On Tue, Aug 10, 2021 at 03:50:58PM -0600, Alex Williamson wrote: > On Tue, 10 Aug 2021 13:55:00 +0100 > Christoph Hellwig <hch@xxxxxxxxxxxxx> wrote: > > > On Tue, Aug 10, 2021 at 08:57:22AM -0300, Jason Gunthorpe wrote: > > > I'm not sure there is a real performance win to chase here? Doesn't > > > this only protect mmap against reset? The mmap isn't performance > > > sensitive, right? > > > > > > If this really needs extra optimization adding a rwsem to the devset > > > and using that across the whole set would surely be sufficient. > > > > Every mmio read or write takes memory_lock. > > Exactly. Ideally we're not using that path often, but I don't think > that's a good excuse to introduce memory access serialization, or even > dependencies between devices. Thanks, But a cross device rwsem seems OK to me?? It won't contend unless we are trying to reset and the upgrade to a percpu rwsem to optimize the atomic doesn't seem warranted since this path already has a vmexit and a syscall in it? Jason
