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_msgbuf.h" 104 #include "opt_vm.h" 105 106 #include <sys/param.h> 107 #include <sys/systm.h> 108 #include <sys/lock.h> 109 #include <sys/kernel.h> 110 #include <sys/limits.h> 111 #include <sys/malloc.h> 112 #include <sys/msgbuf.h> 113 #include <sys/mutex.h> 114 #include <sys/proc.h> 115 #include <sys/sysctl.h> 116 #include <sys/vmmeter.h> 117 #include <sys/vnode.h> 118 119 #include <vm/vm.h> 120 #include <vm/pmap.h> 121 #include <vm/vm_param.h> 122 #include <vm/vm_kern.h> 123 #include <vm/vm_object.h> 124 #include <vm/vm_page.h> 125 #include <vm/vm_pageout.h> 126 #include <vm/vm_pager.h> 127 #include <vm/vm_phys.h> 128 #include <vm/vm_reserv.h> 129 #include <vm/vm_extern.h> 130 #include <vm/uma.h> 131 #include <vm/uma_int.h> 132 133 #include <machine/md_var.h> 134 135 #if defined(__amd64__) || defined (__i386__) 136 extern struct sysctl_oid_list sysctl__vm_pmap_children; 137 #else 138 SYSCTL_NODE(_vm, OID_AUTO, pmap, CTLFLAG_RD, 0, "VM/pmap parameters"); 139 #endif 140 141 static uint64_t pmap_tryrelock_calls; 142 SYSCTL_QUAD(_vm_pmap, OID_AUTO, tryrelock_calls, CTLFLAG_RD, 143 &pmap_tryrelock_calls, 0, "Number of tryrelock calls"); 144 145 static int pmap_tryrelock_restart; 146 SYSCTL_INT(_vm_pmap, OID_AUTO, tryrelock_restart, CTLFLAG_RD, 147 &pmap_tryrelock_restart, 0, "Number of tryrelock restarts"); 148 149 static int pmap_tryrelock_race; 150 SYSCTL_INT(_vm_pmap, OID_AUTO, tryrelock_race, CTLFLAG_RD, 151 &pmap_tryrelock_race, 0, "Number of tryrelock pmap race cases"); 152 153 /* 154 * Associated with page of user-allocatable memory is a 155 * page structure. 156 */ 157 158 struct vpgqueues vm_page_queues[PQ_COUNT]; 159 struct vpglocks vm_page_queue_lock; 160 struct vpglocks vm_page_queue_free_lock; 161 162 struct vpglocks pa_lock[PA_LOCK_COUNT] __aligned(CACHE_LINE_SIZE); 163 164 vm_page_t vm_page_array = 0; 165 int vm_page_array_size = 0; 166 long first_page = 0; 167 int vm_page_zero_count = 0; 168 169 static int boot_pages = UMA_BOOT_PAGES; 170 TUNABLE_INT("vm.boot_pages", &boot_pages); 171 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RD, &boot_pages, 0, 172 "number of pages allocated for bootstrapping the VM system"); 173 174 static void vm_page_clear_dirty_mask(vm_page_t m, int pagebits); 175 static void vm_page_queue_remove(int queue, vm_page_t m); 176 static void vm_page_enqueue(int queue, vm_page_t m); 177 178 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */ 179 #if PAGE_SIZE == 32768 180 #ifdef CTASSERT 181 CTASSERT(sizeof(u_long) >= 8); 182 #endif 183 #endif 184 185 /* 186 * Try to acquire a physical address lock while a pmap is locked. If we 187 * fail to trylock we unlock and lock the pmap directly and cache the 188 * locked pa in *locked. The caller should then restart their loop in case 189 * the virtual to physical mapping has changed. 190 */ 191 int 192 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked) 193 { 194 vm_paddr_t lockpa; 195 uint32_t gen_count; 196 197 gen_count = pmap->pm_gen_count; 198 atomic_add_long((volatile long *)&pmap_tryrelock_calls, 1); 199 lockpa = *locked; 200 *locked = pa; 201 if (lockpa) { 202 PA_LOCK_ASSERT(lockpa, MA_OWNED); 203 if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa)) 204 return (0); 205 PA_UNLOCK(lockpa); 206 } 207 if (PA_TRYLOCK(pa)) 208 return (0); 209 PMAP_UNLOCK(pmap); 210 atomic_add_int((volatile int *)&pmap_tryrelock_restart, 1); 211 PA_LOCK(pa); 212 PMAP_LOCK(pmap); 213 214 if (pmap->pm_gen_count != gen_count + 1) { 215 pmap->pm_retries++; 216 atomic_add_int((volatile int *)&pmap_tryrelock_race, 1); 217 return (EAGAIN); 218 } 219 return (0); 220 } 221 222 /* 223 * vm_set_page_size: 224 * 225 * Sets the page size, perhaps based upon the memory 226 * size. Must be called before any use of page-size 227 * dependent functions. 228 */ 229 void 230 vm_set_page_size(void) 231 { 232 if (cnt.v_page_size == 0) 233 cnt.v_page_size = PAGE_SIZE; 234 if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0) 235 panic("vm_set_page_size: page size not a power of two"); 236 } 237 238 /* 239 * vm_page_blacklist_lookup: 240 * 241 * See if a physical address in this page has been listed 242 * in the blacklist tunable. Entries in the tunable are 243 * separated by spaces or commas. If an invalid integer is 244 * encountered then the rest of the string is skipped. 245 */ 246 static int 247 vm_page_blacklist_lookup(char *list, vm_paddr_t pa) 248 { 249 vm_paddr_t bad; 250 char *cp, *pos; 251 252 for (pos = list; *pos != '\0'; pos = cp) { 253 bad = strtoq(pos, &cp, 0); 254 if (*cp != '\0') { 255 if (*cp == ' ' || *cp == ',') { 256 cp++; 257 if (cp == pos) 258 continue; 259 } else 260 break; 261 } 262 if (pa == trunc_page(bad)) 263 return (1); 264 } 265 return (0); 266 } 267 268 /* 269 * vm_page_startup: 270 * 271 * Initializes the resident memory module. 272 * 273 * Allocates memory for the page cells, and 274 * for the object/offset-to-page hash table headers. 275 * Each page cell is initialized and placed on the free list. 276 */ 277 vm_offset_t 278 vm_page_startup(vm_offset_t vaddr) 279 { 280 vm_offset_t mapped; 281 vm_paddr_t page_range; 282 vm_paddr_t new_end; 283 int i; 284 vm_paddr_t pa; 285 int nblocks; 286 vm_paddr_t last_pa; 287 char *list; 288 289 /* the biggest memory array is the second group of pages */ 290 vm_paddr_t end; 291 vm_paddr_t biggestsize; 292 vm_paddr_t low_water, high_water; 293 int biggestone; 294 295 biggestsize = 0; 296 biggestone = 0; 297 nblocks = 0; 298 vaddr = round_page(vaddr); 299 300 for (i = 0; phys_avail[i + 1]; i += 2) { 301 phys_avail[i] = round_page(phys_avail[i]); 302 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 303 } 304 305 low_water = phys_avail[0]; 306 high_water = phys_avail[1]; 307 308 for (i = 0; phys_avail[i + 1]; i += 2) { 309 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i]; 310 311 if (size > biggestsize) { 312 biggestone = i; 313 biggestsize = size; 314 } 315 if (phys_avail[i] < low_water) 316 low_water = phys_avail[i]; 317 if (phys_avail[i + 1] > high_water) 318 high_water = phys_avail[i + 1]; 319 ++nblocks; 320 } 321 322 #ifdef XEN 323 low_water = 0; 324 #endif 325 326 end = phys_avail[biggestone+1]; 327 328 /* 329 * Initialize the locks. 330 */ 331 mtx_init(&vm_page_queue_mtx, "vm page queue mutex", NULL, MTX_DEF | 332 MTX_RECURSE); 333 mtx_init(&vm_page_queue_free_mtx, "vm page queue free mutex", NULL, 334 MTX_DEF); 335 336 /* Setup page locks. */ 337 for (i = 0; i < PA_LOCK_COUNT; i++) 338 mtx_init(&pa_lock[i].data, "page lock", NULL, 339 MTX_DEF | MTX_RECURSE | MTX_DUPOK); 340 341 /* 342 * Initialize the queue headers for the hold queue, the active queue, 343 * and the inactive queue. 344 */ 345 for (i = 0; i < PQ_COUNT; i++) 346 TAILQ_INIT(&vm_page_queues[i].pl); 347 vm_page_queues[PQ_INACTIVE].cnt = &cnt.v_inactive_count; 348 vm_page_queues[PQ_ACTIVE].cnt = &cnt.v_active_count; 349 vm_page_queues[PQ_HOLD].cnt = &cnt.v_active_count; 350 351 /* 352 * Allocate memory for use when boot strapping the kernel memory 353 * allocator. 354 */ 355 new_end = end - (boot_pages * UMA_SLAB_SIZE); 356 new_end = trunc_page(new_end); 357 mapped = pmap_map(&vaddr, new_end, end, 358 VM_PROT_READ | VM_PROT_WRITE); 359 bzero((void *)mapped, end - new_end); 360 uma_startup((void *)mapped, boot_pages); 361 362 #if defined(__amd64__) || defined(__i386__) || defined(__arm__) || \ 363 defined(__mips__) 364 /* 365 * Allocate a bitmap to indicate that a random physical page 366 * needs to be included in a minidump. 367 * 368 * The amd64 port needs this to indicate which direct map pages 369 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 370 * 371 * However, i386 still needs this workspace internally within the 372 * minidump code. In theory, they are not needed on i386, but are 373 * included should the sf_buf code decide to use them. 374 */ 375 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE; 376 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 377 new_end -= vm_page_dump_size; 378 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, 379 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); 380 bzero((void *)vm_page_dump, vm_page_dump_size); 381 #endif 382 #ifdef __amd64__ 383 /* 384 * Request that the physical pages underlying the message buffer be 385 * included in a crash dump. Since the message buffer is accessed 386 * through the direct map, they are not automatically included. 387 */ 388 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); 389 last_pa = pa + round_page(MSGBUF_SIZE); 390 while (pa < last_pa) { 391 dump_add_page(pa); 392 pa += PAGE_SIZE; 393 } 394 #endif 395 /* 396 * Compute the number of pages of memory that will be available for 397 * use (taking into account the overhead of a page structure per 398 * page). 