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->type == OBJT_DEVICE || 1418 object->type == OBJT_SG || 1419 (object->flags & OBJ_COLORED) == 0 || 1420 (m = vm_reserv_alloc_page(object, pindex)) == NULL) { 1421 #else 1422 } else { 1423 #endif 1424 m = vm_phys_alloc_pages(object != NULL ? 1425 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0); 1426 #if VM_NRESERVLEVEL > 0 1427 if (m == NULL && vm_reserv_reclaim_inactive()) { 1428 m = vm_phys_alloc_pages(object != NULL ? 1429 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 1430 0); 1431 } 1432 #endif 1433 } 1434 } else { 1435 /* 1436 * Not allocatable, give up. 1437 */ 1438 mtx_unlock(&vm_page_queue_free_mtx); 1439 atomic_add_int(&vm_pageout_deficit, 1440 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 1441 pagedaemon_wakeup(); 1442 return (NULL); 1443 } 1444 1445 /* 1446 * At this point we had better have found a good page. 1447 */ 1448 KASSERT(m != NULL, ("vm_page_alloc: missing page")); 1449 KASSERT(m->queue == PQ_NONE, 1450 ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue)); 1451 KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m)); 1452 KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m)); 1453 KASSERT(m->busy == 0, ("vm_page_alloc: page %p is busy", m)); 1454 KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m)); 1455 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 1456 ("vm_page_alloc: page %p has unexpected memattr %d", m, 1457 pmap_page_get_memattr(m))); 1458 if ((m->flags & PG_CACHED) != 0) { 1459 KASSERT((m->flags & PG_ZERO) == 0, 1460 ("vm_page_alloc: cached page %p is PG_ZERO", m)); 1461 KASSERT(m->valid != 0, 1462 ("vm_page_alloc: cached page %p is invalid", m)); 1463 if (m->object == object && m->pindex == pindex) 1464 cnt.v_reactivated++; 1465 else 1466 m->valid = 0; 1467 m_object = m->object; 1468 vm_page_cache_remove(m); 1469 if (m_object->type == OBJT_VNODE && m_object->cache == NULL) 1470 vp = m_object->handle; 1471 } else { 1472 KASSERT(VM_PAGE_IS_FREE(m), 1473 ("vm_page_alloc: page %p is not free", m)); 1474 KASSERT(m->valid == 0, 1475 ("vm_page_alloc: free page %p is valid", m)); 1476 cnt.v_free_count--; 1477 } 1478 1479 /* 1480 * Only the PG_ZERO flag is inherited. The PG_CACHED or PG_FREE flag 1481 * must be cleared before the free page queues lock is released. 1482 */ 1483 flags = 0; 1484 if (m->flags & PG_ZERO) { 1485 vm_page_zero_count--; 1486 if (req & VM_ALLOC_ZERO) 1487 flags = PG_ZERO; 1488 } 1489 if (req & VM_ALLOC_NODUMP) 1490 flags |= PG_NODUMP; 1491 m->flags = flags; 1492 mtx_unlock(&vm_page_queue_free_mtx); 1493 m->aflags = 0; 1494 if (object == NULL || object->type == OBJT_PHYS) 1495 m->oflags = VPO_UNMANAGED; 1496 else 1497 m->oflags = 0; 1498 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ)) == 0) 1499 m->oflags |= VPO_BUSY; 1500 if (req & VM_ALLOC_WIRED) { 1501 /* 1502 * The page lock is not required for wiring a page until that 1503 * page is inserted into the object. 1504 */ 1505 atomic_add_int(&cnt.v_wire_count, 1); 1506 m->wire_count = 1; 1507 } 1508 m->act_count = 0; 1509 1510 if (object != NULL) { 1511 /* Ignore device objects; the pager sets "memattr" for them. */ 1512 if (object->memattr != VM_MEMATTR_DEFAULT && 1513 object->type != OBJT_DEVICE && object->type != OBJT_SG) 1514 pmap_page_set_memattr(m, object->memattr); 1515 vm_page_insert(m, object, pindex); 1516 } else 1517 m->pindex = pindex; 1518 1519 /* 1520 * The following call to vdrop() must come after the above call 1521 * to vm_page_insert() in case both affect the same object and 1522 * vnode. Otherwise, the affected vnode's hold count could 1523 * temporarily become zero. 1524 */ 1525 if (vp != NULL) 1526 vdrop(vp); 1527 1528 /* 1529 * Don't wakeup too often - wakeup the pageout daemon when 1530 * we would be nearly out of memory. 1531 */ 1532 if (vm_paging_needed()) 1533 pagedaemon_wakeup(); 1534 1535 return (m); 1536 } 1537 1538 /* 1539 * vm_page_alloc_contig: 1540 * 1541 * Allocate a contiguous set of physical pages of the given size "npages" 1542 * from the free lists. All of the physical pages must be at or above 1543 * the given physical address "low" and below the given physical address 1544 * "high". The given value "alignment" determines the alignment of the 1545 * first physical page in the set. If the given value "boundary" is 1546 * non-zero, then the set of physical pages cannot cross any physical 1547 * address boundary that is a multiple of that value. Both "alignment" 1548 * and "boundary" must be a power of two. 1549 * 1550 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, 1551 * then the memory attribute setting for the physical pages is configured 1552 * to the object's memory attribute setting. Otherwise, the memory 1553 * attribute setting for the physical pages is configured to "memattr", 1554 * overriding the object's memory attribute setting. However, if the 1555 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the 1556 * memory attribute setting for the physical pages cannot be configured 1557 * to VM_MEMATTR_DEFAULT. 1558 * 1559 * The caller must always specify an allocation class. 1560 * 1561 * allocation classes: 1562 * VM_ALLOC_NORMAL normal process request 1563 * VM_ALLOC_SYSTEM system *really* needs a page 1564 * VM_ALLOC_INTERRUPT interrupt time request 1565 * 1566 * optional allocation flags: 1567 * VM_ALLOC_NOBUSY do not set the flag VPO_BUSY on the page 1568 * VM_ALLOC_NOOBJ page is not associated with an object and 1569 * should not have the flag VPO_BUSY set 1570 * VM_ALLOC_WIRED wire the allocated page 1571 * VM_ALLOC_ZERO prefer a zeroed page 1572 * 1573 * This routine may not sleep. 