xref: /netbsd-src/sys/fs/nfs/client/nfs_clnfsiod.c (revision 3be2222bbe56192eb75b39905cef98ac451a3d70)
1 /*	$NetBSD: nfs_clnfsiod.c,v 1.2 2016/12/13 22:17:33 pgoyette Exp $	*/
2 /*-
3  * Copyright (c) 1989, 1993
4  *	The Regents of the University of California.  All rights reserved.
5  *
6  * This code is derived from software contributed to Berkeley by
7  * Rick Macklem at The University of Guelph.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  * 4. Neither the name of the University nor the names of its contributors
18  *    may be used to endorse or promote products derived from this software
19  *    without specific prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  *
33  *	from nfs_syscalls.c	8.5 (Berkeley) 3/30/95
34  */
35 
36 #include <sys/cdefs.h>
37 /* __FBSDID("FreeBSD: head/sys/fs/nfsclient/nfs_clnfsiod.c 249630 2013-04-18 23:20:16Z rmacklem "); */
38 __RCSID("$NetBSD: nfs_clnfsiod.c,v 1.2 2016/12/13 22:17:33 pgoyette Exp $");
39 
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/sysproto.h>
43 #include <sys/kernel.h>
44 #include <sys/sysctl.h>
45 #include <sys/file.h>
46 #include <sys/filedesc.h>
47 #include <sys/vnode.h>
48 #include <sys/malloc.h>
49 #include <sys/mount.h>
50 #include <sys/proc.h>
51 #include <sys/bio.h>
52 #include <sys/buf.h>
53 #include <sys/mbuf.h>
54 #include <sys/socket.h>
55 #include <sys/socketvar.h>
56 #include <sys/domain.h>
57 #include <sys/protosw.h>
58 #include <sys/namei.h>
59 #include <sys/unistd.h>
60 #include <sys/kthread.h>
61 #include <sys/fcntl.h>
62 #include <sys/lockf.h>
63 #include <sys/mutex.h>
64 #include <sys/taskqueue.h>
65 
66 #include <netinet/in.h>
67 #include <netinet/tcp.h>
68 
69 #include <fs/nfs/common/nfsport.h>
70 #include <fs/nfs/client/nfsmount.h>
71 #include <fs/nfs/client/nfs.h>
72 #include <fs/nfs/client/nfsnode.h>
73 
74 extern struct mtx	ncl_iod_mutex;
75 extern struct task	ncl_nfsiodnew_task;
76 
77 int ncl_numasync;
78 enum nfsiod_state ncl_iodwant[NFS_MAXASYNCDAEMON];
79 struct nfsmount *ncl_iodmount[NFS_MAXASYNCDAEMON];
80 
81 static void	nfssvc_iod(void *);
82 
83 static int nfs_asyncdaemon[NFS_MAXASYNCDAEMON];
84 
85 SYSCTL_DECL(_vfs_nfs);
86 
87 /* Maximum number of seconds a nfsiod kthread will sleep before exiting */
88 static unsigned int nfs_iodmaxidle = 120;
89 SYSCTL_UINT(_vfs_nfs, OID_AUTO, iodmaxidle, CTLFLAG_RW, &nfs_iodmaxidle, 0,
90     "Max number of seconds an nfsiod kthread will sleep before exiting");
91 
92 /* Maximum number of nfsiod kthreads */
93 unsigned int ncl_iodmax = 20;
94 
95 /* Minimum number of nfsiod kthreads to keep as spares */
96 static unsigned int nfs_iodmin = 0;
97 
98 static int nfs_nfsiodnew_sync(void);
99 
100 static int
sysctl_iodmin(SYSCTL_HANDLER_ARGS)101 sysctl_iodmin(SYSCTL_HANDLER_ARGS)
102 {
103 	int error, i;
104 	int newmin;
105 
106 	newmin = nfs_iodmin;
107 	error = sysctl_handle_int(oidp, &newmin, 0, req);
108 	if (error || (req->newptr == NULL))
109 		return (error);
110 	mtx_lock(&ncl_iod_mutex);
111 	if (newmin > ncl_iodmax) {
112 		error = EINVAL;
113 		goto out;
114 	}
115 	nfs_iodmin = newmin;
116 	if (ncl_numasync >= nfs_iodmin)
117 		goto out;
118 	/*
119 	 * If the current number of nfsiod is lower
120 	 * than the new minimum, create some more.
