fs/xfs/linux-2.6/xfs_sync.c

   1 /*
   2  * Copyright (c) 2000-2005 Silicon Graphics, Inc.
   3  * All Rights Reserved.
   4  *
   5  * This program is free software; you can redistribute it and/or
   6  * modify it under the terms of the GNU General Public License as
   7  * published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope that it would be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write the Free Software Foundation,
  16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  17  */
  18 #include "xfs.h"
  19 #include "xfs_fs.h"
  20 #include "xfs_types.h"
  21 #include "xfs_bit.h"
  22 #include "xfs_log.h"
  23 #include "xfs_inum.h"
  24 #include "xfs_trans.h"
  25 #include "xfs_sb.h"
  26 #include "xfs_ag.h"
  27 #include "xfs_dir2.h"
  28 #include "xfs_dmapi.h"
  29 #include "xfs_mount.h"
  30 #include "xfs_bmap_btree.h"
  31 #include "xfs_alloc_btree.h"
  32 #include "xfs_ialloc_btree.h"
  33 #include "xfs_btree.h"
  34 #include "xfs_dir2_sf.h"
  35 #include "xfs_attr_sf.h"
  36 #include "xfs_inode.h"
  37 #include "xfs_dinode.h"
  38 #include "xfs_error.h"
  39 #include "xfs_mru_cache.h"
  40 #include "xfs_filestream.h"
  41 #include "xfs_vnodeops.h"
  42 #include "xfs_utils.h"
  43 #include "xfs_buf_item.h"
  44 #include "xfs_inode_item.h"
  45 #include "xfs_rw.h"
  46
  47 #include <linux/kthread.h>
  48 #include <linux/freezer.h>
  49
  50 /*
  51  * xfs_sync flushes any pending I/O to file system vfsp.
  52  *
  53  * This routine is called by vfs_sync() to make sure that things make it
  54  * out to disk eventually, on sync() system calls to flush out everything,
  55  * and when the file system is unmounted.  For the vfs_sync() case, all
  56  * we really need to do is sync out the log to make all of our meta-data
  57  * updates permanent (except for timestamps).  For calls from pflushd(),
  58  * dirty pages are kept moving by calling pdflush() on the inodes
  59  * containing them.  We also flush the inodes that we can lock without
  60  * sleeping and the superblock if we can lock it without sleeping from
  61  * vfs_sync() so that items at the tail of the log are always moving out.
  62  *
  63  * Flags:
  64  *      SYNC_BDFLUSH - We're being called from vfs_sync() so we don't want
  65  *                     to sleep if we can help it.  All we really need
  66  *                     to do is ensure that the log is synced at least
  67  *                     periodically.  We also push the inodes and
  68  *                     superblock if we can lock them without sleeping
  69  *                      and they are not pinned.
  70  *      SYNC_ATTR    - We need to flush the inodes.  If SYNC_BDFLUSH is not
  71  *                     set, then we really want to lock each inode and flush
  72  *                     it.
  73  *      SYNC_WAIT    - All the flushes that take place in this call should
  74  *                     be synchronous.
  75  *      SYNC_DELWRI  - This tells us to push dirty pages associated with
  76  *                     inodes.  SYNC_WAIT and SYNC_BDFLUSH are used to
  77  *                     determine if they should be flushed sync, async, or
  78  *                     delwri.
  79  *      SYNC_CLOSE   - This flag is passed when the system is being
  80  *                     unmounted.  We should sync and invalidate everything.
  81  *      SYNC_FSDATA  - This indicates that the caller would like to make
  82  *                     sure the superblock is safe on disk.  We can ensure
  83  *                     this by simply making sure the log gets flushed
  84  *                     if SYNC_BDFLUSH is set, and by actually writing it
  85  *                     out otherwise.
  86  *      SYNC_IOWAIT  - The caller wants us to wait for all data I/O to complete
  87  *                     before we return (including direct I/O). Forms the drain
  88  *                     side of the write barrier needed to safely quiesce the
  89  *                     filesystem.
  90  *
  91  */
  92 int
  93 xfs_sync(
  94         xfs_mount_t     *mp,
  95         int             flags)
  96 {
  97         int             error;
  98
  99         /*
 100          * Get the Quota Manager to flush the dquots.
