Re: [Intel-wired-lan] FW: [PATCH bpf-next 2/4] xsk: allow AF_XDP sockets to receive packets directly from a queue
From: "Samudrala, Sridhar" <sridhar.samudrala@intel.com>
Date: 2019-10-31 22:38:44
Also in:
bpf, intel-wired-lan
[...]
quoted
With mitigations ON ------------------- Samples: 6K of event 'cycles', 4000 Hz, Event count (approx.): 5646512726 bpf_prog_3c8251c7e0fef8db bpf_prog_3c8251c7e0fef8db [Percent: local period] 45.05 push %rbp 0.02 mov %rsp,%rbp 0.03 sub $0x8,%rsp 22.09 push %rbx
[...]quoted
Do you see any issues with this data? With mitigations ON push %rbp and push %rbx overhead seems to be pretty high.
That's sample skid from the retpoline thunk when entring the XDP program. Pretty expensive push otherwise! :-)
Another thought; Disclaimer: I'm no spectrev2 expert, and probably not getting the mitigations well enough. So this is me trying to swim at the deep end! Would it be possible to avoid the retpoline when entering the XDP program. At least for some XDP program that can be proved "safe"? If so, PeterZ's upcoming static_call could be used from the driver side.
Alexei, Jakub Do you think it will be possible to avoid this overhead when mitigations are turned ON? The other part of the overhead is going through the redirect path. Can i assume that your silence as an indication that you are now okay with optional bypass flag as long as it doesn't effect the normal XDP datapath. If so, i will respin and submit the patches against the latest bpf-next -Sridhar