perf: xgene: Add APM X-Gene SoC Performance Monitoring Unit driver
authorTai Nguyen <ttnguyen@apm.com>
Fri, 15 Jul 2016 17:38:04 +0000 (10:38 -0700)
committerDuc Dang <dhdang@apm.com>
Thu, 15 Sep 2016 18:20:55 +0000 (11:20 -0700)
This patch adds a driver for the SoC-wide (AKA uncore) PMU hardware
found in APM X-Gene SoCs.

Signed-off-by: Tai Nguyen <ttnguyen@apm.com>
Reviewed-by: Mark Rutland <mark.rutland@arm.com>
Documentation/perf/xgene-pmu.txt [new file with mode: 0644]
drivers/perf/Kconfig
drivers/perf/Makefile
drivers/perf/xgene_pmu.c [new file with mode: 0644]

diff --git a/Documentation/perf/xgene-pmu.txt b/Documentation/perf/xgene-pmu.txt
new file mode 100644 (file)
index 0000000..d7cff44
--- /dev/null
@@ -0,0 +1,48 @@
+APM X-Gene SoC Performance Monitoring Unit (PMU)
+================================================
+
+X-Gene SoC PMU consists of various independent system device PMUs such as
+L3 cache(s), I/O bridge(s), memory controller bridge(s) and memory
+controller(s). These PMU devices are loosely architected to follow the
+same model as the PMU for ARM cores. The PMUs share the same top level
+interrupt and status CSR region.
+
+PMU (perf) driver
+-----------------
+
+The xgene-pmu driver registers several perf PMU drivers. Each of the perf
+driver provides description of its available events and configuration options
+in sysfs, see /sys/devices/<l3cX/iobX/mcbX/mcX>/.
+
+The "format" directory describes format of the config (event ID),
+config1 (agent ID) fields of the perf_event_attr structure. The "events"
+directory provides configuration templates for all supported event types that
+can be used with perf tool. For example, "l3c0/bank-fifo-full/" is an
+equivalent of "l3c0/config=0x0b/".
+
+Most of the SoC PMU has a specific list of agent ID used for monitoring
+performance of a specific datapath. For example, agents of a L3 cache can be
+a specific CPU or an I/O bridge. Each PMU has a set of 2 registers capable of
+masking the agents from which the request come from. If the bit with
+the bit number corresponding to the agent is set, the event is counted only if
+it is caused by a request from that agent. Each agent ID bit is inversely mapped
+to a corresponding bit in "config1" field. By default, the event will be
+counted for all agent requests (config1 = 0x0). For all the supported agents of
+each PMU, please refer to APM X-Gene User Manual.
+
+Each perf driver also provides a "cpumask" sysfs attribute, which contains a
+single CPU ID of the processor which will be used to handle all the PMU events.
+
+Example for perf tool use:
+
+ / # perf list | grep -e l3c -e iob -e mcb -e mc
+   l3c0/ackq-full/                                    [Kernel PMU event]
+ <...>
+   mcb1/mcb-csw-stall/                                [Kernel PMU event]
+
+ / # perf stat -a -e l3c0/read-miss/,mcb1/csw-write-request/ sleep 1
+
+ / # perf stat -a -e l3c0/read-miss,config1=0xfffffffffffffffe/ sleep 1
+
+The driver does not support sampling, therefore "perf record" will
+not work. Per-task (without "-a") perf sessions are not supported.
index 04e2653bb8c02cd9aede61bca14c9638da62c432..4d5c5f9f0dbd4ac337d6e5063dc5cfd7207d1292 100644 (file)
@@ -12,4 +12,11 @@ config ARM_PMU
          Say y if you want to use CPU performance monitors on ARM-based
          systems.
 
+config XGENE_PMU
+        depends on PERF_EVENTS && ARCH_XGENE
+        bool "APM X-Gene SoC PMU"
+        default n
+        help
+          Say y if you want to use APM X-Gene SoC performance monitors.
+
 endmenu
index acd2397ded94dda71cbb0b8db5c82dde80d72803..b116e982810b202880d00d98f09775898f17f3f1 100644 (file)
@@ -1 +1,2 @@
 obj-$(CONFIG_ARM_PMU) += arm_pmu.o
+obj-$(CONFIG_XGENE_PMU) += xgene_pmu.o
diff --git a/drivers/perf/xgene_pmu.c b/drivers/perf/xgene_pmu.c
new file mode 100644 (file)
index 0000000..c2ac764
--- /dev/null
@@ -0,0 +1,1398 @@
+/*
+ * APM X-Gene SoC PMU (Performance Monitor Unit)
+ *
+ * Copyright (c) 2016, Applied Micro Circuits Corporation
+ * Author: Hoan Tran <hotran@apm.com>
+ *         Tai Nguyen <ttnguyen@apm.com>
+ *
+ * This program is free software; you can redistribute  it and/or modify it
+ * under  the terms of  the GNU General  Public License as published by the
+ * Free Software Foundation;  either version 2 of the  License, or (at your
+ * option) any later version.
+ *
+ * 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.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program.  If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#include <linux/acpi.h>
+#include <linux/clk.h>
+#include <linux/cpumask.h>
+#include <linux/interrupt.h>
+#include <linux/io.h>
+#include <linux/mfd/syscon.h>
+#include <linux/of_address.h>
+#include <linux/of_fdt.h>
+#include <linux/of_irq.h>
+#include <linux/of_platform.h>
+#include <linux/perf_event.h>
+#include <linux/platform_device.h>
+#include <linux/regmap.h>
+#include <linux/slab.h>
+
+#define CSW_CSWCR                       0x0000
+#define  CSW_CSWCR_DUALMCB_MASK         BIT(0)
+#define MCBADDRMR                       0x0000
+#define  MCBADDRMR_DUALMCU_MODE_MASK    BIT(2)
+
+#define PCPPMU_INTSTATUS_REG   0x000
+#define PCPPMU_INTMASK_REG     0x004
+#define  PCPPMU_INTMASK                0x0000000F
+#define  PCPPMU_INTENMASK      0xFFFFFFFF
+#define  PCPPMU_INTCLRMASK     0xFFFFFFF0
+#define  PCPPMU_INT_MCU                BIT(0)
+#define  PCPPMU_INT_MCB                BIT(1)
+#define  PCPPMU_INT_L3C                BIT(2)
+#define  PCPPMU_INT_IOB                BIT(3)
+
+#define PMU_MAX_COUNTERS       4
+#define PMU_CNT_MAX_PERIOD     0x100000000ULL
+#define PMU_OVERFLOW_MASK      0xF
+#define PMU_PMCR_E             BIT(0)
+#define PMU_PMCR_P             BIT(1)
+
+#define PMU_PMEVCNTR0          0x000
+#define PMU_PMEVCNTR1          0x004
+#define PMU_PMEVCNTR2          0x008
+#define PMU_PMEVCNTR3          0x00C
+#define PMU_PMEVTYPER0         0x400
+#define PMU_PMEVTYPER1         0x404
+#define PMU_PMEVTYPER2         0x408
