1 /*- 2 * SPDX-License-Identifier: (BSD-4-Clause AND MIT-CMU) 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/blockcount.h> 86 #include <sys/eventhandler.h> 87 #include <sys/lock.h> 88 #include <sys/mutex.h> 89 #include <sys/proc.h> 90 #include <sys/kthread.h> 91 #include <sys/ktr.h> 92 #include <sys/mount.h> 93 #include <sys/racct.h> 94 #include <sys/resourcevar.h> 95 #include <sys/sched.h> 96 #include <sys/sdt.h> 97 #include <sys/signalvar.h> 98 #include <sys/smp.h> 99 #include <sys/time.h> 100 #include <sys/vnode.h> 101 #include <sys/vmmeter.h> 102 #include <sys/rwlock.h> 103 #include <sys/sx.h> 104 #include <sys/sysctl.h> 105 106 #include <vm/vm.h> 107 #include <vm/vm_param.h> 108 #include <vm/vm_object.h> 109 #include <vm/vm_page.h> 110 #include <vm/vm_map.h> 111 #include <vm/vm_pageout.h> 112 #include <vm/vm_pager.h> 113 #include <vm/vm_phys.h> 114 #include <vm/vm_pagequeue.h> 115 #include <vm/swap_pager.h> 116 #include <vm/vm_extern.h> 117 #include <vm/uma.h> 118 119 /* 120 * System initialization 121 */ 122 123 /* the kernel process "vm_pageout"*/ 124 static void vm_pageout(void); 125 static void vm_pageout_init(void); 126 static int vm_pageout_clean(vm_page_t m, int *numpagedout); 127 static int vm_pageout_cluster(vm_page_t m); 128 static void vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage, 129 int starting_page_shortage); 130 131 SYSINIT(pagedaemon_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, vm_pageout_init, 132 NULL); 133 134 struct proc *pageproc; 135 136 static struct kproc_desc page_kp = { 137 "pagedaemon", 138 vm_pageout, 139 &pageproc 140 }; 141 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start, 142 &page_kp); 143 144 SDT_PROVIDER_DEFINE(vm); 145 SDT_PROBE_DEFINE(vm, , , vm__lowmem_scan); 146 147 /* Pagedaemon activity rates, in subdivisions of one second. */ 148 #define VM_LAUNDER_RATE 10 149 #define VM_INACT_SCAN_RATE 10 150 151 static int vm_pageout_oom_seq = 12; 152 153 static int vm_pageout_update_period; 154 static int disable_swap_pageouts; 155 static int lowmem_period = 10; 156 static int swapdev_enabled; 157 158 static int vm_panic_on_oom = 0; 159 160 SYSCTL_INT(_vm, OID_AUTO, panic_on_oom, 161 CTLFLAG_RWTUN, &vm_panic_on_oom, 0, 162 "Panic on the given number of out-of-memory errors instead of killing the largest process"); 163 164 SYSCTL_INT(_vm, OID_AUTO, pageout_update_period, 165 CTLFLAG_RWTUN, &vm_pageout_update_period, 0, 166 "Maximum active LRU update period"); 167 168 static int pageout_cpus_per_thread = 16; 169 SYSCTL_INT(_vm, OID_AUTO, pageout_cpus_per_thread, CTLFLAG_RDTUN, 170 &pageout_cpus_per_thread, 0, 171 "Number of CPUs per pagedaemon worker thread"); 172 173 SYSCTL_INT(_vm, OID_AUTO, lowmem_period, CTLFLAG_RWTUN, &lowmem_period, 0, 174 "Low memory callback period"); 175 176 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts, 177 CTLFLAG_RWTUN, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages"); 178 179 static int pageout_lock_miss; 180 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss, 181 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout"); 182 183 SYSCTL_INT(_vm, OID_AUTO, pageout_oom_seq, 184 CTLFLAG_RWTUN, &vm_pageout_oom_seq, 0, 185 "back-to-back calls to oom detector to start OOM"); 186 187 static int act_scan_laundry_weight = 3; 188 SYSCTL_INT(_vm, OID_AUTO, act_scan_laundry_weight, CTLFLAG_RWTUN, 189 &act_scan_laundry_weight, 0, 190 "weight given to clean vs. dirty pages in active queue scans"); 191 192 static u_int vm_background_launder_rate = 4096; 193 SYSCTL_UINT(_vm, OID_AUTO, background_launder_rate, CTLFLAG_RWTUN, 194 &vm_background_launder_rate, 0, 195 "background laundering rate, in kilobytes per second"); 196 197 static u_int vm_background_launder_max = 20 * 1024; 198 SYSCTL_UINT(_vm, OID_AUTO, background_launder_max, CTLFLAG_RWTUN, 199 &vm_background_launder_max, 0, "background laundering cap, in kilobytes"); 200 201 int vm_pageout_page_count = 32; 202 203 u_long vm_page_max_user_wired; 204 SYSCTL_ULONG(_vm, OID_AUTO, max_user_wired, CTLFLAG_RW, 205 &vm_page_max_user_wired, 0, 206 "system-wide limit to user-wired page count"); 207 208 static u_int isqrt(u_int num); 209 static int vm_pageout_launder(struct vm_domain *vmd, int launder, 210 bool in_shortfall); 211 static void vm_pageout_laundry_worker(void *arg); 212 213 struct scan_state { 214 struct vm_batchqueue bq; 215 struct vm_pagequeue *pq; 216 vm_page_t marker; 217 int maxscan; 218 int scanned; 219 }; 220 221 static void 222 vm_pageout_init_scan(struct scan_state *ss, struct vm_pagequeue *pq, 223 vm_page_t marker, vm_page_t after, int maxscan) 224 { 225 226 vm_pagequeue_assert_locked(pq); 227 KASSERT((marker->a.flags & PGA_ENQUEUED) == 0, 228 ("marker %p already enqueued", marker)); 229 230 if (after == NULL) 231 TAILQ_INSERT_HEAD(&pq->pq_pl, marker, plinks.q); 232 else 233 TAILQ_INSERT_AFTER(&pq->pq_pl, after, marker, plinks.q); 234 vm_page_aflag_set(marker, PGA_ENQUEUED); 235 236 vm_batchqueue_init(&ss->bq); 237 ss->pq = pq; 238 ss->marker = marker; 239 ss->maxscan = maxscan; 240 ss->scanned = 0; 241 vm_pagequeue_unlock(pq); 242 } 243 244 static void 245 vm_pageout_end_scan(struct scan_state *ss) 246 { 247 struct vm_pagequeue *pq; 248 249 pq = ss->pq; 250 vm_pagequeue_assert_locked(pq); 251 KASSERT((ss->marker->a.flags & PGA_ENQUEUED) != 0, 252 ("marker %p not enqueued", ss->marker)); 253 254 TAILQ_REMOVE(&pq->pq_pl, ss->marker, plinks.q); 255 vm_page_aflag_clear(ss->marker, PGA_ENQUEUED); 256 pq->pq_pdpages += ss->scanned; 257 } 258 259 /* 260 * Add a small number of queued pages to a batch queue for later processing 261 * without the corresponding queue lock held. The caller must have enqueued a 262 * marker page at the desired start point for the scan. Pages will be 263 * physically dequeued if the caller so requests. Otherwise, the returned 264 * batch may contain marker pages, and it is up to the caller to handle them. 265 * 266 * When processing the batch queue, vm_pageout_defer() must be used to 267 * determine whether the page has been logically dequeued since the batch was 268 * collected. 269 */ 270 static __always_inline void 271 vm_pageout_collect_batch(struct scan_state *ss, const bool dequeue) 272 { 273 struct vm_pagequeue *pq; 274 vm_page_t m, marker, n; 275 276 marker = ss->marker; 277 pq = ss->pq; 278 279 KASSERT((marker->a.flags & PGA_ENQUEUED) != 0, 280 ("marker %p not enqueued", ss->marker)); 281 282 vm_pagequeue_lock(pq); 283 for (m = TAILQ_NEXT(marker, plinks.q); m != NULL && 284 ss->scanned < ss->maxscan && ss->bq.bq_cnt < VM_BATCHQUEUE_SIZE; 285 m = n, ss->scanned++) { 286 n = TAILQ_NEXT(m, plinks.q); 287 if ((m->flags & PG_MARKER) == 0) { 288 KASSERT((m->a.flags & PGA_ENQUEUED) != 0, 289 ("page %p not enqueued", m)); 290 KASSERT((m->flags & PG_FICTITIOUS) == 0, 291 ("Fictitious page %p cannot be in page queue", m)); 292 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 293 ("Unmanaged page %p cannot be in page queue", m)); 294 } else if (dequeue) 295 continue; 296 297 (void)vm_batchqueue_insert(&ss->bq, m); 298 if (dequeue) { 299 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 300 vm_page_aflag_clear(m, PGA_ENQUEUED); 301 } 302 } 303 TAILQ_REMOVE(&pq->pq_pl, marker, plinks.q); 304 if (__predict_true(m != NULL)) 305 TAILQ_INSERT_BEFORE(m, marker, plinks.q); 306 else 307 TAILQ_INSERT_TAIL(&pq->pq_pl, marker, plinks.q); 308 if (dequeue) 309 vm_pagequeue_cnt_add(pq, -ss->bq.bq_cnt); 310 vm_pagequeue_unlock(pq); 311 } 312 313 /* 314 * Return the next page to be scanned, or NULL if the scan is complete. 315 */ 316 static __always_inline vm_page_t 317 vm_pageout_next(struct scan_state *ss, const bool dequeue) 318 { 319 320 if (ss->bq.bq_cnt == 0) 321 vm_pageout_collect_batch(ss, dequeue); 322 return (vm_batchqueue_pop(&ss->bq)); 323 } 324 325 /* 326 * Determine whether processing of a page should be deferred and ensure that any 327 * outstanding queue operations are processed. 328 */ 329 static __always_inline bool 330 vm_pageout_defer(vm_page_t m, const uint8_t queue, const bool enqueued) 331 { 332 vm_page_astate_t as; 333 334 as = vm_page_astate_load(m); 335 if (__predict_false(as.queue != queue || 336 ((as.flags & PGA_ENQUEUED) != 0) != enqueued)) 337 return (true); 338 if ((as.flags & PGA_QUEUE_OP_MASK) != 0) { 339 vm_page_pqbatch_submit(m, queue); 340 return (true); 341 } 342 return (false); 343 } 344 345 /* 346 * Scan for pages at adjacent offsets within the given page's object that are 347 * eligible for laundering, form a cluster of these pages and the given page, 348 * and launder that cluster. 349 */ 350 static int 351 vm_pageout_cluster(vm_page_t m) 352 { 353 vm_object_t object; 354 vm_page_t mc[2 * vm_pageout_page_count], p, pb, ps; 355 vm_pindex_t pindex; 356 int ib, is, page_base, pageout_count; 357 358 object = m->object; 359 VM_OBJECT_ASSERT_WLOCKED(object); 360 pindex = m->pindex; 361 362 vm_page_assert_xbusied(m); 363 364 mc[vm_pageout_page_count] = pb = ps = m; 365 pageout_count = 1; 366 page_base = vm_pageout_page_count; 367 ib = 1; 368 is = 1; 369 370 /* 371 * We can cluster only if the page is not clean, busy, or held, and 372 * the page is in the laundry queue. 