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 pageq mutex is required when adding or removing a page from a 67 * page queue (vm_page_queue[]), regardless of other mutexes or the 68 * busy state of a page. 69 * 70 * - The object mutex is held when inserting or removing 71 * pages from an object (vm_page_insert() or vm_page_remove()). 72 * 73 */ 74 75 /* 76 * Resident memory management module. 77 */ 78 79 #include <sys/cdefs.h> 80 __FBSDID("$FreeBSD$"); 81 82 #include "opt_vm.h" 83 84 #include <sys/param.h> 85 #include <sys/systm.h> 86 #include <sys/lock.h> 87 #include <sys/kernel.h> 88 #include <sys/limits.h> 89 #include <sys/malloc.h> 90 #include <sys/msgbuf.h> 91 #include <sys/mutex.h> 92 #include <sys/proc.h> 93 #include <sys/sysctl.h> 94 #include <sys/vmmeter.h> 95 #include <sys/vnode.h> 96 97 #include <vm/vm.h> 98 #include <vm/pmap.h> 99 #include <vm/vm_param.h> 100 #include <vm/vm_kern.h> 101 #include <vm/vm_object.h> 102 #include <vm/vm_page.h> 103 #include <vm/vm_pageout.h> 104 #include <vm/vm_pager.h> 105 #include <vm/vm_phys.h> 106 #include <vm/vm_reserv.h> 107 #include <vm/vm_extern.h> 108 #include <vm/uma.h> 109 #include <vm/uma_int.h> 110 111 #include <machine/md_var.h> 112 113 /* 114 * Associated with page of user-allocatable memory is a 115 * page structure. 116 */ 117 118 struct vpgqueues vm_page_queues[PQ_COUNT]; 119 struct vpglocks vm_page_queue_lock; 120 struct vpglocks vm_page_queue_free_lock; 121 122 struct vpglocks pa_lock[PA_LOCK_COUNT]; 123 124 vm_page_t vm_page_array; 125 long vm_page_array_size; 126 long first_page; 127 int vm_page_zero_count; 128 129 static int boot_pages = UMA_BOOT_PAGES; 130 TUNABLE_INT("vm.boot_pages", &boot_pages); 131 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RD, &boot_pages, 0, 132 "number of pages allocated for bootstrapping the VM system"); 133 134 static int pa_tryrelock_restart; 135 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD, 136 &pa_tryrelock_restart, 0, "Number of tryrelock restarts"); 137 138 static uma_zone_t fakepg_zone; 139 140 static struct vnode *vm_page_alloc_init(vm_page_t m); 141 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits); 142 static void vm_page_queue_remove(int queue, vm_page_t m); 143 static void vm_page_enqueue(int queue, vm_page_t m); 144 static void vm_page_init_fakepg(void *dummy); 145 146 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init_fakepg, NULL); 147 148 static void 149 vm_page_init_fakepg(void *dummy) 150 { 151 152 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL, 153 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM); 154 } 155 156 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */ 157 #if PAGE_SIZE == 32768 158 #ifdef CTASSERT 159 CTASSERT(sizeof(u_long) >= 8); 160 #endif 161 #endif 162 163 /* 164 * Try to acquire a physical address lock while a pmap is locked. If we 165 * fail to trylock we unlock and lock the pmap directly and cache the 166 * locked pa in *locked. The caller should then restart their loop in case 167 * the virtual to physical mapping has changed. 168 */ 169 int 170 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked) 171 { 172 vm_paddr_t lockpa; 173 174 lockpa = *locked; 175 *locked = pa; 176 if (lockpa) { 177 PA_LOCK_ASSERT(lockpa, MA_OWNED); 178 if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa)) 179 return (0); 180 PA_UNLOCK(lockpa); 181 } 182 if (PA_TRYLOCK(pa)) 183 return (0); 184 PMAP_UNLOCK(pmap); 185 atomic_add_int(&pa_tryrelock_restart, 1); 186 PA_LOCK(pa); 187 PMAP_LOCK(pmap); 188 return (EAGAIN); 189 } 190 191 /* 192 * vm_set_page_size: 193 * 194 * Sets the page size, perhaps based upon the memory 195 * size. Must be called before any use of page-size 196 * dependent functions. 197 */ 198 void 199 vm_set_page_size(void) 200 { 201 if (cnt.v_page_size == 0) 202 cnt.v_page_size = PAGE_SIZE; 203 if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0) 204 panic("vm_set_page_size: page size not a power of two"); 205 } 206 207 /* 208 * vm_page_blacklist_lookup: 209 * 210 * See if a physical address in this page has been listed 211 * in the blacklist tunable. Entries in the tunable are 212 * separated by spaces or commas. If an invalid integer is 213 * encountered then the rest of the string is skipped. 214 */ 215 static int 216 vm_page_blacklist_lookup(char *list, vm_paddr_t pa) 217 { 218 vm_paddr_t bad; 219 char *cp, *pos; 220 221 for (pos = list; *pos != '\0'; pos = cp) { 222 bad = strtoq(pos, &cp, 0); 223 if (*cp != '\0') { 224 if (*cp == ' ' || *cp == ',') { 225 cp++; 226 if (cp == pos) 227 continue; 228 } else 229 break; 230 } 231 if (pa == trunc_page(bad)) 232 return (1); 233 } 234 return (0); 235 } 236 237 /* 238 * vm_page_startup: 239 * 240 * Initializes the resident memory module. 241 * 242 * Allocates memory for the page cells, and 243 * for the object/offset-to-page hash table headers. 244 * Each page cell is initialized and placed on the free list. 245 */ 246 vm_offset_t 247 vm_page_startup(vm_offset_t vaddr) 248 { 249 vm_offset_t mapped; 250 vm_paddr_t page_range; 251 vm_paddr_t new_end; 252 int i; 253 vm_paddr_t pa; 254 vm_paddr_t last_pa; 255 char *list; 256 257 /* the biggest memory array is the second group of pages */ 258 vm_paddr_t end; 259 vm_paddr_t biggestsize; 260 vm_paddr_t low_water, high_water; 261 int biggestone; 262 263 biggestsize = 0; 264 biggestone = 0; 265 vaddr = round_page(vaddr); 266 267 for (i = 0; phys_avail[i + 1]; i += 2) { 268 phys_avail[i] = round_page(phys_avail[i]); 269 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 270 } 271 272 low_water = phys_avail[0]; 273 high_water = phys_avail[1]; 274 275 for (i = 0; phys_avail[i + 1]; i += 2) { 276 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i]; 277 278 if (size > biggestsize) { 279 biggestone = i; 280 biggestsize = size; 281 } 282 if (phys_avail[i] < low_water) 283 low_water = phys_avail[i]; 284 if (phys_avail[i + 1] > high_water) 285 high_water = phys_avail[i + 1]; 286 } 287 288 #ifdef XEN 289 low_water = 0; 290 #endif 291 292 end = phys_avail[biggestone+1]; 293 294 /* 295 * Initialize the page and queue locks. 296 */ 297 mtx_init(&vm_page_queue_mtx, "vm page queue", NULL, MTX_DEF | 298 MTX_RECURSE); 299 mtx_init(&vm_page_queue_free_mtx, "vm page free queue", NULL, MTX_DEF); 300 for (i = 0; i < PA_LOCK_COUNT; i++) 301 mtx_init(&pa_lock[i].data, "vm page", NULL, MTX_DEF); 302 303 /* 304 * Initialize the queue headers for the hold queue, the active queue, 305 * and the inactive queue. 306 */ 307 for (i = 0; i < PQ_COUNT; i++) 308 TAILQ_INIT(&vm_page_queues[i].pl); 309 vm_page_queues[PQ_INACTIVE].cnt = &cnt.v_inactive_count; 310 vm_page_queues[PQ_ACTIVE].cnt = &cnt.v_active_count; 311 vm_page_queues[PQ_HOLD].cnt = &cnt.v_active_count; 312 313 /* 314 * Allocate memory for use when boot strapping the kernel memory 315 * allocator. 316 */ 317 new_end = end - (boot_pages * UMA_SLAB_SIZE); 318 new_end = trunc_page(new_end); 319 mapped = pmap_map(&vaddr, new_end, end, 320 VM_PROT_READ | VM_PROT_WRITE); 321 bzero((void *)mapped, end - new_end); 322 uma_startup((void *)mapped, boot_pages); 323 324 #if defined(__amd64__) || defined(__i386__) || defined(__arm__) || \ 325 defined(__mips__) 326 /* 327 * Allocate a bitmap to indicate that a random physical page 328 * needs to be included in a minidump. 329 * 330 * The amd64 port needs this to indicate which direct map pages 331 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 332 * 333 * However, i386 still needs this workspace internally within the 334 * minidump code. In theory, they are not needed on i386, but are 335 * included should the sf_buf code decide to use them. 336 */ 337 last_pa = 0; 338 for (i = 0; dump_avail[i + 1] != 0; i += 2) 339 if (dump_avail[i + 1] > last_pa) 340 last_pa = dump_avail[i + 1]; 341 page_range = last_pa / PAGE_SIZE; 342 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 343 new_end -= vm_page_dump_size; 344 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, 345 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); 346 bzero((void *)vm_page_dump, vm_page_dump_size); 347 #endif 348 #ifdef __amd64__ 349 /* 350 * Request that the physical pages underlying the message buffer be 351 * included in a crash dump. Since the message buffer is accessed 352 * through the direct map, they are not automatically included. 353 */ 354 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); 355 last_pa = pa + round_page(msgbufsize); 356 while (pa < last_pa) { 357 dump_add_page(pa); 358 pa += PAGE_SIZE; 359 } 360 #endif 361 /* 362 * Compute the number of pages of memory that will be available for 363 * use (taking into account the overhead of a page structure per 364 * page). 365 */ 366 first_page = low_water / PAGE_SIZE; 367 #ifdef VM_PHYSSEG_SPARSE 368 page_range = 0; 369 for (i = 0; phys_avail[i + 1] != 0; i += 2) 370 page_range += atop(phys_avail[i + 1] - phys_avail[i]); 371 #elif defined(VM_PHYSSEG_DENSE) 372 page_range = high_water / PAGE_SIZE - first_page; 373 #else 374 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 375 #endif 376 end = new_end; 377 378 /* 379 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 380 */ 381 vaddr += PAGE_SIZE; 382 383 /* 384 * Initialize the mem entry structures now, and put them in the free 385 * queue. 386 */ 387 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 388 mapped = pmap_map(&vaddr, new_end, end, 389 VM_PROT_READ | VM_PROT_WRITE); 390 vm_page_array = (vm_page_t) mapped; 391 #if VM_NRESERVLEVEL > 0 392 /* 393 * Allocate memory for the reservation management system's data 394 * structures. 395 */ 396 new_end = vm_reserv_startup(&vaddr, new_end, high_water); 397 #endif 398 #if defined(__amd64__) || defined(__mips__) 399 /* 400 * pmap_map on amd64 and mips can come out of the direct-map, not kvm 401 * like i386, so the pages must be tracked for a crashdump to include 402 * this data. This includes the vm_page_array and the early UMA 403 * bootstrap pages. 404 */ 405 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE) 406 dump_add_page(pa); 407 #endif 408 phys_avail[biggestone + 1] = new_end; 409 410 /* 411 * Clear all of the page structures 412 */ 413 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 414 for (i = 0; i < page_range; i++) 415 vm_page_array[i].order = VM_NFREEORDER; 416 vm_page_array_size = page_range; 417 418 /* 419 * Initialize the physical memory allocator. 420 */ 421 vm_phys_init(); 422 423 /* 424 * Add every available physical page that is not blacklisted to 425 * the free lists. 