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 1.36 2008/07/01 02:02:56 dillon 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 int 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 vm_load to slow down faulting processes. 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 int 386 vm_pageout_flush(vm_page_t *mc, int count, int flags) 387 { 388 vm_object_t object; 389 int pageout_status[count]; 390 int numpagedout = 0; 391 int i; 392 393 /* 394 * Initiate I/O. Bump the vm_page_t->busy counter. 395 */ 396 for (i = 0; i < count; i++) { 397 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, ("vm_pageout_flush page %p index %d/%d: partially invalid page", mc[i], i, count)); 398 vm_page_io_start(mc[i]); 399 } 400 401 /* 402 * We must make the pages read-only. This will also force the 403 * modified bit in the related pmaps to be cleared. The pager 404 * cannot clear the bit for us since the I/O completion code 405 * typically runs from an interrupt. The act of making the page 406 * read-only handles the case for us. 407 */ 408 for (i = 0; i < count; i++) { 409 vm_page_protect(mc[i], VM_PROT_READ); 410 } 411 412 object = mc[0]->object; 413 vm_object_pip_add(object, count); 414 415 vm_pager_put_pages(object, mc, count, 416 (flags | ((object == &kernel_object) ? VM_PAGER_PUT_SYNC : 0)), 417 pageout_status); 418 419 for (i = 0; i < count; i++) { 420 vm_page_t mt = mc[i]; 421 422 switch (pageout_status[i]) { 423 case VM_PAGER_OK: 424 numpagedout++; 425 break; 426 case VM_PAGER_PEND: 427 numpagedout++; 428 break; 429 case VM_PAGER_BAD: 430 /* 431 * Page outside of range of object. Right now we 432 * essentially lose the changes by pretending it 433 * worked. 434 */ 435 pmap_clear_modify(mt); 436 vm_page_undirty(mt); 437 break; 438 case VM_PAGER_ERROR: 439 case VM_PAGER_FAIL: 440 /* 441 * A page typically cannot be paged out when we 442 * have run out of swap. We leave the page 443 * marked inactive and will try to page it out 444 * again later. 445 * 446 * Starvation of the active page list is used to 447 * determine when the system is massively memory 448 * starved. 449 */ 450 break; 451 case VM_PAGER_AGAIN: 452 break; 453 } 454 455 /* 456 * If the operation is still going, leave the page busy to 457 * block all other accesses. Also, leave the paging in 458 * progress indicator set so that we don't attempt an object 459 * collapse. 460 * 461 * For any pages which have completed synchronously, 462 * deactivate the page if we are under a severe deficit. 463 * Do not try to enter them into the cache, though, they 464 * might still be read-heavy. 465 */ 466 if (pageout_status[i] != VM_PAGER_PEND) { 467 vm_object_pip_wakeup(object); 468 vm_page_io_finish(mt); 469 if (vm_page_count_severe()) 470 vm_page_deactivate(mt); 471 #if 0 472 if (!vm_page_count_severe() || !vm_page_try_to_cache(mt)) 473 vm_page_protect(mt, VM_PROT_READ); 474 #endif 475 } 476 } 477 return numpagedout; 478 } 479 480 #if !defined(NO_SWAPPING) 481 /* 482 * vm_pageout_object_deactivate_pages 483 * 484 * deactivate enough pages to satisfy the inactive target 485 * requirements or if vm_page_proc_limit is set, then 486 * deactivate all of the pages in the object and its 487 * backing_objects. 488 * 489 * The object and map must be locked. 490 */ 491 static int vm_pageout_object_deactivate_pages_callback(vm_page_t, void *); 492 493 static void 494 vm_pageout_object_deactivate_pages(vm_map_t map, vm_object_t object, 495 vm_pindex_t desired, int map_remove_only) 496 { 497 struct rb_vm_page_scan_info info; 498 int remove_mode; 499 500 if (object->type == OBJT_DEVICE || object->type == OBJT_PHYS) 501 return; 502 503 while (object) { 504 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 505 return; 506 if (object->paging_in_progress) 507 return; 508 509 remove_mode = map_remove_only; 510 if (object->shadow_count > 1) 511 remove_mode = 1; 512 513 /* 514 * scan the objects entire memory queue. spl protection is 515 * required to avoid an interrupt unbusy/free race against 516 * our busy check. 