[PATCH v6 0/2] Optimize performance of update hash-map when free is zero

[PATCH v6 0/2] Optimize performance of update hash-map when free is zero

[Feng zhou]

From: Feng Zhou <zhoufeng.zf@bytedance.com>

We encountered bad case on big system with 96 CPUs that
alloc_htab_elem() would last for 1ms. The reason is that after the
prealloc hashtab has no free elems, when trying to update, it will still
grab spin_locks of all cpus. If there are multiple update users, the
competition is very serious.

0001: Use head->first to check whether the free list is empty or not before taking
the lock.
0002: Add benchmark to reproduce this worst case.

Changelog:
v5->v6: Addressed comments from Alexei Starovoitov.
- Adjust the commit log.
some details in here:
https://lore.kernel.org/all/20220608021050.47279-1-zhoufeng.zf@bytedance.com/

v4->v5: Addressed comments from Alexei Starovoitov.
- Use head->first.
- Use cpu+max_entries.
some details in here:
https://lore.kernel.org/bpf/20220601084149.13097-1-zhoufeng.zf@bytedance.com/

v3->v4: Addressed comments from Daniel Borkmann.
- Use READ_ONCE/WRITE_ONCE.
some details in here:
https://lore.kernel.org/all/20220530091340.53443-1-zhoufeng.zf@bytedance.com/

v2->v3: Addressed comments from Alexei Starovoitov, Andrii Nakryiko.
- Adjust the way the benchmark is tested.
- Adjust the code format.
some details in here:
https://lore.kernel.org/all/20220524075306.32306-1-zhoufeng.zf@bytedance.com/T/

v1->v2: Addressed comments from Alexei Starovoitov.
- add a benchmark to reproduce the issue.
- Adjust the code format that avoid adding indent.
some details in here:
https://lore.kernel.org/all/877ac441-045b-1844-6938-fcaee5eee7f2@bytedance.com/T/

Feng Zhou (2):
  bpf: avoid grabbing spin_locks of all cpus when no free elems
  selftest/bpf/benchs: Add bpf_map benchmark

 kernel/bpf/percpu_freelist.c                  | 20 ++--
 tools/testing/selftests/bpf/Makefile          |  4 +-
 tools/testing/selftests/bpf/bench.c           |  2 +
 .../benchs/bench_bpf_hashmap_full_update.c    | 96 +++++++++++++++++++
 .../run_bench_bpf_hashmap_full_update.sh      | 11 +++
 .../bpf/progs/bpf_hashmap_full_update_bench.c | 40 ++++++++
 6 files changed, 166 insertions(+), 7 deletions(-)
 create mode 100644 tools/testing/selftests/bpf/benchs/bench_bpf_hashmap_full_update.c
 create mode 100755 tools/testing/selftests/bpf/benchs/run_bench_bpf_hashmap_full_update.sh
 create mode 100644 tools/testing/selftests/bpf/progs/bpf_hashmap_full_update_bench.c

-- 
2.20.1

[PATCH v6 1/2] bpf: avoid grabbing spin_locks of all cpus when no free elems

[Feng zhou]

From: Feng Zhou <zhoufeng.zf@bytedance.com>

This patch use head->first in pcpu_freelist_head to check freelist
having free or not. If having, grab spin_lock, or check next cpu's
freelist.

Before patch: hash_map performance
./map_perf_test 1
0:hash_map_perf pre-alloc 1043397 events per sec
...
The average of the test results is around 1050000 events per sec.

hash_map the worst: no free
./run_bench_bpf_hashmap_full_update.sh
Setting up benchmark 'bpf-hashmap-ful-update'...
Benchmark 'bpf-hashmap-ful-update' started.
1:hash_map_full_perf 15687 events per sec
...
The average of the test results is around 16000 events per sec.

ftrace trace:
0)               |  htab_map_update_elem() {
0)               |      __pcpu_freelist_pop() {
0)               |        _raw_spin_lock()
0)               |        _raw_spin_unlock()
0)               |        ...
0) + 25.188 us   |      }
0) + 28.439 us   |  }

