On 31/08/23 10:18, Dixit, Ashutosh wrote:

Hi Ashutosh,
On Tue, 29 Aug 2023 22:15:44 -0700, Aravind Iddamsetty wrote:

      
Hi Aravind,

diff --git a/drivers/gpu/drm/xe/xe_pmu.c b/drivers/gpu/drm/xe/xe_pmu.c
new file mode 100644
index 000000000000..41dd422812ff
--- /dev/null
+++ b/drivers/gpu/drm/xe/xe_pmu.c
@@ -0,0 +1,679 @@
+// SPDX-License-Identifier: MIT
+/*
+ * Copyright © 2023 Intel Corporation
+ */
+
+#include <drm/drm_drv.h>
+#include <drm/drm_managed.h>
+#include <drm/xe_drm.h>
+
+#include "regs/xe_gt_regs.h"
+#include "xe_device.h"
+#include "xe_gt_clock.h"
+#include "xe_mmio.h"
+
+static cpumask_t xe_pmu_cpumask;
+static unsigned int xe_pmu_target_cpu = -1;
+
+static unsigned int config_gt_id(const u64 config)
+{
+	return config >> __XE_PMU_GT_SHIFT;
+}
+
+static u64 config_counter(const u64 config)
+{
+	return config & ~(~0ULL << __XE_PMU_GT_SHIFT);
+}
+
+static int engine_busyness_sample_type(u64 config)
+{
+	int type = 0;
Why initialize? The switch statement should have a default with a BUG/WARN_ON
below? Also see the comment below.

+
+	switch (config) {
+	case XE_PMU_RENDER_GROUP_BUSY(0):
+		type = __XE_SAMPLE_RENDER_GROUP_BUSY;
+		break;
+	case XE_PMU_COPY_GROUP_BUSY(0):
+		type = __XE_SAMPLE_COPY_GROUP_BUSY;
+		break;
+	case XE_PMU_MEDIA_GROUP_BUSY(0):
+		type = __XE_SAMPLE_MEDIA_GROUP_BUSY;
+		break;
+	case XE_PMU_ANY_ENGINE_GROUP_BUSY(0):
+		type = __XE_SAMPLE_ANY_ENGINE_GROUP_BUSY;
+		break;
+	}
+
+	return type;
+}
I am thinking this function is not really needed. We can just do:

	int sample_type = config - 1;

or

	int sample_type = config_counter(config) - 1;
It might not always be true in future, the configs can start from any range.

in engine_group_busyness_read? See comment at __xe_pmu_event_read below.

+
+static void xe_pmu_event_destroy(struct perf_event *event)
+{
+	struct xe_device *xe =
+		container_of(event->pmu, typeof(*xe), pmu.base);
+
+	drm_WARN_ON(&xe->drm, event->parent);
+
+	drm_dev_put(&xe->drm);
+}
+
+static u64 __engine_group_busyness_read(struct xe_gt *gt, int sample_type)
+{
+	u64 val = 0;
+
+	switch (sample_type) {
+	case __XE_SAMPLE_RENDER_GROUP_BUSY:
+		val = xe_mmio_read32(gt, XE_OAG_RENDER_BUSY_FREE);
+		break;
+	case __XE_SAMPLE_COPY_GROUP_BUSY:
+		val = xe_mmio_read32(gt, XE_OAG_BLT_BUSY_FREE);
+		break;
+	case __XE_SAMPLE_MEDIA_GROUP_BUSY:
+		val = xe_mmio_read32(gt, XE_OAG_ANY_MEDIA_FF_BUSY_FREE);
+		break;
+	case __XE_SAMPLE_ANY_ENGINE_GROUP_BUSY:
+		val = xe_mmio_read32(gt, XE_OAG_RC0_ANY_ENGINE_BUSY_FREE);
+		break;
+	default:
+		drm_warn(&gt->tile->xe->drm, "unknown pmu event\n");
+	}
+
+	return xe_gt_clock_cycles_to_ns(gt, val * 16);
+}
+
+static u64 engine_group_busyness_read(struct xe_gt *gt, u64 config)
+{
+	int sample_type = engine_busyness_sample_type(config);
If config is event->attr.config, this can just be 'config_counter(config) - 1'.
See comment at __xe_pmu_event_read below.

+	const unsigned int gt_id = gt->info.id;
+	struct xe_device *xe = gt->tile->xe;
+	struct xe_pmu *pmu = &xe->pmu;
+	unsigned long flags;
+	bool device_awake;
+	u64 val;
+
+	device_awake = xe_device_mem_access_get_if_ongoing(xe);
+	if (device_awake) {
+		XE_WARN_ON(xe_force_wake_get(gt_to_fw(gt), XE_FW_GT));
+		val = __engine_group_busyness_read(gt, sample_type);
+		XE_WARN_ON(xe_force_wake_put(gt_to_fw(gt), XE_FW_GT));
+		xe_device_mem_access_put(xe);
+	}
+
+	spin_lock_irqsave(&pmu->lock, flags);
+
+	if (device_awake)
+		pmu->sample[gt_id][sample_type] = val;
+	else
+		val = pmu->sample[gt_id][sample_type];
+
+	spin_unlock_irqrestore(&pmu->lock, flags);
+
+	return val;
+}
+
+static void engine_group_busyness_store(struct xe_gt *gt)
+{
+	struct xe_pmu *pmu = &gt->tile->xe->pmu;
+	unsigned int gt_id = gt->info.id;
+	unsigned long flags;
+	int i;
+
+	spin_lock_irqsave(&pmu->lock, flags);
+
+	for (i = __XE_SAMPLE_RENDER_GROUP_BUSY; i <= __XE_SAMPLE_ANY_ENGINE_GROUP_BUSY; i++) {
+		pmu->sample[gt_id][i] = __engine_group_busyness_read(gt, i);
This is not quite right. At the minimum we need to take forcewake
here. Also since this is called in both suspend and runtime_suspend code
paths we might also need to the take the runtime_pm reference.
The pm reference and forcewake are already taken in suspend paths hence didn't
add here again as this is called only from those paths.

check xe_gt_suspend.

I think the simplest might be to just construct 'config'
(event->attr.config) here and call engine_group_busyness_read? Something
like:

	for (i = __XE_SAMPLE_RENDER_GROUP_BUSY; i <= __XE_SAMPLE_ANY_ENGINE_GROUP_BUSY; i++) {
		config = ; // Construct config using gt_id and i
		engine_group_busyness_read(gt, i);
	}

This will automatically save the read values in pmu->sample[][] if the
device is awake. Thoughts?
I think this is best kept separate from usual read paths(which are atomic) didn't want to club them.
Also because forcewakes and pm reference are taken separately in suspend path.

+	}
+
+	spin_unlock_irqrestore(&pmu->lock, flags);
+}
+
+static int
+config_status(struct xe_device *xe, u64 config)
+{
+	unsigned int max_gt_id = xe->info.gt_count > 1 ? 1 : 0;
What is this for? See below.
reminiscent of my previous code, will clean it up.

+	unsigned int gt_id = config_gt_id(config);
+	struct xe_gt *gt = xe_device_get_gt(xe, gt_id);
+
+	if (gt_id > max_gt_id)
Maybe this can just be:

	if (gt_id >= XE_PMU_MAX_GT)

+		return -ENOENT;
+
+	switch (config_counter(config)) {
+	case XE_PMU_INTERRUPTS(0):
+		if (gt_id)
+			return -ENOENT;
+		break;
+	case XE_PMU_RENDER_GROUP_BUSY(0):
+	case XE_PMU_COPY_GROUP_BUSY(0):
+	case XE_PMU_ANY_ENGINE_GROUP_BUSY(0):
+		if (gt->info.type == XE_GT_TYPE_MEDIA)
+			return -ENOENT;
+		break;
+	case XE_PMU_MEDIA_GROUP_BUSY(0):
+		if (!(gt->info.engine_mask & (BIT(XE_HW_ENGINE_VCS0) | BIT(XE_HW_ENGINE_VECS0))))
+			return -ENOENT;
+		break;
+	default:
+		return -ENOENT;
+	}
+
+	return 0;
+}
+
+static int xe_pmu_event_init(struct perf_event *event)
+{
+	struct xe_device *xe =
+		container_of(event->pmu, typeof(*xe), pmu.base);
+	struct xe_pmu *pmu = &xe->pmu;
+	int ret;
+
+	if (pmu->closed)
+		return -ENODEV;
+
+	if (event->attr.type != event->pmu->type)
+		return -ENOENT;
+
+	/* unsupported modes and filters */
+	if (event->attr.sample_period) /* no sampling */
+		return -EINVAL;
+
+	if (has_branch_stack(event))
+		return -EOPNOTSUPP;
+
+	if (event->cpu < 0)
+		return -EINVAL;
+
+	/* only allow running on one cpu at a time */
+	if (!cpumask_test_cpu(event->cpu, &xe_pmu_cpumask))
+		return -EINVAL;
+
+	ret = config_status(xe, event->attr.config);
+	if (ret)
+		return ret;
+
+	if (!event->parent) {
+		drm_dev_get(&xe->drm);
+		event->destroy = xe_pmu_event_destroy;
+	}
+
+	return 0;
+}
+
+static u64 __xe_pmu_event_read(struct perf_event *event)
+{
+	struct xe_device *xe =
+		container_of(event->pmu, typeof(*xe), pmu.base);
+	const unsigned int gt_id = config_gt_id(event->attr.config);
+	const u64 config = config_counter(event->attr.config);
Probably nit but this config being different from event->attr.config is
confusing. Let's use 'event->attr.config' throughout as argument to
functions and use config_counter() to get rid of gt_id. So get rid of this
config variable.

+	struct xe_gt *gt = xe_device_get_gt(xe, gt_id);
+	struct xe_pmu *pmu = &xe->pmu;
+	u64 val = 0;
+
+	switch (config) {
switch (config_counter(event->attr.config))

+	case XE_PMU_INTERRUPTS(0):
+		val = READ_ONCE(pmu->irq_count);
+		break;
+	case XE_PMU_RENDER_GROUP_BUSY(0):
+	case XE_PMU_COPY_GROUP_BUSY(0):
+	case XE_PMU_ANY_ENGINE_GROUP_BUSY(0):
+	case XE_PMU_MEDIA_GROUP_BUSY(0):
+		val = engine_group_busyness_read(gt, config);
engine_group_busyness_read(gt, event->attr.config);
hmmm ok.

Also, need a default case.

+	}
+
+	return val;
+}
+
+static void xe_pmu_event_read(struct perf_event *event)
+{
+	struct xe_device *xe =
+		container_of(event->pmu, typeof(*xe), pmu.base);
+	struct hw_perf_event *hwc = &event->hw;
+	struct xe_pmu *pmu = &xe->pmu;
+	u64 prev, new;
+
+	if (pmu->closed) {
+		event->hw.state = PERF_HES_STOPPED;
+		return;
+	}
+again:
+	prev = local64_read(&hwc->prev_count);
+	new = __xe_pmu_event_read(event);
+
+	if (local64_cmpxchg(&hwc->prev_count, prev, new) != prev)
+		goto again;
+
+	local64_add(new - prev, &event->count);
+}
+
+static void xe_pmu_enable(struct perf_event *event)
+{
+	/*
+	 * Store the current counter value so we can report the correct delta
+	 * for all listeners. Even when the event was already enabled and has
+	 * an existing non-zero value.
+	 */
+	local64_set(&event->hw.prev_count, __xe_pmu_event_read(event));
Right now nothing is being enabled here (unlike i915) so the function name
xe_pmu_enable looks weird. Not sure, maybe leave as is for when things get
added in the future?

+static int xe_pmu_cpu_online(unsigned int cpu, struct hlist_node *node)
+{
+	struct xe_pmu *pmu = hlist_entry_safe(node, typeof(*pmu), cpuhp.node);
+
+	BUG_ON(!pmu->base.event_init);
+
+	/* Select the first online CPU as a designated reader. */
+	if (cpumask_empty(&xe_pmu_cpumask))
+		cpumask_set_cpu(cpu, &xe_pmu_cpumask);
+
+	return 0;
+}
+
+static int xe_pmu_cpu_offline(unsigned int cpu, struct hlist_node *node)
+{
+	struct xe_pmu *pmu = hlist_entry_safe(node, typeof(*pmu), cpuhp.node);
+	unsigned int target = xe_pmu_target_cpu;
+
+	BUG_ON(!pmu->base.event_init);
nit but wondering if we should remove these two BUG_ON's (and save a couple
of checkpatch warnings even though the BUG_ON's are in i915) and just do
something like:

	  if (!pmu->base.event_init)
		return 0;

The reason for the BUG_ON's seems to be that these functions can be called
after module_init but before probe.

xe_pmu_cpu_online() doesn't depend on pmu at all so looks like the BUG_ON
can just be dropped?
the  xe_pmu_cpu_online/offline are not invoked when they are registered with
cpuhp_setup_state_multi, but rather when cpuhp_state_add_instance() is called
which is done post the PMU is initialized hence the check for BUG_ON.

Thanks,
Aravind.


Thanks.
--
Ashutosh