[LU-918] ensure that BRW requests prevent lock timeout

• Status: Closed
• Project: Lustre
• Type: Technical task
• Priority: Minor
• Reporter: Andreas Dilger
• Assignee: Jinshan Xiong (Inactive)
• Resolution: Fixed

Description

To avoid lock timeouts in LU-874, one point of discussion was to ensure that BRW requests under a DLM extent lock ensure that the client does not get evicted, even under heavy OST load. The best option is if a client sends a BRW request under a DLM lock that the lock timeout is stopped entirely for that lock until the IO has completed. That would avoid overhead from continually refresh the DLM lock during operation, but may be more complex to implement than having the DLM lock timeout first check for BRW requests in the hpreq queue or in-progress by an OSS thread that would refresh it.

Issue Links

• is related to: LU-15 strange slow IO messages and bad perf...
• is related to: LU-410 Performance concern with Shrink file_...

Comments

[Andreas Dilger]

Add dependency on OSS read cache issues

[Peter Jones]

Jinshan

Did you cover this already under one of your LU874 patches?

Peter

[Jinshan Xiong (Inactive)]

I solved this problem by refreshing lock timeout each time.

Yes, I have ever had a patch to take locks out of waiting list if a covering RPC is coming. However, this way will have to modify the state of dlm lock, for example, to remember how many active RPCs existing. That's a stateful implementation.

After thinking about it, I decided to stay with current stateless implementation because of:
1. less possibilities of producing bugs;
2. lock timeout is rare event in the system, so performance shouldn't be a problem.

How do you think?

[Andreas Dilger]

I think if we have a simple solution to the problem that works, then we don't need a complex solution to the problem. If there is nothing here left to be fixed, then this bug can be closed.