[dpdk-dev] lpm patches

[Vladimir Medvedkin]

Hi all,

2015-10-30 20:59 GMT+03:00 Matthew Hall <mhall at mhcomputing.net>:

> On Fri, Oct 30, 2015 at 12:00:18PM +0000, Bruce Richardson wrote:
> > Matthew's patches were attachments, I don't think they came through in
> patchwork
> > correctly :-(, but that is the relevant link there anyway.]
>
> Let me know if there is something I can do better there. I was having a
> difficult time figuring out how to preserve the thread ID in the middle of
> the
> thread and not cause a new thread. The git email workflows are very
> confusing
> and I figured it was better to send something as soon as I could.
>
> > * Some patches increase the next-hop to 16 bits, others to 24-bits. In
> both cases
> >   a single entry still only occupies 32-bits, so can be read/written to
> >   atomically
>
> I went with 24 because it was the biggest amount I could get that still had
> this property.
>
> > * Only Michal's set appears to take into account ABI versioning, which is
> >   a difficult problem for this lib, with inlined functions.
>
> Agreed. His patches are the most professional from this perspective. This
> is
> why I was trying to contribute to you and to him so we get the most
> professional result for the customers.
>
> > * Matthew's patchset moves the lookup functions to be non-inlined, which
> will
> >   make future updates easier from ABI compatibility - at the cost of
> lookup
> >   performance.
>
> This point is optional for me. I did it, because without it, it was totally
> impossible for me to work on the code in a debugger as I am a security
> engineering guy not a crazy embedded C coder or kernel hacker.
>
> > * Vladimir's patchset merges in the tbl24 and tbl8 entries into a single
> data
> >   type.
>
> I really liked this feature of Vladimir's patches, it makes it easier to
> maintain and less confusing. I had a lot of headaches keeping all those
> structs straight with the separate types, but I didn't know we had the
> chance
> for a great big MEGA-REFACTOR. I love this community!
>
> > * That patchset also introduces an extra optional 32-bit field "as_num",
> allowing
> >   64-bit lpm table entries - obviously at a cost of increased
> memory/cache
> >   footprint.
>
> Is there a way we could test it? Vladimir, did you test the performance? If
> so, what happened?
>
Performance regression depends on the traffic pattern and cache size.

>
> > * Stephen's patchset includes a range of other fixes e.g. for more
> efficient
> >   management of the rules array, and dynamic allocation of the TBL8s.
> > * Matthew's patchset also includes change to LPM for IPv6, which I'm
> considering
> >   out-of-scope for now, so as to focus on LPM v4 only.
>
> Any chance that is inconsistent betwen LPM4 and LPM6 really hoses me,
> because
> I am writing green-field code which treats both protocols as first-class
> citizens and I'd really not like to have totally inconsistent and inferior
> support in one versus the other.
>
> > * Increase next hops to be the full 24 bits, so as to allow maximum
> flexibility
> >   and not waste the extra 8 bits of space in the 32-bit entries.
>
> +1
>
+1. Split of next hop and forwarding class I can do in app logic.

>
> > * Move the lookup functions which work on multiple packets to be
> non-inlined
>
> Open to opinions on the performance of this. I am not an expert on this
> area.
>
> > * Merge in the tbl24 and tbl8 structures to make the code that little
> bit shorter
>
> +1
>
> > * Look to pull in as many of Stephen's other improvements as possible -
> though
> >   this may be in a separate patchset to the other changes.
>
> +1. Perhaps if we get a pre-release on a branch with everything else, we
> could
> see if Stephen is willing to rebase his non-duplicate changes.
>
> > * I'm uncertain as to the extra 32-bit as_num field. Adding it as an
> extra
> >   #define is trivial, but adds to the compile-time config. Having it as
> a run-time
> >   option is possible, but likely will make the code a lot more
> complicated, as
> >   we no longer have arrays of a fixed size.
> >
> > Naturally, with whatever solution is come up with, ABI compatibility
> must be
> > taken into account and functions versionned appropriately!
>
> Normally I am not a big define guy. But it seems like a define is good
> here.
> Somebody is going to need to know beforehand if they are making a Core
> Router
> where they want this, or a Security Inspection system like mine, etc.
>
For example in case of core router as_num feature can be necessary for
netflow. It can be necessary in case of security device such as my ddos
mitigation system.

>
> So it seems easier than doing a bunch of crazy size-juggling in the code.
>
> > do we want to have some of these changes in 2.2?
>
> Personally I am OK to wait as I have it working in my copy. I am just
> trying
> to be a good citizen of the community and contribute back when I see some
> core
> engineers going after the same code.
>
> In particular, for me, having LPM4 only with no LPM6 is not worth much so
> I'd
> be happy to wait for a single upgrade to both of them.
>
> > Matthew, Stephen, Vladimir, Michal, Thomas - thoughts on this?
> > [do I accurately sum up the situation?]
>
> This email was top-quality and very well done by you guys.
>
> Matthew.
>

Regards,
Vladimir