vm_pageout.c revision 107185
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 107185 2002-11-23 08:08:54Z 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 pmap_remove(vm_map_pmap(map), vm_map_min(map), 603 vm_map_max(map)); 604 vm_map_unlock(map); 605 return; 606} 607#endif /* !defined(NO_SWAPPING) */ 608 609/* 610 * Don't try to be fancy - being fancy can lead to VOP_LOCK's and therefore 611 * to vnode deadlocks. We only do it for OBJT_DEFAULT and OBJT_SWAP objects 612 * which we know can be trivially freed. 613 */ 614void 615vm_pageout_page_free(vm_page_t m) 616{ 617 vm_object_t object = m->object; 618 int type = object->type; 619 620 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 621 if (type == OBJT_SWAP || type == OBJT_DEFAULT) 622 vm_object_reference(object); 623 vm_page_busy(m); 624 pmap_remove_all(m); 625 vm_page_free(m); 626 cnt.v_dfree++; 627 if (type == OBJT_SWAP || type == OBJT_DEFAULT) 628 vm_object_deallocate(object); 629} 630 631/* 632 * This routine is very drastic, but can save the system 633 * in a pinch. 634 */ 635static void 636vm_pageout_pmap_collect(void) 637{ 638 int i; 639 vm_page_t m; 640 static int warningdone; 641 642 if (pmap_pagedaemon_waken == 0) 643 return; 644 if (warningdone < 5) { 645 printf("collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n"); 646 warningdone++; 647 } 648 vm_page_lock_queues(); 649 for (i = 0; i < vm_page_array_size; i++) { 650 m = &vm_page_array[i]; 651 if (m->wire_count || m->hold_count || m->busy || 652 (m->flags & (PG_BUSY | PG_UNMANAGED))) 653 continue; 654 pmap_remove_all(m); 655 } 656 vm_page_unlock_queues(); 657 pmap_pagedaemon_waken = 0; 658} 659 660/* 661 * vm_pageout_scan does the dirty work for the pageout daemon. 662 */ 663static void 664vm_pageout_scan(int pass) 665{ 666 vm_page_t m, next; 667 struct vm_page marker; 668 int save_page_shortage; 669 int save_inactive_count; 670 int page_shortage, maxscan, pcount; 671 int addl_page_shortage, addl_page_shortage_init; 672 struct proc *p, *bigproc; 673 vm_offset_t size, bigsize; 674 vm_object_t object; 675 int actcount; 676 int vnodes_skipped = 0; 677 int maxlaunder; 678 int s; 679 struct thread *td; 680 681 GIANT_REQUIRED; 682 /* 683 * Decrease registered cache sizes. 684 */ 685 EVENTHANDLER_INVOKE(vm_lowmem, 0); 686 /* 687 * We do this explicitly after the caches have been drained above. 688 */ 689 uma_reclaim(); 690 /* 691 * Do whatever cleanup that the pmap code can. 692 */ 693 vm_pageout_pmap_collect(); 694 695 addl_page_shortage_init = vm_pageout_deficit; 696 vm_pageout_deficit = 0; 697 698 /* 699 * Calculate the number of pages we want to either free or move 700 * to the cache. 701 */ 702 page_shortage = vm_paging_target() + addl_page_shortage_init; 703 save_page_shortage = page_shortage; 704 save_inactive_count = cnt.v_inactive_count; 705 706 /* 707 * Initialize our marker 708 */ 709 bzero(&marker, sizeof(marker)); 710 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER; 711 marker.queue = PQ_INACTIVE; 712 marker.wire_count = 1; 713 714 /* 715 * Start scanning the inactive queue for pages we can move to the 716 * cache or free. The scan will stop when the target is reached or 717 * we have scanned the entire inactive queue. Note that m->act_count 718 * is not used to form decisions for the inactive queue, only for the 719 * active queue. 