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 * $FreeBSD$ 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/lock.h> 80 #include <sys/mutex.h> 81 #include <sys/proc.h> 82 #include <sys/kthread.h> 83 #include <sys/ktr.h> 84 #include <sys/resourcevar.h> 85 #include <sys/signalvar.h> 86 #include <sys/vnode.h> 87 #include <sys/vmmeter.h> 88 #include <sys/sx.h> 89 #include <sys/sysctl.h> 90 91 #include <vm/vm.h> 92 #include <vm/vm_param.h> 93 #include <vm/vm_object.h> 94 #include <vm/vm_page.h> 95 #include <vm/vm_map.h> 96 #include <vm/vm_pageout.h> 97 #include <vm/vm_pager.h> 98 #include <vm/swap_pager.h> 99 #include <vm/vm_extern.h> 100 #include <vm/uma.h> 101 102 #include <machine/mutex.h> 103 104 /* 105 * System initialization 106 */ 107 108 /* the kernel process "vm_pageout"*/ 109 static void vm_pageout(void); 110 static int vm_pageout_clean(vm_page_t); 111 static void vm_pageout_scan(int pass); 112 static int vm_pageout_free_page_calc(vm_size_t count); 113 struct proc *pageproc; 114 115 static struct kproc_desc page_kp = { 116 "pagedaemon", 117 vm_pageout, 118 &pageproc 119 }; 120 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp) 121 122 #if !defined(NO_SWAPPING) 123 /* the kernel process "vm_daemon"*/ 124 static void vm_daemon(void); 125 static struct proc *vmproc; 126 127 static struct kproc_desc vm_kp = { 128 "vmdaemon", 129 vm_daemon, 130 &vmproc 131 }; 132 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp) 133 #endif 134 135 136 int vm_pages_needed=0; /* Event on which pageout daemon sleeps */ 137 int vm_pageout_deficit=0; /* Estimated number of pages deficit */ 138 int vm_pageout_pages_needed=0; /* flag saying that the pageout daemon needs pages */ 139 140 #if !defined(NO_SWAPPING) 141 static int vm_pageout_req_swapout; /* XXX */ 142 static int vm_daemon_needed; 143 #endif 144 extern int vm_swap_size; 145 static int vm_max_launder = 32; 146 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0; 147 static int vm_pageout_full_stats_interval = 0; 148 static int vm_pageout_stats_free_max=0, vm_pageout_algorithm=0; 149 static int defer_swap_pageouts=0; 150 static int disable_swap_pageouts=0; 151 152 #if defined(NO_SWAPPING) 153 static int vm_swap_enabled=0; 154 static int vm_swap_idle_enabled=0; 155 #else 156 static int vm_swap_enabled=1; 157 static int vm_swap_idle_enabled=0; 158 #endif 159 160 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm, 161 CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt"); 162 163 SYSCTL_INT(_vm, OID_AUTO, max_launder, 164 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout"); 165 166 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max, 167 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length"); 168 169 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval, 170 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan"); 171 172 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval, 173 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan"); 174 175 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_free_max, 176 CTLFLAG_RW, &vm_pageout_stats_free_max, 0, "Not implemented"); 177 178 #if defined(NO_SWAPPING) 179 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 180 CTLFLAG_RD, &vm_swap_enabled, 0, ""); 181 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 182 CTLFLAG_RD, &vm_swap_idle_enabled, 0, ""); 183 #else 184 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 185 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout"); 186 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 187 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 188 #endif 189 190 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts, 191 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem"); 192 193 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts, 194 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages"); 195 196 static int pageout_lock_miss; 197 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss, 198 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout"); 199 200 #define VM_PAGEOUT_PAGE_COUNT 16 201 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT; 202 203 int vm_page_max_wired; /* XXX max # of wired pages system-wide */ 204 205 #if !defined(NO_SWAPPING) 206 typedef void freeer_fcn_t(vm_map_t, vm_object_t, vm_pindex_t, int); 207 static void vm_pageout_map_deactivate_pages(vm_map_t, vm_pindex_t); 208 static freeer_fcn_t vm_pageout_object_deactivate_pages; 209 static void vm_req_vmdaemon(void); 210 #endif 211 static void vm_pageout_page_stats(void); 212 213 /* 214 * vm_pageout_clean: 215 * 216 * Clean the page and remove it from the laundry. 217 * 218 * We set the busy bit to cause potential page faults on this page to 219 * block. Note the careful timing, however, the busy bit isn't set till 220 * late and we cannot do anything that will mess with the page. 221 */ 222 static int 223 vm_pageout_clean(m) 224 vm_page_t m; 225 { 226 vm_object_t object; 227 vm_page_t mc[2*vm_pageout_page_count]; 228 int pageout_count; 229 int ib, is, page_base; 230 vm_pindex_t pindex = m->pindex; 231 232 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 233 234 object = m->object; 235 236 /* 237 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP 238 * with the new swapper, but we could have serious problems paging 239 * out other object types if there is insufficient memory. 