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 "opt_kdtrace.h" 80 #include <sys/param.h> 81 #include <sys/systm.h> 82 #include <sys/kernel.h> 83 #include <sys/eventhandler.h> 84 #include <sys/lock.h> 85 #include <sys/mutex.h> 86 #include <sys/proc.h> 87 #include <sys/kthread.h> 88 #include <sys/ktr.h> 89 #include <sys/mount.h> 90 #include <sys/racct.h> 91 #include <sys/resourcevar.h> 92 #include <sys/sched.h> 93 #include <sys/sdt.h> 94 #include <sys/signalvar.h> 95 #include <sys/smp.h> 96 #include <sys/time.h> 97 #include <sys/vnode.h> 98 #include <sys/vmmeter.h> 99 #include <sys/rwlock.h> 100 #include <sys/sx.h> 101 #include <sys/sysctl.h> 102 103 #include <vm/vm.h> 104 #include <vm/vm_param.h> 105 #include <vm/vm_object.h> 106 #include <vm/vm_page.h> 107 #include <vm/vm_map.h> 108 #include <vm/vm_pageout.h> 109 #include <vm/vm_pager.h> 110 #include <vm/vm_phys.h> 111 #include <vm/swap_pager.h> 112 #include <vm/vm_extern.h> 113 #include <vm/uma.h> 114 115 /* 116 * System initialization 117 */ 118 119 /* the kernel process "vm_pageout"*/ 120 static void vm_pageout(void); 121 static void vm_pageout_init(void); 122 static int vm_pageout_clean(vm_page_t m); 123 static int vm_pageout_cluster(vm_page_t m); 124 static void vm_pageout_scan(struct vm_domain *vmd, int pass); 125 static void vm_pageout_mightbe_oom(struct vm_domain *vmd, int pass); 126 127 SYSINIT(pagedaemon_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, vm_pageout_init, 128 NULL); 129 130 struct proc *pageproc; 131 132 static struct kproc_desc page_kp = { 133 "pagedaemon", 134 vm_pageout, 135 &pageproc 136 }; 137 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start, 138 &page_kp); 139 140 SDT_PROVIDER_DEFINE(vm); 141 SDT_PROBE_DEFINE(vm, , , vm__lowmem_cache); 142 SDT_PROBE_DEFINE(vm, , , vm__lowmem_scan); 143 144 #if !defined(NO_SWAPPING) 145 /* the kernel process "vm_daemon"*/ 146 static void vm_daemon(void); 147 static struct proc *vmproc; 148 149 static struct kproc_desc vm_kp = { 150 "vmdaemon", 151 vm_daemon, 152 &vmproc 153 }; 154 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp); 155 #endif 156 157 158 int vm_pages_needed; /* Event on which pageout daemon sleeps */ 159 int vm_pageout_deficit; /* Estimated number of pages deficit */ 160 int vm_pageout_wakeup_thresh; 161 162 #if !defined(NO_SWAPPING) 163 static int vm_pageout_req_swapout; /* XXX */ 164 static int vm_daemon_needed; 165 static struct mtx vm_daemon_mtx; 166 /* Allow for use by vm_pageout before vm_daemon is initialized. */ 167 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF); 168 #endif 169 static int vm_max_launder = 32; 170 static int vm_pageout_update_period; 171 static int defer_swap_pageouts; 172 static int disable_swap_pageouts; 173 static int lowmem_period = 10; 174 static time_t lowmem_uptime; 175 176 #if defined(NO_SWAPPING) 177 static int vm_swap_enabled = 0; 178 static int vm_swap_idle_enabled = 0; 179 #else 180 static int vm_swap_enabled = 1; 181 static int vm_swap_idle_enabled = 0; 182 #endif 183 184 static int vm_panic_on_oom = 0; 185 186 SYSCTL_INT(_vm, OID_AUTO, panic_on_oom, 187 CTLFLAG_RWTUN, &vm_panic_on_oom, 0, 188 "panic on out of memory instead of killing the largest process"); 189 190 SYSCTL_INT(_vm, OID_AUTO, pageout_wakeup_thresh, 191 CTLFLAG_RW, &vm_pageout_wakeup_thresh, 0, 192 "free page threshold for waking up the pageout daemon"); 193 194 SYSCTL_INT(_vm, OID_AUTO, max_launder, 195 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout"); 196 197 SYSCTL_INT(_vm, OID_AUTO, pageout_update_period, 198 CTLFLAG_RW, &vm_pageout_update_period, 0, 199 "Maximum active LRU update period"); 200 201 SYSCTL_INT(_vm, OID_AUTO, lowmem_period, CTLFLAG_RW, &lowmem_period, 0, 202 "Low memory callback period"); 203 204 #if defined(NO_SWAPPING) 205 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 206 CTLFLAG_RD, &vm_swap_enabled, 0, "Enable entire process swapout"); 207 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 208 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 209 #else 210 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 211 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout"); 212 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 213 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 214 #endif 215 216 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts, 217 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem"); 218 219 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts, 220 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages"); 221 222 static int pageout_lock_miss; 223 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss, 224 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout"); 225 226 #define VM_PAGEOUT_PAGE_COUNT 16 227 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT; 228 229 int vm_page_max_wired; /* XXX max # of wired pages system-wide */ 230 SYSCTL_INT(_vm, OID_AUTO, max_wired, 231 CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count"); 232 233 static boolean_t vm_pageout_fallback_object_lock(vm_page_t, vm_page_t *); 234 static boolean_t vm_pageout_launder(struct vm_pagequeue *pq, int, vm_paddr_t, 235 vm_paddr_t); 236 #if !defined(NO_SWAPPING) 237 static void vm_pageout_map_deactivate_pages(vm_map_t, long); 238 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long); 239 static void vm_req_vmdaemon(int req); 240 #endif 241 static boolean_t vm_pageout_page_lock(vm_page_t, vm_page_t *); 242 243 /* 244 * Initialize a dummy page for marking the caller's place in the specified 245 * paging queue. In principle, this function only needs to set the flag 246 * PG_MARKER. Nonetheless, it wirte busies and initializes the hold count 247 * to one as safety precautions. 248 */ 249 static void 250 vm_pageout_init_marker(vm_page_t marker, u_short queue) 251 { 252 253 bzero(marker, sizeof(*marker)); 254 marker->flags = PG_MARKER; 255 marker->busy_lock = VPB_SINGLE_EXCLUSIVER; 256 marker->queue = queue; 257 marker->hold_count = 1; 258 } 259 260 /* 261 * vm_pageout_fallback_object_lock: 262 * 263 * Lock vm object currently associated with `m'. VM_OBJECT_TRYWLOCK is 264 * known to have failed and page queue must be either PQ_ACTIVE or 265 * PQ_INACTIVE. To avoid lock order violation, unlock the page queues 266 * while locking the vm object. Use marker page to detect page queue 267 * changes and maintain notion of next page on page queue. Return 268 * TRUE if no changes were detected, FALSE otherwise. vm object is 269 * locked on return. 270 * 271 * This function depends on both the lock portion of struct vm_object 272 * and normal struct vm_page being type stable. 