vm_pageout.c revision 207694
137985Sphk/*- 237985Sphk * Copyright (c) 1991 Regents of the University of California. 337985Sphk * All rights reserved. 437985Sphk * Copyright (c) 1994 John S. Dyson 537985Sphk * All rights reserved. 637985Sphk * Copyright (c) 1994 David Greenman 737985Sphk * All rights reserved. 837985Sphk * Copyright (c) 2005 Yahoo! Technologies Norway AS 937985Sphk * All rights reserved. 1037985Sphk * 1137985Sphk * This code is derived from software contributed to Berkeley by 1237985Sphk * The Mach Operating System project at Carnegie-Mellon University. 1337985Sphk * 1437985Sphk * Redistribution and use in source and binary forms, with or without 1537985Sphk * modification, are permitted provided that the following conditions 1637985Sphk * are met: 1737985Sphk * 1. Redistributions of source code must retain the above copyright 1837985Sphk * notice, this list of conditions and the following disclaimer. 1937985Sphk * 2. Redistributions in binary form must reproduce the above copyright 2037985Sphk * notice, this list of conditions and the following disclaimer in the 2137985Sphk * documentation and/or other materials provided with the distribution. 2237985Sphk * 3. All advertising materials mentioning features or use of this software 2337985Sphk * must display the following acknowledgement: 2437985Sphk * This product includes software developed by the University of 2537985Sphk * California, Berkeley and its contributors. 2637985Sphk * 4. Neither the name of the University nor the names of its contributors 2737985Sphk * may be used to endorse or promote products derived from this software 2850476Speter * without specific prior written permission. 2948794Snik * 30350835Skib * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 3137985Sphk * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 3279531Sru * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 3337985Sphk * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 3437985Sphk * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 3537985Sphk * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 3659501Sphantom * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 37124535Sru * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 3837985Sphk * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 3984306Sru * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 4037985Sphk * SUCH DAMAGE. 4137985Sphk * 4237985Sphk * from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91 4337985Sphk * 4437985Sphk * 4537985Sphk * Copyright (c) 1987, 1990 Carnegie-Mellon University. 4637985Sphk * All rights reserved. 4737985Sphk * 4837985Sphk * Authors: Avadis Tevanian, Jr., Michael Wayne Young 4937985Sphk * 5037985Sphk * Permission to use, copy, modify and distribute this software and 5137985Sphk * its documentation is hereby granted, provided that both the copyright 5237985Sphk * notice and this permission notice appear in all copies of the 5337985Sphk * software, derivative works or modified versions, and any portions 5437985Sphk * thereof, and that both notices appear in supporting documentation. 55112542Scharnier * 5637985Sphk * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 57112542Scharnier * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 58300043Skib * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 5937985Sphk * 6037985Sphk * Carnegie Mellon requests users of this software to return to 6137985Sphk * 6237985Sphk * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 6337985Sphk * School of Computer Science 6437985Sphk * Carnegie Mellon University 6537985Sphk * Pittsburgh PA 15213-3890 66300043Skib * 67300043Skib * any improvements or extensions that they make and grant Carnegie the 68300043Skib * rights to redistribute these changes. 69350835Skib */ 70350835Skib 71300043Skib/* 72300043Skib * The proverbial page-out daemon. 