nfs_kq.c revision 1.4 
1/*	$NetBSD: nfs_kq.c,v 1.4 2003/03/27 19:58:26 jdolecek 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 1.4 2003/03/27 19:58:26 jdolecek 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
62struct 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};
73SLIST_HEAD(kevqlist, kevq);
74
75static struct lock nfskevq_lock;
76static struct proc *pnfskq;
77static struct kevqlist kevlist = SLIST_HEAD_INITIALIZER(kevlist);
78
79void
80nfs_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 */
100static void
101nfs_kqpoll(void *arg)
102{
103	struct kevq *ke;
104	struct vattr attr;
105	int error;
106	struct proc *p = pnfskq;
107	u_quad_t osize;
108
109	for(;;) {
110		lockmgr(&nfskevq_lock, LK_EXCLUSIVE, NULL);
111		SLIST_FOREACH(ke, &kevlist, kev_link) {
112			/* skip if still in attrcache */
113			if (nfs_getattrcache(ke->vp, &attr) != ENOENT)
114				continue;
115
116			/*
117			 * Mark entry busy, release lock and check
118			 * for changes.
119			 */
120			ke->flags |= KEVQ_BUSY;
121			lockmgr(&nfskevq_lock, LK_RELEASE, NULL);
122
123			/* save v_size, nfs_getattr() updates it */
124			osize = ke->vp->v_size;
125
126			error = VOP_GETATTR(ke->vp, &attr, p->p_ucred, p);
127
128			/* following is a bit fragile, but about best
129			 * we can get */
130			if (attr.va_size != osize) {
131				int extended = (attr.va_size > osize);
132				VN_KNOTE(ke->vp, NOTE_WRITE
133					| (extended ? NOTE_EXTEND : 0));
134				ke->omtime = attr.va_mtime;
135			} else if (attr.va_mtime.tv_sec != ke->omtime.tv_sec
136			    || attr.va_mtime.tv_nsec != ke->omtime.tv_nsec) {
137				VN_KNOTE(ke->vp, NOTE_WRITE);
138				ke->omtime = attr.va_mtime;
139			}
140
141			if (attr.va_ctime.tv_sec != ke->octime.tv_sec
142			    || attr.va_ctime.tv_nsec != ke->octime.tv_nsec) {
143				VN_KNOTE(ke->vp, NOTE_ATTRIB);
144				ke->octime = attr.va_ctime;
145			}
146
147			if (attr.va_nlink != ke->onlink) {
148				VN_KNOTE(ke->vp, NOTE_LINK);
149				ke->onlink = attr.va_nlink;
150			}
151
152			lockmgr(&nfskevq_lock, LK_EXCLUSIVE, NULL);
153			ke->flags &= ~KEVQ_BUSY;
154			if (ke->flags & KEVQ_WANT) {
155				ke->flags &= ~KEVQ_WANT;
156				wakeup(ke);
157			}
158		}
159
160		if (SLIST_EMPTY(&kevlist)) {
161			/* Nothing more to watch, exit */
162			pnfskq = NULL;
163			lockmgr(&nfskevq_lock, LK_RELEASE, NULL);
164			kthread_exit(0);
165		}
166		lockmgr(&nfskevq_lock, LK_RELEASE, NULL);
167
168		/* wait a while before checking for changes again */
169		tsleep(pnfskq, PSOCK, "nfskqpw",
170			NFS_MINATTRTIMO * hz / 2);
171
172	}
173}
174
175static void
176filt_nfsdetach(struct knote *kn)
177{
178	struct vnode *vp = (struct vnode *)kn->kn_hook;
179	struct kevq *ke;
180
181	/* XXXLUKEM lock the struct? */
182	SLIST_REMOVE(&vp->v_klist, kn, knote, kn_selnext);
183
184	/* Remove the vnode from watch list */
185	lockmgr(&nfskevq_lock, LK_EXCLUSIVE, NULL);
186	SLIST_FOREACH(ke, &kevlist, kev_link) {
187		if (ke->vp == vp) {
188			while (ke->flags & KEVQ_BUSY) {
189				ke->flags |= KEVQ_WANT;
190				lockmgr(&nfskevq_lock, LK_RELEASE, NULL);
191				(void) tsleep(ke, PSOCK, "nfskqdet", 0);
192				lockmgr(&nfskevq_lock, LK_EXCLUSIVE, NULL);
193			}
194
195			if (ke->usecount > 1) {
196				/* keep, other kevents need this */
197				ke->usecount--;
198			} else {
199				/* last user, g/c */
200				SLIST_REMOVE(&kevlist, ke, kevq, kev_link);
201				FREE(ke, M_KEVENT);
202			}
203			break;
204		}
205	}
206	lockmgr(&nfskevq_lock, LK_RELEASE, NULL);
207}
208
209static int
210filt_nfsread(struct knote *kn, long hint)
211{
212	struct vnode *vp = (struct vnode *)kn->kn_hook;
213
214	/*
215	 * filesystem is gone, so set the EOF flag and schedule
216	 * the knote for deletion.
