xref: /freebsd/sys/vm/vm_swapout.c (revision af23369a6deaaeb612ab266eb88b8bb8d560c322)
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_kern.h>
108 #include <vm/vm_object.h>
109 #include <vm/vm_page.h>
110 #include <vm/vm_map.h>
111 #include <vm/vm_pageout.h>
112 #include <vm/vm_pager.h>
113 #include <vm/vm_phys.h>
114 #include <vm/swap_pager.h>
115 #include <vm/vm_extern.h>
116 #include <vm/uma.h>
117 
118 /* the kernel process "vm_daemon" */
119 static void vm_daemon(void);
120 static struct proc *vmproc;
121 
122 static struct kproc_desc vm_kp = {
123 	"vmdaemon",
124 	vm_daemon,
125 	&vmproc
126 };
127 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
128 
129 static int vm_swap_enabled = 1;
130 static int vm_swap_idle_enabled = 0;
131 
132 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, CTLFLAG_RW,
133     &vm_swap_enabled, 0,
134     "Enable entire process swapout");
135 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, CTLFLAG_RW,
136     &vm_swap_idle_enabled, 0,
137     "Allow swapout on idle criteria");
138 
139 /*
140  * Swap_idle_threshold1 is the guaranteed swapped in time for a process
141  */
142 static int swap_idle_threshold1 = 2;
143 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1, CTLFLAG_RW,
144     &swap_idle_threshold1, 0,
145     "Guaranteed swapped in time for a process");
146 
147 /*
148  * Swap_idle_threshold2 is the time that a process can be idle before
149  * it will be swapped out, if idle swapping is enabled.
150  */
151 static int swap_idle_threshold2 = 10;
152 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2, CTLFLAG_RW,
153     &swap_idle_threshold2, 0,
154     "Time before a process will be swapped out");
155 
156 static int vm_daemon_timeout = 0;
157 SYSCTL_INT(_vm, OID_AUTO, vmdaemon_timeout, CTLFLAG_RW,
158     &vm_daemon_timeout, 0,
159     "Time between vmdaemon runs");
160 
161 static int vm_pageout_req_swapout;	/* XXX */
162 static int vm_daemon_needed;
163 static struct mtx vm_daemon_mtx;
164 /* Allow for use by vm_pageout before vm_daemon is initialized. */
165 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
166 
167 static int swapped_cnt;
168 static int swap_inprogress;	/* Pending swap-ins done outside swapper. */
169 static int last_swapin;
170 
171 static void swapclear(struct proc *);
172 static int swapout(struct proc *);
173 static void vm_swapout_map_deactivate_pages(vm_map_t, long);
174 static void vm_swapout_object_deactivate(pmap_t, vm_object_t, long);
175 static void swapout_procs(int action);
176 static void vm_req_vmdaemon(int req);
177 static void vm_thread_swapout(struct thread *td);
178 
179 static void
180 vm_swapout_object_deactivate_page(pmap_t pmap, vm_page_t m, bool unmap)
181 {
182 
183 	/*
184 	 * Ignore unreclaimable wired pages.  Repeat the check after busying
185 	 * since a busy holder may wire the page.
186 	 */
187 	if (vm_page_wired(m) || !vm_page_tryxbusy(m))
188 		return;
189 
190 	if (vm_page_wired(m) || !pmap_page_exists_quick(pmap, m)) {
191 		vm_page_xunbusy(m);
192 		return;
193 	}
194 	if (!pmap_is_referenced(m)) {
195 		if (!vm_page_active(m))
196 			(void)vm_page_try_remove_all(m);
197 		else if (unmap && vm_page_try_remove_all(m))
198 			vm_page_deactivate(m);
199 	}
200 	vm_page_xunbusy(m);
201 }
202 
203 /*
204  *	vm_swapout_object_deactivate
205  *
206  *	Deactivate enough pages to satisfy the inactive target
207  *	requirements.
208  *
209  *	The object and map must be locked.
