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 (vm_pagequeues[]), regardless of other locks or the 68 * busy state of a page. 69 * 70 * * In general, no thread besides the page daemon can acquire or 71 * hold more than one page queue lock at a time. 72 * 73 * * The page daemon can acquire and hold any pair of page queue 74 * locks in any order. 75 * 76 * - The object mutex is held when inserting or removing 77 * pages from an object (vm_page_insert() or vm_page_remove()). 78 * 79 */ 80 81 /* 82 * Resident memory management module. 83 */ 84 85 #include <sys/cdefs.h> 86 __FBSDID("$FreeBSD$"); 87 88 #include "opt_vm.h" 89 90 #include <sys/param.h> 91 #include <sys/systm.h> 92 #include <sys/lock.h> 93 #include <sys/kernel.h> 94 #include <sys/limits.h> 95 #include <sys/malloc.h> 96 #include <sys/msgbuf.h> 97 #include <sys/mutex.h> 98 #include <sys/proc.h> 99 #include <sys/sysctl.h> 100 #include <sys/vmmeter.h> 101 #include <sys/vnode.h> 102 103 #include <vm/vm.h> 104 #include <vm/pmap.h> 105 #include <vm/vm_param.h> 106 #include <vm/vm_kern.h> 107 #include <vm/vm_object.h> 108 #include <vm/vm_page.h> 109 #include <vm/vm_pageout.h> 110 #include <vm/vm_pager.h> 111 #include <vm/vm_phys.h> 112 #include <vm/vm_reserv.h> 113 #include <vm/vm_extern.h> 114 #include <vm/uma.h> 115 #include <vm/uma_int.h> 116 117 #include <machine/md_var.h> 118 119 /* 120 * Associated with page of user-allocatable memory is a 121 * page structure. 122 */ 123 124 struct vm_pagequeue vm_pagequeues[PQ_COUNT] = { 125 [PQ_INACTIVE] = { 126 .pq_pl = TAILQ_HEAD_INITIALIZER( 127 vm_pagequeues[PQ_INACTIVE].pq_pl), 128 .pq_cnt = &cnt.v_inactive_count, 129 .pq_name = "vm inactive pagequeue" 130 }, 131 [PQ_ACTIVE] = { 132 .pq_pl = TAILQ_HEAD_INITIALIZER( 133 vm_pagequeues[PQ_ACTIVE].pq_pl), 134 .pq_cnt = &cnt.v_active_count, 135 .pq_name = "vm active pagequeue" 136 } 137 }; 138 struct mtx_padalign vm_page_queue_free_mtx; 139 140 struct mtx_padalign pa_lock[PA_LOCK_COUNT]; 141 142 vm_page_t vm_page_array; 143 long vm_page_array_size; 144 long first_page; 145 int vm_page_zero_count; 146 147 static int boot_pages = UMA_BOOT_PAGES; 148 TUNABLE_INT("vm.boot_pages", &boot_pages); 149 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RD, &boot_pages, 0, 150 "number of pages allocated for bootstrapping the VM system"); 151 152 static int pa_tryrelock_restart; 153 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD, 154 &pa_tryrelock_restart, 0, "Number of tryrelock restarts"); 155 156 static uma_zone_t fakepg_zone; 157 158 static struct vnode *vm_page_alloc_init(vm_page_t m); 159 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits); 160 static void vm_page_enqueue(int queue, vm_page_t m); 161 static void vm_page_init_fakepg(void *dummy); 162 163 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init_fakepg, NULL); 164 165 static void 166 vm_page_init_fakepg(void *dummy) 167 { 168 169 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL, 170 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM); 171 } 172 173 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */ 174 #if PAGE_SIZE == 32768 175 #ifdef CTASSERT 176 CTASSERT(sizeof(u_long) >= 8); 177 #endif 178 #endif 179 180 /* 181 * Try to acquire a physical address lock while a pmap is locked. If we 182 * fail to trylock we unlock and lock the pmap directly and cache the 183 * locked pa in *locked. The caller should then restart their loop in case 184 * the virtual to physical mapping has changed. 185 */ 186 int 187 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked) 188 { 189 vm_paddr_t lockpa; 190 191 lockpa = *locked; 192 *locked = pa; 193 if (lockpa) { 194 PA_LOCK_ASSERT(lockpa, MA_OWNED); 195 if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa)) 196 return (0); 197 PA_UNLOCK(lockpa); 198 } 199 if (PA_TRYLOCK(pa)) 200 return (0); 201 PMAP_UNLOCK(pmap); 202 atomic_add_int(&pa_tryrelock_restart, 1); 203 PA_LOCK(pa); 204 PMAP_LOCK(pmap); 205 return (EAGAIN); 206 } 207 208 /* 209 * vm_set_page_size: 210 * 211 * Sets the page size, perhaps based upon the memory 212 * size. Must be called before any use of page-size 213 * dependent functions. 214 */ 215 void 216 vm_set_page_size(void) 217 { 218 if (cnt.v_page_size == 0) 219 cnt.v_page_size = PAGE_SIZE; 220 if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0) 221 panic("vm_set_page_size: page size not a power of two"); 222 } 223 224 /* 225 * vm_page_blacklist_lookup: 226 * 227 * See if a physical address in this page has been listed 228 * in the blacklist tunable. Entries in the tunable are 229 * separated by spaces or commas. If an invalid integer is 230 * encountered then the rest of the string is skipped. 231 */ 232 static int 233 vm_page_blacklist_lookup(char *list, vm_paddr_t pa) 234 { 235 vm_paddr_t bad; 236 char *cp, *pos; 237 238 for (pos = list; *pos != '\0'; pos = cp) { 239 bad = strtoq(pos, &cp, 0); 240 if (*cp != '\0') { 241 if (*cp == ' ' || *cp == ',') { 242 cp++; 243 if (cp == pos) 244 continue; 245 } else 246 break; 247 } 248 if (pa == trunc_page(bad)) 249 return (1); 250 } 251 return (0); 252 } 253 254 /* 255 * vm_page_startup: 256 * 257 * Initializes the resident memory module. 258 * 259 * Allocates memory for the page cells, and 260 * for the object/offset-to-page hash table headers. 261 * Each page cell is initialized and placed on the free list. 262 */ 263 vm_offset_t 264 vm_page_startup(vm_offset_t vaddr) 265 { 266 vm_offset_t mapped; 267 vm_paddr_t page_range; 268 vm_paddr_t new_end; 269 int i; 270 vm_paddr_t pa; 271 vm_paddr_t last_pa; 272 char *list; 273 274 /* the biggest memory array is the second group of pages */ 275 vm_paddr_t end; 276 vm_paddr_t biggestsize; 277 vm_paddr_t low_water, high_water; 278 int biggestone; 279 280 biggestsize = 0; 281 biggestone = 0; 282 vaddr = round_page(vaddr); 283 284 for (i = 0; phys_avail[i + 1]; i += 2) { 285 phys_avail[i] = round_page(phys_avail[i]); 286 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 287 } 288 289 low_water = phys_avail[0]; 290 high_water = phys_avail[1]; 291 292 for (i = 0; phys_avail[i + 1]; i += 2) { 293 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i]; 294 295 if (size > biggestsize) { 296 biggestone = i; 297 biggestsize = size; 298 } 299 if (phys_avail[i] < low_water) 300 low_water = phys_avail[i]; 301 if (phys_avail[i + 1] > high_water) 302 high_water = phys_avail[i + 1]; 303 } 304 305 #ifdef XEN 306 low_water = 0; 307 #endif 308 309 end = phys_avail[biggestone+1]; 310 311 /* 312 * Initialize the page and queue locks. 313 */ 314 mtx_init(&vm_page_queue_free_mtx, "vm page free queue", NULL, MTX_DEF); 315 for (i = 0; i < PA_LOCK_COUNT; i++) 316 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF); 317 for (i = 0; i < PQ_COUNT; i++) 318 vm_pagequeue_init_lock(&vm_pagequeues[i]); 319 320 /* 321 * Allocate memory for use when boot strapping the kernel memory 322 * allocator. 323 */ 324 new_end = end - (boot_pages * UMA_SLAB_SIZE); 325 new_end = trunc_page(new_end); 326 mapped = pmap_map(&vaddr, new_end, end, 327 VM_PROT_READ | VM_PROT_WRITE); 328 bzero((void *)mapped, end - new_end); 329 uma_startup((void *)mapped, boot_pages); 330 331 #if defined(__amd64__) || defined(__i386__) || defined(__arm__) || \ 332 defined(__mips__) 333 /* 334 * Allocate a bitmap to indicate that a random physical page 335 * needs to be included in a minidump. 336 * 337 * The amd64 port needs this to indicate which direct map pages 338 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 339 * 340 * However, i386 still needs this workspace internally within the 341 * minidump code. In theory, they are not needed on i386, but are 342 * included should the sf_buf code decide to use them. 343 */ 344 last_pa = 0; 345 for (i = 0; dump_avail[i + 1] != 0; i += 2) 346 if (dump_avail[i + 1] > last_pa) 347 last_pa = dump_avail[i + 1]; 348 page_range = last_pa / PAGE_SIZE; 349 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 350 new_end -= vm_page_dump_size; 351 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, 352 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); 353 bzero((void *)vm_page_dump, vm_page_dump_size); 354 #endif 355 #ifdef __amd64__ 356 /* 357 * Request that the physical pages underlying the message buffer be 358 * included in a crash dump. Since the message buffer is accessed 359 * through the direct map, they are not automatically included. 360 */ 361 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); 362 last_pa = pa + round_page(msgbufsize); 363 while (pa < last_pa) { 364 dump_add_page(pa); 365 pa += PAGE_SIZE; 366 } 367 #endif 368 /* 369 * Compute the number of pages of memory that will be available for 370 * use (taking into account the overhead of a page structure per 371 * page). 372 */ 373 first_page = low_water / PAGE_SIZE; 374 #ifdef VM_PHYSSEG_SPARSE 375 page_range = 0; 376 for (i = 0; phys_avail[i + 1] != 0; i += 2) 377 page_range += atop(phys_avail[i + 1] - phys_avail[i]); 378 #elif defined(VM_PHYSSEG_DENSE) 379 page_range = high_water / PAGE_SIZE - first_page; 380 #else 381 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 382 #endif 383 end = new_end; 384 385 /* 386 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 387 */ 388 vaddr += PAGE_SIZE; 389 390 /* 391 * Initialize the mem entry structures now, and put them in the free 392 * queue. 393 */ 394 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 395 mapped = pmap_map(&vaddr, new_end, end, 396 VM_PROT_READ | VM_PROT_WRITE); 397 vm_page_array = (vm_page_t) mapped; 398 #if VM_NRESERVLEVEL > 0 399 /* 400 * Allocate memory for the reservation management system's data 401 * structures. 402 */ 403 new_end = vm_reserv_startup(&vaddr, new_end, high_water); 404 #endif 405 #if defined(__amd64__) || defined(__mips__) 406 /* 407 * pmap_map on amd64 and mips can come out of the direct-map, not kvm 408 * like i386, so the pages must be tracked for a crashdump to include 409 * this data. This includes the vm_page_array and the early UMA 410 * bootstrap pages. 411 */ 412 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE) 413 dump_add_page(pa); 414 #endif 415 phys_avail[biggestone + 1] = new_end; 416 417 /* 418 * Clear all of the page structures 419 */ 420 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 421 for (i = 0; i < page_range; i++) 422 vm_page_array[i].order = VM_NFREEORDER; 423 vm_page_array_size = page_range; 424 425 /* 426 * Initialize the physical memory allocator. 427 */ 428 vm_phys_init(); 429 430 /* 431 * Add every available physical page that is not blacklisted to 432 * the free lists. 433 */ 434 cnt.v_page_count = 0; 435 cnt.v_free_count = 0; 436 list = getenv("vm.blacklist"); 437 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 438 pa = phys_avail[i]; 439 last_pa = phys_avail[i + 1]; 440 while (pa < last_pa) { 441 if (list != NULL && 442 vm_page_blacklist_lookup(list, pa)) 443 printf("Skipping page with pa 0x%jx\n", 444 (uintmax_t)pa); 445 else 446 vm_phys_add_page(pa); 447 pa += PAGE_SIZE; 448 } 449 } 450 freeenv(list); 451 #if VM_NRESERVLEVEL > 0 452 /* 453 * Initialize the reservation management system. 