Description of problem: An I/O error happens when files are being created and written to a disperse volume when a replace-brick is executed. Version-Release number of selected component (if applicable): mainline How reproducible: Always Steps to Reproduce: 1. Create a disperse volume 2. Kill one brick 3. Open fd on a subdirectory 4. Do a replace brick of the killed brick 5. Write on the previous file Actual results: The write fails with I/O error Expected results: The write should succeed Additional info:
The problem happens because a reopen is attempted on all available bricks and any error it finds is propagated to the main fop. Basically, when a write fop is sent and ec discovers that there's a brick that has come up again but doesn't have the fd open, it tries to open it. It could happen that the file was created when the brick was down and self-heal has not yet recovered it. In this case the open will fail with ENOENT. This should be ok, since the other bricks are perfectly fine to successfully process the write with enough quorum, but this error is not ignored and it's propagated to the main fop, causing it to fail even before attempting the write.
REVIEW: https://review.gluster.org/22558 (cluster/ec: fix fd reopen) posted (#2) for review on master by Xavi Hernandez
REVIEW: https://review.gluster.org/22574 (tests: Heal should fail when read/write fails) merged (#2) on master by Pranith Kumar Karampuri
REVIEW: https://review.gluster.org/22558 (cluster/ec: fix fd reopen) merged (#7) on master by Pranith Kumar Karampuri