perf: qcom: Add L3 cache PMU driver

This adds a new dynamic PMU to the Perf Events framework to program
and control the L3 cache PMUs in some Qualcomm Technologies SOCs.

The driver supports a distributed cache architecture where the overall
cache for a socket is comprised of multiple slices each with its own PMU.
Access to each individual PMU is provided even though all CPUs share all
the slices. User space needs to aggregate to individual counts to provide
a global picture.

The driver exports formatting and event information to sysfs so it can
be used by the perf user space tools with the syntaxes:
   perf stat -a -e l3cache_0_0/read-miss/
   perf stat -a -e l3cache_0_0/event=0x21/
Acked-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Agustin Vega-Frias <agustinv@codeaurora.org>
[will: fixed sparse issues]
Signed-off-by: Will Deacon <will.deacon@arm.com>

Documentation/perf/qcom_l3_pmu.txt

+Qualcomm Datacenter Technologies L3 Cache Performance Monitoring Unit (PMU)
+===========================================================================
+
+This driver supports the L3 cache PMUs found in Qualcomm Datacenter Technologies
+Centriq SoCs. The L3 cache on these SOCs is composed of multiple slices, shared
+by all cores within a socket. Each slice is exposed as a separate uncore perf
+PMU with device name l3cache_<socket>_<instance>. User space is responsible
+for aggregating across slices.
+
+The driver provides a description of its available events and configuration
+options in sysfs, see /sys/devices/l3cache*. Given that these are uncore PMUs
+the driver also exposes a "cpumask" sysfs attribute which contains a mask
+consisting of one CPU per socket which will be used to handle all the PMU
+events on that socket.
+
+The hardware implements 32bit event counters and has a flat 8bit event space
+exposed via the "event" format attribute. In addition to the 32bit physical
+counters the driver supports virtual 64bit hardware counters by using hardware
+counter chaining. This feature is exposed via the "lc" (long counter) format
+flag. E.g.:
+
+  perf stat -e l3cache_0_0/read-miss,lc/
+
+Given that these are uncore PMUs the driver does not support sampling, therefore
+"perf record" will not work. Per-task perf sessions are not supported.

drivers/perf/Kconfig

@@ -21,6 +21,16 @@ config QCOM_L2_PMU
 	  Adds the L2 cache PMU into the perf events subsystem for
 	  monitoring L2 cache events.
 
+config QCOM_L3_PMU
+	bool "Qualcomm Technologies L3-cache PMU"
+	depends on ARCH_QCOM && ARM64 && PERF_EVENTS && ACPI
+	select QCOM_IRQ_COMBINER
+	help
+	   Provides support for the L3 cache performance monitor unit (PMU)
+	   in Qualcomm Technologies processors.
+	   Adds the L3 cache PMU into the perf events subsystem for
+	   monitoring L3 cache events.
+
 config XGENE_PMU
         depends on PERF_EVENTS && ARCH_XGENE
         bool "APM X-Gene SoC PMU"

drivers/perf/Makefile

 obj-$(CONFIG_ARM_PMU) += arm_pmu.o
 obj-$(CONFIG_QCOM_L2_PMU)	+= qcom_l2_pmu.o
+obj-$(CONFIG_QCOM_L3_PMU) += qcom_l3_pmu.o
 obj-$(CONFIG_XGENE_PMU) += xgene_pmu.o

drivers/perf/qcom_l3_pmu.c

+/*
+ * Driver for the L3 cache PMUs in Qualcomm Technologies chips.
+ *
+ * The driver supports a distributed cache architecture where the overall
+ * cache for a socket is comprised of multiple slices each with its own PMU.
+ * Access to each individual PMU is provided even though all CPUs share all
+ * the slices. User space needs to aggregate to individual counts to provide
+ * a global picture.
+ *
+ * See Documentation/perf/qcom_l3_pmu.txt for more details.
+ *
+ * Copyright (c) 2015-2017, The Linux Foundation. All rights reserved.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 and
+ * only version 2 as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ */
+
+#include <linux/acpi.h>
+#include <linux/bitops.h>
+#include <linux/interrupt.h>
+#include <linux/io.h>
+#include <linux/list.h>
+#include <linux/module.h>
+#include <linux/perf_event.h>
+#include <linux/platform_device.h>
+
+/*
+ * General constants
+ */
+
+/* Number of counters on each PMU */
+#define L3_NUM_COUNTERS  8
+/* Mask for the event type field within perf_event_attr.config and EVTYPE reg */
+#define L3_EVTYPE_MASK   0xFF
+/*
+ * Bit position of the 'long counter' flag within perf_event_attr.config.
+ * Reserve some space between the event type and this flag to allow expansion
+ * in the event type field.
