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: src/sys/vm/vm_pageout.c,v 1.151.2.15 2002/12/29 18:21:04 dillon Exp $ 69 * $DragonFly: src/sys/vm/vm_pageout.c,v 1.18 2006/03/15 07:58:37 dillon Exp $ 70 */ 71 72 /* 73 * The proverbial page-out daemon. 74 */ 75 76 #include "opt_vm.h" 77 #include <sys/param.h> 78 #include <sys/systm.h> 79 #include <sys/kernel.h> 80 #include <sys/proc.h> 81 #include <sys/kthread.h> 82 #include <sys/resourcevar.h> 83 #include <sys/signalvar.h> 84 #include <sys/vnode.h> 85 #include <sys/vmmeter.h> 86 #include <sys/sysctl.h> 87 88 #include <vm/vm.h> 89 #include <vm/vm_param.h> 90 #include <sys/lock.h> 91 #include <vm/vm_object.h> 92 #include <vm/vm_page.h> 93 #include <vm/vm_map.h> 94 #include <vm/vm_pageout.h> 95 #include <vm/vm_pager.h> 96 #include <vm/swap_pager.h> 97 #include <vm/vm_extern.h> 98 99 #include <sys/thread2.h> 100 #include <vm/vm_page2.h> 101 102 /* 103 * System initialization 104 */ 105 106 /* the kernel process "vm_pageout"*/ 107 static void vm_pageout (void); 108 static int vm_pageout_clean (vm_page_t); 109 static void vm_pageout_scan (int pass); 110 static int vm_pageout_free_page_calc (vm_size_t count); 111 struct thread *pagethread; 112 113 static struct kproc_desc page_kp = { 114 "pagedaemon", 115 vm_pageout, 116 &pagethread 117 }; 118 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp) 119 120 #if !defined(NO_SWAPPING) 121 /* the kernel process "vm_daemon"*/ 122 static void vm_daemon (void); 123 static struct thread *vmthread; 124 125 static struct kproc_desc vm_kp = { 126 "vmdaemon", 127 vm_daemon, 128 &vmthread 129 }; 130 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp) 131 #endif 132 133 134 int vm_pages_needed=0; /* Event on which pageout daemon sleeps */ 135 int vm_pageout_deficit=0; /* Estimated number of pages deficit */ 136 int vm_pageout_pages_needed=0; /* flag saying that the pageout daemon needs pages */ 137 138 #if !defined(NO_SWAPPING) 139 static int vm_pageout_req_swapout; /* XXX */ 140 static int vm_daemon_needed; 141 #endif 142 extern int vm_swap_size; 143 static int vm_max_launder = 32; 144 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0; 145 static int vm_pageout_full_stats_interval = 0; 146 static int vm_pageout_stats_free_max=0, vm_pageout_algorithm=0; 147 static int defer_swap_pageouts=0; 148 static int disable_swap_pageouts=0; 149 150 #if defined(NO_SWAPPING) 151 static int vm_swap_enabled=0; 152 static int vm_swap_idle_enabled=0; 153 #else 154 static int vm_swap_enabled=1; 155 static int vm_swap_idle_enabled=0; 156 #endif 157 158 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm, 159 CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt"); 160 161 SYSCTL_INT(_vm, OID_AUTO, max_launder, 162 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout"); 163 164 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max, 165 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length"); 166 167 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval, 168 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan"); 169 170 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval, 171 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan"); 172 173 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_free_max, 174 CTLFLAG_RW, &vm_pageout_stats_free_max, 0, "Not implemented"); 175 176 #if defined(NO_SWAPPING) 177 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 178 CTLFLAG_RD, &vm_swap_enabled, 0, ""); 179 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 180 CTLFLAG_RD, &vm_swap_idle_enabled, 0, ""); 181 #else 182 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 183 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout"); 184 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 185 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 186 #endif 187 188 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts, 189 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem"); 190 191 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts, 192 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages"); 193 194 static int pageout_lock_miss; 195 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss, 196 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout"); 197 198 int vm_load; 199 SYSCTL_INT(_vm, OID_AUTO, vm_load, 200 CTLFLAG_RD, &vm_load, 0, "load on the VM system"); 201 int vm_load_enable = 1; 202 SYSCTL_INT(_vm, OID_AUTO, vm_load_enable, 203 CTLFLAG_RW, &vm_load_enable, 0, "enable vm_load rate limiting"); 204 #ifdef INVARIANTS 205 int vm_load_debug; 206 SYSCTL_INT(_vm, OID_AUTO, vm_load_debug, 207 CTLFLAG_RW, &vm_load_debug, 0, "debug vm_load"); 208 #endif 209 210 #define VM_PAGEOUT_PAGE_COUNT 16 211 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT; 212 213 int vm_page_max_wired; /* XXX max # of wired pages system-wide */ 214 215 #if !defined(NO_SWAPPING) 216 typedef void freeer_fcn_t (vm_map_t, vm_object_t, vm_pindex_t, int); 217 static void vm_pageout_map_deactivate_pages (vm_map_t, vm_pindex_t); 218 static freeer_fcn_t vm_pageout_object_deactivate_pages; 219 static void vm_req_vmdaemon (void); 220 #endif 221 static void vm_pageout_page_stats(void); 222 223 /* 