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 1.35 2008/05/09 07:24:48 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 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. 396 */ 397 for (i = 0; i < count; i++) { 398 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, ("vm_pageout_flush page %p index %d/%d: partially invalid page", mc[i], i, count)); 399 vm_page_io_start(mc[i]); 400 } 401 402 /* 403 * We must make the pages read-only. This will also force the 404 * modified bit in the related pmaps to be cleared. The pager 405 * cannot clear the bit for us since the I/O completion code 406 * typically runs from an interrupt. The act of making the page 407 * read-only handles the case for us. 408 */ 409 for (i = 0; i < count; i++) { 410 vm_page_protect(mc[i], VM_PROT_READ); 411 } 412 413 object = mc[0]->object; 414 vm_object_pip_add(object, count); 415 416 vm_pager_put_pages(object, mc, count, 417 (flags | ((object == &kernel_object) ? VM_PAGER_PUT_SYNC : 0)), 418 pageout_status); 419 420 for (i = 0; i < count; i++) { 421 vm_page_t mt = mc[i]; 422 423 switch (pageout_status[i]) { 424 case VM_PAGER_OK: 425 numpagedout++; 426 break; 427 case VM_PAGER_PEND: 428 numpagedout++; 429 break; 430 case VM_PAGER_BAD: 431 /* 432 * Page outside of range of object. Right now we 433 * essentially lose the changes by pretending it 434 * worked. 435 */ 436 pmap_clear_modify(mt); 437 vm_page_undirty(mt); 438 break; 439 case VM_PAGER_ERROR: 440 case VM_PAGER_FAIL: 441 /* 442 * If page couldn't be paged out, then reactivate the 443 * page so it doesn't clog the inactive list. (We 444 * will try paging out it again later). 445 */ 446 vm_page_activate(mt); 447 break; 448 case VM_PAGER_AGAIN: 449 break; 450 } 451 452 /* 453 * If the operation is still going, leave the page busy to 454 * block all other accesses. Also, leave the paging in 455 * progress indicator set so that we don't attempt an object 456 * collapse. 457 * 458 * For any pages which have completed synchronously, 459 * deactivate the page if we are under a severe deficit. 460 * Do not try to enter them into the cache, though, they 461 * might still be read-heavy. 462 */ 463 if (pageout_status[i] != VM_PAGER_PEND) { 464 vm_object_pip_wakeup(object); 465 vm_page_io_finish(mt); 466 if (vm_page_count_severe()) 467 vm_page_deactivate(mt); 468 #if 0 469 if (!vm_page_count_severe() || !vm_page_try_to_cache(mt)) 470 vm_page_protect(mt, VM_PROT_READ); 471 #endif 472 } 473 } 474 return numpagedout; 475 } 476 477 #if !defined(NO_SWAPPING) 478 /* 479 * vm_pageout_object_deactivate_pages 480 * 481 * deactivate enough pages to satisfy the inactive target 482 * requirements or if vm_page_proc_limit is set, then 483 * deactivate all of the pages in the object and its 484 * backing_objects. 485 * 486 * The object and map must be locked. 487 */ 488 static int vm_pageout_object_deactivate_pages_callback(vm_page_t, void *); 489 490 static void 491 vm_pageout_object_deactivate_pages(vm_map_t map, vm_object_t object, 492 vm_pindex_t desired, int map_remove_only) 493 { 494 struct rb_vm_page_scan_info info; 495 int remove_mode; 496 497 if (object->type == OBJT_DEVICE || object->type == OBJT_PHYS) 498 return; 499 500 while (object) { 501 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 502 return; 503 if (object->paging_in_progress) 504 return; 505 506 remove_mode = map_remove_only; 507 if (object->shadow_count > 1) 508 remove_mode = 1; 509 510 /* 511 * scan the objects entire memory queue. spl protection is 512 * required to avoid an interrupt unbusy/free race against 513 * our busy check. 514 */ 515 crit_enter(); 516 info.limit = remove_mode; 517 info.map = map; 518 info.desired = desired; 519 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL, 520 vm_pageout_object_deactivate_pages_callback, 521 &info 522 ); 523 crit_exit(); 524 object = object->backing_object; 525 } 526 } 527 528 static int 529 vm_pageout_object_deactivate_pages_callback(vm_page_t p, void *data) 530 { 531 struct rb_vm_page_scan_info *info = data; 532 int actcount; 533 534 if (pmap_resident_count(vm_map_pmap(info->map)) <= info->desired) { 535 return(-1); 536 } 537 mycpu->gd_cnt.v_pdpages++; 538 if (p->wire_count != 0 || p->hold_count != 0 || p->busy != 0 || 539 (p->flags & (PG_BUSY|PG_UNMANAGED)) || 540 !pmap_page_exists_quick(vm_map_pmap(info->map), p)) { 541 return(0); 542 } 543 544 actcount = pmap_ts_referenced(p); 545 if (actcount) { 546 vm_page_flag_set(p, PG_REFERENCED); 547 } else if (p->flags & PG_REFERENCED) { 548 actcount = 1; 549 } 550 551 if ((p->queue != PQ_ACTIVE) && 552 (p->flags & PG_REFERENCED)) { 553 vm_page_activate(p); 554 p->act_count += actcount; 555 vm_page_flag_clear(p, PG_REFERENCED); 556 } else if (p->queue == PQ_ACTIVE) { 557 if ((p->flags & PG_REFERENCED) == 0) { 558 p->act_count -= min(p->act_count, ACT_DECLINE); 559 if (!info->limit && (vm_pageout_algorithm || (p->act_count == 0))) { 560 vm_page_busy(p); 561 vm_page_protect(p, VM_PROT_NONE); 562 vm_page_wakeup(p); 563 vm_page_deactivate(p); 564 } else { 565 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, p, pageq); 566 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, p, pageq); 567 } 568 } else { 569 vm_page_activate(p); 570 vm_page_flag_clear(p, PG_REFERENCED); 571 if (p->act_count < (ACT_MAX - ACT_ADVANCE)) 572 p->act_count += ACT_ADVANCE; 573 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, p, pageq); 574 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, p, pageq); 575 } 576 } else if (p->queue == PQ_INACTIVE) { 577 vm_page_busy(p); 578 vm_page_protect(p, VM_PROT_NONE); 579 vm_page_wakeup(p); 580 } 581 return(0); 582 } 583 584 /* 585 * deactivate some number of pages in a map, try to do it fairly, but 586 * that is really hard to do. 587 */ 588 static void 589 vm_pageout_map_deactivate_pages(vm_map_t map, vm_pindex_t desired) 590 { 591 vm_map_entry_t tmpe; 592 vm_object_t obj, bigobj; 593 int nothingwired; 594 595 if (lockmgr(&map->lock, LK_EXCLUSIVE | LK_NOWAIT)) { 596 return; 597 } 598 599 bigobj = NULL; 600 nothingwired = TRUE; 601 602 /* 603 * first, search out the biggest object, and try to free pages from 604 * that. 605 */ 606 tmpe = map->header.next; 607 while (tmpe != &map->header) { 608 switch(tmpe->maptype) { 609 case VM_MAPTYPE_NORMAL: 610 case VM_MAPTYPE_VPAGETABLE: 611 obj = tmpe->object.vm_object; 612 if ((obj != NULL) && (obj->shadow_count <= 1) && 613 ((bigobj == NULL) || 614 (bigobj->resident_page_count < obj->resident_page_count))) { 615 bigobj = obj; 616 } 617 break; 618 default: 619 break; 620 } 621 if (tmpe->wired_count > 0) 622 nothingwired = FALSE; 623 tmpe = tmpe->next; 624 } 625 626 if (bigobj) 627 vm_pageout_object_deactivate_pages(map, bigobj, desired, 0); 628 629 /* 630 * Next, hunt around for other pages to deactivate. We actually 631 * do this search sort of wrong -- .text first is not the best idea. 632 */ 633 tmpe = map->header.next; 634 while (tmpe != &map->header) { 635 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 636 break; 637 switch(tmpe->maptype) { 638 case VM_MAPTYPE_NORMAL: 639 case VM_MAPTYPE_VPAGETABLE: 640 obj = tmpe->object.vm_object; 641 if (obj) 642 vm_pageout_object_deactivate_pages(map, obj, desired, 0); 643 break; 644 default: 645 break; 646 } 647 tmpe = tmpe->next; 648 }; 649 650 /* 651 * Remove all mappings if a process is swapped out, this will free page 652 * table pages. 653 */ 654 if (desired == 0 && nothingwired) 655 pmap_remove(vm_map_pmap(map), 656 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS); 657 vm_map_unlock(map); 658 } 659 #endif 660 661 /* 662 * Don't try to be fancy - being fancy can lead to vnode deadlocks. We 663 * only do it for OBJT_DEFAULT and OBJT_SWAP objects which we know can 664 * be trivially freed. 665 */ 666 void 667 vm_pageout_page_free(vm_page_t m) 668 { 669 vm_object_t object = m->object; 670 int type = object->type; 671 672 if (type == OBJT_SWAP || type == OBJT_DEFAULT) 673 vm_object_reference(object); 674 vm_page_busy(m); 675 vm_page_protect(m, VM_PROT_NONE); 676 vm_page_free(m); 677 if (type == OBJT_SWAP || type == OBJT_DEFAULT) 678 vm_object_deallocate(object); 679 } 680 681 /* 682 * vm_pageout_scan does the dirty work for the pageout daemon. 