On Mon, 16 Dec 2024 15:07:15 +0000, Mark Rutland mark.rutland@arm.com wrote:
On Mon, Dec 16, 2024 at 02:31:43PM +0000, Marc Zyngier wrote:
On Mon, 16 Dec 2024 12:38:17 +0000, Mark Rutland mark.rutland@arm.com wrote:
I think that what we did in commit:
892f7237b3ff ("arm64: Delay initialisation of cpuinfo_arm64::reg_{zcr,smcr}")
... introduces an anti-pattern that'd be nice to avoid. That broke the existing split of __cpuinfo_store_cpu() and init_cpu_features(), where the former read the ID regs, and the latter set up the features *without* altering the copy of the ID regs that was read. i.e. init_cpu_features() shouldn't write to its info argument at all.
I understand that we have to do something as a bodge for broken FW which traps SME, but I'd much rather we did that within __cpuinfo_store_cpu().
Honestly, I'd rather revert that patch, together with b3000e2133d8 ("arm64: Add the arm64.nosme command line option"). I'm getting tired of the FW nonsense, and we are only allowing vendors to ship untested crap.
Furthermore, given the state of SME in the kernel, I don't think this is makes any difference. So maybe this is the right time to reset everything to a sane state.
Looking again, a revert does look to be the best option.
We removed reg_zcr and reg_smcr in v6.7 in commits:
abef0695f9665c3d ("arm64/sve: Remove ZCR pseudo register from cpufeature code") 391208485c3ad50f ("arm64/sve: Remove SMCR pseudo register from cpufeature code")
As of those commits, ZCR and SCMR no longer matter to __cpuinfo_store_cpu(), and only SMIDR_EL1 remains...
Per ARM DDI 0487 L.a, accesses to SMIDR_EL1 never trap to EL3, so we can read that safely as long as ID_AA64PFR1_EL1.SME indicates that SME is implemented.
Which is to say that if we revert the remaining portion of 892f7237b3ff and restore the read of SMIDR, that should be good as far back as v6.7, which sounds good to me.
Sounds reasonable indeed. I guess *someone* will want it for the previous kernel versions, but they can have fun with the backport on their own.
M.