Skip to contentSkip to conversation

perf/amd: Zen4 IBS extensions support (tool changes)

PATCH Applied Ravi Bangoria 4 participants 13 messages Jun 4, 2022 – Jul 27, 2022

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

v6 cover perf/amd: Zen4 IBS extensions support (tool changes)

Kernel side of changes have already been applied to linus/master (except amd-ibs.h header). This series contains perf tool changes.

v5: https://lore.kernel.org/lkml/[email protected]
v5->v6:
- Use macros instead of magic numbers for IBS l3missonly bits
- Use asprintf() instead of allocating memory and copying data manually
- Add Reviewed-by Kan Liang (patch 2-5).

Patches prepared on acme/perf/core (9dde6cadb92b5)

Original cover letter:

IBS support has been enhanced with two new features in upcoming uarch:
1. DataSrc extension and 2. L3 Miss Filtering capability. Both are indicated by CPUID_Fn8000001B_EAX bit 11.

DataSrc extension provides additional data source details for tagged load/store operations. Add support for these new bits in perf report/ script raw-dump.

IBS L3 miss filtering works by tagging an instruction on IBS counter overflow and generating an NMI if the tagged instruction causes an L3 miss. Samples without an L3 miss are discarded and counter is reset with random value (between 1-15 for fetch pmu and 1-127 for op pmu). This helps in reducing sampling overhead when user is interested only in such samples. One of the use case of such filtered samples is to feed data to page-migration daemon in tiered memory systems.

Add support for L3 miss filtering in IBS driver via new pmu attribute "l3missonly". Example usage:

  # perf record -a -e ibs_op/l3missonly=1/ --raw-samples sleep 5
  # perf report -D

Some important points to keep in mind while using L3 miss filtering:
1. Hw internally reset sampling period when tagged instruction does not cause L3 miss. But there is no way to reconstruct aggregated sampling period when this happens.
2. L3 miss is not the actual event being counted. Rather, IBS will count fetch, cycles or uOps depending on the configuration. Thus sampling period have no direct connection to L3 misses.

1st causes sampling period skew. Thus, I've added warning message at perf record:

  # perf record -c 10000 -C 0 -e ibs_op/l3missonly=1/
  WARNING: Hw internally resets sampling period when L3 Miss Filtering is enabled
  and tagged operation does not cause L3 Miss. This causes sampling period skew.

User can configure smaller sampling period to get more samples while using l3missonly.

Ravi Bangoria (8):
perf record ibs: Warn about sampling period skew perf tool: Parse pmu caps sysfs only once
perf headers: Pass "cpu" pmu name while printing caps perf headers: Store pmu caps in an array of strings perf headers: Record non-cpu pmu capabilities perf/x86/ibs: Add new IBS register bits into header perf tool ibs: Sync amd ibs header file
perf script ibs: Support new IBS bits in raw trace dump

 arch/x86/include/asm/amd-ibs.h                |  16 +-
 tools/arch/x86/include/asm/amd-ibs.h          |  16 +-
 .../Documentation/perf.data-file-format.txt   |  10 +-
 tools/perf/arch/x86/util/evsel.c              |  52 +++++
 tools/perf/builtin-inject.c                   |   2 +-
 tools/perf/util/amd-sample-raw.c              |  68 ++++++-
 tools/perf/util/env.c                         |  62 +++++-
 tools/perf/util/env.h                         |  14 +-
 tools/perf/util/evsel.c                       |   7 +
 tools/perf/util/evsel.h                       |   1 +
 tools/perf/util/header.c                      | 189 ++++++++++--------
 tools/perf/util/header.h                      |   2 +-
 tools/perf/util/pmu.c                         |  15 +-
 tools/perf/util/pmu.h                         |   2 +
 14 files changed, 329 insertions(+), 127 deletions(-)
Signature · 1 line
2.31.1
#1 lore raw

PATCH v6 1/8 perf record ibs: Warn about sampling period skew

Samples without an L3 miss are discarded and counter is reset with random value (between 1-15 for fetch pmu and 1-127 for op pmu) when IBS L3 miss filtering is enabled. This causes a sampling period skew but there is no way to reconstruct aggregated sampling period. So print a warning at perf record if user sets l3missonly=1.

Ex:

  # perf record -c 10000 -C 0 -e ibs_op/l3missonly=1/
  WARNING: Hw internally resets sampling period when L3 Miss Filtering is enabled
  and tagged operation does not cause L3 Miss. This causes sampling period skew.
---
 tools/perf/arch/x86/util/evsel.c | 52 ++++++++++++++++++++++++++++++++
 tools/perf/util/evsel.c          |  7 +++++
 tools/perf/util/evsel.h          |  1 +
 3 files changed, 60 insertions(+)
tools/perf/arch/x86/util/evsel.c +52 -0
index 88306183d629..53763e583804 100644
@@ -5,6 +5,10 @@
 #include "util/env.h"
 #include "util/pmu.h"
 #include "linux/string.h"
+#include "util/debug.h"
+
+#define IBS_FETCH_L3MISSONLY	(1ULL << 59)
+#define IBS_OP_L3MISSONLY	(1ULL << 16)
 
