On Tue, Nov 18, 2014 at 12:36 PM, Masao Uebayashi <uebay...@gmail.com> wrote: > On Tue, Nov 18, 2014 at 11:24 AM, Ryota Ozaki <ozak...@netbsd.org> wrote: >> First I'm not against restructuring, though I hoped minimum >> restructuring on non-performance-sensitive paths. > > My understanding is that, non-performance-sensitive paths (e.g. > ioctl()'s) also touches performance-sensitive data structures (e.g. > ifnet list), that is the problem. Non-performance-sensitive paths > have to be more careful to not disturb performance-sensitive data > structures and paths. For ioctl()'s, they have to allocate > kernel-wired memory to not sleep (via copyout()) while iterating > performance-critical ifnet list. Thus I don't think you can avoid > restructuring those codes.
pserialize can prevent slow paths (IFNET_LOCK) from disturbing fast paths (IFNET_RENTER) because IFNET_RENTER (pserialize_read_enter) is just a splsoftserial and can run regardless of IFNET_LOCK. Of course slow paths including ioctl() can work against each other. > > (I'd want to also encourage you to *not* become too nervous about > changing those codes. Let's progress bravely and rely on high-quality > NetBSD "end" users for testing. :) You know well :) I'm now positive to do so. Thanks! ozaki-r > >> BTW, do you think we eventually introduce "all packet processing in >> a LWP context"-like restructuring? I'm inclined to run Layer2 and bpf >> in softint. If we also end up doing so, I want to do it early. >> Of course, we have to address performance issues somehow at some point >> though. >> >> Thanks, >> ozaki-r >> >>> >>> Dave >>> >>> -- >>> David Young >>> dyo...@pobox.com Urbana, IL (217) 721-9981
