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 69 /* 70 * The proverbial page-out daemon. 71 */ 72 73 #include <sys/cdefs.h> 74 __FBSDID("$FreeBSD$"); 75 76 #include "opt_vm.h" 77 #include <sys/param.h> 78 #include <sys/systm.h> 79 #include <sys/kernel.h> 80 #include <sys/eventhandler.h> 81 #include <sys/lock.h> 82 #include <sys/mutex.h> 83 #include <sys/proc.h> 84 #include <sys/kthread.h> 85 #include <sys/ktr.h> 86 #include <sys/resourcevar.h> 87 #include <sys/sched.h> 88 #include <sys/signalvar.h> 89 #include <sys/vnode.h> 90 #include <sys/vmmeter.h> 91 #include <sys/sx.h> 92 #include <sys/sysctl.h> 93 94 #include <vm/vm.h> 95 #include <vm/vm_param.h> 96 #include <vm/vm_object.h> 97 #include <vm/vm_page.h> 98 #include <vm/vm_map.h> 99 #include <vm/vm_pageout.h> 100 #include <vm/vm_pager.h> 101 #include <vm/swap_pager.h> 102 #include <vm/vm_extern.h> 103 #include <vm/uma.h> 104 105 #include <machine/mutex.h> 106 107 /* 108 * System initialization 109 */ 110 111 /* the kernel process "vm_pageout"*/ 112 static void vm_pageout(void); 113 static int vm_pageout_clean(vm_page_t); 114 static void vm_pageout_pmap_collect(void); 115 static void vm_pageout_scan(int pass); 116 117 struct proc *pageproc; 118 119 static struct kproc_desc page_kp = { 120 "pagedaemon", 121 vm_pageout, 122 &pageproc 123 }; 124 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp) 125 126 #if !defined(NO_SWAPPING) 127 /* the kernel process "vm_daemon"*/ 128 static void vm_daemon(void); 129 static struct proc *vmproc; 130 131 static struct kproc_desc vm_kp = { 132 "vmdaemon", 133 vm_daemon, 134 &vmproc 135 }; 136 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp) 137 #endif 138 139 140 int vm_pages_needed; /* Event on which pageout daemon sleeps */ 141 int vm_pageout_deficit; /* Estimated number of pages deficit */ 142 int vm_pageout_pages_needed; /* flag saying that the pageout daemon needs pages */ 143 144 #if !defined(NO_SWAPPING) 145 static int vm_pageout_req_swapout; /* XXX */ 146 static int vm_daemon_needed; 147 #endif 148 static int vm_max_launder = 32; 149 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0; 150 static int vm_pageout_full_stats_interval = 0; 151 static int vm_pageout_stats_free_max=0, vm_pageout_algorithm=0; 152 static int defer_swap_pageouts=0; 153 static int disable_swap_pageouts=0; 154 155 #if defined(NO_SWAPPING) 156 static int vm_swap_enabled=0; 157 static int vm_swap_idle_enabled=0; 158 #else 159 static int vm_swap_enabled=1; 160 static int vm_swap_idle_enabled=0; 161 #endif 162 163 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm, 164 CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt"); 165 166 SYSCTL_INT(_vm, OID_AUTO, max_launder, 167 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout"); 168 169 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max, 170 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length"); 171 172 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval, 173 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan"); 174 175 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval, 176 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan"); 177 178 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_free_max, 179 CTLFLAG_RW, &vm_pageout_stats_free_max, 0, "Not implemented"); 180 181 #if defined(NO_SWAPPING) 182 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 183 CTLFLAG_RD, &vm_swap_enabled, 0, ""); 184 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 185 CTLFLAG_RD, &vm_swap_idle_enabled, 0, ""); 186 #else 187 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 188 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout"); 189 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 190 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 191 #endif 192 193 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts, 194 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem"); 195 196 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts, 197 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages"); 198 199 static int pageout_lock_miss; 200 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss, 201 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout"); 202 203 #define VM_PAGEOUT_PAGE_COUNT 16 204 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT; 205 206 int vm_page_max_wired; /* XXX max # of wired pages system-wide */ 207 208 #if !defined(NO_SWAPPING) 209 static void vm_pageout_map_deactivate_pages(vm_map_t, long); 210 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long); 211 static void vm_req_vmdaemon(void); 212 #endif 213 static void vm_pageout_page_stats(void); 214 215 /* 216 * vm_pageout_clean: 217 * 218 * Clean the page and remove it from the laundry. 219 * 220 * We set the busy bit to cause potential page faults on this page to 221 * block. Note the careful timing, however, the busy bit isn't set till 222 * late and we cannot do anything that will mess with the page. 223 */ 224 static int 225 vm_pageout_clean(m) 226 vm_page_t m; 227 { 228 vm_object_t object; 229 vm_page_t mc[2*vm_pageout_page_count]; 230 int pageout_count; 231 int ib, is, page_base; 232 vm_pindex_t pindex = m->pindex; 233 234 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 235 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 236 237 /* 238 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP 239 * with the new swapper, but we could have serious problems paging 240 * out other object types if there is insufficient memory. 