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/resourcevar.h> 90 #include <sys/sched.h> 91 #include <sys/signalvar.h> 92 #include <sys/vnode.h> 93 #include <sys/vmmeter.h> 94 #include <sys/sx.h> 95 #include <sys/sysctl.h> 96 97 #include <vm/vm.h> 98 #include <vm/vm_param.h> 99 #include <vm/vm_object.h> 100 #include <vm/vm_page.h> 101 #include <vm/vm_map.h> 102 #include <vm/vm_pageout.h> 103 #include <vm/vm_pager.h> 104 #include <vm/swap_pager.h> 105 #include <vm/vm_extern.h> 106 #include <vm/uma.h> 107 108 /* 109 * System initialization 110 */ 111 112 /* the kernel process "vm_pageout"*/ 113 static void vm_pageout(void); 114 static int vm_pageout_clean(vm_page_t); 115 static void vm_pageout_scan(int pass); 116 117 struct proc *pageproc; 118 119 static struct kproc_desc page_kp = { 120 "pagedaemon", 121 vm_pageout, 122 &pageproc 123 }; 124 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, 125 &page_kp); 126 127 #if !defined(NO_SWAPPING) 128 /* the kernel process "vm_daemon"*/ 129 static void vm_daemon(void); 130 static struct proc *vmproc; 131 132 static struct kproc_desc vm_kp = { 133 "vmdaemon", 134 vm_daemon, 135 &vmproc 136 }; 137 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp); 138 #endif 139 140 141 int vm_pages_needed; /* Event on which pageout daemon sleeps */ 142 int vm_pageout_deficit; /* Estimated number of pages deficit */ 143 int vm_pageout_pages_needed; /* flag saying that the pageout daemon needs pages */ 144 145 #if !defined(NO_SWAPPING) 146 static int vm_pageout_req_swapout; /* XXX */ 147 static int vm_daemon_needed; 148 static struct mtx vm_daemon_mtx; 149 /* Allow for use by vm_pageout before vm_daemon is initialized. */ 150 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF); 151 #endif 152 static int vm_max_launder = 32; 153 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0; 154 static int vm_pageout_full_stats_interval = 0; 155 static int vm_pageout_algorithm=0; 156 static int defer_swap_pageouts=0; 157 static int disable_swap_pageouts=0; 158 159 #if defined(NO_SWAPPING) 160 static int vm_swap_enabled=0; 161 static int vm_swap_idle_enabled=0; 162 #else 163 static int vm_swap_enabled=1; 164 static int vm_swap_idle_enabled=0; 165 #endif 166 167 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm, 168 CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt"); 169 170 SYSCTL_INT(_vm, OID_AUTO, max_launder, 171 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout"); 172 173 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max, 174 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length"); 175 176 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval, 177 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan"); 178 179 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval, 180 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan"); 181 182 #if defined(NO_SWAPPING) 183 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 184 CTLFLAG_RD, &vm_swap_enabled, 0, "Enable entire process swapout"); 185 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 186 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 187 #else 188 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 189 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout"); 190 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 191 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 192 #endif 193 194 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts, 195 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem"); 196 197 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts, 198 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages"); 199 200 static int pageout_lock_miss; 201 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss, 202 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout"); 203 204 #define VM_PAGEOUT_PAGE_COUNT 16 205 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT; 206 207 int vm_page_max_wired; /* XXX max # of wired pages system-wide */ 208 SYSCTL_INT(_vm, OID_AUTO, max_wired, 209 CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count"); 210 211 #if !defined(NO_SWAPPING) 212 static void vm_pageout_map_deactivate_pages(vm_map_t, long); 213 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long); 214 static void vm_req_vmdaemon(int req); 215 #endif 216 static void vm_pageout_page_stats(void); 217 218 /* 219 * vm_pageout_fallback_object_lock: 220 * 221 * Lock vm object currently associated with `m'. VM_OBJECT_TRYLOCK is 222 * known to have failed and page queue must be either PQ_ACTIVE or 223 * PQ_INACTIVE. To avoid lock order violation, unlock the page queues 224 * while locking the vm object. Use marker page to detect page queue 225 * changes and maintain notion of next page on page queue. Return 226 * TRUE if no changes were detected, FALSE otherwise. vm object is 227 * locked on return. 228 * 229 * This function depends on both the lock portion of struct vm_object 230 * and normal struct vm_page being type stable. 231 */ 232 boolean_t 233 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next) 234 { 235 struct vm_page marker; 236 boolean_t unchanged; 237 u_short queue; 238 vm_object_t object; 239 240 /* 241 * Initialize our marker 242 */ 243 bzero(&marker, sizeof(marker)); 244 marker.flags = PG_FICTITIOUS | PG_MARKER; 245 marker.oflags = VPO_BUSY; 246 marker.queue = m->queue; 247 marker.wire_count = 1; 248 249 queue = m->queue; 250 object = m->object; 251 252 TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl, 253 m, &marker, pageq); 254 vm_page_unlock_queues(); 255 VM_OBJECT_LOCK(object); 256 vm_page_lock_queues(); 257 258 /* Page queue might have changed. */ 259 *next = TAILQ_NEXT(&marker, pageq); 260 unchanged = (m->queue == queue && 261 m->object == object && 262 &marker == TAILQ_NEXT(m, pageq)); 263 TAILQ_REMOVE(&vm_page_queues[queue].pl, 264 &marker, pageq); 265 return (unchanged); 266 } 267 268 /* 269 * vm_pageout_clean: 270 * 271 * Clean the page and remove it from the laundry. 