Re: [RFC PATCH v1 21/37] perf evsel: move ignore_missing_thread to fallback code

[Arnaldo Carvalho de Melo]

Em Sat, Aug 21, 2021 at 11:19:27AM +0200, Riccardo Mancini escreveu:
> This patch moves ignore_missing_thread outside the perf_event_open loop.
> Doing so, we need to move the retry_open flag a few places higher, with
> minimal impact. Furthermore, thread need not be decreased since it won't
> get increased by the for loop (since we're jumping back inside), but we
> need to check that the nthreads decrease didn't put thread out of range.
> 
> The goal is to have fallbacks handled in one place only, since in the
> future parallel code, these would be handled separately.

Thanks, looks ok, applied.

- Arnaldo

 
> Signed-off-by: Riccardo Mancini <rickyman7@xxxxxxxxx>
> ---
>  tools/perf/util/evsel.c | 29 +++++++++++++----------------
>  tools/perf/util/evsel.h |  5 +++++
>  2 files changed, 18 insertions(+), 16 deletions(-)
> 
> diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
> index 916930ea31450265..a1a88607fd59efcb 100644
> --- a/tools/perf/util/evsel.c
> +++ b/tools/perf/util/evsel.c
> @@ -1656,7 +1656,7 @@ static int update_fds(struct evsel *evsel,
>  	return 0;
>  }
>  
> -static bool ignore_missing_thread(struct evsel *evsel,
> +bool evsel__ignore_missing_thread(struct evsel *evsel,
>  				  int nr_cpus, int cpu,
>  				  struct perf_thread_map *threads,
>  				  int thread, int err)
> @@ -1993,12 +1993,15 @@ static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
>  
>  		for (thread = 0; thread < nthreads; thread++) {
>  			int fd, group_fd;
> +retry_open:
> +			if (thread >= nthreads)
> +				break;
>  
>  			if (!evsel->cgrp && !evsel->core.system_wide)
>  				pid = perf_thread_map__pid(threads, thread);
>  
>  			group_fd = get_group_fd(evsel, cpu, thread);
> -retry_open:
> +
>  			test_attr__ready();
>  
>  			fd = perf_event_open(evsel, pid, cpus->map[cpu],
> @@ -2016,20 +2019,6 @@ static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
>  			if (fd < 0) {
>  				err = -errno;
>  
> -				if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
> -					/*
> -					 * We just removed 1 thread, so take a step
> -					 * back on thread index and lower the upper
> -					 * nthreads limit.
> -					 */
> -					nthreads--;
> -					thread--;
> -
> -					/* ... and pretend like nothing have happened. */
> -					err = 0;
> -					continue;
> -				}
> -
>  				pr_debug2_peo("\nsys_perf_event_open failed, error %d\n",
>  					  err);
>  				goto try_fallback;
> @@ -2069,6 +2058,14 @@ static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
>  	return 0;
>  
>  try_fallback:
> +	if (evsel__ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
> +		/* We just removed 1 thread, so lower the upper nthreads limit. */
> +		nthreads--;
> +
> +		/* ... and pretend like nothing have happened. */
> +		err = 0;
> +		goto retry_open;
> +	}
>  	/*
>  	 * perf stat needs between 5 and 22 fds per CPU. When we run out
>  	 * of them try to increase the limits.
> diff --git a/tools/perf/util/evsel.h b/tools/perf/util/evsel.h
> index bf9abd9a5cbf9852..661d2677a19b6248 100644
> --- a/tools/perf/util/evsel.h
> +++ b/tools/perf/util/evsel.h
> @@ -293,6 +293,11 @@ bool evsel__detect_missing_features(struct evsel *evsel);
>  enum rlimit_action { NO_CHANGE, SET_TO_MAX, INCREASED_MAX };
>  bool evsel__increase_rlimit(enum rlimit_action *set_rlimit);
>  
> +bool evsel__ignore_missing_thread(struct evsel *evsel,
> +				  int nr_cpus, int cpu,
> +				  struct perf_thread_map *threads,
> +				  int thread, int err);
> +
>  struct perf_sample;
>  
>  void *evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, const char *name);
> -- 
> 2.31.1

-- 

- Arnaldo