Re: [PATCH rdma-next 4/4] RDMA/mlx5: Use query_special_contexts for mkeys

From: Leon Romanovsky <leon@kernel.org>
To: Saeed Mahameed <saeed@kernel.org>
Cc: Jason Gunthorpe <jgg@nvidia.com>,
	Or Har-Toov <ohartoov@nvidia.com>,
	"David S. Miller" <davem@davemloft.net>,
	Eric Dumazet <edumazet@google.com>,
	Jakub Kicinski <kuba@kernel.org>,
	linux-rdma@vger.kernel.org,
	Michael Guralnik <michaelgur@nvidia.com>,
	netdev@vger.kernel.org, Paolo Abeni <pabeni@redhat.com>,
	Saeed Mahameed <saeedm@nvidia.com>
Subject: Re: [PATCH rdma-next 4/4] RDMA/mlx5: Use query_special_contexts for mkeys
Date: Tue, 10 Jan 2023 10:45:49 +0200	[thread overview]
Message-ID: <Y70lvUCHeHhNnneq@unreal> (raw)
In-Reply-To: <Y7yUHJukwiIcL5BJ@x130>

On Mon, Jan 09, 2023 at 02:24:28PM -0800, Saeed Mahameed wrote:
> On 08 Jan 12:21, Leon Romanovsky wrote:
> > On Thu, Jan 05, 2023 at 05:04:03PM -0800, Saeed Mahameed wrote:
> > > On 04 Jan 16:03, Leon Romanovsky wrote:
> > > > On Wed, Jan 04, 2023 at 09:56:21AM -0400, Jason Gunthorpe wrote:
> > > > > On Wed, Jan 04, 2023 at 03:55:25PM +0200, Leon Romanovsky wrote:
> > > > > > On Wed, Jan 04, 2023 at 09:13:10AM -0400, Jason Gunthorpe wrote:
> > > > > > > On Wed, Jan 04, 2023 at 03:09:54PM +0200, Leon Romanovsky wrote:
> > > > > > > > On Wed, Jan 04, 2023 at 09:03:06AM -0400, Jason Gunthorpe wrote:
> > > > > > > > > On Wed, Jan 04, 2023 at 10:11:25AM +0200, Leon Romanovsky wrote:
> > > > > > > > > > -int mlx5_cmd_null_mkey(struct mlx5_core_dev *dev, u32 *null_mkey)
> > > > > > > > > > -{
> > > > > > > > > > -	u32 out[MLX5_ST_SZ_DW(query_special_contexts_out)] = {};
> > > > > > > > > > -	u32 in[MLX5_ST_SZ_DW(query_special_contexts_in)] = {};
> > > > > > > > > > -	int err;
> > > > > > > > > > +	err = mlx5_cmd_exec_inout(dev->mdev, query_special_contexts, in, out);
> > > > > > > > > > +	if (err)
> > > > > > > > > > +		return err;
> > > > > > > > > >
> > > > > > > > > > -	MLX5_SET(query_special_contexts_in, in, opcode,
> > > > > > > > > > -		 MLX5_CMD_OP_QUERY_SPECIAL_CONTEXTS);
> > > > > > > > > > -	err = mlx5_cmd_exec_inout(dev, query_special_contexts, in, out);
> > > > > > > > > > -	if (!err)
> > > > > > > > > > -		*null_mkey = MLX5_GET(query_special_contexts_out, out,
> > > > > > > > > > -				      null_mkey);
> > > > > > > > > > -	return err;
> > > > > > > > > > +	if (MLX5_CAP_GEN(dev->mdev, dump_fill_mkey))
> > > > > > > > > > +		dev->mkeys.dump_fill_mkey = MLX5_GET(query_special_contexts_out,
> > > > > > > > > > +						     out, dump_fill_mkey);
> > > > > > > > > > +
> > > > > > > > > > +	if (MLX5_CAP_GEN(dev->mdev, null_mkey))
> > > > > > > > > > +		dev->mkeys.null_mkey = cpu_to_be32(
> > > > > > > > > > +			MLX5_GET(query_special_contexts_out, out, null_mkey));
> > > > > > > > > > +
> > > > > > > > > > +	if (MLX5_CAP_GEN(dev->mdev, terminate_scatter_list_mkey)) {
> > > > > > > > > > +		dev->mkeys.terminate_scatter_list_mkey =
> > > > > > > > > > +			cpu_to_be32(MLX5_GET(query_special_contexts_out, out,
> > > > > > > > > > +					     terminate_scatter_list_mkey));
> > > > > > > > > > +		return 0;
> > > > > > > > > > +	}
> > > > > > > > > > +	dev->mkeys.terminate_scatter_list_mkey =
> > > > > > > > > > +		MLX5_TERMINATE_SCATTER_LIST_LKEY;
> > > > > > > > >
> > > > > > > > > This is already stored in the core dev, why are you recalculating it
> > > > > > > > > here?
> > > > > > > >
> > > > > > > > It is not recalculating but setting default value. In core dev, we will
> > > > > > > > have value only if MLX5_CAP_GEN(dev->mdev, terminate_scatter_list_mkey)
> > > > > > > > is true.
> > > > > > >
> > > > > > > No, it has the identical code:
> > > > > > >
> > > > > > > +static int mlx5_get_terminate_scatter_list_mkey(struct mlx5_core_dev *dev)
> > > > > > > +{
> > > > > > > +       if (MLX5_CAP_GEN(dev, terminate_scatter_list_mkey)) {
> > > > > > > +               dev->terminate_scatter_list_mkey =
> > > > > > > +                       cpu_to_be32(MLX5_GET(query_special_contexts_out, out,
> > > > > > > +                                            terminate_scatter_list_mkey));
> > > > > > > +               return 0;
> > > > > > > +       }
> > > > > > > +       dev->terminate_scatter_list_mkey = MLX5_TERMINATE_SCATTER_LIST_LKEY;
> > > > > >
> > > > > > Ahh, you are talking about that.
> > > > > > terminate_scatter_list_mkey is part of an output from MLX5_CMD_OP_QUERY_SPECIAL_CONTEXTS,
> > > > > > which is needed to get other mkeys. So instead of doing special logic
> > > > > > for the terminate_scatter_list_mkey, we decided to use same pattern as
> > > > > > for other mkeys, which don't belong to core.
> > > > >
> > > > > Regardless, don't duplicate the code and maybe don't even duplicate
> > > > > the storage of the terminate_scatter_list_mkey
> > > >
> > > > ok, will update and resend.
> > > >
> > > 
> > > Please provide a helper mlx5_get_xyz function, avoid assuming mlx5_core will store
> > > it in dev->xyz.
> > 
> > This helper was used in this version of patch and it is:
> > MLX5_GET(query_special_contexts_out, out, terminate_scatter_list_mkey))
> > Which was dropped in favor of Jason's request to rely on mlx5_core_dev
> > as this value already known:
> > https://lore.kernel.org/all/cover.1672917578.git.leonro@nvidia.com
> > 
> > I don't like the idea to add obfuscation function just for the sake of
> > obfuscation. If you don't want access to mlx5_core_dev, please argue with
> > Jason to accept first variant.
> 
> I think you are confused, 1st version duplicated the logic that already
> existed in mlx5 core, didn't just access with MLX5_GET.

No, the end result is the same: call to the memory to get terminate_scatter_list_mkey
value. MLX5_GET(..) will return exactly the same value as dev->terminate_scatter_list_mkey
or some other function call:

int mlx5_get_terminate_scatter_list_mkey() {
 return dev->terminate_scatter_list_mkey;
}

> 
> 