vm_pageout.c revision 108011
1117397Skan/* 2117397Skan * Copyright (c) 1991 Regents of the University of California. 3169691Skan * All rights reserved. 4117397Skan * Copyright (c) 1994 John S. Dyson 5117397Skan * All rights reserved. 6117397Skan * Copyright (c) 1994 David Greenman 7117397Skan * All rights reserved. 8117397Skan * 9117397Skan * This code is derived from software contributed to Berkeley by 10117397Skan * The Mach Operating System project at Carnegie-Mellon University. 11117397Skan * 12117397Skan * Redistribution and use in source and binary forms, with or without 13117397Skan * modification, are permitted provided that the following conditions 14117397Skan * are met: 15117397Skan * 1. Redistributions of source code must retain the above copyright 16117397Skan * notice, this list of conditions and the following disclaimer. 17117397Skan * 2. Redistributions in binary form must reproduce the above copyright 18169691Skan * notice, this list of conditions and the following disclaimer in the 19117397Skan * documentation and/or other materials provided with the distribution. 20117397Skan * 3. All advertising materials mentioning features or use of this software 21117397Skan * must display the following acknowledgement: 22117397Skan * This product includes software developed by the University of 23117397Skan * California, Berkeley and its contributors. 24117397Skan * 4. Neither the name of the University nor the names of its contributors 25117397Skan * may be used to endorse or promote products derived from this software 26117397Skan * without specific prior written permission. 27117397Skan * 28117397Skan * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 29117397Skan * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 30132720Skan * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 31132720Skan * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 32117397Skan * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 33117397Skan * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 34117397Skan * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 35117397Skan * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 36117397Skan * 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 108011 2002-12-18 04:02:02Z alc $ 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_pmap_collect(void); 114static void vm_pageout_scan(int pass); 115static int vm_pageout_free_page_calc(vm_size_t count); 116struct proc *pageproc; 117 118static struct kproc_desc page_kp = { 119 "pagedaemon", 120 vm_pageout, 121 &pageproc 122}; 123SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp) 124 125#if !defined(NO_SWAPPING) 126/* the kernel process "vm_daemon"*/ 127static void vm_daemon(void); 128static struct proc *vmproc; 129 130static struct kproc_desc vm_kp = { 131 "vmdaemon", 132 vm_daemon, 133 &vmproc 134}; 135SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp) 136#endif 137 138 139int vm_pages_needed=0; /* Event on which pageout daemon sleeps */ 140int vm_pageout_deficit=0; /* Estimated number of pages deficit */ 141int vm_pageout_pages_needed=0; /* flag saying that the pageout daemon needs pages */ 142 143#if !defined(NO_SWAPPING) 144static int vm_pageout_req_swapout; /* XXX */ 145static int vm_daemon_needed; 146#endif 147extern int vm_swap_size; 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, int); 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); 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) 365 vm_page_t *mc; 366 int count; 367 int flags; 368{ 369 vm_object_t object; 370 int pageout_status[count]; 371 int numpagedout = 0; 372 int i; 373 374 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 375 /* 376 * Initiate I/O. Bump the vm_page_t->busy counter and 377 * mark the pages read-only. 378 * 379 * We do not have to fixup the clean/dirty bits here... we can 380 * allow the pager to do it after the I/O completes. 381 * 382 * NOTE! mc[i]->dirty may be partial or fragmented due to an 383 * edge case with file fragments. 384 */ 385 for (i = 0; i < count; i++) { 386 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, ("vm_pageout_flush page %p index %d/%d: partially invalid page", mc[i], i, count)); 387 vm_page_io_start(mc[i]); 388 pmap_page_protect(mc[i], VM_PROT_READ); 389 } 390 object = mc[0]->object; 391 vm_page_unlock_queues(); 392 vm_object_pip_add(object, count); 393 394 vm_pager_put_pages(object, mc, count, 395 (flags | ((object == kernel_object) ? OBJPC_SYNC : 0)), 396 pageout_status); 397 398 vm_page_lock_queues(); 399 for (i = 0; i < count; i++) { 400 vm_page_t mt = mc[i]; 401 402 switch (pageout_status[i]) { 403 case VM_PAGER_OK: 404 numpagedout++; 405 break; 406 case VM_PAGER_PEND: 407 numpagedout++; 408 break; 409 case VM_PAGER_BAD: 410 /* 411 * Page outside of range of object. Right now we 412 * essentially lose the changes by pretending it 413 * worked. 