Skip to contentSkip to conversation

perf: add support for capturing skid IP

PATCH Superseded Stephane Eranian 4 participants 11 messages Nov 2, 2017 – Nov 6, 2017

Subsystems (2)
x86 perf events

v2 of 4 Superseded Latest is v4 →

lore
Share & tools
lore

Stephane Eranian

author
#0 lore raw

v2 cover perf: add support for capturing skid IP

This patchs adds a new sample record type called PERF_SAMPLE_SKID_IP. The goal is to record the unmodified interrupted instruction pointer (IP) as seen by the kernel and reflected in the machine state.

On some architectures, it is possible to avoid the IP skid using hardware support. For instance, on Intel x86, the use of PEBS helps eliminate the skid on Haswell and later processors. On older Intel processor, software, i.e., the kernel, may succeed in eliminating the skid.

Without this patch, on Haswell processors, if you set:
- attr.precise = 0, then you get the skid IP
- attr.precise = 1, then you get the skid PEBS ip (off-by-1)
- attr.precise = 2, then you get the skidless PEBS ip

The IP is captured when the event has PERF_SAMPLE_IP set in sample_type. However, there are certain measurements where you need to have BOTH the skidless IP and the skid IP. For instance, when studying branches, the skid IP usually points to the target of the branch while the skidless IP points to the branch instruction itself. Today, it is not possible to retrieve both at the same time. This patch makes this possible by specifying PERF_SAMPLE_IP|PERF_SAMPLE_SKID_IP.

As an example, consider the following code snipet:

 37.51 42c2ed    je     42c2f3
       42c2ef    add    $0x1,%rdx
       42c2f3    sub    $0x1,%rax

When using PEBS (precise=2) and sampling on BR_INST_RETIRED.CONDITIONAL, the IP always points to 0x42c2ed. With precise=1, the IP would point to 0x42c2f3. It is interesting to collect both IPs in a single run to determine how often the conditional branch is taken vs. non-taken.

Understanding the skid is also interesting for other precise events.

In V2, we rebased to 10d94ff4d558 (v4.14-rc7).

Stephane Eranian (5):
perf/core: add PERF_RECORD_SAMPLE_SKID_IP record type perf/x86: add PERF_SAMPLE_SKID_IP support for X86 PEBS perf/tools: add support for PERF_SAMPLE_SKID_IP perf/record: add support for sampling skid ip perf/script: add support for skid ip

 arch/x86/events/intel/ds.c               |  7 +++++++
 include/linux/perf_event.h               |  2 ++
 include/uapi/linux/perf_event.h          |  4 +++-
 kernel/events/core.c                     | 14 ++++++++++++++
 tools/include/uapi/linux/perf_event.h    |  4 +++-
 tools/perf/Documentation/perf-record.txt |  8 ++++++++
 tools/perf/builtin-record.c              |  2 ++
 tools/perf/builtin-script.c              |  6 ++++++
 tools/perf/perf.h                        |  1 +
 tools/perf/util/event.h                  |  1 +
 tools/perf/util/evsel.c                  | 10 ++++++++++
 tools/perf/util/session.c                |  3 +++
 12 files changed, 60 insertions(+), 2 deletions(-)
Signature · 1 line
2.7.4
author
#1 lore raw

PATCH v2 1/5 perf/core: add PERF_RECORD_SAMPLE_SKID_IP record type

From: Stephane Eranian <[email protected]>

This patchs adds a new sample record type. The goal
is to record the interrupted instruction pointer (IP)
as seen by the kernel and reflected in the machine state (pt_regs).

On some architectures, it is possible to avoid the IP skid using hardware support. For instance, on Intel x86, the use of PEBS helps eliminate the skid on Haswell and later processors.

