1 /* 2 * (MPSAFE) 3 * 4 * Copyright (c) 1991 Regents of the University of California. 5 * All rights reserved. 6 * Copyright (c) 1994 John S. Dyson 7 * All rights reserved. 8 * Copyright (c) 1994 David Greenman 9 * All rights reserved. 10 * 11 * This code is derived from software contributed to Berkeley by 12 * The Mach Operating System project at Carnegie-Mellon University. 13 * 14 * Redistribution and use in source and binary forms, with or without 15 * modification, are permitted provided that the following conditions 16 * are met: 17 * 1. Redistributions of source code must retain the above copyright 18 * notice, this list of conditions and the following disclaimer. 19 * 2. Redistributions in binary form must reproduce the above copyright 20 * notice, this list of conditions and the following disclaimer in the 21 * documentation and/or other materials provided with the distribution. 22 * 3. All advertising materials mentioning features or use of this software 23 * must display the following acknowledgement: 24 * This product includes software developed by the University of 25 * California, Berkeley and its contributors. 26 * 4. Neither the name of the University nor the names of its contributors 27 * may be used to endorse or promote products derived from this software 28 * without specific prior written permission. 29 * 30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 40 * SUCH DAMAGE. 41 * 42 * from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91 43 * 44 * 45 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 46 * All rights reserved. 47 * 48 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 49 * 50 * Permission to use, copy, modify and distribute this software and 51 * its documentation is hereby granted, provided that both the copyright 52 * notice and this permission notice appear in all copies of the 53 * software, derivative works or modified versions, and any portions 54 * thereof, and that both notices appear in supporting documentation. 55 * 56 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 57 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 58 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 59 * 60 * Carnegie Mellon requests users of this software to return to 61 * 62 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 63 * School of Computer Science 64 * Carnegie Mellon University 65 * Pittsburgh PA 15213-3890 66 * 67 * any improvements or extensions that they make and grant Carnegie the 68 * rights to redistribute these changes. 69 * 70 * $FreeBSD: src/sys/vm/vm_pageout.c,v 1.151.2.15 2002/12/29 18:21:04 dillon Exp $ 71 * $DragonFly: src/sys/vm/vm_pageout.c,v 1.36 2008/07/01 02:02:56 dillon Exp $ 72 */ 73 74 /* 75 * The proverbial page-out daemon. 76 */ 77 78 #include "opt_vm.h" 79 #include <sys/param.h> 80 #include <sys/systm.h> 81 #include <sys/kernel.h> 82 #include <sys/proc.h> 83 #include <sys/kthread.h> 84 #include <sys/resourcevar.h> 85 #include <sys/signalvar.h> 86 #include <sys/vnode.h> 87 #include <sys/vmmeter.h> 88 #include <sys/sysctl.h> 89 90 #include <vm/vm.h> 91 #include <vm/vm_param.h> 92 #include <sys/lock.h> 93 #include <vm/vm_object.h> 94 #include <vm/vm_page.h> 95 #include <vm/vm_map.h> 96 #include <vm/vm_pageout.h> 97 #include <vm/vm_pager.h> 98 #include <vm/swap_pager.h> 99 #include <vm/vm_extern.h> 100 101 #include <sys/thread2.h> 102 #include <vm/vm_page2.h> 103 104 /* 105 * System initialization 106 */ 107 108 /* the kernel process "vm_pageout"*/ 109 static void vm_pageout (void); 110 static int vm_pageout_clean (vm_page_t); 111 static int vm_pageout_scan (int pass); 112 static int vm_pageout_free_page_calc (vm_size_t count); 113 struct thread *pagethread; 114 115 static struct kproc_desc page_kp = { 116 "pagedaemon", 117 vm_pageout, 118 &pagethread 119 }; 120 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp) 121 122 #if !defined(NO_SWAPPING) 123 /* the kernel process "vm_daemon"*/ 124 static void vm_daemon (void); 125 static struct thread *vmthread; 126 127 static struct kproc_desc vm_kp = { 128 "vmdaemon", 129 vm_daemon, 130 &vmthread 131 }; 132 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp) 133 #endif 134 135 136 int vm_pages_needed=0; /* Event on which pageout daemon sleeps */ 137 int vm_pageout_deficit=0; /* Estimated number of pages deficit */ 138 int vm_pageout_pages_needed=0; /* flag saying that the pageout daemon needs pages */ 139 140 #if !defined(NO_SWAPPING) 141 static int vm_pageout_req_swapout; /* XXX */ 142 static int vm_daemon_needed; 143 #endif 144 static int vm_max_launder = 32; 145 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0; 146 static int vm_pageout_full_stats_interval = 0; 147 static int vm_pageout_stats_free_max=0, vm_pageout_algorithm=0; 148 static int defer_swap_pageouts=0; 149 static int disable_swap_pageouts=0; 150 151 #if defined(NO_SWAPPING) 152 static int vm_swap_enabled=0; 153 static int vm_swap_idle_enabled=0; 154 #else 155 static int vm_swap_enabled=1; 156 static int vm_swap_idle_enabled=0; 157 #endif 158 159 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm, 160 CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt"); 161 162 SYSCTL_INT(_vm, OID_AUTO, max_launder, 163 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout"); 164 165 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max, 166 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length"); 167 168 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval, 169 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan"); 170 171 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval, 172 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan"); 173 174 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_free_max, 175 CTLFLAG_RW, &vm_pageout_stats_free_max, 0, "Not implemented"); 176 177 #if defined(NO_SWAPPING) 178 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 179 CTLFLAG_RD, &vm_swap_enabled, 0, ""); 180 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 181 CTLFLAG_RD, &vm_swap_idle_enabled, 0, ""); 182 #else 183 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 184 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout"); 185 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 186 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 187 #endif 188 189 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts, 190 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem"); 191 192 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts, 193 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages"); 194 195 static int pageout_lock_miss; 196 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss, 197 