273 */ 274 static boolean_t 275 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next) 276 { 277 struct vm_page marker; 278 struct vm_pagequeue *pq; 279 boolean_t unchanged; 280 u_short queue; 281 vm_object_t object; 282 283 queue = m->queue; 284 vm_pageout_init_marker(&marker, queue); 285 pq = vm_page_pagequeue(m); 286 object = m->object; 287 288 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q); 289 vm_pagequeue_unlock(pq); 290 vm_page_unlock(m); 291 VM_OBJECT_WLOCK(object); 292 vm_page_lock(m); 293 vm_pagequeue_lock(pq); 294 295 /* Page queue might have changed. */ 296 *next = TAILQ_NEXT(&marker, plinks.q); 297 unchanged = (m->queue == queue && 298 m->object == object && 299 &marker == TAILQ_NEXT(m, plinks.q)); 300 TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q); 301 return (unchanged); 302 } 303 304 /* 305 * Lock the page while holding the page queue lock. Use marker page 306 * to detect page queue changes and maintain notion of next page on 307 * page queue. Return TRUE if no changes were detected, FALSE 308 * otherwise. The page is locked on return. The page queue lock might 309 * be dropped and reacquired. 310 * 311 * This function depends on normal struct vm_page being type stable. 312 */ 313 static boolean_t 314 vm_pageout_page_lock(vm_page_t m, vm_page_t *next) 315 { 316 struct vm_page marker; 317 struct vm_pagequeue *pq; 318 boolean_t unchanged; 319 u_short queue; 320 321 vm_page_lock_assert(m, MA_NOTOWNED); 322 if (vm_page_trylock(m)) 323 return (TRUE); 324 325 queue = m->queue; 326 vm_pageout_init_marker(&marker, queue); 327 pq = vm_page_pagequeue(m); 328 329 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q); 330 vm_pagequeue_unlock(pq); 331 vm_page_lock(m); 332 vm_pagequeue_lock(pq); 333 334 /* Page queue might have changed. */ 335 *next = TAILQ_NEXT(&marker, plinks.q); 336 unchanged = (m->queue == queue && &marker == TAILQ_NEXT(m, plinks.q)); 337 TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q); 338 return (unchanged); 339 } 340 341 /* 342 * vm_pageout_clean: 343 * 344 * Clean the page and remove it from the laundry. 345 * 346 * We set the busy bit to cause potential page faults on this page to 347 * block. Note the careful timing, however, the busy bit isn't set till 348 * late and we cannot do anything that will mess with the page. 349 */ 350 static int 351 vm_pageout_cluster(vm_page_t m) 352 { 353 vm_object_t object; 354 vm_page_t mc[2*vm_pageout_page_count], pb, ps; 355 int pageout_count; 356 int ib, is, page_base; 357 vm_pindex_t pindex = m->pindex; 358 359 vm_page_lock_assert(m, MA_OWNED); 360 object = m->object; 361 VM_OBJECT_ASSERT_WLOCKED(object); 362 363 /* 364 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP 365 * with the new swapper, but we could have serious problems paging 366 * out other object types if there is insufficient memory. 367 * 368 * Unfortunately, checking free memory here is far too late, so the 369 * check has been moved up a procedural level. 370 */ 371 372 /* 373 * Can't clean the page if it's busy or held. 374 */ 375 vm_page_assert_unbusied(m); 376 KASSERT(m->hold_count == 0, ("vm_pageout_clean: page %p is held", m)); 377 vm_page_unlock(m); 378 379 mc[vm_pageout_page_count] = pb = ps = m; 380 pageout_count = 1; 381 page_base = vm_pageout_page_count; 382 ib = 1; 383 is = 1; 384 385 /* 386 * Scan object for clusterable pages. 387 * 388 * We can cluster ONLY if: ->> the page is NOT 389 * clean, wired, busy, held, or mapped into a 390 * buffer, and one of the following: 391 * 1) The page is inactive, or a seldom used 392 * active page. 393 * -or- 394 * 2) we force the issue. 395 * 396 * During heavy mmap/modification loads the pageout 397 * daemon can really fragment the underlying file 398 * due to flushing pages out of order and not trying 399 * align the clusters (which leave sporatic out-of-order 400 * holes). To solve this problem we do the reverse scan 401 * first and attempt to align our cluster, then do a 402 * forward scan if room remains. 403 */ 404 more: 405 while (ib && pageout_count < vm_pageout_page_count) { 406 vm_page_t p; 407 408 if (ib > pindex) { 409 ib = 0; 410 break; 411 } 412 413 if ((p = vm_page_prev(pb)) == NULL || vm_page_busied(p)) { 414 ib = 0; 415 break; 416 } 417 vm_page_test_dirty(p); 418 if (p->dirty == 0) { 419 ib = 0; 420 break; 421 } 422 vm_page_lock(p); 423 if (p->queue != PQ_INACTIVE || 424 p->hold_count != 0) { /* may be undergoing I/O */ 425 vm_page_unlock(p); 426 ib = 0; 427 break; 428 } 429 vm_page_unlock(p); 430 mc[--page_base] = pb = p; 431 ++pageout_count; 432 ++ib; 433 /* 434 * alignment boundry, stop here and switch directions. Do 435 * not clear ib. 436 */ 437 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0) 438 break; 439 } 440 441 while (pageout_count < vm_pageout_page_count && 442 pindex + is < object->size) { 443 vm_page_t p; 444 445 if ((p = vm_page_next(ps)) == NULL || vm_page_busied(p)) 446 break; 447 vm_page_test_dirty(p); 448 if (p->dirty == 0) 449 break; 450 vm_page_lock(p); 451 if (p->queue != PQ_INACTIVE || 452 p->hold_count != 0) { /* may be undergoing I/O */ 453 vm_page_unlock(p); 454 break; 455 } 456 vm_page_unlock(p); 457 mc[page_base + pageout_count] = ps = p; 458 ++pageout_count; 459 ++is; 460 } 461 462 /* 463 * If we exhausted our forward scan, continue with the reverse scan 464 * when possible, even past a page boundry. This catches boundry 465 * conditions. 466 */ 467 if (ib && pageout_count < vm_pageout_page_count) 468 goto more; 469 470 /* 471 * we allow reads during pageouts... 472 */ 473 return (vm_pageout_flush(&mc[page_base], pageout_count, 0, 0, NULL, 474 NULL)); 475 } 476 477 /* 478 * vm_pageout_flush() - launder the given pages 479 * 480 * The given pages are laundered. Note that we setup for the start of 481 * I/O ( i.e. busy the page ), mark it read-only, and bump the object 482 * reference count all in here rather then in the parent. If we want 483 * the parent to do more sophisticated things we may have to change 484 * the ordering. 485 * 486 * Returned runlen is the count of pages between mreq and first 487 * page after mreq with status VM_PAGER_AGAIN. 488 * *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL 489 * for any page in runlen set. 490 */ 491 int 492 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen, 493 boolean_t *eio) 494 { 495 vm_object_t object = mc[0]->object; 496 int pageout_status[count]; 497 int numpagedout = 0; 498 int i, runlen; 499 500 VM_OBJECT_ASSERT_WLOCKED(object); 501 502 /* 503 * Initiate I/O. Bump the vm_page_t->busy counter and 504 * mark the pages read-only. 505 * 506 * We do not have to fixup the clean/dirty bits here... we can 507 * allow the pager to do it after the I/O completes. 508 * 509 * NOTE! mc[i]->dirty may be partial or fragmented due to an 510 * edge case with file fragments. 