[RFC PATCH] perf: Add load latency monitoring on IntelNehalem/Westmere

From: Lin Ming
Date: Wed Dec 22 2010 - 03:09:21 EST

Next message: Richard Cochran: "Re: [PATCH V7 4/8] posix clocks: hook dynamic clocks into systemcalls"
Previous message: Michel Dänzer: "Re: [PATCH] drm/radeon: Remove DRIVER_DATE for UMS and KMS"
Next in thread: Peter Zijlstra: "Re: [RFC PATCH] perf: Add load latency monitoring on IntelNehalem/Westmere"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

Hi, all

This patch adds load latency monitoring on Intel Nehalem/Westmere.
It's applied on top of tip/master(3ea1f4f89) and Andi's offcore
patchsets are needed.

Updated perf offcore patchkit
http://marc.info/?l=linux-kernel&m=129103647731356&w=2

The load latency on Intel Nehalem/Westmere is monitored by event
MEM_INST_RETIRED.LATENCY_ABOVE_THRESHOLD(0x100b). It measures latency
from micro-operation (uop) dispatch to when data is globally observable
(GO).

To monitor load latency, both the PEBS_EN_CTRx and LL_EN_CTRx bits must
be set in IA32_PEBS_ENABLE register. And an extra MSR
MSR_PEBS_LD_LAT_THRESHOLD must be programmed with the desired latency
threshold in core clock cycles. Loads with latencies greater than this
value are counted.

The latency threshold is encoded in the upper 32 bits of
perf_event_attr::config and 'p' modifier must be used to enabel PEBS.

The default latency threshold is 3, as Intel manual says, "The minimum
value that may be programmed in this register is 3 (the minimum
detectable load latency is 4 core clock cycles)."

Here are some example outputs.

# perf top -e r100b:p

----------------------------------------------------------------------------------------------------
PerfTop: 1800 irqs/sec kernel:41.9% exact: 0.0% [1000Hz raw 0x100b], (all, 4 CPUs)
----------------------------------------------------------------------------------------------------

samples pcnt function DSO
_______ _____ _______________________________ _________________________________________

643.00 13.0% __lock_acquire [kernel.kallsyms]
328.00 6.6% check_chain_key [kernel.kallsyms]
308.00 6.2% mark_lock [kernel.kallsyms]
286.00 5.8% lock_release [kernel.kallsyms]
280.00 5.7% check_flags [kernel.kallsyms]
239.00 4.8% use_config /home/mlin/linux-2.6/scripts/basic/fixdep
201.00 4.1% trace_hardirqs_on_caller [kernel.kallsyms]
193.00 3.9% lock_acquire [kernel.kallsyms]
182.00 3.7% validate_chain [kernel.kallsyms]
162.00 3.3% __gconv_transform_utf8_internal /lib64/libc-2.8.so
145.00 2.9% trace_hardirqs_off_caller [kernel.kallsyms]
115.00 2.3% mark_held_locks [kernel.kallsyms]
102.00 2.1% __GI_mbrtowc /lib64/libc-2.8.so
76.00 1.5% do_raw_spin_lock [kernel.kallsyms]
73.00 1.5% parse_dep_file /home/mlin/linux-2.6/scripts/basic/fixdep
71.00 1.4% restore [kernel.kallsyms]
59.00 1.2% _int_malloc /lib64/libc-2.8.so

#Monitor load latency > 51 cycles
# perf top -e r510000100b:p

----------------------------------------------------------------------------------------------------
PerfTop: 2055 irqs/sec kernel:47.0% exact: 0.0% [1000Hz raw 0x510000100b], (all, 4 CPUs)
----------------------------------------------------------------------------------------------------

samples pcnt function DSO
_______ _____ _________________________ _________________________________________

