io apic + tsc = slowdown (bugreport + possible fix)

io apic + tsc = slowdown (bugreport + possible fix)

[Frederic Krueger]

Hi,

just wanted to point out the patch I imported from an older version of 
the 2.6.x kernel (not done by me at all ;-)).
It's basically causing the system to a near zero slowdown after one hour 
of uptime. And the reason for it must be the io_apic irq0 handling with 
TSC enabled systems (seems to show up on ibm pre-built boxes more often 
than on others though).
Funnily pressing any keys _all the time_ works around for this for the 
time a key is pressed ;)

The bug report entry contains a patch for arch/i386/kernel/io_apic.c  
which fixed it for me.
Maybe this should make it into the cvs tree since it's a _really_ 
annoying bug, basically rendering the computer unusable if present.

The link to the bug report with more detailed information can be found here:
http://bugme.osdl.org/show_bug.cgi?id=2964

Thanks for checking and considering :-)

Bye,
Frederic

Re: io apic + tsc = slowdown (bugreport + possible fix)

[john stultz]

On Mon, 2004-06-28 at 00:27, Frederic Krueger wrote:
> just wanted to point out the patch I imported from an older version of 
> the 2.6.x kernel (not done by me at all ;-)).
> It's basically causing the system to a near zero slowdown after one hour 
> of uptime. And the reason for it must be the io_apic irq0 handling with 
> TSC enabled systems (seems to show up on ibm pre-built boxes more often 
> than on others though).
> Funnily pressing any keys _all the time_ works around for this for the 
> time a key is pressed ;)

I duped the bug to 2544. I've got the BIOS folks looking at this issue
on the x305 for the proper fix. 

Looking closer at the proposed workaround by Maciej posted here:
http://linux.derkeiler.com/Mailing-Lists/Kernel/2004-04/3174.html

Why exactly are we using cpu_has_tsc to switch this? I'm not sure I'm
following how this is TSC dependent. Additionally the comment change
looks to be from the 2.4 era. 

Maciej, could you further explain? 

thanks
-john

Re: io apic + tsc = slowdown (bugreport + possible fix)

[Maciej W. Rozycki]

On Mon, 28 Jun 2004, john stultz wrote:

> Looking closer at the proposed workaround by Maciej posted here:
> http://linux.derkeiler.com/Mailing-Lists/Kernel/2004-04/3174.html
> 
> Why exactly are we using cpu_has_tsc to switch this? I'm not sure I'm
> following how this is TSC dependent. Additionally the comment change
> looks to be from the 2.4 era. 

 One of the two users of timer_ack is do_slow_gettimeoffset().  When the
TSC is selected for use as a high-precision timer do_fast_gettimeoffset()
is used instead.

 Please folks do read the sources sometimes -- I've been repeatedly
clarifying these bits while they are all documented in the sources,
sigh...

  Maciej

Re: io apic + tsc = slowdown (bugreport + possible fix)

[john stultz]

On Tue, 2004-06-29 at 05:07, Maciej W. Rozycki wrote:
> On Mon, 28 Jun 2004, john stultz wrote:
> 
> > Looking closer at the proposed workaround by Maciej posted here:
> > http://linux.derkeiler.com/Mailing-Lists/Kernel/2004-04/3174.html
> > 
> > Why exactly are we using cpu_has_tsc to switch this? I'm not sure I'm
> > following how this is TSC dependent. Additionally the comment change
> > looks to be from the 2.4 era. 
> 
>  One of the two users of timer_ack is do_slow_gettimeoffset().  When the
> TSC is selected for use as a high-precision timer do_fast_gettimeoffset()
> is used instead.
> 
>  Please folks do read the sources sometimes -- I've been repeatedly
> clarifying these bits while they are all documented in the sources,
> sigh...

Whoops, sorry, I had the impression the patch was targeted against 2.6,
where there is no do_slow_gettimeoffset() and the only user of timer_ack
is in do_timer_interrupt(). 

So am I incorrect that the TSC bits can be dropped for the 2.6 version
of the patch?

thanks
-john

Re: io apic + tsc = slowdown (bugreport + possible fix)

[Maciej W. Rozycki]

On Tue, 29 Jun 2004, john stultz wrote:

> >  Please folks do read the sources sometimes -- I've been repeatedly
> > clarifying these bits while they are all documented in the sources,
> > sigh...
> 
> Whoops, sorry, I had the impression the patch was targeted against 2.6,

 It works the same way for both series.

> where there is no do_slow_gettimeoffset() and the only user of timer_ack
> is in do_timer_interrupt(). 

 A quick search reveals the bits were moved to do_timer_overflow() in
include/asm-i386/mach-default/do_timer.h which is used by
arch/i386/kernel/timers/timer_pit.c -- I suppose the comment in 
arch/i386/kernel/time.c needs an update then.

> So am I incorrect that the TSC bits can be dropped for the 2.6 version
> of the patch?

 You are, sorry.

  Maciej