373 * 374 * During heavy mmap/modification loads the pageout 375 * daemon can really fragment the underlying file 376 * due to flushing pages out of order and not trying to 377 * align the clusters (which leaves sporadic out-of-order 378 * holes). To solve this problem we do the reverse scan 379 * first and attempt to align our cluster, then do a 380 * forward scan if room remains. 381 */ 382 more: 383 while (ib != 0 && pageout_count < vm_pageout_page_count) { 384 if (ib > pindex) { 385 ib = 0; 386 break; 387 } 388 if ((p = vm_page_prev(pb)) == NULL || 389 vm_page_tryxbusy(p) == 0) { 390 ib = 0; 391 break; 392 } 393 if (vm_page_wired(p)) { 394 ib = 0; 395 vm_page_xunbusy(p); 396 break; 397 } 398 vm_page_test_dirty(p); 399 if (p->dirty == 0) { 400 ib = 0; 401 vm_page_xunbusy(p); 402 break; 403 } 404 if (!vm_page_in_laundry(p) || !vm_page_try_remove_write(p)) { 405 vm_page_xunbusy(p); 406 ib = 0; 407 break; 408 } 409 mc[--page_base] = pb = p; 410 ++pageout_count; 411 ++ib; 412 413 /* 414 * We are at an alignment boundary. Stop here, and switch 415 * directions. Do not clear ib. 416 */ 417 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0) 418 break; 419 } 420 while (pageout_count < vm_pageout_page_count && 421 pindex + is < object->size) { 422 if ((p = vm_page_next(ps)) == NULL || 423 vm_page_tryxbusy(p) == 0) 424 break; 425 if (vm_page_wired(p)) { 426 vm_page_xunbusy(p); 427 break; 428 } 429 vm_page_test_dirty(p); 430 if (p->dirty == 0) { 431 vm_page_xunbusy(p); 432 break; 433 } 434 if (!vm_page_in_laundry(p) || !vm_page_try_remove_write(p)) { 435 vm_page_xunbusy(p); 436 break; 437 } 438 mc[page_base + pageout_count] = ps = p; 439 ++pageout_count; 440 ++is; 441 } 442 443 /* 444 * If we exhausted our forward scan, continue with the reverse scan 445 * when possible, even past an alignment boundary. This catches 446 * boundary conditions. 447 */ 448 if (ib != 0 && pageout_count < vm_pageout_page_count) 449 goto more; 450 451 return (vm_pageout_flush(&mc[page_base], pageout_count, 452 VM_PAGER_PUT_NOREUSE, 0, NULL, NULL)); 453 } 454 455 /* 456 * vm_pageout_flush() - launder the given pages 457 * 458 * The given pages are laundered. Note that we setup for the start of 459 * I/O ( i.e. busy the page ), mark it read-only, and bump the object 460 * reference count all in here rather then in the parent. If we want 461 * the parent to do more sophisticated things we may have to change 462 * the ordering. 463 * 464 * Returned runlen is the count of pages between mreq and first 465 * page after mreq with status VM_PAGER_AGAIN. 466 * *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL 467 * for any page in runlen set. 468 */ 469 int 470 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen, 471 boolean_t *eio) 472 { 473 vm_object_t object = mc[0]->object; 474 int pageout_status[count]; 475 int numpagedout = 0; 476 int i, runlen; 477 478 VM_OBJECT_ASSERT_WLOCKED(object); 479 480 /* 481 * Initiate I/O. Mark the pages shared busy and verify that they're 482 * valid and read-only. 483 * 484 * We do not have to fixup the clean/dirty bits here... we can 485 * allow the pager to do it after the I/O completes. 486 * 487 * NOTE! mc[i]->dirty may be partial or fragmented due to an 488 * edge case with file fragments. 489 */ 490 for (i = 0; i < count; i++) { 491 KASSERT(vm_page_all_valid(mc[i]), 492 ("vm_pageout_flush: partially invalid page %p index %d/%d", 493 mc[i], i, count)); 494 KASSERT((mc[i]->a.flags & PGA_WRITEABLE) == 0, 495 ("vm_pageout_flush: writeable page %p", mc[i])); 496 vm_page_busy_downgrade(mc[i]); 497 } 498 vm_object_pip_add(object, count); 499 500 vm_pager_put_pages(object, mc, count, flags, pageout_status); 501 502 runlen = count - mreq; 503 if (eio != NULL) 504 *eio = FALSE; 505 for (i = 0; i < count; i++) { 506 vm_page_t mt = mc[i]; 507 508 KASSERT(pageout_status[i] == VM_PAGER_PEND || 509 !pmap_page_is_write_mapped(mt), 510 ("vm_pageout_flush: page %p is not write protected", mt)); 511 switch (pageout_status[i]) { 512 case VM_PAGER_OK: 513 /* 514 * The page may have moved since laundering started, in 515 * which case it should be left alone. 516 */ 517 if (vm_page_in_laundry(mt)) 518 vm_page_deactivate_noreuse(mt); 519 /* FALLTHROUGH */ 520 case VM_PAGER_PEND: 521 numpagedout++; 522 break; 523 case VM_PAGER_BAD: 524 /* 525 * The page is outside the object's range. We pretend 526 * that the page out worked and clean the page, so the 527 * changes will be lost if the page is reclaimed by 528 * the page daemon. 529 */ 530 vm_page_undirty(mt); 531 if (vm_page_in_laundry(mt)) 532 vm_page_deactivate_noreuse(mt); 533 break; 534 case VM_PAGER_ERROR: 535 case VM_PAGER_FAIL: 536 /* 537 * If the page couldn't be paged out to swap because the 538 * pager wasn't able to find space, place the page in 539 * the PQ_UNSWAPPABLE holding queue. This is an 540 * optimization that prevents the page daemon from 541 * wasting CPU cycles on pages that cannot be reclaimed 542 * becase no swap device is configured. 543 * 544 * Otherwise, reactivate the page so that it doesn't 545 * clog the laundry and inactive queues. (We will try 546 * paging it out again later.) 547 */ 548 if (object->type == OBJT_SWAP && 549 pageout_status[i] == VM_PAGER_FAIL) { 550 vm_page_unswappable(mt); 551 numpagedout++; 552 } else 553 vm_page_activate(mt); 554 if (eio != NULL && i >= mreq && i - mreq < runlen) 555 *eio = TRUE; 556 break; 557 case VM_PAGER_AGAIN: 558 if (i >= mreq && i - mreq < runlen) 559 runlen = i - mreq; 560 break; 561 } 562 563 /* 564 * If the operation is still going, leave the page busy to 565 * block all other accesses. Also, leave the paging in 566 * progress indicator set so that we don't attempt an object 567 * collapse. 568 */ 569 if (pageout_status[i] != VM_PAGER_PEND) { 570 vm_object_pip_wakeup(object); 571 vm_page_sunbusy(mt); 572 } 573 } 574 if (prunlen != NULL) 575 *prunlen = runlen; 576 return (numpagedout); 577 } 578 579 static void 580 vm_pageout_swapon(void *arg __unused, struct swdevt *sp __unused) 581 { 582 583 atomic_store_rel_int(&swapdev_enabled, 1); 584 } 585 586 static void 587 vm_pageout_swapoff(void *arg __unused, struct swdevt *sp __unused) 588 { 589 590 if (swap_pager_nswapdev() == 1) 591 atomic_store_rel_int(&swapdev_enabled, 0); 592 } 593 594 /* 595 * Attempt to acquire all of the necessary locks to launder a page and 596 * then call through the clustering layer to PUTPAGES. Wait a short 597 * time for a vnode lock. 598 * 599 * Requires the page and object lock on entry, releases both before return. 600 * Returns 0 on success and an errno otherwise. 601 */ 602 static int 603 vm_pageout_clean(vm_page_t m, int *numpagedout) 604 { 605 struct vnode *vp; 606 struct mount *mp; 607 vm_object_t object; 608 vm_pindex_t pindex; 609 int error, lockmode; 610 611 object = m->object; 612 VM_OBJECT_ASSERT_WLOCKED(object); 613 error = 0; 614 vp = NULL; 615 mp = NULL; 616 617 /* 618 * The object is already known NOT to be dead. It 619 * is possible for the vget() to block the whole 620 * pageout daemon, but the new low-memory handling 621 * code should prevent it. 622 * 623 * We can't wait forever for the vnode lock, we might 624 * deadlock due to a vn_read() getting stuck in 625 * vm_wait while holding this vnode. We skip the 626 * vnode if we can't get it in a reasonable amount 627 * of time. 628 */ 629 if (object->type == OBJT_VNODE) { 630 vm_page_xunbusy(m); 631 vp = object->handle; 632 if (vp->v_type == VREG && 633 vn_start_write(vp, &mp, V_NOWAIT) != 0) { 634 mp = NULL; 635 error = EDEADLK; 636 goto unlock_all; 637 } 638 KASSERT(mp != NULL, 639 ("vp %p with NULL v_mount", vp)); 640 vm_object_reference_locked(object); 641 pindex = m->pindex; 642 VM_OBJECT_WUNLOCK(object); 643 lockmode = MNT_SHARED_WRITES(vp->v_mount) ? 644 LK_SHARED : LK_EXCLUSIVE; 645 if (vget(vp, lockmode | LK_TIMELOCK)) { 646 vp = NULL; 647 error = EDEADLK; 648 goto unlock_mp; 649 } 650 VM_OBJECT_WLOCK(object); 651 652 /* 653 * Ensure that the object and vnode were not disassociated 654 * while locks were dropped. 655 */ 656 if (vp->v_object != object) { 657 error = ENOENT; 658 goto unlock_all; 659 } 660 661 /* 662 * While the object was unlocked, the page may have been: 663 * (1) moved to a different queue, 664 * (2) reallocated to a different object, 665 * (3) reallocated to a different offset, or 666 * (4) cleaned. 667 */ 668 if (!vm_page_in_laundry(m) || m->object != object || 669 m->pindex != pindex || m->dirty == 0) { 670 error = ENXIO; 671 goto unlock_all; 672 } 673 674 /* 675 * The page may have been busied while the object lock was 676 * released. 677 */ 678 if (vm_page_tryxbusy(m) == 0) { 679 error = EBUSY; 680 goto unlock_all; 681 } 682 } 683 684 /* 685 * Remove all writeable mappings, failing if the page is wired. 686 */ 687 if (!vm_page_try_remove_write(m)) { 688 vm_page_xunbusy(m); 689 error = EBUSY; 690 goto unlock_all; 691 } 692 693 /* 694 * If a page is dirty, then it is either being washed 695 * (but not yet cleaned) or it is still in the 696 * laundry. If it is still in the laundry, then we 697 * start the cleaning operation. 