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 #include <machine/mutex.h> 109 110 /* 111 * System initialization 112 */ 113 114 /* the kernel process "vm_pageout"*/ 115 static void vm_pageout(void); 116 static int vm_pageout_clean(vm_page_t); 117 static void vm_pageout_scan(int pass); 118 119 struct proc *pageproc; 120 121 static struct kproc_desc page_kp = { 122 "pagedaemon", 123 vm_pageout, 124 &pageproc 125 }; 126 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, 127 &page_kp); 128 129 #if !defined(NO_SWAPPING) 130 /* the kernel process "vm_daemon"*/ 131 static void vm_daemon(void); 132 static struct proc *vmproc; 133 134 static struct kproc_desc vm_kp = { 135 "vmdaemon", 136 vm_daemon, 137 &vmproc 138 }; 139 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp); 140 #endif 141 142 143 int vm_pages_needed; /* Event on which pageout daemon sleeps */ 144 int vm_pageout_deficit; /* Estimated number of pages deficit */ 145 int vm_pageout_pages_needed; /* flag saying that the pageout daemon needs pages */ 146 147 #if !defined(NO_SWAPPING) 148 static int vm_pageout_req_swapout; /* XXX */ 149 static int vm_daemon_needed; 150 static struct mtx vm_daemon_mtx; 151 /* Allow for use by vm_pageout before vm_daemon is initialized. */ 152 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF); 153 #endif 154 static int vm_max_launder = 32; 155 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0; 156 static int vm_pageout_full_stats_interval = 0; 157 static int vm_pageout_algorithm=0; 158 static int defer_swap_pageouts=0; 159 static int disable_swap_pageouts=0; 160 161 #if defined(NO_SWAPPING) 162 static int vm_swap_enabled=0; 163 static int vm_swap_idle_enabled=0; 164 #else 165 static int vm_swap_enabled=1; 166 static int vm_swap_idle_enabled=0; 167 #endif 168 169 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm, 170 CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt"); 171 172 SYSCTL_INT(_vm, OID_AUTO, max_launder, 173 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout"); 174 175 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max, 176 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length"); 177 178 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval, 179 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan"); 180 181 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval, 182 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan"); 183 184 #if defined(NO_SWAPPING) 185 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 186 CTLFLAG_RD, &vm_swap_enabled, 0, "Enable entire process swapout"); 187 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 188 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 189 #else 190 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 191 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout"); 192 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 193 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 194 #endif 195 196 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts, 197 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem"); 198 199 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts, 200 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages"); 201 202 static int pageout_lock_miss; 203 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss, 204 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout"); 205 206 #define VM_PAGEOUT_PAGE_COUNT 16 207 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT; 208 209 int vm_page_max_wired; /* XXX max # of wired pages system-wide */ 210 SYSCTL_INT(_vm, OID_AUTO, max_wired, 211 CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count"); 212 213 #if !defined(NO_SWAPPING) 214 static void vm_pageout_map_deactivate_pages(vm_map_t, long); 215 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long); 216 static void vm_req_vmdaemon(int req); 217 #endif 218 static void vm_pageout_page_stats(void); 219 220 /* 221 * vm_pageout_fallback_object_lock: 222 * 223 * Lock vm object currently associated with `m'. VM_OBJECT_TRYLOCK is 224 * known to have failed and page queue must be either PQ_ACTIVE or 225 * PQ_INACTIVE. To avoid lock order violation, unlock the page queues 226 * while locking the vm object. Use marker page to detect page queue 227 * changes and maintain notion of next page on page queue. Return 228 * TRUE if no changes were detected, FALSE otherwise. vm object is 229 * locked on return. 230 * 231 * This function depends on both the lock portion of struct vm_object 232 * and normal struct vm_page being type stable. 233 */ 234 boolean_t 235 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next) 236 { 237 struct vm_page marker; 238 boolean_t unchanged; 239 u_short queue; 240 vm_object_t object; 241 242 /* 243 * Initialize our marker 244 */ 245 bzero(&marker, sizeof(marker)); 246 marker.flags = PG_FICTITIOUS | PG_MARKER; 247 marker.oflags = VPO_BUSY; 248 marker.queue = m->queue; 249 marker.wire_count = 1; 250 251 queue = m->queue; 252 object = m->object; 253 254 TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl, 255 m, &marker, pageq); 256 vm_page_unlock_queues(); 257 VM_OBJECT_LOCK(object); 258 vm_page_lock_queues(); 259 260 /* Page queue might have changed. */ 261 *next = TAILQ_NEXT(&marker, pageq); 262 unchanged = (m->queue == queue && 263 m->object == object && 264 &marker == TAILQ_NEXT(m, pageq)); 265 TAILQ_REMOVE(&vm_page_queues[queue].pl, 266 &marker, pageq); 267 return (unchanged); 268 } 269 270 /* 271 * vm_pageout_clean: 272 * 273 * Clean the page and remove it from the laundry. 