+#define PMU_PMEVTYPER3         0x40C
+#define PMU_PMAMR0             0xA00
+#define PMU_PMAMR1             0xA04
+#define PMU_PMCNTENSET         0xC00
+#define PMU_PMCNTENCLR         0xC20
+#define PMU_PMINTENSET         0xC40
+#define PMU_PMINTENCLR         0xC60
+#define PMU_PMOVSR             0xC80
+#define PMU_PMCR               0xE04
+
+#define to_pmu_dev(p)     container_of(p, struct xgene_pmu_dev, pmu)
+#define GET_CNTR(ev)      (ev->hw.idx)
+#define GET_EVENTID(ev)   (ev->hw.config & 0xFFULL)
+#define GET_AGENTID(ev)   (ev->hw.config_base & 0xFFFFFFFFUL)
+#define GET_AGENT1ID(ev)  ((ev->hw.config_base >> 32) & 0xFFFFFFFFUL)
+
+struct hw_pmu_info {
+       u32 type;
+       u32 enable_mask;
+       void __iomem *csr;
+};
+
+struct xgene_pmu_dev {
+       struct hw_pmu_info *inf;
+       struct xgene_pmu *parent;
+       struct pmu pmu;
+       u8 max_counters;
+       DECLARE_BITMAP(cntr_assign_mask, PMU_MAX_COUNTERS);
+       u64 max_period;
+       const struct attribute_group **attr_groups;
+       struct perf_event *pmu_counter_event[PMU_MAX_COUNTERS];
+};
+
+struct xgene_pmu {
+       struct device *dev;
+       int version;
+       void __iomem *pcppmu_csr;
+       u32 mcb_active_mask;
+       u32 mc_active_mask;
+       cpumask_t cpu;
+       raw_spinlock_t lock;
+       struct list_head l3cpmus;
+       struct list_head iobpmus;
+       struct list_head mcbpmus;
+       struct list_head mcpmus;
+};
+
+struct xgene_pmu_dev_ctx {
+       char *name;
+       struct list_head next;
+       struct xgene_pmu_dev *pmu_dev;
+       struct hw_pmu_info inf;
+};
+
+struct xgene_pmu_data {
+       int id;
+       u32 data;
+};
+
+enum xgene_pmu_version {
+       PCP_PMU_V1 = 1,
+       PCP_PMU_V2,
+};
+
+enum xgene_pmu_dev_type {
+       PMU_TYPE_L3C = 0,
+       PMU_TYPE_IOB,
+       PMU_TYPE_MCB,
+       PMU_TYPE_MC,
+};
+
+/*
+ * sysfs format attributes
+ */
+static ssize_t xgene_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 XGENE_PMU_FORMAT_ATTR(_name, _config)          \
+       (&((struct dev_ext_attribute[]) {               \
+               { .attr = __ATTR(_name, S_IRUGO, xgene_pmu_format_show, NULL), \
+                 .var = (void *) _config, }            \
+       })[0].attr.attr)
+
+static struct attribute *l3c_pmu_format_attrs[] = {
+       XGENE_PMU_FORMAT_ATTR(l3c_eventid, "config:0-7"),
+       XGENE_PMU_FORMAT_ATTR(l3c_agentid, "config1:0-9"),
+       NULL,
+};
+
+static struct attribute *iob_pmu_format_attrs[] = {
+       XGENE_PMU_FORMAT_ATTR(iob_eventid, "config:0-7"),
+       XGENE_PMU_FORMAT_ATTR(iob_agentid, "config1:0-63"),
+       NULL,
+};
+
+static struct attribute *mcb_pmu_format_attrs[] = {
+       XGENE_PMU_FORMAT_ATTR(mcb_eventid, "config:0-5"),
+       XGENE_PMU_FORMAT_ATTR(mcb_agentid, "config1:0-9"),
+       NULL,
+};
+
+static struct attribute *mc_pmu_format_attrs[] = {
+       XGENE_PMU_FORMAT_ATTR(mc_eventid, "config:0-28"),
+       NULL,
+};
+
+static const struct attribute_group l3c_pmu_format_attr_group = {
+       .name = "format",
+       .attrs = l3c_pmu_format_attrs,
+};
+
+static const struct attribute_group iob_pmu_format_attr_group = {
+       .name = "format",
+       .attrs = iob_pmu_format_attrs,
+};
+
+static const struct attribute_group mcb_pmu_format_attr_group = {
+       .name = "format",
+       .attrs = mcb_pmu_format_attrs,
+};
+
+static const struct attribute_group mc_pmu_format_attr_group = {
+       .name = "format",
+       .attrs = mc_pmu_format_attrs,
+};
+
+/*
+ * sysfs event attributes
+ */
+static ssize_t xgene_pmu_event_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, "config=0x%lx\n", (unsigned long) eattr->var);
+}
+
+#define XGENE_PMU_EVENT_ATTR(_name, _config)           \
+       (&((struct dev_ext_attribute[]) {               \
+               { .attr = __ATTR(_name, S_IRUGO, xgene_pmu_event_show, NULL), \
+                 .var = (void *) _config, }            \
+        })[0].attr.attr)
+
+static struct attribute *l3c_pmu_events_attrs[] = {
+       XGENE_PMU_EVENT_ATTR(cycle-count,                       0x00),
+       XGENE_PMU_EVENT_ATTR(cycle-count-div-64,                0x01),
+       XGENE_PMU_EVENT_ATTR(read-hit,                          0x02),
+       XGENE_PMU_EVENT_ATTR(read-miss,                         0x03),
+       XGENE_PMU_EVENT_ATTR(write-need-replacement,            0x06),
+       XGENE_PMU_EVENT_ATTR(write-not-need-replacement,        0x07),
+       XGENE_PMU_EVENT_ATTR(tq-full,                           0x08),
+       XGENE_PMU_EVENT_ATTR(ackq-full,                         0x09),
+       XGENE_PMU_EVENT_ATTR(wdb-full,                          0x0a),
+       XGENE_PMU_EVENT_ATTR(bank-fifo-full,                    0x0b),
+       XGENE_PMU_EVENT_ATTR(odb-full,                          0x0c),
+       XGENE_PMU_EVENT_ATTR(wbq-full,                          0x0d),
+       XGENE_PMU_EVENT_ATTR(bank-conflict-fifo-issue,          0x0e),
+       XGENE_PMU_EVENT_ATTR(bank-fifo-issue,                   0x0f),
+       NULL,
+};
+
+static struct attribute *iob_pmu_events_attrs[] = {
+       XGENE_PMU_EVENT_ATTR(cycle-count,                       0x00),
+       XGENE_PMU_EVENT_ATTR(cycle-count-div-64,                0x01),
+       XGENE_PMU_EVENT_ATTR(axi0-read,                         0x02),
+       XGENE_PMU_EVENT_ATTR(axi0-read-partial,                 0x03),
+       XGENE_PMU_EVENT_ATTR(axi1-read,                         0x04),
+       XGENE_PMU_EVENT_ATTR(axi1-read-partial,                 0x05),
+       XGENE_PMU_EVENT_ATTR(csw-read-block,                    0x06),
+       XGENE_PMU_EVENT_ATTR(csw-read-partial,                  0x07),
+       XGENE_PMU_EVENT_ATTR(axi0-write,                        0x10),
+       XGENE_PMU_EVENT_ATTR(axi0-write-partial,                0x11),
+       XGENE_PMU_EVENT_ATTR(axi1-write,                        0x13),
+       XGENE_PMU_EVENT_ATTR(axi1-write-partial,                0x14),
+       XGENE_PMU_EVENT_ATTR(csw-inbound-dirty,                 0x16),
+       NULL,
