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 *
45 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
46 * All rights reserved.
47 *
48 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
49 *
50 * Permission to use, copy, modify and distribute this software and
51 * its documentation is hereby granted, provided that both the copyright
52 * notice and this permission notice appear in all copies of the
53 * software, derivative works or modified versions, and any portions
54 * thereof, and that both notices appear in supporting documentation.
55 *
56 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
57 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
58 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
59 *
60 * Carnegie Mellon requests users of this software to return to
61 *
62 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
63 * School of Computer Science
64 * Carnegie Mellon University
65 * Pittsburgh PA 15213-3890
66 *
67 * any improvements or extensions that they make and grant Carnegie the
68 * rights to redistribute these changes.
69 */
70
71 /*
72 * The proverbial page-out daemon.
73 */
74
75 #include <sys/cdefs.h>
76 #include "opt_vm.h"
77
78 #include <sys/param.h>
79 #include <sys/systm.h>
80 #include <sys/kernel.h>
81 #include <sys/blockcount.h>
82 #include <sys/eventhandler.h>
83 #include <sys/lock.h>
84 #include <sys/mutex.h>
85 #include <sys/proc.h>
86 #include <sys/kthread.h>
87 #include <sys/ktr.h>
88 #include <sys/mount.h>
89 #include <sys/racct.h>
90 #include <sys/resourcevar.h>
91 #include <sys/sched.h>
92 #include <sys/sdt.h>
93 #include <sys/signalvar.h>
94 #include <sys/smp.h>
95 #include <sys/time.h>
96 #include <sys/vnode.h>
97 #include <sys/vmmeter.h>
98 #include <sys/rwlock.h>
99 #include <sys/sx.h>
100 #include <sys/sysctl.h>
101
102 #include <vm/vm.h>
103 #include <vm/vm_param.h>
104 #include <vm/vm_object.h>
105 #include <vm/vm_page.h>
106 #include <vm/vm_map.h>
107 #include <vm/vm_pageout.h>
108 #include <vm/vm_pager.h>
109 #include <vm/vm_phys.h>
110 #include <vm/vm_pagequeue.h>
111 #include <vm/swap_pager.h>
112 #include <vm/vm_extern.h>
113 #include <vm/uma.h>
114
115 /*
116 * System initialization
117 */
118
119 /* the kernel process "vm_pageout"*/
120 static void vm_pageout(void);
121 static void vm_pageout_init(void);
122 static int vm_pageout_clean(vm_page_t m, int *numpagedout);
123 static int vm_pageout_cluster(vm_page_t m);
124 static void vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
125 int starting_page_shortage);
126
127 SYSINIT(pagedaemon_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, vm_pageout_init,
128 NULL);
129
130 struct proc *pageproc;
131
132 static struct kproc_desc page_kp = {
133 "pagedaemon",
134 vm_pageout,
135 &pageproc
136 };
137 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start,
138 &page_kp);
139
140 SDT_PROVIDER_DEFINE(vm);
141 SDT_PROBE_DEFINE(vm, , , vm__lowmem_scan);
142
143 /* Pagedaemon activity rates, in subdivisions of one second. */
144 #define VM_LAUNDER_RATE 10
145 #define VM_INACT_SCAN_RATE 10
146
147 static int swapdev_enabled;
148 int vm_pageout_page_count = 32;
149
150 static int vm_panic_on_oom = 0;
151 SYSCTL_INT(_vm, OID_AUTO, panic_on_oom,
152 CTLFLAG_RWTUN, &vm_panic_on_oom, 0,
153 "Panic on the given number of out-of-memory errors instead of "
154 "killing the largest process");
155
156 static int vm_pageout_update_period;
157 SYSCTL_INT(_vm, OID_AUTO, pageout_update_period,
158 CTLFLAG_RWTUN, &vm_pageout_update_period, 0,
159 "Maximum active LRU update period");
160
161 static int pageout_cpus_per_thread = 16;
162 SYSCTL_INT(_vm, OID_AUTO, pageout_cpus_per_thread, CTLFLAG_RDTUN,
163 &pageout_cpus_per_thread, 0,
164 "Number of CPUs per pagedaemon worker thread");
165
166 static int lowmem_period = 10;
167 SYSCTL_INT(_vm, OID_AUTO, lowmem_period, CTLFLAG_RWTUN, &lowmem_period, 0,
168 "Low memory callback period");
169
170 static int disable_swap_pageouts;
171 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
172 CTLFLAG_RWTUN, &disable_swap_pageouts, 0,
173 "Disallow swapout of dirty pages");
174
175 static int pageout_lock_miss;
176 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
177 CTLFLAG_RD, &pageout_lock_miss, 0,
178 "vget() lock misses during pageout");
179
180 static int vm_pageout_oom_seq = 12;
181 SYSCTL_INT(_vm, OID_AUTO, pageout_oom_seq,
182 CTLFLAG_RWTUN, &vm_pageout_oom_seq, 0,
183 "back-to-back calls to oom detector to start OOM");
184
185 static int act_scan_laundry_weight = 3;
186 SYSCTL_INT(_vm, OID_AUTO, act_scan_laundry_weight, CTLFLAG_RWTUN,
187 &act_scan_laundry_weight, 0,
188 "weight given to clean vs. dirty pages in active queue scans");
189
190 static u_int vm_background_launder_rate = 4096;
191 SYSCTL_UINT(_vm, OID_AUTO, background_launder_rate, CTLFLAG_RWTUN,
192 &vm_background_launder_rate, 0,
193 "background laundering rate, in kilobytes per second");
194
195 static u_int vm_background_launder_max = 20 * 1024;
196 SYSCTL_UINT(_vm, OID_AUTO, background_launder_max, CTLFLAG_RWTUN,
197 &vm_background_launder_max, 0,
198 "background laundering cap, in kilobytes");
199
200 u_long vm_page_max_user_wired;
201 SYSCTL_ULONG(_vm, OID_AUTO, max_user_wired, CTLFLAG_RW,
202 &vm_page_max_user_wired, 0,
203 "system-wide limit to user-wired page count");
204
205 static u_int isqrt(u_int num);
206 static int vm_pageout_launder(struct vm_domain *vmd, int launder,
207 bool in_shortfall);
208 static void vm_pageout_laundry_worker(void *arg);
209
210 struct scan_state {
211 struct vm_batchqueue bq;
212 struct vm_pagequeue *pq;
213 vm_page_t marker;
214 int maxscan;
215 int scanned;
216 };
217
218 static void
vm_pageout_init_scan(struct scan_state * ss,struct vm_pagequeue * pq,vm_page_t marker,vm_page_t after,int maxscan)219 vm_pageout_init_scan(struct scan_state *ss, struct vm_pagequeue *pq,
220 vm_page_t marker, vm_page_t after, int maxscan)
221 {
222
223 vm_pagequeue_assert_locked(pq);
224 KASSERT((marker->a.flags & PGA_ENQUEUED) == 0,
225 ("marker %p already enqueued", marker));
226
227 if (after == NULL)
228 TAILQ_INSERT_HEAD(&pq->pq_pl, marker, plinks.q);
229 else
230 TAILQ_INSERT_AFTER(&pq->pq_pl, after, marker, plinks.q);
231 vm_page_aflag_set(marker, PGA_ENQUEUED);
232
233 vm_batchqueue_init(&ss->bq);
234 ss->pq = pq;
235 ss->marker = marker;
236 ss->maxscan = maxscan;
237 ss->scanned = 0;
238 vm_pagequeue_unlock(pq);
239 }
240
241 static void
vm_pageout_end_scan(struct scan_state * ss)242 vm_pageout_end_scan(struct scan_state *ss)
243 {
244 struct vm_pagequeue *pq;
245
246 pq = ss->pq;
247 vm_pagequeue_assert_locked(pq);
248 KASSERT((ss->marker->a.flags & PGA_ENQUEUED) != 0,
249 ("marker %p not enqueued", ss->marker));
250
251 TAILQ_REMOVE(&pq->pq_pl, ss->marker, plinks.q);
252 vm_page_aflag_clear(ss->marker, PGA_ENQUEUED);
253 pq->pq_pdpages += ss->scanned;
254 }
255
256 /*
257 * Add a small number of queued pages to a batch queue for later processing
258 * without the corresponding queue lock held. The caller must have enqueued a
259 * marker page at the desired start point for the scan. Pages will be
260 * physically dequeued if the caller so requests. Otherwise, the returned
261 * batch may contain marker pages, and it is up to the caller to handle them.
262 *
263 * When processing the batch queue, vm_pageout_defer() must be used to
264 * determine whether the page has been logically dequeued since the batch was
265 * collected.
266 */
267 static __always_inline void
vm_pageout_collect_batch(struct scan_state * ss,const bool dequeue)268 vm_pageout_collect_batch(struct scan_state *ss, const bool dequeue)
269 {
270 struct vm_pagequeue *pq;
271 vm_page_t m, marker, n;
272
273 marker = ss->marker;
274 pq = ss->pq;
275
276 KASSERT((marker->a.flags & PGA_ENQUEUED) != 0,
277 ("marker %p not enqueued", ss->marker));
278
279 vm_pagequeue_lock(pq);
280 for (m = TAILQ_NEXT(marker, plinks.q); m != NULL &&
281 ss->scanned < ss->maxscan && ss->bq.bq_cnt < VM_BATCHQUEUE_SIZE;
282 m = n, ss->scanned++) {
283 n = TAILQ_NEXT(m, plinks.q);
284 if ((m->flags & PG_MARKER) == 0) {
285 KASSERT((m->a.flags & PGA_ENQUEUED) != 0,
286 ("page %p not enqueued", m));
287 KASSERT((m->flags & PG_FICTITIOUS) == 0,
288 ("Fictitious page %p cannot be in page queue", m));
289 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
290 ("Unmanaged page %p cannot be in page queue", m));
291 } else if (dequeue)
292 continue;
293
294 (void)vm_batchqueue_insert(&ss->bq, m);
295 if (dequeue) {
296 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
297 vm_page_aflag_clear(m, PGA_ENQUEUED);
298 }
299 }
300 TAILQ_REMOVE(&pq->pq_pl, marker, plinks.q);
301 if (__predict_true(m != NULL))
302 TAILQ_INSERT_BEFORE(m, marker, plinks.q);
303 else
304 TAILQ_INSERT_TAIL(&pq->pq_pl, marker, plinks.q);
305 if (dequeue)
306 vm_pagequeue_cnt_add(pq, -ss->bq.bq_cnt);
307 vm_pagequeue_unlock(pq);
308 }
309
310 /*
311 * Return the next page to be scanned, or NULL if the scan is complete.
312 */
313 static __always_inline vm_page_t
vm_pageout_next(struct scan_state * ss,const bool dequeue)314 vm_pageout_next(struct scan_state *ss, const bool dequeue)
315 {
316
317 if (ss->bq.bq_cnt == 0)
318 vm_pageout_collect_batch(ss, dequeue);
319 return (vm_batchqueue_pop(&ss->bq));
320 }
321
322 /*
323 * Determine whether processing of a page should be deferred and ensure that any
324 * outstanding queue operations are processed.
325 */
326 static __always_inline bool
vm_pageout_defer(vm_page_t m,const uint8_t queue,const bool enqueued)327 vm_pageout_defer(vm_page_t m, const uint8_t queue, const bool enqueued)
328 {
329 vm_page_astate_t as;
330
331 as = vm_page_astate_load(m);
332 if (__predict_false(as.queue != queue ||
333 ((as.flags & PGA_ENQUEUED) != 0) != enqueued))
334 return (true);
335 if ((as.flags & PGA_QUEUE_OP_MASK) != 0) {
336 vm_page_pqbatch_submit(m, queue);
337 return (true);
338 }
339 return (false);
340 }
341
342 /*
343 * Scan for pages at adjacent offsets within the given page's object that are
344 * eligible for laundering, form a cluster of these pages and the given page,
345 * and launder that cluster.
346 */
347 static int
vm_pageout_cluster(vm_page_t m)348 vm_pageout_cluster(vm_page_t m)
349 {
350 vm_object_t object;
351 vm_page_t mc[2 * vm_pageout_page_count], p, pb, ps;
352 vm_pindex_t pindex;
353 int ib, is, page_base, pageout_count;
354
355 object = m->object;
356 VM_OBJECT_ASSERT_WLOCKED(object);
357 pindex = m->pindex;
358
359 vm_page_assert_xbusied(m);
360
361 mc[vm_pageout_page_count] = pb = ps = m;
362 pageout_count = 1;
363 page_base = vm_pageout_page_count;
364 ib = 1;
365 is = 1;
366
367 /*
368 * We can cluster only if the page is not clean, busy, or held, and
369 * the page is in the laundry queue.
370 *
371 * During heavy mmap/modification loads the pageout
372 * daemon can really fragment the underlying file
373 * due to flushing pages out of order and not trying to
374 * align the clusters (which leaves sporadic out-of-order
375 * holes). To solve this problem we do the reverse scan
376 * first and attempt to align our cluster, then do a
377 * forward scan if room remains.
378 */
379 more:
380 while (ib != 0 && pageout_count < vm_pageout_page_count) {
381 if (ib > pindex) {
382 ib = 0;
383 break;
384 }
385 if ((p = vm_page_prev(pb)) == NULL ||
386 vm_page_tryxbusy(p) == 0) {
387 ib = 0;
388 break;
389 }
390 if (vm_page_wired(p)) {
391 ib = 0;
392 vm_page_xunbusy(p);
393 break;
394 }
395 vm_page_test_dirty(p);
396 if (p->dirty == 0) {
397 ib = 0;
398 vm_page_xunbusy(p);
399 break;
400 }
401 if (!vm_page_in_laundry(p) || !vm_page_try_remove_write(p)) {
402 vm_page_xunbusy(p);
403 ib = 0;
404 break;
405 }
406 mc[--page_base] = pb = p;
407 ++pageout_count;
408 ++ib;
409
410 /*
411 * We are at an alignment boundary. Stop here, and switch
412 * directions. Do not clear ib.
413 */
414 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
415 break;
416 }
417 while (pageout_count < vm_pageout_page_count &&
418 pindex + is < object->size) {
419 if ((p = vm_page_next(ps)) == NULL ||
420 vm_page_tryxbusy(p) == 0)
421 break;
422 if (vm_page_wired(p)) {
423 vm_page_xunbusy(p);
424 break;
425 }
426 vm_page_test_dirty(p);
427 if (p->dirty == 0) {
428 vm_page_xunbusy(p);
429 break;
430 }
431 if (!vm_page_in_laundry(p) || !vm_page_try_remove_write(p)) {
432 vm_page_xunbusy(p);
433 break;
434 }
435 mc[page_base + pageout_count] = ps = p;
436 ++pageout_count;
437 ++is;
438 }
439
440 /*
441 * If we exhausted our forward scan, continue with the reverse scan
442 * when possible, even past an alignment boundary. This catches
443 * boundary conditions.
444 */
445 if (ib != 0 && pageout_count < vm_pageout_page_count)
446 goto more;
447
448 return (vm_pageout_flush(&mc[page_base], pageout_count,
449 VM_PAGER_PUT_NOREUSE, 0, NULL, NULL));
450 }
451
452 /*
453 * vm_pageout_flush() - launder the given pages
454 *
455 * The given pages are laundered. Note that we setup for the start of
456 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
457 * reference count all in here rather then in the parent. If we want
458 * the parent to do more sophisticated things we may have to change
459 * the ordering.
460 *
461 * Returned runlen is the count of pages between mreq and first
462 * page after mreq with status VM_PAGER_AGAIN.
463 * *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL
464 * for any page in runlen set.
465 */
466 int
vm_pageout_flush(vm_page_t * mc,int count,int flags,int mreq,int * prunlen,boolean_t * eio)467 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen,
468 boolean_t *eio)
469 {
470 vm_object_t object = mc[0]->object;
471 int pageout_status[count];
472 int numpagedout = 0;
473 int i, runlen;
474
475 VM_OBJECT_ASSERT_WLOCKED(object);
476
477 /*
478 * Initiate I/O. Mark the pages shared busy and verify that they're
479 * valid and read-only.
480 *
481 * We do not have to fixup the clean/dirty bits here... we can
482 * allow the pager to do it after the I/O completes.
483 *
484 * NOTE! mc[i]->dirty may be partial or fragmented due to an
485 * edge case with file fragments.
486 */
487 for (i = 0; i < count; i++) {
488 KASSERT(vm_page_all_valid(mc[i]),
489 ("vm_pageout_flush: partially invalid page %p index %d/%d",
490 mc[i], i, count));
491 KASSERT((mc[i]->a.flags & PGA_WRITEABLE) == 0,
492 ("vm_pageout_flush: writeable page %p", mc[i]));
493 vm_page_busy_downgrade(mc[i]);
494 }
495 vm_object_pip_add(object, count);
496
497 vm_pager_put_pages(object, mc, count, flags, pageout_status);
498
499 runlen = count - mreq;
500 if (eio != NULL)
501 *eio = FALSE;
502 for (i = 0; i < count; i++) {
503 vm_page_t mt = mc[i];
504
505 KASSERT(pageout_status[i] == VM_PAGER_PEND ||
506 !pmap_page_is_write_mapped(mt),
507 ("vm_pageout_flush: page %p is not write protected", mt));
508 switch (pageout_status[i]) {
509 case VM_PAGER_OK:
510 /*
511 * The page may have moved since laundering started, in
512 * which case it should be left alone.
513 */
514 if (vm_page_in_laundry(mt))
515 vm_page_deactivate_noreuse(mt);
516 /* FALLTHROUGH */
517 case VM_PAGER_PEND:
518 numpagedout++;
519 break;
520 case VM_PAGER_BAD:
521 /*
522 * The page is outside the object's range. We pretend
523 * that the page out worked and clean the page, so the
524 * changes will be lost if the page is reclaimed by
525 * the page daemon.
526 */
527 vm_page_undirty(mt);
528 if (vm_page_in_laundry(mt))
529 vm_page_deactivate_noreuse(mt);
530 break;
531 case VM_PAGER_ERROR:
532 case VM_PAGER_FAIL:
533 /*
534 * If the page couldn't be paged out to swap because the
535 * pager wasn't able to find space, place the page in
536 * the PQ_UNSWAPPABLE holding queue. This is an
537 * optimization that prevents the page daemon from
538 * wasting CPU cycles on pages that cannot be reclaimed
539 * because no swap device is configured.
540 *
541 * Otherwise, reactivate the page so that it doesn't
542 * clog the laundry and inactive queues. (We will try
543 * paging it out again later.)
544 */
545 if ((object->flags & OBJ_SWAP) != 0 &&
546 pageout_status[i] == VM_PAGER_FAIL) {
547 vm_page_unswappable(mt);
548 numpagedout++;
549 } else
550 vm_page_activate(mt);
551 if (eio != NULL && i >= mreq && i - mreq < runlen)
552 *eio = TRUE;
553 break;
554 case VM_PAGER_AGAIN:
555 if (i >= mreq && i - mreq < runlen)
556 runlen = i - mreq;
557 break;
558 }
559
560 /*
561 * If the operation is still going, leave the page busy to
562 * block all other accesses. Also, leave the paging in
563 * progress indicator set so that we don't attempt an object
564 * collapse.
565 */
566 if (pageout_status[i] != VM_PAGER_PEND) {
567 vm_object_pip_wakeup(object);
568 vm_page_sunbusy(mt);
569 }
570 }
571 if (prunlen != NULL)
572 *prunlen = runlen;
573 return (numpagedout);
574 }
575
576 static void
vm_pageout_swapon(void * arg __unused,struct swdevt * sp __unused)577 vm_pageout_swapon(void *arg __unused, struct swdevt *sp __unused)
578 {
579
580 atomic_store_rel_int(&swapdev_enabled, 1);
581 }
582
583 static void
vm_pageout_swapoff(void * arg __unused,struct swdevt * sp __unused)584 vm_pageout_swapoff(void *arg __unused, struct swdevt *sp __unused)
585 {
586
587 if (swap_pager_nswapdev() == 1)
588 atomic_store_rel_int(&swapdev_enabled, 0);
589 }
590
591 /*
592 * Attempt to acquire all of the necessary locks to launder a page and
593 * then call through the clustering layer to PUTPAGES. Wait a short
594 * time for a vnode lock.
595 *
596 * Requires the page and object lock on entry, releases both before return.
597 * Returns 0 on success and an errno otherwise.
598 */
599 static int
vm_pageout_clean(vm_page_t m,int * numpagedout)600 vm_pageout_clean(vm_page_t m, int *numpagedout)
601 {
602 struct vnode *vp;
603 struct mount *mp;
604 vm_object_t object;
605 vm_pindex_t pindex;
606 int error;
607
608 object = m->object;
609 VM_OBJECT_ASSERT_WLOCKED(object);
610 error = 0;
611 vp = NULL;
612 mp = NULL;
613
614 /*
615 * The object is already known NOT to be dead. It
616 * is possible for the vget() to block the whole
617 * pageout daemon, but the new low-memory handling
618 * code should prevent it.
619 *
620 * We can't wait forever for the vnode lock, we might
621 * deadlock due to a vn_read() getting stuck in
622 * vm_wait while holding this vnode. We skip the
623 * vnode if we can't get it in a reasonable amount
624 * of time.
625 */
626 if (object->type == OBJT_VNODE) {
627 vm_page_xunbusy(m);
628 vp = object->handle;
629 if (vp->v_type == VREG &&
630 vn_start_write(vp, &mp, V_NOWAIT) != 0) {
631 mp = NULL;
632 error = EDEADLK;
633 goto unlock_all;
634 }
635 KASSERT(mp != NULL,
636 ("vp %p with NULL v_mount", vp));
637 vm_object_reference_locked(object);
638 pindex = m->pindex;
639 VM_OBJECT_WUNLOCK(object);
640 if (vget(vp, vn_lktype_write(NULL, vp) | LK_TIMELOCK) != 0) {
641 vp = NULL;
642 error = EDEADLK;
643 goto unlock_mp;
644 }
645 VM_OBJECT_WLOCK(object);
646
647 /*
648 * Ensure that the object and vnode were not disassociated
649 * while locks were dropped.
650 */
651 if (vp->v_object != object) {
652 error = ENOENT;
653 goto unlock_all;
654 }
655
656 /*
657 * While the object was unlocked, the page may have been:
658 * (1) moved to a different queue,
659 * (2) reallocated to a different object,
660 * (3) reallocated to a different offset, or
661 * (4) cleaned.
