bus: ARM CCN PMU driver
authorPawel Moll <pawel.moll@arm.com>
Tue, 22 Jul 2014 17:32:59 +0000 (18:32 +0100)
committerArnd Bergmann <arnd@arndb.de>
Wed, 23 Jul 2014 20:14:43 +0000 (22:14 +0200)
Driver providing perf backend for ARM Cache Coherent Network
interconnect. Supports counting all hardware events and crosspoint
watchpoints.

Currently works with CCN-504 only, although there should be
no changes required for CCN-508 (just impossible to test it now).

Signed-off-by: Pawel Moll <pawel.moll@arm.com>
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Documentation/arm/CCN.txt [new file with mode: 0644]
Documentation/devicetree/bindings/arm/ccn.txt [new file with mode: 0644]
drivers/bus/Kconfig
drivers/bus/Makefile
drivers/bus/arm-ccn.c [new file with mode: 0644]

diff --git a/Documentation/arm/CCN.txt b/Documentation/arm/CCN.txt
new file mode 100644 (file)
index 0000000..0632b3a
--- /dev/null
@@ -0,0 +1,52 @@
+ARM Cache Coherent Network
+==========================
+
+CCN-504 is a ring-bus interconnect consisting of 11 crosspoints
+(XPs), with each crosspoint supporting up to two device ports,
+so nodes (devices) 0 and 1 are connected to crosspoint 0,
+nodes 2 and 3 to crosspoint 1 etc.
+
+PMU (perf) driver
+-----------------
+
+The CCN driver registers a perf PMU driver, which provides
+description of available events and configuration options
+in sysfs, see /sys/bus/event_source/devices/ccn*.
+
+The "format" directory describes format of the config, config1
+and config2 fields of the perf_event_attr structure. The "events"
+directory provides configuration templates for all documented
+events, that can be used with perf tool. For example "xp_valid_flit"
+is an equivalent of "type=0x8,event=0x4". Other parameters must be
+explicitly specified. For events originating from device, "node"
+defines its index. All crosspoint events require "xp" (index),
+"port" (device port number) and "vc" (virtual channel ID) and
+"dir" (direction). Watchpoints (special "event" value 0xfe) also
+require comparator values ("cmp_l" and "cmp_h") and "mask", being
+index of the comparator mask.
+
+Masks are defined separately from the event description
+(due to limited number of the config values) in the "cmp_mask"
+directory, with first 8 configurable by user and additional
+4 hardcoded for the most frequent use cases.
+
+Cycle counter is described by a "type" value 0xff and does
+not require any other settings.
+
+Example of perf tool use:
+
+/ # perf list | grep ccn
+  ccn/cycles/                                        [Kernel PMU event]
+<...>
+  ccn/xp_valid_flit/                                 [Kernel PMU event]
+<...>
+
+/ # perf stat -C 0 -e ccn/cycles/,ccn/xp_valid_flit,xp=1,port=0,vc=1,dir=1/ \
+                                                                       sleep 1
+
+The driver does not support sampling, therefore "perf record" will
+not work. Also notice that only single cpu is being selected
+("-C 0") - this is because perf framework does not support
+"non-CPU related" counters (yet?) so system-wide session ("-a")
+would try (and in most cases fail) to set up the same event
+per each CPU.
diff --git a/Documentation/devicetree/bindings/arm/ccn.txt b/Documentation/devicetree/bindings/arm/ccn.txt
new file mode 100644 (file)
index 0000000..b100d38
--- /dev/null
@@ -0,0 +1,21 @@
+* ARM CCN (Cache Coherent Network)
+
+Required properties:
+
+- compatible: (standard compatible string) should be one of:
+       "arm,ccn-504"
+       "arm,ccn-508"
+
+- reg: (standard registers property) physical address and size
+       (16MB) of the configuration registers block
+
+- interrupts: (standard interrupt property) single interrupt
+       generated by the control block
+
+Example:
+
+       ccn@0x2000000000 {
+               compatible = "arm,ccn-504";
+               reg = <0x20 0x00000000 0 0x1000000>;
+               interrupts = <0 181 4>;
+       };
index 1f37d9870e7a034fce68cc80efd4ab5bb1aed1e8..5c0c2764839f703053be59e44628ced6e8071e26 100644 (file)
@@ -50,6 +50,13 @@ config ARM_CCI
          Driver supporting the CCI cache coherent interconnect for ARM
          platforms.
 
+config ARM_CCN
+       bool "ARM CCN driver support"
+       depends on ARM || ARM64
+       help
+         PMU (perf) driver supporting the ARM CCN (Cache Coherent Network)
+         interconnect.
+
 config VEXPRESS_CONFIG
        bool "Versatile Express configuration bus"
        default y if ARCH_VEXPRESS
index 6a4ea7e4af1a9d8f46bb6ebaf0521f2df252db29..2973c18cbcc2781626645ec02434393baf0b2e1b 100644 (file)
@@ -9,7 +9,9 @@ obj-$(CONFIG_OMAP_OCP2SCP)      += omap-ocp2scp.o
 
 # Interconnect bus driver for OMAP SoCs.
 obj-$(CONFIG_OMAP_INTERCONNECT)        += omap_l3_smx.o omap_l3_noc.o
-# CCI cache coherent interconnect for ARM platforms
+
+# Interconnect bus drivers for ARM platforms
 obj-$(CONFIG_ARM_CCI)          += arm-cci.o
+obj-$(CONFIG_ARM_CCN)          += arm-ccn.o
 
 obj-$(CONFIG_VEXPRESS_CONFIG)  += vexpress-config.o
diff --git a/drivers/bus/arm-ccn.c b/drivers/bus/arm-ccn.c
new file mode 100644 (file)
index 0000000..4f86bbb
--- /dev/null
@@ -0,0 +1,1390 @@
+/*
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License 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.
+ *
+ * Copyright (C) 2014 ARM Limited
+ */
+
+#include <linux/ctype.h>
+#include <linux/hrtimer.h>
+#include <linux/idr.h>
+#include <linux/interrupt.h>
+#include <linux/io.h>
+#include <linux/module.h>
+#include <linux/perf_event.h>
+#include <linux/platform_device.h>
+#include <linux/slab.h>
+
+#define CCN_NUM_XP_PORTS 2
+#define CCN_NUM_VCS 4
+#define CCN_NUM_REGIONS        256
+#define CCN_REGION_SIZE        0x10000
+
+#define CCN_ALL_OLY_ID                 0xff00
+#define CCN_ALL_OLY_ID__OLY_ID__SHIFT                  0
+#define CCN_ALL_OLY_ID__OLY_ID__MASK                   0x1f
+#define CCN_ALL_OLY_ID__NODE_ID__SHIFT                 8
+#define CCN_ALL_OLY_ID__NODE_ID__MASK                  0x3f
+
+#define CCN_MN_ERRINT_STATUS           0x0008
+#define CCN_MN_ERRINT_STATUS__INTREQ__DESSERT          0x11
+#define CCN_MN_ERRINT_STATUS__ALL_ERRORS__ENABLE       0x02
+#define CCN_MN_ERRINT_STATUS__ALL_ERRORS__DISABLED     0x20
+#define CCN_MN_ERRINT_STATUS__ALL_ERRORS__DISABLE      0x22
+#define CCN_MN_ERRINT_STATUS__CORRECTED_ERRORS_ENABLE  0x04
+#define CCN_MN_ERRINT_STATUS__CORRECTED_ERRORS_DISABLED        0x40
+#define CCN_MN_ERRINT_STATUS__CORRECTED_ERRORS_DISABLE 0x44
+#define CCN_MN_ERRINT_STATUS__PMU_EVENTS__ENABLE       0x08
+#define CCN_MN_ERRINT_STATUS__PMU_EVENTS__DISABLED     0x80
+#define CCN_MN_ERRINT_STATUS__PMU_EVENTS__DISABLE      0x88
+#define CCN_MN_OLY_COMP_LIST_63_0      0x01e0
+#define CCN_MN_ERR_SIG_VAL_63_0                0x0300
+#define CCN_MN_ERR_SIG_VAL_63_0__DT                    (1 << 1)
+
+#define CCN_DT_ACTIVE_DSM              0x0000
+#define CCN_DT_ACTIVE_DSM__DSM_ID__SHIFT(n)            ((n) * 8)
+#define CCN_DT_ACTIVE_DSM__DSM_ID__MASK                        0xff
+#define CCN_DT_CTL                     0x0028
+#define CCN_DT_CTL__DT_EN                              (1 << 0)
+#define CCN_DT_PMEVCNT(n)              (0x0100 + (n) * 0x8)
+#define CCN_DT_PMCCNTR                 0x0140
+#define CCN_DT_PMCCNTRSR               0x0190
+#define CCN_DT_PMOVSR                  0x0198
+#define CCN_DT_PMOVSR_CLR              0x01a0
+#define CCN_DT_PMCR                    0x01a8
+#define CCN_DT_PMCR__OVFL_INTR_EN                      (1 << 6)
+#define CCN_DT_PMCR__PMU_EN                            (1 << 0)
+#define CCN_DT_PMSR                    0x01b0
+#define CCN_DT_PMSR_REQ                        0x01b8
+#define CCN_DT_PMSR_CLR                        0x01c0
+
+#define CCN_HNF_PMU_EVENT_SEL          0x0600
+#define CCN_HNF_PMU_EVENT_SEL__ID__SHIFT(n)            ((n) * 4)
+#define CCN_HNF_PMU_EVENT_SEL__ID__MASK                        0xf
+
+#define CCN_XP_DT_CONFIG               0x0300
+#define CCN_XP_DT_CONFIG__DT_CFG__SHIFT(n)             ((n) * 4)
+#define CCN_XP_DT_CONFIG__DT_CFG__MASK                 0xf
+#define CCN_XP_DT_CONFIG__DT_CFG__PASS_THROUGH         0x0
+#define CCN_XP_DT_CONFIG__DT_CFG__WATCHPOINT_0_OR_1    0x1
+#define CCN_XP_DT_CONFIG__DT_CFG__WATCHPOINT(n)                (0x2 + (n))
+#define CCN_XP_DT_CONFIG__DT_CFG__XP_PMU_EVENT(n)      (0x4 + (n))
+#define CCN_XP_DT_CONFIG__DT_CFG__DEVICE_PMU_EVENT(d, n) (0x8 + (d) * 4 + (n))
+#define CCN_XP_DT_INTERFACE_SEL                0x0308
+#define CCN_XP_DT_INTERFACE_SEL__DT_IO_SEL__SHIFT(n)   (0 + (n) * 8)
+#define CCN_XP_DT_INTERFACE_SEL__DT_IO_SEL__MASK       0x1
+#define CCN_XP_DT_INTERFACE_SEL__DT_DEV_SEL__SHIFT(n)  (1 + (n) * 8)
+#define CCN_XP_DT_INTERFACE_SEL__DT_DEV_SEL__MASK      0x1
+#define CCN_XP_DT_INTERFACE_SEL__DT_VC_SEL__SHIFT(n)   (2 + (n) * 8)
+#define CCN_XP_DT_INTERFACE_SEL__DT_VC_SEL__MASK       0x3
+#define CCN_XP_DT_CMP_VAL_L(n)         (0x0310 + (n) * 0x40)
+#define CCN_XP_DT_CMP_VAL_H(n)         (0x0318 + (n) * 0x40)
+#define CCN_XP_DT_CMP_MASK_L(n)                (0x0320 + (n) * 0x40)
+#define CCN_XP_DT_CMP_MASK_H(n)                (0x0328 + (n) * 0x40)
+#define CCN_XP_DT_CONTROL              0x0370
+#define CCN_XP_DT_CONTROL__DT_ENABLE                   (1 << 0)
+#define CCN_XP_DT_CONTROL__WP_ARM_SEL__SHIFT(n)                (12 + (n) * 4)
+#define CCN_XP_DT_CONTROL__WP_ARM_SEL__MASK            0xf
+#define CCN_XP_DT_CONTROL__WP_ARM_SEL__ALWAYS          0xf
+#define CCN_XP_PMU_EVENT_SEL           0x0600
+#define CCN_XP_PMU_EVENT_SEL__ID__SHIFT(n)             ((n) * 7)
+#define CCN_XP_PMU_EVENT_SEL__ID__MASK                 0x3f
+
+#define CCN_SBAS_PMU_EVENT_SEL         0x0600
+#define CCN_SBAS_PMU_EVENT_SEL__ID__SHIFT(n)           ((n) * 4)
+#define CCN_SBAS_PMU_EVENT_SEL__ID__MASK               0xf
+
+#define CCN_RNI_PMU_EVENT_SEL          0x0600
+#define CCN_RNI_PMU_EVENT_SEL__ID__SHIFT(n)            ((n) * 4)
+#define CCN_RNI_PMU_EVENT_SEL__ID__MASK                        0xf
+
+#define CCN_TYPE_MN    0x01
+#define CCN_TYPE_DT    0x02
+#define CCN_TYPE_HNF   0x04
+#define CCN_TYPE_HNI   0x05
+#define CCN_TYPE_XP    0x08
+#define CCN_TYPE_SBSX  0x0c
+#define CCN_TYPE_SBAS  0x10
+#define CCN_TYPE_RNI_1P        0x14
+#define CCN_TYPE_RNI_2P        0x15
+#define CCN_TYPE_RNI_3P        0x16
+#define CCN_TYPE_RND_1P        0x18 /* RN-D = RN-I + DVM */
+#define CCN_TYPE_RND_2P        0x19
+#define CCN_TYPE_RND_3P        0x1a
+#define CCN_TYPE_CYCLES        0xff /* Pseudotype */
+
+#define CCN_EVENT_WATCHPOINT 0xfe /* Pseudoevent */
+
+#define CCN_NUM_PMU_EVENTS             4
+#define CCN_NUM_XP_WATCHPOINTS         2 /* See DT.dbg_id.num_watchpoints */
+#define CCN_NUM_PMU_EVENT_COUNTERS     8 /* See DT.dbg_id.num_pmucntr */
+#define CCN_IDX_PMU_CYCLE_COUNTER      CCN_NUM_PMU_EVENT_COUNTERS
+
+#define CCN_NUM_PREDEFINED_MASKS       4
+#define CCN_IDX_MASK_ANY               (CCN_NUM_PMU_EVENT_COUNTERS + 0)
+#define CCN_IDX_MASK_EXACT             (CCN_NUM_PMU_EVENT_COUNTERS + 1)
+#define CCN_IDX_MASK_ORDER             (CCN_NUM_PMU_EVENT_COUNTERS + 2)
+#define CCN_IDX_MASK_OPCODE            (CCN_NUM_PMU_EVENT_COUNTERS + 3)
+
+struct arm_ccn_component {
+       void __iomem *base;
+       u32 type;
+
+       DECLARE_BITMAP(pmu_events_mask, CCN_NUM_PMU_EVENTS);
+       union {
+               struct {
+                       DECLARE_BITMAP(dt_cmp_mask, CCN_NUM_XP_WATCHPOINTS);
+               } xp;
+       };
+};
+
+#define pmu_to_arm_ccn(_pmu) container_of(container_of(_pmu, \
+       struct arm_ccn_dt, pmu), struct arm_ccn, dt)
+
+struct arm_ccn_dt {
+       int id;
+       void __iomem *base;
+
+       spinlock_t config_lock;
+
+       DECLARE_BITMAP(pmu_counters_mask, CCN_NUM_PMU_EVENT_COUNTERS + 1);
+       struct {
+               struct arm_ccn_component *source;
+               struct perf_event *event;
+       } pmu_counters[CCN_NUM_PMU_EVENT_COUNTERS + 1];
+
+       struct {
+              u64 l, h;
+       } cmp_mask[CCN_NUM_PMU_EVENT_COUNTERS + CCN_NUM_PREDEFINED_MASKS];
+
+       struct hrtimer hrtimer;
+
+       struct pmu pmu;
+};
+
+struct arm_ccn {
+       struct device *dev;
+       void __iomem *base;
+       unsigned irq_used:1;
+       unsigned sbas_present:1;
+       unsigned sbsx_present:1;
+
+       int num_nodes;
+       struct arm_ccn_component *node;
+
+       int num_xps;
+       struct arm_ccn_component *xp;
+
+       struct arm_ccn_dt dt;
+};
+
+
+static int arm_ccn_node_to_xp(int node)
+{
+       return node / CCN_NUM_XP_PORTS;
+}
+
+static int arm_ccn_node_to_xp_port(int node)
+{
+       return node % CCN_NUM_XP_PORTS;
+}
+
+
+/*
+ * Bit shifts and masks in these defines must be kept in sync with
+ * arm_ccn_pmu_config_set() and CCN_FORMAT_ATTRs below!
+ */
+#define CCN_CONFIG_NODE(_config)       (((_config) >> 0) & 0xff)
+#define CCN_CONFIG_XP(_config)         (((_config) >> 0) & 0xff)
+#define CCN_CONFIG_TYPE(_config)       (((_config) >> 8) & 0xff)
+#define CCN_CONFIG_EVENT(_config)      (((_config) >> 16) & 0xff)
+#define CCN_CONFIG_PORT(_config)       (((_config) >> 24) & 0x3)
+#define CCN_CONFIG_VC(_config)         (((_config) >> 26) & 0x7)
+#define CCN_CONFIG_DIR(_config)                (((_config) >> 29) & 0x1)
+#define CCN_CONFIG_MASK(_config)       (((_config) >> 30) & 0xf)
+
+static void arm_ccn_pmu_config_set(u64 *config, u32 node_xp, u32 type, u32 port)
+{
+       *config &= ~((0xff << 0) | (0xff << 8) | (0xff << 24));
+       *config |= (node_xp << 0) | (type << 8) | (port << 24);
+}
+
+static ssize_t arm_ccn_pmu_format_show(struct device *dev,
+               struct device_attribute *attr, char *buf)
+{
+       struct dev_ext_attribute *ea = container_of(attr,
+                       struct dev_ext_attribute, attr);
+
+       return snprintf(buf, PAGE_SIZE, "%s\n", (char *)ea->var);
+}
+
+#define CCN_FORMAT_ATTR(_name, _config) \
+       struct dev_ext_attribute arm_ccn_pmu_format_attr_##_name = \
+                       { __ATTR(_name, S_IRUGO, arm_ccn_pmu_format_show, \
+                       NULL), _config }
+
+static CCN_FORMAT_ATTR(node, "config:0-7");
+static CCN_FORMAT_ATTR(xp, "config:0-7");
+static CCN_FORMAT_ATTR(type, "config:8-15");
+static CCN_FORMAT_ATTR(event, "config:16-23");
+static CCN_FORMAT_ATTR(port, "config:24-25");
+static CCN_FORMAT_ATTR(vc, "config:26-28");
+static CCN_FORMAT_ATTR(dir, "config:29-29");
+static CCN_FORMAT_ATTR(mask, "config:30-33");
+static CCN_FORMAT_ATTR(cmp_l, "config1:0-62");
+static CCN_FORMAT_ATTR(cmp_h, "config2:0-59");
+
+static struct attribute *arm_ccn_pmu_format_attrs[] = {
+       &arm_ccn_pmu_format_attr_node.attr.attr,
+       &arm_ccn_pmu_format_attr_xp.attr.attr,
+       &arm_ccn_pmu_format_attr_type.attr.attr,
+       &arm_ccn_pmu_format_attr_event.attr.attr,
+       &arm_ccn_pmu_format_attr_port.attr.attr,
+       &arm_ccn_pmu_format_attr_vc.attr.attr,
+       &arm_ccn_pmu_format_attr_dir.attr.attr,
+       &arm_ccn_pmu_format_attr_mask.attr.attr,
+       &arm_ccn_pmu_format_attr_cmp_l.attr.attr,
+       &arm_ccn_pmu_format_attr_cmp_h.attr.attr,
+       NULL
+};
+
+static struct attribute_group arm_ccn_pmu_format_attr_group = {
+       .name = "format",
+       .attrs = arm_ccn_pmu_format_attrs,
+};
+
+
+struct arm_ccn_pmu_event {
+       struct device_attribute attr;
+       u32 type;
+       u32 event;
+       int num_ports;
+       int num_vcs;
+       const char *def;
+       int mask;
+};
+
+#define CCN_EVENT_ATTR(_name) \
+       __ATTR(_name, S_IRUGO, arm_ccn_pmu_event_show, NULL)
+
+/*
+ * Events defined in TRM for MN, HN-I and SBSX are actually watchpoints set on
+ * their ports in XP they are connected to. For the sake of usability they are
+ * explicitly defined here (and translated into a relevant watchpoint in
+ * arm_ccn_pmu_event_init()) so the user can easily request them without deep
+ * knowledge of the flit format.
+ */
+
+#define CCN_EVENT_MN(_name, _def, _mask) { .attr = CCN_EVENT_ATTR(mn_##_name), \
+               .type = CCN_TYPE_MN, .event = CCN_EVENT_WATCHPOINT, \
+               .num_ports = CCN_NUM_XP_PORTS, .num_vcs = CCN_NUM_VCS, \
+               .def = _def, .mask = _mask, }
+
+#define CCN_EVENT_HNI(_name, _def, _mask) { \
+               .attr = CCN_EVENT_ATTR(hni_##_name), .type = CCN_TYPE_HNI, \
+               .event = CCN_EVENT_WATCHPOINT, .num_ports = CCN_NUM_XP_PORTS, \
+               .num_vcs = CCN_NUM_VCS, .def = _def, .mask = _mask, }
+
+#define CCN_EVENT_SBSX(_name, _def, _mask) { \
+               .attr = CCN_EVENT_ATTR(sbsx_##_name), .type = CCN_TYPE_SBSX, \
+               .event = CCN_EVENT_WATCHPOINT, .num_ports = CCN_NUM_XP_PORTS, \
+               .num_vcs = CCN_NUM_VCS, .def = _def, .mask = _mask, }
+
+#define CCN_EVENT_HNF(_name, _event) { .attr = CCN_EVENT_ATTR(hnf_##_name), \
+               .type = CCN_TYPE_HNF, .event = _event, }
+
+#define CCN_EVENT_XP(_name, _event) { .attr = CCN_EVENT_ATTR(xp_##_name), \
+               .type = CCN_TYPE_XP, .event = _event, \
+               .num_ports = CCN_NUM_XP_PORTS, .num_vcs = CCN_NUM_VCS, }
+
+/*
+ * RN-I & RN-D (RN-D = RN-I + DVM) nodes have different type ID depending
+ * on configuration. One of them is picked to represent the whole group,
+ * as they all share the same event types.
+ */
+#define CCN_EVENT_RNI(_name, _event) { .attr = CCN_EVENT_ATTR(rni_##_name), \
+               .type = CCN_TYPE_RNI_3P, .event = _event, }
+
+#define CCN_EVENT_SBAS(_name, _event) { .attr = CCN_EVENT_ATTR(sbas_##_name), \
+               .type = CCN_TYPE_SBAS, .event = _event, }
+
+#define CCN_EVENT_CYCLES(_name) { .attr = CCN_EVENT_ATTR(_name), \
+               .type = CCN_TYPE_CYCLES }
+
+
+static ssize_t arm_ccn_pmu_event_show(struct device *dev,
+               struct device_attribute *attr, char *buf)
+{
+       struct arm_ccn_pmu_event *event = container_of(attr,
+                       struct arm_ccn_pmu_event, attr);
+       ssize_t res;
+
+       res = snprintf(buf, PAGE_SIZE, "type=0x%x", event->type);
+       if (event->event)
+               res += snprintf(buf + res, PAGE_SIZE - res, ",event=0x%x",
+                               event->event);
+       if (event->def)
+               res += snprintf(buf + res, PAGE_SIZE - res, ",%s",
+                               event->def);
+       if (event->mask)
+               res += snprintf(buf + res, PAGE_SIZE - res, ",mask=0x%x",
+                               event->mask);
+       res += snprintf(buf + res, PAGE_SIZE - res, "\n");
+
+       return res;
+}
+
+static umode_t arm_ccn_pmu_events_is_visible(struct kobject *kobj,
+                                    struct attribute *attr, int index)
+{
+       struct device *dev = kobj_to_dev(kobj);
+       struct arm_ccn *ccn = pmu_to_arm_ccn(dev_get_drvdata(dev));
+       struct device_attribute *dev_attr = container_of(attr,
+                       struct device_attribute, attr);
+       struct arm_ccn_pmu_event *event = container_of(dev_attr,
+                       struct arm_ccn_pmu_event, attr);
+
+       if (event->type == CCN_TYPE_SBAS && !ccn->sbas_present)
+               return 0;
+       if (event->type == CCN_TYPE_SBSX && !ccn->sbsx_present)
+               return 0;
+
+       return attr->mode;
+}
+
+static struct arm_ccn_pmu_event arm_ccn_pmu_events[] = {
+       CCN_EVENT_MN(eobarrier, "dir=0,vc=0,cmp_h=0x1c00", CCN_IDX_MASK_OPCODE),
+       CCN_EVENT_MN(ecbarrier, "dir=0,vc=0,cmp_h=0x1e00", CCN_IDX_MASK_OPCODE),
