vm_pageout.c revision 170816
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 170816 2007-06-16 04:57:06Z 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; 686 int vnodes_skipped = 0; 687 int maxlaunder; 688 689 mtx_lock(&Giant); 690 /* 691 * Decrease registered cache sizes. 692 */ 693 EVENTHANDLER_INVOKE(vm_lowmem, 0); 694 /* 695 * We do this explicitly after the caches have been drained above. 696 */ 697 uma_reclaim(); 698 699 addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit); 700 701 /* 702 * Calculate the number of pages we want to either free or move 703 * to the cache. 704 */ 705 page_shortage = vm_paging_target() + addl_page_shortage_init; 706 707 /* 708 * Initialize our marker 709 */ 710 bzero(&marker, sizeof(marker)); 711 marker.flags = PG_FICTITIOUS | PG_MARKER; 712 marker.oflags = VPO_BUSY; 713 marker.queue = PQ_INACTIVE; 714 marker.wire_count = 1; 715 716 /* 717 * Start scanning the inactive queue for pages we can move to the 718 * cache or free. The scan will stop when the target is reached or 719 * we have scanned the entire inactive queue. Note that m->act_count 720 * is not used to form decisions for the inactive queue, only for the 721 * active queue. 722 * 723 * maxlaunder limits the number of dirty pages we flush per scan. 724 * For most systems a smaller value (16 or 32) is more robust under 725 * extreme memory and disk pressure because any unnecessary writes 726 * to disk can result in extreme performance degredation. However, 727 * systems with excessive dirty pages (especially when MAP_NOSYNC is 728 * used) will die horribly with limited laundering. If the pageout 729 * daemon cannot clean enough pages in the first pass, we let it go 730 * all out in succeeding passes. 731 */ 732 if ((maxlaunder = vm_max_launder) <= 1) 733 maxlaunder = 1; 734 if (pass) 735 maxlaunder = 10000; 736 vm_page_lock_queues(); 737rescan0: 738 addl_page_shortage = addl_page_shortage_init; 739 maxscan = cnt.v_inactive_count; 740 741 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl); 742 m != NULL && maxscan-- > 0 && page_shortage > 0; 743 m = next) { 744 745 cnt.v_pdpages++; 746 747 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE) { 748 goto rescan0; 749 } 750 751 next = TAILQ_NEXT(m, pageq); 752 object = m->object; 753 754 /* 755 * skip marker pages 756 */ 757 if (m->flags & PG_MARKER) 758 continue; 759 760 /* 761 * A held page may be undergoing I/O, so skip it. 762 */ 763 if (m->hold_count) { 764 vm_pageq_requeue(m); 765 addl_page_shortage++; 766 continue; 767 } 768 /* 769 * Don't mess with busy pages, keep in the front of the 770 * queue, most likely are being paged out. 771 */ 772 if (!VM_OBJECT_TRYLOCK(object) && 773 (!vm_pageout_fallback_object_lock(m, &next) || 774 m->hold_count != 0)) { 775 VM_OBJECT_UNLOCK(object); 776 addl_page_shortage++; 777 continue; 778 } 779 if (m->busy || (m->oflags & VPO_BUSY)) { 780 VM_OBJECT_UNLOCK(object); 781 addl_page_shortage++; 782 continue; 783 } 784 785 /* 786 * If the object is not being used, we ignore previous 787 * references. 788 */ 789 if (object->ref_count == 0) { 790 vm_page_flag_clear(m, PG_REFERENCED); 791 pmap_clear_reference(m); 792 793 /* 794 * Otherwise, if the page has been referenced while in the 795 * inactive queue, we bump the "activation count" upwards, 796 * making it less likely that the page will be added back to 797 * the inactive queue prematurely again. Here we check the 798 * page tables (or emulated bits, if any), given the upper 799 * level VM system not knowing anything about existing 800 * references. 801 */ 802 } else if (((m->flags & PG_REFERENCED) == 0) && 803 (actcount = pmap_ts_referenced(m))) { 804 vm_page_activate(m); 805 VM_OBJECT_UNLOCK(object); 806 m->act_count += (actcount + ACT_ADVANCE); 807 continue; 808 } 809 810 /* 811 * If the upper level VM system knows about any page 812 * references, we activate the page. We also set the 813 * "activation count" higher than normal so that we will less 814 * likely place pages back onto the inactive queue again. 815 */ 816 if ((m->flags & PG_REFERENCED) != 0) { 817 vm_page_flag_clear(m, PG_REFERENCED); 818 actcount = pmap_ts_referenced(m); 819 vm_page_activate(m); 820 VM_OBJECT_UNLOCK(object); 821 m->act_count += (actcount + ACT_ADVANCE + 1); 822 continue; 823 } 824 825 /* 826 * If the upper level VM system doesn't know anything about 827 * the page being dirty, we have to check for it again. As 828 * far as the VM code knows, any partially dirty pages are 829 * fully dirty. 830 */ 831 if (m->dirty == 0 && !pmap_is_modified(m)) { 832 /* 833 * Avoid a race condition: Unless write access is 834 * removed from the page, another processor could 835 * modify it before all access is removed by the call 836 * to vm_page_cache() below. If vm_page_cache() finds 837 * that the page has been modified when it removes all 838 * access, it panics because it cannot cache dirty 839 * pages. In principle, we could eliminate just write 840 * access here rather than all access. In the expected 841 * case, when there are no last instant modifications 842 * to the page, removing all access will be cheaper 843 * overall. 844 */ 845 if ((m->flags & PG_WRITEABLE) != 0) 846 pmap_remove_all(m); 847 } else { 848 vm_page_dirty(m); 849 } 850 851 if (m->valid == 0) { 852 /* 853 * Invalid pages can be easily freed 854 */ 855 vm_page_free(m); 856 cnt.v_dfree++; 857 --page_shortage; 858 } else if (m->dirty == 0) { 859 /* 860 * Clean pages can be placed onto the cache queue. 861 * This effectively frees them. 862 */ 863 vm_page_cache(m); 864 --page_shortage; 865 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) { 866 /* 867 * Dirty pages need to be paged out, but flushing 868 * a page is extremely expensive verses freeing 869 * a clean page. Rather then artificially limiting 870 * the number of pages we can flush, we instead give 871 * dirty pages extra priority on the inactive queue 872 * by forcing them to be cycled through the queue 873 * twice before being flushed, after which the 874 * (now clean) page will cycle through once more 875 * before being freed. This significantly extends 876 * the thrash point for a heavily loaded machine. 877 */ 878 vm_page_flag_set(m, PG_WINATCFLS); 879 vm_pageq_requeue(m); 880 } else if (maxlaunder > 0) { 881 /* 882 * We always want to try to flush some dirty pages if 883 * we encounter them, to keep the system stable. 884 * Normally this number is small, but under extreme 885 * pressure where there are insufficient clean pages 886 * on the inactive queue, we may have to go all out. 887 */ 888 int swap_pageouts_ok; 889 struct vnode *vp = NULL; 890 struct mount *mp; 891 892 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 893 swap_pageouts_ok = 1; 894 } else { 895 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 896 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 897 vm_page_count_min()); 898 899 } 900 901 /* 902 * We don't bother paging objects that are "dead". 