1 /*- 2 * SPDX-License-Identifier: BSD-4-Clause 3 * 4 * Copyright (c) 1991 Regents of the University of California. 5 * All rights reserved. 6 * Copyright (c) 1994 John S. Dyson 7 * All rights reserved. 8 * Copyright (c) 1994 David Greenman 9 * All rights reserved. 10 * Copyright (c) 2005 Yahoo! Technologies Norway AS 11 * All rights reserved. 12 * 13 * This code is derived from software contributed to Berkeley by 14 * The Mach Operating System project at Carnegie-Mellon University. 15 * 16 * Redistribution and use in source and binary forms, with or without 17 * modification, are permitted provided that the following conditions 18 * are met: 19 * 1. Redistributions of source code must retain the above copyright 20 * notice, this list of conditions and the following disclaimer. 21 * 2. Redistributions in binary form must reproduce the above copyright 22 * notice, this list of conditions and the following disclaimer in the 23 * documentation and/or other materials provided with the distribution. 24 * 3. All advertising materials mentioning features or use of this software 25 * must display the following acknowledgement: 26 * This product includes software developed by the University of 27 * California, Berkeley and its contributors. 28 * 4. Neither the name of the University nor the names of its contributors 29 * may be used to endorse or promote products derived from this software 30 * without specific prior written permission. 31 * 32 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 33 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 34 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 35 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 36 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 37 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 38 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 39 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 40 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 41 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 42 * SUCH DAMAGE. 43 * 44 * from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91 45 * 46 * 47 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 48 * All rights reserved. 49 * 50 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 51 * 52 * Permission to use, copy, modify and distribute this software and 53 * its documentation is hereby granted, provided that both the copyright 54 * notice and this permission notice appear in all copies of the 55 * software, derivative works or modified versions, and any portions 56 * thereof, and that both notices appear in supporting documentation. 57 * 58 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 59 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 60 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 61 * 62 * Carnegie Mellon requests users of this software to return to 63 * 64 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 65 * School of Computer Science 66 * Carnegie Mellon University 67 * Pittsburgh PA 15213-3890 68 * 69 * any improvements or extensions that they make and grant Carnegie the 70 * rights to redistribute these changes. 71 */ 72 73 /* 74 * The proverbial page-out daemon. 75 */ 76 77 #include <sys/cdefs.h> 78 __FBSDID("$FreeBSD$"); 79 80 #include "opt_vm.h" 81 82 #include <sys/param.h> 83 #include <sys/systm.h> 84 #include <sys/kernel.h> 85 #include <sys/eventhandler.h> 86 #include <sys/lock.h> 87 #include <sys/mutex.h> 88 #include <sys/proc.h> 89 #include <sys/kthread.h> 90 #include <sys/ktr.h> 91 #include <sys/mount.h> 92 #include <sys/racct.h> 93 #include <sys/resourcevar.h> 94 #include <sys/sched.h> 95 #include <sys/sdt.h> 96 #include <sys/signalvar.h> 97 #include <sys/smp.h> 98 #include <sys/time.h> 99 #include <sys/vnode.h> 100 #include <sys/vmmeter.h> 101 #include <sys/rwlock.h> 102 #include <sys/sx.h> 103 #include <sys/sysctl.h> 104 105 #include <vm/vm.h> 106 #include <vm/vm_param.h> 107 #include <vm/vm_object.h> 108 #include <vm/vm_page.h> 109 #include <vm/vm_map.h> 110 #include <vm/vm_pageout.h> 111 #include <vm/vm_pager.h> 112 #include <vm/vm_phys.h> 113 #include <vm/swap_pager.h> 114 #include <vm/vm_extern.h> 115 #include <vm/uma.h> 116 117 /* 118 * System initialization 119 */ 120 121 /* the kernel process "vm_pageout"*/ 122 static void vm_pageout(void); 123 static void vm_pageout_init(void); 124 static int vm_pageout_clean(vm_page_t m, int *numpagedout); 125 static int vm_pageout_cluster(vm_page_t m); 126 static bool vm_pageout_scan(struct vm_domain *vmd, int pass); 127 static void vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage, 128 int starting_page_shortage); 129 130 SYSINIT(pagedaemon_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, vm_pageout_init, 131 NULL); 132 133 struct proc *pageproc; 134 135 static struct kproc_desc page_kp = { 136 "pagedaemon", 137 vm_pageout, 138 &pageproc 139 }; 140 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start, 141 &page_kp); 142 143 SDT_PROVIDER_DEFINE(vm); 144 SDT_PROBE_DEFINE(vm, , , vm__lowmem_scan); 145 146 /* Pagedaemon activity rates, in subdivisions of one second. */ 147 #define VM_LAUNDER_RATE 10 148 #define VM_INACT_SCAN_RATE 2 149 150 int vm_pageout_deficit; /* Estimated number of pages deficit */ 151 u_int vm_pageout_wakeup_thresh; 152 static int vm_pageout_oom_seq = 12; 153 bool vm_pageout_wanted; /* Event on which pageout daemon sleeps */ 154 bool vm_pages_needed; /* Are threads waiting for free pages? */ 155 156 /* Pending request for dirty page laundering. */ 157 static enum { 158 VM_LAUNDRY_IDLE, 159 VM_LAUNDRY_BACKGROUND, 160 VM_LAUNDRY_SHORTFALL 161 } vm_laundry_request = VM_LAUNDRY_IDLE; 162 163 static int vm_pageout_update_period; 164 static int disable_swap_pageouts; 165 static int lowmem_period = 10; 166 static time_t lowmem_uptime; 167 static int swapdev_enabled; 168 169 static int vm_panic_on_oom = 0; 170 171 SYSCTL_INT(_vm, OID_AUTO, panic_on_oom, 172 CTLFLAG_RWTUN, &vm_panic_on_oom, 0, 173 "panic on out of memory instead of killing the largest process"); 174 175 SYSCTL_INT(_vm, OID_AUTO, pageout_wakeup_thresh, 176 CTLFLAG_RWTUN, &vm_pageout_wakeup_thresh, 0, 177 "free page threshold for waking up the pageout daemon"); 178 179 SYSCTL_INT(_vm, OID_AUTO, pageout_update_period, 180 CTLFLAG_RWTUN, &vm_pageout_update_period, 0, 181 "Maximum active LRU update period"); 182 183 SYSCTL_INT(_vm, OID_AUTO, lowmem_period, CTLFLAG_RWTUN, &lowmem_period, 0, 184 "Low memory callback period"); 185 186 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts, 187 CTLFLAG_RWTUN, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages"); 188 189 static int pageout_lock_miss; 190 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss, 191 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout"); 192 193 SYSCTL_INT(_vm, OID_AUTO, pageout_oom_seq, 194 CTLFLAG_RWTUN, &vm_pageout_oom_seq, 0, 195 "back-to-back calls to oom detector to start OOM"); 196 197 static int act_scan_laundry_weight = 3; 198 SYSCTL_INT(_vm, OID_AUTO, act_scan_laundry_weight, CTLFLAG_RWTUN, 199 &act_scan_laundry_weight, 0, 200 "weight given to clean vs. dirty pages in active queue scans"); 201 202 static u_int vm_background_launder_target; 203 SYSCTL_UINT(_vm, OID_AUTO, background_launder_target, CTLFLAG_RWTUN, 204 &vm_background_launder_target, 0, 205 "background laundering target, in pages"); 206 207 static u_int vm_background_launder_rate = 4096; 208 SYSCTL_UINT(_vm, OID_AUTO, background_launder_rate, CTLFLAG_RWTUN, 209 &vm_background_launder_rate, 0, 210 "background laundering rate, in kilobytes per second"); 211 212 static u_int vm_background_launder_max = 20 * 1024; 213 SYSCTL_UINT(_vm, OID_AUTO, background_launder_max, CTLFLAG_RWTUN, 214 &vm_background_launder_max, 0, "background laundering cap, in kilobytes"); 215 216 int vm_pageout_page_count = 32; 217 218 int vm_page_max_wired; /* XXX max # of wired pages system-wide */ 219 SYSCTL_INT(_vm, OID_AUTO, max_wired, 220 CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count"); 221 222 static u_int isqrt(u_int num); 223 static boolean_t vm_pageout_fallback_object_lock(vm_page_t, vm_page_t *); 224 static int vm_pageout_launder(struct vm_domain *vmd, int launder, 225 bool in_shortfall); 226 static void vm_pageout_laundry_worker(void *arg); 227 static boolean_t vm_pageout_page_lock(vm_page_t, vm_page_t *); 228 229 /* 230 * Initialize a dummy page for marking the caller's place in the specified 231 * paging queue. In principle, this function only needs to set the flag 232 * PG_MARKER. Nonetheless, it write busies and initializes the hold count 233 * to one as safety precautions. 