+ */
+#define L3_EVENT_LC_BIT  32
+
+/*
+ * Register offsets
+ */
+
+/* Perfmon registers */
+#define L3_HML3_PM_CR       0x000
+#define L3_HML3_PM_EVCNTR(__cntr) (0x420 + ((__cntr) & 0x7) * 8)
+#define L3_HML3_PM_CNTCTL(__cntr) (0x120 + ((__cntr) & 0x7) * 8)
+#define L3_HML3_PM_EVTYPE(__cntr) (0x220 + ((__cntr) & 0x7) * 8)
+#define L3_HML3_PM_FILTRA   0x300
+#define L3_HML3_PM_FILTRB   0x308
+#define L3_HML3_PM_FILTRC   0x310
+#define L3_HML3_PM_FILTRAM  0x304
+#define L3_HML3_PM_FILTRBM  0x30C
+#define L3_HML3_PM_FILTRCM  0x314
+
+/* Basic counter registers */
+#define L3_M_BC_CR         0x500
+#define L3_M_BC_SATROLL_CR 0x504
+#define L3_M_BC_CNTENSET   0x508
+#define L3_M_BC_CNTENCLR   0x50C
+#define L3_M_BC_INTENSET   0x510
+#define L3_M_BC_INTENCLR   0x514
+#define L3_M_BC_GANG       0x718
+#define L3_M_BC_OVSR       0x740
+#define L3_M_BC_IRQCTL     0x96C
+
+/*
+ * Bit field definitions
+ */
+
+/* L3_HML3_PM_CR */
+#define PM_CR_RESET           (0)
+
+/* L3_HML3_PM_XCNTCTL/L3_HML3_PM_CNTCTLx */
+#define PMCNT_RESET           (0)
+
+/* L3_HML3_PM_EVTYPEx */
+#define EVSEL(__val)          ((__val) & L3_EVTYPE_MASK)
+
+/* Reset value for all the filter registers */
+#define PM_FLTR_RESET         (0)
+
+/* L3_M_BC_CR */
+#define BC_RESET              (1UL << 1)
+#define BC_ENABLE             (1UL << 0)
+
+/* L3_M_BC_SATROLL_CR */
+#define BC_SATROLL_CR_RESET   (0)
+
+/* L3_M_BC_CNTENSET */
+#define PMCNTENSET(__cntr)    (1UL << ((__cntr) & 0x7))
+
+/* L3_M_BC_CNTENCLR */
+#define PMCNTENCLR(__cntr)    (1UL << ((__cntr) & 0x7))
+#define BC_CNTENCLR_RESET     (0xFF)
+
+/* L3_M_BC_INTENSET */
+#define PMINTENSET(__cntr)    (1UL << ((__cntr) & 0x7))
+
+/* L3_M_BC_INTENCLR */
+#define PMINTENCLR(__cntr)    (1UL << ((__cntr) & 0x7))
+#define BC_INTENCLR_RESET     (0xFF)
+
+/* L3_M_BC_GANG */
+#define GANG_EN(__cntr)       (1UL << ((__cntr) & 0x7))
+#define BC_GANG_RESET         (0)
+
+/* L3_M_BC_OVSR */
+#define PMOVSRCLR(__cntr)     (1UL << ((__cntr) & 0x7))
+#define PMOVSRCLR_RESET       (0xFF)
+
+/* L3_M_BC_IRQCTL */
+#define PMIRQONMSBEN(__cntr)  (1UL << ((__cntr) & 0x7))
+#define BC_IRQCTL_RESET       (0x0)
+
+/*
+ * Events
+ */
+
+#define L3_EVENT_CYCLES		0x01
+#define L3_EVENT_READ_HIT		0x20
+#define L3_EVENT_READ_MISS		0x21
+#define L3_EVENT_READ_HIT_D		0x22
+#define L3_EVENT_READ_MISS_D		0x23
+#define L3_EVENT_WRITE_HIT		0x24
+#define L3_EVENT_WRITE_MISS		0x25
+
+/*
+ * Decoding of settings from perf_event_attr
+ *
+ * The config format for perf events is:
+ * - config: bits 0-7: event type
+ *           bit  32:  HW counter size requested, 0: 32 bits, 1: 64 bits
+ */
+
+static inline u32 get_event_type(struct perf_event *event)
+{
+	return (event->attr.config) & L3_EVTYPE_MASK;
+}
+
+static inline bool event_uses_long_counter(struct perf_event *event)
+{
+	return !!(event->attr.config & BIT_ULL(L3_EVENT_LC_BIT));
+}
+
+static inline int event_num_counters(struct perf_event *event)
+{
+	return event_uses_long_counter(event) ? 2 : 1;
+}
+
+/*
+ * Main PMU, inherits from the core perf PMU type
+ */
+struct l3cache_pmu {
+	struct pmu		pmu;
+	struct hlist_node	node;
+	void __iomem		*regs;
+	struct perf_event	*events[L3_NUM_COUNTERS];
+	unsigned long		used_mask[BITS_TO_LONGS(L3_NUM_COUNTERS)];
+	cpumask_t		cpumask;
+};
+
+#define to_l3cache_pmu(p) (container_of(p, struct l3cache_pmu, pmu))
+
+/*
+ * Type used to group hardware counter operations
+ *
+ * Used to implement two types of hardware counters, standard (32bits) and
+ * long (64bits). The hardware supports counter chaining which we use to
+ * implement long counters. This support is exposed via the 'lc' flag field
+ * in perf_event_attr.config.