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout"); 198 199 int vm_load; 200 SYSCTL_INT(_vm, OID_AUTO, vm_load, 201 CTLFLAG_RD, &vm_load, 0, "load on the VM system"); 202 int vm_load_enable = 1; 203 SYSCTL_INT(_vm, OID_AUTO, vm_load_enable, 204 CTLFLAG_RW, &vm_load_enable, 0, "enable vm_load rate limiting"); 205 #ifdef INVARIANTS 206 int vm_load_debug; 207 SYSCTL_INT(_vm, OID_AUTO, vm_load_debug, 208 CTLFLAG_RW, &vm_load_debug, 0, "debug vm_load"); 209 #endif 210 211 #define VM_PAGEOUT_PAGE_COUNT 16 212 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT; 213 214 int vm_page_max_wired; /* XXX max # of wired pages system-wide */ 215 216 #if !defined(NO_SWAPPING) 217 typedef void freeer_fcn_t (vm_map_t, vm_object_t, vm_pindex_t, int); 218 static void vm_pageout_map_deactivate_pages (vm_map_t, vm_pindex_t); 219 static freeer_fcn_t vm_pageout_object_deactivate_pages; 220 static void vm_req_vmdaemon (void); 221 #endif 222 static void vm_pageout_page_stats(void); 223 224 /* 225 * Update vm_load to slow down faulting processes. 226 * 227 * SMP races ok. 228 * No requirements. 229 */ 230 void 231 vm_fault_ratecheck(void) 232 { 233 if (vm_pages_needed) { 234 if (vm_load < 1000) 235 ++vm_load; 236 } else { 237 if (vm_load > 0) 238 --vm_load; 239 } 240 } 241 242 /* 243 * vm_pageout_clean: 244 * 245 * Clean the page and remove it from the laundry. The page must not be 246 * busy on-call. 247 * 248 * We set the busy bit to cause potential page faults on this page to 249 * block. Note the careful timing, however, the busy bit isn't set till 250 * late and we cannot do anything that will mess with the page. 251 * 252 * The caller must hold vm_token. 253 */ 254 static int 255 vm_pageout_clean(vm_page_t m) 256 { 257 vm_object_t object; 258 vm_page_t mc[2*vm_pageout_page_count]; 259 int pageout_count; 260 int ib, is, page_base; 261 vm_pindex_t pindex = m->pindex; 262 263 object = m->object; 264 265 /* 266 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP 267 * with the new swapper, but we could have serious problems paging 268 * out other object types if there is insufficient memory. 269 * 270 * Unfortunately, checking free memory here is far too late, so the 271 * check has been moved up a procedural level. 272 */ 273 274 /* 275 * Don't mess with the page if it's busy, held, or special 276 */ 277 if ((m->hold_count != 0) || 278 ((m->busy != 0) || (m->flags & (PG_BUSY|PG_UNMANAGED)))) { 279 return 0; 280 } 281 282 mc[vm_pageout_page_count] = m; 283 pageout_count = 1; 284 page_base = vm_pageout_page_count; 285 ib = 1; 286 is = 1; 287 288 /* 289 * Scan object for clusterable pages. 290 * 291 * We can cluster ONLY if: ->> the page is NOT 292 * clean, wired, busy, held, or mapped into a 293 * buffer, and one of the following: 294 * 1) The page is inactive, or a seldom used 295 * active page. 296 * -or- 297 * 2) we force the issue. 298 * 299 * During heavy mmap/modification loads the pageout 300 * daemon can really fragment the underlying file 301 * due to flushing pages out of order and not trying 302 * align the clusters (which leave sporatic out-of-order 303 * holes). To solve this problem we do the reverse scan 304 * first and attempt to align our cluster, then do a 305 * forward scan if room remains. 306 */ 307 308 more: 309 while (ib && pageout_count < vm_pageout_page_count) { 310 vm_page_t p; 311 312 if (ib > pindex) { 313 ib = 0; 314 break; 315 } 316 317 if ((p = vm_page_lookup(object, pindex - ib)) == NULL) { 318 ib = 0; 319 break; 320 } 321 if (((p->queue - p->pc) == PQ_CACHE) || 322 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) { 323 ib = 0; 324 break; 325 } 326 vm_page_test_dirty(p); 327 if ((p->dirty & p->valid) == 0 || 328 p->queue != PQ_INACTIVE || 329 p->wire_count != 0 || /* may be held by buf cache */ 330 p->hold_count != 0) { /* may be undergoing I/O */ 331 ib = 0; 332 break; 333 } 334 mc[--page_base] = p; 335 ++pageout_count; 336 ++ib; 337 /* 338 * alignment boundry, stop here and switch directions. Do 339 * not clear ib. 340 */ 341 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0) 342 break; 343 } 344 345 while (pageout_count < vm_pageout_page_count && 346 pindex + is < object->size) { 347 vm_page_t p; 348 349 if ((p = vm_page_lookup(object, pindex + is)) == NULL) 350 break; 351 if (((p->queue - p->pc) == PQ_CACHE) || 352 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) { 353 break; 354 } 355 vm_page_test_dirty(p); 356 if ((p->dirty & p->valid) == 0 || 357 p->queue != PQ_INACTIVE || 358 p->wire_count != 0 || /* may be held by buf cache */ 359 p->hold_count != 0) { /* may be undergoing I/O */ 360 break; 361 } 362 mc[page_base + pageout_count] = p; 363 ++pageout_count; 364 ++is; 365 } 366 367 /* 368 * If we exhausted our forward scan, continue with the reverse scan 369 * when possible, even past a page boundry. This catches boundry 370 * conditions. 371 */ 372 if (ib && pageout_count < vm_pageout_page_count) 373 goto more; 374 375 /* 376 * we allow reads during pageouts... 377 */ 378 return vm_pageout_flush(&mc[page_base], pageout_count, 0); 379 } 380 381 /* 382 * vm_pageout_flush() - launder the given pages 383 * 384 * The given pages are laundered. Note that we setup for the start of 385 * I/O ( i.e. busy the page ), mark it read-only, and bump the object 386 * reference count all in here rather then in the parent. If we want 387 * the parent to do more sophisticated things we may have to change 388 * the ordering. 389 * 390 * The caller must hold vm_token. 391 */ 392 int 393 vm_pageout_flush(vm_page_t *mc, int count, int flags) 394 { 395 vm_object_t object; 396 int pageout_status[count]; 397 int numpagedout = 0; 398 int i; 399 400 ASSERT_LWKT_TOKEN_HELD(&vm_token); 401 402 /* 403 * Initiate I/O. Bump the vm_page_t->busy counter. 404 */ 405 for (i = 0; i < count; i++) { 406 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, ("vm_pageout_flush page %p index %d/%d: partially invalid page", mc[i], i, count)); 407 vm_page_io_start(mc[i]); 408 } 409 410 /* 411 * We must make the pages read-only. This will also force the 412 * modified bit in the related pmaps to be cleared. The pager 413 * cannot clear the bit for us since the I/O completion code 414 * typically runs from an interrupt. The act of making the page 415 * read-only handles the case for us. 416 */ 417 for (i = 0; i < count; i++) { 418 vm_page_protect(mc[i], VM_PROT_READ); 419 } 420 421 object = mc[0]->object; 422 vm_object_pip_add(object, count); 423 424 vm_pager_put_pages(object, mc, count, 425 (flags | ((object == &kernel_object) ? VM_PAGER_PUT_SYNC : 0)), 426 pageout_status); 427 428 for (i = 0; i < count; i++) { 429 vm_page_t mt = mc[i]; 430 431 switch (pageout_status[i]) { 432 case VM_PAGER_OK: 433 numpagedout++; 434 break; 435 case VM_PAGER_PEND: 436 numpagedout++; 437 break; 438 case VM_PAGER_BAD: 439 /* 440 * Page outside of range of object. Right now we 441 * essentially lose the changes by pretending it 442 * worked. 