1574 */ 1575 vm_page_t 1576 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, 1577 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 1578 vm_paddr_t boundary, vm_memattr_t memattr) 1579 { 1580 struct vnode *drop; 1581 vm_page_t deferred_vdrop_list, m, m_ret; 1582 u_int flags, oflags; 1583 int req_class; 1584 1585 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0), 1586 ("vm_page_alloc_contig: inconsistent object/req")); 1587 if (object != NULL) { 1588 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1589 KASSERT(object->type == OBJT_PHYS, 1590 ("vm_page_alloc_contig: object %p isn't OBJT_PHYS", 1591 object)); 1592 } 1593 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 1594 req_class = req & VM_ALLOC_CLASS_MASK; 1595 1596 /* 1597 * The page daemon is allowed to dig deeper into the free page list. 1598 */ 1599 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1600 req_class = VM_ALLOC_SYSTEM; 1601 1602 deferred_vdrop_list = NULL; 1603 mtx_lock(&vm_page_queue_free_mtx); 1604 if (cnt.v_free_count + cnt.v_cache_count >= npages + 1605 cnt.v_free_reserved || (req_class == VM_ALLOC_SYSTEM && 1606 cnt.v_free_count + cnt.v_cache_count >= npages + 1607 cnt.v_interrupt_free_min) || (req_class == VM_ALLOC_INTERRUPT && 1608 cnt.v_free_count + cnt.v_cache_count >= npages)) { 1609 #if VM_NRESERVLEVEL > 0 1610 retry: 1611 if (object == NULL || (object->flags & OBJ_COLORED) == 0 || 1612 (m_ret = vm_reserv_alloc_contig(object, pindex, npages, 1613 low, high, alignment, boundary)) == NULL) 1614 #endif 1615 m_ret = vm_phys_alloc_contig(npages, low, high, 1616 alignment, boundary); 1617 } else { 1618 mtx_unlock(&vm_page_queue_free_mtx); 1619 atomic_add_int(&vm_pageout_deficit, npages); 1620 pagedaemon_wakeup(); 1621 return (NULL); 1622 } 1623 if (m_ret != NULL) 1624 for (m = m_ret; m < &m_ret[npages]; m++) { 1625 drop = vm_page_alloc_init(m); 1626 if (drop != NULL) { 1627 /* 1628 * Enqueue the vnode for deferred vdrop(). 1629 * 1630 * Once the pages are removed from the free 1631 * page list, "pageq" can be safely abused to 1632 * construct a short-lived list of vnodes. 1633 */ 1634 m->pageq.tqe_prev = (void *)drop; 1635 m->pageq.tqe_next = deferred_vdrop_list; 1636 deferred_vdrop_list = m; 1637 } 1638 } 1639 else { 1640 #if VM_NRESERVLEVEL > 0 1641 if (vm_reserv_reclaim_contig(npages, low, high, alignment, 1642 boundary)) 1643 goto retry; 1644 #endif 1645 } 1646 mtx_unlock(&vm_page_queue_free_mtx); 1647 if (m_ret == NULL) 1648 return (NULL); 1649 1650 /* 1651 * Initialize the pages. Only the PG_ZERO flag is inherited. 1652 */ 1653 flags = 0; 1654 if ((req & VM_ALLOC_ZERO) != 0) 1655 flags = PG_ZERO; 1656 if ((req & VM_ALLOC_NODUMP) != 0) 1657 flags |= PG_NODUMP; 1658 if ((req & VM_ALLOC_WIRED) != 0) 1659 atomic_add_int(&cnt.v_wire_count, npages); 1660 oflags = VPO_UNMANAGED; 1661 if (object != NULL) { 1662 if ((req & VM_ALLOC_NOBUSY) == 0) 1663 oflags |= VPO_BUSY; 1664 if (object->memattr != VM_MEMATTR_DEFAULT && 1665 memattr == VM_MEMATTR_DEFAULT) 1666 memattr = object->memattr; 1667 } 1668 for (m = m_ret; m < &m_ret[npages]; m++) { 1669 m->aflags = 0; 1670 m->flags = (m->flags | PG_NODUMP) & flags; 1671 if ((req & VM_ALLOC_WIRED) != 0) 1672 m->wire_count = 1; 1673 /* Unmanaged pages don't use "act_count". */ 1674 m->oflags = oflags; 1675 if (memattr != VM_MEMATTR_DEFAULT) 1676 pmap_page_set_memattr(m, memattr); 1677 if (object != NULL) 1678 vm_page_insert(m, object, pindex); 1679 else 1680 m->pindex = pindex; 1681 pindex++; 1682 } 1683 while (deferred_vdrop_list != NULL) { 1684 vdrop((struct vnode *)deferred_vdrop_list->pageq.tqe_prev); 1685 deferred_vdrop_list = deferred_vdrop_list->pageq.tqe_next; 1686 } 1687 if (vm_paging_needed()) 1688 pagedaemon_wakeup(); 1689 return (m_ret); 1690 } 1691 1692 /* 1693 * Initialize a page that has been freshly dequeued from a freelist. 1694 * The caller has to drop the vnode returned, if it is not NULL. 1695 * 1696 * This function may only be used to initialize unmanaged pages. 1697 * 1698 * To be called with vm_page_queue_free_mtx held. 1699 */ 1700 static struct vnode * 1701 vm_page_alloc_init(vm_page_t m) 1702 { 1703 struct vnode *drop; 1704 vm_object_t m_object; 1705 1706 KASSERT(m->queue == PQ_NONE, 1707 ("vm_page_alloc_init: page %p has unexpected queue %d", 1708 m, m->queue)); 1709 KASSERT(m->wire_count == 0, 1710 ("vm_page_alloc_init: page %p is wired", m)); 1711 KASSERT(m->hold_count == 0, 1712 ("vm_page_alloc_init: page %p is held", m)); 1713 KASSERT(m->busy == 0, 1714 ("vm_page_alloc_init: page %p is busy", m)); 1715 KASSERT(m->dirty == 0, 1716 ("vm_page_alloc_init: page %p is dirty", m)); 1717 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 1718 ("vm_page_alloc_init: page %p has unexpected memattr %d", 1719 m, pmap_page_get_memattr(m))); 1720 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1721 drop = NULL; 1722 if ((m->flags & PG_CACHED) != 0) { 1723 KASSERT((m->flags & PG_ZERO) == 0, 1724 ("vm_page_alloc_init: cached page %p is PG_ZERO", m)); 1725 m->valid = 0; 1726 m_object = m->object; 1727 vm_page_cache_remove(m); 1728 if (m_object->type == OBJT_VNODE && m_object->cache == NULL) 1729 drop = m_object->handle; 1730 } else { 1731 KASSERT(VM_PAGE_IS_FREE(m), 1732 ("vm_page_alloc_init: page %p is not free", m)); 1733 KASSERT(m->valid == 0, 1734 ("vm_page_alloc_init: free page %p is valid", m)); 1735 cnt.v_free_count--; 1736 if ((m->flags & PG_ZERO) != 0) 1737 vm_page_zero_count--; 1738 } 1739 /* Don't clear the PG_ZERO flag; we'll need it later. */ 1740 m->flags &= PG_ZERO; 1741 return (drop); 1742 } 1743 1744 /* 1745 * vm_page_alloc_freelist: 1746 * 1747 * Allocate a physical page from the specified free page list. 1748 * 1749 * The caller must always specify an allocation class. 1750 * 1751 * allocation classes: 1752 * VM_ALLOC_NORMAL normal process request 1753 * VM_ALLOC_SYSTEM system *really* needs a page 1754 * VM_ALLOC_INTERRUPT interrupt time request 1755 * 1756 * optional allocation flags: 1757 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1758 * intends to allocate 1759 * VM_ALLOC_WIRED wire the allocated page 1760 * VM_ALLOC_ZERO prefer a zeroed page 1761 * 1762 * This routine may not sleep. 1763 */ 1764 vm_page_t 1765 vm_page_alloc_freelist(int flind, int req) 1766 { 1767 struct vnode *drop; 1768 vm_page_t m; 1769 u_int flags; 1770 int req_class; 1771 1772 req_class = req & VM_ALLOC_CLASS_MASK; 1773 1774 /* 1775 * The page daemon is allowed to dig deeper into the free page list. 1776 */ 1777 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1778 req_class = VM_ALLOC_SYSTEM; 1779 1780 /* 1781 * Do not allocate reserved pages unless the req has asked for it. 1782 */ 1783 mtx_lock(&vm_page_queue_free_mtx); 1784 if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved || 1785 (req_class == VM_ALLOC_SYSTEM && 1786 cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) || 1787 (req_class == VM_ALLOC_INTERRUPT && 1788 cnt.v_free_count + cnt.v_cache_count > 0)) 1789 m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0); 1790 else { 1791 mtx_unlock(&vm_page_queue_free_mtx); 1792 atomic_add_int(&vm_pageout_deficit, 1793 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 1794 pagedaemon_wakeup(); 1795 return (NULL); 1796 } 1797 if (m == NULL) { 1798 mtx_unlock(&vm_page_queue_free_mtx); 1799 return (NULL); 1800 } 1801 drop = vm_page_alloc_init(m); 1802 mtx_unlock(&vm_page_queue_free_mtx); 1803 1804 /* 1805 * Initialize the page. Only the PG_ZERO flag is inherited. 