xref: /freebsd/sys/vm/vm_swapout.c (revision ce6a89e27cd190313be39bb479880aeda4778436)
1 /*-
2  * SPDX-License-Identifier: (BSD-4-Clause AND MIT-CMU)
3  *
4  * Copyright (c) 1991 Regents of the University of California.
5  * All rights reserved.
6  * Copyright (c) 1994 John S. Dyson
7  * All rights reserved.
8  * Copyright (c) 1994 David Greenman
9  * All rights reserved.
10  * Copyright (c) 2005 Yahoo! Technologies Norway AS
11  * All rights reserved.
12  *
13  * This code is derived from software contributed to Berkeley by
14  * The Mach Operating System project at Carnegie-Mellon University.
15  *
16  * Redistribution and use in source and binary forms, with or without
17  * modification, are permitted provided that the following conditions
18  * are met:
19  * 1. Redistributions of source code must retain the above copyright
20  *    notice, this list of conditions and the following disclaimer.
21  * 2. Redistributions in binary form must reproduce the above copyright
22  *    notice, this list of conditions and the following disclaimer in the
23  *    documentation and/or other materials provided with the distribution.
24  * 3. All advertising materials mentioning features or use of this software
25  *    must display the following acknowledgement:
26  *	This product includes software developed by the University of
27  *	California, Berkeley and its contributors.
28  * 4. Neither the name of the University nor the names of its contributors
29  *    may be used to endorse or promote products derived from this software
30  *    without specific prior written permission.
31  *
32  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
33  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
34  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
35  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
36  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
37  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
38  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
39  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
40  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
41  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
42  * SUCH DAMAGE.
43  *
44  *	from: @(#)vm_pageout.c	7.4 (Berkeley) 5/7/91
45  *
46  *
47  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
48  * All rights reserved.
49  *
50  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
51  *
52  * Permission to use, copy, modify and distribute this software and
53  * its documentation is hereby granted, provided that both the copyright
54  * notice and this permission notice appear in all copies of the
55  * software, derivative works or modified versions, and any portions
56  * thereof, and that both notices appear in supporting documentation.
57  *
58  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
59  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
60  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
61  *
62  * Carnegie Mellon requests users of this software to return to
63  *
64  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
65  *  School of Computer Science
66  *  Carnegie Mellon University
67  *  Pittsburgh PA 15213-3890
68  *
69  * any improvements or extensions that they make and grant Carnegie the
70  * rights to redistribute these changes.
71  */
72 
73 #include <sys/cdefs.h>
74 __FBSDID("$FreeBSD$");
75 
76 #include "opt_kstack_pages.h"
77 #include "opt_kstack_max_pages.h"
78 #include "opt_vm.h"
79 
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>
85 #include <sys/lock.h>
86 #include <sys/mutex.h>
87 #include <sys/proc.h>
88 #include <sys/kthread.h>
89 #include <sys/ktr.h>
90 #include <sys/mount.h>
91 #include <sys/racct.h>
92 #include <sys/resourcevar.h>
93 #include <sys/refcount.h>
94 #include <sys/sched.h>
95 #include <sys/sdt.h>
96 #include <sys/signalvar.h>
97 #include <sys/smp.h>
98 #include <sys/time.h>
99 #include <sys/vnode.h>
100 #include <sys/vmmeter.h>
101 #include <sys/rwlock.h>
102 #include <sys/sx.h>
103 #include <sys/sysctl.h>
104 
105 #include <vm/vm.h>
106 #include <vm/vm_param.h>
107 #include <vm/vm_object.h>
108 #include <vm/vm_page.h>
109 #include <vm/vm_map.h>
110 #include <vm/vm_pageout.h>
111 #include <vm/vm_pager.h>
112 #include <vm/vm_phys.h>
113 #include <vm/swap_pager.h>
114 #include <vm/vm_extern.h>
115 #include <vm/uma.h>
116 
117 /* the kernel process "vm_daemon" */
118 static void vm_daemon(void);
119 static struct proc *vmproc;
120 
121 static struct kproc_desc vm_kp = {
122 	"vmdaemon",
123 	vm_daemon,
124 	&vmproc
125 };
126 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
127 
128 static int vm_swap_enabled = 1;
129 static int vm_swap_idle_enabled = 0;
130 
131 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, CTLFLAG_RW,
132     &vm_swap_enabled, 0,
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");
137 
138 /*
139  * Swap_idle_threshold1 is the guaranteed swapped in time for a process
140  */
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");
145 
146 /*
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.
