Re: [PATCH 3/4] rseq: Make rseq work with protection keys

[Mathieu Desnoyers <mathieu.desnoyers@efficios.com>]

On 2025-02-18 02:55, Dmitry Vyukov wrote:
> On Mon, 17 Feb 2025 at 21:21, Mathieu Desnoyers
[...]
>>
>>    we still let this function read the rseq_cs.
>>> +      * It's unclear what benefits the resticted code gets by doing this
>>
>> restricted
>>
>>> +      * (it probably already hijacked control flow at this point), and
>>> +      * presumably any sane sandbox should prohibit restricted code
>>> +      * from accessing struct rseq, and this is still better than
>>> +      * terminating the app unconditionally (it always has a choice
>>> +      * of not using rseq and pkeys together).
>>
>> Note that because userspace can complete an rseq critical section
>> without clearing the rseq_cs pointer, this could happen simply because
>> the kernel is preempting the task after it has:
>>
>> 1) completed an rseq critical section, without clearing rseq_cs,
>> 2) changed pkey.
>>
>> So allowing this is important, and I would remove the comment about
>> hijacked control flow and such. This can happen with normal use of the
>> ABI.
> 
> Thanks for the review!
> 
> I've addressed all comments in the series in v2.
> 
> I've reworded this paragraph to simplify sentences, but I still kept
> the note aboud malicious rseq_cs.
> 
> If we would not be circumventing normal protection, then, yes, these
> cases would be the same. But since we are circumventing protection
> that otherwise exists, I think it's important to think about
> potentially malicious cases. In this context inaccessible rseq_cs
> values that resulted from normal execution are very different from
> malicious onces. Normal ones will point to a fixed set of real
> well-formed rseq_cs objects, while malicious ones may point to
> god-knows-where in an attempt of an attacker to do things we can't
> even imagine right now (e.g. rseq_cs overlapping with protected crypto
> keys).
> 
> It's as if a particular instance of copy_to_user would allow
> user-space to write arbitrary kernel memory, and memory of other
> processes circumventing all normal protections. In that context we
> would need to be very careful regarding what we actually allow.

I'm considering that we should clear the rseq_cs pointer whenever
userspace issues pkey_mprotect.

This would ensure that no legitimate scenario can trigger a load
from a rseq_cs area which has the wrong pkey, and therefore we
could accept read/write from/to a struct rseq which has the wrong
pkey, but kill the process if trying to read/write from a
struct rseq_cs with the wrong key. This would prevent userspace
from making the kernel read/write from/to memory with the wrong
pkey through a pointer it controls (rseq_cs pointer).

Thoughts ?

Thanks,

Mathieu

-- 
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com