Skip to contentSkip to conversation

perf mem: Add new output fields for data source (v1)

PATCH In review Namhyung Kim 3 participants 23 messages Apr 30, 2025 – May 12, 2025 perf events

lore
Share & tools
lore

Namhyung Kim

author maintainer
#0 lore raw

cover perf mem: Add new output fields for data source (v1)

Hello,

The perf mem uses PERF_SAMPLE_DATA_SRC which has a lot of information for memory access. It has various sort keys to group related samples together but it's still cumbersome to see the result. While perf c2c command provides a way to investigate the data in a specific way, I'd like to add more generic ways using new output fields.

For example, the following is the 'cache' output field which breaks down the sample weights into different level of caches.

  $ perf mem record -a sleep 1
  $ perf mem report -F cache,dso,sym --stdio
  ...
  #
  # -------------- Cache --------------
  #      L1     L2     L3 L1-buf  Other  Shared Object                                  Symbol
  # ...................................  .....................................  .........................................
  #
       0.0%   0.0%   0.0%   0.0% 100.0%  [kernel.kallsyms]                      [k] ioread8
     100.0%   0.0%   0.0%   0.0%   0.0%  [kernel.kallsyms]                      [k] _raw_spin_lock_irq
       0.0%   0.0%   0.0%   0.0% 100.0%  [xhci_hcd]                             [k] xhci_update_erst_dequeue
       0.0%   0.0%   0.0%  95.8%   4.2%  [kernel.kallsyms]                      [k] smaps_account
       0.6%   1.8%  22.7%  45.5%  29.5%  [kernel.kallsyms]                      [k] sched_balance_update_blocked_averages
      29.4%   0.0%   1.6%  58.8%  10.2%  [kernel.kallsyms]                      [k] __update_load_avg_cfs_rq
       0.0%   8.5%   4.3%   0.0%  87.2%  [kernel.kallsyms]                      [k] copy_mc_enhanced_fast_string
      63.9%   0.0%   8.0%  23.8%   4.3%  [kernel.kallsyms]                      [k] psi_group_change
       3.9%   0.0%   9.3%  35.7%  51.1%  [kernel.kallsyms]                      [k] timerqueue_add
      35.9%  10.9%   0.0%  39.0%  14.2%  [kernel.kallsyms]                      [k] memcpy
      94.1%   0.0%   0.0%   5.9%   0.0%  [kernel.kallsyms]                      [k] unmap_page_range
      25.7%   0.0%   4.9%  51.0%  18.4%  [kernel.kallsyms]                      [k] __update_load_avg_se
       0.0%  24.9%  19.4%   9.6%  46.1%  [kernel.kallsyms]                      [k] _copy_to_iter
      12.9%   0.0%   0.0%  87.1%   0.0%  [kernel.kallsyms]                      [k] next_uptodate_folio
      36.8%   0.0%   9.5%  16.6%  37.1%  [kernel.kallsyms]                      [k] update_curr
     100.0%   0.0%   0.0%   0.0%   0.0%  bpf_prog_b9611ccbbb3d1833_dfs_iter     [k] bpf_prog_b9611ccbbb3d1833_dfs_iter
      45.4%   1.8%  20.4%  23.6%   8.8%  [kernel.kallsyms]                      [k] audit_filter_rules.isra.0
      92.8%   0.0%   0.0%   7.2%   0.0%  [kernel.kallsyms]                      [k] filemap_map_pages
      10.6%   0.0%   0.0%  89.4%   0.0%  [kernel.kallsyms]                      [k] smaps_page_accumulate
      38.3%   0.0%  29.6%  27.1%   5.0%  [kernel.kallsyms]                      [k] __schedule

Please see the description of each commit for other fields.

New mem_stat field was added to the hist_entry to save this information. It's a generic data structure (array) to handle different type of information like cache-level, memory location, snoop-result, etc.

The first patch is a fix for the hierarchy mode and it was sent separately. I just add it here not to break the hierarchy mode. The second patch is to enable SAMPLE_DATA_SRC without SAMPLE_ADDR and perf_event_attr.mmap_data which generate a lot more data.

The name of some new fields are the same as the corresponding sort keys (mem, op, snoop) so I had to change the order whether it's applied as an output field or a sort key. Maybe it's better to name them differently but I couldn't come up with better ideas.

That means, you need to use -F/--fields option to specify those fields and the sort keys you want. Maybe we can change the default output and sort keys for perf mem report with this.

The code is available at 'perf/mem-field-v1' branch in

git://git.kernel.org/pub/scm/linux/kernel/git/namhyung/linux-perf.git

Thanks,
Namhyung

Namhyung Kim (11):
perf hist: Remove output field from sort-list properly perf record: Add --sample-mem-info option perf hist: Support multi-line header
perf hist: Add struct he_mem_stat
perf hist: Basic support for mem_stat accounting perf hist: Implement output fields for mem stats perf mem: Add 'op' output field
perf hist: Hide unused mem stat columns
perf mem: Add 'cache' and 'memory' output fields perf mem: Add 'snoop' output field
perf mem: Add 'dtlb' output field

 tools/perf/Documentation/perf-record.txt |   7 +-
 tools/perf/builtin-record.c              |   6 +
 tools/perf/ui/browsers/hists.c           |  50 ++++-
 tools/perf/ui/hist.c                     | 272 ++++++++++++++++++++++-
 tools/perf/ui/stdio/hist.c               |  57 +++--
 tools/perf/util/evsel.c                  |   2 +-
 tools/perf/util/hist.c                   |  78 +++++++
 tools/perf/util/hist.h                   |  22 ++
 tools/perf/util/mem-events.c             | 183 ++++++++++++++-
 tools/perf/util/mem-events.h             |  57 +++++
 tools/perf/util/record.h                 |   1 +
 tools/perf/util/sort.c                   |  42 +++-
 12 files changed, 718 insertions(+), 59 deletions(-)
Signature · 1 line
2.49.0.906.g1f30a19c02-goog
author maintainer
#1 lore raw

PATCH 1/11 perf hist: Remove output field from sort-list properly

When it removes an output format for cancelled children or latency, it should delete itself from the sort list as well. Otherwise assertion in fmt_free() will fire.

  $ perf report -H --stdio
  perf: ui/hist.c:603: fmt_free: Assertion `!(!list_empty(&fmt->sort_list))' failed.
  Aborted (core dumped)

Also convert to perf_hpp__column_unregister() for the same open codes.

Fixes dbd11b6bdab12f60 ("perf hist: Remove formats in hierarchy when cancel children")
---
 tools/perf/ui/hist.c | 15 +++++----------
 1 file changed, 5 insertions(+), 10 deletions(-)
tools/perf/ui/hist.c +5 -10
index 3ffce69fc823e0bf..bc0689fceeb18bde 100644
@@ -696,6 +696,7 @@ void perf_hpp_list__prepend_sort_field(struct perf_hpp_list *list,
 static void perf_hpp__column_unregister(struct perf_hpp_fmt *format)
 {
 	list_del_init(&format->list);
+	list_del_init(&format->sort_list);
 	fmt_free(format);
 }
 
@@ -818,18 +819,12 @@ void perf_hpp__reset_output_field(struct perf_hpp_list *list)
 	struct perf_hpp_fmt *fmt, *tmp;
 
 	/* reset output fields */
-	perf_hpp_list__for_each_format_safe(list, fmt, tmp) {
-		list_del_init(&fmt->list);
-		list_del_init(&fmt->sort_list);
-		fmt_free(fmt);
-	}
+	perf_hpp_list__for_each_format_safe(list, fmt, tmp)
+		perf_hpp__column_unregister(fmt);
 
 	/* reset sort keys */
-	perf_hpp_list__for_each_sort_list_safe(list, fmt, tmp) {
-		list_del_init(&fmt->list);
-		list_del_init(&fmt->sort_list);
-		fmt_free(fmt);
-	}
+	perf_hpp_list__for_each_sort_list_safe(list, fmt, tmp)
+		perf_hpp__column_unregister(fmt);
 }
 
 /*
Signature · 1 line
2.49.0.906.g1f30a19c02-goog
author maintainer
#2 lore raw

PATCH 2/11 perf record: Add --sample-mem-info option

In reply to Namhyung Kim

There's no way to enable PERF_SAMPLE_DATA_SRC without PERF_SAMPLE_ADDR which brings a lot of overhead due to the number of MMAP[2] records.

Let's add a new option to enable this information separately.

---
 tools/perf/Documentation/perf-record.txt | 7 ++++++-
 tools/perf/builtin-record.c              | 6 ++++++
 tools/perf/util/evsel.c                  | 2 +-
 tools/perf/util/record.h                 | 1 +
 4 files changed, 14 insertions(+), 2 deletions(-)
tools/perf/Documentation/perf-record.txt +6 -1
index c7fc1ba265e2755d..c59f1e79f2b4a6f8 100644
@@ -340,7 +340,7 @@ OPTIONS
 
 -d::
 --data::
-	Record the sample virtual addresses.
+	Record the sample virtual addresses.  Implies --sample-mem-info.
 
 --phys-data::
 	Record the sample physical addresses.
@@ -368,6 +368,11 @@ OPTIONS
 	the sample_type member of the struct perf_event_attr argument to the
 	perf_event_open system call.
 
+--sample-mem-info::
+	Record the sample data source information for memory operations.
+	It requires hardware supports and may work on specific events only.
+	Please consider using 'perf mem record' instead if you're not sure.
+
 -n::
 --no-samples::
 	Don't sample.
tools/perf/builtin-record.c +6 -0
index ba20bf7c011d7765..6637a3acb1f1295f 100644
@@ -3436,6 +3436,8 @@ static struct option __record_options[] = {
 		    "Record the sampled data address data page size"),
 	OPT_BOOLEAN(0, "code-page-size", &record.opts.sample_code_page_size,
 		    "Record the sampled code address (ip) page size"),
+	OPT_BOOLEAN(0, "sample-mem-info", &record.opts.sample_data_src,
+		    "Record the data source for memory operations"),
 	OPT_BOOLEAN(0, "sample-cpu", &record.opts.sample_cpu, "Record the sample cpu"),
 	OPT_BOOLEAN(0, "sample-identifier", &record.opts.sample_identifier,
 		    "Record the sample identifier"),
@@ -4130,6 +4132,10 @@ int cmd_record(int argc, const char **argv)
 		goto out_opts;
 	}
 
+	/* For backward compatibility, -d implies --mem-info */
+	if (rec->opts.sample_address)
+		rec->opts.sample_data_src = true;
+
 	/*
 	 * Allow aliases to facilitate the lookup of symbols for address
 	 * filters. Refer to auxtrace_parse_filters().
tools/perf/util/evsel.c +1 -1
index 1d79ffecd41f10ec..0f86df259c822799 100644
@@ -1425,7 +1425,7 @@ void evsel__config(struct evsel *evsel, struct record_opts *opts,
 		evsel__set_sample_bit(evsel, CPU);
 	}
 
-	if (opts->sample_address)
+	if (opts->sample_data_src)
 		evsel__set_sample_bit(evsel, DATA_SRC);
 
 	if (opts->sample_phys_addr)
tools/perf/util/record.h +1 -0
index a6566134e09e5b19..f1956c4db3195070 100644
@@ -28,6 +28,7 @@ struct record_opts {
 	bool	      sample_time_set;
 	bool	      sample_cpu;
 	bool	      sample_identifier;
+	bool	      sample_data_src;
 	bool	      period;
 	bool	      period_set;
 	bool	      running_time;
Signature · 1 line
2.49.0.906.g1f30a19c02-goog
author maintainer
#3 lore raw

PATCH 3/11 perf hist: Support multi-line header

In reply to Namhyung Kim

This is a preparation to support multi-line headers in perf mem report. Normal sort keys and output fields that don't have contents for multi- line will print the header string at the last line only.

As we don't use multi-line headers normally, it should not have any changes in the output.

---
 tools/perf/ui/browsers/hists.c | 24 +++++++++-----
 tools/perf/ui/hist.c           |  9 ++++--
 tools/perf/ui/stdio/hist.c     | 57 +++++++++++++++++++++-------------
 tools/perf/util/sort.c         |  8 +++--
 4 files changed, 64 insertions(+), 34 deletions(-)
tools/perf/ui/browsers/hists.c +16 -8
index cf022e92d06b9b28..67cbdec90d0bf0ea 100644
@@ -1686,7 +1686,8 @@ hists_browser__scnprintf_headers(struct hist_browser *browser, char *buf,
 	return ret;
 }
 
-static int hists_browser__scnprintf_hierarchy_headers(struct hist_browser *browser, char *buf, size_t size)
+static int hists_browser__scnprintf_hierarchy_headers(struct hist_browser *browser,
+						      char *buf, size_t size, int line)
 {
 	struct hists *hists = browser->hists;
 	struct perf_hpp dummy_hpp = {
@@ -1712,7 +1713,7 @@ static int hists_browser__scnprintf_hierarchy_headers(struct hist_browser *brows
 		if (column++ < browser->b.horiz_scroll)
 			continue;
 
-		ret = fmt->header(fmt, &dummy_hpp, hists, 0, NULL);
+		ret = fmt->header(fmt, &dummy_hpp, hists, line, NULL);
 		if (advance_hpp_check(&dummy_hpp, ret))
 			break;
 
@@ -1723,6 +1724,9 @@ static int hists_browser__scnprintf_hierarchy_headers(struct hist_browser *brows
 		first_node = false;
 	}
 
+	if (line < hists->hpp_list->nr_header_lines - 1)
+		return ret;
+
 	if (!first_node) {
 		ret = scnprintf(dummy_hpp.buf, dummy_hpp.size, "%*s",
 				indent * HIERARCHY_INDENT, "");
@@ -1753,7 +1757,7 @@ static int hists_browser__scnprintf_hierarchy_headers(struct hist_browser *brows
 			}
 			first_col = false;
 
-			ret = fmt->header(fmt, &dummy_hpp, hists, 0, NULL);
+			ret = fmt->header(fmt, &dummy_hpp, hists, line, NULL);
 			dummy_hpp.buf[ret] = '\0';
 
 			start = strim(dummy_hpp.buf);
@@ -1772,14 +1776,18 @@ static int hists_browser__scnprintf_hierarchy_headers(struct hist_browser *brows
 
 static void hists_browser__hierarchy_headers(struct hist_browser *browser)
 {
+	struct perf_hpp_list *hpp_list = browser->hists->hpp_list;
 	char headers[1024];
+	int line;
 
-	hists_browser__scnprintf_hierarchy_headers(browser, headers,
-						   sizeof(headers));
+	for (line = 0; line < hpp_list->nr_header_lines; line++) {
+		hists_browser__scnprintf_hierarchy_headers(browser, headers,
+							   sizeof(headers), line);
 
-	ui_browser__gotorc_title(&browser->b, 0, 0);
-	ui_browser__set_color(&browser->b, HE_COLORSET_ROOT);
-	ui_browser__write_nstring(&browser->b, headers, browser->b.width + 1);
+		ui_browser__gotorc_title(&browser->b, line, 0);
+		ui_browser__set_color(&browser->b, HE_COLORSET_ROOT);
+		ui_browser__write_nstring(&browser->b, headers, browser->b.width + 1);
+	}
 }
 
 static void hists_browser__headers(struct hist_browser *browser)
tools/perf/ui/hist.c +7 -2
index bc0689fceeb18bde..ec44633207aa3aba 100644
@@ -321,11 +321,16 @@ static int hpp__width_fn(struct perf_hpp_fmt *fmt,
 }
 
 static int hpp__header_fn(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
-			  struct hists *hists, int line __maybe_unused,
+			  struct hists *hists, int line,
 			  int *span __maybe_unused)
 {
 	int len = hpp__width_fn(fmt, hpp, hists);
-	return scnprintf(hpp->buf, hpp->size, "%*s", len, fmt->name);
+	const char *hdr = "";
+
+	if (line == hists->hpp_list->nr_header_lines - 1)
+		hdr = fmt->name;
+
+	return scnprintf(hpp->buf, hpp->size, "%*s", len, hdr);
 }
 
 int hpp_color_scnprintf(struct perf_hpp *hpp, const char *fmt, ...)
tools/perf/ui/stdio/hist.c +35 -22

80 lines not on this page · raw message

tools/perf/util/sort.c +6 -2
index 594b75ca95bf72b2..ae8b8ceb82f3d00b 100644
@@ -2641,18 +2641,22 @@ void perf_hpp__reset_sort_width(struct perf_hpp_fmt *fmt, struct hists *hists)
 }
 
 static int __sort__hpp_header(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
-			      struct hists *hists, int line __maybe_unused,
+			      struct hists *hists, int line,
 			      int *span __maybe_unused)
 {
 	struct hpp_sort_entry *hse;
 	size_t len = fmt->user_len;
+	const char *hdr = "";
+
+	if (line == hists->hpp_list->nr_header_lines - 1)
+		hdr = fmt->name;
 
 	hse = container_of(fmt, struct hpp_sort_entry, hpp);
 
 	if (!len)
 		len = hists__col_len(hists, hse->se->se_width_idx);
 
-	return scnprintf(hpp->buf, hpp->size, "%-*.*s", len, len, fmt->name);
+	return scnprintf(hpp->buf, hpp->size, "%-*.*s", len, len, hdr);
 }
 
 static int __sort__hpp_width(struct perf_hpp_fmt *fmt,
Signature · 1 line
2.49.0.906.g1f30a19c02-goog
author maintainer
#4 lore raw

PATCH 4/11 perf hist: Add struct he_mem_stat

In reply to Namhyung Kim

The struct he_mem_stat is to save detailed information about memory instruction. It'll be used to show breakdown of various data from PERF_SAMPLE_DATA_SRC. Note that this structure is generic and the contents will be different depending on actual data it'll use later.

The information about the actual data will be saved in struct hists and its length is in nr_mem_stats. This commit just adds ground works and does nothing since hists->nr_mem_stats is 0 for now.