272 * 273 * We set the busy bit to cause potential page faults on this page to 274 * block. Note the careful timing, however, the busy bit isn't set till 275 * late and we cannot do anything that will mess with the page. 276 */ 277 static int 278 vm_pageout_clean(m) 279 vm_page_t m; 280 { 281 vm_object_t object; 282 vm_page_t mc[2*vm_pageout_page_count]; 283 int pageout_count; 284 int ib, is, page_base; 285 vm_pindex_t pindex = m->pindex; 286 287 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 288 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 289 290 /* 291 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP 292 * with the new swapper, but we could have serious problems paging 293 * out other object types if there is insufficient memory. 294 * 295 * Unfortunately, checking free memory here is far too late, so the 296 * check has been moved up a procedural level. 297 */ 298 299 /* 300 * Can't clean the page if it's busy or held. 301 */ 302 if ((m->hold_count != 0) || 303 ((m->busy != 0) || (m->oflags & VPO_BUSY))) { 304 return 0; 305 } 306 307 mc[vm_pageout_page_count] = m; 308 pageout_count = 1; 309 page_base = vm_pageout_page_count; 310 ib = 1; 311 is = 1; 312 313 /* 314 * Scan object for clusterable pages. 315 * 316 * We can cluster ONLY if: ->> the page is NOT 317 * clean, wired, busy, held, or mapped into a 318 * buffer, and one of the following: 319 * 1) The page is inactive, or a seldom used 320 * active page. 321 * -or- 322 * 2) we force the issue. 323 * 324 * During heavy mmap/modification loads the pageout 325 * daemon can really fragment the underlying file 326 * due to flushing pages out of order and not trying 327 * align the clusters (which leave sporatic out-of-order 328 * holes). To solve this problem we do the reverse scan 329 * first and attempt to align our cluster, then do a 330 * forward scan if room remains. 331 */ 332 object = m->object; 333 more: 334 while (ib && pageout_count < vm_pageout_page_count) { 335 vm_page_t p; 336 337 if (ib > pindex) { 338 ib = 0; 339 break; 340 } 341 342 if ((p = vm_page_lookup(object, pindex - ib)) == NULL) { 343 ib = 0; 344 break; 345 } 346 if ((p->oflags & VPO_BUSY) || p->busy) { 347 ib = 0; 348 break; 349 } 350 vm_page_test_dirty(p); 351 if (p->dirty == 0 || 352 p->queue != PQ_INACTIVE || 353 p->wire_count != 0 || /* may be held by buf cache */ 354 p->hold_count != 0) { /* may be undergoing I/O */ 355 ib = 0; 356 break; 357 } 358 mc[--page_base] = p; 359 ++pageout_count; 360 ++ib; 361 /* 362 * alignment boundry, stop here and switch directions. Do 363 * not clear ib. 364 */ 365 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0) 366 break; 367 } 368 369 while (pageout_count < vm_pageout_page_count && 370 pindex + is < object->size) { 371 vm_page_t p; 372 373 if ((p = vm_page_lookup(object, pindex + is)) == NULL) 374 break; 375 if ((p->oflags & VPO_BUSY) || p->busy) { 376 break; 377 } 378 vm_page_test_dirty(p); 379 if (p->dirty == 0 || 380 p->queue != PQ_INACTIVE || 381 p->wire_count != 0 || /* may be held by buf cache */ 382 p->hold_count != 0) { /* may be undergoing I/O */ 383 break; 384 } 385 mc[page_base + pageout_count] = p; 386 ++pageout_count; 387 ++is; 388 } 389 390 /* 391 * If we exhausted our forward scan, continue with the reverse scan 392 * when possible, even past a page boundry. This catches boundry 393 * conditions. 394 */ 395 if (ib && pageout_count < vm_pageout_page_count) 396 goto more; 397 398 /* 399 * we allow reads during pageouts... 400 */ 401 return (vm_pageout_flush(&mc[page_base], pageout_count, 0)); 402 } 403 404 /* 405 * vm_pageout_flush() - launder the given pages 406 * 407 * The given pages are laundered. Note that we setup for the start of 408 * I/O ( i.e. busy the page ), mark it read-only, and bump the object 409 * reference count all in here rather then in the parent. If we want 410 * the parent to do more sophisticated things we may have to change 411 * the ordering. 412 */ 413 int 414 vm_pageout_flush(vm_page_t *mc, int count, int flags) 415 { 416 vm_object_t object = mc[0]->object; 417 int pageout_status[count]; 418 int numpagedout = 0; 419 int i; 420 421 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 422 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 423 /* 424 * Initiate I/O. Bump the vm_page_t->busy counter and 425 * mark the pages read-only. 426 * 427 * We do not have to fixup the clean/dirty bits here... we can 428 * allow the pager to do it after the I/O completes. 429 * 430 * NOTE! mc[i]->dirty may be partial or fragmented due to an 431 * edge case with file fragments. 432 */ 433 for (i = 0; i < count; i++) { 434 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, 435 ("vm_pageout_flush: partially invalid page %p index %d/%d", 436 mc[i], i, count)); 437 vm_page_io_start(mc[i]); 438 pmap_remove_write(mc[i]); 439 } 440 vm_page_unlock_queues(); 441 vm_object_pip_add(object, count); 442 443 vm_pager_put_pages(object, mc, count, flags, pageout_status); 444 445 vm_page_lock_queues(); 446 for (i = 0; i < count; i++) { 447 vm_page_t mt = mc[i]; 448 449 KASSERT(pageout_status[i] == VM_PAGER_PEND || 450 (mt->flags & PG_WRITEABLE) == 0, 451 ("vm_pageout_flush: page %p is not write protected", mt)); 452 switch (pageout_status[i]) { 453 case VM_PAGER_OK: 454 case VM_PAGER_PEND: 455 numpagedout++; 456 break; 457 case VM_PAGER_BAD: 458 /* 459 * Page outside of range of object. Right now we 460 * essentially lose the changes by pretending it 461 * worked. 