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 } 450 451 /* 452 * vm_pageout_flush() - launder the given pages 453 * 454 * The given pages are laundered. Note that we setup for the start of 455 * I/O ( i.e. busy the page ), mark it read-only, and bump the object 456 * reference count all in here rather then in the parent. If we want 457 * the parent to do more sophisticated things we may have to change 458 * the ordering. 459 * 460 * Returned runlen is the count of pages between mreq and first 461 * page after mreq with status VM_PAGER_AGAIN. 462 */ 463 int 464 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen) 465 { 466 vm_object_t object = mc[0]->object; 467 int pageout_status[count]; 468 int numpagedout = 0; 469 int i, runlen; 470 471 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 472 mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED); 473 474 /* 475 * Initiate I/O. Bump the vm_page_t->busy counter and 476 * mark the pages read-only. 477 * 478 * We do not have to fixup the clean/dirty bits here... we can 479 * allow the pager to do it after the I/O completes. 480 * 481 * NOTE! mc[i]->dirty may be partial or fragmented due to an 482 * edge case with file fragments. 483 */ 484 for (i = 0; i < count; i++) { 485 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, 486 ("vm_pageout_flush: partially invalid page %p index %d/%d", 487 mc[i], i, count)); 488 vm_page_io_start(mc[i]); 489 pmap_remove_write(mc[i]); 490 } 491 vm_object_pip_add(object, count); 492 493 vm_pager_put_pages(object, mc, count, flags, pageout_status); 494 495 runlen = count - mreq; 496 for (i = 0; i < count; i++) { 497 vm_page_t mt = mc[i]; 498 499 KASSERT(pageout_status[i] == VM_PAGER_PEND || 500 (mt->aflags & PGA_WRITEABLE) == 0, 501 ("vm_pageout_flush: page %p is not write protected", mt)); 502 switch (pageout_status[i]) { 503 case VM_PAGER_OK: 504 case VM_PAGER_PEND: 505 numpagedout++; 506 break; 507 case VM_PAGER_BAD: 508 /* 509 * Page outside of range of object. Right now we 510 * essentially lose the changes by pretending it 511 * worked. 512 */ 513 vm_page_undirty(mt); 514 break; 515 case VM_PAGER_ERROR: 516 case VM_PAGER_FAIL: 517 /* 518 * If page couldn't be paged out, then reactivate the 519 * page so it doesn't clog the inactive list. (We 520 * will try paging out it again later). 521 */ 522 vm_page_lock(mt); 523 vm_page_activate(mt); 524 vm_page_unlock(mt); 525 break; 526 case VM_PAGER_AGAIN: 527 if (i >= mreq && i - mreq < runlen) 528 runlen = i - mreq; 529 break; 530 } 531 532 /* 533 * If the operation is still going, leave the page busy to 534 * block all other accesses. Also, leave the paging in 535 * progress indicator set so that we don't attempt an object 536 * collapse. 537 */ 538 if (pageout_status[i] != VM_PAGER_PEND) { 539 vm_object_pip_wakeup(object); 540 vm_page_io_finish(mt); 541 if (vm_page_count_severe()) { 542 vm_page_lock(mt); 543 vm_page_try_to_cache(mt); 544 vm_page_unlock(mt); 545 } 546 } 547 } 548 if (prunlen != NULL) 549 *prunlen = runlen; 550 return (numpagedout); 551 } 552 553 #if !defined(NO_SWAPPING) 554 /* 555 * vm_pageout_object_deactivate_pages 556 * 557 * Deactivate enough pages to satisfy the inactive target 558 * requirements. 559 * 560 * The object and map must be locked. 561 */ 562 static void 563 vm_pageout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object, 564 long desired) 565 { 566 vm_object_t backing_object, object; 567 vm_page_t p; 568 int actcount, remove_mode; 569 570 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED); 571 if (first_object->type == OBJT_DEVICE || 572 first_object->type == OBJT_SG) 573 return; 574 for (object = first_object;; object = backing_object) { 575 if (pmap_resident_count(pmap) <= desired) 576 goto unlock_return; 577 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 578 if (object->type == OBJT_PHYS || object->paging_in_progress) 579 goto unlock_return; 580 581 remove_mode = 0; 582 if (object->shadow_count > 1) 583 remove_mode = 1; 584 /* 585 * Scan the object's entire memory queue. 