 101          *
 102          * If XFS quota support is not enabled or this filesystem
 103          * instance does not use quotas XFS_QM_DQSYNC will always
 104          * return zero.
 105          */
 106         error = XFS_QM_DQSYNC(mp, flags);
 107         if (error) {
 108                 /*
 109                  * If we got an IO error, we will be shutting down.
 110                  * So, there's nothing more for us to do here.
 111                  */
 112                 ASSERT(error != EIO || XFS_FORCED_SHUTDOWN(mp));
 113                 if (XFS_FORCED_SHUTDOWN(mp))
 114                         return XFS_ERROR(error);
 115         }
 116
 117         if (flags & SYNC_IOWAIT)
 118                 xfs_filestream_flush(mp);
 119
 120         return xfs_syncsub(mp, flags, NULL);
 121 }
 122
 123 /*
 124  * xfs sync routine for internal use
 125  *
 126  * This routine supports all of the flags defined for the generic vfs_sync
 127  * interface as explained above under xfs_sync.
 128  *
 129  */
 130 int
 131 xfs_sync_inodes(
 132         xfs_mount_t     *mp,
 133         int             flags,
 134         int             *bypassed)
 135 {
 136         xfs_inode_t     *ip = NULL;
 137         struct inode    *vp = NULL;
 138         int             error;
 139         int             last_error;
 140         uint64_t        fflag;
 141         uint            lock_flags;
 142         uint            base_lock_flags;
 143         boolean_t       mount_locked;
 144         boolean_t       vnode_refed;
 145         int             preempt;
 146         xfs_iptr_t      *ipointer;
 147 #ifdef DEBUG
 148         boolean_t       ipointer_in = B_FALSE;
 149
 150 #define IPOINTER_SET    ipointer_in = B_TRUE
 151 #define IPOINTER_CLR    ipointer_in = B_FALSE
 152 #else
 153 #define IPOINTER_SET
 154 #define IPOINTER_CLR
 155 #endif
 156
 157
 158 /* Insert a marker record into the inode list after inode ip. The list
 159  * must be locked when this is called. After the call the list will no
 160  * longer be locked.
 161  */
 162 #define IPOINTER_INSERT(ip, mp) { \
 163                 ASSERT(ipointer_in == B_FALSE); \
 164                 ipointer->ip_mnext = ip->i_mnext; \
 165                 ipointer->ip_mprev = ip; \
 166                 ip->i_mnext = (xfs_inode_t *)ipointer; \
 167                 ipointer->ip_mnext->i_mprev = (xfs_inode_t *)ipointer; \
 168                 preempt = 0; \
 169                 XFS_MOUNT_IUNLOCK(mp); \
 170                 mount_locked = B_FALSE; \
 171                 IPOINTER_SET; \
 172         }
 173
 174 /* Remove the marker from the inode list. If the marker was the only item
 175  * in the list then there are no remaining inodes and we should zero out
 176  * the whole list. If we are the current head of the list then move the head
 177  * past us.
 178  */
 179 #define IPOINTER_REMOVE(ip, mp) { \
 180                 ASSERT(ipointer_in == B_TRUE); \
 181                 if (ipointer->ip_mnext != (xfs_inode_t *)ipointer) { \
 182                         ip = ipointer->ip_mnext; \
 183                         ip->i_mprev = ipointer->ip_mprev; \
 184                         ipointer->ip_mprev->i_mnext = ip; \
 185                         if (mp->m_inodes == (xfs_inode_t *)ipointer) { \
 186                                 mp->m_inodes = ip; \
 187                         } \
 188                 } else { \
 189                         ASSERT(mp->m_inodes == (xfs_inode_t *)ipointer); \
 190                         mp->m_inodes = NULL; \
 191                         ip = NULL; \
 192                 } \
 193                 IPOINTER_CLR; \
 194         }
 195
 196 #define XFS_PREEMPT_MASK        0x7f
 197
 198         ASSERT(!(flags & SYNC_BDFLUSH));
 199
 200         if (bypassed)
 201                 *bypassed = 0;
 202         if (mp->m_flags & XFS_MOUNT_RDONLY)
 203                 return 0;
 204         error = 0;
 205         last_error = 0;
 206         preempt = 0;
 207
 208         /* Allocate a reference marker */
 209         ipointer = (xfs_iptr_t *)kmem_zalloc(sizeof(xfs_iptr_t), KM_SLEEP);
 210
 211         fflag = XFS_B_ASYNC;            /* default is don't wait */
 212         if (flags & SYNC_DELWRI)
 213                 fflag = XFS_B_DELWRI;
 214         if (flags & SYNC_WAIT)
 215                 fflag = 0;              /* synchronous overrides all */
 216
 217         base_lock_flags = XFS_ILOCK_SHARED;
 218         if (flags & (SYNC_DELWRI | SYNC_CLOSE)) {
 219                 /*
 220                  * We need the I/O lock if we're going to call any of
 221                  * the flush/inval routines.