517 */ 518 crit_enter(); 519 info.limit = remove_mode; 520 info.map = map; 521 info.desired = desired; 522 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL, 523 vm_pageout_object_deactivate_pages_callback, 524 &info 525 ); 526 crit_exit(); 527 object = object->backing_object; 528 } 529 } 530 531 static int 532 vm_pageout_object_deactivate_pages_callback(vm_page_t p, void *data) 533 { 534 struct rb_vm_page_scan_info *info = data; 535 int actcount; 536 537 if (pmap_resident_count(vm_map_pmap(info->map)) <= info->desired) { 538 return(-1); 539 } 540 mycpu->gd_cnt.v_pdpages++; 541 if (p->wire_count != 0 || p->hold_count != 0 || p->busy != 0 || 542 (p->flags & (PG_BUSY|PG_UNMANAGED)) || 543 !pmap_page_exists_quick(vm_map_pmap(info->map), p)) { 544 return(0); 545 } 546 547 actcount = pmap_ts_referenced(p); 548 if (actcount) { 549 vm_page_flag_set(p, PG_REFERENCED); 550 } else if (p->flags & PG_REFERENCED) { 551 actcount = 1; 552 } 553 554 if ((p->queue != PQ_ACTIVE) && 555 (p->flags & PG_REFERENCED)) { 556 vm_page_activate(p); 557 p->act_count += actcount; 558 vm_page_flag_clear(p, PG_REFERENCED); 559 } else if (p->queue == PQ_ACTIVE) { 560 if ((p->flags & PG_REFERENCED) == 0) { 561 p->act_count -= min(p->act_count, ACT_DECLINE); 562 if (!info->limit && (vm_pageout_algorithm || (p->act_count == 0))) { 563 vm_page_busy(p); 564 vm_page_protect(p, VM_PROT_NONE); 565 vm_page_wakeup(p); 566 vm_page_deactivate(p); 567 } else { 568 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, p, pageq); 569 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, p, pageq); 570 } 571 } else { 572 vm_page_activate(p); 573 vm_page_flag_clear(p, PG_REFERENCED); 574 if (p->act_count < (ACT_MAX - ACT_ADVANCE)) 575 p->act_count += ACT_ADVANCE; 576 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, p, pageq); 577 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, p, pageq); 578 } 579 } else if (p->queue == PQ_INACTIVE) { 580 vm_page_busy(p); 581 vm_page_protect(p, VM_PROT_NONE); 582 vm_page_wakeup(p); 583 } 584 return(0); 585 } 586 587 /* 588 * deactivate some number of pages in a map, try to do it fairly, but 589 * that is really hard to do. 590 */ 591 static void 592 vm_pageout_map_deactivate_pages(vm_map_t map, vm_pindex_t desired) 593 { 594 vm_map_entry_t tmpe; 595 vm_object_t obj, bigobj; 596 int nothingwired; 597 598 if (lockmgr(&map->lock, LK_EXCLUSIVE | LK_NOWAIT)) { 599 return; 600 } 601 602 bigobj = NULL; 603 nothingwired = TRUE; 604 605 /* 606 * first, search out the biggest object, and try to free pages from 607 * that. 608 */ 609 tmpe = map->header.next; 610 while (tmpe != &map->header) { 611 switch(tmpe->maptype) { 612 case VM_MAPTYPE_NORMAL: 613 case VM_MAPTYPE_VPAGETABLE: 614 obj = tmpe->object.vm_object; 615 if ((obj != NULL) && (obj->shadow_count <= 1) && 616 ((bigobj == NULL) || 617 (bigobj->resident_page_count < obj->resident_page_count))) { 618 bigobj = obj; 619 } 620 break; 621 default: 622 break; 623 } 624 if (tmpe->wired_count > 0) 625 nothingwired = FALSE; 626 tmpe = tmpe->next; 627 } 628 629 if (bigobj) 630 vm_pageout_object_deactivate_pages(map, bigobj, desired, 0); 631 632 /* 633 * Next, hunt around for other pages to deactivate. We actually 634 * do this search sort of wrong -- .text first is not the best idea. 635 */ 636 tmpe = map->header.next; 637 while (tmpe != &map->header) { 638 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 639 break; 640 switch(tmpe->maptype) { 641 case VM_MAPTYPE_NORMAL: 642 case VM_MAPTYPE_VPAGETABLE: 643 obj = tmpe->object.vm_object; 644 if (obj) 645 vm_pageout_object_deactivate_pages(map, obj, desired, 0); 646 break; 647 default: 648 break; 649 } 650 tmpe = tmpe->next; 651 }; 652 653 /* 654 * Remove all mappings if a process is swapped out, this will free page 655 * table pages. 656 */ 657 if (desired == 0 && nothingwired) 658 pmap_remove(vm_map_pmap(map), 659 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS); 660 vm_map_unlock(map); 661 } 662 #endif 663 664 /* 665 * Don't try to be fancy - being fancy can lead to vnode deadlocks. We 666 * only do it for OBJT_DEFAULT and OBJT_SWAP objects which we know can 667 * be trivially freed. 668 */ 669 void 670 vm_pageout_page_free(vm_page_t m) 671 { 672 vm_object_t object = m->object; 673 int type = object->type; 674 675 if (type == OBJT_SWAP || type == OBJT_DEFAULT) 676 vm_object_reference(object); 677 vm_page_busy(m); 678 vm_page_protect(m, VM_PROT_NONE); 679 vm_page_free(m); 680 if (type == OBJT_SWAP || type == OBJT_DEFAULT) 681 vm_object_deallocate(object); 682 } 683 684 /* 685 * vm_pageout_scan does the dirty work for the pageout daemon. 686 */ 687 struct vm_pageout_scan_info { 688 struct proc *bigproc; 689 vm_offset_t bigsize; 690 }; 691 692 static int vm_pageout_scan_callback(struct proc *p, void *data); 693 694 static int 695 vm_pageout_scan(int pass) 696 { 697 struct vm_pageout_scan_info info; 698 vm_page_t m, next; 699 struct vm_page marker; 700 int maxscan, pcount; 701 int recycle_count; 702 int inactive_shortage, active_shortage; 703 vm_object_t object; 704 int actcount; 705 int vnodes_skipped = 0; 706 int maxlaunder; 707 708 /* 709 * Do whatever cleanup that the pmap code can. 710 */ 711 pmap_collect(); 712 713 /* 714 * Calculate our target for the number of free+cache pages we 715 * want to get to. This is higher then the number that causes 716 * allocations to stall (severe) in order to provide hysteresis, 717 * and if we don't make it all the way but get to the minimum 718 * we're happy. 