Re: [PATCH 3/4] perf tool: Fix build when sysconf doesn't support cache line size

[Arnaldo Carvalho de Melo <acme@kernel.org>]

Em Mon, Jul 04, 2016 at 05:19:20PM -0700, Chris Phlipot escreveu:
> 
> 
> On 07/04/2016 03:48 PM, Arnaldo Carvalho de Melo wrote:
> > Em Thu, Jun 30, 2016 at 10:12:34PM -0700, Chris Phlipot escreveu:
> > > Enable perf to build on libc implementations where sysconf() doesn't
> > > support _SC_LEVEL1_DCACHE_LINESIZE as a parameter.
> > > 
> > > For example, the Bionic implementation does not support this as a
> > > paremter. Older versions of Bionic will throw an error when this is passed
> > > in as a parameter, and more recent versions will just return 0 as the
> > > cache line size.
> > > 
> > > Signed-off-by: Chris Phlipot <cphlipot0@gmail.com>
> > > ---
> > >   tools/perf/perf.c | 4 ++++
> > >   1 file changed, 4 insertions(+)
> > > 
> > > diff --git a/tools/perf/perf.c b/tools/perf/perf.c
> > > index 8f21922..113ca5b 100644
> > > --- a/tools/perf/perf.c
> > > +++ b/tools/perf/perf.c
> > > @@ -509,7 +509,11 @@ int main(int argc, const char **argv)
> > > 
> > >   	/* The page_size is placed in util object. */
> > >   	page_size = sysconf(_SC_PAGE_SIZE);
> > > +#ifdef _SC_LEVEL1_DCACHE_LINESIZE
> > >   	cacheline_size = sysconf(_SC_LEVEL1_DCACHE_LINESIZE);
> > > +#else
> > > +	cacheline_size = 0;
> > > +#endif
> > 
> > Couldn't we instead fallback to:
> > 
> > sysfs__read_int("devices/system/cpu/cpu0/cache/index0/coherency_line_size", &cacheline_size)
> > 
> > ?
> 
> I agree that in general this would be a better fallback, but in  all Android
> images I have tested so far, "devices/system/cpu/cpu0/cache" does not exist.
> I know not know of a good way to retrieve cache line size in this case.
> 
> I would be ok with attempting to get cacheline size using using the
> following methods, unless you have other ideas:
> 
> 1. attempt to use sysconf(_SC_LEVEL1_DCACHE_LINESIZE)
> 2. attempt to use
> sysfs__read_int("devices/system/cpu/cpu0/cache/index0/coherency_line_size",
> &cacheline_size)
> 3. set to zero if both of the above fail.

Ok, but perhaps we should have some sort of warning in places using
this?

- Arnaldo
 
> 
> > 
> > >   	if (sysctl__read_int("kernel/perf_event_max_stack", &value) == 0)
> > >   		sysctl_perf_event_max_stack = value;
> > > --
> > > 2.7.4