1 /* $OpenBSD: vfs_sync.c,v 1.64 2020/06/24 22:03:41 cheloha 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 52 #include <sys/kernel.h> 53 #include <sys/sched.h> 54 55 #ifdef FFS_SOFTUPDATES 56 int softdep_process_worklist(struct mount *); 57 #endif 58 59 /* 60 * The workitem queue. 61 */ 62 #define SYNCER_MAXDELAY 32 /* maximum sync delay time */ 63 #define SYNCER_DEFAULT 30 /* default sync delay time */ 64 int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */ 65 int syncdelay = SYNCER_DEFAULT; /* time to delay syncing vnodes */ 66 67 int rushjob = 0; /* number of slots to run ASAP */ 68 int stat_rush_requests = 0; /* number of rush requests */ 69 70 int syncer_delayno = 0; 71 long syncer_mask; 72 LIST_HEAD(synclist, vnode); 73 static struct synclist *syncer_workitem_pending; 74 75 struct proc *syncerproc; 76 77 /* 78 * The workitem queue. 79 * 80 * It is useful to delay writes of file data and filesystem metadata 81 * for tens of seconds so that quickly created and deleted files need 82 * not waste disk bandwidth being created and removed. To realize this, 83 * we append vnodes to a "workitem" queue. When running with a soft 84 * updates implementation, most pending metadata dependencies should 85 * not wait for more than a few seconds. Thus, mounted block devices 86 * are delayed only about half the time that file data is delayed. 87 * Similarly, directory updates are more critical, so are only delayed 88 * about a third the time that file data is delayed. Thus, there are 89 * SYNCER_MAXDELAY queues that are processed round-robin at a rate of 90 * one each second (driven off the filesystem syncer process). The 91 * syncer_delayno variable indicates the next queue that is to be processed. 92 * Items that need to be processed soon are placed in this queue: 93 * 94 * syncer_workitem_pending[syncer_delayno] 95 * 96 * A delay of fifteen seconds is done by placing the request fifteen 97 * entries later in the queue: 98 * 99 * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask] 100 * 101 */ 102 103 void 104 vn_initialize_syncerd(void) 105 { 106 syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE, M_WAITOK, 107 &syncer_mask); 108 syncer_maxdelay = syncer_mask + 1; 109 } 110 111 /* 112 * Add an item to the syncer work queue. 113 */ 114 void 115 vn_syncer_add_to_worklist(struct vnode *vp, int delay) 116 { 117 int s, slot; 118 119 if (delay > syncer_maxdelay - 2) 120 delay = syncer_maxdelay - 2; 121 slot = (syncer_delayno + delay) & syncer_mask; 122 123 s = splbio(); 124 if (vp->v_bioflag & VBIOONSYNCLIST) 125 LIST_REMOVE(vp, v_synclist); 126 127 vp->v_bioflag |= VBIOONSYNCLIST; 128 LIST_INSERT_HEAD(&syncer_workitem_pending[slot], vp, v_synclist); 129 splx(s); 130 } 131 132 /* 133 * System filesystem synchronizer daemon. 134 */ 135 void 136 syncer_thread(void *arg) 137 { 138 struct proc *p = curproc; 139 struct synclist *slp; 140 struct vnode *vp; 141 time_t starttime; 142 int s; 143 144 for (;;) { 145 starttime = gettime(); 146 147 /* 148 * Push files whose dirty time has expired. 149 */ 150 s = splbio(); 151 slp = &syncer_workitem_pending[syncer_delayno]; 152 153 syncer_delayno += 1; 154 if (syncer_delayno == syncer_maxdelay) 155 syncer_delayno = 0; 156 157 while ((vp = LIST_FIRST(slp)) != NULL) { 158 if (vget(vp, LK_EXCLUSIVE | LK_NOWAIT)) { 159 /* 160 * If we fail to get the lock, we move this 161 * vnode one second ahead in time. 162 * XXX - no good, but the best we can do. 163 */ 164 vn_syncer_add_to_worklist(vp, 1); 165 continue; 166 } 167 splx(s); 168 (void) VOP_FSYNC(vp, p->p_ucred, MNT_LAZY, p); 169 vput(vp); 170 s = splbio(); 171 if (LIST_FIRST(slp) == vp) { 172 /* 173 * Note: disk vps can remain on the 174 * worklist too with no dirty blocks, but 175 * since sync_fsync() moves it to a different 176 * slot we are safe. 177 */ 178 #ifdef DIAGNOSTIC 179 if (LIST_FIRST(&vp->v_dirtyblkhd) == NULL && 180 vp->v_type != VBLK) { 181 vprint("fsync failed", vp); 182 if (vp->v_mount != NULL) 183 printf("mounted on: %s\n", 184 vp->v_mount->mnt_stat.f_mntonname); 185 panic("%s: fsync failed", __func__); 186 } 187 #endif /* DIAGNOSTIC */ 188 /* 189 * Put us back on the worklist. The worklist 190 * routine will remove us from our current 191 * position and then add us back in at a later 192 * position. 193 */ 194 vn_syncer_add_to_worklist(vp, syncdelay); 195 } 196 197 sched_pause(yield); 198 } 199 200 splx(s); 201 202 #ifdef FFS_SOFTUPDATES 203 /* 204 * Do soft update processing. 