On Fri, Nov 13, 2020 at 09:36:30AM +0000, Will Deacon wrote:

On Wed, Nov 11, 2020 at 07:10:44PM +0000, Catalin Marinas wrote:

quoted

On Mon, Nov 09, 2020 at 09:30:18PM +0000, Will Deacon wrote:

quoted

+static bool has_32bit_el0(const struct arm64_cpu_capabilities *entry, int scope)
+{
+	if (!has_cpuid_feature(entry, scope))
+		return allow_mismatched_32bit_el0;

I still don't like overriding the cpufeature mechanism in this way. What about
something like below? It still doesn't fit perfectly but at least the
capability represents what was detected in the system. We then decide in
system_supports_32bit_el0() whether to allow asymmetry. There is an
extra trick to park a non-AArch32 capable CPU in has_32bit_el0() if it
comes up late and the feature has already been advertised with
!allow_mismatched_32bit_el0.

I deliberately allow late onlining of 64-bit-only cores and I don't think
this is something we should forbid

I agree and my patch allows this. That's a property of
WEAK_LOCAL_CPUFEATURE.

(although it's not clear from your patch when
allow_mismatched_32bit_el0 gets set).

It doesn't, that was meant as a discussion point around the cpufeature
framework but still relying on your other patches for cpumask, cmdline
argument.

Furthermore, killing CPUs from the matches callback feels _very_ dodgy
to me, as it's invoked indirectly by things such as
this_cpu_has_cap().

Yeah, this part is not nice. What we want here is for a late 64-bit only
CPU to be parked if !allow_mismatched_32bit_el0 and we detected 32-bit
already (symmetric).

quoted

I find it clearer, though I probably stared at it more than at your
patch ;).

Yeah, swings and roundabouts...

I think we're quibbling on implementation details a bit here whereas we
should probably be focussing on what to do about execve() and CPU hotplug.

Do we need to solve the execve() problem? If yes, we have to get the
scheduler involved. The hotplug case I think is simpler, just make sure
we have a last standing 32-bit capable CPU, no ABI changes.

For execve(), arguably we don't necessarily break the execve() ABI as
the affinity change may be done later when scheduling the task. But
given that it's a user opt-in for this feature anyway, we just document
the ABI changes.

Your patch doesn't apply on top of my series or replace this one, so there's
not an awful lot I can do with it. I'm about to post a v3 with a tentative
solution for execve(), so please could you demonstrate your idea on top of
that so I can see how it fits together?

I'll give it a go and if it looks any nicer, I'll post something. As you
say, it's more like personal preference on the implementation details.

On the functional aspects, we must preserve the current behaviour like
detecting symmetry and parking late CPUs if they don't comply in the
!allow_mismatched_32bit_el0 case. In the allow_mismatched_32bit_el0
case, we relax this to the equivalent of a weak feature (either missing
or present for early/late CPUs) but only report 32-bit if we found an
early 32-bit capable CPU (I ended up with a truth table on a piece of
paper).

-- 
Catalin

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel