1 /* 2 * Copyright (c) 1991 Regents of the University of California. 3 * All rights reserved. 4 * Copyright (c) 1994 John S. Dyson 5 * All rights reserved. 6 * Copyright (c) 1994 David Greenman 7 * All rights reserved. 8 * 9 * This code is derived from software contributed to Berkeley by 10 * The Mach Operating System project at Carnegie-Mellon University. 11 * 12 * Redistribution and use in source and binary forms, with or without 13 * modification, are permitted provided that the following conditions 14 * are met: 15 * 1. Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in the 19 * documentation and/or other materials provided with the distribution. 20 * 3. All advertising materials mentioning features or use of this software 21 * must display the following acknowledgement: 22 * This product includes software developed by the University of 23 * California, Berkeley and its contributors. 24 * 4. Neither the name of the University nor the names of its contributors 25 * may be used to endorse or promote products derived from this software 26 * without specific prior written permission. 27 * 28 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 29 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 30 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 31 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 32 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 33 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 34 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 35 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 36 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 37 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 38 * SUCH DAMAGE. 39 * 40 * from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91 41 * 42 * 43 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 44 * All rights reserved. 45 * 46 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 47 * 48 * Permission to use, copy, modify and distribute this software and 49 * its documentation is hereby granted, provided that both the copyright 50 * notice and this permission notice appear in all copies of the 51 * software, derivative works or modified versions, and any portions 52 * thereof, and that both notices appear in supporting documentation. 53 * 54 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 55 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 56 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 57 * 58 * Carnegie Mellon requests users of this software to return to 59 * 60 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 61 * School of Computer Science 62 * Carnegie Mellon University 63 * Pittsburgh PA 15213-3890 64 * 65 * any improvements or extensions that they make and grant Carnegie the 66 * rights to redistribute these changes. 67 * 68 * $FreeBSD: src/sys/vm/vm_pageout.c,v 1.151.2.15 2002/12/29 18:21:04 dillon Exp $ 69 * $DragonFly: src/sys/vm/vm_pageout.c,v 1.31 2007/02/16 23:11:40 corecode Exp $ 70 */ 71 72 /* 73 * The proverbial page-out daemon. 74 */ 75 76 #include "opt_vm.h" 77 #include <sys/param.h> 78 #include <sys/systm.h> 79 #include <sys/kernel.h> 80 #include <sys/proc.h> 81 #include <sys/kthread.h> 82 #include <sys/resourcevar.h> 83 #include <sys/signalvar.h> 84 #include <sys/vnode.h> 85 #include <sys/vmmeter.h> 86 #include <sys/sysctl.h> 87 88 #include <vm/vm.h> 89 #include <vm/vm_param.h> 90 #include <sys/lock.h> 91 #include <vm/vm_object.h> 92 #include <vm/vm_page.h> 93 #include <vm/vm_map.h> 94 #include <vm/vm_pageout.h> 95 #include <vm/vm_pager.h> 96 #include <vm/swap_pager.h> 97 #include <vm/vm_extern.h> 98 99 #include <sys/thread2.h> 100 #include <vm/vm_page2.h> 101 102 /* 103 * System initialization 104 */ 105 106 /* the kernel process "vm_pageout"*/ 107 static void vm_pageout (void); 108 static int vm_pageout_clean (vm_page_t); 109 static void vm_pageout_scan (int pass); 110 static int vm_pageout_free_page_calc (vm_size_t count); 111 struct thread *pagethread; 112 113 static struct kproc_desc page_kp = { 114 "pagedaemon", 115 vm_pageout, 116 &pagethread 117 }; 118 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp) 119 120 #if !defined(NO_SWAPPING) 121 /* the kernel process "vm_daemon"*/ 122 static void vm_daemon (void); 123 static struct thread *vmthread; 124 125 static struct kproc_desc vm_kp = { 126 "vmdaemon", 127 vm_daemon, 128 &vmthread 129 }; 130 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp) 131 #endif 132 133 134 int vm_pages_needed=0; /* Event on which pageout daemon sleeps */ 135 int vm_pageout_deficit=0; /* Estimated number of pages deficit */ 136 int vm_pageout_pages_needed=0; /* flag saying that the pageout daemon needs pages */ 137 138 #if !defined(NO_SWAPPING) 139 static int vm_pageout_req_swapout; /* XXX */ 140 static int vm_daemon_needed; 141 #endif 142 extern int vm_swap_size; 143 static int vm_max_launder = 32; 144 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0; 145 static int vm_pageout_full_stats_interval = 0; 146 static int vm_pageout_stats_free_max=0, vm_pageout_algorithm=0; 147 static int defer_swap_pageouts=0; 148 static int disable_swap_pageouts=0; 149 150 #if defined(NO_SWAPPING) 151 static int vm_swap_enabled=0; 152 static int vm_swap_idle_enabled=0; 153 #else 154 static int vm_swap_enabled=1; 155 static int vm_swap_idle_enabled=0; 156 #endif 157 158 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm, 159 CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt"); 160 161 SYSCTL_INT(_vm, OID_AUTO, max_launder, 162 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout"); 163 164 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max, 165 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length"); 166 167 