 222                  */
 223                 base_lock_flags |= XFS_IOLOCK_SHARED;
 224         }
 225
 226         XFS_MOUNT_ILOCK(mp);
 227
 228         ip = mp->m_inodes;
 229
 230         mount_locked = B_TRUE;
 231         vnode_refed  = B_FALSE;
 232
 233         IPOINTER_CLR;
 234
 235         do {
 236                 ASSERT(ipointer_in == B_FALSE);
 237                 ASSERT(vnode_refed == B_FALSE);
 238
 239                 lock_flags = base_lock_flags;
 240
 241                 /*
 242                  * There were no inodes in the list, just break out
 243                  * of the loop.
 244                  */
 245                 if (ip == NULL) {
 246                         break;
 247                 }
 248
 249                 /*
 250                  * We found another sync thread marker - skip it
 251                  */
 252                 if (ip->i_mount == NULL) {
 253                         ip = ip->i_mnext;
 254                         continue;
 255                 }
 256
 257                 vp = VFS_I(ip);
 258
 259                 /*
 260                  * If the vnode is gone then this is being torn down,
 261                  * call reclaim if it is flushed, else let regular flush
 262                  * code deal with it later in the loop.
 263                  */
 264
 265                 if (vp == NULL) {
 266                         /* Skip ones already in reclaim */
 267                         if (ip->i_flags & XFS_IRECLAIM) {
 268                                 ip = ip->i_mnext;
 269                                 continue;
 270                         }
 271                         if (xfs_ilock_nowait(ip, XFS_ILOCK_EXCL) == 0) {
 272                                 ip = ip->i_mnext;
 273                         } else if ((xfs_ipincount(ip) == 0) &&
 274                                     xfs_iflock_nowait(ip)) {
 275                                 IPOINTER_INSERT(ip, mp);
 276
 277                                 xfs_finish_reclaim(ip, 1,
 278                                                 XFS_IFLUSH_DELWRI_ELSE_ASYNC);
 279
 280                                 XFS_MOUNT_ILOCK(mp);
 281                                 mount_locked = B_TRUE;
 282                                 IPOINTER_REMOVE(ip, mp);
 283                         } else {
 284                                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 285                                 ip = ip->i_mnext;
 286                         }
 287                         continue;
 288                 }
 289
 290                 if (VN_BAD(vp)) {
 291                         ip = ip->i_mnext;
 292                         continue;
 293                 }
 294
 295                 if (XFS_FORCED_SHUTDOWN(mp) && !(flags & SYNC_CLOSE)) {
 296                         XFS_MOUNT_IUNLOCK(mp);
 297                         kmem_free(ipointer);
 298                         return 0;
 299                 }
 300
 301                 /*
 302                  * Try to lock without sleeping.  We're out of order with
 303                  * the inode list lock here, so if we fail we need to drop
 304                  * the mount lock and try again.  If we're called from
 305                  * bdflush() here, then don't bother.
 306                  *
 307                  * The inode lock here actually coordinates with the
 308                  * almost spurious inode lock in xfs_ireclaim() to prevent
 309                  * the vnode we handle here without a reference from
 310                  * being freed while we reference it.  If we lock the inode
 311                  * while it's on the mount list here, then the spurious inode
 312                  * lock in xfs_ireclaim() after the inode is pulled from
 313                  * the mount list will sleep until we release it here.
 314                  * This keeps the vnode from being freed while we reference
 315                  * it.
