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 */ 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/proc.h> 80 #include <sys/kthread.h> 81 #include <sys/resourcevar.h> 82 #include <sys/signalvar.h> 83 #include <sys/vnode.h> 84 #include <sys/vmmeter.h> 85 #include <sys/sysctl.h> 86 87 #include <vm/vm.h> 88 #include <vm/vm_param.h> 89 #include <sys/lock.h> 90 #include <vm/vm_object.h> 91 #include <vm/vm_page.h> 92 #include <vm/vm_map.h> 93 #include <vm/vm_pageout.h> 94 #include <vm/vm_pager.h> 95 #include <vm/swap_pager.h> 96 #include <vm/vm_extern.h> 97 98 /* 99 * System initialization 100 */ 101 102 /* the kernel process "vm_pageout"*/ 103 static void vm_pageout __P((void)); 104 static int vm_pageout_clean __P((vm_page_t)); 105 static void vm_pageout_scan __P((int pass)); 106 static int vm_pageout_free_page_calc __P((vm_size_t count)); 107 struct proc *pageproc; 108 109 static struct kproc_desc page_kp = { 110 "pagedaemon", 111 vm_pageout, 112 &pageproc 113 }; 114 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp) 115 116 #if !defined(NO_SWAPPING) 117 /* the kernel process "vm_daemon"*/ 118 static void vm_daemon __P((void)); 119 static struct proc *vmproc; 120 121 static struct kproc_desc vm_kp = { 122 "vmdaemon", 123 vm_daemon, 124 &vmproc 125 }; 126 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp) 127 #endif 128 129 130 int vm_pages_needed=0; /* Event on which pageout daemon sleeps */ 131 int vm_pageout_deficit=0; /* Estimated number of pages deficit */ 132 int vm_pageout_pages_needed=0; /* flag saying that the pageout daemon needs pages */ 133 134 #if !defined(NO_SWAPPING) 135 static int vm_pageout_req_swapout; /* XXX */ 136 static int vm_daemon_needed; 137 #endif 138 extern int vm_swap_size; 139 static int vm_max_launder = 32; 140 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0; 141 static int vm_pageout_full_stats_interval = 0; 142 static int vm_pageout_stats_free_max=0, vm_pageout_algorithm=0; 143 static int defer_swap_pageouts=0; 144 static int disable_swap_pageouts=0; 145 146 #if defined(NO_SWAPPING) 147 static int vm_swap_enabled=0; 148 static int vm_swap_idle_enabled=0; 149 #else 150 static int vm_swap_enabled=1; 151 static int vm_swap_idle_enabled=0; 152 #endif 153 154 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm, 155 CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt"); 156 157 SYSCTL_INT(_vm, OID_AUTO, max_launder, 158 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout"); 159 160 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max, 161 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length"); 162 163 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval, 164 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan"); 165 166 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval, 167 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan"); 168 169 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_free_max, 170 CTLFLAG_RW, &vm_pageout_stats_free_max, 0, "Not implemented"); 171 172 #if defined(NO_SWAPPING) 173 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 174 CTLFLAG_RD, &vm_swap_enabled, 0, ""); 175 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 176 CTLFLAG_RD, &vm_swap_idle_enabled, 0, ""); 177 #else 178 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 179 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout"); 180 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 181 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 182 #endif 183 184 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts, 185 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem"); 186 187 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts, 188 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages"); 189 190 static int pageout_lock_miss; 191 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss, 192 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout"); 193 194 #define VM_PAGEOUT_PAGE_COUNT 16 195 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT; 196 197 int vm_page_max_wired; /* XXX max # of wired pages system-wide */ 198 199 #if !defined(NO_SWAPPING) 200 typedef void freeer_fcn_t __P((vm_map_t, vm_object_t, vm_pindex_t, int)); 201 static void vm_pageout_map_deactivate_pages __P((vm_map_t, vm_pindex_t)); 202 static freeer_fcn_t vm_pageout_object_deactivate_pages; 203 static void vm_req_vmdaemon __P((void)); 204 #endif 205 static void vm_pageout_page_stats(void); 206 207 /* 208 * vm_pageout_clean: 209 * 210 * Clean the page and remove it from the laundry. 211 * 212 * We set the busy bit to cause potential page faults on this page to 213 * block. Note the careful timing, however, the busy bit isn't set till 214 * late and we cannot do anything that will mess with the page. 215 */ 216 217 static int 218 vm_pageout_clean(m) 219 vm_page_t m; 220 { 221 register vm_object_t object; 222 vm_page_t mc[2*vm_pageout_page_count]; 223 int pageout_count; 224 int ib, is, page_base; 225 vm_pindex_t pindex = m->pindex; 226 227 object = m->object; 228 229 /* 230 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP 231 * with the new swapper, but we could have serious problems paging 232 * out other object types if there is insufficient memory. 