224 * Update 225 */ 226 void 227 vm_fault_ratecheck(void) 228 { 229 if (vm_pages_needed) { 230 if (vm_load < 1000) 231 ++vm_load; 232 } else { 233 if (vm_load > 0) 234 --vm_load; 235 } 236 } 237 238 /* 239 * vm_pageout_clean: 240 * 241 * Clean the page and remove it from the laundry. The page must not be 242 * busy on-call. 243 * 244 * We set the busy bit to cause potential page faults on this page to 245 * block. Note the careful timing, however, the busy bit isn't set till 246 * late and we cannot do anything that will mess with the page. 247 */ 248 249 static int 250 vm_pageout_clean(vm_page_t m) 251 { 252 vm_object_t object; 253 vm_page_t mc[2*vm_pageout_page_count]; 254 int pageout_count; 255 int ib, is, page_base; 256 vm_pindex_t pindex = m->pindex; 257 258 object = m->object; 259 260 /* 261 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP 262 * with the new swapper, but we could have serious problems paging 263 * out other object types if there is insufficient memory. 264 * 265 * Unfortunately, checking free memory here is far too late, so the 266 * check has been moved up a procedural level. 267 */ 268 269 /* 270 * Don't mess with the page if it's busy, held, or special 271 */ 272 if ((m->hold_count != 0) || 273 ((m->busy != 0) || (m->flags & (PG_BUSY|PG_UNMANAGED)))) { 274 return 0; 275 } 276 277 mc[vm_pageout_page_count] = m; 278 pageout_count = 1; 279 page_base = vm_pageout_page_count; 280 ib = 1; 281 is = 1; 282 283 /* 284 * Scan object for clusterable pages. 285 * 286 * We can cluster ONLY if: ->> the page is NOT 287 * clean, wired, busy, held, or mapped into a 288 * buffer, and one of the following: 289 * 1) The page is inactive, or a seldom used 290 * active page. 291 * -or- 292 * 2) we force the issue. 293 * 294 * During heavy mmap/modification loads the pageout 295 * daemon can really fragment the underlying file 296 * due to flushing pages out of order and not trying 297 * align the clusters (which leave sporatic out-of-order 298 * holes). To solve this problem we do the reverse scan 299 * first and attempt to align our cluster, then do a 300 * forward scan if room remains. 301 */ 302 303 more: 304 while (ib && pageout_count < vm_pageout_page_count) { 305 vm_page_t p; 306 307 if (ib > pindex) { 308 ib = 0; 309 break; 310 } 311 312 if ((p = vm_page_lookup(object, pindex - ib)) == NULL) { 313 ib = 0; 314 break; 315 } 316 if (((p->queue - p->pc) == PQ_CACHE) || 317 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) { 318 ib = 0; 319 break; 320 } 321 vm_page_test_dirty(p); 322 if ((p->dirty & p->valid) == 0 || 323 p->queue != PQ_INACTIVE || 324 p->wire_count != 0 || /* may be held by buf cache */ 325 p->hold_count != 0) { /* may be undergoing I/O */ 326 ib = 0; 327 break; 328 } 329 mc[--page_base] = p; 330 ++pageout_count; 331 ++ib; 332 /* 333 * alignment boundry, stop here and switch directions. Do 334 * not clear ib. 335 */ 336 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0) 337 break; 338 } 339 340 while (pageout_count < vm_pageout_page_count && 341 pindex + is < object->size) { 342 vm_page_t p; 343 344 if ((p = vm_page_lookup(object, pindex + is)) == NULL) 345 break; 346 if (((p->queue - p->pc) == PQ_CACHE) || 347 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) { 348 break; 349 } 350 vm_page_test_dirty(p); 351 if ((p->dirty & p->valid) == 0 || 352 p->queue != PQ_INACTIVE || 353 p->wire_count != 0 || /* may be held by buf cache */ 354 p->hold_count != 0) { /* may be undergoing I/O */ 355 break; 356 } 357 mc[page_base + pageout_count] = p; 358 ++pageout_count; 359 ++is; 360 } 361 362 /* 363 * If we exhausted our forward scan, continue with the reverse scan 364 * when possible, even past a page boundry. This catches boundry 365 * conditions. 366 */ 367 if (ib && pageout_count < vm_pageout_page_count) 368 goto more; 369 370 /* 371 * we allow reads during pageouts... 372 */ 373 return vm_pageout_flush(&mc[page_base], pageout_count, 0); 374 } 375 376 /* 377 * vm_pageout_flush() - launder the given pages 378 * 379 * The given pages are laundered. Note that we setup for the start of 380 * I/O ( i.e. busy the page ), mark it read-only, and bump the object 381 * reference count all in here rather then in the parent. If we want 382 * the parent to do more sophisticated things we may have to change 383 * the ordering. 384 */ 385 386 int 387 vm_pageout_flush(vm_page_t *mc, int count, int flags) 388 { 389 vm_object_t object; 390 int pageout_status[count]; 391 int numpagedout = 0; 392 int i; 393 394 /* 395 * Initiate I/O. Bump the vm_page_t->busy counter and 396 * mark the pages read-only. 397 * 398 * We do not have to fixup the clean/dirty bits here... we can 399 * allow the pager to do it after the I/O completes. 400 */ 401 402 for (i = 0; i < count; i++) { 403 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, ("vm_pageout_flush page %p index %d/%d: partially invalid page", mc[i], i, count)); 404 vm_page_io_start(mc[i]); 405 vm_page_protect(mc[i], VM_PROT_READ); 406 } 407 408 object = mc[0]->object; 409 vm_object_pip_add(object, count); 410 411 vm_pager_put_pages(object, mc, count, 412 (flags | ((object == kernel_object) ? VM_PAGER_PUT_SYNC : 0)), 413 pageout_status); 414 415 for (i = 0; i < count; i++) { 416 vm_page_t mt = mc[i]; 417 418 switch (pageout_status[i]) { 419 case VM_PAGER_OK: 420 numpagedout++; 421 break; 422 case VM_PAGER_PEND: 423 numpagedout++; 424 break; 425 case VM_PAGER_BAD: 426 /* 427 * Page outside of range of object. Right now we 428 * essentially lose the changes by pretending it 429 * worked. 