[LU-587] mgs have corrupted config llog but forbid to re-register a target

• Status: Closed
• Project: Lustre
• Type: Bug
• Priority: Minor
• Reporter: Alexey Lyashkov
• Assignee: Andreas Dilger
• Resolution: Incomplete

Environment

lustre 2.1/RHEL6

• Severity: 2

Description

after hard reset mgs configuration llog can truncated. in case OST config llog also corrupted it's prevent ost to be started. Same situation can be occurred with MDT also.
it's easy replicated by truncate mgs config llog to zero.
In that case mgs have bit set - 'that target registered' and forbid to re-register target, but none configuration send to OST.

Comments

[Peter Jones]

Shadow, can you please resubmit your patch - http://review.whamcloud.com/#change,1209 - again (after merging with current master so that the hash changes). Unfortunately it was submitted at unfortunate time where our testing infrastructure was having problems and os it was not test-built and tested. Thanks

[Alexey Lyashkov]

Peter,

Done

[Peter Jones]

Thx

[Andreas Dilger]

I'm trying to understand this bug and the associated patch, but there is not nearly enough explanation in either to understand what is broken, how it is breaking, or why the patch would fix it. Error messages from the console or kernel debug log would definitely help, as to an explanation of why opening a file read-write would allow it to be corrupted on a hard reset, and why opening it read-only prevents that corruption.

Is there a specific test case that causes this corruption? I can't imagine that the local config llog gets corrupted after a hard reset, regardless of the open mode, unless the reset happens immediately after the llog is written and has not yet made it to disk.

[Andreas Dilger]

There isn't any useful information in this bug, or the patch. Closing until some information about how/when this problem is being seen is provided.