Re: [PATCH v4 2/2] ThunderX2: Add Cavium ThunderX2 SoC UNCORE PMU driver

[Mark Rutland]

Hi,

On Wed, Apr 25, 2018 at 02:30:47PM +0530, Ganapatrao Kulkarni wrote:
> +
> +/* L3c and DMC has 16 and 8 channels per socket respectively.
> + * Each Channel supports UNCORE PMU device and consists of
> + * 4 independent programmable counters. Counters are 32 bit
> + * and does not support overflow interrupt, they needs to be
> + * sampled before overflow(i.e, at every 2 seconds).
> + */
> +
> +#define UNCORE_MAX_COUNTERS		4
> +#define UNCORE_L3_MAX_TILES		16
> +#define UNCORE_DMC_MAX_CHANNELS		8
> +
> +#define UNCORE_HRTIMER_INTERVAL		(2 * NSEC_PER_SEC)

How was a period of two seconds chosen?

What's the maximum clock speed for the L3C and DMC?

Given that all channels compete for access to the muxed register
interface, I suspect we need to try more often than once every 2
seconds...

[...]

> +struct active_timer {
> +	struct perf_event *event;
> +	struct hrtimer hrtimer;
> +};
> +
> +/*
> + * pmu on each socket has 2 uncore devices(dmc and l3),
> + * each uncore device has up to 16 channels, each channel can sample
> + * events independently with counters up to 4.
> + *
> + * struct thunderx2_pmu_uncore_channel created per channel.
> + * struct thunderx2_pmu_uncore_dev per uncore device.
> + */
> +struct thunderx2_pmu_uncore_channel {
> +	struct thunderx2_pmu_uncore_dev *uncore_dev;
> +	struct pmu pmu;

Can we put the pmu first in the struct, please?

> +	int counter;

AFAICT, this counter field is never used.

> +	int channel;
> +	DECLARE_BITMAP(counter_mask, UNCORE_MAX_COUNTERS);
> +	struct active_timer *active_timers;

You should only need a single timer per channel, rather than one per
event.

I think you can get rid of the active_timer structure, and have:

	struct perf_event *events[UNCORE_MAX_COUNTERS];
	struct hrtimer timer;

> +	/* to sync counter alloc/release */
> +	raw_spinlock_t lock;
> +};
> +
> +struct thunderx2_pmu_uncore_dev {
> +	char *name;
> +	struct device *dev;
> +	enum thunderx2_uncore_type type;
> +	unsigned long base;

This should be:

void __iomem *base;

[...]

> +static ssize_t cpumask_show(struct device *dev,
> +				struct device_attribute *attr, char *buf)
> +{
> +	struct cpumask cpu_mask;
> +	struct thunderx2_pmu_uncore_channel *pmu_uncore =
> +		pmu_to_thunderx2_pmu_uncore(dev_get_drvdata(dev));
> +
> +	/* Pick first online cpu from the node */
> +	cpumask_clear(&cpu_mask);
> +	cpumask_set_cpu(cpumask_first(
> +			cpumask_of_node(pmu_uncore->uncore_dev->node)),
> +			&cpu_mask);
> +
> +	return cpumap_print_to_pagebuf(true, buf, &cpu_mask);
> +}

AFAICT cpumask_of_node() returns a mask that can include offline CPUs.

Regardless, I don't think that you can keep track of the management CPU
this way. Please keep track of the CPU the PMU should be managed by,
either with a cpumask or int embedded within the PMU structure.

At hotplug time, you'll need to update the management CPU, calling
perf_pmu_migrate_context() when it is offlined.

[...]

> +static int alloc_counter(struct thunderx2_pmu_uncore_channel *pmu_uncore)
> +{
> +	int counter;
> +
> +	raw_spin_lock(&pmu_uncore->lock);
> +	counter = find_first_zero_bit(pmu_uncore->counter_mask,
> +				pmu_uncore->uncore_dev->max_counters);
> +	if (counter == pmu_uncore->uncore_dev->max_counters) {
> +		raw_spin_unlock(&pmu_uncore->lock);
> +		return -ENOSPC;
> +	}
> +	set_bit(counter, pmu_uncore->counter_mask);
> +	raw_spin_unlock(&pmu_uncore->lock);
> +	return counter;
> +}
> +
> +static void free_counter(struct thunderx2_pmu_uncore_channel *pmu_uncore,
> +					int counter)
> +{
> +	raw_spin_lock(&pmu_uncore->lock);
> +	clear_bit(counter, pmu_uncore->counter_mask);
> +	raw_spin_unlock(&pmu_uncore->lock);
> +}

I don't believe that locking is required in either of these, as the perf
core serializes pmu::add() and pmu::del(), where these get called.

> +
> +/*
> + * DMC and L3 counter interface is muxed across all channels.
> + * hence we need to select the channel before accessing counter
> + * data/control registers.