430 */ 431 pmap_clear_modify(mt); 432 vm_page_undirty(mt); 433 break; 434 case VM_PAGER_ERROR: 435 case VM_PAGER_FAIL: 436 /* 437 * If page couldn't be paged out, then reactivate the 438 * page so it doesn't clog the inactive list. (We 439 * will try paging out it again later). 440 */ 441 vm_page_activate(mt); 442 break; 443 case VM_PAGER_AGAIN: 444 break; 445 } 446 447 /* 448 * If the operation is still going, leave the page busy to 449 * block all other accesses. Also, leave the paging in 450 * progress indicator set so that we don't attempt an object 451 * collapse. 452 */ 453 if (pageout_status[i] != VM_PAGER_PEND) { 454 vm_object_pip_wakeup(object); 455 vm_page_io_finish(mt); 456 if (!vm_page_count_severe() || !vm_page_try_to_cache(mt)) 457 vm_page_protect(mt, VM_PROT_READ); 458 } 459 } 460 return numpagedout; 461 } 462 463 #if !defined(NO_SWAPPING) 464 /* 465 * vm_pageout_object_deactivate_pages 466 * 467 * deactivate enough pages to satisfy the inactive target 468 * requirements or if vm_page_proc_limit is set, then 469 * deactivate all of the pages in the object and its 470 * backing_objects. 471 * 472 * The object and map must be locked. 473 */ 474 static void 475 vm_pageout_object_deactivate_pages(vm_map_t map, vm_object_t object, 476 vm_pindex_t desired, int map_remove_only) 477 { 478 vm_page_t p, next; 479 int rcount; 480 int remove_mode; 481 482 if (object->type == OBJT_DEVICE || object->type == OBJT_PHYS) 483 return; 484 485 while (object) { 486 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 487 return; 488 if (object->paging_in_progress) 489 return; 490 491 remove_mode = map_remove_only; 492 if (object->shadow_count > 1) 493 remove_mode = 1; 494 495 /* 496 * scan the objects entire memory queue. spl protection is 497 * required to avoid an interrupt unbusy/free race against 498 * our busy check. 499 */ 500 crit_enter(); 501 rcount = object->resident_page_count; 502 p = TAILQ_FIRST(&object->memq); 503 504 while (p && (rcount-- > 0)) { 505 int actcount; 506 if (pmap_resident_count(vm_map_pmap(map)) <= desired) { 507 crit_exit(); 508 return; 509 } 510 next = TAILQ_NEXT(p, listq); 511 mycpu->gd_cnt.v_pdpages++; 512 if (p->wire_count != 0 || 513 p->hold_count != 0 || 514 p->busy != 0 || 515 (p->flags & (PG_BUSY|PG_UNMANAGED)) || 516 !pmap_page_exists_quick(vm_map_pmap(map), p)) { 517 p = next; 518 continue; 519 } 520 521 actcount = pmap_ts_referenced(p); 522 if (actcount) { 523 vm_page_flag_set(p, PG_REFERENCED); 524 } else if (p->flags & PG_REFERENCED) { 525 actcount = 1; 526 } 527 528 if ((p->queue != PQ_ACTIVE) && 529 (p->flags & PG_REFERENCED)) { 530 vm_page_activate(p); 531 p->act_count += actcount; 532 vm_page_flag_clear(p, PG_REFERENCED); 533 } else if (p->queue == PQ_ACTIVE) { 534 if ((p->flags & PG_REFERENCED) == 0) { 535 p->act_count -= min(p->act_count, ACT_DECLINE); 536 if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) { 537 vm_page_protect(p, VM_PROT_NONE); 538 vm_page_deactivate(p); 539 } else { 540 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, p, pageq); 541 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, p, pageq); 542 } 543 } else { 544 vm_page_activate(p); 545 vm_page_flag_clear(p, PG_REFERENCED); 546 if (p->act_count < (ACT_MAX - ACT_ADVANCE)) 547 p->act_count += ACT_ADVANCE; 548 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, p, pageq); 549 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, p, pageq); 550 } 551 } else if (p->queue == PQ_INACTIVE) { 552 vm_page_protect(p, VM_PROT_NONE); 553 } 554 p = next; 555 } 556 crit_exit(); 557 object = object->backing_object; 558 } 559 } 560 561 /* 562 * deactivate some number of pages in a map, try to do it fairly, but 563 * that is really hard to do. 564 */ 565 static void 566 vm_pageout_map_deactivate_pages(vm_map_t map, vm_pindex_t desired) 567 { 568 vm_map_entry_t tmpe; 569 vm_object_t obj, bigobj; 570 int nothingwired; 571 572 if (lockmgr(&map->lock, LK_EXCLUSIVE | LK_NOWAIT, NULL, curthread)) { 573 return; 574 } 575 576 bigobj = NULL; 577 nothingwired = TRUE; 578 579 /* 580 * first, search out the biggest object, and try to free pages from 581 * that. 582 */ 583 tmpe = map->header.next; 584 while (tmpe != &map->header) { 585 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 586 obj = tmpe->object.vm_object; 587 if ((obj != NULL) && (obj->shadow_count <= 1) && 588 ((bigobj == NULL) || 589 (bigobj->resident_page_count < obj->resident_page_count))) { 590 bigobj = obj; 591 } 592 } 593 if (tmpe->wired_count > 0) 594 nothingwired = FALSE; 595 tmpe = tmpe->next; 596 } 597 598 if (bigobj) 599 vm_pageout_object_deactivate_pages(map, bigobj, desired, 0); 600 601 /* 602 * Next, hunt around for other pages to deactivate. We actually 603 * do this search sort of wrong -- .text first is not the best idea. 604 */ 605 tmpe = map->header.next; 606 while (tmpe != &map->header) { 607 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 608 break; 609 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 610 obj = tmpe->object.vm_object; 611 if (obj) 612 vm_pageout_object_deactivate_pages(map, obj, desired, 0); 613 } 614 tmpe = tmpe->next; 615 }; 616 617 /* 618 * Remove all mappings if a process is swapped out, this will free page 619 * table pages. 