vm_pageout.c revision 49937
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 1.144 1999/07/04 00:25:37 mckusick Exp $ 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/proc.h> 80#include <sys/kthread.h> 81#include <sys/resourcevar.h> 82#include <sys/signalvar.h> 83#include <sys/vnode.h> 84#include <sys/vmmeter.h> 85#include <sys/sysctl.h> 86 87#include <vm/vm.h> 88#include <vm/vm_param.h> 89#include <vm/vm_prot.h> 90#include <sys/lock.h> 91#include <vm/vm_object.h> 92#include <vm/vm_page.h> 93#include <vm/vm_map.h> 94#include <vm/vm_pageout.h> 95#include <vm/vm_pager.h> 96#include <vm/swap_pager.h> 97#include <vm/vm_extern.h> 98 99/* 100 * System initialization 101 */ 102 103/* the kernel process "vm_pageout"*/ 104static void vm_pageout __P((void)); 105static int vm_pageout_clean __P((vm_page_t)); 106static int vm_pageout_scan __P((void)); 107static int vm_pageout_free_page_calc __P((vm_size_t count)); 108struct proc *pageproc; 109 110static struct kproc_desc page_kp = { 111 "pagedaemon", 112 vm_pageout, 113 &pageproc 114}; 115SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp) 116 117#if !defined(NO_SWAPPING) 118/* the kernel process "vm_daemon"*/ 119static void vm_daemon __P((void)); 120static struct proc *vmproc; 121 122static struct kproc_desc vm_kp = { 123 "vmdaemon", 124 vm_daemon, 125 &vmproc 126}; 127SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp) 128#endif 129 130 131int vm_pages_needed=0; /* Event on which pageout daemon sleeps */ 132int vm_pageout_deficit=0; /* Estimated number of pages deficit */ 133int vm_pageout_pages_needed=0; /* flag saying that the pageout daemon needs pages */ 134 135extern int npendingio; 136#if !defined(NO_SWAPPING) 137static int vm_pageout_req_swapout; /* XXX */ 138static int vm_daemon_needed; 139#endif 140extern int nswiodone; 141extern int vm_swap_size; 142static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0; 143static int vm_pageout_full_stats_interval = 0; 144static int vm_pageout_stats_free_max=0, vm_pageout_algorithm_lru=0; 145static int defer_swap_pageouts=0; 146static int disable_swap_pageouts=0; 147 148static int max_page_launder=100; 149#if defined(NO_SWAPPING) 150static int vm_swap_enabled=0; 151static int vm_swap_idle_enabled=0; 152#else 153static int vm_swap_enabled=1; 154static int vm_swap_idle_enabled=0; 155#endif 156 157SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm, 158 CTLFLAG_RW, &vm_pageout_algorithm_lru, 0, "LRU page mgmt"); 159 160SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max, 161 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length"); 162 163SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval, 164 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan"); 165 166SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval, 167 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan"); 168 169SYSCTL_INT(_vm, OID_AUTO, pageout_stats_free_max, 170 CTLFLAG_RW, &vm_pageout_stats_free_max, 0, "Not implemented"); 171 172#if defined(NO_SWAPPING) 173SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 174 CTLFLAG_RD, &vm_swap_enabled, 0, ""); 175SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 176 CTLFLAG_RD, &vm_swap_idle_enabled, 0, ""); 177#else 178SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 179 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout"); 180SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 181 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 182#endif 183 184SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts, 185 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem"); 186 187SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts, 188 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages"); 189 190SYSCTL_INT(_vm, OID_AUTO, max_page_launder, 191 CTLFLAG_RW, &max_page_launder, 0, "Maximum number of pages to clean per pass"); 192 193 194#define VM_PAGEOUT_PAGE_COUNT 16 195int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT; 196 197int vm_page_max_wired; /* XXX max # of wired pages system-wide */ 198 199#if !defined(NO_SWAPPING) 200typedef void freeer_fcn_t __P((vm_map_t, vm_object_t, vm_pindex_t, int)); 201static void vm_pageout_map_deactivate_pages __P((vm_map_t, vm_pindex_t)); 202static freeer_fcn_t vm_pageout_object_deactivate_pages; 203static void vm_req_vmdaemon __P((void)); 204#endif 205static void vm_pageout_page_stats(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. Note the careful timing, however, the busy bit isn't set till 214 * late and we cannot do anything that will mess with the page. 215 */ 216 217static int 218vm_pageout_clean(m) 219 vm_page_t m; 220{ 221 register vm_object_t object; 222 vm_page_t mc[2*vm_pageout_page_count]; 223 int pageout_count; 224 int i, forward_okay, backward_okay, page_base; 225 vm_pindex_t pindex = m->pindex; 226 227 object = m->object; 228 229 /* 230 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP 231 * with the new swapper, but we could have serious problems paging 232 * out other object types if there is insufficient memory. 