[LU-6160] all test builds should enable SPL/ZFS debugging

• Type: Improvement
• Resolution: Fixed
• Priority: Major
• Fix Version/s: Lustre 2.12.0
• Labels: zfs
• Assignee: Oleg Drokin
• Reporter: Isaac Huang (Inactive)

Description

ASSERTs in SPL/ZFS are off by default. It'd be a good idea to enable SPL/ZFS assertions at least for test builds, which'd pinpoint problems much earlier before they manifest as hard-to-diagnose symptoms.

I have a patch that enables SPL/ZFS debugging:
http://review.whamcloud.com/#/c/13431/

But there's build failures for i686 in SPL/ZFS. So upstream SPL/ZFS would need a fix for that.

Issue Links

• is related to: LU-6155 osd_count_not_mapped() calls dbuf_hold_impl() without the lock
• is related to: LU-6195 osd-zfs: osd_declare_object_destroy() calls dmu_tx_hold_zap() with wrong keys

Comments

[Olaf Faaland]

I see that Oleg landed the patch to allow Lustre to build against zfs packages with debug enabled.

But were the appropriate change(s) made to your build system to enable debug in zfs for test builds?

[Peter Jones]

Landed for 2.12

[Gerrit Updater]

Oleg Drokin ([e-mail]) merged in patch https://review.whamcloud.com/28544/
Subject: LU-6160 osd-zfs: Fix refcount_add call
Project: fs/lustre-release
Branch: master
Current Patch Set:
Commit: ad7e62cc15e9e90d33a7302308d47566e4af3593

[Giuseppe Di Natale (Inactive)]

I'm submitting https://review.whamcloud.com/28544 just so lustre builds against zfs packages that have debug enabled.

[Gerrit Updater]

Giuseppe Di Natale ([e-mail]) uploaded a new patch: https://review.whamcloud.com/28544
Subject: LU-6160 osd-zfs: Fix refcount_add call
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 510f1536aa60e11271d621e2171dc19bcf8c7a5b

[Peter Jones]

Assigining to Oleg for his comment

[Alex Zhuravlev]

"ran" doesn't exactly mean it's reliable.. not that I'm against running with debug enabled, but there were known issues we never fixed, some can't be fixed outside of ZFS code.

[Olaf Faaland]

I hear that Oleg successfully ran Lustre using ZFS built with debug enabled, and submitted patches for some bugs he found that way.

Was Lustre patched to enable this, so that this ticket can now go forward? Or did he just make some one-off change in his environment, perhaps to ZFS, to enable this for testing purposes?

[Alex Zhuravlev]

well, I don't remember exact lines, but basically in debug mode DMU tracks all writes are declared properly (including offset/range), while LLOG doesn't know exact offset at declare.

[Isaac Huang (Inactive)]

Alex, can you please point me to the code?

[Alex Zhuravlev]

I think Lustre isn't quite ready for this yet. in some cases (like append to llog) we can't declare in the way ZFS's debugging expects.