 316                  */
 317                 if (xfs_ilock_nowait(ip, lock_flags) == 0) {
 318                         if (vp == NULL) {
 319                                 ip = ip->i_mnext;
 320                                 continue;
 321                         }
 322
 323                         vp = vn_grab(vp);
 324                         if (vp == NULL) {
 325                                 ip = ip->i_mnext;
 326                                 continue;
 327                         }
 328
 329                         IPOINTER_INSERT(ip, mp);
 330                         xfs_ilock(ip, lock_flags);
 331
 332                         ASSERT(vp == VFS_I(ip));
 333                         ASSERT(ip->i_mount == mp);
 334
 335                         vnode_refed = B_TRUE;
 336                 }
 337
 338                 /* From here on in the loop we may have a marker record
 339                  * in the inode list.
 340                  */
 341
 342                 /*
 343                  * If we have to flush data or wait for I/O completion
 344                  * we need to drop the ilock that we currently hold.
 345                  * If we need to drop the lock, insert a marker if we
 346                  * have not already done so.
 347                  */
 348                 if ((flags & (SYNC_CLOSE|SYNC_IOWAIT)) ||
 349                     ((flags & SYNC_DELWRI) && VN_DIRTY(vp))) {
 350                         if (mount_locked) {
 351                                 IPOINTER_INSERT(ip, mp);
 352                         }
 353                         xfs_iunlock(ip, XFS_ILOCK_SHARED);
 354
 355                         if (flags & SYNC_CLOSE) {
 356                                 /* Shutdown case. Flush and invalidate. */
 357                                 if (XFS_FORCED_SHUTDOWN(mp))
 358                                         xfs_tosspages(ip, 0, -1,
 359                                                              FI_REMAPF);
 360                                 else
 361                                         error = xfs_flushinval_pages(ip,
 362                                                         0, -1, FI_REMAPF);
 363                         } else if ((flags & SYNC_DELWRI) && VN_DIRTY(vp)) {
 364                                 error = xfs_flush_pages(ip, 0,
 365                                                         -1, fflag, FI_NONE);
 366                         }
 367
 368                         /*
 369                          * When freezing, we need to wait ensure all I/O (including direct
 370                          * I/O) is complete to ensure no further data modification can take
 371                          * place after this point
 372                          */
 373                         if (flags & SYNC_IOWAIT)
 374                                 vn_iowait(ip);
 375
 376                         xfs_ilock(ip, XFS_ILOCK_SHARED);
 377                 }
 378
 379                 if ((flags & SYNC_ATTR) &&
 380                     (ip->i_update_core ||
 381                      (ip->i_itemp && ip->i_itemp->ili_format.ilf_fields))) {
 382                         if (mount_locked)
 383                                 IPOINTER_INSERT(ip, mp);
 384
 385                         if (flags & SYNC_WAIT) {
 386                                 xfs_iflock(ip);
 387                                 error = xfs_iflush(ip, XFS_IFLUSH_SYNC);
 388
 389                         /*
 390                          * If we can't acquire the flush lock, then the inode
 391                          * is already being flushed so don't bother waiting.
 392                          *
 393                          * If we can lock it then do a delwri flush so we can
 394                          * combine multiple inode flushes in each disk write.
 395                          */
 396                         } else if (xfs_iflock_nowait(ip)) {
 397                                 error = xfs_iflush(ip, XFS_IFLUSH_DELWRI);
 398                         } else if (bypassed) {
 399                                 (*bypassed)++;
 400                         }
 401                 }
 402
 403                 if (lock_flags != 0) {
 404                         xfs_iunlock(ip, lock_flags);
 405                 }
 406
 407                 if (vnode_refed) {
 408                         /*
 409                          * If we had to take a reference on the vnode
 410                          * above, then wait until after we've unlocked
 411                          * the inode to release the reference.  This is
 412                          * because we can be already holding the inode
 413                          * lock when IRELE() calls xfs_inactive().
 414                          *
 415                          * Make sure to drop the mount lock before calling
 416                          * IRELE() so that we don't trip over ourselves if
 417                          * we have to go for the mount lock again in the
 418                          * inactive code.
 419                          */
 420                         if (mount_locked) {
 421                                 IPOINTER_INSERT(ip, mp);
 422                         }
 423
 424                         IRELE(ip);
 425
 426                         vnode_refed = B_FALSE;
 427                 }
 428
 429                 if (error) {
 430                         last_error = error;
 431                 }
 432
 433                 /*
 434                  * bail out if the filesystem is corrupted.
 435                  */
 436                 if (error == EFSCORRUPTED)  {
 437                         if (!mount_locked) {
 438                                 XFS_MOUNT_ILOCK(mp);
 439                                 IPOINTER_REMOVE(ip, mp);
 440                         }
 441                         XFS_MOUNT_IUNLOCK(mp);
 442                         ASSERT(ipointer_in == B_FALSE);
 443                         kmem_free(ipointer);
 444                         return XFS_ERROR(error);
 445                 }
 446
 447                 /* Let other threads have a chance at the mount lock
 448                  * if we have looped many times without dropping the
 449                  * lock.
 450                  */
 451                 if ((++preempt & XFS_PREEMPT_MASK) == 0) {
 452                         if (mount_locked) {
 453                                 IPOINTER_INSERT(ip, mp);
 454                         }
 455                 }
 456
 457                 if (mount_locked == B_FALSE) {
 458                         XFS_MOUNT_ILOCK(mp);
 459                         mount_locked = B_TRUE;
 460                         IPOINTER_REMOVE(ip, mp);
 461                         continue;
 462                 }
 463
 464                 ASSERT(ipointer_in == B_FALSE);
 465                 ip = ip->i_mnext;
 466
 467         } while (ip != mp->m_inodes);
 468
 469         XFS_MOUNT_IUNLOCK(mp);
 470
 471         ASSERT(ipointer_in == B_FALSE);
 472
 473         kmem_free(ipointer);
 474         return XFS_ERROR(last_error);
 475 }
 476
 477 /*
 478  * xfs sync routine for internal use
 479  *
 480  * This routine supports all of the flags defined for the generic vfs_sync
 481  * interface as explained above under xfs_sync.
 482  *
 483  */
 484 int
 485 xfs_syncsub(
 486         xfs_mount_t     *mp,
 487         int             flags,
 488         int             *bypassed)
 489 {
 490         int             error = 0;
 491         int             last_error = 0;
 492         uint            log_flags = XFS_LOG_FORCE;
 493         xfs_buf_t       *bp;
 494         xfs_buf_log_item_t      *bip;
 495
 496         /*
 497          * Sync out the log.  This ensures that the log is periodically
 498          * flushed even if there is not enough activity to fill it up.
 499          */
 500         if (flags & SYNC_WAIT)
 501                 log_flags |= XFS_LOG_SYNC;
 502
 503         xfs_log_force(mp, (xfs_lsn_t)0, log_flags);
 504
 505         if (flags & (SYNC_ATTR|SYNC_DELWRI)) {
 506                 if (flags & SYNC_BDFLUSH)
 507                         xfs_finish_reclaim_all(mp, 1, XFS_IFLUSH_DELWRI_ELSE_ASYNC);
 508                 else
 509                         error = xfs_sync_inodes(mp, flags, bypassed);
 510         }
 511
 512         /*
 513          * Flushing out dirty data above probably generated more
 514          * log activity, so if this isn't vfs_sync() then flush
 515          * the log again.
 516          */
 517         if (flags & SYNC_DELWRI) {
 518                 xfs_log_force(mp, (xfs_lsn_t)0, log_flags);
 519         }
 520
 521         if (flags & SYNC_FSDATA) {
 522                 /*
 523                  * If this is vfs_sync() then only sync the superblock
 524                  * if we can lock it without sleeping and it is not pinned.
 525                  */
 526                 if (flags & SYNC_BDFLUSH) {
 527                         bp = xfs_getsb(mp, XFS_BUF_TRYLOCK);
 528                         if (bp != NULL) {
 529                                 bip = XFS_BUF_FSPRIVATE(bp,xfs_buf_log_item_t*);
 530                                 if ((bip != NULL) &&
 531                                     xfs_buf_item_dirty(bip)) {
 532                                         if (!(XFS_BUF_ISPINNED(bp))) {
 533                                                 XFS_BUF_ASYNC(bp);
 534                                                 error = xfs_bwrite(mp, bp);
 535                                         } else {
 536                                                 xfs_buf_relse(bp);
 537                                         }
 538                                 } else {
 539                                         xfs_buf_relse(bp);
 540                                 }
 541                         }
 542                 } else {
 543                         bp = xfs_getsb(mp, 0);
 544                         /*
 545                          * If the buffer is pinned then push on the log so
 546                          * we won't get stuck waiting in the write for
 547                          * someone, maybe ourselves, to flush the log.
