Re: [PATCH 2/6] arm64: Allow mismatched 32-bit EL0 support

[Will Deacon <will@kernel.org>]

On Wed, Nov 11, 2020 at 04:27:00PM +0000, Qais Yousef wrote:
> On 11/09/20 13:52, Will Deacon wrote:
> > On Fri, Nov 06, 2020 at 02:48:35PM +0000, Qais Yousef wrote:
> > > On 11/06/20 13:00, Will Deacon wrote:
> > > > On Fri, Nov 06, 2020 at 12:54:25PM +0000, Qais Yousef wrote:
> > > > > FWIW I have my v3 over here in case it's of any help. It solves the problem of
> > > > > HWCAP discovery when late AArch32 CPU is booted by populating boot_cpu_date
> > > > > with 32bit features then.
> > > > > 
> > > > > 	git clone https://git.gitlab.arm.com/linux-arm/linux-qy.git -b asym-aarch32-upstream-v3 origin/asym-aarch32-upstream-v3
> > > > 
> > > > Cheers, I've done something similar. I was hoping to post it today, but I've
> > > > been side-tracked with bug fixing this morning. The main headache I ended up
> > > > with was allowing late-onlining of 64-bit-only CPUs if all the boot CPUs
> > > > are 32-bit capable. What do you do in that case?
> > > 
> > > Do you mean if CPUs 0-3 were 32bit capable and we boot with maxcpus=4 then
> > > attempt to bring the remaining 64bit-only cpus online later?
> > 
> > Right. I think we will refuse to online them. I'll post my attempt at
> > handling that shortly.
> 
> Sorry for the delayed response.
> 
> You're right, I tried that and they refuse to come online. We missed that tbh.
> 
> Haven't thought what we should do yet. I tried your v2 and it failed similarly.

Hmm, it shouldn't do. Please could you provide the log? My hunch is that you
are blatting 32-bit EL1 support as well, and we can't handle a mismatch for
that with a late CPU. Do you know if the CPUs being integrated into these
broken designs have a mismatch at EL1 as well?

> I usually have a similar hunk in my testing to check how the kernel perceives
> the 32bit support when I execute a binary:
> 
> 	diff --git a/arch/arm64/include/asm/cpufeature.h b/arch/arm64/include/asm/cpufeature.h
> 	index f447d313a9c5..a9549e55a6c8 100644
> 	--- a/arch/arm64/include/asm/cpufeature.h
> 	+++ b/arch/arm64/include/asm/cpufeature.h
> 	@@ -611,6 +611,9 @@ static inline bool system_supports_32bit_el0(void)
> 	 {
> 		u64 pfr0 = read_sanitised_ftr_reg(SYS_ID_AA64PFR0_EL1);
> 
> 	+       pr_err("System supports symmetric 32bit el0: %d\n", id_aa64pfr0_32bit_el0(pfr0));
> 	+       pr_err("System supports Asymmetric 32bit el0: %ld\n", static_branch_unlikely(&arm64_mismatched_32bit_el0));
> 	+
> 		return id_aa64pfr0_32bit_el0(pfr0) ||
> 		       static_branch_unlikely(&arm64_mismatched_32bit_el0);
> 	 }
> 
> In your v2 both conditions are true. In my series we see the system as
> symmetric if we boot the 32bit capable cpus _only_.

The "arm64_mismatched_32bit_el0" key drives both the creation of the sysfs
file and the allocation of the cpu mask. See the comment in cpufeature.c
That file should be created whenever the command-line is passed to enable
this feature, because a late CPU could come up and set bits in there. The
presence of the file can therefore inform userspace that this can happen.

Will