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 * $FreeBSD$ 69 */ 70 71 /* 72 * The proverbial page-out daemon. 73 */ 74 75 #include "opt_vm.h" 76 #include <sys/param.h> 77 #include <sys/systm.h> 78 #include <sys/kernel.h> 79 #include <sys/lock.h> 80 #include <sys/mutex.h> 81 #include <sys/proc.h> 82 #include <sys/kthread.h> 83 #include <sys/ktr.h> 84 #include <sys/resourcevar.h> 85 #include <sys/signalvar.h> 86 #include <sys/vnode.h> 87 #include <sys/vmmeter.h> 88 #include <sys/sx.h> 89 #include <sys/sysctl.h> 90 91 #include <vm/vm.h> 92 #include <vm/vm_param.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/vm_zone.h> 99 #include <vm/swap_pager.h> 100 #include <vm/vm_extern.h> 101 102 #include <machine/mutex.h> 103 104 /* 105 * System initialization 106 */ 107 108 /* the kernel process "vm_pageout"*/ 109 static void vm_pageout __P((void)); 110 static int vm_pageout_clean __P((vm_page_t)); 111 static void vm_pageout_scan __P((int pass)); 112 static int vm_pageout_free_page_calc __P((vm_size_t count)); 113 struct proc *pageproc; 114 115 static struct kproc_desc page_kp = { 116 "pagedaemon", 117 vm_pageout, 118 &pageproc 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 __P((void)); 125 static struct proc *vmproc; 126 127 static struct kproc_desc vm_kp = { 128 "vmdaemon", 129 vm_daemon, 130 &vmproc 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 extern int vm_swap_size; 145 static int vm_max_launder = 32; 146 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0; 147 static int vm_pageout_full_stats_interval = 0; 148 static int vm_pageout_stats_free_max=0, vm_pageout_algorithm=0; 149 static int defer_swap_pageouts=0; 150 static int disable_swap_pageouts=0; 151 152 #if defined(NO_SWAPPING) 153 static int vm_swap_enabled=0; 154 static int vm_swap_idle_enabled=0; 155 #else 156 static int vm_swap_enabled=1; 157 static int vm_swap_idle_enabled=0; 158 #endif 159 160 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm, 161 CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt"); 162 163 SYSCTL_INT(_vm, OID_AUTO, max_launder, 164 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout"); 165 166 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max, 167 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length"); 168 169 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval, 170 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan"); 171 172 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval, 173 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan"); 174 175 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_free_max, 176 CTLFLAG_RW, &vm_pageout_stats_free_max, 0, "Not implemented"); 177 178 #if defined(NO_SWAPPING) 179 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 180 CTLFLAG_RD, &vm_swap_enabled, 0, ""); 181 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 182 CTLFLAG_RD, &vm_swap_idle_enabled, 0, ""); 183 #else 184 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 185 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout"); 186 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 187 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 188 #endif 189 190 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts, 191 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem"); 192 193 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts, 194 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages"); 195 196 #define VM_PAGEOUT_PAGE_COUNT 16 197 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT; 198 199 int vm_page_max_wired; /* XXX max # of wired pages system-wide */ 200 201 #if !defined(NO_SWAPPING) 202 typedef void freeer_fcn_t __P((vm_map_t, vm_object_t, vm_pindex_t, int)); 203 static void vm_pageout_map_deactivate_pages __P((vm_map_t, vm_pindex_t)); 204 static freeer_fcn_t vm_pageout_object_deactivate_pages; 205 static void vm_req_vmdaemon __P((void)); 206 #endif 207 static void vm_pageout_page_stats(void); 208 209 /* 210 * vm_pageout_clean: 211 * 212 * Clean the page and remove it from the laundry. 213 * 214 * We set the busy bit to cause potential page faults on this page to 215 * block. Note the careful timing, however, the busy bit isn't set till 216 * late and we cannot do anything that will mess with the page. 217 */ 218 219 static int 220 vm_pageout_clean(m) 221 vm_page_t m; 222 { 223 vm_object_t object; 224 vm_page_t mc[2*vm_pageout_page_count]; 225 int pageout_count; 226 int ib, is, page_base; 227 vm_pindex_t pindex = m->pindex; 228 229 GIANT_REQUIRED; 230 231 object = m->object; 232 233 /* 234 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP 235 * with the new swapper, but we could have serious problems paging 236 * out other object types if there is insufficient memory. 237 * 238 * Unfortunately, checking free memory here is far too late, so the 239 * check has been moved up a procedural level. 