+ */
+struct l3cache_event_ops {
+	/* Called to start event monitoring */
+	void (*start)(struct perf_event *event);
+	/* Called to stop event monitoring */
+	void (*stop)(struct perf_event *event, int flags);
+	/* Called to update the perf_event */
+	void (*update)(struct perf_event *event);
+};
+
+/*
+ * Implementation of long counter operations
+ *
+ * 64bit counters are implemented by chaining two of the 32bit physical
+ * counters. The PMU only supports chaining of adjacent even/odd pairs
+ * and for simplicity the driver always configures the odd counter to
+ * count the overflows of the lower-numbered even counter. Note that since
+ * the resulting hardware counter is 64bits no IRQs are required to maintain
+ * the software counter which is also 64bits.
+ */
+
+static void qcom_l3_cache__64bit_counter_start(struct perf_event *event)
+{
+	struct l3cache_pmu *l3pmu = to_l3cache_pmu(event->pmu);
+	int idx = event->hw.idx;
+	u32 evsel = get_event_type(event);
+	u32 gang;
+
+	/* Set the odd counter to count the overflows of the even counter */
+	gang = readl_relaxed(l3pmu->regs + L3_M_BC_GANG);
+	gang |= GANG_EN(idx + 1);
+	writel_relaxed(gang, l3pmu->regs + L3_M_BC_GANG);
+
+	/* Initialize the hardware counters and reset prev_count*/
+	local64_set(&event->hw.prev_count, 0);
+	writel_relaxed(0, l3pmu->regs + L3_HML3_PM_EVCNTR(idx + 1));
+	writel_relaxed(0, l3pmu->regs + L3_HML3_PM_EVCNTR(idx));
+
+	/*
+	 * Set the event types, the upper half must use zero and the lower
+	 * half the actual event type
+	 */
+	writel_relaxed(EVSEL(0), l3pmu->regs + L3_HML3_PM_EVTYPE(idx + 1));
+	writel_relaxed(EVSEL(evsel), l3pmu->regs + L3_HML3_PM_EVTYPE(idx));
+
+	/* Finally, enable the counters */
+	writel_relaxed(PMCNT_RESET, l3pmu->regs + L3_HML3_PM_CNTCTL(idx + 1));
+	writel_relaxed(PMCNTENSET(idx + 1), l3pmu->regs + L3_M_BC_CNTENSET);
+	writel_relaxed(PMCNT_RESET, l3pmu->regs + L3_HML3_PM_CNTCTL(idx));
+	writel_relaxed(PMCNTENSET(idx), l3pmu->regs + L3_M_BC_CNTENSET);
+}
+
+static void qcom_l3_cache__64bit_counter_stop(struct perf_event *event,
+					      int flags)
+{
+	struct l3cache_pmu *l3pmu = to_l3cache_pmu(event->pmu);
+	int idx = event->hw.idx;
+	u32 gang = readl_relaxed(l3pmu->regs + L3_M_BC_GANG);
+
+	/* Disable the counters */
+	writel_relaxed(PMCNTENCLR(idx), l3pmu->regs + L3_M_BC_CNTENCLR);
+	writel_relaxed(PMCNTENCLR(idx + 1), l3pmu->regs + L3_M_BC_CNTENCLR);
+
+	/* Disable chaining */
+	writel_relaxed(gang & ~GANG_EN(idx + 1), l3pmu->regs + L3_M_BC_GANG);
+}
+
+static void qcom_l3_cache__64bit_counter_update(struct perf_event *event)
+{
+	struct l3cache_pmu *l3pmu = to_l3cache_pmu(event->pmu);
+	int idx = event->hw.idx;
+	u32 hi, lo;
+	u64 prev, new;
+
+	do {
+		prev = local64_read(&event->hw.prev_count);
+		do {
+			hi = readl_relaxed(l3pmu->regs + L3_HML3_PM_EVCNTR(idx + 1));
+			lo = readl_relaxed(l3pmu->regs + L3_HML3_PM_EVCNTR(idx));
+		} while (hi != readl_relaxed(l3pmu->regs + L3_HML3_PM_EVCNTR(idx + 1)));
+		new = ((u64)hi << 32) | lo;
+	} while (local64_cmpxchg(&event->hw.prev_count, prev, new) != prev);
+
+	local64_add(new - prev, &event->count);
+}
+
+static const struct l3cache_event_ops event_ops_long = {
+	.start = qcom_l3_cache__64bit_counter_start,
+	.stop = qcom_l3_cache__64bit_counter_stop,
+	.update = qcom_l3_cache__64bit_counter_update,
+};
+
+/*
+ * Implementation of standard counter operations
+ *
+ * 32bit counters use a single physical counter and a hardware feature that
+ * asserts the overflow IRQ on the toggling of the most significant bit in
+ * the counter. This feature allows the counters to be left free-running
+ * without needing the usual reprogramming required to properly handle races
+ * during concurrent calls to update.
