1/* $NetBSD: nfs_kq.c,v 1.25 2011/10/24 11:43:30 hannken 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 1.25 2011/10/24 11:43:30 hannken 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
57struct 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};
68SLIST_HEAD(kevqlist, kevq);
69
70static kmutex_t nfskq_lock;
71static struct lwp *nfskq_thread;
72static kcondvar_t nfskq_cv;
73static struct kevqlist kevlist = SLIST_HEAD_INITIALIZER(kevlist);
74static bool nfskq_thread_exit;
75
76void
77nfs_kqinit(void)
78{
79
80 mutex_init(&nfskq_lock, MUTEX_DEFAULT, IPL_NONE);
81 cv_init(&nfskq_cv, "nfskqpw");
82}
83
84void
85nfs_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 */
116static void
117nfs_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 memset(&attr, 0, sizeof(attr));
142 vn_lock(ke->vp, LK_SHARED | LK_RETRY);
143 (void) VOP_GETATTR(ke->vp, &attr, l->l_cred);
144 VOP_UNLOCK(ke->vp);
145
146 /* following is a bit fragile, but about best
147 * we can get */
148 if (attr.va_size != osize) {
149 int extended = (attr.va_size > osize);
150 VN_KNOTE(ke->vp, NOTE_WRITE
151 | (extended ? NOTE_EXTEND : 0));
152 ke->omtime = attr.va_mtime;
153 } else if (attr.va_mtime.tv_sec != ke->omtime.tv_sec
154 || attr.va_mtime.tv_nsec != ke->omtime.tv_nsec) {
155 VN_KNOTE(ke->vp, NOTE_WRITE);
156 ke->omtime = attr.va_mtime;
157 }
158
159 if (attr.va_ctime.tv_sec != ke->octime.tv_sec
160 || attr.va_ctime.tv_nsec != ke->octime.tv_nsec) {
161 VN_KNOTE(ke->vp, NOTE_ATTRIB);
162 ke->octime = attr.va_ctime;
163 }
164
165 if (attr.va_nlink != ke->onlink) {
166 VN_KNOTE(ke->vp, NOTE_LINK);
167 ke->onlink = attr.va_nlink;
168 }
169
170 mutex_enter(&nfskq_lock);
171 ke->flags &= ~KEVQ_BUSY;
172 cv_signal(&ke->cv);
173 }
174
175 if (SLIST_EMPTY(&kevlist)) {
176 /* Nothing more to watch, exit */
177 nfskq_thread = NULL;
178 mutex_exit(&nfskq_lock);
179 kthread_exit(0);
180 }
181
182 /* wait a while before checking for changes again */
183 cv_timedwait(&nfskq_cv, &nfskq_lock,
184 NFS_MINATTRTIMO * hz / 2);
185 }
186 nfskq_thread = NULL;
187 cv_broadcast(&nfskq_cv);
188 mutex_exit(&nfskq_lock);
189}
190
191static void
192filt_nfsdetach(struct knote *kn)
193{
194 struct vnode *vp = (struct vnode *)kn->kn_hook;
195 struct kevq *ke;
196
197 mutex_enter(vp->v_interlock);
198 SLIST_REMOVE(&vp->v_klist, kn, knote, kn_selnext);
199 mutex_exit(vp->v_interlock);
200
201 /* Remove the vnode from watch list */
202 mutex_enter(&nfskq_lock);
203 SLIST_FOREACH(ke, &kevlist, kev_link) {
204 if (ke->vp == vp) {
205 while (ke->flags & KEVQ_BUSY) {
206 cv_wait(&ke->cv, &nfskq_lock);
207 }
208
209 if (ke->usecount > 1) {
210 /* keep, other kevents need this */
211 ke->usecount--;
212 } else {
213 /* last user, g/c */
214 cv_destroy(&ke->cv);
215 SLIST_REMOVE(&kevlist, ke, kevq, kev_link);
216 kmem_free(ke, sizeof(*ke));
217 }
218 break;
219 }
220 }
221 mutex_exit(&nfskq_lock);
222}
223
224static int
225filt_nfsread(struct knote *kn, long hint)
226{
227 struct vnode *vp = (struct vnode *)kn->kn_hook;
228 int rv;
229
230 /*
231 * filesystem is gone, so set the EOF flag and schedule
232 * the knote for deletion.
