vm_pageout.c revision 172207
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 172207 2007-09-17 05:31:39Z jeff $"); 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#include <machine/mutex.h> 109 110/* 111 * System initialization 112 */ 113 114/* the kernel process "vm_pageout"*/ 115static void vm_pageout(void); 116static int vm_pageout_clean(vm_page_t); 117static void vm_pageout_scan(int pass); 118 119struct proc *pageproc; 120 121static struct kproc_desc page_kp = { 122 "pagedaemon", 123 vm_pageout, 124 &pageproc 125}; 126SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp) 127 128#if !defined(NO_SWAPPING) 129/* the kernel process "vm_daemon"*/ 130static void vm_daemon(void); 131static struct proc *vmproc; 132 133static struct kproc_desc vm_kp = { 134 "vmdaemon", 135 vm_daemon, 136 &vmproc 137}; 138SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp) 139#endif 140 141 142int vm_pages_needed; /* Event on which pageout daemon sleeps */ 143int vm_pageout_deficit; /* Estimated number of pages deficit */ 144int vm_pageout_pages_needed; /* flag saying that the pageout daemon needs pages */ 145 146#if !defined(NO_SWAPPING) 147static int vm_pageout_req_swapout; /* XXX */ 148static int vm_daemon_needed; 149static struct mtx vm_daemon_mtx; 150/* Allow for use by vm_pageout before vm_daemon is initialized. */ 151MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF); 152#endif 153static int vm_max_launder = 32; 154static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0; 155static int vm_pageout_full_stats_interval = 0; 156static int vm_pageout_algorithm=0; 157static int defer_swap_pageouts=0; 158static int disable_swap_pageouts=0; 159 160#if defined(NO_SWAPPING) 161static int vm_swap_enabled=0; 162static int vm_swap_idle_enabled=0; 163#else 164static int vm_swap_enabled=1; 165static int vm_swap_idle_enabled=0; 166#endif 167 168SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm, 169 CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt"); 170 171SYSCTL_INT(_vm, OID_AUTO, max_launder, 172 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout"); 173 174SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max, 175 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length"); 176 177SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval, 178 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan"); 179 180SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval, 181 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan"); 182 183#if defined(NO_SWAPPING) 184SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 185 CTLFLAG_RD, &vm_swap_enabled, 0, ""); 186SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 187 CTLFLAG_RD, &vm_swap_idle_enabled, 0, ""); 188#else 189SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 190 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout"); 191SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 192 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 193#endif 194 195SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts, 196 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem"); 197 198SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts, 199 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages"); 200 201static int pageout_lock_miss; 202SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss, 203 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout"); 204 205#define VM_PAGEOUT_PAGE_COUNT 16 206int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT; 207 208int vm_page_max_wired; /* XXX max # of wired pages system-wide */ 209 210#if !defined(NO_SWAPPING) 211static void vm_pageout_map_deactivate_pages(vm_map_t, long); 212static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long); 213static void vm_req_vmdaemon(int req); 214#endif 215static void vm_pageout_page_stats(void); 216 217/* 218 * vm_pageout_fallback_object_lock: 219 * 220 * Lock vm object currently associated with `m'. VM_OBJECT_TRYLOCK is 221 * known to have failed and page queue must be either PQ_ACTIVE or 222 * PQ_INACTIVE. To avoid lock order violation, unlock the page queues 223 * while locking the vm object. Use marker page to detect page queue 224 * changes and maintain notion of next page on page queue. Return 225 * TRUE if no changes were detected, FALSE otherwise. vm object is 226 * locked on return. 227 * 228 * This function depends on both the lock portion of struct vm_object 229 * and normal struct vm_page being type stable. 230 */ 231static boolean_t 232vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next) 233{ 234 struct vm_page marker; 235 boolean_t unchanged; 236 u_short queue; 237 vm_object_t object; 238 239 /* 240 * Initialize our marker 241 */ 242 bzero(&marker, sizeof(marker)); 243 marker.flags = PG_FICTITIOUS | PG_MARKER; 244 marker.oflags = VPO_BUSY; 245 marker.queue = m->queue; 246 marker.wire_count = 1; 247 248 queue = m->queue; 249 object = m->object; 250 251 TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl, 252 m, &marker, pageq); 253 vm_page_unlock_queues(); 254 VM_OBJECT_LOCK(object); 255 vm_page_lock_queues(); 256 257 /* Page queue might have changed. */ 258 *next = TAILQ_NEXT(&marker, pageq); 259 unchanged = (m->queue == queue && 260 m->object == object && 261 &marker == TAILQ_NEXT(m, pageq)); 262 TAILQ_REMOVE(&vm_page_queues[queue].pl, 263 &marker, pageq); 264 return (unchanged); 265} 266 267/* 268 * vm_pageout_clean: 269 * 270 * Clean the page and remove it from the laundry. 