903 * Those objects are in a "rundown" state. 904 */ 905 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 906 VM_OBJECT_UNLOCK(object); 907 vm_pageq_requeue(m); 908 continue; 909 } 910 911 /* 912 * Following operations may unlock 913 * vm_page_queue_mtx, invalidating the 'next' 914 * pointer. To prevent an inordinate number 915 * of restarts we use our marker to remember 916 * our place. 917 * 918 */ 919 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, 920 m, &marker, pageq); 921 /* 922 * The object is already known NOT to be dead. It 923 * is possible for the vget() to block the whole 924 * pageout daemon, but the new low-memory handling 925 * code should prevent it. 926 * 927 * The previous code skipped locked vnodes and, worse, 928 * reordered pages in the queue. This results in 929 * completely non-deterministic operation and, on a 930 * busy system, can lead to extremely non-optimal 931 * pageouts. For example, it can cause clean pages 932 * to be freed and dirty pages to be moved to the end 933 * of the queue. Since dirty pages are also moved to 934 * the end of the queue once-cleaned, this gives 935 * way too large a weighting to defering the freeing 936 * of dirty pages. 937 * 938 * We can't wait forever for the vnode lock, we might 939 * deadlock due to a vn_read() getting stuck in 940 * vm_wait while holding this vnode. We skip the 941 * vnode if we can't get it in a reasonable amount 942 * of time. 943 */ 944 if (object->type == OBJT_VNODE) { 945 vp = object->handle; 946 mp = NULL; 947 if (vp->v_type == VREG && 948 vn_start_write(vp, &mp, V_NOWAIT) != 0) { 949 ++pageout_lock_miss; 950 if (object->flags & OBJ_MIGHTBEDIRTY) 951 vnodes_skipped++; 952 vp = NULL; 953 goto unlock_and_continue; 954 } 955 vm_page_unlock_queues(); 956 VI_LOCK(vp); 957 VM_OBJECT_UNLOCK(object); 958 if (vget(vp, LK_EXCLUSIVE | LK_INTERLOCK | 959 LK_TIMELOCK, curthread)) { 960 VM_OBJECT_LOCK(object); 961 vm_page_lock_queues(); 962 ++pageout_lock_miss; 963 vn_finished_write(mp); 964 if (object->flags & OBJ_MIGHTBEDIRTY) 965 vnodes_skipped++; 966 vp = NULL; 967 goto unlock_and_continue; 968 } 969 VM_OBJECT_LOCK(object); 970 vm_page_lock_queues(); 971 /* 972 * The page might have been moved to another 973 * queue during potential blocking in vget() 974 * above. The page might have been freed and 975 * reused for another vnode. The object might 976 * have been reused for another vnode. 977 */ 978 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE || 979 m->object != object || 980 object->handle != vp || 981 TAILQ_NEXT(m, pageq) != &marker) { 982 if (object->flags & OBJ_MIGHTBEDIRTY) 983 vnodes_skipped++; 984 goto unlock_and_continue; 985 } 986 987 /* 988 * The page may have been busied during the 989 * blocking in vput(); We don't move the 990 * page back onto the end of the queue so that 991 * statistics are more correct if we don't. 992 */ 993 if (m->busy || (m->oflags & VPO_BUSY)) { 994 goto unlock_and_continue; 995 } 996 997 /* 998 * If the page has become held it might 999 * be undergoing I/O, so skip it 1000 */ 1001 if (m->hold_count) { 1002 vm_pageq_requeue(m); 1003 if (object->flags & OBJ_MIGHTBEDIRTY) 1004 vnodes_skipped++; 1005 goto unlock_and_continue; 1006 } 1007 } 1008 1009 /* 1010 * If a page is dirty, then it is either being washed 1011 * (but not yet cleaned) or it is still in the 1012 * laundry. If it is still in the laundry, then we 1013 * start the cleaning operation. 