443 */ 444 pmap_clear_modify(mt); 445 vm_page_undirty(mt); 446 break; 447 case VM_PAGER_ERROR: 448 case VM_PAGER_FAIL: 449 /* 450 * A page typically cannot be paged out when we 451 * have run out of swap. We leave the page 452 * marked inactive and will try to page it out 453 * again later. 454 * 455 * Starvation of the active page list is used to 456 * determine when the system is massively memory 457 * starved. 458 */ 459 break; 460 case VM_PAGER_AGAIN: 461 break; 462 } 463 464 /* 465 * If the operation is still going, leave the page busy to 466 * block all other accesses. Also, leave the paging in 467 * progress indicator set so that we don't attempt an object 468 * collapse. 469 * 470 * For any pages which have completed synchronously, 471 * deactivate the page if we are under a severe deficit. 472 * Do not try to enter them into the cache, though, they 473 * might still be read-heavy. 474 */ 475 if (pageout_status[i] != VM_PAGER_PEND) { 476 vm_object_pip_wakeup(object); 477 vm_page_io_finish(mt); 478 if (vm_page_count_severe()) 479 vm_page_deactivate(mt); 480 #if 0 481 if (!vm_page_count_severe() || !vm_page_try_to_cache(mt)) 482 vm_page_protect(mt, VM_PROT_READ); 483 #endif 484 } 485 } 486 return numpagedout; 487 } 488 489 #if !defined(NO_SWAPPING) 490 /* 491 * vm_pageout_object_deactivate_pages 492 * 493 * deactivate enough pages to satisfy the inactive target 494 * requirements or if vm_page_proc_limit is set, then 495 * deactivate all of the pages in the object and its 496 * backing_objects. 497 * 498 * The map must be locked. 499 * The caller must hold vm_token. 500 */ 501 static int vm_pageout_object_deactivate_pages_callback(vm_page_t, void *); 502 503 static void 504 vm_pageout_object_deactivate_pages(vm_map_t map, vm_object_t object, 505 vm_pindex_t desired, int map_remove_only) 506 { 507 struct rb_vm_page_scan_info info; 508 int remove_mode; 509 510 if (object->type == OBJT_DEVICE || object->type == OBJT_PHYS) 511 return; 512 513 while (object) { 514 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 515 return; 516 if (object->paging_in_progress) 517 return; 518 519 remove_mode = map_remove_only; 520 if (object->shadow_count > 1) 521 remove_mode = 1; 522 523 /* 524 * scan the objects entire memory queue. spl protection is 525 * required to avoid an interrupt unbusy/free race against 526 * our busy check. 527 */ 528 crit_enter(); 529 info.limit = remove_mode; 530 info.map = map; 531 info.desired = desired; 532 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL, 533 vm_pageout_object_deactivate_pages_callback, 534 &info 535 ); 536 crit_exit(); 537 object = object->backing_object; 538 } 539 } 540 541 /* 542 * The caller must hold vm_token. 543 */ 544 static int 545 vm_pageout_object_deactivate_pages_callback(vm_page_t p, void *data) 546 { 547 struct rb_vm_page_scan_info *info = data; 548 int actcount; 549 550 if (pmap_resident_count(vm_map_pmap(info->map)) <= info->desired) { 551 return(-1); 552 } 553 mycpu->gd_cnt.v_pdpages++; 554 if (p->wire_count != 0 || p->hold_count != 0 || p->busy != 0 || 555 (p->flags & (PG_BUSY|PG_UNMANAGED)) || 556 !pmap_page_exists_quick(vm_map_pmap(info->map), p)) { 557 return(0); 558 } 559 560 actcount = pmap_ts_referenced(p); 561 if (actcount) { 562 vm_page_flag_set(p, PG_REFERENCED); 563 } else if (p->flags & PG_REFERENCED) { 564 actcount = 1; 565 } 566 567 if ((p->queue != PQ_ACTIVE) && 568 (p->flags & PG_REFERENCED)) { 569 vm_page_activate(p); 570 p->act_count += actcount; 571 vm_page_flag_clear(p, PG_REFERENCED); 572 } else if (p->queue == PQ_ACTIVE) { 573 if ((p->flags & PG_REFERENCED) == 0) { 574 p->act_count -= min(p->act_count, ACT_DECLINE); 575 if (!info->limit && (vm_pageout_algorithm || (p->act_count == 0))) { 576 vm_page_busy(p); 577 vm_page_protect(p, VM_PROT_NONE); 578 vm_page_wakeup(p); 579 vm_page_deactivate(p); 580 } else { 581 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, p, pageq); 582 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, p, pageq); 583 } 584 } else { 585 vm_page_activate(p); 586 vm_page_flag_clear(p, PG_REFERENCED); 587 if (p->act_count < (ACT_MAX - ACT_ADVANCE)) 588 p->act_count += ACT_ADVANCE; 589 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, p, pageq); 590 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, p, pageq); 591 } 592 } else if (p->queue == PQ_INACTIVE) { 593 vm_page_busy(p); 594 vm_page_protect(p, VM_PROT_NONE); 595 vm_page_wakeup(p); 596 } 597 return(0); 598 } 599 600 /* 601 * Deactivate some number of pages in a map, try to do it fairly, but 602 * that is really hard to do. 603 * 604 * The caller must hold vm_token. 605 */ 606 static void 607 vm_pageout_map_deactivate_pages(vm_map_t map, vm_pindex_t desired) 608 { 609 vm_map_entry_t tmpe; 610 vm_object_t obj, bigobj; 611 int nothingwired; 612 613 if (lockmgr(&map->lock, LK_EXCLUSIVE | LK_NOWAIT)) { 614 return; 615 } 616 617 bigobj = NULL; 618 nothingwired = TRUE; 619 620 /* 621 * first, search out the biggest object, and try to free pages from 622 * that. 623 */ 624 tmpe = map->header.next; 625 while (tmpe != &map->header) { 626 switch(tmpe->maptype) { 627 case VM_MAPTYPE_NORMAL: 628 case VM_MAPTYPE_VPAGETABLE: 629 obj = tmpe->object.vm_object; 630 if ((obj != NULL) && (obj->shadow_count <= 1) && 631 ((bigobj == NULL) || 632 (bigobj->resident_page_count < obj->resident_page_count))) { 633 bigobj = obj; 634 } 635 break; 636 default: 637 break; 638 } 639 if (tmpe->wired_count > 0) 640 nothingwired = FALSE; 641 tmpe = tmpe->next; 642 } 643 644 if (bigobj) 645 vm_pageout_object_deactivate_pages(map, bigobj, desired, 0); 646 647 /* 648 * Next, hunt around for other pages to deactivate. We actually 649 * do this search sort of wrong -- .text first is not the best idea. 650 */ 651 tmpe = map->header.next; 652 while (tmpe != &map->header) { 653 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 654 break; 655 switch(tmpe->maptype) { 656 case VM_MAPTYPE_NORMAL: 657 case VM_MAPTYPE_VPAGETABLE: 658 obj = tmpe->object.vm_object; 659 if (obj) 660 vm_pageout_object_deactivate_pages(map, obj, desired, 0); 661 break; 662 default: 663 break; 664 } 665 tmpe = tmpe->next; 666 }; 667 668 /* 669 * Remove all mappings if a process is swapped out, this will free page 670 * table pages. 