233 * 234 * Unfortunately, checking free memory here is far too late, so the 235 * check has been moved up a procedural level. 236 */ 237 238 /* 239 * Don't mess with the page if it's busy, held, or special 240 */ 241 if ((m->hold_count != 0) || 242 ((m->busy != 0) || (m->flags & (PG_BUSY|PG_UNMANAGED)))) { 243 return 0; 244 } 245 246 mc[vm_pageout_page_count] = m; 247 pageout_count = 1; 248 page_base = vm_pageout_page_count; 249 ib = 1; 250 is = 1; 251 252 /* 253 * Scan object for clusterable pages. 254 * 255 * We can cluster ONLY if: ->> the page is NOT 256 * clean, wired, busy, held, or mapped into a 257 * buffer, and one of the following: 258 * 1) The page is inactive, or a seldom used 259 * active page. 260 * -or- 261 * 2) we force the issue. 262 * 263 * During heavy mmap/modification loads the pageout 264 * daemon can really fragment the underlying file 265 * due to flushing pages out of order and not trying 266 * align the clusters (which leave sporatic out-of-order 267 * holes). To solve this problem we do the reverse scan 268 * first and attempt to align our cluster, then do a 269 * forward scan if room remains. 270 */ 271 272 more: 273 while (ib && pageout_count < vm_pageout_page_count) { 274 vm_page_t p; 275 276 if (ib > pindex) { 277 ib = 0; 278 break; 279 } 280 281 if ((p = vm_page_lookup(object, pindex - ib)) == NULL) { 282 ib = 0; 283 break; 284 } 285 if (((p->queue - p->pc) == PQ_CACHE) || 286 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) { 287 ib = 0; 288 break; 289 } 290 vm_page_test_dirty(p); 291 if ((p->dirty & p->valid) == 0 || 292 p->queue != PQ_INACTIVE || 293 p->wire_count != 0 || /* may be held by buf cache */ 294 p->hold_count != 0) { /* may be undergoing I/O */ 295 ib = 0; 296 break; 297 } 298 mc[--page_base] = p; 299 ++pageout_count; 300 ++ib; 301 /* 302 * alignment boundry, stop here and switch directions. Do 303 * not clear ib. 304 */ 305 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0) 306 break; 307 } 308 309 while (pageout_count < vm_pageout_page_count && 310 pindex + is < object->size) { 311 vm_page_t p; 312 313 if ((p = vm_page_lookup(object, pindex + is)) == NULL) 314 break; 315 if (((p->queue - p->pc) == PQ_CACHE) || 316 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) { 317 break; 318 } 319 vm_page_test_dirty(p); 320 if ((p->dirty & p->valid) == 0 || 321 p->queue != PQ_INACTIVE || 322 p->wire_count != 0 || /* may be held by buf cache */ 323 p->hold_count != 0) { /* may be undergoing I/O */ 324 break; 325 } 326 mc[page_base + pageout_count] = p; 327 ++pageout_count; 328 ++is; 329 } 330 331 /* 332 * If we exhausted our forward scan, continue with the reverse scan 333 * when possible, even past a page boundry. This catches boundry 334 * conditions. 335 */ 336 if (ib && pageout_count < vm_pageout_page_count) 337 goto more; 338 339 /* 340 * we allow reads during pageouts... 341 */ 342 return vm_pageout_flush(&mc[page_base], pageout_count, 0); 343 } 344 345 /* 346 * vm_pageout_flush() - launder the given pages 347 * 348 * The given pages are laundered. Note that we setup for the start of 349 * I/O ( i.e. busy the page ), mark it read-only, and bump the object 350 * reference count all in here rather then in the parent. If we want 351 * the parent to do more sophisticated things we may have to change 352 * the ordering. 353 */ 354 355 int 356 vm_pageout_flush(mc, count, flags) 357 vm_page_t *mc; 358 int count; 359 int flags; 360 { 361 register vm_object_t object; 362 int pageout_status[count]; 363 int numpagedout = 0; 364 int i; 365 366 /* 367 * Initiate I/O. Bump the vm_page_t->busy counter and 368 * mark the pages read-only. 369 * 370 * We do not have to fixup the clean/dirty bits here... we can 371 * allow the pager to do it after the I/O completes. 372 */ 373 374 for (i = 0; i < count; i++) { 375 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, ("vm_pageout_flush page %p index %d/%d: partially invalid page", mc[i], i, count)); 376 vm_page_io_start(mc[i]); 377 vm_page_protect(mc[i], VM_PROT_READ); 378 } 379 380 object = mc[0]->object; 381 vm_object_pip_add(object, count); 382 383 vm_pager_put_pages(object, mc, count, 384 (flags | ((object == kernel_object) ? VM_PAGER_PUT_SYNC : 0)), 385 pageout_status); 386 387 for (i = 0; i < count; i++) { 388 vm_page_t mt = mc[i]; 389 390 switch (pageout_status[i]) { 391 case VM_PAGER_OK: 392 numpagedout++; 393 break; 394 case VM_PAGER_PEND: 395 numpagedout++; 396 break; 397 case VM_PAGER_BAD: 398 /* 399 * Page outside of range of object. Right now we 400 * essentially lose the changes by pretending it 401 * worked. 402 */ 403 pmap_clear_modify(mt); 404 vm_page_undirty(mt); 405 break; 406 case VM_PAGER_ERROR: 407 case VM_PAGER_FAIL: 408 /* 409 * If page couldn't be paged out, then reactivate the 410 * page so it doesn't clog the inactive list. (We 411 * will try paging out it again later). 