Commit e5a5886d authored Jan 23, 2016 by Sukadev Bhattiprolu Committed by Michael Ellerman Mar 10, 2016

powerpc/perf/hv-24x7: Fix usage with chip events.



24x7 counters can belong to different domains (core, chip, virtual CPU
etc). For events in the 'chip' domain, sysfs entry currently looks like:

	$ cd /sys/bus/event_source/devices/hv_24x7/events
	$ cat PM_XLINK_CYCLES__PHYS_CHIP
	domain=0x1,offset=0x230,core=?,lpar=0x0

where the required parameter, 'core=?' is specified with perf as:

	perf stat -C 0 -e hv_24x7/PM_XLINK_CYCLES__PHYS_CHIP,core=1/ \
		/bin/true

This is inconsistent in that 'core' is a required parameter for a chip
event.  Instead, have the the sysfs entry display 'chip=?' for chip
events:

	$ cd /sys/bus/event_source/devices/hv_24x7/events
	$ cat PM_XLINK_CYCLES__PHYS_CHIP
	domain=0x1,offset=0x230,chip=?,lpar=0x0

We also need to add a 'chip' entry in the sysfs format directory:

	$ ls /sys/bus/event_source/devices/hv_24x7/format
	chip  core  domain  lpar  offset  vcpu
	^^^^
	(new)

so the perf tool can automatically check usage and format the chip
parameter correctly:

	$ perf stat -C 0 -v -e hv_24x7/PM_XLINK_CYCLES__PHYS_CHIP/ \
		/bin/true
	Required parameter 'chip' not specified
	invalid or unsupported event: 'hv_24x7/PM_XLINK_CYCLES__PHYS_CHIP/'

	$ perf stat -C 0 -v -e hv_24x7/PM_XLINK_CYCLES__PHYS_CHIP,chip=1/ \
		/bin/true
	hv_24x7/PM_XLINK_CYCLES__PHYS_CHIP,chip=1/: 0 6628908 6628908

	 Performance counter stats for 'CPU(s) 0':

	         0      hv_24x7/PM_XLINK_CYCLES__PHYS_CHIP,chip=1/

	    0.006606970 seconds time elapsed

Signed-off-by: Sukadev Bhattiprolu <sukadev@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

parent e0728b50

arch/powerpc/perf/hv-24x7.c

+18 −4

Original line number	Diff line number	Diff line
		@@ -101,6 +101,7 @@ static bool catalog_entry_domain_is_valid(unsigned domain)
		EVENT_DEFINE_RANGE_FORMAT(domain, config, 0, 3);
		/* u16 */
		EVENT_DEFINE_RANGE_FORMAT(core, config, 16, 31);
		EVENT_DEFINE_RANGE_FORMAT(chip, config, 16, 31);
		EVENT_DEFINE_RANGE_FORMAT(vcpu, config, 16, 31);
		/* u32, see "data_offset" */
		EVENT_DEFINE_RANGE_FORMAT(offset, config, 32, 63);
		@@ -115,6 +116,7 @@ static struct attribute *format_attrs[] = {
		&format_attr_domain.attr,
		&format_attr_offset.attr,
		&format_attr_core.attr,
		&format_attr_chip.attr,
		&format_attr_vcpu.attr,
		&format_attr_lpar.attr,
		NULL,
		@@ -289,10 +291,16 @@ static char event_fmt(struct hv_24x7_event_data event, unsigned domain)
		const char *sindex;
		const char *lpar;

		if (is_physical_domain(domain)) {
		switch (domain) {
		case HV_PERF_DOMAIN_PHYS_CHIP:
		lpar = "0x0";
		sindex = "chip";
		break;
		case HV_PERF_DOMAIN_PHYS_CORE:
		lpar = "0x0";
		sindex = "core";
		} else {
		break;
		default:
		lpar = "?";
		sindex = "vcpu";
		}
		@@ -1089,10 +1097,16 @@ static int add_event_to_24x7_request(struct perf_event *event,
		return -EINVAL;
		}

		if (is_physical_domain(event_get_domain(event)))
		switch (event_get_domain(event)) {
		case HV_PERF_DOMAIN_PHYS_CHIP:
		idx = event_get_chip(event);
		break;
		case HV_PERF_DOMAIN_PHYS_CORE:
		idx = event_get_core(event);
		else
		break;
		default:
		idx = event_get_vcpu(event);
		}

		i = request_buffer->num_requests++;
		req = &request_buffer->requests[i];