RE: The EAL is bloated

["Morten Brørup" <mb@smartsharesystems.com>]

> From: Bruce Richardson [mailto:bruce.richardson@intel.com]
> Sent: Friday, 26 August 2022 12.46
> 
> On Fri, Aug 26, 2022 at 10:58:15AM +0200, Morten Brørup wrote:
> > Dear all,
> >
> > <rant> The "Environment Abstraction Layer" is expanding far beyond
> its
> > purpose...
> >
> > It not only includes abstractions for the underlying CPU Arch and
> O/S,
> > but also a bunch of generic utility functions. In an ideal world,
> these
> > belong in a Utility library; but I can live with them staying in the
> EAL
> > library.
> >
> > However, since the Utility features are also considered part of the
> EAL
> > library, some features get misclassified as Utilities and thus sneak
> into
> > the EAL library, regardless that they are completely independent of
> the
> > underlying CPU Arch and O/S. E.g.: Service Cores, Trace, and soon the
> > Lcore Poll Busyness library.
> >
> > The EAL is not a catch-all library, and we should not allow the EAL
> to
> > grow like this!  </rant>
> >
> > If this misbehavior doesn't stop naturally, I propose that adding any
> new
> > feature to the EAL requires techboard approval.
> >
> I don't disagree with you that it is indeed becoming ever bigger, and
> that
> we need to do something to do some cleanup on EAL. However, IMHO this
> is
> not a simple problem to fix or even to draft up a solution for. I
> actually
> did some prototyping work in the recent past to try and see if or how
> much
> the EAL could be split up to make it more modular. On the plus side,
> some
> things like logging, for example, could be fairly easily pulled out of
> it
> and put into a separate library. On the other hand, really splitting
> things
> up beyond pulling out a few easy things was a massive undertaking - at
> least in my tests.
> 
> I also think trying to classify contents between abstractions and
> utilities
> is overly simplistic. To my mind we also need to have a category for
> DPDK
> initialization code, which is a lot of what complicates things - and
> may
> well be the cause of a lot of the "scope creep" in EAL.

For the initialization code, it should make it easier that DPDK is no longer compiled to a bunch of independent libraries (.so files) with independent versions, but compiled into one big library.

Regarding libraries that are part of DPDK, e.g. Logging, I don't mind calling their required init functions from the various relevant locations inside the EAL. I assume that the various libraries need to be initialized at different stages throughout EAL, so a single point of "library init callout" in the EAL is not sufficient.

Some of the DPDK Core libraries, e.g. Mempool and Ring, are outside the EAL; this shows that they don't have to be part of EAL, although a DPDK application without them would be pretty silly.

Long term, the EAL could also include some hooks or similar, to allow external libraries to be initialized from EAL. The DPDK application startup sequence is well documented [1]; perhaps something can be added to this "power-on sequence" (borrowing a term from the hardware world). For inspiration, C loadable libraries expose an _init() function, which is called when the library is loaded. I don't have a solution to this; just thinking out aloud. The libraries included in DPDK could also use such a generic initialization hooking mechanism, if present.

[1] https://doc.dpdk.org/guides/prog_guide/env_abstraction_layer.html#eal-in-a-linux-userland-execution-environment

> 
> Given the scope of the problem - and the fact that splitting EAL has
> been
> discussed before and nothing came of it in the community - I'm not sure
> of
> the best approach here. Maybe we can start by splitting out what we can
> of
> the easy stuff, and work iteratively from there. Alternatively if
> someone
> has time for a big-bang rework of EAL, that would be great too.

I wish! But I think small steps are more realistic.

Splitting out the easy parts would be a good start. And it would certainly raise awareness against EAL scope creep.

> 
> Regards,
> /Bruce