241 * 242 * Unfortunately, checking free memory here is far too late, so the 243 * check has been moved up a procedural level. 244 */ 245 246 /* 247 * Don't mess with the page if it's busy, held, or special 248 */ 249 if ((m->hold_count != 0) || 250 ((m->busy != 0) || (m->flags & (PG_BUSY|PG_UNMANAGED)))) { 251 return 0; 252 } 253 254 mc[vm_pageout_page_count] = m; 255 pageout_count = 1; 256 page_base = vm_pageout_page_count; 257 ib = 1; 258 is = 1; 259 260 /* 261 * Scan object for clusterable pages. 262 * 263 * We can cluster ONLY if: ->> the page is NOT 264 * clean, wired, busy, held, or mapped into a 265 * buffer, and one of the following: 266 * 1) The page is inactive, or a seldom used 267 * active page. 268 * -or- 269 * 2) we force the issue. 270 * 271 * During heavy mmap/modification loads the pageout 272 * daemon can really fragment the underlying file 273 * due to flushing pages out of order and not trying 274 * align the clusters (which leave sporatic out-of-order 275 * holes). To solve this problem we do the reverse scan 276 * first and attempt to align our cluster, then do a 277 * forward scan if room remains. 278 */ 279 object = m->object; 280 more: 281 while (ib && pageout_count < vm_pageout_page_count) { 282 vm_page_t p; 283 284 if (ib > pindex) { 285 ib = 0; 286 break; 287 } 288 289 if ((p = vm_page_lookup(object, pindex - ib)) == NULL) { 290 ib = 0; 291 break; 292 } 293 if (((p->queue - p->pc) == PQ_CACHE) || 294 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) { 295 ib = 0; 296 break; 297 } 298 vm_page_test_dirty(p); 299 if ((p->dirty & p->valid) == 0 || 300 p->queue != PQ_INACTIVE || 301 p->wire_count != 0 || /* may be held by buf cache */ 302 p->hold_count != 0) { /* may be undergoing I/O */ 303 ib = 0; 304 break; 305 } 306 mc[--page_base] = p; 307 ++pageout_count; 308 ++ib; 309 /* 310 * alignment boundry, stop here and switch directions. Do 311 * not clear ib. 312 */ 313 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0) 314 break; 315 } 316 317 while (pageout_count < vm_pageout_page_count && 318 pindex + is < object->size) { 319 vm_page_t p; 320 321 if ((p = vm_page_lookup(object, pindex + is)) == NULL) 322 break; 323 if (((p->queue - p->pc) == PQ_CACHE) || 324 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) { 325 break; 326 } 327 vm_page_test_dirty(p); 328 if ((p->dirty & p->valid) == 0 || 329 p->queue != PQ_INACTIVE || 330 p->wire_count != 0 || /* may be held by buf cache */ 331 p->hold_count != 0) { /* may be undergoing I/O */ 332 break; 333 } 334 mc[page_base + pageout_count] = p; 335 ++pageout_count; 336 ++is; 337 } 338 339 /* 340 * If we exhausted our forward scan, continue with the reverse scan 341 * when possible, even past a page boundry. This catches boundry 342 * conditions. 343 */ 344 if (ib && pageout_count < vm_pageout_page_count) 345 goto more; 346 347 /* 348 * we allow reads during pageouts... 349 */ 350 return (vm_pageout_flush(&mc[page_base], pageout_count, 0)); 351 } 352 353 /* 354 * vm_pageout_flush() - launder the given pages 355 * 356 * The given pages are laundered. Note that we setup for the start of 357 * I/O ( i.e. busy the page ), mark it read-only, and bump the object 358 * reference count all in here rather then in the parent. If we want 359 * the parent to do more sophisticated things we may have to change 360 * the ordering. 361 */ 362 int 363 vm_pageout_flush(vm_page_t *mc, int count, int flags) 364 { 365 vm_object_t object = mc[0]->object; 366 int pageout_status[count]; 367 int numpagedout = 0; 368 int i; 369 370 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 371 VM_OBJECT_LOCK_ASSERT(object, 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, 384 ("vm_pageout_flush: partially invalid page %p index %d/%d", 385 mc[i], i, count)); 386 vm_page_io_start(mc[i]); 387 pmap_page_protect(mc[i], VM_PROT_READ); 388 } 389 vm_page_unlock_queues(); 390 vm_object_pip_add(object, count); 391 392 vm_pager_put_pages(object, mc, count, 393 (flags | ((object == kernel_object) ? VM_PAGER_PUT_SYNC : 0)), 394 pageout_status); 395 396 vm_page_lock_queues(); 397 for (i = 0; i < count; i++) { 398 vm_page_t mt = mc[i]; 399 400 switch (pageout_status[i]) { 401 case VM_PAGER_OK: 402 case VM_PAGER_PEND: 403 numpagedout++; 404 break; 405 case VM_PAGER_BAD: 406 /* 407 * Page outside of range of object. Right now we 408 * essentially lose the changes by pretending it 409 * worked. 410 */ 411 pmap_clear_modify(mt); 412 vm_page_undirty(mt); 413 break; 414 case VM_PAGER_ERROR: 415 case VM_PAGER_FAIL: 416 /* 417 * If page couldn't be paged out, then reactivate the 418 * page so it doesn't clog the inactive list. (We 419 * will try paging out it again later). 420 */ 421 vm_page_activate(mt); 422 break; 423 case VM_PAGER_AGAIN: 424 break; 425 } 426 427 /* 428 * If the operation is still going, leave the page busy to 429 * block all other accesses. Also, leave the paging in 430 * progress indicator set so that we don't attempt an object 431 * collapse. 