210  */
211 static void
212 vm_swapout_object_deactivate(pmap_t pmap, vm_object_t first_object,
213     long desired)
214 {
215 	vm_object_t backing_object, object;
216 	vm_page_t m;
217 	bool unmap;
218 
219 	VM_OBJECT_ASSERT_LOCKED(first_object);
220 	if ((first_object->flags & OBJ_FICTITIOUS) != 0)
221 		return;
222 	for (object = first_object;; object = backing_object) {
223 		if (pmap_resident_count(pmap) <= desired)
224 			goto unlock_return;
225 		VM_OBJECT_ASSERT_LOCKED(object);
226 		if ((object->flags & OBJ_UNMANAGED) != 0 ||
227 		    blockcount_read(&object->paging_in_progress) > 0)
228 			goto unlock_return;
229 
230 		unmap = true;
231 		if (object->shadow_count > 1)
232 			unmap = false;
233 
234 		/*
235 		 * Scan the object's entire memory queue.
236 		 */
237 		TAILQ_FOREACH(m, &object->memq, listq) {
238 			if (pmap_resident_count(pmap) <= desired)
239 				goto unlock_return;
240 			if (should_yield())
241 				goto unlock_return;
242 			vm_swapout_object_deactivate_page(pmap, m, unmap);
243 		}
244 		if ((backing_object = object->backing_object) == NULL)
245 			goto unlock_return;
246 		VM_OBJECT_RLOCK(backing_object);
247 		if (object != first_object)
248 			VM_OBJECT_RUNLOCK(object);
249 	}
250 unlock_return:
251 	if (object != first_object)
252 		VM_OBJECT_RUNLOCK(object);
253 }
254 
255 /*
256  * deactivate some number of pages in a map, try to do it fairly, but
257  * that is really hard to do.
258  */
259 static void
260 vm_swapout_map_deactivate_pages(vm_map_t map, long desired)
261 {
262 	vm_map_entry_t tmpe;
263 	vm_object_t obj, bigobj;
264 	int nothingwired;
265 
266 	if (!vm_map_trylock_read(map))
267 		return;
268 
269 	bigobj = NULL;
270 	nothingwired = TRUE;
271 
272 	/*
273 	 * first, search out the biggest object, and try to free pages from
274 	 * that.
275 	 */
276 	VM_MAP_ENTRY_FOREACH(tmpe, map) {
277 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
278 			obj = tmpe->object.vm_object;
279 			if (obj != NULL && VM_OBJECT_TRYRLOCK(obj)) {
280 				if (obj->shadow_count <= 1 &&
281 				    (bigobj == NULL ||
282 				     bigobj->resident_page_count <
283 				     obj->resident_page_count)) {
284 					if (bigobj != NULL)
285 						VM_OBJECT_RUNLOCK(bigobj);
286 					bigobj = obj;
287 				} else
288 					VM_OBJECT_RUNLOCK(obj);
289 			}
290 		}
291 		if (tmpe->wired_count > 0)
292 			nothingwired = FALSE;
293 	}
294 
295 	if (bigobj != NULL) {
296 		vm_swapout_object_deactivate(map->pmap, bigobj, desired);
297 		VM_OBJECT_RUNLOCK(bigobj);
298 	}
299 	/*
300 	 * Next, hunt around for other pages to deactivate.  We actually
301 	 * do this search sort of wrong -- .text first is not the best idea.
302 	 */
303 	VM_MAP_ENTRY_FOREACH(tmpe, map) {
304 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
305 			break;
306 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
307 			obj = tmpe->object.vm_object;
308 			if (obj != NULL) {
309 				VM_OBJECT_RLOCK(obj);
310 				vm_swapout_object_deactivate(map->pmap, obj,
311 				    desired);
312 				VM_OBJECT_RUNLOCK(obj);
313 			}
314 		}
315 	}
316 
317 	/*
318 	 * Remove all mappings if a process is swapped out, this will free page
319 	 * table pages.
320 	 */
321 	if (desired == 0 && nothingwired) {
322 		pmap_remove(vm_map_pmap(map), vm_map_min(map),
323 		    vm_map_max(map));
324 	}
325 
326 	vm_map_unlock_read(map);
327 }
328 
329 /*
330  * Swap out requests
331  */
332 #define VM_SWAP_NORMAL 1
333 #define VM_SWAP_IDLE 2
334 
335 void
336 vm_swapout_run(void)
337 {
338 
339 	if (vm_swap_enabled)
340 		vm_req_vmdaemon(VM_SWAP_NORMAL);
341 }
342 
343 /*
344  * Idle process swapout -- run once per second when pagedaemons are
345  * reclaiming pages.
