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("$FreeBSD$"); 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/racct.h> 90 #include <sys/resourcevar.h> 91 #include <sys/sched.h> 92 #include <sys/signalvar.h> 93 #include <sys/vnode.h> 94 #include <sys/vmmeter.h> 95 #include <sys/sx.h> 96 #include <sys/sysctl.h> 97 98 #include <vm/vm.h> 99 #include <vm/vm_param.h> 100 #include <vm/vm_object.h> 101 #include <vm/vm_page.h> 102 #include <vm/vm_map.h> 103 #include <vm/vm_pageout.h> 104 #include <vm/vm_pager.h> 105 #include <vm/swap_pager.h> 106 #include <vm/vm_extern.h> 107 #include <vm/uma.h> 108 109 /* 110 * System initialization 111 */ 112 113 /* the kernel process "vm_pageout"*/ 114 static void vm_pageout(void); 115 static int vm_pageout_clean(vm_page_t); 116 static void vm_pageout_scan(int pass); 117 118 struct proc *pageproc; 119 120 static struct kproc_desc page_kp = { 121 "pagedaemon", 122 vm_pageout, 123 &pageproc 124 }; 125 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, 126 &page_kp); 127 128 #if !defined(NO_SWAPPING) 129 /* the kernel process "vm_daemon"*/ 130 static void vm_daemon(void); 131 static struct proc *vmproc; 132 133 static struct kproc_desc vm_kp = { 134 "vmdaemon", 135 vm_daemon, 136 &vmproc 137 }; 138 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp); 139 #endif 140 141 142 int vm_pages_needed; /* Event on which pageout daemon sleeps */ 143 int vm_pageout_deficit; /* Estimated number of pages deficit */ 144 int vm_pageout_pages_needed; /* flag saying that the pageout daemon needs pages */ 145 146 #if !defined(NO_SWAPPING) 147 static int vm_pageout_req_swapout; /* XXX */ 148 static int vm_daemon_needed; 149 static struct mtx vm_daemon_mtx; 150 /* Allow for use by vm_pageout before vm_daemon is initialized. */ 151 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF); 152 #endif 153 static int vm_max_launder = 32; 154 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0; 155 static int vm_pageout_full_stats_interval = 0; 156 static int vm_pageout_algorithm=0; 157 static int defer_swap_pageouts=0; 158 static int disable_swap_pageouts=0; 159 160 #if defined(NO_SWAPPING) 161 static int vm_swap_enabled=0; 162 static int vm_swap_idle_enabled=0; 163 #else 164 static int vm_swap_enabled=1; 165 static int vm_swap_idle_enabled=0; 166 #endif 167 168 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm, 169 CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt"); 170 171 SYSCTL_INT(_vm, OID_AUTO, max_launder, 172 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout"); 173 174 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max, 175 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length"); 176 177 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval, 178 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan"); 179 180 SYSCTL_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) 184 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 185 CTLFLAG_RD, &vm_swap_enabled, 0, "Enable entire process swapout"); 186 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 187 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 188 #else 189 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 190 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout"); 191 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 192 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 193 #endif 194 195 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts, 196 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem"); 197 198 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts, 199 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages"); 200 201 static int pageout_lock_miss; 202 SYSCTL_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 16 206 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT; 207 208 int vm_page_max_wired; /* XXX max # of wired pages system-wide */ 209 SYSCTL_INT(_vm, OID_AUTO, max_wired, 210 CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count"); 211 212 #if !defined(NO_SWAPPING) 213 static void vm_pageout_map_deactivate_pages(vm_map_t, long); 214 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long); 215 static void vm_req_vmdaemon(int req); 216 #endif 217 static void vm_pageout_page_stats(void); 218 219 /* 220 * Initialize a dummy page for marking the caller's place in the specified 221 * paging queue. In principle, this function only needs to set the flag 222 * PG_MARKER. Nonetheless, it sets the flag VPO_BUSY and initializes the hold 223 * count to one as safety precautions. 224 */ 225 static void 226 vm_pageout_init_marker(vm_page_t marker, u_short queue) 227 { 228 229 bzero(marker, sizeof(*marker)); 230 marker->flags = PG_MARKER; 231 marker->oflags = VPO_BUSY; 232 marker->queue = queue; 233 marker->hold_count = 1; 234 } 235 236 /* 237 * vm_pageout_fallback_object_lock: 238 * 239 * Lock vm object currently associated with `m'. VM_OBJECT_TRYLOCK is 240 * known to have failed and page queue must be either PQ_ACTIVE or 241 * PQ_INACTIVE. To avoid lock order violation, unlock the page queues 242 * while locking the vm object. Use marker page to detect page queue 243 * changes and maintain notion of next page on page queue. Return 244 * TRUE if no changes were detected, FALSE otherwise. vm object is 245 * locked on return. 246 * 247 * This function depends on both the lock portion of struct vm_object 248 * and normal struct vm_page being type stable. 249 */ 250 boolean_t 251 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next) 252 { 253 struct vm_page marker; 254 boolean_t unchanged; 255 u_short queue; 256 vm_object_t object; 257 258 queue = m->queue; 259 vm_pageout_init_marker(&marker, queue); 260 object = m->object; 261 262 TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl, 263 m, &marker, pageq); 264 vm_page_unlock_queues(); 265 vm_page_unlock(m); 266 VM_OBJECT_LOCK(object); 267 vm_page_lock(m); 268 vm_page_lock_queues(); 269 270 /* Page queue might have changed. */ 271 *next = TAILQ_NEXT(&marker, pageq); 272 unchanged = (m->queue == queue && 273 m->object == object && 274 &marker == TAILQ_NEXT(m, pageq)); 275 TAILQ_REMOVE(&vm_page_queues[queue].pl, 276 &marker, pageq); 277 return (unchanged); 278 } 279 280 /* 281 * Lock the page while holding the page queue lock. Use marker page 282 * to detect page queue changes and maintain notion of next page on 283 * page queue. Return TRUE if no changes were detected, FALSE 284 * otherwise. The page is locked on return. The page queue lock might 285 * be dropped and reacquired. 286 * 287 * This function depends on normal struct vm_page being type stable. 288 */ 289 boolean_t 290 vm_pageout_page_lock(vm_page_t m, vm_page_t *next) 291 { 292 struct vm_page marker; 293 boolean_t unchanged; 294 u_short queue; 295 296 vm_page_lock_assert(m, MA_NOTOWNED); 297 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 298 299 if (vm_page_trylock(m)) 300 return (TRUE); 301 302 queue = m->queue; 303 vm_pageout_init_marker(&marker, queue); 304 305 TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl, m, &marker, pageq); 306 vm_page_unlock_queues(); 307 vm_page_lock(m); 308 vm_page_lock_queues(); 309 310 /* Page queue might have changed. */ 311 *next = TAILQ_NEXT(&marker, pageq); 312 unchanged = (m->queue == queue && &marker == TAILQ_NEXT(m, pageq)); 313 TAILQ_REMOVE(&vm_page_queues[queue].pl, &marker, pageq); 314 return (unchanged); 315 } 316 317 /* 318 * vm_pageout_clean: 319 * 320 * Clean the page and remove it from the laundry. 321 * 322 * We set the busy bit to cause potential page faults on this page to 323 * block. Note the careful timing, however, the busy bit isn't set till 324 * late and we cannot do anything that will mess with the page. 325 */ 326 static int 327 vm_pageout_clean(vm_page_t m) 328 { 329 vm_object_t object; 330 vm_page_t mc[2*vm_pageout_page_count], pb, ps; 331 int pageout_count; 332 int ib, is, page_base; 333 vm_pindex_t pindex = m->pindex; 334 335 vm_page_lock_assert(m, MA_OWNED); 336 object = m->object; 337 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 338 339 /* 340 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP 341 * with the new swapper, but we could have serious problems paging 342 * out other object types if there is insufficient memory. 343 * 344 * Unfortunately, checking free memory here is far too late, so the 345 * check has been moved up a procedural level. 346 */ 347 348 /* 349 * Can't clean the page if it's busy or held. 350 */ 351 KASSERT(m->busy == 0 && (m->oflags & VPO_BUSY) == 0, 352 ("vm_pageout_clean: page %p is busy", m)); 353 KASSERT(m->hold_count == 0, ("vm_pageout_clean: page %p is held", m)); 354 vm_page_unlock(m); 355 356 mc[vm_pageout_page_count] = pb = ps = m; 357 pageout_count = 1; 358 page_base = vm_pageout_page_count; 359 ib = 1; 360 is = 1; 361 362 /* 363 * Scan object for clusterable pages. 364 * 365 * We can cluster ONLY if: ->> the page is NOT 366 * clean, wired, busy, held, or mapped into a 367 * buffer, and one of the following: 368 * 1) The page is inactive, or a seldom used 369 * active page. 370 * -or- 371 * 2) we force the issue. 372 * 373 * During heavy mmap/modification loads the pageout 374 * daemon can really fragment the underlying file 375 * due to flushing pages out of order and not trying 376 * align the clusters (which leave sporatic out-of-order 377 * holes). To solve this problem we do the reverse scan 378 * first and attempt to align our cluster, then do a 379 * forward scan if room remains. 380 */ 381 more: 382 while (ib && pageout_count < vm_pageout_page_count) { 383 vm_page_t p; 384 385 if (ib > pindex) { 386 ib = 0; 387 break; 388 } 389 390 if ((p = vm_page_prev(pb)) == NULL || 391 (p->oflags & VPO_BUSY) != 0 || p->busy != 0) { 392 ib = 0; 393 break; 394 } 395 vm_page_lock(p); 396 vm_page_test_dirty(p); 397 if (p->dirty == 0 || 398 p->queue != PQ_INACTIVE || 399 p->hold_count != 0) { /* may be undergoing I/O */ 400 vm_page_unlock(p); 401 ib = 0; 402 break; 403 } 404 vm_page_unlock(p); 405 mc[--page_base] = pb = p; 406 ++pageout_count; 407 ++ib; 408 /* 409 * alignment boundry, stop here and switch directions. Do 410 * not clear ib. 411 */ 412 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0) 413 break; 414 } 415 416 while (pageout_count < vm_pageout_page_count && 417 pindex + is < object->size) { 418 vm_page_t p; 419 420 if ((p = vm_page_next(ps)) == NULL || 421 (p->oflags & VPO_BUSY) != 0 || p->busy != 0) 422 break; 423 vm_page_lock(p); 424 vm_page_test_dirty(p); 425 if (p->dirty == 0 || 426 p->queue != PQ_INACTIVE || 427 p->hold_count != 0) { /* may be undergoing I/O */ 428 vm_page_unlock(p); 429 break; 430 } 431 vm_page_unlock(p); 432 mc[page_base + pageout_count] = ps = p; 433 ++pageout_count; 434 ++is; 435 } 436 437 /* 438 * If we exhausted our forward scan, continue with the reverse scan 439 * when possible, even past a page boundry. This catches boundry 440 * conditions. 441 */ 442 if (ib && pageout_count < vm_pageout_page_count) 443 goto more; 444 445 /* 446 * we allow reads during pageouts... 447 */ 448 return (vm_pageout_flush(&mc[page_base], pageout_count, 0, 0, NULL, 449 NULL)); 450 } 451 452 /* 453 * vm_pageout_flush() - launder the given pages 454 * 455 * The given pages are laundered. Note that we setup for the start of 456 * I/O ( i.e. busy the page ), mark it read-only, and bump the object 457 * reference count all in here rather then in the parent. If we want 458 * the parent to do more sophisticated things we may have to change 459 * the ordering. 460 * 461 * Returned runlen is the count of pages between mreq and first 462 * page after mreq with status VM_PAGER_AGAIN. 463 * *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL 464 * for any page in runlen set. 465 */ 466 int 467 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen, 468 boolean_t *eio) 469 { 470 vm_object_t object = mc[0]->object; 471 int pageout_status[count]; 472 int numpagedout = 0; 473 int i, runlen; 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 pmap_remove_write(mc[i]); 494 } 495 vm_object_pip_add(object, count); 496 497 vm_pager_put_pages(object, mc, count, flags, pageout_status); 498 499 runlen = count - mreq; 500 if (eio != NULL) 501 *eio = FALSE; 502 for (i = 0; i < count; i++) { 503 vm_page_t mt = mc[i]; 504 505 KASSERT(pageout_status[i] == VM_PAGER_PEND || 506 (mt->aflags & PGA_WRITEABLE) == 0, 507 ("vm_pageout_flush: page %p is not write protected", mt)); 508 switch (pageout_status[i]) { 509 case VM_PAGER_OK: 510 case VM_PAGER_PEND: 511 numpagedout++; 512 break; 513 case VM_PAGER_BAD: 514 /* 515 * Page outside of range of object. Right now we 516 * essentially lose the changes by pretending it 517 * worked. 518 */ 519 vm_page_undirty(mt); 520 break; 521 case VM_PAGER_ERROR: 522 case VM_PAGER_FAIL: 523 /* 524 * If page couldn't be paged out, then reactivate the 525 * page so it doesn't clog the inactive list. (We 526 * will try paging out it again later). 527 */ 528 vm_page_lock(mt); 529 vm_page_activate(mt); 530 vm_page_unlock(mt); 531 if (eio != NULL && i >= mreq && i - mreq < runlen) 532 *eio = TRUE; 533 break; 534 case VM_PAGER_AGAIN: 535 if (i >= mreq && i - mreq < runlen) 536 runlen = i - mreq; 537 break; 538 } 539 540 /* 541 * If the operation is still going, leave the page busy to 542 * block all other accesses. Also, leave the paging in 543 * progress indicator set so that we don't attempt an object 544 * collapse. 545 */ 546 if (pageout_status[i] != VM_PAGER_PEND) { 547 vm_object_pip_wakeup(object); 548 vm_page_io_finish(mt); 549 if (vm_page_count_severe()) { 550 vm_page_lock(mt); 551 vm_page_try_to_cache(mt); 552 vm_page_unlock(mt); 553 } 554 } 555 } 556 if (prunlen != NULL) 557 *prunlen = runlen; 558 return (numpagedout); 559 } 560 561 #if !defined(NO_SWAPPING) 562 /* 563 * vm_pageout_object_deactivate_pages 564 * 565 * Deactivate enough pages to satisfy the inactive target 566 * requirements. 567 * 568 * The object and map must be locked. 569 */ 570 static void 571 vm_pageout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object, 572 long desired) 573 { 574 vm_object_t backing_object, object; 575 vm_page_t p; 576 int actcount, remove_mode; 577 578 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED); 579 if (first_object->type == OBJT_DEVICE || 580 first_object->type == OBJT_SG) 581 return; 582 for (object = first_object;; object = backing_object) { 583 if (pmap_resident_count(pmap) <= desired) 584 goto unlock_return; 585 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 586 if (object->type == OBJT_PHYS || object->paging_in_progress) 587 goto unlock_return; 588 589 remove_mode = 0; 590 if (object->shadow_count > 1) 591 remove_mode = 1; 592 /* 593 * Scan the object's entire memory queue. 594 */ 595 TAILQ_FOREACH(p, &object->memq, listq) { 596 if (pmap_resident_count(pmap) <= desired) 597 goto unlock_return; 598 if ((p->oflags & VPO_BUSY) != 0 || p->busy != 0) 599 continue; 600 PCPU_INC(cnt.v_pdpages); 601 vm_page_lock(p); 602 if (p->wire_count != 0 || p->hold_count != 0 || 603 !pmap_page_exists_quick(pmap, p)) { 604 vm_page_unlock(p); 605 continue; 606 } 607 actcount = pmap_ts_referenced(p); 608 if ((p->aflags & PGA_REFERENCED) != 0) { 609 if (actcount == 0) 610 actcount = 1; 611 vm_page_aflag_clear(p, PGA_REFERENCED); 612 } 613 if (p->queue != PQ_ACTIVE && actcount != 0) { 614 vm_page_activate(p); 615 p->act_count += actcount; 616 } else if (p->queue == PQ_ACTIVE) { 617 if (actcount == 0) { 618 p->act_count -= min(p->act_count, 619 ACT_DECLINE); 620 if (!remove_mode && 621 (vm_pageout_algorithm || 622 p->act_count == 0)) { 623 pmap_remove_all(p); 624 vm_page_deactivate(p); 625 } else { 626 vm_page_lock_queues(); 627 vm_page_requeue(p); 628 vm_page_unlock_queues(); 629 } 630 } else { 631 vm_page_activate(p); 632 if (p->act_count < ACT_MAX - 633 ACT_ADVANCE) 634 p->act_count += ACT_ADVANCE; 635 vm_page_lock_queues(); 636 vm_page_requeue(p); 637 vm_page_unlock_queues(); 638 } 639 } else if (p->queue == PQ_INACTIVE) 640 pmap_remove_all(p); 641 vm_page_unlock(p); 642 } 643 if ((backing_object = object->backing_object) == NULL) 644 goto unlock_return; 645 VM_OBJECT_LOCK(backing_object); 646 if (object != first_object) 647 VM_OBJECT_UNLOCK(object); 648 } 649 unlock_return: 650 if (object != first_object) 651 VM_OBJECT_UNLOCK(object); 652 } 653 654 /* 655 * deactivate some number of pages in a map, try to do it fairly, but 656 * that is really hard to do. 657 */ 658 static void 659 vm_pageout_map_deactivate_pages(map, desired) 660 vm_map_t map; 661 long desired; 662 { 663 vm_map_entry_t tmpe; 664 vm_object_t obj, bigobj; 665 int nothingwired; 666 667 if (!vm_map_trylock(map)) 668 return; 669 670 bigobj = NULL; 671 nothingwired = TRUE; 672 673 /* 674 * first, search out the biggest object, and try to free pages from 675 * that. 676 */ 677 tmpe = map->header.next; 678 while (tmpe != &map->header) { 679 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 680 obj = tmpe->object.vm_object; 681 if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) { 682 if (obj->shadow_count <= 1 && 683 (bigobj == NULL || 684 bigobj->resident_page_count < obj->resident_page_count)) { 685 if (bigobj != NULL) 686 VM_OBJECT_UNLOCK(bigobj); 687 bigobj = obj; 688 } else 689 VM_OBJECT_UNLOCK(obj); 690 } 691 } 692 if (tmpe->wired_count > 0) 693 nothingwired = FALSE; 694 tmpe = tmpe->next; 695 } 696 697 if (bigobj != NULL) { 698 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired); 699 VM_OBJECT_UNLOCK(bigobj); 700 } 701 /* 702 * Next, hunt around for other pages to deactivate. We actually 703 * do this search sort of wrong -- .text first is not the best idea. 704 */ 705 tmpe = map->header.next; 706 while (tmpe != &map->header) { 707 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 708 break; 709 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 710 obj = tmpe->object.vm_object; 711 if (obj != NULL) { 712 VM_OBJECT_LOCK(obj); 713 vm_pageout_object_deactivate_pages(map->pmap, obj, desired); 714 VM_OBJECT_UNLOCK(obj); 715 } 716 } 717 tmpe = tmpe->next; 718 } 719 720 /* 721 * Remove all mappings if a process is swapped out, this will free page 722 * table pages. 