683 */ 684 685 struct vm_pageout_scan_info { 686 struct proc *bigproc; 687 vm_offset_t bigsize; 688 }; 689 690 static int vm_pageout_scan_callback(struct proc *p, void *data); 691 692 static void 693 vm_pageout_scan(int pass) 694 { 695 struct vm_pageout_scan_info info; 696 vm_page_t m, next; 697 struct vm_page marker; 698 int page_shortage, maxscan, pcount; 699 int addl_page_shortage, addl_page_shortage_init; 700 vm_object_t object; 701 int actcount; 702 int vnodes_skipped = 0; 703 int maxlaunder; 704 705 /* 706 * Do whatever cleanup that the pmap code can. 707 */ 708 pmap_collect(); 709 710 addl_page_shortage_init = vm_pageout_deficit; 711 vm_pageout_deficit = 0; 712 713 /* 714 * Calculate the number of pages we want to either free or move 715 * to the cache. 716 */ 717 page_shortage = vm_paging_target() + addl_page_shortage_init; 718 719 /* 720 * Initialize our marker 721 */ 722 bzero(&marker, sizeof(marker)); 723 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER; 724 marker.queue = PQ_INACTIVE; 725 marker.wire_count = 1; 726 727 /* 728 * Start scanning the inactive queue for pages we can move to the 729 * cache or free. The scan will stop when the target is reached or 730 * we have scanned the entire inactive queue. Note that m->act_count 731 * is not used to form decisions for the inactive queue, only for the 732 * active queue. 733 * 734 * maxlaunder limits the number of dirty pages we flush per scan. 735 * For most systems a smaller value (16 or 32) is more robust under 736 * extreme memory and disk pressure because any unnecessary writes 737 * to disk can result in extreme performance degredation. However, 738 * systems with excessive dirty pages (especially when MAP_NOSYNC is 739 * used) will die horribly with limited laundering. If the pageout 740 * daemon cannot clean enough pages in the first pass, we let it go 741 * all out in succeeding passes. 742 */ 743 if ((maxlaunder = vm_max_launder) <= 1) 744 maxlaunder = 1; 745 if (pass) 746 maxlaunder = 10000; 747 748 /* 749 * We will generally be in a critical section throughout the 750 * scan, but we can release it temporarily when we are sitting on a 751 * non-busy page without fear. this is required to prevent an 752 * interrupt from unbusying or freeing a page prior to our busy 753 * check, leaving us on the wrong queue or checking the wrong 754 * page. 755 */ 756 crit_enter(); 757 rescan0: 758 addl_page_shortage = addl_page_shortage_init; 759 maxscan = vmstats.v_inactive_count; 760 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl); 761 m != NULL && maxscan-- > 0 && page_shortage > 0; 762 m = next 763 ) { 764 mycpu->gd_cnt.v_pdpages++; 765 766 /* 767 * Give interrupts a chance 768 */ 769 crit_exit(); 770 crit_enter(); 771 772 /* 773 * It's easier for some of the conditions below to just loop 774 * and catch queue changes here rather then check everywhere 775 * else. 776 */ 777 if (m->queue != PQ_INACTIVE) 778 goto rescan0; 779 next = TAILQ_NEXT(m, pageq); 780 781 /* 782 * skip marker pages 783 */ 784 if (m->flags & PG_MARKER) 785 continue; 786 787 /* 788 * A held page may be undergoing I/O, so skip it. 789 */ 790 if (m->hold_count) { 791 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 792 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 793 addl_page_shortage++; 794 continue; 795 } 796 797 /* 798 * Dont mess with busy pages, keep in the front of the 799 * queue, most likely are being paged out. 800 */ 801 if (m->busy || (m->flags & PG_BUSY)) { 802 addl_page_shortage++; 803 continue; 804 } 805 806 if (m->object->ref_count == 0) { 807 /* 808 * If the object is not being used, we ignore previous 809 * references. 810 */ 811 vm_page_flag_clear(m, PG_REFERENCED); 812 pmap_clear_reference(m); 813 814 } else if (((m->flags & PG_REFERENCED) == 0) && 815 (actcount = pmap_ts_referenced(m))) { 816 /* 817 * Otherwise, if the page has been referenced while 818 * in the inactive queue, we bump the "activation 819 * count" upwards, making it less likely that the 820 * page will be added back to the inactive queue 821 * prematurely again. Here we check the page tables 822 * (or emulated bits, if any), given the upper level 823 * VM system not knowing anything about existing 824 * references. 825 */ 826 vm_page_activate(m); 827 m->act_count += (actcount + ACT_ADVANCE); 828 continue; 829 } 830 831 /* 832 * If the upper level VM system knows about any page 833 * references, we activate the page. We also set the 834 * "activation count" higher than normal so that we will less 835 * likely place pages back onto the inactive queue again. 