121 	 */
122 	for (i = nfs_iodmin - ncl_numasync; i > 0; i--)
123 		nfs_nfsiodnew_sync();
124 out:
125 	mtx_unlock(&ncl_iod_mutex);
126 	return (0);
127 }
128 SYSCTL_PROC(_vfs_nfs, OID_AUTO, iodmin, CTLTYPE_UINT | CTLFLAG_RW, 0,
129     sizeof (nfs_iodmin), sysctl_iodmin, "IU",
130     "Min number of nfsiod kthreads to keep as spares");
131 
132 static int
sysctl_iodmax(SYSCTL_HANDLER_ARGS)133 sysctl_iodmax(SYSCTL_HANDLER_ARGS)
134 {
135 	int error, i;
136 	int iod, newmax;
137 
138 	newmax = ncl_iodmax;
139 	error = sysctl_handle_int(oidp, &newmax, 0, req);
140 	if (error || (req->newptr == NULL))
141 		return (error);
142 	if (newmax > NFS_MAXASYNCDAEMON)
143 		return (EINVAL);
144 	mtx_lock(&ncl_iod_mutex);
145 	ncl_iodmax = newmax;
146 	if (ncl_numasync <= ncl_iodmax)
147 		goto out;
148 	/*
149 	 * If there are some asleep nfsiods that should
150 	 * exit, wakeup() them so that they check ncl_iodmax
151 	 * and exit.  Those who are active will exit as
152 	 * soon as they finish I/O.
153 	 */
154 	iod = ncl_numasync - 1;
155 	for (i = 0; i < ncl_numasync - ncl_iodmax; i++) {
156 		if (ncl_iodwant[iod] == NFSIOD_AVAILABLE)
157 			wakeup(&ncl_iodwant[iod]);
158 		iod--;
159 	}
160 out:
161 	mtx_unlock(&ncl_iod_mutex);
162 	return (0);
163 }
164 SYSCTL_PROC(_vfs_nfs, OID_AUTO, iodmax, CTLTYPE_UINT | CTLFLAG_RW, 0,
165     sizeof (ncl_iodmax), sysctl_iodmax, "IU",
166     "Max number of nfsiod kthreads");
167 
168 static int
nfs_nfsiodnew_sync(void)169 nfs_nfsiodnew_sync(void)
170 {
171 	int error, i;
172 
173 	mtx_assert(&ncl_iod_mutex, MA_OWNED);
174 	for (i = 0; i < ncl_iodmax; i++) {
175 		if (nfs_asyncdaemon[i] == 0) {
176 			nfs_asyncdaemon[i] = 1;
177 			break;
178 		}
179 	}
180 	if (i == ncl_iodmax)
181 		return (0);
182 	mtx_unlock(&ncl_iod_mutex);
183 	error = kproc_create(nfssvc_iod, nfs_asyncdaemon + i, NULL,
184 	    RFHIGHPID, 0, "newnfs %d", i);
185 	mtx_lock(&ncl_iod_mutex);
186 	if (error == 0) {
187 		ncl_numasync++;
188 		ncl_iodwant[i] = NFSIOD_AVAILABLE;
189 	} else
190 		nfs_asyncdaemon[i] = 0;
191 	return (error);
192 }
193 
194 void
ncl_nfsiodnew_tq(__unused void * arg,int pending)195 ncl_nfsiodnew_tq(__unused void *arg, int pending)
196 {
197 
198 	mtx_lock(&ncl_iod_mutex);
199 	while (pending > 0) {
200 		pending--;
201 		nfs_nfsiodnew_sync();
202 	}
203 	mtx_unlock(&ncl_iod_mutex);
204 }
205 
206 void
ncl_nfsiodnew(void)207 ncl_nfsiodnew(void)
208 {
209 
210 	mtx_assert(&ncl_iod_mutex, MA_OWNED);
211 	taskqueue_enqueue(taskqueue_thread, &ncl_nfsiodnew_task);
212 }
213 
214 static void
nfsiod_setup(void * dummy)215 nfsiod_setup(void *dummy)
216 {
217 	int error;
218 
219 	TUNABLE_INT_FETCH("vfs.nfs.iodmin", &nfs_iodmin);
220 	nfscl_init();
221 	mtx_lock(&ncl_iod_mutex);
222 	/* Silently limit the start number of nfsiod's */
223 	if (nfs_iodmin > NFS_MAXASYNCDAEMON)
224 		nfs_iodmin = NFS_MAXASYNCDAEMON;
225 
226 	while (ncl_numasync < nfs_iodmin) {
227 		error = nfs_nfsiodnew_sync();
228 		if (error == -1)
229 			panic("nfsiod_setup: nfs_nfsiodnew failed");
230 	}
231 	mtx_unlock(&ncl_iod_mutex);
232 }
233 SYSINIT(newnfsiod, SI_SUB_KTHREAD_IDLE, SI_ORDER_ANY, nfsiod_setup, NULL);
234 
235 static int nfs_defect = 0;
236 SYSCTL_INT(_vfs_nfs, OID_AUTO, defect, CTLFLAG_RW, &nfs_defect, 0,
237     "Allow nfsiods to migrate serving different mounts");
238 
239 /*
240  * Asynchronous I/O daemons for client nfs.
