xref: /freebsd/sys/vm/vm_swapout.c (revision 499fe48de8938d4c7b0a91e20eb6c16db9d55633)
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_pages(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 /*
174  *	vm_swapout_object_deactivate_pages
175  *
176  *	Deactivate enough pages to satisfy the inactive target
177  *	requirements.
178  *
179  *	The object and map must be locked.
180  */
181 static void
182 vm_swapout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object,
183     long desired)
184 {
185 	vm_object_t backing_object, object;
186 	vm_page_t p;
187 	int act_delta, remove_mode;
188 
189 	VM_OBJECT_ASSERT_LOCKED(first_object);
190 	if ((first_object->flags & OBJ_FICTITIOUS) != 0)
191 		return;
192 	for (object = first_object;; object = backing_object) {
193 		if (pmap_resident_count(pmap) <= desired)
194 			goto unlock_return;
195 		VM_OBJECT_ASSERT_LOCKED(object);
196 		if ((object->flags & OBJ_UNMANAGED) != 0 ||
197 		    REFCOUNT_COUNT(object->paging_in_progress) > 0)
198 			goto unlock_return;
199 
200 		remove_mode = 0;
201 		if (object->shadow_count > 1)
202 			remove_mode = 1;
203 		/*
204 		 * Scan the object's entire memory queue.
205 		 */
206 		TAILQ_FOREACH(p, &object->memq, listq) {
207 			if (pmap_resident_count(pmap) <= desired)
208 				goto unlock_return;
209 			if (should_yield())
210 				goto unlock_return;
211 			if (vm_page_tryxbusy(p) == 0)
212 				continue;
213 			VM_CNT_INC(v_pdpages);
214 
215 			/*
216 			 * The page may acquire a wiring after this check.
217 			 * The page daemon handles wired pages, so there is
218 			 * no harm done if a wiring appears while we are
219 			 * attempting to deactivate the page.
220 			 */
221 			if (vm_page_wired(p) || !pmap_page_exists_quick(pmap, p)) {
222 				vm_page_xunbusy(p);
223 				continue;
224 			}
225 			act_delta = pmap_ts_referenced(p);
226 			vm_page_lock(p);
227 			if ((p->aflags & PGA_REFERENCED) != 0) {
228 				if (act_delta == 0)
229 					act_delta = 1;
230 				vm_page_aflag_clear(p, PGA_REFERENCED);
231 			}
232 			if (!vm_page_active(p) && act_delta != 0) {
233 				vm_page_activate(p);
234 				p->act_count += act_delta;
235 			} else if (vm_page_active(p)) {
236 				/*
237 				 * The page daemon does not requeue pages
238 				 * after modifying their activation count.
239 				 */
240 				if (act_delta == 0) {
241 					p->act_count -= min(p->act_count,
242 					    ACT_DECLINE);
243 					if (!remove_mode && p->act_count == 0) {
244 						(void)vm_page_try_remove_all(p);
245 						vm_page_deactivate(p);
246 					}
247 				} else {
248 					vm_page_activate(p);
249 					if (p->act_count < ACT_MAX -
250 					    ACT_ADVANCE)
251 						p->act_count += ACT_ADVANCE;
252 				}
253 			} else if (vm_page_inactive(p))
254 				(void)vm_page_try_remove_all(p);
255 			vm_page_unlock(p);
256 			vm_page_xunbusy(p);
257 		}
258 		if ((backing_object = object->backing_object) == NULL)
259 			goto unlock_return;
260 		VM_OBJECT_RLOCK(backing_object);
261 		if (object != first_object)
262 			VM_OBJECT_RUNLOCK(object);
263 	}
264 unlock_return:
265 	if (object != first_object)
266 		VM_OBJECT_RUNLOCK(object);
267 }
268 
269 /*
270  * deactivate some number of pages in a map, try to do it fairly, but
271  * that is really hard to do.