 void arch_evsel__set_sample_weight(struct evsel *evsel)
 {
@@ -60,3 +64,51 @@ bool arch_evsel__must_be_in_group(const struct evsel *evsel)
 		(!strcasecmp(evsel->name, "slots") ||
 		 strcasestr(evsel->name, "topdown"));
 }
+
+static void ibs_l3miss_warn(void)
+{
+	pr_warning(
+"WARNING: Hw internally resets sampling period when L3 Miss Filtering is enabled\n"
+"and tagged operation does not cause L3 Miss. This causes sampling period skew.\n");
+}
+
+void arch__post_evsel_config(struct evsel *evsel, struct perf_event_attr *attr)
+{
+	struct perf_pmu *evsel_pmu, *ibs_fetch_pmu, *ibs_op_pmu;
+	static int warned_once;
+	/* 0: Uninitialized, 1: Yes, -1: No */
+	static int is_amd;
+
+	if (warned_once || is_amd == -1)
+		return;
+
+	if (!is_amd) {
+		struct perf_env *env = evsel__env(evsel);
+
+		if (!perf_env__cpuid(env) || !env->cpuid ||
+		    !strstarts(env->cpuid, "AuthenticAMD")) {
+			is_amd = -1;
+			return;
+		}
+		is_amd = 1;
+	}
+
+	evsel_pmu = evsel__find_pmu(evsel);
+	if (!evsel_pmu)
+		return;
+
+	ibs_fetch_pmu = perf_pmu__find("ibs_fetch");
+	ibs_op_pmu = perf_pmu__find("ibs_op");
+
+	if (ibs_fetch_pmu && ibs_fetch_pmu->type == evsel_pmu->type) {
+		if (attr->config & IBS_FETCH_L3MISSONLY) {
+			ibs_l3miss_warn();
+			warned_once = 1;
+		}
+	} else if (ibs_op_pmu && ibs_op_pmu->type == evsel_pmu->type) {
+		if (attr->config & IBS_OP_L3MISSONLY) {
+			ibs_l3miss_warn();
+			warned_once = 1;
+		}
+	}
+}
tools/perf/util/evsel.c +7 -0
index ce499c5da8d7..8fea51a9cd90 100644
@@ -1091,6 +1091,11 @@ void __weak arch_evsel__fixup_new_cycles(struct perf_event_attr *attr __maybe_un
 {
 }
 
+void __weak arch__post_evsel_config(struct evsel *evsel __maybe_unused,
+				    struct perf_event_attr *attr __maybe_unused)
+{
+}
+
 static void evsel__set_default_freq_period(struct record_opts *opts,
 					   struct perf_event_attr *attr)
 {
@@ -1366,6 +1371,8 @@ void evsel__config(struct evsel *evsel, struct record_opts *opts,
 	 */
 	if (evsel__is_dummy_event(evsel))
 		evsel__reset_sample_bit(evsel, BRANCH_STACK);
+
+	arch__post_evsel_config(evsel, attr);
 }
 
 int evsel__set_filter(struct evsel *evsel, const char *filter)
tools/perf/util/evsel.h +1 -0
index 73ea48e94079..92bed8e2f7d8 100644
@@ -297,6 +297,7 @@ void evsel__set_sample_id(struct evsel *evsel, bool use_sample_identifier);
 
 void arch_evsel__set_sample_weight(struct evsel *evsel);
 void arch_evsel__fixup_new_cycles(struct perf_event_attr *attr);
+void arch__post_evsel_config(struct evsel *evsel, struct perf_event_attr *attr);
 
 int evsel__set_filter(struct evsel *evsel, const char *filter);
 int evsel__append_tp_filter(struct evsel *evsel, const char *filter);
Signature · 1 line
2.31.1
#2 lore raw

PATCH v6 2/8 perf tool: Parse pmu caps sysfs only once

In reply to Ravi Bangoria

In addition to returning nr_caps, cache it locally in struct perf_pmu. Similarly, cache status of whether caps sysfs has already been parsed or not. These will help to avoid parsing sysfs every time the function gets called.

---
 tools/perf/util/pmu.c | 15 +++++++++++----
 tools/perf/util/pmu.h |  2 ++
 2 files changed, 13 insertions(+), 4 deletions(-)
tools/perf/util/pmu.c +11 -4
index 9a1c7e63e663..0112e1c36418 100644
@@ -1890,7 +1890,11 @@ int perf_pmu__caps_parse(struct perf_pmu *pmu)
 	const char *sysfs = sysfs__mountpoint();
 	DIR *caps_dir;
 	struct dirent *evt_ent;
-	int nr_caps = 0;
+
+	if (pmu->caps_initialized)
+		return pmu->nr_caps;
+
+	pmu->nr_caps = 0;
 
 	if (!sysfs)
 		return -1;
@@ -1898,8 +1902,10 @@ int perf_pmu__caps_parse(struct perf_pmu *pmu)
 	snprintf(caps_path, PATH_MAX,
 		 "%s" EVENT_SOURCE_DEVICE_PATH "%s/caps", sysfs, pmu->name);
 
-	if (stat(caps_path, &st) < 0)
+	if (stat(caps_path, &st) < 0) {
+		pmu->caps_initialized = true;
 		return 0;	/* no error if caps does not exist */
+	}
 
 	caps_dir = opendir(caps_path);
 	if (!caps_dir)
@@ -1926,13 +1932,14 @@ int perf_pmu__caps_parse(struct perf_pmu *pmu)
 			continue;
 		}
 
-		nr_caps++;
+		pmu->nr_caps++;
 		fclose(file);
 	}
 
 	closedir(caps_dir);
 
-	return nr_caps;
+	pmu->caps_initialized = true;
+	return pmu->nr_caps;
 }
 
 void perf_pmu__warn_invalid_config(struct perf_pmu *pmu, __u64 config,
tools/perf/util/pmu.h +2 -0
index 541889fa9f9c..4b45fd8da5a3 100644
@@ -46,6 +46,8 @@ struct perf_pmu {
 	struct perf_cpu_map *cpus;
 	struct list_head format;  /* HEAD struct perf_pmu_format -> list */
 	struct list_head aliases; /* HEAD struct perf_pmu_alias -> list */
+	bool caps_initialized;
+	u32 nr_caps;
 	struct list_head caps;    /* HEAD struct perf_pmu_caps -> list */
 	struct list_head list;    /* ELEM */
 	struct list_head hybrid_list;
Signature · 1 line
2.31.1
#3 lore raw

PATCH v6 3/8 perf headers: Pass "cpu" pmu name while printing caps

In reply to Ravi Bangoria

Avoid unnecessary conditional code to check if pmu name is NULL or not by passing "cpu" pmu name to the printing function.

