linux-kernel - [PATCH] perf_event use rdpmc rather than rdmsr when possible in kernel

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [thread-next>] [day] [month] [year] [list]

Date:	Mon, 20 Feb 2012 17:38:49 -0500
From:	Vince Weaver <vweaver1@...s.utk.edu>
To:	<linux-kernel@...r.kernel.org>
CC:	<mingo@...e.hu>, <a.p.zijlstra@...llo.nl>,
	Paul Mackerras <paulus@...ba.org>,
	Arnaldo Carvalho de Melo <acme@...stprotocols.net>,
	Stephane Eranian <eranian@...il.com>
Subject: [PATCH] perf_event use rdpmc rather than rdmsr when possible in
 kernel

Hello

The perfctr interface uses rdpmc rather than rdmsr when possible in the 
kernel, as rdpmc tends to have lower latencies.  (One can look in 
etc/costs in the perfctr-2.6 package to see a historical list of the
overhead).

I have done some tests on a 3.2 kernel:

                    rdmsr            rdpmc
Core2 T9900:      203.9 cycles     30.9 cycles
AMD fam0fh:        56.2 cycles      9.8 cycles
Atom 6/28/2:      129.7 cycles     50.6 cycles

As you can see the speedup of using rdpmc is large, although granted
it really is a drop in the bucket compared to the other overheads 
involved.

I've attached a kernel module that can be used to test this on your own 
personal x86 machine.

Below is a patch that changes perf_event to use rdpmc rather than rdmsr 
when possible.  It's probably possible (and desirable) to do this without 
requiring a new field in the hw_perf_event structure, but the fixed events 
make this tricky.

Signed-off-by: Vince Weaver <vweaver1@...s.utk.edu>

diff --git a/arch/x86/kernel/cpu/perf_event.c b/arch/x86/kernel/cpu/perf_event.c
index 5adce10..5550047 100644
--- a/arch/x86/kernel/cpu/perf_event.c
+++ b/arch/x86/kernel/cpu/perf_event.c
@@ -85,7 +85,7 @@ u64 x86_perf_event_update(struct perf_event *event)
 	 */
 again:
 	prev_raw_count = local64_read(&hwc->prev_count);
-	rdmsrl(hwc->event_base, new_raw_count);
+	new_raw_count=native_read_pmc(hwc->event_base_rdpmc);
 
 	if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
 					new_raw_count) != prev_raw_count)
@@ -768,9 +768,11 @@ static inline void x86_assign_hw_event(struct perf_event *event,
 	} else if (hwc->idx >= X86_PMC_IDX_FIXED) {
 		hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
 		hwc->event_base = MSR_ARCH_PERFMON_FIXED_CTR0 + (hwc->idx - X86_PMC_IDX_FIXED);
+		hwc->event_base_rdpmc = (hwc->idx - X86_PMC_IDX_FIXED) | 1<<30;
 	} else {
 		hwc->config_base = x86_pmu_config_addr(hwc->idx);
 		hwc->event_base  = x86_pmu_event_addr(hwc->idx);
+		hwc->event_base_rdpmc = x86_pmu_addr_offset(hwc->idx);
 	}
 }
 
diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
index abb2776..432ac69 100644
--- a/include/linux/perf_event.h
+++ b/include/linux/perf_event.h
@@ -562,6 +562,7 @@ struct hw_perf_event {
 			u64		last_tag;
 			unsigned long	config_base;
 			unsigned long	event_base;
+			unsigned long	event_base_rdpmc;
 			int		idx;
 			int		last_cpu;
 			struct hw_perf_event_extra extra_reg;
View attachment "Makefile" of type "text/plain" (197 bytes)

View attachment "rdpmc-module.c" of type "text/x-csrc" (7352 bytes)