586 */ 587 TAILQ_FOREACH(p, &object->memq, listq) { 588 if (pmap_resident_count(pmap) <= desired) 589 goto unlock_return; 590 if ((p->oflags & VPO_BUSY) != 0 || p->busy != 0) 591 continue; 592 PCPU_INC(cnt.v_pdpages); 593 vm_page_lock(p); 594 if (p->wire_count != 0 || p->hold_count != 0 || 595 !pmap_page_exists_quick(pmap, p)) { 596 vm_page_unlock(p); 597 continue; 598 } 599 actcount = pmap_ts_referenced(p); 600 if ((p->aflags & PGA_REFERENCED) != 0) { 601 if (actcount == 0) 602 actcount = 1; 603 vm_page_aflag_clear(p, PGA_REFERENCED); 604 } 605 if (p->queue != PQ_ACTIVE && actcount != 0) { 606 vm_page_activate(p); 607 p->act_count += actcount; 608 } else if (p->queue == PQ_ACTIVE) { 609 if (actcount == 0) { 610 p->act_count -= min(p->act_count, 611 ACT_DECLINE); 612 if (!remove_mode && 613 (vm_pageout_algorithm || 614 p->act_count == 0)) { 615 pmap_remove_all(p); 616 vm_page_deactivate(p); 617 } else { 618 vm_page_lock_queues(); 619 vm_page_requeue(p); 620 vm_page_unlock_queues(); 621 } 622 } else { 623 vm_page_activate(p); 624 if (p->act_count < ACT_MAX - 625 ACT_ADVANCE) 626 p->act_count += ACT_ADVANCE; 627 vm_page_lock_queues(); 628 vm_page_requeue(p); 629 vm_page_unlock_queues(); 630 } 631 } else if (p->queue == PQ_INACTIVE) 632 pmap_remove_all(p); 633 vm_page_unlock(p); 634 } 635 if ((backing_object = object->backing_object) == NULL) 636 goto unlock_return; 637 VM_OBJECT_LOCK(backing_object); 638 if (object != first_object) 639 VM_OBJECT_UNLOCK(object); 640 } 641 unlock_return: 642 if (object != first_object) 643 VM_OBJECT_UNLOCK(object); 644 } 645 646 /* 647 * deactivate some number of pages in a map, try to do it fairly, but 648 * that is really hard to do. 649 */ 650 static void 651 vm_pageout_map_deactivate_pages(map, desired) 652 vm_map_t map; 653 long desired; 654 { 655 vm_map_entry_t tmpe; 656 vm_object_t obj, bigobj; 657 int nothingwired; 658 659 if (!vm_map_trylock(map)) 660 return; 661 662 bigobj = NULL; 663 nothingwired = TRUE; 664 665 /* 666 * first, search out the biggest object, and try to free pages from 667 * that. 668 */ 669 tmpe = map->header.next; 670 while (tmpe != &map->header) { 671 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 672 obj = tmpe->object.vm_object; 673 if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) { 674 if (obj->shadow_count <= 1 && 675 (bigobj == NULL || 676 bigobj->resident_page_count < obj->resident_page_count)) { 677 if (bigobj != NULL) 678 VM_OBJECT_UNLOCK(bigobj); 679 bigobj = obj; 680 } else 681 VM_OBJECT_UNLOCK(obj); 682 } 683 } 684 if (tmpe->wired_count > 0) 685 nothingwired = FALSE; 686 tmpe = tmpe->next; 687 } 688 689 if (bigobj != NULL) { 690 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired); 691 VM_OBJECT_UNLOCK(bigobj); 692 } 693 /* 694 * Next, hunt around for other pages to deactivate. We actually 695 * do this search sort of wrong -- .text first is not the best idea. 696 */ 697 tmpe = map->header.next; 698 while (tmpe != &map->header) { 699 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 700 break; 701 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 702 obj = tmpe->object.vm_object; 703 if (obj != NULL) { 704 VM_OBJECT_LOCK(obj); 705 vm_pageout_object_deactivate_pages(map->pmap, obj, desired); 706 VM_OBJECT_UNLOCK(obj); 707 } 708 } 709 tmpe = tmpe->next; 710 } 711 712 /* 713 * Remove all mappings if a process is swapped out, this will free page 714 * table pages. 715 */ 716 if (desired == 0 && nothingwired) { 717 tmpe = map->header.next; 718 while (tmpe != &map->header) { 719 pmap_remove(vm_map_pmap(map), tmpe->start, tmpe->end); 720 tmpe = tmpe->next; 721 } 722 } 723 vm_map_unlock(map); 724 } 725 #endif /* !defined(NO_SWAPPING) */ 726 727 /* 728 * vm_pageout_scan does the dirty work for the pageout daemon. 729 */ 730 static void 731 vm_pageout_scan(int pass) 732 { 733 vm_page_t m, next; 734 struct vm_page marker; 735 int page_shortage, maxscan, pcount; 736 int addl_page_shortage, addl_page_shortage_init; 737 vm_object_t object; 738 int actcount; 739 int vnodes_skipped = 0; 740 int maxlaunder; 741 742 /* 743 * Decrease registered cache sizes. 744 */ 745 EVENTHANDLER_INVOKE(vm_lowmem, 0); 746 /* 747 * We do this explicitly after the caches have been drained above. 748 */ 749 uma_reclaim(); 750 751 addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit); 752 753 /* 754 * Calculate the number of pages we want to either free or move 755 * to the cache. 756 */ 757 page_shortage = vm_paging_target() + addl_page_shortage_init; 758 759 vm_pageout_init_marker(&marker, PQ_INACTIVE); 760 761 /* 762 * Start scanning the inactive queue for pages we can move to the 763 * cache or free. The scan will stop when the target is reached or 764 * we have scanned the entire inactive queue. Note that m->act_count 765 * is not used to form decisions for the inactive queue, only for the 766 * active queue. 767 * 768 * maxlaunder limits the number of dirty pages we flush per scan. 769 * For most systems a smaller value (16 or 32) is more robust under 770 * extreme memory and disk pressure because any unnecessary writes 771 * to disk can result in extreme performance degredation. However, 772 * systems with excessive dirty pages (especially when MAP_NOSYNC is 773 * used) will die horribly with limited laundering. If the pageout 774 * daemon cannot clean enough pages in the first pass, we let it go 775 * all out in succeeding passes. 776 */ 777 if ((maxlaunder = vm_max_launder) <= 1) 778 maxlaunder = 1; 779 if (pass) 780 maxlaunder = 10000; 781 vm_page_lock_queues(); 782 rescan0: 783 addl_page_shortage = addl_page_shortage_init; 784 maxscan = cnt.v_inactive_count; 785 786 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl); 787 m != NULL && maxscan-- > 0 && page_shortage > 0; 788 m = next) { 789 790 cnt.v_pdpages++; 791 792 if (m->queue != PQ_INACTIVE) 793 goto rescan0; 794 795 next = TAILQ_NEXT(m, pageq); 796 797 /* 798 * skip marker pages 799 */ 800 if (m->flags & PG_MARKER) 801 continue; 802 803 /* 804 * Lock the page. 805 */ 806 if (!vm_pageout_page_lock(m, &next)) { 807 vm_page_unlock(m); 808 addl_page_shortage++; 809 continue; 810 } 811 812 /* 813 * A held page may be undergoing I/O, so skip it. 814 */ 815 if (m->hold_count) { 816 vm_page_unlock(m); 817 vm_page_requeue(m); 818 addl_page_shortage++; 819 continue; 820 } 821 822 /* 823 * Don't mess with busy pages, keep in the front of the 824 * queue, most likely are being paged out. 825 */ 826 object = m->object; 827 if (!VM_OBJECT_TRYLOCK(object) && 828 (!vm_pageout_fallback_object_lock(m, &next) || 829 m->hold_count != 0)) { 830 VM_OBJECT_UNLOCK(object); 831 vm_page_unlock(m); 832 addl_page_shortage++; 833 continue; 834 } 835 if (m->busy || (m->oflags & VPO_BUSY)) { 836 vm_page_unlock(m); 837 VM_OBJECT_UNLOCK(object); 838 addl_page_shortage++; 839 continue; 840 } 841 842 /* 843 * If the object is not being used, we ignore previous 844 * references. 845 */ 846 if (object->ref_count == 0) { 847 vm_page_aflag_clear(m, PGA_REFERENCED); 848 KASSERT(!pmap_page_is_mapped(m), 849 ("vm_pageout_scan: page %p is mapped", m)); 850 851 /* 852 * Otherwise, if the page has been referenced while in the 853 * inactive queue, we bump the "activation count" upwards, 854 * making it less likely that the page will be added back to 855 * the inactive queue prematurely again. Here we check the 856 * page tables (or emulated bits, if any), given the upper 857 * level VM system not knowing anything about existing 858 * references. 859 */ 860 } else if (((m->aflags & PGA_REFERENCED) == 0) && 861 (actcount = pmap_ts_referenced(m))) { 862 vm_page_activate(m); 863 vm_page_unlock(m); 864 m->act_count += actcount + ACT_ADVANCE; 865 VM_OBJECT_UNLOCK(object); 866 continue; 867 } 868 869 /* 870 * If the upper level VM system knows about any page 871 * references, we activate the page. We also set the 872 * "activation count" higher than normal so that we will less 873 * likely place pages back onto the inactive queue again. 874 */ 875 if ((m->aflags & PGA_REFERENCED) != 0) { 876 vm_page_aflag_clear(m, PGA_REFERENCED); 877 actcount = pmap_ts_referenced(m); 878 vm_page_activate(m); 879 vm_page_unlock(m); 880 m->act_count += actcount + ACT_ADVANCE + 1; 881 VM_OBJECT_UNLOCK(object); 882 continue; 883 } 884 885 /* 886 * If the upper level VM system does not believe that the page 887 * is fully dirty, but it is mapped for write access, then we 888 * consult the pmap to see if the page's dirty status should 889 * be updated. 890 */ 891 if (m->dirty != VM_PAGE_BITS_ALL && 892 (m->aflags & PGA_WRITEABLE) != 0) { 893 /* 894 * Avoid a race condition: Unless write access is 895 * removed from the page, another processor could 896 * modify it before all access is removed by the call 897 * to vm_page_cache() below. If vm_page_cache() finds 898 * that the page has been modified when it removes all 899 * access, it panics because it cannot cache dirty 900 * pages. In principle, we could eliminate just write 901 * access here rather than all access. In the expected 902 * case, when there are no last instant modifications 903 * to the page, removing all access will be cheaper 904 * overall. 905 */ 906 if (pmap_is_modified(m)) 907 vm_page_dirty(m); 908 else if (m->dirty == 0) 909 pmap_remove_all(m); 910 } 911 912 if (m->valid == 0) { 913 /* 914 * Invalid pages can be easily freed 915 */ 916 vm_page_free(m); 917 cnt.v_dfree++; 918 --page_shortage; 919 } else if (m->dirty == 0) { 920 /* 921 * Clean pages can be placed onto the cache queue. 922 * This effectively frees them. 923 */ 924 vm_page_cache(m); 925 --page_shortage; 926 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) { 927 /* 928 * Dirty pages need to be paged out, but flushing 929 * a page is extremely expensive verses freeing 930 * a clean page. Rather then artificially limiting 931 * the number of pages we can flush, we instead give 932 * dirty pages extra priority on the inactive queue 933 * by forcing them to be cycled through the queue 934 * twice before being flushed, after which the 935 * (now clean) page will cycle through once more 936 * before being freed. This significantly extends 937 * the thrash point for a heavily loaded machine. 