[LU-19015] Possible records skipping during changelog processing when an ENOSPC occurred while writing a record

• Type: Bug
• Resolution: Fixed
• Priority: Minor
• Fix Version/s: Lustre 2.17.0
• Severity: 3
• Assignee: Alexander Boyko
• Reporter: Alexander Boyko

Description

Multiple threads get into mdd_changelog_write_rec(), they all get their own offsets and now start dt_record_write() in parallel.
Due to scheduling one thread with an offset corresponding to block X+1 starts first and allocates block X+1, few more threads now can write to X+1.
Another thread with block X can't allocate it due to ENOSPC.

This situation leads to a sparse file for the changelog. When processing, dt_read() will get a zeroed block for an offset that has not been written. And llog_process_thread() will skip an entire chunk (2 blocks), losing changelog records from the second block. I understand that there is a small chance of this happening, but it is still possible that regression could occur after LU-18218.

Issue Links

• is related to: LU-18218 Performance impact on MDT performance with changelogs enabled

Comments

[Peter Jones]

Merged for 2.17

[Gerrit Updater]

"Oleg Drokin <[e-mail]>" merged in patch https://review.whamcloud.com/c/fs/lustre-release/+/59267/
Subject: LU-19015 llog: logic for skipping a zeroed record
Project: fs/lustre-release
Branch: master
Current Patch Set:
Commit: 919c5d25fc45121466ae0ea803558039a2162538

[Gerrit Updater]

"Alexander Boyko <[e-mail]>" uploaded a new patch: https://review.whamcloud.com/c/fs/lustre-release/+/59267
Subject: LU-19015 llog: logic for skipping a zeroed record
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 89c2d5b51cc436311b755ba90bb3408450c719eb