On Mon, Dec 16, 2024 at 12:17:54PM +0000, Mark Brown wrote: > On Sat, Dec 14, 2024 at 10:56:13AM +0000, Marc Zyngier wrote: > > > I don't understand the need to single out SMIDR_EL1. It seems to only > > make things even more fragile than they already are by adding more > > synchronisation phases. > > > Why isn't the following a good enough fix? It makes it plain that > > boot_cpu_data is only a copy of CPU0's initial boot state. > > That would work but it's not clear to me that that is what the intent is > here. The current ordering seemed like a strange enough decision to be > deliberate, though I couldn't identify the reasoning. The original intent was that __cpuinfo_store_cpu() read *all* of a CPU's implemented ID regs, and init_cpu_features() initialised the expected system features based on the boot CPU's ID regs. The expectation was that init_cpu_features() would only consume the register values, and would not alter the cpuinfo_arm64 values, so the order of: boot_cpu_data = *info; init_cpu_features(&boot_cpu_data); ... didn't matter either way, and using '&boot_cpu_data' was intended to make it clear that the features were based on the boot CPU's info, even if you just grepped for that and didn't see the surrounding context. I think the real fix here is to move the reading back into __cpuinfo_store_cpu(), but to have an explicit check that SME has been disabled on the commandline, with a comment explaining that this is a bodge for broken FW which traps the SME ID regs. Mark.