234 */ 235 static void 236 vm_pageout_init_marker(vm_page_t marker, u_short queue) 237 { 238 239 bzero(marker, sizeof(*marker)); 240 marker->flags = PG_MARKER; 241 marker->busy_lock = VPB_SINGLE_EXCLUSIVER; 242 marker->queue = queue; 243 marker->hold_count = 1; 244 } 245 246 /* 247 * vm_pageout_fallback_object_lock: 248 * 249 * Lock vm object currently associated with `m'. VM_OBJECT_TRYWLOCK is 250 * known to have failed and page queue must be either PQ_ACTIVE or 251 * PQ_INACTIVE. To avoid lock order violation, unlock the page queue 252 * while locking the vm object. Use marker page to detect page queue 253 * changes and maintain notion of next page on page queue. Return 254 * TRUE if no changes were detected, FALSE otherwise. vm object is 255 * locked on return. 256 * 257 * This function depends on both the lock portion of struct vm_object 258 * and normal struct vm_page being type stable. 259 */ 260 static boolean_t 261 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next) 262 { 263 struct vm_page marker; 264 struct vm_pagequeue *pq; 265 boolean_t unchanged; 266 u_short queue; 267 vm_object_t object; 268 269 queue = m->queue; 270 vm_pageout_init_marker(&marker, queue); 271 pq = vm_page_pagequeue(m); 272 object = m->object; 273 274 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q); 275 vm_pagequeue_unlock(pq); 276 vm_page_unlock(m); 277 VM_OBJECT_WLOCK(object); 278 vm_page_lock(m); 279 vm_pagequeue_lock(pq); 280 281 /* 282 * The page's object might have changed, and/or the page might 283 * have moved from its original position in the queue. If the 284 * page's object has changed, then the caller should abandon 285 * processing the page because the wrong object lock was 286 * acquired. Use the marker's plinks.q, not the page's, to 287 * determine if the page has been moved. The state of the 288 * page's plinks.q can be indeterminate; whereas, the marker's 289 * plinks.q must be valid. 290 */ 291 *next = TAILQ_NEXT(&marker, plinks.q); 292 unchanged = m->object == object && 293 m == TAILQ_PREV(&marker, pglist, plinks.q); 294 KASSERT(!unchanged || m->queue == queue, 295 ("page %p queue %d %d", m, queue, m->queue)); 296 TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q); 297 return (unchanged); 298 } 299 300 /* 301 * Lock the page while holding the page queue lock. Use marker page 302 * to detect page queue changes and maintain notion of next page on 303 * page queue. Return TRUE if no changes were detected, FALSE 304 * otherwise. The page is locked on return. The page queue lock might 305 * be dropped and reacquired. 306 * 307 * This function depends on normal struct vm_page being type stable. 308 */ 309 static boolean_t 310 vm_pageout_page_lock(vm_page_t m, vm_page_t *next) 311 { 312 struct vm_page marker; 313 struct vm_pagequeue *pq; 314 boolean_t unchanged; 315 u_short queue; 316 317 vm_page_lock_assert(m, MA_NOTOWNED); 318 if (vm_page_trylock(m)) 319 return (TRUE); 320 321 queue = m->queue; 322 vm_pageout_init_marker(&marker, queue); 323 pq = vm_page_pagequeue(m); 324 325 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q); 326 vm_pagequeue_unlock(pq); 327 vm_page_lock(m); 328 vm_pagequeue_lock(pq); 329 330 /* Page queue might have changed. */ 331 *next = TAILQ_NEXT(&marker, plinks.q); 332 unchanged = m == TAILQ_PREV(&marker, pglist, plinks.q); 333 KASSERT(!unchanged || m->queue == queue, 334 ("page %p queue %d %d", m, queue, m->queue)); 335 TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q); 336 return (unchanged); 337 } 338 339 /* 340 * Scan for pages at adjacent offsets within the given page's object that are 341 * eligible for laundering, form a cluster of these pages and the given page, 342 * and launder that cluster. 343 */ 344 static int 345 vm_pageout_cluster(vm_page_t m) 346 { 347 vm_object_t object; 348 vm_page_t mc[2 * vm_pageout_page_count], p, pb, ps; 349 vm_pindex_t pindex; 350 int ib, is, page_base, pageout_count; 351 352 vm_page_assert_locked(m); 353 object = m->object; 354 VM_OBJECT_ASSERT_WLOCKED(object); 355 pindex = m->pindex; 356 357 /* 358 * We can't clean the page if it is busy or held. 359 */ 360 vm_page_assert_unbusied(m); 361 KASSERT(m->hold_count == 0, ("page %p is held", m)); 362 363 pmap_remove_write(m); 364 vm_page_unlock(m); 365 366 mc[vm_pageout_page_count] = pb = ps = m; 367 pageout_count = 1; 368 page_base = vm_pageout_page_count; 369 ib = 1; 370 is = 1; 371 372 /* 373 * We can cluster only if the page is not clean, busy, or held, and 374 * the page is in the laundry queue. 375 * 376 * During heavy mmap/modification loads the pageout 377 * daemon can really fragment the underlying file 378 * due to flushing pages out of order and not trying to 379 * align the clusters (which leaves sporadic out-of-order 380 * holes). To solve this problem we do the reverse scan 381 * first and attempt to align our cluster, then do a 382 * forward scan if room remains. 383 */ 384 more: 385 while (ib != 0 && pageout_count < vm_pageout_page_count) { 386 if (ib > pindex) { 387 ib = 0; 388 break; 389 } 390 if ((p = vm_page_prev(pb)) == NULL || vm_page_busied(p)) { 391 ib = 0; 392 break; 393 } 394 vm_page_test_dirty(p); 395 if (p->dirty == 0) { 396 ib = 0; 397 break; 398 } 399 vm_page_lock(p); 400 if (!vm_page_in_laundry(p) || 401 p->hold_count != 0) { /* may be undergoing I/O */ 402 vm_page_unlock(p); 403 ib = 0; 404 break; 405 } 406 pmap_remove_write(p); 407 vm_page_unlock(p); 408 mc[--page_base] = pb = p; 409 ++pageout_count; 410 ++ib; 411 412 /* 413 * We are at an alignment boundary. Stop here, and switch 414 * directions. Do not clear ib. 415 */ 416 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0) 417 break; 418 } 419 while (pageout_count < vm_pageout_page_count && 420 pindex + is < object->size) { 421 if ((p = vm_page_next(ps)) == NULL || vm_page_busied(p)) 422 break; 423 vm_page_test_dirty(p); 424 if (p->dirty == 0) 425 break; 426 vm_page_lock(p); 427 if (!vm_page_in_laundry(p) || 428 p->hold_count != 0) { /* may be undergoing I/O */ 429 vm_page_unlock(p); 430 break; 431 } 432 pmap_remove_write(p); 433 vm_page_unlock(p); 434 mc[page_base + pageout_count] = ps = p; 435 ++pageout_count; 436 ++is; 437 } 438 439 /* 440 * If we exhausted our forward scan, continue with the reverse scan 441 * when possible, even past an alignment boundary. This catches 442 * boundary conditions. 443 */ 444 if (ib != 0 && pageout_count < vm_pageout_page_count) 445 goto more; 446 447 return (vm_pageout_flush(&mc[page_base], pageout_count, 448 VM_PAGER_PUT_NOREUSE, 0, NULL, NULL)); 449 } 450 451 /* 452 * vm_pageout_flush() - launder the given pages 453 * 454 * The given pages are laundered. Note that we setup for the start of 455 * I/O ( i.e. busy the page ), mark it read-only, and bump the object 456 * reference count all in here rather then in the parent. If we want 457 * the parent to do more sophisticated things we may have to change 458 * the ordering. 