Improve IXGBE driver performance

[jiegec]

Based on iperf.

Current status (tcp, maximum observed): rcore -> linux 1.24Gb/s -> 1.40Gb/s -> 2.18Gb/s -> 2.62Gb/s -> 3.50Gb/s linux -> rcore 308Mb/s -> 320Mb/s -> 379Mb/s -> 407Mb/s -> 1.40Gb/s -> 2.47Gb/s -> 2.53Gb/s -> 3.74Gb/s -> 4.44Gb/s

Udp is not working because smoltcp does not support segmentation.

Possible ways to improve:

• [ ] IP checksum TX
• [x] IP checksum RX: little improvement
• [ ] UDP checksum TX
• [ ] UDP checksum RX
• [ ] TCP checksum TX
• [ ] TCP checksum RX
• [ ] TCP segmentation offload
• [x] Interrupt throttling: good for tx, bad for rx without spinning
• [x] Spin for some times before waiting for condvar: observable improvement for rx
• [x] Enlarge TCP recv/trans buffers: observable improvement for rx
• [x] Replace linked_list_allocator with buddy_system_allocator: a little improvement for rx
• [x] Enable jumbo frame: observable improvement

Manually disable RX checksum for smoltcp seems to have little improvement.

I think we have met smoltcp's performance limit.

Test under linux:

Put two interfaces into two netns, then use iperf to test:

tcp: 9.38Gb/s

[jiegec]

Idea: Check performance of sending raw ethernet packets, skipping network stack.