723 */ 724 if (desired == 0 && nothingwired) { 725 pmap_remove(vm_map_pmap(map), vm_map_min(map), 726 vm_map_max(map)); 727 } 728 vm_map_unlock(map); 729 } 730 #endif /* !defined(NO_SWAPPING) */ 731 732 /* 733 * vm_pageout_scan does the dirty work for the pageout daemon. 734 */ 735 static void 736 vm_pageout_scan(int pass) 737 { 738 vm_page_t m, next; 739 struct vm_page marker; 740 int page_shortage, maxscan, pcount; 741 int addl_page_shortage, addl_page_shortage_init; 742 vm_object_t object; 743 int actcount; 744 int vnodes_skipped = 0; 745 int maxlaunder; 746 747 /* 748 * Decrease registered cache sizes. 749 */ 750 EVENTHANDLER_INVOKE(vm_lowmem, 0); 751 /* 752 * We do this explicitly after the caches have been drained above. 753 */ 754 uma_reclaim(); 755 756 addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit); 757 758 /* 759 * Calculate the number of pages we want to either free or move 760 * to the cache. 761 */ 762 page_shortage = vm_paging_target() + addl_page_shortage_init; 763 764 vm_pageout_init_marker(&marker, PQ_INACTIVE); 765 766 /* 767 * Start scanning the inactive queue for pages we can move to the 768 * cache or free. The scan will stop when the target is reached or 769 * we have scanned the entire inactive queue. Note that m->act_count 770 * is not used to form decisions for the inactive queue, only for the 771 * active queue. 772 * 773 * maxlaunder limits the number of dirty pages we flush per scan. 774 * For most systems a smaller value (16 or 32) is more robust under 775 * extreme memory and disk pressure because any unnecessary writes 776 * to disk can result in extreme performance degredation. However, 777 * systems with excessive dirty pages (especially when MAP_NOSYNC is 778 * used) will die horribly with limited laundering. If the pageout 779 * daemon cannot clean enough pages in the first pass, we let it go 780 * all out in succeeding passes. 781 */ 782 if ((maxlaunder = vm_max_launder) <= 1) 783 maxlaunder = 1; 784 if (pass) 785 maxlaunder = 10000; 786 vm_page_lock_queues(); 787 rescan0: 788 addl_page_shortage = addl_page_shortage_init; 789 maxscan = cnt.v_inactive_count; 790 791 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl); 792 m != NULL && maxscan-- > 0 && page_shortage > 0; 793 m = next) { 794 795 cnt.v_pdpages++; 796 797 if (m->queue != PQ_INACTIVE) 798 goto rescan0; 799 800 next = TAILQ_NEXT(m, pageq); 801 802 /* 803 * skip marker pages 804 */ 805 if (m->flags & PG_MARKER) 806 continue; 807 808 KASSERT((m->flags & PG_FICTITIOUS) == 0, 809 ("Fictitious page %p cannot be in inactive queue", m)); 810 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 811 ("Unmanaged page %p cannot be in inactive queue", m)); 812 813 /* 814 * Lock the page. 815 */ 816 if (!vm_pageout_page_lock(m, &next)) { 817 vm_page_unlock(m); 818 addl_page_shortage++; 819 continue; 820 } 821 822 /* 823 * A held page may be undergoing I/O, so skip it. 824 */ 825 if (m->hold_count) { 826 vm_page_unlock(m); 827 vm_page_requeue(m); 828 addl_page_shortage++; 829 continue; 830 } 831 832 /* 833 * Don't mess with busy pages, keep in the front of the 834 * queue, most likely are being paged out. 835 */ 836 object = m->object; 837 if (!VM_OBJECT_TRYLOCK(object) && 838 (!vm_pageout_fallback_object_lock(m, &next) || 839 m->hold_count != 0)) { 840 VM_OBJECT_UNLOCK(object); 841 vm_page_unlock(m); 842 addl_page_shortage++; 843 continue; 844 } 845 if (m->busy || (m->oflags & VPO_BUSY)) { 846 vm_page_unlock(m); 847 VM_OBJECT_UNLOCK(object); 848 addl_page_shortage++; 849 continue; 850 } 851 852 /* 853 * If the object is not being used, we ignore previous 854 * references. 855 */ 856 if (object->ref_count == 0) { 857 vm_page_aflag_clear(m, PGA_REFERENCED); 858 KASSERT(!pmap_page_is_mapped(m), 859 ("vm_pageout_scan: page %p is mapped", m)); 860 861 /* 862 * Otherwise, if the page has been referenced while in the 863 * inactive queue, we bump the "activation count" upwards, 864 * making it less likely that the page will be added back to 865 * the inactive queue prematurely again. Here we check the 866 * page tables (or emulated bits, if any), given the upper 867 * level VM system not knowing anything about existing 868 * references. 869 */ 870 } else if (((m->aflags & PGA_REFERENCED) == 0) && 871 (actcount = pmap_ts_referenced(m))) { 872 vm_page_activate(m); 873 vm_page_unlock(m); 874 m->act_count += actcount + ACT_ADVANCE; 875 VM_OBJECT_UNLOCK(object); 876 continue; 877 } 878 879 /* 880 * If the upper level VM system knows about any page 881 * references, we activate the page. We also set the 882 * "activation count" higher than normal so that we will less 883 * likely place pages back onto the inactive queue again. 884 */ 885 if ((m->aflags & PGA_REFERENCED) != 0) { 886 vm_page_aflag_clear(m, PGA_REFERENCED); 887 actcount = pmap_ts_referenced(m); 888 vm_page_activate(m); 889 vm_page_unlock(m); 890 m->act_count += actcount + ACT_ADVANCE + 1; 891 VM_OBJECT_UNLOCK(object); 892 continue; 893 } 894 895 /* 896 * If the upper level VM system does not believe that the page 897 * is fully dirty, but it is mapped for write access, then we 898 * consult the pmap to see if the page's dirty status should 899 * be updated. 900 */ 901 if (m->dirty != VM_PAGE_BITS_ALL && 902 (m->aflags & PGA_WRITEABLE) != 0) { 903 /* 904 * Avoid a race condition: Unless write access is 905 * removed from the page, another processor could 906 * modify it before all access is removed by the call 907 * to vm_page_cache() below. If vm_page_cache() finds 908 * that the page has been modified when it removes all 909 * access, it panics because it cannot cache dirty 910 * pages. In principle, we could eliminate just write 911 * access here rather than all access. In the expected 912 * case, when there are no last instant modifications 913 * to the page, removing all access will be cheaper 914 * overall. 915 */ 916 if (pmap_is_modified(m)) 917 vm_page_dirty(m); 918 else if (m->dirty == 0) 919 pmap_remove_all(m); 920 } 921 922 if (m->valid == 0) { 923 /* 924 * Invalid pages can be easily freed 925 */ 926 vm_page_free(m); 927 cnt.v_dfree++; 928 --page_shortage; 929 } else if (m->dirty == 0) { 930 /* 931 * Clean pages can be placed onto the cache queue. 