454 */ 455 vm_reserv_init(); 456 #endif 457 return (vaddr); 458 } 459 460 void 461 vm_page_reference(vm_page_t m) 462 { 463 464 vm_page_aflag_set(m, PGA_REFERENCED); 465 } 466 467 void 468 vm_page_busy(vm_page_t m) 469 { 470 471 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 472 KASSERT((m->oflags & VPO_BUSY) == 0, 473 ("vm_page_busy: page already busy!!!")); 474 m->oflags |= VPO_BUSY; 475 } 476 477 /* 478 * vm_page_flash: 479 * 480 * wakeup anyone waiting for the page. 481 */ 482 void 483 vm_page_flash(vm_page_t m) 484 { 485 486 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 487 if (m->oflags & VPO_WANTED) { 488 m->oflags &= ~VPO_WANTED; 489 wakeup(m); 490 } 491 } 492 493 /* 494 * vm_page_wakeup: 495 * 496 * clear the VPO_BUSY flag and wakeup anyone waiting for the 497 * page. 498 * 499 */ 500 void 501 vm_page_wakeup(vm_page_t m) 502 { 503 504 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 505 KASSERT(m->oflags & VPO_BUSY, ("vm_page_wakeup: page not busy!!!")); 506 m->oflags &= ~VPO_BUSY; 507 vm_page_flash(m); 508 } 509 510 void 511 vm_page_io_start(vm_page_t m) 512 { 513 514 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 515 m->busy++; 516 } 517 518 void 519 vm_page_io_finish(vm_page_t m) 520 { 521 522 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 523 KASSERT(m->busy > 0, ("vm_page_io_finish: page %p is not busy", m)); 524 m->busy--; 525 if (m->busy == 0) 526 vm_page_flash(m); 527 } 528 529 /* 530 * Keep page from being freed by the page daemon 531 * much of the same effect as wiring, except much lower 532 * overhead and should be used only for *very* temporary 533 * holding ("wiring"). 534 */ 535 void 536 vm_page_hold(vm_page_t mem) 537 { 538 539 vm_page_lock_assert(mem, MA_OWNED); 540 mem->hold_count++; 541 } 542 543 void 544 vm_page_unhold(vm_page_t mem) 545 { 546 547 vm_page_lock_assert(mem, MA_OWNED); 548 --mem->hold_count; 549 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!")); 550 if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0) 551 vm_page_free_toq(mem); 552 } 553 554 /* 555 * vm_page_unhold_pages: 556 * 557 * Unhold each of the pages that is referenced by the given array. 558 */ 559 void 560 vm_page_unhold_pages(vm_page_t *ma, int count) 561 { 562 struct mtx *mtx, *new_mtx; 563 564 mtx = NULL; 565 for (; count != 0; count--) { 566 /* 567 * Avoid releasing and reacquiring the same page lock. 568 */ 569 new_mtx = vm_page_lockptr(*ma); 570 if (mtx != new_mtx) { 571 if (mtx != NULL) 572 mtx_unlock(mtx); 573 mtx = new_mtx; 574 mtx_lock(mtx); 575 } 576 vm_page_unhold(*ma); 577 ma++; 578 } 579 if (mtx != NULL) 580 mtx_unlock(mtx); 581 } 582 583 vm_page_t 584 PHYS_TO_VM_PAGE(vm_paddr_t pa) 585 { 586 vm_page_t m; 587 588 #ifdef VM_PHYSSEG_SPARSE 589 m = vm_phys_paddr_to_vm_page(pa); 590 if (m == NULL) 591 m = vm_phys_fictitious_to_vm_page(pa); 592 return (m); 593 #elif defined(VM_PHYSSEG_DENSE) 594 long pi; 595 596 pi = atop(pa); 597 if (pi >= first_page && (pi - first_page) < vm_page_array_size) { 598 m = &vm_page_array[pi - first_page]; 599 return (m); 600 } 601 return (vm_phys_fictitious_to_vm_page(pa)); 602 #else 603 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 604 #endif 605 } 606 607 /* 608 * vm_page_getfake: 609 * 610 * Create a fictitious page with the specified physical address and 611 * memory attribute. The memory attribute is the only the machine- 612 * dependent aspect of a fictitious page that must be initialized. 613 */ 614 vm_page_t 615 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr) 616 { 617 vm_page_t m; 618 619 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO); 620 vm_page_initfake(m, paddr, memattr); 621 return (m); 622 } 623 624 void 625 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 626 { 627 628 if ((m->flags & PG_FICTITIOUS) != 0) { 629 /* 630 * The page's memattr might have changed since the 631 * previous initialization. Update the pmap to the 632 * new memattr. 633 */ 634 goto memattr; 635 } 636 m->phys_addr = paddr; 637 m->queue = PQ_NONE; 638 /* Fictitious pages don't use "segind". */ 639 m->flags = PG_FICTITIOUS; 640 /* Fictitious pages don't use "order" or "pool". */ 641 m->oflags = VPO_BUSY | VPO_UNMANAGED; 642 m->wire_count = 1; 643 memattr: 644 pmap_page_set_memattr(m, memattr); 645 } 646 647 /* 648 * vm_page_putfake: 649 * 650 * Release a fictitious page. 651 */ 652 void 653 vm_page_putfake(vm_page_t m) 654 { 655 656 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m)); 657 KASSERT((m->flags & PG_FICTITIOUS) != 0, 658 ("vm_page_putfake: bad page %p", m)); 659 uma_zfree(fakepg_zone, m); 660 } 661 662 /* 663 * vm_page_updatefake: 664 * 665 * Update the given fictitious page to the specified physical address and 666 * memory attribute. 667 */ 668 void 669 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 670 { 671 672 KASSERT((m->flags & PG_FICTITIOUS) != 0, 673 ("vm_page_updatefake: bad page %p", m)); 674 m->phys_addr = paddr; 675 pmap_page_set_memattr(m, memattr); 676 } 677 678 /* 679 * vm_page_free: 680 * 681 * Free a page. 682 */ 683 void 684 vm_page_free(vm_page_t m) 685 { 686 687 m->flags &= ~PG_ZERO; 688 vm_page_free_toq(m); 689 } 690 691 /* 692 * vm_page_free_zero: 693 * 694 * Free a page to the zerod-pages queue 695 */ 696 void 697 vm_page_free_zero(vm_page_t m) 698 { 699 700 m->flags |= PG_ZERO; 701 vm_page_free_toq(m); 702 } 703 704 /* 705 * Unbusy and handle the page queueing for a page from the VOP_GETPAGES() 706 * array which is not the request page. 707 */ 708 void 709 vm_page_readahead_finish(vm_page_t m) 710 { 711 712 if (m->valid != 0) { 713 /* 714 * Since the page is not the requested page, whether 715 * it should be activated or deactivated is not 716 * obvious. Empirical results have shown that 717 * deactivating the page is usually the best choice, 718 * unless the page is wanted by another thread. 719 */ 720 if (m->oflags & VPO_WANTED) { 721 vm_page_lock(m); 722 vm_page_activate(m); 723 vm_page_unlock(m); 724 } else { 725 vm_page_lock(m); 726 vm_page_deactivate(m); 727 vm_page_unlock(m); 728 } 729 vm_page_wakeup(m); 730 } else { 731 /* 732 * Free the completely invalid page. Such page state 733 * occurs due to the short read operation which did 734 * not covered our page at all, or in case when a read 735 * error happens. 736 */ 737 vm_page_lock(m); 738 vm_page_free(m); 739 vm_page_unlock(m); 740 } 741 } 742 743 /* 744 * vm_page_sleep: 745 * 746 * Sleep and release the page lock. 747 * 748 * The object containing the given page must be locked. 749 */ 750 void 751 vm_page_sleep(vm_page_t m, const char *msg) 752 { 753 754 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 755 if (mtx_owned(vm_page_lockptr(m))) 756 vm_page_unlock(m); 757 758 /* 759 * It's possible that while we sleep, the page will get 760 * unbusied and freed. If we are holding the object 761 * lock, we will assume we hold a reference to the object 762 * such that even if m->object changes, we can re-lock 763 * it. 764 */ 765 m->oflags |= VPO_WANTED; 766 msleep(m, VM_OBJECT_MTX(m->object), PVM, msg, 0); 767 } 768 769 /* 770 * vm_page_dirty_KBI: [ internal use only ] 771 * 772 * Set all bits in the page's dirty field. 773 * 774 * The object containing the specified page must be locked if the 775 * call is made from the machine-independent layer. 776 * 777 * See vm_page_clear_dirty_mask(). 778 * 779 * This function should only be called by vm_page_dirty(). 780 */ 781 void 782 vm_page_dirty_KBI(vm_page_t m) 783 { 784 785 /* These assertions refer to this operation by its public name. */ 786 KASSERT((m->flags & PG_CACHED) == 0, 787 ("vm_page_dirty: page in cache!")); 788 KASSERT(!VM_PAGE_IS_FREE(m), 789 ("vm_page_dirty: page is free!")); 790 KASSERT(m->valid == VM_PAGE_BITS_ALL, 791 ("vm_page_dirty: page is invalid!")); 792 m->dirty = VM_PAGE_BITS_ALL; 793 } 794 795 /* 796 * vm_page_splay: 797 * 798 * Implements Sleator and Tarjan's top-down splay algorithm. Returns 799 * the vm_page containing the given pindex. If, however, that 800 * pindex is not found in the vm_object, returns a vm_page that is 801 * adjacent to the pindex, coming before or after it. 802 */ 803 vm_page_t 804 vm_page_splay(vm_pindex_t pindex, vm_page_t root) 805 { 806 struct vm_page dummy; 807 vm_page_t lefttreemax, righttreemin, y; 808 809 if (root == NULL) 810 return (root); 811 lefttreemax = righttreemin = &dummy; 812 for (;; root = y) { 813 if (pindex < root->pindex) { 814 if ((y = root->left) == NULL) 815 break; 816 if (pindex < y->pindex) { 817 /* Rotate right. */ 818 root->left = y->right; 819 y->right = root; 820 root = y; 821 if ((y = root->left) == NULL) 822 break; 823 } 824 /* Link into the new root's right tree. */ 825 righttreemin->left = root; 826 righttreemin = root; 827 } else if (pindex > root->pindex) { 828 if ((y = root->right) == NULL) 829 break; 830 if (pindex > y->pindex) { 831 /* Rotate left. */ 832 root->right = y->left; 833 y->left = root; 834 root = y; 835 if ((y = root->right) == NULL) 836 break; 837 } 838 /* Link into the new root's left tree. */ 839 lefttreemax->right = root; 840 lefttreemax = root; 841 } else 842 break; 843 } 844 /* Assemble the new root. */ 845 lefttreemax->right = root->left; 846 righttreemin->left = root->right; 847 root->left = dummy.right; 848 root->right = dummy.left; 849 return (root); 850 } 851 852 /* 853 * vm_page_insert: [ internal use only ] 854 * 855 * Inserts the given mem entry into the object and object list. 856 * 857 * The pagetables are not updated but will presumably fault the page 858 * in if necessary, or if a kernel page the caller will at some point 859 * enter the page into the kernel's pmap. We are not allowed to sleep 860 * here so we *can't* do this anyway. 861 * 862 * The object must be locked. 863 */ 864 void 865 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 866 { 867 vm_page_t root; 868 869 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 870 if (m->object != NULL) 871 panic("vm_page_insert: page already inserted"); 872 873 /* 874 * Record the object/offset pair in this page 875 */ 876 m->object = object; 877 m->pindex = pindex; 878 879 /* 880 * Now link into the object's ordered list of backed pages. 881 */ 882 root = object->root; 883 if (root == NULL) { 884 m->left = NULL; 885 m->right = NULL; 886 TAILQ_INSERT_TAIL(&object->memq, m, listq); 887 } else { 888 root = vm_page_splay(pindex, root); 889 if (pindex < root->pindex) { 890 m->left = root->left; 891 m->right = root; 892 root->left = NULL; 893 TAILQ_INSERT_BEFORE(root, m, listq); 894 } else if (pindex == root->pindex) 895 panic("vm_page_insert: offset already allocated"); 896 else { 897 m->right = root->right; 898 m->left = root; 899 root->right = NULL; 900 TAILQ_INSERT_AFTER(&object->memq, root, m, listq); 901 } 902 } 903 object->root = m; 904 905 /* 906 * Show that the object has one more resident page. 907 */ 908 object->resident_page_count++; 909 910 /* 911 * Hold the vnode until the last page is released. 