+       CCN_EVENT_MN(dvmop, "dir=0,vc=0,cmp_h=0x2800", CCN_IDX_MASK_OPCODE),
+       CCN_EVENT_HNI(txdatflits, "dir=1,vc=3", CCN_IDX_MASK_ANY),
+       CCN_EVENT_HNI(rxdatflits, "dir=0,vc=3", CCN_IDX_MASK_ANY),
+       CCN_EVENT_HNI(txreqflits, "dir=1,vc=0", CCN_IDX_MASK_ANY),
+       CCN_EVENT_HNI(rxreqflits, "dir=0,vc=0", CCN_IDX_MASK_ANY),
+       CCN_EVENT_HNI(rxreqflits_order, "dir=0,vc=0,cmp_h=0x8000",
+                       CCN_IDX_MASK_ORDER),
+       CCN_EVENT_SBSX(txdatflits, "dir=1,vc=3", CCN_IDX_MASK_ANY),
+       CCN_EVENT_SBSX(rxdatflits, "dir=0,vc=3", CCN_IDX_MASK_ANY),
+       CCN_EVENT_SBSX(txreqflits, "dir=1,vc=0", CCN_IDX_MASK_ANY),
+       CCN_EVENT_SBSX(rxreqflits, "dir=0,vc=0", CCN_IDX_MASK_ANY),
+       CCN_EVENT_SBSX(rxreqflits_order, "dir=0,vc=0,cmp_h=0x8000",
+                       CCN_IDX_MASK_ORDER),
+       CCN_EVENT_HNF(cache_miss, 0x1),
+       CCN_EVENT_HNF(l3_sf_cache_access, 0x02),
+       CCN_EVENT_HNF(cache_fill, 0x3),
+       CCN_EVENT_HNF(pocq_retry, 0x4),
+       CCN_EVENT_HNF(pocq_reqs_recvd, 0x5),
+       CCN_EVENT_HNF(sf_hit, 0x6),
+       CCN_EVENT_HNF(sf_evictions, 0x7),
+       CCN_EVENT_HNF(snoops_sent, 0x8),
+       CCN_EVENT_HNF(snoops_broadcast, 0x9),
+       CCN_EVENT_HNF(l3_eviction, 0xa),
+       CCN_EVENT_HNF(l3_fill_invalid_way, 0xb),
+       CCN_EVENT_HNF(mc_retries, 0xc),
+       CCN_EVENT_HNF(mc_reqs, 0xd),
+       CCN_EVENT_HNF(qos_hh_retry, 0xe),
+       CCN_EVENT_RNI(rdata_beats_p0, 0x1),
+       CCN_EVENT_RNI(rdata_beats_p1, 0x2),
+       CCN_EVENT_RNI(rdata_beats_p2, 0x3),
+       CCN_EVENT_RNI(rxdat_flits, 0x4),
+       CCN_EVENT_RNI(txdat_flits, 0x5),
+       CCN_EVENT_RNI(txreq_flits, 0x6),
+       CCN_EVENT_RNI(txreq_flits_retried, 0x7),
+       CCN_EVENT_RNI(rrt_full, 0x8),
+       CCN_EVENT_RNI(wrt_full, 0x9),
+       CCN_EVENT_RNI(txreq_flits_replayed, 0xa),
+       CCN_EVENT_XP(upload_starvation, 0x1),
+       CCN_EVENT_XP(download_starvation, 0x2),
+       CCN_EVENT_XP(respin, 0x3),
+       CCN_EVENT_XP(valid_flit, 0x4),
+       CCN_EVENT_XP(watchpoint, CCN_EVENT_WATCHPOINT),
+       CCN_EVENT_SBAS(rdata_beats_p0, 0x1),
+       CCN_EVENT_SBAS(rxdat_flits, 0x4),
+       CCN_EVENT_SBAS(txdat_flits, 0x5),
+       CCN_EVENT_SBAS(txreq_flits, 0x6),
+       CCN_EVENT_SBAS(txreq_flits_retried, 0x7),
+       CCN_EVENT_SBAS(rrt_full, 0x8),
+       CCN_EVENT_SBAS(wrt_full, 0x9),
+       CCN_EVENT_SBAS(txreq_flits_replayed, 0xa),
+       CCN_EVENT_CYCLES(cycles),
+};
+
+/* Populated in arm_ccn_init() */
+static struct attribute
+               *arm_ccn_pmu_events_attrs[ARRAY_SIZE(arm_ccn_pmu_events) + 1];
+
+static struct attribute_group arm_ccn_pmu_events_attr_group = {
+       .name = "events",
+       .is_visible = arm_ccn_pmu_events_is_visible,
+       .attrs = arm_ccn_pmu_events_attrs,
+};
+
+
+static u64 *arm_ccn_pmu_get_cmp_mask(struct arm_ccn *ccn, const char *name)
+{
+       unsigned long i;
+
+       if (WARN_ON(!name || !name[0] || !isxdigit(name[0]) || !name[1]))
+               return NULL;
+       i = isdigit(name[0]) ? name[0] - '0' : 0xa + tolower(name[0]) - 'a';
+
+       switch (name[1]) {
+       case 'l':
+               return &ccn->dt.cmp_mask[i].l;
+       case 'h':
+               return &ccn->dt.cmp_mask[i].h;
+       default:
+               return NULL;
+       }
+}
+
+static ssize_t arm_ccn_pmu_cmp_mask_show(struct device *dev,
+               struct device_attribute *attr, char *buf)
+{
+       struct arm_ccn *ccn = pmu_to_arm_ccn(dev_get_drvdata(dev));
+       u64 *mask = arm_ccn_pmu_get_cmp_mask(ccn, attr->attr.name);
+
+       return mask ? snprintf(buf, PAGE_SIZE, "0x%016llx\n", *mask) : -EINVAL;
+}
+
+static ssize_t arm_ccn_pmu_cmp_mask_store(struct device *dev,
+               struct device_attribute *attr, const char *buf, size_t count)
+{
+       struct arm_ccn *ccn = pmu_to_arm_ccn(dev_get_drvdata(dev));
+       u64 *mask = arm_ccn_pmu_get_cmp_mask(ccn, attr->attr.name);
+       int err = -EINVAL;
+
+       if (mask)
+               err = kstrtoull(buf, 0, mask);
+
+       return err ? err : count;
+}
+
+#define CCN_CMP_MASK_ATTR(_name) \
+       struct device_attribute arm_ccn_pmu_cmp_mask_attr_##_name = \
+                       __ATTR(_name, S_IRUGO | S_IWUSR, \
+                       arm_ccn_pmu_cmp_mask_show, arm_ccn_pmu_cmp_mask_store)
+
+#define CCN_CMP_MASK_ATTR_RO(_name) \
+       struct device_attribute arm_ccn_pmu_cmp_mask_attr_##_name = \
+                       __ATTR(_name, S_IRUGO, arm_ccn_pmu_cmp_mask_show, NULL)
+
+static CCN_CMP_MASK_ATTR(0l);
+static CCN_CMP_MASK_ATTR(0h);
+static CCN_CMP_MASK_ATTR(1l);
+static CCN_CMP_MASK_ATTR(1h);
+static CCN_CMP_MASK_ATTR(2l);
+static CCN_CMP_MASK_ATTR(2h);
+static CCN_CMP_MASK_ATTR(3l);
+static CCN_CMP_MASK_ATTR(3h);
+static CCN_CMP_MASK_ATTR(4l);
+static CCN_CMP_MASK_ATTR(4h);
+static CCN_CMP_MASK_ATTR(5l);
+static CCN_CMP_MASK_ATTR(5h);
+static CCN_CMP_MASK_ATTR(6l);
+static CCN_CMP_MASK_ATTR(6h);
+static CCN_CMP_MASK_ATTR(7l);
+static CCN_CMP_MASK_ATTR(7h);
+static CCN_CMP_MASK_ATTR_RO(8l);
+static CCN_CMP_MASK_ATTR_RO(8h);
+static CCN_CMP_MASK_ATTR_RO(9l);
+static CCN_CMP_MASK_ATTR_RO(9h);
+static CCN_CMP_MASK_ATTR_RO(al);
+static CCN_CMP_MASK_ATTR_RO(ah);
+static CCN_CMP_MASK_ATTR_RO(bl);
+static CCN_CMP_MASK_ATTR_RO(bh);
+
+static struct attribute *arm_ccn_pmu_cmp_mask_attrs[] = {
+       &arm_ccn_pmu_cmp_mask_attr_0l.attr, &arm_ccn_pmu_cmp_mask_attr_0h.attr,
+       &arm_ccn_pmu_cmp_mask_attr_1l.attr, &arm_ccn_pmu_cmp_mask_attr_1h.attr,
+       &arm_ccn_pmu_cmp_mask_attr_2l.attr, &arm_ccn_pmu_cmp_mask_attr_2h.attr,
+       &arm_ccn_pmu_cmp_mask_attr_3l.attr, &arm_ccn_pmu_cmp_mask_attr_3h.attr,
+       &arm_ccn_pmu_cmp_mask_attr_4l.attr, &arm_ccn_pmu_cmp_mask_attr_4h.attr,
+       &arm_ccn_pmu_cmp_mask_attr_5l.attr, &arm_ccn_pmu_cmp_mask_attr_5h.attr,
+       &arm_ccn_pmu_cmp_mask_attr_6l.attr, &arm_ccn_pmu_cmp_mask_attr_6h.attr,
+       &arm_ccn_pmu_cmp_mask_attr_7l.attr, &arm_ccn_pmu_cmp_mask_attr_7h.attr,
+       &arm_ccn_pmu_cmp_mask_attr_8l.attr, &arm_ccn_pmu_cmp_mask_attr_8h.attr,
+       &arm_ccn_pmu_cmp_mask_attr_9l.attr, &arm_ccn_pmu_cmp_mask_attr_9h.attr,
+       &arm_ccn_pmu_cmp_mask_attr_al.attr, &arm_ccn_pmu_cmp_mask_attr_ah.attr,
+       &arm_ccn_pmu_cmp_mask_attr_bl.attr, &arm_ccn_pmu_cmp_mask_attr_bh.attr,
+       NULL
+};
+
+static struct attribute_group arm_ccn_pmu_cmp_mask_attr_group = {
+       .name = "cmp_mask",
+       .attrs = arm_ccn_pmu_cmp_mask_attrs,
+};
+
+
+/*
+ * Default poll period is 10ms, which is way over the top anyway,
+ * as in the worst case scenario (an event every cycle), with 1GHz
+ * clocked bus, the smallest, 32 bit counter will overflow in
+ * more than 4s.
