[ofw][mlx4] SRQ attached QP issue

[Alex Estrin]

Hi,

Mthca doesn't have that problem.

Thanks,
Alex.

> -----Original Message-----
> From: Anatoly Greenblatt [mailto:anatolyg at voltaire.com] 
> Sent: Saturday, September 27, 2008 3:40 AM
> To: Alex Estrin; Leonid Keller
> Cc: ofw at lists.openfabrics.org
> Subject: RE: [ofw][mlx4] SRQ attached QP issue
> 
> Hi,
>  
> What about mthca? Is this thing allowed in non connectx hcas?
>  
> Thanks,
> Anatoly.
> 
> ________________________________
> 
> From: ofw-bounces at lists.openfabrics.org on behalf of Alex Estrin
> Sent: Sat 9/27/2008 1:20 AM
> To: Leonid Keller
> Cc: ofw at lists.openfabrics.org
> Subject: [ofw][mlx4] SRQ attached QP issue
> 
> 
> 
> Hello,
> 
> This patch seem fixed the create-QP-attached-to-SRQ issue.
> I didn't verify though, if code allowes send queue zero size. Included
> check anyway.
> Please review.
> 
> Thanks,
> Alex.
> 
> Index: mlx4/kernel/bus/ib/qp.c
> ===================================================================
> --- mlx4/kernel/bus/ib/qp.c     (revision 1622)
> +++ mlx4/kernel/bus/ib/qp.c     (working copy)
> @@ -411,14 +411,20 @@
>                 err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, 
> &qp->buf);
>                 if (err)
>                         goto err_mtt;
> -
> -               qp->sq.wrid  = kmalloc(qp->sq.wqe_cnt * sizeof (u64),
> GFP_KERNEL);
> -               qp->rq.wrid  = kmalloc(qp->rq.wqe_cnt * sizeof (u64),
> GFP_KERNEL);
> -
> -               if (!qp->sq.wrid || !qp->rq.wrid) {
> -                       err = -ENOMEM;
> -                       goto err_wrid;
> +               if (qp->sq.wqe_cnt) {
> +                       qp->sq.wrid  = kmalloc(qp->sq.wqe_cnt * sizeof
> (u64), GFP_KERNEL);
> +                       if (!qp->sq.wrid) {
> +                               err = -ENOMEM;
> +                               goto err_wrid;
> +                       }
>                 }
> +               if (qp->rq.wqe_cnt) {
> +                       qp->rq.wrid  = kmalloc(qp->rq.wqe_cnt * sizeof
> (u64), GFP_KERNEL);
> +                       if (!qp->rq.wrid) {
> +                               err = -ENOMEM;
> +                               goto err_wrid;
> +                       }
> +               }
>         }
> 
>         if (!sqpn)
> @@ -452,8 +458,10 @@
>                         
> mlx4_ib_db_unmap_user(to_mucontext(pd->p_uctx),
>                                               &qp->db);
>         } else {
> -               kfree(qp->sq.wrid);
> -               kfree(qp->rq.wrid);
> +               if (qp->sq.wrid)
> +                       kfree(qp->sq.wrid);
> +               if (qp->rq.wrid)
> +                       kfree(qp->rq.wrid);
>         }
> 
>  err_mtt:
> 
> 
> 
> > Hello,
> >
> > It seem there is a controversy exist in mlx4 driver that
> > won't allow to create QP if it was attached to SRQ.
> >
> > At first it doesn't allow to set any recv queue size (which is fine)
> >
> > mlx4\kernel\bus\ib\qp.c@ line 219:
> >               if (cap->max_recv_wr)
> >                       return -EINVAL;
> >
> >               qp->rq.wqe_cnt = qp->rq.max_gs = 0;
> >
> > But then there is assertion for zero size memory request in 
> 'kmalloc':
> > mlx4\kernel\bus\ib\qp.c @ line 416:
> >
> >               qp->rq.wrid  = kmalloc(qp->rq.wqe_cnt * sizeof
> > (u64), GFP_KERNEL);
> >              
> > mlx4\kernel\inc\l2w_memory.h @ line 81:
> >
> >               static inline void * kmalloc( SIZE_T bsize,
> > gfp_t gfp_mask)
> >               {
> >                       void *ptr;
> >                       ASSERT( KeGetCurrentIrql() <= DISPATCH_LEVEL);
> >                       ASSERT(bsize);
> >               .....
> >
> > I believe it was added for a reason.
> > Please let me know if I missed anything.
> >
> > Thanks,
> > Alex.
> >
> >
> >
> 
> 
>