+};
+
+static struct attribute *mcb_pmu_events_attrs[] = {
+       XGENE_PMU_EVENT_ATTR(cycle-count,                       0x00),
+       XGENE_PMU_EVENT_ATTR(cycle-count-div-64,                0x01),
+       XGENE_PMU_EVENT_ATTR(csw-read,                          0x02),
+       XGENE_PMU_EVENT_ATTR(csw-write-request,                 0x03),
+       XGENE_PMU_EVENT_ATTR(mcb-csw-stall,                     0x04),
+       XGENE_PMU_EVENT_ATTR(cancel-read-gack,                  0x05),
+       NULL,
+};
+
+static struct attribute *mc_pmu_events_attrs[] = {
+       XGENE_PMU_EVENT_ATTR(cycle-count,                       0x00),
+       XGENE_PMU_EVENT_ATTR(cycle-count-div-64,                0x01),
+       XGENE_PMU_EVENT_ATTR(act-cmd-sent,                      0x02),
+       XGENE_PMU_EVENT_ATTR(pre-cmd-sent,                      0x03),
+       XGENE_PMU_EVENT_ATTR(rd-cmd-sent,                       0x04),
+       XGENE_PMU_EVENT_ATTR(rda-cmd-sent,                      0x05),
+       XGENE_PMU_EVENT_ATTR(wr-cmd-sent,                       0x06),
+       XGENE_PMU_EVENT_ATTR(wra-cmd-sent,                      0x07),
+       XGENE_PMU_EVENT_ATTR(pde-cmd-sent,                      0x08),
+       XGENE_PMU_EVENT_ATTR(sre-cmd-sent,                      0x09),
+       XGENE_PMU_EVENT_ATTR(prea-cmd-sent,                     0x0a),
+       XGENE_PMU_EVENT_ATTR(ref-cmd-sent,                      0x0b),
+       XGENE_PMU_EVENT_ATTR(rd-rda-cmd-sent,                   0x0c),
+       XGENE_PMU_EVENT_ATTR(wr-wra-cmd-sent,                   0x0d),
+       XGENE_PMU_EVENT_ATTR(in-rd-collision,                   0x0e),
+       XGENE_PMU_EVENT_ATTR(in-wr-collision,                   0x0f),
+       XGENE_PMU_EVENT_ATTR(collision-queue-not-empty,         0x10),
+       XGENE_PMU_EVENT_ATTR(collision-queue-full,              0x11),
+       XGENE_PMU_EVENT_ATTR(mcu-request,                       0x12),
+       XGENE_PMU_EVENT_ATTR(mcu-rd-request,                    0x13),
+       XGENE_PMU_EVENT_ATTR(mcu-hp-rd-request,                 0x14),
+       XGENE_PMU_EVENT_ATTR(mcu-wr-request,                    0x15),
+       XGENE_PMU_EVENT_ATTR(mcu-rd-proceed-all,                0x16),
+       XGENE_PMU_EVENT_ATTR(mcu-rd-proceed-cancel,             0x17),
+       XGENE_PMU_EVENT_ATTR(mcu-rd-response,                   0x18),
+       XGENE_PMU_EVENT_ATTR(mcu-rd-proceed-speculative-all,    0x19),
+       XGENE_PMU_EVENT_ATTR(mcu-rd-proceed-speculative-cancel, 0x1a),
+       XGENE_PMU_EVENT_ATTR(mcu-wr-proceed-all,                0x1b),
+       XGENE_PMU_EVENT_ATTR(mcu-wr-proceed-cancel,             0x1c),
+       NULL,
+};
+
+static const struct attribute_group l3c_pmu_events_attr_group = {
+       .name = "events",
+       .attrs = l3c_pmu_events_attrs,
+};
+
+static const struct attribute_group iob_pmu_events_attr_group = {
+       .name = "events",
+       .attrs = iob_pmu_events_attrs,
+};
+
+static const struct attribute_group mcb_pmu_events_attr_group = {
+       .name = "events",
+       .attrs = mcb_pmu_events_attrs,
+};
+
+static const struct attribute_group mc_pmu_events_attr_group = {
+       .name = "events",
+       .attrs = mc_pmu_events_attrs,
+};
+
+/*
+ * sysfs cpumask attributes
+ */
+static ssize_t xgene_pmu_cpumask_show(struct device *dev,
+                                     struct device_attribute *attr, char *buf)
+{
+       struct xgene_pmu_dev *pmu_dev = to_pmu_dev(dev_get_drvdata(dev));
+
+       return cpumap_print_to_pagebuf(true, buf, &pmu_dev->parent->cpu);
+}
+
+static DEVICE_ATTR(cpumask, S_IRUGO, xgene_pmu_cpumask_show, NULL);
+
+static struct attribute *xgene_pmu_cpumask_attrs[] = {
+       &dev_attr_cpumask.attr,
+       NULL,
+};
+
+static const struct attribute_group pmu_cpumask_attr_group = {
+       .attrs = xgene_pmu_cpumask_attrs,
+};
+
+/*
+ * Per PMU device attribute groups
+ */
+static const struct attribute_group *l3c_pmu_attr_groups[] = {
+       &l3c_pmu_format_attr_group,
+       &pmu_cpumask_attr_group,
+       &l3c_pmu_events_attr_group,
+       NULL
+};
+
+static const struct attribute_group *iob_pmu_attr_groups[] = {
+       &iob_pmu_format_attr_group,
+       &pmu_cpumask_attr_group,
+       &iob_pmu_events_attr_group,
+       NULL
+};
+
+static const struct attribute_group *mcb_pmu_attr_groups[] = {
+       &mcb_pmu_format_attr_group,
+       &pmu_cpumask_attr_group,
+       &mcb_pmu_events_attr_group,
+       NULL
+};
+
+static const struct attribute_group *mc_pmu_attr_groups[] = {
+       &mc_pmu_format_attr_group,
+       &pmu_cpumask_attr_group,
+       &mc_pmu_events_attr_group,
+       NULL
+};
+
+static int get_next_avail_cntr(struct xgene_pmu_dev *pmu_dev)
+{
+       int cntr;
+
+       cntr = find_first_zero_bit(pmu_dev->cntr_assign_mask,
+                               pmu_dev->max_counters);
+       if (cntr == pmu_dev->max_counters)
+               return -ENOSPC;
+       set_bit(cntr, pmu_dev->cntr_assign_mask);
+
+       return cntr;
+}
+
+static void clear_avail_cntr(struct xgene_pmu_dev *pmu_dev, int cntr)
+{
+       clear_bit(cntr, pmu_dev->cntr_assign_mask);
+}
+
+static inline void xgene_pmu_mask_int(struct xgene_pmu *xgene_pmu)
+{
+       writel(PCPPMU_INTENMASK, xgene_pmu->pcppmu_csr + PCPPMU_INTMASK_REG);
+}
+
+static inline void xgene_pmu_unmask_int(struct xgene_pmu *xgene_pmu)
+{
+       writel(PCPPMU_INTCLRMASK, xgene_pmu->pcppmu_csr + PCPPMU_INTMASK_REG);
+}
+
+static inline u32 xgene_pmu_read_counter(struct xgene_pmu_dev *pmu_dev, int idx)
+{
+       return readl(pmu_dev->inf->csr + PMU_PMEVCNTR0 + (4 * idx));
+}
+
+static inline void
+xgene_pmu_write_counter(struct xgene_pmu_dev *pmu_dev, int idx, u32 val)
+{
+       writel(val, pmu_dev->inf->csr + PMU_PMEVCNTR0 + (4 * idx));
+}
+
+static inline void
+xgene_pmu_write_evttype(struct xgene_pmu_dev *pmu_dev, int idx, u32 val)
+{
+       writel(val, pmu_dev->inf->csr + PMU_PMEVTYPER0 + (4 * idx));
+}
+
+static inline void
+xgene_pmu_write_agentmsk(struct xgene_pmu_dev *pmu_dev, u32 val)
+{
