On Sun, Nov 20, 2016 at 08:22:25AM -0800, Andy Lutomirski wrote: > This makes me nervous: don't some CPUs actually need the cpuid > instruction when patching alternatives?
Nope, we use boot_cpu_has() in apply_alternatives() if that is what you mean. > And with this approach, we won't have the cpuid instruction there > until after patching. We will have set (or not) the X86_FEATURE_CPUID bit at early_identify_cpu() time. Looking at the code, we do call sync_core() pretty early. :-\ Hmm, that #ifdef CONFIG_M486 is there for a reason. > Why not change this function entirely: > > write_cr2(0); > > CR2 should be available on all 32-bit CPUs. It clobbers fewer > registers. Yap, just one. And not only that - it clobbers a register which gcc doesn't have to reload at all. > More usefully, CPUID causes an exit when running under > most hypervisors, and that's quite slow. The only case I can think of > where CPUID should be faster than MOV to CR2 is on Xen PV before Ivy > Bridge, and I'm not sure I care about performance there. > > (On Xen PV, it will do a hypercall instead, but the hypercall should > be good enough to serialize, too.) A nop hypercall or whatever... But yeah, pending a nod from hw people, this one sounds nice too. You can do it basically on every CPU which supports paging. And that should be all we support in Linux anyway. > Or we could do it dynamically: > > bt $X86_FEATURE_CPUID, CPU_FLAGS(boot_cpu_data) # or whatever -- I > think we need to add an asm offset > jnc 1f # here's our jump > cpuid > 1: We could... we did move the X86_FEATURE* things to a separate header so that they can be used in asm too. write_cr2(0) doesn't sound so bad either. Except what happens if someone decides to sync_core() before the first line of do_page_fault() executes... I know, it is unlikely but we do unlikely things :) But yeah, the write_cr2() sounds better if one considers the lower register pressure. Which is nice. -- Regards/Gruss, Boris. Good mailing practices for 400: avoid top-posting and trim the reply.