vm_pageout.c revision 33784
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 * $Id: vm_pageout.c,v 1.115 1998/02/23 08:22:37 dyson Exp $ 69 */ 70 71/* 72 * The proverbial page-out daemon. 73 */ 74 75#include <sys/param.h> 76#include <sys/systm.h> 77#include <sys/kernel.h> 78#include <sys/proc.h> 79#include <sys/resourcevar.h> 80#include <sys/signalvar.h> 81#include <sys/vnode.h> 82#include <sys/vmmeter.h> 83#include <sys/sysctl.h> 84 85#include <vm/vm.h> 86#include <vm/vm_param.h> 87#include <vm/vm_prot.h> 88#include <sys/lock.h> 89#include <vm/vm_object.h> 90#include <vm/vm_page.h> 91#include <vm/vm_map.h> 92#include <vm/vm_pageout.h> 93#include <vm/vm_pager.h> 94#include <vm/swap_pager.h> 95#include <vm/vm_extern.h> 96 97/* 98 * System initialization 99 */ 100 101/* the kernel process "vm_pageout"*/ 102static void vm_pageout __P((void)); 103static int vm_pageout_clean __P((vm_page_t, int)); 104static int vm_pageout_scan __P((void)); 105static int vm_pageout_free_page_calc __P((vm_size_t count)); 106struct proc *pageproc; 107 108static struct kproc_desc page_kp = { 109 "pagedaemon", 110 vm_pageout, 111 &pageproc 112}; 113SYSINIT_KT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp) 114 115#if !defined(NO_SWAPPING) 116/* the kernel process "vm_daemon"*/ 117static void vm_daemon __P((void)); 118static struct proc *vmproc; 119 120static struct kproc_desc vm_kp = { 121 "vmdaemon", 122 vm_daemon, 123 &vmproc 124}; 125SYSINIT_KT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp) 126#endif 127 128 129int vm_pages_needed=0; /* Event on which pageout daemon sleeps */ 130int vm_pageout_deficit=0; /* Estimated number of pages deficit */ 131int vm_pageout_pages_needed=0; /* flag saying that the pageout daemon needs pages */ 132 133extern int npendingio; 134#if !defined(NO_SWAPPING) 135static int vm_pageout_req_swapout; /* XXX */ 136static int vm_daemon_needed; 137#endif 138extern int nswiodone; 139extern int vm_swap_size; 140extern int vfs_update_wakeup; 141static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0; 142static int vm_pageout_full_stats_interval = 0; 143static int vm_pageout_stats_free_max=0, vm_pageout_algorithm_lru=0; 144static int defer_swap_pageouts=0; 145static int disable_swap_pageouts=0; 146 147static int max_page_launder=100; 148#if defined(NO_SWAPPING) 149static int vm_swap_enabled=0; 150static int vm_swap_idle_enabled=0; 151#else 152static int vm_swap_enabled=1; 153static int vm_swap_idle_enabled=0; 154#endif 155 156SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm, 157 CTLFLAG_RW, &vm_pageout_algorithm_lru, 0, ""); 158 159SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max, 160 CTLFLAG_RW, &vm_pageout_stats_max, 0, ""); 161 162SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval, 163 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, ""); 164 165SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval, 166 CTLFLAG_RW, &vm_pageout_stats_interval, 0, ""); 167 168SYSCTL_INT(_vm, OID_AUTO, pageout_stats_free_max, 169 CTLFLAG_RW, &vm_pageout_stats_free_max, 0, ""); 170 171#if defined(NO_SWAPPING) 172SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 173 CTLFLAG_RD, &vm_swap_enabled, 0, ""); 174SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 175 CTLFLAG_RD, &vm_swap_idle_enabled, 0, ""); 176#else 177SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 178 CTLFLAG_RW, &vm_swap_enabled, 0, ""); 179SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 180 CTLFLAG_RW, &vm_swap_idle_enabled, 0, ""); 181#endif 182 183SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts, 184 CTLFLAG_RW, &defer_swap_pageouts, 0, ""); 185 186SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts, 187 CTLFLAG_RW, &disable_swap_pageouts, 0, ""); 188 189SYSCTL_INT(_vm, OID_AUTO, max_page_launder, 190 CTLFLAG_RW, &max_page_launder, 0, ""); 191 192 193#define VM_PAGEOUT_PAGE_COUNT 8 194int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT; 195 196int vm_page_max_wired; /* XXX max # of wired pages system-wide */ 197 198#if !defined(NO_SWAPPING) 199typedef void freeer_fcn_t __P((vm_map_t, vm_object_t, vm_pindex_t, int)); 200static void vm_pageout_map_deactivate_pages __P((vm_map_t, vm_pindex_t)); 201static freeer_fcn_t vm_pageout_object_deactivate_pages; 202static void vm_req_vmdaemon __P((void)); 203#endif 204static void vm_pageout_page_stats(void); 205void pmap_collect(void); 206 207/* 208 * vm_pageout_clean: 209 * 210 * Clean the page and remove it from the laundry. 