240 * 241 * Unfortunately, checking free memory here is far too late, so the 242 * check has been moved up a procedural level. 243 */ 244 245 /* 246 * Don't mess with the page if it's busy, held, or special 247 */ 248 if ((m->hold_count != 0) || 249 ((m->busy != 0) || (m->flags & (PG_BUSY|PG_UNMANAGED)))) { 250 return 0; 251 } 252 253 mc[vm_pageout_page_count] = m; 254 pageout_count = 1; 255 page_base = vm_pageout_page_count; 256 ib = 1; 257 is = 1; 258 259 /* 260 * Scan object for clusterable pages. 261 * 262 * We can cluster ONLY if: ->> the page is NOT 263 * clean, wired, busy, held, or mapped into a 264 * buffer, and one of the following: 265 * 1) The page is inactive, or a seldom used 266 * active page. 267 * -or- 268 * 2) we force the issue. 269 * 270 * During heavy mmap/modification loads the pageout 271 * daemon can really fragment the underlying file 272 * due to flushing pages out of order and not trying 273 * align the clusters (which leave sporatic out-of-order 274 * holes). To solve this problem we do the reverse scan 275 * first and attempt to align our cluster, then do a 276 * forward scan if room remains. 277 */ 278 more: 279 while (ib && pageout_count < vm_pageout_page_count) { 280 vm_page_t p; 281 282 if (ib > pindex) { 283 ib = 0; 284 break; 285 } 286 287 if ((p = vm_page_lookup(object, pindex - ib)) == NULL) { 288 ib = 0; 289 break; 290 } 291 if (((p->queue - p->pc) == PQ_CACHE) || 292 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) { 293 ib = 0; 294 break; 295 } 296 vm_page_test_dirty(p); 297 if ((p->dirty & p->valid) == 0 || 298 p->queue != PQ_INACTIVE || 299 p->wire_count != 0 || /* may be held by buf cache */ 300 p->hold_count != 0) { /* may be undergoing I/O */ 301 ib = 0; 302 break; 303 } 304 mc[--page_base] = p; 305 ++pageout_count; 306 ++ib; 307 /* 308 * alignment boundry, stop here and switch directions. Do 309 * not clear ib. 310 */ 311 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0) 312 break; 313 } 314 315 while (pageout_count < vm_pageout_page_count && 316 pindex + is < object->size) { 317 vm_page_t p; 318 319 if ((p = vm_page_lookup(object, pindex + is)) == NULL) 320 break; 321 if (((p->queue - p->pc) == PQ_CACHE) || 322 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) { 323 break; 324 } 325 vm_page_test_dirty(p); 326 if ((p->dirty & p->valid) == 0 || 327 p->queue != PQ_INACTIVE || 328 p->wire_count != 0 || /* may be held by buf cache */ 329 p->hold_count != 0) { /* may be undergoing I/O */ 330 break; 331 } 332 mc[page_base + pageout_count] = p; 333 ++pageout_count; 334 ++is; 335 } 336 337 /* 338 * If we exhausted our forward scan, continue with the reverse scan 339 * when possible, even past a page boundry. This catches boundry 340 * conditions. 341 */ 342 if (ib && pageout_count < vm_pageout_page_count) 343 goto more; 344 345 /* 346 * we allow reads during pageouts... 347 */ 348 return vm_pageout_flush(&mc[page_base], pageout_count, 0); 349 } 350 351 /* 352 * vm_pageout_flush() - launder the given pages 353 * 354 * The given pages are laundered. Note that we setup for the start of 355 * I/O ( i.e. busy the page ), mark it read-only, and bump the object 356 * reference count all in here rather then in the parent. If we want 357 * the parent to do more sophisticated things we may have to change 358 * the ordering. 359 */ 360 int 361 vm_pageout_flush(mc, count, flags) 362 vm_page_t *mc; 363 int count; 364 int flags; 365 { 366 vm_object_t object; 367 int pageout_status[count]; 368 int numpagedout = 0; 369 int i; 370 371 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 372 /* 373 * Initiate I/O. Bump the vm_page_t->busy counter and 374 * mark the pages read-only. 375 * 376 * We do not have to fixup the clean/dirty bits here... we can 377 * allow the pager to do it after the I/O completes. 378 * 379 * NOTE! mc[i]->dirty may be partial or fragmented due to an 380 * edge case with file fragments. 381 */ 382 for (i = 0; i < count; i++) { 383 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, ("vm_pageout_flush page %p index %d/%d: partially invalid page", mc[i], i, count)); 384 vm_page_io_start(mc[i]); 385 vm_page_protect(mc[i], VM_PROT_READ); 386 } 387 object = mc[0]->object; 388 vm_page_unlock_queues(); 389 vm_object_pip_add(object, count); 390 391 vm_pager_put_pages(object, mc, count, 392 (flags | ((object == kernel_object) ? OBJPC_SYNC : 0)), 393 pageout_status); 394 395 vm_page_lock_queues(); 396 for (i = 0; i < count; i++) { 397 vm_page_t mt = mc[i]; 398 399 switch (pageout_status[i]) { 400 case VM_PAGER_OK: 401 numpagedout++; 402 break; 403 case VM_PAGER_PEND: 404 numpagedout++; 405 break; 406 case VM_PAGER_BAD: 407 /* 408 * Page outside of range of object. Right now we 409 * essentially lose the changes by pretending it 410 * worked. 411 */ 412 pmap_clear_modify(mt); 413 vm_page_undirty(mt); 414 break; 415 case VM_PAGER_ERROR: 416 case VM_PAGER_FAIL: 417 /* 418 * If page couldn't be paged out, then reactivate the 419 * page so it doesn't clog the inactive list. (We 420 * will try paging out it again later). 