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 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 vaddr = round_page(vaddr); 297 298 for (i = 0; phys_avail[i + 1]; i += 2) { 299 phys_avail[i] = round_page(phys_avail[i]); 300 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 301 } 302 303 low_water = phys_avail[0]; 304 high_water = phys_avail[1]; 305 306 for (i = 0; phys_avail[i + 1]; i += 2) { 307 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i]; 308 309 if (size > biggestsize) { 310 biggestone = i; 311 biggestsize = size; 312 } 313 if (phys_avail[i] < low_water) 314 low_water = phys_avail[i]; 315 if (phys_avail[i + 1] > high_water) 316 high_water = phys_avail[i + 1]; 317 } 318 319 #ifdef XEN 320 low_water = 0; 321 #endif 322 323 end = phys_avail[biggestone+1]; 324 325 /* 326 * Initialize the locks. 327 */ 328 mtx_init(&vm_page_queue_mtx, "vm page queue mutex", NULL, MTX_DEF | 329 MTX_RECURSE); 330 mtx_init(&vm_page_queue_free_mtx, "vm page queue free mutex", NULL, 331 MTX_DEF); 332 333 /* Setup page locks. */ 334 for (i = 0; i < PA_LOCK_COUNT; i++) 335 mtx_init(&pa_lock[i].data, "page lock", NULL, 336 MTX_DEF | MTX_RECURSE | MTX_DUPOK); 337 338 /* 339 * Initialize the queue headers for the hold queue, the active queue, 340 * and the inactive queue. 341 */ 342 for (i = 0; i < PQ_COUNT; i++) 343 TAILQ_INIT(&vm_page_queues[i].pl); 344 vm_page_queues[PQ_INACTIVE].cnt = &cnt.v_inactive_count; 345 vm_page_queues[PQ_ACTIVE].cnt = &cnt.v_active_count; 346 vm_page_queues[PQ_HOLD].cnt = &cnt.v_active_count; 347 348 /* 349 * Allocate memory for use when boot strapping the kernel memory 350 * allocator. 351 */ 352 new_end = end - (boot_pages * UMA_SLAB_SIZE); 353 new_end = trunc_page(new_end); 354 mapped = pmap_map(&vaddr, new_end, end, 355 VM_PROT_READ | VM_PROT_WRITE); 356 bzero((void *)mapped, end - new_end); 357 uma_startup((void *)mapped, boot_pages); 358 359 #if defined(__amd64__) || defined(__i386__) || defined(__arm__) || \ 360 defined(__mips__) 361 /* 362 * Allocate a bitmap to indicate that a random physical page 363 * needs to be included in a minidump. 364 * 365 * The amd64 port needs this to indicate which direct map pages 366 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 367 * 368 * However, i386 still needs this workspace internally within the 369 * minidump code. In theory, they are not needed on i386, but are 370 * included should the sf_buf code decide to use them. 371 */ 372 last_pa = 0; 373 for (i = 0; dump_avail[i + 1] != 0; i += 2) 374 if (dump_avail[i + 1] > last_pa) 375 last_pa = dump_avail[i + 1]; 376 page_range = last_pa / PAGE_SIZE; 377 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 378 new_end -= vm_page_dump_size; 379 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, 380 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); 381 bzero((void *)vm_page_dump, vm_page_dump_size); 382 #endif 383 #ifdef __amd64__ 384 /* 385 * Request that the physical pages underlying the message buffer be 386 * included in a crash dump. Since the message buffer is accessed 387 * through the direct map, they are not automatically included. 388 */ 389 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); 390 last_pa = pa + round_page(MSGBUF_SIZE); 391 while (pa < last_pa) { 392 dump_add_page(pa); 393 pa += PAGE_SIZE; 394 } 395 #endif 396 /* 397 * Compute the number of pages of memory that will be available for 398 * use (taking into account the overhead of a page structure per 399 * page). 400 */ 401 first_page = low_water / PAGE_SIZE; 402 #ifdef VM_PHYSSEG_SPARSE 403 page_range = 0; 404 for (i = 0; phys_avail[i + 1] != 0; i += 2) 405 page_range += atop(phys_avail[i + 1] - phys_avail[i]); 406 #elif defined(VM_PHYSSEG_DENSE) 407 page_range = high_water / PAGE_SIZE - first_page; 408 #else 409 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 410 #endif 411 end = new_end; 412 413 /* 414 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 415 */ 416 vaddr += PAGE_SIZE; 417 418 /* 419 * Initialize the mem entry structures now, and put them in the free 420 * queue. 421 */ 422 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 423 mapped = pmap_map(&vaddr, new_end, end, 424 VM_PROT_READ | VM_PROT_WRITE); 425 vm_page_array = (vm_page_t) mapped; 426 #if VM_NRESERVLEVEL > 0 427 /* 428 * Allocate memory for the reservation management system's data 429 * structures. 430 */ 431 new_end = vm_reserv_startup(&vaddr, new_end, high_water); 432 #endif 433 #if defined(__amd64__) || defined(__mips__) 434 /* 435 * pmap_map on amd64 and mips can come out of the direct-map, not kvm 436 * like i386, so the pages must be tracked for a crashdump to include 437 * this data. This includes the vm_page_array and the early UMA 438 * bootstrap pages. 439 */ 440 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE) 441 dump_add_page(pa); 442 #endif 443 phys_avail[biggestone + 1] = new_end; 444 445 /* 446 * Clear all of the page structures 447 */ 448 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 449 for (i = 0; i < page_range; i++) 450 vm_page_array[i].order = VM_NFREEORDER; 451 vm_page_array_size = page_range; 452 453 /* 454 * Initialize the physical memory allocator. 455 */ 456 vm_phys_init(); 457 458 /* 459 * Add every available physical page that is not blacklisted to 460 * the free lists. 461 */ 462 cnt.v_page_count = 0; 463 cnt.v_free_count = 0; 464 list = getenv("vm.blacklist"); 465 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 466 pa = phys_avail[i]; 467 last_pa = phys_avail[i + 1]; 468 while (pa < last_pa) { 469 if (list != NULL && 470 vm_page_blacklist_lookup(list, pa)) 471 printf("Skipping page with pa 0x%jx\n", 472 (uintmax_t)pa); 473 else 474 vm_phys_add_page(pa); 475 pa += PAGE_SIZE; 476 } 477 } 478 freeenv(list); 479 #if VM_NRESERVLEVEL > 0 480 /* 481 * Initialize the reservation management system. 482 */ 483 vm_reserv_init(); 484 #endif 485 return (vaddr); 486 } 487 488 void 489 vm_page_flag_set(vm_page_t m, unsigned short bits) 490 { 491 492 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 493 /* 494 * The PG_WRITEABLE flag can only be set if the page is managed and 495 * VPO_BUSY. Currently, this flag is only set by pmap_enter(). 496 */ 497 KASSERT((bits & PG_WRITEABLE) == 0 || 498 ((m->flags & (PG_UNMANAGED | PG_FICTITIOUS)) == 0 && 499 (m->oflags & VPO_BUSY) != 0), ("PG_WRITEABLE and !VPO_BUSY")); 500 m->flags |= bits; 501 } 502 503 void 504 vm_page_flag_clear(vm_page_t m, unsigned short bits) 505 { 506 507 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 508 /* 509 * The PG_REFERENCED flag can only be cleared if the object 510 * containing the page is locked. 511 */ 512 KASSERT((bits & PG_REFERENCED) == 0 || VM_OBJECT_LOCKED(m->object), 513 ("PG_REFERENCED and !VM_OBJECT_LOCKED")); 514 m->flags &= ~bits; 515 } 516 517 void 518 vm_page_busy(vm_page_t m) 519 { 520 521 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 522 KASSERT((m->oflags & VPO_BUSY) == 0, 523 ("vm_page_busy: page already busy!!!")); 524 m->oflags |= VPO_BUSY; 525 } 526 527 /* 528 * vm_page_flash: 529 * 530 * wakeup anyone waiting for the page. 531 */ 532 void 533 vm_page_flash(vm_page_t m) 534 { 535 536 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 537 if (m->oflags & VPO_WANTED) { 538 m->oflags &= ~VPO_WANTED; 539 wakeup(m); 540 } 541 } 542 543 /* 544 * vm_page_wakeup: 545 * 546 * clear the VPO_BUSY flag and wakeup anyone waiting for the 547 * page. 548 * 549 */ 550 void 551 vm_page_wakeup(vm_page_t m) 552 { 553 554 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 555 KASSERT(m->oflags & VPO_BUSY, ("vm_page_wakeup: page not busy!!!")); 556 m->oflags &= ~VPO_BUSY; 557 vm_page_flash(m); 558 } 559 560 void 561 vm_page_io_start(vm_page_t m) 562 { 563 564 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 565 m->busy++; 566 } 567 568 void 569 vm_page_io_finish(vm_page_t m) 570 { 571 572 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 573 m->busy--; 574 if (m->busy == 0) 575 vm_page_flash(m); 576 } 577 578 /* 579 * Keep page from being freed by the page daemon 580 * much of the same effect as wiring, except much lower 581 * overhead and should be used only for *very* temporary 582 * holding ("wiring"). 583 */ 584 void 585 vm_page_hold(vm_page_t mem) 586 { 587 588 vm_page_lock_assert(mem, MA_OWNED); 589 mem->hold_count++; 590 } 591 592 void 593 vm_page_unhold(vm_page_t mem) 594 { 595 596 vm_page_lock_assert(mem, MA_OWNED); 597 --mem->hold_count; 598 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!")); 599 if (mem->hold_count == 0 && mem->queue == PQ_HOLD) 600 vm_page_free_toq(mem); 601 } 602 603 /* 604 * vm_page_unhold_pages: 605 * 606 * Unhold each of the pages that is referenced by the given array. 607 */ 608 void 609 vm_page_unhold_pages(vm_page_t *ma, int count) 610 { 611 struct mtx *mtx, *new_mtx; 612 613 mtx = NULL; 614 for (; count != 0; count--) { 615 /* 616 * Avoid releasing and reacquiring the same page lock. 617 */ 618 new_mtx = vm_page_lockptr(*ma); 619 if (mtx != new_mtx) { 620 if (mtx != NULL) 621 mtx_unlock(mtx); 622 mtx = new_mtx; 623 mtx_lock(mtx); 624 } 625 vm_page_unhold(*ma); 626 ma++; 627 } 628 if (mtx != NULL) 629 mtx_unlock(mtx); 630 } 631 632 /* 633 * vm_page_free: 634 * 635 * Free a page. 636 */ 637 void 638 vm_page_free(vm_page_t m) 639 { 640 641 m->flags &= ~PG_ZERO; 642 vm_page_free_toq(m); 643 } 644 645 /* 646 * vm_page_free_zero: 647 * 648 * Free a page to the zerod-pages queue 649 */ 650 void 651 vm_page_free_zero(vm_page_t m) 652 { 653 654 m->flags |= PG_ZERO; 655 vm_page_free_toq(m); 656 } 657 658 /* 659 * vm_page_sleep: 660 * 661 * Sleep and release the page and page queues locks. 662 * 663 * The object containing the given page must be locked. 664 */ 665 void 666 vm_page_sleep(vm_page_t m, const char *msg) 667 { 668 669 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 670 if (mtx_owned(&vm_page_queue_mtx)) 671 vm_page_unlock_queues(); 672 if (mtx_owned(vm_page_lockptr(m))) 673 vm_page_unlock(m); 674 675 /* 676 * It's possible that while we sleep, the page will get 677 * unbusied and freed. If we are holding the object 678 * lock, we will assume we hold a reference to the object 679 * such that even if m->object changes, we can re-lock 680 * it. 681 */ 682 m->oflags |= VPO_WANTED; 683 msleep(m, VM_OBJECT_MTX(m->object), PVM, msg, 0); 684 } 685 686 /* 687 * vm_page_dirty: 688 * 689 * make page all dirty 690 */ 691 void 692 vm_page_dirty(vm_page_t m) 693 { 694 695 KASSERT((m->flags & PG_CACHED) == 0, 696 ("vm_page_dirty: page in cache!")); 697 KASSERT(!VM_PAGE_IS_FREE(m), 698 ("vm_page_dirty: page is free!")); 699 KASSERT(m->valid == VM_PAGE_BITS_ALL, 700 ("vm_page_dirty: page is invalid!")); 701 m->dirty = VM_PAGE_BITS_ALL; 702 } 703 704 /* 705 * vm_page_splay: 706 * 707 * Implements Sleator and Tarjan's top-down splay algorithm. Returns 708 * the vm_page containing the given pindex. If, however, that 709 * pindex is not found in the vm_object, returns a vm_page that is 710 * adjacent to the pindex, coming before or after it. 711 */ 712 vm_page_t 713 vm_page_splay(vm_pindex_t pindex, vm_page_t root) 714 { 715 struct vm_page dummy; 716 vm_page_t lefttreemax, righttreemin, y; 717 718 if (root == NULL) 719 return (root); 720 lefttreemax = righttreemin = &dummy; 721 for (;; root = y) { 722 if (pindex < root->pindex) { 723 if ((y = root->left) == NULL) 724 break; 725 if (pindex < y->pindex) { 726 /* Rotate right. */ 727 root->left = y->right; 728 y->right = root; 729 root = y; 730 if ((y = root->left) == NULL) 731 break; 732 } 733 /* Link into the new root's right tree. */ 734 righttreemin->left = root; 735 righttreemin = root; 736 } else if (pindex > root->pindex) { 737 if ((y = root->right) == NULL) 738 break; 739 if (pindex > y->pindex) { 740 /* Rotate left. */ 741 root->right = y->left; 742 y->left = root; 743 root = y; 744 if ((y = root->right) == NULL) 745 break; 746 } 747 /* Link into the new root's left tree. */ 748 lefttreemax->right = root; 749 lefttreemax = root; 750 } else 751 break; 752 } 753 /* Assemble the new root. */ 754 lefttreemax->right = root->left; 755 righttreemin->left = root->right; 756 root->left = dummy.right; 757 root->right = dummy.left; 758 return (root); 759 } 760 761 /* 762 * vm_page_insert: [ internal use only ] 763 * 764 * Inserts the given mem entry into the object and object list. 765 * 766 * The pagetables are not updated but will presumably fault the page 767 * in if necessary, or if a kernel page the caller will at some point 768 * enter the page into the kernel's pmap. We are not allowed to block 769 * here so we *can't* do this anyway. 770 * 771 * The object and page must be locked. 772 * This routine may not block. 773 */ 774 void 775 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 776 { 777 vm_page_t root; 778 779 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 780 if (m->object != NULL) 781 panic("vm_page_insert: page already inserted"); 782 783 /* 784 * Record the object/offset pair in this page 785 */ 786 m->object = object; 787 m->pindex = pindex; 788 789 /* 790 * Now link into the object's ordered list of backed pages. 791 */ 792 root = object->root; 793 if (root == NULL) { 794 m->left = NULL; 795 m->right = NULL; 796 TAILQ_INSERT_TAIL(&object->memq, m, listq); 797 } else { 798 root = vm_page_splay(pindex, root); 799 if (pindex < root->pindex) { 800 m->left = root->left; 801 m->right = root; 802 root->left = NULL; 803 TAILQ_INSERT_BEFORE(root, m, listq); 804 } else if (pindex == root->pindex) 805 panic("vm_page_insert: offset already allocated"); 806 else { 807 m->right = root->right; 808 m->left = root; 809 root->right = NULL; 810 TAILQ_INSERT_AFTER(&object->memq, root, m, listq); 811 } 812 } 813 object->root = m; 814 object->generation++; 815 816 /* 817 * show that the object has one more resident page. 818 */ 819 object->resident_page_count++; 820 /* 821 * Hold the vnode until the last page is released. 822 */ 823 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 824 vhold((struct vnode *)object->handle); 825 826 /* 827 * Since we are inserting a new and possibly dirty page, 828 * update the object's OBJ_MIGHTBEDIRTY flag. 829 */ 830 if (m->flags & PG_WRITEABLE) 831 vm_object_set_writeable_dirty(object); 832 } 833 834 /* 835 * vm_page_remove: 836 * NOTE: used by device pager as well -wfj 837 * 838 * Removes the given mem entry from the object/offset-page 839 * table and the object page list, but do not invalidate/terminate 840 * the backing store. 841 * 842 * The object and page must be locked. 843 * The underlying pmap entry (if any) is NOT removed here. 844 * This routine may not block. 845 */ 846 void 847 vm_page_remove(vm_page_t m) 848 { 849 vm_object_t object; 850 vm_page_t root; 851 852 if ((m->flags & PG_UNMANAGED) == 0) 853 vm_page_lock_assert(m, MA_OWNED); 854 if ((object = m->object) == NULL) 855 return; 856 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 857 if (m->oflags & VPO_BUSY) { 858 m->oflags &= ~VPO_BUSY; 859 vm_page_flash(m); 860 } 861 862 /* 863 * Now remove from the object's list of backed pages. 864 */ 865 if (m != object->root) 866 vm_page_splay(m->pindex, object->root); 867 if (m->left == NULL) 868 root = m->right; 869 else { 870 root = vm_page_splay(m->pindex, m->left); 871 root->right = m->right; 872 } 873 object->root = root; 874 TAILQ_REMOVE(&object->memq, m, listq); 875 876 /* 877 * And show that the object has one fewer resident page. 878 */ 879 object->resident_page_count--; 880 /* 881 * The vnode may now be recycled. 882 */ 883 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 884 vdrop((struct vnode *)object->handle); 885 886 m->object = NULL; 887 } 888 889 /* 890 * vm_page_lookup: 891 * 892 * Returns the page associated with the object/offset 893 * pair specified; if none is found, NULL is returned. 894 * 895 * The object must be locked. 896 * This routine may not block. 897 * This is a critical path routine 898 */ 899 vm_page_t 900 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 901 { 902 vm_page_t m; 903 904 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 905 if ((m = object->root) != NULL && m->pindex != pindex) { 906 m = vm_page_splay(pindex, m); 907 if ((object->root = m)->pindex != pindex) 908 m = NULL; 909 } 910 return (m); 911 } 912 913 /* 914 * vm_page_find_least: 915 * 916 * Returns the page associated with the object with least pindex 917 * greater than or equal to the parameter pindex, or NULL. 918 * 919 * The object must be locked. 920 * The routine may not block. 921 */ 922 vm_page_t 923 vm_page_find_least(vm_object_t object, vm_pindex_t pindex) 924 { 925 vm_page_t m; 926 927 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 928 if ((m = TAILQ_FIRST(&object->memq)) != NULL) { 929 if (m->pindex < pindex) { 930 m = vm_page_splay(pindex, object->root); 931 if ((object->root = m)->pindex < pindex) 932 m = TAILQ_NEXT(m, listq); 933 } 934 } 935 return (m); 936 } 937 938 /* 939 * Returns the given page's successor (by pindex) within the object if it is 940 * resident; if none is found, NULL is returned. 