On Thu, 19 Apr 2018 13:44:41 +0100
Quentin Monnet <quentin.mon...@netronome.com> wrote:
> 2018-04-18 17:43 UTC+0200 ~ Jesper Dangaard Brouer <bro...@redhat.com>
> > On Wed, 18 Apr 2018 15:09:41 +0100
> > Quentin Monnet <quentin.mon...@netronome.com> wrote:
> >
> >> 2018-04-18 15:34 UTC+0200 ~ Jesper Dangaard Brouer <bro...@redhat.com>
> >>> On Tue, 17 Apr 2018 15:34:38 +0100
> >>> Quentin Monnet <quentin.mon...@netronome.com> wrote:
> >>>
> >>>> diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
> >>>> index 350459c583de..3d329538498f 100644
> >>>> --- a/include/uapi/linux/bpf.h
> >>>> +++ b/include/uapi/linux/bpf.h
> >>>> @@ -1276,6 +1276,50 @@ union bpf_attr {
> >>>> * Return
> >>>> * 0 on success, or a negative error in case of failure.
> >>>> *
> >>>> + * int bpf_redirect_map(struct bpf_map *map, u32 key, u64 flags)
> >>>> + * Description
> >>>> + * Redirect the packet to the endpoint referenced by *map*
> >>>> at
> >>>> + * index *key*. Depending on its type, his *map* can
> >>>> contain
> >>> ^^^
> >>>
> >>> "his" -> "this"
> >>
> >> Thanks!
> >>
> >>>> + * references to net devices (for forwarding packets
> >>>> through other
> >>>> + * ports), or to CPUs (for redirecting XDP frames to
> >>>> another CPU;
> >>>> + * but this is only implemented for native XDP (with driver
> >>>> + * support) as of this writing).
> >>>> + *
> >>>> + * All values for *flags* are reserved for future usage,
> >>>> and must
> >>>> + * be left at zero.
> >>>> + * Return
> >>>> + * **XDP_REDIRECT** on success, or **XDP_ABORT** on error.
> >>>> + *
> >>>
> >>> "XDP_ABORT" -> "XDP_ABORTED"
> >>
> >> Ouch. And I did the same for bpf_redirect(). Thanks for the catch.
> >>
> >>>
> >>> I don't know if it's worth mentioning in the doc/man-page; that for XDP
> >>> using bpf_redirect_map() is a HUGE performance advantage, compared to
> >>> the bpf_redirect() call ?
> >>
> >> It seems worth to me. How would you simply explain the reason for this
> >> difference?
> >
> > The basic reason is "bulking effect", as devmap avoids the NIC
> > tailptr/doorbell update on every packet... how to write that in a doc
> > format?
> >
> > I wrote about why XDP_REDIRECT with maps are smart here:
> >
> > http://vger.kernel.org/netconf2017_files/XDP_devel_update_NetConf2017_Seoul.pdf
> >
> > Using maps for redirect, hopefully makes XDP_REDIRECT the last driver
> > XDP action code we need. As new types of redirect can be introduced
> > without driver changes. See that AF_XDP also uses a map.
> >
> > It is more subtle, but maps also function as a sorting step. Imagine
> > your XDP program need to redirect out different interfaces (or CPUs in
> > cpumap case), and packets arrive intermixed. Packets get sorted into
> > the different map indexes, and the xdp_do_flush_map() will trigger the
> > flush operation.
> >
> >
> > Happened to have an i40e NIC benchmark setup, and ran a single flow pktgen
> > test.
> >
> > Results with 'xdp_redirect_map'
> > 13589297 pps (13,589,297)
> >
> > Results with 'xdp_redirect' NOT using devmap:
> > 7567575 pps (7,567,575)
> >
> > Just to point out the performance benefit of devmap...
>
>
> Thanks for those details! This is an impressive change in performance
> indeed.
>
> I think I will just keep it simple for the documentation. I will add the
> following for bpf_redirect_map():
>
> When used to redirect packets to net devices, this helper
> provides a high performance increase over **bpf_redirect**\ ().
> This is due to various implementation details of the underlying
> mechanisms, one of which is the fact that **bpf_redirect_map**\ ()
> tries to send packet as a "bulk" to the device.
>
> And also append the following to bpf_redirect():
>
> The same effect can be attained with the more generic
> **bpf_redirect_map**\ (), which requires specific maps
> to be used but offers better performance.
This sounds good to me! :-)
--
Best regards,
Jesper Dangaard Brouer
MSc.CS, Principal Kernel Engineer at Red Hat
LinkedIn: http://www.linkedin.com/in/brouer
