); 2418 vm_pagequeue_unlock(pq); 2419 } 2420 } 2421 2422 /* 2423 * Move the specified page to the inactive queue. 2424 * 2425 * The page must be locked. 2426 */ 2427 void 2428 vm_page_deactivate(vm_page_t m) 2429 { 2430 2431 _vm_page_deactivate(m, 0); 2432 } 2433 2434 /* 2435 * vm_page_try_to_cache: 2436 * 2437 * Returns 0 on failure, 1 on success 2438 */ 2439 int 2440 vm_page_try_to_cache(vm_page_t m) 2441 { 2442 2443 vm_page_lock_assert(m, MA_OWNED); 2444 VM_OBJECT_ASSERT_WLOCKED(m->object); 2445 if (m->dirty || m->hold_count || m->wire_count || 2446 (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m)) 2447 return (0); 2448 pmap_remove_all(m); 2449 if (m->dirty) 2450 return (0); 2451 vm_page_cache(m); 2452 return (1); 2453 } 2454 2455 /* 2456 * vm_page_try_to_free() 2457 * 2458 * Attempt to free the page. If we cannot free it, we do nothing. 2459 * 1 is returned on success, 0 on failure. 2460 */ 2461 int 2462 vm_page_try_to_free(vm_page_t m) 2463 { 2464 2465 vm_page_lock_assert(m, MA_OWNED); 2466 if (m->object != NULL) 2467 VM_OBJECT_ASSERT_WLOCKED(m->object); 2468 if (m->dirty || m->hold_count || m->wire_count || 2469 (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m)) 2470 return (0); 2471 pmap_remove_all(m); 2472 if (m->dirty) 2473 return (0); 2474 vm_page_free(m); 2475 return (1); 2476 } 2477 2478 /* 2479 * vm_page_cache 2480 * 2481 * Put the specified page onto the page cache queue (if appropriate). 2482 * 2483 * The object and page must be locked. 2484 */ 2485 void 2486 vm_page_cache(vm_page_t m) 2487 { 2488 vm_object_t object; 2489 boolean_t cache_was_empty; 2490 2491 vm_page_lock_assert(m, MA_OWNED); 2492 object = m->object; 2493 VM_OBJECT_ASSERT_WLOCKED(object); 2494 if (vm_page_busied(m) || (m->oflags & VPO_UNMANAGED) || 2495 m->hold_count || m->wire_count) 2496 panic("vm_page_cache: attempting to cache busy page"); 2497 KASSERT(!pmap_page_is_mapped(m), 2498 ("vm_page_cache: page %p is mapped", m)); 2499 KASSERT(m->dirty == 0, ("vm_page_cache: page %p is dirty", m)); 2500 if (m->valid == 0 || object->type == OBJT_DEFAULT || 2501 (object->type == OBJT_SWAP && 2502 !vm_pager_has_page(object, m->pindex, NULL, NULL))) { 2503 /* 2504 * Hypothesis: A cache-eligible page belonging to a 2505 * default object or swap object but without a backing 2506 * store must be zero filled. 2507 */ 2508 vm_page_free(m); 2509 return; 2510 } 2511 KASSERT((m->flags & PG_CACHED) == 0, 2512 ("vm_page_cache: page %p is already cached", m)); 2513 2514 /* 2515 * Remove the page from the paging queues. 2516 */ 2517 vm_page_remque(m); 2518 2519 /* 2520 * Remove the page from the object's collection of resident 2521 * pages. 2522 */ 2523 vm_radix_remove(&object->rtree, m->pindex); 2524 TAILQ_REMOVE(&object->memq, m, listq); 2525 object->resident_page_count--; 2526 2527 /* 2528 * Restore the default memory attribute to the page. 2529 */ 2530 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 2531 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 2532 2533 /* 2534 * Insert the page into the object's collection of cached pages 2535 * and the physical memory allocator's cache/free page queues. 2536 */ 2537 m->flags &= ~PG_ZERO; 2538 mtx_lock(&vm_page_queue_free_mtx); 2539 cache_was_empty = vm_radix_is_empty(&object->cache); 2540 if (vm_radix_insert(&object->cache, m)) { 2541 mtx_unlock(&vm_page_queue_free_mtx); 2542 if (object->resident_page_count == 0) 2543 vdrop(object->handle); 2544 m->object = NULL; 2545 vm_page_free(m); 2546 return; 2547 } 2548 2549 /* 2550 * The above call to vm_radix_insert() could reclaim the one pre- 2551 * existing cached page from this object, resulting in a call to 2552 * vdrop(). 