The test machine is 16C, trying to get spin_lock 17 times, in addition
to 16c, there is an extralist.

after patch: hash_map performance
./map_perf_test 1
0:hash_map_perf pre-alloc 1053298 events per sec
...
The average of the test results is around 1050000 events per sec.

hash_map worst: no free
./run_bench_bpf_hashmap_full_update.sh
Setting up benchmark 'bpf-hashmap-ful-update'...
Benchmark 'bpf-hashmap-ful-update' started.
1:hash_map_full_perf 555830 events per sec
...
The average of the test results is around 550000 events per sec.

ftrace trace:
0)               |  htab_map_update_elem() {
0)               |    alloc_htab_elem() {
0)   0.586 us    |      __pcpu_freelist_pop();
0)   0.945 us    |    }
0)   8.669 us    |  }

It can be seen that after adding this patch, the map performance is
almost not degraded, and when free=0, first check head->first instead of
directly acquiring spin_lock.

Co-developed-by: Chengming Zhou <zhouchengming@bytedance.com>
Signed-off-by: Chengming Zhou <zhouchengming@bytedance.com>
Signed-off-by: Feng Zhou <zhoufeng.zf@bytedance.com>
---
 kernel/bpf/percpu_freelist.c | 20 ++++++++++++++------
 1 file changed, 14 insertions(+), 6 deletions(-)

diff --git a/kernel/bpf/percpu_freelist.c b/kernel/bpf/percpu_freelist.c
index 3d897de89061..00b874c8e889 100644
--- a/kernel/bpf/percpu_freelist.c
+++ b/kernel/bpf/percpu_freelist.c
@@ -31,7 +31,7 @@ static inline void pcpu_freelist_push_node(struct pcpu_freelist_head *head,
 					   struct pcpu_freelist_node *node)
 {
 	node->next = head->first;
-	head->first = node;
+	WRITE_ONCE(head->first, node);
 }
 
 static inline void ___pcpu_freelist_push(struct pcpu_freelist_head *head,
@@ -130,14 +130,17 @@ static struct pcpu_freelist_node *___pcpu_freelist_pop(struct pcpu_freelist *s)
 	orig_cpu = cpu = raw_smp_processor_id();
 	while (1) {
 		head = per_cpu_ptr(s->freelist, cpu);
+		if (!READ_ONCE(head->first))
+			goto next_cpu;
 		raw_spin_lock(&head->lock);
 		node = head->first;
 		if (node) {
-			head->first = node->next;
+			WRITE_ONCE(head->first, node->next);
 			raw_spin_unlock(&head->lock);
 			return node;
 		}
 		raw_spin_unlock(&head->lock);
+next_cpu:
 		cpu = cpumask_next(cpu, cpu_possible_mask);
 		if (cpu >= nr_cpu_ids)
 			cpu = 0;
@@ -146,10 +149,12 @@ static struct pcpu_freelist_node *___pcpu_freelist_pop(struct pcpu_freelist *s)
 	}
 
 	/* per cpu lists are all empty, try extralist */
+	if (!READ_ONCE(s->extralist.first))
+		return NULL;
 	raw_spin_lock(&s->extralist.lock);
 	node = s->extralist.first;
 	if (node)
-		s->extralist.first = node->next;
+		WRITE_ONCE(s->extralist.first, node->next);
 	raw_spin_unlock(&s->extralist.lock);
 	return node;
 }
@@ -164,15 +169,18 @@ ___pcpu_freelist_pop_nmi(struct pcpu_freelist *s)
 	orig_cpu = cpu = raw_smp_processor_id();
 	while (1) {
 		head = per_cpu_ptr(s->freelist, cpu);
+		if (!READ_ONCE(head->first))
+			goto next_cpu;
 		if (raw_spin_trylock(&head->lock)) {
 			node = head->first;
 			if (node) {
-				head->first = node->next;
+				WRITE_ONCE(head->first, node->next);
 				raw_spin_unlock(&head->lock);
 				return node;
 			}
 			raw_spin_unlock(&head->lock);
 		}
+next_cpu:
 		cpu = cpumask_next(cpu, cpu_possible_mask);
 		if (cpu >= nr_cpu_ids)
 			cpu = 0;
@@ -181,11 +189,11 @@ ___pcpu_freelist_pop_nmi(struct pcpu_freelist *s)
 	}
 
