vm_pageout.c revision 170658
1/*- 2 * Copyright (c) 1991 Regents of the University of California. 3 * All rights reserved. 4 * Copyright (c) 1994 John S. Dyson 5 * All rights reserved. 6 * Copyright (c) 1994 David Greenman 7 * All rights reserved. 8 * Copyright (c) 2005 Yahoo! Technologies Norway AS 9 * All rights reserved. 10 * 11 * This code is derived from software contributed to Berkeley by 12 * The Mach Operating System project at Carnegie-Mellon University. 13 * 14 * Redistribution and use in source and binary forms, with or without 15 * modification, are permitted provided that the following conditions 16 * are met: 17 * 1. Redistributions of source code must retain the above copyright 18 * notice, this list of conditions and the following disclaimer. 19 * 2. Redistributions in binary form must reproduce the above copyright 20 * notice, this list of conditions and the following disclaimer in the 21 * documentation and/or other materials provided with the distribution. 22 * 3. All advertising materials mentioning features or use of this software 23 * must display the following acknowledgement: 24 * This product includes software developed by the University of 25 * California, Berkeley and its contributors. 26 * 4. Neither the name of the University nor the names of its contributors 27 * may be used to endorse or promote products derived from this software 28 * without specific prior written permission. 29 * 30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 40 * SUCH DAMAGE. 41 * 42 * from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91 43 * 44 * 45 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 46 * All rights reserved. 47 * 48 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 49 * 50 * Permission to use, copy, modify and distribute this software and 51 * its documentation is hereby granted, provided that both the copyright 52 * notice and this permission notice appear in all copies of the 53 * software, derivative works or modified versions, and any portions 54 * thereof, and that both notices appear in supporting documentation. 55 * 56 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 57 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 58 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 59 * 60 * Carnegie Mellon requests users of this software to return to 61 * 62 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 63 * School of Computer Science 64 * Carnegie Mellon University 65 * Pittsburgh PA 15213-3890 66 * 67 * any improvements or extensions that they make and grant Carnegie the 68 * rights to redistribute these changes. 69 */ 70 71/* 72 * The proverbial page-out daemon. 73 */ 74 75#include <sys/cdefs.h> 76__FBSDID("$FreeBSD: head/sys/vm/vm_pageout.c 170658 2007-06-13 06:10:10Z alc $"); 77 78#include "opt_vm.h" 79#include <sys/param.h> 80#include <sys/systm.h> 81#include <sys/kernel.h> 82#include <sys/eventhandler.h> 83#include <sys/lock.h> 84#include <sys/mutex.h> 85#include <sys/proc.h> 86#include <sys/kthread.h> 87#include <sys/ktr.h> 88#include <sys/resourcevar.h> 89#include <sys/sched.h> 90#include <sys/signalvar.h> 91#include <sys/vnode.h> 92#include <sys/vmmeter.h> 93#include <sys/sx.h> 94#include <sys/sysctl.h> 95 96#include <vm/vm.h> 97#include <vm/vm_param.h> 98#include <vm/vm_object.h> 99#include <vm/vm_page.h> 100#include <vm/vm_map.h> 101#include <vm/vm_pageout.h> 102#include <vm/vm_pager.h> 103#include <vm/swap_pager.h> 104#include <vm/vm_extern.h> 105#include <vm/uma.h> 106 107#include <machine/mutex.h> 108 109/* 110 * System initialization 111 */ 112 113/* the kernel process "vm_pageout"*/ 114static void vm_pageout(void); 115static int vm_pageout_clean(vm_page_t); 116static void vm_pageout_scan(int pass); 117 118struct proc *pageproc; 119 120static struct kproc_desc page_kp = { 121 "pagedaemon", 122 vm_pageout, 123 &pageproc 124}; 125SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp) 126 127#if !defined(NO_SWAPPING) 128/* the kernel process "vm_daemon"*/ 129static void vm_daemon(void); 130static struct proc *vmproc; 131 132static struct kproc_desc vm_kp = { 133 "vmdaemon", 134 vm_daemon, 135 &vmproc 136}; 137SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp) 138#endif 139 140 141int vm_pages_needed; /* Event on which pageout daemon sleeps */ 142int vm_pageout_deficit; /* Estimated number of pages deficit */ 143int vm_pageout_pages_needed; /* flag saying that the pageout daemon needs pages */ 144 145#if !defined(NO_SWAPPING) 146static int vm_pageout_req_swapout; /* XXX */ 147static int vm_daemon_needed; 148#endif 149static int vm_max_launder = 32; 150static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0; 151static int vm_pageout_full_stats_interval = 0; 152static int vm_pageout_algorithm=0; 153static int defer_swap_pageouts=0; 154static int disable_swap_pageouts=0; 155 156#if defined(NO_SWAPPING) 157static int vm_swap_enabled=0; 158static int vm_swap_idle_enabled=0; 159#else 160static int vm_swap_enabled=1; 161static int vm_swap_idle_enabled=0; 162#endif 163 164SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm, 165 CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt"); 166 167SYSCTL_INT(_vm, OID_AUTO, max_launder, 168 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout"); 169 170SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max, 171 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length"); 172 173SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval, 174 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan"); 175 176SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval, 177 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan"); 178 179#if defined(NO_SWAPPING) 180SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 181 CTLFLAG_RD, &vm_swap_enabled, 0, ""); 182SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 183 CTLFLAG_RD, &vm_swap_idle_enabled, 0, ""); 184#else 185SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 186 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout"); 187SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 188 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 189#endif 190 191SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts, 192 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem"); 193 194SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts, 195 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages"); 196 197static int pageout_lock_miss; 198SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss, 199 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout"); 200 201#define VM_PAGEOUT_PAGE_COUNT 16 202int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT; 203 204int vm_page_max_wired; /* XXX max # of wired pages system-wide */ 205 206#if !defined(NO_SWAPPING) 207static void vm_pageout_map_deactivate_pages(vm_map_t, long); 208static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long); 209static void vm_req_vmdaemon(void); 210#endif 211static void vm_pageout_page_stats(void); 212 213/* 214 * vm_pageout_fallback_object_lock: 215 * 216 * Lock vm object currently associated with `m'. VM_OBJECT_TRYLOCK is 217 * known to have failed and page queue must be either PQ_ACTIVE or 218 * PQ_INACTIVE. To avoid lock order violation, unlock the page queues 219 * while locking the vm object. Use marker page to detect page queue 220 * changes and maintain notion of next page on page queue. Return 221 * TRUE if no changes were detected, FALSE otherwise. vm object is 222 * locked on return. 223 * 224 * This function depends on both the lock portion of struct vm_object 225 * and normal struct vm_page being type stable. 226 */ 227static boolean_t 228vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next) 229{ 230 struct vm_page marker; 231 boolean_t unchanged; 232 u_short queue; 233 vm_object_t object; 234 235 /* 236 * Initialize our marker 237 */ 238 bzero(&marker, sizeof(marker)); 239 marker.flags = PG_FICTITIOUS | PG_MARKER; 240 marker.oflags = VPO_BUSY; 241 marker.queue = m->queue; 242 marker.wire_count = 1; 243 244 queue = m->queue; 245 object = m->object; 246 247 TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl, 248 m, &marker, pageq); 249 vm_page_unlock_queues(); 250 VM_OBJECT_LOCK(object); 251 vm_page_lock_queues(); 252 253 /* Page queue might have changed. */ 254 *next = TAILQ_NEXT(&marker, pageq); 255 unchanged = (m->queue == queue && 256 m->object == object && 257 &marker == TAILQ_NEXT(m, pageq)); 258 TAILQ_REMOVE(&vm_page_queues[queue].pl, 259 &marker, pageq); 260 return (unchanged); 261} 262 263/* 264 * vm_pageout_clean: 265 * 266 * Clean the page and remove it from the laundry. 267 * 268 * We set the busy bit to cause potential page faults on this page to 269 * block. Note the careful timing, however, the busy bit isn't set till 270 * late and we cannot do anything that will mess with the page. 271 */ 272static int 273vm_pageout_clean(m) 274 vm_page_t m; 275{ 276 vm_object_t object; 277 vm_page_t mc[2*vm_pageout_page_count]; 278 int pageout_count; 279 int ib, is, page_base; 280 vm_pindex_t pindex = m->pindex; 281 282 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 283 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 284 285 /* 286 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP 287 * with the new swapper, but we could have serious problems paging 288 * out other object types if there is insufficient memory. 289 * 290 * Unfortunately, checking free memory here is far too late, so the 291 * check has been moved up a procedural level. 292 */ 293 294 /* 295 * Don't mess with the page if it's busy, held, or special 296 */ 297 if ((m->hold_count != 0) || 298 ((m->busy != 0) || (m->oflags & VPO_BUSY) || 299 (m->flags & PG_UNMANAGED))) { 300 return 0; 301 } 302 303 mc[vm_pageout_page_count] = m; 304 pageout_count = 1; 305 page_base = vm_pageout_page_count; 306 ib = 1; 307 is = 1; 308 309 /* 310 * Scan object for clusterable pages. 311 * 312 * We can cluster ONLY if: ->> the page is NOT 313 * clean, wired, busy, held, or mapped into a 314 * buffer, and one of the following: 315 * 1) The page is inactive, or a seldom used 316 * active page. 317 * -or- 318 * 2) we force the issue. 319 * 320 * During heavy mmap/modification loads the pageout 321 * daemon can really fragment the underlying file 322 * due to flushing pages out of order and not trying 323 * align the clusters (which leave sporatic out-of-order 324 * holes). To solve this problem we do the reverse scan 325 * first and attempt to align our cluster, then do a 326 * forward scan if room remains. 327 */ 328 object = m->object; 329more: 330 while (ib && pageout_count < vm_pageout_page_count) { 331 vm_page_t p; 332 333 if (ib > pindex) { 334 ib = 0; 335 break; 336 } 337 338 if ((p = vm_page_lookup(object, pindex - ib)) == NULL) { 339 ib = 0; 340 break; 341 } 342 if (VM_PAGE_INQUEUE1(p, PQ_CACHE) || 343 (p->oflags & VPO_BUSY) || p->busy || 344 (p->flags & PG_UNMANAGED)) { 345 ib = 0; 346 break; 347 } 348 vm_page_test_dirty(p); 349 if ((p->dirty & p->valid) == 0 || 350 p->queue != PQ_INACTIVE || 351 p->wire_count != 0 || /* may be held by buf cache */ 352 p->hold_count != 0) { /* may be undergoing I/O */ 353 ib = 0; 354 break; 355 } 356 mc[--page_base] = p; 357 ++pageout_count; 358 ++ib; 359 /* 360 * alignment boundry, stop here and switch directions. Do 361 * not clear ib. 362 */ 363 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0) 364 break; 365 } 366 367 while (pageout_count < vm_pageout_page_count && 368 pindex + is < object->size) { 369 vm_page_t p; 370 371 if ((p = vm_page_lookup(object, pindex + is)) == NULL) 372 break; 373 if (VM_PAGE_INQUEUE1(p, PQ_CACHE) || 374 (p->oflags & VPO_BUSY) || p->busy || 375 (p->flags & PG_UNMANAGED)) { 376 break; 377 } 378 vm_page_test_dirty(p); 379 if ((p->dirty & p->valid) == 0 || 380 p->queue != PQ_INACTIVE || 381 p->wire_count != 0 || /* may be held by buf cache */ 382 p->hold_count != 0) { /* may be undergoing I/O */ 383 break; 384 } 385 mc[page_base + pageout_count] = p; 386 ++pageout_count; 387 ++is; 388 } 389 390 /* 391 * If we exhausted our forward scan, continue with the reverse scan 392 * when possible, even past a page boundry. This catches boundry 393 * conditions. 