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 KASSERT((mt->flags & PG_WRITEABLE) == 0, 401 ("vm_pageout_flush: page %p is not write protected", mt)); 402 switch (pageout_status[i]) { 403 case VM_PAGER_OK: 404 case VM_PAGER_PEND: 405 numpagedout++; 406 break; 407 case VM_PAGER_BAD: 408 /* 409 * Page outside of range of object. Right now we 410 * essentially lose the changes by pretending it 411 * worked. 412 */ 413 pmap_clear_modify(mt); 414 vm_page_undirty(mt); 415 break; 416 case VM_PAGER_ERROR: 417 case VM_PAGER_FAIL: 418 /* 419 * If page couldn't be paged out, then reactivate the 420 * page so it doesn't clog the inactive list. (We 421 * will try paging out it again later). 422 */ 423 vm_page_activate(mt); 424 break; 425 case VM_PAGER_AGAIN: 426 break; 427 } 428 429 /* 430 * If the operation is still going, leave the page busy to 431 * block all other accesses. Also, leave the paging in 432 * progress indicator set so that we don't attempt an object 433 * collapse. 434 */ 435 if (pageout_status[i] != VM_PAGER_PEND) { 436 vm_object_pip_wakeup(object); 437 vm_page_io_finish(mt); 438 if (vm_page_count_severe()) 439 vm_page_try_to_cache(mt); 440 } 441 } 442 return numpagedout; 443 } 444 445 #if !defined(NO_SWAPPING) 446 /* 447 * vm_pageout_object_deactivate_pages 448 * 449 * deactivate enough pages to satisfy the inactive target 450 * requirements or if vm_page_proc_limit is set, then 451 * deactivate all of the pages in the object and its 452 * backing_objects. 453 * 454 * The object and map must be locked. 455 */ 456 static void 457 vm_pageout_object_deactivate_pages(pmap, first_object, desired) 458 pmap_t pmap; 459 vm_object_t first_object; 460 long desired; 461 { 462 vm_object_t backing_object, object; 463 vm_page_t p, next; 464 int actcount, rcount, remove_mode; 465 466 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED); 467 if (first_object->type == OBJT_DEVICE || first_object->type == OBJT_PHYS) 468 return; 469 for (object = first_object;; object = backing_object) { 470 if (pmap_resident_count(pmap) <= desired) 471 goto unlock_return; 472 if (object->paging_in_progress) 473 goto unlock_return; 474 475 remove_mode = 0; 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(pmap) <= desired) { 486 vm_page_unlock_queues(); 487 goto unlock_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(pmap, 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 pmap_remove_all(p); 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 pmap_remove_all(p); 528 } 529 p = next; 530 } 531 vm_page_unlock_queues(); 532 if ((backing_object = object->backing_object) == NULL) 533 goto unlock_return; 534 VM_OBJECT_LOCK(backing_object); 535 if (object != first_object) 536 VM_OBJECT_UNLOCK(object); 537 } 538 unlock_return: 539 if (object != first_object) 540 VM_OBJECT_UNLOCK(object); 541 } 542 543 /* 544 * deactivate some number of pages in a map, try to do it fairly, but 545 * that is really hard to do. 546 */ 547 static void 548 vm_pageout_map_deactivate_pages(map, desired) 549 vm_map_t map; 550 long desired; 551 { 552 vm_map_entry_t tmpe; 553 vm_object_t obj, bigobj; 554 int nothingwired; 555 556 if (!vm_map_trylock(map)) 557 return; 558 559 bigobj = NULL; 560 nothingwired = TRUE; 561 562 /* 563 * first, search out the biggest object, and try to free pages from 564 * that. 565 */ 566 tmpe = map->header.next; 567 while (tmpe != &map->header) { 568 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 569 obj = tmpe->object.vm_object; 570 if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) { 571 if (obj->shadow_count <= 1 && 572 (bigobj == NULL || 573 bigobj->resident_page_count < obj->resident_page_count)) { 