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/swap_pager.h> 99 #include <vm/vm_extern.h> 100 #include <vm/uma.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(void); 110 static int vm_pageout_clean(vm_page_t); 111 static void vm_pageout_scan(int pass); 112 static int vm_pageout_free_page_calc(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(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 static int pageout_lock_miss; 197 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss, 198 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout"); 199 200 #define VM_PAGEOUT_PAGE_COUNT 16 201 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT; 202 203 int vm_page_max_wired; /* XXX max # of wired pages system-wide */ 204 205 #if !defined(NO_SWAPPING) 206 typedef void freeer_fcn_t(vm_map_t, vm_object_t, vm_pindex_t, int); 207 static void vm_pageout_map_deactivate_pages(vm_map_t, vm_pindex_t); 208 static freeer_fcn_t vm_pageout_object_deactivate_pages; 209 static void vm_req_vmdaemon(void); 210 #endif 211 static void vm_pageout_page_stats(void); 212 213 /* 214 * vm_pageout_clean: 215 * 216 * Clean the page and remove it from the laundry. 217 * 218 * We set the busy bit to cause potential page faults on this page to 219 * block. Note the careful timing, however, the busy bit isn't set till 220 * late and we cannot do anything that will mess with the page. 221 */ 222 static int 223 vm_pageout_clean(m) 224 vm_page_t m; 225 { 226 vm_object_t object; 227 vm_page_t mc[2*vm_pageout_page_count]; 228 int pageout_count; 229 int ib, is, page_base; 230 vm_pindex_t pindex = m->pindex; 231 232 GIANT_REQUIRED; 233 234 object = m->object; 235 236 /* 237 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP 238 * with the new swapper, but we could have serious problems paging 239 * out other object types if there is insufficient memory. 240 * 241 * Unfortunately, checking free memory here is far too late, so the 242 * check has been moved up a procedural level. 243 */ 244 245 /* 246 * Don't mess with the page if it's busy, held, or special 247 */ 248 if ((m->hold_count != 0) || 249 ((m->busy != 0) || (m->flags & (PG_BUSY|PG_UNMANAGED)))) { 250 return 0; 251 } 252 253 mc[vm_pageout_page_count] = m; 254 pageout_count = 1; 255 page_base = vm_pageout_page_count; 256 ib = 1; 257 is = 1; 258 259 /* 260 * Scan object for clusterable pages. 261 * 262 * We can cluster ONLY if: ->> the page is NOT 263 * clean, wired, busy, held, or mapped into a 264 * buffer, and one of the following: 265 * 1) The page is inactive, or a seldom used 266 * active page. 267 * -or- 268 * 2) we force the issue. 269 * 270 * During heavy mmap/modification loads the pageout 271 * daemon can really fragment the underlying file 272 * due to flushing pages out of order and not trying 273 * align the clusters (which leave sporatic out-of-order 274 * holes). To solve this problem we do the reverse scan 275 * first and attempt to align our cluster, then do a 276 * forward scan if room remains. 277 */ 278 more: 279 while (ib && pageout_count < vm_pageout_page_count) { 280 vm_page_t p; 281 282 if (ib > pindex) { 283 ib = 0; 284 break; 285 } 286 287 if ((p = vm_page_lookup(object, pindex - ib)) == NULL) { 288 ib = 0; 289 break; 290 } 291 if (((p->queue - p->pc) == PQ_CACHE) || 292 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) { 293 ib = 0; 294 break; 295 } 296 vm_page_test_dirty(p); 297 if ((p->dirty & p->valid) == 0 || 298 p->queue != PQ_INACTIVE || 299 p->wire_count != 0 || /* may be held by buf cache */ 300 p->hold_count != 0) { /* may be undergoing I/O */ 301 ib = 0; 302 break; 303 } 304 mc[--page_base] = p; 305 ++pageout_count; 306 ++ib; 307 /* 308 * alignment boundry, stop here and switch directions. Do 309 * not clear ib. 310 */ 311 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0) 312 break; 313 } 314 315 while (pageout_count < vm_pageout_page_count && 316 pindex + is < object->size) { 317 vm_page_t p; 318 319 if ((p = vm_page_lookup(object, pindex + is)) == NULL) 320 break; 321 if (((p->queue - p->pc) == PQ_CACHE) || 322 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) { 323 break; 324 } 325 vm_page_test_dirty(p); 326 if ((p->dirty & p->valid) == 0 || 327 p->queue != PQ_INACTIVE || 328 p->wire_count != 0 || /* may be held by buf cache */ 329 p->hold_count != 0) { /* may be undergoing I/O */ 330 break; 331 } 332 mc[page_base + pageout_count] = p; 333 ++pageout_count; 334 ++is; 335 } 336 337 /* 338 * If we exhausted our forward scan, continue with the reverse scan 339 * when possible, even past a page boundry. This catches boundry 340 * conditions. 