Message-ID: <CAM0EoMnmEJEPmCVaBQZkWRT95CkKyzYd0g3vpETDqyRj7KHMcw@mail.gmail.com>
Date: Mon, 20 Nov 2023 17:13:05 -0500
From: Jamal Hadi Salim <jhs@...atatu.com>
To: John Fastabend <john.fastabend@...il.com>
Cc: Jiri Pirko <jiri@...nulli.us>, 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, daniel@...earbox.net, 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 Mon, Nov 20, 2023 at 3:41 PM John Fastabend <john.fastabend@...il.com> wrote:
>
> 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:
> > > >On Mon, Nov 20, 2023 at 4:39 AM Jiri Pirko <jiri@...nulli.us> wrote:
> > > >>
> > > >> Fri, Nov 17, 2023 at 09:46:11PM CET, jhs@...atatu.com wrote:
> > > >> >On Fri, Nov 17, 2023 at 1:37 PM John Fastabend <john.fastabend@...il.com> wrote:
> > > >> >>
> > > >> >> Jamal Hadi Salim wrote:
> > > >> >> > On Fri, Nov 17, 2023 at 1:27 AM John Fastabend <john.fastabend@...il.com> wrote:
> > > >> >> > >
> > > >> >> > > Jamal Hadi Salim wrote:
> > > >>
> > > >> [...]
> > > >>
> > > >>
> > > >> >>
> > > >> >> I think I'm judging the technical work here. Bullet points.
> > > >> >>
> > > >> >> 1. p4c-tc implementation looks like it should be slower than a
> > > >> >>    in terms of pkts/sec than a bpf implementation. Meaning
> > > >> >>    I suspect pipeline and objects laid out like this will lose
> > > >> >>    to a BPF program with an parser and single lookup. The p4c-ebpf
> > > >> >>    compiler should look to create optimized EBPF code not some
> > > >> >>    emulated switch topology.
> > > >> >>
> > > >> >
> > > >> >The parser is ebpf based. The other objects which require control
> > > >> >plane interaction are not - those interact via netlink.
> > > >> >We published perf data a while back - presented at the P4 workshop
> > > >> >back in April (was in the cover letter)
> > > >> >https://github.com/p4tc-dev/docs/blob/main/p4-conference-2023/2023P4WorkshopP4TC.pdf
> > > >> >But do note: the correct abstraction is the first priority.
> > > >> >Optimization is something we can teach the compiler over time. But
> > > >> >even with the minimalist code generation you can see that our approach
> > > >> >always beats ebpf in LPM and ternary. The other ones I am pretty sure
> > > >>
> > > >> Any idea why? Perhaps the existing eBPF maps are not that suitable for
> > > >> this kinds of lookups? I mean in theory, eBPF should be always faster.
> > > >
> > > >We didnt look closely; however, that is not the point - the point is
> > > >the perf difference if there is one, is not big with the big win being
> > > >proper P4 abstraction. For LPM for sure our algorithmic approach is
> > > >better. For ternary the compute intensity in looping is better done in
> > > >C. And for exact i believe that ebpf uses better hashing.
> > > >Again, that is not the point we were trying to validate in those experiments..
>
> If you compared your implementation to the bpf lpm_trie its a bit
> misleading. The data structure is a rhashtable vs a Trie doing LPM.
>
> Also I can't see how __p4tc_table_entry_lookup() is going to scale?
> That looks like a bucket per key? If so that wont scale well with
> 1000's of entries and lots of duplicate masks.

I think you are misreading the code - there are no duplicate masks;
iiuc, by scale you mean performance of lookup and the numbers we got
show very different results (the more entries and masks the better
numbers we showed).
Again - i dont want to make this a topic. The issue is not whether we
beat you or you beat us in numbers is not relevant to begin with.

>I did a quick scan
> of code, but would be nice to detail the algorithm in the commit
> msg so we can disect it.
>
> This doesn't look what we would want though for an LPM unless
> I've dropped this out of context.
>
> +static struct p4tc_table_entry *
> +__p4tc_entry_lookup(struct p4tc_table *table, struct p4tc_table_entry_key *key)
> +       __must_hold(RCU)
> +{
> +       struct p4tc_table_entry *entry = NULL;
> +       struct rhlist_head *tmp, *bucket_list;
> +       struct p4tc_table_entry *entry_curr;
> +       u32 smallest_prio = U32_MAX;
> +
> +       bucket_list =
> +               rhltable_lookup(&table->tbl_entries, key, entry_hlt_params);
> +       if (!bucket_list)
> +               return NULL;
> +
> +       rhl_for_each_entry_rcu(entry_curr, tmp, bucket_list, ht_node) {
> +               struct p4tc_table_entry_value *value =
> +                       p4tc_table_entry_value(entry_curr);
> +               if (value->prio <= smallest_prio) {
> +                       smallest_prio = value->prio;
> +                       entry = entry_curr;
> +               }
> +       }
> +
> +       return entry;
> +}

You are quoting the ternary (not LPM) matching code. It iterates all
entries (we could only do ~190 when we tested in plain ebpf, thats why
our test was restricted to that number).

> Also I don't know what 'better done in C' matters the TCAM data structure
> can be written in C and used as a BPF map. At least that is how we would
> normally approach it from BPF side.

See the code you quoted - you have to loop and pick the best of N
matches, where N could be arbitrarily large.

> > > >
> > > >On your point of "maps are not that suitable" P4 tables tend to have
> > > >very specific attributes (examples associated meters, counters,
> > > >default hit and miss actions, etc).
>
> The typical way we handle this from BPF is to either use the 0 entry
> for stats, annotations, etc. or create a blob of memory (another map,
> variables, global struct, ...) and stash the info there. If we care
> about performance we make those per cpu and deal with it in user
> land.
>

Back to the abstraction overhead in user space being high. The whole
point is to minimize all that..

> > > >
> > > >> >we can optimize over time.
> > > >> >Your view of "single lookup" is true for simple programs but if you
> > > >> >have 10 tables trying to model a 5G function then it doesnt make sense
> > > >> >(and i think the data we published was clear that you gain no
> > > >> >advantage using ebpf - as a matter of fact there was no perf
> > > >> >difference between XDP and tc in such cases).
> > > >> >
> > > >> >> 2. p4c-tc control plan looks slower than a directly mmaped bpf
> > > >> >>    map. Doing a simple update vs a netlink msg. The argument
> > > >> >>    that BPF can't do CRUD (which we had offlist) seems incorrect
> > > >> >>    to me. Correct me if I'm wrong with details about why.
> > > >> >>
> > > >> >
> > > >> >So let me see....
> > > >> >you want me to replace netlink and all its features and rewrite it
> > > >> >using the ebpf system calls? Congestion control, event handling,
> > > >> >arbitrary message crafting, etc and the years of work that went into
> > > >> >netlink? NO to the HELL.
> > > >>
> > > >> Wait, I don't think John suggests anything like that. He just suggests
> > > >> to have the tables as eBPF maps.
> > > >
> > > >What's the difference? Unless maps can do netlink.
> > > >
>
> I'm going to argue map update time matters and we should use the fastest
> updates possible. If it complicates user space side some I would prefer
> that to slow updates. I don't think you can get much faster than a
> mmaped block of memory. Or even syscall updates are probably faster than
> netlink msgs.

So lets put this to rest:
It's about the P4 abstraction first (as i mentioned earlier) - i am
sure mmaping would be faster, but that is secondary - correct
abstraction first.
I am ok with some level of abstraction wrangling (example match-action
in P4 to match-value in ebpf) but there is a limit.

