1 /* $OpenBSD: nfs_kq.c,v 1.37 2024/05/01 13:15:59 jsg Exp $ */ 2 /* $NetBSD: nfs_kq.c,v 1.7 2003/10/30 01:43:10 simonb Exp $ */ 3 4 /*- 5 * Copyright (c) 2002 The NetBSD Foundation, Inc. 6 * All rights reserved. 7 * 8 * This code is derived from software contributed to The NetBSD Foundation 9 * by Jaromir Dolecek. 10 * 11 * Redistribution and use in source and binary forms, with or without 12 * modification, are permitted provided that the following conditions 13 * are met: 14 * 1. Redistributions of source code must retain the above copyright 15 * notice, this list of conditions and the following disclaimer. 16 * 2. Redistributions in binary form must reproduce the above copyright 17 * notice, this list of conditions and the following disclaimer in the 18 * documentation and/or other materials provided with the distribution. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS 21 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 22 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 23 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS 24 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 25 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 26 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 27 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 28 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 29 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 30 * POSSIBILITY OF SUCH DAMAGE. 31 */ 32 33 #include <sys/param.h> 34 #include <sys/systm.h> 35 #include <sys/proc.h> 36 #include <sys/mount.h> 37 #include <sys/malloc.h> 38 #include <sys/vnode.h> 39 #include <sys/file.h> 40 #include <sys/kthread.h> 41 #include <sys/rwlock.h> 42 #include <sys/queue.h> 43 44 #include <nfs/nfsproto.h> 45 #include <nfs/nfs.h> 46 #include <nfs/nfsnode.h> 47 #include <nfs/nfs_var.h> 48 49 void nfs_kqpoll(void *); 50 int nfs_kqwatch(struct vnode *); 51 void nfs_kqunwatch(struct vnode *); 52 53 void filt_nfsdetach(struct knote *); 54 int filt_nfsread(struct knote *, long); 55 int filt_nfswrite(struct knote *, long); 56 int filt_nfsvnode(struct knote *, long); 57 58 struct kevq { 59 SLIST_ENTRY(kevq) kev_link; 60 struct vnode *vp; 61 u_int usecount; 62 u_int flags; 63 #define KEVQ_BUSY 0x01 /* currently being processed */ 64 #define KEVQ_WANT 0x02 /* want to change this entry */ 65 struct timespec omtime; /* old modification time */ 66 struct timespec octime; /* old change time */ 67 nlink_t onlink; /* old number of references to file */ 68 }; 69 SLIST_HEAD(kevqlist, kevq); 70 71 struct rwlock nfskevq_lock = RWLOCK_INITIALIZER("nfskqlk"); 72 struct proc *pnfskq; 73 struct kevqlist kevlist = SLIST_HEAD_INITIALIZER(kevlist); 74 75 /* 76 * This quite simplistic routine periodically checks for server changes 77 * of any of the watched files every NFS_MINATTRTIMO/2 seconds. 78 * Only changes in size, modification time, change time and nlinks 79 * are being checked, everything else is ignored. 80 * The routine only calls VOP_GETATTR() when it's likely it would get 81 * some new data, i.e. when the vnode expires from attrcache. This 82 * should give same result as periodically running stat(2) from userland, 83 * while keeping CPU/network usage low, and still provide proper kevent 84 * semantics. 85 * The poller thread is created when first vnode is added to watch list, 86 * and exits when the watch list is empty. The overhead of thread creation 87 * isn't really important, neither speed of attach and detach of knote. 