Merge branch 'master' into for-linus
authorAlex Elder <aelder@sgi.com>
Thu, 8 Oct 2009 18:53:44 +0000 (13:53 -0500)
committerAlex Elder <aelder@sgi.com>
Thu, 8 Oct 2009 18:53:44 +0000 (13:53 -0500)
13 files changed:
fs/xfs/linux-2.6/xfs_aops.c
fs/xfs/linux-2.6/xfs_file.c
fs/xfs/linux-2.6/xfs_iops.c
fs/xfs/linux-2.6/xfs_lrw.c
fs/xfs/linux-2.6/xfs_super.c
fs/xfs/linux-2.6/xfs_sync.c
fs/xfs/xfs_dfrag.c
fs/xfs/xfs_dir2_leaf.c
fs/xfs/xfs_inode.c
fs/xfs/xfs_inode.h
fs/xfs/xfs_inode_item.c
fs/xfs/xfs_itable.c
fs/xfs/xfs_vnodeops.c

index d5e5559e31db3774ba6f76e350302009ccf831eb..6a2910ee8a28b0382efe29bc5d14ebc08afae662 100644 (file)
@@ -185,6 +185,24 @@ xfs_destroy_ioend(
        mempool_free(ioend, xfs_ioend_pool);
 }
 
+/*
+ * If the end of the current ioend is beyond the current EOF,
+ * return the new EOF value, otherwise zero.
+ */
+STATIC xfs_fsize_t
+xfs_ioend_new_eof(
+       xfs_ioend_t             *ioend)
+{
+       xfs_inode_t             *ip = XFS_I(ioend->io_inode);
+       xfs_fsize_t             isize;
+       xfs_fsize_t             bsize;
+
+       bsize = ioend->io_offset + ioend->io_size;
+       isize = MAX(ip->i_size, ip->i_new_size);
+       isize = MIN(isize, bsize);
+       return isize > ip->i_d.di_size ? isize : 0;
+}
+
 /*
  * Update on-disk file size now that data has been written to disk.
  * The current in-memory file size is i_size.  If a write is beyond
@@ -192,13 +210,13 @@ xfs_destroy_ioend(
  * updated.  If this write does not extend all the way to the valid
  * file size then restrict this update to the end of the write.
  */
+
 STATIC void
 xfs_setfilesize(
        xfs_ioend_t             *ioend)
 {
        xfs_inode_t             *ip = XFS_I(ioend->io_inode);
        xfs_fsize_t             isize;
-       xfs_fsize_t             bsize;
 
        ASSERT((ip->i_d.di_mode & S_IFMT) == S_IFREG);
        ASSERT(ioend->io_type != IOMAP_READ);
@@ -206,16 +224,10 @@ xfs_setfilesize(
        if (unlikely(ioend->io_error))
                return;
 
-       bsize = ioend->io_offset + ioend->io_size;
-
        xfs_ilock(ip, XFS_ILOCK_EXCL);
-
-       isize = MAX(ip->i_size, ip->i_new_size);
-       isize = MIN(isize, bsize);
-
-       if (ip->i_d.di_size < isize) {
+       isize = xfs_ioend_new_eof(ioend);
+       if (isize) {
                ip->i_d.di_size = isize;
-               ip->i_update_core = 1;
                xfs_mark_inode_dirty_sync(ip);
        }
 
@@ -404,10 +416,16 @@ xfs_submit_ioend_bio(
        struct bio      *bio)
 {
        atomic_inc(&ioend->io_remaining);
-
        bio->bi_private = ioend;
        bio->bi_end_io = xfs_end_bio;
 