> > > >> Honestly, I don't understand the
> > > >> fixation on netlink. Its socket messaging, memcpies, processing
> > > >> overhead, etc can't keep up with mmaped memory access at scale. Measure
> > > >> that and I bet you'll get drastically different results.
> > > >>
> > > >> I mean, netlink is good for a lot of things, but does not mean it is an
> > > >> universal answer to userspace<->kernel data passing.
> > > >
> > > >Here's a small sample of our requirements that are satisfied by
> > > >netlink for P4 object hierarchy[1]:
> > > >1. Msg construction/parsing
> > > >2. Multi-user request/response messaging
> > >
> > > What is actually a usecase for having multiple users program p4 pipeline
> > > in parallel?
> >
> > First of all - this is Linux, multiple users is a way of life, you
> > shouldnt have to ask that question unless you are trying to be
> > socratic. Meaning multiple control plane apps can be allowed to
> > program different parts and even different tables - think multi-tier
> > pipeline.
>
> Linux is always been opinionated and rejects code all the time because
> its not the "right" way. I've been on the reject your stuff side before.
>
> Partitioning ownershiip of the pipeline is different than multiple
> users of the same elements. From BPF side (to show its doable) is
> done by pinning maps to files and giving that file to different
> programs. The DDOS thing can own the DDOS map and the router can own
> its router tables. BPF handles this using the file systems mostly.
>

And with tc it just fits right in without any of those tricks...

> >
> > > >3. Multi-user event subscribe/publish messaging
> > >
> > > Same here. What is the usecase for multiple users receiving p4 events?
> >
> > Same thing.
> > Note: Events are really not part of P4 but we added them for
> > flexibility - and as you well know they are useful.
>
> Per above I wouldn't sacrafice update perf for this. Also its doable
> from userspace if you need to. Other thing I've come to dislike a bit
> is teaching the kernel a specific DSL. P4 is my favorte, but still
> going so far as to encode a specific P4 spec into the kernel seems
> unnecessary. Also now will we have to have kernel X supports P4.16 and
> kernel X+N supports P4.18 it seems like a pain.
>

I believe you are misunderstanding, let me explain. While our focus is on PNA:
There is no change in the kernel infra (upstream code) for PNA or PSA
or PXXXX. The compiler may end up generating different code depending
on the architecture selected at the compile command line. The control
constructs are very static with their hierarchy IDs.
In regards to what you prophesize above when the language goes from
P4.16 to P4.18 - i dont mean to be rude, but: kettle look at pot
much?;-> i.e what happens when eBPF ISA gets extended? The safety
feature we have for P4TC is externs - most of these will be
implemented as self-fulfilling kfuncs.

> >
> > >
> > > >
> > > >I dont think i need to provide an explanation on the differences here
> > > >visavis what ebpf system calls provide vs what netlink provides and
> > > >how netlink is a clear fit. If it is not clear i can give more
> > >
> > > It is not :/
> >
> > I thought it was obvious for someone like you, but fine - here goes for those 3:
> >
> > 1. Msg construction/parsing: A lot of infra for sending attributes
> > back and forth is already built into netlink. I would have to create
> > mine from scratch for ebpf.  This will include not just the
> > construction/parsing but all the detailed attribute content policy
> > validations(even in the presence of hierarchies) that comes with it.
> > And not to forget the state transform between kernel and user space.
>
> But the series here does that as well probably could reuse that on
> top of BPF. We have lots of libraries to deal with ebpf to help.
> I don't see anything problematic here for BPF.

Which library does all these (netlink features) in eBPF and has
something matching it in the kernel? We did try to write our own but
it was a huge waste of time.

> >
> > 2. Multi-user request/response messaging
> > If you can write all the code for #1 above then this should work fine for ebpf
> >
> > 3. Event publish subscribe
> > You would have to create mechanisms for ebpf which either are non
> > trivial or non complete: Example 1: you can put surgeries in the ebpf
> > code to look at map manipulations and then interface it to some event
> > management scheme which checks for subscribed users. Example 2: It may
> > also be feasible to create your own map for subscription vs something
> > like perf ring for event publication(something i have done in the
> > past), but that is also limited in many ways.
>
> I would just push them out over a single perf ring and build the
> subscription on top of GRPC (pick your protocol of choice).
>


Why - just so i could use ebpf? Ive never understood that single user
mode perf ring thing.