+ */
+static unsigned int arm_ccn_pmu_poll_period_us = 10000;
+module_param_named(pmu_poll_period_us, arm_ccn_pmu_poll_period_us, uint,
+               S_IRUGO | S_IWUSR);
+
+static ktime_t arm_ccn_pmu_timer_period(void)
+{
+       return ns_to_ktime((u64)arm_ccn_pmu_poll_period_us * 1000);
+}
+
+
+static const struct attribute_group *arm_ccn_pmu_attr_groups[] = {
+       &arm_ccn_pmu_events_attr_group,
+       &arm_ccn_pmu_format_attr_group,
+       &arm_ccn_pmu_cmp_mask_attr_group,
+       NULL
+};
+
+
+static int arm_ccn_pmu_alloc_bit(unsigned long *bitmap, unsigned long size)
+{
+       int bit;
+
+       do {
+               bit = find_first_zero_bit(bitmap, size);
+               if (bit >= size)
+                       return -EAGAIN;
+       } while (test_and_set_bit(bit, bitmap));
+
+       return bit;
+}
+
+/* All RN-I and RN-D nodes have identical PMUs */
+static int arm_ccn_pmu_type_eq(u32 a, u32 b)
+{
+       if (a == b)
+               return 1;
+
+       switch (a) {
+       case CCN_TYPE_RNI_1P:
+       case CCN_TYPE_RNI_2P:
+       case CCN_TYPE_RNI_3P:
+       case CCN_TYPE_RND_1P:
+       case CCN_TYPE_RND_2P:
+       case CCN_TYPE_RND_3P:
+               switch (b) {
+               case CCN_TYPE_RNI_1P:
+               case CCN_TYPE_RNI_2P:
+               case CCN_TYPE_RNI_3P:
+               case CCN_TYPE_RND_1P:
+               case CCN_TYPE_RND_2P:
+               case CCN_TYPE_RND_3P:
+                       return 1;
+               }
+               break;
+       }
+
+       return 0;
+}
+
+static int arm_ccn_pmu_event_init(struct perf_event *event)
+{
+       struct arm_ccn *ccn;
+       struct hw_perf_event *hw = &event->hw;
+       u32 node_xp, type, event_id;
+       int valid;
+       struct arm_ccn_component *source;
+       int i;
+
+       if (event->attr.type != event->pmu->type)
+               return -ENOENT;
+
+       ccn = pmu_to_arm_ccn(event->pmu);
+
+       if (hw->sample_period) {
+               dev_warn(ccn->dev, "Sampling not supported!\n");
+               return -EOPNOTSUPP;
+       }
+
+       if (has_branch_stack(event) || event->attr.exclude_user ||
+                       event->attr.exclude_kernel || event->attr.exclude_hv ||
+                       event->attr.exclude_idle) {
+               dev_warn(ccn->dev, "Can't exclude execution levels!\n");
+               return -EOPNOTSUPP;
+       }
+
+       if (event->cpu < 0) {
+               dev_warn(ccn->dev, "Can't provide per-task data!\n");
+               return -EOPNOTSUPP;
+       }
+
+       node_xp = CCN_CONFIG_NODE(event->attr.config);
+       type = CCN_CONFIG_TYPE(event->attr.config);
+       event_id = CCN_CONFIG_EVENT(event->attr.config);
+
+       /* Validate node/xp vs topology */
+       switch (type) {
+       case CCN_TYPE_XP:
+               if (node_xp >= ccn->num_xps) {
+                       dev_warn(ccn->dev, "Invalid XP ID %d!\n", node_xp);
+                       return -EINVAL;
+               }
+               break;
+       case CCN_TYPE_CYCLES:
+               break;
+       default:
+               if (node_xp >= ccn->num_nodes) {
+                       dev_warn(ccn->dev, "Invalid node ID %d!\n", node_xp);
+                       return -EINVAL;
+               }
+               if (!arm_ccn_pmu_type_eq(type, ccn->node[node_xp].type)) {
+                       dev_warn(ccn->dev, "Invalid type 0x%x for node %d!\n",
+                                       type, node_xp);
+                       return -EINVAL;
+               }
+               break;
+       }
+
+       /* Validate event ID vs available for the type */
+       for (i = 0, valid = 0; i < ARRAY_SIZE(arm_ccn_pmu_events) && !valid;
+                       i++) {
+               struct arm_ccn_pmu_event *e = &arm_ccn_pmu_events[i];
+               u32 port = CCN_CONFIG_PORT(event->attr.config);
+               u32 vc = CCN_CONFIG_VC(event->attr.config);
+
+               if (!arm_ccn_pmu_type_eq(type, e->type))
+                       continue;
+               if (event_id != e->event)
+                       continue;
+               if (e->num_ports && port >= e->num_ports) {
+                       dev_warn(ccn->dev, "Invalid port %d for node/XP %d!\n",
+                                       port, node_xp);
+                       return -EINVAL;
+               }
+               if (e->num_vcs && vc >= e->num_vcs) {
+                       dev_warn(ccn->dev, "Invalid vc %d for node/XP %d!\n",
+                                       port, node_xp);
+                       return -EINVAL;
+               }
+               valid = 1;
+       }
+       if (!valid) {
+               dev_warn(ccn->dev, "Invalid event 0x%x for node/XP %d!\n",
+                               event_id, node_xp);
+               return -EINVAL;
+       }
+
+       /* Watchpoint-based event for a node is actually set on XP */
+       if (event_id == CCN_EVENT_WATCHPOINT && type != CCN_TYPE_XP) {
+               u32 port;
+
+               type = CCN_TYPE_XP;
+               port = arm_ccn_node_to_xp_port(node_xp);
+               node_xp = arm_ccn_node_to_xp(node_xp);
+
+               arm_ccn_pmu_config_set(&event->attr.config,
+                               node_xp, type, port);
+       }
+
+       /* Allocate the cycle counter */
+       if (type == CCN_TYPE_CYCLES) {
+               if (test_and_set_bit(CCN_IDX_PMU_CYCLE_COUNTER,
+                               ccn->dt.pmu_counters_mask))
+                       return -EAGAIN;
+
+               hw->idx = CCN_IDX_PMU_CYCLE_COUNTER;
+               ccn->dt.pmu_counters[CCN_IDX_PMU_CYCLE_COUNTER].event = event;
+
+               return 0;
+       }
+
+       /* Allocate an event counter */
+       hw->idx = arm_ccn_pmu_alloc_bit(ccn->dt.pmu_counters_mask,
+                       CCN_NUM_PMU_EVENT_COUNTERS);
+       if (hw->idx < 0) {
+               dev_warn(ccn->dev, "No more counters available!\n");
+               return -EAGAIN;
+       }
+
+       if (type == CCN_TYPE_XP)
+               source = &ccn->xp[node_xp];
+       else
+               source = &ccn->node[node_xp];
+       ccn->dt.pmu_counters[hw->idx].source = source;
+
+       /* Allocate an event source or a watchpoint */
+       if (type == CCN_TYPE_XP && event_id == CCN_EVENT_WATCHPOINT)
+               hw->config_base = arm_ccn_pmu_alloc_bit(source->xp.dt_cmp_mask,
+                               CCN_NUM_XP_WATCHPOINTS);
+       else
+               hw->config_base = arm_ccn_pmu_alloc_bit(source->pmu_events_mask,
+                               CCN_NUM_PMU_EVENTS);
+       if (hw->config_base < 0) {
+               dev_warn(ccn->dev, "No more event sources/watchpoints on node/XP %d!\n",
+                               node_xp);
+               clear_bit(hw->idx, ccn->dt.pmu_counters_mask);
+               return -EAGAIN;
+       }
+
+       ccn->dt.pmu_counters[hw->idx].event = event;
+
+       return 0;
+}
+
+static void arm_ccn_pmu_event_free(struct perf_event *event)
+{
+       struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
+       struct hw_perf_event *hw = &event->hw;
+
+       if (hw->idx == CCN_IDX_PMU_CYCLE_COUNTER) {
+               clear_bit(CCN_IDX_PMU_CYCLE_COUNTER, ccn->dt.pmu_counters_mask);
+       } else {
+               struct arm_ccn_component *source =
+                               ccn->dt.pmu_counters[hw->idx].source;
+
+               if (CCN_CONFIG_TYPE(event->attr.config) == CCN_TYPE_XP &&
+                               CCN_CONFIG_EVENT(event->attr.config) ==
+                               CCN_EVENT_WATCHPOINT)
+                       clear_bit(hw->config_base, source->xp.dt_cmp_mask);
+               else
+                       clear_bit(hw->config_base, source->pmu_events_mask);
+               clear_bit(hw->idx, ccn->dt.pmu_counters_mask);
+       }
+
+       ccn->dt.pmu_counters[hw->idx].source = NULL;
+       ccn->dt.pmu_counters[hw->idx].event = NULL;
+}
+
+static u64 arm_ccn_pmu_read_counter(struct arm_ccn *ccn, int idx)
+{
+       u64 res;
+
+       if (idx == CCN_IDX_PMU_CYCLE_COUNTER) {
+#ifdef readq
+               res = readq(ccn->dt.base + CCN_DT_PMCCNTR);
+#else
+               /* 40 bit counter, can do snapshot and read in two parts */
+               writel(0x1, ccn->dt.base + CCN_DT_PMSR_REQ);
+               while (!(readl(ccn->dt.base + CCN_DT_PMSR) & 0x1))
+                       ;
+               writel(0x1, ccn->dt.base + CCN_DT_PMSR_CLR);
+               res = readl(ccn->dt.base + CCN_DT_PMCCNTRSR + 4) & 0xff;
+               res <<= 32;
+               res |= readl(ccn->dt.base + CCN_DT_PMCCNTRSR);
+#endif
+       } else {
+               res = readl(ccn->dt.base + CCN_DT_PMEVCNT(idx));
+       }
+
+       return res;
+}
+
+static void arm_ccn_pmu_event_update(struct perf_event *event)
+{
+       struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
+       struct hw_perf_event *hw = &event->hw;
+       u64 prev_count, new_count, mask;
+
+       do {
+               prev_count = local64_read(&hw->prev_count);
+               new_count = arm_ccn_pmu_read_counter(ccn, hw->idx);
+       } while (local64_xchg(&hw->prev_count, new_count) != prev_count);
+
+       mask = (1LLU << (hw->idx == CCN_IDX_PMU_CYCLE_COUNTER ? 40 : 32)) - 1;
+
+       local64_add((new_count - prev_count) & mask, &event->count);
+}
+
+static void arm_ccn_pmu_xp_dt_config(struct perf_event *event, int enable)
+{
+       struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
+       struct hw_perf_event *hw = &event->hw;
+       struct arm_ccn_component *xp;
+       u32 val, dt_cfg;
+
+       if (CCN_CONFIG_TYPE(event->attr.config) == CCN_TYPE_XP)
+               xp = &ccn->xp[CCN_CONFIG_XP(event->attr.config)];
+       else
+               xp = &ccn->xp[arm_ccn_node_to_xp(
+                               CCN_CONFIG_NODE(event->attr.config))];
+
+       if (enable)
+               dt_cfg = hw->event_base;
+       else
+               dt_cfg = CCN_XP_DT_CONFIG__DT_CFG__PASS_THROUGH;
+
+       spin_lock(&ccn->dt.config_lock);
+
+       val = readl(xp->base + CCN_XP_DT_CONFIG);
+       val &= ~(CCN_XP_DT_CONFIG__DT_CFG__MASK <<
+                       CCN_XP_DT_CONFIG__DT_CFG__SHIFT(hw->idx));
+       val |= dt_cfg << CCN_XP_DT_CONFIG__DT_CFG__SHIFT(hw->idx);
+       writel(val, xp->base + CCN_XP_DT_CONFIG);
+
+       spin_unlock(&ccn->dt.config_lock);
+}
+
+static void arm_ccn_pmu_event_start(struct perf_event *event, int flags)
+{
+       struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
+       struct hw_perf_event *hw = &event->hw;
+
+       local64_set(&event->hw.prev_count,
+                       arm_ccn_pmu_read_counter(ccn, hw->idx));
+       hw->state = 0;
+
+       if (!ccn->irq_used)
+               hrtimer_start(&ccn->dt.hrtimer, arm_ccn_pmu_timer_period(),
+                               HRTIMER_MODE_REL);
+
+       /* Set the DT bus input, engaging the counter */
+       arm_ccn_pmu_xp_dt_config(event, 1);
+}
+
+static void arm_ccn_pmu_event_stop(struct perf_event *event, int flags)
+{
+       struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
+       struct hw_perf_event *hw = &event->hw;
+       u64 timeout;
+
+       /* Disable counting, setting the DT bus to pass-through mode */
+       arm_ccn_pmu_xp_dt_config(event, 0);
+
+       if (!ccn->irq_used)
+               hrtimer_cancel(&ccn->dt.hrtimer);
+
+       /* Let the DT bus drain */
+       timeout = arm_ccn_pmu_read_counter(ccn, CCN_IDX_PMU_CYCLE_COUNTER) +
+                       ccn->num_xps;
+       while (arm_ccn_pmu_read_counter(ccn, CCN_IDX_PMU_CYCLE_COUNTER) <
+                       timeout)
+               cpu_relax();
+
+       if (flags & PERF_EF_UPDATE)
+               arm_ccn_pmu_event_update(event);
+
+       hw->state |= PERF_HES_STOPPED;
+}
+
+static void arm_ccn_pmu_xp_watchpoint_config(struct perf_event *event)
+{
+       struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
+       struct hw_perf_event *hw = &event->hw;
+       struct arm_ccn_component *source =
+                       ccn->dt.pmu_counters[hw->idx].source;
+       unsigned long wp = hw->config_base;
+       u32 val;
+       u64 cmp_l = event->attr.config1;
+       u64 cmp_h = event->attr.config2;
+       u64 mask_l = ccn->dt.cmp_mask[CCN_CONFIG_MASK(event->attr.config)].l;
+       u64 mask_h = ccn->dt.cmp_mask[CCN_CONFIG_MASK(event->attr.config)].h;
+
+       hw->event_base = CCN_XP_DT_CONFIG__DT_CFG__WATCHPOINT(wp);
+
+       /* Direction (RX/TX), device (port) & virtual channel */
+       val = readl(source->base + CCN_XP_DT_INTERFACE_SEL);
+       val &= ~(CCN_XP_DT_INTERFACE_SEL__DT_IO_SEL__MASK <<
+                       CCN_XP_DT_INTERFACE_SEL__DT_IO_SEL__SHIFT(wp));
+       val |= CCN_CONFIG_DIR(event->attr.config) <<
+                       CCN_XP_DT_INTERFACE_SEL__DT_IO_SEL__SHIFT(wp);
+       val &= ~(CCN_XP_DT_INTERFACE_SEL__DT_DEV_SEL__MASK <<
+                       CCN_XP_DT_INTERFACE_SEL__DT_DEV_SEL__SHIFT(wp));
+       val |= CCN_CONFIG_PORT(event->attr.config) <<
+                       CCN_XP_DT_INTERFACE_SEL__DT_DEV_SEL__SHIFT(wp);
+       val &= ~(CCN_XP_DT_INTERFACE_SEL__DT_VC_SEL__MASK <<
+                       CCN_XP_DT_INTERFACE_SEL__DT_VC_SEL__SHIFT(wp));
+       val |= CCN_CONFIG_VC(event->attr.config) <<
+                       CCN_XP_DT_INTERFACE_SEL__DT_VC_SEL__SHIFT(wp);
+       writel(val, source->base + CCN_XP_DT_INTERFACE_SEL);
+
+       /* Comparison values */
+       writel(cmp_l & 0xffffffff, source->base + CCN_XP_DT_CMP_VAL_L(wp));
+       writel((cmp_l >> 32) & 0xefffffff,
+                       source->base + CCN_XP_DT_CMP_VAL_L(wp) + 4);
+       writel(cmp_h & 0xffffffff, source->base + CCN_XP_DT_CMP_VAL_H(wp));
+       writel((cmp_h >> 32) & 0x0fffffff,
+                       source->base + CCN_XP_DT_CMP_VAL_H(wp) + 4);
+
+       /* Mask */
+       writel(mask_l & 0xffffffff, source->base + CCN_XP_DT_CMP_MASK_L(wp));
+       writel((mask_l >> 32) & 0xefffffff,
+                       source->base + CCN_XP_DT_CMP_MASK_L(wp) + 4);
+       writel(mask_h & 0xffffffff, source->base + CCN_XP_DT_CMP_MASK_H(wp));
+       writel((mask_h >> 32) & 0x0fffffff,
+                       source->base + CCN_XP_DT_CMP_MASK_H(wp) + 4);
+}
+
+static void arm_ccn_pmu_xp_event_config(struct perf_event *event)
+{
+       struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
+       struct hw_perf_event *hw = &event->hw;
+       struct arm_ccn_component *source =
+                       ccn->dt.pmu_counters[hw->idx].source;
+       u32 val, id;
+
+       hw->event_base = CCN_XP_DT_CONFIG__DT_CFG__XP_PMU_EVENT(hw->config_base);
+
+       id = (CCN_CONFIG_VC(event->attr.config) << 4) |
+                       (CCN_CONFIG_PORT(event->attr.config) << 3) |
+                       (CCN_CONFIG_EVENT(event->attr.config) << 0);
+
+       val = readl(source->base + CCN_XP_PMU_EVENT_SEL);
+       val &= ~(CCN_XP_PMU_EVENT_SEL__ID__MASK <<
+                       CCN_XP_PMU_EVENT_SEL__ID__SHIFT(hw->config_base));
+       val |= id << CCN_XP_PMU_EVENT_SEL__ID__SHIFT(hw->config_base);
+       writel(val, source->base + CCN_XP_PMU_EVENT_SEL);
+}
+
+static void arm_ccn_pmu_node_event_config(struct perf_event *event)
+{
+       struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
+       struct hw_perf_event *hw = &event->hw;