719 */ 720 inactive_shortage = vm_paging_target() + vm_pageout_deficit; 721 vm_pageout_deficit = 0; 722 723 /* 724 * Initialize our marker 725 */ 726 bzero(&marker, sizeof(marker)); 727 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER; 728 marker.queue = PQ_INACTIVE; 729 marker.wire_count = 1; 730 731 /* 732 * Start scanning the inactive queue for pages we can move to the 733 * cache or free. The scan will stop when the target is reached or 734 * we have scanned the entire inactive queue. Note that m->act_count 735 * is not used to form decisions for the inactive queue, only for the 736 * active queue. 737 * 738 * maxlaunder limits the number of dirty pages we flush per scan. 739 * For most systems a smaller value (16 or 32) is more robust under 740 * extreme memory and disk pressure because any unnecessary writes 741 * to disk can result in extreme performance degredation. However, 742 * systems with excessive dirty pages (especially when MAP_NOSYNC is 743 * used) will die horribly with limited laundering. If the pageout 744 * daemon cannot clean enough pages in the first pass, we let it go 745 * all out in succeeding passes. 746 */ 747 if ((maxlaunder = vm_max_launder) <= 1) 748 maxlaunder = 1; 749 if (pass) 750 maxlaunder = 10000; 751 752 /* 753 * We will generally be in a critical section throughout the 754 * scan, but we can release it temporarily when we are sitting on a 755 * non-busy page without fear. this is required to prevent an 756 * interrupt from unbusying or freeing a page prior to our busy 757 * check, leaving us on the wrong queue or checking the wrong 758 * page. 759 */ 760 crit_enter(); 761 rescan0: 762 maxscan = vmstats.v_inactive_count; 763 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl); 764 m != NULL && maxscan-- > 0 && inactive_shortage > 0; 765 m = next 766 ) { 767 mycpu->gd_cnt.v_pdpages++; 768 769 /* 770 * Give interrupts a chance 771 */ 772 crit_exit(); 773 crit_enter(); 774 775 /* 776 * It's easier for some of the conditions below to just loop 777 * and catch queue changes here rather then check everywhere 778 * else. 779 */ 780 if (m->queue != PQ_INACTIVE) 781 goto rescan0; 782 next = TAILQ_NEXT(m, pageq); 783 784 /* 785 * skip marker pages 786 */ 787 if (m->flags & PG_MARKER) 788 continue; 789 790 /* 791 * A held page may be undergoing I/O, so skip it. 792 */ 793 if (m->hold_count) { 794 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 795 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 796 continue; 797 } 798 799 /* 800 * Dont mess with busy pages, keep in the front of the 801 * queue, most likely are being paged out. 802 */ 803 if (m->busy || (m->flags & PG_BUSY)) { 804 continue; 805 } 806 807 if (m->object->ref_count == 0) { 808 /* 809 * If the object is not being used, we ignore previous 810 * references. 811 */ 812 vm_page_flag_clear(m, PG_REFERENCED); 813 pmap_clear_reference(m); 814 815 } else if (((m->flags & PG_REFERENCED) == 0) && 816 (actcount = pmap_ts_referenced(m))) { 817 /* 818 * Otherwise, if the page has been referenced while 819 * in the inactive queue, we bump the "activation 820 * count" upwards, making it less likely that the 821 * page will be added back to the inactive queue 822 * prematurely again. Here we check the page tables 823 * (or emulated bits, if any), given the upper level 824 * VM system not knowing anything about existing 825 * references. 826 */ 827 vm_page_activate(m); 828 m->act_count += (actcount + ACT_ADVANCE); 829 continue; 830 } 831 832 /* 833 * If the upper level VM system knows about any page 834 * references, we activate the page. We also set the 835 * "activation count" higher than normal so that we will less 836 * likely place pages back onto the inactive queue again. 837 */ 838 if ((m->flags & PG_REFERENCED) != 0) { 839 vm_page_flag_clear(m, PG_REFERENCED); 840 actcount = pmap_ts_referenced(m); 841 vm_page_activate(m); 842 m->act_count += (actcount + ACT_ADVANCE + 1); 843 continue; 844 } 845 846 /* 847 * If the upper level VM system doesn't know anything about 848 * the page being dirty, we have to check for it again. As 849 * far as the VM code knows, any partially dirty pages are 850 * fully dirty. 851 * 852 * Pages marked PG_WRITEABLE may be mapped into the user 853 * address space of a process running on another cpu. A 854 * user process (without holding the MP lock) running on 855 * another cpu may be able to touch the page while we are 856 * trying to remove it. vm_page_cache() will handle this 857 * case for us. 858 */ 859 if (m->dirty == 0) { 860 vm_page_test_dirty(m); 861 } else { 862 vm_page_dirty(m); 863 } 864 865 if (m->valid == 0) { 866 /* 867 * Invalid pages can be easily freed 868 */ 869 vm_pageout_page_free(m); 870 mycpu->gd_cnt.v_dfree++; 871 --inactive_shortage; 872 } else if (m->dirty == 0) { 873 /* 874 * Clean pages can be placed onto the cache queue. 