xref: /freebsd/sys/vm/vm_swapout.c (revision 1b5e2d70766989a593709debbf46837a415d168f)
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  *
45  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
46  * All rights reserved.
47  *
48  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
49  *
50  * Permission to use, copy, modify and distribute this software and
51  * its documentation is hereby granted, provided that both the copyright
52  * notice and this permission notice appear in all copies of the
53  * software, derivative works or modified versions, and any portions
54  * thereof, and that both notices appear in supporting documentation.
55  *
56  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
57  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
58  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
59  *
60  * Carnegie Mellon requests users of this software to return to
61  *
62  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
63  *  School of Computer Science
64  *  Carnegie Mellon University
65  *  Pittsburgh PA 15213-3890
66  *
67  * any improvements or extensions that they make and grant Carnegie the
68  * rights to redistribute these changes.
69  */
70 
71 #include "opt_kstack_pages.h"
72 #include "opt_kstack_max_pages.h"
73 #include "opt_vm.h"
74 
75 #include <sys/param.h>
76 #include <sys/systm.h>
77 #include <sys/limits.h>
78 #include <sys/kernel.h>
79 #include <sys/eventhandler.h>
80 #include <sys/lock.h>
81 #include <sys/mutex.h>
82 #include <sys/proc.h>
83 #include <sys/kthread.h>
84 #include <sys/ktr.h>
85 #include <sys/mount.h>
86 #include <sys/racct.h>
87 #include <sys/resourcevar.h>
88 #include <sys/refcount.h>
89 #include <sys/sched.h>
90 #include <sys/sdt.h>
91 #include <sys/signalvar.h>
92 #include <sys/smp.h>
93 #include <sys/time.h>
94 #include <sys/vnode.h>
95 #include <sys/vmmeter.h>
96 #include <sys/rwlock.h>
97 #include <sys/sx.h>
98 #include <sys/sysctl.h>
99 
100 #include <vm/vm.h>
101 #include <vm/vm_param.h>
102 #include <vm/vm_kern.h>
103 #include <vm/vm_object.h>
104 #include <vm/vm_page.h>
105 #include <vm/vm_map.h>
106 #include <vm/vm_pageout.h>
107 #include <vm/vm_pager.h>
108 #include <vm/vm_phys.h>
109 #include <vm/vm_radix.h>
110 #include <vm/swap_pager.h>
111 #include <vm/vm_extern.h>
112 #include <vm/uma.h>
113 
114 /* the kernel process "vm_daemon" */
115 static void vm_daemon(void);
116 static struct proc *vmproc;
117 
118 static struct kproc_desc vm_kp = {
119 	"vmdaemon",
120 	vm_daemon,
121 	&vmproc
122 };
123 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
124 
125 static int vm_daemon_timeout = 0;
126 SYSCTL_INT(_vm, OID_AUTO, vmdaemon_timeout, CTLFLAG_RW,
127     &vm_daemon_timeout, 0,
128     "Time between vmdaemon runs");
129 
130 static int vm_daemon_needed;
131 static struct mtx vm_daemon_mtx;
132 /* Allow for use by vm_pageout before vm_daemon is initialized. */
133 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
134 
135 static void vm_swapout_map_deactivate_pages(vm_map_t, long);
136 static void vm_swapout_object_deactivate(pmap_t, vm_object_t, long);
137 
138 static void
vm_swapout_object_deactivate_page(pmap_t pmap,vm_page_t m,bool unmap)139 vm_swapout_object_deactivate_page(pmap_t pmap, vm_page_t m, bool unmap)
140 {
141 
142 	/*
143 	 * Ignore unreclaimable wired pages.  Repeat the check after busying
144 	 * since a busy holder may wire the page.
145 	 */
146 	if (vm_page_wired(m) || !vm_page_tryxbusy(m))
147 		return;
148 
149 	if (vm_page_wired(m) || !pmap_page_exists_quick(pmap, m)) {
150 		vm_page_xunbusy(m);
151 		return;
152 	}
153 	if (!pmap_is_referenced(m)) {
154 		if (!vm_page_active(m))
155 			(void)vm_page_try_remove_all(m);
156 		else if (unmap && vm_page_try_remove_all(m))
157 			vm_page_deactivate(m);
158 	}
159 	vm_page_xunbusy(m);
160 }
161 
162 /*
163  *	vm_swapout_object_deactivate
164  *
165  *	Deactivate enough pages to satisfy the inactive target
166  *	requirements.
167  *
168  *	The object and map must be locked.