511 */ 512 for (i = 0; i < count; i++) { 513 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, 514 ("vm_pageout_flush: partially invalid page %p index %d/%d", 515 mc[i], i, count)); 516 vm_page_sbusy(mc[i]); 517 pmap_remove_write(mc[i]); 518 } 519 vm_object_pip_add(object, count); 520 521 vm_pager_put_pages(object, mc, count, flags, pageout_status); 522 523 runlen = count - mreq; 524 if (eio != NULL) 525 *eio = FALSE; 526 for (i = 0; i < count; i++) { 527 vm_page_t mt = mc[i]; 528 529 KASSERT(pageout_status[i] == VM_PAGER_PEND || 530 !pmap_page_is_write_mapped(mt), 531 ("vm_pageout_flush: page %p is not write protected", mt)); 532 switch (pageout_status[i]) { 533 case VM_PAGER_OK: 534 case VM_PAGER_PEND: 535 numpagedout++; 536 break; 537 case VM_PAGER_BAD: 538 /* 539 * Page outside of range of object. Right now we 540 * essentially lose the changes by pretending it 541 * worked. 542 */ 543 vm_page_undirty(mt); 544 break; 545 case VM_PAGER_ERROR: 546 case VM_PAGER_FAIL: 547 /* 548 * If page couldn't be paged out, then reactivate the 549 * page so it doesn't clog the inactive list. (We 550 * will try paging out it again later). 551 */ 552 vm_page_lock(mt); 553 vm_page_activate(mt); 554 vm_page_unlock(mt); 555 if (eio != NULL && i >= mreq && i - mreq < runlen) 556 *eio = TRUE; 557 break; 558 case VM_PAGER_AGAIN: 559 if (i >= mreq && i - mreq < runlen) 560 runlen = i - mreq; 561 break; 562 } 563 564 /* 565 * If the operation is still going, leave the page busy to 566 * block all other accesses. Also, leave the paging in 567 * progress indicator set so that we don't attempt an object 568 * collapse. 569 */ 570 if (pageout_status[i] != VM_PAGER_PEND) { 571 vm_object_pip_wakeup(object); 572 vm_page_sunbusy(mt); 573 } 574 } 575 if (prunlen != NULL) 576 *prunlen = runlen; 577 return (numpagedout); 578 } 579 580 static boolean_t 581 vm_pageout_launder(struct vm_pagequeue *pq, int tries, vm_paddr_t low, 582 vm_paddr_t high) 583 { 584 struct mount *mp; 585 struct vnode *vp; 586 vm_object_t object; 587 vm_paddr_t pa; 588 vm_page_t m, m_tmp, next; 589 int lockmode; 590 591 vm_pagequeue_lock(pq); 592 TAILQ_FOREACH_SAFE(m, &pq->pq_pl, plinks.q, next) { 593 if ((m->flags & PG_MARKER) != 0) 594 continue; 595 pa = VM_PAGE_TO_PHYS(m); 596 if (pa < low || pa + PAGE_SIZE > high) 597 continue; 598 if (!vm_pageout_page_lock(m, &next) || m->hold_count != 0) { 599 vm_page_unlock(m); 600 continue; 601 } 602 object = m->object; 603 if ((!VM_OBJECT_TRYWLOCK(object) && 604 (!vm_pageout_fallback_object_lock(m, &next) || 605 m->hold_count != 0)) || vm_page_busied(m)) { 606 vm_page_unlock(m); 607 VM_OBJECT_WUNLOCK(object); 608 continue; 609 } 610 vm_page_test_dirty(m); 611 if (m->dirty == 0 && object->ref_count != 0) 612 pmap_remove_all(m); 613 if (m->dirty != 0) { 614 vm_page_unlock(m); 615 if (tries == 0 || (object->flags & OBJ_DEAD) != 0) { 616 VM_OBJECT_WUNLOCK(object); 617 continue; 618 } 619 if (object->type == OBJT_VNODE) { 620 vm_pagequeue_unlock(pq); 621 vp = object->handle; 622 vm_object_reference_locked(object); 623 VM_OBJECT_WUNLOCK(object); 624 (void)vn_start_write(vp, &mp, V_WAIT); 625 lockmode = MNT_SHARED_WRITES(vp->v_mount) ? 626 LK_SHARED : LK_EXCLUSIVE; 627 vn_lock(vp, lockmode | LK_RETRY); 628 VM_OBJECT_WLOCK(object); 629 vm_object_page_clean(object, 0, 0, OBJPC_SYNC); 630 VM_OBJECT_WUNLOCK(object); 631 VOP_UNLOCK(vp, 0); 632 vm_object_deallocate(object); 633 vn_finished_write(mp); 634 return (TRUE); 635 } else if (object->type == OBJT_SWAP || 636 object->type == OBJT_DEFAULT) { 637 vm_pagequeue_unlock(pq); 638 m_tmp = m; 639 vm_pageout_flush(&m_tmp, 1, VM_PAGER_PUT_SYNC, 640 0, NULL, NULL); 641 VM_OBJECT_WUNLOCK(object); 642 return (TRUE); 643 } 644 } else { 645 /* 646 * Dequeue here to prevent lock recursion in 647 * vm_page_cache(). 648 */ 649 vm_page_dequeue_locked(m); 650 vm_page_cache(m); 651 vm_page_unlock(m); 652 } 653 VM_OBJECT_WUNLOCK(object); 654 } 655 vm_pagequeue_unlock(pq); 656 return (FALSE); 657 } 658 659 /* 660 * Increase the number of cached pages. The specified value, "tries", 661 * determines which categories of pages are cached: 662 * 663 * 0: All clean, inactive pages within the specified physical address range 664 * are cached. Will not sleep. 665 * 1: The vm_lowmem handlers are called. All inactive pages within 666 * the specified physical address range are cached. May sleep. 667 * 2: The vm_lowmem handlers are called. All inactive and active pages 668 * within the specified physical address range are cached. May sleep. 669 */ 670 void 671 vm_pageout_grow_cache(int tries, vm_paddr_t low, vm_paddr_t high) 672 { 673 int actl, actmax, inactl, inactmax, dom, initial_dom; 674 static int start_dom = 0; 675 676 if (tries > 0) { 677 /* 678 * Decrease registered cache sizes. The vm_lowmem handlers 679 * may acquire locks and/or sleep, so they can only be invoked 680 * when "tries" is greater than zero. 681 */ 682 SDT_PROBE0(vm, , , vm__lowmem_cache); 683 EVENTHANDLER_INVOKE(vm_lowmem, 0); 684 685 /* 686 * We do this explicitly after the caches have been drained 687 * above. 688 */ 689 uma_reclaim(); 690 } 691 692 /* 693 * Make the next scan start on the next domain. 694 */ 695 initial_dom = atomic_fetchadd_int(&start_dom, 1) % vm_ndomains; 696 697 inactl = 0; 698 inactmax = vm_cnt.v_inactive_count; 699 actl = 0; 700 actmax = tries < 2 ? 0 : vm_cnt.v_active_count; 701 dom = initial_dom; 702 703 /* 704 * Scan domains in round-robin order, first inactive queues, 705 * then active. Since domain usually owns large physically 706 * contiguous chunk of memory, it makes sense to completely 707 * exhaust one domain before switching to next, while growing 708 * the pool of contiguous physical pages. 709 * 710 * Do not even start launder a domain which cannot contain 711 * the specified address range, as indicated by segments 712 * constituting the domain. 713 */ 714 again: 715 if (inactl < inactmax) { 716 if (vm_phys_domain_intersects(vm_dom[dom].vmd_segs, 717 low, high) && 718 vm_pageout_launder(&vm_dom[dom].vmd_pagequeues[PQ_INACTIVE], 719 tries, low, high)) { 720 inactl++; 721 goto again; 722 } 723 if (++dom == vm_ndomains) 724 dom = 0; 725 if (dom != initial_dom) 726 goto again; 727 } 728 if (actl < actmax) { 729 if (vm_phys_domain_intersects(vm_dom[dom].vmd_segs, 730 low, high) && 731 vm_pageout_launder(&vm_dom[dom].vmd_pagequeues[PQ_ACTIVE], 732 tries, low, high)) { 733 actl++; 734 goto again; 735 } 736 if (++dom == vm_ndomains) 737 dom = 0; 738 if (dom != initial_dom) 739 goto again; 740 } 741 } 742 743 #if !defined(NO_SWAPPING) 744 /* 745 * vm_pageout_object_deactivate_pages 746 * 747 * Deactivate enough pages to satisfy the inactive target 748 * requirements. 749 * 750 * The object and map must be locked. 