426 */ 427 cnt.v_page_count = 0; 428 cnt.v_free_count = 0; 429 list = getenv("vm.blacklist"); 430 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 431 pa = phys_avail[i]; 432 last_pa = phys_avail[i + 1]; 433 while (pa < last_pa) { 434 if (list != NULL && 435 vm_page_blacklist_lookup(list, pa)) 436 printf("Skipping page with pa 0x%jx\n", 437 (uintmax_t)pa); 438 else 439 vm_phys_add_page(pa); 440 pa += PAGE_SIZE; 441 } 442 } 443 freeenv(list); 444 #if VM_NRESERVLEVEL > 0 445 /* 446 * Initialize the reservation management system. 447 */ 448 vm_reserv_init(); 449 #endif 450 return (vaddr); 451 } 452 453 void 454 vm_page_reference(vm_page_t m) 455 { 456 457 vm_page_aflag_set(m, PGA_REFERENCED); 458 } 459 460 void 461 vm_page_busy(vm_page_t m) 462 { 463 464 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 465 KASSERT((m->oflags & VPO_BUSY) == 0, 466 ("vm_page_busy: page already busy!!!")); 467 m->oflags |= VPO_BUSY; 468 } 469 470 /* 471 * vm_page_flash: 472 * 473 * wakeup anyone waiting for the page. 474 */ 475 void 476 vm_page_flash(vm_page_t m) 477 { 478 479 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 480 if (m->oflags & VPO_WANTED) { 481 m->oflags &= ~VPO_WANTED; 482 wakeup(m); 483 } 484 } 485 486 /* 487 * vm_page_wakeup: 488 * 489 * clear the VPO_BUSY flag and wakeup anyone waiting for the 490 * page. 491 * 492 */ 493 void 494 vm_page_wakeup(vm_page_t m) 495 { 496 497 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 498 KASSERT(m->oflags & VPO_BUSY, ("vm_page_wakeup: page not busy!!!")); 499 m->oflags &= ~VPO_BUSY; 500 vm_page_flash(m); 501 } 502 503 void 504 vm_page_io_start(vm_page_t m) 505 { 506 507 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 508 m->busy++; 509 } 510 511 void 512 vm_page_io_finish(vm_page_t m) 513 { 514 515 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 516 KASSERT(m->busy > 0, ("vm_page_io_finish: page %p is not busy", m)); 517 m->busy--; 518 if (m->busy == 0) 519 vm_page_flash(m); 520 } 521 522 /* 523 * Keep page from being freed by the page daemon 524 * much of the same effect as wiring, except much lower 525 * overhead and should be used only for *very* temporary 526 * holding ("wiring"). 527 */ 528 void 529 vm_page_hold(vm_page_t mem) 530 { 531 532 vm_page_lock_assert(mem, MA_OWNED); 533 mem->hold_count++; 534 } 535 536 void 537 vm_page_unhold(vm_page_t mem) 538 { 539 540 vm_page_lock_assert(mem, MA_OWNED); 541 --mem->hold_count; 542 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!")); 543 if (mem->hold_count == 0 && mem->queue == PQ_HOLD) 544 vm_page_free_toq(mem); 545 } 546 547 /* 548 * vm_page_unhold_pages: 549 * 550 * Unhold each of the pages that is referenced by the given array. 551 */ 552 void 553 vm_page_unhold_pages(vm_page_t *ma, int count) 554 { 555 struct mtx *mtx, *new_mtx; 556 557 mtx = NULL; 558 for (; count != 0; count--) { 559 /* 560 * Avoid releasing and reacquiring the same page lock. 561 */ 562 new_mtx = vm_page_lockptr(*ma); 563 if (mtx != new_mtx) { 564 if (mtx != NULL) 565 mtx_unlock(mtx); 566 mtx = new_mtx; 567 mtx_lock(mtx); 568 } 569 vm_page_unhold(*ma); 570 ma++; 571 } 572 if (mtx != NULL) 573 mtx_unlock(mtx); 574 } 575 576 vm_page_t 577 PHYS_TO_VM_PAGE(vm_paddr_t pa) 578 { 579 vm_page_t m; 580 581 #ifdef VM_PHYSSEG_SPARSE 582 m = vm_phys_paddr_to_vm_page(pa); 583 if (m == NULL) 584 m = vm_phys_fictitious_to_vm_page(pa); 585 return (m); 586 #elif defined(VM_PHYSSEG_DENSE) 587 long pi; 588 589 pi = atop(pa); 590 if (pi >= first_page && (pi - first_page) < vm_page_array_size) { 591 m = &vm_page_array[pi - first_page]; 592 return (m); 593 } 594 return (vm_phys_fictitious_to_vm_page(pa)); 595 #else 596 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 597 #endif 598 } 599 600 /* 601 * vm_page_getfake: 602 * 603 * Create a fictitious page with the specified physical address and 604 * memory attribute. The memory attribute is the only the machine- 605 * dependent aspect of a fictitious page that must be initialized. 606 */ 607 vm_page_t 608 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr) 609 { 610 vm_page_t m; 611 612 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO); 613 vm_page_initfake(m, paddr, memattr); 614 return (m); 615 } 616 617 void 618 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 619 { 620 621 if ((m->flags & PG_FICTITIOUS) != 0) { 622 /* 623 * The page's memattr might have changed since the 624 * previous initialization. Update the pmap to the 625 * new memattr. 626 */ 627 goto memattr; 628 } 629 m->phys_addr = paddr; 630 m->queue = PQ_NONE; 631 /* Fictitious pages don't use "segind". */ 632 m->flags = PG_FICTITIOUS; 633 /* Fictitious pages don't use "order" or "pool". */ 634 m->oflags = VPO_BUSY | VPO_UNMANAGED; 635 m->wire_count = 1; 636 memattr: 637 pmap_page_set_memattr(m, memattr); 638 } 639 640 /* 641 * vm_page_putfake: 642 * 643 * Release a fictitious page. 644 */ 645 void 646 vm_page_putfake(vm_page_t m) 647 { 648 649 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m)); 650 KASSERT((m->flags & PG_FICTITIOUS) != 0, 651 ("vm_page_putfake: bad page %p", m)); 652 uma_zfree(fakepg_zone, m); 653 } 654 655 /* 656 * vm_page_updatefake: 657 * 658 * Update the given fictitious page to the specified physical address and 659 * memory attribute. 660 */ 661 void 662 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 663 { 664 665 KASSERT((m->flags & PG_FICTITIOUS) != 0, 666 ("vm_page_updatefake: bad page %p", m)); 667 m->phys_addr = paddr; 668 pmap_page_set_memattr(m, memattr); 669 } 670 671 /* 672 * vm_page_free: 673 * 674 * Free a page. 675 */ 676 void 677 vm_page_free(vm_page_t m) 678 { 679 680 m->flags &= ~PG_ZERO; 681 vm_page_free_toq(m); 682 } 683 684 /* 685 * vm_page_free_zero: 686 * 687 * Free a page to the zerod-pages queue 688 */ 689 void 690 vm_page_free_zero(vm_page_t m) 691 { 692 693 m->flags |= PG_ZERO; 694 vm_page_free_toq(m); 695 } 696 697 /* 698 * Unbusy and handle the page queueing for a page from the VOP_GETPAGES() 699 * array which is not the request page. 700 */ 701 void 702 vm_page_readahead_finish(vm_page_t m) 703 { 704 705 if (m->valid != 0) { 706 /* 707 * Since the page is not the requested page, whether 708 * it should be activated or deactivated is not 709 * obvious. Empirical results have shown that 710 * deactivating the page is usually the best choice, 711 * unless the page is wanted by another thread. 712 */ 713 if (m->oflags & VPO_WANTED) { 714 vm_page_lock(m); 715 vm_page_activate(m); 716 vm_page_unlock(m); 717 } else { 718 vm_page_lock(m); 719 vm_page_deactivate(m); 720 vm_page_unlock(m); 721 } 722 vm_page_wakeup(m); 723 } else { 724 /* 725 * Free the completely invalid page. Such page state 726 * occurs due to the short read operation which did 727 * not covered our page at all, or in case when a read 728 * error happens. 729 */ 730 vm_page_lock(m); 731 vm_page_free(m); 732 vm_page_unlock(m); 733 } 734 } 735 736 /* 737 * vm_page_sleep: 738 * 739 * Sleep and release the page and page queues locks. 740 * 741 * The object containing the given page must be locked. 742 */ 743 void 744 vm_page_sleep(vm_page_t m, const char *msg) 745 { 746 747 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 748 if (mtx_owned(&vm_page_queue_mtx)) 749 vm_page_unlock_queues(); 750 if (mtx_owned(vm_page_lockptr(m))) 751 vm_page_unlock(m); 752 753 /* 754 * It's possible that while we sleep, the page will get 755 * unbusied and freed. If we are holding the object 756 * lock, we will assume we hold a reference to the object 757 * such that even if m->object changes, we can re-lock 758 * it. 759 */ 760 m->oflags |= VPO_WANTED; 761 msleep(m, VM_OBJECT_MTX(m->object), PVM, msg, 0); 762 } 763 764 /* 765 * vm_page_dirty_KBI: [ internal use only ] 766 * 767 * Set all bits in the page's dirty field. 768 * 769 * The object containing the specified page must be locked if the 770 * call is made from the machine-independent layer. 771 * 772 * See vm_page_clear_dirty_mask(). 773 * 774 * This function should only be called by vm_page_dirty(). 775 */ 776 void 777 vm_page_dirty_KBI(vm_page_t m) 778 { 779 780 /* These assertions refer to this operation by its public name. */ 781 KASSERT((m->flags & PG_CACHED) == 0, 782 ("vm_page_dirty: page in cache!")); 783 KASSERT(!VM_PAGE_IS_FREE(m), 784 ("vm_page_dirty: page is free!")); 785 KASSERT(m->valid == VM_PAGE_BITS_ALL, 786 ("vm_page_dirty: page is invalid!")); 787 m->dirty = VM_PAGE_BITS_ALL; 788 } 789 790 /* 791 * vm_page_splay: 792 * 793 * Implements Sleator and Tarjan's top-down splay algorithm. Returns 794 * the vm_page containing the given pindex. If, however, that 795 * pindex is not found in the vm_object, returns a vm_page that is 796 * adjacent to the pindex, coming before or after it. 797 */ 798 vm_page_t 799 vm_page_splay(vm_pindex_t pindex, vm_page_t root) 800 { 801 struct vm_page dummy; 802 vm_page_t lefttreemax, righttreemin, y; 803 804 if (root == NULL) 805 return (root); 806 lefttreemax = righttreemin = &dummy; 807 for (;; root = y) { 808 if (pindex < root->pindex) { 809 if ((y = root->left) == NULL) 810 break; 811 if (pindex < y->pindex) { 812 /* Rotate right. */ 813 root->left = y->right; 814 y->right = root; 815 root = y; 816 if ((y = root->left) == NULL) 817 break; 818 } 819 /* Link into the new root's right tree. */ 820 righttreemin->left = root; 821 righttreemin = root; 822 } else if (pindex > root->pindex) { 823 if ((y = root->right) == NULL) 824 break; 825 if (pindex > y->pindex) { 826 /* Rotate left. */ 827 root->right = y->left; 828 y->left = root; 829 root = y; 830 if ((y = root->right) == NULL) 831 break; 832 } 833 /* Link into the new root's left tree. */ 834 lefttreemax->right = root; 835 lefttreemax = root; 836 } else 837 break; 838 } 839 /* Assemble the new root. */ 840 lefttreemax->right = root->left; 841 righttreemin->left = root->right; 842 root->left = dummy.right; 843 root->right = dummy.left; 844 return (root); 845 } 846 847 /* 848 * vm_page_insert: [ internal use only ] 849 * 850 * Inserts the given mem entry into the object and object list. 851 * 852 * The pagetables are not updated but will presumably fault the page 853 * in if necessary, or if a kernel page the caller will at some point 854 * enter the page into the kernel's pmap. We are not allowed to block 855 * here so we *can't* do this anyway. 856 * 857 * The object and page must be locked. 858 * This routine may not block. 859 */ 860 void 861 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 862 { 863 vm_page_t root; 864 865 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 866 if (m->object != NULL) 867 panic("vm_page_insert: page already inserted"); 868 869 /* 870 * Record the object/offset pair in this page 871 */ 872 m->object = object; 873 m->pindex = pindex; 874 875 /* 876 * Now link into the object's ordered list of backed pages. 