Re: [PATCH v4] perf bench: Add atomic operations benchmark

[Namhyung Kim <namhyung@kernel.org>]

On Tue, Oct 13, 2026 at 04:56:48PM +0800, Changbin Du wrote:
> Add a new 'atomic' collection to perf bench for benchmarking
> atomic read-modify-write operations with multi-threaded
> contention testing. It provides two subcommands:
> 
>   - atomic cas: contended compare-and-swap operations
>     (__atomic_compare_exchange_n)
>   - atomic inc: contended fetch-add operations (__atomic_fetch_add)
> 
> Both support a configurable thread count and runtime to measure
> atomic contention effects.
> 
> Why this benchmark is needed:
> - Atomic read-modify-write operations are fundamental to lock-free
>   algorithms and data structures. Understanding their performance
>   characteristics under contention is critical for designing
>   high-performance concurrent applications.
> - The benchmark helps identify atomic operation latency and
>   scalability issues across different thread counts, revealing
>   contention patterns that are not visible in single-threaded tests.
> - Useful for evaluating atomic implementation quality on different
>   architectures and for regression testing after changes to atomic
>   primitives or memory ordering.
> 
> Measurement methodology:
> - By default two threads contend for the counter, enough to expose
>   the contention effect; use -t to scale the count.
> - Each thread starts a private timer (clock_gettime CLOCK_MONOTONIC)
>   after synchronizing on a pthread_barrier, ensuring all threads
>   begin simultaneously.
> - Each thread performs a hot loop of the selected atomic operation
>   (compare-and-swap or fetch-add) on a shared unsigned long counter
>   for a fixed runtime. It counts its own completed operations in a
>   local variable, and the loop runs in batches so the termination
>   flag is only polled between batches.
> - The main thread sleeps for the runtime and then sets the flag; the
>   throughput of a repeat is the total operation count divided by the
>   measured wall time (the max per-thread runtime).
> - The shared counter is cache-line aligned (64 bytes) to isolate
>   contention to the target cache line and avoid false sharing.
> - The per-thread operation counts of the last repeat show whether
>   any thread failed to make progress under contention.
> - The first repeat is excluded from statistics as a warmup phase
>   to avoid cache-cold effects.
> 
> Example usage:
>   $ perf bench atomic all
>   # Running atomic/cas benchmark...
>     Threads: 2, runtime: 2 sec, repeats: 10 (warmup: 1)
>     Avg throughput: 68,599,954 ops/sec (+- 0.13%)
>     Per-thread ops (last repeat):
>       fastest:  74,842,112 ops
>       slowest:  63,176,704 ops
>       avg:      69,009,408 ops
> 
>   # Running atomic/inc benchmark...
>     Threads: 2, runtime: 2 sec, repeats: 10 (warmup: 1)
>     Avg throughput: 128,428,772 ops/sec (+- 0.54%)
>     Per-thread ops (last repeat):
>       fastest:  146,230,272 ops
>       slowest:  112,825,344 ops
>       avg:      129,527,808 ops
> 
> Output fields explained:
> - Threads: number of contending threads
> - runtime: seconds each thread runs per repeat
> - repeats: number of test runs (first is warmup)
> - Avg throughput: mean aggregate ops/sec across repeats
> - +- %: relative standard error of the mean throughput across repeats
> - Per-thread ops: fastest/slowest/avg operations per thread in the
>   last repeat (shows scheduling imbalance and lack of progress)
> 
> Assisted-by: opencode:DeepSeek-V4-Pro
> Signed-off-by: Changbin Du <changbin.du@gmail.com>
> ---
> Changes in v4:
> - Default to 2 threads (1 on a single-CPU machine).
> - Replace the iteration count with a fixed runtime per repeat.
> - Report per-thread operation counts to show lack of progress.
> 
> Changes in v3:
> - Use native word size (unsigned long) for the shared counter instead
>   of u64 to avoid libatomic on 32-bit targets.
> - Include <stdint.h> explicitly for UINT64_MAX.
> 
> Changes in v2:
> - Add an 'atomic inc' subcommand for contended __atomic_fetch_add.
> - Hoist the load out of the CAS retry loop; the failed
>   __atomic_compare_exchange_n already returns the current value.
> - Cast the seconds delta to u64 before multiplying by NSEC_PER_SEC to
>   avoid 32-bit overflow.
