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 * Copyright (c) 2005 Yahoo! Technologies Norway AS 9 * All rights reserved. 10 * 11 * This code is derived from software contributed to Berkeley by 12 * The Mach Operating System project at Carnegie-Mellon University. 13 * 14 * Redistribution and use in source and binary forms, with or without 15 * modification, are permitted provided that the following conditions 16 * are met: 17 * 1. Redistributions of source code must retain the above copyright 18 * notice, this list of conditions and the following disclaimer. 19 * 2. Redistributions in binary form must reproduce the above copyright 20 * notice, this list of conditions and the following disclaimer in the 21 * documentation and/or other materials provided with the distribution. 22 * 3. All advertising materials mentioning features or use of this software 23 * must display the following acknowledgement: 24 * This product includes software developed by the University of 25 * California, Berkeley and its contributors. 26 * 4. Neither the name of the University nor the names of its contributors 27 * may be used to endorse or promote products derived from this software 28 * without specific prior written permission. 29 * 30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 40 * SUCH DAMAGE. 41 * 42 * from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91 43 * 44 * 45 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 46 * All rights reserved. 47 * 48 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 49 * 50 * Permission to use, copy, modify and distribute this software and 51 * its documentation is hereby granted, provided that both the copyright 52 * notice and this permission notice appear in all copies of the 53 * software, derivative works or modified versions, and any portions 54 * thereof, and that both notices appear in supporting documentation. 55 * 56 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 57 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 58 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 59 * 60 * Carnegie Mellon requests users of this software to return to 61 * 62 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 63 * School of Computer Science 64 * Carnegie Mellon University 65 * Pittsburgh PA 15213-3890 66 * 67 * any improvements or extensions that they make and grant Carnegie the 68 * rights to redistribute these changes. 69 */ 70 71 /* 72 * The proverbial page-out daemon. 73 */ 74 75 #include <sys/cdefs.h> 76 __FBSDID("$FreeBSD$"); 77 78 #include "opt_vm.h" 79 #include <sys/param.h> 80 #include <sys/systm.h> 81 #include <sys/kernel.h> 82 #include <sys/eventhandler.h> 83 #include <sys/lock.h> 84 #include <sys/mutex.h> 85 #include <sys/proc.h> 86 #include <sys/kthread.h> 87 #include <sys/ktr.h> 88 #include <sys/mount.h> 89 #include <sys/racct.h> 90 #include <sys/resourcevar.h> 91 #include <sys/sched.h> 92 #include <sys/signalvar.h> 93 #include <sys/vnode.h> 94 #include <sys/vmmeter.h> 95 #include <sys/sx.h> 96 #include <sys/sysctl.h> 97 98 #include <vm/vm.h> 99 #include <vm/vm_param.h> 100 #include <vm/vm_object.h> 101 #include <vm/vm_page.h> 102 #include <vm/vm_map.h> 103 #include <vm/vm_pageout.h> 104 #include <vm/vm_pager.h> 105 #include <vm/swap_pager.h> 106 #include <vm/vm_extern.h> 107 #include <vm/uma.h> 108 109 /* 110 * System initialization 111 */ 112 113 /* the kernel process "vm_pageout"*/ 114 static void vm_pageout(void); 115 static int vm_pageout_clean(vm_page_t); 116 static void vm_pageout_scan(int pass); 117 118 struct proc *pageproc; 119 120 static struct kproc_desc page_kp = { 121 "pagedaemon", 122 vm_pageout, 123 &pageproc 124 }; 125 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, 126 &page_kp); 127 128 #if !defined(NO_SWAPPING) 129 /* the kernel process "vm_daemon"*/ 130 static void vm_daemon(void); 131 static struct proc *vmproc; 132 133 static struct kproc_desc vm_kp = { 134 "vmdaemon", 135 vm_daemon, 136 &vmproc 137 }; 138 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp); 139 #endif 140 141 142 int vm_pages_needed; /* Event on which pageout daemon sleeps */ 143 int vm_pageout_deficit; /* Estimated number of pages deficit */ 144 int vm_pageout_pages_needed; /* flag saying that the pageout daemon needs pages */ 145 146 #if !defined(NO_SWAPPING) 147 static int vm_pageout_req_swapout; /* XXX */ 148 static int vm_daemon_needed; 149 static struct mtx vm_daemon_mtx; 150 /* Allow for use by vm_pageout before vm_daemon is initialized. */ 151 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF); 152 #endif 153 static int vm_max_launder = 32; 154 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0; 155 static int vm_pageout_full_stats_interval = 0; 156 static int vm_pageout_algorithm=0; 157 static int defer_swap_pageouts=0; 158 static int disable_swap_pageouts=0; 159 160 #if defined(NO_SWAPPING) 161 static int vm_swap_enabled=0; 162 static int vm_swap_idle_enabled=0; 163 #else 164 static int vm_swap_enabled=1; 165 static int vm_swap_idle_enabled=0; 166 #endif 167 168 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm, 169 CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt"); 170 171 SYSCTL_INT(_vm, OID_AUTO, max_launder, 172 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout"); 173 174 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max, 175 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length"); 176 177 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval, 178 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan"); 179 180 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval, 181 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan"); 182 183 #if defined(NO_SWAPPING) 184 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 185 CTLFLAG_RD, &vm_swap_enabled, 0, "Enable entire process swapout"); 186 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 187 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 188 #else 189 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 190 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout"); 191 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 192 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 193 #endif 194 195 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts, 196 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem"); 197 198 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts, 199 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages"); 200 201 static int pageout_lock_miss; 202 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss, 203 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout"); 204 205 #define VM_PAGEOUT_PAGE_COUNT 16 206 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT; 207 208 int vm_page_max_wired; /* XXX max # of wired pages system-wide */ 209 SYSCTL_INT(_vm, OID_AUTO, max_wired, 210 CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count"); 211 212 static boolean_t vm_pageout_fallback_object_lock(vm_page_t, vm_page_t *); 213 static boolean_t vm_pageout_launder(int, int, vm_paddr_t, vm_paddr_t); 214 #if !defined(NO_SWAPPING) 215 static void vm_pageout_map_deactivate_pages(vm_map_t, long); 216 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long); 217 static void vm_req_vmdaemon(int req); 218 #endif 219 static boolean_t vm_pageout_page_lock(vm_page_t, vm_page_t *); 220 static void vm_pageout_page_stats(void); 221 static void vm_pageout_requeue(vm_page_t m); 222 223 /* 224 * Initialize a dummy page for marking the caller's place in the specified 225 * paging queue. In principle, this function only needs to set the flag 226 * PG_MARKER. Nonetheless, it sets the flag VPO_BUSY and initializes the hold 227 * count to one as safety precautions. 