412 */ 413 vm_page_activate(mt); 414 break; 415 case VM_PAGER_AGAIN: 416 break; 417 } 418 419 /* 420 * If the operation is still going, leave the page busy to 421 * block all other accesses. Also, leave the paging in 422 * progress indicator set so that we don't attempt an object 423 * collapse. 424 */ 425 if (pageout_status[i] != VM_PAGER_PEND) { 426 vm_object_pip_wakeup(object); 427 vm_page_io_finish(mt); 428 if (!vm_page_count_severe() || !vm_page_try_to_cache(mt)) 429 vm_page_protect(mt, VM_PROT_READ); 430 } 431 } 432 return numpagedout; 433 } 434 435 #if !defined(NO_SWAPPING) 436 /* 437 * vm_pageout_object_deactivate_pages 438 * 439 * deactivate enough pages to satisfy the inactive target 440 * requirements or if vm_page_proc_limit is set, then 441 * deactivate all of the pages in the object and its 442 * backing_objects. 443 * 444 * The object and map must be locked. 445 */ 446 static void 447 vm_pageout_object_deactivate_pages(map, object, desired, map_remove_only) 448 vm_map_t map; 449 vm_object_t object; 450 vm_pindex_t desired; 451 int map_remove_only; 452 { 453 register vm_page_t p, next; 454 int rcount; 455 int remove_mode; 456 int s; 457 458 if (object->type == OBJT_DEVICE || object->type == OBJT_PHYS) 459 return; 460 461 while (object) { 462 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 463 return; 464 if (object->paging_in_progress) 465 return; 466 467 remove_mode = map_remove_only; 468 if (object->shadow_count > 1) 469 remove_mode = 1; 470 /* 471 * scan the objects entire memory queue 472 */ 473 rcount = object->resident_page_count; 474 p = TAILQ_FIRST(&object->memq); 475 while (p && (rcount-- > 0)) { 476 int actcount; 477 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 478 return; 479 next = TAILQ_NEXT(p, listq); 480 cnt.v_pdpages++; 481 if (p->wire_count != 0 || 482 p->hold_count != 0 || 483 p->busy != 0 || 484 (p->flags & (PG_BUSY|PG_UNMANAGED)) || 485 !pmap_page_exists_quick(vm_map_pmap(map), p)) { 486 p = next; 487 continue; 488 } 489 490 actcount = pmap_ts_referenced(p); 491 if (actcount) { 492 vm_page_flag_set(p, PG_REFERENCED); 493 } else if (p->flags & PG_REFERENCED) { 494 actcount = 1; 495 } 496 497 if ((p->queue != PQ_ACTIVE) && 498 (p->flags & PG_REFERENCED)) { 499 vm_page_activate(p); 500 p->act_count += actcount; 501 vm_page_flag_clear(p, PG_REFERENCED); 502 } else if (p->queue == PQ_ACTIVE) { 503 if ((p->flags & PG_REFERENCED) == 0) { 504 p->act_count -= min(p->act_count, ACT_DECLINE); 505 if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) { 506 vm_page_protect(p, VM_PROT_NONE); 507 vm_page_deactivate(p); 508 } else { 509 s = splvm(); 510 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, p, pageq); 511 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, p, pageq); 512 splx(s); 513 } 514 } else { 515 vm_page_activate(p); 516 vm_page_flag_clear(p, PG_REFERENCED); 517 if (p->act_count < (ACT_MAX - ACT_ADVANCE)) 518 p->act_count += ACT_ADVANCE; 519 s = splvm(); 520 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, p, pageq); 521 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, p, pageq); 522 splx(s); 523 } 524 } else if (p->queue == PQ_INACTIVE) { 525 vm_page_protect(p, VM_PROT_NONE); 526 } 527 p = next; 528 } 529 object = object->backing_object; 530 } 531 return; 532 } 533 534 /* 535 * deactivate some number of pages in a map, try to do it fairly, but 536 * that is really hard to do. 537 */ 538 static void 539 vm_pageout_map_deactivate_pages(map, desired) 540 vm_map_t map; 541 vm_pindex_t desired; 542 { 543 vm_map_entry_t tmpe; 544 vm_object_t obj, bigobj; 545 int nothingwired; 546 547 if (lockmgr(&map->lock, LK_EXCLUSIVE | LK_NOWAIT, (void *)0, curproc)) { 548 return; 549 } 550 551 bigobj = NULL; 552 nothingwired = TRUE; 553 554 /* 555 * first, search out the biggest object, and try to free pages from 556 * that. 557 */ 558 tmpe = map->header.next; 559 while (tmpe != &map->header) { 560 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 561 obj = tmpe->object.vm_object; 562 if ((obj != NULL) && (obj->shadow_count <= 1) && 563 ((bigobj == NULL) || 564 (bigobj->resident_page_count < obj->resident_page_count))) { 565 bigobj = obj; 566 } 567 } 568 if (tmpe->wired_count > 0) 569 nothingwired = FALSE; 570 tmpe = tmpe->next; 571 } 572 573 if (bigobj) 574 vm_pageout_object_deactivate_pages(map, bigobj, desired, 0); 575 576 /* 577 * Next, hunt around for other pages to deactivate. We actually 578 * do this search sort of wrong -- .text first is not the best idea. 579 */ 580 tmpe = map->header.next; 581 while (tmpe != &map->header) { 582 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 583 break; 584 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 585 obj = tmpe->object.vm_object; 586 if (obj) 587 vm_pageout_object_deactivate_pages(map, obj, desired, 0); 588 } 589 tmpe = tmpe->next; 590 }; 591 592 /* 593 * Remove all mappings if a process is swapped out, this will free page 594 * table pages. 