698 */ 699 if ((*numpagedout = vm_pageout_cluster(m)) == 0) 700 error = EIO; 701 702 unlock_all: 703 VM_OBJECT_WUNLOCK(object); 704 705 unlock_mp: 706 if (mp != NULL) { 707 if (vp != NULL) 708 vput(vp); 709 vm_object_deallocate(object); 710 vn_finished_write(mp); 711 } 712 713 return (error); 714 } 715 716 /* 717 * Attempt to launder the specified number of pages. 718 * 719 * Returns the number of pages successfully laundered. 720 */ 721 static int 722 vm_pageout_launder(struct vm_domain *vmd, int launder, bool in_shortfall) 723 { 724 struct scan_state ss; 725 struct vm_pagequeue *pq; 726 vm_object_t object; 727 vm_page_t m, marker; 728 vm_page_astate_t new, old; 729 int act_delta, error, numpagedout, queue, refs, starting_target; 730 int vnodes_skipped; 731 bool pageout_ok; 732 733 object = NULL; 734 starting_target = launder; 735 vnodes_skipped = 0; 736 737 /* 738 * Scan the laundry queues for pages eligible to be laundered. We stop 739 * once the target number of dirty pages have been laundered, or once 740 * we've reached the end of the queue. A single iteration of this loop 741 * may cause more than one page to be laundered because of clustering. 742 * 743 * As an optimization, we avoid laundering from PQ_UNSWAPPABLE when no 744 * swap devices are configured. 745 */ 746 if (atomic_load_acq_int(&swapdev_enabled)) 747 queue = PQ_UNSWAPPABLE; 748 else 749 queue = PQ_LAUNDRY; 750 751 scan: 752 marker = &vmd->vmd_markers[queue]; 753 pq = &vmd->vmd_pagequeues[queue]; 754 vm_pagequeue_lock(pq); 755 vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt); 756 while (launder > 0 && (m = vm_pageout_next(&ss, false)) != NULL) { 757 if (__predict_false((m->flags & PG_MARKER) != 0)) 758 continue; 759 760 /* 761 * Don't touch a page that was removed from the queue after the 762 * page queue lock was released. Otherwise, ensure that any 763 * pending queue operations, such as dequeues for wired pages, 764 * are handled. 765 */ 766 if (vm_pageout_defer(m, queue, true)) 767 continue; 768 769 /* 770 * Lock the page's object. 771 */ 772 if (object == NULL || object != m->object) { 773 if (object != NULL) 774 VM_OBJECT_WUNLOCK(object); 775 object = atomic_load_ptr(&m->object); 776 if (__predict_false(object == NULL)) 777 /* The page is being freed by another thread. */ 778 continue; 779 780 /* Depends on type-stability. */ 781 VM_OBJECT_WLOCK(object); 782 if (__predict_false(m->object != object)) { 783 VM_OBJECT_WUNLOCK(object); 784 object = NULL; 785 continue; 786 } 787 } 788 789 if (vm_page_tryxbusy(m) == 0) 790 continue; 791 792 /* 793 * Check for wirings now that we hold the object lock and have 794 * exclusively busied the page. If the page is mapped, it may 795 * still be wired by pmap lookups. The call to 796 * vm_page_try_remove_all() below atomically checks for such 797 * wirings and removes mappings. If the page is unmapped, the 798 * wire count is guaranteed not to increase after this check. 799 */ 800 if (__predict_false(vm_page_wired(m))) 801 goto skip_page; 802 803 /* 804 * Invalid pages can be easily freed. They cannot be 805 * mapped; vm_page_free() asserts this. 806 */ 807 if (vm_page_none_valid(m)) 808 goto free_page; 809 810 refs = object->ref_count != 0 ? pmap_ts_referenced(m) : 0; 811 812 for (old = vm_page_astate_load(m);;) { 813 /* 814 * Check to see if the page has been removed from the 815 * queue since the first such check. Leave it alone if 816 * so, discarding any references collected by 817 * pmap_ts_referenced(). 818 */ 819 if (__predict_false(_vm_page_queue(old) == PQ_NONE)) 820 goto skip_page; 821 822 new = old; 823 act_delta = refs; 824 if ((old.flags & PGA_REFERENCED) != 0) { 825 new.flags &= ~PGA_REFERENCED; 826 act_delta++; 827 } 828 if (act_delta == 0) { 829 ; 830 } else if (object->ref_count != 0) { 831 /* 832 * Increase the activation count if the page was 833 * referenced while in the laundry queue. This 834 * makes it less likely that the page will be 835 * returned prematurely to the laundry queue. 836 */ 837 new.act_count += ACT_ADVANCE + 838 act_delta; 839 if (new.act_count > ACT_MAX) 840 new.act_count = ACT_MAX; 841 842 new.flags &= ~PGA_QUEUE_OP_MASK; 843 new.flags |= PGA_REQUEUE; 844 new.queue = PQ_ACTIVE; 845 if (!vm_page_pqstate_commit(m, &old, new)) 846 continue; 847 848 /* 849 * If this was a background laundering, count 850 * activated pages towards our target. The 851 * purpose of background laundering is to ensure 852 * that pages are eventually cycled through the 853 * laundry queue, and an activation is a valid 854 * way out. 855 */ 856 if (!in_shortfall) 857 launder--; 858 VM_CNT_INC(v_reactivated); 859 goto skip_page; 860 } else if ((object->flags & OBJ_DEAD) == 0) { 861 new.flags |= PGA_REQUEUE; 862 if (!vm_page_pqstate_commit(m, &old, new)) 863 continue; 864 goto skip_page; 865 } 866 break; 867 } 868 869 /* 870 * If the page appears to be clean at the machine-independent 871 * layer, then remove all of its mappings from the pmap in 872 * anticipation of freeing it. If, however, any of the page's 873 * mappings allow write access, then the page may still be 874 * modified until the last of those mappings are removed. 875 */ 876 if (object->ref_count != 0) { 877 vm_page_test_dirty(m); 878 if (m->dirty == 0 && !vm_page_try_remove_all(m)) 879 goto skip_page; 880 } 881 882 /* 883 * Clean pages are freed, and dirty pages are paged out unless 884 * they belong to a dead object. Requeueing dirty pages from 885 * dead objects is pointless, as they are being paged out and 886 * freed by the thread that destroyed the object. 887 */ 888 if (m->dirty == 0) { 889 free_page: 890 /* 891 * Now we are guaranteed that no other threads are 892 * manipulating the page, check for a last-second 893 * reference. 894 */ 895 if (vm_pageout_defer(m, queue, true)) 896 goto skip_page; 897 vm_page_free(m); 898 VM_CNT_INC(v_dfree); 899 } else if ((object->flags & OBJ_DEAD) == 0) { 900 if (object->type != OBJT_SWAP && 901 object->type != OBJT_DEFAULT) 902 pageout_ok = true; 903 else if (disable_swap_pageouts) 904 pageout_ok = false; 905 else 906 pageout_ok = true; 907 if (!pageout_ok) { 908 vm_page_launder(m); 909 goto skip_page; 910 } 911 912 /* 913 * Form a cluster with adjacent, dirty pages from the 914 * same object, and page out that entire cluster. 915 * 916 * The adjacent, dirty pages must also be in the 917 * laundry. However, their mappings are not checked 918 * for new references. Consequently, a recently 919 * referenced page may be paged out. However, that 920 * page will not be prematurely reclaimed. After page 921 * out, the page will be placed in the inactive queue, 922 * where any new references will be detected and the 923 * page reactivated. 924 */ 925 error = vm_pageout_clean(m, &numpagedout); 926 if (error == 0) { 927 launder -= numpagedout; 928 ss.scanned += numpagedout; 929 } else if (error == EDEADLK) { 930 pageout_lock_miss++; 931 vnodes_skipped++; 932 } 933 object = NULL; 934 } else { 935 skip_page: 936 vm_page_xunbusy(m); 937 } 938 } 939 if (object != NULL) { 940 VM_OBJECT_WUNLOCK(object); 941 object = NULL; 942 } 943 vm_pagequeue_lock(pq); 944 vm_pageout_end_scan(&ss); 945 vm_pagequeue_unlock(pq); 946 947 if (launder > 0 && queue == PQ_UNSWAPPABLE) { 948 queue = PQ_LAUNDRY; 949 goto scan; 950 } 951 952 /* 953 * Wakeup the sync daemon if we skipped a vnode in a writeable object 954 * and we didn't launder enough pages. 955 */ 956 if (vnodes_skipped > 0 && launder > 0) 957 (void)speedup_syncer(); 958 959 return (starting_target - launder); 960 } 961 962 /* 963 * Compute the integer square root. 964 */ 965 static u_int 966 isqrt(u_int num) 967 { 968 u_int bit, root, tmp; 969 970 bit = num != 0 ? (1u << ((fls(num) - 1) & ~1)) : 0; 971 root = 0; 972 while (bit != 0) { 973 tmp = root + bit; 974 root >>= 1; 975 if (num >= tmp) { 976 num -= tmp; 977 root += bit; 978 } 979 bit >>= 2; 980 } 981 return (root); 982 } 983 984 /* 985 * Perform the work of the laundry thread: periodically wake up and determine 986 * whether any pages need to be laundered. If so, determine the number of pages 987 * that need to be laundered, and launder them. 988 */ 989 static void 990 vm_pageout_laundry_worker(void *arg) 991 { 992 struct vm_domain *vmd; 993 struct vm_pagequeue *pq; 994 uint64_t nclean, ndirty, nfreed; 995 int domain, last_target, launder, shortfall, shortfall_cycle, target; 996 bool in_shortfall; 997 998 domain = (uintptr_t)arg; 999 vmd = VM_DOMAIN(domain); 1000 pq = &vmd->vmd_pagequeues[PQ_LAUNDRY]; 1001 KASSERT(vmd->vmd_segs != 0, ("domain without segments")); 1002 1003 shortfall = 0; 1004 in_shortfall = false; 1005 shortfall_cycle = 0; 1006 last_target = target = 0; 1007 nfreed = 0; 1008 1009 /* 1010 * Calls to these handlers are serialized by the swap syscall lock. 1011 */ 1012 (void)EVENTHANDLER_REGISTER(swapon, vm_pageout_swapon, vmd, 1013 EVENTHANDLER_PRI_ANY); 1014 (void)EVENTHANDLER_REGISTER(swapoff, vm_pageout_swapoff, vmd, 1015 EVENTHANDLER_PRI_ANY); 1016 1017 /* 1018 * The pageout laundry worker is never done, so loop forever. 