+       /*
+        * If the I/O is beyond EOF we mark the inode dirty immediately
+        * but don't update the inode size until I/O completion.
+        */
+       if (xfs_ioend_new_eof(ioend))
+               xfs_mark_inode_dirty_sync(XFS_I(ioend->io_inode));
+
        submit_bio(WRITE, bio);
        ASSERT(!bio_flagged(bio, BIO_EOPNOTSUPP));
        bio_put(bio);
index 988d8f87bc0f80b8a23d18397bcbac8f3116647e..6d65baf15a14f17aba87a879e6397cfedfcdd260 100644 (file)
@@ -176,14 +176,7 @@ xfs_file_fsync(
        struct dentry           *dentry,
        int                     datasync)
 {
-       struct inode            *inode = dentry->d_inode;
-       struct xfs_inode        *ip = XFS_I(inode);
-       int                     error;
-
-       /* capture size updates in I/O completion before writing the inode. */
-       error = filemap_fdatawait(inode->i_mapping);
-       if (error)
-               return error;
+       struct xfs_inode        *ip = XFS_I(dentry->d_inode);
 
        xfs_iflags_clear(ip, XFS_ITRUNCATED);
        return -xfs_fsync(ip);
index da0159d99f82795cd28277138a1f3d5e06916b03..cd42ef78f6b54e705ad604e3ccfc4eb1a3df3743 100644 (file)
 #include <linux/fiemap.h>
 
 /*
- * Bring the atime in the XFS inode uptodate.
- * Used before logging the inode to disk or when the Linux inode goes away.
+ * Bring the timestamps in the XFS inode uptodate.
+ *
+ * Used before writing the inode to disk.
  */
 void
-xfs_synchronize_atime(
+xfs_synchronize_times(
        xfs_inode_t     *ip)
 {
        struct inode    *inode = VFS_I(ip);
 
-       if (!(inode->i_state & I_CLEAR)) {
-               ip->i_d.di_atime.t_sec = (__int32_t)inode->i_atime.tv_sec;
-               ip->i_d.di_atime.t_nsec = (__int32_t)inode->i_atime.tv_nsec;
-       }
+       ip->i_d.di_atime.t_sec = (__int32_t)inode->i_atime.tv_sec;
+       ip->i_d.di_atime.t_nsec = (__int32_t)inode->i_atime.tv_nsec;
+       ip->i_d.di_ctime.t_sec = (__int32_t)inode->i_ctime.tv_sec;
+       ip->i_d.di_ctime.t_nsec = (__int32_t)inode->i_ctime.tv_nsec;
+       ip->i_d.di_mtime.t_sec = (__int32_t)inode->i_mtime.tv_sec;
+       ip->i_d.di_mtime.t_nsec = (__int32_t)inode->i_mtime.tv_nsec;
 }
 
 /*
@@ -106,32 +109,20 @@ xfs_ichgtime(
        if ((flags & XFS_ICHGTIME_MOD) &&
            !timespec_equal(&inode->i_mtime, &tv)) {
                inode->i_mtime = tv;
-               ip->i_d.di_mtime.t_sec = (__int32_t)tv.tv_sec;
-               ip->i_d.di_mtime.t_nsec = (__int32_t)tv.tv_nsec;
                sync_it = 1;
        }
        if ((flags & XFS_ICHGTIME_CHG) &&
            !timespec_equal(&inode->i_ctime, &tv)) {
                inode->i_ctime = tv;
-               ip->i_d.di_ctime.t_sec = (__int32_t)tv.tv_sec;
-               ip->i_d.di_ctime.t_nsec = (__int32_t)tv.tv_nsec;
                sync_it = 1;
        }
 
        /*
-        * We update the i_update_core field _after_ changing
-        * the timestamps in order to coordinate properly with
-        * xfs_iflush() so that we don't lose timestamp updates.
-        * This keeps us from having to hold the inode lock
-        * while doing this.  We use the SYNCHRONIZE macro to
-        * ensure that the compiler does not reorder the update
-        * of i_update_core above the timestamp updates above.
+        * Update complete - now make sure everyone knows that the inode
+        * is dirty.
         */
-       if (sync_it) {
-               SYNCHRONIZE();
-               ip->i_update_core = 1;
+       if (sync_it)
                xfs_mark_inode_dirty_sync(ip);
-       }
 }
 
