On Thu, Aug 16, 2018 at 07:52:06AM +1000, Benjamin Herrenschmidt wrote: > I would also like to use this rather than the bitmap atomics for is_added > etc... (Hari's fix) in the long run. Atomics aren't significantly cheaper > and imho makes thing even messier. PCI device enablement isn't a hotpath, so the performance saving of acquire/release memory barriers vs. full memory barriers is hardly a strong argument. Atomic bitops have the advantage of being usable in atomic context, unlike a struct mutex. And they don't require any spinning if the variable is accessed concurrently. Last not least an atomic bitop needs just 1 line of code versus 3 lines of code to acquire a lock, update the variable and release the lock. "Elegance is not optional." That said, the race *you're* dealing with appears to be distinct from Hari's in that a lock is unavoidable because (AFAIUI) atomic_inc_return() is called before enablement actually happens, yet both needs to happen atomically. Thanks, Lukas