240 */ 241 242 /* 243 * Don't mess with the page if it's busy, held, or special 244 */ 245 if ((m->hold_count != 0) || 246 ((m->busy != 0) || (m->flags & (PG_BUSY|PG_UNMANAGED)))) { 247 return 0; 248 } 249 250 mc[vm_pageout_page_count] = m; 251 pageout_count = 1; 252 page_base = vm_pageout_page_count; 253 ib = 1; 254 is = 1; 255 256 /* 257 * Scan object for clusterable pages. 258 * 259 * We can cluster ONLY if: ->> the page is NOT 260 * clean, wired, busy, held, or mapped into a 261 * buffer, and one of the following: 262 * 1) The page is inactive, or a seldom used 263 * active page. 264 * -or- 265 * 2) we force the issue. 266 * 267 * During heavy mmap/modification loads the pageout 268 * daemon can really fragment the underlying file 269 * due to flushing pages out of order and not trying 270 * align the clusters (which leave sporatic out-of-order 271 * holes). To solve this problem we do the reverse scan 272 * first and attempt to align our cluster, then do a 273 * forward scan if room remains. 274 */ 275 276 more: 277 while (ib && pageout_count < vm_pageout_page_count) { 278 vm_page_t p; 279 280 if (ib > pindex) { 281 ib = 0; 282 break; 283 } 284 285 if ((p = vm_page_lookup(object, pindex - ib)) == NULL) { 286 ib = 0; 287 break; 288 } 289 if (((p->queue - p->pc) == PQ_CACHE) || 290 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) { 291 ib = 0; 292 break; 293 } 294 vm_page_test_dirty(p); 295 if ((p->dirty & p->valid) == 0 || 296 p->queue != PQ_INACTIVE || 297 p->wire_count != 0 || 298 p->hold_count != 0) { 299 ib = 0; 300 break; 301 } 302 mc[--page_base] = p; 303 ++pageout_count; 304 ++ib; 305 /* 306 * alignment boundry, stop here and switch directions. Do 307 * not clear ib. 308 */ 309 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0) 310 break; 311 } 312 313 while (pageout_count < vm_pageout_page_count && 314 pindex + is < object->size) { 315 vm_page_t p; 316 317 if ((p = vm_page_lookup(object, pindex + is)) == NULL) 318 break; 319 if (((p->queue - p->pc) == PQ_CACHE) || 320 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) { 321 break; 322 } 323 vm_page_test_dirty(p); 324 if ((p->dirty & p->valid) == 0 || 325 p->queue != PQ_INACTIVE || 326 p->wire_count != 0 || 327 p->hold_count != 0) { 328 break; 329 } 330 mc[page_base + pageout_count] = p; 331 ++pageout_count; 332 ++is; 333 } 334 335 /* 336 * If we exhausted our forward scan, continue with the reverse scan 337 * when possible, even past a page boundry. This catches boundry 338 * conditions. 339 */ 340 if (ib && pageout_count < vm_pageout_page_count) 341 goto more; 342 343 /* 344 * we allow reads during pageouts... 345 */ 346 return vm_pageout_flush(&mc[page_base], pageout_count, 0); 347 } 348 349 /* 350 * vm_pageout_flush() - launder the given pages 351 * 352 * The given pages are laundered. Note that we setup for the start of 353 * I/O ( i.e. busy the page ), mark it read-only, and bump the object 354 * reference count all in here rather then in the parent. If we want 355 * the parent to do more sophisticated things we may have to change 356 * the ordering. 357 */ 358 359 int 360 vm_pageout_flush(mc, count, flags) 361 vm_page_t *mc; 362 int count; 363 int flags; 364 { 365 vm_object_t object; 366 int pageout_status[count]; 367 int numpagedout = 0; 368 int i; 369 370 GIANT_REQUIRED; 371 /* 372 * Initiate I/O. Bump the vm_page_t->busy counter and 373 * mark the pages read-only. 374 * 375 * We do not have to fixup the clean/dirty bits here... we can 376 * allow the pager to do it after the I/O completes. 377 * 378 * NOTE! mc[i]->dirty may be partial or fragmented due to an 379 * edge case with file fragments. 380 */ 381 382 for (i = 0; i < count; i++) { 383 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, ("vm_pageout_flush page %p index %d/%d: partially invalid page", mc[i], i, count)); 384 vm_page_io_start(mc[i]); 385 vm_page_protect(mc[i], VM_PROT_READ); 386 } 387 388 object = mc[0]->object; 389 vm_object_pip_add(object, count); 390 391 vm_pager_put_pages(object, mc, count, 392 (flags | ((object == kernel_object) ? OBJPC_SYNC : 0)), 393 pageout_status); 394 395 for (i = 0; i < count; i++) { 396 vm_page_t mt = mc[i]; 397 398 switch (pageout_status[i]) { 399 case VM_PAGER_OK: 400 numpagedout++; 401 break; 402 case VM_PAGER_PEND: 403 numpagedout++; 404 break; 405 case VM_PAGER_BAD: 406 /* 407 * Page outside of range of object. Right now we 408 * essentially lose the changes by pretending it 409 * worked. 410 */ 411 pmap_clear_modify(mt); 412 vm_page_undirty(mt); 413 break; 414 case VM_PAGER_ERROR: 415 case VM_PAGER_FAIL: 416 /* 417 * If page couldn't be paged out, then reactivate the 418 * page so it doesn't clog the inactive list. (We 419 * will try paging out it again later). 420 */ 421 vm_page_activate(mt); 422 break; 423 case VM_PAGER_AGAIN: 424 break; 425 } 426 427 /* 428 * If the operation is still going, leave the page busy to 429 * block all other accesses. Also, leave the paging in 430 * progress indicator set so that we don't attempt an object 431 * collapse. 