877 */ 878 root = object->root; 879 if (root == NULL) { 880 m->left = NULL; 881 m->right = NULL; 882 TAILQ_INSERT_TAIL(&object->memq, m, listq); 883 } else { 884 root = vm_page_splay(pindex, root); 885 if (pindex < root->pindex) { 886 m->left = root->left; 887 m->right = root; 888 root->left = NULL; 889 TAILQ_INSERT_BEFORE(root, m, listq); 890 } else if (pindex == root->pindex) 891 panic("vm_page_insert: offset already allocated"); 892 else { 893 m->right = root->right; 894 m->left = root; 895 root->right = NULL; 896 TAILQ_INSERT_AFTER(&object->memq, root, m, listq); 897 } 898 } 899 object->root = m; 900 901 /* 902 * show that the object has one more resident page. 903 */ 904 object->resident_page_count++; 905 /* 906 * Hold the vnode until the last page is released. 907 */ 908 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 909 vhold((struct vnode *)object->handle); 910 911 /* 912 * Since we are inserting a new and possibly dirty page, 913 * update the object's OBJ_MIGHTBEDIRTY flag. 914 */ 915 if (pmap_page_is_write_mapped(m)) 916 vm_object_set_writeable_dirty(object); 917 } 918 919 /* 920 * vm_page_remove: 921 * NOTE: used by device pager as well -wfj 922 * 923 * Removes the given mem entry from the object/offset-page 924 * table and the object page list, but do not invalidate/terminate 925 * the backing store. 926 * 927 * The object and page must be locked. 928 * The underlying pmap entry (if any) is NOT removed here. 929 * This routine may not block. 930 */ 931 void 932 vm_page_remove(vm_page_t m) 933 { 934 vm_object_t object; 935 vm_page_t next, prev, root; 936 937 if ((m->oflags & VPO_UNMANAGED) == 0) 938 vm_page_lock_assert(m, MA_OWNED); 939 if ((object = m->object) == NULL) 940 return; 941 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 942 if (m->oflags & VPO_BUSY) { 943 m->oflags &= ~VPO_BUSY; 944 vm_page_flash(m); 945 } 946 947 /* 948 * Now remove from the object's list of backed pages. 949 */ 950 if ((next = TAILQ_NEXT(m, listq)) != NULL && next->left == m) { 951 /* 952 * Since the page's successor in the list is also its parent 953 * in the tree, its right subtree must be empty. 954 */ 955 next->left = m->left; 956 KASSERT(m->right == NULL, 957 ("vm_page_remove: page %p has right child", m)); 958 } else if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL && 959 prev->right == m) { 960 /* 961 * Since the page's predecessor in the list is also its parent 962 * in the tree, its left subtree must be empty. 963 */ 964 KASSERT(m->left == NULL, 965 ("vm_page_remove: page %p has left child", m)); 966 prev->right = m->right; 967 } else { 968 if (m != object->root) 969 vm_page_splay(m->pindex, object->root); 970 if (m->left == NULL) 971 root = m->right; 972 else if (m->right == NULL) 973 root = m->left; 974 else { 975 /* 976 * Move the page's successor to the root, because 977 * pages are usually removed in ascending order. 978 */ 979 if (m->right != next) 980 vm_page_splay(m->pindex, m->right); 981 next->left = m->left; 982 root = next; 983 } 984 object->root = root; 985 } 986 TAILQ_REMOVE(&object->memq, m, listq); 987 988 /* 989 * And show that the object has one fewer resident page. 990 */ 991 object->resident_page_count--; 992 /* 993 * The vnode may now be recycled. 994 */ 995 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 996 vdrop((struct vnode *)object->handle); 997 998 m->object = NULL; 999 } 1000 1001 /* 1002 * vm_page_lookup: 1003 * 1004 * Returns the page associated with the object/offset 1005 * pair specified; if none is found, NULL is returned. 1006 * 1007 * The object must be locked. 1008 * This routine may not block. 1009 * This is a critical path routine 1010 */ 1011 vm_page_t 1012 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1013 { 1014 vm_page_t m; 1015 1016 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1017 if ((m = object->root) != NULL && m->pindex != pindex) { 1018 m = vm_page_splay(pindex, m); 1019 if ((object->root = m)->pindex != pindex) 1020 m = NULL; 1021 } 1022 return (m); 1023 } 1024 1025 /* 1026 * vm_page_find_least: 1027 * 1028 * Returns the page associated with the object with least pindex 1029 * greater than or equal to the parameter pindex, or NULL. 1030 * 1031 * The object must be locked. 1032 * The routine may not block. 1033 */ 1034 vm_page_t 1035 vm_page_find_least(vm_object_t object, vm_pindex_t pindex) 1036 { 1037 vm_page_t m; 1038 1039 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1040 if ((m = TAILQ_FIRST(&object->memq)) != NULL) { 1041 if (m->pindex < pindex) { 1042 m = vm_page_splay(pindex, object->root); 1043 if ((object->root = m)->pindex < pindex) 1044 m = TAILQ_NEXT(m, listq); 1045 } 1046 } 1047 return (m); 1048 } 1049 1050 /* 1051 * Returns the given page's successor (by pindex) within the object if it is 1052 * resident; if none is found, NULL is returned. 1053 * 1054 * The object must be locked. 1055 */ 1056 vm_page_t 1057 vm_page_next(vm_page_t m) 1058 { 1059 vm_page_t next; 1060 1061 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1062 if ((next = TAILQ_NEXT(m, listq)) != NULL && 1063 next->pindex != m->pindex + 1) 1064 next = NULL; 1065 return (next); 1066 } 1067 1068 /* 1069 * Returns the given page's predecessor (by pindex) within the object if it is 1070 * resident; if none is found, NULL is returned. 1071 * 1072 * The object must be locked. 1073 */ 1074 vm_page_t 1075 vm_page_prev(vm_page_t m) 1076 { 1077 vm_page_t prev; 1078 1079 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1080 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL && 1081 prev->pindex != m->pindex - 1) 1082 prev = NULL; 1083 return (prev); 1084 } 1085 1086 /* 1087 * vm_page_rename: 1088 * 1089 * Move the given memory entry from its 1090 * current object to the specified target object/offset. 1091 * 1092 * The object must be locked. 1093 * This routine may not block. 1094 * 1095 * Note: swap associated with the page must be invalidated by the move. We 1096 * have to do this for several reasons: (1) we aren't freeing the 1097 * page, (2) we are dirtying the page, (3) the VM system is probably 1098 * moving the page from object A to B, and will then later move 1099 * the backing store from A to B and we can't have a conflict. 1100 * 1101 * Note: we *always* dirty the page. It is necessary both for the 1102 * fact that we moved it, and because we may be invalidating 1103 * swap. If the page is on the cache, we have to deactivate it 1104 * or vm_page_dirty() will panic. Dirty pages are not allowed 1105 * on the cache. 1106 */ 1107 void 1108 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1109 { 1110 1111 vm_page_remove(m); 1112 vm_page_insert(m, new_object, new_pindex); 1113 vm_page_dirty(m); 1114 } 1115 1116 /* 1117 * Convert all of the given object's cached pages that have a 1118 * pindex within the given range into free pages. If the value 1119 * zero is given for "end", then the range's upper bound is 1120 * infinity. If the given object is backed by a vnode and it 1121 * transitions from having one or more cached pages to none, the 1122 * vnode's hold count is reduced. 1123 */ 1124 void 1125 vm_page_cache_free(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 1126 { 1127 vm_page_t m, m_next; 1128 boolean_t empty; 1129 1130 mtx_lock(&vm_page_queue_free_mtx); 1131 if (__predict_false(object->cache == NULL)) { 1132 mtx_unlock(&vm_page_queue_free_mtx); 1133 return; 1134 } 1135 m = object->cache = vm_page_splay(start, object->cache); 1136 if (m->pindex < start) { 1137 if (m->right == NULL) 1138 m = NULL; 1139 else { 1140 m_next = vm_page_splay(start, m->right); 1141 m_next->left = m; 1142 m->right = NULL; 1143 m = object->cache = m_next; 1144 } 1145 } 1146 1147 /* 1148 * At this point, "m" is either (1) a reference to the page 1149 * with the least pindex that is greater than or equal to 1150 * "start" or (2) NULL. 1151 */ 1152 for (; m != NULL && (m->pindex < end || end == 0); m = m_next) { 1153 /* 1154 * Find "m"'s successor and remove "m" from the 1155 * object's cache. 1156 */ 1157 if (m->right == NULL) { 1158 object->cache = m->left; 1159 m_next = NULL; 1160 } else { 1161 m_next = vm_page_splay(start, m->right); 1162 m_next->left = m->left; 1163 object->cache = m_next; 1164 } 1165 /* Convert "m" to a free page. */ 1166 m->object = NULL; 1167 m->valid = 0; 1168 /* Clear PG_CACHED and set PG_FREE. */ 1169 m->flags ^= PG_CACHED | PG_FREE; 1170 KASSERT((m->flags & (PG_CACHED | PG_FREE)) == PG_FREE, 1171 ("vm_page_cache_free: page %p has inconsistent flags", m)); 1172 cnt.v_cache_count--; 1173 cnt.v_free_count++; 1174 } 1175 empty = object->cache == NULL; 1176 mtx_unlock(&vm_page_queue_free_mtx); 1177 if (object->type == OBJT_VNODE && empty) 1178 vdrop(object->handle); 1179 } 1180 1181 /* 1182 * Returns the cached page that is associated with the given 1183 * object and offset. If, however, none exists, returns NULL. 1184 * 1185 * The free page queue must be locked. 1186 */ 1187 static inline vm_page_t 1188 vm_page_cache_lookup(vm_object_t object, vm_pindex_t pindex) 1189 { 1190 vm_page_t m; 1191 1192 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1193 if ((m = object->cache) != NULL && m->pindex != pindex) { 1194 m = vm_page_splay(pindex, m); 1195 if ((object->cache = m)->pindex != pindex) 1196 m = NULL; 1197 } 1198 return (m); 1199 } 1200 1201 /* 1202 * Remove the given cached page from its containing object's 1203 * collection of cached pages. 1204 * 1205 * The free page queue must be locked. 1206 */ 1207 static void 1208 vm_page_cache_remove(vm_page_t m) 1209 { 1210 vm_object_t object; 1211 vm_page_t root; 1212 1213 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1214 KASSERT((m->flags & PG_CACHED) != 0, 1215 ("vm_page_cache_remove: page %p is not cached", m)); 1216 object = m->object; 1217 if (m != object->cache) { 1218 root = vm_page_splay(m->pindex, object->cache); 1219 KASSERT(root == m, 1220 ("vm_page_cache_remove: page %p is not cached in object %p", 1221 m, object)); 1222 } 1223 if (m->left == NULL) 1224 root = m->right; 1225 else if (m->right == NULL) 1226 root = m->left; 1227 else { 1228 root = vm_page_splay(m->pindex, m->left); 1229 root->right = m->right; 1230 } 1231 object->cache = root; 1232 m->object = NULL; 1233 cnt.v_cache_count--; 1234 } 1235 1236 /* 1237 * Transfer all of the cached pages with offset greater than or 1238 * equal to 'offidxstart' from the original object's cache to the 1239 * new object's cache. However, any cached pages with offset 1240 * greater than or equal to the new object's size are kept in the 1241 * original object. Initially, the new object's cache must be 1242 * empty. Offset 'offidxstart' in the original object must 1243 * correspond to offset zero in the new object. 1244 * 1245 * The new object must be locked. 