272  */
273 static void
274 vm_swapout_map_deactivate_pages(vm_map_t map, long desired)
275 {
276 	vm_map_entry_t tmpe;
277 	vm_object_t obj, bigobj;
278 	int nothingwired;
279 
280 	if (!vm_map_trylock_read(map))
281 		return;
282 
283 	bigobj = NULL;
284 	nothingwired = TRUE;
285 
286 	/*
287 	 * first, search out the biggest object, and try to free pages from
288 	 * that.
289 	 */
290 	VM_MAP_ENTRY_FOREACH(tmpe, map) {
291 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
292 			obj = tmpe->object.vm_object;
293 			if (obj != NULL && VM_OBJECT_TRYRLOCK(obj)) {
294 				if (obj->shadow_count <= 1 &&
295 				    (bigobj == NULL ||
296 				     bigobj->resident_page_count <
297 				     obj->resident_page_count)) {
298 					if (bigobj != NULL)
299 						VM_OBJECT_RUNLOCK(bigobj);
300 					bigobj = obj;
301 				} else
302 					VM_OBJECT_RUNLOCK(obj);
303 			}
304 		}
305 		if (tmpe->wired_count > 0)
306 			nothingwired = FALSE;
307 	}
308 
309 	if (bigobj != NULL) {
310 		vm_swapout_object_deactivate_pages(map->pmap, bigobj, desired);
311 		VM_OBJECT_RUNLOCK(bigobj);
312 	}
313 	/*
314 	 * Next, hunt around for other pages to deactivate.  We actually
315 	 * do this search sort of wrong -- .text first is not the best idea.
316 	 */
317 	VM_MAP_ENTRY_FOREACH(tmpe, map) {
318 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
319 			break;
320 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
321 			obj = tmpe->object.vm_object;
322 			if (obj != NULL) {
323 				VM_OBJECT_RLOCK(obj);
324 				vm_swapout_object_deactivate_pages(map->pmap,
325 				    obj, desired);
326 				VM_OBJECT_RUNLOCK(obj);
327 			}
328 		}
329 	}
330 
331 	/*
332 	 * Remove all mappings if a process is swapped out, this will free page
333 	 * table pages.
334 	 */
335 	if (desired == 0 && nothingwired) {
336 		pmap_remove(vm_map_pmap(map), vm_map_min(map),
337 		    vm_map_max(map));
338 	}
339 
340 	vm_map_unlock_read(map);
341 }
342 
343 /*
344  * Swap out requests
345  */
346 #define VM_SWAP_NORMAL 1
347 #define VM_SWAP_IDLE 2
348 
349 void
350 vm_swapout_run(void)
351 {
352 
353 	if (vm_swap_enabled)
354 		vm_req_vmdaemon(VM_SWAP_NORMAL);
355 }
356 
357 /*
358  * Idle process swapout -- run once per second when pagedaemons are
359  * reclaiming pages.
360  */
361 void
362 vm_swapout_run_idle(void)
363 {
364 	static long lsec;
365 
366 	if (!vm_swap_idle_enabled || time_second == lsec)
367 		return;
368 	vm_req_vmdaemon(VM_SWAP_IDLE);
369 	lsec = time_second;
370 }
371 
372 static void
373 vm_req_vmdaemon(int req)
374 {
375 	static int lastrun = 0;
376 
377 	mtx_lock(&vm_daemon_mtx);
378 	vm_pageout_req_swapout |= req;
379 	if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
380 		wakeup(&vm_daemon_needed);
381 		lastrun = ticks;
382 	}
383 	mtx_unlock(&vm_daemon_mtx);
384 }
385 
386 static void
387 vm_daemon(void)
388 {
389 	struct rlimit rsslim;
390 	struct proc *p;
391 	struct thread *td;
392 	struct vmspace *vm;
393 	int breakout, swapout_flags, tryagain, attempts;
394 #ifdef RACCT
395 	uint64_t rsize, ravailable;
396 #endif
397 
398 	while (TRUE) {
399 		mtx_lock(&vm_daemon_mtx);
400 		msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep",
401 #ifdef RACCT
402 		    racct_enable ? hz : 0
403 #else
404 		    0
405 #endif
406 		);
407 		swapout_flags = vm_pageout_req_swapout;
408 		vm_pageout_req_swapout = 0;
409 		mtx_unlock(&vm_daemon_mtx);
410 		if (swapout_flags != 0) {
411 			/*
412 			 * Drain the per-CPU page queue batches as a deadlock
413 			 * avoidance measure.
