Description of problem: ------------------------ `cat' on a file that was in split-brain condition was successful. This should ideally fail with `Input/output error'. Version-Release number of selected component (if applicable): -------------------------------------------------------------- glusterfs-3.7.0beta1-0.69.git1a32479.el6.x86_64 How reproducible: ------------------ Always Steps to Reproduce: -------------------- 1. Create a distributed-replicate volume and mount it via fuse. 2. Create a file `1' on the mount point - # touch 1 3. Bring down one brick in the replica pair where `1' resides. #kill -9 <pid-of-brick-process> 4. Write to the file - # echo "Hello" > 1 5. Start volume with force option. 6. Bring down the other brick in the replica pair and write to the file again - # echo "World" > 1 7. `cat' the file - # cat 1 Actual results: ---------------- # cat 1 World Expected results: ------------------ `cat' should fail with `Input/output error'. Additional info: ----------------- The volume configuration - # gluster volume info 2-test Volume Name: 2-test Type: Distributed-Replicate Volume ID: 0e312bd3-0473-4fdc-ba2f-7df53b9e9683 Status: Started Number of Bricks: 6 x 2 = 12 Transport-type: tcp Bricks: Brick1: dhcp37-126.lab.eng.blr.redhat.com:/rhs/brick4/b1 Brick2: dhcp37-123.lab.eng.blr.redhat.com:/rhs/brick4/b1 Brick3: dhcp37-98.lab.eng.blr.redhat.com:/rhs/brick4/b1 Brick4: dhcp37-54.lab.eng.blr.redhat.com:/rhs/brick4/b1 Brick5: dhcp37-210.lab.eng.blr.redhat.com:/rhs/brick4/b1 Brick6: dhcp37-59.lab.eng.blr.redhat.com:/rhs/brick4/b1 Brick7: dhcp37-126.lab.eng.blr.redhat.com:/rhs/brick5/b1 Brick8: dhcp37-123.lab.eng.blr.redhat.com:/rhs/brick5/b1 Brick9: dhcp37-98.lab.eng.blr.redhat.com:/rhs/brick5/b1 Brick10: dhcp37-54.lab.eng.blr.redhat.com:/rhs/brick5/b1 Brick11: dhcp37-210.lab.eng.blr.redhat.com:/rhs/brick5/b1 Brick12: dhcp37-59.lab.eng.blr.redhat.com:/rhs/brick5/b1 Options Reconfigured: performance.readdir-ahead: on cluster.self-heal-daemon: off cluster.entry-self-heal: off cluster.data-self-heal: off cluster.metadata-self-heal: off features.uss: enable features.quota: on performance.write-behind: off performance.read-ahead: off performance.io-cache: off performance.stat-prefetch: off performance.quick-read: off performance.open-behind: off features.bitrot: on features.scrub: Active diagnostics.client-log-level: DEBUG
Observations from debugging the setup. When debugging the mount process with gdb, it was observed that in afr_lookup_done, we do afr_inode_read_subvol_reset() and consequently when afr_read_txn(), afr_read_txn_refresh_done() is called, we bail out because there are no read subvols and the client gets EIO. When no gdb was attached, the client again began reading stale data. On further examination, it was observed that fuse sends the following FOPS when 'cat' was performed on the mount: 1)fuse_fop_resume-->fuse_lookup_resume 2)fuse_fop_resume-->fuse_open_resume 3)fuse_fop_resume-->fuse_getattr_resume--->afr_fstat-->afr_read_txn-->bail out with EIO. 4)fuse_fop_resume-->fuse_flush_resume However when 'cat' was done in rapid succession, (3) was not being called. i.e only fuse_lookup_resume, fuse_open_resume and fuse_flush_resume were being called. Since the getattr was not sent by fuse, it did not get the EIO and was serving data from kernel cache. It was noted that this data returned was always the one written to the latest brick, "World" in this case. I don't think we should hit the issue if we perform a 1) drop_caches on the existing mount, or 2) do a remount or 3)mount with the options attribute-timeout and entry-timeout set to zero to begin with.
> > I don't think we should hit the issue if we perform a 1) drop_caches on the > existing mount, or 2) do a remount or 3)mount with the options > attribute-timeout and entry-timeout set to zero to begin with. Tried each of the above 3 and did not hit the issue.
Can be closed now that it is proved it kernel cache in action? or can be this taken as a feature? Ravi, I guess you can decide.
Raghavendra G has suggested a fix where we can set attribute-timeout to zero for the files that are in split-brain forcing fuse to send a fuse_getattr_resume(). I'll send a patch for it, let us see if it is acceptable. Keeping the bug open until then.
Closing this as a duplicate of 1229226 (instead of the other way around) because there's more discussion there. *** This bug has been marked as a duplicate of bug 1229226 ***