Re: [v4.17-rcx] Lost IBPB, IBRS_FW support for spectre_v2 mitigation.

From: Linus Torvalds
Date: Tue May 01 2018 - 16:14:54 EST


On Tue, May 1, 2018 at 5:59 AM Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:

> Then I really have no idea how reverting the patch you pointed out would
> fix it.

So I do think that the original patch is buggy.

What I think *may* be going on is:

- first we do that

get_cpu_cap(c);
get_cpu_address_sizes(c);

but at that point, CPU levels may be masked, and that 0x80000008 leaf
isn't seen

- then we do

if (this_cpu->c_early_init)
this_cpu->c_early_init(c);

which calls early_init_intel(), which does that

if (msr_clear_bit(MSR_IA32_MISC_ENABLE,
MSR_IA32_MISC_ENABLE_LIMIT_CPUID_BIT) >
0) {

which now raises the cpuid_level.

- then we do

get_cpu_cap(c);

again, because the cpuid level has been raised, and _now_ it used to get
that 0x80000008 leaf information.

But with the change, that second call to get_cpu_cap() didn't do anything,
because the 0x80000008 leaf handling had been moved away.

However, I agree that your patch to just do that CPUID_8000_0008_EBX in
get_cpu_cap() should have fixed it, and it's possible that JÃrg mis-tested
it.

JÃrg, are you sure you didn't somehow get the wrong microcode? Because
another way for those bits to be cleared again is if
bad_spectre_microcode() triggers. That should show up in dmesg as "Intel
Spectre v2 broken microcode detected" though.

Linus