73300043Skib */ 74300043Skib 75300043Skib#include <sys/cdefs.h> 76300043Skib__FBSDID("$FreeBSD: head/sys/vm/vm_pageout.c 207694 2010-05-06 04:57:33Z kib $"); 7737985Sphk 7837985Sphk#include "opt_vm.h" 79300043Skib#include <sys/param.h> 8038702Swosch#include <sys/systm.h> 8138702Swosch#include <sys/kernel.h> 8237985Sphk#include <sys/eventhandler.h> 8338702Swosch#include <sys/lock.h> 8437985Sphk#include <sys/mutex.h> 85112542Scharnier#include <sys/proc.h> 8637985Sphk#include <sys/kthread.h> 87112542Scharnier#include <sys/ktr.h> 8873093Sru#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 /* 769 * Initialize our marker 770 */ 771 bzero(&marker, sizeof(marker)); 772 marker.flags = PG_FICTITIOUS | PG_MARKER; 773 marker.oflags = VPO_BUSY; 774 marker.queue = PQ_INACTIVE; 775 marker.wire_count = 1; 776 777 /* 778 * Start scanning the inactive queue for pages we can move to the 779 * cache or free. The scan will stop when the target is reached or 780 * we have scanned the entire inactive queue. Note that m->act_count 781 * is not used to form decisions for the inactive queue, only for the 782 * active queue. 783 * 784 * maxlaunder limits the number of dirty pages we flush per scan. 785 * For most systems a smaller value (16 or 32) is more robust under 786 * extreme memory and disk pressure because any unnecessary writes 787 * to disk can result in extreme performance degredation. However, 788 * systems with excessive dirty pages (especially when MAP_NOSYNC is 789 * used) will die horribly with limited laundering. If the pageout 790 * daemon cannot clean enough pages in the first pass, we let it go 791 * all out in succeeding passes. 792 */ 793 if ((maxlaunder = vm_max_launder) <= 1) 794 maxlaunder = 1; 795 if (pass) 796 maxlaunder = 10000; 797 vm_page_lock_queues(); 798rescan0: 799 addl_page_shortage = addl_page_shortage_init; 800 maxscan = cnt.v_inactive_count; 801 802 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl); 803 m != NULL && maxscan-- > 0 && page_shortage > 0; 804 m = next) { 805 806 cnt.v_pdpages++; 807 808 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE) { 809 goto rescan0; 810 } 811 812 next = TAILQ_NEXT(m, pageq); 813 814 /* 815 * skip marker pages 816 */ 817 if (m->flags & PG_MARKER) 818 continue; 819 820 /* 821 * Lock the page. 822 */ 823 if (!vm_pageout_page_lock(m, &next)) { 824 vm_page_unlock(m); 825 addl_page_shortage++; 826 continue; 827 } 828 829 /* 830 * A held page may be undergoing I/O, so skip it. 831 */ 832 if (m->hold_count || (object = m->object) == NULL) { 833 vm_page_unlock(m); 834 vm_page_requeue(m); 835 addl_page_shortage++; 836 continue; 837 } 838 839 /* 840 * Don't mess with busy pages, keep in the front of the 841 * queue, most likely are being paged out. 842 */ 843 if (!VM_OBJECT_TRYLOCK(object) && 844 (!vm_pageout_fallback_object_lock(m, &next) || 845 m->hold_count != 0)) { 846 VM_OBJECT_UNLOCK(object); 847 vm_page_unlock(m); 848 addl_page_shortage++; 849 continue; 850 } 851 if (m->busy || (m->oflags & VPO_BUSY)) { 852 vm_page_unlock(m); 853 VM_OBJECT_UNLOCK(object); 854 addl_page_shortage++; 855 continue; 856 } 857 858 /* 859 * If the object is not being used, we ignore previous 860 * references. 861 */ 862 if (object->ref_count == 0) { 863 vm_page_flag_clear(m, PG_REFERENCED); 864 KASSERT(!pmap_page_is_mapped(m), 865 ("vm_pageout_scan: page %p is mapped", m)); 866 867 /* 868 * Otherwise, if the page has been referenced while in the 869 * inactive queue, we bump the "activation count" upwards, 870 * making it less likely that the page will be added back to 871 * the inactive queue prematurely again. Here we check the 872 * page tables (or emulated bits, if any), given the upper 873 * level VM system not knowing anything about existing 874 * references. 875 */ 876 } else if (((m->flags & PG_REFERENCED) == 0) && 877 (actcount = pmap_ts_referenced(m))) { 878 vm_page_activate(m); 879 VM_OBJECT_UNLOCK(object); 880 m->act_count += (actcount + ACT_ADVANCE); 881 vm_page_unlock(m); 882 continue; 883 } 884 885 /* 886 * If the upper level VM system knows about any page 887 * references, we activate the page. We also set the 888 * "activation count" higher than normal so that we will less 889 * likely place pages back onto the inactive queue again. 