836 */ 837 if ((m->flags & PG_REFERENCED) != 0) { 838 vm_page_flag_clear(m, PG_REFERENCED); 839 actcount = pmap_ts_referenced(m); 840 vm_page_activate(m); 841 m->act_count += (actcount + ACT_ADVANCE + 1); 842 continue; 843 } 844 845 /* 846 * If the upper level VM system doesn't know anything about 847 * the page being dirty, we have to check for it again. As 848 * far as the VM code knows, any partially dirty pages are 849 * fully dirty. 850 * 851 * Pages marked PG_WRITEABLE may be mapped into the user 852 * address space of a process running on another cpu. A 853 * user process (without holding the MP lock) running on 854 * another cpu may be able to touch the page while we are 855 * trying to remove it. vm_page_cache() will handle this 856 * case for us. 857 */ 858 if (m->dirty == 0) { 859 vm_page_test_dirty(m); 860 #if 0 861 if (m->dirty == 0 && (m->flags & PG_WRITEABLE) != 0) 862 pmap_remove_all(m); 863 #endif 864 } else { 865 vm_page_dirty(m); 866 } 867 868 if (m->valid == 0) { 869 /* 870 * Invalid pages can be easily freed 871 */ 872 vm_pageout_page_free(m); 873 mycpu->gd_cnt.v_dfree++; 874 --page_shortage; 875 } else if (m->dirty == 0) { 876 /* 877 * Clean pages can be placed onto the cache queue. 878 * This effectively frees them. 879 */ 880 vm_page_cache(m); 881 --page_shortage; 882 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) { 883 /* 884 * Dirty pages need to be paged out, but flushing 885 * a page is extremely expensive verses freeing 886 * a clean page. Rather then artificially limiting 887 * the number of pages we can flush, we instead give 888 * dirty pages extra priority on the inactive queue 889 * by forcing them to be cycled through the queue 890 * twice before being flushed, after which the 891 * (now clean) page will cycle through once more 892 * before being freed. This significantly extends 893 * the thrash point for a heavily loaded machine. 894 */ 895 vm_page_flag_set(m, PG_WINATCFLS); 896 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 897 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 898 } else if (maxlaunder > 0) { 899 /* 900 * We always want to try to flush some dirty pages if 901 * we encounter them, to keep the system stable. 902 * Normally this number is small, but under extreme 903 * pressure where there are insufficient clean pages 904 * on the inactive queue, we may have to go all out. 905 */ 906 int swap_pageouts_ok; 907 struct vnode *vp = NULL; 908 909 object = m->object; 910 911 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 912 swap_pageouts_ok = 1; 913 } else { 914 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 915 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 916 vm_page_count_min()); 917 918 } 919 920 /* 921 * We don't bother paging objects that are "dead". 922 * Those objects are in a "rundown" state. 923 */ 924 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 925 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 926 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 927 continue; 928 } 929 930 /* 931 * The object is already known NOT to be dead. It 932 * is possible for the vget() to block the whole 933 * pageout daemon, but the new low-memory handling 934 * code should prevent it. 935 * 936 * The previous code skipped locked vnodes and, worse, 937 * reordered pages in the queue. This results in 938 * completely non-deterministic operation because, 939 * quite often, a vm_fault has initiated an I/O and 940 * is holding a locked vnode at just the point where 941 * the pageout daemon is woken up. 942 * 943 * We can't wait forever for the vnode lock, we might 944 * deadlock due to a vn_read() getting stuck in 945 * vm_wait while holding this vnode. We skip the 946 * vnode if we can't get it in a reasonable amount 947 * of time. 948 */ 949 950 if (object->type == OBJT_VNODE) { 951 vp = object->handle; 952 953 if (vget(vp, LK_EXCLUSIVE|LK_NOOBJ|LK_TIMELOCK)) { 954 ++pageout_lock_miss; 955 if (object->flags & OBJ_MIGHTBEDIRTY) 956 vnodes_skipped++; 957 continue; 958 } 959 960 /* 961 * The page might have been moved to another 962 * queue during potential blocking in vget() 963 * above. The page might have been freed and 964 * reused for another vnode. The object might 965 * have been reused for another vnode. 