938 */ 939 m->flags |= PG_WINATCFLS; 940 vm_page_requeue(m); 941 } else if (maxlaunder > 0) { 942 /* 943 * We always want to try to flush some dirty pages if 944 * we encounter them, to keep the system stable. 945 * Normally this number is small, but under extreme 946 * pressure where there are insufficient clean pages 947 * on the inactive queue, we may have to go all out. 948 */ 949 int swap_pageouts_ok, vfslocked = 0; 950 struct vnode *vp = NULL; 951 struct mount *mp = NULL; 952 953 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 954 swap_pageouts_ok = 1; 955 } else { 956 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 957 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 958 vm_page_count_min()); 959 960 } 961 962 /* 963 * We don't bother paging objects that are "dead". 964 * Those objects are in a "rundown" state. 965 */ 966 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 967 vm_page_unlock(m); 968 VM_OBJECT_UNLOCK(object); 969 vm_page_requeue(m); 970 continue; 971 } 972 973 /* 974 * Following operations may unlock 975 * vm_page_queue_mtx, invalidating the 'next' 976 * pointer. To prevent an inordinate number 977 * of restarts we use our marker to remember 978 * our place. 979 * 980 */ 981 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, 982 m, &marker, pageq); 983 /* 984 * The object is already known NOT to be dead. It 985 * is possible for the vget() to block the whole 986 * pageout daemon, but the new low-memory handling 987 * code should prevent it. 988 * 989 * The previous code skipped locked vnodes and, worse, 990 * reordered pages in the queue. This results in 991 * completely non-deterministic operation and, on a 992 * busy system, can lead to extremely non-optimal 993 * pageouts. For example, it can cause clean pages 994 * to be freed and dirty pages to be moved to the end 995 * of the queue. Since dirty pages are also moved to 996 * the end of the queue once-cleaned, this gives 997 * way too large a weighting to defering the freeing 998 * of dirty pages. 999 * 1000 * We can't wait forever for the vnode lock, we might 1001 * deadlock due to a vn_read() getting stuck in 1002 * vm_wait while holding this vnode. We skip the 1003 * vnode if we can't get it in a reasonable amount 1004 * of time. 1005 */ 1006 if (object->type == OBJT_VNODE) { 1007 vm_page_unlock_queues(); 1008 vm_page_unlock(m); 1009 vp = object->handle; 1010 if (vp->v_type == VREG && 1011 vn_start_write(vp, &mp, V_NOWAIT) != 0) { 1012 mp = NULL; 1013 ++pageout_lock_miss; 1014 if (object->flags & OBJ_MIGHTBEDIRTY) 1015 vnodes_skipped++; 1016 vm_page_lock_queues(); 1017 goto unlock_and_continue; 1018 } 1019 KASSERT(mp != NULL, 1020 ("vp %p with NULL v_mount", vp)); 1021 vm_object_reference_locked(object); 1022 VM_OBJECT_UNLOCK(object); 1023 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 1024 if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK, 1025 curthread)) { 1026 VM_OBJECT_LOCK(object); 1027 vm_page_lock_queues(); 1028 ++pageout_lock_miss; 1029 if (object->flags & OBJ_MIGHTBEDIRTY) 1030 vnodes_skipped++; 1031 vp = NULL; 1032 goto unlock_and_continue; 1033 } 1034 VM_OBJECT_LOCK(object); 1035 vm_page_lock(m); 1036 vm_page_lock_queues(); 1037 /* 1038 * The page might have been moved to another 1039 * queue during potential blocking in vget() 1040 * above. The page might have been freed and 1041 * reused for another vnode. 1042 */ 1043 if (m->queue != PQ_INACTIVE || 1044 m->object != object || 1045 TAILQ_NEXT(m, pageq) != &marker) { 1046 vm_page_unlock(m); 1047 if (object->flags & OBJ_MIGHTBEDIRTY) 1048 vnodes_skipped++; 1049 goto unlock_and_continue; 1050 } 1051 1052 /* 1053 * The page may have been busied during the 1054 * blocking in vget(). We don't move the 1055 * page back onto the end of the queue so that 1056 * statistics are more correct if we don't. 1057 */ 1058 if (m->busy || (m->oflags & VPO_BUSY)) { 1059 vm_page_unlock(m); 1060 goto unlock_and_continue; 1061 } 1062 1063 /* 1064 * If the page has become held it might 1065 * be undergoing I/O, so skip it 1066 */ 1067 if (m->hold_count) { 1068 vm_page_unlock(m); 1069 vm_page_requeue(m); 1070 if (object->flags & OBJ_MIGHTBEDIRTY) 1071 vnodes_skipped++; 1072 goto unlock_and_continue; 1073 } 1074 } 1075 1076 /* 1077 * If a page is dirty, then it is either being washed 1078 * (but not yet cleaned) or it is still in the 1079 * laundry. If it is still in the laundry, then we 1080 * start the cleaning operation. 