Re: [RFC PATCH v1 24/37] perf evsel: handle precise_ip fallback in evsel__open_cpu

[Arnaldo Carvalho de Melo <acme@kernel.org>]

Em Sat, Aug 21, 2021 at 11:19:30AM +0200, Riccardo Mancini escreveu:
> This is another patch in the effort to separate the fallback mechanisms
> from the open itself.
> 
> In case of precise_ip fallback, the original precise_ip will be stored
> in the evsel (it was stored in a local variable) and the open will be
> retried. Since the precise_ip fallback will be the first in the chain of
> fallbacks, there should be no functional change with this patch.

Thanks, looks ok, applied.

- Arnaldo

 
> Signed-off-by: Riccardo Mancini <rickyman7@gmail.com>
> ---
>  tools/perf/util/evsel.c | 59 ++++++++++++++++++-----------------------
>  tools/perf/util/evsel.h |  2 ++
>  2 files changed, 28 insertions(+), 33 deletions(-)
> 
> diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
> index 3e556afed8dd396c..2e95416b8320c6b9 100644
> --- a/tools/perf/util/evsel.c
> +++ b/tools/perf/util/evsel.c
> @@ -1709,42 +1709,29 @@ static void display_attr(struct perf_event_attr *attr)
>  	}
>  }
>  
> -static int perf_event_open(struct evsel *evsel,
> -			   pid_t pid, int cpu, int group_fd)
> +bool evsel__precise_ip_fallback(struct evsel *evsel)
>  {
> -	int precise_ip = evsel->core.attr.precise_ip;
> -	int fd;
> -
> -	while (1) {
> -		pr_debug2_peo("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
> -			  pid, cpu, group_fd, evsel->open_flags);
> -
> -		fd = sys_perf_event_open(&evsel->core.attr, pid, cpu, group_fd, evsel->open_flags);
> -		if (fd >= 0)
> -			break;
> -
> -		/* Do not try less precise if not requested. */
> -		if (!evsel->precise_max)
> -			break;
> -
> -		/*
> -		 * We tried all the precise_ip values, and it's
> -		 * still failing, so leave it to standard fallback.
> -		 */
> -		if (!evsel->core.attr.precise_ip) {
> -			evsel->core.attr.precise_ip = precise_ip;
> -			break;
> -		}
> +	/* Do not try less precise if not requested. */
> +	if (!evsel->precise_max)
> +		return false;
>  
> -		pr_debug2_peo("\nsys_perf_event_open failed, error %d\n", -ENOTSUP);
> -		evsel->core.attr.precise_ip--;
> -		pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
> -		display_attr(&evsel->core.attr);
> +	/*
> +	 * We tried all the precise_ip values, and it's
> +	 * still failing, so leave it to standard fallback.
> +	 */
> +	if (!evsel->core.attr.precise_ip) {
> +		evsel->core.attr.precise_ip = evsel->precise_ip_original;
> +		return false;
>  	}
>  
> -	return fd;
> -}
> +	if (!evsel->precise_ip_original)
> +		evsel->precise_ip_original = evsel->core.attr.precise_ip;
>  
> +	evsel->core.attr.precise_ip--;
> +	pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
> +	display_attr(&evsel->core.attr);
> +	return true;
> +}
>  
>  static struct perf_cpu_map *empty_cpu_map;
>  static struct perf_thread_map *empty_thread_map;
> @@ -2004,8 +1991,11 @@ static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
>  
>  			test_attr__ready();
>  
> -			fd = perf_event_open(evsel, pid, cpus->map[cpu],
> -					     group_fd);
> +			pr_debug2_peo("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
> +				pid, cpus->map[cpu], group_fd, evsel->open_flags);
> +
> +			fd = sys_perf_event_open(&evsel->core.attr, pid, cpus->map[cpu],
> +						group_fd, evsel->open_flags);
>  
>  			FD(evsel, cpu, thread) = fd;
>  
> @@ -2058,6 +2048,9 @@ static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
>  	return 0;
>  
>  try_fallback:
> +	if (evsel__precise_ip_fallback(evsel))
> +		goto retry_open;
> +
>  	if (evsel__ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
>  		/* We just removed 1 thread, so lower the upper nthreads limit. */
>  		nthreads--;
> diff --git a/tools/perf/util/evsel.h b/tools/perf/util/evsel.h
> index 661d2677a19b6248..0a245afab2d87d74 100644
> --- a/tools/perf/util/evsel.h
> +++ b/tools/perf/util/evsel.h
> @@ -150,6 +150,7 @@ struct evsel {
>  		struct bperf_follower_bpf *follower_skel;
>  	};
>  	unsigned long		open_flags;
> +	int			precise_ip_original;
>  };
>  
>  struct perf_missing_features {
> @@ -297,6 +298,7 @@ bool evsel__ignore_missing_thread(struct evsel *evsel,
>  				  int nr_cpus, int cpu,
>  				  struct perf_thread_map *threads,
>  				  int thread, int err);
> +bool evsel__precise_ip_fallback(struct evsel *evsel);
>  
>  struct perf_sample;
>  
> -- 
> 2.31.1

-- 

- Arnaldo