-
Bug
-
Resolution: Not a Bug
-
Medium
-
None
-
Lustre 2.17.0
-
None
-
3
-
9223372036854775807
On a Lustre 2.17.0 cluster, metadata operations (mount, unmount, touch) are experiencing extreme delays (2–4 minutes). While root can eventually complete a touch, non-root users receive "Permission denied" after a significant hang.
Setting identity_upcall to NONE bypasses the delay and restores normal performance, indicating a bottleneck or failure in the l_getidentity_nss upcall path.
Environment
- Lustre Version: 2.17.0_RC4_4_g81318f8
- Upcall Configuration: mdt.* MDT*.identity_upcall=/usr/sbin/l_getidentity_nss
- Identity Source: Remote LDAP via NSS
Symptoms
1. Metadata Latency: ** A touch as root takes over 4 minutes:
# time touch /lus/fs_name/test_file real 4m1.679s user 0m0.001s sys 0m0.000s
2. Permission Denied for Non-Root: ** Users encounter EACCES after a ~2-minute timeout:
$ time touch /lus/fs_name/user_file
touch: cannot touch '/lus/fs_name/user_file': Permission denied
real 2m1.124s
3. MDS Kernel Errors: The MDS nodes report upcall cache acquisition failures with error -110:
LustreError: [PID]:0:(upcall_cache.c:342:upcall_cache_get_entry()) [FS]-MDT0000: acquire for key [UID] after 30: rc = -110
The error rc = -110 in upcall_cache_get_entry suggests that the Lustre MDS is successfully triggering the upcall, but the user space process (l_getidentity_nss) is failing to return identity/group information within the 30-second window.
Steps to Reproduce
- Configure Lustre MDT to use /usr/sbin/l_getidentity_nss.
- Ensure an LDAP-backed NSS configuration is active on the MDS.
- Perform metadata operations from a client.
- Observe rc = -110 in dmesg on the MDS and high real time on the client.
- is related to
-
LU-21264 mdt: identity upcall has no limit on helper processes and does not cache a failed lookup
-
- Open
-