vm_pageout.c revision 207759
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 207759 2010-05-07 21:06:52Z jkim $"); 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 218static void 219vm_pageout_init_marker(vm_page_t marker, u_short queue) 220{ 221 222 bzero(marker, sizeof(*marker)); 223 marker->flags = PG_FICTITIOUS | PG_MARKER; 224 marker->oflags = VPO_BUSY; 225 marker->queue = queue; 226 marker->wire_count = 1; 227} 228 229/* 230 * vm_pageout_fallback_object_lock: 231 * 232 * Lock vm object currently associated with `m'. VM_OBJECT_TRYLOCK is 233 * known to have failed and page queue must be either PQ_ACTIVE or 234 * PQ_INACTIVE. To avoid lock order violation, unlock the page queues 235 * while locking the vm object. Use marker page to detect page queue 236 * changes and maintain notion of next page on page queue. Return 237 * TRUE if no changes were detected, FALSE otherwise. vm object is 238 * locked on return. 239 * 240 * This function depends on both the lock portion of struct vm_object 241 * and normal struct vm_page being type stable. 242 */ 243boolean_t 244vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next) 245{ 246 struct vm_page marker; 247 boolean_t unchanged; 248 u_short queue; 249 vm_object_t object; 250 251 queue = m->queue; 252 vm_pageout_init_marker(&marker, queue); 253 object = m->object; 254 255 TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl, 256 m, &marker, pageq); 257 vm_page_unlock_queues(); 258 vm_page_unlock(m); 259 VM_OBJECT_LOCK(object); 260 vm_page_lock(m); 261 vm_page_lock_queues(); 262 263 /* Page queue might have changed. */ 264 *next = TAILQ_NEXT(&marker, pageq); 265 unchanged = (m->queue == queue && 266 m->object == object && 267 &marker == TAILQ_NEXT(m, pageq)); 268 TAILQ_REMOVE(&vm_page_queues[queue].pl, 269 &marker, pageq); 270 return (unchanged); 271} 272 273/* 274 * Lock the page while holding the page queue lock. Use marker page 275 * to detect page queue changes and maintain notion of next page on 276 * page queue. Return TRUE if no changes were detected, FALSE 277 * otherwise. The page is locked on return. The page queue lock might 278 * be dropped and reacquired. 279 * 280 * This function depends on normal struct vm_page being type stable. 281 */ 282boolean_t 283vm_pageout_page_lock(vm_page_t m, vm_page_t *next) 284{ 285 struct vm_page marker; 286 boolean_t unchanged; 287 u_short queue; 288 289 vm_page_lock_assert(m, MA_NOTOWNED); 290 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 291 292 if (vm_page_trylock(m)) 293 return (TRUE); 294 295 queue = m->queue; 296 vm_pageout_init_marker(&marker, queue); 297 298 TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl, m, &marker, pageq); 299 vm_page_unlock_queues(); 300 vm_page_lock(m); 301 vm_page_lock_queues(); 302 303 /* Page queue might have changed. */ 304 *next = TAILQ_NEXT(&marker, pageq); 305 unchanged = (m->queue == queue && &marker == TAILQ_NEXT(m, pageq)); 306 TAILQ_REMOVE(&vm_page_queues[queue].pl, &marker, pageq); 307 return (unchanged); 308} 309 310/* 311 * vm_pageout_clean: 312 * 313 * Clean the page and remove it from the laundry. 314 * 315 * We set the busy bit to cause potential page faults on this page to 316 * block. Note the careful timing, however, the busy bit isn't set till 317 * late and we cannot do anything that will mess with the page. 318 */ 319static int 320vm_pageout_clean(vm_page_t m) 321{ 322 vm_object_t object; 323 vm_page_t mc[2*vm_pageout_page_count]; 324 int pageout_count; 325 int ib, is, page_base; 326 vm_pindex_t pindex = m->pindex; 327 328 vm_page_lock_assert(m, MA_NOTOWNED); 329 vm_page_lock(m); 330 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 331 332 /* 333 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP 334 * with the new swapper, but we could have serious problems paging 335 * out other object types if there is insufficient memory. 336 * 337 * Unfortunately, checking free memory here is far too late, so the 338 * check has been moved up a procedural level. 339 */ 340 341 /* 342 * Can't clean the page if it's busy or held. 343 */ 344 if ((m->hold_count != 0) || 345 ((m->busy != 0) || (m->oflags & VPO_BUSY))) { 346 vm_page_unlock(m); 347 return 0; 348 } 349 350 mc[vm_pageout_page_count] = m; 351 pageout_count = 1; 352 page_base = vm_pageout_page_count; 353 ib = 1; 354 is = 1; 355 356 /* 357 * Scan object for clusterable pages. 358 * 359 * We can cluster ONLY if: ->> the page is NOT 360 * clean, wired, busy, held, or mapped into a 361 * buffer, and one of the following: 362 * 1) The page is inactive, or a seldom used 363 * active page. 364 * -or- 365 * 2) we force the issue. 366 * 367 * During heavy mmap/modification loads the pageout 368 * daemon can really fragment the underlying file 369 * due to flushing pages out of order and not trying 370 * align the clusters (which leave sporatic out-of-order 371 * holes). To solve this problem we do the reverse scan 372 * first and attempt to align our cluster, then do a 373 * forward scan if room remains. 374 */ 375 object = m->object; 376more: 377 while (ib && pageout_count < vm_pageout_page_count) { 378 vm_page_t p; 379 380 if (ib > pindex) { 381 ib = 0; 382 break; 383 } 384 385 if ((p = vm_page_lookup(object, pindex - ib)) == NULL) { 386 ib = 0; 387 break; 388 } 389 if ((p->oflags & VPO_BUSY) || p->busy) { 390 ib = 0; 391 break; 392 } 393 vm_page_lock(p); 394 vm_page_lock_queues(); 395 vm_page_test_dirty(p); 396 if (p->dirty == 0 || 397 p->queue != PQ_INACTIVE || 398 p->hold_count != 0) { /* may be undergoing I/O */ 399 vm_page_unlock(p); 400 vm_page_unlock_queues(); 401 ib = 0; 402 break; 403 } 404 vm_page_unlock_queues(); 405 vm_page_unlock(p); 406 mc[--page_base] = p; 407 ++pageout_count; 408 ++ib; 409 /* 410 * alignment boundry, stop here and switch directions. Do 411 * not clear ib. 412 */ 413 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0) 414 break; 415 } 416 417 while (pageout_count < vm_pageout_page_count && 418 pindex + is < object->size) { 419 vm_page_t p; 420 421 if ((p = vm_page_lookup(object, pindex + is)) == NULL) 422 break; 423 if ((p->oflags & VPO_BUSY) || p->busy) { 424 break; 425 } 426 vm_page_lock(p); 427 vm_page_lock_queues(); 428 vm_page_test_dirty(p); 429 if (p->dirty == 0 || 430 p->queue != PQ_INACTIVE || 431 p->hold_count != 0) { /* may be undergoing I/O */ 432 vm_page_unlock_queues(); 433 vm_page_unlock(p); 434 break; 435 } 436 vm_page_unlock_queues(); 437 vm_page_unlock(p); 438 mc[page_base + pageout_count] = p; 439 ++pageout_count; 440 ++is; 441 } 442 443 /* 444 * If we exhausted our forward scan, continue with the reverse scan 445 * when possible, even past a page boundry. This catches boundry 446 * conditions. 447 */ 448 if (ib && pageout_count < vm_pageout_page_count) 449 goto more; 450 451 vm_page_unlock(m); 452 /* 453 * we allow reads during pageouts... 454 */ 455 return (vm_pageout_flush(&mc[page_base], pageout_count, 0)); 456} 457 458/* 459 * vm_pageout_flush() - launder the given pages 460 * 461 * The given pages are laundered. Note that we setup for the start of 462 * I/O ( i.e. busy the page ), mark it read-only, and bump the object 463 * reference count all in here rather then in the parent. If we want 464 * the parent to do more sophisticated things we may have to change 465 * the ordering. 466 */ 467int 468vm_pageout_flush(vm_page_t *mc, int count, int flags) 469{ 470 vm_object_t object = mc[0]->object; 471 int pageout_status[count]; 472 int numpagedout = 0; 473 int i; 474 475 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 476 mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED); 477 478 /* 479 * Initiate I/O. Bump the vm_page_t->busy counter and 480 * mark the pages read-only. 481 * 482 * We do not have to fixup the clean/dirty bits here... we can 483 * allow the pager to do it after the I/O completes. 484 * 485 * NOTE! mc[i]->dirty may be partial or fragmented due to an 486 * edge case with file fragments. 487 */ 488 for (i = 0; i < count; i++) { 489 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, 490 ("vm_pageout_flush: partially invalid page %p index %d/%d", 491 mc[i], i, count)); 492 vm_page_io_start(mc[i]); 493 vm_page_lock(mc[i]); 494 vm_page_lock_queues(); 495 pmap_remove_write(mc[i]); 496 vm_page_unlock(mc[i]); 497 vm_page_unlock_queues(); 498 } 499 vm_object_pip_add(object, count); 500 501 vm_pager_put_pages(object, mc, count, flags, pageout_status); 502 503 for (i = 0; i < count; i++) { 504 vm_page_t mt = mc[i]; 505 506 vm_page_lock(mt); 507 vm_page_lock_queues(); 508 KASSERT(pageout_status[i] == VM_PAGER_PEND || 509 (mt->flags & PG_WRITEABLE) == 0, 510 ("vm_pageout_flush: page %p is not write protected", mt)); 511 switch (pageout_status[i]) { 512 case VM_PAGER_OK: 513 case VM_PAGER_PEND: 514 numpagedout++; 515 break; 516 case VM_PAGER_BAD: 517 /* 518 * Page outside of range of object. Right now we 519 * essentially lose the changes by pretending it 520 * worked. 521 */ 522 vm_page_undirty(mt); 523 break; 524 case VM_PAGER_ERROR: 525 case VM_PAGER_FAIL: 526 /* 527 * If page couldn't be paged out, then reactivate the 528 * page so it doesn't clog the inactive list. (We 529 * will try paging out it again later). 530 */ 531 vm_page_activate(mt); 532 break; 533 case VM_PAGER_AGAIN: 534 break; 535 } 536 537 /* 538 * If the operation is still going, leave the page busy to 539 * block all other accesses. Also, leave the paging in 540 * progress indicator set so that we don't attempt an object 541 * collapse. 542 */ 543 if (pageout_status[i] != VM_PAGER_PEND) { 544 vm_object_pip_wakeup(object); 545 vm_page_io_finish(mt); 546 if (vm_page_count_severe()) 547 vm_page_try_to_cache(mt); 548 } 549 vm_page_unlock_queues(); 550 vm_page_unlock(mt); 551 } 552 return numpagedout; 553} 554 555#if !defined(NO_SWAPPING) 556/* 557 * vm_pageout_object_deactivate_pages 558 * 559 * deactivate enough pages to satisfy the inactive target 560 * requirements or if vm_page_proc_limit is set, then 561 * deactivate all of the pages in the object and its 562 * backing_objects. 563 * 564 * The object and map must be locked. 