On Mon, Aug 23, 2021 at 03:06:21PM +0000, Trond Myklebust wrote: > On Mon, 2021-08-23 at 10:48 -0400, J. Bruce Fields wrote: > > I'm seeing generic/154 failing sometimes. > > > > It does a "cp --reflink" (which uses FI_CLONE, which results in a > > ->remap_file_range call that NFS maps to the CLONE operation), then > > overwriting parts of the fire, and checking free blocks (with "stat -f > > /path -c "%f") at various points, and failing when the number of free > > blocks is outside an expected range. > > > > I don't know if it might be some caching issue, or something about how > > NFS reports free blocks. > > > > Honestly it looks unlikely to be critical, so for now I'm ignoring > > it.... > > > > XFS backend? Yes. > It could be speculative preallocation. The fact that NFS > can defer closing the file (either due to delegations or due to the > NFSv3 file cache) typically results in it taking longer for XFS to free > up the blocks it preallocated. That again means it takes longer for the > 'space used' to settle to the correct final value. > > https://linux-xfs.oss.sgi.narkive.com/jjjfnyI1/faq-xfs-speculative-preallocation That sounds plausible. --b.