1014 * 1015 * decrement page_shortage on success to account for 1016 * the (future) cleaned page. Otherwise we could wind 1017 * up laundering or cleaning too many pages. 1018 */ 1019 if (vm_pageout_clean(m) != 0) { 1020 --page_shortage; 1021 --maxlaunder; 1022 } 1023unlock_and_continue: 1024 VM_OBJECT_UNLOCK(object); 1025 if (vp) { 1026 vm_page_unlock_queues(); 1027 vput(vp); 1028 vn_finished_write(mp); 1029 vm_page_lock_queues(); 1030 } 1031 next = TAILQ_NEXT(&marker, pageq); 1032 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, 1033 &marker, pageq); 1034 continue; 1035 } 1036 VM_OBJECT_UNLOCK(object); 1037 } 1038 1039 /* 1040 * Compute the number of pages we want to try to move from the 1041 * active queue to the inactive queue. 1042 */ 1043 page_shortage = vm_paging_target() + 1044 cnt.v_inactive_target - cnt.v_inactive_count; 1045 page_shortage += addl_page_shortage; 1046 1047 /* 1048 * Scan the active queue for things we can deactivate. We nominally 1049 * track the per-page activity counter and use it to locate 1050 * deactivation candidates. 1051 */ 1052 pcount = cnt.v_active_count; 1053 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1054 1055 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) { 1056 1057 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE), 1058 ("vm_pageout_scan: page %p isn't active", m)); 1059 1060 next = TAILQ_NEXT(m, pageq); 1061 object = m->object; 1062 if ((m->flags & PG_MARKER) != 0) { 1063 m = next; 1064 continue; 1065 } 1066 if (!VM_OBJECT_TRYLOCK(object) && 1067 !vm_pageout_fallback_object_lock(m, &next)) { 1068 VM_OBJECT_UNLOCK(object); 1069 m = next; 1070 continue; 1071 } 1072 1073 /* 1074 * Don't deactivate pages that are busy. 1075 */ 1076 if ((m->busy != 0) || 1077 (m->oflags & VPO_BUSY) || 1078 (m->hold_count != 0)) { 1079 VM_OBJECT_UNLOCK(object); 1080 vm_pageq_requeue(m); 1081 m = next; 1082 continue; 1083 } 1084 1085 /* 1086 * The count for pagedaemon pages is done after checking the 1087 * page for eligibility... 1088 */ 1089 cnt.v_pdpages++; 1090 1091 /* 1092 * Check to see "how much" the page has been used. 1093 */ 1094 actcount = 0; 1095 if (object->ref_count != 0) { 1096 if (m->flags & PG_REFERENCED) { 1097 actcount += 1; 1098 } 1099 actcount += pmap_ts_referenced(m); 1100 if (actcount) { 1101 m->act_count += ACT_ADVANCE + actcount; 1102 if (m->act_count > ACT_MAX) 1103 m->act_count = ACT_MAX; 1104 } 1105 } 1106 1107 /* 1108 * Since we have "tested" this bit, we need to clear it now. 1109 */ 1110 vm_page_flag_clear(m, PG_REFERENCED); 1111 1112 /* 1113 * Only if an object is currently being used, do we use the 1114 * page activation count stats. 1115 */ 1116 if (actcount && (object->ref_count != 0)) { 1117 vm_pageq_requeue(m); 1118 } else { 1119 m->act_count -= min(m->act_count, ACT_DECLINE); 1120 if (vm_pageout_algorithm || 1121 object->ref_count == 0 || 1122 m->act_count == 0) { 1123 page_shortage--; 1124 if (object->ref_count == 0) { 1125 pmap_remove_all(m); 1126 if (m->dirty == 0) 1127 vm_page_cache(m); 1128 else 1129 vm_page_deactivate(m); 1130 } else { 1131 vm_page_deactivate(m); 1132 } 1133 } else { 1134 vm_pageq_requeue(m); 1135 } 1136 } 1137 VM_OBJECT_UNLOCK(object); 1138 m = next; 1139 } 1140 1141 /* 1142 * We try to maintain some *really* free pages, this allows interrupt 1143 * code to be guaranteed space. Since both cache and free queues 1144 * are considered basically 'free', moving pages from cache to free 1145 * does not effect other calculations. 