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 1.141 1999/04/23 20:29:57 dt 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/resourcevar.h> 81 #include <sys/signalvar.h> 82 #include <sys/vnode.h> 83 #include <sys/vmmeter.h> 84 #include <sys/sysctl.h> 85 86 #include <vm/vm.h> 87 #include <vm/vm_param.h> 88 #include <vm/vm_prot.h> 89 #include <sys/lock.h> 90 #include <vm/vm_object.h> 91 #include <vm/vm_page.h> 92 #include <vm/vm_map.h> 93 #include <vm/vm_pageout.h> 94 #include <vm/vm_pager.h> 95 #include <vm/swap_pager.h> 96 #include <vm/vm_extern.h> 97 98 /* 99 * System initialization 100 */ 101 102 /* the kernel process "vm_pageout"*/ 103 static void vm_pageout __P((void)); 104 static int vm_pageout_clean __P((vm_page_t)); 105 static int vm_pageout_scan __P((void)); 106 static int vm_pageout_free_page_calc __P((vm_size_t count)); 107 struct proc *pageproc; 108 109 static struct kproc_desc page_kp = { 110 "pagedaemon", 111 vm_pageout, 112 &pageproc 113 }; 114 SYSINIT_KT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp) 115 116 #if !defined(NO_SWAPPING) 117 /* the kernel process "vm_daemon"*/ 118 static void vm_daemon __P((void)); 119 static struct proc *vmproc; 120 121 static struct kproc_desc vm_kp = { 122 "vmdaemon", 123 vm_daemon, 124 &vmproc 125 }; 126 SYSINIT_KT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp) 127 #endif 128 129 130 int vm_pages_needed=0; /* Event on which pageout daemon sleeps */ 131 int vm_pageout_deficit=0; /* Estimated number of pages deficit */ 132 int vm_pageout_pages_needed=0; /* flag saying that the pageout daemon needs pages */ 133 134 extern int npendingio; 135 #if !defined(NO_SWAPPING) 136 static int vm_pageout_req_swapout; /* XXX */ 137 static int vm_daemon_needed; 138 #endif 139 extern int nswiodone; 140 extern int vm_swap_size; 141 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0; 142 static int vm_pageout_full_stats_interval = 0; 143 static int vm_pageout_stats_free_max=0, vm_pageout_algorithm_lru=0; 144 static int defer_swap_pageouts=0; 145 static int disable_swap_pageouts=0; 146 147 static int max_page_launder=100; 148 #if defined(NO_SWAPPING) 149 static int vm_swap_enabled=0; 150 static int vm_swap_idle_enabled=0; 151 #else 152 static int vm_swap_enabled=1; 153 static int vm_swap_idle_enabled=0; 154 #endif 155 156 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm, 157 CTLFLAG_RW, &vm_pageout_algorithm_lru, 0, "LRU page mgmt"); 158 159 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max, 160 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length"); 161 162 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval, 163 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan"); 164 165 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval, 166 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan"); 167 168 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_free_max, 169 CTLFLAG_RW, &vm_pageout_stats_free_max, 0, "Not implemented"); 170 171 #if defined(NO_SWAPPING) 172 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 173 CTLFLAG_RD, &vm_swap_enabled, 0, ""); 174 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 175 CTLFLAG_RD, &vm_swap_idle_enabled, 0, ""); 176 #else 177 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 178 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout"); 179 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 180 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 181 #endif 182 183 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts, 184 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem"); 185 186 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts, 187 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages"); 188 189 SYSCTL_INT(_vm, OID_AUTO, max_page_launder, 190 CTLFLAG_RW, &max_page_launder, 0, "Maximum number of pages to clean per pass"); 191 192 193 #define VM_PAGEOUT_PAGE_COUNT 16 194 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT; 195 196 int vm_page_max_wired; /* XXX max # of wired pages system-wide */ 197 198 #if !defined(NO_SWAPPING) 199 typedef void freeer_fcn_t __P((vm_map_t, vm_object_t, vm_pindex_t, int)); 200 static void vm_pageout_map_deactivate_pages __P((vm_map_t, vm_pindex_t)); 201 static freeer_fcn_t vm_pageout_object_deactivate_pages; 202 static void vm_req_vmdaemon __P((void)); 203 #endif 204 static void vm_pageout_page_stats(void); 205 206 /* 207 * vm_pageout_clean: 208 * 209 * Clean the page and remove it from the laundry. 