[PATCH v2 01/11] arm64: use RET instruction for exiting the trampoline

[Ard Biesheuvel]

On 8 January 2018 at 14:45, Will Deacon <will.deacon at arm.com> wrote:
> On Mon, Jan 08, 2018 at 02:38:00PM +0000, Ard Biesheuvel wrote:
>> On 8 January 2018 at 14:33, Will Deacon <will.deacon at arm.com> wrote:
>> > On Sat, Jan 06, 2018 at 01:13:23PM +0000, Ard Biesheuvel wrote:
>> >> On 5 January 2018 at 13:12, Will Deacon <will.deacon at arm.com> wrote:
>> >> > Speculation attacks against the entry trampoline can potentially resteer
>> >> > the speculative instruction stream through the indirect branch and into
>> >> > arbitrary gadgets within the kernel.
>> >> >
>> >> > This patch defends against these attacks by forcing a misprediction
>> >> > through the return stack: a dummy BL instruction loads an entry into
>> >> > the stack, so that the predicted program flow of the subsequent RET
>> >> > instruction is to a branch-to-self instruction which is finally resolved
>> >> > as a branch to the kernel vectors with speculation suppressed.
>> >> >
>> >>
>> >> How safe is it to assume that every microarchitecture will behave as
>> >> expected here? Wouldn't it be safer in general not to rely on a memory
>> >> load for x30 in the first place? (see below) Or may the speculative
>> >> execution still branch anywhere even if the branch target is
>> >> guaranteed to be known by that time?
>> >
>> > The main problem with this approach is that EL0 can read out the text and
>> > find the kaslr offset.
>>
>> Not really - the CONFIG_RANDOMIZE_BASE path puts the movz/movk
>> sequence in the next page, but that does involve an unconditional
>> branch.
>
> Ah sorry, I had missed that. The unconditional branch may still be attacked,
> however.
>

Yeah, I was surprised by that. How on earth is there ever a point to
using a branch predictor to [potentially mis]predict unconditional
branches.

>> > The memory load is fine, because the data page is
>> > unmapped along with the kernel text. I'm not aware of any
>> > micro-architectures where this patch doesn't do what we need.
>> >
>>
>> Well, the memory load is what may incur the delay, creating the window
>> for speculative execution of the indirect branch. What I don't have
>> enough of a handle on is whether this speculative execution may still
>> branch to wherever the branch predictor is pointing even if the
>> register containing the branch target is already available.
>
> For the micro-architectures I'm aware of, the return stack predictor will
> always safely mispredict the jump into the kernel vectors with this patch
> applied.
>

OK, fair enough. What I am asking is really whether there is a way
where we don't have to force a misprediction, by ensuring that x30 has
assumed its final value by the time the indirect branch is
[speculatively] executed. But if unconditional branches may be
mispredicted as well, I guess this doesn't fly either.