1146 */ 1147 while (cnt.v_free_count < cnt.v_free_reserved) { 1148 TAILQ_FOREACH(m, &vm_page_queues[PQ_CACHE].pl, pageq) { 1149 KASSERT(m->dirty == 0, 1150 ("Found dirty cache page %p", m)); 1151 KASSERT(!pmap_page_is_mapped(m), 1152 ("Found mapped cache page %p", m)); 1153 KASSERT((m->flags & PG_UNMANAGED) == 0, 1154 ("Found unmanaged cache page %p", m)); 1155 KASSERT(m->wire_count == 0, 1156 ("Found wired cache page %p", m)); 1157 if (m->hold_count == 0 && VM_OBJECT_TRYLOCK(object = 1158 m->object)) { 1159 KASSERT((m->oflags & VPO_BUSY) == 0 && 1160 m->busy == 0, ("Found busy cache page %p", 1161 m)); 1162 vm_page_free(m); 1163 VM_OBJECT_UNLOCK(object); 1164 cnt.v_dfree++; 1165 break; 1166 } 1167 } 1168 if (m == NULL) 1169 break; 1170 } 1171 vm_page_unlock_queues(); 1172#if !defined(NO_SWAPPING) 1173 /* 1174 * Idle process swapout -- run once per second. 1175 */ 1176 if (vm_swap_idle_enabled) { 1177 static long lsec; 1178 if (time_second != lsec) { 1179 vm_pageout_req_swapout |= VM_SWAP_IDLE; 1180 vm_req_vmdaemon(); 1181 lsec = time_second; 1182 } 1183 } 1184#endif 1185 1186 /* 1187 * If we didn't get enough free pages, and we have skipped a vnode 1188 * in a writeable object, wakeup the sync daemon. And kick swapout 1189 * if we did not get enough free pages. 1190 */ 1191 if (vm_paging_target() > 0) { 1192 if (vnodes_skipped && vm_page_count_min()) 1193 (void) speedup_syncer(); 1194#if !defined(NO_SWAPPING) 1195 if (vm_swap_enabled && vm_page_count_target()) { 1196 vm_req_vmdaemon(); 1197 vm_pageout_req_swapout |= VM_SWAP_NORMAL; 1198 } 1199#endif 1200 } 1201 1202 /* 1203 * If we are critically low on one of RAM or swap and low on 1204 * the other, kill the largest process. However, we avoid 1205 * doing this on the first pass in order to give ourselves a 1206 * chance to flush out dirty vnode-backed pages and to allow 1207 * active pages to be moved to the inactive queue and reclaimed. 1208 * 1209 * We keep the process bigproc locked once we find it to keep anyone 1210 * from messing with it; however, there is a possibility of 1211 * deadlock if process B is bigproc and one of it's child processes 1212 * attempts to propagate a signal to B while we are waiting for A's 1213 * lock while walking this list. To avoid this, we don't block on 1214 * the process lock but just skip a process if it is already locked. 1215 */ 1216 if (pass != 0 && 1217 ((swap_pager_avail < 64 && vm_page_count_min()) || 1218 (swap_pager_full && vm_paging_target() > 0))) { 1219 bigproc = NULL; 1220 bigsize = 0; 1221 sx_slock(&allproc_lock); 1222 FOREACH_PROC_IN_SYSTEM(p) { 1223 int breakout; 1224 1225 if (PROC_TRYLOCK(p) == 0) 1226 continue; 1227 /* 1228 * If this is a system or protected process, skip it. 1229 */ 1230 if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) || 1231 (p->p_flag & P_PROTECTED) || 1232 ((p->p_pid < 48) && (swap_pager_avail != 0))) { 1233 PROC_UNLOCK(p); 1234 continue; 1235 } 1236 /* 1237 * If the process is in a non-running type state, 1238 * don't touch it. Check all the threads individually. 