875 * This effectively frees them. 876 */ 877 vm_page_cache(m); 878 --inactive_shortage; 879 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) { 880 /* 881 * Dirty pages need to be paged out, but flushing 882 * a page is extremely expensive verses freeing 883 * a clean page. Rather then artificially limiting 884 * the number of pages we can flush, we instead give 885 * dirty pages extra priority on the inactive queue 886 * by forcing them to be cycled through the queue 887 * twice before being flushed, after which the 888 * (now clean) page will cycle through once more 889 * before being freed. This significantly extends 890 * the thrash point for a heavily loaded machine. 891 */ 892 vm_page_flag_set(m, PG_WINATCFLS); 893 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 894 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 895 } else if (maxlaunder > 0) { 896 /* 897 * We always want to try to flush some dirty pages if 898 * we encounter them, to keep the system stable. 899 * Normally this number is small, but under extreme 900 * pressure where there are insufficient clean pages 901 * on the inactive queue, we may have to go all out. 902 */ 903 int swap_pageouts_ok; 904 struct vnode *vp = NULL; 905 906 object = m->object; 907 908 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 909 swap_pageouts_ok = 1; 910 } else { 911 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 912 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 913 vm_page_count_min(0)); 914 915 } 916 917 /* 918 * We don't bother paging objects that are "dead". 919 * Those objects are in a "rundown" state. 920 */ 921 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 922 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 923 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 924 continue; 925 } 926 927 /* 928 * The object is already known NOT to be dead. It 929 * is possible for the vget() to block the whole 930 * pageout daemon, but the new low-memory handling 931 * code should prevent it. 932 * 933 * The previous code skipped locked vnodes and, worse, 934 * reordered pages in the queue. This results in 935 * completely non-deterministic operation because, 936 * quite often, a vm_fault has initiated an I/O and 937 * is holding a locked vnode at just the point where 938 * the pageout daemon is woken up. 939 * 940 * We can't wait forever for the vnode lock, we might 941 * deadlock due to a vn_read() getting stuck in 942 * vm_wait while holding this vnode. We skip the 943 * vnode if we can't get it in a reasonable amount 944 * of time. 945 */ 946 947 if (object->type == OBJT_VNODE) { 948 vp = object->handle; 949 950 if (vget(vp, LK_EXCLUSIVE|LK_NOOBJ|LK_TIMELOCK)) { 951 ++pageout_lock_miss; 952 if (object->flags & OBJ_MIGHTBEDIRTY) 953 vnodes_skipped++; 954 continue; 955 } 956 957 /* 958 * The page might have been moved to another 959 * queue during potential blocking in vget() 960 * above. The page might have been freed and 961 * reused for another vnode. The object might 962 * have been reused for another vnode. 963 */ 964 if (m->queue != PQ_INACTIVE || 965 m->object != object || 966 object->handle != vp) { 967 if (object->flags & OBJ_MIGHTBEDIRTY) 968 vnodes_skipped++; 969 vput(vp); 970 continue; 971 } 972 973 /* 974 * The page may have been busied during the 975 * blocking in vput(); We don't move the 976 * page back onto the end of the queue so that 977 * statistics are more correct if we don't. 978 */ 979 if (m->busy || (m->flags & PG_BUSY)) { 980 vput(vp); 981 continue; 982 } 983 984 /* 985 * If the page has become held it might 986 * be undergoing I/O, so skip it 987 */ 988 if (m->hold_count) { 989 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 990 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 991 if (object->flags & OBJ_MIGHTBEDIRTY) 992 vnodes_skipped++; 993 vput(vp); 994 continue; 995 } 996 } 997 998 /* 999 * If a page is dirty, then it is either being washed 1000 * (but not yet cleaned) or it is still in the 1001 * laundry. If it is still in the laundry, then we 1002 * start the cleaning operation. 1003 * 1004 * This operation may cluster, invalidating the 'next' 1005 * pointer. To prevent an inordinate number of 1006 * restarts we use our marker to remember our place. 1007 * 1008 * decrement inactive_shortage on success to account 1009 * for the (future) cleaned page. Otherwise we 1010 * could wind up laundering or cleaning too many 1011 * pages. 