PROBLEM:
Consider a volume with granular-entry-heal and sharding enabled. When
a replica is down and a shard is created as part of a write, the name
index is correctly created under indices/entry-changes/<dot-shard-gfid>.
Now when a read on the same region triggers another MKNOD, the fop
fails on the online bricks with EEXIST. By virtue of this being a
symmetric error, the failed_subvols[] array is reset to all zeroes.
Because of this, before post-op, the GF_XATTROP_ENTRY_OUT_KEY will be
set, causing the name index, which was created in the previous MKNOD
operation, to be wrongly deleted in THIS MKNOD operation.
FIX:
The ideal fix would have been for a transaction to delete the name
index ONLY if it knows it is the one that created the index in the first
place. This would involve gathering information as to whether THIS xattrop
created the index from individual bricks, aggregating their responses and
based on the various posisble combinations of responses, decide whether to
delete the index or not. This is rather complex. Simpler fix would be
for post-op to examine local->op_ret in the event of no failed_subvols
to figure out whether to delete the name index or not. This can occasionally
lead to creation of stale name indices but they won't be affecting the IO path
or mess with pending changelogs in any way and self-heal in its crawl of
"entry-changes" directory would take care to delete such indices.
Change-Id: Ic1b5257f4dc9c20cb740a866b9598cf785a1affa
BUG: 1408712
Signed-off-by: Krutika Dhananjay <kdhananj@redhat.com>
Reviewed-on: http://review.gluster.org/16286
Smoke: Gluster Build System <jenkins@build.gluster.org>
Reviewed-by: Pranith Kumar Karampuri <pkarampu@redhat.com>
NetBSD-regression: NetBSD Build System <jenkins@build.gluster.org>
CentOS-regression: Gluster Build System <jenkins@build.gluster.org>