+ */
+
+static void qcom_l3_cache__32bit_counter_start(struct perf_event *event)
+{
+	struct l3cache_pmu *l3pmu = to_l3cache_pmu(event->pmu);
+	int idx = event->hw.idx;
+	u32 evsel = get_event_type(event);
+	u32 irqctl = readl_relaxed(l3pmu->regs + L3_M_BC_IRQCTL);
+
+	/* Set the counter to assert the overflow IRQ on MSB toggling */
+	writel_relaxed(irqctl | PMIRQONMSBEN(idx), l3pmu->regs + L3_M_BC_IRQCTL);
+
+	/* Initialize the hardware counter and reset prev_count*/
+	local64_set(&event->hw.prev_count, 0);
+	writel_relaxed(0, l3pmu->regs + L3_HML3_PM_EVCNTR(idx));
+
+	/* Set the event type */
+	writel_relaxed(EVSEL(evsel), l3pmu->regs + L3_HML3_PM_EVTYPE(idx));
+
+	/* Enable interrupt generation by this counter */
+	writel_relaxed(PMINTENSET(idx), l3pmu->regs + L3_M_BC_INTENSET);
+
+	/* Finally, enable the counter */
+	writel_relaxed(PMCNT_RESET, l3pmu->regs + L3_HML3_PM_CNTCTL(idx));
+	writel_relaxed(PMCNTENSET(idx), l3pmu->regs + L3_M_BC_CNTENSET);
+}
+
+static void qcom_l3_cache__32bit_counter_stop(struct perf_event *event,
+					      int flags)
+{
+	struct l3cache_pmu *l3pmu = to_l3cache_pmu(event->pmu);
+	int idx = event->hw.idx;
+	u32 irqctl = readl_relaxed(l3pmu->regs + L3_M_BC_IRQCTL);
+
+	/* Disable the counter */
+	writel_relaxed(PMCNTENCLR(idx), l3pmu->regs + L3_M_BC_CNTENCLR);
+
+	/* Disable interrupt generation by this counter */
+	writel_relaxed(PMINTENCLR(idx), l3pmu->regs + L3_M_BC_INTENCLR);
+
+	/* Set the counter to not assert the overflow IRQ on MSB toggling */
+	writel_relaxed(irqctl & ~PMIRQONMSBEN(idx), l3pmu->regs + L3_M_BC_IRQCTL);
+}
+
+static void qcom_l3_cache__32bit_counter_update(struct perf_event *event)
+{
+	struct l3cache_pmu *l3pmu = to_l3cache_pmu(event->pmu);
+	int idx = event->hw.idx;
+	u32 prev, new;
+
+	do {
+		prev = local64_read(&event->hw.prev_count);
+		new = readl_relaxed(l3pmu->regs + L3_HML3_PM_EVCNTR(idx));
+	} while (local64_cmpxchg(&event->hw.prev_count, prev, new) != prev);
+
+	local64_add(new - prev, &event->count);
+}
+
+static const struct l3cache_event_ops event_ops_std = {
+	.start = qcom_l3_cache__32bit_counter_start,
+	.stop = qcom_l3_cache__32bit_counter_stop,
+	.update = qcom_l3_cache__32bit_counter_update,
+};
+
+/* Retrieve the appropriate operations for the given event */
+static
+const struct l3cache_event_ops *l3cache_event_get_ops(struct perf_event *event)
+{
+	if (event_uses_long_counter(event))
+		return &event_ops_long;
+	else
+		return &event_ops_std;
+}
+
+/*
+ * Top level PMU functions.