751 */ 752 static void 753 vm_pageout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object, 754 long desired) 755 { 756 vm_object_t backing_object, object; 757 vm_page_t p; 758 int act_delta, remove_mode; 759 760 VM_OBJECT_ASSERT_LOCKED(first_object); 761 if ((first_object->flags & OBJ_FICTITIOUS) != 0) 762 return; 763 for (object = first_object;; object = backing_object) { 764 if (pmap_resident_count(pmap) <= desired) 765 goto unlock_return; 766 VM_OBJECT_ASSERT_LOCKED(object); 767 if ((object->flags & OBJ_UNMANAGED) != 0 || 768 object->paging_in_progress != 0) 769 goto unlock_return; 770 771 remove_mode = 0; 772 if (object->shadow_count > 1) 773 remove_mode = 1; 774 /* 775 * Scan the object's entire memory queue. 776 */ 777 TAILQ_FOREACH(p, &object->memq, listq) { 778 if (pmap_resident_count(pmap) <= desired) 779 goto unlock_return; 780 if (vm_page_busied(p)) 781 continue; 782 PCPU_INC(cnt.v_pdpages); 783 vm_page_lock(p); 784 if (p->wire_count != 0 || p->hold_count != 0 || 785 !pmap_page_exists_quick(pmap, p)) { 786 vm_page_unlock(p); 787 continue; 788 } 789 act_delta = pmap_ts_referenced(p); 790 if ((p->aflags & PGA_REFERENCED) != 0) { 791 if (act_delta == 0) 792 act_delta = 1; 793 vm_page_aflag_clear(p, PGA_REFERENCED); 794 } 795 if (p->queue != PQ_ACTIVE && act_delta != 0) { 796 vm_page_activate(p); 797 p->act_count += act_delta; 798 } else if (p->queue == PQ_ACTIVE) { 799 if (act_delta == 0) { 800 p->act_count -= min(p->act_count, 801 ACT_DECLINE); 802 if (!remove_mode && p->act_count == 0) { 803 pmap_remove_all(p); 804 vm_page_deactivate(p); 805 } else 806 vm_page_requeue(p); 807 } else { 808 vm_page_activate(p); 809 if (p->act_count < ACT_MAX - 810 ACT_ADVANCE) 811 p->act_count += ACT_ADVANCE; 812 vm_page_requeue(p); 813 } 814 } else if (p->queue == PQ_INACTIVE) 815 pmap_remove_all(p); 816 vm_page_unlock(p); 817 } 818 if ((backing_object = object->backing_object) == NULL) 819 goto unlock_return; 820 VM_OBJECT_RLOCK(backing_object); 821 if (object != first_object) 822 VM_OBJECT_RUNLOCK(object); 823 } 824 unlock_return: 825 if (object != first_object) 826 VM_OBJECT_RUNLOCK(object); 827 } 828 829 /* 830 * deactivate some number of pages in a map, try to do it fairly, but 831 * that is really hard to do. 832 */ 833 static void 834 vm_pageout_map_deactivate_pages(map, desired) 835 vm_map_t map; 836 long desired; 837 { 838 vm_map_entry_t tmpe; 839 vm_object_t obj, bigobj; 840 int nothingwired; 841 842 if (!vm_map_trylock(map)) 843 return; 844 845 bigobj = NULL; 846 nothingwired = TRUE; 847 848 /* 849 * first, search out the biggest object, and try to free pages from 850 * that. 851 */ 852 tmpe = map->header.next; 853 while (tmpe != &map->header) { 854 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 855 obj = tmpe->object.vm_object; 856 if (obj != NULL && VM_OBJECT_TRYRLOCK(obj)) { 857 if (obj->shadow_count <= 1 && 858 (bigobj == NULL || 859 bigobj->resident_page_count < obj->resident_page_count)) { 860 if (bigobj != NULL) 861 VM_OBJECT_RUNLOCK(bigobj); 862 bigobj = obj; 863 } else 864 VM_OBJECT_RUNLOCK(obj); 865 } 866 } 867 if (tmpe->wired_count > 0) 868 nothingwired = FALSE; 869 tmpe = tmpe->next; 870 } 871 872 if (bigobj != NULL) { 873 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired); 874 VM_OBJECT_RUNLOCK(bigobj); 875 } 876 /* 877 * Next, hunt around for other pages to deactivate. We actually 878 * do this search sort of wrong -- .text first is not the best idea. 879 */ 880 tmpe = map->header.next; 881 while (tmpe != &map->header) { 882 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 883 break; 884 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 885 obj = tmpe->object.vm_object; 886 if (obj != NULL) { 887 VM_OBJECT_RLOCK(obj); 888 vm_pageout_object_deactivate_pages(map->pmap, obj, desired); 889 VM_OBJECT_RUNLOCK(obj); 890 } 891 } 892 tmpe = tmpe->next; 893 } 894 895 /* 896 * Remove all mappings if a process is swapped out, this will free page 897 * table pages. 898 */ 899 if (desired == 0 && nothingwired) { 900 pmap_remove(vm_map_pmap(map), vm_map_min(map), 901 vm_map_max(map)); 902 } 903 904 vm_map_unlock(map); 905 } 906 #endif /* !defined(NO_SWAPPING) */ 907 908 /* 909 * Attempt to acquire all of the necessary locks to launder a page and 910 * then call through the clustering layer to PUTPAGES. Wait a short 911 * time for a vnode lock. 912 * 913 * Requires the page and object lock on entry, releases both before return. 914 * Returns 0 on success and an errno otherwise. 915 */ 916 static int 917 vm_pageout_clean(vm_page_t m) 918 { 919 struct vnode *vp; 920 struct mount *mp; 921 vm_object_t object; 922 vm_pindex_t pindex; 923 int error, lockmode; 924 925 vm_page_assert_locked(m); 926 object = m->object; 927 VM_OBJECT_ASSERT_WLOCKED(object); 928 error = 0; 929 vp = NULL; 930 mp = NULL; 931 932 /* 933 * The object is already known NOT to be dead. It 934 * is possible for the vget() to block the whole 935 * pageout daemon, but the new low-memory handling 936 * code should prevent it. 937 * 938 * We can't wait forever for the vnode lock, we might 939 * deadlock due to a vn_read() getting stuck in 940 * vm_wait while holding this vnode. We skip the 941 * vnode if we can't get it in a reasonable amount 942 * of time. 943 */ 944 if (object->type == OBJT_VNODE) { 945 vm_page_unlock(m); 946 vp = object->handle; 947 if (vp->v_type == VREG && 948 vn_start_write(vp, &mp, V_NOWAIT) != 0) { 949 mp = NULL; 950 error = EDEADLK; 951 goto unlock_all; 952 } 953 KASSERT(mp != NULL, 954 ("vp %p with NULL v_mount", vp)); 955 vm_object_reference_locked(object); 956 pindex = m->pindex; 957 VM_OBJECT_WUNLOCK(object); 958 lockmode = MNT_SHARED_WRITES(vp->v_mount) ? 959 LK_SHARED : LK_EXCLUSIVE; 960 if (vget(vp, lockmode | LK_TIMELOCK, curthread)) { 961 vp = NULL; 962 error = EDEADLK; 963 goto unlock_mp; 964 } 965 VM_OBJECT_WLOCK(object); 966 vm_page_lock(m); 967 /* 968 * While the object and page were unlocked, the page 969 * may have been: 970 * (1) moved to a different queue, 971 * (2) reallocated to a different object, 972 * (3) reallocated to a different offset, or 973 * (4) cleaned. 974 */ 975 if (m->queue != PQ_INACTIVE || m->object != object || 976 m->pindex != pindex || m->dirty == 0) { 977 vm_page_unlock(m); 978 error = ENXIO; 979 goto unlock_all; 980 } 981 982 /* 983 * The page may have been busied or held while the object 984 * and page locks were released. 985 */ 986 if (vm_page_busied(m) || m->hold_count != 0) { 987 vm_page_unlock(m); 988 error = EBUSY; 989 goto unlock_all; 990 } 991 } 992 993 /* 994 * If a page is dirty, then it is either being washed 995 * (but not yet cleaned) or it is still in the 996 * laundry. If it is still in the laundry, then we 997 * start the cleaning operation. 