On Wed, Aug 04, 2021 at 03:29:54PM +0800, liuqi (BA) wrote:

Hi Will,

quoted

Hmm, I was hoping that you would expose all the events as proper perf_events
and get rid of the subevents entirely.

Then userspace could do things like:

   // Count number of RX memory reads
   $ perf stat -e hisi_pcie0_0/rx_memory_read/

   // Count delay cycles
   $ perf stat -e hisi_pcie0_0/latency/

   // Count both of the above (events must be in the same group)
   $ perf stat -g -e hisi_pcie0_0/latency/ -e hisi_pcie0_0/rx_memory_read/

Note that in all three of these cases the hardware will be programmed in
the same way and both HISI_PCIE_CNT and HISI_PCIE_EXT_CNT are allocated!

So for example, doing this (i.e. without the '-g'):

   $ perf stat -e hisi_pcie0_0/latency/ -e hisi_pcie0_0/rx_memory_read/

would fail because the first event would allocate both of the counters.

I'm confused with this situation when getting rid of subevent:

$ perf stat -e hisi_pcie0_0/latency/ -e hisi_pcie0_0/rx_memory_read/

In this case, driver checks the relationship of "latency" and
"rx_memory_read" in pmu->add() function and return a -EINVAL, but this seems
lead to time division multiplexing.

	if (event->pmu->add(event, PERF_EF_START)) {
		perf_event_set_state(event, PERF_EVENT_STATE_INACTIVE);
		event->oncpu = -1;
		ret = -EAGAIN;
		goto out;
	}
	...
out:
	perf_pmu_enable(event->pmu);

This result doesn't meet our expection, do I miss something here?

This is how perf works. If you don't want multiplexing, put the events in a
group. What's the problem with that?

Will