Description of problem: Occasionally, dd on a sharded file in tests/bugs/shard/bug-1251824.t fails with EIO. Turns out this is caused by gfid-mismatch between the replicas. On investigation, it was found that this is due to a race between posix mknod and posix lookup. posix mknod has 3 important stages, among other operations: 1. creation of the file itself 2. setting the gfid xattr on the file, and 3. creating the gfid link under .glusterfs. Now assume the thread doing posix mknod has executed steps 1 and 2 and is on its way to executing 3. And a parallel lookup from another thread sees that loc->inode->gfid is NULL, so it tries to perform gfid_heal and also attempts to create the gfid link under .glusterfs. Assume lookup wins the race and creates the gfid link. posix_gfid_set() through mknod fails with EEXIST. In the older code, mknod under such conditions was NOT being treated as a failure. But ever since the following commit was merged: <commit-msg> Parent: 788cda4c (glusterd: fix some coverity issues) Author: karthik-us <ksubrahm> AuthorDate: 2018-08-03 15:55:18 +0530 Commit: Amar Tumballi <amarts> CommitDate: 2018-08-20 12:14:22 +0000 posix: Delete the entry if gfid link creation fails Problem: If the gfid link file inside .glusterfs is not present for a file, the operations which are dependent on the gfid will fail, complaining the link file does not exists inside .glusterfs. Fix: If the link file creation fails, fail the entry creation operation and delete the original file. Change-Id: Id767511de2da46b1f45aea45cb68b98d965ac96d fixes: bz#1612037 Signed-off-by: karthik-us <ksubrahm> </commit-msg> ... this behavior changes and the mknod is treated as failure and the subsequent entry deleted. When sometime in future, shard sends another mknod on the shard, the file is created, although this time with a new gfid (since "gfid-req" that is passed now is a new UUID. This leads to a gfid-mismatch across the replicas. Version-Release number of selected component (if applicable): How reproducible: Fairly consistently. Just run the test tests/bugs/shard/bug-1251824.t in a loop on your laptop. I was able to hit it in less than 5 mins time. Steps to Reproduce: 1. 2. 3. Actual results: Expected results: Additional info:
REVIEW: https://review.gluster.org/21436 (storage/posix: Do not fail entry creation fops if gfid handle already exists) posted (#1) for review on master by Krutika Dhananjay
COMMIT: https://review.gluster.org/21436 committed in master by "Raghavendra Bhat" <raghavendra> with a commit message- storage/posix: Do not fail entry creation fops if gfid handle already exists PROBLEM: tests/bugs/shard/bug-1251824.t fails occasionally with EIO due to gfid mismatch across replicas on the same shard when dd is executed. CAUSE: Turns out this is due to a race between posix_mknod() and posix_lookup(). posix mknod does 3 operations, among other things: 1. creation of the entry itself under its parent directory 2. setting the gfid xattr on the file, and 3. creating the gfid link under .glusterfs. Consider a case where the thread doing posix_mknod() (initiated by shard) has executed steps 1 and 2 and is on its way to executing 3. And a parallel LOOKUP from another thread on noting that loc->inode->gfid is NULL, tries to perform gfid_heal where it attempts to create the gfid link under .glusterfs and succeeds. As a result, posix_gfid_set() through MKNOD (step 3) fails with EEXIST. In the older code, MKNOD under such conditions was NOT being treated as a failure. But commit e37ee6d changes this behavior by failing MKNOD, causing the entry creation to be undone in posix_mknod() (it's another matter that the stale gfid handle gets left behind if lookup has gone ahead and gfid-healed it). All of this happens on only one replica while on the other MKNOD succeeds. Now if a parallel write causes shard translator to send another MKNOD of the same shard (shortly after AFR releases entrylk from the first MKNOD), the file is created on the other replica too, although with a new gfid (since "gfid-req" that is passed now is a new UUID. This leads to a gfid-mismatch across the replicas. FIX: The solution is to not fail MKNOD (or any other entry fop for that matter that does posix_gfid_set()) if the .glusterfs link creation fails with EEXIST. Change-Id: I84a5e54d214b6c47ed85671a880bb1c767a29f4d fixes: bz#1638453 Signed-off-by: Krutika Dhananjay <kdhananj>
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-6.0, please open a new bug report. glusterfs-6.0 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] https://lists.gluster.org/pipermail/announce/2019-March/000120.html [2] https://www.gluster.org/pipermail/gluster-users/