---
 tools/perf/util/header.c | 12 +++---------
 1 file changed, 3 insertions(+), 9 deletions(-)
tools/perf/util/header.c +3 -9
index 53332da100e8..ee7ccd94e272 100644
@@ -2058,17 +2058,11 @@ static void print_per_cpu_pmu_caps(FILE *fp, int nr_caps, char *cpu_pmu_caps,
 	char *str, buf[128];
 
 	if (!nr_caps) {
-		if (!pmu_name)
-			fprintf(fp, "# cpu pmu capabilities: not available\n");
-		else
-			fprintf(fp, "# %s pmu capabilities: not available\n", pmu_name);
+		fprintf(fp, "# %s pmu capabilities: not available\n", pmu_name);
 		return;
 	}
 
-	if (!pmu_name)
-		scnprintf(buf, sizeof(buf), "# cpu pmu capabilities: ");
-	else
-		scnprintf(buf, sizeof(buf), "# %s pmu capabilities: ", pmu_name);
+	scnprintf(buf, sizeof(buf), "# %s pmu capabilities: ", pmu_name);
 
 	delimiter = buf;
 
@@ -2085,7 +2079,7 @@ static void print_per_cpu_pmu_caps(FILE *fp, int nr_caps, char *cpu_pmu_caps,
 static void print_cpu_pmu_caps(struct feat_fd *ff, FILE *fp)
 {
 	print_per_cpu_pmu_caps(fp, ff->ph->env.nr_cpu_pmu_caps,
-			       ff->ph->env.cpu_pmu_caps, NULL);
+			       ff->ph->env.cpu_pmu_caps, (char *)"cpu");
 }
 
 static void print_hybrid_cpu_pmu_caps(struct feat_fd *ff, FILE *fp)
Signature · 1 line
2.31.1
#4 lore raw

PATCH v6 4/8 perf headers: Store pmu caps in an array of strings

In reply to Ravi Bangoria

Currently all capabilities are stored in a single string separated by NULL character. Instead, store them in an array which makes searching of capability easier.

---
 tools/perf/util/env.c    |  6 +++-
 tools/perf/util/env.h    |  4 +--
 tools/perf/util/header.c | 63 ++++++++++++++++++++++------------------
 3 files changed, 41 insertions(+), 32 deletions(-)
tools/perf/util/env.c +5 -1
index 579e44c59914..7d3aeb2e4622 100644
@@ -179,7 +179,7 @@ static void perf_env__purge_bpf(struct perf_env *env __maybe_unused)
 
 void perf_env__exit(struct perf_env *env)
 {
-	int i;
+	int i, j;
 
 	perf_env__purge_bpf(env);
 	perf_env__purge_cgroups(env);
@@ -196,6 +196,8 @@ void perf_env__exit(struct perf_env *env)
 	zfree(&env->sibling_threads);
 	zfree(&env->pmu_mappings);
 	zfree(&env->cpu);
+	for (i = 0; i < env->nr_cpu_pmu_caps; i++)
+		zfree(&env->cpu_pmu_caps[i]);
 	zfree(&env->cpu_pmu_caps);
 	zfree(&env->numa_map);
 
@@ -218,6 +220,8 @@ void perf_env__exit(struct perf_env *env)
 	zfree(&env->hybrid_nodes);
 
 	for (i = 0; i < env->nr_hybrid_cpc_nodes; i++) {
+		for (j = 0; j < env->hybrid_cpc_nodes[i].nr_cpu_pmu_caps; j++)
+			zfree(&env->hybrid_cpc_nodes[i].cpu_pmu_caps[j]);
 		zfree(&env->hybrid_cpc_nodes[i].cpu_pmu_caps);
 		zfree(&env->hybrid_cpc_nodes[i].pmu_name);
 	}
tools/perf/util/env.h +2 -2
index a3541f98e1fc..43aab59f7322 100644
@@ -46,7 +46,7 @@ struct hybrid_node {
 struct hybrid_cpc_node {
 	int		nr_cpu_pmu_caps;
 	unsigned int    max_branches;
-	char            *cpu_pmu_caps;
+	char            **cpu_pmu_caps;
 	char            *pmu_name;
 };
 
@@ -81,7 +81,7 @@ struct perf_env {
 	char			*sibling_dies;
 	char			*sibling_threads;
 	char			*pmu_mappings;
-	char			*cpu_pmu_caps;
+	char			**cpu_pmu_caps;
 	struct cpu_topology_map	*cpu;
 	struct cpu_cache_level	*caches;
 	int			 caches_cnt;
tools/perf/util/header.c +34 -29

140 lines not on this page · raw message

Signature · 1 line
2.31.1
#5 lore raw

PATCH v6 5/8 perf headers: Record non-cpu pmu capabilities

In reply to Ravi Bangoria

Pmus advertise their capabilities via sysfs attribute files but perf tool currently parses only core(cpu) or hybrid core pmu capabilities. Add support of recording non-core pmu capabilities int perf.data header.

Note that a newly proposed HEADER_PMU_CAPS is replacing existing HEADER_HYBRID_CPU_PMU_CAPS. Special care is taken for hybrid core pmus by writing their capabilities first in the perf.data header to make sure new perf.data file being read by old perf tool does not break.