233 * 234 * Unfortunately, checking free memory here is far too late, so the 235 * check has been moved up a procedural level. 236 */ 237 238 /* 239 * Don't mess with the page if it's busy. 240 */ 241 if ((m->hold_count != 0) || 242 ((m->busy != 0) || (m->flags & PG_BUSY))) 243 return 0; 244 245 mc[vm_pageout_page_count] = m; 246 pageout_count = 1; 247 page_base = vm_pageout_page_count; 248 forward_okay = TRUE; 249 if (pindex != 0) 250 backward_okay = TRUE; 251 else 252 backward_okay = FALSE; 253 /* 254 * Scan object for clusterable pages. 255 * 256 * We can cluster ONLY if: ->> the page is NOT 257 * clean, wired, busy, held, or mapped into a 258 * buffer, and one of the following: 259 * 1) The page is inactive, or a seldom used 260 * active page. 261 * -or- 262 * 2) we force the issue. 263 */ 264 for (i = 1; (i < vm_pageout_page_count) && (forward_okay || backward_okay); i++) { 265 vm_page_t p; 266 267 /* 268 * See if forward page is clusterable. 269 */ 270 if (forward_okay) { 271 /* 272 * Stop forward scan at end of object. 273 */ 274 if ((pindex + i) > object->size) { 275 forward_okay = FALSE; 276 goto do_backward; 277 } 278 p = vm_page_lookup(object, pindex + i); 279 if (p) { 280 if (((p->queue - p->pc) == PQ_CACHE) || 281 (p->flags & PG_BUSY) || p->busy) { 282 forward_okay = FALSE; 283 goto do_backward; 284 } 285 vm_page_test_dirty(p); 286 if ((p->dirty & p->valid) != 0 && 287 (p->queue == PQ_INACTIVE) && 288 (p->wire_count == 0) && 289 (p->hold_count == 0)) { 290 mc[vm_pageout_page_count + i] = p; 291 pageout_count++; 292 if (pageout_count == vm_pageout_page_count) 293 break; 294 } else { 295 forward_okay = FALSE; 296 } 297 } else { 298 forward_okay = FALSE; 299 } 300 } 301do_backward: 302 /* 303 * See if backward page is clusterable. 304 */ 305 if (backward_okay) { 306 /* 307 * Stop backward scan at beginning of object. 308 */ 309 if ((pindex - i) == 0) { 310 backward_okay = FALSE; 311 } 312 p = vm_page_lookup(object, pindex - i); 313 if (p) { 314 if (((p->queue - p->pc) == PQ_CACHE) || 315 (p->flags & PG_BUSY) || p->busy) { 316 backward_okay = FALSE; 317 continue; 318 } 319 vm_page_test_dirty(p); 320 if ((p->dirty & p->valid) != 0 && 321 (p->queue == PQ_INACTIVE) && 322 (p->wire_count == 0) && 323 (p->hold_count == 0)) { 324 mc[vm_pageout_page_count - i] = p; 325 pageout_count++; 326 page_base--; 327 if (pageout_count == vm_pageout_page_count) 328 break; 329 } else { 330 backward_okay = FALSE; 331 } 332 } else { 333 backward_okay = FALSE; 334 } 335 } 336 } 337 338 /* 339 * we allow reads during pageouts... 340 */ 341 return vm_pageout_flush(&mc[page_base], pageout_count, 0); 342} 343 344/* 345 * vm_pageout_flush() - launder the given pages 346 * 347 * The given pages are laundered. Note that we setup for the start of 348 * I/O ( i.e. busy the page ), mark it read-only, and bump the object 349 * reference count all in here rather then in the parent. If we want 350 * the parent to do more sophisticated things we may have to change 351 * the ordering. 352 */ 353 354int 355vm_pageout_flush(mc, count, flags) 356 vm_page_t *mc; 357 int count; 358 int flags; 359{ 360 register vm_object_t object; 361 int pageout_status[count]; 362 int numpagedout = 0; 363 int i; 364 365 /* 366 * Initiate I/O. Bump the vm_page_t->busy counter and 367 * mark the pages read-only. 368 * 369 * We do not have to fixup the clean/dirty bits here... we can 370 * allow the pager to do it after the I/O completes. 371 */ 372 373 for (i = 0; i < count; i++) { 374 vm_page_io_start(mc[i]); 375 vm_page_protect(mc[i], VM_PROT_READ); 376 } 377 378 object = mc[0]->object; 379 vm_object_pip_add(object, count); 380 381 vm_pager_put_pages(object, mc, count, 382 (flags | ((object == kernel_object) ? OBJPC_SYNC : 0)), 383 pageout_status); 384 385 for (i = 0; i < count; i++) { 386 vm_page_t mt = mc[i]; 387 388 switch (pageout_status[i]) { 389 case VM_PAGER_OK: 390 numpagedout++; 391 break; 392 case VM_PAGER_PEND: 393 numpagedout++; 394 break; 395 case VM_PAGER_BAD: 396 /* 397 * Page outside of range of object. Right now we 398 * essentially lose the changes by pretending it 399 * worked. 