> On Aug 19, 2016, at 10:50 AM, J. Bruce Fields <bfields@xxxxxxxxxx> wrote:
>
> On Thu, Aug 18, 2016 at 06:11:43PM -0400, Chuck Lever wrote:
>>
>>> On Aug 18, 2016, at 5:56 PM, J. Bruce Fields <bfields@xxxxxxxxxxxx> wrote:
>>>
>>> On Wed, Aug 03, 2016 at 03:40:11PM -0400, Chuck Lever wrote:
>>>>
>>>>> On Aug 3, 2016, at 1:47 PM, bfields@xxxxxxxxxxxx wrote:
>>>>>
>>>>> On Wed, Aug 03, 2016 at 11:27:47AM -0400, Chuck Lever wrote:
>>>>>> Hi Bruce-
>>>>>>
>>>>>> I see that commit 39a9beab5acb83176e8b9a4f0778749a09341f1f
>>>>>> Author: J. Bruce Fields <bfields@xxxxxxxxxx>
>>>>>> AuthorDate: Tue May 17 12:38:21 2016 -0400
>>>>>>
>>>>>> rpc: share one xps between all backchannels
>>>>>>
>>>>>> has added this piece of code:
>>>>>>
>>>>>> @@ -452,10 +452,20 @@ static struct rpc_clnt *rpc_create_xprt(struct rpc_create_args *args,
>>>>>> struct rpc_clnt *clnt = NULL;
>>>>>> struct rpc_xprt_switch *xps;
>>>>>>
>>>>>> - xps = xprt_switch_alloc(xprt, GFP_KERNEL);
>>>>>> - if (xps == NULL) {
>>>>>> - xprt_put(xprt);
>>>>>> - return ERR_PTR(-ENOMEM);
>>>>>> + if (args->bc_xprt && args->bc_xprt->xpt_bc_xps) {
>>>>>> + WARN_ON(args->protocol != XPRT_TRANSPORT_BC_TCP);
>>>>>> + xps = args->bc_xprt->xpt_bc_xps;
>>>>>> + xprt_switch_get(xps);
>>>>>> + } else {
>>>>>>
>>>>>>
>>>>>> the WARN_ON here fires on the server whenever I use NFSv4.1 on RDMA.
>>>>>>
>>>>>> Can you say why it was added? Is there something RPC/RDMA needs to
>>>>>> do to make the code safe?
>>>>>
>>>>> What is args->protocol in this case?
>>>>>
>>>>> Digging around... OK, I missed that BC_TCP and BC_RDMA were defined as
>>>>> OR's of an XPRT_TRANSPORT_BC bit with the identifier of the underlying
>>>>> transport. That makes sense.
>>>>>
>>>>> So, I should have just used XPRT_TRANSPORT_BC there--I think all I meant
>>>>> was "is this a backchannel".
>>>>>
>>>>> Does that fix the problem?
>>>>
>>>> This simple fix eliminates the log noise:
>>>>
>>>> diff --git a/net/sunrpc/clnt.c b/net/sunrpc/clnt.c
>>>> index 2808d55..f94caf7 100644
>>>> --- a/net/sunrpc/clnt.c
>>>> +++ b/net/sunrpc/clnt.c
>>>> @@ -520,7 +520,7 @@ struct rpc_clnt *rpc_create(struct rpc_create_args *args)
>>>> char servername[48];
>>>>
>>>> if (args->bc_xprt) {
>>>> - WARN_ON(args->protocol != XPRT_TRANSPORT_BC_TCP);
>>>> + WARN_ON(!(args->protocol & XPRT_TRANSPORT_BC));
>>>> xprt = args->bc_xprt->xpt_bc_xprt;
>>>> if (xprt) {
>>>> xprt_get(xprt);
>>>>
>>>>
>>>> This code seems to come from:
>>>>
>>>> commit d50039ea5ee63c589b0434baa5ecf6e5075bb6f9
>>>> Author: J. Bruce Fields <bfields@xxxxxxxxxx>
>>>> AuthorDate: Mon May 16 17:03:42 2016 -0400
>>>>
>>>> nfsd4/rpc: move backchannel create logic into rpc code
>>>>
>>>>
>>>> Where it may have been copied from:
>>>>
>>>> -static struct rpc_clnt *create_backchannel_client(struct rpc_create_args *args)
>>>> -{
>>>> - struct rpc_xprt *xprt;
>>>> -
>>>> - if (args->protocol != XPRT_TRANSPORT_BC_TCP)
>>>> - return rpc_create(args);
>>>> -
>>>> - xprt = args->bc_xprt->xpt_bc_xprt;
>>>> - if (xprt) {
>>>> - xprt_get(xprt);
>>>> - return rpc_create_xprt(args, xprt);
>>>> - }
>>>> -
>>>> - return rpc_create(args);
>>>> -}
>>>>
>>>> There's no warning here. In fact, protocol != BC_TCP seems to
>>>> be expected.
>>>
>>> The protocol should be BC_TCP (OK, actually just BC) if and only if
>>> bc_xprt is set.
>>>
>>> (The BC_TCP case is the 4.1+ case, the other is the 4.0 case. In the
>>> 4.1+ case, the new client uses an existing (client-initiated)
>>> connection, in the 4.0 case, the new client must also have a new
>>> connection.
>>>
>>> In the 4.0 case we'll always create a new xprt, in the 4.1 case we might
>>> or might not--depends on whether that particular connection has been
>>> used for a backchannel previously.)
>>
>> OK, but why is a WARN_ON needed here? Why not return -EINVAL,
>> for example (once you've corrected BC_TCP -> BC) ?
>
> Well, it would be a programming bug, so I'd want a WARN_ON or similar
> somewhere, I don't care particularly where it is if you see a better way
> to organize things.
The way it works now, the WARN_ON fires, but the logic goes ahead
and creates the transport anyway.
If this is a programming bug, it should fail and return an error,
no transport should be created. I can see a WARN_ON being useful
because it displays a backtrace which identifies the broken
caller.
If it is not a programming bug (which is implied by the fact that
a transport is created anyway) then no WARN_ON is needed.
If you think it is correct that a WARN_ON fires _and_ a transport
is created, could a comment be added explaining that? The new
logic seems less straightforward to me than what it replaces.
--
Chuck Lever
--
To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
