[PATCH v1 1/2] perf sharded_mutex: Introduce sharded_mutex

[PATCH v1 1/2] perf sharded_mutex: Introduce sharded_mutex

[Ian Rogers]

Per object mutexes may come with significant memory cost while a
global mutex can suffer from unnecessary contention. A sharded mutex
is a compromise where objects are hashed and then a particular mutex
for the hash of the object used. Contention can be controlled by the
number of shards.

Signed-off-by: Ian Rogers <irogers@google.com>
---
 tools/perf/util/Build           |  1 +
 tools/perf/util/sharded_mutex.c | 28 ++++++++++++++++++++++++++++
 tools/perf/util/sharded_mutex.h | 27 +++++++++++++++++++++++++++
 3 files changed, 56 insertions(+)
 create mode 100644 tools/perf/util/sharded_mutex.c
 create mode 100644 tools/perf/util/sharded_mutex.h

diff --git a/tools/perf/util/Build b/tools/perf/util/Build
index ff2fd1a36bb8..96f4ea1d45c5 100644
--- a/tools/perf/util/Build
+++ b/tools/perf/util/Build
@@ -145,6 +145,7 @@ perf-y += mem2node.o
 perf-y += clockid.o
 perf-y += list_sort.o
 perf-y += mutex.o
+perf-y += sharded_mutex.o
 
 perf-$(CONFIG_LIBBPF) += bpf-loader.o
 perf-$(CONFIG_LIBBPF) += bpf_map.o
diff --git a/tools/perf/util/sharded_mutex.c b/tools/perf/util/sharded_mutex.c
new file mode 100644
index 000000000000..2c6a1d0ac4f7
--- /dev/null
+++ b/tools/perf/util/sharded_mutex.c
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: GPL-2.0
+#include "sharded_mutex.h"
+
+#include <stdlib.h>
+
+struct sharded_mutex *sharded_mutex__new(size_t num_shards)
+{
+	struct sharded_mutex *result;
+	size_t size = sizeof(*result) + sizeof(struct mutex) * num_shards;
+
+	result = malloc(size);
+	if (!result)
+		return NULL;
+
+	result->num_shards = num_shards;
+	for (size_t i = 0; i < num_shards; i++)
+		mutex_init(&result->mutexes[i]);
+
+	return result;
+}
+
+void sharded_mutex__delete(struct sharded_mutex *sm)
+{
+	for (size_t i = 0; i < sm->num_shards; i++)
+		mutex_destroy(&sm->mutexes[i]);
+
+	free(sm);
+}
diff --git a/tools/perf/util/sharded_mutex.h b/tools/perf/util/sharded_mutex.h
new file mode 100644
index 000000000000..aa649a63a54f
--- /dev/null
+++ b/tools/perf/util/sharded_mutex.h
@@ -0,0 +1,27 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef PERF_SHARDED_MUTEX_H
+#define PERF_SHARDED_MUTEX_H
+
+#include "mutex.h"
+
+/*
+ * In a situation where a lock is needed per object, having a mutex can be
+ * relatively memory expensive (40 bytes on x86-64). If the object can be
+ * constantly hashed, a sharded mutex is an alternative global pool of mutexes
+ * where the mutex is looked up from a hash value. This can lead to collisions
+ * if the number of shards isn't large enough.
+ */
+struct sharded_mutex {
+	size_t num_shards;
+	struct mutex mutexes[];
+};
+
+struct sharded_mutex *sharded_mutex__new(size_t num_shards);
+void sharded_mutex__delete(struct sharded_mutex *sm);
+
+static inline struct mutex *sharded_mutex__get_mutex(struct sharded_mutex *sm, size_t hash)
+{
+	return &sm->mutexes[hash % sm->num_shards];
+}
+
+#endif  /* PERF_SHARDED_MUTEX_H */
-- 
2.41.0.162.gfafddb0af9-goog

[PATCH v1 2/2] perf annotation: Switch lock from a mutex to a sharded_mutex

[Ian Rogers]

Remove the "struct mutex lock" variable from annotation that is
allocated per symbol. This removes in the region of 40 bytes per
symbol allocation. Use a sharded mutex where the number of shards is
set to the number of CPUs. Assuming good hashing of the annotation
(done based on the pointer), this means in order to contend there
needs to be more threads than CPUs, which is not currently true in any
perf command. Were contention an issue it is straightforward to
increase the number of shards in the mutex.

On my Debian/glibc based machine, this reduces the size of struct
annotation from 136 bytes to 96 bytes, or nearly 30%.