941 * 942 * The object must be locked. 943 */ 944 vm_page_t 945 vm_page_next(vm_page_t m) 946 { 947 vm_page_t next; 948 949 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 950 if ((next = TAILQ_NEXT(m, listq)) != NULL && 951 next->pindex != m->pindex + 1) 952 next = NULL; 953 return (next); 954 } 955 956 /* 957 * Returns the given page's predecessor (by pindex) within the object if it is 958 * resident; if none is found, NULL is returned. 959 * 960 * The object must be locked. 961 */ 962 vm_page_t 963 vm_page_prev(vm_page_t m) 964 { 965 vm_page_t prev; 966 967 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 968 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL && 969 prev->pindex != m->pindex - 1) 970 prev = NULL; 971 return (prev); 972 } 973 974 /* 975 * vm_page_rename: 976 * 977 * Move the given memory entry from its 978 * current object to the specified target object/offset. 979 * 980 * The object must be locked. 981 * This routine may not block. 982 * 983 * Note: swap associated with the page must be invalidated by the move. We 984 * have to do this for several reasons: (1) we aren't freeing the 985 * page, (2) we are dirtying the page, (3) the VM system is probably 986 * moving the page from object A to B, and will then later move 987 * the backing store from A to B and we can't have a conflict. 988 * 989 * Note: we *always* dirty the page. It is necessary both for the 990 * fact that we moved it, and because we may be invalidating 991 * swap. If the page is on the cache, we have to deactivate it 992 * or vm_page_dirty() will panic. Dirty pages are not allowed 993 * on the cache. 994 */ 995 void 996 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 997 { 998 999 vm_page_remove(m); 1000 vm_page_insert(m, new_object, new_pindex); 1001 vm_page_dirty(m); 1002 } 1003 1004 /* 1005 * Convert all of the given object's cached pages that have a 1006 * pindex within the given range into free pages. If the value 1007 * zero is given for "end", then the range's upper bound is 1008 * infinity. If the given object is backed by a vnode and it 1009 * transitions from having one or more cached pages to none, the 1010 * vnode's hold count is reduced. 1011 */ 1012 void 1013 vm_page_cache_free(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 1014 { 1015 vm_page_t m, m_next; 1016 boolean_t empty; 1017 1018 mtx_lock(&vm_page_queue_free_mtx); 1019 if (__predict_false(object->cache == NULL)) { 1020 mtx_unlock(&vm_page_queue_free_mtx); 1021 return; 1022 } 1023 m = object->cache = vm_page_splay(start, object->cache); 1024 if (m->pindex < start) { 1025 if (m->right == NULL) 1026 m = NULL; 1027 else { 1028 m_next = vm_page_splay(start, m->right); 1029 m_next->left = m; 1030 m->right = NULL; 1031 m = object->cache = m_next; 1032 } 1033 } 1034 1035 /* 1036 * At this point, "m" is either (1) a reference to the page 1037 * with the least pindex that is greater than or equal to 1038 * "start" or (2) NULL. 1039 */ 1040 for (; m != NULL && (m->pindex < end || end == 0); m = m_next) { 1041 /* 1042 * Find "m"'s successor and remove "m" from the 1043 * object's cache. 1044 */ 1045 if (m->right == NULL) { 1046 object->cache = m->left; 1047 m_next = NULL; 1048 } else { 1049 m_next = vm_page_splay(start, m->right); 1050 m_next->left = m->left; 1051 object->cache = m_next; 1052 } 1053 /* Convert "m" to a free page. */ 1054 m->object = NULL; 1055 m->valid = 0; 1056 /* Clear PG_CACHED and set PG_FREE. */ 1057 m->flags ^= PG_CACHED | PG_FREE; 1058 KASSERT((m->flags & (PG_CACHED | PG_FREE)) == PG_FREE, 1059 ("vm_page_cache_free: page %p has inconsistent flags", m)); 1060 cnt.v_cache_count--; 1061 cnt.v_free_count++; 1062 } 1063 empty = object->cache == NULL; 1064 mtx_unlock(&vm_page_queue_free_mtx); 1065 if (object->type == OBJT_VNODE && empty) 1066 vdrop(object->handle); 1067 } 1068 1069 /* 1070 * Returns the cached page that is associated with the given 1071 * object and offset. If, however, none exists, returns NULL. 1072 * 1073 * The free page queue must be locked. 1074 */ 1075 static inline vm_page_t 1076 vm_page_cache_lookup(vm_object_t object, vm_pindex_t pindex) 1077 { 1078 vm_page_t m; 1079 1080 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1081 if ((m = object->cache) != NULL && m->pindex != pindex) { 1082 m = vm_page_splay(pindex, m); 1083 if ((object->cache = m)->pindex != pindex) 1084 m = NULL; 1085 } 1086 return (m); 1087 } 1088 1089 /* 1090 * Remove the given cached page from its containing object's 1091 * collection of cached pages. 1092 * 1093 * The free page queue must be locked. 1094 */ 1095 void 1096 vm_page_cache_remove(vm_page_t m) 1097 { 1098 vm_object_t object; 1099 vm_page_t root; 1100 1101 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1102 KASSERT((m->flags & PG_CACHED) != 0, 1103 ("vm_page_cache_remove: page %p is not cached", m)); 1104 object = m->object; 1105 if (m != object->cache) { 1106 root = vm_page_splay(m->pindex, object->cache); 1107 KASSERT(root == m, 1108 ("vm_page_cache_remove: page %p is not cached in object %p", 1109 m, object)); 1110 } 1111 if (m->left == NULL) 1112 root = m->right; 1113 else if (m->right == NULL) 1114 root = m->left; 1115 else { 1116 root = vm_page_splay(m->pindex, m->left); 1117 root->right = m->right; 1118 } 1119 object->cache = root; 1120 m->object = NULL; 1121 cnt.v_cache_count--; 1122 } 1123 1124 /* 1125 * Transfer all of the cached pages with offset greater than or 1126 * equal to 'offidxstart' from the original object's cache to the 1127 * new object's cache. However, any cached pages with offset 1128 * greater than or equal to the new object's size are kept in the 1129 * original object. Initially, the new object's cache must be 1130 * empty. Offset 'offidxstart' in the original object must 1131 * correspond to offset zero in the new object. 1132 * 1133 * The new object must be locked. 1134 */ 1135 void 1136 vm_page_cache_transfer(vm_object_t orig_object, vm_pindex_t offidxstart, 1137 vm_object_t new_object) 1138 { 1139 vm_page_t m, m_next; 1140 1141 /* 1142 * Insertion into an object's collection of cached pages 1143 * requires the object to be locked. In contrast, removal does 1144 * not. 1145 */ 1146 VM_OBJECT_LOCK_ASSERT(new_object, MA_OWNED); 1147 KASSERT(new_object->cache == NULL, 1148 ("vm_page_cache_transfer: object %p has cached pages", 1149 new_object)); 1150 mtx_lock(&vm_page_queue_free_mtx); 1151 if ((m = orig_object->cache) != NULL) { 1152 /* 1153 * Transfer all of the pages with offset greater than or 1154 * equal to 'offidxstart' from the original object's 1155 * cache to the new object's cache. 1156 */ 1157 m = vm_page_splay(offidxstart, m); 1158 if (m->pindex < offidxstart) { 1159 orig_object->cache = m; 1160 new_object->cache = m->right; 1161 m->right = NULL; 1162 } else { 1163 orig_object->cache = m->left; 1164 new_object->cache = m; 1165 m->left = NULL; 1166 } 1167 while ((m = new_object->cache) != NULL) { 1168 if ((m->pindex - offidxstart) >= new_object->size) { 1169 /* 1170 * Return all of the cached pages with 1171 * offset greater than or equal to the 1172 * new object's size to the original 1173 * object's cache. 1174 */ 1175 new_object->cache = m->left; 1176 m->left = orig_object->cache; 1177 orig_object->cache = m; 1178 break; 1179 } 1180 m_next = vm_page_splay(m->pindex, m->right); 1181 /* Update the page's object and offset. */ 1182 m->object = new_object; 1183 m->pindex -= offidxstart; 1184 if (m_next == NULL) 1185 break; 1186 m->right = NULL; 1187 m_next->left = m; 1188 new_object->cache = m_next; 1189 } 1190 KASSERT(new_object->cache == NULL || 1191 new_object->type == OBJT_SWAP, 1192 ("vm_page_cache_transfer: object %p's type is incompatible" 1193 " with cached pages", new_object)); 1194 } 1195 mtx_unlock(&vm_page_queue_free_mtx); 1196 } 1197 1198 /* 1199 * vm_page_alloc: 1200 * 1201 * Allocate and return a memory cell associated 1202 * with this VM object/offset pair. 1203 * 1204 * The caller must always specify an allocation class. 