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval, 168 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan"); 169 170 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval, 171 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan"); 172 173 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_free_max, 174 CTLFLAG_RW, &vm_pageout_stats_free_max, 0, "Not implemented"); 175 176 #if defined(NO_SWAPPING) 177 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 178 CTLFLAG_RD, &vm_swap_enabled, 0, ""); 179 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 180 CTLFLAG_RD, &vm_swap_idle_enabled, 0, ""); 181 #else 182 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 183 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout"); 184 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 185 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 186 #endif 187 188 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts, 189 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem"); 190 191 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts, 192 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages"); 193 194 static int pageout_lock_miss; 195 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss, 196 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout"); 197 198 int vm_load; 199 SYSCTL_INT(_vm, OID_AUTO, vm_load, 200 CTLFLAG_RD, &vm_load, 0, "load on the VM system"); 201 int vm_load_enable = 1; 202 SYSCTL_INT(_vm, OID_AUTO, vm_load_enable, 203 CTLFLAG_RW, &vm_load_enable, 0, "enable vm_load rate limiting"); 204 #ifdef INVARIANTS 205 int vm_load_debug; 206 SYSCTL_INT(_vm, OID_AUTO, vm_load_debug, 207 CTLFLAG_RW, &vm_load_debug, 0, "debug vm_load"); 208 #endif 209 210 #define VM_PAGEOUT_PAGE_COUNT 16 211 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT; 212 213 int vm_page_max_wired; /* XXX max # of wired pages system-wide */ 214 215 #if !defined(NO_SWAPPING) 216 typedef void freeer_fcn_t (vm_map_t, vm_object_t, vm_pindex_t, int); 217 static void vm_pageout_map_deactivate_pages (vm_map_t, vm_pindex_t); 218 static freeer_fcn_t vm_pageout_object_deactivate_pages; 219 static void vm_req_vmdaemon (void); 220 #endif 221 static void vm_pageout_page_stats(void); 222 223 /* 224 * Update 225 */ 226 void 227 vm_fault_ratecheck(void) 228 { 229 if (vm_pages_needed) { 230 if (vm_load < 1000) 231 ++vm_load; 232 } else { 233 if (vm_load > 0) 234 --vm_load; 235 } 236 } 237 238 /* 239 * vm_pageout_clean: 240 * 241 * Clean the page and remove it from the laundry. The page must not be 242 * busy on-call. 243 * 244 * We set the busy bit to cause potential page faults on this page to 245 * block. Note the careful timing, however, the busy bit isn't set till 246 * late and we cannot do anything that will mess with the page. 247 */ 248 249 static int 250 vm_pageout_clean(vm_page_t m) 251 { 252 vm_object_t object; 253 vm_page_t mc[2*vm_pageout_page_count]; 254 int pageout_count; 255 int ib, is, page_base; 256 vm_pindex_t pindex = m->pindex; 257 258 object = m->object; 259 260 /* 261 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP 262 * with the new swapper, but we could have serious problems paging 263 * out other object types if there is insufficient memory. 264 * 265 * Unfortunately, checking free memory here is far too late, so the 266 * check has been moved up a procedural level. 267 */ 268 269 /* 270 * Don't mess with the page if it's busy, held, or special 271 */ 272 if ((m->hold_count != 0) || 273 ((m->busy != 0) || (m->flags & (PG_BUSY|PG_UNMANAGED)))) { 274 return 0; 275 } 276 277 mc[vm_pageout_page_count] = m; 278 pageout_count = 1; 279 page_base = vm_pageout_page_count; 280 ib = 1; 281 is = 1; 282 283 /* 284 * Scan object for clusterable pages. 285 * 286 * We can cluster ONLY if: ->> the page is NOT 287 * clean, wired, busy, held, or mapped into a 288 * buffer, and one of the following: 289 * 1) The page is inactive, or a seldom used 290 * active page. 291 * -or- 292 * 2) we force the issue. 293 * 294 * During heavy mmap/modification loads the pageout 295 * daemon can really fragment the underlying file 296 * due to flushing pages out of order and not trying 297 * align the clusters (which leave sporatic out-of-order 298 * holes). To solve this problem we do the reverse scan 299 * first and attempt to align our cluster, then do a 300 * forward scan if room remains. 301 */ 302 303 more: 304 while (ib && pageout_count < vm_pageout_page_count) { 305 vm_page_t p; 306 307 if (ib > pindex) { 308 ib = 0; 309 break; 310 } 311 312 if ((p = vm_page_lookup(object, pindex - ib)) == NULL) { 313 ib = 0; 314 break; 315 } 316 if (((p->queue - p->pc) == PQ_CACHE) || 317 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) { 318 ib = 0; 319 break; 320 } 321 vm_page_test_dirty(p); 322 if ((p->dirty & p->valid) == 0 || 323 p->queue != PQ_INACTIVE || 324 p->wire_count != 0 || /* may be held by buf cache */ 325 p->hold_count != 0) { /* may be undergoing I/O */ 326 ib = 0; 327 break; 328 } 329 mc[--page_base] = p; 330 ++pageout_count; 331 ++ib; 332 /* 333 * alignment boundry, stop here and switch directions. Do 334 * not clear ib. 335 */ 336 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0) 337 break; 338 } 339 340 while (pageout_count < vm_pageout_page_count && 341 pindex + is < object->size) { 342 vm_page_t p; 343 344 if ((p = vm_page_lookup(object, pindex + is)) == NULL) 345 break; 346 if (((p->queue - p->pc) == PQ_CACHE) || 347 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) { 348 break; 349 } 350 vm_page_test_dirty(p); 351 if ((p->dirty & p->valid) == 0 || 352 p->queue != PQ_INACTIVE || 353 p->wire_count != 0 || /* may be held by buf cache */ 354 p->hold_count != 0) { /* may be undergoing I/O */ 355 break; 356 } 357 mc[page_base + pageout_count] = p; 358 ++pageout_count; 359 ++is; 360 } 361 362 /* 363 * If we exhausted our forward scan, continue with the reverse scan 364 * when possible, even past a page boundry. This catches boundry 365 * conditions. 