Re: [PATCH] rust: RFC/demo of safe API for Dpdk Eal, Eth and Rxq

[Owen Hilyard <Owen.Hilyard@unh.edu>]

[-- Attachment #1: Type: text/plain, Size: 15732 bytes --]

From: Van Haaren, Harry <harry.van.haaren@intel.com>
Sent: Friday, May 2, 2025 9:58 AM
To: Etelson, Gregory <getelson@nvidia.com>; Richardson, Bruce <bruce.richardson@intel.com>
Cc: dev@dpdk.org <dev@dpdk.org>; Owen Hilyard <owen.hilyard@unh.edu>
Subject: Re: [PATCH] rust: RFC/demo of safe API for Dpdk Eal, Eth and Rxq

> From: Etelson, Gregory
> Sent: Friday, May 02, 2025 1:46 PM
> To: Richardson, Bruce
> Cc: Gregory Etelson; Van Haaren, Harry; dev@dpdk.org; owen.hilyard@unh.edu
> Subject: Re: [PATCH] rust: RFC/demo of safe API for Dpdk Eal, Eth and Rxq
>
> Hello Bruce,

Hi All,
Hi All,

> > Thanks for sharing. However, IMHO using EAL for thread management in rust
> > is the wrong interface to expose.
>
> EAL is a singleton object in DPDK architecture.
> I see it as a hub for other resources.

Yep, i tend to agree here; EAL is central to the rest of DPDK working correctly.
And given EALs implementation is heavily relying on global static variables, it is
certainly a "singleton" instance, yes.
I think a singleton one way to implement this, but then you lose some of the RAII/automatic resource management behavior. It would, however, make some APIs inherently unsafe or very unergonomic unless we were to force rte_eal_cleanup to be run via atexit(3) or the platform equivalent and forbid the user from running it themselves. For a lot of Rust runtimes similar to the EAL (tokio, glommio, etc), once you spawn a runtime it's around until process exit. The other option is to have a handle which represents the state of the EAL on the Rust side and runs rte_eal_init on creation and rte_eal_cleanup on destruction. There are two ways we can make that safe. First, reference counting, once the handles are created, they can be passed around easily, and the last one runs rte_eal_cleanup when it gets dropped.  This avoids having tons of complicated lifetimes and I think that, everywhere that it shouldn't affect fast path performance, we should use refcounting. The other option is to use lifetimes. This is doable, but is going to force people who are more likely to primarily be C or C++ developers to dive deep into Rust's type system if they want to build abstractions over it. If we add async into the mix, as many people are going to want to do, it's going to become much, much harder. As a result, I'd advocate for only using it for data path components where refcounting isn't an option.

> Following that idea, the EAL structure can be divided to hold the
> "original" resources inherited from librte_eal and new resources
> introduced in Rust EAL.

Here we can look from different perspectives. Should "Rust EAL" even exist?
If so, why? The DPDK C APIs were designed in baremetal/linux days, where
certain "best-practices" didn't exist yet, and Rust language was pre 1.0 release.

Of course, certain parts of Rust API must depend on EAL being initialized.
There is a logical flow to DPDK initialization, these must be kept for correct functionality.

I guess I'm saying, perhaps we can do better than mirroring the concept of
"DPDK EAL in C" in to "DPDK EAL in Rust".

I think that there will need to be some kind of runtime exposed by the library. A lot of the existing EAL abstractions may need to be reworked, especially those dealing with memory, but I think a lot of things can be layered on top of the C API. However, I think many of the invariants in the EAL could be enforced at compile time for free, which may mean the creation of a lot of "unchecked" function variants which skip over null checks and other validation.

As was mentioned before, it may also make sense for some abstractions in the C EAL to be lifted to compile time. I've spent a lot of time thinking about how to use something like Rust's traits for "it just works" capabilities where you can declare what features you want (ex: scatter/gather) and it will either be done in hardware or fall back to software, since you were going to need to do it anyway. This might lead to parameterizing a lot of user code on the devices they expect to interact with and then having some "dyn EthDev" as a fallback, which should be roughly equivalent to what we have now. I can explain that in more detail if there's interest.

> > Instead, I believe we should be
> > encouraging native rust thread management, and not exposing any DPDK
> > threading APIs except those necessary to have rust threads work with DPDK,
> > i.e. with an lcore ID. Many years ago when DPDK started, and in the C
> > world, having DPDK as a runtime environment made sense, but times have
> > changed and for Rust, there is a whole ecosystem out there already that we
> > need to "play nice with", so having Rust (not DPDK) do all thread
> > management is the way to go (again IMHO).
> >
>
> I'm not sure what exposed DPDK API you refer to.

