Re: [RFC PATCH v4 1/1] perf cs-etm: Split Coresight decode by aux records

[Leo Yan]

Hi James,

On Tue, Jun 01, 2021 at 04:07:51PM +0300, James Clark wrote:
> Populate the auxtrace queues using AUX records rather than whole
> auxtrace buffers.
> 
> This is similar to the auxtrace_queues__process_index() ->
> auxtrace_queues__add_indexed_event() flow where
> perf_session__peek_event() is used to read AUXTRACE events out of random
> positions in the file based on the auxtrace index.
> 
> Now there is a second pass using perf_session__peek_events() which loops
> over all PERF_RECORD_AUX events, finds the corresponding auxtrace
> buffer, and adds a fragment of that buffer to the auxtrace queues,
> rather than the whole buffer.
> 
> This is a completely new implementation compared to v1 and v2 of the
> patchset where significant modifications were made to the decoding flow,
> including saving all of the AUX records, and changing the point where
> decoding started. These changes had effects on the outcome of decoding,
> but with this implementation, the output of decoding is identical to
> perf/core, except in the case where decoding failed due to the issue
> that this change addresses. No changes to decoding were made, apart
> from populating the auxtrace queues.

IMHO, this patch is much better than previous versions! :)

When went through the whole patch, the most code actually is general
enough and can consider to place into into util/auxtrace.c.  If there
have any objection, we can come back to add the code for only cs-etm
specific.

> Because the decoder is already reset between each auxtrace buffer,
> populating the queues with fragments of whole buffers based on AUX
> records has the same affect as resetting the decoder on the AUX record
> bounds during decode. But this approach only requires a change to one
> part of the whole flow, rather than many parts.
> 
> There is still one TODO about how buffers are found. Currently there is
> a plain loop as a search, but this could be improved with a binary
> search or keeping a record of which buffers have already been added and
> starting the search from the last used place. AUX records and buffers
> will come in similar orders in the file so the latter would be simpler
> and probably be good enough. But I'm looking for feedback about the
> approach as a whole before ironing out all of the details.
> 
> Other issues that were fixed by the previous patchsets such as fixing
> the case where the first aux record comes before the first MMAP record
> can now be fixed as completely independent issues.
> 
> Testing
> =======
> 
> Testing was done with the following script, to diff the decoding results
> between the patched and un-patched versions of perf:
> 
> 	#!/bin/bash
> 	set -ex
> 
> 	$1 script -i $3 $4 > split.script
> 	$2 script -i $3 $4 > default.script
> 
> 	diff split.script default.script | head -n 20
> 
> And it was run like this, with various itrace options depending on the
> quantity of synthesised events:
> 
> 	compare.sh ./perf-patched ./perf-default perf-per-cpu-2-threads.data --itrace=i100000ns
> 
> No changes in output were observed in the following scenarios:
> 
> * Simple per-cpu
> 	perf record -e cs_etm/@tmc_etr0/u top
> 
> * Per-thread, single thread
> 	perf record -e cs_etm/@tmc_etr0/u --per-thread ./threads_C
> 
> * Per-thread multiple threads (but only one thread collected data):
> 	perf record -e cs_etm/@tmc_etr0/u --per-thread --pid 4596,4597
> 
> * Per-thread multiple threads (both threads collected data):
> 	perf record -e cs_etm/@tmc_etr0/u --per-thread --pid 4596,4597
> 
> * Per-cpu explicit threads:
> 	perf record -e cs_etm/@tmc_etr0/u --pid 853,854
> 
> * System-wide (per-cpu):
>     perf record -e cs_etm/@tmc_etr0/u -a
> 
> * No data collected (no aux buffers)
> 	Can happen with any command when run for a short period
> 
> * Containing truncated records
> 	Can happen with any command
> 
> * Containing aux records with 0 size
> 	Can happen with any command
> 
> * Snapshot mode
> 	perf record -e cs_etm/@tmc_etr0/u -a --snapshot
> 
> Signed-off-by: James Clark <james.clark@xxxxxxx>
> ---
>  tools/perf/util/cs-etm.c | 159 ++++++++++++++++++++++++++++++++++++++-
>  1 file changed, 156 insertions(+), 3 deletions(-)
> 
> diff --git a/tools/perf/util/cs-etm.c b/tools/perf/util/cs-etm.c
> index 64536a6ed10a..053aad4b28cf 100644
> --- a/tools/perf/util/cs-etm.c
> +++ b/tools/perf/util/cs-etm.c
> @@ -2679,6 +2679,161 @@ static u64 *cs_etm__create_meta_blk(u64 *buff_in, int *buff_in_offset,
>  	return metadata;
>  }
>  
> +/**
> + * Puts a fragment of an auxtrace buffer into the auxtrace queues based
> + * on the bounds of aux_event, if it matches with the buffer that's at
> + * file_offset.
> + *
> + * Normally, whole auxtrace buffers would be added to the queue. But we
> + * want to reset the decoder for every PERF_RECORD_AUX event, and the decoder
> + * is reset across each buffer, so splitting the buffers up in advance has
> + * the same effect.
> + */
> +static int cs_etm__queue_aux_fragment(struct perf_session *session,
> +						off_t file_offset, size_t sz,
> +						struct perf_record_aux *aux_event,
> +						struct perf_sample *sample)
> +{
> +	int err;
> +	char buf[PERF_SAMPLE_MAX_SIZE];
> +	union perf_event *auxtrace_event_union;
> +	struct perf_record_auxtrace *auxtrace_event;
> +	union perf_event auxtrace_fragment;
> +	bool matchesCpuPid;

