Re: [pnfs] [PATCH 2/2] nfs41: fix race on session reset

[Benny Halevy <bhalevy@panasas.com>]

On Oct. 22, 2009, 15:32 +0200, "William A. (Andy) Adamson" <androsadamson@gmail.com> wrote:
> On Thu, Oct 22, 2009 at 9:20 AM, Benny Halevy <bhalevy@panasas.com> wrote:
>> On Oct. 21, 2009, 20:24 +0200, andros@netapp.com wrote:
>>> From: Andy Adamson <andros@netapp.com>
>>>
>>> Do not clear the NFS4CLNT_SESSION_SETUP bit until after the session has been
>>> reset (a possible EXCHANGE_ID, a DESTROY_SESSION, and a CREATE_SESSION) to
>>> prevent a race with nfs41_setup_sequence assigning a slot on a
>>> partially reset session.
>>>
>>> Signed-off-by: Andy Adamson <andros@netapp.com>
>>> ---
>>>  fs/nfs/nfs4proc.c  |    3 +++
>>>  fs/nfs/nfs4state.c |   15 +++++++++------
>>>  2 files changed, 12 insertions(+), 6 deletions(-)
>>>
>>> diff --git a/fs/nfs/nfs4proc.c b/fs/nfs/nfs4proc.c
>>> index eb245a1..80764e2 100644
>>> --- a/fs/nfs/nfs4proc.c
>>> +++ b/fs/nfs/nfs4proc.c
>>> @@ -4788,6 +4788,9 @@ int nfs4_proc_create_session(struct nfs_client *clp, int reset)
>>>       if (status)
>>>               goto out;
>>>
>>> +     /* Signal nfs41_setup_sequence that the session is ready for use */
>>> +     clear_bit(NFS4CLNT_SESSION_SETUP, &clp->cl_state);
>>> +
>>>       ptr = (unsigned *)&session->sess_id.data[0];
>>>       dprintk("%s client>seqid %d sessionid %u:%u:%u:%u\n", __func__,
>>>               clp->cl_seqid, ptr[0], ptr[1], ptr[2], ptr[3]);
>>> diff --git a/fs/nfs/nfs4state.c b/fs/nfs/nfs4state.c
>>> index 1394dfb..09ca30b 100644
>>> --- a/fs/nfs/nfs4state.c
>>> +++ b/fs/nfs/nfs4state.c
>>> @@ -1252,17 +1252,20 @@ static void nfs4_state_manager(struct nfs_client *clp)
>>>                               continue;
>>>               }
>>>               /* Initialize or reset the session */
>>> -             if (test_and_clear_bit(NFS4CLNT_SESSION_SETUP, &clp->cl_state)
>>> -                && nfs4_has_session(clp)) {
>>> +             if (test_bit(NFS4CLNT_SESSION_SETUP, &clp->cl_state)
>>> +                 && nfs4_has_session(clp)) {
>>>                       if (clp->cl_cons_state == NFS_CS_SESSION_INITING)
>>>                               status = nfs4_initialize_session(clp);
>>>                       else
>>>                               status = nfs4_reset_session(clp);
>>> -                     if (status) {
>>> -                             if (status == -NFS4ERR_STALE_CLIENTID)
>>> -                                     continue;
>>> +                     if (status == -NFS4ERR_STALE_CLIENTID)
>>> +                             continue;
>>> +                     /* For error case. On success the bit is cleared in
>>> +                      * nfs4_proc_create_session */
>> Why the separation?
>> Why not handle both success and error cases here?
> 
> We don't clear the bit on NFS4ERR_STALE_CLIENTID because we first need
> an EXCHANGE_ID and then session reset and we don't want any rpc's
> queued on the slot_tbl_waitq to run. The check for
> NFS4ERR_STALE_CLIENTID is in the nfs4_state_manager. Also, all other
> cases in the state manager clear the bits used - so I thought it best
> to follow suit.

I agree, so why clear the NFS4CLNT_SESSION_SETUP bit on success
in nfs4_proc_create_session()
Also, nfs4_proc_create_session() clears NFS4CLNT_LEASE_EXPIRED
which is test_and_cleared by the state manager...

Benny

> 
> -->Andy
> 
>> Benny
>>
>>> +                     clear_bit(NFS4CLNT_SESSION_SETUP, &clp->cl_state);
>>> +
>>> +                     if (status)
>>>                               goto out_error;
>>> -                     }
>>>               }
>>>               /* First recover reboot state... */
>>>               if (test_and_clear_bit(NFS4CLNT_RECLAIM_REBOOT, &clp->cl_state)) {
>> _______________________________________________
>> pNFS mailing list
>> pNFS@linux-nfs.org
>> http://linux-nfs.org/cgi-bin/mailman/listinfo/pnfs
>>