211 * 212 * We set the busy bit to cause potential page faults on this page to 213 * block. 214 * 215 * And we set pageout-in-progress to keep the object from disappearing 216 * during pageout. This guarantees that the page won't move from the 217 * inactive queue. (However, any other page on the inactive queue may 218 * move!) 219 */ 220static int 221vm_pageout_clean(m, sync) 222 vm_page_t m; 223 int sync; 224{ 225 register vm_object_t object; 226 vm_page_t mc[2*vm_pageout_page_count]; 227 int pageout_count; 228 int i, forward_okay, backward_okay, page_base; 229 vm_pindex_t pindex = m->pindex; 230 231 object = m->object; 232 233 /* 234 * If not OBJT_SWAP, additional memory may be needed to do the pageout. 235 * Try to avoid the deadlock. 236 */ 237 if ((sync != VM_PAGEOUT_FORCE) && 238 (object->type == OBJT_DEFAULT) && 239 ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_pageout_free_min)) 240 return 0; 241 242 /* 243 * Don't mess with the page if it's busy. 244 */ 245 if ((!sync && m->hold_count != 0) || 246 ((m->busy != 0) || (m->flags & PG_BUSY))) 247 return 0; 248 249 /* 250 * Try collapsing before it's too late. 251 */ 252 if (!sync && object->backing_object) { 253 vm_object_collapse(object); 254 } 255 256 mc[vm_pageout_page_count] = m; 257 pageout_count = 1; 258 page_base = vm_pageout_page_count; 259 forward_okay = TRUE; 260 if (pindex != 0) 261 backward_okay = TRUE; 262 else 263 backward_okay = FALSE; 264 /* 265 * Scan object for clusterable pages. 266 * 267 * We can cluster ONLY if: ->> the page is NOT 268 * clean, wired, busy, held, or mapped into a 269 * buffer, and one of the following: 270 * 1) The page is inactive, or a seldom used 271 * active page. 272 * -or- 273 * 2) we force the issue. 274 */ 275 for (i = 1; (i < vm_pageout_page_count) && (forward_okay || backward_okay); i++) { 276 vm_page_t p; 277 278 /* 279 * See if forward page is clusterable. 280 */ 281 if (forward_okay) { 282 /* 283 * Stop forward scan at end of object. 284 */ 285 if ((pindex + i) > object->size) { 286 forward_okay = FALSE; 287 goto do_backward; 288 } 289 p = vm_page_lookup(object, pindex + i); 290 if (p) { 291 if (((p->queue - p->pc) == PQ_CACHE) || 292 (p->flags & PG_BUSY) || p->busy) { 293 forward_okay = FALSE; 294 goto do_backward; 295 } 296 vm_page_test_dirty(p); 297 if ((p->dirty & p->valid) != 0 && 298 ((p->queue == PQ_INACTIVE) || 299 (sync == VM_PAGEOUT_FORCE)) && 300 (p->wire_count == 0) && 301 (p->hold_count == 0)) { 302 mc[vm_pageout_page_count + i] = p; 303 pageout_count++; 304 if (pageout_count == vm_pageout_page_count) 305 break; 306 } else { 307 forward_okay = FALSE; 308 } 309 } else { 310 forward_okay = FALSE; 311 } 312 } 313do_backward: 314 /* 315 * See if backward page is clusterable. 316 */ 317 if (backward_okay) { 318 /* 319 * Stop backward scan at beginning of object. 320 */ 321 if ((pindex - i) == 0) { 322 backward_okay = FALSE; 323 } 324 p = vm_page_lookup(object, pindex - i); 325 if (p) { 326 if (((p->queue - p->pc) == PQ_CACHE) || 327 (p->flags & PG_BUSY) || p->busy) { 328 backward_okay = FALSE; 329 continue; 330 } 331 vm_page_test_dirty(p); 332 if ((p->dirty & p->valid) != 0 && 333 ((p->queue == PQ_INACTIVE) || 334 (sync == VM_PAGEOUT_FORCE)) && 335 (p->wire_count == 0) && 336 (p->hold_count == 0)) { 337 mc[vm_pageout_page_count - i] = p; 338 pageout_count++; 339 page_base--; 340 if (pageout_count == vm_pageout_page_count) 341 break; 342 } else { 343 backward_okay = FALSE; 344 } 345 } else { 346 backward_okay = FALSE; 347 } 348 } 349 } 350 351 /* 352 * we allow reads during pageouts... 353 */ 354 for (i = page_base; i < (page_base + pageout_count); i++) { 355 mc[i]->flags |= PG_BUSY; 356 vm_page_protect(mc[i], VM_PROT_READ); 357 } 358 359 return vm_pageout_flush(&mc[page_base], pageout_count, sync); 360} 361 362int 363vm_pageout_flush(mc, count, sync) 364 vm_page_t *mc; 365 int count; 366 int sync; 367{ 368 register vm_object_t object; 369 int pageout_status[count]; 370 int numpagedout = 0; 371 int i; 372 373 object = mc[0]->object; 374 object->paging_in_progress += count; 375 376 vm_pager_put_pages(object, mc, count, 377 ((sync || (object == kernel_object)) ? TRUE : FALSE), 378 pageout_status); 379 380 for (i = 0; i < count; i++) { 381 vm_page_t mt = mc[i]; 382 383 switch (pageout_status[i]) { 384 case VM_PAGER_OK: 385 numpagedout++; 386 break; 387 case VM_PAGER_PEND: 388 numpagedout++; 389 break; 390 case VM_PAGER_BAD: 391 /* 392 * Page outside of range of object. Right now we 393 * essentially lose the changes by pretending it 394 * worked. 