On Wed, Oct 08, 2014 at 05:10:59AM -0500, Chuck Ebbert wrote:
> On Wed, 8 Oct 2014 10:03:36 +0100
> Chris Wilson <ch...@chris-wilson.co.uk> wrote:
>
> >
> > I ran into a problem on a Sandybridge i5-2500s whilst measuring the
> > performance of GTT write-combining access. I found subsequent runs were
> > about 10-40x slower than the first. For example,
> >
> > igt/gem_gtt_speed:
> >
> > Time to read 16k through a GTT map: 325.285µs
> > Time to write 16k through a GTT map: 4.729µs
> > Time to clear 16k through a GTT map: 4.584µs
> > Time to clear 16k through a cached GTT map: 1.342µs
> >
> > on the second run became:
> >
> > Time to read 16k through a GTT map: 332.148µs
> > Time to write 16k through a GTT map: 209.411µs
> > Time to clear 16k through a GTT map: 56.460µs
> > Time to clear 16k through a cached GTT map: 50.897µs
> >
> > Naively I would say that we lost the wc on our ioremap.
> > /sys/kernel/debug/x86/pat_memtype_list remained the same across repeated
> > runs.
> >
> > A bisection pointed to
> >
> > commit ea8596bb2d8d37957f3e92db9511c50801689180
> > Author: Masami Hiramatsu <masami.hiramatsu...@hitachi.com>
> > Date: Thu Jul 18 20:47:53 2013 +0900
> >
> > kprobes/x86: Remove unused text_poke_smp() and text_poke_smp_batch()
> > functions
> >
> > of which the active ingredient was just
> >
> > diff --git a/arch/x86/Kconfig b/arch/x86/Kconfig
> > index b32ebf9..f4001e0 100644
> > --- a/arch/x86/Kconfig
> > +++ b/arch/x86/Kconfig
> > @@ -2334,7 +2334,6 @@ config HAVE_ATOMIC_IOMAP
> >
> > config HAVE_TEXT_POKE_SMP
> > bool
> > - select STOP_MACHINE if SMP
> >
> > config X86_DEV_DMA_OPS
> > bool
> >
> > and adding that back into the current build, e.g.
>
> Hmm, set_mtrr() uses stop_machine(). I wonder if your MTRRs are out of
> sync and your results depend on which CPU the test runs on?
Indeed, this appears to be the explanation. (And here I thought PAT
superseded mtrrs - i915.ko stopped trying to use assign an mtrr for its
GTT quite a while ago.)
Replacing the stop_machine there with on_each_cpu does the trick:
diff --git a/arch/x86/kernel/cpu/mtrr/main.c b/arch/x86/kernel/cpu/mtrr/main.c
index f961de9..c0e37d5 100644
--- a/arch/x86/kernel/cpu/mtrr/main.c
+++ b/arch/x86/kernel/cpu/mtrr/main.c
@@ -151,7 +151,7 @@ struct set_mtrr_data {
*
* Returns nothing.
*/
-static int mtrr_rendezvous_handler(void *info)
+static void mtrr_rendezvous_handler(void *info)
{
struct set_mtrr_data *data = info;
@@ -174,7 +174,6 @@ static int mtrr_rendezvous_handler(void *info)
} else if (mtrr_aps_delayed_init || !cpu_online(smp_processor_id())) {
mtrr_if->set_all();
}
- return 0;
}
static inline int types_compatible(mtrr_type type1, mtrr_type type2)
@@ -228,7 +227,7 @@ set_mtrr(unsigned int reg, unsigned long base, unsigned
long size, mtrr_type typ
.smp_type = type
};
- stop_machine(mtrr_rendezvous_handler, &data, cpu_online_mask);
+ on_each_cpu_mask(cpu_online_mask, mtrr_rendezvous_handler, &data, true);
}
static void set_mtrr_from_inactive_cpu(unsigned int reg, unsigned long base,
@@ -240,8 +239,7 @@ static void set_mtrr_from_inactive_cpu(unsigned int reg,
unsigned long base,
.smp_type = type
};
- stop_machine_from_inactive_cpu(mtrr_rendezvous_handler, &data,
- cpu_callout_mask);
+ on_each_cpu_mask(cpu_callout_mask, mtrr_rendezvous_handler, &data,
true);
}
/**
--
Chris Wilson, Intel Open Source Technology Centre
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