 548                          * Even though we just pushed the log above, we
 549                          * did not have the superblock buffer locked at
 550                          * that point so it can become pinned in between
 551                          * there and here.
 552                          */
 553                         if (XFS_BUF_ISPINNED(bp))
 554                                 xfs_log_force(mp, (xfs_lsn_t)0, XFS_LOG_FORCE);
 555                         if (flags & SYNC_WAIT)
 556                                 XFS_BUF_UNASYNC(bp);
 557                         else
 558                                 XFS_BUF_ASYNC(bp);
 559                         error = xfs_bwrite(mp, bp);
 560                 }
 561                 if (error) {
 562                         last_error = error;
 563                 }
 564         }
 565
 566         /*
 567          * Now check to see if the log needs a "dummy" transaction.
 568          */
 569         if (!(flags & SYNC_REMOUNT) && xfs_log_need_covered(mp)) {
 570                 xfs_trans_t *tp;
 571                 xfs_inode_t *ip;
 572
 573                 /*
 574                  * Put a dummy transaction in the log to tell
 575                  * recovery that all others are OK.
 576                  */
 577                 tp = xfs_trans_alloc(mp, XFS_TRANS_DUMMY1);
 578                 if ((error = xfs_trans_reserve(tp, 0,
 579                                 XFS_ICHANGE_LOG_RES(mp),
 580                                 0, 0, 0)))  {
 581                         xfs_trans_cancel(tp, 0);
 582                         return error;
 583                 }
 584
 585                 ip = mp->m_rootip;
 586                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 587
 588                 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
 589                 xfs_trans_ihold(tp, ip);
 590                 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
 591                 error = xfs_trans_commit(tp, 0);
 592                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 593                 xfs_log_force(mp, (xfs_lsn_t)0, log_flags);
 594         }
 595
 596         /*
 597          * When shutting down, we need to insure that the AIL is pushed
 598          * to disk or the filesystem can appear corrupt from the PROM.
 599          */
 600         if ((flags & (SYNC_CLOSE|SYNC_WAIT)) == (SYNC_CLOSE|SYNC_WAIT)) {
 601                 XFS_bflush(mp->m_ddev_targp);
 602                 if (mp->m_rtdev_targp) {
 603                         XFS_bflush(mp->m_rtdev_targp);
 604                 }
 605         }
 606
 607         return XFS_ERROR(last_error);
 608 }
 609
 610 /*
 611  * Enqueue a work item to be picked up by the vfs xfssyncd thread.
 612  * Doing this has two advantages:
 613  * - It saves on stack space, which is tight in certain situations
 614  * - It can be used (with care) as a mechanism to avoid deadlocks.
 615  * Flushing while allocating in a full filesystem requires both.
 616  */
 617 STATIC void
 618 xfs_syncd_queue_work(
 619         struct xfs_mount *mp,
 620         void            *data,
 621         void            (*syncer)(struct xfs_mount *, void *))
 622 {
 623         struct bhv_vfs_sync_work *work;
 624
 625         work = kmem_alloc(sizeof(struct bhv_vfs_sync_work), KM_SLEEP);
 626         INIT_LIST_HEAD(&work->w_list);
 627         work->w_syncer = syncer;
 628         work->w_data = data;
 629         work->w_mount = mp;
 630         spin_lock(&mp->m_sync_lock);
 631         list_add_tail(&work->w_list, &mp->m_sync_list);
 632         spin_unlock(&mp->m_sync_lock);
 633         wake_up_process(mp->m_sync_task);
 634 }
 635
 636 /*
 637  * Flush delayed allocate data, attempting to free up reserved space
 638  * from existing allocations.  At this point a new allocation attempt
 639  * has failed with ENOSPC and we are in the process of scratching our
 640  * heads, looking about for more room...
 641  */
 642 STATIC void
 643 xfs_flush_inode_work(
 644         struct xfs_mount *mp,
 645         void            *arg)
 646 {
 647         struct inode    *inode = arg;
 648         filemap_flush(inode->i_mapping);
 649         iput(inode);
 650 }
 651
 652 void
 653 xfs_flush_inode(
 654         xfs_inode_t     *ip)
 655 {
 656         struct inode    *inode = VFS_I(ip);
 657
 658         igrab(inode);
 659         xfs_syncd_queue_work(ip->i_mount, inode, xfs_flush_inode_work);
 660         delay(msecs_to_jiffies(500));
 661 }
 662
 663 /*
 664  * This is the "bigger hammer" version of xfs_flush_inode_work...