395 */ 396 pmap_clear_modify(VM_PAGE_TO_PHYS(mt)); 397 mt->dirty = 0; 398 break; 399 case VM_PAGER_ERROR: 400 case VM_PAGER_FAIL: 401 /* 402 * If page couldn't be paged out, then reactivate the 403 * page so it doesn't clog the inactive list. (We 404 * will try paging out it again later). 405 */ 406 vm_page_activate(mt); 407 break; 408 case VM_PAGER_AGAIN: 409 break; 410 } 411 412 413 /* 414 * If the operation is still going, leave the page busy to 415 * block all other accesses. Also, leave the paging in 416 * progress indicator set so that we don't attempt an object 417 * collapse. 418 */ 419 if (pageout_status[i] != VM_PAGER_PEND) { 420 vm_object_pip_wakeup(object); 421 PAGE_WAKEUP(mt); 422 } 423 } 424 return numpagedout; 425} 426 427#if !defined(NO_SWAPPING) 428/* 429 * vm_pageout_object_deactivate_pages 430 * 431 * deactivate enough pages to satisfy the inactive target 432 * requirements or if vm_page_proc_limit is set, then 433 * deactivate all of the pages in the object and its 434 * backing_objects. 435 * 436 * The object and map must be locked. 437 */ 438static void 439vm_pageout_object_deactivate_pages(map, object, desired, map_remove_only) 440 vm_map_t map; 441 vm_object_t object; 442 vm_pindex_t desired; 443 int map_remove_only; 444{ 445 register vm_page_t p, next; 446 int rcount; 447 int remove_mode; 448 int s; 449 450 if (object->type == OBJT_DEVICE) 451 return; 452 453 while (object) { 454 if (vm_map_pmap(map)->pm_stats.resident_count <= desired) 455 return; 456 if (object->paging_in_progress) 457 return; 458 459 remove_mode = map_remove_only; 460 if (object->shadow_count > 1) 461 remove_mode = 1; 462 /* 463 * scan the objects entire memory queue 464 */ 465 rcount = object->resident_page_count; 466 p = TAILQ_FIRST(&object->memq); 467 while (p && (rcount-- > 0)) { 468 int actcount; 469 if (vm_map_pmap(map)->pm_stats.resident_count <= desired) 470 return; 471 next = TAILQ_NEXT(p, listq); 472 cnt.v_pdpages++; 473 if (p->wire_count != 0 || 474 p->hold_count != 0 || 475 p->busy != 0 || 476 (p->flags & PG_BUSY) || 477 !pmap_page_exists(vm_map_pmap(map), VM_PAGE_TO_PHYS(p))) { 478 p = next; 479 continue; 480 } 481 482 actcount = pmap_ts_referenced(VM_PAGE_TO_PHYS(p)); 483 if (actcount) { 484 p->flags |= PG_REFERENCED; 485 } else if (p->flags & PG_REFERENCED) { 486 actcount = 1; 487 } 488 489 if ((p->queue != PQ_ACTIVE) && 490 (p->flags & PG_REFERENCED)) { 491 vm_page_activate(p); 492 p->act_count += actcount; 493 p->flags &= ~PG_REFERENCED; 494 } else if (p->queue == PQ_ACTIVE) { 495 if ((p->flags & PG_REFERENCED) == 0) { 496 p->act_count -= min(p->act_count, ACT_DECLINE); 497 if (!remove_mode && (vm_pageout_algorithm_lru || (p->act_count == 0))) { 498 vm_page_protect(p, VM_PROT_NONE); 499 vm_page_deactivate(p); 500 } else { 501 s = splvm(); 502 TAILQ_REMOVE(&vm_page_queue_active, p, pageq); 503 TAILQ_INSERT_TAIL(&vm_page_queue_active, p, pageq); 504 splx(s); 505 } 506 } else { 507 vm_page_activate(p); 508 p->flags &= ~PG_REFERENCED; 509 if (p->act_count < (ACT_MAX - ACT_ADVANCE)) 510 p->act_count += ACT_ADVANCE; 511 s = splvm(); 512 TAILQ_REMOVE(&vm_page_queue_active, p, pageq); 513 TAILQ_INSERT_TAIL(&vm_page_queue_active, p, pageq); 514 splx(s); 515 } 516 } else if (p->queue == PQ_INACTIVE) { 517 vm_page_protect(p, VM_PROT_NONE); 518 } 519 p = next; 520 } 521 object = object->backing_object; 522 } 523 return; 524} 525 526/* 527 * deactivate some number of pages in a map, try to do it fairly, but 528 * that is really hard to do. 529 */ 530static void 531vm_pageout_map_deactivate_pages(map, desired) 532 vm_map_t map; 533 vm_pindex_t desired; 534{ 535 vm_map_entry_t tmpe; 536 vm_object_t obj, bigobj; 537 538 if (lockmgr(&map->lock, LK_EXCLUSIVE | LK_NOWAIT, (void *)0, curproc)) { 539 return; 540 } 541 542 bigobj = NULL; 543 544 /* 545 * first, search