---
 tools/perf/util/hist.c | 74 ++++++++++++++++++++++++++++++++++++++++++
 tools/perf/util/hist.h |  9 +++++
 2 files changed, 83 insertions(+)
tools/perf/util/hist.c +74 -0
index d65228c1141251fb..fcb9f0db0c92a229 100644
@@ -336,6 +336,67 @@ static void he_stat__decay(struct he_stat *he_stat)
 	he_stat->latency = (he_stat->latency * 7) / 8;
 }
 
+static int hists__update_mem_stat(struct hists *hists, struct hist_entry *he,
+				  struct mem_info *mi, u64 period)
+{
+	if (hists->nr_mem_stats == 0)
+		return 0;
+
+	if (he->mem_stat == NULL) {
+		he->mem_stat = calloc(hists->nr_mem_stats, sizeof(*he->mem_stat));
+		if (he->mem_stat == NULL)
+			return -1;
+	}
+
+	for (int i = 0; i < hists->nr_mem_stats; i++) {
+		int idx = 0; /* TODO: get correct index from mem info */
+
+		(void)mi;
+		he->mem_stat[i].entries[idx] += period;
+	}
+	return 0;
+}
+
+static void hists__add_mem_stat(struct hists *hists, struct hist_entry *dst,
+				struct hist_entry *src)
+{
+	if (hists->nr_mem_stats == 0)
+		return;
+
+	for (int i = 0; i < hists->nr_mem_stats; i++) {
+		for (int k = 0; k < MEM_STAT_LEN; k++)
+			dst->mem_stat[i].entries[k] += src->mem_stat[i].entries[k];
+	}
+}
+
+static int hists__clone_mem_stat(struct hists *hists, struct hist_entry *dst,
+				  struct hist_entry *src)
+{
+	if (hists->nr_mem_stats == 0)
+		return 0;
+
+	dst->mem_stat = calloc(hists->nr_mem_stats, sizeof(*dst->mem_stat));
+	if (dst->mem_stat == NULL)
+		return -1;
+
+	for (int i = 0; i < hists->nr_mem_stats; i++) {
+		for (int k = 0; k < MEM_STAT_LEN; k++)
+			dst->mem_stat[i].entries[k] = src->mem_stat[i].entries[k];
+	}
+	return 0;
+}
+
+static void hists__decay_mem_stat(struct hists *hists, struct hist_entry *he)
+{
+	if (hists->nr_mem_stats == 0)
+		return;
+
+	for (int i = 0; i < hists->nr_mem_stats; i++) {
+		for (int k = 0; k < MEM_STAT_LEN; k++)
+			he->mem_stat[i].entries[k] = (he->mem_stat[i].entries[k] * 7) / 8;
+	}
+}
+
 static void hists__delete_entry(struct hists *hists, struct hist_entry *he);
 
 static bool hists__decay_entry(struct hists *hists, struct hist_entry *he)
@@ -350,6 +411,7 @@ static bool hists__decay_entry(struct hists *hists, struct hist_entry *he)
 	if (symbol_conf.cumulate_callchain)
 		he_stat__decay(he->stat_acc);
 	decay_callchain(he->callchain);
+	hists__decay_mem_stat(hists, he);
 
 	if (!he->depth) {
 		u64 period_diff = prev_period - he->stat.period;
@@ -693,6 +755,10 @@ static struct hist_entry *hists__findnew_entry(struct hists *hists,
 		he_stat__add_cpumode_period(&he->stat, al->cpumode, period);
 	if (symbol_conf.cumulate_callchain)
 		he_stat__add_cpumode_period(he->stat_acc, al->cpumode, period);
+	if (hists__update_mem_stat(hists, he, entry->mem_info, period) < 0) {
+		hist_entry__delete(he);
+		return NULL;
+	}
 	return he;
 }
 
@@ -1423,6 +1489,7 @@ void hist_entry__delete(struct hist_entry *he)
 	free_callchain(he->callchain);
 	zfree(&he->trace_output);
 	zfree(&he->raw_data);
+	zfree(&he->mem_stat);
 	ops->free(he);
 }
 
@@ -1572,6 +1639,7 @@ static struct hist_entry *hierarchy_insert_entry(struct hists *hists,
 		cmp = hist_entry__collapse_hierarchy(hpp_list, iter, he);
 		if (!cmp) {
 			he_stat__add_stat(&iter->stat, &he->stat);
+			hists__add_mem_stat(hists, iter, he);
 			return iter;
 		}
 
@@ -1613,6 +1681,11 @@ static struct hist_entry *hierarchy_insert_entry(struct hists *hists,
 			new->srcfile = NULL;
 	}
 
+	if (hists__clone_mem_stat(hists, new, he) < 0) {
+		hist_entry__delete(new);
+		return NULL;
+	}
+
 	rb_link_node(&new->rb_node_in, parent, p);
 	rb_insert_color_cached(&new->rb_node_in, root, leftmost);
 	return new;
@@ -1695,6 +1768,7 @@ static int hists__collapse_insert_entry(struct hists *hists,
 			he_stat__add_stat(&iter->stat, &he->stat);
 			if (symbol_conf.cumulate_callchain)
 				he_stat__add_stat(iter->stat_acc, he->stat_acc);
+			hists__add_mem_stat(hists, iter, he);
 
 			if (hist_entry__has_callchains(he) && symbol_conf.use_callchain) {
 				struct callchain_cursor *cursor = get_tls_callchain_cursor();
tools/perf/util/hist.h +9 -0
index 76efd8952507a561..aba1d84ca074f27b 100644
@@ -100,6 +100,13 @@ enum hist_column {
 struct thread;
 struct dso;
 
+#define MEM_STAT_LEN  8
+
+struct he_mem_stat {
+	/* meaning of entries depends on enum mem_stat_type */
+	u64			entries[MEM_STAT_LEN];
+};
+
 struct hists {
 	struct rb_root_cached	entries_in_array[2];
 	struct rb_root_cached	*entries_in;
@@ -125,6 +132,7 @@ struct hists {
 	struct perf_hpp_list	*hpp_list;
 	struct list_head	hpp_formats;
 	int			nr_hpp_node;
+	int			nr_mem_stats;
 };
 
 #define hists__has(__h, __f) (__h)->hpp_list->__f
@@ -232,6 +240,7 @@ struct hist_entry {
 	} pairs;
 	struct he_stat		stat;
 	struct he_stat		*stat_acc;
+	struct he_mem_stat	*mem_stat;
 	struct map_symbol	ms;
 	struct thread		*thread;
 	struct comm		*comm;
Signature · 1 line
2.49.0.906.g1f30a19c02-goog
author maintainer
#5 lore raw

PATCH 5/11 perf hist: Basic support for mem_stat accounting

In reply to Namhyung Kim

Add a logic to account he->mem_stat based on mem_stat_type in hists. Each mem_stat entry will have different meaning based on the type so the index in the array is calculated at runtime using the corresponding value in the sample.data_src.

Still hists has no mem_stat_types yet so this code won't work for now. Later hists->mem_stat_types will be allocated based on what users want in the output actually.

---
 tools/perf/ui/hist.c         | 39 ++++++++++++++++++++++++++++++++++++
 tools/perf/util/hist.c       |  6 ++++--
 tools/perf/util/hist.h       |  4 ++++
 tools/perf/util/mem-events.c | 18 +++++++++++++++++
 tools/perf/util/mem-events.h |  6 ++++++
 tools/perf/util/sort.c       |  4 ++++
 6 files changed, 75 insertions(+), 2 deletions(-)
tools/perf/ui/hist.c +39 -0
index ec44633207aa3aba..2aad46bbd2ed4d93 100644
@@ -11,6 +11,7 @@
 #include "../util/sort.h"
 #include "../util/evsel.h"
 #include "../util/evlist.h"
+#include "../util/mem-events.h"
 #include "../util/thread.h"
 #include "../util/util.h"
 
@@ -500,6 +501,12 @@ static int64_t hpp__nop_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 	return 0;
 }
 
+static bool perf_hpp__is_mem_stat_entry(struct perf_hpp_fmt *fmt)
+{
+	(void)fmt;
+	return false;
+}
+
 static bool perf_hpp__is_hpp_entry(struct perf_hpp_fmt *a)
 {
 	return a->header == hpp__header_fn;
@@ -1022,3 +1029,35 @@ int perf_hpp__setup_hists_formats(struct perf_hpp_list *list,
 
 	return 0;
 }
+
+int perf_hpp__alloc_mem_stats(struct perf_hpp_list *list, struct evlist *evlist)
+{
+	struct perf_hpp_fmt *fmt;
+	struct evsel *evsel;
+	enum mem_stat_type mst[16];
+	unsigned nr_mem_stats = 0;
+
+	perf_hpp_list__for_each_format(list, fmt) {
+		if (!perf_hpp__is_mem_stat_entry(fmt))
+			continue;
+
+		assert(nr_mem_stats < ARRAY_SIZE(mst));
+		mst[nr_mem_stats++] = PERF_MEM_STAT_UNKNOWN;
+	}
+
+	if (nr_mem_stats == 0)
+		return 0;
+
+	evlist__for_each_entry(evlist, evsel) {
+		struct hists *hists = evsel__hists(evsel);
+
+		hists->mem_stat_types = calloc(nr_mem_stats,
+					       sizeof(*hists->mem_stat_types));
+		if (hists->mem_stat_types == NULL)
+			return -ENOMEM;
+
+		memcpy(hists->mem_stat_types, mst, nr_mem_stats * sizeof(*mst));
+		hists->nr_mem_stats = nr_mem_stats;
+	}
+	return 0;
+}
tools/perf/util/hist.c +4 -2
index fcb9f0db0c92a229..7759c1818c1ad168 100644
@@ -349,9 +349,10 @@ static int hists__update_mem_stat(struct hists *hists, struct hist_entry *he,
 	}
 
 	for (int i = 0; i < hists->nr_mem_stats; i++) {
-		int idx = 0; /* TODO: get correct index from mem info */
+		int idx = mem_stat_index(hists->mem_stat_types[i],
+					 mem_info__const_data_src(mi)->val);
 
-		(void)mi;
+		assert(0 <= idx && idx < MEM_STAT_LEN);
 		he->mem_stat[i].entries[idx] += period;
 	}
 	return 0;
@@ -3052,6 +3053,7 @@ static void hists_evsel__exit(struct evsel *evsel)
 	struct perf_hpp_list_node *node, *tmp;
 
 	hists__delete_all_entries(hists);
+	zfree(&hists->mem_stat_types);
 
 	list_for_each_entry_safe(node, tmp, &hists->hpp_formats, list) {
 		perf_hpp_list__for_each_format_safe(&node->hpp, fmt, pos) {
tools/perf/util/hist.h +4 -0
index aba1d84ca074f27b..509af09691b84e10 100644
@@ -9,6 +9,7 @@
 #include "events_stats.h"
 #include "evsel.h"
 #include "map_symbol.h"
+#include "mem-events.h"
 #include "mutex.h"
 #include "sample.h"
 #include "spark.h"
@@ -133,6 +134,7 @@ struct hists {
 	struct list_head	hpp_formats;
 	int			nr_hpp_node;
 	int			nr_mem_stats;
+	enum mem_stat_type	*mem_stat_types;
 };
 
 #define hists__has(__h, __f) (__h)->hpp_list->__f
@@ -597,6 +599,8 @@ void perf_hpp__reset_output_field(struct perf_hpp_list *list);
 void perf_hpp__append_sort_keys(struct perf_hpp_list *list);
 int perf_hpp__setup_hists_formats(struct perf_hpp_list *list,
 				  struct evlist *evlist);
+int perf_hpp__alloc_mem_stats(struct perf_hpp_list *list,
+			      struct evlist *evlist);
 
 
 bool perf_hpp__is_sort_entry(struct perf_hpp_fmt *format);
tools/perf/util/mem-events.c +18 -0
index 884d9aebce9199c0..1bc60ad3dc312542 100644
@@ -799,3 +799,21 @@ void c2c_add_stats(struct c2c_stats *stats, struct c2c_stats *add)
 	stats->nomap		+= add->nomap;
 	stats->noparse		+= add->noparse;
 }
+
+/*
+ * It returns an index in hist_entry->mem_stat array for the given val which
+ * represents a data-src based on the mem_stat_type.
+ *
+ * For example, when mst is about cache level, the index can be 1 for L1, 2 for
+ * L2 and so on.
+ */
+int mem_stat_index(const enum mem_stat_type mst, const u64 val)
+{
+	switch (mst) {
+	case PERF_MEM_STAT_UNKNOWN:  /* placeholder */
+	default:
+		break;
+	}
+	(void)val;
+	return -1;
+}
tools/perf/util/mem-events.h +6 -0
index a5c19d39ee37147b..2604464f985815f6 100644
@@ -89,4 +89,10 @@ struct hist_entry;
 int c2c_decode_stats(struct c2c_stats *stats, struct mem_info *mi);
 void c2c_add_stats(struct c2c_stats *stats, struct c2c_stats *add);
 
+enum mem_stat_type {
+	PERF_MEM_STAT_UNKNOWN,  /* placeholder */
+};
+
+int mem_stat_index(const enum mem_stat_type mst, const u64 data_src);
+
 #endif /* __PERF_MEM_EVENTS_H */
tools/perf/util/sort.c +4 -0
index ae8b8ceb82f3d00b..6024f588f66f3156 100644
@@ -4163,6 +4163,10 @@ int setup_sorting(struct evlist *evlist)
 	if (err < 0)
 		return err;
 
+	err = perf_hpp__alloc_mem_stats(&perf_hpp_list, evlist);
+	if (err < 0)
+		return err;
+
 	/* copy sort keys to output fields */
 	perf_hpp__setup_output_field(&perf_hpp_list);
 	/* and then copy output fields to sort keys */
Signature · 1 line
2.49.0.906.g1f30a19c02-goog
author maintainer
#6 lore raw

PATCH 6/11 perf hist: Implement output fields for mem stats

In reply to Namhyung Kim

This is a preparation for later changes to support mem_stat output. The new fields will need two lines for the header - the first line will show type of mem stat and the second line will show the name of each item which is returned by mem_stat_name().

Each element in the mem_stat array will be printed in percentage for the hist_entry and their sum would be 100%.

Add new output field dimension only for SORT_MODE__MEM using mem_stat. To handle possible name conflict with existing sort keys, move the order of checking output field dimensions after the sort dimensions when it looks for sort keys.

---
 tools/perf/ui/browsers/hists.c |  11 +++
 tools/perf/ui/hist.c           | 158 ++++++++++++++++++++++++++++++++-
 tools/perf/util/hist.h         |   4 +
 tools/perf/util/mem-events.c   |  12 +++
 tools/perf/util/mem-events.h   |   3 +
 tools/perf/util/sort.c         |  26 ++++--
 6 files changed, 202 insertions(+), 12 deletions(-)
tools/perf/ui/browsers/hists.c +11 -0
index 67cbdec90d0bf0ea..f6ab1310a0bdd6c4 100644
@@ -1266,6 +1266,16 @@ hist_browser__hpp_color_##_type(struct perf_hpp_fmt *fmt,		\
 			_fmttype);					\
 }
 
+#define __HPP_COLOR_MEM_STAT_FN(_name, _type)				\
+static int								\
+hist_browser__hpp_color_mem_stat_##_name(struct perf_hpp_fmt *fmt,	\
+					 struct perf_hpp *hpp,		\
+					 struct hist_entry *he)		\
+{									\
+	return hpp__fmt_mem_stat(fmt, hpp, he, PERF_MEM_STAT_##_type,	\
+				 " %5.1f%%", __hpp__slsmg_color_printf);\
+}
+
 __HPP_COLOR_PERCENT_FN(overhead, period, PERF_HPP_FMT_TYPE__PERCENT)
 __HPP_COLOR_PERCENT_FN(latency, latency, PERF_HPP_FMT_TYPE__LATENCY)
 __HPP_COLOR_PERCENT_FN(overhead_sys, period_sys, PERF_HPP_FMT_TYPE__PERCENT)
@@ -1277,6 +1287,7 @@ __HPP_COLOR_ACC_PERCENT_FN(latency_acc, latency, PERF_HPP_FMT_TYPE__LATENCY)
 
 #undef __HPP_COLOR_PERCENT_FN
 #undef __HPP_COLOR_ACC_PERCENT_FN
+#undef __HPP_COLOR_MEM_STAT_FN
 
 void hist_browser__init_hpp(void)
 {
tools/perf/ui/hist.c +155 -3

258 lines not on this page · raw message

tools/perf/util/hist.h +4 -0
index 509af09691b84e10..18c696d8d568a9fa 100644
@@ -587,6 +587,7 @@ enum {
 	PERF_HPP__WEIGHT1,
 	PERF_HPP__WEIGHT2,
 	PERF_HPP__WEIGHT3,
+	PERF_HPP__MEM_STAT_UNKNOWN,  /* placeholder */
 
 	PERF_HPP__MAX_INDEX
 };
@@ -656,6 +657,9 @@ int hpp__fmt_acc(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 		 struct hist_entry *he, hpp_field_fn get_field,
 		 const char *fmtstr, hpp_snprint_fn print_fn,
 		 enum perf_hpp_fmt_type fmtype);
+int hpp__fmt_mem_stat(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
+		      struct hist_entry *he, enum mem_stat_type mst,
+		      const char *fmtstr, hpp_snprint_fn print_fn);
 
 static inline void advance_hpp(struct perf_hpp *hpp, int inc)
 {
tools/perf/util/mem-events.c +12 -0
index 1bc60ad3dc312542..a4c1e42de30f8307 100644
@@ -817,3 +817,15 @@ int mem_stat_index(const enum mem_stat_type mst, const u64 val)
 	(void)val;
 	return -1;
 }
+
+/* To align output, returned string should be shorter than MEM_STAT_PRINT_LEN */
+const char *mem_stat_name(const enum mem_stat_type mst, const int idx)
+{
+	switch (mst) {
+	case PERF_MEM_STAT_UNKNOWN:
+	default:
+		break;
+	}
+	(void)idx;
+	return "N/A";
+}
tools/perf/util/mem-events.h +3 -0
index 2604464f985815f6..7aeb4c5fefc89698 100644
@@ -93,6 +93,9 @@ enum mem_stat_type {
 	PERF_MEM_STAT_UNKNOWN,  /* placeholder */
 };
 
+#define MEM_STAT_PRINT_LEN  7  /* 1 space + 5 digits + 1 percent sign */
+
 int mem_stat_index(const enum mem_stat_type mst, const u64 data_src);
+const char *mem_stat_name(const enum mem_stat_type mst, const int idx);
 
 #endif /* __PERF_MEM_EVENTS_H */
tools/perf/util/sort.c +17 -9
index 6024f588f66f3156..7c669ea27af247e5 100644
@@ -2598,9 +2598,11 @@ struct hpp_dimension {
 	struct perf_hpp_fmt	*fmt;
 	int			taken;
 	int			was_taken;
+	int			mem_mode;
 };
 
 #define DIM(d, n) { .name = n, .fmt = &perf_hpp__format[d], }
+#define DIM_MEM(d, n) { .name = n, .fmt = &perf_hpp__format[d], .mem_mode = 1, }
 
 static struct hpp_dimension hpp_sort_dimensions[] = {
 	DIM(PERF_HPP__OVERHEAD, "overhead"),
@@ -2620,8 +2622,11 @@ static struct hpp_dimension hpp_sort_dimensions[] = {
 	DIM(PERF_HPP__WEIGHT2, "ins_lat"),
 	DIM(PERF_HPP__WEIGHT3, "retire_lat"),
 	DIM(PERF_HPP__WEIGHT3, "p_stage_cyc"),
+	/* used for output only when SORT_MODE__MEM */
+	DIM_MEM(PERF_HPP__MEM_STAT_UNKNOWN, "unknown"),  /* placeholder */
 };
 
+#undef DIM_MEM
 #undef DIM
 
 struct hpp_sort_entry {
@@ -3608,15 +3613,6 @@ int sort_dimension__add(struct perf_hpp_list *list, const char *tok,
 		return __sort_dimension__add(sd, list, level);
 	}
 
-	for (i = 0; i < ARRAY_SIZE(hpp_sort_dimensions); i++) {
-		struct hpp_dimension *hd = &hpp_sort_dimensions[i];
-
-		if (strncasecmp(tok, hd->name, strlen(tok)))
-			continue;
-
-		return __hpp_dimension__add(hd, list, level);
-	}
-
 	for (i = 0; i < ARRAY_SIZE(bstack_sort_dimensions); i++) {
 		struct sort_dimension *sd = &bstack_sort_dimensions[i];
 
@@ -3658,6 +3654,15 @@ int sort_dimension__add(struct perf_hpp_list *list, const char *tok,
 		return 0;
 	}
 
+	for (i = 0; i < ARRAY_SIZE(hpp_sort_dimensions); i++) {
+		struct hpp_dimension *hd = &hpp_sort_dimensions[i];
+
+		if (strncasecmp(tok, hd->name, strlen(tok)))
+			continue;
+
+		return __hpp_dimension__add(hd, list, level);
+	}
+
 	if (!add_dynamic_entry(evlist, tok, level))
 		return 0;
 
@@ -4020,6 +4025,9 @@ int output_field_add(struct perf_hpp_list *list, const char *tok, int *level)
 		if (!strcasecmp(tok, "weight"))
 			ui__warning("--fields weight shows the average value unlike in the --sort key.\n");
 
+		if (hd->mem_mode && sort__mode != SORT_MODE__MEMORY)
+			continue;
+
 		return __hpp_dimension__add_output(list, hd, *level);
 	}
 
Signature · 1 line
2.49.0.906.g1f30a19c02-goog
author maintainer
#7 lore raw

PATCH 7/11 perf mem: Add 'op' output field

In reply to Namhyung Kim

This is an actual example of the he_mem_stat based sample breakdown. It uses 'mem_op' field of union perf_mem_data_src which means memory operations.

It'd have basically 'load' or 'store' which can be useful if PMU doesn't have separate events for them like IBS or SPE. In addition, there's an entry in case load and store happen at the same time. Also adds entries for prefetching and execution.

