xref: /netbsd-src/sys/nfs/nfs_kq.c (revision ce099b40997c43048fb78bd578195f81d2456523)
1 /*	$NetBSD: nfs_kq.c,v 1.22 2008/04/28 20:24:10 martin Exp $	*/
2 
3 /*-
4  * Copyright (c) 2002, 2008 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Jaromir Dolecek.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29  * POSSIBILITY OF SUCH DAMAGE.
30  */
31 
32 #include <sys/cdefs.h>
33 __KERNEL_RCSID(0, "$NetBSD: nfs_kq.c,v 1.22 2008/04/28 20:24:10 martin Exp $");
34 
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/condvar.h>
38 #include <sys/kernel.h>
39 #include <sys/proc.h>
40 #include <sys/kmem.h>
41 #include <sys/mount.h>
42 #include <sys/mutex.h>
43 #include <sys/vnode.h>
44 #include <sys/unistd.h>
45 #include <sys/file.h>
46 #include <sys/kthread.h>
47 
48 #include <uvm/uvm_extern.h>
49 #include <uvm/uvm.h>
50 
51 #include <nfs/rpcv2.h>
52 #include <nfs/nfsproto.h>
53 #include <nfs/nfs.h>
54 #include <nfs/nfsnode.h>
55 #include <nfs/nfs_var.h>
56 
57 struct kevq {
58 	SLIST_ENTRY(kevq)	kev_link;
59 	struct vnode		*vp;
60 	u_int			usecount;
61 	u_int			flags;
62 #define KEVQ_BUSY	0x01	/* currently being processed */
63 	struct timespec		omtime;	/* old modification time */
64 	struct timespec		octime;	/* old change time */
65 	nlink_t			onlink;	/* old number of references to file */
66 	kcondvar_t		cv;
67 };
68 SLIST_HEAD(kevqlist, kevq);
69 
70 static kmutex_t nfskevq_lock;
71 static struct lwp *nfskq_thread;
72 static kcondvar_t nfskq_cv;
73 static struct kevqlist kevlist = SLIST_HEAD_INITIALIZER(kevlist);
74 
75 void
76 nfs_kqinit(void)
77 {
78 
79 	mutex_init(&nfskevq_lock, MUTEX_DEFAULT, IPL_NONE);
80 	cv_init(&nfskq_cv, "nfskqpw");
81 }
82 
83 /*
84  * This quite simplistic routine periodically checks for server changes
85  * of any of the watched files every NFS_MINATTRTIMO/2 seconds.
86  * Only changes in size, modification time, change time and nlinks
87  * are being checked, everything else is ignored.
88  * The routine only calls VOP_GETATTR() when it's likely it would get
89  * some new data, i.e. when the vnode expires from attrcache. This
90  * should give same result as periodically running stat(2) from userland,
91  * while keeping CPU/network usage low, and still provide proper kevent
92  * semantics.
93  * The poller thread is created when first vnode is added to watch list,
94  * and exits when the watch list is empty. The overhead of thread creation
95  * isn't really important, neither speed of attach and detach of knote.
96  */
97 /* ARGSUSED */
98 static void
99 nfs_kqpoll(void *arg)
100 {
101 	struct kevq *ke;
102 	struct vattr attr;
103 	struct lwp *l = curlwp;
104 	u_quad_t osize;
105 
106 	mutex_enter(&nfskevq_lock);
107 	for (;;) {
108 		SLIST_FOREACH(ke, &kevlist, kev_link) {
109 			/* skip if still in attrcache */
110 			if (nfs_getattrcache(ke->vp, &attr) != ENOENT)
111 				continue;
112 
113 			/*
114 			 * Mark entry busy, release lock and check
115 			 * for changes.