210 * 211 * We set the busy bit to cause potential page faults on this page to 212 * block. Note the careful timing, however, the busy bit isn't set till 213 * late and we cannot do anything that will mess with the page. 214 */ 215 216 static int 217 vm_pageout_clean(m) 218 vm_page_t m; 219 { 220 register vm_object_t object; 221 vm_page_t mc[2*vm_pageout_page_count]; 222 int pageout_count; 223 int i, forward_okay, backward_okay, page_base; 224 vm_pindex_t pindex = m->pindex; 225 226 object = m->object; 227 228 /* 229 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP 230 * with the new swapper, but we could have serious problems paging 231 * out other object types if there is insufficient memory. 232 * 233 * Unfortunately, checking free memory here is far too late, so the 234 * check has been moved up a procedural level. 235 */ 236 237 #if 0 238 /* 239 * If not OBJT_SWAP, additional memory may be needed to do the pageout. 240 * Try to avoid the deadlock. 241 */ 242 if ((object->type == OBJT_DEFAULT) && 243 ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_pageout_free_min)) 244 return 0; 245 #endif 246 247 /* 248 * Don't mess with the page if it's busy. 249 */ 250 if ((m->hold_count != 0) || 251 ((m->busy != 0) || (m->flags & PG_BUSY))) 252 return 0; 253 254 #if 0 255 /* 256 * XXX REMOVED XXX. vm_object_collapse() can block, which can 257 * change the page state. Calling vm_object_collapse() might also 258 * destroy or rename the page because we have not busied it yet!!! 259 * So this code segment is removed. 260 */ 261 /* 262 * Try collapsing before it's too late. XXX huh? Why are we doing 263 * this here? 264 */ 265 if (object->backing_object) { 266 vm_object_collapse(object); 267 } 268 #endif 269 270 mc[vm_pageout_page_count] = m; 271 pageout_count = 1; 272 page_base = vm_pageout_page_count; 273 forward_okay = TRUE; 274 if (pindex != 0) 275 backward_okay = TRUE; 276 else 277 backward_okay = FALSE; 278 /* 279 * Scan object for clusterable pages. 280 * 281 * We can cluster ONLY if: ->> the page is NOT 282 * clean, wired, busy, held, or mapped into a 283 * buffer, and one of the following: 284 * 1) The page is inactive, or a seldom used 285 * active page. 286 * -or- 287 * 2) we force the issue. 288 */ 289 for (i = 1; (i < vm_pageout_page_count) && (forward_okay || backward_okay); i++) { 290 vm_page_t p; 291 292 /* 293 * See if forward page is clusterable. 294 */ 295 if (forward_okay) { 296 /* 297 * Stop forward scan at end of object. 298 */ 299 if ((pindex + i) > object->size) { 300 forward_okay = FALSE; 301 goto do_backward; 302 } 303 p = vm_page_lookup(object, pindex + i); 304 if (p) { 305 if (((p->queue - p->pc) == PQ_CACHE) || 306 (p->flags & PG_BUSY) || p->busy) { 307 forward_okay = FALSE; 308 goto do_backward; 309 } 310 vm_page_test_dirty(p); 311 if ((p->dirty & p->valid) != 0 && 312 (p->queue == PQ_INACTIVE) && 313 (p->wire_count == 0) && 314 (p->hold_count == 0)) { 315 mc[vm_pageout_page_count + i] = p; 316 pageout_count++; 317 if (pageout_count == vm_pageout_page_count) 318 break; 319 } else { 320 forward_okay = FALSE; 321 } 322 } else { 323 forward_okay = FALSE; 324 } 325 } 326 do_backward: 327 /* 328 * See if backward page is clusterable. 329 */ 330 if (backward_okay) { 331 /* 332 * Stop backward scan at beginning of object. 333 */ 334 if ((pindex - i) == 0) { 335 backward_okay = FALSE; 336 } 337 p = vm_page_lookup(object, pindex - i); 338 if (p) { 339 if (((p->queue - p->pc) == PQ_CACHE) || 340 (p->flags & PG_BUSY) || p->busy) { 341 backward_okay = FALSE; 342 continue; 343 } 344 vm_page_test_dirty(p); 345 if ((p->dirty & p->valid) != 0 && 346 (p->queue == PQ_INACTIVE) && 347 (p->wire_count == 0) && 348 (p->hold_count == 0)) { 349 mc[vm_pageout_page_count - i] = p; 350 pageout_count++; 351 page_base--; 352 if (pageout_count == vm_pageout_page_count) 353 break; 354 } else { 355 backward_okay = FALSE; 356 } 357 } else { 358 backward_okay = FALSE; 359 } 360 } 361 } 362 363 /* 364 * we allow reads during pageouts... 365 */ 366 return vm_pageout_flush(&mc[page_base], pageout_count, 0); 367 } 368 369 /* 370 * vm_pageout_flush() - launder the given pages 371 * 372 * The given pages are laundered. Note that we setup for the start of 373 * I/O ( i.e. busy the page ), mark it read-only, and bump the object 374 * reference count all in here rather then in the parent. If we want 375 * the parent to do more sophisticated things we may have to change 376 * the ordering. 377 */ 378 379 int 380 vm_pageout_flush(mc, count, flags) 381 vm_page_t *mc; 382 int count; 383 int flags; 384 { 385 register vm_object_t object; 386 int pageout_status[count]; 387 int numpagedout = 0; 388 int i; 389 390 /* 391 * Initiate I/O. Bump the vm_page_t->busy counter and 392 * mark the pages read-only. 393 * 394 * We do not have to fixup the clean/dirty bits here... we can 395 * allow the pager to do it after the I/O completes. 396 */ 397 398 for (i = 0; i < count; i++) { 399 vm_page_io_start(mc[i]); 400 vm_page_protect(mc[i], VM_PROT_READ); 401 } 402 403 object = mc[0]->object; 404 vm_object_pip_add(object, count); 405 406 vm_pager_put_pages(object, mc, count, 407 (flags | ((object == kernel_object) ? OBJPC_SYNC : 0)), 408 pageout_status); 409 410 for (i = 0; i < count; i++) { 411 vm_page_t mt = mc[i]; 412 413 switch (pageout_status[i]) { 414 case VM_PAGER_OK: 415 numpagedout++; 416 break; 417 case VM_PAGER_PEND: 418 numpagedout++; 419 break; 420 case VM_PAGER_BAD: 421 /* 422 * Page outside of range of object. Right now we 423 * essentially lose the changes by pretending it 424 * worked. 425 */ 426 pmap_clear_modify(VM_PAGE_TO_PHYS(mt)); 427 mt->dirty = 0; 428 break; 429 case VM_PAGER_ERROR: 430 case VM_PAGER_FAIL: 431 /* 432 * If page couldn't be paged out, then reactivate the 433 * page so it doesn't clog the inactive list. (We 434 * will try paging out it again later). 435 */ 436 vm_page_activate(mt); 437 break; 438 case VM_PAGER_AGAIN: 439 break; 440 } 441 442 /* 443 * If the operation is still going, leave the page busy to 444 * block all other accesses. Also, leave the paging in 445 * progress indicator set so that we don't attempt an object 446 * collapse. 447 */ 448 if (pageout_status[i] != VM_PAGER_PEND) { 449 vm_object_pip_wakeup(object); 450 vm_page_io_finish(mt); 451 } 452 } 453 return numpagedout; 454 } 455 456 #if !defined(NO_SWAPPING) 457 /* 458 * vm_pageout_object_deactivate_pages 459 * 460 * deactivate enough pages to satisfy the inactive target 461 * requirements or if vm_page_proc_limit is set, then 462 * deactivate all of the pages in the object and its 463 * backing_objects. 464 * 465 * The object and map must be locked. 466 */ 467 static void 468 vm_pageout_object_deactivate_pages(map, object, desired, map_remove_only) 469 vm_map_t map; 470 vm_object_t object; 471 vm_pindex_t desired; 472 int map_remove_only; 473 { 474 register vm_page_t p, next; 475 int rcount; 476 int remove_mode; 477 int s; 478 479 if (object->type == OBJT_DEVICE) 480 return; 481 482 while (object) { 483 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 484 return; 485 if (object->paging_in_progress) 486 return; 487 488 remove_mode = map_remove_only; 489 if (object->shadow_count > 1) 490 remove_mode = 1; 491 /* 492 * scan the objects entire memory queue 493 */ 494 rcount = object->resident_page_count; 495 p = TAILQ_FIRST(&object->memq); 496 while (p && (rcount-- > 0)) { 497 int actcount; 498 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 499 return; 500 next = TAILQ_NEXT(p, listq); 501 cnt.v_pdpages++; 502 if (p->wire_count != 0 || 503 p->hold_count != 0 || 504 p->busy != 0 || 505 (p->flags & PG_BUSY) || 506 !pmap_page_exists(vm_map_pmap(map), VM_PAGE_TO_PHYS(p))) { 507 p = next; 508 continue; 509 } 510 511 actcount = pmap_ts_referenced(VM_PAGE_TO_PHYS(p)); 512 if (actcount) { 513 vm_page_flag_set(p, PG_REFERENCED); 514 } else if (p->flags & PG_REFERENCED) { 515 actcount = 1; 516 } 517 518 if ((p->queue != PQ_ACTIVE) && 519 (p->flags & PG_REFERENCED)) { 520 vm_page_activate(p); 521 p->act_count += actcount; 522 vm_page_flag_clear(p, PG_REFERENCED); 523 } else if (p->queue == PQ_ACTIVE) { 524 if ((p->flags & PG_REFERENCED) == 0) { 525 p->act_count -= min(p->act_count, ACT_DECLINE); 526 if (!remove_mode && (vm_pageout_algorithm_lru || (p->act_count == 0))) { 527 vm_page_protect(p, VM_PROT_NONE); 528 vm_page_deactivate(p); 529 } else { 530 s = splvm(); 531 TAILQ_REMOVE(&vm_page_queue_active, p, pageq); 532 TAILQ_INSERT_TAIL(&vm_page_queue_active, p, pageq); 533 splx(s); 534 } 535 } else { 536 vm_page_activate(p); 537 vm_page_flag_clear(p, PG_REFERENCED); 538 if (p->act_count < (ACT_MAX - ACT_ADVANCE)) 539 p->act_count += ACT_ADVANCE; 540 s = splvm(); 541 TAILQ_REMOVE(&vm_page_queue_active, p, pageq); 542 TAILQ_INSERT_TAIL(&vm_page_queue_active, p, pageq); 543 splx(s); 544 } 545 } else if (p->queue == PQ_INACTIVE) { 546 vm_page_protect(p, VM_PROT_NONE); 547 } 548 p = next; 549 } 550 object = object->backing_object; 551 } 552 return; 553 } 554 555 /* 556 * deactivate some number of pages in a map, try to do it fairly, but 557 * that is really hard to do. 558 */ 559 static void 560 vm_pageout_map_deactivate_pages(map, desired) 561 vm_map_t map; 562 vm_pindex_t desired; 563 { 564 vm_map_entry_t tmpe; 565 vm_object_t obj, bigobj; 566 567 if (lockmgr(&map->lock, LK_EXCLUSIVE | LK_NOWAIT, (void *)0, curproc)) { 568 return; 569 } 570 571 bigobj = NULL; 572 573 /* 574 * first, search out the biggest object, and try to free pages from 575 * that. 576 */ 577 tmpe = map->header.next; 578 while (tmpe != &map->header) { 579 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 580 obj = tmpe->object.vm_object; 581 if ((obj != NULL) && (obj->shadow_count <= 1) && 582 ((bigobj == NULL) || 583 (bigobj->resident_page_count < obj->resident_page_count))) { 584 bigobj = obj; 585 } 586 } 587 tmpe = tmpe->next; 588 } 589 590 if (bigobj) 591 vm_pageout_object_deactivate_pages(map, bigobj, desired, 0); 592 593 /* 594 * Next, hunt around for other pages to deactivate. We actually 595 * do this search sort of wrong -- .text first is not the best idea. 596 */ 597 tmpe = map->header.next; 598 while (tmpe != &map->header) { 599 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 600 break; 601 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 602 obj = tmpe->object.vm_object; 603 if (obj) 604 vm_pageout_object_deactivate_pages(map, obj, desired, 0); 605 } 606 tmpe = tmpe->next; 607 }; 608 609 /* 610 * Remove all mappings if a process is swapped out, this will free page 611 * table pages. 