432 */ 433 if (pageout_status[i] != VM_PAGER_PEND) { 434 vm_object_pip_wakeup(object); 435 vm_page_io_finish(mt); 436 if (!vm_page_count_severe() || !vm_page_try_to_cache(mt)) 437 vm_page_protect(mt, VM_PROT_READ); 438 } 439 } 440 return numpagedout; 441 } 442 443 #if !defined(NO_SWAPPING) 444 /* 445 * vm_pageout_object_deactivate_pages 446 * 447 * deactivate enough pages to satisfy the inactive target 448 * requirements or if vm_page_proc_limit is set, then 449 * deactivate all of the pages in the object and its 450 * backing_objects. 451 * 452 * The object and map must be locked. 453 */ 454 static void 455 vm_pageout_object_deactivate_pages(map, object, desired, map_remove_only) 456 vm_map_t map; 457 vm_object_t object; 458 vm_pindex_t desired; 459 int map_remove_only; 460 { 461 vm_page_t p, next; 462 int rcount; 463 int remove_mode; 464 465 GIANT_REQUIRED; 466 if (object->type == OBJT_DEVICE || object->type == OBJT_PHYS) 467 return; 468 469 while (object) { 470 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 471 return; 472 if (object->paging_in_progress) 473 return; 474 475 remove_mode = map_remove_only; 476 if (object->shadow_count > 1) 477 remove_mode = 1; 478 /* 479 * scan the objects entire memory queue 480 */ 481 rcount = object->resident_page_count; 482 p = TAILQ_FIRST(&object->memq); 483 while (p && (rcount-- > 0)) { 484 int actcount; 485 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 486 return; 487 next = TAILQ_NEXT(p, listq); 488 cnt.v_pdpages++; 489 if (p->wire_count != 0 || 490 p->hold_count != 0 || 491 p->busy != 0 || 492 (p->flags & (PG_BUSY|PG_UNMANAGED)) || 493 !pmap_page_exists(vm_map_pmap(map), p)) { 494 p = next; 495 continue; 496 } 497 498 actcount = pmap_ts_referenced(p); 499 if (actcount) { 500 vm_page_flag_set(p, PG_REFERENCED); 501 } else if (p->flags & PG_REFERENCED) { 502 actcount = 1; 503 } 504 505 if ((p->queue != PQ_ACTIVE) && 506 (p->flags & PG_REFERENCED)) { 507 vm_page_activate(p); 508 p->act_count += actcount; 509 vm_page_flag_clear(p, PG_REFERENCED); 510 } else if (p->queue == PQ_ACTIVE) { 511 if ((p->flags & PG_REFERENCED) == 0) { 512 p->act_count -= min(p->act_count, ACT_DECLINE); 513 if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) { 514 vm_page_protect(p, VM_PROT_NONE); 515 vm_page_deactivate(p); 516 } else { 517 vm_pageq_requeue(p); 518 } 519 } else { 520 vm_page_activate(p); 521 vm_page_flag_clear(p, PG_REFERENCED); 522 if (p->act_count < (ACT_MAX - ACT_ADVANCE)) 523 p->act_count += ACT_ADVANCE; 524 vm_pageq_requeue(p); 525 } 526 } else if (p->queue == PQ_INACTIVE) { 527 vm_page_protect(p, VM_PROT_NONE); 528 } 529 p = next; 530 } 531 object = object->backing_object; 532 } 533 return; 534 } 535 536 /* 537 * deactivate some number of pages in a map, try to do it fairly, but 538 * that is really hard to do. 539 */ 540 static void 541 vm_pageout_map_deactivate_pages(map, desired) 542 vm_map_t map; 543 vm_pindex_t desired; 544 { 545 vm_map_entry_t tmpe; 546 vm_object_t obj, bigobj; 547 548 GIANT_REQUIRED; 549 if (lockmgr(&map->lock, LK_EXCLUSIVE | LK_NOWAIT, (void *)0, curthread)) { 550 return; 551 } 552 553 bigobj = NULL; 554 555 /* 556 * first, search out the biggest object, and try to free pages from 557 * that. 558 */ 559 tmpe = map->header.next; 560 while (tmpe != &map->header) { 561 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 562 obj = tmpe->object.vm_object; 563 if ((obj != NULL) && (obj->shadow_count <= 1) && 564 ((bigobj == NULL) || 565 (bigobj->resident_page_count < obj->resident_page_count))) { 566 bigobj = obj; 567 } 568 } 569 tmpe = tmpe->next; 570 } 571 572 if (bigobj) 573 vm_pageout_object_deactivate_pages(map, bigobj, desired, 0); 574 575 /* 576 * Next, hunt around for other pages to deactivate. We actually 577 * do this search sort of wrong -- .text first is not the best idea. 578 */ 579 tmpe = map->header.next; 580 while (tmpe != &map->header) { 581 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 582 break; 583 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 584 obj = tmpe->object.vm_object; 585 if (obj) 586 vm_pageout_object_deactivate_pages(map, obj, desired, 0); 587 } 588 tmpe = tmpe->next; 589 }; 590 591 /* 592 * Remove all mappings if a process is swapped out, this will free page 593 * table pages. 594 */ 595 if (desired == 0) 596 pmap_remove(vm_map_pmap(map), 597 VM_MIN_ADDRESS, VM_MAXUSER_ADDRESS); 598 vm_map_unlock(map); 599 return; 600 } 601 #endif 602 603 /* 604 * Don't try to be fancy - being fancy can lead to VOP_LOCK's and therefore 605 * to vnode deadlocks. We only do it for OBJT_DEFAULT and OBJT_SWAP objects 606 * which we know can be trivially freed. 607 */ 608 609 void 610 vm_pageout_page_free(vm_page_t m) { 611 vm_object_t object = m->object; 612 int type = object->type; 613 614 GIANT_REQUIRED; 615 if (type == OBJT_SWAP || type == OBJT_DEFAULT) 616 vm_object_reference(object); 617 vm_page_busy(m); 618 vm_page_protect(m, VM_PROT_NONE); 619 vm_page_free(m); 620 if (type == OBJT_SWAP || type == OBJT_DEFAULT) 621 vm_object_deallocate(object); 622 } 623 624 /* 625 * vm_pageout_scan does the dirty work for the pageout daemon. 