720 * 721 * maxlaunder limits the number of dirty pages we flush per scan. 722 * For most systems a smaller value (16 or 32) is more robust under 723 * extreme memory and disk pressure because any unnecessary writes 724 * to disk can result in extreme performance degredation. However, 725 * systems with excessive dirty pages (especially when MAP_NOSYNC is 726 * used) will die horribly with limited laundering. If the pageout 727 * daemon cannot clean enough pages in the first pass, we let it go 728 * all out in succeeding passes. 729 */ 730 if ((maxlaunder = vm_max_launder) <= 1) 731 maxlaunder = 1; 732 if (pass) 733 maxlaunder = 10000; 734rescan0: 735 addl_page_shortage = addl_page_shortage_init; 736 maxscan = cnt.v_inactive_count; 737 738 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl); 739 m != NULL && maxscan-- > 0 && page_shortage > 0; 740 m = next) { 741 742 cnt.v_pdpages++; 743 744 if (m->queue != PQ_INACTIVE) { 745 goto rescan0; 746 } 747 748 next = TAILQ_NEXT(m, pageq); 749 750 /* 751 * skip marker pages 752 */ 753 if (m->flags & PG_MARKER) 754 continue; 755 756 /* 757 * A held page may be undergoing I/O, so skip it. 758 */ 759 if (m->hold_count) { 760 vm_pageq_requeue(m); 761 addl_page_shortage++; 762 continue; 763 } 764 /* 765 * Don't mess with busy pages, keep in the front of the 766 * queue, most likely are being paged out. 767 */ 768 if (m->busy || (m->flags & PG_BUSY)) { 769 addl_page_shortage++; 770 continue; 771 } 772 773 /* 774 * If the object is not being used, we ignore previous 775 * references. 776 */ 777 if (m->object->ref_count == 0) { 778 vm_page_flag_clear(m, PG_REFERENCED); 779 pmap_clear_reference(m); 780 781 /* 782 * Otherwise, if the page has been referenced while in the 783 * inactive queue, we bump the "activation count" upwards, 784 * making it less likely that the page will be added back to 785 * the inactive queue prematurely again. Here we check the 786 * page tables (or emulated bits, if any), given the upper 787 * level VM system not knowing anything about existing 788 * references. 789 */ 790 } else if (((m->flags & PG_REFERENCED) == 0) && 791 (actcount = pmap_ts_referenced(m))) { 792 vm_page_lock_queues(); 793 vm_page_activate(m); 794 vm_page_unlock_queues(); 795 m->act_count += (actcount + ACT_ADVANCE); 796 continue; 797 } 798 799 /* 800 * If the upper level VM system knows about any page 801 * references, we activate the page. We also set the 802 * "activation count" higher than normal so that we will less 803 * likely place pages back onto the inactive queue again. 804 */ 805 if ((m->flags & PG_REFERENCED) != 0) { 806 vm_page_flag_clear(m, PG_REFERENCED); 807 actcount = pmap_ts_referenced(m); 808 vm_page_lock_queues(); 809 vm_page_activate(m); 810 vm_page_unlock_queues(); 811 m->act_count += (actcount + ACT_ADVANCE + 1); 812 continue; 813 } 814 815 /* 816 * If the upper level VM system doesn't know anything about 817 * the page being dirty, we have to check for it again. As 818 * far as the VM code knows, any partially dirty pages are 819 * fully dirty. 