169  */
170 static void
vm_swapout_object_deactivate(pmap_t pmap,vm_object_t first_object,long desired)171 vm_swapout_object_deactivate(pmap_t pmap, vm_object_t first_object,
172     long desired)
173 {
174 	struct pctrie_iter pages;
175 	vm_object_t backing_object, object;
176 	vm_page_t m;
177 	bool unmap;
178 
179 	VM_OBJECT_ASSERT_LOCKED(first_object);
180 	if ((first_object->flags & OBJ_FICTITIOUS) != 0)
181 		return;
182 	for (object = first_object;; object = backing_object) {
183 		if (pmap_resident_count(pmap) <= desired)
184 			goto unlock_return;
185 		VM_OBJECT_ASSERT_LOCKED(object);
186 		if ((object->flags & OBJ_UNMANAGED) != 0 ||
187 		    blockcount_read(&object->paging_in_progress) > 0)
188 			goto unlock_return;
189 
190 		unmap = true;
191 		if (object->shadow_count > 1)
192 			unmap = false;
193 
194 		/*
195 		 * Scan the object's entire memory queue.
196 		 */
197 		vm_page_iter_init(&pages, object);
198 		VM_RADIX_FOREACH(m, &pages) {
199 			if (pmap_resident_count(pmap) <= desired)
200 				goto unlock_return;
201 			if (should_yield())
202 				goto unlock_return;
203 			vm_swapout_object_deactivate_page(pmap, m, unmap);
204 		}
205 		if ((backing_object = object->backing_object) == NULL)
206 			goto unlock_return;
207 		VM_OBJECT_RLOCK(backing_object);
208 		if (object != first_object)
209 			VM_OBJECT_RUNLOCK(object);
210 	}
211 unlock_return:
212 	if (object != first_object)
213 		VM_OBJECT_RUNLOCK(object);
214 }
215 
216 /*
217  * deactivate some number of pages in a map, try to do it fairly, but
218  * that is really hard to do.
219  */
220 static void
vm_swapout_map_deactivate_pages(vm_map_t map,long desired)221 vm_swapout_map_deactivate_pages(vm_map_t map, long desired)
222 {
223 	vm_map_entry_t tmpe;
224 	vm_object_t obj, bigobj;
225 	int nothingwired;
226 
227 	if (!vm_map_trylock_read(map))
228 		return;
229 
230 	bigobj = NULL;
231 	nothingwired = TRUE;
232 
233 	/*
234 	 * first, search out the biggest object, and try to free pages from
235 	 * that.
236 	 */
237 	VM_MAP_ENTRY_FOREACH(tmpe, map) {
238 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
239 			obj = tmpe->object.vm_object;
240 			if (obj != NULL && VM_OBJECT_TRYRLOCK(obj)) {
241 				if (obj->shadow_count <= 1 &&
242 				    (bigobj == NULL ||
243 				     bigobj->resident_page_count <
244 				     obj->resident_page_count)) {
245 					if (bigobj != NULL)
246 						VM_OBJECT_RUNLOCK(bigobj);
247 					bigobj = obj;
248 				} else
249 					VM_OBJECT_RUNLOCK(obj);
250 			}
251 		}
252 		if (tmpe->wired_count > 0)
253 			nothingwired = FALSE;
254 	}
255 
256 	if (bigobj != NULL) {
257 		vm_swapout_object_deactivate(map->pmap, bigobj, desired);
258 		VM_OBJECT_RUNLOCK(bigobj);
259 	}
260 	/*
261 	 * Next, hunt around for other pages to deactivate.  We actually
262 	 * do this search sort of wrong -- .text first is not the best idea.
263 	 */
264 	VM_MAP_ENTRY_FOREACH(tmpe, map) {
265 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
266 			break;
267 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
268 			obj = tmpe->object.vm_object;
269 			if (obj != NULL) {
270 				VM_OBJECT_RLOCK(obj);
271 				vm_swapout_object_deactivate(map->pmap, obj,
272 				    desired);
273 				VM_OBJECT_RUNLOCK(obj);
274 			}
275 		}
276 	}
277 
278 	/*
279 	 * Remove all mappings if a process is swapped out, this will free page
280 	 * table pages.