2553 */ 2554 if (!cache_was_empty) 2555 cache_was_empty = vm_radix_is_singleton(&object->cache); 2556 2557 m->flags |= PG_CACHED; 2558 vm_cnt.v_cache_count++; 2559 PCPU_INC(cnt.v_tcached); 2560 #if VM_NRESERVLEVEL > 0 2561 if (!vm_reserv_free_page(m)) { 2562 #else 2563 if (TRUE) { 2564 #endif 2565 vm_phys_set_pool(VM_FREEPOOL_CACHE, m, 0); 2566 vm_phys_free_pages(m, 0); 2567 } 2568 vm_page_free_wakeup(); 2569 mtx_unlock(&vm_page_queue_free_mtx); 2570 2571 /* 2572 * Increment the vnode's hold count if this is the object's only 2573 * cached page. Decrement the vnode's hold count if this was 2574 * the object's only resident page. 2575 */ 2576 if (object->type == OBJT_VNODE) { 2577 if (cache_was_empty && object->resident_page_count != 0) 2578 vhold(object->handle); 2579 else if (!cache_was_empty && object->resident_page_count == 0) 2580 vdrop(object->handle); 2581 } 2582 } 2583 2584 /* 2585 * vm_page_advise 2586 * 2587 * Cache, deactivate, or do nothing as appropriate. This routine 2588 * is used by madvise(). 2589 * 2590 * Generally speaking we want to move the page into the cache so 2591 * it gets reused quickly. However, this can result in a silly syndrome 2592 * due to the page recycling too quickly. Small objects will not be 2593 * fully cached. On the other hand, if we move the page to the inactive 2594 * queue we wind up with a problem whereby very large objects 2595 * unnecessarily blow away our inactive and cache queues. 2596 * 2597 * The solution is to move the pages based on a fixed weighting. We 2598 * either leave them alone, deactivate them, or move them to the cache, 2599 * where moving them to the cache has the highest weighting. 2600 * By forcing some pages into other queues we eventually force the 2601 * system to balance the queues, potentially recovering other unrelated 2602 * space from active. The idea is to not force this to happen too 2603 * often. 2604 * 2605 * The object and page must be locked. 2606 */ 2607 void 2608 vm_page_advise(vm_page_t m, int advice) 2609 { 2610 int dnw, head; 2611 2612 vm_page_assert_locked(m); 2613 VM_OBJECT_ASSERT_WLOCKED(m->object); 2614 if (advice == MADV_FREE) { 2615 /* 2616 * Mark the page clean. This will allow the page to be freed 2617 * up by the system. However, such pages are often reused 2618 * quickly by malloc() so we do not do anything that would 2619 * cause a page fault if we can help it. 2620 * 2621 * Specifically, we do not try to actually free the page now 2622 * nor do we try to put it in the cache (which would cause a 2623 * page fault on reuse). 2624 * 2625 * But we do make the page is freeable as we can without 2626 * actually taking the step of unmapping it. 2627 */ 2628 m->dirty = 0; 2629 m->act_count = 0; 2630 } else if (advice != MADV_DONTNEED) 2631 return; 2632 dnw = PCPU_GET(dnweight); 2633 PCPU_INC(dnweight); 2634 2635 /* 2636 * Occasionally leave the page alone. 2637 */ 2638 if ((dnw & 0x01F0) == 0 || m->queue == PQ_INACTIVE) { 2639 if (m->act_count >= ACT_INIT) 2640 --m->act_count; 2641 return; 2642 } 2643 2644 /* 2645 * Clear any references to the page. Otherwise, the page daemon will 2646 * immediately reactivate the page. 2647 */ 2648 vm_page_aflag_clear(m, PGA_REFERENCED); 2649 2650 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) 2651 vm_page_dirty(m); 2652 2653 if (m->dirty || (dnw & 0x0070) == 0) { 2654 /* 2655 * Deactivate the page 3 times out of 32. 