 /*
@@ -506,10 +497,8 @@ xfs_vn_getattr(
        stat->gid = ip->i_d.di_gid;
        stat->ino = ip->i_ino;
        stat->atime = inode->i_atime;
-       stat->mtime.tv_sec = ip->i_d.di_mtime.t_sec;
-       stat->mtime.tv_nsec = ip->i_d.di_mtime.t_nsec;
-       stat->ctime.tv_sec = ip->i_d.di_ctime.t_sec;
-       stat->ctime.tv_nsec = ip->i_d.di_ctime.t_nsec;
+       stat->mtime = inode->i_mtime;
+       stat->ctime = inode->i_ctime;
        stat->blocks =
                XFS_FSB_TO_BB(mp, ip->i_d.di_nblocks + ip->i_delayed_blks);
 
index 49e4a6aea73c3c85f7d388d56d4371867c930178..072050f8d34634d59e4d200dfe14218bcf0e75ad 100644 (file)
@@ -667,7 +667,7 @@ start:
                xip->i_new_size = new_size;
 
        if (likely(!(ioflags & IO_INVIS)))
-               xfs_ichgtime(xip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
+               file_update_time(file);
 
        /*
         * If the offset is beyond the size of the file, we have a couple
index 5d7c60ac77b4967d262fbca08037993109c9c39b..eefea0d9d03d6c92b1d8edd2a0846b3b438092fc 100644 (file)
@@ -976,6 +976,28 @@ xfs_fs_inode_init_once(
        mrlock_init(&ip->i_iolock, MRLOCK_BARRIER, "xfsio", ip->i_ino);
 }
 
+/*
+ * Dirty the XFS inode when mark_inode_dirty_sync() is called so that
+ * we catch unlogged VFS level updates to the inode. Care must be taken
+ * here - the transaction code calls mark_inode_dirty_sync() to mark the
+ * VFS inode dirty in a transaction and clears the i_update_core field;
+ * it must clear the field after calling mark_inode_dirty_sync() to
+ * correctly indicate that the dirty state has been propagated into the
+ * inode log item.
+ *
+ * We need the barrier() to maintain correct ordering between unlogged
+ * updates and the transaction commit code that clears the i_update_core
+ * field. This requires all updates to be completed before marking the
+ * inode dirty.
+ */
+STATIC void
+xfs_fs_dirty_inode(
+       struct inode    *inode)
+{
+       barrier();
+       XFS_I(inode)->i_update_core = 1;
+}
+
 /*
  * Attempt to flush the inode, this will actually fail
  * if the inode is pinned, but we dirty the inode again
@@ -1126,7 +1148,7 @@ xfs_fs_put_super(
 }
 
 STATIC int
-xfs_fs_sync_super(
+xfs_fs_sync_fs(
        struct super_block      *sb,
        int                     wait)
 {
@@ -1134,23 +1156,23 @@ xfs_fs_sync_super(
        int                     error;
 
        /*
-        * Treat a sync operation like a freeze.  This is to work
-        * around a race in sync_inodes() which works in two phases
-        * - an asynchronous flush, which can write out an inode
-        * without waiting for file size updates to complete, and a
-        * synchronous flush, which wont do anything because the
-        * async flush removed the inode's dirty flag.  Also
-        * sync_inodes() will not see any files that just have
-        * outstanding transactions to be flushed because we don't
-        * dirty the Linux inode until after the transaction I/O
-        * completes.
+        * Not much we can do for the first async pass.  Writing out the
+        * superblock would be counter-productive as we are going to redirty
+        * when writing out other data and metadata (and writing out a single
+        * block is quite fast anyway).
+        *
+        * Try to asynchronously kick off quota syncing at least.
         */
-       if (wait || unlikely(sb->s_frozen == SB_FREEZE_WRITE))
-               error = xfs_quiesce_data(mp);
-       else
-               error = xfs_sync_fsdata(mp, 0);
+       if (!wait) {
+               xfs_qm_sync(mp, SYNC_TRYLOCK);
+               return 0;
+       }
+
+       error = xfs_quiesce_data(mp);
+       if (error)
+               return -error;
 
-       if (unlikely(laptop_mode)) {
+       if (laptop_mode) {
                int     prev_sync_seq = mp->m_sync_seq;
 
