[LU-9959] hsm: cannot schedule two different requests on the same fid

• Status: Open
• Project: Lustre
• Type: Bug
• Priority: Minor
• Reporter: CEA
• Assignee: Quentin Bouget
• Resolution: Unresolved
• Labels: HSM
• Severity: 3

Description

mdt_hsm_add_actions() makes the assumption that hsm_find_compatible() will set the hai_cookie field of incoming requests to something other than 0 only if :

• the incoming request is a CANCEL request, and there is a request to cancel;
• the incoming request is a NONE request, and there is a request scheduled for the given fid;
• the incoming request is an ARCHIVE or a RESTORE or a REMOVE, and the same request is already scheduled.

The third assumption is used to detect duplicate requests, but hsm_find_compatible() does not exactly work like this. If the incoming request is an ARCHIVE/RESTORE/REMOVE, hsm_find_compatible() will set the hai_cookie field if it finds any request that applies to the same fid.
This means that an ARCHIVE/RESTORE/REMOVE request on a given fid cannot be scheduled on the coordinator while any other request is still scheduled for that same fid.

This makes sense in some cases:

• one cannot run both an ARCHIVE and a RESTORE on the same file (either the archive is not needed, or the file is not released)
• one cannot run both a RESTORE and a REMOVE on the same file (either the file is not released, or its archive in the backend cannot be purged)

But not in others:

• one can want to ARCHIVE the new version of a file while the previous version is being REMOVEd from the hsm backend.
• one can want to re-run an {ARCHIVE,RESTORE,REMOVE}
  request while another is being canceled.

Issue Links

• is related to: LU-7988 HSM: high lock contention for cdt_llo...

Comments

[Peter Jones]

Quentin

Is this issue something that you are planning to work on or something you are documenting in the hope that someone else will work on it?

Peter

[Quentin Bouget]

Hi Peter,

A little of both. Ben Evans is working on the LU-7988 patch series and the last patch of the series is about calling hsm_find_compatible() less often. I am hoping we can fix both issues at once. If not, I will eventually submit a patch, but probably not until October or November.

Quentin

[Peter Jones]

In that case I think that it makes sense for you to own this ticket for the time being. I also added a link to LU-7988.