On Tue, May 25, 2021 at 10:07:46AM -0400, Jason J. Herne wrote: > On 5/25/21 9:26 AM, Jason Gunthorpe wrote: > > On Tue, May 25, 2021 at 09:24:59AM -0400, Jason J. Herne wrote: > > > change the function pointer to point to vfio_ap_ops:handle_pqap(). When we > > > unload the module we change the function pointer back to the stub. The > > > updates should be atomic operations so no lock needed, right? > > > > No > > > > Jason > > > > Okay... Would you be willing to elaborate, please? A counter argument, or a > simple explanation would be appreciated. A simple "no" does not really do > much to advance the discussion :). Go back and review the earlier thread, the issue was never the atomicity of the function pointer but the locking of the data that function is accessing. > I'm fairly sure that a 64-bit pointer would be updated atomically. A reader > of this value is either going to see value A or value B, not the high half > of A and the low half of B. Maybe we also need a memory barrier to prevent > stale values from being seen on another core? You need to use special macros in Linux to follow this memory model Jason