Without this patch, on Haswell processors, if you set:
- attr.precise = 0, then you get the skid IP
- attr.precise > 0, then you get the PEBS ip corrected for skid

The IP normally comes when the event has PERF_RECORD_SAMPLE_IP set. However, there are certain measuremewnts where you need to have BOTH the corrected IP and the skid IP. For instance, when studying branches, the skid IP usually points to the target of the branch while the corrected IP point to the branch instruction itself. Today, it is not possible to retrieve both at the same time. This patch makes this possible by specifying PERF_SAMPLE_IP|PERF_SAMPLE_SKID_IP.

Signed-off-by Stephane Eranian <[email protected]>
---
 include/linux/perf_event.h      |  2 ++
 include/uapi/linux/perf_event.h |  4 +++-
 kernel/events/core.c            | 14 ++++++++++++++
 3 files changed, 19 insertions(+), 1 deletion(-)
include/linux/perf_event.h +2 -0
index 874b71a70058..772530501025 100644
@@ -917,6 +917,7 @@ struct perf_sample_data {
 	u64				stack_user_size;
 
 	u64				phys_addr;
+	u64				skid_ip;
 } ____cacheline_aligned;
 
 /* default value for data source */

@@ -937,6 +938,7 @@ static inline void perf_sample_data_init(struct perf_sample_data *data,
 	data->weight = 0;
 	data->data_src.val = PERF_MEM_NA;
 	data->txn = 0;
+	data->skid_ip = 0; /* mark as uinitialized */
 }
 
 extern void perf_output_sample(struct perf_output_handle *handle,
include/uapi/linux/perf_event.h +3 -1
index 140ae638cfd6..93e1970e9421 100644
@@ -140,8 +140,9 @@ enum perf_event_sample_format {
 	PERF_SAMPLE_TRANSACTION			= 1U << 17,
 	PERF_SAMPLE_REGS_INTR			= 1U << 18,
 	PERF_SAMPLE_PHYS_ADDR			= 1U << 19,
+	PERF_SAMPLE_SKID_IP			= 1U << 20,
 
-	PERF_SAMPLE_MAX = 1U << 20,		/* non-ABI */
+	PERF_SAMPLE_MAX = 1U << 21,		/* non-ABI */
 };
 
 /*

@@ -816,6 +817,7 @@ enum perf_event_type {
 	 *	{ u64			abi; # enum perf_sample_regs_abi
 	 *	  u64			regs[weight(mask)]; } && PERF_SAMPLE_REGS_INTR
 	 *	{ u64			phys_addr;} && PERF_SAMPLE_PHYS_ADDR
+	 *	{ u64			skid_ip;  } && PERF_SAMPLE_SKID_IP
 	 * };
 	 */
 	PERF_RECORD_SAMPLE			= 9,
kernel/events/core.c +14 -0
index 0649a84204e6..40f2839c8b94 100644
@@ -1565,6 +1565,9 @@ static void __perf_event_header_size(struct perf_event *event, u64 sample_type)
 	if (sample_type & PERF_SAMPLE_PHYS_ADDR)
 		size += sizeof(data->phys_addr);
 
+	if (sample_type & PERF_SAMPLE_SKID_IP)
+		size += sizeof(data->skid_ip);
+
 	event->header_size = size;
 }
 

@@ -5934,6 +5937,9 @@ void perf_output_sample(struct perf_output_handle *handle,
 	if (sample_type & PERF_SAMPLE_PHYS_ADDR)
 		perf_output_put(handle, data->phys_addr);
 
+	if (sample_type & PERF_SAMPLE_SKID_IP)
+		perf_output_put(handle, data->skid_ip);
+
 	if (!event->attr.watermark) {
 		int wakeup_events = event->attr.wakeup_events;
 

@@ -5999,6 +6005,14 @@ void perf_prepare_sample(struct perf_event_header *header,
 	if (sample_type & PERF_SAMPLE_IP)
 		data->ip = perf_instruction_pointer(regs);
 