271 * 272 * We set the busy bit to cause potential page faults on this page to 273 * block. Note the careful timing, however, the busy bit isn't set till 274 * late and we cannot do anything that will mess with the page. 275 */ 276static int 277vm_pageout_clean(m) 278 vm_page_t m; 279{ 280 vm_object_t object; 281 vm_page_t mc[2*vm_pageout_page_count]; 282 int pageout_count; 283 int ib, is, page_base; 284 vm_pindex_t pindex = m->pindex; 285 286 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 287 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 288 289 /* 290 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP 291 * with the new swapper, but we could have serious problems paging 292 * out other object types if there is insufficient memory. 293 * 294 * Unfortunately, checking free memory here is far too late, so the 295 * check has been moved up a procedural level. 296 */ 297 298 /* 299 * Can't clean the page if it's busy or held. 300 */ 301 if ((m->hold_count != 0) || 302 ((m->busy != 0) || (m->oflags & VPO_BUSY))) { 303 return 0; 304 } 305 306 mc[vm_pageout_page_count] = m; 307 pageout_count = 1; 308 page_base = vm_pageout_page_count; 309 ib = 1; 310 is = 1; 311 312 /* 313 * Scan object for clusterable pages. 314 * 315 * We can cluster ONLY if: ->> the page is NOT 316 * clean, wired, busy, held, or mapped into a 317 * buffer, and one of the following: 318 * 1) The page is inactive, or a seldom used 319 * active page. 320 * -or- 321 * 2) we force the issue. 322 * 323 * During heavy mmap/modification loads the pageout 324 * daemon can really fragment the underlying file 325 * due to flushing pages out of order and not trying 326 * align the clusters (which leave sporatic out-of-order 327 * holes). To solve this problem we do the reverse scan 328 * first and attempt to align our cluster, then do a 329 * forward scan if room remains. 330 */ 331 object = m->object; 332more: 333 while (ib && pageout_count < vm_pageout_page_count) { 334 vm_page_t p; 335 336 if (ib > pindex) { 337 ib = 0; 338 break; 339 } 340 341 if ((p = vm_page_lookup(object, pindex - ib)) == NULL) { 342 ib = 0; 343 break; 344 } 345 if (VM_PAGE_INQUEUE1(p, PQ_CACHE) || 346 (p->oflags & VPO_BUSY) || p->busy) { 347 ib = 0; 348 break; 349 } 350 vm_page_test_dirty(p); 351 if ((p->dirty & p->valid) == 0 || 352 p->queue != PQ_INACTIVE || 353 p->wire_count != 0 || /* may be held by buf cache */ 354 p->hold_count != 0) { /* may be undergoing I/O */ 355 ib = 0; 356 break; 357 } 358 mc[--page_base] = p; 359 ++pageout_count; 360 ++ib; 361 /* 362 * alignment boundry, stop here and switch directions. Do 363 * not clear ib. 364 */ 365 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0) 366 break; 367 } 368 369 while (pageout_count < vm_pageout_page_count && 370 pindex + is < object->size) { 371 vm_page_t p; 372 373 if ((p = vm_page_lookup(object, pindex + is)) == NULL) 374 break; 375 if (VM_PAGE_INQUEUE1(p, PQ_CACHE) || 376 (p->oflags & VPO_BUSY) || p->busy) { 377 break; 378 } 379 vm_page_test_dirty(p); 380 if ((p->dirty & p->valid) == 0 || 381 p->queue != PQ_INACTIVE || 382 p->wire_count != 0 || /* may be held by buf cache */ 383 p->hold_count != 0) { /* may be undergoing I/O */ 384 break; 385 } 386 mc[page_base + pageout_count] = p; 387 ++pageout_count; 388 ++is; 389 } 390 391 /* 392 * If we exhausted our forward scan, continue with the reverse scan 393 * when possible, even past a page boundry. This catches boundry 394 * conditions. 395 */ 396 if (ib && pageout_count < vm_pageout_page_count) 397 goto more; 398 399 /* 400 * we allow reads during pageouts... 401 */ 402 return (vm_pageout_flush(&mc[page_base], pageout_count, 0)); 403} 404 405/* 406 * vm_pageout_flush() - launder the given pages 407 * 408 * The given pages are laundered. Note that we setup for the start of 409 * I/O ( i.e. busy the page ), mark it read-only, and bump the object 410 * reference count all in here rather then in the parent. If we want 411 * the parent to do more sophisticated things we may have to change 412 * the ordering. 413 */ 414int 415vm_pageout_flush(vm_page_t *mc, int count, int flags) 416{ 417 vm_object_t object = mc[0]->object; 418 int pageout_status[count]; 419 int numpagedout = 0; 420 int i; 421 422 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 423 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 424 /* 425 * Initiate I/O. Bump the vm_page_t->busy counter and 426 * mark the pages read-only. 427 * 428 * We do not have to fixup the clean/dirty bits here... we can 429 * allow the pager to do it after the I/O completes. 430 * 431 * NOTE! mc[i]->dirty may be partial or fragmented due to an 432 * edge case with file fragments. 433 */ 434 for (i = 0; i < count; i++) { 435 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, 436 ("vm_pageout_flush: partially invalid page %p index %d/%d", 437 mc[i], i, count)); 438 vm_page_io_start(mc[i]); 439 pmap_remove_write(mc[i]); 440 } 441 vm_page_unlock_queues(); 442 vm_object_pip_add(object, count); 443 444 vm_pager_put_pages(object, mc, count, flags, pageout_status); 445 446 vm_page_lock_queues(); 447 for (i = 0; i < count; i++) { 448 vm_page_t mt = mc[i]; 449 450 KASSERT(pageout_status[i] == VM_PAGER_PEND || 451 (mt->flags & PG_WRITEABLE) == 0, 452 ("vm_pageout_flush: page %p is not write protected", mt)); 453 switch (pageout_status[i]) { 454 case VM_PAGER_OK: 455 case VM_PAGER_PEND: 456 numpagedout++; 457 break; 458 case VM_PAGER_BAD: 459 /* 460 * Page outside of range of object. Right now we 461 * essentially lose the changes by pretending it 462 * worked. 