[LU-12967] sanity test 80 silently fails to get sync_on_lock_cancel parameter

• Type: Bug
• Resolution: Fixed
• Priority: Minor
• Fix Version/s: Lustre 2.14.0, Lustre 2.12.4
• Affects Version/s: Lustre 2.12.0, Lustre 2.13.0, Lustre 2.14.0
• Severity: 3
• Assignee: James Nunez (Inactive)
• Reporter: James Nunez (Inactive)

Description

There are actually two issues with sanity test_80 as see here in the suite_log to https://testing.whamcloud.com/test_sets/92b52b08-05c5-11ea-9487-52540065bddc

== sanity test 80: Page eviction is equally fast at high offsets too  ================================ 19:09:39 (1573585779)
CMD: trevis-34vm3 lctl get_param -n obdfilter.*.sync_on_lock_cancel
trevis-34vm3: error: get_param: param_path 'obdfilter/*/sync_on_lock_cancel': No such file or directory
CMD: trevis-34vm3trevis-34vm3trevis-34vm3trevis-34vm3trevis-34vm3trevis-34vm3trevis-34vm3trevis-34vm3 lctl set_param obdfilter.*.sync_on_lock_cancel=never
pdsh@trevis-34vm1: gethostbyname("trevis-34vm3trevis-34vm3trevis-34vm3trevis-34vm3trevis-34vm3trevis-34vm3trevis-34vm3trevis-34vm3") failed
1+0 records in
1+0 records out
1048576 bytes (1.0 MB) copied, 0.00475709 s, 220 MB/s
CMD: trevis-34vm3trevis-34vm3trevis-34vm3trevis-34vm3trevis-34vm3trevis-34vm3trevis-34vm3trevis-34vm3 lctl set_param obdfilter.*.sync_on_lock_cancel=
pdsh@trevis-34vm1: gethostbyname("trevis-34vm3trevis-34vm3trevis-34vm3trevis-34vm3trevis-34vm3trevis-34vm3trevis-34vm3trevis-34vm3") failed

One issue is that it looks like the obdfilter.*.sync_on_lock_cancel parameter no longer exists and the host name composed in the test is problematic.

The test does not fail when it hits these issues, but most likely, the test is not working as intended.

Issue Links

• is related to: LU-15106 deprecated obdfilter parameter warn on read
• is related to: LU-10496 sanity-dom sanity test_39k: FAIL: mtime is lost on close: 1515705427, should be 1484169395
• is related to: LU-8066 Move lustre procfs handling to sysfs and debugfs.

Comments

[Gerrit Updater]

Oleg Drokin ([e-mail]) merged in patch https://review.whamcloud.com/37038/
Subject: LU-12967 tgt: clean up sync_on_cancel references
Project: fs/lustre-release
Branch: b2_12
Current Patch Set:
Commit: 1e43d3c4fc8432bebacfa7a8a320163e3b15448d

[Gerrit Updater]

Oleg Drokin ([e-mail]) merged in patch https://review.whamcloud.com/37037/
Subject: LU-12967 ofd: restore sync_on_lock_cancel tunable
Project: fs/lustre-release
Branch: b2_12
Current Patch Set:
Commit: cdffffb73080ad1100549afbcdbb09d3ee7a1c50

[Gerrit Updater]

Minh Diep ([e-mail]) uploaded a new patch: https://review.whamcloud.com/37038
Subject: LU-12967 tgt: clean up sync_on_cancel references
Project: fs/lustre-release
Branch: b2_12
Current Patch Set: 1
Commit: 894d051d1161b6a2462fdac59df2ed4668315854

[Gerrit Updater]

Minh Diep ([e-mail]) uploaded a new patch: https://review.whamcloud.com/37037
Subject: LU-12967 ofd: restore sync_on_lock_cancel tunable
Project: fs/lustre-release
Branch: b2_12
Current Patch Set: 1
Commit: 7e3ac15524e1178674e0c3e50945215b82883b0f

[Peter Jones]

Landed for 2.14

[Gerrit Updater]

Oleg Drokin ([e-mail]) merged in patch https://review.whamcloud.com/36754/
Subject: LU-12967 tgt: clean up sync_on_cancel references
Project: fs/lustre-release
Branch: master
Current Patch Set:
Commit: 52a5981be4df863088168b3ea41fac9e29ddf060

[Gerrit Updater]

Oleg Drokin ([e-mail]) merged in patch https://review.whamcloud.com/36748/
Subject: LU-12967 ofd: restore sync_on_lock_cancel tunable
Project: fs/lustre-release
Branch: master
Current Patch Set:
Commit: 7df7347b7b188e7168e094304fd6d2d985f7f274

[Gerrit Updater]

Andreas Dilger ([e-mail]) uploaded a new patch: https://review.whamcloud.com/36754
Subject: LU-12967 tgt: clean up sync_on_cancel references
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: bb49bfca3ea3434a72e54fc6001f2d71c52d761d

[Andreas Dilger]

It would make sense to write a conf-sanity test that extracts the "lctl list_param -R '*'" output on the client, MDS, and OSS, removes duplicate entries (e.g. removes instance (e.g. "-ffff012344567"), filesystem name (e.g. "lustre-"), NIDs (e.g. "192.168.10.1@tcp"), target index numbers (e.g. "OST0000" from parameters that make them unique to a specific mount/fsname/configuration), then sorts and removes duplicates and compares with a previously-saved parameter list.

It should not be considered an error if new parameters are added, but it should be an error to remove existing parameters (or at least we will be notified of this and can make a proper decision about it). The lists should be saved with the version running on the current node (e.g. OSS, MDS, client) and not on the version of the test script, so that this does not cause errors during interop. The list can be updated right before a release so that we don't have to update the saved parameter list continually (though that is not bad either), but we also don't accidentally lose parameters between releases.

[Andreas Dilger]

It looks like the "ofd.*.sync_on_lock_cancel" tunable was broken by patch https://review.whamcloud.com/33059 "LU-8066 ofd: migrate from proc to sysfs" due to implicit use of the function name as the parameter name. That patch was landed and is part of the 2.12.0 release, so we can't just revert the tunable name to "sync_on_lock_cancel". It also isn't just a matter of restoring the old tunable name, since the "mdt.*.sync_lock_cancel" name is also used since 2.8, and the code for the two tunables was recently consolidated in the server target code in patch https://review.whamcloud.com/34190 "LU-10496 tgt: move FMD handling from OFD to target", and in the long run it is better to have a single tunable name for both.

Instead, I think the best path forward is to keep the common "sync_lock_cancel" tunable name for both MDT and OST, and add backward compatibility for "ofd.*.sync_on_lock_cancel" for a number of releases, and print a deprecation warning if the old name is used.