149  */
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");
154 
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);
160 
161 static int swapped_cnt;
162 static int swap_inprogress;	/* Pending swap-ins done outside swapper. */
163 static int last_swapin;
164 
165 static void swapclear(struct proc *);
166 static int swapout(struct proc *);
167 static void vm_swapout_map_deactivate_pages(vm_map_t, long);
168 static void vm_swapout_object_deactivate(pmap_t, vm_object_t, long);
169 static void swapout_procs(int action);
170 static void vm_req_vmdaemon(int req);
171 static void vm_thread_swapout(struct thread *td);
172 
173 static void
174 vm_swapout_object_deactivate_page(pmap_t pmap, vm_page_t m, bool unmap)
175 {
176 
177 	/*
178 	 * Ignore unreclaimable wired pages.  Repeat the check after busying
179 	 * since a busy holder may wire the page.
180 	 */
181 	if (vm_page_wired(m) || !vm_page_tryxbusy(m))
182 		return;
183 
184 	if (vm_page_wired(m) || !pmap_page_exists_quick(pmap, m)) {
185 		vm_page_xunbusy(m);
186 		return;
187 	}
188 	if (!pmap_is_referenced(m)) {
189 		if (!vm_page_active(m))
190 			(void)vm_page_try_remove_all(m);
191 		else if (unmap && vm_page_try_remove_all(m))
192 			vm_page_deactivate(m);
193 	}
194 	vm_page_xunbusy(m);
195 }
196 
197 /*
198  *	vm_swapout_object_deactivate
199  *
200  *	Deactivate enough pages to satisfy the inactive target
201  *	requirements.
202  *
203  *	The object and map must be locked.
204  */
205 static void
206 vm_swapout_object_deactivate(pmap_t pmap, vm_object_t first_object,
207     long desired)
208 {
209 	vm_object_t backing_object, object;
210 	vm_page_t m;
211 	bool unmap;
212 
213 	VM_OBJECT_ASSERT_LOCKED(first_object);
214 	if ((first_object->flags & OBJ_FICTITIOUS) != 0)
215 		return;
216 	for (object = first_object;; object = backing_object) {
217 		if (pmap_resident_count(pmap) <= desired)
218 			goto unlock_return;
219 		VM_OBJECT_ASSERT_LOCKED(object);
220 		if ((object->flags & OBJ_UNMANAGED) != 0 ||
221 		    blockcount_read(&object->paging_in_progress) > 0)
222 			goto unlock_return;
223 
224 		unmap = true;
225 		if (object->shadow_count > 1)
226 			unmap = false;
227 
228 		/*
229 		 * Scan the object's entire memory queue.
230 		 */
231 		TAILQ_FOREACH(m, &object->memq, listq) {
232 			if (pmap_resident_count(pmap) <= desired)
233 				goto unlock_return;
234 			if (should_yield())
235 				goto unlock_return;
236 			vm_swapout_object_deactivate_page(pmap, m, unmap);
237 		}
238 		if ((backing_object = object->backing_object) == NULL)
239 			goto unlock_return;
240 		VM_OBJECT_RLOCK(backing_object);
241 		if (object != first_object)
242 			VM_OBJECT_RUNLOCK(object);
243 	}
244 unlock_return:
245 	if (object != first_object)
246 		VM_OBJECT_RUNLOCK(object);
247 }
248 
249 /*
250  * deactivate some number of pages in a map, try to do it fairly, but
251  * that is really hard to do.