1019 */ 1020 for (;;) { 1021 KASSERT(target >= 0, ("negative target %d", target)); 1022 KASSERT(shortfall_cycle >= 0, 1023 ("negative cycle %d", shortfall_cycle)); 1024 launder = 0; 1025 1026 /* 1027 * First determine whether we need to launder pages to meet a 1028 * shortage of free pages. 1029 */ 1030 if (shortfall > 0) { 1031 in_shortfall = true; 1032 shortfall_cycle = VM_LAUNDER_RATE / VM_INACT_SCAN_RATE; 1033 target = shortfall; 1034 } else if (!in_shortfall) 1035 goto trybackground; 1036 else if (shortfall_cycle == 0 || vm_laundry_target(vmd) <= 0) { 1037 /* 1038 * We recently entered shortfall and began laundering 1039 * pages. If we have completed that laundering run 1040 * (and we are no longer in shortfall) or we have met 1041 * our laundry target through other activity, then we 1042 * can stop laundering pages. 1043 */ 1044 in_shortfall = false; 1045 target = 0; 1046 goto trybackground; 1047 } 1048 launder = target / shortfall_cycle--; 1049 goto dolaundry; 1050 1051 /* 1052 * There's no immediate need to launder any pages; see if we 1053 * meet the conditions to perform background laundering: 1054 * 1055 * 1. The ratio of dirty to clean inactive pages exceeds the 1056 * background laundering threshold, or 1057 * 2. we haven't yet reached the target of the current 1058 * background laundering run. 1059 * 1060 * The background laundering threshold is not a constant. 1061 * Instead, it is a slowly growing function of the number of 1062 * clean pages freed by the page daemon since the last 1063 * background laundering. Thus, as the ratio of dirty to 1064 * clean inactive pages grows, the amount of memory pressure 1065 * required to trigger laundering decreases. We ensure 1066 * that the threshold is non-zero after an inactive queue 1067 * scan, even if that scan failed to free a single clean page. 1068 */ 1069 trybackground: 1070 nclean = vmd->vmd_free_count + 1071 vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt; 1072 ndirty = vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt; 1073 if (target == 0 && ndirty * isqrt(howmany(nfreed + 1, 1074 vmd->vmd_free_target - vmd->vmd_free_min)) >= nclean) { 1075 target = vmd->vmd_background_launder_target; 1076 } 1077 1078 /* 1079 * We have a non-zero background laundering target. If we've 1080 * laundered up to our maximum without observing a page daemon 1081 * request, just stop. This is a safety belt that ensures we 1082 * don't launder an excessive amount if memory pressure is low 1083 * and the ratio of dirty to clean pages is large. Otherwise, 1084 * proceed at the background laundering rate. 1085 */ 1086 if (target > 0) { 1087 if (nfreed > 0) { 1088 nfreed = 0; 1089 last_target = target; 1090 } else if (last_target - target >= 1091 vm_background_launder_max * PAGE_SIZE / 1024) { 1092 target = 0; 1093 } 1094 launder = vm_background_launder_rate * PAGE_SIZE / 1024; 1095 launder /= VM_LAUNDER_RATE; 1096 if (launder > target) 1097 launder = target; 1098 } 1099 1100 dolaundry: 1101 if (launder > 0) { 1102 /* 1103 * Because of I/O clustering, the number of laundered 1104 * pages could exceed "target" by the maximum size of 1105 * a cluster minus one. 1106 */ 1107 target -= min(vm_pageout_launder(vmd, launder, 1108 in_shortfall), target); 1109 pause("laundp", hz / VM_LAUNDER_RATE); 1110 } 1111 1112 /* 1113 * If we're not currently laundering pages and the page daemon 1114 * hasn't posted a new request, sleep until the page daemon 1115 * kicks us. 1116 */ 1117 vm_pagequeue_lock(pq); 1118 if (target == 0 && vmd->vmd_laundry_request == VM_LAUNDRY_IDLE) 1119 (void)mtx_sleep(&vmd->vmd_laundry_request, 1120 vm_pagequeue_lockptr(pq), PVM, "launds", 0); 1121 1122 /* 1123 * If the pagedaemon has indicated that it's in shortfall, start 1124 * a shortfall laundering unless we're already in the middle of 1125 * one. This may preempt a background laundering. 1126 */ 1127 if (vmd->vmd_laundry_request == VM_LAUNDRY_SHORTFALL && 1128 (!in_shortfall || shortfall_cycle == 0)) { 1129 shortfall = vm_laundry_target(vmd) + 1130 vmd->vmd_pageout_deficit; 1131 target = 0; 1132 } else 1133 shortfall = 0; 1134 1135 if (target == 0) 1136 vmd->vmd_laundry_request = VM_LAUNDRY_IDLE; 1137 nfreed += vmd->vmd_clean_pages_freed; 1138 vmd->vmd_clean_pages_freed = 0; 1139 vm_pagequeue_unlock(pq); 1140 } 1141 } 1142 1143 /* 1144 * Compute the number of pages we want to try to move from the 1145 * active queue to either the inactive or laundry queue. 1146 * 1147 * When scanning active pages during a shortage, we make clean pages 1148 * count more heavily towards the page shortage than dirty pages. 1149 * This is because dirty pages must be laundered before they can be 1150 * reused and thus have less utility when attempting to quickly 1151 * alleviate a free page shortage. However, this weighting also 1152 * causes the scan to deactivate dirty pages more aggressively, 1153 * improving the effectiveness of clustering. 1154 */ 1155 static int 1156 vm_pageout_active_target(struct vm_domain *vmd) 1157 { 1158 int shortage; 1159 1160 shortage = vmd->vmd_inactive_target + vm_paging_target(vmd) - 1161 (vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt + 1162 vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt / act_scan_laundry_weight); 1163 shortage *= act_scan_laundry_weight; 1164 return (shortage); 1165 } 1166 1167 /* 1168 * Scan the active queue. If there is no shortage of inactive pages, scan a 1169 * small portion of the queue in order to maintain quasi-LRU. 1170 */ 1171 static void 1172 vm_pageout_scan_active(struct vm_domain *vmd, int page_shortage) 1173 { 1174 struct scan_state ss; 1175 vm_object_t object; 1176 vm_page_t m, marker; 1177 struct vm_pagequeue *pq; 1178 vm_page_astate_t old, new; 1179 long min_scan; 1180 int act_delta, max_scan, ps_delta, refs, scan_tick; 1181 uint8_t nqueue; 1182 1183 marker = &vmd->vmd_markers[PQ_ACTIVE]; 1184 pq = &vmd->vmd_pagequeues[PQ_ACTIVE]; 1185 vm_pagequeue_lock(pq); 1186 1187 /* 1188 * If we're just idle polling attempt to visit every 1189 * active page within 'update_period' seconds. 1190 */ 1191 scan_tick = ticks; 1192 if (vm_pageout_update_period != 0) { 1193 min_scan = pq->pq_cnt; 1194 min_scan *= scan_tick - vmd->vmd_last_active_scan; 1195 min_scan /= hz * vm_pageout_update_period; 1196 } else 1197 min_scan = 0; 1198 if (min_scan > 0 || (page_shortage > 0 && pq->pq_cnt > 0)) 1199 vmd->vmd_last_active_scan = scan_tick; 1200 1201 /* 1202 * Scan the active queue for pages that can be deactivated. Update 1203 * the per-page activity counter and use it to identify deactivation 1204 * candidates. Held pages may be deactivated. 1205 * 1206 * To avoid requeuing each page that remains in the active queue, we 1207 * implement the CLOCK algorithm. To keep the implementation of the 1208 * enqueue operation consistent for all page queues, we use two hands, 1209 * represented by marker pages. Scans begin at the first hand, which 1210 * precedes the second hand in the queue. When the two hands meet, 1211 * they are moved back to the head and tail of the queue, respectively, 1212 * and scanning resumes. 1213 */ 1214 max_scan = page_shortage > 0 ? pq->pq_cnt : min_scan; 1215 act_scan: 1216 vm_pageout_init_scan(&ss, pq, marker, &vmd->vmd_clock[0], max_scan); 1217 while ((m = vm_pageout_next(&ss, false)) != NULL) { 1218 if (__predict_false(m == &vmd->vmd_clock[1])) { 1219 vm_pagequeue_lock(pq); 1220 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q); 1221 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[1], plinks.q); 1222 TAILQ_INSERT_HEAD(&pq->pq_pl, &vmd->vmd_clock[0], 1223 plinks.q); 1224 TAILQ_INSERT_TAIL(&pq->pq_pl, &vmd->vmd_clock[1], 1225 plinks.q); 1226 max_scan -= ss.scanned; 1227 vm_pageout_end_scan(&ss); 1228 goto act_scan; 1229 } 1230 if (__predict_false((m->flags & PG_MARKER) != 0)) 1231 continue; 1232 1233 /* 1234 * Don't touch a page that was removed from the queue after the 1235 * page queue lock was released. Otherwise, ensure that any 1236 * pending queue operations, such as dequeues for wired pages, 1237 * are handled. 1238 */ 1239 if (vm_pageout_defer(m, PQ_ACTIVE, true)) 1240 continue; 1241 1242 /* 1243 * A page's object pointer may be set to NULL before 1244 * the object lock is acquired. 1245 */ 1246 object = atomic_load_ptr(&m->object); 1247 if (__predict_false(object == NULL)) 1248 /* 1249 * The page has been removed from its object. 1250 */ 1251 continue; 1252 1253 /* Deferred free of swap space. */ 1254 if ((m->a.flags & PGA_SWAP_FREE) != 0 && 1255 VM_OBJECT_TRYWLOCK(object)) { 1256 if (m->object == object) 1257 vm_pager_page_unswapped(m); 1258 VM_OBJECT_WUNLOCK(object); 1259 } 1260 1261 /* 1262 * Check to see "how much" the page has been used. 1263 * 1264 * Test PGA_REFERENCED after calling pmap_ts_referenced() so 1265 * that a reference from a concurrently destroyed mapping is 1266 * observed here and now. 1267 * 1268 * Perform an unsynchronized object ref count check. While 1269 * the page lock ensures that the page is not reallocated to 1270 * another object, in particular, one with unmanaged mappings 1271 * that cannot support pmap_ts_referenced(), two races are, 1272 * nonetheless, possible: 1273 * 1) The count was transitioning to zero, but we saw a non- 1274 * zero value. pmap_ts_referenced() will return zero 1275 * because the page is not mapped. 