671 */ 672 if (desired == 0 && nothingwired) 673 pmap_remove(vm_map_pmap(map), 674 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS); 675 vm_map_unlock(map); 676 } 677 #endif 678 679 /* 680 * Don't try to be fancy - being fancy can lead to vnode deadlocks. We 681 * only do it for OBJT_DEFAULT and OBJT_SWAP objects which we know can 682 * be trivially freed. 683 * 684 * The caller must hold vm_token. 685 */ 686 static void 687 vm_pageout_page_free(vm_page_t m) 688 { 689 vm_object_t object = m->object; 690 int type = object->type; 691 692 if (type == OBJT_SWAP || type == OBJT_DEFAULT) 693 vm_object_reference(object); 694 vm_page_busy(m); 695 vm_page_protect(m, VM_PROT_NONE); 696 vm_page_free(m); 697 if (type == OBJT_SWAP || type == OBJT_DEFAULT) 698 vm_object_deallocate(object); 699 } 700 701 /* 702 * vm_pageout_scan does the dirty work for the pageout daemon. 703 */ 704 struct vm_pageout_scan_info { 705 struct proc *bigproc; 706 vm_offset_t bigsize; 707 }; 708 709 static int vm_pageout_scan_callback(struct proc *p, void *data); 710 711 /* 712 * The caller must hold vm_token. 713 */ 714 static int 715 vm_pageout_scan(int pass) 716 { 717 struct vm_pageout_scan_info info; 718 vm_page_t m, next; 719 struct vm_page marker; 720 struct vnode *vpfailed; /* warning, allowed to be stale */ 721 int maxscan, pcount; 722 int recycle_count; 723 int inactive_shortage, active_shortage; 724 int inactive_original_shortage; 725 vm_object_t object; 726 int actcount; 727 int vnodes_skipped = 0; 728 int maxlaunder; 729 730 /* 731 * Do whatever cleanup that the pmap code can. 732 */ 733 pmap_collect(); 734 735 /* 736 * Calculate our target for the number of free+cache pages we 737 * want to get to. This is higher then the number that causes 738 * allocations to stall (severe) in order to provide hysteresis, 739 * and if we don't make it all the way but get to the minimum 740 * we're happy. 741 */ 742 inactive_shortage = vm_paging_target() + vm_pageout_deficit; 743 inactive_original_shortage = inactive_shortage; 744 vm_pageout_deficit = 0; 745 746 /* 747 * Initialize our marker 748 */ 749 bzero(&marker, sizeof(marker)); 750 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER; 751 marker.queue = PQ_INACTIVE; 752 marker.wire_count = 1; 753 754 /* 755 * Start scanning the inactive queue for pages we can move to the 756 * cache or free. The scan will stop when the target is reached or 757 * we have scanned the entire inactive queue. Note that m->act_count 758 * is not used to form decisions for the inactive queue, only for the 759 * active queue. 760 * 761 * maxlaunder limits the number of dirty pages we flush per scan. 762 * For most systems a smaller value (16 or 32) is more robust under 763 * extreme memory and disk pressure because any unnecessary writes 764 * to disk can result in extreme performance degredation. However, 765 * systems with excessive dirty pages (especially when MAP_NOSYNC is 766 * used) will die horribly with limited laundering. If the pageout 767 * daemon cannot clean enough pages in the first pass, we let it go 768 * all out in succeeding passes. 769 */ 770 if ((maxlaunder = vm_max_launder) <= 1) 771 maxlaunder = 1; 772 if (pass) 773 maxlaunder = 10000; 774 775 /* 776 * We will generally be in a critical section throughout the 777 * scan, but we can release it temporarily when we are sitting on a 778 * non-busy page without fear. this is required to prevent an 779 * interrupt from unbusying or freeing a page prior to our busy 780 * check, leaving us on the wrong queue or checking the wrong 781 * page. 782 */ 783 crit_enter(); 784 rescan0: 785 vpfailed = NULL; 786 maxscan = vmstats.v_inactive_count; 787 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl); 788 m != NULL && maxscan-- > 0 && inactive_shortage > 0; 789 m = next 790 ) { 791 mycpu->gd_cnt.v_pdpages++; 792 793 /* 794 * Give interrupts a chance 795 */ 796 crit_exit(); 797 crit_enter(); 798 799 /* 800 * It's easier for some of the conditions below to just loop 801 * and catch queue changes here rather then check everywhere 802 * else. 803 */ 804 if (m->queue != PQ_INACTIVE) 805 goto rescan0; 806 next = TAILQ_NEXT(m, pageq); 807 808 /* 809 * skip marker pages 810 */ 811 if (m->flags & PG_MARKER) 812 continue; 813 814 /* 815 * A held page may be undergoing I/O, so skip it. 816 */ 817 if (m->hold_count) { 818 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 819 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 820 ++vm_swapcache_inactive_heuristic; 821 continue; 822 } 823 824 /* 825 * Dont mess with busy pages, keep in the front of the 826 * queue, most likely are being paged out. 827 */ 828 if (m->busy || (m->flags & PG_BUSY)) { 829 continue; 830 } 831 832 if (m->object->ref_count == 0) { 833 /* 834 * If the object is not being used, we ignore previous 835 * references. 836 */ 837 vm_page_flag_clear(m, PG_REFERENCED); 838 pmap_clear_reference(m); 839 840 } else if (((m->flags & PG_REFERENCED) == 0) && 841 (actcount = pmap_ts_referenced(m))) { 842 /* 843 * Otherwise, if the page has been referenced while 844 * in the inactive queue, we bump the "activation 845 * count" upwards, making it less likely that the 846 * page will be added back to the inactive queue 847 * prematurely again. Here we check the page tables 848 * (or emulated bits, if any), given the upper level 849 * VM system not knowing anything about existing 850 * references. 851 */ 852 vm_page_activate(m); 853 m->act_count += (actcount + ACT_ADVANCE); 854 continue; 855 } 856 857 /* 858 * If the upper level VM system knows about any page 859 * references, we activate the page. We also set the 860 * "activation count" higher than normal so that we will less 861 * likely place pages back onto the inactive queue again. 862 */ 863 if ((m->flags & PG_REFERENCED) != 0) { 864 vm_page_flag_clear(m, PG_REFERENCED); 865 actcount = pmap_ts_referenced(m); 866 vm_page_activate(m); 867 m->act_count += (actcount + ACT_ADVANCE + 1); 868 continue; 869 } 870 871 /* 872 * If the upper level VM system doesn't know anything about 873 * the page being dirty, we have to check for it again. As 874 * far as the VM code knows, any partially dirty pages are 875 * fully dirty. 876 * 877 * Pages marked PG_WRITEABLE may be mapped into the user 878 * address space of a process running on another cpu. A 879 * user process (without holding the MP lock) running on 880 * another cpu may be able to touch the page while we are 881 * trying to remove it. vm_page_cache() will handle this 882 * case for us. 883 */ 884 if (m->dirty == 0) { 885 vm_page_test_dirty(m); 886 } else { 887 vm_page_dirty(m); 888 } 889 890 if (m->valid == 0) { 891 /* 892 * Invalid pages can be easily freed 893 */ 894 vm_pageout_page_free(m); 895 mycpu->gd_cnt.v_dfree++; 896 --inactive_shortage; 897 } else if (m->dirty == 0) { 898 /* 899 * Clean pages can be placed onto the cache queue. 