890 */ 891 if ((m->flags & PG_REFERENCED) != 0) { 892 vm_page_flag_clear(m, PG_REFERENCED); 893 actcount = pmap_ts_referenced(m); 894 vm_page_activate(m); 895 VM_OBJECT_UNLOCK(object); 896 m->act_count += (actcount + ACT_ADVANCE + 1); 897 vm_page_unlock(m); 898 continue; 899 } 900 901 /* 902 * If the upper level VM system does not believe that the page 903 * is fully dirty, but it is mapped for write access, then we 904 * consult the pmap to see if the page's dirty status should 905 * be updated. 906 */ 907 if (m->dirty != VM_PAGE_BITS_ALL && 908 (m->flags & PG_WRITEABLE) != 0) { 909 /* 910 * Avoid a race condition: Unless write access is 911 * removed from the page, another processor could 912 * modify it before all access is removed by the call 913 * to vm_page_cache() below. If vm_page_cache() finds 914 * that the page has been modified when it removes all 915 * access, it panics because it cannot cache dirty 916 * pages. In principle, we could eliminate just write 917 * access here rather than all access. In the expected 918 * case, when there are no last instant modifications 919 * to the page, removing all access will be cheaper 920 * overall. 921 */ 922 if (pmap_is_modified(m)) 923 vm_page_dirty(m); 924 else if (m->dirty == 0) 925 pmap_remove_all(m); 926 } 927 928 if (m->valid == 0) { 929 /* 930 * Invalid pages can be easily freed 931 */ 932 vm_page_free(m); 933 cnt.v_dfree++; 934 --page_shortage; 935 } else if (m->dirty == 0) { 936 /* 937 * Clean pages can be placed onto the cache queue. 938 * This effectively frees them. 939 */ 940 vm_page_cache(m); 941 --page_shortage; 942 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) { 943 /* 944 * Dirty pages need to be paged out, but flushing 945 * a page is extremely expensive verses freeing 946 * a clean page. Rather then artificially limiting 947 * the number of pages we can flush, we instead give 948 * dirty pages extra priority on the inactive queue 949 * by forcing them to be cycled through the queue 950 * twice before being flushed, after which the 951 * (now clean) page will cycle through once more 952 * before being freed. This significantly extends 953 * the thrash point for a heavily loaded machine. 954 */ 955 vm_page_flag_set(m, PG_WINATCFLS); 956 vm_page_requeue(m); 957 } else if (maxlaunder > 0) { 958 /* 959 * We always want to try to flush some dirty pages if 960 * we encounter them, to keep the system stable. 961 * Normally this number is small, but under extreme 962 * pressure where there are insufficient clean pages 963 * on the inactive queue, we may have to go all out. 964 */ 965 int swap_pageouts_ok, vfslocked = 0; 966 struct vnode *vp = NULL; 967 struct mount *mp = NULL; 968 969 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 970 swap_pageouts_ok = 1; 971 } else { 972 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 973 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 974 vm_page_count_min()); 975 976 } 977 978 /* 979 * We don't bother paging objects that are "dead". 980 * Those objects are in a "rundown" state. 981 */ 982 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 983 vm_page_unlock(m); 984 VM_OBJECT_UNLOCK(object); 985 vm_page_requeue(m); 986 continue; 987 } 988 989 /* 990 * Following operations may unlock 991 * vm_page_queue_mtx, invalidating the 'next' 992 * pointer. To prevent an inordinate number 993 * of restarts we use our marker to remember 994 * our place. 995 * 996 */ 997 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, 998 m, &marker, pageq); 999 /* 1000 * The object is already known NOT to be dead. It 1001 * is possible for the vget() to block the whole 1002 * pageout daemon, but the new low-memory handling 1003 * code should prevent it. 1004 * 1005 * The previous code skipped locked vnodes and, worse, 1006 * reordered pages in the queue. This results in 1007 * completely non-deterministic operation and, on a 1008 * busy system, can lead to extremely non-optimal 1009 * pageouts. For example, it can cause clean pages 1010 * to be freed and dirty pages to be moved to the end 1011 * of the queue. Since dirty pages are also moved to 1012 * the end of the queue once-cleaned, this gives 1013 * way too large a weighting to defering the freeing 1014 * of dirty pages. 