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, ""); 187 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 188 CTLFLAG_RD, &vm_swap_idle_enabled, 0, ""); 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 pmap_clear_modify(mt); 466 vm_page_undirty(mt); 467 break; 468 case VM_PAGER_ERROR: 469 case VM_PAGER_FAIL: 470 /* 471 * If page couldn't be paged out, then reactivate the 472 * page so it doesn't clog the inactive list. (We 473 * will try paging out it again later). 474 */ 475 vm_page_activate(mt); 476 break; 477 case VM_PAGER_AGAIN: 478 break; 479 } 480 481 /* 482 * If the operation is still going, leave the page busy to 483 * block all other accesses. Also, leave the paging in 484 * progress indicator set so that we don't attempt an object 485 * collapse. 486 */ 487 if (pageout_status[i] != VM_PAGER_PEND) { 488 vm_object_pip_wakeup(object); 489 vm_page_io_finish(mt); 490 if (vm_page_count_severe()) 491 vm_page_try_to_cache(mt); 492 } 493 } 494 return numpagedout; 495 } 496 497 #if !defined(NO_SWAPPING) 498 /* 499 * vm_pageout_object_deactivate_pages 500 * 501 * deactivate enough pages to satisfy the inactive target 502 * requirements or if vm_page_proc_limit is set, then 503 * deactivate all of the pages in the object and its 504 * backing_objects. 505 * 506 * The object and map must be locked. 507 */ 508 static void 509 vm_pageout_object_deactivate_pages(pmap, first_object, desired) 510 pmap_t pmap; 511 vm_object_t first_object; 512 long desired; 513 { 514 vm_object_t backing_object, object; 515 vm_page_t p, next; 516 int actcount, rcount, remove_mode; 517 518 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED); 519 if (first_object->type == OBJT_DEVICE || first_object->type == OBJT_PHYS) 520 return; 521 for (object = first_object;; object = backing_object) { 522 if (pmap_resident_count(pmap) <= desired) 523 goto unlock_return; 524 if (object->paging_in_progress) 525 goto unlock_return; 526 527 remove_mode = 0; 528 if (object->shadow_count > 1) 529 remove_mode = 1; 530 /* 531 * scan the objects entire memory queue 532 */ 533 rcount = object->resident_page_count; 534 p = TAILQ_FIRST(&object->memq); 535 vm_page_lock_queues(); 536 while (p && (rcount-- > 0)) { 537 if (pmap_resident_count(pmap) <= desired) { 538 vm_page_unlock_queues(); 539 goto unlock_return; 540 } 541 next = TAILQ_NEXT(p, listq); 542 cnt.v_pdpages++; 543 if (p->wire_count != 0 || 544 p->hold_count != 0 || 545 p->busy != 0 || 546 (p->oflags & VPO_BUSY) || 547 (p->flags & PG_UNMANAGED) || 548 !pmap_page_exists_quick(pmap, p)) { 549 p = next; 550 continue; 551 } 552 actcount = pmap_ts_referenced(p); 553 if (actcount) { 554 vm_page_flag_set(p, PG_REFERENCED); 555 } else if (p->flags & PG_REFERENCED) { 556 actcount = 1; 557 } 558 if ((p->queue != PQ_ACTIVE) && 559 (p->flags & PG_REFERENCED)) { 560 vm_page_activate(p); 561 p->act_count += actcount; 562 vm_page_flag_clear(p, PG_REFERENCED); 563 } else if (p->queue == PQ_ACTIVE) { 564 if ((p->flags & PG_REFERENCED) == 0) { 565 p->act_count -= min(p->act_count, ACT_DECLINE); 566 if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) { 567 pmap_remove_all(p); 568 vm_page_deactivate(p); 569 } else { 570 vm_page_requeue(p); 571 } 572 } else { 573 vm_page_activate(p); 574 vm_page_flag_clear(p, PG_REFERENCED); 575 if (p->act_count < (ACT_MAX - ACT_ADVANCE)) 576 p->act_count += ACT_ADVANCE; 577 vm_page_requeue(p); 578 } 579 } else if (p->queue == PQ_INACTIVE) { 580 pmap_remove_all(p); 581 } 582 p = next; 583 } 584 vm_page_unlock_queues(); 585 if ((backing_object = object->backing_object) == NULL) 586 goto unlock_return; 587 VM_OBJECT_LOCK(backing_object); 588 if (object != first_object) 589 VM_OBJECT_UNLOCK(object); 590 } 591 unlock_return: 592 if (object != first_object) 593 VM_OBJECT_UNLOCK(object); 594 } 595 596 /* 597 * deactivate some number of pages in a map, try to do it fairly, but 598 * that is really hard to do. 599 */ 600 static void 601 vm_pageout_map_deactivate_pages(map, desired) 602 vm_map_t map; 603 long desired; 604 { 605 vm_map_entry_t tmpe; 606 vm_object_t obj, bigobj; 607 int