On Thu, 2013-07-04 at 05:20 +0100, Al Viro wrote:
> On Wed, Jul 03, 2013 at 04:25:32PM -0400, Waiman Long wrote:
> > There is no change in logic and everything should just work.
>
> > - spin_lock(&file->f_path.dentry->d_lock);
> > + d_lock(file->f_path.dentry);
> > if (!d_unhashed(file->f_path.dentry))
> > clnt = RPC_I(inode)->private;
> > if (clnt != NULL && atomic_inc_not_zero(&clnt->cl_count)) {
> > - spin_unlock(&file->f_path.dentry->d_lock);
> > + d_unlock(file->f_path.dentry);
>
> Could somebody explain WTF is being protected here? It's not ->private -
> that gets set (and, more importantly, cleared) without ->d_lock in sight.
> Trond, that seems to be your code from about three years ago (introduced
> in "SUNRPC: Fix a race in rpc_info_open"). What's going on there?
AFAICR we're using the fact that the dentry will remain hashed until
we're in the process of freeing up the rpc_client. By testing that the
dentry is hashed under the dentry->d_lock, we are assured that the
non-NULL 'clnt' is still pointing to a valid rpc_client, and that it is
safe to access clnt->cl_count.
--
Trond Myklebust
Linux NFS client maintainer
NetApp
trond.mykleb...@netapp.com
www.netapp.com