1015 * 1016 * We can't wait forever for the vnode lock, we might 1017 * deadlock due to a vn_read() getting stuck in 1018 * vm_wait while holding this vnode. We skip the 1019 * vnode if we can't get it in a reasonable amount 1020 * of time. 1021 */ 1022 if (object->type == OBJT_VNODE) { 1023 vm_page_unlock_queues(); 1024 vm_page_unlock(m); 1025 vp = object->handle; 1026 if (vp->v_type == VREG && 1027 vn_start_write(vp, &mp, V_NOWAIT) != 0) { 1028 mp = NULL; 1029 ++pageout_lock_miss; 1030 if (object->flags & OBJ_MIGHTBEDIRTY) 1031 vnodes_skipped++; 1032 vm_page_lock_queues(); 1033 goto unlock_and_continue; 1034 } 1035 KASSERT(mp != NULL, 1036 ("vp %p with NULL v_mount", vp)); 1037 vm_object_reference_locked(object); 1038 VM_OBJECT_UNLOCK(object); 1039 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 1040 if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK, 1041 curthread)) { 1042 VM_OBJECT_LOCK(object); 1043 vm_page_lock_queues(); 1044 ++pageout_lock_miss; 1045 if (object->flags & OBJ_MIGHTBEDIRTY) 1046 vnodes_skipped++; 1047 vp = NULL; 1048 goto unlock_and_continue; 1049 } 1050 VM_OBJECT_LOCK(object); 1051 vm_page_lock(m); 1052 vm_page_lock_queues(); 1053 /* 1054 * The page might have been moved to another 1055 * queue during potential blocking in vget() 1056 * above. The page might have been freed and 1057 * reused for another vnode. 1058 */ 1059 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE || 1060 m->object != object || 1061 TAILQ_NEXT(m, pageq) != &marker) { 1062 vm_page_unlock(m); 1063 if (object->flags & OBJ_MIGHTBEDIRTY) 1064 vnodes_skipped++; 1065 goto unlock_and_continue; 1066 } 1067 1068 /* 1069 * The page may have been busied during the 1070 * blocking in vget(). We don't move the 1071 * page back onto the end of the queue so that 1072 * statistics are more correct if we don't. 1073 */ 1074 if (m->busy || (m->oflags & VPO_BUSY)) { 1075 vm_page_unlock(m); 1076 goto unlock_and_continue; 1077 } 1078 1079 /* 1080 * If the page has become held it might 1081 * be undergoing I/O, so skip it 1082 */ 1083 if (m->hold_count) { 1084 vm_page_unlock(m); 1085 vm_page_requeue(m); 1086 if (object->flags & OBJ_MIGHTBEDIRTY) 1087 vnodes_skipped++; 1088 goto unlock_and_continue; 1089 } 1090 } 1091 vm_page_unlock(m); 1092 1093 /* 1094 * If a page is dirty, then it is either being washed 1095 * (but not yet cleaned) or it is still in the 1096 * laundry. If it is still in the laundry, then we 1097 * start the cleaning operation. 1098 * 1099 * decrement page_shortage on success to account for 1100 * the (future) cleaned page. Otherwise we could wind 1101 * up laundering or cleaning too many pages. 1102 */ 1103 vm_page_unlock_queues(); 1104 if (vm_pageout_clean(m) != 0) { 1105 --page_shortage; 1106 --maxlaunder; 1107 } 1108 vm_page_lock_queues(); 1109unlock_and_continue: 1110 vm_page_lock_assert(m, MA_NOTOWNED); 1111 VM_OBJECT_UNLOCK(object); 1112 if (mp != NULL) { 1113 vm_page_unlock_queues(); 1114 if (vp != NULL) 1115 vput(vp); 1116 VFS_UNLOCK_GIANT(vfslocked); 1117 vm_object_deallocate(object); 1118 vn_finished_write(mp); 1119 vm_page_lock_queues(); 1120 } 1121 next = TAILQ_NEXT(&marker, pageq); 1122 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, 1123 &marker, pageq); 1124 vm_page_lock_assert(m, MA_NOTOWNED); 1125 continue; 1126 } 1127 vm_page_unlock(m); 1128 VM_OBJECT_UNLOCK(object); 1129 } 1130 1131 /* 1132 * Compute the number of pages we want to try to move from the 1133 * active queue to the inactive queue. 1134 */ 1135 page_shortage = vm_paging_target() + 1136 cnt.v_inactive_target - cnt.v_inactive_count; 1137 page_shortage += addl_page_shortage; 1138 1139 /* 1140 * Scan the active queue for things we can deactivate. We nominally 1141 * track the per-page activity counter and use it to locate 1142 * deactivation candidates. 1143 */ 1144 pcount = cnt.v_active_count; 1145 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1146 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1147 1148 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) { 1149 1150 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE), 1151 ("vm_pageout_scan: page %p isn't active", m)); 1152 1153 next = TAILQ_NEXT(m, pageq); 1154 object = m->object; 1155 if ((m->flags & PG_MARKER) != 0) { 1156 m = next; 1157 continue; 1158 } 1159 if (!vm_pageout_page_lock(m, &next) || 1160 (object = m->object) == NULL) { 1161 vm_page_unlock(m); 1162 m = next; 1163 continue; 1164 } 1165 if (!VM_OBJECT_TRYLOCK(object) && 1166 !vm_pageout_fallback_object_lock(m, &next)) { 1167 VM_OBJECT_UNLOCK(object); 1168 vm_page_unlock(m); 1169 m = next; 1170 continue; 1171 } 1172 1173 /* 1174 * Don't deactivate pages that are busy. 1175 */ 1176 if ((m->busy != 0) || 1177 (m->oflags & VPO_BUSY) || 1178 (m->hold_count != 0)) { 1179 vm_page_unlock(m); 1180 VM_OBJECT_UNLOCK(object); 1181 vm_page_requeue(m); 1182 m = next; 1183 continue; 1184 } 1185 1186 /* 1187 * The count for pagedaemon pages is done after checking the 1188 * page for eligibility... 1189 */ 1190 cnt.v_pdpages++; 1191 1192 /* 1193 * Check to see "how much" the page has been used. 1194 */ 1195 actcount = 0; 1196 if (object->ref_count != 0) { 1197 if (m->flags & PG_REFERENCED) { 1198 actcount += 1; 1199 } 1200 actcount += pmap_ts_referenced(m); 1201 if (actcount) { 1202 m->act_count += ACT_ADVANCE + actcount; 1203 if (m->act_count > ACT_MAX) 1204 m->act_count = ACT_MAX; 1205 } 1206 } 1207 1208 /* 1209 * Since we have "tested" this bit, we need to clear it now. 1210 */ 1211 vm_page_flag_clear(m, PG_REFERENCED); 1212 1213 /* 1214 * Only if an object is currently being used, do we use the 1215 * page activation count stats. 1216 */ 1217 if (actcount && (object->ref_count != 0)) { 1218 vm_page_requeue(m); 1219 } else { 1220 m->act_count -= min(m->act_count, ACT_DECLINE); 1221 if (vm_pageout_algorithm || 1222 object->ref_count == 0 || 1223 m->act_count == 0) { 1224 page_shortage--; 1225 if (object->ref_count == 0) { 1226 KASSERT(!pmap_page_is_mapped(m), 1227 ("vm_pageout_scan: page %p is mapped", m)); 1228 if (m->dirty == 0) 1229 vm_page_cache(m); 1230 else 1231 vm_page_deactivate(m); 1232 } else { 1233 vm_page_deactivate(m); 1234 } 1235 } else { 1236 vm_page_requeue(m); 1237 } 1238 } 1239 vm_page_unlock(m); 1240 VM_OBJECT_UNLOCK(object); 1241 m = next; 1242 } 1243 vm_page_unlock_queues(); 1244#if !defined(NO_SWAPPING) 1245 /* 1246 * Idle process swapout -- run once per second. 1247 */ 1248 if (vm_swap_idle_enabled) { 1249 static long lsec; 1250 if (time_second != lsec) { 1251 vm_req_vmdaemon(VM_SWAP_IDLE); 1252 lsec = time_second; 1253 } 1254 } 1255#endif 1256 1257 /* 1258 * If we didn't get enough free pages, and we have skipped a vnode 1259 * in a writeable object, wakeup the sync daemon. And kick swapout 1260 * if we did not get enough free pages. 1261 */ 1262 if (vm_paging_target() > 0) { 1263 if (vnodes_skipped && vm_page_count_min()) 1264 (void) speedup_syncer(); 1265#if !defined(NO_SWAPPING) 1266 if (vm_swap_enabled && vm_page_count_target()) 1267 vm_req_vmdaemon(VM_SWAP_NORMAL); 1268#endif 1269 } 1270 1271 /* 1272 * If we are critically low on one of RAM or swap and low on 1273 * the other, kill the largest process. However, we avoid 1274 * doing this on the first pass in order to give ourselves a 1275 * chance to flush out dirty vnode-backed pages and to allow 1276 * active pages to be moved to the inactive queue and reclaimed. 1277 */ 1278 if (pass != 0 && 1279 ((swap_pager_avail < 64 && vm_page_count_min()) || 1280 (swap_pager_full && vm_paging_target() > 0))) 1281 vm_pageout_oom(VM_OOM_MEM); 1282} 1283 1284 1285void 1286vm_pageout_oom(int shortage) 1287{ 1288 struct proc *p, *bigproc; 1289 vm_offset_t size, bigsize; 1290 struct thread *td; 1291 struct vmspace *vm; 1292 1293 /* 1294 * We keep the process bigproc locked once we find it to keep anyone 1295 * from messing with it; however, there is a possibility of 1296 * deadlock if process B is bigproc and one of it's child processes 1297 * attempts to propagate a signal to B while we are waiting for A's 1298 * lock while walking this list. To avoid this, we don't block on 1299 * the process lock but just skip a process if it is already locked. 