+       writel(val, pmu_dev->inf->csr + PMU_PMAMR0);
+}
+
+static inline void
+xgene_pmu_write_agent1msk(struct xgene_pmu_dev *pmu_dev, u32 val)
+{
+       writel(val, pmu_dev->inf->csr + PMU_PMAMR1);
+}
+
+static inline void
+xgene_pmu_enable_counter(struct xgene_pmu_dev *pmu_dev, int idx)
+{
+       u32 val;
+
+       val = readl(pmu_dev->inf->csr + PMU_PMCNTENSET);
+       val |= 1 << idx;
+       writel(val, pmu_dev->inf->csr + PMU_PMCNTENSET);
+}
+
+static inline void
+xgene_pmu_disable_counter(struct xgene_pmu_dev *pmu_dev, int idx)
+{
+       u32 val;
+
+       val = readl(pmu_dev->inf->csr + PMU_PMCNTENCLR);
+       val |= 1 << idx;
+       writel(val, pmu_dev->inf->csr + PMU_PMCNTENCLR);
+}
+
+static inline void
+xgene_pmu_enable_counter_int(struct xgene_pmu_dev *pmu_dev, int idx)
+{
+       u32 val;
+
+       val = readl(pmu_dev->inf->csr + PMU_PMINTENSET);
+       val |= 1 << idx;
+       writel(val, pmu_dev->inf->csr + PMU_PMINTENSET);
+}
+
+static inline void
+xgene_pmu_disable_counter_int(struct xgene_pmu_dev *pmu_dev, int idx)
+{
+       u32 val;
+
+       val = readl(pmu_dev->inf->csr + PMU_PMINTENCLR);
+       val |= 1 << idx;
+       writel(val, pmu_dev->inf->csr + PMU_PMINTENCLR);
+}
+
+static inline void xgene_pmu_reset_counters(struct xgene_pmu_dev *pmu_dev)
+{
+       u32 val;
+
+       val = readl(pmu_dev->inf->csr + PMU_PMCR);
+       val |= PMU_PMCR_P;
+       writel(val, pmu_dev->inf->csr + PMU_PMCR);
+}
+
+static inline void xgene_pmu_start_counters(struct xgene_pmu_dev *pmu_dev)
+{
+       u32 val;
+
+       val = readl(pmu_dev->inf->csr + PMU_PMCR);
+       val |= PMU_PMCR_E;
+       writel(val, pmu_dev->inf->csr + PMU_PMCR);
+}
+
+static inline void xgene_pmu_stop_counters(struct xgene_pmu_dev *pmu_dev)
+{
+       u32 val;
+
+       val = readl(pmu_dev->inf->csr + PMU_PMCR);
+       val &= ~PMU_PMCR_E;
+       writel(val, pmu_dev->inf->csr + PMU_PMCR);
+}
+
+static void xgene_perf_pmu_enable(struct pmu *pmu)
+{
+       struct xgene_pmu_dev *pmu_dev = to_pmu_dev(pmu);
+       int enabled = bitmap_weight(pmu_dev->cntr_assign_mask,
+                       pmu_dev->max_counters);
+
+       if (!enabled)
+               return;
+
+       xgene_pmu_start_counters(pmu_dev);
+}
+
+static void xgene_perf_pmu_disable(struct pmu *pmu)
+{
+       struct xgene_pmu_dev *pmu_dev = to_pmu_dev(pmu);
+
+       xgene_pmu_stop_counters(pmu_dev);
+}
+
+static int xgene_perf_event_init(struct perf_event *event)
+{
+       struct xgene_pmu_dev *pmu_dev = to_pmu_dev(event->pmu);
+       struct hw_perf_event *hw = &event->hw;
+       struct perf_event *sibling;
+
+       /* Test the event attr type check for PMU enumeration */
+       if (event->attr.type != event->pmu->type)
+               return -ENOENT;
+
+       /*
+        * SOC PMU counters are shared across all cores.
+        * Therefore, it does not support per-process mode.
+        * Also, it does not support event sampling mode.
+        */
+       if (is_sampling_event(event) || event->attach_state & PERF_ATTACH_TASK)
+               return -EINVAL;
+
+       /* SOC counters do not have usr/os/guest/host bits */
+       if (event->attr.exclude_user || event->attr.exclude_kernel ||
+           event->attr.exclude_host || event->attr.exclude_guest)
+               return -EINVAL;
+
+       if (event->cpu < 0)
+               return -EINVAL;
+       /*
+        * 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, where each
+        * event could be theoretically assigned to a different CPU. To
+        * mitigate this, we enforce CPU assignment to one, selected
+        * processor (the one described in the "cpumask" attribute).
+        */
+       event->cpu = cpumask_first(&pmu_dev->parent->cpu);
+
+       hw->config = event->attr.config;
+       /*
+        * Each bit of the config1 field represents an agent from which the
+        * request of the event come. The event is counted only if it's caused
+        * by a request of an agent has the bit cleared.
+        * By default, the event is counted for all agents.
+        */
+       hw->config_base = event->attr.config1;
+
+       /*
+        * We must NOT create groups containing mixed PMUs, although software
+        * events are acceptable
+        */
+       if (event->group_leader->pmu != event->pmu &&
+                       !is_software_event(event->group_leader))
+               return -EINVAL;
+
+       list_for_each_entry(sibling, &event->group_leader->sibling_list,
+                       group_entry)
+               if (sibling->pmu != event->pmu &&
+                               !is_software_event(sibling))
+                       return -EINVAL;
+
+       return 0;
+}
+
+static void xgene_perf_enable_event(struct perf_event *event)
+{
+       struct xgene_pmu_dev *pmu_dev = to_pmu_dev(event->pmu);
+
+       xgene_pmu_write_evttype(pmu_dev, GET_CNTR(event), GET_EVENTID(event));
+       xgene_pmu_write_agentmsk(pmu_dev, ~((u32)GET_AGENTID(event)));
+       if (pmu_dev->inf->type == PMU_TYPE_IOB)
+               xgene_pmu_write_agent1msk(pmu_dev, ~((u32)GET_AGENT1ID(event)));
+
+       xgene_pmu_enable_counter(pmu_dev, GET_CNTR(event));
+       xgene_pmu_enable_counter_int(pmu_dev, GET_CNTR(event));
+}
+
+static void xgene_perf_disable_event(struct perf_event *event)
+{
+       struct xgene_pmu_dev *pmu_dev = to_pmu_dev(event->pmu);
+
+       xgene_pmu_disable_counter(pmu_dev, GET_CNTR(event));
+       xgene_pmu_disable_counter_int(pmu_dev, GET_CNTR(event));
+}
+
+static void xgene_perf_event_set_period(struct perf_event *event)
+{
+       struct xgene_pmu_dev *pmu_dev = to_pmu_dev(event->pmu);
+       struct hw_perf_event *hw = &event->hw;
+       /*
+        * The X-Gene PMU counters have a period of 2^32. To account for the
+        * possiblity of extreme interrupt latency we program for a period of
+        * half that. Hopefully we can handle the interrupt before another 2^31
+        * events occur and the counter overtakes its previous value.