252  */
253 static void
254 vm_swapout_map_deactivate_pages(vm_map_t map, long desired)
255 {
256 	vm_map_entry_t tmpe;
257 	vm_object_t obj, bigobj;
258 	int nothingwired;
259 
260 	if (!vm_map_trylock_read(map))
261 		return;
262 
263 	bigobj = NULL;
264 	nothingwired = TRUE;
265 
266 	/*
267 	 * first, search out the biggest object, and try to free pages from
268 	 * that.
269 	 */
270 	VM_MAP_ENTRY_FOREACH(tmpe, map) {
271 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
272 			obj = tmpe->object.vm_object;
273 			if (obj != NULL && VM_OBJECT_TRYRLOCK(obj)) {
274 				if (obj->shadow_count <= 1 &&
275 				    (bigobj == NULL ||
276 				     bigobj->resident_page_count <
277 				     obj->resident_page_count)) {
278 					if (bigobj != NULL)
279 						VM_OBJECT_RUNLOCK(bigobj);
280 					bigobj = obj;
281 				} else
282 					VM_OBJECT_RUNLOCK(obj);
283 			}
284 		}
285 		if (tmpe->wired_count > 0)
286 			nothingwired = FALSE;
287 	}
288 
289 	if (bigobj != NULL) {
290 		vm_swapout_object_deactivate(map->pmap, bigobj, desired);
291 		VM_OBJECT_RUNLOCK(bigobj);
292 	}
293 	/*
294 	 * Next, hunt around for other pages to deactivate.  We actually
295 	 * do this search sort of wrong -- .text first is not the best idea.
296 	 */
297 	VM_MAP_ENTRY_FOREACH(tmpe, map) {
298 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
299 			break;
300 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
301 			obj = tmpe->object.vm_object;
302 			if (obj != NULL) {
303 				VM_OBJECT_RLOCK(obj);
304 				vm_swapout_object_deactivate(map->pmap, obj,
305 				    desired);
306 				VM_OBJECT_RUNLOCK(obj);
307 			}
308 		}
309 	}
310 
311 	/*
312 	 * Remove all mappings if a process is swapped out, this will free page
313 	 * table pages.
314 	 */
315 	if (desired == 0 && nothingwired) {
316 		pmap_remove(vm_map_pmap(map), vm_map_min(map),
317 		    vm_map_max(map));
318 	}
319 
320 	vm_map_unlock_read(map);
321 }
322 
323 /*
324  * Swap out requests
325  */
326 #define VM_SWAP_NORMAL 1
327 #define VM_SWAP_IDLE 2
328 
329 void
330 vm_swapout_run(void)
331 {
332 
333 	if (vm_swap_enabled)
334 		vm_req_vmdaemon(VM_SWAP_NORMAL);
335 }
336 
337 /*
338  * Idle process swapout -- run once per second when pagedaemons are
339  * reclaiming pages.
340  */
341 void
342 vm_swapout_run_idle(void)
343 {
344 	static long lsec;
345 
346 	if (!vm_swap_idle_enabled || time_second == lsec)
347 		return;
348 	vm_req_vmdaemon(VM_SWAP_IDLE);
349 	lsec = time_second;
350 }
351 
352 static void
353 vm_req_vmdaemon(int req)
354 {
355 	static int lastrun = 0;
356 
357 	mtx_lock(&vm_daemon_mtx);
358 	vm_pageout_req_swapout |= req;
359 	if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
360 		wakeup(&vm_daemon_needed);
361 		lastrun = ticks;
362 	}
363 	mtx_unlock(&vm_daemon_mtx);
364 }
365 
366 static void
367 vm_daemon(void)
368 {
369 	struct rlimit rsslim;
370 	struct proc *p;
371 	struct thread *td;
372 	struct vmspace *vm;
373 	int breakout, swapout_flags, tryagain, attempts;
374 #ifdef RACCT
375 	uint64_t rsize, ravailable;
376 #endif
377 
378 	while (TRUE) {
379 		mtx_lock(&vm_daemon_mtx);
380 		msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep",
381 #ifdef RACCT
382 		    racct_enable ? hz : 0
383 #else
384 		    0
385 #endif
386 		);
387 		swapout_flags = vm_pageout_req_swapout;
388 		vm_pageout_req_swapout = 0;
389 		mtx_unlock(&vm_daemon_mtx);
390 		if (swapout_flags != 0) {
391 			/*
392 			 * Drain the per-CPU page queue batches as a deadlock
393 			 * avoidance measure.