+       struct arm_ccn_component *source =
+                       ccn->dt.pmu_counters[hw->idx].source;
+       u32 type = CCN_CONFIG_TYPE(event->attr.config);
+       u32 val, port;
+
+       port = arm_ccn_node_to_xp_port(CCN_CONFIG_NODE(event->attr.config));
+       hw->event_base = CCN_XP_DT_CONFIG__DT_CFG__DEVICE_PMU_EVENT(port,
+                       hw->config_base);
+
+       /* These *_event_sel regs should be identical, but let's make sure... */
+       BUILD_BUG_ON(CCN_HNF_PMU_EVENT_SEL != CCN_SBAS_PMU_EVENT_SEL);
+       BUILD_BUG_ON(CCN_SBAS_PMU_EVENT_SEL != CCN_RNI_PMU_EVENT_SEL);
+       BUILD_BUG_ON(CCN_HNF_PMU_EVENT_SEL__ID__SHIFT(1) !=
+                       CCN_SBAS_PMU_EVENT_SEL__ID__SHIFT(1));
+       BUILD_BUG_ON(CCN_SBAS_PMU_EVENT_SEL__ID__SHIFT(1) !=
+                       CCN_RNI_PMU_EVENT_SEL__ID__SHIFT(1));
+       BUILD_BUG_ON(CCN_HNF_PMU_EVENT_SEL__ID__MASK !=
+                       CCN_SBAS_PMU_EVENT_SEL__ID__MASK);
+       BUILD_BUG_ON(CCN_SBAS_PMU_EVENT_SEL__ID__MASK !=
+                       CCN_RNI_PMU_EVENT_SEL__ID__MASK);
+       if (WARN_ON(type != CCN_TYPE_HNF && type != CCN_TYPE_SBAS &&
+                       !arm_ccn_pmu_type_eq(type, CCN_TYPE_RNI_3P)))
+               return;
+
+       /* Set the event id for the pre-allocated counter */
+       val = readl(source->base + CCN_HNF_PMU_EVENT_SEL);
+       val &= ~(CCN_HNF_PMU_EVENT_SEL__ID__MASK <<
+               CCN_HNF_PMU_EVENT_SEL__ID__SHIFT(hw->config_base));
+       val |= CCN_CONFIG_EVENT(event->attr.config) <<
+               CCN_HNF_PMU_EVENT_SEL__ID__SHIFT(hw->config_base);
+       writel(val, source->base + CCN_HNF_PMU_EVENT_SEL);
+}
+
+static void arm_ccn_pmu_event_config(struct perf_event *event)
+{
+       struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
+       struct hw_perf_event *hw = &event->hw;
+       u32 xp, offset, val;
+
+       /* Cycle counter requires no setup */
+       if (hw->idx == CCN_IDX_PMU_CYCLE_COUNTER)
+               return;
+
+       if (CCN_CONFIG_TYPE(event->attr.config) == CCN_TYPE_XP)
+               xp = CCN_CONFIG_XP(event->attr.config);
+       else
+               xp = arm_ccn_node_to_xp(CCN_CONFIG_NODE(event->attr.config));
+
+       spin_lock(&ccn->dt.config_lock);
+
+       /* Set the DT bus "distance" register */
+       offset = (hw->idx / 4) * 4;
+       val = readl(ccn->dt.base + CCN_DT_ACTIVE_DSM + offset);
+       val &= ~(CCN_DT_ACTIVE_DSM__DSM_ID__MASK <<
+                       CCN_DT_ACTIVE_DSM__DSM_ID__SHIFT(hw->idx % 4));
+       val |= xp << CCN_DT_ACTIVE_DSM__DSM_ID__SHIFT(hw->idx % 4);
+       writel(val, ccn->dt.base + CCN_DT_ACTIVE_DSM + offset);
+
+       if (CCN_CONFIG_TYPE(event->attr.config) == CCN_TYPE_XP) {
+               if (CCN_CONFIG_EVENT(event->attr.config) ==
+                               CCN_EVENT_WATCHPOINT)
+                       arm_ccn_pmu_xp_watchpoint_config(event);
+               else
+                       arm_ccn_pmu_xp_event_config(event);
+       } else {
+               arm_ccn_pmu_node_event_config(event);
+       }
+
+       spin_unlock(&ccn->dt.config_lock);
+}
+
+static int arm_ccn_pmu_event_add(struct perf_event *event, int flags)
+{
+       struct hw_perf_event *hw = &event->hw;
+
+       arm_ccn_pmu_event_config(event);
+
+       hw->state = PERF_HES_STOPPED;
+
+       if (flags & PERF_EF_START)
+               arm_ccn_pmu_event_start(event, PERF_EF_UPDATE);
+
+       return 0;
+}
+
+static void arm_ccn_pmu_event_del(struct perf_event *event, int flags)
+{
+       arm_ccn_pmu_event_stop(event, PERF_EF_UPDATE);
+
+       arm_ccn_pmu_event_free(event);
+}
+
+static void arm_ccn_pmu_event_read(struct perf_event *event)
+{
+       arm_ccn_pmu_event_update(event);
+}
+
+static irqreturn_t arm_ccn_pmu_overflow_handler(struct arm_ccn_dt *dt)
+{
+       u32 pmovsr = readl(dt->base + CCN_DT_PMOVSR);
+       int idx;
+
+       if (!pmovsr)
+               return IRQ_NONE;
+
+       writel(pmovsr, dt->base + CCN_DT_PMOVSR_CLR);
+
+       BUILD_BUG_ON(CCN_IDX_PMU_CYCLE_COUNTER != CCN_NUM_PMU_EVENT_COUNTERS);
+
+       for (idx = 0; idx < CCN_NUM_PMU_EVENT_COUNTERS + 1; idx++) {
+               struct perf_event *event = dt->pmu_counters[idx].event;
+               int overflowed = pmovsr & BIT(idx);
+
+               WARN_ON_ONCE(overflowed && !event);
+
+               if (!event || !overflowed)
+                       continue;
+
+               arm_ccn_pmu_event_update(event);
+       }
+
+       return IRQ_HANDLED;
+}
+
+static enum hrtimer_restart arm_ccn_pmu_timer_handler(struct hrtimer *hrtimer)
+{
+       struct arm_ccn_dt *dt = container_of(hrtimer, struct arm_ccn_dt,
+                       hrtimer);
+       unsigned long flags;
+
+       local_irq_save(flags);
+       arm_ccn_pmu_overflow_handler(dt);
+       local_irq_restore(flags);
+
+       hrtimer_forward_now(hrtimer, arm_ccn_pmu_timer_period());
+       return HRTIMER_RESTART;
+}
+
+
+static DEFINE_IDA(arm_ccn_pmu_ida);
+
+static int arm_ccn_pmu_init(struct arm_ccn *ccn)
+{
+       int i;
+       char *name;
+
+       /* Initialize DT subsystem */
+       ccn->dt.base = ccn->base + CCN_REGION_SIZE;
+       spin_lock_init(&ccn->dt.config_lock);
+       writel(CCN_DT_CTL__DT_EN, ccn->dt.base + CCN_DT_CTL);
+       writel(CCN_DT_PMCR__OVFL_INTR_EN | CCN_DT_PMCR__PMU_EN,
+                       ccn->dt.base + CCN_DT_PMCR);
+       writel(0x1, ccn->dt.base + CCN_DT_PMSR_CLR);
+       for (i = 0; i < ccn->num_xps; i++) {
+               writel(0, ccn->xp[i].base + CCN_XP_DT_CONFIG);
+               writel((CCN_XP_DT_CONTROL__WP_ARM_SEL__ALWAYS <<
+                               CCN_XP_DT_CONTROL__WP_ARM_SEL__SHIFT(0)) |
+                               (CCN_XP_DT_CONTROL__WP_ARM_SEL__ALWAYS <<
+                               CCN_XP_DT_CONTROL__WP_ARM_SEL__SHIFT(1)) |
+                               CCN_XP_DT_CONTROL__DT_ENABLE,
+                               ccn->xp[i].base + CCN_XP_DT_CONTROL);
+       }
+       ccn->dt.cmp_mask[CCN_IDX_MASK_ANY].l = ~0;
+       ccn->dt.cmp_mask[CCN_IDX_MASK_ANY].h = ~0;
+       ccn->dt.cmp_mask[CCN_IDX_MASK_EXACT].l = 0;
+       ccn->dt.cmp_mask[CCN_IDX_MASK_EXACT].h = 0;
+       ccn->dt.cmp_mask[CCN_IDX_MASK_ORDER].l = ~0;
+       ccn->dt.cmp_mask[CCN_IDX_MASK_ORDER].h = ~(0x1 << 15);
+       ccn->dt.cmp_mask[CCN_IDX_MASK_OPCODE].l = ~0;
+       ccn->dt.cmp_mask[CCN_IDX_MASK_OPCODE].h = ~(0x1f << 9);
+
+       /* Get a convenient /sys/event_source/devices/ name */
+       ccn->dt.id = ida_simple_get(&arm_ccn_pmu_ida, 0, 0, GFP_KERNEL);
+       if (ccn->dt.id == 0) {
+               name = "ccn";
+       } else {
+               int len = snprintf(NULL, 0, "ccn_%d", ccn->dt.id);
+
+               name = devm_kzalloc(ccn->dev, len + 1, GFP_KERNEL);
+               snprintf(name, len + 1, "ccn_%d", ccn->dt.id);
+       }
+
+       /* Perf driver registration */
+       ccn->dt.pmu = (struct pmu) {
+               .attr_groups = arm_ccn_pmu_attr_groups,
+               .task_ctx_nr = perf_invalid_context,
+               .event_init = arm_ccn_pmu_event_init,
+               .add = arm_ccn_pmu_event_add,
+               .del = arm_ccn_pmu_event_del,
+               .start = arm_ccn_pmu_event_start,
+               .stop = arm_ccn_pmu_event_stop,
+               .read = arm_ccn_pmu_event_read,
+       };
+
+       /* No overflow interrupt? Have to use a timer instead. */
+       if (!ccn->irq_used) {
+               dev_info(ccn->dev, "No access to interrupts, using timer.\n");
+               hrtimer_init(&ccn->dt.hrtimer, CLOCK_MONOTONIC,
+                               HRTIMER_MODE_REL);
+               ccn->dt.hrtimer.function = arm_ccn_pmu_timer_handler;
