xfs: add missing ilock around dio write last extent alignment
authorBrian Foster <bfoster@redhat.com>
Mon, 12 Oct 2015 04:34:20 +0000 (15:34 +1100)
committerDave Chinner <david@fromorbit.com>
Mon, 12 Oct 2015 04:34:20 +0000 (15:34 +1100)
The iomap codepath (via get_blocks()) acquires and release the inode
lock in the case of a direct write that requires block allocation. This
is because xfs_iomap_write_direct() allocates a transaction, which means
the ilock must be dropped and reacquired after the transaction is
allocated and reserved.

xfs_iomap_write_direct() invokes xfs_iomap_eof_align_last_fsb() before
the transaction is created and thus before the ilock is reacquired. This
can lead to calls to xfs_iread_extents() and reads of the in-core extent
list without any synchronization (via xfs_bmap_eof() and
xfs_bmap_last_extent()). xfs_iread_extents() assert fails if the ilock
is not held, but this is not currently seen in practice as the current
callers had already invoked xfs_bmapi_read().

What has been seen in practice are reports of crashes down in the
xfs_bmap_eof() codepath on direct writes due to seemingly bogus pointer
references from xfs_iext_get_ext(). While an explicit reproducer is not
currently available to confirm the cause of the problem, crash analysis
and code inspection from David Jeffrey had identified the insufficient
locking.

xfs_iomap_eof_align_last_fsb() is called from other contexts with the
inode lock already held, so we cannot acquire it therein.
__xfs_get_blocks() acquires and drops the ilock with variable flags to
cover the event that the extent list must be read in. The common case is
that __xfs_get_blocks() acquires the shared ilock. To provide locking
around the last extent alignment call without adding more lock cycles to
the dio path, update xfs_iomap_write_direct() to expect the shared ilock
held on entry and do the extent alignment under its protection. Demote
the lock, if necessary, from __xfs_get_blocks() and push the
xfs_qm_dqattach() call outside of the shared lock critical section.
Also, add an assert to document that the extent list is always expected
to be present in this path. Otherwise, we risk a call to
xfs_iread_extents() while under the shared ilock. This is safe as all
current callers have executed an xfs_bmapi_read() call under the current
iolock context.

Reported-by: David Jeffery <djeffery@redhat.com>
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
fs/xfs/xfs_aops.c
fs/xfs/xfs_iomap.c
fs/xfs/xfs_pnfs.c

index e485e31813fa9903cbbcc8b5c773f7c5c89e5c3c..e4fff5898c1c0eefd4fb053b028bd357e6dd69d8 100644 (file)
@@ -1408,12 +1408,12 @@ __xfs_get_blocks(
              imap.br_startblock == DELAYSTARTBLOCK))) {
                if (direct || xfs_get_extsz_hint(ip)) {
                        /*
-                        * Drop the ilock in preparation for starting the block
-                        * allocation transaction.  It will be retaken
-                        * exclusively inside xfs_iomap_write_direct for the
-                        * actual allocation.
+                        * xfs_iomap_write_direct() expects the shared lock. It
+                        * is unlocked on return.
                         */
-                       xfs_iunlock(ip, lockmode);
+                       if (lockmode == XFS_ILOCK_EXCL)
+                               xfs_ilock_demote(ip, lockmode);
+
                        error = xfs_iomap_write_direct(ip, offset, size,
                                                       &imap, nimaps);
                        if (error)
index 1f86033171c84ef4b12a201e112c9c37c7c69bdf..1beda331d8a7a9d1e3ea3e655427b85cda9313a8 100644 (file)
@@ -131,20 +131,29 @@ xfs_iomap_write_direct(
        uint            qblocks, resblks, resrtextents;
        int             committed;
        int             error;
-
-       error = xfs_qm_dqattach(ip, 0);
-       if (error)
-               return error;
+       int             lockmode;
 
        rt = XFS_IS_REALTIME_INODE(ip);
        extsz = xfs_get_extsz_hint(ip);
+       lockmode = XFS_ILOCK_SHARED;    /* locked by caller */
+
+       ASSERT(xfs_isilocked(ip, lockmode));
 
        offset_fsb = XFS_B_TO_FSBT(mp, offset);
        last_fsb = XFS_B_TO_FSB(mp, ((xfs_ufsize_t)(offset + count)));
        if ((offset + count) > XFS_ISIZE(ip)) {
+               /*
+                * Assert that the in-core extent list is present since this can
+                * call xfs_iread_extents() and we only have the ilock shared.
+                * This should be safe because the lock was held around a bmapi
+                * call in the caller and we only need it to access the in-core
+                * list.
+                */
+               ASSERT(XFS_IFORK_PTR(ip, XFS_DATA_FORK)->if_flags &
+                                                               XFS_IFEXTENTS);
                error = xfs_iomap_eof_align_last_fsb(mp, ip, extsz, &last_fsb);
                if (error)
-                       return error;
+                       goto out_unlock;
        } else {
                if (nmaps && (imap->br_startblock == HOLESTARTBLOCK))
                        last_fsb = MIN(last_fsb, (xfs_fileoff_t)
@@ -173,6 +182,15 @@ xfs_iomap_write_direct(
                quota_flag = XFS_QMOPT_RES_REGBLKS;
        }
 
+       /*
+        * Drop the shared lock acquired by the caller, attach the dquot if
+        * necessary and move on to transaction setup.
+        */
+       xfs_iunlock(ip, lockmode);
+       error = xfs_qm_dqattach(ip, 0);
+       if (error)
+               return error;
+
        /*
         * Allocate and setup the transaction
         */
@@ -187,7 +205,8 @@ xfs_iomap_write_direct(
                return error;
        }
 
-       xfs_ilock(ip, XFS_ILOCK_EXCL);
+       lockmode = XFS_ILOCK_EXCL;
+       xfs_ilock(ip, lockmode);
 
        error = xfs_trans_reserve_quota_nblks(tp, ip, qblocks, 0, quota_flag);
        if (error)
@@ -229,7 +248,7 @@ xfs_iomap_write_direct(
                error = xfs_alert_fsblock_zero(ip, imap);
 
 out_unlock:
-       xfs_iunlock(ip, XFS_ILOCK_EXCL);
+       xfs_iunlock(ip, lockmode);
        return error;
 
 out_bmap_cancel:
index ab4a6066f7ca7b24fbe4cdf2db0f2c1ddfdab3df..dc6221942b85f437767decfca9bf7ca9b0a94026 100644 (file)
@@ -181,6 +181,11 @@ xfs_fs_map_blocks(
                ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
 
                if (!nimaps || imap.br_startblock == HOLESTARTBLOCK) {
+                       /*
+                        * xfs_iomap_write_direct() expects to take ownership of
+                        * the shared ilock.
+                        */
+                       xfs_ilock(ip, XFS_ILOCK_SHARED);
                        error = xfs_iomap_write_direct(ip, offset, length,
                                                       &imap, nimaps);
                        if (error)