574 if (bigobj != NULL) 575 VM_OBJECT_UNLOCK(bigobj); 576 bigobj = obj; 577 } else 578 VM_OBJECT_UNLOCK(obj); 579 } 580 } 581 if (tmpe->wired_count > 0) 582 nothingwired = FALSE; 583 tmpe = tmpe->next; 584 } 585 586 if (bigobj != NULL) { 587 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired); 588 VM_OBJECT_UNLOCK(bigobj); 589 } 590 /* 591 * Next, hunt around for other pages to deactivate. We actually 592 * do this search sort of wrong -- .text first is not the best idea. 593 */ 594 tmpe = map->header.next; 595 while (tmpe != &map->header) { 596 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 597 break; 598 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 599 obj = tmpe->object.vm_object; 600 if (obj != NULL) { 601 VM_OBJECT_LOCK(obj); 602 vm_pageout_object_deactivate_pages(map->pmap, obj, desired); 603 VM_OBJECT_UNLOCK(obj); 604 } 605 } 606 tmpe = tmpe->next; 607 } 608 609 /* 610 * Remove all mappings if a process is swapped out, this will free page 611 * table pages. 612 */ 613 if (desired == 0 && nothingwired) { 614 GIANT_REQUIRED; 615 vm_page_lock_queues(); 616 pmap_remove(vm_map_pmap(map), vm_map_min(map), 617 vm_map_max(map)); 618 vm_page_unlock_queues(); 619 } 620 vm_map_unlock(map); 621 } 622 #endif /* !defined(NO_SWAPPING) */ 623 624 /* 625 * This routine is very drastic, but can save the system 626 * in a pinch. 627 */ 628 static void 629 vm_pageout_pmap_collect(void) 630 { 631 int i; 632 vm_page_t m; 633 static int warningdone; 634 635 if (pmap_pagedaemon_waken == 0) 636 return; 637 if (warningdone < 5) { 638 printf("collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n"); 639 warningdone++; 640 } 641 vm_page_lock_queues(); 642 for (i = 0; i < vm_page_array_size; i++) { 643 m = &vm_page_array[i]; 644 if (m->wire_count || m->hold_count || m->busy || 645 (m->flags & (PG_BUSY | PG_UNMANAGED))) 646 continue; 647 pmap_remove_all(m); 648 } 649 vm_page_unlock_queues(); 650 pmap_pagedaemon_waken = 0; 651 } 652 653 /* 654 * vm_pageout_scan does the dirty work for the pageout daemon. 655 */ 656 static void 657 vm_pageout_scan(int pass) 658 { 659 vm_page_t m, next; 660 struct vm_page marker; 661 int page_shortage, maxscan, pcount; 662 int addl_page_shortage, addl_page_shortage_init; 663 struct proc *p, *bigproc; 664 struct thread *td; 665 vm_offset_t size, bigsize; 666 vm_object_t object; 667 int actcount; 668 int vnodes_skipped = 0; 669 int maxlaunder; 670 671 mtx_lock(&Giant); 672 /* 673 * Decrease registered cache sizes. 674 */ 675 EVENTHANDLER_INVOKE(vm_lowmem, 0); 676 /* 677 * We do this explicitly after the caches have been drained above. 678 */ 679 uma_reclaim(); 680 /* 681 * Do whatever cleanup that the pmap code can. 682 */ 683 vm_pageout_pmap_collect(); 684 685 addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit); 686 687 /* 688 * Calculate the number of pages we want to either free or move 689 * to the cache. 690 */ 691 page_shortage = vm_paging_target() + addl_page_shortage_init; 692 693 /* 694 * Initialize our marker 695 */ 696 bzero(&marker, sizeof(marker)); 697 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER; 698 marker.queue = PQ_INACTIVE; 699 marker.wire_count = 1; 700 701 /* 702 * Start scanning the inactive queue for pages we can move to the 703 * cache or free. The scan will stop when the target is reached or 704 * we have scanned the entire inactive queue. Note that m->act_count 705 * is not used to form decisions for the inactive queue, only for the 706 * active queue. 707 * 708 * maxlaunder limits the number of dirty pages we flush per scan. 709 * For most systems a smaller value (16 or 32) is more robust under 710 * extreme memory and disk pressure because any unnecessary writes 711 * to disk can result in extreme performance degredation. However, 712 * systems with excessive dirty pages (especially when MAP_NOSYNC is 713 * used) will die horribly with limited laundering. If the pageout 714 * daemon cannot clean enough pages in the first pass, we let it go 715 * all out in succeeding passes. 716 */ 717 if ((maxlaunder = vm_max_launder) <= 1) 718 maxlaunder = 1; 719 if (pass) 720 maxlaunder = 10000; 721 vm_page_lock_queues(); 722 rescan0: 723 addl_page_shortage = addl_page_shortage_init; 724 maxscan = cnt.v_inactive_count; 725 726 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl); 727 m != NULL && maxscan-- > 0 && page_shortage > 0; 728 m = next) { 729 730 cnt.v_pdpages++; 731 732 if (m->queue != PQ_INACTIVE) { 733 goto rescan0; 734 } 735 736 next = TAILQ_NEXT(m, pageq); 737 738 /* 739 * skip marker pages 740 */ 741 if (m->flags & PG_MARKER) 742 continue; 743 744 /* 745 * A held page may be undergoing I/O, so skip it. 746 */ 747 if (m->hold_count) { 748 vm_pageq_requeue(m); 749 addl_page_shortage++; 750 continue; 751 } 752 /* 753 * Don't mess with busy pages, keep in the front of the 754 * queue, most likely are being paged out. 755 */ 756 if (m->busy || (m->flags & PG_BUSY)) { 757 addl_page_shortage++; 758 continue; 759 } 760 761 /* 762 * If the object is not being used, we ignore previous 763 * references. 764 */ 765 if (m->object->ref_count == 0) { 766 vm_page_flag_clear(m, PG_REFERENCED); 767 pmap_clear_reference(m); 768 769 /* 770 * Otherwise, if the page has been referenced while in the 771 * inactive queue, we bump the "activation count" upwards, 772 * making it less likely that the page will be added back to 773 * the inactive queue prematurely again. Here we check the 774 * page tables (or emulated bits, if any), given the upper 775 * level VM system not knowing anything about existing 776 * references. 777 */ 778 } else if (((m->flags & PG_REFERENCED) == 0) && 779 (actcount = pmap_ts_referenced(m))) { 780 vm_page_activate(m); 781 m->act_count += (actcount + ACT_ADVANCE); 782 continue; 783 } 784 785 /* 786 * If the upper level VM system knows about any page 787 * references, we activate the page. We also set the 788 * "activation count" higher than normal so that we will less 789 * likely place pages back onto the inactive queue again. 790 */ 791 if ((m->flags & PG_REFERENCED) != 0) { 792 vm_page_flag_clear(m, PG_REFERENCED); 793 actcount = pmap_ts_referenced(m); 794 vm_page_activate(m); 795 m->act_count += (actcount + ACT_ADVANCE + 1); 796 continue; 797 } 798 799 /* 800 * If the upper level VM system doesn't know anything about 801 * the page being dirty, we have to check for it again. As 802 * far as the VM code knows, any partially dirty pages are 803 * fully dirty. 804 */ 805 if (m->dirty == 0 && !pmap_is_modified(m)) { 806 /* 807 * Avoid a race condition: Unless write access is 808 * removed from the page, another processor could 809 * modify it before all access is removed by the call 810 * to vm_page_cache() below. If vm_page_cache() finds 811 * that the page has been modified when it removes all 812 * access, it panics because it cannot cache dirty 813 * pages. In principle, we could eliminate just write 814 * access here rather than all access. In the expected 815 * case, when there are no last instant modifications 816 * to the page, removing all access will be cheaper 817 * overall. 818 */ 819 if ((m->flags & PG_WRITEABLE) != 0) 820 pmap_remove_all(m); 821 } else { 822 vm_page_dirty(m); 823 } 824 825 object = m->object; 826 if (!VM_OBJECT_TRYLOCK(object)) 827 continue; 828 if (m->valid == 0) { 829 /* 830 * Invalid pages can be easily freed 831 */ 832 vm_page_busy(m); 833 pmap_remove_all(m); 834 vm_page_free(m); 835 cnt.v_dfree++; 836 --page_shortage; 837 } else if (m->dirty == 0) { 838 /* 839 * Clean pages can be placed onto the cache queue. 