366 */ 367 if (ib && pageout_count < vm_pageout_page_count) 368 goto more; 369 370 /* 371 * we allow reads during pageouts... 372 */ 373 return vm_pageout_flush(&mc[page_base], pageout_count, 0); 374 } 375 376 /* 377 * vm_pageout_flush() - launder the given pages 378 * 379 * The given pages are laundered. Note that we setup for the start of 380 * I/O ( i.e. busy the page ), mark it read-only, and bump the object 381 * reference count all in here rather then in the parent. If we want 382 * the parent to do more sophisticated things we may have to change 383 * the ordering. 384 */ 385 386 int 387 vm_pageout_flush(vm_page_t *mc, int count, int flags) 388 { 389 vm_object_t object; 390 int pageout_status[count]; 391 int numpagedout = 0; 392 int i; 393 394 /* 395 * Initiate I/O. Bump the vm_page_t->busy counter and 396 * mark the pages read-only. 397 * 398 * We do not have to fixup the clean/dirty bits here... we can 399 * allow the pager to do it after the I/O completes. 400 */ 401 402 for (i = 0; i < count; i++) { 403 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, ("vm_pageout_flush page %p index %d/%d: partially invalid page", mc[i], i, count)); 404 vm_page_io_start(mc[i]); 405 vm_page_protect(mc[i], VM_PROT_READ); 406 } 407 408 object = mc[0]->object; 409 vm_object_pip_add(object, count); 410 411 vm_pager_put_pages(object, mc, count, 412 (flags | ((object == &kernel_object) ? VM_PAGER_PUT_SYNC : 0)), 413 pageout_status); 414 415 for (i = 0; i < count; i++) { 416 vm_page_t mt = mc[i]; 417 418 switch (pageout_status[i]) { 419 case VM_PAGER_OK: 420 numpagedout++; 421 break; 422 case VM_PAGER_PEND: 423 numpagedout++; 424 break; 425 case VM_PAGER_BAD: 426 /* 427 * Page outside of range of object. Right now we 428 * essentially lose the changes by pretending it 429 * worked. 430 */ 431 pmap_clear_modify(mt); 432 vm_page_undirty(mt); 433 break; 434 case VM_PAGER_ERROR: 435 case VM_PAGER_FAIL: 436 /* 437 * If page couldn't be paged out, then reactivate the 438 * page so it doesn't clog the inactive list. (We 439 * will try paging out it again later). 440 */ 441 vm_page_activate(mt); 442 break; 443 case VM_PAGER_AGAIN: 444 break; 445 } 446 447 /* 448 * If the operation is still going, leave the page busy to 449 * block all other accesses. Also, leave the paging in 450 * progress indicator set so that we don't attempt an object 451 * collapse. 452 */ 453 if (pageout_status[i] != VM_PAGER_PEND) { 454 vm_object_pip_wakeup(object); 455 vm_page_io_finish(mt); 456 if (!vm_page_count_severe() || !vm_page_try_to_cache(mt)) 457 vm_page_protect(mt, VM_PROT_READ); 458 } 459 } 460 return numpagedout; 461 } 462 463 #if !defined(NO_SWAPPING) 464 /* 465 * vm_pageout_object_deactivate_pages 466 * 467 * deactivate enough pages to satisfy the inactive target 468 * requirements or if vm_page_proc_limit is set, then 469 * deactivate all of the pages in the object and its 470 * backing_objects. 471 * 472 * The object and map must be locked. 473 */ 474 static int vm_pageout_object_deactivate_pages_callback(vm_page_t, void *); 475 476 static void 477 vm_pageout_object_deactivate_pages(vm_map_t map, vm_object_t object, 478 vm_pindex_t desired, int map_remove_only) 479 { 480 struct rb_vm_page_scan_info info; 481 int remove_mode; 482 483 if (object->type == OBJT_DEVICE || object->type == OBJT_PHYS) 484 return; 485 486 while (object) { 487 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 488 return; 489 if (object->paging_in_progress) 490 return; 491 492 remove_mode = map_remove_only; 493 if (object->shadow_count > 1) 494 remove_mode = 1; 495 496 /* 497 * scan the objects entire memory queue. spl protection is 498 * required to avoid an interrupt unbusy/free race against 499 * our busy check. 500 */ 501 crit_enter(); 502 info.limit = remove_mode; 503 info.map = map; 504 info.desired = desired; 505 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL, 506 vm_pageout_object_deactivate_pages_callback, 507 &info 508 ); 509 crit_exit(); 510 object = object->backing_object; 511 } 512 } 513 514 static int 515 vm_pageout_object_deactivate_pages_callback(vm_page_t p, void *data) 516 { 517 struct rb_vm_page_scan_info *info = data; 518 int actcount; 519 520 if (pmap_resident_count(vm_map_pmap(info->map)) <= info->desired) { 521 return(-1); 522 } 523 mycpu->gd_cnt.v_pdpages++; 524 if (p->wire_count != 0 || p->hold_count != 0 || p->busy != 0 || 525 (p->flags & (PG_BUSY|PG_UNMANAGED)) || 526 !pmap_page_exists_quick(vm_map_pmap(info->map), p)) { 527 return(0); 528 } 529 530 actcount = pmap_ts_referenced(p); 531 if (actcount) { 532 vm_page_flag_set(p, PG_REFERENCED); 533 } else if (p->flags & PG_REFERENCED) { 534 actcount = 1; 535 } 536 537 if ((p->queue != PQ_ACTIVE) && 538 (p->flags & PG_REFERENCED)) { 539 vm_page_activate(p); 540 p->act_count += actcount; 541 vm_page_flag_clear(p, PG_REFERENCED); 542 } else if (p->queue == PQ_ACTIVE) { 543 if ((p->flags & PG_REFERENCED) == 0) { 544 p->act_count -= min(p->act_count, ACT_DECLINE); 545 if (!info->limit && (vm_pageout_algorithm || (p->act_count == 0))) { 546 vm_page_protect(p, VM_PROT_NONE); 547 vm_page_deactivate(p); 548 } else { 549 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, p, pageq); 550 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, p, pageq); 551 } 552 } else { 553 vm_page_activate(p); 554 vm_page_flag_clear(p, PG_REFERENCED); 555 if (p->act_count < (ACT_MAX - ACT_ADVANCE)) 556 p->act_count += ACT_ADVANCE; 557 