out the biggest object, and try to free pages from 546 * that. 547 */ 548 tmpe = map->header.next; 549 while (tmpe != &map->header) { 550 if ((tmpe->eflags & (MAP_ENTRY_IS_A_MAP|MAP_ENTRY_IS_SUB_MAP)) == 0) { 551 obj = tmpe->object.vm_object; 552 if ((obj != NULL) && (obj->shadow_count <= 1) && 553 ((bigobj == NULL) || 554 (bigobj->resident_page_count < obj->resident_page_count))) { 555 bigobj = obj; 556 } 557 } 558 tmpe = tmpe->next; 559 } 560 561 if (bigobj) 562 vm_pageout_object_deactivate_pages(map, bigobj, desired, 0); 563 564 /* 565 * Next, hunt around for other pages to deactivate. We actually 566 * do this search sort of wrong -- .text first is not the best idea. 567 */ 568 tmpe = map->header.next; 569 while (tmpe != &map->header) { 570 if (vm_map_pmap(map)->pm_stats.resident_count <= desired) 571 break; 572 if ((tmpe->eflags & (MAP_ENTRY_IS_A_MAP|MAP_ENTRY_IS_SUB_MAP)) == 0) { 573 obj = tmpe->object.vm_object; 574 if (obj) 575 vm_pageout_object_deactivate_pages(map, obj, desired, 0); 576 } 577 tmpe = tmpe->next; 578 }; 579 580 /* 581 * Remove all mappings if a process is swapped out, this will free page 582 * table pages. 583 */ 584 if (desired == 0) 585 pmap_remove(vm_map_pmap(map), 586 VM_MIN_ADDRESS, VM_MAXUSER_ADDRESS); 587 vm_map_unlock(map); 588 return; 589} 590#endif 591 592void 593vm_pageout_page_free(vm_page_t m) { 594 struct vnode *vp; 595 vm_object_t object; 596 597 object = m->object; 598 object->ref_count++; 599 600 if (object->type == OBJT_VNODE) { 601 vp = object->handle; 602 vp->v_usecount++; 603 if (VSHOULDBUSY(vp)) 604 vbusy(vp); 605 } 606 607 m->flags |= PG_BUSY; 608 vm_page_protect(m, VM_PROT_NONE); 609 vm_page_free(m); 610 vm_object_deallocate(object); 611} 612 613/* 614 * vm_pageout_scan does the dirty work for the pageout daemon. 615 */ 616static int 617vm_pageout_scan() 618{ 619 vm_page_t m, next; 620 int page_shortage, addl_page_shortage, maxscan, pcount; 621 int maxlaunder; 622 int pages_freed; 623 struct proc *p, *bigproc; 624 vm_offset_t size, bigsize; 625 vm_object_t object; 626 int force_wakeup = 0; 627 int actcount; 628 int vnodes_skipped = 0; 629 int s; 630 631 /* 632 * Do whatever cleanup that the pmap code can. 633 */ 634 pmap_collect(); 635 636 /* 637 * Start scanning the inactive queue for pages we can free. We keep 638 * scanning until we have enough free pages or we have scanned through 639 * the entire queue. If we encounter dirty pages, we start cleaning 640 * them. 641 */ 642 643 pages_freed = 0; 644 addl_page_shortage = vm_pageout_deficit; 645 vm_pageout_deficit = 0; 646 647 if (max_page_launder == 0) 648 max_page_launder = 1; 649 maxlaunder = (cnt.v_inactive_target > max_page_launder) ? 650 max_page_launder : cnt.v_inactive_target; 651 652rescan0: 653 maxscan = cnt.v_inactive_count; 654 for( m = TAILQ_FIRST(&vm_page_queue_inactive); 655 656 (m != NULL) && (maxscan-- > 0) && 657 ((cnt.v_cache_count + cnt.v_free_count) < 658 (cnt.v_cache_min + cnt.v_free_target)); 659 660 m = next) { 661 662 cnt.v_pdpages++; 663 664 if (m->queue != PQ_INACTIVE) { 665 goto rescan0; 666 } 667 668 next = TAILQ_NEXT(m, pageq); 669 670 if (m->hold_count) { 671 s = splvm(); 672 TAILQ_REMOVE(&vm_page_queue_inactive, m, pageq); 673 TAILQ_INSERT_TAIL(&vm_page_queue_inactive, m, pageq); 674 splx(s); 675 addl_page_shortage++; 676 continue; 677 } 678 /* 679 * Dont mess with busy pages, keep in the front of the 680 * queue, most likely are being paged out. 681 */ 682 if (m->busy || (m->flags & PG_BUSY)) { 683 addl_page_shortage++; 684 continue; 685 } 686 687 /* 688 * If the object is not being used, we ignore previous references. 689 */ 690 if (m->object->ref_count == 0) { 691 m->flags &= ~PG_REFERENCED; 692 pmap_clear_reference(VM_PAGE_TO_PHYS(m)); 693 694 /* 695 * Otherwise, if the page has been referenced while in the inactive 696 * queue, we bump the "activation count" upwards, making it less 697 * likely that the page will be added back to the inactive queue 698 * prematurely again. Here we check the page tables (or emulated 699 * bits, if any), given the upper level VM system not knowing anything 700 * about existing references. 