1276 * 2) The count was transitioning to one, but we saw zero. 1277 * This race delays the detection of a new reference. At 1278 * worst, we will deactivate and reactivate the page. 1279 */ 1280 refs = object->ref_count != 0 ? pmap_ts_referenced(m) : 0; 1281 1282 old = vm_page_astate_load(m); 1283 do { 1284 /* 1285 * Check to see if the page has been removed from the 1286 * queue since the first such check. Leave it alone if 1287 * so, discarding any references collected by 1288 * pmap_ts_referenced(). 1289 */ 1290 if (__predict_false(_vm_page_queue(old) == PQ_NONE)) 1291 break; 1292 1293 /* 1294 * Advance or decay the act_count based on recent usage. 1295 */ 1296 new = old; 1297 act_delta = refs; 1298 if ((old.flags & PGA_REFERENCED) != 0) { 1299 new.flags &= ~PGA_REFERENCED; 1300 act_delta++; 1301 } 1302 if (act_delta != 0) { 1303 new.act_count += ACT_ADVANCE + act_delta; 1304 if (new.act_count > ACT_MAX) 1305 new.act_count = ACT_MAX; 1306 } else { 1307 new.act_count -= min(new.act_count, 1308 ACT_DECLINE); 1309 } 1310 1311 if (new.act_count > 0) { 1312 /* 1313 * Adjust the activation count and keep the page 1314 * in the active queue. The count might be left 1315 * unchanged if it is saturated. The page may 1316 * have been moved to a different queue since we 1317 * started the scan, in which case we move it 1318 * back. 1319 */ 1320 ps_delta = 0; 1321 if (old.queue != PQ_ACTIVE) { 1322 new.flags &= ~PGA_QUEUE_OP_MASK; 1323 new.flags |= PGA_REQUEUE; 1324 new.queue = PQ_ACTIVE; 1325 } 1326 } else { 1327 /* 1328 * When not short for inactive pages, let dirty 1329 * pages go through the inactive queue before 1330 * moving to the laundry queue. This gives them 1331 * some extra time to be reactivated, 1332 * potentially avoiding an expensive pageout. 1333 * However, during a page shortage, the inactive 1334 * queue is necessarily small, and so dirty 1335 * pages would only spend a trivial amount of 1336 * time in the inactive queue. Therefore, we 1337 * might as well place them directly in the 1338 * laundry queue to reduce queuing overhead. 1339 * 1340 * Calling vm_page_test_dirty() here would 1341 * require acquisition of the object's write 1342 * lock. However, during a page shortage, 1343 * directing dirty pages into the laundry queue 1344 * is only an optimization and not a 1345 * requirement. Therefore, we simply rely on 1346 * the opportunistic updates to the page's dirty 1347 * field by the pmap. 1348 */ 1349 if (page_shortage <= 0) { 1350 nqueue = PQ_INACTIVE; 1351 ps_delta = 0; 1352 } else if (m->dirty == 0) { 1353 nqueue = PQ_INACTIVE; 1354 ps_delta = act_scan_laundry_weight; 1355 } else { 1356 nqueue = PQ_LAUNDRY; 1357 ps_delta = 1; 1358 } 1359 1360 new.flags &= ~PGA_QUEUE_OP_MASK; 1361 new.flags |= PGA_REQUEUE; 1362 new.queue = nqueue; 1363 } 1364 } while (!vm_page_pqstate_commit(m, &old, new)); 1365 1366 page_shortage -= ps_delta; 1367 } 1368 vm_pagequeue_lock(pq); 1369 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q); 1370 TAILQ_INSERT_AFTER(&pq->pq_pl, marker, &vmd->vmd_clock[0], plinks.q); 1371 vm_pageout_end_scan(&ss); 1372 vm_pagequeue_unlock(pq); 1373 } 1374 1375 static int 1376 vm_pageout_reinsert_inactive_page(struct vm_pagequeue *pq, vm_page_t marker, 1377 vm_page_t m) 1378 { 1379 vm_page_astate_t as; 1380 1381 vm_pagequeue_assert_locked(pq); 1382 1383 as = vm_page_astate_load(m); 1384 if (as.queue != PQ_INACTIVE || (as.flags & PGA_ENQUEUED) != 0) 1385 return (0); 1386 vm_page_aflag_set(m, PGA_ENQUEUED); 1387 TAILQ_INSERT_BEFORE(marker, m, plinks.q); 1388 return (1); 1389 } 1390 1391 /* 1392 * Re-add stuck pages to the inactive queue. We will examine them again 1393 * during the next scan. If the queue state of a page has changed since 1394 * it was physically removed from the page queue in 1395 * vm_pageout_collect_batch(), don't do anything with that page. 1396 */ 1397 static void 1398 vm_pageout_reinsert_inactive(struct scan_state *ss, struct vm_batchqueue *bq, 1399 vm_page_t m) 1400 { 1401 struct vm_pagequeue *pq; 1402 vm_page_t marker; 1403 int delta; 1404 1405 delta = 0; 1406 marker = ss->marker; 1407 pq = ss->pq; 1408 1409 if (m != NULL) { 1410 if (vm_batchqueue_insert(bq, m)) 1411 return; 1412 vm_pagequeue_lock(pq); 1413 delta += vm_pageout_reinsert_inactive_page(pq, marker, m); 1414 } else 1415 vm_pagequeue_lock(pq); 1416 while ((m = vm_batchqueue_pop(bq)) != NULL) 1417 delta += vm_pageout_reinsert_inactive_page(pq, marker, m); 1418 vm_pagequeue_cnt_add(pq, delta); 1419 vm_pagequeue_unlock(pq); 1420 vm_batchqueue_init(bq); 1421 } 1422 1423 static void 1424 vm_pageout_scan_inactive(struct vm_domain *vmd, int page_shortage) 1425 { 1426 struct timeval start, end; 1427 struct scan_state ss; 1428 struct vm_batchqueue rq; 1429 struct vm_page marker_page; 1430 vm_page_t m, marker; 1431 struct vm_pagequeue *pq; 1432 vm_object_t object; 1433 vm_page_astate_t old, new; 1434 int act_delta, addl_page_shortage, starting_page_shortage, refs; 1435 1436 object = NULL; 1437 vm_batchqueue_init(&rq); 1438 getmicrouptime(&start); 1439 1440 /* 1441 * The addl_page_shortage is an estimate of the number of temporarily 1442 * stuck pages in the inactive queue. In other words, the 1443 * number of pages from the inactive count that should be 1444 * discounted in setting the target for the active queue scan. 1445 */ 1446 addl_page_shortage = 0; 1447 1448 /* 1449 * Start scanning the inactive queue for pages that we can free. The 1450 * scan will stop when we reach the target or we have scanned the 1451 * entire queue. (Note that m->a.act_count is not used to make 1452 * decisions for the inactive queue, only for the active queue.) 1453 */ 1454 starting_page_shortage = page_shortage; 1455 marker = &marker_page; 1456 vm_page_init_marker(marker, PQ_INACTIVE, 0); 1457 pq = &vmd->vmd_pagequeues[PQ_INACTIVE]; 1458 vm_pagequeue_lock(pq); 1459 vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt); 1460 while (page_shortage > 0 && (m = vm_pageout_next(&ss, true)) != NULL) { 1461 KASSERT((m->flags & PG_MARKER) == 0, 1462 ("marker page %p was dequeued", m)); 1463 1464 /* 1465 * Don't touch a page that was removed from the queue after the 1466 * page queue lock was released. Otherwise, ensure that any 1467 * pending queue operations, such as dequeues for wired pages, 1468 * are handled. 1469 */ 1470 if (vm_pageout_defer(m, PQ_INACTIVE, false)) 1471 continue; 1472 1473 /* 1474 * Lock the page's object. 1475 */ 1476 if (object == NULL || object != m->object) { 1477 if (object != NULL) 1478 VM_OBJECT_WUNLOCK(object); 1479 object = atomic_load_ptr(&m->object); 1480 if (__predict_false(object == NULL)) 1481 /* The page is being freed by another thread. */ 1482 continue; 1483 1484 /* Depends on type-stability. */ 1485 VM_OBJECT_WLOCK(object); 1486 if (__predict_false(m->object != object)) { 1487 VM_OBJECT_WUNLOCK(object); 1488 object = NULL; 1489 goto reinsert; 1490 } 1491 } 1492 1493 if (vm_page_tryxbusy(m) == 0) { 1494 /* 1495 * Don't mess with busy pages. Leave them at 1496 * the front of the queue. Most likely, they 1497 * are being paged out and will leave the 1498 * queue shortly after the scan finishes. So, 1499 * they ought to be discounted from the 1500 * inactive count. 1501 */ 1502 addl_page_shortage++; 1503 goto reinsert; 1504 } 1505 1506 /* Deferred free of swap space. */ 1507 if ((m->a.flags & PGA_SWAP_FREE) != 0) 1508 vm_pager_page_unswapped(m); 1509 1510 /* 1511 * Check for wirings now that we hold the object lock and have 1512 * exclusively busied the page. If the page is mapped, it may 1513 * still be wired by pmap lookups. The call to 1514 * vm_page_try_remove_all() below atomically checks for such 1515 * wirings and removes mappings. If the page is unmapped, the 1516 * wire count is guaranteed not to increase after this check. 1517 */ 1518 if (__predict_false(vm_page_wired(m))) 1519 goto skip_page; 1520 1521 /* 1522 * Invalid pages can be easily freed. They cannot be 1523 * mapped, vm_page_free() asserts this. 1524 */ 1525 if (vm_page_none_valid(m)) 1526 goto free_page; 1527 1528 refs = object->ref_count != 0 ? pmap_ts_referenced(m) : 0; 1529 1530 for (old = vm_page_astate_load(m);;) { 1531 /* 1532 * Check to see if the page has been removed from the 1533 * queue since the first such check. Leave it alone if 1534 * so, discarding any references collected by 1535 * pmap_ts_referenced(). 1536 */ 1537 if (__predict_false(_vm_page_queue(old) == PQ_NONE)) 1538 goto skip_page; 1539 1540 new = old; 1541 act_delta = refs; 1542 if ((old.flags & PGA_REFERENCED) != 0) { 1543 new.flags &= ~PGA_REFERENCED; 1544 act_delta++; 1545 } 1546 if (act_delta == 0) { 1547 ; 1548 } else if (object->ref_count != 0) { 1549 /* 1550 * Increase the activation count if the 1551 * page was referenced while in the 1552 * inactive queue. This makes it less 1553 * likely that the page will be returned 1554 * prematurely to the inactive queue. 