A file created a few seconds after an OST was unmounted was still given an object on that OST (from the OSP precreate pool) in 2 of 4 test runs, and every write() to the new file then blocked for the whole outage (60s and 300s). In the other runs the new file avoided the OST. See the LU-20212 test results.

The MDS OSP import goes to DISCONN at once, but osp_statfs() keeps returning cached statfs and LOD keeps the OST ACTIVE (lod.*.target_obd at T0+20s) until the precreate status changes (prealloc_status=-19). New objects should not be allocated on an OST whose OSP import is not FULL, at least when other OSTs can satisfy the layout, since an EC data stripe cannot be written while its OST is unavailable.