432 */ 433 if (pageout_status[i] != VM_PAGER_PEND) { 434 vm_object_pip_wakeup(object); 435 vm_page_io_finish(mt); 436 if (!vm_page_count_severe() || !vm_page_try_to_cache(mt)) 437 pmap_page_protect(mt, VM_PROT_READ); 438 } 439 } 440 return numpagedout; 441 } 442 443 #if !defined(NO_SWAPPING) 444 /* 445 * vm_pageout_object_deactivate_pages 446 * 447 * deactivate enough pages to satisfy the inactive target 448 * requirements or if vm_page_proc_limit is set, then 449 * deactivate all of the pages in the object and its 450 * backing_objects. 451 * 452 * The object and map must be locked. 453 */ 454 static void 455 vm_pageout_object_deactivate_pages(pmap, first_object, desired) 456 pmap_t pmap; 457 vm_object_t first_object; 458 long desired; 459 { 460 vm_object_t backing_object, object; 461 vm_page_t p, next; 462 int actcount, rcount, remove_mode; 463 464 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED); 465 if (first_object->type == OBJT_DEVICE || first_object->type == OBJT_PHYS) 466 return; 467 for (object = first_object;; object = backing_object) { 468 if (pmap_resident_count(pmap) <= desired) 469 goto unlock_return; 470 if (object->paging_in_progress) 471 goto unlock_return; 472 473 remove_mode = 0; 474 if (object->shadow_count > 1) 475 remove_mode = 1; 476 /* 477 * scan the objects entire memory queue 478 */ 479 rcount = object->resident_page_count; 480 p = TAILQ_FIRST(&object->memq); 481 vm_page_lock_queues(); 482 while (p && (rcount-- > 0)) { 483 if (pmap_resident_count(pmap) <= desired) { 484 vm_page_unlock_queues(); 485 goto unlock_return; 486 } 487 next = TAILQ_NEXT(p, listq); 488 cnt.v_pdpages++; 489 if (p->wire_count != 0 || 490 p->hold_count != 0 || 491 p->busy != 0 || 492 (p->flags & (PG_BUSY|PG_UNMANAGED)) || 493 !pmap_page_exists_quick(pmap, p)) { 494 p = next; 495 continue; 496 } 497 actcount = pmap_ts_referenced(p); 498 if (actcount) { 499 vm_page_flag_set(p, PG_REFERENCED); 500 } else if (p->flags & PG_REFERENCED) { 501 actcount = 1; 502 } 503 if ((p->queue != PQ_ACTIVE) && 504 (p->flags & PG_REFERENCED)) { 505 vm_page_activate(p); 506 p->act_count += actcount; 507 vm_page_flag_clear(p, PG_REFERENCED); 508 } else if (p->queue == PQ_ACTIVE) { 509 if ((p->flags & PG_REFERENCED) == 0) { 510 p->act_count -= min(p->act_count, ACT_DECLINE); 511 if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) { 512 pmap_remove_all(p); 513 vm_page_deactivate(p); 514 } else { 515 vm_pageq_requeue(p); 516 } 517 } else { 518 vm_page_activate(p); 519 vm_page_flag_clear(p, PG_REFERENCED); 520 if (p->act_count < (ACT_MAX - ACT_ADVANCE)) 521 p->act_count += ACT_ADVANCE; 522 vm_pageq_requeue(p); 523 } 524 } else if (p->queue == PQ_INACTIVE) { 525 pmap_remove_all(p); 526 } 527 p = next; 528 } 529 vm_page_unlock_queues(); 530 if ((backing_object = object->backing_object) == NULL) 531 goto unlock_return; 532 VM_OBJECT_LOCK(backing_object); 533 if (object != first_object) 534 VM_OBJECT_UNLOCK(object); 535 } 536 unlock_return: 537 if (object != first_object) 538 VM_OBJECT_UNLOCK(object); 539 } 540 541 /* 542 * deactivate some number of pages in a map, try to do it fairly, but 543 * that is really hard to do. 544 */ 545 static void 546 vm_pageout_map_deactivate_pages(map, desired) 547 vm_map_t map; 548 long desired; 549 { 550 vm_map_entry_t tmpe; 551 vm_object_t obj, bigobj; 552 int nothingwired; 553 554 if (!vm_map_trylock(map)) 555 return; 556 557 bigobj = NULL; 558 nothingwired = TRUE; 559 560 /* 561 * first, search out the biggest object, and try to free pages from 562 * that. 563 */ 564 tmpe = map->header.next; 565 while (tmpe != &map->header) { 566 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 567 obj = tmpe->object.vm_object; 568 if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) { 569 if (obj->shadow_count <= 1 && 570 (bigobj == NULL || 571 bigobj->resident_page_count < obj->resident_page_count)) { 572 if (bigobj != NULL) 573 VM_OBJECT_UNLOCK(bigobj); 574 bigobj = obj; 575 } else 576 VM_OBJECT_UNLOCK(obj); 577 } 578 } 579 if (tmpe->wired_count > 0) 580 nothingwired = FALSE; 581 tmpe = tmpe->next; 582 } 583 584 if (bigobj != NULL) { 585 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired); 586 VM_OBJECT_UNLOCK(bigobj); 587 } 588 /* 589 * Next, hunt around for other pages to deactivate. We actually 590 * do this search sort of wrong -- .text first is not the best idea. 591 */ 592 tmpe = map->header.next; 593 while (tmpe != &map->header) { 594 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 595 break; 596 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 597 obj = tmpe->object.vm_object; 598 if (obj != NULL) { 599 VM_OBJECT_LOCK(obj); 600 vm_pageout_object_deactivate_pages(map->pmap, obj, desired); 601 VM_OBJECT_UNLOCK(obj); 602 } 603 } 604 tmpe = tmpe->next; 605 } 606 607 /* 608 * Remove all mappings if a process is swapped out, this will free page 609 * table pages. 610 */ 611 if (desired == 0 && nothingwired) { 612 GIANT_REQUIRED; 613 vm_page_lock_queues(); 614 pmap_remove(vm_map_pmap(map), vm_map_min(map), 615 vm_map_max(map)); 616 vm_page_unlock_queues(); 617 } 618 vm_map_unlock(map); 619 } 620 #endif /* !defined(NO_SWAPPING) */ 621 622 /* 623 * This routine is very drastic, but can save the system 624 * in a pinch. 