On Fri, Nov 27, 2020 at 01:41:22PM +0000, Qais Yousef wrote:
> On 11/24/20 15:50, Will Deacon wrote:
> > If we want to support 32-bit applications, then when we identify a CPU
> > with mismatched 32-bit EL0 support we must ensure that we will always
> > have an active 32-bit CPU available to us from then on. This is important
> > for the scheduler, because is_cpu_allowed() will be constrained to 32-bit
> > CPUs for compat tasks and forced migration due to a hotplug event will
> > hang if no 32-bit CPUs are available.
> >
> > On detecting a mismatch, prevent offlining of either the mismatching CPU
> > if it is 32-bit capable, or find the first active 32-bit capable CPU
> > otherwise.
> ^^^^^
>
> You use cpumask_any_and(). Better use cpumask_first_and()? We have a truly
> random function now, cpumask_any_and_distribute(), if you'd like to pick
> something 'truly' random.
I think cpumask_any_and() is better, because it makes it clear that I don't
care about which CPU is chosen (and under the hood it ends up calling
cpumask_first_and() _anyway_). So this is purely cosmetic.
> > diff --git a/arch/arm64/kernel/cpufeature.c b/arch/arm64/kernel/cpufeature.c
> > index 29017cbb6c8e..fe470683b43e 100644
> > --- a/arch/arm64/kernel/cpufeature.c
> > +++ b/arch/arm64/kernel/cpufeature.c
> > @@ -1237,6 +1237,8 @@ has_cpuid_feature(const struct arm64_cpu_capabilities *entry, int scope)
> >
> > static int enable_mismatched_32bit_el0(unsigned int cpu)
> > {
> > + static int lucky_winner = -1;
> > +
> > struct cpuinfo_arm64 *info = &per_cpu(cpu_data, cpu);
> > bool cpu_32bit = id_aa64pfr0_32bit_el0(info->reg_id_aa64pfr0);
> >
> > @@ -1245,6 +1247,22 @@ static int enable_mismatched_32bit_el0(unsigned int cpu)
> > static_branch_enable_cpuslocked(&arm64_mismatched_32bit_el0);
> > }
> >
> > + if (cpumask_test_cpu(0, cpu_32bit_el0_mask) == cpu_32bit)
> > + return 0;
>
> Hmm I'm struggling to get what you're doing here. You're treating CPU0 (the
> boot CPU) specially here, but I don't get why?
If our ability to execute 32-bit code is the same as the boot CPU then we
don't have to do anything. That way, we can postpone nominating the lucky
winner until we really need to.
> > + if (lucky_winner >= 0)
> > + return 0;
> > +
> > + /*
> > + * We've detected a mismatch. We need to keep one of our CPUs with
> > + * 32-bit EL0 online so that is_cpu_allowed() doesn't end up rejecting
> > + * every CPU in the system for a 32-bit task.
> > + */
> > + lucky_winner = cpu_32bit ? cpu : cpumask_any_and(cpu_32bit_el0_mask,
> > + cpu_active_mask);
>
> cpumask_any_and() could return an error. It could be hard or even impossible to
> trigger, but better check if lucky_winner is not >= nr_cpu_ids before calling
> get_cpu_device(lucky_winner) to stay in the safe side and avoid a potential
> splat?
I don't see how it can return an error here. There are two cases to
consider:
1. The CPU being brought online is the first 32-bit-capable CPU. In which
case, we don't use cpumask_any_and() at all.
2. The CPU being brought online is the first 64-bit-only CPU. In which
case, the CPU doing the onlining is 32-bit capable and will be in
the active mask.
> We can do better by the way and do smarter check in remove_cpu() to block
> offlining the last aarch32 capable CPU without 'hardcoding' a specific cpu. But
> won't insist and happy to wait for someone to come complaining this is not good
> enough first.
I couldn't find a satisfactory way to do this without the possibility of
subtle races, so I'd prefer to keep it simple for the moment. In particular,
I wanted to make sure that somebody iterating over the cpu_possible_mask
and calling is_cpu_allowed(p, cpu) for each CPU and a 32-bit task can not
reach the end of the mask without ever getting a value of 'true'.
I'm open to revisiting this once some of this is merged, but right now
I don't think it's needed and it certainly adds complexity.
> Some vendors play games with hotplug to help with saving power. They might want
> to dynamically nominate the last man standing 32bit capable CPU. Again, we can
> wait for someone to complain first I guess.
The reality is that either all "big" cores or all "little" cores will be the
ones that are 32-bit capable, so I doubt it matters an awful lot which one
of the cluster is left online from a PM perspective. The real problem is
that a core has to be left online at all, but I don't think we can avoid
that.
Will