2656 */ 2657 head = 0; 2658 } else { 2659 /* 2660 * Cache the page 28 times out of every 32. Note that 2661 * the page is deactivated instead of cached, but placed 2662 * at the head of the queue instead of the tail. 2663 */ 2664 head = 1; 2665 } 2666 _vm_page_deactivate(m, head); 2667 } 2668 2669 /* 2670 * Grab a page, waiting until we are waken up due to the page 2671 * changing state. We keep on waiting, if the page continues 2672 * to be in the object. If the page doesn't exist, first allocate it 2673 * and then conditionally zero it. 2674 * 2675 * This routine may sleep. 2676 * 2677 * The object must be locked on entry. The lock will, however, be released 2678 * and reacquired if the routine sleeps. 2679 */ 2680 vm_page_t 2681 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 2682 { 2683 vm_page_t m; 2684 int sleep; 2685 2686 VM_OBJECT_ASSERT_WLOCKED(object); 2687 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 2688 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 2689 ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 2690 retrylookup: 2691 if ((m = vm_page_lookup(object, pindex)) != NULL) { 2692 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ? 2693 vm_page_xbusied(m) : vm_page_busied(m); 2694 if (sleep) { 2695 /* 2696 * Reference the page before unlocking and 2697 * sleeping so that the page daemon is less 2698 * likely to reclaim it. 2699 */ 2700 vm_page_aflag_set(m, PGA_REFERENCED); 2701 vm_page_lock(m); 2702 VM_OBJECT_WUNLOCK(object); 2703 vm_page_busy_sleep(m, "pgrbwt"); 2704 VM_OBJECT_WLOCK(object); 2705 goto retrylookup; 2706 } else { 2707 if ((allocflags & VM_ALLOC_WIRED) != 0) { 2708 vm_page_lock(m); 2709 vm_page_wire(m); 2710 vm_page_unlock(m); 2711 } 2712 if ((allocflags & 2713 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 2714 vm_page_xbusy(m); 2715 if ((allocflags & VM_ALLOC_SBUSY) != 0) 2716 vm_page_sbusy(m); 2717 return (m); 2718 } 2719 } 2720 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_IGN_SBUSY); 2721 if (m == NULL) { 2722 VM_OBJECT_WUNLOCK(object); 2723 VM_WAIT; 2724 VM_OBJECT_WLOCK(object); 2725 goto retrylookup; 2726 } else if (m->valid != 0) 2727 return (m); 2728 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 2729 pmap_zero_page(m); 2730 return (m); 2731 } 2732 2733 /* 2734 * Mapping function for valid or dirty bits in a page. 2735 * 2736 * Inputs are required to range within a page. 2737 */ 2738 vm_page_bits_t 2739 vm_page_bits(int base, int size) 2740 { 2741 int first_bit; 2742 int last_bit; 2743 2744 KASSERT( 2745 base + size <= PAGE_SIZE, 2746 ("vm_page_bits: illegal base/size %d/%d", base, size) 2747 ); 2748 2749 if (size == 0) /* handle degenerate case */ 2750 return (0); 2751 2752 first_bit = base >> DEV_BSHIFT; 2753 last_bit = (base + size - 1) >> DEV_BSHIFT; 2754 2755 return (((vm_page_bits_t)2 << last_bit) - 2756 ((vm_page_bits_t)1 << first_bit)); 2757 } 2758 2759 /* 2760 * vm_page_set_valid_range: 2761 * 2762 * Sets portions of a page valid. The arguments are expected 2763 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2764 * of any partial chunks touched by the range. The invalid portion of 2765 * such chunks will be zeroed. 2766 * 2767 * (base + size) must be less then or equal to PAGE_SIZE. 2768 */ 2769 void 2770 vm_page_set_valid_range(vm_page_t m, int base, int size) 2771 { 2772 int endoff, frag; 2773 2774 VM_OBJECT_ASSERT_WLOCKED(m->object); 2775 if (size == 0) /* handle degenerate case */ 2776 return; 2777 2778 /* 2779 * If the base is not DEV_BSIZE aligned and the valid 2780 * bit is clear, we have to zero out a portion of the 2781 * first block. 2782 */ 2783 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 2784 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 2785 pmap_zero_page_area(m, frag, base - frag); 2786 2787 /* 2788 * If the ending offset is not DEV_BSIZE aligned and the 2789 * valid bit is clear, we have to zero out a portion of 2790 * the last block. 2791 */ 2792 endoff = base + size; 2793 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 2794 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 2795 pmap_zero_page_area(m, endoff, 2796 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 2797 2798 /* 2799 * Assert that no previously invalid block that is now being validated 2800 * is already dirty. 2801 */ 2802 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 2803 ("vm_page_set_valid_range: page %p is dirty", m)); 2804 2805 /* 2806 * Set valid bits inclusive of any overlap. 2807 */ 2808 m->valid |= vm_page_bits(base, size); 2809 } 2810 2811 /* 2812 * Clear the given bits from the specified page's dirty field. 2813 */ 2814 static __inline void 2815 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) 2816 { 2817 uintptr_t addr; 2818 #if PAGE_SIZE < 16384 2819 int shift; 2820 #endif 2821 2822 /* 2823 * If the object is locked and the page is neither exclusive busy nor 2824 * write mapped, then the page's dirty field cannot possibly be 2825 * set by a concurrent pmap operation. 2826 */ 2827 VM_OBJECT_ASSERT_WLOCKED(m->object); 2828 if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) 2829 m->dirty &= ~pagebits; 2830 else { 2831 /* 2832 * The pmap layer can call vm_page_dirty() without 2833 * holding a distinguished lock. The combination of 2834 * the object's lock and an atomic operation suffice 2835 * to guarantee consistency of the page dirty field. 2836 * 2837 * For PAGE_SIZE == 32768 case, compiler already 2838 * properly aligns the dirty field, so no forcible 2839 * alignment is needed. Only require existence of 2840 * atomic_clear_64 when page size is 32768. 2841 */ 2842 addr = (uintptr_t)&m->dirty; 2843 #if PAGE_SIZE == 32768 2844 atomic_clear_64((uint64_t *)addr, pagebits); 2845 #elif PAGE_SIZE == 16384 2846 atomic_clear_32((uint32_t *)addr, pagebits); 2847 #else /* PAGE_SIZE <= 8192 */ 2848 /* 2849 * Use a trick to perform a 32-bit atomic on the 2850 * containing aligned word, to not depend on the existence 2851 * of atomic_clear_{8, 16}. 2852 */ 2853 shift = addr & (sizeof(uint32_t) - 1); 2854 #if BYTE_ORDER == BIG_ENDIAN 2855 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY; 2856 #else 2857 shift *= NBBY; 2858 #endif 2859 addr &= ~(sizeof(uint32_t) - 1); 2860 atomic_clear_32((uint32_t *)addr, pagebits << shift); 2861 #endif /* PAGE_SIZE */ 2862 } 2863 } 2864 2865 /* 2866 * vm_page_set_validclean: 2867 * 2868 * Sets portions of a page valid and clean. The arguments are expected 2869 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2870 * of any partial chunks touched by the range. The invalid portion of 2871 * such chunks will be zero'd. 2872 * 2873 * (base + size) must be less then or equal to PAGE_SIZE. 2874 */ 2875 void 2876 vm_page_set_validclean(vm_page_t m, int base, int size) 2877 { 2878 vm_page_bits_t oldvalid, pagebits; 2879 int endoff, frag; 2880 2881 VM_OBJECT_ASSERT_WLOCKED(m->object); 2882 if (size == 0) /* handle degenerate case */ 2883 return; 2884 2885 /* 2886 * If the base is not DEV_BSIZE aligned and the valid 2887 * bit is clear, we have to zero out a portion of the 2888 * first block. 2889 */ 2890 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 2891 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) 2892 pmap_zero_page_area(m, frag, base - frag); 2893 2894 /* 2895 * If the ending offset is not DEV_BSIZE aligned and the 2896 * valid bit is clear, we have to zero out a portion of 2897 * the last block. 2898 */ 2899 endoff = base + size; 2900 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 2901 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) 2902 pmap_zero_page_area(m, endoff, 2903 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 2904 2905 /* 2906 * Set valid, clear dirty bits. If validating the entire 2907 * page we can safely clear the pmap modify bit. We also 2908 * use this opportunity to clear the VPO_NOSYNC flag. If a process 2909 * takes a write fault on a MAP_NOSYNC memory area the flag will 2910 * be set again. 2911 * 2912 * We set valid bits inclusive of any overlap, but we can only 2913 * clear dirty bits for DEV_BSIZE chunks that are fully within 2914 * the range. 2915 */ 2916 oldvalid = m->valid; 2917 pagebits = vm_page_bits(base, size); 2918 m->valid |= pagebits; 2919 #if 0 /* NOT YET */ 2920 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 2921 frag = DEV_BSIZE - frag; 2922 base += frag; 2923 size -= frag; 2924 if (size < 0) 2925 size = 0; 2926 } 2927 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 2928 #endif 2929 if (base == 0 && size == PAGE_SIZE) { 2930 /* 2931 * The page can only be modified within the pmap if it is 2932 * mapped, and it can only be mapped if it was previously 2933 * fully valid. 2934 */ 2935 if (oldvalid == VM_PAGE_BITS_ALL) 2936 /* 2937 * Perform the pmap_clear_modify() first. Otherwise, 2938 * a concurrent pmap operation, such as 2939 * pmap_protect(), could clear a modification in the 2940 * pmap and set the dirty field on the page before 2941 * pmap_clear_modify() had begun and after the dirty 2942 * field was cleared here. 2943 */ 2944 pmap_clear_modify(m); 2945 m->dirty = 0; 2946 m->oflags &= ~VPO_NOSYNC; 2947 } else if (oldvalid != VM_PAGE_BITS_ALL) 2948 m->dirty &= ~pagebits; 2949 else 2950 vm_page_clear_dirty_mask(m, pagebits); 2951 } 2952 2953 void 2954 vm_page_clear_dirty(vm_page_t m, int base, int size) 2955 { 2956 2957 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 2958 } 2959 2960 /* 2961 * vm_page_set_invalid: 2962 * 2963 * Invalidates DEV_BSIZE'd chunks within a page. Both the 2964 * valid and dirty bits for the effected areas are cleared. 2965 */ 2966 void 2967 vm_page_set_invalid(vm_page_t m, int base, int size) 2968 { 2969 vm_page_bits_t bits; 2970 vm_object_t object; 2971 2972 object = m->object; 2973 VM_OBJECT_ASSERT_WLOCKED(object); 2974 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + 2975 size >= object->un_pager.vnp.vnp_size) 2976 bits = VM_PAGE_BITS_ALL; 2977 else 2978 bits = vm_page_bits(base, size); 2979 if (m->valid == VM_PAGE_BITS_ALL && bits != 0) 2980 pmap_remove_all(m); 2981 KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) || 2982 !pmap_page_is_mapped(m), 2983 ("vm_page_set_invalid: page %p is mapped", m)); 2984 m->valid &= ~bits; 2985 m->dirty &= ~bits; 2986 } 2987 2988 /* 2989 * vm_page_zero_invalid() 2990 * 2991 * The kernel assumes that the invalid portions of a page contain 2992 * garbage, but such pages can be mapped into memory by user code. 2993 * When this occurs, we must zero out the non-valid portions of the 2994 * page so user code sees what it expects. 