932 * This effectively frees them. 933 */ 934 vm_page_cache(m); 935 --page_shortage; 936 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) { 937 /* 938 * Dirty pages need to be paged out, but flushing 939 * a page is extremely expensive verses freeing 940 * a clean page. Rather then artificially limiting 941 * the number of pages we can flush, we instead give 942 * dirty pages extra priority on the inactive queue 943 * by forcing them to be cycled through the queue 944 * twice before being flushed, after which the 945 * (now clean) page will cycle through once more 946 * before being freed. This significantly extends 947 * the thrash point for a heavily loaded machine. 948 */ 949 m->flags |= PG_WINATCFLS; 950 vm_page_requeue(m); 951 } else if (maxlaunder > 0) { 952 /* 953 * We always want to try to flush some dirty pages if 954 * we encounter them, to keep the system stable. 955 * Normally this number is small, but under extreme 956 * pressure where there are insufficient clean pages 957 * on the inactive queue, we may have to go all out. 958 */ 959 int swap_pageouts_ok, vfslocked = 0; 960 struct vnode *vp = NULL; 961 struct mount *mp = NULL; 962 963 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 964 swap_pageouts_ok = 1; 965 } else { 966 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 967 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 968 vm_page_count_min()); 969 970 } 971 972 /* 973 * We don't bother paging objects that are "dead". 974 * Those objects are in a "rundown" state. 975 */ 976 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 977 vm_page_unlock(m); 978 VM_OBJECT_UNLOCK(object); 979 vm_page_requeue(m); 980 continue; 981 } 982 983 /* 984 * Following operations may unlock 985 * vm_page_queue_mtx, invalidating the 'next' 986 * pointer. To prevent an inordinate number 987 * of restarts we use our marker to remember 988 * our place. 989 * 990 */ 991 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, 992 m, &marker, pageq); 993 /* 994 * The object is already known NOT to be dead. It 995 * is possible for the vget() to block the whole 996 * pageout daemon, but the new low-memory handling 997 * code should prevent it. 998 * 999 * The previous code skipped locked vnodes and, worse, 1000 * reordered pages in the queue. This results in 1001 * completely non-deterministic operation and, on a 1002 * busy system, can lead to extremely non-optimal 1003 * pageouts. For example, it can cause clean pages 1004 * to be freed and dirty pages to be moved to the end 1005 * of the queue. Since dirty pages are also moved to 1006 * the end of the queue once-cleaned, this gives 1007 * way too large a weighting to defering the freeing 1008 * of dirty pages. 1009 * 1010 * We can't wait forever for the vnode lock, we might 1011 * deadlock due to a vn_read() getting stuck in 1012 * vm_wait while holding this vnode. We skip the 1013 * vnode if we can't get it in a reasonable amount 1014 * of time. 1015 */ 1016 if (object->type == OBJT_VNODE) { 1017 vm_page_unlock_queues(); 1018 vm_page_unlock(m); 1019 vp = object->handle; 1020 if (vp->v_type == VREG && 1021 vn_start_write(vp, &mp, V_NOWAIT) != 0) { 1022 mp = NULL; 1023 ++pageout_lock_miss; 1024 if (object->flags & OBJ_MIGHTBEDIRTY) 1025 vnodes_skipped++; 1026 vm_page_lock_queues(); 1027 goto unlock_and_continue; 1028 } 1029 KASSERT(mp != NULL, 1030 ("vp %p with NULL v_mount", vp)); 1031 vm_object_reference_locked(object); 1032 VM_OBJECT_UNLOCK(object); 1033 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 1034 if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK, 1035 curthread)) { 1036 VM_OBJECT_LOCK(object); 1037 vm_page_lock_queues(); 1038 ++pageout_lock_miss; 1039 if (object->flags & OBJ_MIGHTBEDIRTY) 1040 vnodes_skipped++; 1041 vp = NULL; 1042 goto unlock_and_continue; 1043 } 1044 VM_OBJECT_LOCK(object); 1045 vm_page_lock(m); 1046 vm_page_lock_queues(); 1047 /* 1048 * The page might have been moved to another 1049 * queue during potential blocking in vget() 1050 * above. The page might have been freed and 1051 * reused for another vnode. 1052 */ 1053 if (m->queue != PQ_INACTIVE || 1054 m->object != object || 1055 TAILQ_NEXT(m, pageq) != &marker) { 1056 vm_page_unlock(m); 1057 if (object->flags & OBJ_MIGHTBEDIRTY) 1058 vnodes_skipped++; 1059 goto unlock_and_continue; 1060 } 1061 1062 /* 1063 * The page may have been busied during the 1064 * blocking in vget(). We don't move the 1065 * page back onto the end of the queue so that 1066 * statistics are more correct if we don't. 1067 */ 1068 if (m->busy || (m->oflags & VPO_BUSY)) { 1069 vm_page_unlock(m); 1070 goto unlock_and_continue; 1071 } 1072 1073 /* 1074 * If the page has become held it might 1075 * be undergoing I/O, so skip it 1076 */ 1077 if (m->hold_count) { 1078 vm_page_unlock(m); 1079 vm_page_requeue(m); 1080 if (object->flags & OBJ_MIGHTBEDIRTY) 1081 vnodes_skipped++; 1082 goto unlock_and_continue; 1083 } 1084 } 1085 1086 /* 1087 * If a page is dirty, then it is either being washed 1088 * (but not yet cleaned) or it is still in the 1089 * laundry. If it is still in the laundry, then we 1090 * start the cleaning operation. 1091 * 1092 * decrement page_shortage on success to account for 1093 * the (future) cleaned page. Otherwise we could wind 1094 * up laundering or cleaning too many pages. 1095 */ 1096 vm_page_unlock_queues(); 1097 if (vm_pageout_clean(m) != 0) { 1098 --page_shortage; 1099 --maxlaunder; 1100 } 1101 vm_page_lock_queues(); 1102 unlock_and_continue: 1103 vm_page_lock_assert(m, MA_NOTOWNED); 1104 VM_OBJECT_UNLOCK(object); 1105 if (mp != NULL) { 1106 vm_page_unlock_queues(); 1107 if (vp != NULL) 1108 vput(vp); 1109 VFS_UNLOCK_GIANT(vfslocked); 1110 vm_object_deallocate(object); 1111 vn_finished_write(mp); 1112 vm_page_lock_queues(); 1113 } 1114 next = TAILQ_NEXT(&marker, pageq); 1115 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, 1116 &marker, pageq); 1117 vm_page_lock_assert(m, MA_NOTOWNED); 1118 continue; 1119 } 1120 vm_page_unlock(m); 1121 VM_OBJECT_UNLOCK(object); 1122 } 1123 1124 /* 1125 * Compute the number of pages we want to try to move from the 1126 * active queue to the inactive queue. 