620 */ 621 if (desired == 0 && nothingwired) 622 pmap_remove(vm_map_pmap(map), 623 VM_MIN_ADDRESS, VM_MAXUSER_ADDRESS); 624 vm_map_unlock(map); 625 } 626 #endif 627 628 /* 629 * Don't try to be fancy - being fancy can lead to VOP_LOCK's and therefore 630 * to vnode deadlocks. We only do it for OBJT_DEFAULT and OBJT_SWAP objects 631 * which we know can be trivially freed. 632 */ 633 634 void 635 vm_pageout_page_free(vm_page_t m) { 636 vm_object_t object = m->object; 637 int type = object->type; 638 639 if (type == OBJT_SWAP || type == OBJT_DEFAULT) 640 vm_object_reference(object); 641 vm_page_busy(m); 642 vm_page_protect(m, VM_PROT_NONE); 643 vm_page_free(m); 644 if (type == OBJT_SWAP || type == OBJT_DEFAULT) 645 vm_object_deallocate(object); 646 } 647 648 /* 649 * vm_pageout_scan does the dirty work for the pageout daemon. 650 */ 651 static void 652 vm_pageout_scan(int pass) 653 { 654 vm_page_t m, next; 655 struct vm_page marker; 656 int page_shortage, maxscan, pcount; 657 int addl_page_shortage, addl_page_shortage_init; 658 struct proc *p, *bigproc; 659 vm_offset_t size, bigsize; 660 vm_object_t object; 661 int actcount; 662 int vnodes_skipped = 0; 663 int maxlaunder; 664 665 /* 666 * Do whatever cleanup that the pmap code can. 667 */ 668 pmap_collect(); 669 670 addl_page_shortage_init = vm_pageout_deficit; 671 vm_pageout_deficit = 0; 672 673 /* 674 * Calculate the number of pages we want to either free or move 675 * to the cache. 676 */ 677 page_shortage = vm_paging_target() + addl_page_shortage_init; 678 679 /* 680 * Initialize our marker 681 */ 682 bzero(&marker, sizeof(marker)); 683 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER; 684 marker.queue = PQ_INACTIVE; 685 marker.wire_count = 1; 686 687 /* 688 * Start scanning the inactive queue for pages we can move to the 689 * cache or free. The scan will stop when the target is reached or 690 * we have scanned the entire inactive queue. Note that m->act_count 691 * is not used to form decisions for the inactive queue, only for the 692 * active queue. 693 * 694 * maxlaunder limits the number of dirty pages we flush per scan. 695 * For most systems a smaller value (16 or 32) is more robust under 696 * extreme memory and disk pressure because any unnecessary writes 697 * to disk can result in extreme performance degredation. However, 698 * systems with excessive dirty pages (especially when MAP_NOSYNC is 699 * used) will die horribly with limited laundering. If the pageout 700 * daemon cannot clean enough pages in the first pass, we let it go 701 * all out in succeeding passes. 702 */ 703 if ((maxlaunder = vm_max_launder) <= 1) 704 maxlaunder = 1; 705 if (pass) 706 maxlaunder = 10000; 707 708 /* 709 * We will generally be in a critical section throughout the 710 * scan, but we can release it temporarily when we are sitting on a 711 * non-busy page without fear. this is required to prevent an 712 * interrupt from unbusying or freeing a page prior to our busy 713 * check, leaving us on the wrong queue or checking the wrong 714 * page. 715 */ 716 crit_enter(); 717 rescan0: 718 addl_page_shortage = addl_page_shortage_init; 719 maxscan = vmstats.v_inactive_count; 720 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl); 721 m != NULL && maxscan-- > 0 && page_shortage > 0; 722 m = next 723 ) { 724 mycpu->gd_cnt.v_pdpages++; 725 726 /* 727 * Give interrupts a chance 728 */ 729 crit_exit(); 730 crit_enter(); 731 732 /* 733 * It's easier for some of the conditions below to just loop 734 * and catch queue changes here rather then check everywhere 735 * else. 736 */ 737 if (m->queue != PQ_INACTIVE) 738 goto rescan0; 739 next = TAILQ_NEXT(m, pageq); 740 741 /* 742 * skip marker pages 743 */ 744 if (m->flags & PG_MARKER) 745 continue; 746 747 /* 748 * A held page may be undergoing I/O, so skip it. 749 */ 750 if (m->hold_count) { 751 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 752 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 753 addl_page_shortage++; 754 continue; 755 } 756 757 /* 758 * Dont mess with busy pages, keep in the front of the 759 * queue, most likely are being paged out. 760 */ 761 if (m->busy || (m->flags & PG_BUSY)) { 762 addl_page_shortage++; 763 continue; 764 } 765 766 if (m->object->ref_count == 0) { 767 /* 768 * If the object is not being used, we ignore previous 769 * references. 770 */ 771 vm_page_flag_clear(m, PG_REFERENCED); 772 pmap_clear_reference(m); 773 774 } else if (((m->flags & PG_REFERENCED) == 0) && 775 (actcount = pmap_ts_referenced(m))) { 776 /* 777 * Otherwise, if the page has been referenced while 778 * in the inactive queue, we bump the "activation 779 * count" upwards, making it less likely that the 780 * page will be added back to the inactive queue 781 * prematurely again. Here we check the page tables 782 * (or emulated bits, if any), given the upper level 783 * VM system not knowing anything about existing 784 * references. 