912 */ 913 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 914 vhold(object->handle); 915 916 /* 917 * Since we are inserting a new and possibly dirty page, 918 * update the object's OBJ_MIGHTBEDIRTY flag. 919 */ 920 if (pmap_page_is_write_mapped(m)) 921 vm_object_set_writeable_dirty(object); 922 } 923 924 /* 925 * vm_page_remove: 926 * 927 * Removes the given mem entry from the object/offset-page 928 * table and the object page list, but do not invalidate/terminate 929 * the backing store. 930 * 931 * The underlying pmap entry (if any) is NOT removed here. 932 * 933 * The object must be locked. The page must be locked if it is managed. 934 */ 935 void 936 vm_page_remove(vm_page_t m) 937 { 938 vm_object_t object; 939 vm_page_t next, prev, root; 940 941 if ((m->oflags & VPO_UNMANAGED) == 0) 942 vm_page_lock_assert(m, MA_OWNED); 943 if ((object = m->object) == NULL) 944 return; 945 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 946 if (m->oflags & VPO_BUSY) { 947 m->oflags &= ~VPO_BUSY; 948 vm_page_flash(m); 949 } 950 951 /* 952 * Now remove from the object's list of backed pages. 953 */ 954 if ((next = TAILQ_NEXT(m, listq)) != NULL && next->left == m) { 955 /* 956 * Since the page's successor in the list is also its parent 957 * in the tree, its right subtree must be empty. 958 */ 959 next->left = m->left; 960 KASSERT(m->right == NULL, 961 ("vm_page_remove: page %p has right child", m)); 962 } else if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL && 963 prev->right == m) { 964 /* 965 * Since the page's predecessor in the list is also its parent 966 * in the tree, its left subtree must be empty. 967 */ 968 KASSERT(m->left == NULL, 969 ("vm_page_remove: page %p has left child", m)); 970 prev->right = m->right; 971 } else { 972 if (m != object->root) 973 vm_page_splay(m->pindex, object->root); 974 if (m->left == NULL) 975 root = m->right; 976 else if (m->right == NULL) 977 root = m->left; 978 else { 979 /* 980 * Move the page's successor to the root, because 981 * pages are usually removed in ascending order. 982 */ 983 if (m->right != next) 984 vm_page_splay(m->pindex, m->right); 985 next->left = m->left; 986 root = next; 987 } 988 object->root = root; 989 } 990 TAILQ_REMOVE(&object->memq, m, listq); 991 992 /* 993 * And show that the object has one fewer resident page. 994 */ 995 object->resident_page_count--; 996 997 /* 998 * The vnode may now be recycled. 999 */ 1000 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 1001 vdrop(object->handle); 1002 1003 m->object = NULL; 1004 } 1005 1006 /* 1007 * vm_page_lookup: 1008 * 1009 * Returns the page associated with the object/offset 1010 * pair specified; if none is found, NULL is returned. 1011 * 1012 * The object must be locked. 1013 */ 1014 vm_page_t 1015 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1016 { 1017 vm_page_t m; 1018 1019 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1020 if ((m = object->root) != NULL && m->pindex != pindex) { 1021 m = vm_page_splay(pindex, m); 1022 if ((object->root = m)->pindex != pindex) 1023 m = NULL; 1024 } 1025 return (m); 1026 } 1027 1028 /* 1029 * vm_page_find_least: 1030 * 1031 * Returns the page associated with the object with least pindex 1032 * greater than or equal to the parameter pindex, or NULL. 1033 * 1034 * The object must be locked. 1035 */ 1036 vm_page_t 1037 vm_page_find_least(vm_object_t object, vm_pindex_t pindex) 1038 { 1039 vm_page_t m; 1040 1041 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1042 if ((m = TAILQ_FIRST(&object->memq)) != NULL) { 1043 if (m->pindex < pindex) { 1044 m = vm_page_splay(pindex, object->root); 1045 if ((object->root = m)->pindex < pindex) 1046 m = TAILQ_NEXT(m, listq); 1047 } 1048 } 1049 return (m); 1050 } 1051 1052 /* 1053 * Returns the given page's successor (by pindex) within the object if it is 1054 * resident; if none is found, NULL is returned. 1055 * 1056 * The object must be locked. 1057 */ 1058 vm_page_t 1059 vm_page_next(vm_page_t m) 1060 { 1061 vm_page_t next; 1062 1063 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1064 if ((next = TAILQ_NEXT(m, listq)) != NULL && 1065 next->pindex != m->pindex + 1) 1066 next = NULL; 1067 return (next); 1068 } 1069 1070 /* 1071 * Returns the given page's predecessor (by pindex) within the object if it is 1072 * resident; if none is found, NULL is returned. 1073 * 1074 * The object must be locked. 1075 */ 1076 vm_page_t 1077 vm_page_prev(vm_page_t m) 1078 { 1079 vm_page_t prev; 1080 1081 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1082 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL && 1083 prev->pindex != m->pindex - 1) 1084 prev = NULL; 1085 return (prev); 1086 } 1087 1088 /* 1089 * vm_page_rename: 1090 * 1091 * Move the given memory entry from its 1092 * current object to the specified target object/offset. 1093 * 1094 * Note: swap associated with the page must be invalidated by the move. We 1095 * have to do this for several reasons: (1) we aren't freeing the 1096 * page, (2) we are dirtying the page, (3) the VM system is probably 1097 * moving the page from object A to B, and will then later move 1098 * the backing store from A to B and we can't have a conflict. 1099 * 1100 * Note: we *always* dirty the page. It is necessary both for the 1101 * fact that we moved it, and because we may be invalidating 1102 * swap. If the page is on the cache, we have to deactivate it 1103 * or vm_page_dirty() will panic. Dirty pages are not allowed 1104 * on the cache. 1105 * 1106 * The objects must be locked. The page must be locked if it is managed. 1107 */ 1108 void 1109 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1110 { 1111 1112 vm_page_remove(m); 1113 vm_page_insert(m, new_object, new_pindex); 1114 vm_page_dirty(m); 1115 } 1116 1117 /* 1118 * Convert all of the given object's cached pages that have a 1119 * pindex within the given range into free pages. If the value 1120 * zero is given for "end", then the range's upper bound is 1121 * infinity. If the given object is backed by a vnode and it 1122 * transitions from having one or more cached pages to none, the 1123 * vnode's hold count is reduced. 1124 */ 1125 void 1126 vm_page_cache_free(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 1127 { 1128 vm_page_t m, m_next; 1129 boolean_t empty; 1130 1131 mtx_lock(&vm_page_queue_free_mtx); 1132 if (__predict_false(object->cache == NULL)) { 1133 mtx_unlock(&vm_page_queue_free_mtx); 1134 return; 1135 } 1136 m = object->cache = vm_page_splay(start, object->cache); 1137 if (m->pindex < start) { 1138 if (m->right == NULL) 1139 m = NULL; 1140 else { 1141 m_next = vm_page_splay(start, m->right); 1142 m_next->left = m; 1143 m->right = NULL; 1144 m = object->cache = m_next; 1145 } 1146 } 1147 1148 /* 1149 * At this point, "m" is either (1) a reference to the page 1150 * with the least pindex that is greater than or equal to 1151 * "start" or (2) NULL. 1152 */ 1153 for (; m != NULL && (m->pindex < end || end == 0); m = m_next) { 1154 /* 1155 * Find "m"'s successor and remove "m" from the 1156 * object's cache. 1157 */ 1158 if (m->right == NULL) { 1159 object->cache = m->left; 1160 m_next = NULL; 1161 } else { 1162 m_next = vm_page_splay(start, m->right); 1163 m_next->left = m->left; 1164 object->cache = m_next; 1165 } 1166 /* Convert "m" to a free page. */ 1167 m->object = NULL; 1168 m->valid = 0; 1169 /* Clear PG_CACHED and set PG_FREE. */ 1170 m->flags ^= PG_CACHED | PG_FREE; 1171 KASSERT((m->flags & (PG_CACHED | PG_FREE)) == PG_FREE, 1172 ("vm_page_cache_free: page %p has inconsistent flags", m)); 1173 cnt.v_cache_count--; 1174 cnt.v_free_count++; 1175 } 1176 empty = object->cache == NULL; 1177 mtx_unlock(&vm_page_queue_free_mtx); 1178 if (object->type == OBJT_VNODE && empty) 1179 vdrop(object->handle); 1180 } 1181 1182 /* 1183 * Returns the cached page that is associated with the given 1184 * object and offset. If, however, none exists, returns NULL. 1185 * 1186 * The free page queue must be locked. 1187 */ 1188 static inline vm_page_t 1189 vm_page_cache_lookup(vm_object_t object, vm_pindex_t pindex) 1190 { 1191 vm_page_t m; 1192 1193 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1194 if ((m = object->cache) != NULL && m->pindex != pindex) { 1195 m = vm_page_splay(pindex, m); 1196 if ((object->cache = m)->pindex != pindex) 1197 m = NULL; 1198 } 1199 return (m); 1200 } 1201 1202 /* 1203 * Remove the given cached page from its containing object's 1204 * collection of cached pages. 1205 * 1206 * The free page queue must be locked. 1207 */ 1208 static void 1209 vm_page_cache_remove(vm_page_t m) 1210 { 1211 vm_object_t object; 1212 vm_page_t root; 1213 1214 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1215 KASSERT((m->flags & PG_CACHED) != 0, 1216 ("vm_page_cache_remove: page %p is not cached", m)); 1217 object = m->object; 1218 if (m != object->cache) { 1219 root = vm_page_splay(m->pindex, object->cache); 1220 KASSERT(root == m, 1221 ("vm_page_cache_remove: page %p is not cached in object %p", 1222 m, object)); 1223 } 1224 if (m->left == NULL) 1225 root = m->right; 1226 else if (m->right == NULL) 1227 root = m->left; 1228 else { 1229 root = vm_page_splay(m->pindex, m->left); 1230 root->right = m->right; 1231 } 1232 object->cache = root; 1233 m->object = NULL; 1234 cnt.v_cache_count--; 1235 } 1236 1237 /* 1238 * Transfer all of the cached pages with offset greater than or 1239 * equal to 'offidxstart' from the original object's cache to the 1240 * new object's cache. However, any cached pages with offset 1241 * greater than or equal to the new object's size are kept in the 1242 * original object. Initially, the new object's cache must be 1243 * empty. Offset 'offidxstart' in the original object must 1244 * correspond to offset zero in the new object. 1245 * 1246 * The new object must be locked. 1247 */ 1248 void 1249 vm_page_cache_transfer(vm_object_t orig_object, vm_pindex_t offidxstart, 1250 vm_object_t new_object) 1251 { 1252 vm_page_t m, m_next; 1253 1254 /* 1255 * Insertion into an object's collection of cached pages 1256 * requires the object to be locked. In contrast, removal does 1257 * not. 1258 */ 1259 VM_OBJECT_LOCK_ASSERT(new_object, MA_OWNED); 1260 KASSERT(new_object->cache == NULL, 1261 ("vm_page_cache_transfer: object %p has cached pages", 1262 new_object)); 1263 mtx_lock(&vm_page_queue_free_mtx); 1264 if ((m = orig_object->cache) != NULL) { 1265 /* 1266 * Transfer all of the pages with offset greater than or 1267 * equal to 'offidxstart' from the original object's 1268 * cache to the new object's cache. 1269 */ 1270 m = vm_page_splay(offidxstart, m); 1271 if (m->pindex < offidxstart) { 1272 orig_object->cache = m; 1273 new_object->cache = m->right; 1274 m->right = NULL; 1275 } else { 1276 orig_object->cache = m->left; 1277 new_object->cache = m; 1278 m->left = NULL; 1279 } 1280 while ((m = new_object->cache) != NULL) { 1281 if ((m->pindex - offidxstart) >= new_object->size) { 1282 /* 1283 * Return all of the cached pages with 1284 * offset greater than or equal to the 1285 * new object's size to the original 1286 * object's cache. 