1246 */ 1247 void 1248 vm_page_cache_transfer(vm_object_t orig_object, vm_pindex_t offidxstart, 1249 vm_object_t new_object) 1250 { 1251 vm_page_t m, m_next; 1252 1253 /* 1254 * Insertion into an object's collection of cached pages 1255 * requires the object to be locked. In contrast, removal does 1256 * not. 1257 */ 1258 VM_OBJECT_LOCK_ASSERT(new_object, MA_OWNED); 1259 KASSERT(new_object->cache == NULL, 1260 ("vm_page_cache_transfer: object %p has cached pages", 1261 new_object)); 1262 mtx_lock(&vm_page_queue_free_mtx); 1263 if ((m = orig_object->cache) != NULL) { 1264 /* 1265 * Transfer all of the pages with offset greater than or 1266 * equal to 'offidxstart' from the original object's 1267 * cache to the new object's cache. 1268 */ 1269 m = vm_page_splay(offidxstart, m); 1270 if (m->pindex < offidxstart) { 1271 orig_object->cache = m; 1272 new_object->cache = m->right; 1273 m->right = NULL; 1274 } else { 1275 orig_object->cache = m->left; 1276 new_object->cache = m; 1277 m->left = NULL; 1278 } 1279 while ((m = new_object->cache) != NULL) { 1280 if ((m->pindex - offidxstart) >= new_object->size) { 1281 /* 1282 * Return all of the cached pages with 1283 * offset greater than or equal to the 1284 * new object's size to the original 1285 * object's cache. 1286 */ 1287 new_object->cache = m->left; 1288 m->left = orig_object->cache; 1289 orig_object->cache = m; 1290 break; 1291 } 1292 m_next = vm_page_splay(m->pindex, m->right); 1293 /* Update the page's object and offset. */ 1294 m->object = new_object; 1295 m->pindex -= offidxstart; 1296 if (m_next == NULL) 1297 break; 1298 m->right = NULL; 1299 m_next->left = m; 1300 new_object->cache = m_next; 1301 } 1302 KASSERT(new_object->cache == NULL || 1303 new_object->type == OBJT_SWAP, 1304 ("vm_page_cache_transfer: object %p's type is incompatible" 1305 " with cached pages", new_object)); 1306 } 1307 mtx_unlock(&vm_page_queue_free_mtx); 1308 } 1309 1310 /* 1311 * Returns TRUE if a cached page is associated with the given object and 1312 * offset, and FALSE otherwise. 1313 * 1314 * The object must be locked. 1315 */ 1316 boolean_t 1317 vm_page_is_cached(vm_object_t object, vm_pindex_t pindex) 1318 { 1319 vm_page_t m; 1320 1321 /* 1322 * Insertion into an object's collection of cached pages requires the 1323 * object to be locked. Therefore, if the object is locked and the 1324 * object's collection is empty, there is no need to acquire the free 1325 * page queues lock in order to prove that the specified page doesn't 1326 * exist. 1327 */ 1328 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1329 if (__predict_true(object->cache == NULL)) 1330 return (FALSE); 1331 mtx_lock(&vm_page_queue_free_mtx); 1332 m = vm_page_cache_lookup(object, pindex); 1333 mtx_unlock(&vm_page_queue_free_mtx); 1334 return (m != NULL); 1335 } 1336 1337 /* 1338 * vm_page_alloc: 1339 * 1340 * Allocate and return a page that is associated with the specified 1341 * object and offset pair. By default, this page has the flag VPO_BUSY 1342 * set. 1343 * 1344 * The caller must always specify an allocation class. 1345 * 1346 * allocation classes: 1347 * VM_ALLOC_NORMAL normal process request 1348 * VM_ALLOC_SYSTEM system *really* needs a page 1349 * VM_ALLOC_INTERRUPT interrupt time request 1350 * 1351 * optional allocation flags: 1352 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1353 * intends to allocate 1354 * VM_ALLOC_IFCACHED return page only if it is cached 1355 * VM_ALLOC_IFNOTCACHED return NULL, do not reactivate if the page 1356 * is cached 1357 * VM_ALLOC_NOBUSY do not set the flag VPO_BUSY on the page 1358 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1359 * VM_ALLOC_NOOBJ page is not associated with an object and 1360 * should not have the flag VPO_BUSY set 1361 * VM_ALLOC_WIRED wire the allocated page 1362 * VM_ALLOC_ZERO prefer a zeroed page 1363 * 1364 * This routine may not sleep. 1365 */ 1366 vm_page_t 1367 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 1368 { 1369 struct vnode *vp = NULL; 1370 vm_object_t m_object; 1371 vm_page_t m; 1372 int flags, req_class; 1373 1374 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0), 1375 ("vm_page_alloc: inconsistent object/req")); 1376 if (object != NULL) 1377 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1378 1379 req_class = req & VM_ALLOC_CLASS_MASK; 1380 1381 /* 1382 * The page daemon is allowed to dig deeper into the free page list. 1383 */ 1384 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1385 req_class = VM_ALLOC_SYSTEM; 1386 1387 mtx_lock(&vm_page_queue_free_mtx); 1388 if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved || 1389 (req_class == VM_ALLOC_SYSTEM && 1390 cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) || 1391 (req_class == VM_ALLOC_INTERRUPT && 1392 cnt.v_free_count + cnt.v_cache_count > 0)) { 1393 /* 1394 * Allocate from the free queue if the number of free pages 1395 * exceeds the minimum for the request class. 1396 */ 1397 if (object != NULL && 1398 (m = vm_page_cache_lookup(object, pindex)) != NULL) { 1399 if ((req & VM_ALLOC_IFNOTCACHED) != 0) { 1400 mtx_unlock(&vm_page_queue_free_mtx); 1401 return (NULL); 1402 } 1403 if (vm_phys_unfree_page(m)) 1404 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0); 1405 #if VM_NRESERVLEVEL > 0 1406 else if (!vm_reserv_reactivate_page(m)) 1407 #else 1408 else 1409 #endif 1410 panic("vm_page_alloc: cache page %p is missing" 1411 " from the free queue", m); 1412 } else if ((req & VM_ALLOC_IFCACHED) != 0) { 1413 mtx_unlock(&vm_page_queue_free_mtx); 1414 return (NULL); 1415 #if VM_NRESERVLEVEL > 0 1416 } else if (object == NULL || object->type == OBJT_DEVICE || 1417 object->type == OBJT_SG || 1418 (object->flags & OBJ_COLORED) == 0 || 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 (req & VM_ALLOC_NODUMP) 1484 flags |= PG_NODUMP; 1485 if (m->flags & PG_ZERO) { 1486 vm_page_zero_count--; 1487 if (req & VM_ALLOC_ZERO) 1488 flags = PG_ZERO; 1489 } 1490 m->flags = flags; 1491 mtx_unlock(&vm_page_queue_free_mtx); 1492 m->aflags = 0; 1493 if (object == NULL || object->type == OBJT_PHYS) 1494 m->oflags = VPO_UNMANAGED; 1495 else 1496 m->oflags = 0; 1497 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ)) == 0) 1498 m->oflags |= VPO_BUSY; 1499 if (req & VM_ALLOC_WIRED) { 1500 /* 1501 * The page lock is not required for wiring a page until that 1502 * page is inserted into the object. 1503 */ 1504 atomic_add_int(&cnt.v_wire_count, 1); 1505 m->wire_count = 1; 1506 } 1507 m->act_count = 0; 1508 1509 if (object != NULL) { 1510 /* Ignore device objects; the pager sets "memattr" for them. */ 1511 if (object->memattr != VM_MEMATTR_DEFAULT && 1512 object->type != OBJT_DEVICE && object->type != OBJT_SG) 1513 pmap_page_set_memattr(m, object->memattr); 1514 vm_page_insert(m, object, pindex); 1515 } else 1516 m->pindex = pindex; 1517 1518 /* 1519 * The following call to vdrop() must come after the above call 1520 * to vm_page_insert() in case both affect the same object and 1521 * vnode. Otherwise, the affected vnode's hold count could 1522 * temporarily become zero. 1523 */ 1524 if (vp != NULL) 1525 vdrop(vp); 1526 1527 /* 1528 * Don't wakeup too often - wakeup the pageout daemon when 1529 * we would be nearly out of memory. 1530 */ 1531 if (vm_paging_needed()) 1532 pagedaemon_wakeup(); 1533 1534 return (m); 1535 } 1536 1537 /* 1538 * vm_page_alloc_contig: 1539 * 1540 * Allocate a contiguous set of physical pages of the given size "npages" 1541 * from the free lists. All of the physical pages must be at or above 1542 * the given physical address "low" and below the given physical address 1543 * "high". The given value "alignment" determines the alignment of the 1544 * first physical page in the set. If the given value "boundary" is 1545 * non-zero, then the set of physical pages cannot cross any physical 1546 * address boundary that is a multiple of that value. Both "alignment" 1547 * and "boundary" must be a power of two. 1548 * 1549 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, 1550 * then the memory attribute setting for the physical pages is configured 1551 * to the object's memory attribute setting. Otherwise, the memory 1552 * attribute setting for the physical pages is configured to "memattr", 1553 * overriding the object's memory attribute setting. However, if the 1554 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the 1555 * memory attribute setting for the physical pages cannot be configured 1556 * to VM_MEMATTR_DEFAULT. 1557 * 1558 * The caller must always specify an allocation class. 1559 * 1560 * allocation classes: 1561 * VM_ALLOC_NORMAL normal process request 1562 * VM_ALLOC_SYSTEM system *really* needs a page 1563 * VM_ALLOC_INTERRUPT interrupt time request 1564 * 1565 * optional allocation flags: 1566 * VM_ALLOC_NOBUSY do not set the flag VPO_BUSY on the page 1567 * VM_ALLOC_NOOBJ page is not associated with an object and 1568 * should not have the flag VPO_BUSY set 1569 * VM_ALLOC_WIRED wire the allocated page 1570 * VM_ALLOC_ZERO prefer a zeroed page 1571 * 1572 * This routine may not sleep. 1573 */ 1574 vm_page_t 1575 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, 1576 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 1577 vm_paddr_t boundary, vm_memattr_t memattr) 1578 { 1579 struct vnode *drop; 1580 vm_page_t deferred_vdrop_list, m, m_ret; 1581 u_int flags, oflags; 1582 int req_class; 1583 1584 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0), 1585 ("vm_page_alloc_contig: inconsistent object/req")); 1586 if (object != NULL) { 1587 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1588 KASSERT(object->type == OBJT_PHYS, 1589 ("vm_page_alloc_contig: object %p isn't OBJT_PHYS", 1590 object)); 1591 } 1592 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 1593 req_class = req & VM_ALLOC_CLASS_MASK; 1594 1595 /* 1596 * The page daemon is allowed to dig deeper into the free page list. 1597 */ 1598 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1599 req_class = VM_ALLOC_SYSTEM; 1600 1601 deferred_vdrop_list = NULL; 1602 mtx_lock(&vm_page_queue_free_mtx); 1603 if (cnt.v_free_count + cnt.v_cache_count >= npages + 1604 cnt.v_free_reserved || (req_class == VM_ALLOC_SYSTEM && 1605 cnt.v_free_count + cnt.v_cache_count >= npages + 1606 cnt.v_interrupt_free_min) || (req_class == VM_ALLOC_INTERRUPT && 1607 cnt.v_free_count + cnt.v_cache_count >= npages)) { 1608 #if VM_NRESERVLEVEL > 0 1609 retry: 1610 if (object == NULL || (object->flags & OBJ_COLORED) == 0 || 1611 (m_ret = vm_reserv_alloc_contig(object, pindex, npages, 1612 low, high, alignment, boundary)) == NULL) 1613 #endif 1614 m_ret = vm_phys_alloc_contig(npages, low, high, 1615 alignment, boundary); 1616 } else { 1617 mtx_unlock(&vm_page_queue_free_mtx); 1618 atomic_add_int(&vm_pageout_deficit, npages); 1619 pagedaemon_wakeup(); 1620 return (NULL); 1621 } 1622 if (m_ret != NULL) 1623 for (m = m_ret; m < &m_ret[npages]; m++) { 1624 drop = vm_page_alloc_init(m); 1625 if (drop != NULL) { 1626 /* 1627 * Enqueue the vnode for deferred vdrop(). 