346  */
347 void
348 vm_swapout_run_idle(void)
349 {
350 	static long lsec;
351 
352 	if (!vm_swap_idle_enabled || time_second == lsec)
353 		return;
354 	vm_req_vmdaemon(VM_SWAP_IDLE);
355 	lsec = time_second;
356 }
357 
358 static void
359 vm_req_vmdaemon(int req)
360 {
361 	static int lastrun = 0;
362 
363 	mtx_lock(&vm_daemon_mtx);
364 	vm_pageout_req_swapout |= req;
365 	if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
366 		wakeup(&vm_daemon_needed);
367 		lastrun = ticks;
368 	}
369 	mtx_unlock(&vm_daemon_mtx);
370 }
371 
372 static void
373 vm_daemon(void)
374 {
375 	struct rlimit rsslim;
376 	struct proc *p;
377 	struct thread *td;
378 	struct vmspace *vm;
379 	int breakout, swapout_flags, tryagain, attempts;
380 #ifdef RACCT
381 	uint64_t rsize, ravailable;
382 
383 	if (racct_enable && vm_daemon_timeout == 0)
384 		vm_daemon_timeout = hz;
385 #endif
386 
387 	while (TRUE) {
388 		mtx_lock(&vm_daemon_mtx);
389 		msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep",
390 		    vm_daemon_timeout);
391 		swapout_flags = vm_pageout_req_swapout;
392 		vm_pageout_req_swapout = 0;
393 		mtx_unlock(&vm_daemon_mtx);
394 		if (swapout_flags != 0) {
395 			/*
396 			 * Drain the per-CPU page queue batches as a deadlock
397 			 * avoidance measure.
398 			 */
399 			if ((swapout_flags & VM_SWAP_NORMAL) != 0)
400 				vm_page_pqbatch_drain();
401 			swapout_procs(swapout_flags);
402 		}
403 
404 		/*
405 		 * scan the processes for exceeding their rlimits or if
406 		 * process is swapped out -- deactivate pages
407 		 */
408 		tryagain = 0;
409 		attempts = 0;
410 again:
411 		attempts++;
412 		sx_slock(&allproc_lock);
413 		FOREACH_PROC_IN_SYSTEM(p) {
414 			vm_pindex_t limit, size;
415 
416 			/*
417 			 * if this is a system process or if we have already
418 			 * looked at this process, skip it.
419 			 */
420 			PROC_LOCK(p);
421 			if (p->p_state != PRS_NORMAL ||
422 			    p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
423 				PROC_UNLOCK(p);
424 				continue;
425 			}
426 			/*
427 			 * if the process is in a non-running type state,
428 			 * don't touch it.
429 			 */
430 			breakout = 0;
431 			FOREACH_THREAD_IN_PROC(p, td) {
432 				thread_lock(td);
433 				if (!TD_ON_RUNQ(td) &&
434 				    !TD_IS_RUNNING(td) &&
435 				    !TD_IS_SLEEPING(td) &&
436 				    !TD_IS_SUSPENDED(td)) {
437 					thread_unlock(td);
438 					breakout = 1;
439 					break;
440 				}
441 				thread_unlock(td);
442 			}
443 			if (breakout) {
444 				PROC_UNLOCK(p);
445 				continue;
446 			}
447 			/*
448 			 * get a limit
449 			 */
450 			lim_rlimit_proc(p, RLIMIT_RSS, &rsslim);
451 			limit = OFF_TO_IDX(
452 			    qmin(rsslim.rlim_cur, rsslim.rlim_max));
453 
454 			/*
455 			 * let processes that are swapped out really be
456 			 * swapped out set the limit to nothing (will force a
457 			 * swap-out.)
458 			 */
459 			if ((p->p_flag & P_INMEM) == 0)
460 				limit = 0;	/* XXX */
461 			vm = vmspace_acquire_ref(p);
462 			_PHOLD_LITE(p);
463 			PROC_UNLOCK(p);
464 			if (vm == NULL) {
465 				PRELE(p);
466 				continue;
467 			}
468 			sx_sunlock(&allproc_lock);
469 
470 			size = vmspace_resident_count(vm);
471 			if (size >= limit) {
472 				vm_swapout_map_deactivate_pages(
473 				    &vm->vm_map, limit);
474 				size = vmspace_resident_count(vm);
475 			}
476 #ifdef RACCT
477 			if (racct_enable) {
478 				rsize = IDX_TO_OFF(size);
479 				PROC_LOCK(p);
480 				if (p->p_state == PRS_NORMAL)
481 					racct_set(p, RACCT_RSS, rsize);
482 				ravailable = racct_get_available(p, RACCT_RSS);
483 				PROC_UNLOCK(p);
484 				if (rsize > ravailable) {
485 					/*
486 					 * Don't be overly aggressive; this
487 					 * might be an innocent process,
488 					 * and the limit could've been exceeded
489 					 * by some memory hog.  Don't try
490 					 * to deactivate more than 1/4th
491 					 * of process' resident set size.
