On Jul 22, 2009, at 4:20 PM, Trond Myklebust wrote:
> On Wed, 2009-07-22 at 15:49 -0400, Andy Adamson wrote:
>> On Jul 21, 2009, at 5:17 PM, Trond Myklebust wrote:
>>> Note that there is a bug remaining inside nfs4_init_session(): we
>>> shouldn't be copying the rsize/wsize into the nfs_client if the
>>> latter
>>> was already initialised.
>>
>> The rsize/wsize is copied into the session prior to the
>> create_session
>> call (triggered by the state management code you moved), and is used
>> for session negotiation. At this point the nfs_client cl_cons_state
>> is
>> set to NFS_CS_SESSION_INITING (see nfs4_alloc_session), so the
>> nfs_client is not initialized. The cl_cons_state is set to
>> NFS_CS_READY after a successful create_session call.
>
> The call to nfs4_init_session() is in nfs4_create_server(). It can be
> called several times _after_ the nfs_client has been initialised when
> you mount more than one partition from the same NFS server.
>
> If that is the case, and if you use different rsize/wsize values on
> those different mounts, then you will end up clobbering the values of
> fc_attrs.max_rqst_sz, and fc_attrs.max_resp_sz, having set them to the
> wsize/rsize that was set by the very last mount call.
You are right, this is a bug.
>
>
> AFAICS, what you _should_ be doing in nfs4_init_session, is something
> like
>
> if (clp->cl_session->fc_attrs.max_rqst_sz < server->wsize)
> clp->cl_session->fc_attrs.max_rqst_sz = server->wsize;
Currently, we only support one session per nfs_client - one session
for all partition mounts to the same NFS server.
For NFSv4.1 the per partition struct nfs_server rsize/wsize is used to
construct rpc requests that are sent over the session so the rsize/
wsize must not be larger than the session max_resp_sz/max_rqst_sz.
nfs4_init_session should simply return if the nfs_client cl_cons_state
is not NFS_CS_SESSION_INITING.
I shouldn't be trying to set the session max_resp_sz/max_rqst_sz to
the rsize/wsize, but rather to the maximum rsize/wsize supported by
the client.
If the server accepts or increases the max_resp_sz/max_rqst_sz then
all is well.
If the server reduces the max_resp_sz/max_rqst_sz, the maximum rsize/
wsize available for NFSv4.1 partition mounts to the server needs to be
reduced accordingly. So the nfs_server rsize/wsize needs to be bound
by the session max_resp_sz/max_rqst_sz as well as by the maximum
supported size.
If you think this is correct, I'll send a patch.
-->Andy
>
>
> Trond
>
On Wed, 2009-07-22 at 17:32 -0400, Andy Adamson wrote:
> nfs4_init_session should simply return if the nfs_client cl_cons_state
> is not NFS_CS_SESSION_INITING.
> I shouldn't be trying to set the session max_resp_sz/max_rqst_sz to
> the rsize/wsize, but rather to the maximum rsize/wsize supported by
> the client.
> If the server accepts or increases the max_resp_sz/max_rqst_sz then
> all is well.
> If the server reduces the max_resp_sz/max_rqst_sz, the maximum rsize/
> wsize available for NFSv4.1 partition mounts to the server needs to be
> reduced accordingly. So the nfs_server rsize/wsize needs to be bound
> by the session max_resp_sz/max_rqst_sz as well as by the maximum
> supported size.
Well.... The rsize/wsize is one thing, but how about acls? We simply
don't know how big they may become.
An alternative would be to let the server choose the
max_rqst_sz/max_resp_sz by always requesting the maximum allowed value.
Trond