840 * This effectively frees them. 841 */ 842 vm_page_cache(m); 843 --page_shortage; 844 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) { 845 /* 846 * Dirty pages need to be paged out, but flushing 847 * a page is extremely expensive verses freeing 848 * a clean page. Rather then artificially limiting 849 * the number of pages we can flush, we instead give 850 * dirty pages extra priority on the inactive queue 851 * by forcing them to be cycled through the queue 852 * twice before being flushed, after which the 853 * (now clean) page will cycle through once more 854 * before being freed. This significantly extends 855 * the thrash point for a heavily loaded machine. 856 */ 857 vm_page_flag_set(m, PG_WINATCFLS); 858 vm_pageq_requeue(m); 859 } else if (maxlaunder > 0) { 860 /* 861 * We always want to try to flush some dirty pages if 862 * we encounter them, to keep the system stable. 863 * Normally this number is small, but under extreme 864 * pressure where there are insufficient clean pages 865 * on the inactive queue, we may have to go all out. 866 */ 867 int swap_pageouts_ok; 868 struct vnode *vp = NULL; 869 struct mount *mp; 870 871 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 872 swap_pageouts_ok = 1; 873 } else { 874 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 875 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 876 vm_page_count_min()); 877 878 } 879 880 /* 881 * We don't bother paging objects that are "dead". 882 * Those objects are in a "rundown" state. 883 */ 884 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 885 VM_OBJECT_UNLOCK(object); 886 vm_pageq_requeue(m); 887 continue; 888 } 889 890 /* 891 * The object is already known NOT to be dead. It 892 * is possible for the vget() to block the whole 893 * pageout daemon, but the new low-memory handling 894 * code should prevent it. 895 * 896 * The previous code skipped locked vnodes and, worse, 897 * reordered pages in the queue. This results in 898 * completely non-deterministic operation and, on a 899 * busy system, can lead to extremely non-optimal 900 * pageouts. For example, it can cause clean pages 901 * to be freed and dirty pages to be moved to the end 902 * of the queue. Since dirty pages are also moved to 903 * the end of the queue once-cleaned, this gives 904 * way too large a weighting to defering the freeing 905 * of dirty pages. 906 * 907 * We can't wait forever for the vnode lock, we might 908 * deadlock due to a vn_read() getting stuck in 909 * vm_wait while holding this vnode. We skip the 910 * vnode if we can't get it in a reasonable amount 911 * of time. 912 */ 913 if (object->type == OBJT_VNODE) { 914 vp = object->handle; 915 mp = NULL; 916 if (vp->v_type == VREG) 917 vn_start_write(vp, &mp, V_NOWAIT); 918 vm_page_unlock_queues(); 919 VI_LOCK(vp); 920 VM_OBJECT_UNLOCK(object); 921 if (vget(vp, LK_EXCLUSIVE | LK_INTERLOCK | 922 LK_TIMELOCK, curthread)) { 923 VM_OBJECT_LOCK(object); 924 vm_page_lock_queues(); 925 ++pageout_lock_miss; 926 vn_finished_write(mp); 927 if (object->flags & OBJ_MIGHTBEDIRTY) 928 vnodes_skipped++; 929 VM_OBJECT_UNLOCK(object); 930 continue; 931 } 932 VM_OBJECT_LOCK(object); 933 vm_page_lock_queues(); 934 /* 935 * The page might have been moved to another 936 * queue during potential blocking in vget() 937 * above. The page might have been freed and 938 * reused for another vnode. The object might 939 * have been reused for another vnode. 