xref: /freebsd/sys/vm/vm_swapout.c (revision eb69d1f144a6fcc765d1b9d44a5ae8082353e70b)
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/_kstack_cache.h>
89 #include <sys/kthread.h>
90 #include <sys/ktr.h>
91 #include <sys/mount.h>
92 #include <sys/racct.h>
93 #include <sys/resourcevar.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 void swapclear(struct proc *);
162 static int swapout(struct proc *);
163 static void vm_swapout_map_deactivate_pages(vm_map_t, long);
164 static void vm_swapout_object_deactivate_pages(pmap_t, vm_object_t, long);
165 static void swapout_procs(int action);
166 static void vm_req_vmdaemon(int req);
167 static void vm_thread_swapin(struct thread *td);
168 static void vm_thread_swapout(struct thread *td);
169 
170 /*
171  *	vm_swapout_object_deactivate_pages
172  *
173  *	Deactivate enough pages to satisfy the inactive target
174  *	requirements.
175  *
176  *	The object and map must be locked.
177  */
178 static void
179 vm_swapout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object,
180     long desired)
181 {
182 	vm_object_t backing_object, object;
183 	vm_page_t p;
184 	int act_delta, remove_mode;
185 
186 	VM_OBJECT_ASSERT_LOCKED(first_object);
187 	if ((first_object->flags & OBJ_FICTITIOUS) != 0)
188 		return;
189 	for (object = first_object;; object = backing_object) {
190 		if (pmap_resident_count(pmap) <= desired)
191 			goto unlock_return;
192 		VM_OBJECT_ASSERT_LOCKED(object);
193 		if ((object->flags & OBJ_UNMANAGED) != 0 ||
194 		    object->paging_in_progress != 0)
195 			goto unlock_return;
196 
197 		remove_mode = 0;
198 		if (object->shadow_count > 1)
199 			remove_mode = 1;
200 		/*
201 		 * Scan the object's entire memory queue.
202 		 */
203 		TAILQ_FOREACH(p, &object->memq, listq) {
204 			if (pmap_resident_count(pmap) <= desired)
205 				goto unlock_return;
206 			if (should_yield())
207 				goto unlock_return;
208 			if (vm_page_busied(p))
209 				continue;
210 			VM_CNT_INC(v_pdpages);
211 			vm_page_lock(p);
212 			if (vm_page_held(p) ||
213 			    !pmap_page_exists_quick(pmap, p)) {
214 				vm_page_unlock(p);
215 				continue;
216 			}
217 			act_delta = pmap_ts_referenced(p);
218 			if ((p->aflags & PGA_REFERENCED) != 0) {
219 				if (act_delta == 0)
220 					act_delta = 1;
221 				vm_page_aflag_clear(p, PGA_REFERENCED);
222 			}
223 			if (!vm_page_active(p) && act_delta != 0) {
224 				vm_page_activate(p);
225 				p->act_count += act_delta;
226 			} else if (vm_page_active(p)) {
227 				if (act_delta == 0) {
228 					p->act_count -= min(p->act_count,
229 					    ACT_DECLINE);
230 					if (!remove_mode && p->act_count == 0) {
231 						pmap_remove_all(p);
232 						vm_page_deactivate(p);
233 					} else
234 						vm_page_requeue(p);
235 				} else {
236 					vm_page_activate(p);
237 					if (p->act_count < ACT_MAX -
238 					    ACT_ADVANCE)
239 						p->act_count += ACT_ADVANCE;
240 					vm_page_requeue(p);
241 				}
242 			} else if (vm_page_inactive(p))
243 				pmap_remove_all(p);
244 			vm_page_unlock(p);
245 		}
246 		if ((backing_object = object->backing_object) == NULL)
247 			goto unlock_return;
248 		VM_OBJECT_RLOCK(backing_object);
249 		if (object != first_object)
250 			VM_OBJECT_RUNLOCK(object);
251 	}
252 unlock_return:
253 	if (object != first_object)
254 		VM_OBJECT_RUNLOCK(object);
255 }
256 
257 /*
258  * deactivate some number of pages in a map, try to do it fairly, but
259  * that is really hard to do.
