On Thu, Nov 25, 2010 at 11:23:42PM +0100, Mark Kettenis wrote:
> > Date: Thu, 25 Nov 2010 12:30:39 +0000
> > From: Owain Ainsworth <zer...@googlemail.com>
> >
> > On Thu, Nov 25, 2010 at 11:50:06AM +0100, Claudio Jeker wrote:
> > > On Wed, Nov 24, 2010 at 05:42:59PM +0100, Mike Belopuhov wrote:
> > > > On Wed, Nov 24, 2010 at 17:06 +0100, Claudio Jeker wrote:
> > > > > This diff was made by oga@ some time ago -- I just fixed a few
> > > > > conflicts
> > > > > and I would really like to see this going in.
> > > > >
> > > > > netisr needs to be made a normal C function and not this madness it
> > > > > currently is. This is necessary to imporve the packet scheduling.
> > > > >
> > > > > Tested myself on i386, amd64, sparc64, sparc, alpha and hppa
> > > >
> > > > hi,
> > > >
> > > > i like the diff, but why not get rid of netisr_dispatch.h?
> > >
> > > I planned to this after this diff goes in. I did not want to change
> > > Owain's initial diff too much so I left that out for now.
> >
> > Miod pointed out to me when I first sent this out that some archs use
> > something slightly different than this generic version.
> >
> > For example hppa uses a specially aligned netisr variable so it can use
> > the pa-risc real atomic instructions which are quite a bit faster.
>
> Especially on MP kernels.
>
> > Similarly x86 uses compare-and-swap instead of test, then clear bits.
> >
> > I meant to add some MD defines so that this could be kept (falling back
> > to the implementation in this diff) since miod claims that it is useful
> > and significantly faster (I did not check this). I've cced miod so he
> > can answer with somewhat more authority.
>
> The hppa and i386/amd64 approach are rather similar. Both do an
> atomic load and clear, which happens to be the only atomic instruction
> available on hppa, and is easy to implement as an atomic swap with 0
> on i386/amd64. But implementing this operation on other architectures
> shouldn't be much more difficult, and should work just as well as the
> proposed code that uses atomic_clearbits_int(). The MI code would
> look something like:
>
> void
> netintr(void *unused)
> {
> int n;
>
> n = atomic_load_and_clear(&netisr);
> #define DONETISR(bit, fn) \
> do { \
> if (n & 1 << (bit)) \
> fn(); \
> } while ( /* CONSTCOND */ 0)
> #include <net/netisr_dispatch.h>
> }
The only other thing to note is that you may need to define a type (or
whatever) since iirc the hppa one needs a specific alignment (cacheline
to itself).
Otherwise I like this approach, it is cleaner than what I was thinking
(not knowing exactly what hppa was doing other than it was faster didn't
help ;).
-0-
--
Sometimes I worry about being a success in a mediocre world.
-- Lily Tomlin
