xref: /netbsd-src/sys/nfs/nfs_kq.c (revision b1c86f5f087524e68db12794ee9c3e3da1ab17a0)
1 /*	$NetBSD: nfs_kq.c,v 1.23 2008/11/19 18:36:09 ad 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 1.23 2008/11/19 18:36:09 ad 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 nfskq_lock;
71 static struct lwp *nfskq_thread;
72 static kcondvar_t nfskq_cv;
73 static struct kevqlist kevlist = SLIST_HEAD_INITIALIZER(kevlist);
74 static bool nfskq_thread_exit;
75 
76 void
77 nfs_kqinit(void)
78 {
79 
80 	mutex_init(&nfskq_lock, MUTEX_DEFAULT, IPL_NONE);
81 	cv_init(&nfskq_cv, "nfskqpw");
82 }
83 
84 void
85 nfs_kqfini(void)
86 {
87 
88 	if (nfskq_thread != NULL) {
89 		mutex_enter(&nfskq_lock);
90 		nfskq_thread_exit = true;
91 		cv_broadcast(&nfskq_cv);
92 		do {
93 			cv_wait(&nfskq_cv, &nfskq_lock);
94 		} while (nfskq_thread != NULL);
95 		mutex_exit(&nfskq_lock);
96 	}
97 	mutex_destroy(&nfskq_lock);
98 	cv_destroy(&nfskq_cv);
99 }
100 
101 /*
102  * This quite simplistic routine periodically checks for server changes
103  * of any of the watched files every NFS_MINATTRTIMO/2 seconds.
104  * Only changes in size, modification time, change time and nlinks
105  * are being checked, everything else is ignored.
106  * The routine only calls VOP_GETATTR() when it's likely it would get
107  * some new data, i.e. when the vnode expires from attrcache. This
108  * should give same result as periodically running stat(2) from userland,
109  * while keeping CPU/network usage low, and still provide proper kevent
110  * semantics.
111  * The poller thread is created when first vnode is added to watch list,
112  * and exits when the watch list is empty. The overhead of thread creation
113  * isn't really important, neither speed of attach and detach of knote.
114  */
115 /* ARGSUSED */
116 static void
117 nfs_kqpoll(void *arg)
118 {
119 	struct kevq *ke;
120 	struct vattr attr;
121 	struct lwp *l = curlwp;
122 	u_quad_t osize;
123 
124 	mutex_enter(&nfskq_lock);
125 	while (!nfskq_thread_exit) {
126 		SLIST_FOREACH(ke, &kevlist, kev_link) {
127 			/* skip if still in attrcache */
128 			if (nfs_getattrcache(ke->vp, &attr) != ENOENT)
129 				continue;
130 
131 			/*
132 			 * Mark entry busy, release lock and check
133 			 * for changes.
134 			 */
135 			ke->flags |= KEVQ_BUSY;
136 			mutex_exit(&nfskq_lock);
137 
138 			/* save v_size, nfs_getattr() updates it */
139 			osize = ke->vp->v_size;
140 
141 			(void) VOP_GETATTR(ke->vp, &attr, l->l_cred);
142 
143 			/* following is a bit fragile, but about best
144 			 * we can get */
145 			if (attr.va_size != osize) {
146 				int extended = (attr.va_size > osize);
147 				VN_KNOTE(ke->vp, NOTE_WRITE
148 					| (extended ? NOTE_EXTEND : 0));
149 				ke->omtime = attr.va_mtime;
150 			} else if (attr.va_mtime.tv_sec != ke->omtime.tv_sec
151 			    || attr.va_mtime.tv_nsec != ke->omtime.tv_nsec) {
152 				VN_KNOTE(ke->vp, NOTE_WRITE);
153 				ke->omtime = attr.va_mtime;
154 			}
155 
156 			if (attr.va_ctime.tv_sec != ke->octime.tv_sec
157 			    || attr.va_ctime.tv_nsec != ke->octime.tv_nsec) {
158 				VN_KNOTE(ke->vp, NOTE_ATTRIB);
159 				ke->octime = attr.va_ctime;
160 			}
161 
162 			if (attr.va_nlink != ke->onlink) {
163 				VN_KNOTE(ke->vp, NOTE_LINK);
164 				ke->onlink = attr.va_nlink;
165 			}
166 
167 			mutex_enter(&nfskq_lock);
168 			ke->flags &= ~KEVQ_BUSY;
169 			cv_signal(&ke->cv);
170 		}
171 
172 		if (SLIST_EMPTY(&kevlist)) {
173 			/* Nothing more to watch, exit */
174 			nfskq_thread = NULL;
175 			mutex_exit(&nfskq_lock);
176 			kthread_exit(0);
177 		}
178 
179 		/* wait a while before checking for changes again */
180 		cv_timedwait(&nfskq_cv, &nfskq_lock,
181 		    NFS_MINATTRTIMO * hz / 2);
182 	}
183 	nfskq_thread = NULL;
184 	cv_broadcast(&nfskq_cv);
185 	mutex_exit(&nfskq_lock);
186 }
187 
188 static void
189 filt_nfsdetach(struct knote *kn)
190 {
191 	struct vnode *vp = (struct vnode *)kn->kn_hook;
192 	struct kevq *ke;
193 
194 	mutex_enter(&vp->v_interlock);
195 	SLIST_REMOVE(&vp->v_klist, kn, knote, kn_selnext);
196 	mutex_exit(&vp->v_interlock);
197 
198 	/* Remove the vnode from watch list */
199 	mutex_enter(&nfskq_lock);
200 	SLIST_FOREACH(ke, &kevlist, kev_link) {
201 		if (ke->vp == vp) {
202 			while (ke->flags & KEVQ_BUSY) {
203 				cv_wait(&ke->cv, &nfskq_lock);
204 			}
205 
206 			if (ke->usecount > 1) {
207 				/* keep, other kevents need this */
208 				ke->usecount--;
209 			} else {
210 				/* last user, g/c */
211 				cv_destroy(&ke->cv);
212 				SLIST_REMOVE(&kevlist, ke, kevq, kev_link);
213 				kmem_free(ke, sizeof(*ke));
214 			}
215 			break;
216 		}
217 	}
218 	mutex_exit(&nfskq_lock);
219 }
220 
221 static int
222 filt_nfsread(struct knote *kn, long hint)
223 {
224 	struct vnode *vp = (struct vnode *)kn->kn_hook;
225 	int rv;
226 
227 	/*
228 	 * filesystem is gone, so set the EOF flag and schedule
229 	 * the knote for deletion.