88 */ 89 void 90 nfs_kqpoll(void *arg) 91 { 92 struct kevq *ke; 93 struct vattr attr; 94 struct proc *p = pnfskq; 95 u_quad_t osize; 96 int error; 97 98 for(;;) { 99 rw_enter_write(&nfskevq_lock); 100 SLIST_FOREACH(ke, &kevlist, kev_link) { 101 struct nfsnode *np = VTONFS(ke->vp); 102 103 #ifdef DEBUG 104 printf("nfs_kqpoll on: "); 105 VOP_PRINT(ke->vp); 106 #endif 107 /* skip if still in attrcache */ 108 if (nfs_getattrcache(ke->vp, &attr) != ENOENT) 109 continue; 110 111 /* 112 * Mark entry busy, release lock and check 113 * for changes. 114 */ 115 ke->flags |= KEVQ_BUSY; 116 rw_exit_write(&nfskevq_lock); 117 118 /* save v_size, nfs_getattr() updates it */ 119 osize = np->n_size; 120 121 error = VOP_GETATTR(ke->vp, &attr, p->p_ucred, p); 122 if (error == ESTALE) { 123 NFS_INVALIDATE_ATTRCACHE(np); 124 VN_KNOTE(ke->vp, NOTE_DELETE); 125 goto next; 126 } 127 128 /* following is a bit fragile, but about best 129 * we can get */ 130 if (attr.va_size != osize) { 131 int flags = NOTE_WRITE; 132 133 if (attr.va_size > osize) 134 flags |= NOTE_EXTEND; 135 else 136 flags |= NOTE_TRUNCATE; 137 138 VN_KNOTE(ke->vp, flags); 139 ke->omtime = attr.va_mtime; 140 } else if (attr.va_mtime.tv_sec != ke->omtime.tv_sec 141 || attr.va_mtime.tv_nsec != ke->omtime.tv_nsec) { 142 VN_KNOTE(ke->vp, NOTE_WRITE); 143 ke->omtime = attr.va_mtime; 144 } 145 146 if (attr.va_ctime.tv_sec != ke->octime.tv_sec 147 || attr.va_ctime.tv_nsec != ke->octime.tv_nsec) { 148 VN_KNOTE(ke->vp, NOTE_ATTRIB); 149 ke->octime = attr.va_ctime; 150 } 151 152 if (attr.va_nlink != ke->onlink) { 153 VN_KNOTE(ke->vp, NOTE_LINK); 154 ke->onlink = attr.va_nlink; 155 } 156 157 next: 158 rw_enter_write(&nfskevq_lock); 159 ke->flags &= ~KEVQ_BUSY; 160 if (ke->flags & KEVQ_WANT) { 161 ke->flags &= ~KEVQ_WANT; 162 wakeup(ke); 163 } 164 } 165 166 if (SLIST_EMPTY(&kevlist)) { 167 /* Nothing more to watch, exit */ 168 pnfskq = NULL; 169 rw_exit_write(&nfskevq_lock); 170 kthread_exit(0); 171 } 172 rw_exit_write(&nfskevq_lock); 173 174 /* wait a while before checking for changes again */ 175 tsleep_nsec(pnfskq, PSOCK, "nfskqpw", 176 SEC_TO_NSEC(NFS_MINATTRTIMO) / 2); 177 } 178 } 179 180 void 181 filt_nfsdetach(struct knote *kn) 182 { 183 struct vnode *vp = (struct vnode *)kn->kn_hook; 184 185 klist_remove_locked(&vp->v_klist, kn); 186 187 /* Remove the vnode from watch list */ 188 if ((kn->kn_flags & (__EV_POLL | __EV_SELECT)) == 0) 189 nfs_kqunwatch(vp); 190 } 191 192 void 193 nfs_kqunwatch(struct vnode *vp) 194 { 195 struct kevq *ke; 196 197 rw_enter_write(&nfskevq_lock); 198 SLIST_FOREACH(ke, &kevlist, kev_link) { 199 if (ke->vp == vp) { 200 while (ke->flags & KEVQ_BUSY) { 201 ke->flags |= KEVQ_WANT; 202 rw_exit_write(&nfskevq_lock); 203 tsleep_nsec(ke, PSOCK, "nfskqdet", INFSLP); 204 rw_enter_write(&nfskevq_lock); 205 } 206 207 if (ke->usecount > 1) { 208 /* keep, other kevents need this */ 209 ke->usecount--; 210 } else { 211 /* last user, g/c */ 212 SLIST_REMOVE(&kevlist, ke, kevq, kev_link); 213 free(ke, M_KEVENT, sizeof(*ke)); 214 } 215 break; 216 } 217 } 218 rw_exit_write(&nfskevq_lock); 219 } 220 221 int 222 filt_nfsread(struct knote *kn, long hint) 223 { 224 struct vnode *vp = (struct vnode *)kn->kn_hook; 225 struct nfsnode *np = VTONFS(vp); 226 227 /* 228 * filesystem is gone, so set the EOF flag and schedule 229 * the knote for deletion. 