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