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, p, pageq); 558 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, p, pageq); 559 } 560 } else if (p->queue == PQ_INACTIVE) { 561 vm_page_protect(p, VM_PROT_NONE); 562 } 563 return(0); 564 } 565 566 /* 567 * deactivate some number of pages in a map, try to do it fairly, but 568 * that is really hard to do. 569 */ 570 static void 571 vm_pageout_map_deactivate_pages(vm_map_t map, vm_pindex_t desired) 572 { 573 vm_map_entry_t tmpe; 574 vm_object_t obj, bigobj; 575 int nothingwired; 576 577 if (lockmgr(&map->lock, LK_EXCLUSIVE | LK_NOWAIT)) { 578 return; 579 } 580 581 bigobj = NULL; 582 nothingwired = TRUE; 583 584 /* 585 * first, search out the biggest object, and try to free pages from 586 * that. 587 */ 588 tmpe = map->header.next; 589 while (tmpe != &map->header) { 590 switch(tmpe->maptype) { 591 case VM_MAPTYPE_NORMAL: 592 case VM_MAPTYPE_VPAGETABLE: 593 obj = tmpe->object.vm_object; 594 if ((obj != NULL) && (obj->shadow_count <= 1) && 595 ((bigobj == NULL) || 596 (bigobj->resident_page_count < obj->resident_page_count))) { 597 bigobj = obj; 598 } 599 break; 600 default: 601 break; 602 } 603 if (tmpe->wired_count > 0) 604 nothingwired = FALSE; 605 tmpe = tmpe->next; 606 } 607 608 if (bigobj) 609 vm_pageout_object_deactivate_pages(map, bigobj, desired, 0); 610 611 /* 612 * Next, hunt around for other pages to deactivate. We actually 613 * do this search sort of wrong -- .text first is not the best idea. 614 */ 615 tmpe = map->header.next; 616 while (tmpe != &map->header) { 617 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 618 break; 619 switch(tmpe->maptype) { 620 case VM_MAPTYPE_NORMAL: 621 case VM_MAPTYPE_VPAGETABLE: 622 obj = tmpe->object.vm_object; 623 if (obj) 624 vm_pageout_object_deactivate_pages(map, obj, desired, 0); 625 break; 626 default: 627 break; 628 } 629 tmpe = tmpe->next; 630 }; 631 632 /* 633 * Remove all mappings if a process is swapped out, this will free page 634 * table pages. 635 */ 636 if (desired == 0 && nothingwired) 637 pmap_remove(vm_map_pmap(map), 638 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS); 639 vm_map_unlock(map); 640 } 641 #endif 642 643 /* 644 * Don't try to be fancy - being fancy can lead to vnode deadlocks. We 645 * only do it for OBJT_DEFAULT and OBJT_SWAP objects which we know can 646 * be trivially freed. 647 */ 648 void 649 vm_pageout_page_free(vm_page_t m) { 650 vm_object_t object = m->object; 651 int type = object->type; 652 653 if (type == OBJT_SWAP || type == OBJT_DEFAULT) 654 vm_object_reference(object); 655 vm_page_busy(m); 656 vm_page_protect(m, VM_PROT_NONE); 657 vm_page_free(m); 658 if (type == OBJT_SWAP || type == OBJT_DEFAULT) 659 vm_object_deallocate(object); 660 } 661 662 /* 663 * vm_pageout_scan does the dirty work for the pageout daemon. 664 */ 665 666 struct vm_pageout_scan_info { 667 struct proc *bigproc; 668 vm_offset_t bigsize; 669 }; 670 671 static int vm_pageout_scan_callback(struct proc *p, void *data); 672 673 static void 674 vm_pageout_scan(int pass) 675 { 676 struct vm_pageout_scan_info info; 677 vm_page_t m, next; 678 struct vm_page marker; 679 int page_shortage, maxscan, pcount; 680 int addl_page_shortage, addl_page_shortage_init; 681 vm_object_t object; 682 int actcount; 683 int vnodes_skipped = 0; 684 int maxlaunder; 685 686 /* 687 * Do whatever cleanup that the pmap code can. 688 */ 689 pmap_collect(); 690 691 addl_page_shortage_init = vm_pageout_deficit; 692 vm_pageout_deficit = 0; 693 694 /* 695 * Calculate the number of pages we want to either free or move 696 * to the cache. 697 */ 698 page_shortage = vm_paging_target() + addl_page_shortage_init; 699 700 /* 701 * Initialize our marker 702 */ 703 bzero(&marker, sizeof(marker)); 704 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER; 705 marker.queue = PQ_INACTIVE; 706 marker.wire_count = 1; 707 708 /* 709 * Start scanning the inactive queue for pages we can move to the 710 * cache or free. The scan will stop when the target is reached or 711 * we have scanned the entire inactive queue. Note that m->act_count 712 * is not used to form decisions for the inactive queue, only for the 713 * active queue. 714 * 715 * maxlaunder limits the number of dirty pages we flush per scan. 716 * For most systems a smaller value (16 or 32) is more robust under 717 * extreme memory and disk pressure because any unnecessary writes 718 * to disk can result in extreme performance degredation. However, 719 * systems with excessive dirty pages (especially when MAP_NOSYNC is 720 * used) will die horribly with limited laundering. If the pageout 721 * daemon cannot clean enough pages in the first pass, we let it go 722 * all out in succeeding passes. 723 */ 724 if ((maxlaunder = vm_max_launder) <= 1) 725 maxlaunder = 1; 726 if (pass) 727 maxlaunder = 10000; 728 729 /* 730 * We will generally be in a critical section throughout the 731 * scan, but we can release it temporarily when we are sitting on a 732 * non-busy page without fear. this is required to prevent an 733 * interrupt from unbusying or freeing a page prior to our busy 734 * check, leaving us on the wrong queue or checking the wrong 735 * page. 736 */ 737 crit_enter(); 738 rescan0: 739 addl_page_shortage = addl_page_shortage_init; 740 maxscan = vmstats.v_inactive_count; 741 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl); 742 m != NULL && maxscan-- > 0 && page_shortage > 0; 743 m = next 744 ) { 745 mycpu->gd_cnt.v_pdpages++; 746 747 /* 748 * Give interrupts a chance 749 */ 750 crit_exit(); 751 crit_enter(); 752 753 /* 754 * It's easier for some of the conditions below to just loop 755 * and catch queue changes here rather then check everywhere 756 * else. 757 */ 758 if (m->queue != PQ_INACTIVE) 759 goto rescan0; 760 next = TAILQ_NEXT(m, pageq); 761 762 /* 763 * skip marker pages 764 */ 765 if (m->flags & PG_MARKER) 766 continue; 767 768 /* 769 * A held page may be undergoing I/O, so skip it. 770 */ 771 if (m->hold_count) { 772 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 773 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 774 addl_page_shortage++; 775 continue; 776 } 777 778 /* 779 * Dont mess with busy pages, keep in the front of the 780 * queue, most likely are being paged out. 781 */ 782 if (m->busy || (m->flags & PG_BUSY)) { 783 addl_page_shortage++; 784 continue; 785 } 786 787 if (m->object->ref_count == 0) { 788 /* 789 * If the object is not being used, we ignore previous 790 * references. 