xref: /netbsd-src/sys/nfs/nfs_kq.c (revision 23c8222edbfb0f0932d88a8351d3a0cf817dfb9e)
1 /*	$NetBSD: nfs_kq.c,v 1.7 2003/10/30 01:43:10 simonb Exp $	*/
2 
3 /*-
4  * Copyright (c) 2002 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  * 3. All advertising materials mentioning features or use of this software
19  *    must display the following acknowledgement:
20  *        This product includes software developed by the NetBSD
21  *        Foundation, Inc. and its contributors.
22  * 4. Neither the name of The NetBSD Foundation nor the names of its
23  *    contributors may be used to endorse or promote products derived
24  *    from this software without specific prior written permission.
25  *
26  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
27  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
28  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
29  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
30  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
31  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
32  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
33  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
34  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
35  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
36  * POSSIBILITY OF SUCH DAMAGE.
37  */
38 
39 #include <sys/cdefs.h>
40 __KERNEL_RCSID(0, "$NetBSD: nfs_kq.c,v 1.7 2003/10/30 01:43:10 simonb Exp $");
41 
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/kernel.h>
45 #include <sys/proc.h>
46 #include <sys/mount.h>
47 #include <sys/malloc.h>
48 #include <sys/vnode.h>
49 #include <sys/unistd.h>
50 #include <sys/file.h>
51 #include <sys/kthread.h>
52 
53 #include <uvm/uvm_extern.h>
54 #include <uvm/uvm.h>
55 
56 #include <nfs/rpcv2.h>
57 #include <nfs/nfsproto.h>
58 #include <nfs/nfs.h>
59 #include <nfs/nfsnode.h>
60 #include <nfs/nfs_var.h>
61 
62 struct kevq {
63 	SLIST_ENTRY(kevq)	kev_link;
64 	struct vnode		*vp;
65 	u_int			usecount;
66 	u_int			flags;
67 #define KEVQ_BUSY	0x01	/* currently being processed */
68 #define KEVQ_WANT	0x02	/* want to change this entry */
69 	struct timespec		omtime;	/* old modification time */
70 	struct timespec		octime;	/* old change time */
71 	nlink_t			onlink;	/* old number of references to file */
72 };
73 SLIST_HEAD(kevqlist, kevq);
74 
75 static struct lock nfskevq_lock;
76 static struct proc *pnfskq;
77 static struct kevqlist kevlist = SLIST_HEAD_INITIALIZER(kevlist);
78 
79 void
80 nfs_kqinit(void)
81 {
82 	lockinit(&nfskevq_lock, PSOCK, "nfskqlck", 0, 0);
83 }
84 
85 /*
86  * This quite simplistic routine periodically checks for server changes
87  * of any of the watched files every NFS_MINATTRTIMO/2 seconds.
88  * Only changes in size, modification time, change time and nlinks
89  * are being checked, everything else is ignored.
90  * The routine only calls VOP_GETATTR() when it's likely it would get
91  * some new data, i.e. when the vnode expires from attrcache. This
92  * should give same result as periodically running stat(2) from userland,
93  * while keeping CPU/network usage low, and still provide proper kevent
94  * semantics.
95  * The poller thread is created when first vnode is added to watch list,
96  * and exits when the watch list is empty. The overhead of thread creation
97  * isn't really important, neither speed of attach and detach of knote.
98  */
99 /* ARGSUSED */
100 static void
101 nfs_kqpoll(void *arg)
102 {
103 	struct kevq *ke;
104 	struct vattr attr;
105 	struct proc *p = pnfskq;
106 	u_quad_t osize;
107 
108 	for(;;) {
109 		lockmgr(&nfskevq_lock, LK_EXCLUSIVE, NULL);
110 		SLIST_FOREACH(ke, &kevlist, kev_link) {
111 			/* skip if still in attrcache */
112 			if (nfs_getattrcache(ke->vp, &attr) != ENOENT)
113 				continue;
114 
115 			/*
116 			 * Mark entry busy, release lock and check
117 			 * for changes.