+ */
+
+static inline void qcom_l3_cache__init(struct l3cache_pmu *l3pmu)
+{
+	int i;
+
+	writel_relaxed(BC_RESET, l3pmu->regs + L3_M_BC_CR);
+
+	/*
+	 * Use writel for the first programming command to ensure the basic
+	 * counter unit is stopped before proceeding
+	 */
+	writel(BC_SATROLL_CR_RESET, l3pmu->regs + L3_M_BC_SATROLL_CR);
+
+	writel_relaxed(BC_CNTENCLR_RESET, l3pmu->regs + L3_M_BC_CNTENCLR);
+	writel_relaxed(BC_INTENCLR_RESET, l3pmu->regs + L3_M_BC_INTENCLR);
+	writel_relaxed(PMOVSRCLR_RESET, l3pmu->regs + L3_M_BC_OVSR);
+	writel_relaxed(BC_GANG_RESET, l3pmu->regs + L3_M_BC_GANG);
+	writel_relaxed(BC_IRQCTL_RESET, l3pmu->regs + L3_M_BC_IRQCTL);
+	writel_relaxed(PM_CR_RESET, l3pmu->regs + L3_HML3_PM_CR);
+
+	for (i = 0; i < L3_NUM_COUNTERS; ++i) {
+		writel_relaxed(PMCNT_RESET, l3pmu->regs + L3_HML3_PM_CNTCTL(i));
+		writel_relaxed(EVSEL(0), l3pmu->regs + L3_HML3_PM_EVTYPE(i));
+	}
+
+	writel_relaxed(PM_FLTR_RESET, l3pmu->regs + L3_HML3_PM_FILTRA);
+	writel_relaxed(PM_FLTR_RESET, l3pmu->regs + L3_HML3_PM_FILTRAM);
+	writel_relaxed(PM_FLTR_RESET, l3pmu->regs + L3_HML3_PM_FILTRB);
+	writel_relaxed(PM_FLTR_RESET, l3pmu->regs + L3_HML3_PM_FILTRBM);
+	writel_relaxed(PM_FLTR_RESET, l3pmu->regs + L3_HML3_PM_FILTRC);
+	writel_relaxed(PM_FLTR_RESET, l3pmu->regs + L3_HML3_PM_FILTRCM);
+
+	/*
+	 * Use writel here to ensure all programming commands are done
+	 *  before proceeding
+	 */
+	writel(BC_ENABLE, l3pmu->regs + L3_M_BC_CR);
+}
+
+static irqreturn_t qcom_l3_cache__handle_irq(int irq_num, void *data)
+{
+	struct l3cache_pmu *l3pmu = data;
+	/* Read the overflow status register */
+	long status = readl_relaxed(l3pmu->regs + L3_M_BC_OVSR);
+	int idx;
+
+	if (status == 0)
+		return IRQ_NONE;
+
+	/* Clear the bits we read on the overflow status register */
+	writel_relaxed(status, l3pmu->regs + L3_M_BC_OVSR);
+
+	for_each_set_bit(idx, &status, L3_NUM_COUNTERS) {
+		struct perf_event *event;
+		const struct l3cache_event_ops *ops;
+
+		event = l3pmu->events[idx];
+		if (!event)
+			continue;
+
+		/*
+		 * Since the IRQ is not enabled for events using long counters
+		 * we should never see one of those here, however, be consistent
+		 * and use the ops indirections like in the other operations.
+		 */
+
+		ops = l3cache_event_get_ops(event);
+		ops->update(event);
+	}
+
+	return IRQ_HANDLED;
+}
+
+/*
+ * Implementation of abstract pmu functionality required by
+ * the core perf events code.
+ */
+
+static void qcom_l3_cache__pmu_enable(struct pmu *pmu)
+{
+	struct l3cache_pmu *l3pmu = to_l3cache_pmu(pmu);
+
+	/* Ensure the other programming commands are observed before enabling */
+	wmb();
+
+	writel_relaxed(BC_ENABLE, l3pmu->regs + L3_M_BC_CR);
+}
+
+static void qcom_l3_cache__pmu_disable(struct pmu *pmu)
+{
+	struct l3cache_pmu *l3pmu = to_l3cache_pmu(pmu);
+
+	writel_relaxed(0, l3pmu->regs + L3_M_BC_CR);
+
+	/* Ensure the basic counter unit is stopped before proceeding */
+	wmb();
+}
+
+/*
+ * We must NOT create groups containing events from multiple hardware PMUs,
+ * although mixing different software and hardware PMUs is allowed.
+ */
+static bool qcom_l3_cache__validate_event_group(struct perf_event *event)
+{
+	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 = event_num_counters(event);
+	counters += event_num_counters(leader);
+
+	list_for_each_entry(sibling, &leader->sibling_list, group_entry) {
+		if (is_software_event(sibling))
+			continue;
+		if (sibling->pmu != event->pmu)
+			return false;
+		counters += event_num_counters(sibling);
+	}
+
+	/*
+	 * If the group requires more counters than the HW has, it
+	 * cannot ever be scheduled.
+	 */
+	return counters <= L3_NUM_COUNTERS;
+}
+
+static int qcom_l3_cache__event_init(struct perf_event *event)
+{
+	struct l3cache_pmu *l3pmu = to_l3cache_pmu(event->pmu);
+	struct hw_perf_event *hwc = &event->hw;
+
+	/*
+	 * Is the event for this PMU?
+	 */
+	if (event->attr.type != event->pmu->type)
+		return -ENOENT;
+
+	/*
+	 * There are no per-counter mode filters in the PMU.