230 */ 231 if (hint == NOTE_REVOKE) { 232 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 233 return (1); 234 } 235 236 kn->kn_data = np->n_size - foffset(kn->kn_fp); 237 #ifdef DEBUG 238 printf("nfsread event. %lld\n", kn->kn_data); 239 #endif 240 if (kn->kn_data == 0 && kn->kn_sfflags & NOTE_EOF) { 241 kn->kn_fflags |= NOTE_EOF; 242 return (1); 243 } 244 245 if (kn->kn_flags & (__EV_POLL | __EV_SELECT)) 246 return (1); 247 248 return (kn->kn_data != 0); 249 } 250 251 int 252 filt_nfswrite(struct knote *kn, long hint) 253 { 254 /* 255 * filesystem is gone, so set the EOF flag and schedule 256 * the knote for deletion. 257 */ 258 if (hint == NOTE_REVOKE) { 259 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 260 return (1); 261 } 262 263 kn->kn_data = 0; 264 return (1); 265 } 266 267 int 268 filt_nfsvnode(struct knote *kn, long hint) 269 { 270 if (kn->kn_sfflags & hint) 271 kn->kn_fflags |= hint; 272 if (hint == NOTE_REVOKE) { 273 kn->kn_flags |= EV_EOF; 274 return (1); 275 } 276 return (kn->kn_fflags != 0); 277 } 278 279 static const struct filterops nfsread_filtops = { 280 .f_flags = FILTEROP_ISFD, 281 .f_attach = NULL, 282 .f_detach = filt_nfsdetach, 283 .f_event = filt_nfsread, 284 }; 285 286 static const struct filterops nfswrite_filtops = { 287 .f_flags = FILTEROP_ISFD, 288 .f_attach = NULL, 289 .f_detach = filt_nfsdetach, 290 .f_event = filt_nfswrite, 291 }; 292 293 static const struct filterops nfsvnode_filtops = { 294 .f_flags = FILTEROP_ISFD, 295 .f_attach = NULL, 296 .f_detach = filt_nfsdetach, 297 .f_event = filt_nfsvnode, 298 }; 299 300 int 301 nfs_kqfilter(void *v) 302 { 303 struct vop_kqfilter_args *ap = v; 304 struct vnode *vp; 305 struct knote *kn; 306 307 vp = ap->a_vp; 308 kn = ap->a_kn; 309 310 #ifdef DEBUG 311 printf("nfs_kqfilter(%d) on: ", kn->kn_filter); 312 VOP_PRINT(vp); 313 #endif 314 315 switch (kn->kn_filter) { 316 case EVFILT_READ: 317 kn->kn_fop = &nfsread_filtops; 318 break; 319 case EVFILT_WRITE: 320 kn->kn_fop = &nfswrite_filtops; 321 break; 322 case EVFILT_VNODE: 323 kn->kn_fop = &nfsvnode_filtops; 324 break; 325 default: 326 return (EINVAL); 327 } 328 329 kn->kn_hook = vp; 330 331 /* 332 * Put the vnode to watched list. 333 */ 334 if ((kn->kn_flags & (__EV_POLL | __EV_SELECT)) == 0) { 335 int error; 336 337 error = nfs_kqwatch(vp); 338 if (error) 339 return (error); 340 } 341 342 klist_insert_locked(&vp->v_klist, kn); 343 344 return (0); 345 } 346 347 int 348 nfs_kqwatch(struct vnode *vp) 349 { 350 struct proc *p = curproc; /* XXX */ 351 struct vattr attr; 352 struct kevq *ke; 353 int error = 0; 354 355 /* 356 * Fetch current attributes. It's only needed when the vnode 357 * is not watched yet, but we need to do this without lock 358 * held. This is likely cheap due to attrcache, so do it now. 359 */ 360 memset(&attr, 0, sizeof(attr)); 361 (void) VOP_GETATTR(vp, &attr, p->p_ucred, p); 362 363 rw_enter_write(&nfskevq_lock); 364 365 /* ensure the poller is running */ 366 if (!pnfskq) { 367 error = kthread_create(nfs_kqpoll, NULL, &pnfskq, 368 "nfskqpoll"); 369 if (error) 370 goto out; 371 } 372 373 SLIST_FOREACH(ke, &kevlist, kev_link) 374 if (ke->vp == vp) 375 break; 376 377 if (ke) { 378 /* already watched, so just bump usecount */ 379 ke->usecount++; 380 } else { 381 /* need a new one */ 382 ke = malloc(sizeof(*ke), M_KEVENT, M_WAITOK); 383 ke->vp = vp; 384 ke->usecount = 1; 385 ke->flags = 0; 386 ke->omtime = attr.va_mtime; 387 ke->octime = attr.va_ctime; 388 ke->onlink = attr.va_nlink; 389 SLIST_INSERT_HEAD(&kevlist, ke, kev_link); 390 } 391 392 /* kick the poller */ 393 wakeup(pnfskq); 394 395 out: 396 rw_exit_write(&nfskevq_lock); 397 return (error); 398 } 399