399 */ 400 first_page = low_water / PAGE_SIZE; 401 #ifdef VM_PHYSSEG_SPARSE 402 page_range = 0; 403 for (i = 0; phys_avail[i + 1] != 0; i += 2) 404 page_range += atop(phys_avail[i + 1] - phys_avail[i]); 405 #elif defined(VM_PHYSSEG_DENSE) 406 page_range = high_water / PAGE_SIZE - first_page; 407 #else 408 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 409 #endif 410 end = new_end; 411 412 /* 413 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 414 */ 415 vaddr += PAGE_SIZE; 416 417 /* 418 * Initialize the mem entry structures now, and put them in the free 419 * queue. 420 */ 421 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 422 mapped = pmap_map(&vaddr, new_end, end, 423 VM_PROT_READ | VM_PROT_WRITE); 424 vm_page_array = (vm_page_t) mapped; 425 #if VM_NRESERVLEVEL > 0 426 /* 427 * Allocate memory for the reservation management system's data 428 * structures. 429 */ 430 new_end = vm_reserv_startup(&vaddr, new_end, high_water); 431 #endif 432 #ifdef __amd64__ 433 /* 434 * pmap_map on amd64 comes out of the direct-map, not kvm like i386, 435 * so the pages must be tracked for a crashdump to include this data. 436 * This includes the vm_page_array and the early UMA bootstrap pages. 437 */ 438 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE) 439 dump_add_page(pa); 440 #endif 441 phys_avail[biggestone + 1] = new_end; 442 443 /* 444 * Clear all of the page structures 445 */ 446 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 447 for (i = 0; i < page_range; i++) 448 vm_page_array[i].order = VM_NFREEORDER; 449 vm_page_array_size = page_range; 450 451 /* 452 * Initialize the physical memory allocator. 453 */ 454 vm_phys_init(); 455 456 /* 457 * Add every available physical page that is not blacklisted to 458 * the free lists. 459 */ 460 cnt.v_page_count = 0; 461 cnt.v_free_count = 0; 462 list = getenv("vm.blacklist"); 463 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 464 pa = phys_avail[i]; 465 last_pa = phys_avail[i + 1]; 466 while (pa < last_pa) { 467 if (list != NULL && 468 vm_page_blacklist_lookup(list, pa)) 469 printf("Skipping page with pa 0x%jx\n", 470 (uintmax_t)pa); 471 else 472 vm_phys_add_page(pa); 473 pa += PAGE_SIZE; 474 } 475 } 476 freeenv(list); 477 #if VM_NRESERVLEVEL > 0 478 /* 479 * Initialize the reservation management system. 480 */ 481 vm_reserv_init(); 482 #endif 483 return (vaddr); 484 } 485 486 void 487 vm_page_flag_set(vm_page_t m, unsigned short bits) 488 { 489 490 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 491 /* 492 * The PG_WRITEABLE flag can only be set if the page is managed and 493 * VPO_BUSY. Currently, this flag is only set by pmap_enter(). 494 */ 495 KASSERT((bits & PG_WRITEABLE) == 0 || 496 ((m->flags & (PG_UNMANAGED | PG_FICTITIOUS)) == 0 && 497 (m->oflags & VPO_BUSY) != 0), ("PG_WRITEABLE and !VPO_BUSY")); 498 m->flags |= bits; 499 } 500 501 void 502 vm_page_flag_clear(vm_page_t m, unsigned short bits) 503 { 504 505 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 506 /* 507 * The PG_REFERENCED flag can only be cleared if the object 508 * containing the page is locked. 509 */ 510 KASSERT((bits & PG_REFERENCED) == 0 || VM_OBJECT_LOCKED(m->object), 511 ("PG_REFERENCED and !VM_OBJECT_LOCKED")); 512 m->flags &= ~bits; 513 } 514 515 void 516 vm_page_busy(vm_page_t m) 517 { 518 519 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 520 KASSERT((m->oflags & VPO_BUSY) == 0, 521 ("vm_page_busy: page already busy!!!")); 522 m->oflags |= VPO_BUSY; 523 } 524 525 /* 526 * vm_page_flash: 527 * 528 * wakeup anyone waiting for the page. 529 */ 530 void 531 vm_page_flash(vm_page_t m) 532 { 533 534 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 535 if (m->oflags & VPO_WANTED) { 536 m->oflags &= ~VPO_WANTED; 537 wakeup(m); 538 } 539 } 540 541 /* 542 * vm_page_wakeup: 543 * 544 * clear the VPO_BUSY flag and wakeup anyone waiting for the 545 * page. 546 * 547 */ 548 void 549 vm_page_wakeup(vm_page_t m) 550 { 551 552 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 553 KASSERT(m->oflags & VPO_BUSY, ("vm_page_wakeup: page not busy!!!")); 554 m->oflags &= ~VPO_BUSY; 555 vm_page_flash(m); 556 } 557 558 void 559 vm_page_io_start(vm_page_t m) 560 { 561 562 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 563 m->busy++; 564 } 565 566 void 567 vm_page_io_finish(vm_page_t m) 568 { 569 570 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 571 m->busy--; 572 if (m->busy == 0) 573 vm_page_flash(m); 574 } 575 576 /* 577 * Keep page from being freed by the page daemon 578 * much of the same effect as wiring, except much lower 579 * overhead and should be used only for *very* temporary 580 * holding ("wiring"). 581 */ 582 void 583 vm_page_hold(vm_page_t mem) 584 { 585 586 vm_page_lock_assert(mem, MA_OWNED); 587 mem->hold_count++; 588 } 589 590 void 591 vm_page_unhold(vm_page_t mem) 592 { 593 594 vm_page_lock_assert(mem, MA_OWNED); 595 --mem->hold_count; 596 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!")); 597 if (mem->hold_count == 0 && mem->queue == PQ_HOLD) 598 vm_page_free_toq(mem); 599 } 600 601 /* 602 * vm_page_free: 603 * 604 * Free a page. 605 */ 606 void 607 vm_page_free(vm_page_t m) 608 { 609 610 m->flags &= ~PG_ZERO; 611 vm_page_free_toq(m); 612 } 613 614 /* 615 * vm_page_free_zero: 616 * 617 * Free a page to the zerod-pages queue 618 */ 619 void 620 vm_page_free_zero(vm_page_t m) 621 { 622 623 m->flags |= PG_ZERO; 624 vm_page_free_toq(m); 625 } 626 627 /* 628 * vm_page_sleep: 629 * 630 * Sleep and release the page and page queues locks. 631 * 632 * The object containing the given page must be locked. 633 */ 634 void 635 vm_page_sleep(vm_page_t m, const char *msg) 636 { 637 638 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 639 if (mtx_owned(&vm_page_queue_mtx)) 640 vm_page_unlock_queues(); 641 if (mtx_owned(vm_page_lockptr(m))) 642 vm_page_unlock(m); 643 644 /* 645 * It's possible that while we sleep, the page will get 646 * unbusied and freed. If we are holding the object 647 * lock, we will assume we hold a reference to the object 648 * such that even if m->object changes, we can re-lock 649 * it. 650 */ 651 m->oflags |= VPO_WANTED; 652 msleep(m, VM_OBJECT_MTX(m->object), PVM, msg, 0); 653 } 654 655 /* 656 * vm_page_dirty: 657 * 658 * make page all dirty 659 */ 660 void 661 vm_page_dirty(vm_page_t m) 662 { 663 664 KASSERT((m->flags & PG_CACHED) == 0, 665 ("vm_page_dirty: page in cache!")); 666 KASSERT(!VM_PAGE_IS_FREE(m), 667 ("vm_page_dirty: page is free!")); 668 KASSERT(m->valid == VM_PAGE_BITS_ALL, 669 ("vm_page_dirty: page is invalid!")); 670 m->dirty = VM_PAGE_BITS_ALL; 671 } 672 673 /* 674 * vm_page_splay: 675 * 676 * Implements Sleator and Tarjan's top-down splay algorithm. Returns 677 * the vm_page containing the given pindex. If, however, that 678 * pindex is not found in the vm_object, returns a vm_page that is 679 * adjacent to the pindex, coming before or after it. 680 */ 681 vm_page_t 682 vm_page_splay(vm_pindex_t pindex, vm_page_t root) 683 { 684 struct vm_page dummy; 685 vm_page_t lefttreemax, righttreemin, y; 686 687 if (root == NULL) 688 return (root); 689 lefttreemax = righttreemin = &dummy; 690 for (;; root = y) { 691 if (pindex < root->pindex) { 692 if ((y = root->left) == NULL) 693 break; 694 if (pindex < y->pindex) { 695 /* Rotate right. */ 696 root->left = y->right; 697 y->right = root; 698 root = y; 699 if ((y = root->left) == NULL) 700 break; 701 } 702 /* Link into the new root's right tree. */ 703 righttreemin->left = root; 704 righttreemin = root; 705 } else if (pindex > root->pindex) { 706 if ((y = root->right) == NULL) 707 break; 708 if (pindex > y->pindex) { 709 /* Rotate left. */ 710 root->right = y->left; 711 y->left = root; 712 root = y; 713 if ((y = root->right) == NULL) 714 break; 715 } 716 /* Link into the new root's left tree. */ 717 lefttreemax->right = root; 718 lefttreemax = root; 719 } else 720 break; 721 } 722 /* Assemble the new root. */ 723 lefttreemax->right = root->left; 724 righttreemin->left = root->right; 725 root->left = dummy.right; 726 root->right = dummy.left; 727 return (root); 728 } 729 730 /* 731 * vm_page_insert: [ internal use only ] 732 * 733 * Inserts the given mem entry into the object and object list. 734 * 735 * The pagetables are not updated but will presumably fault the page 736 * in if necessary, or if a kernel page the caller will at some point 737 * enter the page into the kernel's pmap. We are not allowed to block 738 * here so we *can't* do this anyway. 739 * 740 * The object and page must be locked. 741 * This routine may not block. 742 */ 743 void 744 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 745 { 746 vm_page_t root; 747 748 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 749 if (m->object != NULL) 750 panic("vm_page_insert: page already inserted"); 751 752 /* 753 * Record the object/offset pair in this page 754 */ 755 m->object = object; 756 m->pindex = pindex; 757 758 /* 759 * Now link into the object's ordered list of backed pages. 