1205 * 1206 * allocation classes: 1207 * VM_ALLOC_NORMAL normal process request 1208 * VM_ALLOC_SYSTEM system *really* needs a page 1209 * VM_ALLOC_INTERRUPT interrupt time request 1210 * 1211 * optional allocation flags: 1212 * VM_ALLOC_ZERO prefer a zeroed page 1213 * VM_ALLOC_WIRED wire the allocated page 1214 * VM_ALLOC_NOOBJ page is not associated with a vm object 1215 * VM_ALLOC_NOBUSY do not set the page busy 1216 * VM_ALLOC_IFCACHED return page only if it is cached 1217 * VM_ALLOC_IFNOTCACHED return NULL, do not reactivate if the page 1218 * is cached 1219 * 1220 * This routine may not sleep. 1221 */ 1222 vm_page_t 1223 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 1224 { 1225 struct vnode *vp = NULL; 1226 vm_object_t m_object; 1227 vm_page_t m; 1228 int flags, page_req; 1229 1230 page_req = req & VM_ALLOC_CLASS_MASK; 1231 KASSERT(curthread->td_intr_nesting_level == 0 || 1232 page_req == VM_ALLOC_INTERRUPT, 1233 ("vm_page_alloc(NORMAL|SYSTEM) in interrupt context")); 1234 1235 if ((req & VM_ALLOC_NOOBJ) == 0) { 1236 KASSERT(object != NULL, 1237 ("vm_page_alloc: NULL object.")); 1238 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1239 } 1240 1241 /* 1242 * The pager is allowed to eat deeper into the free page list. 1243 */ 1244 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) { 1245 page_req = VM_ALLOC_SYSTEM; 1246 }; 1247 1248 mtx_lock(&vm_page_queue_free_mtx); 1249 if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved || 1250 (page_req == VM_ALLOC_SYSTEM && 1251 cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) || 1252 (page_req == VM_ALLOC_INTERRUPT && 1253 cnt.v_free_count + cnt.v_cache_count > 0)) { 1254 /* 1255 * Allocate from the free queue if the number of free pages 1256 * exceeds the minimum for the request class. 1257 */ 1258 if (object != NULL && 1259 (m = vm_page_cache_lookup(object, pindex)) != NULL) { 1260 if ((req & VM_ALLOC_IFNOTCACHED) != 0) { 1261 mtx_unlock(&vm_page_queue_free_mtx); 1262 return (NULL); 1263 } 1264 if (vm_phys_unfree_page(m)) 1265 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0); 1266 #if VM_NRESERVLEVEL > 0 1267 else if (!vm_reserv_reactivate_page(m)) 1268 #else 1269 else 1270 #endif 1271 panic("vm_page_alloc: cache page %p is missing" 1272 " from the free queue", m); 1273 } else if ((req & VM_ALLOC_IFCACHED) != 0) { 1274 mtx_unlock(&vm_page_queue_free_mtx); 1275 return (NULL); 1276 #if VM_NRESERVLEVEL > 0 1277 } else if (object == NULL || object->type == OBJT_DEVICE || 1278 object->type == OBJT_SG || 1279 (object->flags & OBJ_COLORED) == 0 || 1280 (m = vm_reserv_alloc_page(object, pindex)) == NULL) { 1281 #else 1282 } else { 1283 #endif 1284 m = vm_phys_alloc_pages(object != NULL ? 1285 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0); 1286 #if VM_NRESERVLEVEL > 0 1287 if (m == NULL && vm_reserv_reclaim_inactive()) { 1288 m = vm_phys_alloc_pages(object != NULL ? 1289 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 1290 0); 1291 } 1292 #endif 1293 } 1294 } else { 1295 /* 1296 * Not allocatable, give up. 1297 */ 1298 mtx_unlock(&vm_page_queue_free_mtx); 1299 atomic_add_int(&vm_pageout_deficit, 1300 MAX((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 1301 pagedaemon_wakeup(); 1302 return (NULL); 1303 } 1304 1305 /* 1306 * At this point we had better have found a good page. 1307 */ 1308 1309 KASSERT(m != NULL, ("vm_page_alloc: missing page")); 1310 KASSERT(m->queue == PQ_NONE, 1311 ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue)); 1312 KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m)); 1313 KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m)); 1314 KASSERT(m->busy == 0, ("vm_page_alloc: page %p is busy", m)); 1315 KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m)); 1316 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 1317 ("vm_page_alloc: page %p has unexpected memattr %d", m, 1318 pmap_page_get_memattr(m))); 1319 if ((m->flags & PG_CACHED) != 0) { 1320 KASSERT(m->valid != 0, 1321 ("vm_page_alloc: cached page %p is invalid", m)); 1322 if (m->object == object && m->pindex == pindex) 1323 cnt.v_reactivated++; 1324 else 1325 m->valid = 0; 1326 m_object = m->object; 1327 vm_page_cache_remove(m); 1328 if (m_object->type == OBJT_VNODE && m_object->cache == NULL) 1329 vp = m_object->handle; 1330 } else { 1331 KASSERT(VM_PAGE_IS_FREE(m), 1332 ("vm_page_alloc: page %p is not free", m)); 1333 KASSERT(m->valid == 0, 1334 ("vm_page_alloc: free page %p is valid", m)); 1335 cnt.v_free_count--; 1336 } 1337 1338 /* 1339 * Initialize structure. Only the PG_ZERO flag is inherited. 1340 */ 1341 flags = 0; 1342 if (m->flags & PG_ZERO) { 1343 vm_page_zero_count--; 1344 if (req & VM_ALLOC_ZERO) 1345 flags = PG_ZERO; 1346 } 1347 if (object == NULL || object->type == OBJT_PHYS) 1348 flags |= PG_UNMANAGED; 1349 m->flags = flags; 1350 if (req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ)) 1351 m->oflags = 0; 1352 else 1353 m->oflags = VPO_BUSY; 1354 if (req & VM_ALLOC_WIRED) { 1355 atomic_add_int(&cnt.v_wire_count, 1); 1356 m->wire_count = 1; 1357 } 1358 m->act_count = 0; 1359 mtx_unlock(&vm_page_queue_free_mtx); 1360 1361 if (object != NULL) { 1362 /* Ignore device objects; the pager sets "memattr" for them. */ 1363 if (object->memattr != VM_MEMATTR_DEFAULT && 1364 object->type != OBJT_DEVICE && object->type != OBJT_SG) 1365 pmap_page_set_memattr(m, object->memattr); 1366 vm_page_insert(m, object, pindex); 1367 } else 1368 m->pindex = pindex; 1369 1370 /* 1371 * The following call to vdrop() must come after the above call 1372 * to vm_page_insert() in case both affect the same object and 1373 * vnode. Otherwise, the affected vnode's hold count could 1374 * temporarily become zero. 1375 */ 1376 if (vp != NULL) 1377 vdrop(vp); 1378 1379 /* 1380 * Don't wakeup too often - wakeup the pageout daemon when 1381 * we would be nearly out of memory. 1382 */ 1383 if (vm_paging_needed()) 1384 pagedaemon_wakeup(); 1385 1386 return (m); 1387 } 1388 1389 /* 1390 * Initialize a page that has been freshly dequeued from a freelist. 1391 * The caller has to drop the vnode returned, if it is not NULL. 1392 * 1393 * To be called with vm_page_queue_free_mtx held. 1394 */ 1395 struct vnode * 1396 vm_page_alloc_init(vm_page_t m) 1397 { 1398 struct vnode *drop; 1399 vm_object_t m_object; 1400 1401 KASSERT(m->queue == PQ_NONE, 1402 ("vm_page_alloc_init: page %p has unexpected queue %d", 1403 m, m->queue)); 1404 KASSERT(m->wire_count == 0, 1405 ("vm_page_alloc_init: page %p is wired", m)); 1406 KASSERT(m->hold_count == 0, 1407 ("vm_page_alloc_init: page %p is held", m)); 1408 KASSERT(m->busy == 0, 1409 ("vm_page_alloc_init: page %p is busy", m)); 1410 KASSERT(m->dirty == 0, 1411 ("vm_page_alloc_init: page %p is dirty", m)); 1412 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 1413 ("vm_page_alloc_init: page %p has unexpected memattr %d", 1414 m, pmap_page_get_memattr(m))); 1415 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1416 drop = NULL; 1417 if ((m->flags & PG_CACHED) != 0) { 1418 m->valid = 0; 1419 m_object = m->object; 1420 vm_page_cache_remove(m); 1421 if (m_object->type == OBJT_VNODE && 1422 m_object->cache == NULL) 1423 drop = m_object->handle; 1424 } else { 1425 KASSERT(VM_PAGE_IS_FREE(m), 1426 ("vm_page_alloc_init: page %p is not free", m)); 1427 KASSERT(m->valid == 0, 1428 ("vm_page_alloc_init: free page %p is valid", m)); 1429 cnt.v_free_count--; 1430 } 1431 if (m->flags & PG_ZERO) 1432 vm_page_zero_count--; 1433 /* Don't clear the PG_ZERO flag; we'll need it later. */ 1434 m->flags = PG_UNMANAGED | (m->flags & PG_ZERO); 1435 m->oflags = 0; 1436 /* Unmanaged pages don't use "act_count". */ 1437 return (drop); 1438 } 1439 1440 /* 1441 * vm_page_alloc_freelist: 1442 * 1443 * Allocate a page from the specified freelist. 1444 * Only the ALLOC_CLASS values in req are honored, other request flags 1445 * are ignored. 1446 */ 1447 vm_page_t 1448 vm_page_alloc_freelist(int flind, int req) 1449 { 1450 struct vnode *drop; 1451 vm_page_t m; 1452 int page_req; 1453 1454 m = NULL; 1455 page_req = req & VM_ALLOC_CLASS_MASK; 1456 mtx_lock(&vm_page_queue_free_mtx); 1457 /* 1458 * Do not allocate reserved pages unless the req has asked for it. 1459 */ 1460 if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved || 1461 (page_req == VM_ALLOC_SYSTEM && 1462 cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) || 1463 (page_req == VM_ALLOC_INTERRUPT && 1464 cnt.v_free_count + cnt.v_cache_count > 0)) { 1465 m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0); 1466 } 1467 if (m == NULL) { 1468 mtx_unlock(&vm_page_queue_free_mtx); 1469 return (NULL); 1470 } 1471 drop = vm_page_alloc_init(m); 1472 mtx_unlock(&vm_page_queue_free_mtx); 1473 if (drop) 1474 vdrop(drop); 1475 return (m); 1476 } 1477 1478 /* 1479 * vm_wait: (also see VM_WAIT macro) 1480 * 1481 * Block until free pages are available for allocation 1482 * - Called in various places before memory allocations. 