Re: [for-next PATCH v2 1/2] RDMA/rxe: Remove unnecessary mr testing

[Bob Pearson <rpearsonhpe@gmail.com>]

On 10/23/22 21:25, Zhu Yanjun wrote:
> On Mon, Oct 24, 2022 at 2:05 AM Bob Pearson <rpearsonhpe@gmail.com> wrote:
>>
>> On 10/21/22 20:09, Li Zhijian wrote:
>>>
>>>
>>> On 21/10/2022 22:39, Zhu Yanjun wrote:
>>>> On Fri, Oct 21, 2022 at 3:53 PM Li Zhijian <lizhijian@fujitsu.com> wrote:
>>>>> Before the testing, we already passed it to rxe_mr_copy() where mr could
>>>>> be dereferenced. so this checking is not exactly correct.
>>>>>
>>>>> I tried to figure out the details how/when mr could be NULL, but failed
>>>>> at last. Add a WARN_ON(!mr) to that path to tell us more when it
>>>>> happends.
>>>> If I get you correctly, you confronted a problem,
>>> Not exactly,  I removed the mr checking since i think this checking is not correct.
>>> the newly added WARN_ON(!mr) is the only once place where the mr can be NULL but not handled correctly.
>>> At least with/without this patch, once WARN_ON(!mr) is triggered, kernel will go something wrong.
>>>
>>> so i want to place this  WARN_ON(!mr) to point to the problem.
>>>
>>> Thanks
>>> Zhijian
>>>
>>>>   but you can not figure it out.
>>>> So you send it upstream as a patch?
>>>>
>>>> I am not sure if it is a good idea.
>>>>
>>>> Zhu Yanjun
>>>>
>>>>> Signed-off-by: Li Zhijian <lizhijian@fujitsu.com>
>>>>> ---
>>>>>   drivers/infiniband/sw/rxe/rxe_resp.c | 4 ++--
>>>>>   1 file changed, 2 insertions(+), 2 deletions(-)
>>>>>
>>>>> diff --git a/drivers/infiniband/sw/rxe/rxe_resp.c b/drivers/infiniband/sw/rxe/rxe_resp.c
>>>>> index ed5a09e86417..218c14fb07c6 100644
>>>>> --- a/drivers/infiniband/sw/rxe/rxe_resp.c
>>>>> +++ b/drivers/infiniband/sw/rxe/rxe_resp.c
>>>>> @@ -778,6 +778,7 @@ static enum resp_states read_reply(struct rxe_qp *qp,
>>>>>          if (res->state == rdatm_res_state_new) {
>>>>>                  if (!res->replay) {
>>>>>                          mr = qp->resp.mr;
>>>>> +                       WARN_ON(!mr);
>>>>>                          qp->resp.mr = NULL;
>>>>>                  } else {
>>>>>                          mr = rxe_recheck_mr(qp, res->read.rkey);
>>>>> @@ -811,8 +812,7 @@ static enum resp_states read_reply(struct rxe_qp *qp,
>>>>>
>>>>>          rxe_mr_copy(mr, res->read.va, payload_addr(&ack_pkt),
>>>>>                      payload, RXE_FROM_MR_OBJ);
>>>>> -       if (mr)
>>>>> -               rxe_put(mr);
>>>>> +       rxe_put(mr);
>>>>>
>>>>>          if (bth_pad(&ack_pkt)) {
>>>>>                  u8 *pad = payload_addr(&ack_pkt) + payload
>>>>> --
>>>>> 2.31.1
>>>>>
>>>
>>
>> Li is correct that the only way mr could be NULL is if qp->resp.mr == NULL. So the
> 
> What I am concerned about is if "WARN_ON(!mr);" should be added or not.
> IMO, if the root cause remains unclear, this should be a problem.
> Currently this problem is not fixed. It is useless to send a debug
> statement to the maillist.

Li was fixing a bug that no one ever saw. mr is not NULL in this case.

Bob
> 
> Zhu Yanjun
> 
>> 'if (mr)' is not needed if that is the case. The read_reply subroutine is reached
>> from a new rdma read operation after going through check_rkey or from a previous
>> rdma read operations from get_req if qp->resp.res != NULL or from a duplicate request
>> where the previous responder resource is found. In all these cases the mr is set.
>> Initially in check_rkey where if it can't find the mr it causes an RKEY_VIOLATION.
>> Thereafter the rkey is stored in the responder resources and looked up for each
>> packet to get an mr or cause an RKEY_VIOLATION. So the mr can't be NULL. I think
>> you can leave out the WARN and just drop the if (mr).
>>
>> Bob
>>