2 * SPDX-License-Identifier: (BSD-4-Clause AND MIT-CMU)
4 * Copyright (c) 1991 Regents of the University of California.
6 * Copyright (c) 1994 John S. Dyson
8 * Copyright (c) 1994 David Greenman
10 * Copyright (c) 2005 Yahoo! Technologies Norway AS
11 * All rights reserved.
13 * This code is derived from software contributed to Berkeley by
14 * The Mach Operating System project at Carnegie-Mellon University.
16 * Redistribution and use in source and binary forms, with or without
17 * modification, are permitted provided that the following conditions
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.
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
44 * from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91
47 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
48 * All rights reserved.
50 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
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.
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.
62 * Carnegie Mellon requests users of this software to return to
64 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
65 * School of Computer Science
66 * Carnegie Mellon University
67 * Pittsburgh PA 15213-3890
69 * any improvements or extensions that they make and grant Carnegie the
70 * rights to redistribute these changes.
73 #include <sys/cdefs.h>
74 __FBSDID("$FreeBSD$");
76 #include "opt_kstack_pages.h"
77 #include "opt_kstack_max_pages.h"
80 #include <sys/param.h>
81 #include <sys/systm.h>
82 #include <sys/limits.h>
83 #include <sys/kernel.h>
84 #include <sys/eventhandler.h>
86 #include <sys/mutex.h>
88 #include <sys/_kstack_cache.h>
89 #include <sys/kthread.h>
91 #include <sys/mount.h>
92 #include <sys/racct.h>
93 #include <sys/resourcevar.h>
94 #include <sys/sched.h>
96 #include <sys/signalvar.h>
99 #include <sys/vnode.h>
100 #include <sys/vmmeter.h>
101 #include <sys/rwlock.h>
103 #include <sys/sysctl.h>
106 #include <vm/vm_param.h>
107 #include <vm/vm_object.h>
108 #include <vm/vm_page.h>
109 #include <vm/vm_map.h>
110 #include <vm/vm_pageout.h>
111 #include <vm/vm_pager.h>
112 #include <vm/vm_phys.h>
113 #include <vm/swap_pager.h>
114 #include <vm/vm_extern.h>
117 /* the kernel process "vm_daemon" */
118 static void vm_daemon(void);
119 static struct proc *vmproc;
121 static struct kproc_desc vm_kp = {
126 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
128 static int vm_swap_enabled = 1;
129 static int vm_swap_idle_enabled = 0;
131 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, CTLFLAG_RW,
133 "Enable entire process swapout");
134 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, CTLFLAG_RW,
135 &vm_swap_idle_enabled, 0,
136 "Allow swapout on idle criteria");
139 * Swap_idle_threshold1 is the guaranteed swapped in time for a process
141 static int swap_idle_threshold1 = 2;
142 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1, CTLFLAG_RW,
143 &swap_idle_threshold1, 0,
144 "Guaranteed swapped in time for a process");
147 * Swap_idle_threshold2 is the time that a process can be idle before
148 * it will be swapped out, if idle swapping is enabled.
150 static int swap_idle_threshold2 = 10;
151 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2, CTLFLAG_RW,
152 &swap_idle_threshold2, 0,
153 "Time before a process will be swapped out");
155 static int vm_pageout_req_swapout; /* XXX */
156 static int vm_daemon_needed;
157 static struct mtx vm_daemon_mtx;
158 /* Allow for use by vm_pageout before vm_daemon is initialized. */
159 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
161 static int swapped_cnt;
163 static void swapclear(struct proc *);
164 static int swapout(struct proc *);
165 static void vm_swapout_map_deactivate_pages(vm_map_t, long);
166 static void vm_swapout_object_deactivate_pages(pmap_t, vm_object_t, long);
167 static void swapout_procs(int action);
168 static void vm_req_vmdaemon(int req);
169 static void vm_thread_swapout(struct thread *td);
172 * vm_swapout_object_deactivate_pages
174 * Deactivate enough pages to satisfy the inactive target
177 * The object and map must be locked.