1127 */ 1128 page_shortage = vm_paging_target() + 1129 cnt.v_inactive_target - cnt.v_inactive_count; 1130 page_shortage += addl_page_shortage; 1131 1132 /* 1133 * Scan the active queue for things we can deactivate. We nominally 1134 * track the per-page activity counter and use it to locate 1135 * deactivation candidates. 1136 */ 1137 pcount = cnt.v_active_count; 1138 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1139 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1140 1141 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) { 1142 1143 KASSERT(m->queue == PQ_ACTIVE, 1144 ("vm_pageout_scan: page %p isn't active", m)); 1145 1146 next = TAILQ_NEXT(m, pageq); 1147 if ((m->flags & PG_MARKER) != 0) { 1148 m = next; 1149 continue; 1150 } 1151 KASSERT((m->flags & PG_FICTITIOUS) == 0, 1152 ("Fictitious page %p cannot be in active queue", m)); 1153 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 1154 ("Unmanaged page %p cannot be in active queue", m)); 1155 if (!vm_pageout_page_lock(m, &next)) { 1156 vm_page_unlock(m); 1157 m = next; 1158 continue; 1159 } 1160 object = m->object; 1161 if (!VM_OBJECT_TRYLOCK(object) && 1162 !vm_pageout_fallback_object_lock(m, &next)) { 1163 VM_OBJECT_UNLOCK(object); 1164 vm_page_unlock(m); 1165 m = next; 1166 continue; 1167 } 1168 1169 /* 1170 * Don't deactivate pages that are busy. 1171 */ 1172 if ((m->busy != 0) || 1173 (m->oflags & VPO_BUSY) || 1174 (m->hold_count != 0)) { 1175 vm_page_unlock(m); 1176 VM_OBJECT_UNLOCK(object); 1177 vm_page_requeue(m); 1178 m = next; 1179 continue; 1180 } 1181 1182 /* 1183 * The count for pagedaemon pages is done after checking the 1184 * page for eligibility... 1185 */ 1186 cnt.v_pdpages++; 1187 1188 /* 1189 * Check to see "how much" the page has been used. 1190 */ 1191 actcount = 0; 1192 if (object->ref_count != 0) { 1193 if (m->aflags & PGA_REFERENCED) { 1194 actcount += 1; 1195 } 1196 actcount += pmap_ts_referenced(m); 1197 if (actcount) { 1198 m->act_count += ACT_ADVANCE + actcount; 1199 if (m->act_count > ACT_MAX) 1200 m->act_count = ACT_MAX; 1201 } 1202 } 1203 1204 /* 1205 * Since we have "tested" this bit, we need to clear it now. 1206 */ 1207 vm_page_aflag_clear(m, PGA_REFERENCED); 1208 1209 /* 1210 * Only if an object is currently being used, do we use the 1211 * page activation count stats. 1212 */ 1213 if (actcount && (object->ref_count != 0)) { 1214 vm_page_requeue(m); 1215 } else { 1216 m->act_count -= min(m->act_count, ACT_DECLINE); 1217 if (vm_pageout_algorithm || 1218 object->ref_count == 0 || 1219 m->act_count == 0) { 1220 page_shortage--; 1221 if (object->ref_count == 0) { 1222 KASSERT(!pmap_page_is_mapped(m), 1223 ("vm_pageout_scan: page %p is mapped", m)); 1224 if (m->dirty == 0) 1225 vm_page_cache(m); 1226 else 1227 vm_page_deactivate(m); 1228 } else { 1229 vm_page_deactivate(m); 1230 } 1231 } else { 1232 vm_page_requeue(m); 1233 } 1234 } 1235 vm_page_unlock(m); 1236 VM_OBJECT_UNLOCK(object); 1237 m = next; 1238 } 1239 vm_page_unlock_queues(); 1240 #if !defined(NO_SWAPPING) 1241 /* 1242 * Idle process swapout -- run once per second. 1243 */ 1244 if (vm_swap_idle_enabled) { 1245 static long lsec; 1246 if (time_second != lsec) { 1247 vm_req_vmdaemon(VM_SWAP_IDLE); 1248 lsec = time_second; 1249 } 1250 } 1251 #endif 1252 1253 /* 1254 * If we didn't get enough free pages, and we have skipped a vnode 1255 * in a writeable object, wakeup the sync daemon. And kick swapout 1256 * if we did not get enough free pages. 1257 */ 1258 if (vm_paging_target() > 0) { 1259 if (vnodes_skipped && vm_page_count_min()) 1260 (void) speedup_syncer(); 1261 #if !defined(NO_SWAPPING) 1262 if (vm_swap_enabled && vm_page_count_target()) 1263 vm_req_vmdaemon(VM_SWAP_NORMAL); 1264 #endif 1265 } 1266 1267 /* 1268 * If we are critically low on one of RAM or swap and low on 1269 * the other, kill the largest process. However, we avoid 1270 * doing this on the first pass in order to give ourselves a 1271 * chance to flush out dirty vnode-backed pages and to allow 1272 * active pages to be moved to the inactive queue and reclaimed. 1273 */ 1274 if (pass != 0 && 1275 ((swap_pager_avail < 64 && vm_page_count_min()) || 1276 (swap_pager_full && vm_paging_target() > 0))) 1277 vm_pageout_oom(VM_OOM_MEM); 1278 } 1279 1280 1281 void 1282 vm_pageout_oom(int shortage) 1283 { 1284 struct proc *p, *bigproc; 1285 vm_offset_t size, bigsize; 1286 struct thread *td; 1287 struct vmspace *vm; 1288 1289 /* 1290 * We keep the process bigproc locked once we find it to keep anyone 1291 * from messing with it; however, there is a possibility of 1292 * deadlock if process B is bigproc and one of it's child processes 1293 * attempts to propagate a signal to B while we are waiting for A's 1294 * lock while walking this list. To avoid this, we don't block on 1295 * the process lock but just skip a process if it is already locked. 1296 */ 1297 bigproc = NULL; 1298 bigsize = 0; 1299 sx_slock(&allproc_lock); 1300 FOREACH_PROC_IN_SYSTEM(p) { 1301 int breakout; 1302 1303 if (PROC_TRYLOCK(p) == 0) 1304 continue; 1305 /* 1306 * If this is a system, protected or killed process, skip it. 1307 */ 1308 if (p->p_state != PRS_NORMAL || 1309 (p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM)) || 1310 (p->p_pid == 1) || P_KILLED(p) || 1311 ((p->p_pid < 48) && (swap_pager_avail != 0))) { 1312 PROC_UNLOCK(p); 1313 continue; 1314 } 1315 /* 1316 * If the process is in a non-running type state, 1317 * don't touch it. Check all the threads individually. 