vm_pageout.c revision 115146
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 * 9 * This code is derived from software contributed to Berkeley by 10 * The Mach Operating System project at Carnegie-Mellon University. 11 * 12 * Redistribution and use in source and binary forms, with or without 13 * modification, are permitted provided that the following conditions 14 * are met: 15 * 1. Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in the 19 * documentation and/or other materials provided with the distribution. 20 * 3. All advertising materials mentioning features or use of this software 21 * must display the following acknowledgement: 22 * This product includes software developed by the University of 23 * California, Berkeley and its contributors. 24 * 4. Neither the name of the University nor the names of its contributors 25 * may be used to endorse or promote products derived from this software 26 * without specific prior written permission. 27 * 28 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 29 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 30 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 31 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 32 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 33 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 34 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 35 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 36 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 37 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 38 * SUCH DAMAGE. 39 * 40 * from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91 41 * 42 * 43 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 44 * All rights reserved. 45 * 46 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 47 * 48 * Permission to use, copy, modify and distribute this software and 49 * its documentation is hereby granted, provided that both the copyright 50 * notice and this permission notice appear in all copies of the 51 * software, derivative works or modified versions, and any portions 52 * thereof, and that both notices appear in supporting documentation. 53 * 54 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 55 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 56 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 57 * 58 * Carnegie Mellon requests users of this software to return to 59 * 60 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 61 * School of Computer Science 62 * Carnegie Mellon University 63 * Pittsburgh PA 15213-3890 64 * 65 * any improvements or extensions that they make and grant Carnegie the 66 * rights to redistribute these changes. 67 * 68 * $FreeBSD: head/sys/vm/vm_pageout.c 115146 2003-05-19 00:51:07Z das $ 69 */ 70 71/* 72 * The proverbial page-out daemon. 73 */ 74 75#include "opt_vm.h" 76#include <sys/param.h> 77#include <sys/systm.h> 78#include <sys/kernel.h> 79#include <sys/eventhandler.h> 80#include <sys/lock.h> 81#include <sys/mutex.h> 82#include <sys/proc.h> 83#include <sys/kthread.h> 84#include <sys/ktr.h> 85#include <sys/resourcevar.h> 86#include <sys/sched.h> 87#include <sys/signalvar.h> 88#include <sys/vnode.h> 89#include <sys/vmmeter.h> 90#include <sys/sx.h> 91#include <sys/sysctl.h> 92 93#include <vm/vm.h> 94#include <vm/vm_param.h> 95#include <vm/vm_object.h> 96#include <vm/vm_page.h> 97#include <vm/vm_map.h> 98#include <vm/vm_pageout.h> 99#include <vm/vm_pager.h> 100#include <vm/swap_pager.h> 101#include <vm/vm_extern.h> 102#include <vm/uma.h> 103 104#include <machine/mutex.h> 105 106/* 107 * System initialization 108 */ 109 110/* the kernel process "vm_pageout"*/ 111static void vm_pageout(void); 112static int vm_pageout_clean(vm_page_t); 113static void vm_pageout_page_free(vm_page_t); 114static void vm_pageout_pmap_collect(void); 115static void vm_pageout_scan(int pass); 116static int vm_pageout_free_page_calc(vm_size_t count); 117struct proc *pageproc; 118 119static struct kproc_desc page_kp = { 120 "pagedaemon", 121 vm_pageout, 122 &pageproc 123}; 124SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp) 125 126#if !defined(NO_SWAPPING) 127/* the kernel process "vm_daemon"*/ 128static void vm_daemon(void); 129static struct proc *vmproc; 130 131static struct kproc_desc vm_kp = { 132 "vmdaemon", 133 vm_daemon, 134 &vmproc 135}; 136SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp) 137#endif 138 139 140int vm_pages_needed; /* Event on which pageout daemon sleeps */ 141int vm_pageout_deficit; /* Estimated number of pages deficit */ 142int vm_pageout_pages_needed; /* flag saying that the pageout daemon needs pages */ 143 144#if !defined(NO_SWAPPING) 145static int vm_pageout_req_swapout; /* XXX */ 146static int vm_daemon_needed; 147#endif 148static int vm_max_launder = 32; 149static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0; 150static int vm_pageout_full_stats_interval = 0; 151static int vm_pageout_stats_free_max=0, vm_pageout_algorithm=0; 152static int defer_swap_pageouts=0; 153static int disable_swap_pageouts=0; 154 155#if defined(NO_SWAPPING) 156static int vm_swap_enabled=0; 157static int vm_swap_idle_enabled=0; 158#else 159static int vm_swap_enabled=1; 160static int vm_swap_idle_enabled=0; 161#endif 162 163SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm, 164 CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt"); 165 166SYSCTL_INT(_vm, OID_AUTO, max_launder, 167 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout"); 