[LU-1517] no retry for the bulk operation

• Type: Bug
• Resolution: Fixed
• Priority: Critical
• Fix Version/s: Lustre 2.4.0, Lustre 2.1.4, Lustre 1.8.9
• Affects Version/s: Lustre 1.8.8
• Labels: patch
• Severity: 3
• Assignee: Keith Mannthey (Inactive)
• Reporter: Alexander Boyko

Description

lnd_cb.c:558:kgnilnd_setup_phys_buffer()) failed to allocate tx_phys
[2012-04-07 02:08:24][c5-0c0s5n2]LNet: 29099:0:(gnilnd_cb.c:1068:kgnilnd_tx_done()) $$ error -12 on tx 0xffff88000fe06b40-><?> id 0/0 state GNILND_TX_ALLOCD age 17481575s  msg@0xffff88000fe06bc0 m/v/ty/ck/pck/pl b00fbabe/8/3/0/78db/0 x0:GNILND_MSG_PUT_REQ
[2012-04-07 02:08:24][c5-0c0s5n2]LustreError: 29099:0:(events.c:198:client_bulk_callback()) event type 0, status -5, desc ffff880627c24000

The error is detected on both client and server; the server expects the client to retry but it doesn't. In the mean time, the OSS issues a lock callback to the client, but the client does not respond because it is waiting for the I/O to complete. Eventually the OSS evicts the client. Lustre does not retry the bulk op when it detects the error.

Issue Links

• Trackbacks: Lustre 1.8.x known issues tracker While testing against Lustre b18 branch, we would hit known bugs which were already reported in Lustre Bugzilla https://bugzilla.lustre.org/. In order to move away from relying on Bugzilla, we would create a JIRA

Comments

[Keith Mannthey (Inactive)]

Ok it seems the patches for 1.8,2.3 and Master has been merged. 2.2 and 2.3 are dead branches at this point.

I think this issue is safe to close. Please reopen if you disagree.

[Cory Spitz]

I think that we should at least land the master patch.

[Keith Mannthey (Inactive)]

The patch ported easily as expected so I sent it to the various branches.

http://review.whamcloud.com/4296 <- b2_1
http://review.whamcloud.com/4297 <- b2_2
http://review.whamcloud.com/4298 <- b2_3
http://review.whamcloud.com/4299 <- Master

[Johann Lombardi (Inactive)]

Sure, let's push review 3102 to all branches first (b1_8, b2_1, b2_2, b2_3 and master). Then more intrusive changes (like 4092) can be considered on master.

[Cory Spitz]

Yes, it is curious that we landed this fix to b1_8 before master. FYI, Cray has been using this fix on our 2.2 for some time and testing has gone well. We should push it to master now while we wait for LU-901.

[Cory Spitz]

Johann, for master, have you seen LU-901 and change #4092? Or does it only partially address the fault?

[Johann Lombardi (Inactive)]

I'm fine with landing the patch on b1_8. That said, it seems that master suffers from the same issue, right?
If so, it would be great to push a patch against master.
Thanks in advance!

[Cory Spitz]

Understood. Trigerring this bug results in client eviction though. Just FYI.