Re: [PATCH 9/9] crypto: qce - switch to using a mutex

[Eric Biggers <ebiggers@kernel.org>]

On Thu, Feb 20, 2025 at 10:14:20AM +0100, Bartosz Golaszewski wrote:
> On Mon, Jan 20, 2025 at 2:46 PM Bartosz Golaszewski <brgl@bgdev.pl> wrote:
> >
> > On Sat, Jan 18, 2025 at 6:55 PM Eric Biggers <ebiggers@kernel.org> wrote:
> > >
> > > On Sat, Jan 18, 2025 at 10:28:26AM +0100, Bartosz Golaszewski wrote:
> > > > I was testing with kcapi-speed and cryptsetup benchmark. I've never
> > > > seen any errors.
> > > >
> > > > Is this after my changes only or did it exist before? You're testing
> > > > with the tcrypt module? How are you inserting it exactly? What params?
> > >
> > > Those are all benchmarks, not tests.  The tests run at registration time if you
> > > just enable the kconfig options for them:
> > >
> > >     # CONFIG_CRYPTO_MANAGER_DISABLE_TESTS is not set
> > >     CONFIG_CRYPTO_MANAGER_EXTRA_TESTS=y
> > >
> > > The test failures and KASAN error occur on mainline too, so yes they occur
> > > before your patchset too.
> > >
> > > > >
> > > > > I personally still struggle to understand how this driver could plausibly be
> > > > > useful when the software crypto has no issues, is much faster, and is much
> > > > > better tested.  What is motivating having this driver in the kernel?
> > > >
> > > > We want to use it in conjunction with the upcoming scminvoke (for
> > > > loading TAs and invoking objects - used to program the keys into the
> > > > QCE) to support the DRM use-case for decrypting streaming data inside
> > > > secure buffers upstream.
> > >
> > > Notably lacking is any claim that any of the current features of the driver are
> > > actually useful.
> > >
> >
> > Noted. I'm still quite early into working on the upstream-bound code
> > supporting the streaming use-case but I will consider a proposal to
> > remove existing features that are better provided by ARM CE.
> >
> > Thanks,
> > Bartosz
> 
> Just an FYI, I was informed by Qualcomm that upcoming platforms will
> contain an upgrade to this IP and it will be up to 3x faster than ARM
> CE.

I suspect that is measured under some ideal condition that won't be reached in
the real world, but we'll see.

- Eric