  $ perf mem report -F +op -s comm --stdio
  # To display the perf.data header info, please use --header/--header-only options.
  #
  #
  # Total Lost Samples: 0
  #
  # Samples: 4K of event 'ibs_op//'
  # Total weight : 9559
  # Sort order   : comm
  #
  #                         ------------------------ Mem Op ------------------------
  # Overhead       Samples     Load  Store  Ld+St Pfetch   Exec  Other    N/A    N/A  Command
  # ........  ............  ........................................................  ...............
  #
      44.85%          4077    21.1%  30.7%   0.0%   0.0%   0.0%  48.3%   0.0%   0.0%  swapper
      26.82%            45    98.8%   0.3%   0.0%   0.0%   0.0%   0.9%   0.0%   0.0%  netsli-prober
       7.19%           442    51.7%  13.7%   0.0%   0.0%   0.0%  34.6%   0.0%   0.0%  perf
       5.81%            75    89.7%   2.2%   0.0%   0.0%   0.0%   8.1%   0.0%   0.0%  qemu-system-ppc
       4.77%             1   100.0%   0.0%   0.0%   0.0%   0.0%   0.0%   0.0%   0.0%  notifications_c
       1.77%            10    95.9%   1.2%   0.0%   0.0%   0.0%   3.0%   0.0%   0.0%  MemoryReleaser
       0.77%            32    71.6%   4.1%   0.0%   0.0%   0.0%  24.3%   0.0%   0.0%  DefaultEventMan
       0.19%            10    66.7%  22.2%   0.0%   0.0%   0.0%  11.1%   0.0%   0.0%  gnome-shell
---
 tools/perf/ui/browsers/hists.c |  3 +++
 tools/perf/ui/hist.c           | 12 ++++++---
 tools/perf/util/hist.h         |  2 +-
 tools/perf/util/mem-events.c   | 48 ++++++++++++++++++++++++++--------
 tools/perf/util/mem-events.h   | 11 +++++++-
 tools/perf/util/sort.c         |  2 +-
 6 files changed, 61 insertions(+), 17 deletions(-)
tools/perf/ui/browsers/hists.c +3 -0
index f6ab1310a0bdd6c4..66a4c769b2d76436 100644
@@ -1284,6 +1284,7 @@ __HPP_COLOR_PERCENT_FN(overhead_guest_sys, period_guest_sys, PERF_HPP_FMT_TYPE__
 __HPP_COLOR_PERCENT_FN(overhead_guest_us, period_guest_us, PERF_HPP_FMT_TYPE__PERCENT)
 __HPP_COLOR_ACC_PERCENT_FN(overhead_acc, period, PERF_HPP_FMT_TYPE__PERCENT)
 __HPP_COLOR_ACC_PERCENT_FN(latency_acc, latency, PERF_HPP_FMT_TYPE__LATENCY)
+__HPP_COLOR_MEM_STAT_FN(op, OP)
 
 #undef __HPP_COLOR_PERCENT_FN
 #undef __HPP_COLOR_ACC_PERCENT_FN
@@ -1307,6 +1308,8 @@ void hist_browser__init_hpp(void)
 				hist_browser__hpp_color_overhead_acc;
 	perf_hpp__format[PERF_HPP__LATENCY_ACC].color =
 				hist_browser__hpp_color_latency_acc;
+	perf_hpp__format[PERF_HPP__MEM_STAT_OP].color =
+				hist_browser__hpp_color_mem_stat_op;
 
 	res_sample_init();
 }
tools/perf/ui/hist.c +9 -3
index 2a5c9f2b328b2c5c..427ce687ad815a62 100644
@@ -342,6 +342,12 @@ static enum mem_stat_type hpp__mem_stat_type(struct perf_hpp_fmt *fmt)
 	if (!perf_hpp__is_mem_stat_entry(fmt))
 		return -1;
 
+	switch (fmt->idx) {
+	case PERF_HPP__MEM_STAT_OP:
+		return PERF_MEM_STAT_OP;
+	default:
+		break;
+	}
 	pr_debug("Should not reach here\n");
 	return -1;
 }
@@ -611,7 +617,7 @@ HPP_AVERAGE_FNS(weight1, weight1)
 HPP_AVERAGE_FNS(weight2, weight2)
 HPP_AVERAGE_FNS(weight3, weight3)
 
-HPP_MEM_STAT_FNS(unknown, UNKNOWN)  /* placeholder */
+HPP_MEM_STAT_FNS(op, OP)
 
 static int64_t hpp__nop_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 			    struct hist_entry *a __maybe_unused,
@@ -715,7 +721,7 @@ struct perf_hpp_fmt perf_hpp__format[] = {
 	HPP__PRINT_FNS("Weight1", weight1, WEIGHT1),
 	HPP__PRINT_FNS("Weight2", weight2, WEIGHT2),
 	HPP__PRINT_FNS("Weight3", weight3, WEIGHT3),
-	HPP__MEM_STAT_PRINT_FNS("Unknown", unknown, UNKNOWN),  /* placeholder */
+	HPP__MEM_STAT_PRINT_FNS("Mem Op", op, OP),
 };
 
 struct perf_hpp_list perf_hpp_list = {
@@ -1070,7 +1076,7 @@ void perf_hpp__reset_width(struct perf_hpp_fmt *fmt, struct hists *hists)
 		fmt->len = 8;
 		break;
 
-	case PERF_HPP__MEM_STAT_UNKNOWN:  /* placeholder */
+	case PERF_HPP__MEM_STAT_OP:
 		fmt->len = MEM_STAT_LEN * MEM_STAT_PRINT_LEN;
 		break;
 
tools/perf/util/hist.h +1 -1
index 18c696d8d568a9fa..3990cfc21b1615ae 100644
@@ -587,7 +587,7 @@ enum {
 	PERF_HPP__WEIGHT1,
 	PERF_HPP__WEIGHT2,
 	PERF_HPP__WEIGHT3,
-	PERF_HPP__MEM_STAT_UNKNOWN,  /* placeholder */
+	PERF_HPP__MEM_STAT_OP,
 
 	PERF_HPP__MAX_INDEX
 };
tools/perf/util/mem-events.c +37 -11

80 lines not on this page · raw message

tools/perf/util/mem-events.h +10 -1
index 7aeb4c5fefc89698..55e5e2607fb732b4 100644
@@ -90,7 +90,16 @@ int c2c_decode_stats(struct c2c_stats *stats, struct mem_info *mi);
 void c2c_add_stats(struct c2c_stats *stats, struct c2c_stats *add);
 
 enum mem_stat_type {
-	PERF_MEM_STAT_UNKNOWN,  /* placeholder */
+	PERF_MEM_STAT_OP,
+};
+
+enum mem_stat_op {
+	MEM_STAT_OP_LOAD,
+	MEM_STAT_OP_STORE,
+	MEM_STAT_OP_LDST,
+	MEM_STAT_OP_PFETCH,
+	MEM_STAT_OP_EXEC,
+	MEM_STAT_OP_OTHER,
 };
 
 #define MEM_STAT_PRINT_LEN  7  /* 1 space + 5 digits + 1 percent sign */
tools/perf/util/sort.c +1 -1
index 7c669ea27af247e5..53fcb9191ea0cdc3 100644
@@ -2623,7 +2623,7 @@ static struct hpp_dimension hpp_sort_dimensions[] = {
 	DIM(PERF_HPP__WEIGHT3, "retire_lat"),
 	DIM(PERF_HPP__WEIGHT3, "p_stage_cyc"),
 	/* used for output only when SORT_MODE__MEM */
-	DIM_MEM(PERF_HPP__MEM_STAT_UNKNOWN, "unknown"),  /* placeholder */
+	DIM_MEM(PERF_HPP__MEM_STAT_OP, "op"),
 };
 
 #undef DIM_MEM
Signature · 1 line
2.49.0.906.g1f30a19c02-goog
author maintainer
#8 lore raw

PATCH 8/11 perf hist: Hide unused mem stat columns

In reply to Namhyung Kim

Some mem_stat types don't use all 8 columns. And there are cases only samples in certain kinds of mem_stat types are available only. For that case hide columns which has no samples.

The new output for the previous data would be:

  $ perf mem report -F overhead,op,comm --stdio
  ...
  #           ------ Mem Op -------
  # Overhead     Load  Store  Other  Command
  # ........  .....................  ...............
  #
      44.85%    21.1%  30.7%  48.3%  swapper
      26.82%    98.8%   0.3%   0.9%  netsli-prober
       7.19%    51.7%  13.7%  34.6%  perf
       5.81%    89.7%   2.2%   8.1%  qemu-system-ppc
       4.77%   100.0%   0.0%   0.0%  notifications_c
       1.77%    95.9%   1.2%   3.0%  MemoryReleaser
       0.77%    71.6%   4.1%  24.3%  DefaultEventMan
       0.19%    66.7%  22.2%  11.1%  gnome-shell
       ...

On Intel machines, the event is only for loads or stores so it'll have only one columns like below:

# Mem Op
# Overhead Load Command

  # ........  .......  ...............
  #
      20.55%   100.0%  swapper
      17.13%   100.0%  chrome
       9.02%   100.0%  data-loop.0
       6.26%   100.0%  pipewire-pulse
       5.63%   100.0%  threaded-ml
       5.47%   100.0%  GraphRunner
       5.37%   100.0%  AudioIP~allback
       5.30%   100.0%  Chrome_ChildIOT
       3.17%   100.0%  Isolated Web Co
       ...
---
 tools/perf/ui/hist.c   | 35 +++++++++++++++++++++++++++++++++--
 tools/perf/util/hist.c |  2 ++
 tools/perf/util/hist.h |  1 +
 3 files changed, 36 insertions(+), 2 deletions(-)
tools/perf/ui/hist.c +33 -2
index 427ce687ad815a62..661922c4d7863224 100644
@@ -178,6 +178,9 @@ int hpp__fmt_mem_stat(struct perf_hpp_fmt *fmt __maybe_unused, struct perf_hpp *
 	for (int i = 0; i < MEM_STAT_LEN; i++) {
 		u64 val = he->mem_stat[mem_stat_idx].entries[i];
 
+		if (hists->mem_stat_total[mem_stat_idx].entries[i] == 0)
+			continue;
+
 		ret += hpp__call_print_fn(hpp, print_fn, fmtstr, 100.0 * val / total);
 	}
 
@@ -405,12 +408,31 @@ static int hpp__header_mem_stat_fn(struct perf_hpp_fmt *fmt, struct perf_hpp *hp
 	int ret = 0;
 	int len;
 	enum mem_stat_type mst = hpp__mem_stat_type(fmt);
+	int mem_stat_idx = -1;
+
+	for (int i = 0; i < hists->nr_mem_stats; i++) {
+		if (hists->mem_stat_types[i] == mst) {
+			mem_stat_idx = i;
+			break;
+		}
+	}
+	assert(mem_stat_idx != -1);
 
-	(void)hists;
 	if (line == 0) {
 		int left, right;
 
-		len = fmt->len;
+		len = 0;
+		/* update fmt->len for acutally used columns only */
+		for (int i = 0; i < MEM_STAT_LEN; i++) {
+			if (hists->mem_stat_total[mem_stat_idx].entries[i])
+				len += MEM_STAT_PRINT_LEN;
+		}
+		fmt->len = len;
+
+		/* print header directly if single column only */
+		if (len == MEM_STAT_PRINT_LEN)
+			return scnprintf(hpp->buf, hpp->size, "%*s", len, fmt->name);
+
 		left = (len - strlen(fmt->name)) / 2 - 1;
 		right = len - left - strlen(fmt->name) - 2;
 
@@ -423,10 +445,14 @@ static int hpp__header_mem_stat_fn(struct perf_hpp_fmt *fmt, struct perf_hpp *hp
 				 left, graph_dotted_line, fmt->name, right, graph_dotted_line);
 	}
 
+
 	len = hpp->size;
 	for (int i = 0; i < MEM_STAT_LEN; i++) {
 		int printed;
 
+		if (hists->mem_stat_total[mem_stat_idx].entries[i] == 0)
+			continue;
+
 		printed = scnprintf(buf, len, "%*s", MEM_STAT_PRINT_LEN,
 				    mem_stat_name(mst, i));
 		ret += printed;
@@ -1214,6 +1240,11 @@ int perf_hpp__alloc_mem_stats(struct perf_hpp_list *list, struct evlist *evlist)
 		if (hists->mem_stat_types == NULL)
 			return -ENOMEM;
 
+		hists->mem_stat_total = calloc(nr_mem_stats,
+					       sizeof(*hists->mem_stat_total));
+		if (hists->mem_stat_total == NULL)
+			return -ENOMEM;
+
 		memcpy(hists->mem_stat_types, mst, nr_mem_stats * sizeof(*mst));
 		hists->nr_mem_stats = nr_mem_stats;
 	}
tools/perf/util/hist.c +2 -0
index 7759c1818c1ad168..afc6855327ab0de6 100644
@@ -354,6 +354,7 @@ static int hists__update_mem_stat(struct hists *hists, struct hist_entry *he,
 
 		assert(0 <= idx && idx < MEM_STAT_LEN);
 		he->mem_stat[i].entries[idx] += period;
+		hists->mem_stat_total[i].entries[idx] += period;
 	}
 	return 0;
 }
@@ -3054,6 +3055,7 @@ static void hists_evsel__exit(struct evsel *evsel)
 
 	hists__delete_all_entries(hists);
 	zfree(&hists->mem_stat_types);
+	zfree(&hists->mem_stat_total);
 
 	list_for_each_entry_safe(node, tmp, &hists->hpp_formats, list) {
 		perf_hpp_list__for_each_format_safe(&node->hpp, fmt, pos) {
tools/perf/util/hist.h +1 -0
index 3990cfc21b1615ae..fa5e886e5b04ec9b 100644
@@ -135,6 +135,7 @@ struct hists {
 	int			nr_hpp_node;
 	int			nr_mem_stats;
 	enum mem_stat_type	*mem_stat_types;
+	struct he_mem_stat	*mem_stat_total;
 };
 
 #define hists__has(__h, __f) (__h)->hpp_list->__f
Signature · 1 line
2.49.0.906.g1f30a19c02-goog
#9 lore raw

Re: [PATCH 08/11] perf hist: Hide unused mem stat columns

Quoted 27 lines · Namhyung Kim

Some mem_stat types don't use all 8 columns. And there are cases only samples in certain kinds of mem_stat types are available only. For that case hide columns which has no samples.

The new output for the previous data would be:

  $ perf mem report -F overhead,op,comm --stdio
  ...
  #           ------ Mem Op -------
  # Overhead     Load  Store  Other  Command
  # ........  .....................  ...............
  #
      44.85%    21.1%  30.7%  48.3%  swapper
      26.82%    98.8%   0.3%   0.9%  netsli-prober
       7.19%    51.7%  13.7%  34.6%  perf
       5.81%    89.7%   2.2%   8.1%  qemu-system-ppc
       4.77%   100.0%   0.0%   0.0%  notifications_c
       1.77%    95.9%   1.2%   3.0%  MemoryReleaser
       0.77%    71.6%   4.1%  24.3%  DefaultEventMan
       0.19%    66.7%  22.2%  11.1%  gnome-shell
       ...

On Intel machines, the event is only for loads or stores so it'll have only one columns like below:

# Mem Op
# Overhead Load Command

Quoted 12 lines
  # ........  .......  ...............
  #
      20.55%   100.0%  swapper
      17.13%   100.0%  chrome
       9.02%   100.0%  data-loop.0
       6.26%   100.0%  pipewire-pulse
       5.63%   100.0%  threaded-ml
       5.47%   100.0%  GraphRunner
       5.37%   100.0%  AudioIP~allback
       5.30%   100.0%  Chrome_ChildIOT
       3.17%   100.0%  Isolated Web Co
       ...
  # grep "model name" -m1 /proc/cpuinfo
  model name    : AMD Ryzen 9 9950X3D 16-Core Processo
  # perf mem report -F overhead,op,comm --stdio
  # Total Lost Samples: 0
  #
  # Samples: 2K of event 'cycles:P'
  # Total weight : 2637
  # Sort order   : local_weight,mem,sym,dso,symbol_daddr,dso_daddr,snoop,tlb,locked,blocked,local_ins_lat,local_p_stage_cyc
  #
  #           ------ Mem Op -------
  # Overhead     Load  Store  Other  Command
  # ........  .....................  ...............
  #
      61.02%    14.4%  25.5%  60.1%  swapper
       5.61%    26.4%  13.5%  60.1%  Isolated Web Co
       5.50%    21.4%  29.7%  49.0%  perf
       4.74%    27.2%  15.2%  57.6%  gnome-shell
       4.63%    33.6%  11.5%  54.9%  mdns_service
       4.29%    28.3%  12.4%  59.3%  ptyxis
       2.16%    24.6%  19.3%  56.1%  DOM Worker
       0.99%    23.1%  34.6%  42.3%  firefox
       0.72%    26.3%  15.8%  57.9%  IPC I/O Parent
       0.61%    12.5%  12.5%  75.0%  kworker/u130:20
       0.61%    37.5%  18.8%  43.8%  podman
       0.57%    33.3%   6.7%  60.0%  Timer
       0.53%    14.3%   7.1%  78.6%  KMS thread
       0.49%    30.8%   7.7%  61.5%  kworker/u130:3-
       0.46%    41.7%  33.3%  25.0%  IPDL Background
#10 lore raw

Re: [PATCH 08/11] perf hist: Hide unused mem stat columns

In reply to Namhyung Kim
Quoted 6 lines · Namhyung Kim

Some mem_stat types don't use all 8 columns. And there are cases only samples in certain kinds of mem_stat types are available only. For that case hide columns which has no samples.

The new output for the previous data would be:

Quoted 8 lines
  $ perf mem report -F overhead,op,comm --stdio
  ...
  #           ------ Mem Op -------
  # Overhead     Load  Store  Other  Command
  # ........  .....................  ...............
  #
      44.85%    21.1%  30.7%  48.3%  swapper
      26.82%    98.8%   0.3%   0.9%  netsli-prober

/me curious about this "Other" column.

Maps to MEM_STAT_OP_OTHER, that comes from mem_stat_index, that comes from:

int mem_stat_index(const enum mem_stat_type mst, const u64 val)

{
        union perf_mem_data_src src = {
                .val = val,
        };

                int idx = mem_stat_index(hists->mem_stat_types[i],
                                         mem_info__const_data_src(mi)->val);

struct mem_info *mi

union perf_mem_data_src {
__u64 val;
struct {
__u64 mem_op:5, /* type of opcode */
mem_lvl:14, /* memory hierarchy level */
mem_snoop:5, /* snoop mode */
mem_lock:2, /* lock instr */
mem_dtlb:7, /* tlb access */
mem_lvl_num:4, /* memory hierarchy level number */
mem_remote:1, /* remote */
mem_snoopx:2, /* snoop mode, ext */
mem_blk:3, /* access blocked */
mem_hops:3, /* hop level */
mem_rsvd:18;
};
};

As the percentage for "Other" is so high I think some other patch in this series will elucidate that :-)

Lemme continue testing...

- Arnaldo

Quoted 142 lines
       7.19%    51.7%  13.7%  34.6%  perf
       5.81%    89.7%   2.2%   8.1%  qemu-system-ppc
       4.77%   100.0%   0.0%   0.0%  notifications_c
       1.77%    95.9%   1.2%   3.0%  MemoryReleaser
       0.77%    71.6%   4.1%  24.3%  DefaultEventMan
       0.19%    66.7%  22.2%  11.1%  gnome-shell
       ...

On Intel machines, the event is only for loads or stores so it'll have only one columns like below:

# Mem Op
# Overhead Load Command

  # ........  .......  ...............
  #
      20.55%   100.0%  swapper
      17.13%   100.0%  chrome
       9.02%   100.0%  data-loop.0
       6.26%   100.0%  pipewire-pulse
       5.63%   100.0%  threaded-ml
       5.47%   100.0%  GraphRunner
       5.37%   100.0%  AudioIP~allback
       5.30%   100.0%  Chrome_ChildIOT
       3.17%   100.0%  Isolated Web Co
       ...
---
 tools/perf/ui/hist.c   | 35 +++++++++++++++++++++++++++++++++--
 tools/perf/util/hist.c |  2 ++
 tools/perf/util/hist.h |  1 +
 3 files changed, 36 insertions(+), 2 deletions(-)
tools/perf/ui/hist.c +33 -2
index 427ce687ad815a62..661922c4d7863224 100644
@@ -178,6 +178,9 @@ int hpp__fmt_mem_stat(struct perf_hpp_fmt *fmt __maybe_unused, struct perf_hpp *
 	for (int i = 0; i < MEM_STAT_LEN; i++) {
 		u64 val = he->mem_stat[mem_stat_idx].entries[i];
 
+		if (hists->mem_stat_total[mem_stat_idx].entries[i] == 0)
+			continue;
+
 		ret += hpp__call_print_fn(hpp, print_fn, fmtstr, 100.0 * val / total);
 	}
 
@@ -405,12 +408,31 @@ static int hpp__header_mem_stat_fn(struct perf_hpp_fmt *fmt, struct perf_hpp *hp
 	int ret = 0;
 	int len;
 	enum mem_stat_type mst = hpp__mem_stat_type(fmt);
+	int mem_stat_idx = -1;
+
+	for (int i = 0; i < hists->nr_mem_stats; i++) {
+		if (hists->mem_stat_types[i] == mst) {
+			mem_stat_idx = i;
+			break;
+		}
+	}
+	assert(mem_stat_idx != -1);
 
-	(void)hists;
 	if (line == 0) {
 		int left, right;
 
-		len = fmt->len;
+		len = 0;
+		/* update fmt->len for acutally used columns only */
+		for (int i = 0; i < MEM_STAT_LEN; i++) {
+			if (hists->mem_stat_total[mem_stat_idx].entries[i])
+				len += MEM_STAT_PRINT_LEN;
+		}
+		fmt->len = len;
+
+		/* print header directly if single column only */
+		if (len == MEM_STAT_PRINT_LEN)
+			return scnprintf(hpp->buf, hpp->size, "%*s", len, fmt->name);
+
 		left = (len - strlen(fmt->name)) / 2 - 1;
 		right = len - left - strlen(fmt->name) - 2;
 
@@ -423,10 +445,14 @@ static int hpp__header_mem_stat_fn(struct perf_hpp_fmt *fmt, struct perf_hpp *hp
 				 left, graph_dotted_line, fmt->name, right, graph_dotted_line);
 	}
 
+
 	len = hpp->size;
 	for (int i = 0; i < MEM_STAT_LEN; i++) {
 		int printed;
 
+		if (hists->mem_stat_total[mem_stat_idx].entries[i] == 0)
+			continue;
+
 		printed = scnprintf(buf, len, "%*s", MEM_STAT_PRINT_LEN,
 				    mem_stat_name(mst, i));
 		ret += printed;
@@ -1214,6 +1240,11 @@ int perf_hpp__alloc_mem_stats(struct perf_hpp_list *list, struct evlist *evlist)
 		if (hists->mem_stat_types == NULL)
 			return -ENOMEM;
 
+		hists->mem_stat_total = calloc(nr_mem_stats,
+					       sizeof(*hists->mem_stat_total));
+		if (hists->mem_stat_total == NULL)
+			return -ENOMEM;
+
 		memcpy(hists->mem_stat_types, mst, nr_mem_stats * sizeof(*mst));
 		hists->nr_mem_stats = nr_mem_stats;
 	}
tools/perf/util/hist.c +2 -0
index 7759c1818c1ad168..afc6855327ab0de6 100644
@@ -354,6 +354,7 @@ static int hists__update_mem_stat(struct hists *hists, struct hist_entry *he,
 
 		assert(0 <= idx && idx < MEM_STAT_LEN);
 		he->mem_stat[i].entries[idx] += period;
+		hists->mem_stat_total[i].entries[idx] += period;
 	}
 	return 0;
 }
@@ -3054,6 +3055,7 @@ static void hists_evsel__exit(struct evsel *evsel)
 
 	hists__delete_all_entries(hists);
 	zfree(&hists->mem_stat_types);
+	zfree(&hists->mem_stat_total);
 
 	list_for_each_entry_safe(node, tmp, &hists->hpp_formats, list) {
 		perf_hpp_list__for_each_format_safe(&node->hpp, fmt, pos) {
tools/perf/util/hist.h +1 -0
index 3990cfc21b1615ae..fa5e886e5b04ec9b 100644
@@ -135,6 +135,7 @@ struct hists {
 	int			nr_hpp_node;
 	int			nr_mem_stats;
 	enum mem_stat_type	*mem_stat_types;
+	struct he_mem_stat	*mem_stat_total;
 };
 
 #define hists__has(__h, __f) (__h)->hpp_list->__f
Signature · 1 line
2.49.0.906.g1f30a19c02-goog
author maintainer
#11 lore raw

Re: [PATCH 08/11] perf hist: Hide unused mem stat columns

Quoted 16 lines · Arnaldo Carvalho de Melo
Quoted 6 lines · Namhyung Kim

Some mem_stat types don't use all 8 columns. And there are cases only samples in certain kinds of mem_stat types are available only. For that case hide columns which has no samples.