966 */ 967 if (m->queue != PQ_INACTIVE || 968 m->object != object || 969 object->handle != vp) { 970 if (object->flags & OBJ_MIGHTBEDIRTY) 971 vnodes_skipped++; 972 vput(vp); 973 continue; 974 } 975 976 /* 977 * The page may have been busied during the 978 * blocking in vput(); We don't move the 979 * page back onto the end of the queue so that 980 * statistics are more correct if we don't. 981 */ 982 if (m->busy || (m->flags & PG_BUSY)) { 983 vput(vp); 984 continue; 985 } 986 987 /* 988 * If the page has become held it might 989 * be undergoing I/O, so skip it 990 */ 991 if (m->hold_count) { 992 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 993 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 994 if (object->flags & OBJ_MIGHTBEDIRTY) 995 vnodes_skipped++; 996 vput(vp); 997 continue; 998 } 999 } 1000 1001 /* 1002 * If a page is dirty, then it is either being washed 1003 * (but not yet cleaned) or it is still in the 1004 * laundry. If it is still in the laundry, then we 1005 * start the cleaning operation. 1006 * 1007 * This operation may cluster, invalidating the 'next' 1008 * pointer. To prevent an inordinate number of 1009 * restarts we use our marker to remember our place. 1010 * 1011 * decrement page_shortage on success to account for 1012 * the (future) cleaned page. Otherwise we could wind 1013 * up laundering or cleaning too many pages. 1014 */ 1015 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, m, &marker, pageq); 1016 if (vm_pageout_clean(m) != 0) { 1017 --page_shortage; 1018 --maxlaunder; 1019 } 1020 next = TAILQ_NEXT(&marker, pageq); 1021 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, &marker, pageq); 1022 if (vp != NULL) 1023 vput(vp); 1024 } 1025 } 1026 1027 /* 1028 * Compute the number of pages we want to try to move from the 1029 * active queue to the inactive queue. 1030 */ 1031 page_shortage = vm_paging_target() + 1032 vmstats.v_inactive_target - vmstats.v_inactive_count; 1033 page_shortage += addl_page_shortage; 1034 1035 /* 1036 * Scan the active queue for things we can deactivate. We nominally 1037 * track the per-page activity counter and use it to locate 1038 * deactivation candidates. 1039 * 1040 * NOTE: we are still in a critical section. 1041 */ 1042 pcount = vmstats.v_active_count; 1043 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1044 1045 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) { 1046 /* 1047 * Give interrupts a chance. 1048 */ 1049 crit_exit(); 1050 crit_enter(); 1051 1052 /* 1053 * If the page was ripped out from under us, just stop. 1054 */ 1055 if (m->queue != PQ_ACTIVE) 1056 break; 1057 next = TAILQ_NEXT(m, pageq); 1058 1059 /* 1060 * Don't deactivate pages that are busy. 1061 */ 1062 if ((m->busy != 0) || 1063 (m->flags & PG_BUSY) || 1064 (m->hold_count != 0)) { 1065 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1066 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1067 m = next; 1068 continue; 1069 } 1070 1071 /* 1072 * The count for pagedaemon pages is done after checking the 1073 * page for eligibility... 1074 */ 1075 mycpu->gd_cnt.v_pdpages++; 1076 1077 /* 1078 * Check to see "how much" the page has been used. 1079 */ 1080 actcount = 0; 1081 if (m->object->ref_count != 0) { 1082 if (m->flags & PG_REFERENCED) { 1083 actcount += 1; 1084 } 1085 actcount += pmap_ts_referenced(m); 1086 if (actcount) { 1087 m->act_count += ACT_ADVANCE + actcount; 1088 if (m->act_count > ACT_MAX) 1089 m->act_count = ACT_MAX; 1090 } 1091 } 1092 1093 /* 1094 * Since we have "tested" this bit, we need to clear it now. 1095 */ 1096 vm_page_flag_clear(m, PG_REFERENCED); 1097 1098 /* 1099 * Only if an object is currently being used, do we use the 1100 * page activation count stats. 1101 */ 1102 if (actcount && (m->object->ref_count != 0)) { 1103 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1104 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1105 } else { 1106 m->act_count -= min(m->act_count, ACT_DECLINE); 1107 if (vm_pageout_algorithm || 1108 m->object->ref_count == 0 || 1109 m->act_count < pass) { 1110 page_shortage--; 1111 if (m->object->ref_count == 0) { 1112 vm_page_busy(m); 1113 vm_page_protect(m, VM_PROT_NONE); 1114 vm_page_wakeup(m); 1115 if (m->dirty == 0) 1116 vm_page_cache(m); 1117 else 1118 vm_page_deactivate(m); 1119 } else { 1120 vm_page_deactivate(m); 1121 } 1122 } else { 1123 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1124 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1125 } 1126 } 1127 m = next; 1128 } 1129 1130 /* 1131 * We try to maintain some *really* free pages, this allows interrupt 1132 * code to be guaranteed space. Since both cache and free queues 1133 * are considered basically 'free', moving pages from cache to free 1134 * does not effect other calculations. 1135 * 1136 * NOTE: we are still in a critical section. 1137 */ 1138 1139 while (vmstats.v_free_count < vmstats.v_free_reserved) { 1140 static int cache_rover = 0; 1141 m = vm_page_list_find(PQ_CACHE, cache_rover, FALSE); 1142 if (!m) 1143 break; 1144 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || 1145 m->busy || 1146 m->hold_count || 1147 m->wire_count) { 1148 #ifdef INVARIANTS 1149 kprintf("Warning: busy page %p found in cache\n", m); 1150 #endif 1151 vm_page_deactivate(m); 1152 continue; 1153 } 1154 KKASSERT((m->flags & PG_MAPPED) == 0); 1155 KKASSERT(m->dirty == 0); 1156 cache_rover = (cache_rover + PQ_PRIME2) & PQ_L2_MASK; 1157 vm_pageout_page_free(m); 1158 mycpu->gd_cnt.v_dfree++; 1159 } 1160 1161 crit_exit(); 1162 1163 #if !defined(NO_SWAPPING) 1164 /* 1165 * Idle process swapout -- run once per second. 1166 */ 1167 if (vm_swap_idle_enabled) { 1168 static long lsec; 1169 if (time_second != lsec) { 1170 vm_pageout_req_swapout |= VM_SWAP_IDLE; 1171 vm_req_vmdaemon(); 1172 lsec = time_second; 1173 } 1174 } 1175 #endif 1176 1177 /* 1178 * If we didn't get enough free pages, and we have skipped a vnode 1179 * in a writeable object, wakeup the sync daemon. And kick swapout 1180 * if we did not get enough free pages. 1181 */ 1182 if (vm_paging_target() > 0) { 1183 if (vnodes_skipped && vm_page_count_min()) 1184 speedup_syncer(); 1185 #if !defined(NO_SWAPPING) 1186 if (vm_swap_enabled && vm_page_count_target()) { 1187 vm_req_vmdaemon(); 1188 vm_pageout_req_swapout |= VM_SWAP_NORMAL; 1189 } 1190 #endif 1191 } 1192 1193 /* 1194 * If we are out of swap and were not able to reach our paging 1195 * target, kill the largest process. 1196 */ 1197 if ((vm_swap_size < 64 && vm_page_count_min()) || 1198 (swap_pager_full && vm_paging_target() > 0)) { 1199 #if 0 1200 if ((vm_swap_size < 64 || swap_pager_full) && vm_page_count_min()) { 1201 #endif 1202 info.bigproc = NULL; 1203 info.bigsize = 0; 1204 allproc_scan(vm_pageout_scan_callback, &info); 1205 if (info.bigproc != NULL) { 1206 killproc(info.bigproc, "out of swap space"); 1207 info.bigproc->p_nice = PRIO_MIN; 1208 info.bigproc->p_usched->resetpriority( 1209 FIRST_LWP_IN_PROC(info.bigproc)); 1210 wakeup(&vmstats.v_free_count); 1211 PRELE(info.bigproc); 1212 } 1213 } 1214 } 1215 1216 static int 1217 vm_pageout_scan_callback(struct proc *p, void *data) 1218 { 1219 struct vm_pageout_scan_info *info = data; 1220 vm_offset_t size; 1221 1222 /* 1223 * if this is a system process, skip it 1224 */ 1225 if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) || 1226 ((p->p_pid < 48) && (vm_swap_size != 0))) { 1227 return (0); 1228 } 1229 1230 /* 1231 * if the process is in a non-running type state, 1232 * don't touch it. 1233 */ 1234 if (p->p_stat != SACTIVE && p->p_stat != SSTOP) { 1235 return (0); 1236 } 1237 1238 /* 1239 * get the process size 1240 */ 1241 size = vmspace_resident_count(p->p_vmspace) + 1242 vmspace_swap_count(p->p_vmspace); 1243 1244 /* 1245 * If the this process is bigger than the biggest one 1246 * remember it. 1247 */ 1248 if (size > info->bigsize) { 1249 if (info->bigproc) 1250 PRELE(info->bigproc); 1251 PHOLD(p); 1252 info->bigproc = p; 1253 info->bigsize = size; 1254 } 1255 return(0); 1256 } 1257 1258 /* 1259 * This routine tries to maintain the pseudo LRU active queue, 1260 * so that during long periods of time where there is no paging, 1261 * that some statistic accumulation still occurs. This code 1262 * helps the situation where paging just starts to occur. 