462 */ 463 vm_page_undirty(mt); 464 break; 465 case VM_PAGER_ERROR: 466 case VM_PAGER_FAIL: 467 /* 468 * If page couldn't be paged out, then reactivate the 469 * page so it doesn't clog the inactive list. (We 470 * will try paging out it again later). 471 */ 472 vm_page_activate(mt); 473 break; 474 case VM_PAGER_AGAIN: 475 break; 476 } 477 478 /* 479 * If the operation is still going, leave the page busy to 480 * block all other accesses. Also, leave the paging in 481 * progress indicator set so that we don't attempt an object 482 * collapse. 483 */ 484 if (pageout_status[i] != VM_PAGER_PEND) { 485 vm_object_pip_wakeup(object); 486 vm_page_io_finish(mt); 487 if (vm_page_count_severe()) 488 vm_page_try_to_cache(mt); 489 } 490 } 491 return numpagedout; 492 } 493 494 #if !defined(NO_SWAPPING) 495 /* 496 * vm_pageout_object_deactivate_pages 497 * 498 * deactivate enough pages to satisfy the inactive target 499 * requirements or if vm_page_proc_limit is set, then 500 * deactivate all of the pages in the object and its 501 * backing_objects. 502 * 503 * The object and map must be locked. 504 */ 505 static void 506 vm_pageout_object_deactivate_pages(pmap, first_object, desired) 507 pmap_t pmap; 508 vm_object_t first_object; 509 long desired; 510 { 511 vm_object_t backing_object, object; 512 vm_page_t p, next; 513 int actcount, rcount, remove_mode; 514 515 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED); 516 if (first_object->type == OBJT_DEVICE || first_object->type == OBJT_PHYS) 517 return; 518 for (object = first_object;; object = backing_object) { 519 if (pmap_resident_count(pmap) <= desired) 520 goto unlock_return; 521 if (object->paging_in_progress) 522 goto unlock_return; 523 524 remove_mode = 0; 525 if (object->shadow_count > 1) 526 remove_mode = 1; 527 /* 528 * scan the objects entire memory queue 529 */ 530 rcount = object->resident_page_count; 531 p = TAILQ_FIRST(&object->memq); 532 vm_page_lock_queues(); 533 while (p && (rcount-- > 0)) { 534 if (pmap_resident_count(pmap) <= desired) { 535 vm_page_unlock_queues(); 536 goto unlock_return; 537 } 538 next = TAILQ_NEXT(p, listq); 539 cnt.v_pdpages++; 540 if (p->wire_count != 0 || 541 p->hold_count != 0 || 542 p->busy != 0 || 543 (p->oflags & VPO_BUSY) || 544 (p->flags & PG_UNMANAGED) || 545 !pmap_page_exists_quick(pmap, p)) { 546 p = next; 547 continue; 548 } 549 actcount = pmap_ts_referenced(p); 550 if (actcount) { 551 vm_page_flag_set(p, PG_REFERENCED); 552 } else if (p->flags & PG_REFERENCED) { 553 actcount = 1; 554 } 555 if ((p->queue != PQ_ACTIVE) && 556 (p->flags & PG_REFERENCED)) { 557 vm_page_activate(p); 558 p->act_count += actcount; 559 vm_page_flag_clear(p, PG_REFERENCED); 560 } else if (p->queue == PQ_ACTIVE) { 561 if ((p->flags & PG_REFERENCED) == 0) { 562 p->act_count -= min(p->act_count, ACT_DECLINE); 563 if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) { 564 pmap_remove_all(p); 565 vm_page_deactivate(p); 566 } else { 567 vm_page_requeue(p); 568 } 569 } else { 570 vm_page_activate(p); 571 vm_page_flag_clear(p, PG_REFERENCED); 572 if (p->act_count < (ACT_MAX - ACT_ADVANCE)) 573 p->act_count += ACT_ADVANCE; 574 vm_page_requeue(p); 575 } 576 } else if (p->queue == PQ_INACTIVE) { 577 pmap_remove_all(p); 578 } 579 p = next; 580 } 581 vm_page_unlock_queues(); 582 if ((backing_object = object->backing_object) == NULL) 583 goto unlock_return; 584 VM_OBJECT_LOCK(backing_object); 585 if (object != first_object) 586 VM_OBJECT_UNLOCK(object); 587 } 588 unlock_return: 589 if (object != first_object) 590 VM_OBJECT_UNLOCK(object); 591 } 592 593 /* 594 * deactivate some number of pages in a map, try to do it fairly, but 595 * that is really hard to do. 596 */ 597 static void 598 vm_pageout_map_deactivate_pages(map, desired) 599 vm_map_t map; 600 long desired; 601 { 602 vm_map_entry_t tmpe; 603 vm_object_t obj, bigobj; 604 int nothingwired; 605 606 if (!vm_map_trylock(map)) 607 return; 608 609 bigobj = NULL; 610 nothingwired = TRUE; 611 612 /* 613 * first, search out the biggest object, and try to free pages from 614 * that. 615 */ 616 tmpe = map->header.next; 617 while (tmpe != &map->header) { 618 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 619 obj = tmpe->object.vm_object; 620 if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) { 621 if (obj->shadow_count <= 1 && 622 (bigobj == NULL || 623 bigobj->resident_page_count < obj->resident_page_count)) { 624 if (bigobj != NULL) 625 VM_OBJECT_UNLOCK(bigobj); 626 bigobj = obj; 627 } else 628 VM_OBJECT_UNLOCK(obj); 629 } 630 } 631 if (tmpe->wired_count > 0) 632 nothingwired = FALSE; 633 tmpe = tmpe->next; 634 } 635 636 if (bigobj != NULL) { 637 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired); 638 VM_OBJECT_UNLOCK(bigobj); 639 } 640 /* 641 * Next, hunt around for other pages to deactivate. We actually 642 * do this search sort of wrong -- .text first is not the best idea. 643 */ 644 tmpe = map->header.next; 645 while (tmpe != &map->header) { 646 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 647 break; 648 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 649 obj = tmpe->object.vm_object; 650 if (obj != NULL) { 651 VM_OBJECT_LOCK(obj); 652 vm_pageout_object_deactivate_pages(map->pmap, obj, desired); 653 VM_OBJECT_UNLOCK(obj); 654 } 655 } 656 tmpe = tmpe->next; 657 } 658 659 /* 660 * Remove all mappings if a process is swapped out, this will free page 661 * table pages. 