The problem:

If your application native code was generated with a NDK release older than r7b
and uses any of the following functions defined in the <sys/atomics.h>


Then the corresponding machine code is not guaranteed to work properly on
some multi-core Android ARM-based devices (x86 ones are not affected).

The solution:

The <sys/atomics.h> header has been updated in NDK r7b. Simply recompiling
your _unmodified_ sources with this version of the NDK should be enough to
completely eliminate the problem.

If you can't use NDK r7b or later for some reason, read the section below.

More details:

The main issue is that the implementation of these functions, as provided
by the C library, did not provide any associated memory barriers. This is
by design, because the platform code that uses them does insert explicit
barriers around these operations.

The functions were only exposed through the NDK by mistake, they were not
supposed to be used from applications. Any application code that use them
without inserting its own barriers may experiment incorrect behaviour,
which can result in bugs that are very hard to reproduce and diagnose.

Not all multi-core devices are affected though. Certain OEMs enforce a
policy where all threads of a single process are forced to run on the same
core. In this case, the bug cannot occur, unless you're directly accessing
shared memory between two processes.

The problem is only likely to be seen on devices running Android 3.0 to
Android 4.1. The C library implementation in 4.1 has been updated to provide
full memory barriers as well. This ensures existing native code keeps
working correctly on this platform and future ones, even if they were not

We still strongly recommend recompiling your native code to ensure you'll
never have to debug this issue (life is short). In the case where this would
not be possible (e.g. you're using an older version of the NDK for some
reason, or a custom build system / toolchain), we recommend stopping from
using these functions entirely. Very fortunately, GCC provides handy
intrinsics functions that work with very reasonable performance and always
provide a full barrier.

  __sync_fetch_and_add         instead of __atomic_inc
  __sync_fetch_and_sub         instead of __atomic_sub
  __sync_val_compare_and_swap  instead of __atomic_cmpxchg

See the content of platforms/android-3/arch-arm/usr/include/sys/atomics.h
to see how these can be used.

See the following URL for more GCC-related information: