Re: [PATCH v2 net-next 19/21] net/mlx5e: NVMEoTCP, data-path for DDP offload

From: David Ahern <dsahern@gmail.com>
To: Boris Pismenny <borispismenny@gmail.com>,
	Boris Pismenny <borisp@mellanox.com>,
	kuba@kernel.org, davem@davemloft.net, saeedm@nvidia.com,
	hch@lst.de, sagi@grimberg.me, axboe@fb.com, kbusch@kernel.org,
	viro@zeniv.linux.org.uk, edumazet@google.com, smalin@marvell.com
Cc: boris.pismenny@gmail.com, linux-nvme@lists.infradead.org,
	netdev@vger.kernel.org, benishay@nvidia.com, ogerlitz@nvidia.com,
	yorayz@nvidia.com, Or Gerlitz <ogerlitz@mellanox.com>,
	Yoray Zack <yorayz@mellanox.com>
Subject: Re: [PATCH v2 net-next 19/21] net/mlx5e: NVMEoTCP, data-path for DDP offload
Date: Mon, 18 Jan 2021 21:36:30 -0700	[thread overview]
Message-ID: <2a0bfce0-6226-7b9a-95b1-15f4f1f321e8@gmail.com> (raw)
In-Reply-To: <15248743-82bf-4283-d8c6-99f2210e42ae@gmail.com>

On 1/17/21 1:42 AM, Boris Pismenny wrote:
> This is needed for a few reasons that are explained in detail
> in the tcp-ddp offload documentation. See patch 21 overview
> and rx-data-path sections. Our reasons are as follows:

I read the documentation patch, and it does not explain it and really
should not since this is very mlx specific based on the changes.
Different h/w will have different limitations. Given that, it would be
best to enhance the patch description to explain why these gymnastics
are needed for the skb.

> 1) Each SKB may contain multiple PDUs. DDP offload doesn't operate on
> PDU headers, so these are written in the receive ring. Therefore, we
> need to rebuild the SKB to account for it. Additionally, due to HW
> limitations, we will only offload the first PDU in the SKB.

Are you referring to LRO skbs here? I can't imagine going through this
for 1500 byte packets that have multiple PDUs.

> 2) The newly constructed SKB represents the original data as it is on
> the wire, such that the network stack is oblivious to the offload.
> 3) We decided not to modify all of the mlx5e_skb_from_cqe* functions
> because it would make the offload harder to distinguish, and it would
> add overhead to the existing data-path fucntions. Therefore, we opted
> for this modular approach.
> 
> If we only had generic header-data split, then we just couldn't
> provide this offload. It is not enough to place payload into some
> buffer without TCP headers because RPC protocols and advanced storage
> protocols, such as nvme-tcp, reorder their responses and require data
> to be placed into application/pagecache buffers, which are anything
> but anonymous. In other words, header-data split alone writes data
> to the wrong buffers (reordering), or to anonymous buffers that
> can't be page-flipped to replace application/pagecache buffers.
> 