168 169SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max, 170 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length"); 171 172SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval, 173 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan"); 174 175SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval, 176 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan"); 177 178SYSCTL_INT(_vm, OID_AUTO, pageout_stats_free_max, 179 CTLFLAG_RW, &vm_pageout_stats_free_max, 0, "Not implemented"); 180 181#if defined(NO_SWAPPING) 182SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 183 CTLFLAG_RD, &vm_swap_enabled, 0, ""); 184SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 185 CTLFLAG_RD, &vm_swap_idle_enabled, 0, ""); 186#else 187SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 188 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout"); 189SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 190 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 191#endif 192 193SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts, 194 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem"); 195 196SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts, 197 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages"); 198 199static int pageout_lock_miss; 200SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss, 201 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout"); 202 203#define VM_PAGEOUT_PAGE_COUNT 16 204int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT; 205 206int vm_page_max_wired; /* XXX max # of wired pages system-wide */ 207 208#if !defined(NO_SWAPPING) 209typedef void freeer_fcn_t(vm_map_t, vm_object_t, vm_pindex_t); 210static void vm_pageout_map_deactivate_pages(vm_map_t, vm_pindex_t); 211static freeer_fcn_t vm_pageout_object_deactivate_pages; 212static void vm_req_vmdaemon(void); 213#endif 214static void vm_pageout_page_stats(void); 215 216/* 217 * vm_pageout_clean: 218 * 219 * Clean the page and remove it from the laundry. 220 * 221 * We set the busy bit to cause potential page faults on this page to 222 * block. Note the careful timing, however, the busy bit isn't set till 223 * late and we cannot do anything that will mess with the page. 224 */ 225static int 226vm_pageout_clean(m) 227 vm_page_t m; 228{ 229 vm_object_t object; 230 vm_page_t mc[2*vm_pageout_page_count]; 231 int pageout_count; 232 int ib, is, page_base; 233 vm_pindex_t pindex = m->pindex; 234 235 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 236 237 object = m->object; 238 239 /* 240 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP 241 * with the new swapper, but we could have serious problems paging 242 * out other object types if there is insufficient memory. 243 * 244 * Unfortunately, checking free memory here is far too late, so the 245 * check has been moved up a procedural level. 246 */ 247 248 /* 249 * Don't mess with the page if it's busy, held, or special 250 */ 251 if ((m->hold_count != 0) || 252 ((m->busy != 0) || (m->flags & (PG_BUSY|PG_UNMANAGED)))) { 253 return 0; 254 } 255 256 mc[vm_pageout_page_count] = m; 257 pageout_count = 1; 258 page_base = vm_pageout_page_count; 259 ib = 1; 260 is = 1; 261 262 /* 263 * Scan object for clusterable pages. 264 * 265 * We can cluster ONLY if: ->> the page is NOT 266 * clean, wired, busy, held, or mapped into a 267 * buffer, and one of the following: 268 * 1) The page is inactive, or a seldom used 269 * active page. 270 * -or- 271 * 2) we force the issue. 272 * 273 * During heavy mmap/modification loads the pageout 274 * daemon can really fragment the underlying file 275 * due to flushing pages out of order and not trying 276 * align the clusters (which leave sporatic out-of-order 277 * holes). To solve this problem we do the reverse scan 278 * first and attempt to align our cluster, then do a 279 * forward scan if room remains. 280 */ 281more: 282 while (ib && pageout_count < vm_pageout_page_count) { 283 vm_page_t p; 284 285 if (ib > pindex) { 286 ib = 0; 287 break; 288 } 289 290 if ((p = vm_page_lookup(object, pindex - ib)) == NULL) { 291 ib = 0; 292 break; 293 } 294 if (((p->queue - p->pc) == PQ_CACHE) || 295 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) { 296 ib = 0; 297 break; 298 } 299 vm_page_test_dirty(p); 300 if ((p->dirty & p->valid) == 0 || 301 p->queue != PQ_INACTIVE || 302 p->wire_count != 0 || /* may be held by buf cache */ 303 p->hold_count != 0) { /* may be undergoing I/O */ 304 ib = 0; 305 break; 306 } 307 mc[--page_base] = p; 308 ++pageout_count; 309 ++ib; 310 /* 311 * alignment boundry, stop here and switch directions. Do 312 * not clear ib. 313 */ 314 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0) 315 break; 316 } 317 318 while (pageout_count < vm_pageout_page_count && 319 pindex + is < object->size) { 320 vm_page_t p; 321 322 if ((p = vm_page_lookup(object, pindex + is)) == NULL) 323 break; 324 if (((p->queue - p->pc) == PQ_CACHE) || 325 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) { 326 break; 327 } 328 vm_page_test_dirty(p); 329 if ((p->dirty & p->valid) == 0 || 330 p->queue != PQ_INACTIVE || 331 p->wire_count != 0 || /* may be held by buf cache */ 332 p->hold_count != 0) { /* may be undergoing I/O */ 333 break; 334 } 335 mc[page_base + pageout_count] = p; 336 ++pageout_count; 337 ++is; 338 } 339 340 /* 341 * If we exhausted our forward scan, continue with the reverse scan 342 * when possible, even past a page boundry. This catches boundry 343 * conditions. 