1483 */ 1484 void 1485 vm_wait(void) 1486 { 1487 1488 mtx_lock(&vm_page_queue_free_mtx); 1489 if (curproc == pageproc) { 1490 vm_pageout_pages_needed = 1; 1491 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx, 1492 PDROP | PSWP, "VMWait", 0); 1493 } else { 1494 if (!vm_pages_needed) { 1495 vm_pages_needed = 1; 1496 wakeup(&vm_pages_needed); 1497 } 1498 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM, 1499 "vmwait", 0); 1500 } 1501 } 1502 1503 /* 1504 * vm_waitpfault: (also see VM_WAITPFAULT macro) 1505 * 1506 * Block until free pages are available for allocation 1507 * - Called only in vm_fault so that processes page faulting 1508 * can be easily tracked. 1509 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 1510 * processes will be able to grab memory first. Do not change 1511 * this balance without careful testing first. 1512 */ 1513 void 1514 vm_waitpfault(void) 1515 { 1516 1517 mtx_lock(&vm_page_queue_free_mtx); 1518 if (!vm_pages_needed) { 1519 vm_pages_needed = 1; 1520 wakeup(&vm_pages_needed); 1521 } 1522 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER, 1523 "pfault", 0); 1524 } 1525 1526 /* 1527 * vm_page_requeue: 1528 * 1529 * Move the given page to the tail of its present page queue. 1530 * 1531 * The page queues must be locked. 1532 */ 1533 void 1534 vm_page_requeue(vm_page_t m) 1535 { 1536 struct vpgqueues *vpq; 1537 int queue; 1538 1539 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1540 queue = m->queue; 1541 KASSERT(queue != PQ_NONE, 1542 ("vm_page_requeue: page %p is not queued", m)); 1543 vpq = &vm_page_queues[queue]; 1544 TAILQ_REMOVE(&vpq->pl, m, pageq); 1545 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq); 1546 } 1547 1548 /* 1549 * vm_page_queue_remove: 1550 * 1551 * Remove the given page from the specified queue. 1552 * 1553 * The page and page queues must be locked. 1554 */ 1555 static __inline void 1556 vm_page_queue_remove(int queue, vm_page_t m) 1557 { 1558 struct vpgqueues *pq; 1559 1560 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1561 vm_page_lock_assert(m, MA_OWNED); 1562 pq = &vm_page_queues[queue]; 1563 TAILQ_REMOVE(&pq->pl, m, pageq); 1564 (*pq->cnt)--; 1565 } 1566 1567 /* 1568 * vm_pageq_remove: 1569 * 1570 * Remove a page from its queue. 1571 * 1572 * The given page must be locked. 1573 * This routine may not block. 1574 */ 1575 void 1576 vm_pageq_remove(vm_page_t m) 1577 { 1578 int queue; 1579 1580 vm_page_lock_assert(m, MA_OWNED); 1581 if ((queue = m->queue) != PQ_NONE) { 1582 vm_page_lock_queues(); 1583 m->queue = PQ_NONE; 1584 vm_page_queue_remove(queue, m); 1585 vm_page_unlock_queues(); 1586 } 1587 } 1588 1589 /* 1590 * vm_page_enqueue: 1591 * 1592 * Add the given page to the specified queue. 1593 * 1594 * The page queues must be locked. 1595 */ 1596 static void 1597 vm_page_enqueue(int queue, vm_page_t m) 1598 { 1599 struct vpgqueues *vpq; 1600 1601 vpq = &vm_page_queues[queue]; 1602 m->queue = queue; 1603 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq); 1604 ++*vpq->cnt; 1605 } 1606 1607 /* 1608 * vm_page_activate: 1609 * 1610 * Put the specified page on the active list (if appropriate). 1611 * Ensure that act_count is at least ACT_INIT but do not otherwise 1612 * mess with it. 1613 * 1614 * The page must be locked. 1615 * This routine may not block. 1616 */ 1617 void 1618 vm_page_activate(vm_page_t m) 1619 { 1620 int queue; 1621 1622 vm_page_lock_assert(m, MA_OWNED); 1623 if ((queue = m->queue) != PQ_ACTIVE) { 1624 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1625 if (m->act_count < ACT_INIT) 1626 m->act_count = ACT_INIT; 1627 vm_page_lock_queues(); 1628 if (queue != PQ_NONE) 1629 vm_page_queue_remove(queue, m); 1630 vm_page_enqueue(PQ_ACTIVE, m); 1631 vm_page_unlock_queues(); 1632 } else 1633 KASSERT(queue == PQ_NONE, 1634 ("vm_page_activate: wired page %p is queued", m)); 1635 } else { 1636 if (m->act_count < ACT_INIT) 1637 m->act_count = ACT_INIT; 1638 } 1639 } 1640 1641 /* 1642 * vm_page_free_wakeup: 1643 * 1644 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 1645 * routine is called when a page has been added to the cache or free 1646 * queues. 1647 * 1648 * The page queues must be locked. 1649 * This routine may not block. 1650 */ 1651 static inline void 1652 vm_page_free_wakeup(void) 1653 { 1654 1655 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1656 /* 1657 * if pageout daemon needs pages, then tell it that there are 1658 * some free. 1659 */ 1660 if (vm_pageout_pages_needed && 1661 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 1662 wakeup(&vm_pageout_pages_needed); 1663 vm_pageout_pages_needed = 0; 1664 } 1665 /* 1666 * wakeup processes that are waiting on memory if we hit a 1667 * high water mark. And wakeup scheduler process if we have 1668 * lots of memory. this process will swapin processes. 1669 */ 1670 if (vm_pages_needed && !vm_page_count_min()) { 1671 vm_pages_needed = 0; 1672 wakeup(&cnt.v_free_count); 1673 } 1674 } 1675 1676 /* 1677 * vm_page_free_toq: 1678 * 1679 * Returns the given page to the free list, 1680 * disassociating it with any VM object. 1681 * 1682 * Object and page must be locked prior to entry. 1683 * This routine may not block. 1684 */ 1685 1686 void 1687 vm_page_free_toq(vm_page_t m) 1688 { 1689 1690 if ((m->flags & PG_UNMANAGED) == 0) { 1691 vm_page_lock_assert(m, MA_OWNED); 1692 KASSERT(!pmap_page_is_mapped(m), 1693 ("vm_page_free_toq: freeing mapped page %p", m)); 1694 } 1695 PCPU_INC(cnt.v_tfree); 1696 1697 if (VM_PAGE_IS_FREE(m)) 1698 panic("vm_page_free: freeing free page %p", m); 1699 else if (m->busy != 0) 1700 panic("vm_page_free: freeing busy page %p", m); 1701 1702 /* 1703 * unqueue, then remove page. Note that we cannot destroy 1704 * the page here because we do not want to call the pager's 1705 * callback routine until after we've put the page on the 1706 * appropriate free queue. 1707 */ 1708 if ((m->flags & PG_UNMANAGED) == 0) 1709 vm_pageq_remove(m); 1710 vm_page_remove(m); 1711 1712 /* 1713 * If fictitious remove object association and 1714 * return, otherwise delay object association removal. 1715 */ 1716 if ((m->flags & PG_FICTITIOUS) != 0) { 1717 return; 1718 } 1719 1720 m->valid = 0; 1721 vm_page_undirty(m); 1722 1723 if (m->wire_count != 0) 1724 panic("vm_page_free: freeing wired page %p", m); 1725 if (m->hold_count != 0) { 1726 m->flags &= ~PG_ZERO; 1727 vm_page_lock_queues(); 1728 vm_page_enqueue(PQ_HOLD, m); 1729 vm_page_unlock_queues(); 1730 } else { 1731 /* 1732 * Restore the default memory attribute to the page. 1733 */ 1734 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 1735 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 1736 1737 /* 1738 * Insert the page into the physical memory allocator's 1739 * cache/free page queues. 1740 */ 1741 mtx_lock(&vm_page_queue_free_mtx); 1742 m->flags |= PG_FREE; 1743 cnt.v_free_count++; 1744 #if VM_NRESERVLEVEL > 0 1745 if (!vm_reserv_free_page(m)) 1746 #else 1747 if (TRUE) 1748 #endif 1749 vm_phys_free_pages(m, 0); 1750 if ((m->flags & PG_ZERO) != 0) 1751 ++vm_page_zero_count; 1752 else 1753 vm_page_zero_idle_wakeup(); 1754 vm_page_free_wakeup(); 1755 mtx_unlock(&vm_page_queue_free_mtx); 1756 } 1757 } 1758 1759 /* 1760 * vm_page_wire: 1761 * 1762 * Mark this page as wired down by yet 1763 * another map, removing it from paging queues 1764 * as necessary. 1765 * 1766 * If the page is fictitious, then its wire count must remain one. 1767 * 1768 * The page must be locked. 1769 * This routine may not block. 1770 */ 1771 void 1772 vm_page_wire(vm_page_t m) 1773 { 1774 1775 /* 1776 * Only bump the wire statistics if the page is not already wired, 1777 * and only unqueue the page if it is on some queue (if it is unmanaged 1778 * it is already off the queues). 1779 */ 1780 vm_page_lock_assert(m, MA_OWNED); 1781 if ((m->flags & PG_FICTITIOUS) != 0) { 1782 KASSERT(m->wire_count == 1, 1783 ("vm_page_wire: fictitious page %p's wire count isn't one", 1784 m)); 1785 return; 1786 } 1787 if (m->wire_count == 0) { 1788 if ((m->flags & PG_UNMANAGED) == 0) 1789 vm_pageq_remove(m); 1790 atomic_add_int(&cnt.v_wire_count, 1); 1791 } 1792 m->wire_count++; 1793 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 1794 } 1795 1796 /* 1797 * vm_page_unwire: 1798 * 1799 * Release one wiring of the specified page, potentially enabling it to be 1800 * paged again. If paging is enabled, then the value of the parameter 1801 * "activate" determines to which queue the page is added. If "activate" is 1802 * non-zero, then the page is added to the active queue. Otherwise, it is 1803 * added to the inactive queue. 1804 * 1805 * However, unless the page belongs to an object, it is not enqueued because 1806 * it cannot be paged out. 1807 * 1808 * If a page is fictitious, then its wire count must alway be one. 1809 * 1810 * A managed page must be locked. 1811 */ 1812 void 1813 vm_page_unwire(vm_page_t m, int activate) 1814 { 1815 1816 if ((m->flags & PG_UNMANAGED) == 0) 1817 vm_page_lock_assert(m, MA_OWNED); 1818 if ((m->flags & PG_FICTITIOUS) != 0) { 1819 KASSERT(m->wire_count == 1, 1820 ("vm_page_unwire: fictitious page %p's wire count isn't one", m)); 1821 return; 1822 } 1823 if (m->wire_count > 0) { 1824 m->wire_count--; 1825 if (m->wire_count == 0) { 1826 atomic_subtract_int(&cnt.v_wire_count, 1); 1827 if ((m->flags & PG_UNMANAGED) != 0 || 1828 m->object == NULL) 1829 return; 1830 vm_page_lock_queues(); 1831 if (activate) 1832 vm_page_enqueue(PQ_ACTIVE, m); 1833 else { 1834 vm_page_flag_clear(m, PG_WINATCFLS); 1835 vm_page_enqueue(PQ_INACTIVE, m); 1836 } 1837 vm_page_unlock_queues(); 1838 } 1839 } else 1840 panic("vm_page_unwire: page %p's wire count is zero", m); 1841 } 1842 1843 /* 1844 * Move the specified page to the inactive queue. 1845 * 1846 * Many pages placed on the inactive queue should actually go 1847 * into the cache, but it is difficult to figure out which. What 1848 * we do instead, if the inactive target is well met, is to put 1849 * clean pages at the head of the inactive queue instead of the tail. 1850 * This will cause them to be moved to the cache more quickly and 1851 * if not actively re-referenced, reclaimed more quickly. If we just 1852 * stick these pages at the end of the inactive queue, heavy filesystem 1853 * meta-data accesses can cause an unnecessary paging load on memory bound 1854 * processes. This optimization causes one-time-use metadata to be 1855 * reused more quickly. 1856 * 1857 * Normally athead is 0 resulting in LRU operation. athead is set 1858 * to 1 if we want this page to be 'as if it were placed in the cache', 1859 * except without unmapping it from the process address space. 1860 * 1861 * This routine may not block. 1862 */ 1863 static inline void 1864 _vm_page_deactivate(vm_page_t m, int athead) 1865 { 1866 int queue; 1867 1868 vm_page_lock_assert(m, MA_OWNED); 1869 1870 /* 1871 * Ignore if already inactive. 1872 */ 1873 if ((queue = m->queue) == PQ_INACTIVE) 1874 return; 1875 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1876 vm_page_lock_queues(); 1877 vm_page_flag_clear(m, PG_WINATCFLS); 1878 if (queue != PQ_NONE) 1879 vm_page_queue_remove(queue, m); 1880 if (athead) 1881 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, 1882 pageq); 1883 else 1884 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, 1885 pageq); 1886 m->queue = PQ_INACTIVE; 1887 cnt.v_inactive_count++; 1888 vm_page_unlock_queues(); 1889 } 1890 } 1891 1892 /* 1893 * Move the specified page to the inactive queue. 1894 * 1895 * The page must be locked. 1896 */ 1897 void 1898 vm_page_deactivate(vm_page_t m) 1899 { 1900 1901 _vm_page_deactivate(m, 0); 1902 } 1903 1904 /* 1905 * vm_page_try_to_cache: 1906 * 1907 * Returns 0 on failure, 1 on success 1908 */ 1909 int 1910 vm_page_try_to_cache(vm_page_t m) 1911 { 1912 1913 vm_page_lock_assert(m, MA_OWNED); 1914 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1915 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1916 (m->oflags & VPO_BUSY) || (m->flags & PG_UNMANAGED)) 1917 return (0); 1918 pmap_remove_all(m); 1919 if (m->dirty) 1920 return (0); 1921 vm_page_cache(m); 1922 return (1); 1923 } 1924 1925 /* 1926 * vm_page_try_to_free() 1927 * 1928 * Attempt to free the page. If we cannot free it, we do nothing. 1929 * 1 is returned on success, 0 on failure. 1930 */ 1931 int 1932 vm_page_try_to_free(vm_page_t m) 1933 { 1934 1935 vm_page_lock_assert(m, MA_OWNED); 1936 if (m->object != NULL) 1937 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1938 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1939 (m->oflags & VPO_BUSY) || (m->flags & PG_UNMANAGED)) 1940 return (0); 1941 pmap_remove_all(m); 1942 if (m->dirty) 1943 return (0); 1944 vm_page_free(m); 1945 return (1); 1946 } 1947 1948 /* 1949 * vm_page_cache 1950 * 1951 * Put the specified page onto the page cache queue (if appropriate). 1952 * 1953 * This routine may not block. 1954 */ 1955 void 1956 vm_page_cache(vm_page_t m) 1957 { 1958 vm_object_t object; 1959 vm_page_t root; 1960 1961 vm_page_lock_assert(m, MA_OWNED); 1962 object = m->object; 1963 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1964 if ((m->flags & PG_UNMANAGED) || (m->oflags & VPO_BUSY) || m->busy || 1965 m->hold_count || m->wire_count) 1966 panic("vm_page_cache: attempting to cache busy page"); 1967 pmap_remove_all(m); 1968 if (m->dirty != 0) 1969 panic("vm_page_cache: page %p is dirty", m); 1970 if (m->valid == 0 || object->type == OBJT_DEFAULT || 1971 (object->type == OBJT_SWAP && 1972 !vm_pager_has_page(object, m->pindex, NULL, NULL))) { 1973 /* 1974 * Hypothesis: A cache-elgible page belonging to a 1975 * default object or swap object but without a backing 1976 * store must be zero filled. 1977 */ 1978 vm_page_free(m); 1979 return; 1980 } 1981 KASSERT((m->flags & PG_CACHED) == 0, 1982 ("vm_page_cache: page %p is already cached", m)); 1983 PCPU_INC(cnt.v_tcached); 1984 1985 /* 1986 * Remove the page from the paging queues. 1987 */ 1988 vm_pageq_remove(m); 1989 1990 /* 1991 * Remove the page from the object's collection of resident 1992 * pages. 1993 */ 1994 if (m != object->root) 1995 vm_page_splay(m->pindex, object->root); 1996 if (m->left == NULL) 1997 root = m->right; 1998 else { 1999 root = vm_page_splay(m->pindex, m->left); 2000 root->right = m->right; 2001 } 2002 object->root = root; 2003 TAILQ_REMOVE(&object->memq, m, listq); 2004 object->resident_page_count--; 2005 2006 /* 2007 * Restore the default memory attribute to the page. 2008 */ 2009 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 2010 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 2011 2012 /* 2013 * Insert the page into the object's collection of cached pages 2014 * and the physical memory allocator's cache/free page queues. 2015 */ 2016 m->flags &= ~PG_ZERO; 2017 mtx_lock(&vm_page_queue_free_mtx); 2018 m->flags |= PG_CACHED; 2019 cnt.v_cache_count++; 2020 root = object->cache; 2021 if (root == NULL) { 2022 m->left = NULL; 2023 m->right = NULL; 2024 } else { 2025 root = vm_page_splay(m->pindex, root); 2026 if (m->pindex < root->pindex) { 2027 m->left = root->left; 2028 m->right = root; 2029 root->left = NULL; 2030 } else if (__predict_false(m->pindex == root->pindex)) 2031 panic("vm_page_cache: offset already cached"); 2032 else { 2033 m->right = root->right; 2034 m->left = root; 2035 root->right = NULL; 2036 } 2037 } 2038 object->cache = m; 2039 #if VM_NRESERVLEVEL > 0 2040 if (!vm_reserv_free_page(m)) { 2041 #else 2042 if (TRUE) { 2043 #endif 2044 vm_phys_set_pool(VM_FREEPOOL_CACHE, m, 0); 2045 vm_phys_free_pages(m, 0); 2046 } 2047 vm_page_free_wakeup(); 2048 mtx_unlock(&vm_page_queue_free_mtx); 2049 2050 /* 2051 * Increment the vnode's hold count if this is the object's only 2052 * cached page. Decrement the vnode's hold count if this was 2053 * the object's only resident page. 2054 */ 2055 if (object->type == OBJT_VNODE) { 2056 if (root == NULL && object->resident_page_count != 0) 2057 vhold(object->handle); 2058 else if (root != NULL && object->resident_page_count == 0) 2059 vdrop(object->handle); 2060 } 2061 } 2062 2063 /* 2064 * vm_page_dontneed 2065 * 2066 * Cache, deactivate, or do nothing as appropriate. This routine 2067 * is typically used by madvise() MADV_DONTNEED. 