940 */ 941 if (m->queue != PQ_INACTIVE || 942 m->object != object || 943 object->handle != vp) { 944 if (object->flags & OBJ_MIGHTBEDIRTY) 945 vnodes_skipped++; 946 goto unlock_and_continue; 947 } 948 949 /* 950 * The page may have been busied during the 951 * blocking in vput(); We don't move the 952 * page back onto the end of the queue so that 953 * statistics are more correct if we don't. 954 */ 955 if (m->busy || (m->flags & PG_BUSY)) { 956 goto unlock_and_continue; 957 } 958 959 /* 960 * If the page has become held it might 961 * be undergoing I/O, so skip it 962 */ 963 if (m->hold_count) { 964 vm_pageq_requeue(m); 965 if (object->flags & OBJ_MIGHTBEDIRTY) 966 vnodes_skipped++; 967 goto unlock_and_continue; 968 } 969 } 970 971 /* 972 * If a page is dirty, then it is either being washed 973 * (but not yet cleaned) or it is still in the 974 * laundry. If it is still in the laundry, then we 975 * start the cleaning operation. 976 * 977 * This operation may cluster, invalidating the 'next' 978 * pointer. To prevent an inordinate number of 979 * restarts we use our marker to remember our place. 980 * 981 * decrement page_shortage on success to account for 982 * the (future) cleaned page. Otherwise we could wind 983 * up laundering or cleaning too many pages. 984 */ 985 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, m, &marker, pageq); 986 if (vm_pageout_clean(m) != 0) { 987 --page_shortage; 988 --maxlaunder; 989 } 990 next = TAILQ_NEXT(&marker, pageq); 991 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, &marker, pageq); 992 unlock_and_continue: 993 VM_OBJECT_UNLOCK(object); 994 if (vp) { 995 vm_page_unlock_queues(); 996 vput(vp); 997 vn_finished_write(mp); 998 vm_page_lock_queues(); 999 } 1000 continue; 1001 } 1002 VM_OBJECT_UNLOCK(object); 1003 } 1004 1005 /* 1006 * Compute the number of pages we want to try to move from the 1007 * active queue to the inactive queue. 1008 */ 1009 page_shortage = vm_paging_target() + 1010 cnt.v_inactive_target - cnt.v_inactive_count; 1011 page_shortage += addl_page_shortage; 1012 1013 /* 1014 * Scan the active queue for things we can deactivate. We nominally 1015 * track the per-page activity counter and use it to locate 1016 * deactivation candidates. 1017 */ 1018 pcount = cnt.v_active_count; 1019 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1020 1021 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) { 1022 1023 KASSERT(m->queue == PQ_ACTIVE, 1024 ("vm_pageout_scan: page %p isn't active", m)); 1025 1026 next = TAILQ_NEXT(m, pageq); 1027 /* 1028 * Don't deactivate pages that are busy. 1029 */ 1030 if ((m->busy != 0) || 1031 (m->flags & PG_BUSY) || 1032 (m->hold_count != 0)) { 1033 vm_pageq_requeue(m); 1034 m = next; 1035 continue; 1036 } 1037 1038 /* 1039 * The count for pagedaemon pages is done after checking the 1040 * page for eligibility... 1041 */ 1042 cnt.v_pdpages++; 1043 1044 /* 1045 * Check to see "how much" the page has been used. 1046 */ 1047 actcount = 0; 1048 if (m->object->ref_count != 0) { 1049 if (m->flags & PG_REFERENCED) { 1050 actcount += 1; 1051 } 1052 actcount += pmap_ts_referenced(m); 1053 if (actcount) { 1054 m->act_count += ACT_ADVANCE + actcount; 1055 if (m->act_count > ACT_MAX) 1056 m->act_count = ACT_MAX; 1057 } 1058 } 1059 1060 /* 1061 * Since we have "tested" this bit, we need to clear it now. 1062 */ 1063 vm_page_flag_clear(m, PG_REFERENCED); 1064 1065 /* 1066 * Only if an object is currently being used, do we use the 1067 * page activation count stats. 