From: Toke Høiland-Jørgensen <toke@xxxxxxxxxx> Date: Tue, 18 Feb 2020 23:23:22 +0100 > Jakub Kicinski <kuba@xxxxxxxxxx> writes: > >> On Tue, 18 Feb 2020 01:14:29 +0100 Lorenzo Bianconi wrote: >>> Introduce "rx" prefix in the name scheme for xdp counters >>> on rx path. >>> Differentiate between XDP_TX and ndo_xdp_xmit counters >>> >>> Signed-off-by: Lorenzo Bianconi <lorenzo@xxxxxxxxxx> >> >> Sorry for coming in late. >> >> I thought the ability to attach a BPF program to a fexit of another BPF >> program will put an end to these unnecessary statistics. IOW I maintain >> my position that there should be no ethtool stats for XDP. >> >> As discussed before real life BPF progs will maintain their own stats >> at the granularity of their choosing, so we're just wasting datapath >> cycles. >> >> The previous argument that the BPF prog stats are out of admin control >> is no longer true with the fexit option (IIUC how that works). > > So you're proposing an admin that wants to keep track of XDP has to > (permantently?) attach an fexit program to every running XDP program and > use that to keep statistics? But presumably he'd first need to discover > that XDP is enabled; which the ethtool stats is a good hint for :) Really, mistakes happen and a poorly implemented or inserted fexit module should not be a reason to not have access to accurate and working statistics for fundamental events. I am therefore totally against requiring fexit for this functionality. If you want more sophisticated events or custome ones, sure, but not for this baseline stuff. I do, however, think we need a way to turn off these counter bumps if the user wishes to do so for maximum performance.
