[ofa-general] [RFC] XRC -- make receiving XRC QP independent of any one user process

[Tang, Changqing]

You mentioned you need a kernel thread to scan all the XRC domain to cleanup the QP, is there any existing timeout event or other event to trigger such a scan ?


--CQ

> -----Original Message-----
> From: Jack Morgenstein [mailto:jackm at dev.mellanox.co.il]
> Sent: Saturday, December 22, 2007 2:05 AM
> To: Tang, Changqing
> Cc: pasha at dev.mellanox.co.il; general at lists.openfabrics.org
> Subject: Re: [ofa-general] [RFC] XRC -- make receiving XRC QP
> independent of any one user process
>
> On Saturday 22 December 2007 04:51, Tang, Changqing wrote:
> > We need to ask Roland to confirm this.
> >
> I'll speak to the firmware guys here for clarification.
> >
> >
> > I did not use zero byte rdma_read, I only use zero byte rdma_write.
> >
> > Here is our code:
> >
> >                 sr.next = NULL;
> >                 sr.wr_id = (uint64_t)(AULONG)rdmahdr;
> >
> >                 sr.sg_list = &ssg;
> >                 sr.num_sge = 0;
> >                 sr.opcode = IBV_WR_RDMA_WRITE;
> >                 sr.send_flags = IBV_SEND_INLINE|IBV_SEND_SIGNALED;
> >
> >                 err =
> ibv_post_send(ibvproc->connection[i].qp_hndl, &sr, &bad_sr);
> >                 if (err != 0) {
> >                         hpmp_printf("ibv_post_send() failed");
> >                         return (-1);
> >                 }
> >
> > Note, ssg is not initialized (Maybe we can set sr.sg_list = NULL ?)
> You don't need to initialize ssg -- you have set num_sge to
> zero, so sg_list is not relevant.
> >
> I notice that the rdma fields are also not initialized.
> That implies that no validity checking is done on the remote
> host side once the length is zero.
>
> I just looked at the ConnectX PRM, revision 0.35, section
> 8.4.1.11.  It specifically addresses the issue of sending
> 0-byte RDMA reads/writes: send a wqe with NO data segments,
> and you will get a 0-length RDMA-write.
> However, it does not mention whether the remote address and
> rkey provided in the WQE must be valid or not.
>
> I'll check with F/W regarding this issue, too.
>
> - Jack
>