421 */ 422 vm_page_activate(mt); 423 break; 424 case VM_PAGER_AGAIN: 425 break; 426 } 427 428 /* 429 * If the operation is still going, leave the page busy to 430 * block all other accesses. Also, leave the paging in 431 * progress indicator set so that we don't attempt an object 432 * collapse. 433 */ 434 if (pageout_status[i] != VM_PAGER_PEND) { 435 vm_object_pip_wakeup(object); 436 vm_page_io_finish(mt); 437 if (!vm_page_count_severe() || !vm_page_try_to_cache(mt)) 438 vm_page_protect(mt, VM_PROT_READ); 439 } 440 } 441 return numpagedout; 442 } 443 444 #if !defined(NO_SWAPPING) 445 /* 446 * vm_pageout_object_deactivate_pages 447 * 448 * deactivate enough pages to satisfy the inactive target 449 * requirements or if vm_page_proc_limit is set, then 450 * deactivate all of the pages in the object and its 451 * backing_objects. 452 * 453 * The object and map must be locked. 454 */ 455 static void 456 vm_pageout_object_deactivate_pages(map, object, desired, map_remove_only) 457 vm_map_t map; 458 vm_object_t object; 459 vm_pindex_t desired; 460 int map_remove_only; 461 { 462 vm_page_t p, next; 463 int actcount, rcount, remove_mode; 464 465 GIANT_REQUIRED; 466 if (object->type == OBJT_DEVICE || object->type == OBJT_PHYS) 467 return; 468 469 while (object) { 470 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 471 return; 472 if (object->paging_in_progress) 473 return; 474 475 remove_mode = map_remove_only; 476 if (object->shadow_count > 1) 477 remove_mode = 1; 478 /* 479 * scan the objects entire memory queue 480 */ 481 rcount = object->resident_page_count; 482 p = TAILQ_FIRST(&object->memq); 483 vm_page_lock_queues(); 484 while (p && (rcount-- > 0)) { 485 if (pmap_resident_count(map->pmap) <= desired) { 486 vm_page_unlock_queues(); 487 return; 488 } 489 next = TAILQ_NEXT(p, listq); 490 cnt.v_pdpages++; 491 if (p->wire_count != 0 || 492 p->hold_count != 0 || 493 p->busy != 0 || 494 (p->flags & (PG_BUSY|PG_UNMANAGED)) || 495 !pmap_page_exists_quick(vm_map_pmap(map), p)) { 496 p = next; 497 continue; 498 } 499 actcount = pmap_ts_referenced(p); 500 if (actcount) { 501 vm_page_flag_set(p, PG_REFERENCED); 502 } else if (p->flags & PG_REFERENCED) { 503 actcount = 1; 504 } 505 if ((p->queue != PQ_ACTIVE) && 506 (p->flags & PG_REFERENCED)) { 507 vm_page_activate(p); 508 p->act_count += actcount; 509 vm_page_flag_clear(p, PG_REFERENCED); 510 } else if (p->queue == PQ_ACTIVE) { 511 if ((p->flags & PG_REFERENCED) == 0) { 512 p->act_count -= min(p->act_count, ACT_DECLINE); 513 if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) { 514 vm_page_protect(p, VM_PROT_NONE); 515 vm_page_deactivate(p); 516 } else { 517 vm_pageq_requeue(p); 518 } 519 } else { 520 vm_page_activate(p); 521 vm_page_flag_clear(p, PG_REFERENCED); 522 if (p->act_count < (ACT_MAX - ACT_ADVANCE)) 523 p->act_count += ACT_ADVANCE; 524 vm_pageq_requeue(p); 525 } 526 } else if (p->queue == PQ_INACTIVE) { 527 vm_page_protect(p, VM_PROT_NONE); 528 } 529 p = next; 530 } 531 vm_page_unlock_queues(); 532 object = object->backing_object; 533 } 534 } 535 536 /* 537 * deactivate some number of pages in a map, try to do it fairly, but 538 * that is really hard to do. 539 */ 540 static void 541 vm_pageout_map_deactivate_pages(map, desired) 542 vm_map_t map; 543 vm_pindex_t desired; 544 { 545 vm_map_entry_t tmpe; 546 vm_object_t obj, bigobj; 547 int nothingwired; 548 549 GIANT_REQUIRED; 550 if (!vm_map_trylock(map)) 551 return; 552 553 bigobj = NULL; 554 nothingwired = TRUE; 555 556 /* 557 * first, search out the biggest object, and try to free pages from 558 * that. 559 */ 560 tmpe = map->header.next; 561 while (tmpe != &map->header) { 562 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 563 obj = tmpe->object.vm_object; 564 if ((obj != NULL) && (obj->shadow_count <= 1) && 565 ((bigobj == NULL) || 566 (bigobj->resident_page_count < obj->resident_page_count))) { 567 bigobj = obj; 568 } 569 } 570 if (tmpe->wired_count > 0) 571 nothingwired = FALSE; 572 tmpe = tmpe->next; 573 } 574 575 if (bigobj) 576 vm_pageout_object_deactivate_pages(map, bigobj, desired, 0); 577 578 /* 579 * Next, hunt around for other pages to deactivate. We actually 580 * do this search sort of wrong -- .text first is not the best idea. 581 */ 582 tmpe = map->header.next; 583 while (tmpe != &map->header) { 584 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 585 break; 586 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 587 obj = tmpe->object.vm_object; 588 if (obj) 589 vm_pageout_object_deactivate_pages(map, obj, desired, 0); 590 } 591 tmpe = tmpe->next; 592 }; 593 594 /* 595 * Remove all mappings if a process is swapped out, this will free page 596 * table pages. 597 */ 598 if (desired == 0 && nothingwired) 599 pmap_remove(vm_map_pmap(map), 600 VM_MIN_ADDRESS, VM_MAXUSER_ADDRESS); 601 vm_map_unlock(map); 602 return; 603 } 604 #endif /* !defined(NO_SWAPPING) */ 605 606 /* 607 * Don't try to be fancy - being fancy can lead to VOP_LOCK's and therefore 608 * to vnode deadlocks. We only do it for OBJT_DEFAULT and OBJT_SWAP objects 609 * which we know can be trivially freed. 610 */ 611 void 612 vm_pageout_page_free(vm_page_t m) { 613 vm_object_t object = m->object; 614 int type = object->type; 615 616 GIANT_REQUIRED; 617 if (type == OBJT_SWAP || type == OBJT_DEFAULT) 618 vm_object_reference(object); 619 vm_page_busy(m); 620 vm_page_protect(m, VM_PROT_NONE); 621 vm_page_free(m); 622 cnt.v_dfree++; 623 if (type == OBJT_SWAP || type == OBJT_DEFAULT) 624 vm_object_deallocate(object); 625 } 626 627 /* 628 * vm_pageout_scan does the dirty work for the pageout daemon. 629 */ 630 static void 631 vm_pageout_scan(int pass) 632 { 633 vm_page_t m, next; 634 struct vm_page marker; 635 int save_page_shortage; 636 int save_inactive_count; 637 int page_shortage, maxscan, pcount; 638 int addl_page_shortage, addl_page_shortage_init; 639 struct proc *p, *bigproc; 640 vm_offset_t size, bigsize; 641 vm_object_t object; 642 int actcount; 643 int vnodes_skipped = 0; 644 int maxlaunder; 645 int s; 646 struct thread *td; 647 648 GIANT_REQUIRED; 649 /* 650 * Do whatever cleanup that the pmap code can. 651 */ 652 pmap_collect(); 653 uma_reclaim(); 654 655 addl_page_shortage_init = vm_pageout_deficit; 656 vm_pageout_deficit = 0; 657 658 /* 659 * Calculate the number of pages we want to either free or move 660 * to the cache. 661 */ 662 page_shortage = vm_paging_target() + addl_page_shortage_init; 663 save_page_shortage = page_shortage; 664 save_inactive_count = cnt.v_inactive_count; 665 666 /* 667 * Initialize our marker 668 */ 669 bzero(&marker, sizeof(marker)); 670 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER; 671 marker.queue = PQ_INACTIVE; 672 marker.wire_count = 1; 673 674 /* 675 * Start scanning the inactive queue for pages we can move to the 676 * cache or free. The scan will stop when the target is reached or 677 * we have scanned the entire inactive queue. Note that m->act_count 678 * is not used to form decisions for the inactive queue, only for the 679 * active queue. 680 * 681 * maxlaunder limits the number of dirty pages we flush per scan. 682 * For most systems a smaller value (16 or 32) is more robust under 683 * extreme memory and disk pressure because any unnecessary writes 684 * to disk can result in extreme performance degredation. However, 685 * systems with excessive dirty pages (especially when MAP_NOSYNC is 686 * used) will die horribly with limited laundering. If the pageout 687 * daemon cannot clean enough pages in the first pass, we let it go 688 * all out in succeeding passes. 689 */ 690 if ((maxlaunder = vm_max_launder) <= 1) 691 maxlaunder = 1; 692 if (pass) 693 maxlaunder = 10000; 694 rescan0: 695 addl_page_shortage = addl_page_shortage_init; 696 maxscan = cnt.v_inactive_count; 697 698 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl); 699 m != NULL && maxscan-- > 0 && page_shortage > 0; 700 m = next) { 701 702 cnt.v_pdpages++; 703 704 if (m->queue != PQ_INACTIVE) { 705 goto rescan0; 706 } 707 708 next = TAILQ_NEXT(m, pageq); 709 710 /* 711 * skip marker pages 712 */ 713 if (m->flags & PG_MARKER) 714 continue; 715 716 /* 717 * A held page may be undergoing I/O, so skip it. 718 */ 719 if (m->hold_count) { 720 vm_pageq_requeue(m); 721 addl_page_shortage++; 722 continue; 723 } 724 /* 725 * Don't mess with busy pages, keep in the front of the 726 * queue, most likely are being paged out. 727 */ 728 if (m->busy || (m->flags & PG_BUSY)) { 729 addl_page_shortage++; 730 continue; 731 } 732 733 /* 734 * If the object is not being used, we ignore previous 735 * references. 736 */ 737 if (m->object->ref_count == 0) { 738 vm_page_flag_clear(m, PG_REFERENCED); 739 pmap_clear_reference(m); 740 741 /* 742 * Otherwise, if the page has been referenced while in the 743 * inactive queue, we bump the "activation count" upwards, 744 * making it less likely that the page will be added back to 745 * the inactive queue prematurely again. Here we check the 746 * page tables (or emulated bits, if any), given the upper 747 * level VM system not knowing anything about existing 748 * references. 749 */ 750 } else if (((m->flags & PG_REFERENCED) == 0) && 751 (actcount = pmap_ts_referenced(m))) { 752 vm_page_lock_queues(); 753 vm_page_activate(m); 754 vm_page_unlock_queues(); 755 m->act_count += (actcount + ACT_ADVANCE); 756 continue; 757 } 758 759 /* 760 * If the upper level VM system knows about any page 761 * references, we activate the page. We also set the 762 * "activation count" higher than normal so that we will less 763 * likely place pages back onto the inactive queue again. 764 */ 765 if ((m->flags & PG_REFERENCED) != 0) { 766 vm_page_flag_clear(m, PG_REFERENCED); 767 actcount = pmap_ts_referenced(m); 768 vm_page_lock_queues(); 769 vm_page_activate(m); 770 vm_page_unlock_queues(); 771 m->act_count += (actcount + ACT_ADVANCE + 1); 772 continue; 773 } 774 775 /* 776 * If the upper level VM system doesn't know anything about 777 * the page being dirty, we have to check for it again. As 778 * far as the VM code knows, any partially dirty pages are 779 * fully dirty. 