Device Memory TCP

Intro

Device memory TCP (devmem TCP) enables receiving data directly into device memory (dmabuf). The feature is currently implemented for TCP sockets.

Opportunity

A large number of data transfers have device memory as the source and/or destination. Accelerators drastically increased the prevalence of such transfers. Some examples include:

  • Distributed training, where ML accelerators, such as GPUs on different hosts, exchange data.

  • Distributed raw block storage applications transfer large amounts of data with remote SSDs. Much of this data does not require host processing.

Typically the Device-to-Device data transfers in the network are implemented as the following low-level operations: Device-to-Host copy, Host-to-Host network transfer, and Host-to-Device copy.

The flow involving host copies is suboptimal, especially for bulk data transfers, and can put significant strains on system resources such as host memory bandwidth and PCIe bandwidth.

Devmem TCP optimizes this use case by implementing socket APIs that enable the user to receive incoming network packets directly into device memory.

Packet payloads go directly from the NIC to device memory.

Packet headers go to host memory and are processed by the TCP/IP stack normally. The NIC must support header split to achieve this.

Advantages:

  • Alleviate host memory bandwidth pressure, compared to existing network-transfer + device-copy semantics.

  • Alleviate PCIe bandwidth pressure, by limiting data transfer to the lowest level of the PCIe tree, compared to the traditional path which sends data through the root complex.

More Info

Interface

Example

tools/testing/selftests/net/ncdevmem.c:do_server shows an example of setting up the RX path of this API.

NIC Setup

Header split, flow steering, & RSS are required features for devmem TCP.

Header split is used to split incoming packets into a header buffer in host memory, and a payload buffer in device memory.

Flow steering & RSS are used to ensure that only flows targeting devmem land on an RX queue bound to devmem.

Enable header split & flow steering:

# enable header split
ethtool -G eth1 tcp-data-split on


# enable flow steering
ethtool -K eth1 ntuple on

Configure RSS to steer all traffic away from the target RX queue (queue 15 in this example):

ethtool --set-rxfh-indir eth1 equal 15

The user must bind a dmabuf to any number of RX queues on a given NIC using the netlink API:

/* Bind dmabuf to NIC RX queue 15 */
struct netdev_queue *queues;
queues = malloc(sizeof(*queues) * 1);

queues[0]._present.type = 1;
queues[0]._present.idx = 1;
queues[0].type = NETDEV_RX_QUEUE_TYPE_RX;
queues[0].idx = 15;

*ys = ynl_sock_create(&ynl_netdev_family, &yerr);

req = netdev_bind_rx_req_alloc();
netdev_bind_rx_req_set_ifindex(req, 1 /* ifindex */);
netdev_bind_rx_req_set_dmabuf_fd(req, dmabuf_fd);
__netdev_bind_rx_req_set_queues(req, queues, n_queue_index);

rsp = netdev_bind_rx(*ys, req);

dmabuf_id = rsp->dmabuf_id;

The netlink API returns a dmabuf_id: a unique ID that refers to this dmabuf that has been bound.

The user can unbind the dmabuf from the netdevice by closing the netlink socket that established the binding. We do this so that the binding is automatically unbound even if the userspace process crashes.

Note that any reasonably well-behaved dmabuf from any exporter should work with devmem TCP, even if the dmabuf is not actually backed by devmem. An example of this is udmabuf, which wraps user memory (non-devmem) in a dmabuf.

Socket Setup

The socket must be flow steered to the dmabuf bound RX queue:

ethtool -N eth1 flow-type tcp4 ... queue 15

Receiving data

The user application must signal to the kernel that it is capable of receiving devmem data by passing the MSG_SOCK_DEVMEM flag to recvmsg:

ret = recvmsg(fd, &msg, MSG_SOCK_DEVMEM);

Applications that do not specify the MSG_SOCK_DEVMEM flag will receive an EFAULT on devmem data.

Devmem data is received directly into the dmabuf bound to the NIC in ‘NIC Setup’, and the kernel signals such to the user via the SCM_DEVMEM_* cmsgs:

for (cm = CMSG_FIRSTHDR(&msg); cm; cm = CMSG_NXTHDR(&msg, cm)) {
        if (cm->cmsg_level != SOL_SOCKET ||
                (cm->cmsg_type != SCM_DEVMEM_DMABUF &&
                 cm->cmsg_type != SCM_DEVMEM_LINEAR))
                continue;

        dmabuf_cmsg = (struct dmabuf_cmsg *)CMSG_DATA(cm);

        if (cm->cmsg_type == SCM_DEVMEM_DMABUF) {
                /* Frag landed in dmabuf.
                 *
                 * dmabuf_cmsg->dmabuf_id is the dmabuf the
                 * frag landed on.
                 *
                 * dmabuf_cmsg->frag_offset is the offset into
                 * the dmabuf where the frag starts.
                 *
                 * dmabuf_cmsg->frag_size is the size of the
                 * frag.
                 *
                 * dmabuf_cmsg->frag_token is a token used to
                 * refer to this frag for later freeing.
                 */

                struct dmabuf_token token;
                token.token_start = dmabuf_cmsg->frag_token;
                token.token_count = 1;
                continue;
        }

        if (cm->cmsg_type == SCM_DEVMEM_LINEAR)
                /* Frag landed in linear buffer.
                 *
                 * dmabuf_cmsg->frag_size is the size of the
                 * frag.
                 */
                continue;

}

Applications may receive 2 cmsgs:

  • SCM_DEVMEM_DMABUF: this indicates the fragment landed in the dmabuf indicated by dmabuf_id.

  • SCM_DEVMEM_LINEAR: this indicates the fragment landed in the linear buffer. This typically happens when the NIC is unable to split the packet at the header boundary, such that part (or all) of the payload landed in host memory.

Applications may receive no SO_DEVMEM_* cmsgs. That indicates non-devmem, regular TCP data that landed on an RX queue not bound to a dmabuf.

Freeing frags

Frags received via SCM_DEVMEM_DMABUF are pinned by the kernel while the user processes the frag. The user must return the frag to the kernel via SO_DEVMEM_DONTNEED:

ret = setsockopt(client_fd, SOL_SOCKET, SO_DEVMEM_DONTNEED, &token,
                 sizeof(token));

The user must ensure the tokens are returned to the kernel in a timely manner. Failure to do so will exhaust the limited dmabuf that is bound to the RX queue and will lead to packet drops.

The user must pass no more than 128 tokens, with no more than 1024 total frags among the token->token_count across all the tokens. If the user provides more than 1024 frags, the kernel will free up to 1024 frags and return early.

The kernel returns the number of actual frags freed. The number of frags freed can be less than the tokens provided by the user in case of:

  1. an internal kernel leak bug.

  2. the user passed more than 1024 frags.

Implementation & Caveats

Unreadable skbs

Devmem payloads are inaccessible to the kernel processing the packets. This results in a few quirks for payloads of devmem skbs:

  • Loopback is not functional. Loopback relies on copying the payload, which is not possible with devmem skbs.

  • Software checksum calculation fails.

  • TCP Dump and bpf can’t access devmem packet payloads.

Testing

More realistic example code can be found in the kernel source under tools/testing/selftests/net/ncdevmem.c

ncdevmem is a devmem TCP netcat. It works very similarly to netcat, but receives data directly into a udmabuf.

To run ncdevmem, you need to run it on a server on the machine under test, and you need to run netcat on a peer to provide the TX data.

ncdevmem has a validation mode as well that expects a repeating pattern of incoming data and validates it as such. For example, you can launch ncdevmem on the server by:

ncdevmem -s <server IP> -c <client IP> -f eth1 -d 3 -n 0000:06:00.0 -l \
         -p 5201 -v 7

On client side, use regular netcat to send TX data to ncdevmem process on the server:

yes $(echo -e \\x01\\x02\\x03\\x04\\x05\\x06) | \
        tr \\n \\0 | head -c 5G | nc <server IP> 5201 -p 5201