+	/*
+	 * if skid_ip has not been set by arch specific code, then
+	 * we initialize it to IP as interrupt-based sampling has
+	 * skid
+	 */
+	if (!data->skid_ip && sample_type & PERF_SAMPLE_SKID_IP)
+		data->skid_ip = perf_instruction_pointer(regs);
+
 	if (sample_type & PERF_SAMPLE_CALLCHAIN) {
 		int size = 1;
Signature · 1 line
2.7.4
maintainer
#2 lore raw

Re: [PATCH v2 1/5] perf/core: add PERF_RECORD_SAMPLE_SKID_IP record type

Quoted 13 lines · Stephane Eranian
include/linux/perf_event.h +1 -0
index 874b71a70058..772530501025 100644
@@ -917,6 +917,7 @@ struct perf_sample_data {
 	u64				stack_user_size;
 
 	u64				phys_addr;
+	u64				skid_ip;
 } ____cacheline_aligned;
 
 /* default value for data source */

include/linux/perf_event.h static inline void perf_sample_data_init(struct perf_sample_data *data, · line 942

Quoted 5 lines
include/linux/perf_event.h +1 -0
@@ -937,6 +938,7 @@ static inline void perf_sample_data_init(struct perf_sample_data *data,
 	data->weight = 0;
 	data->data_src.val = PERF_MEM_NA;
 	data->txn = 0;
+	data->skid_ip = 0; /* mark as uinitialized */
 }

I keep forgetting to write the comment explaining how this function only touches a single cacheline; you just wrecked that :/

author
#3 lore raw

Re: [PATCH v2 1/5] perf/core: add PERF_RECORD_SAMPLE_SKID_IP record type

Quoted 13 lines · Peter Zijlstra
Quoted 11 lines · Stephane Eranian
include/linux/perf_event.h +1 -0
index 874b71a70058..772530501025 100644
@@ -917,6 +917,7 @@ struct perf_sample_data {
      u64                             stack_user_size;
 
      u64                             phys_addr;
+     u64                             skid_ip;
 } ____cacheline_aligned;
Quoted 10 lines

include/linux/perf_event.h static inline void perf_sample_data_init(struct perf_sample_data *data, · line 942

Quoted 7 lines
include/linux/perf_event.h +1 -0
 /* default value for data source */

@@ -937,6 +938,7 @@ static inline void perf_sample_data_init(struct perf_sample_data *data,
      data->weight = 0;
      data->data_src.val = PERF_MEM_NA;
      data->txn = 0;
+     data->skid_ip = 0; /* mark as uinitialized */
 }

I keep forgetting to write the comment explaining how this function only touches a single cacheline; you just wrecked that :/

Ok, but I need to add that field. So are you saying there is a pad I could use instead?

author
#4 lore raw

PATCH v2 2/5 perf/x86: add PERF_SAMPLE_SKID_IP support for X86 PEBS

In reply to Stephane Eranian

From: Stephane Eranian <[email protected]>

This atch adds support for SKID_IP to Intel x86 processors in PEBS mode.

Signed-off-by Stephane Eranian <[email protected]>
---
 arch/x86/events/intel/ds.c | 7 +++++++
 1 file changed, 7 insertions(+)
arch/x86/events/intel/ds.c +7 -0
index e1965e5ff570..b7afdf88f441 100644
@@ -1189,6 +1189,13 @@ static void setup_pebs_sample_data(struct perf_event *event,
 	    x86_pmu.intel_cap.pebs_format >= 1)
 		data->addr = pebs->dla;
 
+	/*
+	 * unmodified, skid IP which is guaranteed to be the next
+	 * dyanmic instruction
+	 */
+	if (sample_type & PERF_SAMPLE_SKID_IP)
+		data->skid_ip = pebs->ip;
+
 	if (x86_pmu.intel_cap.pebs_format >= 2) {
 		/* Only set the TSX weight when no memory weight. */
 		if ((sample_type & PERF_SAMPLE_WEIGHT) && !fll)
Signature · 1 line
2.7.4
#5 lore raw