+        */
+       u64 val = 1ULL << 31;
+
+       local64_set(&hw->prev_count, val);
+       xgene_pmu_write_counter(pmu_dev, hw->idx, (u32) val);
+}
+
+static void xgene_perf_event_update(struct perf_event *event)
+{
+       struct xgene_pmu_dev *pmu_dev = to_pmu_dev(event->pmu);
+       struct hw_perf_event *hw = &event->hw;
+       u64 delta, prev_raw_count, new_raw_count;
+
+again:
+       prev_raw_count = local64_read(&hw->prev_count);
+       new_raw_count = xgene_pmu_read_counter(pmu_dev, GET_CNTR(event));
+
+       if (local64_cmpxchg(&hw->prev_count, prev_raw_count,
+                           new_raw_count) != prev_raw_count)
+               goto again;
+
+       delta = (new_raw_count - prev_raw_count) & pmu_dev->max_period;
+
+       local64_add(delta, &event->count);
+}
+
+static void xgene_perf_read(struct perf_event *event)
+{
+       xgene_perf_event_update(event);
+}
+
+static void xgene_perf_start(struct perf_event *event, int flags)
+{
+       struct xgene_pmu_dev *pmu_dev = to_pmu_dev(event->pmu);
+       struct hw_perf_event *hw = &event->hw;
+
+       if (WARN_ON_ONCE(!(hw->state & PERF_HES_STOPPED)))
+               return;
+
+       WARN_ON_ONCE(!(hw->state & PERF_HES_UPTODATE));
+       hw->state = 0;
+
+       xgene_perf_event_set_period(event);
+
+       if (flags & PERF_EF_RELOAD) {
+               u64 prev_raw_count =  local64_read(&hw->prev_count);
+
+               xgene_pmu_write_counter(pmu_dev, GET_CNTR(event),
+                                       (u32) prev_raw_count);
+       }
+
+       xgene_perf_enable_event(event);
+       perf_event_update_userpage(event);
+}
+
+static void xgene_perf_stop(struct perf_event *event, int flags)
+{
+       struct hw_perf_event *hw = &event->hw;
+       u64 config;
+
+       if (hw->state & PERF_HES_UPTODATE)
+               return;
+
+       xgene_perf_disable_event(event);
+       WARN_ON_ONCE(hw->state & PERF_HES_STOPPED);
+       hw->state |= PERF_HES_STOPPED;
+
+       if (hw->state & PERF_HES_UPTODATE)
+               return;
+
+       config = hw->config;
+       xgene_perf_read(event);
+       hw->state |= PERF_HES_UPTODATE;
+}
+
+static int xgene_perf_add(struct perf_event *event, int flags)
+{
+       struct xgene_pmu_dev *pmu_dev = to_pmu_dev(event->pmu);
+       struct hw_perf_event *hw = &event->hw;
+
+       hw->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
+
+       /* Allocate an event counter */
+       hw->idx = get_next_avail_cntr(pmu_dev);
+       if (hw->idx < 0)
+               return -EAGAIN;
+
+       /* Update counter event pointer for Interrupt handler */
+       pmu_dev->pmu_counter_event[hw->idx] = event;
+
+       if (flags & PERF_EF_START)
+               xgene_perf_start(event, PERF_EF_RELOAD);
+
+       return 0;
+}
+
+static void xgene_perf_del(struct perf_event *event, int flags)
+{
+       struct xgene_pmu_dev *pmu_dev = to_pmu_dev(event->pmu);
+       struct hw_perf_event *hw = &event->hw;
+
+       xgene_perf_stop(event, PERF_EF_UPDATE);
+
+       /* clear the assigned counter */
+       clear_avail_cntr(pmu_dev, GET_CNTR(event));
+
+       perf_event_update_userpage(event);
+       pmu_dev->pmu_counter_event[hw->idx] = NULL;
+}
+
+static int xgene_init_perf(struct xgene_pmu_dev *pmu_dev, char *name)
+{
+       struct xgene_pmu *xgene_pmu;
+
+       pmu_dev->max_period = PMU_CNT_MAX_PERIOD - 1;
+       /* First version PMU supports only single event counter */
+       xgene_pmu = pmu_dev->parent;
+       if (xgene_pmu->version == PCP_PMU_V1)
+               pmu_dev->max_counters = 1;
+       else
+               pmu_dev->max_counters = PMU_MAX_COUNTERS;
+
+       /* Perf driver registration */
+       pmu_dev->pmu = (struct pmu) {
+               .attr_groups    = pmu_dev->attr_groups,
+               .task_ctx_nr    = perf_invalid_context,
+               .pmu_enable     = xgene_perf_pmu_enable,
+               .pmu_disable    = xgene_perf_pmu_disable,
+               .event_init     = xgene_perf_event_init,
+               .add            = xgene_perf_add,
+               .del            = xgene_perf_del,
+               .start          = xgene_perf_start,
+               .stop           = xgene_perf_stop,
+               .read           = xgene_perf_read,
+       };
+
+       /* Hardware counter init */
+       xgene_pmu_stop_counters(pmu_dev);
+       xgene_pmu_reset_counters(pmu_dev);
+
+       return perf_pmu_register(&pmu_dev->pmu, name, -1);
+}
+
+static int
+xgene_pmu_dev_add(struct xgene_pmu *xgene_pmu, struct xgene_pmu_dev_ctx *ctx)
+{
+       struct device *dev = xgene_pmu->dev;
+       struct xgene_pmu_dev *pmu;
+       int rc;
+
+       pmu = devm_kzalloc(dev, sizeof(*pmu), GFP_KERNEL);
+       if (!pmu)
+               return -ENOMEM;
+       pmu->parent = xgene_pmu;
+       pmu->inf = &ctx->inf;
+       ctx->pmu_dev = pmu;
+
+       switch (pmu->inf->type) {
+       case PMU_TYPE_L3C:
+               pmu->attr_groups = l3c_pmu_attr_groups;
+               break;
+       case PMU_TYPE_IOB:
+               pmu->attr_groups = iob_pmu_attr_groups;
+               break;
+       case PMU_TYPE_MCB:
+               if (!(xgene_pmu->mcb_active_mask & pmu->inf->enable_mask))
+                       goto dev_err;
+               pmu->attr_groups = mcb_pmu_attr_groups;
+               break;
+       case PMU_TYPE_MC:
+               if (!(xgene_pmu->mc_active_mask & pmu->inf->enable_mask))
+                       goto dev_err;
+               pmu->attr_groups = mc_pmu_attr_groups;
+               break;
+       default:
+               return -EINVAL;
+       }
+
+       rc = xgene_init_perf(pmu, ctx->name);
+       if (rc) {
+               dev_err(dev, "%s PMU: Failed to init perf driver\n", ctx->name);
+               goto dev_err;
+       }
+
+       dev_info(dev, "%s PMU registered\n", ctx->name);
+
+       return rc;
+
+dev_err:
+       devm_kfree(dev, pmu);
+       return -ENODEV;
+}
+
+static void _xgene_pmu_isr(int irq, struct xgene_pmu_dev *pmu_dev)
+{
+       struct xgene_pmu *xgene_pmu = pmu_dev->parent;
+       u32 pmovsr;
+       int idx;
+
+       pmovsr = readl(pmu_dev->inf->csr + PMU_PMOVSR) & PMU_OVERFLOW_MASK;
+       if (!pmovsr)
+               return;
+
+       /* Clear interrupt flag */
+       if (xgene_pmu->version == PCP_PMU_V1)
+               writel(0x0, pmu_dev->inf->csr + PMU_PMOVSR);
+       else
+               writel(pmovsr, pmu_dev->inf->csr + PMU_PMOVSR);
+
+       for (idx = 0; idx < PMU_MAX_COUNTERS; idx++) {
+               struct perf_event *event = pmu_dev->pmu_counter_event[idx];
+               int overflowed = pmovsr & BIT(idx);
+
+               /* Ignore if we don't have an event. */