791 */ 792 vm_page_flag_clear(m, PG_REFERENCED); 793 pmap_clear_reference(m); 794 795 } else if (((m->flags & PG_REFERENCED) == 0) && 796 (actcount = pmap_ts_referenced(m))) { 797 /* 798 * Otherwise, if the page has been referenced while 799 * in the inactive queue, we bump the "activation 800 * count" upwards, making it less likely that the 801 * page will be added back to the inactive queue 802 * prematurely again. Here we check the page tables 803 * (or emulated bits, if any), given the upper level 804 * VM system not knowing anything about existing 805 * references. 806 */ 807 vm_page_activate(m); 808 m->act_count += (actcount + ACT_ADVANCE); 809 continue; 810 } 811 812 /* 813 * If the upper level VM system knows about any page 814 * references, we activate the page. We also set the 815 * "activation count" higher than normal so that we will less 816 * likely place pages back onto the inactive queue again. 817 */ 818 if ((m->flags & PG_REFERENCED) != 0) { 819 vm_page_flag_clear(m, PG_REFERENCED); 820 actcount = pmap_ts_referenced(m); 821 vm_page_activate(m); 822 m->act_count += (actcount + ACT_ADVANCE + 1); 823 continue; 824 } 825 826 /* 827 * If the upper level VM system doesn't know anything about 828 * the page being dirty, we have to check for it again. As 829 * far as the VM code knows, any partially dirty pages are 830 * fully dirty. 831 * 832 * Pages marked PG_WRITEABLE may be mapped into the user 833 * address space of a process running on another cpu. A 834 * user process (without holding the MP lock) running on 835 * another cpu may be able to touch the page while we are 836 * trying to remove it. To prevent this from occuring we 837 * must call pmap_remove_all() or otherwise make the page 838 * read-only. If the race occured pmap_remove_all() is 839 * responsible for setting m->dirty. 840 */ 841 if (m->dirty == 0) { 842 vm_page_test_dirty(m); 843 #if 0 844 if (m->dirty == 0 && (m->flags & PG_WRITEABLE) != 0) 845 pmap_remove_all(m); 846 #endif 847 } else { 848 vm_page_dirty(m); 849 } 850 851 if (m->valid == 0) { 852 /* 853 * Invalid pages can be easily freed 854 */ 855 vm_pageout_page_free(m); 856 mycpu->gd_cnt.v_dfree++; 857 --page_shortage; 858 } else if (m->dirty == 0) { 859 /* 860 * Clean pages can be placed onto the cache queue. 861 * This effectively frees them. 862 */ 863 vm_page_cache(m); 864 --page_shortage; 865 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) { 866 /* 867 * Dirty pages need to be paged out, but flushing 868 * a page is extremely expensive verses freeing 869 * a clean page. Rather then artificially limiting 870 * the number of pages we can flush, we instead give 871 * dirty pages extra priority on the inactive queue 872 * by forcing them to be cycled through the queue 873 * twice before being flushed, after which the 874 * (now clean) page will cycle through once more 875 * before being freed. This significantly extends 876 * the thrash point for a heavily loaded machine. 877 */ 878 vm_page_flag_set(m, PG_WINATCFLS); 879 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 880 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 881 } else if (maxlaunder > 0) { 882 /* 883 * We always want to try to flush some dirty pages if 884 * we encounter them, to keep the system stable. 885 * Normally this number is small, but under extreme 886 * pressure where there are insufficient clean pages 887 * on the inactive queue, we may have to go all out. 888 */ 889 int swap_pageouts_ok; 890 struct vnode *vp = NULL; 891 892 object = m->object; 893 894 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 895 swap_pageouts_ok = 1; 896 } else { 897 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 898 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 899 vm_page_count_min()); 900 901 } 902 903 /* 904 * We don't bother paging objects that are "dead". 905 * Those objects are in a "rundown" state. 906 */ 907 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 908 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 909 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 910 continue; 911 } 912 913 /* 914 * The object is already known NOT to be dead. It 915 * is possible for the vget() to block the whole 916 * pageout daemon, but the new low-memory handling 917 * code should prevent it. 918 * 919 * The previous code skipped locked vnodes and, worse, 920 * reordered pages in the queue. This results in 921 * completely non-deterministic operation because, 922 * quite often, a vm_fault has initiated an I/O and 923 * is holding a locked vnode at just the point where 924 * the pageout daemon is woken up. 925 * 926 * We can't wait forever for the vnode lock, we might 927 * deadlock due to a vn_read() getting stuck in 928 * vm_wait while holding this vnode. We skip the 929 * vnode if we can't get it in a reasonable amount 930 * of time. 931 */ 932 933 if (object->type == OBJT_VNODE) { 934 vp = object->handle; 935 936 if (vget(vp, LK_EXCLUSIVE|LK_NOOBJ|LK_TIMELOCK)) { 937 ++pageout_lock_miss; 938 if (object->flags & OBJ_MIGHTBEDIRTY) 939 vnodes_skipped++; 940 continue; 941 } 942 943 /* 944 * The page might have been moved to another 945 * queue during potential blocking in vget() 946 * above. The page might have been freed and 947 * reused for another vnode. The object might 948 * have been reused for another vnode. 