-
Bug
-
Resolution: Unresolved
-
Medium
-
None
-
None
-
None
-
3
-
9223372036854775807
sanity test_27Ca created "lfs setstripe -i 1 -c -1" file with only 7 of 8 stripes: OST index list "1 5 2 6 7 3 4", OST0000 missing, so the test reported "Can not find 0 in obdidx 1 5 2 6 7 3 4" (and the awk parse printed the obdidx header as one element, since getstripe only listed 7 OST rows).
The MDS debug log of the failed test shows lod_prepare_create() -> lod_qos_statfs_update() -> lod_statfs_and_check() returning rc = -11 (-EAGAIN) for the first OST in the walk (OST0000, statfs not yet available a few seconds after setup), while the other seven OSTs return 0. OST0000 was therefore skipped by wide striping (-c -1) for that create. This is an MDS allocator/statfs race (see also LU-20914 on LOD statfs age), not a getstripe parsing problem, and is unrelated to the osc/mdc truncate change it was seen on (change 70505, ps2, review-ldiskfs-dne-arm, DNE with 8 OSTs, aarch64 clients).
Rate: one failure in ~19400 sanity runs of lustre-reviews in 30 days for this error; 1 of 605 runs of review-ldiskfs-dne-arm on master. No earlier instance in 730 days.
Possible fix: have test_27Ca wait for all OSTs to be usable (e.g. wait_osts_up / wait_osp_statfs) before the loop, or have LOD retry statfs for an OST with a pending statfs rather than skipping it for -c -1.
Raised by Patch Watcher run pw-engineer-70505-ps2-3a752c70b1d2 on https://review.whamcloud.com/70505