820 */ 821 if (m->dirty == 0) { 822 vm_page_test_dirty(m); 823 } else { 824 vm_page_dirty(m); 825 } 826 827 /* 828 * Invalid pages can be easily freed 829 */ 830 if (m->valid == 0) { 831 vm_page_lock_queues(); 832 vm_pageout_page_free(m); 833 vm_page_unlock_queues(); 834 --page_shortage; 835 836 /* 837 * Clean pages can be placed onto the cache queue. This 838 * effectively frees them. 839 */ 840 } else if (m->dirty == 0) { 841 vm_page_lock_queues(); 842 vm_page_cache(m); 843 vm_page_unlock_queues(); 844 --page_shortage; 845 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) { 846 /* 847 * Dirty pages need to be paged out, but flushing 848 * a page is extremely expensive verses freeing 849 * a clean page. Rather then artificially limiting 850 * the number of pages we can flush, we instead give 851 * dirty pages extra priority on the inactive queue 852 * by forcing them to be cycled through the queue 853 * twice before being flushed, after which the 854 * (now clean) page will cycle through once more 855 * before being freed. This significantly extends 856 * the thrash point for a heavily loaded machine. 857 */ 858 vm_page_flag_set(m, PG_WINATCFLS); 859 vm_pageq_requeue(m); 860 } else if (maxlaunder > 0) { 861 /* 862 * We always want to try to flush some dirty pages if 863 * we encounter them, to keep the system stable. 864 * Normally this number is small, but under extreme 865 * pressure where there are insufficient clean pages 866 * on the inactive queue, we may have to go all out. 867 */ 868 int swap_pageouts_ok; 869 struct vnode *vp = NULL; 870 struct mount *mp; 871 872 object = m->object; 873 874 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 875 swap_pageouts_ok = 1; 876 } else { 877 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 878 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 879 vm_page_count_min()); 880 881 } 882 883 /* 884 * We don't bother paging objects that are "dead". 885 * Those objects are in a "rundown" state. 886 */ 887 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 888 vm_pageq_requeue(m); 889 continue; 890 } 891 892 /* 893 * The object is already known NOT to be dead. It 894 * is possible for the vget() to block the whole 895 * pageout daemon, but the new low-memory handling 896 * code should prevent it. 897 * 898 * The previous code skipped locked vnodes and, worse, 899 * reordered pages in the queue. This results in 900 * completely non-deterministic operation and, on a 901 * busy system, can lead to extremely non-optimal 902 * pageouts. For example, it can cause clean pages 903 * to be freed and dirty pages to be moved to the end 904 * of the queue. Since dirty pages are also moved to 905 * the end of the queue once-cleaned, this gives 906 * way too large a weighting to defering the freeing 907 * of dirty pages. 908 * 909 * We can't wait forever for the vnode lock, we might 910 * deadlock due to a vn_read() getting stuck in 911 * vm_wait while holding this vnode. We skip the 912 * vnode if we can't get it in a reasonable amount 913 * of time. 914 */ 915 if (object->type == OBJT_VNODE) { 916 vp = object->handle; 917 918 mp = NULL; 919 if (vp->v_type == VREG) 920 vn_start_write(vp, &mp, V_NOWAIT); 921 if (vget(vp, LK_EXCLUSIVE|LK_TIMELOCK, curthread)) { 922 ++pageout_lock_miss; 923 vn_finished_write(mp); 924 if (object->flags & OBJ_MIGHTBEDIRTY) 925 vnodes_skipped++; 926 continue; 927 } 928 929 /* 930 * The page might have been moved to another 931 * queue during potential blocking in vget() 932 * above. The page might have been freed and 933 * reused for another vnode. The object might 934 * have been reused for another vnode. 