1300 */ 1301 bigproc = NULL; 1302 bigsize = 0; 1303 sx_slock(&allproc_lock); 1304 FOREACH_PROC_IN_SYSTEM(p) { 1305 int breakout; 1306 1307 if (PROC_TRYLOCK(p) == 0) 1308 continue; 1309 /* 1310 * If this is a system, protected or killed process, skip it. 1311 */ 1312 if ((p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM)) || 1313 (p->p_pid == 1) || P_KILLED(p) || 1314 ((p->p_pid < 48) && (swap_pager_avail != 0))) { 1315 PROC_UNLOCK(p); 1316 continue; 1317 } 1318 /* 1319 * If the process is in a non-running type state, 1320 * don't touch it. Check all the threads individually. 1321 */ 1322 breakout = 0; 1323 FOREACH_THREAD_IN_PROC(p, td) { 1324 thread_lock(td); 1325 if (!TD_ON_RUNQ(td) && 1326 !TD_IS_RUNNING(td) && 1327 !TD_IS_SLEEPING(td)) { 1328 thread_unlock(td); 1329 breakout = 1; 1330 break; 1331 } 1332 thread_unlock(td); 1333 } 1334 if (breakout) { 1335 PROC_UNLOCK(p); 1336 continue; 1337 } 1338 /* 1339 * get the process size 1340 */ 1341 vm = vmspace_acquire_ref(p); 1342 if (vm == NULL) { 1343 PROC_UNLOCK(p); 1344 continue; 1345 } 1346 if (!vm_map_trylock_read(&vm->vm_map)) { 1347 vmspace_free(vm); 1348 PROC_UNLOCK(p); 1349 continue; 1350 } 1351 size = vmspace_swap_count(vm); 1352 vm_map_unlock_read(&vm->vm_map); 1353 if (shortage == VM_OOM_MEM) 1354 size += vmspace_resident_count(vm); 1355 vmspace_free(vm); 1356 /* 1357 * if the this process is bigger than the biggest one 1358 * remember it. 1359 */ 1360 if (size > bigsize) { 1361 if (bigproc != NULL) 1362 PROC_UNLOCK(bigproc); 1363 bigproc = p; 1364 bigsize = size; 1365 } else 1366 PROC_UNLOCK(p); 1367 } 1368 sx_sunlock(&allproc_lock); 1369 if (bigproc != NULL) { 1370 killproc(bigproc, "out of swap space"); 1371 sched_nice(bigproc, PRIO_MIN); 1372 PROC_UNLOCK(bigproc); 1373 wakeup(&cnt.v_free_count); 1374 } 1375} 1376 1377/* 1378 * This routine tries to maintain the pseudo LRU active queue, 1379 * so that during long periods of time where there is no paging, 1380 * that some statistic accumulation still occurs. This code 1381 * helps the situation where paging just starts to occur. 1382 */ 1383static void 1384vm_pageout_page_stats() 1385{ 1386 vm_object_t object; 1387 vm_page_t m,next; 1388 int pcount,tpcount; /* Number of pages to check */ 1389 static int fullintervalcount = 0; 1390 int page_shortage; 1391 1392 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1393 page_shortage = 1394 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) - 1395 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count); 1396 1397 if (page_shortage <= 0) 1398 return; 1399 1400 pcount = cnt.v_active_count; 1401 fullintervalcount += vm_pageout_stats_interval; 1402 if (fullintervalcount < vm_pageout_full_stats_interval) { 1403 tpcount = (int64_t)vm_pageout_stats_max * cnt.v_active_count / 1404 cnt.v_page_count; 1405 if (pcount > tpcount) 1406 pcount = tpcount; 1407 } else { 1408 fullintervalcount = 0; 1409 } 1410 1411 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1412 while ((m != NULL) && (pcount-- > 0)) { 1413 int actcount; 1414 1415 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE), 1416 ("vm_pageout_page_stats: page %p isn't active", m)); 1417 1418 next = TAILQ_NEXT(m, pageq); 1419 if ((m->flags & PG_MARKER) != 0) { 1420 m = next; 1421 continue; 1422 } 1423 vm_page_lock_assert(m, MA_NOTOWNED); 1424 if (!vm_pageout_page_lock(m, &next) || 1425 (object = m->object) == NULL) { 1426 vm_page_unlock(m); 1427 m = next; 1428 continue; 1429 } 1430 if (!VM_OBJECT_TRYLOCK(object) && 1431 !vm_pageout_fallback_object_lock(m, &next)) { 1432 VM_OBJECT_UNLOCK(object); 1433 vm_page_unlock(m); 1434 m = next; 1435 continue; 1436 } 1437 1438 /* 1439 * Don't deactivate