626 */ 627 static void 628 vm_pageout_scan(int pass) 629 { 630 vm_page_t m, next; 631 struct vm_page marker; 632 int save_page_shortage; 633 int save_inactive_count; 634 int page_shortage, maxscan, pcount; 635 int addl_page_shortage, addl_page_shortage_init; 636 struct proc *p, *bigproc; 637 vm_offset_t size, bigsize; 638 vm_object_t object; 639 int actcount; 640 int vnodes_skipped = 0; 641 int maxlaunder; 642 int s; 643 644 GIANT_REQUIRED; 645 /* 646 * Do whatever cleanup that the pmap code can. 647 */ 648 pmap_collect(); 649 650 addl_page_shortage_init = vm_pageout_deficit; 651 vm_pageout_deficit = 0; 652 653 /* 654 * Calculate the number of pages we want to either free or move 655 * to the cache. 656 */ 657 page_shortage = vm_paging_target() + addl_page_shortage_init; 658 save_page_shortage = page_shortage; 659 save_inactive_count = cnt.v_inactive_count; 660 661 /* 662 * Initialize our marker 663 */ 664 bzero(&marker, sizeof(marker)); 665 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER; 666 marker.queue = PQ_INACTIVE; 667 marker.wire_count = 1; 668 669 /* 670 * Start scanning the inactive queue for pages we can move to the 671 * cache or free. The scan will stop when the target is reached or 672 * we have scanned the entire inactive queue. Note that m->act_count 673 * is not used to form decisions for the inactive queue, only for the 674 * active queue. 675 * 676 * maxlaunder limits the number of dirty pages we flush per scan. 677 * For most systems a smaller value (16 or 32) is more robust under 678 * extreme memory and disk pressure because any unnecessary writes 679 * to disk can result in extreme performance degredation. However, 680 * systems with excessive dirty pages (especially when MAP_NOSYNC is 681 * used) will die horribly with limited laundering. If the pageout 682 * daemon cannot clean enough pages in the first pass, we let it go 683 * all out in succeeding passes. 684 */ 685 686 if ((maxlaunder = vm_max_launder) <= 1) 687 maxlaunder = 1; 688 if (pass) 689 maxlaunder = 10000; 690 691 rescan0: 692 addl_page_shortage = addl_page_shortage_init; 693 maxscan = cnt.v_inactive_count; 694 695 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl); 696 m != NULL && maxscan-- > 0 && page_shortage > 0; 697 m = next) { 698 699 cnt.v_pdpages++; 700 701 if (m->queue != PQ_INACTIVE) { 702 goto rescan0; 703 } 704 705 next = TAILQ_NEXT(m, pageq); 706 707 /* 708 * skip marker pages 709 */ 710 if (m->flags & PG_MARKER) 711 continue; 712 713 if (m->hold_count) { 714 vm_pageq_requeue(m); 715 addl_page_shortage++; 716 continue; 717 } 718 /* 719 * Dont mess with busy pages, keep in the front of the 720 * queue, most likely are being paged out. 721 */ 722 if (m->busy || (m->flags & PG_BUSY)) { 723 addl_page_shortage++; 724 continue; 725 } 726 727 /* 728 * If the object is not being used, we ignore previous 729 * references. 730 */ 731 if (m->object->ref_count == 0) { 732 vm_page_flag_clear(m, PG_REFERENCED); 733 pmap_clear_reference(m); 734 735 /* 736 * Otherwise, if the page has been referenced while in the 737 * inactive queue, we bump the "activation count" upwards, 738 * making it less likely that the page will be added back to 739 * the inactive queue prematurely again. Here we check the 740 * page tables (or emulated bits, if any), given the upper 741 * level VM system not knowing anything about existing 742 * references. 743 */ 744 } else if (((m->flags & PG_REFERENCED) == 0) && 745 (actcount = pmap_ts_referenced(m))) { 746 vm_page_activate(m); 747 m->act_count += (actcount + ACT_ADVANCE); 748 continue; 749 } 750 751 /* 752 * If the upper level VM system knows about any page 753 * references, we activate the page. We also set the 754 * "activation count" higher than normal so that we will less 755 * likely place pages back onto the inactive queue again. 756 */ 757 if ((m->flags & PG_REFERENCED) != 0) { 758 vm_page_flag_clear(m, PG_REFERENCED); 759 actcount = pmap_ts_referenced(m); 760 vm_page_activate(m); 761 m->act_count += (actcount + ACT_ADVANCE + 1); 762 continue; 763 } 764 765 /* 766 * If the upper level VM system doesn't know anything about 767 * the page being dirty, we have to check for it again. As 768 * far as the VM code knows, any partially dirty pages are 769 * fully dirty. 