Re: [GIT PULL] Second set of RISC-V updates for v5.5-rc1

[Anup Patel <anup@brainfault.org>]

On Thu, Dec 5, 2019 at 1:20 AM Alistair Francis
<Alistair.Francis@wdc.com> wrote:
>
> On Wed, 2019-12-04 at 11:38 -0800, Paul Walmsley wrote:
> > Alistair, Anup,
> >
> > On Wed, 4 Dec 2019, Alistair Francis wrote:
> >
> > > On Wed, 2019-12-04 at 18:22 +0530, Anup Patel wrote:
> > >
> > > > I had commented on your patch but my comments are still
> > > > not addressed.
> > > >
> > > > Various debug options enabled by this patch have performance
> > > > impact. Instead of enabling these debug options in primary
> > > > defconfigs, I suggest to have separate debug defconfigs with
> > > > these options enabled.
> > >
> > > +1
> > >
> > > OE uses the defconfig (as I'm sure other distros do) and slowing
> > > down
> > > users seems like a bad idea.
> >
> > While I respect your points of view, our defconfigs are oriented
> > towards
> > kernel developers.  This is particularly important when right now the
> > only
>
> That is just not what happens though.
>
> It is too much to expect every distro to maintain a defconfig for RISC-
> V. There are constantly new features that need to be enabled/disabled
> in the configs and it isn't always clear to outsiders. Which is why we
> currently use the defconfig as a base and apply extra features that
> distro want on top.
>
> Expecting every distro to have a kernel developers level of knowledge
> about configuring Kconfigs is just unrealistic.
>
> > RISC-V hardware on the market are test chips.  Our expectation is
> > that
>
> Treating RISC-V as a test architecture seems like a good way to make
> sure that is all it ever is.
>
> > distros and benchmarkers will create their own Kconfigs for their
> > needs.
>
> Like I said, that isn't true. After this patch is applied (and it makes
> it to a release) all OE users will now have a slower RISC-V kernel.
> This also applies to buildroot and probably other distos.
>
> Now image some company wants to investigate using a RISC-V chip for
> their embedded project. They use OE/buildroot to build a quick test
> setup and boot Linux. It now runs significantly slower then some other
> architecture and they don't choose RISC-V.
>
> Slowing down all users to help kernel developers debug seems like the
> wrong direction. Kernel developers should know enough to be able to
> turn on the required configs, why does this need to be the default?

I quickly tried hackbench on SiFive Unleashed board with latest Linus
tree master branch (having your patch) and I am seeing 12% slowdown.

I am sure if I try more heavier benchmarks (such as stress-ng) then
the slowdown will be even more.

Here are the detailed numbers:

Command: ./hackbench 32
Number of Tasks: 32*40 (== 1280)
Average Time (without debug options): 3.10525
Average Time (with debug options): 3.471 (11.78% slower)

Command: ./hackbench 64
Number of Tasks: 64*40 (== 2560)
Average Time (without debug options): 6.3015
Average Time (with debug options): 7.05875 (12.017% slower)

Command: ./hackbench 128
Number of Tasks: 128*40 (== 5120)
Average Time (without debug options): 12.6275
Average Time (with debug options): 14.1455 (12.0214% slower)

It is this performance impact due to which other architectures (such as
x86 and ARM64) don't have these debug options enabled in their defconfigs.

I will send a patch to move these debug options to separate debug
defconfigs so that people have a way to build a debug kernel.

Regards,
Anup