Can we use zero-copy for TCP send/recv with the default linux TCP/IP-stack?

Question

Can we use zero-copy for TCP send/recv with the default linux TCP/IP-stack?

• As known, we can re-map socket buffer from kernel-space to the user-space for RAW-sockets: https://www.kernel.org/doc/Documentation/networking/packet_mmap.txt

Example:

int packet_socket = socket(AF_PACKET, SOCK_RAW, htons(ETH_P_ALL)); // raw-sockets

struct tpacket_req3 req;  
setsockopt(packet_socket, SOL_PACKET , PACKET_RX_RING , (void*)&req , sizeof(req));
mapped_buffer = (uint8_t*)mmap(NULL, req.tp_block_size * req.tp_block_nr,
    PROT_READ | PROT_WRITE, MAP_SHARED | MAP_LOCKED, packet_socket, 0);

• Also known, that payload doesn't change while lifting up using the TCP-stack: https://www.informatix-sol.com/docs/TCP_bypass.pdf

So can we map part of socket-buffer with received payload from kernel-space to user-space to avoid zero-copy?

Answer 1

4.18 merged zero-copy receive support for the in-kernel TCP stack, see:

• https://lwn.net/Articles/754681/
• https://lwn.net/Articles/752188/

Answer 2

Technically it is impossible to implement zero-copy TCP.

Assume there are multiple application using TCP stack, each app owns a memory region where tcp data is supposed to be written.

When a packet is coming, the NIC hardware doesn't know which connection the packet is belonged to, so the NIC can't determine which memory region to write the packet.

So the only way is first writes to a kernel region and then copy to the app region later.