998 */ 999 if (vm_pageout_cluster(m) == 0) 1000 error = EIO; 1001 1002 unlock_all: 1003 VM_OBJECT_WUNLOCK(object); 1004 1005 unlock_mp: 1006 vm_page_lock_assert(m, MA_NOTOWNED); 1007 if (mp != NULL) { 1008 if (vp != NULL) 1009 vput(vp); 1010 vm_object_deallocate(object); 1011 vn_finished_write(mp); 1012 } 1013 1014 return (error); 1015 } 1016 1017 /* 1018 * vm_pageout_scan does the dirty work for the pageout daemon. 1019 * 1020 * pass 0 - Update active LRU/deactivate pages 1021 * pass 1 - Move inactive to cache or free 1022 * pass 2 - Launder dirty pages 1023 */ 1024 static void 1025 vm_pageout_scan(struct vm_domain *vmd, int pass) 1026 { 1027 vm_page_t m, next; 1028 struct vm_pagequeue *pq; 1029 vm_object_t object; 1030 long min_scan; 1031 int act_delta, addl_page_shortage, deficit, error, maxlaunder, maxscan; 1032 int page_shortage, scan_tick, scanned, vnodes_skipped; 1033 boolean_t pageout_ok, queues_locked; 1034 1035 /* 1036 * If we need to reclaim memory ask kernel caches to return 1037 * some. We rate limit to avoid thrashing. 1038 */ 1039 if (vmd == &vm_dom[0] && pass > 0 && 1040 (time_uptime - lowmem_uptime) >= lowmem_period) { 1041 /* 1042 * Decrease registered cache sizes. 1043 */ 1044 SDT_PROBE0(vm, , , vm__lowmem_scan); 1045 EVENTHANDLER_INVOKE(vm_lowmem, 0); 1046 /* 1047 * We do this explicitly after the caches have been 1048 * drained above. 1049 */ 1050 uma_reclaim(); 1051 lowmem_uptime = time_uptime; 1052 } 1053 1054 /* 1055 * The addl_page_shortage is the number of temporarily 1056 * stuck pages in the inactive queue. In other words, the 1057 * number of pages from the inactive count that should be 1058 * discounted in setting the target for the active queue scan. 1059 */ 1060 addl_page_shortage = 0; 1061 1062 /* 1063 * Calculate the number of pages we want to either free or move 1064 * to the cache. 1065 */ 1066 if (pass > 0) { 1067 deficit = atomic_readandclear_int(&vm_pageout_deficit); 1068 page_shortage = vm_paging_target() + deficit; 1069 } else 1070 page_shortage = deficit = 0; 1071 1072 /* 1073 * maxlaunder limits the number of dirty pages we flush per scan. 1074 * For most systems a smaller value (16 or 32) is more robust under 1075 * extreme memory and disk pressure because any unnecessary writes 1076 * to disk can result in extreme performance degredation. However, 1077 * systems with excessive dirty pages (especially when MAP_NOSYNC is 1078 * used) will die horribly with limited laundering. If the pageout 1079 * daemon cannot clean enough pages in the first pass, we let it go 1080 * all out in succeeding passes. 1081 */ 1082 if ((maxlaunder = vm_max_launder) <= 1) 1083 maxlaunder = 1; 1084 if (pass > 1) 1085 maxlaunder = 10000; 1086 1087 vnodes_skipped = 0; 1088 1089 /* 1090 * Start scanning the inactive queue for pages we can move to the 1091 * cache or free. The scan will stop when the target is reached or 1092 * we have scanned the entire inactive queue. Note that m->act_count 1093 * is not used to form decisions for the inactive queue, only for the 1094 * active queue. 1095 */ 1096 pq = &vmd->vmd_pagequeues[PQ_INACTIVE]; 1097 maxscan = pq->pq_cnt; 1098 vm_pagequeue_lock(pq); 1099 queues_locked = TRUE; 1100 for (m = TAILQ_FIRST(&pq->pq_pl); 1101 m != NULL && maxscan-- > 0 && page_shortage > 0; 1102 m = next) { 1103 vm_pagequeue_assert_locked(pq); 1104 KASSERT(queues_locked, ("unlocked queues")); 1105 KASSERT(m->queue == PQ_INACTIVE, ("Inactive queue %p", m)); 1106 1107 PCPU_INC(cnt.v_pdpages); 1108 next = TAILQ_NEXT(m, plinks.q); 1109 1110 /* 1111 * skip marker pages 1112 */ 1113 if (m->flags & PG_MARKER) 1114 continue; 1115 1116 KASSERT((m->flags & PG_FICTITIOUS) == 0, 1117 ("Fictitious page %p cannot be in inactive queue", m)); 1118 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 1119 ("Unmanaged page %p cannot be in inactive queue", m)); 1120 1121 /* 1122 * The page or object lock acquisitions fail if the 1123 * page was removed from the queue or moved to a 1124 * different position within the queue. In either 1125 * case, addl_page_shortage should not be incremented. 1126 */ 1127 if (!vm_pageout_page_lock(m, &next)) 1128 goto unlock_page; 1129 else if (m->hold_count != 0) { 1130 /* 1131 * Held pages are essentially stuck in the 1132 * queue. So, they ought to be discounted 1133 * from the inactive count. See the 1134 * calculation of the page_shortage for the 1135 * loop over the active queue below. 1136 */ 1137 addl_page_shortage++; 1138 goto unlock_page; 1139 } 1140 object = m->object; 1141 if (!VM_OBJECT_TRYWLOCK(object)) { 1142 if (!vm_pageout_fallback_object_lock(m, &next)) 1143 goto unlock_object; 1144 else if (m->hold_count != 0) { 1145 addl_page_shortage++; 1146 goto unlock_object; 1147 } 1148 } 1149 if (vm_page_busied(m)) { 1150 /* 1151 * Don't mess with busy pages. Leave them at 1152 * the front of the queue. Most likely, they 1153 * are being paged out and will leave the 1154 * queue shortly after the scan finishes. So, 1155 * they ought to be discounted from the 1156 * inactive count. 1157 */ 1158 addl_page_shortage++; 1159 unlock_object: 1160 VM_OBJECT_WUNLOCK(object); 1161 unlock_page: 1162 vm_page_unlock(m); 1163 continue; 1164 } 1165 KASSERT(m->hold_count == 0, ("Held page %p", m)); 1166 1167 /* 1168 * We unlock the inactive page queue, invalidating the 1169 * 'next' pointer. Use our marker to remember our 1170 * place. 1171 */ 1172 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &vmd->vmd_marker, plinks.q); 1173 vm_pagequeue_unlock(pq); 1174 queues_locked = FALSE; 1175 1176 /* 1177 * Invalid pages can be easily freed. They cannot be 1178 * mapped, vm_page_free() asserts this. 1179 */ 1180 if (m->valid == 0) 1181 goto free_page; 1182 1183 /* 1184 * If the page has been referenced and the object is not dead, 1185 * reactivate or requeue the page depending on whether the 1186 * object is mapped. 1187 */ 1188 if ((m->aflags & PGA_REFERENCED) != 0) { 1189 vm_page_aflag_clear(m, PGA_REFERENCED); 1190 act_delta = 1; 1191 } else 1192 act_delta = 0; 1193 if (object->ref_count != 0) { 1194 act_delta += pmap_ts_referenced(m); 1195 } else { 1196 KASSERT(!pmap_page_is_mapped(m), 1197 ("vm_pageout_scan: page %p is mapped", m)); 1198 } 1199 if (act_delta != 0) { 1200 if (object->ref_count != 0) { 1201 vm_page_activate(m); 1202 1203 /* 1204 * Increase the activation count if the page 1205 * was referenced while in the inactive queue. 1206 * This makes it less likely that the page will 1207 * be returned prematurely to the inactive 1208 * queue. 1209 */ 1210 m->act_count += act_delta + ACT_ADVANCE; 1211 goto drop_page; 1212 } else if ((object->flags & OBJ_DEAD) == 0) 1213 goto requeue_page; 1214 } 1215 1216 /* 1217 * If the page appears to be clean at the machine-independent 1218 * layer, then remove all of its mappings from the pmap in 1219 * anticipation of placing it onto the cache queue. If, 1220 * however, any of the page's mappings allow write access, 1221 * then the page may still be modified until the last of those 1222 * mappings are removed. 