2068 * 2069 * Generally speaking we want to move the page into the cache so 2070 * it gets reused quickly. However, this can result in a silly syndrome 2071 * due to the page recycling too quickly. Small objects will not be 2072 * fully cached. On the otherhand, if we move the page to the inactive 2073 * queue we wind up with a problem whereby very large objects 2074 * unnecessarily blow away our inactive and cache queues. 2075 * 2076 * The solution is to move the pages based on a fixed weighting. We 2077 * either leave them alone, deactivate them, or move them to the cache, 2078 * where moving them to the cache has the highest weighting. 2079 * By forcing some pages into other queues we eventually force the 2080 * system to balance the queues, potentially recovering other unrelated 2081 * space from active. The idea is to not force this to happen too 2082 * often. 2083 */ 2084 void 2085 vm_page_dontneed(vm_page_t m) 2086 { 2087 int dnw; 2088 int head; 2089 2090 vm_page_lock_assert(m, MA_OWNED); 2091 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2092 dnw = PCPU_GET(dnweight); 2093 PCPU_INC(dnweight); 2094 2095 /* 2096 * Occasionally leave the page alone. 2097 */ 2098 if ((dnw & 0x01F0) == 0 || m->queue == PQ_INACTIVE) { 2099 if (m->act_count >= ACT_INIT) 2100 --m->act_count; 2101 return; 2102 } 2103 2104 /* 2105 * Clear any references to the page. Otherwise, the page daemon will 2106 * immediately reactivate the page. 2107 * 2108 * Perform the pmap_clear_reference() first. Otherwise, a concurrent 2109 * pmap operation, such as pmap_remove(), could clear a reference in 2110 * the pmap and set PG_REFERENCED on the page before the 2111 * pmap_clear_reference() had completed. Consequently, the page would 2112 * appear referenced based upon an old reference that occurred before 2113 * this function ran. 2114 */ 2115 pmap_clear_reference(m); 2116 vm_page_lock_queues(); 2117 vm_page_flag_clear(m, PG_REFERENCED); 2118 vm_page_unlock_queues(); 2119 2120 if (m->dirty == 0 && pmap_is_modified(m)) 2121 vm_page_dirty(m); 2122 2123 if (m->dirty || (dnw & 0x0070) == 0) { 2124 /* 2125 * Deactivate the page 3 times out of 32. 2126 */ 2127 head = 0; 2128 } else { 2129 /* 2130 * Cache the page 28 times out of every 32. Note that 2131 * the page is deactivated instead of cached, but placed 2132 * at the head of the queue instead of the tail. 2133 */ 2134 head = 1; 2135 } 2136 _vm_page_deactivate(m, head); 2137 } 2138 2139 /* 2140 * Grab a page, waiting until we are waken up due to the page 2141 * changing state. We keep on waiting, if the page continues 2142 * to be in the object. If the page doesn't exist, first allocate it 2143 * and then conditionally zero it. 2144 * 2145 * The caller must always specify the VM_ALLOC_RETRY flag. This is intended 2146 * to facilitate its eventual removal. 2147 * 2148 * This routine may block. 2149 */ 2150 vm_page_t 2151 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 2152 { 2153 vm_page_t m; 2154 2155 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2156 KASSERT((allocflags & VM_ALLOC_RETRY) != 0, 2157 ("vm_page_grab: VM_ALLOC_RETRY is required")); 2158 retrylookup: 2159 if ((m = vm_page_lookup(object, pindex)) != NULL) { 2160 if ((m->oflags & VPO_BUSY) != 0 || 2161 ((allocflags & VM_ALLOC_IGN_SBUSY) == 0 && m->busy != 0)) { 2162 /* 2163 * Reference the page before unlocking and 2164 * sleeping so that the page daemon is less 2165 * likely to reclaim it. 2166 */ 2167 vm_page_lock_queues(); 2168 vm_page_flag_set(m, PG_REFERENCED); 2169 vm_page_sleep(m, "pgrbwt"); 2170 goto retrylookup; 2171 } else { 2172 if ((allocflags & VM_ALLOC_WIRED) != 0) { 2173 vm_page_lock(m); 2174 vm_page_wire(m); 2175 vm_page_unlock(m); 2176 } 2177 if ((allocflags & VM_ALLOC_NOBUSY) == 0) 2178 vm_page_busy(m); 2179 return (m); 2180 } 2181 } 2182 m = vm_page_alloc(object, pindex, allocflags & ~(VM_ALLOC_RETRY | 2183 VM_ALLOC_IGN_SBUSY)); 2184 if (m == NULL) { 2185 VM_OBJECT_UNLOCK(object); 2186 VM_WAIT; 2187 VM_OBJECT_LOCK(object); 2188 goto retrylookup; 2189 } else if (m->valid != 0) 2190 return (m); 2191 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 2192 pmap_zero_page(m); 2193 return (m); 2194 } 2195 2196 /* 2197 * Mapping function for valid bits or for dirty bits in 2198 * a page. May not block. 2199 * 2200 * Inputs are required to range within a page. 2201 */ 2202 int 2203 vm_page_bits(int base, int size) 2204 { 2205 int first_bit; 2206 int last_bit; 2207 2208 KASSERT( 2209 base + size <= PAGE_SIZE, 2210 ("vm_page_bits: illegal base/size %d/%d", base, size) 2211 ); 2212 2213 if (size == 0) /* handle degenerate case */ 2214 return (0); 2215 2216 first_bit = base >> DEV_BSHIFT; 2217 last_bit = (base + size - 1) >> DEV_BSHIFT; 2218 2219 return ((2 << last_bit) - (1 << first_bit)); 2220 } 2221 2222 /* 2223 * vm_page_set_valid: 2224 * 2225 * Sets portions of a page valid. The arguments are expected 2226 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2227 * of any partial chunks touched by the range. The invalid portion of 2228 * such chunks will be zeroed. 2229 * 2230 * (base + size) must be less then or equal to PAGE_SIZE. 2231 */ 2232 void 2233 vm_page_set_valid(vm_page_t m, int base, int size) 2234 { 2235 int endoff, frag; 2236 2237 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2238 if (size == 0) /* handle degenerate case */ 2239 return; 2240 2241 /* 2242 * If the base is not DEV_BSIZE aligned and the valid 2243 * bit is clear, we have to zero out a portion of the 2244 * first block. 2245 */ 2246 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 2247 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 2248 pmap_zero_page_area(m, frag, base - frag); 2249 2250 /* 2251 * If the ending offset is not DEV_BSIZE aligned and the 2252 * valid bit is clear, we have to zero out a portion of 2253 * the last block. 2254 */ 2255 endoff = base + size; 2256 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 2257 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 2258 pmap_zero_page_area(m, endoff, 2259 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 2260 2261 /* 2262 * Assert that no previously invalid block that is now being validated 2263 * is already dirty. 2264 */ 2265 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 2266 ("vm_page_set_valid: page %p is dirty", m)); 2267 2268 /* 2269 * Set valid bits inclusive of any overlap. 2270 */ 2271 m->valid |= vm_page_bits(base, size); 2272 } 2273 2274 /* 2275 * Clear the given bits from the specified page's dirty field. 2276 */ 2277 static __inline void 2278 vm_page_clear_dirty_mask(vm_page_t m, int pagebits) 2279 { 2280 2281 /* 2282 * If the object is locked and the page is neither VPO_BUSY nor 2283 * PG_WRITEABLE, then the page's dirty field cannot possibly be 2284 * modified by a concurrent pmap operation. 2285 */ 2286 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2287 if ((m->oflags & VPO_BUSY) == 0 && (m->flags & PG_WRITEABLE) == 0) 2288 m->dirty &= ~pagebits; 2289 else { 2290 vm_page_lock_queues(); 2291 m->dirty &= ~pagebits; 2292 vm_page_unlock_queues(); 2293 } 2294 } 2295 2296 /* 2297 * vm_page_set_validclean: 2298 * 2299 * Sets portions of a page valid and clean. The arguments are expected 2300 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 2301 * of any partial chunks touched by the range. The invalid portion of 2302 * such chunks will be zero'd. 2303 * 2304 * This routine may not block. 2305 * 2306 * (base + size) must be less then or equal to PAGE_SIZE. 2307 */ 2308 void 2309 vm_page_set_validclean(vm_page_t m, int base, int size) 2310 { 2311 u_long oldvalid; 2312 int endoff, frag, pagebits; 2313 2314 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2315 if (size == 0) /* handle degenerate case */ 2316 return; 2317 2318 /* 2319 * If the base is not DEV_BSIZE aligned and the valid 2320 * bit is clear, we have to zero out a portion of the 2321 * first block. 2322 */ 2323 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 2324 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 2325 pmap_zero_page_area(m, frag, base - frag); 2326 2327 /* 2328 * If the ending offset is not DEV_BSIZE aligned and the 2329 * valid bit is clear, we have to zero out a portion of 2330 * the last block. 2331 */ 2332 endoff = base + size; 2333 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 2334 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 2335 pmap_zero_page_area(m, endoff, 2336 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 2337 2338 /* 2339 * Set valid, clear dirty bits. If validating the entire 2340 * page we can safely clear the pmap modify bit. We also 2341 * use this opportunity to clear the VPO_NOSYNC flag. If a process 2342 * takes a write fault on a MAP_NOSYNC memory area the flag will 2343 * be set again. 