701 */ 702 } else if (((m->flags & PG_REFERENCED) == 0) && 703 (actcount = pmap_ts_referenced(VM_PAGE_TO_PHYS(m)))) { 704 vm_page_activate(m); 705 m->act_count += (actcount + ACT_ADVANCE); 706 continue; 707 } 708 709 /* 710 * If the upper level VM system knows about any page references, 711 * we activate the page. We also set the "activation count" higher 712 * than normal so that we will less likely place pages back onto the 713 * inactive queue again. 714 */ 715 if ((m->flags & PG_REFERENCED) != 0) { 716 m->flags &= ~PG_REFERENCED; 717 actcount = pmap_ts_referenced(VM_PAGE_TO_PHYS(m)); 718 vm_page_activate(m); 719 m->act_count += (actcount + ACT_ADVANCE + 1); 720 continue; 721 } 722 723 /* 724 * If the upper level VM system doesn't know anything about the 725 * page being dirty, we have to check for it again. As far as the 726 * VM code knows, any partially dirty pages are fully dirty. 727 */ 728 if (m->dirty == 0) { 729 vm_page_test_dirty(m); 730 } else if (m->dirty != 0) { 731 m->dirty = VM_PAGE_BITS_ALL; 732 } 733 734 /* 735 * Invalid pages can be easily freed 736 */ 737 if (m->valid == 0) { 738 vm_pageout_page_free(m); 739 cnt.v_dfree++; 740 pages_freed++; 741 742 /* 743 * Clean pages can be placed onto the cache queue. 744 */ 745 } else if (m->dirty == 0) { 746 vm_page_cache(m); 747 pages_freed++; 748 749 /* 750 * Dirty pages need to be paged out. Note that we clean 751 * only a limited number of pages per pagedaemon pass. 752 */ 753 } else if (maxlaunder > 0) { 754 int written; 755 int swap_pageouts_ok; 756 struct vnode *vp = NULL; 757 758 object = m->object; 759 760 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 761 swap_pageouts_ok = 1; 762 } else { 763 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 764 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 765 (cnt.v_free_count + cnt.v_cache_count) < cnt.v_free_min); 766 767 } 768 769 /* 770 * We don't bother paging objects that are "dead". Those 771 * objects are in a "rundown" state. 772 */ 773 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 774 s = splvm(); 775 TAILQ_REMOVE(&vm_page_queue_inactive, m, pageq); 776 TAILQ_INSERT_TAIL(&vm_page_queue_inactive, m, pageq); 777 splx(s); 778 continue; 779 } 780 781 if ((object->type == OBJT_VNODE) && 782 (object->flags & OBJ_DEAD) == 0) { 783 vp = object->handle; 784 if (VOP_ISLOCKED(vp) || 785 vget(vp, LK_EXCLUSIVE|LK_NOOBJ, curproc)) { 786 if ((m->queue == PQ_INACTIVE) && 787 (m->hold_count == 0) && 788 (m->busy == 0) && 789 (m->flags & PG_BUSY) == 0) { 790 s = splvm(); 791 TAILQ_REMOVE(&vm_page_queue_inactive, m, pageq); 792 TAILQ_INSERT_TAIL(&vm_page_queue_inactive, m, pageq); 793 splx(s); 794 } 795 if (object->flags & OBJ_MIGHTBEDIRTY) 796 vnodes_skipped++; 797 continue; 798 } 799 800 /* 801 * The page might have been moved to another queue 802 * during potential blocking in vget() above. 803 */ 804 if (m->queue != PQ_INACTIVE) { 805 if (object->flags & OBJ_MIGHTBEDIRTY) 806 vnodes_skipped++; 807 vput(vp); 808 continue; 809 } 810 811 /* 812 * The page may have been busied during the blocking in 813 * vput(); We don't move the page back onto the end of 814 * the queue so that statistics are more correct if we don't. 815 */ 816 if (m->busy || (m->flags & PG_BUSY)) { 817 vput(vp); 818 continue; 819 } 820 821 /* 822 * If the page has become held, then skip it 823 */ 824 if (m->hold_count) { 825 s = splvm(); 826 TAILQ_REMOVE(&vm_page_queue_inactive, m, pageq); 827 TAILQ_INSERT_TAIL(&vm_page_queue_inactive, m, pageq); 828 splx(s); 829 if (object->flags & OBJ_MIGHTBEDIRTY) 830 vnodes_skipped++; 831 vput(vp); 832 continue; 833 } 834 } 835 836 /* 837 * If a page is dirty, then it is either being washed 838 * (but not yet cleaned) or it is still in the 839 * laundry. If it is still in the laundry, then we 840 * start the cleaning operation. 841 */ 842 written = vm_pageout_clean(m, 0); 843 if (vp) 844 vput(vp); 845 846 maxlaunder -= written; 847 } 848 } 849 850 /* 851 * Compute the page shortage. If we are still very low on memory be 852 * sure that we will move a minimal amount of pages from active to 853 * inactive. 854 */ 855 page_shortage = (cnt.v_inactive_target + cnt.v_cache_min) - 856 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count); 857 if (page_shortage <= 0) { 858 if (pages_freed == 0) { 859 page_shortage = cnt.v_free_min - cnt.v_free_count; 860 } else { 861 page_shortage = 1; 862 } 863 } 864 865 /* 866 * If the "inactive" loop finds that there is a shortage over and 867 * above the page statistics variables, then we need to accomodate 868 * that. This avoids potential deadlocks due to pages being temporarily 869 * busy for I/O or other types of temporary wiring. 