xfs: remove an extent from the rmap btree
authorDarrick J. Wong <darrick.wong@oracle.com>
Wed, 3 Aug 2016 01:45:12 +0000 (11:45 +1000)
committerDave Chinner <david@fromorbit.com>
Wed, 3 Aug 2016 01:45:12 +0000 (11:45 +1000)
Originally-From: Dave Chinner <dchinner@redhat.com>

Now that we have records in the rmap btree, we need to remove them
when extents are freed. This needs to find the relevant record in
the btree and remove/trim/split it accordingly.

[darrick.wong@oracle.com: make rmap routines handle the enlarged keyspace]
[dchinner: remove remaining unused debug printks]
[darrick: fix a bug when growfs in an AG with an rmap ending at EOFS]

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
fs/xfs/libxfs/xfs_rmap.c

index 1c0e2a8943a4b6a031ff2dc464ae354066f98ad8..1e48a0b62ba6682d4819475b87476b7863ff0af4 100644 (file)
@@ -133,6 +133,212 @@ xfs_rmap_get_rec(
        return xfs_rmap_btrec_to_irec(rec, irec);
 }
 
+/*
+ * Find the extent in the rmap btree and remove it.
+ *
+ * The record we find should always be an exact match for the extent that we're
+ * looking for, since we insert them into the btree without modification.
+ *
+ * Special Case #1: when growing the filesystem, we "free" an extent when
+ * growing the last AG. This extent is new space and so it is not tracked as
+ * used space in the btree. The growfs code will pass in an owner of
+ * XFS_RMAP_OWN_NULL to indicate that it expected that there is no owner of this
+ * extent. We verify that - the extent lookup result in a record that does not
+ * overlap.
+ *
+ * Special Case #2: EFIs do not record the owner of the extent, so when
+ * recovering EFIs from the log we pass in XFS_RMAP_OWN_UNKNOWN to tell the rmap
+ * btree to ignore the owner (i.e. wildcard match) so we don't trigger
+ * corruption checks during log recovery.
+ */
+STATIC int
+xfs_rmap_unmap(
+       struct xfs_btree_cur    *cur,
+       xfs_agblock_t           bno,
+       xfs_extlen_t            len,
+       bool                    unwritten,
+       struct xfs_owner_info   *oinfo)
+{
+       struct xfs_mount        *mp = cur->bc_mp;
+       struct xfs_rmap_irec    ltrec;
+       uint64_t                ltoff;
+       int                     error = 0;
+       int                     i;
+       uint64_t                owner;
+       uint64_t                offset;
+       unsigned int            flags;
+       bool                    ignore_off;
+
+       xfs_owner_info_unpack(oinfo, &owner, &offset, &flags);
+       ignore_off = XFS_RMAP_NON_INODE_OWNER(owner) ||
+                       (flags & XFS_RMAP_BMBT_BLOCK);
+       if (unwritten)
+               flags |= XFS_RMAP_UNWRITTEN;
+       trace_xfs_rmap_unmap(mp, cur->bc_private.a.agno, bno, len,
+                       unwritten, oinfo);
+
+       /*
+        * We should always have a left record because there's a static record
+        * for the AG headers at rm_startblock == 0 created by mkfs/growfs that
+        * will not ever be removed from the tree.
+        */
+       error = xfs_rmap_lookup_le(cur, bno, len, owner, offset, flags, &i);
+       if (error)
+               goto out_error;
+       XFS_WANT_CORRUPTED_GOTO(mp, i == 1, out_error);
+
+       error = xfs_rmap_get_rec(cur, &ltrec, &i);
+       if (error)
+               goto out_error;
+       XFS_WANT_CORRUPTED_GOTO(mp, i == 1, out_error);
+       trace_xfs_rmap_lookup_le_range_result(cur->bc_mp,
+                       cur->bc_private.a.agno, ltrec.rm_startblock,
+                       ltrec.rm_blockcount, ltrec.rm_owner,
+                       ltrec.rm_offset, ltrec.rm_flags);
+       ltoff = ltrec.rm_offset;
+
+       /*
+        * For growfs, the incoming extent must be beyond the left record we
+        * just found as it is new space and won't be used by anyone. This is
+        * just a corruption check as we don't actually do anything with this
+        * extent.  Note that we need to use >= instead of > because it might
+        * be the case that the "left" extent goes all the way to EOFS.
+        */
+       if (owner == XFS_RMAP_OWN_NULL) {
+               XFS_WANT_CORRUPTED_GOTO(mp, bno >= ltrec.rm_startblock +
+                                               ltrec.rm_blockcount, out_error);
+               goto out_done;
+       }
+
+       /* Make sure the unwritten flag matches. */
+       XFS_WANT_CORRUPTED_GOTO(mp, (flags & XFS_RMAP_UNWRITTEN) ==
+                       (ltrec.rm_flags & XFS_RMAP_UNWRITTEN), out_error);
+
+       /* Make sure the extent we found covers the entire freeing range. */
+       XFS_WANT_CORRUPTED_GOTO(mp, ltrec.rm_startblock <= bno &&
+               ltrec.rm_startblock + ltrec.rm_blockcount >=
+               bno + len, out_error);
+
+       /* Make sure the owner matches what we expect to find in the tree. */
+       XFS_WANT_CORRUPTED_GOTO(mp, owner == ltrec.rm_owner ||
+                                   XFS_RMAP_NON_INODE_OWNER(owner), out_error);
+
+       /* Check the offset, if necessary. */
+       if (!XFS_RMAP_NON_INODE_OWNER(owner)) {
+               if (flags & XFS_RMAP_BMBT_BLOCK) {
+                       XFS_WANT_CORRUPTED_GOTO(mp,
+                                       ltrec.rm_flags & XFS_RMAP_BMBT_BLOCK,
+                                       out_error);
+               } else {
+                       XFS_WANT_CORRUPTED_GOTO(mp,
+                                       ltrec.rm_offset <= offset, out_error);
+                       XFS_WANT_CORRUPTED_GOTO(mp,