662 */ 663 if (desired == 0 && nothingwired) { 664 pmap_remove(vm_map_pmap(map), vm_map_min(map), 665 vm_map_max(map)); 666 } 667 vm_map_unlock(map); 668 } 669 #endif /* !defined(NO_SWAPPING) */ 670 671 /* 672 * vm_pageout_scan does the dirty work for the pageout daemon. 673 */ 674 static void 675 vm_pageout_scan(int pass) 676 { 677 vm_page_t m, next; 678 struct vm_page marker; 679 int page_shortage, maxscan, pcount; 680 int addl_page_shortage, addl_page_shortage_init; 681 vm_object_t object; 682 int actcount; 683 int vnodes_skipped = 0; 684 int maxlaunder; 685 686 /* 687 * Decrease registered cache sizes. 688 */ 689 EVENTHANDLER_INVOKE(vm_lowmem, 0); 690 /* 691 * We do this explicitly after the caches have been drained above. 692 */ 693 uma_reclaim(); 694 695 addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit); 696 697 /* 698 * Calculate the number of pages we want to either free or move 699 * to the cache. 700 */ 701 page_shortage = vm_paging_target() + addl_page_shortage_init; 702 703 /* 704 * Initialize our marker 705 */ 706 bzero(&marker, sizeof(marker)); 707 marker.flags = PG_FICTITIOUS | PG_MARKER; 708 marker.oflags = VPO_BUSY; 709 marker.queue = PQ_INACTIVE; 710 marker.wire_count = 1; 711 712 /* 713 * Start scanning the inactive queue for pages we can move to the 714 * cache or free. The scan will stop when the target is reached or 715 * we have scanned the entire inactive queue. Note that m->act_count 716 * is not used to form decisions for the inactive queue, only for the 717 * active queue. 718 * 719 * maxlaunder limits the number of dirty pages we flush per scan. 720 * For most systems a smaller value (16 or 32) is more robust under 721 * extreme memory and disk pressure because any unnecessary writes 722 * to disk can result in extreme performance degredation. However, 723 * systems with excessive dirty pages (especially when MAP_NOSYNC is 724 * used) will die horribly with limited laundering. If the pageout 725 * daemon cannot clean enough pages in the first pass, we let it go 726 * all out in succeeding passes. 727 */ 728 if ((maxlaunder = vm_max_launder) <= 1) 729 maxlaunder = 1; 730 if (pass) 731 maxlaunder = 10000; 732 vm_page_lock_queues(); 733 rescan0: 734 addl_page_shortage = addl_page_shortage_init; 735 maxscan = cnt.v_inactive_count; 736 737 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl); 738 m != NULL && maxscan-- > 0 && page_shortage > 0; 739 m = next) { 740 741 cnt.v_pdpages++; 742 743 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE) { 744 goto rescan0; 745 } 746 747 next = TAILQ_NEXT(m, pageq); 748 object = m->object; 749 750 /* 751 * skip marker pages 752 */ 753 if (m->flags & PG_MARKER) 754 continue; 755 756 /* 757 * A held page may be undergoing I/O, so skip it. 758 */ 759 if (m->hold_count) { 760 vm_page_requeue(m); 761 addl_page_shortage++; 762 continue; 763 } 764 /* 765 * Don't mess with busy pages, keep in the front of the 766 * queue, most likely are being paged out. 767 */ 768 if (!VM_OBJECT_TRYLOCK(object) && 769 (!vm_pageout_fallback_object_lock(m, &next) || 770 m->hold_count != 0)) { 771 VM_OBJECT_UNLOCK(object); 772 addl_page_shortage++; 773 continue; 774 } 775 if (m->busy || (m->oflags & VPO_BUSY)) { 776 VM_OBJECT_UNLOCK(object); 777 addl_page_shortage++; 778 continue; 779 } 780 781 /* 782 * If the object is not being used, we ignore previous 783 * references. 784 */ 785 if (object->ref_count == 0) { 786 vm_page_flag_clear(m, PG_REFERENCED); 787 KASSERT(!pmap_page_is_mapped(m), 788 ("vm_pageout_scan: page %p is mapped", m)); 789 790 /* 791 * Otherwise, if the page has been referenced while in the 792 * inactive queue, we bump the "activation count" upwards, 793 * making it less likely that the page will be added back to 794 * the inactive queue prematurely again. Here we check the 795 * page tables (or emulated bits, if any), given the upper 796 * level VM system not knowing anything about existing 797 * references. 798 */ 799 } else if (((m->flags & PG_REFERENCED) == 0) && 800 (actcount = pmap_ts_referenced(m))) { 801 vm_page_activate(m); 802 VM_OBJECT_UNLOCK(object); 803 m->act_count += (actcount + ACT_ADVANCE); 804 continue; 805 } 806 807 /* 808 * If the upper level VM system knows about any page 809 * references, we activate the page. We also set the 810 * "activation count" higher than normal so that we will less 811 * likely place pages back onto the inactive queue again. 812 */ 813 if ((m->flags & PG_REFERENCED) != 0) { 814 vm_page_flag_clear(m, PG_REFERENCED); 815 actcount = pmap_ts_referenced(m); 816 vm_page_activate(m); 817 VM_OBJECT_UNLOCK(object); 818 m->act_count += (actcount + ACT_ADVANCE + 1); 819 continue; 820 } 821 822 /* 823 * If the upper level VM system does not believe that the page 824 * is fully dirty, but it is mapped for write access, then we 825 * consult the pmap to see if the page's dirty status should 826 * be updated. 827 */ 828 if (m->dirty != VM_PAGE_BITS_ALL && 829 (m->flags & PG_WRITEABLE) != 0) { 830 /* 831 * Avoid a race condition: Unless write access is 832 * removed from the page, another processor could 833 * modify it before all access is removed by the call 834 * to vm_page_cache() below. If vm_page_cache() finds 835 * that the page has been modified when it removes all 836 * access, it panics because it cannot cache dirty 837 * pages. In principle, we could eliminate just write 838 * access here rather than all access. In the expected 839 * case, when there are no last instant modifications 840 * to the page, removing all access will be cheaper 841 * overall. 