1068 */ 1069 if (actcount && (m->object->ref_count != 0)) { 1070 vm_pageq_requeue(m); 1071 } else { 1072 m->act_count -= min(m->act_count, ACT_DECLINE); 1073 if (vm_pageout_algorithm || 1074 m->object->ref_count == 0 || 1075 m->act_count == 0) { 1076 page_shortage--; 1077 if (m->object->ref_count == 0) { 1078 pmap_remove_all(m); 1079 if (m->dirty == 0) 1080 vm_page_cache(m); 1081 else 1082 vm_page_deactivate(m); 1083 } else { 1084 vm_page_deactivate(m); 1085 } 1086 } else { 1087 vm_pageq_requeue(m); 1088 } 1089 } 1090 m = next; 1091 } 1092 1093 /* 1094 * We try to maintain some *really* free pages, this allows interrupt 1095 * code to be guaranteed space. Since both cache and free queues 1096 * are considered basically 'free', moving pages from cache to free 1097 * does not effect other calculations. 1098 */ 1099 while (cnt.v_free_count < cnt.v_free_reserved) { 1100 static int cache_rover = 0; 1101 1102 if ((m = vm_page_select_cache(cache_rover)) == NULL) 1103 break; 1104 cache_rover = (m->pc + PQ_PRIME2) & PQ_L2_MASK; 1105 object = m->object; 1106 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1107 vm_page_busy(m); 1108 vm_page_free(m); 1109 VM_OBJECT_UNLOCK(object); 1110 cnt.v_dfree++; 1111 } 1112 vm_page_unlock_queues(); 1113 #if !defined(NO_SWAPPING) 1114 /* 1115 * Idle process swapout -- run once per second. 1116 */ 1117 if (vm_swap_idle_enabled) { 1118 static long lsec; 1119 if (time_second != lsec) { 1120 vm_pageout_req_swapout |= VM_SWAP_IDLE; 1121 vm_req_vmdaemon(); 1122 lsec = time_second; 1123 } 1124 } 1125 #endif 1126 1127 /* 1128 * If we didn't get enough free pages, and we have skipped a vnode 1129 * in a writeable object, wakeup the sync daemon. And kick swapout 1130 * if we did not get enough free pages. 1131 */ 1132 if (vm_paging_target() > 0) { 1133 if (vnodes_skipped && vm_page_count_min()) 1134 (void) speedup_syncer(); 1135 #if !defined(NO_SWAPPING) 1136 if (vm_swap_enabled && vm_page_count_target()) { 1137 vm_req_vmdaemon(); 1138 vm_pageout_req_swapout |= VM_SWAP_NORMAL; 1139 } 1140 #endif 1141 } 1142 1143 /* 1144 * If we are critically low on one of RAM or swap and low on 1145 * the other, kill the largest process. However, we avoid 1146 * doing this on the first pass in order to give ourselves a 1147 * chance to flush out dirty vnode-backed pages and to allow 1148 * active pages to be moved to the inactive queue and reclaimed. 1149 * 1150 * We keep the process bigproc locked once we find it to keep anyone 1151 * from messing with it; however, there is a possibility of 1152 * deadlock if process B is bigproc and one of it's child processes 1153 * attempts to propagate a signal to B while we are waiting for A's 1154 * lock while walking this list. To avoid this, we don't block on 1155 * the process lock but just skip a process if it is already locked. 1156 */ 1157 if (pass != 0 && 1158 ((swap_pager_avail < 64 && vm_page_count_min()) || 1159 (swap_pager_full && vm_paging_target() > 0))) { 1160 bigproc = NULL; 1161 bigsize = 0; 1162 sx_slock(&allproc_lock); 1163 FOREACH_PROC_IN_SYSTEM(p) { 1164 int breakout; 1165 1166 if (PROC_TRYLOCK(p) == 0) 1167 continue; 1168 /* 1169 * If this is a system or protected process, skip it. 1170 */ 1171 if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) || 1172 (p->p_flag & P_PROTECTED) || 1173 ((p->p_pid < 48) && (swap_pager_avail != 0))) { 1174 PROC_UNLOCK(p); 1175 continue; 1176 } 1177 /* 1178 * If the process is in a non-running type state, 1179 * don't touch it. Check all the threads individually. 