1318 */ 1319 breakout = 0; 1320 FOREACH_THREAD_IN_PROC(p, td) { 1321 thread_lock(td); 1322 if (!TD_ON_RUNQ(td) && 1323 !TD_IS_RUNNING(td) && 1324 !TD_IS_SLEEPING(td) && 1325 !TD_IS_SUSPENDED(td)) { 1326 thread_unlock(td); 1327 breakout = 1; 1328 break; 1329 } 1330 thread_unlock(td); 1331 } 1332 if (breakout) { 1333 PROC_UNLOCK(p); 1334 continue; 1335 } 1336 /* 1337 * get the process size 1338 */ 1339 vm = vmspace_acquire_ref(p); 1340 if (vm == NULL) { 1341 PROC_UNLOCK(p); 1342 continue; 1343 } 1344 if (!vm_map_trylock_read(&vm->vm_map)) { 1345 vmspace_free(vm); 1346 PROC_UNLOCK(p); 1347 continue; 1348 } 1349 size = vmspace_swap_count(vm); 1350 vm_map_unlock_read(&vm->vm_map); 1351 if (shortage == VM_OOM_MEM) 1352 size += vmspace_resident_count(vm); 1353 vmspace_free(vm); 1354 /* 1355 * if the this process is bigger than the biggest one 1356 * remember it. 1357 */ 1358 if (size > bigsize) { 1359 if (bigproc != NULL) 1360 PROC_UNLOCK(bigproc); 1361 bigproc = p; 1362 bigsize = size; 1363 } else 1364 PROC_UNLOCK(p); 1365 } 1366 sx_sunlock(&allproc_lock); 1367 if (bigproc != NULL) { 1368 killproc(bigproc, "out of swap space"); 1369 sched_nice(bigproc, PRIO_MIN); 1370 PROC_UNLOCK(bigproc); 1371 wakeup(&cnt.v_free_count); 1372 } 1373 } 1374 1375 /* 1376 * This routine tries to maintain the pseudo LRU active queue, 1377 * so that during long periods of time where there is no paging, 1378 * that some statistic accumulation still occurs. This code 1379 * helps the situation where paging just starts to occur. 1380 */ 1381 static void 1382 vm_pageout_page_stats() 1383 { 1384 vm_object_t object; 1385 vm_page_t m,next; 1386 int pcount,tpcount; /* Number of pages to check */ 1387 static int fullintervalcount = 0; 1388 int page_shortage; 1389 1390 page_shortage = 1391 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) - 1392 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count); 1393 1394 if (page_shortage <= 0) 1395 return; 1396 1397 vm_page_lock_queues(); 1398 pcount = cnt.v_active_count; 1399 fullintervalcount += vm_pageout_stats_interval; 1400 if (fullintervalcount < vm_pageout_full_stats_interval) { 1401 tpcount = (int64_t)vm_pageout_stats_max * cnt.v_active_count / 1402 cnt.v_page_count; 1403 if (pcount > tpcount) 1404 pcount = tpcount; 1405 } else { 1406 fullintervalcount = 0; 1407 } 1408 1409 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1410 while ((m != NULL) && (pcount-- > 0)) { 1411 int actcount; 1412 1413 KASSERT(m->queue == PQ_ACTIVE, 1414 ("vm_pageout_page_stats: page %p isn't active", m)); 1415 1416 next = TAILQ_NEXT(m, pageq); 1417 if ((m->flags & PG_MARKER) != 0) { 1418 m = next; 1419 continue; 1420 } 1421 vm_page_lock_assert(m, MA_NOTOWNED); 1422 if (!vm_pageout_page_lock(m, &next)) { 1423 vm_page_unlock(m); 1424 m = next; 1425 continue; 1426 } 1427 object = m->object; 1428 if (!VM_OBJECT_TRYLOCK(object) && 1429 !vm_pageout_fallback_object_lock(m, &next)) { 1430 VM_OBJECT_UNLOCK(object); 1431 vm_page_unlock(m); 1432 m = next; 1433 continue; 1434 } 1435 1436 /* 1437 * Don't deactivate pages that are busy. 1438 */ 1439 if ((m->busy != 0) || 1440 (m->oflags & VPO_BUSY) || 1441 (m->hold_count != 0)) { 1442 vm_page_unlock(m); 1443 VM_OBJECT_UNLOCK(object); 1444 vm_page_requeue(m); 1445 m = next; 1446 continue; 1447 } 1448 1449 actcount = 0; 1450 if (m->aflags & PGA_REFERENCED) { 1451 vm_page_aflag_clear(m, PGA_REFERENCED); 1452 actcount += 1; 1453 } 1454 1455 actcount += pmap_ts_referenced(m); 1456 if (actcount) { 1457 m->act_count += ACT_ADVANCE + actcount; 1458 if (m->act_count > ACT_MAX) 1459 m->act_count = ACT_MAX; 1460 vm_page_requeue(m); 1461 } else { 1462 if (m->act_count == 0) { 1463 /* 1464 * We turn off page access, so that we have 1465 * more accurate RSS stats. We don't do this 1466 * in the normal page deactivation when the 1467 * system is loaded VM wise, because the 1468 * cost of the large number of page protect 1469 * operations would be higher than the value 1470 * of doing the operation. 1471 */ 1472 pmap_remove_all(m); 1473 vm_page_deactivate(m); 1474 } else { 1475 m->act_count -= min(m->act_count, ACT_DECLINE); 1476 vm_page_requeue(m); 1477 } 1478 } 1479 vm_page_unlock(m); 1480 VM_OBJECT_UNLOCK(object); 1481 m = next; 1482 } 1483 vm_page_unlock_queues(); 1484 } 1485 1486 /* 1487 * vm_pageout is the high level pageout daemon. 1488 */ 1489 static void 1490 vm_pageout() 1491 { 1492 int error, pass; 1493 1494 /* 1495 * Initialize some paging parameters. 1496 */ 1497 cnt.v_interrupt_free_min = 2; 1498 if (cnt.v_page_count < 2000) 1499 vm_pageout_page_count = 8; 1500 1501 /* 1502 * v_free_reserved needs to include enough for the largest 1503 * swap pager structures plus enough for any pv_entry structs 1504 * when paging. 1505 */ 1506 if (cnt.v_page_count > 1024) 1507 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200; 1508 else 1509 cnt.v_free_min = 4; 1510 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1511 cnt.v_interrupt_free_min; 1512 cnt.v_free_reserved = vm_pageout_page_count + 1513 cnt.v_pageout_free_min + (cnt.v_page_count / 768); 1514 cnt.v_free_severe = cnt.v_free_min / 2; 1515 cnt.v_free_min += cnt.v_free_reserved; 1516 cnt.v_free_severe += cnt.v_free_reserved; 1517 1518 /* 1519 * v_free_target and v_cache_min control pageout hysteresis. Note 1520 * that these are more a measure of the VM cache queue hysteresis 1521 * then the VM free queue. Specifically, v_free_target is the 1522 * high water mark (free+cache pages). 1523 * 1524 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the 1525 * low water mark, while v_free_min is the stop. v_cache_min must 1526 * be big enough to handle memory needs while the pageout daemon 1527 * is signalled and run to free more pages. 1528 */ 1529 if (cnt.v_free_count > 6144) 1530 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved; 1531 else 1532 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved; 1533 1534 if (cnt.v_free_count > 2048) { 1535 cnt.v_cache_min = cnt.v_free_target; 1536 cnt.v_cache_max = 2 * cnt.v_cache_min; 1537 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2; 1538 } else { 1539 cnt.v_cache_min = 0; 1540 cnt.v_cache_max = 0; 1541 cnt.v_inactive_target = cnt.v_free_count / 4; 1542 } 1543 if (cnt.v_inactive_target > cnt.v_free_count / 3) 1544 cnt.v_inactive_target = cnt.v_free_count / 3; 1545 1546 /* XXX does not really belong here */ 1547 if (vm_page_max_wired == 0) 1548 vm_page_max_wired = cnt.v_free_count / 3; 1549 1550 if (vm_pageout_stats_max == 0) 1551 vm_pageout_stats_max = cnt.v_free_target; 1552 1553 /* 1554 * Set interval in seconds for stats scan. 1555 */ 1556 if (vm_pageout_stats_interval == 0) 1557 vm_pageout_stats_interval = 5; 1558 if (vm_pageout_full_stats_interval == 0) 1559 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4; 1560 1561 swap_pager_swap_init(); 1562 pass = 0; 1563 /* 1564 * The pageout daemon is never done, so loop forever. 