Re: [PATCH 1/4] glibc: Perform rseq(2) registration at C startup and thread creation (v7)

[Florian Weimer]

* Carlos O'Donell:

> On 4/2/19 3:08 AM, Florian Weimer wrote:
>> * Michael Ellerman:
>>
>>> I'm a bit vague on what we're trying to do here.
>>>
>>> But it seems like you want some sort of "eye catcher" prior to the branch?
>>>
>>> That value is a valid instruction on current CPUs (rlwimi.
>>> r5,r24,6,1,9), and even if it wasn't it could become one in future.
>>>
>>> If you change it to 0x8053530 that is both a valid instruction and is a
>>> nop (conditional trap immediate but with no conditions set).
>>
>> I think we need something that is very unlikely to appear in the
>> instruction stream.  It's just a marker.  The instruction will never be
>> executed, and it does not have to be a trap, either (I believe that a
>> standard trap instruction would be a bad choice).
>
> I assume you want to avoid a standard trap instruction because it would
> be common, and so not meet the intent of the RSEQ_SIG choice as being something
> that is *uncommon* right?

Ideally, RSEQ_SIG would be something that does not show up in the
instruction stream at all, so that it is a reliable marker for the start
of an rseq handler.  I assume the intent here is that the kernel
provides some validation on the program counter it reads from the rseq
area, so that we do not end up with some easily-abused gadget in every
process image.

> It is valuable that it be a trap, particularly for constant pools because
> it means that a jump into the constant pool will trap.

Sorry, I don't understand why this matters in this context.  Would you
please elaborate?

Thanks,
Florian