Signed-off-by: Ian Rogers <irogers@google.com>
---
 tools/perf/builtin-top.c          | 14 +++----
 tools/perf/ui/browsers/annotate.c | 10 ++---
 tools/perf/util/annotate.c        | 66 ++++++++++++++++++++++++++-----
 tools/perf/util/annotate.h        | 11 ++++--
 4 files changed, 77 insertions(+), 24 deletions(-)

diff --git a/tools/perf/builtin-top.c b/tools/perf/builtin-top.c
index c363c04e16df..1baa2acb3ced 100644
--- a/tools/perf/builtin-top.c
+++ b/tools/perf/builtin-top.c
@@ -137,10 +137,10 @@ static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
 	}
 
 	notes = symbol__annotation(sym);
-	mutex_lock(&notes->lock);
+	annotation__lock(notes);
 
 	if (!symbol__hists(sym, top->evlist->core.nr_entries)) {
-		mutex_unlock(&notes->lock);
+		annotation__unlock(notes);
 		pr_err("Not enough memory for annotating '%s' symbol!\n",
 		       sym->name);
 		sleep(1);
@@ -156,7 +156,7 @@ static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
 		pr_err("Couldn't annotate %s: %s\n", sym->name, msg);
 	}
 
-	mutex_unlock(&notes->lock);
+	annotation__unlock(notes);
 	return err;
 }
 
@@ -211,12 +211,12 @@ static void perf_top__record_precise_ip(struct perf_top *top,
 
 	notes = symbol__annotation(sym);
 
-	if (!mutex_trylock(&notes->lock))
+	if (!annotation__trylock(notes))
 		return;
 
 	err = hist_entry__inc_addr_samples(he, sample, evsel, ip);
 
-	mutex_unlock(&notes->lock);
+	annotation__unlock(notes);
 
 	if (unlikely(err)) {
 		/*
@@ -253,7 +253,7 @@ static void perf_top__show_details(struct perf_top *top)
 	symbol = he->ms.sym;
 	notes = symbol__annotation(symbol);
 
-	mutex_lock(&notes->lock);
+	annotation__lock(notes);
 
 	symbol__calc_percent(symbol, evsel);
 
@@ -274,7 +274,7 @@ static void perf_top__show_details(struct perf_top *top)
 	if (more != 0)
 		printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 out_unlock:
-	mutex_unlock(&notes->lock);
+	annotation__unlock(notes);
 }
 
 static void perf_top__resort_hists(struct perf_top *t)
diff --git a/tools/perf/ui/browsers/annotate.c b/tools/perf/ui/browsers/annotate.c
index 70bad42b807b..ccdb2cd11fbf 100644
--- a/tools/perf/ui/browsers/annotate.c
+++ b/tools/perf/ui/browsers/annotate.c
@@ -314,7 +314,7 @@ static void annotate_browser__calc_percent(struct annotate_browser *browser,
 
 	browser->entries = RB_ROOT;
 
-	mutex_lock(&notes->lock);
+	annotation__lock(notes);
 
 	symbol__calc_percent(sym, evsel);
 
@@ -343,7 +343,7 @@ static void annotate_browser__calc_percent(struct annotate_browser *browser,
 		}
 		disasm_rb_tree__insert(browser, &pos->al);
 	}
-	mutex_unlock(&notes->lock);
+	annotation__unlock(notes);
 
 	browser->curr_hot = rb_last(&browser->entries);
 }
@@ -470,10 +470,10 @@ static bool annotate_browser__callq(struct annotate_browser *browser,
 	}
 
 	notes = symbol__annotation(dl->ops.target.sym);
-	mutex_lock(&notes->lock);
+	annotation__lock(notes);
 
 	if (!symbol__hists(dl->ops.target.sym, evsel->evlist->core.nr_entries)) {
-		mutex_unlock(&notes->lock);
+		annotation__unlock(notes);
 		ui__warning("Not enough memory for annotating '%s' symbol!\n",
 			    dl->ops.target.sym->name);
 		return true;
@@ -482,7 +482,7 @@ static bool annotate_browser__callq(struct annotate_browser *browser,
 	target_ms.maps = ms->maps;
 	target_ms.map = ms->map;
 	target_ms.sym = dl->ops.target.sym;
-	mutex_unlock(&notes->lock);
+	annotation__unlock(notes);
 	symbol__tui_annotate(&target_ms, evsel, hbt, browser->opts);
 	sym_title(ms->sym, ms->map, title, sizeof(title), browser->opts->percent_type);
 	ui_browser__show_title(&browser->b, title);
diff --git a/tools/perf/util/annotate.c b/tools/perf/util/annotate.c
index bde890cfa620..310dde7b3419 100644
--- a/tools/perf/util/annotate.c
+++ b/tools/perf/util/annotate.c
@@ -32,6 +32,7 @@
 #include "block-range.h"
 #include "string2.h"
 #include "util/event.h"
+#include "util/sharded_mutex.h"
 #include "arch/common.h"
 #include "namespaces.h"
 #include <regex.h>
@@ -856,7 +857,7 @@ void symbol__annotate_zero_histograms(struct symbol *sym)
 {
 	struct annotation *notes = symbol__annotation(sym);
 
