[LU-2303] Bulk write replay & grant interaction

• Status: Resolved
• Project: Lustre
• Affects Version/s: Lustre 2.4.0
• Type: Bug
• Priority: Major
• Reporter: Mikhail Pershin
• Assignee: Johann Lombardi (Inactive)
• Resolution: Not a Bug
• Labels: zfs
• Story Points: 1
• Severity: 3
• Project: Orion

Description

Grant management when replaying bulk writes need to be reviewed since we might end up with a different result after replay than the original one.

Comments

[Andreas Dilger]

Johann, can you please provide some details about what this bug is for? It is a clone from ORI-384, so possibly it relates to interop with old clients related to "OBD_BRW_FROM_GRANT in the rnb_flags"?

[Johann Lombardi (Inactive)]

No, it is not an interop issue and not something new in 2.4 (it has been around since bulk write replay is enabled).
The issue is that we return an array of rc when processing bulk write and this array isn't resent to the server on replay. As a result, we might end up with a different result when replaying bulk writes than the original one.
e.g.: page 1 was written successfully while page 2 failed with ENOSPC might end up being the two pages are successfully written during replay.