[LU-5604] Lots of FAIL_ID checking are lost

• Type: Bug
• Resolution: Fixed
• Priority: Critical
• Fix Version/s: Lustre 2.7.0, Lustre 2.13.0
• Affects Version/s: Lustre 2.6.0, Lustre 2.7.0
• Labels: MB
• Severity: 3
• Assignee: Mikhail Pershin
• Reporter: Niu Yawei (Inactive)

Description

It looks to me that lots of FAIL_ID checking are lost from time to time, take the replay-single.sh as an example:

• test_73c() checked OBD_FAIL_TGT_LAST_REPLAY, but this FAIL_ID is never being checked in Lustre code from the day one it was introduced.

• test_73b() checked OBD_FAIL_LDLM_REPLY, but this FAIL_ID is now only checked in mdt_reint_open(), I think it should be checked for every lock enqueue as well.

• test_73a() checked OBD_FAIL_LDLM_ENQUEUE_NET, but this FAIL_ID is not being checked in Lustre code anymore.

• test_80c() checked OBD_FAIL_UPDATE_OBJ_NET_REP, but this FAIL_ID has been removed from Lustre code.

• test_83a() checked OBD_FAIL_MDS_FAIL_LOV_LOG_ADD, but this FAIL_ID isn't checked in Lustre code.
  ...

To make sure the error injection test working as expected, I think we'd go through all the fail IDs, and add back all the missed fail_id checking. If some FAIL_ID is obsolete already, we'd remove or improve the corresponding test case.

Issue Links

• is blocking: LU-5579 MDS crashed by "mdt_check_resent_lock()) ASSERTION( lock != NULL ) failed"
• is duplicated by: LU-5709 "unified request handlers" commits broke tests
• is related to: LU-10816 fix usage of drop_ldlm_reply() in tests