[LU-27] (mds_open.c:1667:mds_close()) @@@ no handle for file close ino

• Status: Resolved
• Project: Lustre
• Affects Version/s: Lustre 1.8.6
• Fix Version/s: Lustre 1.8.6
• Type: Bug
• Priority: Minor
• Reporter: Shuichi Ihara (Inactive)
• Assignee: Niu Yawei (Inactive)
• Resolution: Fixed
• Severity: 3

Description

We are hitting frequent MDS hangs at Titech due to LBUG caused by "no handle for file close in mds_close()".
it looks like similar bug 22104 and 22528, but no solution and patches yet to address this problem.
could you have a look at attachment and any suggestions?

Comments

[Liang Zhen (Inactive)]

I think the LBUG is because error handler in mds_verify_child() is not quite right:

switch(cleanup_phase) {
case 2:
if (child_res_id->name[0] != 0)
ldlm_lock_decref(child_lockh, child_mode);

we do have chance to get "child_res_id->name[0] != 0" and "child_lockh == NULL" at here, probably Niu can look into it to see whether these is any other issue, as he is also working on open/close related code for 2.x?

[Dan Ferber (Inactive)]

Assigned to Niu, per Liang's comments and suggestion.

[Niu Yawei (Inactive)]

Yes, there are some defects in the mds_verify_child():

• Wrongly decref child lock in the "no child lock wanted" case;
• Wrongly decref parent lock in the "reget child lock successfully" case;

This bug isn't necessarily caused by the "no handle for file close in mds_close()", so I think it's not similar to bug 22104 and 22528.

Will post a patch for review soon.

[Shuichi Ihara (Inactive)]

Niu, just confirmation. you did file this on bugzilla as 24360, then moving forward to review patches, right?

[Niu Yawei (Inactive)]

Yes, the patch has been posted on BZ and Gerrit for review.

[Peter Jones]

It looks like this patch has landed on the Oracle 1.8.6. Is the same fix needed for master?

[Niu Yawei (Inactive)]

No, the master doesn't have this bug.

[Peter Jones]

Great! So, does any work remain or can we mark this issue as resolved?

[Niu Yawei (Inactive)]

Yes, I think we can mark this as resolved, not sure if it should be marked by reporter or assignee.

[Peter Jones]

ok I will mark it as resolved. Ihara, please reopen if you feel that this is inappropriate