565 */ 566static void 567vm_pageout_object_deactivate_pages(pmap, first_object, desired) 568 pmap_t pmap; 569 vm_object_t first_object; 570 long desired; 571{ 572 vm_object_t backing_object, object; 573 vm_page_t p, next; 574 int actcount, remove_mode; 575 576 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED); 577 if (first_object->type == OBJT_DEVICE || 578 first_object->type == OBJT_SG) 579 return; 580 for (object = first_object;; object = backing_object) { 581 if (pmap_resident_count(pmap) <= desired) 582 goto unlock_return; 583 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 584 if (object->type == OBJT_PHYS || object->paging_in_progress) 585 goto unlock_return; 586 587 remove_mode = 0; 588 if (object->shadow_count > 1) 589 remove_mode = 1; 590 /* 591 * scan the objects entire memory queue 592 */ 593 p = TAILQ_FIRST(&object->memq); 594 while (p != NULL) { 595 if (pmap_resident_count(pmap) <= desired) 596 goto unlock_return; 597 next = TAILQ_NEXT(p, listq); 598 if ((p->oflags & VPO_BUSY) != 0 || p->busy != 0) { 599 p = next; 600 continue; 601 } 602 vm_page_lock(p); 603 vm_page_lock_queues(); 604 cnt.v_pdpages++; 605 if (p->wire_count != 0 || 606 p->hold_count != 0 || 607 !pmap_page_exists_quick(pmap, p)) { 608 vm_page_unlock_queues(); 609 vm_page_unlock(p); 610 p = next; 611 continue; 612 } 613 actcount = pmap_ts_referenced(p); 614 if (actcount) { 615 vm_page_flag_set(p, PG_REFERENCED); 616 } else if (p->flags & PG_REFERENCED) { 617 actcount = 1; 618 } 619 if ((p->queue != PQ_ACTIVE) && 620 (p->flags & PG_REFERENCED)) { 621 vm_page_activate(p); 622 p->act_count += actcount; 623 vm_page_flag_clear(p, PG_REFERENCED); 624 } else if (p->queue == PQ_ACTIVE) { 625 if ((p->flags & PG_REFERENCED) == 0) { 626 p->act_count -= min(p->act_count, ACT_DECLINE); 627 if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) { 628 pmap_remove_all(p); 629 vm_page_deactivate(p); 630 } else { 631 vm_page_requeue(p); 632 } 633 } else { 634 vm_page_activate(p); 635 vm_page_flag_clear(p, PG_REFERENCED); 636 if (p->act_count < (ACT_MAX - ACT_ADVANCE)) 637 p->act_count += ACT_ADVANCE; 638 vm_page_requeue(p); 639 } 640 } else if (p->queue == PQ_INACTIVE) { 641 pmap_remove_all(p); 642 } 643 vm_page_unlock_queues(); 644 vm_page_unlock(p); 645 p = next; 646 } 647 if ((backing_object = object->backing_object) == NULL) 648 goto unlock_return; 649 VM_OBJECT_LOCK(backing_object); 650 if (object != first_object) 651 VM_OBJECT_UNLOCK(object); 652 } 653unlock_return: 654 if (object != first_object) 655 VM_OBJECT_UNLOCK(object); 656} 657 658/* 659 * deactivate some number of pages in a map, try to do it fairly, but 660 * that is really hard to do. 661 */ 662static void 663vm_pageout_map_deactivate_pages(map, desired) 664 vm_map_t map; 665 long desired; 666{ 667 vm_map_entry_t tmpe; 668 vm_object_t obj, bigobj; 669 int nothingwired; 670 671 if (!vm_map_trylock(map)) 672 return; 673 674 bigobj = NULL; 675 nothingwired = TRUE; 676 677 /* 678 * first, search out the biggest object, and try to free pages from 679 * that. 680 */ 681 tmpe = map->header.next; 682 while (tmpe != &map->header) { 683 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 684 obj = tmpe->object.vm_object; 685 if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) { 686 if (obj->shadow_count <= 1 && 687 (bigobj == NULL || 688 bigobj->resident_page_count < obj->resident_page_count)) { 689 if (bigobj != NULL) 690 VM_OBJECT_UNLOCK(bigobj); 691 bigobj = obj; 692 } else 693 VM_OBJECT_UNLOCK(obj); 694 } 695 } 696 if (tmpe->wired_count > 0) 697 nothingwired = FALSE; 698 tmpe = tmpe->next; 699 } 700 701 if (bigobj != NULL) { 702 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired); 703 VM_OBJECT_UNLOCK(bigobj); 704 } 705 /* 706 * Next, hunt around for other pages to deactivate. We actually 707 * do this search sort of wrong -- .text first is not the best idea. 708 */ 709 tmpe = map->header.next; 710 while (tmpe != &map->header) { 711 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 712 break; 713 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 714 obj = tmpe->object.vm_object; 715 if (obj != NULL) { 716 VM_OBJECT_LOCK(obj); 717 vm_pageout_object_deactivate_pages(map->pmap, obj, desired); 718 VM_OBJECT_UNLOCK(obj); 719 } 720 } 721 tmpe = tmpe->next; 722 } 723 724 /* 725 * Remove all mappings if a process is swapped out, this will free page 726 * table pages. 727 */ 728 if (desired == 0 && nothingwired) { 729 pmap_remove(vm_map_pmap(map), vm_map_min(map), 730 vm_map_max(map)); 731 } 732 vm_map_unlock(map); 733} 734#endif /* !defined(NO_SWAPPING) */ 735 736/* 737 * vm_pageout_scan does the dirty work for the pageout daemon. 738 */ 739static void 740vm_pageout_scan(int pass) 741{ 742 vm_page_t m, next; 743 struct vm_page marker; 744 int page_shortage, maxscan, pcount; 745 int addl_page_shortage, addl_page_shortage_init; 746 vm_object_t object; 747 int actcount; 748 int vnodes_skipped = 0; 749 int maxlaunder; 750 751 /* 752 * Decrease registered cache sizes. 753 */ 754 EVENTHANDLER_INVOKE(vm_lowmem, 0); 755 /* 756 * We do this explicitly after the caches have been drained above. 757 */ 758 uma_reclaim(); 759 760 addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit); 761 762 /* 763 * Calculate the number of pages we want to either free or move 764 * to the cache. 765 */ 766 page_shortage = vm_paging_target() + addl_page_shortage_init; 767 768 vm_pageout_init_marker(&marker, PQ_INACTIVE); 769 770 /* 771 * Start scanning the inactive queue for pages we can move to the 772 * cache or free. The scan will stop when the target is reached or 773 * we have scanned the entire inactive queue. Note that m->act_count 774 * is not used to form decisions for the inactive queue, only for the 775 * active queue. 776 * 777 * maxlaunder limits the number of dirty pages we flush per scan. 778 * For most systems a smaller value (16 or 32) is more robust under 779 * extreme memory and disk pressure because any unnecessary writes 780 * to disk can result in extreme performance degredation. However, 781 * systems with excessive dirty pages (especially when MAP_NOSYNC is 782 * used) will die horribly with limited laundering. If the pageout 783 * daemon cannot clean enough pages in the first pass, we let it go 784 * all out in succeeding passes. 