1287 */ 1288 new_object->cache = m->left; 1289 m->left = orig_object->cache; 1290 orig_object->cache = m; 1291 break; 1292 } 1293 m_next = vm_page_splay(m->pindex, m->right); 1294 /* Update the page's object and offset. */ 1295 m->object = new_object; 1296 m->pindex -= offidxstart; 1297 if (m_next == NULL) 1298 break; 1299 m->right = NULL; 1300 m_next->left = m; 1301 new_object->cache = m_next; 1302 } 1303 KASSERT(new_object->cache == NULL || 1304 new_object->type == OBJT_SWAP, 1305 ("vm_page_cache_transfer: object %p's type is incompatible" 1306 " with cached pages", new_object)); 1307 } 1308 mtx_unlock(&vm_page_queue_free_mtx); 1309 } 1310 1311 /* 1312 * Returns TRUE if a cached page is associated with the given object and 1313 * offset, and FALSE otherwise. 1314 * 1315 * The object must be locked. 1316 */ 1317 boolean_t 1318 vm_page_is_cached(vm_object_t object, vm_pindex_t pindex) 1319 { 1320 vm_page_t m; 1321 1322 /* 1323 * Insertion into an object's collection of cached pages requires the 1324 * object to be locked. Therefore, if the object is locked and the 1325 * object's collection is empty, there is no need to acquire the free 1326 * page queues lock in order to prove that the specified page doesn't 1327 * exist. 1328 */ 1329 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1330 if (__predict_true(object->cache == NULL)) 1331 return (FALSE); 1332 mtx_lock(&vm_page_queue_free_mtx); 1333 m = vm_page_cache_lookup(object, pindex); 1334 mtx_unlock(&vm_page_queue_free_mtx); 1335 return (m != NULL); 1336 } 1337 1338 /* 1339 * vm_page_alloc: 1340 * 1341 * Allocate and return a page that is associated with the specified 1342 * object and offset pair. By default, this page has the flag VPO_BUSY 1343 * set. 1344 * 1345 * The caller must always specify an allocation class. 1346 * 1347 * allocation classes: 1348 * VM_ALLOC_NORMAL normal process request 1349 * VM_ALLOC_SYSTEM system *really* needs a page 1350 * VM_ALLOC_INTERRUPT interrupt time request 1351 * 1352 * optional allocation flags: 1353 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1354 * intends to allocate 1355 * VM_ALLOC_IFCACHED return page only if it is cached 1356 * VM_ALLOC_IFNOTCACHED return NULL, do not reactivate if the page 1357 * is cached 1358 * VM_ALLOC_NOBUSY do not set the flag VPO_BUSY on the page 1359 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1360 * VM_ALLOC_NOOBJ page is not associated with an object and 1361 * should not have the flag VPO_BUSY set 1362 * VM_ALLOC_WIRED wire the allocated page 1363 * VM_ALLOC_ZERO prefer a zeroed page 1364 * 1365 * This routine may not sleep. 1366 */ 1367 vm_page_t 1368 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 1369 { 1370 struct vnode *vp = NULL; 1371 vm_object_t m_object; 1372 vm_page_t m; 1373 int flags, req_class; 1374 1375 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0), 1376 ("vm_page_alloc: inconsistent object/req")); 1377 if (object != NULL) 1378 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1379 1380 req_class = req & VM_ALLOC_CLASS_MASK; 1381 1382 /* 1383 * The page daemon is allowed to dig deeper into the free page list. 1384 */ 1385 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1386 req_class = VM_ALLOC_SYSTEM; 1387 1388 mtx_lock(&vm_page_queue_free_mtx); 1389 if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved || 1390 (req_class == VM_ALLOC_SYSTEM && 1391 cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) || 1392 (req_class == VM_ALLOC_INTERRUPT && 1393 cnt.v_free_count + cnt.v_cache_count > 0)) { 1394 /* 1395 * Allocate from the free queue if the number of free pages 1396 * exceeds the minimum for the request class. 1397 */ 1398 if (object != NULL && 1399 (m = vm_page_cache_lookup(object, pindex)) != NULL) { 1400 if ((req & VM_ALLOC_IFNOTCACHED) != 0) { 1401 mtx_unlock(&vm_page_queue_free_mtx); 1402 return (NULL); 1403 } 1404 if (vm_phys_unfree_page(m)) 1405 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0); 1406 #if VM_NRESERVLEVEL > 0 1407 else if (!vm_reserv_reactivate_page(m)) 1408 #else 1409 else 1410 #endif 1411 panic("vm_page_alloc: cache page %p is missing" 1412 " from the free queue", m); 1413 } else if ((req & VM_ALLOC_IFCACHED) != 0) { 1414 mtx_unlock(&vm_page_queue_free_mtx); 1415 return (NULL); 1416 #if VM_NRESERVLEVEL > 0 1417 } else if (object == NULL || (object->flags & (OBJ_COLORED | 1418 OBJ_FICTITIOUS)) != OBJ_COLORED || 1419 (m = vm_reserv_alloc_page(object, pindex)) == NULL) { 1420 #else 1421 } else { 1422 #endif 1423 m = vm_phys_alloc_pages(object != NULL ? 1424 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0); 1425 #if VM_NRESERVLEVEL > 0 1426 if (m == NULL && vm_reserv_reclaim_inactive()) { 1427 m = vm_phys_alloc_pages(object != NULL ? 1428 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 1429 0); 1430 } 1431 #endif 1432 } 1433 } else { 1434 /* 1435 * Not allocatable, give up. 1436 */ 1437 mtx_unlock(&vm_page_queue_free_mtx); 1438 atomic_add_int(&vm_pageout_deficit, 1439 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 1440 pagedaemon_wakeup(); 1441 return (NULL); 1442 } 1443 1444 /* 1445 * At this point we had better have found a good page. 1446 */ 1447 KASSERT(m != NULL, ("vm_page_alloc: missing page")); 1448 KASSERT(m->queue == PQ_NONE, 1449 ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue)); 1450 KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m)); 1451 KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m)); 1452 KASSERT(m->busy == 0, ("vm_page_alloc: page %p is busy", m)); 1453 KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m)); 1454 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 1455 ("vm_page_alloc: page %p has unexpected memattr %d", m, 1456 pmap_page_get_memattr(m))); 1457 if ((m->flags & PG_CACHED) != 0) { 1458 KASSERT((m->flags & PG_ZERO) == 0, 1459 ("vm_page_alloc: cached page %p is PG_ZERO", m)); 1460 KASSERT(m->valid != 0, 1461 ("vm_page_alloc: cached page %p is invalid", m)); 1462 if (m->object == object && m->pindex == pindex) 1463 cnt.v_reactivated++; 1464 else 1465 m->valid = 0; 1466 m_object = m->object; 1467 vm_page_cache_remove(m); 1468 if (m_object->type == OBJT_VNODE && m_object->cache == NULL) 1469 vp = m_object->handle; 1470 } else { 1471 KASSERT(VM_PAGE_IS_FREE(m), 1472 ("vm_page_alloc: page %p is not free", m)); 1473 KASSERT(m->valid == 0, 1474 ("vm_page_alloc: free page %p is valid", m)); 1475 cnt.v_free_count--; 1476 } 1477 1478 /* 1479 * Only the PG_ZERO flag is inherited. The PG_CACHED or PG_FREE flag 1480 * must be cleared before the free page queues lock is released. 1481 */ 1482 flags = 0; 1483 if (m->flags & PG_ZERO) { 1484 vm_page_zero_count--; 1485 if (req & VM_ALLOC_ZERO) 1486 flags = PG_ZERO; 1487 } 1488 if (req & VM_ALLOC_NODUMP) 1489 flags |= PG_NODUMP; 1490 m->flags = flags; 1491 mtx_unlock(&vm_page_queue_free_mtx); 1492 m->aflags = 0; 1493 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? 1494 VPO_UNMANAGED : 0; 1495 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ)) == 0) 1496 m->oflags |= VPO_BUSY; 1497 if (req & VM_ALLOC_WIRED) { 1498 /* 1499 * The page lock is not required for wiring a page until that 1500 * page is inserted into the object. 1501 */ 1502 atomic_add_int(&cnt.v_wire_count, 1); 1503 m->wire_count = 1; 1504 } 1505 m->act_count = 0; 1506 1507 if (object != NULL) { 1508 /* Ignore device objects; the pager sets "memattr" for them. */ 1509 if (object->memattr != VM_MEMATTR_DEFAULT && 1510 (object->flags & OBJ_FICTITIOUS) == 0) 1511 pmap_page_set_memattr(m, object->memattr); 1512 vm_page_insert(m, object, pindex); 1513 } else 1514 m->pindex = pindex; 1515 1516 /* 1517 * The following call to vdrop() must come after the above call 1518 * to vm_page_insert() in case both affect the same object and 1519 * vnode. Otherwise, the affected vnode's hold count could 1520 * temporarily become zero. 1521 */ 1522 if (vp != NULL) 1523 vdrop(vp); 1524 1525 /* 1526 * Don't wakeup too often - wakeup the pageout daemon when 1527 * we would be nearly out of memory. 1528 */ 1529 if (vm_paging_needed()) 1530 pagedaemon_wakeup(); 1531 1532 return (m); 1533 } 1534 1535 /* 1536 * vm_page_alloc_contig: 1537 * 1538 * Allocate a contiguous set of physical pages of the given size "npages" 1539 * from the free lists. All of the physical pages must be at or above 1540 * the given physical address "low" and below the given physical address 1541 * "high". The given value "alignment" determines the alignment of the 1542 * first physical page in the set. If the given value "boundary" is 1543 * non-zero, then the set of physical pages cannot cross any physical 1544 * address boundary that is a multiple of that value. Both "alignment" 1545 * and "boundary" must be a power of two. 1546 * 1547 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, 1548 * then the memory attribute setting for the physical pages is configured 1549 * to the object's memory attribute setting. Otherwise, the memory 1550 * attribute setting for the physical pages is configured to "memattr", 1551 * overriding the object's memory attribute setting. However, if the 1552 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the 1553 * memory attribute setting for the physical pages cannot be configured 1554 * to VM_MEMATTR_DEFAULT. 1555 * 1556 * The caller must always specify an allocation class. 1557 * 1558 * allocation classes: 1559 * VM_ALLOC_NORMAL normal process request 1560 * VM_ALLOC_SYSTEM system *really* needs a page 1561 * VM_ALLOC_INTERRUPT interrupt time request 1562 * 1563 * optional allocation flags: 1564 * VM_ALLOC_NOBUSY do not set the flag VPO_BUSY on the page 1565 * VM_ALLOC_NOOBJ page is not associated with an object and 1566 * should not have the flag VPO_BUSY set 1567 * VM_ALLOC_WIRED wire the allocated page 1568 * VM_ALLOC_ZERO prefer a zeroed page 1569 * 1570 * This routine may not sleep. 1571 */ 1572 vm_page_t 1573 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, 1574 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 1575 vm_paddr_t boundary, vm_memattr_t memattr) 1576 { 1577 struct vnode *drop; 1578 vm_page_t deferred_vdrop_list, m, m_ret; 1579 u_int flags, oflags; 1580 int req_class; 1581 1582 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0), 1583 ("vm_page_alloc_contig: inconsistent object/req")); 1584 if (object != NULL) { 1585 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1586 KASSERT(object->type == OBJT_PHYS, 1587 ("vm_page_alloc_contig: object %p isn't OBJT_PHYS", 1588 object)); 1589 } 1590 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 1591 req_class = req & VM_ALLOC_CLASS_MASK; 1592 1593 /* 1594 * The page daemon is allowed to dig deeper into the free page list. 1595 */ 1596 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1597 req_class = VM_ALLOC_SYSTEM; 1598 1599 deferred_vdrop_list = NULL; 1600 mtx_lock(&vm_page_queue_free_mtx); 1601 if (cnt.v_free_count + cnt.v_cache_count >= npages + 1602 cnt.v_free_reserved || (req_class == VM_ALLOC_SYSTEM && 1603 cnt.v_free_count + cnt.v_cache_count >= npages + 1604 cnt.v_interrupt_free_min) || (req_class == VM_ALLOC_INTERRUPT && 1605 cnt.v_free_count + cnt.v_cache_count >= npages)) { 1606 #if VM_NRESERVLEVEL > 0 1607 retry: 1608 if (object == NULL || (object->flags & OBJ_COLORED) == 0 || 1609 (m_ret = vm_reserv_alloc_contig(object, pindex, npages, 1610 low, high, alignment, boundary)) == NULL) 1611 #endif 1612 m_ret = vm_phys_alloc_contig(npages, low, high, 1613 alignment, boundary); 1614 } else { 1615 mtx_unlock(&vm_page_queue_free_mtx); 1616 atomic_add_int(&vm_pageout_deficit, npages); 1617 pagedaemon_wakeup(); 1618 return (NULL); 1619 } 1620 if (m_ret != NULL) 1621 for (m = m_ret; m < &m_ret[npages]; m++) { 1622 drop = vm_page_alloc_init(m); 1623 if (drop != NULL) { 1624 /* 1625 * Enqueue the vnode for deferred vdrop(). 