662 */
663 if (!vm_page_in_laundry(m) || m->object != object ||
664 m->pindex != pindex || m->dirty == 0) {
665 error = ENXIO;
666 goto unlock_all;
667 }
668
669 /*
670 * The page may have been busied while the object lock was
671 * released.
672 */
673 if (vm_page_tryxbusy(m) == 0) {
674 error = EBUSY;
675 goto unlock_all;
676 }
677 }
678
679 /*
680 * Remove all writeable mappings, failing if the page is wired.
681 */
682 if (!vm_page_try_remove_write(m)) {
683 vm_page_xunbusy(m);
684 error = EBUSY;
685 goto unlock_all;
686 }
687
688 /*
689 * If a page is dirty, then it is either being washed
690 * (but not yet cleaned) or it is still in the
691 * laundry. If it is still in the laundry, then we
692 * start the cleaning operation.
693 */
694 if ((*numpagedout = vm_pageout_cluster(m)) == 0)
695 error = EIO;
696
697 unlock_all:
698 VM_OBJECT_WUNLOCK(object);
699
700 unlock_mp:
701 if (mp != NULL) {
702 if (vp != NULL)
703 vput(vp);
704 vm_object_deallocate(object);
705 vn_finished_write(mp);
706 }
707
708 return (error);
709 }
710
711 /*
712 * Attempt to launder the specified number of pages.
713 *
714 * Returns the number of pages successfully laundered.
715 */
716 static int
vm_pageout_launder(struct vm_domain * vmd,int launder,bool in_shortfall)717 vm_pageout_launder(struct vm_domain *vmd, int launder, bool in_shortfall)
718 {
719 struct scan_state ss;
720 struct vm_pagequeue *pq;
721 vm_object_t object;
722 vm_page_t m, marker;
723 vm_page_astate_t new, old;
724 int act_delta, error, numpagedout, queue, refs, starting_target;
725 int vnodes_skipped;
726 bool pageout_ok;
727
728 object = NULL;
729 starting_target = launder;
730 vnodes_skipped = 0;
731
732 /*
733 * Scan the laundry queues for pages eligible to be laundered. We stop
734 * once the target number of dirty pages have been laundered, or once
735 * we've reached the end of the queue. A single iteration of this loop
736 * may cause more than one page to be laundered because of clustering.
737 *
738 * As an optimization, we avoid laundering from PQ_UNSWAPPABLE when no
739 * swap devices are configured.
740 */
741 if (atomic_load_acq_int(&swapdev_enabled))
742 queue = PQ_UNSWAPPABLE;
743 else
744 queue = PQ_LAUNDRY;
745
746 scan:
747 marker = &vmd->vmd_markers[queue];
748 pq = &vmd->vmd_pagequeues[queue];
749 vm_pagequeue_lock(pq);
750 vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt);
751 while (launder > 0 && (m = vm_pageout_next(&ss, false)) != NULL) {
752 if (__predict_false((m->flags & PG_MARKER) != 0))
753 continue;
754
755 /*
756 * Don't touch a page that was removed from the queue after the
757 * page queue lock was released. Otherwise, ensure that any
758 * pending queue operations, such as dequeues for wired pages,
759 * are handled.
760 */
761 if (vm_pageout_defer(m, queue, true))
762 continue;
763
764 /*
765 * Lock the page's object.
766 */
767 if (object == NULL || object != m->object) {
768 if (object != NULL)
769 VM_OBJECT_WUNLOCK(object);
770 object = atomic_load_ptr(&m->object);
771 if (__predict_false(object == NULL))
772 /* The page is being freed by another thread. */
773 continue;
774
775 /* Depends on type-stability. */
776 VM_OBJECT_WLOCK(object);
777 if (__predict_false(m->object != object)) {
778 VM_OBJECT_WUNLOCK(object);
779 object = NULL;
780 continue;
781 }
782 }
783
784 if (vm_page_tryxbusy(m) == 0)
785 continue;
786
787 /*
788 * Check for wirings now that we hold the object lock and have
789 * exclusively busied the page. If the page is mapped, it may
790 * still be wired by pmap lookups. The call to
791 * vm_page_try_remove_all() below atomically checks for such
792 * wirings and removes mappings. If the page is unmapped, the
793 * wire count is guaranteed not to increase after this check.
794 */
795 if (__predict_false(vm_page_wired(m)))
796 goto skip_page;
797
798 /*
799 * Invalid pages can be easily freed. They cannot be
800 * mapped; vm_page_free() asserts this.
801 */
802 if (vm_page_none_valid(m))
803 goto free_page;
804
805 refs = object->ref_count != 0 ? pmap_ts_referenced(m) : 0;
806
807 for (old = vm_page_astate_load(m);;) {
808 /*
809 * Check to see if the page has been removed from the
810 * queue since the first such check. Leave it alone if
811 * so, discarding any references collected by
812 * pmap_ts_referenced().
813 */
814 if (__predict_false(_vm_page_queue(old) == PQ_NONE))
815 goto skip_page;
816
817 new = old;
818 act_delta = refs;
819 if ((old.flags & PGA_REFERENCED) != 0) {
820 new.flags &= ~PGA_REFERENCED;
821 act_delta++;
822 }
823 if (act_delta == 0) {
824 ;
825 } else if (object->ref_count != 0) {
826 /*
827 * Increase the activation count if the page was
828 * referenced while in the laundry queue. This
829 * makes it less likely that the page will be
830 * returned prematurely to the laundry queue.
831 */
832 new.act_count += ACT_ADVANCE +
833 act_delta;
834 if (new.act_count > ACT_MAX)
835 new.act_count = ACT_MAX;
836
837 new.flags &= ~PGA_QUEUE_OP_MASK;
838 new.flags |= PGA_REQUEUE;
839 new.queue = PQ_ACTIVE;
840 if (!vm_page_pqstate_commit(m, &old, new))
841 continue;
842
843 /*
844 * If this was a background laundering, count
845 * activated pages towards our target. The
846 * purpose of background laundering is to ensure
847 * that pages are eventually cycled through the
848 * laundry queue, and an activation is a valid
849 * way out.
850 */
851 if (!in_shortfall)
852 launder--;
853 VM_CNT_INC(v_reactivated);
854 goto skip_page;
855 } else if ((object->flags & OBJ_DEAD) == 0) {
856 new.flags |= PGA_REQUEUE;
857 if (!vm_page_pqstate_commit(m, &old, new))
858 continue;
859 goto skip_page;
860 }
861 break;
862 }
863
864 /*
865 * If the page appears to be clean at the machine-independent
866 * layer, then remove all of its mappings from the pmap in
867 * anticipation of freeing it. If, however, any of the page's
868 * mappings allow write access, then the page may still be
869 * modified until the last of those mappings are removed.
870 */
871 if (object->ref_count != 0) {
872 vm_page_test_dirty(m);
873 if (m->dirty == 0 && !vm_page_try_remove_all(m))
874 goto skip_page;
875 }
876
877 /*
878 * Clean pages are freed, and dirty pages are paged out unless
879 * they belong to a dead object. Requeueing dirty pages from
880 * dead objects is pointless, as they are being paged out and
881 * freed by the thread that destroyed the object.
882 */
883 if (m->dirty == 0) {
884 free_page:
885 /*
886 * Now we are guaranteed that no other threads are
887 * manipulating the page, check for a last-second
888 * reference.
889 */
890 if (vm_pageout_defer(m, queue, true))
891 goto skip_page;
892 vm_page_free(m);
893 VM_CNT_INC(v_dfree);
894 } else if ((object->flags & OBJ_DEAD) == 0) {
895 if ((object->flags & OBJ_SWAP) != 0)
896 pageout_ok = disable_swap_pageouts == 0;
897 else
898 pageout_ok = true;
899 if (!pageout_ok) {
900 vm_page_launder(m);
901 goto skip_page;
902 }
903
904 /*
905 * Form a cluster with adjacent, dirty pages from the
906 * same object, and page out that entire cluster.
907 *
908 * The adjacent, dirty pages must also be in the
909 * laundry. However, their mappings are not checked
910 * for new references. Consequently, a recently
911 * referenced page may be paged out. However, that
912 * page will not be prematurely reclaimed. After page
913 * out, the page will be placed in the inactive queue,
914 * where any new references will be detected and the
915 * page reactivated.
916 */
917 error = vm_pageout_clean(m, &numpagedout);
918 if (error == 0) {
919 launder -= numpagedout;
920 ss.scanned += numpagedout;
921 } else if (error == EDEADLK) {
922 pageout_lock_miss++;
923 vnodes_skipped++;
924 }
925 object = NULL;
926 } else {
927 skip_page:
928 vm_page_xunbusy(m);
929 }
930 }
931 if (object != NULL) {
932 VM_OBJECT_WUNLOCK(object);
933 object = NULL;
934 }
935 vm_pagequeue_lock(pq);
936 vm_pageout_end_scan(&ss);
937 vm_pagequeue_unlock(pq);
938
939 if (launder > 0 && queue == PQ_UNSWAPPABLE) {
940 queue = PQ_LAUNDRY;
941 goto scan;
942 }
943
944 /*
945 * Wakeup the sync daemon if we skipped a vnode in a writeable object
946 * and we didn't launder enough pages.
947 */
948 if (vnodes_skipped > 0 && launder > 0)
949 (void)speedup_syncer();
950
951 return (starting_target - launder);
952 }
953
954 /*
955 * Compute the integer square root.
956 */
957 static u_int
isqrt(u_int num)958 isqrt(u_int num)
959 {
960 u_int bit, root, tmp;
961
962 bit = num != 0 ? (1u << ((fls(num) - 1) & ~1)) : 0;
963 root = 0;
964 while (bit != 0) {
965 tmp = root + bit;
966 root >>= 1;
967 if (num >= tmp) {
968 num -= tmp;
969 root += bit;
970 }
971 bit >>= 2;
972 }
973 return (root);
974 }
975
976 /*
977 * Perform the work of the laundry thread: periodically wake up and determine
978 * whether any pages need to be laundered. If so, determine the number of pages
979 * that need to be laundered, and launder them.
980 */
981 static void
vm_pageout_laundry_worker(void * arg)982 vm_pageout_laundry_worker(void *arg)
983 {
984 struct vm_domain *vmd;
985 struct vm_pagequeue *pq;
986 uint64_t nclean, ndirty, nfreed;
987 int domain, last_target, launder, shortfall, shortfall_cycle, target;
988 bool in_shortfall;
989
990 domain = (uintptr_t)arg;
991 vmd = VM_DOMAIN(domain);
992 pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
993 KASSERT(vmd->vmd_segs != 0, ("domain without segments"));
994
995 shortfall = 0;
996 in_shortfall = false;
997 shortfall_cycle = 0;
998 last_target = target = 0;
999 nfreed = 0;
1000
1001 /*
1002 * Calls to these handlers are serialized by the swap syscall lock.
1003 */
1004 (void)EVENTHANDLER_REGISTER(swapon, vm_pageout_swapon, vmd,
1005 EVENTHANDLER_PRI_ANY);
1006 (void)EVENTHANDLER_REGISTER(swapoff, vm_pageout_swapoff, vmd,
1007 EVENTHANDLER_PRI_ANY);
1008
1009 /*
1010 * The pageout laundry worker is never done, so loop forever.