260  */
261 static void
262 vm_swapout_map_deactivate_pages(vm_map_t map, long desired)
263 {
264 	vm_map_entry_t tmpe;
265 	vm_object_t obj, bigobj;
266 	int nothingwired;
267 
268 	if (!vm_map_trylock_read(map))
269 		return;
270 
271 	bigobj = NULL;
272 	nothingwired = TRUE;
273 
274 	/*
275 	 * first, search out the biggest object, and try to free pages from
276 	 * that.
277 	 */
278 	tmpe = map->header.next;
279 	while (tmpe != &map->header) {
280 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
281 			obj = tmpe->object.vm_object;
282 			if (obj != NULL && VM_OBJECT_TRYRLOCK(obj)) {
283 				if (obj->shadow_count <= 1 &&
284 				    (bigobj == NULL ||
285 				     bigobj->resident_page_count <
286 				     obj->resident_page_count)) {
287 					if (bigobj != NULL)
288 						VM_OBJECT_RUNLOCK(bigobj);
289 					bigobj = obj;
290 				} else
291 					VM_OBJECT_RUNLOCK(obj);
292 			}
293 		}
294 		if (tmpe->wired_count > 0)
295 			nothingwired = FALSE;
296 		tmpe = tmpe->next;
297 	}
298 
299 	if (bigobj != NULL) {
300 		vm_swapout_object_deactivate_pages(map->pmap, bigobj, desired);
301 		VM_OBJECT_RUNLOCK(bigobj);
302 	}
303 	/*
304 	 * Next, hunt around for other pages to deactivate.  We actually
305 	 * do this search sort of wrong -- .text first is not the best idea.
306 	 */
307 	tmpe = map->header.next;
308 	while (tmpe != &map->header) {
309 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
310 			break;
311 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
312 			obj = tmpe->object.vm_object;
313 			if (obj != NULL) {
314 				VM_OBJECT_RLOCK(obj);
315 				vm_swapout_object_deactivate_pages(map->pmap,
316 				    obj, desired);
317 				VM_OBJECT_RUNLOCK(obj);
318 			}
319 		}
320 		tmpe = tmpe->next;
321 	}
322 
323 	/*
324 	 * Remove all mappings if a process is swapped out, this will free page
325 	 * table pages.
326 	 */
327 	if (desired == 0 && nothingwired) {
328 		pmap_remove(vm_map_pmap(map), vm_map_min(map),
329 		    vm_map_max(map));
330 	}
331 
332 	vm_map_unlock_read(map);
333 }
334 
335 /*
336  * Swap out requests
337  */
338 #define VM_SWAP_NORMAL 1
339 #define VM_SWAP_IDLE 2
340 
341 void
342 vm_swapout_run(void)
343 {
344 
345 	if (vm_swap_enabled)
346 		vm_req_vmdaemon(VM_SWAP_NORMAL);
347 }
348 
349 /*
350  * Idle process swapout -- run once per second when pagedaemons are
351  * reclaiming pages.
352  */
353 void
354 vm_swapout_run_idle(void)
355 {
356 	static long lsec;
357 
358 	if (!vm_swap_idle_enabled || time_second == lsec)
359 		return;
360 	vm_req_vmdaemon(VM_SWAP_IDLE);
361 	lsec = time_second;
362 }
363 
364 static void
365 vm_req_vmdaemon(int req)
366 {
367 	static int lastrun = 0;
368 
369 	mtx_lock(&vm_daemon_mtx);
370 	vm_pageout_req_swapout |= req;
371 	if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
372 		wakeup(&vm_daemon_needed);
373 		lastrun = ticks;
374 	}
375 	mtx_unlock(&vm_daemon_mtx);
376 }
377 
378 static void
379 vm_daemon(void)
380 {
381 	struct rlimit rsslim;
382 	struct proc *p;
383 	struct thread *td;
384 	struct vmspace *vm;
385 	int breakout, swapout_flags, tryagain, attempts;
386 #ifdef RACCT
387 	uint64_t rsize, ravailable;
388 #endif
389 
390 	while (TRUE) {
391 		mtx_lock(&vm_daemon_mtx);
392 		msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep",
393 #ifdef RACCT
394 		    racct_enable ? hz : 0
395 #else
396 		    0
397 #endif
398 		);
399 		swapout_flags = vm_pageout_req_swapout;
400 		vm_pageout_req_swapout = 0;
401 		mtx_unlock(&vm_daemon_mtx);
402 		if (swapout_flags)
403 			swapout_procs(swapout_flags);
404 
405 		/*
406 		 * scan the processes for exceeding their rlimits or if
407 		 * process is swapped out -- deactivate pages
408 		 */
409 		tryagain = 0;
410 		attempts = 0;
411 again:
412 		attempts++;
413 		sx_slock(&allproc_lock);
414 		FOREACH_PROC_IN_SYSTEM(p) {
415 			vm_pindex_t limit, size;
416 
417 			/*
418 			 * if this is a system process or if we have already
419 			 * looked at this process, skip it.