394 			 */
395 			if ((swapout_flags & VM_SWAP_NORMAL) != 0)
396 				vm_page_pqbatch_drain();
397 			swapout_procs(swapout_flags);
398 		}
399 
400 		/*
401 		 * scan the processes for exceeding their rlimits or if
402 		 * process is swapped out -- deactivate pages
403 		 */
404 		tryagain = 0;
405 		attempts = 0;
406 again:
407 		attempts++;
408 		sx_slock(&allproc_lock);
409 		FOREACH_PROC_IN_SYSTEM(p) {
410 			vm_pindex_t limit, size;
411 
412 			/*
413 			 * if this is a system process or if we have already
414 			 * looked at this process, skip it.
415 			 */
416 			PROC_LOCK(p);
417 			if (p->p_state != PRS_NORMAL ||
418 			    p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
419 				PROC_UNLOCK(p);
420 				continue;
421 			}
422 			/*
423 			 * if the process is in a non-running type state,
424 			 * don't touch it.
425 			 */
426 			breakout = 0;
427 			FOREACH_THREAD_IN_PROC(p, td) {
428 				thread_lock(td);
429 				if (!TD_ON_RUNQ(td) &&
430 				    !TD_IS_RUNNING(td) &&
431 				    !TD_IS_SLEEPING(td) &&
432 				    !TD_IS_SUSPENDED(td)) {
433 					thread_unlock(td);
434 					breakout = 1;
435 					break;
436 				}
437 				thread_unlock(td);
438 			}
439 			if (breakout) {
440 				PROC_UNLOCK(p);
441 				continue;
442 			}
443 			/*
444 			 * get a limit
445 			 */
446 			lim_rlimit_proc(p, RLIMIT_RSS, &rsslim);
447 			limit = OFF_TO_IDX(
448 			    qmin(rsslim.rlim_cur, rsslim.rlim_max));
449 
450 			/*
451 			 * let processes that are swapped out really be
452 			 * swapped out set the limit to nothing (will force a
453 			 * swap-out.)
454 			 */
455 			if ((p->p_flag & P_INMEM) == 0)
456 				limit = 0;	/* XXX */
457 			vm = vmspace_acquire_ref(p);
458 			_PHOLD_LITE(p);
459 			PROC_UNLOCK(p);
460 			if (vm == NULL) {
461 				PRELE(p);
462 				continue;
463 			}
464 			sx_sunlock(&allproc_lock);
465 
466 			size = vmspace_resident_count(vm);
467 			if (size >= limit) {
468 				vm_swapout_map_deactivate_pages(
469 				    &vm->vm_map, limit);
470 				size = vmspace_resident_count(vm);
471 			}
472 #ifdef RACCT
473 			if (racct_enable) {
474 				rsize = IDX_TO_OFF(size);
475 				PROC_LOCK(p);
476 				if (p->p_state == PRS_NORMAL)
477 					racct_set(p, RACCT_RSS, rsize);
478 				ravailable = racct_get_available(p, RACCT_RSS);
479 				PROC_UNLOCK(p);
480 				if (rsize > ravailable) {
481 					/*
482 					 * Don't be overly aggressive; this
483 					 * might be an innocent process,
484 					 * and the limit could've been exceeded
485 					 * by some memory hog.  Don't try
486 					 * to deactivate more than 1/4th
487 					 * of process' resident set size.