400 */ 401 pmap_clear_modify(VM_PAGE_TO_PHYS(mt)); 402 mt->dirty = 0; 403 break; 404 case VM_PAGER_ERROR: 405 case VM_PAGER_FAIL: 406 /* 407 * If page couldn't be paged out, then reactivate the 408 * page so it doesn't clog the inactive list. (We 409 * will try paging out it again later). 410 */ 411 vm_page_activate(mt); 412 break; 413 case VM_PAGER_AGAIN: 414 break; 415 } 416 417 /* 418 * If the operation is still going, leave the page busy to 419 * block all other accesses. Also, leave the paging in 420 * progress indicator set so that we don't attempt an object 421 * collapse. 422 */ 423 if (pageout_status[i] != VM_PAGER_PEND) { 424 vm_object_pip_wakeup(object); 425 vm_page_io_finish(mt); 426 } 427 } 428 return numpagedout; 429} 430 431#if !defined(NO_SWAPPING) 432/* 433 * vm_pageout_object_deactivate_pages 434 * 435 * deactivate enough pages to satisfy the inactive target 436 * requirements or if vm_page_proc_limit is set, then 437 * deactivate all of the pages in the object and its 438 * backing_objects. 439 * 440 * The object and map must be locked. 441 */ 442static void 443vm_pageout_object_deactivate_pages(map, object, desired, map_remove_only) 444 vm_map_t map; 445 vm_object_t object; 446 vm_pindex_t desired; 447 int map_remove_only; 448{ 449 register vm_page_t p, next; 450 int rcount; 451 int remove_mode; 452 int s; 453 454 if (object->type == OBJT_DEVICE) 455 return; 456 457 while (object) { 458 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 459 return; 460 if (object->paging_in_progress) 461 return; 462 463 remove_mode = map_remove_only; 464 if (object->shadow_count > 1) 465 remove_mode = 1; 466 /* 467 * scan the objects entire memory queue 468 */ 469 rcount = object->resident_page_count; 470 p = TAILQ_FIRST(&object->memq); 471 while (p && (rcount-- > 0)) { 472 int actcount; 473 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 474 return; 475 next = TAILQ_NEXT(p, listq); 476 cnt.v_pdpages++; 477 if (p->wire_count != 0 || 478 p->hold_count != 0 || 479 p->busy != 0 || 480 (p->flags & PG_BUSY) || 481 !pmap_page_exists(vm_map_pmap(map), VM_PAGE_TO_PHYS(p))) { 482 p = next; 483 continue; 484 } 485 486 actcount = pmap_ts_referenced(VM_PAGE_TO_PHYS(p)); 487 if (actcount) { 488 vm_page_flag_set(p, PG_REFERENCED); 489 } else if (p->flags & PG_REFERENCED) { 490 actcount = 1; 491 } 492 493 if ((p->queue != PQ_ACTIVE) && 494 (p->flags & PG_REFERENCED)) { 495 vm_page_activate(p); 496 p->act_count += actcount; 497 vm_page_flag_clear(p, PG_REFERENCED); 498 } else if (p->queue == PQ_ACTIVE) { 499 if ((p->flags & PG_REFERENCED) == 0) { 500 p->act_count -= min(p->act_count, ACT_DECLINE); 501 if (!remove_mode && (vm_pageout_algorithm_lru || (p->act_count == 0))) { 502 vm_page_protect(p, VM_PROT_NONE); 503 vm_page_deactivate(p); 504 } else { 505 s = splvm(); 506 TAILQ_REMOVE(&vm_page_queue_active, p, pageq); 507 TAILQ_INSERT_TAIL(&vm_page_queue_active, p, pageq); 508 splx(s); 509 } 510 } else { 511 vm_page_activate(p); 512 vm_page_flag_clear(p, PG_REFERENCED); 513 if (p->act_count < (ACT_MAX - ACT_ADVANCE)) 514 p->act_count += ACT_ADVANCE; 515 s = splvm(); 516 TAILQ_REMOVE(&vm_page_queue_active, p, pageq); 517 TAILQ_INSERT_TAIL(&vm_page_queue_active, p, pageq); 518 splx(s); 519 } 520 } else if (p->queue == PQ_INACTIVE) { 521 vm_page_protect(p, VM_PROT_NONE); 522 } 523 p = next; 524 } 525 object = object->backing_object; 526 } 527 return; 528} 529 530/* 531 * deactivate some number of pages in a map, try to do it fairly, but 532 * that is really hard to do. 533 */ 534static void 535vm_pageout_map_deactivate_pages(map, desired) 536 vm_map_t map; 537 vm_pindex_t desired; 538{ 539 vm_map_entry_t tmpe; 540 vm_object_t obj, bigobj; 541 542 if (lockmgr(&map->lock, LK_EXCLUSIVE | LK_NOWAIT, (void *)0, curproc)) { 543 return; 544 } 545 546 bigobj = NULL; 547 548 /* 549 * first, search out the biggest object, and try to free pages from 550 * that. 551 */ 552 tmpe = map->header.next; 553 while (tmpe != &map->header) { 554 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 555 obj = tmpe->object.vm_object; 556 if ((obj != NULL) && (obj->shadow_count <= 1) && 557 ((bigobj == NULL) || 558 (bigobj->resident_page_count < obj->resident_page_count))) { 559 bigobj = obj; 560 } 561 } 562 tmpe = tmpe->next; 563 } 564 565 if (bigobj) 566 vm_pageout_object_deactivate_pages(map, bigobj, desired, 0); 567 568 /* 569 * Next, hunt around for other pages to deactivate. We actually 570 * do this search sort of wrong -- .text first is not the best idea. 571 */ 572 tmpe = map->header.next; 573 while (tmpe != &map->header) { 574 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 575 break; 576 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 577 obj = tmpe->object.vm_object; 578 if (obj) 579 vm_pageout_object_deactivate_pages(map, obj, desired, 0); 580 } 581 tmpe = tmpe->next; 582 }; 583 584 /* 585 * Remove all mappings if a process is swapped out, this will free page 586 * table pages. 