625 */ 626 static void 627 vm_pageout_pmap_collect(void) 628 { 629 int i; 630 vm_page_t m; 631 static int warningdone; 632 633 if (pmap_pagedaemon_waken == 0) 634 return; 635 if (warningdone < 5) { 636 printf("collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n"); 637 warningdone++; 638 } 639 vm_page_lock_queues(); 640 for (i = 0; i < vm_page_array_size; i++) { 641 m = &vm_page_array[i]; 642 if (m->wire_count || m->hold_count || m->busy || 643 (m->flags & (PG_BUSY | PG_UNMANAGED))) 644 continue; 645 pmap_remove_all(m); 646 } 647 vm_page_unlock_queues(); 648 pmap_pagedaemon_waken = 0; 649 } 650 651 /* 652 * vm_pageout_scan does the dirty work for the pageout daemon. 653 */ 654 static void 655 vm_pageout_scan(int pass) 656 { 657 vm_page_t m, next; 658 struct vm_page marker; 659 int page_shortage, maxscan, pcount; 660 int addl_page_shortage, addl_page_shortage_init; 661 struct proc *p, *bigproc; 662 vm_offset_t size, bigsize; 663 vm_object_t object; 664 int actcount; 665 int vnodes_skipped = 0; 666 int maxlaunder; 667 int s; 668 struct thread *td; 669 670 mtx_lock(&Giant); 671 /* 672 * Decrease registered cache sizes. 673 */ 674 EVENTHANDLER_INVOKE(vm_lowmem, 0); 675 /* 676 * We do this explicitly after the caches have been drained above. 677 */ 678 uma_reclaim(); 679 /* 680 * Do whatever cleanup that the pmap code can. 681 */ 682 vm_pageout_pmap_collect(); 683 684 addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit); 685 686 /* 687 * Calculate the number of pages we want to either free or move 688 * to the cache. 689 */ 690 page_shortage = vm_paging_target() + addl_page_shortage_init; 691 692 /* 693 * Initialize our marker 694 */ 695 bzero(&marker, sizeof(marker)); 696 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER; 697 marker.queue = PQ_INACTIVE; 698 marker.wire_count = 1; 699 700 /* 701 * Start scanning the inactive queue for pages we can move to the 702 * cache or free. The scan will stop when the target is reached or 703 * we have scanned the entire inactive queue. Note that m->act_count 704 * is not used to form decisions for the inactive queue, only for the 705 * active queue. 706 * 707 * maxlaunder limits the number of dirty pages we flush per scan. 708 * For most systems a smaller value (16 or 32) is more robust under 709 * extreme memory and disk pressure because any unnecessary writes 710 * to disk can result in extreme performance degredation. However, 711 * systems with excessive dirty pages (especially when MAP_NOSYNC is 712 * used) will die horribly with limited laundering. If the pageout 713 * daemon cannot clean enough pages in the first pass, we let it go 714 * all out in succeeding passes. 715 */ 716 if ((maxlaunder = vm_max_launder) <= 1) 717 maxlaunder = 1; 718 if (pass) 719 maxlaunder = 10000; 720 vm_page_lock_queues(); 721 rescan0: 722 addl_page_shortage = addl_page_shortage_init; 723 maxscan = cnt.v_inactive_count; 724 725 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl); 726 m != NULL && maxscan-- > 0 && page_shortage > 0; 727 m = next) { 728 729 cnt.v_pdpages++; 730 731 if (m->queue != PQ_INACTIVE) { 732 goto rescan0; 733 } 734 735 next = TAILQ_NEXT(m, pageq); 736 737 /* 738 * skip marker pages 739 */ 740 if (m->flags & PG_MARKER) 741 continue; 742 743 /* 744 * A held page may be undergoing I/O, so skip it. 745 */ 746 if (m->hold_count) { 747 vm_pageq_requeue(m); 748 addl_page_shortage++; 749 continue; 750 } 751 /* 752 * Don't mess with busy pages, keep in the front of the 753 * queue, most likely are being paged out. 754 */ 755 if (m->busy || (m->flags & PG_BUSY)) { 756 addl_page_shortage++; 757 continue; 758 } 759 760 /* 761 * If the object is not being used, we ignore previous 762 * references. 763 */ 764 if (m->object->ref_count == 0) { 765 vm_page_flag_clear(m, PG_REFERENCED); 766 pmap_clear_reference(m); 767 768 /* 769 * Otherwise, if the page has been referenced while in the 770 * inactive queue, we bump the "activation count" upwards, 771 * making it less likely that the page will be added back to 772 * the inactive queue prematurely again. Here we check the 773 * page tables (or emulated bits, if any), given the upper 774 * level VM system not knowing anything about existing 775 * references. 776 */ 777 } else if (((m->flags & PG_REFERENCED) == 0) && 778 (actcount = pmap_ts_referenced(m))) { 779 vm_page_activate(m); 780 m->act_count += (actcount + ACT_ADVANCE); 781 continue; 782 } 783 784 /* 785 * If the upper level VM system knows about any page 786 * references, we activate the page. We also set the 787 * "activation count" higher than normal so that we will less 788 * likely place pages back onto the inactive queue again. 789 */ 790 if ((m->flags & PG_REFERENCED) != 0) { 791 vm_page_flag_clear(m, PG_REFERENCED); 792 actcount = pmap_ts_referenced(m); 793 vm_page_activate(m); 794 m->act_count += (actcount + ACT_ADVANCE + 1); 795 continue; 796 } 797 798 /* 799 * If the upper level VM system doesn't know anything about 800 * the page being dirty, we have to check for it again. As 801 * far as the VM code knows, any partially dirty pages are 802 * fully dirty. 803 */ 804 if (m->dirty == 0) { 805 vm_page_test_dirty(m); 806 } else { 807 vm_page_dirty(m); 808 } 809 object = m->object; 810 if (!VM_OBJECT_TRYLOCK(object)) 811 continue; 812 if (m->valid == 0) { 813 /* 814 * Invalid pages can be easily freed 815 */ 816 vm_page_busy(m); 817 pmap_remove_all(m); 818 vm_page_free(m); 819 cnt.v_dfree++; 820 --page_shortage; 821 } else if (m->dirty == 0) { 822 /* 823 * Clean pages can be placed onto the cache queue. 