2344 * 2345 * We set valid bits inclusive of any overlap, but we can only 2346 * clear dirty bits for DEV_BSIZE chunks that are fully within 2347 * the range. 2348 */ 2349 oldvalid = m->valid; 2350 pagebits = vm_page_bits(base, size); 2351 m->valid |= pagebits; 2352 #if 0 /* NOT YET */ 2353 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 2354 frag = DEV_BSIZE - frag; 2355 base += frag; 2356 size -= frag; 2357 if (size < 0) 2358 size = 0; 2359 } 2360 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 2361 #endif 2362 if (base == 0 && size == PAGE_SIZE) { 2363 /* 2364 * The page can only be modified within the pmap if it is 2365 * mapped, and it can only be mapped if it was previously 2366 * fully valid. 2367 */ 2368 if (oldvalid == VM_PAGE_BITS_ALL) 2369 /* 2370 * Perform the pmap_clear_modify() first. Otherwise, 2371 * a concurrent pmap operation, such as 2372 * pmap_protect(), could clear a modification in the 2373 * pmap and set the dirty field on the page before 2374 * pmap_clear_modify() had begun and after the dirty 2375 * field was cleared here. 2376 */ 2377 pmap_clear_modify(m); 2378 m->dirty = 0; 2379 m->oflags &= ~VPO_NOSYNC; 2380 } else if (oldvalid != VM_PAGE_BITS_ALL) 2381 m->dirty &= ~pagebits; 2382 else 2383 vm_page_clear_dirty_mask(m, pagebits); 2384 } 2385 2386 void 2387 vm_page_clear_dirty(vm_page_t m, int base, int size) 2388 { 2389 2390 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 2391 } 2392 2393 /* 2394 * vm_page_set_invalid: 2395 * 2396 * Invalidates DEV_BSIZE'd chunks within a page. Both the 2397 * valid and dirty bits for the effected areas are cleared. 2398 * 2399 * May not block. 2400 */ 2401 void 2402 vm_page_set_invalid(vm_page_t m, int base, int size) 2403 { 2404 int bits; 2405 2406 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2407 KASSERT((m->oflags & VPO_BUSY) == 0, 2408 ("vm_page_set_invalid: page %p is busy", m)); 2409 bits = vm_page_bits(base, size); 2410 if (m->valid == VM_PAGE_BITS_ALL && bits != 0) 2411 pmap_remove_all(m); 2412 KASSERT(!pmap_page_is_mapped(m), 2413 ("vm_page_set_invalid: page %p is mapped", m)); 2414 m->valid &= ~bits; 2415 m->dirty &= ~bits; 2416 } 2417 2418 /* 2419 * vm_page_zero_invalid() 2420 * 2421 * The kernel assumes that the invalid portions of a page contain 2422 * garbage, but such pages can be mapped into memory by user code. 2423 * When this occurs, we must zero out the non-valid portions of the 2424 * page so user code sees what it expects. 2425 * 2426 * Pages are most often semi-valid when the end of a file is mapped 2427 * into memory and the file's size is not page aligned. 2428 */ 2429 void 2430 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 2431 { 2432 int b; 2433 int i; 2434 2435 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2436 /* 2437 * Scan the valid bits looking for invalid sections that 2438 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 2439 * valid bit may be set ) have already been zerod by 2440 * vm_page_set_validclean(). 2441 */ 2442 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 2443 if (i == (PAGE_SIZE / DEV_BSIZE) || 2444 (m->valid & (1 << i)) 2445 ) { 2446 if (i > b) { 2447 pmap_zero_page_area(m, 2448 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 2449 } 2450 b = i + 1; 2451 } 2452 } 2453 2454 /* 2455 * setvalid is TRUE when we can safely set the zero'd areas 2456 * as being valid. We can do this if there are no cache consistancy 2457 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 2458 */ 2459 if (setvalid) 2460 m->valid = VM_PAGE_BITS_ALL; 2461 } 2462 2463 /* 2464 * vm_page_is_valid: 2465 * 2466 * Is (partial) page valid? Note that the case where size == 0 2467 * will return FALSE in the degenerate case where the page is 2468 * entirely invalid, and TRUE otherwise. 2469 * 2470 * May not block. 2471 */ 2472 int 2473 vm_page_is_valid(vm_page_t m, int base, int size) 2474 { 2475 int bits = vm_page_bits(base, size); 2476 2477 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2478 if (m->valid && ((m->valid & bits) == bits)) 2479 return 1; 2480 else 2481 return 0; 2482 } 2483 2484 /* 2485 * update dirty bits from pmap/mmu. May not block. 2486 */ 2487 void 2488 vm_page_test_dirty(vm_page_t m) 2489 { 2490 2491 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2492 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 2493 vm_page_dirty(m); 2494 } 2495 2496 int so_zerocp_fullpage = 0; 2497 2498 /* 2499 * Replace the given page with a copy. The copied page assumes 2500 * the portion of the given page's "wire_count" that is not the 2501 * responsibility of this copy-on-write mechanism. 2502 * 2503 * The object containing the given page must have a non-zero 2504 * paging-in-progress count and be locked. 2505 */ 2506 void 2507 vm_page_cowfault(vm_page_t m) 2508 { 2509 vm_page_t mnew; 2510 vm_object_t object; 2511 vm_pindex_t pindex; 2512 2513 mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED); 2514 vm_page_lock_assert(m, MA_OWNED); 2515 object = m->object; 2516 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2517 KASSERT(object->paging_in_progress != 0, 2518 ("vm_page_cowfault: object %p's paging-in-progress count is zero.", 2519 object)); 2520 pindex = m->pindex; 2521 2522 retry_alloc: 2523 pmap_remove_all(m); 2524 vm_page_remove(m); 2525 mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY); 2526 if (mnew == NULL) { 2527 vm_page_insert(m, object, pindex); 2528 vm_page_unlock(m); 2529 VM_OBJECT_UNLOCK(object); 2530 VM_WAIT; 2531 VM_OBJECT_LOCK(object); 2532 if (m == vm_page_lookup(object, pindex)) { 2533 vm_page_lock(m); 2534 goto retry_alloc; 2535 } else { 2536 /* 2537 * Page disappeared during the wait. 2538 */ 2539 return; 2540 } 2541 } 2542 2543 if (m->cow == 0) { 2544 /* 2545 * check to see if we raced with an xmit complete when 2546 * waiting to allocate a page. If so, put things back 2547 * the way they were 2548 */ 2549 vm_page_unlock(m); 2550 vm_page_lock(mnew); 2551 vm_page_free(mnew); 2552 vm_page_unlock(mnew); 2553 vm_page_insert(m, object, pindex); 2554 } else { /* clear COW & copy page */ 2555 if (!so_zerocp_fullpage) 2556 pmap_copy_page(m, mnew); 2557 mnew->valid = VM_PAGE_BITS_ALL; 2558 vm_page_dirty(mnew); 2559 mnew->wire_count = m->wire_count - m->cow; 2560 m->wire_count = m->cow; 2561 vm_page_unlock(m); 2562 } 2563 } 2564 2565 void 2566 vm_page_cowclear(vm_page_t m) 2567 { 2568 2569 vm_page_lock_assert(m, MA_OWNED); 2570 if (m->cow) { 2571 m->cow--; 2572 /* 2573 * let vm_fault add back write permission lazily 2574 */ 2575 } 2576 /* 2577 * sf_buf_free() will free the page, so we needn't do it here 2578 */ 2579 } 2580 2581 int 2582 vm_page_cowsetup(vm_page_t m) 2583 { 2584 2585 vm_page_lock_assert(m, MA_OWNED); 2586 if ((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) != 0 || 2587 m->cow == USHRT_MAX - 1 || !VM_OBJECT_TRYLOCK(m->object)) 2588 return (EBUSY); 2589 m->cow++; 2590 pmap_remove_write(m); 2591 VM_OBJECT_UNLOCK(m->object); 2592 return (0); 2593 } 2594 2595 #include "opt_ddb.h" 2596 #ifdef DDB 2597 #include <sys/kernel.h> 2598 2599 #include <ddb/ddb.h> 2600 2601 DB_SHOW_COMMAND(page, vm_page_print_page_info) 2602 { 2603 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 2604 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 2605 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 2606 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 2607 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 2608 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 2609 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 2610 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 2611 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 2612 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 2613 } 2614 2615 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 2616 { 2617 2618 db_printf("PQ_FREE:"); 2619 db_printf(" %d", cnt.v_free_count); 2620 db_printf("\n"); 2621 2622 db_printf("PQ_CACHE:"); 2623 db_printf(" %d", cnt.v_cache_count); 2624 db_printf("\n"); 2625 2626 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 2627 *vm_page_queues[PQ_ACTIVE].cnt, 2628 *vm_page_queues[PQ_INACTIVE].cnt); 2629 } 2630 #endif /* DDB */ 2631