[LU-11889] sanity test 156 fails on ZFS: roc_hit not safe to use

• Status: Open
• Project: Lustre
• Affects Version/s: Lustre 2.10.7, Lustre 2.12.1
• Type: Bug
• Priority: Trivial
• Reporter: Alex Zhuravlev
• Assignee: Alex Zhuravlev
• Resolution: Unresolved

Environment

zfs

• Severity: 3

Description

because osd-zfs doesn't maintain cache hit/miss stats, there is just no regular API for this at the moment.

Issue Links

• is related to: LU-2261 Add cache stats to zfs osd
• is related to: LU-11607 Reduce repeated function calls in Lus...
• is related to: LU-5375 Failure on test suite sanity test_151...

Comments

[Gerrit Updater]

Alex Zhuravlev ([e-mail]) uploaded a new patch: https://review.whamcloud.com/34114
Subject: LU-11889 tests: disable sanity/156 for ZFS OST
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: e138216b361324b2fb4da7384b07f6f086d29649

[Andreas Dilger]

Are there examples of this failing in autotest?

[Alex Zhuravlev]

yes, I found number of failures like this:
https://testing.whamcloud.com/test_sets/f337b3f4-20cb-11e9-af70-52540065bddc

[Andreas Dilger]

It looks like this bug originated in LU-2261, which enabled sanity test_156 for ZFS when brw_stats was enabled for ZFS, even though the cache_hits tunable is not actually being tracked for ZFS. The error went unnoticed because there was a syntax error in the "skip" logic until patch "LU-11607 tests: create routine to get Lustre env" landed, which properly enabled the test for ZFS and proceeded to cause all the ZFS tests to fail.

[Gerrit Updater]

Andreas Dilger ([e-mail]) merged in patch https://review.whamcloud.com/34114/
Subject: LU-11889 tests: disable sanity/156 for ZFS OST
Project: fs/lustre-release
Branch: master
Current Patch Set:
Commit: 42c4dab3c817f9f03efe457fd33e946ed68fab14

[Andreas Dilger]

Alex, is it possible to easily implement the cache_hits tracking for ZFS?

[Alex Zhuravlev]

adilger I checked ZFS sources and it looks that there is no easy way to get stats for Lustre-related dbufs - all stats ZFS collects are global. We could pre-lookup dbufs and check its status, but that would be +1 lookup for every dbuf which isn't free. not sure we want to pay that price.