1626 * 1627 * Once the pages are removed from the free 1628 * page list, "pageq" can be safely abused to 1629 * construct a short-lived list of vnodes. 1630 */ 1631 m->pageq.tqe_prev = (void *)drop; 1632 m->pageq.tqe_next = deferred_vdrop_list; 1633 deferred_vdrop_list = m; 1634 } 1635 } 1636 else { 1637 #if VM_NRESERVLEVEL > 0 1638 if (vm_reserv_reclaim_contig(npages, low, high, alignment, 1639 boundary)) 1640 goto retry; 1641 #endif 1642 } 1643 mtx_unlock(&vm_page_queue_free_mtx); 1644 if (m_ret == NULL) 1645 return (NULL); 1646 1647 /* 1648 * Initialize the pages. Only the PG_ZERO flag is inherited. 1649 */ 1650 flags = 0; 1651 if ((req & VM_ALLOC_ZERO) != 0) 1652 flags = PG_ZERO; 1653 if ((req & VM_ALLOC_NODUMP) != 0) 1654 flags |= PG_NODUMP; 1655 if ((req & VM_ALLOC_WIRED) != 0) 1656 atomic_add_int(&cnt.v_wire_count, npages); 1657 oflags = VPO_UNMANAGED; 1658 if (object != NULL) { 1659 if ((req & VM_ALLOC_NOBUSY) == 0) 1660 oflags |= VPO_BUSY; 1661 if (object->memattr != VM_MEMATTR_DEFAULT && 1662 memattr == VM_MEMATTR_DEFAULT) 1663 memattr = object->memattr; 1664 } 1665 for (m = m_ret; m < &m_ret[npages]; m++) { 1666 m->aflags = 0; 1667 m->flags = (m->flags | PG_NODUMP) & flags; 1668 if ((req & VM_ALLOC_WIRED) != 0) 1669 m->wire_count = 1; 1670 /* Unmanaged pages don't use "act_count". */ 1671 m->oflags = oflags; 1672 if (memattr != VM_MEMATTR_DEFAULT) 1673 pmap_page_set_memattr(m, memattr); 1674 if (object != NULL) 1675 vm_page_insert(m, object, pindex); 1676 else 1677 m->pindex = pindex; 1678 pindex++; 1679 } 1680 while (deferred_vdrop_list != NULL) { 1681 vdrop((struct vnode *)deferred_vdrop_list->pageq.tqe_prev); 1682 deferred_vdrop_list = deferred_vdrop_list->pageq.tqe_next; 1683 } 1684 if (vm_paging_needed()) 1685 pagedaemon_wakeup(); 1686 return (m_ret); 1687 } 1688 1689 /* 1690 * Initialize a page that has been freshly dequeued from a freelist. 1691 * The caller has to drop the vnode returned, if it is not NULL. 1692 * 1693 * This function may only be used to initialize unmanaged pages. 1694 * 1695 * To be called with vm_page_queue_free_mtx held. 1696 */ 1697 static struct vnode * 1698 vm_page_alloc_init(vm_page_t m) 1699 { 1700 struct vnode *drop; 1701 vm_object_t m_object; 1702 1703 KASSERT(m->queue == PQ_NONE, 1704 ("vm_page_alloc_init: page %p has unexpected queue %d", 1705 m, m->queue)); 1706 KASSERT(m->wire_count == 0, 1707 ("vm_page_alloc_init: page %p is wired", m)); 1708 KASSERT(m->hold_count == 0, 1709 ("vm_page_alloc_init: page %p is held", m)); 1710 KASSERT(m->busy == 0, 1711 ("vm_page_alloc_init: page %p is busy", m)); 1712 KASSERT(m->dirty == 0, 1713 ("vm_page_alloc_init: page %p is dirty", m)); 1714 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 1715 ("vm_page_alloc_init: page %p has unexpected memattr %d", 1716 m, pmap_page_get_memattr(m))); 1717 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1718 drop = NULL; 1719 if ((m->flags & PG_CACHED) != 0) { 1720 KASSERT((m->flags & PG_ZERO) == 0, 1721 ("vm_page_alloc_init: cached page %p is PG_ZERO", m)); 1722 m->valid = 0; 1723 m_object = m->object; 1724 vm_page_cache_remove(m); 1725 if (m_object->type == OBJT_VNODE && m_object->cache == NULL) 1726 drop = m_object->handle; 1727 } else { 1728 KASSERT(VM_PAGE_IS_FREE(m), 1729 ("vm_page_alloc_init: page %p is not free", m)); 1730 KASSERT(m->valid == 0, 1731 ("vm_page_alloc_init: free page %p is valid", m)); 1732 cnt.v_free_count--; 1733 if ((m->flags & PG_ZERO) != 0) 1734 vm_page_zero_count--; 1735 } 1736 /* Don't clear the PG_ZERO flag; we'll need it later. */ 1737 m->flags &= PG_ZERO; 1738 return (drop); 1739 } 1740 1741 /* 1742 * vm_page_alloc_freelist: 1743 * 1744 * Allocate a physical page from the specified free page list. 1745 * 1746 * The caller must always specify an allocation class. 1747 * 1748 * allocation classes: 1749 * VM_ALLOC_NORMAL normal process request 1750 * VM_ALLOC_SYSTEM system *really* needs a page 1751 * VM_ALLOC_INTERRUPT interrupt time request 1752 * 1753 * optional allocation flags: 1754 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1755 * intends to allocate 1756 * VM_ALLOC_WIRED wire the allocated page 1757 * VM_ALLOC_ZERO prefer a zeroed page 1758 * 1759 * This routine may not sleep. 1760 */ 1761 vm_page_t 1762 vm_page_alloc_freelist(int flind, int req) 1763 { 1764 struct vnode *drop; 1765 vm_page_t m; 1766 u_int flags; 1767 int req_class; 1768 1769 req_class = req & VM_ALLOC_CLASS_MASK; 1770 1771 /* 1772 * The page daemon is allowed to dig deeper into the free page list. 1773 */ 1774 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1775 req_class = VM_ALLOC_SYSTEM; 1776 1777 /* 1778 * Do not allocate reserved pages unless the req has asked for it. 1779 */ 1780 mtx_lock(&vm_page_queue_free_mtx); 1781 if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved || 1782 (req_class == VM_ALLOC_SYSTEM && 1783 cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) || 1784 (req_class == VM_ALLOC_INTERRUPT && 1785 cnt.v_free_count + cnt.v_cache_count > 0)) 1786 m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0); 1787 else { 1788 mtx_unlock(&vm_page_queue_free_mtx); 1789 atomic_add_int(&vm_pageout_deficit, 1790 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 1791 pagedaemon_wakeup(); 1792 return (NULL); 1793 } 1794 if (m == NULL) { 1795 mtx_unlock(&vm_page_queue_free_mtx); 1796 return (NULL); 1797 } 1798 drop = vm_page_alloc_init(m); 1799 mtx_unlock(&vm_page_queue_free_mtx); 1800 1801 /* 1802 * Initialize the page. Only the PG_ZERO flag is inherited. 1803 */ 1804 m->aflags = 0; 1805 flags = 0; 1806 if ((req & VM_ALLOC_ZERO) != 0) 1807 flags = PG_ZERO; 1808 m->flags &= flags; 1809 if ((req & VM_ALLOC_WIRED) != 0) { 1810 /* 1811 * The page lock is not required for wiring a page that does 1812 * not belong to an object. 1813 */ 1814 atomic_add_int(&cnt.v_wire_count, 1); 1815 m->wire_count = 1; 1816 } 1817 /* Unmanaged pages don't use "act_count". */ 1818 m->oflags = VPO_UNMANAGED; 1819 if (drop != NULL) 1820 vdrop(drop); 1821 if (vm_paging_needed()) 1822 pagedaemon_wakeup(); 1823 return (m); 1824 } 1825 1826 /* 1827 * vm_wait: (also see VM_WAIT macro) 1828 * 1829 * Sleep until free pages are available for allocation. 1830 * - Called in various places before memory allocations. 1831 */ 1832 void 1833 vm_wait(void) 1834 { 1835 1836 mtx_lock(&vm_page_queue_free_mtx); 1837 if (curproc == pageproc) { 1838 vm_pageout_pages_needed = 1; 1839 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx, 1840 PDROP | PSWP, "VMWait", 0); 1841 } else { 1842 if (!vm_pages_needed) { 1843 vm_pages_needed = 1; 1844 wakeup(&vm_pages_needed); 1845 } 1846 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM, 1847 "vmwait", 0); 1848 } 1849 } 1850 1851 /* 1852 * vm_waitpfault: (also see VM_WAITPFAULT macro) 1853 * 1854 * Sleep until free pages are available for allocation. 1855 * - Called only in vm_fault so that processes page faulting 1856 * can be easily tracked. 1857 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 1858 * processes will be able to grab memory first. Do not change 1859 * this balance without careful testing first. 1860 */ 1861 void 1862 vm_waitpfault(void) 1863 { 1864 1865 mtx_lock(&vm_page_queue_free_mtx); 1866 if (!vm_pages_needed) { 1867 vm_pages_needed = 1; 1868 wakeup(&vm_pages_needed); 1869 } 1870 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER, 1871 "pfault", 0); 1872 } 1873 1874 /* 1875 * vm_page_dequeue: 1876 * 1877 * Remove the given page from its current page queue. 1878 * 1879 * The page must be locked. 1880 */ 1881 void 1882 vm_page_dequeue(vm_page_t m) 1883 { 1884 struct vm_pagequeue *pq; 1885 1886 vm_page_lock_assert(m, MA_OWNED); 1887 KASSERT(m->queue != PQ_NONE, 1888 ("vm_page_dequeue: page %p is not queued", m)); 1889 pq = &vm_pagequeues[m->queue]; 1890 vm_pagequeue_lock(pq); 1891 m->queue = PQ_NONE; 1892 TAILQ_REMOVE(&pq->pq_pl, m, pageq); 1893 (*pq->pq_cnt)--; 1894 vm_pagequeue_unlock(pq); 1895 } 1896 1897 /* 1898 * vm_page_dequeue_locked: 1899 * 1900 * Remove the given page from its current page queue. 1901 * 1902 * The page and page queue must be locked. 1903 */ 1904 void 1905 vm_page_dequeue_locked(vm_page_t m) 1906 { 1907 struct vm_pagequeue *pq; 1908 1909 vm_page_lock_assert(m, MA_OWNED); 1910 pq = &vm_pagequeues[m->queue]; 1911 vm_pagequeue_assert_locked(pq); 1912 m->queue = PQ_NONE; 1913 TAILQ_REMOVE(&pq->pq_pl, m, pageq); 1914 (*pq->pq_cnt)--; 1915 } 1916 1917 /* 1918 * vm_page_enqueue: 1919 * 1920 * Add the given page to the specified page queue. 1921 * 1922 * The page must be locked. 1923 */ 1924 static void 1925 vm_page_enqueue(int queue, vm_page_t m) 1926 { 1927 struct vm_pagequeue *pq; 1928 1929 vm_page_lock_assert(m, MA_OWNED); 1930 pq = &vm_pagequeues[queue]; 1931 vm_pagequeue_lock(pq); 1932 m->queue = queue; 1933 TAILQ_INSERT_TAIL(&pq->pq_pl, m, pageq); 1934 ++*pq->pq_cnt; 1935 vm_pagequeue_unlock(pq); 1936 } 1937 1938 /* 1939 * vm_page_requeue: 1940 * 1941 * Move the given page to the tail of its current page queue. 1942 * 1943 * The page must be locked. 1944 */ 1945 void 1946 vm_page_requeue(vm_page_t m) 1947 { 1948 struct vm_pagequeue *pq; 1949 1950 vm_page_lock_assert(m, MA_OWNED); 1951 KASSERT(m->queue != PQ_NONE, 1952 ("vm_page_requeue: page %p is not queued", m)); 1953 pq = &vm_pagequeues[m->queue]; 1954 vm_pagequeue_lock(pq); 1955 TAILQ_REMOVE(&pq->pq_pl, m, pageq); 1956 TAILQ_INSERT_TAIL(&pq->pq_pl, m, pageq); 1957 vm_pagequeue_unlock(pq); 1958 } 1959 1960 /* 1961 * vm_page_requeue_locked: 1962 * 1963 * Move the given page to the tail of its current page queue. 1964 * 1965 * The page queue must be locked. 1966 */ 1967 void 1968 vm_page_requeue_locked(vm_page_t m) 1969 { 1970 struct vm_pagequeue *pq; 1971 1972 KASSERT(m->queue != PQ_NONE, 1973 ("vm_page_requeue_locked: page %p is not queued", m)); 1974 pq = &vm_pagequeues[m->queue]; 1975 vm_pagequeue_assert_locked(pq); 1976 TAILQ_REMOVE(&pq->pq_pl, m, pageq); 1977 TAILQ_INSERT_TAIL(&pq->pq_pl, m, pageq); 1978 } 1979 1980 /* 1981 * vm_page_activate: 1982 * 1983 * Put the specified page on the active list (if appropriate). 1984 * Ensure that act_count is at least ACT_INIT but do not otherwise 1985 * mess with it. 1986 * 1987 * The page must be locked. 