1180 */ 1181 mtx_lock_spin(&sched_lock); 1182 breakout = 0; 1183 FOREACH_THREAD_IN_PROC(p, td) { 1184 if (!TD_ON_RUNQ(td) && 1185 !TD_IS_RUNNING(td) && 1186 !TD_IS_SLEEPING(td)) { 1187 breakout = 1; 1188 break; 1189 } 1190 } 1191 if (breakout) { 1192 mtx_unlock_spin(&sched_lock); 1193 PROC_UNLOCK(p); 1194 continue; 1195 } 1196 mtx_unlock_spin(&sched_lock); 1197 /* 1198 * get the process size 1199 */ 1200 if (!vm_map_trylock_read(&p->p_vmspace->vm_map)) { 1201 PROC_UNLOCK(p); 1202 continue; 1203 } 1204 size = vmspace_swap_count(p->p_vmspace); 1205 vm_map_unlock_read(&p->p_vmspace->vm_map); 1206 size += vmspace_resident_count(p->p_vmspace); 1207 /* 1208 * if the this process is bigger than the biggest one 1209 * remember it. 1210 */ 1211 if (size > bigsize) { 1212 if (bigproc != NULL) 1213 PROC_UNLOCK(bigproc); 1214 bigproc = p; 1215 bigsize = size; 1216 } else 1217 PROC_UNLOCK(p); 1218 } 1219 sx_sunlock(&allproc_lock); 1220 if (bigproc != NULL) { 1221 killproc(bigproc, "out of swap space"); 1222 mtx_lock_spin(&sched_lock); 1223 sched_nice(bigproc, PRIO_MIN); 1224 mtx_unlock_spin(&sched_lock); 1225 PROC_UNLOCK(bigproc); 1226 wakeup(&cnt.v_free_count); 1227 } 1228 } 1229 mtx_unlock(&Giant); 1230 } 1231 1232 /* 1233 * This routine tries to maintain the pseudo LRU active queue, 1234 * so that during long periods of time where there is no paging, 1235 * that some statistic accumulation still occurs. This code 1236 * helps the situation where paging just starts to occur. 1237 */ 1238 static void 1239 vm_pageout_page_stats() 1240 { 1241 vm_page_t m,next; 1242 int pcount,tpcount; /* Number of pages to check */ 1243 static int fullintervalcount = 0; 1244 int page_shortage; 1245 1246 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1247 page_shortage = 1248 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) - 1249 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count); 1250 1251 if (page_shortage <= 0) 1252 return; 1253 1254 pcount = cnt.v_active_count; 1255 fullintervalcount += vm_pageout_stats_interval; 1256 if (fullintervalcount < vm_pageout_full_stats_interval) { 1257 tpcount = (vm_pageout_stats_max * cnt.v_active_count) / cnt.v_page_count; 1258 if (pcount > tpcount) 1259 pcount = tpcount; 1260 } else { 1261 fullintervalcount = 0; 1262 } 1263 1264 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1265 while ((m != NULL) && (pcount-- > 0)) { 1266 int actcount; 1267 1268 KASSERT(m->queue == PQ_ACTIVE, 1269 ("vm_pageout_page_stats: page %p isn't active", m)); 1270 1271 next = TAILQ_NEXT(m, pageq); 1272 /* 1273 * Don't deactivate pages that are busy. 1274 */ 1275 if ((m->busy != 0) || 1276 (m->flags & PG_BUSY) || 1277 (m->hold_count != 0)) { 1278 vm_pageq_requeue(m); 1279 m = next; 1280 continue; 1281 } 1282 1283 actcount = 0; 1284 if (m->flags & PG_REFERENCED) { 1285 vm_page_flag_clear(m, PG_REFERENCED); 1286 actcount += 1; 1287 } 1288 1289 actcount += pmap_ts_referenced(m); 1290 if (actcount) { 1291 m->act_count += ACT_ADVANCE + actcount; 1292 if (m->act_count > ACT_MAX) 1293 m->act_count = ACT_MAX; 1294 vm_pageq_requeue(m); 1295 } else { 1296 if (m->act_count == 0) { 1297 /* 1298 * We turn off page access, so that we have 1299 * more accurate RSS stats. We don't do this 1300 * in the normal page deactivation when the 1301 * system is loaded VM wise, because the 1302 * cost of the large number of page protect 1303 * operations would be higher than the value 1304 * of doing the operation. 1305 */ 1306 pmap_remove_all(m); 1307 vm_page_deactivate(m); 1308 } else { 1309 m->act_count -= min(m->act_count, ACT_DECLINE); 1310 vm_pageq_requeue(m); 1311 } 1312 } 1313 1314 m = next; 1315 } 1316 } 1317 1318 /* 1319 * vm_pageout is the high level pageout daemon. 