                /*
@@ -1169,7 +1191,7 @@ xfs_fs_sync_super(
                                mp->m_sync_seq != prev_sync_seq);
        }
 
-       return -error;
+       return 0;
 }
 
 STATIC int
@@ -1539,10 +1561,11 @@ xfs_fs_get_sb(
 static struct super_operations xfs_super_operations = {
        .alloc_inode            = xfs_fs_alloc_inode,
        .destroy_inode          = xfs_fs_destroy_inode,
+       .dirty_inode            = xfs_fs_dirty_inode,
        .write_inode            = xfs_fs_write_inode,
        .clear_inode            = xfs_fs_clear_inode,
        .put_super              = xfs_fs_put_super,
-       .sync_fs                = xfs_fs_sync_super,
+       .sync_fs                = xfs_fs_sync_fs,
        .freeze_fs              = xfs_fs_freeze,
        .statfs                 = xfs_fs_statfs,
        .remount_fs             = xfs_fs_remount,
index 320be6aea4924ed6fdcf4ac1e581c0c616d04c83..961df0a22c7837e9a969a8ab73331c594d5a6116 100644 (file)
@@ -309,11 +309,15 @@ xfs_sync_attr(
 STATIC int
 xfs_commit_dummy_trans(
        struct xfs_mount        *mp,
-       uint                    log_flags)
+       uint                    flags)
 {
        struct xfs_inode        *ip = mp->m_rootip;
        struct xfs_trans        *tp;
        int                     error;
+       int                     log_flags = XFS_LOG_FORCE;
+
+       if (flags & SYNC_WAIT)
+               log_flags |= XFS_LOG_SYNC;
 
        /*
         * Put a dummy transaction in the log to tell recovery
@@ -331,13 +335,12 @@ xfs_commit_dummy_trans(
        xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
        xfs_trans_ihold(tp, ip);
        xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
-       /* XXX(hch): ignoring the error here.. */
        error = xfs_trans_commit(tp, 0);
-
        xfs_iunlock(ip, XFS_ILOCK_EXCL);
 
+       /* the log force ensures this transaction is pushed to disk */
        xfs_log_force(mp, 0, log_flags);
-       return 0;
+       return error;
 }
 
 int
@@ -385,7 +388,20 @@ xfs_sync_fsdata(
        else
                XFS_BUF_ASYNC(bp);
 
-       return xfs_bwrite(mp, bp);
+       error = xfs_bwrite(mp, bp);
+       if (error)
+               return error;
+
+       /*
+        * If this is a data integrity sync make sure all pending buffers
+        * are flushed out for the log coverage check below.
+        */
+       if (flags & SYNC_WAIT)
+               xfs_flush_buftarg(mp->m_ddev_targp, 1);
+
+       if (xfs_log_need_covered(mp))
+               error = xfs_commit_dummy_trans(mp, flags);
+       return error;
 
  out_brelse:
        xfs_buf_relse(bp);
@@ -419,14 +435,16 @@ xfs_quiesce_data(
        /* push non-blocking */
        xfs_sync_data(mp, 0);
        xfs_qm_sync(mp, SYNC_TRYLOCK);
-       xfs_filestream_flush(mp);
 
-       /* push and block */
+       /* push and block till complete */
        xfs_sync_data(mp, SYNC_WAIT);
        xfs_qm_sync(mp, SYNC_WAIT);
 
+       /* drop inode references pinned by filestreams */
+       xfs_filestream_flush(mp);
+
        /* write superblock and hoover up shutdown errors */
-       error = xfs_sync_fsdata(mp, 0);
+       error = xfs_sync_fsdata(mp, SYNC_WAIT);
 