463 */ 464 pmap_clear_modify(mt); 465 vm_page_undirty(mt); 466 break; 467 case VM_PAGER_ERROR: 468 case VM_PAGER_FAIL: 469 /* 470 * If page couldn't be paged out, then reactivate the 471 * page so it doesn't clog the inactive list. (We 472 * will try paging out it again later). 473 */ 474 vm_page_activate(mt); 475 break; 476 case VM_PAGER_AGAIN: 477 break; 478 } 479 480 /* 481 * If the operation is still going, leave the page busy to 482 * block all other accesses. Also, leave the paging in 483 * progress indicator set so that we don't attempt an object 484 * collapse. 485 */ 486 if (pageout_status[i] != VM_PAGER_PEND) { 487 vm_object_pip_wakeup(object); 488 vm_page_io_finish(mt); 489 if (vm_page_count_severe()) 490 vm_page_try_to_cache(mt); 491 } 492 } 493 return numpagedout; 494} 495 496#if !defined(NO_SWAPPING) 497/* 498 * vm_pageout_object_deactivate_pages 499 * 500 * deactivate enough pages to satisfy the inactive target 501 * requirements or if vm_page_proc_limit is set, then 502 * deactivate all of the pages in the object and its 503 * backing_objects. 504 * 505 * The object and map must be locked. 506 */ 507static void 508vm_pageout_object_deactivate_pages(pmap, first_object, desired) 509 pmap_t pmap; 510 vm_object_t first_object; 511 long desired; 512{ 513 vm_object_t backing_object, object; 514 vm_page_t p, next; 515 int actcount, rcount, remove_mode; 516 517 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED); 518 if (first_object->type == OBJT_DEVICE || first_object->type == OBJT_PHYS) 519 return; 520 for (object = first_object;; object = backing_object) { 521 if (pmap_resident_count(pmap) <= desired) 522 goto unlock_return; 523 if (object->paging_in_progress) 524 goto unlock_return; 525 526 remove_mode = 0; 527 if (object->shadow_count > 1) 528 remove_mode = 1; 529 /* 530 * scan the objects entire memory queue 531 */ 532 rcount = object->resident_page_count; 533 p = TAILQ_FIRST(&object->memq); 534 vm_page_lock_queues(); 535 while (p && (rcount-- > 0)) { 536 if (pmap_resident_count(pmap) <= desired) { 537 vm_page_unlock_queues(); 538 goto unlock_return; 539 } 540 next = TAILQ_NEXT(p, listq); 541 cnt.v_pdpages++; 542 if (p->wire_count != 0 || 543 p->hold_count != 0 || 544 p->busy != 0 || 545 (p->oflags & VPO_BUSY) || 546 (p->flags & PG_UNMANAGED) || 547 !pmap_page_exists_quick(pmap, p)) { 548 p = next; 549 continue; 550 } 551 actcount = pmap_ts_referenced(p); 552 if (actcount) { 553 vm_page_flag_set(p, PG_REFERENCED); 554 } else if (p->flags & PG_REFERENCED) { 555 actcount = 1; 556 } 557 if ((p->queue != PQ_ACTIVE) && 558 (p->flags & PG_REFERENCED)) { 559 vm_page_activate(p); 560 p->act_count += actcount; 561 vm_page_flag_clear(p, PG_REFERENCED); 562 } else if (p->queue == PQ_ACTIVE) { 563 if ((p->flags & PG_REFERENCED) == 0) { 564 p->act_count -= min(p->act_count, ACT_DECLINE); 565 if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) { 566 pmap_remove_all(p); 567 vm_page_deactivate(p); 568 } else { 569 vm_pageq_requeue(p); 570 } 571 } else { 572 vm_page_activate(p); 573 vm_page_flag_clear(p, PG_REFERENCED); 574 if (p->act_count < (ACT_MAX - ACT_ADVANCE)) 575 p->act_count += ACT_ADVANCE; 576 vm_pageq_requeue(p); 577 } 578 } else if (p->queue == PQ_INACTIVE) { 579 pmap_remove_all(p); 580 } 581 p = next; 582 } 583 vm_page_unlock_queues(); 584 if ((backing_object = object->backing_object) == NULL) 585 goto unlock_return; 586 VM_OBJECT_LOCK(backing_object); 587 if (object != first_object) 588 VM_OBJECT_UNLOCK(object); 589 } 590unlock_return: 591 if (object != first_object) 592 VM_OBJECT_UNLOCK(object); 593} 594 595/* 596 * deactivate some number of pages in a map, try to do it fairly, but 597 * that is really hard to do. 598 */ 599static void 600vm_pageout_map_deactivate_pages(map, desired) 601 vm_map_t map; 602 long desired; 603{ 604 vm_map_entry_t tmpe; 605 vm_object_t obj, bigobj; 606 int nothingwired; 607 608 if (!vm_map_trylock(map)) 609 return; 610 611 bigobj = NULL; 612 nothingwired = TRUE; 613 614 /* 615 * first, search out the biggest object, and try to free pages from 616 * that. 617 */ 618 tmpe = map->header.next; 619 while (tmpe != &map->header) { 620 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 621 obj = tmpe->object.vm_object; 622 if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) { 623 if (obj->shadow_count <= 1 && 624 (bigobj == NULL || 625 bigobj->resident_page_count < obj->resident_page_count)) { 626 if (bigobj != NULL) 627 VM_OBJECT_UNLOCK(bigobj); 628 bigobj = obj; 629 } else 630 VM_OBJECT_UNLOCK(obj); 631 } 632 } 633 if (tmpe->wired_count > 0) 634 nothingwired = FALSE; 635 tmpe = tmpe->next; 636 } 637 638 if (bigobj != NULL) { 639 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired); 640 VM_OBJECT_UNLOCK(bigobj); 641 } 642 /* 643 * Next, hunt around for other pages to deactivate. We actually 644 * do this search sort of wrong -- .text first is not the best idea. 645 */ 646 tmpe = map->header.next; 647 while (tmpe != &map->header) { 648 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 649 break; 650 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 651 obj = tmpe->object.vm_object; 652 if (obj != NULL) { 653 VM_OBJECT_LOCK(obj); 654 vm_pageout_object_deactivate_pages(map->pmap, obj, desired); 655 VM_OBJECT_UNLOCK(obj); 656 } 657 } 658 tmpe = tmpe->next; 659 } 660 661 /* 662 * Remove all mappings if a process is swapped out, this will free page 663 * table pages. 664 */ 665 if (desired == 0 && nothingwired) { 666 pmap_remove(vm_map_pmap(map), vm_map_min(map), 667 vm_map_max(map)); 668 } 669 vm_map_unlock(map); 670} 671#endif /* !defined(NO_SWAPPING) */ 672 673/* 674 * vm_pageout_scan does the dirty work for the pageout daemon. 