842 */ 843 if (pmap_is_modified(m)) 844 vm_page_dirty(m); 845 else if (m->dirty == 0) 846 pmap_remove_all(m); 847 } 848 849 if (m->valid == 0) { 850 /* 851 * Invalid pages can be easily freed 852 */ 853 vm_page_free(m); 854 cnt.v_dfree++; 855 --page_shortage; 856 } else if (m->dirty == 0) { 857 /* 858 * Clean pages can be placed onto the cache queue. 859 * This effectively frees them. 860 */ 861 vm_page_cache(m); 862 --page_shortage; 863 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) { 864 /* 865 * Dirty pages need to be paged out, but flushing 866 * a page is extremely expensive verses freeing 867 * a clean page. Rather then artificially limiting 868 * the number of pages we can flush, we instead give 869 * dirty pages extra priority on the inactive queue 870 * by forcing them to be cycled through the queue 871 * twice before being flushed, after which the 872 * (now clean) page will cycle through once more 873 * before being freed. This significantly extends 874 * the thrash point for a heavily loaded machine. 875 */ 876 vm_page_flag_set(m, PG_WINATCFLS); 877 vm_page_requeue(m); 878 } else if (maxlaunder > 0) { 879 /* 880 * We always want to try to flush some dirty pages if 881 * we encounter them, to keep the system stable. 882 * Normally this number is small, but under extreme 883 * pressure where there are insufficient clean pages 884 * on the inactive queue, we may have to go all out. 885 */ 886 int swap_pageouts_ok, vfslocked = 0; 887 struct vnode *vp = NULL; 888 struct mount *mp = NULL; 889 890 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 891 swap_pageouts_ok = 1; 892 } else { 893 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 894 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 895 vm_page_count_min()); 896 897 } 898 899 /* 900 * We don't bother paging objects that are "dead". 901 * Those objects are in a "rundown" state. 902 */ 903 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 904 VM_OBJECT_UNLOCK(object); 905 vm_page_requeue(m); 906 continue; 907 } 908 909 /* 910 * Following operations may unlock 911 * vm_page_queue_mtx, invalidating the 'next' 912 * pointer. To prevent an inordinate number 913 * of restarts we use our marker to remember 914 * our place. 915 * 916 */ 917 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, 918 m, &marker, pageq); 919 /* 920 * The object is already known NOT to be dead. It 921 * is possible for the vget() to block the whole 922 * pageout daemon, but the new low-memory handling 923 * code should prevent it. 924 * 925 * The previous code skipped locked vnodes and, worse, 926 * reordered pages in the queue. This results in 927 * completely non-deterministic operation and, on a 928 * busy system, can lead to extremely non-optimal 929 * pageouts. For example, it can cause clean pages 930 * to be freed and dirty pages to be moved to the end 931 * of the queue. Since dirty pages are also moved to 932 * the end of the queue once-cleaned, this gives 933 * way too large a weighting to defering the freeing 934 * of dirty pages. 935 * 936 * We can't wait forever for the vnode lock, we might 937 * deadlock due to a vn_read() getting stuck in 938 * vm_wait while holding this vnode. We skip the 939 * vnode if we can't get it in a reasonable amount 940 * of time. 941 */ 942 if (object->type == OBJT_VNODE) { 943 vp = object->handle; 944 if (vp->v_type == VREG && 945 vn_start_write(vp, &mp, V_NOWAIT) != 0) { 946 mp = NULL; 947 ++pageout_lock_miss; 948 if (object->flags & OBJ_MIGHTBEDIRTY) 949 vnodes_skipped++; 950 goto unlock_and_continue; 951 } 952 vm_page_unlock_queues(); 953 vm_object_reference_locked(object); 954 VM_OBJECT_UNLOCK(object); 955 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 956 if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK, 957 curthread)) { 958 VM_OBJECT_LOCK(object); 959 vm_page_lock_queues(); 960 ++pageout_lock_miss; 961 if (object->flags & OBJ_MIGHTBEDIRTY) 962 vnodes_skipped++; 963 vp = NULL; 964 goto unlock_and_continue; 965 } 966 VM_OBJECT_LOCK(object); 967 vm_page_lock_queues(); 968 /* 969 * The page might have been moved to another 970 * queue during potential blocking in vget() 971 * above. The page might have been freed and 972 * reused for another vnode. 973 */ 974 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE || 975 m->object != object || 976 TAILQ_NEXT(m, pageq) != &marker) { 977 if (object->flags & OBJ_MIGHTBEDIRTY) 978 vnodes_skipped++; 979 goto unlock_and_continue; 980 } 981 982 /* 983 * The page may have been busied during the 984 * blocking in vget(). We don't move the 985 * page back onto the end of the queue so that 986 * statistics are more correct if we don't. 987 */ 988 if (m->busy || (m->oflags & VPO_BUSY)) { 989 goto unlock_and_continue; 990 } 991 992 /* 993 * If the page has become held it might 994 * be undergoing I/O, so skip it 995 */ 996 if (m->hold_count) { 997 vm_page_requeue(m); 998 if (object->flags & OBJ_MIGHTBEDIRTY) 999 vnodes_skipped++; 1000 goto unlock_and_continue; 1001 } 1002 } 1003 1004 /* 1005 * If a page is dirty, then it is either being washed 1006 * (but not yet cleaned) or it is still in the 1007 * laundry. If it is still in the laundry, then we 1008 * start the cleaning operation. 1009 * 1010 * decrement page_shortage on success to account for 1011 * the (future) cleaned page. Otherwise we could wind 1012 * up laundering or cleaning too many pages. 