414 			 */
415 			if ((swapout_flags & VM_SWAP_NORMAL) != 0)
416 				vm_page_pqbatch_drain();
417 			swapout_procs(swapout_flags);
418 		}
419 
420 		/*
421 		 * scan the processes for exceeding their rlimits or if
422 		 * process is swapped out -- deactivate pages
423 		 */
424 		tryagain = 0;
425 		attempts = 0;
426 again:
427 		attempts++;
428 		sx_slock(&allproc_lock);
429 		FOREACH_PROC_IN_SYSTEM(p) {
430 			vm_pindex_t limit, size;
431 
432 			/*
433 			 * if this is a system process or if we have already
434 			 * looked at this process, skip it.
435 			 */
436 			PROC_LOCK(p);
437 			if (p->p_state != PRS_NORMAL ||
438 			    p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
439 				PROC_UNLOCK(p);
440 				continue;
441 			}
442 			/*
443 			 * if the process is in a non-running type state,
444 			 * don't touch it.
445 			 */
446 			breakout = 0;
447 			FOREACH_THREAD_IN_PROC(p, td) {
448 				thread_lock(td);
449 				if (!TD_ON_RUNQ(td) &&
450 				    !TD_IS_RUNNING(td) &&
451 				    !TD_IS_SLEEPING(td) &&
452 				    !TD_IS_SUSPENDED(td)) {
453 					thread_unlock(td);
454 					breakout = 1;
455 					break;
456 				}
457 				thread_unlock(td);
458 			}
459 			if (breakout) {
460 				PROC_UNLOCK(p);
461 				continue;
462 			}
463 			/*
464 			 * get a limit
465 			 */
466 			lim_rlimit_proc(p, RLIMIT_RSS, &rsslim);
467 			limit = OFF_TO_IDX(
468 			    qmin(rsslim.rlim_cur, rsslim.rlim_max));
469 
470 			/*
471 			 * let processes that are swapped out really be
472 			 * swapped out set the limit to nothing (will force a
473 			 * swap-out.)
474 			 */
475 			if ((p->p_flag & P_INMEM) == 0)
476 				limit = 0;	/* XXX */
477 			vm = vmspace_acquire_ref(p);
478 			_PHOLD_LITE(p);
479 			PROC_UNLOCK(p);
480 			if (vm == NULL) {
481 				PRELE(p);
482 				continue;
483 			}
484 			sx_sunlock(&allproc_lock);
485 
486 			size = vmspace_resident_count(vm);
487 			if (size >= limit) {
488 				vm_swapout_map_deactivate_pages(
489 				    &vm->vm_map, limit);
490 				size = vmspace_resident_count(vm);
491 			}
492 #ifdef RACCT
493 			if (racct_enable) {
494 				rsize = IDX_TO_OFF(size);
495 				PROC_LOCK(p);
496 				if (p->p_state == PRS_NORMAL)
497 					racct_set(p, RACCT_RSS, rsize);
498 				ravailable = racct_get_available(p, RACCT_RSS);
499 				PROC_UNLOCK(p);
500 				if (rsize > ravailable) {
501 					/*
502 					 * Don't be overly aggressive; this
503 					 * might be an innocent process,
504 					 * and the limit could've been exceeded
505 					 * by some memory hog.  Don't try
506 					 * to deactivate more than 1/4th
507 					 * of process' resident set size.