675 */ 676static void 677vm_pageout_scan(int pass) 678{ 679 vm_page_t m, next; 680 struct vm_page marker; 681 int page_shortage, maxscan, pcount; 682 int addl_page_shortage, addl_page_shortage_init; 683 struct proc *p, *bigproc; 684 struct thread *td; 685 vm_offset_t size, bigsize; 686 vm_object_t object; 687 int actcount; 688 int vnodes_skipped = 0; 689 int maxlaunder; 690 691 /* 692 * Decrease registered cache sizes. 693 */ 694 EVENTHANDLER_INVOKE(vm_lowmem, 0); 695 /* 696 * We do this explicitly after the caches have been drained above. 697 */ 698 uma_reclaim(); 699 700 addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit); 701 702 /* 703 * Calculate the number of pages we want to either free or move 704 * to the cache. 705 */ 706 page_shortage = vm_paging_target() + addl_page_shortage_init; 707 708 /* 709 * Initialize our marker 710 */ 711 bzero(&marker, sizeof(marker)); 712 marker.flags = PG_FICTITIOUS | PG_MARKER; 713 marker.oflags = VPO_BUSY; 714 marker.queue = PQ_INACTIVE; 715 marker.wire_count = 1; 716 717 /* 718 * Start scanning the inactive queue for pages we can move to the 719 * cache or free. The scan will stop when the target is reached or 720 * we have scanned the entire inactive queue. Note that m->act_count 721 * is not used to form decisions for the inactive queue, only for the 722 * active queue. 723 * 724 * maxlaunder limits the number of dirty pages we flush per scan. 725 * For most systems a smaller value (16 or 32) is more robust under 726 * extreme memory and disk pressure because any unnecessary writes 727 * to disk can result in extreme performance degredation. However, 728 * systems with excessive dirty pages (especially when MAP_NOSYNC is 729 * used) will die horribly with limited laundering. If the pageout 730 * daemon cannot clean enough pages in the first pass, we let it go 731 * all out in succeeding passes. 732 */ 733 if ((maxlaunder = vm_max_launder) <= 1) 734 maxlaunder = 1; 735 if (pass) 736 maxlaunder = 10000; 737 vm_page_lock_queues(); 738rescan0: 739 addl_page_shortage = addl_page_shortage_init; 740 maxscan = cnt.v_inactive_count; 741 742 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl); 743 m != NULL && maxscan-- > 0 && page_shortage > 0; 744 m = next) { 745 746 cnt.v_pdpages++; 747 748 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE) { 749 goto rescan0; 750 } 751 752 next = TAILQ_NEXT(m, pageq); 753 object = m->object; 754 755 /* 756 * skip marker pages 757 */ 758 if (m->flags & PG_MARKER) 759 continue; 760 761 /* 762 * A held page may be undergoing I/O, so skip it. 763 */ 764 if (m->hold_count) { 765 vm_pageq_requeue(m); 766 addl_page_shortage++; 767 continue; 768 } 769 /* 770 * Don't mess with busy pages, keep in the front of the 771 * queue, most likely are being paged out. 772 */ 773 if (!VM_OBJECT_TRYLOCK(object) && 774 (!vm_pageout_fallback_object_lock(m, &next) || 775 m->hold_count != 0)) { 776 VM_OBJECT_UNLOCK(object); 777 addl_page_shortage++; 778 continue; 779 } 780 if (m->busy || (m->oflags & VPO_BUSY)) { 781 VM_OBJECT_UNLOCK(object); 782 addl_page_shortage++; 783 continue; 784 } 785 786 /* 787 * If the object is not being used, we ignore previous 788 * references. 789 */ 790 if (object->ref_count == 0) { 791 vm_page_flag_clear(m, PG_REFERENCED); 792 pmap_clear_reference(m); 793 794 /* 795 * Otherwise, if the page has been referenced while in the 796 * inactive queue, we bump the "activation count" upwards, 797 * making it less likely that the page will be added back to 798 * the inactive queue prematurely again. Here we check the 799 * page tables (or emulated bits, if any), given the upper 800 * level VM system not knowing anything about existing 801 * references. 802 */ 803 } else if (((m->flags & PG_REFERENCED) == 0) && 804 (actcount = pmap_ts_referenced(m))) { 805 vm_page_activate(m); 806 VM_OBJECT_UNLOCK(object); 807 m->act_count += (actcount + ACT_ADVANCE); 808 continue; 809 } 810 811 /* 812 * If the upper level VM system knows about any page 813 * references, we activate the page. We also set the 814 * "activation count" higher than normal so that we will less 815 * likely place pages back onto the inactive queue again. 816 */ 817 if ((m->flags & PG_REFERENCED) != 0) { 818 vm_page_flag_clear(m, PG_REFERENCED); 819 actcount = pmap_ts_referenced(m); 820 vm_page_activate(m); 821 VM_OBJECT_UNLOCK(object); 822 m->act_count += (actcount + ACT_ADVANCE + 1); 823 continue; 824 } 825 826 /* 827 * If the upper level VM system doesn't know anything about 828 * the page being dirty, we have to check for it again. As 829 * far as the VM code knows, any partially dirty pages are 830 * fully dirty. 831 */ 832 if (m->dirty == 0 && !pmap_is_modified(m)) { 833 /* 834 * Avoid a race condition: Unless write access is 835 * removed from the page, another processor could 836 * modify it before all access is removed by the call 837 * to vm_page_cache() below. If vm_page_cache() finds 838 * that the page has been modified when it removes all 839 * access, it panics because it cannot cache dirty 840 * pages. In principle, we could eliminate just write 841 * access here rather than all access. In the expected 842 * case, when there are no last instant modifications 843 * to the page, removing all access will be cheaper 844 * overall. 845 */ 846 if ((m->flags & PG_WRITEABLE) != 0) 847 pmap_remove_all(m); 848 } else { 849 vm_page_dirty(m); 850 } 851 852 if (m->valid == 0) { 853 /* 854 * Invalid pages can be easily freed 855 */ 856 vm_page_free(m); 857 cnt.v_dfree++; 858 --page_shortage; 859 } else if (m->dirty == 0) { 860 /* 861 * Clean pages can be placed onto the cache queue. 