Message-ID: <CAM0EoM=RR6kcdHsGhFNUeDc96rSDa8S7SP7GQOeXrZBN_P7jtQ@mail.gmail.com>
Date: Wed, 22 Nov 2023 14:35:21 -0500
From: Jamal Hadi Salim <jhs@...atatu.com>
To: Jiri Pirko <jiri@...nulli.us>
Cc: Daniel Borkmann <daniel@...earbox.net>, John Fastabend <john.fastabend@...il.com>, 
	netdev@...r.kernel.org, deb.chatterjee@...el.com, anjali.singhai@...el.com, 
	Vipin.Jain@....com, namrata.limaye@...el.com, tom@...anda.io, 
	mleitner@...hat.com, Mahesh.Shirshyad@....com, tomasz.osinski@...el.com, 
	xiyou.wangcong@...il.com, davem@...emloft.net, edumazet@...gle.com, 
	kuba@...nel.org, pabeni@...hat.com, vladbu@...dia.com, horms@...nel.org, 
	bpf@...r.kernel.org, khalidm@...dia.com, toke@...hat.com, mattyk@...dia.com, 
	dan.daly@...el.com, chris.sommers@...sight.com, john.andy.fingerhut@...el.com
Subject: Re: [PATCH net-next v8 00/15] Introducing P4TC

