vm_pageout.c revision 207452
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 207452 2010-04-30 22:31:37Z kmacy $"); 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/mount.h> 89#include <sys/resourcevar.h> 90#include <sys/sched.h> 91#include <sys/signalvar.h> 92#include <sys/vnode.h> 93#include <sys/vmmeter.h> 94#include <sys/sx.h> 95#include <sys/sysctl.h> 96 97#include <vm/vm.h> 98#include <vm/vm_param.h> 99#include <vm/vm_object.h> 100#include <vm/vm_page.h> 101#include <vm/vm_map.h> 102#include <vm/vm_pageout.h> 103#include <vm/vm_pager.h> 104#include <vm/swap_pager.h> 105#include <vm/vm_extern.h> 106#include <vm/uma.h> 107 108/* 109 * System initialization 110 */ 111 112/* the kernel process "vm_pageout"*/ 113static void vm_pageout(void); 114static int vm_pageout_clean(vm_page_t); 115static void vm_pageout_scan(int pass); 116 117struct proc *pageproc; 118 119static struct kproc_desc page_kp = { 120 "pagedaemon", 121 vm_pageout, 122 &pageproc 123}; 124SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, 125 &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; 148static struct mtx vm_daemon_mtx; 149/* Allow for use by vm_pageout before vm_daemon is initialized. */ 150MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF); 151#endif 152static int vm_max_launder = 32; 153static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0; 154static int vm_pageout_full_stats_interval = 0; 155static int vm_pageout_algorithm=0; 156static int defer_swap_pageouts=0; 157static int disable_swap_pageouts=0; 158 159#if defined(NO_SWAPPING) 160static int vm_swap_enabled=0; 161static int vm_swap_idle_enabled=0; 162#else 163static int vm_swap_enabled=1; 164static int vm_swap_idle_enabled=0; 165#endif 166 167SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm, 168 CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt"); 169 170SYSCTL_INT(_vm, OID_AUTO, max_launder, 171 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout"); 172 173SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max, 174 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length"); 175 176SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval, 177 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan"); 178 179SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval, 180 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan"); 181 182#if defined(NO_SWAPPING) 183SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 184 CTLFLAG_RD, &vm_swap_enabled, 0, "Enable entire process swapout"); 185SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 186 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 187#else 188SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 189 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout"); 190SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 191 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 192#endif 193 194SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts, 195 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem"); 196 197SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts, 198 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages"); 199 200static int pageout_lock_miss; 201SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss, 202 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout"); 203 204#define VM_PAGEOUT_PAGE_COUNT 16 205int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT; 206 207int vm_page_max_wired; /* XXX max # of wired pages system-wide */ 208SYSCTL_INT(_vm, OID_AUTO, max_wired, 209 CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count"); 210 211#if !defined(NO_SWAPPING) 212static void vm_pageout_map_deactivate_pages(vm_map_t, long); 213static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long); 214static void vm_req_vmdaemon(int req); 215#endif 216static void vm_pageout_page_stats(void); 217 218/* 219 * vm_pageout_fallback_object_lock: 220 * 221 * Lock vm object currently associated with `m'. VM_OBJECT_TRYLOCK is 222 * known to have failed and page queue must be either PQ_ACTIVE or 223 * PQ_INACTIVE. To avoid lock order violation, unlock the page queues 224 * while locking the vm object. Use marker page to detect page queue 225 * changes and maintain notion of next page on page queue. Return 226 * TRUE if no changes were detected, FALSE otherwise. vm object is 227 * locked on return. 228 * 229 * This function depends on both the lock portion of struct vm_object 230 * and normal struct vm_page being type stable. 231 */ 232boolean_t 233vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next) 234{ 235 struct vm_page marker; 236 boolean_t unchanged; 237 u_short queue; 238 vm_object_t object; 239 240 /* 241 * Initialize our marker 242 */ 243 bzero(&marker, sizeof(marker)); 244 marker.flags = PG_FICTITIOUS | PG_MARKER; 245 marker.oflags = VPO_BUSY; 246 marker.queue = m->queue; 247 marker.wire_count = 1; 248 249 queue = m->queue; 250 object = m->object; 251 252 TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl, 253 m, &marker, pageq); 254 vm_page_unlock_queues(); 255 vm_page_unlock(m); 256 VM_OBJECT_LOCK(object); 257 vm_page_lock(m); 258 vm_page_lock_queues(); 259 260 /* Page queue might have changed. */ 261 *next = TAILQ_NEXT(&marker, pageq); 262 unchanged = (m->queue == queue && 263 m->object == object && 264 &marker == TAILQ_NEXT(m, pageq)); 265 TAILQ_REMOVE(&vm_page_queues[queue].pl, 266 &marker, pageq); 267 return (unchanged); 268} 269 270/* 271 * vm_pageout_clean: 272 * 273 * Clean the page and remove it from the laundry. 