xfs: fix getbmap vs mmap deadlock
authorChristoph Hellwig <hch@infradead.org>
Tue, 24 Feb 2009 13:39:02 +0000 (08:39 -0500)
committerFelix Blyakher <felixb@sgi.com>
Wed, 29 Apr 2009 18:25:29 +0000 (13:25 -0500)
xfs_getbmap (or rather the formatters called by it) copy out the getbmap
structures under the ilock, which can deadlock against mmap.  This has
been reported via bugzilla a while ago (#717) and has recently also
shown up via lockdep.

So allocate a temporary buffer to format the kernel getbmap structures
into and then copy them out after dropping the locks.

A little problem with this is that we limit the number of extents we
can copy out by the maximum allocation size, but I see no real way
around that.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Eric Sandeen <sandeen@sandeen.net>
Reviewed-by: Felix Blyakher <felixb@sgi.com>
Signed-off-by: Felix Blyakher <felixb@sgi.com>
fs/xfs/xfs_bmap.c

index abe42448b1c0fa85c1a6556d99be3e60cd2c8df7..ca7c6005a487cab8a9d7b700a9c34b937fc6bc8c 100644 (file)
@@ -5890,12 +5890,13 @@ xfs_getbmap(
        int                     nexleft;        /* # of user extents left */
        int                     subnex;         /* # of bmapi's can do */
        int                     nmap;           /* number of map entries */
-       struct getbmapx         out;            /* output structure */
+       struct getbmapx         *out;           /* output structure */
        int                     whichfork;      /* data or attr fork */
        int                     prealloced;     /* this is a file with
                                                 * preallocated data space */
        int                     iflags;         /* interface flags */
        int                     bmapi_flags;    /* flags for xfs_bmapi */
+       int                     cur_ext = 0;
 
        mp = ip->i_mount;
        iflags = bmv->bmv_iflags;
@@ -5971,6 +5972,13 @@ xfs_getbmap(
                return XFS_ERROR(EINVAL);
        bmvend = bmv->bmv_offset + bmv->bmv_length;
 
+
+       if (bmv->bmv_count > ULONG_MAX / sizeof(struct getbmapx))
+               return XFS_ERROR(ENOMEM);
+       out = kmem_zalloc(bmv->bmv_count * sizeof(struct getbmapx), KM_MAYFAIL);
+       if (!out)
+               return XFS_ERROR(ENOMEM);
+
        xfs_ilock(ip, XFS_IOLOCK_SHARED);
        if (whichfork == XFS_DATA_FORK && !(iflags & BMV_IF_DELALLOC)) {
                if (ip->i_delayed_blks || ip->i_size > ip->i_d.di_size) {
@@ -6025,39 +6033,39 @@ xfs_getbmap(
                ASSERT(nmap <= subnex);
 
                for (i = 0; i < nmap && nexleft && bmv->bmv_length; i++) {
-                       int full = 0;   /* user array is full */
-
-                       out.bmv_oflags = 0;
+                       out[cur_ext].bmv_oflags = 0;
                        if (map[i].br_state == XFS_EXT_UNWRITTEN)
-                               out.bmv_oflags |= BMV_OF_PREALLOC;
+                               out[cur_ext].bmv_oflags |= BMV_OF_PREALLOC;
                        else if (map[i].br_startblock == DELAYSTARTBLOCK)
-                               out.bmv_oflags |= BMV_OF_DELALLOC;
-                       out.bmv_offset = XFS_FSB_TO_BB(mp, map[i].br_startoff);
-                       out.bmv_length = XFS_FSB_TO_BB(mp, map[i].br_blockcount);
-                       out.bmv_unused1 = out.bmv_unused2 = 0;
+                               out[cur_ext].bmv_oflags |= BMV_OF_DELALLOC;
+                       out[cur_ext].bmv_offset =
+                               XFS_FSB_TO_BB(mp, map[i].br_startoff);
+                       out[cur_ext].bmv_length =
+                               XFS_FSB_TO_BB(mp, map[i].br_blockcount);
+                       out[cur_ext].bmv_unused1 = 0;
+                       out[cur_ext].bmv_unused2 = 0;
                        ASSERT(((iflags & BMV_IF_DELALLOC) != 0) ||
                              (map[i].br_startblock != DELAYSTARTBLOCK));
                         if (map[i].br_startblock == HOLESTARTBLOCK &&
                            whichfork == XFS_ATTR_FORK) {
                                /* came to the end of attribute fork */
-                               out.bmv_oflags |= BMV_OF_LAST;
+                               out[cur_ext].bmv_oflags |= BMV_OF_LAST;
                                goto out_free_map;
                        }
 
-                       if (!xfs_getbmapx_fix_eof_hole(ip, &out, prealloced,
-                                       bmvend, map[i].br_startblock))
+                       if (!xfs_getbmapx_fix_eof_hole(ip, &out[cur_ext],
+                                       prealloced, bmvend,
+                                       map[i].br_startblock))
                                goto out_free_map;
 
-                       /* format results & advance arg */
-                       error = formatter(&arg, &out, &full);
-                       if (error || full)
-                               goto out_free_map;
                        nexleft--;
                        bmv->bmv_offset =
-                               out.bmv_offset + out.bmv_length;
+                               out[cur_ext].bmv_offset +
+                               out[cur_ext].bmv_length;
                        bmv->bmv_length =
                                max_t(__int64_t, 0, bmvend - bmv->bmv_offset);
                        bmv->bmv_entries++;
+                       cur_ext++;
                }
        } while (nmap && nexleft && bmv->bmv_length);
 
@@ -6067,6 +6075,16 @@ xfs_getbmap(
        xfs_iunlock_map_shared(ip, lock);
  out_unlock_iolock:
        xfs_iunlock(ip, XFS_IOLOCK_SHARED);
+
+       for (i = 0; i < cur_ext; i++) {
+               int full = 0;   /* user array is full */
+
+               /* format results & advance arg */
+               error = formatter(&arg, &out[i], &full);
+               if (error || full)
+                       break;
+       }
+
        return error;
 }