nothingwired; 608 609 if (!vm_map_trylock(map)) 610 return; 611 612 bigobj = NULL; 613 nothingwired = TRUE; 614 615 /* 616 * first, search out the biggest object, and try to free pages from 617 * that. 618 */ 619 tmpe = map->header.next; 620 while (tmpe != &map->header) { 621 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 622 obj = tmpe->object.vm_object; 623 if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) { 624 if (obj->shadow_count <= 1 && 625 (bigobj == NULL || 626 bigobj->resident_page_count < obj->resident_page_count)) { 627 if (bigobj != NULL) 628 VM_OBJECT_UNLOCK(bigobj); 629 bigobj = obj; 630 } else 631 VM_OBJECT_UNLOCK(obj); 632 } 633 } 634 if (tmpe->wired_count > 0) 635 nothingwired = FALSE; 636 tmpe = tmpe->next; 637 } 638 639 if (bigobj != NULL) { 640 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired); 641 VM_OBJECT_UNLOCK(bigobj); 642 } 643 /* 644 * Next, hunt around for other pages to deactivate. We actually 645 * do this search sort of wrong -- .text first is not the best idea. 646 */ 647 tmpe = map->header.next; 648 while (tmpe != &map->header) { 649 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 650 break; 651 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 652 obj = tmpe->object.vm_object; 653 if (obj != NULL) { 654 VM_OBJECT_LOCK(obj); 655 vm_pageout_object_deactivate_pages(map->pmap, obj, desired); 656 VM_OBJECT_UNLOCK(obj); 657 } 658 } 659 tmpe = tmpe->next; 660 } 661 662 /* 663 * Remove all mappings if a process is swapped out, this will free page 664 * table pages. 665 */ 666 if (desired == 0 && nothingwired) { 667 pmap_remove(vm_map_pmap(map), vm_map_min(map), 668 vm_map_max(map)); 669 } 670 vm_map_unlock(map); 671 } 672 #endif /* !defined(NO_SWAPPING) */ 673 674 /* 675 * vm_pageout_scan does the dirty work for the pageout daemon. 676 */ 677 static void 678 vm_pageout_scan(int pass) 679 { 680 vm_page_t m, next; 681 struct vm_page marker; 682 int page_shortage, maxscan, pcount; 683 int addl_page_shortage, addl_page_shortage_init; 684 struct proc *p, *bigproc; 685 struct thread *td; 686 vm_offset_t size, bigsize; 687 vm_object_t object; 688 int actcount; 689 int vnodes_skipped = 0; 690 int maxlaunder; 691 692 /* 693 * Decrease registered cache sizes. 694 */ 695 EVENTHANDLER_INVOKE(vm_lowmem, 0); 696 /* 697 * We do this explicitly after the caches have been drained above. 698 */ 699 uma_reclaim(); 700 701 addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit); 702 703 /* 704 * Calculate the number of pages we want to either free or move 705 * to the cache. 706 */ 707 page_shortage = vm_paging_target() + addl_page_shortage_init; 708 709 /* 710 * Initialize our marker 711 */ 712 bzero(&marker, sizeof(marker)); 713 marker.flags = PG_FICTITIOUS | PG_MARKER; 714 marker.oflags = VPO_BUSY; 715 marker.queue = PQ_INACTIVE; 716 marker.wire_count = 1; 717 718 /* 719 * Start scanning the inactive queue for pages we can move to the 720 * cache or free. The scan will stop when the target is reached or 721 * we have scanned the entire inactive queue. Note that m->act_count 722 * is not used to form decisions for the inactive queue, only for the 723 * active queue. 724 * 725 * maxlaunder limits the number of dirty pages we flush per scan. 726 * For most systems a smaller value (16 or 32) is more robust under 727 * extreme memory and disk pressure because any unnecessary writes 728 * to disk can result in extreme performance degredation. However, 729 * systems with excessive dirty pages (especially when MAP_NOSYNC is 730 * used) will die horribly with limited laundering. If the pageout 731 * daemon cannot clean enough pages in the first pass, we let it go 732 * all out in succeeding passes. 