1081 * 1082 * decrement page_shortage on success to account for 1083 * the (future) cleaned page. Otherwise we could wind 1084 * up laundering or cleaning too many pages. 1085 */ 1086 vm_page_unlock_queues(); 1087 if (vm_pageout_clean(m) != 0) { 1088 --page_shortage; 1089 --maxlaunder; 1090 } 1091 vm_page_lock_queues(); 1092 unlock_and_continue: 1093 vm_page_lock_assert(m, MA_NOTOWNED); 1094 VM_OBJECT_UNLOCK(object); 1095 if (mp != NULL) { 1096 vm_page_unlock_queues(); 1097 if (vp != NULL) 1098 vput(vp); 1099 VFS_UNLOCK_GIANT(vfslocked); 1100 vm_object_deallocate(object); 1101 vn_finished_write(mp); 1102 vm_page_lock_queues(); 1103 } 1104 next = TAILQ_NEXT(&marker, pageq); 1105 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, 1106 &marker, pageq); 1107 vm_page_lock_assert(m, MA_NOTOWNED); 1108 continue; 1109 } 1110 vm_page_unlock(m); 1111 VM_OBJECT_UNLOCK(object); 1112 } 1113 1114 /* 1115 * Compute the number of pages we want to try to move from the 1116 * active queue to the inactive queue. 1117 */ 1118 page_shortage = vm_paging_target() + 1119 cnt.v_inactive_target - cnt.v_inactive_count; 1120 page_shortage += addl_page_shortage; 1121 1122 /* 1123 * Scan the active queue for things we can deactivate. We nominally 1124 * track the per-page activity counter and use it to locate 1125 * deactivation candidates. 1126 */ 1127 pcount = cnt.v_active_count; 1128 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1129 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1130 1131 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) { 1132 1133 KASSERT(m->queue == PQ_ACTIVE, 1134 ("vm_pageout_scan: page %p isn't active", m)); 1135 1136 next = TAILQ_NEXT(m, pageq); 1137 if ((m->flags & PG_MARKER) != 0) { 1138 m = next; 1139 continue; 1140 } 1141 if (!vm_pageout_page_lock(m, &next)) { 1142 vm_page_unlock(m); 1143 m = next; 1144 continue; 1145 } 1146 object = m->object; 1147 if (!VM_OBJECT_TRYLOCK(object) && 1148 !vm_pageout_fallback_object_lock(m, &next)) { 1149 VM_OBJECT_UNLOCK(object); 1150 vm_page_unlock(m); 1151 m = next; 1152 continue; 1153 } 1154 1155 /* 1156 * Don't deactivate pages that are busy. 1157 */ 1158 if ((m->busy != 0) || 1159 (m->oflags & VPO_BUSY) || 1160 (m->hold_count != 0)) { 1161 vm_page_unlock(m); 1162 VM_OBJECT_UNLOCK(object); 1163 vm_page_requeue(m); 1164 m = next; 1165 continue; 1166 } 1167 1168 /* 1169 * The count for pagedaemon pages is done after checking the 1170 * page for eligibility... 1171 */ 1172 cnt.v_pdpages++; 1173 1174 /* 1175 * Check to see "how much" the page has been used. 1176 */ 1177 actcount = 0; 1178 if (object->ref_count != 0) { 1179 if (m->aflags & PGA_REFERENCED) { 1180 actcount += 1; 1181 } 1182 actcount += pmap_ts_referenced(m); 1183 if (actcount) { 1184 m->act_count += ACT_ADVANCE + actcount; 1185 if (m->act_count > ACT_MAX) 1186 m->act_count = ACT_MAX; 1187 } 1188 } 1189 1190 /* 1191 * Since we have "tested" this bit, we need to clear it now. 1192 */ 1193 vm_page_aflag_clear(m, PGA_REFERENCED); 1194 1195 /* 1196 * Only if an object is currently being used, do we use the 1197 * page activation count stats. 1198 */ 1199 if (actcount && (object->ref_count != 0)) { 1200 vm_page_requeue(m); 1201 } else { 1202 m->act_count -= min(m->act_count, ACT_DECLINE); 1203 if (vm_pageout_algorithm || 1204 object->ref_count == 0 || 1205 m->act_count == 0) { 1206 page_shortage--; 1207 if (object->ref_count == 0) { 1208 KASSERT(!pmap_page_is_mapped(m), 1209 ("vm_pageout_scan: page %p is mapped", m)); 1210 if (m->dirty == 0) 1211 vm_page_cache(m); 1212 else 1213 vm_page_deactivate(m); 1214 } else { 1215 vm_page_deactivate(m); 1216 } 1217 } else { 1218 vm_page_requeue(m); 1219 } 1220 } 1221 vm_page_unlock(m); 1222 VM_OBJECT_UNLOCK(object); 1223 m = next; 1224 } 1225 vm_page_unlock_queues(); 1226 #if !defined(NO_SWAPPING) 1227 /* 1228 * Idle process swapout -- run once per second. 1229 */ 1230 if (vm_swap_idle_enabled) { 1231 static long lsec; 1232 if (time_second != lsec) { 1233 vm_req_vmdaemon(VM_SWAP_IDLE); 1234 lsec = time_second; 1235 } 1236 } 1237 #endif 1238 1239 /* 1240 * If we didn't get enough free pages, and we have skipped a vnode 1241 * in a writeable object, wakeup the sync daemon. And kick swapout 1242 * if we did not get enough free pages. 1243 */ 1244 if (vm_paging_target() > 0) { 1245 if (vnodes_skipped && vm_page_count_min()) 1246 (void) speedup_syncer(); 1247 #if !defined(NO_SWAPPING) 1248 if (vm_swap_enabled && vm_page_count_target()) 1249 vm_req_vmdaemon(VM_SWAP_NORMAL); 1250 #endif 1251 } 1252 1253 /* 1254 * If we are critically low on one of RAM or swap and low on 1255 * the other, kill the largest process. However, we avoid 1256 * doing this on the first pass in order to give ourselves a 1257 * chance to flush out dirty vnode-backed pages and to allow 1258 * active pages to be moved to the inactive queue and reclaimed. 