274 * 275 * We set the busy bit to cause potential page faults on this page to 276 * block. Note the careful timing, however, the busy bit isn't set till 277 * late and we cannot do anything that will mess with the page. 278 */ 279 static int 280 vm_pageout_clean(m) 281 vm_page_t m; 282 { 283 vm_object_t object; 284 vm_page_t mc[2*vm_pageout_page_count]; 285 int pageout_count; 286 int ib, is, page_base; 287 vm_pindex_t pindex = m->pindex; 288 289 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 290 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 291 292 /* 293 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP 294 * with the new swapper, but we could have serious problems paging 295 * out other object types if there is insufficient memory. 296 * 297 * Unfortunately, checking free memory here is far too late, so the 298 * check has been moved up a procedural level. 299 */ 300 301 /* 302 * Can't clean the page if it's busy or held. 303 */ 304 if ((m->hold_count != 0) || 305 ((m->busy != 0) || (m->oflags & VPO_BUSY))) { 306 return 0; 307 } 308 309 mc[vm_pageout_page_count] = m; 310 pageout_count = 1; 311 page_base = vm_pageout_page_count; 312 ib = 1; 313 is = 1; 314 315 /* 316 * Scan object for clusterable pages. 317 * 318 * We can cluster ONLY if: ->> the page is NOT 319 * clean, wired, busy, held, or mapped into a 320 * buffer, and one of the following: 321 * 1) The page is inactive, or a seldom used 322 * active page. 323 * -or- 324 * 2) we force the issue. 325 * 326 * During heavy mmap/modification loads the pageout 327 * daemon can really fragment the underlying file 328 * due to flushing pages out of order and not trying 329 * align the clusters (which leave sporatic out-of-order 330 * holes). To solve this problem we do the reverse scan 331 * first and attempt to align our cluster, then do a 332 * forward scan if room remains. 333 */ 334 object = m->object; 335 more: 336 while (ib && pageout_count < vm_pageout_page_count) { 337 vm_page_t p; 338 339 if (ib > pindex) { 340 ib = 0; 341 break; 342 } 343 344 if ((p = vm_page_lookup(object, pindex - ib)) == NULL) { 345 ib = 0; 346 break; 347 } 348 if ((p->oflags & VPO_BUSY) || p->busy) { 349 ib = 0; 350 break; 351 } 352 vm_page_test_dirty(p); 353 if ((p->dirty & p->valid) == 0 || 354 p->queue != PQ_INACTIVE || 355 p->wire_count != 0 || /* may be held by buf cache */ 356 p->hold_count != 0) { /* may be undergoing I/O */ 357 ib = 0; 358 break; 359 } 360 mc[--page_base] = p; 361 ++pageout_count; 362 ++ib; 363 /* 364 * alignment boundry, stop here and switch directions. Do 365 * not clear ib. 366 */ 367 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0) 368 break; 369 } 370 371 while (pageout_count < vm_pageout_page_count && 372 pindex + is < object->size) { 373 vm_page_t p; 374 375 if ((p = vm_page_lookup(object, pindex + is)) == NULL) 376 break; 377 if ((p->oflags & VPO_BUSY) || p->busy) { 378 break; 379 } 380 vm_page_test_dirty(p); 381 if ((p->dirty & p->valid) == 0 || 382 p->queue != PQ_INACTIVE || 383 p->wire_count != 0 || /* may be held by buf cache */ 384 p->hold_count != 0) { /* may be undergoing I/O */ 385 break; 386 } 387 mc[page_base + pageout_count] = p; 388 ++pageout_count; 389 ++is; 390 } 391 392 /* 393 * If we exhausted our forward scan, continue with the reverse scan 394 * when possible, even past a page boundry. This catches boundry 395 * conditions. 396 */ 397 if (ib && pageout_count < vm_pageout_page_count) 398 goto more; 399 400 /* 401 * we allow reads during pageouts... 402 */ 403 return (vm_pageout_flush(&mc[page_base], pageout_count, 0)); 404 } 405 406 /* 407 * vm_pageout_flush() - launder the given pages 408 * 409 * The given pages are laundered. Note that we setup for the start of 410 * I/O ( i.e. busy the page ), mark it read-only, and bump the object 411 * reference count all in here rather then in the parent. If we want 412 * the parent to do more sophisticated things we may have to change 413 * the ordering. 414 */ 415 int 416 vm_pageout_flush(vm_page_t *mc, int count, int flags) 417 { 418 vm_object_t object = mc[0]->object; 419 int pageout_status[count]; 420 int numpagedout = 0; 421 int i; 422 423 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 424 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 425 /* 426 * Initiate I/O. Bump the vm_page_t->busy counter and 427 * mark the pages read-only. 428 * 429 * We do not have to fixup the clean/dirty bits here... we can 430 * allow the pager to do it after the I/O completes. 431 * 432 * NOTE! mc[i]->dirty may be partial or fragmented due to an 433 * edge case with file fragments. 434 */ 435 for (i = 0; i < count; i++) { 436 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, 437 ("vm_pageout_flush: partially invalid page %p index %d/%d", 438 mc[i], i, count)); 439 vm_page_io_start(mc[i]); 440 pmap_remove_write(mc[i]); 441 } 442 vm_page_unlock_queues(); 443 vm_object_pip_add(object, count); 444 445 vm_pager_put_pages(object, mc, count, flags, pageout_status); 446 447 vm_page_lock_queues(); 448 for (i = 0; i < count; i++) { 449 vm_page_t mt = mc[i]; 450 451 KASSERT(pageout_status[i] == VM_PAGER_PEND || 452 (mt->flags & PG_WRITEABLE) == 0, 453 ("vm_pageout_flush: page %p is not write protected", mt)); 454 switch (pageout_status[i]) { 455 case VM_PAGER_OK: 456 case VM_PAGER_PEND: 457 numpagedout++; 458 break; 459 case VM_PAGER_BAD: 460 /* 461 * Page outside of range of object. Right now we 462 * essentially lose the changes by pretending it 463 * worked. 