1988 */ 1989 void 1990 vm_page_activate(vm_page_t m) 1991 { 1992 int queue; 1993 1994 vm_page_lock_assert(m, MA_OWNED); 1995 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1996 if ((queue = m->queue) != PQ_ACTIVE) { 1997 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 1998 if (m->act_count < ACT_INIT) 1999 m->act_count = ACT_INIT; 2000 if (queue != PQ_NONE) 2001 vm_page_dequeue(m); 2002 vm_page_enqueue(PQ_ACTIVE, m); 2003 } else 2004 KASSERT(queue == PQ_NONE, 2005 ("vm_page_activate: wired page %p is queued", m)); 2006 } else { 2007 if (m->act_count < ACT_INIT) 2008 m->act_count = ACT_INIT; 2009 } 2010 } 2011 2012 /* 2013 * vm_page_free_wakeup: 2014 * 2015 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 2016 * routine is called when a page has been added to the cache or free 2017 * queues. 2018 * 2019 * The page queues must be locked. 2020 */ 2021 static inline void 2022 vm_page_free_wakeup(void) 2023 { 2024 2025 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 2026 /* 2027 * if pageout daemon needs pages, then tell it that there are 2028 * some free. 2029 */ 2030 if (vm_pageout_pages_needed && 2031 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 2032 wakeup(&vm_pageout_pages_needed); 2033 vm_pageout_pages_needed = 0; 2034 } 2035 /* 2036 * wakeup processes that are waiting on memory if we hit a 2037 * high water mark. And wakeup scheduler process if we have 2038 * lots of memory. this process will swapin processes. 2039 */ 2040 if (vm_pages_needed && !vm_page_count_min()) { 2041 vm_pages_needed = 0; 2042 wakeup(&cnt.v_free_count); 2043 } 2044 } 2045 2046 /* 2047 * vm_page_free_toq: 2048 * 2049 * Returns the given page to the free list, 2050 * disassociating it with any VM object. 2051 * 2052 * The object must be locked. The page must be locked if it is managed. 2053 */ 2054 void 2055 vm_page_free_toq(vm_page_t m) 2056 { 2057 2058 if ((m->oflags & VPO_UNMANAGED) == 0) { 2059 vm_page_lock_assert(m, MA_OWNED); 2060 KASSERT(!pmap_page_is_mapped(m), 2061 ("vm_page_free_toq: freeing mapped page %p", m)); 2062 } else 2063 KASSERT(m->queue == PQ_NONE, 2064 ("vm_page_free_toq: unmanaged page %p is queued", m)); 2065 PCPU_INC(cnt.v_tfree); 2066 2067 if (VM_PAGE_IS_FREE(m)) 2068 panic("vm_page_free: freeing free page %p", m); 2069 else if (m->busy != 0) 2070 panic("vm_page_free: freeing busy page %p", m); 2071 2072 /* 2073 * Unqueue, then remove page. Note that we cannot destroy 2074 * the page here because we do not want to call the pager's 2075 * callback routine until after we've put the page on the 2076 * appropriate free queue. 2077 */ 2078 vm_page_remque(m); 2079 vm_page_remove(m); 2080 2081 /* 2082 * If fictitious remove object association and 2083 * return, otherwise delay object association removal. 2084 */ 2085 if ((m->flags & PG_FICTITIOUS) != 0) { 2086 return; 2087 } 2088 2089 m->valid = 0; 2090 vm_page_undirty(m); 2091 2092 if (m->wire_count != 0) 2093 panic("vm_page_free: freeing wired page %p", m); 2094 if (m->hold_count != 0) { 2095 m->flags &= ~PG_ZERO; 2096 KASSERT((m->flags & PG_UNHOLDFREE) == 0, 2097 ("vm_page_free: freeing PG_UNHOLDFREE page %p", m)); 2098 m->flags |= PG_UNHOLDFREE; 2099 } else { 2100 /* 2101 * Restore the default memory attribute to the page. 2102 */ 2103 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 2104 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 2105 2106 /* 2107 * Insert the page into the physical memory allocator's 2108 * cache/free page queues. 2109 */ 2110 mtx_lock(&vm_page_queue_free_mtx); 2111 m->flags |= PG_FREE; 2112 cnt.v_free_count++; 2113 #if VM_NRESERVLEVEL > 0 2114 if (!vm_reserv_free_page(m)) 2115 #else 2116 if (TRUE) 2117 #endif 2118 vm_phys_free_pages(m, 0); 2119 if ((m->flags & PG_ZERO) != 0) 2120 ++vm_page_zero_count; 2121 else 2122 vm_page_zero_idle_wakeup(); 2123 vm_page_free_wakeup(); 2124 mtx_unlock(&vm_page_queue_free_mtx); 2125 } 2126 } 2127 2128 /* 2129 * vm_page_wire: 2130 * 2131 * Mark this page as wired down by yet 2132 * another map, removing it from paging queues 2133 * as necessary. 2134 * 2135 * If the page is fictitious, then its wire count must remain one. 2136 * 2137 * The page must be locked. 2138 */ 2139 void 2140 vm_page_wire(vm_page_t m) 2141 { 2142 2143 /* 2144 * Only bump the wire statistics if the page is not already wired, 2145 * and only unqueue the page if it is on some queue (if it is unmanaged 2146 * it is already off the queues). 2147 */ 2148 vm_page_lock_assert(m, MA_OWNED); 2149 if ((m->flags & PG_FICTITIOUS) != 0) { 2150 KASSERT(m->wire_count == 1, 2151 ("vm_page_wire: fictitious page %p's wire count isn't one", 2152 m)); 2153 return; 2154 } 2155 if (m->wire_count == 0) { 2156 KASSERT((m->oflags & VPO_UNMANAGED) == 0 || 2157 m->queue == PQ_NONE, 2158 ("vm_page_wire: unmanaged page %p is queued", m)); 2159 vm_page_remque(m); 2160 atomic_add_int(&cnt.v_wire_count, 1); 2161 } 2162 m->wire_count++; 2163 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 2164 } 2165 2166 /* 2167 * vm_page_unwire: 2168 * 2169 * Release one wiring of the specified page, potentially enabling it to be 2170 * paged again. If paging is enabled, then the value of the parameter 2171 * "activate" determines to which queue the page is added. If "activate" is 2172 * non-zero, then the page is added to the active queue. Otherwise, it is 2173 * added to the inactive queue. 2174 * 2175 * However, unless the page belongs to an object, it is not enqueued because 2176 * it cannot be paged out. 2177 * 2178 * If a page is fictitious, then its wire count must alway be one. 2179 * 2180 * A managed page must be locked. 2181 */ 2182 void 2183 vm_page_unwire(vm_page_t m, int activate) 2184 { 2185 2186 if ((m->oflags & VPO_UNMANAGED) == 0) 2187 vm_page_lock_assert(m, MA_OWNED); 2188 if ((m->flags & PG_FICTITIOUS) != 0) { 2189 KASSERT(m->wire_count == 1, 2190 ("vm_page_unwire: fictitious page %p's wire count isn't one", m)); 2191 return; 2192 } 2193 if (m->wire_count > 0) { 2194 m->wire_count--; 2195 if (m->wire_count == 0) { 2196 atomic_subtract_int(&cnt.v_wire_count, 1); 2197 if ((m->oflags & VPO_UNMANAGED) != 0 || 2198 m->object == NULL) 2199 return; 2200 if (!activate) 2201 m->flags &= ~PG_WINATCFLS; 2202 vm_page_enqueue(activate ? PQ_ACTIVE : PQ_INACTIVE, m); 2203 } 2204 } else 2205 panic("vm_page_unwire: page %p's wire count is zero", m); 2206 } 2207 2208 /* 2209 * Move the specified page to the inactive queue. 2210 * 2211 * Many pages placed on the inactive queue should actually go 2212 * into the cache, but it is difficult to figure out which. What 2213 * we do instead, if the inactive target is well met, is to put 2214 * clean pages at the head of the inactive queue instead of the tail. 2215 * This will cause them to be moved to the cache more quickly and 2216 * if not actively re-referenced, reclaimed more quickly. If we just 2217 * stick these pages at the end of the inactive queue, heavy filesystem 2218 * meta-data accesses can cause an unnecessary paging load on memory bound 2219 * processes. This optimization causes one-time-use metadata to be 2220 * reused more quickly. 2221 * 2222 * Normally athead is 0 resulting in LRU operation. athead is set 2223 * to 1 if we want this page to be 'as if it were placed in the cache', 2224 * except without unmapping it from the process address space. 2225 * 2226 * The page must be locked. 2227 */ 2228 static inline void 2229 _vm_page_deactivate(vm_page_t m, int athead) 2230 { 2231 struct vm_pagequeue *pq; 2232 int queue; 2233 2234 vm_page_lock_assert(m, MA_OWNED); 2235 2236 /* 2237 * Ignore if already inactive. 2238 */ 2239 if ((queue = m->queue) == PQ_INACTIVE) 2240 return; 2241 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 2242 if (queue != PQ_NONE) 2243 vm_page_dequeue(m); 2244 m->flags &= ~PG_WINATCFLS; 2245 pq = &vm_pagequeues[PQ_INACTIVE]; 2246 vm_pagequeue_lock(pq); 2247 m->queue = PQ_INACTIVE; 2248 if (athead) 2249 TAILQ_INSERT_HEAD(&pq->pq_pl, m, pageq); 2250 else 2251 TAILQ_INSERT_TAIL(&pq->pq_pl, m, pageq); 2252 cnt.v_inactive_count++; 2253 vm_pagequeue_unlock(pq); 2254 } 2255 } 2256 2257 /* 2258 * Move the specified page to the inactive queue. 2259 * 2260 * The page must be locked. 2261 */ 2262 void 2263 vm_page_deactivate(vm_page_t m) 2264 { 2265 2266 _vm_page_deactivate(m, 0); 2267 } 2268 2269 /* 2270 * vm_page_try_to_cache: 2271 * 2272 * Returns 0 on failure, 1 on success 2273 */ 2274 int 2275 vm_page_try_to_cache(vm_page_t m) 2276 { 2277 2278 vm_page_lock_assert(m, MA_OWNED); 2279 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2280 if (m->dirty || m->hold_count || m->busy || m->wire_count || 2281 (m->oflags & (VPO_BUSY | VPO_UNMANAGED)) != 0) 2282 return (0); 2283 pmap_remove_all(m); 2284 if (m->dirty) 2285 return (0); 2286 vm_page_cache(m); 2287 return (1); 2288 } 2289 2290 /* 2291 * vm_page_try_to_free() 2292 * 2293 * Attempt to free the page. If we cannot free it, we do nothing. 2294 * 1 is returned on success, 0 on failure. 2295 */ 2296 int 2297 vm_page_try_to_free(vm_page_t m) 2298 { 2299 2300 vm_page_lock_assert(m, MA_OWNED); 2301 if (m->object != NULL) 2302 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2303 if (m->dirty || m->hold_count || m->busy || m->wire_count || 2304 (m->oflags & (VPO_BUSY | VPO_UNMANAGED)) != 0) 2305 return (0); 2306 pmap_remove_all(m); 2307 if (m->dirty) 2308 return (0); 2309 vm_page_free(m); 2310 return (1); 2311 } 2312 2313 /* 2314 * vm_page_cache 2315 * 2316 * Put the specified page onto the page cache queue (if appropriate). 2317 * 2318 * The object and page must be locked. 2319 */ 2320 void 2321 vm_page_cache(vm_page_t m) 2322 { 2323 vm_object_t object; 2324 vm_page_t next, prev, root; 2325 2326 vm_page_lock_assert(m, MA_OWNED); 2327 object = m->object; 2328 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2329 if ((m->oflags & (VPO_UNMANAGED | VPO_BUSY)) || m->busy || 2330 m->hold_count || m->wire_count) 2331 panic("vm_page_cache: attempting to cache busy page"); 2332 KASSERT(!pmap_page_is_mapped(m), 2333 ("vm_page_cache: page %p is mapped", m)); 2334 KASSERT(m->dirty == 0, ("vm_page_cache: page %p is dirty", m)); 2335 if (m->valid == 0 || object->type == OBJT_DEFAULT || 2336 (object->type == OBJT_SWAP && 2337 !vm_pager_has_page(object, m->pindex, NULL, NULL))) { 2338 /* 2339 * Hypothesis: A cache-elgible page belonging to a 2340 * default object or swap object but without a backing 2341 * store must be zero filled. 2342 */ 2343 vm_page_free(m); 2344 return; 2345 } 2346 KASSERT((m->flags & PG_CACHED) == 0, 2347 ("vm_page_cache: page %p is already cached", m)); 2348 PCPU_INC(cnt.v_tcached); 2349 2350 /* 2351 * Remove the page from the paging queues. 2352 */ 2353 vm_page_remque(m); 2354 2355 /* 2356 * Remove the page from the object's collection of resident 2357 * pages. 2358 */ 2359 if ((next = TAILQ_NEXT(m, listq)) != NULL && next->left == m) { 2360 /* 2361 * Since the page's successor in the list is also its parent 2362 * in the tree, its right subtree must be empty. 