+               if (!event || !overflowed)
+                       continue;
+               xgene_perf_event_update(event);
+               xgene_perf_event_set_period(event);
+       }
+}
+
+static irqreturn_t xgene_pmu_isr(int irq, void *dev_id)
+{
+       struct xgene_pmu_dev_ctx *ctx;
+       struct xgene_pmu *xgene_pmu = dev_id;
+       unsigned long flags;
+       u32 val;
+
+       raw_spin_lock_irqsave(&xgene_pmu->lock, flags);
+
+       /* Get Interrupt PMU source */
+       val = readl(xgene_pmu->pcppmu_csr + PCPPMU_INTSTATUS_REG);
+       if (val & PCPPMU_INT_MCU) {
+               list_for_each_entry(ctx, &xgene_pmu->mcpmus, next) {
+                       _xgene_pmu_isr(irq, ctx->pmu_dev);
+               }
+       }
+       if (val & PCPPMU_INT_MCB) {
+               list_for_each_entry(ctx, &xgene_pmu->mcbpmus, next) {
+                       _xgene_pmu_isr(irq, ctx->pmu_dev);
+               }
+       }
+       if (val & PCPPMU_INT_L3C) {
+               list_for_each_entry(ctx, &xgene_pmu->l3cpmus, next) {
+                       _xgene_pmu_isr(irq, ctx->pmu_dev);
+               }
+       }
+       if (val & PCPPMU_INT_IOB) {
+               list_for_each_entry(ctx, &xgene_pmu->iobpmus, next) {
+                       _xgene_pmu_isr(irq, ctx->pmu_dev);
+               }
+       }
+
+       raw_spin_unlock_irqrestore(&xgene_pmu->lock, flags);
+
+       return IRQ_HANDLED;
+}
+
+static int acpi_pmu_probe_active_mcb_mcu(struct xgene_pmu *xgene_pmu,
+                                        struct platform_device *pdev)
+{
+       void __iomem *csw_csr, *mcba_csr, *mcbb_csr;
+       struct resource *res;
+       unsigned int reg;
+
+       res = platform_get_resource(pdev, IORESOURCE_MEM, 1);
+       csw_csr = devm_ioremap_resource(&pdev->dev, res);
+       if (IS_ERR(csw_csr)) {
+               dev_err(&pdev->dev, "ioremap failed for CSW CSR resource\n");
+               return PTR_ERR(csw_csr);
+       }
+
+       res = platform_get_resource(pdev, IORESOURCE_MEM, 2);
+       mcba_csr = devm_ioremap_resource(&pdev->dev, res);
+       if (IS_ERR(mcba_csr)) {
+               dev_err(&pdev->dev, "ioremap failed for MCBA CSR resource\n");
+               return PTR_ERR(mcba_csr);
+       }
+
+       res = platform_get_resource(pdev, IORESOURCE_MEM, 3);
+       mcbb_csr = devm_ioremap_resource(&pdev->dev, res);
+       if (IS_ERR(mcbb_csr)) {
+               dev_err(&pdev->dev, "ioremap failed for MCBB CSR resource\n");
+               return PTR_ERR(mcbb_csr);
+       }
+
+       reg = readl(csw_csr + CSW_CSWCR);
+       if (reg & CSW_CSWCR_DUALMCB_MASK) {
+               /* Dual MCB active */
+               xgene_pmu->mcb_active_mask = 0x3;
+               /* Probe all active MC(s) */
+               reg = readl(mcbb_csr + CSW_CSWCR);
+               xgene_pmu->mc_active_mask =
+                       (reg & MCBADDRMR_DUALMCU_MODE_MASK) ? 0xF : 0x5;
+       } else {
+               /* Single MCB active */
+               xgene_pmu->mcb_active_mask = 0x1;
+               /* Probe all active MC(s) */
+               reg = readl(mcba_csr + CSW_CSWCR);
+               xgene_pmu->mc_active_mask =
+                       (reg & MCBADDRMR_DUALMCU_MODE_MASK) ? 0x3 : 0x1;
+       }
+
+       return 0;
+}
+
+static int fdt_pmu_probe_active_mcb_mcu(struct xgene_pmu *xgene_pmu,
+                                       struct platform_device *pdev)
+{
+       struct regmap *csw_map, *mcba_map, *mcbb_map;
+       struct device_node *np = pdev->dev.of_node;
+       unsigned int reg;
+
+       csw_map = syscon_regmap_lookup_by_phandle(np, "regmap-csw");
+       if (IS_ERR(csw_map)) {
+               dev_err(&pdev->dev, "unable to get syscon regmap csw\n");
+               return PTR_ERR(csw_map);
+       }
+
+       mcba_map = syscon_regmap_lookup_by_phandle(np, "regmap-mcba");
+       if (IS_ERR(mcba_map)) {
+               dev_err(&pdev->dev, "unable to get syscon regmap mcba\n");
+               return PTR_ERR(mcba_map);
+       }
+
+       mcbb_map = syscon_regmap_lookup_by_phandle(np, "regmap-mcbb");
+       if (IS_ERR(mcbb_map)) {
+               dev_err(&pdev->dev, "unable to get syscon regmap mcbb\n");
+               return PTR_ERR(mcbb_map);
+       }
+
+       if (regmap_read(csw_map, CSW_CSWCR, &reg))
+               return -EINVAL;
+
+       if (reg & CSW_CSWCR_DUALMCB_MASK) {
+               /* Dual MCB active */
+               xgene_pmu->mcb_active_mask = 0x3;
+               /* Probe all active MC(s) */
+               if (regmap_read(mcbb_map, MCBADDRMR, &reg))
+                       return 0;
+               xgene_pmu->mc_active_mask =
+                       (reg & MCBADDRMR_DUALMCU_MODE_MASK) ? 0xF : 0x5;
+       } else {
+               /* Single MCB active */
+               xgene_pmu->mcb_active_mask = 0x1;
+               /* Probe all active MC(s) */
+               if (regmap_read(mcba_map, MCBADDRMR, &reg))
+                       return 0;
+               xgene_pmu->mc_active_mask =
+                       (reg & MCBADDRMR_DUALMCU_MODE_MASK) ? 0x3 : 0x1;
+       }
+
+       return 0;
+}
+
+static int xgene_pmu_probe_active_mcb_mcu(struct xgene_pmu *xgene_pmu,
+                                         struct platform_device *pdev)
+{
+       if (has_acpi_companion(&pdev->dev))
+               return acpi_pmu_probe_active_mcb_mcu(xgene_pmu, pdev);
+       return fdt_pmu_probe_active_mcb_mcu(xgene_pmu, pdev);
+}
+
+static char *xgene_pmu_dev_name(struct device *dev, u32 type, int id)
+{
+       switch (type) {
+       case PMU_TYPE_L3C:
+               return devm_kasprintf(dev, GFP_KERNEL, "l3c%d", id);
+       case PMU_TYPE_IOB:
+               return devm_kasprintf(dev, GFP_KERNEL, "iob%d", id);
+       case PMU_TYPE_MCB:
+               return devm_kasprintf(dev, GFP_KERNEL, "mcb%d", id);
+       case PMU_TYPE_MC:
+               return devm_kasprintf(dev, GFP_KERNEL, "mc%d", id);
+       default:
+               return devm_kasprintf(dev, GFP_KERNEL, "unknown");
+       }
+}
+
+#if defined(CONFIG_ACPI)
+static int acpi_pmu_dev_add_resource(struct acpi_resource *ares, void *data)
+{
+       struct resource *res = data;
+
+       if (ares->type == ACPI_RESOURCE_TYPE_FIXED_MEMORY32)
+               acpi_dev_resource_memory(ares, res);
+
+       /* Always tell the ACPI core to skip this resource */
+       return 1;
+}
+
+static struct
+xgene_pmu_dev_ctx *acpi_get_pmu_hw_inf(struct xgene_pmu *xgene_pmu,
+                                      struct acpi_device *adev, u32 type)
+{
+       struct device *dev = xgene_pmu->dev;
+       struct list_head resource_list;
+       struct xgene_pmu_dev_ctx *ctx;
+       const union acpi_object *obj;