464 */ 465 vm_page_undirty(mt); 466 break; 467 case VM_PAGER_ERROR: 468 case VM_PAGER_FAIL: 469 /* 470 * If page couldn't be paged out, then reactivate the 471 * page so it doesn't clog the inactive list. (We 472 * will try paging out it again later). 473 */ 474 vm_page_activate(mt); 475 break; 476 case VM_PAGER_AGAIN: 477 break; 478 } 479 480 /* 481 * If the operation is still going, leave the page busy to 482 * block all other accesses. Also, leave the paging in 483 * progress indicator set so that we don't attempt an object 484 * collapse. 485 */ 486 if (pageout_status[i] != VM_PAGER_PEND) { 487 vm_object_pip_wakeup(object); 488 vm_page_io_finish(mt); 489 if (vm_page_count_severe()) 490 vm_page_try_to_cache(mt); 491 } 492 } 493 return numpagedout; 494 } 495 496 #if !defined(NO_SWAPPING) 497 /* 498 * vm_pageout_object_deactivate_pages 499 * 500 * deactivate enough pages to satisfy the inactive target 501 * requirements or if vm_page_proc_limit is set, then 502 * deactivate all of the pages in the object and its 503 * backing_objects. 504 * 505 * The object and map must be locked. 506 */ 507 static void 508 vm_pageout_object_deactivate_pages(pmap, first_object, desired) 509 pmap_t pmap; 510 vm_object_t first_object; 511 long desired; 512 { 513 vm_object_t backing_object, object; 514 vm_page_t p, next; 515 int actcount, rcount, remove_mode; 516 517 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED); 518 if (first_object->type == OBJT_DEVICE || first_object->type == OBJT_PHYS) 519 return; 520 for (object = first_object;; object = backing_object) { 521 if (pmap_resident_count(pmap) <= desired) 522 goto unlock_return; 523 if (object->paging_in_progress) 524 goto unlock_return; 525 526 remove_mode = 0; 527 if (object->shadow_count > 1) 528 remove_mode = 1; 529 /* 530 * scan the objects entire memory queue 531 */ 532 rcount = object->resident_page_count; 533 p = TAILQ_FIRST(&object->memq); 534 vm_page_lock_queues(); 535 while (p && (rcount-- > 0)) { 536 if (pmap_resident_count(pmap) <= desired) { 537 vm_page_unlock_queues(); 538 goto unlock_return; 539 } 540 next = TAILQ_NEXT(p, listq); 541 cnt.v_pdpages++; 542 if (p->wire_count != 0 || 543 p->hold_count != 0 || 544 p->busy != 0 || 545 (p->oflags & VPO_BUSY) || 546 (p->flags & PG_UNMANAGED) || 547 !pmap_page_exists_quick(pmap, p)) { 548 p = next; 549 continue; 550 } 551 actcount = pmap_ts_referenced(p); 552 if (actcount) { 553 vm_page_flag_set(p, PG_REFERENCED); 554 } else if (p->flags & PG_REFERENCED) { 555 actcount = 1; 556 } 557 if ((p->queue != PQ_ACTIVE) && 558 (p->flags & PG_REFERENCED)) { 559 vm_page_activate(p); 560 p->act_count += actcount; 561 vm_page_flag_clear(p, PG_REFERENCED); 562 } else if (p->queue == PQ_ACTIVE) { 563 if ((p->flags & PG_REFERENCED) == 0) { 564 p->act_count -= min(p->act_count, ACT_DECLINE); 565 if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) { 566 pmap_remove_all(p); 567 vm_page_deactivate(p); 568 } else { 569 vm_page_requeue(p); 570 } 571 } else { 572 vm_page_activate(p); 573 vm_page_flag_clear(p, PG_REFERENCED); 574 if (p->act_count < (ACT_MAX - ACT_ADVANCE)) 575 p->act_count += ACT_ADVANCE; 576 vm_page_requeue(p); 577 } 578 } else if (p->queue == PQ_INACTIVE) { 579 pmap_remove_all(p); 580 } 581 p = next; 582 } 583 vm_page_unlock_queues(); 584 if ((backing_object = object->backing_object) == NULL) 585 goto unlock_return; 586 VM_OBJECT_LOCK(backing_object); 587 if (object != first_object) 588 VM_OBJECT_UNLOCK(object); 589 } 590 unlock_return: 591 if (object != first_object) 592 VM_OBJECT_UNLOCK(object); 593 } 594 595 /* 596 * deactivate some number of pages in a map, try to do it fairly, but 597 * that is really hard to do. 598 */ 599 static void 600 vm_pageout_map_deactivate_pages(map, desired) 601 vm_map_t map; 602 long desired; 603 { 604 vm_map_entry_t tmpe; 605 vm_object_t obj, bigobj; 606 int nothingwired; 607 608 if (!vm_map_trylock(map)) 609 return; 610 611 bigobj = NULL; 612 nothingwired = TRUE; 613 614 /* 615 * first, search out the biggest object, and try to free pages from 616 * that. 617 */ 618 tmpe = map->header.next; 619 while (tmpe != &map->header) { 620 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 621 obj = tmpe->object.vm_object; 622 if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) { 623 if (obj->shadow_count <= 1 && 624 (bigobj == NULL || 625 bigobj->resident_page_count < obj->resident_page_count)) { 626 if (bigobj != NULL) 627 VM_OBJECT_UNLOCK(bigobj); 628 bigobj = obj; 629 } else 630 VM_OBJECT_UNLOCK(obj); 631 } 632 } 633 if (tmpe->wired_count > 0) 634 nothingwired = FALSE; 635 tmpe = tmpe->next; 636 } 637 638 if (bigobj != NULL) { 639 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired); 640 VM_OBJECT_UNLOCK(bigobj); 641 } 642 /* 643 * Next, hunt around for other pages to deactivate. We actually 644 * do this search sort of wrong -- .text first is not the best idea. 645 */ 646 tmpe = map->header.next; 647 while (tmpe != &map->header) { 648 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 649 break; 650 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 651 obj = tmpe->object.vm_object; 652 if (obj != NULL) { 653 VM_OBJECT_LOCK(obj); 654 vm_pageout_object_deactivate_pages(map->pmap, obj, desired); 655 VM_OBJECT_UNLOCK(obj); 656 } 657 } 658 tmpe = tmpe->next; 659 } 660 661 /* 662 * Remove all mappings if a process is swapped out, this will free page 663 * table pages. 