Re: [PATCH net-next v3 1/2] net/smc: make wr buffer count configurable

[Dust Li <dust.li@linux.alibaba.com>]

On 2025-09-28 10:02:43, Dust Li wrote:
>On 2025-09-28 00:55:15, Halil Pasic wrote:
>>On Thu, 25 Sep 2025 13:25:40 +0200
>>Halil Pasic <pasic@linux.ibm.com> wrote:
>>
>>> > [...]  
>>> > > @@ -683,6 +678,8 @@ int smc_ib_create_queue_pair(struct smc_link *lnk)
>>> > >  	};
>>> > >  	int rc;
>>> > >  
>>> > > +	qp_attr.cap.max_send_wr = 3 * lnk->lgr->max_send_wr;
>>> > > +	qp_attr.cap.max_recv_wr = lnk->lgr->max_recv_wr;    
>>> > 
>>> > Possibly:
>>> > 
>>> > 	cap = max(3 * lnk->lgr->max_send_wr, lnk->lgr->max_recv_wr);
>>> > 	qp_attr.cap.max_send_wr = cap;
>>> > 	qp_attr.cap.max_recv_wr = cap
>>> > 
>>> > to avoid assumption on `max_send_wr`, `max_recv_wr` relative values.  
>>> 
>>> Can you explain a little more. I'm happy to do the change, but I would
>>> prefer to understand why is keeping qp_attr.cap.max_send_wr ==
>>> qp_attr.cap.max_recv_wr better? But if you tell: "Just trust me!" I will.
>>
>>Due to a little accident we ended up having a private conversation
>>on this, which I'm going to sum up quickly.
>>
>>Paolo stated that he has no strong preference and that I should at
>>least add a comment, which I will do for v4. 
>>
>>Unfortunately I don't quite understand why qp_attr.cap.max_send_wr is 3
>>times the number of send WR buffers we allocate. My understanding
>>is that qp_attr.cap.max_send_wr is about the number of send WQEs.
>
>We have at most 2 RDMA Write for 1 RDMA send. So 3 times is necessary.
>That is explained in the original comments. Maybe it's better to keep it.
>
>```
>.cap = {
>                /* include unsolicited rdma_writes as well,
>                 * there are max. 2 RDMA_WRITE per 1 WR_SEND
>                 */
>        .max_send_wr = SMC_WR_BUF_CNT * 3,
>        .max_recv_wr = SMC_WR_BUF_CNT * 3,
>        .max_send_sge = SMC_IB_MAX_SEND_SGE,
>        .max_recv_sge = lnk->wr_rx_sge_cnt,
>        .max_inline_data = 0,
>},
>```
>
>>I assume that qp_attr.cap.max_send_wr == qp_attr.cap.max_recv_wr
>>is not something we would want to preserve.
>
>IIUC, RDMA Write won't consume any RX wqe on the receive side, so I think
>the .max_recv_wr can be SMC_WR_BUF_CNT if we don't use RDMA_WRITE_IMM.

I kept thinking about this a bit more, and I realized that max_recv_wr
should be larger than SMC_WR_BUF_CNT.

Since receive WQEs are posted in a softirq context, their posting may be
delayed. Meanwhile, the sender might already have received the TX
completion (CQE) and continue sending new messages. In this case, if the
receiver’s post_recv() (i.e., posting of RX WQEs) is delayed, an RNR
(Receiver Not Ready) can easily occur.

Best regards,
Dust