[Devel] Re: [PATCH v2 1/4] SUNRPC: release per-net clients lock before calling PipeFS dentries creation

[Myklebust, Trond]

On Mon, 2012-02-27 at 20:55 +0400, Stanislav Kinsbursky wrote:
> 27.02.2012 20:21, Myklebust, Trond пишет:
> > On Mon, 2012-02-27 at 19:50 +0400, Stanislav Kinsbursky wrote:
> >> Lockdep is sad otherwise, because inode mutex is taken on PipeFS dentry
> >> creation, which can be called on mount notification, where this per-net client
> >> lock is taken on clients list walk.
> >>
> >> Signed-off-by: Stanislav Kinsbursky<skinsbursky at parallels.com>
> >>
> >> ---
> >>   net/sunrpc/clnt.c |   10 +++++++---
> >>   1 files changed, 7 insertions(+), 3 deletions(-)
> >>
> >> diff --git a/net/sunrpc/clnt.c b/net/sunrpc/clnt.c
> >> index bb7ed2f3..ddb5741 100644
> >> --- a/net/sunrpc/clnt.c
> >> +++ b/net/sunrpc/clnt.c
> >> @@ -84,7 +84,7 @@ static void rpc_register_client(struct rpc_clnt *clnt)
> >>   	struct sunrpc_net *sn = net_generic(clnt->cl_xprt->xprt_net, sunrpc_net_id);
> >>
> >>   	spin_lock(&sn->rpc_client_lock);
> >> -	list_add(&clnt->cl_clients,&sn->all_clients);
> >> +	list_add_tail(&clnt->cl_clients,&sn->all_clients);
> >>   	spin_unlock(&sn->rpc_client_lock);
> >>   }
> >>
> >> @@ -208,15 +208,19 @@ static int rpc_pipefs_event(struct notifier_block *nb, unsigned long event,
> >>   			    void *ptr)
> >>   {
> >>   	struct super_block *sb = ptr;
> >> -	struct rpc_clnt *clnt;
> >> +	struct rpc_clnt *clnt, *tmp;
> >>   	int error = 0;
> >>   	struct sunrpc_net *sn = net_generic(sb->s_fs_info, sunrpc_net_id);
> >>
> >>   	spin_lock(&sn->rpc_client_lock);
> >> -	list_for_each_entry(clnt,&sn->all_clients, cl_clients) {
> >> +	list_for_each_entry_safe(clnt, tmp,&sn->all_clients, cl_clients) {
> >> +		atomic_inc(&clnt->cl_count);
> >> +		spin_unlock(&sn->rpc_client_lock);
> >>   		error = __rpc_pipefs_event(clnt, event, sb);
> >> +		rpc_release_client(clnt);
> >>   		if (error)
> >>   			break;
> >> +		spin_lock(&sn->rpc_client_lock);
> >>   	}
> >>   	spin_unlock(&sn->rpc_client_lock);
> >>   	return error;
> >>
> >
> > This won't be safe. Nothing guarantees that 'tmp' remains valid after
> > you drop the spin_lock.
> >
> > I think you rather need to add a check for whether clnt->cl_dentry is in
> > the right state (NULL if RPC_PIPEFS_UMOUNT or non-NULL if
> > RPC_PIPEFS_MOUNT) before deciding whether or not to atomic_inc() and
> > drop the lock, so that you can restart the loop after calling
> > __rpc_pipefs_event().
> >
> 
> Gmmm.
> Please, correct me, if I'm wrong, that you are proposing something like this:
> 
>    	spin_lock(&sn->rpc_client_lock);
> again:
> 	list_for_each_entry(clnt,&sn->all_clients, cl_clients) {
> 		if ((event == RPC_PIPEFS_MOUNT) && clnt->cl_dentry) ||
> 		    (event == RPC_PIPEFS_UMOUNT) && !clnt->cl_dentry))
> 			continue;
> 		atomic_inc(&clnt->cl_count);
> 		spin_unlock(&sn->rpc_client_lock);
>     		error = __rpc_pipefs_event(clnt, event, sb);
> 		rpc_release_client(clnt);
> 		spin_lock(&sn->rpc_client_lock);
>     		if (error)
>     			break;
> 		goto again;
>     	}
>     	spin_unlock(&sn->rpc_client_lock);

Something along those lines, yes... Alternatively, you could do as David
proposes, and grab a reference to the next rpc_client before calling
rpc_release_client. Either works for me...

-- 
Trond Myklebust
Linux NFS client maintainer

NetApp
Trond.Myklebust at netapp.com
www.netapp.com