824 * This effectively frees them. 825 */ 826 vm_page_cache(m); 827 --page_shortage; 828 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) { 829 /* 830 * Dirty pages need to be paged out, but flushing 831 * a page is extremely expensive verses freeing 832 * a clean page. Rather then artificially limiting 833 * the number of pages we can flush, we instead give 834 * dirty pages extra priority on the inactive queue 835 * by forcing them to be cycled through the queue 836 * twice before being flushed, after which the 837 * (now clean) page will cycle through once more 838 * before being freed. This significantly extends 839 * the thrash point for a heavily loaded machine. 840 */ 841 vm_page_flag_set(m, PG_WINATCFLS); 842 vm_pageq_requeue(m); 843 } else if (maxlaunder > 0) { 844 /* 845 * We always want to try to flush some dirty pages if 846 * we encounter them, to keep the system stable. 847 * Normally this number is small, but under extreme 848 * pressure where there are insufficient clean pages 849 * on the inactive queue, we may have to go all out. 850 */ 851 int swap_pageouts_ok; 852 struct vnode *vp = NULL; 853 struct mount *mp; 854 855 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 856 swap_pageouts_ok = 1; 857 } else { 858 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 859 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 860 vm_page_count_min()); 861 862 } 863 864 /* 865 * We don't bother paging objects that are "dead". 866 * Those objects are in a "rundown" state. 867 */ 868 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 869 VM_OBJECT_UNLOCK(object); 870 vm_pageq_requeue(m); 871 continue; 872 } 873 874 /* 875 * The object is already known NOT to be dead. It 876 * is possible for the vget() to block the whole 877 * pageout daemon, but the new low-memory handling 878 * code should prevent it. 879 * 880 * The previous code skipped locked vnodes and, worse, 881 * reordered pages in the queue. This results in 882 * completely non-deterministic operation and, on a 883 * busy system, can lead to extremely non-optimal 884 * pageouts. For example, it can cause clean pages 885 * to be freed and dirty pages to be moved to the end 886 * of the queue. Since dirty pages are also moved to 887 * the end of the queue once-cleaned, this gives 888 * way too large a weighting to defering the freeing 889 * of dirty pages. 890 * 891 * We can't wait forever for the vnode lock, we might 892 * deadlock due to a vn_read() getting stuck in 893 * vm_wait while holding this vnode. We skip the 894 * vnode if we can't get it in a reasonable amount 895 * of time. 896 */ 897 if (object->type == OBJT_VNODE) { 898 vp = object->handle; 899 mp = NULL; 900 if (vp->v_type == VREG) 901 vn_start_write(vp, &mp, V_NOWAIT); 902 vm_page_unlock_queues(); 903 VI_LOCK(vp); 904 VM_OBJECT_UNLOCK(object); 905 if (vget(vp, LK_EXCLUSIVE | LK_INTERLOCK | 906 LK_TIMELOCK, curthread)) { 907 VM_OBJECT_LOCK(object); 908 vm_page_lock_queues(); 909 ++pageout_lock_miss; 910 vn_finished_write(mp); 911 if (object->flags & OBJ_MIGHTBEDIRTY) 912 vnodes_skipped++; 913 VM_OBJECT_UNLOCK(object); 914 continue; 915 } 916 VM_OBJECT_LOCK(object); 917 vm_page_lock_queues(); 918 /* 919 * The page might have been moved to another 920 * queue during potential blocking in vget() 921 * above. The page might have been freed and 922 * reused for another vnode. The object might 923 * have been reused for another vnode. 924 */ 925 if (m->queue != PQ_INACTIVE || 926 m->object != object || 927 object->handle != vp) { 928 if (object->flags & OBJ_MIGHTBEDIRTY) 929 vnodes_skipped++; 930 goto unlock_and_continue; 931 } 932 933 /* 934 * The page may have been busied during the 935 * blocking in vput(); We don't move the 936 * page back onto the end of the queue so that 937 * statistics are more correct if we don't. 938 */ 939 if (m->busy || (m->flags & PG_BUSY)) { 940 goto unlock_and_continue; 941 } 942 943 /* 944 * If the page has become held it might 945 * be undergoing I/O, so skip it 946 */ 947 if (m->hold_count) { 948 vm_pageq_requeue(m); 949 if (object->flags & OBJ_MIGHTBEDIRTY) 950 vnodes_skipped++; 951 goto unlock_and_continue; 952 } 953 } 954 955 /* 956 * If a page is dirty, then it is either being washed 957 * (but not yet cleaned) or it is still in the 958 * laundry. If it is still in the laundry, then we 959 * start the cleaning operation. 960 * 961 * This operation may cluster, invalidating the 'next' 962 * pointer. To prevent an inordinate number of 963 * restarts we use our marker to remember our place. 964 * 965 * decrement page_shortage on success to account for 966 * the (future) cleaned page. Otherwise we could wind 967 * up laundering or cleaning too many pages. 968 */ 969 s = splvm(); 970 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, m, &marker, pageq); 971 splx(s); 972 if (vm_pageout_clean(m) != 0) { 973 --page_shortage; 974 --maxlaunder; 975 } 976 s = splvm(); 977 next = TAILQ_NEXT(&marker, pageq); 978 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, &marker, pageq); 979 splx(s); 980 unlock_and_continue: 981 VM_OBJECT_UNLOCK(object); 982 if (vp) { 983 vm_page_unlock_queues(); 984 vput(vp); 985 vn_finished_write(mp); 986 vm_page_lock_queues(); 987 } 988 continue; 989 } 990 VM_OBJECT_UNLOCK(object); 991 } 992 993 /* 994 * Compute the number of pages we want to try to move from the 995 * active queue to the inactive queue. 