780 */ 781 if (m->dirty == 0) { 782 vm_page_test_dirty(m); 783 } else { 784 vm_page_dirty(m); 785 } 786 787 /* 788 * Invalid pages can be easily freed 789 */ 790 if (m->valid == 0) { 791 vm_page_lock_queues(); 792 vm_pageout_page_free(m); 793 vm_page_unlock_queues(); 794 --page_shortage; 795 796 /* 797 * Clean pages can be placed onto the cache queue. This 798 * effectively frees them. 799 */ 800 } else if (m->dirty == 0) { 801 vm_page_lock_queues(); 802 vm_page_cache(m); 803 vm_page_unlock_queues(); 804 --page_shortage; 805 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) { 806 /* 807 * Dirty pages need to be paged out, but flushing 808 * a page is extremely expensive verses freeing 809 * a clean page. Rather then artificially limiting 810 * the number of pages we can flush, we instead give 811 * dirty pages extra priority on the inactive queue 812 * by forcing them to be cycled through the queue 813 * twice before being flushed, after which the 814 * (now clean) page will cycle through once more 815 * before being freed. This significantly extends 816 * the thrash point for a heavily loaded machine. 817 */ 818 vm_page_flag_set(m, PG_WINATCFLS); 819 vm_pageq_requeue(m); 820 } else if (maxlaunder > 0) { 821 /* 822 * We always want to try to flush some dirty pages if 823 * we encounter them, to keep the system stable. 824 * Normally this number is small, but under extreme 825 * pressure where there are insufficient clean pages 826 * on the inactive queue, we may have to go all out. 827 */ 828 int swap_pageouts_ok; 829 struct vnode *vp = NULL; 830 struct mount *mp; 831 832 object = m->object; 833 834 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 835 swap_pageouts_ok = 1; 836 } else { 837 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 838 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 839 vm_page_count_min()); 840 841 } 842 843 /* 844 * We don't bother paging objects that are "dead". 845 * Those objects are in a "rundown" state. 846 */ 847 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 848 vm_pageq_requeue(m); 849 continue; 850 } 851 852 /* 853 * The object is already known NOT to be dead. It 854 * is possible for the vget() to block the whole 855 * pageout daemon, but the new low-memory handling 856 * code should prevent it. 857 * 858 * The previous code skipped locked vnodes and, worse, 859 * reordered pages in the queue. This results in 860 * completely non-deterministic operation and, on a 861 * busy system, can lead to extremely non-optimal 862 * pageouts. For example, it can cause clean pages 863 * to be freed and dirty pages to be moved to the end 864 * of the queue. Since dirty pages are also moved to 865 * the end of the queue once-cleaned, this gives 866 * way too large a weighting to defering the freeing 867 * of dirty pages. 868 * 869 * We can't wait forever for the vnode lock, we might 870 * deadlock due to a vn_read() getting stuck in 871 * vm_wait while holding this vnode. We skip the 872 * vnode if we can't get it in a reasonable amount 873 * of time. 874 */ 875 if (object->type == OBJT_VNODE) { 876 vp = object->handle; 877 878 mp = NULL; 879 if (vp->v_type == VREG) 880 vn_start_write(vp, &mp, V_NOWAIT); 881 if (vget(vp, LK_EXCLUSIVE|LK_NOOBJ|LK_TIMELOCK, curthread)) { 882 ++pageout_lock_miss; 883 vn_finished_write(mp); 884 if (object->flags & OBJ_MIGHTBEDIRTY) 885 vnodes_skipped++; 886 continue; 887 } 888 889 /* 890 * The page might have been moved to another 891 * queue during potential blocking in vget() 892 * above. The page might have been freed and 893 * reused for another vnode. The object might 894 * have been reused for another vnode. 895 */ 896 if (m->queue != PQ_INACTIVE || 897 m->object != object || 898 object->handle != vp) { 899 if (object->flags & OBJ_MIGHTBEDIRTY) 900 vnodes_skipped++; 901 vput(vp); 902 vn_finished_write(mp); 903 continue; 904 } 905 906 /* 907 * The page may have been busied during the 908 * blocking in vput(); We don't move the 909 * page back onto the end of the queue so that 910 * statistics are more correct if we don't. 911 */ 912 if (m->busy || (m->flags & PG_BUSY)) { 913 vput(vp); 914 vn_finished_write(mp); 915 continue; 916 } 917 918 /* 919 * If the page has become held it might 920 * be undergoing I/O, so skip it 921 */ 922 if (m->hold_count) { 923 vm_pageq_requeue(m); 924 if (object->flags & OBJ_MIGHTBEDIRTY) 925 vnodes_skipped++; 926 vput(vp); 927 vn_finished_write(mp); 928 continue; 929 } 930 } 931 932 /* 933 * If a page is dirty, then it is either being washed 934 * (but not yet cleaned) or it is still in the 935 * laundry. If it is still in the laundry, then we 936 * start the cleaning operation. 937 * 938 * This operation may cluster, invalidating the 'next' 939 * pointer. To prevent an inordinate number of 940 * restarts we use our marker to remember our place. 941 * 942 * decrement page_shortage on success to account for 943 * the (future) cleaned page. Otherwise we could wind 944 * up laundering or cleaning too many pages. 945 */ 946 vm_page_lock_queues(); 947 s = splvm(); 948 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, m, &marker, pageq); 949 splx(s); 950 if (vm_pageout_clean(m) != 0) { 951 --page_shortage; 952 --maxlaunder; 953 } 954 s = splvm(); 955 next = TAILQ_NEXT(&marker, pageq); 956 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, &marker, pageq); 957 splx(s); 958 vm_page_unlock_queues(); 959 if (vp) { 960 vput(vp); 961 vn_finished_write(mp); 962 } 963 } 964 } 965 966 /* 967 * Compute the number of pages we want to try to move from the 968 * active queue to the inactive queue. 