1259 */ 1260 if (pass != 0 && 1261 ((swap_pager_avail < 64 && vm_page_count_min()) || 1262 (swap_pager_full && vm_paging_target() > 0))) 1263 vm_pageout_oom(VM_OOM_MEM); 1264 } 1265 1266 1267 void 1268 vm_pageout_oom(int shortage) 1269 { 1270 struct proc *p, *bigproc; 1271 vm_offset_t size, bigsize; 1272 struct thread *td; 1273 struct vmspace *vm; 1274 1275 /* 1276 * We keep the process bigproc locked once we find it to keep anyone 1277 * from messing with it; however, there is a possibility of 1278 * deadlock if process B is bigproc and one of it's child processes 1279 * attempts to propagate a signal to B while we are waiting for A's 1280 * lock while walking this list. To avoid this, we don't block on 1281 * the process lock but just skip a process if it is already locked. 1282 */ 1283 bigproc = NULL; 1284 bigsize = 0; 1285 sx_slock(&allproc_lock); 1286 FOREACH_PROC_IN_SYSTEM(p) { 1287 int breakout; 1288 1289 if (PROC_TRYLOCK(p) == 0) 1290 continue; 1291 /* 1292 * If this is a system, protected or killed process, skip it. 1293 */ 1294 if (p->p_state != PRS_NORMAL || 1295 (p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM)) || 1296 (p->p_pid == 1) || P_KILLED(p) || 1297 ((p->p_pid < 48) && (swap_pager_avail != 0))) { 1298 PROC_UNLOCK(p); 1299 continue; 1300 } 1301 /* 1302 * If the process is in a non-running type state, 1303 * don't touch it. Check all the threads individually. 1304 */ 1305 breakout = 0; 1306 FOREACH_THREAD_IN_PROC(p, td) { 1307 thread_lock(td); 1308 if (!TD_ON_RUNQ(td) && 1309 !TD_IS_RUNNING(td) && 1310 !TD_IS_SLEEPING(td) && 1311 !TD_IS_SUSPENDED(td)) { 1312 thread_unlock(td); 1313 breakout = 1; 1314 break; 1315 } 1316 thread_unlock(td); 1317 } 1318 if (breakout) { 1319 PROC_UNLOCK(p); 1320 continue; 1321 } 1322 /* 1323 * get the process size 1324 */ 1325 vm = vmspace_acquire_ref(p); 1326 if (vm == NULL) { 1327 PROC_UNLOCK(p); 1328 continue; 1329 } 1330 if (!vm_map_trylock_read(&vm->vm_map)) { 1331 vmspace_free(vm); 1332 PROC_UNLOCK(p); 1333 continue; 1334 } 1335 size = vmspace_swap_count(vm); 1336 vm_map_unlock_read(&vm->vm_map); 1337 if (shortage == VM_OOM_MEM) 1338 size += vmspace_resident_count(vm); 1339 vmspace_free(vm); 1340 /* 1341 * if the this process is bigger than the biggest one 1342 * remember it. 1343 */ 1344 if (size > bigsize) { 1345 if (bigproc != NULL) 1346 PROC_UNLOCK(bigproc); 1347 bigproc = p; 1348 bigsize = size; 1349 } else 1350 PROC_UNLOCK(p); 1351 } 1352 sx_sunlock(&allproc_lock); 1353 if (bigproc != NULL) { 1354 killproc(bigproc, "out of swap space"); 1355 sched_nice(bigproc, PRIO_MIN); 1356 PROC_UNLOCK(bigproc); 1357 wakeup(&cnt.v_free_count); 1358 } 1359 } 1360 1361 /* 1362 * This routine tries to maintain the pseudo LRU active queue, 1363 * so that during long periods of time where there is no paging, 1364 * that some statistic accumulation still occurs. This code 1365 * helps the situation where paging just starts to occur. 1366 */ 1367 static void 1368 vm_pageout_page_stats() 1369 { 1370 vm_object_t object; 1371 vm_page_t m,next; 1372 int pcount,tpcount; /* Number of pages to check */ 1373 static int fullintervalcount = 0; 1374 int page_shortage; 1375 1376 page_shortage = 1377 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) - 1378 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count); 1379 1380 if (page_shortage <= 0) 1381 return; 1382 1383 vm_page_lock_queues(); 1384 pcount = cnt.v_active_count; 1385 fullintervalcount += vm_pageout_stats_interval; 1386 if (fullintervalcount < vm_pageout_full_stats_interval) { 1387 tpcount = (int64_t)vm_pageout_stats_max * cnt.v_active_count / 1388 cnt.v_page_count; 1389 if (pcount > tpcount) 1390 pcount = tpcount; 1391 } else { 1392 fullintervalcount = 0; 1393 } 1394 1395 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1396 while ((m != NULL) && (pcount-- > 0)) { 1397 int actcount; 1398 1399 KASSERT(m->queue == PQ_ACTIVE, 1400 ("vm_pageout_page_stats: page %p isn't active", m)); 1401 1402 next = TAILQ_NEXT(m, pageq); 1403 if ((m->flags & PG_MARKER) != 0) { 1404 m = next; 1405 continue; 1406 } 1407 vm_page_lock_assert(m, MA_NOTOWNED); 1408 if (!vm_pageout_page_lock(m, &next)) { 1409 vm_page_unlock(m); 1410 m = next; 1411 continue; 1412 } 1413 object = m->object; 1414 if (!VM_OBJECT_TRYLOCK(object) && 1415 !vm_pageout_fallback_object_lock(m, &next)) { 1416 VM_OBJECT_UNLOCK(object); 1417 vm_page_unlock(m); 1418 m = next; 1419 continue; 1420 } 1421 1422 /* 1423 * Don't deactivate pages that are busy. 