Re: [PATCH bpf-next] selftests/bpf: Add benchmark for local_storage RCU Tasks Trace usage

[Martin KaFai Lau <kafai@fb.com>]

On Thu, Jun 23, 2022 at 04:46:09PM -0700, Dave Marchevsky wrote:
> +static void report_progress(int iter, struct bench_res *res, long delta_ns)
> +{
> +	if (ctx.skel->bss->unexpected) {
> +		fprintf(stderr, "Error: Unexpected order of bpf prog calls (postgp after pregp).");
> +		fprintf(stderr, "Data can't be trusted, exiting\n");
> +		exit(1);
> +	}
> +
> +	if (args.quiet)
> +		return;
> +
> +	printf("Iter %d\t avg tasks_trace grace period latency\t%lf ns\n",
> +	       iter, res->gp_ns / (double)res->gp_ct);
> +	printf("Iter %d\t avg ticks per tasks_trace grace period\t%lf\n",
> +	       iter, res->stime / (double)res->gp_ct);
> +}
> +

[ ... ]

> diff --git a/tools/testing/selftests/bpf/benchs/run_bench_local_storage_rcu_tasks_trace.sh b/tools/testing/selftests/bpf/benchs/run_bench_local_storage_rcu_tasks_trace.sh
> new file mode 100755
> index 000000000000..5dac1f02892c
> --- /dev/null
> +++ b/tools/testing/selftests/bpf/benchs/run_bench_local_storage_rcu_tasks_trace.sh
> @@ -0,0 +1,11 @@
> +#!/bin/bash
> +# SPDX-License-Identifier: GPL-2.0
> +
> +kthread_pid=`pgrep rcu_tasks_trace_kthread`
> +
> +if [ -z $kthread_pid ]; then
> +	echo "error: Couldn't find rcu_tasks_trace_kthread"
> +	exit 1
> +fi
> +
> +./bench --nr_procs 15000 --kthread_pid $kthread_pid -d 600 --quiet 1 local-storage-tasks-trace
> diff --git a/tools/testing/selftests/bpf/progs/local_storage_rcu_tasks_trace_bench.c b/tools/testing/selftests/bpf/progs/local_storage_rcu_tasks_trace_bench.c
> new file mode 100644
> index 000000000000..9b11342b19a0
> --- /dev/null
> +++ b/tools/testing/selftests/bpf/progs/local_storage_rcu_tasks_trace_bench.c
> @@ -0,0 +1,65 @@
> +// SPDX-License-Identifier: GPL-2.0
> +/* Copyright (c) 2022 Meta Platforms, Inc. and affiliates. */
> +
> +#include "vmlinux.h"
> +#include <bpf/bpf_helpers.h>
> +#include "bpf_misc.h"
> +
> +struct {
> +	__uint(type, BPF_MAP_TYPE_TASK_STORAGE);
> +	__uint(map_flags, BPF_F_NO_PREALLOC);
> +	__type(key, int);
> +	__type(value, int);
> +} task_storage SEC(".maps");
> +
> +long hits;
> +long gp_hits;
> +long gp_times;
> +long current_gp_start;
> +long unexpected;
> +
> +SEC("fentry/" SYS_PREFIX "sys_getpgid")
> +int get_local(void *ctx)
> +{
> +	struct task_struct *task;
> +	int idx;
> +	int *s;
> +
> +	idx = 0;
> +	task = bpf_get_current_task_btf();
> +	s = bpf_task_storage_get(&task_storage, task, &idx,
> +				 BPF_LOCAL_STORAGE_GET_F_CREATE);
> +	if (!s)
> +		return 0;
> +
> +	*s = 3;
> +	bpf_task_storage_delete(&task_storage, task);
> +	__sync_add_and_fetch(&hits, 1);
> +	return 0;
> +}
> +
> +SEC("kprobe/rcu_tasks_trace_pregp_step")
nit.  Similar to the fentry sys_getpgid above.
may as well use fentry for everything.

> +int pregp_step(struct pt_regs *ctx)
> +{
> +	current_gp_start = bpf_ktime_get_ns();
> +	return 0;
> +}
> +
> +SEC("kprobe/rcu_tasks_trace_postgp")
> +int postgp(struct pt_regs *ctx)
> +{
> +	if (!current_gp_start) {
> +		/* Will only happen if prog tracing rcu_tasks_trace_pregp_step doesn't
> +		 * execute before this prog
> +		 */
> +		__sync_add_and_fetch(&unexpected, 1);
I consistently hit this:
./bench --nr_procs 1500 --kthread_pid ... -d 60 --quiet 1 local-storage-tasks-trace
Setting up benchmark 'local-storage-tasks-trace'...
Spun up 1500 procs (our pid 28351)
Benchmark 'local-storage-tasks-trace' started.
Error: Unexpected order of bpf prog calls (postgp after pregp).Data can't be trusted, exiting

May be there is a chance for the very first postgp being called
before the pregp_step?

Thanks for working on this!