900 * This effectively frees them. 901 */ 902 vm_page_cache(m); 903 --inactive_shortage; 904 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) { 905 /* 906 * Dirty pages need to be paged out, but flushing 907 * a page is extremely expensive verses freeing 908 * a clean page. Rather then artificially limiting 909 * the number of pages we can flush, we instead give 910 * dirty pages extra priority on the inactive queue 911 * by forcing them to be cycled through the queue 912 * twice before being flushed, after which the 913 * (now clean) page will cycle through once more 914 * before being freed. This significantly extends 915 * the thrash point for a heavily loaded machine. 916 */ 917 vm_page_flag_set(m, PG_WINATCFLS); 918 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 919 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 920 ++vm_swapcache_inactive_heuristic; 921 } else if (maxlaunder > 0) { 922 /* 923 * We always want to try to flush some dirty pages if 924 * we encounter them, to keep the system stable. 925 * Normally this number is small, but under extreme 926 * pressure where there are insufficient clean pages 927 * on the inactive queue, we may have to go all out. 928 */ 929 int swap_pageouts_ok; 930 struct vnode *vp = NULL; 931 932 object = m->object; 933 934 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 935 swap_pageouts_ok = 1; 936 } else { 937 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 938 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 939 vm_page_count_min(0)); 940 941 } 942 943 /* 944 * We don't bother paging objects that are "dead". 945 * Those objects are in a "rundown" state. 946 */ 947 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 948 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 949 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 950 ++vm_swapcache_inactive_heuristic; 951 continue; 952 } 953 954 /* 955 * The object is already known NOT to be dead. It 956 * is possible for the vget() to block the whole 957 * pageout daemon, but the new low-memory handling 958 * code should prevent it. 959 * 960 * The previous code skipped locked vnodes and, worse, 961 * reordered pages in the queue. This results in 962 * completely non-deterministic operation because, 963 * quite often, a vm_fault has initiated an I/O and 964 * is holding a locked vnode at just the point where 965 * the pageout daemon is woken up. 966 * 967 * We can't wait forever for the vnode lock, we might 968 * deadlock due to a vn_read() getting stuck in 969 * vm_wait while holding this vnode. We skip the 970 * vnode if we can't get it in a reasonable amount 971 * of time. 972 * 973 * vpfailed is used to (try to) avoid the case where 974 * a large number of pages are associated with a 975 * locked vnode, which could cause the pageout daemon 976 * to stall for an excessive amount of time. 977 */ 978 if (object->type == OBJT_VNODE) { 979 int flags; 980 981 vp = object->handle; 982 flags = LK_EXCLUSIVE | LK_NOOBJ; 983 if (vp == vpfailed) 984 flags |= LK_NOWAIT; 985 else 986 flags |= LK_TIMELOCK; 987 if (vget(vp, flags) != 0) { 988 vpfailed = vp; 989 ++pageout_lock_miss; 990 if (object->flags & OBJ_MIGHTBEDIRTY) 991 vnodes_skipped++; 992 continue; 993 } 994 995 /* 996 * The page might have been moved to another 997 * queue during potential blocking in vget() 998 * above. The page might have been freed and 999 * reused for another vnode. The object might 1000 * have been reused for another vnode. 1001 */ 1002 if (m->queue != PQ_INACTIVE || 1003 m->object != object || 1004 object->handle != vp) { 1005 if (object->flags & OBJ_MIGHTBEDIRTY) 1006 vnodes_skipped++; 1007 vput(vp); 1008 continue; 1009 } 1010 1011 /* 1012 * The page may have been busied during the 1013 * blocking in vput(); We don't move the 1014 * page back onto the end of the queue so that 1015 * statistics are more correct if we don't. 1016 */ 1017 if (m->busy || (m->flags & PG_BUSY)) { 1018 vput(vp); 1019 continue; 1020 } 1021 1022 /* 1023 * If the page has become held it might 1024 * be undergoing I/O, so skip it 1025 */ 1026 if (m->hold_count) { 1027 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1028 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1029 ++vm_swapcache_inactive_heuristic; 1030 if (object->flags & OBJ_MIGHTBEDIRTY) 1031 vnodes_skipped++; 1032 vput(vp); 1033 continue; 1034 } 1035 } 1036 1037 /* 1038 * If a page is dirty, then it is either being washed 1039 * (but not yet cleaned) or it is still in the 1040 * laundry. If it is still in the laundry, then we 1041 * start the cleaning operation. 1042 * 1043 * This operation may cluster, invalidating the 'next' 1044 * pointer. To prevent an inordinate number of 1045 * restarts we use our marker to remember our place. 1046 * 1047 * decrement inactive_shortage on success to account 1048 * for the (future) cleaned page. Otherwise we 1049 * could wind up laundering or cleaning too many 1050 * pages. 1051 */ 1052 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, m, &marker, pageq); 1053 if (vm_pageout_clean(m) != 0) { 1054 --inactive_shortage; 1055 --maxlaunder; 1056 } 1057 next = TAILQ_NEXT(&marker, pageq); 1058 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, &marker, pageq); 1059 if (vp != NULL) 1060 vput(vp); 1061 } 1062 } 1063 1064 /* 1065 * We want to move pages from the active queue to the inactive 1066 * queue to get the inactive queue to the inactive target. If 1067 * we still have a page shortage from above we try to directly free 1068 * clean pages instead of moving them. 1069 * 1070 * If we do still have a shortage we keep track of the number of 1071 * pages we free or cache (recycle_count) as a measure of thrashing 1072 * between the active and inactive queues. 1073 * 1074 * If we were able to completely satisfy the free+cache targets 1075 * from the inactive pool we limit the number of pages we move 1076 * from the active pool to the inactive pool to 2x the pages we 1077 * had removed from the inactive pool (with a minimum of 1/5 the 1078 * inactive target). If we were not able to completely satisfy 1079 * the free+cache targets we go for the whole target aggressively. 1080 * 1081 * NOTE: Both variables can end up negative. 1082 * NOTE: We are still in a critical section. 1083 */ 1084 active_shortage = vmstats.v_inactive_target - vmstats.v_inactive_count; 1085 if (inactive_original_shortage < vmstats.v_inactive_target / 10) 1086 inactive_original_shortage = vmstats.v_inactive_target / 10; 1087 if (inactive_shortage <= 0 && 1088 active_shortage > inactive_original_shortage * 2) { 1089 active_shortage = inactive_original_shortage * 2; 1090 } 1091 1092 pcount = vmstats.v_active_count; 1093 recycle_count = 0; 1094 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1095 1096 while ((m != NULL) && (pcount-- > 0) && 1097 (inactive_shortage > 0 || active_shortage > 0) 1098 ) { 1099 /* 1100 * Give interrupts a chance. 