870 */ 871 if (addl_page_shortage) { 872 if (page_shortage < 0) 873 page_shortage = 0; 874 page_shortage += addl_page_shortage; 875 } 876 877 pcount = cnt.v_active_count; 878 m = TAILQ_FIRST(&vm_page_queue_active); 879 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) { 880 881 /* 882 * This is a consistancy check, and should likely be a panic 883 * or warning. 884 */ 885 if (m->queue != PQ_ACTIVE) { 886 break; 887 } 888 889 next = TAILQ_NEXT(m, pageq); 890 /* 891 * Don't deactivate pages that are busy. 892 */ 893 if ((m->busy != 0) || 894 (m->flags & PG_BUSY) || 895 (m->hold_count != 0)) { 896 s = splvm(); 897 TAILQ_REMOVE(&vm_page_queue_active, m, pageq); 898 TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq); 899 splx(s); 900 m = next; 901 continue; 902 } 903 904 /* 905 * The count for pagedaemon pages is done after checking the 906 * page for eligbility... 907 */ 908 cnt.v_pdpages++; 909 910 /* 911 * Check to see "how much" the page has been used. 912 */ 913 actcount = 0; 914 if (m->object->ref_count != 0) { 915 if (m->flags & PG_REFERENCED) { 916 actcount += 1; 917 } 918 actcount += pmap_ts_referenced(VM_PAGE_TO_PHYS(m)); 919 if (actcount) { 920 m->act_count += ACT_ADVANCE + actcount; 921 if (m->act_count > ACT_MAX) 922 m->act_count = ACT_MAX; 923 } 924 } 925 926 /* 927 * Since we have "tested" this bit, we need to clear it now. 928 */ 929 m->flags &= ~PG_REFERENCED; 930 931 /* 932 * Only if an object is currently being used, do we use the 933 * page activation count stats. 934 */ 935 if (actcount && (m->object->ref_count != 0)) { 936 s = splvm(); 937 TAILQ_REMOVE(&vm_page_queue_active, m, pageq); 938 TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq); 939 splx(s); 940 } else { 941 m->act_count -= min(m->act_count, ACT_DECLINE); 942 if (vm_pageout_algorithm_lru || 943 (m->object->ref_count == 0) || (m->act_count == 0)) { 944 page_shortage--; 945 if (m->object->ref_count == 0) { 946 vm_page_protect(m, VM_PROT_NONE); 947 if (m->dirty == 0) 948 vm_page_cache(m); 949 else 950 vm_page_deactivate(m); 951 } else { 952 vm_page_deactivate(m); 953 } 954 } else { 955 s = splvm(); 956 TAILQ_REMOVE(&vm_page_queue_active, m, pageq); 957 TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq); 958 splx(s); 959 } 960 } 961 m = next; 962 } 963 964 s = splvm(); 965 /* 966 * We try to maintain some *really* free pages, this allows interrupt 967 * code to be guaranteed space. 968 */ 969 while (cnt.v_free_count < cnt.v_free_reserved) { 970 static int cache_rover = 0; 971 m = vm_page_list_find(PQ_CACHE, cache_rover); 972 if (!m) 973 break; 974 cache_rover = (cache_rover + PQ_PRIME2) & PQ_L2_MASK; 975 vm_pageout_page_free(m); 976 cnt.v_dfree++; 977 } 978 splx(s); 979 980#if !defined(NO_SWAPPING) 981 /* 982 * Idle process swapout -- run once per second. 983 */ 984 if (vm_swap_idle_enabled) { 985 static long lsec; 986 if (time.tv_sec != lsec) { 987 vm_pageout_req_swapout |= VM_SWAP_IDLE; 988 vm_req_vmdaemon(); 989 lsec = time.tv_sec; 990 } 991 } 992#endif 993 994 /* 995 * If we didn't get enough free pages, and we have skipped a vnode 996 * in a writeable object, wakeup the sync daemon. And kick swapout 997 * if we did not get enough free pages. 998 */ 999 if ((cnt.v_cache_count + cnt.v_free_count) < 1000 (cnt.v_free_target + cnt.v_cache_min) ) { 1001 if (vnodes_skipped && 1002 (cnt.v_cache_count + cnt.v_free_count) < cnt.v_free_min) { 1003 if (!vfs_update_wakeup) { 1004 vfs_update_wakeup = 1; 1005 wakeup(&vfs_update_wakeup); 1006 } 1007 } 1008#if !defined(NO_SWAPPING) 1009 if (vm_swap_enabled && 1010 (cnt.v_free_count + cnt.v_cache_count < cnt.v_free_target)) { 1011 vm_req_vmdaemon(); 1012 vm_pageout_req_swapout |= VM_SWAP_NORMAL; 1013 } 1014#endif 1015 } 1016 1017 1018 /* 1019 * make sure that we have swap space -- if we are low on memory and 1020 * swap -- then kill the biggest process. 