785 */ 786 vm_page_activate(m); 787 m->act_count += (actcount + ACT_ADVANCE); 788 continue; 789 } 790 791 /* 792 * If the upper level VM system knows about any page 793 * references, we activate the page. We also set the 794 * "activation count" higher than normal so that we will less 795 * likely place pages back onto the inactive queue again. 796 */ 797 if ((m->flags & PG_REFERENCED) != 0) { 798 vm_page_flag_clear(m, PG_REFERENCED); 799 actcount = pmap_ts_referenced(m); 800 vm_page_activate(m); 801 m->act_count += (actcount + ACT_ADVANCE + 1); 802 continue; 803 } 804 805 /* 806 * If the upper level VM system doesn't know anything about 807 * the page being dirty, we have to check for it again. As 808 * far as the VM code knows, any partially dirty pages are 809 * fully dirty. 810 * 811 * Pages marked PG_WRITEABLE may be mapped into the user 812 * address space of a process running on another cpu. A 813 * user process (without holding the MP lock) running on 814 * another cpu may be able to touch the page while we are 815 * trying to remove it. To prevent this from occuring we 816 * must call pmap_remove_all() or otherwise make the page 817 * read-only. If the race occured pmap_remove_all() is 818 * responsible for setting m->dirty. 819 */ 820 if (m->dirty == 0) { 821 vm_page_test_dirty(m); 822 #if 0 823 if (m->dirty == 0 && (m->flags & PG_WRITEABLE) != 0) 824 pmap_remove_all(m); 825 #endif 826 } else { 827 vm_page_dirty(m); 828 } 829 830 if (m->valid == 0) { 831 /* 832 * Invalid pages can be easily freed 833 */ 834 vm_pageout_page_free(m); 835 mycpu->gd_cnt.v_dfree++; 836 --page_shortage; 837 } else if (m->dirty == 0) { 838 /* 839 * Clean pages can be placed onto the cache queue. 840 * This effectively frees them. 841 */ 842 vm_page_cache(m); 843 --page_shortage; 844 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) { 845 /* 846 * Dirty pages need to be paged out, but flushing 847 * a page is extremely expensive verses freeing 848 * a clean page. Rather then artificially limiting 849 * the number of pages we can flush, we instead give 850 * dirty pages extra priority on the inactive queue 851 * by forcing them to be cycled through the queue 852 * twice before being flushed, after which the 853 * (now clean) page will cycle through once more 854 * before being freed. This significantly extends 855 * the thrash point for a heavily loaded machine. 856 */ 857 vm_page_flag_set(m, PG_WINATCFLS); 858 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 859 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 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; 870 871 object = m->object; 872 873 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 874 swap_pageouts_ok = 1; 875 } else { 876 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 877 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 878 vm_page_count_min()); 879 880 } 881 882 /* 883 * We don't bother paging objects that are "dead". 884 * Those objects are in a "rundown" state. 885 */ 886 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 887 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 888 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 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 because, 901 * quite often, a vm_fault has initiated an I/O and 902 * is holding a locked vnode at just the point where 903 * the pageout daemon is woken up. 904 * 905 * We can't wait forever for the vnode lock, we might 906 * deadlock due to a vn_read() getting stuck in 907 * vm_wait while holding this vnode. We skip the 908 * vnode if we can't get it in a reasonable amount 909 * of time. 910 */ 911 912 if (object->type == OBJT_VNODE) { 913 vp = object->handle; 914 915 if (vget(vp, LK_EXCLUSIVE|LK_NOOBJ|LK_TIMELOCK, curthread)) { 916 ++pageout_lock_miss; 917 if (object->flags & OBJ_MIGHTBEDIRTY) 918 vnodes_skipped++; 919 continue; 920 } 921 922 /* 923 * The page might have been moved to another 924 * queue during potential blocking in vget() 925 * above. The page might have been freed and 926 * reused for another vnode. The object might 927 * have been reused for another vnode. 928 */ 929 if (m->queue != PQ_INACTIVE || 930 m->object != object || 931 object->handle != vp) { 932 if (object->flags & OBJ_MIGHTBEDIRTY) 933 vnodes_skipped++; 934 vput(vp); 935 continue; 936 } 937 938 /* 939 * The page may have been busied during the 940 * blocking in vput(); We don't move the 941 * page back onto the end of the queue so that 942 * statistics are more correct if we don't. 943 */ 944 if (m->busy || (m->flags & PG_BUSY)) { 945 vput(vp); 946 continue; 947 } 948 949 /* 950 * If the page has become held it might 951 * be undergoing I/O, so skip it 952 */ 953 if (m->hold_count) { 954 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 955 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 956 if (object->flags & OBJ_MIGHTBEDIRTY) 957 vnodes_skipped++; 958 vput(vp); 959 continue; 960 } 961 } 962 963 /* 964 * If a page is dirty, then it is either being washed 965 * (but not yet cleaned) or it is still in the 966 * laundry. If it is still in the laundry, then we 967 * start the cleaning operation. 