180 vm_swapout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object,
183 vm_object_t backing_object, object;
185 int act_delta, remove_mode;
187 VM_OBJECT_ASSERT_LOCKED(first_object);
188 if ((first_object->flags & OBJ_FICTITIOUS) != 0)
190 for (object = first_object;; object = backing_object) {
191 if (pmap_resident_count(pmap) <= desired)
193 VM_OBJECT_ASSERT_LOCKED(object);
194 if ((object->flags & OBJ_UNMANAGED) != 0 ||
195 object->paging_in_progress != 0)
199 if (object->shadow_count > 1)
202 * Scan the object's entire memory queue.
204 TAILQ_FOREACH(p, &object->memq, listq) {
205 if (pmap_resident_count(pmap) <= desired)
209 if (vm_page_busied(p))
211 VM_CNT_INC(v_pdpages);
213 if (vm_page_held(p) ||
214 !pmap_page_exists_quick(pmap, p)) {
218 act_delta = pmap_ts_referenced(p);
219 if ((p->aflags & PGA_REFERENCED) != 0) {
222 vm_page_aflag_clear(p, PGA_REFERENCED);
224 if (!vm_page_active(p) && act_delta != 0) {
226 p->act_count += act_delta;
227 } else if (vm_page_active(p)) {
228 if (act_delta == 0) {
229 p->act_count -= min(p->act_count,
231 if (!remove_mode && p->act_count == 0) {
233 vm_page_deactivate(p);
238 if (p->act_count < ACT_MAX -
240 p->act_count += ACT_ADVANCE;
243 } else if (vm_page_inactive(p))
247 if ((backing_object = object->backing_object) == NULL)
249 VM_OBJECT_RLOCK(backing_object);
250 if (object != first_object)
251 VM_OBJECT_RUNLOCK(object);
254 if (object != first_object)
255 VM_OBJECT_RUNLOCK(object);
259 * deactivate some number of pages in a map, try to do it fairly, but
260 * that is really hard to do.
263 vm_swapout_map_deactivate_pages(vm_map_t map, long desired)
266 vm_object_t obj, bigobj;
269 if (!vm_map_trylock_read(map))
276 * first, search out the biggest object, and try to free pages from
279 tmpe = map->header.next;
280 while (tmpe != &map->header) {
281 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
282 obj = tmpe->object.vm_object;
283 if (obj != NULL && VM_OBJECT_TRYRLOCK(obj)) {
284 if (obj->shadow_count <= 1 &&
286 bigobj->resident_page_count <
287 obj->resident_page_count)) {
289 VM_OBJECT_RUNLOCK(bigobj);
292 VM_OBJECT_RUNLOCK(obj);
295 if (tmpe->wired_count > 0)
296 nothingwired = FALSE;
300 if (bigobj != NULL) {
301 vm_swapout_object_deactivate_pages(map->pmap, bigobj, desired);
302 VM_OBJECT_RUNLOCK(bigobj);
305 * Next, hunt around for other pages to deactivate. We actually
306 * do this search sort of wrong -- .text first is not the best idea.
308 tmpe = map->header.next;
309 while (tmpe != &map->header) {
310 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
312 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
313 obj = tmpe->object.vm_object;
315 VM_OBJECT_RLOCK(obj);
316 vm_swapout_object_deactivate_pages(map->pmap,
318 VM_OBJECT_RUNLOCK(obj);
325 * Remove all mappings if a process is swapped out, this will free page
328 if (desired == 0 && nothingwired) {
329 pmap_remove(vm_map_pmap(map), vm_map_min(map),
333 vm_map_unlock_read(map);
339 #define VM_SWAP_NORMAL 1
340 #define VM_SWAP_IDLE 2
347 vm_req_vmdaemon(VM_SWAP_NORMAL);
351 * Idle process swapout -- run once per second when pagedaemons are
355 vm_swapout_run_idle(void)
359 if (!vm_swap_idle_enabled || time_second == lsec)
361 vm_req_vmdaemon(VM_SWAP_IDLE);
366 vm_req_vmdaemon(int req)
368 static int lastrun = 0;
370 mtx_lock(&vm_daemon_mtx);
371 vm_pageout_req_swapout |= req;
372 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
373 wakeup(&vm_daemon_needed);
376 mtx_unlock(&vm_daemon_mtx);
382 struct rlimit rsslim;
386 int breakout, swapout_flags, tryagain, attempts;
388 uint64_t rsize, ravailable;
392 mtx_lock(&vm_daemon_mtx);
393 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep",
395 racct_enable ? hz : 0
400 swapout_flags = vm_pageout_req_swapout;
401 vm_pageout_req_swapout = 0;
402 mtx_unlock(&vm_daemon_mtx);
403 if (swapout_flags != 0) {
405 * Drain the per-CPU page queue batches as a deadlock
408 if ((swapout_flags & VM_SWAP_NORMAL) != 0)
409 vm_page_drain_pqbatch();
410 swapout_procs(swapout_flags);
414 * scan the processes for exceeding their rlimits or if
415 * process is swapped out -- deactivate pages
421 sx_slock(&allproc_lock);
422 FOREACH_PROC_IN_SYSTEM(p) {
423 vm_pindex_t limit, size;
426 * if this is a system process or if we have already
427 * looked at this process, skip it.