+       struct hw_pmu_info *inf;
+       void __iomem *dev_csr;
+       struct resource res;
+       int enable_bit;
+       int rc;
+
+       ctx = devm_kzalloc(dev, sizeof(*ctx), GFP_KERNEL);
+       if (!ctx)
+               return NULL;
+
+       INIT_LIST_HEAD(&resource_list);
+       rc = acpi_dev_get_resources(adev, &resource_list,
+                                   acpi_pmu_dev_add_resource, &res);
+       acpi_dev_free_resource_list(&resource_list);
+       if (rc < 0 || IS_ERR(&res)) {
+               dev_err(dev, "PMU type %d: No resource address found\n", type);
+               goto err;
+       }
+
+       dev_csr = devm_ioremap_resource(dev, &res);
+       if (IS_ERR(dev_csr)) {
+               dev_err(dev, "PMU type %d: Fail to map resource\n", type);
+               goto err;
+       }
+
+       /* A PMU device node without enable-bit-index is always enabled */
+       rc = acpi_dev_get_property(adev, "enable-bit-index",
+                                  ACPI_TYPE_INTEGER, &obj);
+       if (rc < 0)
+               enable_bit = 0;
+       else
+               enable_bit = (int) obj->integer.value;
+
+       ctx->name = xgene_pmu_dev_name(dev, type, enable_bit);
+       if (!ctx->name) {
+               dev_err(dev, "PMU type %d: Fail to get device name\n", type);
+               goto err;
+       }
+       inf = &ctx->inf;
+       inf->type = type;
+       inf->csr = dev_csr;
+       inf->enable_mask = 1 << enable_bit;
+
+       return ctx;
+err:
+       devm_kfree(dev, ctx);
+       return NULL;
+}
+
+static acpi_status acpi_pmu_dev_add(acpi_handle handle, u32 level,
+                                   void *data, void **return_value)
+{
+       struct xgene_pmu *xgene_pmu = data;
+       struct xgene_pmu_dev_ctx *ctx;
+       struct acpi_device *adev;
+
+       if (acpi_bus_get_device(handle, &adev))
+               return AE_OK;
+       if (acpi_bus_get_status(adev) || !adev->status.present)
+               return AE_OK;
+
+       if (!strcmp(acpi_device_hid(adev), "APMC0D5D"))
+               ctx = acpi_get_pmu_hw_inf(xgene_pmu, adev, PMU_TYPE_L3C);
+       else if (!strcmp(acpi_device_hid(adev), "APMC0D5E"))
+               ctx = acpi_get_pmu_hw_inf(xgene_pmu, adev, PMU_TYPE_IOB);
+       else if (!strcmp(acpi_device_hid(adev), "APMC0D5F"))
+               ctx = acpi_get_pmu_hw_inf(xgene_pmu, adev, PMU_TYPE_MCB);
+       else if (!strcmp(acpi_device_hid(adev), "APMC0D60"))
+               ctx = acpi_get_pmu_hw_inf(xgene_pmu, adev, PMU_TYPE_MC);
+       else
+               ctx = NULL;
+
+       if (!ctx)
+               return AE_OK;
+
+       if (xgene_pmu_dev_add(xgene_pmu, ctx)) {
+               /* Can't add the PMU device, skip it */
+               devm_kfree(xgene_pmu->dev, ctx);
+               return AE_OK;
+       }
+
+       switch (ctx->inf.type) {
+       case PMU_TYPE_L3C:
+               list_add(&ctx->next, &xgene_pmu->l3cpmus);
+               break;
+       case PMU_TYPE_IOB:
+               list_add(&ctx->next, &xgene_pmu->iobpmus);
+               break;
+       case PMU_TYPE_MCB:
+               list_add(&ctx->next, &xgene_pmu->mcbpmus);
+               break;
+       case PMU_TYPE_MC:
+               list_add(&ctx->next, &xgene_pmu->mcpmus);
+               break;
+       }
+       return AE_OK;
+}
+
+static int acpi_pmu_probe_pmu_dev(struct xgene_pmu *xgene_pmu,
+                                 struct platform_device *pdev)
+{
+       struct device *dev = xgene_pmu->dev;
+       acpi_handle handle;
+       acpi_status status;
+
+       handle = ACPI_HANDLE(dev);
+       if (!handle)
+               return -EINVAL;
+
+       status = acpi_walk_namespace(ACPI_TYPE_DEVICE, handle, 1,
+                                    acpi_pmu_dev_add, NULL, xgene_pmu, NULL);
+       if (ACPI_FAILURE(status)) {
+               dev_err(dev, "failed to probe PMU devices\n");
+               return -ENODEV;
+       }
+
+       return 0;
+}
+#else
+static int acpi_pmu_probe_pmu_dev(struct xgene_pmu *xgene_pmu,
+                                 struct platform_device *pdev)
+{
+       return 0;
+}
+#endif
+
+static struct
+xgene_pmu_dev_ctx *fdt_get_pmu_hw_inf(struct xgene_pmu *xgene_pmu,
+                                     struct device_node *np, u32 type)
+{
+       struct device *dev = xgene_pmu->dev;
+       struct xgene_pmu_dev_ctx *ctx;
+       struct hw_pmu_info *inf;
+       void __iomem *dev_csr;
+       struct resource res;
+       int enable_bit;
+       int rc;
+
+       ctx = devm_kzalloc(dev, sizeof(*ctx), GFP_KERNEL);
+       if (!ctx)
+               return NULL;
+       rc = of_address_to_resource(np, 0, &res);
+       if (rc < 0) {
+               dev_err(dev, "PMU type %d: No resource address found\n", type);
+               goto err;
+       }
+       dev_csr = devm_ioremap_resource(dev, &res);
+       if (IS_ERR(dev_csr)) {
+               dev_err(dev, "PMU type %d: Fail to map resource\n", type);
+               goto err;
+       }
+
+       /* A PMU device node without enable-bit-index is always enabled */
+       if (of_property_read_u32(np, "enable-bit-index", &enable_bit))
+               enable_bit = 0;
+
+       ctx->name = xgene_pmu_dev_name(dev, type, enable_bit);
+       if (!ctx->name) {
+               dev_err(dev, "PMU type %d: Fail to get device name\n", type);
+               goto err;
+       }
+       inf = &ctx->inf;
+       inf->type = type;
+       inf->csr = dev_csr;
+       inf->enable_mask = 1 << enable_bit;
+
+       return ctx;
+err:
+       devm_kfree(dev, ctx);
+       return NULL;
+}
+
+static int fdt_pmu_probe_pmu_dev(struct xgene_pmu *xgene_pmu,
+                                struct platform_device *pdev)
+{
+       struct xgene_pmu_dev_ctx *ctx;
+       struct device_node *np;
+
+       for_each_child_of_node(pdev->dev.of_node, np) {
+               if (!of_device_is_available(np))
+                       continue;
+
+               if (of_device_is_compatible(np, "apm,xgene-pmu-l3c"))
+                       ctx = fdt_get_pmu_hw_inf(xgene_pmu, np, PMU_TYPE_L3C);
+               else if (of_device_is_compatible(np, "apm,xgene-pmu-iob"))
+                       ctx = fdt_get_pmu_hw_inf(xgene_pmu, np, PMU_TYPE_IOB);
+               else if (of_device_is_compatible(np, "apm,xgene-pmu-mcb"))
+                       ctx = fdt_get_pmu_hw_inf(xgene_pmu, np, PMU_TYPE_MCB);
+               else if (of_device_is_compatible(np, "apm,xgene-pmu-mc"))
+                       ctx = fdt_get_pmu_hw_inf(xgene_pmu, np, PMU_TYPE_MC);
+               else
+                       ctx = NULL;
+
+               if (!ctx)
+                       continue;
+
+               if (xgene_pmu_dev_add(xgene_pmu, ctx)) {
