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