414 */ 415 pmap_clear_modify(mt); 416 vm_page_undirty(mt); 417 break; 418 case VM_PAGER_ERROR: 419 case VM_PAGER_FAIL: 420 /* 421 * If page couldn't be paged out, then reactivate the 422 * page so it doesn't clog the inactive list. (We 423 * will try paging out it again later). 424 */ 425 vm_page_activate(mt); 426 break; 427 case VM_PAGER_AGAIN: 428 break; 429 } 430 431 /* 432 * If the operation is still going, leave the page busy to 433 * block all other accesses. Also, leave the paging in 434 * progress indicator set so that we don't attempt an object 435 * collapse. 436 */ 437 if (pageout_status[i] != VM_PAGER_PEND) { 438 vm_object_pip_wakeup(object); 439 vm_page_io_finish(mt); 440 if (!vm_page_count_severe() || !vm_page_try_to_cache(mt)) 441 pmap_page_protect(mt, VM_PROT_READ); 442 } 443 } 444 return numpagedout; 445} 446 447#if !defined(NO_SWAPPING) 448/* 449 * vm_pageout_object_deactivate_pages 450 * 451 * deactivate enough pages to satisfy the inactive target 452 * requirements or if vm_page_proc_limit is set, then 453 * deactivate all of the pages in the object and its 454 * backing_objects. 455 * 456 * The object and map must be locked. 457 */ 458static void 459vm_pageout_object_deactivate_pages(map, object, desired, map_remove_only) 460 vm_map_t map; 461 vm_object_t object; 462 vm_pindex_t desired; 463 int map_remove_only; 464{ 465 vm_page_t p, next; 466 int actcount, rcount, remove_mode; 467 468 GIANT_REQUIRED; 469 if (object->type == OBJT_DEVICE || object->type == OBJT_PHYS) 470 return; 471 472 while (object) { 473 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 474 return; 475 if (object->paging_in_progress) 476 return; 477 478 remove_mode = map_remove_only; 479 if (object->shadow_count > 1) 480 remove_mode = 1; 481 /* 482 * scan the objects entire memory queue 483 */ 484 rcount = object->resident_page_count; 485 p = TAILQ_FIRST(&object->memq); 486 vm_page_lock_queues(); 487 while (p && (rcount-- > 0)) { 488 if (pmap_resident_count(map->pmap) <= desired) { 489 vm_page_unlock_queues(); 490 return; 491 } 492 next = TAILQ_NEXT(p, listq); 493 cnt.v_pdpages++; 494 if (p->wire_count != 0 || 495 p->hold_count != 0 || 496 p->busy != 0 || 497 (p->flags & (PG_BUSY|PG_UNMANAGED)) || 498 !pmap_page_exists_quick(vm_map_pmap(map), p)) { 499 p = next; 500 continue; 501 } 502 actcount = pmap_ts_referenced(p); 503 if (actcount) { 504 vm_page_flag_set(p, PG_REFERENCED); 505 } else if (p->flags & PG_REFERENCED) { 506 actcount = 1; 507 } 508 if ((p->queue != PQ_ACTIVE) && 509 (p->flags & PG_REFERENCED)) { 510 vm_page_activate(p); 511 p->act_count += actcount; 512 vm_page_flag_clear(p, PG_REFERENCED); 513 } else if (p->queue == PQ_ACTIVE) { 514 if ((p->flags & PG_REFERENCED) == 0) { 515 p->act_count -= min(p->act_count, ACT_DECLINE); 516 if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) { 517 pmap_remove_all(p); 518 vm_page_deactivate(p); 519 } else { 520 vm_pageq_requeue(p); 521 } 522 } else { 523 vm_page_activate(p); 524 vm_page_flag_clear(p, PG_REFERENCED); 525 if (p->act_count < (ACT_MAX - ACT_ADVANCE)) 526 p->act_count += ACT_ADVANCE; 527 vm_pageq_requeue(p); 528 } 529 } else if (p->queue == PQ_INACTIVE) { 530 pmap_remove_all(p); 531 } 532 p = next; 533 } 534 vm_page_unlock_queues(); 535 object = object->backing_object; 536 } 537} 538 539/* 540 * deactivate some number of pages in a map, try to do it fairly, but 541 * that is really hard to do. 542 */ 543static void 544vm_pageout_map_deactivate_pages(map, desired) 545 vm_map_t map; 546 vm_pindex_t desired; 547{ 548 vm_map_entry_t tmpe; 549 vm_object_t obj, bigobj; 550 int nothingwired; 551 552 GIANT_REQUIRED; 553 if (!vm_map_trylock(map)) 554 return; 555 556 bigobj = NULL; 557 nothingwired = TRUE; 558 559 /* 560 * first, search out the biggest object, and try to free pages from 561 * that. 562 */ 563 tmpe = map->header.next; 564 while (tmpe != &map->header) { 565 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 566 obj = tmpe->object.vm_object; 567 if ((obj != NULL) && (obj->shadow_count <= 1) && 568 ((bigobj == NULL) || 569 (bigobj->resident_page_count < obj->resident_page_count))) { 570 bigobj = obj; 571 } 572 } 573 if (tmpe->wired_count > 0) 574 nothingwired = FALSE; 575 tmpe = tmpe->next; 576 } 577 578 if (bigobj) 579 vm_pageout_object_deactivate_pages(map, bigobj, desired, 0); 580 581 /* 582 * Next, hunt around for other pages to deactivate. We actually 583 * do this search sort of wrong -- .text first is not the best idea. 584 */ 585 tmpe = map->header.next; 586 while (tmpe != &map->header) { 587 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 588 break; 589 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 590 obj = tmpe->object.vm_object; 591 if (obj) 592 vm_pageout_object_deactivate_pages(map, obj, desired, 0); 593 } 594 tmpe = tmpe->next; 595 } 596 597 /* 598 * Remove all mappings if a process is swapped out, this will free page 599 * table pages. 