935 */ 936 if (m->queue != PQ_INACTIVE || 937 m->object != object || 938 object->handle != vp) { 939 if (object->flags & OBJ_MIGHTBEDIRTY) 940 vnodes_skipped++; 941 vput(vp); 942 vn_finished_write(mp); 943 continue; 944 } 945 946 /* 947 * The page may have been busied during the 948 * blocking in vput(); We don't move the 949 * page back onto the end of the queue so that 950 * statistics are more correct if we don't. 951 */ 952 if (m->busy || (m->flags & PG_BUSY)) { 953 vput(vp); 954 vn_finished_write(mp); 955 continue; 956 } 957 958 /* 959 * If the page has become held it might 960 * be undergoing I/O, so skip it 961 */ 962 if (m->hold_count) { 963 vm_pageq_requeue(m); 964 if (object->flags & OBJ_MIGHTBEDIRTY) 965 vnodes_skipped++; 966 vput(vp); 967 vn_finished_write(mp); 968 continue; 969 } 970 } 971 972 /* 973 * If a page is dirty, then it is either being washed 974 * (but not yet cleaned) or it is still in the 975 * laundry. If it is still in the laundry, then we 976 * start the cleaning operation. 977 * 978 * This operation may cluster, invalidating the 'next' 979 * pointer. To prevent an inordinate number of 980 * restarts we use our marker to remember our place. 981 * 982 * decrement page_shortage on success to account for 983 * the (future) cleaned page. Otherwise we could wind 984 * up laundering or cleaning too many pages. 985 */ 986 vm_page_lock_queues(); 987 s = splvm(); 988 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, m, &marker, pageq); 989 splx(s); 990 if (vm_pageout_clean(m) != 0) { 991 --page_shortage; 992 --maxlaunder; 993 } 994 s = splvm(); 995 next = TAILQ_NEXT(&marker, pageq); 996 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, &marker, pageq); 997 splx(s); 998 vm_page_unlock_queues(); 999 if (vp) { 1000 vput(vp); 1001 vn_finished_write(mp); 1002 } 1003 } 1004 } 1005 1006 /* 1007 * Compute the number of pages we want to try to move from the 1008 * active queue to the inactive queue. 1009 */ 1010 page_shortage = vm_paging_target() + 1011 cnt.v_inactive_target - cnt.v_inactive_count; 1012 page_shortage += addl_page_shortage; 1013 1014 vm_page_lock_queues(); 1015 /* 1016 * Scan the active queue for things we can deactivate. We nominally 1017 * track the per-page activity counter and use it to locate 1018 * deactivation candidates. 1019 */ 1020 pcount = cnt.v_active_count; 1021 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1022 1023 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) { 1024 1025 /* 1026 * This is a consistency check, and should likely be a panic 1027 * or warning. 1028 */ 1029 if (m->queue != PQ_ACTIVE) { 1030 break; 1031 } 1032 1033 next = TAILQ_NEXT(m, pageq); 1034 /* 1035 * Don't deactivate pages that are busy. 1036 */ 1037 if ((m->busy != 0) || 1038 (m->flags & PG_BUSY) || 1039 (m->hold_count != 0)) { 1040 vm_pageq_requeue(m); 1041 m = next; 1042 continue; 1043 } 1044 1045 /* 1046 * The count for pagedaemon pages is done after checking the 1047 * page for eligibility... 1048 */ 1049 cnt.v_pdpages++; 1050 1051 /* 1052 * Check to see "how much" the page has been used. 1053 */ 1054 actcount = 0; 1055 if (m->object->ref_count != 0) { 1056 if (m->flags & PG_REFERENCED) { 1057 actcount += 1; 1058 } 1059 actcount += pmap_ts_referenced(m); 1060 if (actcount) { 1061 m->act_count += ACT_ADVANCE + actcount; 1062 if (m->act_count > ACT_MAX) 1063 m->act_count = ACT_MAX; 1064 } 1065 } 1066 1067 /* 1068 * Since we have "tested" this bit, we need to clear it now. 1069 */ 1070 vm_page_flag_clear(m, PG_REFERENCED); 1071 1072 /* 1073 * Only if an object is currently being used, do we use the 1074 * page activation count stats. 