sys/kern/vfs_sync.c

   1 /*
   2  * Copyright (c) 1989, 1993
   3  *      The Regents of the University of California.  All rights reserved.
   4  * (c) UNIX System Laboratories, Inc.
   5  * All or some portions of this file are derived from material licensed
   6  * to the University of California by American Telephone and Telegraph
   7  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
   8  * the permission of UNIX System Laboratories, Inc.
   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  * 3. All advertising materials mentioning features or use of this software
  19  *    must display the following acknowledgement:
  20  *      This product includes software developed by the University of
  21  *      California, Berkeley and its contributors.
  22  * 4. 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  *      @(#)vfs_subr.c  8.31 (Berkeley) 5/26/95
  39  * $FreeBSD: src/sys/kern/vfs_subr.c,v 1.249.2.30 2003/04/04 20:35:57 tegge Exp $
  40  * $DragonFly: src/sys/kern/vfs_sync.c,v 1.18 2008/05/18 05:54:25 dillon Exp $
  41  */
  42
  43 /*
  44  * External virtual filesystem routines
  45  */
  46 #include "opt_ddb.h"
  47
  48 #include <sys/param.h>
  49 #include <sys/systm.h>
  50 #include <sys/buf.h>
  51 #include <sys/conf.h>
  52 #include <sys/dirent.h>
  53 #include <sys/domain.h>
  54 #include <sys/eventhandler.h>
  55 #include <sys/fcntl.h>
  56 #include <sys/kernel.h>
  57 #include <sys/kthread.h>
  58 #include <sys/malloc.h>
  59 #include <sys/mbuf.h>
  60 #include <sys/mount.h>
  61 #include <sys/proc.h>
  62 #include <sys/namei.h>
  63 #include <sys/reboot.h>
  64 #include <sys/socket.h>
  65 #include <sys/stat.h>
  66 #include <sys/sysctl.h>
  67 #include <sys/syslog.h>
  68 #include <sys/vmmeter.h>
  69 #include <sys/vnode.h>
  70
  71 #include <machine/limits.h>
  72
  73 #include <vm/vm.h>
  74 #include <vm/vm_object.h>
  75 #include <vm/vm_extern.h>
  76 #include <vm/vm_kern.h>
  77 #include <vm/pmap.h>
  78 #include <vm/vm_map.h>
  79 #include <vm/vm_page.h>
  80 #include <vm/vm_pager.h>
  81 #include <vm/vnode_pager.h>
  82
  83 #include <sys/buf2.h>
  84 #include <sys/thread2.h>
  85
  86 /*
  87  * The workitem queue.
  88  */
  89 #define SYNCER_MAXDELAY         32
  90 static int syncer_maxdelay = SYNCER_MAXDELAY;   /* maximum delay time */
  91 time_t syncdelay = 30;          /* max time to delay syncing data */
  92 SYSCTL_INT(_kern, OID_AUTO, syncdelay, CTLFLAG_RW,
  93                 &syncdelay, 0, "VFS data synchronization delay");
  94 time_t filedelay = 30;          /* time to delay syncing files */
  95 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW,
  96                 &filedelay, 0, "File synchronization delay");
  97 time_t dirdelay = 29;           /* time to delay syncing directories */
  98 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW,
  99                 &dirdelay, 0, "Directory synchronization delay");
 100 time_t metadelay = 28;          /* time to delay syncing metadata */
 101 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW,
 102                 &metadelay, 0, "VFS metadata synchronization delay");
 103 static int rushjob;                     /* number of slots to run ASAP */
 104 static int stat_rush_requests;  /* number of times I/O speeded up */
 105 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW,
 106                 &stat_rush_requests, 0, "");
 107
 108 static int syncer_delayno = 0;
 109 static long syncer_mask;
 110 LIST_HEAD(synclist, vnode);
 111 static struct synclist *syncer_workitem_pending;
 112
 113 /*
 114  * Called from vfsinit()
 115  */
 116 void
 117 vfs_sync_init(void)
 118 {
 119         syncer_workitem_pending = hashinit(syncer_maxdelay, M_DEVBUF,
 120                                             &syncer_mask);
 121         syncer_maxdelay = syncer_mask + 1;
 122 }
 123
 124 /*
 125  * The workitem queue.
 126  *
 127  * It is useful to delay writes of file data and filesystem metadata
 128  * for tens of seconds so that quickly created and deleted files need
 129  * not waste disk bandwidth being created and removed. To realize this,
 130  * we append vnodes to a "workitem" queue. When running with a soft
 131  * updates implementation, most pending metadata dependencies should
 132  * not wait for more than a few seconds. Thus, mounted on block devices
 133  * are delayed only about a half the time that file data is delayed.
 134  * Similarly, directory updates are more critical, so are only delayed
 135  * about a third the time that file data is delayed. Thus, there are
 136  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
 137  * one each second (driven off the filesystem syncer process). The
 138  * syncer_delayno variable indicates the next queue that is to be processed.
 139  * Items that need to be processed soon are placed in this queue:
 140  *
 141  *      syncer_workitem_pending[syncer_delayno]
 142  *
 143  * A delay of fifteen seconds is done by placing the request fifteen
 144  * entries later in the queue:
 145  *
 146  *      syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
 147  *
 148  */
 149
 150 /*
 151  * Add an item to the syncer work queue.
 152  */
 153 void
 154 vn_syncer_add_to_worklist(struct vnode *vp, int delay)
 155 {
 156         int slot;
 157
 158         crit_enter();
 159
 160         if (vp->v_flag & VONWORKLST) {
 161                 LIST_REMOVE(vp, v_synclist);
 162         }
 163
 164         if (delay > syncer_maxdelay - 2)
 165                 delay = syncer_maxdelay - 2;
 166         slot = (syncer_delayno + delay) & syncer_mask;
 167
 168         LIST_INSERT_HEAD(&syncer_workitem_pending[slot], vp, v_synclist);
 169         vp->v_flag |= VONWORKLST;
 170         crit_exit();
 171 }
 172
 173 struct  thread *updatethread;
 174 static void sched_sync (void);
 175 static struct kproc_desc up_kp = {
 176         "syncer",
 177         sched_sync,
 178         &updatethread
 179 };
 180 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp)
 181
 182 /*
 183  * System filesystem synchronizer daemon.
 184  */
 185 void
 186 sched_sync(void)
 187 {
 188         struct synclist *slp;
 189         struct vnode *vp;
 190         long starttime;
 191         struct thread *td = curthread;
 192
 193         EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, td,
 194             SHUTDOWN_PRI_LAST);
 195
 196         for (;;) {
 197                 kproc_suspend_loop();
 198
 199                 starttime = time_second;
 200
 201                 /*
 202                  * Push files whose dirty time has expired.  Be careful
 203                  * of interrupt race on slp queue.
 204                  */
 205                 crit_enter();
 206                 slp = &syncer_workitem_pending[syncer_delayno];
 207                 syncer_delayno += 1;
 208                 if (syncer_delayno == syncer_maxdelay)
 209                         syncer_delayno = 0;
 210                 crit_exit();
 211
 212                 while ((vp = LIST_FIRST(slp)) != NULL) {
 213                         if (vget(vp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
 214                                 VOP_FSYNC(vp, MNT_LAZY);
 215                                 vput(vp);
 216                         }
 217                         crit_enter();
 218
 219                         /*
 220                          * If the vnode is still at the head of the list
 221                          * we were not able to completely flush it.  To
 222                          * give other vnodes a fair shake we move it to
 223                          * a later slot.
 224                          *
 225                          * Note that v_tag VT_VFS vnodes can remain on the
 226                          * worklist with no dirty blocks, but sync_fsync()
 227                          * moves it to a later slot so we will never see it
 228                          * here.
 229                          */
 230                         if (LIST_FIRST(slp) == vp) {
 231                                 if (RB_EMPTY(&vp->v_rbdirty_tree) &&
 232                                     !vn_isdisk(vp, NULL)) {
 233                                         panic("sched_sync: fsync failed vp %p tag %d", vp, vp->v_tag);
 234                                 }
 235                                 vn_syncer_add_to_worklist(vp, syncdelay);
 236                         }
 237                         crit_exit();
 238                 }
 239
 240                 /*
 241                  * Do sync processing for each mount.
