On Thu, Apr 08, 2021 at 06:23:38AM +0100, Matthew Wilcox (Oracle) wrote: > bit_spinlocks are horrible on RT because there's absolutely nowhere > to put the mutex to sleep on. They also do not participate in lockdep > because there's nowhere to put the map. > > Most (all?) bit spinlocks are actually a split lock; logically they > could be treated as a single spinlock, but for performance, we want to > split the lock over many objects. Introduce the split_lock as somewhere > to store the lockdep map and as somewhere that the RT kernel can put > a mutex. It may also let us store a ticket lock for better performance > on non-RT kernels in the future, but I have left the current cpu_relax() > implementation intact for now.
I think I like it, but I'm not sure it'll work for RT as is. It's a bit like qrwlock in that it only uses the internal (split) lock for contention, but that doesn't work for PI. I've not recently looked at RT, but I think they simply used to bloat a number of the data structures with a real lock. Sebastian and Thomas will know better.