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