I think that's the point :) Perhaps the Rust application should decide how/when to
create threads, and how to schedule & pin them. Not the "DPDK crate for Rust".
To give a more concrete examples, lets look at Tokio (or Monoio, or Glommio, or .. )
which are prominent players in the Rust ecosystem, particularly for networking workloads
where request/response patterns are well served by the "async" programming model (e.g HTTP server).
Rust doesn't really care about threads that much. Yes, it has std::thread as a pthread equivalent, but on Linux those literally call pthread. Enforcing the correctness of the Send and Sync traits (responsible for helping enforce thread safety) in APIs is left to library authors. I've used Rust with EAL threads and it's fine, although a slightly nicer API for launching based on a closure (which is a function pointer and a struct with the captured inputs) would be nice. In Rust, I'd say that async and threads are orthogonal concepts, except where runtimes force them to mix. Async is a way to write a state machine or (with some more abstraction) an execution graph, and Rust the language doesn't care whether a library decides to run some dependencies in parallel. What I think Rust is more likely to want is thread per core and then running either a single async runtime over all of them or an async runtime per core.

Lets focus on Tokio first: it is an "async runtime" (two links for future readers)
    <snip>
So an async runtime can run "async" Rust functions (called Futures, or Tasks when run independently..)
There are lots of words/concepts, but I'll focus only on the thread creation/control aspect, given the DPDK EAL lcore context.

Tokio is a work-stealing scheduler. It spawns "worker" threads, and then gives these "tasks"
to various worker cores (similar to how Golang does its work-stealing scheduling). Some
DPDK crate users might like this type of workflow, where e.g. RXQ polling is a task, and the
"tokio runtime" figures out which worker to run it on. "Spawning" a task causes the "Future"
to start executing. (technical Rust note: notice the "Send" bound on Future: https://docs.rs/tokio/latest/tokio/task/fn.spawn.html )
The work stealing aspect of Tokio has also led to some issues in the Rust ecosystem. What it effectively means is that every "await" is a place where you might get moved to another thread. This means that it would be unsound to, for example, have a queue handle on devices without MT-safe queues unless we want to put a mutex on top of all of the device queues. I personally think this is a lot of the source of people thinking that Rust async is hard, because Tokio forces you to be thread safe at really weird places in your code and has issues like not being able to hold a mutex over an await point.

Other users might prefer the "thread-per-core" and CPU pinning approach (like DPDK itself would do).
nit: Tokio also spawns a thread per core, it just freely moves tasks between cores. It doesn't pin because it's designed to interoperate with the normal kernel scheduler more nicely. I think that not needing pinned cores is nice, but we want the ability to pin for performance reasons, especially on NUMA/NUCA systems (NUCA = Non-Uniform Cache Architecture, almost every AMD EPYC above 8 cores, higher core count Intel Xeons for 3 generations, etc).
Monoio and Glommio both serve these use cases (but in slightly different ways!). They both spawn threads and do CPU pinning.
Monoio and Glommio say "tasks will always remain on the local thread". In Rust techie terms: "Futures are !Send and !Sync"
    https://docs.rs/monoio/latest/monoio/fn.spawn.html
    https://docs.rs/glommio/latest/glommio/fn.spawn_local.html
There is also another option, one which would eliminate "service cores". We provide both a work stealing pool of tasks that have to deal with being yanked between cores/EAL threads at any time, but aren't data plane tasks, and then a different API for spawning tasks onto the local thread/core for data plane tasks (ex: something to manage a particular HTTP connection). This might make writing the runtime harder, but it should provide the best of both worlds provided we can build in a feature (Rust provides a way to "ifdef out" code via features) to disable one or the other if someone doesn't want the overhead.

So there are at least 3 different async runtimes (and I haven't even talked about async-std, smol, embassy, ...) which
all have different use-cases, and methods of running "tasks" on threads. These runtimes exist, and are widely used,
and applications make use of their thread-scheduling capabilities.

So "async runtimes" do thread creation (and optionally CPU pinning) for the user.
Other libraries like "Rayon" are thread-pool managers, those also have various CPU thread-create/pinning capabilities.
If DPDK *also* wants to do thread creation/management and CPU-thread-to-core pinning for the user, that creates tension.
The other problem is that most of these async runtimes have IO very tightly integrated into them. A large portion of Tokio had to be forked and rewritten for io_uring support, and DPDK is a rather stark departure from what they were all designed for. I know that both Tokio and Glommio have "start a new async runtime on this thread" functions, and I think that Tokio has an "add this thread to a multithreaded runtime" somewhere.