---
 .../Documentation/perf.data-file-format.txt   |  10 +-
 tools/perf/builtin-inject.c                   |   2 +-
 tools/perf/util/env.c                         |  60 ++++++-
 tools/perf/util/env.h                         |  12 +-
 tools/perf/util/header.c                      | 160 ++++++++++--------
 tools/perf/util/header.h                      |   2 +-
 6 files changed, 158 insertions(+), 88 deletions(-)
tools/perf/Documentation/perf.data-file-format.txt +6 -4
index f56d0e0fbff6..2233534e508a 100644
@@ -419,18 +419,20 @@ Example:
   cpu_core cpu list : 0-15
   cpu_atom cpu list : 16-23
 
-	HEADER_HYBRID_CPU_PMU_CAPS = 31,
+	HEADER_PMU_CAPS = 31,
 
-	A list of hybrid CPU PMU capabilities.
+	List of pmu capabilities (except cpu pmu which is already
+	covered by HEADER_CPU_PMU_CAPS). Note that hybrid cpu pmu
+	capabilities are also stored here.
 
 struct {
 	u32 nr_pmu;
 	struct {
-		u32 nr_cpu_pmu_caps;
+		u32 nr_caps;
 		{
 			char	name[];
 			char	value[];
-		} [nr_cpu_pmu_caps];
+		} [nr_caps];
 		char pmu_name[];
 	} [nr_pmu];
 };
tools/perf/builtin-inject.c +1 -1
index a75bf11585b5..05bc0cfccf83 100644
@@ -809,7 +809,7 @@ static bool keep_feat(int feat)
 	case HEADER_CPU_PMU_CAPS:
 	case HEADER_CLOCK_DATA:
 	case HEADER_HYBRID_TOPOLOGY:
-	case HEADER_HYBRID_CPU_PMU_CAPS:
+	case HEADER_PMU_CAPS:
 		return true;
 	/* Information that can be updated */
 	case HEADER_BUILD_ID:
tools/perf/util/env.c +54 -6

70 lines not on this page · raw message

tools/perf/util/env.h +7 -5
index 43aab59f7322..4566c51f2fd9 100644
@@ -43,10 +43,10 @@ struct hybrid_node {
 	char	*cpus;
 };
 
-struct hybrid_cpc_node {
-	int		nr_cpu_pmu_caps;
+struct pmu_caps {
+	int		nr_caps;
 	unsigned int    max_branches;
-	char            **cpu_pmu_caps;
+	char            **caps;
 	char            *pmu_name;
 };
 
@@ -74,7 +74,7 @@ struct perf_env {
 	int			nr_groups;
 	int			nr_cpu_pmu_caps;
 	int			nr_hybrid_nodes;
-	int			nr_hybrid_cpc_nodes;
+	int			nr_pmus_with_caps;
 	char			*cmdline;
 	const char		**cmdline_argv;
 	char			*sibling_cores;
@@ -94,7 +94,7 @@ struct perf_env {
 	struct memory_node	*memory_nodes;
 	unsigned long long	 memory_bsize;
 	struct hybrid_node	*hybrid_nodes;
-	struct hybrid_cpc_node	*hybrid_cpc_nodes;
+	struct pmu_caps		*pmu_caps;
 #ifdef HAVE_LIBBPF_SUPPORT
 	/*
 	 * bpf_info_lock protects bpf rbtrees. This is needed because the
@@ -172,4 +172,6 @@ bool perf_env__insert_btf(struct perf_env *env, struct btf_node *btf_node);
 struct btf_node *perf_env__find_btf(struct perf_env *env, __u32 btf_id);
 
 int perf_env__numa_node(struct perf_env *env, struct perf_cpu cpu);
+char *perf_env__find_pmu_cap(struct perf_env *env, const char *pmu_name,
+			     const char *cap);
 #endif /* __PERF_ENV_H */
tools/perf/util/header.c +89 -71

296 lines not on this page · raw message

tools/perf/util/header.h +1 -1
index 08563c1f1bff..5790bf556b90 100644
@@ -46,7 +46,7 @@ enum {
 	HEADER_CPU_PMU_CAPS,
 	HEADER_CLOCK_DATA,
 	HEADER_HYBRID_TOPOLOGY,
-	HEADER_HYBRID_CPU_PMU_CAPS,
+	HEADER_PMU_CAPS,
 	HEADER_LAST_FEATURE,
 	HEADER_FEAT_BITS	= 256,
 };
Signature · 1 line
2.31.1
#6 lore raw

PATCH v6 6/8 perf/x86/ibs: Add new IBS register bits into header

In reply to Ravi Bangoria

IBS support has been enhanced with two new features in upcoming uarch:
1. DataSrc extension and 2. L3 miss filtering. Additional set of bits has been introduced in IBS registers to exploit these features. Define these new bits into arch/x86/ header.

---
 arch/x86/include/asm/amd-ibs.h | 16 ++++++++++------
 1 file changed, 10 insertions(+), 6 deletions(-)
arch/x86/include/asm/amd-ibs.h +10 -6
index aabdbb5ab920..f3eb098d63d4 100644
@@ -29,7 +29,10 @@ union ibs_fetch_ctl {
 			rand_en:1,	/* 57: random tagging enable */
 			fetch_l2_miss:1,/* 58: L2 miss for sampled fetch
 					 *      (needs IbsFetchComp) */
-			reserved:5;	/* 59-63: reserved */
+			l3_miss_only:1,	/* 59: Collect L3 miss samples only */
+			fetch_oc_miss:1,/* 60: Op cache miss for the sampled fetch */
+			fetch_l3_miss:1,/* 61: L3 cache miss for the sampled fetch */
+			reserved:2;	/* 62-63: reserved */
 	};
 };
 
@@ -38,14 +41,14 @@ union ibs_op_ctl {
 	__u64 val;
 	struct {
 		__u64	opmaxcnt:16,	/* 0-15: periodic op max. count */
-			reserved0:1,	/* 16: reserved */
+			l3_miss_only:1,	/* 16: Collect L3 miss samples only */
 			op_en:1,	/* 17: op sampling enable */
 			op_val:1,	/* 18: op sample valid */
 			cnt_ctl:1,	/* 19: periodic op counter control */
 			opmaxcnt_ext:7,	/* 20-26: upper 7 bits of periodic op maximum count */
-			reserved1:5,	/* 27-31: reserved */
+			reserved0:5,	/* 27-31: reserved */
 			opcurcnt:27,	/* 32-58: periodic op counter current count */
-			reserved2:5;	/* 59-63: reserved */
+			reserved1:5;	/* 59-63: reserved */
 	};
 };
 