595 */ 596 if (desired == 0 && nothingwired) 597 pmap_remove(vm_map_pmap(map), 598 VM_MIN_ADDRESS, VM_MAXUSER_ADDRESS); 599 vm_map_unlock(map); 600 return; 601 } 602 #endif 603 604 /* 605 * Don't try to be fancy - being fancy can lead to VOP_LOCK's and therefore 606 * to vnode deadlocks. We only do it for OBJT_DEFAULT and OBJT_SWAP objects 607 * which we know can be trivially freed. 608 */ 609 610 void 611 vm_pageout_page_free(vm_page_t m) { 612 vm_object_t object = m->object; 613 int type = object->type; 614 615 if (type == OBJT_SWAP || type == OBJT_DEFAULT) 616 vm_object_reference(object); 617 vm_page_busy(m); 618 vm_page_protect(m, VM_PROT_NONE); 619 vm_page_free(m); 620 if (type == OBJT_SWAP || type == OBJT_DEFAULT) 621 vm_object_deallocate(object); 622 } 623 624 /* 625 * vm_pageout_scan does the dirty work for the pageout daemon. 626 */ 627 static void 628 vm_pageout_scan(int pass) 629 { 630 vm_page_t m, next; 631 struct vm_page marker; 632 int page_shortage, maxscan, pcount; 633 int addl_page_shortage, addl_page_shortage_init; 634 struct proc *p, *bigproc; 635 vm_offset_t size, bigsize; 636 vm_object_t object; 637 int actcount; 638 int vnodes_skipped = 0; 639 int maxlaunder; 640 int s; 641 642 /* 643 * Do whatever cleanup that the pmap code can. 644 */ 645 pmap_collect(); 646 647 addl_page_shortage_init = vm_pageout_deficit; 648 vm_pageout_deficit = 0; 649 650 /* 651 * Calculate the number of pages we want to either free or move 652 * to the cache. 653 */ 654 page_shortage = vm_paging_target() + addl_page_shortage_init; 655 656 /* 657 * Initialize our marker 658 */ 659 bzero(&marker, sizeof(marker)); 660 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER; 661 marker.queue = PQ_INACTIVE; 662 marker.wire_count = 1; 663 664 /* 665 * Start scanning the inactive queue for pages we can move to the 666 * cache or free. The scan will stop when the target is reached or 667 * we have scanned the entire inactive queue. Note that m->act_count 668 * is not used to form decisions for the inactive queue, only for the 669 * active queue. 670 * 671 * maxlaunder limits the number of dirty pages we flush per scan. 672 * For most systems a smaller value (16 or 32) is more robust under 673 * extreme memory and disk pressure because any unnecessary writes 674 * to disk can result in extreme performance degredation. However, 675 * systems with excessive dirty pages (especially when MAP_NOSYNC is 676 * used) will die horribly with limited laundering. If the pageout 677 * daemon cannot clean enough pages in the first pass, we let it go 678 * all out in succeeding passes. 679 */ 680 if ((maxlaunder = vm_max_launder) <= 1) 681 maxlaunder = 1; 682 if (pass) 683 maxlaunder = 10000; 684 685 rescan0: 686 addl_page_shortage = addl_page_shortage_init; 687 maxscan = cnt.v_inactive_count; 688 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl); 689 m != NULL && maxscan-- > 0 && page_shortage > 0; 690 m = next) { 691 692 cnt.v_pdpages++; 693 694 if (m->queue != PQ_INACTIVE) { 695 goto rescan0; 696 } 697 698 next = TAILQ_NEXT(m, pageq); 699 700 /* 701 * skip marker pages 702 */ 703 if (m->flags & PG_MARKER) 704 continue; 705 706 /* 707 * A held page may be undergoing I/O, so skip it. 708 */ 709 if (m->hold_count) { 710 s = splvm(); 711 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 712 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 713 splx(s); 714 addl_page_shortage++; 715 continue; 716 } 717 /* 718 * Dont mess with busy pages, keep in the front of the 719 * queue, most likely are being paged out. 720 */ 721 if (m->busy || (m->flags & PG_BUSY)) { 722 addl_page_shortage++; 723 continue; 724 } 725 726 /* 727 * If the object is not being used, we ignore previous 728 * references. 729 */ 730 if (m->object->ref_count == 0) { 731 vm_page_flag_clear(m, PG_REFERENCED); 732 pmap_clear_reference(m); 733 734 /* 735 * Otherwise, if the page has been referenced while in the 736 * inactive queue, we bump the "activation count" upwards, 737 * making it less likely that the page will be added back to 738 * the inactive queue prematurely again. Here we check the 739 * page tables (or emulated bits, if any), given the upper 740 * level VM system not knowing anything about existing 741 * references. 742 */ 743 } else if (((m->flags & PG_REFERENCED) == 0) && 744 (actcount = pmap_ts_referenced(m))) { 745 vm_page_activate(m); 746 m->act_count += (actcount + ACT_ADVANCE); 747 continue; 748 } 749 750 /* 751 * If the upper level VM system knows about any page 752 * references, we activate the page. We also set the 753 * "activation count" higher than normal so that we will less 754 * likely place pages back onto the inactive queue again. 755 */ 756 if ((m->flags & PG_REFERENCED) != 0) { 757 vm_page_flag_clear(m, PG_REFERENCED); 758 actcount = pmap_ts_referenced(m); 759 vm_page_activate(m); 760 m->act_count += (actcount + ACT_ADVANCE + 1); 761 continue; 762 } 763 764 /* 765 * If the upper level VM system doesn't know anything about 766 * the page being dirty, we have to check for it again. As 767 * far as the VM code knows, any partially dirty pages are 768 * fully dirty. 