760 */ 761 root = object->root; 762 if (root == NULL) { 763 m->left = NULL; 764 m->right = NULL; 765 TAILQ_INSERT_TAIL(&object->memq, m, listq); 766 } else { 767 root = vm_page_splay(pindex, root); 768 if (pindex < root->pindex) { 769 m->left = root->left; 770 m->right = root; 771 root->left = NULL; 772 TAILQ_INSERT_BEFORE(root, m, listq); 773 } else if (pindex == root->pindex) 774 panic("vm_page_insert: offset already allocated"); 775 else { 776 m->right = root->right; 777 m->left = root; 778 root->right = NULL; 779 TAILQ_INSERT_AFTER(&object->memq, root, m, listq); 780 } 781 } 782 object->root = m; 783 object->generation++; 784 785 /* 786 * show that the object has one more resident page. 787 */ 788 object->resident_page_count++; 789 /* 790 * Hold the vnode until the last page is released. 791 */ 792 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 793 vhold((struct vnode *)object->handle); 794 795 /* 796 * Since we are inserting a new and possibly dirty page, 797 * update the object's OBJ_MIGHTBEDIRTY flag. 798 */ 799 if (m->flags & PG_WRITEABLE) 800 vm_object_set_writeable_dirty(object); 801 } 802 803 /* 804 * vm_page_remove: 805 * NOTE: used by device pager as well -wfj 806 * 807 * Removes the given mem entry from the object/offset-page 808 * table and the object page list, but do not invalidate/terminate 809 * the backing store. 810 * 811 * The object and page must be locked. 812 * The underlying pmap entry (if any) is NOT removed here. 813 * This routine may not block. 814 */ 815 void 816 vm_page_remove(vm_page_t m) 817 { 818 vm_object_t object; 819 vm_page_t root; 820 821 if ((m->flags & PG_UNMANAGED) == 0) 822 vm_page_lock_assert(m, MA_OWNED); 823 if ((object = m->object) == NULL) 824 return; 825 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 826 if (m->oflags & VPO_BUSY) { 827 m->oflags &= ~VPO_BUSY; 828 vm_page_flash(m); 829 } 830 831 /* 832 * Now remove from the object's list of backed pages. 833 */ 834 if (m != object->root) 835 vm_page_splay(m->pindex, object->root); 836 if (m->left == NULL) 837 root = m->right; 838 else { 839 root = vm_page_splay(m->pindex, m->left); 840 root->right = m->right; 841 } 842 object->root = root; 843 TAILQ_REMOVE(&object->memq, m, listq); 844 845 /* 846 * And show that the object has one fewer resident page. 847 */ 848 object->resident_page_count--; 849 object->generation++; 850 /* 851 * The vnode may now be recycled. 852 */ 853 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 854 vdrop((struct vnode *)object->handle); 855 856 m->object = NULL; 857 } 858 859 /* 860 * vm_page_lookup: 861 * 862 * Returns the page associated with the object/offset 863 * pair specified; if none is found, NULL is returned. 864 * 865 * The object must be locked. 866 * This routine may not block. 867 * This is a critical path routine 868 */ 869 vm_page_t 870 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 871 { 872 vm_page_t m; 873 874 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 875 if ((m = object->root) != NULL && m->pindex != pindex) { 876 m = vm_page_splay(pindex, m); 877 if ((object->root = m)->pindex != pindex) 878 m = NULL; 879 } 880 return (m); 881 } 882 883 /* 884 * vm_page_find_least: 885 * 886 * Returns the page associated with the object with least pindex 887 * greater than or equal to the parameter pindex, or NULL. 888 * 889 * The object must be locked. 890 * The routine may not block. 891 */ 892 vm_page_t 893 vm_page_find_least(vm_object_t object, vm_pindex_t pindex) 894 { 895 vm_page_t m; 896 897 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 898 if ((m = TAILQ_FIRST(&object->memq)) != NULL) { 899 if (m->pindex < pindex) { 900 m = vm_page_splay(pindex, object->root); 901 if ((object->root = m)->pindex < pindex) 902 m = TAILQ_NEXT(m, listq); 903 } 904 } 905 return (m); 906 } 907 908 /* 909 * Returns the given page's successor (by pindex) within the object if it is 910 * resident; if none is found, NULL is returned. 911 * 912 * The object must be locked. 913 */ 914 vm_page_t 915 vm_page_next(vm_page_t m) 916 { 917 vm_page_t next; 918 919 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 920 if ((next = TAILQ_NEXT(m, listq)) != NULL && 921 next->pindex != m->pindex + 1) 922 next = NULL; 923 return (next); 924 } 925 926 /* 927 * Returns the given page's predecessor (by pindex) within the object if it is 928 * resident; if none is found, NULL is returned. 929 * 930 * The object must be locked. 931 */ 932 vm_page_t 933 vm_page_prev(vm_page_t m) 934 { 935 vm_page_t prev; 936 937 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 938 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL && 939 prev->pindex != m->pindex - 1) 940 prev = NULL; 941 return (prev); 942 } 943 944 /* 945 * vm_page_rename: 946 * 947 * Move the given memory entry from its 948 * current object to the specified target object/offset. 949 * 950 * The object must be locked. 951 * This routine may not block. 952 * 953 * Note: swap associated with the page must be invalidated by the move. We 954 * have to do this for several reasons: (1) we aren't freeing the 955 * page, (2) we are dirtying the page, (3) the VM system is probably 956 * moving the page from object A to B, and will then later move 957 * the backing store from A to B and we can't have a conflict. 958 * 959 * Note: we *always* dirty the page. It is necessary both for the 960 * fact that we moved it, and because we may be invalidating 961 * swap. If the page is on the cache, we have to deactivate it 962 * or vm_page_dirty() will panic. Dirty pages are not allowed 963 * on the cache. 964 */ 965 void 966 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 967 { 968 969 vm_page_remove(m); 970 vm_page_insert(m, new_object, new_pindex); 971 vm_page_dirty(m); 972 } 973 974 /* 975 * Convert all of the given object's cached pages that have a 976 * pindex within the given range into free pages. If the value 977 * zero is given for "end", then the range's upper bound is 978 * infinity. If the given object is backed by a vnode and it 979 * transitions from having one or more cached pages to none, the 980 * vnode's hold count is reduced. 981 */ 982 void 983 vm_page_cache_free(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 984 { 985 vm_page_t m, m_next; 986 boolean_t empty; 987 988 mtx_lock(&vm_page_queue_free_mtx); 989 if (__predict_false(object->cache == NULL)) { 990 mtx_unlock(&vm_page_queue_free_mtx); 991 return; 992 } 993 m = object->cache = vm_page_splay(start, object->cache); 994 if (m->pindex < start) { 995 if (m->right == NULL) 996 m = NULL; 997 else { 998 m_next = vm_page_splay(start, m->right); 999 m_next->left = m; 1000 m->right = NULL; 1001 m = object->cache = m_next; 1002 } 1003 } 1004 1005 /* 1006 * At this point, "m" is either (1) a reference to the page 1007 * with the least pindex that is greater than or equal to 1008 * "start" or (2) NULL. 1009 */ 1010 for (; m != NULL && (m->pindex < end || end == 0); m = m_next) { 1011 /* 1012 * Find "m"'s successor and remove "m" from the 1013 * object's cache. 1014 */ 1015 if (m->right == NULL) { 1016 object->cache = m->left; 1017 m_next = NULL; 1018 } else { 1019 m_next = vm_page_splay(start, m->right); 1020 m_next->left = m->left; 1021 object->cache = m_next; 1022 } 1023 /* Convert "m" to a free page. */ 1024 m->object = NULL; 1025 m->valid = 0; 1026 /* Clear PG_CACHED and set PG_FREE. */ 1027 m->flags ^= PG_CACHED | PG_FREE; 1028 KASSERT((m->flags & (PG_CACHED | PG_FREE)) == PG_FREE, 1029 ("vm_page_cache_free: page %p has inconsistent flags", m)); 1030 cnt.v_cache_count--; 1031 cnt.v_free_count++; 1032 } 1033 empty = object->cache == NULL; 1034 mtx_unlock(&vm_page_queue_free_mtx); 1035 if (object->type == OBJT_VNODE && empty) 1036 vdrop(object->handle); 1037 } 1038 1039 /* 1040 * Returns the cached page that is associated with the given 1041 * object and offset. If, however, none exists, returns NULL. 1042 * 1043 * The free page queue must be locked. 1044 */ 1045 static inline vm_page_t 1046 vm_page_cache_lookup(vm_object_t object, vm_pindex_t pindex) 1047 { 1048 vm_page_t m; 1049 1050 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1051 if ((m = object->cache) != NULL && m->pindex != pindex) { 1052 m = vm_page_splay(pindex, m); 1053 if ((object->cache = m)->pindex != pindex) 1054 m = NULL; 1055 } 1056 return (m); 1057 } 1058 1059 /* 1060 * Remove the given cached page from its containing object's 1061 * collection of cached pages. 1062 * 1063 * The free page queue must be locked. 1064 */ 1065 void 1066 vm_page_cache_remove(vm_page_t m) 1067 { 1068 vm_object_t object; 1069 vm_page_t root; 1070 1071 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1072 KASSERT((m->flags & PG_CACHED) != 0, 1073 ("vm_page_cache_remove: page %p is not cached", m)); 1074 object = m->object; 1075 if (m != object->cache) { 1076 root = vm_page_splay(m->pindex, object->cache); 1077 KASSERT(root == m, 1078 ("vm_page_cache_remove: page %p is not cached in object %p", 1079 m, object)); 1080 } 1081 if (m->left == NULL) 1082 root = m->right; 1083 else if (m->right == NULL) 1084 root = m->left; 1085 else { 1086 root = vm_page_splay(m->pindex, m->left); 1087 root->right = m->right; 1088 } 1089 object->cache = root; 1090 m->object = NULL; 1091 cnt.v_cache_count--; 1092 } 1093 1094 /* 1095 * Transfer all of the cached pages with offset greater than or 1096 * equal to 'offidxstart' from the original object's cache to the 1097 * new object's cache. However, any cached pages with offset 1098 * greater than or equal to the new object's size are kept in the 1099 * original object. Initially, the new object's cache must be 1100 * empty. Offset 'offidxstart' in the original object must 1101 * correspond to offset zero in the new object. 