664 */ 665 if (desired == 0 && nothingwired) { 666 pmap_remove(vm_map_pmap(map), vm_map_min(map), 667 vm_map_max(map)); 668 } 669 vm_map_unlock(map); 670 } 671 #endif /* !defined(NO_SWAPPING) */ 672 673 /* 674 * vm_pageout_scan does the dirty work for the pageout daemon. 675 */ 676 static void 677 vm_pageout_scan(int pass) 678 { 679 vm_page_t m, next; 680 struct vm_page marker; 681 int page_shortage, maxscan, pcount; 682 int addl_page_shortage, addl_page_shortage_init; 683 vm_object_t object; 684 int actcount; 685 int vnodes_skipped = 0; 686 int maxlaunder; 687 688 /* 689 * Decrease registered cache sizes. 690 */ 691 EVENTHANDLER_INVOKE(vm_lowmem, 0); 692 /* 693 * We do this explicitly after the caches have been drained above. 694 */ 695 uma_reclaim(); 696 697 addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit); 698 699 /* 700 * Calculate the number of pages we want to either free or move 701 * to the cache. 702 */ 703 page_shortage = vm_paging_target() + addl_page_shortage_init; 704 705 /* 706 * Initialize our marker 707 */ 708 bzero(&marker, sizeof(marker)); 709 marker.flags = PG_FICTITIOUS | PG_MARKER; 710 marker.oflags = VPO_BUSY; 711 marker.queue = PQ_INACTIVE; 712 marker.wire_count = 1; 713 714 /* 715 * Start scanning the inactive queue for pages we can move to the 716 * cache or free. The scan will stop when the target is reached or 717 * we have scanned the entire inactive queue. Note that m->act_count 718 * is not used to form decisions for the inactive queue, only for the 719 * active queue. 720 * 721 * maxlaunder limits the number of dirty pages we flush per scan. 722 * For most systems a smaller value (16 or 32) is more robust under 723 * extreme memory and disk pressure because any unnecessary writes 724 * to disk can result in extreme performance degredation. However, 725 * systems with excessive dirty pages (especially when MAP_NOSYNC is 726 * used) will die horribly with limited laundering. If the pageout 727 * daemon cannot clean enough pages in the first pass, we let it go 728 * all out in succeeding passes. 729 */ 730 if ((maxlaunder = vm_max_launder) <= 1) 731 maxlaunder = 1; 732 if (pass) 733 maxlaunder = 10000; 734 vm_page_lock_queues(); 735 rescan0: 736 addl_page_shortage = addl_page_shortage_init; 737 maxscan = cnt.v_inactive_count; 738 739 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl); 740 m != NULL && maxscan-- > 0 && page_shortage > 0; 741 m = next) { 742 743 cnt.v_pdpages++; 744 745 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE) { 746 goto rescan0; 747 } 748 749 next = TAILQ_NEXT(m, pageq); 750 object = m->object; 751 752 /* 753 * skip marker pages 754 */ 755 if (m->flags & PG_MARKER) 756 continue; 757 758 /* 759 * A held page may be undergoing I/O, so skip it. 760 */ 761 if (m->hold_count) { 762 vm_page_requeue(m); 763 addl_page_shortage++; 764 continue; 765 } 766 /* 767 * Don't mess with busy pages, keep in the front of the 768 * queue, most likely are being paged out. 769 */ 770 if (!VM_OBJECT_TRYLOCK(object) && 771 (!vm_pageout_fallback_object_lock(m, &next) || 772 m->hold_count != 0)) { 773 VM_OBJECT_UNLOCK(object); 774 addl_page_shortage++; 775 continue; 776 } 777 if (m->busy || (m->oflags & VPO_BUSY)) { 778 VM_OBJECT_UNLOCK(object); 779 addl_page_shortage++; 780 continue; 781 } 782 783 /* 784 * If the object is not being used, we ignore previous 785 * references. 786 */ 787 if (object->ref_count == 0) { 788 vm_page_flag_clear(m, PG_REFERENCED); 789 KASSERT(!pmap_page_is_mapped(m), 790 ("vm_pageout_scan: page %p is mapped", m)); 791 792 /* 793 * Otherwise, if the page has been referenced while in the 794 * inactive queue, we bump the "activation count" upwards, 795 * making it less likely that the page will be added back to 796 * the inactive queue prematurely again. Here we check the 797 * page tables (or emulated bits, if any), given the upper 798 * level VM system not knowing anything about existing 799 * references. 800 */ 801 } else if (((m->flags & PG_REFERENCED) == 0) && 802 (actcount = pmap_ts_referenced(m))) { 803 vm_page_activate(m); 804 VM_OBJECT_UNLOCK(object); 805 m->act_count += (actcount + ACT_ADVANCE); 806 continue; 807 } 808 809 /* 810 * If the upper level VM system knows about any page 811 * references, we activate the page. We also set the 812 * "activation count" higher than normal so that we will less 813 * likely place pages back onto the inactive queue again. 814 */ 815 if ((m->flags & PG_REFERENCED) != 0) { 816 vm_page_flag_clear(m, PG_REFERENCED); 817 actcount = pmap_ts_referenced(m); 818 vm_page_activate(m); 819 VM_OBJECT_UNLOCK(object); 820 m->act_count += (actcount + ACT_ADVANCE + 1); 821 continue; 822 } 823 824 /* 825 * If the upper level VM system doesn't know anything about 826 * the page being dirty, we have to check for it again. As 827 * far as the VM code knows, any partially dirty pages are 828 * fully dirty. 829 */ 830 if (m->dirty == 0 && !pmap_is_modified(m)) { 831 /* 832 * Avoid a race condition: Unless write access is 833 * removed from the page, another processor could 834 * modify it before all access is removed by the call 835 * to vm_page_cache() below. If vm_page_cache() finds 836 * that the page has been modified when it removes all 837 * access, it panics because it cannot cache dirty 838 * pages. In principle, we could eliminate just write 839 * access here rather than all access. In the expected 840 * case, when there are no last instant modifications 841 * to the page, removing all access will be cheaper 842 * overall. 