+	 */
+	if (event->attr.exclude_user || event->attr.exclude_kernel ||
+	    event->attr.exclude_hv || event->attr.exclude_idle)
+		return -EINVAL;
+
+	/*
+	 * Sampling not supported since these events are not core-attributable.
+	 */
+	if (hwc->sample_period)
+		return -EINVAL;
+
+	/*
+	 * Task mode not available, we run the counters as socket counters,
+	 * not attributable to any CPU and therefore cannot attribute per-task.
+	 */
+	if (event->cpu < 0)
+		return -EINVAL;
+
+	/* Validate the group */
+	if (!qcom_l3_cache__validate_event_group(event))
+		return -EINVAL;
+
+	hwc->idx = -1;
+
+	/*
+	 * Many perf core operations (eg. events rotation) operate on a
+	 * single CPU context. This is obvious for CPU PMUs, where one
+	 * expects the same sets of events being observed on all CPUs,
+	 * but can lead to issues for off-core PMUs, like this one, where
+	 * each event could be theoretically assigned to a different CPU.
+	 * To mitigate this, we enforce CPU assignment to one designated
+	 * processor (the one described in the "cpumask" attribute exported
+	 * by the PMU device). perf user space tools honor this and avoid
+	 * opening more than one copy of the events.
+	 */
+	event->cpu = cpumask_first(&l3pmu->cpumask);
+
+	return 0;
+}
+
+static void qcom_l3_cache__event_start(struct perf_event *event, int flags)
+{
+	struct hw_perf_event *hwc = &event->hw;
+	const struct l3cache_event_ops *ops = l3cache_event_get_ops(event);
+
+	hwc->state = 0;
+	ops->start(event);
+}
+
+static void qcom_l3_cache__event_stop(struct perf_event *event, int flags)
+{
+	struct hw_perf_event *hwc = &event->hw;
+	const struct l3cache_event_ops *ops = l3cache_event_get_ops(event);
+
+	if (hwc->state & PERF_HES_STOPPED)
+		return;
+
+	ops->stop(event, flags);
+	if (flags & PERF_EF_UPDATE)
+		ops->update(event);
+	hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
+}
+
+static int qcom_l3_cache__event_add(struct perf_event *event, int flags)
+{
+	struct l3cache_pmu *l3pmu = to_l3cache_pmu(event->pmu);
+	struct hw_perf_event *hwc = &event->hw;
+	int order = event_uses_long_counter(event) ? 1 : 0;
+	int idx;
+
+	/*
+	 * Try to allocate a counter.
+	 */
+	idx = bitmap_find_free_region(l3pmu->used_mask, L3_NUM_COUNTERS, order);
+	if (idx < 0)
+		/* The counters are all in use. */
+		return -EAGAIN;
+
+	hwc->idx = idx;
+	hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
+	l3pmu->events[idx] = event;
+
+	if (flags & PERF_EF_START)
+		qcom_l3_cache__event_start(event, 0);
+
+	/* Propagate changes to the userspace mapping. */
+	perf_event_update_userpage(event);
+
+	return 0;
+}
+
+static void qcom_l3_cache__event_del(struct perf_event *event, int flags)
+{
+	struct l3cache_pmu *l3pmu = to_l3cache_pmu(event->pmu);
+	struct hw_perf_event *hwc = &event->hw;
+	int order = event_uses_long_counter(event) ? 1 : 0;
+
+	/* Stop and clean up */
+	qcom_l3_cache__event_stop(event,  flags | PERF_EF_UPDATE);
+	l3pmu->events[hwc->idx] = NULL;
+	bitmap_release_region(l3pmu->used_mask, hwc->idx, order);
+
+	/* Propagate changes to the userspace mapping. */
+	perf_event_update_userpage(event);
+}
+
+static void qcom_l3_cache__event_read(struct perf_event *event)
+{
+	const struct l3cache_event_ops *ops = l3cache_event_get_ops(event);
+
