On Tue, Jan 9, 2018 at 1:41 PM, Josh Poimboeuf <jpoim...@redhat.com> wrote:
> On Fri, Jan 05, 2018 at 06:52:07PM -0800, Linus Torvalds wrote:
>> On Fri, Jan 5, 2018 at 5:10 PM, Dan Williams <dan.j.willi...@intel.com>
>> wrote:
>> > From: Andi Kleen <a...@linux.intel.com>
>> >
>> > When access_ok fails we should always stop speculating.
>> > Add the required barriers to the x86 access_ok macro.
>>
>> Honestly, this seems completely bogus.
>>
>> The description is pure garbage afaik.
>>
>> The fact is, we have to stop speculating when access_ok() does *not*
>> fail - because that's when we'll actually do the access. And it's that
>> access that needs to be non-speculative.
>>
>> That actually seems to be what the code does (it stops speculation
>> when __range_not_ok() returns false, but access_ok() is
>> !__range_not_ok()). But the explanation is crap, and dangerous.
>
> The description also seems to be missing the "why", as it's not
> self-evident (to me, at least).
>
> Isn't this (access_ok/uaccess_begin/ASM_STAC) too early for the lfence?
>
> i.e., wouldn't the pattern be:
>
> get_user(uval, uptr);
> if (uval < array_size) {
> lfence();
> foo = a2[a1[uval] * 256];
> }
>
> Shouldn't the lfence come much later, *after* reading the variable and
> comparing it and branching accordingly?
The goal of putting the lfence in uaccess_begin() is to prevent
speculation past access_ok(). You are correct that the cpu could later
mis-speculate on uval, that's where taint analysis tooling needs to
come into play to track uval to where it is used. That's where the
nospec_array_ptr() patches come into play.
