Re: [PATCH for-next] RDMA/rxe fixed bug in rxe_requester

[Bob Pearson <rpearsonhpe@gmail.com>]

On 10/29/20 4:27 AM, Leon Romanovsky wrote:
> On Tue, Oct 13, 2020 at 12:07:42PM -0500, Bob Pearson wrote:
>> The code which limited the number of unacknowledged PSNs was incorrect.
>> The PSNs are limited to 24 bits and wrap back to zero from 0x00ffffff.
>> The test was computing a 32 bit value which wraps at 32 bits so that
>> qp->req.psn can appear smaller than the limit when it is actually larger.
>>
>> Replace '>' test with psn_compare which is used for other PSN comparisons
>> and correctly handles the 24 bit size.
>>
>> Fixes: 8700e3e7c485 ("Soft RoCE (RXE) - The software RoCE driver")
>> Signed-off-by: Bob Pearson <rpearson@hpe.com>
>> ---
>>  drivers/infiniband/sw/rxe/rxe_req.c | 3 ++-
>>  1 file changed, 2 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/infiniband/sw/rxe/rxe_req.c b/drivers/infiniband/sw/rxe/rxe_req.c
>> index af3923bf0a36..d4917646641a 100644
>> --- a/drivers/infiniband/sw/rxe/rxe_req.c
>> +++ b/drivers/infiniband/sw/rxe/rxe_req.c
>> @@ -634,7 +634,8 @@ int rxe_requester(void *arg)
>>  	}
>>
>>  	if (unlikely(qp_type(qp) == IB_QPT_RC &&
>> -		     qp->req.psn > (qp->comp.psn + RXE_MAX_UNACKED_PSNS))) {
>> +		psn_compare(qp->req.psn, (qp->comp.psn +
>> +				RXE_MAX_UNACKED_PSNS)) > 0)) {
> 
> qp->comp.psn is u32, so you are checking that
> qp->comp.psn + RXE_MAX_UNACKED_PSNS != 0, am I right?
> 
>>  		qp->req.wait_psn = 1;
>>  		goto exit;
>>  	}
>> --
>> 2.25.1

First, qp->comp.psn is a 24 bit unsigned quantity as is qp->req.psn.

RXE_MAX_UNACKED_PSNS is a reasonably small number e.g. 128 for now.

So qp->comp.psn + RXE_MAX_UNACKED_PSNS which is a 32 bit number never wraps to zero and remains in the
range [RXE_MAX_UNACKED_PSNS, RXE_MAX_UNACKED_PSNS + 2^24 -1]. The upper limit will not wrap back zero unless
RXE_MAX_UNACKED_PSNS is > 2^32 - 2^24 which would be a grossly unreasonable upper limit. You would have long
since run out of memory.

psn_compare(a, b) = (a - b) << 8 and is a signed 32 bit number.

This correctly determines the magnitude and sign of the difference between a and b as long as that difference
is less than 2^23.

Bob