pages that are busy. 1440 */ 1441 if ((m->busy != 0) || 1442 (m->oflags & VPO_BUSY) || 1443 (m->hold_count != 0)) { 1444 vm_page_unlock(m); 1445 VM_OBJECT_UNLOCK(object); 1446 vm_page_requeue(m); 1447 m = next; 1448 continue; 1449 } 1450 1451 actcount = 0; 1452 if (m->flags & PG_REFERENCED) { 1453 vm_page_flag_clear(m, PG_REFERENCED); 1454 actcount += 1; 1455 } 1456 1457 actcount += pmap_ts_referenced(m); 1458 if (actcount) { 1459 m->act_count += ACT_ADVANCE + actcount; 1460 if (m->act_count > ACT_MAX) 1461 m->act_count = ACT_MAX; 1462 vm_page_requeue(m); 1463 } else { 1464 if (m->act_count == 0) { 1465 /* 1466 * We turn off page access, so that we have 1467 * more accurate RSS stats. We don't do this 1468 * in the normal page deactivation when the 1469 * system is loaded VM wise, because the 1470 * cost of the large number of page protect 1471 * operations would be higher than the value 1472 * of doing the operation. 1473 */ 1474 pmap_remove_all(m); 1475 vm_page_deactivate(m); 1476 } else { 1477 m->act_count -= min(m->act_count, ACT_DECLINE); 1478 vm_page_requeue(m); 1479 } 1480 } 1481 vm_page_unlock(m); 1482 VM_OBJECT_UNLOCK(object); 1483 m = next; 1484 } 1485} 1486 1487/* 1488 * vm_pageout is the high level pageout daemon. 1489 */ 1490static void 1491vm_pageout() 1492{ 1493 int error, pass; 1494 1495 /* 1496 * Initialize some paging parameters. 1497 */ 1498 cnt.v_interrupt_free_min = 2; 1499 if (cnt.v_page_count < 2000) 1500 vm_pageout_page_count = 8; 1501 1502 /* 1503 * v_free_reserved needs to include enough for the largest 1504 * swap pager structures plus enough for any pv_entry structs 1505 * when paging. 1506 */ 1507 if (cnt.v_page_count > 1024) 1508 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200; 1509 else 1510 cnt.v_free_min = 4; 1511 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1512 cnt.v_interrupt_free_min; 1513 cnt.v_free_reserved = vm_pageout_page_count + 1514 cnt.v_pageout_free_min + (cnt.v_page_count / 768); 1515 cnt.v_free_severe = cnt.v_free_min / 2; 1516 cnt.v_free_min += cnt.v_free_reserved; 1517 cnt.v_free_severe += cnt.v_free_reserved; 1518 1519 /* 1520 * v_free_target and v_cache_min control pageout hysteresis. Note 1521 * that these are more a measure of the VM cache queue hysteresis 1522 * then the VM free queue. Specifically, v_free_target is the 1523 * high water mark (free+cache pages). 1524 * 1525 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the 1526 * low water mark, while v_free_min is the stop. v_cache_min must 1527 * be big enough to handle memory needs while the pageout daemon 1528 * is signalled and run to free more pages. 1529 */ 1530 if (cnt.v_free_count > 6144) 1531 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved; 1532 else 1533 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved; 1534 1535 if (cnt.v_free_count > 2048) { 1536 cnt.v_cache_min = cnt.v_free_target; 1537 cnt.v_cache_max = 2 * cnt.v_cache_min; 1538 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2; 1539 } else { 1540 cnt.v_cache_min = 0; 1541 cnt.v_cache_max = 0; 1542 cnt.v_inactive_target = cnt.v_free_count / 4; 1543 } 1544 if (cnt.v_inactive_target > cnt.v_free_count / 3) 1545 cnt.v_inactive_target = cnt.v_free_count / 3; 1546 1547 /* XXX does not really belong here */ 1548 if (vm_page_max_wired == 0) 1549 vm_page_max_wired = cnt.v_free_count / 3; 1550 1551 if (vm_pageout_stats_max == 0) 1552 vm_pageout_stats_max = cnt.v_free_target; 1553 1554 /* 1555 * Set interval in seconds for stats scan. 1556 */ 1557 if (vm_pageout_stats_interval == 0) 1558 vm_pageout_stats_interval = 5; 1559 if (vm_pageout_full_stats_interval == 0) 1560 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4; 1561 1562 swap_pager_swap_init(); 1563 pass = 0; 1564 /* 1565 * The pageout daemon is never done, so loop forever. 