 665  * (IOW, "If at first you don't succeed, use a Bigger Hammer").
 666  */
 667 STATIC void
 668 xfs_flush_device_work(
 669         struct xfs_mount *mp,
 670         void            *arg)
 671 {
 672         struct inode    *inode = arg;
 673         sync_blockdev(mp->m_super->s_bdev);
 674         iput(inode);
 675 }
 676
 677 void
 678 xfs_flush_device(
 679         xfs_inode_t     *ip)
 680 {
 681         struct inode    *inode = VFS_I(ip);
 682
 683         igrab(inode);
 684         xfs_syncd_queue_work(ip->i_mount, inode, xfs_flush_device_work);
 685         delay(msecs_to_jiffies(500));
 686         xfs_log_force(ip->i_mount, (xfs_lsn_t)0, XFS_LOG_FORCE|XFS_LOG_SYNC);
 687 }
 688
 689 STATIC void
 690 xfs_sync_worker(
 691         struct xfs_mount *mp,
 692         void            *unused)
 693 {
 694         int             error;
 695
 696         if (!(mp->m_flags & XFS_MOUNT_RDONLY))
 697                 error = xfs_sync(mp, SYNC_FSDATA | SYNC_BDFLUSH | SYNC_ATTR);
 698         mp->m_sync_seq++;
 699         wake_up(&mp->m_wait_single_sync_task);
 700 }
 701
 702 STATIC int
 703 xfssyncd(
 704         void                    *arg)
 705 {
 706         struct xfs_mount        *mp = arg;
 707         long                    timeleft;
 708         bhv_vfs_sync_work_t     *work, *n;
 709         LIST_HEAD               (tmp);
 710
 711         set_freezable();
 712         timeleft = xfs_syncd_centisecs * msecs_to_jiffies(10);
 713         for (;;) {
 714                 timeleft = schedule_timeout_interruptible(timeleft);
 715                 /* swsusp */
 716                 try_to_freeze();
 717                 if (kthread_should_stop() && list_empty(&mp->m_sync_list))
 718                         break;
 719
 720                 spin_lock(&mp->m_sync_lock);
 721                 /*
 722                  * We can get woken by laptop mode, to do a sync -
 723                  * that's the (only!) case where the list would be
 724                  * empty with time remaining.
 725                  */
 726                 if (!timeleft || list_empty(&mp->m_sync_list)) {
 727                         if (!timeleft)
 728                                 timeleft = xfs_syncd_centisecs *
 729                                                         msecs_to_jiffies(10);
 730                         INIT_LIST_HEAD(&mp->m_sync_work.w_list);
 731                         list_add_tail(&mp->m_sync_work.w_list,
 732                                         &mp->m_sync_list);
 733                 }
 734                 list_for_each_entry_safe(work, n, &mp->m_sync_list, w_list)
 735                         list_move(&work->w_list, &tmp);
 736                 spin_unlock(&mp->m_sync_lock);
 737
 738                 list_for_each_entry_safe(work, n, &tmp, w_list) {
 739                         (*work->w_syncer)(mp, work->w_data);
 740                         list_del(&work->w_list);
 741                         if (work == &mp->m_sync_work)
 742                                 continue;
 743                         kmem_free(work);
 744                 }
 745         }
 746
 747         return 0;
 748 }
 749
 750 int
 751 xfs_syncd_init(
 752         struct xfs_mount        *mp)
 753 {
 754         mp->m_sync_work.w_syncer = xfs_sync_worker;
 755         mp->m_sync_work.w_mount = mp;
 756         mp->m_sync_task = kthread_run(xfssyncd, mp, "xfssyncd");
 757         if (IS_ERR(mp->m_sync_task))
 758                 return -PTR_ERR(mp->m_sync_task);
 759         return 0;
 760 }
 761
 762 void
 763 xfs_syncd_stop(
 764         struct xfs_mount        *mp)
 765 {
 766         kthread_stop(mp->m_sync_task);
 767 }
 768