I think the main thing that DPDK would need to be concerned about is that many of these runtimes use thread locals, and I'm not sure if that would be transparently handled by the EAL thread runtime since I've always used thread per core and then used the Rust runtime to multiplex between tasks instead of spawning more EAL threads.

Rayon should probably be thought of in a similar vein to OpenMP, since it's mainly designed for batch processing. Unless someone is doing some fairly heavy computation (the kind where "do we want a GPU to accelerate this?" becomes a question) inside of their DPDK application, I'm having trouble thinking of a use case that would want both DPDK and Rayon.

> Bruce wrote: "so having Rust (not DPDK) do all thread management is the way to go (again IMHO)."

I think I agree here, in order to make the Rust DPDK crate usable from the Rust ecosystem,
it must align itself with the existing Rust networking ecosystem.

That means, the DPDK Rust crate should not FORCE the usage of lcore pinnings and mappings.
Allowing a Rust application to decide how to best handle threading (via Rayon, Tokio, Monoio, etc)
will allow much more "native" or "ergonomic" integration of DPDK into Rust applications.
I'm not sure that using DPDK from Rust will be possible without either serious performance sacrifices or rewrites of a lot of the networking libraries. Tokio continues to mimic the BSD sockets API for IO, even with the io_uring version, as does glommio. The idea of the "recv" giving you a buffer without you passing one in isn't really used outside of some lower-level io_uring crates. At a bare minimum, even if DPDK managed to offer an API that works exactly the same ways as io_uring or epoll, we would still need to go to all of the async runtimes and get them to plumb DPDK support in or approve someone from the DPDK community maintaining support. If we don't offer that API, then we either need rewrites inside of the async runtimes or for individual libraries to provide DPDK support, which is going to be even more difficult.

I agree that forcing lcore pinnings and mappings isn't good, but I think that DPDK is well within its rights to build its own async runtime which exposes a standard API. For one thing, the first thing Rust users will ask for is a TCP stack, which the community has been discussing and debating for a long time. I think we should figure out whether the goal is to allow DPDK applications to be written in Rust, or to allow generic Rust applications to use DPDK. The former means that the audience would likely be Rust-fluent people who would have used DPDK regardless, and are fine dealing with mempools, mbufs, the eal, and ethdev configuration. The latter is a much larger audience who is likely going to be less tolerant of dpdk-rs exposing the true complexity of using DPDK. Yes, Rust can help make the abstractions better, but there's an amount of inherent complexity in "Your NIC can handle IPSec for you and can also direct all IPv6 traffic to one core" that I don't think we can remove.

I personally think that making an API for DPDK applications to be written in Rust, and then steadily adding abstractions on top of that until we arrive at something that someone who has never looked at a TCP header can use without too much confusion. That was part of the goal of the Iris project I pitched (and then had to go finish another project so the design is still WIP). I think that a move to DPDK is going to be as radical of a change as a move to io_uring, however, DPDK is fast enough that I think it may be possible to convince people to do a rewrite once we arrive at that high level API. "Swap out your sockets and rework the functions that do network IO for a 5x performance increase" is a very, very attractive offer, but for us to get there I think we need to have DPDK's full potential available in Rust, and then build as many zero-overhead (zero cost or you couldn't write it better yourself) abstractions as we can on top. I want to avoid a situation where we build up to the high-level APIs as fast as we can and then end up in a situation where you have "Easy Mode" and then "C DPDK written in Rust" as your two options.
> Regards,
> Gregory

Apologies for the long-form, "wall of text" email, but I hope it captures the nuance of threading and
async runtimes, which I believe in the long term will be very nice to capture "async offload" use-cases
for DPDK. To put it another way, lookaside processing can be hidden behind async functions & runtimes,
if we design the APIs right: and that would be really cool for making async-offload code easy to write correctly!

Regards, -Harry

Sorry for my own walls of text. As a consequence of working on Iris I've spent a lot of time thinking about how to make DPDK easier to use while keeping the performance intact, and I was already thinking in Rust since it provides one of the better options for these kinds of abstractions (the other option I see is Mojo, which isn't ready yet). I want to see DPDK become more accessible, but the performance and access to hardware is one of the main things that make DPDK special, so I don't want to compromise that. I definitely agree that we need to force DPDK's existing APIs to justify themselves in the face of the new capabilities of Rust, but I think that starting from "How are Rust applications written today?" is a mistake.

Regards,
Owen

[-- Attachment #2: Type: text/html, Size: 36913 bytes --]