lkml.org 
[lkml]   [2008]   [Nov]   [26]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [patch 21/24] perfmon: Intel architectural PMU support (x86)
On Wed, 26 Nov 2008, stephane eranian wrote:
> In anycase, the idea is to encapsulate as much as possible code
> related into a PMU model
> into each module. That is why you are seing some redundancy.

Makes sense.

> There is a difference between enable_mask and used_pmcs. The used_pmcs
> bitmasks shows
> all the config registers in use. Whereas enable_mask shows the all
> config registers which have
> start/stop capabilities. For the basic AMD64 PMU (4 counters)
> used_pmcs and enable_mask
> are equivalent, but that is not the case on Barcelona once we support
> IBS and sampling. So
> for now, I could clean this up and drop enable_mask to use plain used_pmcs.

Understood. If we need that in the near future then it's ok to keep
it, it just did not make any sense from the current code.

But I think you should do this once when you set up the context and
keep that as a separate mask. Right now you evaluate enable_mask and
used_pmcs over and over again.

> >> + count = pfm_arch_bv_weight(used_mask, max_enable);
> >
> > So we have:
> >
> > set->used_pmcs and enable_mask and max_enable.
> >
> > Why can set->used_pmcs contain bits which are not in the enable_mask
> > in the first place ? Why does the arch code not tell the generic code
> > which pmcs are available so we can avoid all this mask, weight
> > whatever magic ?
> >
>
> Because used_pmcs is part of generic code and enable_mask is a x86 construct.
> As I said above, for now, I could drop enable_mask.
> The arch code already export the list of available pmcs and pmds in
> impl_pmcs and impl_pmds.

See above.

> > Why are the counters enabled at all when an overflow is pending, which
> > stopped the counters anyway ?
> >
> Because on Intel and AMD64, counters are not automatically frozen on interrupt.
> On Intel X86, they can be configured to do so, but it is an all or
> nothing setting.
> I am not using this option because we would then have a problem with the NMI
> watchdog given that it is also using a counter.

Well, my question was: why do we have to stop the counters when an
overflow is pending already ?

The overflow pending is set inside of stop_save() and cleared
somewhere else.

stop_save() is called from pfm_arch_stop() and
pfm_arch_ctxswout_thread(). The first thing it does is to disable the
counters.

Now at some points the counters are obviously reenabled for this
context, but why are they reenabled _before_ the pending overflow has
been resolved ? For N counters that N * 2 wrmsrl() overhead.

Thanks,

tglx


\
 
 \ /
  Last update: 2008-11-26 17:13    [from the cache]
©2003-2014 Jasper Spaans. hosted at Digital Ocean