344 */ 345 if (ib && pageout_count < vm_pageout_page_count) 346 goto more; 347 348 /* 349 * we allow reads during pageouts... 350 */ 351 return vm_pageout_flush(&mc[page_base], pageout_count, 0, FALSE); 352} 353 354/* 355 * vm_pageout_flush() - launder the given pages 356 * 357 * The given pages are laundered. Note that we setup for the start of 358 * I/O ( i.e. busy the page ), mark it read-only, and bump the object 359 * reference count all in here rather then in the parent. If we want 360 * the parent to do more sophisticated things we may have to change 361 * the ordering. 362 */ 363int 364vm_pageout_flush(mc, count, flags, is_object_locked) 365 vm_page_t *mc; 366 int count; 367 int flags; 368 int is_object_locked; 369{ 370 vm_object_t object; 371 int pageout_status[count]; 372 int numpagedout = 0; 373 int i; 374 375 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 376 /* 377 * Initiate I/O. Bump the vm_page_t->busy counter and 378 * mark the pages read-only. 379 * 380 * We do not have to fixup the clean/dirty bits here... we can 381 * allow the pager to do it after the I/O completes. 382 * 383 * NOTE! mc[i]->dirty may be partial or fragmented due to an 384 * edge case with file fragments. 385 */ 386 for (i = 0; i < count; i++) { 387 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, ("vm_pageout_flush page %p index %d/%d: partially invalid page", mc[i], i, count)); 388 vm_page_io_start(mc[i]); 389 pmap_page_protect(mc[i], VM_PROT_READ); 390 } 391 object = mc[0]->object; 392 vm_page_unlock_queues(); 393 if (!is_object_locked) 394 VM_OBJECT_LOCK(object); 395 vm_object_pip_add(object, count); 396 VM_OBJECT_UNLOCK(object); 397 398 vm_pager_put_pages(object, mc, count, 399 (flags | ((object == kernel_object) ? VM_PAGER_PUT_SYNC : 0)), 400 pageout_status); 401 402 VM_OBJECT_LOCK(object); 403 vm_page_lock_queues(); 404 for (i = 0; i < count; i++) { 405 vm_page_t mt = mc[i]; 406 407 switch (pageout_status[i]) { 408 case VM_PAGER_OK: 409 case VM_PAGER_PEND: 410 numpagedout++; 411 break; 412 case VM_PAGER_BAD: 413 /* 414 * Page outside of range of object. Right now we 415 * essentially lose the changes by pretending it 416 * worked. 417 */ 418 pmap_clear_modify(mt); 419 vm_page_undirty(mt); 420 break; 421 case VM_PAGER_ERROR: 422 case VM_PAGER_FAIL: 423 /* 424 * If page couldn't be paged out, then reactivate the 425 * page so it doesn't clog the inactive list. (We 426 * will try paging out it again later). 427 */ 428 vm_page_activate(mt); 429 break; 430 case VM_PAGER_AGAIN: 431 break; 432 } 433 434 /* 435 * If the operation is still going, leave the page busy to 436 * block all other accesses. Also, leave the paging in 437 * progress indicator set so that we don't attempt an object 438 * collapse. 439 */ 440 if (pageout_status[i] != VM_PAGER_PEND) { 441 vm_object_pip_wakeup(object); 442 vm_page_io_finish(mt); 443 if (!vm_page_count_severe() || !vm_page_try_to_cache(mt)) 444 pmap_page_protect(mt, VM_PROT_READ); 445 } 446 } 447 if (!is_object_locked) 448 VM_OBJECT_UNLOCK(object); 449 return numpagedout; 450} 451 452#if !defined(NO_SWAPPING) 453/* 454 * vm_pageout_object_deactivate_pages 455 * 456 * deactivate enough pages to satisfy the inactive target 457 * requirements or if vm_page_proc_limit is set, then 458 * deactivate all of the pages in the object and its 459 * backing_objects. 460 * 461 * The object and map must be locked. 462 */ 463static void 464vm_pageout_object_deactivate_pages(map, object, desired) 465 vm_map_t map; 466 vm_object_t object; 467 vm_pindex_t desired; 468{ 469 vm_page_t p, next; 470 int actcount, rcount, remove_mode; 471 472 GIANT_REQUIRED; 473 if (object->type == OBJT_DEVICE || object->type == OBJT_PHYS) 474 return; 475 476 while (object) { 477 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 478 return; 479 if (object->paging_in_progress) 480 return; 481 482 remove_mode = 0; 483 if (object->shadow_count > 1) 484 remove_mode = 1; 485 /* 486 * scan the objects entire memory queue 487 */ 488 rcount = object->resident_page_count; 489 p = TAILQ_FIRST(&object->memq); 490 vm_page_lock_queues(); 491 while (p && (rcount-- > 0)) { 492 if (pmap_resident_count(map->pmap) <= desired) { 493 vm_page_unlock_queues(); 494 return; 495 } 496 next = TAILQ_NEXT(p, listq); 497 cnt.v_pdpages++; 498 if (p->wire_count != 0 || 499 p->hold_count != 0 || 500 p->busy != 0 || 501 (p->flags & (PG_BUSY|PG_UNMANAGED)) || 502 !pmap_page_exists_quick(vm_map_pmap(map), p)) { 503 p = next; 504 continue; 505 } 506 actcount = pmap_ts_referenced(p); 507 if (actcount) { 508 vm_page_flag_set(p, PG_REFERENCED); 509 } else if (p->flags & PG_REFERENCED) { 510 actcount = 1; 511 } 512 if ((p->queue != PQ_ACTIVE) && 513 (p->flags & PG_REFERENCED)) { 514 vm_page_activate(p); 515 p->act_count += actcount; 516 vm_page_flag_clear(p, PG_REFERENCED); 517 } else if (p->queue == PQ_ACTIVE) { 518 if ((p->flags & PG_REFERENCED) == 0) { 519 p->act_count -= min(p->act_count, ACT_DECLINE); 520 if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) { 521 pmap_remove_all(p); 522 vm_page_deactivate(p); 523 } else { 524 vm_pageq_requeue(p); 525 } 526 } else { 527 vm_page_activate(p); 528 vm_page_flag_clear(p, PG_REFERENCED); 529 if (p->act_count < (ACT_MAX - ACT_ADVANCE)) 530 p->act_count += ACT_ADVANCE; 531 vm_pageq_requeue(p); 532 } 533 } else if (p->queue == PQ_INACTIVE) { 534 