> ---
>  tools/perf/Documentation/perf-bench.txt |  38 +++
>  tools/perf/bench/Build                  |   1 +
>  tools/perf/bench/atomic.c               | 295 ++++++++++++++++++++++++
>  tools/perf/bench/bench.h                |   2 +
>  tools/perf/builtin-bench.c              |   9 +
>  5 files changed, 345 insertions(+)
>  create mode 100644 tools/perf/bench/atomic.c
> 
> diff --git a/tools/perf/Documentation/perf-bench.txt b/tools/perf/Documentation/perf-bench.txt
> index c5913cf59c98..2a1206fea0c9 100644
> --- a/tools/perf/Documentation/perf-bench.txt
> +++ b/tools/perf/Documentation/perf-bench.txt
> @@ -58,6 +58,9 @@ SUBSYSTEM
>  'numa'::
>  	NUMA scheduling and MM benchmarks.
>  
> +'atomic'::
> +	Atomic operation benchmarks.
> +
>  'futex'::
>  	Futex stressing benchmarks.
>  
> @@ -283,6 +286,41 @@ SUITES FOR 'numa'
>  *mem*::
>  Suite for evaluating NUMA workloads.
>  
> +SUITES FOR 'atomic'
> +~~~~~~~~~~~~~~~~~~~
> +*cas*::
> +Suite for evaluating compare-and-swap (CAS) atomic operations under
> +multi-threaded contention.
> +
> +*inc*::
> +Suite for evaluating atomic fetch-add (increment) operations under
> +multi-threaded contention.
> +
> +Options of *cas* and *inc*
> +^^^^^^^^^^^^^^^^^^^^^^^^^^
> +-t::
> +--threads=::
> +Number of threads contending for the shared counter (default: 2, or
> +1 on a single-CPU machine).
> +
> +-r::
> +--runtime=::
> +Runtime of each repeat in seconds (default: 2).
> +
> +The per-thread fastest/slowest/avg ops summary is only printed when
> +running with more than one thread.
> +
> +The average throughput is annotated with its relative deviation
> +("+- x%"), i.e. the standard error of the mean over the measured
> +repeats. A small value means the average is reproducible; it does not
> +reflect per-thread fairness, which the per-thread ops summary shows.
> +
> +Note that the cas benchmark counts successful compare-and-swap
> +operations only; retries and the loads between them are additional
> +atomic instructions not included in the ops/sec count, so cas and
> +inc throughputs are not directly comparable at the instruction
> +level.
> +
>  SUITES FOR 'futex'
>  ~~~~~~~~~~~~~~~~~~
>  *hash*::
> diff --git a/tools/perf/bench/Build b/tools/perf/bench/Build
> index 67b76fe20ba6..c64c52468d3d 100644
> --- a/tools/perf/bench/Build
> +++ b/tools/perf/bench/Build
> @@ -3,6 +3,7 @@ perf-bench-y += sched-pipe.o
>  perf-bench-y += sched-seccomp-notify.o
>  perf-bench-y += syscall.o
>  perf-bench-y += mem-functions.o
> +perf-bench-y += atomic.o
>  perf-bench-y += futex.o
>  perf-bench-y += futex-hash.o
>  perf-bench-y += futex-wake.o
> diff --git a/tools/perf/bench/atomic.c b/tools/perf/bench/atomic.c
> new file mode 100644
> index 000000000000..2da74bf078bc
> --- /dev/null
> +++ b/tools/perf/bench/atomic.c
> @@ -0,0 +1,295 @@
> +// SPDX-License-Identifier: GPL-2.0
> +/*
> + * Copyright (C) 2026 Changbin Du <changbin.du@gmail.com>
> + *
> + * Benchmarks for atomic read-modify-write operations.
> + *
> + * Measures throughput (ops/sec) of contended atomic operations on a
> + * single shared counter across multiple threads. Each thread performs
> + * a hot loop of the selected operation -- compare-and-swap (cas) or
> + * fetch-add (inc) -- competing with all other threads for the same
> + * cache line. Threads run for a fixed runtime, so a system that
> + * cannot make progress under heavy contention reports a low (or
> + * zero) throughput instead of hanging.