600 */ 601 if (desired == 0 && nothingwired) { 602 vm_page_lock_queues(); 603 pmap_remove(vm_map_pmap(map), vm_map_min(map), 604 vm_map_max(map)); 605 vm_page_unlock_queues(); 606 } 607 vm_map_unlock(map); 608} 609#endif /* !defined(NO_SWAPPING) */ 610 611/* 612 * Don't try to be fancy - being fancy can lead to VOP_LOCK's and therefore 613 * to vnode deadlocks. We only do it for OBJT_DEFAULT and OBJT_SWAP objects 614 * which we know can be trivially freed. 615 */ 616void 617vm_pageout_page_free(vm_page_t m) 618{ 619 vm_object_t object = m->object; 620 int type = object->type; 621 622 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 623 if (type == OBJT_SWAP || type == OBJT_DEFAULT) 624 vm_object_reference(object); 625 vm_page_busy(m); 626 pmap_remove_all(m); 627 vm_page_free(m); 628 cnt.v_dfree++; 629 if (type == OBJT_SWAP || type == OBJT_DEFAULT) 630 vm_object_deallocate(object); 631} 632 633/* 634 * This routine is very drastic, but can save the system 635 * in a pinch. 636 */ 637static void 638vm_pageout_pmap_collect(void) 639{ 640 int i; 641 vm_page_t m; 642 static int warningdone; 643 644 if (pmap_pagedaemon_waken == 0) 645 return; 646 if (warningdone < 5) { 647 printf("collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n"); 648 warningdone++; 649 } 650 vm_page_lock_queues(); 651 for (i = 0; i < vm_page_array_size; i++) { 652 m = &vm_page_array[i]; 653 if (m->wire_count || m->hold_count || m->busy || 654 (m->flags & (PG_BUSY | PG_UNMANAGED))) 655 continue; 656 pmap_remove_all(m); 657 } 658 vm_page_unlock_queues(); 659 pmap_pagedaemon_waken = 0; 660} 661 662/* 663 * vm_pageout_scan does the dirty work for the pageout daemon. 664 */ 665static void 666vm_pageout_scan(int pass) 667{ 668 vm_page_t m, next; 669 struct vm_page marker; 670 int save_page_shortage; 671 int save_inactive_count; 672 int page_shortage, maxscan, pcount; 673 int addl_page_shortage, addl_page_shortage_init; 674 struct proc *p, *bigproc; 675 vm_offset_t size, bigsize; 676 vm_object_t object; 677 int actcount; 678 int vnodes_skipped = 0; 679 int maxlaunder; 680 int s; 681 struct thread *td; 682 683 GIANT_REQUIRED; 684 /* 685 * Decrease registered cache sizes. 686 */ 687 EVENTHANDLER_INVOKE(vm_lowmem, 0); 688 /* 689 * We do this explicitly after the caches have been drained above. 690 */ 691 uma_reclaim(); 692 /* 693 * Do whatever cleanup that the pmap code can. 694 */ 695 vm_pageout_pmap_collect(); 696 697 addl_page_shortage_init = vm_pageout_deficit; 698 vm_pageout_deficit = 0; 699 700 /* 701 * Calculate the number of pages we want to either free or move 702 * to the cache. 703 */ 704 page_shortage = vm_paging_target() + addl_page_shortage_init; 705 save_page_shortage = page_shortage; 706 save_inactive_count = cnt.v_inactive_count; 707 708 /* 709 * Initialize our marker 710 */ 711 bzero(&marker, sizeof(marker)); 712 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER; 713 marker.queue = PQ_INACTIVE; 714 marker.wire_count = 1; 715 716 /* 717 * Start scanning the inactive queue for pages we can move to the 718 * cache or free. The scan will stop when the target is reached or 719 * we have scanned the entire inactive queue. Note that m->act_count 720 * is not used to form decisions for the inactive queue, only for the 721 * active queue. 722 * 723 * maxlaunder limits the number of dirty pages we flush per scan. 724 * For most systems a smaller value (16 or 32) is more robust under 725 * extreme memory and disk pressure because any unnecessary writes 726 * to disk can result in extreme performance degredation. However, 727 * systems with excessive dirty pages (especially when MAP_NOSYNC is 728 * used) will die horribly with limited laundering. If the pageout 729 * daemon cannot clean enough pages in the first pass, we let it go 730 * all out in succeeding passes. 731 */ 732 if ((maxlaunder = vm_max_launder) <= 1) 733 maxlaunder = 1; 734 if (pass) 735 maxlaunder = 10000; 736rescan0: 737 addl_page_shortage = addl_page_shortage_init; 738 maxscan = cnt.v_inactive_count; 739 740 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl); 741 m != NULL && maxscan-- > 0 && page_shortage > 0; 742 m = next) { 743 744 cnt.v_pdpages++; 745 746 if (m->queue != PQ_INACTIVE) { 747 goto rescan0; 748 } 749 750 next = TAILQ_NEXT(m, pageq); 751 752 /* 753 * skip marker pages 754 */ 755 if (m->flags & PG_MARKER) 756 continue; 757 758 /* 759 * A held page may be undergoing I/O, so skip it. 760 */ 761 if (m->hold_count) { 762 vm_pageq_requeue(m); 763 addl_page_shortage++; 764 continue; 765 } 766 /* 767 * Don't mess with busy pages, keep in the front of the 768 * queue, most likely are being paged out. 769 */ 770 if (m->busy || (m->flags & PG_BUSY)) { 771 addl_page_shortage++; 772 continue; 773 } 774 775 vm_page_lock_queues(); 776 /* 777 * If the object is not being used, we ignore previous 778 * references. 779 */ 780 if (m->object->ref_count == 0) { 781 vm_page_flag_clear(m, PG_REFERENCED); 782 pmap_clear_reference(m); 783 784 /* 785 * Otherwise, if the page has been referenced while in the 786 * inactive queue, we bump the "activation count" upwards, 787 * making it less likely that the page will be added back to 788 * the inactive queue prematurely again. Here we check the 789 * page tables (or emulated bits, if any), given the upper 790 * level VM system not knowing anything about existing 791 * references. 792 */ 793 } else if (((m->flags & PG_REFERENCED) == 0) && 794 (actcount = pmap_ts_referenced(m))) { 795 vm_page_activate(m); 796 vm_page_unlock_queues(); 797 m->act_count += (actcount + ACT_ADVANCE); 798 continue; 799 } 800 801 /* 802 * If the upper level VM system knows about any page 803 * references, we activate the page. We also set the 804 * "activation count" higher than normal so that we will less 805 * likely place pages back onto the inactive queue again. 806 */ 807 if ((m->flags & PG_REFERENCED) != 0) { 808 vm_page_flag_clear(m, PG_REFERENCED); 809 actcount = pmap_ts_referenced(m); 810 vm_page_activate(m); 811 vm_page_unlock_queues(); 812 m->act_count += (actcount + ACT_ADVANCE + 1); 813 continue; 814 } 815 816 /* 817 * If the upper level VM system doesn't know anything about 818 * the page being dirty, we have to check for it again. As 819 * far as the VM code knows, any partially dirty pages are 820 * fully dirty. 821 */ 822 if (m->dirty == 0) { 823 vm_page_test_dirty(m); 824 } else { 825 vm_page_dirty(m); 826 } 827 vm_page_unlock_queues(); 828 829 /* 830 * Invalid pages can be easily freed 831 */ 832 if (m->valid == 0) { 833 vm_page_lock_queues(); 834 vm_pageout_page_free(m); 835 vm_page_unlock_queues(); 836 --page_shortage; 837 838 /* 839 * Clean pages can be placed onto the cache queue. This 840 * effectively frees them. 841 */ 842 } else if (m->dirty == 0) { 843 vm_page_lock_queues(); 844 vm_page_cache(m); 845 vm_page_unlock_queues(); 846 --page_shortage; 847 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) { 848 /* 849 * Dirty pages need to be paged out, but flushing 850 * a page is extremely expensive verses freeing 851 * a clean page. Rather then artificially limiting 852 * the number of pages we can flush, we instead give 853 * dirty pages extra priority on the inactive queue 854 * by forcing them to be cycled through the queue 855 * twice before being flushed, after which the 856 * (now clean) page will cycle through once more 857 * before being freed. This significantly extends 858 * the thrash point for a heavily loaded machine. 859 */ 860 vm_page_lock_queues(); 861 vm_page_flag_set(m, PG_WINATCFLS); 862 vm_pageq_requeue(m); 863 vm_page_unlock_queues(); 864 } else if (maxlaunder > 0) { 865 /* 866 * We always want to try to flush some dirty pages if 867 * we encounter them, to keep the system stable. 868 * Normally this number is small, but under extreme 869 * pressure where there are insufficient clean pages 870 * on the inactive queue, we may have to go all out. 871 */ 872 int swap_pageouts_ok; 873 struct vnode *vp = NULL; 874 struct mount *mp; 875 876 object = m->object; 877 878 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 879 swap_pageouts_ok = 1; 880 } else { 881 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 882 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 883 vm_page_count_min()); 884 885 } 886 887 /* 888 * We don't bother paging objects that are "dead". 889 * Those objects are in a "rundown" state. 890 */ 891 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 892 vm_pageq_requeue(m); 893 continue; 894 } 895 896 /* 897 * The object is already known NOT to be dead. It 898 * is possible for the vget() to block the whole 899 * pageout daemon, but the new low-memory handling 900 * code should prevent it. 901 * 902 * The previous code skipped locked vnodes and, worse, 903 * reordered pages in the queue. This results in 904 * completely non-deterministic operation and, on a 905 * busy system, can lead to extremely non-optimal 906 * pageouts. For example, it can cause clean pages 907 * to be freed and dirty pages to be moved to the end 908 * of the queue. Since dirty pages are also moved to 909 * the end of the queue once-cleaned, this gives 910 * way too large a weighting to defering the freeing 911 * of dirty pages. 912 * 913 * We can't wait forever for the vnode lock, we might 914 * deadlock due to a vn_read() getting stuck in 915 * vm_wait while holding this vnode. We skip the 916 * vnode if we can't get it in a reasonable amount 917 * of time. 918 */ 919 if (object->type == OBJT_VNODE) { 920 vp = object->handle; 921 922 mp = NULL; 923 if (vp->v_type == VREG) 924 vn_start_write(vp, &mp, V_NOWAIT); 925 if (vget(vp, LK_EXCLUSIVE|LK_TIMELOCK, curthread)) { 926 ++pageout_lock_miss; 927 vn_finished_write(mp); 928 if (object->flags & OBJ_MIGHTBEDIRTY) 929 vnodes_skipped++; 930 continue; 931 } 932 933 /* 934 * The page might have been moved to another 935 * queue during potential blocking in vget() 936 * above. The page might have been freed and 937 * reused for another vnode. The object might 938 * have been reused for another vnode. 939 */ 940 if (m->queue != PQ_INACTIVE || 941 m->object != object || 942 object->handle != vp) { 943 if (object->flags & OBJ_MIGHTBEDIRTY) 944 vnodes_skipped++; 945 vput(vp); 946 vn_finished_write(mp); 947 continue; 948 } 949 950 /* 951 * The page may have been busied during the 952 * blocking in vput(); We don't move the 953 * page back onto the end of the queue so that 954 * statistics are more correct if we don't. 955 */ 956 if (m->busy || (m->flags & PG_BUSY)) { 957 vput(vp); 958 vn_finished_write(mp); 959 continue; 960 } 961 962 /* 963 * If the page has become held it might 964 * be undergoing I/O, so skip it 965 */ 966 if (m->hold_count) { 967 vm_pageq_requeue(m); 968 if (object->flags & OBJ_MIGHTBEDIRTY) 969 vnodes_skipped++; 970 vput(vp); 971 vn_finished_write(mp); 972 continue; 973 } 974 } 975 976 /* 977 * If a page is dirty, then it is either being washed 978 * (but not yet cleaned) or it is still in the 979 * laundry. If it is still in the laundry, then we 980 * start the cleaning operation. 981 * 982 * This operation may cluster, invalidating the 'next' 983 * pointer. To prevent an inordinate number of 984 * restarts we use our marker to remember our place. 985 * 986 * decrement page_shortage on success to account for 987 * the (future) cleaned page. Otherwise we could wind 988 * up laundering or cleaning too many pages. 989 */ 990 vm_page_lock_queues(); 991 s = splvm(); 992 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, m, &marker, pageq); 993 splx(s); 994 if (vm_pageout_clean(m) != 0) { 995 --page_shortage; 996 --maxlaunder; 997 } 998 s = splvm(); 999 next = TAILQ_NEXT(&marker, pageq); 1000 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, &marker, pageq); 1001 splx(s); 1002 vm_page_unlock_queues(); 1003 if (vp) { 1004 vput(vp); 1005 vn_finished_write(mp); 1006 } 1007 } 1008 } 1009 1010 /* 1011 * Compute the number of pages we want to try to move from the 1012 * active queue to the inactive queue. 1013 */ 1014 page_shortage = vm_paging_target() + 1015 cnt.v_inactive_target - cnt.v_inactive_count; 1016 page_shortage += addl_page_shortage; 1017 1018 vm_page_lock_queues(); 1019 /* 1020 * Scan the active queue for things we can deactivate. We nominally 1021 * track the per-page activity counter and use it to locate 1022 * deactivation candidates. 1023 */ 1024 pcount = cnt.v_active_count; 1025 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1026 1027 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) { 1028 1029 /* 1030 * This is a consistency check, and should likely be a panic 1031 * or warning. 1032 */ 1033 if (m->queue != PQ_ACTIVE) { 1034 break; 1035 } 1036 1037 next = TAILQ_NEXT(m, pageq); 1038 /* 1039 * Don't deactivate pages that are busy. 1040 */ 1041 if ((m->busy != 0) || 1042 (m->flags & PG_BUSY) || 1043 (m->hold_count != 0)) { 1044 vm_pageq_requeue(m); 1045 m = next; 1046 continue; 1047 } 1048 1049 /* 1050 * The count for pagedaemon pages is done after checking the 1051 * page for eligibility... 