492 					 */
493 					if (attempts <= 8) {
494 						if (ravailable < rsize -
495 						    (rsize / 4)) {
496 							ravailable = rsize -
497 							    (rsize / 4);
498 						}
499 					}
500 					vm_swapout_map_deactivate_pages(
501 					    &vm->vm_map,
502 					    OFF_TO_IDX(ravailable));
503 					/* Update RSS usage after paging out. */
504 					size = vmspace_resident_count(vm);
505 					rsize = IDX_TO_OFF(size);
506 					PROC_LOCK(p);
507 					if (p->p_state == PRS_NORMAL)
508 						racct_set(p, RACCT_RSS, rsize);
509 					PROC_UNLOCK(p);
510 					if (rsize > ravailable)
511 						tryagain = 1;
512 				}
513 			}
514 #endif
515 			vmspace_free(vm);
516 			sx_slock(&allproc_lock);
517 			PRELE(p);
518 		}
519 		sx_sunlock(&allproc_lock);
520 		if (tryagain != 0 && attempts <= 10) {
521 			maybe_yield();
522 			goto again;
523 		}
524 	}
525 }
526 
527 /*
528  * Allow a thread's kernel stack to be paged out.
529  */
530 static void
531 vm_thread_swapout(struct thread *td)
532 {
533 	vm_page_t m;
534 	vm_offset_t kaddr;
535 	vm_pindex_t pindex;
536 	int i, pages;
537 
538 	cpu_thread_swapout(td);
539 	kaddr = td->td_kstack;
540 	pages = td->td_kstack_pages;
541 	pindex = atop(kaddr - VM_MIN_KERNEL_ADDRESS);
542 	pmap_qremove(kaddr, pages);
543 	VM_OBJECT_WLOCK(kstack_object);
544 	for (i = 0; i < pages; i++) {
545 		m = vm_page_lookup(kstack_object, pindex + i);
546 		if (m == NULL)
547 			panic("vm_thread_swapout: kstack already missing?");
548 		vm_page_dirty(m);
549 		vm_page_xunbusy_unchecked(m);
550 		vm_page_unwire(m, PQ_LAUNDRY);
551 	}
552 	VM_OBJECT_WUNLOCK(kstack_object);
553 }
554 
555 /*
556  * Bring the kernel stack for a specified thread back in.
557  */
558 static void
559 vm_thread_swapin(struct thread *td, int oom_alloc)
560 {
561 	vm_page_t ma[KSTACK_MAX_PAGES];
562 	vm_offset_t kaddr;
563 	int a, count, i, j, pages, rv __diagused;
564 
565 	kaddr = td->td_kstack;
566 	pages = td->td_kstack_pages;
567 	vm_thread_stack_back(td->td_domain.dr_policy, kaddr, ma, pages,
568 	    oom_alloc);
569 	for (i = 0; i < pages;) {
570 		vm_page_assert_xbusied(ma[i]);
571 		if (vm_page_all_valid(ma[i])) {
572 			i++;
573 			continue;
574 		}
575 		vm_object_pip_add(kstack_object, 1);
576 		for (j = i + 1; j < pages; j++)
577 			if (vm_page_all_valid(ma[j]))
578 				break;
579 		VM_OBJECT_WLOCK(kstack_object);
580 		rv = vm_pager_has_page(kstack_object, ma[i]->pindex, NULL, &a);
581 		VM_OBJECT_WUNLOCK(kstack_object);
582 		KASSERT(rv == 1, ("%s: missing page %p", __func__, ma[i]));
583 		count = min(a + 1, j - i);
584 		rv = vm_pager_get_pages(kstack_object, ma + i, count, NULL, NULL);
585 		KASSERT(rv == VM_PAGER_OK, ("%s: cannot get kstack for proc %d",
586 		    __func__, td->td_proc->p_pid));
587 		vm_object_pip_wakeup(kstack_object);
588 		i += count;
589 	}
590 	pmap_qenter(kaddr, ma, pages);
591 	cpu_thread_swapin(td);
592 }
593 
594 void
595 faultin(struct proc *p)
596 {
597 	struct thread *td;
598 	int oom_alloc;
599 
600 	PROC_LOCK_ASSERT(p, MA_OWNED);
601 
602 	/*
603 	 * If another process is swapping in this process,
604 	 * just wait until it finishes.
605 	 */
606 	if (p->p_flag & P_SWAPPINGIN) {
607 		while (p->p_flag & P_SWAPPINGIN)
608 			msleep(&p->p_flag, &p->p_mtx, PVM, "faultin", 0);
609 		return;
610 	}
611 
612 	if ((p->p_flag & P_INMEM) == 0) {
613 		oom_alloc = (p->p_flag & P_WKILLED) != 0 ? VM_ALLOC_SYSTEM :
614 		    VM_ALLOC_NORMAL;
615 
616 		/*
617 		 * Don't let another thread swap process p out while we are
618 		 * busy swapping it in.
619 		 */
620 		++p->p_lock;
621 		p->p_flag |= P_SWAPPINGIN;
622 		PROC_UNLOCK(p);
623 		sx_xlock(&allproc_lock);
624 		MPASS(swapped_cnt > 0);
625 		swapped_cnt--;
626 		if (curthread != &thread0)
627 			swap_inprogress++;
628 		sx_xunlock(&allproc_lock);
629 
630 		/*
631 		 * We hold no lock here because the list of threads
632 		 * can not change while all threads in the process are
633 		 * swapped out.
634 		 */
635 		FOREACH_THREAD_IN_PROC(p, td)
636 			vm_thread_swapin(td, oom_alloc);
637 
638 		if (curthread != &thread0) {
639 			sx_xlock(&allproc_lock);
640 			MPASS(swap_inprogress > 0);
641 			swap_inprogress--;
642 			last_swapin = ticks;
643 			sx_xunlock(&allproc_lock);
644 		}
645 		PROC_LOCK(p);
646 		swapclear(p);
647 		p->p_swtick = ticks;
648 
649 		/* Allow other threads to swap p out now. */
650 		wakeup(&p->p_flag);
651 		--p->p_lock;
652 	}
653 }
654 
655 /*
656  * This swapin algorithm attempts to swap-in processes only if there
657  * is enough space for them.  Of course, if a process waits for a long
658  * time, it will be swapped in anyway.
659  */
660 
661 static struct proc *
662 swapper_selector(bool wkilled_only)
663 {
664 	struct proc *p, *res;
665 	struct thread *td;
666 	int ppri, pri, slptime, swtime;
667 
668 	sx_assert(&allproc_lock, SA_SLOCKED);
669 	if (swapped_cnt == 0)
670 		return (NULL);
671 	res = NULL;
672 	ppri = INT_MIN;
673 	FOREACH_PROC_IN_SYSTEM(p) {
674 		PROC_LOCK(p);
675 		if (p->p_state == PRS_NEW || (p->p_flag & (P_SWAPPINGOUT |
676 		    P_SWAPPINGIN | P_INMEM)) != 0) {
677 			PROC_UNLOCK(p);
678 			continue;
679 		}
680 		if (p->p_state == PRS_NORMAL && (p->p_flag & P_WKILLED) != 0) {
681 			/*
682 			 * A swapped-out process might have mapped a
683 			 * large portion of the system's pages as
684 			 * anonymous memory.  There is no other way to
685 			 * release the memory other than to kill the
686 			 * process, for which we need to swap it in.
687 			 */
688 			return (p);
689 		}
690 		if (wkilled_only) {
691 			PROC_UNLOCK(p);
692 			continue;
693 		}
694 		swtime = (ticks - p->p_swtick) / hz;
695 		FOREACH_THREAD_IN_PROC(p, td) {
696 			/*
697 			 * An otherwise runnable thread of a process
698 			 * swapped out has only the TDI_SWAPPED bit set.
699 			 */
700 			thread_lock(td);
701 			if (td->td_inhibitors == TDI_SWAPPED) {
702 				slptime = (ticks - td->td_slptick) / hz;
703 				pri = swtime + slptime;
704 				if ((td->td_flags & TDF_SWAPINREQ) == 0)
705 					pri -= p->p_nice * 8;
706 				/*
707 				 * if this thread is higher priority
708 				 * and there is enough space, then select
709 				 * this process instead of the previous
710 				 * selection.