+	ops->update(event);
+}
+
+/*
+ * Add sysfs attributes
+ *
+ * We export:
+ * - formats, used by perf user space and other tools to configure events
+ * - events, used by perf user space and other tools to create events
+ *   symbolically, e.g.:
+ *     perf stat -a -e l3cache_0_0/event=read-miss/ ls
+ *     perf stat -a -e l3cache_0_0/event=0x21/ ls
+ * - cpumask, used by perf user space and other tools to know on which CPUs
+ *   to open the events
+ */
+
+/* formats */
+
+static ssize_t l3cache_pmu_format_show(struct device *dev,
+				       struct device_attribute *attr, char *buf)
+{
+	struct dev_ext_attribute *eattr;
+
+	eattr = container_of(attr, struct dev_ext_attribute, attr);
+	return sprintf(buf, "%s\n", (char *) eattr->var);
+}
+
+#define L3CACHE_PMU_FORMAT_ATTR(_name, _config)				      \
+	(&((struct dev_ext_attribute[]) {				      \
+		{ .attr = __ATTR(_name, 0444, l3cache_pmu_format_show, NULL), \
+		  .var = (void *) _config, }				      \
+	})[0].attr.attr)
+
+static struct attribute *qcom_l3_cache_pmu_formats[] = {
+	L3CACHE_PMU_FORMAT_ATTR(event, "config:0-7"),
+	L3CACHE_PMU_FORMAT_ATTR(lc, "config:" __stringify(L3_EVENT_LC_BIT)),
+	NULL,
+};
+
+static struct attribute_group qcom_l3_cache_pmu_format_group = {
+	.name = "format",
+	.attrs = qcom_l3_cache_pmu_formats,
+};
+
+/* events */
+
+static ssize_t l3cache_pmu_event_show(struct device *dev,
+				     struct device_attribute *attr, char *page)
+{
+	struct perf_pmu_events_attr *pmu_attr;
+
+	pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr);
+	return sprintf(page, "event=0x%02llx\n", pmu_attr->id);
+}
+
+#define L3CACHE_EVENT_ATTR(_name, _id)					     \
+	(&((struct perf_pmu_events_attr[]) {				     \
+		{ .attr = __ATTR(_name, 0444, l3cache_pmu_event_show, NULL), \
+		  .id = _id, }						     \
+	})[0].attr.attr)
+
+static struct attribute *qcom_l3_cache_pmu_events[] = {
+	L3CACHE_EVENT_ATTR(cycles, L3_EVENT_CYCLES),
+	L3CACHE_EVENT_ATTR(read-hit, L3_EVENT_READ_HIT),
+	L3CACHE_EVENT_ATTR(read-miss, L3_EVENT_READ_MISS),
+	L3CACHE_EVENT_ATTR(read-hit-d-side, L3_EVENT_READ_HIT_D),
+	L3CACHE_EVENT_ATTR(read-miss-d-side, L3_EVENT_READ_MISS_D),
+	L3CACHE_EVENT_ATTR(write-hit, L3_EVENT_WRITE_HIT),
+	L3CACHE_EVENT_ATTR(write-miss, L3_EVENT_WRITE_MISS),
+	NULL
+};
+
+static struct attribute_group qcom_l3_cache_pmu_events_group = {
+	.name = "events",
+	.attrs = qcom_l3_cache_pmu_events,
+};
+
+/* cpumask */
+
+static ssize_t qcom_l3_cache_pmu_cpumask_show(struct device *dev,
+				     struct device_attribute *attr, char *buf)
+{
+	struct l3cache_pmu *l3pmu = to_l3cache_pmu(dev_get_drvdata(dev));
+
+	return cpumap_print_to_pagebuf(true, buf, &l3pmu->cpumask);
+}
+
+static DEVICE_ATTR(cpumask, 0444, qcom_l3_cache_pmu_cpumask_show, NULL);
+
+static struct attribute *qcom_l3_cache_pmu_cpumask_attrs[] = {
+	&dev_attr_cpumask.attr,
+	NULL,
+};
+
+static struct attribute_group qcom_l3_cache_pmu_cpumask_attr_group = {
+	.attrs = qcom_l3_cache_pmu_cpumask_attrs,
+};
+
+/*
+ * Per PMU device attribute groups
+ */
+static const struct attribute_group *qcom_l3_cache_pmu_attr_grps[] = {
+	&qcom_l3_cache_pmu_format_group,
+	&qcom_l3_cache_pmu_events_group,
+	&qcom_l3_cache_pmu_cpumask_attr_group,
+	NULL,
+};
+
+/*
+ * Probing functions and data.