It's better to avoid using camel naming.

> +	__u64 aux_offset;
> +	struct cs_etm_auxtrace *etm = container_of(session->auxtrace,
> +						   struct cs_etm_auxtrace,
> +						   auxtrace);
> +
> +	/*
> +	 * There should be a PERF_RECORD_AUXTRACE event at the file_offset that we got
> +	 * from looping through the auxtrace index.
> +	 */
> +	err = perf_session__peek_event(session, file_offset, buf,
> +				       PERF_SAMPLE_MAX_SIZE, &auxtrace_event_union, NULL);
> +	if (err)
> +		return err;
> +	auxtrace_event = &auxtrace_event_union->auxtrace;
> +	if (auxtrace_event->header.type != PERF_RECORD_AUXTRACE)
> +		return -EINVAL;
> +
> +	if (auxtrace_event->header.size < sizeof(struct perf_record_auxtrace) ||
> +		auxtrace_event->header.size != sz) {
> +		return -EINVAL;
> +	}
> +
> +	/*
> +	 * In per-thread mode, CPU is set to -1, but TID will be set instead.
> +	 * See auxtrace_mmap_params__set_idx()
> +	 */
> +	if (auxtrace_event->cpu == (__u32) -1)
> +		matchesCpuPid = auxtrace_event->tid == sample->tid;
> +	else
> +		matchesCpuPid = auxtrace_event->cpu == sample->cpu;

If "matchesCpuPid" is false, can directly bail out at here, so can
avoid the below unnecessary calculations.

> +
> +	/*
> +	 * In snapshot/overwrite mode, the head points to the end of the buffer so aux_offset needs
> +	 * to have the size subtracted so it points to the beginning as in normal mode.
> +	 */
> +	if (aux_event->flags & PERF_AUX_FLAG_OVERWRITE)
> +		aux_offset = aux_event->aux_offset - aux_event->aux_size;
> +	else
> +		aux_offset = aux_event->aux_offset;
> +
> +	if (matchesCpuPid &&
> +	    aux_offset >= auxtrace_event->offset &&
> +	    aux_offset + aux_event->aux_size <= auxtrace_event->offset + auxtrace_event->size) {

It's possible that an event PERF_RECORD_AUXTRACE saves AUX trace data
for multiple events PERF_RECORD_AUX (if we consider watermark).

So here the logic makes sense for me for checking the buffer range and
I don't find any issue.

