Gal Pressman

I ran the same command with mlx5 (UD instead of SRD), it works both with and without this patch, should be safe to merge.

@mrgolin @shuki-zanyovka can you please summarize what you're trying to fix here? You started with a custom provider that bypasses qp limits check and it got more and more complicated...

``` if (attr->cap.max_recv_wr > ctx->max_rq_wr) { verbs_err(&ctx->ibvctx, "Max receive WR %u > %u\n", attr->cap.max_recv_wr, ctx->max_rq_wr); return EINVAL; } ``` For RQs max_rq_wr is misleading.. The driver reports the max RQ...

> > max_rq_wr > > Does it mean we need to compare it to: roundup_pow_of_two(cap->max_recv_sge * cap->max_recv_wr) ? > > I.e. > > max_recv_descs = roundup_pow_of_two(attr->cap.max_recv_sge * attr->cap.max_recv_wr); if (max_recv_descs...

> > > > max_rq_wr > > > > > > > > > Does it mean we need to compare it to: roundup_pow_of_two(cap->max_recv_sge * cap->max_recv_wr) ? > > >...

I don't see how your suggested code solves anything. There are some inconsistencies in the efa_wq area, but I believe the root cause for all is that the size of...