1102 * 1103 * The new object must be locked. 1104 */ 1105 void 1106 vm_page_cache_transfer(vm_object_t orig_object, vm_pindex_t offidxstart, 1107 vm_object_t new_object) 1108 { 1109 vm_page_t m, m_next; 1110 1111 /* 1112 * Insertion into an object's collection of cached pages 1113 * requires the object to be locked. In contrast, removal does 1114 * not. 1115 */ 1116 VM_OBJECT_LOCK_ASSERT(new_object, MA_OWNED); 1117 KASSERT(new_object->cache == NULL, 1118 ("vm_page_cache_transfer: object %p has cached pages", 1119 new_object)); 1120 mtx_lock(&vm_page_queue_free_mtx); 1121 if ((m = orig_object->cache) != NULL) { 1122 /* 1123 * Transfer all of the pages with offset greater than or 1124 * equal to 'offidxstart' from the original object's 1125 * cache to the new object's cache. 1126 */ 1127 m = vm_page_splay(offidxstart, m); 1128 if (m->pindex < offidxstart) { 1129 orig_object->cache = m; 1130 new_object->cache = m->right; 1131 m->right = NULL; 1132 } else { 1133 orig_object->cache = m->left; 1134 new_object->cache = m; 1135 m->left = NULL; 1136 } 1137 while ((m = new_object->cache) != NULL) { 1138 if ((m->pindex - offidxstart) >= new_object->size) { 1139 /* 1140 * Return all of the cached pages with 1141 * offset greater than or equal to the 1142 * new object's size to the original 1143 * object's cache. 1144 */ 1145 new_object->cache = m->left; 1146 m->left = orig_object->cache; 1147 orig_object->cache = m; 1148 break; 1149 } 1150 m_next = vm_page_splay(m->pindex, m->right); 1151 /* Update the page's object and offset. */ 1152 m->object = new_object; 1153 m->pindex -= offidxstart; 1154 if (m_next == NULL) 1155 break; 1156 m->right = NULL; 1157 m_next->left = m; 1158 new_object->cache = m_next; 1159 } 1160 KASSERT(new_object->cache == NULL || 1161 new_object->type == OBJT_SWAP, 1162 ("vm_page_cache_transfer: object %p's type is incompatible" 1163 " with cached pages", new_object)); 1164 } 1165 mtx_unlock(&vm_page_queue_free_mtx); 1166 } 1167 1168 /* 1169 * vm_page_alloc: 1170 * 1171 * Allocate and return a memory cell associated 1172 * with this VM object/offset pair. 1173 * 1174 * The caller must always specify an allocation class. 1175 * 1176 * allocation classes: 1177 * VM_ALLOC_NORMAL normal process request 1178 * VM_ALLOC_SYSTEM system *really* needs a page 1179 * VM_ALLOC_INTERRUPT interrupt time request 1180 * 1181 * optional allocation flags: 1182 * VM_ALLOC_ZERO prefer a zeroed page 1183 * VM_ALLOC_WIRED wire the allocated page 1184 * VM_ALLOC_NOOBJ page is not associated with a vm object 1185 * VM_ALLOC_NOBUSY do not set the page busy 1186 * VM_ALLOC_IFCACHED return page only if it is cached 1187 * VM_ALLOC_IFNOTCACHED return NULL, do not reactivate if the page 1188 * is cached 1189 * 1190 * This routine may not sleep. 1191 */ 1192 vm_page_t 1193 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 1194 { 1195 struct vnode *vp = NULL; 1196 vm_object_t m_object; 1197 vm_page_t m; 1198 int flags, page_req; 1199 1200 page_req = req & VM_ALLOC_CLASS_MASK; 1201 KASSERT(curthread->td_intr_nesting_level == 0 || 1202 page_req == VM_ALLOC_INTERRUPT, 1203 ("vm_page_alloc(NORMAL|SYSTEM) in interrupt context")); 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 /* 1212 * The pager is allowed to eat deeper into the free page list. 1213 */ 1214 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) { 1215 page_req = VM_ALLOC_SYSTEM; 1216 }; 1217 1218 mtx_lock(&vm_page_queue_free_mtx); 1219 if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved || 1220 (page_req == VM_ALLOC_SYSTEM && 1221 cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) || 1222 (page_req == VM_ALLOC_INTERRUPT && 1223 cnt.v_free_count + cnt.v_cache_count > 0)) { 1224 /* 1225 * Allocate from the free queue if the number of free pages 1226 * exceeds the minimum for the request class. 1227 */ 1228 if (object != NULL && 1229 (m = vm_page_cache_lookup(object, pindex)) != NULL) { 1230 if ((req & VM_ALLOC_IFNOTCACHED) != 0) { 1231 mtx_unlock(&vm_page_queue_free_mtx); 1232 return (NULL); 1233 } 1234 if (vm_phys_unfree_page(m)) 1235 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0); 1236 #if VM_NRESERVLEVEL > 0 1237 else if (!vm_reserv_reactivate_page(m)) 1238 #else 1239 else 1240 #endif 1241 panic("vm_page_alloc: cache page %p is missing" 1242 " from the free queue", m); 1243 } else if ((req & VM_ALLOC_IFCACHED) != 0) { 1244 mtx_unlock(&vm_page_queue_free_mtx); 1245 return (NULL); 1246 #if VM_NRESERVLEVEL > 0 1247 } else if (object == NULL || object->type == OBJT_DEVICE || 1248 object->type == OBJT_SG || 1249 (object->flags & OBJ_COLORED) == 0 || 1250 (m = vm_reserv_alloc_page(object, pindex)) == NULL) { 1251 #else 1252 } else { 1253 #endif 1254 m = vm_phys_alloc_pages(object != NULL ? 1255 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0); 1256 #if VM_NRESERVLEVEL > 0 1257 if (m == NULL && vm_reserv_reclaim_inactive()) { 1258 m = vm_phys_alloc_pages(object != NULL ? 1259 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 1260 0); 1261 } 1262 #endif 1263 } 1264 } else { 1265 /* 1266 * Not allocatable, give up. 1267 */ 1268 mtx_unlock(&vm_page_queue_free_mtx); 1269 atomic_add_int(&vm_pageout_deficit, 1270 MAX((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 1271 pagedaemon_wakeup(); 1272 return (NULL); 1273 } 1274 1275 /* 1276 * At this point we had better have found a good page. 1277 */ 1278 1279 KASSERT(m != NULL, ("vm_page_alloc: missing page")); 1280 KASSERT(m->queue == PQ_NONE, 1281 ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue)); 1282 KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m)); 1283 KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m)); 1284 KASSERT(m->busy == 0, ("vm_page_alloc: page %p is busy", m)); 1285 KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m)); 1286 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 1287 ("vm_page_alloc: page %p has unexpected memattr %d", m, 1288 pmap_page_get_memattr(m))); 1289 if ((m->flags & PG_CACHED) != 0) { 1290 KASSERT(m->valid != 0, 1291 ("vm_page_alloc: cached page %p is invalid", m)); 1292 if (m->object == object && m->pindex == pindex) 1293 cnt.v_reactivated++; 1294 else 1295 m->valid = 0; 1296 m_object = m->object; 1297 vm_page_cache_remove(m); 1298 if (m_object->type == OBJT_VNODE && m_object->cache == NULL) 1299 vp = m_object->handle; 1300 } else { 1301 KASSERT(VM_PAGE_IS_FREE(m), 1302 ("vm_page_alloc: page %p is not free", m)); 1303 KASSERT(m->valid == 0, 1304 ("vm_page_alloc: free page %p is valid", m)); 1305 cnt.v_free_count--; 1306 } 1307 1308 /* 1309 * Initialize structure. Only the PG_ZERO flag is inherited. 1310 */ 1311 flags = 0; 1312 if (m->flags & PG_ZERO) { 1313 vm_page_zero_count--; 1314 if (req & VM_ALLOC_ZERO) 1315 flags = PG_ZERO; 1316 } 1317 if (object == NULL || object->type == OBJT_PHYS) 1318 flags |= PG_UNMANAGED; 1319 m->flags = flags; 1320 if (req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ)) 1321 m->oflags = 0; 1322 else 1323 m->oflags = VPO_BUSY; 1324 if (req & VM_ALLOC_WIRED) { 1325 atomic_add_int(&cnt.v_wire_count, 1); 1326 m->wire_count = 1; 1327 } 1328 m->act_count = 0; 1329 mtx_unlock(&vm_page_queue_free_mtx); 1330 1331 if (object != NULL) { 1332 /* Ignore device objects; the pager sets "memattr" for them. */ 1333 if (object->memattr != VM_MEMATTR_DEFAULT && 1334 object->type != OBJT_DEVICE && object->type != OBJT_SG) 1335 pmap_page_set_memattr(m, object->memattr); 1336 vm_page_insert(m, object, pindex); 1337 } else 1338 m->pindex = pindex; 1339 1340 /* 1341 * The following call to vdrop() must come after the above call 1342 * to vm_page_insert() in case both affect the same object and 1343 * vnode. Otherwise, the affected vnode's hold count could 1344 * temporarily become zero. 1345 */ 1346 if (vp != NULL) 1347 vdrop(vp); 1348 1349 /* 1350 * Don't wakeup too often - wakeup the pageout daemon when 1351 * we would be nearly out of memory. 1352 */ 1353 if (vm_paging_needed()) 1354 pagedaemon_wakeup(); 1355 1356 return (m); 1357 } 1358 1359 /* 1360 * Initialize a page that has been freshly dequeued from a freelist. 