1806 */ 1807 m->aflags = 0; 1808 flags = 0; 1809 if ((req & VM_ALLOC_ZERO) != 0) 1810 flags = PG_ZERO; 1811 m->flags &= flags; 1812 if ((req & VM_ALLOC_WIRED) != 0) { 1813 /* 1814 * The page lock is not required for wiring a page that does 1815 * not belong to an object. 1816 */ 1817 atomic_add_int(&cnt.v_wire_count, 1); 1818 m->wire_count = 1; 1819 } 1820 /* Unmanaged pages don't use "act_count". */ 1821 m->oflags = VPO_UNMANAGED; 1822 if (drop != NULL) 1823 vdrop(drop); 1824 if (vm_paging_needed()) 1825 pagedaemon_wakeup(); 1826 return (m); 1827 } 1828 1829 /* 1830 * vm_wait: (also see VM_WAIT macro) 1831 * 1832 * Sleep until free pages are available for allocation. 1833 * - Called in various places before memory allocations. 1834 */ 1835 void 1836 vm_wait(void) 1837 { 1838 1839 mtx_lock(&vm_page_queue_free_mtx); 1840 if (curproc == pageproc) { 1841 vm_pageout_pages_needed = 1; 1842 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx, 1843 PDROP | PSWP, "VMWait", 0); 1844 } else { 1845 if (!vm_pages_needed) { 1846 vm_pages_needed = 1; 1847 wakeup(&vm_pages_needed); 1848 } 1849 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM, 1850 "vmwait", 0); 1851 } 1852 } 1853 1854 /* 1855 * vm_waitpfault: (also see VM_WAITPFAULT macro) 1856 * 1857 * Sleep until free pages are available for allocation. 1858 * - Called only in vm_fault so that processes page faulting 1859 * can be easily tracked. 1860 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 1861 * processes will be able to grab memory first. Do not change 1862 * this balance without careful testing first. 1863 */ 1864 void 1865 vm_waitpfault(void) 1866 { 1867 1868 mtx_lock(&vm_page_queue_free_mtx); 1869 if (!vm_pages_needed) { 1870 vm_pages_needed = 1; 1871 wakeup(&vm_pages_needed); 1872 } 1873 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER, 1874 "pfault", 0); 1875 } 1876 1877 /* 1878 * vm_page_dequeue: 1879 * 1880 * Remove the given page from its current page queue. 1881 * 1882 * The page must be locked. 1883 */ 1884 void 1885 vm_page_dequeue(vm_page_t m) 1886 { 1887 struct vm_pagequeue *pq; 1888 1889 vm_page_lock_assert(m, MA_OWNED); 1890 KASSERT(m->queue != PQ_NONE, 1891 ("vm_page_dequeue: page %p is not queued", m)); 1892 pq = &vm_pagequeues[m->queue]; 1893 vm_pagequeue_lock(pq); 1894 m->queue = PQ_NONE; 1895 TAILQ_REMOVE(&pq->pq_pl, m, pageq); 1896 (*pq->pq_cnt)--; 1897 vm_pagequeue_unlock(pq); 1898 } 1899 1900 /* 1901 * vm_page_dequeue_locked: 1902 * 1903 * Remove the given page from its current page queue. 1904 * 1905 * The page and page queue must be locked. 1906 */ 1907 void 1908 vm_page_dequeue_locked(vm_page_t m) 1909 { 1910 struct vm_pagequeue *pq; 1911 1912 vm_page_lock_assert(m, MA_OWNED); 1913 pq = &vm_pagequeues[m->queue]; 1914 vm_pagequeue_assert_locked(pq); 1915 m->queue = PQ_NONE; 1916 TAILQ_REMOVE(&pq->pq_pl, m, pageq); 1917 (*pq->pq_cnt)--; 1918 } 1919 1920 /* 1921 * vm_page_enqueue: 1922 * 1923 * Add the given page to the specified page queue. 1924 * 1925 * The page must be locked. 1926 */ 1927 static void 1928 vm_page_enqueue(int queue, vm_page_t m) 1929 { 1930 struct vm_pagequeue *pq; 1931 1932 vm_page_lock_assert(m, MA_OWNED); 1933 pq = &vm_pagequeues[queue]; 1934 vm_pagequeue_lock(pq); 1935 m->queue = queue; 1936 TAILQ_INSERT_TAIL(&pq->pq_pl, m, pageq); 1937 ++*pq->pq_cnt; 1938 vm_pagequeue_unlock(pq); 1939 } 1940 1941 /* 1942 * vm_page_requeue: 1943 * 1944 * Move the given page to the tail of its current page queue. 1945 * 1946 * The page must be locked. 1947 */ 1948 void 1949 vm_page_requeue(vm_page_t m) 1950 { 1951 struct vm_pagequeue *pq; 1952 1953 vm_page_lock_assert(m, MA_OWNED); 1954 KASSERT(m->queue != PQ_NONE, 1955 ("vm_page_requeue: page %p is not queued", m)); 1956 pq = &vm_pagequeues[m->queue]; 1957 vm_pagequeue_lock(pq); 1958 TAILQ_REMOVE(&pq->pq_pl, m, pageq); 1959 TAILQ_INSERT_TAIL(&pq->pq_pl, m, pageq); 1960 vm_pagequeue_unlock(pq); 1961 } 1962 1963 /* 1964 * vm_page_requeue_locked: 1965 * 1966 * Move the given page to the tail of its current page queue. 1967 * 1968 * The page queue must be locked. 1969 */ 1970 void 1971 vm_page_requeue_locked(vm_page_t m) 1972 { 1973 struct vm_pagequeue *pq; 1974 1975 KASSERT(m->queue != PQ_NONE, 1976 ("vm_page_requeue_locked: page %p is not queued", m)); 1977 pq = &vm_pagequeues[m->queue]; 1978 vm_pagequeue_assert_locked(pq); 1979 TAILQ_REMOVE(&pq->pq_pl, m, pageq); 1980 TAILQ_INSERT_TAIL(&pq->pq_pl, m, pageq); 1981 } 1982 1983 /* 1984 * vm_page_activate: 1985 * 1986 * Put the specified page on the active list (if appropriate). 1987 * Ensure that act_count is at least ACT_INIT but do not otherwise 1988 * mess with it. 1989 * 1990 * The page must be locked. 1991 */ 1992 void 1993 vm_page_activate(vm_page_t m) 1994 { 1995 int queue; 1996 1997 vm_page_lock_assert(m, MA_OWNED); 1998 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1999 if ((queue = m->queue) != PQ_ACTIVE) { 2000 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 2001 if (m->act_count < ACT_INIT) 2002 m->act_count = ACT_INIT; 2003 if (queue != PQ_NONE) 2004 vm_page_dequeue(m); 2005 vm_page_enqueue(PQ_ACTIVE, m); 2006 } else 2007 KASSERT(queue == PQ_NONE, 2008 ("vm_page_activate: wired page %p is queued", m)); 2009 } else { 2010 if (m->act_count < ACT_INIT) 2011 m->act_count = ACT_INIT; 2012 } 2013 } 2014 2015 /* 2016 * vm_page_free_wakeup: 2017 * 2018 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 2019 * routine is called when a page has been added to the cache or free 2020 * queues. 2021 * 2022 * The page queues must be locked. 2023 */ 2024 static inline void 2025 vm_page_free_wakeup(void) 2026 { 2027 2028 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 2029 /* 2030 * if pageout daemon needs pages, then tell it that there are 2031 * some free. 2032 */ 2033 if (vm_pageout_pages_needed && 2034 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 2035 wakeup(&vm_pageout_pages_needed); 2036 vm_pageout_pages_needed = 0; 2037 } 2038 /* 2039 * wakeup processes that are waiting on memory if we hit a 2040 * high water mark. And wakeup scheduler process if we have 2041 * lots of memory. this process will swapin processes. 