xref: /openbsd/sys/kern/vfs_sync.c (revision 771fbea0)
1 /*       $OpenBSD: vfs_sync.c,v 1.65 2021/01/14 03:32:01 cheloha Exp $  */
2 
3 /*
4  *  Portions of this code are:
5  *
6  * Copyright (c) 1989, 1993
7  *	The Regents of the University of California.  All rights reserved.
8  * (c) UNIX System Laboratories, Inc.
9  * All or some portions of this file are derived from material licensed
10  * to the University of California by American Telephone and Telegraph
11  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
12  * the permission of UNIX System Laboratories, Inc.
13  *
14  * Redistribution and use in source and binary forms, with or without
15  * modification, are permitted provided that the following conditions
16  * are met:
17  * 1. Redistributions of source code must retain the above copyright
18  *    notice, this list of conditions and the following disclaimer.
19  * 2. Redistributions in binary form must reproduce the above copyright
20  *    notice, this list of conditions and the following disclaimer in the
21  *    documentation and/or other materials provided with the distribution.
22  * 3. Neither the name of the University nor the names of its contributors
23  *    may be used to endorse or promote products derived from this software
24  *    without specific prior written permission.
25  *
26  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
27  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
28  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
29  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
30  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36  * SUCH DAMAGE.
37  */
38 
39 /*
40  * Syncer daemon
41  */
42 
43 #include <sys/queue.h>
44 #include <sys/param.h>
45 #include <sys/systm.h>
46 #include <sys/proc.h>
47 #include <sys/mount.h>
48 #include <sys/vnode.h>
49 #include <sys/lock.h>
50 #include <sys/malloc.h>
51 #include <sys/time.h>
52 
53 #include <sys/kernel.h>
54 #include <sys/sched.h>
55 
56 #ifdef FFS_SOFTUPDATES
57 int   softdep_process_worklist(struct mount *);
58 #endif
59 
60 /*
61  * The workitem queue.
62  */
63 #define SYNCER_MAXDELAY	32		/* maximum sync delay time */
64 #define SYNCER_DEFAULT 30		/* default sync delay time */
65 int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
66 int syncdelay = SYNCER_DEFAULT;		/* time to delay syncing vnodes */
67 
68 int rushjob = 0;			/* number of slots to run ASAP */
69 int stat_rush_requests = 0;		/* number of rush requests */
70 
71 int syncer_delayno = 0;
72 long syncer_mask;
73 LIST_HEAD(synclist, vnode);
74 static struct synclist *syncer_workitem_pending;
75 
76 struct proc *syncerproc;
77 int syncer_chan;
78 
79 /*
80  * The workitem queue.
81  *
82  * It is useful to delay writes of file data and filesystem metadata
83  * for tens of seconds so that quickly created and deleted files need
84  * not waste disk bandwidth being created and removed. To realize this,
85  * we append vnodes to a "workitem" queue. When running with a soft
86  * updates implementation, most pending metadata dependencies should
87  * not wait for more than a few seconds. Thus, mounted block devices
88  * are delayed only about half the time that file data is delayed.
89  * Similarly, directory updates are more critical, so are only delayed
90  * about a third the time that file data is delayed. Thus, there are
91  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
92  * one each second (driven off the filesystem syncer process). The
93  * syncer_delayno variable indicates the next queue that is to be processed.
94  * Items that need to be processed soon are placed in this queue:
95  *
96  *	syncer_workitem_pending[syncer_delayno]
97  *
98  * A delay of fifteen seconds is done by placing the request fifteen
99  * entries later in the queue:
100  *
101  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
102  *
103  */
104 
105 void
106 vn_initialize_syncerd(void)
107 {
108 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE, M_WAITOK,
109 	    &syncer_mask);
110 	syncer_maxdelay = syncer_mask + 1;
111 }
112 
113 /*
114  * Add an item to the syncer work queue.
115  */
116 void
117 vn_syncer_add_to_worklist(struct vnode *vp, int delay)
118 {
119 	int s, slot;
120 
121 	if (delay > syncer_maxdelay - 2)
122 		delay = syncer_maxdelay - 2;
123 	slot = (syncer_delayno + delay) & syncer_mask;
124 
125 	s = splbio();
126 	if (vp->v_bioflag & VBIOONSYNCLIST)
127 		LIST_REMOVE(vp, v_synclist);
128 
129 	vp->v_bioflag |= VBIOONSYNCLIST;
130 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], vp, v_synclist);
131 	splx(s);
132 }
133 
134 /*
135  * TODO Move getnsecuptime() to kern_tc.c and document it when we have
136  * more users in the kernel.