116 			 */
117 			ke->flags |= KEVQ_BUSY;
118 			mutex_exit(&nfskevq_lock);
119 
120 			/* save v_size, nfs_getattr() updates it */
121 			osize = ke->vp->v_size;
122 
123 			(void) VOP_GETATTR(ke->vp, &attr, l->l_cred);
124 
125 			/* following is a bit fragile, but about best
126 			 * we can get */
127 			if (attr.va_size != osize) {
128 				int extended = (attr.va_size > osize);
129 				VN_KNOTE(ke->vp, NOTE_WRITE
130 					| (extended ? NOTE_EXTEND : 0));
131 				ke->omtime = attr.va_mtime;
132 			} else if (attr.va_mtime.tv_sec != ke->omtime.tv_sec
133 			    || attr.va_mtime.tv_nsec != ke->omtime.tv_nsec) {
134 				VN_KNOTE(ke->vp, NOTE_WRITE);
135 				ke->omtime = attr.va_mtime;
136 			}
137 
138 			if (attr.va_ctime.tv_sec != ke->octime.tv_sec
139 			    || attr.va_ctime.tv_nsec != ke->octime.tv_nsec) {
140 				VN_KNOTE(ke->vp, NOTE_ATTRIB);
141 				ke->octime = attr.va_ctime;
142 			}
143 
144 			if (attr.va_nlink != ke->onlink) {
145 				VN_KNOTE(ke->vp, NOTE_LINK);
146 				ke->onlink = attr.va_nlink;
147 			}
148 
149 			mutex_enter(&nfskevq_lock);
150 			ke->flags &= ~KEVQ_BUSY;
151 			cv_signal(&ke->cv);
152 		}
153 
154 		if (SLIST_EMPTY(&kevlist)) {
155 			/* Nothing more to watch, exit */
156 			nfskq_thread = NULL;
157 			mutex_exit(&nfskevq_lock);
158 			kthread_exit(0);
159 		}
160 
161 		/* wait a while before checking for changes again */
162 		cv_timedwait(&nfskq_cv, &nfskevq_lock,
163 		    NFS_MINATTRTIMO * hz / 2);
164 	}
165 }
166 
167 static void
168 filt_nfsdetach(struct knote *kn)
169 {
170 	struct vnode *vp = (struct vnode *)kn->kn_hook;
171 	struct kevq *ke;
172 
173 	mutex_enter(&vp->v_interlock);
174 	SLIST_REMOVE(&vp->v_klist, kn, knote, kn_selnext);
175 	mutex_exit(&vp->v_interlock);
176 
177 	/* Remove the vnode from watch list */
178 	mutex_enter(&nfskevq_lock);
179 	SLIST_FOREACH(ke, &kevlist, kev_link) {
180 		if (ke->vp == vp) {
181 			while (ke->flags & KEVQ_BUSY) {
182 				cv_wait(&ke->cv, &nfskevq_lock);
183 			}
184 
185 			if (ke->usecount > 1) {
186 				/* keep, other kevents need this */
187 				ke->usecount--;
188 			} else {
189 				/* last user, g/c */
190 				cv_destroy(&ke->cv);
191 				SLIST_REMOVE(&kevlist, ke, kevq, kev_link);
192 				kmem_free(ke, sizeof(*ke));
193 			}
194 			break;
195 		}
196 	}
197 	mutex_exit(&nfskevq_lock);
198 }
199 
200 static int
201 filt_nfsread(struct knote *kn, long hint)
202 {
203 	struct vnode *vp = (struct vnode *)kn->kn_hook;
204 	int rv;
205 
206 	/*
207 	 * filesystem is gone, so set the EOF flag and schedule
208 	 * the knote for deletion.