1075 */ 1076 if (actcount && (m->object->ref_count != 0)) { 1077 vm_pageq_requeue(m); 1078 } else { 1079 m->act_count -= min(m->act_count, ACT_DECLINE); 1080 if (vm_pageout_algorithm || 1081 m->object->ref_count == 0 || 1082 m->act_count == 0) { 1083 page_shortage--; 1084 if (m->object->ref_count == 0) { 1085 pmap_remove_all(m); 1086 if (m->dirty == 0) 1087 vm_page_cache(m); 1088 else 1089 vm_page_deactivate(m); 1090 } else { 1091 vm_page_deactivate(m); 1092 } 1093 } else { 1094 vm_pageq_requeue(m); 1095 } 1096 } 1097 m = next; 1098 } 1099 s = splvm(); 1100 1101 /* 1102 * We try to maintain some *really* free pages, this allows interrupt 1103 * code to be guaranteed space. Since both cache and free queues 1104 * are considered basically 'free', moving pages from cache to free 1105 * does not effect other calculations. 1106 */ 1107 while (cnt.v_free_count < cnt.v_free_reserved) { 1108 static int cache_rover = 0; 1109 m = vm_pageq_find(PQ_CACHE, cache_rover, FALSE); 1110 if (!m) 1111 break; 1112 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || 1113 m->busy || 1114 m->hold_count || 1115 m->wire_count) { 1116#ifdef INVARIANTS 1117 printf("Warning: busy page %p found in cache\n", m); 1118#endif 1119 vm_page_deactivate(m); 1120 continue; 1121 } 1122 cache_rover = (cache_rover + PQ_PRIME2) & PQ_L2_MASK; 1123 vm_pageout_page_free(m); 1124 } 1125 splx(s); 1126 vm_page_unlock_queues(); 1127#if !defined(NO_SWAPPING) 1128 /* 1129 * Idle process swapout -- run once per second. 1130 */ 1131 if (vm_swap_idle_enabled) { 1132 static long lsec; 1133 if (time_second != lsec) { 1134 vm_pageout_req_swapout |= VM_SWAP_IDLE; 1135 vm_req_vmdaemon(); 1136 lsec = time_second; 1137 } 1138 } 1139#endif 1140 1141 /* 1142 * If we didn't get enough free pages, and we have skipped a vnode 1143 * in a writeable object, wakeup the sync daemon. And kick swapout 1144 * if we did not get enough free pages. 1145 */ 1146 if (vm_paging_target() > 0) { 1147 if (vnodes_skipped && vm_page_count_min()) 1148 (void) speedup_syncer(); 1149#if !defined(NO_SWAPPING) 1150 if (vm_swap_enabled && vm_page_count_target()) { 1151 vm_req_vmdaemon(); 1152 vm_pageout_req_swapout |= VM_SWAP_NORMAL; 1153 } 1154#endif 1155 } 1156 1157 /* 1158 * If we are out of swap and were not able to reach our paging 1159 * target, kill the largest process. 1160 * 1161 * We keep the process bigproc locked once we find it to keep anyone 1162 * from messing with it; however, there is a possibility of 1163 * deadlock if process B is bigproc and one of it's child processes 1164 * attempts to propagate a signal to B while we are waiting for A's 1165 * lock while walking this list. To avoid this, we don't block on 1166 * the process lock but just skip a process if it is already locked. 1167 */ 1168 if ((vm_swap_size < 64 && vm_page_count_min()) || 1169 (swap_pager_full && vm_paging_target() > 0)) { 1170#if 0 1171 if ((vm_swap_size < 64 || swap_pager_full) && vm_page_count_min()) { 1172#endif 1173 bigproc = NULL; 1174 bigsize = 0; 1175 sx_slock(&allproc_lock); 1176 FOREACH_PROC_IN_SYSTEM(p) { 1177 int breakout; 1178 /* 1179 * If this process is already locked, skip it. 1180 */ 1181 if (PROC_TRYLOCK(p) == 0) 1182 continue; 1183 /* 1184 * if this is a system process, skip it 1185 */ 1186 if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) || 1187 ((p->p_pid < 48) && (vm_swap_size != 0))) { 1188 PROC_UNLOCK(p); 1189 continue; 1190 } 1191 /* 1192 * if the process is in a non-running type state, 1193 * don't touch it. Check all the threads individually. 