2042 */ 2043 if (vm_pages_needed && !vm_page_count_min()) { 2044 vm_pages_needed = 0; 2045 wakeup(&cnt.v_free_count); 2046 } 2047 } 2048 2049 /* 2050 * vm_page_free_toq: 2051 * 2052 * Returns the given page to the free list, 2053 * disassociating it with any VM object. 2054 * 2055 * The object must be locked. The page must be locked if it is managed. 2056 */ 2057 void 2058 vm_page_free_toq(vm_page_t m) 2059 { 2060 2061 if ((m->oflags & VPO_UNMANAGED) == 0) { 2062 vm_page_lock_assert(m, MA_OWNED); 2063 KASSERT(!pmap_page_is_mapped(m), 2064 ("vm_page_free_toq: freeing mapped page %p", m)); 2065 } else 2066 KASSERT(m->queue == PQ_NONE, 2067 ("vm_page_free_toq: unmanaged page %p is queued", m)); 2068 PCPU_INC(cnt.v_tfree); 2069 2070 if (VM_PAGE_IS_FREE(m)) 2071 panic("vm_page_free: freeing free page %p", m); 2072 else if (m->busy != 0) 2073 panic("vm_page_free: freeing busy page %p", m); 2074 2075 /* 2076 * Unqueue, then remove page. Note that we cannot destroy 2077 * the page here because we do not want to call the pager's 2078 * callback routine until after we've put the page on the 2079 * appropriate free queue. 2080 */ 2081 vm_page_remque(m); 2082 vm_page_remove(m); 2083 2084 /* 2085 * If fictitious remove object association and 2086 * return, otherwise delay object association removal. 2087 */ 2088 if ((m->flags & PG_FICTITIOUS) != 0) { 2089 return; 2090 } 2091 2092 m->valid = 0; 2093 vm_page_undirty(m); 2094 2095 if (m->wire_count != 0) 2096 panic("vm_page_free: freeing wired page %p", m); 2097 if (m->hold_count != 0) { 2098 m->flags &= ~PG_ZERO; 2099 KASSERT((m->flags & PG_UNHOLDFREE) == 0, 2100 ("vm_page_free: freeing PG_UNHOLDFREE page %p", m)); 2101 m->flags |= PG_UNHOLDFREE; 2102 } else { 2103 /* 2104 * Restore the default memory attribute to the page. 2105 */ 2106 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 2107 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 2108 2109 /* 2110 * Insert the page into the physical memory allocator's 2111 * cache/free page queues. 2112 */ 2113 mtx_lock(&vm_page_queue_free_mtx); 2114 m->flags |= PG_FREE; 2115 cnt.v_free_count++; 2116 #if VM_NRESERVLEVEL > 0 2117 if (!vm_reserv_free_page(m)) 2118 #else 2119 if (TRUE) 2120 #endif 2121 vm_phys_free_pages(m, 0); 2122 if ((m->flags & PG_ZERO) != 0) 2123 ++vm_page_zero_count; 2124 else 2125 vm_page_zero_idle_wakeup(); 2126 vm_page_free_wakeup(); 2127 mtx_unlock(&vm_page_queue_free_mtx); 2128 } 2129 } 2130 2131 /* 2132 * vm_page_wire: 2133 * 2134 * Mark this page as wired down by yet 2135 * another map, removing it from paging queues 2136 * as necessary. 2137 * 2138 * If the page is fictitious, then its wire count must remain one. 2139 * 2140 * The page must be locked. 2141 */ 2142 void 2143 vm_page_wire(vm_page_t m) 2144 { 2145 2146 /* 2147 * Only bump the wire statistics if the page is not already wired, 2148 * and only unqueue the page if it is on some queue (if it is unmanaged 2149 * it is already off the queues). 2150 */ 2151 vm_page_lock_assert(m, MA_OWNED); 2152 if ((m->flags & PG_FICTITIOUS) != 0) { 2153 KASSERT(m->wire_count == 1, 2154 ("vm_page_wire: fictitious page %p's wire count isn't one", 2155 m)); 2156 return; 2157 } 2158 if (m->wire_count == 0) { 2159 KASSERT((m->oflags & VPO_UNMANAGED) == 0 || 2160 m->queue == PQ_NONE, 2161 ("vm_page_wire: unmanaged page %p is queued", m)); 2162 vm_page_remque(m); 2163 atomic_add_int(&cnt.v_wire_count, 1); 2164 } 2165 m->wire_count++; 2166 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 2167 } 2168 2169 /* 2170 * vm_page_unwire: 2171 * 2172 * Release one wiring of the specified page, potentially enabling it to be 2173 * paged again. If paging is enabled, then the value of the parameter 2174 * "activate" determines to which queue the page is added. If "activate" is 2175 * non-zero, then the page is added to the active queue. Otherwise, it is 2176 * added to the inactive queue. 2177 * 2178 * However, unless the page belongs to an object, it is not enqueued because 2179 * it cannot be paged out. 2180 * 2181 * If a page is fictitious, then its wire count must alway be one. 2182 * 2183 * A managed page must be locked. 2184 */ 2185 void 2186 vm_page_unwire(vm_page_t m, int activate) 2187 { 2188 2189 if ((m->oflags & VPO_UNMANAGED) == 0) 2190 vm_page_lock_assert(m, MA_OWNED); 2191 if ((m->flags & PG_FICTITIOUS) != 0) { 2192 KASSERT(m->wire_count == 1, 2193 ("vm_page_unwire: fictitious page %p's wire count isn't one", m)); 2194 return; 2195 } 2196 if (m->wire_count > 0) { 2197 m->wire_count--; 2198 if (m->wire_count == 0) { 2199 atomic_subtract_int(&cnt.v_wire_count, 1); 2200 if ((m->oflags & VPO_UNMANAGED) != 0 || 2201 m->object == NULL) 2202 return; 2203 if (!activate) 2204 m->flags &= ~PG_WINATCFLS; 2205 vm_page_enqueue(activate ? PQ_ACTIVE : PQ_INACTIVE, m); 2206 } 2207 } else 2208 panic("vm_page_unwire: page %p's wire count is zero", m); 2209 } 2210 2211 /* 2212 * Move the specified page to the inactive queue. 2213 * 2214 * Many pages placed on the inactive queue should actually go 2215 * into the cache, but it is difficult to figure out which. What 2216 * we do instead, if the inactive target is well met, is to put 2217 * clean pages at the head of the inactive queue instead of the tail. 2218 * This will cause them to be moved to the cache more quickly and 2219 * if not actively re-referenced, reclaimed more quickly. If we just 2220 * stick these pages at the end of the inactive queue, heavy filesystem 2221 * meta-data accesses can cause an unnecessary paging load on memory bound 2222 * processes. This optimization causes one-time-use metadata to be 2223 * reused more quickly. 2224 * 2225 * Normally athead is 0 resulting in LRU operation. athead is set 2226 * to 1 if we want this page to be 'as if it were placed in the cache', 2227 * except without unmapping it from the process address space. 2228 * 2229 * The page must be locked. 2230 */ 2231 static inline void 2232 _vm_page_deactivate(vm_page_t m, int athead) 2233 { 2234 struct vm_pagequeue *pq; 2235 int queue; 2236 2237 vm_page_lock_assert(m, MA_OWNED); 2238 2239 /* 2240 * Ignore if already inactive. 