217	 */
218	if (hint == NOTE_REVOKE) {
219		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
220		return (1);
221	}
222
223	/* XXXLUKEM lock the struct? */
224	kn->kn_data = vp->v_size - kn->kn_fp->f_offset;
225        return (kn->kn_data != 0);
226}
227
228static int
229filt_nfsvnode(struct knote *kn, long hint)
230{
231
232	if (kn->kn_sfflags & hint)
233		kn->kn_fflags |= hint;
234	if (hint == NOTE_REVOKE) {
235		kn->kn_flags |= EV_EOF;
236		return (1);
237	}
238	return (kn->kn_fflags != 0);
239}
240
241static const struct filterops nfsread_filtops =
242	{ 1, NULL, filt_nfsdetach, filt_nfsread };
243static const struct filterops nfsvnode_filtops =
244	{ 1, NULL, filt_nfsdetach, filt_nfsvnode };
245
246int
247nfs_kqfilter(void *v)
248{
249	struct vop_kqfilter_args /* {
250		struct vnode	*a_vp;
251		struct knote	*a_kn;
252	} */ *ap = v;
253	struct vnode *vp;
254	struct knote *kn;
255	struct kevq *ke;
256	int error = 0;
257	struct vattr attr;
258	struct proc *p = curproc;	/* XXX */
259
260	vp = ap->a_vp;
261	kn = ap->a_kn;
262	switch (kn->kn_filter) {
263	case EVFILT_READ:
264		kn->kn_fop = &nfsread_filtops;
265		break;
266	case EVFILT_VNODE:
267		kn->kn_fop = &nfsvnode_filtops;
268		break;
269	default:
270		return (1);
271	}
272
273	kn->kn_hook = vp;
274
275	/*
276	 * Put the vnode to watched list.
277	 */
278
279	/*
280	 * Fetch current attributes. It's only needed when the vnode
281	 * is not watched yet, but we need to do this without lock
282	 * held. This is likely cheap due to attrcache, so do it now.
283	 */
284	memset(&attr, 0, sizeof(attr));
285	(void) VOP_GETATTR(vp, &attr, p->p_ucred, p);
286
287	lockmgr(&nfskevq_lock, LK_EXCLUSIVE, NULL);
288
289	/* ensure the poller is running */
290	if (!pnfskq) {
291		error = kthread_create1(nfs_kqpoll, NULL, &pnfskq,
292				"nfskqpoll");
293		if (error)
294			goto out;
295	}
296
297	SLIST_FOREACH(ke, &kevlist, kev_link) {
298		if (ke->vp == vp)
299			break;
300	}
301
302	if (ke) {
303		/* already watched, so just bump usecount */
304		ke->usecount++;
305	} else {
306		/* need a new one */
307		MALLOC(ke, struct kevq *, sizeof(struct kevq), M_KEVENT,
308			M_WAITOK);
309		ke->vp = vp;
310		ke->usecount = 1;
311		ke->flags = 0;
312		ke->omtime = attr.va_mtime;
313		ke->octime = attr.va_ctime;
314		ke->onlink = attr.va_nlink;
315		SLIST_INSERT_HEAD(&kevlist, ke, kev_link);
316	}
317
318	/* kick the poller */
319	wakeup(pnfskq);
320
321	/* XXXLUKEM lock the struct? */
322	SLIST_INSERT_HEAD(&vp->v_klist, kn, kn_selnext);
323
324    out:
325	lockmgr(&nfskevq_lock, LK_RELEASE, NULL);
326
327	return (error);
328}
329