pmap_remove_all(p); 535 } 536 p = next; 537 } 538 vm_page_unlock_queues(); 539 object = object->backing_object; 540 } 541} 542 543/* 544 * deactivate some number of pages in a map, try to do it fairly, but 545 * that is really hard to do. 546 */ 547static void 548vm_pageout_map_deactivate_pages(map, desired) 549 vm_map_t map; 550 vm_pindex_t desired; 551{ 552 vm_map_entry_t tmpe; 553 vm_object_t obj, bigobj; 554 int nothingwired; 555 556 GIANT_REQUIRED; 557 if (!vm_map_trylock(map)) 558 return; 559 560 bigobj = NULL; 561 nothingwired = TRUE; 562 563 /* 564 * first, search out the biggest object, and try to free pages from 565 * that. 566 */ 567 tmpe = map->header.next; 568 while (tmpe != &map->header) { 569 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 570 obj = tmpe->object.vm_object; 571 if ((obj != NULL) && (obj->shadow_count <= 1) && 572 ((bigobj == NULL) || 573 (bigobj->resident_page_count < obj->resident_page_count))) { 574 bigobj = obj; 575 } 576 } 577 if (tmpe->wired_count > 0) 578 nothingwired = FALSE; 579 tmpe = tmpe->next; 580 } 581 582 if (bigobj) 583 vm_pageout_object_deactivate_pages(map, bigobj, desired); 584 585 /* 586 * Next, hunt around for other pages to deactivate. We actually 587 * do this search sort of wrong -- .text first is not the best idea. 588 */ 589 tmpe = map->header.next; 590 while (tmpe != &map->header) { 591 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 592 break; 593 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 594 obj = tmpe->object.vm_object; 595 if (obj) 596 vm_pageout_object_deactivate_pages(map, obj, desired); 597 } 598 tmpe = tmpe->next; 599 } 600 601 /* 602 * Remove all mappings if a process is swapped out, this will free page 603 * table pages. 604 */ 605 if (desired == 0 && nothingwired) { 606 vm_page_lock_queues(); 607 pmap_remove(vm_map_pmap(map), vm_map_min(map), 608 vm_map_max(map)); 609 vm_page_unlock_queues(); 610 } 611 vm_map_unlock(map); 612} 613#endif /* !defined(NO_SWAPPING) */ 614 615/* 616 * Warning! The page queue lock is released and reacquired. 617 */ 618static void 619vm_pageout_page_free(vm_page_t m) 620{ 621 vm_object_t object = m->object; 622 623 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 624 vm_page_busy(m); 625 vm_page_unlock_queues(); 626 /* 627 * Avoid a lock order reversal. The page must be busy. 628 */ 629 VM_OBJECT_LOCK(object); 630 vm_page_lock_queues(); 631 pmap_remove_all(m); 632 vm_page_free(m); 633 VM_OBJECT_UNLOCK(object); 634 cnt.v_dfree++; 635} 636 637/* 638 * This routine is very drastic, but can save the system 639 * in a pinch. 640 */ 641static void 642vm_pageout_pmap_collect(void) 643{ 644 int i; 645 vm_page_t m; 646 static int warningdone; 647 648 if (pmap_pagedaemon_waken == 0) 649 return; 650 if (warningdone < 5) { 651 printf("collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n"); 652 warningdone++; 653 } 654 vm_page_lock_queues(); 655 for (i = 0; i < vm_page_array_size; i++) { 656 m = &vm_page_array[i]; 657 if (m->wire_count || m->hold_count || m->busy || 658 (m->flags & (PG_BUSY | PG_UNMANAGED))) 659 continue; 660 pmap_remove_all(m); 661 } 662 vm_page_unlock_queues(); 663 pmap_pagedaemon_waken = 0; 664} 665 666/* 667 * vm_pageout_scan does the dirty work for the pageout daemon. 668 */ 669static void 670vm_pageout_scan(int pass) 671{ 672 vm_page_t m, next; 673 struct vm_page marker; 674 int save_page_shortage; 675 int save_inactive_count; 676 int page_shortage, maxscan, pcount; 677 int addl_page_shortage, addl_page_shortage_init; 678 struct proc *p, *bigproc; 679 vm_offset_t size, bigsize; 680 vm_object_t object; 681 int actcount; 682 int vnodes_skipped = 0; 683 int maxlaunder; 684 int s; 685 struct thread *td; 686 687 GIANT_REQUIRED; 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 * Do whatever cleanup that the pmap code can. 698 */ 699 vm_pageout_pmap_collect(); 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 save_page_shortage = page_shortage; 709 save_inactive_count = cnt.v_inactive_count; 710 711 /* 712 * Initialize our marker 713 */ 714 bzero(&marker, sizeof(marker)); 715 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER; 716 marker.queue = PQ_INACTIVE; 717 marker.wire_count = 1; 718 719 /* 720 * Start scanning the inactive queue for pages we can move to the 721 * cache or free. The scan will stop when the target is reached or 722 * we have scanned the entire inactive queue. Note that m->act_count 723 * is not used to form decisions for the inactive queue, only for the 724 * active queue. 725 * 726 * maxlaunder limits the number of dirty pages we flush per scan. 727 * For most systems a smaller value (16 or 32) is more robust under 728 * extreme memory and disk pressure because any unnecessary writes 729 * to disk can result in extreme performance degredation. However, 730 * systems with excessive dirty pages (especially when MAP_NOSYNC is 731 * used) will die horribly with limited laundering. If the pageout 732 * daemon cannot clean enough pages in the first pass, we let it go 733 * all out in succeeding passes. 