508 					 */
509 					if (attempts <= 8) {
510 						if (ravailable < rsize -
511 						    (rsize / 4)) {
512 							ravailable = rsize -
513 							    (rsize / 4);
514 						}
515 					}
516 					vm_swapout_map_deactivate_pages(
517 					    &vm->vm_map,
518 					    OFF_TO_IDX(ravailable));
519 					/* Update RSS usage after paging out. */
520 					size = vmspace_resident_count(vm);
521 					rsize = IDX_TO_OFF(size);
522 					PROC_LOCK(p);
523 					if (p->p_state == PRS_NORMAL)
524 						racct_set(p, RACCT_RSS, rsize);
525 					PROC_UNLOCK(p);
526 					if (rsize > ravailable)
527 						tryagain = 1;
528 				}
529 			}
530 #endif
531 			vmspace_free(vm);
532 			sx_slock(&allproc_lock);
533 			PRELE(p);
534 		}
535 		sx_sunlock(&allproc_lock);
536 		if (tryagain != 0 && attempts <= 10) {
537 			maybe_yield();
538 			goto again;
539 		}
540 	}
541 }
542 
543 /*
544  * Allow a thread's kernel stack to be paged out.
545  */
546 static void
547 vm_thread_swapout(struct thread *td)
548 {
549 	vm_object_t ksobj;
550 	vm_page_t m;
551 	int i, pages;
552 
553 	cpu_thread_swapout(td);
554 	pages = td->td_kstack_pages;
555 	ksobj = td->td_kstack_obj;
556 	pmap_qremove(td->td_kstack, pages);
557 	VM_OBJECT_WLOCK(ksobj);
558 	for (i = 0; i < pages; i++) {
559 		m = vm_page_lookup(ksobj, i);
560 		if (m == NULL)
561 			panic("vm_thread_swapout: kstack already missing?");
562 		vm_page_dirty(m);
563 		vm_page_unwire(m, PQ_LAUNDRY);
564 	}
565 	VM_OBJECT_WUNLOCK(ksobj);
566 }
567 
568 /*
569  * Bring the kernel stack for a specified thread back in.
570  */
571 static void
572 vm_thread_swapin(struct thread *td, int oom_alloc)
573 {
574 	vm_object_t ksobj;
575 	vm_page_t ma[KSTACK_MAX_PAGES];
576 	int a, count, i, j, pages, rv;
577 
578 	pages = td->td_kstack_pages;
579 	ksobj = td->td_kstack_obj;
580 	VM_OBJECT_WLOCK(ksobj);
581 	(void)vm_page_grab_pages(ksobj, 0, oom_alloc | VM_ALLOC_WIRED, ma,
582 	    pages);
583 	for (i = 0; i < pages;) {
584 		vm_page_assert_xbusied(ma[i]);
585 		if (vm_page_all_valid(ma[i])) {
586 			vm_page_xunbusy(ma[i]);
587 			i++;
588 			continue;
589 		}
590 		vm_object_pip_add(ksobj, 1);
591 		for (j = i + 1; j < pages; j++)
592 			if (vm_page_all_valid(ma[j]))
593 				break;
594 		rv = vm_pager_has_page(ksobj, ma[i]->pindex, NULL, &a);
595 		KASSERT(rv == 1, ("%s: missing page %p", __func__, ma[i]));
596 		count = min(a + 1, j - i);
597 		rv = vm_pager_get_pages(ksobj, ma + i, count, NULL, NULL);
598 		KASSERT(rv == VM_PAGER_OK, ("%s: cannot get kstack for proc %d",
599 		    __func__, td->td_proc->p_pid));
600 		vm_object_pip_wakeup(ksobj);
601 		for (j = i; j < i + count; j++)
602 			vm_page_xunbusy(ma[j]);
603 		i += count;
604 	}
605 	VM_OBJECT_WUNLOCK(ksobj);
606 	pmap_qenter(td->td_kstack, ma, pages);
607 	cpu_thread_swapin(td);
608 }
609 
610 void
611 faultin(struct proc *p)
612 {
613 	struct thread *td;
614 	int oom_alloc;
615 
616 	PROC_LOCK_ASSERT(p, MA_OWNED);
617 
618 	/*
619 	 * If another process is swapping in this process,
620 	 * just wait until it finishes.