On Wed, Nov 22, 2023 at 1:31 PM Jiri Pirko <jiri@...nulli.us> wrote:
>
> Wed, Nov 22, 2023 at 04:14:02PM CET, jhs@...atatu.com wrote:
> >On Wed, Nov 22, 2023 at 4:25 AM Jiri Pirko <jiri@...nulli.us> wrote:
> >>
> >> Tue, Nov 21, 2023 at 04:21:44PM CET, jhs@...atatu.com wrote:
> >> >On Tue, Nov 21, 2023 at 9:19 AM Jiri Pirko <jiri@...nulli.us> wrote:
> >> >>
> >> >> Tue, Nov 21, 2023 at 02:47:40PM CET, jhs@...atatu.com wrote:
> >> >> >On Tue, Nov 21, 2023 at 8:06 AM Jiri Pirko <jiri@...nulli.us> wrote:
> >> >> >>
> >> >> >> Mon, Nov 20, 2023 at 11:56:50PM CET, jhs@...atatu.com wrote:
> >> >> >> >On Mon, Nov 20, 2023 at 4:49 PM Daniel Borkmann <daniel@...earbox.net> wrote:
> >> >> >> >>
> >> >> >> >> On 11/20/23 8:56 PM, Jamal Hadi Salim wrote:
> >> >> >> >> > On Mon, Nov 20, 2023 at 1:10 PM Jiri Pirko <jiri@...nulli.us> wrote:
> >> >> >> >> >> Mon, Nov 20, 2023 at 03:23:59PM CET, jhs@...atatu.com wrote:
> >> >> >>
> >> >> >> [...]
> >> >> >>
> >> >> >> >
> >> >> >> >> tc BPF and XDP already have widely used infrastructure and can be developed
> >> >> >> >> against libbpf or other user space libraries for a user space control plane.
> >> >> >> >> With 'control plane' you refer here to the tc / netlink shim you've built,
> >> >> >> >> but looking at the tc command line examples, this doesn't really provide a
> >> >> >> >> good user experience (you call it p4 but people load bpf obj files). If the
> >> >> >> >> expectation is that an operator should run tc commands, then neither it's
> >> >> >> >> a nice experience for p4 nor for BPF folks. From a BPF PoV, we moved over
> >> >> >> >> to bpf_mprog and plan to also extend this for XDP to have a common look and
> >> >> >> >> feel wrt networking for developers. Why can't this be reused?
> >> >> >> >
> >> >> >> >The filter loading which loads the program is considered pipeline
> >> >> >> >instantiation - consider it as "provisioning" more than "control"
> >> >> >> >which runs at runtime. "control" is purely netlink based. The iproute2
> >> >> >> >code we use links libbpf for example for the filter. If we can achieve
> >> >> >> >the same with bpf_mprog then sure - we just dont want to loose
> >> >> >> >functionality though.  off top of my head, some sample space:
> >> >> >> >- we could have multiple pipelines with different priorities (which tc
> >> >> >> >provides to us) - and each pipeline may have its own logic with many
> >> >> >> >tables etc (and the choice to iterate the next one is essentially
> >> >> >> >encoded in the tc action codes)
> >> >> >> >- we use tc block to map groups of ports (which i dont think bpf has
> >> >> >> >internal access of)
> >> >> >> >
> >> >> >> >In regards to usability: no i dont expect someone doing things at
> >> >> >> >scale to use command line tc. The APIs are via netlink. But the tc cli
> >> >> >> >is must for the rest of the masses per our traditions. Also i really
> >> >> >>
> >> >> >> I don't follow. You repeatedly mention "the must of the traditional tc
> >> >> >> cli", but what of the existing traditional cli you use for p4tc?
> >> >> >> If I look at the examples, pretty much everything looks new to me.
> >> >> >> Example:
> >> >> >>
> >> >> >>   tc p4ctrl create myprog/table/mytable dstAddr 10.0.1.2/32 \
> >> >> >>     action send_to_port param port eno1
> >> >> >>
> >> >> >> This is just TC/RTnetlink used as a channel to pass new things over. If
> >> >> >> that is the case, what's traditional here?
> >> >> >>
> >> >> >
> >> >> >
> >> >> >What is not traditional about it?
> >> >>
> >> >> Okay, so in that case, the following example communitating with
> >> >> userspace deamon using imaginary "p4ctrl" app is equally traditional:
> >> >>   $ p4ctrl create myprog/table/mytable dstAddr 10.0.1.2/32 \
> >> >>      action send_to_port param port eno1
> >> >
> >> >Huh? Thats just an application - classical tc which part of iproute2
> >> >that is sending to the kernel, no different than "tc flower.."
> >> >Where do you get the "userspace" daemon part? Yes, you can write a
> >> >daemon but it will use the same APIs as tc.
> >>
> >> Okay, so which part is the "tradition"?
> >>
> >
> >Provides tooling via tc cli that _everyone_ in the tc world is
> >familiar with - which uses the same syntax as other tc extensions do,
> >same expectations (eg events, request responses, familiar commands for
> >dumping, flushing etc). Basically someone familiar with tc will pick
> >this up and operate it very quickly and would have an easier time
> >debugging it.
> >There are caveats - as will be with all new classifiers - but those
> >are within reason.
>
> Okay, so syntax familiarity wise, what's the difference between
> following 2 approaches:
> $ tc p4ctrl create myprog/table/mytable dstAddr 10.0.1.2/32 \
>       action send_to_port param port eno1
> $ p4ctrl create myprog/table/mytable dstAddr 10.0.1.2/32 \
>       action send_to_port param port eno1
> ?
>
>
> >
> >> >>
> >> >> >
> >> >> >>
> >> >> >> >didnt even want to use ebpf at all for operator experience reasons -
> >> >> >> >it requires a compilation of the code and an extra loading compared to
> >> >> >> >what our original u32/pedit code offered.
> >> >> >> >
> >> >> >> >> I don't quite follow why not most of this could be implemented entirely in
> >> >> >> >> user space without the detour of this and you would provide a developer
> >> >> >> >> library which could then be integrated into a p4 runtime/frontend? This
> >> >> >> >> way users never interface with ebpf parts nor tc given they also shouldn't
> >> >> >> >> have to - it's an implementation detail. This is what John was also pointing
> >> >> >> >> out earlier.
> >> >> >> >>
> >> >> >> >
> >> >> >> >Netlink is the API. We will provide a library for object manipulation
> >> >> >> >which abstracts away the need to know netlink. Someone who for their
> >> >> >> >own reasons wants to use p4runtime or TDI could write on top of this.
> >> >> >> >I would not design a kernel interface to just meet p4runtime (we
> >> >> >> >already have TDI which came later which does things differently). So i
> >> >> >> >expect us to support both those two. And if i was to do something on
> >> >> >> >SDN that was more robust i would write my own that still uses these
> >> >> >> >netlink interfaces.
> >> >> >>
> >> >> >> Actually, what Daniel says about the p4 library used as a backend to p4
> >> >> >> frontend is pretty much aligned what I claimed on the p4 calls couple of
> >> >> >> times. If you have this p4 userspace tooling, it is easy for offloads to
> >> >> >> replace the backed by vendor-specific library which allows p4 offload
> >> >> >> suitable for all vendors (your plan of p4tc offload does not work well
> >> >> >> for our hw, as we repeatedly claimed).
> >> >> >>
> >> >> >
> >> >> >That's you - NVIDIA. You have chosen a path away from the kernel
> >> >> >towards DOCA. I understand NVIDIA's frustration with dealing with
> >> >> >upstream process (which has been cited to me as a good reason for
> >> >> >DOCA) but please dont impose these values and your politics on other
> >> >> >vendors(Intel, AMD for example) who are more than willing to invest