137  */
138 static uint64_t
139 getnsecuptime(void)
140 {
141 	struct timespec now;
142 
143 	getnanouptime(&now);
144 	return TIMESPEC_TO_NSEC(&now);
145 }
146 
147 /*
148  * System filesystem synchronizer daemon.
149  */
150 void
151 syncer_thread(void *arg)
152 {
153 	uint64_t elapsed, start;
154 	struct proc *p = curproc;
155 	struct synclist *slp;
156 	struct vnode *vp;
157 	int s;
158 
159 	for (;;) {
160 		start = getnsecuptime();
161 
162 		/*
163 		 * Push files whose dirty time has expired.
164 		 */
165 		s = splbio();
166 		slp = &syncer_workitem_pending[syncer_delayno];
167 
168 		syncer_delayno += 1;
169 		if (syncer_delayno == syncer_maxdelay)
170 			syncer_delayno = 0;
171 
172 		while ((vp = LIST_FIRST(slp)) != NULL) {
173 			if (vget(vp, LK_EXCLUSIVE | LK_NOWAIT)) {
174 				/*
175 				 * If we fail to get the lock, we move this
176 				 * vnode one second ahead in time.
177 				 * XXX - no good, but the best we can do.
178 				 */
179 				vn_syncer_add_to_worklist(vp, 1);
180 				continue;
181 			}
182 			splx(s);
183 			(void) VOP_FSYNC(vp, p->p_ucred, MNT_LAZY, p);
184 			vput(vp);
185 			s = splbio();
186 			if (LIST_FIRST(slp) == vp) {
187 				/*
188 				 * Note: disk vps can remain on the
189 				 * worklist too with no dirty blocks, but
190 				 * since sync_fsync() moves it to a different
191 				 * slot we are safe.
192 				 */
193 #ifdef DIAGNOSTIC
194 				if (LIST_FIRST(&vp->v_dirtyblkhd) == NULL &&
195 				    vp->v_type != VBLK) {
196 					vprint("fsync failed", vp);
197 					if (vp->v_mount != NULL)
198 						printf("mounted on: %s\n",
199 						    vp->v_mount->mnt_stat.f_mntonname);
200 					panic("%s: fsync failed", __func__);
201 				}
202 #endif /* DIAGNOSTIC */
203 				/*
204 				 * Put us back on the worklist.  The worklist
205 				 * routine will remove us from our current
206 				 * position and then add us back in at a later
207 				 * position.
208 				 */
209 				vn_syncer_add_to_worklist(vp, syncdelay);
210 			}
211 
212 			sched_pause(yield);
213 		}
214 
215 		splx(s);
216 
217 #ifdef FFS_SOFTUPDATES
218 		/*
219 		 * Do soft update processing.
220 		 */
221 		softdep_process_worklist(NULL);
222 #endif
223 
224 		/*
225 		 * The variable rushjob allows the kernel to speed up the
226 		 * processing of the filesystem syncer process. A rushjob
227 		 * value of N tells the filesystem syncer to process the next
228 		 * N seconds worth of work on its queue ASAP. Currently rushjob
229 		 * is used by the soft update code to speed up the filesystem
230 		 * syncer process when the incore state is getting so far
231 		 * ahead of the disk that the kernel memory pool is being
232 		 * threatened with exhaustion.
233 		 */
234 		if (rushjob > 0) {
235 			rushjob -= 1;
236 			continue;
237 		}
238 
239 		/*
240 		 * If it has taken us less than a second to process the
241 		 * current work, then wait. Otherwise start right over
242 		 * again. We can still lose time if any single round
243 		 * takes more than two seconds, but it does not really
244 		 * matter as we are just trying to generally pace the
245 		 * filesystem activity.
246 		 */
247 		elapsed = getnsecuptime() - start;
248 		if (elapsed < SEC_TO_NSEC(1)) {
249 			tsleep_nsec(&syncer_chan, PPAUSE, "syncer",
250 			    SEC_TO_NSEC(1) - elapsed);
251 		}
252 	}
253 }
254 
255 /*
256  * Request the syncer daemon to speed up its work.
257  * We never push it to speed up more than half of its
258  * normal turn time, otherwise it could take over the cpu.
