Re: [PATCH v2] x86: disable non-instrumented version of copy_mc when KMSAN is enabled

[Thomas Gleixner]

On Tue, Mar 05 2024 at 07:21, Dave Hansen wrote:
> On 3/1/24 14:52, Tetsuo Handa wrote:
>> -	if (static_cpu_has(X86_FEATURE_ERMS)) {
>> +	if (!IS_ENABLED(CONFIG_KMSAN) && static_cpu_has(X86_FEATURE_ERMS)) {
>>  		__uaccess_begin();
>>  		ret = copy_mc_enhanced_fast_string((__force void *)dst, src, len);
>>  		__uaccess_end();
>
> Where does the false positive _come_ from?  Can we fix copy_mc_fragile()
> and copy_mc_enhanced_fast_string() instead of just not using them?

All it takes is a variant of __msan_memcpy() which uses a variant of
copy_mc_to_kernel() instead of __memcpy(). It's not rocket science.

Aside of that, this:

@@ -74,14 +74,14 @@ unsigned long __must_check copy_mc_to_user(void __user *dst, const void *src, un
 {
	unsigned long ret;

-	if (copy_mc_fragile_enabled) {
+	if (!IS_ENABLED(CONFIG_KMSAN) && copy_mc_fragile_enabled) {
		__uaccess_begin();

is completely bogus. copy_user_generic() is not at all covered by
KMSAN. So why fiddling with it in the first place? Just because it has
the same pattern as copy_mc_to_kernel()?

> The three enable_copy_mc_fragile() are presumably doing so for a
> reason.

Very much so. It's for MCE recovery purposes.

And yes, the changelog and the non-existing comments should explain why
this is "correct" when KMSAN is enabled. Hint: It is NOT.

Thanks,

        tglx