241  * They do read-ahead and write-behind operations on the block I/O cache.
242  * Returns if we hit the timeout defined by the iodmaxidle sysctl.
243  */
244 static void
nfssvc_iod(void * instance)245 nfssvc_iod(void *instance)
246 {
247 	struct buf *bp;
248 	struct nfsmount *nmp;
249 	int myiod, timo;
250 	int error = 0;
251 
252 	mtx_lock(&ncl_iod_mutex);
253 	myiod = (int *)instance - nfs_asyncdaemon;
254 	/*
255 	 * Main loop
256 	 */
257 	for (;;) {
258 	    while (((nmp = ncl_iodmount[myiod]) == NULL)
259 		   || !TAILQ_FIRST(&nmp->nm_bufq)) {
260 		if (myiod >= ncl_iodmax)
261 			goto finish;
262 		if (nmp)
263 			nmp->nm_bufqiods--;
264 		if (ncl_iodwant[myiod] == NFSIOD_NOT_AVAILABLE)
265 			ncl_iodwant[myiod] = NFSIOD_AVAILABLE;
266 		ncl_iodmount[myiod] = NULL;
267 		/*
268 		 * Always keep at least nfs_iodmin kthreads.
269 		 */
270 		timo = (myiod < nfs_iodmin) ? 0 : nfs_iodmaxidle * hz;
271 		error = msleep(&ncl_iodwant[myiod], &ncl_iod_mutex, PWAIT | PCATCH,
272 		    "-", timo);
273 		if (error) {
274 			nmp = ncl_iodmount[myiod];
275 			/*
276 			 * Rechecking the nm_bufq closes a rare race where the
277 			 * nfsiod is woken up at the exact time the idle timeout
278 			 * fires
279 			 */
280 			if (nmp && TAILQ_FIRST(&nmp->nm_bufq))
281 				error = 0;
282 			break;
283 		}
284 	    }
285 	    if (error)
286 		    break;
287 	    while ((bp = TAILQ_FIRST(&nmp->nm_bufq)) != NULL) {
288 		/* Take one off the front of the list */
289 		TAILQ_REMOVE(&nmp->nm_bufq, bp, b_freelist);
290 		nmp->nm_bufqlen--;
291 		if (nmp->nm_bufqwant && nmp->nm_bufqlen <= ncl_numasync) {
292 		    nmp->nm_bufqwant = 0;
293 		    wakeup(&nmp->nm_bufq);
294 		}
295 		mtx_unlock(&ncl_iod_mutex);
296 		if (bp->b_flags & B_DIRECT) {
297 			KASSERT((bp->b_iocmd == BIO_WRITE), ("nfscvs_iod: BIO_WRITE not set"));
298 			(void)ncl_doio_directwrite(bp);
299 		} else {
300 			if (bp->b_iocmd == BIO_READ)
301 				(void) ncl_doio(bp->b_vp, bp, bp->b_rcred,
302 				    NULL, 0);
303 			else
304 				(void) ncl_doio(bp->b_vp, bp, bp->b_wcred,
305 				    NULL, 0);
306 		}
307 		mtx_lock(&ncl_iod_mutex);
308 		/*
309 		 * Make sure the nmp hasn't been dismounted as soon as
310 		 * ncl_doio() completes for the last buffer.
311 		 */
312 		nmp = ncl_iodmount[myiod];
313 		if (nmp == NULL)
314 			break;
315 
316 		/*
317 		 * If there are more than one iod on this mount, then defect
318 		 * so that the iods can be shared out fairly between the mounts
319 		 */
320 		if (nfs_defect && nmp->nm_bufqiods > 1) {
321 		    NFS_DPF(ASYNCIO,
322 			    ("nfssvc_iod: iod %d defecting from mount %p\n",
323 			     myiod, nmp));
324 		    ncl_iodmount[myiod] = NULL;
325 		    nmp->nm_bufqiods--;
326 		    break;
327 		}
328 	    }
329 	}
330 finish:
331 	nfs_asyncdaemon[myiod] = 0;
332 	if (nmp)
333 	    nmp->nm_bufqiods--;
334 	ncl_iodwant[myiod] = NFSIOD_NOT_AVAILABLE;
335 	ncl_iodmount[myiod] = NULL;
336 	/* Someone may be waiting for the last nfsiod to terminate. */
337 	if (--ncl_numasync == 0)
338 		wakeup(&ncl_numasync);
339 	mtx_unlock(&ncl_iod_mutex);
340 	if ((error == 0) || (error == EWOULDBLOCK))
341 		kproc_exit(0);
342 	/* Abnormal termination */
343 	kproc_exit(1);
344 }
345