Are there separate interfaces for all-dmc-channels and all-l3c-channels?

... or is a single interface used by all-dmc-and-l3c-channels?

> + *
> + *  L3 Tile and DMC channel selection is through SMC call
> + *  SMC call arguments,
> + *	x0 = THUNDERX2_SMC_CALL_ID	(Vendor SMC call Id)
> + *	x1 = THUNDERX2_SMC_SET_CHANNEL	(Id to set DMC/L3C channel)
> + *	x2 = Node id

How do we map Linux node IDs to the firmware's view of node IDs?

I don't believe the two are necessarily the same -- Linux's node IDs are
a Linux-specific construct.

It would be much nicer if we could pass something based on the MPIDR,
which is a known HW construct, or if this implicitly affected the
current node.

It would be vastly more sane for this to not be muxed at all. :/

> + *	x3 = DMC(1)/L3C(0)
> + *	x4 = channel Id
> + */
> +static void uncore_select_channel(struct perf_event *event)
> +{
> +	struct arm_smccc_res res;
> +	struct thunderx2_pmu_uncore_channel *pmu_uncore =
> +		pmu_to_thunderx2_pmu_uncore(event->pmu);
> +
> +	arm_smccc_smc(THUNDERX2_SMC_CALL_ID, THUNDERX2_SMC_SET_CHANNEL,
> +			pmu_uncore->uncore_dev->node,
> +			pmu_uncore->uncore_dev->type,
> +			pmu_uncore->channel, 0, 0, 0, &res);
> +
> +}

Why aren't we checking the return value of the SMC call?

The muxing and SMC sound quite scary. :/

> +
> +static void uncore_start_event_l3c(struct perf_event *event, int flags)
> +{
> +	u32 val;
> +	struct hw_perf_event *hwc = &event->hw;
> +
> +	/* event id encoded in bits [07:03] */
> +	val = GET_EVENTID(event) << 3;
> +	reg_writel(val, hwc->config_base);
> +
> +	if (flags & PERF_EF_RELOAD) {
> +		u64 prev_raw_count =
> +			local64_read(&event->hw.prev_count);
> +		reg_writel(prev_raw_count, hwc->event_base);
> +	}
> +	local64_set(&event->hw.prev_count,
> +			reg_readl(hwc->event_base));

It would be simpler to ignore PERF_EF_RELOAD, and always reprogram the
prev_count and HW to zero.

> +}
> +
> +static void uncore_start_event_dmc(struct perf_event *event, int flags)
> +{
> +	u32 val, event_shift = 8;
> +	struct hw_perf_event *hwc = &event->hw;
> +
> +	/* enable and start counters and read current value in prev_count */
> +	val = reg_readl(hwc->config_base);
> +
> +	/* 8 bits for each counter,
> +	 * bits [05:01] of a counter to set event type.
> +	 */
> +	reg_writel((val & ~(0x1f << (((GET_COUNTERID(event)) *
> +		event_shift) + 1))) |
> +		(GET_EVENTID(event) <<
> +		 (((GET_COUNTERID(event)) * event_shift) + 1)),
> +		hwc->config_base);

This would be far more legible if val were constructed before the
reg_writel(), especially if there were a helper for the event field
shifting, e.g.

	#define DMC_EVENT_CFG(idx, val)	((val) << (((idx) * 8) + 1))

	int id = GET_COUNTERID(event);
	int event = GET_EVENTID(event);

	val = reg_readl(hwc->config_base);
	val &= ~DMC_EVENT_CFG(id, 0x1f);
	val |= DMCC_EVENT_CFG(id, event);
	reg_writel(val, hwc->config_base));

What are bits 7:6 and 1 used for?

> +
> +	if (flags & PERF_EF_RELOAD) {
> +		u64 prev_raw_count =
> +			local64_read(&event->hw.prev_count);
> +		reg_writel(prev_raw_count, hwc->event_base);
> +	}
> +	local64_set(&event->hw.prev_count,
> +			reg_readl(hwc->event_base));


As with the L3C events, it would be simpler to always reprogram the
prev_count and HW to zero.

> +}
> +
> +static void uncore_stop_event_l3c(struct perf_event *event)
> +{
> +	reg_writel(0, event->hw.config_base);
> +}
> +
> +static void uncore_stop_event_dmc(struct perf_event *event)
> +{
> +	u32 val, event_shift = 8;
> +	struct hw_perf_event *hwc = &event->hw;
> +
> +	val = reg_readl(hwc->config_base);
> +	reg_writel((val & ~(0xff << ((GET_COUNTERID(event)) * event_shift))),
> +			hwc->config_base);


This could be simplified with the helper proposed above.