341 */ 342 if (ib && pageout_count < vm_pageout_page_count) 343 goto more; 344 345 /* 346 * we allow reads during pageouts... 347 */ 348 return vm_pageout_flush(&mc[page_base], pageout_count, 0); 349 } 350 351 /* 352 * vm_pageout_flush() - launder the given pages 353 * 354 * The given pages are laundered. Note that we setup for the start of 355 * I/O ( i.e. busy the page ), mark it read-only, and bump the object 356 * reference count all in here rather then in the parent. If we want 357 * the parent to do more sophisticated things we may have to change 358 * the ordering. 359 */ 360 int 361 vm_pageout_flush(mc, count, flags) 362 vm_page_t *mc; 363 int count; 364 int flags; 365 { 366 vm_object_t object; 367 int pageout_status[count]; 368 int numpagedout = 0; 369 int i; 370 371 GIANT_REQUIRED; 372 /* 373 * Initiate I/O. Bump the vm_page_t->busy counter and 374 * mark the pages read-only. 375 * 376 * We do not have to fixup the clean/dirty bits here... we can 377 * allow the pager to do it after the I/O completes. 378 * 379 * NOTE! mc[i]->dirty may be partial or fragmented due to an 380 * edge case with file fragments. 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_quick(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 int nothingwired; 548 549 GIANT_REQUIRED; 550 if (!vm_map_trylock(map)) 551 return; 552 553 bigobj = NULL; 554 nothingwired = TRUE; 555 556 /* 557 * first, search out the biggest object, and try to free pages from 558 * that. 559 */ 560 tmpe = map->header.next; 561 while (tmpe != &map->header) { 562 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 563 obj = tmpe->object.vm_object; 564 if ((obj != NULL) && (obj->shadow_count <= 1) && 565 ((bigobj == NULL) || 566 (bigobj->resident_page_count < obj->resident_page_count))) { 567 bigobj = obj; 568 } 569 } 570 if (tmpe->wired_count > 0) 571 nothingwired = FALSE; 572 tmpe = tmpe->next; 573 } 574 575 if (bigobj) 576 vm_pageout_object_deactivate_pages(map, bigobj, desired, 0); 577 578 /* 579 * Next, hunt around for other pages to deactivate. We actually 580 * do this search sort of wrong -- .text first is not the best idea. 581 */ 582 tmpe = map->header.next; 583 while (tmpe != &map->header) { 584 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 585 break; 586 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 587 obj = tmpe->object.vm_object; 588 if (obj) 589 vm_pageout_object_deactivate_pages(map, obj, desired, 0); 590 } 591 tmpe = tmpe->next; 592 }; 593 594 /* 595 * Remove all mappings if a process is swapped out, this will free page 596 * table pages. 597 */ 598 if (desired == 0 && nothingwired) 599 pmap_remove(vm_map_pmap(map), 600 VM_MIN_ADDRESS, VM_MAXUSER_ADDRESS); 601 vm_map_unlock(map); 602 return; 603 } 604 #endif /* !defined(NO_SWAPPING) */ 605 606 /* 607 * Don't try to be fancy - being fancy can lead to VOP_LOCK's and therefore 608 * to vnode deadlocks. We only do it for OBJT_DEFAULT and OBJT_SWAP objects 609 * which we know can be trivially freed. 610 */ 611 void 612 vm_pageout_page_free(vm_page_t m) { 613 vm_object_t object = m->object; 614 int type = object->type; 615 616 GIANT_REQUIRED; 617 if (type == OBJT_SWAP || type == OBJT_DEFAULT) 618 vm_object_reference(object); 619 vm_page_busy(m); 620 vm_page_protect(m, VM_PROT_NONE); 621 vm_page_free(m); 622 if (type == OBJT_SWAP || type == OBJT_DEFAULT) 623 vm_object_deallocate(object); 624 } 625 626 /* 627 * vm_pageout_scan does the dirty work for the pageout daemon. 