> >> >> >into making the kernel interfaces the path forward. Your choice.
> >> >>
> >> >> No, you are missing the point. This has nothing to do with DOCA.
> >> >
> >> >Right Jiri ;->
> >> >
> >> >> This
> >> >> has to do with the simple limitation of your offload assuming there are
> >> >> no runtime changes in the compiled pipeline. For Intel, maybe they
> >> >> aren't, and it's a good fit for them. All I say is, that it is not the
> >> >> good fit for everyone.
> >> >
> >> > a) it is not part of the P4 spec to dynamically make changes to the
> >> >datapath pipeline after it is create and we are discussing a P4
> >>
> >> Isn't this up to the implementation? I mean from the p4 perspective,
> >> everything is static. Hw might need to reshuffle the pipeline internally
> >> during rule insertion/remove in order to optimize the layout.
> >>
> >
> >But do note: the focus here is on P4 (hence the name P4TC).
> >
> >> >implementation not an extension that would add more value b) We are
> >> >more than happy to add extensions in the future to accomodate for
> >> >features but first _P4 spec_ must be met c) we had longer discussions
> >> >with Matty, Khalid and the Rice folks who wrote a paper on that topic
> >> >which you probably didnt attend and everything that needs to be done
> >> >can be from user space today for all those optimizations.
> >> >
> >> >Conclusion is: For what you need to do (which i dont believe is a
> >> >limitation in your hardware rather a design decision on your part) run
> >> >your user space daemon, do optimizations and update the datapath.
> >> >Everybody is happy.
> >>
> >> Should the userspace daemon listen on inserted rules to be offloade
> >> over netlink?
> >>
> >
> >I mean you could if you wanted to given this is just traditional
> >netlink which emits events (with some filtering when we integrate the
> >filter approach). But why?
>
> Nevermind.
>
>
> >
> >> >
> >> >>
> >> >> >Nobody is stopping you from offering your customers proprietary
> >> >> >solutions which include a specific ebpf approach alongside DOCA. We
> >> >> >believe that a singular interface regardless of the vendor is the
> >> >> >right way forward. IMHO, this siloing that unfortunately is also added
> >> >> >by eBPF being a double edged sword is not good for the community.
> >> >> >
> >> >> >> As I also said on the p4 call couple of times, I don't see the kernel
> >> >> >> as the correct place to do the p4 abstractions. Why don't you do it in
> >> >> >> userspace and give vendors possiblity to have p4 backends with compilers,
> >> >> >> runtime optimizations etc in userspace, talking to the HW in the
> >> >> >> vendor-suitable way too. Then the SW implementation could be easily eBPF
> >> >> >> and the main reason (I believe) why you need to have this is TC
> >> >> >> (offload) is then void.
> >> >> >>
> >> >> >> The "everyone wants to use TC/netlink" claim does not seem correct
> >> >> >> to me. Why not to have one Linux p4 solution that fits everyones needs?
> >> >> >
> >> >> >You mean more fitting to the DOCA world? no, because iam a kernel
> >> >>
> >> >> Again, this has 0 relation to DOCA.
> >> >>
> >> >>
> >> >> >first person and kernel interfaces are good for everyone.
> >> >>
> >> >> Yeah, not really. Not always the kernel is the right answer. Your/Intel
> >> >> plan to handle the offload by:
> >> >> 1) abuse devlink to flash p4 binary
> >> >> 2) parse the binary in kernel to match to the table ids of rules coming
> >> >>    from p4tc ndo_setup_tc
> >> >> 3) abuse devlink to flash p4 binary for tc-flower
> >> >> 4) parse the binary in kernel to match to the table ids of rules coming
> >> >>    from tc-flower ndo_setup_tc
> >> >> is really something that is making me a little bit nauseous.
> >> >>
> >> >> If you don't have a feasible plan to do the offload, p4tc does not make
> >> >> sense to me to be honest.
> >> >
> >> >You mean if there's no plan to match your (NVIDIA?)  point of view.
> >> >For #1 - how's this different from DDP? Wasnt that your suggestion to
> >>
> >> I doubt that. Any flashing-blob-parsing-in-kernel is something I'm
> >> opposed to from day 1.
> >>
> >>
> >
> >Oh well - it is in the kernel and it works fine tbh.
> >
> >> >begin with? For #2 Nobody is proposing to do anything of the sort. The
> >> >ndo is passed IDs for the objects and associated contents. For #3+#4
> >>
> >> During offload, you need to parse the blob in driver to be able to match
> >> the ids with blob entities. That was presented by you/Intel in the past
> >> IIRC.
> >>
> >
> >You are correct - in case of offload the netlink IDs will have to be
> >authenticated against what the hardware can accept, but the devlink
> >flash use i believe was from you as a compromise.
>
> Definitelly not. I'm against devlink abuse for this from day 1.
>
>
> >
> >>
> >> >tc flower thing has nothing to do with P4TC that was just some random
> >> >proposal someone made seeing if they could ride on top of P4TC.
> >>
> >> Yeah, it's not yet merged and already mentally used for abuse. I love
> >> that :)
> >>
> >> >
> >> >Besides this nobody really has to satisfy your point of view - like i
> >> >said earlier feel free to provide proprietary solutions. From a
> >> >consumer perspective  I would not want to deal with 4 different
> >> >vendors with 4 different proprietary approaches. The kernel is the
> >> >unifying part. You seemed happier with tc flower just not with the
> >>
> >> Yeah, that is my point, why the unifying part can't be a userspace
> >> daemon/library with multiple backends (p4tc, bpf, vendorX, vendorY, ..)?
> >>
> >> I just don't see the kernel as a good fit for abstraction here,
> >> given the fact that the vendor compilers does not run in kernel.
> >> That is breaking your model.
> >>
> >
> >Jiri - we want to support P4, first. Like you said the P4 pipeline,
> >once installed is static.
> >P4 doesnt allow dynamic update of the pipeline. For example, once you
> >say "here are my 14 tables and their associated actions and here's how
> >the pipeline main control (on how to iterate the tables etc) is going
> >to be" and after you instantiate/activate that pipeline, you dont go
> >back 5 minutes later and say "sorry, please introduce table 15, which
> >i want you to walk to after you visit table 3 if metadata foo is 5" or
> >"shoot, let's change that table 5 to be exact instead of LPM". It's
> >not anywhere in the spec.
> >That doesnt mean it is not useful thing to have - but it is an
> >invention that has _nothing to do with the P4 spec_; so saying a P4
> >implementation must support it is a bit out of scope and there are
> >vendors with hardware who support P4 today that dont need any of this.
>
> I'm not talking about the spec. I'm talking about the offload
> implemetation, the offload compiler the offload runtime manager. You
> don't have those in kernel. That is the issue. The runtime manager is
> the one to decide and reshuffle the hw internals. Again, this has
> nothing to do with p4 frontend. This is offload implementation.
>
> And that is why I believe your p4 kernel implementation is unoffloadable.
> And if it is unoffloadable, do we really need it? IDK.
>

Say what?
It's not offloadable in your hardware, you mean? Because i have beside
me here an intel e2000 which offloads just fine (and the AMD folks
seem fine too).
If your view is that all these runtime optimization surmount to a
compiler in the kernel/driver that is your, well, your view. In my
view (and others have said this to you already) the P4C compiler is
responsible for resource optimizations. The hardware supports P4, you
give it constraints and it knows what to do. At runtime, anything a
driver needs to do for resource optimization (resorting, reshuffling
etc), that is not a P4 problem - sorry if you have issues in your
architecture approach.

> >In my opinion that is a feature that could be added later out of
> >necessity (there is some good niche value in being able to add some
> >"dynamicism" to any pipeline) and influence the P4 standards on why it
> >is needed.
> >It should be doable today in a brute force way (this is just one
> >suggestion that came to me when Rice University/Nvidia presented[1]);
> >i am sure there are other approaches and the idea is by no means
> >proven.
> >
> >1) User space Creates/compiles/Adds/activate your program that has 14
> >tables at tc prio X chain Y
> >2) a) 5 minutes later user space decides it wants to change and add
> >table 3 after table 15, visited when metadata foo=5
> >    b) your compiler in user space compiles a brand new program which
> >satisfies #2a (how this program was authored is out of scope of
> >discussion)
> >    c) user space adds the new program at tc prio X+1 chain Y or another chain Z
> >    d) user space delete tc prio X chain Y (and make sure your packets
> >entry point is whatever #c is)
>
> I never suggested anything like what you describe. I'm not sure why you
> think so.

It's the same class of problems - the paper i pointed to (coauthored
by Matty and others) has runtime resource optimizations which are
tantamount to changing the nature of the pipeline. We may need to
profile in the kernel but all those optimizations can be derived in
user space using the approach I described.

cheers,
jamal


> >[1] https://www.cs.rice.edu/~eugeneng/papers/SIGCOMM23-Pipeleon.pdf
> >
> >>
> >> >kernel process - which is ironically the same thing we are going
> >> >through here ;->
> >> >
> >> >cheers,
> >> >jamal
> >> >
> >> >>
> >> >> >
> >> >> >cheers,
> >> >> >jamal

Powered by blists - more mailing lists