1011 */
1012 for (;;) {
1013 KASSERT(target >= 0, ("negative target %d", target));
1014 KASSERT(shortfall_cycle >= 0,
1015 ("negative cycle %d", shortfall_cycle));
1016 launder = 0;
1017
1018 /*
1019 * First determine whether we need to launder pages to meet a
1020 * shortage of free pages.
1021 */
1022 if (shortfall > 0) {
1023 in_shortfall = true;
1024 shortfall_cycle = VM_LAUNDER_RATE / VM_INACT_SCAN_RATE;
1025 target = shortfall;
1026 } else if (!in_shortfall)
1027 goto trybackground;
1028 else if (shortfall_cycle == 0 || vm_laundry_target(vmd) <= 0) {
1029 /*
1030 * We recently entered shortfall and began laundering
1031 * pages. If we have completed that laundering run
1032 * (and we are no longer in shortfall) or we have met
1033 * our laundry target through other activity, then we
1034 * can stop laundering pages.
1035 */
1036 in_shortfall = false;
1037 target = 0;
1038 goto trybackground;
1039 }
1040 launder = target / shortfall_cycle--;
1041 goto dolaundry;
1042
1043 /*
1044 * There's no immediate need to launder any pages; see if we
1045 * meet the conditions to perform background laundering:
1046 *
1047 * 1. The ratio of dirty to clean inactive pages exceeds the
1048 * background laundering threshold, or
1049 * 2. we haven't yet reached the target of the current
1050 * background laundering run.
1051 *
1052 * The background laundering threshold is not a constant.
1053 * Instead, it is a slowly growing function of the number of
1054 * clean pages freed by the page daemon since the last
1055 * background laundering. Thus, as the ratio of dirty to
1056 * clean inactive pages grows, the amount of memory pressure
1057 * required to trigger laundering decreases. We ensure
1058 * that the threshold is non-zero after an inactive queue
1059 * scan, even if that scan failed to free a single clean page.
1060 */
1061 trybackground:
1062 nclean = vmd->vmd_free_count +
1063 vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt;
1064 ndirty = vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt;
1065 if (target == 0 && ndirty * isqrt(howmany(nfreed + 1,
1066 vmd->vmd_free_target - vmd->vmd_free_min)) >= nclean) {
1067 target = vmd->vmd_background_launder_target;
1068 }
1069
1070 /*
1071 * We have a non-zero background laundering target. If we've
1072 * laundered up to our maximum without observing a page daemon
1073 * request, just stop. This is a safety belt that ensures we
1074 * don't launder an excessive amount if memory pressure is low
1075 * and the ratio of dirty to clean pages is large. Otherwise,
1076 * proceed at the background laundering rate.
1077 */
1078 if (target > 0) {
1079 if (nfreed > 0) {
1080 nfreed = 0;
1081 last_target = target;
1082 } else if (last_target - target >=
1083 vm_background_launder_max * PAGE_SIZE / 1024) {
1084 target = 0;
1085 }
1086 launder = vm_background_launder_rate * PAGE_SIZE / 1024;
1087 launder /= VM_LAUNDER_RATE;
1088 if (launder > target)
1089 launder = target;
1090 }
1091
1092 dolaundry:
1093 if (launder > 0) {
1094 /*
1095 * Because of I/O clustering, the number of laundered
1096 * pages could exceed "target" by the maximum size of
1097 * a cluster minus one.
1098 */
1099 target -= min(vm_pageout_launder(vmd, launder,
1100 in_shortfall), target);
1101 pause("laundp", hz / VM_LAUNDER_RATE);
1102 }
1103
1104 /*
1105 * If we're not currently laundering pages and the page daemon
1106 * hasn't posted a new request, sleep until the page daemon
1107 * kicks us.
1108 */
1109 vm_pagequeue_lock(pq);
1110 if (target == 0 && vmd->vmd_laundry_request == VM_LAUNDRY_IDLE)
1111 (void)mtx_sleep(&vmd->vmd_laundry_request,
1112 vm_pagequeue_lockptr(pq), PVM, "launds", 0);
1113
1114 /*
1115 * If the pagedaemon has indicated that it's in shortfall, start
1116 * a shortfall laundering unless we're already in the middle of
1117 * one. This may preempt a background laundering.
1118 */
1119 if (vmd->vmd_laundry_request == VM_LAUNDRY_SHORTFALL &&
1120 (!in_shortfall || shortfall_cycle == 0)) {
1121 shortfall = vm_laundry_target(vmd) +
1122 vmd->vmd_pageout_deficit;
1123 target = 0;
1124 } else
1125 shortfall = 0;
1126
1127 if (target == 0)
1128 vmd->vmd_laundry_request = VM_LAUNDRY_IDLE;
1129 nfreed += vmd->vmd_clean_pages_freed;
1130 vmd->vmd_clean_pages_freed = 0;
1131 vm_pagequeue_unlock(pq);
1132 }
1133 }
1134
1135 /*
1136 * Compute the number of pages we want to try to move from the
1137 * active queue to either the inactive or laundry queue.
1138 *
1139 * When scanning active pages during a shortage, we make clean pages
1140 * count more heavily towards the page shortage than dirty pages.
1141 * This is because dirty pages must be laundered before they can be
1142 * reused and thus have less utility when attempting to quickly
1143 * alleviate a free page shortage. However, this weighting also
1144 * causes the scan to deactivate dirty pages more aggressively,
1145 * improving the effectiveness of clustering.
1146 */
1147 static int
vm_pageout_active_target(struct vm_domain * vmd)1148 vm_pageout_active_target(struct vm_domain *vmd)
1149 {
1150 int shortage;
1151
1152 shortage = vmd->vmd_inactive_target + vm_paging_target(vmd) -
1153 (vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt +
1154 vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt / act_scan_laundry_weight);
1155 shortage *= act_scan_laundry_weight;
1156 return (shortage);
1157 }
1158
1159 /*
1160 * Scan the active queue. If there is no shortage of inactive pages, scan a
1161 * small portion of the queue in order to maintain quasi-LRU.
1162 */
1163 static void
vm_pageout_scan_active(struct vm_domain * vmd,int page_shortage)1164 vm_pageout_scan_active(struct vm_domain *vmd, int page_shortage)
1165 {
1166 struct scan_state ss;
1167 vm_object_t object;
1168 vm_page_t m, marker;
1169 struct vm_pagequeue *pq;
1170 vm_page_astate_t old, new;
1171 long min_scan;
1172 int act_delta, max_scan, ps_delta, refs, scan_tick;
1173 uint8_t nqueue;
1174
1175 marker = &vmd->vmd_markers[PQ_ACTIVE];
1176 pq = &vmd->vmd_pagequeues[PQ_ACTIVE];
1177 vm_pagequeue_lock(pq);
1178
1179 /*
1180 * If we're just idle polling attempt to visit every
1181 * active page within 'update_period' seconds.
1182 */
1183 scan_tick = ticks;
1184 if (vm_pageout_update_period != 0) {
1185 min_scan = pq->pq_cnt;
1186 min_scan *= scan_tick - vmd->vmd_last_active_scan;
1187 min_scan /= hz * vm_pageout_update_period;
1188 } else
1189 min_scan = 0;
1190 if (min_scan > 0 || (page_shortage > 0 && pq->pq_cnt > 0))
1191 vmd->vmd_last_active_scan = scan_tick;
1192
1193 /*
1194 * Scan the active queue for pages that can be deactivated. Update
1195 * the per-page activity counter and use it to identify deactivation
1196 * candidates. Held pages may be deactivated.
1197 *
1198 * To avoid requeuing each page that remains in the active queue, we
1199 * implement the CLOCK algorithm. To keep the implementation of the
1200 * enqueue operation consistent for all page queues, we use two hands,
1201 * represented by marker pages. Scans begin at the first hand, which
1202 * precedes the second hand in the queue. When the two hands meet,
1203 * they are moved back to the head and tail of the queue, respectively,
1204 * and scanning resumes.
1205 */
1206 max_scan = page_shortage > 0 ? pq->pq_cnt : min_scan;
1207 act_scan:
1208 vm_pageout_init_scan(&ss, pq, marker, &vmd->vmd_clock[0], max_scan);
1209 while ((m = vm_pageout_next(&ss, false)) != NULL) {
1210 if (__predict_false(m == &vmd->vmd_clock[1])) {
1211 vm_pagequeue_lock(pq);
1212 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q);
1213 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[1], plinks.q);
1214 TAILQ_INSERT_HEAD(&pq->pq_pl, &vmd->vmd_clock[0],
1215 plinks.q);
1216 TAILQ_INSERT_TAIL(&pq->pq_pl, &vmd->vmd_clock[1],
1217 plinks.q);
1218 max_scan -= ss.scanned;
1219 vm_pageout_end_scan(&ss);
1220 goto act_scan;
1221 }
1222 if (__predict_false((m->flags & PG_MARKER) != 0))
1223 continue;
1224
1225 /*
1226 * Don't touch a page that was removed from the queue after the
1227 * page queue lock was released. Otherwise, ensure that any
1228 * pending queue operations, such as dequeues for wired pages,
1229 * are handled.
1230 */
1231 if (vm_pageout_defer(m, PQ_ACTIVE, true))
1232 continue;
1233
1234 /*
1235 * A page's object pointer may be set to NULL before
1236 * the object lock is acquired.
1237 */
1238 object = atomic_load_ptr(&m->object);
1239 if (__predict_false(object == NULL))
1240 /*
1241 * The page has been removed from its object.
1242 */
1243 continue;
1244
1245 /* Deferred free of swap space. */
1246 if ((m->a.flags & PGA_SWAP_FREE) != 0 &&
1247 VM_OBJECT_TRYWLOCK(object)) {
1248 if (m->object == object)
1249 vm_pager_page_unswapped(m);
1250 VM_OBJECT_WUNLOCK(object);
1251 }
1252
1253 /*
1254 * Check to see "how much" the page has been used.
1255 *
1256 * Test PGA_REFERENCED after calling pmap_ts_referenced() so
1257 * that a reference from a concurrently destroyed mapping is
1258 * observed here and now.
1259 *
1260 * Perform an unsynchronized object ref count check. While
1261 * the page lock ensures that the page is not reallocated to
1262 * another object, in particular, one with unmanaged mappings
1263 * that cannot support pmap_ts_referenced(), two races are,
1264 * nonetheless, possible:
1265 * 1) The count was transitioning to zero, but we saw a non-
1266 * zero value. pmap_ts_referenced() will return zero
1267 * because the page is not mapped.
1268 * 2) The count was transitioning to one, but we saw zero.
1269 * This race delays the detection of a new reference. At
1270 * worst, we will deactivate and reactivate the page.
1271 */
1272 refs = object->ref_count != 0 ? pmap_ts_referenced(m) : 0;
1273
1274 old = vm_page_astate_load(m);
1275 do {
1276 /*
1277 * Check to see if the page has been removed from the
1278 * queue since the first such check. Leave it alone if
1279 * so, discarding any references collected by
1280 * pmap_ts_referenced().
1281 */
1282 if (__predict_false(_vm_page_queue(old) == PQ_NONE)) {
1283 ps_delta = 0;
1284 break;
1285 }
1286
1287 /*
1288 * Advance or decay the act_count based on recent usage.