733 */ 734 if ((maxlaunder = vm_max_launder) <= 1) 735 maxlaunder = 1; 736 if (pass) 737 maxlaunder = 10000; 738 vm_page_lock_queues(); 739 rescan0: 740 addl_page_shortage = addl_page_shortage_init; 741 maxscan = cnt.v_inactive_count; 742 743 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl); 744 m != NULL && maxscan-- > 0 && page_shortage > 0; 745 m = next) { 746 747 cnt.v_pdpages++; 748 749 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE) { 750 goto rescan0; 751 } 752 753 next = TAILQ_NEXT(m, pageq); 754 object = m->object; 755 756 /* 757 * skip marker pages 758 */ 759 if (m->flags & PG_MARKER) 760 continue; 761 762 /* 763 * A held page may be undergoing I/O, so skip it. 764 */ 765 if (m->hold_count) { 766 vm_page_requeue(m); 767 addl_page_shortage++; 768 continue; 769 } 770 /* 771 * Don't mess with busy pages, keep in the front of the 772 * queue, most likely are being paged out. 773 */ 774 if (!VM_OBJECT_TRYLOCK(object) && 775 (!vm_pageout_fallback_object_lock(m, &next) || 776 m->hold_count != 0)) { 777 VM_OBJECT_UNLOCK(object); 778 addl_page_shortage++; 779 continue; 780 } 781 if (m->busy || (m->oflags & VPO_BUSY)) { 782 VM_OBJECT_UNLOCK(object); 783 addl_page_shortage++; 784 continue; 785 } 786 787 /* 788 * If the object is not being used, we ignore previous 789 * references. 790 */ 791 if (object->ref_count == 0) { 792 vm_page_flag_clear(m, PG_REFERENCED); 793 pmap_clear_reference(m); 794 795 /* 796 * Otherwise, if the page has been referenced while in the 797 * inactive queue, we bump the "activation count" upwards, 798 * making it less likely that the page will be added back to 799 * the inactive queue prematurely again. Here we check the 800 * page tables (or emulated bits, if any), given the upper 801 * level VM system not knowing anything about existing 802 * references. 803 */ 804 } else if (((m->flags & PG_REFERENCED) == 0) && 805 (actcount = pmap_ts_referenced(m))) { 806 vm_page_activate(m); 807 VM_OBJECT_UNLOCK(object); 808 m->act_count += (actcount + ACT_ADVANCE); 809 continue; 810 } 811 812 /* 813 * If the upper level VM system knows about any page 814 * references, we activate the page. We also set the 815 * "activation count" higher than normal so that we will less 816 * likely place pages back onto the inactive queue again. 817 */ 818 if ((m->flags & PG_REFERENCED) != 0) { 819 vm_page_flag_clear(m, PG_REFERENCED); 820 actcount = pmap_ts_referenced(m); 821 vm_page_activate(m); 822 VM_OBJECT_UNLOCK(object); 823 m->act_count += (actcount + ACT_ADVANCE + 1); 824 continue; 825 } 826 827 /* 828 * If the upper level VM system doesn't know anything about 829 * the page being dirty, we have to check for it again. As 830 * far as the VM code knows, any partially dirty pages are 831 * fully dirty. 832 */ 833 if (m->dirty == 0 && !pmap_is_modified(m)) { 834 /* 835 * Avoid a race condition: Unless write access is 836 * removed from the page, another processor could 837 * modify it before all access is removed by the call 838 * to vm_page_cache() below. If vm_page_cache() finds 839 * that the page has been modified when it removes all 840 * access, it panics because it cannot cache dirty 841 * pages. In principle, we could eliminate just write 842 * access here rather than all access. In the expected 843 * case, when there are no last instant modifications 844 * to the page, removing all access will be cheaper 845 * overall. 846 */ 847 if ((m->flags & PG_WRITEABLE) != 0) 848 pmap_remove_all(m); 849 } else { 850 vm_page_dirty(m); 851 } 852 853 if (m->valid == 0) { 854 /* 855 * Invalid pages can be easily freed 856 */ 857 vm_page_free(m); 858 cnt.v_dfree++; 859 --page_shortage; 860 } else if (m->dirty == 0) { 861 /* 862 * Clean pages can be placed onto the cache queue. 863 * This effectively frees them. 864 */ 865 vm_page_cache(m); 866 --page_shortage; 867 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) { 868 /* 869 * Dirty pages need to be paged out, but flushing 870 * a page is extremely expensive verses freeing 871 * a clean page. Rather then artificially limiting 872 * the number of pages we can flush, we instead give 873 * dirty pages extra priority on the inactive queue 874 * by forcing them to be cycled through the queue 875 * twice before being flushed, after which the 876 * (now clean) page will cycle through once more 877 * before being freed. This significantly extends 878 * the thrash point for a heavily loaded machine. 