274 * 275 * We set the busy bit to cause potential page faults on this page to 276 * block. Note the careful timing, however, the busy bit isn't set till 277 * late and we cannot do anything that will mess with the page. 278 */ 279static int 280vm_pageout_clean(vm_page_t m) 281{ 282 vm_object_t object; 283 vm_page_t mc[2*vm_pageout_page_count]; 284 int pageout_count; 285 int ib, is, page_base; 286 vm_pindex_t pindex = m->pindex; 287 288 vm_page_lock_assert(m, MA_NOTOWNED); 289 vm_page_lock(m); 290 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 291 292 /* 293 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP 294 * with the new swapper, but we could have serious problems paging 295 * out other object types if there is insufficient memory. 296 * 297 * Unfortunately, checking free memory here is far too late, so the 298 * check has been moved up a procedural level. 299 */ 300 301 /* 302 * Can't clean the page if it's busy or held. 303 */ 304 if ((m->hold_count != 0) || 305 ((m->busy != 0) || (m->oflags & VPO_BUSY))) { 306 vm_page_unlock(m); 307 return 0; 308 } 309 310 mc[vm_pageout_page_count] = m; 311 pageout_count = 1; 312 page_base = vm_pageout_page_count; 313 ib = 1; 314 is = 1; 315 316 /* 317 * Scan object for clusterable pages. 318 * 319 * We can cluster ONLY if: ->> the page is NOT 320 * clean, wired, busy, held, or mapped into a 321 * buffer, and one of the following: 322 * 1) The page is inactive, or a seldom used 323 * active page. 324 * -or- 325 * 2) we force the issue. 326 * 327 * During heavy mmap/modification loads the pageout 328 * daemon can really fragment the underlying file 329 * due to flushing pages out of order and not trying 330 * align the clusters (which leave sporatic out-of-order 331 * holes). To solve this problem we do the reverse scan 332 * first and attempt to align our cluster, then do a 333 * forward scan if room remains. 334 */ 335 object = m->object; 336more: 337 while (ib && pageout_count < vm_pageout_page_count) { 338 vm_page_t p; 339 340 if (ib > pindex) { 341 ib = 0; 342 break; 343 } 344 345 if ((p = vm_page_lookup(object, pindex - ib)) == NULL) { 346 ib = 0; 347 break; 348 } 349 if ((p->oflags & VPO_BUSY) || p->busy) { 350 ib = 0; 351 break; 352 } 353 vm_page_lock(p); 354 vm_page_lock_queues(); 355 vm_page_test_dirty(p); 356 if (p->dirty == 0 || 357 p->queue != PQ_INACTIVE || 358 p->hold_count != 0) { /* may be undergoing I/O */ 359 vm_page_unlock(p); 360 vm_page_unlock_queues(); 361 ib = 0; 362 break; 363 } 364 vm_page_unlock_queues(); 365 vm_page_unlock(p); 366 mc[--page_base] = p; 367 ++pageout_count; 368 ++ib; 369 /* 370 * alignment boundry, stop here and switch directions. Do 371 * not clear ib. 372 */ 373 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0) 374 break; 375 } 376 377 while (pageout_count < vm_pageout_page_count && 378 pindex + is < object->size) { 379 vm_page_t p; 380 381 if ((p = vm_page_lookup(object, pindex + is)) == NULL) 382 break; 383 if ((p->oflags & VPO_BUSY) || p->busy) { 384 break; 385 } 386 vm_page_lock(p); 387 vm_page_lock_queues(); 388 vm_page_test_dirty(p); 389 if (p->dirty == 0 || 390 p->queue != PQ_INACTIVE || 391 p->hold_count != 0) { /* may be undergoing I/O */ 392 vm_page_unlock_queues(); 393 vm_page_unlock(p); 394 break; 395 } 396 vm_page_unlock_queues(); 397 vm_page_unlock(p); 398 mc[page_base + pageout_count] = p; 399 ++pageout_count; 400 ++is; 401 } 402 403 /* 404 * If we exhausted our forward scan, continue with the reverse scan 405 * when possible, even past a page boundry. This catches boundry 406 * conditions. 407 */ 408 if (ib && pageout_count < vm_pageout_page_count) 409 goto more; 410 411 vm_page_unlock(m); 412 /* 413 * we allow reads during pageouts... 414 */ 415 return (vm_pageout_flush(&mc[page_base], pageout_count, 0)); 416} 417 418/* 419 * vm_pageout_flush() - launder the given pages 420 * 421 * The given pages are laundered. Note that we setup for the start of 422 * I/O ( i.e. busy the page ), mark it read-only, and bump the object 423 * reference count all in here rather then in the parent. If we want 424 * the parent to do more sophisticated things we may have to change 425 * the ordering. 426 */ 427int 428vm_pageout_flush(vm_page_t *mc, int count, int flags) 429{ 430 vm_object_t object = mc[0]->object; 431 int pageout_status[count]; 432 int numpagedout = 0; 433 int i; 434 435 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 436 mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED); 437 438 /* 439 * Initiate I/O. Bump the vm_page_t->busy counter and 440 * mark the pages read-only. 441 * 442 * We do not have to fixup the clean/dirty bits here... we can 443 * allow the pager to do it after the I/O completes. 444 * 445 * NOTE! mc[i]->dirty may be partial or fragmented due to an 446 * edge case with file fragments. 447 */ 448 for (i = 0; i < count; i++) { 449 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, 450 ("vm_pageout_flush: partially invalid page %p index %d/%d", 451 mc[i], i, count)); 452 vm_page_io_start(mc[i]); 453 vm_page_lock(mc[i]); 454 vm_page_lock_queues(); 455 pmap_remove_write(mc[i]); 456 vm_page_unlock(mc[i]); 457 vm_page_unlock_queues(); 458 } 459 vm_object_pip_add(object, count); 460 461 vm_pager_put_pages(object, mc, count, flags, pageout_status); 462 463 for (i = 0; i < count; i++) { 464 vm_page_t mt = mc[i]; 465 466 vm_page_lock(mt); 467 vm_page_lock_queues(); 468 KASSERT(pageout_status[i] == VM_PAGER_PEND || 469 (mt->flags & PG_WRITEABLE) == 0, 470 ("vm_pageout_flush: page %p is not write protected", mt)); 471 switch (pageout_status[i]) { 472 case VM_PAGER_OK: 473 case VM_PAGER_PEND: 474 numpagedout++; 475 break; 476 case VM_PAGER_BAD: 477 /* 478 * Page outside of range of object. Right now we 479 * essentially lose the changes by pretending it 480 * worked. 