1628 * 1629 * Once the pages are removed from the free 1630 * page list, "pageq" can be safely abused to 1631 * construct a short-lived list of vnodes. 1632 */ 1633 m->pageq.tqe_prev = (void *)drop; 1634 m->pageq.tqe_next = deferred_vdrop_list; 1635 deferred_vdrop_list = m; 1636 } 1637 } 1638 else { 1639 #if VM_NRESERVLEVEL > 0 1640 if (vm_reserv_reclaim_contig(npages, low, high, alignment, 1641 boundary)) 1642 goto retry; 1643 #endif 1644 } 1645 mtx_unlock(&vm_page_queue_free_mtx); 1646 if (m_ret == NULL) 1647 return (NULL); 1648 1649 /* 1650 * Initialize the pages. Only the PG_ZERO flag is inherited. 1651 */ 1652 flags = 0; 1653 if ((req & VM_ALLOC_ZERO) != 0) 1654 flags = PG_ZERO; 1655 if ((req & VM_ALLOC_NODUMP) != 0) 1656 flags |= PG_NODUMP; 1657 if ((req & VM_ALLOC_WIRED) != 0) 1658 atomic_add_int(&cnt.v_wire_count, npages); 1659 oflags = VPO_UNMANAGED; 1660 if (object != NULL) { 1661 if ((req & VM_ALLOC_NOBUSY) == 0) 1662 oflags |= VPO_BUSY; 1663 if (object->memattr != VM_MEMATTR_DEFAULT && 1664 memattr == VM_MEMATTR_DEFAULT) 1665 memattr = object->memattr; 1666 } 1667 for (m = m_ret; m < &m_ret[npages]; m++) { 1668 m->aflags = 0; 1669 m->flags = (m->flags | PG_NODUMP) & flags; 1670 if ((req & VM_ALLOC_WIRED) != 0) 1671 m->wire_count = 1; 1672 /* Unmanaged pages don't use "act_count". */ 1673 m->oflags = oflags; 1674 if (memattr != VM_MEMATTR_DEFAULT) 1675 pmap_page_set_memattr(m, memattr); 1676 if (object != NULL) 1677 vm_page_insert(m, object, pindex); 1678 else 1679 m->pindex = pindex; 1680 pindex++; 1681 } 1682 while (deferred_vdrop_list != NULL) { 1683 vdrop((struct vnode *)deferred_vdrop_list->pageq.tqe_prev); 1684 deferred_vdrop_list = deferred_vdrop_list->pageq.tqe_next; 1685 } 1686 if (vm_paging_needed()) 1687 pagedaemon_wakeup(); 1688 return (m_ret); 1689 } 1690 1691 /* 1692 * Initialize a page that has been freshly dequeued from a freelist. 1693 * The caller has to drop the vnode returned, if it is not NULL. 1694 * 1695 * This function may only be used to initialize unmanaged pages. 1696 * 1697 * To be called with vm_page_queue_free_mtx held. 1698 */ 1699 static struct vnode * 1700 vm_page_alloc_init(vm_page_t m) 1701 { 1702 struct vnode *drop; 1703 vm_object_t m_object; 1704 1705 KASSERT(m->queue == PQ_NONE, 1706 ("vm_page_alloc_init: page %p has unexpected queue %d", 1707 m, m->queue)); 1708 KASSERT(m->wire_count == 0, 1709 ("vm_page_alloc_init: page %p is wired", m)); 1710 KASSERT(m->hold_count == 0, 1711 ("vm_page_alloc_init: page %p is held", m)); 1712 KASSERT(m->busy == 0, 1713 ("vm_page_alloc_init: page %p is busy", m)); 1714 KASSERT(m->dirty == 0, 1715 ("vm_page_alloc_init: page %p is dirty", m)); 1716 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 1717 ("vm_page_alloc_init: page %p has unexpected memattr %d", 1718 m, pmap_page_get_memattr(m))); 1719 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1720 drop = NULL; 1721 if ((m->flags & PG_CACHED) != 0) { 1722 KASSERT((m->flags & PG_ZERO) == 0, 1723 ("vm_page_alloc_init: cached page %p is PG_ZERO", m)); 1724 m->valid = 0; 1725 m_object = m->object; 1726 vm_page_cache_remove(m); 1727 if (m_object->type == OBJT_VNODE && m_object->cache == NULL) 1728 drop = m_object->handle; 1729 } else { 1730 KASSERT(VM_PAGE_IS_FREE(m), 1731 ("vm_page_alloc_init: page %p is not free", m)); 1732 KASSERT(m->valid == 0, 1733 ("vm_page_alloc_init: free page %p is valid", m)); 1734 cnt.v_free_count--; 1735 if ((m->flags & PG_ZERO) != 0) 1736 vm_page_zero_count--; 1737 } 1738 /* Don't clear the PG_ZERO flag; we'll need it later. */ 1739 m->flags &= PG_ZERO; 1740 return (drop); 1741 } 1742 1743 /* 1744 * vm_page_alloc_freelist: 1745 * 1746 * Allocate a physical page from the specified free page list. 1747 * 1748 * The caller must always specify an allocation class. 1749 * 1750 * allocation classes: 1751 * VM_ALLOC_NORMAL normal process request 1752 * VM_ALLOC_SYSTEM system *really* needs a page 1753 * VM_ALLOC_INTERRUPT interrupt time request 1754 * 1755 * optional allocation flags: 1756 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1757 * intends to allocate 1758 * VM_ALLOC_WIRED wire the allocated page 1759 * VM_ALLOC_ZERO prefer a zeroed page 1760 * 1761 * This routine may not sleep. 1762 */ 1763 vm_page_t 1764 vm_page_alloc_freelist(int flind, int req) 1765 { 1766 struct vnode *drop; 1767 vm_page_t m; 1768 u_int flags; 1769 int req_class; 1770 1771 req_class = req & VM_ALLOC_CLASS_MASK; 1772 1773 /* 1774 * The page daemon is allowed to dig deeper into the free page list. 1775 */ 1776 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1777 req_class = VM_ALLOC_SYSTEM; 1778 1779 /* 1780 * Do not allocate reserved pages unless the req has asked for it. 1781 */ 1782 mtx_lock(&vm_page_queue_free_mtx); 1783 if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved || 1784 (req_class == VM_ALLOC_SYSTEM && 1785 cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) || 1786 (req_class == VM_ALLOC_INTERRUPT && 1787 cnt.v_free_count + cnt.v_cache_count > 0)) 1788 m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0); 1789 else { 1790 mtx_unlock(&vm_page_queue_free_mtx); 1791 atomic_add_int(&vm_pageout_deficit, 1792 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 1793 pagedaemon_wakeup(); 1794 return (NULL); 1795 } 1796 if (m == NULL) { 1797 mtx_unlock(&vm_page_queue_free_mtx); 1798 return (NULL); 1799 } 1800 drop = vm_page_alloc_init(m); 1801 mtx_unlock(&vm_page_queue_free_mtx); 1802 1803 /* 1804 * Initialize the page. Only the PG_ZERO flag is inherited. 1805 */ 1806 m->aflags = 0; 1807 flags = 0; 1808 if ((req & VM_ALLOC_ZERO) != 0) 1809 flags = PG_ZERO; 1810 m->flags &= flags; 1811 if ((req & VM_ALLOC_WIRED) != 0) { 1812 /* 1813 * The page lock is not required for wiring a page that does 1814 * not belong to an object. 1815 */ 1816 atomic_add_int(&cnt.v_wire_count, 1); 1817 m->wire_count = 1; 1818 } 1819 /* Unmanaged pages don't use "act_count". */ 1820 m->oflags = VPO_UNMANAGED; 1821 if (drop != NULL) 1822 vdrop(drop); 1823 if (vm_paging_needed()) 1824 pagedaemon_wakeup(); 1825 return (m); 1826 } 1827 1828 /* 1829 * vm_wait: (also see VM_WAIT macro) 1830 * 1831 * Block until free pages are available for allocation 1832 * - Called in various places before memory allocations. 1833 */ 1834 void 1835 vm_wait(void) 1836 { 1837 1838 mtx_lock(&vm_page_queue_free_mtx); 1839 if (curproc == pageproc) { 1840 vm_pageout_pages_needed = 1; 1841 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx, 1842 PDROP | PSWP, "VMWait", 0); 1843 } else { 1844 if (!vm_pages_needed) { 1845 vm_pages_needed = 1; 1846 wakeup(&vm_pages_needed); 1847 } 1848 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM, 1849 "vmwait", 0); 1850 } 1851 } 1852 1853 /* 1854 * vm_waitpfault: (also see VM_WAITPFAULT macro) 1855 * 1856 * Block until free pages are available for allocation 1857 * - Called only in vm_fault so that processes page faulting 1858 * can be easily tracked. 1859 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 1860 * processes will be able to grab memory first. Do not change 1861 * this balance without careful testing first. 1862 */ 1863 void 1864 vm_waitpfault(void) 1865 { 1866 1867 mtx_lock(&vm_page_queue_free_mtx); 1868 if (!vm_pages_needed) { 1869 vm_pages_needed = 1; 1870 wakeup(&vm_pages_needed); 1871 } 1872 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER, 1873 "pfault", 0); 1874 } 1875 1876 /* 1877 * vm_page_requeue: 1878 * 1879 * Move the given page to the tail of its present page queue. 1880 * 1881 * The page queues must be locked. 1882 */ 1883 void 1884 vm_page_requeue(vm_page_t m) 1885 { 1886 struct vpgqueues *vpq; 1887 int queue; 1888 1889 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1890 queue = m->queue; 1891 KASSERT(queue != PQ_NONE, 1892 ("vm_page_requeue: page %p is not queued", m)); 1893 vpq = &vm_page_queues[queue]; 1894 TAILQ_REMOVE(&vpq->pl, m, pageq); 1895 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq); 1896 } 1897 1898 /* 1899 * vm_page_queue_remove: 1900 * 1901 * Remove the given page from the specified queue. 1902 * 1903 * The page and page queues must be locked. 1904 */ 1905 static __inline void 1906 vm_page_queue_remove(int queue, vm_page_t m) 1907 { 1908 struct vpgqueues *pq; 1909 1910 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1911 vm_page_lock_assert(m, MA_OWNED); 1912 pq = &vm_page_queues[queue]; 1913 TAILQ_REMOVE(&pq->pl, m, pageq); 1914 (*pq->cnt)--; 1915 } 1916 1917 /* 1918 * vm_pageq_remove: 1919 * 1920 * Remove a page from its queue. 1921 * 1922 * The given page must be locked. 1923 * This routine may not block. 1924 */ 1925 void 1926 vm_pageq_remove(vm_page_t m) 1927 { 1928 int queue; 1929 1930 vm_page_lock_assert(m, MA_OWNED); 1931 if ((queue = m->queue) != PQ_NONE) { 1932 vm_page_lock_queues(); 1933 m->queue = PQ_NONE; 1934 vm_page_queue_remove(queue, m); 1935 vm_page_unlock_queues(); 1936 } 1937 } 1938 1939 /* 1940 * vm_page_enqueue: 1941 * 1942 * Add the given page to the specified queue. 1943 * 1944 * The page queues must be locked. 1945 */ 1946 static void 1947 vm_page_enqueue(int queue, vm_page_t m) 1948 { 1949 struct vpgqueues *vpq; 1950 1951 vpq = &vm_page_queues[queue]; 1952 m->queue = queue; 1953 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq); 1954 ++*vpq->cnt; 1955 } 1956 1957 /* 1958 * vm_page_activate: 1959 * 1960 * Put the specified page on the active list (if appropriate). 1961 * Ensure that act_count is at least ACT_INIT but do not otherwise 1962 * mess with it. 1963 * 1964 * The page must be locked. 1965 * This routine may not block. 1966 */ 1967 void 1968 vm_page_activate(vm_page_t m) 1969 { 1970 int queue; 1971 1972 vm_page_lock_assert(m, MA_OWNED); 1973 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1974 if ((queue = m->queue) != PQ_ACTIVE) { 1975 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 1976 if (m->act_count < ACT_INIT) 1977 m->act_count = ACT_INIT; 1978 vm_page_lock_queues(); 1979 if (queue != PQ_NONE) 1980 vm_page_queue_remove(queue, m); 1981 vm_page_enqueue(PQ_ACTIVE, m); 1982 vm_page_unlock_queues(); 1983 } else 1984 KASSERT(queue == PQ_NONE, 1985 ("vm_page_activate: wired page %p is queued", m)); 1986 } else { 1987 if (m->act_count < ACT_INIT) 1988 m->act_count = ACT_INIT; 1989 } 1990 } 1991 1992 /* 1993 * vm_page_free_wakeup: 1994 * 1995 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 1996 * routine is called when a page has been added to the cache or free 1997 * queues. 1998 * 1999 * The page queues must be locked. 2000 * This routine may not block. 2001 */ 2002 static inline void 2003 vm_page_free_wakeup(void) 2004 { 2005 2006 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 2007 /* 2008 * if pageout daemon needs pages, then tell it that there are 2009 * some free. 2010 */ 2011 if (vm_pageout_pages_needed && 2012 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 2013 wakeup(&vm_pageout_pages_needed); 2014 vm_pageout_pages_needed = 0; 2015 } 2016 /* 2017 * wakeup processes that are waiting on memory if we hit a 2018 * high water mark. And wakeup scheduler process if we have 2019 * lots of memory. this process will swapin processes. 2020 */ 2021 if (vm_pages_needed && !vm_page_count_min()) { 2022 vm_pages_needed = 0; 2023 wakeup(&cnt.v_free_count); 2024 } 2025 } 2026 2027 /* 2028 * vm_page_free_toq: 2029 * 2030 * Returns the given page to the free list, 2031 * disassociating it with any VM object. 2032 * 2033 * Object and page must be locked prior to entry. 2034 * This routine may not block. 2035 */ 2036 2037 void 2038 vm_page_free_toq(vm_page_t m) 2039 { 2040 2041 if ((m->oflags & VPO_UNMANAGED) == 0) { 2042 vm_page_lock_assert(m, MA_OWNED); 2043 KASSERT(!pmap_page_is_mapped(m), 2044 ("vm_page_free_toq: freeing mapped page %p", m)); 2045 } 2046 PCPU_INC(cnt.v_tfree); 2047 2048 if (VM_PAGE_IS_FREE(m)) 2049 panic("vm_page_free: freeing free page %p", m); 2050 else if (m->busy != 0) 2051 panic("vm_page_free: freeing busy page %p", m); 2052 2053 /* 2054 * unqueue, then remove page. Note that we cannot destroy 2055 * the page here because we do not want to call the pager's 2056 * callback routine until after we've put the page on the 2057 * appropriate free queue. 2058 */ 2059 if ((m->oflags & VPO_UNMANAGED) == 0) 2060 vm_pageq_remove(m); 2061 vm_page_remove(m); 2062 2063 /* 2064 * If fictitious remove object association and 2065 * return, otherwise delay object association removal. 2066 */ 2067 if ((m->flags & PG_FICTITIOUS) != 0) { 2068 return; 2069 } 2070 2071 m->valid = 0; 2072 vm_page_undirty(m); 2073 2074 if (m->wire_count != 0) 2075 panic("vm_page_free: freeing wired page %p", m); 2076 if (m->hold_count != 0) { 2077 m->flags &= ~PG_ZERO; 2078 vm_page_lock_queues(); 2079 vm_page_enqueue(PQ_HOLD, m); 2080 vm_page_unlock_queues(); 2081 } else { 2082 /* 2083 * Restore the default memory attribute to the page. 2084 */ 2085 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 2086 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 2087 2088 /* 2089 * Insert the page into the physical memory allocator's 2090 * cache/free page queues. 2091 */ 2092 mtx_lock(&vm_page_queue_free_mtx); 2093 m->flags |= PG_FREE; 2094 cnt.v_free_count++; 2095 #if VM_NRESERVLEVEL > 0 2096 if (!vm_reserv_free_page(m)) 2097 #else 2098 if (TRUE) 2099 #endif 2100 vm_phys_free_pages(m, 0); 2101 if ((m->flags & PG_ZERO) != 0) 2102 ++vm_page_zero_count; 2103 else 2104 vm_page_zero_idle_wakeup(); 2105 vm_page_free_wakeup(); 2106 mtx_unlock(&vm_page_queue_free_mtx); 2107 } 2108 } 2109 2110 /* 2111 * vm_page_wire: 2112 * 2113 * Mark this page as wired down by yet 2114 * another map, removing it from paging queues 2115 * as necessary. 2116 * 2117 * If the page is fictitious, then its wire count must remain one. 2118 * 2119 * The page must be locked. 2120 * This routine may not block. 2121 */ 2122 void 2123 vm_page_wire(vm_page_t m) 2124 { 2125 2126 /* 2127 * Only bump the wire statistics if the page is not already wired, 2128 * and only unqueue the page if it is on some queue (if it is unmanaged 2129 * it is already off the queues). 2130 */ 2131 vm_page_lock_assert(m, MA_OWNED); 2132 if ((m->flags & PG_FICTITIOUS) != 0) { 2133 KASSERT(m->wire_count == 1, 2134 ("vm_page_wire: fictitious page %p's wire count isn't one", 2135 m)); 2136 return; 2137 } 2138 if (m->wire_count == 0) { 2139 if ((m->oflags & VPO_UNMANAGED) == 0) 2140 vm_pageq_remove(m); 2141 atomic_add_int(&cnt.v_wire_count, 1); 2142 } 2143 m->wire_count++; 2144 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 2145 } 2146 2147 /* 2148 * vm_page_unwire: 2149 * 2150 * Release one wiring of the specified page, potentially enabling it to be 2151 * paged again. If paging is enabled, then the value of the parameter 2152 * "activate" determines to which queue the page is added. If "activate" is 2153 * non-zero, then the page is added to the active queue. Otherwise, it is 2154 * added to the inactive queue. 2155 * 2156 * However, unless the page belongs to an object, it is not enqueued because 2157 * it cannot be paged out. 2158 * 2159 * If a page is fictitious, then its wire count must alway be one. 2160 * 2161 * A managed page must be locked. 2162 */ 2163 void 2164 vm_page_unwire(vm_page_t m, int activate) 2165 { 2166 2167 if ((m->oflags & VPO_UNMANAGED) == 0) 2168 vm_page_lock_assert(m, MA_OWNED); 2169 if ((m->flags & PG_FICTITIOUS) != 0) { 2170 KASSERT(m->wire_count == 1, 2171 ("vm_page_unwire: fictitious page %p's wire count isn't one", m)); 2172 return; 2173 } 2174 if (m->wire_count > 0) { 2175 m->wire_count--; 2176 if (m->wire_count == 0) { 2177 atomic_subtract_int(&cnt.v_wire_count, 1); 2178 if ((m->oflags & VPO_UNMANAGED) != 0 || 2179 m->object == NULL) 2180 return; 2181 if (!activate) 2182 m->flags &= ~PG_WINATCFLS; 2183 vm_page_lock_queues(); 2184 vm_page_enqueue(activate ? PQ_ACTIVE : PQ_INACTIVE, m); 2185 vm_page_unlock_queues(); 2186 } 2187 } else 2188 panic("vm_page_unwire: page %p's wire count is zero", m); 2189 } 2190 2191 /* 2192 * Move the specified page to the inactive queue. 2193 * 2194 * Many pages placed on the inactive queue should actually go 2195 * into the cache, but it is difficult to figure out which. What 2196 * we do instead, if the inactive target is well met, is to put 2197 * clean pages at the head of the inactive queue instead of the tail. 2198 * This will cause them to be moved to the cache more quickly and 2199 * if not actively re-referenced, reclaimed more quickly. If we just 2200 * stick these pages at the end of the inactive queue, heavy filesystem 2201 * meta-data accesses can cause an unnecessary paging load on memory bound 2202 * processes. This optimization causes one-time-use metadata to be 2203 * reused more quickly. 2204 * 2205 * Normally athead is 0 resulting in LRU operation. athead is set 2206 * to 1 if we want this page to be 'as if it were placed in the cache', 2207 * except without unmapping it from the process address space. 2208 * 2209 * This routine may not block. 2210 */ 2211 static inline void 2212 _vm_page_deactivate(vm_page_t m, int athead) 2213 { 2214 int queue; 2215 2216 vm_page_lock_assert(m, MA_OWNED); 2217 2218 /* 2219 * Ignore if already inactive. 2220 */ 2221 if ((queue = m->queue) == PQ_INACTIVE) 2222 return; 2223 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 2224 m->flags &= ~PG_WINATCFLS; 2225 vm_page_lock_queues(); 2226 if (queue != PQ_NONE) 2227 vm_page_queue_remove(queue, m); 2228 if (athead) 2229 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, 2230 pageq); 2231 else 2232 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, 2233 pageq); 2234 m->queue = PQ_INACTIVE; 2235 cnt.v_inactive_count++; 2236 vm_page_unlock_queues(); 2237 } 2238 } 2239 2240 /* 2241 * Move the specified page to the inactive queue. 2242 * 2243 * The page must be locked. 2244 */ 2245 void 2246 vm_page_deactivate(vm_page_t m) 2247 { 2248 2249 _vm_page_deactivate(m, 0); 2250 } 2251 2252 /* 2253 * vm_page_try_to_cache: 2254 * 2255 * Returns 0 on failure, 1 on success 2256 */ 2257 int 2258 vm_page_try_to_cache(vm_page_t m) 2259 { 2260 2261 vm_page_lock_assert(m, MA_OWNED); 2262 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2263 if (m->dirty || m->hold_count || m->busy || m->wire_count || 2264 (m->oflags & (VPO_BUSY | VPO_UNMANAGED)) != 0) 2265 return (0); 2266 pmap_remove_all(m); 2267 if (m->dirty) 2268 return (0); 2269 vm_page_cache(m); 2270 return (1); 2271 } 2272 2273 /* 2274 * vm_page_try_to_free() 2275 * 2276 * Attempt to free the page. If we cannot free it, we do nothing. 2277 * 1 is returned on success, 0 on failure. 2278 */ 2279 int 2280 vm_page_try_to_free(vm_page_t m) 2281 { 2282 2283 vm_page_lock_assert(m, MA_OWNED); 2284 if (m->object != NULL) 2285 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2286 if (m->dirty || m->hold_count || m->busy || m->wire_count || 2287 (m->oflags & (VPO_BUSY | VPO_UNMANAGED)) != 0) 2288 return (0); 2289 pmap_remove_all(m); 2290 if (m->dirty) 2291 return (0); 2292 vm_page_free(m); 2293 return (1); 2294 } 2295 2296 /* 2297 * vm_page_cache 2298 * 2299 * Put the specified page onto the page cache queue (if appropriate). 2300 * 2301 * This routine may not block. 2302 */ 2303 void 2304 vm_page_cache(vm_page_t m) 2305 { 2306 vm_object_t object; 2307 vm_page_t next, prev, root; 2308 2309 vm_page_lock_assert(m, MA_OWNED); 2310 object = m->object; 2311 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2312 if ((m->oflags & (VPO_UNMANAGED | VPO_BUSY)) || m->busy || 2313 m->hold_count || m->wire_count) 2314 panic("vm_page_cache: attempting to cache busy page"); 2315 pmap_remove_all(m); 2316 if (m->dirty != 0) 2317 panic("vm_page_cache: page %p is dirty", m); 2318 if (m->valid == 0 || object->type == OBJT_DEFAULT || 2319 (object->type == OBJT_SWAP && 2320 !vm_pager_has_page(object, m->pindex, NULL, NULL))) { 2321 /* 2322 * Hypothesis: A cache-elgible page belonging to a 2323 * default object or swap object but without a backing 2324 * store must be zero filled. 2325 */ 2326 vm_page_free(m); 2327 return; 2328 } 2329 KASSERT((m->flags & PG_CACHED) == 0, 2330 ("vm_page_cache: page %p is already cached", m)); 2331 PCPU_INC(cnt.v_tcached); 2332 2333 /* 2334 * Remove the page from the paging queues. 2335 */ 2336 vm_pageq_remove(m); 2337 2338 /* 2339 * Remove the page from the object's collection of resident 2340 * pages. 2341 */ 2342 if ((next = TAILQ_NEXT(m, listq)) != NULL && next->left == m) { 2343 /* 2344 * Since the page's successor in the list is also its parent 2345 * in the tree, its right subtree must be empty. 2346 */ 2347 next->left = m->left; 2348 KASSERT(m->right == NULL, 2349 ("vm_page_cache: page %p has right child", m)); 2350 } else if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL && 2351 prev->right == m) { 2352 /* 2353 * Since the page's predecessor in the list is also its parent 2354 * in the tree, its left subtree must be empty. 2355 */ 2356 KASSERT(m->left == NULL, 2357 ("vm_page_cache: page %p has left child", m)); 2358 prev->right = m->right; 2359 } else { 2360 if (m != object->root) 2361 vm_page_splay(m->pindex, object->root); 2362 if (m->left == NULL) 2363 root = m->right; 2364 else if (m->right == NULL) 2365 root = m->left; 2366 else { 2367 /* 2368 * Move the page's successor to the root, because 2369 * pages are usually removed in ascending order. 