Re: [PATCH] seccomp: fix populating a0-a5 syscall args in 32-bit x86 BPF

[Andy Lutomirski <luto@amacapital.net>]

On Mon, Apr 14, 2014 at 11:31 PM, Alexei Starovoitov <ast@plumgrid.com> wrote:
> On Mon, Apr 14, 2014 at 1:28 PM, Andy Lutomirski <luto@amacapital.net> wrote:
>> On Mon, Apr 14, 2014 at 1:24 PM, David Miller <davem@davemloft.net> wrote:
>>> From: Andy Lutomirski <luto@amacapital.net>
>>> Date: Mon, 14 Apr 2014 13:13:45 -0700
>>>
>>>> I think this description is wrong.  (unsigned long *) &sd->args[1] is
>>>> the right location, at least on 32-bit little-endian architectures.
>>>
>>> It absolutely is not.
>>
>> Huh?  It's a pointer to the right address, but the type is wrong.
>>
>> The changelog says "on 32-bit x86 (or any other 32bit arch), it would
>> result in storing a0-a5 at wrong offsets in args[] member".  Unless
>> I'm mistaken, this is incorrect: a0-a5 are are the correct offsets,
>> but they are stored with the wrong type, so the other bits in there
>> are garbage.
>
> agree. your above description is more correct than the log.
> We were focusing on the bug itself and the log came a bit misleading
> as a result of multiple iterations back and forth between me and Daniel.
>
> also the log says:
> "gcc is clever and optimizes the copy away in other cases, e.g. x86_64"
> since we actually checked assembler, so the fix doesn't pessimize
> 64-bit architectures :)
> This function is in critical path for seccomp, so performance definitely
> matters.

Yeah, I'm not entirely sure what I was thinking when I wrote that
part.  The new code should actually be much better than the old code
for weird architectures like ia-64.

For reference, ia-64 uses the unwinder (!) to look up arguments, so
the fewer times it gets invoked, the better.

--Andy