481 */ 482 vm_page_undirty(mt); 483 break; 484 case VM_PAGER_ERROR: 485 case VM_PAGER_FAIL: 486 /* 487 * If page couldn't be paged out, then reactivate the 488 * page so it doesn't clog the inactive list. (We 489 * will try paging out it again later). 490 */ 491 vm_page_activate(mt); 492 break; 493 case VM_PAGER_AGAIN: 494 break; 495 } 496 497 /* 498 * If the operation is still going, leave the page busy to 499 * block all other accesses. Also, leave the paging in 500 * progress indicator set so that we don't attempt an object 501 * collapse. 502 */ 503 if (pageout_status[i] != VM_PAGER_PEND) { 504 vm_object_pip_wakeup(object); 505 vm_page_io_finish(mt); 506 if (vm_page_count_severe()) 507 vm_page_try_to_cache(mt); 508 } 509 vm_page_unlock_queues(); 510 vm_page_unlock(mt); 511 } 512 return numpagedout; 513} 514 515#if !defined(NO_SWAPPING) 516/* 517 * vm_pageout_object_deactivate_pages 518 * 519 * deactivate enough pages to satisfy the inactive target 520 * requirements or if vm_page_proc_limit is set, then 521 * deactivate all of the pages in the object and its 522 * backing_objects. 523 * 524 * The object and map must be locked. 525 */ 526static void 527vm_pageout_object_deactivate_pages(pmap, first_object, desired) 528 pmap_t pmap; 529 vm_object_t first_object; 530 long desired; 531{ 532 vm_object_t backing_object, object; 533 vm_page_t p, next; 534 int actcount, remove_mode; 535 536 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED); 537 if (first_object->type == OBJT_DEVICE || 538 first_object->type == OBJT_SG) 539 return; 540 for (object = first_object;; object = backing_object) { 541 if (pmap_resident_count(pmap) <= desired) 542 goto unlock_return; 543 if (object->type == OBJT_PHYS || object->paging_in_progress) 544 goto unlock_return; 545 546 remove_mode = 0; 547 if (object->shadow_count > 1) 548 remove_mode = 1; 549 /* 550 * scan the objects entire memory queue 551 */ 552 p = TAILQ_FIRST(&object->memq); 553 while (p != NULL) { 554 vm_page_lock(p); 555 vm_page_lock_queues(); 556 if (pmap_resident_count(pmap) <= desired) { 557 vm_page_unlock_queues(); 558 vm_page_unlock(p); 559 goto unlock_return; 560 } 561 next = TAILQ_NEXT(p, listq); 562 cnt.v_pdpages++; 563 if (p->wire_count != 0 || 564 p->hold_count != 0 || 565 p->busy != 0 || 566 (p->oflags & VPO_BUSY) || 567 !pmap_page_exists_quick(pmap, p)) { 568 vm_page_unlock_queues(); 569 vm_page_unlock(p); 570 p = next; 571 continue; 572 } 573 actcount = pmap_ts_referenced(p); 574 if (actcount) { 575 vm_page_flag_set(p, PG_REFERENCED); 576 } else if (p->flags & PG_REFERENCED) { 577 actcount = 1; 578 } 579 if ((p->queue != PQ_ACTIVE) && 580 (p->flags & PG_REFERENCED)) { 581 vm_page_activate(p); 582 p->act_count += actcount; 583 vm_page_flag_clear(p, PG_REFERENCED); 584 } else if (p->queue == PQ_ACTIVE) { 585 if ((p->flags & PG_REFERENCED) == 0) { 586 p->act_count -= min(p->act_count, ACT_DECLINE); 587 if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) { 588 pmap_remove_all(p); 589 vm_page_deactivate(p); 590 } else { 591 vm_page_requeue(p); 592 } 593 } else { 594 vm_page_activate(p); 595 vm_page_flag_clear(p, PG_REFERENCED); 596 if (p->act_count < (ACT_MAX - ACT_ADVANCE)) 597 p->act_count += ACT_ADVANCE; 598 vm_page_requeue(p); 599 } 600 } else if (p->queue == PQ_INACTIVE) { 601 pmap_remove_all(p); 602 } 603 vm_page_unlock_queues(); 604 vm_page_unlock(p); 605 p = next; 606 } 607 if ((backing_object = object->backing_object) == NULL) 608 goto unlock_return; 609 VM_OBJECT_LOCK(backing_object); 610 if (object != first_object) 611 VM_OBJECT_UNLOCK(object); 612 } 613unlock_return: 614 if (object != first_object) 615 VM_OBJECT_UNLOCK(object); 616} 617 618/* 619 * deactivate some number of pages in a map, try to do it fairly, but 620 * that is really hard to do. 621 */ 622static void 623vm_pageout_map_deactivate_pages(map, desired) 624 vm_map_t map; 625 long desired; 626{ 627 vm_map_entry_t tmpe; 628 vm_object_t obj, bigobj; 629 int nothingwired; 630 631 if (!vm_map_trylock(map)) 632 return; 633 634 bigobj = NULL; 635 nothingwired = TRUE; 636 637 /* 638 * first, search out the biggest object, and try to free pages from 639 * that. 640 */ 641 tmpe = map->header.next; 642 while (tmpe != &map->header) { 643 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 644 obj = tmpe->object.vm_object; 645 if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) { 646 if (obj->shadow_count <= 1 && 647 (bigobj == NULL || 648 bigobj->resident_page_count < obj->resident_page_count)) { 649 if (bigobj != NULL) 650 VM_OBJECT_UNLOCK(bigobj); 651 bigobj = obj; 652 } else 653 VM_OBJECT_UNLOCK(obj); 654 } 655 } 656 if (tmpe->wired_count > 0) 657 nothingwired = FALSE; 658 tmpe = tmpe->next; 659 } 660 661 if (bigobj != NULL) { 662 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired); 663 VM_OBJECT_UNLOCK(bigobj); 664 } 665 /* 666 * Next, hunt around for other pages to deactivate. We actually 667 * do this search sort of wrong -- .text first is not the best idea. 668 */ 669 tmpe = map->header.next; 670 while (tmpe != &map->header) { 671 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 672 break; 673 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 674 obj = tmpe->object.vm_object; 675 if (obj != NULL) { 676 VM_OBJECT_LOCK(obj); 677 vm_pageout_object_deactivate_pages(map->pmap, obj, desired); 678 VM_OBJECT_UNLOCK(obj); 679 } 680 } 681 tmpe = tmpe->next; 682 } 683 684 /* 685 * Remove all mappings if a process is swapped out, this will free page 686 * table pages. 687 */ 688 if (desired == 0 && nothingwired) { 689 pmap_remove(vm_map_pmap(map), vm_map_min(map), 690 vm_map_max(map)); 691 } 692 vm_map_unlock(map); 693} 694#endif /* !defined(NO_SWAPPING) */ 695 696/* 697 * vm_pageout_scan does the dirty work for the pageout daemon. 698 */ 699static void 700vm_pageout_scan(int pass) 701{ 702 vm_page_t m, next; 703 struct vm_page marker; 704 int page_shortage, maxscan, pcount; 705 int addl_page_shortage, addl_page_shortage_init; 706 vm_object_t object; 707 int actcount; 708 int vnodes_skipped = 0; 709 int maxlaunder; 710 711 /* 712 * Decrease registered cache sizes. 713 */ 714 EVENTHANDLER_INVOKE(vm_lowmem, 0); 715 /* 716 * We do this explicitly after the caches have been drained above. 717 */ 718 uma_reclaim(); 719 720 addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit); 721 722 /* 723 * Calculate the number of pages we want to either free or move 724 * to the cache. 725 */ 726 page_shortage = vm_paging_target() + addl_page_shortage_init; 727 728 /* 729 * Initialize our marker 730 */ 731 bzero(&marker, sizeof(marker)); 732 marker.flags = PG_FICTITIOUS | PG_MARKER; 733 marker.oflags = VPO_BUSY; 734 marker.queue = PQ_INACTIVE; 735 marker.wire_count = 1; 736 737 /* 738 * Start scanning the inactive queue for pages we can move to the 739 * cache or free. The scan will stop when the target is reached or 740 * we have scanned the entire inactive queue. Note that m->act_count 741 * is not used to form decisions for the inactive queue, only for the 742 * active queue. 743 * 744 * maxlaunder limits the number of dirty pages we flush per scan. 745 * For most systems a smaller value (16 or 32) is more robust under 746 * extreme memory and disk pressure because any unnecessary writes 747 * to disk can result in extreme performance degredation. However, 748 * systems with excessive dirty pages (especially when MAP_NOSYNC is 749 * used) will die horribly with limited laundering. If the pageout 750 * daemon cannot clean enough pages in the first pass, we let it go 751 * all out in succeeding passes. 752 */ 753 if ((maxlaunder = vm_max_launder) <= 1) 754 maxlaunder = 1; 755 if (pass) 756 maxlaunder = 10000; 757 vm_page_lock_queues(); 758rescan0: 759 addl_page_shortage = addl_page_shortage_init; 760 maxscan = cnt.v_inactive_count; 761 762 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl); 763 m != NULL && maxscan-- > 0 && page_shortage > 0; 764 m = next) { 765 766 cnt.v_pdpages++; 767 768 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE) { 769 goto rescan0; 770 } 771 772 next = TAILQ_NEXT(m, pageq); 773 774 /* 775 * skip marker pages 776 */ 777 if (m->flags & PG_MARKER) 778 continue; 779 780 if (!vm_page_trylock(m)) { 781 addl_page_shortage++; 782 continue; 783 } 784 785 /* 786 * A held page may be undergoing I/O, so skip it. 787 */ 788 if (m->hold_count || (object = m->object) == NULL) { 789 vm_page_unlock(m); 790 vm_page_requeue(m); 791 addl_page_shortage++; 792 continue; 793 } 794 795 /* 796 * Don't mess with busy pages, keep in the front of the 797 * queue, most likely are being paged out. 798 */ 799 if (!VM_OBJECT_TRYLOCK(object) && 800 (!vm_pageout_fallback_object_lock(m, &next) || 801 m->hold_count != 0)) { 802 VM_OBJECT_UNLOCK(object); 803 vm_page_unlock(m); 804 addl_page_shortage++; 805 continue; 806 } 807 if (m->busy || (m->oflags & VPO_BUSY)) { 808 vm_page_unlock(m); 809 VM_OBJECT_UNLOCK(object); 810 addl_page_shortage++; 811 continue; 812 } 813 814 /* 815 * If the object is not being used, we ignore previous 816 * references. 817 */ 818 if (object->ref_count == 0) { 819 vm_page_flag_clear(m, PG_REFERENCED); 820 KASSERT(!pmap_page_is_mapped(m), 821 ("vm_pageout_scan: page %p is mapped", m)); 822 823 /* 824 * Otherwise, if the page has been referenced while in the 825 * inactive queue, we bump the "activation count" upwards, 826 * making it less likely that the page will be added back to 827 * the inactive queue prematurely again. Here we check the 828 * page tables (or emulated bits, if any), given the upper 829 * level VM system not knowing anything about existing 830 * references. 831 */ 832 } else if (((m->flags & PG_REFERENCED) == 0) && 833 (actcount = pmap_ts_referenced(m))) { 834 vm_page_activate(m); 835 VM_OBJECT_UNLOCK(object); 836 m->act_count += (actcount + ACT_ADVANCE); 837 vm_page_unlock(m); 838 continue; 839 } 840 841 /* 842 * If the upper level VM system knows about any page 843 * references, we activate the page. We also set the 844 * "activation count" higher than normal so that we will less 845 * likely place pages back onto the inactive queue again. 846 */ 847 if ((m->flags & PG_REFERENCED) != 0) { 848 vm_page_flag_clear(m, PG_REFERENCED); 849 actcount = pmap_ts_referenced(m); 850 vm_page_activate(m); 851 VM_OBJECT_UNLOCK(object); 852 m->act_count += (actcount + ACT_ADVANCE + 1); 853 vm_page_unlock(m); 854 continue; 855 } 856 857 /* 858 * If the upper level VM system does not believe that the page 859 * is fully dirty, but it is mapped for write access, then we 860 * consult the pmap to see if the page's dirty status should 861 * be updated. 862 */ 863 if (m->dirty != VM_PAGE_BITS_ALL && 864 (m->flags & PG_WRITEABLE) != 0) { 865 /* 866 * Avoid a race condition: Unless write access is 867 * removed from the page, another processor could 868 * modify it before all access is removed by the call 869 * to vm_page_cache() below. If vm_page_cache() finds 870 * that the page has been modified when it removes all 871 * access, it panics because it cannot cache dirty 872 * pages. In principle, we could eliminate just write 873 * access here rather than all access. In the expected 874 * case, when there are no last instant modifications 875 * to the page, removing all access will be cheaper 876 * overall. 877 */ 878 if (pmap_is_modified(m)) 879 vm_page_dirty(m); 880 else if (m->dirty == 0) 881 pmap_remove_all(m); 882 } 883 884 if (m->valid == 0) { 885 /* 886 * Invalid pages can be easily freed 887 */ 888 vm_page_free(m); 889 cnt.v_dfree++; 890 --page_shortage; 891 } else if (m->dirty == 0) { 892 /* 893 * Clean pages can be placed onto the cache queue. 894 * This effectively frees them. 895 */ 896 vm_page_cache(m); 897 --page_shortage; 898 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) { 899 /* 900 * Dirty pages need to be paged out, but flushing 901 * a page is extremely expensive verses freeing 902 * a clean page. Rather then artificially limiting 903 * the number of pages we can flush, we instead give 904 * dirty pages extra priority on the inactive queue 905 * by forcing them to be cycled through the queue 906 * twice before being flushed, after which the 907 * (now clean) page will cycle through once more 908 * before being freed. This significantly extends 909 * the thrash point for a heavily loaded machine. 910 */ 911 vm_page_flag_set(m, PG_WINATCFLS); 912 vm_page_requeue(m); 913 } else if (maxlaunder > 0) { 914 /* 915 * We always want to try to flush some dirty pages if 916 * we encounter them, to keep the system stable. 917 * Normally this number is small, but under extreme 918 * pressure where there are insufficient clean pages 919 * on the inactive queue, we may have to go all out. 920 */ 921 int swap_pageouts_ok, vfslocked = 0; 922 struct vnode *vp = NULL; 923 struct mount *mp = NULL; 924 925 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 926 swap_pageouts_ok = 1; 927 } else { 928 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 929 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 930 vm_page_count_min()); 931 932 } 933 934 /* 935 * We don't bother paging objects that are "dead". 936 * Those objects are in a "rundown" state. 937 */ 938 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 939 vm_page_unlock(m); 940 VM_OBJECT_UNLOCK(object); 941 vm_page_requeue(m); 942 continue; 943 } 944 945 /* 946 * Following operations may unlock 947 * vm_page_queue_mtx, invalidating the 'next' 948 * pointer. To prevent an inordinate number 949 * of restarts we use our marker to remember 950 * our place. 951 * 952 */ 953 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, 954 m, &marker, pageq); 955 /* 956 * The object is already known NOT to be dead. It 957 * is possible for the vget() to block the whole 958 * pageout daemon, but the new low-memory handling 959 * code should prevent it. 960 * 961 * The previous code skipped locked vnodes and, worse, 962 * reordered pages in the queue. This results in 963 * completely non-deterministic operation and, on a 964 * busy system, can lead to extremely non-optimal 965 * pageouts. For example, it can cause clean pages 966 * to be freed and dirty pages to be moved to the end 967 * of the queue. Since dirty pages are also moved to 968 * the end of the queue once-cleaned, this gives 969 * way too large a weighting to defering the freeing 970 * of dirty pages. 971 * 972 * We can't wait forever for the vnode lock, we might 973 * deadlock due to a vn_read() getting stuck in 974 * vm_wait while holding this vnode. We skip the 975 * vnode if we can't get it in a reasonable amount 976 * of time. 977 */ 978 if (object->type == OBJT_VNODE) { 979 vm_page_unlock_queues(); 980 vm_page_unlock(m); 981 vp = object->handle; 982 if (vp->v_type == VREG && 983 vn_start_write(vp, &mp, V_NOWAIT) != 0) { 984 mp = NULL; 985 ++pageout_lock_miss; 986 if (object->flags & OBJ_MIGHTBEDIRTY) 987 vnodes_skipped++; 988 vm_page_lock_queues(); 989 goto unlock_and_continue; 990 } 991 KASSERT(mp != NULL, 992 ("vp %p with NULL v_mount", vp)); 993 vm_object_reference_locked(object); 994 VM_OBJECT_UNLOCK(object); 995 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 996 if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK, 997 curthread)) { 998 VM_OBJECT_LOCK(object); 999 vm_page_lock_queues(); 1000 ++pageout_lock_miss; 1001 if (object->flags & OBJ_MIGHTBEDIRTY) 1002 vnodes_skipped++; 1003 vp = NULL; 1004 goto unlock_and_continue; 1005 } 1006 VM_OBJECT_LOCK(object); 1007 vm_page_lock(m); 1008 vm_page_lock_queues(); 1009 /* 1010 * The page might have been moved to another 1011 * queue during potential blocking in vget() 1012 * above. The page might have been freed and 1013 * reused for another vnode. 1014 */ 1015 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE || 1016 m->object != object || 1017 TAILQ_NEXT(m, pageq) != &marker) { 1018 vm_page_unlock(m); 1019 if (object->flags & OBJ_MIGHTBEDIRTY) 1020 vnodes_skipped++; 1021 goto unlock_and_continue; 1022 } 1023 1024 /* 1025 * The page may have been busied during the 1026 * blocking in vget(). We don't move the 1027 * page back onto the end of the queue so that 1028 * statistics are more correct if we don't. 1029 */ 1030 if (m->busy || (m->oflags & VPO_BUSY)) { 1031 vm_page_unlock(m); 1032 goto unlock_and_continue; 1033 } 1034 1035 /* 1036 * If the page has become held it might 1037 * be undergoing I/O, so skip it 1038 */ 1039 if (m->hold_count) { 1040 vm_page_unlock(m); 1041 vm_page_requeue(m); 1042 if (object->flags & OBJ_MIGHTBEDIRTY) 1043 vnodes_skipped++; 1044 goto unlock_and_continue; 1045 } 1046 } 1047 vm_page_unlock(m); 1048 1049 /* 1050 * If a page is dirty, then it is either being washed 1051 * (but not yet cleaned) or it is still in the 1052 * laundry. If it is still in the laundry, then we 1053 * start the cleaning operation. 1054 * 1055 * decrement page_shortage on success to account for 1056 * the (future) cleaned page. Otherwise we could wind 1057 * up laundering or cleaning too many pages. 1058 */ 1059 vm_page_unlock_queues(); 1060 if (vm_pageout_clean(m) != 0) { 1061 --page_shortage; 1062 --maxlaunder; 1063 } 1064 vm_page_lock_queues(); 1065unlock_and_continue: 1066 vm_page_lock_assert(m, MA_NOTOWNED); 1067 VM_OBJECT_UNLOCK(object); 1068 if (mp != NULL) { 1069 vm_page_unlock_queues(); 1070 if (vp != NULL) 1071 vput(vp); 1072 VFS_UNLOCK_GIANT(vfslocked); 1073 vm_object_deallocate(object); 1074 vn_finished_write(mp); 1075 vm_page_lock_queues(); 1076 } 1077 next = TAILQ_NEXT(&marker, pageq); 1078 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, 1079 &marker, pageq); 1080 vm_page_lock_assert(m, MA_NOTOWNED); 1081 continue; 1082 } 1083 vm_page_unlock(m); 1084 VM_OBJECT_UNLOCK(object); 1085 } 1086 1087 /* 1088 * Compute the number of pages we want to try to move from the 1089 * active queue to the inactive queue. 1090 */ 1091 page_shortage = vm_paging_target() + 1092 cnt.v_inactive_target - cnt.v_inactive_count; 1093 page_shortage += addl_page_shortage; 1094 1095 /* 1096 * Scan the active queue for things we can deactivate. We nominally 1097 * track the per-page activity counter and use it to locate 1098 * deactivation candidates. 1099 */ 1100 pcount = cnt.v_active_count; 1101 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1102 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1103 1104 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) { 1105 1106 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE), 1107 ("vm_pageout_scan: page %p isn't active", m)); 1108 1109 next = TAILQ_NEXT(m, pageq); 1110 object = m->object; 1111 if ((m->flags & PG_MARKER) != 0) { 1112 m = next; 1113 continue; 1114 } 1115 if (!vm_page_trylock(m) || (object = m->object) == NULL) { 1116 m = next; 1117 continue; 1118 } 1119 if (!VM_OBJECT_TRYLOCK(object) && 1120 !vm_pageout_fallback_object_lock(m, &next)) { 1121 VM_OBJECT_UNLOCK(object); 1122 vm_page_unlock(m); 1123 m = next; 1124 continue; 1125 } 1126 1127 /* 1128 * Don't deactivate pages that are busy. 1129 */ 1130 if ((m->busy != 0) || 1131 (m->oflags & VPO_BUSY) || 1132 (m->hold_count != 0)) { 1133 vm_page_unlock(m); 1134 VM_OBJECT_UNLOCK(object); 1135 vm_page_requeue(m); 1136 m = next; 1137 continue; 1138 } 1139 1140 /* 1141 * The count for pagedaemon pages is done after checking the 1142 * page for eligibility... 1143 */ 1144 cnt.v_pdpages++; 1145 1146 /* 1147 * Check to see "how much" the page has been used. 1148 */ 1149 actcount = 0; 1150 if (object->ref_count != 0) { 1151 if (m->flags & PG_REFERENCED) { 1152 actcount += 1; 1153 } 1154 actcount += pmap_ts_referenced(m); 1155 if (actcount) { 1156 m->act_count += ACT_ADVANCE + actcount; 1157 if (m->act_count > ACT_MAX) 1158 m->act_count = ACT_MAX; 1159 } 1160 } 1161 1162 /* 1163 * Since we have "tested" this bit, we need to clear it now. 1164 */ 1165 vm_page_flag_clear(m, PG_REFERENCED); 1166 1167 /* 1168 * Only if an object is currently being used, do we use the 1169 * page activation count stats. 1170 */ 1171 if (actcount && (object->ref_count != 0)) { 1172 vm_page_requeue(m); 1173 } else { 1174 m->act_count -= min(m->act_count, ACT_DECLINE); 1175 if (vm_pageout_algorithm || 1176 object->ref_count == 0 || 1177 m->act_count == 0) { 1178 page_shortage--; 1179 if (object->ref_count == 0) { 1180 KASSERT(!pmap_page_is_mapped(m), 1181 ("vm_pageout_scan: page %p is mapped", m)); 1182 if (m->dirty == 0) 1183 vm_page_cache(m); 1184 else 1185 vm_page_deactivate(m); 1186 } else { 1187 vm_page_deactivate(m); 1188 } 1189 } else { 1190 vm_page_requeue(m); 1191 } 1192 } 1193 vm_page_unlock(m); 1194 VM_OBJECT_UNLOCK(object); 1195 m = next; 1196 } 1197 vm_page_unlock_queues(); 1198#if !defined(NO_SWAPPING) 1199 /* 1200 * Idle process swapout -- run once per second. 1201 */ 1202 if (vm_swap_idle_enabled) { 1203 static long lsec; 1204 if (time_second != lsec) { 1205 vm_req_vmdaemon(VM_SWAP_IDLE); 1206 lsec = time_second; 1207 } 1208 } 1209#endif 1210 1211 /* 1212 * If we didn't get enough free pages, and we have skipped a vnode 1213 * in a writeable object, wakeup the sync daemon. And kick swapout 1214 * if we did not get enough free pages. 1215 */ 1216 if (vm_paging_target() > 0) { 1217 if (vnodes_skipped && vm_page_count_min()) 1218 (void) speedup_syncer(); 1219#if !defined(NO_SWAPPING) 1220 if (vm_swap_enabled && vm_page_count_target()) 1221 vm_req_vmdaemon(VM_SWAP_NORMAL); 1222#endif 1223 } 1224 1225 /* 1226 * If we are critically low on one of RAM or swap and low on 1227 * the other, kill the largest process. However, we avoid 1228 * doing this on the first pass in order to give ourselves a 1229 * chance to flush out dirty vnode-backed pages and to allow 1230 * active pages to be moved to the inactive queue and reclaimed. 1231 */ 1232 if (pass != 0 && 1233 ((swap_pager_avail < 64 && vm_page_count_min()) || 1234 (swap_pager_full && vm_paging_target() > 0))) 1235 vm_pageout_oom(VM_OOM_MEM); 1236} 1237 1238 1239void 1240vm_pageout_oom(int shortage) 1241{ 1242 struct proc *p, *bigproc; 1243 vm_offset_t size, bigsize; 1244 struct thread *td; 1245 struct vmspace *vm; 1246 1247 /* 1248 * We keep the process bigproc locked once we find it to keep anyone 1249 * from messing with it; however, there is a possibility of 1250 * deadlock if process B is bigproc and one of it's child processes 1251 * attempts to propagate a signal to B while we are waiting for A's 1252 * lock while walking this list. To avoid this, we don't block on 1253 * the process lock but just skip a process if it is already locked. 1254 */ 1255 bigproc = NULL; 1256 bigsize = 0; 1257 sx_slock(&allproc_lock); 1258 FOREACH_PROC_IN_SYSTEM(p) { 1259 int breakout; 1260 1261 if (PROC_TRYLOCK(p) == 0) 1262 continue; 1263 /* 1264 * If this is a system, protected or killed process, skip it. 1265 */ 1266 if ((p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM)) || 1267 (p->p_pid == 1) || P_KILLED(p) || 1268 ((p->p_pid < 48) && (swap_pager_avail != 0))) { 1269 PROC_UNLOCK(p); 1270 continue; 1271 } 1272 /* 1273 * If the process is in a non-running type state, 1274 * don't touch it. Check all the threads individually. 