Re: Perf and ftrace [was Re: PyTimechart]

[Mathieu Desnoyers <mathieu.desnoyers@efficios.com>]

* Peter Zijlstra (peterz@infradead.org) wrote:
> On Wed, 2010-05-12 at 13:07 -0400, Mathieu Desnoyers wrote:
> > > Its mostly an interface/api question. You cannot easily splice() a
> > > mmap()'ed buffer on machines that have address constraints like sparc.
> > 
> > Ah ? Can you explain this issue a bit more ? There is possibly a concern I don't
> > quite see here. 
> 
> IIRC Sparc has virtually tagged D-caches on the lower 9 bits of the pfn,
> so to avoid cache aliasing both mappings (kernel and user) need to be
> aligned.

I see.

> 
> Since splice needs to swap pages it needs to allocate replacement pages
> with the exact right alignment, which is rather expensive (in either
> time or space).
> 

Hrm ? Why does splice() _need_ to swap pages exactly ? Or is it purely just the
way the current Ftrace implementation happens to work ? ;)

Would you be fine with an alternative implementation and API that support both
mmap() and splice(), without any need to swap pages ?

Thanks,

Mathieu


-- 
Mathieu Desnoyers
Operating System Efficiency R&D Consultant
EfficiOS Inc.
http://www.efficios.com