1239 */ 1240 PROC_SLOCK(p); 1241 breakout = 0; 1242 FOREACH_THREAD_IN_PROC(p, td) { 1243 thread_lock(td); 1244 if (!TD_ON_RUNQ(td) && 1245 !TD_IS_RUNNING(td) && 1246 !TD_IS_SLEEPING(td)) { 1247 thread_unlock(td); 1248 breakout = 1; 1249 break; 1250 } 1251 thread_unlock(td); 1252 } 1253 PROC_SUNLOCK(p); 1254 if (breakout) { 1255 PROC_UNLOCK(p); 1256 continue; 1257 } 1258 /* 1259 * get the process size 1260 */ 1261 if (!vm_map_trylock_read(&p->p_vmspace->vm_map)) { 1262 PROC_UNLOCK(p); 1263 continue; 1264 } 1265 size = vmspace_swap_count(p->p_vmspace); 1266 vm_map_unlock_read(&p->p_vmspace->vm_map); 1267 size += vmspace_resident_count(p->p_vmspace); 1268 /* 1269 * if the this process is bigger than the biggest one 1270 * remember it. 1271 */ 1272 if (size > bigsize) { 1273 if (bigproc != NULL) 1274 PROC_UNLOCK(bigproc); 1275 bigproc = p; 1276 bigsize = size; 1277 } else 1278 PROC_UNLOCK(p); 1279 } 1280 sx_sunlock(&allproc_lock); 1281 if (bigproc != NULL) { 1282 killproc(bigproc, "out of swap space"); 1283 PROC_SLOCK(bigproc); 1284 sched_nice(bigproc, PRIO_MIN); 1285 PROC_SUNLOCK(bigproc); 1286 PROC_UNLOCK(bigproc); 1287 wakeup(&cnt.v_free_count); 1288 } 1289 } 1290 mtx_unlock(&Giant); 1291} 1292 1293/* 1294 * This routine tries to maintain the pseudo LRU active queue, 1295 * so that during long periods of time where there is no paging, 1296 * that some statistic accumulation still occurs. This code 1297 * helps the situation where paging just starts to occur. 1298 */ 1299static void 1300vm_pageout_page_stats() 1301{ 1302 vm_object_t object; 1303 vm_page_t m,next; 1304 int pcount,tpcount; /* Number of pages to check */ 1305 static int fullintervalcount = 0; 1306 int page_shortage; 1307 1308 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1309 page_shortage = 1310 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) - 1311 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count); 1312 1313 if (page_shortage <= 0) 1314 return; 1315 1316 pcount = cnt.v_active_count; 1317 fullintervalcount += vm_pageout_stats_interval; 1318 if (fullintervalcount < vm_pageout_full_stats_interval) { 1319 tpcount = (vm_pageout_stats_max * cnt.v_active_count) / cnt.v_page_count; 1320 if (pcount > tpcount) 1321 pcount = tpcount; 1322 } else { 1323 fullintervalcount = 0; 1324 } 1325 1326 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1327 while ((m != NULL) && (pcount-- > 0)) { 1328 int actcount; 1329 1330 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE), 1331 ("vm_pageout_page_stats: page %p isn't active", m)); 1332 1333 next = TAILQ_NEXT(m, pageq); 1334 object = m->object; 1335 1336 if ((m->flags & PG_MARKER) != 0) { 1337 m = next; 1338 continue; 1339 } 1340 if (!VM_OBJECT_TRYLOCK(object) && 1341 !vm_pageout_fallback_object_lock(m, &next)) { 1342 VM_OBJECT_UNLOCK(object); 1343 m = next; 1344 continue; 1345 } 1346 1347 /* 1348 * Don't deactivate pages that are busy. 1349 */ 1350 if ((m->busy != 0) || 1351 (m->oflags & VPO_BUSY) || 1352 (m->hold_count != 0)) { 1353 VM_OBJECT_UNLOCK(object); 1354 vm_pageq_requeue(m); 1355 m = next; 1356 continue; 1357 } 1358 1359 actcount = 0; 1360 if (m->flags & PG_REFERENCED) { 1361 vm_page_flag_clear(m, PG_REFERENCED); 1362 actcount += 1; 1363 } 1364 1365 actcount += pmap_ts_referenced(m); 1366 if (actcount) { 1367 m->act_count += ACT_ADVANCE + actcount; 1368 if (m->act_count > ACT_MAX) 1369 m->act_count = ACT_MAX; 1370 vm_pageq_requeue(m); 1371 } else { 1372 if (m->act_count == 0) { 1373 /* 1374 * We turn off page access, so that we have 1375 * more accurate RSS stats. We don't do this 1376 * in the normal page deactivation when the 1377 * system is loaded VM wise, because the 1378 * cost of the large number of page protect 1379 * operations would be higher than the value 1380 * of doing the operation. 