1013 */ 1014 if (vm_pageout_clean(m) != 0) { 1015 --page_shortage; 1016 --maxlaunder; 1017 } 1018 unlock_and_continue: 1019 VM_OBJECT_UNLOCK(object); 1020 if (mp != NULL) { 1021 vm_page_unlock_queues(); 1022 if (vp != NULL) 1023 vput(vp); 1024 VFS_UNLOCK_GIANT(vfslocked); 1025 vm_object_deallocate(object); 1026 vn_finished_write(mp); 1027 vm_page_lock_queues(); 1028 } 1029 next = TAILQ_NEXT(&marker, pageq); 1030 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, 1031 &marker, pageq); 1032 continue; 1033 } 1034 VM_OBJECT_UNLOCK(object); 1035 } 1036 1037 /* 1038 * Compute the number of pages we want to try to move from the 1039 * active queue to the inactive queue. 1040 */ 1041 page_shortage = vm_paging_target() + 1042 cnt.v_inactive_target - cnt.v_inactive_count; 1043 page_shortage += addl_page_shortage; 1044 1045 /* 1046 * Scan the active queue for things we can deactivate. We nominally 1047 * track the per-page activity counter and use it to locate 1048 * deactivation candidates. 1049 */ 1050 pcount = cnt.v_active_count; 1051 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1052 1053 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) { 1054 1055 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE), 1056 ("vm_pageout_scan: page %p isn't active", m)); 1057 1058 next = TAILQ_NEXT(m, pageq); 1059 object = m->object; 1060 if ((m->flags & PG_MARKER) != 0) { 1061 m = next; 1062 continue; 1063 } 1064 if (!VM_OBJECT_TRYLOCK(object) && 1065 !vm_pageout_fallback_object_lock(m, &next)) { 1066 VM_OBJECT_UNLOCK(object); 1067 m = next; 1068 continue; 1069 } 1070 1071 /* 1072 * Don't deactivate pages that are busy. 1073 */ 1074 if ((m->busy != 0) || 1075 (m->oflags & VPO_BUSY) || 1076 (m->hold_count != 0)) { 1077 VM_OBJECT_UNLOCK(object); 1078 vm_page_requeue(m); 1079 m = next; 1080 continue; 1081 } 1082 1083 /* 1084 * The count for pagedaemon pages is done after checking the 1085 * page for eligibility... 1086 */ 1087 cnt.v_pdpages++; 1088 1089 /* 1090 * Check to see "how much" the page has been used. 1091 */ 1092 actcount = 0; 1093 if (object->ref_count != 0) { 1094 if (m->flags & PG_REFERENCED) { 1095 actcount += 1; 1096 } 1097 actcount += pmap_ts_referenced(m); 1098 if (actcount) { 1099 m->act_count += ACT_ADVANCE + actcount; 1100 if (m->act_count > ACT_MAX) 1101 m->act_count = ACT_MAX; 1102 } 1103 } 1104 1105 /* 1106 * Since we have "tested" this bit, we need to clear it now. 1107 */ 1108 vm_page_flag_clear(m, PG_REFERENCED); 1109 1110 /* 1111 * Only if an object is currently being used, do we use the 1112 * page activation count stats. 1113 */ 1114 if (actcount && (object->ref_count != 0)) { 1115 vm_page_requeue(m); 1116 } else { 1117 m->act_count -= min(m->act_count, ACT_DECLINE); 1118 if (vm_pageout_algorithm || 1119 object->ref_count == 0 || 1120 m->act_count == 0) { 1121 page_shortage--; 1122 if (object->ref_count == 0) { 1123 pmap_remove_all(m); 1124 if (m->dirty == 0) 1125 vm_page_cache(m); 1126 else 1127 vm_page_deactivate(m); 1128 } else { 1129 vm_page_deactivate(m); 1130 } 1131 } else { 1132 vm_page_requeue(m); 1133 } 1134 } 1135 VM_OBJECT_UNLOCK(object); 1136 m = next; 1137 } 1138 vm_page_unlock_queues(); 1139 #if !defined(NO_SWAPPING) 1140 /* 1141 * Idle process swapout -- run once per second. 1142 */ 1143 if (vm_swap_idle_enabled) { 1144 static long lsec; 1145 if (time_second != lsec) { 1146 vm_req_vmdaemon(VM_SWAP_IDLE); 1147 lsec = time_second; 1148 } 1149 } 1150 #endif 1151 1152 /* 1153 * If we didn't get enough free pages, and we have skipped a vnode 1154 * in a writeable object, wakeup the sync daemon. And kick swapout 1155 * if we did not get enough free pages. 1156 */ 1157 if (vm_paging_target() > 0) { 1158 if (vnodes_skipped && vm_page_count_min()) 1159 (void) speedup_syncer(); 1160 #if !defined(NO_SWAPPING) 1161 if (vm_swap_enabled && vm_page_count_target()) 1162 vm_req_vmdaemon(VM_SWAP_NORMAL); 1163 #endif 1164 } 1165 1166 /* 1167 * If we are critically low on one of RAM or swap and low on 1168 * the other, kill the largest process. However, we avoid 1169 * doing this on the first pass in order to give ourselves a 1170 * chance to flush out dirty vnode-backed pages and to allow 1171 * active pages to be moved to the inactive queue and reclaimed. 1172 */ 1173 if (pass != 0 && 1174 ((swap_pager_avail < 64 && vm_page_count_min()) || 1175 (swap_pager_full && vm_paging_target() > 0))) 1176 vm_pageout_oom(VM_OOM_MEM); 1177 } 1178 1179 1180 void 1181 vm_pageout_oom(int shortage) 1182 { 1183 struct proc *p, *bigproc; 1184 vm_offset_t size, bigsize; 1185 struct thread *td; 1186 struct vmspace *vm; 1187 1188 /* 1189 * We keep the process bigproc locked once we find it to keep anyone 1190 * from messing with it; however, there is a possibility of 1191 * deadlock if process B is bigproc and one of it's child processes 1192 * attempts to propagate a signal to B while we are waiting for A's 1193 * lock while walking this list. To avoid this, we don't block on 1194 * the process lock but just skip a process if it is already locked. 1195 */ 1196 bigproc = NULL; 1197 bigsize = 0; 1198 sx_slock(&allproc_lock); 1199 FOREACH_PROC_IN_SYSTEM(p) { 1200 int breakout; 1201 1202 if (PROC_TRYLOCK(p) == 0) 1203 continue; 1204 /* 1205 * If this is a system or protected process, skip it. 1206 */ 1207 if ((p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM)) || 1208 (p->p_pid == 1) || 1209 ((p->p_pid < 48) && (swap_pager_avail != 0))) { 1210 PROC_UNLOCK(p); 1211 continue; 1212 } 1213 /* 1214 * If the process is in a non-running type state, 1215 * don't touch it. Check all the threads individually. 