From: Jakub Kicinski <kuba@xxxxxxxxxx>
Date: Tue, 30 Nov 2021 08:12:07 -0800
> On Tue, 30 Nov 2021 16:56:12 +0100 Alexander Lobakin wrote:
> > 3. XDP and XSK ctrs separately or not.
> >
> > My PoV is that those are two quite different worlds.
> > However, stats for actions on XSK really make a little sense since
> > 99% of time we have xskmap redirect. So I think it'd be fine to just
> > expand stats structure with xsk_{rx,tx}_{packets,bytes} and count
> > the rest (actions, errors) together with XDP.
> >
> >
> > Rest:
> > - don't create a separate `ip` command and report under `-s`;
> > - save some RTNL skb space by skipping zeroed counters.
>
> Let me ruin this point of clarity for you. I think that stats should
> be skipped when they are not collected (see ETHTOOL_STAT_NOT_SET).
> If messages get large user should use the GETSTATS call and avoid
> the problem more effectively.
Well, it was Dave's thought here: [0]
> Another thought on this patch: with individual attributes you could save
> some overhead by not sending 0 counters to userspace. e.g., define a
> helper that does:
I know about ETHTOOL_STAT_NOT_SET, but RTNL xstats doesn't use this,
does it?
GETSTATS is another thing, and I'll use it, thanks.
>
> > Also, regarding that I count all on the stack and then add to the
> > storage once in a polling cycle -- most drivers don't do that and
> > just increment the values in the storage directly, but this can be
> > less performant for frequently updated stats (or it's just my
> > embedded past).
> > Re u64 vs u64_stats_t -- the latter is more universal and
> > architecture-friendly, the former is used directly in most of the
> > drivers primarily because those drivers and the corresponding HW
> > are being run on 64-bit systems in the vast majority of cases, and
> > Ethtools stats themselves are not so critical to guard them with
> > anti-tearing. Anyways, local64_t is cheap on ARM64/x86_64 I guess?
[0] https://lore.kernel.org/netdev/a4602b15-25b1-c388-73b4-1f97f6f0e555@xxxxxxxxx/
Al