711 				 */
712 				if (pri > ppri) {
713 					res = p;
714 					ppri = pri;
715 				}
716 			}
717 			thread_unlock(td);
718 		}
719 		PROC_UNLOCK(p);
720 	}
721 
722 	if (res != NULL)
723 		PROC_LOCK(res);
724 	return (res);
725 }
726 
727 #define	SWAPIN_INTERVAL	(MAXSLP * hz / 2)
728 
729 /*
730  * Limit swapper to swap in one non-WKILLED process in MAXSLP/2
731  * interval, assuming that there is:
732  * - at least one domain that is not suffering from a shortage of free memory;
733  * - no parallel swap-ins;
734  * - no other swap-ins in the current SWAPIN_INTERVAL.
735  */
736 static bool
737 swapper_wkilled_only(void)
738 {
739 
740 	return (vm_page_count_min_set(&all_domains) || swap_inprogress > 0 ||
741 	    (u_int)(ticks - last_swapin) < SWAPIN_INTERVAL);
742 }
743 
744 void
745 swapper(void)
746 {
747 	struct proc *p;
748 
749 	for (;;) {
750 		sx_slock(&allproc_lock);
751 		p = swapper_selector(swapper_wkilled_only());
752 		sx_sunlock(&allproc_lock);
753 
754 		if (p == NULL) {
755 			tsleep(&proc0, PVM, "swapin", SWAPIN_INTERVAL);
756 		} else {
757 			PROC_LOCK_ASSERT(p, MA_OWNED);
758 
759 			/*
760 			 * Another process may be bringing or may have
761 			 * already brought this process in while we
762 			 * traverse all threads.  Or, this process may
763 			 * have exited or even being swapped out
764 			 * again.
765 			 */
766 			if (p->p_state == PRS_NORMAL && (p->p_flag & (P_INMEM |
767 			    P_SWAPPINGOUT | P_SWAPPINGIN)) == 0) {
768 				faultin(p);
769 			}
770 			PROC_UNLOCK(p);
771 		}
772 	}
773 }
774 
775 /*
776  * First, if any processes have been sleeping or stopped for at least
777  * "swap_idle_threshold1" seconds, they are swapped out.  If, however,
778  * no such processes exist, then the longest-sleeping or stopped
779  * process is swapped out.  Finally, and only as a last resort, if
780  * there are no sleeping or stopped processes, the longest-resident
781  * process is swapped out.
782  */
783 static void
784 swapout_procs(int action)
785 {
786 	struct proc *p;
787 	struct thread *td;
788 	int slptime;
789 	bool didswap, doswap;
790 
791 	MPASS((action & (VM_SWAP_NORMAL | VM_SWAP_IDLE)) != 0);
792 
793 	didswap = false;
794 	sx_slock(&allproc_lock);
795 	FOREACH_PROC_IN_SYSTEM(p) {
796 		/*
797 		 * Filter out not yet fully constructed processes.  Do
798 		 * not swap out held processes.  Avoid processes which
799 		 * are system, exiting, execing, traced, already swapped
800 		 * out or are in the process of being swapped in or out.
801 		 */
802 		PROC_LOCK(p);
803 		if (p->p_state != PRS_NORMAL || p->p_lock != 0 || (p->p_flag &
804 		    (P_SYSTEM | P_WEXIT | P_INEXEC | P_STOPPED_SINGLE |
805 		    P_TRACED | P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) !=
806 		    P_INMEM) {
807 			PROC_UNLOCK(p);
808 			continue;
809 		}
810 
811 		/*
812 		 * Further consideration of this process for swap out
813 		 * requires iterating over its threads.  We release
814 		 * allproc_lock here so that process creation and
815 		 * destruction are not blocked while we iterate.
816 		 *
817 		 * To later reacquire allproc_lock and resume
818 		 * iteration over the allproc list, we will first have
819 		 * to release the lock on the process.  We place a
820 		 * hold on the process so that it remains in the
821 		 * allproc list while it is unlocked.
822 		 */
823 		_PHOLD_LITE(p);
824 		sx_sunlock(&allproc_lock);
825 
826 		/*
827 		 * Do not swapout a realtime process.