228 */ 229 static void 230 vm_pageout_init_marker(vm_page_t marker, u_short queue) 231 { 232 233 bzero(marker, sizeof(*marker)); 234 marker->flags = PG_MARKER; 235 marker->oflags = VPO_BUSY; 236 marker->queue = queue; 237 marker->hold_count = 1; 238 } 239 240 /* 241 * vm_pageout_fallback_object_lock: 242 * 243 * Lock vm object currently associated with `m'. VM_OBJECT_TRYLOCK is 244 * known to have failed and page queue must be either PQ_ACTIVE or 245 * PQ_INACTIVE. To avoid lock order violation, unlock the page queues 246 * while locking the vm object. Use marker page to detect page queue 247 * changes and maintain notion of next page on page queue. Return 248 * TRUE if no changes were detected, FALSE otherwise. vm object is 249 * locked on return. 250 * 251 * This function depends on both the lock portion of struct vm_object 252 * and normal struct vm_page being type stable. 253 */ 254 static boolean_t 255 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next) 256 { 257 struct vm_page marker; 258 boolean_t unchanged; 259 u_short queue; 260 vm_object_t object; 261 262 queue = m->queue; 263 vm_pageout_init_marker(&marker, queue); 264 object = m->object; 265 266 TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl, 267 m, &marker, pageq); 268 vm_page_unlock_queues(); 269 vm_page_unlock(m); 270 VM_OBJECT_LOCK(object); 271 vm_page_lock(m); 272 vm_page_lock_queues(); 273 274 /* Page queue might have changed. */ 275 *next = TAILQ_NEXT(&marker, pageq); 276 unchanged = (m->queue == queue && 277 m->object == object && 278 &marker == TAILQ_NEXT(m, pageq)); 279 TAILQ_REMOVE(&vm_page_queues[queue].pl, 280 &marker, pageq); 281 return (unchanged); 282 } 283 284 /* 285 * Lock the page while holding the page queue lock. Use marker page 286 * to detect page queue changes and maintain notion of next page on 287 * page queue. Return TRUE if no changes were detected, FALSE 288 * otherwise. The page is locked on return. The page queue lock might 289 * be dropped and reacquired. 290 * 291 * This function depends on normal struct vm_page being type stable. 292 */ 293 static boolean_t 294 vm_pageout_page_lock(vm_page_t m, vm_page_t *next) 295 { 296 struct vm_page marker; 297 boolean_t unchanged; 298 u_short queue; 299 300 vm_page_lock_assert(m, MA_NOTOWNED); 301 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 302 303 if (vm_page_trylock(m)) 304 return (TRUE); 305 306 queue = m->queue; 307 vm_pageout_init_marker(&marker, queue); 308 309 TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl, m, &marker, pageq); 310 vm_page_unlock_queues(); 311 vm_page_lock(m); 312 vm_page_lock_queues(); 313 314 /* Page queue might have changed. */ 315 *next = TAILQ_NEXT(&marker, pageq); 316 unchanged = (m->queue == queue && &marker == TAILQ_NEXT(m, pageq)); 317 TAILQ_REMOVE(&vm_page_queues[queue].pl, &marker, pageq); 318 return (unchanged); 319 } 320 321 /* 322 * vm_pageout_clean: 323 * 324 * Clean the page and remove it from the laundry. 325 * 326 * We set the busy bit to cause potential page faults on this page to 327 * block. Note the careful timing, however, the busy bit isn't set till 328 * late and we cannot do anything that will mess with the page. 329 */ 330 static int 331 vm_pageout_clean(vm_page_t m) 332 { 333 vm_object_t object; 334 vm_page_t mc[2*vm_pageout_page_count], pb, ps; 335 int pageout_count; 336 int ib, is, page_base; 337 vm_pindex_t pindex = m->pindex; 338 339 vm_page_lock_assert(m, MA_OWNED); 340 object = m->object; 341 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 342 343 /* 344 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP 345 * with the new swapper, but we could have serious problems paging 346 * out other object types if there is insufficient memory. 347 * 348 * Unfortunately, checking free memory here is far too late, so the 349 * check has been moved up a procedural level. 350 */ 351 352 /* 353 * Can't clean the page if it's busy or held. 354 */ 355 KASSERT(m->busy == 0 && (m->oflags & VPO_BUSY) == 0, 356 ("vm_pageout_clean: page %p is busy", m)); 357 KASSERT(m->hold_count == 0, ("vm_pageout_clean: page %p is held", m)); 358 vm_page_unlock(m); 359 360 mc[vm_pageout_page_count] = pb = ps = m; 361 pageout_count = 1; 362 page_base = vm_pageout_page_count; 363 ib = 1; 364 is = 1; 365 366 /* 367 * Scan object for clusterable pages. 368 * 369 * We can cluster ONLY if: ->> the page is NOT 370 * clean, wired, busy, held, or mapped into a 371 * buffer, and one of the following: 372 * 1) The page is inactive, or a seldom used 373 * active page. 374 * -or- 375 * 2) we force the issue. 376 * 377 * During heavy mmap/modification loads the pageout 378 * daemon can really fragment the underlying file 379 * due to flushing pages out of order and not trying 380 * align the clusters (which leave sporatic out-of-order 381 * holes). To solve this problem we do the reverse scan 382 * first and attempt to align our cluster, then do a 383 * forward scan if room remains. 384 */ 385 more: 386 while (ib && pageout_count < vm_pageout_page_count) { 387 vm_page_t p; 388 389 if (ib > pindex) { 390 ib = 0; 391 break; 392 } 393 394 if ((p = vm_page_prev(pb)) == NULL || 395 (p->oflags & VPO_BUSY) != 0 || p->busy != 0) { 396 ib = 0; 397 break; 398 } 399 vm_page_lock(p); 400 vm_page_test_dirty(p); 401 if (p->dirty == 0 || 402 p->queue != PQ_INACTIVE || 403 p->hold_count != 0) { /* may be undergoing I/O */ 404 vm_page_unlock(p); 405 ib = 0; 406 break; 407 } 408 vm_page_unlock(p); 409 mc[--page_base] = pb = p; 410 ++pageout_count; 411 ++ib; 412 /* 413 * alignment boundry, stop here and switch directions. Do 414 * not clear ib. 415 */ 416 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0) 417 break; 418 } 419 420 while (pageout_count < vm_pageout_page_count && 421 pindex + is < object->size) { 422 vm_page_t p; 423 424 if ((p = vm_page_next(ps)) == NULL || 425 (p->oflags & VPO_BUSY) != 0 || p->busy != 0) 426 break; 427 vm_page_lock(p); 428 vm_page_test_dirty(p); 429 if (p->dirty == 0 || 430 p->queue != PQ_INACTIVE || 431 p->hold_count != 0) { /* may be undergoing I/O */ 432 vm_page_unlock(p); 433 break; 434 } 435 vm_page_unlock(p); 436 mc[page_base + pageout_count] = ps = p; 437 ++pageout_count; 438 ++is; 439 } 440 441 /* 442 * If we exhausted our forward scan, continue with the reverse scan 443 * when possible, even past a page boundry. This catches boundry 444 * conditions. 445 */ 446 if (ib && pageout_count < vm_pageout_page_count) 447 goto more; 448 449 /* 450 * we allow reads during pageouts... 