The new output for the previous data would be:

Quoted 8 lines
  $ perf mem report -F overhead,op,comm --stdio
  ...
  #           ------ Mem Op -------
  # Overhead     Load  Store  Other  Command
  # ........  .....................  ...............
  #
      44.85%    21.1%  30.7%  48.3%  swapper
      26.82%    98.8%   0.3%   0.9%  netsli-prober
Quoted 1 line

/me curious about this "Other" column.

They are instructions that don't have memory operations.

Quoted 33 lines

Maps to MEM_STAT_OP_OTHER, that comes from mem_stat_index, that comes from:

int mem_stat_index(const enum mem_stat_type mst, const u64 val)

{
        union perf_mem_data_src src = {
                .val = val,
        };

                int idx = mem_stat_index(hists->mem_stat_types[i],
                                         mem_info__const_data_src(mi)->val);

struct mem_info *mi

union perf_mem_data_src {
__u64 val;
struct {
__u64 mem_op:5, /* type of opcode */
mem_lvl:14, /* memory hierarchy level */
mem_snoop:5, /* snoop mode */
mem_lock:2, /* lock instr */
mem_dtlb:7, /* tlb access */
mem_lvl_num:4, /* memory hierarchy level number */
mem_remote:1, /* remote */
mem_snoopx:2, /* snoop mode, ext */
mem_blk:3, /* access blocked */
mem_hops:3, /* hop level */
mem_rsvd:18;
};
};

Quoted 2 lines

As the percentage for "Other" is so high I think some other patch in this series will elucidate that :-)

IIUC AMD IBS cannot sample memory instructions specifically. It'd just pick random uops/instructions and capture the data. So it's natural to see large 'Other' operations on AMD.

Quoted 147 lines

Lemme continue testing...

- Arnaldo

Quoted 142 lines
       7.19%    51.7%  13.7%  34.6%  perf
       5.81%    89.7%   2.2%   8.1%  qemu-system-ppc
       4.77%   100.0%   0.0%   0.0%  notifications_c
       1.77%    95.9%   1.2%   3.0%  MemoryReleaser
       0.77%    71.6%   4.1%  24.3%  DefaultEventMan
       0.19%    66.7%  22.2%  11.1%  gnome-shell
       ...

On Intel machines, the event is only for loads or stores so it'll have only one columns like below:

# Mem Op
# Overhead Load Command

  # ........  .......  ...............
  #
      20.55%   100.0%  swapper
      17.13%   100.0%  chrome
       9.02%   100.0%  data-loop.0
       6.26%   100.0%  pipewire-pulse
       5.63%   100.0%  threaded-ml
       5.47%   100.0%  GraphRunner
       5.37%   100.0%  AudioIP~allback
       5.30%   100.0%  Chrome_ChildIOT
       3.17%   100.0%  Isolated Web Co
       ...
---
 tools/perf/ui/hist.c   | 35 +++++++++++++++++++++++++++++++++--
 tools/perf/util/hist.c |  2 ++
 tools/perf/util/hist.h |  1 +
 3 files changed, 36 insertions(+), 2 deletions(-)
tools/perf/ui/hist.c +33 -2
index 427ce687ad815a62..661922c4d7863224 100644
@@ -178,6 +178,9 @@ int hpp__fmt_mem_stat(struct perf_hpp_fmt *fmt __maybe_unused, struct perf_hpp *
 	for (int i = 0; i < MEM_STAT_LEN; i++) {
 		u64 val = he->mem_stat[mem_stat_idx].entries[i];
 
+		if (hists->mem_stat_total[mem_stat_idx].entries[i] == 0)
+			continue;
+
 		ret += hpp__call_print_fn(hpp, print_fn, fmtstr, 100.0 * val / total);
 	}
 
@@ -405,12 +408,31 @@ static int hpp__header_mem_stat_fn(struct perf_hpp_fmt *fmt, struct perf_hpp *hp
 	int ret = 0;
 	int len;
 	enum mem_stat_type mst = hpp__mem_stat_type(fmt);
+	int mem_stat_idx = -1;
+
+	for (int i = 0; i < hists->nr_mem_stats; i++) {
+		if (hists->mem_stat_types[i] == mst) {
+			mem_stat_idx = i;
+			break;
+		}
+	}
+	assert(mem_stat_idx != -1);
 
-	(void)hists;
 	if (line == 0) {
 		int left, right;
 
-		len = fmt->len;
+		len = 0;
+		/* update fmt->len for acutally used columns only */
+		for (int i = 0; i < MEM_STAT_LEN; i++) {
+			if (hists->mem_stat_total[mem_stat_idx].entries[i])
+				len += MEM_STAT_PRINT_LEN;
+		}
+		fmt->len = len;
+
+		/* print header directly if single column only */
+		if (len == MEM_STAT_PRINT_LEN)
+			return scnprintf(hpp->buf, hpp->size, "%*s", len, fmt->name);
+
 		left = (len - strlen(fmt->name)) / 2 - 1;
 		right = len - left - strlen(fmt->name) - 2;
 
@@ -423,10 +445,14 @@ static int hpp__header_mem_stat_fn(struct perf_hpp_fmt *fmt, struct perf_hpp *hp
 				 left, graph_dotted_line, fmt->name, right, graph_dotted_line);
 	}
 
+
 	len = hpp->size;
 	for (int i = 0; i < MEM_STAT_LEN; i++) {
 		int printed;
 
+		if (hists->mem_stat_total[mem_stat_idx].entries[i] == 0)
+			continue;
+
 		printed = scnprintf(buf, len, "%*s", MEM_STAT_PRINT_LEN,
 				    mem_stat_name(mst, i));
 		ret += printed;
@@ -1214,6 +1240,11 @@ int perf_hpp__alloc_mem_stats(struct perf_hpp_list *list, struct evlist *evlist)
 		if (hists->mem_stat_types == NULL)
 			return -ENOMEM;
 
+		hists->mem_stat_total = calloc(nr_mem_stats,
+					       sizeof(*hists->mem_stat_total));
+		if (hists->mem_stat_total == NULL)
+			return -ENOMEM;
+
 		memcpy(hists->mem_stat_types, mst, nr_mem_stats * sizeof(*mst));
 		hists->nr_mem_stats = nr_mem_stats;
 	}
tools/perf/util/hist.c +2 -0
index 7759c1818c1ad168..afc6855327ab0de6 100644
@@ -354,6 +354,7 @@ static int hists__update_mem_stat(struct hists *hists, struct hist_entry *he,
 
 		assert(0 <= idx && idx < MEM_STAT_LEN);
 		he->mem_stat[i].entries[idx] += period;
+		hists->mem_stat_total[i].entries[idx] += period;
 	}
 	return 0;
 }
@@ -3054,6 +3055,7 @@ static void hists_evsel__exit(struct evsel *evsel)
 
 	hists__delete_all_entries(hists);
 	zfree(&hists->mem_stat_types);
+	zfree(&hists->mem_stat_total);
 
 	list_for_each_entry_safe(node, tmp, &hists->hpp_formats, list) {
 		perf_hpp_list__for_each_format_safe(&node->hpp, fmt, pos) {
tools/perf/util/hist.h +1 -0
index 3990cfc21b1615ae..fa5e886e5b04ec9b 100644
@@ -135,6 +135,7 @@ struct hists {
 	int			nr_hpp_node;
 	int			nr_mem_stats;
 	enum mem_stat_type	*mem_stat_types;
+	struct he_mem_stat	*mem_stat_total;
 };
 
 #define hists__has(__h, __f) (__h)->hpp_list->__f
Signature · 1 line
2.49.0.906.g1f30a19c02-goog
author maintainer
#12 lore raw

PATCH 9/11 perf mem: Add 'cache' and 'memory' output fields

In reply to Namhyung Kim

This is a breakdown of perf_mem_data_src.mem_lvl_num. But it's also divided into two parts because the combination is bigger than 8.

Since there are many entries for different cache levels, 'cache' field focuses on them. I generalized buffers like LFB, MAB and MHB to L1-buf and L2-buf.

The rest goes to 'memory' field which can be RAM, CXL, PMEM, IO, etc.

  $ perf mem report -F cache,mem,dso --stdio
  ...
  #
  # -------------- Cache --------------  --- Memory ---
  #      L1     L2     L3 L1-buf  Other      RAM  Other  Shared Object
  # ...................................  ..............  ....................................
  #
      53.9%   3.6%  16.2%  21.6%   4.8%     4.8%  95.2%  [kernel.kallsyms]
      64.7%   1.7%   3.5%  17.4%  12.8%    12.8%  87.2%  chrome (deleted)
      78.3%   2.8%   0.0%   1.0%  17.9%    17.9%  82.1%  libc.so.6
      39.6%   1.5%   0.0%   5.7%  53.2%    53.2%  46.8%  libxul.so
      26.2%   0.0%   0.0%   0.0%  73.8%    73.8%  26.2%  [unknown]
      85.5%   0.0%   0.0%  14.5%   0.0%     0.0% 100.0%  libspa-audioconvert.so
      66.3%   4.4%   0.0%  29.4%   0.0%     0.0% 100.0%  libglib-2.0.so.0.8200.1 (deleted)
       1.9%   0.0%   0.0%   0.0%  98.1%    98.1%   1.9%  libmutter-cogl-15.so.0.0.0 (deleted)
      10.6%   0.0%   0.0%  89.4%   0.0%     0.0% 100.0%  libpulsecommon-16.1.so
       0.0%   0.0%   0.0% 100.0%   0.0%     0.0% 100.0%  libfreeblpriv3.so (deleted)
       ...
---
 tools/perf/ui/browsers/hists.c |  6 +++
 tools/perf/ui/hist.c           | 10 +++++
 tools/perf/util/hist.h         |  2 +
 tools/perf/util/mem-events.c   | 71 +++++++++++++++++++++++++++++++++-
 tools/perf/util/mem-events.h   | 24 +++++++++++-
 tools/perf/util/sort.c         |  2 +
 6 files changed, 113 insertions(+), 2 deletions(-)
tools/perf/ui/browsers/hists.c +6 -0
index 66a4c769b2d76436..675dd64067747126 100644
@@ -1285,6 +1285,8 @@ __HPP_COLOR_PERCENT_FN(overhead_guest_us, period_guest_us, PERF_HPP_FMT_TYPE__PE
 __HPP_COLOR_ACC_PERCENT_FN(overhead_acc, period, PERF_HPP_FMT_TYPE__PERCENT)
 __HPP_COLOR_ACC_PERCENT_FN(latency_acc, latency, PERF_HPP_FMT_TYPE__LATENCY)
 __HPP_COLOR_MEM_STAT_FN(op, OP)
+__HPP_COLOR_MEM_STAT_FN(cache, CACHE)
+__HPP_COLOR_MEM_STAT_FN(memory, MEMORY)
 
 #undef __HPP_COLOR_PERCENT_FN
 #undef __HPP_COLOR_ACC_PERCENT_FN
@@ -1310,6 +1312,10 @@ void hist_browser__init_hpp(void)
 				hist_browser__hpp_color_latency_acc;
 	perf_hpp__format[PERF_HPP__MEM_STAT_OP].color =
 				hist_browser__hpp_color_mem_stat_op;
+	perf_hpp__format[PERF_HPP__MEM_STAT_CACHE].color =
+				hist_browser__hpp_color_mem_stat_cache;
+	perf_hpp__format[PERF_HPP__MEM_STAT_MEMORY].color =
+				hist_browser__hpp_color_mem_stat_memory;
 
 	res_sample_init();
 }
tools/perf/ui/hist.c +10 -0
index 661922c4d7863224..7fc09c738ed02acb 100644
@@ -348,6 +348,10 @@ static enum mem_stat_type hpp__mem_stat_type(struct perf_hpp_fmt *fmt)
 	switch (fmt->idx) {
 	case PERF_HPP__MEM_STAT_OP:
 		return PERF_MEM_STAT_OP;
+	case PERF_HPP__MEM_STAT_CACHE:
+		return PERF_MEM_STAT_CACHE;
+	case PERF_HPP__MEM_STAT_MEMORY:
+		return PERF_MEM_STAT_MEMORY;
 	default:
 		break;
 	}
@@ -644,6 +648,8 @@ HPP_AVERAGE_FNS(weight2, weight2)
 HPP_AVERAGE_FNS(weight3, weight3)
 
 HPP_MEM_STAT_FNS(op, OP)
+HPP_MEM_STAT_FNS(cache, CACHE)
+HPP_MEM_STAT_FNS(memory, MEMORY)
 
 static int64_t hpp__nop_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 			    struct hist_entry *a __maybe_unused,
@@ -748,6 +754,8 @@ struct perf_hpp_fmt perf_hpp__format[] = {
 	HPP__PRINT_FNS("Weight2", weight2, WEIGHT2),
 	HPP__PRINT_FNS("Weight3", weight3, WEIGHT3),
 	HPP__MEM_STAT_PRINT_FNS("Mem Op", op, OP),
+	HPP__MEM_STAT_PRINT_FNS("Cache", cache, CACHE),
+	HPP__MEM_STAT_PRINT_FNS("Memory", memory, MEMORY),
 };
 
 struct perf_hpp_list perf_hpp_list = {
@@ -1103,6 +1111,8 @@ void perf_hpp__reset_width(struct perf_hpp_fmt *fmt, struct hists *hists)
 		break;
 
 	case PERF_HPP__MEM_STAT_OP:
+	case PERF_HPP__MEM_STAT_CACHE:
+	case PERF_HPP__MEM_STAT_MEMORY:
 		fmt->len = MEM_STAT_LEN * MEM_STAT_PRINT_LEN;
 		break;
 
tools/perf/util/hist.h +2 -0
index fa5e886e5b04ec9b..9de50d929ad1268c 100644
@@ -589,6 +589,8 @@ enum {
 	PERF_HPP__WEIGHT2,
 	PERF_HPP__WEIGHT3,
 	PERF_HPP__MEM_STAT_OP,
+	PERF_HPP__MEM_STAT_CACHE,
+	PERF_HPP__MEM_STAT_MEMORY,
 
 	PERF_HPP__MAX_INDEX
 };
tools/perf/util/mem-events.c +70 -1

85 lines not on this page · raw message

tools/perf/util/mem-events.h +23 -1
index 55e5e2607fb732b4..002e2772400e3dda 100644
@@ -91,8 +91,12 @@ void c2c_add_stats(struct c2c_stats *stats, struct c2c_stats *add);
 
 enum mem_stat_type {
 	PERF_MEM_STAT_OP,
+	PERF_MEM_STAT_CACHE,
+	PERF_MEM_STAT_MEMORY,
 };
 
+#define MEM_STAT_PRINT_LEN  7  /* 1 space + 5 digits + 1 percent sign */
+
 enum mem_stat_op {
 	MEM_STAT_OP_LOAD,
 	MEM_STAT_OP_STORE,
@@ -102,7 +106,25 @@ enum mem_stat_op {
 	MEM_STAT_OP_OTHER,
 };
 
-#define MEM_STAT_PRINT_LEN  7  /* 1 space + 5 digits + 1 percent sign */
+enum mem_stat_cache {
+	MEM_STAT_CACHE_L1,
+	MEM_STAT_CACHE_L2,
+	MEM_STAT_CACHE_L3,
+	MEM_STAT_CACHE_L4,
+	MEM_STAT_CACHE_L1_BUF,
+	MEM_STAT_CACHE_L2_BUF,
+	MEM_STAT_CACHE_OTHER,
+};
+
+enum mem_stat_memory {
+	MEM_STAT_MEMORY_RAM,
+	MEM_STAT_MEMORY_MSC,
+	MEM_STAT_MEMORY_UNC,
+	MEM_STAT_MEMORY_CXL,
+	MEM_STAT_MEMORY_IO,
+	MEM_STAT_MEMORY_PMEM,
+	MEM_STAT_MEMORY_OTHER,
+};
 
 int mem_stat_index(const enum mem_stat_type mst, const u64 data_src);
 const char *mem_stat_name(const enum mem_stat_type mst, const int idx);
tools/perf/util/sort.c +2 -0
index 53fcb9191ea0cdc3..2ad88f7de95a2247 100644
@@ -2624,6 +2624,8 @@ static struct hpp_dimension hpp_sort_dimensions[] = {
 	DIM(PERF_HPP__WEIGHT3, "p_stage_cyc"),
 	/* used for output only when SORT_MODE__MEM */
 	DIM_MEM(PERF_HPP__MEM_STAT_OP, "op"),
+	DIM_MEM(PERF_HPP__MEM_STAT_CACHE, "cache"),
+	DIM_MEM(PERF_HPP__MEM_STAT_MEMORY, "memory"),
 };
 
 #undef DIM_MEM
Signature · 1 line
2.49.0.906.g1f30a19c02-goog
author maintainer
#13 lore raw

PATCH 10/11 perf mem: Add 'snoop' output field

In reply to Namhyung Kim

This is a breakdown of perf_mem_data_src.mem_snoop values. For now, it doesn't use mem_snoopx values like FWD and PEER.