628 */ 629 static void 630 vm_pageout_scan(int pass) 631 { 632 vm_page_t m, next; 633 struct vm_page marker; 634 int save_page_shortage; 635 int save_inactive_count; 636 int page_shortage, maxscan, pcount; 637 int addl_page_shortage, addl_page_shortage_init; 638 struct proc *p, *bigproc; 639 vm_offset_t size, bigsize; 640 vm_object_t object; 641 int actcount; 642 int vnodes_skipped = 0; 643 int maxlaunder; 644 int s; 645 struct thread *td; 646 647 GIANT_REQUIRED; 648 /* 649 * Do whatever cleanup that the pmap code can. 650 */ 651 pmap_collect(); 652 uma_reclaim(); 653 654 addl_page_shortage_init = vm_pageout_deficit; 655 vm_pageout_deficit = 0; 656 657 /* 658 * Calculate the number of pages we want to either free or move 659 * to the cache. 660 */ 661 page_shortage = vm_paging_target() + addl_page_shortage_init; 662 save_page_shortage = page_shortage; 663 save_inactive_count = cnt.v_inactive_count; 664 665 /* 666 * Initialize our marker 667 */ 668 bzero(&marker, sizeof(marker)); 669 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER; 670 marker.queue = PQ_INACTIVE; 671 marker.wire_count = 1; 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. Note that m->act_count 677 * is not used to form decisions for the inactive queue, only for the 678 * active queue. 679 * 680 * maxlaunder limits the number of dirty pages we flush per scan. 681 * For most systems a smaller value (16 or 32) is more robust under 682 * extreme memory and disk pressure because any unnecessary writes 683 * to disk can result in extreme performance degredation. However, 684 * systems with excessive dirty pages (especially when MAP_NOSYNC is 685 * used) will die horribly with limited laundering. If the pageout 686 * daemon cannot clean enough pages in the first pass, we let it go 687 * all out in succeeding passes. 688 */ 689 if ((maxlaunder = vm_max_launder) <= 1) 690 maxlaunder = 1; 691 if (pass) 692 maxlaunder = 10000; 693 rescan0: 694 addl_page_shortage = addl_page_shortage_init; 695 maxscan = cnt.v_inactive_count; 696 697 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl); 698 m != NULL && maxscan-- > 0 && page_shortage > 0; 699 m = next) { 700 701 cnt.v_pdpages++; 702 703 if (m->queue != PQ_INACTIVE) { 704 goto rescan0; 705 } 706 707 next = TAILQ_NEXT(m, pageq); 708 709 /* 710 * skip marker pages 711 */ 712 if (m->flags & PG_MARKER) 713 continue; 714 715 /* 716 * A held page may be undergoing I/O, so skip it. 717 */ 718 if (m->hold_count) { 719 vm_pageq_requeue(m); 720 addl_page_shortage++; 721 continue; 722 } 723 /* 724 * Don't mess with busy pages, keep in the front of the 725 * queue, most likely are being paged out. 726 */ 727 if (m->busy || (m->flags & PG_BUSY)) { 728 addl_page_shortage++; 729 continue; 730 } 731 732 /* 733 * If the object is not being used, we ignore previous 734 * references. 735 */ 736 if (m->object->ref_count == 0) { 737 vm_page_flag_clear(m, PG_REFERENCED); 738 pmap_clear_reference(m); 739 740 /* 741 * Otherwise, if the page has been referenced while in the 742 * inactive queue, we bump the "activation count" upwards, 743 * making it less likely that the page will be added back to 744 * the inactive queue prematurely again. Here we check the 745 * page tables (or emulated bits, if any), given the upper 746 * level VM system not knowing anything about existing 747 * references. 748 */ 749 } else if (((m->flags & PG_REFERENCED) == 0) && 750 (actcount = pmap_ts_referenced(m))) { 751 vm_page_activate(m); 752 m->act_count += (actcount + ACT_ADVANCE); 753 continue; 754 } 755 756 /* 757 * If the upper level VM system knows about any page 758 * references, we activate the page. We also set the 759 * "activation count" higher than normal so that we will less 760 * likely place pages back onto the inactive queue again. 761 */ 762 if ((m->flags & PG_REFERENCED) != 0) { 763 vm_page_flag_clear(m, PG_REFERENCED); 764 actcount = pmap_ts_referenced(m); 765 vm_page_activate(m); 766 m->act_count += (actcount + ACT_ADVANCE + 1); 767 continue; 768 } 769 770 /* 771 * If the upper level VM system doesn't know anything about 772 * the page being dirty, we have to check for it again. As 773 * far as the VM code knows, any partially dirty pages are 774 * fully dirty. 