394 */ 395 if (ib && pageout_count < vm_pageout_page_count) 396 goto more; 397 398 /* 399 * we allow reads during pageouts... 400 */ 401 return (vm_pageout_flush(&mc[page_base], pageout_count, 0)); 402} 403 404/* 405 * vm_pageout_flush() - launder the given pages 406 * 407 * The given pages are laundered. Note that we setup for the start of 408 * I/O ( i.e. busy the page ), mark it read-only, and bump the object 409 * reference count all in here rather then in the parent. If we want 410 * the parent to do more sophisticated things we may have to change 411 * the ordering. 412 */ 413int 414vm_pageout_flush(vm_page_t *mc, int count, int flags) 415{ 416 vm_object_t object = mc[0]->object; 417 int pageout_status[count]; 418 int numpagedout = 0; 419 int i; 420 421 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 422 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 423 /* 424 * Initiate I/O. Bump the vm_page_t->busy counter and 425 * mark the pages read-only. 426 * 427 * We do not have to fixup the clean/dirty bits here... we can 428 * allow the pager to do it after the I/O completes. 429 * 430 * NOTE! mc[i]->dirty may be partial or fragmented due to an 431 * edge case with file fragments. 432 */ 433 for (i = 0; i < count; i++) { 434 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, 435 ("vm_pageout_flush: partially invalid page %p index %d/%d", 436 mc[i], i, count)); 437 vm_page_io_start(mc[i]); 438 pmap_remove_write(mc[i]); 439 } 440 vm_page_unlock_queues(); 441 vm_object_pip_add(object, count); 442 443 vm_pager_put_pages(object, mc, count, flags, pageout_status); 444 445 vm_page_lock_queues(); 446 for (i = 0; i < count; i++) { 447 vm_page_t mt = mc[i]; 448 449 KASSERT((mt->flags & PG_WRITEABLE) == 0, 450 ("vm_pageout_flush: page %p is not write protected", mt)); 451 switch (pageout_status[i]) { 452 case VM_PAGER_OK: 453 case VM_PAGER_PEND: 454 numpagedout++; 455 break; 456 case VM_PAGER_BAD: 457 /* 458 * Page outside of range of object. Right now we 459 * essentially lose the changes by pretending it 460 * worked. 461 */ 462 pmap_clear_modify(mt); 463 vm_page_undirty(mt); 464 break; 465 case VM_PAGER_ERROR: 466 case VM_PAGER_FAIL: 467 /* 468 * If page couldn't be paged out, then reactivate the 469 * page so it doesn't clog the inactive list. (We 470 * will try paging out it again later). 471 */ 472 vm_page_activate(mt); 473 break; 474 case VM_PAGER_AGAIN: 475 break; 476 } 477 478 /* 479 * If the operation is still going, leave the page busy to 480 * block all other accesses. Also, leave the paging in 481 * progress indicator set so that we don't attempt an object 482 * collapse. 483 */ 484 if (pageout_status[i] != VM_PAGER_PEND) { 485 vm_object_pip_wakeup(object); 486 vm_page_io_finish(mt); 487 if (vm_page_count_severe()) 488 vm_page_try_to_cache(mt); 489 } 490 } 491 return numpagedout; 492} 493 494#if !defined(NO_SWAPPING) 495/* 496 * vm_pageout_object_deactivate_pages 497 * 498 * deactivate enough pages to satisfy the inactive target 499 * requirements or if vm_page_proc_limit is set, then 500 * deactivate all of the pages in the object and its 501 * backing_objects. 502 * 503 * The object and map must be locked. 504 */ 505static void 506vm_pageout_object_deactivate_pages(pmap, first_object, desired) 507 pmap_t pmap; 508 vm_object_t first_object; 509 long desired; 510{ 511 vm_object_t backing_object, object; 512 vm_page_t p, next; 513 int actcount, rcount, remove_mode; 514 515 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED); 516 if (first_object->type == OBJT_DEVICE || first_object->type == OBJT_PHYS) 517 return; 518 for (object = first_object;; object = backing_object) { 519 if (pmap_resident_count(pmap) <= desired) 520 goto unlock_return; 521 if (object->paging_in_progress) 522 goto unlock_return; 523 524 remove_mode = 0; 525 if (object->shadow_count > 1) 526 remove_mode = 1; 527 /* 528 * scan the objects entire memory queue 529 */ 530 rcount = object->resident_page_count; 531 p = TAILQ_FIRST(&object->memq); 532 vm_page_lock_queues(); 533 while (p && (rcount-- > 0)) { 534 if (pmap_resident_count(pmap) <= desired) { 535 vm_page_unlock_queues(); 536 goto unlock_return; 537 } 538 next = TAILQ_NEXT(p, listq); 539 cnt.v_pdpages++; 540 if (p->wire_count != 0 || 541 p->hold_count != 0 || 542 p->busy != 0 || 543 (p->oflags & VPO_BUSY) || 544 (p->flags & PG_UNMANAGED) || 545 !pmap_page_exists_quick(pmap, p)) { 546 p = next; 547 continue; 548 } 549 actcount = pmap_ts_referenced(p); 550 if (actcount) { 551 vm_page_flag_set(p, PG_REFERENCED); 552 } else if (p->flags & PG_REFERENCED) { 553 actcount = 1; 554 } 555 if ((p->queue != PQ_ACTIVE) && 556 (p->flags & PG_REFERENCED)) { 557 vm_page_activate(p); 558 p->act_count += actcount; 559 vm_page_flag_clear(p, PG_REFERENCED); 560 } else if (p->queue == PQ_ACTIVE) { 561 if ((p->flags & PG_REFERENCED) == 0) { 562 p->act_count -= min(p->act_count, ACT_DECLINE); 563 if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) { 564 pmap_remove_all(p); 565 vm_page_deactivate(p); 566 } else { 567 vm_pageq_requeue(p); 568 } 569 } else { 570 vm_page_activate(p); 571 vm_page_flag_clear(p, PG_REFERENCED); 572 if (p->act_count < (ACT_MAX - ACT_ADVANCE)) 573 p->act_count += ACT_ADVANCE; 574 vm_pageq_requeue(p); 575 } 576 } else if (p->queue == PQ_INACTIVE) { 577 pmap_remove_all(p); 578 } 579 p = next; 580 } 581 vm_page_unlock_queues(); 582 if ((backing_object = object->backing_object) == NULL) 583 goto unlock_return; 584 VM_OBJECT_LOCK(backing_object); 585 if (object != first_object) 586 VM_OBJECT_UNLOCK(object); 587 } 588unlock_return: 589 if (object != first_object) 590 VM_OBJECT_UNLOCK(object); 591} 592 593/* 594 * deactivate some number of pages in a map, try to do it fairly, but 595 * that is really hard to do. 596 */ 597static void 598vm_pageout_map_deactivate_pages(map, desired) 