1275 */ 1276 breakout = 0; 1277 FOREACH_THREAD_IN_PROC(p, td) { 1278 thread_lock(td); 1279 if (!TD_ON_RUNQ(td) && 1280 !TD_IS_RUNNING(td) && 1281 !TD_IS_SLEEPING(td)) { 1282 thread_unlock(td); 1283 breakout = 1; 1284 break; 1285 } 1286 thread_unlock(td); 1287 } 1288 if (breakout) { 1289 PROC_UNLOCK(p); 1290 continue; 1291 } 1292 /* 1293 * get the process size 1294 */ 1295 vm = vmspace_acquire_ref(p); 1296 if (vm == NULL) { 1297 PROC_UNLOCK(p); 1298 continue; 1299 } 1300 if (!vm_map_trylock_read(&vm->vm_map)) { 1301 vmspace_free(vm); 1302 PROC_UNLOCK(p); 1303 continue; 1304 } 1305 size = vmspace_swap_count(vm); 1306 vm_map_unlock_read(&vm->vm_map); 1307 if (shortage == VM_OOM_MEM) 1308 size += vmspace_resident_count(vm); 1309 vmspace_free(vm); 1310 /* 1311 * if the this process is bigger than the biggest one 1312 * remember it. 1313 */ 1314 if (size > bigsize) { 1315 if (bigproc != NULL) 1316 PROC_UNLOCK(bigproc); 1317 bigproc = p; 1318 bigsize = size; 1319 } else 1320 PROC_UNLOCK(p); 1321 } 1322 sx_sunlock(&allproc_lock); 1323 if (bigproc != NULL) { 1324 killproc(bigproc, "out of swap space"); 1325 sched_nice(bigproc, PRIO_MIN); 1326 PROC_UNLOCK(bigproc); 1327 wakeup(&cnt.v_free_count); 1328 } 1329} 1330 1331/* 1332 * This routine tries to maintain the pseudo LRU active queue, 1333 * so that during long periods of time where there is no paging, 1334 * that some statistic accumulation still occurs. This code 1335 * helps the situation where paging just starts to occur. 1336 */ 1337static void 1338vm_pageout_page_stats() 1339{ 1340 vm_object_t object; 1341 vm_page_t m,next; 1342 int pcount,tpcount; /* Number of pages to check */ 1343 static int fullintervalcount = 0; 1344 int page_shortage; 1345 1346 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1347 page_shortage = 1348 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) - 1349 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count); 1350 1351 if (page_shortage <= 0) 1352 return; 1353 1354 pcount = cnt.v_active_count; 1355 fullintervalcount += vm_pageout_stats_interval; 1356 if (fullintervalcount < vm_pageout_full_stats_interval) { 1357 tpcount = (int64_t)vm_pageout_stats_max * cnt.v_active_count / 1358 cnt.v_page_count; 1359 if (pcount > tpcount) 1360 pcount = tpcount; 1361 } else { 1362 fullintervalcount = 0; 1363 } 1364 1365 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1366 while ((m != NULL) && (pcount-- > 0)) { 1367 int actcount; 1368 1369 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE), 1370 ("vm_pageout_page_stats: page %p isn't active", m)); 1371 1372 next = TAILQ_NEXT(m, pageq); 1373 object = m->object; 1374 1375 if ((m->flags & PG_MARKER) != 0) { 1376 m = next; 1377 continue; 1378 } 1379 vm_page_lock_assert(m, MA_NOTOWNED); 1380 if (vm_page_trylock(m) == 0 || (object = m->object) == NULL) { 1381 m = next; 1382 continue; 1383 } 1384 if (!VM_OBJECT_TRYLOCK(object) && 1385 !vm_pageout_fallback_object_lock(m, &next)) { 1386 VM_OBJECT_UNLOCK(object); 1387 vm_page_unlock(m); 1388 m = next; 1389 continue; 1390 } 1391 1392 /* 1393 * Don't deactivate pages that are busy. 1394 */ 1395 if ((m->busy != 0) || 1396 (m->oflags & VPO_BUSY) || 1397 (m->hold_count != 0)) { 1398 vm_page_unlock(m); 1399 VM_OBJECT_UNLOCK(object); 1400 vm_page_requeue(m); 1401 m = next; 1402 continue; 1403 } 1404 1405 actcount = 0; 1406 if (m->flags & PG_REFERENCED) { 1407 vm_page_flag_clear(m, PG_REFERENCED); 1408 actcount += 1; 1409 } 1410 1411 actcount += pmap_ts_referenced(m); 1412 if (actcount) { 1413 m->act_count += ACT_ADVANCE + actcount; 1414 if (m->act_count > ACT_MAX) 1415 m->act_count = ACT_MAX; 1416 vm_page_requeue(m); 1417 } else { 1418 if (m->act_count == 0) { 1419 /* 1420 * We turn off page access, so that we have 1421 * more accurate RSS stats. We don't do this 1422 * in the normal page deactivation when the 1423 * system is loaded VM wise, because the 1424 * cost of the large number of page protect 1425 * operations would be higher than the value 1426 * of doing the operation. 1427 */ 1428 pmap_remove_all(m); 1429 vm_page_deactivate(m); 1430 } else { 1431 m->act_count -= min(m->act_count, ACT_DECLINE); 1432 vm_page_requeue(m); 1433 } 1434 } 1435 vm_page_unlock(m); 1436 VM_OBJECT_UNLOCK(object); 1437 m = next; 1438 } 1439} 1440 1441/* 1442 * vm_pageout is the high level pageout daemon. 1443 */ 1444static void 1445vm_pageout() 1446{ 1447 int error, pass; 1448 1449 /* 1450 * Initialize some paging parameters. 1451 */ 1452 cnt.v_interrupt_free_min = 2; 1453 if (cnt.v_page_count < 2000) 1454 vm_pageout_page_count = 8; 1455 1456 /* 1457 * v_free_reserved needs to include enough for the largest 1458 * swap pager structures plus enough for any pv_entry structs 1459 * when paging. 1460 */ 1461 if (cnt.v_page_count > 1024) 1462 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200; 1463 else 1464 cnt.v_free_min = 4; 1465 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1466 cnt.v_interrupt_free_min; 1467 cnt.v_free_reserved = vm_pageout_page_count + 1468 cnt.v_pageout_free_min + (cnt.v_page_count / 768); 1469 cnt.v_free_severe = cnt.v_free_min / 2; 1470 cnt.v_free_min += cnt.v_free_reserved; 1471 cnt.v_free_severe += cnt.v_free_reserved; 1472 1473 /* 1474 * v_free_target and v_cache_min control pageout hysteresis. Note 1475 * that these are more a measure of the VM cache queue hysteresis 1476 * then the VM free queue. Specifically, v_free_target is the 1477 * high water mark (free+cache pages). 