770 */ 771 if (m->dirty == 0) { 772 vm_page_test_dirty(m); 773 } else { 774 vm_page_dirty(m); 775 } 776 777 /* 778 * Invalid pages can be easily freed 779 */ 780 if (m->valid == 0) { 781 vm_pageout_page_free(m); 782 cnt.v_dfree++; 783 --page_shortage; 784 785 /* 786 * Clean pages can be placed onto the cache queue. This 787 * effectively frees them. 788 */ 789 } else if (m->dirty == 0) { 790 vm_page_cache(m); 791 --page_shortage; 792 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) { 793 /* 794 * Dirty pages need to be paged out, but flushing 795 * a page is extremely expensive verses freeing 796 * a clean page. Rather then artificially limiting 797 * the number of pages we can flush, we instead give 798 * dirty pages extra priority on the inactive queue 799 * by forcing them to be cycled through the queue 800 * twice before being flushed, after which the 801 * (now clean) page will cycle through once more 802 * before being freed. This significantly extends 803 * the thrash point for a heavily loaded machine. 804 */ 805 vm_page_flag_set(m, PG_WINATCFLS); 806 vm_pageq_requeue(m); 807 } else if (maxlaunder > 0) { 808 /* 809 * We always want to try to flush some dirty pages if 810 * we encounter them, to keep the system stable. 811 * Normally this number is small, but under extreme 812 * pressure where there are insufficient clean pages 813 * on the inactive queue, we may have to go all out. 814 */ 815 int swap_pageouts_ok; 816 struct vnode *vp = NULL; 817 struct mount *mp; 818 819 object = m->object; 820 821 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 822 swap_pageouts_ok = 1; 823 } else { 824 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 825 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 826 vm_page_count_min()); 827 828 } 829 830 /* 831 * We don't bother paging objects that are "dead". 832 * Those objects are in a "rundown" state. 833 */ 834 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 835 vm_pageq_requeue(m); 836 continue; 837 } 838 839 /* 840 * The object is already known NOT to be dead. It 841 * is possible for the vget() to block the whole 842 * pageout daemon, but the new low-memory handling 843 * code should prevent it. 844 * 845 * The previous code skipped locked vnodes and, worse, 846 * reordered pages in the queue. This results in 847 * completely non-deterministic operation and, on a 848 * busy system, can lead to extremely non-optimal 849 * pageouts. For example, it can cause clean pages 850 * to be freed and dirty pages to be moved to the end 851 * of the queue. Since dirty pages are also moved to 852 * the end of the queue once-cleaned, this gives 853 * way too large a weighting to defering the freeing 854 * of dirty pages. 855 * 856 * XXX we need to be able to apply a timeout to the 857 * vget() lock attempt. 858 */ 859 860 if (object->type == OBJT_VNODE) { 861 vp = object->handle; 862 863 mp = NULL; 864 if (vp->v_type == VREG) 865 vn_start_write(vp, &mp, V_NOWAIT); 866 if (vget(vp, LK_EXCLUSIVE|LK_NOOBJ, curthread)) { 867 vn_finished_write(mp); 868 if (object->flags & OBJ_MIGHTBEDIRTY) 869 vnodes_skipped++; 870 continue; 871 } 872 873 /* 874 * The page might have been moved to another 875 * queue during potential blocking in vget() 876 * above. The page might have been freed and 877 * reused for another vnode. The object might 878 * have been reused for another vnode. 879 */ 880 if (m->queue != PQ_INACTIVE || 881 m->object != object || 882 object->handle != vp) { 883 if (object->flags & OBJ_MIGHTBEDIRTY) 884 vnodes_skipped++; 885 vput(vp); 886 vn_finished_write(mp); 887 continue; 888 } 889 890 /* 891 * The page may have been busied during the 892 * blocking in vput(); We don't move the 893 * page back onto the end of the queue so that 894 * statistics are more correct if we don't. 895 */ 896 if (m->busy || (m->flags & PG_BUSY)) { 897 vput(vp); 898 vn_finished_write(mp); 899 continue; 900 } 901 902 /* 903 * If the page has become held, then skip it 904 */ 905 if (m->hold_count) { 906 vm_pageq_requeue(m); 907 if (object->flags & OBJ_MIGHTBEDIRTY) 908 vnodes_skipped++; 909 vput(vp); 910 vn_finished_write(mp); 911 continue; 912 } 913 } 914 915 /* 916 * If a page is dirty, then it is either being washed 917 * (but not yet cleaned) or it is still in the 918 * laundry. If it is still in the laundry, then we 919 * start the cleaning operation. 920 * 921 * This operation may cluster, invalidating the 'next' 922 * pointer. To prevent an inordinate number of 923 * restarts we use our marker to remember our place. 924 * 925 * decrement page_shortage on success to account for 926 * the (future) cleaned page. Otherwise we could wind 927 * up laundering or cleaning too many pages. 928 */ 929 s = splvm(); 930 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, m, &marker, pageq); 931 splx(s); 932 if (vm_pageout_clean(m) != 0) { 933 --page_shortage; 934 --maxlaunder; 935 } 936 s = splvm(); 937 next = TAILQ_NEXT(&marker, pageq); 938 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, &marker, pageq); 939 splx(s); 940 if (vp) { 941 vput(vp); 942 vn_finished_write(mp); 943 } 944 } 945 } 946 947 /* 948 * Compute the number of pages we want to try to move from the 949 * active queue to the inactive queue. 