[PATCH 10/14] tools/perf: Allocate and setup aux buffer queue to help co-relate with other events across CPU's

[Adrian Hunter]

On 15/08/2025 11:34, Athira Rajeev wrote:
> When the Dispatch Trace Log data is collected along with other events
> like sched tracepoint events, it needs to be correlated and present
> interleaved along with these events. Perf events can be collected
> parallely across the CPUs. Hence it needs to be ensured events/dtl
> entries are processed in timestamp order.
> 
> An auxtrace_queue is created for each CPU. Data within each queue is in
> increasing order of timestamp. Each auxtrace queue has a array/list of
> auxtrace buffers. When processing the auxtrace buffer, the data is
> mmapp'ed. All auxtrace queues is maintained in auxtrace heap. Each queue
> has a queue number and a timestamp. The queues are sorted/added to head
> based on the time stamp. So always the lowest timestamp (entries to be
> processed first) is on top of the heap.
> 
> The auxtrace queue needs to be allocated and heap needs to be populated
> in the sorted order of timestamp. The queue needs to be filled with data
> only once via powerpc_vpadtl__update_queues() function.
> powerpc_vpadtl__setup_queues() iterates through all the entries to
> allocate and setup the auxtrace queue. To add to auxtrace heap, it is
> required to fetch the timebase of first entry for each of the queue.
> 
> The first entry in the queue for VPA DTL PMU has the boot timebase,
> frequency details which are needed to get timestamp which is required to
> correlate with other events. The very next entry is the actual trace data
> that provides timestamp for occurrence of DTL event. Formula used to get
> the timestamp from dtl entry is:
> 
> ((timbase from DTL entry - boot time) / frequency) * 1000000000
> 
> powerpc_vpadtl_decode() adds the boot time and frequency as part of
> powerpc_vpadtl_queue structure so that it can be reused. Each of the
> dtl_entry is of 48 bytes size. Sometimes it could happen that one buffer
> is only partially processed (if the timestamp of occurrence of another
> event is more than currently processed element in queue, it will move on
> to next event). Inorder to keep track of position of buffer, additional

Inorder -> In order

> fields is added to powerpc_vpadtl_queue structure.
> 
> Signed-off-by: Athira Rajeev <atrajeev at linux.ibm.com>
> ---
>  tools/perf/util/powerpc-vpadtl.c | 219 ++++++++++++++++++++++++++++++-
>  1 file changed, 218 insertions(+), 1 deletion(-)
> 
> diff --git a/tools/perf/util/powerpc-vpadtl.c b/tools/perf/util/powerpc-vpadtl.c
> index 36c02821cf0a..299927901c9d 100644
> --- a/tools/perf/util/powerpc-vpadtl.c
> +++ b/tools/perf/util/powerpc-vpadtl.c
> @@ -28,6 +28,7 @@
>  #include "map.h"
>  #include "symbol_conf.h"
>  #include "symbol.h"
> +#include "tool.h"
>  
>  /*
>   * The DTL entries are of below format
> @@ -72,6 +73,14 @@ struct powerpc_vpadtl_queue {
>  	struct auxtrace_buffer	*buffer;
>  	struct thread		*thread;
>  	bool			on_heap;
> +	struct dtl_entry	*dtl;
> +	u64			timestamp;
> +	unsigned long		pkt_len;
> +	unsigned long		buf_len;
> +	u64			boot_tb;
> +	u64			tb_freq;
> +	unsigned int		tb_buffer;
> +	unsigned int		size;
>  	bool			done;
>  	pid_t			pid;
>  	pid_t			tid;
> @@ -151,12 +160,217 @@ static void powerpc_vpadtl_dump_event(struct powerpc_vpadtl *vpa, unsigned char
>  	powerpc_vpadtl_dump(vpa, buf, len);
>  }
>  
> +static int powerpc_vpadtl_get_buffer(struct powerpc_vpadtl_queue *vpaq)
> +{
> +	struct auxtrace_buffer *buffer = vpaq->buffer;
> +	struct auxtrace_queues *queues = &vpaq->vpa->queues;
> +	struct auxtrace_queue *queue;
> +
> +	queue = &queues->queue_array[vpaq->queue_nr];
> +	buffer = auxtrace_buffer__next(queue, buffer);
> +
> +	if (!buffer)
> +		return 0;
> +
> +	vpaq->buffer = buffer;
> +	vpaq->size = buffer->size;
> +
> +	/* If the aux_buffer doesn't have data associated, try to load it */
> +	if (!buffer->data) {
> +		/* get the file desc associated with the perf data file */
> +		int fd = perf_data__fd(vpaq->vpa->session->data);
> +
> +		buffer->data = auxtrace_buffer__get_data(buffer, fd);
> +		if (!buffer->data)
> +			return -ENOMEM;
> +	}
> +
> +	vpaq->buf_len = buffer->size;
> +
> +	if (buffer->size % dtl_entry_size)
> +		vpaq->buf_len = buffer->size - (buffer->size % dtl_entry_size);
> +
> +	if (vpaq->tb_buffer != buffer->buffer_nr) {
> +		vpaq->pkt_len = 0;
> +		vpaq->tb_buffer = 0;
> +	}
> +
> +	return 1;
> +}
> +
> +/*
> + * The first entry in the queue for VPA DTL PMU has the boot timebase,
> + * frequency details which are needed to get timestamp which is required to
> + * correlate with other events. Save the boot_tb and tb_freq as part of
> + * powerpc_vpadtl_queue. The very next entry is the actual trace data to
> + * be returned.
> + */
> +static int powerpc_vpadtl_decode(struct powerpc_vpadtl_queue *vpaq)
> +{
> +	int ret;
> +	char *buf;
> +	struct boottb_freq *boottb;
> +
> +	ret = powerpc_vpadtl_get_buffer(vpaq);
> +	if (ret <= 0)
> +		return ret;
> +
> +	boottb = (struct boottb_freq *)vpaq->buffer->data;
> +	if (boottb->timebase == 0) {
> +		vpaq->boot_tb = boottb->boot_tb;
> +		vpaq->tb_freq = boottb->tb_freq;
> +		vpaq->pkt_len += dtl_entry_size;
> +	}
> +
> +	buf = vpaq->buffer->data;
> +	buf += vpaq->pkt_len;
> +	vpaq->dtl = (struct dtl_entry *)buf;
> +
> +	vpaq->tb_buffer = vpaq->buffer->buffer_nr;
> +	vpaq->buffer = NULL;
> +	vpaq->buf_len = 0;
> +
> +	return 1;
> +}
> +
> +static struct powerpc_vpadtl_queue *powerpc_vpadtl__alloc_queue(struct powerpc_vpadtl *vpa,
> +						unsigned int queue_nr)
> +{
> +	struct powerpc_vpadtl_queue *vpaq;
> +
> +	vpaq = zalloc(sizeof(*vpaq));
> +	if (!vpaq)
> +		return NULL;
> +
> +	vpaq->vpa = vpa;
> +	vpaq->queue_nr = queue_nr;
> +
> +	return vpaq;
> +}
> +
> +/*
> + * When the Dispatch Trace Log data is collected along with other events
> + * like sched tracepoint events, it needs to be correlated and present
> + * interleaved along with these events. Perf events can be collected
> + * parallely across the CPUs.
> + *
> + * An auxtrace_queue is created for each CPU. Data within each queue is in
> + * increasing order of timestamp. Allocate and setup auxtrace queues here.
> + * All auxtrace queues is maintained in auxtrace heap in the increasing order
> + * of timestamp. So always the lowest timestamp (entries to be processed first)
> + * is on top of the heap.
> + *
> + * To add to auxtrace heap, fetch the timestamp from first DTL entry
> + * for each of the queue.
> + */
> +static int powerpc_vpadtl__setup_queue(struct powerpc_vpadtl *vpa,
> +		struct auxtrace_queue *queue,
> +		unsigned int queue_nr)
> +{
> +	struct powerpc_vpadtl_queue *vpaq = queue->priv;
> +	struct dtl_entry *record;
> +	double result, div;
> +	double boot_freq;
> +	unsigned long long boot_tb;
> +	unsigned long long diff;
> +	unsigned long long save = 0;
> +
> +	if (list_empty(&queue->head) || vpaq)
> +		return 0;
> +
> +	vpaq = powerpc_vpadtl__alloc_queue(vpa, queue_nr);
> +	if (!vpaq)
> +		return -ENOMEM;
> +
> +	queue->priv = vpaq;
> +
> +	if (queue->cpu != -1)
> +		vpaq->cpu = queue->cpu;
> +
> +	if (!vpaq->on_heap) {
> +		int ret;
> +retry:
> +		ret = powerpc_vpadtl_decode(vpaq);
> +		if (!ret)
> +			return 0;
> +
> +		if (ret < 0)
> +			goto retry;
> +
> +		record = vpaq->dtl;
> +		/*
> +		 * Formula used to get timestamp that can be co-related with
> +		 * other perf events:
> +		 * ((timbase from DTL entry - boot time) / frequency) * 1000000000
> +		 */
> +		if (record->timebase) {
> +			boot_tb = vpaq->boot_tb;
> +			boot_freq = vpaq->tb_freq;
> +			diff = be64_to_cpu(record->timebase) - boot_tb;
> +			div = diff / boot_freq;
> +			result = div;
> +			result = result * 1000000000;
> +			save = result;