587 */ 588 if (desired == 0) 589 pmap_remove(vm_map_pmap(map), 590 VM_MIN_ADDRESS, VM_MAXUSER_ADDRESS); 591 vm_map_unlock(map); 592 return; 593} 594#endif 595 596/* 597 * Don't try to be fancy - being fancy can lead to VOP_LOCK's and therefore 598 * to vnode deadlocks. We only do it for OBJT_DEFAULT and OBJT_SWAP objects 599 * which we know can be trivially freed. 600 */ 601 602void 603vm_pageout_page_free(vm_page_t m) { 604 vm_object_t object = m->object; 605 int type = object->type; 606 607 if (type == OBJT_SWAP || type == OBJT_DEFAULT) 608 vm_object_reference(object); 609 vm_page_busy(m); 610 vm_page_protect(m, VM_PROT_NONE); 611 vm_page_free(m); 612 if (type == OBJT_SWAP || type == OBJT_DEFAULT) 613 vm_object_deallocate(object); 614} 615 616/* 617 * vm_pageout_scan does the dirty work for the pageout daemon. 618 */ 619static int 620vm_pageout_scan() 621{ 622 vm_page_t m, next; 623 int page_shortage, maxscan, pcount; 624 int addl_page_shortage, addl_page_shortage_init; 625 int maxlaunder; 626 int launder_loop = 0; 627 struct proc *p, *bigproc; 628 vm_offset_t size, bigsize; 629 vm_object_t object; 630 int force_wakeup = 0; 631 int actcount; 632 int vnodes_skipped = 0; 633 int s; 634 635 /* 636 * Do whatever cleanup that the pmap code can. 637 */ 638 pmap_collect(); 639 640 addl_page_shortage_init = vm_pageout_deficit; 641 vm_pageout_deficit = 0; 642 643 if (max_page_launder == 0) 644 max_page_launder = 1; 645 646 /* 647 * Calculate the number of pages we want to either free or move 648 * to the cache. 649 */ 650 651 page_shortage = (cnt.v_free_target + cnt.v_cache_min) - 652 (cnt.v_free_count + cnt.v_cache_count); 653 page_shortage += addl_page_shortage_init; 654 655 /* 656 * Figure out what to do with dirty pages when they are encountered. 657 * Assume that 1/3 of the pages on the inactive list are clean. If 658 * we think we can reach our target, disable laundering (do not 659 * clean any dirty pages). If we miss the target we will loop back 660 * up and do a laundering run. 661 */ 662 663 if (cnt.v_inactive_count / 3 > page_shortage) { 664 maxlaunder = 0; 665 launder_loop = 0; 666 } else { 667 maxlaunder = 668 (cnt.v_inactive_target > max_page_launder) ? 669 max_page_launder : cnt.v_inactive_target; 670 launder_loop = 1; 671 } 672 673 /* 674 * Start scanning the inactive queue for pages we can move to the 675 * cache or free. The scan will stop when the target is reached or 676 * we have scanned the entire inactive queue. 677 */ 678 679rescan0: 680 addl_page_shortage = addl_page_shortage_init; 681 maxscan = cnt.v_inactive_count; 682 for (m = TAILQ_FIRST(&vm_page_queue_inactive); 683 m != NULL && maxscan-- > 0 && page_shortage > 0; 684 m = next) { 685 686 cnt.v_pdpages++; 687 688 if (m->queue != PQ_INACTIVE) { 689 goto rescan0; 690 } 691 692 next = TAILQ_NEXT(m, pageq); 693 694 if (m->hold_count) { 695 s = splvm(); 696 TAILQ_REMOVE(&vm_page_queue_inactive, m, pageq); 697 TAILQ_INSERT_TAIL(&vm_page_queue_inactive, m, pageq); 698 splx(s); 699 addl_page_shortage++; 700 continue; 701 } 702 /* 703 * Dont mess with busy pages, keep in the front of the 704 * queue, most likely are being paged out. 705 */ 706 if (m->busy || (m->flags & PG_BUSY)) { 707 addl_page_shortage++; 708 continue; 709 } 710 711 /* 712 * If the object is not being used, we ignore previous 713 * references. 714 */ 715 if (m->object->ref_count == 0) { 716 vm_page_flag_clear(m, PG_REFERENCED); 717 pmap_clear_reference(VM_PAGE_TO_PHYS(m)); 718 719 /* 720 * Otherwise, if the page has been referenced while in the 721 * inactive queue, we bump the "activation count" upwards, 722 * making it less likely that the page will be added back to 723 * the inactive queue prematurely again. Here we check the 724 * page tables (or emulated bits, if any), given the upper 725 * level VM system not knowing anything about existing 726 * references. 727 */ 728 } else if (((m->flags & PG_REFERENCED) == 0) && 729 (actcount = pmap_ts_referenced(VM_PAGE_TO_PHYS(m)))) { 730 vm_page_activate(m); 731 m->act_count += (actcount + ACT_ADVANCE); 732 continue; 733 } 734 735 /* 736 * If the upper level VM system knows about any page 737 * references, we activate the page. We also set the 738 * "activation count" higher than normal so that we will less 739 * likely place pages back onto the inactive queue again. 