843 */ 844 if ((m->flags & PG_WRITEABLE) != 0) 845 pmap_remove_all(m); 846 } else { 847 vm_page_dirty(m); 848 } 849 850 if (m->valid == 0) { 851 /* 852 * Invalid pages can be easily freed 853 */ 854 vm_page_free(m); 855 cnt.v_dfree++; 856 --page_shortage; 857 } else if (m->dirty == 0) { 858 /* 859 * Clean pages can be placed onto the cache queue. 860 * This effectively frees them. 861 */ 862 vm_page_cache(m); 863 --page_shortage; 864 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) { 865 /* 866 * Dirty pages need to be paged out, but flushing 867 * a page is extremely expensive verses freeing 868 * a clean page. Rather then artificially limiting 869 * the number of pages we can flush, we instead give 870 * dirty pages extra priority on the inactive queue 871 * by forcing them to be cycled through the queue 872 * twice before being flushed, after which the 873 * (now clean) page will cycle through once more 874 * before being freed. This significantly extends 875 * the thrash point for a heavily loaded machine. 876 */ 877 vm_page_flag_set(m, PG_WINATCFLS); 878 vm_page_requeue(m); 879 } else if (maxlaunder > 0) { 880 /* 881 * We always want to try to flush some dirty pages if 882 * we encounter them, to keep the system stable. 883 * Normally this number is small, but under extreme 884 * pressure where there are insufficient clean pages 885 * on the inactive queue, we may have to go all out. 886 */ 887 int swap_pageouts_ok, vfslocked = 0; 888 struct vnode *vp = NULL; 889 struct mount *mp = NULL; 890 891 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 892 swap_pageouts_ok = 1; 893 } else { 894 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 895 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 896 vm_page_count_min()); 897 898 } 899 900 /* 901 * We don't bother paging objects that are "dead". 902 * Those objects are in a "rundown" state. 903 */ 904 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 905 VM_OBJECT_UNLOCK(object); 906 vm_page_requeue(m); 907 continue; 908 } 909 910 /* 911 * Following operations may unlock 912 * vm_page_queue_mtx, invalidating the 'next' 913 * pointer. To prevent an inordinate number 914 * of restarts we use our marker to remember 915 * our place. 916 * 917 */ 918 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, 919 m, &marker, pageq); 920 /* 921 * The object is already known NOT to be dead. It 922 * is possible for the vget() to block the whole 923 * pageout daemon, but the new low-memory handling 924 * code should prevent it. 925 * 926 * The previous code skipped locked vnodes and, worse, 927 * reordered pages in the queue. This results in 928 * completely non-deterministic operation and, on a 929 * busy system, can lead to extremely non-optimal 930 * pageouts. For example, it can cause clean pages 931 * to be freed and dirty pages to be moved to the end 932 * of the queue. Since dirty pages are also moved to 933 * the end of the queue once-cleaned, this gives 934 * way too large a weighting to defering the freeing 935 * of dirty pages. 936 * 937 * We can't wait forever for the vnode lock, we might 938 * deadlock due to a vn_read() getting stuck in 939 * vm_wait while holding this vnode. We skip the 940 * vnode if we can't get it in a reasonable amount 941 * of time. 942 */ 943 if (object->type == OBJT_VNODE) { 944 vp = object->handle; 945 if (vp->v_type == VREG && 946 vn_start_write(vp, &mp, V_NOWAIT) != 0) { 947 mp = NULL; 948 ++pageout_lock_miss; 949 if (object->flags & OBJ_MIGHTBEDIRTY) 950 vnodes_skipped++; 951 goto unlock_and_continue; 952 } 953 vm_page_unlock_queues(); 954 vm_object_reference_locked(object); 955 VM_OBJECT_UNLOCK(object); 956 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 957 if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK, 958 curthread)) { 959 VM_OBJECT_LOCK(object); 960 vm_page_lock_queues(); 961 ++pageout_lock_miss; 962 if (object->flags & OBJ_MIGHTBEDIRTY) 963 vnodes_skipped++; 964 vp = NULL; 965 goto unlock_and_continue; 966 } 967 VM_OBJECT_LOCK(object); 968 vm_page_lock_queues(); 969 /* 970 * The page might have been moved to another 971 * queue during potential blocking in vget() 972 * above. The page might have been freed and 973 * reused for another vnode. 974 */ 975 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE || 976 m->object != object || 977 TAILQ_NEXT(m, pageq) != &marker) { 978 if (object->flags & OBJ_MIGHTBEDIRTY) 979 vnodes_skipped++; 980 goto unlock_and_continue; 981 } 982 983 /* 984 * The page may have been busied during the 985 * blocking in vget(). We don't move the 986 * page back onto the end of the queue so that 987 * statistics are more correct if we don't. 988 */ 989 if (m->busy || (m->oflags & VPO_BUSY)) { 990 goto unlock_and_continue; 991 } 992 993 /* 994 * If the page has become held it might 995 * be undergoing I/O, so skip it 996 */ 997 if (m->hold_count) { 998 vm_page_requeue(m); 999 if (object->flags & OBJ_MIGHTBEDIRTY) 1000 vnodes_skipped++; 1001 goto unlock_and_continue; 1002 } 1003 } 1004 1005 /* 1006 * If a page is dirty, then it is either being washed 1007 * (but not yet cleaned) or it is still in the 1008 * laundry. If it is still in the laundry, then we 1009 * start the cleaning operation. 1010 * 1011 * decrement page_shortage on success to account for 1012 * the (future) cleaned page. Otherwise we could wind 1013 * up laundering or cleaning too many pages. 