-	mutex_lock(&notes->lock);
+	annotation__lock(notes);
 	if (notes->src != NULL) {
 		memset(notes->src->histograms, 0,
 		       notes->src->nr_histograms * notes->src->sizeof_sym_hist);
@@ -864,7 +865,7 @@ void symbol__annotate_zero_histograms(struct symbol *sym)
 			memset(notes->src->cycles_hist, 0,
 				symbol__size(sym) * sizeof(struct cyc_hist));
 	}
-	mutex_unlock(&notes->lock);
+	annotation__unlock(notes);
 }
 
 static int __symbol__account_cycles(struct cyc_hist *ch,
@@ -1121,7 +1122,7 @@ void annotation__compute_ipc(struct annotation *notes, size_t size)
 	notes->hit_insn = 0;
 	notes->cover_insn = 0;
 
-	mutex_lock(&notes->lock);
+	annotation__lock(notes);
 	for (offset = size - 1; offset >= 0; --offset) {
 		struct cyc_hist *ch;
 
@@ -1140,7 +1141,7 @@ void annotation__compute_ipc(struct annotation *notes, size_t size)
 			notes->have_cycles = true;
 		}
 	}
-	mutex_unlock(&notes->lock);
+	annotation__unlock(notes);
 }
 
 int addr_map_symbol__inc_samples(struct addr_map_symbol *ams, struct perf_sample *sample,
@@ -1291,17 +1292,64 @@ int disasm_line__scnprintf(struct disasm_line *dl, char *bf, size_t size, bool r
 	return ins__scnprintf(&dl->ins, bf, size, &dl->ops, max_ins_name);
 }
 
-void annotation__init(struct annotation *notes)
+void annotation__exit(struct annotation *notes)
 {
-	mutex_init(&notes->lock);
+	annotated_source__delete(notes->src);
 }
 
-void annotation__exit(struct annotation *notes)
+static struct sharded_mutex *sharded_mutex;
+
+static void annotation__init_sharded_mutex(void)
 {
-	annotated_source__delete(notes->src);
-	mutex_destroy(&notes->lock);
+	/* As many mutexes as there are CPUs. */
+	sharded_mutex = sharded_mutex__new(cpu__max_present_cpu().cpu);
+}
+
+static size_t annotation__hash(const struct annotation *notes)
+{
+	return ((size_t)notes) >> 4;
 }
 
+static struct mutex *annotation__get_mutex(const struct annotation *notes)
+{
+	static pthread_once_t once = PTHREAD_ONCE_INIT;
+
+	pthread_once(&once, annotation__init_sharded_mutex);
+	if (!sharded_mutex)
+		return NULL;
+
+	return sharded_mutex__get_mutex(sharded_mutex, annotation__hash(notes));
+}
+
+void annotation__lock(struct annotation *notes)
+	NO_THREAD_SAFETY_ANALYSIS
+{
+	struct mutex *mutex = annotation__get_mutex(notes);
+
+	if (mutex)
+		mutex_lock(mutex);
+}
+
+void annotation__unlock(struct annotation *notes)
+	NO_THREAD_SAFETY_ANALYSIS
+{
+	struct mutex *mutex = annotation__get_mutex(notes);
+
+	if (mutex)
+		mutex_unlock(mutex);
+}
+
+bool annotation__trylock(struct annotation *notes)
+{
+	struct mutex *mutex = annotation__get_mutex(notes);
+
+	if (!mutex)
+		return false;
+
+	return mutex_trylock(mutex);
+}
+
+
 static void annotation_line__add(struct annotation_line *al, struct list_head *head)
 {
 	list_add_tail(&al->node, head);
diff --git a/tools/perf/util/annotate.h b/tools/perf/util/annotate.h
index 1c6335b8333a..962780559176 100644
--- a/tools/perf/util/annotate.h
+++ b/tools/perf/util/annotate.h
@@ -271,8 +271,7 @@ struct annotated_source {
 	struct sym_hist	   *histograms;
 };
 