968 * 969 * This operation may cluster, invalidating the 'next' 970 * pointer. To prevent an inordinate number of 971 * restarts we use our marker to remember our place. 972 * 973 * decrement page_shortage on success to account for 974 * the (future) cleaned page. Otherwise we could wind 975 * up laundering or cleaning too many pages. 976 */ 977 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, m, &marker, pageq); 978 if (vm_pageout_clean(m) != 0) { 979 --page_shortage; 980 --maxlaunder; 981 } 982 next = TAILQ_NEXT(&marker, pageq); 983 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, &marker, pageq); 984 if (vp != NULL) 985 vput(vp); 986 } 987 } 988 989 /* 990 * Compute the number of pages we want to try to move from the 991 * active queue to the inactive queue. 992 */ 993 page_shortage = vm_paging_target() + 994 vmstats.v_inactive_target - vmstats.v_inactive_count; 995 page_shortage += addl_page_shortage; 996 997 /* 998 * Scan the active queue for things we can deactivate. We nominally 999 * track the per-page activity counter and use it to locate 1000 * deactivation candidates. 1001 * 1002 * NOTE: we are still in a critical section. 1003 */ 1004 pcount = vmstats.v_active_count; 1005 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1006 1007 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) { 1008 /* 1009 * Give interrupts a chance. 1010 */ 1011 crit_exit(); 1012 crit_enter(); 1013 1014 /* 1015 * If the page was ripped out from under us, just stop. 1016 */ 1017 if (m->queue != PQ_ACTIVE) 1018 break; 1019 next = TAILQ_NEXT(m, pageq); 1020 1021 /* 1022 * Don't deactivate pages that are busy. 1023 */ 1024 if ((m->busy != 0) || 1025 (m->flags & PG_BUSY) || 1026 (m->hold_count != 0)) { 1027 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1028 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1029 m = next; 1030 continue; 1031 } 1032 1033 /* 1034 * The count for pagedaemon pages is done after checking the 1035 * page for eligibility... 1036 */ 1037 mycpu->gd_cnt.v_pdpages++; 1038 1039 /* 1040 * Check to see "how much" the page has been used. 1041 */ 1042 actcount = 0; 1043 if (m->object->ref_count != 0) { 1044 if (m->flags & PG_REFERENCED) { 1045 actcount += 1; 1046 } 1047 actcount += pmap_ts_referenced(m); 1048 if (actcount) { 1049 m->act_count += ACT_ADVANCE + actcount; 1050 if (m->act_count > ACT_MAX) 1051 m->act_count = ACT_MAX; 1052 } 1053 } 1054 1055 /* 1056 * Since we have "tested" this bit, we need to clear it now. 1057 */ 1058 vm_page_flag_clear(m, PG_REFERENCED); 1059 1060 /* 1061 * Only if an object is currently being used, do we use the 1062 * page activation count stats. 1063 */ 1064 if (actcount && (m->object->ref_count != 0)) { 1065 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1066 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1067 } else { 1068 m->act_count -= min(m->act_count, ACT_DECLINE); 1069 if (vm_pageout_algorithm || 1070 m->object->ref_count == 0 || 1071 m->act_count == 0) { 1072 page_shortage--; 1073 if (m->object->ref_count == 0) { 1074 vm_page_protect(m, VM_PROT_NONE); 1075 if (m->dirty == 0) 1076 vm_page_cache(m); 1077 else 1078 vm_page_deactivate(m); 1079 } else { 1080 vm_page_deactivate(m); 1081 } 1082 } else { 1083 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1084 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1085 } 1086 } 1087 m = next; 1088 } 1089 1090 /* 1091 * We try to maintain some *really* free pages, this allows interrupt 1092 * code to be guaranteed space. Since both cache and free queues 1093 * are considered basically 'free', moving pages from cache to free 1094 * does not effect other calculations. 1095 * 1096 * NOTE: we are still in a critical section. 1097 */ 1098 1099 while (vmstats.v_free_count < vmstats.v_free_reserved) { 1100 static int cache_rover = 0; 1101 m = vm_page_list_find(PQ_CACHE, cache_rover, FALSE); 1102 if (!m) 1103 break; 1104 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || 1105 m->busy || 1106 m->hold_count || 1107 m->wire_count) { 1108 #ifdef INVARIANTS 1109 printf("Warning: busy page %p found in cache\n", m); 1110 #endif 1111 vm_page_deactivate(m); 1112 continue; 1113 } 1114 cache_rover = (cache_rover + PQ_PRIME2) & PQ_L2_MASK; 1115 vm_pageout_page_free(m); 1116 mycpu->gd_cnt.v_dfree++; 1117 } 1118 1119 crit_exit(); 1120 1121 #if !defined(NO_SWAPPING) 1122 /* 1123 * Idle process swapout -- run once per second. 1124 */ 1125 if (vm_swap_idle_enabled) { 1126 static long lsec; 1127 if (time_second != lsec) { 1128 vm_pageout_req_swapout |= VM_SWAP_IDLE; 1129 vm_req_vmdaemon(); 1130 lsec = time_second; 1131 } 1132 } 1133 #endif 1134 1135 /* 1136 * If we didn't get enough free pages, and we have skipped a vnode 1137 * in a writeable object, wakeup the sync daemon. And kick swapout 1138 * if we did not get enough free pages. 1139 */ 1140 if (vm_paging_target() > 0) { 1141 if (vnodes_skipped && vm_page_count_min()) 1142 speedup_syncer(); 1143 #if !defined(NO_SWAPPING) 1144 if (vm_swap_enabled && vm_page_count_target()) { 1145 vm_req_vmdaemon(); 1146 vm_pageout_req_swapout |= VM_SWAP_NORMAL; 1147 } 1148 #endif 1149 } 1150 1151 /* 1152 * If we are out of swap and were not able to reach our paging 1153 * target, kill the largest process. 1154 */ 1155 if ((vm_swap_size < 64 && vm_page_count_min()) || 1156 (swap_pager_full && vm_paging_target() > 0)) { 1157 #if 0 1158 if ((vm_swap_size < 64 || swap_pager_full) && vm_page_count_min()) { 1159 #endif 1160 bigproc = NULL; 1161 bigsize = 0; 1162 for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) { 1163 /* 1164 * if this is a system process, skip it 1165 */ 1166 if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) || 1167 ((p->p_pid < 48) && (vm_swap_size != 0))) { 1168 continue; 1169 } 1170 /* 1171 * if the process is in a non-running type state, 1172 * don't touch it. 1173 */ 1174 if (p->p_stat != SRUN && p->p_stat != SSLEEP) { 1175 continue; 1176 } 1177 /* 1178 * get the process size 1179 */ 1180 size = vmspace_resident_count(p->p_vmspace) + 1181 vmspace_swap_count(p->p_vmspace); 1182 /* 1183 * if the this process is bigger than the biggest one 1184 * remember it. 1185 */ 1186 if (size > bigsize) { 1187 bigproc = p; 1188 bigsize = size; 1189 } 1190 } 1191 if (bigproc != NULL) { 1192 killproc(bigproc, "out of swap space"); 1193 bigproc->p_nice = PRIO_MIN; 1194 bigproc->p_usched->resetpriority(&bigproc->p_lwp); 1195 wakeup(&vmstats.v_free_count); 1196 } 1197 } 1198 } 1199 1200 /* 1201 * This routine tries to maintain the pseudo LRU active queue, 1202 * so that during long periods of time where there is no paging, 1203 * that some statistic accumulation still occurs. This code 1204 * helps the situation where paging just starts to occur. 1205 */ 1206 static void 1207 vm_pageout_page_stats(void) 1208 { 1209 vm_page_t m,next; 1210 int pcount,tpcount; /* Number of pages to check */ 1211 static int fullintervalcount = 0; 1212 int page_shortage; 1213 1214 page_shortage = 1215 (vmstats.v_inactive_target + vmstats.v_cache_max + vmstats.v_free_min) - 1216 (vmstats.v_free_count + vmstats.v_inactive_count + vmstats.v_cache_count); 1217 1218 if (page_shortage <= 0) 1219 return; 1220 1221 crit_enter(); 1222 1223 pcount = vmstats.v_active_count; 1224 fullintervalcount += vm_pageout_stats_interval; 1225 if (fullintervalcount < vm_pageout_full_stats_interval) { 1226 tpcount = (vm_pageout_stats_max * vmstats.v_active_count) / vmstats.v_page_count; 1227 if (pcount > tpcount) 1228 pcount = tpcount; 1229 } else { 1230 fullintervalcount = 0; 1231 } 1232 1233 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1234 while ((m != NULL) && (pcount-- > 0)) { 1235 int actcount; 1236 1237 if (m->queue != PQ_ACTIVE) { 1238 break; 1239 } 1240 1241 next = TAILQ_NEXT(m, pageq); 1242 /* 1243 * Don't deactivate pages that are busy. 1244 */ 1245 if ((m->busy != 0) || 1246 (m->flags & PG_BUSY) || 1247 (m->hold_count != 0)) { 1248 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1249 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1250 m = next; 1251 continue; 1252 } 1253 1254 actcount = 0; 1255 if (m->flags & PG_REFERENCED) { 1256 vm_page_flag_clear(m, PG_REFERENCED); 1257 actcount += 1; 1258 } 1259 1260 actcount += pmap_ts_referenced(m); 1261 if (actcount) { 1262 m->act_count += ACT_ADVANCE + actcount; 1263 if (m->act_count > ACT_MAX) 1264 m->act_count = ACT_MAX; 1265 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1266 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1267 } else { 1268 if (m->act_count == 0) { 1269 /* 1270 * We turn off page access, so that we have 1271 * more accurate RSS stats. We don't do this 1272 * in the normal page deactivation when the 1273 * system is loaded VM wise, because the 1274 * cost of the large number of page protect 1275 * operations would be higher than the value 1276 * of doing the operation. 1277 */ 1278 vm_page_protect(m, VM_PROT_NONE); 1279 vm_page_deactivate(m); 1280 } else { 1281 m->act_count -= min(m->act_count, ACT_DECLINE); 1282 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1283 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1284 } 1285 } 1286 1287 m = next; 1288 } 1289 crit_exit(); 1290 } 1291 1292 static int 1293 vm_pageout_free_page_calc(vm_size_t count) 1294 { 1295 if (count < vmstats.v_page_count) 1296 return 0; 1297 /* 1298 * free_reserved needs to include enough for the largest swap pager 1299 * structures plus enough for any pv_entry structs when paging. 1300 */ 1301 if (vmstats.v_page_count > 1024) 1302 vmstats.v_free_min = 4 + (vmstats.v_page_count - 1024) / 200; 1303 else 1304 vmstats.v_free_min = 4; 1305 vmstats.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1306 vmstats.v_interrupt_free_min; 1307 vmstats.v_free_reserved = vm_pageout_page_count + 1308 vmstats.v_pageout_free_min + (count / 768) + PQ_L2_SIZE; 1309 vmstats.v_free_severe = vmstats.v_free_min / 2; 1310 vmstats.v_free_min += vmstats.v_free_reserved; 1311 vmstats.v_free_severe += vmstats.v_free_reserved; 1312 return 1; 1313 } 1314 1315 1316 /* 1317 * vm_pageout is the high level pageout daemon. 1318 */ 1319 static void 1320 vm_pageout(void) 1321 { 1322 int pass; 1323 1324 /* 1325 * Initialize some paging parameters. 