[PATCH v23 05/20] nvme-tcp: Add DDP offload control path

[Sagi Grimberg]

On 28/02/2024 14:57, Aurelien Aptel wrote:
> From: Boris Pismenny <borisp at nvidia.com>
>
> This commit introduces direct data placement offload to NVME
> TCP. There is a context per queue, which is established after the
> handshake using the sk_add/del NDOs.
>
> Additionally, a resynchronization routine is used to assist
> hardware recovery from TCP OOO, and continue the offload.
> Resynchronization operates as follows:
>
> 1. TCP OOO causes the NIC HW to stop the offload
>
> 2. NIC HW identifies a PDU header at some TCP sequence number,
> and asks NVMe-TCP to confirm it.
> This request is delivered from the NIC driver to NVMe-TCP by first
> finding the socket for the packet that triggered the request, and
> then finding the nvme_tcp_queue that is used by this routine.
> Finally, the request is recorded in the nvme_tcp_queue.
>
> 3. When NVMe-TCP observes the requested TCP sequence, it will compare
> it with the PDU header TCP sequence, and report the result to the
> NIC driver (resync), which will update the HW, and resume offload
> when all is successful.
>
> Some HW implementation such as ConnectX-7 assume linear CCID (0...N-1
> for queue of size N) where the linux nvme driver uses part of the 16
> bit CCID for generation counter. To address that, we use the existing
> quirk in the nvme layer when the HW driver advertises if the device is
> not supports the full 16 bit CCID range.
>
> Furthermore, we let the offloading driver advertise what is the max hw
> sectors/segments via ulp_ddp_limits.
>
> A follow-up patch introduces the data-path changes required for this
> offload.
>
> Socket operations need a netdev reference. This reference is
> dropped on NETDEV_GOING_DOWN events to allow the device to go down in
> a follow-up patch.
>
> Signed-off-by: Boris Pismenny <borisp at nvidia.com>
> Signed-off-by: Ben Ben-Ishay <benishay at nvidia.com>
> Signed-off-by: Or Gerlitz <ogerlitz at nvidia.com>
> Signed-off-by: Yoray Zack <yorayz at nvidia.com>
> Signed-off-by: Shai Malin <smalin at nvidia.com>
> Signed-off-by: Aurelien Aptel <aaptel at nvidia.com>
> Signed-off-by: Max Gurtovoy <mgurtovoy at nvidia.com>
> Reviewed-by: Chaitanya Kulkarni <kch at nvidia.com>
> Reviewed-by: Max Gurtovoy <mgurtovoy at nvidia.com>
> ---
>   drivers/nvme/host/tcp.c | 265 ++++++++++++++++++++++++++++++++++++++--
>   1 file changed, 252 insertions(+), 13 deletions(-)
>
> diff --git a/drivers/nvme/host/tcp.c b/drivers/nvme/host/tcp.c
> index a6d596e05602..86a9ad9f679b 100644
> --- a/drivers/nvme/host/tcp.c
> +++ b/drivers/nvme/host/tcp.c
> @@ -21,6 +21,10 @@
>   #include <net/busy_poll.h>
>   #include <trace/events/sock.h>
>   
> +#ifdef CONFIG_ULP_DDP
> +#include <net/ulp_ddp.h>
> +#endif
> +
>   #include "nvme.h"
>   #include "fabrics.h"
>   
> @@ -46,6 +50,16 @@ MODULE_PARM_DESC(tls_handshake_timeout,
>   		 "nvme TLS handshake timeout in seconds (default 10)");
>   #endif
>   
> +#ifdef CONFIG_ULP_DDP
> +/* NVMeTCP direct data placement and data digest offload will not
> + * happen if this parameter false (default), regardless of what the
> + * underlying netdev capabilities are.
> + */
> +static bool ddp_offload;
> +module_param(ddp_offload, bool, 0644);
> +MODULE_PARM_DESC(ddp_offload, "Enable or disable NVMeTCP direct data placement support");
> +#endif
> +
>   #ifdef CONFIG_DEBUG_LOCK_ALLOC
>   /* lockdep can detect a circular dependency of the form
>    *   sk_lock -> mmap_lock (page fault) -> fs locks -> sk_lock
> @@ -119,6 +133,7 @@ enum nvme_tcp_queue_flags {
>   	NVME_TCP_Q_ALLOCATED	= 0,
>   	NVME_TCP_Q_LIVE		= 1,
>   	NVME_TCP_Q_POLLING	= 2,
> +	NVME_TCP_Q_OFF_DDP	= 3,
>   };
>   
>   enum nvme_tcp_recv_state {
> @@ -146,6 +161,18 @@ struct nvme_tcp_queue {
>   	size_t			ddgst_remaining;
>   	unsigned int		nr_cqe;
>   
> +#ifdef CONFIG_ULP_DDP
> +	/*
> +	 * resync_tcp_seq is a speculative PDU header tcp seq number (with
> +	 * an additional flag in the lower 32 bits) that the HW send to
> +	 * the SW, for the SW to verify.
> +	 * - The 32 high bits store the seq number
> +	 * - The 32 low bits are used as a flag to know if a request
> +	 *   is pending (ULP_DDP_RESYNC_PENDING).
> +	 */
> +	atomic64_t		resync_tcp_seq;
> +#endif
> +
>   	/* send state */
>   	struct nvme_tcp_request *request;
>   
> @@ -186,6 +213,13 @@ struct nvme_tcp_ctrl {
>   	struct delayed_work	connect_work;
>   	struct nvme_tcp_request async_req;
>   	u32			io_queues[HCTX_MAX_TYPES];
> +
> +	struct net_device	*ddp_netdev;
> +	netdevice_tracker	netdev_tracker;
> +	u32			ddp_threshold;
> +#ifdef CONFIG_ULP_DDP
> +	struct ulp_ddp_limits	ddp_limits;
> +#endif
>   };
>   
>   static LIST_HEAD(nvme_tcp_ctrl_list);
> @@ -297,6 +331,171 @@ static inline size_t nvme_tcp_pdu_last_send(struct nvme_tcp_request *req,
>   	return nvme_tcp_pdu_data_left(req) <= len;
>   }
>   
> +#ifdef CONFIG_ULP_DDP
> +
> +static struct net_device *
> +nvme_tcp_get_ddp_netdev_with_limits(struct nvme_tcp_ctrl *ctrl)
> +{
> +	struct net_device *netdev;
> +	int ret;
> +
> +	if (!ddp_offload)
> +		return NULL;
> +
> +	/* netdev ref is put in nvme_tcp_stop_admin_queue() */
> +	netdev = get_netdev_for_sock(ctrl->queues[0].sock->sk, &ctrl->netdev_tracker, GFP_KERNEL);
> +	if (!netdev) {
> +		dev_dbg(ctrl->ctrl.device, "netdev not found\n");
> +		return NULL;
> +	}
> +
> +	if (!ulp_ddp_is_cap_active(netdev, ULP_DDP_CAP_NVME_TCP))
> +		goto err;
> +
> +	ret = ulp_ddp_get_limits(netdev, &ctrl->ddp_limits, ULP_DDP_NVME);
> +	if (ret)
> +		goto err;
> +
> +	if (ctrl->ctrl.opts->tls && !ctrl->ddp_limits.tls)
> +		goto err;
> +
> +	return netdev;
> +err:
> +	netdev_put(netdev, &ctrl->netdev_tracker);
> +	return NULL;
> +}
> +
> +static bool nvme_tcp_resync_request(struct sock *sk, u32 seq, u32 flags);
> +static const struct ulp_ddp_ulp_ops nvme_tcp_ddp_ulp_ops = {
> +	.resync_request		= nvme_tcp_resync_request,
> +};
> +
> +static int nvme_tcp_offload_socket(struct nvme_tcp_queue *queue)
> +{
> +	struct ulp_ddp_config config = {.type = ULP_DDP_NVME};
> +	int ret;
> +
> +	config.nvmeotcp.pfv = NVME_TCP_PFV_1_0;
> +	config.nvmeotcp.cpda = 0;
> +	config.nvmeotcp.dgst =
> +		queue->hdr_digest ? NVME_TCP_HDR_DIGEST_ENABLE : 0;
> +	config.nvmeotcp.dgst |=
> +		queue->data_digest ? NVME_TCP_DATA_DIGEST_ENABLE : 0;
> +	config.nvmeotcp.queue_size = queue->ctrl->ctrl.sqsize + 1;
> +	config.nvmeotcp.queue_id = nvme_tcp_queue_id(queue);