785 */ 786 if ((maxlaunder = vm_max_launder) <= 1) 787 maxlaunder = 1; 788 if (pass) 789 maxlaunder = 10000; 790 vm_page_lock_queues(); 791rescan0: 792 addl_page_shortage = addl_page_shortage_init; 793 maxscan = cnt.v_inactive_count; 794 795 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl); 796 m != NULL && maxscan-- > 0 && page_shortage > 0; 797 m = next) { 798 799 cnt.v_pdpages++; 800 801 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE) { 802 goto rescan0; 803 } 804 805 next = TAILQ_NEXT(m, pageq); 806 807 /* 808 * skip marker pages 809 */ 810 if (m->flags & PG_MARKER) 811 continue; 812 813 /* 814 * Lock the page. 815 */ 816 if (!vm_pageout_page_lock(m, &next)) { 817 vm_page_unlock(m); 818 addl_page_shortage++; 819 continue; 820 } 821 822 /* 823 * A held page may be undergoing I/O, so skip it. 824 */ 825 if (m->hold_count || (object = m->object) == NULL) { 826 vm_page_unlock(m); 827 vm_page_requeue(m); 828 addl_page_shortage++; 829 continue; 830 } 831 832 /* 833 * Don't mess with busy pages, keep in the front of the 834 * queue, most likely are being paged out. 835 */ 836 if (!VM_OBJECT_TRYLOCK(object) && 837 (!vm_pageout_fallback_object_lock(m, &next) || 838 m->hold_count != 0)) { 839 VM_OBJECT_UNLOCK(object); 840 vm_page_unlock(m); 841 addl_page_shortage++; 842 continue; 843 } 844 if (m->busy || (m->oflags & VPO_BUSY)) { 845 vm_page_unlock(m); 846 VM_OBJECT_UNLOCK(object); 847 addl_page_shortage++; 848 continue; 849 } 850 851 /* 852 * If the object is not being used, we ignore previous 853 * references. 854 */ 855 if (object->ref_count == 0) { 856 vm_page_flag_clear(m, PG_REFERENCED); 857 KASSERT(!pmap_page_is_mapped(m), 858 ("vm_pageout_scan: page %p is mapped", m)); 859 860 /* 861 * Otherwise, if the page has been referenced while in the 862 * inactive queue, we bump the "activation count" upwards, 863 * making it less likely that the page will be added back to 864 * the inactive queue prematurely again. Here we check the 865 * page tables (or emulated bits, if any), given the upper 866 * level VM system not knowing anything about existing 867 * references. 868 */ 869 } else if (((m->flags & PG_REFERENCED) == 0) && 870 (actcount = pmap_ts_referenced(m))) { 871 vm_page_activate(m); 872 VM_OBJECT_UNLOCK(object); 873 m->act_count += (actcount + ACT_ADVANCE); 874 vm_page_unlock(m); 875 continue; 876 } 877 878 /* 879 * If the upper level VM system knows about any page 880 * references, we activate the page. We also set the 881 * "activation count" higher than normal so that we will less 882 * likely place pages back onto the inactive queue again. 883 */ 884 if ((m->flags & PG_REFERENCED) != 0) { 885 vm_page_flag_clear(m, PG_REFERENCED); 886 actcount = pmap_ts_referenced(m); 887 vm_page_activate(m); 888 VM_OBJECT_UNLOCK(object); 889 m->act_count += (actcount + ACT_ADVANCE + 1); 890 vm_page_unlock(m); 891 continue; 892 } 893 894 /* 895 * If the upper level VM system does not believe that the page 896 * is fully dirty, but it is mapped for write access, then we 897 * consult the pmap to see if the page's dirty status should 898 * be updated. 899 */ 900 if (m->dirty != VM_PAGE_BITS_ALL && 901 (m->flags & PG_WRITEABLE) != 0) { 902 /* 903 * Avoid a race condition: Unless write access is 904 * removed from the page, another processor could 905 * modify it before all access is removed by the call 906 * to vm_page_cache() below. If vm_page_cache() finds 907 * that the page has been modified when it removes all 908 * access, it panics because it cannot cache dirty 909 * pages. In principle, we could eliminate just write 910 * access here rather than all access. In the expected 911 * case, when there are no last instant modifications 912 * to the page, removing all access will be cheaper 913 * overall. 914 */ 915 if (pmap_is_modified(m)) 916 vm_page_dirty(m); 917 else if (m->dirty == 0) 918 pmap_remove_all(m); 919 } 920 921 if (m->valid == 0) { 922 /* 923 * Invalid pages can be easily freed 924 */ 925 vm_page_free(m); 926 cnt.v_dfree++; 927 --page_shortage; 928 } else if (m->dirty == 0) { 929 /* 930 * Clean pages can be placed onto the cache queue. 931 * This effectively frees them. 932 */ 933 vm_page_cache(m); 934 --page_shortage; 935 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) { 936 /* 937 * Dirty pages need to be paged out, but flushing 938 * a page is extremely expensive verses freeing 939 * a clean page. Rather then artificially limiting 940 * the number of pages we can flush, we instead give 941 * dirty pages extra priority on the inactive queue 942 * by forcing them to be cycled through the queue 943 * twice before being flushed, after which the 944 * (now clean) page will cycle through once more 945 * before being freed. This significantly extends 946 * the thrash point for a heavily loaded machine. 947 */ 948 vm_page_flag_set(m, PG_WINATCFLS); 949 vm_page_requeue(m); 950 } else if (maxlaunder > 0) { 951 /* 952 * We always want to try to flush some dirty pages if 953 * we encounter them, to keep the system stable. 