test: failure in journal.sh workunit test

[trociny]

No description provided.

[dillaman]

@trociny The new 'rbd journal client status' seems to overlap with 'rbd journal status'? Also, while I can see it might be a useful tool to have 'rbd journal client unregister', I have a feeling that 'rbd journal client register' could lead to some abuse / confusion. Should this command just be moved to a standalone support tool?

[trociny]

@dillaman Agree, these commands are better to have in a standalone support tool. Moreover, I think all rbd journal commands are better to be moved to a standalone tool because they are not supposed to be used normally (may be except rbd journal status and rbd journal info but this could be moved to rbd status and rbd info).

So, what do you think about creating a rbd-journal-tool, which would support info, status, export, import, reset, client register ... commands, and making rbd journal commands deprecated (invisible in help output and printing a warning on their use)?

Just making a tool for register/unregister does not look very useful...

[dillaman]

@trociny Another possibility is to just write a small Python script that handles this special case (see test_lock_fence.sh for an example). I think moving all the journal-related commands to a new CLI is fine, but probably not worth the effort at this point since in a perfect world it would be nice to have feature parity (where it makes sense) with cephfs-journal-tool.

[trociny]

@dillaman I think a Python script whould have worked if we had had journal API?

[dillaman]

@trociny Whoops -- totally forgot that we didn't implement an API. 😢 How about running the following before bench-write:

rados -p rbd getomapval journal.{journal id} client_ {temp file}

After bench-write you can reset the journal position:

rados -p rbd setomapval journal.{journal id} client_ < {temp file}

[dillaman]

@trociny ... or modify the cls_journal client_register method to only do that min commit position reset if the min set != 0 (i.e. we have had trimming)

[trociny]

@dillaman I changed it to store/restore commit position as you proposed.

Still I think for the future it will be useful to have a support tool to register/unregister client. Then it could be used in journal.sh too to make the test more valuable.

Also, I left the patch " treat empty commit position as minimal" in the PR. The problem was observed when a new (non-master) client was registered on a virgin image, so it had empty commit position. When later the master position was updated, it was incorrectly used as minimal.

[dillaman]

lgtm

[dillaman]

http://pulpito.ceph.com/yuriw-2016-05-26_19:14:40-rbd-wip-yuri-testing2-distro-basic-smithi/