488 					 */
489 					if (attempts <= 8) {
490 						if (ravailable < rsize -
491 						    (rsize / 4)) {
492 							ravailable = rsize -
493 							    (rsize / 4);
494 						}
495 					}
496 					vm_swapout_map_deactivate_pages(
497 					    &vm->vm_map,
498 					    OFF_TO_IDX(ravailable));
499 					/* Update RSS usage after paging out. */
500 					size = vmspace_resident_count(vm);
501 					rsize = IDX_TO_OFF(size);
502 					PROC_LOCK(p);
503 					if (p->p_state == PRS_NORMAL)
504 						racct_set(p, RACCT_RSS, rsize);
505 					PROC_UNLOCK(p);
506 					if (rsize > ravailable)
507 						tryagain = 1;
508 				}
509 			}
510 #endif
511 			vmspace_free(vm);
512 			sx_slock(&allproc_lock);
513 			PRELE(p);
514 		}
515 		sx_sunlock(&allproc_lock);
516 		if (tryagain != 0 && attempts <= 10) {
517 			maybe_yield();
518 			goto again;
519 		}
520 	}
521 }
522 
523 /*
524  * Allow a thread's kernel stack to be paged out.
525  */
526 static void
527 vm_thread_swapout(struct thread *td)
528 {
529 	vm_object_t ksobj;
530 	vm_page_t m;
531 	int i, pages;
532 
533 	cpu_thread_swapout(td);
534 	pages = td->td_kstack_pages;
535 	ksobj = td->td_kstack_obj;
536 	pmap_qremove(td->td_kstack, pages);
537 	VM_OBJECT_WLOCK(ksobj);
538 	for (i = 0; i < pages; i++) {
539 		m = vm_page_lookup(ksobj, i);
540 		if (m == NULL)
541 			panic("vm_thread_swapout: kstack already missing?");
542 		vm_page_dirty(m);
543 		vm_page_xunbusy_unchecked(m);
544 		vm_page_unwire(m, PQ_LAUNDRY);
545 	}
546 	VM_OBJECT_WUNLOCK(ksobj);
547 }
548 
549 /*
550  * Bring the kernel stack for a specified thread back in.
551  */
552 static void
553 vm_thread_swapin(struct thread *td, int oom_alloc)
554 {
555 	vm_object_t ksobj;
556 	vm_page_t ma[KSTACK_MAX_PAGES];
557 	int a, count, i, j, pages, rv;
558 
559 	pages = td->td_kstack_pages;
560 	ksobj = td->td_kstack_obj;
561 	VM_OBJECT_WLOCK(ksobj);
562 	(void)vm_page_grab_pages(ksobj, 0, oom_alloc | VM_ALLOC_WIRED, ma,
563 	    pages);
564 	VM_OBJECT_WUNLOCK(ksobj);
565 	for (i = 0; i < pages;) {
566 		vm_page_assert_xbusied(ma[i]);
567 		if (vm_page_all_valid(ma[i])) {
568 			i++;
569 			continue;
570 		}
571 		vm_object_pip_add(ksobj, 1);
572 		for (j = i + 1; j < pages; j++)
573 			if (vm_page_all_valid(ma[j]))
574 				break;
575 		VM_OBJECT_WLOCK(ksobj);
576 		rv = vm_pager_has_page(ksobj, ma[i]->pindex, NULL, &a);
577 		VM_OBJECT_WUNLOCK(ksobj);
578 		KASSERT(rv == 1, ("%s: missing page %p", __func__, ma[i]));
579 		count = min(a + 1, j - i);
580 		rv = vm_pager_get_pages(ksobj, ma + i, count, NULL, NULL);
581 		KASSERT(rv == VM_PAGER_OK, ("%s: cannot get kstack for proc %d",
582 		    __func__, td->td_proc->p_pid));
583 		vm_object_pip_wakeup(ksobj);
584 		i += count;
585 	}
586 	pmap_qenter(td->td_kstack, ma, pages);
587 	cpu_thread_swapin(td);
588 }
589 
590 void
591 faultin(struct proc *p)
592 {
593 	struct thread *td;
594 	int oom_alloc;
595 
596 	PROC_LOCK_ASSERT(p, MA_OWNED);
597 
598 	/*
599 	 * If another process is swapping in this process,
600 	 * just wait until it finishes.