1194 */ 1195 mtx_lock_spin(&sched_lock); 1196 breakout = 0; 1197 FOREACH_THREAD_IN_PROC(p, td) { 1198 if (!TD_ON_RUNQ(td) && 1199 !TD_IS_RUNNING(td) && 1200 !TD_IS_SLEEPING(td)) { 1201 breakout = 1; 1202 break; 1203 } 1204 } 1205 if (breakout) { 1206 mtx_unlock_spin(&sched_lock); 1207 PROC_UNLOCK(p); 1208 continue; 1209 } 1210 mtx_unlock_spin(&sched_lock); 1211 /* 1212 * get the process size 1213 */ 1214 size = vmspace_resident_count(p->p_vmspace) + 1215 vmspace_swap_count(p->p_vmspace); 1216 /* 1217 * if the this process is bigger than the biggest one 1218 * remember it. 1219 */ 1220 if (size > bigsize) { 1221 if (bigproc != NULL) 1222 PROC_UNLOCK(bigproc); 1223 bigproc = p; 1224 bigsize = size; 1225 } else 1226 PROC_UNLOCK(p); 1227 } 1228 sx_sunlock(&allproc_lock); 1229 if (bigproc != NULL) { 1230 struct ksegrp *kg; 1231 killproc(bigproc, "out of swap space"); 1232 mtx_lock_spin(&sched_lock); 1233 FOREACH_KSEGRP_IN_PROC(bigproc, kg) { 1234 sched_nice(kg, PRIO_MIN); /* XXXKSE ??? */ 1235 } 1236 mtx_unlock_spin(&sched_lock); 1237 PROC_UNLOCK(bigproc); 1238 wakeup(&cnt.v_free_count); 1239 } 1240 } 1241} 1242 1243/* 1244 * This routine tries to maintain the pseudo LRU active queue, 1245 * so that during long periods of time where there is no paging, 1246 * that some statistic accumulation still occurs. This code 1247 * helps the situation where paging just starts to occur. 1248 */ 1249static void 1250vm_pageout_page_stats() 1251{ 1252 vm_page_t m,next; 1253 int pcount,tpcount; /* Number of pages to check */ 1254 static int fullintervalcount = 0; 1255 int page_shortage; 1256 int s0; 1257 1258 page_shortage = 1259 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) - 1260 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count); 1261 1262 if (page_shortage <= 0) 1263 return; 1264 1265 s0 = splvm(); 1266 vm_page_lock_queues(); 1267 pcount = cnt.v_active_count; 1268 fullintervalcount += vm_pageout_stats_interval; 1269 if (fullintervalcount < vm_pageout_full_stats_interval) { 1270 tpcount = (vm_pageout_stats_max * cnt.v_active_count) / cnt.v_page_count; 1271 if (pcount > tpcount) 1272 pcount = tpcount; 1273 } else { 1274 fullintervalcount = 0; 1275 } 1276 1277 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1278 while ((m != NULL) && (pcount-- > 0)) { 1279 int actcount; 1280 1281 if (m->queue != PQ_ACTIVE) { 1282 break; 1283 } 1284 1285 next = TAILQ_NEXT(m, pageq); 1286 /* 1287 * Don't deactivate pages that are busy. 1288 */ 1289 if ((m->busy != 0) || 1290 (m->flags & PG_BUSY) || 1291 (m->hold_count != 0)) { 1292 vm_pageq_requeue(m); 1293 m = next; 1294 continue; 1295 } 1296 1297 actcount = 0; 1298 if (m->flags & PG_REFERENCED) { 1299 vm_page_flag_clear(m, PG_REFERENCED); 1300 actcount += 1; 1301 } 1302 1303 actcount += pmap_ts_referenced(m); 1304 if (actcount) { 1305 m->act_count += ACT_ADVANCE + actcount; 1306 if (m->act_count > ACT_MAX) 1307 m->act_count = ACT_MAX; 1308 vm_pageq_requeue(m); 1309 } else { 1310 if (m->act_count == 0) { 1311 /* 1312 * We turn off page access, so that we have 1313 * more accurate RSS stats. We don't do this 1314 * in the normal page deactivation when the 1315 * system is loaded VM wise, because the 1316 * cost of the large number of page protect 1317 * operations would be higher than the value 1318 * of doing the operation. 