+                       /* Can't add the PMU device, skip it */
+                       devm_kfree(xgene_pmu->dev, ctx);
+                       continue;
+               }
+
+               switch (ctx->inf.type) {
+               case PMU_TYPE_L3C:
+                       list_add(&ctx->next, &xgene_pmu->l3cpmus);
+                       break;
+               case PMU_TYPE_IOB:
+                       list_add(&ctx->next, &xgene_pmu->iobpmus);
+                       break;
+               case PMU_TYPE_MCB:
+                       list_add(&ctx->next, &xgene_pmu->mcbpmus);
+                       break;
+               case PMU_TYPE_MC:
+                       list_add(&ctx->next, &xgene_pmu->mcpmus);
+                       break;
+               }
+       }
+
+       return 0;
+}
+
+static int xgene_pmu_probe_pmu_dev(struct xgene_pmu *xgene_pmu,
+                                  struct platform_device *pdev)
+{
+       if (has_acpi_companion(&pdev->dev))
+               return acpi_pmu_probe_pmu_dev(xgene_pmu, pdev);
+       return fdt_pmu_probe_pmu_dev(xgene_pmu, pdev);
+}
+
+static const struct xgene_pmu_data xgene_pmu_data = {
+       .id   = PCP_PMU_V1,
+};
+
+static const struct xgene_pmu_data xgene_pmu_v2_data = {
+       .id   = PCP_PMU_V2,
+};
+
+static const struct of_device_id xgene_pmu_of_match[] = {
+       { .compatible   = "apm,xgene-pmu",      .data = &xgene_pmu_data },
+       { .compatible   = "apm,xgene-pmu-v2",   .data = &xgene_pmu_v2_data },
+       {},
+};
+MODULE_DEVICE_TABLE(of, xgene_pmu_of_match);
+#ifdef CONFIG_ACPI
+static const struct acpi_device_id xgene_pmu_acpi_match[] = {
+       {"APMC0D5B", PCP_PMU_V1},
+       {"APMC0D5C", PCP_PMU_V2},
+       {},
+};
+MODULE_DEVICE_TABLE(acpi, xgene_pmu_acpi_match);
+#endif
+
+static int xgene_pmu_probe(struct platform_device *pdev)
+{
+       const struct xgene_pmu_data *dev_data;
+       const struct of_device_id *of_id;
+       struct xgene_pmu *xgene_pmu;
+       struct resource *res;
+       int irq, rc;
+       int version;
+
+       xgene_pmu = devm_kzalloc(&pdev->dev, sizeof(*xgene_pmu), GFP_KERNEL);
+       if (!xgene_pmu)
+               return -ENOMEM;
+       xgene_pmu->dev = &pdev->dev;
+       platform_set_drvdata(pdev, xgene_pmu);
+
+       version = -EINVAL;
+       of_id = of_match_device(xgene_pmu_of_match, &pdev->dev);
+       if (of_id) {
+               dev_data = (const struct xgene_pmu_data *) of_id->data;
+               version = dev_data->id;
+       }
+
+#ifdef CONFIG_ACPI
+       if (ACPI_COMPANION(&pdev->dev)) {
+               const struct acpi_device_id *acpi_id;
+
+               acpi_id = acpi_match_device(xgene_pmu_acpi_match, &pdev->dev);
+               if (acpi_id)
+                       version = (int) acpi_id->driver_data;
+       }
+#endif
+       if (version < 0)
+               return -ENODEV;
+
+       INIT_LIST_HEAD(&xgene_pmu->l3cpmus);
+       INIT_LIST_HEAD(&xgene_pmu->iobpmus);
+       INIT_LIST_HEAD(&xgene_pmu->mcbpmus);
+       INIT_LIST_HEAD(&xgene_pmu->mcpmus);
+
+       xgene_pmu->version = version;
+       dev_info(&pdev->dev, "X-Gene PMU version %d\n", xgene_pmu->version);
+
+       res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
+       xgene_pmu->pcppmu_csr = devm_ioremap_resource(&pdev->dev, res);
+       if (IS_ERR(xgene_pmu->pcppmu_csr)) {
+               dev_err(&pdev->dev, "ioremap failed for PCP PMU resource\n");
+               rc = PTR_ERR(xgene_pmu->pcppmu_csr);
+               goto err;
+       }
+
+       irq = platform_get_irq(pdev, 0);
+       if (irq < 0) {
+               dev_err(&pdev->dev, "No IRQ resource\n");
+               rc = -EINVAL;
+               goto err;
+       }
+       rc = devm_request_irq(&pdev->dev, irq, xgene_pmu_isr,
+                               IRQF_NOBALANCING | IRQF_NO_THREAD,
+                               dev_name(&pdev->dev), xgene_pmu);
+       if (rc) {
+               dev_err(&pdev->dev, "Could not request IRQ %d\n", irq);
+               goto err;
+       }
+
+       raw_spin_lock_init(&xgene_pmu->lock);
+
+       /* Check for active MCBs and MCUs */
+       rc = xgene_pmu_probe_active_mcb_mcu(xgene_pmu, pdev);
+       if (rc) {
+               dev_warn(&pdev->dev, "Unknown MCB/MCU active status\n");
+               xgene_pmu->mcb_active_mask = 0x1;
+               xgene_pmu->mc_active_mask = 0x1;
+       }
+
+       /* Pick one core to use for cpumask attributes */
+       cpumask_set_cpu(smp_processor_id(), &xgene_pmu->cpu);
+
+       /* Make sure that the overflow interrupt is handled by this CPU */
+       rc = irq_set_affinity(irq, &xgene_pmu->cpu);
+       if (rc) {
+               dev_err(&pdev->dev, "Failed to set interrupt affinity!\n");
+               goto err;
+       }
+
+       /* Walk through the tree for all PMU perf devices */
+       rc = xgene_pmu_probe_pmu_dev(xgene_pmu, pdev);
+       if (rc) {
+               dev_err(&pdev->dev, "No PMU perf devices found!\n");
+               goto err;
+       }
+
+       /* Enable interrupt */
+       xgene_pmu_unmask_int(xgene_pmu);
+
+       return 0;
+
+err:
+       if (xgene_pmu->pcppmu_csr)
+               devm_iounmap(&pdev->dev, xgene_pmu->pcppmu_csr);
+       devm_kfree(&pdev->dev, xgene_pmu);
+
+       return rc;
+}
+
+static void
+xgene_pmu_dev_cleanup(struct xgene_pmu *xgene_pmu, struct list_head *pmus)
+{
+       struct xgene_pmu_dev_ctx *ctx;
+       struct device *dev = xgene_pmu->dev;
+       struct xgene_pmu_dev *pmu_dev;
+
+       list_for_each_entry(ctx, pmus, next) {
+               pmu_dev = ctx->pmu_dev;
+               if (pmu_dev->inf->csr)
+                       devm_iounmap(dev, pmu_dev->inf->csr);
+               devm_kfree(dev, ctx);
+               devm_kfree(dev, pmu_dev);
+       }
+}
+
+static int xgene_pmu_remove(struct platform_device *pdev)
+{
+       struct xgene_pmu *xgene_pmu = dev_get_drvdata(&pdev->dev);
+
+       xgene_pmu_dev_cleanup(xgene_pmu, &xgene_pmu->l3cpmus);
+       xgene_pmu_dev_cleanup(xgene_pmu, &xgene_pmu->iobpmus);
+       xgene_pmu_dev_cleanup(xgene_pmu, &xgene_pmu->mcbpmus);
+       xgene_pmu_dev_cleanup(xgene_pmu, &xgene_pmu->mcpmus);
+
+       if (xgene_pmu->pcppmu_csr)
+               devm_iounmap(&pdev->dev, xgene_pmu->pcppmu_csr);
+       devm_kfree(&pdev->dev, xgene_pmu);
+
+       return 0;
+}
+
+static struct platform_driver xgene_pmu_driver = {
+       .probe = xgene_pmu_probe,
+       .remove = xgene_pmu_remove,
+       .driver = {
+               .name           = "xgene-pmu",
+               .of_match_table = xgene_pmu_of_match,
+               .acpi_match_table = ACPI_PTR(xgene_pmu_acpi_match),
+       },
+};
+
+builtin_platform_driver(xgene_pmu_driver);