1289 */
1290 new = old;
1291 act_delta = refs;
1292 if ((old.flags & PGA_REFERENCED) != 0) {
1293 new.flags &= ~PGA_REFERENCED;
1294 act_delta++;
1295 }
1296 if (act_delta != 0) {
1297 new.act_count += ACT_ADVANCE + act_delta;
1298 if (new.act_count > ACT_MAX)
1299 new.act_count = ACT_MAX;
1300 } else {
1301 new.act_count -= min(new.act_count,
1302 ACT_DECLINE);
1303 }
1304
1305 if (new.act_count > 0) {
1306 /*
1307 * Adjust the activation count and keep the page
1308 * in the active queue. The count might be left
1309 * unchanged if it is saturated. The page may
1310 * have been moved to a different queue since we
1311 * started the scan, in which case we move it
1312 * back.
1313 */
1314 ps_delta = 0;
1315 if (old.queue != PQ_ACTIVE) {
1316 new.flags &= ~PGA_QUEUE_OP_MASK;
1317 new.flags |= PGA_REQUEUE;
1318 new.queue = PQ_ACTIVE;
1319 }
1320 } else {
1321 /*
1322 * When not short for inactive pages, let dirty
1323 * pages go through the inactive queue before
1324 * moving to the laundry queue. This gives them
1325 * some extra time to be reactivated,
1326 * potentially avoiding an expensive pageout.
1327 * However, during a page shortage, the inactive
1328 * queue is necessarily small, and so dirty
1329 * pages would only spend a trivial amount of
1330 * time in the inactive queue. Therefore, we
1331 * might as well place them directly in the
1332 * laundry queue to reduce queuing overhead.
1333 *
1334 * Calling vm_page_test_dirty() here would
1335 * require acquisition of the object's write
1336 * lock. However, during a page shortage,
1337 * directing dirty pages into the laundry queue
1338 * is only an optimization and not a
1339 * requirement. Therefore, we simply rely on
1340 * the opportunistic updates to the page's dirty
1341 * field by the pmap.
1342 */
1343 if (page_shortage <= 0) {
1344 nqueue = PQ_INACTIVE;
1345 ps_delta = 0;
1346 } else if (m->dirty == 0) {
1347 nqueue = PQ_INACTIVE;
1348 ps_delta = act_scan_laundry_weight;
1349 } else {
1350 nqueue = PQ_LAUNDRY;
1351 ps_delta = 1;
1352 }
1353
1354 new.flags &= ~PGA_QUEUE_OP_MASK;
1355 new.flags |= PGA_REQUEUE;
1356 new.queue = nqueue;
1357 }
1358 } while (!vm_page_pqstate_commit(m, &old, new));
1359
1360 page_shortage -= ps_delta;
1361 }
1362 vm_pagequeue_lock(pq);
1363 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q);
1364 TAILQ_INSERT_AFTER(&pq->pq_pl, marker, &vmd->vmd_clock[0], plinks.q);
1365 vm_pageout_end_scan(&ss);
1366 vm_pagequeue_unlock(pq);
1367 }
1368
1369 static int
vm_pageout_reinsert_inactive_page(struct vm_pagequeue * pq,vm_page_t marker,vm_page_t m)1370 vm_pageout_reinsert_inactive_page(struct vm_pagequeue *pq, vm_page_t marker,
1371 vm_page_t m)
1372 {
1373 vm_page_astate_t as;
1374
1375 vm_pagequeue_assert_locked(pq);
1376
1377 as = vm_page_astate_load(m);
1378 if (as.queue != PQ_INACTIVE || (as.flags & PGA_ENQUEUED) != 0)
1379 return (0);
1380 vm_page_aflag_set(m, PGA_ENQUEUED);
1381 TAILQ_INSERT_BEFORE(marker, m, plinks.q);
1382 return (1);
1383 }
1384
1385 /*
1386 * Re-add stuck pages to the inactive queue. We will examine them again
1387 * during the next scan. If the queue state of a page has changed since
1388 * it was physically removed from the page queue in
1389 * vm_pageout_collect_batch(), don't do anything with that page.
1390 */
1391 static void
vm_pageout_reinsert_inactive(struct scan_state * ss,struct vm_batchqueue * bq,vm_page_t m)1392 vm_pageout_reinsert_inactive(struct scan_state *ss, struct vm_batchqueue *bq,
1393 vm_page_t m)
1394 {
1395 struct vm_pagequeue *pq;
1396 vm_page_t marker;
1397 int delta;
1398
1399 delta = 0;
1400 marker = ss->marker;
1401 pq = ss->pq;
1402
1403 if (m != NULL) {
1404 if (vm_batchqueue_insert(bq, m) != 0)
1405 return;
1406 vm_pagequeue_lock(pq);
1407 delta += vm_pageout_reinsert_inactive_page(pq, marker, m);
1408 } else
1409 vm_pagequeue_lock(pq);
1410 while ((m = vm_batchqueue_pop(bq)) != NULL)
1411 delta += vm_pageout_reinsert_inactive_page(pq, marker, m);
1412 vm_pagequeue_cnt_add(pq, delta);
1413 vm_pagequeue_unlock(pq);
1414 vm_batchqueue_init(bq);
1415 }
1416
1417 static void
vm_pageout_scan_inactive(struct vm_domain * vmd,int page_shortage)1418 vm_pageout_scan_inactive(struct vm_domain *vmd, int page_shortage)
1419 {
1420 struct timeval start, end;
1421 struct scan_state ss;
1422 struct vm_batchqueue rq;
1423 struct vm_page marker_page;
1424 vm_page_t m, marker;
1425 struct vm_pagequeue *pq;
1426 vm_object_t object;
1427 vm_page_astate_t old, new;
1428 int act_delta, addl_page_shortage, starting_page_shortage, refs;
1429
1430 object = NULL;
1431 vm_batchqueue_init(&rq);
1432 getmicrouptime(&start);
1433
1434 /*
1435 * The addl_page_shortage is an estimate of the number of temporarily
1436 * stuck pages in the inactive queue. In other words, the
1437 * number of pages from the inactive count that should be
1438 * discounted in setting the target for the active queue scan.
1439 */
1440 addl_page_shortage = 0;
1441
1442 /*
1443 * Start scanning the inactive queue for pages that we can free. The
1444 * scan will stop when we reach the target or we have scanned the
1445 * entire queue. (Note that m->a.act_count is not used to make
1446 * decisions for the inactive queue, only for the active queue.)
1447 */
1448 starting_page_shortage = page_shortage;
1449 marker = &marker_page;
1450 vm_page_init_marker(marker, PQ_INACTIVE, 0);
1451 pq = &vmd->vmd_pagequeues[PQ_INACTIVE];
1452 vm_pagequeue_lock(pq);
1453 vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt);
1454 while (page_shortage > 0) {
1455 /*
1456 * If we need to refill the scan batch queue, release any
1457 * optimistically held object lock. This gives someone else a
1458 * chance to grab the lock, and also avoids holding it while we
1459 * do unrelated work.
1460 */
1461 if (object != NULL && vm_batchqueue_empty(&ss.bq)) {
1462 VM_OBJECT_WUNLOCK(object);
1463 object = NULL;
1464 }
1465
1466 m = vm_pageout_next(&ss, true);
1467 if (m == NULL)
1468 break;
1469 KASSERT((m->flags & PG_MARKER) == 0,
1470 ("marker page %p was dequeued", m));
1471
1472 /*
1473 * Don't touch a page that was removed from the queue after the
1474 * page queue lock was released. Otherwise, ensure that any
1475 * pending queue operations, such as dequeues for wired pages,
1476 * are handled.
1477 */
1478 if (vm_pageout_defer(m, PQ_INACTIVE, false))
1479 continue;
1480
1481 /*
1482 * Lock the page's object.
1483 */
1484 if (object == NULL || object != m->object) {
1485 if (object != NULL)
1486 VM_OBJECT_WUNLOCK(object);
1487 object = atomic_load_ptr(&m->object);
1488 if (__predict_false(object == NULL))
1489 /* The page is being freed by another thread. */
1490 continue;
1491
1492 /* Depends on type-stability. */
1493 VM_OBJECT_WLOCK(object);
1494 if (__predict_false(m->object != object)) {
1495 VM_OBJECT_WUNLOCK(object);
1496 object = NULL;
1497 goto reinsert;
1498 }
1499 }
1500
1501 if (vm_page_tryxbusy(m) == 0) {
1502 /*
1503 * Don't mess with busy pages. Leave them at
1504 * the front of the queue. Most likely, they
1505 * are being paged out and will leave the
1506 * queue shortly after the scan finishes. So,
1507 * they ought to be discounted from the
1508 * inactive count.
1509 */
1510 addl_page_shortage++;
1511 goto reinsert;
1512 }
1513
1514 /* Deferred free of swap space. */
1515 if ((m->a.flags & PGA_SWAP_FREE) != 0)
1516 vm_pager_page_unswapped(m);
1517
1518 /*
1519 * Check for wirings now that we hold the object lock and have
1520 * exclusively busied the page. If the page is mapped, it may
1521 * still be wired by pmap lookups. The call to
1522 * vm_page_try_remove_all() below atomically checks for such
1523 * wirings and removes mappings. If the page is unmapped, the
1524 * wire count is guaranteed not to increase after this check.
1525 */
1526 if (__predict_false(vm_page_wired(m)))
1527 goto skip_page;
1528
1529 /*
1530 * Invalid pages can be easily freed. They cannot be
1531 * mapped, vm_page_free() asserts this.
1532 */
1533 if (vm_page_none_valid(m))
1534 goto free_page;
1535
1536 refs = object->ref_count != 0 ? pmap_ts_referenced(m) : 0;
1537
1538 for (old = vm_page_astate_load(m);;) {
1539 /*
1540 * Check to see if the page has been removed from the
1541 * queue since the first such check. Leave it alone if
1542 * so, discarding any references collected by
1543 * pmap_ts_referenced().
1544 */
1545 if (__predict_false(_vm_page_queue(old) == PQ_NONE))
1546 goto skip_page;
1547
1548 new = old;
1549 act_delta = refs;
1550 if ((old.flags & PGA_REFERENCED) != 0) {
1551 new.flags &= ~PGA_REFERENCED;
1552 act_delta++;
1553 }
1554 if (act_delta == 0) {
1555 ;
1556 } else if (object->ref_count != 0) {
1557 /*
1558 * Increase the activation count if the
1559 * page was referenced while in the
1560 * inactive queue. This makes it less
1561 * likely that the page will be returned
1562 * prematurely to the inactive queue.
1563 */
1564 new.act_count += ACT_ADVANCE +
1565 act_delta;
1566 if (new.act_count > ACT_MAX)
1567 new.act_count = ACT_MAX;
1568
1569 new.flags &= ~PGA_QUEUE_OP_MASK;
1570 new.flags |= PGA_REQUEUE;
1571 new.queue = PQ_ACTIVE;
1572 if (!vm_page_pqstate_commit(m, &old, new))
1573 continue;
1574
1575 VM_CNT_INC(v_reactivated);
1576 goto skip_page;
1577 } else if ((object->flags & OBJ_DEAD) == 0) {
1578 new.queue = PQ_INACTIVE;
1579 new.flags |= PGA_REQUEUE;
1580 if (!vm_page_pqstate_commit(m, &old, new))
1581 continue;
1582 goto skip_page;
1583 }
1584 break;
1585 }
1586
1587 /*
1588 * If the page appears to be clean at the machine-independent
1589 * layer, then remove all of its mappings from the pmap in
1590 * anticipation of freeing it. If, however, any of the page's
1591 * mappings allow write access, then the page may still be
1592 * modified until the last of those mappings are removed.
1593 */
1594 if (object->ref_count != 0) {
1595 vm_page_test_dirty(m);
1596 if (m->dirty == 0 && !vm_page_try_remove_all(m))
1597 goto skip_page;
1598 }
1599
1600 /*
1601 * Clean pages can be freed, but dirty pages must be sent back
1602 * to the laundry, unless they belong to a dead object.
1603 * Requeueing dirty pages from dead objects is pointless, as
1604 * they are being paged out and freed by the thread that
1605 * destroyed the object.
1606 */
1607 if (m->dirty == 0) {
1608 free_page:
1609 /*
1610 * Now we are guaranteed that no other threads are
1611 * manipulating the page, check for a last-second
1612 * reference that would save it from doom.
1613 */
1614 if (vm_pageout_defer(m, PQ_INACTIVE, false))
1615 goto skip_page;
1616
1617 /*
1618 * Because we dequeued the page and have already checked
1619 * for pending dequeue and enqueue requests, we can
1620 * safely disassociate the page from the inactive queue
1621 * without holding the queue lock.
1622 */
1623 m->a.queue = PQ_NONE;
1624 vm_page_free(m);
1625 page_shortage--;
1626 continue;
1627 }
1628 if ((object->flags & OBJ_DEAD) == 0)
1629 vm_page_launder(m);
1630 skip_page:
1631 vm_page_xunbusy(m);
1632 continue;
1633 reinsert:
1634 vm_pageout_reinsert_inactive(&ss, &rq, m);
1635 }
1636 if (object != NULL)
1637 VM_OBJECT_WUNLOCK(object);
1638 vm_pageout_reinsert_inactive(&ss, &rq, NULL);
1639 vm_pageout_reinsert_inactive(&ss, &ss.bq, NULL);
1640 vm_pagequeue_lock(pq);
1641 vm_pageout_end_scan(&ss);
1642 vm_pagequeue_unlock(pq);
1643
1644 /*
1645 * Record the remaining shortage and the progress and rate it was made.
1646 */
1647 atomic_add_int(&vmd->vmd_addl_shortage, addl_page_shortage);
1648 getmicrouptime(&end);
1649 timevalsub(&end, &start);
1650 atomic_add_int(&vmd->vmd_inactive_us,
1651 end.tv_sec * 1000000 + end.tv_usec);
1652 atomic_add_int(&vmd->vmd_inactive_freed,
1653 starting_page_shortage - page_shortage);
1654 }
1655
1656 /*
1657 * Dispatch a number of inactive threads according to load and collect the
1658 * results to present a coherent view of paging activity on this domain.
1659 */
1660 static int
vm_pageout_inactive_dispatch(struct vm_domain * vmd,int shortage)1661 vm_pageout_inactive_dispatch(struct vm_domain *vmd, int shortage)
1662 {
1663 u_int freed, pps, slop, threads, us;
1664
1665 vmd->vmd_inactive_shortage = shortage;
1666 slop = 0;
1667
1668 /*
1669 * If we have more work than we can do in a quarter of our interval, we
1670 * fire off multiple threads to process it.
1671 */
1672 threads = vmd->vmd_inactive_threads;
1673 if (threads > 1 && vmd->vmd_inactive_pps != 0 &&
1674 shortage > vmd->vmd_inactive_pps / VM_INACT_SCAN_RATE / 4) {
1675 vmd->vmd_inactive_shortage /= threads;
1676 slop = shortage % threads;
1677 vm_domain_pageout_lock(vmd);
1678 blockcount_acquire(&vmd->vmd_inactive_starting, threads - 1);
1679 blockcount_acquire(&vmd->vmd_inactive_running, threads - 1);
1680 wakeup(&vmd->vmd_inactive_shortage);
1681 vm_domain_pageout_unlock(vmd);
1682 }
1683
1684 /* Run the local thread scan. */
1685 vm_pageout_scan_inactive(vmd, vmd->vmd_inactive_shortage + slop);
1686
1687 /*
1688 * Block until helper threads report results and then accumulate
1689 * totals.
1690 */
1691 blockcount_wait(&vmd->vmd_inactive_running, NULL, "vmpoid", PVM);
1692 freed = atomic_readandclear_int(&vmd->vmd_inactive_freed);
1693 VM_CNT_ADD(v_dfree, freed);
1694
1695 /*
1696 * Calculate the per-thread paging rate with an exponential decay of
1697 * prior results. Careful to avoid integer rounding errors with large
1698 * us values.
1699 */
1700 us = max(atomic_readandclear_int(&vmd->vmd_inactive_us), 1);
1701 if (us > 1000000)
1702 /* Keep rounding to tenths */
1703 pps = (freed * 10) / ((us * 10) / 1000000);
1704 else
1705 pps = (1000000 / us) * freed;
1706 vmd->vmd_inactive_pps = (vmd->vmd_inactive_pps / 2) + (pps / 2);
1707
1708 return (shortage - freed);
1709 }
1710
1711 /*
1712 * Attempt to reclaim the requested number of pages from the inactive queue.
1713 * Returns true if the shortage was addressed.
1714 */
1715 static int
vm_pageout_inactive(struct vm_domain * vmd,int shortage,int * addl_shortage)1716 vm_pageout_inactive(struct vm_domain *vmd, int shortage, int *addl_shortage)
1717 {
1718 struct vm_pagequeue *pq;
1719 u_int addl_page_shortage, deficit, page_shortage;
1720 u_int starting_page_shortage;
1721
1722 /*
1723 * vmd_pageout_deficit counts the number of pages requested in
1724 * allocations that failed because of a free page shortage. We assume
1725 * that the allocations will be reattempted and thus include the deficit
1726 * in our scan target.
1727 */
1728 deficit = atomic_readandclear_int(&vmd->vmd_pageout_deficit);
1729 starting_page_shortage = shortage + deficit;
1730
1731 /*
1732 * Run the inactive scan on as many threads as is necessary.
1733 */
1734 page_shortage = vm_pageout_inactive_dispatch(vmd, starting_page_shortage);
1735 addl_page_shortage = atomic_readandclear_int(&vmd->vmd_addl_shortage);
1736
1737 /*
1738 * Wake up the laundry thread so that it can perform any needed
1739 * laundering. If we didn't meet our target, we're in shortfall and
1740 * need to launder more aggressively. If PQ_LAUNDRY is empty and no
1741 * swap devices are configured, the laundry thread has no work to do, so
1742 * don't bother waking it up.
1743 *
1744 * The laundry thread uses the number of inactive queue scans elapsed
1745 * since the last laundering to determine whether to launder again, so
1746 * keep count.
1747 */
1748 if (starting_page_shortage > 0) {
1749 pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
1750 vm_pagequeue_lock(pq);
1751 if (vmd->vmd_laundry_request == VM_LAUNDRY_IDLE &&
1752 (pq->pq_cnt > 0 || atomic_load_acq_int(&swapdev_enabled))) {
1753 if (page_shortage > 0) {
1754 vmd->vmd_laundry_request = VM_LAUNDRY_SHORTFALL;
1755 VM_CNT_INC(v_pdshortfalls);
1756 } else if (vmd->vmd_laundry_request !=
1757 VM_LAUNDRY_SHORTFALL)
1758 vmd->vmd_laundry_request =
1759 VM_LAUNDRY_BACKGROUND;
1760 wakeup(&vmd->vmd_laundry_request);
1761 }
1762 vmd->vmd_clean_pages_freed +=
1763 starting_page_shortage - page_shortage;
1764 vm_pagequeue_unlock(pq);
1765 }
1766
1767 /*
1768 * Wakeup the swapout daemon if we didn't free the targeted number of
1769 * pages.
1770 */
1771 if (page_shortage > 0)
1772 vm_swapout_run();
1773
1774 /*
1775 * If the inactive queue scan fails repeatedly to meet its
1776 * target, kill the largest process.
1777 */
1778 vm_pageout_mightbe_oom(vmd, page_shortage, starting_page_shortage);
1779
1780 /*
1781 * Reclaim pages by swapping out idle processes, if configured to do so.
1782 */
1783 vm_swapout_run_idle();
1784
1785 /*
1786 * See the description of addl_page_shortage above.
1787 */
1788 *addl_shortage = addl_page_shortage + deficit;
1789
1790 return (page_shortage <= 0);
1791 }
1792
1793 static int vm_pageout_oom_vote;
1794
1795 /*
1796 * The pagedaemon threads randlomly select one to perform the
1797 * OOM. Trying to kill processes before all pagedaemons
1798 * failed to reach free target is premature.
1799 */
1800 static void
vm_pageout_mightbe_oom(struct vm_domain * vmd,int page_shortage,int starting_page_shortage)1801 vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
1802 int starting_page_shortage)
1803 {
1804 int old_vote;
1805
1806 if (starting_page_shortage <= 0 || starting_page_shortage !=
1807 page_shortage)
1808 vmd->vmd_oom_seq = 0;
1809 else
1810 vmd->vmd_oom_seq++;
1811 if (vmd->vmd_oom_seq < vm_pageout_oom_seq) {
1812 if (vmd->vmd_oom) {
1813 vmd->vmd_oom = FALSE;
1814 atomic_subtract_int(&vm_pageout_oom_vote, 1);
1815 }
1816 return;
1817 }
1818
1819 /*
1820 * Do not follow the call sequence until OOM condition is
1821 * cleared.
1822 */
1823 vmd->vmd_oom_seq = 0;
1824
1825 if (vmd->vmd_oom)
1826 return;
1827
1828 vmd->vmd_oom = TRUE;
1829 old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1);
1830 if (old_vote != vm_ndomains - 1)
1831 return;
1832
1833 /*
1834 * The current pagedaemon thread is the last in the quorum to
1835 * start OOM. Initiate the selection and signaling of the
1836 * victim.
1837 */
1838 vm_pageout_oom(VM_OOM_MEM);
1839
1840 /*
1841 * After one round of OOM terror, recall our vote. On the
1842 * next pass, current pagedaemon would vote again if the low
1843 * memory condition is still there, due to vmd_oom being
1844 * false.
1845 */
1846 vmd->vmd_oom = FALSE;
1847 atomic_subtract_int(&vm_pageout_oom_vote, 1);
1848 }
1849
1850 /*
1851 * The OOM killer is the page daemon's action of last resort when
1852 * memory allocation requests have been stalled for a prolonged period
1853 * of time because it cannot reclaim memory. This function computes
1854 * the approximate number of physical pages that could be reclaimed if
1855 * the specified address space is destroyed.
1856 *
1857 * Private, anonymous memory owned by the address space is the
1858 * principal resource that we expect to recover after an OOM kill.
1859 * Since the physical pages mapped by the address space's COW entries
1860 * are typically shared pages, they are unlikely to be released and so
1861 * they are not counted.