451 */ 452 return (vm_pageout_flush(&mc[page_base], pageout_count, 0, 0, NULL, 453 NULL)); 454 } 455 456 /* 457 * vm_pageout_flush() - launder the given pages 458 * 459 * The given pages are laundered. Note that we setup for the start of 460 * I/O ( i.e. busy the page ), mark it read-only, and bump the object 461 * reference count all in here rather then in the parent. If we want 462 * the parent to do more sophisticated things we may have to change 463 * the ordering. 464 * 465 * Returned runlen is the count of pages between mreq and first 466 * page after mreq with status VM_PAGER_AGAIN. 467 * *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL 468 * for any page in runlen set. 469 */ 470 int 471 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen, 472 boolean_t *eio) 473 { 474 vm_object_t object = mc[0]->object; 475 int pageout_status[count]; 476 int numpagedout = 0; 477 int i, runlen; 478 479 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 480 481 /* 482 * Initiate I/O. Bump the vm_page_t->busy counter and 483 * mark the pages read-only. 484 * 485 * We do not have to fixup the clean/dirty bits here... we can 486 * allow the pager to do it after the I/O completes. 487 * 488 * NOTE! mc[i]->dirty may be partial or fragmented due to an 489 * edge case with file fragments. 490 */ 491 for (i = 0; i < count; i++) { 492 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, 493 ("vm_pageout_flush: partially invalid page %p index %d/%d", 494 mc[i], i, count)); 495 vm_page_io_start(mc[i]); 496 pmap_remove_write(mc[i]); 497 } 498 vm_object_pip_add(object, count); 499 500 vm_pager_put_pages(object, mc, count, flags, pageout_status); 501 502 runlen = count - mreq; 503 if (eio != NULL) 504 *eio = FALSE; 505 for (i = 0; i < count; i++) { 506 vm_page_t mt = mc[i]; 507 508 KASSERT(pageout_status[i] == VM_PAGER_PEND || 509 !pmap_page_is_write_mapped(mt), 510 ("vm_pageout_flush: page %p is not write protected", mt)); 511 switch (pageout_status[i]) { 512 case VM_PAGER_OK: 513 case VM_PAGER_PEND: 514 numpagedout++; 515 break; 516 case VM_PAGER_BAD: 517 /* 518 * Page outside of range of object. Right now we 519 * essentially lose the changes by pretending it 520 * worked. 521 */ 522 vm_page_undirty(mt); 523 break; 524 case VM_PAGER_ERROR: 525 case VM_PAGER_FAIL: 526 /* 527 * If page couldn't be paged out, then reactivate the 528 * page so it doesn't clog the inactive list. (We 529 * will try paging out it again later). 530 */ 531 vm_page_lock(mt); 532 vm_page_activate(mt); 533 vm_page_unlock(mt); 534 if (eio != NULL && i >= mreq && i - mreq < runlen) 535 *eio = TRUE; 536 break; 537 case VM_PAGER_AGAIN: 538 if (i >= mreq && i - mreq < runlen) 539 runlen = i - mreq; 540 break; 541 } 542 543 /* 544 * If the operation is still going, leave the page busy to 545 * block all other accesses. Also, leave the paging in 546 * progress indicator set so that we don't attempt an object 547 * collapse. 548 */ 549 if (pageout_status[i] != VM_PAGER_PEND) { 550 vm_object_pip_wakeup(object); 551 vm_page_io_finish(mt); 552 if (vm_page_count_severe()) { 553 vm_page_lock(mt); 554 vm_page_try_to_cache(mt); 555 vm_page_unlock(mt); 556 } 557 } 558 } 559 if (prunlen != NULL) 560 *prunlen = runlen; 561 return (numpagedout); 562 } 563 564 static boolean_t 565 vm_pageout_launder(int queue, int tries, vm_paddr_t low, vm_paddr_t high) 566 { 567 struct mount *mp; 568 struct vnode *vp; 569 vm_object_t object; 570 vm_paddr_t pa; 571 vm_page_t m, m_tmp, next; 572 573 vm_page_lock_queues(); 574 TAILQ_FOREACH_SAFE(m, &vm_page_queues[queue].pl, pageq, next) { 575 KASSERT(m->queue == queue, 576 ("vm_pageout_launder: page %p's queue is not %d", m, 577 queue)); 578 if ((m->flags & PG_MARKER) != 0) 579 continue; 580 pa = VM_PAGE_TO_PHYS(m); 581 if (pa < low || pa + PAGE_SIZE > high) 582 continue; 583 if (!vm_pageout_page_lock(m, &next) || m->hold_count != 0) { 584 vm_page_unlock(m); 585 continue; 586 } 587 object = m->object; 588 if ((!VM_OBJECT_TRYLOCK(object) && 589 (!vm_pageout_fallback_object_lock(m, &next) || 590 m->hold_count != 0)) || (m->oflags & VPO_BUSY) != 0 || 591 m->busy != 0) { 592 vm_page_unlock(m); 593 VM_OBJECT_UNLOCK(object); 594 continue; 595 } 596 vm_page_test_dirty(m); 597 if (m->dirty == 0) 598 pmap_remove_all(m); 599 if (m->dirty != 0) { 600 vm_page_unlock(m); 601 if (tries == 0 || (object->flags & OBJ_DEAD) != 0) { 602 VM_OBJECT_UNLOCK(object); 603 continue; 604 } 605 if (object->type == OBJT_VNODE) { 606 vm_page_unlock_queues(); 607 vp = object->handle; 608 vm_object_reference_locked(object); 609 VM_OBJECT_UNLOCK(object); 610 (void)vn_start_write(vp, &mp, V_WAIT); 611 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 612 VM_OBJECT_LOCK(object); 613 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 614 VM_OBJECT_UNLOCK(object); 615 VOP_UNLOCK(vp, 0); 616 vm_object_deallocate(object); 617 vn_finished_write(mp); 618 return (TRUE); 619 } else if (object->type == OBJT_SWAP || 620 object->type == OBJT_DEFAULT) { 621 vm_page_unlock_queues(); 622 m_tmp = m; 623 vm_pageout_flush(&m_tmp, 1, VM_PAGER_PUT_SYNC, 624 0, NULL, NULL); 625 VM_OBJECT_UNLOCK(object); 626 return (TRUE); 627 } 628 } else { 629 vm_page_cache(m); 630 vm_page_unlock(m); 631 } 632 VM_OBJECT_UNLOCK(object); 633 } 634 vm_page_unlock_queues(); 635 return (FALSE); 636 } 637 638 /* 639 * Increase the number of cached pages. The specified value, "tries", 640 * determines which categories of pages are cached: 641 * 642 * 0: All clean, inactive pages within the specified physical address range 643 * are cached. Will not sleep. 644 * 1: The vm_lowmem handlers are called. All inactive pages within 645 * the specified physical address range are cached. May sleep. 646 * 2: The vm_lowmem handlers are called. All inactive and active pages 647 * within the specified physical address range are cached. May sleep. 648 */ 649 void 650 vm_pageout_grow_cache(int tries, vm_paddr_t low, vm_paddr_t high) 651 { 652 int actl, actmax, inactl, inactmax; 653 654 if (tries > 0) { 655 /* 656 * Decrease registered cache sizes. The vm_lowmem handlers 657 * may acquire locks and/or sleep, so they can only be invoked 658 * when "tries" is greater than zero. 659 */ 660 EVENTHANDLER_INVOKE(vm_lowmem, 0); 661 662 /* 663 * We do this explicitly after the caches have been drained 664 * above. 665 */ 666 uma_reclaim(); 667 } 668 inactl = 0; 669 inactmax = cnt.v_inactive_count; 670 actl = 0; 671 actmax = tries < 2 ? 0 : cnt.v_active_count; 672 again: 673 if (inactl < inactmax && vm_pageout_launder(PQ_INACTIVE, tries, low, 674 high)) { 675 inactl++; 676 goto again; 677 } 678 if (actl < actmax && vm_pageout_launder(PQ_ACTIVE, tries, low, high)) { 679 actl++; 680 goto again; 681 } 682 } 683 684 #if !defined(NO_SWAPPING) 685 /* 686 * vm_pageout_object_deactivate_pages 687 * 688 * Deactivate enough pages to satisfy the inactive target 689 * requirements. 690 * 691 * The object and map must be locked. 