949 */ 950 if (m->queue != PQ_INACTIVE || 951 m->object != object || 952 object->handle != vp) { 953 if (object->flags & OBJ_MIGHTBEDIRTY) 954 vnodes_skipped++; 955 vput(vp); 956 continue; 957 } 958 959 /* 960 * The page may have been busied during the 961 * blocking in vput(); We don't move the 962 * page back onto the end of the queue so that 963 * statistics are more correct if we don't. 964 */ 965 if (m->busy || (m->flags & PG_BUSY)) { 966 vput(vp); 967 continue; 968 } 969 970 /* 971 * If the page has become held it might 972 * be undergoing I/O, so skip it 973 */ 974 if (m->hold_count) { 975 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 976 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 977 if (object->flags & OBJ_MIGHTBEDIRTY) 978 vnodes_skipped++; 979 vput(vp); 980 continue; 981 } 982 } 983 984 /* 985 * If a page is dirty, then it is either being washed 986 * (but not yet cleaned) or it is still in the 987 * laundry. If it is still in the laundry, then we 988 * start the cleaning operation. 989 * 990 * This operation may cluster, invalidating the 'next' 991 * pointer. To prevent an inordinate number of 992 * restarts we use our marker to remember our place. 993 * 994 * decrement page_shortage on success to account for 995 * the (future) cleaned page. Otherwise we could wind 996 * up laundering or cleaning too many pages. 997 */ 998 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, m, &marker, pageq); 999 if (vm_pageout_clean(m) != 0) { 1000 --page_shortage; 1001 --maxlaunder; 1002 } 1003 next = TAILQ_NEXT(&marker, pageq); 1004 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, &marker, pageq); 1005 if (vp != NULL) 1006 vput(vp); 1007 } 1008 } 1009 1010 /* 1011 * Compute the number of pages we want to try to move from the 1012 * active queue to the inactive queue. 1013 */ 1014 page_shortage = vm_paging_target() + 1015 vmstats.v_inactive_target - vmstats.v_inactive_count; 1016 page_shortage += addl_page_shortage; 1017 1018 /* 1019 * Scan the active queue for things we can deactivate. We nominally 1020 * track the per-page activity counter and use it to locate 1021 * deactivation candidates. 1022 * 1023 * NOTE: we are still in a critical section. 1024 */ 1025 pcount = vmstats.v_active_count; 1026 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1027 1028 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) { 1029 /* 1030 * Give interrupts a chance. 1031 */ 1032 crit_exit(); 1033 crit_enter(); 1034 1035 /* 1036 * If the page was ripped out from under us, just stop. 1037 */ 1038 if (m->queue != PQ_ACTIVE) 1039 break; 1040 next = TAILQ_NEXT(m, pageq); 1041 1042 /* 1043 * Don't deactivate pages that are busy. 1044 */ 1045 if ((m->busy != 0) || 1046 (m->flags & PG_BUSY) || 1047 (m->hold_count != 0)) { 1048 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1049 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1050 m = next; 1051 continue; 1052 } 1053 1054 /* 1055 * The count for pagedaemon pages is done after checking the 1056 * page for eligibility... 1057 */ 1058 mycpu->gd_cnt.v_pdpages++; 1059 1060 /* 1061 * Check to see "how much" the page has been used. 1062 */ 1063 actcount = 0; 1064 if (m->object->ref_count != 0) { 1065 if (m->flags & PG_REFERENCED) { 1066 actcount += 1; 1067 } 1068 actcount += pmap_ts_referenced(m); 1069 if (actcount) { 1070 m->act_count += ACT_ADVANCE + actcount; 1071 if (m->act_count > ACT_MAX) 1072 m->act_count = ACT_MAX; 1073 } 1074 } 1075 1076 /* 1077 * Since we have "tested" this bit, we need to clear it now. 1078 */ 1079 vm_page_flag_clear(m, PG_REFERENCED); 1080 1081 /* 1082 * Only if an object is currently being used, do we use the 1083 * page activation count stats. 1084 */ 1085 if (actcount && (m->object->ref_count != 0)) { 1086 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1087 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1088 } else { 1089 m->act_count -= min(m->act_count, ACT_DECLINE); 1090 if (vm_pageout_algorithm || 1091 m->object->ref_count == 0 || 1092 m->act_count < pass) { 1093 page_shortage--; 1094 if (m->object->ref_count == 0) { 1095 vm_page_protect(m, VM_PROT_NONE); 1096 if (m->dirty == 0) 1097 vm_page_cache(m); 1098 else 1099 vm_page_deactivate(m); 1100 } else { 1101 vm_page_deactivate(m); 1102 } 1103 } else { 1104 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1105 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1106 } 1107 } 1108 m = next; 1109 } 1110 1111 /* 1112 * We try to maintain some *really* free pages, this allows interrupt 1113 * code to be guaranteed space. Since both cache and free queues 1114 * are considered basically 'free', moving pages from cache to free 1115 * does not effect other calculations. 1116 * 1117 * NOTE: we are still in a critical section. 1118 */ 1119 1120 while (vmstats.v_free_count < vmstats.v_free_reserved) { 1121 static int cache_rover = 0; 1122 m = vm_page_list_find(PQ_CACHE, cache_rover, FALSE); 1123 if (!m) 1124 break; 1125 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || 1126 m->busy || 1127 m->hold_count || 1128 m->wire_count) { 1129 #ifdef INVARIANTS 1130 kprintf("Warning: busy page %p found in cache\n", m); 1131 #endif 1132 vm_page_deactivate(m); 1133 continue; 1134 } 1135 cache_rover = (cache_rover + PQ_PRIME2) & PQ_L2_MASK; 1136 vm_pageout_page_free(m); 1137 mycpu->gd_cnt.v_dfree++; 1138 } 1139 1140 crit_exit(); 1141 1142 #if !defined(NO_SWAPPING) 1143 /* 1144 * Idle process swapout -- run once per second. 