621 	 */
622 	if (p->p_flag & P_SWAPPINGIN) {
623 		while (p->p_flag & P_SWAPPINGIN)
624 			msleep(&p->p_flag, &p->p_mtx, PVM, "faultin", 0);
625 		return;
626 	}
627 
628 	if ((p->p_flag & P_INMEM) == 0) {
629 		oom_alloc = (p->p_flag & P_WKILLED) != 0 ? VM_ALLOC_SYSTEM :
630 		    VM_ALLOC_NORMAL;
631 
632 		/*
633 		 * Don't let another thread swap process p out while we are
634 		 * busy swapping it in.
635 		 */
636 		++p->p_lock;
637 		p->p_flag |= P_SWAPPINGIN;
638 		PROC_UNLOCK(p);
639 		sx_xlock(&allproc_lock);
640 		MPASS(swapped_cnt > 0);
641 		swapped_cnt--;
642 		if (curthread != &thread0)
643 			swap_inprogress++;
644 		sx_xunlock(&allproc_lock);
645 
646 		/*
647 		 * We hold no lock here because the list of threads
648 		 * can not change while all threads in the process are
649 		 * swapped out.
650 		 */
651 		FOREACH_THREAD_IN_PROC(p, td)
652 			vm_thread_swapin(td, oom_alloc);
653 
654 		if (curthread != &thread0) {
655 			sx_xlock(&allproc_lock);
656 			MPASS(swap_inprogress > 0);
657 			swap_inprogress--;
658 			last_swapin = ticks;
659 			sx_xunlock(&allproc_lock);
660 		}
661 		PROC_LOCK(p);
662 		swapclear(p);
663 		p->p_swtick = ticks;
664 
665 		/* Allow other threads to swap p out now. */
666 		wakeup(&p->p_flag);
667 		--p->p_lock;
668 	}
669 }
670 
671 /*
672  * This swapin algorithm attempts to swap-in processes only if there
673  * is enough space for them.  Of course, if a process waits for a long
674  * time, it will be swapped in anyway.
675  */
676 
677 static struct proc *
678 swapper_selector(bool wkilled_only)
679 {
680 	struct proc *p, *res;
681 	struct thread *td;
682 	int ppri, pri, slptime, swtime;
683 
684 	sx_assert(&allproc_lock, SA_SLOCKED);
685 	if (swapped_cnt == 0)
686 		return (NULL);
687 	res = NULL;
688 	ppri = INT_MIN;
689 	FOREACH_PROC_IN_SYSTEM(p) {
690 		PROC_LOCK(p);
691 		if (p->p_state == PRS_NEW || (p->p_flag & (P_SWAPPINGOUT |
692 		    P_SWAPPINGIN | P_INMEM)) != 0) {
693 			PROC_UNLOCK(p);
694 			continue;
695 		}
696 		if (p->p_state == PRS_NORMAL && (p->p_flag & P_WKILLED) != 0) {
697 			/*
698 			 * A swapped-out process might have mapped a
699 			 * large portion of the system's pages as
700 			 * anonymous memory.  There is no other way to
701 			 * release the memory other than to kill the
702 			 * process, for which we need to swap it in.
703 			 */
704 			return (p);
705 		}
706 		if (wkilled_only) {
707 			PROC_UNLOCK(p);
708 			continue;
709 		}
710 		swtime = (ticks - p->p_swtick) / hz;
711 		FOREACH_THREAD_IN_PROC(p, td) {
712 			/*
713 			 * An otherwise runnable thread of a process
714 			 * swapped out has only the TDI_SWAPPED bit set.
715 			 */
716 			thread_lock(td);
717 			if (td->td_inhibitors == TDI_SWAPPED) {
718 				slptime = (ticks - td->td_slptick) / hz;
719 				pri = swtime + slptime;
720 				if ((td->td_flags & TDF_SWAPINREQ) == 0)
721 					pri -= p->p_nice * 8;
722 				/*
723 				 * if this thread is higher priority
724 				 * and there is enough space, then select
725 				 * this process instead of the previous
726 				 * selection.