734 */ 735 if ((maxlaunder = vm_max_launder) <= 1) 736 maxlaunder = 1; 737 if (pass) 738 maxlaunder = 10000; 739rescan0: 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 (m->queue != PQ_INACTIVE) { 750 goto rescan0; 751 } 752 753 next = TAILQ_NEXT(m, pageq); 754 755 /* 756 * skip marker pages 757 */ 758 if (m->flags & PG_MARKER) 759 continue; 760 761 /* 762 * A held page may be undergoing I/O, so skip it. 763 */ 764 if (m->hold_count) { 765 vm_pageq_requeue(m); 766 addl_page_shortage++; 767 continue; 768 } 769 /* 770 * Don't mess with busy pages, keep in the front of the 771 * queue, most likely are being paged out. 772 */ 773 if (m->busy || (m->flags & PG_BUSY)) { 774 addl_page_shortage++; 775 continue; 776 } 777 778 vm_page_lock_queues(); 779 /* 780 * If the object is not being used, we ignore previous 781 * references. 782 */ 783 if (m->object->ref_count == 0) { 784 vm_page_flag_clear(m, PG_REFERENCED); 785 pmap_clear_reference(m); 786 787 /* 788 * Otherwise, if the page has been referenced while in the 789 * inactive queue, we bump the "activation count" upwards, 790 * making it less likely that the page will be added back to 791 * the inactive queue prematurely again. Here we check the 792 * page tables (or emulated bits, if any), given the upper 793 * level VM system not knowing anything about existing 794 * references. 795 */ 796 } else if (((m->flags & PG_REFERENCED) == 0) && 797 (actcount = pmap_ts_referenced(m))) { 798 vm_page_activate(m); 799 vm_page_unlock_queues(); 800 m->act_count += (actcount + ACT_ADVANCE); 801 continue; 802 } 803 804 /* 805 * If the upper level VM system knows about any page 806 * references, we activate the page. We also set the 807 * "activation count" higher than normal so that we will less 808 * likely place pages back onto the inactive queue again. 809 */ 810 if ((m->flags & PG_REFERENCED) != 0) { 811 vm_page_flag_clear(m, PG_REFERENCED); 812 actcount = pmap_ts_referenced(m); 813 vm_page_activate(m); 814 vm_page_unlock_queues(); 815 m->act_count += (actcount + ACT_ADVANCE + 1); 816 continue; 817 } 818 819 /* 820 * If the upper level VM system doesn't know anything about 821 * the page being dirty, we have to check for it again. As 822 * far as the VM code knows, any partially dirty pages are 823 * fully dirty. 824 */ 825 if (m->dirty == 0) { 826 vm_page_test_dirty(m); 827 } else { 828 vm_page_dirty(m); 829 } 830 vm_page_unlock_queues(); 831 832 /* 833 * Invalid pages can be easily freed 834 */ 835 if (m->valid == 0) { 836 vm_page_lock_queues(); 837 vm_pageout_page_free(m); 838 vm_page_unlock_queues(); 839 --page_shortage; 840 841 /* 842 * Clean pages can be placed onto the cache queue. This 843 * effectively frees them. 844 */ 845 } else if (m->dirty == 0) { 846 vm_page_lock_queues(); 847 vm_page_cache(m); 848 vm_page_unlock_queues(); 849 --page_shortage; 850 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) { 851 /* 852 * Dirty pages need to be paged out, but flushing 853 * a page is extremely expensive verses freeing 854 * a clean page. Rather then artificially limiting 855 * the number of pages we can flush, we instead give 856 * dirty pages extra priority on the inactive queue 857 * by forcing them to be cycled through the queue 858 * twice before being flushed, after which the 859 * (now clean) page will cycle through once more 860 * before being freed. This significantly extends 861 * the thrash point for a heavily loaded machine. 862 */ 863 vm_page_lock_queues(); 864 vm_page_flag_set(m, PG_WINATCFLS); 865 vm_pageq_requeue(m); 866 vm_page_unlock_queues(); 867 } else if (maxlaunder > 0) { 868 /* 869 * We always want to try to flush some dirty pages if 870 * we encounter them, to keep the system stable. 871 * Normally this number is small, but under extreme 872 * pressure where there are insufficient clean pages 873 * on the inactive queue, we may have to go all out. 874 */ 875 int swap_pageouts_ok; 876 struct vnode *vp = NULL; 877 struct mount *mp; 878 879 object = m->object; 880 881 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 882 swap_pageouts_ok = 1; 883 } else { 884 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 885 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 886 vm_page_count_min()); 887 888 } 889 890 /* 891 * We don't bother paging objects that are "dead". 892 * Those objects are in a "rundown" state. 893 */ 894 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 895 vm_pageq_requeue(m); 896 continue; 897 } 898 899 /* 900 * The object is already known NOT to be dead. It 901 * is possible for the vget() to block the whole 902 * pageout daemon, but the new low-memory handling 903 * code should prevent it. 904 * 905 * The previous code skipped locked vnodes and, worse, 906 * reordered pages in the queue. This results in 907 * completely non-deterministic operation and, on a 908 * busy system, can lead to extremely non-optimal 909 * pageouts. For example, it can cause clean pages 910 * to be freed and dirty pages to be moved to the end 911 * of the queue. Since dirty pages are also moved to 912 * the end of the queue once-cleaned, this gives 913 * way too large a weighting to defering the freeing 914 * of dirty pages. 915 * 916 * We can't wait forever for the vnode lock, we might 917 * deadlock due to a vn_read() getting stuck in 918 * vm_wait while holding this vnode. We skip the 919 * vnode if we can't get it in a reasonable amount 920 * of time. 921 */ 922 if (object->type == OBJT_VNODE) { 923 vp = object->handle; 924 925 mp = NULL; 926 if (vp->v_type == VREG) 927 vn_start_write(vp, &mp, V_NOWAIT); 928 if (vget(vp, LK_EXCLUSIVE|LK_TIMELOCK, curthread)) { 929 ++pageout_lock_miss; 930 vn_finished_write(mp); 931 if (object->flags & OBJ_MIGHTBEDIRTY) 932 vnodes_skipped++; 933 continue; 934 } 935 936 /* 937 * The page might have been moved to another 938 * queue during potential blocking in vget() 939 * above. The page might have been freed and 940 * reused for another vnode. The object might 941 * have been reused for another vnode. 