828 		 * Guarantee swap_idle_threshold1 time in memory.
829 		 * If the system is under memory stress, or if we are
830 		 * swapping idle processes >= swap_idle_threshold2,
831 		 * then swap the process out.
832 		 */
833 		doswap = true;
834 		FOREACH_THREAD_IN_PROC(p, td) {
835 			thread_lock(td);
836 			slptime = (ticks - td->td_slptick) / hz;
837 			if (PRI_IS_REALTIME(td->td_pri_class) ||
838 			    slptime < swap_idle_threshold1 ||
839 			    !thread_safetoswapout(td) ||
840 			    ((action & VM_SWAP_NORMAL) == 0 &&
841 			    slptime < swap_idle_threshold2))
842 				doswap = false;
843 			thread_unlock(td);
844 			if (!doswap)
845 				break;
846 		}
847 		if (doswap && swapout(p) == 0)
848 			didswap = true;
849 
850 		PROC_UNLOCK(p);
851 		if (didswap) {
852 			sx_xlock(&allproc_lock);
853 			swapped_cnt++;
854 			sx_downgrade(&allproc_lock);
855 		} else
856 			sx_slock(&allproc_lock);
857 		PRELE(p);
858 	}
859 	sx_sunlock(&allproc_lock);
860 
861 	/*
862 	 * If we swapped something out, and another process needed memory,
863 	 * then wakeup the sched process.
864 	 */
865 	if (didswap)
866 		wakeup(&proc0);
867 }
868 
869 static void
870 swapclear(struct proc *p)
871 {
872 	struct thread *td;
873 
874 	PROC_LOCK_ASSERT(p, MA_OWNED);
875 
876 	FOREACH_THREAD_IN_PROC(p, td) {
877 		thread_lock(td);
878 		td->td_flags |= TDF_INMEM;
879 		td->td_flags &= ~TDF_SWAPINREQ;
880 		TD_CLR_SWAPPED(td);
881 		if (TD_CAN_RUN(td)) {
882 			if (setrunnable(td, 0)) {
883 #ifdef INVARIANTS
884 				/*
885 				 * XXX: We just cleared TDI_SWAPPED
886 				 * above and set TDF_INMEM, so this
887 				 * should never happen.
888 				 */
889 				panic("not waking up swapper");
890 #endif
891 			}
892 		} else
893 			thread_unlock(td);
894 	}
895 	p->p_flag &= ~(P_SWAPPINGIN | P_SWAPPINGOUT);
896 	p->p_flag |= P_INMEM;
897 }
898 
899 static int
900 swapout(struct proc *p)
901 {
902 	struct thread *td;
903 
904 	PROC_LOCK_ASSERT(p, MA_OWNED);
905 
906 	/*
907 	 * The states of this process and its threads may have changed
908 	 * by now.  Assuming that there is only one pageout daemon thread,
909 	 * this process should still be in memory.
910 	 */
911 	KASSERT((p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) ==
912 	    P_INMEM, ("swapout: lost a swapout race?"));
913 
914 	/*
915 	 * Remember the resident count.
916 	 */
917 	p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace);
918 
919 	/*
920 	 * Check and mark all threads before we proceed.
921 	 */
922 	p->p_flag &= ~P_INMEM;
923 	p->p_flag |= P_SWAPPINGOUT;
924 	FOREACH_THREAD_IN_PROC(p, td) {
925 		thread_lock(td);
926 		if (!thread_safetoswapout(td)) {
927 			thread_unlock(td);
928 			swapclear(p);
929 			return (EBUSY);
930 		}
931 		td->td_flags &= ~TDF_INMEM;
932 		TD_SET_SWAPPED(td);
933 		thread_unlock(td);
934 	}
935 	td = FIRST_THREAD_IN_PROC(p);
936 	++td->td_ru.ru_nswap;
937 	PROC_UNLOCK(p);
938 
939 	/*
940 	 * This list is stable because all threads are now prevented from
941 	 * running.  The list is only modified in the context of a running
942 	 * thread in this process.
943 	 */
944 	FOREACH_THREAD_IN_PROC(p, td)
945 		vm_thread_swapout(td);
946 
947 	PROC_LOCK(p);
948 	p->p_flag &= ~P_SWAPPINGOUT;
949 	p->p_swtick = ticks;
950 	return (0);
951 }
952