459 * 460 * Returned runlen is the count of pages between mreq and first 461 * page after mreq with status VM_PAGER_AGAIN. 462 * *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL 463 * for any page in runlen set. 464 */ 465 int 466 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen, 467 boolean_t *eio) 468 { 469 vm_object_t object = mc[0]->object; 470 int pageout_status[count]; 471 int numpagedout = 0; 472 int i, runlen; 473 474 VM_OBJECT_ASSERT_WLOCKED(object); 475 476 /* 477 * Initiate I/O. Mark the pages busy and verify that they're valid 478 * and read-only. 479 * 480 * We do not have to fixup the clean/dirty bits here... we can 481 * allow the pager to do it after the I/O completes. 482 * 483 * NOTE! mc[i]->dirty may be partial or fragmented due to an 484 * edge case with file fragments. 485 */ 486 for (i = 0; i < count; i++) { 487 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, 488 ("vm_pageout_flush: partially invalid page %p index %d/%d", 489 mc[i], i, count)); 490 KASSERT((mc[i]->aflags & PGA_WRITEABLE) == 0, 491 ("vm_pageout_flush: writeable page %p", mc[i])); 492 vm_page_sbusy(mc[i]); 493 } 494 vm_object_pip_add(object, count); 495 496 vm_pager_put_pages(object, mc, count, flags, pageout_status); 497 498 runlen = count - mreq; 499 if (eio != NULL) 500 *eio = FALSE; 501 for (i = 0; i < count; i++) { 502 vm_page_t mt = mc[i]; 503 504 KASSERT(pageout_status[i] == VM_PAGER_PEND || 505 !pmap_page_is_write_mapped(mt), 506 ("vm_pageout_flush: page %p is not write protected", mt)); 507 switch (pageout_status[i]) { 508 case VM_PAGER_OK: 509 vm_page_lock(mt); 510 if (vm_page_in_laundry(mt)) 511 vm_page_deactivate_noreuse(mt); 512 vm_page_unlock(mt); 513 /* FALLTHROUGH */ 514 case VM_PAGER_PEND: 515 numpagedout++; 516 break; 517 case VM_PAGER_BAD: 518 /* 519 * The page is outside the object's range. We pretend 520 * that the page out worked and clean the page, so the 521 * changes will be lost if the page is reclaimed by 522 * the page daemon. 523 */ 524 vm_page_undirty(mt); 525 vm_page_lock(mt); 526 if (vm_page_in_laundry(mt)) 527 vm_page_deactivate_noreuse(mt); 528 vm_page_unlock(mt); 529 break; 530 case VM_PAGER_ERROR: 531 case VM_PAGER_FAIL: 532 /* 533 * If the page couldn't be paged out to swap because the 534 * pager wasn't able to find space, place the page in 535 * the PQ_UNSWAPPABLE holding queue. This is an 536 * optimization that prevents the page daemon from 537 * wasting CPU cycles on pages that cannot be reclaimed 538 * becase no swap device is configured. 539 * 540 * Otherwise, reactivate the page so that it doesn't 541 * clog the laundry and inactive queues. (We will try 542 * paging it out again later.) 543 */ 544 vm_page_lock(mt); 545 if (object->type == OBJT_SWAP && 546 pageout_status[i] == VM_PAGER_FAIL) { 547 vm_page_unswappable(mt); 548 numpagedout++; 549 } else 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 void 578 vm_pageout_swapon(void *arg __unused, struct swdevt *sp __unused) 579 { 580 581 atomic_store_rel_int(&swapdev_enabled, 1); 582 } 583 584 static void 585 vm_pageout_swapoff(void *arg __unused, struct swdevt *sp __unused) 586 { 587 588 if (swap_pager_nswapdev() == 1) 589 atomic_store_rel_int(&swapdev_enabled, 0); 590 } 591 592 /* 593 * Attempt to acquire all of the necessary locks to launder a page and 594 * then call through the clustering layer to PUTPAGES. Wait a short 595 * time for a vnode lock. 596 * 597 * Requires the page and object lock on entry, releases both before return. 598 * Returns 0 on success and an errno otherwise. 599 */ 600 static int 601 vm_pageout_clean(vm_page_t m, int *numpagedout) 602 { 603 struct vnode *vp; 604 struct mount *mp; 605 vm_object_t object; 606 vm_pindex_t pindex; 607 int error, lockmode; 608 609 vm_page_assert_locked(m); 610 object = m->object; 611 VM_OBJECT_ASSERT_WLOCKED(object); 612 error = 0; 613 vp = NULL; 614 mp = NULL; 615 616 /* 617 * The object is already known NOT to be dead. It 618 * is possible for the vget() to block the whole 619 * pageout daemon, but the new low-memory handling 620 * code should prevent it. 621 * 622 * We can't wait forever for the vnode lock, we might 623 * deadlock due to a vn_read() getting stuck in 624 * vm_wait while holding this vnode. We skip the 625 * vnode if we can't get it in a reasonable amount 626 * of time. 627 */ 628 if (object->type == OBJT_VNODE) { 629 vm_page_unlock(m); 630 vp = object->handle; 631 if (vp->v_type == VREG && 632 vn_start_write(vp, &mp, V_NOWAIT) != 0) { 633 mp = NULL; 634 error = EDEADLK; 635 goto unlock_all; 636 } 637 KASSERT(mp != NULL, 638 ("vp %p with NULL v_mount", vp)); 639 vm_object_reference_locked(object); 640 pindex = m->pindex; 641 VM_OBJECT_WUNLOCK(object); 642 lockmode = MNT_SHARED_WRITES(vp->v_mount) ? 643 LK_SHARED : LK_EXCLUSIVE; 644 if (vget(vp, lockmode | LK_TIMELOCK, curthread)) { 645 vp = NULL; 646 error = EDEADLK; 647 goto unlock_mp; 648 } 649 VM_OBJECT_WLOCK(object); 650 vm_page_lock(m); 651 /* 652 * While the object and page were unlocked, the page 653 * may have been: 654 * (1) moved to a different queue, 655 * (2) reallocated to a different object, 656 * (3) reallocated to a different offset, or 657 * (4) cleaned. 658 */ 659 if (!vm_page_in_laundry(m) || m->object != object || 660 m->pindex != pindex || m->dirty == 0) { 661 vm_page_unlock(m); 662 error = ENXIO; 663 goto unlock_all; 664 } 665 666 /* 667 * The page may have been busied or held while the object 668 * and page locks were released. 669 */ 670 if (vm_page_busied(m) || m->hold_count != 0) { 671 vm_page_unlock(m); 672 error = EBUSY; 673 goto unlock_all; 674 } 675 } 676 677 /* 678 * If a page is dirty, then it is either being washed 679 * (but not yet cleaned) or it is still in the 680 * laundry. If it is still in the laundry, then we 681 * start the cleaning operation. 682 */ 683 if ((*numpagedout = vm_pageout_cluster(m)) == 0) 684 error = EIO; 685 686 unlock_all: 687 VM_OBJECT_WUNLOCK(object); 688 689 unlock_mp: 690 vm_page_lock_assert(m, MA_NOTOWNED); 691 if (mp != NULL) { 692 if (vp != NULL) 693 vput(vp); 694 vm_object_deallocate(object); 695 vn_finished_write(mp); 696 } 697 698 return (error); 699 } 700 701 /* 702 * Attempt to launder the specified number of pages. 703 * 704 * Returns the number of pages successfully laundered. 705 */ 706 static int 707 vm_pageout_launder(struct vm_domain *vmd, int launder, bool in_shortfall) 708 { 709 struct vm_pagequeue *pq; 710 vm_object_t object; 711 vm_page_t m, next; 712 int act_delta, error, maxscan, numpagedout, starting_target; 713 int vnodes_skipped; 714 bool pageout_ok, queue_locked; 715 716 starting_target = launder; 717 vnodes_skipped = 0; 718 719 /* 720 * Scan the laundry queues for pages eligible to be laundered. We stop 721 * once the target number of dirty pages have been laundered, or once 722 * we've reached the end of the queue. A single iteration of this loop 723 * may cause more than one page to be laundered because of clustering. 724 * 725 * maxscan ensures that we don't re-examine requeued pages. Any 726 * additional pages written as part of a cluster are subtracted from 727 * maxscan since they must be taken from the laundry queue. 728 * 729 * As an optimization, we avoid laundering from PQ_UNSWAPPABLE when no 730 * swap devices are configured. 