1381 */ 1382 pmap_remove_all(m); 1383 vm_page_deactivate(m); 1384 } else { 1385 m->act_count -= min(m->act_count, ACT_DECLINE); 1386 vm_pageq_requeue(m); 1387 } 1388 } 1389 VM_OBJECT_UNLOCK(object); 1390 m = next; 1391 } 1392} 1393 1394/* 1395 * vm_pageout is the high level pageout daemon. 1396 */ 1397static void 1398vm_pageout() 1399{ 1400 int error, pass; 1401 1402 /* 1403 * Initialize some paging parameters. 1404 */ 1405 cnt.v_interrupt_free_min = 2; 1406 if (cnt.v_page_count < 2000) 1407 vm_pageout_page_count = 8; 1408 1409 /* 1410 * v_free_reserved needs to include enough for the largest 1411 * swap pager structures plus enough for any pv_entry structs 1412 * when paging. 1413 */ 1414 if (cnt.v_page_count > 1024) 1415 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200; 1416 else 1417 cnt.v_free_min = 4; 1418 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1419 cnt.v_interrupt_free_min; 1420 cnt.v_free_reserved = vm_pageout_page_count + 1421 cnt.v_pageout_free_min + (cnt.v_page_count / 768); 1422 cnt.v_free_severe = cnt.v_free_min / 2; 1423 cnt.v_free_min += cnt.v_free_reserved; 1424 cnt.v_free_severe += cnt.v_free_reserved; 1425 1426 /* 1427 * v_free_target and v_cache_min control pageout hysteresis. Note 1428 * that these are more a measure of the VM cache queue hysteresis 1429 * then the VM free queue. Specifically, v_free_target is the 1430 * high water mark (free+cache pages). 1431 * 1432 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the 1433 * low water mark, while v_free_min is the stop. v_cache_min must 1434 * be big enough to handle memory needs while the pageout daemon 1435 * is signalled and run to free more pages. 1436 */ 1437 if (cnt.v_free_count > 6144) 1438 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved; 1439 else 1440 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved; 1441 1442 if (cnt.v_free_count > 2048) { 1443 cnt.v_cache_min = cnt.v_free_target; 1444 cnt.v_cache_max = 2 * cnt.v_cache_min; 1445 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2; 1446 } else { 1447 cnt.v_cache_min = 0; 1448 cnt.v_cache_max = 0; 1449 cnt.v_inactive_target = cnt.v_free_count / 4; 1450 } 1451 if (cnt.v_inactive_target > cnt.v_free_count / 3) 1452 cnt.v_inactive_target = cnt.v_free_count / 3; 1453 1454 /* XXX does not really belong here */ 1455 if (vm_page_max_wired == 0) 1456 vm_page_max_wired = cnt.v_free_count / 3; 1457 1458 if (vm_pageout_stats_max == 0) 1459 vm_pageout_stats_max = cnt.v_free_target; 1460 1461 /* 1462 * Set interval in seconds for stats scan. 1463 */ 1464 if (vm_pageout_stats_interval == 0) 1465 vm_pageout_stats_interval = 5; 1466 if (vm_pageout_full_stats_interval == 0) 1467 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4; 1468 1469 swap_pager_swap_init(); 1470 pass = 0; 1471 /* 1472 * The pageout daemon is never done, so loop forever. 1473 */ 1474 while (TRUE) { 1475 /* 1476 * If we have enough free memory, wakeup waiters. Do 1477 * not clear vm_pages_needed until we reach our target, 1478 * otherwise we may be woken up over and over again and 1479 * waste a lot of cpu. 1480 */ 1481 mtx_lock(&vm_page_queue_free_mtx); 1482 if (vm_pages_needed && !vm_page_count_min()) { 1483 if (!vm_paging_needed()) 1484 vm_pages_needed = 0; 1485 wakeup(&cnt.v_free_count); 1486 } 1487 if (vm_pages_needed) { 1488 /* 1489 * Still not done, take a second pass without waiting 1490 * (unlimited dirty cleaning), otherwise sleep a bit 1491 * and try again. 