> + */
> +#include "bench.h"
> +#include <linux/compiler.h>
> +#include "../util/debug.h"
> +
> +#ifndef HAVE_PTHREAD_BARRIER
> +int bench_atomic_cas(int argc __maybe_unused, const char **argv __maybe_unused)
> +{
> +	pr_err("%s: pthread_barrier_t unavailable, disabling this test...\n", __func__);
> +	return 0;
> +}
> +
> +int bench_atomic_inc(int argc __maybe_unused, const char **argv __maybe_unused)
> +{
> +	pr_err("%s: pthread_barrier_t unavailable, disabling this test...\n", __func__);
> +	return 0;
> +}
> +#else /* HAVE_PTHREAD_BARRIER */
> +#include <stdlib.h>
> +#include <stdio.h>
> +#include <stdbool.h>
> +#include <stdint.h>
> +#include <unistd.h>
> +#include <pthread.h>
> +#include <time.h>
> +#include <inttypes.h>
> +#include <err.h>
> +#include <linux/types.h>
> +#include <linux/kernel.h>
> +#include <linux/time64.h>
> +#include <perf/cpumap.h>
> +#include <subcmd/parse-options.h>
> +#include "../util/stat.h"
> +
> +static unsigned int threads;		/* 0: default to min(2, online CPUs) */
> +static unsigned int runtime	= 2;	/* seconds per repeat */
> +
> +static const struct option options[] = {
> +	OPT_UINTEGER('t', "threads", &threads,
> +		     "Number of contending threads (default: 2, or 1 on single CPU)"),
> +	OPT_UINTEGER('r', "runtime", &runtime,

Can you please pick a different short option than 'r'?  perf bench
itself has -r/--repeat option so it would be confusing.  It doesn't
clash for now, but I think it's better to use perf bench options from
'perf bench atomic ...' later.

Otherwise, looks good to me.

Thanks,
Namhyung


> +		     "Runtime of each repeat in seconds (default: 2)"),
> +	OPT_END()
> +};
> +
> +static const char * const cas_usage[] = {
> +	"perf bench atomic cas <options>",
> +	NULL
> +};
> +
> +static const char * const inc_usage[] = {
> +	"perf bench atomic inc <options>",
> +	NULL
> +};
> +
> +enum atomic_op {
> +	OP_CAS,
> +	OP_INC,
> +};
> +
> +static int bench_op = OP_CAS;
> +
> +/*
> + * The termination flag is polled between batches only, so that the
> + * check does not perturb the measured loop.
> + */
> +#define BATCH_OPS	1024
> +
> +static unsigned long shared_counter __aligned(64);
> +static pthread_barrier_t start_barrier __aligned(64);
> +static bool done;
> +
> +struct worker_stats {
> +	u64	runtime_ns;
> +	u64	ops;
> +};
> +
> +static void cas_batch(void)
> +{
> +	unsigned long i;
> +
> +	for (i = 0; i < BATCH_OPS; i++) {
> +		unsigned long old_val, new_val;
> +
> +		old_val = __atomic_load_n(&shared_counter, __ATOMIC_RELAXED);
> +		do {
> +			new_val = old_val + 1;
> +		} while (!__atomic_compare_exchange_n(&shared_counter, &old_val,
> +						      new_val, false,
> +						      __ATOMIC_RELAXED,
> +						      __ATOMIC_RELAXED));
> +	}
> +}
> +
> +static void inc_batch(void)
> +{
> +	unsigned long i;
> +
> +	for (i = 0; i < BATCH_OPS; i++)
> +		__atomic_fetch_add(&shared_counter, 1, __ATOMIC_RELAXED);
> +}
> +
> +static void *worker(void *arg)
> +{
> +	struct worker_stats *ws = (struct worker_stats *)arg;
> +	struct timespec tstart, tend;
> +	u64 ops = 0;
> +
> +	pthread_barrier_wait(&start_barrier);
> +
> +	clock_gettime(CLOCK_MONOTONIC, &tstart);
> +	while (!__atomic_load_n(&done, __ATOMIC_RELAXED)) {
> +		switch (bench_op) {
> +		case OP_CAS:
> +			cas_batch();
> +			break;
> +		case OP_INC:
> +			inc_batch();
> +			break;
> +		default:
> +			BUG_ON(1);
> +			break;
> +		}
> +		ops += BATCH_OPS;
> +	}
> +	clock_gettime(CLOCK_MONOTONIC, &tend);
> +
> +	ws->runtime_ns	= (u64)(tend.tv_sec - tstart.tv_sec) * NSEC_PER_SEC +
> +			  (tend.tv_nsec - tstart.tv_nsec);
> +	ws->ops		= ops;
> +
> +	return NULL;
> +}
> +
> +static void print_default_format(struct worker_stats *wstats,
> +				 struct stats *throughput_stats)
> +{
> +	double thru_avg, thru_stddev;
> +
> +	printf("  Threads: %u, runtime: %u sec, repeats: %u (warmup: 1)\n",
> +	       threads, runtime, bench_repeat);
> +
> +	thru_avg    = avg_stats(throughput_stats);
> +	thru_stddev = stddev_stats(throughput_stats);
> +
> +	printf("  Avg throughput: %'.0f ops/sec (+- %.2f%%)\n",
> +	       thru_avg, rel_stddev_stats(thru_stddev, thru_avg));
> +
> +	/*
> +	 * Per-thread operation counts from the last measured repeat show
> +	 * whether any thread failed to make progress under contention.