1361 * The caller has to drop the vnode returned, if it is not NULL. 1362 * 1363 * To be called with vm_page_queue_free_mtx held. 1364 */ 1365 struct vnode * 1366 vm_page_alloc_init(vm_page_t m) 1367 { 1368 struct vnode *drop; 1369 vm_object_t m_object; 1370 1371 KASSERT(m->queue == PQ_NONE, 1372 ("vm_page_alloc_init: page %p has unexpected queue %d", 1373 m, m->queue)); 1374 KASSERT(m->wire_count == 0, 1375 ("vm_page_alloc_init: page %p is wired", m)); 1376 KASSERT(m->hold_count == 0, 1377 ("vm_page_alloc_init: page %p is held", m)); 1378 KASSERT(m->busy == 0, 1379 ("vm_page_alloc_init: page %p is busy", m)); 1380 KASSERT(m->dirty == 0, 1381 ("vm_page_alloc_init: page %p is dirty", m)); 1382 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 1383 ("vm_page_alloc_init: page %p has unexpected memattr %d", 1384 m, pmap_page_get_memattr(m))); 1385 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1386 drop = NULL; 1387 if ((m->flags & PG_CACHED) != 0) { 1388 m->valid = 0; 1389 m_object = m->object; 1390 vm_page_cache_remove(m); 1391 if (m_object->type == OBJT_VNODE && 1392 m_object->cache == NULL) 1393 drop = m_object->handle; 1394 } else { 1395 KASSERT(VM_PAGE_IS_FREE(m), 1396 ("vm_page_alloc_init: page %p is not free", m)); 1397 KASSERT(m->valid == 0, 1398 ("vm_page_alloc_init: free page %p is valid", m)); 1399 cnt.v_free_count--; 1400 } 1401 if (m->flags & PG_ZERO) 1402 vm_page_zero_count--; 1403 /* Don't clear the PG_ZERO flag; we'll need it later. */ 1404 m->flags = PG_UNMANAGED | (m->flags & PG_ZERO); 1405 m->oflags = 0; 1406 /* Unmanaged pages don't use "act_count". */ 1407 return (drop); 1408 } 1409 1410 /* 1411 * vm_page_alloc_freelist: 1412 * 1413 * Allocate a page from the specified freelist with specified order. 1414 * Only the ALLOC_CLASS values in req are honored, other request flags 1415 * are ignored. 1416 */ 1417 vm_page_t 1418 vm_page_alloc_freelist(int flind, int order, int req) 1419 { 1420 struct vnode *drop; 1421 vm_page_t m; 1422 int page_req; 1423 1424 m = NULL; 1425 page_req = req & VM_ALLOC_CLASS_MASK; 1426 mtx_lock(&vm_page_queue_free_mtx); 1427 /* 1428 * Do not allocate reserved pages unless the req has asked for it. 1429 */ 1430 if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved || 1431 (page_req == VM_ALLOC_SYSTEM && 1432 cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) || 1433 (page_req == VM_ALLOC_INTERRUPT && 1434 cnt.v_free_count + cnt.v_cache_count > 0)) { 1435 m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, order); 1436 } 1437 if (m == NULL) { 1438 mtx_unlock(&vm_page_queue_free_mtx); 1439 return (NULL); 1440 } 1441 drop = vm_page_alloc_init(m); 1442 mtx_unlock(&vm_page_queue_free_mtx); 1443 if (drop) 1444 vdrop(drop); 1445 return (m); 1446 } 1447 1448 /* 1449 * vm_wait: (also see VM_WAIT macro) 1450 * 1451 * Block until free pages are available for allocation 1452 * - Called in various places before memory allocations. 1453 */ 1454 void 1455 vm_wait(void) 1456 { 1457 1458 mtx_lock(&vm_page_queue_free_mtx); 1459 if (curproc == pageproc) { 1460 vm_pageout_pages_needed = 1; 1461 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx, 1462 PDROP | PSWP, "VMWait", 0); 1463 } else { 1464 if (!vm_pages_needed) { 1465 vm_pages_needed = 1; 1466 wakeup(&vm_pages_needed); 1467 } 1468 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM, 1469 "vmwait", 0); 1470 } 1471 } 1472 1473 /* 1474 * vm_waitpfault: (also see VM_WAITPFAULT macro) 1475 * 1476 * Block until free pages are available for allocation 1477 * - Called only in vm_fault so that processes page faulting 1478 * can be easily tracked. 1479 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 1480 * processes will be able to grab memory first. Do not change 1481 * this balance without careful testing first. 1482 */ 1483 void 1484 vm_waitpfault(void) 1485 { 1486 1487 mtx_lock(&vm_page_queue_free_mtx); 1488 if (!vm_pages_needed) { 1489 vm_pages_needed = 1; 1490 wakeup(&vm_pages_needed); 1491 } 1492 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER, 1493 "pfault", 0); 1494 } 1495 1496 /* 1497 * vm_page_requeue: 1498 * 1499 * Move the given page to the tail of its present page queue. 1500 * 1501 * The page queues must be locked. 1502 */ 1503 void 1504 vm_page_requeue(vm_page_t m) 1505 { 1506 struct vpgqueues *vpq; 1507 int queue; 1508 1509 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1510 queue = m->queue; 1511 KASSERT(queue != PQ_NONE, 1512 ("vm_page_requeue: page %p is not queued", m)); 1513 vpq = &vm_page_queues[queue]; 1514 TAILQ_REMOVE(&vpq->pl, m, pageq); 1515 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq); 1516 } 1517 1518 /* 1519 * vm_page_queue_remove: 1520 * 1521 * Remove the given page from the specified queue. 1522 * 1523 * The page and page queues must be locked. 1524 */ 1525 static __inline void 1526 vm_page_queue_remove(int queue, vm_page_t m) 1527 { 1528 struct vpgqueues *pq; 1529 1530 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1531 vm_page_lock_assert(m, MA_OWNED); 1532 pq = &vm_page_queues[queue]; 1533 TAILQ_REMOVE(&pq->pl, m, pageq); 1534 (*pq->cnt)--; 1535 } 1536 1537 /* 1538 * vm_pageq_remove: 1539 * 1540 * Remove a page from its queue. 1541 * 1542 * The given page must be locked. 1543 * This routine may not block. 1544 */ 1545 void 1546 vm_pageq_remove(vm_page_t m) 1547 { 1548 int queue; 1549 1550 vm_page_lock_assert(m, MA_OWNED); 1551 if ((queue = m->queue) != PQ_NONE) { 1552 vm_page_lock_queues(); 1553 m->queue = PQ_NONE; 1554 vm_page_queue_remove(queue, m); 1555 vm_page_unlock_queues(); 1556 } 1557 } 1558 1559 /* 1560 * vm_page_enqueue: 1561 * 1562 * Add the given page to the specified queue. 1563 * 1564 * The page queues must be locked. 1565 */ 1566 static void 1567 vm_page_enqueue(int queue, vm_page_t m) 1568 { 1569 struct vpgqueues *vpq; 1570 1571 vpq = &vm_page_queues[queue]; 1572 m->queue = queue; 1573 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq); 1574 ++*vpq->cnt; 1575 } 1576 1577 /* 1578 * vm_page_activate: 1579 * 1580 * Put the specified page on the active list (if appropriate). 1581 * Ensure that act_count is at least ACT_INIT but do not otherwise 1582 * mess with it. 1583 * 1584 * The page must be locked. 1585 * This routine may not block. 1586 */ 1587 void 1588 vm_page_activate(vm_page_t m) 1589 { 1590 int queue; 1591 1592 vm_page_lock_assert(m, MA_OWNED); 1593 if ((queue = m->queue) != PQ_ACTIVE) { 1594 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1595 if (m->act_count < ACT_INIT) 1596 m->act_count = ACT_INIT; 1597 vm_page_lock_queues(); 1598 if (queue != PQ_NONE) 1599 vm_page_queue_remove(queue, m); 1600 vm_page_enqueue(PQ_ACTIVE, m); 1601 vm_page_unlock_queues(); 1602 } else 1603 KASSERT(queue == PQ_NONE, 1604 ("vm_page_activate: wired page %p is queued", m)); 1605 } else { 1606 if (m->act_count < ACT_INIT) 1607 m->act_count = ACT_INIT; 1608 } 1609 } 1610 1611 /* 1612 * vm_page_free_wakeup: 1613 * 1614 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 1615 * routine is called when a page has been added to the cache or free 1616 * queues. 1617 * 1618 * The page queues must be locked. 1619 * This routine may not block. 1620 */ 1621 static inline void 1622 vm_page_free_wakeup(void) 1623 { 1624 1625 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1626 /* 1627 * if pageout daemon needs pages, then tell it that there are 1628 * some free. 1629 */ 1630 if (vm_pageout_pages_needed && 1631 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 1632 wakeup(&vm_pageout_pages_needed); 1633 vm_pageout_pages_needed = 0; 1634 } 1635 /* 1636 * wakeup processes that are waiting on memory if we hit a 1637 * high water mark. And wakeup scheduler process if we have 1638 * lots of memory. this process will swapin processes. 1639 */ 1640 if (vm_pages_needed && !vm_page_count_min()) { 1641 vm_pages_needed = 0; 1642 wakeup(&cnt.v_free_count); 1643 } 1644 } 1645 1646 /* 1647 * vm_page_free_toq: 1648 * 1649 * Returns the given page to the free list, 1650 * disassociating it with any VM object. 1651 * 1652 * Object and page must be locked prior to entry. 1653 * This routine may not block. 1654 */ 1655 1656 void 1657 vm_page_free_toq(vm_page_t m) 1658 { 1659 1660 if ((m->flags & PG_UNMANAGED) == 0) { 1661 vm_page_lock_assert(m, MA_OWNED); 1662 KASSERT(!pmap_page_is_mapped(m), 1663 ("vm_page_free_toq: freeing mapped page %p", m)); 1664 } 1665 PCPU_INC(cnt.v_tfree); 1666 1667 if (m->busy || VM_PAGE_IS_FREE(m)) { 1668 printf( 1669 "vm_page_free: pindex(%lu), busy(%d), VPO_BUSY(%d), hold(%d)\n", 1670 (u_long)m->pindex, m->busy, (m->oflags & VPO_BUSY) ? 1 : 0, 1671 m->hold_count); 1672 if (VM_PAGE_IS_FREE(m)) 1673 panic("vm_page_free: freeing free page"); 1674 else 1675 panic("vm_page_free: freeing busy page"); 1676 } 1677 1678 /* 1679 * unqueue, then remove page. Note that we cannot destroy 1680 * the page here because we do not want to call the pager's 1681 * callback routine until after we've put the page on the 1682 * appropriate free queue. 1683 */ 1684 if ((m->flags & PG_UNMANAGED) == 0) 1685 vm_pageq_remove(m); 1686 vm_page_remove(m); 1687 1688 /* 1689 * If fictitious remove object association and 1690 * return, otherwise delay object association removal. 