 	/* cannot pop from per cpu lists, try extralist */
-	if (!raw_spin_trylock(&s->extralist.lock))
+	if (!READ_ONCE(s->extralist.first) || !raw_spin_trylock(&s->extralist.lock))
 		return NULL;
 	node = s->extralist.first;
 	if (node)
-		s->extralist.first = node->next;
+		WRITE_ONCE(s->extralist.first, node->next);
 	raw_spin_unlock(&s->extralist.lock);
 	return node;
 }
-- 
2.20.1

[PATCH v6 2/2] selftest/bpf/benchs: Add bpf_map benchmark

[Feng zhou]

From: Feng Zhou <zhoufeng.zf@bytedance.com>

Add benchmark for hash_map to reproduce the worst case
that non-stop update when map's free is zero.

Just like this:
./run_bench_bpf_hashmap_full_update.sh
Setting up benchmark 'bpf-hashmap-ful-update'...
Benchmark 'bpf-hashmap-ful-update' started.
1:hash_map_full_perf 555830 events per sec
...

Signed-off-by: Feng Zhou <zhoufeng.zf@bytedance.com>
---
 tools/testing/selftests/bpf/Makefile          |  4 +-
 tools/testing/selftests/bpf/bench.c           |  2 +
 .../benchs/bench_bpf_hashmap_full_update.c    | 96 +++++++++++++++++++
 .../run_bench_bpf_hashmap_full_update.sh      | 11 +++
 .../bpf/progs/bpf_hashmap_full_update_bench.c | 40 ++++++++
 5 files changed, 152 insertions(+), 1 deletion(-)
 create mode 100644 tools/testing/selftests/bpf/benchs/bench_bpf_hashmap_full_update.c
 create mode 100755 tools/testing/selftests/bpf/benchs/run_bench_bpf_hashmap_full_update.sh
 create mode 100644 tools/testing/selftests/bpf/progs/bpf_hashmap_full_update_bench.c

diff --git a/tools/testing/selftests/bpf/Makefile b/tools/testing/selftests/bpf/Makefile
index 2d3c8c8f558a..8ad7a733a505 100644
--- a/tools/testing/selftests/bpf/Makefile
+++ b/tools/testing/selftests/bpf/Makefile
@@ -560,6 +560,7 @@ $(OUTPUT)/bench_ringbufs.o: $(OUTPUT)/ringbuf_bench.skel.h \
 $(OUTPUT)/bench_bloom_filter_map.o: $(OUTPUT)/bloom_filter_bench.skel.h
 $(OUTPUT)/bench_bpf_loop.o: $(OUTPUT)/bpf_loop_bench.skel.h
 $(OUTPUT)/bench_strncmp.o: $(OUTPUT)/strncmp_bench.skel.h
+$(OUTPUT)/bench_bpf_hashmap_full_update.o: $(OUTPUT)/bpf_hashmap_full_update_bench.skel.h
 $(OUTPUT)/bench.o: bench.h testing_helpers.h $(BPFOBJ)
 $(OUTPUT)/bench: LDLIBS += -lm
 $(OUTPUT)/bench: $(OUTPUT)/bench.o \
@@ -571,7 +572,8 @@ $(OUTPUT)/bench: $(OUTPUT)/bench.o \
 		 $(OUTPUT)/bench_ringbufs.o \
 		 $(OUTPUT)/bench_bloom_filter_map.o \
 		 $(OUTPUT)/bench_bpf_loop.o \
-		 $(OUTPUT)/bench_strncmp.o
+		 $(OUTPUT)/bench_strncmp.o \
+		 $(OUTPUT)/bench_bpf_hashmap_full_update.o
 	$(call msg,BINARY,,$@)
 	$(Q)$(CC) $(CFLAGS) $(LDFLAGS) $(filter %.a %.o,$^) $(LDLIBS) -o $@
 