1326 */ 1327 1328 vmstats.v_interrupt_free_min = 2; 1329 if (vmstats.v_page_count < 2000) 1330 vm_pageout_page_count = 8; 1331 1332 vm_pageout_free_page_calc(vmstats.v_page_count); 1333 /* 1334 * v_free_target and v_cache_min control pageout hysteresis. Note 1335 * that these are more a measure of the VM cache queue hysteresis 1336 * then the VM free queue. Specifically, v_free_target is the 1337 * high water mark (free+cache pages). 1338 * 1339 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the 1340 * low water mark, while v_free_min is the stop. v_cache_min must 1341 * be big enough to handle memory needs while the pageout daemon 1342 * is signalled and run to free more pages. 1343 */ 1344 if (vmstats.v_free_count > 6144) 1345 vmstats.v_free_target = 4 * vmstats.v_free_min + vmstats.v_free_reserved; 1346 else 1347 vmstats.v_free_target = 2 * vmstats.v_free_min + vmstats.v_free_reserved; 1348 1349 if (vmstats.v_free_count > 2048) { 1350 vmstats.v_cache_min = vmstats.v_free_target; 1351 vmstats.v_cache_max = 2 * vmstats.v_cache_min; 1352 vmstats.v_inactive_target = (3 * vmstats.v_free_target) / 2; 1353 } else { 1354 vmstats.v_cache_min = 0; 1355 vmstats.v_cache_max = 0; 1356 vmstats.v_inactive_target = vmstats.v_free_count / 4; 1357 } 1358 if (vmstats.v_inactive_target > vmstats.v_free_count / 3) 1359 vmstats.v_inactive_target = vmstats.v_free_count / 3; 1360 1361 /* XXX does not really belong here */ 1362 if (vm_page_max_wired == 0) 1363 vm_page_max_wired = vmstats.v_free_count / 3; 1364 1365 if (vm_pageout_stats_max == 0) 1366 vm_pageout_stats_max = vmstats.v_free_target; 1367 1368 /* 1369 * Set interval in seconds for stats scan. 1370 */ 1371 if (vm_pageout_stats_interval == 0) 1372 vm_pageout_stats_interval = 5; 1373 if (vm_pageout_full_stats_interval == 0) 1374 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4; 1375 1376 1377 /* 1378 * Set maximum free per pass 1379 */ 1380 if (vm_pageout_stats_free_max == 0) 1381 vm_pageout_stats_free_max = 5; 1382 1383 swap_pager_swap_init(); 1384 pass = 0; 1385 /* 1386 * The pageout daemon is never done, so loop forever. 1387 */ 1388 while (TRUE) { 1389 int error; 1390 1391 /* 1392 * If we have enough free memory, wakeup waiters. Do 1393 * not clear vm_pages_needed until we reach our target, 1394 * otherwise we may be woken up over and over again and 1395 * waste a lot of cpu. 1396 */ 1397 crit_enter(); 1398 if (vm_pages_needed && !vm_page_count_min()) { 1399 if (vm_paging_needed() <= 0) 1400 vm_pages_needed = 0; 1401 wakeup(&vmstats.v_free_count); 1402 } 1403 if (vm_pages_needed) { 1404 /* 1405 * Still not done, take a second pass without waiting 1406 * (unlimited dirty cleaning), otherwise sleep a bit 1407 * and try again. 1408 */ 1409 ++pass; 1410 if (pass > 1) 1411 tsleep(&vm_pages_needed, 0, "psleep", hz/2); 1412 } else { 1413 /* 1414 * Good enough, sleep & handle stats. Prime the pass 1415 * for the next run. 1416 */ 1417 if (pass > 1) 1418 pass = 1; 1419 else 1420 pass = 0; 1421 error = tsleep(&vm_pages_needed, 1422 0, "psleep", vm_pageout_stats_interval * hz); 1423 if (error && !vm_pages_needed) { 1424 crit_exit(); 1425 pass = 0; 1426 vm_pageout_page_stats(); 1427 continue; 1428 } 1429 } 1430 1431 if (vm_pages_needed) 1432 mycpu->gd_cnt.v_pdwakeups++; 1433 crit_exit(); 1434 vm_pageout_scan(pass); 1435 vm_pageout_deficit = 0; 1436 } 1437 } 1438 1439 void 1440 pagedaemon_wakeup(void) 1441 { 1442 if (!vm_pages_needed && curthread != pagethread) { 1443 vm_pages_needed++; 1444 wakeup(&vm_pages_needed); 1445 } 1446 } 1447 1448 #if !defined(NO_SWAPPING) 1449 static void 1450 vm_req_vmdaemon(void) 1451 { 1452 static int lastrun = 0; 1453 1454 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1455 wakeup(&vm_daemon_needed); 1456 lastrun = ticks; 1457 } 1458 } 1459 1460 static void 1461 vm_daemon(void) 1462 { 1463 struct proc *p; 1464 1465 while (TRUE) { 1466 tsleep(&vm_daemon_needed, 0, "psleep", 0); 1467 if (vm_pageout_req_swapout) { 1468 swapout_procs(vm_pageout_req_swapout); 1469 vm_pageout_req_swapout = 0; 1470 } 1471 /* 1472 * scan the processes for exceeding their rlimits or if 1473 * process is swapped out -- deactivate pages 1474 */ 1475 1476 for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) { 1477 vm_pindex_t limit, size; 1478 1479 /* 1480 * if this is a system process or if we have already 1481 * looked at this process, skip it. 1482 */ 1483 if (p->p_flag & (P_SYSTEM | P_WEXIT)) { 1484 continue; 1485 } 1486 /* 1487 * if the process is in a non-running type state, 1488 * don't touch it. 1489 */ 1490 if (p->p_stat != SRUN && p->p_stat != SSLEEP) { 1491 continue; 1492 } 1493 /* 1494 * get a limit 1495 */ 1496 limit = OFF_TO_IDX( 1497 qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur, 1498 p->p_rlimit[RLIMIT_RSS].rlim_max)); 1499 1500 /* 1501 * let processes that are swapped out really be 1502 * swapped out. Set the limit to nothing to get as 1503 * many pages out to swap as possible. 1504 */ 1505 if (p->p_flag & P_SWAPPEDOUT) 1506 limit = 0; 1507 1508 size = vmspace_resident_count(p->p_vmspace); 1509 if (limit >= 0 && size >= limit) { 1510 vm_pageout_map_deactivate_pages( 1511 &p->p_vmspace->vm_map, limit); 1512 } 1513 } 1514 } 1515 } 1516 #endif 1517