969 */ 970 page_shortage = vm_paging_target() + 971 cnt.v_inactive_target - cnt.v_inactive_count; 972 page_shortage += addl_page_shortage; 973 974 vm_page_lock_queues(); 975 /* 976 * Scan the active queue for things we can deactivate. We nominally 977 * track the per-page activity counter and use it to locate 978 * deactivation candidates. 979 */ 980 pcount = cnt.v_active_count; 981 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 982 983 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) { 984 985 /* 986 * This is a consistency check, and should likely be a panic 987 * or warning. 988 */ 989 if (m->queue != PQ_ACTIVE) { 990 break; 991 } 992 993 next = TAILQ_NEXT(m, pageq); 994 /* 995 * Don't deactivate pages that are busy. 996 */ 997 if ((m->busy != 0) || 998 (m->flags & PG_BUSY) || 999 (m->hold_count != 0)) { 1000 vm_pageq_requeue(m); 1001 m = next; 1002 continue; 1003 } 1004 1005 /* 1006 * The count for pagedaemon pages is done after checking the 1007 * page for eligibility... 1008 */ 1009 cnt.v_pdpages++; 1010 1011 /* 1012 * Check to see "how much" the page has been used. 1013 */ 1014 actcount = 0; 1015 if (m->object->ref_count != 0) { 1016 if (m->flags & PG_REFERENCED) { 1017 actcount += 1; 1018 } 1019 actcount += pmap_ts_referenced(m); 1020 if (actcount) { 1021 m->act_count += ACT_ADVANCE + actcount; 1022 if (m->act_count > ACT_MAX) 1023 m->act_count = ACT_MAX; 1024 } 1025 } 1026 1027 /* 1028 * Since we have "tested" this bit, we need to clear it now. 1029 */ 1030 vm_page_flag_clear(m, PG_REFERENCED); 1031 1032 /* 1033 * Only if an object is currently being used, do we use the 1034 * page activation count stats. 1035 */ 1036 if (actcount && (m->object->ref_count != 0)) { 1037 vm_pageq_requeue(m); 1038 } else { 1039 m->act_count -= min(m->act_count, ACT_DECLINE); 1040 if (vm_pageout_algorithm || 1041 m->object->ref_count == 0 || 1042 m->act_count == 0) { 1043 page_shortage--; 1044 if (m->object->ref_count == 0) { 1045 vm_page_protect(m, VM_PROT_NONE); 1046 if (m->dirty == 0) 1047 vm_page_cache(m); 1048 else 1049 vm_page_deactivate(m); 1050 } else { 1051 vm_page_deactivate(m); 1052 } 1053 } else { 1054 vm_pageq_requeue(m); 1055 } 1056 } 1057 m = next; 1058 } 1059 s = splvm(); 1060 1061 /* 1062 * We try to maintain some *really* free pages, this allows interrupt 1063 * code to be guaranteed space. Since both cache and free queues 1064 * are considered basically 'free', moving pages from cache to free 1065 * does not effect other calculations. 1066 */ 1067 while (cnt.v_free_count < cnt.v_free_reserved) { 1068 static int cache_rover = 0; 1069 m = vm_pageq_find(PQ_CACHE, cache_rover, FALSE); 1070 if (!m) 1071 break; 1072 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || 1073 m->busy || 1074 m->hold_count || 1075 m->wire_count) { 1076 #ifdef INVARIANTS 1077 printf("Warning: busy page %p found in cache\n", m); 1078 #endif 1079 vm_page_deactivate(m); 1080 continue; 1081 } 1082 cache_rover = (cache_rover + PQ_PRIME2) & PQ_L2_MASK; 1083 vm_pageout_page_free(m); 1084 } 1085 splx(s); 1086 vm_page_unlock_queues(); 1087 #if !defined(NO_SWAPPING) 1088 /* 1089 * Idle process swapout -- run once per second. 1090 */ 1091 if (vm_swap_idle_enabled) { 1092 static long lsec; 1093 if (time_second != lsec) { 1094 vm_pageout_req_swapout |= VM_SWAP_IDLE; 1095 vm_req_vmdaemon(); 1096 lsec = time_second; 1097 } 1098 } 1099 #endif 1100 1101 /* 1102 * If we didn't get enough free pages, and we have skipped a vnode 1103 * in a writeable object, wakeup the sync daemon. And kick swapout 1104 * if we did not get enough free pages. 1105 */ 1106 if (vm_paging_target() > 0) { 1107 if (vnodes_skipped && vm_page_count_min()) 1108 (void) speedup_syncer(); 1109 #if !defined(NO_SWAPPING) 1110 if (vm_swap_enabled && vm_page_count_target()) { 1111 vm_req_vmdaemon(); 1112 vm_pageout_req_swapout |= VM_SWAP_NORMAL; 1113 } 1114 #endif 1115 } 1116 1117 /* 1118 * If we are out of swap and were not able to reach our paging 1119 * target, kill the largest process. 1120 * 1121 * We keep the process bigproc locked once we find it to keep anyone 1122 * from messing with it; however, there is a possibility of 1123 * deadlock if process B is bigproc and one of it's child processes 1124 * attempts to propagate a signal to B while we are waiting for A's 1125 * lock while walking this list. To avoid this, we don't block on 1126 * the process lock but just skip a process if it is already locked. 1127 */ 1128 if ((vm_swap_size < 64 && vm_page_count_min()) || 1129 (swap_pager_full && vm_paging_target() > 0)) { 1130 #if 0 1131 if ((vm_swap_size < 64 || swap_pager_full) && vm_page_count_min()) { 1132 #endif 1133 bigproc = NULL; 1134 bigsize = 0; 1135 sx_slock(&allproc_lock); 1136 FOREACH_PROC_IN_SYSTEM(p) { 1137 int breakout; 1138 /* 1139 * If this process is already locked, skip it. 1140 */ 1141 if (PROC_TRYLOCK(p) == 0) 1142 continue; 1143 /* 1144 * if this is a system process, skip it 1145 */ 1146 if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) || 1147 ((p->p_pid < 48) && (vm_swap_size != 0))) { 1148 PROC_UNLOCK(p); 1149 continue; 1150 } 1151 /* 1152 * if the process is in a non-running type state, 1153 * don't touch it. Check all the threads individually. 