Re: [PATCH v2 2/5] perf/x86: add PERF_SAMPLE_SKID_IP support for X86 PEBS

Quoted 4 lines · Stephane Eranian

From: Stephane Eranian <[email protected]>

This atch adds support for SKID_IP to Intel x86 processors in PEBS mode.

s/atch/patch/

Quoted 11 lines
Signed-off-by Stephane Eranian <[email protected]>
---
 arch/x86/events/intel/ds.c | 7 +++++++
 1 file changed, 7 insertions(+)
arch/x86/events/intel/ds.c +0 -0
index e1965e5ff570..b7afdf88f441 100644

arch/x86/events/intel/ds.c static void setup_pebs_sample_data(struct perf_event *event, · line 1195

Quoted 7 lines
arch/x86/events/intel/ds.c +4 -0
@@ -1189,6 +1189,13 @@ static void setup_pebs_sample_data(struct perf_event *event,
 	    x86_pmu.intel_cap.pebs_format >= 1)
 		data->addr = pebs->dla;
 
+	/*
+	 * unmodified, skid IP which is guaranteed to be the next
+	 * dyanmic instruction
+	 */

s/dyanmic/dynamic/

jirka

Quoted 9 lines
arch/x86/events/intel/ds.c +3 -0
+	if (sample_type & PERF_SAMPLE_SKID_IP)
+		data->skid_ip = pebs->ip;
+
 	if (x86_pmu.intel_cap.pebs_format >= 2) {
 		/* Only set the TSX weight when no memory weight. */
 		if ((sample_type & PERF_SAMPLE_WEIGHT) && !fll)
Signature · 1 line
2.7.4
author
#6 lore raw

PATCH v2 3/5 perf/tools: add support for PERF_SAMPLE_SKID_IP

In reply to Stephane Eranian

This patch adds the support code to handle the PERF_SAMPLE_SKID_IP record type.

Signed-off-by Stephane Eranian <[email protected]>
---
 tools/include/uapi/linux/perf_event.h |  4 +++-
 tools/perf/perf.h                     |  1 +
 tools/perf/util/event.h               |  1 +
 tools/perf/util/evsel.c               | 10 ++++++++++
 tools/perf/util/session.c             |  3 +++
 5 files changed, 18 insertions(+), 1 deletion(-)
tools/include/uapi/linux/perf_event.h +3 -1
index 140ae638cfd6..fd0f5111e433 100644
@@ -140,8 +140,9 @@ enum perf_event_sample_format {
 	PERF_SAMPLE_TRANSACTION			= 1U << 17,
 	PERF_SAMPLE_REGS_INTR			= 1U << 18,
 	PERF_SAMPLE_PHYS_ADDR			= 1U << 19,
+	PERF_SAMPLE_SKID_IP			= 1U << 20,
 
-	PERF_SAMPLE_MAX = 1U << 20,		/* non-ABI */
+	PERF_SAMPLE_MAX = 1U << 21,		/* non-ABI */
 };
 
 /*

@@ -816,6 +817,7 @@ enum perf_event_type {
 	 *	{ u64			abi; # enum perf_sample_regs_abi
 	 *	  u64			regs[weight(mask)]; } && PERF_SAMPLE_REGS_INTR
 	 *	{ u64			phys_addr;} && PERF_SAMPLE_PHYS_ADDR
+	 *	{ u64			skid_ip; } && PERF_SAMPLE_SKID_IP
 	 * };
 	 */
 	PERF_RECORD_SAMPLE			= 9,
tools/perf/perf.h +1 -0
index fbb0a9cd0ac6..b7637d6b34ec 100644
@@ -59,6 +59,7 @@ struct record_opts {
 	bool	     tail_synthesize;
 	bool	     overwrite;
 	bool	     ignore_missing_thread;
+	bool	     skid_ip;
 	unsigned int freq;
 	unsigned int mmap_pages;
 	unsigned int auxtrace_mmap_pages;
tools/perf/util/event.h +1 -0
index d6cbb0a0d919..e4165e824fdf 100644
@@ -201,6 +201,7 @@ struct perf_sample {
 	u32 raw_size;
 	u64 data_src;
 	u64 phys_addr;
+	u64 skid_ip;
 	u32 flags;
 	u16 insn_len;
 	u8  cpumode;
tools/perf/util/evsel.c +10 -0
index f894893c203d..08549bd92c05 100644
@@ -980,6 +980,9 @@ void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts,
 	if (opts->sample_weight)
 		perf_evsel__set_sample_bit(evsel, WEIGHT);
 