1319 */ 1320 pmap_remove_all(m); 1321 vm_page_deactivate(m); 1322 } else { 1323 m->act_count -= min(m->act_count, ACT_DECLINE); 1324 vm_pageq_requeue(m); 1325 } 1326 } 1327 1328 m = next; 1329 } 1330 vm_page_unlock_queues(); 1331 splx(s0); 1332} 1333 1334static int 1335vm_pageout_free_page_calc(count) 1336vm_size_t count; 1337{ 1338 if (count < cnt.v_page_count) 1339 return 0; 1340 /* 1341 * free_reserved needs to include enough for the largest swap pager 1342 * structures plus enough for any pv_entry structs when paging. 1343 */ 1344 if (cnt.v_page_count > 1024) 1345 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200; 1346 else 1347 cnt.v_free_min = 4; 1348 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1349 cnt.v_interrupt_free_min; 1350 cnt.v_free_reserved = vm_pageout_page_count + 1351 cnt.v_pageout_free_min + (count / 768) + PQ_L2_SIZE; 1352 cnt.v_free_severe = cnt.v_free_min / 2; 1353 cnt.v_free_min += cnt.v_free_reserved; 1354 cnt.v_free_severe += cnt.v_free_reserved; 1355 return 1; 1356} 1357 1358/* 1359 * vm_pageout is the high level pageout daemon. 1360 */ 1361static void 1362vm_pageout() 1363{ 1364 int pass; 1365 1366 mtx_lock(&Giant); 1367 1368 /* 1369 * Initialize some paging parameters. 1370 */ 1371 cnt.v_interrupt_free_min = 2; 1372 if (cnt.v_page_count < 2000) 1373 vm_pageout_page_count = 8; 1374 1375 vm_pageout_free_page_calc(cnt.v_page_count); 1376 /* 1377 * v_free_target and v_cache_min control pageout hysteresis. Note 1378 * that these are more a measure of the VM cache queue hysteresis 1379 * then the VM free queue. Specifically, v_free_target is the 1380 * high water mark (free+cache pages). 1381 * 1382 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the 1383 * low water mark, while v_free_min is the stop. v_cache_min must 1384 * be big enough to handle memory needs while the pageout daemon 1385 * is signalled and run to free more pages. 1386 */ 1387 if (cnt.v_free_count > 6144) 1388 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved; 1389 else 1390 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved; 1391 1392 if (cnt.v_free_count > 2048) { 1393 cnt.v_cache_min = cnt.v_free_target; 1394 cnt.v_cache_max = 2 * cnt.v_cache_min; 1395 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2; 1396 } else { 1397 cnt.v_cache_min = 0; 1398 cnt.v_cache_max = 0; 1399 cnt.v_inactive_target = cnt.v_free_count / 4; 1400 } 1401 if (cnt.v_inactive_target > cnt.v_free_count / 3) 1402 cnt.v_inactive_target = cnt.v_free_count / 3; 1403 1404 /* XXX does not really belong here */ 1405 if (vm_page_max_wired == 0) 1406 vm_page_max_wired = cnt.v_free_count / 3; 1407 1408 if (vm_pageout_stats_max == 0) 1409 vm_pageout_stats_max = cnt.v_free_target; 1410 1411 /* 1412 * Set interval in seconds for stats scan. 1413 */ 1414 if (vm_pageout_stats_interval == 0) 1415 vm_pageout_stats_interval = 5; 1416 if (vm_pageout_full_stats_interval == 0) 1417 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4; 1418 1419 /* 1420 * Set maximum free per pass 1421 */ 1422 if (vm_pageout_stats_free_max == 0) 1423 vm_pageout_stats_free_max = 5; 1424 1425 swap_pager_swap_init(); 1426 pass = 0; 1427 /* 1428 * The pageout daemon is never done, so loop forever. 