233 */
234 switch (hint) {
235 case NOTE_REVOKE:
236 KASSERT(mutex_owned(vp->v_interlock));
237 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
238 return (1);
239 case 0:
240 mutex_enter(vp->v_interlock);
241 kn->kn_data = vp->v_size - ((file_t *)kn->kn_obj)->f_offset;
242 rv = (kn->kn_data != 0);
243 mutex_exit(vp->v_interlock);
244 return rv;
245 default:
246 KASSERT(mutex_owned(vp->v_interlock));
247 kn->kn_data = vp->v_size - ((file_t *)kn->kn_obj)->f_offset;
248 return (kn->kn_data != 0);
249 }
250}
251
252static int
253filt_nfsvnode(struct knote *kn, long hint)
254{
255 struct vnode *vp = (struct vnode *)kn->kn_hook;
256 int fflags;
257
258 switch (hint) {
259 case NOTE_REVOKE:
260 KASSERT(mutex_owned(vp->v_interlock));
261 kn->kn_flags |= EV_EOF;
262 if ((kn->kn_sfflags & hint) != 0)
263 kn->kn_fflags |= hint;
264 return (1);
265 case 0:
266 mutex_enter(vp->v_interlock);
267 fflags = kn->kn_fflags;
268 mutex_exit(vp->v_interlock);
269 break;
270 default:
271 KASSERT(mutex_owned(vp->v_interlock));
272 if ((kn->kn_sfflags & hint) != 0)
273 kn->kn_fflags |= hint;
274 fflags = kn->kn_fflags;
275 break;
276 }
277
278 return (fflags != 0);
279}
280
281
282static const struct filterops nfsread_filtops =
283 { 1, NULL, filt_nfsdetach, filt_nfsread };
284static const struct filterops nfsvnode_filtops =
285 { 1, NULL, filt_nfsdetach, filt_nfsvnode };
286
287int
288nfs_kqfilter(void *v)
289{
290 struct vop_kqfilter_args /* {
291 struct vnode *a_vp;
292 struct knote *a_kn;
293 } */ *ap = v;
294 struct vnode *vp;
295 struct knote *kn;
296 struct kevq *ke;
297 int error = 0;
298 struct vattr attr;
299 struct lwp *l = curlwp;
300
301 vp = ap->a_vp;
302 kn = ap->a_kn;
303 switch (kn->kn_filter) {
304 case EVFILT_READ:
305 kn->kn_fop = &nfsread_filtops;
306 break;
307 case EVFILT_VNODE:
308 kn->kn_fop = &nfsvnode_filtops;
309 break;
310 default:
311 return (EINVAL);
312 }
313
314 /*
315 * Put the vnode to watched list.
316 */
317
318 /*
319 * Fetch current attributes. It's only needed when the vnode
320 * is not watched yet, but we need to do this without lock
321 * held. This is likely cheap due to attrcache, so do it now.
322 */
323 memset(&attr, 0, sizeof(attr));
324 vn_lock(vp, LK_SHARED | LK_RETRY);
325 (void) VOP_GETATTR(vp, &attr, l->l_cred);
326 VOP_UNLOCK(vp);
327
328 mutex_enter(&nfskq_lock);
329
330 /* ensure the poller is running */
331 if (!nfskq_thread) {
332 error = kthread_create(PRI_NONE, 0, NULL, nfs_kqpoll,
333 NULL, &nfskq_thread, "nfskqpoll");
334 if (error) {
335 mutex_exit(&nfskq_lock);
336 return error;
337 }
338 }
339
340 SLIST_FOREACH(ke, &kevlist, kev_link) {
341 if (ke->vp == vp)
342 break;
343 }
344
345 if (ke) {
346 /* already watched, so just bump usecount */
347 ke->usecount++;
348 } else {
349 /* need a new one */
350 ke = kmem_alloc(sizeof(*ke), KM_SLEEP);
351 ke->vp = vp;
352 ke->usecount = 1;
353 ke->flags = 0;
354 ke->omtime = attr.va_mtime;
355 ke->octime = attr.va_ctime;
356 ke->onlink = attr.va_nlink;
357 cv_init(&ke->cv, "nfskqdet");
358 SLIST_INSERT_HEAD(&kevlist, ke, kev_link);
359 }
360
361 mutex_enter(vp->v_interlock);
362 SLIST_INSERT_HEAD(&vp->v_klist, kn, kn_selnext);
363 kn->kn_hook = vp;
364 mutex_exit(vp->v_interlock);
365
366 /* kick the poller */
367 cv_signal(&nfskq_cv);
368 mutex_exit(&nfskq_lock);
369
370 return (error);
371}
372