+       }
+
+       return perf_pmu_register(&ccn->dt.pmu, name, -1);
+}
+
+static void arm_ccn_pmu_cleanup(struct arm_ccn *ccn)
+{
+       int i;
+
+       for (i = 0; i < ccn->num_xps; i++)
+               writel(0, ccn->xp[i].base + CCN_XP_DT_CONTROL);
+       writel(0, ccn->dt.base + CCN_DT_PMCR);
+       perf_pmu_unregister(&ccn->dt.pmu);
+       ida_simple_remove(&arm_ccn_pmu_ida, ccn->dt.id);
+}
+
+
+static int arm_ccn_for_each_valid_region(struct arm_ccn *ccn,
+               int (*callback)(struct arm_ccn *ccn, int region,
+               void __iomem *base, u32 type, u32 id))
+{
+       int region;
+
+       for (region = 0; region < CCN_NUM_REGIONS; region++) {
+               u32 val, type, id;
+               void __iomem *base;
+               int err;
+
+               val = readl(ccn->base + CCN_MN_OLY_COMP_LIST_63_0 +
+                               4 * (region / 32));
+               if (!(val & (1 << (region % 32))))
+                       continue;
+
+               base = ccn->base + region * CCN_REGION_SIZE;
+               val = readl(base + CCN_ALL_OLY_ID);
+               type = (val >> CCN_ALL_OLY_ID__OLY_ID__SHIFT) &
+                               CCN_ALL_OLY_ID__OLY_ID__MASK;
+               id = (val >> CCN_ALL_OLY_ID__NODE_ID__SHIFT) &
+                               CCN_ALL_OLY_ID__NODE_ID__MASK;
+
+               err = callback(ccn, region, base, type, id);
+               if (err)
+                       return err;
+       }
+
+       return 0;
+}
+
+static int arm_ccn_get_nodes_num(struct arm_ccn *ccn, int region,
+               void __iomem *base, u32 type, u32 id)
+{
+
+       if (type == CCN_TYPE_XP && id >= ccn->num_xps)
+               ccn->num_xps = id + 1;
+       else if (id >= ccn->num_nodes)
+               ccn->num_nodes = id + 1;
+
+       return 0;
+}
+
+static int arm_ccn_init_nodes(struct arm_ccn *ccn, int region,
+               void __iomem *base, u32 type, u32 id)
+{
+       struct arm_ccn_component *component;
+
+       dev_dbg(ccn->dev, "Region %d: id=%u, type=0x%02x\n", region, id, type);
+
+       switch (type) {
+       case CCN_TYPE_MN:
+       case CCN_TYPE_DT:
+               return 0;
+       case CCN_TYPE_XP:
+               component = &ccn->xp[id];
+               break;
+       case CCN_TYPE_SBSX:
+               ccn->sbsx_present = 1;
+               component = &ccn->node[id];
+               break;
+       case CCN_TYPE_SBAS:
+               ccn->sbas_present = 1;
+               /* Fall-through */
+       default:
+               component = &ccn->node[id];
+               break;
+       }
+
+       component->base = base;
+       component->type = type;
+
+       return 0;
+}
+
+
+static irqreturn_t arm_ccn_error_handler(struct arm_ccn *ccn,
+               const u32 *err_sig_val)
+{
+       /* This should be really handled by firmware... */
+       dev_err(ccn->dev, "Error reported in %08x%08x%08x%08x%08x%08x.\n",
+                       err_sig_val[5], err_sig_val[4], err_sig_val[3],
+                       err_sig_val[2], err_sig_val[1], err_sig_val[0]);
+       dev_err(ccn->dev, "Disabling interrupt generation for all errors.\n");
+       writel(CCN_MN_ERRINT_STATUS__ALL_ERRORS__DISABLE,
+                       ccn->base + CCN_MN_ERRINT_STATUS);
+
+       return IRQ_HANDLED;
+}
+
+
+static irqreturn_t arm_ccn_irq_handler(int irq, void *dev_id)
+{
+       irqreturn_t res = IRQ_NONE;
+       struct arm_ccn *ccn = dev_id;
+       u32 err_sig_val[6];
+       u32 err_or;
+       int i;
+
+       /* PMU overflow is a special case */
+       err_or = err_sig_val[0] = readl(ccn->base + CCN_MN_ERR_SIG_VAL_63_0);
+       if (err_or & CCN_MN_ERR_SIG_VAL_63_0__DT) {
+               err_or &= ~CCN_MN_ERR_SIG_VAL_63_0__DT;
+               res = arm_ccn_pmu_overflow_handler(&ccn->dt);
+       }
+
+       /* Have to read all err_sig_vals to clear them */
+       for (i = 1; i < ARRAY_SIZE(err_sig_val); i++) {
+               err_sig_val[i] = readl(ccn->base +
+                               CCN_MN_ERR_SIG_VAL_63_0 + i * 4);
+               err_or |= err_sig_val[i];
+       }
+       if (err_or)
+               res |= arm_ccn_error_handler(ccn, err_sig_val);
+
+       if (res != IRQ_NONE)
+               writel(CCN_MN_ERRINT_STATUS__INTREQ__DESSERT,
+                               ccn->base + CCN_MN_ERRINT_STATUS);
+
+       return res;
+}
+
+
+static int arm_ccn_probe(struct platform_device *pdev)
+{
+       struct arm_ccn *ccn;
+       struct resource *res;
+       int err;
+
+       ccn = devm_kzalloc(&pdev->dev, sizeof(*ccn), GFP_KERNEL);
+       if (!ccn)
+               return -ENOMEM;
+       ccn->dev = &pdev->dev;
+       platform_set_drvdata(pdev, ccn);
+
+       res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
+       if (!res)
+               return -EINVAL;
+
+       if (!devm_request_mem_region(ccn->dev, res->start,
+                       resource_size(res), pdev->name))
+               return -EBUSY;
+
+       ccn->base = devm_ioremap(ccn->dev, res->start,
+                               resource_size(res));
+       if (!ccn->base)
+               return -EFAULT;
+
+       res = platform_get_resource(pdev, IORESOURCE_IRQ, 0);
+       if (!res)
+               return -EINVAL;
+
+       /* Check if we can use the interrupt */
+       writel(CCN_MN_ERRINT_STATUS__PMU_EVENTS__DISABLE,
+                       ccn->base + CCN_MN_ERRINT_STATUS);
+       if (readl(ccn->base + CCN_MN_ERRINT_STATUS) &
+                       CCN_MN_ERRINT_STATUS__PMU_EVENTS__DISABLED) {
+               /* Can set 'disable' bits, so can acknowledge interrupts */
+               writel(CCN_MN_ERRINT_STATUS__PMU_EVENTS__ENABLE,
+                               ccn->base + CCN_MN_ERRINT_STATUS);
+               err = devm_request_irq(ccn->dev, res->start,
+                               arm_ccn_irq_handler, 0, dev_name(ccn->dev),
+                               ccn);
+               if (err)
+                       return err;
+
+               ccn->irq_used = 1;
+       }
+
+
+       /* Build topology */
+
+       err = arm_ccn_for_each_valid_region(ccn, arm_ccn_get_nodes_num);
+       if (err)
+               return err;
+
+       ccn->node = devm_kzalloc(ccn->dev, sizeof(*ccn->node) * ccn->num_nodes,
+               GFP_KERNEL);
+       ccn->xp = devm_kzalloc(ccn->dev, sizeof(*ccn->node) * ccn->num_xps,
+               GFP_KERNEL);
+       if (!ccn->node || !ccn->xp)
+               return -ENOMEM;
+
+       err = arm_ccn_for_each_valid_region(ccn, arm_ccn_init_nodes);
+       if (err)
+               return err;
+
+       return arm_ccn_pmu_init(ccn);
+}
+
+static int arm_ccn_remove(struct platform_device *pdev)
+{
+       struct arm_ccn *ccn = platform_get_drvdata(pdev);
+
+       arm_ccn_pmu_cleanup(ccn);
+
+       return 0;
+}
+
+static const struct of_device_id arm_ccn_match[] = {
+       { .compatible = "arm,ccn-504", },
+       {},
+};
+
+static struct platform_driver arm_ccn_driver = {
+       .driver = {
+               .name = "arm-ccn",
+               .of_match_table = arm_ccn_match,
+       },
+       .probe = arm_ccn_probe,
+       .remove = arm_ccn_remove,
+};
+
+static int __init arm_ccn_init(void)
+{
+       int i;
+
+       for (i = 0; i < ARRAY_SIZE(arm_ccn_pmu_events); i++)
+               arm_ccn_pmu_events_attrs[i] = &arm_ccn_pmu_events[i].attr.attr;
+
+       return platform_driver_register(&arm_ccn_driver);
+}
+
+static void __exit arm_ccn_exit(void)
+{
+       platform_driver_unregister(&arm_ccn_driver);
+}
+
+module_init(arm_ccn_init);
+module_exit(arm_ccn_exit);
+
+MODULE_AUTHOR("Pawel Moll <pawel.moll@arm.com>");
+MODULE_LICENSE("GPL");