118 			 */
119 			ke->flags |= KEVQ_BUSY;
120 			lockmgr(&nfskevq_lock, LK_RELEASE, NULL);
121 
122 			/* save v_size, nfs_getattr() updates it */
123 			osize = ke->vp->v_size;
124 
125 			(void) VOP_GETATTR(ke->vp, &attr, p->p_ucred, p);
126 
127 			/* following is a bit fragile, but about best
128 			 * we can get */
129 			if (attr.va_size != osize) {
130 				int extended = (attr.va_size > osize);
131 				VN_KNOTE(ke->vp, NOTE_WRITE
132 					| (extended ? NOTE_EXTEND : 0));
133 				ke->omtime = attr.va_mtime;
134 			} else if (attr.va_mtime.tv_sec != ke->omtime.tv_sec
135 			    || attr.va_mtime.tv_nsec != ke->omtime.tv_nsec) {
136 				VN_KNOTE(ke->vp, NOTE_WRITE);
137 				ke->omtime = attr.va_mtime;
138 			}
139 
140 			if (attr.va_ctime.tv_sec != ke->octime.tv_sec
141 			    || attr.va_ctime.tv_nsec != ke->octime.tv_nsec) {
142 				VN_KNOTE(ke->vp, NOTE_ATTRIB);
143 				ke->octime = attr.va_ctime;
144 			}
145 
146 			if (attr.va_nlink != ke->onlink) {
147 				VN_KNOTE(ke->vp, NOTE_LINK);
148 				ke->onlink = attr.va_nlink;
149 			}
150 
151 			lockmgr(&nfskevq_lock, LK_EXCLUSIVE, NULL);
152 			ke->flags &= ~KEVQ_BUSY;
153 			if (ke->flags & KEVQ_WANT) {
154 				ke->flags &= ~KEVQ_WANT;
155 				wakeup(ke);
156 			}
157 		}
158 
159 		if (SLIST_EMPTY(&kevlist)) {
160 			/* Nothing more to watch, exit */
161 			pnfskq = NULL;
162 			lockmgr(&nfskevq_lock, LK_RELEASE, NULL);
163 			kthread_exit(0);
164 		}
165 		lockmgr(&nfskevq_lock, LK_RELEASE, NULL);
166 
167 		/* wait a while before checking for changes again */
168 		tsleep(pnfskq, PSOCK, "nfskqpw",
169 			NFS_MINATTRTIMO * hz / 2);
170 
171 	}
172 }
173 
174 static void
175 filt_nfsdetach(struct knote *kn)
176 {
177 	struct vnode *vp = (struct vnode *)kn->kn_hook;
178 	struct kevq *ke;
179 
180 	/* XXXLUKEM lock the struct? */
181 	SLIST_REMOVE(&vp->v_klist, kn, knote, kn_selnext);
182 
183 	/* Remove the vnode from watch list */
184 	lockmgr(&nfskevq_lock, LK_EXCLUSIVE, NULL);
185 	SLIST_FOREACH(ke, &kevlist, kev_link) {
186 		if (ke->vp == vp) {
187 			while (ke->flags & KEVQ_BUSY) {
188 				ke->flags |= KEVQ_WANT;
189 				lockmgr(&nfskevq_lock, LK_RELEASE, NULL);
190 				(void) tsleep(ke, PSOCK, "nfskqdet", 0);
191 				lockmgr(&nfskevq_lock, LK_EXCLUSIVE, NULL);
192 			}
193 
194 			if (ke->usecount > 1) {
195 				/* keep, other kevents need this */
196 				ke->usecount--;
197 			} else {
198 				/* last user, g/c */
199 				SLIST_REMOVE(&kevlist, ke, kevq, kev_link);
200 				FREE(ke, M_KEVENT);
201 			}
202 			break;
203 		}
204 	}
205 	lockmgr(&nfskevq_lock, LK_RELEASE, NULL);
206 }
207 
208 static int
209 filt_nfsread(struct knote *kn, long hint)
210 {
211 	struct vnode *vp = (struct vnode *)kn->kn_hook;
212 
213 	/*
214 	 * filesystem is gone, so set the EOF flag and schedule
215 	 * the knote for deletion.
