GCC, unaligned access and UB in the Linux kernel

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]


Hi everyone,

I noticed that in the Linux kernel we have a define called
CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS that's used in a fair amount of places
with the following purpose: when it's set, unaligned access is supported by the
CPU so we can do it directly, otherwise fall-back to some logic where a byte at
the time is read/written. For example, check the implementation of

That's nice and seems to work today as expected, but there's one problem:
unaligned access is UB according to the ISO C standard, no matter if the
architecture supports it or not. Also, GCC doesn't have any option similar to "-
fno-strict-aliasing" to make unaligned access non-UB. At the moment, they won't
consider introducing such an option either. Check this bug:


Starting from GCC 8.x, the compiler introduced some new optimizations that
assume correct alignment which can break some code[1]. However, unaligned access
like the following [from do_strncpy_from_user()]:

    *(unsigned long *)(dst+res) = c;

Still generate correct instructions because:

  1. There's no aliasing involved [1]
  2. SIMD instructions are not allowed in the kernel [2]

But that doesn't mean at all that things won't change in the future. At any
point, some optimization in a newer compiler might generate incorrect code even
for the above-mentioned example. Therefore, while I understand compiler
engineers' point of view (they provide a compiler with an ISO-compliant
behaviour), I'm concerned about the correctness of all the code that assumes
unaligned access (on some architectures) in C is just fine. 

In my understanding, the simplest way to force GCC to emit a single MOV
instruction with unaligned access, without risking any kind of UB, is to use
__builtin_memcpy(). It works great, but it requires fixing the code in many
places. Also, maybe using get_unaligned()/put_unaligned() is the right thing to
do? The problem is that the put_unaligned_* inline functions don't use
__builtin_memcpy() and are defined like:

   static __always_inline void put_unaligned_le32(u32 val, void *p)
   	*((__le32 *)p) = cpu_to_le32(val);

So, still UB. To make the compiler happy, maybe we should make them use

What do you think about all of this? I realize that this is not a big urgent
problem *right now*, but at some point it might become one. How do you believe
this problem should be addressed in Linux? 


[1] If aliasing is involved, even with -fno-strict-aliasing, unaligned access
WILL break some code, today. Check the following example:

   int h(int *p, int *q){
     *p = 1;
     *q = 1;
     return *p;

   typedef __attribute__((__may_alias__)) int I;

   I k(I *p, I *q){
     *p = 1;
     *q = 1;
     return *p;

Starting from GCC 8.1, both h() and k() will always return 1, when compiled with
-O2, even with -fno-strict-aliasing.

[2] Some SIMD instructions have alignment requirements that recent compilers
might just start to assume to be true, in my current understanding. In general,
SIMD instructions can be emitted automatically by the compiler because of auto-
vectorization. But, fortunately, that *cannot* happen in the kernel because we
build with -fno-mmx, -fno-sse, -fno-avx etc.

Vladislav Valtchev (vvaltchev)

[Index of Archives]     [Kernel]     [Linux ARM]     [Yosemite News]     [Linux SCSI]     [Linux Hams]

  Powered by Linux