2995 * 2996 * Pages are most often semi-valid when the end of a file is mapped 2997 * into memory and the file's size is not page aligned. 2998 */ 2999 void 3000 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 3001 { 3002 int b; 3003 int i; 3004 3005 VM_OBJECT_ASSERT_WLOCKED(m->object); 3006 /* 3007 * Scan the valid bits looking for invalid sections that 3008 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 3009 * valid bit may be set ) have already been zerod by 3010 * vm_page_set_validclean(). 3011 */ 3012 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 3013 if (i == (PAGE_SIZE / DEV_BSIZE) || 3014 (m->valid & ((vm_page_bits_t)1 << i))) { 3015 if (i > b) { 3016 pmap_zero_page_area(m, 3017 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 3018 } 3019 b = i + 1; 3020 } 3021 } 3022 3023 /* 3024 * setvalid is TRUE when we can safely set the zero'd areas 3025 * as being valid. We can do this if there are no cache consistancy 3026 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 3027 */ 3028 if (setvalid) 3029 m->valid = VM_PAGE_BITS_ALL; 3030 } 3031 3032 /* 3033 * vm_page_is_valid: 3034 * 3035 * Is (partial) page valid? Note that the case where size == 0 3036 * will return FALSE in the degenerate case where the page is 3037 * entirely invalid, and TRUE otherwise. 3038 */ 3039 int 3040 vm_page_is_valid(vm_page_t m, int base, int size) 3041 { 3042 vm_page_bits_t bits; 3043 3044 VM_OBJECT_ASSERT_LOCKED(m->object); 3045 bits = vm_page_bits(base, size); 3046 return (m->valid != 0 && (m->valid & bits) == bits); 3047 } 3048 3049 /* 3050 * vm_page_ps_is_valid: 3051 * 3052 * Returns TRUE if the entire (super)page is valid and FALSE otherwise. 3053 */ 3054 boolean_t 3055 vm_page_ps_is_valid(vm_page_t m) 3056 { 3057 int i, npages; 3058 3059 VM_OBJECT_ASSERT_LOCKED(m->object); 3060 npages = atop(pagesizes[m->psind]); 3061 3062 /* 3063 * The physically contiguous pages that make up a superpage, i.e., a 3064 * page with a page size index ("psind") greater than zero, will 3065 * occupy adjacent entries in vm_page_array[]. 3066 */ 3067 for (i = 0; i < npages; i++) { 3068 if (m[i].valid != VM_PAGE_BITS_ALL) 3069 return (FALSE); 3070 } 3071 return (TRUE); 3072 } 3073 3074 /* 3075 * Set the page's dirty bits if the page is modified. 3076 */ 3077 void 3078 vm_page_test_dirty(vm_page_t m) 3079 { 3080 3081 VM_OBJECT_ASSERT_WLOCKED(m->object); 3082 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 3083 vm_page_dirty(m); 3084 } 3085 3086 void 3087 vm_page_lock_KBI(vm_page_t m, const char *file, int line) 3088 { 3089 3090 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); 3091 } 3092 3093 void 3094 vm_page_unlock_KBI(vm_page_t m, const char *file, int line) 3095 { 3096 3097 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); 3098 } 3099 3100 int 3101 vm_page_trylock_KBI(vm_page_t m, const char *file, int line) 3102 { 3103 3104 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); 3105 } 3106 3107 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) 3108 void 3109 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line) 3110 { 3111 3112 vm_page_lock_assert_KBI(m, MA_OWNED, file, line); 3113 } 3114 3115 void 3116 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) 3117 { 3118 3119 mtx_assert_(vm_page_lockptr(m), a, file, line); 3120 } 3121 #endif 3122 3123 #ifdef INVARIANTS 3124 void 3125 vm_page_object_lock_assert(vm_page_t m) 3126 { 3127 3128 /* 3129 * Certain of the page's fields may only be modified by the 3130 * holder of the containing object's lock or the exclusive busy. 3131 * holder. Unfortunately, the holder of the write busy is 3132 * not recorded, and thus cannot be checked here. 3133 */ 3134 if (m->object != NULL && !vm_page_xbusied(m)) 3135 VM_OBJECT_ASSERT_WLOCKED(m->object); 3136 } 3137 3138 void 3139 vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits) 3140 { 3141 3142 if ((bits & PGA_WRITEABLE) == 0) 3143 return; 3144 3145 /* 3146 * The PGA_WRITEABLE flag can only be set if the page is 3147 * managed, is exclusively busied or the object is locked. 3148 * Currently, this flag is only set by pmap_enter(). 3149 */ 3150 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 3151 ("PGA_WRITEABLE on unmanaged page")); 3152 if (!vm_page_xbusied(m)) 3153 VM_OBJECT_ASSERT_LOCKED(m->object); 3154 } 3155 #endif 3156 3157 #include "opt_ddb.h" 3158 #ifdef DDB 3159 #include <sys/kernel.h> 3160 3161 #include <ddb/ddb.h> 3162 3163 DB_SHOW_COMMAND(page, vm_page_print_page_info) 3164 { 3165 db_printf("vm_cnt.v_free_count: %d\n", vm_cnt.v_free_count); 3166 db_printf("vm_cnt.v_cache_count: %d\n", vm_cnt.v_cache_count); 3167 db_printf("vm_cnt.v_inactive_count: %d\n", vm_cnt.v_inactive_count); 3168 db_printf("vm_cnt.v_active_count: %d\n", vm_cnt.v_active_count); 3169 db_printf("vm_cnt.v_wire_count: %d\n", vm_cnt.v_wire_count); 3170 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved); 3171 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min); 3172 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target); 3173 db_printf("vm_cnt.v_cache_min: %d\n", vm_cnt.v_cache_min); 3174 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target); 3175 } 3176 3177 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 3178 { 3179 int dom; 3180 3181 db_printf("pq_free %d pq_cache %d\n", 3182 vm_cnt.v_free_count, vm_cnt.v_cache_count); 3183 for (dom = 0; dom < vm_ndomains; dom++) { 3184 db_printf( 3185 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pass %d\n", 3186 dom, 3187 vm_dom[dom].vmd_page_count, 3188 vm_dom[dom].vmd_free_count, 3189 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, 3190 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, 3191 vm_dom[dom].vmd_pass); 3192 } 3193 } 3194 3195 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) 3196 { 3197 vm_page_t m; 3198 boolean_t phys; 3199 3200 if (!have_addr) { 3201 db_printf("show pginfo addr\n"); 3202 return; 3203 } 3204 3205 phys = strchr(modif, 'p') != NULL; 3206 if (phys) 3207 m = PHYS_TO_VM_PAGE(addr); 3208 else 3209 m = (vm_page_t)addr; 3210 db_printf( 3211 "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n" 3212 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", 3213 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, 3214 m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags, 3215 m->flags, m->act_count, m->busy_lock, m->valid, m->dirty); 3216 } 3217 #endif /* DDB */ 3218