1014 */ 1015 if (vm_pageout_clean(m) != 0) { 1016 --page_shortage; 1017 --maxlaunder; 1018 } 1019 unlock_and_continue: 1020 VM_OBJECT_UNLOCK(object); 1021 if (mp != NULL) { 1022 vm_page_unlock_queues(); 1023 if (vp != NULL) 1024 vput(vp); 1025 VFS_UNLOCK_GIANT(vfslocked); 1026 vm_object_deallocate(object); 1027 vn_finished_write(mp); 1028 vm_page_lock_queues(); 1029 } 1030 next = TAILQ_NEXT(&marker, pageq); 1031 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, 1032 &marker, pageq); 1033 continue; 1034 } 1035 VM_OBJECT_UNLOCK(object); 1036 } 1037 1038 /* 1039 * Compute the number of pages we want to try to move from the 1040 * active queue to the inactive queue. 1041 */ 1042 page_shortage = vm_paging_target() + 1043 cnt.v_inactive_target - cnt.v_inactive_count; 1044 page_shortage += addl_page_shortage; 1045 1046 /* 1047 * Scan the active queue for things we can deactivate. We nominally 1048 * track the per-page activity counter and use it to locate 1049 * deactivation candidates. 1050 */ 1051 pcount = cnt.v_active_count; 1052 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1053 1054 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) { 1055 1056 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE), 1057 ("vm_pageout_scan: page %p isn't active", m)); 1058 1059 next = TAILQ_NEXT(m, pageq); 1060 object = m->object; 1061 if ((m->flags & PG_MARKER) != 0) { 1062 m = next; 1063 continue; 1064 } 1065 if (!VM_OBJECT_TRYLOCK(object) && 1066 !vm_pageout_fallback_object_lock(m, &next)) { 1067 VM_OBJECT_UNLOCK(object); 1068 m = next; 1069 continue; 1070 } 1071 1072 /* 1073 * Don't deactivate pages that are busy. 1074 */ 1075 if ((m->busy != 0) || 1076 (m->oflags & VPO_BUSY) || 1077 (m->hold_count != 0)) { 1078 VM_OBJECT_UNLOCK(object); 1079 vm_page_requeue(m); 1080 m = next; 1081 continue; 1082 } 1083 1084 /* 1085 * The count for pagedaemon pages is done after checking the 1086 * page for eligibility... 1087 */ 1088 cnt.v_pdpages++; 1089 1090 /* 1091 * Check to see "how much" the page has been used. 1092 */ 1093 actcount = 0; 1094 if (object->ref_count != 0) { 1095 if (m->flags & PG_REFERENCED) { 1096 actcount += 1; 1097 } 1098 actcount += pmap_ts_referenced(m); 1099 if (actcount) { 1100 m->act_count += ACT_ADVANCE + actcount; 1101 if (m->act_count > ACT_MAX) 1102 m->act_count = ACT_MAX; 1103 } 1104 } 1105 1106 /* 1107 * Since we have "tested" this bit, we need to clear it now. 1108 */ 1109 vm_page_flag_clear(m, PG_REFERENCED); 1110 1111 /* 1112 * Only if an object is currently being used, do we use the 1113 * page activation count stats. 1114 */ 1115 if (actcount && (object->ref_count != 0)) { 1116 vm_page_requeue(m); 1117 } else { 1118 m->act_count -= min(m->act_count, ACT_DECLINE); 1119 if (vm_pageout_algorithm || 1120 object->ref_count == 0 || 1121 m->act_count == 0) { 1122 page_shortage--; 1123 if (object->ref_count == 0) { 1124 pmap_remove_all(m); 1125 if (m->dirty == 0) 1126 vm_page_cache(m); 1127 else 1128 vm_page_deactivate(m); 1129 } else { 1130 vm_page_deactivate(m); 1131 } 1132 } else { 1133 vm_page_requeue(m); 1134 } 1135 } 1136 VM_OBJECT_UNLOCK(object); 1137 m = next; 1138 } 1139 vm_page_unlock_queues(); 1140 #if !defined(NO_SWAPPING) 1141 /* 1142 * Idle process swapout -- run once per second. 1143 */ 1144 if (vm_swap_idle_enabled) { 1145 static long lsec; 1146 if (time_second != lsec) { 1147 vm_req_vmdaemon(VM_SWAP_IDLE); 1148 lsec = time_second; 1149 } 1150 } 1151 #endif 1152 1153 /* 1154 * If we didn't get enough free pages, and we have skipped a vnode 1155 * in a writeable object, wakeup the sync daemon. And kick swapout 1156 * if we did not get enough free pages. 1157 */ 1158 if (vm_paging_target() > 0) { 1159 if (vnodes_skipped && vm_page_count_min()) 1160 (void) speedup_syncer(); 1161 #if !defined(NO_SWAPPING) 1162 if (vm_swap_enabled && vm_page_count_target()) 1163 vm_req_vmdaemon(VM_SWAP_NORMAL); 1164 #endif 1165 } 1166 1167 /* 1168 * If we are critically low on one of RAM or swap and low on 1169 * the other, kill the largest process. However, we avoid 1170 * doing this on the first pass in order to give ourselves a 1171 * chance to flush out dirty vnode-backed pages and to allow 1172 * active pages to be moved to the inactive queue and reclaimed. 1173 */ 1174 if (pass != 0 && 1175 ((swap_pager_avail < 64 && vm_page_count_min()) || 1176 (swap_pager_full && vm_paging_target() > 0))) 1177 vm_pageout_oom(VM_OOM_MEM); 1178 } 1179 1180 1181 void 1182 vm_pageout_oom(int shortage) 1183 { 1184 struct proc *p, *bigproc; 1185 vm_offset_t size, bigsize; 1186 struct thread *td; 1187 struct vmspace *vm; 1188 1189 /* 1190 * We keep the process bigproc locked once we find it to keep anyone 1191 * from messing with it; however, there is a possibility of 1192 * deadlock if process B is bigproc and one of it's child processes 1193 * attempts to propagate a signal to B while we are waiting for A's 1194 * lock while walking this list. To avoid this, we don't block on 1195 * the process lock but just skip a process if it is already locked. 1196 */ 1197 bigproc = NULL; 1198 bigsize = 0; 1199 sx_slock(&allproc_lock); 1200 FOREACH_PROC_IN_SYSTEM(p) { 1201 int breakout; 1202 1203 if (PROC_TRYLOCK(p) == 0) 1204 continue; 1205 /* 1206 * If this is a system or protected process, skip it. 1207 */ 1208 if ((p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM)) || 1209 (p->p_pid == 1) || 1210 ((p->p_pid < 48) && (swap_pager_avail != 0))) { 1211 PROC_UNLOCK(p); 1212 continue; 1213 } 1214 /* 1215 * If the process is in a non-running type state, 1216 * don't touch it. Check all the threads individually. 1217 */ 1218 breakout = 0; 1219 FOREACH_THREAD_IN_PROC(p, td) { 1220 thread_lock(td); 1221 if (!TD_ON_RUNQ(td) && 1222 !TD_IS_RUNNING(td) && 1223 !TD_IS_SLEEPING(td)) { 1224 thread_unlock(td); 1225 breakout = 1; 1226 break; 1227 } 1228 thread_unlock(td); 1229 } 1230 if (breakout) { 1231 PROC_UNLOCK(p); 1232 continue; 1233 } 1234 /* 1235 * get the process size 1236 */ 1237 vm = vmspace_acquire_ref(p); 1238 if (vm == NULL) { 1239 PROC_UNLOCK(p); 1240 continue; 1241 } 1242 if (!vm_map_trylock_read(&vm->vm_map)) { 1243 vmspace_free(vm); 1244 PROC_UNLOCK(p); 1245 continue; 1246 } 1247 size = vmspace_swap_count(vm); 1248 vm_map_unlock_read(&vm->vm_map); 1249 if (shortage == VM_OOM_MEM) 1250 size += vmspace_resident_count(vm); 1251 vmspace_free(vm); 1252 /* 1253 * if the this process is bigger than the biggest one 1254 * remember it. 1255 */ 1256 if (size > bigsize) { 1257 if (bigproc != NULL) 1258 PROC_UNLOCK(bigproc); 1259 bigproc = p; 1260 bigsize = size; 1261 } else 1262 PROC_UNLOCK(p); 1263 } 1264 sx_sunlock(&allproc_lock); 1265 if (bigproc != NULL) { 1266 killproc(bigproc, "out of swap space"); 1267 sched_nice(bigproc, PRIO_MIN); 1268 PROC_UNLOCK(bigproc); 1269 wakeup(&cnt.v_free_count); 1270 } 1271 } 1272 1273 /* 1274 * This routine tries to maintain the pseudo LRU active queue, 1275 * so that during long periods of time where there is no paging, 1276 * that some statistic accumulation still occurs. This code 1277 * helps the situation where paging just starts to occur. 1278 */ 1279 static void 1280 vm_pageout_page_stats() 1281 { 1282 vm_object_t object; 1283 vm_page_t m,next; 1284 int pcount,tpcount; /* Number of pages to check */ 1285 static int fullintervalcount = 0; 1286 int page_shortage; 1287 1288 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1289 page_shortage = 1290 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) - 1291 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count); 1292 1293 if (page_shortage <= 0) 1294 return; 1295 1296 pcount = cnt.v_active_count; 1297 fullintervalcount += vm_pageout_stats_interval; 1298 if (fullintervalcount < vm_pageout_full_stats_interval) { 1299 tpcount = (int64_t)vm_pageout_stats_max * cnt.v_active_count / 1300 cnt.v_page_count; 1301 if (pcount > tpcount) 1302 pcount = tpcount; 1303 } else { 1304 fullintervalcount = 0; 1305 } 1306 1307 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1308 while ((m != NULL) && (pcount-- > 0)) { 1309 int actcount; 1310 1311 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE), 1312 ("vm_pageout_page_stats: page %p isn't active", m)); 1313 1314 next = TAILQ_NEXT(m, pageq); 1315 object = m->object; 1316 1317 if ((m->flags & PG_MARKER) != 0) { 1318 m = next; 1319 continue; 1320 } 1321 if (!VM_OBJECT_TRYLOCK(object) && 1322 !vm_pageout_fallback_object_lock(m, &next)) { 1323 VM_OBJECT_UNLOCK(object); 1324 m = next; 1325 continue; 1326 } 1327 1328 /* 1329 * Don't deactivate pages that are busy. 1330 */ 1331 if ((m->busy != 0) || 1332 (m->oflags & VPO_BUSY) || 1333 (m->hold_count != 0)) { 1334 VM_OBJECT_UNLOCK(object); 1335 vm_page_requeue(m); 1336 m = next; 1337 continue; 1338 } 1339 1340 actcount = 0; 1341 if (m->flags & PG_REFERENCED) { 1342 vm_page_flag_clear(m, PG_REFERENCED); 1343 actcount += 1; 1344 } 1345 1346 actcount += pmap_ts_referenced(m); 1347 if (actcount) { 1348 m->act_count += ACT_ADVANCE + actcount; 1349 if (m->act_count > ACT_MAX) 1350 m->act_count = ACT_MAX; 1351 vm_page_requeue(m); 1352 } else { 1353 if (m->act_count == 0) { 1354 /* 1355 * We turn off page access, so that we have 1356 * more accurate RSS stats. We don't do this 1357 * in the normal page deactivation when the 1358 * system is loaded VM wise, because the 1359 * cost of the large number of page protect 1360 * operations would be higher than the value 1361 * of doing the operation. 1362 */ 1363 pmap_remove_all(m); 1364 vm_page_deactivate(m); 1365 } else { 1366 m->act_count -= min(m->act_count, ACT_DECLINE); 1367 vm_page_requeue(m); 1368 } 1369 } 1370 VM_OBJECT_UNLOCK(object); 1371 m = next; 1372 } 1373 } 1374 1375 /* 1376 * vm_pageout is the high level pageout daemon. 1377 */ 1378 static void 1379 vm_pageout() 1380 { 1381 int error, pass; 1382 1383 /* 1384 * Initialize some paging parameters. 1385 */ 1386 cnt.v_interrupt_free_min = 2; 1387 if (cnt.v_page_count < 2000) 1388 vm_pageout_page_count = 8; 1389 1390 /* 1391 * v_free_reserved needs to include enough for the largest 1392 * swap pager structures plus enough for any pv_entry structs 1393 * when paging. 1394 */ 1395 if (cnt.v_page_count > 1024) 1396 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200; 1397 else 1398 cnt.v_free_min = 4; 1399 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1400 cnt.v_interrupt_free_min; 1401 cnt.v_free_reserved = vm_pageout_page_count + 1402 cnt.v_pageout_free_min + (cnt.v_page_count / 768); 1403 cnt.v_free_severe = cnt.v_free_min / 2; 1404 cnt.v_free_min += cnt.v_free_reserved; 1405 cnt.v_free_severe += cnt.v_free_reserved; 1406 1407 /* 1408 * v_free_target and v_cache_min control pageout hysteresis. Note 1409 * that these are more a measure of the VM cache queue hysteresis 1410 * then the VM free queue. Specifically, v_free_target is the 1411 * high water mark (free+cache pages). 1412 * 1413 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the 1414 * low water mark, while v_free_min is the stop. v_cache_min must 1415 * be big enough to handle memory needs while the pageout daemon 1416 * is signalled and run to free more pages. 