1012 */ 1013 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, m, &marker, pageq); 1014 if (vm_pageout_clean(m) != 0) { 1015 --inactive_shortage; 1016 --maxlaunder; 1017 } 1018 next = TAILQ_NEXT(&marker, pageq); 1019 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, &marker, pageq); 1020 if (vp != NULL) 1021 vput(vp); 1022 } 1023 } 1024 1025 /* 1026 * We want to move pages from the active queue to the inactive 1027 * queue to get the inactive queue to the inactive target. If 1028 * we still have a page shortage from above we try to directly free 1029 * clean pages instead of moving them. 1030 * 1031 * If we do still have a shortage we keep track of the number of 1032 * pages we free or cache (recycle_count) as a measure of thrashing 1033 * between the active and inactive queues. 1034 * 1035 * We do not do this if we were able to satisfy the requirement 1036 * entirely from the inactive queue. 1037 * 1038 * NOTE: Both variables can end up negative. 1039 * NOTE: We are still in a critical section. 1040 */ 1041 active_shortage = vmstats.v_inactive_target - vmstats.v_inactive_count; 1042 if (inactive_shortage <= 0) 1043 active_shortage = 0; 1044 1045 pcount = vmstats.v_active_count; 1046 recycle_count = 0; 1047 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1048 1049 while ((m != NULL) && (pcount-- > 0) && 1050 (inactive_shortage > 0 || active_shortage > 0) 1051 ) { 1052 /* 1053 * Give interrupts a chance. 1054 */ 1055 crit_exit(); 1056 crit_enter(); 1057 1058 /* 1059 * If the page was ripped out from under us, just stop. 1060 */ 1061 if (m->queue != PQ_ACTIVE) 1062 break; 1063 next = TAILQ_NEXT(m, pageq); 1064 1065 /* 1066 * Don't deactivate pages that are busy. 1067 */ 1068 if ((m->busy != 0) || 1069 (m->flags & PG_BUSY) || 1070 (m->hold_count != 0)) { 1071 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1072 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1073 m = next; 1074 continue; 1075 } 1076 1077 /* 1078 * The count for pagedaemon pages is done after checking the 1079 * page for eligibility... 1080 */ 1081 mycpu->gd_cnt.v_pdpages++; 1082 1083 /* 1084 * Check to see "how much" the page has been used and clear 1085 * the tracking access bits. If the object has no references 1086 * don't bother paying the expense. 1087 */ 1088 actcount = 0; 1089 if (m->object->ref_count != 0) { 1090 if (m->flags & PG_REFERENCED) 1091 ++actcount; 1092 actcount += pmap_ts_referenced(m); 1093 if (actcount) { 1094 m->act_count += ACT_ADVANCE + actcount; 1095 if (m->act_count > ACT_MAX) 1096 m->act_count = ACT_MAX; 1097 } 1098 } 1099 vm_page_flag_clear(m, PG_REFERENCED); 1100 1101 /* 1102 * actcount is only valid if the object ref_count is non-zero. 1103 */ 1104 if (actcount && m->object->ref_count != 0) { 1105 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1106 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1107 } else { 1108 m->act_count -= min(m->act_count, ACT_DECLINE); 1109 if (vm_pageout_algorithm || 1110 m->object->ref_count == 0 || 1111 m->act_count < pass + 1 1112 ) { 1113 /* 1114 * Deactivate the page. If we had a 1115 * shortage from our inactive scan try to 1116 * free (cache) the page instead. 1117 */ 1118 --active_shortage; 1119 if (inactive_shortage > 0 || 1120 m->object->ref_count == 0) { 1121 if (inactive_shortage > 0) 1122 ++recycle_count; 1123 vm_page_busy(m); 1124 vm_page_protect(m, VM_PROT_NONE); 1125 vm_page_wakeup(m); 1126 if (m->dirty == 0) { 1127 --inactive_shortage; 1128 vm_page_cache(m); 1129 } else { 1130 vm_page_deactivate(m); 1131 } 1132 } else { 1133 vm_page_deactivate(m); 1134 } 1135 } else { 1136 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1137 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1138 } 1139 } 1140 m = next; 1141 } 1142 1143 /* 1144 * We try to maintain some *really* free pages, this allows interrupt 1145 * code to be guaranteed space. Since both cache and free queues 1146 * are considered basically 'free', moving pages from cache to free 1147 * does not effect other calculations. 1148 * 1149 * NOTE: we are still in a critical section. 1150 * 1151 * Pages moved from PQ_CACHE to totally free are not counted in the 1152 * pages_freed counter. 1153 */ 1154 while (vmstats.v_free_count < vmstats.v_free_reserved) { 1155 static int cache_rover = 0; 1156 m = vm_page_list_find(PQ_CACHE, cache_rover, FALSE); 1157 if (m == NULL) 1158 break; 1159 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || 1160 m->busy || 1161 m->hold_count || 1162 m->wire_count) { 1163 #ifdef INVARIANTS 1164 kprintf("Warning: busy page %p found in cache\n", m); 1165 #endif 1166 vm_page_deactivate(m); 1167 continue; 1168 } 1169 KKASSERT((m->flags & PG_MAPPED) == 0); 1170 KKASSERT(m->dirty == 0); 1171 cache_rover = (cache_rover + PQ_PRIME2) & PQ_L2_MASK; 1172 vm_pageout_page_free(m); 1173 mycpu->gd_cnt.v_dfree++; 1174 } 1175 1176 crit_exit(); 1177 1178 #if !defined(NO_SWAPPING) 1179 /* 1180 * Idle process swapout -- run once per second. 