740 */ 741 if ((m->flags & PG_REFERENCED) != 0) { 742 vm_page_flag_clear(m, PG_REFERENCED); 743 actcount = pmap_ts_referenced(VM_PAGE_TO_PHYS(m)); 744 vm_page_activate(m); 745 m->act_count += (actcount + ACT_ADVANCE + 1); 746 continue; 747 } 748 749 /* 750 * If the upper level VM system doesn't know anything about 751 * the page being dirty, we have to check for it again. As 752 * far as the VM code knows, any partially dirty pages are 753 * fully dirty. 754 */ 755 if (m->dirty == 0) { 756 vm_page_test_dirty(m); 757 } else { 758 vm_page_dirty(m); 759 } 760 761 /* 762 * Invalid pages can be easily freed 763 */ 764 if (m->valid == 0) { 765 vm_pageout_page_free(m); 766 cnt.v_dfree++; 767 --page_shortage; 768 769 /* 770 * Clean pages can be placed onto the cache queue. 771 */ 772 } else if (m->dirty == 0) { 773 vm_page_cache(m); 774 --page_shortage; 775 776 /* 777 * Dirty pages need to be paged out. Note that we clean 778 * only a limited number of pages per pagedaemon pass. 779 */ 780 } else if (maxlaunder > 0) { 781 int written; 782 int swap_pageouts_ok; 783 struct vnode *vp = NULL; 784 785 object = m->object; 786 787 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 788 swap_pageouts_ok = 1; 789 } else { 790 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 791 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 792 (cnt.v_free_count + cnt.v_cache_count) < cnt.v_free_min); 793 794 } 795 796 /* 797 * We don't bother paging objects that are "dead". 798 * Those objects are in a "rundown" state. 799 */ 800 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 801 s = splvm(); 802 TAILQ_REMOVE(&vm_page_queue_inactive, m, pageq); 803 TAILQ_INSERT_TAIL(&vm_page_queue_inactive, m, pageq); 804 splx(s); 805 continue; 806 } 807 808 /* 809 * For now we protect against potential memory 810 * deadlocks by requiring significant memory to be 811 * free if the object is not OBJT_DEFAULT or OBJT_SWAP. 812 * We do not 'trust' any other object type to operate 813 * with low memory, not even OBJT_DEVICE. The VM 814 * allocator will special case allocations done by 815 * the pageout daemon so the check below actually 816 * does have some hysteresis in it. It isn't the best 817 * solution, though. 818 */ 819 820 if (object->type != OBJT_DEFAULT && 821 object->type != OBJT_SWAP && 822 cnt.v_free_count < cnt.v_free_reserved) { 823 s = splvm(); 824 TAILQ_REMOVE(&vm_page_queue_inactive, m, pageq); 825 TAILQ_INSERT_TAIL(&vm_page_queue_inactive, m, 826 pageq); 827 splx(s); 828 continue; 829 } 830 831 /* 832 * Presumably we have sufficient free memory to do 833 * the more sophisticated checks and locking required 834 * for vnodes. 835 * 836 * The object is already known NOT to be dead. The 837 * vget() may still block, though, because 838 * VOP_ISLOCKED() doesn't check to see if an inode 839 * (v_data) is associated with the vnode. If it isn't, 840 * vget() will load in it from disk. Worse, vget() 841 * may actually get stuck waiting on "inode" if another 842 * process is in the process of bringing the inode in. 843 * This is bad news for us either way. 844 * 845 * So for the moment we check v_data == NULL as a 846 * workaround. This means that vnodes which do not 847 * use v_data in the way we expect probably will not 848 * wind up being paged out by the pager and it will be 849 * up to the syncer to get them. That's better then 850 * us blocking here. 851 * 852 * This whole code section is bogus - we need to fix 853 * the vnode pager to handle vm_page_t's without us 854 * having to do any sophisticated VOP tests. 855 */ 856 857 if (object->type == OBJT_VNODE) { 858 vp = object->handle; 859 860 if (VOP_ISLOCKED(vp) || 861 vp->v_data == NULL || 862 vget(vp, LK_EXCLUSIVE|LK_NOOBJ, curproc)) { 863 if ((m->queue == PQ_INACTIVE) && 864 (m->hold_count == 0) && 865 (m->busy == 0) && 866 (m->flags & PG_BUSY) == 0) { 867 s = splvm(); 868 TAILQ_REMOVE(&vm_page_queue_inactive, m, pageq); 869 TAILQ_INSERT_TAIL(&vm_page_queue_inactive, m, pageq); 870 splx(s); 871 } 872 if (object->flags & OBJ_MIGHTBEDIRTY) 873 vnodes_skipped++; 874 continue; 875 } 876 877 /* 878 * The page might have been moved to another queue 879 * during potential blocking in vget() above. 880 */ 881 if (m->queue != PQ_INACTIVE) { 882 if (object->flags & OBJ_MIGHTBEDIRTY) 883 vnodes_skipped++; 884 vput(vp); 885 continue; 886 } 887 888 /* 889 * The page may have been busied during the blocking in 890 * vput(); We don't move the page back onto the end of 891 * the queue so that statistics are more correct if we don't. 892 */ 893 if (m->busy || (m->flags & PG_BUSY)) { 894 vput(vp); 895 continue; 896 } 897 898 /* 899 * If the page has become held, then skip it 900 */ 901 if (m->hold_count) { 902 s = splvm(); 903 TAILQ_REMOVE(&vm_page_queue_inactive, m, pageq); 904 TAILQ_INSERT_TAIL(&vm_page_queue_inactive, m, pageq); 905 splx(s); 906 if (object->flags & OBJ_MIGHTBEDIRTY) 907 vnodes_skipped++; 908 vput(vp); 909 continue; 910 } 911 } 912 913 /* 914 * If a page is dirty, then it is either being washed 915 * (but not yet cleaned) or it is still in the 916 * laundry. If it is still in the laundry, then we 917 * start the cleaning operation. 