  $ perf mem report -F overhead,snoop,comm --stdio
  ...
  #           ---------- Snoop -----------
  # Overhead      Hit   HitM   Miss  Other  Command
  # ........  ............................  ...............
  #
      34.24%     0.6%   0.0%   0.0%  99.4%  gnome-shell
      12.02%     1.0%   0.0%   0.0%  99.0%  chrome
       9.32%     1.0%   0.0%   0.3%  98.7%  Isolated Web Co
       6.85%     1.0%   0.3%   0.0%  98.6%  swapper
       6.30%     0.8%   0.8%   0.0%  98.5%  Xorg
       3.02%     2.4%   0.0%   0.0%  97.6%  VizCompositorTh
       2.35%     0.0%   0.0%   0.0% 100.0%  firefox-esr
       2.04%     0.0%   0.0%   0.0% 100.0%  JS Helper
       1.51%     3.2%   0.0%   0.0%  96.8%  threaded-ml
       1.44%     0.0%   0.0%   0.0% 100.0%  AudioIP~allback
       ...
---
 tools/perf/ui/browsers/hists.c |  3 +++
 tools/perf/ui/hist.c           |  5 +++++
 tools/perf/util/hist.h         |  1 +
 tools/perf/util/mem-events.c   | 23 +++++++++++++++++++++++
 tools/perf/util/mem-events.h   |  8 ++++++++
 tools/perf/util/sort.c         |  1 +
 6 files changed, 41 insertions(+)
tools/perf/ui/browsers/hists.c +3 -0
index 675dd64067747126..5b080f5062440246 100644
@@ -1287,6 +1287,7 @@ __HPP_COLOR_ACC_PERCENT_FN(latency_acc, latency, PERF_HPP_FMT_TYPE__LATENCY)
 __HPP_COLOR_MEM_STAT_FN(op, OP)
 __HPP_COLOR_MEM_STAT_FN(cache, CACHE)
 __HPP_COLOR_MEM_STAT_FN(memory, MEMORY)
+__HPP_COLOR_MEM_STAT_FN(snoop, SNOOP)
 
 #undef __HPP_COLOR_PERCENT_FN
 #undef __HPP_COLOR_ACC_PERCENT_FN
@@ -1316,6 +1317,8 @@ void hist_browser__init_hpp(void)
 				hist_browser__hpp_color_mem_stat_cache;
 	perf_hpp__format[PERF_HPP__MEM_STAT_MEMORY].color =
 				hist_browser__hpp_color_mem_stat_memory;
+	perf_hpp__format[PERF_HPP__MEM_STAT_SNOOP].color =
+				hist_browser__hpp_color_mem_stat_snoop;
 
 	res_sample_init();
 }
tools/perf/ui/hist.c +5 -0
index 7fc09c738ed02acb..94024dfa8dccf9ba 100644
@@ -352,6 +352,8 @@ static enum mem_stat_type hpp__mem_stat_type(struct perf_hpp_fmt *fmt)
 		return PERF_MEM_STAT_CACHE;
 	case PERF_HPP__MEM_STAT_MEMORY:
 		return PERF_MEM_STAT_MEMORY;
+	case PERF_HPP__MEM_STAT_SNOOP:
+		return PERF_MEM_STAT_SNOOP;
 	default:
 		break;
 	}
@@ -650,6 +652,7 @@ HPP_AVERAGE_FNS(weight3, weight3)
 HPP_MEM_STAT_FNS(op, OP)
 HPP_MEM_STAT_FNS(cache, CACHE)
 HPP_MEM_STAT_FNS(memory, MEMORY)
+HPP_MEM_STAT_FNS(snoop, SNOOP)
 
 static int64_t hpp__nop_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 			    struct hist_entry *a __maybe_unused,
@@ -756,6 +759,7 @@ struct perf_hpp_fmt perf_hpp__format[] = {
 	HPP__MEM_STAT_PRINT_FNS("Mem Op", op, OP),
 	HPP__MEM_STAT_PRINT_FNS("Cache", cache, CACHE),
 	HPP__MEM_STAT_PRINT_FNS("Memory", memory, MEMORY),
+	HPP__MEM_STAT_PRINT_FNS("Snoop", snoop, SNOOP),
 };
 
 struct perf_hpp_list perf_hpp_list = {
@@ -1113,6 +1117,7 @@ void perf_hpp__reset_width(struct perf_hpp_fmt *fmt, struct hists *hists)
 	case PERF_HPP__MEM_STAT_OP:
 	case PERF_HPP__MEM_STAT_CACHE:
 	case PERF_HPP__MEM_STAT_MEMORY:
+	case PERF_HPP__MEM_STAT_SNOOP:
 		fmt->len = MEM_STAT_LEN * MEM_STAT_PRINT_LEN;
 		break;
 
tools/perf/util/hist.h +1 -0
index 9de50d929ad1268c..c2d286c4ba395674 100644
@@ -591,6 +591,7 @@ enum {
 	PERF_HPP__MEM_STAT_OP,
 	PERF_HPP__MEM_STAT_CACHE,
 	PERF_HPP__MEM_STAT_MEMORY,
+	PERF_HPP__MEM_STAT_SNOOP,
 
 	PERF_HPP__MAX_INDEX
 };
tools/perf/util/mem-events.c +23 -0
index 6822815278a4b213..ddcfc6500d77a9e6 100644
@@ -857,6 +857,17 @@ int mem_stat_index(const enum mem_stat_type mst, const u64 val)
 		default:
 			return MEM_STAT_MEMORY_OTHER;
 		}
+	case PERF_MEM_STAT_SNOOP:
+		switch (src.mem_snoop) {
+		case PERF_MEM_SNOOP_HIT:
+			return MEM_STAT_SNOOP_HIT;
+		case PERF_MEM_SNOOP_HITM:
+			return MEM_STAT_SNOOP_HITM;
+		case PERF_MEM_SNOOP_MISS:
+			return MEM_STAT_SNOOP_MISS;
+		default:
+			return MEM_STAT_SNOOP_OTHER;
+		}
 	default:
 		break;
 	}
@@ -919,6 +930,18 @@ const char *mem_stat_name(const enum mem_stat_type mst, const int idx)
 		default:
 			return "Other";
 		}
+	case PERF_MEM_STAT_SNOOP:
+		switch (idx) {
+		case MEM_STAT_SNOOP_HIT:
+			return "Hit";
+		case MEM_STAT_SNOOP_HITM:
+			return "HitM";
+		case MEM_STAT_SNOOP_MISS:
+			return "Miss";
+		case MEM_STAT_SNOOP_OTHER:
+		default:
+			return "Other";
+		}
 	default:
 		break;
 	}
tools/perf/util/mem-events.h +8 -0
index 002e2772400e3dda..4d8f18583af42550 100644
@@ -93,6 +93,7 @@ enum mem_stat_type {
 	PERF_MEM_STAT_OP,
 	PERF_MEM_STAT_CACHE,
 	PERF_MEM_STAT_MEMORY,
+	PERF_MEM_STAT_SNOOP,
 };
 
 #define MEM_STAT_PRINT_LEN  7  /* 1 space + 5 digits + 1 percent sign */
@@ -126,6 +127,13 @@ enum mem_stat_memory {
 	MEM_STAT_MEMORY_OTHER,
 };
 
+enum mem_stat_snoop {
+	MEM_STAT_SNOOP_HIT,
+	MEM_STAT_SNOOP_HITM,
+	MEM_STAT_SNOOP_MISS,
+	MEM_STAT_SNOOP_OTHER,
+};
+
 int mem_stat_index(const enum mem_stat_type mst, const u64 data_src);
 const char *mem_stat_name(const enum mem_stat_type mst, const int idx);
 
tools/perf/util/sort.c +1 -0
index 2ad88f7de95a2247..51a210d874327d3a 100644
@@ -2626,6 +2626,7 @@ static struct hpp_dimension hpp_sort_dimensions[] = {
 	DIM_MEM(PERF_HPP__MEM_STAT_OP, "op"),
 	DIM_MEM(PERF_HPP__MEM_STAT_CACHE, "cache"),
 	DIM_MEM(PERF_HPP__MEM_STAT_MEMORY, "memory"),
+	DIM_MEM(PERF_HPP__MEM_STAT_SNOOP, "snoop"),
 };
 
 #undef DIM_MEM
Signature · 1 line
2.49.0.906.g1f30a19c02-goog
author maintainer
#14 lore raw

PATCH 11/11 perf mem: Add 'dtlb' output field

In reply to Namhyung Kim

This is a breakdown of perf_mem_data_src.mem_dtlb values. It assumes PMU drivers would set PERF_MEM_TLB_HIT bit with an appropriate level. And having PERF_MEM_TLB_MISS means that it failed to find one in any levels of TLB. For now, it doesn't use PERF_MEM_TLB_{WK,OS} bits.

Also it seems Intel machines don't distinguish L1 or L2 precisely. So I added ANY_HIT (printed as "L?-Hit") to handle the case.

  $ perf mem report -F overhead,dtlb,dso --stdio
  ...
  #           --- D-TLB ----
  # Overhead   L?-Hit   Miss  Shared Object
  # ........  ..............  .................
  #
      67.03%    99.5%   0.5%  [unknown]
      31.23%    99.2%   0.8%  [kernel.kallsyms]
       1.08%    97.8%   2.2%  [i915]
       0.36%   100.0%   0.0%  [JIT] tid 6853
       0.12%   100.0%   0.0%  [drm]
       0.05%   100.0%   0.0%  [drm_kms_helper]
       0.05%   100.0%   0.0%  [ext4]
       0.02%   100.0%   0.0%  [aesni_intel]
       0.02%   100.0%   0.0%  [crc32c_intel]
       0.02%   100.0%   0.0%  [dm_crypt]
       ...
---
 tools/perf/ui/browsers/hists.c |  3 +++
 tools/perf/ui/hist.c           |  5 +++++
 tools/perf/util/hist.h         |  1 +
 tools/perf/util/mem-events.c   | 27 +++++++++++++++++++++++++++
 tools/perf/util/mem-events.h   |  9 +++++++++
 tools/perf/util/sort.c         |  1 +
 6 files changed, 46 insertions(+)
tools/perf/ui/browsers/hists.c +3 -0
index 5b080f5062440246..d26b925e3d7f46af 100644
@@ -1288,6 +1288,7 @@ __HPP_COLOR_MEM_STAT_FN(op, OP)
 __HPP_COLOR_MEM_STAT_FN(cache, CACHE)
 __HPP_COLOR_MEM_STAT_FN(memory, MEMORY)
 __HPP_COLOR_MEM_STAT_FN(snoop, SNOOP)
+__HPP_COLOR_MEM_STAT_FN(dtlb, DTLB)
 
 #undef __HPP_COLOR_PERCENT_FN
 #undef __HPP_COLOR_ACC_PERCENT_FN
@@ -1319,6 +1320,8 @@ void hist_browser__init_hpp(void)
 				hist_browser__hpp_color_mem_stat_memory;
 	perf_hpp__format[PERF_HPP__MEM_STAT_SNOOP].color =
 				hist_browser__hpp_color_mem_stat_snoop;
+	perf_hpp__format[PERF_HPP__MEM_STAT_DTLB].color =
+				hist_browser__hpp_color_mem_stat_dtlb;
 
 	res_sample_init();
 }
tools/perf/ui/hist.c +5 -0
index 94024dfa8dccf9ba..ed5c40ebd906f076 100644
@@ -354,6 +354,8 @@ static enum mem_stat_type hpp__mem_stat_type(struct perf_hpp_fmt *fmt)
 		return PERF_MEM_STAT_MEMORY;
 	case PERF_HPP__MEM_STAT_SNOOP:
 		return PERF_MEM_STAT_SNOOP;
+	case PERF_HPP__MEM_STAT_DTLB:
+		return PERF_MEM_STAT_DTLB;
 	default:
 		break;
 	}
@@ -653,6 +655,7 @@ HPP_MEM_STAT_FNS(op, OP)
 HPP_MEM_STAT_FNS(cache, CACHE)
 HPP_MEM_STAT_FNS(memory, MEMORY)
 HPP_MEM_STAT_FNS(snoop, SNOOP)
+HPP_MEM_STAT_FNS(dtlb, DTLB)
 
 static int64_t hpp__nop_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 			    struct hist_entry *a __maybe_unused,
@@ -760,6 +763,7 @@ struct perf_hpp_fmt perf_hpp__format[] = {
 	HPP__MEM_STAT_PRINT_FNS("Cache", cache, CACHE),
 	HPP__MEM_STAT_PRINT_FNS("Memory", memory, MEMORY),
 	HPP__MEM_STAT_PRINT_FNS("Snoop", snoop, SNOOP),
+	HPP__MEM_STAT_PRINT_FNS("D-TLB", dtlb, DTLB),
 };
 
 struct perf_hpp_list perf_hpp_list = {
@@ -1118,6 +1122,7 @@ void perf_hpp__reset_width(struct perf_hpp_fmt *fmt, struct hists *hists)
 	case PERF_HPP__MEM_STAT_CACHE:
 	case PERF_HPP__MEM_STAT_MEMORY:
 	case PERF_HPP__MEM_STAT_SNOOP:
+	case PERF_HPP__MEM_STAT_DTLB:
 		fmt->len = MEM_STAT_LEN * MEM_STAT_PRINT_LEN;
 		break;
 
tools/perf/util/hist.h +1 -0
index c2d286c4ba395674..355198fd70281f43 100644
@@ -592,6 +592,7 @@ enum {
 	PERF_HPP__MEM_STAT_CACHE,
 	PERF_HPP__MEM_STAT_MEMORY,
 	PERF_HPP__MEM_STAT_SNOOP,
+	PERF_HPP__MEM_STAT_DTLB,
 
 	PERF_HPP__MAX_INDEX
 };
tools/perf/util/mem-events.c +27 -0
index ddcfc6500d77a9e6..3e9131e05348a996 100644
@@ -868,6 +868,19 @@ int mem_stat_index(const enum mem_stat_type mst, const u64 val)
 		default:
 			return MEM_STAT_SNOOP_OTHER;
 		}
+	case PERF_MEM_STAT_DTLB:
+		switch (src.mem_dtlb) {
+		case PERF_MEM_TLB_L1 | PERF_MEM_TLB_HIT:
+			return MEM_STAT_DTLB_L1_HIT;
+		case PERF_MEM_TLB_L2 | PERF_MEM_TLB_HIT:
+			return MEM_STAT_DTLB_L2_HIT;
+		case PERF_MEM_TLB_L1 | PERF_MEM_TLB_L2 | PERF_MEM_TLB_HIT:
+			return MEM_STAT_DTLB_ANY_HIT;
+		default:
+			if (src.mem_dtlb & PERF_MEM_TLB_MISS)
+				return MEM_STAT_DTLB_MISS;
+			return MEM_STAT_DTLB_OTHER;
+		}
 	default:
 		break;
 	}
@@ -942,6 +955,20 @@ const char *mem_stat_name(const enum mem_stat_type mst, const int idx)
 		default:
 			return "Other";
 		}
+	case PERF_MEM_STAT_DTLB:
+		switch (idx) {
+		case MEM_STAT_DTLB_L1_HIT:
+			return "L1-Hit";
+		case MEM_STAT_DTLB_L2_HIT:
+			return "L2-Hit";
+		case MEM_STAT_DTLB_ANY_HIT:
+			return "L?-Hit";
+		case MEM_STAT_DTLB_MISS:
+			return "Miss";
+		case MEM_STAT_DTLB_OTHER:
+		default:
+			return "Other";
+		}
 	default:
 		break;
 	}
tools/perf/util/mem-events.h +9 -0
index 4d8f18583af42550..5b98076904b0b689 100644
@@ -94,6 +94,7 @@ enum mem_stat_type {
 	PERF_MEM_STAT_CACHE,
 	PERF_MEM_STAT_MEMORY,
 	PERF_MEM_STAT_SNOOP,
+	PERF_MEM_STAT_DTLB,
 };
 
 #define MEM_STAT_PRINT_LEN  7  /* 1 space + 5 digits + 1 percent sign */
@@ -134,6 +135,14 @@ enum mem_stat_snoop {
 	MEM_STAT_SNOOP_OTHER,
 };
 
+enum mem_stat_dtlb {
+	MEM_STAT_DTLB_L1_HIT,
+	MEM_STAT_DTLB_L2_HIT,
+	MEM_STAT_DTLB_ANY_HIT,
+	MEM_STAT_DTLB_MISS,
+	MEM_STAT_DTLB_OTHER,
+};
+
 int mem_stat_index(const enum mem_stat_type mst, const u64 data_src);
 const char *mem_stat_name(const enum mem_stat_type mst, const int idx);
 
tools/perf/util/sort.c +1 -0
index 51a210d874327d3a..8efafa7c10822ee9 100644
@@ -2627,6 +2627,7 @@ static struct hpp_dimension hpp_sort_dimensions[] = {
 	DIM_MEM(PERF_HPP__MEM_STAT_CACHE, "cache"),
 	DIM_MEM(PERF_HPP__MEM_STAT_MEMORY, "memory"),
 	DIM_MEM(PERF_HPP__MEM_STAT_SNOOP, "snoop"),
+	DIM_MEM(PERF_HPP__MEM_STAT_DTLB, "dtlb"),
 };
 
 #undef DIM_MEM
Signature · 1 line
2.49.0.906.g1f30a19c02-goog
#15 lore raw

Re: [PATCH 11/11] perf mem: Add 'dtlb' output field

Quoted 17 lines · Namhyung Kim

This is a breakdown of perf_mem_data_src.mem_dtlb values. It assumes PMU drivers would set PERF_MEM_TLB_HIT bit with an appropriate level. And having PERF_MEM_TLB_MISS means that it failed to find one in any levels of TLB. For now, it doesn't use PERF_MEM_TLB_{WK,OS} bits.

Also it seems Intel machines don't distinguish L1 or L2 precisely. So I added ANY_HIT (printed as "L?-Hit") to handle the case.

  $ perf mem report -F overhead,dtlb,dso --stdio
  ...
  #           --- D-TLB ----
  # Overhead   L?-Hit   Miss  Shared Object
  # ........  ..............  .................
  #
      67.03%    99.5%   0.5%  [unknown]
      31.23%    99.2%   0.8%  [kernel.kallsyms]
       1.08%    97.8%   2.2%  [i915]
Quoted 8 lines
       0.36%   100.0%   0.0%  [JIT] tid 6853
       0.12%   100.0%   0.0%  [drm]
       0.05%   100.0%   0.0%  [drm_kms_helper]
       0.05%   100.0%   0.0%  [ext4]
       0.02%   100.0%   0.0%  [aesni_intel]
       0.02%   100.0%   0.0%  [crc32c_intel]
       0.02%   100.0%   0.0%  [dm_crypt]
       ...
root@number:~# perf report --header | grep cpudesc
# cpudesc : AMD Ryzen 9 9950X3D 16-Core Processor
root@number:~# perf mem report -F overhead,dtlb,dso --stdio | head -20
# To display the perf.data header info, please use --header/--header-only options.
#
#
# Total Lost Samples: 0
#
# Samples: 2K of event 'cycles:P'
# Total weight : 2637
# Sort order   : local_weight,mem,sym,dso,symbol_daddr,dso_daddr,snoop,tlb,locked,blocked,local_ins_lat,local_p_stage_cyc
#
#           ---------- D-TLB -----------                                   
# Overhead   L1-Hit L2-Hit   Miss  Other  Shared Object                    
# ........  ............................  .................................
#
    77.47%    18.4%   0.1%   0.6%  80.9%  [kernel.kallsyms]                
     5.61%    36.5%   0.7%   1.4%  61.5%  libxul.so                        
     2.77%    39.7%   0.0%  12.3%  47.9%  libc.so.6                        
     2.01%    34.0%   1.9%   1.9%  62.3%  libglib-2.0.so.0.8400.1          
     1.93%    31.4%   2.0%   2.0%  64.7%  [amdgpu]                         
     1.63%    48.8%   0.0%   0.0%  51.2%  [JIT] tid 60168                  
     1.14%     3.3%   0.0%   0.0%  96.7%  [vdso]                           
root@number:~#
author maintainer
#16 lore raw

Re: [PATCH 11/11] perf mem: Add 'dtlb' output field

Quoted 43 lines · Arnaldo Carvalho de Melo
Quoted 17 lines · Namhyung Kim

This is a breakdown of perf_mem_data_src.mem_dtlb values. It assumes PMU drivers would set PERF_MEM_TLB_HIT bit with an appropriate level. And having PERF_MEM_TLB_MISS means that it failed to find one in any levels of TLB. For now, it doesn't use PERF_MEM_TLB_{WK,OS} bits.

Also it seems Intel machines don't distinguish L1 or L2 precisely. So I added ANY_HIT (printed as "L?-Hit") to handle the case.

  $ perf mem report -F overhead,dtlb,dso --stdio
  ...
  #           --- D-TLB ----
  # Overhead   L?-Hit   Miss  Shared Object
  # ........  ..............  .................
  #
      67.03%    99.5%   0.5%  [unknown]
      31.23%    99.2%   0.8%  [kernel.kallsyms]
       1.08%    97.8%   2.2%  [i915]
Quoted 8 lines
       0.36%   100.0%   0.0%  [JIT] tid 6853
       0.12%   100.0%   0.0%  [drm]
       0.05%   100.0%   0.0%  [drm_kms_helper]
       0.05%   100.0%   0.0%  [ext4]
       0.02%   100.0%   0.0%  [aesni_intel]
       0.02%   100.0%   0.0%  [crc32c_intel]
       0.02%   100.0%   0.0%  [dm_crypt]
       ...
root@number:~# perf report --header | grep cpudesc
# cpudesc : AMD Ryzen 9 9950X3D 16-Core Processor
root@number:~# perf mem report -F overhead,dtlb,dso --stdio | head -20
# To display the perf.data header info, please use --header/--header-only options.
#
#
# Total Lost Samples: 0
#
# Samples: 2K of event 'cycles:P'
# Total weight : 2637
# Sort order   : local_weight,mem,sym,dso,symbol_daddr,dso_daddr,snoop,tlb,locked,blocked,local_ins_lat,local_p_stage_cyc
#
#           ---------- D-TLB -----------                                   
# Overhead   L1-Hit L2-Hit   Miss  Other  Shared Object                    
# ........  ............................  .................................
#
Quoted 8 lines
    77.47%    18.4%   0.1%   0.6%  80.9%  [kernel.kallsyms]                
     5.61%    36.5%   0.7%   1.4%  61.5%  libxul.so                        
     2.77%    39.7%   0.0%  12.3%  47.9%  libc.so.6                        
     2.01%    34.0%   1.9%   1.9%  62.3%  libglib-2.0.so.0.8400.1          
     1.93%    31.4%   2.0%   2.0%  64.7%  [amdgpu]                         
     1.63%    48.8%   0.0%   0.0%  51.2%  [JIT] tid 60168                  
     1.14%     3.3%   0.0%   0.0%  96.7%  [vdso]                           
root@number:~#

I guess it's because those samples don't have mem_info as they are not memory instructions.