731 */ 732 if (atomic_load_acq_int(&swapdev_enabled)) 733 pq = &vmd->vmd_pagequeues[PQ_UNSWAPPABLE]; 734 else 735 pq = &vmd->vmd_pagequeues[PQ_LAUNDRY]; 736 737 scan: 738 vm_pagequeue_lock(pq); 739 maxscan = pq->pq_cnt; 740 queue_locked = true; 741 for (m = TAILQ_FIRST(&pq->pq_pl); 742 m != NULL && maxscan-- > 0 && launder > 0; 743 m = next) { 744 vm_pagequeue_assert_locked(pq); 745 KASSERT(queue_locked, ("unlocked laundry queue")); 746 KASSERT(vm_page_in_laundry(m), 747 ("page %p has an inconsistent queue", m)); 748 next = TAILQ_NEXT(m, plinks.q); 749 if ((m->flags & PG_MARKER) != 0) 750 continue; 751 KASSERT((m->flags & PG_FICTITIOUS) == 0, 752 ("PG_FICTITIOUS page %p cannot be in laundry queue", m)); 753 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 754 ("VPO_UNMANAGED page %p cannot be in laundry queue", m)); 755 if (!vm_pageout_page_lock(m, &next) || m->hold_count != 0) { 756 vm_page_unlock(m); 757 continue; 758 } 759 object = m->object; 760 if ((!VM_OBJECT_TRYWLOCK(object) && 761 (!vm_pageout_fallback_object_lock(m, &next) || 762 m->hold_count != 0)) || vm_page_busied(m)) { 763 VM_OBJECT_WUNLOCK(object); 764 vm_page_unlock(m); 765 continue; 766 } 767 768 /* 769 * Unlock the laundry queue, invalidating the 'next' pointer. 770 * Use a marker to remember our place in the laundry queue. 771 */ 772 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &vmd->vmd_laundry_marker, 773 plinks.q); 774 vm_pagequeue_unlock(pq); 775 queue_locked = false; 776 777 /* 778 * Invalid pages can be easily freed. They cannot be 779 * mapped; vm_page_free() asserts this. 780 */ 781 if (m->valid == 0) 782 goto free_page; 783 784 /* 785 * If the page has been referenced and the object is not dead, 786 * reactivate or requeue the page depending on whether the 787 * object is mapped. 788 */ 789 if ((m->aflags & PGA_REFERENCED) != 0) { 790 vm_page_aflag_clear(m, PGA_REFERENCED); 791 act_delta = 1; 792 } else 793 act_delta = 0; 794 if (object->ref_count != 0) 795 act_delta += pmap_ts_referenced(m); 796 else { 797 KASSERT(!pmap_page_is_mapped(m), 798 ("page %p is mapped", m)); 799 } 800 if (act_delta != 0) { 801 if (object->ref_count != 0) { 802 VM_CNT_INC(v_reactivated); 803 vm_page_activate(m); 804 805 /* 806 * Increase the activation count if the page 807 * was referenced while in the laundry queue. 808 * This makes it less likely that the page will 809 * be returned prematurely to the inactive 810 * queue. 811 */ 812 m->act_count += act_delta + ACT_ADVANCE; 813 814 /* 815 * If this was a background laundering, count 816 * activated pages towards our target. The 817 * purpose of background laundering is to ensure 818 * that pages are eventually cycled through the 819 * laundry queue, and an activation is a valid 820 * way out. 821 */ 822 if (!in_shortfall) 823 launder--; 824 goto drop_page; 825 } else if ((object->flags & OBJ_DEAD) == 0) 826 goto requeue_page; 827 } 828 829 /* 830 * If the page appears to be clean at the machine-independent 831 * layer, then remove all of its mappings from the pmap in 832 * anticipation of freeing it. If, however, any of the page's 833 * mappings allow write access, then the page may still be 834 * modified until the last of those mappings are removed. 835 */ 836 if (object->ref_count != 0) { 837 vm_page_test_dirty(m); 838 if (m->dirty == 0) 839 pmap_remove_all(m); 840 } 841 842 /* 843 * Clean pages are freed, and dirty pages are paged out unless 844 * they belong to a dead object. Requeueing dirty pages from 845 * dead objects is pointless, as they are being paged out and 846 * freed by the thread that destroyed the object. 847 */ 848 if (m->dirty == 0) { 849 free_page: 850 vm_page_free(m); 851 VM_CNT_INC(v_dfree); 852 } else if ((object->flags & OBJ_DEAD) == 0) { 853 if (object->type != OBJT_SWAP && 854 object->type != OBJT_DEFAULT) 855 pageout_ok = true; 856 else if (disable_swap_pageouts) 857 pageout_ok = false; 858 else 859 pageout_ok = true; 860 if (!pageout_ok) { 861 requeue_page: 862 vm_pagequeue_lock(pq); 863 queue_locked = true; 864 vm_page_requeue_locked(m); 865 goto drop_page; 866 } 867 868 /* 869 * Form a cluster with adjacent, dirty pages from the 870 * same object, and page out that entire cluster. 871 * 872 * The adjacent, dirty pages must also be in the 873 * laundry. However, their mappings are not checked 874 * for new references. Consequently, a recently 875 * referenced page may be paged out. However, that 876 * page will not be prematurely reclaimed. After page 877 * out, the page will be placed in the inactive queue, 878 * where any new references will be detected and the 879 * page reactivated. 880 */ 881 error = vm_pageout_clean(m, &numpagedout); 882 if (error == 0) { 883 launder -= numpagedout; 884 maxscan -= numpagedout - 1; 885 } else if (error == EDEADLK) { 886 pageout_lock_miss++; 887 vnodes_skipped++; 888 } 889 goto relock_queue; 890 } 891 drop_page: 892 vm_page_unlock(m); 893 VM_OBJECT_WUNLOCK(object); 894 relock_queue: 895 if (!queue_locked) { 896 vm_pagequeue_lock(pq); 897 queue_locked = true; 898 } 899 next = TAILQ_NEXT(&vmd->vmd_laundry_marker, plinks.q); 900 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_laundry_marker, plinks.q); 901 } 902 vm_pagequeue_unlock(pq); 903 904 if (launder > 0 && pq == &vmd->vmd_pagequeues[PQ_UNSWAPPABLE]) { 905 pq = &vmd->vmd_pagequeues[PQ_LAUNDRY]; 906 goto scan; 907 } 908 909 /* 910 * Wakeup the sync daemon if we skipped a vnode in a writeable object 911 * and we didn't launder enough pages. 912 */ 913 if (vnodes_skipped > 0 && launder > 0) 914 (void)speedup_syncer(); 915 916 return (starting_target - launder); 917 } 918 919 /* 920 * Compute the integer square root. 921 */ 922 static u_int 923 isqrt(u_int num) 924 { 925 u_int bit, root, tmp; 926 927 bit = 1u << ((NBBY * sizeof(u_int)) - 2); 928 while (bit > num) 929 bit >>= 2; 930 root = 0; 931 while (bit != 0) { 932 tmp = root + bit; 933 root >>= 1; 934 if (num >= tmp) { 935 num -= tmp; 936 root += bit; 937 } 938 bit >>= 2; 939 } 940 return (root); 941 } 942 943 /* 944 * Perform the work of the laundry thread: periodically wake up and determine 945 * whether any pages need to be laundered. If so, determine the number of pages 946 * that need to be laundered, and launder them. 947 */ 948 static void 949 vm_pageout_laundry_worker(void *arg) 950 { 951 struct vm_domain *domain; 952 struct vm_pagequeue *pq; 953 uint64_t nclean, ndirty; 954 u_int last_launder, wakeups; 955 int domidx, last_target, launder, shortfall, shortfall_cycle, target; 956 bool in_shortfall; 957 958 domidx = (uintptr_t)arg; 959 domain = &vm_dom[domidx]; 960 pq = &domain->vmd_pagequeues[PQ_LAUNDRY]; 961 KASSERT(domain->vmd_segs != 0, ("domain without segments")); 962 vm_pageout_init_marker(&domain->vmd_laundry_marker, PQ_LAUNDRY); 963 964 shortfall = 0; 965 in_shortfall = false; 966 shortfall_cycle = 0; 967 target = 0; 968 last_launder = 0; 969 970 /* 971 * Calls to these handlers are serialized by the swap syscall lock. 972 */ 973 (void)EVENTHANDLER_REGISTER(swapon, vm_pageout_swapon, domain, 974 EVENTHANDLER_PRI_ANY); 975 (void)EVENTHANDLER_REGISTER(swapoff, vm_pageout_swapoff, domain, 976 EVENTHANDLER_PRI_ANY); 977 978 /* 979 * The pageout laundry worker is never done, so loop forever. 980 */ 981 for (;;) { 982 KASSERT(target >= 0, ("negative target %d", target)); 983 KASSERT(shortfall_cycle >= 0, 984 ("negative cycle %d", shortfall_cycle)); 985 launder = 0; 986 wakeups = VM_CNT_FETCH(v_pdwakeups); 987 988 /* 989 * First determine whether we need to launder pages to meet a 990 * shortage of free pages. 