1 /* $OpenBSD: vfs_sync.c,v 1.73 2024/10/18 05:52:32 miod Exp $ */ 2 3 /* 4 * Portions of this code are: 5 * 6 * Copyright (c) 1989, 1993 7 * The Regents of the University of California. All rights reserved. 8 * (c) UNIX System Laboratories, Inc. 9 * All or some portions of this file are derived from material licensed 10 * to the University of California by American Telephone and Telegraph 11 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 12 * the permission of UNIX System Laboratories, Inc. 13 * 14 * Redistribution and use in source and binary forms, with or without 15 * modification, are permitted provided that the following conditions 16 * are met: 17 * 1. Redistributions of source code must retain the above copyright 18 * notice, this list of conditions and the following disclaimer. 19 * 2. Redistributions in binary form must reproduce the above copyright 20 * notice, this list of conditions and the following disclaimer in the 21 * documentation and/or other materials provided with the distribution. 22 * 3. Neither the name of the University nor the names of its contributors 23 * may be used to endorse or promote products derived from this software 24 * without specific prior written permission. 25 * 26 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 29 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 36 * SUCH DAMAGE. 37 */ 38 39 /* 40 * Syncer daemon 41 */ 42 43 #include <sys/queue.h> 44 #include <sys/param.h> 45 #include <sys/systm.h> 46 #include <sys/proc.h> 47 #include <sys/mount.h> 48 #include <sys/vnode.h> 49 #include <sys/lock.h> 50 #include <sys/malloc.h> 51 #include <sys/time.h> 52 53 /* 54 * The workitem queue. 55 */ 56 #define SYNCER_MAXDELAY 32 /* maximum sync delay time */ 57 #define SYNCER_DEFAULT 30 /* default sync delay time */ 58 int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */ 59 int syncdelay = SYNCER_DEFAULT; /* time to delay syncing vnodes */ 60 61 int syncer_delayno = 0; 62 long syncer_mask; 63 LIST_HEAD(synclist, vnode); 64 static struct synclist *syncer_workitem_pending; 65 66 struct proc *syncerproc; 67 int syncer_chan; 68 69 /* 70 * The workitem queue. 71 * 72 * It is useful to delay writes of file data and filesystem metadata 73 * for tens of seconds so that quickly created and deleted files need 74 * not waste disk bandwidth being created and removed. To realize this, 75 * we append vnodes to a "workitem" queue. When running with a soft 76 * updates implementation, most pending metadata dependencies should 77 * not wait for more than a few seconds. Thus, mounted block devices 78 * are delayed only about half the time that file data is delayed. 79 * Similarly, directory updates are more critical, so are only delayed 80 * about a third the time that file data is delayed. Thus, there are 81 * SYNCER_MAXDELAY queues that are processed round-robin at a rate of 82 * one each second (driven off the filesystem syncer process). The 83 * syncer_delayno variable indicates the next queue that is to be processed. 84 * Items that need to be processed soon are placed in this queue: 85 * 86 * syncer_workitem_pending[syncer_delayno] 87 * 88 * A delay of fifteen seconds is done by placing the request fifteen 89 * entries later in the queue: 90 * 91 * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask] 92 * 93 */ 94 95 void 96 vn_initialize_syncerd(void) 97 { 98 syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE, M_WAITOK, 99 &syncer_mask); 100 syncer_maxdelay = syncer_mask + 1; 101 } 102 103 /* 104 * Add an item to the syncer work queue. 105 */ 106 void 107 vn_syncer_add_to_worklist(struct vnode *vp, int delay) 108 { 109 int s, slot; 110 111 if (delay > syncer_maxdelay - 2) 112 delay = syncer_maxdelay - 2; 113 slot = (syncer_delayno + delay) & syncer_mask; 114 115 s = splbio(); 116 if (vp->v_bioflag & VBIOONSYNCLIST) 117 LIST_REMOVE(vp, v_synclist); 118 119 vp->v_bioflag |= VBIOONSYNCLIST; 120 LIST_INSERT_HEAD(&syncer_workitem_pending[slot], vp, v_synclist); 121 splx(s); 122 } 123 124 /* 125 * System filesystem synchronizer daemon. 126 */ 127 void 128 syncer_thread(void *arg) 129 { 130 uint64_t elapsed, start; 131 struct proc *p = curproc; 132 struct synclist *slp; 133 struct vnode *vp; 134 int s; 135 136 for (;;) { 137 start = getnsecuptime(); 138 139 /* 140 * Push files whose dirty time has expired. 141 */ 142 s = splbio(); 143 slp = &syncer_workitem_pending[syncer_delayno]; 144 145 syncer_delayno += 1; 146 if (syncer_delayno == syncer_maxdelay) 147 syncer_delayno = 0; 148 149 while ((vp = LIST_FIRST(slp)) != NULL) { 150 if (vget(vp, LK_EXCLUSIVE | LK_NOWAIT)) { 151 /* 152 * If we fail to get the lock, we move this 153 * vnode one second ahead in time. 154 * XXX - no good, but the best we can do. 155 */ 156 vn_syncer_add_to_worklist(vp, 1); 157 continue; 158 } 159 splx(s); 160 (void) VOP_FSYNC(vp, p->p_ucred, MNT_LAZY, p); 161 vput(vp); 162 s = splbio(); 163 if (LIST_FIRST(slp) == vp) { 164 /* 165 * Note: disk vps can remain on the 166 * worklist too with no dirty blocks, but 167 * since sync_fsync() moves it to a different 168 * slot we are safe. 169 */ 170 #ifdef DIAGNOSTIC 171 if (LIST_FIRST(&vp->v_dirtyblkhd) == NULL && 172 vp->v_type != VBLK) { 173 vprint("fsync failed", vp); 174 if (vp->v_mount != NULL) 175 printf("mounted on: %s\n", 176 vp->v_mount->mnt_stat.f_mntonname); 177 panic("%s: fsync failed", __func__); 178 } 179 #endif /* DIAGNOSTIC */ 180 /* 181 * Put us back on the worklist. The worklist 182 * routine will remove us from our current 183 * position and then add us back in at a later 184 * position. 185 */ 186 vn_syncer_add_to_worklist(vp, syncdelay); 187 } 188 189 sched_pause(yield); 190 } 191 192 splx(s); 193 194 /* 195 * If it has taken us less than a second to process the 196 * current work, then wait. Otherwise start right over 197 * again. We can still lose time if any single round 198 * takes more than two seconds, but it does not really 199 * matter as we are just trying to generally pace the 200 * filesystem activity. 201 */ 202 elapsed = getnsecuptime() - start; 203 if (elapsed < SEC_TO_NSEC(1)) { 204 tsleep_nsec(&syncer_chan, PPAUSE, "syncer", 205 SEC_TO_NSEC(1) - elapsed); 206 } 207 } 208 } 209 210 /* Routine to create and manage a filesystem syncer vnode. */ 211 int sync_fsync(void *); 212 int sync_inactive(void *); 213 int sync_print(void *); 214 215 const struct vops sync_vops = { 216 .vop_close = nullop, 217 .vop_fsync = sync_fsync, 218 .vop_inactive = sync_inactive, 219 .vop_reclaim = nullop, 220 .vop_lock = nullop, 221 .vop_unlock = nullop, 222 .vop_islocked = nullop, 223 .vop_print = sync_print, 224 225 .vop_abortop = NULL, 226 .vop_access = NULL, 227 .vop_advlock = NULL, 228 .vop_bmap = NULL, 229 .vop_bwrite = NULL, 230 .vop_create = NULL, 231 .vop_getattr = NULL, 232 .vop_ioctl = NULL, 233 .vop_link = NULL, 234 .vop_lookup = NULL, 235 .vop_mknod = NULL, 236 .vop_open = NULL, 237 .vop_pathconf = NULL, 238 .vop_read = NULL, 239 .vop_readdir = NULL, 240 .vop_readlink = NULL, 241 .vop_remove = eopnotsupp, 242 .vop_rename = NULL, 243 .vop_revoke = NULL, 244 .vop_mkdir = NULL, 245 .vop_rmdir = NULL, 246 .vop_setattr = NULL, 247 .vop_strategy = NULL, 248 .vop_symlink = NULL, 249 .vop_write = NULL, 250 .vop_kqfilter = NULL 251 }; 252 253 /* 254 * Create a new filesystem syncer vnode for the specified mount point. 255 */ 256 int 257 vfs_allocate_syncvnode(struct mount *mp) 258 { 259 struct vnode *vp; 260 static long start, incr, next; 261 int error; 262 263 /* Allocate a new vnode */ 264 if ((error = getnewvnode(VT_VFS, mp, &sync_vops, &vp)) != 0) { 265 mp->mnt_syncer = NULL; 266 return (error); 267 } 268 vp->v_writecount = 1; 269 vp->v_type = VNON; 270 /* 271 * Place the vnode onto the syncer worklist. We attempt to 272 * scatter them about on the list so that they will go off 273 * at evenly distributed times even if all the filesystems 274 * are mounted at once. 275 */ 276 next += incr; 277 if (next == 0 || next > syncer_maxdelay) { 278 start /= 2; 279 incr /= 2; 280 if (start == 0) { 281 start = syncer_maxdelay / 2; 282 incr = syncer_maxdelay; 283 } 284 next = start; 285 } 286 vn_syncer_add_to_worklist(vp, next); 287 mp->mnt_syncer = vp; 288 return (0); 289 } 290 291 /* 292 * Do a lazy sync of the filesystem. 293 */ 294 int 295 sync_fsync(void *v) 296 { 297 struct vop_fsync_args *ap = v; 298 struct vnode *syncvp = ap->a_vp; 299 struct mount *mp = syncvp->v_mount; 300 int asyncflag; 301 302 /* 303 * We only need to do something if this is a lazy evaluation. 304 */ 305 if (ap->a_waitfor != MNT_LAZY) 306 return (0); 307 308 /* 309 * Move ourselves to the back of the sync list. 310 */ 311 vn_syncer_add_to_worklist(syncvp, syncdelay); 312 313 /* 314 * Walk the list of vnodes pushing all that are dirty and 315 * not already on the sync list. 316 */ 317 if (vfs_busy(mp, VB_READ|VB_NOWAIT) == 0) { 318 asyncflag = mp->mnt_flag & MNT_ASYNC; 319 mp->mnt_flag &= ~MNT_ASYNC; 320 VFS_SYNC(mp, MNT_LAZY, 0, ap->a_cred, ap->a_p); 321 if (asyncflag) 322 mp->mnt_flag |= MNT_ASYNC; 323 vfs_unbusy(mp); 324 } 325 326 return (0); 327 } 328 329 /* 330 * The syncer vnode is no longer needed and is being decommissioned. 331 */ 332 int 333 sync_inactive(void *v) 334 { 335 struct vop_inactive_args *ap = v; 336 337 struct vnode *vp = ap->a_vp; 338 int s; 339 340 if (vp->v_usecount == 0) { 341 VOP_UNLOCK(vp); 342 return (0); 343 } 344 345 vp->v_mount->mnt_syncer = NULL; 346 347 s = splbio(); 348 349 LIST_REMOVE(vp, v_synclist); 350 vp->v_bioflag &= ~VBIOONSYNCLIST; 351 352 splx(s); 353 354 vp->v_writecount = 0; 355 vput(vp); 356 357 return (0); 358 } 359 360 /* 361 * Print out a syncer vnode. 362 */ 363 int 364 sync_print(void *v) 365 { 366 #if defined(DEBUG) || defined(DIAGNOSTIC) || defined(VFSLCKDEBUG) 367 printf("syncer vnode\n"); 368 #endif 369 370 return (0); 371 } 372