1021 */ 1022 if ((vm_swap_size == 0 || swap_pager_full) && 1023 ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_free_min)) { 1024 bigproc = NULL; 1025 bigsize = 0; 1026 for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) { 1027 /* 1028 * if this is a system process, skip it 1029 */ 1030 if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) || 1031 ((p->p_pid < 48) && (vm_swap_size != 0))) { 1032 continue; 1033 } 1034 /* 1035 * if the process is in a non-running type state, 1036 * don't touch it. 1037 */ 1038 if (p->p_stat != SRUN && p->p_stat != SSLEEP) { 1039 continue; 1040 } 1041 /* 1042 * get the process size 1043 */ 1044 size = p->p_vmspace->vm_pmap.pm_stats.resident_count; 1045 /* 1046 * if the this process is bigger than the biggest one 1047 * remember it. 1048 */ 1049 if (size > bigsize) { 1050 bigproc = p; 1051 bigsize = size; 1052 } 1053 } 1054 if (bigproc != NULL) { 1055 killproc(bigproc, "out of swap space"); 1056 bigproc->p_estcpu = 0; 1057 bigproc->p_nice = PRIO_MIN; 1058 resetpriority(bigproc); 1059 wakeup(&cnt.v_free_count); 1060 } 1061 } 1062 return force_wakeup; 1063} 1064 1065/* 1066 * This routine tries to maintain the pseudo LRU active queue, 1067 * so that during long periods of time where there is no paging, 1068 * that some statistic accumlation still occurs. This code 1069 * helps the situation where paging just starts to occur. 1070 */ 1071static void 1072vm_pageout_page_stats() 1073{ 1074 int s; 1075 vm_page_t m,next; 1076 int pcount,tpcount; /* Number of pages to check */ 1077 static int fullintervalcount = 0; 1078 1079 pcount = cnt.v_active_count; 1080 fullintervalcount += vm_pageout_stats_interval; 1081 if (fullintervalcount < vm_pageout_full_stats_interval) { 1082 tpcount = (vm_pageout_stats_max * cnt.v_active_count) / cnt.v_page_count; 1083 if (pcount > tpcount) 1084 pcount = tpcount; 1085 } 1086 1087 m = TAILQ_FIRST(&vm_page_queue_active); 1088 while ((m != NULL) && (pcount-- > 0)) { 1089 int actcount; 1090 1091 if (m->queue != PQ_ACTIVE) { 1092 break; 1093 } 1094 1095 next = TAILQ_NEXT(m, pageq); 1096 /* 1097 * Don't deactivate pages that are busy. 1098 */ 1099 if ((m->busy != 0) || 1100 (m->flags & PG_BUSY) || 1101 (m->hold_count != 0)) { 1102 s = splvm(); 1103 TAILQ_REMOVE(&vm_page_queue_active, m, pageq); 1104 TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq); 1105 splx(s); 1106 m = next; 1107 continue; 1108 } 1109 1110 actcount = 0; 1111 if (m->flags & PG_REFERENCED) { 1112 m->flags &= ~PG_REFERENCED; 1113 actcount += 1; 1114 } 1115 1116 actcount += pmap_ts_referenced(VM_PAGE_TO_PHYS(m)); 1117 if (actcount) { 1118 m->act_count += ACT_ADVANCE + actcount; 1119 if (m->act_count > ACT_MAX) 1120 m->act_count = ACT_MAX; 1121 s = splvm(); 1122 TAILQ_REMOVE(&vm_page_queue_active, m, pageq); 1123 TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq); 1124 splx(s); 1125 } else { 1126 if (m->act_count == 0) { 1127 /* 1128 * We turn off page access, so that we have more accurate 1129 * RSS stats. We don't do this in the normal page deactivation 1130 * when the system is loaded VM wise, because the cost of 1131 * the large number of page protect operations would be higher 1132 * than the value of doing the operation. 1133 */ 1134 vm_page_protect(m, VM_PROT_NONE); 1135 vm_page_deactivate(m); 1136 } else { 1137 m->act_count -= min(m->act_count, ACT_DECLINE); 1138 s = splvm(); 1139 TAILQ_REMOVE(&vm_page_queue_active, m, pageq); 1140 TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq); 1141 splx(s); 1142 } 1143 } 1144 1145 m = next; 1146 } 1147} 1148 1149static int 1150vm_pageout_free_page_calc(count) 1151vm_size_t count; 1152{ 1153 if (count < cnt.v_page_count) 1154 return 0; 1155 /* 1156 * free_reserved needs to include enough for the largest swap pager 1157 * structures plus enough for any pv_entry structs when paging. 1158 */ 1159 if (cnt.v_page_count > 1024) 1160 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200; 1161 else 1162 cnt.v_free_min = 4; 1163 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1164 cnt.v_interrupt_free_min; 1165 cnt.v_free_reserved = vm_pageout_page_count + 1166 cnt.v_pageout_free_min + (count / 768) + PQ_L2_SIZE; 1167 cnt.v_free_min += cnt.v_free_reserved; 1168 return 1; 1169} 1170 1171 1172/* 1173 * vm_pageout is the high level pageout daemon. 1174 */ 1175static void 1176vm_pageout() 1177{ 1178 /* 1179 * Initialize some paging parameters. 