+	if (opts->skid_ip)
+		perf_evsel__set_sample_bit(evsel, SKID_IP);
+
 	attr->task  = track;
 	attr->mmap  = track;
 	attr->mmap2 = track && !perf_missing_features.mmap2;

@@ -1478,6 +1481,7 @@ static void __p_sample_type(char *buf, size_t size, u64 value)
 		bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
 		bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
 		bit_name(WEIGHT), bit_name(PHYS_ADDR),
+		bit_name(SKID_IP),
 		{ .name = NULL, }
 	};
 #undef bit_name

@@ -2225,6 +2229,12 @@ int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
 		array++;
 	}
 
+	data->skid_ip = 0;
+	if (type & PERF_SAMPLE_SKID_IP) {
+		data->skid_ip = *array;
+		array++;
+	}
+
 	return 0;
 }
tools/perf/util/session.c +3 -0
index b3fd62f7e4c9..9390cedee6f4 100644
@@ -1130,6 +1130,9 @@ static void dump_sample(struct perf_evsel *evsel, union perf_event *event,
 
 	if (sample_type & PERF_SAMPLE_READ)
 		sample_read__printf(sample, evsel->attr.read_format);
+
+	if (sample_type & PERF_SAMPLE_SKID_IP)
+		printf("... skid_ip: %" PRIu64 "\n", sample->skid_ip);
 }
 
 static void dump_read(struct perf_evsel *evsel, union perf_event *event)
Signature · 1 line
2.7.4
author
#7 lore raw

PATCH v2 4/5 perf/record: add support for sampling skid ip

In reply to Stephane Eranian

This patch adds a new --skid-ip option to perf record to capture the unmodified interrupted instruction pointer in each sample. With this option, the kernel captures both the ip and skid ip. Unless precise mode is enabled both ip are the same. They may be different in precise mode depending on the event and precise sampling mode.

$ perf record --skid-ip .....
Signed-off-by Stephane Eranian <[email protected]>
---
 tools/perf/Documentation/perf-record.txt | 8 ++++++++
 tools/perf/builtin-record.c              | 2 ++
 2 files changed, 10 insertions(+)
tools/perf/Documentation/perf-record.txt +8 -0
index 5a626ef666c2..1b2f37776d90 100644
@@ -483,6 +483,14 @@ config terms. For example: 'cycles/overwrite/' and 'instructions/no-overwrite/'.
 
 Implies --tail-synthesize.
 
+--skid-ip::
+Capture the unmodified interrupt instruction pointer (IP) in each sample. Usually
+with event-based sampling, the ip has skid and rarely point to the instruction which
+caused the event to overflow. On some architectures, the hardware can eliminate the
+skid and perf_events returns it as the IP with precise sampling is enabled. But for
+certain measurements, it may be useful to have both the correct and skid ip. This
+option enable capturing the skid ip in additional to the corrected ip. Default: off
+
 SEE ALSO
 --------
 linkperf:perf-stat[1], linkperf:perf-list[1]
tools/perf/builtin-record.c +2 -0
index a6cbf1640269..d318c0721484 100644
@@ -1585,6 +1585,8 @@ static struct option __record_options[] = {
 			  "signal"),
 	OPT_BOOLEAN(0, "dry-run", &dry_run,
 		    "Parse options then exit"),
+	OPT_BOOLEAN(0, "skid-ip", &record.opts.skid_ip,
+		    "capture skid ip in additional to ip"),
 	OPT_END()
 };
Signature · 1 line
2.7.4
#8 lore raw

Re: [PATCH v2 4/5] perf/record: add support for sampling skid ip

Quoted 8 lines · Stephane Eranian

This patch adds a new --skid-ip option to perf record to capture the unmodified interrupted instruction pointer in each sample. With this option, the kernel captures both the ip and skid ip. Unless precise mode is enabled both ip are the same. They may be different in precise mode depending on the event and precise sampling mode.