996 */ 997 page_shortage = vm_paging_target() + 998 cnt.v_inactive_target - cnt.v_inactive_count; 999 page_shortage += addl_page_shortage; 1000 1001 /* 1002 * Scan the active queue for things we can deactivate. We nominally 1003 * track the per-page activity counter and use it to locate 1004 * deactivation candidates. 1005 */ 1006 pcount = cnt.v_active_count; 1007 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1008 1009 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) { 1010 1011 KASSERT(m->queue == PQ_ACTIVE, 1012 ("vm_pageout_scan: page %p isn't active", m)); 1013 1014 next = TAILQ_NEXT(m, pageq); 1015 /* 1016 * Don't deactivate pages that are busy. 1017 */ 1018 if ((m->busy != 0) || 1019 (m->flags & PG_BUSY) || 1020 (m->hold_count != 0)) { 1021 vm_pageq_requeue(m); 1022 m = next; 1023 continue; 1024 } 1025 1026 /* 1027 * The count for pagedaemon pages is done after checking the 1028 * page for eligibility... 1029 */ 1030 cnt.v_pdpages++; 1031 1032 /* 1033 * Check to see "how much" the page has been used. 1034 */ 1035 actcount = 0; 1036 if (m->object->ref_count != 0) { 1037 if (m->flags & PG_REFERENCED) { 1038 actcount += 1; 1039 } 1040 actcount += pmap_ts_referenced(m); 1041 if (actcount) { 1042 m->act_count += ACT_ADVANCE + actcount; 1043 if (m->act_count > ACT_MAX) 1044 m->act_count = ACT_MAX; 1045 } 1046 } 1047 1048 /* 1049 * Since we have "tested" this bit, we need to clear it now. 1050 */ 1051 vm_page_flag_clear(m, PG_REFERENCED); 1052 1053 /* 1054 * Only if an object is currently being used, do we use the 1055 * page activation count stats. 1056 */ 1057 if (actcount && (m->object->ref_count != 0)) { 1058 vm_pageq_requeue(m); 1059 } else { 1060 m->act_count -= min(m->act_count, ACT_DECLINE); 1061 if (vm_pageout_algorithm || 1062 m->object->ref_count == 0 || 1063 m->act_count == 0) { 1064 page_shortage--; 1065 if (m->object->ref_count == 0) { 1066 pmap_remove_all(m); 1067 if (m->dirty == 0) 1068 vm_page_cache(m); 1069 else 1070 vm_page_deactivate(m); 1071 } else { 1072 vm_page_deactivate(m); 1073 } 1074 } else { 1075 vm_pageq_requeue(m); 1076 } 1077 } 1078 m = next; 1079 } 1080 s = splvm(); 1081 1082 /* 1083 * We try to maintain some *really* free pages, this allows interrupt 1084 * code to be guaranteed space. Since both cache and free queues 1085 * are considered basically 'free', moving pages from cache to free 1086 * does not effect other calculations. 1087 */ 1088 while (cnt.v_free_count < cnt.v_free_reserved) { 1089 static int cache_rover = 0; 1090 1091 if ((m = vm_page_select_cache(cache_rover)) == NULL) 1092 break; 1093 cache_rover = (m->pc + PQ_PRIME2) & PQ_L2_MASK; 1094 object = m->object; 1095 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1096 vm_page_busy(m); 1097 pmap_remove_all(m); 1098 vm_page_free(m); 1099 VM_OBJECT_UNLOCK(object); 1100 cnt.v_dfree++; 1101 } 1102 splx(s); 1103 vm_page_unlock_queues(); 1104 #if !defined(NO_SWAPPING) 1105 /* 1106 * Idle process swapout -- run once per second. 1107 */ 1108 if (vm_swap_idle_enabled) { 1109 static long lsec; 1110 if (time_second != lsec) { 1111 vm_pageout_req_swapout |= VM_SWAP_IDLE; 1112 vm_req_vmdaemon(); 1113 lsec = time_second; 1114 } 1115 } 1116 #endif 1117 1118 /* 1119 * If we didn't get enough free pages, and we have skipped a vnode 1120 * in a writeable object, wakeup the sync daemon. And kick swapout 1121 * if we did not get enough free pages. 1122 */ 1123 if (vm_paging_target() > 0) { 1124 if (vnodes_skipped && vm_page_count_min()) 1125 (void) speedup_syncer(); 1126 #if !defined(NO_SWAPPING) 1127 if (vm_swap_enabled && vm_page_count_target()) { 1128 vm_req_vmdaemon(); 1129 vm_pageout_req_swapout |= VM_SWAP_NORMAL; 1130 } 1131 #endif 1132 } 1133 1134 /* 1135 * If we are critically low on one of RAM or swap and low on 1136 * the other, kill the largest process. However, we avoid 1137 * doing this on the first pass in order to give ourselves a 1138 * chance to flush out dirty vnode-backed pages and to allow 1139 * active pages to be moved to the inactive queue and reclaimed. 1140 * 1141 * We keep the process bigproc locked once we find it to keep anyone 1142 * from messing with it; however, there is a possibility of 1143 * deadlock if process B is bigproc and one of it's child processes 1144 * attempts to propagate a signal to B while we are waiting for A's 1145 * lock while walking this list. To avoid this, we don't block on 1146 * the process lock but just skip a process if it is already locked. 1147 */ 1148 if (pass != 0 && 1149 ((swap_pager_avail < 64 && vm_page_count_min()) || 1150 (swap_pager_full && vm_paging_target() > 0))) { 1151 bigproc = NULL; 1152 bigsize = 0; 1153 sx_slock(&allproc_lock); 1154 FOREACH_PROC_IN_SYSTEM(p) { 1155 int breakout; 1156 /* 1157 * If this process is already locked, skip it. 1158 */ 1159 if (PROC_TRYLOCK(p) == 0) 1160 continue; 1161 /* 1162 * If this is a system or protected process, skip it. 1163 */ 1164 if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) || 1165 (p->p_flag & P_PROTECTED) || 1166 ((p->p_pid < 48) && (swap_pager_avail != 0))) { 1167 PROC_UNLOCK(p); 1168 continue; 1169 } 1170 /* 1171 * if the process is in a non-running type state, 1172 * don't touch it. Check all the threads individually. 