950 */ 951 page_shortage = vm_paging_target() + 952 cnt.v_inactive_target - cnt.v_inactive_count; 953 page_shortage += addl_page_shortage; 954 955 /* 956 * Scan the active queue for things we can deactivate. We nominally 957 * track the per-page activity counter and use it to locate 958 * deactivation candidates. 959 */ 960 961 pcount = cnt.v_active_count; 962 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 963 964 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) { 965 966 /* 967 * This is a consistency check, and should likely be a panic 968 * or warning. 969 */ 970 if (m->queue != PQ_ACTIVE) { 971 break; 972 } 973 974 next = TAILQ_NEXT(m, pageq); 975 /* 976 * Don't deactivate pages that are busy. 977 */ 978 if ((m->busy != 0) || 979 (m->flags & PG_BUSY) || 980 (m->hold_count != 0)) { 981 vm_pageq_requeue(m); 982 m = next; 983 continue; 984 } 985 986 /* 987 * The count for pagedaemon pages is done after checking the 988 * page for eligibility... 989 */ 990 cnt.v_pdpages++; 991 992 /* 993 * Check to see "how much" the page has been used. 994 */ 995 actcount = 0; 996 if (m->object->ref_count != 0) { 997 if (m->flags & PG_REFERENCED) { 998 actcount += 1; 999 } 1000 actcount += pmap_ts_referenced(m); 1001 if (actcount) { 1002 m->act_count += ACT_ADVANCE + actcount; 1003 if (m->act_count > ACT_MAX) 1004 m->act_count = ACT_MAX; 1005 } 1006 } 1007 1008 /* 1009 * Since we have "tested" this bit, we need to clear it now. 1010 */ 1011 vm_page_flag_clear(m, PG_REFERENCED); 1012 1013 /* 1014 * Only if an object is currently being used, do we use the 1015 * page activation count stats. 1016 */ 1017 if (actcount && (m->object->ref_count != 0)) { 1018 vm_pageq_requeue(m); 1019 } else { 1020 m->act_count -= min(m->act_count, ACT_DECLINE); 1021 if (vm_pageout_algorithm || 1022 m->object->ref_count == 0 || 1023 m->act_count == 0) { 1024 page_shortage--; 1025 if (m->object->ref_count == 0) { 1026 vm_page_protect(m, VM_PROT_NONE); 1027 if (m->dirty == 0) 1028 vm_page_cache(m); 1029 else 1030 vm_page_deactivate(m); 1031 } else { 1032 vm_page_deactivate(m); 1033 } 1034 } else { 1035 vm_pageq_requeue(m); 1036 } 1037 } 1038 m = next; 1039 } 1040 1041 s = splvm(); 1042 1043 /* 1044 * We try to maintain some *really* free pages, this allows interrupt 1045 * code to be guaranteed space. Since both cache and free queues 1046 * are considered basically 'free', moving pages from cache to free 1047 * does not effect other calculations. 1048 */ 1049 1050 while (cnt.v_free_count < cnt.v_free_reserved) { 1051 static int cache_rover = 0; 1052 m = vm_pageq_find(PQ_CACHE, cache_rover, FALSE); 1053 if (!m) 1054 break; 1055 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || 1056 m->busy || 1057 m->hold_count || 1058 m->wire_count) { 1059 #ifdef INVARIANTS 1060 printf("Warning: busy page %p found in cache\n", m); 1061 #endif 1062 vm_page_deactivate(m); 1063 continue; 1064 } 1065 cache_rover = (cache_rover + PQ_PRIME2) & PQ_L2_MASK; 1066 vm_pageout_page_free(m); 1067 cnt.v_dfree++; 1068 } 1069 splx(s); 1070 1071 #if !defined(NO_SWAPPING) 1072 /* 1073 * Idle process swapout -- run once per second. 1074 */ 1075 if (vm_swap_idle_enabled) { 1076 static long lsec; 1077 if (time_second != lsec) { 1078 vm_pageout_req_swapout |= VM_SWAP_IDLE; 1079 vm_req_vmdaemon(); 1080 lsec = time_second; 1081 } 1082 } 1083 #endif 1084 1085 /* 1086 * If we didn't get enough free pages, and we have skipped a vnode 1087 * in a writeable object, wakeup the sync daemon. And kick swapout 1088 * if we did not get enough free pages. 1089 */ 1090 if (vm_paging_target() > 0) { 1091 if (vnodes_skipped && vm_page_count_min()) 1092 (void) speedup_syncer(); 1093 #if !defined(NO_SWAPPING) 1094 if (vm_swap_enabled && vm_page_count_target()) { 1095 vm_req_vmdaemon(); 1096 vm_pageout_req_swapout |= VM_SWAP_NORMAL; 1097 } 1098 #endif 1099 } 1100 1101 /* 1102 * If we are out of swap and were not able to reach our paging 1103 * target, kill the largest process. 1104 * 1105 * We keep the process bigproc locked once we find it to keep anyone 1106 * from messing with it; however, there is a possibility of 1107 * deadlock if process B is bigproc and one of it's child processes 1108 * attempts to propagate a signal to B while we are waiting for A's 1109 * lock while walking this list. To avoid this, we don't block on 1110 * the process lock but just skip a process if it is already locked. 