692 */ 693 static void 694 vm_pageout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object, 695 long desired) 696 { 697 vm_object_t backing_object, object; 698 vm_page_t p; 699 int actcount, remove_mode; 700 701 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED); 702 if (first_object->type == OBJT_DEVICE || 703 first_object->type == OBJT_SG) 704 return; 705 for (object = first_object;; object = backing_object) { 706 if (pmap_resident_count(pmap) <= desired) 707 goto unlock_return; 708 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 709 if (object->type == OBJT_PHYS || object->paging_in_progress) 710 goto unlock_return; 711 712 remove_mode = 0; 713 if (object->shadow_count > 1) 714 remove_mode = 1; 715 /* 716 * Scan the object's entire memory queue. 717 */ 718 TAILQ_FOREACH(p, &object->memq, listq) { 719 if (pmap_resident_count(pmap) <= desired) 720 goto unlock_return; 721 if ((p->oflags & VPO_BUSY) != 0 || p->busy != 0) 722 continue; 723 PCPU_INC(cnt.v_pdpages); 724 vm_page_lock(p); 725 if (p->wire_count != 0 || p->hold_count != 0 || 726 !pmap_page_exists_quick(pmap, p)) { 727 vm_page_unlock(p); 728 continue; 729 } 730 actcount = pmap_ts_referenced(p); 731 if ((p->aflags & PGA_REFERENCED) != 0) { 732 if (actcount == 0) 733 actcount = 1; 734 vm_page_aflag_clear(p, PGA_REFERENCED); 735 } 736 if (p->queue != PQ_ACTIVE && actcount != 0) { 737 vm_page_activate(p); 738 p->act_count += actcount; 739 } else if (p->queue == PQ_ACTIVE) { 740 if (actcount == 0) { 741 p->act_count -= min(p->act_count, 742 ACT_DECLINE); 743 if (!remove_mode && 744 (vm_pageout_algorithm || 745 p->act_count == 0)) { 746 pmap_remove_all(p); 747 vm_page_deactivate(p); 748 } else { 749 vm_page_lock_queues(); 750 vm_pageout_requeue(p); 751 vm_page_unlock_queues(); 752 } 753 } else { 754 vm_page_activate(p); 755 if (p->act_count < ACT_MAX - 756 ACT_ADVANCE) 757 p->act_count += ACT_ADVANCE; 758 vm_page_lock_queues(); 759 vm_pageout_requeue(p); 760 vm_page_unlock_queues(); 761 } 762 } else if (p->queue == PQ_INACTIVE) 763 pmap_remove_all(p); 764 vm_page_unlock(p); 765 } 766 if ((backing_object = object->backing_object) == NULL) 767 goto unlock_return; 768 VM_OBJECT_LOCK(backing_object); 769 if (object != first_object) 770 VM_OBJECT_UNLOCK(object); 771 } 772 unlock_return: 773 if (object != first_object) 774 VM_OBJECT_UNLOCK(object); 775 } 776 777 /* 778 * deactivate some number of pages in a map, try to do it fairly, but 779 * that is really hard to do. 780 */ 781 static void 782 vm_pageout_map_deactivate_pages(map, desired) 783 vm_map_t map; 784 long desired; 785 { 786 vm_map_entry_t tmpe; 787 vm_object_t obj, bigobj; 788 int nothingwired; 789 790 if (!vm_map_trylock(map)) 791 return; 792 793 bigobj = NULL; 794 nothingwired = TRUE; 795 796 /* 797 * first, search out the biggest object, and try to free pages from 798 * that. 799 */ 800 tmpe = map->header.next; 801 while (tmpe != &map->header) { 802 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 803 obj = tmpe->object.vm_object; 804 if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) { 805 if (obj->shadow_count <= 1 && 806 (bigobj == NULL || 807 bigobj->resident_page_count < obj->resident_page_count)) { 808 if (bigobj != NULL) 809 VM_OBJECT_UNLOCK(bigobj); 810 bigobj = obj; 811 } else 812 VM_OBJECT_UNLOCK(obj); 813 } 814 } 815 if (tmpe->wired_count > 0) 816 nothingwired = FALSE; 817 tmpe = tmpe->next; 818 } 819 820 if (bigobj != NULL) { 821 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired); 822 VM_OBJECT_UNLOCK(bigobj); 823 } 824 /* 825 * Next, hunt around for other pages to deactivate. We actually 826 * do this search sort of wrong -- .text first is not the best idea. 827 */ 828 tmpe = map->header.next; 829 while (tmpe != &map->header) { 830 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 831 break; 832 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 833 obj = tmpe->object.vm_object; 834 if (obj != NULL) { 835 VM_OBJECT_LOCK(obj); 836 vm_pageout_object_deactivate_pages(map->pmap, obj, desired); 837 VM_OBJECT_UNLOCK(obj); 838 } 839 } 840 tmpe = tmpe->next; 841 } 842 843 /* 844 * Remove all mappings if a process is swapped out, this will free page 845 * table pages. 846 */ 847 if (desired == 0 && nothingwired) { 848 pmap_remove(vm_map_pmap(map), vm_map_min(map), 849 vm_map_max(map)); 850 } 851 vm_map_unlock(map); 852 } 853 #endif /* !defined(NO_SWAPPING) */ 854 855 /* 856 * vm_pageout_requeue: 857 * 858 * Move the specified page to the tail of its present page queue. 859 * 860 * The page queues must be locked. 861 */ 862 static void 863 vm_pageout_requeue(vm_page_t m) 864 { 865 struct vpgqueues *vpq; 866 867 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 868 KASSERT(m->queue != PQ_NONE, 869 ("vm_pageout_requeue: page %p is not queued", m)); 870 vpq = &vm_page_queues[m->queue]; 871 TAILQ_REMOVE(&vpq->pl, m, pageq); 872 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq); 873 } 874 875 /* 876 * vm_pageout_scan does the dirty work for the pageout daemon. 877 */ 878 static void 879 vm_pageout_scan(int pass) 880 { 881 vm_page_t m, next; 882 struct vm_page marker; 883 int page_shortage, maxscan, pcount; 884 int addl_page_shortage; 885 vm_object_t object; 886 int actcount; 887 int vnodes_skipped = 0; 888 int maxlaunder; 889 boolean_t queues_locked; 890 891 /* 892 * Decrease registered cache sizes. 893 */ 894 EVENTHANDLER_INVOKE(vm_lowmem, 0); 895 /* 896 * We do this explicitly after the caches have been drained above. 897 */ 898 uma_reclaim(); 899 900 /* 901 * The addl_page_shortage is the number of temporarily 902 * stuck pages in the inactive queue. In other words, the 903 * number of pages from cnt.v_inactive_count that should be 904 * discounted in setting the target for the active queue scan. 905 */ 906 addl_page_shortage = atomic_readandclear_int(&vm_pageout_deficit); 907 908 /* 909 * Calculate the number of pages we want to either free or move 910 * to the cache. 911 */ 912 page_shortage = vm_paging_target() + addl_page_shortage; 913 914 vm_pageout_init_marker(&marker, PQ_INACTIVE); 915 916 /* 917 * Start scanning the inactive queue for pages we can move to the 918 * cache or free. The scan will stop when the target is reached or 919 * we have scanned the entire inactive queue. Note that m->act_count 920 * is not used to form decisions for the inactive queue, only for the 921 * active queue. 922 * 923 * maxlaunder limits the number of dirty pages we flush per scan. 924 * For most systems a smaller value (16 or 32) is more robust under 925 * extreme memory and disk pressure because any unnecessary writes 926 * to disk can result in extreme performance degredation. However, 927 * systems with excessive dirty pages (especially when MAP_NOSYNC is 928 * used) will die horribly with limited laundering. If the pageout 929 * daemon cannot clean enough pages in the first pass, we let it go 930 * all out in succeeding passes. 931 */ 932 if ((maxlaunder = vm_max_launder) <= 1) 933 maxlaunder = 1; 934 if (pass) 935 maxlaunder = 10000; 936 vm_page_lock_queues(); 937 queues_locked = TRUE; 938 maxscan = cnt.v_inactive_count; 939 940 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl); 941 m != NULL && maxscan-- > 0 && page_shortage > 0; 942 m = next) { 943 KASSERT(queues_locked, ("unlocked queues")); 944 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 945 KASSERT(m->queue == PQ_INACTIVE, ("Inactive queue %p", m)); 946 947 cnt.v_pdpages++; 948 next = TAILQ_NEXT(m, pageq); 949 950 /* 951 * skip marker pages 952 */ 953 if (m->flags & PG_MARKER) 954 continue; 955 956 KASSERT((m->flags & PG_FICTITIOUS) == 0, 957 ("Fictitious page %p cannot be in inactive queue", m)); 958 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 959 ("Unmanaged page %p cannot be in inactive queue", m)); 960 961 /* 962 * The page or object lock acquisitions fail if the 963 * page was removed from the queue or moved to a 964 * different position within the queue. In either 965 * case, addl_page_shortage should not be incremented. 