[PATCH v2 1/2] eal: provide leading and trailing zero bit count abstraction

[Morten Brørup]

> From: Tyler Retzlaff [mailto:roretzla at linux.microsoft.com]
> Sent: Thursday, 5 January 2023 23.06
> 
> On Thu, Jan 05, 2023 at 10:34:55PM +0100, Morten Brørup wrote:
> > > From: Tyler Retzlaff [mailto:roretzla at linux.microsoft.com]
> > > Sent: Thursday, 5 January 2023 21.58
> > >
> > > On Thu, Jan 05, 2023 at 09:27:12AM -0800, Tyler Retzlaff wrote:
> > > > On Thu, Jan 05, 2023 at 09:23:49AM -0800, Tyler Retzlaff wrote:
> > > > > > > oh! not a problem. i'm very keen to catch any mistakes,
> thought
> > > i had
> > > > > > > missed something.
> > > > > >
> > > > > > I think we should move all bit-related functions together.
> > > > > > Please could you add another patch to your series
> > > > > > moving "ms1b"/"bsf"/"fls" functions in this file?
> > > > > >
> > > > > >
> > > > >
> > > > > okay, so there is already a rte_bitops.h. i guess everything
> should
> > > go
> > > > > there including the leading/trailing count functions instead of
> > > adding a
> > > > > new header.
> > > > >
> > > > > i'll introduce a new patch to the series that gathers the
> existing
> > > > > functions into rte_bitops.h and place the new functions there
> too.
> > > > >
> > > > > thanks
> > > >
> > > > just as a further follow up, you do understand that this is
> > > technically
> > > > an api break?
> > > >
> > > > moving functions from rte_common.h to rte_bitops.h will make
> > > translation
> > > > units that included only rte_common.h but used these functions
> will
> > > > fail to compile without being updated to include rte_bitops.h.
> > > >
> > > > anyway, i'll submit v3 with this change anyway.
> > >
> > > so when attempting to do this it became immediately obvious that
> moving
> > > just the bit op functions out is going to create a circular
> dependency
> > > between rte_common.h, rte_bitops.h
> > >
> > > once the bit ops are moved out of common there are still other
> inline
> > > functions that remain in comman that require bringing bitops back
> in,
> > > but bitops depends on common.
> > >
> > > my compromise will be to break log2 and pow2 inline functions into
> > > their
> > > own files to break the cycle (common no longer depends on bitops).
> i'll
> > > submit patches for this but it ends up touching a lot more of the
> > > tree to add back includes for log/pow inline use.
> > >
> > > alternatively i can just not move the remaining bit manipulation
> > > functions, let me know which is preferred.
> >
> > It seems that no perfect solution exists, so we will have to live
> with a compromise. Here is another proposal for a compromise, for yours
> and Thomas's consideration:
> >
> > I noticed that rte_bitops.h is mainly for setting/getting bits, used
> for accessing hardware.
> >
> > Your functions are mathematical functions, and so are the similar
> functions in rte_common.h (which is why it makes sense to keep them
> together with yours). If we cannot clean up rte_common.h by moving them
> out, perhaps we should accept the current situation (until we find a
> way to move them out) and just add your mathematical functions where
> the existing mathematical functions reside, i.e. in rte_common.h.
> >
> > This proposal only makes the existing mess slightly larger; it
> doesn't create a new kind of mess.
> 
> so i fudged around a bit to see if i could get a happy medium. i ended
> up with this.
> 
> remove include of rte_debug.h from rte_bitops.h
> 
>   * had to remove the RTE_ASSERT from existing rte_bitops.h functions
>   * this breaks a good piece of the cycle debug -> log -> common ->
> bitops -> debug
>   * deal breaker? i don't think it was right that we were getting all
>     of log, common just for using bitops anyway.
> 
> move pow2 functions from rte_common.h -> rte_pow2ops.h
>   * new header includes rte_bitops.h
> 
> move log2 functions from rte_common.h -> rte_log2ops.h
>   * new header includes rte_bitops.h, rte_pow2ops.h
> 
> include rte_bitops.h, rte_pow2ops.h and rte_log2ops.h back into
> rte_common.h
> 
>   * this is done to reduce the impact of compatibility break by
>     continuing to expose the pow2/log2/bitops via rte_common.h
> 
> so we end up with 3 standalone headers, where the whole tree builds
> without having to add a pile of includes for the new headers. we can
> later deprecate the exposure of the inline functions when including
> rte_common.h
> 
>   * one caveat is that there was some contamination coming in via the
>     removed rte_debug.h where rte_bitops.h was used. so technically
>     a break of api too.
> 
> objections?
> 
> if this is no good i'll just fold my new functions into rte_common.h
> and
> leave the mess for the next person, though i am trying not to do that.
> 
> thanks for the discussion.

Here's some long term thinking: EAL has grown into a trashcan where too much is thrown in. It should only be a thin shim to abstract the underlying hardware and O/S environment. A step in that direction could be splitting the current EAL into a true EAL and a Utils library. Not now, but perhaps some day in a rosy future.

Your proposal effectively makes rte_common.h even bigger by including rte_bitops.h (which was intended for accessing hardware). I am not sure it is a step in the right direction. On the other hand, introducing yet another header file for bit-mathematical functions is probably worse than adding them to rte_bitops.h.

I can't come up with something good myself, but I lean towards simply adding your functions to rte_common.h and live with the existing mess. If you think your proposal is better, I will not object. I'm only voicing my thoughts.

@Thomas may have another perspective on the matter.

Thanks for all the work you put into this.

-Morten