230 	 */
231 	switch (hint) {
232 	case NOTE_REVOKE:
233 		KASSERT(mutex_owned(&vp->v_interlock));
234 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
235 		return (1);
236 	case 0:
237 		mutex_enter(&vp->v_interlock);
238 		kn->kn_data = vp->v_size - ((file_t *)kn->kn_obj)->f_offset;
239 		rv = (kn->kn_data != 0);
240 		mutex_exit(&vp->v_interlock);
241 		return rv;
242 	default:
243 		KASSERT(mutex_owned(&vp->v_interlock));
244 		kn->kn_data = vp->v_size - ((file_t *)kn->kn_obj)->f_offset;
245 		return (kn->kn_data != 0);
246 	}
247 }
248 
249 static int
250 filt_nfsvnode(struct knote *kn, long hint)
251 {
252 	struct vnode *vp = (struct vnode *)kn->kn_hook;
253 	int fflags;
254 
255 	switch (hint) {
256 	case NOTE_REVOKE:
257 		KASSERT(mutex_owned(&vp->v_interlock));
258 		kn->kn_flags |= EV_EOF;
259 		if ((kn->kn_sfflags & hint) != 0)
260 			kn->kn_fflags |= hint;
261 		return (1);
262 	case 0:
263 		mutex_enter(&vp->v_interlock);
264 		fflags = kn->kn_fflags;
265 		mutex_exit(&vp->v_interlock);
266 		break;
267 	default:
268 		KASSERT(mutex_owned(&vp->v_interlock));
269 		if ((kn->kn_sfflags & hint) != 0)
270 			kn->kn_fflags |= hint;
271 		fflags = kn->kn_fflags;
272 		break;
273 	}
274 
275 	return (fflags != 0);
276 }
277 
278 
279 static const struct filterops nfsread_filtops =
280 	{ 1, NULL, filt_nfsdetach, filt_nfsread };
281 static const struct filterops nfsvnode_filtops =
282 	{ 1, NULL, filt_nfsdetach, filt_nfsvnode };
283 
284 int
285 nfs_kqfilter(void *v)
286 {
287 	struct vop_kqfilter_args /* {
288 		struct vnode	*a_vp;
289 		struct knote	*a_kn;
290 	} */ *ap = v;
291 	struct vnode *vp;
292 	struct knote *kn;
293 	struct kevq *ke;
294 	int error = 0;
295 	struct vattr attr;
296 	struct lwp *l = curlwp;
297 
298 	vp = ap->a_vp;
299 	kn = ap->a_kn;
300 	switch (kn->kn_filter) {
301 	case EVFILT_READ:
302 		kn->kn_fop = &nfsread_filtops;
303 		break;
304 	case EVFILT_VNODE:
305 		kn->kn_fop = &nfsvnode_filtops;
306 		break;
307 	default:
308 		return (EINVAL);
309 	}
310 
311 	/*
312 	 * Put the vnode to watched list.
313 	 */
314 
315 	/*
316 	 * Fetch current attributes. It's only needed when the vnode
317 	 * is not watched yet, but we need to do this without lock
318 	 * held. This is likely cheap due to attrcache, so do it now.
319 	 */
320 	memset(&attr, 0, sizeof(attr));
321 	(void) VOP_GETATTR(vp, &attr, l->l_cred);
322 
323 	mutex_enter(&nfskq_lock);
324 
325 	/* ensure the poller is running */
326 	if (!nfskq_thread) {
327 		error = kthread_create(PRI_NONE, 0, NULL, nfs_kqpoll,
328 		    NULL, &nfskq_thread, "nfskqpoll");
329 		if (error) {
330 			mutex_exit(&nfskq_lock);
331 			return error;
332 		}
333 	}
334 
335 	SLIST_FOREACH(ke, &kevlist, kev_link) {
336 		if (ke->vp == vp)
337 			break;
338 	}
339 
340 	if (ke) {
341 		/* already watched, so just bump usecount */
342 		ke->usecount++;
343 	} else {
344 		/* need a new one */
345 		ke = kmem_alloc(sizeof(*ke), KM_SLEEP);
346 		ke->vp = vp;
347 		ke->usecount = 1;
348 		ke->flags = 0;
349 		ke->omtime = attr.va_mtime;
350 		ke->octime = attr.va_ctime;
351 		ke->onlink = attr.va_nlink;
352 		cv_init(&ke->cv, "nfskqdet");
353 		SLIST_INSERT_HEAD(&kevlist, ke, kev_link);
354 	}
355 
356 	mutex_enter(&vp->v_interlock);
357 	SLIST_INSERT_HEAD(&vp->v_klist, kn, kn_selnext);
358 	kn->kn_hook = vp;
359 	mutex_exit(&vp->v_interlock);
360 
361 	/* kick the poller */
362 	cv_signal(&nfskq_cv);
363 	mutex_exit(&nfskq_lock);
364 
365 	return (error);
366 }
367