On Tue, Nov 17, 2020 at 07:24:41PM +0200, Yishai Hadas wrote:
> > +void mlx5_query_device_ctx(struct mlx5_context *mctx)
> > +{
> > + struct ibv_device_attr_ex device_attr;
> > + struct mlx5_query_device_ex_resp resp;
> > + size_t resp_size = sizeof(resp);
> > +
> > + get_lag_caps(mctx);
> > +
> > + if (!(mctx->cmds_supp_uhw & MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE))
> > + return;
> > +
> Any reason to not read some applicable context fields (e.g. max_dm_size)
> over uverbs by reducing the resp_size to the ib part ?
No, I missed the device_attr detail
> > + if (ibv_cmd_query_device_any(&mctx->ibv_ctx.context, NULL, &device_attr,
> > + sizeof(device_attr), &resp.ibv_resp,
> > + &resp_size))
> > + return;
> > +
> > + mctx->cached_device_cap_flags = device_attr.orig_attr.device_cap_flags;
> > + mctx->atomic_cap = device_attr.orig_attr.atomic_cap;
> > + mctx->cached_tso_caps = device_attr.tso_caps;
>
> The device_attr.tso_caps is not set over uverbs / cmd_device.c, it
> comes as UHW. The below should be done instead.
Got it
I updated the github
Thanks,
Jason