@@ -71,11 +74,12 @@ union ibs_op_data {
 union ibs_op_data2 {
 	__u64 val;
 	struct {
-		__u64	data_src:3,	/* 0-2: data source */
+		__u64	data_src_lo:3,	/* 0-2: data source low */
 			reserved0:1,	/* 3: reserved */
 			rmt_node:1,	/* 4: destination node */
 			cache_hit_st:1,	/* 5: cache hit state */
-			reserved1:57;	/* 5-63: reserved */
+			data_src_hi:2,	/* 6-7: data source high */
+			reserved1:56;	/* 8-63: reserved */
 	};
 };
 
Signature · 1 line
2.31.1
merge bot Applied
#7 lore raw

[tip: perf/core] perf/x86/ibs: Add new IBS register bits into header

The following commit has been merged into the perf/core branch of tip:

The following commit has been merged into the perf/core branch of tip:

Commit-ID:     326ecc15c61c349cd49d1700ff9e3e31c6fd1cd5
Gitweb:        https://git.kernel.org/tip/326ecc15c61c349cd49d1700ff9e3e31c6fd1cd5
Author:        Ravi Bangoria <[email protected]>
AuthorDate:    Sat, 04 Jun 2022 10:15:17 +05:30
Committer:     Borislav Petkov <[email protected]>
CommitterDate: Wed, 27 Jul 2022 13:54:38 +02:00

perf/x86/ibs: Add new IBS register bits into header

IBS support has been enhanced with two new features in upcoming uarch:

1. DataSrc extension and
2. L3 miss filtering.

Additional set of bits has been introduced in IBS registers to use these features. Define these new bits into arch/x86/ header.

[ bp: Massage commit message. ]

---
 arch/x86/include/asm/amd-ibs.h | 16 ++++++++++------
 1 file changed, 10 insertions(+), 6 deletions(-)
arch/x86/include/asm/amd-ibs.h +10 -6
index aabdbb5..f3eb098 100644
@@ -29,7 +29,10 @@ union ibs_fetch_ctl {
 			rand_en:1,	/* 57: random tagging enable */
 			fetch_l2_miss:1,/* 58: L2 miss for sampled fetch
 					 *      (needs IbsFetchComp) */
-			reserved:5;	/* 59-63: reserved */
+			l3_miss_only:1,	/* 59: Collect L3 miss samples only */
+			fetch_oc_miss:1,/* 60: Op cache miss for the sampled fetch */
+			fetch_l3_miss:1,/* 61: L3 cache miss for the sampled fetch */
+			reserved:2;	/* 62-63: reserved */
 	};
 };
 
@@ -38,14 +41,14 @@ union ibs_op_ctl {
 	__u64 val;
 	struct {
 		__u64	opmaxcnt:16,	/* 0-15: periodic op max. count */
-			reserved0:1,	/* 16: reserved */
+			l3_miss_only:1,	/* 16: Collect L3 miss samples only */
 			op_en:1,	/* 17: op sampling enable */
 			op_val:1,	/* 18: op sample valid */
 			cnt_ctl:1,	/* 19: periodic op counter control */
 			opmaxcnt_ext:7,	/* 20-26: upper 7 bits of periodic op maximum count */
-			reserved1:5,	/* 27-31: reserved */
+			reserved0:5,	/* 27-31: reserved */
 			opcurcnt:27,	/* 32-58: periodic op counter current count */
-			reserved2:5;	/* 59-63: reserved */
+			reserved1:5;	/* 59-63: reserved */
 	};
 };
 
@@ -71,11 +74,12 @@ union ibs_op_data {
 union ibs_op_data2 {
 	__u64 val;
 	struct {
-		__u64	data_src:3,	/* 0-2: data source */
+		__u64	data_src_lo:3,	/* 0-2: data source low */
 			reserved0:1,	/* 3: reserved */
 			rmt_node:1,	/* 4: destination node */
 			cache_hit_st:1,	/* 5: cache hit state */
-			reserved1:57;	/* 5-63: reserved */
+			data_src_hi:2,	/* 6-7: data source high */
+			reserved1:56;	/* 8-63: reserved */
 	};
 };
 
#8 lore raw

PATCH v6 7/8 perf tool ibs: Sync amd ibs header file

In reply to Ravi Bangoria

IBS support has been enhanced with two new features in upcoming uarch:
1. DataSrc extension and 2. L3 miss filtering. Additional set of bits has been introduced in IBS registers to exploit these features. New bits are already defining in arch/x86/ header. Sync it with tools header file. Also rename existing ibs_op_data field 'data_src' to 'data_src_lo'.

---
 tools/arch/x86/include/asm/amd-ibs.h | 16 ++++++++++------
 tools/perf/util/amd-sample-raw.c     |  4 ++--
 2 files changed, 12 insertions(+), 8 deletions(-)
tools/arch/x86/include/asm/amd-ibs.h +10 -6
index 765e9e752d03..9a3312e12e2e 100644
@@ -29,7 +29,10 @@ union ibs_fetch_ctl {
 			rand_en:1,	/* 57: random tagging enable */
 			fetch_l2_miss:1,/* 58: L2 miss for sampled fetch
 					 *      (needs IbsFetchComp) */
-			reserved:5;	/* 59-63: reserved */
+			l3_miss_only:1,	/* 59: Collect L3 miss samples only */
+			fetch_oc_miss:1,/* 60: Op cache miss for the sampled fetch */
+			fetch_l3_miss:1,/* 61: L3 cache miss for the sampled fetch */
+			reserved:2;	/* 62-63: reserved */
 	};
 };
 