769 */ 770 if (m->dirty == 0) { 771 vm_page_test_dirty(m); 772 } else { 773 vm_page_dirty(m); 774 } 775 776 /* 777 * Invalid pages can be easily freed 778 */ 779 if (m->valid == 0) { 780 vm_pageout_page_free(m); 781 cnt.v_dfree++; 782 --page_shortage; 783 784 /* 785 * Clean pages can be placed onto the cache queue. This 786 * effectively frees them. 787 */ 788 } else if (m->dirty == 0) { 789 /* 790 * Clean pages can be immediately freed to the cache. 791 */ 792 vm_page_cache(m); 793 --page_shortage; 794 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) { 795 /* 796 * Dirty pages need to be paged out, but flushing 797 * a page is extremely expensive verses freeing 798 * a clean page. Rather then artificially limiting 799 * the number of pages we can flush, we instead give 800 * dirty pages extra priority on the inactive queue 801 * by forcing them to be cycled through the queue 802 * twice before being flushed, after which the 803 * (now clean) page will cycle through once more 804 * before being freed. This significantly extends 805 * the thrash point for a heavily loaded machine. 806 */ 807 s = splvm(); 808 vm_page_flag_set(m, PG_WINATCFLS); 809 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 810 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 811 splx(s); 812 } else if (maxlaunder > 0) { 813 /* 814 * We always want to try to flush some dirty pages if 815 * we encounter them, to keep the system stable. 816 * Normally this number is small, but under extreme 817 * pressure where there are insufficient clean pages 818 * on the inactive queue, we may have to go all out. 819 */ 820 int swap_pageouts_ok; 821 struct vnode *vp = NULL; 822 823 object = m->object; 824 825 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 826 swap_pageouts_ok = 1; 827 } else { 828 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 829 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 830 vm_page_count_min()); 831 832 } 833 834 /* 835 * We don't bother paging objects that are "dead". 836 * Those objects are in a "rundown" state. 837 */ 838 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 839 s = splvm(); 840 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 841 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 842 splx(s); 843 continue; 844 } 845 846 /* 847 * The object is already known NOT to be dead. It 848 * is possible for the vget() to block the whole 849 * pageout daemon, but the new low-memory handling 850 * code should prevent it. 851 * 852 * The previous code skipped locked vnodes and, worse, 853 * reordered pages in the queue. This results in 854 * completely non-deterministic operation because, 855 * quite often, a vm_fault has initiated an I/O and 856 * is holding a locked vnode at just the point where 857 * the pageout daemon is woken up. 858 * 859 * We can't wait forever for the vnode lock, we might 860 * deadlock due to a vn_read() getting stuck in 861 * vm_wait while holding this vnode. We skip the 862 * vnode if we can't get it in a reasonable amount 863 * of time. 864 */ 865 866 if (object->type == OBJT_VNODE) { 867 vp = object->handle; 868 869 if (vget(vp, LK_EXCLUSIVE|LK_NOOBJ|LK_TIMELOCK, curproc)) { 870 ++pageout_lock_miss; 871 if (object->flags & OBJ_MIGHTBEDIRTY) 872 vnodes_skipped++; 873 continue; 874 } 875 876 /* 877 * The page might have been moved to another 878 * queue during potential blocking in vget() 879 * above. The page might have been freed and 880 * reused for another vnode. The object might 881 * have been reused for another vnode. 882 */ 883 if (m->queue != PQ_INACTIVE || 884 m->object != object || 885 object->handle != vp) { 886 if (object->flags & OBJ_MIGHTBEDIRTY) 887 vnodes_skipped++; 888 vput(vp); 889 continue; 890 } 891 892 /* 893 * The page may have been busied during the 894 * blocking in vput(); We don't move the 895 * page back onto the end of the queue so that 896 * statistics are more correct if we don't. 897 */ 898 if (m->busy || (m->flags & PG_BUSY)) { 899 vput(vp); 900 continue; 901 } 902 903 /* 904 * If the page has become held it might 905 * be undergoing I/O, so skip it 906 */ 907 if (m->hold_count) { 908 s = splvm(); 909 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 910 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 911 splx(s); 912 if (object->flags & OBJ_MIGHTBEDIRTY) 913 vnodes_skipped++; 914 vput(vp); 915 continue; 916 } 917 } 918 919 /* 920 * If a page is dirty, then it is either being washed 921 * (but not yet cleaned) or it is still in the 922 * laundry. If it is still in the laundry, then we 923 * start the cleaning operation. 924 * 925 * This operation may cluster, invalidating the 'next' 926 * pointer. To prevent an inordinate number of 927 * restarts we use our marker to remember our place. 