1555 */ 1556 new.act_count += ACT_ADVANCE + 1557 act_delta; 1558 if (new.act_count > ACT_MAX) 1559 new.act_count = ACT_MAX; 1560 1561 new.flags &= ~PGA_QUEUE_OP_MASK; 1562 new.flags |= PGA_REQUEUE; 1563 new.queue = PQ_ACTIVE; 1564 if (!vm_page_pqstate_commit(m, &old, new)) 1565 continue; 1566 1567 VM_CNT_INC(v_reactivated); 1568 goto skip_page; 1569 } else if ((object->flags & OBJ_DEAD) == 0) { 1570 new.queue = PQ_INACTIVE; 1571 new.flags |= PGA_REQUEUE; 1572 if (!vm_page_pqstate_commit(m, &old, new)) 1573 continue; 1574 goto skip_page; 1575 } 1576 break; 1577 } 1578 1579 /* 1580 * If the page appears to be clean at the machine-independent 1581 * layer, then remove all of its mappings from the pmap in 1582 * anticipation of freeing it. If, however, any of the page's 1583 * mappings allow write access, then the page may still be 1584 * modified until the last of those mappings are removed. 1585 */ 1586 if (object->ref_count != 0) { 1587 vm_page_test_dirty(m); 1588 if (m->dirty == 0 && !vm_page_try_remove_all(m)) 1589 goto skip_page; 1590 } 1591 1592 /* 1593 * Clean pages can be freed, but dirty pages must be sent back 1594 * to the laundry, unless they belong to a dead object. 1595 * Requeueing dirty pages from dead objects is pointless, as 1596 * they are being paged out and freed by the thread that 1597 * destroyed the object. 1598 */ 1599 if (m->dirty == 0) { 1600 free_page: 1601 /* 1602 * Now we are guaranteed that no other threads are 1603 * manipulating the page, check for a last-second 1604 * reference that would save it from doom. 1605 */ 1606 if (vm_pageout_defer(m, PQ_INACTIVE, false)) 1607 goto skip_page; 1608 1609 /* 1610 * Because we dequeued the page and have already checked 1611 * for pending dequeue and enqueue requests, we can 1612 * safely disassociate the page from the inactive queue 1613 * without holding the queue lock. 1614 */ 1615 m->a.queue = PQ_NONE; 1616 vm_page_free(m); 1617 page_shortage--; 1618 continue; 1619 } 1620 if ((object->flags & OBJ_DEAD) == 0) 1621 vm_page_launder(m); 1622 skip_page: 1623 vm_page_xunbusy(m); 1624 continue; 1625 reinsert: 1626 vm_pageout_reinsert_inactive(&ss, &rq, m); 1627 } 1628 if (object != NULL) 1629 VM_OBJECT_WUNLOCK(object); 1630 vm_pageout_reinsert_inactive(&ss, &rq, NULL); 1631 vm_pageout_reinsert_inactive(&ss, &ss.bq, NULL); 1632 vm_pagequeue_lock(pq); 1633 vm_pageout_end_scan(&ss); 1634 vm_pagequeue_unlock(pq); 1635 1636 /* 1637 * Record the remaining shortage and the progress and rate it was made. 1638 */ 1639 atomic_add_int(&vmd->vmd_addl_shortage, addl_page_shortage); 1640 getmicrouptime(&end); 1641 timevalsub(&end, &start); 1642 atomic_add_int(&vmd->vmd_inactive_us, 1643 end.tv_sec * 1000000 + end.tv_usec); 1644 atomic_add_int(&vmd->vmd_inactive_freed, 1645 starting_page_shortage - page_shortage); 1646 } 1647 1648 /* 1649 * Dispatch a number of inactive threads according to load and collect the 1650 * results to prevent a coherent (CEM: incoherent?) view of paging activity on 1651 * this domain. 1652 */ 1653 static int 1654 vm_pageout_inactive_dispatch(struct vm_domain *vmd, int shortage) 1655 { 1656 u_int freed, pps, threads, us; 1657 1658 vmd->vmd_inactive_shortage = shortage; 1659 1660 /* 1661 * If we have more work than we can do in a quarter of our interval, we 1662 * fire off multiple threads to process it. 1663 */ 1664 if (vmd->vmd_inactive_threads > 1 && vmd->vmd_inactive_pps != 0 && 1665 shortage > vmd->vmd_inactive_pps / VM_INACT_SCAN_RATE / 4) { 1666 threads = vmd->vmd_inactive_threads; 1667 vm_domain_pageout_lock(vmd); 1668 vmd->vmd_inactive_shortage /= threads; 1669 blockcount_acquire(&vmd->vmd_inactive_starting, threads - 1); 1670 blockcount_acquire(&vmd->vmd_inactive_running, threads - 1); 1671 wakeup(&vmd->vmd_inactive_shortage); 1672 vm_domain_pageout_unlock(vmd); 1673 } 1674 1675 /* Run the local thread scan. */ 1676 vm_pageout_scan_inactive(vmd, vmd->vmd_inactive_shortage); 1677 1678 /* 1679 * Block until helper threads report results and then accumulate 1680 * totals. 1681 */ 1682 blockcount_wait(&vmd->vmd_inactive_running, NULL, "vmpoid", PVM); 1683 freed = atomic_readandclear_int(&vmd->vmd_inactive_freed); 1684 VM_CNT_ADD(v_dfree, freed); 1685 1686 /* 1687 * Calculate the per-thread paging rate with an exponential decay of 1688 * prior results. Careful to avoid integer rounding errors with large 1689 * us values. 1690 */ 1691 us = max(atomic_readandclear_int(&vmd->vmd_inactive_us), 1); 1692 if (us > 1000000) 1693 /* Keep rounding to tenths */ 1694 pps = (freed * 10) / ((us * 10) / 1000000); 1695 else 1696 pps = (1000000 / us) * freed; 1697 vmd->vmd_inactive_pps = (vmd->vmd_inactive_pps / 2) + (pps / 2); 1698 1699 return (shortage - freed); 1700 } 1701 1702 /* 1703 * Attempt to reclaim the requested number of pages from the inactive queue. 1704 * Returns true if the shortage was addressed. 1705 */ 1706 static int 1707 vm_pageout_inactive(struct vm_domain *vmd, int shortage, int *addl_shortage) 1708 { 1709 struct vm_pagequeue *pq; 1710 u_int addl_page_shortage, deficit, page_shortage; 1711 u_int starting_page_shortage; 1712 1713 /* 1714 * vmd_pageout_deficit counts the number of pages requested in 1715 * allocations that failed because of a free page shortage. We assume 1716 * that the allocations will be reattempted and thus include the deficit 1717 * in our scan target. 1718 */ 1719 deficit = atomic_readandclear_int(&vmd->vmd_pageout_deficit); 1720 starting_page_shortage = shortage + deficit; 1721 1722 /* 1723 * Run the inactive scan on as many threads as is necessary. 1724 */ 1725 page_shortage = vm_pageout_inactive_dispatch(vmd, starting_page_shortage); 1726 addl_page_shortage = atomic_readandclear_int(&vmd->vmd_addl_shortage); 1727 1728 /* 1729 * Wake up the laundry thread so that it can perform any needed 1730 * laundering. If we didn't meet our target, we're in shortfall and 1731 * need to launder more aggressively. If PQ_LAUNDRY is empty and no 1732 * swap devices are configured, the laundry thread has no work to do, so 1733 * don't bother waking it up. 1734 * 1735 * The laundry thread uses the number of inactive queue scans elapsed 1736 * since the last laundering to determine whether to launder again, so 1737 * keep count. 1738 */ 1739 if (starting_page_shortage > 0) { 1740 pq = &vmd->vmd_pagequeues[PQ_LAUNDRY]; 1741 vm_pagequeue_lock(pq); 1742 if (vmd->vmd_laundry_request == VM_LAUNDRY_IDLE && 1743 (pq->pq_cnt > 0 || atomic_load_acq_int(&swapdev_enabled))) { 1744 if (page_shortage > 0) { 1745 vmd->vmd_laundry_request = VM_LAUNDRY_SHORTFALL; 1746 VM_CNT_INC(v_pdshortfalls); 1747 } else if (vmd->vmd_laundry_request != 1748 VM_LAUNDRY_SHORTFALL) 1749 vmd->vmd_laundry_request = 1750 VM_LAUNDRY_BACKGROUND; 1751 wakeup(&vmd->vmd_laundry_request); 1752 } 1753 vmd->vmd_clean_pages_freed += 1754 starting_page_shortage - page_shortage; 1755 vm_pagequeue_unlock(pq); 1756 } 1757 1758 /* 1759 * Wakeup the swapout daemon if we didn't free the targeted number of 1760 * pages. 1761 */ 1762 if (page_shortage > 0) 1763 vm_swapout_run(); 1764 1765 /* 1766 * If the inactive queue scan fails repeatedly to meet its 1767 * target, kill the largest process. 1768 */ 1769 vm_pageout_mightbe_oom(vmd, page_shortage, starting_page_shortage); 1770 1771 /* 1772 * Reclaim pages by swapping out idle processes, if configured to do so. 1773 */ 1774 vm_swapout_run_idle(); 1775 1776 /* 1777 * See the description of addl_page_shortage above. 1778 */ 1779 *addl_shortage = addl_page_shortage + deficit; 1780 1781 return (page_shortage <= 0); 1782 } 1783 1784 static int vm_pageout_oom_vote; 1785 1786 /* 1787 * The pagedaemon threads randlomly select one to perform the 1788 * OOM. Trying to kill processes before all pagedaemons 1789 * failed to reach free target is premature. 1790 */ 1791 static void 1792 vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage, 1793 int starting_page_shortage) 1794 { 1795 int old_vote; 1796 1797 if (starting_page_shortage <= 0 || starting_page_shortage != 1798 page_shortage) 1799 vmd->vmd_oom_seq = 0; 1800 else 1801 vmd->vmd_oom_seq++; 1802 if (vmd->vmd_oom_seq < vm_pageout_oom_seq) { 1803 if (vmd->vmd_oom) { 1804 vmd->vmd_oom = FALSE; 1805 atomic_subtract_int(&vm_pageout_oom_vote, 1); 1806 } 1807 return; 1808 } 1809 1810 /* 1811 * Do not follow the call sequence until OOM condition is 1812 * cleared. 1813 */ 1814 vmd->vmd_oom_seq = 0; 1815 1816 if (vmd->vmd_oom) 1817 return; 1818 1819 vmd->vmd_oom = TRUE; 1820 old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1); 1821 if (old_vote != vm_ndomains - 1) 1822 return; 1823 1824 /* 1825 * The current pagedaemon thread is the last in the quorum to 1826 * start OOM. Initiate the selection and signaling of the 1827 * victim. 