Re: [PATCH 4/4] perf: arm_cspmu: ampere_cspmu: Add support for Ampere SoC PMU

On Wed, 21 Jun 2023 18:11:41 -0700
Ilkka Koskinen [off-list ref] wrote:

Ampere SoC PMU follows CoreSight PMU architecture. It uses implementation
specific registers to filter events rather than PMEVFILTnR registers.

Signed-off-by: Ilkka Koskinen <redacted>

Hi Ilkka,

Drive by review so not super detailed (I was curious) but a few questions/comments inline.

Jonathan

+
+Example for perf tool use::
+
+  / # perf list ampere
+
+    ampere_mcu_pmu_0/act_sent/                         [Kernel PMU event]
+    <...>
+    ampere_mcu_pmu_1/rd_sent/                          [Kernel PMU event]
+    <...>
+
+  / # perf stat -a -e ampere_mcu_pmu_0/act_sent,filter_enable=3,bank=5,rank=3,threshold=2/,ampere_mcu_pmu_1/rd_sent/ \
+        sleep 1

Why filter_enable=3?

+static u32 ampere_cspmu_event_filter(const struct perf_event *event)
+{

Whilst lots of other comments on this - perhaps add another one here to
why this is a noop.

+	return 0;
+}
+
+static void ampere_cspmu_set_ev_filter(struct arm_cspmu *cspmu,
+				       struct hw_perf_event *hwc,
+				       u32 filter)
+{
+	struct perf_event *event;
+	unsigned int idx;
+	u32 threshold, rank, bank;
+
+	/*
+	 * At this point, all the events have the same filter settings.
+	 * Therefore, take the first event and use its configuration.
+	 */
+	idx = find_first_bit(cspmu->hw_events.used_ctrs,
+			     cspmu->cycle_counter_logical_idx);
+
+	event = cspmu->hw_events.events[idx];
+
+	threshold	= get_threshold(event);
+	rank		= get_rank(event);
+	bank		= get_bank(event);
+
+	writel(threshold, cspmu->base0 + PMAUXR0);
+	writel(rank, cspmu->base0 + PMAUXR1);
+	writel(bank, cspmu->base0 + PMAUXR2);
+}
+
+static int ampere_cspmu_validate_configs(struct perf_event *event,
+					 struct perf_event *event2)
+{
+	if (get_threshold(event) != get_threshold(event2) ||
+	    get_rank(event) != get_rank(event2) ||
+	    get_bank(event) != get_bank(event2))
+		return -EINVAL;
+
+	return 0;
+}
+
+static int ampere_cspmu_validate_event(struct arm_cspmu *cspmu,
+				       struct perf_event *new)
+{
+	struct perf_event *curr, *leader = new->group_leader;
+	unsigned int idx;
+	int ret;
+
+	ret = ampere_cspmu_validate_configs(new, leader);
+	if (ret)
+		return ret;
+
+	/* We compare the global filter settings to existing events */
+	idx = find_first_bit(cspmu->hw_events.used_ctrs,
+			     cspmu->cycle_counter_logical_idx);
+
+	/* This is the first event */

Maybe add why that matters to the comment?

+	if (idx == cspmu->cycle_counter_logical_idx)
+		return 0;
+
+	curr = cspmu->hw_events.events[idx];
+
+	return ampere_cspmu_validate_configs(curr, new);
+}
+
+static char *ampere_cspmu_format_name(const struct arm_cspmu *cspmu,
+				      const char *name_pattern)
+{
+	struct device *dev = cspmu->dev;
+	static atomic_t pmu_generic_idx = {0};

Why not an ida?

If the pmu drivers ever become easy to unbind then you won't get ID
reusage like this an eventually you will run into overflow problems.

+
+	return devm_kasprintf(dev, GFP_KERNEL, name_pattern,
+			      atomic_fetch_inc(&pmu_generic_idx));
+}
+
+int ampere_cspmu_init_ops(struct arm_cspmu *cspmu)
+{
+	struct device *dev = cspmu->dev;
+	struct ampere_cspmu_ctx *ctx;
+	struct arm_cspmu_impl_ops *impl_ops = &cspmu->impl.ops;
+
+	ctx = devm_kzalloc(dev, sizeof(struct ampere_cspmu_ctx), GFP_KERNEL);
+	if (!ctx)
+		return -ENOMEM;
+
+	ctx->event_attr		= ampereone_mcu_pmu_event_attrs;
+	ctx->format_attr	= ampereone_mcu_format_attrs;
+	ctx->name		= ampere_cspmu_format_name(cspmu,
+							   "ampere_mcu_pmu_%u");

Long line and need to break avoided if you don't bother trying to align the = signs...
Personally I don't like this style as it causes a lot of churn as drivers
evolve, but meh, it's up to you.

Given the result is confusing if the allocation fails (name not what is expected)
I would also check that allocation and error out if it fails.  Obviously it won't
under realistic circumstances, but a bit of paranoia never hurt anyone.

+	cspmu->impl.ctx = ctx;
+
+	impl_ops->event_filter		= ampere_cspmu_event_filter;
+	impl_ops->set_ev_filter		= ampere_cspmu_set_ev_filter;
+	impl_ops->validate_event	= ampere_cspmu_validate_event;
+	impl_ops->get_name		= ampere_cspmu_get_name;
+	impl_ops->get_event_attrs	= ampere_cspmu_get_event_attrs;
+	impl_ops->get_format_attrs	= ampere_cspmu_get_format_attrs;
+
+	return 0;
+}
+
+MODULE_LICENSE("GPL v2");

...

I'd be tempted to keep the generic header in a separate block then
follow with the vendor ones.  Not particularly important though.

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel