xfs: buffer type changes in xfs_da_root_join From: Dave Chinner <dchinner@xxxxxxxxxx> When a root join occurs, a node format tree can be turne dback into a leaf format tree by copying the last remaining leaf block into the root block. At this point, the root block changes from a xfs_da_node to a xfs_attr_leaf, and the magic number changes. However, xfs_da_root_join treats this block as a xfs_da_node and sets the verifier up as such. This is the wrong verifier when a leaf block is moved into the root. Fix this by copying the verifier from the buffer being copied into the root block. This ensures that the leaf verifier is moved with the leaf data, and avoids verifier failures. Signed-off-by: Dave Chinner <dchinner@xxxxxxxxxx> --- I posted this as separate patch rather than going back an integrating it and then rebasing all the other patches on top it afterwards. Simpler, faster, and the buf being fixed is more obvious and easier to review, too. Found from debugging occassional 070 and 117 failures when run in a tight loop for a couple of hours.. fs/xfs/xfs_da_btree.c | 8 ++++++++ 1 file changed, 8 insertions(+) diff --git a/fs/xfs/xfs_da_btree.c b/fs/xfs/xfs_da_btree.c index 43db040..eecdf55 100644 --- a/fs/xfs/xfs_da_btree.c +++ b/fs/xfs/xfs_da_btree.c @@ -835,7 +835,15 @@ xfs_da_root_join(xfs_da_state_t *state, xfs_da_state_blk_t *root_blk) xfs_da_blkinfo_onlychild_validate(bp->b_addr, be16_to_cpu(oldroot->hdr.level)); + /* + * This could be copying a leaf back into the root block in the case of + * there only being a single leaf block left in the tree. Hence we have + * to update the buf ops pointer as well to match the buffer type change + * that could occur. + */ memcpy(root_blk->bp->b_addr, bp->b_addr, state->blocksize); + root_blk->bp->b_ops = bp->b_ops; + xfs_trans_log_buf(args->trans, root_blk->bp, 0, state->blocksize - 1); error = xfs_da_shrink_inode(args, child, bp); return(error); _______________________________________________ xfs mailing list xfs@xxxxxxxxxxx http://oss.sgi.com/mailman/listinfo/xfs