430 if (p->p_state != PRS_NORMAL ||
431 p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
436 * if the process is in a non-running type state,
440 FOREACH_THREAD_IN_PROC(p, td) {
442 if (!TD_ON_RUNQ(td) &&
443 !TD_IS_RUNNING(td) &&
444 !TD_IS_SLEEPING(td) &&
445 !TD_IS_SUSPENDED(td)) {
459 lim_rlimit_proc(p, RLIMIT_RSS, &rsslim);
461 qmin(rsslim.rlim_cur, rsslim.rlim_max));
464 * let processes that are swapped out really be
465 * swapped out set the limit to nothing (will force a
468 if ((p->p_flag & P_INMEM) == 0)
470 vm = vmspace_acquire_ref(p);
477 sx_sunlock(&allproc_lock);
479 size = vmspace_resident_count(vm);
481 vm_swapout_map_deactivate_pages(
483 size = vmspace_resident_count(vm);
487 rsize = IDX_TO_OFF(size);
489 if (p->p_state == PRS_NORMAL)
490 racct_set(p, RACCT_RSS, rsize);
491 ravailable = racct_get_available(p, RACCT_RSS);
493 if (rsize > ravailable) {
495 * Don't be overly aggressive; this
496 * might be an innocent process,
497 * and the limit could've been exceeded
498 * by some memory hog. Don't try
499 * to deactivate more than 1/4th
500 * of process' resident set size.
503 if (ravailable < rsize -
509 vm_swapout_map_deactivate_pages(
511 OFF_TO_IDX(ravailable));
512 /* Update RSS usage after paging out. */
513 size = vmspace_resident_count(vm);
514 rsize = IDX_TO_OFF(size);
516 if (p->p_state == PRS_NORMAL)
517 racct_set(p, RACCT_RSS, rsize);
519 if (rsize > ravailable)
525 sx_slock(&allproc_lock);
528 sx_sunlock(&allproc_lock);
529 if (tryagain != 0 && attempts <= 10) {
537 * Allow a thread's kernel stack to be paged out.
540 vm_thread_swapout(struct thread *td)
546 cpu_thread_swapout(td);
547 pages = td->td_kstack_pages;
548 ksobj = td->td_kstack_obj;
549 pmap_qremove(td->td_kstack, pages);
550 VM_OBJECT_WLOCK(ksobj);
551 for (i = 0; i < pages; i++) {
552 m = vm_page_lookup(ksobj, i);
554 panic("vm_thread_swapout: kstack already missing?");
557 vm_page_unwire(m, PQ_LAUNDRY);
560 VM_OBJECT_WUNLOCK(ksobj);
564 * Bring the kernel stack for a specified thread back in.
