1 /* $OpenBSD: vfs_sync.c,v 1.12 2000/03/23 15:57:33 art 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. All advertising materials mentioning features or use of this software 23 * must display the following acknowledgement: 24 * This product includes software developed by the University of 25 * California, Berkeley and its contributors. 26 * 4. Neither the name of the University nor the names of its contributors 27 * may be used to endorse or promote products derived from this software 28 * without specific prior written permission. 29 * 30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 40 * SUCH DAMAGE. 41 */ 42 43 /* 44 * Syncer daemon 45 */ 46 47 #include <sys/queue.h> 48 #include <sys/param.h> 49 #include <sys/systm.h> 50 #include <sys/proc.h> 51 #include <sys/mount.h> 52 #include <sys/vnode.h> 53 #include <sys/buf.h> 54 #include <sys/malloc.h> 55 56 #include <sys/kernel.h> 57 58 /* 59 * The workitem queue. 60 */ 61 #define SYNCER_MAXDELAY 32 /* maximum sync delay time */ 62 #define SYNCER_DEFAULT 30 /* default sync delay time */ 63 int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */ 64 time_t syncdelay = SYNCER_DEFAULT; /* time to delay syncing vnodes */ 65 66 int rushjob = 0; /* number of slots to run ASAP */ 67 int stat_rush_requests = 0; /* number of rush requests */ 68 69 static int syncer_delayno = 0; 70 static long syncer_last; 71 LIST_HEAD(synclist, vnode); 72 static struct synclist *syncer_workitem_pending; 73 74 extern struct simplelock mountlist_slock; 75 76 struct proc *syncerproc; 77 78 /* 79 * The workitem queue. 80 * 81 * It is useful to delay writes of file data and filesystem metadata 82 * for tens of seconds so that quickly created and deleted files need 83 * not waste disk bandwidth being created and removed. To realize this, 84 * we append vnodes to a "workitem" queue. When running with a soft 85 * updates implementation, most pending metadata dependencies should 86 * not wait for more than a few seconds. Thus, mounted on block devices 87 * are delayed only about a half the time that file data is delayed. 88 * Similarly, directory updates are more critical, so are only delayed 89 * about a third the time that file data is delayed. Thus, there are 90 * SYNCER_MAXDELAY queues that are processed round-robin at a rate of 91 * one each second (driven off the filesystem syner process). The 92 * syncer_delayno variable indicates the next queue that is to be processed. 93 * Items that need to be processed soon are placed in this queue: 94 * 95 * syncer_workitem_pending[syncer_delayno] 96 * 97 * A delay of fifteen seconds is done by placing the request fifteen 98 * entries later in the queue: 99 * 100 * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask] 101 * 102 */ 103 104 void 105 vn_initialize_syncerd() 106 107 { 108 int i; 109 110 syncer_last = SYNCER_MAXDELAY + 2; 111 112 syncer_workitem_pending = 113 malloc(syncer_last * sizeof(struct synclist), 114 M_VNODE, M_WAITOK); 115 116 for (i = 0; i < syncer_last; i++) 117 LIST_INIT(&syncer_workitem_pending[i]); 118 } 119 120 /* 121 * Add an item to the syncer work queue. 