Avoid kernel_recvmsg() about af_ktls

[fridex]

I think the best approach would be to:

• extend tcp_read_sock() with MSG_PEEK flag
• introduce udp_read_sock() with MSG_PEEK support for UDP

Using directly skbuffs is not nice, since there should be appropriate operations on UDP/TCP sockets to encapsulate such logic (and make it possible to reuse these operations in other parts of the kernel).

[fridex]

When run "splice echo time" scenario for 2 seconds a simple ping-pong with server [1]:

splice(ksd, NULL, pipe, NULL, 1400, 0);
splice(pie, NULL, ksd, NULL, 1400, 0);

With MTU 1400:

I am getting following results:

• 44.24% of total time spent in kernel_sendmsg()
  • 38.28% of total time spent in tcp_push - on actual sending
  • 1.15% of total time spent in allocation socket buffers skb_stream_alloc_skb
  • cca 2% on copy from kernel vector (copy_from_iter, memcpy_erms)
• 33.14% of total time spent in tls_splice_read
  • 13.14% of total time spent in kernel_recvmsg
  • cca 2% on copy and allocation (skb_copy_datagram_iter, copy_page_to_iter)

With MTU 16000:

I am getting following results:

• 22.29% of total time spent in kernel_sendmsg()
  • 16.30% of total time spent in tcp_push - on actual sending
  • 0.69% of total time spent in allocation socket buffers skb_stream_alloc_skb
  • 3.03% on copy from kernel vector (copy_from_iter, memcpy_erms)
• 42.25% of total time spent in tls_splice_read
  • 9.02% of total time spent in kernel_recvmsg
  • 4.02 % on copy and allocation (skb_copy_datagram_iter, copy_page_to_iter)

Ideally we could save:

• for 1400 MTU:
  • cca 2% by avoiding kernel_recvmsg()
  • cca 3.15% by avoiding kernel_sendmsg()
• for 16000 MTU:
  • 3.72% by avoiding kernel_sendmsg()
  • 4.02% by avoiding kernel_recvmsg()

We have to consider addional logic within kernel_sendmsg() and kernel_recvmsg() (locking, ...). Using kernel_sendpage() and tcp_read_sock() (udp_read_sock()) can have different logic which could have positive/negative impact as well.

perf reporting context switches not expensive at all (0.30% of total)

related: https://github.com/fridex/af_ktls/issues/22

[1] https://github.com/fridex/af_ktls-tool/blob/master/action.c#L795

[djwatson]

fixed by #62