420 			 */
421 			PROC_LOCK(p);
422 			if (p->p_state != PRS_NORMAL ||
423 			    p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
424 				PROC_UNLOCK(p);
425 				continue;
426 			}
427 			/*
428 			 * if the process is in a non-running type state,
429 			 * don't touch it.
430 			 */
431 			breakout = 0;
432 			FOREACH_THREAD_IN_PROC(p, td) {
433 				thread_lock(td);
434 				if (!TD_ON_RUNQ(td) &&
435 				    !TD_IS_RUNNING(td) &&
436 				    !TD_IS_SLEEPING(td) &&
437 				    !TD_IS_SUSPENDED(td)) {
438 					thread_unlock(td);
439 					breakout = 1;
440 					break;
441 				}
442 				thread_unlock(td);
443 			}
444 			if (breakout) {
445 				PROC_UNLOCK(p);
446 				continue;
447 			}
448 			/*
449 			 * get a limit
450 			 */
451 			lim_rlimit_proc(p, RLIMIT_RSS, &rsslim);
452 			limit = OFF_TO_IDX(
453 			    qmin(rsslim.rlim_cur, rsslim.rlim_max));
454 
455 			/*
456 			 * let processes that are swapped out really be
457 			 * swapped out set the limit to nothing (will force a
458 			 * swap-out.)
459 			 */
460 			if ((p->p_flag & P_INMEM) == 0)
461 				limit = 0;	/* XXX */
462 			vm = vmspace_acquire_ref(p);
463 			_PHOLD_LITE(p);
464 			PROC_UNLOCK(p);
465 			if (vm == NULL) {
466 				PRELE(p);
467 				continue;
468 			}
469 			sx_sunlock(&allproc_lock);
470 
471 			size = vmspace_resident_count(vm);
472 			if (size >= limit) {
473 				vm_swapout_map_deactivate_pages(
474 				    &vm->vm_map, limit);
475 				size = vmspace_resident_count(vm);
476 			}
477 #ifdef RACCT
478 			if (racct_enable) {
479 				rsize = IDX_TO_OFF(size);
480 				PROC_LOCK(p);
481 				if (p->p_state == PRS_NORMAL)
482 					racct_set(p, RACCT_RSS, rsize);
483 				ravailable = racct_get_available(p, RACCT_RSS);
484 				PROC_UNLOCK(p);
485 				if (rsize > ravailable) {
486 					/*
487 					 * Don't be overly aggressive; this
488 					 * might be an innocent process,
489 					 * and the limit could've been exceeded
490 					 * by some memory hog.  Don't try
491 					 * to deactivate more than 1/4th
492 					 * of process' resident set size.
493 					 */
494 					if (attempts <= 8) {
495 						if (ravailable < rsize -
496 						    (rsize / 4)) {
497 							ravailable = rsize -
498 							    (rsize / 4);
499 						}
500 					}
501 					vm_swapout_map_deactivate_pages(
502 					    &vm->vm_map,
503 					    OFF_TO_IDX(ravailable));
504 					/* Update RSS usage after paging out. */
505 					size = vmspace_resident_count(vm);
506 					rsize = IDX_TO_OFF(size);
507 					PROC_LOCK(p);
508 					if (p->p_state == PRS_NORMAL)
509 						racct_set(p, RACCT_RSS, rsize);
510 					PROC_UNLOCK(p);
511 					if (rsize > ravailable)
512 						tryagain = 1;
513 				}
514 			}
515 #endif
516 			vmspace_free(vm);
517 			sx_slock(&allproc_lock);
518 			PRELE(p);
519 		}
520 		sx_sunlock(&allproc_lock);
521 		if (tryagain != 0 && attempts <= 10) {
522 			maybe_yield();
523 			goto again;
524 		}
525 	}
526 }
527 
528 /*
529  * Allow a thread's kernel stack to be paged out.
530  */
531 static void
532 vm_thread_swapout(struct thread *td)
533 {
534 	vm_object_t ksobj;
535 	vm_page_t m;
536 	int i, pages;
537 
538 	cpu_thread_swapout(td);
539 	pages = td->td_kstack_pages;
540 	ksobj = td->td_kstack_obj;
541 	pmap_qremove(td->td_kstack, pages);
542 	VM_OBJECT_WLOCK(ksobj);
543 	for (i = 0; i < pages; i++) {
544 		m = vm_page_lookup(ksobj, i);
545 		if (m == NULL)
546 			panic("vm_thread_swapout: kstack already missing?");
547 		vm_page_dirty(m);
548 		vm_page_lock(m);
549 		vm_page_unwire(m, PQ_LAUNDRY);
550 		vm_page_unlock(m);
551 	}
552 	VM_OBJECT_WUNLOCK(ksobj);
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)
560 {
561 	vm_object_t ksobj;
562 	vm_page_t ma[KSTACK_MAX_PAGES];
563 	int a, count, i, j, pages, rv;
564 
565 	pages = td->td_kstack_pages;
566 	ksobj = td->td_kstack_obj;
567 	VM_OBJECT_WLOCK(ksobj);
568 	(void)vm_page_grab_pages(ksobj, 0, VM_ALLOC_NORMAL | VM_ALLOC_WIRED, ma,
569 	    pages);
570 	for (i = 0; i < pages;) {
571 		vm_page_assert_xbusied(ma[i]);
572 		if (ma[i]->valid == VM_PAGE_BITS_ALL) {
573 			vm_page_xunbusy(ma[i]);
574 			i++;
575 			continue;
576 		}
577 		vm_object_pip_add(ksobj, 1);
578 		for (j = i + 1; j < pages; j++)
579 			if (ma[j]->valid == VM_PAGE_BITS_ALL)
580 				break;
581 		rv = vm_pager_has_page(ksobj, ma[i]->pindex, NULL, &a);
582 		KASSERT(rv == 1, ("%s: missing page %p", __func__, ma[i]));
583 		count = min(a + 1, j - i);
584 		rv = vm_pager_get_pages(ksobj, 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(ksobj);
588 		for (j = i; j < i + count; j++)
589 			vm_page_xunbusy(ma[j]);
590 		i += count;
591 	}
592 	VM_OBJECT_WUNLOCK(ksobj);
593 	pmap_qenter(td->td_kstack, ma, pages);
594 	cpu_thread_swapin(td);
595 }
596 
597 void
598 faultin(struct proc *p)
599 {
600 	struct thread *td;
601 
602 	PROC_LOCK_ASSERT(p, MA_OWNED);
603 	/*
604 	 * If another process is swapping in this process,
605 	 * just wait until it finishes.
606 	 */
607 	if (p->p_flag & P_SWAPPINGIN) {
608 		while (p->p_flag & P_SWAPPINGIN)
609 			msleep(&p->p_flag, &p->p_mtx, PVM, "faultin", 0);
610 		return;
611 	}
612 	if ((p->p_flag & P_INMEM) == 0) {
613 		/*
614 		 * Don't let another thread swap process p out while we are
615 		 * busy swapping it in.
616 		 */
617 		++p->p_lock;
618 		p->p_flag |= P_SWAPPINGIN;
619 		PROC_UNLOCK(p);
620 
621 		/*
622 		 * We hold no lock here because the list of threads
623 		 * can not change while all threads in the process are
624 		 * swapped out.
625 		 */
626 		FOREACH_THREAD_IN_PROC(p, td)
627 			vm_thread_swapin(td);
628 		PROC_LOCK(p);
629 		swapclear(p);
630 		p->p_swtick = ticks;
631 
632 		wakeup(&p->p_flag);
633 
634 		/* Allow other threads to swap p out now. */
635 		--p->p_lock;
636 	}
637 }
638 
639 /*
640  * This swapin algorithm attempts to swap-in processes only if there
641  * is enough space for them.  Of course, if a process waits for a long
642  * time, it will be swapped in anyway.
643  */
644 void
645 swapper(void)
646 {
647 	struct proc *p, *pp;
648 	struct thread *td;
649 	int ppri, pri, slptime, swtime;
650 
651 loop:
652 	if (vm_page_count_min()) {
653 		vm_wait_min();
654 		goto loop;
655 	}
656 
657 	pp = NULL;
658 	ppri = INT_MIN;
659 	sx_slock(&allproc_lock);
660 	FOREACH_PROC_IN_SYSTEM(p) {
661 		PROC_LOCK(p);
662 		if (p->p_state == PRS_NEW ||
663 		    p->p_flag & (P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) {
664 			PROC_UNLOCK(p);
665 			continue;
666 		}
667 		swtime = (ticks - p->p_swtick) / hz;
668 		FOREACH_THREAD_IN_PROC(p, td) {
669 			/*
670 			 * An otherwise runnable thread of a process
671 			 * swapped out has only the TDI_SWAPPED bit set.
672 			 */
673 			thread_lock(td);
674 			if (td->td_inhibitors == TDI_SWAPPED) {
675 				slptime = (ticks - td->td_slptick) / hz;
676 				pri = swtime + slptime;
677 				if ((td->td_flags & TDF_SWAPINREQ) == 0)
678 					pri -= p->p_nice * 8;
679 				/*
680 				 * if this thread is higher priority
681 				 * and there is enough space, then select
682 				 * this process instead of the previous
683 				 * selection.
684 				 */
685 				if (pri > ppri) {
686 					pp = p;
687 					ppri = pri;
688 				}
689 			}
690 			thread_unlock(td);
691 		}
692 		PROC_UNLOCK(p);
693 	}
694 	sx_sunlock(&allproc_lock);
695 
696 	/*
697 	 * Nothing to do, back to sleep.
698 	 */
699 	if ((p = pp) == NULL) {
700 		tsleep(&proc0, PVM, "swapin", MAXSLP * hz / 2);
701 		goto loop;
702 	}
703 	PROC_LOCK(p);
704 
705 	/*
706 	 * Another process may be bringing or may have already
707 	 * brought this process in while we traverse all threads.
708 	 * Or, this process may even be being swapped out again.
709 	 */
710 	if (p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) {
711 		PROC_UNLOCK(p);
712 		goto loop;
713 	}
714 
715 	/*
716 	 * We would like to bring someone in.
717 	 */
718 	faultin(p);
719 	PROC_UNLOCK(p);
720 	goto loop;
721 }
722 
723 /*
724  * First, if any processes have been sleeping or stopped for at least
725  * "swap_idle_threshold1" seconds, they are swapped out.  If, however,
726  * no such processes exist, then the longest-sleeping or stopped
727  * process is swapped out.  Finally, and only as a last resort, if
728  * there are no sleeping or stopped processes, the longest-resident
729  * process is swapped out.
730  */
731 static void
732 swapout_procs(int action)
733 {
734 	struct proc *p;
735 	struct thread *td;
736 	int slptime;
737 	bool didswap, doswap;
738 
739 	MPASS((action & (VM_SWAP_NORMAL | VM_SWAP_IDLE)) != 0);
740 
741 	didswap = false;
742 	sx_slock(&allproc_lock);
743 	FOREACH_PROC_IN_SYSTEM(p) {
744 		/*
745 		 * Filter out not yet fully constructed processes.  Do
746 		 * not swap out held processes.  Avoid processes which
747 		 * are system, exiting, execing, traced, already swapped
748 		 * out or are in the process of being swapped in or out.
749 		 */
750 		PROC_LOCK(p);
751 		if (p->p_state != PRS_NORMAL || p->p_lock != 0 || (p->p_flag &
752 		    (P_SYSTEM | P_WEXIT | P_INEXEC | P_STOPPED_SINGLE |
753 		    P_TRACED | P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) !=
754 		    P_INMEM) {
755 			PROC_UNLOCK(p);
756 			continue;
757 		}
758 
759 		/*
760 		 * Further consideration of this process for swap out
761 		 * requires iterating over its threads.  We release
762 		 * allproc_lock here so that process creation and
763 		 * destruction are not blocked while we iterate.
764 		 *
765 		 * To later reacquire allproc_lock and resume
766 		 * iteration over the allproc list, we will first have
767 		 * to release the lock on the process.  We place a
768 		 * hold on the process so that it remains in the
769 		 * allproc list while it is unlocked.
770 		 */
771 		_PHOLD_LITE(p);
772 		sx_sunlock(&allproc_lock);
773 
774 		/*
775 		 * Do not swapout a realtime process.
776 		 * Guarantee swap_idle_threshold1 time in memory.
777 		 * If the system is under memory stress, or if we are
778 		 * swapping idle processes >= swap_idle_threshold2,
779 		 * then swap the process out.
780 		 */
781 		doswap = true;
782 		FOREACH_THREAD_IN_PROC(p, td) {
783 			thread_lock(td);
784 			slptime = (ticks - td->td_slptick) / hz;
785 			if (PRI_IS_REALTIME(td->td_pri_class) ||
786 			    slptime < swap_idle_threshold1 ||
787 			    !thread_safetoswapout(td) ||
788 			    ((action & VM_SWAP_NORMAL) == 0 &&
789 			    slptime < swap_idle_threshold2))
790 				doswap = false;
791 			thread_unlock(td);
792 			if (!doswap)
793 				break;
794 		}
795 		if (doswap && swapout(p) == 0)
796 			didswap = true;
797 
798 		PROC_UNLOCK(p);
799 		sx_slock(&allproc_lock);
800 		PRELE(p);
801 	}
802 	sx_sunlock(&allproc_lock);
803 
804 	/*
805 	 * If we swapped something out, and another process needed memory,
806 	 * then wakeup the sched process.
807 	 */
808 	if (didswap)
809 		wakeup(&proc0);
810 }
811 
812 static void
813 swapclear(struct proc *p)
814 {
815 	struct thread *td;
816 
817 	PROC_LOCK_ASSERT(p, MA_OWNED);
818 
819 	FOREACH_THREAD_IN_PROC(p, td) {
820 		thread_lock(td);
821 		td->td_flags |= TDF_INMEM;
822 		td->td_flags &= ~TDF_SWAPINREQ;
823 		TD_CLR_SWAPPED(td);
824 		if (TD_CAN_RUN(td))
825 			if (setrunnable(td)) {
826 #ifdef INVARIANTS
827 				/*
828 				 * XXX: We just cleared TDI_SWAPPED
829 				 * above and set TDF_INMEM, so this
830 				 * should never happen.
831 				 */
832 				panic("not waking up swapper");
833 #endif
834 			}
835 		thread_unlock(td);
836 	}
837 	p->p_flag &= ~(P_SWAPPINGIN | P_SWAPPINGOUT);
838 	p->p_flag |= P_INMEM;
839 }
840 
841 static int
842 swapout(struct proc *p)
843 {
844 	struct thread *td;
845 
846 	PROC_LOCK_ASSERT(p, MA_OWNED);
847 
848 	/*
849 	 * The states of this process and its threads may have changed
850 	 * by now.  Assuming that there is only one pageout daemon thread,
851 	 * this process should still be in memory.
852 	 */
853 	KASSERT((p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) ==
854 	    P_INMEM, ("swapout: lost a swapout race?"));
855 
856 	/*
857 	 * Remember the resident count.
858 	 */
859 	p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace);
860 
861 	/*
862 	 * Check and mark all threads before we proceed.
863 	 */
864 	p->p_flag &= ~P_INMEM;
865 	p->p_flag |= P_SWAPPINGOUT;
866 	FOREACH_THREAD_IN_PROC(p, td) {
867 		thread_lock(td);
868 		if (!thread_safetoswapout(td)) {
869 			thread_unlock(td);
870 			swapclear(p);
871 			return (EBUSY);
872 		}
873 		td->td_flags &= ~TDF_INMEM;
874 		TD_SET_SWAPPED(td);
875 		thread_unlock(td);
876 	}
877 	td = FIRST_THREAD_IN_PROC(p);
878 	++td->td_ru.ru_nswap;
879 	PROC_UNLOCK(p);
880 
881 	/*
882 	 * This list is stable because all threads are now prevented from
883 	 * running.  The list is only modified in the context of a running
884 	 * thread in this process.
885 	 */
886 	FOREACH_THREAD_IN_PROC(p, td)
887 		vm_thread_swapout(td);
888 
889 	PROC_LOCK(p);
890 	p->p_flag &= ~P_SWAPPINGOUT;
891 	p->p_swtick = ticks;
892 	return (0);
893 }
894