727 				 */
728 				if (pri > ppri) {
729 					res = p;
730 					ppri = pri;
731 				}
732 			}
733 			thread_unlock(td);
734 		}
735 		PROC_UNLOCK(p);
736 	}
737 
738 	if (res != NULL)
739 		PROC_LOCK(res);
740 	return (res);
741 }
742 
743 #define	SWAPIN_INTERVAL	(MAXSLP * hz / 2)
744 
745 /*
746  * Limit swapper to swap in one non-WKILLED process in MAXSLP/2
747  * interval, assuming that there is:
748  * - at least one domain that is not suffering from a shortage of free memory;
749  * - no parallel swap-ins;
750  * - no other swap-ins in the current SWAPIN_INTERVAL.
751  */
752 static bool
753 swapper_wkilled_only(void)
754 {
755 
756 	return (vm_page_count_min_set(&all_domains) || swap_inprogress > 0 ||
757 	    (u_int)(ticks - last_swapin) < SWAPIN_INTERVAL);
758 }
759 
760 void
761 swapper(void)
762 {
763 	struct proc *p;
764 
765 	for (;;) {
766 		sx_slock(&allproc_lock);
767 		p = swapper_selector(swapper_wkilled_only());
768 		sx_sunlock(&allproc_lock);
769 
770 		if (p == NULL) {
771 			tsleep(&proc0, PVM, "swapin", SWAPIN_INTERVAL);
772 		} else {
773 			PROC_LOCK_ASSERT(p, MA_OWNED);
774 
775 			/*
776 			 * Another process may be bringing or may have
777 			 * already brought this process in while we
778 			 * traverse all threads.  Or, this process may
779 			 * have exited or even being swapped out
780 			 * again.
781 			 */
782 			if (p->p_state == PRS_NORMAL && (p->p_flag & (P_INMEM |
783 			    P_SWAPPINGOUT | P_SWAPPINGIN)) == 0) {
784 				faultin(p);
785 			}
786 			PROC_UNLOCK(p);
787 		}
788 	}
789 }
790 
791 /*
792  * First, if any processes have been sleeping or stopped for at least
793  * "swap_idle_threshold1" seconds, they are swapped out.  If, however,
794  * no such processes exist, then the longest-sleeping or stopped
795  * process is swapped out.  Finally, and only as a last resort, if
796  * there are no sleeping or stopped processes, the longest-resident
797  * process is swapped out.
798  */
799 static void
800 swapout_procs(int action)
801 {
802 	struct proc *p;
803 	struct thread *td;
804 	int slptime;
805 	bool didswap, doswap;
806 
807 	MPASS((action & (VM_SWAP_NORMAL | VM_SWAP_IDLE)) != 0);
808 
809 	didswap = false;
810 	sx_slock(&allproc_lock);
811 	FOREACH_PROC_IN_SYSTEM(p) {
812 		/*
813 		 * Filter out not yet fully constructed processes.  Do
814 		 * not swap out held processes.  Avoid processes which
815 		 * are system, exiting, execing, traced, already swapped
816 		 * out or are in the process of being swapped in or out.
817 		 */
818 		PROC_LOCK(p);
819 		if (p->p_state != PRS_NORMAL || p->p_lock != 0 || (p->p_flag &
820 		    (P_SYSTEM | P_WEXIT | P_INEXEC | P_STOPPED_SINGLE |
821 		    P_TRACED | P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) !=
822 		    P_INMEM) {
823 			PROC_UNLOCK(p);
824 			continue;
825 		}
826 
827 		/*
828 		 * Further consideration of this process for swap out
829 		 * requires iterating over its threads.  We release
830 		 * allproc_lock here so that process creation and
831 		 * destruction are not blocked while we iterate.
832 		 *
833 		 * To later reacquire allproc_lock and resume
834 		 * iteration over the allproc list, we will first have
835 		 * to release the lock on the process.  We place a
836 		 * hold on the process so that it remains in the
837 		 * allproc list while it is unlocked.
838 		 */
839 		_PHOLD_LITE(p);
840 		sx_sunlock(&allproc_lock);
841 
842 		/*
843 		 * Do not swapout a realtime process.
844 		 * Guarantee swap_idle_threshold1 time in memory.
845 		 * If the system is under memory stress, or if we are
846 		 * swapping idle processes >= swap_idle_threshold2,
847 		 * then swap the process out.