1478 * 1479 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the 1480 * low water mark, while v_free_min is the stop. v_cache_min must 1481 * be big enough to handle memory needs while the pageout daemon 1482 * is signalled and run to free more pages. 1483 */ 1484 if (cnt.v_free_count > 6144) 1485 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved; 1486 else 1487 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved; 1488 1489 if (cnt.v_free_count > 2048) { 1490 cnt.v_cache_min = cnt.v_free_target; 1491 cnt.v_cache_max = 2 * cnt.v_cache_min; 1492 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2; 1493 } else { 1494 cnt.v_cache_min = 0; 1495 cnt.v_cache_max = 0; 1496 cnt.v_inactive_target = cnt.v_free_count / 4; 1497 } 1498 if (cnt.v_inactive_target > cnt.v_free_count / 3) 1499 cnt.v_inactive_target = cnt.v_free_count / 3; 1500 1501 /* XXX does not really belong here */ 1502 if (vm_page_max_wired == 0) 1503 vm_page_max_wired = cnt.v_free_count / 3; 1504 1505 if (vm_pageout_stats_max == 0) 1506 vm_pageout_stats_max = cnt.v_free_target; 1507 1508 /* 1509 * Set interval in seconds for stats scan. 1510 */ 1511 if (vm_pageout_stats_interval == 0) 1512 vm_pageout_stats_interval = 5; 1513 if (vm_pageout_full_stats_interval == 0) 1514 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4; 1515 1516 swap_pager_swap_init(); 1517 pass = 0; 1518 /* 1519 * The pageout daemon is never done, so loop forever. 1520 */ 1521 while (TRUE) { 1522 /* 1523 * If we have enough free memory, wakeup waiters. Do 1524 * not clear vm_pages_needed until we reach our target, 1525 * otherwise we may be woken up over and over again and 1526 * waste a lot of cpu. 1527 */ 1528 mtx_lock(&vm_page_queue_free_mtx); 1529 if (vm_pages_needed && !vm_page_count_min()) { 1530 if (!vm_paging_needed()) 1531 vm_pages_needed = 0; 1532 wakeup(&cnt.v_free_count); 1533 } 1534 if (vm_pages_needed) { 1535 /* 1536 * Still not done, take a second pass without waiting 1537 * (unlimited dirty cleaning), otherwise sleep a bit 1538 * and try again. 1539 */ 1540 ++pass; 1541 if (pass > 1) 1542 msleep(&vm_pages_needed, 1543 &vm_page_queue_free_mtx, PVM, "psleep", 1544 hz / 2); 1545 } else { 1546 /* 1547 * Good enough, sleep & handle stats. Prime the pass 1548 * for the next run. 1549 */ 1550 if (pass > 1) 1551 pass = 1; 1552 else 1553 pass = 0; 1554 error = msleep(&vm_pages_needed, 1555 &vm_page_queue_free_mtx, PVM, "psleep", 1556 vm_pageout_stats_interval * hz); 1557 if (error && !vm_pages_needed) { 1558 mtx_unlock(&vm_page_queue_free_mtx); 1559 pass = 0; 1560 vm_page_lock_queues(); 1561 vm_pageout_page_stats(); 1562 vm_page_unlock_queues(); 1563 continue; 1564 } 1565 } 1566 if (vm_pages_needed) 1567 cnt.v_pdwakeups++; 1568 mtx_unlock(&vm_page_queue_free_mtx); 1569 vm_pageout_scan(pass); 1570 } 1571} 1572 1573/* 1574 * Unless the free page queue lock is held by the caller, this function 1575 * should be regarded as advisory. Specifically, the caller should 1576 * not msleep() on &cnt.v_free_count following this function unless 1577 * the free page queue lock is held until the msleep() is performed. 1578 */ 1579void 1580pagedaemon_wakeup() 1581{ 1582 1583 if (!vm_pages_needed && curthread->td_proc != pageproc) { 1584 vm_pages_needed = 1; 1585 wakeup(&vm_pages_needed); 1586 } 1587} 1588 1589#if !defined(NO_SWAPPING) 1590static void 1591vm_req_vmdaemon(int req) 1592{ 1593 static int lastrun = 0; 1594 1595 mtx_lock(&vm_daemon_mtx); 1596 vm_pageout_req_swapout |= req; 1597 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1598 wakeup(&vm_daemon_needed); 1599 lastrun = ticks; 1600 } 1601 mtx_unlock(&vm_daemon_mtx); 1602} 1603 1604static void 1605vm_daemon() 1606{ 1607 struct rlimit rsslim; 1608 struct proc *p; 1609 struct thread *td; 1610 struct vmspace *vm; 1611 int breakout, swapout_flags; 1612 1613 while (TRUE) { 1614 mtx_lock(&vm_daemon_mtx); 1615 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0); 1616 swapout_flags = vm_pageout_req_swapout; 1617 vm_pageout_req_swapout = 0; 1618 mtx_unlock(&vm_daemon_mtx); 1619 if (swapout_flags) 1620 swapout_procs(swapout_flags); 1621 1622 /* 1623 * scan the processes for exceeding their rlimits or if 1624 * process is swapped out -- deactivate pages 1625 */ 1626 sx_slock(&allproc_lock); 1627 FOREACH_PROC_IN_SYSTEM(p) { 1628 vm_pindex_t limit, size; 1629 1630 /* 1631 * if this is a system process or if we have already 1632 * looked at this process, skip it. 1633 */ 1634 PROC_LOCK(p); 1635 if (p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) { 1636 PROC_UNLOCK(p); 1637 continue; 1638 } 1639 /* 1640 * if the process is in a non-running type state, 1641 * don't touch it. 1642 */ 1643 breakout = 0; 1644 FOREACH_THREAD_IN_PROC(p, td) { 1645 thread_lock(td); 1646 if (!TD_ON_RUNQ(td) && 1647 !TD_IS_RUNNING(td) && 1648 !TD_IS_SLEEPING(td)) { 1649 thread_unlock(td); 1650 breakout = 1; 1651 break; 1652 } 1653 thread_unlock(td); 1654 } 1655 if (breakout) { 1656 PROC_UNLOCK(p); 1657 continue; 1658 } 1659 /* 1660 * get a limit 1661 */ 1662 lim_rlimit(p, RLIMIT_RSS, &rsslim); 1663 limit = OFF_TO_IDX( 1664 qmin(rsslim.rlim_cur, rsslim.rlim_max)); 1665 1666 /* 1667 * let processes that are swapped out really be 1668 * swapped out set the limit to nothing (will force a 1669 * swap-out.) 1670 */ 1671 if ((p->p_flag & P_INMEM) == 0) 1672 limit = 0; /* XXX */ 1673 vm = vmspace_acquire_ref(p); 1674 PROC_UNLOCK(p); 1675 if (vm == NULL) 1676 continue; 1677 1678 size = vmspace_resident_count(vm); 1679 if (limit >= 0 && size >= limit) { 1680 vm_pageout_map_deactivate_pages( 1681 &vm->vm_map, limit); 1682 } 1683 vmspace_free(vm); 1684 } 1685 sx_sunlock(&allproc_lock); 1686 } 1687} 1688#endif /* !defined(NO_SWAPPING) */ 1689