2241 */ 2242 if ((queue = m->queue) == PQ_INACTIVE) 2243 return; 2244 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 2245 if (queue != PQ_NONE) 2246 vm_page_dequeue(m); 2247 m->flags &= ~PG_WINATCFLS; 2248 pq = &vm_pagequeues[PQ_INACTIVE]; 2249 vm_pagequeue_lock(pq); 2250 m->queue = PQ_INACTIVE; 2251 if (athead) 2252 TAILQ_INSERT_HEAD(&pq->pq_pl, m, pageq); 2253 else 2254 TAILQ_INSERT_TAIL(&pq->pq_pl, m, pageq); 2255 cnt.v_inactive_count++; 2256 vm_pagequeue_unlock(pq); 2257 } 2258 } 2259 2260 /* 2261 * Move the specified page to the inactive queue. 2262 * 2263 * The page must be locked. 2264 */ 2265 void 2266 vm_page_deactivate(vm_page_t m) 2267 { 2268 2269 _vm_page_deactivate(m, 0); 2270 } 2271 2272 /* 2273 * vm_page_try_to_cache: 2274 * 2275 * Returns 0 on failure, 1 on success 2276 */ 2277 int 2278 vm_page_try_to_cache(vm_page_t m) 2279 { 2280 2281 vm_page_lock_assert(m, MA_OWNED); 2282 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2283 if (m->dirty || m->hold_count || m->busy || m->wire_count || 2284 (m->oflags & (VPO_BUSY | VPO_UNMANAGED)) != 0) 2285 return (0); 2286 pmap_remove_all(m); 2287 if (m->dirty) 2288 return (0); 2289 vm_page_cache(m); 2290 return (1); 2291 } 2292 2293 /* 2294 * vm_page_try_to_free() 2295 * 2296 * Attempt to free the page. If we cannot free it, we do nothing. 2297 * 1 is returned on success, 0 on failure. 2298 */ 2299 int 2300 vm_page_try_to_free(vm_page_t m) 2301 { 2302 2303 vm_page_lock_assert(m, MA_OWNED); 2304 if (m->object != NULL) 2305 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2306 if (m->dirty || m->hold_count || m->busy || m->wire_count || 2307 (m->oflags & (VPO_BUSY | VPO_UNMANAGED)) != 0) 2308 return (0); 2309 pmap_remove_all(m); 2310 if (m->dirty) 2311 return (0); 2312 vm_page_free(m); 2313 return (1); 2314 } 2315 2316 /* 2317 * vm_page_cache 2318 * 2319 * Put the specified page onto the page cache queue (if appropriate). 2320 * 2321 * The object and page must be locked. 2322 */ 2323 void 2324 vm_page_cache(vm_page_t m) 2325 { 2326 vm_object_t object; 2327 vm_page_t next, prev, root; 2328 2329 vm_page_lock_assert(m, MA_OWNED); 2330 object = m->object; 2331 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2332 if ((m->oflags & (VPO_UNMANAGED | VPO_BUSY)) || m->busy || 2333 m->hold_count || m->wire_count) 2334 panic("vm_page_cache: attempting to cache busy page"); 2335 KASSERT(!pmap_page_is_mapped(m), 2336 ("vm_page_cache: page %p is mapped", m)); 2337 KASSERT(m->dirty == 0, ("vm_page_cache: page %p is dirty", m)); 2338 if (m->valid == 0 || object->type == OBJT_DEFAULT || 2339 (object->type == OBJT_SWAP && 2340 !vm_pager_has_page(object, m->pindex, NULL, NULL))) { 2341 /* 2342 * Hypothesis: A cache-elgible page belonging to a 2343 * default object or swap object but without a backing 2344 * store must be zero filled. 2345 */ 2346 vm_page_free(m); 2347 return; 2348 } 2349 KASSERT((m->flags & PG_CACHED) == 0, 2350 ("vm_page_cache: page %p is already cached", m)); 2351 PCPU_INC(cnt.v_tcached); 2352 2353 /* 2354 * Remove the page from the paging queues. 2355 */ 2356 vm_page_remque(m); 2357 2358 /* 2359 * Remove the page from the object's collection of resident 2360 * pages. 2361 */ 2362 if ((next = TAILQ_NEXT(m, listq)) != NULL && next->left == m) { 2363 /* 2364 * Since the page's successor in the list is also its parent 2365 * in the tree, its right subtree must be empty. 2366 */ 2367 next->left = m->left; 2368 KASSERT(m->right == NULL, 2369 ("vm_page_cache: page %p has right child", m)); 2370 } else if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL && 2371 prev->right == m) { 2372 /* 2373 * Since the page's predecessor in the list is also its parent 2374 * in the tree, its left subtree must be empty. 2375 */ 2376 KASSERT(m->left == NULL, 2377 ("vm_page_cache: page %p has left child", m)); 2378 prev->right = m->right; 2379 } else { 2380 if (m != object->root) 2381 vm_page_splay(m->pindex, object->root); 2382 if (m->left == NULL) 2383 root = m->right; 2384 else if (m->right == NULL) 2385 root = m->left; 2386 else { 2387 /* 2388 * Move the page's successor to the root, because 2389 * pages are usually removed in ascending order. 2390 */ 2391 if (m->right != next) 2392 vm_page_splay(m->pindex, m->right); 2393 next->left = m->left; 2394 root = next; 2395 } 2396 object->root = root; 2397 } 2398 TAILQ_REMOVE(&object->memq, m, listq); 2399 object->resident_page_count--; 2400 2401 /* 2402 * Restore the default memory attribute to the page. 2403 */ 2404 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 2405 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 2406 2407 /* 2408 * Insert the page into the object's collection of cached pages 2409 * and the physical memory allocator's cache/free page queues. 2410 */ 2411 m->flags &= ~PG_ZERO; 2412 mtx_lock(&vm_page_queue_free_mtx); 2413 m->flags |= PG_CACHED; 2414 cnt.v_cache_count++; 2415 root = object->cache; 2416 if (root == NULL) { 2417 m->left = NULL; 2418 m->right = NULL; 2419 } else { 2420 root = vm_page_splay(m->pindex, root); 2421 if (m->pindex < root->pindex) { 2422 m->left = root->left; 2423 m->right = root; 2424 root->left = NULL; 2425 } else if (__predict_false(m->pindex == root->pindex)) 2426 panic("vm_page_cache: offset already cached"); 2427 else { 2428 m->right = root->right; 2429 m->left = root; 2430 root->right = NULL; 2431 } 2432 } 2433 object->cache = m; 2434 #if VM_NRESERVLEVEL > 0 2435 if (!vm_reserv_free_page(m)) { 2436 #else 2437 if (TRUE) { 2438 #endif 2439 vm_phys_set_pool(VM_FREEPOOL_CACHE, m, 0); 2440 vm_phys_free_pages(m, 0); 2441 } 2442 vm_page_free_wakeup(); 2443 mtx_unlock(&vm_page_queue_free_mtx); 2444 2445 /* 2446 * Increment the vnode's hold count if this is the object's only 2447 * cached page. Decrement the vnode's hold count if this was 2448 * the object's only resident page. 2449 */ 2450 if (object->type == OBJT_VNODE) { 2451 if (root == NULL && object->resident_page_count != 0) 2452 vhold(object->handle); 2453 else if (root != NULL && object->resident_page_count == 0) 2454 vdrop(object->handle); 2455 } 2456 } 2457 2458 /* 2459 * vm_page_dontneed 2460 * 2461 * Cache, deactivate, or do nothing as appropriate. This routine 2462 * is typically used by madvise() MADV_DONTNEED. 2463 * 2464 * Generally speaking we want to move the page into the cache so 2465 * it gets reused quickly. However, this can result in a silly syndrome 2466 * due to the page recycling too quickly. Small objects will not be 2467 * fully cached. On the otherhand, if we move the page to the inactive 2468 * queue we wind up with a problem whereby very large objects 2469 * unnecessarily blow away our inactive and cache queues. 