419.00 9.8% __lock_acquire [kernel.kallsyms]
278.00 6.5% parse_dep_file /home/mlin/linux-2.6/scripts/basic/fixdep
195.00 4.6% find_get_page [kernel.kallsyms]
195.00 4.6% get_page_from_freelist [kernel.kallsyms]
184.00 4.3% __d_lookup [kernel.kallsyms]
180.00 4.2% memcmp [kernel.kallsyms]
135.00 3.2% __rmqueue [kernel.kallsyms]
123.00 2.9% __GI_strcmp /lib64/libc-2.8.so
92.00 2.2% unmap_vmas [kernel.kallsyms]
88.00 2.1% validate_chain [kernel.kallsyms]
85.00 2.0% do_raw_spin_lock [kernel.kallsyms]
78.00 1.8% __wake_up_bit [kernel.kallsyms]
77.00 1.8% __GI_strlen /lib64/libc-2.8.so
73.00 1.7% copy_user_generic_string [kernel.kallsyms]
71.00 1.7% page_remove_rmap [kernel.kallsyms]
70.00 1.6% kmem_cache_free [kernel.kallsyms]
67.00 1.6% free_pcppages_bulk [kernel.kallsyms]
62.00 1.5% kmem_cache_alloc [kernel.kallsyms]
58.00 1.4% ext3_release_file [kernel.kallsyms]
54.00 1.3% radix_tree_lookup_element [kernel.kallsyms]
47.00 1.1% generic_fillattr [kernel.kallsyms]

Any comment is appropriated.

Signed-off-by: Lin Ming <ming.m.lin@xxxxxxxxx>
---
arch/x86/kernel/cpu/perf_event.c | 18 +++++++++++++++---
arch/x86/kernel/cpu/perf_event_intel.c | 4 ++++
arch/x86/kernel/cpu/perf_event_intel_ds.c | 5 +++++
include/linux/perf_event.h | 1 +
4 files changed, 25 insertions(+), 3 deletions(-)

diff --git a/arch/x86/kernel/cpu/perf_event.c b/arch/x86/kernel/cpu/perf_event.c
index ed6ff11..2a02529 100644
--- a/arch/x86/kernel/cpu/perf_event.c
+++ b/arch/x86/kernel/cpu/perf_event.c
@@ -197,18 +197,25 @@ struct extra_reg {
unsigned int extra_shift;
u64 config_mask;
u64 valid_mask;
+ u64 flags;
};

-#define EVENT_EXTRA_REG(e, ms, m, vm, es) { \
+#define EVENT_EXTRA_REG(e, ms, m, vm, es, f) { \
.event = (e), \
.msr = (ms), \
.config_mask = (m), \
.valid_mask = (vm), \
.extra_shift = (es), \
+ .flags = (f), \
}
#define INTEL_EVENT_EXTRA_REG(event, msr, vm, es) \
- EVENT_EXTRA_REG(event, msr, ARCH_PERFMON_EVENTSEL_EVENT, vm, es)
-#define EVENT_EXTRA_END EVENT_EXTRA_REG(0, 0, 0, 0, 0)
+ EVENT_EXTRA_REG(event, msr, ARCH_PERFMON_EVENTSEL_EVENT, vm, es, 0)
+#define INTEL_EVENT_EXTRA_REG2(event, msr, vm, es, f) \
+ EVENT_EXTRA_REG(event, msr, ARCH_PERFMON_EVENTSEL_EVENT | \
+ ARCH_PERFMON_EVENTSEL_UMASK, vm, es, f)
+#define EVENT_EXTRA_END EVENT_EXTRA_REG(0, 0, 0, 0, 0, 0)
+
+#define EXTRA_REG_LD_LAT 0x1

union perf_capabilities {
struct {
@@ -384,6 +391,11 @@ static int x86_pmu_extra_regs(u64 config, struct perf_event *event)
if (extra & ~er->valid_mask)
return -EINVAL;
event->hw.extra_config = extra;
+ event->hw.extra_flags = er->flags;
+
+ /* The minimum value that may be programmed into MSR_PEBS_LD_LAT is 3 */
+ if ((er->flags & EXTRA_REG_LD_LAT) && extra < 3)
+ event->hw.extra_config = 3;
break;
}
return 0;
diff --git a/arch/x86/kernel/cpu/perf_event_intel.c b/arch/x86/kernel/cpu/perf_event_intel.c
index bc4afb1..7e2b873 100644
--- a/arch/x86/kernel/cpu/perf_event_intel.c
+++ b/arch/x86/kernel/cpu/perf_event_intel.c
@@ -89,6 +89,8 @@ static struct event_constraint intel_nehalem_event_constraints[] =
static struct extra_reg intel_nehalem_extra_regs[] =
{
INTEL_EVENT_EXTRA_REG(0xb7, 0x1a6, 0xffff, 32), /* OFFCORE_RESPONSE */
+ /* MEM_INST_RETIRED.LATENCY_ABOVE_THRESHOLD */
+ INTEL_EVENT_EXTRA_REG2(0x100b, 0x3f6, 0xffff, 32, EXTRA_REG_LD_LAT),
EVENT_EXTRA_END
};

