Description of problem: There are few races in remove-brick hardlink migration code path detailed below. A brief about how hardlink migration works: - Different hardlinks (to the same file) may hash to different bricks, but their cached subvol will be same. Rebalance picks up the first hardlink, calculates it's hash(call it TARGET) and set the hashed subvolume as an xattr on the data file. - Now all the hardlinks those come after this will fetch that xattr and will create linkto files on TARGET (all linkto files for the hardlinks will be hardlink to each other on TARGET). - When number of hardlinks on source is equal to the number of hardlinks on TARGET, the data migration will happen. RACE:1 Since rebalance is multi-threaded, the first lookup (which decides where the TARGET subvol should be), can be called by two hardlink migration parallely and they may end up creating linkto files on two different TARGET subvols. Hence, hardlinks won't be migrated. RACE:2 The linkto files on TARGET can be created by other clients also if they are doing lookup on the hardlinks. Consider a scenario where you have 100 hardlinks. When rebalance is migrating 99th hardlink, as a result of continuous lookups from other client, linkcount on TARGET is equal to source linkcount. Rebalance will migrate data on the 99th hardlink itself. On 100th hardlink migration, hardlink will have TARGET as cached subvolume. If it's hash is also the same, then a migration will be triggered from TARGET to TARGET leading to data loss. This is reproducible intermittently. Since this is related to hardlink migration, this happens only with remove-brick process.
REVIEW: https://review.gluster.org/17755 (cluster/rebalance: Fix hardlink migration failures) posted (#1) for review on master by Susant Palai (spalai)
REVIEW: https://review.gluster.org/17755 (cluster/rebalance: Fix hardlink migration failures) posted (#2) for review on master by Susant Palai (spalai)
COMMIT: https://review.gluster.org/17755 committed in master by Raghavendra G (rgowdapp) ------ commit 0d75e39834d4880dce0cb3c79bef4b70bb32874d Author: Susant Palai <spalai> Date: Wed Jul 12 12:01:40 2017 +0530 cluster/rebalance: Fix hardlink migration failures A brief about how hardlink migration works: - Different hardlinks (to the same file) may hash to different bricks, but their cached subvol will be same. Rebalance picks up the first hardlink, calculates it's hash(call it TARGET) and set the hashed subvolume as an xattr on the data file. - Now all the hardlinks those come after this will fetch that xattr and will create linkto files on TARGET (all linkto files for the hardlinks will be hardlink to each other on TARGET). - When number of hardlinks on source is equal to the number of hardlinks on TARGET, the data migration will happen. RACE:1 Since rebalance is multi-threaded, the first lookup (which decides where the TARGET subvol should be), can be called by two hardlink migration parallely and they may end up creating linkto files on two different TARGET subvols. Hence, hardlinks won't be migrated. Fix: Rely on the xattr response of lookup inside gf_defrag_handle_hardlink since it is executed under synclock. RACE:2 The linkto files on TARGET can be created by other clients also if they are doing lookup on the hardlinks. Consider a scenario where you have 100 hardlinks. When rebalance is migrating 99th hardlink, as a result of continuous lookups from other client, linkcount on TARGET is equal to source linkcount. Rebalance will migrate data on the 99th hardlink itself. On 100th hardlink migration, hardlink will have TARGET as cached subvolume. If it's hash is also the same, then a migration will be triggered from TARGET to TARGET leading to data loss. Fix: Make sure before the final data migration, source is not same as destination. RACE:3 Since a hardlink can be migrating to a non-hashed subvolume, a lookup from other client or even the rebalance it self, might delete the linkto file on TARGET leading to hardlinks never getting migrated. This will be addressed in a different patch in future. Change-Id: If0f6852f0e662384ee3875a2ac9d19ac4a6cea98 BUG: 1469964 Signed-off-by: Susant Palai <spalai> Reviewed-on: https://review.gluster.org/17755 Smoke: Gluster Build System <jenkins.org> CentOS-regression: Gluster Build System <jenkins.org> Reviewed-by: N Balachandran <nbalacha> Reviewed-by: Raghavendra G <rgowdapp>
This bug is getting closed because a release has been made available that should address the reported issue. In case the problem is still not fixed with glusterfs-3.12.0, please open a new bug report. glusterfs-3.12.0 has been announced on the Gluster mailinglists [1], packages for several distributions should become available in the near future. Keep an eye on the Gluster Users mailinglist [2] and the update infrastructure for your distribution. [1] http://lists.gluster.org/pipermail/announce/2017-September/000082.html [2] https://www.gluster.org/pipermail/gluster-users/