1862 *
1863 * To get to the point where the page daemon runs the OOM killer, its
1864 * efforts to write-back vnode-backed pages may have stalled. This
1865 * could be caused by a memory allocation deadlock in the write path
1866 * that might be resolved by an OOM kill. Therefore, physical pages
1867 * belonging to vnode-backed objects are counted, because they might
1868 * be freed without being written out first if the address space holds
1869 * the last reference to an unlinked vnode.
1870 *
1871 * Similarly, physical pages belonging to OBJT_PHYS objects are
1872 * counted because the address space might hold the last reference to
1873 * the object.
1874 */
1875 static long
vm_pageout_oom_pagecount(struct vmspace * vmspace)1876 vm_pageout_oom_pagecount(struct vmspace *vmspace)
1877 {
1878 vm_map_t map;
1879 vm_map_entry_t entry;
1880 vm_object_t obj;
1881 long res;
1882
1883 map = &vmspace->vm_map;
1884 KASSERT(!map->system_map, ("system map"));
1885 sx_assert(&map->lock, SA_LOCKED);
1886 res = 0;
1887 VM_MAP_ENTRY_FOREACH(entry, map) {
1888 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
1889 continue;
1890 obj = entry->object.vm_object;
1891 if (obj == NULL)
1892 continue;
1893 if ((entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0 &&
1894 obj->ref_count != 1)
1895 continue;
1896 if (obj->type == OBJT_PHYS || obj->type == OBJT_VNODE ||
1897 (obj->flags & OBJ_SWAP) != 0)
1898 res += obj->resident_page_count;
1899 }
1900 return (res);
1901 }
1902
1903 static int vm_oom_ratelim_last;
1904 static int vm_oom_pf_secs = 10;
1905 SYSCTL_INT(_vm, OID_AUTO, oom_pf_secs, CTLFLAG_RWTUN, &vm_oom_pf_secs, 0,
1906 "");
1907 static struct mtx vm_oom_ratelim_mtx;
1908
1909 void
vm_pageout_oom(int shortage)1910 vm_pageout_oom(int shortage)
1911 {
1912 const char *reason;
1913 struct proc *p, *bigproc;
1914 vm_offset_t size, bigsize;
1915 struct thread *td;
1916 struct vmspace *vm;
1917 int now;
1918 bool breakout;
1919
1920 /*
1921 * For OOM requests originating from vm_fault(), there is a high
1922 * chance that a single large process faults simultaneously in
1923 * several threads. Also, on an active system running many
1924 * processes of middle-size, like buildworld, all of them
1925 * could fault almost simultaneously as well.
1926 *
1927 * To avoid killing too many processes, rate-limit OOMs
1928 * initiated by vm_fault() time-outs on the waits for free
1929 * pages.
1930 */
1931 mtx_lock(&vm_oom_ratelim_mtx);
1932 now = ticks;
1933 if (shortage == VM_OOM_MEM_PF &&
1934 (u_int)(now - vm_oom_ratelim_last) < hz * vm_oom_pf_secs) {
1935 mtx_unlock(&vm_oom_ratelim_mtx);
1936 return;
1937 }
1938 vm_oom_ratelim_last = now;
1939 mtx_unlock(&vm_oom_ratelim_mtx);
1940
1941 /*
1942 * We keep the process bigproc locked once we find it to keep anyone
1943 * from messing with it; however, there is a possibility of
1944 * deadlock if process B is bigproc and one of its child processes
1945 * attempts to propagate a signal to B while we are waiting for A's
1946 * lock while walking this list. To avoid this, we don't block on
1947 * the process lock but just skip a process if it is already locked.
1948 */
1949 bigproc = NULL;
1950 bigsize = 0;
1951 sx_slock(&allproc_lock);
1952 FOREACH_PROC_IN_SYSTEM(p) {
1953 PROC_LOCK(p);
1954
1955 /*
1956 * If this is a system, protected or killed process, skip it.
1957 */
1958 if (p->p_state != PRS_NORMAL || (p->p_flag & (P_INEXEC |
1959 P_PROTECTED | P_SYSTEM | P_WEXIT)) != 0 ||
1960 p->p_pid == 1 || P_KILLED(p) ||
1961 (p->p_pid < 48 && swap_pager_avail != 0)) {
1962 PROC_UNLOCK(p);
1963 continue;
1964 }
1965 /*
1966 * If the process is in a non-running type state,
1967 * don't touch it. Check all the threads individually.
1968 */
1969 breakout = false;
1970 FOREACH_THREAD_IN_PROC(p, td) {
1971 thread_lock(td);
1972 if (!TD_ON_RUNQ(td) &&
1973 !TD_IS_RUNNING(td) &&
1974 !TD_IS_SLEEPING(td) &&
1975 !TD_IS_SUSPENDED(td) &&
1976 !TD_IS_SWAPPED(td)) {
1977 thread_unlock(td);
1978 breakout = true;
1979 break;
1980 }
1981 thread_unlock(td);
1982 }
1983 if (breakout) {
1984 PROC_UNLOCK(p);
1985 continue;
1986 }
1987 /*
1988 * get the process size
1989 */
1990 vm = vmspace_acquire_ref(p);
1991 if (vm == NULL) {
1992 PROC_UNLOCK(p);
1993 continue;
1994 }
1995 _PHOLD_LITE(p);
1996 PROC_UNLOCK(p);
1997 sx_sunlock(&allproc_lock);
1998 if (!vm_map_trylock_read(&vm->vm_map)) {
1999 vmspace_free(vm);
2000 sx_slock(&allproc_lock);
2001 PRELE(p);
2002 continue;
2003 }
2004 size = vmspace_swap_count(vm);
2005 if (shortage == VM_OOM_MEM || shortage == VM_OOM_MEM_PF)
2006 size += vm_pageout_oom_pagecount(vm);
2007 vm_map_unlock_read(&vm->vm_map);
2008 vmspace_free(vm);
2009 sx_slock(&allproc_lock);
2010
2011 /*
2012 * If this process is bigger than the biggest one,
2013 * remember it.
2014 */
2015 if (size > bigsize) {
2016 if (bigproc != NULL)
2017 PRELE(bigproc);
2018 bigproc = p;
2019 bigsize = size;
2020 } else {
2021 PRELE(p);
2022 }
2023 }
2024 sx_sunlock(&allproc_lock);
2025
2026 if (bigproc != NULL) {
2027 switch (shortage) {
2028 case VM_OOM_MEM:
2029 reason = "failed to reclaim memory";
2030 break;
2031 case VM_OOM_MEM_PF:
2032 reason = "a thread waited too long to allocate a page";
2033 break;
2034 case VM_OOM_SWAPZ:
2035 reason = "out of swap space";
2036 break;
2037 default:
2038 panic("unknown OOM reason %d", shortage);
2039 }
2040 if (vm_panic_on_oom != 0 && --vm_panic_on_oom == 0)
2041 panic("%s", reason);
2042 PROC_LOCK(bigproc);
2043 killproc(bigproc, reason);
2044 sched_nice(bigproc, PRIO_MIN);
2045 _PRELE(bigproc);
2046 PROC_UNLOCK(bigproc);
2047 }
2048 }
2049
2050 /*
2051 * Signal a free page shortage to subsystems that have registered an event
2052 * handler. Reclaim memory from UMA in the event of a severe shortage.
2053 * Return true if the free page count should be re-evaluated.
2054 */
2055 static bool
vm_pageout_lowmem(void)2056 vm_pageout_lowmem(void)
2057 {
2058 static int lowmem_ticks = 0;
2059 int last;
2060 bool ret;
2061
2062 ret = false;
2063
2064 last = atomic_load_int(&lowmem_ticks);
2065 while ((u_int)(ticks - last) / hz >= lowmem_period) {
2066 if (atomic_fcmpset_int(&lowmem_ticks, &last, ticks) == 0)
2067 continue;
2068
2069 /*
2070 * Decrease registered cache sizes.
2071 */
2072 SDT_PROBE0(vm, , , vm__lowmem_scan);
2073 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_PAGES);
2074
2075 /*
2076 * We do this explicitly after the caches have been
2077 * drained above.
2078 */
2079 uma_reclaim(UMA_RECLAIM_TRIM);
2080 ret = true;
2081 break;
2082 }
2083
2084 /*
2085 * Kick off an asynchronous reclaim of cached memory if one of the
2086 * page daemons is failing to keep up with demand. Use the "severe"
2087 * threshold instead of "min" to ensure that we do not blow away the
2088 * caches if a subset of the NUMA domains are depleted by kernel memory
2089 * allocations; the domainset iterators automatically skip domains
2090 * below the "min" threshold on the first pass.
2091 *
2092 * UMA reclaim worker has its own rate-limiting mechanism, so don't
2093 * worry about kicking it too often.
2094 */
2095 if (vm_page_count_severe())
2096 uma_reclaim_wakeup();
2097
2098 return (ret);
2099 }
2100
2101 static void
vm_pageout_worker(void * arg)2102 vm_pageout_worker(void *arg)
2103 {
2104 struct vm_domain *vmd;
2105 u_int ofree;
2106 int addl_shortage, domain, shortage;
2107 bool target_met;
2108
2109 domain = (uintptr_t)arg;
2110 vmd = VM_DOMAIN(domain);
2111 shortage = 0;
2112 target_met = true;
2113
2114 /*
2115 * XXXKIB It could be useful to bind pageout daemon threads to
2116 * the cores belonging to the domain, from which vm_page_array
2117 * is allocated.
2118 */
2119
2120 KASSERT(vmd->vmd_segs != 0, ("domain without segments"));
2121 vmd->vmd_last_active_scan = ticks;
2122
2123 /*
2124 * The pageout daemon worker is never done, so loop forever.
2125 */
2126 while (TRUE) {
2127 vm_domain_pageout_lock(vmd);
2128
2129 /*
2130 * We need to clear wanted before we check the limits. This
2131 * prevents races with wakers who will check wanted after they
2132 * reach the limit.
2133 */
2134 atomic_store_int(&vmd->vmd_pageout_wanted, 0);
2135
2136 /*
2137 * Might the page daemon need to run again?
2138 */
2139 if (vm_paging_needed(vmd, vmd->vmd_free_count)) {
2140 /*
2141 * Yes. If the scan failed to produce enough free
2142 * pages, sleep uninterruptibly for some time in the
2143 * hope that the laundry thread will clean some pages.
2144 */
2145 vm_domain_pageout_unlock(vmd);
2146 if (!target_met)
2147 pause("pwait", hz / VM_INACT_SCAN_RATE);
2148 } else {
2149 /*
2150 * No, sleep until the next wakeup or until pages
2151 * need to have their reference stats updated.
2152 */
2153 if (mtx_sleep(&vmd->vmd_pageout_wanted,
2154 vm_domain_pageout_lockptr(vmd), PDROP | PVM,
2155 "psleep", hz / VM_INACT_SCAN_RATE) == 0)
2156 VM_CNT_INC(v_pdwakeups);
2157 }
2158
2159 /* Prevent spurious wakeups by ensuring that wanted is set. */
2160 atomic_store_int(&vmd->vmd_pageout_wanted, 1);
2161
2162 /*
2163 * Use the controller to calculate how many pages to free in
2164 * this interval, and scan the inactive queue. If the lowmem
2165 * handlers appear to have freed up some pages, subtract the
2166 * difference from the inactive queue scan target.