122 */ 123 void 124 vn_syncer_add_to_worklist(vp, delay) 125 struct vnode *vp; 126 int delay; 127 { 128 int s, slot; 129 130 s = splbio(); 131 132 if (vp->v_flag & VONSYNCLIST) 133 LIST_REMOVE(vp, v_synclist); 134 135 if (delay > syncer_maxdelay) 136 delay = syncer_maxdelay; 137 slot = (syncer_delayno + delay) % syncer_last; 138 LIST_INSERT_HEAD(&syncer_workitem_pending[slot], vp, v_synclist); 139 vp->v_flag |= VONSYNCLIST; 140 splx(s); 141 } 142 143 /* 144 * System filesystem synchronizer daemon. 145 */ 146 147 void 148 sched_sync(p) 149 struct proc *p; 150 { 151 struct synclist *slp; 152 struct vnode *vp; 153 long starttime; 154 int s; 155 156 syncerproc = curproc; 157 158 for (;;) { 159 starttime = time.tv_sec; 160 161 /* 162 * Push files whose dirty time has expired. 163 */ 164 s = splbio(); 165 slp = &syncer_workitem_pending[syncer_delayno]; 166 syncer_delayno += 1; 167 if (syncer_delayno >= syncer_last) 168 syncer_delayno = 0; 169 splx(s); 170 while ((vp = LIST_FIRST(slp)) != NULL) { 171 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, p); 172 (void) VOP_FSYNC(vp, p->p_ucred, MNT_LAZY, p); 173 VOP_UNLOCK(vp, 0, p); 174 if (LIST_FIRST(slp) == vp) { 175 if (LIST_FIRST(&vp->v_dirtyblkhd) == NULL && 176 vp->v_type != VBLK) 177 panic("sched_sync: fsync failed"); 178 /* 179 * Move ourselves to the back of the sync list. 180 */ 181 vn_syncer_add_to_worklist(vp, syncdelay); 182 } 183 } 184 185 /* 186 * Do soft update processing. 187 */ 188 if (bioops.io_sync) 189 (*bioops.io_sync)(NULL); 190 191 /* 192 * The variable rushjob allows the kernel to speed up the 193 * processing of the filesystem syncer process. A rushjob 194 * value of N tells the filesystem syncer to process the next 195 * N seconds worth of work on its queue ASAP. Currently rushjob 196 * is used by the soft update code to speed up the filesystem 197 * syncer process when the incore state is getting so far 198 * ahead of the disk that the kernel memory pool is being 199 * threatened with exhaustion. 200 */ 201 if (rushjob > 0) { 202 rushjob -= 1; 203 continue; 204 } 205 /* 206 * If it has taken us less than a second to process the 207 * current work, then wait. Otherwise start right over 208 * again. We can still lose time if any single round 209 * takes more than two seconds, but it does not really 210 * matter as we are just trying to generally pace the 211 * filesystem activity. 212 */ 213 if (time.tv_sec == starttime) 214 tsleep(&lbolt, PPAUSE, "syncer", 0); 215 } 216 } 217 218 /* 219 * Request the syncer daemon to speed up its work. 220 * We never push it to speed up more than half of its 221 * normal turn time, otherwise it could take over the cpu. 222 */ 223 int 224 speedup_syncer() 225 { 226 int s; 227 228 s = splhigh(); 229 if (syncerproc && syncerproc->p_wchan == &lbolt) 230 setrunnable(syncerproc); 231 splx(s); 232 if (rushjob < syncdelay / 2) { 233 rushjob += 1; 234 stat_rush_requests += 1; 235 return 1; 236 } 237 return 0; 238 } 239 240 /* 241 * Routine to create and manage a filesystem syncer vnode. 242 */ 243 #define sync_close nullop 244 int sync_fsync __P((void *)); 245 int sync_inactive __P((void *)); 246 #define sync_reclaim nullop 247 #define sync_lock vop_generic_lock 248 #define sync_unlock vop_generic_unlock 249 int sync_print __P((void *)); 250 #define sync_islocked vop_generic_islocked 251 252 int (**sync_vnodeop_p) __P((void *)); 253 struct vnodeopv_entry_desc sync_vnodeop_entries[] = { 254 { &vop_default_desc, vn_default_error }, 255 { &vop_close_desc, sync_close }, /* close */ 256 { &vop_fsync_desc, sync_fsync }, /* fsync */ 257 { &vop_inactive_desc, sync_inactive }, /* inactive */ 258 { &vop_reclaim_desc, sync_reclaim }, /* reclaim */ 259 { &vop_lock_desc, sync_lock }, /* lock */ 260 { &vop_unlock_desc, sync_unlock }, /* unlock */ 261 { &vop_print_desc, sync_print }, /* print */ 262 { &vop_islocked_desc, sync_islocked }, /* islocked */ 263 { (struct vnodeop_desc*)NULL, (int(*) __P((void *)))NULL } 264 }; 265 struct vnodeopv_desc sync_vnodeop_opv_desc = 266 { &sync_vnodeop_p, sync_vnodeop_entries }; 267 268 /* 269 * Create a new filesystem syncer vnode for the specified mount point. 270 */ 271 int 272 vfs_allocate_syncvnode(mp) 273 struct mount *mp; 274 { 275 struct vnode *vp; 276 static long start, incr, next; 277 int error; 278 279 /* Allocate a new vnode */ 280 if ((error = getnewvnode(VT_VFS, mp, sync_vnodeop_p, &vp)) != 0) { 281 mp->mnt_syncer = NULL; 282 return (error); 283 } 284 vp->v_writecount = 1; 285 vp->v_type = VNON; 286 /* 287 * Place the vnode onto the syncer worklist. We attempt to 288 * scatter them about on the list so that they will go off 289 * at evenly distributed times even if all the filesystems 290 * are mounted at once. 291 */ 292 next += incr; 293 if (next == 0 || next > syncer_maxdelay) { 294 start /= 2; 295 incr /= 2; 296 if (start == 0) { 297 start = syncer_maxdelay / 2; 298 incr = syncer_maxdelay; 299 } 300 next = start; 301 } 302 vn_syncer_add_to_worklist(vp, next); 303 mp->mnt_syncer = vp; 304 return (0); 305 } 306 307 /* 308 * Do a lazy sync of the filesystem. 309 */ 310 int 311 sync_fsync(v) 312 void *v; 313 { 314 struct vop_fsync_args /* { 315 struct vnode *a_vp; 316 struct ucred *a_cred; 317 int a_waitfor; 318 struct proc *a_p; 319 } */ *ap = v; 320 struct vnode *syncvp = ap->a_vp; 321 struct mount *mp = syncvp->v_mount; 322 int asyncflag; 323 324 /* 325 * We only need to do something if this is a lazy evaluation. 326 */ 327 if (ap->a_waitfor != MNT_LAZY) 328 return (0); 329 330 /* 331 * Move ourselves to the back of the sync list. 332 */ 333 vn_syncer_add_to_worklist(syncvp, syncdelay); 334 335 /* 336 * Walk the list of vnodes pushing all that are dirty and 337 * not already on the sync list. 338 */ 339 simple_lock(&mountlist_slock); 340 if (vfs_busy(mp, LK_NOWAIT, &mountlist_slock, ap->a_p) == 0) { 341 asyncflag = mp->mnt_flag & MNT_ASYNC; 342 mp->mnt_flag &= ~MNT_ASYNC; 343 VFS_SYNC(mp, MNT_LAZY, ap->a_cred, ap->a_p); 344 if (asyncflag) 345 mp->mnt_flag |= MNT_ASYNC; 346 vfs_unbusy(mp, ap->a_p); 347 } else 348 simple_unlock(&mountlist_slock); 349 350 return (0); 351 } 352 353 /* 354 * The syncer vnode is no longer needed and is being decommissioned. 355 */ 356 int 357 sync_inactive(v) 358 void *v; 359 { 360 struct vop_inactive_args /* { 361 struct vnode *a_vp; 362 struct proc *a_p; 363 } */ *ap = v; 364 365 struct vnode *vp = ap->a_vp; 366 367 if (vp->v_usecount == 0) { 368 VOP_UNLOCK(vp, 0, ap->a_p); 369 return (0); 370 } 371 vp->v_mount->mnt_syncer = NULL; 372 LIST_REMOVE(vp, v_synclist); 373 vp->v_writecount = 0; 374 vput(vp); 375 return (0); 376 } 377 378 /* 379 * Print out a syncer vnode. 380 */ 381 int 382 sync_print(v) 383 void *v; 384 385 { 386 struct vop_print_args /* { 387 struct vnode *a_vp; 388 } */ *ap = v; 389 struct vnode *vp = ap->a_vp; 390 391 printf("syncer vnode"); 392 if (vp->v_vnlock != NULL) 393 lockmgr_printinfo(vp->v_vnlock); 394 printf("\n"); 395 return (0); 396 } 397