On Mon, 2022-11-07 at 19:37 +0100, Borislav Petkov wrote: > On Mon, Nov 07, 2022 at 06:19:48PM +0000, Edgecombe, Rick P wrote: > > It was to catch if the software user shadow stack feature gets > > disabled > > at boot with the "clearcpuid" command. > > I don't understand. clearcpuid does setup_clear_cpu_cap() too. It > would > eventually clear the bit in boot_cpu_data.x86_capability's AFAICT. > > cpu_feature_enabled() looks at boot_cpu_data too. > > So what's the problem? You are right, there actually is no problem. I thought the apply_forced_caps() was happening too late, but it is not. So this check you highlighted would not be needed if we kept the clearcpuid method. Thanks. > > Oh, and also, you've added that clearcpuid thing to the help docs. > Please remove it. clearcpuid= taints the kernel and we've left it in > because some of your colleagues really wanted it for testing or > whatnot. > But it is crap and it was on its way out at some point so we better > not > proliferate its use any more. > > > Is there a better way to do this? > > Yeah, cpu_feature_enabled() should be enough and if it isn't, then we > need to fix it to be. > > Which reminds me, I'd need to take Maxim's patch too: > > https://lore.kernel.org/r/20220718141123.136106-3-mlevitsk@xxxxxxxxxx > > as it is a simplification. > > > > Here you need to do > > > > > > setup_clear_cpu_cap(X86_FEATURE_IBT); > > > setup_clear_cpu_cap(X86_FEATURE_SHSTK); > > > > This only gets called by kexec way after boot, as kexec is prepping > > to > > transition to the new kernel. Do we want to be clearing feature > > bits at > > that time? > > Hmm, I was under the impression you'll have the usual chicken bit > "noshstk" which gets added with every big feature. So it'll call that > thing here. There was at one point, but there was a suggestion to remove in favor of clearcpuid. I can add it back.