2470 * 2471 * The solution is to move the pages based on a fixed weighting. We 2472 * either leave them alone, deactivate them, or move them to the cache, 2473 * where moving them to the cache has the highest weighting. 2474 * By forcing some pages into other queues we eventually force the 2475 * system to balance the queues, potentially recovering other unrelated 2476 * space from active. The idea is to not force this to happen too 2477 * often. 2478 * 2479 * The object and page must be locked. 2480 */ 2481 void 2482 vm_page_dontneed(vm_page_t m) 2483 { 2484 int dnw; 2485 int head; 2486 2487 vm_page_lock_assert(m, MA_OWNED); 2488 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2489 dnw = PCPU_GET(dnweight); 2490 PCPU_INC(dnweight); 2491 2492 /* 2493 * Occasionally leave the page alone. 2494 */ 2495 if ((dnw & 0x01F0) == 0 || m->queue == PQ_INACTIVE) { 2496 if (m->act_count >= ACT_INIT) 2497 --m->act_count; 2498 return; 2499 } 2500 2501 /* 2502 * Clear any references to the page. Otherwise, the page daemon will 2503 * immediately reactivate the page. 2504 * 2505 * Perform the pmap_clear_reference() first. Otherwise, a concurrent 2506 * pmap operation, such as pmap_remove(), could clear a reference in 2507 * the pmap and set PGA_REFERENCED on the page before the 2508 * pmap_clear_reference() had completed. Consequently, the page would 2509 * appear referenced based upon an old reference that occurred before 2510 * this function ran. 2511 */ 2512 pmap_clear_reference(m); 2513 vm_page_aflag_clear(m, PGA_REFERENCED); 2514 2515 if (m->dirty == 0 && pmap_is_modified(m)) 2516 vm_page_dirty(m); 2517 2518 if (m->dirty || (dnw & 0x0070) == 0) { 2519 /* 2520 * Deactivate the page 3 times out of 32. 2521 */ 2522 head = 0; 2523 } else { 2524 /* 2525 * Cache the page 28 times out of every 32. Note that 2526 * the page is deactivated instead of cached, but placed 2527 * at the head of the queue instead of the tail. 2528 */ 2529 head = 1; 2530 } 2531 _vm_page_deactivate(m, head); 2532 } 2533 2534 /* 2535 * Grab a page, waiting until we are waken up due to the page 2536 * changing state. We keep on waiting, if the page continues 2537 * to be in the object. If the page doesn't exist, first allocate it 2538 * and then conditionally zero it. 2539 * 2540 * The caller must always specify the VM_ALLOC_RETRY flag. This is intended 2541 * to facilitate its eventual removal. 2542 * 2543 * This routine may sleep. 2544 * 2545 * The object must be locked on entry. The lock will, however, be released 2546 * and reacquired if the routine sleeps. 2547 */ 2548 vm_page_t 2549 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 2550 { 2551 vm_page_t m; 2552 2553 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2554 KASSERT((allocflags & VM_ALLOC_RETRY) != 0, 2555 ("vm_page_grab: VM_ALLOC_RETRY is required")); 2556 retrylookup: 2557 if ((m = vm_page_lookup(object, pindex)) != NULL) { 2558 if ((m->oflags & VPO_BUSY) != 0 || 2559 ((allocflags & VM_ALLOC_IGN_SBUSY) == 0 && m->busy != 0)) { 2560 /* 2561 * Reference the page before unlocking and 2562 * sleeping so that the page daemon is less 2563 * likely to reclaim it. 2564 */ 2565 vm_page_aflag_set(m, PGA_REFERENCED); 2566 vm_page_sleep(m, "pgrbwt"); 2567 goto retrylookup; 2568 } else { 2569 if ((allocflags & VM_ALLOC_WIRED) != 0) { 2570 vm_page_lock(m); 2571 vm_page_wire(m); 2572 vm_page_unlock(m); 2573 } 2574 if ((allocflags & VM_ALLOC_NOBUSY) == 0) 2575 vm_page_busy(m); 2576 return (m); 2577 } 2578 } 2579 m = vm_page_alloc(object, pindex, allocflags & ~(VM_ALLOC_RETRY | 2580 VM_ALLOC_IGN_SBUSY)); 2581 if (m == NULL) { 2582 VM_OBJECT_UNLOCK(object); 2583 VM_WAIT; 2584 VM_OBJECT_LOCK(object); 2585 goto retrylookup; 2586 } else if (m->valid != 0) 2587 return (m); 2588 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 2589 pmap_zero_page(m); 2590 return (m); 2591 } 2592 2593 /* 2594 * Mapping function for valid or dirty bits in a page. 2595 * 2596 * Inputs are required to range within a page. 2597 */ 2598 vm_page_bits_t 2599 vm_page_bits(int base, int size) 2600 { 2601 int first_bit; 2602 int last_bit; 2603 2604 KASSERT( 2605 base + size <= PAGE_SIZE, 2606 ("vm_page_bits: illegal base/size %d/%d", base, size) 2607 ); 2608 2609 if (size == 0) /* handle degenerate case */ 2610 return (0); 2611 2612 first_bit = base >> DEV_BSHIFT; 2613 last_bit = (base + size - 1) >> DEV_BSHIFT; 2614 2615 return (((vm_page_bits_t)2 << last_bit) - 2616 ((vm_page_bits_t)1 << first_bit)); 2617 } 2618 2619 /* 2620 * vm_page_set_valid_range: 2621 * 2622 * Sets portions of a page valid. The arguments are expected 2623 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2624 * of any partial chunks touched by the range. The invalid portion of 2625 * such chunks will be zeroed. 2626 * 2627 * (base + size) must be less then or equal to PAGE_SIZE. 2628 */ 2629 void 2630 vm_page_set_valid_range(vm_page_t m, int base, int size) 2631 { 2632 int endoff, frag; 2633 2634 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2635 if (size == 0) /* handle degenerate case */ 2636 return; 2637 2638 /* 2639 * If the base is not DEV_BSIZE aligned and the valid 2640 * bit is clear, we have to zero out a portion of the 2641 * first block. 2642 */ 2643 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 2644 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 2645 pmap_zero_page_area(m, frag, base - frag); 2646 2647 /* 2648 * If the ending offset is not DEV_BSIZE aligned and the 2649 * valid bit is clear, we have to zero out a portion of 2650 * the last block. 2651 */ 2652 endoff = base + size; 2653 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 2654 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 2655 pmap_zero_page_area(m, endoff, 2656 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 2657 2658 /* 2659 * Assert that no previously invalid block that is now being validated 2660 * is already dirty. 2661 */ 2662 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 2663 ("vm_page_set_valid_range: page %p is dirty", m)); 2664 2665 /* 2666 * Set valid bits inclusive of any overlap. 2667 */ 2668 m->valid |= vm_page_bits(base, size); 2669 } 2670 2671 /* 2672 * Clear the given bits from the specified page's dirty field. 