1145 */ 1146 if (vm_swap_idle_enabled) { 1147 static long lsec; 1148 if (time_second != lsec) { 1149 vm_pageout_req_swapout |= VM_SWAP_IDLE; 1150 vm_req_vmdaemon(); 1151 lsec = time_second; 1152 } 1153 } 1154 #endif 1155 1156 /* 1157 * If we didn't get enough free pages, and we have skipped a vnode 1158 * in a writeable object, wakeup the sync daemon. And kick swapout 1159 * if we did not get enough free pages. 1160 */ 1161 if (vm_paging_target() > 0) { 1162 if (vnodes_skipped && vm_page_count_min()) 1163 speedup_syncer(); 1164 #if !defined(NO_SWAPPING) 1165 if (vm_swap_enabled && vm_page_count_target()) { 1166 vm_req_vmdaemon(); 1167 vm_pageout_req_swapout |= VM_SWAP_NORMAL; 1168 } 1169 #endif 1170 } 1171 1172 /* 1173 * If we are out of swap and were not able to reach our paging 1174 * target, kill the largest process. 1175 */ 1176 if ((vm_swap_size < 64 && vm_page_count_min()) || 1177 (swap_pager_full && vm_paging_target() > 0)) { 1178 #if 0 1179 if ((vm_swap_size < 64 || swap_pager_full) && vm_page_count_min()) { 1180 #endif 1181 info.bigproc = NULL; 1182 info.bigsize = 0; 1183 allproc_scan(vm_pageout_scan_callback, &info); 1184 if (info.bigproc != NULL) { 1185 killproc(info.bigproc, "out of swap space"); 1186 info.bigproc->p_nice = PRIO_MIN; 1187 info.bigproc->p_usched->resetpriority( 1188 FIRST_LWP_IN_PROC(info.bigproc)); 1189 wakeup(&vmstats.v_free_count); 1190 PRELE(info.bigproc); 1191 } 1192 } 1193 } 1194 1195 static int 1196 vm_pageout_scan_callback(struct proc *p, void *data) 1197 { 1198 struct vm_pageout_scan_info *info = data; 1199 vm_offset_t size; 1200 1201 /* 1202 * if this is a system process, skip it 1203 */ 1204 if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) || 1205 ((p->p_pid < 48) && (vm_swap_size != 0))) { 1206 return (0); 1207 } 1208 1209 /* 1210 * if the process is in a non-running type state, 1211 * don't touch it. 1212 */ 1213 if (p->p_stat != SACTIVE && p->p_stat != SSTOP) { 1214 return (0); 1215 } 1216 1217 /* 1218 * get the process size 1219 */ 1220 size = vmspace_resident_count(p->p_vmspace) + 1221 vmspace_swap_count(p->p_vmspace); 1222 1223 /* 1224 * If the this process is bigger than the biggest one 1225 * remember it. 1226 */ 1227 if (size > info->bigsize) { 1228 if (info->bigproc) 1229 PRELE(info->bigproc); 1230 PHOLD(p); 1231 info->bigproc = p; 1232 info->bigsize = size; 1233 } 1234 return(0); 1235 } 1236 1237 /* 1238 * This routine tries to maintain the pseudo LRU active queue, 1239 * so that during long periods of time where there is no paging, 1240 * that some statistic accumulation still occurs. This code 1241 * helps the situation where paging just starts to occur. 1242 */ 1243 static void 1244 vm_pageout_page_stats(void) 1245 { 1246 vm_page_t m,next; 1247 int pcount,tpcount; /* Number of pages to check */ 1248 static int fullintervalcount = 0; 1249 int page_shortage; 1250 1251 page_shortage = 1252 (vmstats.v_inactive_target + vmstats.v_cache_max + vmstats.v_free_min) - 1253 (vmstats.v_free_count + vmstats.v_inactive_count + vmstats.v_cache_count); 1254 1255 if (page_shortage <= 0) 1256 return; 1257 1258 crit_enter(); 1259 1260 pcount = vmstats.v_active_count; 1261 fullintervalcount += vm_pageout_stats_interval; 1262 if (fullintervalcount < vm_pageout_full_stats_interval) { 1263 tpcount = (vm_pageout_stats_max * vmstats.v_active_count) / vmstats.v_page_count; 1264 if (pcount > tpcount) 1265 pcount = tpcount; 1266 } else { 1267 fullintervalcount = 0; 1268 } 1269 1270 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1271 while ((m != NULL) && (pcount-- > 0)) { 1272 int actcount; 1273 1274 if (m->queue != PQ_ACTIVE) { 1275 break; 1276 } 1277 1278 next = TAILQ_NEXT(m, pageq); 1279 /* 1280 * Don't deactivate pages that are busy. 1281 */ 1282 if ((m->busy != 0) || 1283 (m->flags & PG_BUSY) || 1284 (m->hold_count != 0)) { 1285 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1286 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1287 m = next; 1288 continue; 1289 } 1290 1291 actcount = 0; 1292 if (m->flags & PG_REFERENCED) { 1293 vm_page_flag_clear(m, PG_REFERENCED); 1294 actcount += 1; 1295 } 1296 1297 actcount += pmap_ts_referenced(m); 1298 if (actcount) { 1299 m->act_count += ACT_ADVANCE + actcount; 1300 if (m->act_count > ACT_MAX) 1301 m->act_count = ACT_MAX; 1302 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1303 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1304 } else { 1305 if (m->act_count == 0) { 1306 /* 1307 * We turn off page access, so that we have 1308 * more accurate RSS stats. We don't do this 1309 * in the normal page deactivation when the 1310 * system is loaded VM wise, because the 1311 * cost of the large number of page protect 1312 * operations would be higher than the value 1313 * of doing the operation. 1314 */ 1315 vm_page_protect(m, VM_PROT_NONE); 1316 vm_page_deactivate(m); 1317 } else { 1318 m->act_count -= min(m->act_count, ACT_DECLINE); 1319 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1320 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1321 } 1322 } 1323 1324 m = next; 1325 } 1326 crit_exit(); 1327 } 1328 1329 static int 1330 vm_pageout_free_page_calc(vm_size_t count) 1331 { 1332 if (count < vmstats.v_page_count) 1333 return 0; 1334 /* 1335 * free_reserved needs to include enough for the largest swap pager 1336 * structures plus enough for any pv_entry structs when paging. 1337 */ 1338 if (vmstats.v_page_count > 1024) 1339 vmstats.v_free_min = 4 + (vmstats.v_page_count - 1024) / 200; 1340 else 1341 vmstats.v_free_min = 4; 1342 vmstats.