1180 */ 1181 1182 cnt.v_interrupt_free_min = 2; 1183 if (cnt.v_page_count < 2000) 1184 vm_pageout_page_count = 8; 1185 1186 vm_pageout_free_page_calc(cnt.v_page_count); 1187 /* 1188 * free_reserved needs to include enough for the largest swap pager 1189 * structures plus enough for any pv_entry structs when paging. 1190 */ 1191 if (cnt.v_free_count > 6144) 1192 cnt.v_free_target = 3 * cnt.v_free_min + cnt.v_free_reserved; 1193 else 1194 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved; 1195 1196 if (cnt.v_free_count > 2048) { 1197 cnt.v_cache_min = cnt.v_free_target; 1198 cnt.v_cache_max = 2 * cnt.v_cache_min; 1199 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2; 1200 } else { 1201 cnt.v_cache_min = 0; 1202 cnt.v_cache_max = 0; 1203 cnt.v_inactive_target = cnt.v_free_count / 4; 1204 } 1205 if (cnt.v_inactive_target > cnt.v_free_count / 3) 1206 cnt.v_inactive_target = cnt.v_free_count / 3; 1207 1208 /* XXX does not really belong here */ 1209 if (vm_page_max_wired == 0) 1210 vm_page_max_wired = cnt.v_free_count / 3; 1211 1212 if (vm_pageout_stats_max == 0) 1213 vm_pageout_stats_max = cnt.v_free_target; 1214 1215 /* 1216 * Set interval in seconds for stats scan. 1217 */ 1218 if (vm_pageout_stats_interval == 0) 1219 vm_pageout_stats_interval = 4; 1220 if (vm_pageout_full_stats_interval == 0) 1221 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4; 1222 1223 1224 /* 1225 * Set maximum free per pass 1226 */ 1227 if (vm_pageout_stats_free_max == 0) 1228 vm_pageout_stats_free_max = 25; 1229 1230 max_page_launder = (cnt.v_page_count > 1800 ? 32 : 16); 1231 1232 swap_pager_swap_init(); 1233 /* 1234 * The pageout daemon is never done, so loop forever. 1235 */ 1236 while (TRUE) { 1237 int inactive_target; 1238 int error; 1239 int s = splvm(); 1240 if (!vm_pages_needed || 1241 ((cnt.v_free_count + cnt.v_cache_count) > cnt.v_free_min)) { 1242 vm_pages_needed = 0; 1243 error = tsleep(&vm_pages_needed, 1244 PVM, "psleep", vm_pageout_stats_interval * hz); 1245 if (error && !vm_pages_needed) { 1246 splx(s); 1247 vm_pageout_page_stats(); 1248 continue; 1249 } 1250 } else if (vm_pages_needed) { 1251 vm_pages_needed = 0; 1252 tsleep(&vm_pages_needed, PVM, "psleep", hz/2); 1253 } 1254 1255 if (vm_pages_needed) 1256 cnt.v_pdwakeups++; 1257 vm_pages_needed = 0; 1258 splx(s); 1259 vm_pager_sync(); 1260 vm_pageout_scan(); 1261 vm_pageout_deficit = 0; 1262 vm_pager_sync(); 1263 wakeup(&cnt.v_free_count); 1264 } 1265} 1266 1267void 1268pagedaemon_wakeup() 1269{ 1270 if (!vm_pages_needed && curproc != pageproc) { 1271 vm_pages_needed++; 1272 wakeup(&vm_pages_needed); 1273 } 1274} 1275 1276#if !defined(NO_SWAPPING) 1277static void 1278vm_req_vmdaemon() 1279{ 1280 static int lastrun = 0; 1281 1282 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1283 wakeup(&vm_daemon_needed); 1284 lastrun = ticks; 1285 } 1286} 1287 1288static void 1289vm_daemon() 1290{ 1291 vm_object_t object; 1292 struct proc *p; 1293 1294 while (TRUE) { 1295 tsleep(&vm_daemon_needed, PUSER, "psleep", 0); 1296 if (vm_pageout_req_swapout) { 1297 swapout_procs(vm_pageout_req_swapout); 1298 vm_pageout_req_swapout = 0; 1299 } 1300 /* 1301 * scan the processes for exceeding their rlimits or if 1302 * process is swapped out -- deactivate pages 1303 */ 1304 1305 for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) { 1306 quad_t limit; 1307 vm_offset_t size; 1308 1309 /* 1310 * if this is a system process or if we have already 1311 * looked at this process, skip it. 1312 */ 1313 if (p->p_flag & (P_SYSTEM | P_WEXIT)) { 1314 continue; 1315 } 1316 /* 1317 * if the process is in a non-running type state, 1318 * don't touch it. 1319 */ 1320 if (p->p_stat != SRUN && p->p_stat != SSLEEP) { 1321 continue; 1322 } 1323 /* 1324 * get a limit 1325 */ 1326 limit = qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur, 1327 p->p_rlimit[RLIMIT_RSS].rlim_max); 1328 1329 /* 1330 * let processes that are swapped out really be 1331 * swapped out set the limit to nothing (will force a 1332 * swap-out.) 1333 */ 1334 if ((p->p_flag & P_INMEM) == 0) 1335 limit = 0; /* XXX */ 1336 1337 size = p->p_vmspace->vm_pmap.pm_stats.resident_count * PAGE_SIZE; 1338 if (limit >= 0 && size >= limit) { 1339 vm_pageout_map_deactivate_pages(&p->p_vmspace->vm_map, 1340 (vm_pindex_t)(limit >> PAGE_SHIFT) ); 1341 } 1342 } 1343 } 1344} 1345#endif 1346