879 */ 880 vm_page_flag_set(m, PG_WINATCFLS); 881 vm_page_requeue(m); 882 } else if (maxlaunder > 0) { 883 /* 884 * We always want to try to flush some dirty pages if 885 * we encounter them, to keep the system stable. 886 * Normally this number is small, but under extreme 887 * pressure where there are insufficient clean pages 888 * on the inactive queue, we may have to go all out. 889 */ 890 int swap_pageouts_ok, vfslocked = 0; 891 struct vnode *vp = NULL; 892 struct mount *mp = NULL; 893 894 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 895 swap_pageouts_ok = 1; 896 } else { 897 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 898 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 899 vm_page_count_min()); 900 901 } 902 903 /* 904 * We don't bother paging objects that are "dead". 905 * Those objects are in a "rundown" state. 906 */ 907 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 908 VM_OBJECT_UNLOCK(object); 909 vm_page_requeue(m); 910 continue; 911 } 912 913 /* 914 * Following operations may unlock 915 * vm_page_queue_mtx, invalidating the 'next' 916 * pointer. To prevent an inordinate number 917 * of restarts we use our marker to remember 918 * our place. 919 * 920 */ 921 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, 922 m, &marker, pageq); 923 /* 924 * The object is already known NOT to be dead. It 925 * is possible for the vget() to block the whole 926 * pageout daemon, but the new low-memory handling 927 * code should prevent it. 928 * 929 * The previous code skipped locked vnodes and, worse, 930 * reordered pages in the queue. This results in 931 * completely non-deterministic operation and, on a 932 * busy system, can lead to extremely non-optimal 933 * pageouts. For example, it can cause clean pages 934 * to be freed and dirty pages to be moved to the end 935 * of the queue. Since dirty pages are also moved to 936 * the end of the queue once-cleaned, this gives 937 * way too large a weighting to defering the freeing 938 * of dirty pages. 939 * 940 * We can't wait forever for the vnode lock, we might 941 * deadlock due to a vn_read() getting stuck in 942 * vm_wait while holding this vnode. We skip the 943 * vnode if we can't get it in a reasonable amount 944 * of time. 945 */ 946 if (object->type == OBJT_VNODE) { 947 vp = object->handle; 948 if (vp->v_type == VREG && 949 vn_start_write(vp, &mp, V_NOWAIT) != 0) { 950 KASSERT(mp == NULL, 951 ("vm_pageout_scan: mp != NULL")); 952 ++pageout_lock_miss; 953 if (object->flags & OBJ_MIGHTBEDIRTY) 954 vnodes_skipped++; 955 goto unlock_and_continue; 956 } 957 vm_page_unlock_queues(); 958 vm_object_reference_locked(object); 959 VM_OBJECT_UNLOCK(object); 960 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 961 if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK, 962 curthread)) { 963 VM_OBJECT_LOCK(object); 964 vm_page_lock_queues(); 965 ++pageout_lock_miss; 966 if (object->flags & OBJ_MIGHTBEDIRTY) 967 vnodes_skipped++; 968 vp = NULL; 969 goto unlock_and_continue; 970 } 971 VM_OBJECT_LOCK(object); 972 vm_page_lock_queues(); 973 /* 974 * The page might have been moved to another 975 * queue during potential blocking in vget() 976 * above. The page might have been freed and 977 * reused for another vnode. 978 */ 979 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE || 980 m->object != object || 981 TAILQ_NEXT(m, pageq) != &marker) { 982 if (object->flags & OBJ_MIGHTBEDIRTY) 983 vnodes_skipped++; 984 goto unlock_and_continue; 985 } 986 987 /* 988 * The page may have been busied during the 989 * blocking in vget(). We don't move the 990 * page back onto the end of the queue so that 991 * statistics are more correct if we don't. 992 */ 993 if (m->busy || (m->oflags & VPO_BUSY)) { 994 goto unlock_and_continue; 995 } 996 997 /* 998 * If the page has become held it might 999 * be undergoing I/O, so skip it 1000 */ 1001 if (m->hold_count) { 1002 vm_page_requeue(m); 1003 if (object->flags & OBJ_MIGHTBEDIRTY) 1004 vnodes_skipped++; 1005 goto unlock_and_continue; 1006 } 1007 } 1008 1009 /* 1010 * If a page is dirty, then it is either being washed 1011 * (but not yet cleaned) or it is still in the 1012 * laundry. If it is still in the laundry, then we 1013 * start the cleaning operation. 1014 * 1015 * decrement page_shortage on success to account for 1016 * the (future) cleaned page. Otherwise we could wind 1017 * up laundering or cleaning too many pages. 1018 */ 1019 if (vm_pageout_clean(m) != 0) { 1020 --page_shortage; 1021 --maxlaunder; 1022 } 1023 unlock_and_continue: 1024 VM_OBJECT_UNLOCK(object); 1025 if (mp != NULL) { 1026 vm_page_unlock_queues(); 1027 if (vp != NULL) 1028 vput(vp); 1029 VFS_UNLOCK_GIANT(vfslocked); 1030 vm_object_deallocate(object); 1031 vn_finished_write(mp); 1032 vm_page_lock_queues(); 1033 } 1034 next = TAILQ_NEXT(&marker, pageq); 1035 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, 1036 &marker, pageq); 1037 continue; 1038 } 1039 VM_OBJECT_UNLOCK(object); 1040 } 1041 1042 /* 1043 * Compute the number of pages we want to try to move from the 1044 * active queue to the inactive queue. 1045 */ 1046 page_shortage = vm_paging_target() + 1047 cnt.v_inactive_target - cnt.v_inactive_count; 1048 page_shortage += addl_page_shortage; 1049 1050 /* 1051 * Scan the active queue for things we can deactivate. We nominally 1052 * track the per-page activity counter and use it to locate 1053 * deactivation candidates. 1054 */ 1055 pcount = cnt.v_active_count; 1056 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1057 1058 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) { 1059 1060 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE), 1061 ("vm_pageout_scan: page %p isn't active", m)); 1062 1063 next = TAILQ_NEXT(m, pageq); 1064 object = m->object; 1065 if ((m->flags & PG_MARKER) != 0) { 1066 m = next; 1067 continue; 1068 } 1069 if (!VM_OBJECT_TRYLOCK(object) && 1070 !vm_pageout_fallback_object_lock(m, &next)) { 1071 VM_OBJECT_UNLOCK(object); 1072 m = next; 1073 continue; 1074 } 1075 1076 /* 1077 * Don't deactivate pages that are busy. 1078 */ 1079 if ((m->busy != 0) || 1080 (m->oflags & VPO_BUSY) || 1081 (m->hold_count != 0)) { 1082 VM_OBJECT_UNLOCK(object); 1083 vm_page_requeue(m); 1084 m = next; 1085 continue; 1086 } 1087 1088 /* 1089 * The count for pagedaemon pages is done after checking the 1090 * page for eligibility... 1091 */ 1092 cnt.v_pdpages++; 1093 1094 /* 1095 * Check to see "how much" the page has been used. 1096 */ 1097 actcount = 0; 1098 if (object->ref_count != 0) { 1099 if (m->flags & PG_REFERENCED) { 1100 actcount += 1; 1101 } 1102 actcount += pmap_ts_referenced(m); 1103 if (actcount) { 1104 m->act_count += ACT_ADVANCE + actcount; 1105 if (m->act_count > ACT_MAX) 1106 m->act_count = ACT_MAX; 1107 } 1108 } 1109 1110 /* 1111 * Since we have "tested" this bit, we need to clear it now. 1112 */ 1113 vm_page_flag_clear(m, PG_REFERENCED); 1114 1115 /* 1116 * Only if an object is currently being used, do we use the 1117 * page activation count stats. 1118 */ 1119 if (actcount && (object->ref_count != 0)) { 1120 vm_page_requeue(m); 1121 } else { 1122 m->act_count -= min(m->act_count, ACT_DECLINE); 1123 if (vm_pageout_algorithm || 1124 object->ref_count == 0 || 1125 m->act_count == 0) { 1126 page_shortage--; 1127 if (object->ref_count == 0) { 1128 pmap_remove_all(m); 1129 if (m->dirty == 0) 1130 vm_page_cache(m); 1131 else 1132 vm_page_deactivate(m); 1133 } else { 1134 vm_page_deactivate(m); 1135 } 1136 } else { 1137 vm_page_requeue(m); 1138 } 1139 } 1140 VM_OBJECT_UNLOCK(object); 1141 m = next; 1142 } 1143 vm_page_unlock_queues(); 1144 #if !defined(NO_SWAPPING) 1145 /* 1146 * Idle process swapout -- run once per second. 1147 */ 1148 if (vm_swap_idle_enabled) { 1149 static long lsec; 1150 if (time_second != lsec) { 1151 vm_req_vmdaemon(VM_SWAP_IDLE); 1152 lsec = time_second; 1153 } 1154 } 1155 #endif 1156 1157 /* 1158 * If we didn't get enough free pages, and we have skipped a vnode 1159 * in a writeable object, wakeup the sync daemon. And kick swapout 1160 * if we did not get enough free pages. 