942 */ 943 if (m->queue != PQ_INACTIVE || 944 m->object != object || 945 object->handle != vp) { 946 if (object->flags & OBJ_MIGHTBEDIRTY) 947 vnodes_skipped++; 948 vput(vp); 949 vn_finished_write(mp); 950 continue; 951 } 952 953 /* 954 * The page may have been busied during the 955 * blocking in vput(); We don't move the 956 * page back onto the end of the queue so that 957 * statistics are more correct if we don't. 958 */ 959 if (m->busy || (m->flags & PG_BUSY)) { 960 vput(vp); 961 vn_finished_write(mp); 962 continue; 963 } 964 965 /* 966 * If the page has become held it might 967 * be undergoing I/O, so skip it 968 */ 969 if (m->hold_count) { 970 vm_pageq_requeue(m); 971 if (object->flags & OBJ_MIGHTBEDIRTY) 972 vnodes_skipped++; 973 vput(vp); 974 vn_finished_write(mp); 975 continue; 976 } 977 } 978 979 /* 980 * If a page is dirty, then it is either being washed 981 * (but not yet cleaned) or it is still in the 982 * laundry. If it is still in the laundry, then we 983 * start the cleaning operation. 984 * 985 * This operation may cluster, invalidating the 'next' 986 * pointer. To prevent an inordinate number of 987 * restarts we use our marker to remember our place. 988 * 989 * decrement page_shortage on success to account for 990 * the (future) cleaned page. Otherwise we could wind 991 * up laundering or cleaning too many pages. 992 */ 993 vm_page_lock_queues(); 994 s = splvm(); 995 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, m, &marker, pageq); 996 splx(s); 997 if (vm_pageout_clean(m) != 0) { 998 --page_shortage; 999 --maxlaunder; 1000 } 1001 s = splvm(); 1002 next = TAILQ_NEXT(&marker, pageq); 1003 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, &marker, pageq); 1004 splx(s); 1005 vm_page_unlock_queues(); 1006 if (vp) { 1007 vput(vp); 1008 vn_finished_write(mp); 1009 } 1010 } 1011 } 1012 1013 /* 1014 * Compute the number of pages we want to try to move from the 1015 * active queue to the inactive queue. 1016 */ 1017 page_shortage = vm_paging_target() + 1018 cnt.v_inactive_target - cnt.v_inactive_count; 1019 page_shortage += addl_page_shortage; 1020 1021 vm_page_lock_queues(); 1022 /* 1023 * Scan the active queue for things we can deactivate. We nominally 1024 * track the per-page activity counter and use it to locate 1025 * deactivation candidates. 1026 */ 1027 pcount = cnt.v_active_count; 1028 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1029 1030 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) { 1031 1032 /* 1033 * This is a consistency check, and should likely be a panic 1034 * or warning. 1035 */ 1036 if (m->queue != PQ_ACTIVE) { 1037 break; 1038 } 1039 1040 next = TAILQ_NEXT(m, pageq); 1041 /* 1042 * Don't deactivate pages that are busy. 1043 */ 1044 if ((m->busy != 0) || 1045 (m->flags & PG_BUSY) || 1046 (m->hold_count != 0)) { 1047 vm_pageq_requeue(m); 1048 m = next; 1049 continue; 1050 } 1051 1052 /* 1053 * The count for pagedaemon pages is done after checking the 1054 * page for eligibility... 1055 */ 1056 cnt.v_pdpages++; 1057 1058 /* 1059 * Check to see "how much" the page has been used. 1060 */ 1061 actcount = 0; 1062 if (m->object->ref_count != 0) { 1063 if (m->flags & PG_REFERENCED) { 1064 actcount += 1; 1065 } 1066 actcount += pmap_ts_referenced(m); 1067 if (actcount) { 1068 m->act_count += ACT_ADVANCE + actcount; 1069 if (m->act_count > ACT_MAX) 1070 m->act_count = ACT_MAX; 1071 } 1072 } 1073 1074 /* 1075 * Since we have "tested" this bit, we need to clear it now. 1076 */ 1077 vm_page_flag_clear(m, PG_REFERENCED); 1078 1079 /* 1080 * Only if an object is currently being used, do we use the 1081 * page activation count stats. 1082 */ 1083 if (actcount && (m->object->ref_count != 0)) { 1084 vm_pageq_requeue(m); 1085 } else { 1086 m->act_count -= min(m->act_count, ACT_DECLINE); 1087 if (vm_pageout_algorithm || 1088 m->object->ref_count == 0 || 1089 m->act_count == 0) { 1090 page_shortage--; 1091 if (m->object->ref_count == 0) { 1092 pmap_remove_all(m); 1093 if (m->dirty == 0) 1094 vm_page_cache(m); 1095 else 1096 vm_page_deactivate(m); 1097 } else { 1098 vm_page_deactivate(m); 1099 } 1100 } else { 1101 vm_pageq_requeue(m); 1102 } 1103 } 1104 m = next; 1105 } 1106 s = splvm(); 1107 1108 /* 1109 * We try to maintain some *really* free pages, this allows interrupt 1110 * code to be guaranteed space. Since both cache and free queues 1111 * are considered basically 'free', moving pages from cache to free 1112 * does not effect other calculations. 1113 */ 1114 while (cnt.v_free_count < cnt.v_free_reserved) { 1115 static int cache_rover = 0; 1116 m = vm_pageq_find(PQ_CACHE, cache_rover, FALSE); 1117 if (!m) 1118 break; 1119 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || 1120 m->busy || 1121 m->hold_count || 1122 m->wire_count) { 1123#ifdef INVARIANTS 1124 printf("Warning: busy page %p found in cache\n", m); 1125#endif 1126 vm_page_deactivate(m); 1127 continue; 1128 } 1129 cache_rover = (cache_rover + PQ_PRIME2) & PQ_L2_MASK; 1130 vm_pageout_page_free(m); 1131 } 1132 splx(s); 1133 vm_page_unlock_queues(); 1134#if !defined(NO_SWAPPING) 1135 /* 1136 * Idle process swapout -- run once per second. 