281 	 */
282 	if (desired == 0 && nothingwired) {
283 		pmap_remove(vm_map_pmap(map), vm_map_min(map),
284 		    vm_map_max(map));
285 	}
286 
287 	vm_map_unlock_read(map);
288 }
289 
290 static void
vm_daemon(void)291 vm_daemon(void)
292 {
293 	struct rlimit rsslim;
294 	struct proc *p;
295 	struct thread *td;
296 	struct vmspace *vm;
297 	int breakout, tryagain, attempts;
298 	uint64_t rsize, ravailable;
299 
300 	if (racct_enable && vm_daemon_timeout == 0)
301 		vm_daemon_timeout = hz;
302 
303 	while (TRUE) {
304 		mtx_lock(&vm_daemon_mtx);
305 		msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep",
306 		    vm_daemon_timeout);
307 		mtx_unlock(&vm_daemon_mtx);
308 
309 		/*
310 		 * scan the processes for exceeding their rlimits or if
311 		 * process is swapped out -- deactivate pages
312 		 */
313 		tryagain = 0;
314 		attempts = 0;
315 again:
316 		attempts++;
317 		sx_slock(&allproc_lock);
318 		FOREACH_PROC_IN_SYSTEM(p) {
319 			vm_pindex_t limit, size;
320 
321 			/*
322 			 * if this is a system process or if we have already
323 			 * looked at this process, skip it.
324 			 */
325 			PROC_LOCK(p);
326 			if (p->p_state != PRS_NORMAL ||
327 			    p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
328 				PROC_UNLOCK(p);
329 				continue;
330 			}
331 			/*
332 			 * if the process is in a non-running type state,
333 			 * don't touch it.
334 			 */
335 			breakout = 0;
336 			FOREACH_THREAD_IN_PROC(p, td) {
337 				thread_lock(td);
338 				if (!TD_ON_RUNQ(td) &&
339 				    !TD_IS_RUNNING(td) &&
340 				    !TD_IS_SLEEPING(td) &&
341 				    !TD_IS_SUSPENDED(td)) {
342 					thread_unlock(td);
343 					breakout = 1;
344 					break;
345 				}
346 				thread_unlock(td);
347 			}
348 			if (breakout) {
349 				PROC_UNLOCK(p);
350 				continue;
351 			}
352 			/*
353 			 * get a limit
354 			 */
355 			lim_rlimit_proc(p, RLIMIT_RSS, &rsslim);
356 			limit = OFF_TO_IDX(
357 			    qmin(rsslim.rlim_cur, rsslim.rlim_max));
358 
359 			vm = vmspace_acquire_ref(p);
360 			_PHOLD(p);
361 			PROC_UNLOCK(p);
362 			if (vm == NULL) {
363 				PRELE(p);
364 				continue;
365 			}
366 			sx_sunlock(&allproc_lock);
367 
368 			size = vmspace_resident_count(vm);
369 			if (size >= limit) {
370 				vm_swapout_map_deactivate_pages(
371 				    &vm->vm_map, limit);
372 				size = vmspace_resident_count(vm);
373 			}
374 			if (racct_enable) {
375 				rsize = IDX_TO_OFF(size);
376 				PROC_LOCK(p);
377 				if (p->p_state == PRS_NORMAL)
378 					racct_set(p, RACCT_RSS, rsize);
379 				ravailable = racct_get_available(p, RACCT_RSS);
380 				PROC_UNLOCK(p);
381 				if (rsize > ravailable) {
382 					/*
383 					 * Don't be overly aggressive; this
384 					 * might be an innocent process,
385 					 * and the limit could've been exceeded
386 					 * by some memory hog.  Don't try
387 					 * to deactivate more than 1/4th
388 					 * of process' resident set size.
389 					 */
390 					if (attempts <= 8) {
391 						if (ravailable < rsize -
392 						    (rsize / 4)) {
393 							ravailable = rsize -
394 							    (rsize / 4);
395 						}
396 					}
397 					vm_swapout_map_deactivate_pages(
398 					    &vm->vm_map,
399 					    OFF_TO_IDX(ravailable));
400 					/* Update RSS usage after paging out. */
401 					size = vmspace_resident_count(vm);
402 					rsize = IDX_TO_OFF(size);
403 					PROC_LOCK(p);
404 					if (p->p_state == PRS_NORMAL)
405 						racct_set(p, RACCT_RSS, rsize);
406 					PROC_UNLOCK(p);
407 					if (rsize > ravailable)
408 						tryagain = 1;
409 				}
410 			}
411 			vmspace_free(vm);
412 			sx_slock(&allproc_lock);
413 			PRELE(p);
414 		}
415 		sx_sunlock(&allproc_lock);
416 		if (tryagain != 0 && attempts <= 10) {
417 			maybe_yield();
418 			goto again;
419 		}
420 	}
421 }
422