1566 */ 1567 while (TRUE) { 1568 /* 1569 * If we have enough free memory, wakeup waiters. Do 1570 * not clear vm_pages_needed until we reach our target, 1571 * otherwise we may be woken up over and over again and 1572 * waste a lot of cpu. 1573 */ 1574 mtx_lock(&vm_page_queue_free_mtx); 1575 if (vm_pages_needed && !vm_page_count_min()) { 1576 if (!vm_paging_needed()) 1577 vm_pages_needed = 0; 1578 wakeup(&cnt.v_free_count); 1579 } 1580 if (vm_pages_needed) { 1581 /* 1582 * Still not done, take a second pass without waiting 1583 * (unlimited dirty cleaning), otherwise sleep a bit 1584 * and try again. 1585 */ 1586 ++pass; 1587 if (pass > 1) 1588 msleep(&vm_pages_needed, 1589 &vm_page_queue_free_mtx, PVM, "psleep", 1590 hz / 2); 1591 } else { 1592 /* 1593 * Good enough, sleep & handle stats. Prime the pass 1594 * for the next run. 1595 */ 1596 if (pass > 1) 1597 pass = 1; 1598 else 1599 pass = 0; 1600 error = msleep(&vm_pages_needed, 1601 &vm_page_queue_free_mtx, PVM, "psleep", 1602 vm_pageout_stats_interval * hz); 1603 if (error && !vm_pages_needed) { 1604 mtx_unlock(&vm_page_queue_free_mtx); 1605 pass = 0; 1606 vm_page_lock_queues(); 1607 vm_pageout_page_stats(); 1608 vm_page_unlock_queues(); 1609 continue; 1610 } 1611 } 1612 if (vm_pages_needed) 1613 cnt.v_pdwakeups++; 1614 mtx_unlock(&vm_page_queue_free_mtx); 1615 vm_pageout_scan(pass); 1616 } 1617} 1618 1619/* 1620 * Unless the free page queue lock is held by the caller, this function 1621 * should be regarded as advisory. Specifically, the caller should 1622 * not msleep() on &cnt.v_free_count following this function unless 1623 * the free page queue lock is held until the msleep() is performed. 1624 */ 1625void 1626pagedaemon_wakeup() 1627{ 1628 1629 if (!vm_pages_needed && curthread->td_proc != pageproc) { 1630 vm_pages_needed = 1; 1631 wakeup(&vm_pages_needed); 1632 } 1633} 1634 1635#if !defined(NO_SWAPPING) 1636static void 1637vm_req_vmdaemon(int req) 1638{ 1639 static int lastrun = 0; 1640 1641 mtx_lock(&vm_daemon_mtx); 1642 vm_pageout_req_swapout |= req; 1643 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1644 wakeup(&vm_daemon_needed); 1645 lastrun = ticks; 1646 } 1647 mtx_unlock(&vm_daemon_mtx); 1648} 1649 1650static void 1651vm_daemon() 1652{ 1653 struct rlimit rsslim; 1654 struct proc *p; 1655 struct thread *td; 1656 struct vmspace *vm; 1657 int breakout, swapout_flags; 1658 1659 while (TRUE) { 1660 mtx_lock(&vm_daemon_mtx); 1661 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0); 1662 swapout_flags = vm_pageout_req_swapout; 1663 vm_pageout_req_swapout = 0; 1664 mtx_unlock(&vm_daemon_mtx); 1665 if (swapout_flags) 1666 swapout_procs(swapout_flags); 1667 1668 /* 1669 * scan the processes for exceeding their rlimits or if 1670 * process is swapped out -- deactivate pages 1671 */ 1672 sx_slock(&allproc_lock); 1673 FOREACH_PROC_IN_SYSTEM(p) { 1674 vm_pindex_t limit, size; 1675 1676 /* 1677 * if this is a system process or if we have already 1678 * looked at this process, skip it. 1679 */ 1680 PROC_LOCK(p); 1681 if (p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) { 1682 PROC_UNLOCK(p); 1683 continue; 1684 } 1685 /* 1686 * if the process is in a non-running type state, 1687 * don't touch it. 1688 */ 1689 breakout = 0; 1690 FOREACH_THREAD_IN_PROC(p, td) { 1691 thread_lock(td); 1692 if (!TD_ON_RUNQ(td) && 1693 !TD_IS_RUNNING(td) && 1694 !TD_IS_SLEEPING(td)) { 1695 thread_unlock(td); 1696 breakout = 1; 1697 break; 1698 } 1699 thread_unlock(td); 1700 } 1701 if (breakout) { 1702 PROC_UNLOCK(p); 1703 continue; 1704 } 1705 /* 1706 * get a limit 1707 */ 1708 lim_rlimit(p, RLIMIT_RSS, &rsslim); 1709 limit = OFF_TO_IDX( 1710 qmin(rsslim.rlim_cur, rsslim.rlim_max)); 1711 1712 /* 1713 * let processes that are swapped out really be 1714 * swapped out set the limit to nothing (will force a 1715 * swap-out.) 1716 */ 1717 if ((p->p_flag & P_INMEM) == 0) 1718 limit = 0; /* XXX */ 1719 vm = vmspace_acquire_ref(p); 1720 PROC_UNLOCK(p); 1721 if (vm == NULL) 1722 continue; 1723 1724 size = vmspace_resident_count(vm); 1725 if (limit >= 0 && size >= limit) { 1726 vm_pageout_map_deactivate_pages( 1727 &vm->vm_map, limit); 1728 } 1729 vmspace_free(vm); 1730 } 1731 sx_sunlock(&allproc_lock); 1732 } 1733} 1734#endif /* !defined(NO_SWAPPING) */ 1735