Re: [PATCH] nfsd: use nfs client rpc callback program

[Benny Halevy <bhalevy@panasas.com>]

On Sep. 19, 2008, 14:51 -0500, Olga Kornievskaia <aglo@citi.umich.edu> wrote:
> 
> Benny Halevy wrote:
>> On Sep. 17, 2008, 18:10 -0500, "J. Bruce Fields" <bfields@fieldses.org> wrote:
>>   
>>> On Wed, Sep 17, 2008 at 02:43:44PM -0500, Benny Halevy wrote:
>>>     
>>>> From: Benny Halevy <bhalevy@panasas.com>
>>>>
>>>> since commit ff7d9756b501744540be65e172d27ee321d86103
>>>> "nfsd: use static memory for callback program and stats"
>>>> do_probe_callback uses a static callback program
>>>> (NFS4_CALLBACK) rather than the one set in clp->cl_callback.cb_prog
>>>> as passed in by the client in setclientid (4.0)
>>>> or create_session (4.1).
>>>>       
>>> Ugh, yes, sorry about that.  (I wonder why pynfs testing didn't catch
>>> this?  Oh, I guess it's because NFS4_CALLBACK is the program number our
>>> client always gives us.)
>>>     
>> Well, Fred (thanks!) added a test today which uses a non-default
>> callback program and he sees a corresponding callback coming back.
>>
>> (Note that this test is not absolutely generic as the server is
>> not required to probe the callback immediately, or at all, after
>> setclientid or create_session.)
>>
>>   
>>>> @@ -371,6 +356,8 @@ static int do_probe_callback(void *data)
>>>>  		.to_maxval	= (NFSD_LEASE_TIME/2) * HZ,
>>>>  		.to_exponential	= 1,
>>>>  	};
>>>> +	static struct rpc_stat	cb_stats;
>>>> +	struct rpc_program	cb_program;
>>>>  	struct rpc_create_args args = {
>>>>  		.protocol	= IPPROTO_TCP,
>>>>  		.address	= (struct sockaddr *)&addr,
>>>> @@ -394,6 +381,20 @@ static int do_probe_callback(void *data)
>>>>  	addr.sin_port = htons(cb->cb_port);
>>>>  	addr.sin_addr.s_addr = htonl(cb->cb_addr);
>>>>  
>>>> +	/* Initialize rpc_program */
>>>> +	memset(&cb_program, 0, sizeof(cb_program));
>>>> +	cb_program.name = "nfs4_cb";
>>>> +	cb_program.number = clp->cl_callback.cb_prog;
>>>> +	cb_program.nrvers = ARRAY_SIZE(nfs_cb_version);
>>>> +	cb_program.version = nfs_cb_version;
>>>> +	cb_program.stats = &cb_stats;
>>>> +	memset(&cb_stats, 0, sizeof(cb_stats));
>>>> +	cb_stats.program = &cb_program;
>>>>       
>>> You don't want a pointer to data on the stack here, do you?
>>>     
>> Hmm, you're right...
>> I went back and forth whether this should be allocated statically,
>> dynamically, or on the stack.  I was mislead by the fact we're doing
>> a sync rpc call, but indeed this needs to live until the nfs client
>> is destroyed.  I'm trying to fully understand what Olga saw
>> before coming up with a new proposal... maybe putting the cb_program
>> back into struct nfs4_callback and just make cb_stats static would
>> provide a solution of the problem Olga witnessed and keep everybody
>> happy.
>>   
> I'm trying really hard to remember what was the issue of not using the 
> structure and instead using static memory. From what I remember the 
> issue was that the memory (clp->cl_callback.cb_prog) was going away.
> 

Yeah... pls see my reply on this thread.  From reading the code it
seems like the root cause is that gss_auth is holding a reference
on the rpc_clnt (for which I haven't seen a kref taken and released
which looks quite worrisome) and in gss_destroying_context() it's
doing rpc_call_null(gss_auth->client ... RPC_TASK_ASYNC) and therfore it needs
the program and stats.  This probably being called via:

nfs_free_client -> put_rpccred -> gss_destroy_cred (via
cred->cr_ops->crdestroy) -> gss_destroying_context.

since gss_destroying_context issues async null rpc
control returns to nfs_free_client which frees up the nfs_client
(and with it it used to free the rpc_program and rpc_stat)

Benny