954 * Normally this number is small, but under extreme 955 * pressure where there are insufficient clean pages 956 * on the inactive queue, we may have to go all out. 957 */ 958 int swap_pageouts_ok, vfslocked = 0; 959 struct vnode *vp = NULL; 960 struct mount *mp = NULL; 961 962 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 963 swap_pageouts_ok = 1; 964 } else { 965 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 966 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 967 vm_page_count_min()); 968 969 } 970 971 /* 972 * We don't bother paging objects that are "dead". 973 * Those objects are in a "rundown" state. 974 */ 975 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 976 vm_page_unlock(m); 977 VM_OBJECT_UNLOCK(object); 978 vm_page_requeue(m); 979 continue; 980 } 981 982 /* 983 * Following operations may unlock 984 * vm_page_queue_mtx, invalidating the 'next' 985 * pointer. To prevent an inordinate number 986 * of restarts we use our marker to remember 987 * our place. 988 * 989 */ 990 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, 991 m, &marker, pageq); 992 /* 993 * The object is already known NOT to be dead. It 994 * is possible for the vget() to block the whole 995 * pageout daemon, but the new low-memory handling 996 * code should prevent it. 997 * 998 * The previous code skipped locked vnodes and, worse, 999 * reordered pages in the queue. This results in 1000 * completely non-deterministic operation and, on a 1001 * busy system, can lead to extremely non-optimal 1002 * pageouts. For example, it can cause clean pages 1003 * to be freed and dirty pages to be moved to the end 1004 * of the queue. Since dirty pages are also moved to 1005 * the end of the queue once-cleaned, this gives 1006 * way too large a weighting to defering the freeing 1007 * of dirty pages. 1008 * 1009 * We can't wait forever for the vnode lock, we might 1010 * deadlock due to a vn_read() getting stuck in 1011 * vm_wait while holding this vnode. We skip the 1012 * vnode if we can't get it in a reasonable amount 1013 * of time. 1014 */ 1015 if (object->type == OBJT_VNODE) { 1016 vm_page_unlock_queues(); 1017 vm_page_unlock(m); 1018 vp = object->handle; 1019 if (vp->v_type == VREG && 1020 vn_start_write(vp, &mp, V_NOWAIT) != 0) { 1021 mp = NULL; 1022 ++pageout_lock_miss; 1023 if (object->flags & OBJ_MIGHTBEDIRTY) 1024 vnodes_skipped++; 1025 vm_page_lock_queues(); 1026 goto unlock_and_continue; 1027 } 1028 KASSERT(mp != NULL, 1029 ("vp %p with NULL v_mount", vp)); 1030 vm_object_reference_locked(object); 1031 VM_OBJECT_UNLOCK(object); 1032 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 1033 if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK, 1034 curthread)) { 1035 VM_OBJECT_LOCK(object); 1036 vm_page_lock_queues(); 1037 ++pageout_lock_miss; 1038 if (object->flags & OBJ_MIGHTBEDIRTY) 1039 vnodes_skipped++; 1040 vp = NULL; 1041 goto unlock_and_continue; 1042 } 1043 VM_OBJECT_LOCK(object); 1044 vm_page_lock(m); 1045 vm_page_lock_queues(); 1046 /* 1047 * The page might have been moved to another 1048 * queue during potential blocking in vget() 1049 * above. The page might have been freed and 1050 * reused for another vnode. 1051 */ 1052 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE || 1053 m->object != object || 1054 TAILQ_NEXT(m, pageq) != &marker) { 1055 vm_page_unlock(m); 1056 if (object->flags & OBJ_MIGHTBEDIRTY) 1057 vnodes_skipped++; 1058 goto unlock_and_continue; 1059 } 1060 1061 /* 1062 * The page may have been busied during the 1063 * blocking in vget(). We don't move the 1064 * page back onto the end of the queue so that 1065 * statistics are more correct if we don't. 1066 */ 1067 if (m->busy || (m->oflags & VPO_BUSY)) { 1068 vm_page_unlock(m); 1069 goto unlock_and_continue; 1070 } 1071 1072 /* 1073 * If the page has become held it might 1074 * be undergoing I/O, so skip it 1075 */ 1076 if (m->hold_count) { 1077 vm_page_unlock(m); 1078 vm_page_requeue(m); 1079 if (object->flags & OBJ_MIGHTBEDIRTY) 1080 vnodes_skipped++; 1081 goto unlock_and_continue; 1082 } 1083 } 1084 vm_page_unlock(m); 1085 1086 /* 1087 * If a page is dirty, then it is either being washed 1088 * (but not yet cleaned) or it is still in the 1089 * laundry. If it is still in the laundry, then we 1090 * start the cleaning operation. 1091 * 1092 * decrement page_shortage on success to account for 1093 * the (future) cleaned page. Otherwise we could wind 1094 * up laundering or cleaning too many pages. 1095 */ 1096 vm_page_unlock_queues(); 1097 if (vm_pageout_clean(m) != 0) { 1098 --page_shortage; 1099 --maxlaunder; 1100 } 1101 vm_page_lock_queues(); 1102unlock_and_continue: 1103 vm_page_lock_assert(m, MA_NOTOWNED); 1104 VM_OBJECT_UNLOCK(object); 1105 if (mp != NULL) { 1106 vm_page_unlock_queues(); 1107 if (vp != NULL) 1108 vput(vp); 1109 VFS_UNLOCK_GIANT(vfslocked); 1110 vm_object_deallocate(object); 1111 vn_finished_write(mp); 1112 vm_page_lock_queues(); 1113 } 1114 next = TAILQ_NEXT(&marker, pageq); 1115 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, 1116 &marker, pageq); 1117 vm_page_lock_assert(m, MA_NOTOWNED); 1118 continue; 1119 } 1120 vm_page_unlock(m); 1121 VM_OBJECT_UNLOCK(object); 1122 } 1123 1124 /* 1125 * Compute the number of pages we want to try to move from the 1126 * active queue to the inactive queue. 1127 */ 1128 page_shortage = vm_paging_target() + 1129 cnt.v_inactive_target - cnt.v_inactive_count; 1130 page_shortage += addl_page_shortage; 1131 1132 /* 1133 * Scan the active queue for things we can deactivate. We nominally 1134 * track the per-page activity counter and use it to locate 1135 * deactivation candidates. 