1424 */ 1425 if ((m->busy != 0) || 1426 (m->oflags & VPO_BUSY) || 1427 (m->hold_count != 0)) { 1428 vm_page_unlock(m); 1429 VM_OBJECT_UNLOCK(object); 1430 vm_page_requeue(m); 1431 m = next; 1432 continue; 1433 } 1434 1435 actcount = 0; 1436 if (m->aflags & PGA_REFERENCED) { 1437 vm_page_aflag_clear(m, PGA_REFERENCED); 1438 actcount += 1; 1439 } 1440 1441 actcount += pmap_ts_referenced(m); 1442 if (actcount) { 1443 m->act_count += ACT_ADVANCE + actcount; 1444 if (m->act_count > ACT_MAX) 1445 m->act_count = ACT_MAX; 1446 vm_page_requeue(m); 1447 } else { 1448 if (m->act_count == 0) { 1449 /* 1450 * We turn off page access, so that we have 1451 * more accurate RSS stats. We don't do this 1452 * in the normal page deactivation when the 1453 * system is loaded VM wise, because the 1454 * cost of the large number of page protect 1455 * operations would be higher than the value 1456 * of doing the operation. 1457 */ 1458 pmap_remove_all(m); 1459 vm_page_deactivate(m); 1460 } else { 1461 m->act_count -= min(m->act_count, ACT_DECLINE); 1462 vm_page_requeue(m); 1463 } 1464 } 1465 vm_page_unlock(m); 1466 VM_OBJECT_UNLOCK(object); 1467 m = next; 1468 } 1469 vm_page_unlock_queues(); 1470 } 1471 1472 /* 1473 * vm_pageout is the high level pageout daemon. 1474 */ 1475 static void 1476 vm_pageout() 1477 { 1478 int error, pass; 1479 1480 /* 1481 * Initialize some paging parameters. 1482 */ 1483 cnt.v_interrupt_free_min = 2; 1484 if (cnt.v_page_count < 2000) 1485 vm_pageout_page_count = 8; 1486 1487 /* 1488 * v_free_reserved needs to include enough for the largest 1489 * swap pager structures plus enough for any pv_entry structs 1490 * when paging. 1491 */ 1492 if (cnt.v_page_count > 1024) 1493 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200; 1494 else 1495 cnt.v_free_min = 4; 1496 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1497 cnt.v_interrupt_free_min; 1498 cnt.v_free_reserved = vm_pageout_page_count + 1499 cnt.v_pageout_free_min + (cnt.v_page_count / 768); 1500 cnt.v_free_severe = cnt.v_free_min / 2; 1501 cnt.v_free_min += cnt.v_free_reserved; 1502 cnt.v_free_severe += cnt.v_free_reserved; 1503 1504 /* 1505 * v_free_target and v_cache_min control pageout hysteresis. Note 1506 * that these are more a measure of the VM cache queue hysteresis 1507 * then the VM free queue. Specifically, v_free_target is the 1508 * high water mark (free+cache pages). 1509 * 1510 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the 1511 * low water mark, while v_free_min is the stop. v_cache_min must 1512 * be big enough to handle memory needs while the pageout daemon 1513 * is signalled and run to free more pages. 1514 */ 1515 if (cnt.v_free_count > 6144) 1516 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved; 1517 else 1518 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved; 1519 1520 if (cnt.v_free_count > 2048) { 1521 cnt.v_cache_min = cnt.v_free_target; 1522 cnt.v_cache_max = 2 * cnt.v_cache_min; 1523 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2; 1524 } else { 1525 cnt.v_cache_min = 0; 1526 cnt.v_cache_max = 0; 1527 cnt.v_inactive_target = cnt.v_free_count / 4; 1528 } 1529 if (cnt.v_inactive_target > cnt.v_free_count / 3) 1530 cnt.v_inactive_target = cnt.v_free_count / 3; 1531 1532 /* XXX does not really belong here */ 1533 if (vm_page_max_wired == 0) 1534 vm_page_max_wired = cnt.v_free_count / 3; 1535 1536 if (vm_pageout_stats_max == 0) 1537 vm_pageout_stats_max = cnt.v_free_target; 1538 1539 /* 1540 * Set interval in seconds for stats scan. 1541 */ 1542 if (vm_pageout_stats_interval == 0) 1543 vm_pageout_stats_interval = 5; 1544 if (vm_pageout_full_stats_interval == 0) 1545 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4; 1546 1547 swap_pager_swap_init(); 1548 pass = 0; 1549 /* 1550 * The pageout daemon is never done, so loop forever. 1551 */ 1552 while (TRUE) { 1553 /* 1554 * If we have enough free memory, wakeup waiters. Do 1555 * not clear vm_pages_needed until we reach our target, 1556 * otherwise we may be woken up over and over again and 1557 * waste a lot of cpu. 1558 */ 1559 mtx_lock(&vm_page_queue_free_mtx); 1560 if (vm_pages_needed && !vm_page_count_min()) { 1561 if (!vm_paging_needed()) 1562 vm_pages_needed = 0; 1563 wakeup(&cnt.v_free_count); 1564 } 1565 if (vm_pages_needed) { 1566 /* 1567 * Still not done, take a second pass without waiting 1568 * (unlimited dirty cleaning), otherwise sleep a bit 1569 * and try again. 1570 */ 1571 ++pass; 1572 if (pass > 1) 1573 msleep(&vm_pages_needed, 1574 &vm_page_queue_free_mtx, PVM, "psleep", 1575 hz / 2); 1576 } else { 1577 /* 1578 * Good enough, sleep & handle stats. Prime the pass 1579 * for the next run. 1580 */ 1581 if (pass > 1) 1582 pass = 1; 1583 else 1584 pass = 0; 1585 error = msleep(&vm_pages_needed, 1586 &vm_page_queue_free_mtx, PVM, "psleep", 1587 vm_pageout_stats_interval * hz); 1588 if (error && !vm_pages_needed) { 1589 mtx_unlock(&vm_page_queue_free_mtx); 1590 pass = 0; 1591 vm_pageout_page_stats(); 1592 continue; 1593 } 1594 } 1595 if (vm_pages_needed) 1596 cnt.v_pdwakeups++; 1597 mtx_unlock(&vm_page_queue_free_mtx); 1598 vm_pageout_scan(pass); 1599 } 1600 } 1601 1602 /* 1603 * Unless the free page queue lock is held by the caller, this function 1604 * should be regarded as advisory. Specifically, the caller should 1605 * not msleep() on &cnt.v_free_count following this function unless 1606 * the free page queue lock is held until the msleep() is performed. 1607 */ 1608 void 1609 pagedaemon_wakeup() 1610 { 1611 1612 if (!vm_pages_needed && curthread->td_proc != pageproc) { 1613 vm_pages_needed = 1; 1614 wakeup(&vm_pages_needed); 1615 } 1616 } 1617 1618 #if !defined(NO_SWAPPING) 1619 static void 1620 vm_req_vmdaemon(int req) 1621 { 1622 static int lastrun = 0; 1623 1624 mtx_lock(&vm_daemon_mtx); 1625 vm_pageout_req_swapout |= req; 1626 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1627 wakeup(&vm_daemon_needed); 1628 lastrun = ticks; 1629 } 1630 mtx_unlock(&vm_daemon_mtx); 1631 } 1632 1633 static void 1634 vm_daemon() 1635 { 1636 struct rlimit rsslim; 1637 struct proc *p; 1638 struct thread *td; 1639 struct vmspace *vm; 1640 int breakout, swapout_flags, tryagain, attempts; 1641 #ifdef RACCT 1642 uint64_t rsize, ravailable; 1643 #endif 1644 1645 while (TRUE) { 1646 mtx_lock(&vm_daemon_mtx); 1647 #ifdef RACCT 1648 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", hz); 1649 #else 1650 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0); 1651 #endif 1652 swapout_flags = vm_pageout_req_swapout; 1653 vm_pageout_req_swapout = 0; 1654 mtx_unlock(&vm_daemon_mtx); 1655 if (swapout_flags) 1656 swapout_procs(swapout_flags); 1657 1658 /* 1659 * scan the processes for exceeding their rlimits or if 1660 * process is swapped out -- deactivate pages 1661 */ 1662 tryagain = 0; 1663 attempts = 0; 1664 again: 1665 attempts++; 1666 sx_slock(&allproc_lock); 1667 FOREACH_PROC_IN_SYSTEM(p) { 1668 vm_pindex_t limit, size; 1669 1670 /* 1671 * if this is a system process or if we have already 1672 * looked at this process, skip it. 1673 */ 1674 PROC_LOCK(p); 1675 if (p->p_state != PRS_NORMAL || 1676 p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) { 1677 PROC_UNLOCK(p); 1678 continue; 1679 } 1680 /* 1681 * if the process is in a non-running type state, 1682 * don't touch it. 1683 */ 1684 breakout = 0; 1685 FOREACH_THREAD_IN_PROC(p, td) { 1686 thread_lock(td); 1687 if (!TD_ON_RUNQ(td) && 1688 !TD_IS_RUNNING(td) && 1689 !TD_IS_SLEEPING(td) && 1690 !TD_IS_SUSPENDED(td)) { 1691 thread_unlock(td); 1692 breakout = 1; 1693 break; 1694 } 1695 thread_unlock(td); 1696 } 1697 if (breakout) { 1698 PROC_UNLOCK(p); 1699 continue; 1700 } 1701 /* 1702 * get a limit 1703 */ 1704 lim_rlimit(p, RLIMIT_RSS, &rsslim); 1705 limit = OFF_TO_IDX( 1706 qmin(rsslim.rlim_cur, rsslim.rlim_max)); 1707 1708 /* 1709 * let processes that are swapped out really be 1710 * swapped out set the limit to nothing (will force a 1711 * swap-out.) 1712 */ 1713 if ((p->p_flag & P_INMEM) == 0) 1714 limit = 0; /* XXX */ 1715 vm = vmspace_acquire_ref(p); 1716 PROC_UNLOCK(p); 1717 if (vm == NULL) 1718 continue; 1719 1720 size = vmspace_resident_count(vm); 1721 if (limit >= 0 && size >= limit) { 1722 vm_pageout_map_deactivate_pages( 1723 &vm->vm_map, limit); 1724 } 1725 #ifdef RACCT 1726 rsize = IDX_TO_OFF(size); 1727 PROC_LOCK(p); 1728 racct_set(p, RACCT_RSS, rsize); 1729 ravailable = racct_get_available(p, RACCT_RSS); 1730 PROC_UNLOCK(p); 1731 if (rsize > ravailable) { 1732 /* 1733 * Don't be overly aggressive; this might be 1734 * an innocent process, and the limit could've 1735 * been exceeded by some memory hog. Don't 1736 * try to deactivate more than 1/4th of process' 1737 * resident set size. 1738 */ 1739 if (attempts <= 8) { 1740 if (ravailable < rsize - (rsize / 4)) 1741 ravailable = rsize - (rsize / 4); 1742 } 1743 vm_pageout_map_deactivate_pages( 1744 &vm->vm_map, OFF_TO_IDX(ravailable)); 1745 /* Update RSS usage after paging out. */ 1746 size = vmspace_resident_count(vm); 1747 rsize = IDX_TO_OFF(size); 1748 PROC_LOCK(p); 1749 racct_set(p, RACCT_RSS, rsize); 1750 PROC_UNLOCK(p); 1751 if (rsize > ravailable) 1752 tryagain = 1; 1753 } 1754 #endif 1755 vmspace_free(vm); 1756 } 1757 sx_sunlock(&allproc_lock); 1758 if (tryagain != 0 && attempts <= 10) 1759 goto again; 1760 } 1761 } 1762 #endif /* !defined(NO_SWAPPING) */ 1763