1181 */ 1182 if (vm_swap_idle_enabled) { 1183 static long lsec; 1184 if (time_second != lsec) { 1185 vm_pageout_req_swapout |= VM_SWAP_IDLE; 1186 vm_req_vmdaemon(); 1187 lsec = time_second; 1188 } 1189 } 1190 #endif 1191 1192 /* 1193 * If we didn't get enough free pages, and we have skipped a vnode 1194 * in a writeable object, wakeup the sync daemon. And kick swapout 1195 * if we did not get enough free pages. 1196 */ 1197 if (vm_paging_target() > 0) { 1198 if (vnodes_skipped && vm_page_count_min(0)) 1199 speedup_syncer(); 1200 #if !defined(NO_SWAPPING) 1201 if (vm_swap_enabled && vm_page_count_target()) { 1202 vm_req_vmdaemon(); 1203 vm_pageout_req_swapout |= VM_SWAP_NORMAL; 1204 } 1205 #endif 1206 } 1207 1208 /* 1209 * Handle catastrophic conditions. Under good conditions we should 1210 * be at the target, well beyond our minimum. If we could not even 1211 * reach our minimum the system is under heavy stress. 1212 * 1213 * Determine whether we have run out of memory. This occurs when 1214 * swap_pager_full is TRUE and the only pages left in the page 1215 * queues are dirty. We will still likely have page shortages. 1216 * 1217 * - swap_pager_full is set if insufficient swap was 1218 * available to satisfy a requested pageout. 1219 * 1220 * - the inactive queue is bloated (4 x size of active queue), 1221 * meaning it is unable to get rid of dirty pages and. 1222 * 1223 * - vm_page_count_min() without counting pages recycled from the 1224 * active queue (recycle_count) means we could not recover 1225 * enough pages to meet bare minimum needs. This test only 1226 * works if the inactive queue is bloated. 1227 * 1228 * - due to a positive inactive_shortage we shifted the remaining 1229 * dirty pages from the active queue to the inactive queue 1230 * trying to find clean ones to free. 1231 */ 1232 if (swap_pager_full && vm_page_count_min(recycle_count)) 1233 kprintf("Warning: system low on memory+swap!\n"); 1234 if (swap_pager_full && vm_page_count_min(recycle_count) && 1235 vmstats.v_inactive_count > vmstats.v_active_count * 4 && 1236 inactive_shortage > 0) { 1237 /* 1238 * Kill something. 1239 */ 1240 info.bigproc = NULL; 1241 info.bigsize = 0; 1242 allproc_scan(vm_pageout_scan_callback, &info); 1243 if (info.bigproc != NULL) { 1244 killproc(info.bigproc, "out of swap space"); 1245 info.bigproc->p_nice = PRIO_MIN; 1246 info.bigproc->p_usched->resetpriority( 1247 FIRST_LWP_IN_PROC(info.bigproc)); 1248 wakeup(&vmstats.v_free_count); 1249 PRELE(info.bigproc); 1250 } 1251 } 1252 return(inactive_shortage); 1253 } 1254 1255 static int 1256 vm_pageout_scan_callback(struct proc *p, void *data) 1257 { 1258 struct vm_pageout_scan_info *info = data; 1259 vm_offset_t size; 1260 1261 /* 1262 * Never kill system processes or init. If we have configured swap 1263 * then try to avoid killing low-numbered pids. 1264 */ 1265 if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) || 1266 ((p->p_pid < 48) && (vm_swap_size != 0))) { 1267 return (0); 1268 } 1269 1270 /* 1271 * if the process is in a non-running type state, 1272 * don't touch it. 1273 */ 1274 if (p->p_stat != SACTIVE && p->p_stat != SSTOP) 1275 return (0); 1276 1277 /* 1278 * Get the approximate process size. Note that anonymous pages 1279 * with backing swap will be counted twice, but there should not 1280 * be too many such pages due to the stress the VM system is 1281 * under at this point. 1282 */ 1283 size = vmspace_anonymous_count(p->p_vmspace) + 1284 vmspace_swap_count(p->p_vmspace); 1285 1286 /* 1287 * If the this process is bigger than the biggest one 1288 * remember it. 1289 */ 1290 if (info->bigsize < size) { 1291 if (info->bigproc) 1292 PRELE(info->bigproc); 1293 PHOLD(p); 1294 info->bigproc = p; 1295 info->bigsize = size; 1296 } 1297 return(0); 1298 } 1299 1300 /* 1301 * This routine tries to maintain the pseudo LRU active queue, 1302 * so that during long periods of time where there is no paging, 1303 * that some statistic accumulation still occurs. This code 1304 * helps the situation where paging just starts to occur. 