2370 */ 2371 if (m->right != next) 2372 vm_page_splay(m->pindex, m->right); 2373 next->left = m->left; 2374 root = next; 2375 } 2376 object->root = root; 2377 } 2378 TAILQ_REMOVE(&object->memq, m, listq); 2379 object->resident_page_count--; 2380 2381 /* 2382 * Restore the default memory attribute to the page. 2383 */ 2384 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 2385 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 2386 2387 /* 2388 * Insert the page into the object's collection of cached pages 2389 * and the physical memory allocator's cache/free page queues. 2390 */ 2391 m->flags &= ~PG_ZERO; 2392 mtx_lock(&vm_page_queue_free_mtx); 2393 m->flags |= PG_CACHED; 2394 cnt.v_cache_count++; 2395 root = object->cache; 2396 if (root == NULL) { 2397 m->left = NULL; 2398 m->right = NULL; 2399 } else { 2400 root = vm_page_splay(m->pindex, root); 2401 if (m->pindex < root->pindex) { 2402 m->left = root->left; 2403 m->right = root; 2404 root->left = NULL; 2405 } else if (__predict_false(m->pindex == root->pindex)) 2406 panic("vm_page_cache: offset already cached"); 2407 else { 2408 m->right = root->right; 2409 m->left = root; 2410 root->right = NULL; 2411 } 2412 } 2413 object->cache = m; 2414 #if VM_NRESERVLEVEL > 0 2415 if (!vm_reserv_free_page(m)) { 2416 #else 2417 if (TRUE) { 2418 #endif 2419 vm_phys_set_pool(VM_FREEPOOL_CACHE, m, 0); 2420 vm_phys_free_pages(m, 0); 2421 } 2422 vm_page_free_wakeup(); 2423 mtx_unlock(&vm_page_queue_free_mtx); 2424 2425 /* 2426 * Increment the vnode's hold count if this is the object's only 2427 * cached page. Decrement the vnode's hold count if this was 2428 * the object's only resident page. 2429 */ 2430 if (object->type == OBJT_VNODE) { 2431 if (root == NULL && object->resident_page_count != 0) 2432 vhold(object->handle); 2433 else if (root != NULL && object->resident_page_count == 0) 2434 vdrop(object->handle); 2435 } 2436 } 2437 2438 /* 2439 * vm_page_dontneed 2440 * 2441 * Cache, deactivate, or do nothing as appropriate. This routine 2442 * is typically used by madvise() MADV_DONTNEED. 2443 * 2444 * Generally speaking we want to move the page into the cache so 2445 * it gets reused quickly. However, this can result in a silly syndrome 2446 * due to the page recycling too quickly. Small objects will not be 2447 * fully cached. On the otherhand, if we move the page to the inactive 2448 * queue we wind up with a problem whereby very large objects 2449 * unnecessarily blow away our inactive and cache queues. 2450 * 2451 * The solution is to move the pages based on a fixed weighting. We 2452 * either leave them alone, deactivate them, or move them to the cache, 2453 * where moving them to the cache has the highest weighting. 2454 * By forcing some pages into other queues we eventually force the 2455 * system to balance the queues, potentially recovering other unrelated 2456 * space from active. The idea is to not force this to happen too 2457 * often. 2458 */ 2459 void 2460 vm_page_dontneed(vm_page_t m) 2461 { 2462 int dnw; 2463 int head; 2464 2465 vm_page_lock_assert(m, MA_OWNED); 2466 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2467 dnw = PCPU_GET(dnweight); 2468 PCPU_INC(dnweight); 2469 2470 /* 2471 * Occasionally leave the page alone. 2472 */ 2473 if ((dnw & 0x01F0) == 0 || m->queue == PQ_INACTIVE) { 2474 if (m->act_count >= ACT_INIT) 2475 --m->act_count; 2476 return; 2477 } 2478 2479 /* 2480 * Clear any references to the page. Otherwise, the page daemon will 2481 * immediately reactivate the page. 2482 * 2483 * Perform the pmap_clear_reference() first. Otherwise, a concurrent 2484 * pmap operation, such as pmap_remove(), could clear a reference in 2485 * the pmap and set PGA_REFERENCED on the page before the 2486 * pmap_clear_reference() had completed. Consequently, the page would 2487 * appear referenced based upon an old reference that occurred before 2488 * this function ran. 2489 */ 2490 pmap_clear_reference(m); 2491 vm_page_aflag_clear(m, PGA_REFERENCED); 2492 2493 if (m->dirty == 0 && pmap_is_modified(m)) 2494 vm_page_dirty(m); 2495 2496 if (m->dirty || (dnw & 0x0070) == 0) { 2497 /* 2498 * Deactivate the page 3 times out of 32. 2499 */ 2500 head = 0; 2501 } else { 2502 /* 2503 * Cache the page 28 times out of every 32. Note that 2504 * the page is deactivated instead of cached, but placed 2505 * at the head of the queue instead of the tail. 2506 */ 2507 head = 1; 2508 } 2509 _vm_page_deactivate(m, head); 2510 } 2511 2512 /* 2513 * Grab a page, waiting until we are waken up due to the page 2514 * changing state. We keep on waiting, if the page continues 2515 * to be in the object. If the page doesn't exist, first allocate it 2516 * and then conditionally zero it. 2517 * 2518 * The caller must always specify the VM_ALLOC_RETRY flag. This is intended 2519 * to facilitate its eventual removal. 2520 * 2521 * This routine may block. 2522 */ 2523 vm_page_t 2524 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 2525 { 2526 vm_page_t m; 2527 2528 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2529 KASSERT((allocflags & VM_ALLOC_RETRY) != 0, 2530 ("vm_page_grab: VM_ALLOC_RETRY is required")); 2531 retrylookup: 2532 if ((m = vm_page_lookup(object, pindex)) != NULL) { 2533 if ((m->oflags & VPO_BUSY) != 0 || 2534 ((allocflags & VM_ALLOC_IGN_SBUSY) == 0 && m->busy != 0)) { 2535 /* 2536 * Reference the page before unlocking and 2537 * sleeping so that the page daemon is less 2538 * likely to reclaim it. 2539 */ 2540 vm_page_aflag_set(m, PGA_REFERENCED); 2541 vm_page_sleep(m, "pgrbwt"); 2542 goto retrylookup; 2543 } else { 2544 if ((allocflags & VM_ALLOC_WIRED) != 0) { 2545 vm_page_lock(m); 2546 vm_page_wire(m); 2547 vm_page_unlock(m); 2548 } 2549 if ((allocflags & VM_ALLOC_NOBUSY) == 0) 2550 vm_page_busy(m); 2551 return (m); 2552 } 2553 } 2554 m = vm_page_alloc(object, pindex, allocflags & ~(VM_ALLOC_RETRY | 2555 VM_ALLOC_IGN_SBUSY)); 2556 if (m == NULL) { 2557 VM_OBJECT_UNLOCK(object); 2558 VM_WAIT; 2559 VM_OBJECT_LOCK(object); 2560 goto retrylookup; 2561 } else if (m->valid != 0) 2562 return (m); 2563 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 2564 pmap_zero_page(m); 2565 return (m); 2566 } 2567 2568 /* 2569 * Mapping function for valid bits or for dirty bits in 2570 * a page. May not block. 2571 * 2572 * Inputs are required to range within a page. 2573 */ 2574 vm_page_bits_t 2575 vm_page_bits(int base, int size) 2576 { 2577 int first_bit; 2578 int last_bit; 2579 2580 KASSERT( 2581 base + size <= PAGE_SIZE, 2582 ("vm_page_bits: illegal base/size %d/%d", base, size) 2583 ); 2584 2585 if (size == 0) /* handle degenerate case */ 2586 return (0); 2587 2588 first_bit = base >> DEV_BSHIFT; 2589 last_bit = (base + size - 1) >> DEV_BSHIFT; 2590 2591 return (((vm_page_bits_t)2 << last_bit) - 2592 ((vm_page_bits_t)1 << first_bit)); 2593 } 2594 2595 /* 2596 * vm_page_set_valid_range: 2597 * 2598 * Sets portions of a page valid. The arguments are expected 2599 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2600 * of any partial chunks touched by the range. The invalid portion of 2601 * such chunks will be zeroed. 2602 * 2603 * (base + size) must be less then or equal to PAGE_SIZE. 2604 */ 2605 void 2606 vm_page_set_valid_range(vm_page_t m, int base, int size) 2607 { 2608 int endoff, frag; 2609 2610 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2611 if (size == 0) /* handle degenerate case */ 2612 return; 2613 2614 /* 2615 * If the base is not DEV_BSIZE aligned and the valid 2616 * bit is clear, we have to zero out a portion of the 2617 * first block. 2618 */ 2619 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 2620 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 2621 pmap_zero_page_area(m, frag, base - frag); 2622 2623 /* 2624 * If the ending offset is not DEV_BSIZE aligned and the 2625 * valid bit is clear, we have to zero out a portion of 2626 * the last block. 2627 */ 2628 endoff = base + size; 2629 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 2630 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 2631 pmap_zero_page_area(m, endoff, 2632 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 2633 2634 /* 2635 * Assert that no previously invalid block that is now being validated 2636 * is already dirty. 2637 */ 2638 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 2639 ("vm_page_set_valid_range: page %p is dirty", m)); 2640 2641 /* 2642 * Set valid bits inclusive of any overlap. 2643 */ 2644 m->valid |= vm_page_bits(base, size); 2645 } 2646 2647 /* 2648 * Clear the given bits from the specified page's dirty field. 2649 */ 2650 static __inline void 2651 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) 2652 { 2653 uintptr_t addr; 2654 #if PAGE_SIZE < 16384 2655 int shift; 2656 #endif 2657 2658 /* 2659 * If the object is locked and the page is neither VPO_BUSY nor 2660 * write mapped, then the page's dirty field cannot possibly be 2661 * set by a concurrent pmap operation. 2662 */ 2663 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2664 if ((m->oflags & VPO_BUSY) == 0 && !pmap_page_is_write_mapped(m)) 2665 m->dirty &= ~pagebits; 2666 else { 2667 /* 2668 * The pmap layer can call vm_page_dirty() without 2669 * holding a distinguished lock. The combination of 2670 * the object's lock and an atomic operation suffice 2671 * to guarantee consistency of the page dirty field. 2672 * 2673 * For PAGE_SIZE == 32768 case, compiler already 2674 * properly aligns the dirty field, so no forcible 2675 * alignment is needed. Only require existence of 2676 * atomic_clear_64 when page size is 32768. 2677 */ 2678 addr = (uintptr_t)&m->dirty; 2679 #if PAGE_SIZE == 32768 2680 atomic_clear_64((uint64_t *)addr, pagebits); 2681 #elif PAGE_SIZE == 16384 2682 atomic_clear_32((uint32_t *)addr, pagebits); 2683 #else /* PAGE_SIZE <= 8192 */ 2684 /* 2685 * Use a trick to perform a 32-bit atomic on the 2686 * containing aligned word, to not depend on the existence 2687 * of atomic_clear_{8, 16}. 2688 */ 2689 shift = addr & (sizeof(uint32_t) - 1); 2690 #if BYTE_ORDER == BIG_ENDIAN 2691 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY; 2692 #else 2693 shift *= NBBY; 2694 #endif 2695 addr &= ~(sizeof(uint32_t) - 1); 2696 atomic_clear_32((uint32_t *)addr, pagebits << shift); 2697 #endif /* PAGE_SIZE */ 2698 } 2699 } 2700 2701 /* 2702 * vm_page_set_validclean: 2703 * 2704 * Sets portions of a page valid and clean. The arguments are expected 2705 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2706 * of any partial chunks touched by the range. The invalid portion of 2707 * such chunks will be zero'd. 2708 * 2709 * This routine may not block. 