918 */ 919 written = vm_pageout_clean(m); 920 if (vp) 921 vput(vp); 922 923 maxlaunder -= written; 924 } 925 } 926 927 /* 928 * If we still have a page shortage and we didn't launder anything, 929 * run the inactive scan again and launder something this time. 930 */ 931 932 if (launder_loop == 0 && page_shortage > 0) { 933 launder_loop = 1; 934 maxlaunder = 935 (cnt.v_inactive_target > max_page_launder) ? 936 max_page_launder : cnt.v_inactive_target; 937 goto rescan0; 938 } 939 940 /* 941 * Compute the page shortage from the point of view of having to 942 * move pages from the active queue to the inactive queue. 943 */ 944 945 page_shortage = (cnt.v_inactive_target + cnt.v_cache_min) - 946 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count); 947 page_shortage += addl_page_shortage; 948 949 /* 950 * Scan the active queue for things we can deactivate 951 */ 952 953 pcount = cnt.v_active_count; 954 m = TAILQ_FIRST(&vm_page_queue_active); 955 956 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) { 957 958 /* 959 * This is a consistancy check, and should likely be a panic 960 * or warning. 961 */ 962 if (m->queue != PQ_ACTIVE) { 963 break; 964 } 965 966 next = TAILQ_NEXT(m, pageq); 967 /* 968 * Don't deactivate pages that are busy. 969 */ 970 if ((m->busy != 0) || 971 (m->flags & PG_BUSY) || 972 (m->hold_count != 0)) { 973 s = splvm(); 974 TAILQ_REMOVE(&vm_page_queue_active, m, pageq); 975 TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq); 976 splx(s); 977 m = next; 978 continue; 979 } 980 981 /* 982 * The count for pagedaemon pages is done after checking the 983 * page for eligbility... 984 */ 985 cnt.v_pdpages++; 986 987 /* 988 * Check to see "how much" the page has been used. 989 */ 990 actcount = 0; 991 if (m->object->ref_count != 0) { 992 if (m->flags & PG_REFERENCED) { 993 actcount += 1; 994 } 995 actcount += pmap_ts_referenced(VM_PAGE_TO_PHYS(m)); 996 if (actcount) { 997 m->act_count += ACT_ADVANCE + actcount; 998 if (m->act_count > ACT_MAX) 999 m->act_count = ACT_MAX; 1000 } 1001 } 1002 1003 /* 1004 * Since we have "tested" this bit, we need to clear it now. 1005 */ 1006 vm_page_flag_clear(m, PG_REFERENCED); 1007 1008 /* 1009 * Only if an object is currently being used, do we use the 1010 * page activation count stats. 1011 */ 1012 if (actcount && (m->object->ref_count != 0)) { 1013 s = splvm(); 1014 TAILQ_REMOVE(&vm_page_queue_active, m, pageq); 1015 TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq); 1016 splx(s); 1017 } else { 1018 m->act_count -= min(m->act_count, ACT_DECLINE); 1019 if (vm_pageout_algorithm_lru || 1020 (m->object->ref_count == 0) || (m->act_count == 0)) { 1021 page_shortage--; 1022 if (m->object->ref_count == 0) { 1023 vm_page_protect(m, VM_PROT_NONE); 1024 if (m->dirty == 0) 1025 vm_page_cache(m); 1026 else 1027 vm_page_deactivate(m); 1028 } else { 1029 vm_page_deactivate(m); 1030 } 1031 } else { 1032 s = splvm(); 1033 TAILQ_REMOVE(&vm_page_queue_active, m, pageq); 1034 TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq); 1035 splx(s); 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_page_list_find(PQ_CACHE, cache_rover, FALSE); 1053 if (!m) 1054 break; 1055 if ((m->flags & PG_BUSY) || m->busy || m->hold_count || m->wire_count) { 1056#ifdef INVARIANTS 1057 printf("Warning: busy page %p found in cache\n", m); 1058#endif 1059 vm_page_deactivate(m); 1060 continue; 1061 } 1062 cache_rover = (cache_rover + PQ_PRIME2) & PQ_L2_MASK; 1063 vm_pageout_page_free(m); 1064 cnt.v_dfree++; 1065 } 1066 splx(s); 1067 1068#if !defined(NO_SWAPPING) 1069 /* 1070 * Idle process swapout -- run once per second. 1071 */ 1072 if (vm_swap_idle_enabled) { 1073 static long lsec; 1074 if (time_second != lsec) { 1075 vm_pageout_req_swapout |= VM_SWAP_IDLE; 1076 vm_req_vmdaemon(); 1077 lsec = time_second; 1078 } 1079 } 1080#endif 1081 1082 /* 1083 * If we didn't get enough free pages, and we have skipped a vnode 1084 * in a writeable object, wakeup the sync daemon. And kick swapout 1085 * if we did not get enough free pages. 1086 */ 1087 if ((cnt.v_cache_count + cnt.v_free_count) < 1088 (cnt.v_free_target + cnt.v_cache_min) ) { 1089 if (vnodes_skipped && 1090 (cnt.v_cache_count + cnt.v_free_count) < cnt.v_free_min) { 1091 (void) speedup_syncer(); 1092 } 1093#if !defined(NO_SWAPPING) 1094 if (vm_swap_enabled && 1095 (cnt.v_free_count + cnt.v_cache_count < cnt.v_free_target)) { 1096 vm_req_vmdaemon(); 1097 vm_pageout_req_swapout |= VM_SWAP_NORMAL; 1098 } 1099#endif 1100 } 1101 1102 /* 1103 * make sure that we have swap space -- if we are low on memory and 1104 * swap -- then kill the biggest process. 1105 */ 1106 if ((vm_swap_size == 0 || swap_pager_full) && 1107 ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_free_min)) { 1108 bigproc = NULL; 1109 bigsize = 0; 1110 for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) { 1111 /* 1112 * if this is a system process, skip it 1113 */ 1114 if ((p->p_flag & P_SYSTEM) || (p->p_lock > 0) || 1115 (p->p_pid == 1) || 1116 ((p->p_pid < 48) && (vm_swap_size != 0))) { 1117 continue; 1118 } 1119 /* 1120 * if the process is in a non-running type state, 1121 * don't touch it. 1122 */ 1123 if (p->p_stat != SRUN && p->p_stat != SSLEEP) { 1124 continue; 1125 } 1126 /* 1127 * get the process size 1128 */ 1129 size = vmspace_resident_count(p->p_vmspace); 1130 /* 1131 * if the this process is bigger than the biggest one 1132 * remember it. 1133 */ 1134 if (size > bigsize) { 1135 bigproc = p; 1136 bigsize = size; 1137 } 1138 } 1139 if (bigproc != NULL) { 1140 killproc(bigproc, "out of swap space"); 1141 bigproc->p_estcpu = 0; 1142 bigproc->p_nice = PRIO_MIN; 1143 resetpriority(bigproc); 1144 wakeup(&cnt.v_free_count); 1145 } 1146 } 1147 return force_wakeup; 1148} 1149 1150/* 1151 * This routine tries to maintain the pseudo LRU active queue, 1152 * so that during long periods of time where there is no paging, 1153 * that some statistic accumlation still occurs. This code 1154 * helps the situation where paging just starts to occur. 1155 */ 1156static void 1157vm_pageout_page_stats() 1158{ 1159 int s; 1160 vm_page_t m,next; 1161 int pcount,tpcount; /* Number of pages to check */ 1162 static int fullintervalcount = 0; 1163 int page_shortage; 1164 1165 page_shortage = (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) - 1166 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count); 1167 if (page_shortage <= 0) 1168 return; 1169 1170 pcount = cnt.v_active_count; 1171 fullintervalcount += vm_pageout_stats_interval; 1172 if (fullintervalcount < vm_pageout_full_stats_interval) { 1173 tpcount = (vm_pageout_stats_max * cnt.v_active_count) / cnt.v_page_count; 1174 if (pcount > tpcount) 1175 pcount = tpcount; 1176 } 1177 1178 m = TAILQ_FIRST(&vm_page_queue_active); 1179 while ((m != NULL) && (pcount-- > 0)) { 1180 int actcount; 1181 1182 if (m->queue != PQ_ACTIVE) { 1183 break; 1184 } 1185 1186 next = TAILQ_NEXT(m, pageq); 1187 /* 1188 * Don't deactivate pages that are busy. 1189 */ 1190 if ((m->busy != 0) || 1191 (m->flags & PG_BUSY) || 1192 (m->hold_count != 0)) { 1193 s = splvm(); 1194 TAILQ_REMOVE(&vm_page_queue_active, m, pageq); 1195 TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq); 1196 splx(s); 1197 m = next; 1198 continue; 1199 } 1200 1201 actcount = 0; 1202 if (m->flags & PG_REFERENCED) { 1203 vm_page_flag_clear(m, PG_REFERENCED); 1204 actcount += 1; 1205 } 1206 1207 actcount += pmap_ts_referenced(VM_PAGE_TO_PHYS(m)); 1208 if (actcount) { 1209 m->act_count += ACT_ADVANCE + actcount; 1210 if (m->act_count > ACT_MAX) 1211 m->act_count = ACT_MAX; 1212 s = splvm(); 1213 TAILQ_REMOVE(&vm_page_queue_active, m, pageq); 1214 TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq); 1215 splx(s); 1216 } else { 1217 if (m->act_count == 0) { 1218 /* 1219 * We turn off page access, so that we have more accurate 1220 * RSS stats. We don't do this in the normal page deactivation 1221 * when the system is loaded VM wise, because the cost of 1222 * the large number of page protect operations would be higher 1223 * than the value of doing the operation. 1224 */ 1225 vm_page_protect(m, VM_PROT_NONE); 1226 vm_page_deactivate(m); 1227 } else { 1228 m->act_count -= min(m->act_count, ACT_DECLINE); 1229 s = splvm(); 1230 TAILQ_REMOVE(&vm_page_queue_active, m, pageq); 1231 TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq); 1232 splx(s); 1233 } 1234 } 1235 1236 m = next; 1237 } 1238} 1239 1240static int 1241vm_pageout_free_page_calc(count) 1242vm_size_t count; 1243{ 1244 if (count < cnt.v_page_count) 1245 return 0; 1246 /* 1247 * free_reserved needs to include enough for the largest swap pager 1248 * structures plus enough for any pv_entry structs when paging. 