2363 */ 2364 next->left = m->left; 2365 KASSERT(m->right == NULL, 2366 ("vm_page_cache: page %p has right child", m)); 2367 } else if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL && 2368 prev->right == m) { 2369 /* 2370 * Since the page's predecessor in the list is also its parent 2371 * in the tree, its left subtree must be empty. 2372 */ 2373 KASSERT(m->left == NULL, 2374 ("vm_page_cache: page %p has left child", m)); 2375 prev->right = m->right; 2376 } else { 2377 if (m != object->root) 2378 vm_page_splay(m->pindex, object->root); 2379 if (m->left == NULL) 2380 root = m->right; 2381 else if (m->right == NULL) 2382 root = m->left; 2383 else { 2384 /* 2385 * Move the page's successor to the root, because 2386 * pages are usually removed in ascending order. 2387 */ 2388 if (m->right != next) 2389 vm_page_splay(m->pindex, m->right); 2390 next->left = m->left; 2391 root = next; 2392 } 2393 object->root = root; 2394 } 2395 TAILQ_REMOVE(&object->memq, m, listq); 2396 object->resident_page_count--; 2397 2398 /* 2399 * Restore the default memory attribute to the page. 2400 */ 2401 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 2402 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 2403 2404 /* 2405 * Insert the page into the object's collection of cached pages 2406 * and the physical memory allocator's cache/free page queues. 2407 */ 2408 m->flags &= ~PG_ZERO; 2409 mtx_lock(&vm_page_queue_free_mtx); 2410 m->flags |= PG_CACHED; 2411 cnt.v_cache_count++; 2412 root = object->cache; 2413 if (root == NULL) { 2414 m->left = NULL; 2415 m->right = NULL; 2416 } else { 2417 root = vm_page_splay(m->pindex, root); 2418 if (m->pindex < root->pindex) { 2419 m->left = root->left; 2420 m->right = root; 2421 root->left = NULL; 2422 } else if (__predict_false(m->pindex == root->pindex)) 2423 panic("vm_page_cache: offset already cached"); 2424 else { 2425 m->right = root->right; 2426 m->left = root; 2427 root->right = NULL; 2428 } 2429 } 2430 object->cache = m; 2431 #if VM_NRESERVLEVEL > 0 2432 if (!vm_reserv_free_page(m)) { 2433 #else 2434 if (TRUE) { 2435 #endif 2436 vm_phys_set_pool(VM_FREEPOOL_CACHE, m, 0); 2437 vm_phys_free_pages(m, 0); 2438 } 2439 vm_page_free_wakeup(); 2440 mtx_unlock(&vm_page_queue_free_mtx); 2441 2442 /* 2443 * Increment the vnode's hold count if this is the object's only 2444 * cached page. Decrement the vnode's hold count if this was 2445 * the object's only resident page. 2446 */ 2447 if (object->type == OBJT_VNODE) { 2448 if (root == NULL && object->resident_page_count != 0) 2449 vhold(object->handle); 2450 else if (root != NULL && object->resident_page_count == 0) 2451 vdrop(object->handle); 2452 } 2453 } 2454 2455 /* 2456 * vm_page_dontneed 2457 * 2458 * Cache, deactivate, or do nothing as appropriate. This routine 2459 * is typically used by madvise() MADV_DONTNEED. 2460 * 2461 * Generally speaking we want to move the page into the cache so 2462 * it gets reused quickly. However, this can result in a silly syndrome 2463 * due to the page recycling too quickly. Small objects will not be 2464 * fully cached. On the otherhand, if we move the page to the inactive 2465 * queue we wind up with a problem whereby very large objects 2466 * unnecessarily blow away our inactive and cache queues. 2467 * 2468 * The solution is to move the pages based on a fixed weighting. We 2469 * either leave them alone, deactivate them, or move them to the cache, 2470 * where moving them to the cache has the highest weighting. 2471 * By forcing some pages into other queues we eventually force the 2472 * system to balance the queues, potentially recovering other unrelated 2473 * space from active. The idea is to not force this to happen too 2474 * often. 2475 * 2476 * The object and page must be locked. 2477 */ 2478 void 2479 vm_page_dontneed(vm_page_t m) 2480 { 2481 int dnw; 2482 int head; 2483 2484 vm_page_lock_assert(m, MA_OWNED); 2485 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2486 dnw = PCPU_GET(dnweight); 2487 PCPU_INC(dnweight); 2488 2489 /* 2490 * Occasionally leave the page alone. 2491 */ 2492 if ((dnw & 0x01F0) == 0 || m->queue == PQ_INACTIVE) { 2493 if (m->act_count >= ACT_INIT) 2494 --m->act_count; 2495 return; 2496 } 2497 2498 /* 2499 * Clear any references to the page. Otherwise, the page daemon will 2500 * immediately reactivate the page. 2501 * 2502 * Perform the pmap_clear_reference() first. Otherwise, a concurrent 2503 * pmap operation, such as pmap_remove(), could clear a reference in 2504 * the pmap and set PGA_REFERENCED on the page before the 2505 * pmap_clear_reference() had completed. Consequently, the page would 2506 * appear referenced based upon an old reference that occurred before 2507 * this function ran. 2508 */ 2509 pmap_clear_reference(m); 2510 vm_page_aflag_clear(m, PGA_REFERENCED); 2511 2512 if (m->dirty == 0 && pmap_is_modified(m)) 2513 vm_page_dirty(m); 2514 2515 if (m->dirty || (dnw & 0x0070) == 0) { 2516 /* 2517 * Deactivate the page 3 times out of 32. 2518 */ 2519 head = 0; 2520 } else { 2521 /* 2522 * Cache the page 28 times out of every 32. Note that 2523 * the page is deactivated instead of cached, but placed 2524 * at the head of the queue instead of the tail. 2525 */ 2526 head = 1; 2527 } 2528 _vm_page_deactivate(m, head); 2529 } 2530 2531 /* 2532 * Grab a page, waiting until we are waken up due to the page 2533 * changing state. We keep on waiting, if the page continues 2534 * to be in the object. If the page doesn't exist, first allocate it 2535 * and then conditionally zero it. 2536 * 2537 * The caller must always specify the VM_ALLOC_RETRY flag. This is intended 2538 * to facilitate its eventual removal. 2539 * 2540 * This routine may sleep. 2541 * 2542 * The object must be locked on entry. The lock will, however, be released 2543 * and reacquired if the routine sleeps. 2544 */ 2545 vm_page_t 2546 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 2547 { 2548 vm_page_t m; 2549 2550 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2551 KASSERT((allocflags & VM_ALLOC_RETRY) != 0, 2552 ("vm_page_grab: VM_ALLOC_RETRY is required")); 2553 retrylookup: 2554 if ((m = vm_page_lookup(object, pindex)) != NULL) { 2555 if ((m->oflags & VPO_BUSY) != 0 || 2556 ((allocflags & VM_ALLOC_IGN_SBUSY) == 0 && m->busy != 0)) { 2557 /* 2558 * Reference the page before unlocking and 2559 * sleeping so that the page daemon is less 2560 * likely to reclaim it. 2561 */ 2562 vm_page_aflag_set(m, PGA_REFERENCED); 2563 vm_page_sleep(m, "pgrbwt"); 2564 goto retrylookup; 2565 } else { 2566 if ((allocflags & VM_ALLOC_WIRED) != 0) { 2567 vm_page_lock(m); 2568 vm_page_wire(m); 2569 vm_page_unlock(m); 2570 } 2571 if ((allocflags & VM_ALLOC_NOBUSY) == 0) 2572 vm_page_busy(m); 2573 return (m); 2574 } 2575 } 2576 m = vm_page_alloc(object, pindex, allocflags & ~(VM_ALLOC_RETRY | 2577 VM_ALLOC_IGN_SBUSY)); 2578 if (m == NULL) { 2579 VM_OBJECT_UNLOCK(object); 2580 VM_WAIT; 2581 VM_OBJECT_LOCK(object); 2582 goto retrylookup; 2583 } else if (m->valid != 0) 2584 return (m); 2585 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 2586 pmap_zero_page(m); 2587 return (m); 2588 } 2589 2590 /* 2591 * Mapping function for valid or dirty bits in a page. 2592 * 2593 * Inputs are required to range within a page. 2594 */ 2595 vm_page_bits_t 2596 vm_page_bits(int base, int size) 2597 { 2598 int first_bit; 2599 int last_bit; 2600 2601 KASSERT( 2602 base + size <= PAGE_SIZE, 2603 ("vm_page_bits: illegal base/size %d/%d", base, size) 2604 ); 2605 2606 if (size == 0) /* handle degenerate case */ 2607 return (0); 2608 2609 first_bit = base >> DEV_BSHIFT; 2610 last_bit = (base + size - 1) >> DEV_BSHIFT; 2611 2612 return (((vm_page_bits_t)2 << last_bit) - 2613 ((vm_page_bits_t)1 << first_bit)); 2614 } 2615 2616 /* 2617 * vm_page_set_valid_range: 2618 * 2619 * Sets portions of a page valid. The arguments are expected 2620 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2621 * of any partial chunks touched by the range. The invalid portion of 2622 * such chunks will be zeroed. 2623 * 2624 * (base + size) must be less then or equal to PAGE_SIZE. 2625 */ 2626 void 2627 vm_page_set_valid_range(vm_page_t m, int base, int size) 2628 { 2629 int endoff, frag; 2630 2631 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2632 if (size == 0) /* handle degenerate case */ 2633 return; 2634 2635 /* 2636 * If the base is not DEV_BSIZE aligned and the valid 2637 * bit is clear, we have to zero out a portion of the 2638 * first block. 2639 */ 2640 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 2641 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 2642 pmap_zero_page_area(m, frag, base - frag); 2643 2644 /* 2645 * If the ending offset is not DEV_BSIZE aligned and the 2646 * valid bit is clear, we have to zero out a portion of 2647 * the last block. 2648 */ 2649 endoff = base + size; 2650 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 2651 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 2652 pmap_zero_page_area(m, endoff, 2653 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 2654 2655 /* 2656 * Assert that no previously invalid block that is now being validated 2657 * is already dirty. 2658 */ 2659 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 2660 ("vm_page_set_valid_range: page %p is dirty", m)); 2661 2662 /* 2663 * Set valid bits inclusive of any overlap. 2664 */ 2665 m->valid |= vm_page_bits(base, size); 2666 } 2667 2668 /* 2669 * Clear the given bits from the specified page's dirty field. 2670 */ 2671 static __inline void 2672 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) 2673 { 2674 uintptr_t addr; 2675 #if PAGE_SIZE < 16384 2676 int shift; 2677 #endif 2678 2679 /* 2680 * If the object is locked and the page is neither VPO_BUSY nor 2681 * write mapped, then the page's dirty field cannot possibly be 2682 * set by a concurrent pmap operation. 2683 */ 2684 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2685 if ((m->oflags & VPO_BUSY) == 0 && !pmap_page_is_write_mapped(m)) 2686 m->dirty &= ~pagebits; 2687 else { 2688 /* 2689 * The pmap layer can call vm_page_dirty() without 2690 * holding a distinguished lock. The combination of 2691 * the object's lock and an atomic operation suffice 2692 * to guarantee consistency of the page dirty field. 2693 * 2694 * For PAGE_SIZE == 32768 case, compiler already 2695 * properly aligns the dirty field, so no forcible 2696 * alignment is needed. Only require existence of 2697 * atomic_clear_64 when page size is 32768. 2698 */ 2699 addr = (uintptr_t)&m->dirty; 2700 #if PAGE_SIZE == 32768 2701 atomic_clear_64((uint64_t *)addr, pagebits); 2702 #elif PAGE_SIZE == 16384 2703 atomic_clear_32((uint32_t *)addr, pagebits); 2704 #else /* PAGE_SIZE <= 8192 */ 2705 /* 2706 * Use a trick to perform a 32-bit atomic on the 2707 * containing aligned word, to not depend on the existence 2708 * of atomic_clear_{8, 16}. 