On Mon, 2025-04-28 at 13:24 -0700, Jeff Layton wrote:
> Sending this as an RFC as I don't have a reliable reproducer for the
> problem that Omar reported. I'm also not sure this is the best fix for
> the problem. There is probably a case to be made that the real bug is in
> the error handling for pnfs_layoutreturn_before_put_layout_hdr().
>
> My guess is that the issue is that we end up with entries on the
> plh_return_segs list just before the network goes down. That causes the
> LAYOUTRETURN to fail with something that looks retryable, and the lsegs
> on the list aren't freed.
>
> It's possible that we just need to catch ENETUNREACH in the LAYOUTRETURN
> error handling, but I'm not sure I correctly understand the problem. If
> entries are racing onto the list just before the refcount decrement,
> then that wouldn't fix it.
>
> The first patch should fix the issue of the leaked lsegs, and the second
> should let us know if it ever crops up again.
>
> Signed-off-by: Jeff Layton <jlayton@kernel.org>
> ---
> Jeff Layton (2):
> nfs: free leftover lsegs before freeing a layout in pnfs_put_layout_hdr
> nfs: pr_warn if plh_segs or plh_return_segs are non-empty when freeing
>
> fs/nfs/pnfs.c | 23 +++++++++++++++++++++--
> 1 file changed, 21 insertions(+), 2 deletions(-)
> ---
> base-commit: 5bc1018675ec28a8a60d83b378d8c3991faa5a27
> change-id: 20250428-nfs-6-16-87062aa2989d
>
> Best regards,
Trond, Anna, ping? This seems like the right thing to do, but I'd
appreciate a second (and third) set of eyes on this.
Thanks,
--
Jeff Layton <jlayton@kernel.org>