1223 */ 1224 if (object->ref_count != 0) { 1225 vm_page_test_dirty(m); 1226 if (m->dirty == 0) 1227 pmap_remove_all(m); 1228 } 1229 1230 if (m->dirty == 0) { 1231 /* 1232 * Clean pages can be freed. 1233 */ 1234 free_page: 1235 vm_page_free(m); 1236 PCPU_INC(cnt.v_dfree); 1237 --page_shortage; 1238 } else if ((object->flags & OBJ_DEAD) != 0) { 1239 /* 1240 * Leave dirty pages from dead objects at the front of 1241 * the queue. They are being paged out and freed by 1242 * the thread that destroyed the object. They will 1243 * leave the queue shortly after the scan finishes, so 1244 * they should be discounted from the inactive count. 1245 */ 1246 addl_page_shortage++; 1247 } else if ((m->flags & PG_WINATCFLS) == 0 && pass < 2) { 1248 /* 1249 * Dirty pages need to be paged out, but flushing 1250 * a page is extremely expensive versus freeing 1251 * a clean page. Rather then artificially limiting 1252 * the number of pages we can flush, we instead give 1253 * dirty pages extra priority on the inactive queue 1254 * by forcing them to be cycled through the queue 1255 * twice before being flushed, after which the 1256 * (now clean) page will cycle through once more 1257 * before being freed. This significantly extends 1258 * the thrash point for a heavily loaded machine. 1259 */ 1260 m->flags |= PG_WINATCFLS; 1261 requeue_page: 1262 vm_pagequeue_lock(pq); 1263 queues_locked = TRUE; 1264 vm_page_requeue_locked(m); 1265 } else if (maxlaunder > 0) { 1266 /* 1267 * We always want to try to flush some dirty pages if 1268 * we encounter them, to keep the system stable. 1269 * Normally this number is small, but under extreme 1270 * pressure where there are insufficient clean pages 1271 * on the inactive queue, we may have to go all out. 1272 */ 1273 1274 if (object->type != OBJT_SWAP && 1275 object->type != OBJT_DEFAULT) 1276 pageout_ok = TRUE; 1277 else if (disable_swap_pageouts) 1278 pageout_ok = FALSE; 1279 else if (defer_swap_pageouts) 1280 pageout_ok = vm_page_count_min(); 1281 else 1282 pageout_ok = TRUE; 1283 if (!pageout_ok) 1284 goto requeue_page; 1285 error = vm_pageout_clean(m); 1286 /* 1287 * Decrement page_shortage on success to account for 1288 * the (future) cleaned page. Otherwise we could wind 1289 * up laundering or cleaning too many pages. 1290 */ 1291 if (error == 0) { 1292 page_shortage--; 1293 maxlaunder--; 1294 } else if (error == EDEADLK) { 1295 pageout_lock_miss++; 1296 vnodes_skipped++; 1297 } else if (error == EBUSY) { 1298 addl_page_shortage++; 1299 } 1300 vm_page_lock_assert(m, MA_NOTOWNED); 1301 goto relock_queues; 1302 } 1303 drop_page: 1304 vm_page_unlock(m); 1305 VM_OBJECT_WUNLOCK(object); 1306 relock_queues: 1307 if (!queues_locked) { 1308 vm_pagequeue_lock(pq); 1309 queues_locked = TRUE; 1310 } 1311 next = TAILQ_NEXT(&vmd->vmd_marker, plinks.q); 1312 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_marker, plinks.q); 1313 } 1314 vm_pagequeue_unlock(pq); 1315 1316 #if !defined(NO_SWAPPING) 1317 /* 1318 * Wakeup the swapout daemon if we didn't cache or free the targeted 1319 * number of pages. 1320 */ 1321 if (vm_swap_enabled && page_shortage > 0) 1322 vm_req_vmdaemon(VM_SWAP_NORMAL); 1323 #endif 1324 1325 /* 1326 * Wakeup the sync daemon if we skipped a vnode in a writeable object 1327 * and we didn't cache or free enough pages. 1328 */ 1329 if (vnodes_skipped > 0 && page_shortage > vm_cnt.v_free_target - 1330 vm_cnt.v_free_min) 1331 (void)speedup_syncer(); 1332 1333 /* 1334 * Compute the number of pages we want to try to move from the 1335 * active queue to the inactive queue. 1336 */ 1337 page_shortage = vm_cnt.v_inactive_target - vm_cnt.v_inactive_count + 1338 vm_paging_target() + deficit + addl_page_shortage; 1339 1340 pq = &vmd->vmd_pagequeues[PQ_ACTIVE]; 1341 vm_pagequeue_lock(pq); 1342 maxscan = pq->pq_cnt; 1343 1344 /* 1345 * If we're just idle polling attempt to visit every 1346 * active page within 'update_period' seconds. 1347 */ 1348 scan_tick = ticks; 1349 if (vm_pageout_update_period != 0) { 1350 min_scan = pq->pq_cnt; 1351 min_scan *= scan_tick - vmd->vmd_last_active_scan; 1352 min_scan /= hz * vm_pageout_update_period; 1353 } else 1354 min_scan = 0; 1355 if (min_scan > 0 || (page_shortage > 0 && maxscan > 0)) 1356 vmd->vmd_last_active_scan = scan_tick; 1357 1358 /* 1359 * Scan the active queue for pages that can be deactivated. Update 1360 * the per-page activity counter and use it to identify deactivation 1361 * candidates. 1362 */ 1363 for (m = TAILQ_FIRST(&pq->pq_pl), scanned = 0; m != NULL && (scanned < 1364 min_scan || (page_shortage > 0 && scanned < maxscan)); m = next, 1365 scanned++) { 1366 1367 KASSERT(m->queue == PQ_ACTIVE, 1368 ("vm_pageout_scan: page %p isn't active", m)); 1369 1370 next = TAILQ_NEXT(m, plinks.q); 1371 if ((m->flags & PG_MARKER) != 0) 1372 continue; 1373 KASSERT((m->flags & PG_FICTITIOUS) == 0, 1374 ("Fictitious page %p cannot be in active queue", m)); 1375 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 1376 ("Unmanaged page %p cannot be in active queue", m)); 1377 if (!vm_pageout_page_lock(m, &next)) { 1378 vm_page_unlock(m); 1379 continue; 1380 } 1381 1382 /* 1383 * The count for pagedaemon pages is done after checking the 1384 * page for eligibility... 1385 */ 1386 PCPU_INC(cnt.v_pdpages); 1387 1388 /* 1389 * Check to see "how much" the page has been used. 1390 */ 1391 if ((m->aflags & PGA_REFERENCED) != 0) { 1392 vm_page_aflag_clear(m, PGA_REFERENCED); 1393 act_delta = 1; 1394 } else 1395 act_delta = 0; 1396 1397 /* 1398 * Unlocked object ref count check. Two races are possible. 1399 * 1) The ref was transitioning to zero and we saw non-zero, 1400 * the pmap bits will be checked unnecessarily. 1401 * 2) The ref was transitioning to one and we saw zero. 1402 * The page lock prevents a new reference to this page so 1403 * we need not check the reference bits. 1404 */ 1405 if (m->object->ref_count != 0) 1406 act_delta += pmap_ts_referenced(m); 1407 1408 /* 1409 * Advance or decay the act_count based on recent usage. 1410 */ 1411 if (act_delta != 0) { 1412 m->act_count += ACT_ADVANCE + act_delta; 1413 if (m->act_count > ACT_MAX) 1414 m->act_count = ACT_MAX; 1415 } else 1416 m->act_count -= min(m->act_count, ACT_DECLINE); 1417 1418 /* 1419 * Move this page to the tail of the active or inactive 1420 * queue depending on usage. 1421 */ 1422 if (m->act_count == 0) { 1423 /* Dequeue to avoid later lock recursion. */ 1424 vm_page_dequeue_locked(m); 1425 vm_page_deactivate(m); 1426 page_shortage--; 1427 } else 1428 vm_page_requeue_locked(m); 1429 vm_page_unlock(m); 1430 } 1431 vm_pagequeue_unlock(pq); 1432 #if !defined(NO_SWAPPING) 1433 /* 1434 * Idle process swapout -- run once per second. 