[KUDU-1857] Retry LBM hole punching on crash

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 1.4.0
• Component/s: fs
• Code Review: https://gerrit.cloudera.org/#/c/6717/
• Assignee: Adar Dembo
• Reporter: Adar Dembo
• Created: 01/Feb/17 03:59
• Updated: 02/May/17 21:13
• Resolved: 02/May/17 21:13

Description

Given the LBM design, it's possible for block deletion to be committed via a metadata write, but for the subsequent hole punch to fail, or for the server to crash in between. As such, we should provide a mechanism to "re-hole-punch" deleted blocks. This could be dumb and hole punch every dead block in the container, or it could be smart and use the FIEMAP ioctl to figure out which dead blocks have live extents and surgically punch them.

Separately, this should be runnable at startup, and perhaps as part of a general filesystem checker in the CLI tool.

Issue Links

• is related to: KUDU-1856 Kudu can consume far more data than it should on XFS