966 */ 967 if (!vm_pageout_page_lock(m, &next)) { 968 vm_page_unlock(m); 969 continue; 970 } 971 object = m->object; 972 if (!VM_OBJECT_TRYLOCK(object) && 973 !vm_pageout_fallback_object_lock(m, &next)) { 974 vm_page_unlock(m); 975 VM_OBJECT_UNLOCK(object); 976 continue; 977 } 978 979 /* 980 * Don't mess with busy pages, keep them at at the 981 * front of the queue, most likely they are being 982 * paged out. Increment addl_page_shortage for busy 983 * pages, because they may leave the inactive queue 984 * shortly after page scan is finished. 985 */ 986 if (m->busy != 0 || (m->oflags & VPO_BUSY) != 0) { 987 vm_page_unlock(m); 988 VM_OBJECT_UNLOCK(object); 989 addl_page_shortage++; 990 continue; 991 } 992 993 /* 994 * We unlock vm_page_queue_mtx, invalidating the 995 * 'next' pointer. Use our marker to remember our 996 * place. 997 */ 998 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, 999 m, &marker, pageq); 1000 vm_page_unlock_queues(); 1001 queues_locked = FALSE; 1002 1003 /* 1004 * If the object is not being used, we ignore previous 1005 * references. 1006 */ 1007 if (object->ref_count == 0) { 1008 vm_page_aflag_clear(m, PGA_REFERENCED); 1009 KASSERT(!pmap_page_is_mapped(m), 1010 ("vm_pageout_scan: page %p is mapped", m)); 1011 1012 /* 1013 * Otherwise, if the page has been referenced while in the 1014 * inactive queue, we bump the "activation count" upwards, 1015 * making it less likely that the page will be added back to 1016 * the inactive queue prematurely again. Here we check the 1017 * page tables (or emulated bits, if any), given the upper 1018 * level VM system not knowing anything about existing 1019 * references. 1020 */ 1021 } else if ((m->aflags & PGA_REFERENCED) == 0 && 1022 (actcount = pmap_ts_referenced(m)) != 0) { 1023 vm_page_activate(m); 1024 vm_page_unlock(m); 1025 m->act_count += actcount + ACT_ADVANCE; 1026 VM_OBJECT_UNLOCK(object); 1027 goto relock_queues; 1028 } 1029 1030 /* 1031 * If the upper level VM system knows about any page 1032 * references, we activate the page. We also set the 1033 * "activation count" higher than normal so that we will less 1034 * likely place pages back onto the inactive queue again. 1035 */ 1036 if ((m->aflags & PGA_REFERENCED) != 0) { 1037 vm_page_aflag_clear(m, PGA_REFERENCED); 1038 actcount = pmap_ts_referenced(m); 1039 vm_page_activate(m); 1040 vm_page_unlock(m); 1041 m->act_count += actcount + ACT_ADVANCE + 1; 1042 VM_OBJECT_UNLOCK(object); 1043 goto relock_queues; 1044 } 1045 1046 if (m->hold_count != 0) { 1047 vm_page_unlock(m); 1048 VM_OBJECT_UNLOCK(object); 1049 1050 /* 1051 * Held pages are essentially stuck in the 1052 * queue. So, they ought to be discounted 1053 * from cnt.v_inactive_count. See the 1054 * calculation of the page_shortage for the 1055 * loop over the active queue below. 1056 */ 1057 addl_page_shortage++; 1058 goto relock_queues; 1059 } 1060 1061 /* 1062 * If the upper level VM system does not believe that the page 1063 * is fully dirty, but it is mapped for write access, then we 1064 * consult the pmap to see if the page's dirty status should 1065 * be updated. 1066 */ 1067 if (m->dirty != VM_PAGE_BITS_ALL && 1068 pmap_page_is_write_mapped(m)) { 1069 /* 1070 * Avoid a race condition: Unless write access is 1071 * removed from the page, another processor could 1072 * modify it before all access is removed by the call 1073 * to vm_page_cache() below. If vm_page_cache() finds 1074 * that the page has been modified when it removes all 1075 * access, it panics because it cannot cache dirty 1076 * pages. In principle, we could eliminate just write 1077 * access here rather than all access. In the expected 1078 * case, when there are no last instant modifications 1079 * to the page, removing all access will be cheaper 1080 * overall. 1081 */ 1082 if (pmap_is_modified(m)) 1083 vm_page_dirty(m); 1084 else if (m->dirty == 0) 1085 pmap_remove_all(m); 1086 } 1087 1088 if (m->valid == 0) { 1089 /* 1090 * Invalid pages can be easily freed 1091 */ 1092 vm_page_free(m); 1093 PCPU_INC(cnt.v_dfree); 1094 --page_shortage; 1095 } else if (m->dirty == 0) { 1096 /* 1097 * Clean pages can be placed onto the cache queue. 1098 * This effectively frees them. 1099 */ 1100 vm_page_cache(m); 1101 --page_shortage; 1102 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) { 1103 /* 1104 * Dirty pages need to be paged out, but flushing 1105 * a page is extremely expensive verses freeing 1106 * a clean page. Rather then artificially limiting 1107 * the number of pages we can flush, we instead give 1108 * dirty pages extra priority on the inactive queue 1109 * by forcing them to be cycled through the queue 1110 * twice before being flushed, after which the 1111 * (now clean) page will cycle through once more 1112 * before being freed. This significantly extends 1113 * the thrash point for a heavily loaded machine. 1114 */ 1115 m->flags |= PG_WINATCFLS; 1116 vm_page_lock_queues(); 1117 queues_locked = TRUE; 1118 vm_pageout_requeue(m); 1119 } else if (maxlaunder > 0) { 1120 /* 1121 * We always want to try to flush some dirty pages if 1122 * we encounter them, to keep the system stable. 1123 * Normally this number is small, but under extreme 1124 * pressure where there are insufficient clean pages 1125 * on the inactive queue, we may have to go all out. 1126 */ 1127 int swap_pageouts_ok; 1128 struct vnode *vp = NULL; 1129 struct mount *mp = NULL; 1130 1131 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 1132 swap_pageouts_ok = 1; 1133 } else { 1134 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 1135 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 1136 vm_page_count_min()); 1137 1138 } 1139 1140 /* 1141 * We don't bother paging objects that are "dead". 1142 * Those objects are in a "rundown" state. 1143 */ 1144 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 1145 vm_page_lock_queues(); 1146 vm_page_unlock(m); 1147 VM_OBJECT_UNLOCK(object); 1148 queues_locked = TRUE; 1149 vm_pageout_requeue(m); 1150 goto relock_queues; 1151 } 1152 1153 /* 1154 * The object is already known NOT to be dead. It 1155 * is possible for the vget() to block the whole 1156 * pageout daemon, but the new low-memory handling 1157 * code should prevent it. 1158 * 1159 * The previous code skipped locked vnodes and, worse, 1160 * reordered pages in the queue. This results in 1161 * completely non-deterministic operation and, on a 1162 * busy system, can lead to extremely non-optimal 1163 * pageouts. For example, it can cause clean pages 1164 * to be freed and dirty pages to be moved to the end 1165 * of the queue. Since dirty pages are also moved to 1166 * the end of the queue once-cleaned, this gives 1167 * way too large a weighting to defering the freeing 1168 * of dirty pages. 