-struct annotation {
-	struct mutex lock;
+struct LOCKABLE annotation {
 	u64			max_coverage;
 	u64			start;
 	u64			hit_cycles;
@@ -298,9 +297,15 @@ struct annotation {
 	struct annotated_source *src;
 };
 
-void annotation__init(struct annotation *notes);
+static inline void annotation__init(struct annotation *notes __maybe_unused)
+{
+}
 void annotation__exit(struct annotation *notes);
 
+void annotation__lock(struct annotation *notes) EXCLUSIVE_LOCK_FUNCTION(*notes);
+void annotation__unlock(struct annotation *notes) UNLOCK_FUNCTION(*notes);
+bool annotation__trylock(struct annotation *notes) EXCLUSIVE_TRYLOCK_FUNCTION(true, *notes);
+
 static inline int annotation__cycles_width(struct annotation *notes)
 {
 	if (notes->have_cycles && notes->options->show_minmax_cycle)
-- 
2.41.0.162.gfafddb0af9-goog

Re: [PATCH v1 2/2] perf annotation: Switch lock from a mutex to a sharded_mutex

[Namhyung Kim]

Hi Ian,

On Sun, Jun 11, 2023 at 12:28 AM Ian Rogers <irogers@google.com> wrote:
>
> Remove the "struct mutex lock" variable from annotation that is
> allocated per symbol. This removes in the region of 40 bytes per
> symbol allocation. Use a sharded mutex where the number of shards is
> set to the number of CPUs. Assuming good hashing of the annotation
> (done based on the pointer), this means in order to contend there
> needs to be more threads than CPUs, which is not currently true in any
> perf command. Were contention an issue it is straightforward to
> increase the number of shards in the mutex.
>
> On my Debian/glibc based machine, this reduces the size of struct
> annotation from 136 bytes to 96 bytes, or nearly 30%.

That's quite a good improvement given the number of symbols
we can have in a report session!

>
> Signed-off-by: Ian Rogers <irogers@google.com>
> ---

[SNIP]
> @@ -1291,17 +1292,64 @@ int disasm_line__scnprintf(struct disasm_line *dl, char *bf, size_t size, bool r
>         return ins__scnprintf(&dl->ins, bf, size, &dl->ops, max_ins_name);
>  }
>
> -void annotation__init(struct annotation *notes)
> +void annotation__exit(struct annotation *notes)
>  {
> -       mutex_init(&notes->lock);
> +       annotated_source__delete(notes->src);
>  }
>
> -void annotation__exit(struct annotation *notes)
> +static struct sharded_mutex *sharded_mutex;
> +
> +static void annotation__init_sharded_mutex(void)
>  {
> -       annotated_source__delete(notes->src);
> -       mutex_destroy(&notes->lock);
> +       /* As many mutexes as there are CPUs. */
> +       sharded_mutex = sharded_mutex__new(cpu__max_present_cpu().cpu);
> +}
> +
> +static size_t annotation__hash(const struct annotation *notes)
> +{
> +       return ((size_t)notes) >> 4;

But I'm afraid it might create more contention depending on the
malloc implementation.  If it always returns 128-byte (or 256)
aligned memory for this struct then it could always collide in the
slot 0 if the number of CPU is 8 or less, right?

Thanks,
Namhyung


>  }
>
> +static struct mutex *annotation__get_mutex(const struct annotation *notes)
> +{
> +       static pthread_once_t once = PTHREAD_ONCE_INIT;
> +
> +       pthread_once(&once, annotation__init_sharded_mutex);
> +       if (!sharded_mutex)
> +               return NULL;
> +
> +       return sharded_mutex__get_mutex(sharded_mutex, annotation__hash(notes));
> +}
> +
> +void annotation__lock(struct annotation *notes)
> +       NO_THREAD_SAFETY_ANALYSIS
> +{
> +       struct mutex *mutex = annotation__get_mutex(notes);
> +
> +       if (mutex)
> +               mutex_lock(mutex);
> +}
> +
> +void annotation__unlock(struct annotation *notes)
> +       NO_THREAD_SAFETY_ANALYSIS
> +{
> +       struct mutex *mutex = annotation__get_mutex(notes);
> +
> +       if (mutex)
> +               mutex_unlock(mutex);
> +}
> +
> +bool annotation__trylock(struct annotation *notes)
> +{
> +       struct mutex *mutex = annotation__get_mutex(notes);
> +
> +       if (!mutex)
> +               return false;
> +
> +       return mutex_trylock(mutex);
> +}
> +
> +
>  static void annotation_line__add(struct annotation_line *al, struct list_head *head)
>  {
>         list_add_tail(&al->node, head);

Re: [PATCH v1 2/2] perf annotation: Switch lock from a mutex to a sharded_mutex

[Ian Rogers]

On Wed, Jun 14, 2023 at 5:34 PM Namhyung Kim <namhyung@kernel.org> wrote:
>
> Hi Ian,
>
> On Sun, Jun 11, 2023 at 12:28 AM Ian Rogers <irogers@google.com> wrote:
> >
> > Remove the "struct mutex lock" variable from annotation that is
> > allocated per symbol. This removes in the region of 40 bytes per
> > symbol allocation. Use a sharded mutex where the number of shards is
> > set to the number of CPUs. Assuming good hashing of the annotation
> > (done based on the pointer), this means in order to contend there
> > needs to be more threads than CPUs, which is not currently true in any
> > perf command. Were contention an issue it is straightforward to
> > increase the number of shards in the mutex.
> >
> > On my Debian/glibc based machine, this reduces the size of struct
> > annotation from 136 bytes to 96 bytes, or nearly 30%.
>
> That's quite a good improvement given the number of symbols
> we can have in a report session!
>
> >
> > Signed-off-by: Ian Rogers <irogers@google.com>
> > ---
>
> [SNIP]
> > @@ -1291,17 +1292,64 @@ int disasm_line__scnprintf(struct disasm_line *dl, char *bf, size_t size, bool r
> >         return ins__scnprintf(&dl->ins, bf, size, &dl->ops, max_ins_name);
> >  }
> >
> > -void annotation__init(struct annotation *notes)
> > +void annotation__exit(struct annotation *notes)
> >  {
> > -       mutex_init(&notes->lock);
> > +       annotated_source__delete(notes->src);
> >  }
> >
> > -void annotation__exit(struct annotation *notes)
> > +static struct sharded_mutex *sharded_mutex;
> > +
> > +static void annotation__init_sharded_mutex(void)
> >  {
> > -       annotated_source__delete(notes->src);
> > -       mutex_destroy(&notes->lock);
> > +       /* As many mutexes as there are CPUs. */
> > +       sharded_mutex = sharded_mutex__new(cpu__max_present_cpu().cpu);
> > +}
> > +
> > +static size_t annotation__hash(const struct annotation *notes)
> > +{
> > +       return ((size_t)notes) >> 4;
>
> But I'm afraid it might create more contention depending on the
> malloc implementation.  If it always returns 128-byte (or 256)
> aligned memory for this struct then it could always collide in the
> slot 0 if the number of CPU is 8 or less, right?

Right, I think we can use a secondary hash and hashmap.h has one lying
around for us:
https://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git/tree/tools/perf/util/hashmap.h?h=tmp.perf-tools-next#n15
it will mean that the sharded locks will need to be a power of 2
capacity. I'll work on a v2. Fwiw, the hash of a pointer for
collection like those in abseil is just the pointer value, so I'll
drop the shift to remove the low-bits once I'm using a more expensive
hash function.

Thanks,
Ian

> Thanks,
> Namhyung
>
>
> >  }
> >
> > +static struct mutex *annotation__get_mutex(const struct annotation *notes)
> > +{
> > +       static pthread_once_t once = PTHREAD_ONCE_INIT;
> > +
> > +       pthread_once(&once, annotation__init_sharded_mutex);
> > +       if (!sharded_mutex)
> > +               return NULL;
> > +
> > +       return sharded_mutex__get_mutex(sharded_mutex, annotation__hash(notes));
> > +}
> > +
> > +void annotation__lock(struct annotation *notes)
> > +       NO_THREAD_SAFETY_ANALYSIS
> > +{
> > +       struct mutex *mutex = annotation__get_mutex(notes);
> > +
> > +       if (mutex)
> > +               mutex_lock(mutex);
> > +}
> > +
> > +void annotation__unlock(struct annotation *notes)
> > +       NO_THREAD_SAFETY_ANALYSIS
> > +{
> > +       struct mutex *mutex = annotation__get_mutex(notes);
> > +
> > +       if (mutex)
> > +               mutex_unlock(mutex);
> > +}
> > +
> > +bool annotation__trylock(struct annotation *notes)
> > +{
> > +       struct mutex *mutex = annotation__get_mutex(notes);
> > +
> > +       if (!mutex)
> > +               return false;
> > +
> > +       return mutex_trylock(mutex);
> > +}
> > +
> > +
> >  static void annotation_line__add(struct annotation_line *al, struct list_head *head)
> >  {
> >         list_add_tail(&al->node, head);