@@ -38,14 +41,14 @@ union ibs_op_ctl {
 	__u64 val;
 	struct {
 		__u64	opmaxcnt:16,	/* 0-15: periodic op max. count */
-			reserved0:1,	/* 16: reserved */
+			l3_miss_only:1,	/* 16: Collect L3 miss samples only */
 			op_en:1,	/* 17: op sampling enable */
 			op_val:1,	/* 18: op sample valid */
 			cnt_ctl:1,	/* 19: periodic op counter control */
 			opmaxcnt_ext:7,	/* 20-26: upper 7 bits of periodic op maximum count */
-			reserved1:5,	/* 27-31: reserved */
+			reserved0:5,	/* 27-31: reserved */
 			opcurcnt:27,	/* 32-58: periodic op counter current count */
-			reserved2:5;	/* 59-63: reserved */
+			reserved1:5;	/* 59-63: reserved */
 	};
 };
 
@@ -71,11 +74,12 @@ union ibs_op_data {
 union ibs_op_data2 {
 	__u64 val;
 	struct {
-		__u64	data_src:3,	/* 0-2: data source */
+		__u64	data_src_lo:3,	/* 0-2: data source low */
 			reserved0:1,	/* 3: reserved */
 			rmt_node:1,	/* 4: destination node */
 			cache_hit_st:1,	/* 5: cache hit state */
-			reserved1:57;	/* 5-63: reserved */
+			data_src_hi:2,	/* 6-7: data source high */
+			reserved1:56;	/* 8-63: reserved */
 	};
 };
 
tools/perf/util/amd-sample-raw.c +2 -2
index d19d765195c5..3b623ea6ee7e 100644
@@ -98,9 +98,9 @@ static void pr_ibs_op_data2(union ibs_op_data2 reg)
 	};
 
 	printf("ibs_op_data2:\t%016llx %sRmtNode %d%s\n", reg.val,
-	       reg.data_src == 2 ? (reg.cache_hit_st ? "CacheHitSt 1=O-State "
+	       reg.data_src_lo == 2 ? (reg.cache_hit_st ? "CacheHitSt 1=O-State "
 						     : "CacheHitSt 0=M-state ") : "",
-	       reg.rmt_node, data_src_str[reg.data_src]);
+	       reg.rmt_node, data_src_str[reg.data_src_lo]);
 }
 
 static void pr_ibs_op_data3(union ibs_op_data3 reg)
Signature · 1 line
2.31.1
#9 lore raw

PATCH v6 8/8 perf script ibs: Support new IBS bits in raw trace dump

In reply to Ravi Bangoria

Interpret Additional set of IBS register bits while doing perf report/script raw dump.

IBS op pmu ex:

  $ sudo ./perf record -c 130 -a -e ibs_op/l3missonly=1/ --raw-samples
  $ sudo ./perf report -D
  ...
  ibs_op_ctl:     0000004500070008 MaxCnt       128 L3MissOnly 1 En 1
        Val 1 CntCtl 0=cycles CurCnt        69
  ibs_op_data:    0000000000710002 CompToRetCtr     2 TagToRetCtr   113
        BrnRet 0  RipInvalid 0 BrnFuse 0 Microcode 0
  ibs_op_data2:   0000000000000002 CacheHitSt 0=M-state RmtNode 0
        DataSrc 2=A peer cache in a near CCX
  ibs_op_data3:   000000681d1700a1 LdOp 1 StOp 0 DcL1TlbMiss 0
        DcL2TlbMiss 0 DcL1TlbHit2M 0 DcL1TlbHit1G 1 DcL2TlbHit2M 0
        DcMiss 1 DcMisAcc 0 DcWcMemAcc 0 DcUcMemAcc 0 DcLockedOp 0
        DcMissNoMabAlloc 1 DcLinAddrValid 1 DcPhyAddrValid 1
        DcL2TlbHit1G 0 L2Miss 1 SwPf 0 OpMemWidth  8 bytes
        OpDcMissOpenMemReqs  7 DcMissLat   104 TlbRefillLat     0

IBS Fetch pmu ex:

  $ sudo ./perf record -c 130 -a -e ibs_fetch/l3missonly=1/ --raw-samples
  $ sudo ./perf report -D
  ...
  ibs_fetch_ctl:  3c1f00c700080008 MaxCnt     128 Cnt     128 Lat   199
        En 1 Val 1 Comp 1 IcMiss 1 PhyAddrValid        1 L1TlbPgSz 4KB
        L1TlbMiss 0 L2TlbMiss 0 RandEn 0 L2Miss 1 L3MissOnly 1
        FetchOcMiss 1 FetchL3Miss 1

With the DataSrc extensions, the source of data can be decoded among:
- Local L3 or other L1/L2 in CCX.
- A peer cache in a near CCX.
- Data returned from DRAM.
- A peer cache in a far CCX.
- DRAM address map with "long latency" bit set.
- Data returned from MMIO/Config/PCI/APIC.
- Extension Memory (S-Link, GenZ, etc - identified by the CS target and/or address map at DF's choice).
- Peer Agent Memory.

---
 tools/perf/util/amd-sample-raw.c | 64 +++++++++++++++++++++++++++++---
 1 file changed, 58 insertions(+), 6 deletions(-)
tools/perf/util/amd-sample-raw.c +58 -6
index 3b623ea6ee7e..238305868644 100644
@@ -18,6 +18,7 @@
 #include "pmu-events/pmu-events.h"
 
 static u32 cpu_family, cpu_model, ibs_fetch_type, ibs_op_type;
+static bool zen4_ibs_extensions;
 
 static void pr_ibs_fetch_ctl(union ibs_fetch_ctl reg)
 {
@@ -39,6 +40,7 @@ static void pr_ibs_fetch_ctl(union ibs_fetch_ctl reg)
 	};
 	const char *ic_miss_str = NULL;
 	const char *l1tlb_pgsz_str = NULL;
+	char l3_miss_str[sizeof(" L3MissOnly _ FetchOcMiss _ FetchL3Miss _")] = "";
 
 	if (cpu_family == 0x19 && cpu_model < 0x10) {
 		/*
@@ -53,12 +55,19 @@ static void pr_ibs_fetch_ctl(union ibs_fetch_ctl reg)
 		ic_miss_str = ic_miss_strs[reg.ic_miss];
 	}
 