1161 */ 1162 if (vm_paging_target() > 0) { 1163 if (vnodes_skipped && vm_page_count_min()) 1164 (void) speedup_syncer(); 1165 #if !defined(NO_SWAPPING) 1166 if (vm_swap_enabled && vm_page_count_target()) 1167 vm_req_vmdaemon(VM_SWAP_NORMAL); 1168 #endif 1169 } 1170 1171 /* 1172 * If we are critically low on one of RAM or swap and low on 1173 * the other, kill the largest process. However, we avoid 1174 * doing this on the first pass in order to give ourselves a 1175 * chance to flush out dirty vnode-backed pages and to allow 1176 * active pages to be moved to the inactive queue and reclaimed. 1177 * 1178 * We keep the process bigproc locked once we find it to keep anyone 1179 * from messing with it; however, there is a possibility of 1180 * deadlock if process B is bigproc and one of it's child processes 1181 * attempts to propagate a signal to B while we are waiting for A's 1182 * lock while walking this list. To avoid this, we don't block on 1183 * the process lock but just skip a process if it is already locked. 1184 */ 1185 if (pass != 0 && 1186 ((swap_pager_avail < 64 && vm_page_count_min()) || 1187 (swap_pager_full && vm_paging_target() > 0))) { 1188 bigproc = NULL; 1189 bigsize = 0; 1190 sx_slock(&allproc_lock); 1191 FOREACH_PROC_IN_SYSTEM(p) { 1192 int breakout; 1193 1194 if (PROC_TRYLOCK(p) == 0) 1195 continue; 1196 /* 1197 * If this is a system or protected process, skip it. 1198 */ 1199 if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) || 1200 (p->p_flag & P_PROTECTED) || 1201 ((p->p_pid < 48) && (swap_pager_avail != 0))) { 1202 PROC_UNLOCK(p); 1203 continue; 1204 } 1205 /* 1206 * If the process is in a non-running type state, 1207 * don't touch it. Check all the threads individually. 1208 */ 1209 breakout = 0; 1210 FOREACH_THREAD_IN_PROC(p, td) { 1211 thread_lock(td); 1212 if (!TD_ON_RUNQ(td) && 1213 !TD_IS_RUNNING(td) && 1214 !TD_IS_SLEEPING(td)) { 1215 thread_unlock(td); 1216 breakout = 1; 1217 break; 1218 } 1219 thread_unlock(td); 1220 } 1221 if (breakout) { 1222 PROC_UNLOCK(p); 1223 continue; 1224 } 1225 /* 1226 * get the process size 1227 */ 1228 if (!vm_map_trylock_read(&p->p_vmspace->vm_map)) { 1229 PROC_UNLOCK(p); 1230 continue; 1231 } 1232 size = vmspace_swap_count(p->p_vmspace); 1233 vm_map_unlock_read(&p->p_vmspace->vm_map); 1234 size += vmspace_resident_count(p->p_vmspace); 1235 /* 1236 * if the this process is bigger than the biggest one 1237 * remember it. 1238 */ 1239 if (size > bigsize) { 1240 if (bigproc != NULL) 1241 PROC_UNLOCK(bigproc); 1242 bigproc = p; 1243 bigsize = size; 1244 } else 1245 PROC_UNLOCK(p); 1246 } 1247 sx_sunlock(&allproc_lock); 1248 if (bigproc != NULL) { 1249 killproc(bigproc, "out of swap space"); 1250 sched_nice(bigproc, PRIO_MIN); 1251 PROC_UNLOCK(bigproc); 1252 wakeup(&cnt.v_free_count); 1253 } 1254 } 1255 } 1256 1257 /* 1258 * This routine tries to maintain the pseudo LRU active queue, 1259 * so that during long periods of time where there is no paging, 1260 * that some statistic accumulation still occurs. This code 1261 * helps the situation where paging just starts to occur. 1262 */ 1263 static void 1264 vm_pageout_page_stats() 1265 { 1266 vm_object_t object; 1267 vm_page_t m,next; 1268 int pcount,tpcount; /* Number of pages to check */ 1269 static int fullintervalcount = 0; 1270 int page_shortage; 1271 1272 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1273 page_shortage = 1274 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) - 1275 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count); 1276 1277 if (page_shortage <= 0) 1278 return; 1279 1280 pcount = cnt.v_active_count; 1281 fullintervalcount += vm_pageout_stats_interval; 1282 if (fullintervalcount < vm_pageout_full_stats_interval) { 1283 tpcount = (vm_pageout_stats_max * cnt.v_active_count) / cnt.v_page_count; 1284 if (pcount > tpcount) 1285 pcount = tpcount; 1286 } else { 1287 fullintervalcount = 0; 1288 } 1289 1290 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1291 while ((m != NULL) && (pcount-- > 0)) { 1292 int actcount; 1293 1294 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE), 1295 ("vm_pageout_page_stats: page %p isn't active", m)); 1296 1297 next = TAILQ_NEXT(m, pageq); 1298 object = m->object; 1299 1300 if ((m->flags & PG_MARKER) != 0) { 1301 m = next; 1302 continue; 1303 } 1304 if (!VM_OBJECT_TRYLOCK(object) && 1305 !vm_pageout_fallback_object_lock(m, &next)) { 1306 VM_OBJECT_UNLOCK(object); 1307 m = next; 1308 continue; 1309 } 1310 1311 /* 1312 * Don't deactivate pages that are busy. 1313 */ 1314 if ((m->busy != 0) || 1315 (m->oflags & VPO_BUSY) || 1316 (m->hold_count != 0)) { 1317 VM_OBJECT_UNLOCK(object); 1318 vm_page_requeue(m); 1319 m = next; 1320 continue; 1321 } 1322 1323 actcount = 0; 1324 if (m->flags & PG_REFERENCED) { 1325 vm_page_flag_clear(m, PG_REFERENCED); 1326 actcount += 1; 1327 } 1328 1329 actcount += pmap_ts_referenced(m); 1330 if (actcount) { 1331 m->act_count += ACT_ADVANCE + actcount; 1332 if (m->act_count > ACT_MAX) 1333 m->act_count = ACT_MAX; 1334 vm_page_requeue(m); 1335 } else { 1336 if (m->act_count == 0) { 1337 /* 1338 * We turn off page access, so that we have 1339 * more accurate RSS stats. We don't do this 1340 * in the normal page deactivation when the 1341 * system is loaded VM wise, because the 1342 * cost of the large number of page protect 1343 * operations would be higher than the value 1344 * of doing the operation. 1345 */ 1346 pmap_remove_all(m); 1347 vm_page_deactivate(m); 1348 } else { 1349 m->act_count -= min(m->act_count, ACT_DECLINE); 1350 vm_page_requeue(m); 1351 } 1352 } 1353 VM_OBJECT_UNLOCK(object); 1354 m = next; 1355 } 1356 } 1357 1358 /* 1359 * vm_pageout is the high level pageout daemon. 1360 */ 1361 static void 1362 vm_pageout() 1363 { 1364 int error, pass; 1365 1366 /* 1367 * Initialize some paging parameters. 1368 */ 1369 cnt.v_interrupt_free_min = 2; 1370 if (cnt.v_page_count < 2000) 1371 vm_pageout_page_count = 8; 1372 1373 /* 1374 * v_free_reserved needs to include enough for the largest 1375 * swap pager structures plus enough for any pv_entry structs 1376 * when paging. 1377 */ 1378 if (cnt.v_page_count > 1024) 1379 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200; 1380 else 1381 cnt.v_free_min = 4; 1382 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1383 cnt.v_interrupt_free_min; 1384 cnt.v_free_reserved = vm_pageout_page_count + 1385 cnt.v_pageout_free_min + (cnt.v_page_count / 768); 1386 cnt.v_free_severe = cnt.v_free_min / 2; 1387 cnt.v_free_min += cnt.v_free_reserved; 1388 cnt.v_free_severe += cnt.v_free_reserved; 1389 1390 /* 1391 * v_free_target and v_cache_min control pageout hysteresis. Note 1392 * that these are more a measure of the VM cache queue hysteresis 1393 * then the VM free queue. Specifically, v_free_target is the 1394 * high water mark (free+cache pages). 1395 * 1396 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the 1397 * low water mark, while v_free_min is the stop. v_cache_min must 1398 * be big enough to handle memory needs while the pageout daemon 1399 * is signalled and run to free more pages. 1400 */ 1401 if (cnt.v_free_count > 6144) 1402 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved; 1403 else 1404 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved; 1405 1406 if (cnt.v_free_count > 2048) { 1407 cnt.v_cache_min = cnt.v_free_target; 1408 cnt.v_cache_max = 2 * cnt.v_cache_min; 1409 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2; 1410 } else { 1411 cnt.v_cache_min = 0; 1412 cnt.v_cache_max = 0; 1413 cnt.v_inactive_target = cnt.v_free_count / 4; 1414 } 1415 if (cnt.v_inactive_target > cnt.v_free_count / 3) 1416 cnt.v_inactive_target = cnt.v_free_count / 3; 1417 1418 /* XXX does not really belong here */ 1419 if (vm_page_max_wired == 0) 1420 vm_page_max_wired = cnt.v_free_count / 3; 1421 1422 if (vm_pageout_stats_max == 0) 1423 vm_pageout_stats_max = cnt.v_free_target; 1424 1425 /* 1426 * Set interval in seconds for stats scan. 1427 */ 1428 if (vm_pageout_stats_interval == 0) 1429 vm_pageout_stats_interval = 5; 1430 if (vm_pageout_full_stats_interval == 0) 1431 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4; 1432 1433 swap_pager_swap_init(); 1434 pass = 0; 1435 /* 1436 * The pageout daemon is never done, so loop forever. 