1249 */ 1250 if (cnt.v_page_count > 1024) 1251 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200; 1252 else 1253 cnt.v_free_min = 4; 1254 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1255 cnt.v_interrupt_free_min; 1256 cnt.v_free_reserved = vm_pageout_page_count + 1257 cnt.v_pageout_free_min + (count / 768) + PQ_L2_SIZE; 1258 cnt.v_free_min += cnt.v_free_reserved; 1259 return 1; 1260} 1261 1262 1263/* 1264 * vm_pageout is the high level pageout daemon. 1265 */ 1266static void 1267vm_pageout() 1268{ 1269 /* 1270 * Initialize some paging parameters. 1271 */ 1272 1273 cnt.v_interrupt_free_min = 2; 1274 if (cnt.v_page_count < 2000) 1275 vm_pageout_page_count = 8; 1276 1277 vm_pageout_free_page_calc(cnt.v_page_count); 1278 /* 1279 * free_reserved needs to include enough for the largest swap pager 1280 * structures plus enough for any pv_entry structs when paging. 1281 */ 1282 if (cnt.v_free_count > 6144) 1283 cnt.v_free_target = 3 * cnt.v_free_min + cnt.v_free_reserved; 1284 else 1285 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved; 1286 1287 if (cnt.v_free_count > 2048) { 1288 cnt.v_cache_min = cnt.v_free_target; 1289 cnt.v_cache_max = 2 * cnt.v_cache_min; 1290 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2; 1291 } else { 1292 cnt.v_cache_min = 0; 1293 cnt.v_cache_max = 0; 1294 cnt.v_inactive_target = cnt.v_free_count / 4; 1295 } 1296 if (cnt.v_inactive_target > cnt.v_free_count / 3) 1297 cnt.v_inactive_target = cnt.v_free_count / 3; 1298 1299 /* XXX does not really belong here */ 1300 if (vm_page_max_wired == 0) 1301 vm_page_max_wired = cnt.v_free_count / 3; 1302 1303 if (vm_pageout_stats_max == 0) 1304 vm_pageout_stats_max = cnt.v_free_target; 1305 1306 /* 1307 * Set interval in seconds for stats scan. 1308 */ 1309 if (vm_pageout_stats_interval == 0) 1310 vm_pageout_stats_interval = 5; 1311 if (vm_pageout_full_stats_interval == 0) 1312 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4; 1313 1314 1315 /* 1316 * Set maximum free per pass 1317 */ 1318 if (vm_pageout_stats_free_max == 0) 1319 vm_pageout_stats_free_max = 5; 1320 1321 max_page_launder = (cnt.v_page_count > 1800 ? 32 : 16); 1322 1323 curproc->p_flag |= P_BUFEXHAUST; 1324 swap_pager_swap_init(); 1325 /* 1326 * The pageout daemon is never done, so loop forever. 1327 */ 1328 while (TRUE) { 1329 int error; 1330 int s = splvm(); 1331 if (!vm_pages_needed || 1332 ((cnt.v_free_count + cnt.v_cache_count) > cnt.v_free_min)) { 1333 vm_pages_needed = 0; 1334 error = tsleep(&vm_pages_needed, 1335 PVM, "psleep", vm_pageout_stats_interval * hz); 1336 if (error && !vm_pages_needed) { 1337 splx(s); 1338 vm_pageout_page_stats(); 1339 continue; 1340 } 1341 } else if (vm_pages_needed) { 1342 vm_pages_needed = 0; 1343 tsleep(&vm_pages_needed, PVM, "psleep", hz/2); 1344 } 1345 1346 if (vm_pages_needed) 1347 cnt.v_pdwakeups++; 1348 vm_pages_needed = 0; 1349 splx(s); 1350 vm_pageout_scan(); 1351 vm_pageout_deficit = 0; 1352 wakeup(&cnt.v_free_count); 1353 } 1354} 1355 1356void 1357pagedaemon_wakeup() 1358{ 1359 if (!vm_pages_needed && curproc != pageproc) { 1360 vm_pages_needed++; 1361 wakeup(&vm_pages_needed); 1362 } 1363} 1364 1365#if !defined(NO_SWAPPING) 1366static void 1367vm_req_vmdaemon() 1368{ 1369 static int lastrun = 0; 1370 1371 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1372 wakeup(&vm_daemon_needed); 1373 lastrun = ticks; 1374 } 1375} 1376 1377static void 1378vm_daemon() 1379{ 1380 struct proc *p; 1381 1382 while (TRUE) { 1383 tsleep(&vm_daemon_needed, PPAUSE, "psleep", 0); 1384 if (vm_pageout_req_swapout) { 1385 swapout_procs(vm_pageout_req_swapout); 1386 vm_pageout_req_swapout = 0; 1387 } 1388 /* 1389 * scan the processes for exceeding their rlimits or if 1390 * process is swapped out -- deactivate pages 1391 */ 1392 1393 for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) { 1394 vm_pindex_t limit, size; 1395 1396 /* 1397 * if this is a system process or if we have already 1398 * looked at this process, skip it. 1399 */ 1400 if (p->p_flag & (P_SYSTEM | P_WEXIT)) { 1401 continue; 1402 } 1403 /* 1404 * if the process is in a non-running type state, 1405 * don't touch it. 1406 */ 1407 if (p->p_stat != SRUN && p->p_stat != SSLEEP) { 1408 continue; 1409 } 1410 /* 1411 * get a limit 1412 */ 1413 limit = OFF_TO_IDX( 1414 qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur, 1415 p->p_rlimit[RLIMIT_RSS].rlim_max)); 1416 1417 /* 1418 * let processes that are swapped out really be 1419 * swapped out set the limit to nothing (will force a 1420 * swap-out.) 1421 */ 1422 if ((p->p_flag & P_INMEM) == 0) 1423 limit = 0; /* XXX */ 1424 1425 size = vmspace_resident_count(p->p_vmspace); 1426 if (limit >= 0 && size >= limit) { 1427 vm_pageout_map_deactivate_pages( 1428 &p->p_vmspace->vm_map, limit); 1429 } 1430 } 1431 } 1432} 1433#endif 1434