> +}
> +
> +static void init_cntr_base_l3c(struct perf_event *event,
> +		struct thunderx2_pmu_uncore_dev *uncore_dev)
> +{
> +
> +	struct hw_perf_event *hwc = &event->hw;
> +
> +	/* counter ctrl/data reg offset at 8 */

Offset 8, or stride 8?

What does the register layout look like?

> +	hwc->config_base = uncore_dev->base
> +		+ L3C_COUNTER_CTL + (8 * GET_COUNTERID(event));
> +	hwc->event_base =  uncore_dev->base
> +		+ L3C_COUNTER_DATA + (8 * GET_COUNTERID(event));
> +}
> +
> +static void init_cntr_base_dmc(struct perf_event *event,
> +		struct thunderx2_pmu_uncore_dev *uncore_dev)
> +{
> +
> +	struct hw_perf_event *hwc = &event->hw;
> +
> +	hwc->config_base = uncore_dev->base
> +		+ DMC_COUNTER_CTL;
> +	/* counter data reg offset at 0xc */

A stride of 0xc seems unusual.

What does the register layout look like?

> +	hwc->event_base =  uncore_dev->base
> +		+ DMC_COUNTER_DATA + (0xc * GET_COUNTERID(event));
> +}
> +
> +static void thunderx2_uncore_update(struct perf_event *event)
> +{
> +	s64 prev, new = 0;
> +	u64 delta;
> +	struct hw_perf_event *hwc = &event->hw;
> +	struct thunderx2_pmu_uncore_channel *pmu_uncore;
> +	enum thunderx2_uncore_type type;
> +
> +	pmu_uncore = pmu_to_thunderx2_pmu_uncore(event->pmu);
> +	type = pmu_uncore->uncore_dev->type;

AFAICT this variable is not used below.

> +
> +	if (pmu_uncore->uncore_dev->select_channel)
> +		pmu_uncore->uncore_dev->select_channel(event);

This should always be non-NULL, right?

[...]

> +static bool thunderx2_uncore_validate_event_group(struct perf_event *event)
> +{
> +	struct pmu *pmu = event->pmu;
> +	struct perf_event *leader = event->group_leader;
> +	struct perf_event *sibling;
> +	int counters = 0;
> +
> +	if (leader->pmu != event->pmu && !is_software_event(leader))
> +		return false;
> +
> +	counters++;

I don't think this is right when event != leader and the leader is a SW
event. In that case, the leader doesn't take a HW counter.

> +
> +	for_each_sibling_event(sibling, event->group_leader) {
> +		if (is_software_event(sibling))
> +			continue;
> +		if (sibling->pmu != pmu)
> +			return false;
> +		counters++;
> +	}
> +
> +	/*
> +	 * If the group requires more counters than the HW has,
> +	 * it cannot ever be scheduled.
> +	 */
> +	return counters <= UNCORE_MAX_COUNTERS;
> +}

[...]

> +static int thunderx2_uncore_event_init(struct perf_event *event)
> +{
> +	struct hw_perf_event *hwc = &event->hw;
> +	struct thunderx2_pmu_uncore_channel *pmu_uncore;

> +	pmu_uncore = pmu_to_thunderx2_pmu_uncore(event->pmu);
> +
> +	if (!pmu_uncore)
> +		return -ENODEV;

This cannot happen, given pmu_to_thunderx2_pmu_uncore() is a wrapper
around container_of().

> +
> +	/* Pick first online cpu from the node */
> +	event->cpu = cpumask_first(
> +			cpumask_of_node(pmu_uncore->uncore_dev->node));

I don't believe this is safe. You must keep track of which CPU is
managing the PMU, with hotplug callbacks.

[...]

> +static void thunderx2_uncore_start(struct perf_event *event, int flags)
> +{
> +	struct hw_perf_event *hwc = &event->hw;
> +	struct thunderx2_pmu_uncore_channel *pmu_uncore;
> +	struct thunderx2_pmu_uncore_dev *uncore_dev;
> +	unsigned long irqflags;
> +	struct active_timer *active_timer;
> +
> +	hwc->state = 0;
> +	pmu_uncore = pmu_to_thunderx2_pmu_uncore(event->pmu);
> +	uncore_dev = pmu_uncore->uncore_dev;
> +
> +	raw_spin_lock_irqsave(&uncore_dev->lock, irqflags);
> +
> +	if (uncore_dev->select_channel)
> +		uncore_dev->select_channel(event);
> +	uncore_dev->start_event(event, flags);
> +	raw_spin_unlock_irqrestore(&uncore_dev->lock, irqflags);
> +
> +	perf_event_update_userpage(event);
> +	active_timer = &pmu_uncore->active_timers[GET_COUNTERID(event)];
> +	active_timer->event = event;
> +
> +	hrtimer_start(&active_timer->hrtimer,
> +			ns_to_ktime(uncore_dev->hrtimer_interval),
> +			HRTIMER_MODE_REL_PINNED);
> +}

Please use a single hrtimer, and update *all* of the events when it
fires.

I *think* that can be done in the pmu::pmu_enable() and
pmu::pmu_disable() callbacks.

Are there control bits to enable/disable all counters, or can that only
be done through the event configuration registers?

> +static void thunderx2_uncore_stop(struct perf_event *event, int flags)
> +{
> +	struct hw_perf_event *hwc = &event->hw;
> +	struct thunderx2_pmu_uncore_channel *pmu_uncore;
> +	struct thunderx2_pmu_uncore_dev *uncore_dev;
> +	unsigned long irqflags;
> +
> +	if (hwc->state & PERF_HES_UPTODATE)
> +		return;
> +
> +	pmu_uncore = pmu_to_thunderx2_pmu_uncore(event->pmu);
> +	uncore_dev = pmu_uncore->uncore_dev;
> +
> +	raw_spin_lock_irqsave(&uncore_dev->lock, irqflags);
> +
> +	if (uncore_dev->select_channel)
> +		uncore_dev->select_channel(event);

AFAICT this cannot be NULL.

[...]

> +static int thunderx2_pmu_uncore_register(
> +		struct thunderx2_pmu_uncore_channel *pmu_uncore)
> +{
> +	struct device *dev = pmu_uncore->uncore_dev->dev;
> +	char *name = pmu_uncore->uncore_dev->name;
> +	int channel = pmu_uncore->channel;
> +
> +	/* Perf event registration */
> +	pmu_uncore->pmu = (struct pmu) {
> +		.attr_groups	= pmu_uncore->uncore_dev->attr_groups,
> +		.task_ctx_nr	= perf_invalid_context,
> +		.event_init	= thunderx2_uncore_event_init,
> +		.add		= thunderx2_uncore_add,
> +		.del		= thunderx2_uncore_del,
> +		.start		= thunderx2_uncore_start,
> +		.stop		= thunderx2_uncore_stop,
> +		.read		= thunderx2_uncore_read,
> +	};
> +
> +	pmu_uncore->pmu.name = devm_kasprintf(dev, GFP_KERNEL,
> +			"%s_%d", name, channel);

Does the channel idx take the NUMA node into account?

[...]

> +static int thunderx2_pmu_uncore_add(struct thunderx2_pmu_uncore_dev *uncore_dev,
> +		int channel)
> +{

> +	/* we can run up to (max_counters * max_channels) events
> +	 * simultaneously, allocate hrtimers per channel.
> +	 */
> +	pmu_uncore->active_timers = devm_kzalloc(uncore_dev->dev,
> +			sizeof(struct active_timer) * uncore_dev->max_counters,
> +			GFP_KERNEL);

Please just fold a single hrtimer into the thunderx2_pmu_uncore_channel
structure, and you can get rid of this allocation...

> +
> +	for (counter = 0; counter < uncore_dev->max_counters; counter++) {
> +		hrtimer_init(&pmu_uncore->active_timers[counter].hrtimer,
> +				CLOCK_MONOTONIC,
> +				HRTIMER_MODE_REL);
> +		pmu_uncore->active_timers[counter].hrtimer.function =
> +				thunderx2_uncore_hrtimer_callback;
> +	}

... and simplify this initialization.

[...]

> +static struct thunderx2_pmu_uncore_dev *init_pmu_uncore_dev(
> +		struct device *dev, acpi_handle handle,
> +		struct acpi_device *adev, u32 type)
> +{
> +	struct thunderx2_pmu_uncore_dev *uncore_dev;
> +	unsigned long base;

> +	base = (unsigned long)devm_ioremap_resource(dev, &res);
> +	if (IS_ERR((void *)base)) {
> +		dev_err(dev, "PMU type %d: Fail to map resource\n", type);
> +		return NULL;
> +	}

Please treat this as a void __iomem *base.

[...]

> +static int thunderx2_uncore_probe(struct platform_device *pdev)
> +{
> +	struct device *dev = &pdev->dev;
> +	struct arm_smccc_res res;
> +
> +	set_dev_node(dev, acpi_get_node(ACPI_HANDLE(dev)));
> +
> +	/* Make sure firmware supports DMC/L3C set channel smc call */
> +	arm_smccc_smc(THUNDERX2_SMC_CALL_ID, THUNDERX2_SMC_SET_CHANNEL,
> +			dev_to_node(dev), 0, 0, 0, 0, 0, &res);
> +	if (res.a0) {
> +		dev_err(dev, "No Firmware support for PMU UNCORE(%d)\n",
> +				dev_to_node(dev));
> +		return -ENODEV;
> +	}

Please re-use the uncore_select_channel() wrapper rather than
open-coding this.

Which FW supports this interface?

Thanks,
Mark.