612 */ 613 if (desired == 0) 614 pmap_remove(vm_map_pmap(map), 615 VM_MIN_ADDRESS, VM_MAXUSER_ADDRESS); 616 vm_map_unlock(map); 617 return; 618 } 619 #endif 620 621 /* 622 * Don't try to be fancy - being fancy can lead to VOP_LOCK's and therefore 623 * to vnode deadlocks. We only do it for OBJT_DEFAULT and OBJT_SWAP objects 624 * which we know can be trivially freed. 625 */ 626 627 void 628 vm_pageout_page_free(vm_page_t m) { 629 vm_object_t object = m->object; 630 int type = object->type; 631 632 if (type == OBJT_SWAP || type == OBJT_DEFAULT) 633 vm_object_reference(object); 634 vm_page_busy(m); 635 vm_page_protect(m, VM_PROT_NONE); 636 vm_page_free(m); 637 if (type == OBJT_SWAP || type == OBJT_DEFAULT) 638 vm_object_deallocate(object); 639 } 640 641 /* 642 * vm_pageout_scan does the dirty work for the pageout daemon. 643 */ 644 static int 645 vm_pageout_scan() 646 { 647 vm_page_t m, next; 648 int page_shortage, maxscan, pcount; 649 int addl_page_shortage, addl_page_shortage_init; 650 int maxlaunder; 651 int launder_loop = 0; 652 struct proc *p, *bigproc; 653 vm_offset_t size, bigsize; 654 vm_object_t object; 655 int force_wakeup = 0; 656 int actcount; 657 int vnodes_skipped = 0; 658 int s; 659 660 /* 661 * Do whatever cleanup that the pmap code can. 662 */ 663 pmap_collect(); 664 665 addl_page_shortage_init = vm_pageout_deficit; 666 vm_pageout_deficit = 0; 667 668 if (max_page_launder == 0) 669 max_page_launder = 1; 670 671 /* 672 * Calculate the number of pages we want to either free or move 673 * to the cache. 674 */ 675 676 page_shortage = (cnt.v_free_target + cnt.v_cache_min) - 677 (cnt.v_free_count + cnt.v_cache_count); 678 page_shortage += addl_page_shortage_init; 679 680 /* 681 * Figure out what to do with dirty pages when they are encountered. 682 * Assume that 1/3 of the pages on the inactive list are clean. If 683 * we think we can reach our target, disable laundering (do not 684 * clean any dirty pages). If we miss the target we will loop back 685 * up and do a laundering run. 686 */ 687 688 if (cnt.v_inactive_count / 3 > page_shortage) { 689 maxlaunder = 0; 690 launder_loop = 0; 691 } else { 692 maxlaunder = 693 (cnt.v_inactive_target > max_page_launder) ? 694 max_page_launder : cnt.v_inactive_target; 695 launder_loop = 1; 696 } 697 698 /* 699 * Start scanning the inactive queue for pages we can move to the 700 * cache or free. The scan will stop when the target is reached or 701 * we have scanned the entire inactive queue. 702 */ 703 704 rescan0: 705 addl_page_shortage = addl_page_shortage_init; 706 maxscan = cnt.v_inactive_count; 707 for ( 708 m = TAILQ_FIRST(&vm_page_queue_inactive); 709 m != NULL && maxscan-- > 0 && page_shortage > 0; 710 m = next 711 ) { 712 713 cnt.v_pdpages++; 714 715 if (m->queue != PQ_INACTIVE) { 716 goto rescan0; 717 } 718 719 next = TAILQ_NEXT(m, pageq); 720 721 if (m->hold_count) { 722 s = splvm(); 723 TAILQ_REMOVE(&vm_page_queue_inactive, m, pageq); 724 TAILQ_INSERT_TAIL(&vm_page_queue_inactive, m, pageq); 725 splx(s); 726 addl_page_shortage++; 727 continue; 728 } 729 /* 730 * Dont mess with busy pages, keep in the front of the 731 * queue, most likely are being paged out. 732 */ 733 if (m->busy || (m->flags & PG_BUSY)) { 734 addl_page_shortage++; 735 continue; 736 } 737 738 /* 739 * If the object is not being used, we ignore previous 740 * references. 741 */ 742 if (m->object->ref_count == 0) { 743 vm_page_flag_clear(m, PG_REFERENCED); 744 pmap_clear_reference(VM_PAGE_TO_PHYS(m)); 745 746 /* 747 * Otherwise, if the page has been referenced while in the 748 * inactive queue, we bump the "activation count" upwards, 749 * making it less likely that the page will be added back to 750 * the inactive queue prematurely again. Here we check the 751 * page tables (or emulated bits, if any), given the upper 752 * level VM system not knowing anything about existing 753 * references. 754 */ 755 } else if (((m->flags & PG_REFERENCED) == 0) && 756 (actcount = pmap_ts_referenced(VM_PAGE_TO_PHYS(m)))) { 757 vm_page_activate(m); 758 m->act_count += (actcount + ACT_ADVANCE); 759 continue; 760 } 761 762 /* 763 * If the upper level VM system knows about any page 764 * references, we activate the page. We also set the 765 * "activation count" higher than normal so that we will less 766 * likely place pages back onto the inactive queue again. 