2.4 client - update d_cache when server reports ENOENT on an NFS remove
PATCH 1/1 2.4 client - update d_cache when server reports ENOENT on an NFS remove
Anuwat Phrukphicharn of HP discovered and patched this problem in RHEL-3, and asked that I push it upstream.
When the 2.4 NFS client does a REMOVE and gets an ENOENT back from the server it does not remove the dentry from the d_cache. This can make it inappropriately keep writing to an inode that has been renamed.
To reproduce, run this on an NFS server with /scratch exported:
#!/bin/sh
dir=/scratch/98201
mkdir -p $dir/recv
cat /dev/null > $dir/recv/x
while true; do
length=wc -l $dir/recv/x | cut -d' ' -f1 if [ ${length} -gt 1 ]; then
echo "problem occured!" date
exit 1
fi
mv $dir/x $dir/recv/x
sleep 1
done
...and then do this on the client:
mount server:/scratch /mnt/server
while true; do
rm -f /mnt/server/98201/x
date >> /mnt/server/98201/x
usleep 100done
The file "x" should never contain more than 1 line, but occasionally, the server will report an ENOENT back to the client (indicating that the server script has renamed the file before the rm could occur). After this, the client will keep appending to the same inode, even though the file has been renamed.
I've not seen this problem in 2.6 kernels, but I've not done any extensive testing there as of yet so I can't confirm whether it's still a problem there or not.
A patch to fix this follows. It just makes ENOENT a special case when handling errors in the nfs_safe_remove function, and lets the client update the dcache as if the remove had succeeded. The ENOENT is still reported back to userspace. This corrected the problem on my test rig and for the reporter as well:
fs/nfs/dir.c renamed from fs/nfs/dir.c.unlink-enoent +4 -1
@@ -1023,7 +1023,10 @@ static int nfs_safe_remove(struct dentryif (inode)NFS_CACHEINV (inode); error =NFS_PROTO (dir)->remove (dir, &dentry->d_name); -if (error <0 ) + +/* if server returned ENOENT, assume that the dentry is already gone +* and update the cache accordingly */ +if (error <0 && (error != -ENOENT )) goto out;if (inode) inode->i_nlink--;
Quoted 58 lines · Jeff Layton
Anuwat Phrukphicharn of HP discovered and patched this problem in RHEL-3, and asked that I push it upstream.
When the 2.4 NFS client does a REMOVE and gets an ENOENT back from the server it does not remove the dentry from the
d_cache. This can make it inappropriately keep writing to an inode that has been renamed.To reproduce, run this on an NFS server with /scratch exported:
#!/bin/sh
dir=/scratch/98201
mkdir -p $dir/recv
cat/dev/null> $dir/recv/x
while true; do
length=wc -l $dir/recv/x | cut -d' ' -f1if [ ${length} -gt 1 ]; then
echo "problem occured!" date
exit 1
fi
mv $dir/x $dir/recv/x
sleep 1
done...and then do this on the client:
mount server:/scratch
/mnt/server
while true; dorm -f /mnt/server/98201/x date >> /mnt/server/98201/x usleep 100done
The file "x" should never contain more than 1 line, but occasionally, the server will report an ENOENT back to the client (indicating that the server script has renamed the file before the rm could occur). After this, the client will keep appending to the same inode, even though the file has been renamed.
I've not seen this problem in 2.6 kernels, but I've not done any extensive testing there as of yet so I can't confirm whether it's still a problem there or not.
A patch to fix this follows. It just makes ENOENT a special case when handling errors in the
nfs_safe_removefunction, and lets the client update the dcache as if the remove had succeeded. The ENOENT is still reported back to userspace. This corrected the problem on my test rig and for the reporter as well:Signed-off-by Jeff Layton <[email protected]>fs/nfs/dir.c renamed from fs/nfs/dir.c.unlink-enoent +0
-0@@ -1023,7 +1023,10 @@ static int nfs_safe_remove(struct dentryif (inode)
fs/nfs/dir.c static int nfs_safe_remove(struct dentry · line 1032
Quoted 10 lines
fs/nfs/dir.c renamed from fs/nfs/dir.c.unlink-enoent +4
-1NFS_CACHEINV (inode); error =NFS_PROTO (dir)->remove (dir, &dentry->d_name); -if (error <0 ) + +/* if server returned ENOENT, assume that the dentry is already gone +* and update the cache accordingly */ +if (error <0 && (error != -ENOENT )) goto out;if (inode) inode->i_nlink--;
Thanks very much for your patch Jeff, I'm queueing it for 2.4.34.
Cheers,
Willy