599 vm_map_t map; 600 long desired; 601{ 602 vm_map_entry_t tmpe; 603 vm_object_t obj, bigobj; 604 int nothingwired; 605 606 if (!vm_map_trylock(map)) 607 return; 608 609 bigobj = NULL; 610 nothingwired = TRUE; 611 612 /* 613 * first, search out the biggest object, and try to free pages from 614 * that. 615 */ 616 tmpe = map->header.next; 617 while (tmpe != &map->header) { 618 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 619 obj = tmpe->object.vm_object; 620 if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) { 621 if (obj->shadow_count <= 1 && 622 (bigobj == NULL || 623 bigobj->resident_page_count < obj->resident_page_count)) { 624 if (bigobj != NULL) 625 VM_OBJECT_UNLOCK(bigobj); 626 bigobj = obj; 627 } else 628 VM_OBJECT_UNLOCK(obj); 629 } 630 } 631 if (tmpe->wired_count > 0) 632 nothingwired = FALSE; 633 tmpe = tmpe->next; 634 } 635 636 if (bigobj != NULL) { 637 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired); 638 VM_OBJECT_UNLOCK(bigobj); 639 } 640 /* 641 * Next, hunt around for other pages to deactivate. We actually 642 * do this search sort of wrong -- .text first is not the best idea. 643 */ 644 tmpe = map->header.next; 645 while (tmpe != &map->header) { 646 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 647 break; 648 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 649 obj = tmpe->object.vm_object; 650 if (obj != NULL) { 651 VM_OBJECT_LOCK(obj); 652 vm_pageout_object_deactivate_pages(map->pmap, obj, desired); 653 VM_OBJECT_UNLOCK(obj); 654 } 655 } 656 tmpe = tmpe->next; 657 } 658 659 /* 660 * Remove all mappings if a process is swapped out, this will free page 661 * table pages. 662 */ 663 if (desired == 0 && nothingwired) { 664 pmap_remove(vm_map_pmap(map), vm_map_min(map), 665 vm_map_max(map)); 666 } 667 vm_map_unlock(map); 668} 669#endif /* !defined(NO_SWAPPING) */ 670 671/* 672 * vm_pageout_scan does the dirty work for the pageout daemon. 673 */ 674static void 675vm_pageout_scan(int pass) 676{ 677 vm_page_t m, next; 678 struct vm_page marker; 679 int page_shortage, maxscan, pcount; 680 int addl_page_shortage, addl_page_shortage_init; 681 struct proc *p, *bigproc; 682 struct thread *td; 683 vm_offset_t size, bigsize; 684 vm_object_t object; 685 int actcount, cache_cur, cache_first_failure; 686 static int cache_last_free; 687 int vnodes_skipped = 0; 688 int maxlaunder; 689 690 mtx_lock(&Giant); 691 /* 692 * Decrease registered cache sizes. 693 */ 694 EVENTHANDLER_INVOKE(vm_lowmem, 0); 695 /* 696 * We do this explicitly after the caches have been drained above. 697 */ 698 uma_reclaim(); 699 700 addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit); 701 702 /* 703 * Calculate the number of pages we want to either free or move 704 * to the cache. 705 */ 706 page_shortage = vm_paging_target() + addl_page_shortage_init; 707 708 /* 709 * Initialize our marker 710 */ 711 bzero(&marker, sizeof(marker)); 712 marker.flags = PG_FICTITIOUS | PG_MARKER; 713 marker.oflags = VPO_BUSY; 714 marker.queue = PQ_INACTIVE; 715 marker.wire_count = 1; 716 717 /* 718 * Start scanning the inactive queue for pages we can move to the 719 * cache or free. The scan will stop when the target is reached or 720 * we have scanned the entire inactive queue. Note that m->act_count 721 * is not used to form decisions for the inactive queue, only for the 722 * active queue. 723 * 724 * maxlaunder limits the number of dirty pages we flush per scan. 725 * For most systems a smaller value (16 or 32) is more robust under 726 * extreme memory and disk pressure because any unnecessary writes 727 * to disk can result in extreme performance degredation. However, 728 * systems with excessive dirty pages (especially when MAP_NOSYNC is 729 * used) will die horribly with limited laundering. If the pageout 730 * daemon cannot clean enough pages in the first pass, we let it go 731 * all out in succeeding passes. 732 */ 733 if ((maxlaunder = vm_max_launder) <= 1) 734 maxlaunder = 1; 735 if (pass) 736 maxlaunder = 10000; 737 vm_page_lock_queues(); 738rescan0: 739 addl_page_shortage = addl_page_shortage_init; 740 maxscan = cnt.v_inactive_count; 741 742 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl); 743 m != NULL && maxscan-- > 0 && page_shortage > 0; 744 m = next) { 745 746 cnt.v_pdpages++; 747 748 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE) { 749 goto rescan0; 750 } 751 752 next = TAILQ_NEXT(m, pageq); 753 object = m->object; 754 755 /* 756 * skip marker pages 757 */ 758 if (m->flags & PG_MARKER) 759 continue; 760 761 /* 762 * A held page may be undergoing I/O, so skip it. 763 */ 764 if (m->hold_count) { 765 vm_pageq_requeue(m); 766 addl_page_shortage++; 767 continue; 768 } 769 /* 770 * Don't mess with busy pages, keep in the front of the 771 * queue, most likely are being paged out. 772 */ 773 if (!VM_OBJECT_TRYLOCK(object) && 774 (!vm_pageout_fallback_object_lock(m, &next) || 775 m->hold_count != 0)) { 776 VM_OBJECT_UNLOCK(object); 777 addl_page_shortage++; 778 continue; 779 } 780 if (m->busy || (m->oflags & VPO_BUSY)) { 781 VM_OBJECT_UNLOCK(object); 782 addl_page_shortage++; 783 continue; 784 } 785 786 /* 787 * If the object is not being used, we ignore previous 788 * references. 789 */ 790 if (object->ref_count == 0) { 791 vm_page_flag_clear(m, PG_REFERENCED); 792 pmap_clear_reference(m); 793 794 /* 795 * Otherwise, if the page has been referenced while in the 796 * inactive queue, we bump the "activation count" upwards, 797 * making it less likely that the page will be added back to 798 * the inactive queue prematurely again. Here we check the 799 * page tables (or emulated bits, if any), given the upper 800 * level VM system not knowing anything about existing 801 * references. 802 */ 803 } else if (((m->flags & PG_REFERENCED) == 0) && 804 (actcount = pmap_ts_referenced(m))) { 805 vm_page_activate(m); 806 VM_OBJECT_UNLOCK(object); 807 m->act_count += (actcount + ACT_ADVANCE); 808 continue; 809 } 810 811 /* 812 * If the upper level VM system knows about any page 813 * references, we activate the page. We also set the 814 * "activation count" higher than normal so that we will less 815 * likely place pages back onto the inactive queue again. 