        /* flush data-only devices */
        if (mp->m_rtdev_targp)
@@ -570,8 +588,6 @@ xfs_sync_worker(
                /* dgc: errors ignored here */
                error = xfs_qm_sync(mp, SYNC_TRYLOCK);
                error = xfs_sync_fsdata(mp, SYNC_TRYLOCK);
-               if (xfs_log_need_covered(mp))
-                       error = xfs_commit_dummy_trans(mp, XFS_LOG_FORCE);
        }
        mp->m_sync_seq++;
        wake_up(&mp->m_wait_single_sync_task);
index 7465f9ee125fcc2b80194043e9c141799a05ffde..ab89a7e94a0fd4b861114f656b99e5fcf0355a14 100644 (file)
@@ -206,10 +206,10 @@ xfs_swap_extents(
         * process that the file was not changed out from
         * under it.
         */
-       if ((sbp->bs_ctime.tv_sec != ip->i_d.di_ctime.t_sec) ||
-           (sbp->bs_ctime.tv_nsec != ip->i_d.di_ctime.t_nsec) ||
-           (sbp->bs_mtime.tv_sec != ip->i_d.di_mtime.t_sec) ||
-           (sbp->bs_mtime.tv_nsec != ip->i_d.di_mtime.t_nsec)) {
+       if ((sbp->bs_ctime.tv_sec != VFS_I(ip)->i_ctime.tv_sec) ||
+           (sbp->bs_ctime.tv_nsec != VFS_I(ip)->i_ctime.tv_nsec) ||
+           (sbp->bs_mtime.tv_sec != VFS_I(ip)->i_mtime.tv_sec) ||
+           (sbp->bs_mtime.tv_nsec != VFS_I(ip)->i_mtime.tv_nsec)) {
                error = XFS_ERROR(EBUSY);
                goto out_unlock;
        }
index fa913e4594421adefc392e80c254d492b8a9ec91..41ad537c49e983a8ec052bfe38c68814f9a0d796 100644 (file)
@@ -854,6 +854,7 @@ xfs_dir2_leaf_getdents(
                         */
                        ra_want = howmany(bufsize + mp->m_dirblksize,
                                          mp->m_sb.sb_blocksize) - 1;
+                       ASSERT(ra_want >= 0);
 
                        /*
                         * If we don't have as many as we want, and we haven't
@@ -1088,7 +1089,8 @@ xfs_dir2_leaf_getdents(
                 */
                ptr += length;
                curoff += length;
-               bufsize -= length;
+               /* bufsize may have just been a guess; don't go negative */
+               bufsize = bufsize > length ? bufsize - length : 0;
        }
 
        /*
index c1dc7ef5a1d853ea935fd954ad0cf4863e701cd4..b92a4fa2a0a12eda161c2ff897bd820f5cbde1e8 100644 (file)
@@ -3068,9 +3068,9 @@ xfs_iflush_int(
        SYNCHRONIZE();
 
        /*
-        * Make sure to get the latest atime from the Linux inode.
+        * Make sure to get the latest timestamps from the Linux inode.
         */
-       xfs_synchronize_atime(ip);
+       xfs_synchronize_times(ip);
 
        if (XFS_TEST_ERROR(be16_to_cpu(dip->di_magic) != XFS_DINODE_MAGIC,
                               mp, XFS_ERRTAG_IFLUSH_1, XFS_RANDOM_IFLUSH_1)) {
index 0b38b9a869ece2136bf798a5329d2a38b8bf98ea..41555de1d1dbd4e6befcd64840712de5beff8235 100644 (file)
@@ -504,7 +504,7 @@ void                xfs_ichgtime(xfs_inode_t *, int);
 void           xfs_lock_inodes(xfs_inode_t **, int, uint);
 void           xfs_lock_two_inodes(xfs_inode_t *, xfs_inode_t *, uint);
 
-void           xfs_synchronize_atime(xfs_inode_t *);
+void           xfs_synchronize_times(xfs_inode_t *);
 void           xfs_mark_inode_dirty_sync(xfs_inode_t *);
 
 #if defined(XFS_INODE_TRACE)
index 47d5b663c37e44a2b3528413edba37ec43eba880..9794b876d6ff65e34336e390a9becca46fa8e330 100644 (file)
@@ -231,6 +231,15 @@ xfs_inode_item_format(
        vecp++;
        nvecs        = 1;
 