 242                  */
 243                 bio_ops_sync(NULL);
 244
 245                 /*
 246                  * The variable rushjob allows the kernel to speed up the
 247                  * processing of the filesystem syncer process. A rushjob
 248                  * value of N tells the filesystem syncer to process the next
 249                  * N seconds worth of work on its queue ASAP. Currently rushjob
 250                  * is used by the soft update code to speed up the filesystem
 251                  * syncer process when the incore state is getting so far
 252                  * ahead of the disk that the kernel memory pool is being
 253                  * threatened with exhaustion.
 254                  */
 255                 if (rushjob > 0) {
 256                         rushjob -= 1;
 257                         continue;
 258                 }
 259                 /*
 260                  * If it has taken us less than a second to process the
 261                  * current work, then wait. Otherwise start right over
 262                  * again. We can still lose time if any single round
 263                  * takes more than two seconds, but it does not really
 264                  * matter as we are just trying to generally pace the
 265                  * filesystem activity.
 266                  */
 267                 if (time_second == starttime)
 268                         tsleep(&lbolt_syncer, 0, "syncer", 0);
 269         }
 270 }
 271
 272 /*
 273  * Request the syncer daemon to speed up its work.
 274  * We never push it to speed up more than half of its
 275  * normal turn time, otherwise it could take over the cpu.
 276  *
 277  * YYY wchan field protected by the BGL.
 278  */
 279 int
 280 speedup_syncer(void)
 281 {
 282         /*
 283          * Don't bother protecting the test.  unsleep_and_wakeup_thread()
 284          * will only do something real if the thread is in the right state.
 285          */
 286         wakeup(&lbolt_syncer);
 287         if (rushjob < syncdelay / 2) {
 288                 rushjob += 1;
 289                 stat_rush_requests += 1;
 290                 return (1);
 291         }
 292         return(0);
 293 }
 294
 295 /*
 296  * Routine to create and manage a filesystem syncer vnode.
 297  */
 298 static int sync_close(struct vop_close_args *);
 299 static int sync_fsync(struct vop_fsync_args *);
 300 static int sync_inactive(struct vop_inactive_args *);
 301 static int sync_reclaim (struct vop_reclaim_args *);
 302 static int sync_print(struct vop_print_args *);
 303
 304 static struct vop_ops sync_vnode_vops = {
 305         .vop_default =  vop_eopnotsupp,
 306         .vop_close =    sync_close,
 307         .vop_fsync =    sync_fsync,
 308         .vop_inactive = sync_inactive,
 309         .vop_reclaim =  sync_reclaim,
 310         .vop_print =    sync_print,
 311 };
 312
 313 static struct vop_ops *sync_vnode_vops_p = &sync_vnode_vops;
 314
 315 VNODEOP_SET(sync_vnode_vops);
 316
 317 /*
 318  * Create a new filesystem syncer vnode for the specified mount point.
 319  * This vnode is placed on the worklist and is responsible for sync'ing
 320  * the filesystem.
 321  *
 322  * NOTE: read-only mounts are also placed on the worklist.  The filesystem
 323  * sync code is also responsible for cleaning up vnodes.
 324  */
 325 int
 326 vfs_allocate_syncvnode(struct mount *mp)
 327 {
 328         struct vnode *vp;
 329         static long start, incr, next;
 330         int error;
 331
 332         /* Allocate a new vnode */
 333         error = getspecialvnode(VT_VFS, mp, &sync_vnode_vops_p, &vp, 0, 0);
 334         if (error) {
 335                 mp->mnt_syncer = NULL;
 336                 return (error);
 337         }
 338         vp->v_type = VNON;
 339         /*
 340          * Place the vnode onto the syncer worklist. We attempt to
 341          * scatter them about on the list so that they will go off
 342          * at evenly distributed times even if all the filesystems
 343          * are mounted at once.