1305 */ 1306 static void 1307 vm_pageout_page_stats(void) 1308 { 1309 vm_page_t m,next; 1310 int pcount,tpcount; /* Number of pages to check */ 1311 static int fullintervalcount = 0; 1312 int page_shortage; 1313 1314 page_shortage = 1315 (vmstats.v_inactive_target + vmstats.v_cache_max + vmstats.v_free_min) - 1316 (vmstats.v_free_count + vmstats.v_inactive_count + vmstats.v_cache_count); 1317 1318 if (page_shortage <= 0) 1319 return; 1320 1321 crit_enter(); 1322 1323 pcount = vmstats.v_active_count; 1324 fullintervalcount += vm_pageout_stats_interval; 1325 if (fullintervalcount < vm_pageout_full_stats_interval) { 1326 tpcount = (vm_pageout_stats_max * vmstats.v_active_count) / vmstats.v_page_count; 1327 if (pcount > tpcount) 1328 pcount = tpcount; 1329 } else { 1330 fullintervalcount = 0; 1331 } 1332 1333 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1334 while ((m != NULL) && (pcount-- > 0)) { 1335 int actcount; 1336 1337 if (m->queue != PQ_ACTIVE) { 1338 break; 1339 } 1340 1341 next = TAILQ_NEXT(m, pageq); 1342 /* 1343 * Don't deactivate pages that are busy. 1344 */ 1345 if ((m->busy != 0) || 1346 (m->flags & PG_BUSY) || 1347 (m->hold_count != 0)) { 1348 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1349 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1350 m = next; 1351 continue; 1352 } 1353 1354 actcount = 0; 1355 if (m->flags & PG_REFERENCED) { 1356 vm_page_flag_clear(m, PG_REFERENCED); 1357 actcount += 1; 1358 } 1359 1360 actcount += pmap_ts_referenced(m); 1361 if (actcount) { 1362 m->act_count += ACT_ADVANCE + actcount; 1363 if (m->act_count > ACT_MAX) 1364 m->act_count = ACT_MAX; 1365 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1366 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1367 } else { 1368 if (m->act_count == 0) { 1369 /* 1370 * We turn off page access, so that we have 1371 * more accurate RSS stats. We don't do this 1372 * in the normal page deactivation when the 1373 * system is loaded VM wise, because the 1374 * cost of the large number of page protect 1375 * operations would be higher than the value 1376 * of doing the operation. 1377 */ 1378 vm_page_busy(m); 1379 vm_page_protect(m, VM_PROT_NONE); 1380 vm_page_wakeup(m); 1381 vm_page_deactivate(m); 1382 } else { 1383 m->act_count -= min(m->act_count, ACT_DECLINE); 1384 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1385 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1386 } 1387 } 1388 1389 m = next; 1390 } 1391 crit_exit(); 1392 } 1393 1394 static int 1395 vm_pageout_free_page_calc(vm_size_t count) 1396 { 1397 if (count < vmstats.v_page_count) 1398 return 0; 1399 /* 1400 * free_reserved needs to include enough for the largest swap pager 1401 * structures plus enough for any pv_entry structs when paging. 1402 */ 1403 if (vmstats.v_page_count > 1024) 1404 vmstats.v_free_min = 4 + (vmstats.v_page_count - 1024) / 200; 1405 else 1406 vmstats.v_free_min = 4; 1407 vmstats.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1408 vmstats.v_interrupt_free_min; 1409 vmstats.v_free_reserved = vm_pageout_page_count + 1410 vmstats.v_pageout_free_min + (count / 768) + PQ_L2_SIZE; 1411 vmstats.v_free_severe = vmstats.v_free_min / 2; 1412 vmstats.v_free_min += vmstats.v_free_reserved; 1413 vmstats.v_free_severe += vmstats.v_free_reserved; 1414 return 1; 1415 } 1416 1417 1418 /* 1419 * vm_pageout is the high level pageout daemon. 1420 */ 1421 static void 1422 vm_pageout(void) 1423 { 1424 int pass; 1425 int inactive_shortage; 1426 1427 /* 1428 * Initialize some paging parameters. 1429 */ 1430 curthread->td_flags |= TDF_SYSTHREAD; 1431 1432 vmstats.v_interrupt_free_min = 2; 1433 if (vmstats.v_page_count < 2000) 1434 vm_pageout_page_count = 8; 1435 1436 vm_pageout_free_page_calc(vmstats.v_page_count); 1437 1438 /* 1439 * v_free_target and v_cache_min control pageout hysteresis. Note 1440 * that these are more a measure of the VM cache queue hysteresis 1441 * then the VM free queue. Specifically, v_free_target is the 1442 * high water mark (free+cache pages). 1443 * 1444 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the 1445 * low water mark, while v_free_min is the stop. v_cache_min must 1446 * be big enough to handle memory needs while the pageout daemon 1447 * is signalled and run to free more pages. 1448 */ 1449 if (vmstats.v_free_count > 6144) 1450 vmstats.v_free_target = 4 * vmstats.v_free_min + vmstats.v_free_reserved; 1451 else 1452 vmstats.v_free_target = 2 * vmstats.v_free_min + vmstats.v_free_reserved; 1453 1454 if (vmstats.v_free_count > 2048) { 1455 vmstats.v_cache_min = vmstats.v_free_target; 1456 vmstats.v_cache_max = 2 * vmstats.v_cache_min; 1457 vmstats.v_inactive_target = (3 * vmstats.v_free_target) / 2; 1458 } else { 1459 vmstats.v_cache_min = 0; 1460 vmstats.v_cache_max = 0; 1461 vmstats.v_inactive_target = vmstats.v_free_count / 4; 1462 } 1463 if (vmstats.v_inactive_target > vmstats.v_free_count / 3) 1464 vmstats.v_inactive_target = vmstats.v_free_count / 3; 1465 1466 /* XXX does not really belong here */ 1467 if (vm_page_max_wired == 0) 1468 vm_page_max_wired = vmstats.v_free_count / 3; 1469 1470 if (vm_pageout_stats_max == 0) 1471 vm_pageout_stats_max = vmstats.v_free_target; 1472 1473 /* 1474 * Set interval in seconds for stats scan. 1475 */ 1476 if (vm_pageout_stats_interval == 0) 1477 vm_pageout_stats_interval = 5; 1478 if (vm_pageout_full_stats_interval == 0) 1479 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4; 1480 1481 1482 /* 1483 * Set maximum free per pass 1484 */ 1485 if (vm_pageout_stats_free_max == 0) 1486 vm_pageout_stats_free_max = 5; 1487 1488 swap_pager_swap_init(); 1489 pass = 0; 1490 1491 /* 1492 * The pageout daemon is never done, so loop forever. 