> +		/*
> +		 * If this AUX event was inside this buffer somewhere, create a new auxtrace event
> +		 * based on the sizes of the aux event, and queue that fragment.
> +		 */
> +		auxtrace_fragment.auxtrace = *auxtrace_event;
> +		auxtrace_fragment.auxtrace.size = aux_event->aux_size;
> +		auxtrace_fragment.auxtrace.offset = aux_offset;
> +		file_offset += aux_offset - auxtrace_event->offset + auxtrace_event->header.size;
> +		return auxtrace_queues__add_event(&etm->queues,
> +					       session,
> +					       &auxtrace_fragment,
> +					       file_offset,
> +					       NULL);
> +	}
> +
> +	/* Wasn't inside this buffer, but there were no parse errors. 1 == 'not found' */
> +	return 1;
> +}
> +
> +static int cs_etm__queue_aux_records_cb(struct perf_session *session,
> +			     union perf_event *event, u64 offset __maybe_unused,
> +			     void *data __maybe_unused)
> +{
> +	struct perf_sample sample;
> +	int ret;
> +	struct auxtrace_index_entry *ent;
> +	struct auxtrace_index *auxtrace_index;
> +	struct evsel *evsel;
> +	size_t i;
> +	struct cs_etm_auxtrace *etm = container_of(session->auxtrace,
> +						   struct cs_etm_auxtrace,
> +						   auxtrace);
> +
> +	/* Don't care about any other events, we're only queuing buffers for AUX events */
> +	if (event->header.type != PERF_RECORD_AUX)
> +		return 0;
> +
> +	if (event->header.size < sizeof(struct perf_record_aux))
> +		return -EINVAL;
> +
> +	/* Truncated Aux records can have 0 size and shouldn't result in anything being queued. */
> +	if (!event->aux.aux_size)
> +		return 0;
> +
> +	/*
> +	 * Parse the sample, we need the sample_id_all data that comes after the event so that the
> +	 * CPU or PID can be matched to an AUXTRACE buffer's CPU or PID.
> +	 */
> +	evsel = evlist__event2evsel(session->evlist, event);
> +	evsel__parse_sample(evsel, event, &sample);

It's good to check the return errors from evsel__parse_sample().

> +
> +	/*
> +	 * Loop throuch the auxtrace index to find the buffer that matches up with this
> +	 * aux event.
> +	 * TODO: binary search?

For binary search, you might need to reorg the auxtrace_index and
its entries, I personally think this is not the priority.

> +	 */
> +	list_for_each_entry(auxtrace_index, &session->auxtrace_index, list) {
> +		for (i = 0; i < auxtrace_index->nr; i++) {
> +			ent = &auxtrace_index->entries[i];
> +			ret = cs_etm__queue_aux_fragment(session, ent->file_offset,
> +							 ent->sz, &event->aux, &sample);
> +			if (!ret) {
> +				etm->data_queued = true;
> +				return 0;
> +			} else if (ret < 0) {
> +				/*
> +				 * Anything other than 1 is an error. Positive values are 'not
> +				 * found' we just want to go onto the next one in that case.
> +				 */
> +				return ret;
> +			}
> +		}
> +	}

It's good to add a warning at here?  If cannot find a matched
AUXRECORD event and fails to queue the fragment, it's likely caused by
the AUX ring buffer's overrun.

> +
> +	/*
> +	 * We would get here if there are no entries in the index (either no auxtrace buffers or
> +	 * no index). Fail silently as there is the possibility of queueing them in
> +	 * cs_etm__process_auxtrace_event() if etm->data_queued is still false;
> +	 *
> +	 * In that scenario, buffers will not be split by AUX records.
> +	 */
> +	return 0;
> +}
> +
> +static int cs_etm__queue_aux_records(struct perf_session *session)
> +{
> +	return perf_session__peek_events(session, session->header.data_offset,
> +					 session->header.data_size,
> +					 cs_etm__queue_aux_records_cb, NULL);
> +}
> +
>  int cs_etm__process_auxtrace_info(union perf_event *event,
>  				  struct perf_session *session)
>  {
> @@ -2879,12 +3034,10 @@ int cs_etm__process_auxtrace_info(union perf_event *event,
>  	if (err)
>  		goto err_delete_thread;
>  
> -	err = auxtrace_queues__process_index(&etm->queues, session);
> +	err = cs_etm__queue_aux_records(session);
>  	if (err)
>  		goto err_delete_thread;
>  
> -	etm->data_queued = etm->queues.populated;

Seems to me it's no reason to remove this sentence.

"etm->queues.populated" will be set when call
auxtrace_queues__add_event(), so we still can assign it to
"etm->data_queued" at here.

Thanks,
Leo

> -
>  	return 0;
>  
>  err_delete_thread:
> -- 
> 2.28.0
>