816 */ 817 if ((m->flags & PG_REFERENCED) != 0) { 818 vm_page_flag_clear(m, PG_REFERENCED); 819 actcount = pmap_ts_referenced(m); 820 vm_page_activate(m); 821 VM_OBJECT_UNLOCK(object); 822 m->act_count += (actcount + ACT_ADVANCE + 1); 823 continue; 824 } 825 826 /* 827 * If the upper level VM system doesn't know anything about 828 * the page being dirty, we have to check for it again. As 829 * far as the VM code knows, any partially dirty pages are 830 * fully dirty. 831 */ 832 if (m->dirty == 0 && !pmap_is_modified(m)) { 833 /* 834 * Avoid a race condition: Unless write access is 835 * removed from the page, another processor could 836 * modify it before all access is removed by the call 837 * to vm_page_cache() below. If vm_page_cache() finds 838 * that the page has been modified when it removes all 839 * access, it panics because it cannot cache dirty 840 * pages. In principle, we could eliminate just write 841 * access here rather than all access. In the expected 842 * case, when there are no last instant modifications 843 * to the page, removing all access will be cheaper 844 * overall. 845 */ 846 if ((m->flags & PG_WRITEABLE) != 0) 847 pmap_remove_all(m); 848 } else { 849 vm_page_dirty(m); 850 } 851 852 if (m->valid == 0) { 853 /* 854 * Invalid pages can be easily freed 855 */ 856 vm_page_free(m); 857 cnt.v_dfree++; 858 --page_shortage; 859 } else if (m->dirty == 0) { 860 /* 861 * Clean pages can be placed onto the cache queue. 862 * This effectively frees them. 863 */ 864 vm_page_cache(m); 865 --page_shortage; 866 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) { 867 /* 868 * Dirty pages need to be paged out, but flushing 869 * a page is extremely expensive verses freeing 870 * a clean page. Rather then artificially limiting 871 * the number of pages we can flush, we instead give 872 * dirty pages extra priority on the inactive queue 873 * by forcing them to be cycled through the queue 874 * twice before being flushed, after which the 875 * (now clean) page will cycle through once more 876 * before being freed. This significantly extends 877 * the thrash point for a heavily loaded machine. 878 */ 879 vm_page_flag_set(m, PG_WINATCFLS); 880 vm_pageq_requeue(m); 881 } else if (maxlaunder > 0) { 882 /* 883 * We always want to try to flush some dirty pages if 884 * we encounter them, to keep the system stable. 885 * Normally this number is small, but under extreme 886 * pressure where there are insufficient clean pages 887 * on the inactive queue, we may have to go all out. 888 */ 889 int swap_pageouts_ok; 890 struct vnode *vp = NULL; 891 struct mount *mp; 892 893 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 894 swap_pageouts_ok = 1; 895 } else { 896 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 897 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 898 vm_page_count_min()); 899 900 } 901 902 /* 903 * We don't bother paging objects that are "dead". 904 * Those objects are in a "rundown" state. 905 */ 906 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 907 VM_OBJECT_UNLOCK(object); 908 vm_pageq_requeue(m); 909 continue; 910 } 911 912 /* 913 * Following operations may unlock 914 * vm_page_queue_mtx, invalidating the 'next' 915 * pointer. To prevent an inordinate number 916 * of restarts we use our marker to remember 917 * our place. 918 * 919 */ 920 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, 921 m, &marker, pageq); 922 /* 923 * The object is already known NOT to be dead. It 924 * is possible for the vget() to block the whole 925 * pageout daemon, but the new low-memory handling 926 * code should prevent it. 927 * 928 * The previous code skipped locked vnodes and, worse, 929 * reordered pages in the queue. This results in 930 * completely non-deterministic operation and, on a 931 * busy system, can lead to extremely non-optimal 932 * pageouts. For example, it can cause clean pages 933 * to be freed and dirty pages to be moved to the end 934 * of the queue. Since dirty pages are also moved to 935 * the end of the queue once-cleaned, this gives 936 * way too large a weighting to defering the freeing 937 * of dirty pages. 938 * 939 * We can't wait forever for the vnode lock, we might 940 * deadlock due to a vn_read() getting stuck in 941 * vm_wait while holding this vnode. We skip the 942 * vnode if we can't get it in a reasonable amount 943 * of time. 944 */ 945 if (object->type == OBJT_VNODE) { 946 vp = object->handle; 947 mp = NULL; 948 if (vp->v_type == VREG && 949 vn_start_write(vp, &mp, V_NOWAIT) != 0) { 950 ++pageout_lock_miss; 951 if (object->flags & OBJ_MIGHTBEDIRTY) 952 vnodes_skipped++; 953 vp = NULL; 954 goto unlock_and_continue; 955 } 956 vm_page_unlock_queues(); 957 VI_LOCK(vp); 958 VM_OBJECT_UNLOCK(object); 959 if (vget(vp, LK_EXCLUSIVE | LK_INTERLOCK | 960 LK_TIMELOCK, curthread)) { 961 VM_OBJECT_LOCK(object); 962 vm_page_lock_queues(); 963 ++pageout_lock_miss; 964 vn_finished_write(mp); 965 if (object->flags & OBJ_MIGHTBEDIRTY) 966 vnodes_skipped++; 967 vp = NULL; 968 goto unlock_and_continue; 969 } 970 VM_OBJECT_LOCK(object); 971 vm_page_lock_queues(); 972 /* 973 * The page might have been moved to another 974 * queue during potential blocking in vget() 975 * above. The page might have been freed and 976 * reused for another vnode. The object might 977 * have been reused for another vnode. 978 */ 979 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE || 980 m->object != object || 981 object->handle != vp || 982 TAILQ_NEXT(m, pageq) != &marker) { 983 if (object->flags & OBJ_MIGHTBEDIRTY) 984 vnodes_skipped++; 985 goto unlock_and_continue; 986 } 987 988 /* 989 * The page may have been busied during the 990 * blocking in vput(); We don't move the 991 * page back onto the end of the queue so that 992 * statistics are more correct if we don't. 993 */ 994 if (m->busy || (m->oflags & VPO_BUSY)) { 995 goto unlock_and_continue; 996 } 997 998 /* 999 * If the page has become held it might 1000 * be undergoing I/O, so skip it 1001 */ 1002 if (m->hold_count) { 1003 vm_pageq_requeue(m); 1004 if (object->flags & OBJ_MIGHTBEDIRTY) 1005 vnodes_skipped++; 1006 goto unlock_and_continue; 1007 } 1008 } 1009 1010 /* 1011 * If a page is dirty, then it is either being washed 1012 * (but not yet cleaned) or it is still in the 1013 * laundry. If it is still in the laundry, then we 1014 * start the cleaning operation. 1015 * 1016 * decrement page_shortage on success to account for 1017 * the (future) cleaned page. Otherwise we could wind 1018 * up laundering or cleaning too many pages. 1019 */ 1020 if (vm_pageout_clean(m) != 0) { 1021 --page_shortage; 1022 --maxlaunder; 1023 } 1024unlock_and_continue: 1025 VM_OBJECT_UNLOCK(object); 1026 if (vp) { 1027 vm_page_unlock_queues(); 1028 vput(vp); 1029 vn_finished_write(mp); 1030 vm_page_lock_queues(); 1031 } 1032 next = TAILQ_NEXT(&marker, pageq); 1033 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, 1034 &marker, pageq); 1035 continue; 1036 } 1037 VM_OBJECT_UNLOCK(object); 1038 } 1039 1040 /* 1041 * Compute the number of pages we want to try to move from the 1042 * active queue to the inactive queue. 1043 */ 1044 page_shortage = vm_paging_target() + 1045 cnt.v_inactive_target - cnt.v_inactive_count; 1046 page_shortage += addl_page_shortage; 1047 1048 /* 1049 * Scan the active queue for things we can deactivate. We nominally 1050 * track the per-page activity counter and use it to locate 1051 * deactivation candidates. 1052 */ 1053 pcount = cnt.v_active_count; 1054 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1055 1056 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) { 1057 1058 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE), 1059 ("vm_pageout_scan: page %p isn't active", m)); 1060 1061 next = TAILQ_NEXT(m, pageq); 1062 object = m->object; 1063 if ((m->flags & PG_MARKER) != 0) { 1064 m = next; 1065 continue; 1066 } 1067 if (!VM_OBJECT_TRYLOCK(object) && 1068 !vm_pageout_fallback_object_lock(m, &next)) { 1069 VM_OBJECT_UNLOCK(object); 1070 m = next; 1071 continue; 1072 } 1073 1074 /* 1075 * Don't deactivate pages that are busy. 1076 */ 1077 if ((m->busy != 0) || 1078 (m->oflags & VPO_BUSY) || 1079 (m->hold_count != 0)) { 1080 VM_OBJECT_UNLOCK(object); 1081 vm_pageq_requeue(m); 1082 m = next; 1083 continue; 1084 } 1085 1086 /* 1087 * The count for pagedaemon pages is done after checking the 1088 * page for eligibility... 1089 */ 1090 cnt.v_pdpages++; 1091 1092 /* 1093 * Check to see "how much" the page has been used. 1094 */ 1095 actcount = 0; 1096 if (object->ref_count != 0) { 1097 if (m->flags & PG_REFERENCED) { 1098 actcount += 1; 1099 } 1100 actcount += pmap_ts_referenced(m); 1101 if (actcount) { 1102 m->act_count += ACT_ADVANCE + actcount; 1103 if (m->act_count > ACT_MAX) 1104 m->act_count = ACT_MAX; 1105 } 1106 } 1107 1108 /* 1109 * Since we have "tested" this bit, we need to clear it now. 1110 */ 1111 vm_page_flag_clear(m, PG_REFERENCED); 1112 1113 /* 1114 * Only if an object is currently being used, do we use the 1115 * page activation count stats. 1116 */ 1117 if (actcount && (object->ref_count != 0)) { 1118 vm_pageq_requeue(m); 1119 } else { 1120 m->act_count -= min(m->act_count, ACT_DECLINE); 1121 if (vm_pageout_algorithm || 1122 object->ref_count == 0 || 1123 m->act_count == 0) { 1124 page_shortage--; 1125 if (object->ref_count == 0) { 1126 pmap_remove_all(m); 1127 if (m->dirty == 0) 1128 vm_page_cache(m); 1129 else 1130 vm_page_deactivate(m); 1131 } else { 1132 vm_page_deactivate(m); 1133 } 1134 } else { 1135 vm_pageq_requeue(m); 1136 } 1137 } 1138 VM_OBJECT_UNLOCK(object); 1139 m = next; 1140 } 1141 1142 /* 1143 * We try to maintain some *really* free pages, this allows interrupt 1144 * code to be guaranteed space. Since both cache and free queues 1145 * are considered basically 'free', moving pages from cache to free 1146 * does not effect other calculations. 1147 */ 1148 cache_cur = cache_last_free; 1149 cache_first_failure = -1; 1150 while (cnt.v_free_count < cnt.v_free_reserved && (cache_cur = 1151 (cache_cur + PQ_PRIME2) & PQ_COLORMASK) != cache_first_failure) { 1152 TAILQ_FOREACH(m, &vm_page_queues[PQ_CACHE + cache_cur].pl, 1153 pageq) { 1154 KASSERT(m->dirty == 0, 1155 ("Found dirty cache page %p", m)); 1156 KASSERT(!pmap_page_is_mapped(m), 1157 ("Found mapped cache page %p", m)); 1158 KASSERT((m->flags & PG_UNMANAGED) == 0, 1159 ("Found unmanaged cache page %p", m)); 1160 KASSERT(m->wire_count == 0, 1161 ("Found wired cache page %p", m)); 1162 if (m->hold_count == 0 && VM_OBJECT_TRYLOCK(object = 1163 m->object)) { 1164 KASSERT((m->oflags & VPO_BUSY) == 0 && 1165 m->busy == 0, ("Found busy cache page %p", 1166 m)); 1167 vm_page_free(m); 1168 VM_OBJECT_UNLOCK(object); 1169 cnt.v_dfree++; 1170 cache_last_free = cache_cur; 1171 cache_first_failure = -1; 1172 break; 1173 } 1174 } 1175 if (m == NULL && cache_first_failure == -1) 1176 cache_first_failure = cache_cur; 1177 } 1178 vm_page_unlock_queues(); 1179#if !defined(NO_SWAPPING) 1180 /* 1181 * Idle process swapout -- run once per second. 1182 */ 1183 if (vm_swap_idle_enabled) { 1184 static long lsec; 1185 if (time_second != lsec) { 1186 vm_pageout_req_swapout |= VM_SWAP_IDLE; 1187 vm_req_vmdaemon(); 1188 lsec = time_second; 1189 } 1190 } 1191#endif 1192 1193 /* 1194 * If we didn't get enough free pages, and we have skipped a vnode 1195 * in a writeable object, wakeup the sync daemon. And kick swapout 1196 * if we did not get enough free pages. 1197 */ 1198 if (vm_paging_target() > 0) { 1199 if (vnodes_skipped && vm_page_count_min()) 1200 (void) speedup_syncer(); 1201#if !defined(NO_SWAPPING) 1202 if (vm_swap_enabled && vm_page_count_target()) { 1203 vm_req_vmdaemon(); 1204 vm_pageout_req_swapout |= VM_SWAP_NORMAL; 1205 } 1206#endif 1207 } 1208 1209 /* 1210 * If we are critically low on one of RAM or swap and low on 1211 * the other, kill the largest process. However, we avoid 1212 * doing this on the first pass in order to give ourselves a 1213 * chance to flush out dirty vnode-backed pages and to allow 1214 * active pages to be moved to the inactive queue and reclaimed. 