From: Mikael Pettersson <mikpe@csd.uu.se>
This patch fixes an AMD K8-specific problem with perfctr's x86
micro-benchmarking code. Due to lack of serialisation in K8's RDTSC, the
"empty loop" benchmark can appear to take longer than "loop of cheap
operation" benchmarks, causing their per- operation costs to be way off
(negative differences divided by number of operations). The workaround
(from AMD's manuals and the x86-64 arch code) is to run CPUID before RDTSC.
Signed-off-by: Mikael Pettersson <mikpe@csd.uu.se>
Signed-off-by: Andrew Morton <akpm@osdl.org>
---
25-akpm/drivers/perfctr/x86_tests.c | 34 +++++++++++++++++++++++++++++++---
1 files changed, 31 insertions(+), 3 deletions(-)
diff -puN drivers/perfctr/x86_tests.c~perfctr-k8-fix-for-internal-benchmarking-code drivers/perfctr/x86_tests.c
--- 25/drivers/perfctr/x86_tests.c~perfctr-k8-fix-for-internal-benchmarking-code 2004-07-26 22:21:46.779934312 -0700
+++ 25-akpm/drivers/perfctr/x86_tests.c 2004-07-26 22:21:46.783933704 -0700
@@ -1,4 +1,4 @@
-/* $Id: x86_tests.c,v 1.28 2004/05/23 23:22:44 mikpe Exp $
+/* $Id: x86_tests.c,v 1.31 2004/07/26 12:02:32 mikpe Exp $
* Performance-monitoring counters driver.
* Optional x86/x86_64-specific init-time tests.
*
@@ -49,6 +49,15 @@
#define apic_write(reg,vector) do{}while(0)
#endif
+#if !defined(__x86_64__)
+/* Avoid speculative execution by the CPU */
+extern inline void sync_core(void)
+{
+ int tmp;
+ asm volatile("cpuid" : "=a" (tmp) : "0" (1) : "ebx","ecx","edx","memory");
+}
+#endif
+
static void __init do_rdpmc(unsigned pmc, unsigned unused2)
{
unsigned i;
@@ -107,6 +116,21 @@ static void __init do_wrlvtpc(unsigned v
}
}
+static void __init do_sync_core(unsigned unused1, unsigned unused2)
+{
+ unsigned i;
+ for(i = 0; i < NITER/8; ++i) {
+ sync_core();
+ sync_core();
+ sync_core();
+ sync_core();
+ sync_core();
+ sync_core();
+ sync_core();
+ sync_core();
+ }
+}
+
static void __init do_empty_loop(unsigned unused1, unsigned unused2)
{
unsigned i;
@@ -118,8 +142,10 @@ static unsigned __init run(void (*doit)(
unsigned arg1, unsigned arg2)
{
unsigned start, dummy, stop;
+ sync_core();
rdtsc(start, dummy);
(*doit)(arg1, arg2); /* should take < 2^32 cycles to complete */
+ sync_core();
rdtsc(stop, dummy);
return stop - start;
}
@@ -143,8 +169,8 @@ measure_overheads(unsigned msr_evntsel0,
unsigned msr_cccr, unsigned cccr_val)
{
int i;
- unsigned int loop, ticks[12];
- const char *name[12];
+ unsigned int loop, ticks[13];
+ const char *name[13];
if (msr_evntsel0)
wrmsr(msr_evntsel0, 0, 0);
@@ -177,6 +203,8 @@ measure_overheads(unsigned msr_evntsel0,
name[11] = "write LVTPC";
ticks[11] = (perfctr_info.cpu_features & PERFCTR_FEATURE_PCINT)
? run(do_wrlvtpc, APIC_DM_NMI|APIC_LVT_MASKED, 0) : 0;
+ name[12] = "sync_core";
+ ticks[12] = run(do_sync_core, 0, 0);
loop = run(do_empty_loop, 0, 0);
_