1154 */ 1155 mtx_lock_spin(&sched_lock); 1156 breakout = 0; 1157 FOREACH_THREAD_IN_PROC(p, td) { 1158 if (td->td_state != TDS_RUNQ && 1159 td->td_state != TDS_RUNNING && 1160 td->td_state != TDS_SLP) { 1161 breakout = 1; 1162 break; 1163 } 1164 } 1165 if (breakout) { 1166 mtx_unlock_spin(&sched_lock); 1167 PROC_UNLOCK(p); 1168 continue; 1169 } 1170 mtx_unlock_spin(&sched_lock); 1171 /* 1172 * get the process size 1173 */ 1174 size = vmspace_resident_count(p->p_vmspace) + 1175 vmspace_swap_count(p->p_vmspace); 1176 /* 1177 * if the this process is bigger than the biggest one 1178 * remember it. 1179 */ 1180 if (size > bigsize) { 1181 if (bigproc != NULL) 1182 PROC_UNLOCK(bigproc); 1183 bigproc = p; 1184 bigsize = size; 1185 } else 1186 PROC_UNLOCK(p); 1187 } 1188 sx_sunlock(&allproc_lock); 1189 if (bigproc != NULL) { 1190 struct ksegrp *kg; 1191 killproc(bigproc, "out of swap space"); 1192 mtx_lock_spin(&sched_lock); 1193 FOREACH_KSEGRP_IN_PROC(bigproc, kg) { 1194 kg->kg_estcpu = 0; 1195 kg->kg_nice = PRIO_MIN; /* XXXKSE ??? */ 1196 resetpriority(kg); 1197 } 1198 mtx_unlock_spin(&sched_lock); 1199 PROC_UNLOCK(bigproc); 1200 wakeup(&cnt.v_free_count); 1201 } 1202 } 1203 } 1204 1205 /* 1206 * This routine tries to maintain the pseudo LRU active queue, 1207 * so that during long periods of time where there is no paging, 1208 * that some statistic accumulation still occurs. This code 1209 * helps the situation where paging just starts to occur. 1210 */ 1211 static void 1212 vm_pageout_page_stats() 1213 { 1214 vm_page_t m,next; 1215 int pcount,tpcount; /* Number of pages to check */ 1216 static int fullintervalcount = 0; 1217 int page_shortage; 1218 int s0; 1219 1220 page_shortage = 1221 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) - 1222 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count); 1223 1224 if (page_shortage <= 0) 1225 return; 1226 1227 s0 = splvm(); 1228 vm_page_lock_queues(); 1229 pcount = cnt.v_active_count; 1230 fullintervalcount += vm_pageout_stats_interval; 1231 if (fullintervalcount < vm_pageout_full_stats_interval) { 1232 tpcount = (vm_pageout_stats_max * cnt.v_active_count) / cnt.v_page_count; 1233 if (pcount > tpcount) 1234 pcount = tpcount; 1235 } else { 1236 fullintervalcount = 0; 1237 } 1238 1239 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1240 while ((m != NULL) && (pcount-- > 0)) { 1241 int actcount; 1242 1243 if (m->queue != PQ_ACTIVE) { 1244 break; 1245 } 1246 1247 next = TAILQ_NEXT(m, pageq); 1248 /* 1249 * Don't deactivate pages that are busy. 1250 */ 1251 if ((m->busy != 0) || 1252 (m->flags & PG_BUSY) || 1253 (m->hold_count != 0)) { 1254 vm_pageq_requeue(m); 1255 m = next; 1256 continue; 1257 } 1258 1259 actcount = 0; 1260 if (m->flags & PG_REFERENCED) { 1261 vm_page_flag_clear(m, PG_REFERENCED); 1262 actcount += 1; 1263 } 1264 1265 actcount += pmap_ts_referenced(m); 1266 if (actcount) { 1267 m->act_count += ACT_ADVANCE + actcount; 1268 if (m->act_count > ACT_MAX) 1269 m->act_count = ACT_MAX; 1270 vm_pageq_requeue(m); 1271 } else { 1272 if (m->act_count == 0) { 1273 /* 1274 * We turn off page access, so that we have 1275 * more accurate RSS stats. We don't do this 1276 * in the normal page deactivation when the 1277 * system is loaded VM wise, because the 1278 * cost of the large number of page protect 1279 * operations would be higher than the value 1280 * of doing the operation. 1281 */ 1282 vm_page_protect(m, VM_PROT_NONE); 1283 vm_page_deactivate(m); 1284 } else { 1285 m->act_count -= min(m->act_count, ACT_DECLINE); 1286 vm_pageq_requeue(m); 1287 } 1288 } 1289 1290 m = next; 1291 } 1292 vm_page_unlock_queues(); 1293 splx(s0); 1294 } 1295 1296 static int 1297 vm_pageout_free_page_calc(count) 1298 vm_size_t count; 1299 { 1300 if (count < cnt.v_page_count) 1301 return 0; 1302 /* 1303 * free_reserved needs to include enough for the largest swap pager 1304 * structures plus enough for any pv_entry structs when paging. 1305 */ 1306 if (cnt.v_page_count > 1024) 1307 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200; 1308 else 1309 cnt.v_free_min = 4; 1310 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1311 cnt.v_interrupt_free_min; 1312 cnt.v_free_reserved = vm_pageout_page_count + 1313 cnt.v_pageout_free_min + (count / 768) + PQ_L2_SIZE; 1314 cnt.v_free_severe = cnt.v_free_min / 2; 1315 cnt.v_free_min += cnt.v_free_reserved; 1316 cnt.v_free_severe += cnt.v_free_reserved; 1317 return 1; 1318 } 1319 1320 /* 1321 * vm_pageout is the high level pageout daemon. 1322 */ 1323 static void 1324 vm_pageout() 1325 { 1326 int pass; 1327 1328 mtx_lock(&Giant); 1329 1330 /* 1331 * Initialize some paging parameters. 1332 */ 1333 cnt.v_interrupt_free_min = 2; 1334 if (cnt.v_page_count < 2000) 1335 vm_pageout_page_count = 8; 1336 1337 vm_pageout_free_page_calc(cnt.v_page_count); 1338 /* 1339 * v_free_target and v_cache_min control pageout hysteresis. Note 1340 * that these are more a measure of the VM cache queue hysteresis 1341 * then the VM free queue. Specifically, v_free_target is the 1342 * high water mark (free+cache pages). 