1215 * 1216 * We keep the process bigproc locked once we find it to keep anyone 1217 * from messing with it; however, there is a possibility of 1218 * deadlock if process B is bigproc and one of it's child processes 1219 * attempts to propagate a signal to B while we are waiting for A's 1220 * lock while walking this list. To avoid this, we don't block on 1221 * the process lock but just skip a process if it is already locked. 1222 */ 1223 if (pass != 0 && 1224 ((swap_pager_avail < 64 && vm_page_count_min()) || 1225 (swap_pager_full && vm_paging_target() > 0))) { 1226 bigproc = NULL; 1227 bigsize = 0; 1228 sx_slock(&allproc_lock); 1229 FOREACH_PROC_IN_SYSTEM(p) { 1230 int breakout; 1231 1232 if (PROC_TRYLOCK(p) == 0) 1233 continue; 1234 /* 1235 * If this is a system or protected process, skip it. 1236 */ 1237 if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) || 1238 (p->p_flag & P_PROTECTED) || 1239 ((p->p_pid < 48) && (swap_pager_avail != 0))) { 1240 PROC_UNLOCK(p); 1241 continue; 1242 } 1243 /* 1244 * If the process is in a non-running type state, 1245 * don't touch it. Check all the threads individually. 1246 */ 1247 PROC_SLOCK(p); 1248 breakout = 0; 1249 FOREACH_THREAD_IN_PROC(p, td) { 1250 thread_lock(td); 1251 if (!TD_ON_RUNQ(td) && 1252 !TD_IS_RUNNING(td) && 1253 !TD_IS_SLEEPING(td)) { 1254 thread_unlock(td); 1255 breakout = 1; 1256 break; 1257 } 1258 thread_unlock(td); 1259 } 1260 PROC_SUNLOCK(p); 1261 if (breakout) { 1262 PROC_UNLOCK(p); 1263 continue; 1264 } 1265 /* 1266 * get the process size 1267 */ 1268 if (!vm_map_trylock_read(&p->p_vmspace->vm_map)) { 1269 PROC_UNLOCK(p); 1270 continue; 1271 } 1272 size = vmspace_swap_count(p->p_vmspace); 1273 vm_map_unlock_read(&p->p_vmspace->vm_map); 1274 size += vmspace_resident_count(p->p_vmspace); 1275 /* 1276 * if the this process is bigger than the biggest one 1277 * remember it. 1278 */ 1279 if (size > bigsize) { 1280 if (bigproc != NULL) 1281 PROC_UNLOCK(bigproc); 1282 bigproc = p; 1283 bigsize = size; 1284 } else 1285 PROC_UNLOCK(p); 1286 } 1287 sx_sunlock(&allproc_lock); 1288 if (bigproc != NULL) { 1289 killproc(bigproc, "out of swap space"); 1290 PROC_SLOCK(bigproc); 1291 sched_nice(bigproc, PRIO_MIN); 1292 PROC_SUNLOCK(bigproc); 1293 PROC_UNLOCK(bigproc); 1294 wakeup(&cnt.v_free_count); 1295 } 1296 } 1297 mtx_unlock(&Giant); 1298} 1299 1300/* 1301 * This routine tries to maintain the pseudo LRU active queue, 1302 * so that during long periods of time where there is no paging, 1303 * that some statistic accumulation still occurs. This code 1304 * helps the situation where paging just starts to occur. 1305 */ 1306static void 1307vm_pageout_page_stats() 1308{ 1309 vm_object_t object; 1310 vm_page_t m,next; 1311 int pcount,tpcount; /* Number of pages to check */ 1312 static int fullintervalcount = 0; 1313 int page_shortage; 1314 1315 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1316 page_shortage = 1317 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) - 1318 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count); 1319 1320 if (page_shortage <= 0) 1321 return; 1322 1323 pcount = cnt.v_active_count; 1324 fullintervalcount += vm_pageout_stats_interval; 1325 if (fullintervalcount < vm_pageout_full_stats_interval) { 1326 tpcount = (vm_pageout_stats_max * cnt.v_active_count) / cnt.v_page_count; 1327 if (pcount > tpcount) 1328 pcount = tpcount; 1329 } else { 1330 fullintervalcount = 0; 1331 } 1332 1333 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1334 while ((m != NULL) && (pcount-- > 0)) { 1335 int actcount; 1336 1337 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE), 1338 ("vm_pageout_page_stats: page %p isn't active", m)); 1339 1340 next = TAILQ_NEXT(m, pageq); 1341 object = m->object; 1342 1343 if ((m->flags & PG_MARKER) != 0) { 1344 m = next; 1345 continue; 1346 } 1347 if (!VM_OBJECT_TRYLOCK(object) && 1348 !vm_pageout_fallback_object_lock(m, &next)) { 1349 VM_OBJECT_UNLOCK(object); 1350 m = next; 1351 continue; 1352 } 1353 1354 /* 1355 * Don't deactivate pages that are busy. 1356 */ 1357 if ((m->busy != 0) || 1358 (m->oflags & VPO_BUSY) || 1359 (m->hold_count != 0)) { 1360 VM_OBJECT_UNLOCK(object); 1361 vm_pageq_requeue(m); 1362 m = next; 1363 continue; 1364 } 1365 1366 actcount = 0; 1367 if (m->flags & PG_REFERENCED) { 1368 vm_page_flag_clear(m, PG_REFERENCED); 1369 actcount += 1; 1370 } 1371 1372 actcount += pmap_ts_referenced(m); 1373 if (actcount) { 1374 m->act_count += ACT_ADVANCE + actcount; 1375 if (m->act_count > ACT_MAX) 1376 m->act_count = ACT_MAX; 1377 vm_pageq_requeue(m); 1378 } else { 1379 if (m->act_count == 0) { 1380 /* 1381 * We turn off page access, so that we have 1382 * more accurate RSS stats. We don't do this 1383 * in the normal page deactivation when the 1384 * system is loaded VM wise, because the 1385 * cost of the large number of page protect 1386 * operations would be higher than the value 1387 * of doing the operation. 1388 */ 1389 pmap_remove_all(m); 1390 vm_page_deactivate(m); 1391 } else { 1392 m->act_count -= min(m->act_count, ACT_DECLINE); 1393 vm_pageq_requeue(m); 1394 } 1395 } 1396 VM_OBJECT_UNLOCK(object); 1397 m = next; 1398 } 1399} 1400 1401/* 1402 * vm_pageout is the high level pageout daemon. 1403 */ 1404static void 1405vm_pageout() 1406{ 1407 int error, pass; 1408 1409 /* 1410 * Initialize some paging parameters. 1411 */ 1412 cnt.v_interrupt_free_min = 2; 1413 if (cnt.v_page_count < 2000) 1414 vm_pageout_page_count = 8; 1415 1416 /* 1417 * v_free_reserved needs to include enough for the largest 1418 * swap pager structures plus enough for any pv_entry structs 1419 * when paging. 