1136 */ 1137 pcount = cnt.v_active_count; 1138 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1139 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1140 1141 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) { 1142 1143 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE), 1144 ("vm_pageout_scan: page %p isn't active", m)); 1145 1146 next = TAILQ_NEXT(m, pageq); 1147 object = m->object; 1148 if ((m->flags & PG_MARKER) != 0) { 1149 m = next; 1150 continue; 1151 } 1152 if (!vm_pageout_page_lock(m, &next) || 1153 (object = m->object) == NULL) { 1154 vm_page_unlock(m); 1155 m = next; 1156 continue; 1157 } 1158 if (!VM_OBJECT_TRYLOCK(object) && 1159 !vm_pageout_fallback_object_lock(m, &next)) { 1160 VM_OBJECT_UNLOCK(object); 1161 vm_page_unlock(m); 1162 m = next; 1163 continue; 1164 } 1165 1166 /* 1167 * Don't deactivate pages that are busy. 1168 */ 1169 if ((m->busy != 0) || 1170 (m->oflags & VPO_BUSY) || 1171 (m->hold_count != 0)) { 1172 vm_page_unlock(m); 1173 VM_OBJECT_UNLOCK(object); 1174 vm_page_requeue(m); 1175 m = next; 1176 continue; 1177 } 1178 1179 /* 1180 * The count for pagedaemon pages is done after checking the 1181 * page for eligibility... 1182 */ 1183 cnt.v_pdpages++; 1184 1185 /* 1186 * Check to see "how much" the page has been used. 1187 */ 1188 actcount = 0; 1189 if (object->ref_count != 0) { 1190 if (m->flags & PG_REFERENCED) { 1191 actcount += 1; 1192 } 1193 actcount += pmap_ts_referenced(m); 1194 if (actcount) { 1195 m->act_count += ACT_ADVANCE + actcount; 1196 if (m->act_count > ACT_MAX) 1197 m->act_count = ACT_MAX; 1198 } 1199 } 1200 1201 /* 1202 * Since we have "tested" this bit, we need to clear it now. 1203 */ 1204 vm_page_flag_clear(m, PG_REFERENCED); 1205 1206 /* 1207 * Only if an object is currently being used, do we use the 1208 * page activation count stats. 1209 */ 1210 if (actcount && (object->ref_count != 0)) { 1211 vm_page_requeue(m); 1212 } else { 1213 m->act_count -= min(m->act_count, ACT_DECLINE); 1214 if (vm_pageout_algorithm || 1215 object->ref_count == 0 || 1216 m->act_count == 0) { 1217 page_shortage--; 1218 if (object->ref_count == 0) { 1219 KASSERT(!pmap_page_is_mapped(m), 1220 ("vm_pageout_scan: page %p is mapped", m)); 1221 if (m->dirty == 0) 1222 vm_page_cache(m); 1223 else 1224 vm_page_deactivate(m); 1225 } else { 1226 vm_page_deactivate(m); 1227 } 1228 } else { 1229 vm_page_requeue(m); 1230 } 1231 } 1232 vm_page_unlock(m); 1233 VM_OBJECT_UNLOCK(object); 1234 m = next; 1235 } 1236 vm_page_unlock_queues(); 1237#if !defined(NO_SWAPPING) 1238 /* 1239 * Idle process swapout -- run once per second. 1240 */ 1241 if (vm_swap_idle_enabled) { 1242 static long lsec; 1243 if (time_second != lsec) { 1244 vm_req_vmdaemon(VM_SWAP_IDLE); 1245 lsec = time_second; 1246 } 1247 } 1248#endif 1249 1250 /* 1251 * If we didn't get enough free pages, and we have skipped a vnode 1252 * in a writeable object, wakeup the sync daemon. And kick swapout 1253 * if we did not get enough free pages. 1254 */ 1255 if (vm_paging_target() > 0) { 1256 if (vnodes_skipped && vm_page_count_min()) 1257 (void) speedup_syncer(); 1258#if !defined(NO_SWAPPING) 1259 if (vm_swap_enabled && vm_page_count_target()) 1260 vm_req_vmdaemon(VM_SWAP_NORMAL); 1261#endif 1262 } 1263 1264 /* 1265 * If we are critically low on one of RAM or swap and low on 1266 * the other, kill the largest process. However, we avoid 1267 * doing this on the first pass in order to give ourselves a 1268 * chance to flush out dirty vnode-backed pages and to allow 1269 * active pages to be moved to the inactive queue and reclaimed. 1270 */ 1271 if (pass != 0 && 1272 ((swap_pager_avail < 64 && vm_page_count_min()) || 1273 (swap_pager_full && vm_paging_target() > 0))) 1274 vm_pageout_oom(VM_OOM_MEM); 1275} 1276 1277 1278void 1279vm_pageout_oom(int shortage) 1280{ 1281 struct proc *p, *bigproc; 1282 vm_offset_t size, bigsize; 1283 struct thread *td; 1284 struct vmspace *vm; 1285 1286 /* 1287 * We keep the process bigproc locked once we find it to keep anyone 1288 * from messing with it; however, there is a possibility of 1289 * deadlock if process B is bigproc and one of it's child processes 1290 * attempts to propagate a signal to B while we are waiting for A's 1291 * lock while walking this list. To avoid this, we don't block on 1292 * the process lock but just skip a process if it is already locked. 1293 */ 1294 bigproc = NULL; 1295 bigsize = 0; 1296 sx_slock(&allproc_lock); 1297 FOREACH_PROC_IN_SYSTEM(p) { 1298 int breakout; 1299 1300 if (PROC_TRYLOCK(p) == 0) 1301 continue; 1302 /* 1303 * If this is a system, protected or killed process, skip it. 1304 */ 1305 if ((p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM)) || 1306 (p->p_pid == 1) || P_KILLED(p) || 1307 ((p->p_pid < 48) && (swap_pager_avail != 0))) { 1308 PROC_UNLOCK(p); 1309 continue; 1310 } 1311 /* 1312 * If the process is in a non-running type state, 1313 * don't touch it. Check all the threads individually. 