 344          */
 345         next += incr;
 346         if (next == 0 || next > syncer_maxdelay) {
 347                 start /= 2;
 348                 incr /= 2;
 349                 if (start == 0) {
 350                         start = syncer_maxdelay / 2;
 351                         incr = syncer_maxdelay;
 352                 }
 353                 next = start;
 354         }
 355         vn_syncer_add_to_worklist(vp, syncdelay > 0 ? next % syncdelay : 0);
 356
 357         /*
 358          * The mnt_syncer field inherits the vnode reference, which is
 359          * held until later decomissioning.
 360          */
 361         mp->mnt_syncer = vp;
 362         vx_unlock(vp);
 363         return (0);
 364 }
 365
 366 static int
 367 sync_close(struct vop_close_args *ap)
 368 {
 369         return (0);
 370 }
 371
 372 /*
 373  * Do a lazy sync of the filesystem.
 374  *
 375  * sync_fsync { struct vnode *a_vp, int a_waitfor }
 376  */
 377 static int
 378 sync_fsync(struct vop_fsync_args *ap)
 379 {
 380         struct vnode *syncvp = ap->a_vp;
 381         struct mount *mp = syncvp->v_mount;
 382         int asyncflag;
 383
 384         /*
 385          * We only need to do something if this is a lazy evaluation.
 386          */
 387         if (ap->a_waitfor != MNT_LAZY)
 388                 return (0);
 389
 390         /*
 391          * Move ourselves to the back of the sync list.
 392          */
 393         vn_syncer_add_to_worklist(syncvp, syncdelay);
 394
 395         /*
 396          * Walk the list of vnodes pushing all that are dirty and
 397          * not already on the sync list, and freeing vnodes which have
 398          * no refs and whos VM objects are empty.  vfs_msync() handles
 399          * the VM issues and must be called whether the mount is readonly
 400          * or not.
 401          */
 402         if (vfs_busy(mp, LK_NOWAIT) != 0)
 403                 return (0);
 404         if (mp->mnt_flag & MNT_RDONLY) {
 405                 vfs_msync(mp, MNT_NOWAIT);
 406         } else {
 407                 asyncflag = mp->mnt_flag & MNT_ASYNC;
 408                 mp->mnt_flag &= ~MNT_ASYNC;     /* ZZZ hack */
 409                 vfs_msync(mp, MNT_NOWAIT);
 410                 VFS_SYNC(mp, MNT_LAZY);
 411                 if (asyncflag)
 412                         mp->mnt_flag |= MNT_ASYNC;
 413         }
 414         vfs_unbusy(mp);
 415         return (0);
 416 }
 417
 418 /*
 419  * The syncer vnode is no longer referenced.
 420  *
 421  * sync_inactive { struct vnode *a_vp, struct proc *a_p }
 422  */
 423 static int
 424 sync_inactive(struct vop_inactive_args *ap)
 425 {
 426         vgone_vxlocked(ap->a_vp);
 427         return (0);
 428 }
 429
 430 /*
 431  * The syncer vnode is no longer needed and is being decommissioned.
 432  * This can only occur when the last reference has been released on
 433  * mp->mnt_syncer, so mp->mnt_syncer had better be NULL.
 434  *
 435  * Modifications to the worklist must be protected with a critical
 436  * section.
 437  *
 438  *      sync_reclaim { struct vnode *a_vp }
 439  */
 440 static int
 441 sync_reclaim(struct vop_reclaim_args *ap)
 442 {
 443         struct vnode *vp = ap->a_vp;
 444
 445         crit_enter();
 446         KKASSERT(vp->v_mount->mnt_syncer != vp);
 447         if (vp->v_flag & VONWORKLST) {
 448                 LIST_REMOVE(vp, v_synclist);
 449                 vp->v_flag &= ~VONWORKLST;
 450         }
 451         crit_exit();
 452
 453         return (0);
 454 }
 455
 456 /*
 457  * Print out a syncer vnode.
 458  *
 459  *      sync_print { struct vnode *a_vp }
 460  */
 461 static int
 462 sync_print(struct vop_print_args *ap)
 463 {
 464         struct vnode *vp = ap->a_vp;
 465
 466         kprintf("syncer vnode");
 467         lockmgr_printinfo(&vp->v_lock);
 468         kprintf("\n");
 469         return (0);
 470 }
 471