991 */ 992 if (shortfall > 0) { 993 in_shortfall = true; 994 shortfall_cycle = VM_LAUNDER_RATE / VM_INACT_SCAN_RATE; 995 target = shortfall; 996 } else if (!in_shortfall) 997 goto trybackground; 998 else if (shortfall_cycle == 0 || vm_laundry_target() <= 0) { 999 /* 1000 * We recently entered shortfall and began laundering 1001 * pages. If we have completed that laundering run 1002 * (and we are no longer in shortfall) or we have met 1003 * our laundry target through other activity, then we 1004 * can stop laundering pages. 1005 */ 1006 in_shortfall = false; 1007 target = 0; 1008 goto trybackground; 1009 } 1010 last_launder = wakeups; 1011 launder = target / shortfall_cycle--; 1012 goto dolaundry; 1013 1014 /* 1015 * There's no immediate need to launder any pages; see if we 1016 * meet the conditions to perform background laundering: 1017 * 1018 * 1. The ratio of dirty to clean inactive pages exceeds the 1019 * background laundering threshold and the pagedaemon has 1020 * been woken up to reclaim pages since our last 1021 * laundering, or 1022 * 2. we haven't yet reached the target of the current 1023 * background laundering run. 1024 * 1025 * The background laundering threshold is not a constant. 1026 * Instead, it is a slowly growing function of the number of 1027 * page daemon wakeups since the last laundering. Thus, as the 1028 * ratio of dirty to clean inactive pages grows, the amount of 1029 * memory pressure required to trigger laundering decreases. 1030 */ 1031 trybackground: 1032 nclean = vm_cnt.v_inactive_count + vm_cnt.v_free_count; 1033 ndirty = vm_cnt.v_laundry_count; 1034 if (target == 0 && wakeups != last_launder && 1035 ndirty * isqrt(wakeups - last_launder) >= nclean) { 1036 target = vm_background_launder_target; 1037 } 1038 1039 /* 1040 * We have a non-zero background laundering target. If we've 1041 * laundered up to our maximum without observing a page daemon 1042 * wakeup, just stop. This is a safety belt that ensures we 1043 * don't launder an excessive amount if memory pressure is low 1044 * and the ratio of dirty to clean pages is large. Otherwise, 1045 * proceed at the background laundering rate. 1046 */ 1047 if (target > 0) { 1048 if (wakeups != last_launder) { 1049 last_launder = wakeups; 1050 last_target = target; 1051 } else if (last_target - target >= 1052 vm_background_launder_max * PAGE_SIZE / 1024) { 1053 target = 0; 1054 } 1055 launder = vm_background_launder_rate * PAGE_SIZE / 1024; 1056 launder /= VM_LAUNDER_RATE; 1057 if (launder > target) 1058 launder = target; 1059 } 1060 1061 dolaundry: 1062 if (launder > 0) { 1063 /* 1064 * Because of I/O clustering, the number of laundered 1065 * pages could exceed "target" by the maximum size of 1066 * a cluster minus one. 1067 */ 1068 target -= min(vm_pageout_launder(domain, launder, 1069 in_shortfall), target); 1070 pause("laundp", hz / VM_LAUNDER_RATE); 1071 } 1072 1073 /* 1074 * If we're not currently laundering pages and the page daemon 1075 * hasn't posted a new request, sleep until the page daemon 1076 * kicks us. 1077 */ 1078 vm_pagequeue_lock(pq); 1079 if (target == 0 && vm_laundry_request == VM_LAUNDRY_IDLE) 1080 (void)mtx_sleep(&vm_laundry_request, 1081 vm_pagequeue_lockptr(pq), PVM, "launds", 0); 1082 1083 /* 1084 * If the pagedaemon has indicated that it's in shortfall, start 1085 * a shortfall laundering unless we're already in the middle of 1086 * one. This may preempt a background laundering. 1087 */ 1088 if (vm_laundry_request == VM_LAUNDRY_SHORTFALL && 1089 (!in_shortfall || shortfall_cycle == 0)) { 1090 shortfall = vm_laundry_target() + vm_pageout_deficit; 1091 target = 0; 1092 } else 1093 shortfall = 0; 1094 1095 if (target == 0) 1096 vm_laundry_request = VM_LAUNDRY_IDLE; 1097 vm_pagequeue_unlock(pq); 1098 } 1099 } 1100 1101 /* 1102 * vm_pageout_scan does the dirty work for the pageout daemon. 1103 * 1104 * pass == 0: Update active LRU/deactivate pages 1105 * pass >= 1: Free inactive pages 1106 * 1107 * Returns true if pass was zero or enough pages were freed by the inactive 1108 * queue scan to meet the target. 1109 */ 1110 static bool 1111 vm_pageout_scan(struct vm_domain *vmd, int pass) 1112 { 1113 vm_page_t m, next; 1114 struct vm_pagequeue *pq; 1115 vm_object_t object; 1116 long min_scan; 1117 int act_delta, addl_page_shortage, deficit, inactq_shortage, maxscan; 1118 int page_shortage, scan_tick, scanned, starting_page_shortage; 1119 boolean_t queue_locked; 1120 1121 /* 1122 * If we need to reclaim memory ask kernel caches to return 1123 * some. We rate limit to avoid thrashing. 1124 */ 1125 if (vmd == &vm_dom[0] && pass > 0 && 1126 (time_uptime - lowmem_uptime) >= lowmem_period) { 1127 /* 1128 * Decrease registered cache sizes. 1129 */ 1130 SDT_PROBE0(vm, , , vm__lowmem_scan); 1131 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_PAGES); 1132 /* 1133 * We do this explicitly after the caches have been 1134 * drained above. 1135 */ 1136 uma_reclaim(); 1137 lowmem_uptime = time_uptime; 1138 } 1139 1140 /* 1141 * The addl_page_shortage is the number of temporarily 1142 * stuck pages in the inactive queue. In other words, the 1143 * number of pages from the inactive count that should be 1144 * discounted in setting the target for the active queue scan. 1145 */ 1146 addl_page_shortage = 0; 1147 1148 /* 1149 * Calculate the number of pages that we want to free. This number 1150 * can be negative if many pages are freed between the wakeup call to 1151 * the page daemon and this calculation. 1152 */ 1153 if (pass > 0) { 1154 deficit = atomic_readandclear_int(&vm_pageout_deficit); 1155 page_shortage = vm_paging_target() + deficit; 1156 } else 1157 page_shortage = deficit = 0; 1158 starting_page_shortage = page_shortage; 1159 1160 /* 1161 * Start scanning the inactive queue for pages that we can free. The 1162 * scan will stop when we reach the target or we have scanned the 1163 * entire queue. (Note that m->act_count is not used to make 1164 * decisions for the inactive queue, only for the active queue.) 1165 */ 1166 pq = &vmd->vmd_pagequeues[PQ_INACTIVE]; 1167 maxscan = pq->pq_cnt; 1168 vm_pagequeue_lock(pq); 1169 queue_locked = TRUE; 1170 for (m = TAILQ_FIRST(&pq->pq_pl); 1171 m != NULL && maxscan-- > 0 && page_shortage > 0; 1172 m = next) { 1173 vm_pagequeue_assert_locked(pq); 1174 KASSERT(queue_locked, ("unlocked inactive queue")); 1175 KASSERT(vm_page_inactive(m), ("Inactive queue %p", m)); 1176 1177 VM_CNT_INC(v_pdpages); 1178 next = TAILQ_NEXT(m, plinks.q); 1179 1180 /* 1181 * skip marker pages 1182 */ 1183 if (m->flags & PG_MARKER) 1184 continue; 1185 1186 KASSERT((m->flags & PG_FICTITIOUS) == 0, 1187 ("Fictitious page %p cannot be in inactive queue", m)); 1188 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 1189 ("Unmanaged page %p cannot be in inactive queue", m)); 1190 1191 /* 1192 * The page or object lock acquisitions fail if the 1193 * page was removed from the queue or moved to a 1194 * different position within the queue. In either 1195 * case, addl_page_shortage should not be incremented. 1196 */ 1197 if (!vm_pageout_page_lock(m, &next)) 1198 goto unlock_page; 1199 else if (m->hold_count != 0) { 1200 /* 1201 * Held pages are essentially stuck in the 1202 * queue. So, they ought to be discounted 1203 * from the inactive count. See the 1204 * calculation of inactq_shortage before the 1205 * loop over the active queue below. 1206 */ 1207 addl_page_shortage++; 1208 goto unlock_page; 1209 } 1210 object = m->object; 1211 if (!VM_OBJECT_TRYWLOCK(object)) { 1212 if (!vm_pageout_fallback_object_lock(m, &next)) 1213 goto unlock_object; 1214 else if (m->hold_count != 0) { 1215 addl_page_shortage++; 1216 goto unlock_object; 1217 } 1218 } 1219 if (vm_page_busied(m)) { 1220 /* 1221 * Don't mess with busy pages. Leave them at 1222 * the front of the queue. Most likely, they 1223 * are being paged out and will leave the 1224 * queue shortly after the scan finishes. So, 1225 * they ought to be discounted from the 1226 * inactive count. 