Re: [RFC PATCH v2 4/4] arm64: Export id_aar64fpr0 via sysfs

[Will Deacon <will@kernel.org>]

On Wed, Oct 21, 2020 at 04:03:13PM +0100, Qais Yousef wrote:
> On 10/21/20 15:41, Will Deacon wrote:
> > On Wed, Oct 21, 2020 at 01:15:59PM +0100, Catalin Marinas wrote:
> > > On Wed, Oct 21, 2020 at 12:09:58PM +0100, Marc Zyngier wrote:
> > > > On 2020-10-21 11:46, Qais Yousef wrote:
> > > > > Example output. I was surprised that the 2nd field (bits[7:4]) is
> > > > > printed out
> > > > > although it's set as FTR_HIDDEN.
> > > > > 
> > > > > # cat /sys/devices/system/cpu/cpu*/regs/identification/id_aa64pfr0
> > > > > 0x0000000000000011
> > > > > 0x0000000000000011
> > > > > 0x0000000000000011
> > > > > 0x0000000000000011
> > > > > 0x0000000000000011
> > > > > 0x0000000000000011
> > > > > 
> > > > > # echo 1 > /proc/sys/kernel/enable_asym_32bit
> > > > > 
> > > > > # cat /sys/devices/system/cpu/cpu*/regs/identification/id_aa64pfr0
> > > > > 0x0000000000000011
> > > > > 0x0000000000000011
> > > > > 0x0000000000000012
> > > > > 0x0000000000000012
> > > > > 0x0000000000000011
> > > > > 0x0000000000000011
> > > > 
> > > > This looks like a terrible userspace interface. It exposes unrelated
> > > > features,
> > > 
> > > Not sure why the EL1 field ended up in here, that's not relevant to the
> > > user.
> > > 
> > > > and doesn't expose the single useful information that the kernel has:
> > > > the cpumask describing the CPUs supporting  AArch32 at EL0. Why not expose
> > > > this synthetic piece of information which requires very little effort from
> > > > userspace and doesn't spit out unrelated stuff?
> > > 
> > > I thought the whole idea is to try and avoid the "very little effort"
> > > part ;).
> > > 
> > > > Not to mention the discrepancy with what userspace gets while reading
> > > > the same register via the MRS emulation.
> > > > 
> > > > Granted, the cpumask doesn't fit the cpu*/regs/identification hierarchy,
> > > > but I don't think this fits either.
> > > 
> > > We already expose MIDR and REVIDR via the current sysfs interface. We
> > > can expand it to include _all_ the other ID_* regs currently available
> > > to user via the MRS emulation and we won't have to debate what a new
> > > interface would look like. The MRS emulation and the sysfs info should
> > > probably match, though that means we need to expose the
> > > ID_AA64PFR0_EL1.EL0 field which we currently don't.
> > > 
> > > I do agree that an AArch32 cpumask is an easier option both from the
> > > kernel implementation perspective and from the application usability
> > > one, though not as easy as automatic task placement by the scheduler (my
> > > first preference, followed by the id_* regs and the aarch32 mask, though
> > > not a strong preference for any).
> > 
> > If a cpumask is easier to implement and easier to use, then I think that's
> > what we should do. It's also then dead easy to disable if necessary by
> > just returning 0. The only alternative I would prefer is not having to
> > expose this information altogether, but I'm not sure that figuring this
> > out from MIDR/REVIDR alone is reliable.
> 
> I did suggest this before, but I'll try gain. If we want to assume a custom
> bootloader and custom user space, we can make them provide the mask.

Who mentioned a custom bootloader? In the context of Android, we're
talking about a user-space that already manages scheduling affinity.

> For example, the new sysctl_enable_asym_32bit could be a cpumask instead of
> a bool as it currently is. Or we can make it a cmdline parameter too.
> In both cases some admin (bootloader or init process) has to ensure to fill it
> correctly for the target platform. The bootloader should be able to read the
> registers to figure out the mask. So more weight to make it a cmdline param.

I think this is adding complexity for the sake of it. I'm much more in
favour of keeping the implementation and ABI as simple as possible: expose
the fact that the system is heterogenous, have an opt-in for userspace to
say it can handle that and let it handle it.

Will