+	if (zen4_ibs_extensions) {
+		snprintf(l3_miss_str, sizeof(l3_miss_str),
+			 " L3MissOnly %d FetchOcMiss %d FetchL3Miss %d",
+			 reg.l3_miss_only, reg.fetch_oc_miss, reg.fetch_l3_miss);
+	}
+
 	printf("ibs_fetch_ctl:\t%016llx MaxCnt %7d Cnt %7d Lat %5d En %d Val %d Comp %d%s "
-	       "PhyAddrValid %d%s L1TlbMiss %d L2TlbMiss %d RandEn %d%s\n",
+		"PhyAddrValid %d%s L1TlbMiss %d L2TlbMiss %d RandEn %d%s%s\n",
 		reg.val, reg.fetch_maxcnt << 4, reg.fetch_cnt << 4, reg.fetch_lat,
 		reg.fetch_en, reg.fetch_val, reg.fetch_comp, ic_miss_str ? : "",
 		reg.phy_addr_valid, l1tlb_pgsz_str ? : "", reg.l1tlb_miss, reg.l2tlb_miss,
-		reg.rand_en, reg.fetch_comp ? (reg.fetch_l2_miss ? " L2Miss 1" : " L2Miss 0") : "");
+		reg.rand_en, reg.fetch_comp ? (reg.fetch_l2_miss ? " L2Miss 1" : " L2Miss 0") : "",
+		l3_miss_str);
 }
 
 static void pr_ic_ibs_extd_ctl(union ic_ibs_extd_ctl reg)
@@ -68,9 +77,15 @@ static void pr_ic_ibs_extd_ctl(union ic_ibs_extd_ctl reg)
 
 static void pr_ibs_op_ctl(union ibs_op_ctl reg)
 {
-	printf("ibs_op_ctl:\t%016llx MaxCnt %9d En %d Val %d CntCtl %d=%s CurCnt %9d\n",
-	       reg.val, ((reg.opmaxcnt_ext << 16) | reg.opmaxcnt) << 4, reg.op_en, reg.op_val,
-	       reg.cnt_ctl, reg.cnt_ctl ? "uOps" : "cycles", reg.opcurcnt);
+	char l3_miss_only[sizeof(" L3MissOnly _")] = "";
+
+	if (zen4_ibs_extensions)
+		snprintf(l3_miss_only, sizeof(l3_miss_only), " L3MissOnly %d", reg.l3_miss_only);
+
+	printf("ibs_op_ctl:\t%016llx MaxCnt %9d%s En %d Val %d CntCtl %d=%s CurCnt %9d\n",
+		reg.val, ((reg.opmaxcnt_ext << 16) | reg.opmaxcnt) << 4, l3_miss_only,
+		reg.op_en, reg.op_val, reg.cnt_ctl,
+		reg.cnt_ctl ? "uOps" : "cycles", reg.opcurcnt);
 }
 
 static void pr_ibs_op_data(union ibs_op_data reg)
@@ -84,7 +99,34 @@ static void pr_ibs_op_data(union ibs_op_data reg)
 		reg.op_brn_ret, reg.op_rip_invalid, reg.op_brn_fuse, reg.op_microcode);
 }
 
-static void pr_ibs_op_data2(union ibs_op_data2 reg)
+static void pr_ibs_op_data2_extended(union ibs_op_data2 reg)
+{
+	static const char * const data_src_str[] = {
+		"",
+		" DataSrc 1=Local L3 or other L1/L2 in CCX",
+		" DataSrc 2=A peer cache in a near CCX",
+		" DataSrc 3=Data returned from DRAM",
+		" DataSrc 4=(reserved)",
+		" DataSrc 5=A peer cache in a far CCX",
+		" DataSrc 6=DRAM address map with \"long latency\" bit set",
+		" DataSrc 7=Data returned from MMIO/Config/PCI/APIC",
+		" DataSrc 8=Extension Memory (S-Link, GenZ, etc)",
+		" DataSrc 9=(reserved)",
+		" DataSrc 10=(reserved)",
+		" DataSrc 11=(reserved)",
+		" DataSrc 12=Peer Agent Memory",
+		/* 13 to 31 are reserved. Avoid printing them. */
+	};
+	int data_src = (reg.data_src_hi << 3) | reg.data_src_lo;
+
+	printf("ibs_op_data2:\t%016llx %sRmtNode %d%s\n", reg.val,
+		(data_src == 1 || data_src == 2 || data_src == 5) ?
+			(reg.cache_hit_st ? "CacheHitSt 1=O-State " : "CacheHitSt 0=M-state ") : "",
+		reg.rmt_node,
+		data_src < (int)ARRAY_SIZE(data_src_str) ? data_src_str[data_src] : "");
+}
+
+static void pr_ibs_op_data2_default(union ibs_op_data2 reg)
 {
 	static const char * const data_src_str[] = {
 		"",
@@ -103,6 +145,13 @@ static void pr_ibs_op_data2(union ibs_op_data2 reg)
 	       reg.rmt_node, data_src_str[reg.data_src_lo]);
 }
 
+static void pr_ibs_op_data2(union ibs_op_data2 reg)
+{
+	if (zen4_ibs_extensions)
+		return pr_ibs_op_data2_extended(reg);
+	pr_ibs_op_data2_default(reg);
+}
+
 static void pr_ibs_op_data3(union ibs_op_data3 reg)
 {
 	char l2_miss_str[sizeof(" L2Miss _")] = "";
@@ -279,6 +328,9 @@ bool evlist__has_amd_ibs(struct evlist *evlist)
 		pmu_mapping += strlen(pmu_mapping) + 1 /* '\0' */;
 	}
 
+	if (perf_env__find_pmu_cap(env, "ibs_op", "zen4_ibs_extensions"))
+		zen4_ibs_extensions = 1;
+
 	if (ibs_fetch_type || ibs_op_type) {
 		if (!cpu_family)
 			parse_cpuid(env);
Signature · 1 line
2.31.1
maintainer Acked-by on the v6 series
#10 lore raw
In reply to Ravi Bangoria

Hi Ravi,

Quoted 55 lines · Ravi Bangoria

Kernel side of changes have already been applied to linus/master (except amd-ibs.h header). This series contains perf tool changes.

v5: https://lore.kernel.org/lkml/[email protected]
v5->v6:
- Use macros instead of magic numbers for IBS l3missonly bits
- Use asprintf() instead of allocating memory and copying data manually
- Add Reviewed-by Kan Liang (patch 2-5).