+       /*
+        * Make sure the linux inode is dirty. We do this before
+        * clearing i_update_core as the VFS will call back into
+        * XFS here and set i_update_core, so we need to dirty the
+        * inode first so that the ordering of i_update_core and
+        * unlogged modifications still works as described below.
+        */
+       xfs_mark_inode_dirty_sync(ip);
+
        /*
         * Clear i_update_core if the timestamps (or any other
         * non-transactional modification) need flushing/logging
@@ -263,14 +272,9 @@ xfs_inode_item_format(
        }
 
        /*
-        * Make sure to get the latest atime from the Linux inode.
+        * Make sure to get the latest timestamps from the Linux inode.
         */
-       xfs_synchronize_atime(ip);
-
-       /*
-        * make sure the linux inode is dirty
-        */
-       xfs_mark_inode_dirty_sync(ip);
+       xfs_synchronize_times(ip);
 
        vecp->i_addr = (xfs_caddr_t)&ip->i_d;
        vecp->i_len  = sizeof(struct xfs_icdinode);
index b68f9107e26cce1e3eb7f4b08f272d493a9d8b06..62efab2f38392e845bf82fc8ba48825f3c016727 100644 (file)
@@ -59,6 +59,7 @@ xfs_bulkstat_one_iget(
 {
        xfs_icdinode_t  *dic;   /* dinode core info pointer */
        xfs_inode_t     *ip;            /* incore inode pointer */
+       struct inode    *inode;
        int             error;
 
        error = xfs_iget(mp, NULL, ino,
@@ -72,6 +73,7 @@ xfs_bulkstat_one_iget(
        ASSERT(ip->i_imap.im_blkno != 0);
 
        dic = &ip->i_d;
+       inode = VFS_I(ip);
 
        /* xfs_iget returns the following without needing
         * further change.
@@ -83,16 +85,19 @@ xfs_bulkstat_one_iget(
        buf->bs_uid = dic->di_uid;
        buf->bs_gid = dic->di_gid;
        buf->bs_size = dic->di_size;
+
        /*
-        * We are reading the atime from the Linux inode because the
-        * dinode might not be uptodate.
+        * We need to read the timestamps from the Linux inode because
+        * the VFS keeps writing directly into the inode structure instead
+        * of telling us about the updates.
         */
-       buf->bs_atime.tv_sec = VFS_I(ip)->i_atime.tv_sec;
-       buf->bs_atime.tv_nsec = VFS_I(ip)->i_atime.tv_nsec;
-       buf->bs_mtime.tv_sec = dic->di_mtime.t_sec;
-       buf->bs_mtime.tv_nsec = dic->di_mtime.t_nsec;
-       buf->bs_ctime.tv_sec = dic->di_ctime.t_sec;
-       buf->bs_ctime.tv_nsec = dic->di_ctime.t_nsec;
+       buf->bs_atime.tv_sec = inode->i_atime.tv_sec;
+       buf->bs_atime.tv_nsec = inode->i_atime.tv_nsec;
+       buf->bs_mtime.tv_sec = inode->i_mtime.tv_sec;
+       buf->bs_mtime.tv_nsec = inode->i_mtime.tv_nsec;
+       buf->bs_ctime.tv_sec = inode->i_ctime.tv_sec;
+       buf->bs_ctime.tv_nsec = inode->i_ctime.tv_nsec;
+
        buf->bs_xflags = xfs_ip2xflags(ip);
        buf->bs_extsize = dic->di_extsize << mp->m_sb.sb_blocklog;
        buf->bs_extents = dic->di_nextents;
index a434f287962d18588e0711110b8b758818b9ceee..b572f7e840e0b82a372225581d87c85c104dfec6 100644 (file)
@@ -2475,12 +2475,6 @@ xfs_reclaim(
 
        ASSERT(XFS_FORCED_SHUTDOWN(ip->i_mount) || ip->i_delayed_blks == 0);
 
-       /*
-        * Make sure the atime in the XFS inode is correct before freeing the
-        * Linux inode.
-        */
-       xfs_synchronize_atime(ip);
-
        /*
         * If we have nothing to flush with this inode then complete the
         * teardown now, otherwise break the link between the xfs inode and the