862 * This effectively frees them. 863 */ 864 vm_page_cache(m); 865 --page_shortage; 866 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) { 867 /* 868 * Dirty pages need to be paged out, but flushing 869 * a page is extremely expensive verses freeing 870 * a clean page. Rather then artificially limiting 871 * the number of pages we can flush, we instead give 872 * dirty pages extra priority on the inactive queue 873 * by forcing them to be cycled through the queue 874 * twice before being flushed, after which the 875 * (now clean) page will cycle through once more 876 * before being freed. This significantly extends 877 * the thrash point for a heavily loaded machine. 878 */ 879 vm_page_flag_set(m, PG_WINATCFLS); 880 vm_pageq_requeue(m); 881 } else if (maxlaunder > 0) { 882 /* 883 * We always want to try to flush some dirty pages if 884 * we encounter them, to keep the system stable. 885 * Normally this number is small, but under extreme 886 * pressure where there are insufficient clean pages 887 * on the inactive queue, we may have to go all out. 888 */ 889 int swap_pageouts_ok, vfslocked = 0; 890 struct vnode *vp = NULL; 891 struct mount *mp = NULL; 892 893 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 894 swap_pageouts_ok = 1; 895 } else { 896 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 897 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 898 vm_page_count_min()); 899 900 } 901 902 /* 903 * We don't bother paging objects that are "dead". 904 * Those objects are in a "rundown" state. 905 */ 906 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 907 VM_OBJECT_UNLOCK(object); 908 vm_pageq_requeue(m); 909 continue; 910 } 911 912 /* 913 * Following operations may unlock 914 * vm_page_queue_mtx, invalidating the 'next' 915 * pointer. To prevent an inordinate number 916 * of restarts we use our marker to remember 917 * our place. 918 * 919 */ 920 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, 921 m, &marker, pageq); 922 /* 923 * The object is already known NOT to be dead. It 924 * is possible for the vget() to block the whole 925 * pageout daemon, but the new low-memory handling 926 * code should prevent it. 927 * 928 * The previous code skipped locked vnodes and, worse, 929 * reordered pages in the queue. This results in 930 * completely non-deterministic operation and, on a 931 * busy system, can lead to extremely non-optimal 932 * pageouts. For example, it can cause clean pages 933 * to be freed and dirty pages to be moved to the end 934 * of the queue. Since dirty pages are also moved to 935 * the end of the queue once-cleaned, this gives 936 * way too large a weighting to defering the freeing 937 * of dirty pages. 938 * 939 * We can't wait forever for the vnode lock, we might 940 * deadlock due to a vn_read() getting stuck in 941 * vm_wait while holding this vnode. We skip the 942 * vnode if we can't get it in a reasonable amount 943 * of time. 944 */ 945 if (object->type == OBJT_VNODE) { 946 vp = object->handle; 947 if (vp->v_type == VREG && 948 vn_start_write(vp, &mp, V_NOWAIT) != 0) { 949 KASSERT(mp == NULL, 950 ("vm_pageout_scan: mp != NULL")); 951 ++pageout_lock_miss; 952 if (object->flags & OBJ_MIGHTBEDIRTY) 953 vnodes_skipped++; 954 goto unlock_and_continue; 955 } 956 vm_page_unlock_queues(); 957 vm_object_reference_locked(object); 958 VM_OBJECT_UNLOCK(object); 959 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 960 if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK, 961 curthread)) { 962 VM_OBJECT_LOCK(object); 963 vm_page_lock_queues(); 964 ++pageout_lock_miss; 965 if (object->flags & OBJ_MIGHTBEDIRTY) 966 vnodes_skipped++; 967 vp = NULL; 968 goto unlock_and_continue; 969 } 970 VM_OBJECT_LOCK(object); 971 vm_page_lock_queues(); 972 /* 973 * The page might have been moved to another 974 * queue during potential blocking in vget() 975 * above. The page might have been freed and 976 * reused for another vnode. 977 */ 978 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE || 979 m->object != object || 980 TAILQ_NEXT(m, pageq) != &marker) { 981 if (object->flags & OBJ_MIGHTBEDIRTY) 982 vnodes_skipped++; 983 goto unlock_and_continue; 984 } 985 986 /* 987 * The page may have been busied during the 988 * blocking in vget(). We don't move the 989 * page back onto the end of the queue so that 990 * statistics are more correct if we don't. 991 */ 992 if (m->busy || (m->oflags & VPO_BUSY)) { 993 goto unlock_and_continue; 994 } 995 996 /* 997 * If the page has become held it might 998 * be undergoing I/O, so skip it 999 */ 1000 if (m->hold_count) { 1001 vm_pageq_requeue(m); 1002 if (object->flags & OBJ_MIGHTBEDIRTY) 1003 vnodes_skipped++; 1004 goto unlock_and_continue; 1005 } 1006 } 1007 1008 /* 1009 * If a page is dirty, then it is either being washed 1010 * (but not yet cleaned) or it is still in the 1011 * laundry. If it is still in the laundry, then we 1012 * start the cleaning operation. 1013 * 1014 * decrement page_shortage on success to account for 1015 * the (future) cleaned page. Otherwise we could wind 1016 * up laundering or cleaning too many pages. 1017 */ 1018 if (vm_pageout_clean(m) != 0) { 1019 --page_shortage; 1020 --maxlaunder; 1021 } 1022unlock_and_continue: 1023 VM_OBJECT_UNLOCK(object); 1024 if (mp != NULL) { 1025 vm_page_unlock_queues(); 1026 if (vp != NULL) 1027 vput(vp); 1028 VFS_UNLOCK_GIANT(vfslocked); 1029 vm_object_deallocate(object); 1030 vn_finished_write(mp); 1031 vm_page_lock_queues(); 1032 } 1033 next = TAILQ_NEXT(&marker, pageq); 1034 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, 1035 &marker, pageq); 1036 continue; 1037 } 1038 VM_OBJECT_UNLOCK(object); 1039 } 1040 1041 /* 1042 * Compute the number of pages we want to try to move from the 1043 * active queue to the inactive queue. 