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