-
Bug
-
Resolution: Unresolved
-
Medium
-
None
-
Lustre 2.16.1
-
None
-
3
-
9223372036854775807
After a successful "lfs mirror extend", the PFID information stored
in the original OST objects may remain inconsistent with the new
layout on the MDT when an OST is full.
mdd_object_pfid_replace() updates the OST objects in a separate
transaction after the layout merge has completed. However,
lod_trans_space_check() rejects this transaction with -ENOSPC
when the target's cached statfs state contains OS_STATFS_ENOSPC
or OS_STATFS_ENOINO.
The mirror extension reports success, but the original objects
retain their old PFID component IDs. Tools inspecting OST objects
can consequently classify them as non-mirrored and skip them
during mirror-based space reclamation.
Steps to reproduce:
1. Create and write a single-stripe file on OST0.
2. Fill OST0 using a separate file and confirm that further writes
fail with ENOSPC.
3. Wait for the MDT's OSP statfs cache to observe the full OST.
4. Extend the file with a mirror on another OST with free space:
lfs mirror extend -N -c1 -o1 <file>
5. Inspect the original object's on-disk filter_fid before further
client I/O can repair it, and compare its component_id with the
corresponding lcme_id in the MDT layout.
Actual result:
Mirror extension succeeds, but the original object's component_id
does not match the updated layout.
Expected result:
PFID replacement should be attempted even when the OST is reported
as full. Updating the existing filter_fid normally reuses existing
inode space, so the LOD space precheck should not prevent it.
Actual backend allocation errors must still be handled.
The proposed fix allows PFID replacement transactions to bypass
the LOD space precheck and adds a regression test for mirror
extension with a full source OST.