1263 */ 1264 static void 1265 vm_pageout_page_stats(void) 1266 { 1267 vm_page_t m,next; 1268 int pcount,tpcount; /* Number of pages to check */ 1269 static int fullintervalcount = 0; 1270 int page_shortage; 1271 1272 page_shortage = 1273 (vmstats.v_inactive_target + vmstats.v_cache_max + vmstats.v_free_min) - 1274 (vmstats.v_free_count + vmstats.v_inactive_count + vmstats.v_cache_count); 1275 1276 if (page_shortage <= 0) 1277 return; 1278 1279 crit_enter(); 1280 1281 pcount = vmstats.v_active_count; 1282 fullintervalcount += vm_pageout_stats_interval; 1283 if (fullintervalcount < vm_pageout_full_stats_interval) { 1284 tpcount = (vm_pageout_stats_max * vmstats.v_active_count) / vmstats.v_page_count; 1285 if (pcount > tpcount) 1286 pcount = tpcount; 1287 } else { 1288 fullintervalcount = 0; 1289 } 1290 1291 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1292 while ((m != NULL) && (pcount-- > 0)) { 1293 int actcount; 1294 1295 if (m->queue != PQ_ACTIVE) { 1296 break; 1297 } 1298 1299 next = TAILQ_NEXT(m, pageq); 1300 /* 1301 * Don't deactivate pages that are busy. 1302 */ 1303 if ((m->busy != 0) || 1304 (m->flags & PG_BUSY) || 1305 (m->hold_count != 0)) { 1306 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1307 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1308 m = next; 1309 continue; 1310 } 1311 1312 actcount = 0; 1313 if (m->flags & PG_REFERENCED) { 1314 vm_page_flag_clear(m, PG_REFERENCED); 1315 actcount += 1; 1316 } 1317 1318 actcount += pmap_ts_referenced(m); 1319 if (actcount) { 1320 m->act_count += ACT_ADVANCE + actcount; 1321 if (m->act_count > ACT_MAX) 1322 m->act_count = ACT_MAX; 1323 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1324 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1325 } else { 1326 if (m->act_count == 0) { 1327 /* 1328 * We turn off page access, so that we have 1329 * more accurate RSS stats. We don't do this 1330 * in the normal page deactivation when the 1331 * system is loaded VM wise, because the 1332 * cost of the large number of page protect 1333 * operations would be higher than the value 1334 * of doing the operation. 1335 */ 1336 vm_page_busy(m); 1337 vm_page_protect(m, VM_PROT_NONE); 1338 vm_page_wakeup(m); 1339 vm_page_deactivate(m); 1340 } else { 1341 m->act_count -= min(m->act_count, ACT_DECLINE); 1342 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1343 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1344 } 1345 } 1346 1347 m = next; 1348 } 1349 crit_exit(); 1350 } 1351 1352 static int 1353 vm_pageout_free_page_calc(vm_size_t count) 1354 { 1355 if (count < vmstats.v_page_count) 1356 return 0; 1357 /* 1358 * free_reserved needs to include enough for the largest swap pager 1359 * structures plus enough for any pv_entry structs when paging. 1360 */ 1361 if (vmstats.v_page_count > 1024) 1362 vmstats.v_free_min = 4 + (vmstats.v_page_count - 1024) / 200; 1363 else 1364 vmstats.v_free_min = 4; 1365 vmstats.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1366 vmstats.v_interrupt_free_min; 1367 vmstats.v_free_reserved = vm_pageout_page_count + 1368 vmstats.v_pageout_free_min + (count / 768) + PQ_L2_SIZE; 1369 vmstats.v_free_severe = vmstats.v_free_min / 2; 1370 vmstats.v_free_min += vmstats.v_free_reserved; 1371 vmstats.v_free_severe += vmstats.v_free_reserved; 1372 return 1; 1373 } 1374 1375 1376 /* 1377 * vm_pageout is the high level pageout daemon. 1378 */ 1379 static void 1380 vm_pageout(void) 1381 { 1382 int pass; 1383 1384 /* 1385 * Initialize some paging parameters. 1386 */ 1387 1388 vmstats.v_interrupt_free_min = 2; 1389 if (vmstats.v_page_count < 2000) 1390 vm_pageout_page_count = 8; 1391 1392 vm_pageout_free_page_calc(vmstats.v_page_count); 1393 /* 1394 * v_free_target and v_cache_min control pageout hysteresis. Note 1395 * that these are more a measure of the VM cache queue hysteresis 1396 * then the VM free queue. Specifically, v_free_target is the 1397 * high water mark (free+cache pages). 1398 * 1399 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the 1400 * low water mark, while v_free_min is the stop. v_cache_min must 1401 * be big enough to handle memory needs while the pageout daemon 1402 * is signalled and run to free more pages. 