Patches prepared on acme/perf/core (9dde6cadb92b5)

Original cover letter:

IBS support has been enhanced with two new features in upcoming uarch:
1. DataSrc extension and 2. L3 Miss Filtering capability. Both are indicated by CPUID_Fn8000001B_EAX bit 11.

DataSrc extension provides additional data source details for tagged load/store operations. Add support for these new bits in perf report/ script raw-dump.

IBS L3 miss filtering works by tagging an instruction on IBS counter overflow and generating an NMI if the tagged instruction causes an L3 miss. Samples without an L3 miss are discarded and counter is reset with random value (between 1-15 for fetch pmu and 1-127 for op pmu). This helps in reducing sampling overhead when user is interested only in such samples. One of the use case of such filtered samples is to feed data to page-migration daemon in tiered memory systems.

Add support for L3 miss filtering in IBS driver via new pmu attribute "l3missonly". Example usage:

  # perf record -a -e ibs_op/l3missonly=1/ --raw-samples sleep 5
  # perf report -D

Some important points to keep in mind while using L3 miss filtering:
1. Hw internally reset sampling period when tagged instruction does not cause L3 miss. But there is no way to reconstruct aggregated sampling period when this happens.
2. L3 miss is not the actual event being counted. Rather, IBS will count fetch, cycles or uOps depending on the configuration. Thus sampling period have no direct connection to L3 misses.

1st causes sampling period skew. Thus, I've added warning message at perf record:

  # perf record -c 10000 -C 0 -e ibs_op/l3missonly=1/
  WARNING: Hw internally resets sampling period when L3 Miss Filtering is enabled
  and tagged operation does not cause L3 Miss. This causes sampling period skew.

User can configure smaller sampling period to get more samples while using l3missonly.

Quoted 9 lines

Ravi Bangoria (8):
perf record ibs: Warn about sampling period skew perf tool: Parse pmu caps sysfs only once
perf headers: Pass "cpu" pmu name while printing caps perf headers: Store pmu caps in an array of strings perf headers: Record non-cpu pmu capabilities perf/x86/ibs: Add new IBS register bits into header perf tool ibs: Sync amd ibs header file
perf script ibs: Support new IBS bits in raw trace dump

Thanks,
Namhyung

Quoted 20 lines
 arch/x86/include/asm/amd-ibs.h                |  16 +-
 tools/arch/x86/include/asm/amd-ibs.h          |  16 +-
 .../Documentation/perf.data-file-format.txt   |  10 +-
 tools/perf/arch/x86/util/evsel.c              |  52 +++++
 tools/perf/builtin-inject.c                   |   2 +-
 tools/perf/util/amd-sample-raw.c              |  68 ++++++-
 tools/perf/util/env.c                         |  62 +++++-
 tools/perf/util/env.h                         |  14 +-
 tools/perf/util/evsel.c                       |   7 +
 tools/perf/util/evsel.h                       |   1 +
 tools/perf/util/header.c                      | 189 ++++++++++--------
 tools/perf/util/header.h                      |   2 +-
 tools/perf/util/pmu.c                         |  15 +-
 tools/perf/util/pmu.h                         |   2 +
 14 files changed, 329 insertions(+), 127 deletions(-)

--
2.31.1

maintainer Applied
#11 lore raw

Em Mon, Jun 06, 2022 at 04:46:57PM -0700, Namhyung Kim escreveu:

Quoted 11 lines · Ravi Bangoria
Quoted 9 lines

Ravi Bangoria (8):
perf record ibs: Warn about sampling period skew perf tool: Parse pmu caps sysfs only once
perf headers: Pass "cpu" pmu name while printing caps perf headers: Store pmu caps in an array of strings perf headers: Record non-cpu pmu capabilities perf/x86/ibs: Add new IBS register bits into header perf tool ibs: Sync amd ibs header file
perf script ibs: Support new IBS bits in raw trace dump

Quoted 1 line

Thanks, applied the series to perf/core, except for "[PATCH v6 6/8] perf/x86/ibs: Add new IBS register bits into header", that is outside tools/.

- Arnaldo

#12 lore raw
Quoted 12 lines · Arnaldo Carvalho de Melo

Em Mon, Jun 06, 2022 at 04:46:57PM -0700, Namhyung Kim escreveu:

Quoted 10 lines · Ravi Bangoria
Quoted 9 lines

Ravi Bangoria (8):
perf record ibs: Warn about sampling period skew perf tool: Parse pmu caps sysfs only once
perf headers: Pass "cpu" pmu name while printing caps perf headers: Store pmu caps in an array of strings perf headers: Record non-cpu pmu capabilities perf/x86/ibs: Add new IBS register bits into header perf tool ibs: Sync amd ibs header file
perf script ibs: Support new IBS bits in raw trace dump

Quoted 5 lines
Quoted 1 line

Thanks, applied the series to perf/core, except for "[PATCH v6 6/8] perf/x86/ibs: Add new IBS register bits into header", that is outside tools/.

Thanks Arnaldo.

Peter, Can you please pick up patch #6.

Thanks,
Ravi