848 		 */
849 		doswap = true;
850 		FOREACH_THREAD_IN_PROC(p, td) {
851 			thread_lock(td);
852 			slptime = (ticks - td->td_slptick) / hz;
853 			if (PRI_IS_REALTIME(td->td_pri_class) ||
854 			    slptime < swap_idle_threshold1 ||
855 			    !thread_safetoswapout(td) ||
856 			    ((action & VM_SWAP_NORMAL) == 0 &&
857 			    slptime < swap_idle_threshold2))
858 				doswap = false;
859 			thread_unlock(td);
860 			if (!doswap)
861 				break;
862 		}
863 		if (doswap && swapout(p) == 0)
864 			didswap = true;
865 
866 		PROC_UNLOCK(p);
867 		if (didswap) {
868 			sx_xlock(&allproc_lock);
869 			swapped_cnt++;
870 			sx_downgrade(&allproc_lock);
871 		} else
872 			sx_slock(&allproc_lock);
873 		PRELE(p);
874 	}
875 	sx_sunlock(&allproc_lock);
876 
877 	/*
878 	 * If we swapped something out, and another process needed memory,
879 	 * then wakeup the sched process.
880 	 */
881 	if (didswap)
882 		wakeup(&proc0);
883 }
884 
885 static void
886 swapclear(struct proc *p)
887 {
888 	struct thread *td;
889 
890 	PROC_LOCK_ASSERT(p, MA_OWNED);
891 
892 	FOREACH_THREAD_IN_PROC(p, td) {
893 		thread_lock(td);
894 		td->td_flags |= TDF_INMEM;
895 		td->td_flags &= ~TDF_SWAPINREQ;
896 		TD_CLR_SWAPPED(td);
897 		if (TD_CAN_RUN(td))
898 			if (setrunnable(td)) {
899 #ifdef INVARIANTS
900 				/*
901 				 * XXX: We just cleared TDI_SWAPPED
902 				 * above and set TDF_INMEM, so this
903 				 * should never happen.
904 				 */
905 				panic("not waking up swapper");
906 #endif
907 			}
908 		thread_unlock(td);
909 	}
910 	p->p_flag &= ~(P_SWAPPINGIN | P_SWAPPINGOUT);
911 	p->p_flag |= P_INMEM;
912 }
913 
914 static int
915 swapout(struct proc *p)
916 {
917 	struct thread *td;
918 
919 	PROC_LOCK_ASSERT(p, MA_OWNED);
920 
921 	/*
922 	 * The states of this process and its threads may have changed
923 	 * by now.  Assuming that there is only one pageout daemon thread,
924 	 * this process should still be in memory.
925 	 */
926 	KASSERT((p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) ==
927 	    P_INMEM, ("swapout: lost a swapout race?"));
928 
929 	/*
930 	 * Remember the resident count.
931 	 */
932 	p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace);
933 
934 	/*
935 	 * Check and mark all threads before we proceed.
936 	 */
937 	p->p_flag &= ~P_INMEM;
938 	p->p_flag |= P_SWAPPINGOUT;
939 	FOREACH_THREAD_IN_PROC(p, td) {
940 		thread_lock(td);
941 		if (!thread_safetoswapout(td)) {
942 			thread_unlock(td);
943 			swapclear(p);
944 			return (EBUSY);
945 		}
946 		td->td_flags &= ~TDF_INMEM;
947 		TD_SET_SWAPPED(td);
948 		thread_unlock(td);
949 	}
950 	td = FIRST_THREAD_IN_PROC(p);
951 	++td->td_ru.ru_nswap;
952 	PROC_UNLOCK(p);
953 
954 	/*
955 	 * This list is stable because all threads are now prevented from
956 	 * running.  The list is only modified in the context of a running
957 	 * thread in this process.
958 	 */
959 	FOREACH_THREAD_IN_PROC(p, td)
960 		vm_thread_swapout(td);
961 
962 	PROC_LOCK(p);
963 	p->p_flag &= ~P_SWAPPINGOUT;
964 	p->p_swtick = ticks;
965 	return (0);
966 }
967