567 vm_thread_swapin(struct thread *td, int oom_alloc)
570 vm_page_t ma[KSTACK_MAX_PAGES];
571 int a, count, i, j, pages, rv;
573 pages = td->td_kstack_pages;
574 ksobj = td->td_kstack_obj;
575 VM_OBJECT_WLOCK(ksobj);
576 (void)vm_page_grab_pages(ksobj, 0, oom_alloc | VM_ALLOC_WIRED, ma,
578 for (i = 0; i < pages;) {
579 vm_page_assert_xbusied(ma[i]);
580 if (ma[i]->valid == VM_PAGE_BITS_ALL) {
581 vm_page_xunbusy(ma[i]);
585 vm_object_pip_add(ksobj, 1);
586 for (j = i + 1; j < pages; j++)
587 if (ma[j]->valid == VM_PAGE_BITS_ALL)
589 rv = vm_pager_has_page(ksobj, ma[i]->pindex, NULL, &a);
590 KASSERT(rv == 1, ("%s: missing page %p", __func__, ma[i]));
591 count = min(a + 1, j - i);
592 rv = vm_pager_get_pages(ksobj, ma + i, count, NULL, NULL);
593 KASSERT(rv == VM_PAGER_OK, ("%s: cannot get kstack for proc %d",
594 __func__, td->td_proc->p_pid));
595 vm_object_pip_wakeup(ksobj);
596 for (j = i; j < i + count; j++)
597 vm_page_xunbusy(ma[j]);
600 VM_OBJECT_WUNLOCK(ksobj);
601 pmap_qenter(td->td_kstack, ma, pages);
602 cpu_thread_swapin(td);
606 faultin(struct proc *p)
611 PROC_LOCK_ASSERT(p, MA_OWNED);
614 * If another process is swapping in this process,
615 * just wait until it finishes.
617 if (p->p_flag & P_SWAPPINGIN) {
618 while (p->p_flag & P_SWAPPINGIN)
619 msleep(&p->p_flag, &p->p_mtx, PVM, "faultin", 0);
623 if ((p->p_flag & P_INMEM) == 0) {
624 oom_alloc = (p->p_flag & P_WKILLED) != 0 ? VM_ALLOC_SYSTEM :
628 * Don't let another thread swap process p out while we are
629 * busy swapping it in.
632 p->p_flag |= P_SWAPPINGIN;
634 sx_xlock(&allproc_lock);
635 MPASS(swapped_cnt > 0);
637 sx_xunlock(&allproc_lock);
640 * We hold no lock here because the list of threads
641 * can not change while all threads in the process are
644 FOREACH_THREAD_IN_PROC(p, td)
645 vm_thread_swapin(td, oom_alloc);
651 /* Allow other threads to swap p out now. */
658 * This swapin algorithm attempts to swap-in processes only if there
659 * is enough space for them. Of course, if a process waits for a long
660 * time, it will be swapped in anyway.
664 swapper_selector(void)
666 struct proc *p, *res;
668 int min_flag, ppri, pri, slptime, swtime;
670 sx_assert(&allproc_lock, SA_SLOCKED);
671 if (swapped_cnt == 0)
675 min_flag = vm_page_count_min();
676 FOREACH_PROC_IN_SYSTEM(p) {
678 if (p->p_state == PRS_NEW || (p->p_flag & (P_SWAPPINGOUT |
679 P_SWAPPINGIN | P_INMEM)) != 0) {
683 if (p->p_state == PRS_NORMAL && (p->p_flag & P_WKILLED) != 0) {
685 * A swapped-out process might have mapped a
686 * large portion of the system's pages as
687 * anonymous memory. There is no other way to
688 * release the memory other than to kill the
689 * process, for which we need to swap it in.
697 swtime = (ticks - p->p_swtick) / hz;
698 FOREACH_THREAD_IN_PROC(p, td) {
700 * An otherwise runnable thread of a process
701 * swapped out has only the TDI_SWAPPED bit set.
704 if (td->td_inhibitors == TDI_SWAPPED) {
705 slptime = (ticks - td->td_slptick) / hz;
706 pri = swtime + slptime;
707 if ((td->td_flags & TDF_SWAPINREQ) == 0)
708 pri -= p->p_nice * 8;
710 * if this thread is higher priority
711 * and there is enough space, then select
712 * this process instead of the previous
735 sx_slock(&allproc_lock);
736 p = swapper_selector();
737 sx_sunlock(&allproc_lock);
740 tsleep(&proc0, PVM, "swapin", MAXSLP * hz / 2);
742 PROC_LOCK_ASSERT(p, MA_OWNED);
745 * Another process may be bringing or may have
746 * already brought this process in while we
747 * traverse all threads. Or, this process may
748 * have exited or even being swapped out
751 if (p->p_state == PRS_NORMAL && (p->p_flag & (P_INMEM |
752 P_SWAPPINGOUT | P_SWAPPINGIN)) == 0) {
761 * First, if any processes have been sleeping or stopped for at least
762 * "swap_idle_threshold1" seconds, they are swapped out. If, however,
763 * no such processes exist, then the longest-sleeping or stopped
764 * process is swapped out. Finally, and only as a last resort, if
765 * there are no sleeping or stopped processes, the longest-resident
766 * process is swapped out.