1052 */ 1053 cnt.v_pdpages++; 1054 1055 /* 1056 * Check to see "how much" the page has been used. 1057 */ 1058 actcount = 0; 1059 if (m->object->ref_count != 0) { 1060 if (m->flags & PG_REFERENCED) { 1061 actcount += 1; 1062 } 1063 actcount += pmap_ts_referenced(m); 1064 if (actcount) { 1065 m->act_count += ACT_ADVANCE + actcount; 1066 if (m->act_count > ACT_MAX) 1067 m->act_count = ACT_MAX; 1068 } 1069 } 1070 1071 /* 1072 * Since we have "tested" this bit, we need to clear it now. 1073 */ 1074 vm_page_flag_clear(m, PG_REFERENCED); 1075 1076 /* 1077 * Only if an object is currently being used, do we use the 1078 * page activation count stats. 1079 */ 1080 if (actcount && (m->object->ref_count != 0)) { 1081 vm_pageq_requeue(m); 1082 } else { 1083 m->act_count -= min(m->act_count, ACT_DECLINE); 1084 if (vm_pageout_algorithm || 1085 m->object->ref_count == 0 || 1086 m->act_count == 0) { 1087 page_shortage--; 1088 if (m->object->ref_count == 0) { 1089 pmap_remove_all(m); 1090 if (m->dirty == 0) 1091 vm_page_cache(m); 1092 else 1093 vm_page_deactivate(m); 1094 } else { 1095 vm_page_deactivate(m); 1096 } 1097 } else { 1098 vm_pageq_requeue(m); 1099 } 1100 } 1101 m = next; 1102 } 1103 s = splvm(); 1104 1105 /* 1106 * We try to maintain some *really* free pages, this allows interrupt 1107 * code to be guaranteed space. Since both cache and free queues 1108 * are considered basically 'free', moving pages from cache to free 1109 * does not effect other calculations. 1110 */ 1111 while (cnt.v_free_count < cnt.v_free_reserved) { 1112 static int cache_rover = 0; 1113 m = vm_pageq_find(PQ_CACHE, cache_rover, FALSE); 1114 if (!m) 1115 break; 1116 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || 1117 m->busy || 1118 m->hold_count || 1119 m->wire_count) { 1120#ifdef INVARIANTS 1121 printf("Warning: busy page %p found in cache\n", m); 1122#endif 1123 vm_page_deactivate(m); 1124 continue; 1125 } 1126 cache_rover = (cache_rover + PQ_PRIME2) & PQ_L2_MASK; 1127 vm_pageout_page_free(m); 1128 } 1129 splx(s); 1130 vm_page_unlock_queues(); 1131#if !defined(NO_SWAPPING) 1132 /* 1133 * Idle process swapout -- run once per second. 1134 */ 1135 if (vm_swap_idle_enabled) { 1136 static long lsec; 1137 if (time_second != lsec) { 1138 vm_pageout_req_swapout |= VM_SWAP_IDLE; 1139 vm_req_vmdaemon(); 1140 lsec = time_second; 1141 } 1142 } 1143#endif 1144 1145 /* 1146 * If we didn't get enough free pages, and we have skipped a vnode 1147 * in a writeable object, wakeup the sync daemon. And kick swapout 1148 * if we did not get enough free pages. 1149 */ 1150 if (vm_paging_target() > 0) { 1151 if (vnodes_skipped && vm_page_count_min()) 1152 (void) speedup_syncer(); 1153#if !defined(NO_SWAPPING) 1154 if (vm_swap_enabled && vm_page_count_target()) { 1155 vm_req_vmdaemon(); 1156 vm_pageout_req_swapout |= VM_SWAP_NORMAL; 1157 } 1158#endif 1159 } 1160 1161 /* 1162 * If we are out of swap and were not able to reach our paging 1163 * target, kill the largest process. 1164 * 1165 * We keep the process bigproc locked once we find it to keep anyone 1166 * from messing with it; however, there is a possibility of 1167 * deadlock if process B is bigproc and one of it's child processes 1168 * attempts to propagate a signal to B while we are waiting for A's 1169 * lock while walking this list. To avoid this, we don't block on 1170 * the process lock but just skip a process if it is already locked. 1171 */ 1172 if ((vm_swap_size < 64 && vm_page_count_min()) || 1173 (swap_pager_full && vm_paging_target() > 0)) { 1174#if 0 1175 if ((vm_swap_size < 64 || swap_pager_full) && vm_page_count_min()) { 1176#endif 1177 bigproc = NULL; 1178 bigsize = 0; 1179 sx_slock(&allproc_lock); 1180 FOREACH_PROC_IN_SYSTEM(p) { 1181 int breakout; 1182 /* 1183 * If this process is already locked, skip it. 1184 */ 1185 if (PROC_TRYLOCK(p) == 0) 1186 continue; 1187 /* 1188 * if this is a system process, skip it 1189 */ 1190 if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) || 1191 ((p->p_pid < 48) && (vm_swap_size != 0))) { 1192 PROC_UNLOCK(p); 1193 continue; 1194 } 1195 /* 1196 * if the process is in a non-running type state, 1197 * don't touch it. Check all the threads individually. 1198 */ 1199 mtx_lock_spin(&sched_lock); 1200 breakout = 0; 1201 FOREACH_THREAD_IN_PROC(p, td) { 1202 if (!TD_ON_RUNQ(td) && 1203 !TD_IS_RUNNING(td) && 1204 !TD_IS_SLEEPING(td)) { 1205 breakout = 1; 1206 break; 1207 } 1208 } 1209 if (breakout) { 1210 mtx_unlock_spin(&sched_lock); 1211 PROC_UNLOCK(p); 1212 continue; 1213 } 1214 mtx_unlock_spin(&sched_lock); 1215 /* 1216 * get the process size 1217 */ 1218 size = vmspace_resident_count(p->p_vmspace) + 1219 vmspace_swap_count(p->p_vmspace); 1220 /* 1221 * if the this process is bigger than the biggest one 1222 * remember it. 1223 */ 1224 if (size > bigsize) { 1225 if (bigproc != NULL) 1226 PROC_UNLOCK(bigproc); 1227 bigproc = p; 1228 bigsize = size; 1229 } else 1230 PROC_UNLOCK(p); 1231 } 1232 sx_sunlock(&allproc_lock); 1233 if (bigproc != NULL) { 1234 struct ksegrp *kg; 1235 killproc(bigproc, "out of swap space"); 1236 mtx_lock_spin(&sched_lock); 1237 FOREACH_KSEGRP_IN_PROC(bigproc, kg) { 1238 sched_nice(kg, PRIO_MIN); /* XXXKSE ??? */ 1239 } 1240 mtx_unlock_spin(&sched_lock); 1241 PROC_UNLOCK(bigproc); 1242 wakeup(&cnt.v_free_count); 1243 } 1244 } 1245} 1246 1247/* 1248 * This routine tries to maintain the pseudo LRU active queue, 1249 * so that during long periods of time where there is no paging, 1250 * that some statistic accumulation still occurs. This code 1251 * helps the situation where paging just starts to occur. 