Can you please re-run the perf record with filters like below?

  $ perf record -aW --sample-mem-info -e cycles:P --filter 'mem_op == load || mem_op == store' sleep 1

Thanks,
Namhyung

#17 lore raw

Re: [PATCH 11/11] perf mem: Add 'dtlb' output field

Quoted 53 lines · Namhyung Kim
Quoted 43 lines · Arnaldo Carvalho de Melo
Quoted 17 lines · Namhyung Kim

This is a breakdown of perf_mem_data_src.mem_dtlb values. It assumes PMU drivers would set PERF_MEM_TLB_HIT bit with an appropriate level. And having PERF_MEM_TLB_MISS means that it failed to find one in any levels of TLB. For now, it doesn't use PERF_MEM_TLB_{WK,OS} bits.

Also it seems Intel machines don't distinguish L1 or L2 precisely. So I added ANY_HIT (printed as "L?-Hit") to handle the case.

  $ perf mem report -F overhead,dtlb,dso --stdio
  ...
  #           --- D-TLB ----
  # Overhead   L?-Hit   Miss  Shared Object
  # ........  ..............  .................
  #
      67.03%    99.5%   0.5%  [unknown]
      31.23%    99.2%   0.8%  [kernel.kallsyms]
       1.08%    97.8%   2.2%  [i915]
Quoted 8 lines
       0.36%   100.0%   0.0%  [JIT] tid 6853
       0.12%   100.0%   0.0%  [drm]
       0.05%   100.0%   0.0%  [drm_kms_helper]
       0.05%   100.0%   0.0%  [ext4]
       0.02%   100.0%   0.0%  [aesni_intel]
       0.02%   100.0%   0.0%  [crc32c_intel]
       0.02%   100.0%   0.0%  [dm_crypt]
       ...
root@number:~# perf report --header | grep cpudesc
# cpudesc : AMD Ryzen 9 9950X3D 16-Core Processor
root@number:~# perf mem report -F overhead,dtlb,dso --stdio | head -20
# To display the perf.data header info, please use --header/--header-only options.
#
#
# Total Lost Samples: 0
#
# Samples: 2K of event 'cycles:P'
# Total weight : 2637
# Sort order   : local_weight,mem,sym,dso,symbol_daddr,dso_daddr,snoop,tlb,locked,blocked,local_ins_lat,local_p_stage_cyc
#
#           ---------- D-TLB -----------                                   
# Overhead   L1-Hit L2-Hit   Miss  Other  Shared Object                    
# ........  ............................  .................................
#
Quoted 8 lines
    77.47%    18.4%   0.1%   0.6%  80.9%  [kernel.kallsyms]                
     5.61%    36.5%   0.7%   1.4%  61.5%  libxul.so                        
     2.77%    39.7%   0.0%  12.3%  47.9%  libc.so.6                        
     2.01%    34.0%   1.9%   1.9%  62.3%  libglib-2.0.so.0.8400.1          
     1.93%    31.4%   2.0%   2.0%  64.7%  [amdgpu]                         
     1.63%    48.8%   0.0%   0.0%  51.2%  [JIT] tid 60168                  
     1.14%     3.3%   0.0%   0.0%  96.7%  [vdso]                           
root@number:~#
Quoted 6 lines

I guess it's because those samples don't have mem_info as they are not memory instructions.

Can you please re-run the perf record with filters like below?

  $ perf record -aW --sample-mem-info -e cycles:P --filter 'mem_op == load || mem_op == store' sleep 1

I tried, it got stuck for more than 1 second and then control+C took also a while and eventually:

root@x1:~# fg
perf record -aW --sample-mem-info -e cycles:P --filter 'mem_op == load || mem_op == store' sleep 1
^X^C^C^C^Clibbpf: prog 'perf_sample_filter': BPF program load failed: -EAGAIN
libbpf: prog 'perf_sample_filter': -- BEGIN PROG LOAD LOG --
arg#0 reference type('UNKNOWN ') size cannot be determined: -22
0: R1=ctx() R10=fp0
; kctx = bpf_cast_to_kern_ctx(ctx); @ sample_filter.bpf.c:215
0: (85) call bpf_cast_to_kern_ctx#72125       ; R0_w=trusted_ptr_bpf_perf_event_data_kern()
1: (7b) *(u64 *)(r10 -48) = r0        ; R0_w=trusted_ptr_bpf_perf_event_data_kern() R10=fp0 fp-48_w=trusted_ptr_bpf_perf_event_data_kern()
2: (b7) r6 = 0                        ; R6_w=0
; k = 0; @ sample_filter.bpf.c:217
3: (63) *(u32 *)(r10 -4) = r6         ; R6_w=0 R10=fp0 fp-8=0000????
; if (use_idx_hash) { @ sample_filter.bpf.c:219
4: (18) r1 = 0xffffa9e081be6000       ; R1_w=map_value(map=sample_f.rodata,ks=4,vs=4)
6: (61) r1 = *(u32 *)(r1 +0)          ; R1_w=0
7: (15) if r1 == 0x0 goto pc+42       ; R1_w=0
; k = *idx; @ sample_filter.bpf.c:239
50: (bf) r2 = r10                     ; R2_w=fp0 R10=fp0
; entry = bpf_map_lookup_elem(&filters, &k); @ sample_filter.bpf.c:244
51: (07) r2 += -4                     ; R2_w=fp-4
52: (18) r1 = 0xffff8ecc22be8000      ; R1_w=map_ptr(map=filters,ks=4,vs=1536)
54: (85) call bpf_map_lookup_elem#1   ; R0_w=map_value_or_null(id=1,map=filters,ks=4,vs=1536)
55: (bf) r6 = r0                      ; R0_w=map_value_or_null(id=1,map=filters,ks=4,vs=1536) R6_w=map_value_or_null(id=1,map=filters,ks=4,vs=1536)
56: (79) r5 = *(u64 *)(r10 -48)       ; R5_w=trusted_ptr_bpf_perf_event_data_kern() R10=fp0 fp-48=trusted_ptr_bpf_perf_event_data_kern()
; if (entry == NULL) @ sample_filter.bpf.c:245
57: (15) if r6 == 0x0 goto pc-21      ; R6_w=map_value(map=filters,ks=4,vs=1536)
58: (b7) r8 = 0                       ; R8_w=0
59: (b7) r9 = 0                       ; R9_w=0
60: (b7) r1 = 0                       ; R1_w=0
61: (7b) *(u64 *)(r10 -40) = r1       ; R1_w=0 R10=fp0 fp-40_w=0
62: (05) goto pc+9
; for (i = 0; i < MAX_FILTERS; i++) { @ sample_filter.bpf.c:248
72: (bf) r7 = r6                      ; R6=map_value(map=filters,ks=4,vs=1536) R7_w=map_value(map=filters,ks=4,vs=1536)
73: (0f) r7 += r8                     ; R7_w=map_value(map=filters,ks=4,vs=1536) R8=0
; struct perf_sample_data___new *data = (void *)kctx->data; @ sample_filter.bpf.c:77

<SNIP tons of verifier lines>

269: (65) if r2 s> 0x1 goto pc+28     ; R2_w=scalar(smin=smin32=0,smax=umax=smax32=umax32=1,var_off=(0x0; 0x1))
270: (15) if r2 == 0x0 goto pc+45     ; R2_w=1
271: (15) if r2 == 0x1 goto pc+1      ; R2_w=1
; CHECK_RESULT(sample_data, !=, entry[i].value) @ sample_filter.bpf.c:256
273: (bf) r2 = r6                     ; R2_w=map_value(map=filters,ks=4,vs=1536) R6=map_value(map=filters,ks=4,vs=1536)
274: (0f) r2 += r8                    ; R2_w=map_value(map=filters,ks=4,vs=1536,off=984) R8_w=984
275: (79) r2 = *(u64 *)(r2 +16)       ; R2_w=scalar()
276: (5d) if r1 != r2 goto pc+101 378: R0_w=1 R1_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff)) R2_w=scalar() R3_w=ptr_perf_sample_data() R4_w=scalar(smin=umin=umin32=1,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff)) R5=trusted_ptr_bpf_perf_event_data_kern() R6=map_value(map=filters,ks=4,vs=1536) R7_w=map_value(map=filters,ks=4,vs=1536,off=984) R8_w=984 R9=0 R10=fp0 fp-8=mmmm???? fp-40=0 fp-48=trusted_ptr_bpf_perf_event_data_kern()
378: (67) r9 <<= 32                   ; R9_w=0
379: (bf) r1 = r9                     ; R1_w=0 R9_w=0
380: (77) r1 >>= 32                   ; R1_w=0
381: (b7) r9 = 0                      ; R9_w=0
382: (15) if r1 == 0x0 goto pc-313    ; R1_w=0
; for (i = 0; i < MAX_FILTERS; i++) { @ sample_filter.bpf.c:248
70: (07) r8 += 24                     ; R8_w=1008
71: (15) if r8 == 0x600 goto pc-25    ; R8_w=1008
72: (bf) r7 = r6                      ; R6=map_value(map=filters,ks=4,vs=1536) R7_w=map_value(map=filters,ks=4,vs=1536)
73: (0f) r7 += r8                     ; R7_w=map_value(map=filters,ks=4,vs=1536,off=1008) R8_w=1008

; struct perf_sample_data___new *data = (void *)kctx->data; @ sample_filter.bpf.c:77 74: (79) r3 = *(u64 *)(r5 +8)
processed 2344 insns (limit 1000000) max_states_per_insn 4 total_states 57 peak_states 57 mark_read 6 -- END PROG LOAD LOG --
libbpf: prog 'perf_sample_filter': failed to load: -EAGAIN
libbpf: failed to load object 'sample_filter_bpf'
libbpf: failed to load BPF skeleton 'sample_filter_bpf': -EAGAIN
Failed to load perf sample-filter BPF skeleton
failed to set filter "BPF" on event cpu_core/cycles/P with 11 (Resource temporarily unavailable) ^Z
[1]+ Stopped perf record -aW --sample-mem-info -e cycles:P --filter 'mem_op == load || mem_op == store' sleep 1
root@x1:~#

root@x1:~# fg
perf record -aW --sample-mem-info -e cycles:P --filter 'mem_op == load || mem_op == store' sleep 1
^C^C^C^C^C

^C^C^C^C^C

Well, I just had it suspended, I'm not being able to stop it:

root@x1:~# ps 
    PID TTY          TIME CMD
1355566 pts/15   00:00:00 sudo
1355567 pts/15   00:00:00 su
1355570 pts/15   00:00:00 bash
1373907 pts/15   00:00:44 perf
1373908 pts/15   00:00:00 perf-exec
1374019 pts/15   00:00:00 ps
root@x1:~# kill -9 1373907
root@x1:~# kill -9 1373907
-bash: kill: (1373907) - No such process
[1]+  Killed                  perf record -aW --sample-mem-info -e cycles:P --filter 'mem_op == load || mem_op == store' sleep 1
root@x1:~# 

Ok, killed.

It gets stuck in that MAP_FREEZE sys_bpf call:

root@x1:~# perf trace perf record -aW --sample-mem-info -e cycles:P --filter 'mem_op == load || mem_op == store' sleep 1
<SNIP>
 16349.552 ( 0.023 ms): perf/1374043 bpf(uattr: (union bpf_attr){(struct){.map_type = (__u32)2,.key_size = (__u32)4,.value_size = (__u32)4,.max_entries = (__u32)1,.map_flags = (__u32)1152,.map_name = (char[16])['s','a','m','p','l','e','_','f','.','r','o','d','a','t','a',],.btf_fd = (__u32)61,.btf_value_type_id = (__u32)59,},(struct){.map_fd = (__u32)2,.key = (__u64)4294967300,(union){.value = (__u64)1152,.next_key = (__u64)1152,},.flags = (__u64)8101238451258523648,},.batch = (struct){.in_batch = (__u64)17179869186,.out_batch = (__u64)4294967300,.keys = (__u64)1152,.values = (__u64)8101238451258523648,.count = (__u32)1717527916,.map_fd = (__u32)1685025326,.elem_flags = (__u64)6386785,.flags = (__u64)61,},(struct){.prog_type = (__u32)2,.insn_cnt = (__u32)4,.insns = (__u64)4294967300,.license = (__u64)1152,.log_size = (__u32)1886216563,.log_buf = (__u64)7237128669819266412,.kern_version = (__u32)6386785,.prog_name = (char[16])['=',';',],.func_info = (__u64)140735998446044,.func_info_cnt = (__u32)520807888,.line_info = (__u64)140735998445848,.line_info_cnt = (__u32)4874802,.attach_btf_id = (__u32)4,(union){.attach_prog_fd = (__u32)2805057752,.attach_btf_obj_fd = (__u32)2805057752,},.core_relo_cnt = (__u32)32767,.fd_array = (__u64)15131936,.core_relos = (__u64)140735998445952,.core_relo_rec_size = (__u32)15049664,.prog_token_fd = (__s32)-1489909376,},(struct){.pathname = (__u64)17179869186,.bpf_fd = (__u32)4,.file_flags = (__u32)1,.path_fd = (__s32)1152,},(struct){(union){.target_fd = (__u32)2,.target_ifindex = (__u32)2,},.attach_bpf_fd = (__u32)4,.attach_type = (__u32)4,.attach_flags = (__u32)1,.replace_bpf_fd = (__u32)1152,.expected_revision = (__u64)8101238451258523648,},.test = (struct){.prog_fd = (__u32)2,.retval = (__u32)4,.data_size_in = (__u32)4,.data_size_out = (__u32)1,.data_in = (__u64)1152,.data_out = (__u64)8101238451258523648,.repeat = (__u32)1717527916,.duration = (__u32)1685025326,.ctx_size_in = (__u32)6386785,.ctx_in = (__u64)61,.ctx_out = (__u64)59,},(struct){(union){.start_id = (__u32)2,.prog_id = (__u32)2,.map_id = (__u32)2,.btf_id ) = 62
 16349.576 ( 0.002 ms): perf/1374043 dup3(oldfd: 62, newfd: 60<anon_inode:bpf-map>, flags: 524288)         = 60
 16349.582 ( 0.001 ms): perf/1374043 close(fd: 62)                                                         = 0
 16349.586 ( 0.006 ms): perf/1374043 bpf(cmd: MAP_UPDATE_ELEM, uattr: (union bpf_attr){(struct){.map_type = (__u32)60,.value_size = (__u32)2805058140,.max_entries = (__u32)32767,.map_flags = (__u32)1994993664,.inner_map_fd = (__u32)32622,.map_ifindex = (__u32)2,.btf_fd = (__u32)2805058000,.btf_key_type_id = (__u32)32767,.btf_value_type_id = (__u32)16,.btf_vmlinux_value_type_id = (__u32)48,.map_extra = (__u64)140735998446240,.value_type_btf_obj_fd = (__s32)-1489909280,.map_token_fd = (__s32)32767,},(struct){.map_fd = (__u32)60,.key = (__u64)140735998446172,(union){.value = (__u64)140112418123776,.next_key = (__u64)140112418123776,},},.batch = (struct){.in_batch = (__u64)60,.out_batch = (__u64)140735998446172,.keys = (__u64)140112418123776,.count = (__u32)7379400,.elem_flags = (__u64)9110741168,.flags = (__u64)140735998446032,},(struct){.prog_type = (__u32)60,.insns = (__u64)140735998446172,.license = (__u64)140112418123776,.log_buf = (__u64)7379400,.kern_version = (__u32)520806576,.prog_flags = (__u32)2,.prog_name = (char[16])[208,201,'1',167,255,127,16,'0',],.prog_ifindex = (__u32)2805058208,.expected_attach_type = (__u32)32767,.prog_btf_fd = (__u32)2805058016,.func_info_rec_size = (__u32)32767,.func_info = (__u64)410826951296,.func_info_cnt = (__u32)4874930,.line_info = (__u64)40,.line_info_cnt = (__u32)1,.attach_btf_id = (__u32)4,(union){.attach_prog_fd = (__u32)520846144,.attach_btf_obj_fd = (__u32)520846144,},.fd_array = (__u64)60,.core_relos = (__u64)140735998445776,.core_relo_rec_size = (__u32)2805057752,.log_true_size = (__u32)524288,},(struct){.pathname = (__u64)60,.bpf_fd = (__u32)2805058140,.file_flags = (__u32)32767,.path_fd = (__s32)1994993664,},(struct){(union){.target_fd = (__u32)60,.target_ifindex = (__u32)60,},.attach_type = (__u32)2805058140,.attach_flags = (__u32)32767,.replace_bpf_fd = (__u32)1994993664,(union){.relative_fd = (__u32)32622,.relative_id = (__u32)32622,},},.test = (struct){.prog_fd = (__u32)60,.data_size_in = (__u32)2805058140,.data_size_out = (__u32)32767,.data_in = (__u64)140112418123776,.repeat = (__u32)7379400,.) = 0

16349.594 ( 0.002 ms): perf/1374043 bpf(cmd: MAP_FREEZE, uattr: (union bpf_attr){(struct){.map_type = (__u32)60,.value_size = (__u32)2805058140,.max_entries = (__u32)32767,.map_flags = (__u32)1994993664,.inner_map_fd = (__u32)32622,.map_ifindex = (__u32)2,.btf_fd = (__u32)2805058000,.btf_key_type_id = (__u32)32767,.btf_value_type_id = (__u32)16,.btf_vmlinux_value_type_id = (__u32)48,.map_extra = (__u64)140735998446240,.value_type_btf_obj_fd = (__s32)-1489909280,.map_token_fd = (__s32)32767,},(struct){.map_fd = (__u32)60,.key = (__u64)140735998446172,(union){.value = (__u64)140112418123776,.next_key = (__u64)140112418123776,},},.batch = (struct){.in_batch = (__u64)60,.out_batch = (__u64)140735998446172,.keys = (__u64)140112418123776,.count = (__u32)7379400,.elem_flags = (__u64)9110741168,.flags = (__u64)140735998446032,},(struct){.prog_type = (__u32)60,.insns = (__u64)140735998446172,.license = (__u64)140112418123776,.log_buf = (__u64)7379400,.kern_version = (__u32)520806576,.prog_flags = (__u32)2,.prog_name = (char[16])[208,201,'1',167,255,127,16,'0',],.prog_ifindex = (__u32)2805058208,.expected_attach_type = (__u32)32767,.prog_btf_fd = (__u32)2805058016,.func_info_rec_size = (__u32)32767,.func_info = (__u64)410826951296,.func_info_cnt = (__u32)4874930,.line_info = (__u64)40,.line_info_cnt = (__u32)1,.attach_btf_id = (__u32)4,(union){.attach_prog_fd = (__u32)520846144,.attach_btf_obj_fd = (__u32)520846144,},.fd_array = (__u64)60,.core_relos = (__u64)140735998445776,.core_relo_rec_size = (__u32)2805057752,.log_true_size = (__u32)524288,},(struct){.pathname = (__u64)60,.bpf_fd = (__u32)2805058140,.file_flags = (__u32)32767,.path_fd = (__s32)1994993664,},(struct){(union){.target_fd = (__u32)60,.target_ifindex = (__u32)60,},.attach_type = (__u32)2805058140,.attach_flags = (__u32)32767,.replace_bpf_fd = (__u32)1994993664,(union){.relative_fd = (__u32)32622,.relative_id = (__u32)32622,},},.test = (struct){.prog_fd = (__u32)60,.data_size_in = (__u32)2805058140,.data_size_out = (__u32)32767,.data_in = (__u64)140112418123776,.repeat = (__u32)7379400,.ctx_s) = 0 o
16349.601 ( 0.050 ms): perf/1374043 mmap(addr: 0x7f6e76e93000, len: 4096, prot: READ, flags: SHARED|FIXED, fd: 60) = 0x7f6e76e93000

root@x1:~# uname -a
Linux x1 6.13.9-100.fc40.x86_64 #1 SMP PREEMPT_DYNAMIC Sat Mar 29 01:27:18 UTC 2025 x86_64 GNU/Linux
root@x1:~#
root@x1:~# grep -m1 "model name" /proc/cpuinfo 
model name	: 13th Gen Intel(R) Core(TM) i7-1365U
root@x1:~# 

Now trying on another machine:

root@number:~# uname -a
Linux number 6.15.0-rc4+ #2 SMP PREEMPT_DYNAMIC Tue Apr 29 15:56:43 -03 2025 x86_64 GNU/Linux
root@number:~# grep -m1 "model name" /proc/cpuinfo
model name	: AMD Ryzen 9 9950X3D 16-Core Processor
root@number:~# perf record -aW --sample-mem-info -e cycles:P --filter 'mem_op == load || mem_op == store' sleep 1
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 1.870 MB perf.data (1199 samples) ]
root@number:~#
root@number:~# perf mem report -F overhead,dtlb,dso --stdio | head -20
# To display the perf.data header info, please use --header/--header-only options.
#
#
# Total Lost Samples: 0
#
# Samples: 1K of event 'cycles:P'
# Total weight : 7879
# Sort order   : local_weight,mem,sym,dso,symbol_daddr,dso_daddr,snoop,tlb,locked,blocked,local_ins_lat,local_p_stage_cyc
#
#           ---------- D-TLB -----------                                   
# Overhead   L1-Hit L2-Hit   Miss  Other  Shared Object                    
# ........  ............................  .................................
#
    48.51%    44.8%  18.7%  28.3%   8.2%  [kernel.kallsyms]                
    11.97%     2.1%   1.4%  96.0%   0.5%  libc.so.6                        
     8.58%    85.7%  14.1%   0.0%   0.3%  libxul.so                        
     6.76%   100.0%   0.0%   0.0%   0.0%  libfreeblpriv3.so                
     6.08%   100.0%   0.0%   0.0%   0.0%  libsystemd-shared-257.5-2.fc42.so
     4.59%   100.0%   0.0%   0.0%   0.0%  firefox                          
     4.33%   100.0%   0.0%   0.0%   0.0%  libgallium-25.0.4.so             
root@number:~#

Looks better :-)

Having the record as an alias seems interesting, ditto for the report.