767 */ 768 if ((m->flags & PG_REFERENCED) != 0) { 769 vm_page_flag_clear(m, PG_REFERENCED); 770 actcount = pmap_ts_referenced(VM_PAGE_TO_PHYS(m)); 771 vm_page_activate(m); 772 m->act_count += (actcount + ACT_ADVANCE + 1); 773 continue; 774 } 775 776 /* 777 * If the upper level VM system doesn't know anything about 778 * the page being dirty, we have to check for it again. As 779 * far as the VM code knows, any partially dirty pages are 780 * fully dirty. 781 */ 782 if (m->dirty == 0) { 783 vm_page_test_dirty(m); 784 } else { 785 vm_page_dirty(m); 786 } 787 788 /* 789 * Invalid pages can be easily freed 790 */ 791 if (m->valid == 0) { 792 vm_pageout_page_free(m); 793 cnt.v_dfree++; 794 --page_shortage; 795 796 /* 797 * Clean pages can be placed onto the cache queue. 798 */ 799 } else if (m->dirty == 0) { 800 vm_page_cache(m); 801 --page_shortage; 802 803 /* 804 * Dirty pages need to be paged out. Note that we clean 805 * only a limited number of pages per pagedaemon pass. 806 */ 807 } else if (maxlaunder > 0) { 808 int written; 809 int swap_pageouts_ok; 810 struct vnode *vp = NULL; 811 812 object = m->object; 813 814 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 815 swap_pageouts_ok = 1; 816 } else { 817 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 818 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 819 (cnt.v_free_count + cnt.v_cache_count) < cnt.v_free_min); 820 821 } 822 823 /* 824 * We don't bother paging objects that are "dead". 825 * Those objects are in a "rundown" state. 826 */ 827 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 828 s = splvm(); 829 TAILQ_REMOVE(&vm_page_queue_inactive, m, pageq); 830 TAILQ_INSERT_TAIL(&vm_page_queue_inactive, m, pageq); 831 splx(s); 832 continue; 833 } 834 835 /* 836 * For now we protect against potential memory 837 * deadlocks by requiring significant memory to be 838 * free if the object is not OBJT_DEFAULT or OBJT_SWAP. 839 * We do not 'trust' any other object type to operate 840 * with low memory, not even OBJT_DEVICE. The VM 841 * allocator will special case allocations done by 842 * the pageout daemon so the check below actually 843 * does have some hysteresis in it. It isn't the best 844 * solution, though. 845 */ 846 847 if ( 848 object->type != OBJT_DEFAULT && 849 object->type != OBJT_SWAP && 850 cnt.v_free_count < cnt.v_free_reserved 851 ) { 852 s = splvm(); 853 TAILQ_REMOVE(&vm_page_queue_inactive, m, pageq); 854 TAILQ_INSERT_TAIL(&vm_page_queue_inactive, m, pageq); 855 splx(s); 856 continue; 857 } 858 859 /* 860 * Presumably we have sufficient free memory to do 861 * the more sophisticated checks and locking required 862 * for vnodes. 863 * 864 * The object is already known NOT to be dead. The 865 * vget() may still block, though, because 866 * VOP_ISLOCKED() doesn't check to see if an inode 867 * (v_data) is associated with the vnode. If it isn't, 868 * vget() will load in it from disk. Worse, vget() 869 * may actually get stuck waiting on "inode" if another 870 * process is in the process of bringing the inode in. 871 * This is bad news for us either way. 872 * 873 * So for the moment we check v_data == NULL as a 874 * workaround. This means that vnodes which do not 875 * use v_data in the way we expect probably will not 876 * wind up being paged out by the pager and it will be 877 * up to the syncer to get them. That's better then 878 * us blocking here. 879 * 880 * This whole code section is bogus - we need to fix 881 * the vnode pager to handle vm_page_t's without us 882 * having to do any sophisticated VOP tests. 883 */ 884 885 if (object->type == OBJT_VNODE) { 886 vp = object->handle; 887 888 if (VOP_ISLOCKED(vp) || 889 vp->v_data == NULL || 890 vget(vp, LK_EXCLUSIVE|LK_NOOBJ, curproc)) { 891 if ((m->queue == PQ_INACTIVE) && 892 (m->hold_count == 0) && 893 (m->busy == 0) && 894 (m->flags & PG_BUSY) == 0) { 895 s = splvm(); 896 TAILQ_REMOVE(&vm_page_queue_inactive, m, pageq); 897 TAILQ_INSERT_TAIL(&vm_page_queue_inactive, m, pageq); 898 splx(s); 899 } 900 if (object->flags & OBJ_MIGHTBEDIRTY) 901 vnodes_skipped++; 902 continue; 903 } 904 905 /* 906 * The page might have been moved to another queue 907 * during potential blocking in vget() above. 