> +	 */
> +	if (threads > 1) {
> +		u64 min_ops = UINT64_MAX, max_ops = 0, total_ops = 0;
> +
> +		for (unsigned int t = 0; t < threads; t++) {
> +			if (wstats[t].ops < min_ops)
> +				min_ops = wstats[t].ops;
> +			if (wstats[t].ops > max_ops)
> +				max_ops = wstats[t].ops;
> +			total_ops += wstats[t].ops;
> +		}
> +
> +		printf("  Per-thread ops (last repeat):\n");
> +		printf("    fastest:  %'" PRIu64 " ops\n", max_ops);
> +		printf("    slowest:  %'" PRIu64 " ops\n", min_ops);
> +		printf("    avg:      %'" PRIu64 " ops\n", total_ops / threads);
> +	}
> +}
> +
> +static int run_benchmark(void)
> +{
> +	pthread_t *thread_ids;
> +	struct worker_stats *wstats;
> +	struct stats throughput_stats;
> +	unsigned int r, t;
> +
> +	thread_ids = calloc(threads, sizeof(pthread_t));
> +	if (!thread_ids)
> +		err(EXIT_FAILURE, "calloc");
> +
> +	wstats = calloc(threads, sizeof(struct worker_stats));
> +	if (!wstats)
> +		err(EXIT_FAILURE, "calloc");
> +
> +	init_stats(&throughput_stats);
> +
> +	for (r = 0; r < bench_repeat + 1; r++) {
> +		u64 max_runtime_ns = 0, total_ops = 0;
> +		double throughput;
> +
> +		shared_counter = 0;
> +		__atomic_store_n(&done, false, __ATOMIC_RELAXED);
> +		pthread_barrier_init(&start_barrier, NULL, threads);
> +
> +		for (t = 0; t < threads; t++) {
> +			if (pthread_create(&thread_ids[t], NULL,
> +					   worker, &wstats[t]))
> +				err(EXIT_FAILURE, "pthread_create");
> +		}
> +
> +		sleep(runtime);
> +		__atomic_store_n(&done, true, __ATOMIC_RELAXED);
> +
> +		for (t = 0; t < threads; t++)
> +			pthread_join(thread_ids[t], NULL);
> +
> +		pthread_barrier_destroy(&start_barrier);
> +
> +		for (t = 0; t < threads; t++) {
> +			total_ops += wstats[t].ops;
> +			if (wstats[t].runtime_ns > max_runtime_ns)
> +				max_runtime_ns = wstats[t].runtime_ns;
> +		}
> +
> +		/*
> +		 * Exclude the first repeat (warmup) from statistics so
> +		 * that cache-cold and lazy-init effects are not counted.