1492 */ 1493 ++pass; 1494 if (pass > 1) 1495 msleep(&vm_pages_needed, 1496 &vm_page_queue_free_mtx, PVM, "psleep", 1497 hz / 2); 1498 } else { 1499 /* 1500 * Good enough, sleep & handle stats. Prime the pass 1501 * for the next run. 1502 */ 1503 if (pass > 1) 1504 pass = 1; 1505 else 1506 pass = 0; 1507 error = msleep(&vm_pages_needed, 1508 &vm_page_queue_free_mtx, PVM, "psleep", 1509 vm_pageout_stats_interval * hz); 1510 if (error && !vm_pages_needed) { 1511 mtx_unlock(&vm_page_queue_free_mtx); 1512 pass = 0; 1513 vm_page_lock_queues(); 1514 vm_pageout_page_stats(); 1515 vm_page_unlock_queues(); 1516 continue; 1517 } 1518 } 1519 if (vm_pages_needed) 1520 cnt.v_pdwakeups++; 1521 mtx_unlock(&vm_page_queue_free_mtx); 1522 vm_pageout_scan(pass); 1523 } 1524} 1525 1526/* 1527 * Unless the free page queue lock is held by the caller, this function 1528 * should be regarded as advisory. Specifically, the caller should 1529 * not msleep() on &cnt.v_free_count following this function unless 1530 * the free page queue lock is held until the msleep() is performed. 1531 */ 1532void 1533pagedaemon_wakeup() 1534{ 1535 1536 if (!vm_pages_needed && curthread->td_proc != pageproc) { 1537 vm_pages_needed = 1; 1538 wakeup(&vm_pages_needed); 1539 } 1540} 1541 1542#if !defined(NO_SWAPPING) 1543static void 1544vm_req_vmdaemon() 1545{ 1546 static int lastrun = 0; 1547 1548 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1549 wakeup(&vm_daemon_needed); 1550 lastrun = ticks; 1551 } 1552} 1553 1554static void 1555vm_daemon() 1556{ 1557 struct rlimit rsslim; 1558 struct proc *p; 1559 struct thread *td; 1560 int breakout; 1561 1562 mtx_lock(&Giant); 1563 while (TRUE) { 1564 tsleep(&vm_daemon_needed, PPAUSE, "psleep", 0); 1565 if (vm_pageout_req_swapout) { 1566 swapout_procs(vm_pageout_req_swapout); 1567 vm_pageout_req_swapout = 0; 1568 } 1569 /* 1570 * scan the processes for exceeding their rlimits or if 1571 * process is swapped out -- deactivate pages 1572 */ 1573 sx_slock(&allproc_lock); 1574 FOREACH_PROC_IN_SYSTEM(p) { 1575 vm_pindex_t limit, size; 1576 1577 /* 1578 * if this is a system process or if we have already 1579 * looked at this process, skip it. 1580 */ 1581 PROC_LOCK(p); 1582 if (p->p_flag & (P_SYSTEM | P_WEXIT)) { 1583 PROC_UNLOCK(p); 1584 continue; 1585 } 1586 /* 1587 * if the process is in a non-running type state, 1588 * don't touch it. 1589 */ 1590 PROC_SLOCK(p); 1591 breakout = 0; 1592 FOREACH_THREAD_IN_PROC(p, td) { 1593 thread_lock(td); 1594 if (!TD_ON_RUNQ(td) && 1595 !TD_IS_RUNNING(td) && 1596 !TD_IS_SLEEPING(td)) { 1597 thread_unlock(td); 1598 breakout = 1; 1599 break; 1600 } 1601 thread_unlock(td); 1602 } 1603 PROC_SUNLOCK(p); 1604 if (breakout) { 1605 PROC_UNLOCK(p); 1606 continue; 1607 } 1608 /* 1609 * get a limit 1610 */ 1611 lim_rlimit(p, RLIMIT_RSS, &rsslim); 1612 limit = OFF_TO_IDX( 1613 qmin(rsslim.rlim_cur, rsslim.rlim_max)); 1614 1615 /* 1616 * let processes that are swapped out really be 1617 * swapped out set the limit to nothing (will force a 1618 * swap-out.) 1619 */ 1620 if ((p->p_sflag & PS_INMEM) == 0) 1621 limit = 0; /* XXX */ 1622 PROC_UNLOCK(p); 1623 1624 size = vmspace_resident_count(p->p_vmspace); 1625 if (limit >= 0 && size >= limit) { 1626 vm_pageout_map_deactivate_pages( 1627 &p->p_vmspace->vm_map, limit); 1628 } 1629 } 1630 sx_sunlock(&allproc_lock); 1631 } 1632} 1633#endif /* !defined(NO_SWAPPING) */ 1634