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