1437 */ 1438 while (TRUE) { 1439 /* 1440 * If we have enough free memory, wakeup waiters. Do 1441 * not clear vm_pages_needed until we reach our target, 1442 * otherwise we may be woken up over and over again and 1443 * waste a lot of cpu. 1444 */ 1445 mtx_lock(&vm_page_queue_free_mtx); 1446 if (vm_pages_needed && !vm_page_count_min()) { 1447 if (!vm_paging_needed()) 1448 vm_pages_needed = 0; 1449 wakeup(&cnt.v_free_count); 1450 } 1451 if (vm_pages_needed) { 1452 /* 1453 * Still not done, take a second pass without waiting 1454 * (unlimited dirty cleaning), otherwise sleep a bit 1455 * and try again. 1456 */ 1457 ++pass; 1458 if (pass > 1) 1459 msleep(&vm_pages_needed, 1460 &vm_page_queue_free_mtx, PVM, "psleep", 1461 hz / 2); 1462 } else { 1463 /* 1464 * Good enough, sleep & handle stats. Prime the pass 1465 * for the next run. 1466 */ 1467 if (pass > 1) 1468 pass = 1; 1469 else 1470 pass = 0; 1471 error = msleep(&vm_pages_needed, 1472 &vm_page_queue_free_mtx, PVM, "psleep", 1473 vm_pageout_stats_interval * hz); 1474 if (error && !vm_pages_needed) { 1475 mtx_unlock(&vm_page_queue_free_mtx); 1476 pass = 0; 1477 vm_page_lock_queues(); 1478 vm_pageout_page_stats(); 1479 vm_page_unlock_queues(); 1480 continue; 1481 } 1482 } 1483 if (vm_pages_needed) 1484 cnt.v_pdwakeups++; 1485 mtx_unlock(&vm_page_queue_free_mtx); 1486 vm_pageout_scan(pass); 1487 } 1488 } 1489 1490 /* 1491 * Unless the free page queue lock is held by the caller, this function 1492 * should be regarded as advisory. Specifically, the caller should 1493 * not msleep() on &cnt.v_free_count following this function unless 1494 * the free page queue lock is held until the msleep() is performed. 1495 */ 1496 void 1497 pagedaemon_wakeup() 1498 { 1499 1500 if (!vm_pages_needed && curthread->td_proc != pageproc) { 1501 vm_pages_needed = 1; 1502 wakeup(&vm_pages_needed); 1503 } 1504 } 1505 1506 #if !defined(NO_SWAPPING) 1507 static void 1508 vm_req_vmdaemon(int req) 1509 { 1510 static int lastrun = 0; 1511 1512 mtx_lock(&vm_daemon_mtx); 1513 vm_pageout_req_swapout |= req; 1514 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1515 wakeup(&vm_daemon_needed); 1516 lastrun = ticks; 1517 } 1518 mtx_unlock(&vm_daemon_mtx); 1519 } 1520 1521 static void 1522 vm_daemon() 1523 { 1524 struct rlimit rsslim; 1525 struct proc *p; 1526 struct thread *td; 1527 int breakout, swapout_flags; 1528 1529 while (TRUE) { 1530 mtx_lock(&vm_daemon_mtx); 1531 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0); 1532 swapout_flags = vm_pageout_req_swapout; 1533 vm_pageout_req_swapout = 0; 1534 mtx_unlock(&vm_daemon_mtx); 1535 if (swapout_flags) 1536 swapout_procs(swapout_flags); 1537 1538 /* 1539 * scan the processes for exceeding their rlimits or if 1540 * process is swapped out -- deactivate pages 1541 */ 1542 sx_slock(&allproc_lock); 1543 FOREACH_PROC_IN_SYSTEM(p) { 1544 vm_pindex_t limit, size; 1545 1546 /* 1547 * if this is a system process or if we have already 1548 * looked at this process, skip it. 1549 */ 1550 PROC_LOCK(p); 1551 if (p->p_flag & (P_SYSTEM | P_WEXIT)) { 1552 PROC_UNLOCK(p); 1553 continue; 1554 } 1555 /* 1556 * if the process is in a non-running type state, 1557 * don't touch it. 1558 */ 1559 breakout = 0; 1560 FOREACH_THREAD_IN_PROC(p, td) { 1561 thread_lock(td); 1562 if (!TD_ON_RUNQ(td) && 1563 !TD_IS_RUNNING(td) && 1564 !TD_IS_SLEEPING(td)) { 1565 thread_unlock(td); 1566 breakout = 1; 1567 break; 1568 } 1569 thread_unlock(td); 1570 } 1571 if (breakout) { 1572 PROC_UNLOCK(p); 1573 continue; 1574 } 1575 /* 1576 * get a limit 1577 */ 1578 lim_rlimit(p, RLIMIT_RSS, &rsslim); 1579 limit = OFF_TO_IDX( 1580 qmin(rsslim.rlim_cur, rsslim.rlim_max)); 1581 1582 /* 1583 * let processes that are swapped out really be 1584 * swapped out set the limit to nothing (will force a 1585 * swap-out.) 1586 */ 1587 if ((p->p_flag & P_INMEM) == 0) 1588 limit = 0; /* XXX */ 1589 PROC_UNLOCK(p); 1590 1591 size = vmspace_resident_count(p->p_vmspace); 1592 if (limit >= 0 && size >= limit) { 1593 vm_pageout_map_deactivate_pages( 1594 &p->p_vmspace->vm_map, limit); 1595 } 1596 } 1597 sx_sunlock(&allproc_lock); 1598 } 1599 } 1600 #endif /* !defined(NO_SWAPPING) */ 1601