908 */ 909 if (m->queue != PQ_INACTIVE) { 910 if (object->flags & OBJ_MIGHTBEDIRTY) 911 vnodes_skipped++; 912 vput(vp); 913 continue; 914 } 915 916 /* 917 * The page may have been busied during the blocking in 918 * vput(); We don't move the page back onto the end of 919 * the queue so that statistics are more correct if we don't. 920 */ 921 if (m->busy || (m->flags & PG_BUSY)) { 922 vput(vp); 923 continue; 924 } 925 926 /* 927 * If the page has become held, then skip it 928 */ 929 if (m->hold_count) { 930 s = splvm(); 931 TAILQ_REMOVE(&vm_page_queue_inactive, m, pageq); 932 TAILQ_INSERT_TAIL(&vm_page_queue_inactive, m, pageq); 933 splx(s); 934 if (object->flags & OBJ_MIGHTBEDIRTY) 935 vnodes_skipped++; 936 vput(vp); 937 continue; 938 } 939 } 940 941 /* 942 * If a page is dirty, then it is either being washed 943 * (but not yet cleaned) or it is still in the 944 * laundry. If it is still in the laundry, then we 945 * start the cleaning operation. 946 */ 947 written = vm_pageout_clean(m); 948 if (vp) 949 vput(vp); 950 951 maxlaunder -= written; 952 } 953 } 954 955 /* 956 * If we still have a page shortage and we didn't launder anything, 957 * run the inactive scan again and launder something this time. 958 */ 959 960 if (launder_loop == 0 && page_shortage > 0) { 961 launder_loop = 1; 962 maxlaunder = 963 (cnt.v_inactive_target > max_page_launder) ? 964 max_page_launder : cnt.v_inactive_target; 965 goto rescan0; 966 } 967 968 /* 969 * Compute the page shortage from the point of view of having to 970 * move pages from the active queue to the inactive queue. 971 */ 972 973 page_shortage = (cnt.v_inactive_target + cnt.v_cache_min) - 974 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count); 975 page_shortage += addl_page_shortage; 976 977 /* 978 * Scan the active queue for things we can deactivate 979 */ 980 981 pcount = cnt.v_active_count; 982 m = TAILQ_FIRST(&vm_page_queue_active); 983 984 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) { 985 986 /* 987 * This is a consistancy check, and should likely be a panic 988 * or warning. 989 */ 990 if (m->queue != PQ_ACTIVE) { 991 break; 992 } 993 994 next = TAILQ_NEXT(m, pageq); 995 /* 996 * Don't deactivate pages that are busy. 997 */ 998 if ((m->busy != 0) || 999 (m->flags & PG_BUSY) || 1000 (m->hold_count != 0)) { 1001 s = splvm(); 1002 TAILQ_REMOVE(&vm_page_queue_active, m, pageq); 1003 TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq); 1004 splx(s); 1005 m = next; 1006 continue; 1007 } 1008 1009 /* 1010 * The count for pagedaemon pages is done after checking the 1011 * page for eligbility... 1012 */ 1013 cnt.v_pdpages++; 1014 1015 /* 1016 * Check to see "how much" the page has been used. 1017 */ 1018 actcount = 0; 1019 if (m->object->ref_count != 0) { 1020 if (m->flags & PG_REFERENCED) { 1021 actcount += 1; 1022 } 1023 actcount += pmap_ts_referenced(VM_PAGE_TO_PHYS(m)); 1024 if (actcount) { 1025 m->act_count += ACT_ADVANCE + actcount; 1026 if (m->act_count > ACT_MAX) 1027 m->act_count = ACT_MAX; 1028 } 1029 } 1030 1031 /* 1032 * Since we have "tested" this bit, we need to clear it now. 1033 */ 1034 vm_page_flag_clear(m, PG_REFERENCED); 1035 1036 /* 1037 * Only if an object is currently being used, do we use the 1038 * page activation count stats. 1039 */ 1040 if (actcount && (m->object->ref_count != 0)) { 1041 s = splvm(); 1042 TAILQ_REMOVE(&vm_page_queue_active, m, pageq); 1043 TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq); 1044 splx(s); 1045 } else { 1046 m->act_count -= min(m->act_count, ACT_DECLINE); 1047 if (vm_pageout_algorithm_lru || 1048 (m->object->ref_count == 0) || (m->act_count == 0)) { 1049 page_shortage--; 1050 if (m->object->ref_count == 0) { 1051 vm_page_protect(m, VM_PROT_NONE); 1052 if (m->dirty == 0) 1053 vm_page_cache(m); 1054 else 1055 vm_page_deactivate(m); 1056 } else { 1057 vm_page_deactivate(m); 1058 } 1059 } else { 1060 s = splvm(); 1061 TAILQ_REMOVE(&vm_page_queue_active, m, pageq); 1062 TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq); 1063 splx(s); 1064 } 1065 } 1066 m = next; 1067 } 1068 1069 s = splvm(); 1070 1071 /* 1072 * We try to maintain some *really* free pages, this allows interrupt 1073 * code to be guaranteed space. Since both cache and free queues 1074 * are considered basically 'free', moving pages from cache to free 1075 * does not effect other calculations. 1076 */ 1077 1078 while (cnt.v_free_count < cnt.v_free_reserved) { 1079 static int cache_rover = 0; 1080 m = vm_page_list_find(PQ_CACHE, cache_rover, FALSE); 1081 if (!m) 1082 break; 1083 if ((m->flags & PG_BUSY) || m->busy || m->hold_count || m->wire_count) { 1084 #ifdef INVARIANTS 1085 printf("Warning: busy page %p found in cache\n", m); 1086 #endif 1087 vm_page_deactivate(m); 1088 continue; 1089 } 1090 cache_rover = (cache_rover + PQ_PRIME2) & PQ_L2_MASK; 1091 vm_pageout_page_free(m); 1092 cnt.v_dfree++; 1093 } 1094 splx(s); 1095 1096 #if !defined(NO_SWAPPING) 1097 /* 1098 * Idle process swapout -- run once per second. 1099 */ 1100 if (vm_swap_idle_enabled) { 1101 static long lsec; 1102 if (time_second != lsec) { 1103 vm_pageout_req_swapout |= VM_SWAP_IDLE; 1104 vm_req_vmdaemon(); 1105 lsec = time_second; 1106 } 1107 } 1108 #endif 1109 1110 /* 1111 * If we didn't get enough free pages, and we have skipped a vnode 1112 * in a writeable object, wakeup the sync daemon. And kick swapout 1113 * if we did not get enough free pages. 