1420 */ 1421 if (cnt.v_page_count > 1024) 1422 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200; 1423 else 1424 cnt.v_free_min = 4; 1425 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1426 cnt.v_interrupt_free_min; 1427 cnt.v_free_reserved = vm_pageout_page_count + 1428 cnt.v_pageout_free_min + (cnt.v_page_count / 768) + PQ_NUMCOLORS; 1429 cnt.v_free_severe = cnt.v_free_min / 2; 1430 cnt.v_free_min += cnt.v_free_reserved; 1431 cnt.v_free_severe += cnt.v_free_reserved; 1432 1433 /* 1434 * v_free_target and v_cache_min control pageout hysteresis. Note 1435 * that these are more a measure of the VM cache queue hysteresis 1436 * then the VM free queue. Specifically, v_free_target is the 1437 * high water mark (free+cache pages). 1438 * 1439 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the 1440 * low water mark, while v_free_min is the stop. v_cache_min must 1441 * be big enough to handle memory needs while the pageout daemon 1442 * is signalled and run to free more pages. 1443 */ 1444 if (cnt.v_free_count > 6144) 1445 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved; 1446 else 1447 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved; 1448 1449 if (cnt.v_free_count > 2048) { 1450 cnt.v_cache_min = cnt.v_free_target; 1451 cnt.v_cache_max = 2 * cnt.v_cache_min; 1452 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2; 1453 } else { 1454 cnt.v_cache_min = 0; 1455 cnt.v_cache_max = 0; 1456 cnt.v_inactive_target = cnt.v_free_count / 4; 1457 } 1458 if (cnt.v_inactive_target > cnt.v_free_count / 3) 1459 cnt.v_inactive_target = cnt.v_free_count / 3; 1460 1461 /* XXX does not really belong here */ 1462 if (vm_page_max_wired == 0) 1463 vm_page_max_wired = cnt.v_free_count / 3; 1464 1465 if (vm_pageout_stats_max == 0) 1466 vm_pageout_stats_max = cnt.v_free_target; 1467 1468 /* 1469 * Set interval in seconds for stats scan. 1470 */ 1471 if (vm_pageout_stats_interval == 0) 1472 vm_pageout_stats_interval = 5; 1473 if (vm_pageout_full_stats_interval == 0) 1474 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4; 1475 1476 swap_pager_swap_init(); 1477 pass = 0; 1478 /* 1479 * The pageout daemon is never done, so loop forever. 1480 */ 1481 while (TRUE) { 1482 /* 1483 * If we have enough free memory, wakeup waiters. Do 1484 * not clear vm_pages_needed until we reach our target, 1485 * otherwise we may be woken up over and over again and 1486 * waste a lot of cpu. 1487 */ 1488 mtx_lock(&vm_page_queue_free_mtx); 1489 if (vm_pages_needed && !vm_page_count_min()) { 1490 if (!vm_paging_needed()) 1491 vm_pages_needed = 0; 1492 wakeup(&cnt.v_free_count); 1493 } 1494 if (vm_pages_needed) { 1495 /* 1496 * Still not done, take a second pass without waiting 1497 * (unlimited dirty cleaning), otherwise sleep a bit 1498 * and try again. 1499 */ 1500 ++pass; 1501 if (pass > 1) 1502 msleep(&vm_pages_needed, 1503 &vm_page_queue_free_mtx, PVM, "psleep", 1504 hz / 2); 1505 } else { 1506 /* 1507 * Good enough, sleep & handle stats. Prime the pass 1508 * for the next run. 1509 */ 1510 if (pass > 1) 1511 pass = 1; 1512 else 1513 pass = 0; 1514 error = msleep(&vm_pages_needed, 1515 &vm_page_queue_free_mtx, PVM, "psleep", 1516 vm_pageout_stats_interval * hz); 1517 if (error && !vm_pages_needed) { 1518 mtx_unlock(&vm_page_queue_free_mtx); 1519 pass = 0; 1520 vm_page_lock_queues(); 1521 vm_pageout_page_stats(); 1522 vm_page_unlock_queues(); 1523 continue; 1524 } 1525 } 1526 if (vm_pages_needed) 1527 cnt.v_pdwakeups++; 1528 mtx_unlock(&vm_page_queue_free_mtx); 1529 vm_pageout_scan(pass); 1530 } 1531} 1532 1533/* 1534 * Unless the free page queue lock is held by the caller, this function 1535 * should be regarded as advisory. Specifically, the caller should 1536 * not msleep() on &cnt.v_free_count following this function unless 1537 * the free page queue lock is held until the msleep() is performed. 1538 */ 1539void 1540pagedaemon_wakeup() 1541{ 1542 1543 if (!vm_pages_needed && curthread->td_proc != pageproc) { 1544 vm_pages_needed = 1; 1545 wakeup(&vm_pages_needed); 1546 } 1547} 1548 1549#if !defined(NO_SWAPPING) 1550static void 1551vm_req_vmdaemon() 1552{ 1553 static int lastrun = 0; 1554 1555 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1556 wakeup(&vm_daemon_needed); 1557 lastrun = ticks; 1558 } 1559} 1560 1561static void 1562vm_daemon() 1563{ 1564 struct rlimit rsslim; 1565 struct proc *p; 1566 struct thread *td; 1567 int breakout; 1568 1569 mtx_lock(&Giant); 1570 while (TRUE) { 1571 tsleep(&vm_daemon_needed, PPAUSE, "psleep", 0); 1572 if (vm_pageout_req_swapout) { 1573 swapout_procs(vm_pageout_req_swapout); 1574 vm_pageout_req_swapout = 0; 1575 } 1576 /* 1577 * scan the processes for exceeding their rlimits or if 1578 * process is swapped out -- deactivate pages 1579 */ 1580 sx_slock(&allproc_lock); 1581 FOREACH_PROC_IN_SYSTEM(p) { 1582 vm_pindex_t limit, size; 1583 1584 /* 1585 * if this is a system process or if we have already 1586 * looked at this process, skip it. 1587 */ 1588 PROC_LOCK(p); 1589 if (p->p_flag & (P_SYSTEM | P_WEXIT)) { 1590 PROC_UNLOCK(p); 1591 continue; 1592 } 1593 /* 1594 * if the process is in a non-running type state, 1595 * don't touch it. 1596 */ 1597 PROC_SLOCK(p); 1598 breakout = 0; 1599 FOREACH_THREAD_IN_PROC(p, td) { 1600 thread_lock(td); 1601 if (!TD_ON_RUNQ(td) && 1602 !TD_IS_RUNNING(td) && 1603 !TD_IS_SLEEPING(td)) { 1604 thread_unlock(td); 1605 breakout = 1; 1606 break; 1607 } 1608 thread_unlock(td); 1609 } 1610 PROC_SUNLOCK(p); 1611 if (breakout) { 1612 PROC_UNLOCK(p); 1613 continue; 1614 } 1615 /* 1616 * get a limit 1617 */ 1618 lim_rlimit(p, RLIMIT_RSS, &rsslim); 1619 limit = OFF_TO_IDX( 1620 qmin(rsslim.rlim_cur, rsslim.rlim_max)); 1621 1622 /* 1623 * let processes that are swapped out really be 1624 * swapped out set the limit to nothing (will force a 1625 * swap-out.) 1626 */ 1627 if ((p->p_sflag & PS_INMEM) == 0) 1628 limit = 0; /* XXX */ 1629 PROC_UNLOCK(p); 1630 1631 size = vmspace_resident_count(p->p_vmspace); 1632 if (limit >= 0 && size >= limit) { 1633 vm_pageout_map_deactivate_pages( 1634 &p->p_vmspace->vm_map, limit); 1635 } 1636 } 1637 sx_sunlock(&allproc_lock); 1638 } 1639} 1640#endif /* !defined(NO_SWAPPING) */ 1641