1565 */ 1566 while (TRUE) { 1567 /* 1568 * If we have enough free memory, wakeup waiters. Do 1569 * not clear vm_pages_needed until we reach our target, 1570 * otherwise we may be woken up over and over again and 1571 * waste a lot of cpu. 1572 */ 1573 mtx_lock(&vm_page_queue_free_mtx); 1574 if (vm_pages_needed && !vm_page_count_min()) { 1575 if (!vm_paging_needed()) 1576 vm_pages_needed = 0; 1577 wakeup(&cnt.v_free_count); 1578 } 1579 if (vm_pages_needed) { 1580 /* 1581 * Still not done, take a second pass without waiting 1582 * (unlimited dirty cleaning), otherwise sleep a bit 1583 * and try again. 1584 */ 1585 ++pass; 1586 if (pass > 1) 1587 msleep(&vm_pages_needed, 1588 &vm_page_queue_free_mtx, PVM, "psleep", 1589 hz / 2); 1590 } else { 1591 /* 1592 * Good enough, sleep & handle stats. Prime the pass 1593 * for the next run. 1594 */ 1595 if (pass > 1) 1596 pass = 1; 1597 else 1598 pass = 0; 1599 error = msleep(&vm_pages_needed, 1600 &vm_page_queue_free_mtx, PVM, "psleep", 1601 vm_pageout_stats_interval * hz); 1602 if (error && !vm_pages_needed) { 1603 mtx_unlock(&vm_page_queue_free_mtx); 1604 pass = 0; 1605 vm_pageout_page_stats(); 1606 continue; 1607 } 1608 } 1609 if (vm_pages_needed) 1610 cnt.v_pdwakeups++; 1611 mtx_unlock(&vm_page_queue_free_mtx); 1612 vm_pageout_scan(pass); 1613 } 1614 } 1615 1616 /* 1617 * Unless the free page queue lock is held by the caller, this function 1618 * should be regarded as advisory. Specifically, the caller should 1619 * not msleep() on &cnt.v_free_count following this function unless 1620 * the free page queue lock is held until the msleep() is performed. 1621 */ 1622 void 1623 pagedaemon_wakeup() 1624 { 1625 1626 if (!vm_pages_needed && curthread->td_proc != pageproc) { 1627 vm_pages_needed = 1; 1628 wakeup(&vm_pages_needed); 1629 } 1630 } 1631 1632 #if !defined(NO_SWAPPING) 1633 static void 1634 vm_req_vmdaemon(int req) 1635 { 1636 static int lastrun = 0; 1637 1638 mtx_lock(&vm_daemon_mtx); 1639 vm_pageout_req_swapout |= req; 1640 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1641 wakeup(&vm_daemon_needed); 1642 lastrun = ticks; 1643 } 1644 mtx_unlock(&vm_daemon_mtx); 1645 } 1646 1647 static void 1648 vm_daemon() 1649 { 1650 struct rlimit rsslim; 1651 struct proc *p; 1652 struct thread *td; 1653 struct vmspace *vm; 1654 int breakout, swapout_flags, tryagain, attempts; 1655 #ifdef RACCT 1656 uint64_t rsize, ravailable; 1657 #endif 1658 1659 while (TRUE) { 1660 mtx_lock(&vm_daemon_mtx); 1661 #ifdef RACCT 1662 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", hz); 1663 #else 1664 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0); 1665 #endif 1666 swapout_flags = vm_pageout_req_swapout; 1667 vm_pageout_req_swapout = 0; 1668 mtx_unlock(&vm_daemon_mtx); 1669 if (swapout_flags) 1670 swapout_procs(swapout_flags); 1671 1672 /* 1673 * scan the processes for exceeding their rlimits or if 1674 * process is swapped out -- deactivate pages 1675 */ 1676 tryagain = 0; 1677 attempts = 0; 1678 again: 1679 attempts++; 1680 sx_slock(&allproc_lock); 1681 FOREACH_PROC_IN_SYSTEM(p) { 1682 vm_pindex_t limit, size; 1683 1684 /* 1685 * if this is a system process or if we have already 1686 * looked at this process, skip it. 1687 */ 1688 PROC_LOCK(p); 1689 if (p->p_state != PRS_NORMAL || 1690 p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) { 1691 PROC_UNLOCK(p); 1692 continue; 1693 } 1694 /* 1695 * if the process is in a non-running type state, 1696 * don't touch it. 1697 */ 1698 breakout = 0; 1699 FOREACH_THREAD_IN_PROC(p, td) { 1700 thread_lock(td); 1701 if (!TD_ON_RUNQ(td) && 1702 !TD_IS_RUNNING(td) && 1703 !TD_IS_SLEEPING(td) && 1704 !TD_IS_SUSPENDED(td)) { 1705 thread_unlock(td); 1706 breakout = 1; 1707 break; 1708 } 1709 thread_unlock(td); 1710 } 1711 if (breakout) { 1712 PROC_UNLOCK(p); 1713 continue; 1714 } 1715 /* 1716 * get a limit 1717 */ 1718 lim_rlimit(p, RLIMIT_RSS, &rsslim); 1719 limit = OFF_TO_IDX( 1720 qmin(rsslim.rlim_cur, rsslim.rlim_max)); 1721 1722 /* 1723 * let processes that are swapped out really be 1724 * swapped out set the limit to nothing (will force a 1725 * swap-out.) 1726 */ 1727 if ((p->p_flag & P_INMEM) == 0) 1728 limit = 0; /* XXX */ 1729 vm = vmspace_acquire_ref(p); 1730 PROC_UNLOCK(p); 1731 if (vm == NULL) 1732 continue; 1733 1734 size = vmspace_resident_count(vm); 1735 if (limit >= 0 && size >= limit) { 1736 vm_pageout_map_deactivate_pages( 1737 &vm->vm_map, limit); 1738 } 1739 #ifdef RACCT 1740 rsize = IDX_TO_OFF(size); 1741 PROC_LOCK(p); 1742 racct_set(p, RACCT_RSS, rsize); 1743 ravailable = racct_get_available(p, RACCT_RSS); 1744 PROC_UNLOCK(p); 1745 if (rsize > ravailable) { 1746 /* 1747 * Don't be overly aggressive; this might be 1748 * an innocent process, and the limit could've 1749 * been exceeded by some memory hog. Don't 1750 * try to deactivate more than 1/4th of process' 1751 * resident set size. 1752 */ 1753 if (attempts <= 8) { 1754 if (ravailable < rsize - (rsize / 4)) 1755 ravailable = rsize - (rsize / 4); 1756 } 1757 vm_pageout_map_deactivate_pages( 1758 &vm->vm_map, OFF_TO_IDX(ravailable)); 1759 /* Update RSS usage after paging out. */ 1760 size = vmspace_resident_count(vm); 1761 rsize = IDX_TO_OFF(size); 1762 PROC_LOCK(p); 1763 racct_set(p, RACCT_RSS, rsize); 1764 PROC_UNLOCK(p); 1765 if (rsize > ravailable) 1766 tryagain = 1; 1767 } 1768 #endif 1769 vmspace_free(vm); 1770 } 1771 sx_sunlock(&allproc_lock); 1772 if (tryagain != 0 && attempts <= 10) 1773 goto again; 1774 } 1775 } 1776 #endif /* !defined(NO_SWAPPING) */ 1777