diff --git a/tools/testing/selftests/bpf/bench.c b/tools/testing/selftests/bpf/bench.c
index f061cc20e776..d8aa62be996b 100644
--- a/tools/testing/selftests/bpf/bench.c
+++ b/tools/testing/selftests/bpf/bench.c
@@ -396,6 +396,7 @@ extern const struct bench bench_hashmap_with_bloom;
 extern const struct bench bench_bpf_loop;
 extern const struct bench bench_strncmp_no_helper;
 extern const struct bench bench_strncmp_helper;
+extern const struct bench bench_bpf_hashmap_full_update;
 
 static const struct bench *benchs[] = {
 	&bench_count_global,
@@ -430,6 +431,7 @@ static const struct bench *benchs[] = {
 	&bench_bpf_loop,
 	&bench_strncmp_no_helper,
 	&bench_strncmp_helper,
+	&bench_bpf_hashmap_full_update,
 };
 
 static void setup_benchmark()
diff --git a/tools/testing/selftests/bpf/benchs/bench_bpf_hashmap_full_update.c b/tools/testing/selftests/bpf/benchs/bench_bpf_hashmap_full_update.c
new file mode 100644
index 000000000000..cec51e0ff4b8
--- /dev/null
+++ b/tools/testing/selftests/bpf/benchs/bench_bpf_hashmap_full_update.c
@@ -0,0 +1,96 @@
+// SPDX-License-Identifier: GPL-2.0
+/* Copyright (c) 2022 Bytedance */
+
+#include <argp.h>
+#include "bench.h"
+#include "bpf_hashmap_full_update_bench.skel.h"
+#include "bpf_util.h"
+
+/* BPF triggering benchmarks */
+static struct ctx {
+	struct bpf_hashmap_full_update_bench *skel;
+} ctx;
+
+#define MAX_LOOP_NUM 10000
+
+static void validate(void)
+{
+	if (env.consumer_cnt != 1) {
+		fprintf(stderr, "benchmark doesn't support multi-consumer!\n");
+		exit(1);
+	}
+}
+
+static void *producer(void *input)
+{
+	while (true) {
+		/* trigger the bpf program */
+		syscall(__NR_getpgid);
+	}
+
+	return NULL;
+}
+
+static void *consumer(void *input)
+{
+	return NULL;
+}
+
+static void measure(struct bench_res *res)
+{
+}
+
+static void setup(void)
+{
+	struct bpf_link *link;
+	int map_fd, i, max_entries;
+
+	setup_libbpf();
+
+	ctx.skel = bpf_hashmap_full_update_bench__open_and_load();
+	if (!ctx.skel) {
+		fprintf(stderr, "failed to open skeleton\n");
+		exit(1);
+	}
+
+	ctx.skel->bss->nr_loops = MAX_LOOP_NUM;
+
+	link = bpf_program__attach(ctx.skel->progs.benchmark);
+	if (!link) {
+		fprintf(stderr, "failed to attach program!\n");
+		exit(1);
+	}
+
+	/* fill hash_map */
+	map_fd = bpf_map__fd(ctx.skel->maps.hash_map_bench);
+	max_entries = bpf_map__max_entries(ctx.skel->maps.hash_map_bench);
+	for (i = 0; i < max_entries; i++)
+		bpf_map_update_elem(map_fd, &i, &i, BPF_ANY);
+}
+
+void hashmap_report_final(struct bench_res res[], int res_cnt)
+{
+	unsigned int nr_cpus = bpf_num_possible_cpus();
+	int i;
+
+	for (i = 0; i < nr_cpus; i++) {
+		u64 time = ctx.skel->bss->percpu_time[i];
+
+		if (!time)
+			continue;
+
+		printf("%d:hash_map_full_perf %lld events per sec\n",
+		       i, ctx.skel->bss->nr_loops * 1000000000ll / time);
+	}
+}
+
+const struct bench bench_bpf_hashmap_full_update = {
+	.name = "bpf-hashmap-ful-update",
+	.validate = validate,
+	.setup = setup,