1114 */ 1115 if ((cnt.v_cache_count + cnt.v_free_count) < 1116 (cnt.v_free_target + cnt.v_cache_min) ) { 1117 if (vnodes_skipped && 1118 (cnt.v_cache_count + cnt.v_free_count) < cnt.v_free_min) { 1119 (void) speedup_syncer(); 1120 } 1121 #if !defined(NO_SWAPPING) 1122 if (vm_swap_enabled && 1123 (cnt.v_free_count + cnt.v_cache_count < cnt.v_free_target)) { 1124 vm_req_vmdaemon(); 1125 vm_pageout_req_swapout |= VM_SWAP_NORMAL; 1126 } 1127 #endif 1128 } 1129 1130 /* 1131 * make sure that we have swap space -- if we are low on memory and 1132 * swap -- then kill the biggest process. 1133 */ 1134 if ((vm_swap_size == 0 || swap_pager_full) && 1135 ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_free_min)) { 1136 bigproc = NULL; 1137 bigsize = 0; 1138 for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) { 1139 /* 1140 * if this is a system process, skip it 1141 */ 1142 if ((p->p_flag & P_SYSTEM) || (p->p_lock > 0) || 1143 (p->p_pid == 1) || 1144 ((p->p_pid < 48) && (vm_swap_size != 0))) { 1145 continue; 1146 } 1147 /* 1148 * if the process is in a non-running type state, 1149 * don't touch it. 1150 */ 1151 if (p->p_stat != SRUN && p->p_stat != SSLEEP) { 1152 continue; 1153 } 1154 /* 1155 * get the process size 1156 */ 1157 size = vmspace_resident_count(p->p_vmspace); 1158 /* 1159 * if the this process is bigger than the biggest one 1160 * remember it. 1161 */ 1162 if (size > bigsize) { 1163 bigproc = p; 1164 bigsize = size; 1165 } 1166 } 1167 if (bigproc != NULL) { 1168 killproc(bigproc, "out of swap space"); 1169 bigproc->p_estcpu = 0; 1170 bigproc->p_nice = PRIO_MIN; 1171 resetpriority(bigproc); 1172 wakeup(&cnt.v_free_count); 1173 } 1174 } 1175 return force_wakeup; 1176 } 1177 1178 /* 1179 * This routine tries to maintain the pseudo LRU active queue, 1180 * so that during long periods of time where there is no paging, 1181 * that some statistic accumlation still occurs. This code 1182 * helps the situation where paging just starts to occur. 1183 */ 1184 static void 1185 vm_pageout_page_stats() 1186 { 1187 int s; 1188 vm_page_t m,next; 1189 int pcount,tpcount; /* Number of pages to check */ 1190 static int fullintervalcount = 0; 1191 int page_shortage; 1192 1193 page_shortage = (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) - 1194 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count); 1195 if (page_shortage <= 0) 1196 return; 1197 1198 pcount = cnt.v_active_count; 1199 fullintervalcount += vm_pageout_stats_interval; 1200 if (fullintervalcount < vm_pageout_full_stats_interval) { 1201 tpcount = (vm_pageout_stats_max * cnt.v_active_count) / cnt.v_page_count; 1202 if (pcount > tpcount) 1203 pcount = tpcount; 1204 } 1205 1206 m = TAILQ_FIRST(&vm_page_queue_active); 1207 while ((m != NULL) && (pcount-- > 0)) { 1208 int actcount; 1209 1210 if (m->queue != PQ_ACTIVE) { 1211 break; 1212 } 1213 1214 next = TAILQ_NEXT(m, pageq); 1215 /* 1216 * Don't deactivate pages that are busy. 1217 */ 1218 if ((m->busy != 0) || 1219 (m->flags & PG_BUSY) || 1220 (m->hold_count != 0)) { 1221 s = splvm(); 1222 TAILQ_REMOVE(&vm_page_queue_active, m, pageq); 1223 TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq); 1224 splx(s); 1225 m = next; 1226 continue; 1227 } 1228 1229 actcount = 0; 1230 if (m->flags & PG_REFERENCED) { 1231 vm_page_flag_clear(m, PG_REFERENCED); 1232 actcount += 1; 1233 } 1234 1235 actcount += pmap_ts_referenced(VM_PAGE_TO_PHYS(m)); 1236 if (actcount) { 1237 m->act_count += ACT_ADVANCE + actcount; 1238 if (m->act_count > ACT_MAX) 1239 m->act_count = ACT_MAX; 1240 s = splvm(); 1241 TAILQ_REMOVE(&vm_page_queue_active, m, pageq); 1242 TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq); 1243 splx(s); 1244 } else { 1245 if (m->act_count == 0) { 1246 /* 1247 * We turn off page access, so that we have more accurate 1248 * RSS stats. We don't do this in the normal page deactivation 1249 * when the system is loaded VM wise, because the cost of 1250 * the large number of page protect operations would be higher 1251 * than the value of doing the operation. 1252 */ 1253 vm_page_protect(m, VM_PROT_NONE); 1254 vm_page_deactivate(m); 1255 } else { 1256 m->act_count -= min(m->act_count, ACT_DECLINE); 1257 s = splvm(); 1258 TAILQ_REMOVE(&vm_page_queue_active, m, pageq); 1259 TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq); 1260 splx(s); 1261 } 1262 } 1263 1264 m = next; 1265 } 1266 } 1267 1268 static int 1269 vm_pageout_free_page_calc(count) 1270 vm_size_t count; 1271 { 1272 if (count < cnt.v_page_count) 1273 return 0; 1274 /* 1275 * free_reserved needs to include enough for the largest swap pager 1276 * structures plus enough for any pv_entry structs when paging. 