1691 */ 1692 if ((m->flags & PG_FICTITIOUS) != 0) { 1693 return; 1694 } 1695 1696 m->valid = 0; 1697 vm_page_undirty(m); 1698 1699 if (m->wire_count != 0) { 1700 if (m->wire_count > 1) { 1701 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1702 m->wire_count, (long)m->pindex); 1703 } 1704 panic("vm_page_free: freeing wired page"); 1705 } 1706 if (m->hold_count != 0) { 1707 m->flags &= ~PG_ZERO; 1708 vm_page_lock_queues(); 1709 vm_page_enqueue(PQ_HOLD, m); 1710 vm_page_unlock_queues(); 1711 } else { 1712 /* 1713 * Restore the default memory attribute to the page. 1714 */ 1715 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 1716 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 1717 1718 /* 1719 * Insert the page into the physical memory allocator's 1720 * cache/free page queues. 1721 */ 1722 mtx_lock(&vm_page_queue_free_mtx); 1723 m->flags |= PG_FREE; 1724 cnt.v_free_count++; 1725 #if VM_NRESERVLEVEL > 0 1726 if (!vm_reserv_free_page(m)) 1727 #else 1728 if (TRUE) 1729 #endif 1730 vm_phys_free_pages(m, 0); 1731 if ((m->flags & PG_ZERO) != 0) 1732 ++vm_page_zero_count; 1733 else 1734 vm_page_zero_idle_wakeup(); 1735 vm_page_free_wakeup(); 1736 mtx_unlock(&vm_page_queue_free_mtx); 1737 } 1738 } 1739 1740 /* 1741 * vm_page_wire: 1742 * 1743 * Mark this page as wired down by yet 1744 * another map, removing it from paging queues 1745 * as necessary. 1746 * 1747 * If the page is fictitious, then its wire count must remain one. 1748 * 1749 * The page must be locked. 1750 * This routine may not block. 1751 */ 1752 void 1753 vm_page_wire(vm_page_t m) 1754 { 1755 1756 /* 1757 * Only bump the wire statistics if the page is not already wired, 1758 * and only unqueue the page if it is on some queue (if it is unmanaged 1759 * it is already off the queues). 1760 */ 1761 vm_page_lock_assert(m, MA_OWNED); 1762 if ((m->flags & PG_FICTITIOUS) != 0) { 1763 KASSERT(m->wire_count == 1, 1764 ("vm_page_wire: fictitious page %p's wire count isn't one", 1765 m)); 1766 return; 1767 } 1768 if (m->wire_count == 0) { 1769 if ((m->flags & PG_UNMANAGED) == 0) 1770 vm_pageq_remove(m); 1771 atomic_add_int(&cnt.v_wire_count, 1); 1772 } 1773 m->wire_count++; 1774 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 1775 } 1776 1777 /* 1778 * vm_page_unwire: 1779 * 1780 * Release one wiring of the specified page, potentially enabling it to be 1781 * paged again. If paging is enabled, then the value of the parameter 1782 * "activate" determines to which queue the page is added. If "activate" is 1783 * non-zero, then the page is added to the active queue. Otherwise, it is 1784 * added to the inactive queue. 1785 * 1786 * However, unless the page belongs to an object, it is not enqueued because 1787 * it cannot be paged out. 1788 * 1789 * If a page is fictitious, then its wire count must alway be one. 1790 * 1791 * A managed page must be locked. 1792 */ 1793 void 1794 vm_page_unwire(vm_page_t m, int activate) 1795 { 1796 1797 if ((m->flags & PG_UNMANAGED) == 0) 1798 vm_page_lock_assert(m, MA_OWNED); 1799 if ((m->flags & PG_FICTITIOUS) != 0) { 1800 KASSERT(m->wire_count == 1, 1801 ("vm_page_unwire: fictitious page %p's wire count isn't one", m)); 1802 return; 1803 } 1804 if (m->wire_count > 0) { 1805 m->wire_count--; 1806 if (m->wire_count == 0) { 1807 atomic_subtract_int(&cnt.v_wire_count, 1); 1808 if ((m->flags & PG_UNMANAGED) != 0 || 1809 m->object == NULL) 1810 return; 1811 vm_page_lock_queues(); 1812 if (activate) 1813 vm_page_enqueue(PQ_ACTIVE, m); 1814 else { 1815 vm_page_flag_clear(m, PG_WINATCFLS); 1816 vm_page_enqueue(PQ_INACTIVE, m); 1817 } 1818 vm_page_unlock_queues(); 1819 } 1820 } else 1821 panic("vm_page_unwire: page %p's wire count is zero", m); 1822 } 1823 1824 /* 1825 * Move the specified page to the inactive queue. 1826 * 1827 * Many pages placed on the inactive queue should actually go 1828 * into the cache, but it is difficult to figure out which. What 1829 * we do instead, if the inactive target is well met, is to put 1830 * clean pages at the head of the inactive queue instead of the tail. 1831 * This will cause them to be moved to the cache more quickly and 1832 * if not actively re-referenced, reclaimed more quickly. If we just 1833 * stick these pages at the end of the inactive queue, heavy filesystem 1834 * meta-data accesses can cause an unnecessary paging load on memory bound 1835 * processes. This optimization causes one-time-use metadata to be 1836 * reused more quickly. 1837 * 1838 * Normally athead is 0 resulting in LRU operation. athead is set 1839 * to 1 if we want this page to be 'as if it were placed in the cache', 1840 * except without unmapping it from the process address space. 1841 * 1842 * This routine may not block. 1843 */ 1844 static inline void 1845 _vm_page_deactivate(vm_page_t m, int athead) 1846 { 1847 int queue; 1848 1849 vm_page_lock_assert(m, MA_OWNED); 1850 1851 /* 1852 * Ignore if already inactive. 1853 */ 1854 if ((queue = m->queue) == PQ_INACTIVE) 1855 return; 1856 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1857 vm_page_lock_queues(); 1858 vm_page_flag_clear(m, PG_WINATCFLS); 1859 if (queue != PQ_NONE) 1860 vm_page_queue_remove(queue, m); 1861 if (athead) 1862 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, 1863 pageq); 1864 else 1865 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, 1866 pageq); 1867 m->queue = PQ_INACTIVE; 1868 cnt.v_inactive_count++; 1869 vm_page_unlock_queues(); 1870 } 1871 } 1872 1873 /* 1874 * Move the specified page to the inactive queue. 1875 * 1876 * The page must be locked. 1877 */ 1878 void 1879 vm_page_deactivate(vm_page_t m) 1880 { 1881 1882 _vm_page_deactivate(m, 0); 1883 } 1884 1885 /* 1886 * vm_page_try_to_cache: 1887 * 1888 * Returns 0 on failure, 1 on success 1889 */ 1890 int 1891 vm_page_try_to_cache(vm_page_t m) 1892 { 1893 1894 vm_page_lock_assert(m, MA_OWNED); 1895 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1896 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1897 (m->oflags & VPO_BUSY) || (m->flags & PG_UNMANAGED)) 1898 return (0); 1899 pmap_remove_all(m); 1900 if (m->dirty) 1901 return (0); 1902 vm_page_cache(m); 1903 return (1); 1904 } 1905 1906 /* 1907 * vm_page_try_to_free() 1908 * 1909 * Attempt to free the page. If we cannot free it, we do nothing. 1910 * 1 is returned on success, 0 on failure. 1911 */ 1912 int 1913 vm_page_try_to_free(vm_page_t m) 1914 { 1915 1916 vm_page_lock_assert(m, MA_OWNED); 1917 if (m->object != NULL) 1918 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1919 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1920 (m->oflags & VPO_BUSY) || (m->flags & PG_UNMANAGED)) 1921 return (0); 1922 pmap_remove_all(m); 1923 if (m->dirty) 1924 return (0); 1925 vm_page_free(m); 1926 return (1); 1927 } 1928 1929 /* 1930 * vm_page_cache 1931 * 1932 * Put the specified page onto the page cache queue (if appropriate). 1933 * 1934 * This routine may not block. 1935 */ 1936 void 1937 vm_page_cache(vm_page_t m) 1938 { 1939 vm_object_t object; 1940 vm_page_t root; 1941 1942 vm_page_lock_assert(m, MA_OWNED); 1943 object = m->object; 1944 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1945 if ((m->flags & PG_UNMANAGED) || (m->oflags & VPO_BUSY) || m->busy || 1946 m->hold_count || m->wire_count) 1947 panic("vm_page_cache: attempting to cache busy page"); 1948 pmap_remove_all(m); 1949 if (m->dirty != 0) 1950 panic("vm_page_cache: page %p is dirty", m); 1951 if (m->valid == 0 || object->type == OBJT_DEFAULT || 1952 (object->type == OBJT_SWAP && 1953 !vm_pager_has_page(object, m->pindex, NULL, NULL))) { 1954 /* 1955 * Hypothesis: A cache-elgible page belonging to a 1956 * default object or swap object but without a backing 1957 * store must be zero filled. 1958 */ 1959 vm_page_free(m); 1960 return; 1961 } 1962 KASSERT((m->flags & PG_CACHED) == 0, 1963 ("vm_page_cache: page %p is already cached", m)); 1964 PCPU_INC(cnt.v_tcached); 1965 1966 /* 1967 * Remove the page from the paging queues. 1968 */ 1969 vm_pageq_remove(m); 1970 1971 /* 1972 * Remove the page from the object's collection of resident 1973 * pages. 1974 */ 1975 if (m != object->root) 1976 vm_page_splay(m->pindex, object->root); 1977 if (m->left == NULL) 1978 root = m->right; 1979 else { 1980 root = vm_page_splay(m->pindex, m->left); 1981 root->right = m->right; 1982 } 1983 object->root = root; 1984 TAILQ_REMOVE(&object->memq, m, listq); 1985 object->resident_page_count--; 1986 object->generation++; 1987 1988 /* 1989 * Restore the default memory attribute to the page. 1990 */ 1991 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 1992 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 1993 1994 /* 1995 * Insert the page into the object's collection of cached pages 1996 * and the physical memory allocator's cache/free page queues. 