1417 */ 1418 if (cnt.v_free_count > 6144) 1419 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved; 1420 else 1421 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved; 1422 1423 if (cnt.v_free_count > 2048) { 1424 cnt.v_cache_min = cnt.v_free_target; 1425 cnt.v_cache_max = 2 * cnt.v_cache_min; 1426 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2; 1427 } else { 1428 cnt.v_cache_min = 0; 1429 cnt.v_cache_max = 0; 1430 cnt.v_inactive_target = cnt.v_free_count / 4; 1431 } 1432 if (cnt.v_inactive_target > cnt.v_free_count / 3) 1433 cnt.v_inactive_target = cnt.v_free_count / 3; 1434 1435 /* XXX does not really belong here */ 1436 if (vm_page_max_wired == 0) 1437 vm_page_max_wired = cnt.v_free_count / 3; 1438 1439 if (vm_pageout_stats_max == 0) 1440 vm_pageout_stats_max = cnt.v_free_target; 1441 1442 /* 1443 * Set interval in seconds for stats scan. 1444 */ 1445 if (vm_pageout_stats_interval == 0) 1446 vm_pageout_stats_interval = 5; 1447 if (vm_pageout_full_stats_interval == 0) 1448 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4; 1449 1450 swap_pager_swap_init(); 1451 pass = 0; 1452 /* 1453 * The pageout daemon is never done, so loop forever. 1454 */ 1455 while (TRUE) { 1456 /* 1457 * If we have enough free memory, wakeup waiters. Do 1458 * not clear vm_pages_needed until we reach our target, 1459 * otherwise we may be woken up over and over again and 1460 * waste a lot of cpu. 1461 */ 1462 mtx_lock(&vm_page_queue_free_mtx); 1463 if (vm_pages_needed && !vm_page_count_min()) { 1464 if (!vm_paging_needed()) 1465 vm_pages_needed = 0; 1466 wakeup(&cnt.v_free_count); 1467 } 1468 if (vm_pages_needed) { 1469 /* 1470 * Still not done, take a second pass without waiting 1471 * (unlimited dirty cleaning), otherwise sleep a bit 1472 * and try again. 1473 */ 1474 ++pass; 1475 if (pass > 1) 1476 msleep(&vm_pages_needed, 1477 &vm_page_queue_free_mtx, PVM, "psleep", 1478 hz / 2); 1479 } else { 1480 /* 1481 * Good enough, sleep & handle stats. Prime the pass 1482 * for the next run. 1483 */ 1484 if (pass > 1) 1485 pass = 1; 1486 else 1487 pass = 0; 1488 error = msleep(&vm_pages_needed, 1489 &vm_page_queue_free_mtx, PVM, "psleep", 1490 vm_pageout_stats_interval * hz); 1491 if (error && !vm_pages_needed) { 1492 mtx_unlock(&vm_page_queue_free_mtx); 1493 pass = 0; 1494 vm_page_lock_queues(); 1495 vm_pageout_page_stats(); 1496 vm_page_unlock_queues(); 1497 continue; 1498 } 1499 } 1500 if (vm_pages_needed) 1501 cnt.v_pdwakeups++; 1502 mtx_unlock(&vm_page_queue_free_mtx); 1503 vm_pageout_scan(pass); 1504 } 1505 } 1506 1507 /* 1508 * Unless the free page queue lock is held by the caller, this function 1509 * should be regarded as advisory. Specifically, the caller should 1510 * not msleep() on &cnt.v_free_count following this function unless 1511 * the free page queue lock is held until the msleep() is performed. 1512 */ 1513 void 1514 pagedaemon_wakeup() 1515 { 1516 1517 if (!vm_pages_needed && curthread->td_proc != pageproc) { 1518 vm_pages_needed = 1; 1519 wakeup(&vm_pages_needed); 1520 } 1521 } 1522 1523 #if !defined(NO_SWAPPING) 1524 static void 1525 vm_req_vmdaemon(int req) 1526 { 1527 static int lastrun = 0; 1528 1529 mtx_lock(&vm_daemon_mtx); 1530 vm_pageout_req_swapout |= req; 1531 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1532 wakeup(&vm_daemon_needed); 1533 lastrun = ticks; 1534 } 1535 mtx_unlock(&vm_daemon_mtx); 1536 } 1537 1538 static void 1539 vm_daemon() 1540 { 1541 struct rlimit rsslim; 1542 struct proc *p; 1543 struct thread *td; 1544 struct vmspace *vm; 1545 int breakout, swapout_flags; 1546 1547 while (TRUE) { 1548 mtx_lock(&vm_daemon_mtx); 1549 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0); 1550 swapout_flags = vm_pageout_req_swapout; 1551 vm_pageout_req_swapout = 0; 1552 mtx_unlock(&vm_daemon_mtx); 1553 if (swapout_flags) 1554 swapout_procs(swapout_flags); 1555 1556 /* 1557 * scan the processes for exceeding their rlimits or if 1558 * process is swapped out -- deactivate pages 1559 */ 1560 sx_slock(&allproc_lock); 1561 FOREACH_PROC_IN_SYSTEM(p) { 1562 vm_pindex_t limit, size; 1563 1564 /* 1565 * if this is a system process or if we have already 1566 * looked at this process, skip it. 1567 */ 1568 PROC_LOCK(p); 1569 if (p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) { 1570 PROC_UNLOCK(p); 1571 continue; 1572 } 1573 /* 1574 * if the process is in a non-running type state, 1575 * don't touch it. 1576 */ 1577 breakout = 0; 1578 FOREACH_THREAD_IN_PROC(p, td) { 1579 thread_lock(td); 1580 if (!TD_ON_RUNQ(td) && 1581 !TD_IS_RUNNING(td) && 1582 !TD_IS_SLEEPING(td)) { 1583 thread_unlock(td); 1584 breakout = 1; 1585 break; 1586 } 1587 thread_unlock(td); 1588 } 1589 if (breakout) { 1590 PROC_UNLOCK(p); 1591 continue; 1592 } 1593 /* 1594 * get a limit 1595 */ 1596 lim_rlimit(p, RLIMIT_RSS, &rsslim); 1597 limit = OFF_TO_IDX( 1598 qmin(rsslim.rlim_cur, rsslim.rlim_max)); 1599 1600 /* 1601 * let processes that are swapped out really be 1602 * swapped out set the limit to nothing (will force a 1603 * swap-out.) 1604 */ 1605 if ((p->p_flag & P_INMEM) == 0) 1606 limit = 0; /* XXX */ 1607 vm = vmspace_acquire_ref(p); 1608 PROC_UNLOCK(p); 1609 if (vm == NULL) 1610 continue; 1611 1612 size = vmspace_resident_count(vm); 1613 if (limit >= 0 && size >= limit) { 1614 vm_pageout_map_deactivate_pages( 1615 &vm->vm_map, limit); 1616 } 1617 vmspace_free(vm); 1618 } 1619 sx_sunlock(&allproc_lock); 1620 } 1621 } 1622 #endif /* !defined(NO_SWAPPING) */ 1623