1429 */ 1430 while (TRUE) { 1431 int error; 1432 int s = splvm(); 1433 1434 /* 1435 * If we have enough free memory, wakeup waiters. Do 1436 * not clear vm_pages_needed until we reach our target, 1437 * otherwise we may be woken up over and over again and 1438 * waste a lot of cpu. 1439 */ 1440 if (vm_pages_needed && !vm_page_count_min()) { 1441 if (vm_paging_needed() <= 0) 1442 vm_pages_needed = 0; 1443 wakeup(&cnt.v_free_count); 1444 } 1445 if (vm_pages_needed) { 1446 /* 1447 * Still not done, take a second pass without waiting 1448 * (unlimited dirty cleaning), otherwise sleep a bit 1449 * and try again. 1450 */ 1451 ++pass; 1452 if (pass > 1) 1453 tsleep(&vm_pages_needed, PVM, 1454 "psleep", hz/2); 1455 } else { 1456 /* 1457 * Good enough, sleep & handle stats. Prime the pass 1458 * for the next run. 1459 */ 1460 if (pass > 1) 1461 pass = 1; 1462 else 1463 pass = 0; 1464 error = tsleep(&vm_pages_needed, PVM, 1465 "psleep", vm_pageout_stats_interval * hz); 1466 if (error && !vm_pages_needed) { 1467 splx(s); 1468 pass = 0; 1469 vm_pageout_page_stats(); 1470 continue; 1471 } 1472 } 1473 1474 if (vm_pages_needed) 1475 cnt.v_pdwakeups++; 1476 splx(s); 1477 vm_pageout_scan(pass); 1478 vm_pageout_deficit = 0; 1479 } 1480} 1481 1482void 1483pagedaemon_wakeup() 1484{ 1485 if (!vm_pages_needed && curthread->td_proc != pageproc) { 1486 vm_pages_needed++; 1487 wakeup(&vm_pages_needed); 1488 } 1489} 1490 1491#if !defined(NO_SWAPPING) 1492static void 1493vm_req_vmdaemon() 1494{ 1495 static int lastrun = 0; 1496 1497 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1498 wakeup(&vm_daemon_needed); 1499 lastrun = ticks; 1500 } 1501} 1502 1503static void 1504vm_daemon() 1505{ 1506 struct proc *p; 1507 int breakout; 1508 struct thread *td; 1509 1510 mtx_lock(&Giant); 1511 while (TRUE) { 1512 tsleep(&vm_daemon_needed, PPAUSE, "psleep", 0); 1513 if (vm_pageout_req_swapout) { 1514 swapout_procs(vm_pageout_req_swapout); 1515 vm_pageout_req_swapout = 0; 1516 } 1517 /* 1518 * scan the processes for exceeding their rlimits or if 1519 * process is swapped out -- deactivate pages 1520 */ 1521 sx_slock(&allproc_lock); 1522 LIST_FOREACH(p, &allproc, p_list) { 1523 vm_pindex_t limit, size; 1524 1525 /* 1526 * if this is a system process or if we have already 1527 * looked at this process, skip it. 1528 */ 1529 if (p->p_flag & (P_SYSTEM | P_WEXIT)) { 1530 continue; 1531 } 1532 /* 1533 * if the process is in a non-running type state, 1534 * don't touch it. 1535 */ 1536 mtx_lock_spin(&sched_lock); 1537 breakout = 0; 1538 FOREACH_THREAD_IN_PROC(p, td) { 1539 if (!TD_ON_RUNQ(td) && 1540 !TD_IS_RUNNING(td) && 1541 !TD_IS_SLEEPING(td)) { 1542 breakout = 1; 1543 break; 1544 } 1545 } 1546 if (breakout) { 1547 mtx_unlock_spin(&sched_lock); 1548 continue; 1549 } 1550 /* 1551 * get a limit 1552 */ 1553 limit = OFF_TO_IDX( 1554 qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur, 1555 p->p_rlimit[RLIMIT_RSS].rlim_max)); 1556 1557 /* 1558 * let processes that are swapped out really be 1559 * swapped out set the limit to nothing (will force a 1560 * swap-out.) 1561 */ 1562 if ((p->p_sflag & PS_INMEM) == 0) 1563 limit = 0; /* XXX */ 1564 mtx_unlock_spin(&sched_lock); 1565 1566 size = vmspace_resident_count(p->p_vmspace); 1567 if (limit >= 0 && size >= limit) { 1568 vm_pageout_map_deactivate_pages( 1569 &p->p_vmspace->vm_map, limit); 1570 } 1571 } 1572 sx_sunlock(&allproc_lock); 1573 } 1574} 1575#endif /* !defined(NO_SWAPPING) */ 1576