1493 */ 1494 while (TRUE) { 1495 int error; 1496 1497 if (vm_pages_needed == 0) { 1498 /* 1499 * Wait for an action request 1500 */ 1501 error = tsleep(&vm_pages_needed, 1502 0, "psleep", 1503 vm_pageout_stats_interval * hz); 1504 if (error && vm_pages_needed == 0) { 1505 vm_pageout_page_stats(); 1506 continue; 1507 } 1508 vm_pages_needed = 1; 1509 } 1510 1511 /* 1512 * If we have enough free memory, wakeup waiters. 1513 */ 1514 crit_enter(); 1515 if (!vm_page_count_min(0)) 1516 wakeup(&vmstats.v_free_count); 1517 mycpu->gd_cnt.v_pdwakeups++; 1518 crit_exit(); 1519 inactive_shortage = vm_pageout_scan(pass); 1520 1521 /* 1522 * Try to avoid thrashing the system with activity. 1523 */ 1524 if (inactive_shortage > 0) { 1525 ++pass; 1526 if (swap_pager_full) { 1527 /* 1528 * Running out of memory, catastrophic back-off 1529 * to one-second intervals. 1530 */ 1531 tsleep(&vm_pages_needed, 0, "pdelay", hz); 1532 } else if (pass < 10 && vm_pages_needed > 1) { 1533 /* 1534 * Normal operation, additional processes 1535 * have already kicked us. Retry immediately. 1536 */ 1537 } else if (pass < 10) { 1538 /* 1539 * Normal operation, fewer processes. Delay 1540 * a bit but allow wakeups. 1541 */ 1542 vm_pages_needed = 0; 1543 tsleep(&vm_pages_needed, 0, "pdelay", hz / 10); 1544 vm_pages_needed = 1; 1545 } else { 1546 /* 1547 * We've taken too many passes, forced delay. 1548 */ 1549 tsleep(&vm_pages_needed, 0, "pdelay", hz / 10); 1550 } 1551 } else { 1552 pass = 0; 1553 vm_pages_needed = 0; 1554 } 1555 } 1556 } 1557 1558 /* 1559 * Called after allocating a page out of the cache or free queue 1560 * to possibly wake the pagedaemon up to replentish our supply. 1561 * 1562 * We try to generate some hysteresis by waking the pagedaemon up 1563 * when our free+cache pages go below the severe level. The pagedaemon 1564 * tries to get the count back up to at least the minimum, and through 1565 * to the target level if possible. 1566 * 1567 * If the pagedaemon is already active bump vm_pages_needed as a hint 1568 * that there are even more requests pending. 1569 */ 1570 void 1571 pagedaemon_wakeup(void) 1572 { 1573 if (vm_page_count_severe() && curthread != pagethread) { 1574 if (vm_pages_needed == 0) { 1575 vm_pages_needed = 1; 1576 wakeup(&vm_pages_needed); 1577 } else if (vm_page_count_min(0)) { 1578 ++vm_pages_needed; 1579 } 1580 } 1581 } 1582 1583 #if !defined(NO_SWAPPING) 1584 static void 1585 vm_req_vmdaemon(void) 1586 { 1587 static int lastrun = 0; 1588 1589 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1590 wakeup(&vm_daemon_needed); 1591 lastrun = ticks; 1592 } 1593 } 1594 1595 static int vm_daemon_callback(struct proc *p, void *data __unused); 1596 1597 static void 1598 vm_daemon(void) 1599 { 1600 while (TRUE) { 1601 tsleep(&vm_daemon_needed, 0, "psleep", 0); 1602 if (vm_pageout_req_swapout) { 1603 swapout_procs(vm_pageout_req_swapout); 1604 vm_pageout_req_swapout = 0; 1605 } 1606 /* 1607 * scan the processes for exceeding their rlimits or if 1608 * process is swapped out -- deactivate pages 1609 */ 1610 allproc_scan(vm_daemon_callback, NULL); 1611 } 1612 } 1613 1614 static int 1615 vm_daemon_callback(struct proc *p, void *data __unused) 1616 { 1617 vm_pindex_t limit, size; 1618 1619 /* 1620 * if this is a system process or if we have already 1621 * looked at this process, skip it. 1622 */ 1623 if (p->p_flag & (P_SYSTEM | P_WEXIT)) 1624 return (0); 1625 1626 /* 1627 * if the process is in a non-running type state, 1628 * don't touch it. 1629 */ 1630 if (p->p_stat != SACTIVE && p->p_stat != SSTOP) 1631 return (0); 1632 1633 /* 1634 * get a limit 1635 */ 1636 limit = OFF_TO_IDX(qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur, 1637 p->p_rlimit[RLIMIT_RSS].rlim_max)); 1638 1639 /* 1640 * let processes that are swapped out really be 1641 * swapped out. Set the limit to nothing to get as 1642 * many pages out to swap as possible. 1643 */ 1644 if (p->p_flag & P_SWAPPEDOUT) 1645 limit = 0; 1646 1647 size = vmspace_resident_count(p->p_vmspace); 1648 if (limit >= 0 && size >= limit) { 1649 vm_pageout_map_deactivate_pages( 1650 &p->p_vmspace->vm_map, limit); 1651 } 1652 return (0); 1653 } 1654 1655 #endif 1656