1277 */ 1278 if (cnt.v_page_count > 1024) 1279 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200; 1280 else 1281 cnt.v_free_min = 4; 1282 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1283 cnt.v_interrupt_free_min; 1284 cnt.v_free_reserved = vm_pageout_page_count + 1285 cnt.v_pageout_free_min + (count / 768) + PQ_L2_SIZE; 1286 cnt.v_free_min += cnt.v_free_reserved; 1287 return 1; 1288 } 1289 1290 1291 /* 1292 * vm_pageout is the high level pageout daemon. 1293 */ 1294 static void 1295 vm_pageout() 1296 { 1297 /* 1298 * Initialize some paging parameters. 1299 */ 1300 1301 cnt.v_interrupt_free_min = 2; 1302 if (cnt.v_page_count < 2000) 1303 vm_pageout_page_count = 8; 1304 1305 vm_pageout_free_page_calc(cnt.v_page_count); 1306 /* 1307 * free_reserved needs to include enough for the largest swap pager 1308 * structures plus enough for any pv_entry structs when paging. 1309 */ 1310 if (cnt.v_free_count > 6144) 1311 cnt.v_free_target = 3 * cnt.v_free_min + cnt.v_free_reserved; 1312 else 1313 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved; 1314 1315 if (cnt.v_free_count > 2048) { 1316 cnt.v_cache_min = cnt.v_free_target; 1317 cnt.v_cache_max = 2 * cnt.v_cache_min; 1318 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2; 1319 } else { 1320 cnt.v_cache_min = 0; 1321 cnt.v_cache_max = 0; 1322 cnt.v_inactive_target = cnt.v_free_count / 4; 1323 } 1324 if (cnt.v_inactive_target > cnt.v_free_count / 3) 1325 cnt.v_inactive_target = cnt.v_free_count / 3; 1326 1327 /* XXX does not really belong here */ 1328 if (vm_page_max_wired == 0) 1329 vm_page_max_wired = cnt.v_free_count / 3; 1330 1331 if (vm_pageout_stats_max == 0) 1332 vm_pageout_stats_max = cnt.v_free_target; 1333 1334 /* 1335 * Set interval in seconds for stats scan. 1336 */ 1337 if (vm_pageout_stats_interval == 0) 1338 vm_pageout_stats_interval = 5; 1339 if (vm_pageout_full_stats_interval == 0) 1340 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4; 1341 1342 1343 /* 1344 * Set maximum free per pass 1345 */ 1346 if (vm_pageout_stats_free_max == 0) 1347 vm_pageout_stats_free_max = 5; 1348 1349 max_page_launder = (cnt.v_page_count > 1800 ? 32 : 16); 1350 1351 swap_pager_swap_init(); 1352 /* 1353 * The pageout daemon is never done, so loop forever. 1354 */ 1355 while (TRUE) { 1356 int error; 1357 int s = splvm(); 1358 if (!vm_pages_needed || 1359 ((cnt.v_free_count + cnt.v_cache_count) > cnt.v_free_min)) { 1360 vm_pages_needed = 0; 1361 error = tsleep(&vm_pages_needed, 1362 PVM, "psleep", vm_pageout_stats_interval * hz); 1363 if (error && !vm_pages_needed) { 1364 splx(s); 1365 vm_pageout_page_stats(); 1366 continue; 1367 } 1368 } else if (vm_pages_needed) { 1369 vm_pages_needed = 0; 1370 tsleep(&vm_pages_needed, PVM, "psleep", hz/2); 1371 } 1372 1373 if (vm_pages_needed) 1374 cnt.v_pdwakeups++; 1375 vm_pages_needed = 0; 1376 splx(s); 1377 vm_pageout_scan(); 1378 vm_pageout_deficit = 0; 1379 wakeup(&cnt.v_free_count); 1380 } 1381 } 1382 1383 void 1384 pagedaemon_wakeup() 1385 { 1386 if (!vm_pages_needed && curproc != pageproc) { 1387 vm_pages_needed++; 1388 wakeup(&vm_pages_needed); 1389 } 1390 } 1391 1392 #if !defined(NO_SWAPPING) 1393 static void 1394 vm_req_vmdaemon() 1395 { 1396 static int lastrun = 0; 1397 1398 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1399 wakeup(&vm_daemon_needed); 1400 lastrun = ticks; 1401 } 1402 } 1403 1404 static void 1405 vm_daemon() 1406 { 1407 struct proc *p; 1408 1409 while (TRUE) { 1410 tsleep(&vm_daemon_needed, PPAUSE, "psleep", 0); 1411 if (vm_pageout_req_swapout) { 1412 swapout_procs(vm_pageout_req_swapout); 1413 vm_pageout_req_swapout = 0; 1414 } 1415 /* 1416 * scan the processes for exceeding their rlimits or if 1417 * process is swapped out -- deactivate pages 1418 */ 1419 1420 for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) { 1421 vm_pindex_t limit, size; 1422 1423 /* 1424 * if this is a system process or if we have already 1425 * looked at this process, skip it. 1426 */ 1427 if (p->p_flag & (P_SYSTEM | P_WEXIT)) { 1428 continue; 1429 } 1430 /* 1431 * if the process is in a non-running type state, 1432 * don't touch it. 1433 */ 1434 if (p->p_stat != SRUN && p->p_stat != SSLEEP) { 1435 continue; 1436 } 1437 /* 1438 * get a limit 1439 */ 1440 limit = OFF_TO_IDX( 1441 qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur, 1442 p->p_rlimit[RLIMIT_RSS].rlim_max)); 1443 1444 /* 1445 * let processes that are swapped out really be 1446 * swapped out set the limit to nothing (will force a 1447 * swap-out.) 1448 */ 1449 if ((p->p_flag & P_INMEM) == 0) 1450 limit = 0; /* XXX */ 1451 1452 size = vmspace_resident_count(p->p_vmspace); 1453 if (limit >= 0 && size >= limit) { 1454 vm_pageout_map_deactivate_pages( 1455 &p->p_vmspace->vm_map, limit); 1456 } 1457 } 1458 } 1459 } 1460 #endif 1461