I forget, why is the queue_id needed? it does not travel the wire outside
of the connect cmd.

> +
> +	ret = ulp_ddp_sk_add(queue->ctrl->ddp_netdev,
> +			     queue->sock->sk,
> +			     &config,
> +			     &nvme_tcp_ddp_ulp_ops);
> +	if (ret)
> +		return ret;
> +
> +	set_bit(NVME_TCP_Q_OFF_DDP, &queue->flags);
> +
> +	return 0;
> +}
> +
> +static void nvme_tcp_unoffload_socket(struct nvme_tcp_queue *queue)
> +{
> +	clear_bit(NVME_TCP_Q_OFF_DDP, &queue->flags);
> +	ulp_ddp_sk_del(queue->ctrl->ddp_netdev, queue->sock->sk);
> +}
> +
> +static void nvme_tcp_ddp_apply_limits(struct nvme_tcp_ctrl *ctrl)
> +{
> +	ctrl->ctrl.max_segments = ctrl->ddp_limits.max_ddp_sgl_len;
> +	ctrl->ctrl.max_hw_sectors =
> +		ctrl->ddp_limits.max_ddp_sgl_len << (ilog2(SZ_4K) - SECTOR_SHIFT);

I think you can use NVME_CTRL_PAGE_SHIFT instead of ilog2(SZ_4K)?

The rest looks fine to me.