> >
> > >
> > > >breakdown. And of course there's more but above is a good sample.
> > > >
> > > >The part that is taken for granted is the control plane code and
> > > >interaction which is an extremely important detail. P4 Abstraction
> > > >requires hierarchies with different compiler generated encoded path
> > > >ids etc. This ID mapping gets exacerbated by having multitudes of  P4
> > >
> > > Why the actual eBFP mapping does not serve the same purpose as ID?
> > > ID:mapping 1 :1
> >
> > An identification of an object requires hierarchical IDs: A
> > pipeline/program ID, A table id, a table entry Identification, an
> > action identification and for each individual action content
> > parameter, an ID etc. These same IDs would be what hardware would
> > recognize as well (in case of offload).  Given the dynamic nature of
> > these IDs it is essentially up to the compiler to define them. These
> > hierarchies  are much easier to validate in netlink.
>
> I'm on board for offloads, but this series says no offloads and we
> have no one with hardware in Linux for offloads yet. If we have a
> series with a P4 driver and NIC I can get my hands on now we have
> an entirely different conversation.

The argument made is that P4 s/w stands on its own merit regardless of
presence of hardware offload (there are s/w versions of DPDK and rust
that I believe are used in production). As an example, the DASH
project quoted in the cover letter uses P4 as a datapath specification
language. The datapath is then verified to be working in s/w. So let's
not argue that there is no merit to a s/w P4 version without h/w
offload.

I do have a NIC(Intel e2000) that does P4 offloads but i am afraid I
can't give it to you. Folks who are doing offloads will present
drivers when they are ready and when/if those patches show there will
be extensions to deal with ndo_tc. But i know you already know and are
on the p4tc mailing list and are quite aware of these developments - I
am not sure I understand your motivation for bringing this up a few
times now. I read it as some sort of insinuation that there is some
secret vendor hardware that is going to benefit from all this secret
trojan we are doing here. Again P4 s/w stands on its own.

> None of this above is a problem in eBPF. Its just mapping ids around.
>
> >
> > We dont want to be constrained to a generic infra like eBPF for these
> > objects. Again eBPF is a means to an end (and not the goal here!).
>
> I don't see any constraints from eBPF above just a list of things
> that of course you would have to code up. But none of that doesn't
> already exist in other projects.
>

And we can agree to disagree.

cheers,
jamal

> >
> > cheers,
> > jamal
> > >
> > >
> > > >programs which have different requirements. Netlink is a natural fit
> > > >for this P4 abstraction. Not to mention the netlink/tc path (and in
> > > >particular the ID mapping) provides a conduit for offload when that is
> > > >needed.
> > > >eBPF is just a tool - and the objects are intended to be generic - and
> > > >i dont see how any of this could be achieved without retooling to make
> > > >it more specific to P4.
> > > >
> > > >cheers,
> > > >jamal
> > > >
> > > >
> > > >
> > > >>
> > > >> >I should note: that there was an interesting talk at netdevconf 0x17
> > > >> >where the speaker showed the challenges of dealing with ebpf on "day
> > > >> >two" - slides or videos are not up yet, but link is:
> > > >> >https://netdevconf.info/0x17/sessions/talk/is-scaling-ebpf-easy-yet-a-small-step-to-one-server-but-giant-leap-to-distributed-network.html
> > > >> >The point the speaker was making is it's always easy to whip an ebpf
> > > >> >program that can slice and dice packets and maybe even flush LEDs but
> > > >> >the real work and challenge is in the control plane. I agree with the
> > > >> >speaker based on my experiences. This discussion of replacing netlink
> > > >> >with ebpf system calls is absolutely a non-starter. Let's just end the
> > > >> >discussion and agree to disagree if you are going to keep insisting on
> > > >> >that.
> > > >>
> > > >>
> > > >> [...]
>
>

Powered by blists - more mailing lists