On Tue, Aug 06, 2019 at 06:17:42PM +0200, Oleg Nesterov wrote: > but this will also wake all the pending readers up. Every reader will burn > CPU for no reason and likely delay the writer. > > In fact I'm afraid this can lead to live-lock, because every reader in turn > will call __percpu_up_read().
I didn't really consider that case important; because of how heavy the write side is, it should be relatively rare. > How about 2 wait queues? That said, I can certainly try that.