1044 */ 1045 page_shortage = vm_paging_target() + 1046 cnt.v_inactive_target - cnt.v_inactive_count; 1047 page_shortage += addl_page_shortage; 1048 1049 /* 1050 * Scan the active queue for things we can deactivate. We nominally 1051 * track the per-page activity counter and use it to locate 1052 * deactivation candidates. 1053 */ 1054 pcount = cnt.v_active_count; 1055 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1056 1057 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) { 1058 1059 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE), 1060 ("vm_pageout_scan: page %p isn't active", m)); 1061 1062 next = TAILQ_NEXT(m, pageq); 1063 object = m->object; 1064 if ((m->flags & PG_MARKER) != 0) { 1065 m = next; 1066 continue; 1067 } 1068 if (!VM_OBJECT_TRYLOCK(object) && 1069 !vm_pageout_fallback_object_lock(m, &next)) { 1070 VM_OBJECT_UNLOCK(object); 1071 m = next; 1072 continue; 1073 } 1074 1075 /* 1076 * Don't deactivate pages that are busy. 1077 */ 1078 if ((m->busy != 0) || 1079 (m->oflags & VPO_BUSY) || 1080 (m->hold_count != 0)) { 1081 VM_OBJECT_UNLOCK(object); 1082 vm_pageq_requeue(m); 1083 m = next; 1084 continue; 1085 } 1086 1087 /* 1088 * The count for pagedaemon pages is done after checking the 1089 * page for eligibility... 1090 */ 1091 cnt.v_pdpages++; 1092 1093 /* 1094 * Check to see "how much" the page has been used. 1095 */ 1096 actcount = 0; 1097 if (object->ref_count != 0) { 1098 if (m->flags & PG_REFERENCED) { 1099 actcount += 1; 1100 } 1101 actcount += pmap_ts_referenced(m); 1102 if (actcount) { 1103 m->act_count += ACT_ADVANCE + actcount; 1104 if (m->act_count > ACT_MAX) 1105 m->act_count = ACT_MAX; 1106 } 1107 } 1108 1109 /* 1110 * Since we have "tested" this bit, we need to clear it now. 1111 */ 1112 vm_page_flag_clear(m, PG_REFERENCED); 1113 1114 /* 1115 * Only if an object is currently being used, do we use the 1116 * page activation count stats. 1117 */ 1118 if (actcount && (object->ref_count != 0)) { 1119 vm_pageq_requeue(m); 1120 } else { 1121 m->act_count -= min(m->act_count, ACT_DECLINE); 1122 if (vm_pageout_algorithm || 1123 object->ref_count == 0 || 1124 m->act_count == 0) { 1125 page_shortage--; 1126 if (object->ref_count == 0) { 1127 pmap_remove_all(m); 1128 if (m->dirty == 0) 1129 vm_page_cache(m); 1130 else 1131 vm_page_deactivate(m); 1132 } else { 1133 vm_page_deactivate(m); 1134 } 1135 } else { 1136 vm_pageq_requeue(m); 1137 } 1138 } 1139 VM_OBJECT_UNLOCK(object); 1140 m = next; 1141 } 1142 1143 /* 1144 * We try to maintain some *really* free pages, this allows interrupt 1145 * code to be guaranteed space. Since both cache and free queues 1146 * are considered basically 'free', moving pages from cache to free 1147 * does not effect other calculations. 1148 */ 1149 while (cnt.v_free_count < cnt.v_free_reserved) { 1150 TAILQ_FOREACH(m, &vm_page_queues[PQ_CACHE].pl, pageq) { 1151 KASSERT(m->dirty == 0, 1152 ("Found dirty cache page %p", m)); 1153 KASSERT(!pmap_page_is_mapped(m), 1154 ("Found mapped cache page %p", m)); 1155 KASSERT((m->flags & PG_UNMANAGED) == 0, 1156 ("Found unmanaged cache page %p", m)); 1157 KASSERT(m->wire_count == 0, 1158 ("Found wired cache page %p", m)); 1159 if (m->hold_count == 0 && VM_OBJECT_TRYLOCK(object = 1160 m->object)) { 1161 KASSERT((m->oflags & VPO_BUSY) == 0 && 1162 m->busy == 0, ("Found busy cache page %p", 1163 m)); 1164 vm_page_free(m); 1165 VM_OBJECT_UNLOCK(object); 1166 cnt.v_dfree++; 1167 break; 1168 } 1169 } 1170 if (m == NULL) 1171 break; 1172 } 1173 vm_page_unlock_queues(); 1174#if !defined(NO_SWAPPING) 1175 /* 1176 * Idle process swapout -- run once per second. 1177 */ 1178 if (vm_swap_idle_enabled) { 1179 static long lsec; 1180 if (time_second != lsec) { 1181 vm_req_vmdaemon(VM_SWAP_IDLE); 1182 lsec = time_second; 1183 } 1184 } 1185#endif 1186 1187 /* 1188 * If we didn't get enough free pages, and we have skipped a vnode 1189 * in a writeable object, wakeup the sync daemon. And kick swapout 1190 * if we did not get enough free pages. 1191 */ 1192 if (vm_paging_target() > 0) { 1193 if (vnodes_skipped && vm_page_count_min()) 1194 (void) speedup_syncer(); 1195#if !defined(NO_SWAPPING) 1196 if (vm_swap_enabled && vm_page_count_target()) 1197 vm_req_vmdaemon(VM_SWAP_NORMAL); 1198#endif 1199 } 1200 1201 /* 1202 * If we are critically low on one of RAM or swap and low on 1203 * the other, kill the largest process. However, we avoid 1204 * doing this on the first pass in order to give ourselves a 1205 * chance to flush out dirty vnode-backed pages and to allow 1206 * active pages to be moved to the inactive queue and reclaimed. 1207 * 1208 * We keep the process bigproc locked once we find it to keep anyone 1209 * from messing with it; however, there is a possibility of 1210 * deadlock if process B is bigproc and one of it's child processes 1211 * attempts to propagate a signal to B while we are waiting for A's 1212 * lock while walking this list. To avoid this, we don't block on 1213 * the process lock but just skip a process if it is already locked. 