1101 */ 1102 crit_exit(); 1103 crit_enter(); 1104 1105 /* 1106 * If the page was ripped out from under us, just stop. 1107 */ 1108 if (m->queue != PQ_ACTIVE) 1109 break; 1110 next = TAILQ_NEXT(m, pageq); 1111 1112 /* 1113 * Don't deactivate pages that are busy. 1114 */ 1115 if ((m->busy != 0) || 1116 (m->flags & PG_BUSY) || 1117 (m->hold_count != 0)) { 1118 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1119 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1120 m = next; 1121 continue; 1122 } 1123 1124 /* 1125 * The count for pagedaemon pages is done after checking the 1126 * page for eligibility... 1127 */ 1128 mycpu->gd_cnt.v_pdpages++; 1129 1130 /* 1131 * Check to see "how much" the page has been used and clear 1132 * the tracking access bits. If the object has no references 1133 * don't bother paying the expense. 1134 */ 1135 actcount = 0; 1136 if (m->object->ref_count != 0) { 1137 if (m->flags & PG_REFERENCED) 1138 ++actcount; 1139 actcount += pmap_ts_referenced(m); 1140 if (actcount) { 1141 m->act_count += ACT_ADVANCE + actcount; 1142 if (m->act_count > ACT_MAX) 1143 m->act_count = ACT_MAX; 1144 } 1145 } 1146 vm_page_flag_clear(m, PG_REFERENCED); 1147 1148 /* 1149 * actcount is only valid if the object ref_count is non-zero. 1150 */ 1151 if (actcount && m->object->ref_count != 0) { 1152 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1153 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1154 } else { 1155 m->act_count -= min(m->act_count, ACT_DECLINE); 1156 if (vm_pageout_algorithm || 1157 m->object->ref_count == 0 || 1158 m->act_count < pass + 1 1159 ) { 1160 /* 1161 * Deactivate the page. If we had a 1162 * shortage from our inactive scan try to 1163 * free (cache) the page instead. 1164 * 1165 * Don't just blindly cache the page if 1166 * we do not have a shortage from the 1167 * inactive scan, that could lead to 1168 * gigabytes being moved. 1169 */ 1170 --active_shortage; 1171 if (inactive_shortage > 0 || 1172 m->object->ref_count == 0) { 1173 if (inactive_shortage > 0) 1174 ++recycle_count; 1175 vm_page_busy(m); 1176 vm_page_protect(m, VM_PROT_NONE); 1177 vm_page_wakeup(m); 1178 if (m->dirty == 0 && 1179 inactive_shortage > 0) { 1180 --inactive_shortage; 1181 vm_page_cache(m); 1182 } else { 1183 vm_page_deactivate(m); 1184 } 1185 } else { 1186 vm_page_deactivate(m); 1187 } 1188 } else { 1189 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1190 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1191 } 1192 } 1193 m = next; 1194 } 1195 1196 /* 1197 * We try to maintain some *really* free pages, this allows interrupt 1198 * code to be guaranteed space. Since both cache and free queues 1199 * are considered basically 'free', moving pages from cache to free 1200 * does not effect other calculations. 1201 * 1202 * NOTE: we are still in a critical section. 1203 * 1204 * Pages moved from PQ_CACHE to totally free are not counted in the 1205 * pages_freed counter. 1206 */ 1207 while (vmstats.v_free_count < vmstats.v_free_reserved) { 1208 static int cache_rover = 0; 1209 m = vm_page_list_find(PQ_CACHE, cache_rover, FALSE); 1210 if (m == NULL) 1211 break; 1212 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || 1213 m->busy || 1214 m->hold_count || 1215 m->wire_count) { 1216 #ifdef INVARIANTS 1217 kprintf("Warning: busy page %p found in cache\n", m); 1218 #endif 1219 vm_page_deactivate(m); 1220 continue; 1221 } 1222 KKASSERT((m->flags & PG_MAPPED) == 0); 1223 KKASSERT(m->dirty == 0); 1224 cache_rover = (cache_rover + PQ_PRIME2) & PQ_L2_MASK; 1225 vm_pageout_page_free(m); 1226 mycpu->gd_cnt.v_dfree++; 1227 } 1228 1229 crit_exit(); 1230 1231 #if !defined(NO_SWAPPING) 1232 /* 1233 * Idle process swapout -- run once per second. 1234 */ 1235 if (vm_swap_idle_enabled) { 1236 static long lsec; 1237 if (time_second != lsec) { 1238 vm_pageout_req_swapout |= VM_SWAP_IDLE; 1239 vm_req_vmdaemon(); 1240 lsec = time_second; 1241 } 1242 } 1243 #endif 1244 1245 /* 1246 * If we didn't get enough free pages, and we have skipped a vnode 1247 * in a writeable object, wakeup the sync daemon. And kick swapout 1248 * if we did not get enough free pages. 1249 */ 1250 if (vm_paging_target() > 0) { 1251 if (vnodes_skipped && vm_page_count_min(0)) 1252 speedup_syncer(); 1253 #if !defined(NO_SWAPPING) 1254 if (vm_swap_enabled && vm_page_count_target()) { 1255 vm_req_vmdaemon(); 1256 vm_pageout_req_swapout |= VM_SWAP_NORMAL; 1257 } 1258 #endif 1259 } 1260 1261 /* 1262 * Handle catastrophic conditions. Under good conditions we should 1263 * be at the target, well beyond our minimum. If we could not even 1264 * reach our minimum the system is under heavy stress. 1265 * 1266 * Determine whether we have run out of memory. This occurs when 1267 * swap_pager_full is TRUE and the only pages left in the page 1268 * queues are dirty. We will still likely have page shortages. 1269 * 1270 * - swap_pager_full is set if insufficient swap was 1271 * available to satisfy a requested pageout. 1272 * 1273 * - the inactive queue is bloated (4 x size of active queue), 1274 * meaning it is unable to get rid of dirty pages and. 1275 * 1276 * - vm_page_count_min() without counting pages recycled from the 1277 * active queue (recycle_count) means we could not recover 1278 * enough pages to meet bare minimum needs. This test only 1279 * works if the inactive queue is bloated. 1280 * 1281 * - due to a positive inactive_shortage we shifted the remaining 1282 * dirty pages from the active queue to the inactive queue 1283 * trying to find clean ones to free. 1284 */ 1285 if (swap_pager_full && vm_page_count_min(recycle_count)) 1286 kprintf("Warning: system low on memory+swap!\n"); 1287 if (swap_pager_full && vm_page_count_min(recycle_count) && 1288 vmstats.v_inactive_count > vmstats.v_active_count * 4 && 1289 inactive_shortage > 0) { 1290 /* 1291 * Kill something. 1292 */ 1293 info.bigproc = NULL; 1294 info.bigsize = 0; 1295 allproc_scan(vm_pageout_scan_callback, &info); 1296 if (info.bigproc != NULL) { 1297 killproc(info.bigproc, "out of swap space"); 1298 info.bigproc->p_nice = PRIO_MIN; 1299 info.bigproc->p_usched->resetpriority( 1300 FIRST_LWP_IN_PROC(info.bigproc)); 1301 wakeup(&vmstats.v_free_count); 1302 PRELE(info.bigproc); 1303 } 1304 } 1305 return(inactive_shortage); 1306 } 1307 1308 /* 1309 * The caller must hold vm_token and proc_token. 