Re: [RFC PATCH] asm/generic: introduce if_nospec and nospec_barrier
From: Dan Williams
Date: Thu Jan 04 2018 - 16:43:20 EST
On Thu, Jan 4, 2018 at 11:26 AM, Pavel Machek <pavel@xxxxxx> wrote:
> Hi!
>
>> >> > What remains to be seen is if there are other patterns that affect
>> >> > different processors.
>> >> >
>> >> > In the longer term the compiler itself needs to know what is and isn't
>> >> > safe (ie you need to be able to write things like
>> >> >
>> >> > void foo(tainted __user int *x)
>> >> >
>> >> > and have the compiler figure out what level of speculation it can do and
>> >> > (on processors with those features like IA64) when it can and can't do
>> >> > various kinds of non-trapping loads.
>> >> >
>> >>
>> >> It would be great if coccinelle and/or smatch could be taught to catch
>> >> some of these case at least as a first pass "please audit this code
>> >> block" type of notification.
>> >>
>> >
>> > What should one be looking for. Do you have a typical example?
>> >
>>
>> See "Exploiting Conditional Branch Misprediction" from the paper [1].
>>
>> The typical example is an attacker controlled index used to trigger a
>> dependent read near a branch. Where an example of "near" from the
>> paper is "up to 188 simple instructions inserted in the source code
>> between the âifâ statement and the line accessing array...".
>>
>> if (attacker_controlled_index < bound)
>> val = array[attacker_controlled_index];
>> else
>> return error;
>>
>> ...when the cpu speculates that the 'index < bound' branch is taken it
>> reads index and uses that value to read array[index]. The result of an
>> 'array' relative read is potentially observable in the cache.
>
> You still need
>
> (void) array2[val];
>
> after that to get something observable, right?
As far as I understand the presence of array2[val] discloses more
information, but in terms of the cpu taking an action that it is
observable in the cache that's already occurred when "val =
array[attacker_controlled_index];" is speculated. Lets err on the side
of caution and shut down all the observable actions that are already
explicitly gated by an input validation check. In other words, a low
bandwidth information leak is still a leak.