@@ -114,6 +116,8 @@ static struct extra_reg intel_westmere_extra_regs[] =
{
INTEL_EVENT_EXTRA_REG(0xb7, 0x1a6, 0xffff, 32), /* OFFCORE_RESPONSE_0 */
INTEL_EVENT_EXTRA_REG(0xbb, 0x1a7, 0xffff, 32), /* OFFCORE_RESPONSE_1 */
+ /* MEM_INST_RETIRED.LATENCY_ABOVE_THRESHOLD */
+ INTEL_EVENT_EXTRA_REG2(0x100b, 0x3f6, 0xffff, 32, EXTRA_REG_LD_LAT),
EVENT_EXTRA_END
};

diff --git a/arch/x86/kernel/cpu/perf_event_intel_ds.c b/arch/x86/kernel/cpu/perf_event_intel_ds.c
index b7dcd9f..d008c40 100644
--- a/arch/x86/kernel/cpu/perf_event_intel_ds.c
+++ b/arch/x86/kernel/cpu/perf_event_intel_ds.c
@@ -376,6 +376,7 @@ static struct event_constraint intel_core_pebs_events[] = {
};

static struct event_constraint intel_nehalem_pebs_events[] = {
+ PEBS_EVENT_CONSTRAINT(0x100b, 0xf), /* MEM_INST_RETIRED.LATENCY_ABOVE_THRESHOLD */
PEBS_EVENT_CONSTRAINT(0x00c0, 0xf), /* INSTR_RETIRED.ANY */
PEBS_EVENT_CONSTRAINT(0xfec1, 0xf), /* X87_OPS_RETIRED.ANY */
PEBS_EVENT_CONSTRAINT(0x00c5, 0xf), /* BR_INST_RETIRED.MISPRED */
@@ -414,6 +415,8 @@ static void intel_pmu_pebs_enable(struct perf_event *event)
hwc->config &= ~ARCH_PERFMON_EVENTSEL_INT;

cpuc->pebs_enabled |= 1ULL << hwc->idx;
+ if (hwc->extra_flags & EXTRA_REG_LD_LAT)
+ cpuc->pebs_enabled |= 1ULL << (hwc->idx + 32);
WARN_ON_ONCE(cpuc->enabled);

if (x86_pmu.intel_cap.pebs_trap && event->attr.precise_ip > 1)
@@ -426,6 +429,8 @@ static void intel_pmu_pebs_disable(struct perf_event *event)
struct hw_perf_event *hwc = &event->hw;

cpuc->pebs_enabled &= ~(1ULL << hwc->idx);
+ if (hwc->extra_flags & EXTRA_REG_LD_LAT)
+ cpuc->pebs_enabled &= ~(1ULL << (hwc->idx + 32));
if (cpuc->enabled)
wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);

diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
index d24d9ab..38bffa4 100644
--- a/include/linux/perf_event.h
+++ b/include/linux/perf_event.h
@@ -541,6 +541,7 @@ struct hw_perf_event {
int last_cpu;
unsigned int extra_reg;
u64 extra_config;
+ u64 extra_flags;
};
struct { /* software */
struct hrtimer hrtimer;

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

Next message: Richard Cochran: "Re: [PATCH V7 4/8] posix clocks: hook dynamic clocks into systemcalls"
Previous message: Michel Dänzer: "Re: [PATCH] drm/radeon: Remove DRIVER_DATE for UMS and KMS"
Next in thread: Peter Zijlstra: "Re: [RFC PATCH] perf: Add load latency monitoring on IntelNehalem/Westmere"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]