1137 */ 1138 if (vm_swap_idle_enabled) { 1139 static long lsec; 1140 if (time_second != lsec) { 1141 vm_pageout_req_swapout |= VM_SWAP_IDLE; 1142 vm_req_vmdaemon(); 1143 lsec = time_second; 1144 } 1145 } 1146#endif 1147 1148 /* 1149 * If we didn't get enough free pages, and we have skipped a vnode 1150 * in a writeable object, wakeup the sync daemon. And kick swapout 1151 * if we did not get enough free pages. 1152 */ 1153 if (vm_paging_target() > 0) { 1154 if (vnodes_skipped && vm_page_count_min()) 1155 (void) speedup_syncer(); 1156#if !defined(NO_SWAPPING) 1157 if (vm_swap_enabled && vm_page_count_target()) { 1158 vm_req_vmdaemon(); 1159 vm_pageout_req_swapout |= VM_SWAP_NORMAL; 1160 } 1161#endif 1162 } 1163 1164 /* 1165 * If we are critically low on one of RAM or swap and low on 1166 * the other, kill the largest process. However, we avoid 1167 * doing this on the first pass in order to give ourselves a 1168 * chance to flush out dirty vnode-backed pages and to allow 1169 * active pages to be moved to the inactive queue and reclaimed. 1170 * 1171 * We keep the process bigproc locked once we find it to keep anyone 1172 * from messing with it; however, there is a possibility of 1173 * deadlock if process B is bigproc and one of it's child processes 1174 * attempts to propagate a signal to B while we are waiting for A's 1175 * lock while walking this list. To avoid this, we don't block on 1176 * the process lock but just skip a process if it is already locked. 1177 */ 1178 if (pass != 0 && 1179 ((vm_swap_size < 64 && vm_page_count_min()) || 1180 (swap_pager_full && vm_paging_target() > 0))) { 1181 bigproc = NULL; 1182 bigsize = 0; 1183 sx_slock(&allproc_lock); 1184 FOREACH_PROC_IN_SYSTEM(p) { 1185 int breakout; 1186 /* 1187 * If this process is already locked, skip it. 1188 */ 1189 if (PROC_TRYLOCK(p) == 0) 1190 continue; 1191 /* 1192 * If this is a system or protected process, skip it. 1193 */ 1194 if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) || 1195 (p->p_flag & P_PROTECTED) || 1196 ((p->p_pid < 48) && (vm_swap_size != 0))) { 1197 PROC_UNLOCK(p); 1198 continue; 1199 } 1200 /* 1201 * if the process is in a non-running type state, 1202 * don't touch it. Check all the threads individually. 1203 */ 1204 mtx_lock_spin(&sched_lock); 1205 breakout = 0; 1206 FOREACH_THREAD_IN_PROC(p, td) { 1207 if (!TD_ON_RUNQ(td) && 1208 !TD_IS_RUNNING(td) && 1209 !TD_IS_SLEEPING(td)) { 1210 breakout = 1; 1211 break; 1212 } 1213 } 1214 if (breakout) { 1215 mtx_unlock_spin(&sched_lock); 1216 PROC_UNLOCK(p); 1217 continue; 1218 } 1219 mtx_unlock_spin(&sched_lock); 1220 /* 1221 * get the process size 1222 */ 1223 if (!vm_map_trylock_read(&p->p_vmspace->vm_map)) { 1224 PROC_UNLOCK(p); 1225 continue; 1226 } 1227 size = vmspace_swap_count(p->p_vmspace); 1228 vm_map_unlock_read(&p->p_vmspace->vm_map); 1229 size += vmspace_resident_count(p->p_vmspace); 1230 /* 1231 * if the this process is bigger than the biggest one 1232 * remember it. 1233 */ 1234 if (size > bigsize) { 1235 if (bigproc != NULL) 1236 PROC_UNLOCK(bigproc); 1237 bigproc = p; 1238 bigsize = size; 1239 } else 1240 PROC_UNLOCK(p); 1241 } 1242 sx_sunlock(&allproc_lock); 1243 if (bigproc != NULL) { 1244 struct ksegrp *kg; 1245 killproc(bigproc, "out of swap space"); 1246 mtx_lock_spin(&sched_lock); 1247 FOREACH_KSEGRP_IN_PROC(bigproc, kg) { 1248 sched_nice(kg, PRIO_MIN); /* XXXKSE ??? */ 1249 } 1250 mtx_unlock_spin(&sched_lock); 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 */ 1263static void 1264vm_pageout_page_stats() 1265{ 1266 vm_page_t m,next; 1267 int pcount,tpcount; /* Number of pages to check */ 1268 static int fullintervalcount = 0; 1269 int page_shortage; 1270 int s0; 1271 1272 page_shortage = 1273 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) - 1274 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count); 1275 1276 if (page_shortage <= 0) 1277 return; 1278 1279 s0 = splvm(); 1280 vm_page_lock_queues(); 1281 pcount = cnt.v_active_count; 1282 fullintervalcount += vm_pageout_stats_interval; 1283 if (fullintervalcount < vm_pageout_full_stats_interval) { 1284 tpcount = (vm_pageout_stats_max * cnt.v_active_count) / cnt.v_page_count; 1285 if (pcount > tpcount) 1286 pcount = tpcount; 1287 } else { 1288 fullintervalcount = 0; 1289 } 1290 1291 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1292 while ((m != NULL) && (pcount-- > 0)) { 1293 int actcount; 1294 1295 if (m->queue != PQ_ACTIVE) { 1296 break; 1297 } 1298 1299 next = TAILQ_NEXT(m, pageq); 1300 /* 1301 * Don't deactivate pages that are busy. 1302 */ 1303 if ((m->busy != 0) || 1304 (m->flags & PG_BUSY) || 1305 (m->hold_count != 0)) { 1306 vm_pageq_requeue(m); 1307 m = next; 1308 continue; 1309 } 1310 1311 actcount = 0; 1312 if (m->flags & PG_REFERENCED) { 1313 vm_page_flag_clear(m, PG_REFERENCED); 1314 actcount += 1; 1315 } 1316 1317 actcount += pmap_ts_referenced(m); 1318 if (actcount) { 1319 m->act_count += ACT_ADVANCE + actcount; 1320 if (m->act_count > ACT_MAX) 1321 m->act_count = ACT_MAX; 1322 vm_pageq_requeue(m); 1323 } else { 1324 if (m->act_count == 0) { 1325 /* 1326 * We turn off page access, so that we have 1327 * more accurate RSS stats. We don't do this 1328 * in the normal page deactivation when the 1329 * system is loaded VM wise, because the 1330 * cost of the large number of page protect 1331 * operations would be higher than the value 1332 * of doing the operation. 