1227 */ 1228 addl_page_shortage++; 1229 unlock_object: 1230 VM_OBJECT_WUNLOCK(object); 1231 unlock_page: 1232 vm_page_unlock(m); 1233 continue; 1234 } 1235 KASSERT(m->hold_count == 0, ("Held page %p", m)); 1236 1237 /* 1238 * Dequeue the inactive page and unlock the inactive page 1239 * queue, invalidating the 'next' pointer. Dequeueing the 1240 * page here avoids a later reacquisition (and release) of 1241 * the inactive page queue lock when vm_page_activate(), 1242 * vm_page_free(), or vm_page_launder() is called. Use a 1243 * marker to remember our place in the inactive queue. 1244 */ 1245 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &vmd->vmd_marker, plinks.q); 1246 vm_page_dequeue_locked(m); 1247 vm_pagequeue_unlock(pq); 1248 queue_locked = FALSE; 1249 1250 /* 1251 * Invalid pages can be easily freed. They cannot be 1252 * mapped, vm_page_free() asserts this. 1253 */ 1254 if (m->valid == 0) 1255 goto free_page; 1256 1257 /* 1258 * If the page has been referenced and the object is not dead, 1259 * reactivate or requeue the page depending on whether the 1260 * object is mapped. 1261 */ 1262 if ((m->aflags & PGA_REFERENCED) != 0) { 1263 vm_page_aflag_clear(m, PGA_REFERENCED); 1264 act_delta = 1; 1265 } else 1266 act_delta = 0; 1267 if (object->ref_count != 0) { 1268 act_delta += pmap_ts_referenced(m); 1269 } else { 1270 KASSERT(!pmap_page_is_mapped(m), 1271 ("vm_pageout_scan: page %p is mapped", m)); 1272 } 1273 if (act_delta != 0) { 1274 if (object->ref_count != 0) { 1275 VM_CNT_INC(v_reactivated); 1276 vm_page_activate(m); 1277 1278 /* 1279 * Increase the activation count if the page 1280 * was referenced while in the inactive queue. 1281 * This makes it less likely that the page will 1282 * be returned prematurely to the inactive 1283 * queue. 1284 */ 1285 m->act_count += act_delta + ACT_ADVANCE; 1286 goto drop_page; 1287 } else if ((object->flags & OBJ_DEAD) == 0) { 1288 vm_pagequeue_lock(pq); 1289 queue_locked = TRUE; 1290 m->queue = PQ_INACTIVE; 1291 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 1292 vm_pagequeue_cnt_inc(pq); 1293 goto drop_page; 1294 } 1295 } 1296 1297 /* 1298 * If the page appears to be clean at the machine-independent 1299 * layer, then remove all of its mappings from the pmap in 1300 * anticipation of freeing it. If, however, any of the page's 1301 * mappings allow write access, then the page may still be 1302 * modified until the last of those mappings are removed. 1303 */ 1304 if (object->ref_count != 0) { 1305 vm_page_test_dirty(m); 1306 if (m->dirty == 0) 1307 pmap_remove_all(m); 1308 } 1309 1310 /* 1311 * Clean pages can be freed, but dirty pages must be sent back 1312 * to the laundry, unless they belong to a dead object. 1313 * Requeueing dirty pages from dead objects is pointless, as 1314 * they are being paged out and freed by the thread that 1315 * destroyed the object. 1316 */ 1317 if (m->dirty == 0) { 1318 free_page: 1319 vm_page_free(m); 1320 VM_CNT_INC(v_dfree); 1321 --page_shortage; 1322 } else if ((object->flags & OBJ_DEAD) == 0) 1323 vm_page_launder(m); 1324 drop_page: 1325 vm_page_unlock(m); 1326 VM_OBJECT_WUNLOCK(object); 1327 if (!queue_locked) { 1328 vm_pagequeue_lock(pq); 1329 queue_locked = TRUE; 1330 } 1331 next = TAILQ_NEXT(&vmd->vmd_marker, plinks.q); 1332 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_marker, plinks.q); 1333 } 1334 vm_pagequeue_unlock(pq); 1335 1336 /* 1337 * Wake up the laundry thread so that it can perform any needed 1338 * laundering. If we didn't meet our target, we're in shortfall and 1339 * need to launder more aggressively. If PQ_LAUNDRY is empty and no 1340 * swap devices are configured, the laundry thread has no work to do, so 1341 * don't bother waking it up. 1342 */ 1343 if (vm_laundry_request == VM_LAUNDRY_IDLE && 1344 starting_page_shortage > 0) { 1345 pq = &vm_dom[0].vmd_pagequeues[PQ_LAUNDRY]; 1346 vm_pagequeue_lock(pq); 1347 if (pq->pq_cnt > 0 || atomic_load_acq_int(&swapdev_enabled)) { 1348 if (page_shortage > 0) { 1349 vm_laundry_request = VM_LAUNDRY_SHORTFALL; 1350 VM_CNT_INC(v_pdshortfalls); 1351 } else if (vm_laundry_request != VM_LAUNDRY_SHORTFALL) 1352 vm_laundry_request = VM_LAUNDRY_BACKGROUND; 1353 wakeup(&vm_laundry_request); 1354 } 1355 vm_pagequeue_unlock(pq); 1356 } 1357 1358 /* 1359 * Wakeup the swapout daemon if we didn't free the targeted number of 1360 * pages. 1361 */ 1362 if (page_shortage > 0) 1363 vm_swapout_run(); 1364 1365 /* 1366 * If the inactive queue scan fails repeatedly to meet its 1367 * target, kill the largest process. 1368 */ 1369 vm_pageout_mightbe_oom(vmd, page_shortage, starting_page_shortage); 1370 1371 /* 1372 * Compute the number of pages we want to try to move from the 1373 * active queue to either the inactive or laundry queue. 1374 * 1375 * When scanning active pages, we make clean pages count more heavily 1376 * towards the page shortage than dirty pages. This is because dirty 1377 * pages must be laundered before they can be reused and thus have less 1378 * utility when attempting to quickly alleviate a shortage. However, 1379 * this weighting also causes the scan to deactivate dirty pages more 1380 * more aggressively, improving the effectiveness of clustering and 1381 * ensuring that they can eventually be reused. 1382 */ 1383 inactq_shortage = vm_cnt.v_inactive_target - (vm_cnt.v_inactive_count + 1384 vm_cnt.v_laundry_count / act_scan_laundry_weight) + 1385 vm_paging_target() + deficit + addl_page_shortage; 1386 page_shortage *= act_scan_laundry_weight; 1387 1388 pq = &vmd->vmd_pagequeues[PQ_ACTIVE]; 1389 vm_pagequeue_lock(pq); 1390 maxscan = pq->pq_cnt; 1391 1392 /* 1393 * If we're just idle polling attempt to visit every 1394 * active page within 'update_period' seconds. 1395 */ 1396 scan_tick = ticks; 1397 if (vm_pageout_update_period != 0) { 1398 min_scan = pq->pq_cnt; 1399 min_scan *= scan_tick - vmd->vmd_last_active_scan; 1400 min_scan /= hz * vm_pageout_update_period; 1401 } else 1402 min_scan = 0; 1403 if (min_scan > 0 || (inactq_shortage > 0 && maxscan > 0)) 1404 vmd->vmd_last_active_scan = scan_tick; 1405 1406 /* 1407 * Scan the active queue for pages that can be deactivated. Update 1408 * the per-page activity counter and use it to identify deactivation 1409 * candidates. Held pages may be deactivated. 1410 */ 1411 for (m = TAILQ_FIRST(&pq->pq_pl), scanned = 0; m != NULL && (scanned < 1412 min_scan || (inactq_shortage > 0 && scanned < maxscan)); m = next, 1413 scanned++) { 1414 KASSERT(m->queue == PQ_ACTIVE, 1415 ("vm_pageout_scan: page %p isn't active", m)); 1416 next = TAILQ_NEXT(m, plinks.q); 1417 if ((m->flags & PG_MARKER) != 0) 1418 continue; 1419 KASSERT((m->flags & PG_FICTITIOUS) == 0, 1420 ("Fictitious page %p cannot be in active queue", m)); 1421 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 1422 ("Unmanaged page %p cannot be in active queue", m)); 1423 if (!vm_pageout_page_lock(m, &next)) { 1424 vm_page_unlock(m); 1425 continue; 1426 } 1427 1428 /* 1429 * The count for page daemon pages is updated after checking 1430 * the page for eligibility. 1431 */ 1432 VM_CNT_INC(v_pdpages); 1433 1434 /* 1435 * Check to see "how much" the page has been used. 1436 */ 1437 if ((m->aflags & PGA_REFERENCED) != 0) { 1438 vm_page_aflag_clear(m, PGA_REFERENCED); 1439 act_delta = 1; 1440 } else 1441 act_delta = 0; 1442 1443 /* 1444 * Perform an unsynchronized object ref count check. While 1445 * the page lock ensures that the page is not reallocated to 1446 * another object, in particular, one with unmanaged mappings 1447 * that cannot support pmap_ts_referenced(), two races are, 1448 * nonetheless, possible: 1449 * 1) The count was transitioning to zero, but we saw a non- 1450 * zero value. pmap_ts_referenced() will return zero 1451 * because the page is not mapped. 1452 * 2) The count was transitioning to one, but we saw zero. 1453 * This race delays the detection of a new reference. At 1454 * worst, we will deactivate and reactivate the page. 1455 */ 1456 if (m->object->ref_count != 0) 1457 act_delta += pmap_ts_referenced(m); 1458 1459 /* 1460 * Advance or decay the act_count based on recent usage. 1461 */ 1462 if (act_delta != 0) { 1463 m->act_count += ACT_ADVANCE + act_delta; 1464 if (m->act_count > ACT_MAX) 1465 m->act_count = ACT_MAX; 1466 } else 1467 m->act_count -= min(m->act_count, ACT_DECLINE); 1468 1469 /* 1470 * Move this page to the tail of the active, inactive or laundry 1471 * queue depending on usage. 1472 */ 1473 if (m->act_count == 0) { 1474 /* Dequeue to avoid later lock recursion. */ 1475 vm_page_dequeue_locked(m); 1476 1477 /* 1478 * When not short for inactive pages, let dirty pages go 1479 * through the inactive queue before moving to the 1480 * laundry queues. This gives them some extra time to 1481 * be reactivated, potentially avoiding an expensive 1482 * pageout. During a page shortage, the inactive queue 1483 * is necessarily small, so we may move dirty pages 1484 * directly to the laundry queue. 1485 */ 1486 if (inactq_shortage <= 0) 1487 vm_page_deactivate(m); 1488 else { 1489 /* 1490 * Calling vm_page_test_dirty() here would 1491 * require acquisition of the object's write 1492 * lock. However, during a page shortage, 1493 * directing dirty pages into the laundry 1494 * queue is only an optimization and not a 1495 * requirement. Therefore, we simply rely on 1496 * the opportunistic updates to the page's 1497 * dirty field by the pmap. 1498 */ 1499 if (m->dirty == 0) { 1500 vm_page_deactivate(m); 1501 inactq_shortage -= 1502 act_scan_laundry_weight; 1503 } else { 1504 vm_page_launder(m); 1505 inactq_shortage--; 1506 } 1507 } 1508 } else 1509 vm_page_requeue_locked(m); 1510 vm_page_unlock(m); 1511 } 1512 vm_pagequeue_unlock(pq); 1513 if (pass > 0) 1514 vm_swapout_run_idle(); 1515 return (page_shortage <= 0); 1516 } 1517 1518 static int vm_pageout_oom_vote; 1519 1520 /* 1521 * The pagedaemon threads randlomly select one to perform the 1522 * OOM. Trying to kill processes before all pagedaemons 1523 * failed to reach free target is premature. 1524 */ 1525 static void 1526 vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage, 1527 int starting_page_shortage) 1528 { 1529 int old_vote; 1530 1531 if (starting_page_shortage <= 0 || starting_page_shortage != 1532 page_shortage) 1533 vmd->vmd_oom_seq = 0; 1534 else 1535 vmd->vmd_oom_seq++; 1536 if (vmd->vmd_oom_seq < vm_pageout_oom_seq) { 1537 if (vmd->vmd_oom) { 1538 vmd->vmd_oom = FALSE; 1539 atomic_subtract_int(&vm_pageout_oom_vote, 1); 1540 } 1541 return; 1542 } 1543 1544 /* 1545 * Do not follow the call sequence until OOM condition is 1546 * cleared. 1547 */ 1548 vmd->vmd_oom_seq = 0; 1549 1550 if (vmd->vmd_oom) 1551 return; 1552 1553 vmd->vmd_oom = TRUE; 1554 old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1); 1555 if (old_vote != vm_ndomains - 1) 1556 return; 1557 1558 /* 1559 * The current pagedaemon thread is the last in the quorum to 1560 * start OOM. Initiate the selection and signaling of the 1561 * victim. 1562 */ 1563 vm_pageout_oom(VM_OOM_MEM); 1564 1565 /* 1566 * After one round of OOM terror, recall our vote. On the 1567 * next pass, current pagedaemon would vote again if the low 1568 * memory condition is still there, due to vmd_oom being 1569 * false. 1570 */ 1571 vmd->vmd_oom = FALSE; 1572 atomic_subtract_int(&vm_pageout_oom_vote, 1); 1573 } 1574 1575 /* 1576 * The OOM killer is the page daemon's action of last resort when 1577 * memory allocation requests have been stalled for a prolonged period 1578 * of time because it cannot reclaim memory. This function computes 1579 * the approximate number of physical pages that could be reclaimed if 1580 * the specified address space is destroyed. 1581 * 1582 * Private, anonymous memory owned by the address space is the 1583 * principal resource that we expect to recover after an OOM kill. 1584 * Since the physical pages mapped by the address space's COW entries 1585 * are typically shared pages, they are unlikely to be released and so 1586 * they are not counted. 1587 * 1588 * To get to the point where the page daemon runs the OOM killer, its 1589 * efforts to write-back vnode-backed pages may have stalled. This 1590 * could be caused by a memory allocation deadlock in the write path 1591 * that might be resolved by an OOM kill. Therefore, physical pages 1592 * belonging to vnode-backed objects are counted, because they might 1593 * be freed without being written out first if the address space holds 1594 * the last reference to an unlinked vnode. 1595 * 1596 * Similarly, physical pages belonging to OBJT_PHYS objects are 1597 * counted because the address space might hold the last reference to 1598 * the object. 1599 */ 1600 static long 1601 vm_pageout_oom_pagecount(struct vmspace *vmspace) 1602 { 1603 vm_map_t map; 1604 vm_map_entry_t entry; 1605 vm_object_t obj; 1606 long res; 1607 1608 map = &vmspace->vm_map; 1609 KASSERT(!map->system_map, ("system map")); 1610 sx_assert(&map->lock, SA_LOCKED); 1611 res = 0; 1612 for (entry = map->header.next; entry != &map->header; 1613 entry = entry->next) { 1614 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) 1615 continue; 1616 obj = entry->object.vm_object; 1617 if (obj == NULL) 1618 continue; 1619 if ((entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0 && 1620 obj->ref_count != 1) 1621 continue; 1622 switch (obj->type) { 1623 case OBJT_DEFAULT: 1624 case OBJT_SWAP: 1625 case OBJT_PHYS: 1626 case OBJT_VNODE: 1627 res += obj->resident_page_count; 1628 break; 1629 } 1630 } 1631 return (res); 1632 } 1633 1634 void 1635 vm_pageout_oom(int shortage) 1636 { 1637 struct proc *p, *bigproc; 1638 vm_offset_t size, bigsize; 1639 struct thread *td; 1640 struct vmspace *vm; 1641 bool breakout; 1642 1643 /* 1644 * We keep the process bigproc locked once we find it to keep anyone 1645 * from messing with it; however, there is a possibility of 1646 * deadlock if process B is bigproc and one of its child processes 1647 * attempts to propagate a signal to B while we are waiting for A's 1648 * lock while walking this list. To avoid this, we don't block on 1649 * the process lock but just skip a process if it is already locked. 1650 */ 1651 bigproc = NULL; 1652 bigsize = 0; 1653 sx_slock(&allproc_lock); 1654 FOREACH_PROC_IN_SYSTEM(p) { 1655 PROC_LOCK(p); 1656 1657 /* 1658 * If this is a system, protected or killed process, skip it. 1659 */ 1660 if (p->p_state != PRS_NORMAL || (p->p_flag & (P_INEXEC | 1661 P_PROTECTED | P_SYSTEM | P_WEXIT)) != 0 || 1662 p->p_pid == 1 || P_KILLED(p) || 1663 (p->p_pid < 48 && swap_pager_avail != 0)) { 1664 PROC_UNLOCK(p); 1665 continue; 1666 } 1667 /* 1668 * If the process is in a non-running type state, 1669 * don't touch it. Check all the threads individually. 