Re: Interrupt latency problems

From: Bosko Milekic <bmilekic_at_unixdaemons.com>
Date: Sat, 10 May 2003 14:02:14 -0400
On Sat, May 10, 2003 at 05:54:10PM +0100, Paul Richards wrote:
[...]
> I'm not sure it's as simple as Giant though. We have a great big Giant
> in 4.X and I don't see the same problem there. Something in -current is
> not working right on SMP (I also don't have a single processor box to
> hand to confirm it's just SMP either) that's making the system virtually
> unuseable as a workstation.
> 
> Surely the remnants of Giant in a lot of the kernel should just degrade
> the behaviour to being like the single big lock we had before, rather
> than causing these big latencies that people are seeing? (of course the
> locking mechanisms we now have are heavier than the simple kernel lock
> we had before but even so, what I'm seeing is really severe).
> 
> Is everyone seeing this? If not then perhaps we can narrow down the
> problem and look at getting things improved before 5-stable, because if
> it stays like this it's not really what I'd consider useable.

  The thing is that in 5.x the Giant lock is more expensive in itself
  and interrupts themselves are blocked on Giant.  Further, you have to
  do something like take the interrupt, schedule the interrupt, context
  switch, run the interrupt thread only to hit Giant, context switch out
  (possibly back to original thread, possibly not), and so on.  Compared
  to 4.x where interrupts were masked and thus the extra context
  switches avoided...

  The 5.x system scales much better, but Giant really has to be pushed
  out.  It's still unfair to compare the two at this point...

> > > Paul Richards <paul_at_freebsd-services.com>
> > > FreeBSD Services Ltd
> > 
> > <soabbox>
> > Now, if you are talking as freebsd-services as this implies, why dont
> > you guys use some of the $$ you make on FreeBSD to sponser work like
> > the above to happen ?
> > Looks to me like it would be beneficial to all parties..
> > </soabbox>
> 
> If we had the $$ to do so I certainly would.
> 
> -- 
> Paul Richards <paul_at_freebsd-services.com>
> FreeBSD Services Ltd
> 
> _______________________________________________
> freebsd-current_at_freebsd.org mailing list
> http://lists.freebsd.org/mailman/listinfo/freebsd-current
> To unsubscribe, send any mail to "freebsd-current-unsubscribe_at_freebsd.org"
> 

-- 
Bosko Milekic
bmilekic_at_unixdaemons.com
bmilekic_at_FreeBSD.org
Received on Sat May 10 2003 - 09:03:37 UTC

This archive was generated by hypermail 2.4.0 : Wed May 19 2021 - 11:37:07 UTC