1343 * 1344 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the 1345 * low water mark, while v_free_min is the stop. v_cache_min must 1346 * be big enough to handle memory needs while the pageout daemon 1347 * is signalled and run to free more pages. 1348 */ 1349 if (cnt.v_free_count > 6144) 1350 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved; 1351 else 1352 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved; 1353 1354 if (cnt.v_free_count > 2048) { 1355 cnt.v_cache_min = cnt.v_free_target; 1356 cnt.v_cache_max = 2 * cnt.v_cache_min; 1357 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2; 1358 } else { 1359 cnt.v_cache_min = 0; 1360 cnt.v_cache_max = 0; 1361 cnt.v_inactive_target = cnt.v_free_count / 4; 1362 } 1363 if (cnt.v_inactive_target > cnt.v_free_count / 3) 1364 cnt.v_inactive_target = cnt.v_free_count / 3; 1365 1366 /* XXX does not really belong here */ 1367 if (vm_page_max_wired == 0) 1368 vm_page_max_wired = cnt.v_free_count / 3; 1369 1370 if (vm_pageout_stats_max == 0) 1371 vm_pageout_stats_max = cnt.v_free_target; 1372 1373 /* 1374 * Set interval in seconds for stats scan. 1375 */ 1376 if (vm_pageout_stats_interval == 0) 1377 vm_pageout_stats_interval = 5; 1378 if (vm_pageout_full_stats_interval == 0) 1379 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4; 1380 1381 /* 1382 * Set maximum free per pass 1383 */ 1384 if (vm_pageout_stats_free_max == 0) 1385 vm_pageout_stats_free_max = 5; 1386 1387 swap_pager_swap_init(); 1388 pass = 0; 1389 /* 1390 * The pageout daemon is never done, so loop forever. 1391 */ 1392 while (TRUE) { 1393 int error; 1394 int s = splvm(); 1395 1396 /* 1397 * If we have enough free memory, wakeup waiters. Do 1398 * not clear vm_pages_needed until we reach our target, 1399 * otherwise we may be woken up over and over again and 1400 * waste a lot of cpu. 1401 */ 1402 if (vm_pages_needed && !vm_page_count_min()) { 1403 if (vm_paging_needed() <= 0) 1404 vm_pages_needed = 0; 1405 wakeup(&cnt.v_free_count); 1406 } 1407 if (vm_pages_needed) { 1408 /* 1409 * Still not done, take a second pass without waiting 1410 * (unlimited dirty cleaning), otherwise sleep a bit 1411 * and try again. 1412 */ 1413 ++pass; 1414 if (pass > 1) 1415 tsleep(&vm_pages_needed, PVM, 1416 "psleep", hz/2); 1417 } else { 1418 /* 1419 * Good enough, sleep & handle stats. Prime the pass 1420 * for the next run. 1421 */ 1422 if (pass > 1) 1423 pass = 1; 1424 else 1425 pass = 0; 1426 error = tsleep(&vm_pages_needed, PVM, 1427 "psleep", vm_pageout_stats_interval * hz); 1428 if (error && !vm_pages_needed) { 1429 splx(s); 1430 pass = 0; 1431 vm_pageout_page_stats(); 1432 continue; 1433 } 1434 } 1435 1436 if (vm_pages_needed) 1437 cnt.v_pdwakeups++; 1438 splx(s); 1439 vm_pageout_scan(pass); 1440 vm_pageout_deficit = 0; 1441 } 1442 } 1443 1444 void 1445 pagedaemon_wakeup() 1446 { 1447 if (!vm_pages_needed && curthread->td_proc != pageproc) { 1448 vm_pages_needed++; 1449 wakeup(&vm_pages_needed); 1450 } 1451 } 1452 1453 #if !defined(NO_SWAPPING) 1454 static void 1455 vm_req_vmdaemon() 1456 { 1457 static int lastrun = 0; 1458 1459 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1460 wakeup(&vm_daemon_needed); 1461 lastrun = ticks; 1462 } 1463 } 1464 1465 static void 1466 vm_daemon() 1467 { 1468 struct proc *p; 1469 int breakout; 1470 struct thread *td; 1471 1472 mtx_lock(&Giant); 1473 while (TRUE) { 1474 tsleep(&vm_daemon_needed, PPAUSE, "psleep", 0); 1475 if (vm_pageout_req_swapout) { 1476 swapout_procs(vm_pageout_req_swapout); 1477 vm_pageout_req_swapout = 0; 1478 } 1479 /* 1480 * scan the processes for exceeding their rlimits or if 1481 * process is swapped out -- deactivate pages 1482 */ 1483 sx_slock(&allproc_lock); 1484 LIST_FOREACH(p, &allproc, p_list) { 1485 vm_pindex_t limit, size; 1486 1487 /* 1488 * if this is a system process or if we have already 1489 * looked at this process, skip it. 1490 */ 1491 if (p->p_flag & (P_SYSTEM | P_WEXIT)) { 1492 continue; 1493 } 1494 /* 1495 * if the process is in a non-running type state, 1496 * don't touch it. 1497 */ 1498 mtx_lock_spin(&sched_lock); 1499 breakout = 0; 1500 FOREACH_THREAD_IN_PROC(p, td) { 1501 if (td->td_state != TDS_RUNQ && 1502 td->td_state != TDS_RUNNING && 1503 td->td_state != TDS_SLP) { 1504 breakout = 1; 1505 break; 1506 } 1507 } 1508 if (breakout) { 1509 mtx_unlock_spin(&sched_lock); 1510 continue; 1511 } 1512 /* 1513 * get a limit 1514 */ 1515 limit = OFF_TO_IDX( 1516 qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur, 1517 p->p_rlimit[RLIMIT_RSS].rlim_max)); 1518 1519 /* 1520 * let processes that are swapped out really be 1521 * swapped out set the limit to nothing (will force a 1522 * swap-out.) 1523 */ 1524 if ((p->p_sflag & PS_INMEM) == 0) 1525 limit = 0; /* XXX */ 1526 mtx_unlock_spin(&sched_lock); 1527 1528 size = vmspace_resident_count(p->p_vmspace); 1529 if (limit >= 0 && size >= limit) { 1530 vm_pageout_map_deactivate_pages( 1531 &p->p_vmspace->vm_map, limit); 1532 } 1533 } 1534 sx_sunlock(&allproc_lock); 1535 } 1536 } 1537 #endif /* !defined(NO_SWAPPING) */ 1538