1333 */ 1334 pmap_remove_all(m); 1335 vm_page_deactivate(m); 1336 } else { 1337 m->act_count -= min(m->act_count, ACT_DECLINE); 1338 vm_pageq_requeue(m); 1339 } 1340 } 1341 1342 m = next; 1343 } 1344 vm_page_unlock_queues(); 1345 splx(s0); 1346} 1347 1348static int 1349vm_pageout_free_page_calc(count) 1350vm_size_t count; 1351{ 1352 if (count < cnt.v_page_count) 1353 return 0; 1354 /* 1355 * free_reserved needs to include enough for the largest swap pager 1356 * structures plus enough for any pv_entry structs when paging. 1357 */ 1358 if (cnt.v_page_count > 1024) 1359 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200; 1360 else 1361 cnt.v_free_min = 4; 1362 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1363 cnt.v_interrupt_free_min; 1364 cnt.v_free_reserved = vm_pageout_page_count + 1365 cnt.v_pageout_free_min + (count / 768) + PQ_L2_SIZE; 1366 cnt.v_free_severe = cnt.v_free_min / 2; 1367 cnt.v_free_min += cnt.v_free_reserved; 1368 cnt.v_free_severe += cnt.v_free_reserved; 1369 return 1; 1370} 1371 1372/* 1373 * vm_pageout is the high level pageout daemon. 1374 */ 1375static void 1376vm_pageout() 1377{ 1378 int error, pass, s; 1379 1380 mtx_lock(&Giant); 1381 1382 /* 1383 * Initialize some paging parameters. 1384 */ 1385 cnt.v_interrupt_free_min = 2; 1386 if (cnt.v_page_count < 2000) 1387 vm_pageout_page_count = 8; 1388 1389 vm_pageout_free_page_calc(cnt.v_page_count); 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 /* 1434 * Set maximum free per pass 1435 */ 1436 if (vm_pageout_stats_free_max == 0) 1437 vm_pageout_stats_free_max = 5; 1438 1439 swap_pager_swap_init(); 1440 pass = 0; 1441 /* 1442 * The pageout daemon is never done, so loop forever. 1443 */ 1444 while (TRUE) { 1445 s = splvm(); 1446 vm_page_lock_queues(); 1447 /* 1448 * If we have enough free memory, wakeup waiters. Do 1449 * not clear vm_pages_needed until we reach our target, 1450 * otherwise we may be woken up over and over again and 1451 * waste a lot of cpu. 1452 */ 1453 if (vm_pages_needed && !vm_page_count_min()) { 1454 if (!vm_paging_needed()) 1455 vm_pages_needed = 0; 1456 wakeup(&cnt.v_free_count); 1457 } 1458 if (vm_pages_needed) { 1459 /* 1460 * Still not done, take a second pass without waiting 1461 * (unlimited dirty cleaning), otherwise sleep a bit 1462 * and try again. 1463 */ 1464 ++pass; 1465 if (pass > 1) 1466 msleep(&vm_pages_needed, &vm_page_queue_mtx, PVM, 1467 "psleep", hz/2); 1468 } else { 1469 /* 1470 * Good enough, sleep & handle stats. Prime the pass 1471 * for the next run. 1472 */ 1473 if (pass > 1) 1474 pass = 1; 1475 else 1476 pass = 0; 1477 error = msleep(&vm_pages_needed, &vm_page_queue_mtx, PVM, 1478 "psleep", vm_pageout_stats_interval * hz); 1479 if (error && !vm_pages_needed) { 1480 vm_page_unlock_queues(); 1481 splx(s); 1482 pass = 0; 1483 vm_pageout_page_stats(); 1484 continue; 1485 } 1486 } 1487 if (vm_pages_needed) 1488 cnt.v_pdwakeups++; 1489 vm_page_unlock_queues(); 1490 splx(s); 1491 vm_pageout_scan(pass); 1492 } 1493} 1494 1495/* 1496 * Unless the page queue lock is held by the caller, this function 1497 * should be regarded as advisory. Specifically, the caller should 1498 * not msleep() on &cnt.v_free_count following this function unless 1499 * the page queue lock is held until the msleep() is performed. 1500 */ 1501void 1502pagedaemon_wakeup() 1503{ 1504 1505 if (!vm_pages_needed && curthread->td_proc != pageproc) { 1506 vm_pages_needed = 1; 1507 wakeup(&vm_pages_needed); 1508 } 1509} 1510 1511#if !defined(NO_SWAPPING) 1512static void 1513vm_req_vmdaemon() 1514{ 1515 static int lastrun = 0; 1516 1517 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1518 wakeup(&vm_daemon_needed); 1519 lastrun = ticks; 1520 } 1521} 1522 1523static void 1524vm_daemon() 1525{ 1526 struct proc *p; 1527 int breakout; 1528 struct thread *td; 1529 1530 mtx_lock(&Giant); 1531 while (TRUE) { 1532 tsleep(&vm_daemon_needed, PPAUSE, "psleep", 0); 1533 if (vm_pageout_req_swapout) { 1534 swapout_procs(vm_pageout_req_swapout); 1535 vm_pageout_req_swapout = 0; 1536 } 1537 /* 1538 * scan the processes for exceeding their rlimits or if 1539 * process is swapped out -- deactivate pages 1540 */ 1541 sx_slock(&allproc_lock); 1542 LIST_FOREACH(p, &allproc, p_list) { 1543 vm_pindex_t limit, size; 1544 1545 /* 1546 * if this is a system process or if we have already 1547 * looked at this process, skip it. 1548 */ 1549 PROC_LOCK(p); 1550 if (p->p_flag & (P_SYSTEM | P_WEXIT)) { 1551 PROC_UNLOCK(p); 1552 continue; 1553 } 1554 /* 1555 * if the process is in a non-running type state, 1556 * don't touch it. 1557 */ 1558 mtx_lock_spin(&sched_lock); 1559 breakout = 0; 1560 FOREACH_THREAD_IN_PROC(p, td) { 1561 if (!TD_ON_RUNQ(td) && 1562 !TD_IS_RUNNING(td) && 1563 !TD_IS_SLEEPING(td)) { 1564 breakout = 1; 1565 break; 1566 } 1567 } 1568 mtx_unlock_spin(&sched_lock); 1569 if (breakout) { 1570 PROC_UNLOCK(p); 1571 continue; 1572 } 1573 /* 1574 * get a limit 1575 */ 1576 limit = OFF_TO_IDX( 1577 qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur, 1578 p->p_rlimit[RLIMIT_RSS].rlim_max)); 1579 1580 /* 1581 * let processes that are swapped out really be 1582 * swapped out set the limit to nothing (will force a 1583 * swap-out.) 1584 */ 1585 if ((p->p_sflag & PS_INMEM) == 0) 1586 limit = 0; /* XXX */ 1587 PROC_UNLOCK(p); 1588 1589 size = vmspace_resident_count(p->p_vmspace); 1590 if (limit >= 0 && size >= limit) { 1591 vm_pageout_map_deactivate_pages( 1592 &p->p_vmspace->vm_map, limit); 1593 } 1594 } 1595 sx_sunlock(&allproc_lock); 1596 } 1597} 1598#endif /* !defined(NO_SWAPPING) */ 1599