1403 */ 1404 if (vmstats.v_free_count > 6144) 1405 vmstats.v_free_target = 4 * vmstats.v_free_min + vmstats.v_free_reserved; 1406 else 1407 vmstats.v_free_target = 2 * vmstats.v_free_min + vmstats.v_free_reserved; 1408 1409 if (vmstats.v_free_count > 2048) { 1410 vmstats.v_cache_min = vmstats.v_free_target; 1411 vmstats.v_cache_max = 2 * vmstats.v_cache_min; 1412 vmstats.v_inactive_target = (3 * vmstats.v_free_target) / 2; 1413 } else { 1414 vmstats.v_cache_min = 0; 1415 vmstats.v_cache_max = 0; 1416 vmstats.v_inactive_target = vmstats.v_free_count / 4; 1417 } 1418 if (vmstats.v_inactive_target > vmstats.v_free_count / 3) 1419 vmstats.v_inactive_target = vmstats.v_free_count / 3; 1420 1421 /* XXX does not really belong here */ 1422 if (vm_page_max_wired == 0) 1423 vm_page_max_wired = vmstats.v_free_count / 3; 1424 1425 if (vm_pageout_stats_max == 0) 1426 vm_pageout_stats_max = vmstats.v_free_target; 1427 1428 /* 1429 * Set interval in seconds for stats scan. 1430 */ 1431 if (vm_pageout_stats_interval == 0) 1432 vm_pageout_stats_interval = 5; 1433 if (vm_pageout_full_stats_interval == 0) 1434 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4; 1435 1436 1437 /* 1438 * Set maximum free per pass 1439 */ 1440 if (vm_pageout_stats_free_max == 0) 1441 vm_pageout_stats_free_max = 5; 1442 1443 swap_pager_swap_init(); 1444 pass = 0; 1445 /* 1446 * The pageout daemon is never done, so loop forever. 1447 */ 1448 while (TRUE) { 1449 int error; 1450 1451 /* 1452 * If we have enough free memory, wakeup waiters. Do 1453 * not clear vm_pages_needed until we reach our target, 1454 * otherwise we may be woken up over and over again and 1455 * waste a lot of cpu. 1456 */ 1457 crit_enter(); 1458 if (vm_pages_needed && !vm_page_count_min()) { 1459 if (vm_paging_needed() <= 0) 1460 vm_pages_needed = 0; 1461 wakeup(&vmstats.v_free_count); 1462 } 1463 if (vm_pages_needed) { 1464 /* 1465 * Still not done, take a second pass without waiting 1466 * (unlimited dirty cleaning), otherwise sleep a bit 1467 * and try again. 1468 */ 1469 ++pass; 1470 if (pass > 1) 1471 tsleep(&vm_pages_needed, 0, "psleep", hz/2); 1472 } else { 1473 /* 1474 * Good enough, sleep & handle stats. Prime the pass 1475 * for the next run. 1476 */ 1477 if (pass > 1) 1478 pass = 1; 1479 else 1480 pass = 0; 1481 error = tsleep(&vm_pages_needed, 1482 0, "psleep", vm_pageout_stats_interval * hz); 1483 if (error && !vm_pages_needed) { 1484 crit_exit(); 1485 pass = 0; 1486 vm_pageout_page_stats(); 1487 continue; 1488 } 1489 } 1490 1491 if (vm_pages_needed) 1492 mycpu->gd_cnt.v_pdwakeups++; 1493 crit_exit(); 1494 vm_pageout_scan(pass); 1495 vm_pageout_deficit = 0; 1496 } 1497 } 1498 1499 void 1500 pagedaemon_wakeup(void) 1501 { 1502 if (!vm_pages_needed && curthread != pagethread) { 1503 vm_pages_needed++; 1504 wakeup(&vm_pages_needed); 1505 } 1506 } 1507 1508 #if !defined(NO_SWAPPING) 1509 static void 1510 vm_req_vmdaemon(void) 1511 { 1512 static int lastrun = 0; 1513 1514 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1515 wakeup(&vm_daemon_needed); 1516 lastrun = ticks; 1517 } 1518 } 1519 1520 static int vm_daemon_callback(struct proc *p, void *data __unused); 1521 1522 static void 1523 vm_daemon(void) 1524 { 1525 while (TRUE) { 1526 tsleep(&vm_daemon_needed, 0, "psleep", 0); 1527 if (vm_pageout_req_swapout) { 1528 swapout_procs(vm_pageout_req_swapout); 1529 vm_pageout_req_swapout = 0; 1530 } 1531 /* 1532 * scan the processes for exceeding their rlimits or if 1533 * process is swapped out -- deactivate pages 1534 */ 1535 allproc_scan(vm_daemon_callback, NULL); 1536 } 1537 } 1538 1539 static int 1540 vm_daemon_callback(struct proc *p, void *data __unused) 1541 { 1542 vm_pindex_t limit, size; 1543 1544 /* 1545 * if this is a system process or if we have already 1546 * looked at this process, skip it. 1547 */ 1548 if (p->p_flag & (P_SYSTEM | P_WEXIT)) 1549 return (0); 1550 1551 /* 1552 * if the process is in a non-running type state, 1553 * don't touch it. 1554 */ 1555 if (p->p_stat != SACTIVE && p->p_stat != SSTOP) 1556 return (0); 1557 1558 /* 1559 * get a limit 1560 */ 1561 limit = OFF_TO_IDX(qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur, 1562 p->p_rlimit[RLIMIT_RSS].rlim_max)); 1563 1564 /* 1565 * let processes that are swapped out really be 1566 * swapped out. Set the limit to nothing to get as 1567 * many pages out to swap as possible. 1568 */ 1569 if (p->p_flag & P_SWAPPEDOUT) 1570 limit = 0; 1571 1572 size = vmspace_resident_count(p->p_vmspace); 1573 if (limit >= 0 && size >= limit) { 1574 vm_pageout_map_deactivate_pages( 1575 &p->p_vmspace->vm_map, limit); 1576 } 1577 return (0); 1578 } 1579 1580 #endif 1581