775 */ 776 if (m->dirty == 0) { 777 vm_page_test_dirty(m); 778 } else { 779 vm_page_dirty(m); 780 } 781 782 /* 783 * Invalid pages can be easily freed 784 */ 785 if (m->valid == 0) { 786 vm_pageout_page_free(m); 787 cnt.v_dfree++; 788 --page_shortage; 789 790 /* 791 * Clean pages can be placed onto the cache queue. This 792 * effectively frees them. 793 */ 794 } else if (m->dirty == 0) { 795 vm_page_cache(m); 796 --page_shortage; 797 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) { 798 /* 799 * Dirty pages need to be paged out, but flushing 800 * a page is extremely expensive verses freeing 801 * a clean page. Rather then artificially limiting 802 * the number of pages we can flush, we instead give 803 * dirty pages extra priority on the inactive queue 804 * by forcing them to be cycled through the queue 805 * twice before being flushed, after which the 806 * (now clean) page will cycle through once more 807 * before being freed. This significantly extends 808 * the thrash point for a heavily loaded machine. 809 */ 810 vm_page_flag_set(m, PG_WINATCFLS); 811 vm_pageq_requeue(m); 812 } else if (maxlaunder > 0) { 813 /* 814 * We always want to try to flush some dirty pages if 815 * we encounter them, to keep the system stable. 816 * Normally this number is small, but under extreme 817 * pressure where there are insufficient clean pages 818 * on the inactive queue, we may have to go all out. 819 */ 820 int swap_pageouts_ok; 821 struct vnode *vp = NULL; 822 struct mount *mp; 823 824 object = m->object; 825 826 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 827 swap_pageouts_ok = 1; 828 } else { 829 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 830 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 831 vm_page_count_min()); 832 833 } 834 835 /* 836 * We don't bother paging objects that are "dead". 837 * Those objects are in a "rundown" state. 838 */ 839 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 840 vm_pageq_requeue(m); 841 continue; 842 } 843 844 /* 845 * The object is already known NOT to be dead. It 846 * is possible for the vget() to block the whole 847 * pageout daemon, but the new low-memory handling 848 * code should prevent it. 849 * 850 * The previous code skipped locked vnodes and, worse, 851 * reordered pages in the queue. This results in 852 * completely non-deterministic operation and, on a 853 * busy system, can lead to extremely non-optimal 854 * pageouts. For example, it can cause clean pages 855 * to be freed and dirty pages to be moved to the end 856 * of the queue. Since dirty pages are also moved to 857 * the end of the queue once-cleaned, this gives 858 * way too large a weighting to defering the freeing 859 * of dirty pages. 860 * 861 * We can't wait forever for the vnode lock, we might 862 * deadlock due to a vn_read() getting stuck in 863 * vm_wait while holding this vnode. We skip the 864 * vnode if we can't get it in a reasonable amount 865 * of time. 866 */ 867 if (object->type == OBJT_VNODE) { 868 vp = object->handle; 869 870 mp = NULL; 871 if (vp->v_type == VREG) 872 vn_start_write(vp, &mp, V_NOWAIT); 873 if (vget(vp, LK_EXCLUSIVE|LK_NOOBJ|LK_TIMELOCK, curthread)) { 874 ++pageout_lock_miss; 875 vn_finished_write(mp); 876 if (object->flags & OBJ_MIGHTBEDIRTY) 877 vnodes_skipped++; 878 continue; 879 } 880 881 /* 882 * The page might have been moved to another 883 * queue during potential blocking in vget() 884 * above. The page might have been freed and 885 * reused for another vnode. The object might 886 * have been reused for another vnode. 887 */ 888 if (m->queue != PQ_INACTIVE || 889 m->object != object || 890 object->handle != vp) { 891 if (object->flags & OBJ_MIGHTBEDIRTY) 892 vnodes_skipped++; 893 vput(vp); 894 vn_finished_write(mp); 895 continue; 896 } 897 898 /* 899 * The page may have been busied during the 900 * blocking in vput(); We don't move the 901 * page back onto the end of the queue so that 902 * statistics are more correct if we don't. 903 */ 904 if (m->busy || (m->flags & PG_BUSY)) { 905 vput(vp); 906 vn_finished_write(mp); 907 continue; 908 } 909 910 /* 911 * If the page has become held it might 912 * be undergoing I/O, so skip it 913 */ 914 if (m->hold_count) { 915 vm_pageq_requeue(m); 916 if (object->flags & OBJ_MIGHTBEDIRTY) 917 vnodes_skipped++; 918 vput(vp); 919 vn_finished_write(mp); 920 continue; 921 } 922 } 923 924 /* 925 * If a page is dirty, then it is either being washed 926 * (but not yet cleaned) or it is still in the 927 * laundry. If it is still in the laundry, then we 928 * start the cleaning operation. 929 * 930 * This operation may cluster, invalidating the 'next' 931 * pointer. To prevent an inordinate number of 932 * restarts we use our marker to remember our place. 933 * 934 * decrement page_shortage on success to account for 935 * the (future) cleaned page. Otherwise we could wind 936 * up laundering or cleaning too many pages. 937 */ 938 s = splvm(); 939 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, m, &marker, pageq); 940 splx(s); 941 if (vm_pageout_clean(m) != 0) { 942 --page_shortage; 943 --maxlaunder; 944 } 945 s = splvm(); 946 next = TAILQ_NEXT(&marker, pageq); 947 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, &marker, pageq); 948 splx(s); 949 if (vp) { 950 vput(vp); 951 vn_finished_write(mp); 952 } 953 } 954 } 955 956 /* 957 * Compute the number of pages we want to try to move from the 958 * active queue to the inactive queue. 959 */ 960 page_shortage = vm_paging_target() + 961 cnt.v_inactive_target - cnt.v_inactive_count; 962 page_shortage += addl_page_shortage; 963 964 /* 965 * Scan the active queue for things we can deactivate. We nominally 966 * track the per-page activity counter and use it to locate 967 * deactivation candidates. 968 */ 969 pcount = cnt.v_active_count; 970 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 971 972 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) { 973 974 /* 975 * This is a consistency check, and should likely be a panic 976 * or warning. 977 */ 978 if (m->queue != PQ_ACTIVE) { 979 break; 980 } 981 982 next = TAILQ_NEXT(m, pageq); 983 /* 984 * Don't deactivate pages that are busy. 985 */ 986 if ((m->busy != 0) || 987 (m->flags & PG_BUSY) || 988 (m->hold_count != 0)) { 989 vm_pageq_requeue(m); 990 m = next; 991 continue; 992 } 993 994 /* 995 * The count for pagedaemon pages is done after checking the 996 * page for eligibility... 997 */ 998 cnt.v_pdpages++; 999 1000 /* 1001 * Check to see "how much" the page has been used. 1002 */ 1003 actcount = 0; 1004 if (m->object->ref_count != 0) { 1005 if (m->flags & PG_REFERENCED) { 1006 actcount += 1; 1007 } 1008 actcount += pmap_ts_referenced(m); 1009 if (actcount) { 1010 m->act_count += ACT_ADVANCE + actcount; 1011 if (m->act_count > ACT_MAX) 1012 m->act_count = ACT_MAX; 1013 } 1014 } 1015 1016 /* 1017 * Since we have "tested" this bit, we need to clear it now. 1018 */ 1019 vm_page_flag_clear(m, PG_REFERENCED); 1020 1021 /* 1022 * Only if an object is currently being used, do we use the 1023 * page activation count stats. 1024 */ 1025 if (actcount && (m->object->ref_count != 0)) { 1026 vm_pageq_requeue(m); 1027 } else { 1028 m->act_count -= min(m->act_count, ACT_DECLINE); 1029 if (vm_pageout_algorithm || 1030 m->object->ref_count == 0 || 1031 m->act_count == 0) { 1032 page_shortage--; 1033 if (m->object->ref_count == 0) { 1034 vm_page_protect(m, VM_PROT_NONE); 1035 if (m->dirty == 0) 1036 vm_page_cache(m); 1037 else 1038 vm_page_deactivate(m); 1039 } else { 1040 vm_page_deactivate(m); 1041 } 1042 } else { 1043 vm_pageq_requeue(m); 1044 } 1045 } 1046 m = next; 1047 } 1048 1049 s = splvm(); 1050 1051 /* 1052 * We try to maintain some *really* free pages, this allows interrupt 1053 * code to be guaranteed space. Since both cache and free queues 1054 * are considered basically 'free', moving pages from cache to free 1055 * does not effect other calculations. 1056 */ 1057 while (cnt.v_free_count < cnt.v_free_reserved) { 1058 static int cache_rover = 0; 1059 m = vm_pageq_find(PQ_CACHE, cache_rover, FALSE); 1060 if (!m) 1061 break; 1062 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || 1063 m->busy || 1064 m->hold_count || 1065 m->wire_count) { 1066 #ifdef INVARIANTS 1067 printf("Warning: busy page %p found in cache\n", m); 1068 #endif 1069 vm_page_deactivate(m); 1070 continue; 1071 } 1072 cache_rover = (cache_rover + PQ_PRIME2) & PQ_L2_MASK; 1073 vm_pageout_page_free(m); 1074 cnt.v_dfree++; 1075 } 1076 splx(s); 1077 1078 #if !defined(NO_SWAPPING) 1079 /* 1080 * Idle process swapout -- run once per second. 