1435 */ 1436 if (vm_swap_idle_enabled) { 1437 static long lsec; 1438 if (time_second != lsec) { 1439 vm_req_vmdaemon(VM_SWAP_IDLE); 1440 lsec = time_second; 1441 } 1442 } 1443 #endif 1444 1445 /* 1446 * If we are critically low on one of RAM or swap and low on 1447 * the other, kill the largest process. However, we avoid 1448 * doing this on the first pass in order to give ourselves a 1449 * chance to flush out dirty vnode-backed pages and to allow 1450 * active pages to be moved to the inactive queue and reclaimed. 1451 */ 1452 vm_pageout_mightbe_oom(vmd, pass); 1453 } 1454 1455 static int vm_pageout_oom_vote; 1456 1457 /* 1458 * The pagedaemon threads randlomly select one to perform the 1459 * OOM. Trying to kill processes before all pagedaemons 1460 * failed to reach free target is premature. 1461 */ 1462 static void 1463 vm_pageout_mightbe_oom(struct vm_domain *vmd, int pass) 1464 { 1465 int old_vote; 1466 1467 if (pass <= 1 || !((swap_pager_avail < 64 && vm_page_count_min()) || 1468 (swap_pager_full && vm_paging_target() > 0))) { 1469 if (vmd->vmd_oom) { 1470 vmd->vmd_oom = FALSE; 1471 atomic_subtract_int(&vm_pageout_oom_vote, 1); 1472 } 1473 return; 1474 } 1475 1476 if (vmd->vmd_oom) 1477 return; 1478 1479 vmd->vmd_oom = TRUE; 1480 old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1); 1481 if (old_vote != vm_ndomains - 1) 1482 return; 1483 1484 /* 1485 * The current pagedaemon thread is the last in the quorum to 1486 * start OOM. Initiate the selection and signaling of the 1487 * victim. 1488 */ 1489 vm_pageout_oom(VM_OOM_MEM); 1490 1491 /* 1492 * After one round of OOM terror, recall our vote. On the 1493 * next pass, current pagedaemon would vote again if the low 1494 * memory condition is still there, due to vmd_oom being 1495 * false. 1496 */ 1497 vmd->vmd_oom = FALSE; 1498 atomic_subtract_int(&vm_pageout_oom_vote, 1); 1499 } 1500 1501 void 1502 vm_pageout_oom(int shortage) 1503 { 1504 struct proc *p, *bigproc; 1505 vm_offset_t size, bigsize; 1506 struct thread *td; 1507 struct vmspace *vm; 1508 1509 /* 1510 * We keep the process bigproc locked once we find it to keep anyone 1511 * from messing with it; however, there is a possibility of 1512 * deadlock if process B is bigproc and one of it's child processes 1513 * attempts to propagate a signal to B while we are waiting for A's 1514 * lock while walking this list. To avoid this, we don't block on 1515 * the process lock but just skip a process if it is already locked. 1516 */ 1517 bigproc = NULL; 1518 bigsize = 0; 1519 sx_slock(&allproc_lock); 1520 FOREACH_PROC_IN_SYSTEM(p) { 1521 int breakout; 1522 1523 PROC_LOCK(p); 1524 1525 /* 1526 * If this is a system, protected or killed process, skip it. 1527 */ 1528 if (p->p_state != PRS_NORMAL || (p->p_flag & (P_INEXEC | 1529 P_PROTECTED | P_SYSTEM | P_WEXIT)) != 0 || 1530 p->p_pid == 1 || P_KILLED(p) || 1531 (p->p_pid < 48 && swap_pager_avail != 0)) { 1532 PROC_UNLOCK(p); 1533 continue; 1534 } 1535 /* 1536 * If the process is in a non-running type state, 1537 * don't touch it. Check all the threads individually. 1538 */ 1539 breakout = 0; 1540 FOREACH_THREAD_IN_PROC(p, td) { 1541 thread_lock(td); 1542 if (!TD_ON_RUNQ(td) && 1543 !TD_IS_RUNNING(td) && 1544 !TD_IS_SLEEPING(td) && 1545 !TD_IS_SUSPENDED(td)) { 1546 thread_unlock(td); 1547 breakout = 1; 1548 break; 1549 } 1550 thread_unlock(td); 1551 } 1552 if (breakout) { 1553 PROC_UNLOCK(p); 1554 continue; 1555 } 1556 /* 1557 * get the process size 1558 */ 1559 vm = vmspace_acquire_ref(p); 1560 if (vm == NULL) { 1561 PROC_UNLOCK(p); 1562 continue; 1563 } 1564 _PHOLD(p); 1565 if (!vm_map_trylock_read(&vm->vm_map)) { 1566 _PRELE(p); 1567 PROC_UNLOCK(p); 1568 vmspace_free(vm); 1569 continue; 1570 } 1571 PROC_UNLOCK(p); 1572 size = vmspace_swap_count(vm); 1573 vm_map_unlock_read(&vm->vm_map); 1574 if (shortage == VM_OOM_MEM) 1575 size += vmspace_resident_count(vm); 1576 vmspace_free(vm); 1577 /* 1578 * if the this process is bigger than the biggest one 1579 * remember it. 1580 */ 1581 if (size > bigsize) { 1582 if (bigproc != NULL) 1583 PRELE(bigproc); 1584 bigproc = p; 1585 bigsize = size; 1586 } else { 1587 PRELE(p); 1588 } 1589 } 1590 sx_sunlock(&allproc_lock); 1591 if (bigproc != NULL) { 1592 if (vm_panic_on_oom != 0) 1593 panic("out of swap space"); 1594 PROC_LOCK(bigproc); 1595 killproc(bigproc, "out of swap space"); 1596 sched_nice(bigproc, PRIO_MIN); 1597 _PRELE(bigproc); 1598 PROC_UNLOCK(bigproc); 1599 wakeup(&vm_cnt.v_free_count); 1600 } 1601 } 1602 1603 static void 1604 vm_pageout_worker(void *arg) 1605 { 1606 struct vm_domain *domain; 1607 int domidx; 1608 1609 domidx = (uintptr_t)arg; 1610 domain = &vm_dom[domidx]; 1611 1612 /* 1613 * XXXKIB It could be useful to bind pageout daemon threads to 1614 * the cores belonging to the domain, from which vm_page_array 1615 * is allocated. 1616 */ 1617 1618 KASSERT(domain->vmd_segs != 0, ("domain without segments")); 1619 domain->vmd_last_active_scan = ticks; 1620 vm_pageout_init_marker(&domain->vmd_marker, PQ_INACTIVE); 1621 1622 /* 1623 * The pageout daemon worker is never done, so loop forever. 1624 */ 1625 while (TRUE) { 1626 /* 1627 * If we have enough free memory, wakeup waiters. Do 1628 * not clear vm_pages_needed until we reach our target, 1629 * otherwise we may be woken up over and over again and 1630 * waste a lot of cpu. 1631 */ 1632 mtx_lock(&vm_page_queue_free_mtx); 1633 if (vm_pages_needed && !vm_page_count_min()) { 1634 if (!vm_paging_needed()) 1635 vm_pages_needed = 0; 1636 wakeup(&vm_cnt.v_free_count); 1637 } 1638 if (vm_pages_needed) { 1639 /* 1640 * We're still not done. Either vm_pages_needed was 1641 * set by another thread during the previous scan 1642 * (typically, this happens during a level 0 scan) or 1643 * vm_pages_needed was already set and the scan failed 1644 * to free enough pages. If we haven't yet performed 1645 * a level >= 2 scan (unlimited dirty cleaning), then 1646 * upgrade the level and scan again now. Otherwise, 1647 * sleep a bit and try again later. While sleeping, 1648 * vm_pages_needed can be cleared. 1649 */ 1650 if (domain->vmd_pass > 1) 1651 msleep(&vm_pages_needed, 1652 &vm_page_queue_free_mtx, PVM, "psleep", 1653 hz / 2); 1654 } else { 1655 /* 1656 * Good enough, sleep until required to refresh 1657 * stats. 1658 */ 1659 msleep(&vm_pages_needed, &vm_page_queue_free_mtx, 1660 PVM, "psleep", hz); 1661 } 1662 if (vm_pages_needed) { 1663 vm_cnt.v_pdwakeups++; 1664 domain->vmd_pass++; 1665 } else 1666 domain->vmd_pass = 0; 1667 mtx_unlock(&vm_page_queue_free_mtx); 1668 vm_pageout_scan(domain, domain->vmd_pass); 1669 } 1670 } 1671 1672 /* 1673 * vm_pageout_init initialises basic pageout daemon settings. 