+	.producer_thread = producer,
+	.consumer_thread = consumer,
+	.measure = measure,
+	.report_progress = NULL,
+	.report_final = hashmap_report_final,
+};
diff --git a/tools/testing/selftests/bpf/benchs/run_bench_bpf_hashmap_full_update.sh b/tools/testing/selftests/bpf/benchs/run_bench_bpf_hashmap_full_update.sh
new file mode 100755
index 000000000000..1e2de838f9fa
--- /dev/null
+++ b/tools/testing/selftests/bpf/benchs/run_bench_bpf_hashmap_full_update.sh
@@ -0,0 +1,11 @@
+#!/bin/bash
+# SPDX-License-Identifier: GPL-2.0
+
+source ./benchs/run_common.sh
+
+set -eufo pipefail
+
+nr_threads=`expr $(cat /proc/cpuinfo | grep "processor"| wc -l) - 1`
+summary=$($RUN_BENCH -p $nr_threads bpf-hashmap-ful-update)
+printf "$summary"
+printf "\n"
diff --git a/tools/testing/selftests/bpf/progs/bpf_hashmap_full_update_bench.c b/tools/testing/selftests/bpf/progs/bpf_hashmap_full_update_bench.c
new file mode 100644
index 000000000000..56957557e3e1
--- /dev/null
+++ b/tools/testing/selftests/bpf/progs/bpf_hashmap_full_update_bench.c
@@ -0,0 +1,40 @@
+// SPDX-License-Identifier: GPL-2.0
+/* Copyright (c) 2022 Bytedance */
+
+#include "vmlinux.h"
+#include <bpf/bpf_helpers.h>
+#include "bpf_misc.h"
+
+char _license[] SEC("license") = "GPL";
+
+#define MAX_ENTRIES 1000
+
+struct {
+	__uint(type, BPF_MAP_TYPE_HASH);
+	__type(key, u32);
+	__type(value, u64);
+	__uint(max_entries, MAX_ENTRIES);
+} hash_map_bench SEC(".maps");
+
+u64 __attribute__((__aligned__(256))) percpu_time[256];
+u64 nr_loops;
+
+static int loop_update_callback(__u32 index, u32 *key)
+{
+	u64 init_val = 1;
+
+	bpf_map_update_elem(&hash_map_bench, key, &init_val, BPF_ANY);
+	return 0;
+}
+
+SEC("fentry/" SYS_PREFIX "sys_getpgid")
+int benchmark(void *ctx)
+{
+	u32 cpu = bpf_get_smp_processor_id();
+	u32 key = cpu + MAX_ENTRIES;
+	u64 start_time = bpf_ktime_get_ns();
+
+	bpf_loop(nr_loops, loop_update_callback, &key, 0);
+	percpu_time[cpu & 255] = bpf_ktime_get_ns() - start_time;
+	return 0;
+}
-- 
2.20.1

Re: [PATCH v6 0/2] Optimize performance of update hash-map when free is zero

[patchwork-bot+netdevbpf]

Hello:

This series was applied to bpf/bpf-next.git (master)
by Alexei Starovoitov <ast@kernel.org>:

On Fri, 10 Jun 2022 10:33:06 +0800 you wrote:
> From: Feng Zhou <zhoufeng.zf@bytedance.com>
> 
> We encountered bad case on big system with 96 CPUs that
> alloc_htab_elem() would last for 1ms. The reason is that after the
> prealloc hashtab has no free elems, when trying to update, it will still
> grab spin_locks of all cpus. If there are multiple update users, the
> competition is very serious.
> 
> [...]

Here is the summary with links:
  - [v6,1/2] bpf: avoid grabbing spin_locks of all cpus when no free elems
    https://git.kernel.org/bpf/bpf-next/c/54a9c3a42d92
  - [v6,2/2] selftest/bpf/benchs: Add bpf_map benchmark
    https://git.kernel.org/bpf/bpf-next/c/89eda98428ce

You are awesome, thank you!
-- 
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html