1214 */ 1215 if (pass != 0 && 1216 ((swap_pager_avail < 64 && vm_page_count_min()) || 1217 (swap_pager_full && vm_paging_target() > 0))) { 1218 bigproc = NULL; 1219 bigsize = 0; 1220 sx_slock(&allproc_lock); 1221 FOREACH_PROC_IN_SYSTEM(p) { 1222 int breakout; 1223 1224 if (PROC_TRYLOCK(p) == 0) 1225 continue; 1226 /* 1227 * If this is a system or protected process, skip it. 1228 */ 1229 if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) || 1230 (p->p_flag & P_PROTECTED) || 1231 ((p->p_pid < 48) && (swap_pager_avail != 0))) { 1232 PROC_UNLOCK(p); 1233 continue; 1234 } 1235 /* 1236 * If the process is in a non-running type state, 1237 * don't touch it. Check all the threads individually. 1238 */ 1239 PROC_SLOCK(p); 1240 breakout = 0; 1241 FOREACH_THREAD_IN_PROC(p, td) { 1242 thread_lock(td); 1243 if (!TD_ON_RUNQ(td) && 1244 !TD_IS_RUNNING(td) && 1245 !TD_IS_SLEEPING(td)) { 1246 thread_unlock(td); 1247 breakout = 1; 1248 break; 1249 } 1250 thread_unlock(td); 1251 } 1252 PROC_SUNLOCK(p); 1253 if (breakout) { 1254 PROC_UNLOCK(p); 1255 continue; 1256 } 1257 /* 1258 * get the process size 1259 */ 1260 if (!vm_map_trylock_read(&p->p_vmspace->vm_map)) { 1261 PROC_UNLOCK(p); 1262 continue; 1263 } 1264 size = vmspace_swap_count(p->p_vmspace); 1265 vm_map_unlock_read(&p->p_vmspace->vm_map); 1266 size += vmspace_resident_count(p->p_vmspace); 1267 /* 1268 * if the this process is bigger than the biggest one 1269 * remember it. 1270 */ 1271 if (size > bigsize) { 1272 if (bigproc != NULL) 1273 PROC_UNLOCK(bigproc); 1274 bigproc = p; 1275 bigsize = size; 1276 } else 1277 PROC_UNLOCK(p); 1278 } 1279 sx_sunlock(&allproc_lock); 1280 if (bigproc != NULL) { 1281 killproc(bigproc, "out of swap space"); 1282 PROC_SLOCK(bigproc); 1283 sched_nice(bigproc, PRIO_MIN); 1284 PROC_SUNLOCK(bigproc); 1285 PROC_UNLOCK(bigproc); 1286 wakeup(&cnt.v_free_count); 1287 } 1288 } 1289} 1290 1291/* 1292 * This routine tries to maintain the pseudo LRU active queue, 1293 * so that during long periods of time where there is no paging, 1294 * that some statistic accumulation still occurs. This code 1295 * helps the situation where paging just starts to occur. 1296 */ 1297static void 1298vm_pageout_page_stats() 1299{ 1300 vm_object_t object; 1301 vm_page_t m,next; 1302 int pcount,tpcount; /* Number of pages to check */ 1303 static int fullintervalcount = 0; 1304 int page_shortage; 1305 1306 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1307 page_shortage = 1308 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) - 1309 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count); 1310 1311 if (page_shortage <= 0) 1312 return; 1313 1314 pcount = cnt.v_active_count; 1315 fullintervalcount += vm_pageout_stats_interval; 1316 if (fullintervalcount < vm_pageout_full_stats_interval) { 1317 tpcount = (vm_pageout_stats_max * cnt.v_active_count) / cnt.v_page_count; 1318 if (pcount > tpcount) 1319 pcount = tpcount; 1320 } else { 1321 fullintervalcount = 0; 1322 } 1323 1324 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1325 while ((m != NULL) && (pcount-- > 0)) { 1326 int actcount; 1327 1328 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE), 1329 ("vm_pageout_page_stats: page %p isn't active", m)); 1330 1331 next = TAILQ_NEXT(m, pageq); 1332 object = m->object; 1333 1334 if ((m->flags & PG_MARKER) != 0) { 1335 m = next; 1336 continue; 1337 } 1338 if (!VM_OBJECT_TRYLOCK(object) && 1339 !vm_pageout_fallback_object_lock(m, &next)) { 1340 VM_OBJECT_UNLOCK(object); 1341 m = next; 1342 continue; 1343 } 1344 1345 /* 1346 * Don't deactivate pages that are busy. 1347 */ 1348 if ((m->busy != 0) || 1349 (m->oflags & VPO_BUSY) || 1350 (m->hold_count != 0)) { 1351 VM_OBJECT_UNLOCK(object); 1352 vm_pageq_requeue(m); 1353 m = next; 1354 continue; 1355 } 1356 1357 actcount = 0; 1358 if (m->flags & PG_REFERENCED) { 1359 vm_page_flag_clear(m, PG_REFERENCED); 1360 actcount += 1; 1361 } 1362 1363 actcount += pmap_ts_referenced(m); 1364 if (actcount) { 1365 m->act_count += ACT_ADVANCE + actcount; 1366 if (m->act_count > ACT_MAX) 1367 m->act_count = ACT_MAX; 1368 vm_pageq_requeue(m); 1369 } else { 1370 if (m->act_count == 0) { 1371 /* 1372 * We turn off page access, so that we have 1373 * more accurate RSS stats. We don't do this 1374 * in the normal page deactivation when the 1375 * system is loaded VM wise, because the 1376 * cost of the large number of page protect 1377 * operations would be higher than the value 1378 * of doing the operation. 1379 */ 1380 pmap_remove_all(m); 1381 vm_page_deactivate(m); 1382 } else { 1383 m->act_count -= min(m->act_count, ACT_DECLINE); 1384 vm_pageq_requeue(m); 1385 } 1386 } 1387 VM_OBJECT_UNLOCK(object); 1388 m = next; 1389 } 1390} 1391 1392/* 1393 * vm_pageout is the high level pageout daemon. 1394 */ 1395static void 1396vm_pageout() 1397{ 1398 int error, pass; 1399 1400 /* 1401 * Initialize some paging parameters. 1402 */ 1403 cnt.v_interrupt_free_min = 2; 1404 if (cnt.v_page_count < 2000) 1405 vm_pageout_page_count = 8; 1406 1407 /* 1408 * v_free_reserved needs to include enough for the largest 1409 * swap pager structures plus enough for any pv_entry structs 1410 * when paging. 1411 */ 1412 if (cnt.v_page_count > 1024) 1413 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200; 1414 else 1415 cnt.v_free_min = 4; 1416 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1417 cnt.v_interrupt_free_min; 1418 cnt.v_free_reserved = vm_pageout_page_count + 1419 cnt.v_pageout_free_min + (cnt.v_page_count / 768); 1420 cnt.v_free_severe = cnt.v_free_min / 2; 1421 cnt.v_free_min += cnt.v_free_reserved; 1422 cnt.v_free_severe += cnt.v_free_reserved; 1423 1424 /* 1425 * v_free_target and v_cache_min control pageout hysteresis. Note 1426 * that these are more a measure of the VM cache queue hysteresis 1427 * then the VM free queue. Specifically, v_free_target is the 1428 * high water mark (free+cache pages). 1429 * 1430 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the 1431 * low water mark, while v_free_min is the stop. v_cache_min must 1432 * be big enough to handle memory needs while the pageout daemon 1433 * is signalled and run to free more pages. 1434 */ 1435 if (cnt.v_free_count > 6144) 1436 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved; 1437 else 1438 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved; 1439 1440 if (cnt.v_free_count > 2048) { 1441 cnt.v_cache_min = cnt.v_free_target; 1442 cnt.v_cache_max = 2 * cnt.v_cache_min; 1443 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2; 1444 } else { 1445 cnt.v_cache_min = 0; 1446 cnt.v_cache_max = 0; 1447 cnt.v_inactive_target = cnt.v_free_count / 4; 1448 } 1449 if (cnt.v_inactive_target > cnt.v_free_count / 3) 1450 cnt.v_inactive_target = cnt.v_free_count / 3; 1451 1452 /* XXX does not really belong here */ 1453 if (vm_page_max_wired == 0) 1454 vm_page_max_wired = cnt.v_free_count / 3; 1455 1456 if (vm_pageout_stats_max == 0) 1457 vm_pageout_stats_max = cnt.v_free_target; 1458 1459 /* 1460 * Set interval in seconds for stats scan. 1461 */ 1462 if (vm_pageout_stats_interval == 0) 1463 vm_pageout_stats_interval = 5; 1464 if (vm_pageout_full_stats_interval == 0) 1465 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4; 1466 1467 swap_pager_swap_init(); 1468 pass = 0; 1469 /* 1470 * The pageout daemon is never done, so loop forever. 1471 */ 1472 while (TRUE) { 1473 /* 1474 * If we have enough free memory, wakeup waiters. Do 1475 * not clear vm_pages_needed until we reach our target, 1476 * otherwise we may be woken up over and over again and 1477 * waste a lot of cpu. 1478 */ 1479 mtx_lock(&vm_page_queue_free_mtx); 1480 if (vm_pages_needed && !vm_page_count_min()) { 1481 if (!vm_paging_needed()) 1482 vm_pages_needed = 0; 1483 wakeup(&cnt.v_free_count); 1484 } 1485 if (vm_pages_needed) { 1486 /* 1487 * Still not done, take a second pass without waiting 1488 * (unlimited dirty cleaning), otherwise sleep a bit 1489 * and try again. 1490 */ 1491 ++pass; 1492 if (pass > 1) 1493 msleep(&vm_pages_needed, 1494 &vm_page_queue_free_mtx, PVM, "psleep", 1495 hz / 2); 1496 } else { 1497 /* 1498 * Good enough, sleep & handle stats. Prime the pass 1499 * for the next run. 1500 */ 1501 if (pass > 1) 1502 pass = 1; 1503 else 1504 pass = 0; 1505 error = msleep(&vm_pages_needed, 1506 &vm_page_queue_free_mtx, PVM, "psleep", 1507 vm_pageout_stats_interval * hz); 1508 if (error && !vm_pages_needed) { 1509 mtx_unlock(&vm_page_queue_free_mtx); 1510 pass = 0; 1511 vm_page_lock_queues(); 1512 vm_pageout_page_stats(); 1513 vm_page_unlock_queues(); 1514 continue; 1515 } 1516 } 1517 if (vm_pages_needed) 1518 cnt.v_pdwakeups++; 1519 mtx_unlock(&vm_page_queue_free_mtx); 1520 vm_pageout_scan(pass); 1521 } 1522} 1523 1524/* 1525 * Unless the free page queue lock is held by the caller, this function 1526 * should be regarded as advisory. Specifically, the caller should 1527 * not msleep() on &cnt.v_free_count following this function unless 1528 * the free page queue lock is held until the msleep() is performed. 1529 */ 1530void 1531pagedaemon_wakeup() 1532{ 1533 1534 if (!vm_pages_needed && curthread->td_proc != pageproc) { 1535 vm_pages_needed = 1; 1536 wakeup(&vm_pages_needed); 1537 } 1538} 1539 1540#if !defined(NO_SWAPPING) 1541static void 1542vm_req_vmdaemon(int req) 1543{ 1544 static int lastrun = 0; 1545 1546 mtx_lock(&vm_daemon_mtx); 1547 vm_pageout_req_swapout |= req; 1548 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1549 wakeup(&vm_daemon_needed); 1550 lastrun = ticks; 1551 } 1552 mtx_unlock(&vm_daemon_mtx); 1553} 1554 1555static void 1556vm_daemon() 1557{ 1558 struct rlimit rsslim; 1559 struct proc *p; 1560 struct thread *td; 1561 int breakout, swapout_flags; 1562 1563 while (TRUE) { 1564 mtx_lock(&vm_daemon_mtx); 1565 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0); 1566 swapout_flags = vm_pageout_req_swapout; 1567 vm_pageout_req_swapout = 0; 1568 mtx_unlock(&vm_daemon_mtx); 1569 if (swapout_flags) 1570 swapout_procs(swapout_flags); 1571 1572 /* 1573 * scan the processes for exceeding their rlimits or if 1574 * process is swapped out -- deactivate pages 1575 */ 1576 sx_slock(&allproc_lock); 1577 FOREACH_PROC_IN_SYSTEM(p) { 1578 vm_pindex_t limit, size; 1579 1580 /* 1581 * if this is a system process or if we have already 1582 * looked at this process, skip it. 1583 */ 1584 PROC_LOCK(p); 1585 if (p->p_flag & (P_SYSTEM | P_WEXIT)) { 1586 PROC_UNLOCK(p); 1587 continue; 1588 } 1589 /* 1590 * if the process is in a non-running type state, 1591 * don't touch it. 1592 */ 1593 PROC_SLOCK(p); 1594 breakout = 0; 1595 FOREACH_THREAD_IN_PROC(p, td) { 1596 thread_lock(td); 1597 if (!TD_ON_RUNQ(td) && 1598 !TD_IS_RUNNING(td) && 1599 !TD_IS_SLEEPING(td)) { 1600 thread_unlock(td); 1601 breakout = 1; 1602 break; 1603 } 1604 thread_unlock(td); 1605 } 1606 PROC_SUNLOCK(p); 1607 if (breakout) { 1608 PROC_UNLOCK(p); 1609 continue; 1610 } 1611 /* 1612 * get a limit 1613 */ 1614 lim_rlimit(p, RLIMIT_RSS, &rsslim); 1615 limit = OFF_TO_IDX( 1616 qmin(rsslim.rlim_cur, rsslim.rlim_max)); 1617 1618 /* 1619 * let processes that are swapped out really be 1620 * swapped out set the limit to nothing (will force a 1621 * swap-out.) 1622 */ 1623 if ((p->p_flag & P_INMEM) == 0) 1624 limit = 0; /* XXX */ 1625 PROC_UNLOCK(p); 1626 1627 size = vmspace_resident_count(p->p_vmspace); 1628 if (limit >= 0 && size >= limit) { 1629 vm_pageout_map_deactivate_pages( 1630 &p->p_vmspace->vm_map, limit); 1631 } 1632 } 1633 sx_sunlock(&allproc_lock); 1634 } 1635} 1636#endif /* !defined(NO_SWAPPING) */ 1637