1173 */ 1174 mtx_lock_spin(&sched_lock); 1175 breakout = 0; 1176 FOREACH_THREAD_IN_PROC(p, td) { 1177 if (!TD_ON_RUNQ(td) && 1178 !TD_IS_RUNNING(td) && 1179 !TD_IS_SLEEPING(td)) { 1180 breakout = 1; 1181 break; 1182 } 1183 } 1184 if (breakout) { 1185 mtx_unlock_spin(&sched_lock); 1186 PROC_UNLOCK(p); 1187 continue; 1188 } 1189 mtx_unlock_spin(&sched_lock); 1190 /* 1191 * get the process size 1192 */ 1193 if (!vm_map_trylock_read(&p->p_vmspace->vm_map)) { 1194 PROC_UNLOCK(p); 1195 continue; 1196 } 1197 size = vmspace_swap_count(p->p_vmspace); 1198 vm_map_unlock_read(&p->p_vmspace->vm_map); 1199 size += vmspace_resident_count(p->p_vmspace); 1200 /* 1201 * if the this process is bigger than the biggest one 1202 * remember it. 1203 */ 1204 if (size > bigsize) { 1205 if (bigproc != NULL) 1206 PROC_UNLOCK(bigproc); 1207 bigproc = p; 1208 bigsize = size; 1209 } else 1210 PROC_UNLOCK(p); 1211 } 1212 sx_sunlock(&allproc_lock); 1213 if (bigproc != NULL) { 1214 struct ksegrp *kg; 1215 killproc(bigproc, "out of swap space"); 1216 mtx_lock_spin(&sched_lock); 1217 FOREACH_KSEGRP_IN_PROC(bigproc, kg) { 1218 sched_nice(kg, PRIO_MIN); /* XXXKSE ??? */ 1219 } 1220 mtx_unlock_spin(&sched_lock); 1221 PROC_UNLOCK(bigproc); 1222 wakeup(&cnt.v_free_count); 1223 } 1224 } 1225 mtx_unlock(&Giant); 1226 } 1227 1228 /* 1229 * This routine tries to maintain the pseudo LRU active queue, 1230 * so that during long periods of time where there is no paging, 1231 * that some statistic accumulation still occurs. This code 1232 * helps the situation where paging just starts to occur. 1233 */ 1234 static void 1235 vm_pageout_page_stats() 1236 { 1237 vm_page_t m,next; 1238 int pcount,tpcount; /* Number of pages to check */ 1239 static int fullintervalcount = 0; 1240 int page_shortage; 1241 int s0; 1242 1243 page_shortage = 1244 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) - 1245 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count); 1246 1247 if (page_shortage <= 0) 1248 return; 1249 1250 s0 = splvm(); 1251 vm_page_lock_queues(); 1252 pcount = cnt.v_active_count; 1253 fullintervalcount += vm_pageout_stats_interval; 1254 if (fullintervalcount < vm_pageout_full_stats_interval) { 1255 tpcount = (vm_pageout_stats_max * cnt.v_active_count) / cnt.v_page_count; 1256 if (pcount > tpcount) 1257 pcount = tpcount; 1258 } else { 1259 fullintervalcount = 0; 1260 } 1261 1262 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1263 while ((m != NULL) && (pcount-- > 0)) { 1264 int actcount; 1265 1266 KASSERT(m->queue == PQ_ACTIVE, 1267 ("vm_pageout_page_stats: page %p isn't active", m)); 1268 1269 next = TAILQ_NEXT(m, pageq); 1270 /* 1271 * Don't deactivate pages that are busy. 1272 */ 1273 if ((m->busy != 0) || 1274 (m->flags & PG_BUSY) || 1275 (m->hold_count != 0)) { 1276 vm_pageq_requeue(m); 1277 m = next; 1278 continue; 1279 } 1280 1281 actcount = 0; 1282 if (m->flags & PG_REFERENCED) { 1283 vm_page_flag_clear(m, PG_REFERENCED); 1284 actcount += 1; 1285 } 1286 1287 actcount += pmap_ts_referenced(m); 1288 if (actcount) { 1289 m->act_count += ACT_ADVANCE + actcount; 1290 if (m->act_count > ACT_MAX) 1291 m->act_count = ACT_MAX; 1292 vm_pageq_requeue(m); 1293 } else { 1294 if (m->act_count == 0) { 1295 /* 1296 * We turn off page access, so that we have 1297 * more accurate RSS stats. We don't do this 1298 * in the normal page deactivation when the 1299 * system is loaded VM wise, because the 1300 * cost of the large number of page protect 1301 * operations would be higher than the value 1302 * of doing the operation. 1303 */ 1304 pmap_remove_all(m); 1305 vm_page_deactivate(m); 1306 } else { 1307 m->act_count -= min(m->act_count, ACT_DECLINE); 1308 vm_pageq_requeue(m); 1309 } 1310 } 1311 1312 m = next; 1313 } 1314 vm_page_unlock_queues(); 1315 splx(s0); 1316 } 1317 1318 /* 1319 * vm_pageout is the high level pageout daemon. 1320 */ 1321 static void 1322 vm_pageout() 1323 { 1324 int error, pass, s; 1325 1326 /* 1327 * Initialize some paging parameters. 1328 */ 1329 cnt.v_interrupt_free_min = 2; 1330 if (cnt.v_page_count < 2000) 1331 vm_pageout_page_count = 8; 1332 1333 /* 1334 * v_free_reserved needs to include enough for the largest 1335 * swap pager structures plus enough for any pv_entry structs 1336 * when paging. 1337 */ 1338 if (cnt.v_page_count > 1024) 1339 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200; 1340 else 1341 cnt.v_free_min = 4; 1342 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1343 cnt.v_interrupt_free_min; 1344 cnt.v_free_reserved = vm_pageout_page_count + 1345 cnt.v_pageout_free_min + (cnt.v_page_count / 768) + PQ_L2_SIZE; 1346 cnt.v_free_severe = cnt.v_free_min / 2; 1347 cnt.v_free_min += cnt.v_free_reserved; 1348 cnt.v_free_severe += cnt.v_free_reserved; 1349 1350 /* 1351 * v_free_target and v_cache_min control pageout hysteresis. Note 1352 * that these are more a measure of the VM cache queue hysteresis 1353 * then the VM free queue. Specifically, v_free_target is the 1354 * high water mark (free+cache pages). 1355 * 1356 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the 1357 * low water mark, while v_free_min is the stop. v_cache_min must 1358 * be big enough to handle memory needs while the pageout daemon 1359 * is signalled and run to free more pages. 