[LU-6475] race between open and migration

• Type: Bug
• Resolution: Fixed
• Priority: Critical
• Fix Version/s: Lustre 2.8.0
• Affects Version/s: Lustre 2.6.0, Lustre 2.7.0
• Labels: dne2
• Severity: 3
• Assignee: Di Wang
• Reporter: Di Wang

Description

After the directory is migrated to another MDT, the object on the original MDT will be destroyed (mark as DEAD object). If another client is trying to open the file under this directory at the same time (sending RPC to the original MDT), it will then find the object has been destroyed, then return ENOENT to the client, which will cause open fails.

See mdt_reint_open().

   if (result == -ENOENT || result == -ESTALE) {
                mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
                if (result == -ESTALE) {
                        /*
                         * -ESTALE means the parent is a dead(unlinked) dir, so
                         * it should return -ENOENT to in accordance with the
                         * original mds implementaion.
                         */
                        GOTO(out_parent, result = -ENOENT);
                }

In this case, we should return the new parent FID to that client and let it retry open.

Issue Links

• is related to: LU-6831 The ticket for tracking all DNE2 bugs
• is related to: LU-6785 Interop 2.7.0<->master sanity test_56w: cannot swap layouts: Device or resource busy
• is related to: LU-6715 sanity-lfsck test_15c: Fail to repair unmatched pair