259  */
260 int
261 speedup_syncer(void)
262 {
263 	if (syncerproc)
264 		wakeup_proc(syncerproc, &syncer_chan);
265 	if (rushjob < syncdelay / 2) {
266 		rushjob += 1;
267 		stat_rush_requests += 1;
268 		return 1;
269 	}
270 	return 0;
271 }
272 
273 /* Routine to create and manage a filesystem syncer vnode. */
274 int   sync_fsync(void *);
275 int   sync_inactive(void *);
276 int   sync_print(void *);
277 
278 const struct vops sync_vops = {
279 	.vop_close	= nullop,
280 	.vop_fsync	= sync_fsync,
281 	.vop_inactive	= sync_inactive,
282 	.vop_reclaim	= nullop,
283 	.vop_lock	= vop_generic_lock,
284 	.vop_unlock	= vop_generic_unlock,
285 	.vop_islocked	= vop_generic_islocked,
286 	.vop_print	= sync_print
287 };
288 
289 /*
290  * Create a new filesystem syncer vnode for the specified mount point.
291  */
292 int
293 vfs_allocate_syncvnode(struct mount *mp)
294 {
295 	struct vnode *vp;
296 	static long start, incr, next;
297 	int error;
298 
299 	/* Allocate a new vnode */
300 	if ((error = getnewvnode(VT_VFS, mp, &sync_vops, &vp)) != 0) {
301 		mp->mnt_syncer = NULL;
302 		return (error);
303 	}
304 	vp->v_writecount = 1;
305 	vp->v_type = VNON;
306 	/*
307 	 * Place the vnode onto the syncer worklist. We attempt to
308 	 * scatter them about on the list so that they will go off
309 	 * at evenly distributed times even if all the filesystems
310 	 * are mounted at once.
311 	 */
312 	next += incr;
313 	if (next == 0 || next > syncer_maxdelay) {
314 		start /= 2;
315 		incr /= 2;
316 		if (start == 0) {
317 			start = syncer_maxdelay / 2;
318 			incr = syncer_maxdelay;
319 		}
320 		next = start;
321 	}
322 	vn_syncer_add_to_worklist(vp, next);
323 	mp->mnt_syncer = vp;
324 	return (0);
325 }
326 
327 /*
328  * Do a lazy sync of the filesystem.
329  */
330 int
331 sync_fsync(void *v)
332 {
333 	struct vop_fsync_args *ap = v;
334 	struct vnode *syncvp = ap->a_vp;
335 	struct mount *mp = syncvp->v_mount;
336 	int asyncflag;
337 
338 	/*
339 	 * We only need to do something if this is a lazy evaluation.
340 	 */
341 	if (ap->a_waitfor != MNT_LAZY)
342 		return (0);
343 
344 	/*
345 	 * Move ourselves to the back of the sync list.
346 	 */
347 	vn_syncer_add_to_worklist(syncvp, syncdelay);
348 
349 	/*
350 	 * Walk the list of vnodes pushing all that are dirty and
351 	 * not already on the sync list.
352 	 */
353 	if (vfs_busy(mp, VB_READ|VB_NOWAIT) == 0) {
354 		asyncflag = mp->mnt_flag & MNT_ASYNC;
355 		mp->mnt_flag &= ~MNT_ASYNC;
356 		VFS_SYNC(mp, MNT_LAZY, 0, ap->a_cred, ap->a_p);
357 		if (asyncflag)
358 			mp->mnt_flag |= MNT_ASYNC;
359 		vfs_unbusy(mp);
360 	}
361 
362 	return (0);
363 }
364 
365 /*
366  * The syncer vnode is no longer needed and is being decommissioned.
367  */
368 int
369 sync_inactive(void *v)
370 {
371 	struct vop_inactive_args *ap = v;
372 
373 	struct vnode *vp = ap->a_vp;
374 	int s;
375 
376 	if (vp->v_usecount == 0) {
377 		VOP_UNLOCK(vp);
378 		return (0);
379 	}
380 
381 	vp->v_mount->mnt_syncer = NULL;
382 
383 	s = splbio();
384 
385 	LIST_REMOVE(vp, v_synclist);
386 	vp->v_bioflag &= ~VBIOONSYNCLIST;
387 
388 	splx(s);
389 
390 	vp->v_writecount = 0;
391 	vput(vp);
392 
393 	return (0);
394 }
395 
396 /*
397  * Print out a syncer vnode.
398  */
399 int
400 sync_print(void *v)
401 {
402 	printf("syncer vnode\n");
403 
404 	return (0);
405 }
406