2167 */
2168 shortage = pidctrl_daemon(&vmd->vmd_pid, vmd->vmd_free_count);
2169 if (shortage > 0) {
2170 ofree = vmd->vmd_free_count;
2171 if (vm_pageout_lowmem() && vmd->vmd_free_count > ofree)
2172 shortage -= min(vmd->vmd_free_count - ofree,
2173 (u_int)shortage);
2174 target_met = vm_pageout_inactive(vmd, shortage,
2175 &addl_shortage);
2176 } else
2177 addl_shortage = 0;
2178
2179 /*
2180 * Scan the active queue. A positive value for shortage
2181 * indicates that we must aggressively deactivate pages to avoid
2182 * a shortfall.
2183 */
2184 shortage = vm_pageout_active_target(vmd) + addl_shortage;
2185 vm_pageout_scan_active(vmd, shortage);
2186 }
2187 }
2188
2189 /*
2190 * vm_pageout_helper runs additional pageout daemons in times of high paging
2191 * activity.
2192 */
2193 static void
vm_pageout_helper(void * arg)2194 vm_pageout_helper(void *arg)
2195 {
2196 struct vm_domain *vmd;
2197 int domain;
2198
2199 domain = (uintptr_t)arg;
2200 vmd = VM_DOMAIN(domain);
2201
2202 vm_domain_pageout_lock(vmd);
2203 for (;;) {
2204 msleep(&vmd->vmd_inactive_shortage,
2205 vm_domain_pageout_lockptr(vmd), PVM, "psleep", 0);
2206 blockcount_release(&vmd->vmd_inactive_starting, 1);
2207
2208 vm_domain_pageout_unlock(vmd);
2209 vm_pageout_scan_inactive(vmd, vmd->vmd_inactive_shortage);
2210 vm_domain_pageout_lock(vmd);
2211
2212 /*
2213 * Release the running count while the pageout lock is held to
2214 * prevent wakeup races.
2215 */
2216 blockcount_release(&vmd->vmd_inactive_running, 1);
2217 }
2218 }
2219
2220 static int
get_pageout_threads_per_domain(const struct vm_domain * vmd)2221 get_pageout_threads_per_domain(const struct vm_domain *vmd)
2222 {
2223 unsigned total_pageout_threads, eligible_cpus, domain_cpus;
2224
2225 if (VM_DOMAIN_EMPTY(vmd->vmd_domain))
2226 return (0);
2227
2228 /*
2229 * Semi-arbitrarily constrain pagedaemon threads to less than half the
2230 * total number of CPUs in the system as an upper limit.
2231 */
2232 if (pageout_cpus_per_thread < 2)
2233 pageout_cpus_per_thread = 2;
2234 else if (pageout_cpus_per_thread > mp_ncpus)
2235 pageout_cpus_per_thread = mp_ncpus;
2236
2237 total_pageout_threads = howmany(mp_ncpus, pageout_cpus_per_thread);
2238 domain_cpus = CPU_COUNT(&cpuset_domain[vmd->vmd_domain]);
2239
2240 /* Pagedaemons are not run in empty domains. */
2241 eligible_cpus = mp_ncpus;
2242 for (unsigned i = 0; i < vm_ndomains; i++)
2243 if (VM_DOMAIN_EMPTY(i))
2244 eligible_cpus -= CPU_COUNT(&cpuset_domain[i]);
2245
2246 /*
2247 * Assign a portion of the total pageout threads to this domain
2248 * corresponding to the fraction of pagedaemon-eligible CPUs in the
2249 * domain. In asymmetric NUMA systems, domains with more CPUs may be
2250 * allocated more threads than domains with fewer CPUs.
2251 */
2252 return (howmany(total_pageout_threads * domain_cpus, eligible_cpus));
2253 }
2254
2255 /*
2256 * Initialize basic pageout daemon settings. See the comment above the
2257 * definition of vm_domain for some explanation of how these thresholds are
2258 * used.
2259 */
2260 static void
vm_pageout_init_domain(int domain)2261 vm_pageout_init_domain(int domain)
2262 {
2263 struct vm_domain *vmd;
2264 struct sysctl_oid *oid;
2265
2266 vmd = VM_DOMAIN(domain);
2267 vmd->vmd_interrupt_free_min = 2;
2268
2269 /*
2270 * v_free_reserved needs to include enough for the largest
2271 * swap pager structures plus enough for any pv_entry structs
2272 * when paging.
2273 */
2274 vmd->vmd_pageout_free_min = 2 * MAXBSIZE / PAGE_SIZE +
2275 vmd->vmd_interrupt_free_min;
2276 vmd->vmd_free_reserved = vm_pageout_page_count +
2277 vmd->vmd_pageout_free_min + vmd->vmd_page_count / 768;
2278 vmd->vmd_free_min = vmd->vmd_page_count / 200;
2279 vmd->vmd_free_severe = vmd->vmd_free_min / 2;
2280 vmd->vmd_free_target = 4 * vmd->vmd_free_min + vmd->vmd_free_reserved;
2281 vmd->vmd_free_min += vmd->vmd_free_reserved;
2282 vmd->vmd_free_severe += vmd->vmd_free_reserved;
2283 vmd->vmd_inactive_target = (3 * vmd->vmd_free_target) / 2;
2284 if (vmd->vmd_inactive_target > vmd->vmd_free_count / 3)
2285 vmd->vmd_inactive_target = vmd->vmd_free_count / 3;
2286
2287 /*
2288 * Set the default wakeup threshold to be 10% below the paging
2289 * target. This keeps the steady state out of shortfall.
2290 */
2291 vmd->vmd_pageout_wakeup_thresh = (vmd->vmd_free_target / 10) * 9;
2292
2293 /*
2294 * Target amount of memory to move out of the laundry queue during a
2295 * background laundering. This is proportional to the amount of system
2296 * memory.
2297 */
2298 vmd->vmd_background_launder_target = (vmd->vmd_free_target -
2299 vmd->vmd_free_min) / 10;
2300
2301 /* Initialize the pageout daemon pid controller. */
2302 pidctrl_init(&vmd->vmd_pid, hz / VM_INACT_SCAN_RATE,
2303 vmd->vmd_free_target, PIDCTRL_BOUND,
2304 PIDCTRL_KPD, PIDCTRL_KID, PIDCTRL_KDD);
2305 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(vmd->vmd_oid), OID_AUTO,
2306 "pidctrl", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2307 pidctrl_init_sysctl(&vmd->vmd_pid, SYSCTL_CHILDREN(oid));
2308
2309 vmd->vmd_inactive_threads = get_pageout_threads_per_domain(vmd);
2310 }
2311
2312 static void
vm_pageout_init(void)2313 vm_pageout_init(void)
2314 {
2315 u_long freecount;
2316 int i;
2317
2318 /*
2319 * Initialize some paging parameters.
2320 */
2321 if (vm_cnt.v_page_count < 2000)
2322 vm_pageout_page_count = 8;
2323
2324 freecount = 0;
2325 for (i = 0; i < vm_ndomains; i++) {
2326 struct vm_domain *vmd;
2327
2328 vm_pageout_init_domain(i);
2329 vmd = VM_DOMAIN(i);
2330 vm_cnt.v_free_reserved += vmd->vmd_free_reserved;
2331 vm_cnt.v_free_target += vmd->vmd_free_target;
2332 vm_cnt.v_free_min += vmd->vmd_free_min;
2333 vm_cnt.v_inactive_target += vmd->vmd_inactive_target;
2334 vm_cnt.v_pageout_free_min += vmd->vmd_pageout_free_min;
2335 vm_cnt.v_interrupt_free_min += vmd->vmd_interrupt_free_min;
2336 vm_cnt.v_free_severe += vmd->vmd_free_severe;
2337 freecount += vmd->vmd_free_count;
2338 }
2339
2340 /*
2341 * Set interval in seconds for active scan. We want to visit each
2342 * page at least once every ten minutes. This is to prevent worst
2343 * case paging behaviors with stale active LRU.
2344 */
2345 if (vm_pageout_update_period == 0)
2346 vm_pageout_update_period = 600;
2347
2348 /*
2349 * Set the maximum number of user-wired virtual pages. Historically the
2350 * main source of such pages was mlock(2) and mlockall(2). Hypervisors
2351 * may also request user-wired memory.
2352 */
2353 if (vm_page_max_user_wired == 0)
2354 vm_page_max_user_wired = 4 * freecount / 5;
2355 }
2356
2357 /*
2358 * vm_pageout is the high level pageout daemon.
2359 */
2360 static void
vm_pageout(void)2361 vm_pageout(void)
2362 {
2363 struct proc *p;
2364 struct thread *td;
2365 int error, first, i, j, pageout_threads;
2366
2367 p = curproc;
2368 td = curthread;
2369
2370 mtx_init(&vm_oom_ratelim_mtx, "vmoomr", NULL, MTX_DEF);
2371 swap_pager_swap_init();
2372 for (first = -1, i = 0; i < vm_ndomains; i++) {
2373 if (VM_DOMAIN_EMPTY(i)) {
2374 if (bootverbose)
2375 printf("domain %d empty; skipping pageout\n",
2376 i);
2377 continue;
2378 }
2379 if (first == -1)
2380 first = i;
2381 else {
2382 error = kthread_add(vm_pageout_worker,
2383 (void *)(uintptr_t)i, p, NULL, 0, 0, "dom%d", i);
2384 if (error != 0)
2385 panic("starting pageout for domain %d: %d\n",
2386 i, error);
2387 }
2388 pageout_threads = VM_DOMAIN(i)->vmd_inactive_threads;
2389 for (j = 0; j < pageout_threads - 1; j++) {
2390 error = kthread_add(vm_pageout_helper,
2391 (void *)(uintptr_t)i, p, NULL, 0, 0,
2392 "dom%d helper%d", i, j);
2393 if (error != 0)
2394 panic("starting pageout helper %d for domain "
2395 "%d: %d\n", j, i, error);
2396 }
2397 error = kthread_add(vm_pageout_laundry_worker,
2398 (void *)(uintptr_t)i, p, NULL, 0, 0, "laundry: dom%d", i);
2399 if (error != 0)
2400 panic("starting laundry for domain %d: %d", i, error);
2401 }
2402 error = kthread_add(uma_reclaim_worker, NULL, p, NULL, 0, 0, "uma");
2403 if (error != 0)
2404 panic("starting uma_reclaim helper, error %d\n", error);
2405
2406 snprintf(td->td_name, sizeof(td->td_name), "dom%d", first);
2407 vm_pageout_worker((void *)(uintptr_t)first);
2408 }
2409
2410 /*
2411 * Perform an advisory wakeup of the page daemon.
2412 */
2413 void
pagedaemon_wakeup(int domain)2414 pagedaemon_wakeup(int domain)
2415 {
2416 struct vm_domain *vmd;
2417
2418 vmd = VM_DOMAIN(domain);
2419 vm_domain_pageout_assert_unlocked(vmd);
2420 if (curproc == pageproc)
2421 return;
2422
2423 if (atomic_fetchadd_int(&vmd->vmd_pageout_wanted, 1) == 0) {
2424 vm_domain_pageout_lock(vmd);
2425 atomic_store_int(&vmd->vmd_pageout_wanted, 1);
2426 wakeup(&vmd->vmd_pageout_wanted);
2427 vm_domain_pageout_unlock(vmd);
2428 }
2429 }
2430