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1343 vmstats.v_interrupt_free_min; 1344 vmstats.v_free_reserved = vm_pageout_page_count + 1345 vmstats.v_pageout_free_min + (count / 768) + PQ_L2_SIZE; 1346 vmstats.v_free_severe = vmstats.v_free_min / 2; 1347 vmstats.v_free_min += vmstats.v_free_reserved; 1348 vmstats.v_free_severe += vmstats.v_free_reserved; 1349 return 1; 1350 } 1351 1352 1353 /* 1354 * vm_pageout is the high level pageout daemon. 1355 */ 1356 static void 1357 vm_pageout(void) 1358 { 1359 int pass; 1360 1361 /* 1362 * Initialize some paging parameters. 1363 */ 1364 1365 vmstats.v_interrupt_free_min = 2; 1366 if (vmstats.v_page_count < 2000) 1367 vm_pageout_page_count = 8; 1368 1369 vm_pageout_free_page_calc(vmstats.v_page_count); 1370 /* 1371 * v_free_target and v_cache_min control pageout hysteresis. Note 1372 * that these are more a measure of the VM cache queue hysteresis 1373 * then the VM free queue. Specifically, v_free_target is the 1374 * high water mark (free+cache pages). 1375 * 1376 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the 1377 * low water mark, while v_free_min is the stop. v_cache_min must 1378 * be big enough to handle memory needs while the pageout daemon 1379 * is signalled and run to free more pages. 1380 */ 1381 if (vmstats.v_free_count > 6144) 1382 vmstats.v_free_target = 4 * vmstats.v_free_min + vmstats.v_free_reserved; 1383 else 1384 vmstats.v_free_target = 2 * vmstats.v_free_min + vmstats.v_free_reserved; 1385 1386 if (vmstats.v_free_count > 2048) { 1387 vmstats.v_cache_min = vmstats.v_free_target; 1388 vmstats.v_cache_max = 2 * vmstats.v_cache_min; 1389 vmstats.v_inactive_target = (3 * vmstats.v_free_target) / 2; 1390 } else { 1391 vmstats.v_cache_min = 0; 1392 vmstats.v_cache_max = 0; 1393 vmstats.v_inactive_target = vmstats.v_free_count / 4; 1394 } 1395 if (vmstats.v_inactive_target > vmstats.v_free_count / 3) 1396 vmstats.v_inactive_target = vmstats.v_free_count / 3; 1397 1398 /* XXX does not really belong here */ 1399 if (vm_page_max_wired == 0) 1400 vm_page_max_wired = vmstats.v_free_count / 3; 1401 1402 if (vm_pageout_stats_max == 0) 1403 vm_pageout_stats_max = vmstats.v_free_target; 1404 1405 /* 1406 * Set interval in seconds for stats scan. 1407 */ 1408 if (vm_pageout_stats_interval == 0) 1409 vm_pageout_stats_interval = 5; 1410 if (vm_pageout_full_stats_interval == 0) 1411 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4; 1412 1413 1414 /* 1415 * Set maximum free per pass 1416 */ 1417 if (vm_pageout_stats_free_max == 0) 1418 vm_pageout_stats_free_max = 5; 1419 1420 swap_pager_swap_init(); 1421 pass = 0; 1422 /* 1423 * The pageout daemon is never done, so loop forever. 1424 */ 1425 while (TRUE) { 1426 int error; 1427 1428 /* 1429 * If we have enough free memory, wakeup waiters. Do 1430 * not clear vm_pages_needed until we reach our target, 1431 * otherwise we may be woken up over and over again and 1432 * waste a lot of cpu. 1433 */ 1434 crit_enter(); 1435 if (vm_pages_needed && !vm_page_count_min()) { 1436 if (vm_paging_needed() <= 0) 1437 vm_pages_needed = 0; 1438 wakeup(&vmstats.v_free_count); 1439 } 1440 if (vm_pages_needed) { 1441 /* 1442 * Still not done, take a second pass without waiting 1443 * (unlimited dirty cleaning), otherwise sleep a bit 1444 * and try again. 1445 */ 1446 ++pass; 1447 if (pass > 1) 1448 tsleep(&vm_pages_needed, 0, "psleep", hz/2); 1449 } else { 1450 /* 1451 * Good enough, sleep & handle stats. Prime the pass 1452 * for the next run. 1453 */ 1454 if (pass > 1) 1455 pass = 1; 1456 else 1457 pass = 0; 1458 error = tsleep(&vm_pages_needed, 1459 0, "psleep", vm_pageout_stats_interval * hz); 1460 if (error && !vm_pages_needed) { 1461 crit_exit(); 1462 pass = 0; 1463 vm_pageout_page_stats(); 1464 continue; 1465 } 1466 } 1467 1468 if (vm_pages_needed) 1469 mycpu->gd_cnt.v_pdwakeups++; 1470 crit_exit(); 1471 vm_pageout_scan(pass); 1472 vm_pageout_deficit = 0; 1473 } 1474 } 1475 1476 void 1477 pagedaemon_wakeup(void) 1478 { 1479 if (!vm_pages_needed && curthread != pagethread) { 1480 vm_pages_needed++; 1481 wakeup(&vm_pages_needed); 1482 } 1483 } 1484 1485 #if !defined(NO_SWAPPING) 1486 static void 1487 vm_req_vmdaemon(void) 1488 { 1489 static int lastrun = 0; 1490 1491 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1492 wakeup(&vm_daemon_needed); 1493 lastrun = ticks; 1494 } 1495 } 1496 1497 static int vm_daemon_callback(struct proc *p, void *data __unused); 1498 1499 static void 1500 vm_daemon(void) 1501 { 1502 while (TRUE) { 1503 tsleep(&vm_daemon_needed, 0, "psleep", 0); 1504 if (vm_pageout_req_swapout) { 1505 swapout_procs(vm_pageout_req_swapout); 1506 vm_pageout_req_swapout = 0; 1507 } 1508 /* 1509 * scan the processes for exceeding their rlimits or if 1510 * process is swapped out -- deactivate pages 1511 */ 1512 allproc_scan(vm_daemon_callback, NULL); 1513 } 1514 } 1515 1516 static int 1517 vm_daemon_callback(struct proc *p, void *data __unused) 1518 { 1519 vm_pindex_t limit, size; 1520 1521 /* 1522 * if this is a system process or if we have already 1523 * looked at this process, skip it. 1524 */ 1525 if (p->p_flag & (P_SYSTEM | P_WEXIT)) 1526 return (0); 1527 1528 /* 1529 * if the process is in a non-running type state, 1530 * don't touch it. 1531 */ 1532 if (p->p_stat != SACTIVE && p->p_stat != SSTOP) 1533 return (0); 1534 1535 /* 1536 * get a limit 1537 */ 1538 limit = OFF_TO_IDX(qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur, 1539 p->p_rlimit[RLIMIT_RSS].rlim_max)); 1540 1541 /* 1542 * let processes that are swapped out really be 1543 * swapped out. Set the limit to nothing to get as 1544 * many pages out to swap as possible. 1545 */ 1546 if (p->p_flag & P_SWAPPEDOUT) 1547 limit = 0; 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 return (0); 1555 } 1556 1557 #endif 1558