2673 */ 2674 static __inline void 2675 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) 2676 { 2677 uintptr_t addr; 2678 #if PAGE_SIZE < 16384 2679 int shift; 2680 #endif 2681 2682 /* 2683 * If the object is locked and the page is neither VPO_BUSY nor 2684 * write mapped, then the page's dirty field cannot possibly be 2685 * set by a concurrent pmap operation. 2686 */ 2687 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2688 if ((m->oflags & VPO_BUSY) == 0 && !pmap_page_is_write_mapped(m)) 2689 m->dirty &= ~pagebits; 2690 else { 2691 /* 2692 * The pmap layer can call vm_page_dirty() without 2693 * holding a distinguished lock. The combination of 2694 * the object's lock and an atomic operation suffice 2695 * to guarantee consistency of the page dirty field. 2696 * 2697 * For PAGE_SIZE == 32768 case, compiler already 2698 * properly aligns the dirty field, so no forcible 2699 * alignment is needed. Only require existence of 2700 * atomic_clear_64 when page size is 32768. 2701 */ 2702 addr = (uintptr_t)&m->dirty; 2703 #if PAGE_SIZE == 32768 2704 atomic_clear_64((uint64_t *)addr, pagebits); 2705 #elif PAGE_SIZE == 16384 2706 atomic_clear_32((uint32_t *)addr, pagebits); 2707 #else /* PAGE_SIZE <= 8192 */ 2708 /* 2709 * Use a trick to perform a 32-bit atomic on the 2710 * containing aligned word, to not depend on the existence 2711 * of atomic_clear_{8, 16}. 2712 */ 2713 shift = addr & (sizeof(uint32_t) - 1); 2714 #if BYTE_ORDER == BIG_ENDIAN 2715 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY; 2716 #else 2717 shift *= NBBY; 2718 #endif 2719 addr &= ~(sizeof(uint32_t) - 1); 2720 atomic_clear_32((uint32_t *)addr, pagebits << shift); 2721 #endif /* PAGE_SIZE */ 2722 } 2723 } 2724 2725 /* 2726 * vm_page_set_validclean: 2727 * 2728 * Sets portions of a page valid and clean. The arguments are expected 2729 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2730 * of any partial chunks touched by the range. The invalid portion of 2731 * such chunks will be zero'd. 2732 * 2733 * (base + size) must be less then or equal to PAGE_SIZE. 2734 */ 2735 void 2736 vm_page_set_validclean(vm_page_t m, int base, int size) 2737 { 2738 vm_page_bits_t oldvalid, pagebits; 2739 int endoff, frag; 2740 2741 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2742 if (size == 0) /* handle degenerate case */ 2743 return; 2744 2745 /* 2746 * If the base is not DEV_BSIZE aligned and the valid 2747 * bit is clear, we have to zero out a portion of the 2748 * first block. 2749 */ 2750 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 2751 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) 2752 pmap_zero_page_area(m, frag, base - frag); 2753 2754 /* 2755 * If the ending offset is not DEV_BSIZE aligned and the 2756 * valid bit is clear, we have to zero out a portion of 2757 * the last block. 2758 */ 2759 endoff = base + size; 2760 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 2761 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) 2762 pmap_zero_page_area(m, endoff, 2763 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 2764 2765 /* 2766 * Set valid, clear dirty bits. If validating the entire 2767 * page we can safely clear the pmap modify bit. We also 2768 * use this opportunity to clear the VPO_NOSYNC flag. If a process 2769 * takes a write fault on a MAP_NOSYNC memory area the flag will 2770 * be set again. 2771 * 2772 * We set valid bits inclusive of any overlap, but we can only 2773 * clear dirty bits for DEV_BSIZE chunks that are fully within 2774 * the range. 2775 */ 2776 oldvalid = m->valid; 2777 pagebits = vm_page_bits(base, size); 2778 m->valid |= pagebits; 2779 #if 0 /* NOT YET */ 2780 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 2781 frag = DEV_BSIZE - frag; 2782 base += frag; 2783 size -= frag; 2784 if (size < 0) 2785 size = 0; 2786 } 2787 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 2788 #endif 2789 if (base == 0 && size == PAGE_SIZE) { 2790 /* 2791 * The page can only be modified within the pmap if it is 2792 * mapped, and it can only be mapped if it was previously 2793 * fully valid. 2794 */ 2795 if (oldvalid == VM_PAGE_BITS_ALL) 2796 /* 2797 * Perform the pmap_clear_modify() first. Otherwise, 2798 * a concurrent pmap operation, such as 2799 * pmap_protect(), could clear a modification in the 2800 * pmap and set the dirty field on the page before 2801 * pmap_clear_modify() had begun and after the dirty 2802 * field was cleared here. 2803 */ 2804 pmap_clear_modify(m); 2805 m->dirty = 0; 2806 m->oflags &= ~VPO_NOSYNC; 2807 } else if (oldvalid != VM_PAGE_BITS_ALL) 2808 m->dirty &= ~pagebits; 2809 else 2810 vm_page_clear_dirty_mask(m, pagebits); 2811 } 2812 2813 void 2814 vm_page_clear_dirty(vm_page_t m, int base, int size) 2815 { 2816 2817 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 2818 } 2819 2820 /* 2821 * vm_page_set_invalid: 2822 * 2823 * Invalidates DEV_BSIZE'd chunks within a page. Both the 2824 * valid and dirty bits for the effected areas are cleared. 2825 */ 2826 void 2827 vm_page_set_invalid(vm_page_t m, int base, int size) 2828 { 2829 vm_page_bits_t bits; 2830 2831 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2832 KASSERT((m->oflags & VPO_BUSY) == 0, 2833 ("vm_page_set_invalid: page %p is busy", m)); 2834 bits = vm_page_bits(base, size); 2835 if (m->valid == VM_PAGE_BITS_ALL && bits != 0) 2836 pmap_remove_all(m); 2837 KASSERT(!pmap_page_is_mapped(m), 2838 ("vm_page_set_invalid: page %p is mapped", m)); 2839 m->valid &= ~bits; 2840 m->dirty &= ~bits; 2841 } 2842 2843 /* 2844 * vm_page_zero_invalid() 2845 * 2846 * The kernel assumes that the invalid portions of a page contain 2847 * garbage, but such pages can be mapped into memory by user code. 2848 * When this occurs, we must zero out the non-valid portions of the 2849 * page so user code sees what it expects. 2850 * 2851 * Pages are most often semi-valid when the end of a file is mapped 2852 * into memory and the file's size is not page aligned. 2853 */ 2854 void 2855 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 2856 { 2857 int b; 2858 int i; 2859 2860 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2861 /* 2862 * Scan the valid bits looking for invalid sections that 2863 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 2864 * valid bit may be set ) have already been zerod by 2865 * vm_page_set_validclean(). 