1674 */ 1675 static void 1676 vm_pageout_init(void) 1677 { 1678 /* 1679 * Initialize some paging parameters. 1680 */ 1681 vm_cnt.v_interrupt_free_min = 2; 1682 if (vm_cnt.v_page_count < 2000) 1683 vm_pageout_page_count = 8; 1684 1685 /* 1686 * v_free_reserved needs to include enough for the largest 1687 * swap pager structures plus enough for any pv_entry structs 1688 * when paging. 1689 */ 1690 if (vm_cnt.v_page_count > 1024) 1691 vm_cnt.v_free_min = 4 + (vm_cnt.v_page_count - 1024) / 200; 1692 else 1693 vm_cnt.v_free_min = 4; 1694 vm_cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1695 vm_cnt.v_interrupt_free_min; 1696 vm_cnt.v_free_reserved = vm_pageout_page_count + 1697 vm_cnt.v_pageout_free_min + (vm_cnt.v_page_count / 768); 1698 vm_cnt.v_free_severe = vm_cnt.v_free_min / 2; 1699 vm_cnt.v_free_target = 4 * vm_cnt.v_free_min + vm_cnt.v_free_reserved; 1700 vm_cnt.v_free_min += vm_cnt.v_free_reserved; 1701 vm_cnt.v_free_severe += vm_cnt.v_free_reserved; 1702 vm_cnt.v_inactive_target = (3 * vm_cnt.v_free_target) / 2; 1703 if (vm_cnt.v_inactive_target > vm_cnt.v_free_count / 3) 1704 vm_cnt.v_inactive_target = vm_cnt.v_free_count / 3; 1705 1706 /* 1707 * Set the default wakeup threshold to be 10% above the minimum 1708 * page limit. This keeps the steady state out of shortfall. 1709 */ 1710 vm_pageout_wakeup_thresh = (vm_cnt.v_free_min / 10) * 11; 1711 1712 /* 1713 * Set interval in seconds for active scan. We want to visit each 1714 * page at least once every ten minutes. This is to prevent worst 1715 * case paging behaviors with stale active LRU. 1716 */ 1717 if (vm_pageout_update_period == 0) 1718 vm_pageout_update_period = 600; 1719 1720 /* XXX does not really belong here */ 1721 if (vm_page_max_wired == 0) 1722 vm_page_max_wired = vm_cnt.v_free_count / 3; 1723 } 1724 1725 /* 1726 * vm_pageout is the high level pageout daemon. 1727 */ 1728 static void 1729 vm_pageout(void) 1730 { 1731 int error; 1732 #if MAXMEMDOM > 1 1733 int i; 1734 #endif 1735 1736 swap_pager_swap_init(); 1737 #if MAXMEMDOM > 1 1738 for (i = 1; i < vm_ndomains; i++) { 1739 error = kthread_add(vm_pageout_worker, (void *)(uintptr_t)i, 1740 curproc, NULL, 0, 0, "dom%d", i); 1741 if (error != 0) { 1742 panic("starting pageout for domain %d, error %d\n", 1743 i, error); 1744 } 1745 } 1746 #endif 1747 error = kthread_add(uma_reclaim_worker, NULL, curproc, NULL, 1748 0, 0, "uma"); 1749 if (error != 0) 1750 panic("starting uma_reclaim helper, error %d\n", error); 1751 vm_pageout_worker((void *)(uintptr_t)0); 1752 } 1753 1754 /* 1755 * Unless the free page queue lock is held by the caller, this function 1756 * should be regarded as advisory. Specifically, the caller should 1757 * not msleep() on &vm_cnt.v_free_count following this function unless 1758 * the free page queue lock is held until the msleep() is performed. 1759 */ 1760 void 1761 pagedaemon_wakeup(void) 1762 { 1763 1764 if (!vm_pages_needed && curthread->td_proc != pageproc) { 1765 vm_pages_needed = 1; 1766 wakeup(&vm_pages_needed); 1767 } 1768 } 1769 1770 #if !defined(NO_SWAPPING) 1771 static void 1772 vm_req_vmdaemon(int req) 1773 { 1774 static int lastrun = 0; 1775 1776 mtx_lock(&vm_daemon_mtx); 1777 vm_pageout_req_swapout |= req; 1778 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1779 wakeup(&vm_daemon_needed); 1780 lastrun = ticks; 1781 } 1782 mtx_unlock(&vm_daemon_mtx); 1783 } 1784 1785 static void 1786 vm_daemon(void) 1787 { 1788 struct rlimit rsslim; 1789 struct proc *p; 1790 struct thread *td; 1791 struct vmspace *vm; 1792 int breakout, swapout_flags, tryagain, attempts; 1793 #ifdef RACCT 1794 uint64_t rsize, ravailable; 1795 #endif 1796 1797 while (TRUE) { 1798 mtx_lock(&vm_daemon_mtx); 1799 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 1800 #ifdef RACCT 1801 racct_enable ? hz : 0 1802 #else 1803 0 1804 #endif 1805 ); 1806 swapout_flags = vm_pageout_req_swapout; 1807 vm_pageout_req_swapout = 0; 1808 mtx_unlock(&vm_daemon_mtx); 1809 if (swapout_flags) 1810 swapout_procs(swapout_flags); 1811 1812 /* 1813 * scan the processes for exceeding their rlimits or if 1814 * process is swapped out -- deactivate pages 1815 */ 1816 tryagain = 0; 1817 attempts = 0; 1818 again: 1819 attempts++; 1820 sx_slock(&allproc_lock); 1821 FOREACH_PROC_IN_SYSTEM(p) { 1822 vm_pindex_t limit, size; 1823 1824 /* 1825 * if this is a system process or if we have already 1826 * looked at this process, skip it. 1827 */ 1828 PROC_LOCK(p); 1829 if (p->p_state != PRS_NORMAL || 1830 p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) { 1831 PROC_UNLOCK(p); 1832 continue; 1833 } 1834 /* 1835 * if the process is in a non-running type state, 1836 * don't touch it. 1837 */ 1838 breakout = 0; 1839 FOREACH_THREAD_IN_PROC(p, td) { 1840 thread_lock(td); 1841 if (!TD_ON_RUNQ(td) && 1842 !TD_IS_RUNNING(td) && 1843 !TD_IS_SLEEPING(td) && 1844 !TD_IS_SUSPENDED(td)) { 1845 thread_unlock(td); 1846 breakout = 1; 1847 break; 1848 } 1849 thread_unlock(td); 1850 } 1851 if (breakout) { 1852 PROC_UNLOCK(p); 1853 continue; 1854 } 1855 /* 1856 * get a limit 1857 */ 1858 lim_rlimit_proc(p, RLIMIT_RSS, &rsslim); 1859 limit = OFF_TO_IDX( 1860 qmin(rsslim.rlim_cur, rsslim.rlim_max)); 1861 1862 /* 1863 * let processes that are swapped out really be 1864 * swapped out set the limit to nothing (will force a 1865 * swap-out.) 1866 */ 1867 if ((p->p_flag & P_INMEM) == 0) 1868 limit = 0; /* XXX */ 1869 vm = vmspace_acquire_ref(p); 1870 PROC_UNLOCK(p); 1871 if (vm == NULL) 1872 continue; 1873 1874 size = vmspace_resident_count(vm); 1875 if (size >= limit) { 1876 vm_pageout_map_deactivate_pages( 1877 &vm->vm_map, limit); 1878 } 1879 #ifdef RACCT 1880 if (racct_enable) { 1881 rsize = IDX_TO_OFF(size); 1882 PROC_LOCK(p); 1883 racct_set(p, RACCT_RSS, rsize); 1884 ravailable = racct_get_available(p, RACCT_RSS); 1885 PROC_UNLOCK(p); 1886 if (rsize > ravailable) { 1887 /* 1888 * Don't be overly aggressive; this 1889 * might be an innocent process, 1890 * and the limit could've been exceeded 1891 * by some memory hog. Don't try 1892 * to deactivate more than 1/4th 1893 * of process' resident set size. 1894 */ 1895 if (attempts <= 8) { 1896 if (ravailable < rsize - 1897 (rsize / 4)) { 1898 ravailable = rsize - 1899 (rsize / 4); 1900 } 1901 } 1902 vm_pageout_map_deactivate_pages( 1903 &vm->vm_map, 1904 OFF_TO_IDX(ravailable)); 1905 /* Update RSS usage after paging out. */ 1906 size = vmspace_resident_count(vm); 1907 rsize = IDX_TO_OFF(size); 1908 PROC_LOCK(p); 1909 racct_set(p, RACCT_RSS, rsize); 1910 PROC_UNLOCK(p); 1911 if (rsize > ravailable) 1912 tryagain = 1; 1913 } 1914 } 1915 #endif 1916 vmspace_free(vm); 1917 } 1918 sx_sunlock(&allproc_lock); 1919 if (tryagain != 0 && attempts <= 10) 1920 goto again; 1921 } 1922 } 1923 #endif /* !defined(NO_SWAPPING) */ 1924