2710 * 2711 * (base + size) must be less then or equal to PAGE_SIZE. 2712 */ 2713 void 2714 vm_page_set_validclean(vm_page_t m, int base, int size) 2715 { 2716 vm_page_bits_t oldvalid, pagebits; 2717 int endoff, frag; 2718 2719 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2720 if (size == 0) /* handle degenerate case */ 2721 return; 2722 2723 /* 2724 * If the base is not DEV_BSIZE aligned and the valid 2725 * bit is clear, we have to zero out a portion of the 2726 * first block. 2727 */ 2728 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 2729 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) 2730 pmap_zero_page_area(m, frag, base - frag); 2731 2732 /* 2733 * If the ending offset is not DEV_BSIZE aligned and the 2734 * valid bit is clear, we have to zero out a portion of 2735 * the last block. 2736 */ 2737 endoff = base + size; 2738 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 2739 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) 2740 pmap_zero_page_area(m, endoff, 2741 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 2742 2743 /* 2744 * Set valid, clear dirty bits. If validating the entire 2745 * page we can safely clear the pmap modify bit. We also 2746 * use this opportunity to clear the VPO_NOSYNC flag. If a process 2747 * takes a write fault on a MAP_NOSYNC memory area the flag will 2748 * be set again. 2749 * 2750 * We set valid bits inclusive of any overlap, but we can only 2751 * clear dirty bits for DEV_BSIZE chunks that are fully within 2752 * the range. 2753 */ 2754 oldvalid = m->valid; 2755 pagebits = vm_page_bits(base, size); 2756 m->valid |= pagebits; 2757 #if 0 /* NOT YET */ 2758 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 2759 frag = DEV_BSIZE - frag; 2760 base += frag; 2761 size -= frag; 2762 if (size < 0) 2763 size = 0; 2764 } 2765 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 2766 #endif 2767 if (base == 0 && size == PAGE_SIZE) { 2768 /* 2769 * The page can only be modified within the pmap if it is 2770 * mapped, and it can only be mapped if it was previously 2771 * fully valid. 2772 */ 2773 if (oldvalid == VM_PAGE_BITS_ALL) 2774 /* 2775 * Perform the pmap_clear_modify() first. Otherwise, 2776 * a concurrent pmap operation, such as 2777 * pmap_protect(), could clear a modification in the 2778 * pmap and set the dirty field on the page before 2779 * pmap_clear_modify() had begun and after the dirty 2780 * field was cleared here. 2781 */ 2782 pmap_clear_modify(m); 2783 m->dirty = 0; 2784 m->oflags &= ~VPO_NOSYNC; 2785 } else if (oldvalid != VM_PAGE_BITS_ALL) 2786 m->dirty &= ~pagebits; 2787 else 2788 vm_page_clear_dirty_mask(m, pagebits); 2789 } 2790 2791 void 2792 vm_page_clear_dirty(vm_page_t m, int base, int size) 2793 { 2794 2795 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 2796 } 2797 2798 /* 2799 * vm_page_set_invalid: 2800 * 2801 * Invalidates DEV_BSIZE'd chunks within a page. Both the 2802 * valid and dirty bits for the effected areas are cleared. 2803 * 2804 * May not block. 2805 */ 2806 void 2807 vm_page_set_invalid(vm_page_t m, int base, int size) 2808 { 2809 vm_page_bits_t bits; 2810 2811 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2812 KASSERT((m->oflags & VPO_BUSY) == 0, 2813 ("vm_page_set_invalid: page %p is busy", m)); 2814 bits = vm_page_bits(base, size); 2815 if (m->valid == VM_PAGE_BITS_ALL && bits != 0) 2816 pmap_remove_all(m); 2817 KASSERT(!pmap_page_is_mapped(m), 2818 ("vm_page_set_invalid: page %p is mapped", m)); 2819 m->valid &= ~bits; 2820 m->dirty &= ~bits; 2821 } 2822 2823 /* 2824 * vm_page_zero_invalid() 2825 * 2826 * The kernel assumes that the invalid portions of a page contain 2827 * garbage, but such pages can be mapped into memory by user code. 2828 * When this occurs, we must zero out the non-valid portions of the 2829 * page so user code sees what it expects. 2830 * 2831 * Pages are most often semi-valid when the end of a file is mapped 2832 * into memory and the file's size is not page aligned. 2833 */ 2834 void 2835 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 2836 { 2837 int b; 2838 int i; 2839 2840 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2841 /* 2842 * Scan the valid bits looking for invalid sections that 2843 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 2844 * valid bit may be set ) have already been zerod by 2845 * vm_page_set_validclean(). 2846 */ 2847 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 2848 if (i == (PAGE_SIZE / DEV_BSIZE) || 2849 (m->valid & ((vm_page_bits_t)1 << i))) { 2850 if (i > b) { 2851 pmap_zero_page_area(m, 2852 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 2853 } 2854 b = i + 1; 2855 } 2856 } 2857 2858 /* 2859 * setvalid is TRUE when we can safely set the zero'd areas 2860 * as being valid. We can do this if there are no cache consistancy 2861 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 2862 */ 2863 if (setvalid) 2864 m->valid = VM_PAGE_BITS_ALL; 2865 } 2866 2867 /* 2868 * vm_page_is_valid: 2869 * 2870 * Is (partial) page valid? Note that the case where size == 0 2871 * will return FALSE in the degenerate case where the page is 2872 * entirely invalid, and TRUE otherwise. 2873 * 2874 * May not block. 2875 */ 2876 int 2877 vm_page_is_valid(vm_page_t m, int base, int size) 2878 { 2879 vm_page_bits_t bits; 2880 2881 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2882 bits = vm_page_bits(base, size); 2883 if (m->valid && ((m->valid & bits) == bits)) 2884 return 1; 2885 else 2886 return 0; 2887 } 2888 2889 /* 2890 * update dirty bits from pmap/mmu. May not block. 2891 */ 2892 void 2893 vm_page_test_dirty(vm_page_t m) 2894 { 2895 2896 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2897 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 2898 vm_page_dirty(m); 2899 } 2900 2901 void 2902 vm_page_lock_KBI(vm_page_t m, const char *file, int line) 2903 { 2904 2905 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); 2906 } 2907 2908 void 2909 vm_page_unlock_KBI(vm_page_t m, const char *file, int line) 2910 { 2911 2912 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); 2913 } 2914 2915 int 2916 vm_page_trylock_KBI(vm_page_t m, const char *file, int line) 2917 { 2918 2919 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); 2920 } 2921 2922 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) 2923 void 2924 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) 2925 { 2926 2927 mtx_assert_(vm_page_lockptr(m), a, file, line); 2928 } 2929 #endif 2930 2931 int so_zerocp_fullpage = 0; 2932 2933 /* 2934 * Replace the given page with a copy. The copied page assumes 2935 * the portion of the given page's "wire_count" that is not the 2936 * responsibility of this copy-on-write mechanism. 2937 * 2938 * The object containing the given page must have a non-zero 2939 * paging-in-progress count and be locked. 2940 */ 2941 void 2942 vm_page_cowfault(vm_page_t m) 2943 { 2944 vm_page_t mnew; 2945 vm_object_t object; 2946 vm_pindex_t pindex; 2947 2948 mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED); 2949 vm_page_lock_assert(m, MA_OWNED); 2950 object = m->object; 2951 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2952 KASSERT(object->paging_in_progress != 0, 2953 ("vm_page_cowfault: object %p's paging-in-progress count is zero.", 2954 object)); 2955 pindex = m->pindex; 2956 2957 retry_alloc: 2958 pmap_remove_all(m); 2959 vm_page_remove(m); 2960 mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY); 2961 if (mnew == NULL) { 2962 vm_page_insert(m, object, pindex); 2963 vm_page_unlock(m); 2964 VM_OBJECT_UNLOCK(object); 2965 VM_WAIT; 2966 VM_OBJECT_LOCK(object); 2967 if (m == vm_page_lookup(object, pindex)) { 2968 vm_page_lock(m); 2969 goto retry_alloc; 2970 } else { 2971 /* 2972 * Page disappeared during the wait. 2973 */ 2974 return; 2975 } 2976 } 2977 2978 if (m->cow == 0) { 2979 /* 2980 * check to see if we raced with an xmit complete when 2981 * waiting to allocate a page. If so, put things back 2982 * the way they were 2983 */ 2984 vm_page_unlock(m); 2985 vm_page_lock(mnew); 2986 vm_page_free(mnew); 2987 vm_page_unlock(mnew); 2988 vm_page_insert(m, object, pindex); 2989 } else { /* clear COW & copy page */ 2990 if (!so_zerocp_fullpage) 2991 pmap_copy_page(m, mnew); 2992 mnew->valid = VM_PAGE_BITS_ALL; 2993 vm_page_dirty(mnew); 2994 mnew->wire_count = m->wire_count - m->cow; 2995 m->wire_count = m->cow; 2996 vm_page_unlock(m); 2997 } 2998 } 2999 3000 void 3001 vm_page_cowclear(vm_page_t m) 3002 { 3003 3004 vm_page_lock_assert(m, MA_OWNED); 3005 if (m->cow) { 3006 m->cow--; 3007 /* 3008 * let vm_fault add back write permission lazily 3009 */ 3010 } 3011 /* 3012 * sf_buf_free() will free the page, so we needn't do it here 3013 */ 3014 } 3015 3016 int 3017 vm_page_cowsetup(vm_page_t m) 3018 { 3019 3020 vm_page_lock_assert(m, MA_OWNED); 3021 if ((m->flags & PG_FICTITIOUS) != 0 || 3022 (m->oflags & VPO_UNMANAGED) != 0 || 3023 m->cow == USHRT_MAX - 1 || !VM_OBJECT_TRYLOCK(m->object)) 3024 return (EBUSY); 3025 m->cow++; 3026 pmap_remove_write(m); 3027 VM_OBJECT_UNLOCK(m->object); 3028 return (0); 3029 } 3030 3031 #ifdef INVARIANTS 3032 void 3033 vm_page_object_lock_assert(vm_page_t m) 3034 { 3035 3036 /* 3037 * Certain of the page's fields may only be modified by the 3038 * holder of the containing object's lock or the setter of the 3039 * page's VPO_BUSY flag. Unfortunately, the setter of the 3040 * VPO_BUSY flag is not recorded, and thus cannot be checked 3041 * here. 3042 */ 3043 if (m->object != NULL && (m->oflags & VPO_BUSY) == 0) 3044 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 3045 } 3046 #endif 3047 3048 #include "opt_ddb.h" 3049 #ifdef DDB 3050 #include <sys/kernel.h> 3051 3052 #include <ddb/ddb.h> 3053 3054 DB_SHOW_COMMAND(page, vm_page_print_page_info) 3055 { 3056 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 3057 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 3058 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 3059 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 3060 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 3061 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 3062 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 3063 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 3064 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 3065 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 3066 } 3067 3068 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 3069 { 3070 3071 db_printf("PQ_FREE:"); 3072 db_printf(" %d", cnt.v_free_count); 3073 db_printf("\n"); 3074 3075 db_printf("PQ_CACHE:"); 3076 db_printf(" %d", cnt.v_cache_count); 3077 db_printf("\n"); 3078 3079 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 3080 *vm_page_queues[PQ_ACTIVE].cnt, 3081 *vm_page_queues[PQ_INACTIVE].cnt); 3082 } 3083 #endif /* DDB */ 3084