1997 */ 1998 m->flags &= ~PG_ZERO; 1999 mtx_lock(&vm_page_queue_free_mtx); 2000 m->flags |= PG_CACHED; 2001 cnt.v_cache_count++; 2002 root = object->cache; 2003 if (root == NULL) { 2004 m->left = NULL; 2005 m->right = NULL; 2006 } else { 2007 root = vm_page_splay(m->pindex, root); 2008 if (m->pindex < root->pindex) { 2009 m->left = root->left; 2010 m->right = root; 2011 root->left = NULL; 2012 } else if (__predict_false(m->pindex == root->pindex)) 2013 panic("vm_page_cache: offset already cached"); 2014 else { 2015 m->right = root->right; 2016 m->left = root; 2017 root->right = NULL; 2018 } 2019 } 2020 object->cache = m; 2021 #if VM_NRESERVLEVEL > 0 2022 if (!vm_reserv_free_page(m)) { 2023 #else 2024 if (TRUE) { 2025 #endif 2026 vm_phys_set_pool(VM_FREEPOOL_CACHE, m, 0); 2027 vm_phys_free_pages(m, 0); 2028 } 2029 vm_page_free_wakeup(); 2030 mtx_unlock(&vm_page_queue_free_mtx); 2031 2032 /* 2033 * Increment the vnode's hold count if this is the object's only 2034 * cached page. Decrement the vnode's hold count if this was 2035 * the object's only resident page. 2036 */ 2037 if (object->type == OBJT_VNODE) { 2038 if (root == NULL && object->resident_page_count != 0) 2039 vhold(object->handle); 2040 else if (root != NULL && object->resident_page_count == 0) 2041 vdrop(object->handle); 2042 } 2043 } 2044 2045 /* 2046 * vm_page_dontneed 2047 * 2048 * Cache, deactivate, or do nothing as appropriate. This routine 2049 * is typically used by madvise() MADV_DONTNEED. 2050 * 2051 * Generally speaking we want to move the page into the cache so 2052 * it gets reused quickly. However, this can result in a silly syndrome 2053 * due to the page recycling too quickly. Small objects will not be 2054 * fully cached. On the otherhand, if we move the page to the inactive 2055 * queue we wind up with a problem whereby very large objects 2056 * unnecessarily blow away our inactive and cache queues. 2057 * 2058 * The solution is to move the pages based on a fixed weighting. We 2059 * either leave them alone, deactivate them, or move them to the cache, 2060 * where moving them to the cache has the highest weighting. 2061 * By forcing some pages into other queues we eventually force the 2062 * system to balance the queues, potentially recovering other unrelated 2063 * space from active. The idea is to not force this to happen too 2064 * often. 2065 */ 2066 void 2067 vm_page_dontneed(vm_page_t m) 2068 { 2069 int dnw; 2070 int head; 2071 2072 vm_page_lock_assert(m, MA_OWNED); 2073 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2074 dnw = PCPU_GET(dnweight); 2075 PCPU_INC(dnweight); 2076 2077 /* 2078 * Occasionally leave the page alone. 2079 */ 2080 if ((dnw & 0x01F0) == 0 || m->queue == PQ_INACTIVE) { 2081 if (m->act_count >= ACT_INIT) 2082 --m->act_count; 2083 return; 2084 } 2085 2086 /* 2087 * Clear any references to the page. Otherwise, the page daemon will 2088 * immediately reactivate the page. 2089 * 2090 * Perform the pmap_clear_reference() first. Otherwise, a concurrent 2091 * pmap operation, such as pmap_remove(), could clear a reference in 2092 * the pmap and set PG_REFERENCED on the page before the 2093 * pmap_clear_reference() had completed. Consequently, the page would 2094 * appear referenced based upon an old reference that occurred before 2095 * this function ran. 2096 */ 2097 pmap_clear_reference(m); 2098 vm_page_lock_queues(); 2099 vm_page_flag_clear(m, PG_REFERENCED); 2100 vm_page_unlock_queues(); 2101 2102 if (m->dirty == 0 && pmap_is_modified(m)) 2103 vm_page_dirty(m); 2104 2105 if (m->dirty || (dnw & 0x0070) == 0) { 2106 /* 2107 * Deactivate the page 3 times out of 32. 2108 */ 2109 head = 0; 2110 } else { 2111 /* 2112 * Cache the page 28 times out of every 32. Note that 2113 * the page is deactivated instead of cached, but placed 2114 * at the head of the queue instead of the tail. 2115 */ 2116 head = 1; 2117 } 2118 _vm_page_deactivate(m, head); 2119 } 2120 2121 /* 2122 * Grab a page, waiting until we are waken up due to the page 2123 * changing state. We keep on waiting, if the page continues 2124 * to be in the object. If the page doesn't exist, first allocate it 2125 * and then conditionally zero it. 2126 * 2127 * The caller must always specify the VM_ALLOC_RETRY flag. This is intended 2128 * to facilitate its eventual removal. 2129 * 2130 * This routine may block. 2131 */ 2132 vm_page_t 2133 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 2134 { 2135 vm_page_t m; 2136 2137 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2138 KASSERT((allocflags & VM_ALLOC_RETRY) != 0, 2139 ("vm_page_grab: VM_ALLOC_RETRY is required")); 2140 retrylookup: 2141 if ((m = vm_page_lookup(object, pindex)) != NULL) { 2142 if ((m->oflags & VPO_BUSY) != 0 || 2143 ((allocflags & VM_ALLOC_IGN_SBUSY) == 0 && m->busy != 0)) { 2144 /* 2145 * Reference the page before unlocking and 2146 * sleeping so that the page daemon is less 2147 * likely to reclaim it. 2148 */ 2149 vm_page_lock_queues(); 2150 vm_page_flag_set(m, PG_REFERENCED); 2151 vm_page_sleep(m, "pgrbwt"); 2152 goto retrylookup; 2153 } else { 2154 if ((allocflags & VM_ALLOC_WIRED) != 0) { 2155 vm_page_lock(m); 2156 vm_page_wire(m); 2157 vm_page_unlock(m); 2158 } 2159 if ((allocflags & VM_ALLOC_NOBUSY) == 0) 2160 vm_page_busy(m); 2161 return (m); 2162 } 2163 } 2164 m = vm_page_alloc(object, pindex, allocflags & ~(VM_ALLOC_RETRY | 2165 VM_ALLOC_IGN_SBUSY)); 2166 if (m == NULL) { 2167 VM_OBJECT_UNLOCK(object); 2168 VM_WAIT; 2169 VM_OBJECT_LOCK(object); 2170 goto retrylookup; 2171 } else if (m->valid != 0) 2172 return (m); 2173 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 2174 pmap_zero_page(m); 2175 return (m); 2176 } 2177 2178 /* 2179 * Mapping function for valid bits or for dirty bits in 2180 * a page. May not block. 2181 * 2182 * Inputs are required to range within a page. 2183 */ 2184 int 2185 vm_page_bits(int base, int size) 2186 { 2187 int first_bit; 2188 int last_bit; 2189 2190 KASSERT( 2191 base + size <= PAGE_SIZE, 2192 ("vm_page_bits: illegal base/size %d/%d", base, size) 2193 ); 2194 2195 if (size == 0) /* handle degenerate case */ 2196 return (0); 2197 2198 first_bit = base >> DEV_BSHIFT; 2199 last_bit = (base + size - 1) >> DEV_BSHIFT; 2200 2201 return ((2 << last_bit) - (1 << first_bit)); 2202 } 2203 2204 /* 2205 * vm_page_set_valid: 2206 * 2207 * Sets portions of a page valid. The arguments are expected 2208 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2209 * of any partial chunks touched by the range. The invalid portion of 2210 * such chunks will be zeroed. 2211 * 2212 * (base + size) must be less then or equal to PAGE_SIZE. 2213 */ 2214 void 2215 vm_page_set_valid(vm_page_t m, int base, int size) 2216 { 2217 int endoff, frag; 2218 2219 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2220 if (size == 0) /* handle degenerate case */ 2221 return; 2222 2223 /* 2224 * If the base is not DEV_BSIZE aligned and the valid 2225 * bit is clear, we have to zero out a portion of the 2226 * first block. 2227 */ 2228 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 2229 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 2230 pmap_zero_page_area(m, frag, base - frag); 2231 2232 /* 2233 * If the ending offset is not DEV_BSIZE aligned and the 2234 * valid bit is clear, we have to zero out a portion of 2235 * the last block. 2236 */ 2237 endoff = base + size; 2238 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 2239 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 2240 pmap_zero_page_area(m, endoff, 2241 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 2242 2243 /* 2244 * Assert that no previously invalid block that is now being validated 2245 * is already dirty. 2246 */ 2247 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 2248 ("vm_page_set_valid: page %p is dirty", m)); 2249 2250 /* 2251 * Set valid bits inclusive of any overlap. 2252 */ 2253 m->valid |= vm_page_bits(base, size); 2254 } 2255 2256 /* 2257 * Clear the given bits from the specified page's dirty field. 2258 */ 2259 static __inline void 2260 vm_page_clear_dirty_mask(vm_page_t m, int pagebits) 2261 { 2262 2263 /* 2264 * If the object is locked and the page is neither VPO_BUSY nor 2265 * PG_WRITEABLE, then the page's dirty field cannot possibly be 2266 * modified by a concurrent pmap operation. 2267 */ 2268 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2269 if ((m->oflags & VPO_BUSY) == 0 && (m->flags & PG_WRITEABLE) == 0) 2270 m->dirty &= ~pagebits; 2271 else { 2272 vm_page_lock_queues(); 2273 m->dirty &= ~pagebits; 2274 vm_page_unlock_queues(); 2275 } 2276 } 2277 2278 /* 2279 * vm_page_set_validclean: 2280 * 2281 * Sets portions of a page valid and clean. The arguments are expected 2282 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2283 * of any partial chunks touched by the range. The invalid portion of 2284 * such chunks will be zero'd. 2285 * 2286 * This routine may not block. 2287 * 2288 * (base + size) must be less then or equal to PAGE_SIZE. 2289 */ 2290 void 2291 vm_page_set_validclean(vm_page_t m, int base, int size) 2292 { 2293 u_long oldvalid; 2294 int endoff, frag, pagebits; 2295 2296 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2297 if (size == 0) /* handle degenerate case */ 2298 return; 2299 2300 /* 2301 * If the base is not DEV_BSIZE aligned and the valid 2302 * bit is clear, we have to zero out a portion of the 2303 * first block. 