1169 * 1170 * We can't wait forever for the vnode lock, we might 1171 * deadlock due to a vn_read() getting stuck in 1172 * vm_wait while holding this vnode. We skip the 1173 * vnode if we can't get it in a reasonable amount 1174 * of time. 1175 */ 1176 if (object->type == OBJT_VNODE) { 1177 vm_page_unlock(m); 1178 vp = object->handle; 1179 if (vp->v_type == VREG && 1180 vn_start_write(vp, &mp, V_NOWAIT) != 0) { 1181 mp = NULL; 1182 ++pageout_lock_miss; 1183 if (object->flags & OBJ_MIGHTBEDIRTY) 1184 vnodes_skipped++; 1185 goto unlock_and_continue; 1186 } 1187 KASSERT(mp != NULL, 1188 ("vp %p with NULL v_mount", vp)); 1189 vm_object_reference_locked(object); 1190 VM_OBJECT_UNLOCK(object); 1191 if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK, 1192 curthread)) { 1193 VM_OBJECT_LOCK(object); 1194 ++pageout_lock_miss; 1195 if (object->flags & OBJ_MIGHTBEDIRTY) 1196 vnodes_skipped++; 1197 vp = NULL; 1198 goto unlock_and_continue; 1199 } 1200 VM_OBJECT_LOCK(object); 1201 vm_page_lock(m); 1202 vm_page_lock_queues(); 1203 queues_locked = TRUE; 1204 /* 1205 * The page might have been moved to another 1206 * queue during potential blocking in vget() 1207 * above. The page might have been freed and 1208 * reused for another vnode. 1209 */ 1210 if (m->queue != PQ_INACTIVE || 1211 m->object != object || 1212 TAILQ_NEXT(m, pageq) != &marker) { 1213 vm_page_unlock(m); 1214 if (object->flags & OBJ_MIGHTBEDIRTY) 1215 vnodes_skipped++; 1216 goto unlock_and_continue; 1217 } 1218 1219 /* 1220 * The page may have been busied during the 1221 * blocking in vget(). We don't move the 1222 * page back onto the end of the queue so that 1223 * statistics are more correct if we don't. 1224 */ 1225 if (m->busy || (m->oflags & VPO_BUSY)) { 1226 vm_page_unlock(m); 1227 goto unlock_and_continue; 1228 } 1229 1230 /* 1231 * If the page has become held it might 1232 * be undergoing I/O, so skip it 1233 */ 1234 if (m->hold_count) { 1235 vm_page_unlock(m); 1236 vm_pageout_requeue(m); 1237 if (object->flags & OBJ_MIGHTBEDIRTY) 1238 vnodes_skipped++; 1239 goto unlock_and_continue; 1240 } 1241 vm_page_unlock_queues(); 1242 queues_locked = FALSE; 1243 } 1244 1245 /* 1246 * If a page is dirty, then it is either being washed 1247 * (but not yet cleaned) or it is still in the 1248 * laundry. If it is still in the laundry, then we 1249 * start the cleaning operation. 1250 * 1251 * decrement page_shortage on success to account for 1252 * the (future) cleaned page. Otherwise we could wind 1253 * up laundering or cleaning too many pages. 1254 */ 1255 if (vm_pageout_clean(m) != 0) { 1256 --page_shortage; 1257 --maxlaunder; 1258 } 1259 unlock_and_continue: 1260 vm_page_lock_assert(m, MA_NOTOWNED); 1261 VM_OBJECT_UNLOCK(object); 1262 if (mp != NULL) { 1263 if (queues_locked) { 1264 vm_page_unlock_queues(); 1265 queues_locked = FALSE; 1266 } 1267 if (vp != NULL) 1268 vput(vp); 1269 vm_object_deallocate(object); 1270 vn_finished_write(mp); 1271 } 1272 vm_page_lock_assert(m, MA_NOTOWNED); 1273 goto relock_queues; 1274 } 1275 vm_page_unlock(m); 1276 VM_OBJECT_UNLOCK(object); 1277 relock_queues: 1278 if (!queues_locked) { 1279 vm_page_lock_queues(); 1280 queues_locked = TRUE; 1281 } 1282 next = TAILQ_NEXT(&marker, pageq); 1283 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, 1284 &marker, pageq); 1285 } 1286 1287 /* 1288 * Compute the number of pages we want to try to move from the 1289 * active queue to the inactive queue. 1290 */ 1291 page_shortage = vm_paging_target() + 1292 cnt.v_inactive_target - cnt.v_inactive_count; 1293 page_shortage += addl_page_shortage; 1294 1295 /* 1296 * Scan the active queue for things we can deactivate. We nominally 1297 * track the per-page activity counter and use it to locate 1298 * deactivation candidates. 1299 */ 1300 pcount = cnt.v_active_count; 1301 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1302 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1303 1304 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) { 1305 1306 KASSERT(m->queue == PQ_ACTIVE, 1307 ("vm_pageout_scan: page %p isn't active", m)); 1308 1309 next = TAILQ_NEXT(m, pageq); 1310 if ((m->flags & PG_MARKER) != 0) { 1311 m = next; 1312 continue; 1313 } 1314 KASSERT((m->flags & PG_FICTITIOUS) == 0, 1315 ("Fictitious page %p cannot be in active queue", m)); 1316 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 1317 ("Unmanaged page %p cannot be in active queue", m)); 1318 if (!vm_pageout_page_lock(m, &next)) { 1319 vm_page_unlock(m); 1320 m = next; 1321 continue; 1322 } 1323 object = m->object; 1324 if (!VM_OBJECT_TRYLOCK(object) && 1325 !vm_pageout_fallback_object_lock(m, &next)) { 1326 VM_OBJECT_UNLOCK(object); 1327 vm_page_unlock(m); 1328 m = next; 1329 continue; 1330 } 1331 1332 /* 1333 * Don't deactivate pages that are busy. 1334 */ 1335 if ((m->busy != 0) || 1336 (m->oflags & VPO_BUSY) || 1337 (m->hold_count != 0)) { 1338 vm_page_unlock(m); 1339 VM_OBJECT_UNLOCK(object); 1340 vm_pageout_requeue(m); 1341 m = next; 1342 continue; 1343 } 1344 1345 /* 1346 * The count for pagedaemon pages is done after checking the 1347 * page for eligibility... 1348 */ 1349 cnt.v_pdpages++; 1350 1351 /* 1352 * Check to see "how much" the page has been used. 1353 */ 1354 actcount = 0; 1355 if (object->ref_count != 0) { 1356 if (m->aflags & PGA_REFERENCED) { 1357 actcount += 1; 1358 } 1359 actcount += pmap_ts_referenced(m); 1360 if (actcount) { 1361 m->act_count += ACT_ADVANCE + actcount; 1362 if (m->act_count > ACT_MAX) 1363 m->act_count = ACT_MAX; 1364 } 1365 } 1366 1367 /* 1368 * Since we have "tested" this bit, we need to clear it now. 1369 */ 1370 vm_page_aflag_clear(m, PGA_REFERENCED); 1371 1372 /* 1373 * Only if an object is currently being used, do we use the 1374 * page activation count stats. 1375 */ 1376 if (actcount && (object->ref_count != 0)) { 1377 vm_pageout_requeue(m); 1378 } else { 1379 m->act_count -= min(m->act_count, ACT_DECLINE); 1380 if (vm_pageout_algorithm || 1381 object->ref_count == 0 || 1382 m->act_count == 0) { 1383 page_shortage--; 1384 if (object->ref_count == 0) { 1385 KASSERT(!pmap_page_is_mapped(m), 1386 ("vm_pageout_scan: page %p is mapped", m)); 1387 if (m->dirty == 0) 1388 vm_page_cache(m); 1389 else 1390 vm_page_deactivate(m); 1391 } else { 1392 vm_page_deactivate(m); 1393 } 1394 } else { 1395 vm_pageout_requeue(m); 1396 } 1397 } 1398 vm_page_unlock(m); 1399 VM_OBJECT_UNLOCK(object); 1400 m = next; 1401 } 1402 vm_page_unlock_queues(); 1403 #if !defined(NO_SWAPPING) 1404 /* 1405 * Idle process swapout -- run once per second. 1406 */ 1407 if (vm_swap_idle_enabled) { 1408 static long lsec; 1409 if (time_second != lsec) { 1410 vm_req_vmdaemon(VM_SWAP_IDLE); 1411 lsec = time_second; 1412 } 1413 } 1414 #endif 1415 1416 /* 1417 * If we didn't get enough free pages, and we have skipped a vnode 1418 * in a writeable object, wakeup the sync daemon. And kick swapout 1419 * if we did not get enough free pages. 