Re: [PATCH] infiniband: avoid overflow warning

[Arnd Bergmann]

On Mon, Jul 31, 2017 at 5:32 PM, Bart Van Assche <Bart.VanAssche@xxxxxxx> wrote:
> On Mon, 2017-07-31 at 09:30 +0200, Arnd Bergmann wrote:
>> On Mon, Jul 31, 2017 at 9:08 AM, Moni Shoua <monis@xxxxxxxxxxxx> wrote:
>> > On Mon, Jul 31, 2017 at 9:50 AM, Arnd Bergmann <arnd@xxxxxxxx> wrote:
>> > > --- a/include/rdma/ib_addr.h
>> > > +++ b/include/rdma/ib_addr.h
>> > > @@ -172,7 +172,8 @@ static inline int rdma_ip2gid(struct sockaddr *addr, union ib_gid *gid)
>> > >                                        (struct in6_addr *)gid);
>> > >                 break;
>> > >         case AF_INET6:
>> > > -               memcpy(gid->raw, &((struct sockaddr_in6 *)addr)->sin6_addr, 16);
>> > > +               *(struct in6_addr *)&gid->raw =
>> > > +                       ((struct sockaddr_in6 *)addr)->sin6_addr;
>> > >                 break;
>> > >         default:
>> > >                 return -EINVAL;
>> >
>> > what happens if you replace 16 with sizeof(struct in6_addr)?
>>
>> Same thing: the problem is that gcc already knows the size of the structure we
>> pass in here, and it is in fact shorter.
>>
>> I also tried changing the struct sockaddr pointer to a sockaddr_storage pointer,
>> without success. Other approaches that do work are:
>>
>> - mark addr_event() as "noinline" to prevent gcc from seeing the true
>> size of the
>>   inetaddr_event stack object in rdma_ip2gid(). I considered this a little ugly.
>>
>> - change inetaddr_event to put a larger structure on the stack, using
>>   sockaddr_storage or sockaddr_in6. This would be less efficient.
>>
>> - define a union of sockaddr_in and sockaddr_in6, and use that as the argument
>>   to rdma_ip2gid/rdma_gid2ip, and change all callers to use that union type.
>>   This is probably the cleanest approach as it gets rid of a lot of questionable
>>   type casts, but it's a relatively large patch and also slightly less
>> efficient as we have
>>   to zero more stack storage in some cases.
>
>
> So inetaddr_event() assigns AF_INET so .sin_family and gcc warns about code
> that is only executed if .sin_family == AF_INET6? Since this warning is the
> result of incorrect interprocedural analysis by gcc, shouldn't this be
> reported as a bug to the gcc authors?

I think the interprocedural analysis here is just a little worse than it could
be, but it's not actually correct.  It's not gcc that prints the warning (if
it did, then I'd agree it would be a gcc bug) but the warning is triggered
intentionally by the fortified version of memcpy in include/linux/string.h.

The problem as I understand it is that gcc cannot guarantee that it
tracks the value of addr->sa_family at  least as far as the size of the
stack object, and it has no strict reason to do so, so the inlined
rdma_ip2gid() will still contain both cases.

      Arnd