928 * 929 * decrement page_shortage on success to account for 930 * the (future) cleaned page. Otherwise we could wind 931 * up laundering or cleaning too many pages. 932 */ 933 s = splvm(); 934 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, m, &marker, pageq); 935 splx(s); 936 if (vm_pageout_clean(m) != 0) { 937 --page_shortage; 938 --maxlaunder; 939 } 940 s = splvm(); 941 next = TAILQ_NEXT(&marker, pageq); 942 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, &marker, pageq); 943 splx(s); 944 if (vp != NULL) 945 vput(vp); 946 } 947 } 948 949 /* 950 * Compute the number of pages we want to try to move from the 951 * active queue to the inactive queue. 952 */ 953 page_shortage = vm_paging_target() + 954 cnt.v_inactive_target - cnt.v_inactive_count; 955 page_shortage += addl_page_shortage; 956 957 /* 958 * Scan the active queue for things we can deactivate. We nominally 959 * track the per-page activity counter and use it to locate 960 * deactivation candidates. 961 */ 962 963 pcount = cnt.v_active_count; 964 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 965 966 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) { 967 968 /* 969 * This is a consistency check, and should likely be a panic 970 * or warning. 971 */ 972 if (m->queue != PQ_ACTIVE) { 973 break; 974 } 975 976 next = TAILQ_NEXT(m, pageq); 977 /* 978 * Don't deactivate pages that are busy. 979 */ 980 if ((m->busy != 0) || 981 (m->flags & PG_BUSY) || 982 (m->hold_count != 0)) { 983 s = splvm(); 984 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 985 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 986 splx(s); 987 m = next; 988 continue; 989 } 990 991 /* 992 * The count for pagedaemon pages is done after checking the 993 * page for eligibility... 994 */ 995 cnt.v_pdpages++; 996 997 /* 998 * Check to see "how much" the page has been used. 999 */ 1000 actcount = 0; 1001 if (m->object->ref_count != 0) { 1002 if (m->flags & PG_REFERENCED) { 1003 actcount += 1; 1004 } 1005 actcount += pmap_ts_referenced(m); 1006 if (actcount) { 1007 m->act_count += ACT_ADVANCE + actcount; 1008 if (m->act_count > ACT_MAX) 1009 m->act_count = ACT_MAX; 1010 } 1011 } 1012 1013 /* 1014 * Since we have "tested" this bit, we need to clear it now. 1015 */ 1016 vm_page_flag_clear(m, PG_REFERENCED); 1017 1018 /* 1019 * Only if an object is currently being used, do we use the 1020 * page activation count stats. 1021 */ 1022 if (actcount && (m->object->ref_count != 0)) { 1023 s = splvm(); 1024 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1025 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1026 splx(s); 1027 } else { 1028 m->act_count -= min(m->act_count, ACT_DECLINE); 1029 if (vm_pageout_algorithm || 1030 m->object->ref_count == 0 || 1031 m->act_count == 0) { 1032 page_shortage--; 1033 if (m->object->ref_count == 0) { 1034 vm_page_protect(m, VM_PROT_NONE); 1035 if (m->dirty == 0) 1036 vm_page_cache(m); 1037 else 1038 vm_page_deactivate(m); 1039 } else { 1040 vm_page_deactivate(m); 1041 } 1042 } else { 1043 s = splvm(); 1044 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1045 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1046 splx(s); 1047 } 1048 } 1049 m = next; 1050 } 1051 1052 s = splvm(); 1053 1054 /* 1055 * We try to maintain some *really* free pages, this allows interrupt 1056 * code to be guaranteed space. Since both cache and free queues 1057 * are considered basically 'free', moving pages from cache to free 1058 * does not effect other calculations. 1059 */ 1060 1061 while (cnt.v_free_count < cnt.v_free_reserved) { 1062 static int cache_rover = 0; 1063 m = vm_page_list_find(PQ_CACHE, cache_rover, FALSE); 1064 if (!m) 1065 break; 1066 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || 1067 m->busy || 1068 m->hold_count || 1069 m->wire_count) { 1070 #ifdef INVARIANTS 1071 printf("Warning: busy page %p found in cache\n", m); 1072 #endif 1073 vm_page_deactivate(m); 1074 continue; 1075 } 1076 cache_rover = (cache_rover + PQ_PRIME2) & PQ_L2_MASK; 1077 vm_pageout_page_free(m); 1078 cnt.v_dfree++; 1079 } 1080 splx(s); 1081 1082 #if !defined(NO_SWAPPING) 1083 /* 1084 * Idle process swapout -- run once per second. 1085 */ 1086 if (vm_swap_idle_enabled) { 1087 static long lsec; 1088 if (time_second != lsec) { 1089 vm_pageout_req_swapout |= VM_SWAP_IDLE; 1090 vm_req_vmdaemon(); 1091 lsec = time_second; 1092 } 1093 } 1094 #endif 1095 1096 /* 1097 * If we didn't get enough free pages, and we have skipped a vnode 1098 * in a writeable object, wakeup the sync daemon. And kick swapout 1099 * if we did not get enough free pages. 1100 */ 1101 if (vm_paging_target() > 0) { 1102 if (vnodes_skipped && vm_page_count_min()) 1103 (void) speedup_syncer(); 1104 #if !defined(NO_SWAPPING) 1105 if (vm_swap_enabled && vm_page_count_target()) { 1106 vm_req_vmdaemon(); 1107 vm_pageout_req_swapout |= VM_SWAP_NORMAL; 1108 } 1109 #endif 1110 } 1111 1112 /* 1113 * If we are out of swap and were not able to reach our paging 1114 * target, kill the largest process. 