2709 */ 2710 shift = addr & (sizeof(uint32_t) - 1); 2711 #if BYTE_ORDER == BIG_ENDIAN 2712 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY; 2713 #else 2714 shift *= NBBY; 2715 #endif 2716 addr &= ~(sizeof(uint32_t) - 1); 2717 atomic_clear_32((uint32_t *)addr, pagebits << shift); 2718 #endif /* PAGE_SIZE */ 2719 } 2720 } 2721 2722 /* 2723 * vm_page_set_validclean: 2724 * 2725 * Sets portions of a page valid and clean. The arguments are expected 2726 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2727 * of any partial chunks touched by the range. The invalid portion of 2728 * such chunks will be zero'd. 2729 * 2730 * (base + size) must be less then or equal to PAGE_SIZE. 2731 */ 2732 void 2733 vm_page_set_validclean(vm_page_t m, int base, int size) 2734 { 2735 vm_page_bits_t oldvalid, pagebits; 2736 int endoff, frag; 2737 2738 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2739 if (size == 0) /* handle degenerate case */ 2740 return; 2741 2742 /* 2743 * If the base is not DEV_BSIZE aligned and the valid 2744 * bit is clear, we have to zero out a portion of the 2745 * first block. 2746 */ 2747 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 2748 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) 2749 pmap_zero_page_area(m, frag, base - frag); 2750 2751 /* 2752 * If the ending offset is not DEV_BSIZE aligned and the 2753 * valid bit is clear, we have to zero out a portion of 2754 * the last block. 2755 */ 2756 endoff = base + size; 2757 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 2758 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) 2759 pmap_zero_page_area(m, endoff, 2760 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 2761 2762 /* 2763 * Set valid, clear dirty bits. If validating the entire 2764 * page we can safely clear the pmap modify bit. We also 2765 * use this opportunity to clear the VPO_NOSYNC flag. If a process 2766 * takes a write fault on a MAP_NOSYNC memory area the flag will 2767 * be set again. 2768 * 2769 * We set valid bits inclusive of any overlap, but we can only 2770 * clear dirty bits for DEV_BSIZE chunks that are fully within 2771 * the range. 2772 */ 2773 oldvalid = m->valid; 2774 pagebits = vm_page_bits(base, size); 2775 m->valid |= pagebits; 2776 #if 0 /* NOT YET */ 2777 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 2778 frag = DEV_BSIZE - frag; 2779 base += frag; 2780 size -= frag; 2781 if (size < 0) 2782 size = 0; 2783 } 2784 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 2785 #endif 2786 if (base == 0 && size == PAGE_SIZE) { 2787 /* 2788 * The page can only be modified within the pmap if it is 2789 * mapped, and it can only be mapped if it was previously 2790 * fully valid. 2791 */ 2792 if (oldvalid == VM_PAGE_BITS_ALL) 2793 /* 2794 * Perform the pmap_clear_modify() first. Otherwise, 2795 * a concurrent pmap operation, such as 2796 * pmap_protect(), could clear a modification in the 2797 * pmap and set the dirty field on the page before 2798 * pmap_clear_modify() had begun and after the dirty 2799 * field was cleared here. 2800 */ 2801 pmap_clear_modify(m); 2802 m->dirty = 0; 2803 m->oflags &= ~VPO_NOSYNC; 2804 } else if (oldvalid != VM_PAGE_BITS_ALL) 2805 m->dirty &= ~pagebits; 2806 else 2807 vm_page_clear_dirty_mask(m, pagebits); 2808 } 2809 2810 void 2811 vm_page_clear_dirty(vm_page_t m, int base, int size) 2812 { 2813 2814 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 2815 } 2816 2817 /* 2818 * vm_page_set_invalid: 2819 * 2820 * Invalidates DEV_BSIZE'd chunks within a page. Both the 2821 * valid and dirty bits for the effected areas are cleared. 2822 */ 2823 void 2824 vm_page_set_invalid(vm_page_t m, int base, int size) 2825 { 2826 vm_page_bits_t bits; 2827 2828 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2829 KASSERT((m->oflags & VPO_BUSY) == 0, 2830 ("vm_page_set_invalid: page %p is busy", m)); 2831 bits = vm_page_bits(base, size); 2832 if (m->valid == VM_PAGE_BITS_ALL && bits != 0) 2833 pmap_remove_all(m); 2834 KASSERT(!pmap_page_is_mapped(m), 2835 ("vm_page_set_invalid: page %p is mapped", m)); 2836 m->valid &= ~bits; 2837 m->dirty &= ~bits; 2838 } 2839 2840 /* 2841 * vm_page_zero_invalid() 2842 * 2843 * The kernel assumes that the invalid portions of a page contain 2844 * garbage, but such pages can be mapped into memory by user code. 2845 * When this occurs, we must zero out the non-valid portions of the 2846 * page so user code sees what it expects. 2847 * 2848 * Pages are most often semi-valid when the end of a file is mapped 2849 * into memory and the file's size is not page aligned. 2850 */ 2851 void 2852 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 2853 { 2854 int b; 2855 int i; 2856 2857 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2858 /* 2859 * Scan the valid bits looking for invalid sections that 2860 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 2861 * valid bit may be set ) have already been zerod by 2862 * vm_page_set_validclean(). 2863 */ 2864 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 2865 if (i == (PAGE_SIZE / DEV_BSIZE) || 2866 (m->valid & ((vm_page_bits_t)1 << i))) { 2867 if (i > b) { 2868 pmap_zero_page_area(m, 2869 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 2870 } 2871 b = i + 1; 2872 } 2873 } 2874 2875 /* 2876 * setvalid is TRUE when we can safely set the zero'd areas 2877 * as being valid. We can do this if there are no cache consistancy 2878 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 2879 */ 2880 if (setvalid) 2881 m->valid = VM_PAGE_BITS_ALL; 2882 } 2883 2884 /* 2885 * vm_page_is_valid: 2886 * 2887 * Is (partial) page valid? Note that the case where size == 0 2888 * will return FALSE in the degenerate case where the page is 2889 * entirely invalid, and TRUE otherwise. 2890 */ 2891 int 2892 vm_page_is_valid(vm_page_t m, int base, int size) 2893 { 2894 vm_page_bits_t bits; 2895 2896 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2897 bits = vm_page_bits(base, size); 2898 if (m->valid && ((m->valid & bits) == bits)) 2899 return 1; 2900 else 2901 return 0; 2902 } 2903 2904 /* 2905 * Set the page's dirty bits if the page is modified. 2906 */ 2907 void 2908 vm_page_test_dirty(vm_page_t m) 2909 { 2910 2911 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2912 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 2913 vm_page_dirty(m); 2914 } 2915 2916 void 2917 vm_page_lock_KBI(vm_page_t m, const char *file, int line) 2918 { 2919 2920 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); 2921 } 2922 2923 void 2924 vm_page_unlock_KBI(vm_page_t m, const char *file, int line) 2925 { 2926 2927 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); 2928 } 2929 2930 int 2931 vm_page_trylock_KBI(vm_page_t m, const char *file, int line) 2932 { 2933 2934 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); 2935 } 2936 2937 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) 2938 void 2939 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) 2940 { 2941 2942 mtx_assert_(vm_page_lockptr(m), a, file, line); 2943 } 2944 #endif 2945 2946 int so_zerocp_fullpage = 0; 2947 2948 /* 2949 * Replace the given page with a copy. The copied page assumes 2950 * the portion of the given page's "wire_count" that is not the 2951 * responsibility of this copy-on-write mechanism. 2952 * 2953 * The object containing the given page must have a non-zero 2954 * paging-in-progress count and be locked. 2955 */ 2956 void 2957 vm_page_cowfault(vm_page_t m) 2958 { 2959 vm_page_t mnew; 2960 vm_object_t object; 2961 vm_pindex_t pindex; 2962 2963 vm_page_lock_assert(m, MA_OWNED); 2964 object = m->object; 2965 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2966 KASSERT(object->paging_in_progress != 0, 2967 ("vm_page_cowfault: object %p's paging-in-progress count is zero.", 2968 object)); 2969 pindex = m->pindex; 2970 2971 retry_alloc: 2972 pmap_remove_all(m); 2973 vm_page_remove(m); 2974 mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY); 2975 if (mnew == NULL) { 2976 vm_page_insert(m, object, pindex); 2977 vm_page_unlock(m); 2978 VM_OBJECT_UNLOCK(object); 2979 VM_WAIT; 2980 VM_OBJECT_LOCK(object); 2981 if (m == vm_page_lookup(object, pindex)) { 2982 vm_page_lock(m); 2983 goto retry_alloc; 2984 } else { 2985 /* 2986 * Page disappeared during the wait. 2987 */ 2988 return; 2989 } 2990 } 2991 2992 if (m->cow == 0) { 2993 /* 2994 * check to see if we raced with an xmit complete when 2995 * waiting to allocate a page. If so, put things back 2996 * the way they were 2997 */ 2998 vm_page_unlock(m); 2999 vm_page_lock(mnew); 3000 vm_page_free(mnew); 3001 vm_page_unlock(mnew); 3002 vm_page_insert(m, object, pindex); 3003 } else { /* clear COW & copy page */ 3004 if (!so_zerocp_fullpage) 3005 pmap_copy_page(m, mnew); 3006 mnew->valid = VM_PAGE_BITS_ALL; 3007 vm_page_dirty(mnew); 3008 mnew->wire_count = m->wire_count - m->cow; 3009 m->wire_count = m->cow; 3010 vm_page_unlock(m); 3011 } 3012 } 3013 3014 void 3015 vm_page_cowclear(vm_page_t m) 3016 { 3017 3018 vm_page_lock_assert(m, MA_OWNED); 3019 if (m->cow) { 3020 m->cow--; 3021 /* 3022 * let vm_fault add back write permission lazily 3023 */ 3024 } 3025 /* 3026 * sf_buf_free() will free the page, so we needn't do it here 3027 */ 3028 } 3029 3030 int 3031 vm_page_cowsetup(vm_page_t m) 3032 { 3033 3034 vm_page_lock_assert(m, MA_OWNED); 3035 if ((m->flags & PG_FICTITIOUS) != 0 || 3036 (m->oflags & VPO_UNMANAGED) != 0 || 3037 m->cow == USHRT_MAX - 1 || !VM_OBJECT_TRYLOCK(m->object)) 3038 return (EBUSY); 3039 m->cow++; 3040 pmap_remove_write(m); 3041 VM_OBJECT_UNLOCK(m->object); 3042 return (0); 3043 } 3044 3045 #ifdef INVARIANTS 3046 void 3047 vm_page_object_lock_assert(vm_page_t m) 3048 { 3049 3050 /* 3051 * Certain of the page's fields may only be modified by the 3052 * holder of the containing object's lock or the setter of the 3053 * page's VPO_BUSY flag. Unfortunately, the setter of the 3054 * VPO_BUSY flag is not recorded, and thus cannot be checked 3055 * here. 3056 */ 3057 if (m->object != NULL && (m->oflags & VPO_BUSY) == 0) 3058 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 3059 } 3060 #endif 3061 3062 #include "opt_ddb.h" 3063 #ifdef DDB 3064 #include <sys/kernel.h> 3065 3066 #include <ddb/ddb.h> 3067 3068 DB_SHOW_COMMAND(page, vm_page_print_page_info) 3069 { 3070 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 3071 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 3072 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 3073 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 3074 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 3075 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 3076 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 3077 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 3078 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 3079 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 3080 } 3081 3082 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 3083 { 3084 3085 db_printf("PQ_FREE:"); 3086 db_printf(" %d", cnt.v_free_count); 3087 db_printf("\n"); 3088 3089 db_printf("PQ_CACHE:"); 3090 db_printf(" %d", cnt.v_cache_count); 3091 db_printf("\n"); 3092 3093 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 3094 *vm_pagequeues[PQ_ACTIVE].pq_cnt, 3095 *vm_pagequeues[PQ_INACTIVE].pq_cnt); 3096 } 3097 #endif /* DDB */ 3098