xfs: unshare a range of blocks via fallocate
authorDarrick J. Wong <darrick.wong@oracle.com>
Mon, 3 Oct 2016 16:11:43 +0000 (09:11 -0700)
committerDarrick J. Wong <darrick.wong@oracle.com>
Wed, 5 Oct 2016 23:26:26 +0000 (16:26 -0700)
Unshare all shared extents if the user calls fallocate with the new
unshare mode flag set, so that we can guarantee that a subsequent
write will not ENOSPC.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
[hch: pass inode instead of file to xfs_reflink_dirty_range,
      use iomap infrastructure for copy up]
Signed-off-by: Christoph Hellwig <hch@lst.de>
fs/xfs/xfs_file.c
fs/xfs/xfs_reflink.c
fs/xfs/xfs_reflink.h

index 39fde9f51303ec69b89fd67d2c822171a55e5111..07f951dd26858b9f29d61f4b644a6c04caa2590b 100644 (file)
@@ -848,7 +848,7 @@ buffered:
 #define        XFS_FALLOC_FL_SUPPORTED                                         \
                (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |           \
                 FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE |      \
-                FALLOC_FL_INSERT_RANGE)
+                FALLOC_FL_INSERT_RANGE | FALLOC_FL_UNSHARE_RANGE)
 
 STATIC long
 xfs_file_fallocate(
@@ -938,9 +938,15 @@ xfs_file_fallocate(
 
                if (mode & FALLOC_FL_ZERO_RANGE)
                        error = xfs_zero_file_space(ip, offset, len);
-               else
+               else {
+                       if (mode & FALLOC_FL_UNSHARE_RANGE) {
+                               error = xfs_reflink_unshare(ip, offset, len);
+                               if (error)
+                                       goto out_unlock;
+                       }
                        error = xfs_alloc_file_space(ip, offset, len,
                                                     XFS_BMAPI_PREALLOC);
+               }
                if (error)
                        goto out_unlock;
        }
index 6b22669421b263e2cf4780fab903b0c779bb4f1f..d4707e627a740b4beecb61b1c5918a10701a01be 100644 (file)
@@ -1345,3 +1345,258 @@ out_error:
                trace_xfs_reflink_remap_range_error(dest, error, _RET_IP_);
        return error;
 }
+
+/*
+ * The user wants to preemptively CoW all shared blocks in this file,
+ * which enables us to turn off the reflink flag.  Iterate all
+ * extents which are not prealloc/delalloc to see which ranges are
+ * mentioned in the refcount tree, then read those blocks into the
+ * pagecache, dirty them, fsync them back out, and then we can update
+ * the inode flag.  What happens if we run out of memory? :)
+ */
+STATIC int
+xfs_reflink_dirty_extents(
+       struct xfs_inode        *ip,
+       xfs_fileoff_t           fbno,
+       xfs_filblks_t           end,
+       xfs_off_t               isize)
+{
+       struct xfs_mount        *mp = ip->i_mount;
+       xfs_agnumber_t          agno;
+       xfs_agblock_t           agbno;
+       xfs_extlen_t            aglen;
+       xfs_agblock_t           rbno;
+       xfs_extlen_t            rlen;
+       xfs_off_t               fpos;
+       xfs_off_t               flen;
+       struct xfs_bmbt_irec    map[2];
+       int                     nmaps;
+       int                     error;
+
+       while (end - fbno > 0) {
+               nmaps = 1;
+               /*
+                * Look for extents in the file.  Skip holes, delalloc, or
+                * unwritten extents; they can't be reflinked.
+                */
+               error = xfs_bmapi_read(ip, fbno, end - fbno, map, &nmaps, 0);
+               if (error)
+                       goto out;
+               if (nmaps == 0)
+                       break;
+               if (map[0].br_startblock == HOLESTARTBLOCK ||
+                   map[0].br_startblock == DELAYSTARTBLOCK ||
+                   ISUNWRITTEN(&map[0]))
+                       goto next;
+
+               map[1] = map[0];
+               while (map[1].br_blockcount) {
+                       agno = XFS_FSB_TO_AGNO(mp, map[1].br_startblock);
+                       agbno = XFS_FSB_TO_AGBNO(mp, map[1].br_startblock);
+                       aglen = map[1].br_blockcount;
+
+                       error = xfs_reflink_find_shared(mp, agno, agbno, aglen,
+                                       &rbno, &rlen, true);
+                       if (error)
+                               goto out;
+                       if (rbno == NULLAGBLOCK)
+                               break;
+
+                       /* Dirty the pages */
+                       xfs_iunlock(ip, XFS_ILOCK_EXCL);
+                       fpos = XFS_FSB_TO_B(mp, map[1].br_startoff +
+                                       (rbno - agbno));
+                       flen = XFS_FSB_TO_B(mp, rlen);
+                       if (fpos + flen > isize)
+                               flen = isize - fpos;
+                       error = iomap_file_dirty(VFS_I(ip), fpos, flen,
+                                       &xfs_iomap_ops);
+                       xfs_ilock(ip, XFS_ILOCK_EXCL);
+                       if (error)
+                               goto out;
+
+                       map[1].br_blockcount -= (rbno - agbno + rlen);
+                       map[1].br_startoff += (rbno - agbno + rlen);
+                       map[1].br_startblock += (rbno - agbno + rlen);
+               }
+
+next:
+               fbno = map[0].br_startoff + map[0].br_blockcount;
+       }
+out:
+       return error;
+}
+
+/* Clear the inode reflink flag if there are no shared extents. */
+int
+xfs_reflink_clear_inode_flag(
+       struct xfs_inode        *ip,
+       struct xfs_trans        **tpp)
+{
+       struct xfs_mount        *mp = ip->i_mount;
+       xfs_fileoff_t           fbno;
+       xfs_filblks_t           end;
+       xfs_agnumber_t          agno;
+       xfs_agblock_t           agbno;
+       xfs_extlen_t            aglen;
+       xfs_agblock_t           rbno;
+       xfs_extlen_t            rlen;
+       struct xfs_bmbt_irec    map[2];
+       int                     nmaps;
+       int                     error = 0;
+
+       if (!(ip->i_d.di_flags2 & XFS_DIFLAG2_REFLINK))
+               return 0;
+
+       fbno = 0;
+       end = XFS_B_TO_FSB(mp, i_size_read(VFS_I(ip)));
+       while (end - fbno > 0) {
+               nmaps = 1;
+               /*
+                * Look for extents in the file.  Skip holes, delalloc, or
+                * unwritten extents; they can't be reflinked.
+                */
+               error = xfs_bmapi_read(ip, fbno, end - fbno, map, &nmaps, 0);
+               if (error)
+                       return error;
+               if (nmaps == 0)
+                       break;
+               if (map[0].br_startblock == HOLESTARTBLOCK ||
+                   map[0].br_startblock == DELAYSTARTBLOCK ||
+                   ISUNWRITTEN(&map[0]))
+                       goto next;
+
+               map[1] = map[0];
+               while (map[1].br_blockcount) {
+                       agno = XFS_FSB_TO_AGNO(mp, map[1].br_startblock);
+                       agbno = XFS_FSB_TO_AGBNO(mp, map[1].br_startblock);
+                       aglen = map[1].br_blockcount;
+
+                       error = xfs_reflink_find_shared(mp, agno, agbno, aglen,
+                                       &rbno, &rlen, false);
+                       if (error)
+                               return error;
+                       /* Is there still a shared block here? */
+                       if (rbno != NULLAGBLOCK)
+                               return 0;
+
+                       map[1].br_blockcount -= aglen;
+                       map[1].br_startoff += aglen;
+                       map[1].br_startblock += aglen;
+               }
+
+next:
+               fbno = map[0].br_startoff + map[0].br_blockcount;
+       }
+
+       /*
+        * We didn't find any shared blocks so turn off the reflink flag.
+        * First, get rid of any leftover CoW mappings.
+        */
+       error = xfs_reflink_cancel_cow_blocks(ip, tpp, 0, NULLFILEOFF);
+       if (error)
+               return error;
+
+       /* Clear the inode flag. */
+       trace_xfs_reflink_unset_inode_flag(ip);
+       ip->i_d.di_flags2 &= ~XFS_DIFLAG2_REFLINK;
+       xfs_trans_ijoin(*tpp, ip, 0);
+       xfs_trans_log_inode(*tpp, ip, XFS_ILOG_CORE);
+
+       return error;
+}
+
+/*
+ * Clear the inode reflink flag if there are no shared extents and the size
+ * hasn't changed.
+ */
+STATIC int
+xfs_reflink_try_clear_inode_flag(
+       struct xfs_inode        *ip,
+       xfs_off_t               old_isize)
+{
+       struct xfs_mount        *mp = ip->i_mount;
+       struct xfs_trans        *tp;
+       int                     error = 0;
+
+       /* Start a rolling transaction to remove the mappings */
+       error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, 0, 0, 0, &tp);
+       if (error)
+               return error;
+
+       xfs_ilock(ip, XFS_ILOCK_EXCL);
+       xfs_trans_ijoin(tp, ip, 0);
+
+       if (old_isize != i_size_read(VFS_I(ip)))
+               goto cancel;
+
+       error = xfs_reflink_clear_inode_flag(ip, &tp);
+       if (error)
+               goto cancel;
+
+       error = xfs_trans_commit(tp);
+       if (error)
+               goto out;
+
+       xfs_iunlock(ip, XFS_ILOCK_EXCL);
+       return 0;
+cancel:
+       xfs_trans_cancel(tp);
+out:
+       xfs_iunlock(ip, XFS_ILOCK_EXCL);
+       return error;
+}
+
+/*
+ * Pre-COW all shared blocks within a given byte range of a file and turn off
+ * the reflink flag if we unshare all of the file's blocks.
+ */
+int
+xfs_reflink_unshare(
+       struct xfs_inode        *ip,
+       xfs_off_t               offset,
+       xfs_off_t               len)
+{
+       struct xfs_mount        *mp = ip->i_mount;
+       xfs_fileoff_t           fbno;
+       xfs_filblks_t           end;
+       xfs_off_t               isize;
+       int                     error;
+
+       if (!xfs_is_reflink_inode(ip))
+               return 0;
+
+       trace_xfs_reflink_unshare(ip, offset, len);
+
+       inode_dio_wait(VFS_I(ip));
+
+       /* Try to CoW the selected ranges */
+       xfs_ilock(ip, XFS_ILOCK_EXCL);
+       fbno = XFS_B_TO_FSB(mp, offset);
+       isize = i_size_read(VFS_I(ip));
+       end = XFS_B_TO_FSB(mp, offset + len);
+       error = xfs_reflink_dirty_extents(ip, fbno, end, isize);
+       if (error)
+               goto out_unlock;
+       xfs_iunlock(ip, XFS_ILOCK_EXCL);
+
+       /* Wait for the IO to finish */
+       error = filemap_write_and_wait(VFS_I(ip)->i_mapping);
+       if (error)
+               goto out;
+
+       /* Turn off the reflink flag if we unshared the whole file */
+       if (offset == 0 && len == isize) {
+               error = xfs_reflink_try_clear_inode_flag(ip, isize);
+               if (error)
+                       goto out;
+       }
+
+       return 0;
+
+out_unlock:
+       xfs_iunlock(ip, XFS_ILOCK_EXCL);
+out:
+       trace_xfs_reflink_unshare_error(ip, error, _RET_IP_);
+       return error;
+}
index df82b2049187d1579a88f10df34a2a5d4dd82c40..bade5a61f3b0fd2ee668c07755afb00f900fea9d 100644 (file)
@@ -48,5 +48,9 @@ extern int xfs_reflink_recover_cow(struct xfs_mount *mp);
 extern int xfs_reflink_remap_range(struct xfs_inode *src, xfs_off_t srcoff,
                struct xfs_inode *dest, xfs_off_t destoff, xfs_off_t len,
                unsigned int flags);
+extern int xfs_reflink_clear_inode_flag(struct xfs_inode *ip,
+               struct xfs_trans **tpp);
+extern int xfs_reflink_unshare(struct xfs_inode *ip, xfs_off_t offset,
+               xfs_off_t len);
 
 #endif /* __XFS_REFLINK_H */