[PATCH] coresight: Don't immediately close events that are run on invalid CPU/sink combos

[Mathieu Poirier]

On Tue, Sep 21, 2021 at 02:02:31PM +0100, James Clark wrote:
> When a traced process runs on a CPU that can't reach the selected sink,
> the event will be stopped with PERF_HES_STOPPED. This means that even if
> the process migrates to a valid CPU, tracing will not resume.
> 
> This can be reproduced (on N1SDP) by using taskset to start the process
> on CPU 0, and then switching it to CPU 2 (ETF 1 is only reachable from
> CPU 2):
> 
>   taskset --cpu-list 0 ./perf record -e cs_etm/@tmc_etf1/ --per-thread -- taskset --cpu-list 2 ls
> 
> This produces a single 0 length AUX record, and then no more trace:
> 
>   0x3c8 [0x30]: PERF_RECORD_AUX offset: 0 size: 0 flags: 0x1 [T]
> 
> After the fix, the same command produces normal AUX records. The perf
> self test "89: Check Arm CoreSight trace data recording and synthesized
> samples" no longer fails intermittently. This was because the taskset in
> the test is after the fork, so there is a period where the task is
> scheduled on a random CPU rather than forced to a valid one.
> 
> Specifically selecting an invalid CPU will still result in a failure to
> open the event because it will never produce trace:
> 
>   ./perf record -C 2 -e cs_etm/@tmc_etf0/
>   failed to mmap with 12 (Cannot allocate memory)
> 
> The only scenario that has changed is if the CPU mask has a valid CPU
> sink combo in it.
> 
> Testing
> =======
> 
> * Coresight self test passes consistently:
>   ./perf test Coresight
> 
> * CPU wide mode still produces trace:
>   ./perf record -e cs_etm// -a
> 
> * Invalid -C options still fail to open:
>   ./perf record -C 2,3 -e cs_etm/@tmc_etf0/
>   failed to mmap with 12 (Cannot allocate memory)
> 
> * Migrating a task to a valid sink/CPU now produces trace:
>   taskset --cpu-list 0 ./perf record -e cs_etm/@tmc_etf1/ --per-thread -- taskset --cpu-list 2 ls
> 
> * If the task remains on an invalid CPU, no trace is emitted:
>   taskset --cpu-list 0 ./perf record -e cs_etm/@tmc_etf1/ --per-thread -- ls
> 
> Signed-off-by: James Clark <james.clark at arm.com>
> ---
>  .../hwtracing/coresight/coresight-etm-perf.c  | 27 +++++++++++++++----
>  1 file changed, 22 insertions(+), 5 deletions(-)

Very interesting corner case - and I like your solution.  Arnaldo, please
consider.

Reviewed-by: Mathieu Poirier <mathieu.poirier at linaro.org>

> 
> diff --git a/drivers/hwtracing/coresight/coresight-etm-perf.c b/drivers/hwtracing/coresight/coresight-etm-perf.c
> index 8ebd728d3a80..79346f0f0e0b 100644
> --- a/drivers/hwtracing/coresight/coresight-etm-perf.c
> +++ b/drivers/hwtracing/coresight/coresight-etm-perf.c
> @@ -452,9 +452,14 @@ static void etm_event_start(struct perf_event *event, int flags)
>  	 * sink from this ETM. We can't do much in this case if
>  	 * the sink was specified or hinted to the driver. For
>  	 * now, simply don't record anything on this ETM.
> +	 *
> +	 * As such we pretend that everything is fine, and let
> +	 * it continue without actually tracing. The event could
> +	 * continue tracing when it moves to a CPU where it is
> +	 * reachable to a sink.
>  	 */
>  	if (!cpumask_test_cpu(cpu, &event_data->mask))
> -		goto fail_end_stop;
> +		goto out;
>  
>  	path = etm_event_cpu_path(event_data, cpu);
>  	/* We need a sink, no need to continue without one */
> @@ -466,16 +471,15 @@ static void etm_event_start(struct perf_event *event, int flags)
>  	if (coresight_enable_path(path, CS_MODE_PERF, handle))
>  		goto fail_end_stop;
>  
> -	/* Tell the perf core the event is alive */
> -	event->hw.state = 0;
> -
>  	/* Finally enable the tracer */
>  	if (source_ops(csdev)->enable(csdev, event, CS_MODE_PERF))
>  		goto fail_disable_path;
>  
> +out:
> +	/* Tell the perf core the event is alive */
> +	event->hw.state = 0;
>  	/* Save the event_data for this ETM */
>  	ctxt->event_data = event_data;
> -out:
>  	return;
>  
>  fail_disable_path:
> @@ -517,6 +521,19 @@ static void etm_event_stop(struct perf_event *event, int mode)
>  	if (WARN_ON(!event_data))
>  		return;
>  
> +	/*
> +	 * Check if this ETM was allowed to trace, as decided at
> +	 * etm_setup_aux(). If it wasn't allowed to trace, then
> +	 * nothing needs to be torn down other than outputting a
> +	 * zero sized record.
> +	 */
> +	if (handle->event && (mode & PERF_EF_UPDATE) &&
> +	    !cpumask_test_cpu(cpu, &event_data->mask)) {
> +		event->hw.state = PERF_HES_STOPPED;
> +		perf_aux_output_end(handle, 0);
> +		return;
> +	}
> +
>  	if (!csdev)
>  		return;
>  
> -- 
> 2.28.0
>