601 	 */
602 	if (p->p_flag & P_SWAPPINGIN) {
603 		while (p->p_flag & P_SWAPPINGIN)
604 			msleep(&p->p_flag, &p->p_mtx, PVM, "faultin", 0);
605 		return;
606 	}
607 
608 	if ((p->p_flag & P_INMEM) == 0) {
609 		oom_alloc = (p->p_flag & P_WKILLED) != 0 ? VM_ALLOC_SYSTEM :
610 		    VM_ALLOC_NORMAL;
611 
612 		/*
613 		 * Don't let another thread swap process p out while we are
614 		 * busy swapping it in.
615 		 */
616 		++p->p_lock;
617 		p->p_flag |= P_SWAPPINGIN;
618 		PROC_UNLOCK(p);
619 		sx_xlock(&allproc_lock);
620 		MPASS(swapped_cnt > 0);
621 		swapped_cnt--;
622 		if (curthread != &thread0)
623 			swap_inprogress++;
624 		sx_xunlock(&allproc_lock);
625 
626 		/*
627 		 * We hold no lock here because the list of threads
628 		 * can not change while all threads in the process are
629 		 * swapped out.
630 		 */
631 		FOREACH_THREAD_IN_PROC(p, td)
632 			vm_thread_swapin(td, oom_alloc);
633 
634 		if (curthread != &thread0) {
635 			sx_xlock(&allproc_lock);
636 			MPASS(swap_inprogress > 0);
637 			swap_inprogress--;
638 			last_swapin = ticks;
639 			sx_xunlock(&allproc_lock);
640 		}
641 		PROC_LOCK(p);
642 		swapclear(p);
643 		p->p_swtick = ticks;
644 
645 		/* Allow other threads to swap p out now. */
646 		wakeup(&p->p_flag);
647 		--p->p_lock;
648 	}
649 }
650 
651 /*
652  * This swapin algorithm attempts to swap-in processes only if there
653  * is enough space for them.  Of course, if a process waits for a long
654  * time, it will be swapped in anyway.
655  */
656 
657 static struct proc *
658 swapper_selector(bool wkilled_only)
659 {
660 	struct proc *p, *res;
661 	struct thread *td;
662 	int ppri, pri, slptime, swtime;
663 
664 	sx_assert(&allproc_lock, SA_SLOCKED);
665 	if (swapped_cnt == 0)
666 		return (NULL);
667 	res = NULL;
668 	ppri = INT_MIN;
669 	FOREACH_PROC_IN_SYSTEM(p) {
670 		PROC_LOCK(p);
671 		if (p->p_state == PRS_NEW || (p->p_flag & (P_SWAPPINGOUT |
672 		    P_SWAPPINGIN | P_INMEM)) != 0) {
673 			PROC_UNLOCK(p);
674 			continue;
675 		}
676 		if (p->p_state == PRS_NORMAL && (p->p_flag & P_WKILLED) != 0) {
677 			/*
678 			 * A swapped-out process might have mapped a
679 			 * large portion of the system's pages as
680 			 * anonymous memory.  There is no other way to
681 			 * release the memory other than to kill the
682 			 * process, for which we need to swap it in.
683 			 */
684 			return (p);
685 		}
686 		if (wkilled_only) {
687 			PROC_UNLOCK(p);
688 			continue;
689 		}
690 		swtime = (ticks - p->p_swtick) / hz;
691 		FOREACH_THREAD_IN_PROC(p, td) {
692 			/*
693 			 * An otherwise runnable thread of a process
694 			 * swapped out has only the TDI_SWAPPED bit set.
695 			 */
696 			thread_lock(td);
697 			if (td->td_inhibitors == TDI_SWAPPED) {
698 				slptime = (ticks - td->td_slptick) / hz;
699 				pri = swtime + slptime;
700 				if ((td->td_flags & TDF_SWAPINREQ) == 0)
701 					pri -= p->p_nice * 8;
702 				/*
703 				 * if this thread is higher priority
704 				 * and there is enough space, then select
705 				 * this process instead of the previous
706 				 * selection.
707 				 */
708 				if (pri > ppri) {
709 					res = p;
710 					ppri = pri;
711 				}
712 			}
713 			thread_unlock(td);
714 		}
715 		PROC_UNLOCK(p);
716 	}
717 
718 	if (res != NULL)
719 		PROC_LOCK(res);
720 	return (res);
721 }
722 
723 #define	SWAPIN_INTERVAL	(MAXSLP * hz / 2)
724 
725 /*
726  * Limit swapper to swap in one non-WKILLED process in MAXSLP/2
727  * interval, assuming that there is:
728  * - at least one domain that is not suffering from a shortage of free memory;
729  * - no parallel swap-ins;
730  * - no other swap-ins in the current SWAPIN_INTERVAL.
731  */
732 static bool
733 swapper_wkilled_only(void)
734 {
735 
736 	return (vm_page_count_min_set(&all_domains) || swap_inprogress > 0 ||
737 	    (u_int)(ticks - last_swapin) < SWAPIN_INTERVAL);
738 }
739 
740 void
741 swapper(void)
742 {
743 	struct proc *p;
744 
745 	for (;;) {
746 		sx_slock(&allproc_lock);
747 		p = swapper_selector(swapper_wkilled_only());
748 		sx_sunlock(&allproc_lock);
749 
750 		if (p == NULL) {
751 			tsleep(&proc0, PVM, "swapin", SWAPIN_INTERVAL);
752 		} else {
753 			PROC_LOCK_ASSERT(p, MA_OWNED);
754 
755 			/*
756 			 * Another process may be bringing or may have
757 			 * already brought this process in while we
758 			 * traverse all threads.  Or, this process may
759 			 * have exited or even being swapped out
760 			 * again.
761 			 */
762 			if (p->p_state == PRS_NORMAL && (p->p_flag & (P_INMEM |
763 			    P_SWAPPINGOUT | P_SWAPPINGIN)) == 0) {
764 				faultin(p);
765 			}
766 			PROC_UNLOCK(p);
767 		}
768 	}
769 }
770 
771 /*
772  * First, if any processes have been sleeping or stopped for at least
773  * "swap_idle_threshold1" seconds, they are swapped out.  If, however,
774  * no such processes exist, then the longest-sleeping or stopped
775  * process is swapped out.  Finally, and only as a last resort, if
776  * there are no sleeping or stopped processes, the longest-resident
777  * process is swapped out.
778  */
779 static void
780 swapout_procs(int action)
781 {
782 	struct proc *p;
783 	struct thread *td;
784 	int slptime;
785 	bool didswap, doswap;
786 
787 	MPASS((action & (VM_SWAP_NORMAL | VM_SWAP_IDLE)) != 0);
788 
789 	didswap = false;
790 	sx_slock(&allproc_lock);
791 	FOREACH_PROC_IN_SYSTEM(p) {
792 		/*
793 		 * Filter out not yet fully constructed processes.  Do
794 		 * not swap out held processes.  Avoid processes which
795 		 * are system, exiting, execing, traced, already swapped
796 		 * out or are in the process of being swapped in or out.
797 		 */
798 		PROC_LOCK(p);
799 		if (p->p_state != PRS_NORMAL || p->p_lock != 0 || (p->p_flag &
800 		    (P_SYSTEM | P_WEXIT | P_INEXEC | P_STOPPED_SINGLE |
801 		    P_TRACED | P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) !=
802 		    P_INMEM) {
803 			PROC_UNLOCK(p);
804 			continue;
805 		}
806 
807 		/*
808 		 * Further consideration of this process for swap out
809 		 * requires iterating over its threads.  We release
810 		 * allproc_lock here so that process creation and
811 		 * destruction are not blocked while we iterate.
812 		 *
813 		 * To later reacquire allproc_lock and resume
814 		 * iteration over the allproc list, we will first have
815 		 * to release the lock on the process.  We place a
816 		 * hold on the process so that it remains in the
817 		 * allproc list while it is unlocked.
818 		 */
819 		_PHOLD_LITE(p);
820 		sx_sunlock(&allproc_lock);
821 
822 		/*
823 		 * Do not swapout a realtime process.
824 		 * Guarantee swap_idle_threshold1 time in memory.
825 		 * If the system is under memory stress, or if we are
826 		 * swapping idle processes >= swap_idle_threshold2,
827 		 * then swap the process out.
