On Thu, Jun 06, 2019 at 03:54:24PM +0000, Jorgen Hansen wrote: > > > > On 6 Jun 2019, at 11:34, Peter Zijlstra <pet...@infradead.org> wrote: > > > > > > The VMCI driver is abusing atomic64_t and atomic_t, there is no actual > > atomic RmW operations around. > > > > Rewrite the code to use a regular u64 with READ_ONCE() and > > WRITE_ONCE() and a cast to 'unsigned long'. This fully preserves > > whatever broken there was (it's not endian-safe for starters, and also > > looks to be missing ordering). > > Thanks for the cleanup. > > This code is only intended for use with the vmci device driver, and > that is X86 only, so during the original upstreaming no effort was > made to make this work correctly on anything else.
Even x86 needs compiler barriers to ensure the compiler emits the instructions in the expected order. > We’ll be updating the vmci device driver to work on other > architectures soonish, so will be adding barriers to enforce ordering > as well at that point. If you want to leave your patch as is, we can > address the type casting then. You might want to have a look at smp_store_release() and smp_load_acquire() I suppose.