1111 */ 1112 if ((vm_swap_size < 64 && vm_page_count_min()) || 1113 (swap_pager_full && vm_paging_target() > 0)) { 1114 #if 0 1115 if ((vm_swap_size < 64 || swap_pager_full) && vm_page_count_min()) { 1116 #endif 1117 bigproc = NULL; 1118 bigsize = 0; 1119 sx_slock(&allproc_lock); 1120 LIST_FOREACH(p, &allproc, p_list) { 1121 /* 1122 * If this process is already locked, skip it. 1123 */ 1124 if (PROC_TRYLOCK(p) == 0) 1125 continue; 1126 /* 1127 * if this is a system process, skip it 1128 */ 1129 if ((p->p_flag & P_SYSTEM) || (p->p_lock > 0) || 1130 (p->p_pid == 1) || 1131 ((p->p_pid < 48) && (vm_swap_size != 0))) { 1132 PROC_UNLOCK(p); 1133 continue; 1134 } 1135 /* 1136 * if the process is in a non-running type state, 1137 * don't touch it. 1138 */ 1139 mtx_lock_spin(&sched_lock); 1140 if (p->p_stat != SRUN && p->p_stat != SSLEEP) { 1141 mtx_unlock_spin(&sched_lock); 1142 PROC_UNLOCK(p); 1143 continue; 1144 } 1145 mtx_unlock_spin(&sched_lock); 1146 /* 1147 * get the process size 1148 */ 1149 size = vmspace_resident_count(p->p_vmspace) + 1150 vmspace_swap_count(p->p_vmspace); 1151 /* 1152 * if the this process is bigger than the biggest one 1153 * remember it. 1154 */ 1155 if (size > bigsize) { 1156 if (bigproc != NULL) 1157 PROC_UNLOCK(bigproc); 1158 bigproc = p; 1159 bigsize = size; 1160 } else 1161 PROC_UNLOCK(p); 1162 } 1163 sx_sunlock(&allproc_lock); 1164 if (bigproc != NULL) { 1165 struct ksegrp *kg; 1166 killproc(bigproc, "out of swap space"); 1167 mtx_lock_spin(&sched_lock); 1168 FOREACH_KSEGRP_IN_PROC(bigproc, kg) { 1169 kg->kg_estcpu = 0; 1170 kg->kg_nice = PRIO_MIN; /* XXXKSE ??? */ 1171 resetpriority(kg); 1172 } 1173 mtx_unlock_spin(&sched_lock); 1174 PROC_UNLOCK(bigproc); 1175 wakeup(&cnt.v_free_count); 1176 } 1177 } 1178 } 1179 1180 /* 1181 * This routine tries to maintain the pseudo LRU active queue, 1182 * so that during long periods of time where there is no paging, 1183 * that some statistic accumulation still occurs. This code 1184 * helps the situation where paging just starts to occur. 1185 */ 1186 static void 1187 vm_pageout_page_stats() 1188 { 1189 vm_page_t m,next; 1190 int pcount,tpcount; /* Number of pages to check */ 1191 static int fullintervalcount = 0; 1192 int page_shortage; 1193 int s0; 1194 1195 page_shortage = 1196 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) - 1197 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count); 1198 1199 if (page_shortage <= 0) 1200 return; 1201 1202 s0 = splvm(); 1203 1204 pcount = cnt.v_active_count; 1205 fullintervalcount += vm_pageout_stats_interval; 1206 if (fullintervalcount < vm_pageout_full_stats_interval) { 1207 tpcount = (vm_pageout_stats_max * cnt.v_active_count) / cnt.v_page_count; 1208 if (pcount > tpcount) 1209 pcount = tpcount; 1210 } else { 1211 fullintervalcount = 0; 1212 } 1213 1214 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1215 while ((m != NULL) && (pcount-- > 0)) { 1216 int actcount; 1217 1218 if (m->queue != PQ_ACTIVE) { 1219 break; 1220 } 1221 1222 next = TAILQ_NEXT(m, pageq); 1223 /* 1224 * Don't deactivate pages that are busy. 1225 */ 1226 if ((m->busy != 0) || 1227 (m->flags & PG_BUSY) || 1228 (m->hold_count != 0)) { 1229 vm_pageq_requeue(m); 1230 m = next; 1231 continue; 1232 } 1233 1234 actcount = 0; 1235 if (m->flags & PG_REFERENCED) { 1236 vm_page_flag_clear(m, PG_REFERENCED); 1237 actcount += 1; 1238 } 1239 1240 actcount += pmap_ts_referenced(m); 1241 if (actcount) { 1242 m->act_count += ACT_ADVANCE + actcount; 1243 if (m->act_count > ACT_MAX) 1244 m->act_count = ACT_MAX; 1245 vm_pageq_requeue(m); 1246 } else { 1247 if (m->act_count == 0) { 1248 /* 1249 * We turn off page access, so that we have 1250 * more accurate RSS stats. We don't do this 1251 * in the normal page deactivation when the 1252 * system is loaded VM wise, because the 1253 * cost of the large number of page protect 1254 * operations would be higher than the value 1255 * of doing the operation. 1256 */ 1257 vm_page_protect(m, VM_PROT_NONE); 1258 vm_page_deactivate(m); 1259 } else { 1260 m->act_count -= min(m->act_count, ACT_DECLINE); 1261 vm_pageq_requeue(m); 1262 } 1263 } 1264 1265 m = next; 1266 } 1267 splx(s0); 1268 } 1269 1270 static int 1271 vm_pageout_free_page_calc(count) 1272 vm_size_t count; 1273 { 1274 if (count < cnt.v_page_count) 1275 return 0; 1276 /* 1277 * free_reserved needs to include enough for the largest swap pager 1278 * structures plus enough for any pv_entry structs when paging. 1279 */ 1280 if (cnt.v_page_count > 1024) 1281 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200; 1282 else 1283 cnt.v_free_min = 4; 1284 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1285 cnt.v_interrupt_free_min; 1286 cnt.v_free_reserved = vm_pageout_page_count + 1287 cnt.v_pageout_free_min + (count / 768) + PQ_L2_SIZE; 1288 cnt.v_free_severe = cnt.v_free_min / 2; 1289 cnt.v_free_min += cnt.v_free_reserved; 1290 cnt.v_free_severe += cnt.v_free_reserved; 1291 return 1; 1292 } 1293 1294 1295 /* 1296 * vm_pageout is the high level pageout daemon. 1297 */ 1298 static void 1299 vm_pageout() 1300 { 1301 int pass; 1302 1303 mtx_lock(&Giant); 1304 1305 /* 1306 * Initialize some paging parameters. 1307 */ 1308 1309 cnt.v_interrupt_free_min = 2; 1310 if (cnt.v_page_count < 2000) 1311 vm_pageout_page_count = 8; 1312 1313 vm_pageout_free_page_calc(cnt.v_page_count); 1314 /* 1315 * v_free_target and v_cache_min control pageout hysteresis. Note 1316 * that these are more a measure of the VM cache queue hysteresis 1317 * then the VM free queue. Specifically, v_free_target is the 1318 * high water mark (free+cache pages). 