1828 */ 1829 vm_pageout_oom(VM_OOM_MEM); 1830 1831 /* 1832 * After one round of OOM terror, recall our vote. On the 1833 * next pass, current pagedaemon would vote again if the low 1834 * memory condition is still there, due to vmd_oom being 1835 * false. 1836 */ 1837 vmd->vmd_oom = FALSE; 1838 atomic_subtract_int(&vm_pageout_oom_vote, 1); 1839 } 1840 1841 /* 1842 * The OOM killer is the page daemon's action of last resort when 1843 * memory allocation requests have been stalled for a prolonged period 1844 * of time because it cannot reclaim memory. This function computes 1845 * the approximate number of physical pages that could be reclaimed if 1846 * the specified address space is destroyed. 1847 * 1848 * Private, anonymous memory owned by the address space is the 1849 * principal resource that we expect to recover after an OOM kill. 1850 * Since the physical pages mapped by the address space's COW entries 1851 * are typically shared pages, they are unlikely to be released and so 1852 * they are not counted. 1853 * 1854 * To get to the point where the page daemon runs the OOM killer, its 1855 * efforts to write-back vnode-backed pages may have stalled. This 1856 * could be caused by a memory allocation deadlock in the write path 1857 * that might be resolved by an OOM kill. Therefore, physical pages 1858 * belonging to vnode-backed objects are counted, because they might 1859 * be freed without being written out first if the address space holds 1860 * the last reference to an unlinked vnode. 1861 * 1862 * Similarly, physical pages belonging to OBJT_PHYS objects are 1863 * counted because the address space might hold the last reference to 1864 * the object. 1865 */ 1866 static long 1867 vm_pageout_oom_pagecount(struct vmspace *vmspace) 1868 { 1869 vm_map_t map; 1870 vm_map_entry_t entry; 1871 vm_object_t obj; 1872 long res; 1873 1874 map = &vmspace->vm_map; 1875 KASSERT(!map->system_map, ("system map")); 1876 sx_assert(&map->lock, SA_LOCKED); 1877 res = 0; 1878 VM_MAP_ENTRY_FOREACH(entry, map) { 1879 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) 1880 continue; 1881 obj = entry->object.vm_object; 1882 if (obj == NULL) 1883 continue; 1884 if ((entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0 && 1885 obj->ref_count != 1) 1886 continue; 1887 switch (obj->type) { 1888 case OBJT_DEFAULT: 1889 case OBJT_SWAP: 1890 case OBJT_PHYS: 1891 case OBJT_VNODE: 1892 res += obj->resident_page_count; 1893 break; 1894 } 1895 } 1896 return (res); 1897 } 1898 1899 static int vm_oom_ratelim_last; 1900 static int vm_oom_pf_secs = 10; 1901 SYSCTL_INT(_vm, OID_AUTO, oom_pf_secs, CTLFLAG_RWTUN, &vm_oom_pf_secs, 0, 1902 ""); 1903 static struct mtx vm_oom_ratelim_mtx; 1904 1905 void 1906 vm_pageout_oom(int shortage) 1907 { 1908 struct proc *p, *bigproc; 1909 vm_offset_t size, bigsize; 1910 struct thread *td; 1911 struct vmspace *vm; 1912 int now; 1913 bool breakout; 1914 1915 /* 1916 * For OOM requests originating from vm_fault(), there is a high 1917 * chance that a single large process faults simultaneously in 1918 * several threads. Also, on an active system running many 1919 * processes of middle-size, like buildworld, all of them 1920 * could fault almost simultaneously as well. 1921 * 1922 * To avoid killing too many processes, rate-limit OOMs 1923 * initiated by vm_fault() time-outs on the waits for free 1924 * pages. 1925 */ 1926 mtx_lock(&vm_oom_ratelim_mtx); 1927 now = ticks; 1928 if (shortage == VM_OOM_MEM_PF && 1929 (u_int)(now - vm_oom_ratelim_last) < hz * vm_oom_pf_secs) { 1930 mtx_unlock(&vm_oom_ratelim_mtx); 1931 return; 1932 } 1933 vm_oom_ratelim_last = now; 1934 mtx_unlock(&vm_oom_ratelim_mtx); 1935 1936 /* 1937 * We keep the process bigproc locked once we find it to keep anyone 1938 * from messing with it; however, there is a possibility of 1939 * deadlock if process B is bigproc and one of its child processes 1940 * attempts to propagate a signal to B while we are waiting for A's 1941 * lock while walking this list. To avoid this, we don't block on 1942 * the process lock but just skip a process if it is already locked. 1943 */ 1944 bigproc = NULL; 1945 bigsize = 0; 1946 sx_slock(&allproc_lock); 1947 FOREACH_PROC_IN_SYSTEM(p) { 1948 PROC_LOCK(p); 1949 1950 /* 1951 * If this is a system, protected or killed process, skip it. 1952 */ 1953 if (p->p_state != PRS_NORMAL || (p->p_flag & (P_INEXEC | 1954 P_PROTECTED | P_SYSTEM | P_WEXIT)) != 0 || 1955 p->p_pid == 1 || P_KILLED(p) || 1956 (p->p_pid < 48 && swap_pager_avail != 0)) { 1957 PROC_UNLOCK(p); 1958 continue; 1959 } 1960 /* 1961 * If the process is in a non-running type state, 1962 * don't touch it. Check all the threads individually. 1963 */ 1964 breakout = false; 1965 FOREACH_THREAD_IN_PROC(p, td) { 1966 thread_lock(td); 1967 if (!TD_ON_RUNQ(td) && 1968 !TD_IS_RUNNING(td) && 1969 !TD_IS_SLEEPING(td) && 1970 !TD_IS_SUSPENDED(td) && 1971 !TD_IS_SWAPPED(td)) { 1972 thread_unlock(td); 1973 breakout = true; 1974 break; 1975 } 1976 thread_unlock(td); 1977 } 1978 if (breakout) { 1979 PROC_UNLOCK(p); 1980 continue; 1981 } 1982 /* 1983 * get the process size 1984 */ 1985 vm = vmspace_acquire_ref(p); 1986 if (vm == NULL) { 1987 PROC_UNLOCK(p); 1988 continue; 1989 } 1990 _PHOLD_LITE(p); 1991 PROC_UNLOCK(p); 1992 sx_sunlock(&allproc_lock); 1993 if (!vm_map_trylock_read(&vm->vm_map)) { 1994 vmspace_free(vm); 1995 sx_slock(&allproc_lock); 1996 PRELE(p); 1997 continue; 1998 } 1999 size = vmspace_swap_count(vm); 2000 if (shortage == VM_OOM_MEM || shortage == VM_OOM_MEM_PF) 2001 size += vm_pageout_oom_pagecount(vm); 2002 vm_map_unlock_read(&vm->vm_map); 2003 vmspace_free(vm); 2004 sx_slock(&allproc_lock); 2005 2006 /* 2007 * If this process is bigger than the biggest one, 2008 * remember it. 2009 */ 2010 if (size > bigsize) { 2011 if (bigproc != NULL) 2012 PRELE(bigproc); 2013 bigproc = p; 2014 bigsize = size; 2015 } else { 2016 PRELE(p); 2017 } 2018 } 2019 sx_sunlock(&allproc_lock); 2020 if (bigproc != NULL) { 2021 if (vm_panic_on_oom != 0 && --vm_panic_on_oom == 0) 2022 panic("out of swap space"); 2023 PROC_LOCK(bigproc); 2024 killproc(bigproc, "out of swap space"); 2025 sched_nice(bigproc, PRIO_MIN); 2026 _PRELE(bigproc); 2027 PROC_UNLOCK(bigproc); 2028 } 2029 } 2030 2031 /* 2032 * Signal a free page shortage to subsystems that have registered an event 2033 * handler. Reclaim memory from UMA in the event of a severe shortage. 2034 * Return true if the free page count should be re-evaluated. 2035 */ 2036 static bool 2037 vm_pageout_lowmem(void) 2038 { 2039 static int lowmem_ticks = 0; 2040 int last; 2041 bool ret; 2042 2043 ret = false; 2044 2045 last = atomic_load_int(&lowmem_ticks); 2046 while ((u_int)(ticks - last) / hz >= lowmem_period) { 2047 if (atomic_fcmpset_int(&lowmem_ticks, &last, ticks) == 0) 2048 continue; 2049 2050 /* 2051 * Decrease registered cache sizes. 2052 */ 2053 SDT_PROBE0(vm, , , vm__lowmem_scan); 2054 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_PAGES); 2055 2056 /* 2057 * We do this explicitly after the caches have been 2058 * drained above. 2059 */ 2060 uma_reclaim(UMA_RECLAIM_TRIM); 2061 ret = true; 2062 break; 2063 } 2064 2065 /* 2066 * Kick off an asynchronous reclaim of cached memory if one of the 2067 * page daemons is failing to keep up with demand. Use the "severe" 2068 * threshold instead of "min" to ensure that we do not blow away the 2069 * caches if a subset of the NUMA domains are depleted by kernel memory 2070 * allocations; the domainset iterators automatically skip domains 2071 * below the "min" threshold on the first pass. 2072 * 2073 * UMA reclaim worker has its own rate-limiting mechanism, so don't 2074 * worry about kicking it too often. 2075 */ 2076 if (vm_page_count_severe()) 2077 uma_reclaim_wakeup(); 2078 2079 return (ret); 2080 } 2081 2082 static void 2083 vm_pageout_worker(void *arg) 2084 { 2085 struct vm_domain *vmd; 2086 u_int ofree; 2087 int addl_shortage, domain, shortage; 2088 bool target_met; 2089 2090 domain = (uintptr_t)arg; 2091 vmd = VM_DOMAIN(domain); 2092 shortage = 0; 2093 target_met = true; 2094 2095 /* 2096 * XXXKIB It could be useful to bind pageout daemon threads to 2097 * the cores belonging to the domain, from which vm_page_array 2098 * is allocated. 2099 */ 2100 2101 KASSERT(vmd->vmd_segs != 0, ("domain without segments")); 2102 vmd->vmd_last_active_scan = ticks; 2103 2104 /* 2105 * The pageout daemon worker is never done, so loop forever. 2106 */ 2107 while (TRUE) { 2108 vm_domain_pageout_lock(vmd); 2109 2110 /* 2111 * We need to clear wanted before we check the limits. This 2112 * prevents races with wakers who will check wanted after they 2113 * reach the limit. 2114 */ 2115 atomic_store_int(&vmd->vmd_pageout_wanted, 0); 2116 2117 /* 2118 * Might the page daemon need to run again? 2119 */ 2120 if (vm_paging_needed(vmd, vmd->vmd_free_count)) { 2121 /* 2122 * Yes. If the scan failed to produce enough free 2123 * pages, sleep uninterruptibly for some time in the 2124 * hope that the laundry thread will clean some pages. 