1314 */ 1315 breakout = 0; 1316 FOREACH_THREAD_IN_PROC(p, td) { 1317 thread_lock(td); 1318 if (!TD_ON_RUNQ(td) && 1319 !TD_IS_RUNNING(td) && 1320 !TD_IS_SLEEPING(td)) { 1321 thread_unlock(td); 1322 breakout = 1; 1323 break; 1324 } 1325 thread_unlock(td); 1326 } 1327 if (breakout) { 1328 PROC_UNLOCK(p); 1329 continue; 1330 } 1331 /* 1332 * get the process size 1333 */ 1334 vm = vmspace_acquire_ref(p); 1335 if (vm == NULL) { 1336 PROC_UNLOCK(p); 1337 continue; 1338 } 1339 if (!vm_map_trylock_read(&vm->vm_map)) { 1340 vmspace_free(vm); 1341 PROC_UNLOCK(p); 1342 continue; 1343 } 1344 size = vmspace_swap_count(vm); 1345 vm_map_unlock_read(&vm->vm_map); 1346 if (shortage == VM_OOM_MEM) 1347 size += vmspace_resident_count(vm); 1348 vmspace_free(vm); 1349 /* 1350 * if the this process is bigger than the biggest one 1351 * remember it. 1352 */ 1353 if (size > bigsize) { 1354 if (bigproc != NULL) 1355 PROC_UNLOCK(bigproc); 1356 bigproc = p; 1357 bigsize = size; 1358 } else 1359 PROC_UNLOCK(p); 1360 } 1361 sx_sunlock(&allproc_lock); 1362 if (bigproc != NULL) { 1363 killproc(bigproc, "out of swap space"); 1364 sched_nice(bigproc, PRIO_MIN); 1365 PROC_UNLOCK(bigproc); 1366 wakeup(&cnt.v_free_count); 1367 } 1368} 1369 1370/* 1371 * This routine tries to maintain the pseudo LRU active queue, 1372 * so that during long periods of time where there is no paging, 1373 * that some statistic accumulation still occurs. This code 1374 * helps the situation where paging just starts to occur. 1375 */ 1376static void 1377vm_pageout_page_stats() 1378{ 1379 vm_object_t object; 1380 vm_page_t m,next; 1381 int pcount,tpcount; /* Number of pages to check */ 1382 static int fullintervalcount = 0; 1383 int page_shortage; 1384 1385 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1386 page_shortage = 1387 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) - 1388 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count); 1389 1390 if (page_shortage <= 0) 1391 return; 1392 1393 pcount = cnt.v_active_count; 1394 fullintervalcount += vm_pageout_stats_interval; 1395 if (fullintervalcount < vm_pageout_full_stats_interval) { 1396 tpcount = (int64_t)vm_pageout_stats_max * cnt.v_active_count / 1397 cnt.v_page_count; 1398 if (pcount > tpcount) 1399 pcount = tpcount; 1400 } else { 1401 fullintervalcount = 0; 1402 } 1403 1404 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1405 while ((m != NULL) && (pcount-- > 0)) { 1406 int actcount; 1407 1408 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE), 1409 ("vm_pageout_page_stats: page %p isn't active", m)); 1410 1411 next = TAILQ_NEXT(m, pageq); 1412 if ((m->flags & PG_MARKER) != 0) { 1413 m = next; 1414 continue; 1415 } 1416 vm_page_lock_assert(m, MA_NOTOWNED); 1417 if (!vm_pageout_page_lock(m, &next) || 1418 (object = m->object) == NULL) { 1419 vm_page_unlock(m); 1420 m = next; 1421 continue; 1422 } 1423 if (!VM_OBJECT_TRYLOCK(object) && 1424 !vm_pageout_fallback_object_lock(m, &next)) { 1425 VM_OBJECT_UNLOCK(object); 1426 vm_page_unlock(m); 1427 m = next; 1428 continue; 1429 } 1430 1431 /* 1432 * Don't deactivate pages that are busy. 1433 */ 1434 if ((m->busy != 0) || 1435 (m->oflags & VPO_BUSY) || 1436 (m->hold_count != 0)) { 1437 vm_page_unlock(m); 1438 VM_OBJECT_UNLOCK(object); 1439 vm_page_requeue(m); 1440 m = next; 1441 continue; 1442 } 1443 1444 actcount = 0; 1445 if (m->flags & PG_REFERENCED) { 1446 vm_page_flag_clear(m, PG_REFERENCED); 1447 actcount += 1; 1448 } 1449 1450 actcount += pmap_ts_referenced(m); 1451 if (actcount) { 1452 m->act_count += ACT_ADVANCE + actcount; 1453 if (m->act_count > ACT_MAX) 1454 m->act_count = ACT_MAX; 1455 vm_page_requeue(m); 1456 } else { 1457 if (m->act_count == 0) { 1458 /* 1459 * We turn off page access, so that we have 1460 * more accurate RSS stats. We don't do this 1461 * in the normal page deactivation when the 1462 * system is loaded VM wise, because the 1463 * cost of the large number of page protect 1464 * operations would be higher than the value 1465 * of doing the operation. 1466 */ 1467 pmap_remove_all(m); 1468 vm_page_deactivate(m); 1469 } else { 1470 m->act_count -= min(m->act_count, ACT_DECLINE); 1471 vm_page_requeue(m); 1472 } 1473 } 1474 vm_page_unlock(m); 1475 VM_OBJECT_UNLOCK(object); 1476 m = next; 1477 } 1478} 1479 1480/* 1481 * vm_pageout is the high level pageout daemon. 1482 */ 1483static void 1484vm_pageout() 1485{ 1486 int error, pass; 1487 1488 /* 1489 * Initialize some paging parameters. 1490 */ 1491 cnt.v_interrupt_free_min = 2; 1492 if (cnt.v_page_count < 2000) 1493 vm_pageout_page_count = 8; 1494 1495 /* 1496 * v_free_reserved needs to include enough for the largest 1497 * swap pager structures plus enough for any pv_entry structs 1498 * when paging. 1499 */ 1500 if (cnt.v_page_count > 1024) 1501 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200; 1502 else 1503 cnt.v_free_min = 4; 1504 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1505 cnt.v_interrupt_free_min; 1506 cnt.v_free_reserved = vm_pageout_page_count + 1507 cnt.v_pageout_free_min + (cnt.v_page_count / 768); 1508 cnt.v_free_severe = cnt.v_free_min / 2; 1509 cnt.v_free_min += cnt.v_free_reserved; 1510 cnt.v_free_severe += cnt.v_free_reserved; 1511 1512 /* 1513 * v_free_target and v_cache_min control pageout hysteresis. Note 1514 * that these are more a measure of the VM cache queue hysteresis 1515 * then the VM free queue. Specifically, v_free_target is the 1516 * high water mark (free+cache pages). 1517 * 1518 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the 1519 * low water mark, while v_free_min is the stop. v_cache_min must 1520 * be big enough to handle memory needs while the pageout daemon 1521 * is signalled and run to free more pages. 1522 */ 1523 if (cnt.v_free_count > 6144) 1524 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved; 1525 else 1526 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved; 1527 1528 if (cnt.v_free_count > 2048) { 1529 cnt.v_cache_min = cnt.v_free_target; 1530 cnt.v_cache_max = 2 * cnt.v_cache_min; 1531 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2; 1532 } else { 1533 cnt.v_cache_min = 0; 1534 cnt.v_cache_max = 0; 1535 cnt.v_inactive_target = cnt.v_free_count / 4; 1536 } 1537 if (cnt.v_inactive_target > cnt.v_free_count / 3) 1538 cnt.v_inactive_target = cnt.v_free_count / 3; 1539 1540 /* XXX does not really belong here */ 1541 if (vm_page_max_wired == 0) 1542 vm_page_max_wired = cnt.v_free_count / 3; 1543 1544 if (vm_pageout_stats_max == 0) 1545 vm_pageout_stats_max = cnt.v_free_target; 1546 1547 /* 1548 * Set interval in seconds for stats scan. 1549 */ 1550 if (vm_pageout_stats_interval == 0) 1551 vm_pageout_stats_interval = 5; 1552 if (vm_pageout_full_stats_interval == 0) 1553 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4; 1554 1555 swap_pager_swap_init(); 1556 pass = 0; 1557 /* 1558 * The pageout daemon is never done, so loop forever. 1559 */ 1560 while (TRUE) { 1561 /* 1562 * If we have enough free memory, wakeup waiters. Do 1563 * not clear vm_pages_needed until we reach our target, 1564 * otherwise we may be woken up over and over again and 1565 * waste a lot of cpu. 1566 */ 1567 mtx_lock(&vm_page_queue_free_mtx); 1568 if (vm_pages_needed && !vm_page_count_min()) { 1569 if (!vm_paging_needed()) 1570 vm_pages_needed = 0; 1571 wakeup(&cnt.v_free_count); 1572 } 1573 if (vm_pages_needed) { 1574 /* 1575 * Still not done, take a second pass without waiting 1576 * (unlimited dirty cleaning), otherwise sleep a bit 1577 * and try again. 1578 */ 1579 ++pass; 1580 if (pass > 1) 1581 msleep(&vm_pages_needed, 1582 &vm_page_queue_free_mtx, PVM, "psleep", 1583 hz / 2); 1584 } else { 1585 /* 1586 * Good enough, sleep & handle stats. Prime the pass 1587 * for the next run. 1588 */ 1589 if (pass > 1) 1590 pass = 1; 1591 else 1592 pass = 0; 1593 error = msleep(&vm_pages_needed, 1594 &vm_page_queue_free_mtx, PVM, "psleep", 1595 vm_pageout_stats_interval * hz); 1596 if (error && !vm_pages_needed) { 1597 mtx_unlock(&vm_page_queue_free_mtx); 1598 pass = 0; 1599 vm_page_lock_queues(); 1600 vm_pageout_page_stats(); 1601 vm_page_unlock_queues(); 1602 continue; 1603 } 1604 } 1605 if (vm_pages_needed) 1606 cnt.v_pdwakeups++; 1607 mtx_unlock(&vm_page_queue_free_mtx); 1608 vm_pageout_scan(pass); 1609 } 1610} 1611 1612/* 1613 * Unless the free page queue lock is held by the caller, this function 1614 * should be regarded as advisory. Specifically, the caller should 1615 * not msleep() on &cnt.v_free_count following this function unless 1616 * the free page queue lock is held until the msleep() is performed. 1617 */ 1618void 1619pagedaemon_wakeup() 1620{ 1621 1622 if (!vm_pages_needed && curthread->td_proc != pageproc) { 1623 vm_pages_needed = 1; 1624 wakeup(&vm_pages_needed); 1625 } 1626} 1627 1628#if !defined(NO_SWAPPING) 1629static void 1630vm_req_vmdaemon(int req) 1631{ 1632 static int lastrun = 0; 1633 1634 mtx_lock(&vm_daemon_mtx); 1635 vm_pageout_req_swapout |= req; 1636 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1637 wakeup(&vm_daemon_needed); 1638 lastrun = ticks; 1639 } 1640 mtx_unlock(&vm_daemon_mtx); 1641} 1642 1643static void 1644vm_daemon() 1645{ 1646 struct rlimit rsslim; 1647 struct proc *p; 1648 struct thread *td; 1649 struct vmspace *vm; 1650 int breakout, swapout_flags; 1651 1652 while (TRUE) { 1653 mtx_lock(&vm_daemon_mtx); 1654 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0); 1655 swapout_flags = vm_pageout_req_swapout; 1656 vm_pageout_req_swapout = 0; 1657 mtx_unlock(&vm_daemon_mtx); 1658 if (swapout_flags) 1659 swapout_procs(swapout_flags); 1660 1661 /* 1662 * scan the processes for exceeding their rlimits or if 1663 * process is swapped out -- deactivate pages 1664 */ 1665 sx_slock(&allproc_lock); 1666 FOREACH_PROC_IN_SYSTEM(p) { 1667 vm_pindex_t limit, size; 1668 1669 /* 1670 * if this is a system process or if we have already 1671 * looked at this process, skip it. 1672 */ 1673 PROC_LOCK(p); 1674 if (p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) { 1675 PROC_UNLOCK(p); 1676 continue; 1677 } 1678 /* 1679 * if the process is in a non-running type state, 1680 * don't touch it. 1681 */ 1682 breakout = 0; 1683 FOREACH_THREAD_IN_PROC(p, td) { 1684 thread_lock(td); 1685 if (!TD_ON_RUNQ(td) && 1686 !TD_IS_RUNNING(td) && 1687 !TD_IS_SLEEPING(td)) { 1688 thread_unlock(td); 1689 breakout = 1; 1690 break; 1691 } 1692 thread_unlock(td); 1693 } 1694 if (breakout) { 1695 PROC_UNLOCK(p); 1696 continue; 1697 } 1698 /* 1699 * get a limit 1700 */ 1701 lim_rlimit(p, RLIMIT_RSS, &rsslim); 1702 limit = OFF_TO_IDX( 1703 qmin(rsslim.rlim_cur, rsslim.rlim_max)); 1704 1705 /* 1706 * let processes that are swapped out really be 1707 * swapped out set the limit to nothing (will force a 1708 * swap-out.) 1709 */ 1710 if ((p->p_flag & P_INMEM) == 0) 1711 limit = 0; /* XXX */ 1712 vm = vmspace_acquire_ref(p); 1713 PROC_UNLOCK(p); 1714 if (vm == NULL) 1715 continue; 1716 1717 size = vmspace_resident_count(vm); 1718 if (limit >= 0 && size >= limit) { 1719 vm_pageout_map_deactivate_pages( 1720 &vm->vm_map, limit); 1721 } 1722 vmspace_free(vm); 1723 } 1724 sx_sunlock(&allproc_lock); 1725 } 1726} 1727#endif /* !defined(NO_SWAPPING) */ 1728