1081 */ 1082 if (vm_swap_idle_enabled) { 1083 static long lsec; 1084 if (time_second != lsec) { 1085 vm_pageout_req_swapout |= VM_SWAP_IDLE; 1086 vm_req_vmdaemon(); 1087 lsec = time_second; 1088 } 1089 } 1090 #endif 1091 1092 /* 1093 * If we didn't get enough free pages, and we have skipped a vnode 1094 * in a writeable object, wakeup the sync daemon. And kick swapout 1095 * if we did not get enough free pages. 1096 */ 1097 if (vm_paging_target() > 0) { 1098 if (vnodes_skipped && vm_page_count_min()) 1099 (void) speedup_syncer(); 1100 #if !defined(NO_SWAPPING) 1101 if (vm_swap_enabled && vm_page_count_target()) { 1102 vm_req_vmdaemon(); 1103 vm_pageout_req_swapout |= VM_SWAP_NORMAL; 1104 } 1105 #endif 1106 } 1107 1108 /* 1109 * If we are out of swap and were not able to reach our paging 1110 * target, kill the largest process. 1111 * 1112 * We keep the process bigproc locked once we find it to keep anyone 1113 * from messing with it; however, there is a possibility of 1114 * deadlock if process B is bigproc and one of it's child processes 1115 * attempts to propagate a signal to B while we are waiting for A's 1116 * lock while walking this list. To avoid this, we don't block on 1117 * the process lock but just skip a process if it is already locked. 1118 */ 1119 if ((vm_swap_size < 64 && vm_page_count_min()) || 1120 (swap_pager_full && vm_paging_target() > 0)) { 1121 #if 0 1122 if ((vm_swap_size < 64 || swap_pager_full) && vm_page_count_min()) { 1123 #endif 1124 bigproc = NULL; 1125 bigsize = 0; 1126 sx_slock(&allproc_lock); 1127 FOREACH_PROC_IN_SYSTEM(p) { 1128 int breakout; 1129 /* 1130 * If this process is already locked, skip it. 1131 */ 1132 if (PROC_TRYLOCK(p) == 0) 1133 continue; 1134 /* 1135 * if this is a system process, skip it 1136 */ 1137 if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) || 1138 ((p->p_pid < 48) && (vm_swap_size != 0))) { 1139 PROC_UNLOCK(p); 1140 continue; 1141 } 1142 /* 1143 * if the process is in a non-running type state, 1144 * don't touch it. Check all the threads individually. 1145 */ 1146 mtx_lock_spin(&sched_lock); 1147 breakout = 0; 1148 FOREACH_THREAD_IN_PROC(p, td) { 1149 if (td->td_state != TDS_RUNQ && 1150 td->td_state != TDS_RUNNING && 1151 td->td_state != TDS_SLP) { 1152 breakout = 1; 1153 break; 1154 } 1155 } 1156 if (breakout) { 1157 mtx_unlock_spin(&sched_lock); 1158 PROC_UNLOCK(p); 1159 continue; 1160 } 1161 mtx_unlock_spin(&sched_lock); 1162 /* 1163 * get the process size 1164 */ 1165 size = vmspace_resident_count(p->p_vmspace) + 1166 vmspace_swap_count(p->p_vmspace); 1167 /* 1168 * if the this process is bigger than the biggest one 1169 * remember it. 1170 */ 1171 if (size > bigsize) { 1172 if (bigproc != NULL) 1173 PROC_UNLOCK(bigproc); 1174 bigproc = p; 1175 bigsize = size; 1176 } else 1177 PROC_UNLOCK(p); 1178 } 1179 sx_sunlock(&allproc_lock); 1180 if (bigproc != NULL) { 1181 struct ksegrp *kg; 1182 killproc(bigproc, "out of swap space"); 1183 mtx_lock_spin(&sched_lock); 1184 FOREACH_KSEGRP_IN_PROC(bigproc, kg) { 1185 kg->kg_estcpu = 0; 1186 kg->kg_nice = PRIO_MIN; /* XXXKSE ??? */ 1187 resetpriority(kg); 1188 } 1189 mtx_unlock_spin(&sched_lock); 1190 PROC_UNLOCK(bigproc); 1191 wakeup(&cnt.v_free_count); 1192 } 1193 } 1194 } 1195 1196 /* 1197 * This routine tries to maintain the pseudo LRU active queue, 1198 * so that during long periods of time where there is no paging, 1199 * that some statistic accumulation still occurs. This code 1200 * helps the situation where paging just starts to occur. 