It would be nicer for the time calculation to be in a separate function.
Also 'save' is an odd choice of variable name for a timestamp.


> +		}
> +
> +		vpaq->timestamp = save;
> +		ret = auxtrace_heap__add(&vpa->heap, queue_nr, vpaq->timestamp);
> +		if (ret)
> +			return ret;
> +		vpaq->on_heap = true;
> +	}
> +
> +	return 0;
> +}
> +
> +static int powerpc_vpadtl__setup_queues(struct powerpc_vpadtl *vpa)
> +{
> +	unsigned int i;
> +	int ret;
> +
> +	for (i = 0; i < vpa->queues.nr_queues; i++) {
> +		ret = powerpc_vpadtl__setup_queue(vpa, &vpa->queues.queue_array[i], i);
> +		if (ret)
> +			return ret;
> +	}
> +
> +	return 0;
> +}
> +
> +static int powerpc_vpadtl__update_queues(struct powerpc_vpadtl *vpa)
> +{
> +	if (vpa->queues.new_data) {
> +		vpa->queues.new_data = false;
> +		return powerpc_vpadtl__setup_queues(vpa);
> +	}
> +
> +	return 0;
> +}
> +
>  static int powerpc_vpadtl_process_event(struct perf_session *session __maybe_unused,
>  				 union perf_event *event __maybe_unused,
>  				 struct perf_sample *sample __maybe_unused,
>  				 const struct perf_tool *tool __maybe_unused)

tool, sample and session are not __maybe_unused

>  {
> -	return 0;
> +	int err = 0;
> +	struct powerpc_vpadtl *vpa = container_of(session->auxtrace,
> +			struct powerpc_vpadtl, auxtrace);

Arranging local variable declarations in order of descending line
length is often more readable

	struct powerpc_vpadtl *vpa = session_to_vpa(session);
	int err = 0;

> +
> +	if (dump_trace)
> +		return 0;
> +
> +	if (!tool->ordered_events) {
> +		pr_err("VPA requires ordered events\n");
> +		return -EINVAL;
> +	}
> +
> +	if (sample->time) {
> +		err = powerpc_vpadtl__update_queues(vpa);
> +		if (err)
> +			return err;
> +	}
> +
> +	return err;
>  }
>  
>  /*
> @@ -181,6 +395,9 @@ static int powerpc_vpadtl_process_auxtrace_event(struct perf_session *session,
>  			return -errno;
>  	}
>  
> +	if (!dump_trace)
> +		return 0;

See comment about auxtrace_queues__add_event() in patch 8.

> +
>  	err = auxtrace_queues__add_event(&vpa->queues, session, event,
>  					 data_offset, &buffer);
>  	if (err)