On Tue, Oct 10, 2017 at 08:14:09PM +0200, Peter Zijlstra wrote: > On Tue, Oct 10, 2017 at 09:56:26AM -0700, Linus Torvalds wrote: > > > So I think the best model would be something like this: > > > > - T1: > > mutex_lock(&lock) > > ... > > mutex_transfer(&lock) > > > > - T2: > > mutex_receive(&lock); > > ... > > mutex_unlock(&lock); > > > > where the "mutex_transfer() -> mutex_receive()" thing really makes it > > obvious that "now thread 1 is transferring the lock to thread 2". > > Ah, but that's not at all what cross-release is about. Nobody really > does wonky ownership transfer of mutexes like that (although there might > be someone doing something with semaphores, I didn't check). Its to > allow detecting this deadlock: > > mutex_lock(&lock) > wait_for_completion(&c); > mutex_lock(&lock); > complete(&c); > > The completion doesn't have an owner to transfer.
Plus, lock_page(). Honestly, I want that to be the main beneficiary when we talking about crossrelease. Actually, I started the crossrelease work to detect deadlocks by lock_page() and expect it's more useful.