1420 */ 1421 if (vm_paging_target() > 0) { 1422 if (vnodes_skipped && vm_page_count_min()) 1423 (void) speedup_syncer(); 1424 #if !defined(NO_SWAPPING) 1425 if (vm_swap_enabled && vm_page_count_target()) 1426 vm_req_vmdaemon(VM_SWAP_NORMAL); 1427 #endif 1428 } 1429 1430 /* 1431 * If we are critically low on one of RAM or swap and low on 1432 * the other, kill the largest process. However, we avoid 1433 * doing this on the first pass in order to give ourselves a 1434 * chance to flush out dirty vnode-backed pages and to allow 1435 * active pages to be moved to the inactive queue and reclaimed. 1436 */ 1437 if (pass != 0 && 1438 ((swap_pager_avail < 64 && vm_page_count_min()) || 1439 (swap_pager_full && vm_paging_target() > 0))) 1440 vm_pageout_oom(VM_OOM_MEM); 1441 } 1442 1443 1444 void 1445 vm_pageout_oom(int shortage) 1446 { 1447 struct proc *p, *bigproc; 1448 vm_offset_t size, bigsize; 1449 struct thread *td; 1450 struct vmspace *vm; 1451 1452 /* 1453 * We keep the process bigproc locked once we find it to keep anyone 1454 * from messing with it; however, there is a possibility of 1455 * deadlock if process B is bigproc and one of it's child processes 1456 * attempts to propagate a signal to B while we are waiting for A's 1457 * lock while walking this list. To avoid this, we don't block on 1458 * the process lock but just skip a process if it is already locked. 1459 */ 1460 bigproc = NULL; 1461 bigsize = 0; 1462 sx_slock(&allproc_lock); 1463 FOREACH_PROC_IN_SYSTEM(p) { 1464 int breakout; 1465 1466 if (PROC_TRYLOCK(p) == 0) 1467 continue; 1468 /* 1469 * If this is a system, protected or killed process, skip it. 1470 */ 1471 if (p->p_state != PRS_NORMAL || 1472 (p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM)) || 1473 (p->p_pid == 1) || P_KILLED(p) || 1474 ((p->p_pid < 48) && (swap_pager_avail != 0))) { 1475 PROC_UNLOCK(p); 1476 continue; 1477 } 1478 /* 1479 * If the process is in a non-running type state, 1480 * don't touch it. Check all the threads individually. 1481 */ 1482 breakout = 0; 1483 FOREACH_THREAD_IN_PROC(p, td) { 1484 thread_lock(td); 1485 if (!TD_ON_RUNQ(td) && 1486 !TD_IS_RUNNING(td) && 1487 !TD_IS_SLEEPING(td) && 1488 !TD_IS_SUSPENDED(td)) { 1489 thread_unlock(td); 1490 breakout = 1; 1491 break; 1492 } 1493 thread_unlock(td); 1494 } 1495 if (breakout) { 1496 PROC_UNLOCK(p); 1497 continue; 1498 } 1499 /* 1500 * get the process size 1501 */ 1502 vm = vmspace_acquire_ref(p); 1503 if (vm == NULL) { 1504 PROC_UNLOCK(p); 1505 continue; 1506 } 1507 if (!vm_map_trylock_read(&vm->vm_map)) { 1508 vmspace_free(vm); 1509 PROC_UNLOCK(p); 1510 continue; 1511 } 1512 size = vmspace_swap_count(vm); 1513 vm_map_unlock_read(&vm->vm_map); 1514 if (shortage == VM_OOM_MEM) 1515 size += vmspace_resident_count(vm); 1516 vmspace_free(vm); 1517 /* 1518 * if the this process is bigger than the biggest one 1519 * remember it. 1520 */ 1521 if (size > bigsize) { 1522 if (bigproc != NULL) 1523 PROC_UNLOCK(bigproc); 1524 bigproc = p; 1525 bigsize = size; 1526 } else 1527 PROC_UNLOCK(p); 1528 } 1529 sx_sunlock(&allproc_lock); 1530 if (bigproc != NULL) { 1531 killproc(bigproc, "out of swap space"); 1532 sched_nice(bigproc, PRIO_MIN); 1533 PROC_UNLOCK(bigproc); 1534 wakeup(&cnt.v_free_count); 1535 } 1536 } 1537 1538 /* 1539 * This routine tries to maintain the pseudo LRU active queue, 1540 * so that during long periods of time where there is no paging, 1541 * that some statistic accumulation still occurs. This code 1542 * helps the situation where paging just starts to occur. 1543 */ 1544 static void 1545 vm_pageout_page_stats() 1546 { 1547 vm_object_t object; 1548 vm_page_t m,next; 1549 int pcount,tpcount; /* Number of pages to check */ 1550 static int fullintervalcount = 0; 1551 int page_shortage; 1552 1553 page_shortage = 1554 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) - 1555 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count); 1556 1557 if (page_shortage <= 0) 1558 return; 1559 1560 vm_page_lock_queues(); 1561 pcount = cnt.v_active_count; 1562 fullintervalcount += vm_pageout_stats_interval; 1563 if (fullintervalcount < vm_pageout_full_stats_interval) { 1564 tpcount = (int64_t)vm_pageout_stats_max * cnt.v_active_count / 1565 cnt.v_page_count; 1566 if (pcount > tpcount) 1567 pcount = tpcount; 1568 } else { 1569 fullintervalcount = 0; 1570 } 1571 1572 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1573 while ((m != NULL) && (pcount-- > 0)) { 1574 int actcount; 1575 1576 KASSERT(m->queue == PQ_ACTIVE, 1577 ("vm_pageout_page_stats: page %p isn't active", m)); 1578 1579 next = TAILQ_NEXT(m, pageq); 1580 if ((m->flags & PG_MARKER) != 0) { 1581 m = next; 1582 continue; 1583 } 1584 vm_page_lock_assert(m, MA_NOTOWNED); 1585 if (!vm_pageout_page_lock(m, &next)) { 1586 vm_page_unlock(m); 1587 m = next; 1588 continue; 1589 } 1590 object = m->object; 1591 if (!VM_OBJECT_TRYLOCK(object) && 1592 !vm_pageout_fallback_object_lock(m, &next)) { 1593 VM_OBJECT_UNLOCK(object); 1594 vm_page_unlock(m); 1595 m = next; 1596 continue; 1597 } 1598 1599 /* 1600 * Don't deactivate pages that are busy. 1601 */ 1602 if ((m->busy != 0) || 1603 (m->oflags & VPO_BUSY) || 1604 (m->hold_count != 0)) { 1605 vm_page_unlock(m); 1606 VM_OBJECT_UNLOCK(object); 1607 vm_pageout_requeue(m); 1608 m = next; 1609 continue; 1610 } 1611 1612 actcount = 0; 1613 if (m->aflags & PGA_REFERENCED) { 1614 vm_page_aflag_clear(m, PGA_REFERENCED); 1615 actcount += 1; 1616 } 1617 1618 actcount += pmap_ts_referenced(m); 1619 if (actcount) { 1620 m->act_count += ACT_ADVANCE + actcount; 1621 if (m->act_count > ACT_MAX) 1622 m->act_count = ACT_MAX; 1623 vm_pageout_requeue(m); 1624 } else { 1625 if (m->act_count == 0) { 1626 /* 1627 * We turn off page access, so that we have 1628 * more accurate RSS stats. We don't do this 1629 * in the normal page deactivation when the 1630 * system is loaded VM wise, because the 1631 * cost of the large number of page protect 1632 * operations would be higher than the value 1633 * of doing the operation. 1634 */ 1635 pmap_remove_all(m); 1636 vm_page_deactivate(m); 1637 } else { 1638 m->act_count -= min(m->act_count, ACT_DECLINE); 1639 vm_pageout_requeue(m); 1640 } 1641 } 1642 vm_page_unlock(m); 1643 VM_OBJECT_UNLOCK(object); 1644 m = next; 1645 } 1646 vm_page_unlock_queues(); 1647 } 1648 1649 /* 1650 * vm_pageout is the high level pageout daemon. 1651 */ 1652 static void 1653 vm_pageout() 1654 { 1655 int error, pass; 1656 1657 /* 1658 * Initialize some paging parameters. 1659 */ 1660 cnt.v_interrupt_free_min = 2; 1661 if (cnt.v_page_count < 2000) 1662 vm_pageout_page_count = 8; 1663 1664 /* 1665 * v_free_reserved needs to include enough for the largest 1666 * swap pager structures plus enough for any pv_entry structs 1667 * when paging. 1668 */ 1669 if (cnt.v_page_count > 1024) 1670 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200; 1671 else 1672 cnt.v_free_min = 4; 1673 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1674 cnt.v_interrupt_free_min; 1675 cnt.v_free_reserved = vm_pageout_page_count + 1676 cnt.v_pageout_free_min + (cnt.v_page_count / 768); 1677 cnt.v_free_severe = cnt.v_free_min / 2; 1678 cnt.v_free_min += cnt.v_free_reserved; 1679 cnt.v_free_severe += cnt.v_free_reserved; 1680 1681 /* 1682 * v_free_target and v_cache_min control pageout hysteresis. Note 1683 * that these are more a measure of the VM cache queue hysteresis 1684 * then the VM free queue. Specifically, v_free_target is the 1685 * high water mark (free+cache pages). 