1360 */ 1361 if (cnt.v_free_count > 6144) 1362 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved; 1363 else 1364 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved; 1365 1366 if (cnt.v_free_count > 2048) { 1367 cnt.v_cache_min = cnt.v_free_target; 1368 cnt.v_cache_max = 2 * cnt.v_cache_min; 1369 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2; 1370 } else { 1371 cnt.v_cache_min = 0; 1372 cnt.v_cache_max = 0; 1373 cnt.v_inactive_target = cnt.v_free_count / 4; 1374 } 1375 if (cnt.v_inactive_target > cnt.v_free_count / 3) 1376 cnt.v_inactive_target = cnt.v_free_count / 3; 1377 1378 /* XXX does not really belong here */ 1379 if (vm_page_max_wired == 0) 1380 vm_page_max_wired = cnt.v_free_count / 3; 1381 1382 if (vm_pageout_stats_max == 0) 1383 vm_pageout_stats_max = cnt.v_free_target; 1384 1385 /* 1386 * Set interval in seconds for stats scan. 1387 */ 1388 if (vm_pageout_stats_interval == 0) 1389 vm_pageout_stats_interval = 5; 1390 if (vm_pageout_full_stats_interval == 0) 1391 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4; 1392 1393 /* 1394 * Set maximum free per pass 1395 */ 1396 if (vm_pageout_stats_free_max == 0) 1397 vm_pageout_stats_free_max = 5; 1398 1399 swap_pager_swap_init(); 1400 pass = 0; 1401 /* 1402 * The pageout daemon is never done, so loop forever. 1403 */ 1404 while (TRUE) { 1405 s = splvm(); 1406 vm_page_lock_queues(); 1407 /* 1408 * If we have enough free memory, wakeup waiters. Do 1409 * not clear vm_pages_needed until we reach our target, 1410 * otherwise we may be woken up over and over again and 1411 * waste a lot of cpu. 1412 */ 1413 if (vm_pages_needed && !vm_page_count_min()) { 1414 if (!vm_paging_needed()) 1415 vm_pages_needed = 0; 1416 wakeup(&cnt.v_free_count); 1417 } 1418 if (vm_pages_needed) { 1419 /* 1420 * Still not done, take a second pass without waiting 1421 * (unlimited dirty cleaning), otherwise sleep a bit 1422 * and try again. 1423 */ 1424 ++pass; 1425 if (pass > 1) 1426 msleep(&vm_pages_needed, &vm_page_queue_mtx, PVM, 1427 "psleep", hz/2); 1428 } else { 1429 /* 1430 * Good enough, sleep & handle stats. Prime the pass 1431 * for the next run. 1432 */ 1433 if (pass > 1) 1434 pass = 1; 1435 else 1436 pass = 0; 1437 error = msleep(&vm_pages_needed, &vm_page_queue_mtx, PVM, 1438 "psleep", vm_pageout_stats_interval * hz); 1439 if (error && !vm_pages_needed) { 1440 vm_page_unlock_queues(); 1441 splx(s); 1442 pass = 0; 1443 vm_pageout_page_stats(); 1444 continue; 1445 } 1446 } 1447 if (vm_pages_needed) 1448 cnt.v_pdwakeups++; 1449 vm_page_unlock_queues(); 1450 splx(s); 1451 vm_pageout_scan(pass); 1452 } 1453 } 1454 1455 /* 1456 * Unless the page queue lock is held by the caller, this function 1457 * should be regarded as advisory. Specifically, the caller should 1458 * not msleep() on &cnt.v_free_count following this function unless 1459 * the page queue lock is held until the msleep() is performed. 1460 */ 1461 void 1462 pagedaemon_wakeup() 1463 { 1464 1465 if (!vm_pages_needed && curthread->td_proc != pageproc) { 1466 vm_pages_needed = 1; 1467 wakeup(&vm_pages_needed); 1468 } 1469 } 1470 1471 #if !defined(NO_SWAPPING) 1472 static void 1473 vm_req_vmdaemon() 1474 { 1475 static int lastrun = 0; 1476 1477 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1478 wakeup(&vm_daemon_needed); 1479 lastrun = ticks; 1480 } 1481 } 1482 1483 static void 1484 vm_daemon() 1485 { 1486 struct proc *p; 1487 int breakout; 1488 struct thread *td; 1489 1490 mtx_lock(&Giant); 1491 while (TRUE) { 1492 tsleep(&vm_daemon_needed, PPAUSE, "psleep", 0); 1493 if (vm_pageout_req_swapout) { 1494 swapout_procs(vm_pageout_req_swapout); 1495 vm_pageout_req_swapout = 0; 1496 } 1497 /* 1498 * scan the processes for exceeding their rlimits or if 1499 * process is swapped out -- deactivate pages 1500 */ 1501 sx_slock(&allproc_lock); 1502 LIST_FOREACH(p, &allproc, p_list) { 1503 vm_pindex_t limit, size; 1504 1505 /* 1506 * if this is a system process or if we have already 1507 * looked at this process, skip it. 1508 */ 1509 PROC_LOCK(p); 1510 if (p->p_flag & (P_SYSTEM | P_WEXIT)) { 1511 PROC_UNLOCK(p); 1512 continue; 1513 } 1514 /* 1515 * if the process is in a non-running type state, 1516 * don't touch it. 1517 */ 1518 mtx_lock_spin(&sched_lock); 1519 breakout = 0; 1520 FOREACH_THREAD_IN_PROC(p, td) { 1521 if (!TD_ON_RUNQ(td) && 1522 !TD_IS_RUNNING(td) && 1523 !TD_IS_SLEEPING(td)) { 1524 breakout = 1; 1525 break; 1526 } 1527 } 1528 mtx_unlock_spin(&sched_lock); 1529 if (breakout) { 1530 PROC_UNLOCK(p); 1531 continue; 1532 } 1533 /* 1534 * get a limit 1535 */ 1536 limit = OFF_TO_IDX( 1537 qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur, 1538 p->p_rlimit[RLIMIT_RSS].rlim_max)); 1539 1540 /* 1541 * let processes that are swapped out really be 1542 * swapped out set the limit to nothing (will force a 1543 * swap-out.) 1544 */ 1545 if ((p->p_sflag & PS_INMEM) == 0) 1546 limit = 0; /* XXX */ 1547 PROC_UNLOCK(p); 1548 1549 size = vmspace_resident_count(p->p_vmspace); 1550 if (limit >= 0 && size >= limit) { 1551 vm_pageout_map_deactivate_pages( 1552 &p->p_vmspace->vm_map, limit); 1553 } 1554 } 1555 sx_sunlock(&allproc_lock); 1556 } 1557 } 1558 #endif /* !defined(NO_SWAPPING) */ 1559