2125 */ 2126 vm_domain_pageout_unlock(vmd); 2127 if (!target_met) 2128 pause("pwait", hz / VM_INACT_SCAN_RATE); 2129 } else { 2130 /* 2131 * No, sleep until the next wakeup or until pages 2132 * need to have their reference stats updated. 2133 */ 2134 if (mtx_sleep(&vmd->vmd_pageout_wanted, 2135 vm_domain_pageout_lockptr(vmd), PDROP | PVM, 2136 "psleep", hz / VM_INACT_SCAN_RATE) == 0) 2137 VM_CNT_INC(v_pdwakeups); 2138 } 2139 2140 /* Prevent spurious wakeups by ensuring that wanted is set. */ 2141 atomic_store_int(&vmd->vmd_pageout_wanted, 1); 2142 2143 /* 2144 * Use the controller to calculate how many pages to free in 2145 * this interval, and scan the inactive queue. If the lowmem 2146 * handlers appear to have freed up some pages, subtract the 2147 * difference from the inactive queue scan target. 2148 */ 2149 shortage = pidctrl_daemon(&vmd->vmd_pid, vmd->vmd_free_count); 2150 if (shortage > 0) { 2151 ofree = vmd->vmd_free_count; 2152 if (vm_pageout_lowmem() && vmd->vmd_free_count > ofree) 2153 shortage -= min(vmd->vmd_free_count - ofree, 2154 (u_int)shortage); 2155 target_met = vm_pageout_inactive(vmd, shortage, 2156 &addl_shortage); 2157 } else 2158 addl_shortage = 0; 2159 2160 /* 2161 * Scan the active queue. A positive value for shortage 2162 * indicates that we must aggressively deactivate pages to avoid 2163 * a shortfall. 2164 */ 2165 shortage = vm_pageout_active_target(vmd) + addl_shortage; 2166 vm_pageout_scan_active(vmd, shortage); 2167 } 2168 } 2169 2170 /* 2171 * vm_pageout_helper runs additional pageout daemons in times of high paging 2172 * activity. 2173 */ 2174 static void 2175 vm_pageout_helper(void *arg) 2176 { 2177 struct vm_domain *vmd; 2178 int domain; 2179 2180 domain = (uintptr_t)arg; 2181 vmd = VM_DOMAIN(domain); 2182 2183 vm_domain_pageout_lock(vmd); 2184 for (;;) { 2185 msleep(&vmd->vmd_inactive_shortage, 2186 vm_domain_pageout_lockptr(vmd), PVM, "psleep", 0); 2187 blockcount_release(&vmd->vmd_inactive_starting, 1); 2188 2189 vm_domain_pageout_unlock(vmd); 2190 vm_pageout_scan_inactive(vmd, vmd->vmd_inactive_shortage); 2191 vm_domain_pageout_lock(vmd); 2192 2193 /* 2194 * Release the running count while the pageout lock is held to 2195 * prevent wakeup races. 2196 */ 2197 blockcount_release(&vmd->vmd_inactive_running, 1); 2198 } 2199 } 2200 2201 static int 2202 get_pageout_threads_per_domain(const struct vm_domain *vmd) 2203 { 2204 unsigned total_pageout_threads, eligible_cpus, domain_cpus; 2205 2206 if (VM_DOMAIN_EMPTY(vmd->vmd_domain)) 2207 return (0); 2208 2209 /* 2210 * Semi-arbitrarily constrain pagedaemon threads to less than half the 2211 * total number of CPUs in the system as an upper limit. 2212 */ 2213 if (pageout_cpus_per_thread < 2) 2214 pageout_cpus_per_thread = 2; 2215 else if (pageout_cpus_per_thread > mp_ncpus) 2216 pageout_cpus_per_thread = mp_ncpus; 2217 2218 total_pageout_threads = howmany(mp_ncpus, pageout_cpus_per_thread); 2219 domain_cpus = CPU_COUNT(&cpuset_domain[vmd->vmd_domain]); 2220 2221 /* Pagedaemons are not run in empty domains. */ 2222 eligible_cpus = mp_ncpus; 2223 for (unsigned i = 0; i < vm_ndomains; i++) 2224 if (VM_DOMAIN_EMPTY(i)) 2225 eligible_cpus -= CPU_COUNT(&cpuset_domain[i]); 2226 2227 /* 2228 * Assign a portion of the total pageout threads to this domain 2229 * corresponding to the fraction of pagedaemon-eligible CPUs in the 2230 * domain. In asymmetric NUMA systems, domains with more CPUs may be 2231 * allocated more threads than domains with fewer CPUs. 2232 */ 2233 return (howmany(total_pageout_threads * domain_cpus, eligible_cpus)); 2234 } 2235 2236 /* 2237 * Initialize basic pageout daemon settings. See the comment above the 2238 * definition of vm_domain for some explanation of how these thresholds are 2239 * used. 2240 */ 2241 static void 2242 vm_pageout_init_domain(int domain) 2243 { 2244 struct vm_domain *vmd; 2245 struct sysctl_oid *oid; 2246 2247 vmd = VM_DOMAIN(domain); 2248 vmd->vmd_interrupt_free_min = 2; 2249 2250 /* 2251 * v_free_reserved needs to include enough for the largest 2252 * swap pager structures plus enough for any pv_entry structs 2253 * when paging. 2254 */ 2255 vmd->vmd_pageout_free_min = 2 * MAXBSIZE / PAGE_SIZE + 2256 vmd->vmd_interrupt_free_min; 2257 vmd->vmd_free_reserved = vm_pageout_page_count + 2258 vmd->vmd_pageout_free_min + vmd->vmd_page_count / 768; 2259 vmd->vmd_free_min = vmd->vmd_page_count / 200; 2260 vmd->vmd_free_severe = vmd->vmd_free_min / 2; 2261 vmd->vmd_free_target = 4 * vmd->vmd_free_min + vmd->vmd_free_reserved; 2262 vmd->vmd_free_min += vmd->vmd_free_reserved; 2263 vmd->vmd_free_severe += vmd->vmd_free_reserved; 2264 vmd->vmd_inactive_target = (3 * vmd->vmd_free_target) / 2; 2265 if (vmd->vmd_inactive_target > vmd->vmd_free_count / 3) 2266 vmd->vmd_inactive_target = vmd->vmd_free_count / 3; 2267 2268 /* 2269 * Set the default wakeup threshold to be 10% below the paging 2270 * target. This keeps the steady state out of shortfall. 2271 */ 2272 vmd->vmd_pageout_wakeup_thresh = (vmd->vmd_free_target / 10) * 9; 2273 2274 /* 2275 * Target amount of memory to move out of the laundry queue during a 2276 * background laundering. This is proportional to the amount of system 2277 * memory. 2278 */ 2279 vmd->vmd_background_launder_target = (vmd->vmd_free_target - 2280 vmd->vmd_free_min) / 10; 2281 2282 /* Initialize the pageout daemon pid controller. */ 2283 pidctrl_init(&vmd->vmd_pid, hz / VM_INACT_SCAN_RATE, 2284 vmd->vmd_free_target, PIDCTRL_BOUND, 2285 PIDCTRL_KPD, PIDCTRL_KID, PIDCTRL_KDD); 2286 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(vmd->vmd_oid), OID_AUTO, 2287 "pidctrl", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, ""); 2288 pidctrl_init_sysctl(&vmd->vmd_pid, SYSCTL_CHILDREN(oid)); 2289 2290 vmd->vmd_inactive_threads = get_pageout_threads_per_domain(vmd); 2291 } 2292 2293 static void 2294 vm_pageout_init(void) 2295 { 2296 u_int freecount; 2297 int i; 2298 2299 /* 2300 * Initialize some paging parameters. 2301 */ 2302 if (vm_cnt.v_page_count < 2000) 2303 vm_pageout_page_count = 8; 2304 2305 freecount = 0; 2306 for (i = 0; i < vm_ndomains; i++) { 2307 struct vm_domain *vmd; 2308 2309 vm_pageout_init_domain(i); 2310 vmd = VM_DOMAIN(i); 2311 vm_cnt.v_free_reserved += vmd->vmd_free_reserved; 2312 vm_cnt.v_free_target += vmd->vmd_free_target; 2313 vm_cnt.v_free_min += vmd->vmd_free_min; 2314 vm_cnt.v_inactive_target += vmd->vmd_inactive_target; 2315 vm_cnt.v_pageout_free_min += vmd->vmd_pageout_free_min; 2316 vm_cnt.v_interrupt_free_min += vmd->vmd_interrupt_free_min; 2317 vm_cnt.v_free_severe += vmd->vmd_free_severe; 2318 freecount += vmd->vmd_free_count; 2319 } 2320 2321 /* 2322 * Set interval in seconds for active scan. We want to visit each 2323 * page at least once every ten minutes. This is to prevent worst 2324 * case paging behaviors with stale active LRU. 2325 */ 2326 if (vm_pageout_update_period == 0) 2327 vm_pageout_update_period = 600; 2328 2329 if (vm_page_max_user_wired == 0) 2330 vm_page_max_user_wired = freecount / 3; 2331 } 2332 2333 /* 2334 * vm_pageout is the high level pageout daemon. 2335 */ 2336 static void 2337 vm_pageout(void) 2338 { 2339 struct proc *p; 2340 struct thread *td; 2341 int error, first, i, j, pageout_threads; 2342 2343 p = curproc; 2344 td = curthread; 2345 2346 mtx_init(&vm_oom_ratelim_mtx, "vmoomr", NULL, MTX_DEF); 2347 swap_pager_swap_init(); 2348 for (first = -1, i = 0; i < vm_ndomains; i++) { 2349 if (VM_DOMAIN_EMPTY(i)) { 2350 if (bootverbose) 2351 printf("domain %d empty; skipping pageout\n", 2352 i); 2353 continue; 2354 } 2355 if (first == -1) 2356 first = i; 2357 else { 2358 error = kthread_add(vm_pageout_worker, 2359 (void *)(uintptr_t)i, p, NULL, 0, 0, "dom%d", i); 2360 if (error != 0) 2361 panic("starting pageout for domain %d: %d\n", 2362 i, error); 2363 } 2364 pageout_threads = VM_DOMAIN(i)->vmd_inactive_threads; 2365 for (j = 0; j < pageout_threads - 1; j++) { 2366 error = kthread_add(vm_pageout_helper, 2367 (void *)(uintptr_t)i, p, NULL, 0, 0, 2368 "dom%d helper%d", i, j); 2369 if (error != 0) 2370 panic("starting pageout helper %d for domain " 2371 "%d: %d\n", j, i, error); 2372 } 2373 error = kthread_add(vm_pageout_laundry_worker, 2374 (void *)(uintptr_t)i, p, NULL, 0, 0, "laundry: dom%d", i); 2375 if (error != 0) 2376 panic("starting laundry for domain %d: %d", i, error); 2377 } 2378 error = kthread_add(uma_reclaim_worker, NULL, p, NULL, 0, 0, "uma"); 2379 if (error != 0) 2380 panic("starting uma_reclaim helper, error %d\n", error); 2381 2382 snprintf(td->td_name, sizeof(td->td_name), "dom%d", first); 2383 vm_pageout_worker((void *)(uintptr_t)first); 2384 } 2385 2386 /* 2387 * Perform an advisory wakeup of the page daemon. 2388 */ 2389 void 2390 pagedaemon_wakeup(int domain) 2391 { 2392 struct vm_domain *vmd; 2393 2394 vmd = VM_DOMAIN(domain); 2395 vm_domain_pageout_assert_unlocked(vmd); 2396 if (curproc == pageproc) 2397 return; 2398 2399 if (atomic_fetchadd_int(&vmd->vmd_pageout_wanted, 1) == 0) { 2400 vm_domain_pageout_lock(vmd); 2401 atomic_store_int(&vmd->vmd_pageout_wanted, 1); 2402 wakeup(&vmd->vmd_pageout_wanted); 2403 vm_domain_pageout_unlock(vmd); 2404 } 2405 } 2406