769 swapout_procs(int action)
774 bool didswap, doswap;
776 MPASS((action & (VM_SWAP_NORMAL | VM_SWAP_IDLE)) != 0);
779 sx_slock(&allproc_lock);
780 FOREACH_PROC_IN_SYSTEM(p) {
782 * Filter out not yet fully constructed processes. Do
783 * not swap out held processes. Avoid processes which
784 * are system, exiting, execing, traced, already swapped
785 * out or are in the process of being swapped in or out.
788 if (p->p_state != PRS_NORMAL || p->p_lock != 0 || (p->p_flag &
789 (P_SYSTEM | P_WEXIT | P_INEXEC | P_STOPPED_SINGLE |
790 P_TRACED | P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) !=
797 * Further consideration of this process for swap out
798 * requires iterating over its threads. We release
799 * allproc_lock here so that process creation and
800 * destruction are not blocked while we iterate.
802 * To later reacquire allproc_lock and resume
803 * iteration over the allproc list, we will first have
804 * to release the lock on the process. We place a
805 * hold on the process so that it remains in the
806 * allproc list while it is unlocked.
809 sx_sunlock(&allproc_lock);
812 * Do not swapout a realtime process.
813 * Guarantee swap_idle_threshold1 time in memory.
814 * If the system is under memory stress, or if we are
815 * swapping idle processes >= swap_idle_threshold2,
816 * then swap the process out.
819 FOREACH_THREAD_IN_PROC(p, td) {
821 slptime = (ticks - td->td_slptick) / hz;
822 if (PRI_IS_REALTIME(td->td_pri_class) ||
823 slptime < swap_idle_threshold1 ||
824 !thread_safetoswapout(td) ||
825 ((action & VM_SWAP_NORMAL) == 0 &&
826 slptime < swap_idle_threshold2))
832 if (doswap && swapout(p) == 0)
837 sx_xlock(&allproc_lock);
839 sx_downgrade(&allproc_lock);
841 sx_slock(&allproc_lock);
844 sx_sunlock(&allproc_lock);
847 * If we swapped something out, and another process needed memory,
848 * then wakeup the sched process.
855 swapclear(struct proc *p)
859 PROC_LOCK_ASSERT(p, MA_OWNED);
861 FOREACH_THREAD_IN_PROC(p, td) {
863 td->td_flags |= TDF_INMEM;
864 td->td_flags &= ~TDF_SWAPINREQ;
867 if (setrunnable(td)) {
870 * XXX: We just cleared TDI_SWAPPED
871 * above and set TDF_INMEM, so this
872 * should never happen.
874 panic("not waking up swapper");
879 p->p_flag &= ~(P_SWAPPINGIN | P_SWAPPINGOUT);
880 p->p_flag |= P_INMEM;
884 swapout(struct proc *p)
888 PROC_LOCK_ASSERT(p, MA_OWNED);
891 * The states of this process and its threads may have changed
892 * by now. Assuming that there is only one pageout daemon thread,
893 * this process should still be in memory.
895 KASSERT((p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) ==
896 P_INMEM, ("swapout: lost a swapout race?"));
899 * Remember the resident count.
901 p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace);
904 * Check and mark all threads before we proceed.
906 p->p_flag &= ~P_INMEM;
907 p->p_flag |= P_SWAPPINGOUT;
908 FOREACH_THREAD_IN_PROC(p, td) {
910 if (!thread_safetoswapout(td)) {
915 td->td_flags &= ~TDF_INMEM;
919 td = FIRST_THREAD_IN_PROC(p);
920 ++td->td_ru.ru_nswap;
924 * This list is stable because all threads are now prevented from
925 * running. The list is only modified in the context of a running
926 * thread in this process.
928 FOREACH_THREAD_IN_PROC(p, td)
929 vm_thread_swapout(td);
932 p->p_flag &= ~P_SWAPPINGOUT;