1686 * 1687 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the 1688 * low water mark, while v_free_min is the stop. v_cache_min must 1689 * be big enough to handle memory needs while the pageout daemon 1690 * is signalled and run to free more pages. 1691 */ 1692 if (cnt.v_free_count > 6144) 1693 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved; 1694 else 1695 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved; 1696 1697 if (cnt.v_free_count > 2048) { 1698 cnt.v_cache_min = cnt.v_free_target; 1699 cnt.v_cache_max = 2 * cnt.v_cache_min; 1700 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2; 1701 } else { 1702 cnt.v_cache_min = 0; 1703 cnt.v_cache_max = 0; 1704 cnt.v_inactive_target = cnt.v_free_count / 4; 1705 } 1706 if (cnt.v_inactive_target > cnt.v_free_count / 3) 1707 cnt.v_inactive_target = cnt.v_free_count / 3; 1708 1709 /* XXX does not really belong here */ 1710 if (vm_page_max_wired == 0) 1711 vm_page_max_wired = cnt.v_free_count / 3; 1712 1713 if (vm_pageout_stats_max == 0) 1714 vm_pageout_stats_max = cnt.v_free_target; 1715 1716 /* 1717 * Set interval in seconds for stats scan. 1718 */ 1719 if (vm_pageout_stats_interval == 0) 1720 vm_pageout_stats_interval = 5; 1721 if (vm_pageout_full_stats_interval == 0) 1722 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4; 1723 1724 swap_pager_swap_init(); 1725 pass = 0; 1726 /* 1727 * The pageout daemon is never done, so loop forever. 1728 */ 1729 while (TRUE) { 1730 /* 1731 * If we have enough free memory, wakeup waiters. Do 1732 * not clear vm_pages_needed until we reach our target, 1733 * otherwise we may be woken up over and over again and 1734 * waste a lot of cpu. 1735 */ 1736 mtx_lock(&vm_page_queue_free_mtx); 1737 if (vm_pages_needed && !vm_page_count_min()) { 1738 if (!vm_paging_needed()) 1739 vm_pages_needed = 0; 1740 wakeup(&cnt.v_free_count); 1741 } 1742 if (vm_pages_needed) { 1743 /* 1744 * Still not done, take a second pass without waiting 1745 * (unlimited dirty cleaning), otherwise sleep a bit 1746 * and try again. 1747 */ 1748 ++pass; 1749 if (pass > 1) 1750 msleep(&vm_pages_needed, 1751 &vm_page_queue_free_mtx, PVM, "psleep", 1752 hz / 2); 1753 } else { 1754 /* 1755 * Good enough, sleep & handle stats. Prime the pass 1756 * for the next run. 1757 */ 1758 if (pass > 1) 1759 pass = 1; 1760 else 1761 pass = 0; 1762 error = msleep(&vm_pages_needed, 1763 &vm_page_queue_free_mtx, PVM, "psleep", 1764 vm_pageout_stats_interval * hz); 1765 if (error && !vm_pages_needed) { 1766 mtx_unlock(&vm_page_queue_free_mtx); 1767 pass = 0; 1768 vm_pageout_page_stats(); 1769 continue; 1770 } 1771 } 1772 if (vm_pages_needed) 1773 cnt.v_pdwakeups++; 1774 mtx_unlock(&vm_page_queue_free_mtx); 1775 vm_pageout_scan(pass); 1776 } 1777 } 1778 1779 /* 1780 * Unless the free page queue lock is held by the caller, this function 1781 * should be regarded as advisory. Specifically, the caller should 1782 * not msleep() on &cnt.v_free_count following this function unless 1783 * the free page queue lock is held until the msleep() is performed. 1784 */ 1785 void 1786 pagedaemon_wakeup() 1787 { 1788 1789 if (!vm_pages_needed && curthread->td_proc != pageproc) { 1790 vm_pages_needed = 1; 1791 wakeup(&vm_pages_needed); 1792 } 1793 } 1794 1795 #if !defined(NO_SWAPPING) 1796 static void 1797 vm_req_vmdaemon(int req) 1798 { 1799 static int lastrun = 0; 1800 1801 mtx_lock(&vm_daemon_mtx); 1802 vm_pageout_req_swapout |= req; 1803 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1804 wakeup(&vm_daemon_needed); 1805 lastrun = ticks; 1806 } 1807 mtx_unlock(&vm_daemon_mtx); 1808 } 1809 1810 static void 1811 vm_daemon() 1812 { 1813 struct rlimit rsslim; 1814 struct proc *p; 1815 struct thread *td; 1816 struct vmspace *vm; 1817 int breakout, swapout_flags, tryagain, attempts; 1818 #ifdef RACCT 1819 uint64_t rsize, ravailable; 1820 #endif 1821 1822 while (TRUE) { 1823 mtx_lock(&vm_daemon_mtx); 1824 #ifdef RACCT 1825 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", hz); 1826 #else 1827 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0); 1828 #endif 1829 swapout_flags = vm_pageout_req_swapout; 1830 vm_pageout_req_swapout = 0; 1831 mtx_unlock(&vm_daemon_mtx); 1832 if (swapout_flags) 1833 swapout_procs(swapout_flags); 1834 1835 /* 1836 * scan the processes for exceeding their rlimits or if 1837 * process is swapped out -- deactivate pages 1838 */ 1839 tryagain = 0; 1840 attempts = 0; 1841 again: 1842 attempts++; 1843 sx_slock(&allproc_lock); 1844 FOREACH_PROC_IN_SYSTEM(p) { 1845 vm_pindex_t limit, size; 1846 1847 /* 1848 * if this is a system process or if we have already 1849 * looked at this process, skip it. 1850 */ 1851 PROC_LOCK(p); 1852 if (p->p_state != PRS_NORMAL || 1853 p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) { 1854 PROC_UNLOCK(p); 1855 continue; 1856 } 1857 /* 1858 * if the process is in a non-running type state, 1859 * don't touch it. 1860 */ 1861 breakout = 0; 1862 FOREACH_THREAD_IN_PROC(p, td) { 1863 thread_lock(td); 1864 if (!TD_ON_RUNQ(td) && 1865 !TD_IS_RUNNING(td) && 1866 !TD_IS_SLEEPING(td) && 1867 !TD_IS_SUSPENDED(td)) { 1868 thread_unlock(td); 1869 breakout = 1; 1870 break; 1871 } 1872 thread_unlock(td); 1873 } 1874 if (breakout) { 1875 PROC_UNLOCK(p); 1876 continue; 1877 } 1878 /* 1879 * get a limit 1880 */ 1881 lim_rlimit(p, RLIMIT_RSS, &rsslim); 1882 limit = OFF_TO_IDX( 1883 qmin(rsslim.rlim_cur, rsslim.rlim_max)); 1884 1885 /* 1886 * let processes that are swapped out really be 1887 * swapped out set the limit to nothing (will force a 1888 * swap-out.) 1889 */ 1890 if ((p->p_flag & P_INMEM) == 0) 1891 limit = 0; /* XXX */ 1892 vm = vmspace_acquire_ref(p); 1893 PROC_UNLOCK(p); 1894 if (vm == NULL) 1895 continue; 1896 1897 size = vmspace_resident_count(vm); 1898 if (limit >= 0 && size >= limit) { 1899 vm_pageout_map_deactivate_pages( 1900 &vm->vm_map, limit); 1901 } 1902 #ifdef RACCT 1903 rsize = IDX_TO_OFF(size); 1904 PROC_LOCK(p); 1905 racct_set(p, RACCT_RSS, rsize); 1906 ravailable = racct_get_available(p, RACCT_RSS); 1907 PROC_UNLOCK(p); 1908 if (rsize > ravailable) { 1909 /* 1910 * Don't be overly aggressive; this might be 1911 * an innocent process, and the limit could've 1912 * been exceeded by some memory hog. Don't 1913 * try to deactivate more than 1/4th of process' 1914 * resident set size. 1915 */ 1916 if (attempts <= 8) { 1917 if (ravailable < rsize - (rsize / 4)) 1918 ravailable = rsize - (rsize / 4); 1919 } 1920 vm_pageout_map_deactivate_pages( 1921 &vm->vm_map, OFF_TO_IDX(ravailable)); 1922 /* Update RSS usage after paging out. */ 1923 size = vmspace_resident_count(vm); 1924 rsize = IDX_TO_OFF(size); 1925 PROC_LOCK(p); 1926 racct_set(p, RACCT_RSS, rsize); 1927 PROC_UNLOCK(p); 1928 if (rsize > ravailable) 1929 tryagain = 1; 1930 } 1931 #endif 1932 vmspace_free(vm); 1933 } 1934 sx_sunlock(&allproc_lock); 1935 if (tryagain != 0 && attempts <= 10) 1936 goto again; 1937 } 1938 } 1939 #endif /* !defined(NO_SWAPPING) */ 1940