2866 */ 2867 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 2868 if (i == (PAGE_SIZE / DEV_BSIZE) || 2869 (m->valid & ((vm_page_bits_t)1 << i))) { 2870 if (i > b) { 2871 pmap_zero_page_area(m, 2872 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 2873 } 2874 b = i + 1; 2875 } 2876 } 2877 2878 /* 2879 * setvalid is TRUE when we can safely set the zero'd areas 2880 * as being valid. We can do this if there are no cache consistancy 2881 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 2882 */ 2883 if (setvalid) 2884 m->valid = VM_PAGE_BITS_ALL; 2885 } 2886 2887 /* 2888 * vm_page_is_valid: 2889 * 2890 * Is (partial) page valid? Note that the case where size == 0 2891 * will return FALSE in the degenerate case where the page is 2892 * entirely invalid, and TRUE otherwise. 2893 */ 2894 int 2895 vm_page_is_valid(vm_page_t m, int base, int size) 2896 { 2897 vm_page_bits_t bits; 2898 2899 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2900 bits = vm_page_bits(base, size); 2901 if (m->valid && ((m->valid & bits) == bits)) 2902 return 1; 2903 else 2904 return 0; 2905 } 2906 2907 /* 2908 * Set the page's dirty bits if the page is modified. 2909 */ 2910 void 2911 vm_page_test_dirty(vm_page_t m) 2912 { 2913 2914 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2915 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 2916 vm_page_dirty(m); 2917 } 2918 2919 void 2920 vm_page_lock_KBI(vm_page_t m, const char *file, int line) 2921 { 2922 2923 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); 2924 } 2925 2926 void 2927 vm_page_unlock_KBI(vm_page_t m, const char *file, int line) 2928 { 2929 2930 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); 2931 } 2932 2933 int 2934 vm_page_trylock_KBI(vm_page_t m, const char *file, int line) 2935 { 2936 2937 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); 2938 } 2939 2940 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) 2941 void 2942 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) 2943 { 2944 2945 mtx_assert_(vm_page_lockptr(m), a, file, line); 2946 } 2947 #endif 2948 2949 int so_zerocp_fullpage = 0; 2950 2951 /* 2952 * Replace the given page with a copy. The copied page assumes 2953 * the portion of the given page's "wire_count" that is not the 2954 * responsibility of this copy-on-write mechanism. 2955 * 2956 * The object containing the given page must have a non-zero 2957 * paging-in-progress count and be locked. 2958 */ 2959 void 2960 vm_page_cowfault(vm_page_t m) 2961 { 2962 vm_page_t mnew; 2963 vm_object_t object; 2964 vm_pindex_t pindex; 2965 2966 vm_page_lock_assert(m, MA_OWNED); 2967 object = m->object; 2968 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2969 KASSERT(object->paging_in_progress != 0, 2970 ("vm_page_cowfault: object %p's paging-in-progress count is zero.", 2971 object)); 2972 pindex = m->pindex; 2973 2974 retry_alloc: 2975 pmap_remove_all(m); 2976 vm_page_remove(m); 2977 mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY); 2978 if (mnew == NULL) { 2979 vm_page_insert(m, object, pindex); 2980 vm_page_unlock(m); 2981 VM_OBJECT_UNLOCK(object); 2982 VM_WAIT; 2983 VM_OBJECT_LOCK(object); 2984 if (m == vm_page_lookup(object, pindex)) { 2985 vm_page_lock(m); 2986 goto retry_alloc; 2987 } else { 2988 /* 2989 * Page disappeared during the wait. 2990 */ 2991 return; 2992 } 2993 } 2994 2995 if (m->cow == 0) { 2996 /* 2997 * check to see if we raced with an xmit complete when 2998 * waiting to allocate a page. If so, put things back 2999 * the way they were 3000 */ 3001 vm_page_unlock(m); 3002 vm_page_lock(mnew); 3003 vm_page_free(mnew); 3004 vm_page_unlock(mnew); 3005 vm_page_insert(m, object, pindex); 3006 } else { /* clear COW & copy page */ 3007 if (!so_zerocp_fullpage) 3008 pmap_copy_page(m, mnew); 3009 mnew->valid = VM_PAGE_BITS_ALL; 3010 vm_page_dirty(mnew); 3011 mnew->wire_count = m->wire_count - m->cow; 3012 m->wire_count = m->cow; 3013 vm_page_unlock(m); 3014 } 3015 } 3016 3017 void 3018 vm_page_cowclear(vm_page_t m) 3019 { 3020 3021 vm_page_lock_assert(m, MA_OWNED); 3022 if (m->cow) { 3023 m->cow--; 3024 /* 3025 * let vm_fault add back write permission lazily 3026 */ 3027 } 3028 /* 3029 * sf_buf_free() will free the page, so we needn't do it here 3030 */ 3031 } 3032 3033 int 3034 vm_page_cowsetup(vm_page_t m) 3035 { 3036 3037 vm_page_lock_assert(m, MA_OWNED); 3038 if ((m->flags & PG_FICTITIOUS) != 0 || 3039 (m->oflags & VPO_UNMANAGED) != 0 || 3040 m->cow == USHRT_MAX - 1 || !VM_OBJECT_TRYLOCK(m->object)) 3041 return (EBUSY); 3042 m->cow++; 3043 pmap_remove_write(m); 3044 VM_OBJECT_UNLOCK(m->object); 3045 return (0); 3046 } 3047 3048 #ifdef INVARIANTS 3049 void 3050 vm_page_object_lock_assert(vm_page_t m) 3051 { 3052 3053 /* 3054 * Certain of the page's fields may only be modified by the 3055 * holder of the containing object's lock or the setter of the 3056 * page's VPO_BUSY flag. Unfortunately, the setter of the 3057 * VPO_BUSY flag is not recorded, and thus cannot be checked 3058 * here. 3059 */ 3060 if (m->object != NULL && (m->oflags & VPO_BUSY) == 0) 3061 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 3062 } 3063 #endif 3064 3065 #include "opt_ddb.h" 3066 #ifdef DDB 3067 #include <sys/kernel.h> 3068 3069 #include <ddb/ddb.h> 3070 3071 DB_SHOW_COMMAND(page, vm_page_print_page_info) 3072 { 3073 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 3074 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 3075 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 3076 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 3077 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 3078 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 3079 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 3080 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 3081 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 3082 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 3083 } 3084 3085 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 3086 { 3087 3088 db_printf("PQ_FREE:"); 3089 db_printf(" %d", cnt.v_free_count); 3090 db_printf("\n"); 3091 3092 db_printf("PQ_CACHE:"); 3093 db_printf(" %d", cnt.v_cache_count); 3094 db_printf("\n"); 3095 3096 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 3097 *vm_pagequeues[PQ_ACTIVE].pq_cnt, 3098 *vm_pagequeues[PQ_INACTIVE].pq_cnt); 3099 } 3100 #endif /* DDB */ 3101