$ perf record --skid-ip .....

perf record already has too many options. I would rather just make it a event qualifier which is more flexible too

perf record -e cpu/cpu-cycles,skidip=1/

-Andi

author
#9 lore raw

PATCH v2 5/5 perf/script: add support for skid ip

In reply to Stephane Eranian

From: Stephane Eranian <[email protected]>

This patch adds a skid_ip field to perf script
to dump the raw value of the PERF_SAMPLE_SKID_IP
field in each sample.

$ perf script -F +ip,+skid_ip ......
Signed-off-by Stephane Eranian <[email protected]>
---
 tools/perf/builtin-script.c | 6 ++++++
 tools/perf/util/session.c   | 2 +-
 2 files changed, 7 insertions(+), 1 deletion(-)
tools/perf/builtin-script.c +6 -0
index a3add2cd7856..a13fbc13373b 100644
@@ -89,6 +89,7 @@ enum perf_output_field {
 	PERF_OUTPUT_SYNTH           = 1U << 25,
 	PERF_OUTPUT_PHYS_ADDR       = 1U << 26,
 	PERF_OUTPUT_UREGS	    = 1U << 27,
+	PERF_OUTPUT_SKID_IP	    = 1U << 28,
 };
 
 struct output_option {

@@ -123,6 +124,7 @@ struct output_option {
 	{.str = "brstackoff", .field = PERF_OUTPUT_BRSTACKOFF},
 	{.str = "synth", .field = PERF_OUTPUT_SYNTH},
 	{.str = "phys_addr", .field = PERF_OUTPUT_PHYS_ADDR},
+	{.str = "skid_ip", .field = PERF_OUTPUT_SKID_IP},
 };
 
 enum {

@@ -1514,6 +1516,10 @@ static void process_event(struct perf_script *script,
 
 	if (PRINT_FIELD(PHYS_ADDR))
 		printf("%16" PRIx64, sample->phys_addr);
+
+	if (PRINT_FIELD(SKID_IP))
+		printf(" %"PRIx64" ", sample->skid_ip);
+
 	printf("\n");
 }
tools/perf/util/session.c +1 -1
index 9390cedee6f4..802227735a89 100644
@@ -1132,7 +1132,7 @@ static void dump_sample(struct perf_evsel *evsel, union perf_event *event,
 		sample_read__printf(sample, evsel->attr.read_format);
 
 	if (sample_type & PERF_SAMPLE_SKID_IP)
-		printf("... skid_ip: %" PRIu64 "\n", sample->skid_ip);
+		printf("... skid_ip: 0x%" PRIx64 "\n", sample->skid_ip);
 }
 
 static void dump_read(struct perf_evsel *evsel, union perf_event *event)
Signature · 1 line
2.7.4
#10 lore raw

Re: [PATCH v2 5/5] perf/script: add support for skid ip

Quoted 6 lines · Stephane Eranian

From: Stephane Eranian <[email protected]>

This patch adds a skid_ip field to perf script
to dump the raw value of the PERF_SAMPLE_SKID_IP
field in each sample.

Quoted 6 lines
$ perf script -F +ip,+skid_ip ......
Signed-off-by Stephane Eranian <[email protected]>
---
 tools/perf/builtin-script.c | 6 ++++++
 tools/perf/util/session.c   | 2 +-

please update the doc as well (-F option section)

thanks,
jirka