216 	 */
217 	if (hint == NOTE_REVOKE) {
218 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
219 		return (1);
220 	}
221 
222 	/* XXXLUKEM lock the struct? */
223 	kn->kn_data = vp->v_size - kn->kn_fp->f_offset;
224         return (kn->kn_data != 0);
225 }
226 
227 static int
228 filt_nfsvnode(struct knote *kn, long hint)
229 {
230 
231 	if (kn->kn_sfflags & hint)
232 		kn->kn_fflags |= hint;
233 	if (hint == NOTE_REVOKE) {
234 		kn->kn_flags |= EV_EOF;
235 		return (1);
236 	}
237 	return (kn->kn_fflags != 0);
238 }
239 
240 static const struct filterops nfsread_filtops =
241 	{ 1, NULL, filt_nfsdetach, filt_nfsread };
242 static const struct filterops nfsvnode_filtops =
243 	{ 1, NULL, filt_nfsdetach, filt_nfsvnode };
244 
245 int
246 nfs_kqfilter(void *v)
247 {
248 	struct vop_kqfilter_args /* {
249 		struct vnode	*a_vp;
250 		struct knote	*a_kn;
251 	} */ *ap = v;
252 	struct vnode *vp;
253 	struct knote *kn;
254 	struct kevq *ke;
255 	int error = 0;
256 	struct vattr attr;
257 	struct proc *p = curproc;	/* XXX */
258 
259 	vp = ap->a_vp;
260 	kn = ap->a_kn;
261 	switch (kn->kn_filter) {
262 	case EVFILT_READ:
263 		kn->kn_fop = &nfsread_filtops;
264 		break;
265 	case EVFILT_VNODE:
266 		kn->kn_fop = &nfsvnode_filtops;
267 		break;
268 	default:
269 		return (1);
270 	}
271 
272 	kn->kn_hook = vp;
273 
274 	/*
275 	 * Put the vnode to watched list.
276 	 */
277 
278 	/*
279 	 * Fetch current attributes. It's only needed when the vnode
280 	 * is not watched yet, but we need to do this without lock
281 	 * held. This is likely cheap due to attrcache, so do it now.
282 	 */
283 	memset(&attr, 0, sizeof(attr));
284 	(void) VOP_GETATTR(vp, &attr, p->p_ucred, p);
285 
286 	lockmgr(&nfskevq_lock, LK_EXCLUSIVE, NULL);
287 
288 	/* ensure the poller is running */
289 	if (!pnfskq) {
290 		error = kthread_create1(nfs_kqpoll, NULL, &pnfskq,
291 				"nfskqpoll");
292 		if (error)
293 			goto out;
294 	}
295 
296 	SLIST_FOREACH(ke, &kevlist, kev_link) {
297 		if (ke->vp == vp)
298 			break;
299 	}
300 
301 	if (ke) {
302 		/* already watched, so just bump usecount */
303 		ke->usecount++;
304 	} else {
305 		/* need a new one */
306 		MALLOC(ke, struct kevq *, sizeof(struct kevq), M_KEVENT,
307 			M_WAITOK);
308 		ke->vp = vp;
309 		ke->usecount = 1;
310 		ke->flags = 0;
311 		ke->omtime = attr.va_mtime;
312 		ke->octime = attr.va_ctime;
313 		ke->onlink = attr.va_nlink;
314 		SLIST_INSERT_HEAD(&kevlist, ke, kev_link);
315 	}
316 
317 	/* kick the poller */
318 	wakeup(pnfskq);
319 
320 	/* XXXLUKEM lock the struct? */
321 	SLIST_INSERT_HEAD(&vp->v_klist, kn, kn_selnext);
322 
323     out:
324 	lockmgr(&nfskevq_lock, LK_RELEASE, NULL);
325 
326 	return (error);
327 }
328