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

[Suzuki K Poulose]

On 21/09/2021 14:02, 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(-)
> 
> 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;

Hi James,

I think this problematic. The "out" label is reached from the "failed"
path. So, we would reset the STOPPED state. I think we should simply
return when we reach the "fail" label.

Please could you resend this patch with the changes.

Kind regards
Suzuki