1310 */ 1311 static int 1312 vm_pageout_scan_callback(struct proc *p, void *data) 1313 { 1314 struct vm_pageout_scan_info *info = data; 1315 vm_offset_t size; 1316 1317 /* 1318 * Never kill system processes or init. If we have configured swap 1319 * then try to avoid killing low-numbered pids. 1320 */ 1321 if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) || 1322 ((p->p_pid < 48) && (vm_swap_size != 0))) { 1323 return (0); 1324 } 1325 1326 /* 1327 * if the process is in a non-running type state, 1328 * don't touch it. 1329 */ 1330 if (p->p_stat != SACTIVE && p->p_stat != SSTOP) 1331 return (0); 1332 1333 /* 1334 * Get the approximate process size. Note that anonymous pages 1335 * with backing swap will be counted twice, but there should not 1336 * be too many such pages due to the stress the VM system is 1337 * under at this point. 1338 */ 1339 size = vmspace_anonymous_count(p->p_vmspace) + 1340 vmspace_swap_count(p->p_vmspace); 1341 1342 /* 1343 * If the this process is bigger than the biggest one 1344 * remember it. 1345 */ 1346 if (info->bigsize < size) { 1347 if (info->bigproc) 1348 PRELE(info->bigproc); 1349 PHOLD(p); 1350 info->bigproc = p; 1351 info->bigsize = size; 1352 } 1353 return(0); 1354 } 1355 1356 /* 1357 * This routine tries to maintain the pseudo LRU active queue, 1358 * so that during long periods of time where there is no paging, 1359 * that some statistic accumulation still occurs. This code 1360 * helps the situation where paging just starts to occur. 1361 * 1362 * The caller must hold vm_token. 1363 */ 1364 static void 1365 vm_pageout_page_stats(void) 1366 { 1367 vm_page_t m,next; 1368 int pcount,tpcount; /* Number of pages to check */ 1369 static int fullintervalcount = 0; 1370 int page_shortage; 1371 1372 page_shortage = 1373 (vmstats.v_inactive_target + vmstats.v_cache_max + vmstats.v_free_min) - 1374 (vmstats.v_free_count + vmstats.v_inactive_count + vmstats.v_cache_count); 1375 1376 if (page_shortage <= 0) 1377 return; 1378 1379 crit_enter(); 1380 1381 pcount = vmstats.v_active_count; 1382 fullintervalcount += vm_pageout_stats_interval; 1383 if (fullintervalcount < vm_pageout_full_stats_interval) { 1384 tpcount = (vm_pageout_stats_max * vmstats.v_active_count) / vmstats.v_page_count; 1385 if (pcount > tpcount) 1386 pcount = tpcount; 1387 } else { 1388 fullintervalcount = 0; 1389 } 1390 1391 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1392 while ((m != NULL) && (pcount-- > 0)) { 1393 int actcount; 1394 1395 if (m->queue != PQ_ACTIVE) { 1396 break; 1397 } 1398 1399 next = TAILQ_NEXT(m, pageq); 1400 /* 1401 * Don't deactivate pages that are busy. 1402 */ 1403 if ((m->busy != 0) || 1404 (m->flags & PG_BUSY) || 1405 (m->hold_count != 0)) { 1406 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1407 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1408 m = next; 1409 continue; 1410 } 1411 1412 actcount = 0; 1413 if (m->flags & PG_REFERENCED) { 1414 vm_page_flag_clear(m, PG_REFERENCED); 1415 actcount += 1; 1416 } 1417 1418 actcount += pmap_ts_referenced(m); 1419 if (actcount) { 1420 m->act_count += ACT_ADVANCE + actcount; 1421 if (m->act_count > ACT_MAX) 1422 m->act_count = ACT_MAX; 1423 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1424 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1425 } else { 1426 if (m->act_count == 0) { 1427 /* 1428 * We turn off page access, so that we have 1429 * more accurate RSS stats. We don't do this 1430 * in the normal page deactivation when the 1431 * system is loaded VM wise, because the 1432 * cost of the large number of page protect 1433 * operations would be higher than the value 1434 * of doing the operation. 1435 */ 1436 vm_page_busy(m); 1437 vm_page_protect(m, VM_PROT_NONE); 1438 vm_page_wakeup(m); 1439 vm_page_deactivate(m); 1440 } else { 1441 m->act_count -= min(m->act_count, ACT_DECLINE); 1442 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1443 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1444 } 1445 } 1446 1447 m = next; 1448 } 1449 crit_exit(); 1450 } 1451 1452 /* 1453 * The caller must hold vm_token. 1454 */ 1455 static int 1456 vm_pageout_free_page_calc(vm_size_t count) 1457 { 1458 if (count < vmstats.v_page_count) 1459 return 0; 1460 /* 1461 * free_reserved needs to include enough for the largest swap pager 1462 * structures plus enough for any pv_entry structs when paging. 1463 */ 1464 if (vmstats.v_page_count > 1024) 1465 vmstats.v_free_min = 4 + (vmstats.v_page_count - 1024) / 200; 1466 else 1467 vmstats.v_free_min = 4; 1468 vmstats.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1469 vmstats.v_interrupt_free_min; 1470 vmstats.v_free_reserved = vm_pageout_page_count + 1471 vmstats.v_pageout_free_min + (count / 768) + PQ_L2_SIZE; 1472 vmstats.v_free_severe = vmstats.v_free_min / 2; 1473 vmstats.v_free_min += vmstats.v_free_reserved; 1474 vmstats.v_free_severe += vmstats.v_free_reserved; 1475 return 1; 1476 } 1477 1478 1479 /* 1480 * vm_pageout is the high level pageout daemon. 1481 * 1482 * No requirements. 1483 */ 1484 static void 1485 vm_pageout(void) 1486 { 1487 int pass; 1488 int inactive_shortage; 1489 1490 /* 1491 * Permanently hold vm_token. 1492 */ 1493 lwkt_gettoken(&vm_token); 1494 1495 /* 1496 * Initialize some paging parameters. 1497 */ 1498 curthread->td_flags |= TDF_SYSTHREAD; 1499 1500 vmstats.v_interrupt_free_min = 2; 1501 if (vmstats.v_page_count < 2000) 1502 vm_pageout_page_count = 8; 1503 1504 vm_pageout_free_page_calc(vmstats.v_page_count); 1505 1506 /* 1507 * v_free_target and v_cache_min control pageout hysteresis. Note 1508 * that these are more a measure of the VM cache queue hysteresis 1509 * then the VM free queue. Specifically, v_free_target is the 1510 * high water mark (free+cache pages). 1511 * 1512 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the 1513 * low water mark, while v_free_min is the stop. v_cache_min must 1514 * be big enough to handle memory needs while the pageout daemon 1515 * is signalled and run to free more pages. 1516 */ 1517 if (vmstats.v_free_count > 6144) 1518 vmstats.v_free_target = 4 * vmstats.v_free_min + vmstats.v_free_reserved; 1519 else 1520 vmstats.v_free_target = 2 * vmstats.v_free_min + vmstats.v_free_reserved; 1521 1522 /* 1523 * NOTE: With the new buffer cache b_act_count we want the default 1524 * inactive target to be a percentage of available memory. 1525 * 1526 * The inactive target essentially determines the minimum 1527 * number of 'temporary' pages capable of caching one-time-use 1528 * files when the VM system is otherwise full of pages 1529 * belonging to multi-time-use files or active program data. 1530 * 1531 * NOTE: The inactive target is aggressively persued only if the 1532 * inactive queue becomes too small. If the inactive queue 1533 * is large enough to satisfy page movement to free+cache 1534 * then it is repopulated more slowly from the active queue. 