1201 */ 1202 static void 1203 vm_pageout_page_stats() 1204 { 1205 vm_page_t m,next; 1206 int pcount,tpcount; /* Number of pages to check */ 1207 static int fullintervalcount = 0; 1208 int page_shortage; 1209 int s0; 1210 1211 page_shortage = 1212 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) - 1213 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count); 1214 1215 if (page_shortage <= 0) 1216 return; 1217 1218 s0 = splvm(); 1219 1220 pcount = cnt.v_active_count; 1221 fullintervalcount += vm_pageout_stats_interval; 1222 if (fullintervalcount < vm_pageout_full_stats_interval) { 1223 tpcount = (vm_pageout_stats_max * cnt.v_active_count) / cnt.v_page_count; 1224 if (pcount > tpcount) 1225 pcount = tpcount; 1226 } else { 1227 fullintervalcount = 0; 1228 } 1229 1230 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1231 while ((m != NULL) && (pcount-- > 0)) { 1232 int actcount; 1233 1234 if (m->queue != PQ_ACTIVE) { 1235 break; 1236 } 1237 1238 next = TAILQ_NEXT(m, pageq); 1239 /* 1240 * Don't deactivate pages that are busy. 1241 */ 1242 if ((m->busy != 0) || 1243 (m->flags & PG_BUSY) || 1244 (m->hold_count != 0)) { 1245 vm_pageq_requeue(m); 1246 m = next; 1247 continue; 1248 } 1249 1250 actcount = 0; 1251 if (m->flags & PG_REFERENCED) { 1252 vm_page_flag_clear(m, PG_REFERENCED); 1253 actcount += 1; 1254 } 1255 1256 actcount += pmap_ts_referenced(m); 1257 if (actcount) { 1258 m->act_count += ACT_ADVANCE + actcount; 1259 if (m->act_count > ACT_MAX) 1260 m->act_count = ACT_MAX; 1261 vm_pageq_requeue(m); 1262 } else { 1263 if (m->act_count == 0) { 1264 /* 1265 * We turn off page access, so that we have 1266 * more accurate RSS stats. We don't do this 1267 * in the normal page deactivation when the 1268 * system is loaded VM wise, because the 1269 * cost of the large number of page protect 1270 * operations would be higher than the value 1271 * of doing the operation. 1272 */ 1273 vm_page_protect(m, VM_PROT_NONE); 1274 vm_page_deactivate(m); 1275 } else { 1276 m->act_count -= min(m->act_count, ACT_DECLINE); 1277 vm_pageq_requeue(m); 1278 } 1279 } 1280 1281 m = next; 1282 } 1283 splx(s0); 1284 } 1285 1286 static int 1287 vm_pageout_free_page_calc(count) 1288 vm_size_t count; 1289 { 1290 if (count < cnt.v_page_count) 1291 return 0; 1292 /* 1293 * free_reserved needs to include enough for the largest swap pager 1294 * structures plus enough for any pv_entry structs when paging. 1295 */ 1296 if (cnt.v_page_count > 1024) 1297 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200; 1298 else 1299 cnt.v_free_min = 4; 1300 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1301 cnt.v_interrupt_free_min; 1302 cnt.v_free_reserved = vm_pageout_page_count + 1303 cnt.v_pageout_free_min + (count / 768) + PQ_L2_SIZE; 1304 cnt.v_free_severe = cnt.v_free_min / 2; 1305 cnt.v_free_min += cnt.v_free_reserved; 1306 cnt.v_free_severe += cnt.v_free_reserved; 1307 return 1; 1308 } 1309 1310 /* 1311 * vm_pageout is the high level pageout daemon. 1312 */ 1313 static void 1314 vm_pageout() 1315 { 1316 int pass; 1317 1318 mtx_lock(&Giant); 1319 1320 /* 1321 * Initialize some paging parameters. 1322 */ 1323 cnt.v_interrupt_free_min = 2; 1324 if (cnt.v_page_count < 2000) 1325 vm_pageout_page_count = 8; 1326 1327 vm_pageout_free_page_calc(cnt.v_page_count); 1328 /* 1329 * v_free_target and v_cache_min control pageout hysteresis. Note 1330 * that these are more a measure of the VM cache queue hysteresis 1331 * then the VM free queue. Specifically, v_free_target is the 1332 * high water mark (free+cache pages). 1333 * 1334 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the 1335 * low water mark, while v_free_min is the stop. v_cache_min must 1336 * be big enough to handle memory needs while the pageout daemon 1337 * is signalled and run to free more pages. 1338 */ 1339 if (cnt.v_free_count > 6144) 1340 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved; 1341 else 1342 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved; 1343 1344 if (cnt.v_free_count > 2048) { 1345 cnt.v_cache_min = cnt.v_free_target; 1346 cnt.v_cache_max = 2 * cnt.v_cache_min; 1347 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2; 1348 } else { 1349 cnt.v_cache_min = 0; 1350 cnt.v_cache_max = 0; 1351 cnt.v_inactive_target = cnt.v_free_count / 4; 1352 } 1353 if (cnt.v_inactive_target > cnt.v_free_count / 3) 1354 cnt.v_inactive_target = cnt.v_free_count / 3; 1355 1356 /* XXX does not really belong here */ 1357 if (vm_page_max_wired == 0) 1358 