1115 */ 1116 if ((vm_swap_size < 64 && vm_page_count_min()) || 1117 (swap_pager_full && vm_paging_target() > 0)) { 1118 #if 0 1119 if ((vm_swap_size < 64 || swap_pager_full) && vm_page_count_min()) { 1120 #endif 1121 bigproc = NULL; 1122 bigsize = 0; 1123 for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) { 1124 /* 1125 * if this is a system process, skip it 1126 */ 1127 if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) || 1128 ((p->p_pid < 48) && (vm_swap_size != 0))) { 1129 continue; 1130 } 1131 /* 1132 * if the process is in a non-running type state, 1133 * don't touch it. 1134 */ 1135 if (p->p_stat != SRUN && p->p_stat != SSLEEP) { 1136 continue; 1137 } 1138 /* 1139 * get the process size 1140 */ 1141 size = vmspace_resident_count(p->p_vmspace) + 1142 vmspace_swap_count(p->p_vmspace); 1143 /* 1144 * if the this process is bigger than the biggest one 1145 * remember it. 1146 */ 1147 if (size > bigsize) { 1148 bigproc = p; 1149 bigsize = size; 1150 } 1151 } 1152 if (bigproc != NULL) { 1153 killproc(bigproc, "out of swap space"); 1154 bigproc->p_estcpu = 0; 1155 bigproc->p_nice = PRIO_MIN; 1156 resetpriority(bigproc); 1157 wakeup(&cnt.v_free_count); 1158 } 1159 } 1160 } 1161 1162 /* 1163 * This routine tries to maintain the pseudo LRU active queue, 1164 * so that during long periods of time where there is no paging, 1165 * that some statistic accumulation still occurs. This code 1166 * helps the situation where paging just starts to occur. 1167 */ 1168 static void 1169 vm_pageout_page_stats() 1170 { 1171 int s; 1172 vm_page_t m,next; 1173 int pcount,tpcount; /* Number of pages to check */ 1174 static int fullintervalcount = 0; 1175 int page_shortage; 1176 int s0; 1177 1178 page_shortage = 1179 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) - 1180 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count); 1181 1182 if (page_shortage <= 0) 1183 return; 1184 1185 s0 = splvm(); 1186 1187 pcount = cnt.v_active_count; 1188 fullintervalcount += vm_pageout_stats_interval; 1189 if (fullintervalcount < vm_pageout_full_stats_interval) { 1190 tpcount = (vm_pageout_stats_max * cnt.v_active_count) / cnt.v_page_count; 1191 if (pcount > tpcount) 1192 pcount = tpcount; 1193 } else { 1194 fullintervalcount = 0; 1195 } 1196 1197 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1198 while ((m != NULL) && (pcount-- > 0)) { 1199 int actcount; 1200 1201 if (m->queue != PQ_ACTIVE) { 1202 break; 1203 } 1204 1205 next = TAILQ_NEXT(m, pageq); 1206 /* 1207 * Don't deactivate pages that are busy. 1208 */ 1209 if ((m->busy != 0) || 1210 (m->flags & PG_BUSY) || 1211 (m->hold_count != 0)) { 1212 s = splvm(); 1213 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1214 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1215 splx(s); 1216 m = next; 1217 continue; 1218 } 1219 1220 actcount = 0; 1221 if (m->flags & PG_REFERENCED) { 1222 vm_page_flag_clear(m, PG_REFERENCED); 1223 actcount += 1; 1224 } 1225 1226 actcount += pmap_ts_referenced(m); 1227 if (actcount) { 1228 m->act_count += ACT_ADVANCE + actcount; 1229 if (m->act_count > ACT_MAX) 1230 m->act_count = ACT_MAX; 1231 s = splvm(); 1232 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1233 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1234 splx(s); 1235 } else { 1236 if (m->act_count == 0) { 1237 /* 1238 * We turn off page access, so that we have 1239 * more accurate RSS stats. We don't do this 1240 * in the normal page deactivation when the 1241 * system is loaded VM wise, because the 1242 * cost of the large number of page protect 1243 * operations would be higher than the value 1244 * of doing the operation. 1245 */ 1246 vm_page_protect(m, VM_PROT_NONE); 1247 vm_page_deactivate(m); 1248 } else { 1249 m->act_count -= min(m->act_count, ACT_DECLINE); 1250 s = splvm(); 1251 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1252 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1253 splx(s); 1254 } 1255 } 1256 1257 m = next; 1258 } 1259 splx(s0); 1260 } 1261 1262 static int 1263 vm_pageout_free_page_calc(count) 1264 vm_size_t count; 1265 { 1266 if (count < cnt.v_page_count) 1267 return 0; 1268 /* 1269 * free_reserved needs to include enough for the largest swap pager 1270 * structures plus enough for any pv_entry structs when paging. 1271 */ 1272 if (cnt.v_page_count > 1024) 1273 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200; 1274 else 1275 cnt.v_free_min = 4; 1276 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1277 cnt.v_interrupt_free_min; 1278 cnt.v_free_reserved = vm_pageout_page_count + 1279 cnt.v_pageout_free_min + (count / 768) + PQ_L2_SIZE; 1280 cnt.v_free_severe = cnt.v_free_min / 2; 1281 cnt.v_free_min += cnt.v_free_reserved; 1282 cnt.v_free_severe += cnt.v_free_reserved; 1283 return 1; 1284 } 1285 1286 1287 /* 1288 * vm_pageout is the high level pageout daemon. 1289 */ 1290 static void 1291 vm_pageout() 1292 { 1293 int pass; 1294 1295 /* 1296 * Initialize some paging parameters. 1297 */ 1298 1299 cnt.v_interrupt_free_min = 2; 1300 if (cnt.v_page_count < 2000) 1301 vm_pageout_page_count = 8; 1302 1303 vm_pageout_free_page_calc(cnt.v_page_count); 1304 /* 1305 * v_free_target and v_cache_min control pageout hysteresis. Note 1306 * that these are more a measure of the VM cache queue hysteresis 1307 * then the VM free queue. Specifically, v_free_target is the 1308 * high water mark (free+cache pages). 1309 * 1310 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the 1311 * low water mark, while v_free_min is the stop. v_cache_min must 1312 * be big enough to handle memory needs while the pageout daemon 1313 * is signalled and run to free more pages. 1314 */ 1315 if (cnt.v_free_count > 6144) 1316 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved; 1317 else 1318 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved; 1319 1320 if (cnt.v_free_count > 2048) { 1321 cnt.v_cache_min = cnt.v_free_target; 1322 cnt.v_cache_max = 2 * cnt.v_cache_min; 1323 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2; 1324 } else { 1325 cnt.v_cache_min = 0; 1326 cnt.v_cache_max = 0; 1327 cnt.v_inactive_target = cnt.v_free_count / 4; 1328 } 1329 if (cnt.v_inactive_target > cnt.v_free_count / 3) 1330 cnt.v_inactive_target = cnt.v_free_count / 3; 1331 1332 /* XXX does not really belong here */ 1333 if (vm_page_max_wired == 0) 1334 vm_page_max_wired = cnt.v_free_count / 3; 1335 1336 if (vm_pageout_stats_max == 0) 1337 vm_pageout_stats_max = cnt.v_free_target; 1338 1339 /* 1340 * Set interval in seconds for stats scan. 1341 */ 1342 if (vm_pageout_stats_interval == 0) 1343 vm_pageout_stats_interval = 5; 1344 if (vm_pageout_full_stats_interval == 0) 1345 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4; 1346 1347 1348 /* 1349 * Set maximum free per pass 1350 */ 1351 if (vm_pageout_stats_free_max == 0) 1352 vm_pageout_stats_free_max = 5; 1353 1354 swap_pager_swap_init(); 1355 pass = 0; 1356 /* 1357 * The pageout daemon is never done, so loop forever. 1358 */ 1359 while (TRUE) { 1360 int error; 1361 int s = splvm(); 1362 1363 /* 1364 * If we have enough free memory, wakeup waiters. Do 1365 * not clear vm_pages_needed until we reach our target, 1366 * otherwise we may be woken up over and over again and 1367 * waste a lot of cpu. 1368 */ 1369 if (vm_pages_needed && !vm_page_count_min()) { 1370 if (vm_paging_needed() <= 0) 1371 vm_pages_needed = 0; 1372 wakeup(&cnt.v_free_count); 1373 } 1374 if (vm_pages_needed) { 1375 /* 1376 * Still not done, take a second pass without waiting 1377 * (unlimited dirty cleaning), otherwise sleep a bit 1378 * and try again. 1379 */ 1380 ++pass; 1381 if (pass > 1) 1382 tsleep(&vm_pages_needed, PVM, "psleep", hz/2); 1383 } else { 1384 /* 1385 * Good enough, sleep & handle stats. Prime the pass 1386 * for the next run. 1387 */ 1388 if (pass > 1) 1389 pass = 1; 1390 else 1391 pass = 0; 1392 error = tsleep(&vm_pages_needed, 1393 PVM, "psleep", vm_pageout_stats_interval * hz); 1394 if (error && !vm_pages_needed) { 1395 splx(s); 1396 pass = 0; 1397 vm_pageout_page_stats(); 1398 continue; 1399 } 1400 } 1401 1402 if (vm_pages_needed) 1403 cnt.v_pdwakeups++; 1404 splx(s); 1405 vm_pageout_scan(pass); 1406 vm_pageout_deficit = 0; 1407 } 1408 } 1409 1410 void 1411 pagedaemon_wakeup() 1412 { 1413 if (!vm_pages_needed && curproc != pageproc) { 1414 vm_pages_needed++; 1415 wakeup(&vm_pages_needed); 1416 } 1417 } 1418 1419 #if !defined(NO_SWAPPING) 1420 static void 1421 vm_req_vmdaemon() 1422 { 1423 static int lastrun = 0; 1424 1425 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1426 wakeup(&vm_daemon_needed); 1427 lastrun = ticks; 1428 } 1429 } 1430 1431 static void 1432 vm_daemon() 1433 { 1434 struct proc *p; 1435 1436 while (TRUE) { 1437 tsleep(&vm_daemon_needed, PPAUSE, "psleep", 0); 1438 if (vm_pageout_req_swapout) { 1439 swapout_procs(vm_pageout_req_swapout); 1440 vm_pageout_req_swapout = 0; 1441 } 1442 /* 1443 * scan the processes for exceeding their rlimits or if 1444 * process is swapped out -- deactivate pages 1445 */ 1446 1447 for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) { 1448 vm_pindex_t limit, size; 1449 1450 /* 1451 * if this is a system process or if we have already 1452 * looked at this process, skip it. 1453 */ 1454 if (p->p_flag & (P_SYSTEM | P_WEXIT)) { 1455 continue; 1456 } 1457 /* 1458 * if the process is in a non-running type state, 1459 * don't touch it. 1460 */ 1461 if (p->p_stat != SRUN && p->p_stat != SSLEEP) { 1462 continue; 1463 } 1464 /* 1465 * get a limit 1466 */ 1467 limit = OFF_TO_IDX( 1468 qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur, 1469 p->p_rlimit[RLIMIT_RSS].rlim_max)); 1470 1471 /* 1472 * let processes that are swapped out really be 1473 * swapped out set the limit to nothing (will force a 1474 * swap-out.) 1475 */ 1476 if ((p->p_flag & P_INMEM) == 0) 1477 limit = 0; /* XXX */ 1478 1479 size = vmspace_resident_count(p->p_vmspace); 1480 if (limit >= 0 && size >= limit) { 1481 vm_pageout_map_deactivate_pages( 1482 &p->p_vmspace->vm_map, limit); 1483 } 1484 } 1485 } 1486 } 1487 #endif 1488