Hi Trond,
I'd like to fix up lookup revalidation for v4.1+ when the client is using
OPEN_CLAIM_FH. The fixes a while back for Stan Hu's case do not seem to
improve things for v4.1, and actually make the behavior a bit worse since we
no longer pass through nfs_lookup_verify_inode(), which would catch the
cases where nlink == 0.
Would you accept work to _always_ revalidate the dentry's parent for
CLAIM_FH? Alternatively, it seems that CLAIM_NULL would be preferable for
this case, though I don't know how the client would know when to decide
between them.
Here's a simple reproducer for convenience, I think we've already all agreed
that the behavior we want is for the second open by `cat` to reflect the
results of the move on the server, or at least eventually later opens would
revalidate the dentry:
#!/bin/bash
set -o xtrace
vers=4.1
exportfs -ua
exportfs -o rw,sec=sys,no_root_squash *:/exports
mkdir /mnt/localhost || true
rm -f /exports/file{1,2}
echo this is file 1 > /exports/file1
echo this is file 2 > /exports/file2
mount -t nfs -ov$vers,sec=sys localhost:/exports /mnt/localhost
tail -f /mnt/localhost/file1 &
sleep 1
# this is file 1
cat /mnt/localhost/file1
# overwrite the file on the server:
mv -f /exports/file2 /exports/file1
# this is file 2
cat /mnt/localhost/file1
killall tail
# this is file 2
cat /mnt/localhost/file1
umount /mnt/localhost
Switching the $vers variable between v4.0 and v4.1 in this script shows the
difference in behavior.
Thanks for any advice,
Ben