209 	 */
210 	switch (hint) {
211 	case NOTE_REVOKE:
212 		KASSERT(mutex_owned(&vp->v_interlock));
213 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
214 		return (1);
215 	case 0:
216 		mutex_enter(&vp->v_interlock);
217 		kn->kn_data = vp->v_size - ((file_t *)kn->kn_obj)->f_offset;
218 		rv = (kn->kn_data != 0);
219 		mutex_exit(&vp->v_interlock);
220 		return rv;
221 	default:
222 		KASSERT(mutex_owned(&vp->v_interlock));
223 		kn->kn_data = vp->v_size - ((file_t *)kn->kn_obj)->f_offset;
224 		return (kn->kn_data != 0);
225 	}
226 }
227 
228 static int
229 filt_nfsvnode(struct knote *kn, long hint)
230 {
231 	struct vnode *vp = (struct vnode *)kn->kn_hook;
232 	int fflags;
233 
234 	switch (hint) {
235 	case NOTE_REVOKE:
236 		KASSERT(mutex_owned(&vp->v_interlock));
237 		kn->kn_flags |= EV_EOF;
238 		if ((kn->kn_sfflags & hint) != 0)
239 			kn->kn_fflags |= hint;
240 		return (1);
241 	case 0:
242 		mutex_enter(&vp->v_interlock);
243 		fflags = kn->kn_fflags;
244 		mutex_exit(&vp->v_interlock);
245 		break;
246 	default:
247 		KASSERT(mutex_owned(&vp->v_interlock));
248 		if ((kn->kn_sfflags & hint) != 0)
249 			kn->kn_fflags |= hint;
250 		fflags = kn->kn_fflags;
251 		break;
252 	}
253 
254 	return (fflags != 0);
255 }
256 
257 
258 static const struct filterops nfsread_filtops =
259 	{ 1, NULL, filt_nfsdetach, filt_nfsread };
260 static const struct filterops nfsvnode_filtops =
261 	{ 1, NULL, filt_nfsdetach, filt_nfsvnode };
262 
263 int
264 nfs_kqfilter(void *v)
265 {
266 	struct vop_kqfilter_args /* {
267 		struct vnode	*a_vp;
268 		struct knote	*a_kn;
269 	} */ *ap = v;
270 	struct vnode *vp;
271 	struct knote *kn;
272 	struct kevq *ke;
273 	int error = 0;
274 	struct vattr attr;
275 	struct lwp *l = curlwp;
276 
277 	vp = ap->a_vp;
278 	kn = ap->a_kn;
279 	switch (kn->kn_filter) {
280 	case EVFILT_READ:
281 		kn->kn_fop = &nfsread_filtops;
282 		break;
283 	case EVFILT_VNODE:
284 		kn->kn_fop = &nfsvnode_filtops;
285 		break;
286 	default:
287 		return (EINVAL);
288 	}
289 
290 	/*
291 	 * Put the vnode to watched list.
292 	 */
293 
294 	/*
295 	 * Fetch current attributes. It's only needed when the vnode
296 	 * is not watched yet, but we need to do this without lock
297 	 * held. This is likely cheap due to attrcache, so do it now.
298 	 */
299 	memset(&attr, 0, sizeof(attr));
300 	(void) VOP_GETATTR(vp, &attr, l->l_cred);
301 
302 	mutex_enter(&nfskevq_lock);
303 
304 	/* ensure the poller is running */
305 	if (!nfskq_thread) {
306 		error = kthread_create(PRI_NONE, 0, NULL, nfs_kqpoll,
307 		    NULL, &nfskq_thread, "nfskqpoll");
308 		if (error) {
309 			mutex_exit(&nfskevq_lock);
310 			return error;
311 		}
312 	}
313 
314 	SLIST_FOREACH(ke, &kevlist, kev_link) {
315 		if (ke->vp == vp)
316 			break;
317 	}
318 
319 	if (ke) {
320 		/* already watched, so just bump usecount */
321 		ke->usecount++;
322 	} else {
323 		/* need a new one */
324 		ke = kmem_alloc(sizeof(*ke), KM_SLEEP);
325 		ke->vp = vp;
326 		ke->usecount = 1;
327 		ke->flags = 0;
328 		ke->omtime = attr.va_mtime;
329 		ke->octime = attr.va_ctime;
330 		ke->onlink = attr.va_nlink;
331 		cv_init(&ke->cv, "nfskqdet");
332 		SLIST_INSERT_HEAD(&kevlist, ke, kev_link);
333 	}
334 
335 	mutex_enter(&vp->v_interlock);
336 	SLIST_INSERT_HEAD(&vp->v_klist, kn, kn_selnext);
337 	kn->kn_hook = vp;
338 	mutex_exit(&vp->v_interlock);
339 
340 	/* kick the poller */
341 	cv_signal(&nfskq_cv);
342 	mutex_exit(&nfskevq_lock);
343 
344 
345 	return (error);
346 }
347