1319 * 1320 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the 1321 * low water mark, while v_free_min is the stop. v_cache_min must 1322 * be big enough to handle memory needs while the pageout daemon 1323 * is signalled and run to free more pages. 1324 */ 1325 if (cnt.v_free_count > 6144) 1326 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved; 1327 else 1328 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved; 1329 1330 if (cnt.v_free_count > 2048) { 1331 cnt.v_cache_min = cnt.v_free_target; 1332 cnt.v_cache_max = 2 * cnt.v_cache_min; 1333 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2; 1334 } else { 1335 cnt.v_cache_min = 0; 1336 cnt.v_cache_max = 0; 1337 cnt.v_inactive_target = cnt.v_free_count / 4; 1338 } 1339 if (cnt.v_inactive_target > cnt.v_free_count / 3) 1340 cnt.v_inactive_target = cnt.v_free_count / 3; 1341 1342 /* XXX does not really belong here */ 1343 if (vm_page_max_wired == 0) 1344 vm_page_max_wired = cnt.v_free_count / 3; 1345 1346 if (vm_pageout_stats_max == 0) 1347 vm_pageout_stats_max = cnt.v_free_target; 1348 1349 /* 1350 * Set interval in seconds for stats scan. 1351 */ 1352 if (vm_pageout_stats_interval == 0) 1353 vm_pageout_stats_interval = 5; 1354 if (vm_pageout_full_stats_interval == 0) 1355 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4; 1356 1357 1358 /* 1359 * Set maximum free per pass 1360 */ 1361 if (vm_pageout_stats_free_max == 0) 1362 vm_pageout_stats_free_max = 5; 1363 1364 PROC_LOCK(curthread->td_proc); 1365 curthread->td_proc->p_flag |= P_BUFEXHAUST; 1366 PROC_UNLOCK(curthread->td_proc); 1367 swap_pager_swap_init(); 1368 pass = 0; 1369 /* 1370 * The pageout daemon is never done, so loop forever. 1371 */ 1372 while (TRUE) { 1373 int error; 1374 int s = splvm(); 1375 1376 /* 1377 * If we have enough free memory, wakeup waiters. Do 1378 * not clear vm_pages_needed until we reach our target, 1379 * otherwise we may be woken up over and over again and 1380 * waste a lot of cpu. 1381 */ 1382 if (vm_pages_needed && !vm_page_count_min()) { 1383 if (vm_paging_needed() <= 0) 1384 vm_pages_needed = 0; 1385 wakeup(&cnt.v_free_count); 1386 } 1387 if (vm_pages_needed) { 1388 /* 1389 * Still not done, take a second pass without waiting 1390 * (unlimited dirty cleaning), otherwise sleep a bit 1391 * and try again. 1392 */ 1393 ++pass; 1394 if (pass > 1) 1395 tsleep(&vm_pages_needed, PVM, 1396 "psleep", hz/2); 1397 } else { 1398 /* 1399 * Good enough, sleep & handle stats. Prime the pass 1400 * for the next run. 1401 */ 1402 if (pass > 1) 1403 pass = 1; 1404 else 1405 pass = 0; 1406 error = tsleep(&vm_pages_needed, PVM, 1407 "psleep", vm_pageout_stats_interval * hz); 1408 if (error && !vm_pages_needed) { 1409 splx(s); 1410 pass = 0; 1411 vm_pageout_page_stats(); 1412 continue; 1413 } 1414 } 1415 1416 if (vm_pages_needed) 1417 cnt.v_pdwakeups++; 1418 splx(s); 1419 vm_pageout_scan(pass); 1420 vm_pageout_deficit = 0; 1421 } 1422 } 1423 1424 void 1425 pagedaemon_wakeup() 1426 { 1427 if (!vm_pages_needed && curthread->td_proc != pageproc) { 1428 vm_pages_needed++; 1429 wakeup(&vm_pages_needed); 1430 } 1431 } 1432 1433 #if !defined(NO_SWAPPING) 1434 static void 1435 vm_req_vmdaemon() 1436 { 1437 static int lastrun = 0; 1438 1439 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1440 wakeup(&vm_daemon_needed); 1441 lastrun = ticks; 1442 } 1443 } 1444 1445 static void 1446 vm_daemon() 1447 { 1448 struct proc *p; 1449 1450 mtx_lock(&Giant); 1451 while (TRUE) { 1452 tsleep(&vm_daemon_needed, PPAUSE, "psleep", 0); 1453 if (vm_pageout_req_swapout) { 1454 swapout_procs(vm_pageout_req_swapout); 1455 vm_pageout_req_swapout = 0; 1456 } 1457 /* 1458 * scan the processes for exceeding their rlimits or if 1459 * process is swapped out -- deactivate pages 1460 */ 1461 1462 sx_slock(&allproc_lock); 1463 LIST_FOREACH(p, &allproc, p_list) { 1464 vm_pindex_t limit, size; 1465 1466 /* 1467 * if this is a system process or if we have already 1468 * looked at this process, skip it. 1469 */ 1470 if (p->p_flag & (P_SYSTEM | P_WEXIT)) { 1471 continue; 1472 } 1473 /* 1474 * if the process is in a non-running type state, 1475 * don't touch it. 1476 */ 1477 mtx_lock_spin(&sched_lock); 1478 if (p->p_stat != SRUN && p->p_stat != SSLEEP) { 1479 mtx_unlock_spin(&sched_lock); 1480 continue; 1481 } 1482 /* 1483 * get a limit 1484 */ 1485 limit = OFF_TO_IDX( 1486 qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur, 1487 p->p_rlimit[RLIMIT_RSS].rlim_max)); 1488 1489 /* 1490 * let processes that are swapped out really be 1491 * swapped out set the limit to nothing (will force a 1492 * swap-out.) 1493 */ 1494 if ((p->p_sflag & PS_INMEM) == 0) 1495 limit = 0; /* XXX */ 1496 mtx_unlock_spin(&sched_lock); 1497 1498 size = vmspace_resident_count(p->p_vmspace); 1499 if (limit >= 0 && size >= limit) { 1500 vm_pageout_map_deactivate_pages( 1501 &p->p_vmspace->vm_map, limit); 1502 } 1503 } 1504 sx_sunlock(&allproc_lock); 1505 } 1506 } 1507 #endif 1508