2304 */ 2305 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 2306 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 2307 pmap_zero_page_area(m, frag, base - frag); 2308 2309 /* 2310 * If the ending offset is not DEV_BSIZE aligned and the 2311 * valid bit is clear, we have to zero out a portion of 2312 * the last block. 2313 */ 2314 endoff = base + size; 2315 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 2316 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 2317 pmap_zero_page_area(m, endoff, 2318 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 2319 2320 /* 2321 * Set valid, clear dirty bits. If validating the entire 2322 * page we can safely clear the pmap modify bit. We also 2323 * use this opportunity to clear the VPO_NOSYNC flag. If a process 2324 * takes a write fault on a MAP_NOSYNC memory area the flag will 2325 * be set again. 2326 * 2327 * We set valid bits inclusive of any overlap, but we can only 2328 * clear dirty bits for DEV_BSIZE chunks that are fully within 2329 * the range. 2330 */ 2331 oldvalid = m->valid; 2332 pagebits = vm_page_bits(base, size); 2333 m->valid |= pagebits; 2334 #if 0 /* NOT YET */ 2335 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 2336 frag = DEV_BSIZE - frag; 2337 base += frag; 2338 size -= frag; 2339 if (size < 0) 2340 size = 0; 2341 } 2342 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 2343 #endif 2344 if (base == 0 && size == PAGE_SIZE) { 2345 /* 2346 * The page can only be modified within the pmap if it is 2347 * mapped, and it can only be mapped if it was previously 2348 * fully valid. 2349 */ 2350 if (oldvalid == VM_PAGE_BITS_ALL) 2351 /* 2352 * Perform the pmap_clear_modify() first. Otherwise, 2353 * a concurrent pmap operation, such as 2354 * pmap_protect(), could clear a modification in the 2355 * pmap and set the dirty field on the page before 2356 * pmap_clear_modify() had begun and after the dirty 2357 * field was cleared here. 2358 */ 2359 pmap_clear_modify(m); 2360 m->dirty = 0; 2361 m->oflags &= ~VPO_NOSYNC; 2362 } else if (oldvalid != VM_PAGE_BITS_ALL) 2363 m->dirty &= ~pagebits; 2364 else 2365 vm_page_clear_dirty_mask(m, pagebits); 2366 } 2367 2368 void 2369 vm_page_clear_dirty(vm_page_t m, int base, int size) 2370 { 2371 2372 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 2373 } 2374 2375 /* 2376 * vm_page_set_invalid: 2377 * 2378 * Invalidates DEV_BSIZE'd chunks within a page. Both the 2379 * valid and dirty bits for the effected areas are cleared. 2380 * 2381 * May not block. 2382 */ 2383 void 2384 vm_page_set_invalid(vm_page_t m, int base, int size) 2385 { 2386 int bits; 2387 2388 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2389 KASSERT((m->oflags & VPO_BUSY) == 0, 2390 ("vm_page_set_invalid: page %p is busy", m)); 2391 bits = vm_page_bits(base, size); 2392 if (m->valid == VM_PAGE_BITS_ALL && bits != 0) 2393 pmap_remove_all(m); 2394 KASSERT(!pmap_page_is_mapped(m), 2395 ("vm_page_set_invalid: page %p is mapped", m)); 2396 m->valid &= ~bits; 2397 m->dirty &= ~bits; 2398 m->object->generation++; 2399 } 2400 2401 /* 2402 * vm_page_zero_invalid() 2403 * 2404 * The kernel assumes that the invalid portions of a page contain 2405 * garbage, but such pages can be mapped into memory by user code. 2406 * When this occurs, we must zero out the non-valid portions of the 2407 * page so user code sees what it expects. 2408 * 2409 * Pages are most often semi-valid when the end of a file is mapped 2410 * into memory and the file's size is not page aligned. 2411 */ 2412 void 2413 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 2414 { 2415 int b; 2416 int i; 2417 2418 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2419 /* 2420 * Scan the valid bits looking for invalid sections that 2421 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 2422 * valid bit may be set ) have already been zerod by 2423 * vm_page_set_validclean(). 2424 */ 2425 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 2426 if (i == (PAGE_SIZE / DEV_BSIZE) || 2427 (m->valid & (1 << i)) 2428 ) { 2429 if (i > b) { 2430 pmap_zero_page_area(m, 2431 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 2432 } 2433 b = i + 1; 2434 } 2435 } 2436 2437 /* 2438 * setvalid is TRUE when we can safely set the zero'd areas 2439 * as being valid. We can do this if there are no cache consistancy 2440 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 2441 */ 2442 if (setvalid) 2443 m->valid = VM_PAGE_BITS_ALL; 2444 } 2445 2446 /* 2447 * vm_page_is_valid: 2448 * 2449 * Is (partial) page valid? Note that the case where size == 0 2450 * will return FALSE in the degenerate case where the page is 2451 * entirely invalid, and TRUE otherwise. 2452 * 2453 * May not block. 2454 */ 2455 int 2456 vm_page_is_valid(vm_page_t m, int base, int size) 2457 { 2458 int bits = vm_page_bits(base, size); 2459 2460 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2461 if (m->valid && ((m->valid & bits) == bits)) 2462 return 1; 2463 else 2464 return 0; 2465 } 2466 2467 /* 2468 * update dirty bits from pmap/mmu. May not block. 2469 */ 2470 void 2471 vm_page_test_dirty(vm_page_t m) 2472 { 2473 2474 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2475 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 2476 vm_page_dirty(m); 2477 } 2478 2479 int so_zerocp_fullpage = 0; 2480 2481 /* 2482 * Replace the given page with a copy. The copied page assumes 2483 * the portion of the given page's "wire_count" that is not the 2484 * responsibility of this copy-on-write mechanism. 2485 * 2486 * The object containing the given page must have a non-zero 2487 * paging-in-progress count and be locked. 2488 */ 2489 void 2490 vm_page_cowfault(vm_page_t m) 2491 { 2492 vm_page_t mnew; 2493 vm_object_t object; 2494 vm_pindex_t pindex; 2495 2496 mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED); 2497 vm_page_lock_assert(m, MA_OWNED); 2498 object = m->object; 2499 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2500 KASSERT(object->paging_in_progress != 0, 2501 ("vm_page_cowfault: object %p's paging-in-progress count is zero.", 2502 object)); 2503 pindex = m->pindex; 2504 2505 retry_alloc: 2506 pmap_remove_all(m); 2507 vm_page_remove(m); 2508 mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY); 2509 if (mnew == NULL) { 2510 vm_page_insert(m, object, pindex); 2511 vm_page_unlock(m); 2512 VM_OBJECT_UNLOCK(object); 2513 VM_WAIT; 2514 VM_OBJECT_LOCK(object); 2515 if (m == vm_page_lookup(object, pindex)) { 2516 vm_page_lock(m); 2517 goto retry_alloc; 2518 } else { 2519 /* 2520 * Page disappeared during the wait. 2521 */ 2522 return; 2523 } 2524 } 2525 2526 if (m->cow == 0) { 2527 /* 2528 * check to see if we raced with an xmit complete when 2529 * waiting to allocate a page. If so, put things back 2530 * the way they were 2531 */ 2532 vm_page_unlock(m); 2533 vm_page_lock(mnew); 2534 vm_page_free(mnew); 2535 vm_page_unlock(mnew); 2536 vm_page_insert(m, object, pindex); 2537 } else { /* clear COW & copy page */ 2538 if (!so_zerocp_fullpage) 2539 pmap_copy_page(m, mnew); 2540 mnew->valid = VM_PAGE_BITS_ALL; 2541 vm_page_dirty(mnew); 2542 mnew->wire_count = m->wire_count - m->cow; 2543 m->wire_count = m->cow; 2544 vm_page_unlock(m); 2545 } 2546 } 2547 2548 void 2549 vm_page_cowclear(vm_page_t m) 2550 { 2551 2552 vm_page_lock_assert(m, MA_OWNED); 2553 if (m->cow) { 2554 m->cow--; 2555 /* 2556 * let vm_fault add back write permission lazily 2557 */ 2558 } 2559 /* 2560 * sf_buf_free() will free the page, so we needn't do it here 2561 */ 2562 } 2563 2564 int 2565 vm_page_cowsetup(vm_page_t m) 2566 { 2567 2568 vm_page_lock_assert(m, MA_OWNED); 2569 if ((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) != 0 || 2570 m->cow == USHRT_MAX - 1 || !VM_OBJECT_TRYLOCK(m->object)) 2571 return (EBUSY); 2572 m->cow++; 2573 pmap_remove_write(m); 2574 VM_OBJECT_UNLOCK(m->object); 2575 return (0); 2576 } 2577 2578 #include "opt_ddb.h" 2579 #ifdef DDB 2580 #include <sys/kernel.h> 2581 2582 #include <ddb/ddb.h> 2583 2584 DB_SHOW_COMMAND(page, vm_page_print_page_info) 2585 { 2586 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 2587 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 2588 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 2589 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 2590 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 2591 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 2592 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 2593 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 2594 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 2595 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 2596 } 2597 2598 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 2599 { 2600 2601 db_printf("PQ_FREE:"); 2602 db_printf(" %d", cnt.v_free_count); 2603 db_printf("\n"); 2604 2605 db_printf("PQ_CACHE:"); 2606 db_printf(" %d", cnt.v_cache_count); 2607 db_printf("\n"); 2608 2609 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 2610 *vm_page_queues[PQ_ACTIVE].cnt, 2611 *vm_page_queues[PQ_INACTIVE].cnt); 2612 } 2613 #endif /* DDB */ 2614