1252 */ 1253static void 1254vm_pageout_page_stats() 1255{ 1256 vm_page_t m,next; 1257 int pcount,tpcount; /* Number of pages to check */ 1258 static int fullintervalcount = 0; 1259 int page_shortage; 1260 int s0; 1261 1262 page_shortage = 1263 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) - 1264 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count); 1265 1266 if (page_shortage <= 0) 1267 return; 1268 1269 s0 = splvm(); 1270 vm_page_lock_queues(); 1271 pcount = cnt.v_active_count; 1272 fullintervalcount += vm_pageout_stats_interval; 1273 if (fullintervalcount < vm_pageout_full_stats_interval) { 1274 tpcount = (vm_pageout_stats_max * cnt.v_active_count) / cnt.v_page_count; 1275 if (pcount > tpcount) 1276 pcount = tpcount; 1277 } else { 1278 fullintervalcount = 0; 1279 } 1280 1281 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1282 while ((m != NULL) && (pcount-- > 0)) { 1283 int actcount; 1284 1285 if (m->queue != PQ_ACTIVE) { 1286 break; 1287 } 1288 1289 next = TAILQ_NEXT(m, pageq); 1290 /* 1291 * Don't deactivate pages that are busy. 1292 */ 1293 if ((m->busy != 0) || 1294 (m->flags & PG_BUSY) || 1295 (m->hold_count != 0)) { 1296 vm_pageq_requeue(m); 1297 m = next; 1298 continue; 1299 } 1300 1301 actcount = 0; 1302 if (m->flags & PG_REFERENCED) { 1303 vm_page_flag_clear(m, PG_REFERENCED); 1304 actcount += 1; 1305 } 1306 1307 actcount += pmap_ts_referenced(m); 1308 if (actcount) { 1309 m->act_count += ACT_ADVANCE + actcount; 1310 if (m->act_count > ACT_MAX) 1311 m->act_count = ACT_MAX; 1312 vm_pageq_requeue(m); 1313 } else { 1314 if (m->act_count == 0) { 1315 /* 1316 * We turn off page access, so that we have 1317 * more accurate RSS stats. We don't do this 1318 * in the normal page deactivation when the 1319 * system is loaded VM wise, because the 1320 * cost of the large number of page protect 1321 * operations would be higher than the value 1322 * of doing the operation. 1323 */ 1324 pmap_remove_all(m); 1325 vm_page_deactivate(m); 1326 } else { 1327 m->act_count -= min(m->act_count, ACT_DECLINE); 1328 vm_pageq_requeue(m); 1329 } 1330 } 1331 1332 m = next; 1333 } 1334 vm_page_unlock_queues(); 1335 splx(s0); 1336} 1337 1338static int 1339vm_pageout_free_page_calc(count) 1340vm_size_t count; 1341{ 1342 if (count < cnt.v_page_count) 1343 return 0; 1344 /* 1345 * free_reserved needs to include enough for the largest swap pager 1346 * structures plus enough for any pv_entry structs when paging. 1347 */ 1348 if (cnt.v_page_count > 1024) 1349 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200; 1350 else 1351 cnt.v_free_min = 4; 1352 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1353 cnt.v_interrupt_free_min; 1354 cnt.v_free_reserved = vm_pageout_page_count + 1355 cnt.v_pageout_free_min + (count / 768) + PQ_L2_SIZE; 1356 cnt.v_free_severe = cnt.v_free_min / 2; 1357 cnt.v_free_min += cnt.v_free_reserved; 1358 cnt.v_free_severe += cnt.v_free_reserved; 1359 return 1; 1360} 1361 1362/* 1363 * vm_pageout is the high level pageout daemon. 1364 */ 1365static void 1366vm_pageout() 1367{ 1368 int pass; 1369 1370 mtx_lock(&Giant); 1371 1372 /* 1373 * Initialize some paging parameters. 1374 */ 1375 cnt.v_interrupt_free_min = 2; 1376 if (cnt.v_page_count < 2000) 1377 vm_pageout_page_count = 8; 1378 1379 vm_pageout_free_page_calc(cnt.v_page_count); 1380 /* 1381 * v_free_target and v_cache_min control pageout hysteresis. Note 1382 * that these are more a measure of the VM cache queue hysteresis 1383 * then the VM free queue. Specifically, v_free_target is the 1384 * high water mark (free+cache pages). 1385 * 1386 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the 1387 * low water mark, while v_free_min is the stop. v_cache_min must 1388 * be big enough to handle memory needs while the pageout daemon 1389 * is signalled and run to free more pages. 1390 */ 1391 if (cnt.v_free_count > 6144) 1392 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved; 1393 else 1394 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved; 1395 1396 if (cnt.v_free_count > 2048) { 1397 cnt.v_cache_min = cnt.v_free_target; 1398 cnt.v_cache_max = 2 * cnt.v_cache_min; 1399 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2; 1400 } else { 1401 cnt.v_cache_min = 0; 1402 cnt.v_cache_max = 0; 1403 cnt.v_inactive_target = cnt.v_free_count / 4; 1404 } 1405 if (cnt.v_inactive_target > cnt.v_free_count / 3) 1406 cnt.v_inactive_target = cnt.v_free_count / 3; 1407 1408 /* XXX does not really belong here */ 1409 if (vm_page_max_wired == 0) 1410 vm_page_max_wired = cnt.v_free_count / 3; 1411 1412 if (vm_pageout_stats_max == 0) 1413 vm_pageout_stats_max = cnt.v_free_target; 1414 1415 /* 1416 * Set interval in seconds for stats scan. 1417 */ 1418 if (vm_pageout_stats_interval == 0) 1419 vm_pageout_stats_interval = 5; 1420 if (vm_pageout_full_stats_interval == 0) 1421 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4; 1422 1423 /* 1424 * Set maximum free per pass 1425 */ 1426 if (vm_pageout_stats_free_max == 0) 1427 vm_pageout_stats_free_max = 5; 1428 1429 swap_pager_swap_init(); 1430 pass = 0; 1431 /* 1432 * The pageout daemon is never done, so loop forever. 1433 */ 1434 while (TRUE) { 1435 int error; 1436 int s = splvm(); 1437 1438 /* 1439 * If we have enough free memory, wakeup waiters. Do 1440 * not clear vm_pages_needed until we reach our target, 1441 * otherwise we may be woken up over and over again and 1442 * waste a lot of cpu. 1443 */ 1444 if (vm_pages_needed && !vm_page_count_min()) { 1445 if (vm_paging_needed() <= 0) 1446 vm_pages_needed = 0; 1447 wakeup(&cnt.v_free_count); 1448 } 1449 if (vm_pages_needed) { 1450 /* 1451 * Still not done, take a second pass without waiting 1452 * (unlimited dirty cleaning), otherwise sleep a bit 1453 * and try again. 1454 */ 1455 ++pass; 1456 if (pass > 1) 1457 tsleep(&vm_pages_needed, PVM, 1458 "psleep", hz/2); 1459 } else { 1460 /* 1461 * Good enough, sleep & handle stats. Prime the pass 1462 * for the next run. 1463 */ 1464 if (pass > 1) 1465 pass = 1; 1466 else 1467 pass = 0; 1468 error = tsleep(&vm_pages_needed, PVM, 1469 "psleep", vm_pageout_stats_interval * hz); 1470 if (error && !vm_pages_needed) { 1471 splx(s); 1472 pass = 0; 1473 vm_pageout_page_stats(); 1474 continue; 1475 } 1476 } 1477 1478 if (vm_pages_needed) 1479 cnt.v_pdwakeups++; 1480 splx(s); 1481 vm_pageout_scan(pass); 1482 vm_pageout_deficit = 0; 1483 } 1484} 1485 1486void 1487pagedaemon_wakeup() 1488{ 1489 if (!vm_pages_needed && curthread->td_proc != pageproc) { 1490 vm_pages_needed++; 1491 wakeup(&vm_pages_needed); 1492 } 1493} 1494 1495#if !defined(NO_SWAPPING) 1496static void 1497vm_req_vmdaemon() 1498{ 1499 static int lastrun = 0; 1500 1501 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1502 wakeup(&vm_daemon_needed); 1503 lastrun = ticks; 1504 } 1505} 1506 1507static void 1508vm_daemon() 1509{ 1510 struct proc *p; 1511 int breakout; 1512 struct thread *td; 1513 1514 mtx_lock(&Giant); 1515 while (TRUE) { 1516 tsleep(&vm_daemon_needed, PPAUSE, "psleep", 0); 1517 if (vm_pageout_req_swapout) { 1518 swapout_procs(vm_pageout_req_swapout); 1519 vm_pageout_req_swapout = 0; 1520 } 1521 /* 1522 * scan the processes for exceeding their rlimits or if 1523 * process is swapped out -- deactivate pages 1524 */ 1525 sx_slock(&allproc_lock); 1526 LIST_FOREACH(p, &allproc, p_list) { 1527 vm_pindex_t limit, size; 1528 1529 /* 1530 * if this is a system process or if we have already 1531 * looked at this process, skip it. 1532 */ 1533 if (p->p_flag & (P_SYSTEM | P_WEXIT)) { 1534 continue; 1535 } 1536 /* 1537 * if the process is in a non-running type state, 1538 * don't touch it. 1539 */ 1540 mtx_lock_spin(&sched_lock); 1541 breakout = 0; 1542 FOREACH_THREAD_IN_PROC(p, td) { 1543 if (!TD_ON_RUNQ(td) && 1544 !TD_IS_RUNNING(td) && 1545 !TD_IS_SLEEPING(td)) { 1546 breakout = 1; 1547 break; 1548 } 1549 } 1550 if (breakout) { 1551 mtx_unlock_spin(&sched_lock); 1552 continue; 1553 } 1554 /* 1555 * get a limit 1556 */ 1557 limit = OFF_TO_IDX( 1558 qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur, 1559 p->p_rlimit[RLIMIT_RSS].rlim_max)); 1560 1561 /* 1562 * let processes that are swapped out really be 1563 * swapped out set the limit to nothing (will force a 1564 * swap-out.) 1565 */ 1566 if ((p->p_sflag & PS_INMEM) == 0) 1567 limit = 0; /* XXX */ 1568 mtx_unlock_spin(&sched_lock); 1569 1570 size = vmspace_resident_count(p->p_vmspace); 1571 if (limit >= 0 && size >= limit) { 1572 vm_pageout_map_deactivate_pages( 1573 &p->p_vmspace->vm_map, limit); 1574 } 1575 } 1576 sx_sunlock(&allproc_lock); 1577 } 1578} 1579#endif /* !defined(NO_SWAPPING) */ 1580