1216 */ 1217 breakout = 0; 1218 FOREACH_THREAD_IN_PROC(p, td) { 1219 thread_lock(td); 1220 if (!TD_ON_RUNQ(td) && 1221 !TD_IS_RUNNING(td) && 1222 !TD_IS_SLEEPING(td)) { 1223 thread_unlock(td); 1224 breakout = 1; 1225 break; 1226 } 1227 thread_unlock(td); 1228 } 1229 if (breakout) { 1230 PROC_UNLOCK(p); 1231 continue; 1232 } 1233 /* 1234 * get the process size 1235 */ 1236 vm = vmspace_acquire_ref(p); 1237 if (vm == NULL) { 1238 PROC_UNLOCK(p); 1239 continue; 1240 } 1241 if (!vm_map_trylock_read(&vm->vm_map)) { 1242 vmspace_free(vm); 1243 PROC_UNLOCK(p); 1244 continue; 1245 } 1246 size = vmspace_swap_count(vm); 1247 vm_map_unlock_read(&vm->vm_map); 1248 if (shortage == VM_OOM_MEM) 1249 size += vmspace_resident_count(vm); 1250 vmspace_free(vm); 1251 /* 1252 * if the this process is bigger than the biggest one 1253 * remember it. 1254 */ 1255 if (size > bigsize) { 1256 if (bigproc != NULL) 1257 PROC_UNLOCK(bigproc); 1258 bigproc = p; 1259 bigsize = size; 1260 } else 1261 PROC_UNLOCK(p); 1262 } 1263 sx_sunlock(&allproc_lock); 1264 if (bigproc != NULL) { 1265 killproc(bigproc, "out of swap space"); 1266 sched_nice(bigproc, PRIO_MIN); 1267 PROC_UNLOCK(bigproc); 1268 wakeup(&cnt.v_free_count); 1269 } 1270 } 1271 1272 /* 1273 * This routine tries to maintain the pseudo LRU active queue, 1274 * so that during long periods of time where there is no paging, 1275 * that some statistic accumulation still occurs. This code 1276 * helps the situation where paging just starts to occur. 1277 */ 1278 static void 1279 vm_pageout_page_stats() 1280 { 1281 vm_object_t object; 1282 vm_page_t m,next; 1283 int pcount,tpcount; /* Number of pages to check */ 1284 static int fullintervalcount = 0; 1285 int page_shortage; 1286 1287 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1288 page_shortage = 1289 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) - 1290 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count); 1291 1292 if (page_shortage <= 0) 1293 return; 1294 1295 pcount = cnt.v_active_count; 1296 fullintervalcount += vm_pageout_stats_interval; 1297 if (fullintervalcount < vm_pageout_full_stats_interval) { 1298 tpcount = (int64_t)vm_pageout_stats_max * cnt.v_active_count / 1299 cnt.v_page_count; 1300 if (pcount > tpcount) 1301 pcount = tpcount; 1302 } else { 1303 fullintervalcount = 0; 1304 } 1305 1306 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1307 while ((m != NULL) && (pcount-- > 0)) { 1308 int actcount; 1309 1310 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE), 1311 ("vm_pageout_page_stats: page %p isn't active", m)); 1312 1313 next = TAILQ_NEXT(m, pageq); 1314 object = m->object; 1315 1316 if ((m->flags & PG_MARKER) != 0) { 1317 m = next; 1318 continue; 1319 } 1320 if (!VM_OBJECT_TRYLOCK(object) && 1321 !vm_pageout_fallback_object_lock(m, &next)) { 1322 VM_OBJECT_UNLOCK(object); 1323 m = next; 1324 continue; 1325 } 1326 1327 /* 1328 * Don't deactivate pages that are busy. 1329 */ 1330 if ((m->busy != 0) || 1331 (m->oflags & VPO_BUSY) || 1332 (m->hold_count != 0)) { 1333 VM_OBJECT_UNLOCK(object); 1334 vm_page_requeue(m); 1335 m = next; 1336 continue; 1337 } 1338 1339 actcount = 0; 1340 if (m->flags & PG_REFERENCED) { 1341 vm_page_flag_clear(m, PG_REFERENCED); 1342 actcount += 1; 1343 } 1344 1345 actcount += pmap_ts_referenced(m); 1346 if (actcount) { 1347 m->act_count += ACT_ADVANCE + actcount; 1348 if (m->act_count > ACT_MAX) 1349 m->act_count = ACT_MAX; 1350 vm_page_requeue(m); 1351 } else { 1352 if (m->act_count == 0) { 1353 /* 1354 * We turn off page access, so that we have 1355 * more accurate RSS stats. We don't do this 1356 * in the normal page deactivation when the 1357 * system is loaded VM wise, because the 1358 * cost of the large number of page protect 1359 * operations would be higher than the value 1360 * of doing the operation. 1361 */ 1362 pmap_remove_all(m); 1363 vm_page_deactivate(m); 1364 } else { 1365 m->act_count -= min(m->act_count, ACT_DECLINE); 1366 vm_page_requeue(m); 1367 } 1368 } 1369 VM_OBJECT_UNLOCK(object); 1370 m = next; 1371 } 1372 } 1373 1374 /* 1375 * vm_pageout is the high level pageout daemon. 1376 */ 1377 static void 1378 vm_pageout() 1379 { 1380 int error, pass; 1381 1382 /* 1383 * Initialize some paging parameters. 1384 */ 1385 cnt.v_interrupt_free_min = 2; 1386 if (cnt.v_page_count < 2000) 1387 vm_pageout_page_count = 8; 1388 1389 /* 1390 * v_free_reserved needs to include enough for the largest 1391 * swap pager structures plus enough for any pv_entry structs 1392 * when paging. 1393 */ 1394 if (cnt.v_page_count > 1024) 1395 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200; 1396 else 1397 cnt.v_free_min = 4; 1398 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1399 cnt.v_interrupt_free_min; 1400 cnt.v_free_reserved = vm_pageout_page_count + 1401 cnt.v_pageout_free_min + (cnt.v_page_count / 768); 1402 cnt.v_free_severe = cnt.v_free_min / 2; 1403 cnt.v_free_min += cnt.v_free_reserved; 1404 cnt.v_free_severe += cnt.v_free_reserved; 1405 1406 /* 1407 * v_free_target and v_cache_min control pageout hysteresis. Note 1408 * that these are more a measure of the VM cache queue hysteresis 1409 * then the VM free queue. Specifically, v_free_target is the 1410 * high water mark (free+cache pages). 1411 * 1412 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the 1413 * low water mark, while v_free_min is the stop. v_cache_min must 1414 * be big enough to handle memory needs while the pageout daemon 1415 * is signalled and run to free more pages. 