1535 * This allows a general inactive_target default to be set. 1536 * 1537 * There is an issue here for processes which sit mostly idle 1538 * 'overnight', such as sshd, tcsh, and X. Any movement from 1539 * the active queue will eventually cause such pages to 1540 * recycle eventually causing a lot of paging in the morning. 1541 * To reduce the incidence of this pages cycled out of the 1542 * buffer cache are moved directly to the inactive queue if 1543 * they were only used once or twice. 1544 * 1545 * The vfs.vm_cycle_point sysctl can be used to adjust this. 1546 * Increasing the value (up to 64) increases the number of 1547 * buffer recyclements which go directly to the inactive queue. 1548 */ 1549 if (vmstats.v_free_count > 2048) { 1550 vmstats.v_cache_min = vmstats.v_free_target; 1551 vmstats.v_cache_max = 2 * vmstats.v_cache_min; 1552 } else { 1553 vmstats.v_cache_min = 0; 1554 vmstats.v_cache_max = 0; 1555 } 1556 vmstats.v_inactive_target = vmstats.v_free_count / 4; 1557 1558 /* XXX does not really belong here */ 1559 if (vm_page_max_wired == 0) 1560 vm_page_max_wired = vmstats.v_free_count / 3; 1561 1562 if (vm_pageout_stats_max == 0) 1563 vm_pageout_stats_max = vmstats.v_free_target; 1564 1565 /* 1566 * Set interval in seconds for stats scan. 1567 */ 1568 if (vm_pageout_stats_interval == 0) 1569 vm_pageout_stats_interval = 5; 1570 if (vm_pageout_full_stats_interval == 0) 1571 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4; 1572 1573 1574 /* 1575 * Set maximum free per pass 1576 */ 1577 if (vm_pageout_stats_free_max == 0) 1578 vm_pageout_stats_free_max = 5; 1579 1580 swap_pager_swap_init(); 1581 pass = 0; 1582 1583 /* 1584 * The pageout daemon is never done, so loop forever. 1585 */ 1586 while (TRUE) { 1587 int error; 1588 1589 /* 1590 * Wait for an action request 1591 */ 1592 crit_enter(); 1593 if (vm_pages_needed == 0) { 1594 error = tsleep(&vm_pages_needed, 1595 0, "psleep", 1596 vm_pageout_stats_interval * hz); 1597 if (error && vm_pages_needed == 0) { 1598 vm_pageout_page_stats(); 1599 continue; 1600 } 1601 vm_pages_needed = 1; 1602 } 1603 crit_exit(); 1604 1605 /* 1606 * If we have enough free memory, wakeup waiters. 1607 * (This is optional here) 1608 */ 1609 crit_enter(); 1610 if (!vm_page_count_min(0)) 1611 wakeup(&vmstats.v_free_count); 1612 mycpu->gd_cnt.v_pdwakeups++; 1613 crit_exit(); 1614 1615 /* 1616 * Scan for pageout. Try to avoid thrashing the system 1617 * with activity. 1618 */ 1619 inactive_shortage = vm_pageout_scan(pass); 1620 if (inactive_shortage > 0) { 1621 ++pass; 1622 if (swap_pager_full) { 1623 /* 1624 * Running out of memory, catastrophic back-off 1625 * to one-second intervals. 1626 */ 1627 tsleep(&vm_pages_needed, 0, "pdelay", hz); 1628 } else if (pass < 10 && vm_pages_needed > 1) { 1629 /* 1630 * Normal operation, additional processes 1631 * have already kicked us. Retry immediately. 1632 */ 1633 } else if (pass < 10) { 1634 /* 1635 * Normal operation, fewer processes. Delay 1636 * a bit but allow wakeups. 1637 */ 1638 vm_pages_needed = 0; 1639 tsleep(&vm_pages_needed, 0, "pdelay", hz / 10); 1640 vm_pages_needed = 1; 1641 } else { 1642 /* 1643 * We've taken too many passes, forced delay. 1644 */ 1645 tsleep(&vm_pages_needed, 0, "pdelay", hz / 10); 1646 } 1647 } else { 1648 /* 1649 * Interlocked wakeup of waiters (non-optional) 1650 */ 1651 pass = 0; 1652 if (vm_pages_needed && !vm_page_count_min(0)) { 1653 wakeup(&vmstats.v_free_count); 1654 vm_pages_needed = 0; 1655 } 1656 } 1657 } 1658 } 1659 1660 /* 1661 * Called after allocating a page out of the cache or free queue 1662 * to possibly wake the pagedaemon up to replentish our supply. 1663 * 1664 * We try to generate some hysteresis by waking the pagedaemon up 1665 * when our free+cache pages go below the severe level. The pagedaemon 1666 * tries to get the count back up to at least the minimum, and through 1667 * to the target level if possible. 1668 * 1669 * If the pagedaemon is already active bump vm_pages_needed as a hint 1670 * that there are even more requests pending. 1671 * 1672 * SMP races ok? 1673 * No requirements. 1674 */ 1675 void 1676 pagedaemon_wakeup(void) 1677 { 1678 if (vm_page_count_severe() && curthread != pagethread) { 1679 if (vm_pages_needed == 0) { 1680 vm_pages_needed = 1; 1681 wakeup(&vm_pages_needed); 1682 } else if (vm_page_count_min(0)) { 1683 ++vm_pages_needed; 1684 } 1685 } 1686 } 1687 1688 #if !defined(NO_SWAPPING) 1689 1690 /* 1691 * SMP races ok? 1692 * No requirements. 1693 */ 1694 static void 1695 vm_req_vmdaemon(void) 1696 { 1697 static int lastrun = 0; 1698 1699 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1700 wakeup(&vm_daemon_needed); 1701 lastrun = ticks; 1702 } 1703 } 1704 1705 static int vm_daemon_callback(struct proc *p, void *data __unused); 1706 1707 /* 1708 * No requirements. 1709 */ 1710 static void 1711 vm_daemon(void) 1712 { 1713 /* 1714 * Permanently hold vm_token. 1715 */ 1716 lwkt_gettoken(&vm_token); 1717 1718 while (TRUE) { 1719 tsleep(&vm_daemon_needed, 0, "psleep", 0); 1720 if (vm_pageout_req_swapout) { 1721 swapout_procs(vm_pageout_req_swapout); 1722 vm_pageout_req_swapout = 0; 1723 } 1724 /* 1725 * scan the processes for exceeding their rlimits or if 1726 * process is swapped out -- deactivate pages 1727 */ 1728 allproc_scan(vm_daemon_callback, NULL); 1729 } 1730 } 1731 1732 /* 1733 * Caller must hold vm_token and proc_token. 1734 */ 1735 static int 1736 vm_daemon_callback(struct proc *p, void *data __unused) 1737 { 1738 vm_pindex_t limit, size; 1739 1740 /* 1741 * if this is a system process or if we have already 1742 * looked at this process, skip it. 1743 */ 1744 if (p->p_flag & (P_SYSTEM | P_WEXIT)) 1745 return (0); 1746 1747 /* 1748 * if the process is in a non-running type state, 1749 * don't touch it. 1750 */ 1751 if (p->p_stat != SACTIVE && p->p_stat != SSTOP) 1752 return (0); 1753 1754 /* 1755 * get a limit 1756 */ 1757 limit = OFF_TO_IDX(qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur, 1758 p->p_rlimit[RLIMIT_RSS].rlim_max)); 1759 1760 /* 1761 * let processes that are swapped out really be 1762 * swapped out. Set the limit to nothing to get as 1763 * many pages out to swap as possible. 1764 */ 1765 if (p->p_flag & P_SWAPPEDOUT) 1766 limit = 0; 1767 1768 size = vmspace_resident_count(p->p_vmspace); 1769 if (limit >= 0 && size >= limit) { 1770 vm_pageout_map_deactivate_pages( 1771 &p->p_vmspace->vm_map, limit); 1772 } 1773 return (0); 1774 } 1775 1776 #endif 1777