- Arnaldo

author maintainer
#18 lore raw

Re: [PATCH 11/11] perf mem: Add 'dtlb' output field

Quoted 57 lines · Arnaldo Carvalho de Melo
Quoted 56 lines · Namhyung Kim
Quoted 43 lines · Arnaldo Carvalho de Melo
Quoted 17 lines · Namhyung Kim

This is a breakdown of perf_mem_data_src.mem_dtlb values. It assumes PMU drivers would set PERF_MEM_TLB_HIT bit with an appropriate level. And having PERF_MEM_TLB_MISS means that it failed to find one in any levels of TLB. For now, it doesn't use PERF_MEM_TLB_{WK,OS} bits.

Also it seems Intel machines don't distinguish L1 or L2 precisely. So I added ANY_HIT (printed as "L?-Hit") to handle the case.

  $ perf mem report -F overhead,dtlb,dso --stdio
  ...
  #           --- D-TLB ----
  # Overhead   L?-Hit   Miss  Shared Object
  # ........  ..............  .................
  #
      67.03%    99.5%   0.5%  [unknown]
      31.23%    99.2%   0.8%  [kernel.kallsyms]
       1.08%    97.8%   2.2%  [i915]
Quoted 8 lines
       0.36%   100.0%   0.0%  [JIT] tid 6853
       0.12%   100.0%   0.0%  [drm]
       0.05%   100.0%   0.0%  [drm_kms_helper]
       0.05%   100.0%   0.0%  [ext4]
       0.02%   100.0%   0.0%  [aesni_intel]
       0.02%   100.0%   0.0%  [crc32c_intel]
       0.02%   100.0%   0.0%  [dm_crypt]
       ...
root@number:~# perf report --header | grep cpudesc
# cpudesc : AMD Ryzen 9 9950X3D 16-Core Processor
root@number:~# perf mem report -F overhead,dtlb,dso --stdio | head -20
# To display the perf.data header info, please use --header/--header-only options.
#
#
# Total Lost Samples: 0
#
# Samples: 2K of event 'cycles:P'
# Total weight : 2637
# Sort order   : local_weight,mem,sym,dso,symbol_daddr,dso_daddr,snoop,tlb,locked,blocked,local_ins_lat,local_p_stage_cyc
#
#           ---------- D-TLB -----------                                   
# Overhead   L1-Hit L2-Hit   Miss  Other  Shared Object                    
# ........  ............................  .................................
#
Quoted 8 lines
    77.47%    18.4%   0.1%   0.6%  80.9%  [kernel.kallsyms]                
     5.61%    36.5%   0.7%   1.4%  61.5%  libxul.so                        
     2.77%    39.7%   0.0%  12.3%  47.9%  libc.so.6                        
     2.01%    34.0%   1.9%   1.9%  62.3%  libglib-2.0.so.0.8400.1          
     1.93%    31.4%   2.0%   2.0%  64.7%  [amdgpu]                         
     1.63%    48.8%   0.0%   0.0%  51.2%  [JIT] tid 60168                  
     1.14%     3.3%   0.0%   0.0%  96.7%  [vdso]                           
root@number:~#

I guess it's because those samples don't have mem_info as they are not memory instructions.

Quoted 6 lines
Quoted 3 lines

Can you please re-run the perf record with filters like below?

  $ perf record -aW --sample-mem-info -e cycles:P --filter 'mem_op == load || mem_op == store' sleep 1

I tried, it got stuck for more than 1 second and then control+C took also a while and eventually:

I also saw some delay when using BPF. I suspect it's BPF loader taking long.

Quoted 93 lines
root@x1:~# fg
perf record -aW --sample-mem-info -e cycles:P --filter 'mem_op == load || mem_op == store' sleep 1
^X^C^C^C^Clibbpf: prog 'perf_sample_filter': BPF program load failed: -EAGAIN
libbpf: prog 'perf_sample_filter': -- BEGIN PROG LOAD LOG --
arg#0 reference type('UNKNOWN ') size cannot be determined: -22
0: R1=ctx() R10=fp0
; kctx = bpf_cast_to_kern_ctx(ctx); @ sample_filter.bpf.c:215
0: (85) call bpf_cast_to_kern_ctx#72125       ; R0_w=trusted_ptr_bpf_perf_event_data_kern()
1: (7b) *(u64 *)(r10 -48) = r0        ; R0_w=trusted_ptr_bpf_perf_event_data_kern() R10=fp0 fp-48_w=trusted_ptr_bpf_perf_event_data_kern()
2: (b7) r6 = 0                        ; R6_w=0
; k = 0; @ sample_filter.bpf.c:217
3: (63) *(u32 *)(r10 -4) = r6         ; R6_w=0 R10=fp0 fp-8=0000????
; if (use_idx_hash) { @ sample_filter.bpf.c:219
4: (18) r1 = 0xffffa9e081be6000       ; R1_w=map_value(map=sample_f.rodata,ks=4,vs=4)
6: (61) r1 = *(u32 *)(r1 +0)          ; R1_w=0
7: (15) if r1 == 0x0 goto pc+42       ; R1_w=0
; k = *idx; @ sample_filter.bpf.c:239
50: (bf) r2 = r10                     ; R2_w=fp0 R10=fp0
; entry = bpf_map_lookup_elem(&filters, &k); @ sample_filter.bpf.c:244
51: (07) r2 += -4                     ; R2_w=fp-4
52: (18) r1 = 0xffff8ecc22be8000      ; R1_w=map_ptr(map=filters,ks=4,vs=1536)
54: (85) call bpf_map_lookup_elem#1   ; R0_w=map_value_or_null(id=1,map=filters,ks=4,vs=1536)
55: (bf) r6 = r0                      ; R0_w=map_value_or_null(id=1,map=filters,ks=4,vs=1536) R6_w=map_value_or_null(id=1,map=filters,ks=4,vs=1536)
56: (79) r5 = *(u64 *)(r10 -48)       ; R5_w=trusted_ptr_bpf_perf_event_data_kern() R10=fp0 fp-48=trusted_ptr_bpf_perf_event_data_kern()
; if (entry == NULL) @ sample_filter.bpf.c:245
57: (15) if r6 == 0x0 goto pc-21      ; R6_w=map_value(map=filters,ks=4,vs=1536)
58: (b7) r8 = 0                       ; R8_w=0
59: (b7) r9 = 0                       ; R9_w=0
60: (b7) r1 = 0                       ; R1_w=0
61: (7b) *(u64 *)(r10 -40) = r1       ; R1_w=0 R10=fp0 fp-40_w=0
62: (05) goto pc+9
; for (i = 0; i < MAX_FILTERS; i++) { @ sample_filter.bpf.c:248
72: (bf) r7 = r6                      ; R6=map_value(map=filters,ks=4,vs=1536) R7_w=map_value(map=filters,ks=4,vs=1536)
73: (0f) r7 += r8                     ; R7_w=map_value(map=filters,ks=4,vs=1536) R8=0
; struct perf_sample_data___new *data = (void *)kctx->data; @ sample_filter.bpf.c:77

<SNIP tons of verifier lines>

269: (65) if r2 s> 0x1 goto pc+28     ; R2_w=scalar(smin=smin32=0,smax=umax=smax32=umax32=1,var_off=(0x0; 0x1))
270: (15) if r2 == 0x0 goto pc+45     ; R2_w=1
271: (15) if r2 == 0x1 goto pc+1      ; R2_w=1
; CHECK_RESULT(sample_data, !=, entry[i].value) @ sample_filter.bpf.c:256
273: (bf) r2 = r6                     ; R2_w=map_value(map=filters,ks=4,vs=1536) R6=map_value(map=filters,ks=4,vs=1536)
274: (0f) r2 += r8                    ; R2_w=map_value(map=filters,ks=4,vs=1536,off=984) R8_w=984
275: (79) r2 = *(u64 *)(r2 +16)       ; R2_w=scalar()
276: (5d) if r1 != r2 goto pc+101 378: R0_w=1 R1_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff)) R2_w=scalar() R3_w=ptr_perf_sample_data() R4_w=scalar(smin=umin=umin32=1,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff)) R5=trusted_ptr_bpf_perf_event_data_kern() R6=map_value(map=filters,ks=4,vs=1536) R7_w=map_value(map=filters,ks=4,vs=1536,off=984) R8_w=984 R9=0 R10=fp0 fp-8=mmmm???? fp-40=0 fp-48=trusted_ptr_bpf_perf_event_data_kern()
378: (67) r9 <<= 32                   ; R9_w=0
379: (bf) r1 = r9                     ; R1_w=0 R9_w=0
380: (77) r1 >>= 32                   ; R1_w=0
381: (b7) r9 = 0                      ; R9_w=0
382: (15) if r1 == 0x0 goto pc-313    ; R1_w=0
; for (i = 0; i < MAX_FILTERS; i++) { @ sample_filter.bpf.c:248
70: (07) r8 += 24                     ; R8_w=1008
71: (15) if r8 == 0x600 goto pc-25    ; R8_w=1008
72: (bf) r7 = r6                      ; R6=map_value(map=filters,ks=4,vs=1536) R7_w=map_value(map=filters,ks=4,vs=1536)
73: (0f) r7 += r8                     ; R7_w=map_value(map=filters,ks=4,vs=1536,off=1008) R8_w=1008

; struct perf_sample_data___new *data = (void *)kctx->data; @ sample_filter.bpf.c:77 74: (79) r3 = *(u64 *)(r5 +8)
processed 2344 insns (limit 1000000) max_states_per_insn 4 total_states 57 peak_states 57 mark_read 6 -- END PROG LOAD LOG --
libbpf: prog 'perf_sample_filter': failed to load: -EAGAIN
libbpf: failed to load object 'sample_filter_bpf'
libbpf: failed to load BPF skeleton 'sample_filter_bpf': -EAGAIN
Failed to load perf sample-filter BPF skeleton
failed to set filter "BPF" on event cpu_core/cycles/P with 11 (Resource temporarily unavailable) ^Z
[1]+ Stopped perf record -aW --sample-mem-info -e cycles:P --filter 'mem_op == load || mem_op == store' sleep 1
root@x1:~#

root@x1:~# fg
perf record -aW --sample-mem-info -e cycles:P --filter 'mem_op == load || mem_op == store' sleep 1
^C^C^C^C^C

^C^C^C^C^C

Well, I just had it suspended, I'm not being able to stop it:

root@x1:~# ps 
    PID TTY          TIME CMD
1355566 pts/15   00:00:00 sudo
1355567 pts/15   00:00:00 su
1355570 pts/15   00:00:00 bash
1373907 pts/15   00:00:44 perf
1373908 pts/15   00:00:00 perf-exec
1374019 pts/15   00:00:00 ps
root@x1:~# kill -9 1373907
root@x1:~# kill -9 1373907
-bash: kill: (1373907) - No such process
[1]+  Killed                  perf record -aW --sample-mem-info -e cycles:P --filter 'mem_op == load || mem_op == store' sleep 1
root@x1:~# 
Quoted 3 lines

Ok, killed.

It gets stuck in that MAP_FREEZE sys_bpf call:

Interesting.

Quoted 4 lines
root@x1:~# perf trace perf record -aW --sample-mem-info -e cycles:P --filter 'mem_op == load || mem_op == store' sleep 1
<SNIP>
 16349.552 ( 0.023 ms): perf/1374043 bpf(uattr: (union bpf_attr){(struct){.map_type = (__u32)2,.key_size = (__u32)4,.value_size = (__u32)4,.max_entries = (__u32)1,.map_flags = (__u32)1152,.map_name = (char[16])['s','a','m','p','l','e','_','f','.','r','o','d','a','t','a',],.btf_fd = (__u32)61,.btf_value_type_id = (__u32)59,},(struct){.map_fd = (__u32)2,.key = (__u64)4294967300,(union){.value = (__u64)1152,.next_key = (__u64)1152,},.flags = (__u64)8101238451258523648,},.batch = (struct){.in_batch = (__u64)17179869186,.out_batch = (__u64)4294967300,.keys = (__u64)1152,.values = (__u64)8101238451258523648,.count = (__u32)1717527916,.map_fd = (__u32)1685025326,.elem_flags = (__u64)6386785,.flags = (__u64)61,},(struct){.prog_type = (__u32)2,.insn_cnt = (__u32)4,.insns = (__u64)4294967300,.license = (__u64)1152,.log_size = (__u32)1886216563,.log_buf = (__u64)7237128669819266412,.kern_version = (__u32)6386785,.prog_name = (char[16])['=',';',],.func_info = (__u64)140735998446044,.func_info_cnt = (__u32)520807888,.line_info = (__u64)140735998445848,.line_info_cnt = (__u32)4874802,.attach_btf_id = (__u32)4,(union){.attach_prog_fd = (__u32)2805057752,.attach_btf_obj_fd = (__u32)2805057752,},.core_relo_cnt = (__u32)32767,.fd_array = (__u64)15131936,.core_relos = (__u64)140735998445952,.core_relo_rec_size = (__u32)15049664,.prog_token_fd = (__s32)-1489909376,},(struct){.pathname = (__u64)17179869186,.bpf_fd = (__u32)4,.file_flags = (__u32)1,.path_fd = (__s32)1152,},(struct){(union){.target_fd = (__u32)2,.target_ifindex = (__u32)2,},.attach_bpf_fd = (__u32)4,.attach_type = (__u32)4,.attach_flags = (__u32)1,.replace_bpf_fd = (__u32)1152,.expected_revision = (__u64)8101238451258523648,},.test = (struct){.prog_fd = (__u32)2,.retval = (__u32)4,.data_size_in = (__u32)4,.data_size_out = (__u32)1,.data_in = (__u64)1152,.data_out = (__u64)8101238451258523648,.repeat = (__u32)1717527916,.duration = (__u32)1685025326,.ctx_size_in = (__u32)6386785,.ctx_in = (__u64)61,.ctx_out = (__u64)59,},(struct){(union){.start_id = (__u32)2,.prog_id = (__u32)2,.map_id = (__u32)2,.btf_id ) = 62

This line misses 'cmd:' part. Probably it was 0 (MAP_CREATE). It seems we need to print the value even if it's 0.

Quoted 8 lines

16349.576 ( 0.002 ms): perf/1374043 dup3(oldfd: 62, newfd: 60<anon_inode:bpf-map>, flags: 524288) = 60
16349.582 ( 0.001 ms): perf/1374043 close(fd: 62) = 0
16349.586 ( 0.006 ms): perf/1374043 bpf(cmd: MAP_UPDATE_ELEM, uattr: (union bpf_attr){(struct){.map_type = (__u32)60,.value_size = (__u32)2805058140,.max_entries = (__u32)32767,.map_flags = (__u32)1994993664,.inner_map_fd = (__u32)32622,.map_ifindex = (__u32)2,.btf_fd = (__u32)2805058000,.btf_key_type_id = (__u32)32767,.btf_value_type_id = (__u32)16,.btf_vmlinux_value_type_id = (__u32)48,.map_extra = (__u64)140735998446240,.value_type_btf_obj_fd = (__s32)-1489909280,.map_token_fd = (__s32)32767,},(struct){.map_fd = (__u32)60,.key = (__u64)140735998446172,(union){.value = (__u64)140112418123776,.next_key = (__u64)140112418123776,},},.batch = (struct){.in_batch = (__u64)60,.out_batch = (__u64)140735998446172,.keys = (__u64)140112418123776,.count = (__u32)7379400,.elem_flags = (__u64)9110741168,.flags = (__u64)140735998446032,},(struct){.prog_type = (__u32)60,.insns = (__u64)140735998446172,.license = (__u64)140112418123776,.log_buf = (__u64)7379400,.kern_version = (__u32)520806576,.prog_flags = (__u32)2,.prog_name = (char[16])[208,201,'1',167,255,127,16,'0',],.prog_ifindex = (__u32)2805058208,.expected_attach_type = (__u32)32767,.prog_btf_fd = (__u32)2805058016,.func_info_rec_size = (__u32)32767,.func_info = (__u64)410826951296,.func_info_cnt = (__u32)4874930,.line_info = (__u64)40,.line_info_cnt = (__u32)1,.attach_btf_id = (__u32)4,(union){.attach_prog_fd = (__u32)520846144,.attach_btf_obj_fd = (__u32)520846144,},.fd_array = (__u64)60,.core_relos = (__u64)140735998445776,.core_relo_rec_size = (__u32)2805057752,.log_true_size = (__u32)524288,},(struct){.pathname = (__u64)60,.bpf_fd = (__u32)2805058140,.file_flags = (__u32)32767,.path_fd = (__s32)1994993664,},(struct){(union){.target_fd = (__u32)60,.target_ifindex = (__u32)60,},.attach_type = (__u32)2805058140,.attach_flags = (__u32)32767,.replace_bpf_fd = (__u32)1994993664,(union){.relative_fd = (__u32)32622,.relative_id = (__u32)32622,},},.test = (struct){.prog_fd = (__u32)60,.data_size_in = (__u32)2805058140,.data_size_out = (__u32)32767,.data_in = (__u64)140112418123776,.repeat = (__u32)7379400,.) = 0

16349.594 ( 0.002 ms): perf/1374043 bpf(cmd: MAP_FREEZE, uattr: (union bpf_attr){(struct){.map_type = (__u32)60,.value_size = (__u32)2805058140,.max_entries = (__u32)32767,.map_flags = (__u32)1994993664,.inner_map_fd = (__u32)32622,.map_ifindex = (__u32)2,.btf_fd = (__u32)2805058000,.btf_key_type_id = (__u32)32767,.btf_value_type_id = (__u32)16,.btf_vmlinux_value_type_id = (__u32)48,.map_extra = (__u64)140735998446240,.value_type_btf_obj_fd = (__s32)-1489909280,.map_token_fd = (__s32)32767,},(struct){.map_fd = (__u32)60,.key = (__u64)140735998446172,(union){.value = (__u64)140112418123776,.next_key = (__u64)140112418123776,},},.batch = (struct){.in_batch = (__u64)60,.out_batch = (__u64)140735998446172,.keys = (__u64)140112418123776,.count = (__u32)7379400,.elem_flags = (__u64)9110741168,.flags = (__u64)140735998446032,},(struct){.prog_type = (__u32)60,.insns = (__u64)140735998446172,.license = (__u64)140112418123776,.log_buf = (__u64)7379400,.kern_version = (__u32)520806576,.prog_flags = (__u32)2,.prog_name = (char[16])[208,201,'1',167,255,127,16,'0',],.prog_ifindex = (__u32)2805058208,.expected_attach_type = (__u32)32767,.prog_btf_fd = (__u32)2805058016,.func_info_rec_size = (__u32)32767,.func_info = (__u64)410826951296,.func_info_cnt = (__u32)4874930,.line_info = (__u64)40,.line_info_cnt = (__u32)1,.attach_btf_id = (__u32)4,(union){.attach_prog_fd = (__u32)520846144,.attach_btf_obj_fd = (__u32)520846144,},.fd_array = (__u64)60,.core_relos = (__u64)140735998445776,.core_relo_rec_size = (__u32)2805057752,.log_true_size = (__u32)524288,},(struct){.pathname = (__u64)60,.bpf_fd = (__u32)2805058140,.file_flags = (__u32)32767,.path_fd = (__s32)1994993664,},(struct){(union){.target_fd = (__u32)60,.target_ifindex = (__u32)60,},.attach_type = (__u32)2805058140,.attach_flags = (__u32)32767,.replace_bpf_fd = (__u32)1994993664,(union){.relative_fd = (__u32)32622,.relative_id = (__u32)32622,},},.test = (struct){.prog_fd = (__u32)60,.data_size_in = (__u32)2805058140,.data_size_out = (__u32)32767,.data_in = (__u64)140112418123776,.repeat = (__u32)7379400,.ctx_s) = 0 o
16349.601 ( 0.050 ms): perf/1374043 mmap(addr: 0x7f6e76e93000, len: 4096, prot: READ, flags: SHARED|FIXED, fd: 60) = 0x7f6e76e93000

Quoted 7 lines
root@x1:~# uname -a
Linux x1 6.13.9-100.fc40.x86_64 #1 SMP PREEMPT_DYNAMIC Sat Mar 29 01:27:18 UTC 2025 x86_64 GNU/Linux
root@x1:~#
root@x1:~# grep -m1 "model name" /proc/cpuinfo 
model name	: 13th Gen Intel(R) Core(TM) i7-1365U
root@x1:~# 

Well.. you don't need that on Intel and just use 'perf mem record'. :)

Quoted 35 lines

Now trying on another machine:

root@number:~# uname -a
Linux number 6.15.0-rc4+ #2 SMP PREEMPT_DYNAMIC Tue Apr 29 15:56:43 -03 2025 x86_64 GNU/Linux
root@number:~# grep -m1 "model name" /proc/cpuinfo
model name	: AMD Ryzen 9 9950X3D 16-Core Processor
root@number:~# perf record -aW --sample-mem-info -e cycles:P --filter 'mem_op == load || mem_op == store' sleep 1
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 1.870 MB perf.data (1199 samples) ]
root@number:~#
root@number:~# perf mem report -F overhead,dtlb,dso --stdio | head -20
# To display the perf.data header info, please use --header/--header-only options.
#
#
# Total Lost Samples: 0
#
# Samples: 1K of event 'cycles:P'
# Total weight : 7879
# Sort order   : local_weight,mem,sym,dso,symbol_daddr,dso_daddr,snoop,tlb,locked,blocked,local_ins_lat,local_p_stage_cyc
#
#           ---------- D-TLB -----------                                   
# Overhead   L1-Hit L2-Hit   Miss  Other  Shared Object                    
# ........  ............................  .................................
#
    48.51%    44.8%  18.7%  28.3%   8.2%  [kernel.kallsyms]                
    11.97%     2.1%   1.4%  96.0%   0.5%  libc.so.6                        
     8.58%    85.7%  14.1%   0.0%   0.3%  libxul.so                        
     6.76%   100.0%   0.0%   0.0%   0.0%  libfreeblpriv3.so                
     6.08%   100.0%   0.0%   0.0%   0.0%  libsystemd-shared-257.5-2.fc42.so
     4.59%   100.0%   0.0%   0.0%   0.0%  firefox                          
     4.33%   100.0%   0.0%   0.0%   0.0%  libgallium-25.0.4.so             
root@number:~#
Quoted 3 lines

Looks better :-)

Having the record as an alias seems interesting, ditto for the report.

Yep, you can update 'perf mem record'.

Thanks,
Namhyung

#19 lore raw
In reply to Namhyung Kim
Quoted 10 lines · Namhyung Kim

Hello,

The perf mem uses PERF_SAMPLE_DATA_SRC which has a lot of information for memory access. It has various sort keys to group related samples together but it's still cumbersome to see the result. While perf c2c command provides a way to investigate the data in a specific way, I'd like to add more generic ways using new output fields.

For example, the following is the 'cache' output field which breaks down the sample weights into different level of caches.

Super cool!

Quoted 37 lines
  $ perf mem record -a sleep 1
  $ perf mem report -F cache,dso,sym --stdio
  ...
  #
  # -------------- Cache --------------
  #      L1     L2     L3 L1-buf  Other  Shared Object                                  Symbol
  # ...................................  .....................................  .........................................
  #
       0.0%   0.0%   0.0%   0.0% 100.0%  [kernel.kallsyms]                      [k] ioread8
     100.0%   0.0%   0.0%   0.0%   0.0%  [kernel.kallsyms]                      [k] _raw_spin_lock_irq
       0.0%   0.0%   0.0%   0.0% 100.0%  [xhci_hcd]                             [k] xhci_update_erst_dequeue
       0.0%   0.0%   0.0%  95.8%   4.2%  [kernel.kallsyms]                      [k] smaps_account
       0.6%   1.8%  22.7%  45.5%  29.5%  [kernel.kallsyms]                      [k] sched_balance_update_blocked_averages
      29.4%   0.0%   1.6%  58.8%  10.2%  [kernel.kallsyms]                      [k] __update_load_avg_cfs_rq
       0.0%   8.5%   4.3%   0.0%  87.2%  [kernel.kallsyms]                      [k] copy_mc_enhanced_fast_string
      63.9%   0.0%   8.0%  23.8%   4.3%  [kernel.kallsyms]                      [k] psi_group_change
       3.9%   0.0%   9.3%  35.7%  51.1%  [kernel.kallsyms]                      [k] timerqueue_add
      35.9%  10.9%   0.0%  39.0%  14.2%  [kernel.kallsyms]                      [k] memcpy
      94.1%   0.0%   0.0%   5.9%   0.0%  [kernel.kallsyms]                      [k] unmap_page_range
      25.7%   0.0%   4.9%  51.0%  18.4%  [kernel.kallsyms]                      [k] __update_load_avg_se
       0.0%  24.9%  19.4%   9.6%  46.1%  [kernel.kallsyms]                      [k] _copy_to_iter
      12.9%   0.0%   0.0%  87.1%   0.0%  [kernel.kallsyms]                      [k] next_uptodate_folio
      36.8%   0.0%   9.5%  16.6%  37.1%  [kernel.kallsyms]                      [k] update_curr
     100.0%   0.0%   0.0%   0.0%   0.0%  bpf_prog_b9611ccbbb3d1833_dfs_iter     [k] bpf_prog_b9611ccbbb3d1833_dfs_iter
      45.4%   1.8%  20.4%  23.6%   8.8%  [kernel.kallsyms]                      [k] audit_filter_rules.isra.0
      92.8%   0.0%   0.0%   7.2%   0.0%  [kernel.kallsyms]                      [k] filemap_map_pages
      10.6%   0.0%   0.0%  89.4%   0.0%  [kernel.kallsyms]                      [k] smaps_page_accumulate
      38.3%   0.0%  29.6%  27.1%   5.0%  [kernel.kallsyms]                      [k] __schedule

Please see the description of each commit for other fields.

New mem_stat field was added to the hist_entry to save this information. It's a generic data structure (array) to handle different type of information like cache-level, memory location, snoop-result, etc.

Quoted 4 lines

The first patch is a fix for the hierarchy mode and it was sent separately. I just add it here not to break the hierarchy mode. The second patch is to enable SAMPLE_DATA_SRC without SAMPLE_ADDR and perf_event_attr.mmap_data which generate a lot more data.

I merged it and added a test for the hierachy mode as mentioned in my reply to that patch.

Quoted 4 lines

The name of some new fields are the same as the corresponding sort keys (mem, op, snoop) so I had to change the order whether it's applied as an output field or a sort key. Maybe it's better to name them differently but I couldn't come up with better ideas.

Looks ok at first sight.

Quoted 3 lines

That means, you need to use -F/--fields option to specify those fields and the sort keys you want. Maybe we can change the default output and sort keys for perf mem report with this.

Maybe we can come up with aliases to help using these new features without having to create a long command line, maybe:

perf cache

Or some other more suitable name.

That would just be translated into the long command line for 'perf report', kinda like 'perf kvm', but maybe we can do it like with 'perf archive', i.e. just a shell wrapper?

Quoted 1 line

The code is available at 'perf/mem-field-v1' branch in

I'll test it, and I'm CCing Joe Mario, who I think will be very much interesting in trying this!

- Arnaldo

Quoted 34 lines

git://git.kernel.org/pub/scm/linux/kernel/git/namhyung/linux-perf.git

Thanks,
Namhyung

Namhyung Kim (11):
perf hist: Remove output field from sort-list properly perf record: Add --sample-mem-info option perf hist: Support multi-line header
perf hist: Add struct he_mem_stat
perf hist: Basic support for mem_stat accounting perf hist: Implement output fields for mem stats perf mem: Add 'op' output field
perf hist: Hide unused mem stat columns
perf mem: Add 'cache' and 'memory' output fields perf mem: Add 'snoop' output field
perf mem: Add 'dtlb' output field

 tools/perf/Documentation/perf-record.txt |   7 +-
 tools/perf/builtin-record.c              |   6 +
 tools/perf/ui/browsers/hists.c           |  50 ++++-
 tools/perf/ui/hist.c                     | 272 ++++++++++++++++++++++-
 tools/perf/ui/stdio/hist.c               |  57 +++--
 tools/perf/util/evsel.c                  |   2 +-
 tools/perf/util/hist.c                   |  78 +++++++
 tools/perf/util/hist.h                   |  22 ++
 tools/perf/util/mem-events.c             | 183 ++++++++++++++-
 tools/perf/util/mem-events.h             |  57 +++++
 tools/perf/util/record.h                 |   1 +
 tools/perf/util/sort.c                   |  42 +++-
 12 files changed, 718 insertions(+), 59 deletions(-)
Signature · 1 line
2.49.0.906.g1f30a19c02-goog
#20 lore raw
In reply to Namhyung Kim

Hi Namhyung,

I feel the overall idea is good. Running few simple perf-mem commands on AMD works fine too. Few general feedback below.

Quoted 4 lines

The name of some new fields are the same as the corresponding sort keys (mem, op, snoop) so I had to change the order whether it's applied as an output field or a sort key. Maybe it's better to name them differently but I couldn't come up with better ideas.

1) These semantic changes of the field name seems counter intuitive (to me). Example:

-F mem:

     Without patch:

     $ perf mem report -F overhead,sample,mem --stdio
     # Overhead       Samples  Memory access
         39.29%             1  L3 hit
         37.50%            21  N/A
         23.21%            13  L1 hit

     With patch:

     $ perf mem report -F overhead,sample,mem --stdio
     #                          Memory
     # Overhead       Samples    Other
        100.00%            35   100.0%

-F 'snoop':

     Without patch:

     $ perf mem report -F overhead,sample,snoop --stdio
     # Overhead       Samples  Snoop
         60.71%            34  N/A
         39.29%             1  HitM
   
     With patchset:

     $ perf mem report -F overhead,sample,snoop --stdio
     #                         --- Snoop ----
     # Overhead       Samples     HitM  Other
        100.00%            35    39.3%  60.7%

2) It was not intuitive (to me:)) that perf-mem overhead is calculated using sample->weight by overwriting sample->period. I also don't see it documented anywhere (or did I miss it?)

perf report:

     $ perf report -F overhead,sample,period,dso --stdio
     # Overhead  Samples   Period  Shared Object
         80.00%       28  2800000  [kernel.kallsyms]
          5.71%        2   200000  ld-linux-x86-64.so.2
          5.71%        2   200000  libc.so.6
          5.71%        2   200000  ls
          2.86%        1   100000  libpcre2-8.so.0.11.2

perf mem report:

     $ perf mem report -F overhead,sample,period,dso --stdio
     # Overhead  Samples   Period  Shared Object
         87.50%       28       49  [kernel.kallsyms]
          3.57%        2        2  ld-linux-x86-64.so.2
          3.57%        2        2  libc.so.6
          3.57%        2        2  ls
          1.79%        1        1  libpcre2-8.so.0.11.2

3) Similarly, it was not intuitive (again, to me:)) that -F op/snoop/dtlb percentages are calculated based on sample->weight.

4) I've similar recommended perf-mem command in perf-amd-ibs man page. Can you please update alternate command there.
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/perf/Documentation/perf-amd-ibs.txt?h=v6.15-rc5#n167

Please correct me if I'm missing anything.

Thanks,
Ravi

author maintainer
#21 lore raw

Hi Ravi,

Quoted 4 lines · Ravi Bangoria

Hi Namhyung,

I feel the overall idea is good. Running few simple perf-mem commands on AMD works fine too. Few general feedback below.

Thanks for your review!

Quoted 9 lines
Quoted 5 lines

The name of some new fields are the same as the corresponding sort keys (mem, op, snoop) so I had to change the order whether it's applied as an output field or a sort key. Maybe it's better to name them differently but I couldn't come up with better ideas.

1) These semantic changes of the field name seems counter intuitive (to me). Example:

Quoted 16 lines
   -F mem:

     Without patch:

     $ perf mem report -F overhead,sample,mem --stdio
     # Overhead       Samples  Memory access
         39.29%             1  L3 hit
         37.50%            21  N/A
         23.21%            13  L1 hit

     With patch:

     $ perf mem report -F overhead,sample,mem --stdio
     #                          Memory
     # Overhead       Samples    Other
        100.00%            35   100.0%

Yep, that's because I split the 'mem' part to 'cache' and 'mem' because he_mem_stat can handle up to 8 entries. As your samples hit mostly in the caches, you'd get the similar result when you run:

  $ perf mem report -F overhead,sample,cache --stdio
Quoted 16 lines
   -F 'snoop':

     Without patch:

     $ perf mem report -F overhead,sample,snoop --stdio
     # Overhead       Samples  Snoop
         60.71%            34  N/A
         39.29%             1  HitM
   
     With patchset:

     $ perf mem report -F overhead,sample,snoop --stdio
     #                         --- Snoop ----
     # Overhead       Samples     HitM  Other
        100.00%            35    39.3%  60.7%

This matches to 'Overhead' distribution without patch, right?

Quoted 4 lines

2) It was not intuitive (to me:)) that perf-mem overhead is calculated using sample->weight by overwriting sample->period. I also don't see it documented anywhere (or did I miss it?)

I don't see the documentation and I also find it confusing. Sometimes I think the weight is better but sometimes not. :( At least we could add and option to control that (like --use-weight ?).

Also we now have 'weight' output field so users can see it, althought it shows averages.

Quoted 21 lines
   perf report:

     $ perf report -F overhead,sample,period,dso --stdio
     # Overhead  Samples   Period  Shared Object
         80.00%       28  2800000  [kernel.kallsyms]
          5.71%        2   200000  ld-linux-x86-64.so.2
          5.71%        2   200000  libc.so.6
          5.71%        2   200000  ls
          2.86%        1   100000  libpcre2-8.so.0.11.2

   perf mem report:

     $ perf mem report -F overhead,sample,period,dso --stdio
     # Overhead  Samples   Period  Shared Object
         87.50%       28       49  [kernel.kallsyms]
          3.57%        2        2  ld-linux-x86-64.so.2
          3.57%        2        2  libc.so.6
          3.57%        2        2  ls
          1.79%        1        1  libpcre2-8.so.0.11.2
Quoted 2 lines

3) Similarly, it was not intuitive (again, to me:)) that -F op/snoop/dtlb percentages are calculated based on sample->weight.

Hmm.. ok. Maybe better to use the original period for percentage breakdown in the new output fields. For examples, in the above result you have 13 samples for L1 and 1 sample for L3 but the weight of L3 access is bigger. But I guess users probably want to see L1 access was dominant.

Quoted 4 lines

4) I've similar recommended perf-mem command in perf-amd-ibs man page. Can you please update alternate command there.
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/perf/Documentation/perf-amd-ibs.txt?h=v6.15-rc5#n167

Sure will do.

Thanks,
Namhyung

Quoted 5 lines

Please correct me if I'm missing anything.

Thanks,
Ravi

#22 lore raw

Hi Namhyung,

Quoted 20 lines
Quoted 20 lines
Quoted 4 lines

The name of some new fields are the same as the corresponding sort keys (mem, op, snoop) so I had to change the order whether it's applied as an output field or a sort key. Maybe it's better to name them differently but I couldn't come up with better ideas.

1) These semantic changes of the field name seems counter intuitive (to me). Example:

   -F mem:

     Without patch:

     $ perf mem report -F overhead,sample,mem --stdio
     # Overhead       Samples  Memory access
         39.29%             1  L3 hit
         37.50%            21  N/A
         23.21%            13  L1 hit

     With patch:
Quoted 7 lines
Quoted 4 lines
     $ perf mem report -F overhead,sample,mem --stdio
     #                          Memory
     # Overhead       Samples    Other
        100.00%            35   100.0%

Yep, that's because I split the 'mem' part to 'cache' and 'mem' because he_mem_stat can handle up to 8 entries.

+1.

Quoted 15 lines

As your samples hit mostly in
the caches, you'd get the similar result when you run:

  $ perf mem report -F overhead,sample,cache --stdio
Quoted 10 lines
   -F 'snoop':

     Without patch:

     $ perf mem report -F overhead,sample,snoop --stdio
     # Overhead       Samples  Snoop
         60.71%            34  N/A
         39.29%             1  HitM
Quoted 8 lines
Quoted 6 lines

With patchset:

$ perf mem report -F overhead,sample,snoop --stdio
# --- Snoop ----
# Overhead Samples HitM Other
100.00% 35 39.3% 60.7%

This matches to 'Overhead' distribution without patch, right?

Right, it does.

Quoted 7 lines
Quoted 3 lines

2) It was not intuitive (to me:)) that perf-mem overhead is calculated using sample->weight by overwriting sample->period. I also don't see it documented anywhere (or did I miss it?)

I don't see the documentation and I also find it confusing. Sometimes I think the weight is better but sometimes not. :( At least we could add and option to control that (like --use-weight ?).

this and below ...

Quoted 24 lines

Also we now have 'weight' output field so users can see it, althought it shows averages.

Quoted 21 lines
   perf report:

     $ perf report -F overhead,sample,period,dso --stdio
     # Overhead  Samples   Period  Shared Object
         80.00%       28  2800000  [kernel.kallsyms]
          5.71%        2   200000  ld-linux-x86-64.so.2
          5.71%        2   200000  libc.so.6
          5.71%        2   200000  ls
          2.86%        1   100000  libpcre2-8.so.0.11.2

   perf mem report:

     $ perf mem report -F overhead,sample,period,dso --stdio
     # Overhead  Samples   Period  Shared Object
         87.50%       28       49  [kernel.kallsyms]
          3.57%        2        2  ld-linux-x86-64.so.2
          3.57%        2        2  libc.so.6
          3.57%        2        2  ls
          1.79%        1        1  libpcre2-8.so.0.11.2
Quoted 8 lines
Quoted 2 lines

3) Similarly, it was not intuitive (again, to me:)) that -F op/snoop/dtlb percentages are calculated based on sample->weight.

Hmm.. ok. Maybe better to use the original period for percentage breakdown in the new output fields. For examples, in the above result you have 13 samples for L1 and 1 sample for L3 but the weight of L3 access is bigger. But I guess users probably want to see L1 access was dominant.

... I'm also not sure. Logically, it makes sense to use weight as overhead. Also it dates back to ~2014 and nobody has complained so far. So I'm just being pedantic 🙂. For now, how about just document it in the perf-mem man page and leave it. Attaching the patch at the end.

Quoted 5 lines
Quoted 3 lines

4) I've similar recommended perf-mem command in perf-amd-ibs man page. Can you please update alternate command there.
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/perf/Documentation/perf-amd-ibs.txt?h=v6.15-rc5#n167

Sure will do.

Thanks!

------------><---------------
From 7e4393ab7b20f8d89a5dece08fdd925e3e50b15a Mon Sep 17 00:00:00 2001
From: Ravi Bangoria <[email protected]>
Date: Mon, 12 May 2025 06:22:57 +0000
Subject: [PATCH] perf mem doc: Describe overhead calculation in brief

Unlike perf-report which uses sample period for overhead calculation, perf-mem overhead is calculated using sample weight. Describe perf-mem overhead calculation method in it's man page.

---
 tools/perf/Documentation/perf-mem.txt | 19 +++++++++++++++++++
 1 file changed, 19 insertions(+)
tools/perf/Documentation/perf-mem.txt +19 -0
index a9e3c71a2205..965e73d37772 100644
@@ -137,6 +137,25 @@ REPORT OPTIONS
 In addition, for report all perf report options are valid, and for record
 all perf record options.
 
+OVERHEAD CALCULATION
+--------------------
+Unlike linkperf:perf-report[1], which calculates overhead from the actual
+sample period, perf-mem overhead is calculated using sample weight. E.g.
+there are two samples in perf.data file, both with the same sample period,
+but one sample with weight 180 and the other with weight 20:
+
+  $ perf script -F period,data_src,weight,ip,sym
+  100000    629080842 |OP LOAD|LVL L3 hit|...     20       7e69b93ca524 strcmp
+  100000   1a29081042 |OP LOAD|LVL RAM hit|...   180   ffffffff82429168 memcpy
+
+  $ perf report -F overhead,symbol
+  50%   [.] strcmp
+  50%   [k] memcpy
+
+  $ perf mem report -F overhead,symbol
+  90%   [k] memcpy
+  10%   [.] strcmp
+
 SEE ALSO
 --------
 linkperf:perf-record[1], linkperf:perf-report[1], linkperf:perf-arm-spe[1]
Signature · 4 lines
2.43.0

Thanks,
Ravi