+ */
+
+static int qcom_l3_cache_pmu_online_cpu(unsigned int cpu, struct hlist_node *node)
+{
+	struct l3cache_pmu *l3pmu = hlist_entry_safe(node, struct l3cache_pmu, node);
+
+	/* If there is not a CPU/PMU association pick this CPU */
+	if (cpumask_empty(&l3pmu->cpumask))
+		cpumask_set_cpu(cpu, &l3pmu->cpumask);
+
+	return 0;
+}
+
+static int qcom_l3_cache_pmu_offline_cpu(unsigned int cpu, struct hlist_node *node)
+{
+	struct l3cache_pmu *l3pmu = hlist_entry_safe(node, struct l3cache_pmu, node);
+	unsigned int target;
+
+	if (!cpumask_test_and_clear_cpu(cpu, &l3pmu->cpumask))
+		return 0;
+	target = cpumask_any_but(cpu_online_mask, cpu);
+	if (target >= nr_cpu_ids)
+		return 0;
+	perf_pmu_migrate_context(&l3pmu->pmu, cpu, target);
+	cpumask_set_cpu(target, &l3pmu->cpumask);
+	return 0;
+}
+
+static int qcom_l3_cache_pmu_probe(struct platform_device *pdev)
+{
+	struct l3cache_pmu *l3pmu;
+	struct acpi_device *acpi_dev;
+	struct resource *memrc;
+	int ret;
+	char *name;
+
+	/* Initialize the PMU data structures */
+
+	acpi_dev = ACPI_COMPANION(&pdev->dev);
+	if (!acpi_dev)
+		return -ENODEV;
+
+	l3pmu = devm_kzalloc(&pdev->dev, sizeof(*l3pmu), GFP_KERNEL);
+	name = devm_kasprintf(&pdev->dev, GFP_KERNEL, "l3cache_%s_%s",
+		      acpi_dev->parent->pnp.unique_id, acpi_dev->pnp.unique_id);
+	if (!l3pmu || !name)
+		return -ENOMEM;
+
+	l3pmu->pmu = (struct pmu) {
+		.task_ctx_nr	= perf_invalid_context,
+
+		.pmu_enable	= qcom_l3_cache__pmu_enable,
+		.pmu_disable	= qcom_l3_cache__pmu_disable,
+		.event_init	= qcom_l3_cache__event_init,
+		.add		= qcom_l3_cache__event_add,
+		.del		= qcom_l3_cache__event_del,
+		.start		= qcom_l3_cache__event_start,
+		.stop		= qcom_l3_cache__event_stop,
+		.read		= qcom_l3_cache__event_read,
+
+		.attr_groups	= qcom_l3_cache_pmu_attr_grps,
+	};
+
+	memrc = platform_get_resource(pdev, IORESOURCE_MEM, 0);
+	l3pmu->regs = devm_ioremap_resource(&pdev->dev, memrc);
+	if (IS_ERR(l3pmu->regs)) {
+		dev_err(&pdev->dev, "Can't map PMU @%pa\n", &memrc->start);
+		return PTR_ERR(l3pmu->regs);
+	}
+
+	qcom_l3_cache__init(l3pmu);
+
+	ret = platform_get_irq(pdev, 0);
+	if (ret <= 0)
+		return ret;
+
+	ret = devm_request_irq(&pdev->dev, ret, qcom_l3_cache__handle_irq, 0,
+			       name, l3pmu);
+	if (ret) {
+		dev_err(&pdev->dev, "Request for IRQ failed for slice @%pa\n",
+			&memrc->start);
+		return ret;
+	}
+
+	/* Add this instance to the list used by the offline callback */
+	ret = cpuhp_state_add_instance(CPUHP_AP_PERF_ARM_QCOM_L3_ONLINE, &l3pmu->node);
+	if (ret) {
+		dev_err(&pdev->dev, "Error %d registering hotplug", ret);
+		return ret;
+	}
+
+	ret = perf_pmu_register(&l3pmu->pmu, name, -1);
+	if (ret < 0) {
+		dev_err(&pdev->dev, "Failed to register L3 cache PMU (%d)\n", ret);
+		return ret;
+	}
+
+	dev_info(&pdev->dev, "Registered %s, type: %d\n", name, l3pmu->pmu.type);
+
+	return 0;
+}
+
+static const struct acpi_device_id qcom_l3_cache_pmu_acpi_match[] = {
+	{ "QCOM8081", },
+	{ }
+};
+MODULE_DEVICE_TABLE(acpi, qcom_l3_cache_pmu_acpi_match);
+
+static struct platform_driver qcom_l3_cache_pmu_driver = {
+	.driver = {
+		.name = "qcom-l3cache-pmu",
+		.acpi_match_table = ACPI_PTR(qcom_l3_cache_pmu_acpi_match),
+	},
+	.probe = qcom_l3_cache_pmu_probe,
+};
+
+static int __init register_qcom_l3_cache_pmu_driver(void)
+{
+	int ret;
+
+	/* Install a hook to update the reader CPU in case it goes offline */
+	ret = cpuhp_setup_state_multi(CPUHP_AP_PERF_ARM_QCOM_L3_ONLINE,
+				      "perf/qcom/l3cache:online",
+				      qcom_l3_cache_pmu_online_cpu,
+				      qcom_l3_cache_pmu_offline_cpu);
+	if (ret)
+		return ret;
+
+	return platform_driver_register(&qcom_l3_cache_pmu_driver);
+}
+device_initcall(register_qcom_l3_cache_pmu_driver);

include/linux/cpuhotplug.h

@@ -137,6 +137,7 @@ enum cpuhp_state {
 	CPUHP_AP_PERF_ARM_CCN_ONLINE,
 	CPUHP_AP_PERF_ARM_L2X0_ONLINE,
 	CPUHP_AP_PERF_ARM_QCOM_L2_ONLINE,
+	CPUHP_AP_PERF_ARM_QCOM_L3_ONLINE,
 	CPUHP_AP_WORKQUEUE_ONLINE,
 	CPUHP_AP_RCUTREE_ONLINE,
 	CPUHP_AP_ONLINE_DYN,