> +		 */
> +		if (r > 0 && max_runtime_ns) {
> +			throughput = (double)total_ops /
> +				     ((double)max_runtime_ns / NSEC_PER_SEC);
> +			update_stats(&throughput_stats, (u64)throughput);
> +		}
> +	}
> +
> +	switch (bench_format) {
> +	case BENCH_FORMAT_DEFAULT:
> +		print_default_format(wstats, &throughput_stats);
> +		break;
> +
> +	case BENCH_FORMAT_SIMPLE:
> +		printf("%.0f\n", avg_stats(&throughput_stats));
> +		break;
> +
> +	default:
> +		fprintf(stderr, "Unknown format: %d\n", bench_format);
> +		exit(EXIT_FAILURE);
> +	}
> +
> +	free(thread_ids);
> +	free(wstats);
> +
> +	return 0;
> +}
> +
> +static int run_atomic_bench(int argc, const char **argv,
> +			    const char * const *usage)
> +{
> +	struct perf_cpu_map *cpu;
> +
> +	if (parse_options(argc, argv, options, usage, 0)) {
> +		usage_with_options(usage, options);
> +		exit(EXIT_FAILURE);
> +	}
> +
> +	if (!threads) {	/* default to 2 threads (1 on a single-CPU machine) */
> +		cpu = perf_cpu_map__new_online_cpus();
> +		threads = (cpu && perf_cpu_map__nr(cpu) > 1) ? 2 : 1;
> +		perf_cpu_map__put(cpu);
> +	}
> +
> +	if (runtime < 1) {
> +		fprintf(stderr, "Invalid runtime: %u\n", runtime);
> +		return 1;
> +	}
> +
> +	return run_benchmark();
> +}
> +
> +int bench_atomic_cas(int argc, const char **argv)
> +{
> +	bench_op = OP_CAS;
> +	return run_atomic_bench(argc, argv, cas_usage);
> +}
> +
> +int bench_atomic_inc(int argc, const char **argv)
> +{
> +	bench_op = OP_INC;
> +	return run_atomic_bench(argc, argv, inc_usage);
> +}
> +#endif /* HAVE_PTHREAD_BARRIER */
> diff --git a/tools/perf/bench/bench.h b/tools/perf/bench/bench.h
> index 8519eb5a42fa..85e3ac471875 100644
> --- a/tools/perf/bench/bench.h
> +++ b/tools/perf/bench/bench.h
> @@ -30,6 +30,8 @@ int bench_mem_memcpy(int argc, const char **argv);
>  int bench_mem_memset(int argc, const char **argv);
>  int bench_mem_mmap(int argc, const char **argv);
>  int bench_mem_find_bit(int argc, const char **argv);
> +int bench_atomic_cas(int argc, const char **argv);
> +int bench_atomic_inc(int argc, const char **argv);
>  int bench_futex_hash(int argc, const char **argv);
>  int bench_futex_wake(int argc, const char **argv);
>  int bench_futex_wake_parallel(int argc, const char **argv);
> diff --git a/tools/perf/builtin-bench.c b/tools/perf/builtin-bench.c
> index 02d47913cc6a..62bebdedd3b8 100644
> --- a/tools/perf/builtin-bench.c
> +++ b/tools/perf/builtin-bench.c
> @@ -14,6 +14,7 @@
>   *  syscall ... System call performance
>   *  mem   ... memory access performance
>   *  numa  ... NUMA scheduling and MM performance
> + *  atomic ... Atomic operation performance
>   *  futex ... Futex performance
>   *  epoll ... Event poll performance
>   */
> @@ -115,6 +116,13 @@ static const struct bench uprobe_benchmarks[] = {
>  	{ NULL,	NULL, NULL },
>  };
>  
> +static const struct bench atomic_benchmarks[] = {
> +	{ "cas",	"Benchmark CAS (compare-and-swap) operations",	bench_atomic_cas	},
> +	{ "inc",	"Benchmark atomic fetch-add operations",	bench_atomic_inc	},
> +	{ "all",	"Run all atomic benchmarks",			NULL			},
> +	{ NULL,		NULL,						NULL			}
> +};
> +
>  struct collection {
>  	const char		*name;
>  	const char		*summary;
> @@ -128,6 +136,7 @@ static const struct collection collections[] = {
>  #ifdef HAVE_LIBNUMA_SUPPORT
>  	{ "numa",	"NUMA scheduling and MM benchmarks",		numa_benchmarks		},
>  #endif
> +	{ "atomic",	"Atomic operation benchmarks",			atomic_benchmarks	},
>  	{"futex",       "Futex stressing benchmarks",                   futex_benchmarks        },
>  #ifdef HAVE_EVENTFD_SUPPORT
>  	{"epoll",       "Epoll stressing benchmarks",                   epoll_benchmarks        },
> -- 
> 2.53.0
>