828 		 */
829 		doswap = true;
830 		FOREACH_THREAD_IN_PROC(p, td) {
831 			thread_lock(td);
832 			slptime = (ticks - td->td_slptick) / hz;
833 			if (PRI_IS_REALTIME(td->td_pri_class) ||
834 			    slptime < swap_idle_threshold1 ||
835 			    !thread_safetoswapout(td) ||
836 			    ((action & VM_SWAP_NORMAL) == 0 &&
837 			    slptime < swap_idle_threshold2))
838 				doswap = false;
839 			thread_unlock(td);
840 			if (!doswap)
841 				break;
842 		}
843 		if (doswap && swapout(p) == 0)
844 			didswap = true;
845 
846 		PROC_UNLOCK(p);
847 		if (didswap) {
848 			sx_xlock(&allproc_lock);
849 			swapped_cnt++;
850 			sx_downgrade(&allproc_lock);
851 		} else
852 			sx_slock(&allproc_lock);
853 		PRELE(p);
854 	}
855 	sx_sunlock(&allproc_lock);
856 
857 	/*
858 	 * If we swapped something out, and another process needed memory,
859 	 * then wakeup the sched process.
860 	 */
861 	if (didswap)
862 		wakeup(&proc0);
863 }
864 
865 static void
866 swapclear(struct proc *p)
867 {
868 	struct thread *td;
869 
870 	PROC_LOCK_ASSERT(p, MA_OWNED);
871 
872 	FOREACH_THREAD_IN_PROC(p, td) {
873 		thread_lock(td);
874 		td->td_flags |= TDF_INMEM;
875 		td->td_flags &= ~TDF_SWAPINREQ;
876 		TD_CLR_SWAPPED(td);
877 		if (TD_CAN_RUN(td)) {
878 			if (setrunnable(td, 0)) {
879 #ifdef INVARIANTS
880 				/*
881 				 * XXX: We just cleared TDI_SWAPPED
882 				 * above and set TDF_INMEM, so this
883 				 * should never happen.
884 				 */
885 				panic("not waking up swapper");
886 #endif
887 			}
888 		} else
889 			thread_unlock(td);
890 	}
891 	p->p_flag &= ~(P_SWAPPINGIN | P_SWAPPINGOUT);
892 	p->p_flag |= P_INMEM;
893 }
894 
895 static int
896 swapout(struct proc *p)
897 {
898 	struct thread *td;
899 
900 	PROC_LOCK_ASSERT(p, MA_OWNED);
901 
902 	/*
903 	 * The states of this process and its threads may have changed
904 	 * by now.  Assuming that there is only one pageout daemon thread,
905 	 * this process should still be in memory.
906 	 */
907 	KASSERT((p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) ==
908 	    P_INMEM, ("swapout: lost a swapout race?"));
909 
910 	/*
911 	 * Remember the resident count.
912 	 */
913 	p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace);
914 
915 	/*
916 	 * Check and mark all threads before we proceed.
917 	 */
918 	p->p_flag &= ~P_INMEM;
919 	p->p_flag |= P_SWAPPINGOUT;
920 	FOREACH_THREAD_IN_PROC(p, td) {
921 		thread_lock(td);
922 		if (!thread_safetoswapout(td)) {
923 			thread_unlock(td);
924 			swapclear(p);
925 			return (EBUSY);
926 		}
927 		td->td_flags &= ~TDF_INMEM;
928 		TD_SET_SWAPPED(td);
929 		thread_unlock(td);
930 	}
931 	td = FIRST_THREAD_IN_PROC(p);
932 	++td->td_ru.ru_nswap;
933 	PROC_UNLOCK(p);
934 
935 	/*
936 	 * This list is stable because all threads are now prevented from
937 	 * running.  The list is only modified in the context of a running
938 	 * thread in this process.
939 	 */
940 	FOREACH_THREAD_IN_PROC(p, td)
941 		vm_thread_swapout(td);
942 
943 	PROC_LOCK(p);
944 	p->p_flag &= ~P_SWAPPINGOUT;
945 	p->p_swtick = ticks;
946 	return (0);
947 }
948