On Tue, Jul 28, 2015 at 11:14:40AM -0700, Davidlohr Bueso wrote: > On Mon, 2015-07-27 at 13:31 -0700, josh@xxxxxxxxxxxxxxxx wrote: > > That sounds interesting! mmap_sem is definitely a performance > > bottleneck. How do you handle writes versus reads? > > The idea is to make vmas srcu aware, such that their lookups in the > vmacache are lockless and can survive the entire fault path, among > others we have ->vm_file. We simply handle cases when the > vma/page-tables have changed between when the lookup was done and when > we grab the pte lock with mmap_sem. These invalidations are a pain, > albeit non fatal in some cases. SOunds promising. > > > Yes, you can argue that they're not published all you want, > > > but I'm talking beyond my specific use case. Linux VM is known to scale, > > > why should we hide a core scalability tool from it? > > > > In the case of mmap_sem, does it help at all if tiny kernels were 1) > > non-preemptible and 2) non-SMP? Tiny kernels don't necessarily care > > about scaling. > > Yes, I believe it would! I actually assumed tiny kernels were already > UP. I don't think it makes much sense to have it at that level. Same > with preemption. OK. So, would you consider making it possible to compile out SRCU in that specific case, while depending on SRCU if either SMP or PREEMPT? Because that's also the case that allows tiny RCU rather than full RCU. - Josh Triplett -- To unsubscribe from this list: send the line "unsubscribe linux-next" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
