[PATCH net-next] net: bpf: arm: make hole-faulting more robust

[dborkman@redhat.com (Daniel Borkmann)]

On 09/19/2014 01:11 AM, Russell King - ARM Linux wrote:
> On Fri, Sep 19, 2014 at 12:57:03AM +0200, Daniel Borkmann wrote:
>> Will Deacon pointed out, that the currently used opcode for filling holes,
>> that is 0xe7ffffff, seems not robust enough ...
>
> If you're after a single 32-bit word which will fault if executed in
> ARM or Thumb mode, and you only want it to raise an undefined
> instruction exception (iow, you're not using it as a breakpoint or
> similar), then may I suggest the poison value I chose for the vectors
> page, designed to trap userspace branches to locations in there?
>
> 0xe7fddef1
>
>> Similarly, ptrace, kprobes, kgdb, bug and uprobes make use of such instruction
>> as well to trap. Given mentioned section from the specification, we can find
>> such a universe as (where 'x' denotes 'don't care'):
>>
>>    ARM:    xxxx 0111 1111 xxxx xxxx xxxx 1111 xxxx
>>    Thumb:  1101 1110 xxxx xxxx
>
> You'll notice that the value conforms to the ARM undefined instruction
> space.  You'll also notice that the low 16 bits correspond to the
> Thumb case.  The only question is, what is 0xe7fd as a Thumb instruction...
>
> 00000000 <a>:
>     0:   def1                            ; <UNDEFINED> instruction: 0xdef1
>     2:   e7fd            b.n     0 <a>
>
> So, if either 0 or 2 gets branched to, we end up at the Thumb UDF
> instruction.  (Sorry, my binutils doesn't know about UDF.)

Yes, that should keep the code even simpler! Will try that out tomorrow
and respin the patch.

Thanks Russell!