1416 */ 1417 if (cnt.v_free_count > 6144) 1418 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved; 1419 else 1420 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved; 1421 1422 if (cnt.v_free_count > 2048) { 1423 cnt.v_cache_min = cnt.v_free_target; 1424 cnt.v_cache_max = 2 * cnt.v_cache_min; 1425 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2; 1426 } else { 1427 cnt.v_cache_min = 0; 1428 cnt.v_cache_max = 0; 1429 cnt.v_inactive_target = cnt.v_free_count / 4; 1430 } 1431 if (cnt.v_inactive_target > cnt.v_free_count / 3) 1432 cnt.v_inactive_target = cnt.v_free_count / 3; 1433 1434 /* XXX does not really belong here */ 1435 if (vm_page_max_wired == 0) 1436 vm_page_max_wired = cnt.v_free_count / 3; 1437 1438 if (vm_pageout_stats_max == 0) 1439 vm_pageout_stats_max = cnt.v_free_target; 1440 1441 /* 1442 * Set interval in seconds for stats scan. 1443 */ 1444 if (vm_pageout_stats_interval == 0) 1445 vm_pageout_stats_interval = 5; 1446 if (vm_pageout_full_stats_interval == 0) 1447 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4; 1448 1449 swap_pager_swap_init(); 1450 pass = 0; 1451 /* 1452 * The pageout daemon is never done, so loop forever. 1453 */ 1454 while (TRUE) { 1455 /* 1456 * If we have enough free memory, wakeup waiters. Do 1457 * not clear vm_pages_needed until we reach our target, 1458 * otherwise we may be woken up over and over again and 1459 * waste a lot of cpu. 1460 */ 1461 mtx_lock(&vm_page_queue_free_mtx); 1462 if (vm_pages_needed && !vm_page_count_min()) { 1463 if (!vm_paging_needed()) 1464 vm_pages_needed = 0; 1465 wakeup(&cnt.v_free_count); 1466 } 1467 if (vm_pages_needed) { 1468 /* 1469 * Still not done, take a second pass without waiting 1470 * (unlimited dirty cleaning), otherwise sleep a bit 1471 * and try again. 1472 */ 1473 ++pass; 1474 if (pass > 1) 1475 msleep(&vm_pages_needed, 1476 &vm_page_queue_free_mtx, PVM, "psleep", 1477 hz / 2); 1478 } else { 1479 /* 1480 * Good enough, sleep & handle stats. Prime the pass 1481 * for the next run. 1482 */ 1483 if (pass > 1) 1484 pass = 1; 1485 else 1486 pass = 0; 1487 error = msleep(&vm_pages_needed, 1488 &vm_page_queue_free_mtx, PVM, "psleep", 1489 vm_pageout_stats_interval * hz); 1490 if (error && !vm_pages_needed) { 1491 mtx_unlock(&vm_page_queue_free_mtx); 1492 pass = 0; 1493 vm_page_lock_queues(); 1494 vm_pageout_page_stats(); 1495 vm_page_unlock_queues(); 1496 continue; 1497 } 1498 } 1499 if (vm_pages_needed) 1500 cnt.v_pdwakeups++; 1501 mtx_unlock(&vm_page_queue_free_mtx); 1502 vm_pageout_scan(pass); 1503 } 1504 } 1505 1506 /* 1507 * Unless the free page queue lock is held by the caller, this function 1508 * should be regarded as advisory. Specifically, the caller should 1509 * not msleep() on &cnt.v_free_count following this function unless 1510 * the free page queue lock is held until the msleep() is performed. 1511 */ 1512 void 1513 pagedaemon_wakeup() 1514 { 1515 1516 if (!vm_pages_needed && curthread->td_proc != pageproc) { 1517 vm_pages_needed = 1; 1518 wakeup(&vm_pages_needed); 1519 } 1520 } 1521 1522 #if !defined(NO_SWAPPING) 1523 static void 1524 vm_req_vmdaemon(int req) 1525 { 1526 static int lastrun = 0; 1527 1528 mtx_lock(&vm_daemon_mtx); 1529 vm_pageout_req_swapout |= req; 1530 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1531 wakeup(&vm_daemon_needed); 1532 lastrun = ticks; 1533 } 1534 mtx_unlock(&vm_daemon_mtx); 1535 } 1536 1537 static void 1538 vm_daemon() 1539 { 1540 struct rlimit rsslim; 1541 struct proc *p; 1542 struct thread *td; 1543 struct vmspace *vm; 1544 int breakout, swapout_flags; 1545 1546 while (TRUE) { 1547 mtx_lock(&vm_daemon_mtx); 1548 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0); 1549 swapout_flags = vm_pageout_req_swapout; 1550 vm_pageout_req_swapout = 0; 1551 mtx_unlock(&vm_daemon_mtx); 1552 if (swapout_flags) 1553 swapout_procs(swapout_flags); 1554 1555 /* 1556 * scan the processes for exceeding their rlimits or if 1557 * process is swapped out -- deactivate pages 1558 */ 1559 sx_slock(&allproc_lock); 1560 FOREACH_PROC_IN_SYSTEM(p) { 1561 vm_pindex_t limit, size; 1562 1563 /* 1564 * if this is a system process or if we have already 1565 * looked at this process, skip it. 1566 */ 1567 PROC_LOCK(p); 1568 if (p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) { 1569 PROC_UNLOCK(p); 1570 continue; 1571 } 1572 /* 1573 * if the process is in a non-running type state, 1574 * don't touch it. 1575 */ 1576 breakout = 0; 1577 FOREACH_THREAD_IN_PROC(p, td) { 1578 thread_lock(td); 1579 if (!TD_ON_RUNQ(td) && 1580 !TD_IS_RUNNING(td) && 1581 !TD_IS_SLEEPING(td)) { 1582 thread_unlock(td); 1583 breakout = 1; 1584 break; 1585 } 1586 thread_unlock(td); 1587 } 1588 if (breakout) { 1589 PROC_UNLOCK(p); 1590 continue; 1591 } 1592 /* 1593 * get a limit 1594 */ 1595 lim_rlimit(p, RLIMIT_RSS, &rsslim); 1596 limit = OFF_TO_IDX( 1597 qmin(rsslim.rlim_cur, rsslim.rlim_max)); 1598 1599 /* 1600 * let processes that are swapped out really be 1601 * swapped out set the limit to nothing (will force a 1602 * swap-out.) 1603 */ 1604 if ((p->p_flag & P_INMEM) == 0) 1605 limit = 0; /* XXX */ 1606 vm = vmspace_acquire_ref(p); 1607 PROC_UNLOCK(p); 1608 if (vm == NULL) 1609 continue; 1610 1611 size = vmspace_resident_count(vm); 1612 if (limit >= 0 && size >= limit) { 1613 vm_pageout_map_deactivate_pages( 1614 &vm->vm_map, limit); 1615 } 1616 vmspace_free(vm); 1617 } 1618 sx_sunlock(&allproc_lock); 1619 } 1620 } 1621 #endif /* !defined(NO_SWAPPING) */ 1622