+                                       ltoff + ltrec.rm_blockcount >= offset + len,
+                                       out_error);
+               }
+       }
+
+       if (ltrec.rm_startblock == bno && ltrec.rm_blockcount == len) {
+               /* exact match, simply remove the record from rmap tree */
+               trace_xfs_rmap_delete(mp, cur->bc_private.a.agno,
+                               ltrec.rm_startblock, ltrec.rm_blockcount,
+                               ltrec.rm_owner, ltrec.rm_offset,
+                               ltrec.rm_flags);
+               error = xfs_btree_delete(cur, &i);
+               if (error)
+                       goto out_error;
+               XFS_WANT_CORRUPTED_GOTO(mp, i == 1, out_error);
+       } else if (ltrec.rm_startblock == bno) {
+               /*
+                * overlap left hand side of extent: move the start, trim the
+                * length and update the current record.
+                *
+                *       ltbno                ltlen
+                * Orig:    |oooooooooooooooooooo|
+                * Freeing: |fffffffff|
+                * Result:            |rrrrrrrrrr|
+                *         bno       len
+                */
+               ltrec.rm_startblock += len;
+               ltrec.rm_blockcount -= len;
+               if (!ignore_off)
+                       ltrec.rm_offset += len;
+               error = xfs_rmap_update(cur, &ltrec);
+               if (error)
+                       goto out_error;
+       } else if (ltrec.rm_startblock + ltrec.rm_blockcount == bno + len) {
+               /*
+                * overlap right hand side of extent: trim the length and update
+                * the current record.
+                *
+                *       ltbno                ltlen
+                * Orig:    |oooooooooooooooooooo|
+                * Freeing:            |fffffffff|
+                * Result:  |rrrrrrrrrr|
+                *                    bno       len
+                */
+               ltrec.rm_blockcount -= len;
+               error = xfs_rmap_update(cur, &ltrec);
+               if (error)
+                       goto out_error;
+       } else {
+
+               /*
+                * overlap middle of extent: trim the length of the existing
+                * record to the length of the new left-extent size, increment
+                * the insertion position so we can insert a new record
+                * containing the remaining right-extent space.
+                *
+                *       ltbno                ltlen
+                * Orig:    |oooooooooooooooooooo|
+                * Freeing:       |fffffffff|
+                * Result:  |rrrrr|         |rrrr|
+                *               bno       len
+                */
+               xfs_extlen_t    orig_len = ltrec.rm_blockcount;
+
+               ltrec.rm_blockcount = bno - ltrec.rm_startblock;
+               error = xfs_rmap_update(cur, &ltrec);
+               if (error)
+                       goto out_error;
+
+               error = xfs_btree_increment(cur, 0, &i);
+               if (error)
+                       goto out_error;
+
+               cur->bc_rec.r.rm_startblock = bno + len;
+               cur->bc_rec.r.rm_blockcount = orig_len - len -
+                                                    ltrec.rm_blockcount;
+               cur->bc_rec.r.rm_owner = ltrec.rm_owner;
+               if (ignore_off)
+                       cur->bc_rec.r.rm_offset = 0;
+               else
+                       cur->bc_rec.r.rm_offset = offset + len;
+               cur->bc_rec.r.rm_flags = flags;
+               trace_xfs_rmap_insert(mp, cur->bc_private.a.agno,
+                               cur->bc_rec.r.rm_startblock,
+                               cur->bc_rec.r.rm_blockcount,
+                               cur->bc_rec.r.rm_owner,
+                               cur->bc_rec.r.rm_offset,
+                               cur->bc_rec.r.rm_flags);
+               error = xfs_btree_insert(cur, &i);
+               if (error)
+                       goto out_error;
+       }
+
+out_done:
+       trace_xfs_rmap_unmap_done(mp, cur->bc_private.a.agno, bno, len,
+                       unwritten, oinfo);
+out_error:
+       if (error)
+               trace_xfs_rmap_unmap_error(mp, cur->bc_private.a.agno,
+                               error, _RET_IP_);
+       return error;
+}
+
+/*
+ * Remove a reference to an extent in the rmap btree.
+ */
 int
 xfs_rmap_free(
        struct xfs_trans        *tp,
@@ -143,19 +349,23 @@ xfs_rmap_free(
        struct xfs_owner_info   *oinfo)
 {
        struct xfs_mount        *mp = tp->t_mountp;
-       int                     error = 0;
+       struct xfs_btree_cur    *cur;
+       int                     error;
 
        if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
                return 0;
 
-       trace_xfs_rmap_unmap(mp, agno, bno, len, false, oinfo);
-       if (1)
+       cur = xfs_rmapbt_init_cursor(mp, tp, agbp, agno);
+
+       error = xfs_rmap_unmap(cur, bno, len, false, oinfo);
+       if (error)
                goto out_error;
-       trace_xfs_rmap_unmap_done(mp, agno, bno, len, false, oinfo);
+
+       xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
        return 0;
 
 out_error:
-       trace_xfs_rmap_unmap_error(mp, agno, error, _RET_IP_);
+       xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
        return error;
 }