1670 */ 1671 breakout = false; 1672 FOREACH_THREAD_IN_PROC(p, td) { 1673 thread_lock(td); 1674 if (!TD_ON_RUNQ(td) && 1675 !TD_IS_RUNNING(td) && 1676 !TD_IS_SLEEPING(td) && 1677 !TD_IS_SUSPENDED(td) && 1678 !TD_IS_SWAPPED(td)) { 1679 thread_unlock(td); 1680 breakout = true; 1681 break; 1682 } 1683 thread_unlock(td); 1684 } 1685 if (breakout) { 1686 PROC_UNLOCK(p); 1687 continue; 1688 } 1689 /* 1690 * get the process size 1691 */ 1692 vm = vmspace_acquire_ref(p); 1693 if (vm == NULL) { 1694 PROC_UNLOCK(p); 1695 continue; 1696 } 1697 _PHOLD_LITE(p); 1698 PROC_UNLOCK(p); 1699 sx_sunlock(&allproc_lock); 1700 if (!vm_map_trylock_read(&vm->vm_map)) { 1701 vmspace_free(vm); 1702 sx_slock(&allproc_lock); 1703 PRELE(p); 1704 continue; 1705 } 1706 size = vmspace_swap_count(vm); 1707 if (shortage == VM_OOM_MEM) 1708 size += vm_pageout_oom_pagecount(vm); 1709 vm_map_unlock_read(&vm->vm_map); 1710 vmspace_free(vm); 1711 sx_slock(&allproc_lock); 1712 1713 /* 1714 * If this process is bigger than the biggest one, 1715 * remember it. 1716 */ 1717 if (size > bigsize) { 1718 if (bigproc != NULL) 1719 PRELE(bigproc); 1720 bigproc = p; 1721 bigsize = size; 1722 } else { 1723 PRELE(p); 1724 } 1725 } 1726 sx_sunlock(&allproc_lock); 1727 if (bigproc != NULL) { 1728 if (vm_panic_on_oom != 0) 1729 panic("out of swap space"); 1730 PROC_LOCK(bigproc); 1731 killproc(bigproc, "out of swap space"); 1732 sched_nice(bigproc, PRIO_MIN); 1733 _PRELE(bigproc); 1734 PROC_UNLOCK(bigproc); 1735 wakeup(&vm_cnt.v_free_count); 1736 } 1737 } 1738 1739 static void 1740 vm_pageout_worker(void *arg) 1741 { 1742 struct vm_domain *domain; 1743 int domidx, pass; 1744 bool target_met; 1745 1746 domidx = (uintptr_t)arg; 1747 domain = &vm_dom[domidx]; 1748 pass = 0; 1749 target_met = true; 1750 1751 /* 1752 * XXXKIB It could be useful to bind pageout daemon threads to 1753 * the cores belonging to the domain, from which vm_page_array 1754 * is allocated. 1755 */ 1756 1757 KASSERT(domain->vmd_segs != 0, ("domain without segments")); 1758 domain->vmd_last_active_scan = ticks; 1759 vm_pageout_init_marker(&domain->vmd_marker, PQ_INACTIVE); 1760 vm_pageout_init_marker(&domain->vmd_inacthead, PQ_INACTIVE); 1761 TAILQ_INSERT_HEAD(&domain->vmd_pagequeues[PQ_INACTIVE].pq_pl, 1762 &domain->vmd_inacthead, plinks.q); 1763 1764 /* 1765 * The pageout daemon worker is never done, so loop forever. 1766 */ 1767 while (TRUE) { 1768 mtx_lock(&vm_page_queue_free_mtx); 1769 1770 /* 1771 * Generally, after a level >= 1 scan, if there are enough 1772 * free pages to wakeup the waiters, then they are already 1773 * awake. A call to vm_page_free() during the scan awakened 1774 * them. However, in the following case, this wakeup serves 1775 * to bound the amount of time that a thread might wait. 1776 * Suppose a thread's call to vm_page_alloc() fails, but 1777 * before that thread calls VM_WAIT, enough pages are freed by 1778 * other threads to alleviate the free page shortage. The 1779 * thread will, nonetheless, wait until another page is freed 1780 * or this wakeup is performed. 1781 */ 1782 if (vm_pages_needed && !vm_page_count_min()) { 1783 vm_pages_needed = false; 1784 wakeup(&vm_cnt.v_free_count); 1785 } 1786 1787 /* 1788 * Do not clear vm_pageout_wanted until we reach our free page 1789 * target. Otherwise, we may be awakened over and over again, 1790 * wasting CPU time. 1791 */ 1792 if (vm_pageout_wanted && target_met) 1793 vm_pageout_wanted = false; 1794 1795 /* 1796 * Might the page daemon receive a wakeup call? 1797 */ 1798 if (vm_pageout_wanted) { 1799 /* 1800 * No. Either vm_pageout_wanted was set by another 1801 * thread during the previous scan, which must have 1802 * been a level 0 scan, or vm_pageout_wanted was 1803 * already set and the scan failed to free enough 1804 * pages. If we haven't yet performed a level >= 1 1805 * (page reclamation) scan, then increase the level 1806 * and scan again now. Otherwise, sleep a bit and 1807 * try again later. 1808 */ 1809 mtx_unlock(&vm_page_queue_free_mtx); 1810 if (pass >= 1) 1811 pause("psleep", hz / VM_INACT_SCAN_RATE); 1812 pass++; 1813 } else { 1814 /* 1815 * Yes. Sleep until pages need to be reclaimed or 1816 * have their reference stats updated. 1817 */ 1818 if (mtx_sleep(&vm_pageout_wanted, 1819 &vm_page_queue_free_mtx, PDROP | PVM, "psleep", 1820 hz) == 0) { 1821 VM_CNT_INC(v_pdwakeups); 1822 pass = 1; 1823 } else 1824 pass = 0; 1825 } 1826 1827 target_met = vm_pageout_scan(domain, pass); 1828 } 1829 } 1830 1831 /* 1832 * vm_pageout_init initialises basic pageout daemon settings. 1833 */ 1834 static void 1835 vm_pageout_init(void) 1836 { 1837 /* 1838 * Initialize some paging parameters. 1839 */ 1840 vm_cnt.v_interrupt_free_min = 2; 1841 if (vm_cnt.v_page_count < 2000) 1842 vm_pageout_page_count = 8; 1843 1844 /* 1845 * v_free_reserved needs to include enough for the largest 1846 * swap pager structures plus enough for any pv_entry structs 1847 * when paging. 1848 */ 1849 if (vm_cnt.v_page_count > 1024) 1850 vm_cnt.v_free_min = 4 + (vm_cnt.v_page_count - 1024) / 200; 1851 else 1852 vm_cnt.v_free_min = 4; 1853 vm_cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1854 vm_cnt.v_interrupt_free_min; 1855 vm_cnt.v_free_reserved = vm_pageout_page_count + 1856 vm_cnt.v_pageout_free_min + (vm_cnt.v_page_count / 768); 1857 vm_cnt.v_free_severe = vm_cnt.v_free_min / 2; 1858 vm_cnt.v_free_target = 4 * vm_cnt.v_free_min + vm_cnt.v_free_reserved; 1859 vm_cnt.v_free_min += vm_cnt.v_free_reserved; 1860 vm_cnt.v_free_severe += vm_cnt.v_free_reserved; 1861 vm_cnt.v_inactive_target = (3 * vm_cnt.v_free_target) / 2; 1862 if (vm_cnt.v_inactive_target > vm_cnt.v_free_count / 3) 1863 vm_cnt.v_inactive_target = vm_cnt.v_free_count / 3; 1864 1865 /* 1866 * Set the default wakeup threshold to be 10% above the minimum 1867 * page limit. This keeps the steady state out of shortfall. 1868 */ 1869 vm_pageout_wakeup_thresh = (vm_cnt.v_free_min / 10) * 11; 1870 1871 /* 1872 * Set interval in seconds for active scan. We want to visit each 1873 * page at least once every ten minutes. This is to prevent worst 1874 * case paging behaviors with stale active LRU. 1875 */ 1876 if (vm_pageout_update_period == 0) 1877 vm_pageout_update_period = 600; 1878 1879 /* XXX does not really belong here */ 1880 if (vm_page_max_wired == 0) 1881 vm_page_max_wired = vm_cnt.v_free_count / 3; 1882 1883 /* 1884 * Target amount of memory to move out of the laundry queue during a 1885 * background laundering. This is proportional to the amount of system 1886 * memory. 1887 */ 1888 vm_background_launder_target = (vm_cnt.v_free_target - 1889 vm_cnt.v_free_min) / 10; 1890 } 1891 1892 /* 1893 * vm_pageout is the high level pageout daemon. 1894 */ 1895 static void 1896 vm_pageout(void) 1897 { 1898 int error; 1899 #ifdef VM_NUMA_ALLOC 1900 int i; 1901 #endif 1902 1903 swap_pager_swap_init(); 1904 error = kthread_add(vm_pageout_laundry_worker, NULL, curproc, NULL, 1905 0, 0, "laundry: dom0"); 1906 if (error != 0) 1907 panic("starting laundry for domain 0, error %d", error); 1908 #ifdef VM_NUMA_ALLOC 1909 for (i = 1; i < vm_ndomains; i++) { 1910 error = kthread_add(vm_pageout_worker, (void *)(uintptr_t)i, 1911 curproc, NULL, 0, 0, "dom%d", i); 1912 if (error != 0) { 1913 panic("starting pageout for domain %d, error %d\n", 1914 i, error); 1915 } 1916 } 1917 #endif 1918 error = kthread_add(uma_reclaim_worker, NULL, curproc, NULL, 1919 0, 0, "uma"); 1920 if (error != 0) 1921 panic("starting uma_reclaim helper, error %d\n", error); 1922 vm_pageout_worker((void *)(uintptr_t)0); 1923 } 1924 1925 /* 1926 * Unless the free page queue lock is held by the caller, this function 1927 * should be regarded as advisory. Specifically, the caller should 1928 * not msleep() on &vm_cnt.v_free_count following this function unless 1929 * the free page queue lock is held until the msleep() is performed. 1930 */ 1931 void 1932 pagedaemon_wakeup(void) 1933 { 1934 1935 if (!vm_pageout_wanted && curthread->td_proc != pageproc) { 1936 vm_pageout_wanted = true; 1937 wakeup(&vm_pageout_wanted); 1938 } 1939 } 1940