vm_page_max_wired = cnt.v_free_count / 3; 1359 1360 if (vm_pageout_stats_max == 0) 1361 vm_pageout_stats_max = cnt.v_free_target; 1362 1363 /* 1364 * Set interval in seconds for stats scan. 1365 */ 1366 if (vm_pageout_stats_interval == 0) 1367 vm_pageout_stats_interval = 5; 1368 if (vm_pageout_full_stats_interval == 0) 1369 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4; 1370 1371 /* 1372 * Set maximum free per pass 1373 */ 1374 if (vm_pageout_stats_free_max == 0) 1375 vm_pageout_stats_free_max = 5; 1376 1377 swap_pager_swap_init(); 1378 pass = 0; 1379 /* 1380 * The pageout daemon is never done, so loop forever. 1381 */ 1382 while (TRUE) { 1383 int error; 1384 int s = splvm(); 1385 1386 /* 1387 * If we have enough free memory, wakeup waiters. Do 1388 * not clear vm_pages_needed until we reach our target, 1389 * otherwise we may be woken up over and over again and 1390 * waste a lot of cpu. 1391 */ 1392 if (vm_pages_needed && !vm_page_count_min()) { 1393 if (vm_paging_needed() <= 0) 1394 vm_pages_needed = 0; 1395 wakeup(&cnt.v_free_count); 1396 } 1397 if (vm_pages_needed) { 1398 /* 1399 * Still not done, take a second pass without waiting 1400 * (unlimited dirty cleaning), otherwise sleep a bit 1401 * and try again. 1402 */ 1403 ++pass; 1404 if (pass > 1) 1405 tsleep(&vm_pages_needed, PVM, 1406 "psleep", hz/2); 1407 } else { 1408 /* 1409 * Good enough, sleep & handle stats. Prime the pass 1410 * for the next run. 1411 */ 1412 if (pass > 1) 1413 pass = 1; 1414 else 1415 pass = 0; 1416 error = tsleep(&vm_pages_needed, PVM, 1417 "psleep", vm_pageout_stats_interval * hz); 1418 if (error && !vm_pages_needed) { 1419 splx(s); 1420 pass = 0; 1421 vm_pageout_page_stats(); 1422 continue; 1423 } 1424 } 1425 1426 if (vm_pages_needed) 1427 cnt.v_pdwakeups++; 1428 splx(s); 1429 vm_pageout_scan(pass); 1430 vm_pageout_deficit = 0; 1431 } 1432 } 1433 1434 void 1435 pagedaemon_wakeup() 1436 { 1437 if (!vm_pages_needed && curthread->td_proc != pageproc) { 1438 vm_pages_needed++; 1439 wakeup(&vm_pages_needed); 1440 } 1441 } 1442 1443 #if !defined(NO_SWAPPING) 1444 static void 1445 vm_req_vmdaemon() 1446 { 1447 static int lastrun = 0; 1448 1449 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1450 wakeup(&vm_daemon_needed); 1451 lastrun = ticks; 1452 } 1453 } 1454 1455 static void 1456 vm_daemon() 1457 { 1458 struct proc *p; 1459 int breakout; 1460 struct thread *td; 1461 1462 mtx_lock(&Giant); 1463 while (TRUE) { 1464 tsleep(&vm_daemon_needed, PPAUSE, "psleep", 0); 1465 if (vm_pageout_req_swapout) { 1466 swapout_procs(vm_pageout_req_swapout); 1467 vm_pageout_req_swapout = 0; 1468 } 1469 /* 1470 * scan the processes for exceeding their rlimits or if 1471 * process is swapped out -- deactivate pages 1472 */ 1473 sx_slock(&allproc_lock); 1474 LIST_FOREACH(p, &allproc, p_list) { 1475 vm_pindex_t limit, size; 1476 1477 /* 1478 * if this is a system process or if we have already 1479 * looked at this process, skip it. 1480 */ 1481 if (p->p_flag & (P_SYSTEM | P_WEXIT)) { 1482 continue; 1483 } 1484 /* 1485 * if the process is in a non-running type state, 1486 * don't touch it. 1487 */ 1488 mtx_lock_spin(&sched_lock); 1489 breakout = 0; 1490 FOREACH_THREAD_IN_PROC(p, td) { 1491 if (td->td_state != TDS_RUNQ && 1492 td->td_state != TDS_RUNNING && 1493 td->td_state != TDS_SLP) { 1494 breakout = 1; 1495 break; 1496 } 1497 } 1498 if (breakout) { 1499 mtx_unlock_spin(&sched_lock); 1500 continue; 1501 } 1502 /* 1503 * get a limit 1504 */ 1505 limit = OFF_TO_IDX( 1506 qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur, 1507 p->p_rlimit[RLIMIT_RSS].rlim_max)); 1508 1509 /* 1510 * let processes that are swapped out really be 1511 * swapped out set the limit to nothing (will force a 1512 * swap-out.) 1513 */ 1514 if ((p->p_sflag & PS_INMEM) == 0) 1515 limit = 0; /* XXX */ 1516 mtx_unlock_spin(&sched_lock); 1517 1518 size = vmspace_resident_count(p->p_vmspace); 1519 if (limit >= 0 && size >= limit) { 1520 vm_pageout_map_deactivate_pages( 1521 &p->p_vmspace->vm_map, limit); 1522 } 1523 } 1524 sx_sunlock(&allproc_lock); 1525 } 1526 } 1527 #endif /* !defined(NO_SWAPPING) */ 1528