1320 */ 1321 static void 1322 vm_pageout() 1323 { 1324 int error, pass; 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 vm_page_lock_queues(); 1406 /* 1407 * If we have enough free memory, wakeup waiters. Do 1408 * not clear vm_pages_needed until we reach our target, 1409 * otherwise we may be woken up over and over again and 1410 * waste a lot of cpu. 1411 */ 1412 if (vm_pages_needed && !vm_page_count_min()) { 1413 if (!vm_paging_needed()) 1414 vm_pages_needed = 0; 1415 wakeup(&cnt.v_free_count); 1416 } 1417 if (vm_pages_needed) { 1418 /* 1419 * Still not done, take a second pass without waiting 1420 * (unlimited dirty cleaning), otherwise sleep a bit 1421 * and try again. 1422 */ 1423 ++pass; 1424 if (pass > 1) 1425 msleep(&vm_pages_needed, &vm_page_queue_mtx, PVM, 1426 "psleep", hz/2); 1427 } else { 1428 /* 1429 * Good enough, sleep & handle stats. Prime the pass 1430 * for the next run. 1431 */ 1432 if (pass > 1) 1433 pass = 1; 1434 else 1435 pass = 0; 1436 error = msleep(&vm_pages_needed, &vm_page_queue_mtx, PVM, 1437 "psleep", vm_pageout_stats_interval * hz); 1438 if (error && !vm_pages_needed) { 1439 pass = 0; 1440 vm_pageout_page_stats(); 1441 vm_page_unlock_queues(); 1442 continue; 1443 } 1444 } 1445 if (vm_pages_needed) 1446 cnt.v_pdwakeups++; 1447 vm_page_unlock_queues(); 1448 vm_pageout_scan(pass); 1449 } 1450 } 1451 1452 /* 1453 * Unless the page queue lock is held by the caller, this function 1454 * should be regarded as advisory. Specifically, the caller should 1455 * not msleep() on &cnt.v_free_count following this function unless 1456 * the page queue lock is held until the msleep() is performed. 1457 */ 1458 void 1459 pagedaemon_wakeup() 1460 { 1461 1462 if (!vm_pages_needed && curthread->td_proc != pageproc) { 1463 vm_pages_needed = 1; 1464 wakeup(&vm_pages_needed); 1465 } 1466 } 1467 1468 #if !defined(NO_SWAPPING) 1469 static void 1470 vm_req_vmdaemon() 1471 { 1472 static int lastrun = 0; 1473 1474 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1475 wakeup(&vm_daemon_needed); 1476 lastrun = ticks; 1477 } 1478 } 1479 1480 static void 1481 vm_daemon() 1482 { 1483 struct rlimit rsslim; 1484 struct proc *p; 1485 struct thread *td; 1486 int breakout; 1487 1488 mtx_lock(&Giant); 1489 while (TRUE) { 1490 tsleep(&vm_daemon_needed, PPAUSE, "psleep", 0); 1491 if (vm_pageout_req_swapout) { 1492 swapout_procs(vm_pageout_req_swapout); 1493 vm_pageout_req_swapout = 0; 1494 } 1495 /* 1496 * scan the processes for exceeding their rlimits or if 1497 * process is swapped out -- deactivate pages 1498 */ 1499 sx_slock(&allproc_lock); 1500 LIST_FOREACH(p, &allproc, p_list) { 1501 vm_pindex_t limit, size; 1502 1503 /* 1504 * if this is a system process or if we have already 1505 * looked at this process, skip it. 1506 */ 1507 PROC_LOCK(p); 1508 if (p->p_flag & (P_SYSTEM | P_WEXIT)) { 1509 PROC_UNLOCK(p); 1510 continue; 1511 } 1512 /* 1513 * if the process is in a non-running type state, 1514 * don't touch it. 1515 */ 1516 mtx_lock_spin(&sched_lock); 1517 breakout = 0; 1518 FOREACH_THREAD_IN_PROC(p, td) { 1519 if (!TD_ON_RUNQ(td) && 1520 !TD_IS_RUNNING(td) && 1521 !TD_IS_SLEEPING(td)) { 1522 breakout = 1; 1523 break; 1524 } 1525 } 1526 mtx_unlock_spin(&sched_lock); 1527 if (breakout) { 1528 PROC_UNLOCK(p); 1529 continue; 1530 } 1531 /* 1532 * get a limit 1533 */ 1534 lim_rlimit(p, RLIMIT_RSS, &rsslim); 1535 limit = OFF_TO_IDX( 1536 qmin(rsslim.rlim_cur, rsslim.rlim_max)); 1537 1538 /* 1539 * let processes that are swapped out really be 1540 * swapped out set the limit to nothing (will force a 1541 * swap-out.) 1542 */ 1543 if ((p->p_sflag & PS_INMEM) == 0) 1544 limit = 0; /* XXX */ 1545 PROC_UNLOCK(p); 1546 1547 size = vmspace_resident_count(p->p_vmspace); 1548 if (limit >= 0 && size >= limit) { 1549 vm_pageout_map_deactivate_pages( 1550 &p->p_vmspace->vm_map, limit); 1551 } 1552 } 1553 sx_sunlock(&allproc_lock); 1554 } 1555 } 1556 #endif /* !defined(NO_SWAPPING) */ 1557