205 */ 206 softdep_process_worklist(NULL); 207 #endif 208 209 /* 210 * The variable rushjob allows the kernel to speed up the 211 * processing of the filesystem syncer process. A rushjob 212 * value of N tells the filesystem syncer to process the next 213 * N seconds worth of work on its queue ASAP. Currently rushjob 214 * is used by the soft update code to speed up the filesystem 215 * syncer process when the incore state is getting so far 216 * ahead of the disk that the kernel memory pool is being 217 * threatened with exhaustion. 218 */ 219 if (rushjob > 0) { 220 rushjob -= 1; 221 continue; 222 } 223 /* 224 * If it has taken us less than a second to process the 225 * current work, then wait. Otherwise start right over 226 * again. We can still lose time if any single round 227 * takes more than two seconds, but it does not really 228 * matter as we are just trying to generally pace the 229 * filesystem activity. 230 */ 231 if (gettime() == starttime) 232 tsleep_nsec(&lbolt, PPAUSE, "syncer", INFSLP); 233 } 234 } 235 236 /* 237 * Request the syncer daemon to speed up its work. 238 * We never push it to speed up more than half of its 239 * normal turn time, otherwise it could take over the cpu. 240 */ 241 int 242 speedup_syncer(void) 243 { 244 if (syncerproc) 245 wakeup_proc(syncerproc, &lbolt); 246 if (rushjob < syncdelay / 2) { 247 rushjob += 1; 248 stat_rush_requests += 1; 249 return 1; 250 } 251 return 0; 252 } 253 254 /* Routine to create and manage a filesystem syncer vnode. */ 255 int sync_fsync(void *); 256 int sync_inactive(void *); 257 int sync_print(void *); 258 259 const struct vops sync_vops = { 260 .vop_close = nullop, 261 .vop_fsync = sync_fsync, 262 .vop_inactive = sync_inactive, 263 .vop_reclaim = nullop, 264 .vop_lock = vop_generic_lock, 265 .vop_unlock = vop_generic_unlock, 266 .vop_islocked = vop_generic_islocked, 267 .vop_print = sync_print 268 }; 269 270 /* 271 * Create a new filesystem syncer vnode for the specified mount point. 272 */ 273 int 274 vfs_allocate_syncvnode(struct mount *mp) 275 { 276 struct vnode *vp; 277 static long start, incr, next; 278 int error; 279 280 /* Allocate a new vnode */ 281 if ((error = getnewvnode(VT_VFS, mp, &sync_vops, &vp)) != 0) { 282 mp->mnt_syncer = NULL; 283 return (error); 284 } 285 vp->v_writecount = 1; 286 vp->v_type = VNON; 287 /* 288 * Place the vnode onto the syncer worklist. We attempt to 289 * scatter them about on the list so that they will go off 290 * at evenly distributed times even if all the filesystems 291 * are mounted at once. 292 */ 293 next += incr; 294 if (next == 0 || next > syncer_maxdelay) { 295 start /= 2; 296 incr /= 2; 297 if (start == 0) { 298 start = syncer_maxdelay / 2; 299 incr = syncer_maxdelay; 300 } 301 next = start; 302 } 303 vn_syncer_add_to_worklist(vp, next); 304 mp->mnt_syncer = vp; 305 return (0); 306 } 307 308 /* 309 * Do a lazy sync of the filesystem. 310 */ 311 int 312 sync_fsync(void *v) 313 { 314 struct vop_fsync_args *ap = v; 315 struct vnode *syncvp = ap->a_vp; 316 struct mount *mp = syncvp->v_mount; 317 int asyncflag; 318 319 /* 320 * We only need to do something if this is a lazy evaluation. 321 */ 322 if (ap->a_waitfor != MNT_LAZY) 323 return (0); 324 325 /* 326 * Move ourselves to the back of the sync list. 327 */ 328 vn_syncer_add_to_worklist(syncvp, syncdelay); 329 330 /* 331 * Walk the list of vnodes pushing all that are dirty and 332 * not already on the sync list. 333 */ 334 if (vfs_busy(mp, VB_READ|VB_NOWAIT) == 0) { 335 asyncflag = mp->mnt_flag & MNT_ASYNC; 336 mp->mnt_flag &= ~MNT_ASYNC; 337 VFS_SYNC(mp, MNT_LAZY, 0, ap->a_cred, ap->a_p); 338 if (asyncflag) 339 mp->mnt_flag |= MNT_ASYNC; 340 vfs_unbusy(mp); 341 } 342 343 return (0); 344 } 345 346 /* 347 * The syncer vnode is no longer needed and is being decommissioned. 348 */ 349 int 350 sync_inactive(void *v) 351 { 352 struct vop_inactive_args *ap = v; 353 354 struct vnode *vp = ap->a_vp; 355 int s; 356 357 if (vp->v_usecount == 0) { 358 VOP_UNLOCK(vp); 359 return (0); 360 } 361 362 vp->v_mount->mnt_syncer = NULL; 363 364 s = splbio(); 365 366 LIST_REMOVE(vp, v_synclist); 367 vp->v_bioflag &= ~VBIOONSYNCLIST; 368 369 splx(s); 370 371 vp->v_writecount = 0; 372 vput(vp); 373 374 return (0); 375 } 376 377 /* 378 * Print out a syncer vnode. 379 */ 380 int 381 sync_print(void *v) 382 { 383 printf("syncer vnode\n"); 384 385 return (0); 386 } 387