Add support for various PMU counters found on the Cavium ThunderX and
OcteonTx SoC.
The driver provides common "uncore" functions to avoid code duplication and
support adding more device PMUs (like L2 cache) in the future.
Probe and removal of the PMUs is done by hooking into the ThunderX EDAC
driver as this driver owns all the PCI devices containing the PMU counters.
Patches are on top of 4.12-rc6.
Changes to v5:
- Only allow built-in CONFIG_CAVIUM_PMU
- Drop unneeded export of perf_event_update_userpage()
- Simplify callbacks in edac code, move CONFIG_CAVIUM_PMU check
to header file
- Fix some sparse static warnings
- Add documentation
- Fix OCX TLK event_valid check
- Add group validation in event_init
- Add a guard variable to prevent calling init twice
- Use kasprintf and fix pmu name allocation
- Remove unneeded check for embedded pmu
- Loop around local64_cmpxchg
- Simplify cvm_pmu_lmc_event_valid
- Fix list_add error case
Jan Glauber (3):
perf: cavium: Support memory controller PMU counters
perf: cavium: Support transmit-link PMU counters
perf: cavium: Add Documentation
Documentation/perf/cavium-pmu.txt | 74 +++++
drivers/edac/thunderx_edac.c | 13 +-
drivers/perf/Kconfig | 9 +
drivers/perf/Makefile | 1 +
drivers/perf/cavium_pmu.c | 643 ++++++++++++++++++++++++++++++++++++++
include/linux/cpuhotplug.h | 1 +
include/linux/perf/cavium_pmu.h | 54 ++++
7 files changed, 794 insertions(+), 1 deletion(-)
create mode 100644 Documentation/perf/cavium-pmu.txt
create mode 100644 drivers/perf/cavium_pmu.c
create mode 100644 include/linux/perf/cavium_pmu.h
--
2.9.0.rc0.21.g7777322
Add support for the PMU counters on Cavium SOC memory controllers.
This patch also adds generic functions to allow supporting more
devices with PMU counters.
Properties of the LMC PMU counters:
- not stoppable
- fixed purpose
- read-only
- one PCI device per memory controller
Signed-off-by: Jan Glauber <[email protected]>
---
drivers/edac/thunderx_edac.c | 7 +-
drivers/perf/Kconfig | 9 +
drivers/perf/Makefile | 1 +
drivers/perf/cavium_pmu.c | 418 ++++++++++++++++++++++++++++++++++++++++
include/linux/cpuhotplug.h | 1 +
include/linux/perf/cavium_pmu.h | 53 +++++
6 files changed, 488 insertions(+), 1 deletion(-)
create mode 100644 drivers/perf/cavium_pmu.c
create mode 100644 include/linux/perf/cavium_pmu.h
diff --git a/drivers/edac/thunderx_edac.c b/drivers/edac/thunderx_edac.c
index 86d585c..884499a 100644
--- a/drivers/edac/thunderx_edac.c
+++ b/drivers/edac/thunderx_edac.c
@@ -20,7 +20,7 @@
#include <linux/atomic.h>
#include <linux/bitfield.h>
#include <linux/circ_buf.h>
-
+#include <linux/perf/cavium_pmu.h>
#include <asm/page.h>
#include "edac_module.h"
@@ -209,6 +209,8 @@ struct thunderx_lmc {
struct lmc_err_ctx err_ctx[RING_ENTRIES];
unsigned long ring_head;
unsigned long ring_tail;
+
+ void *pmu_data;
};
#define ring_pos(pos, size) ((pos) & (size - 1))
@@ -810,6 +812,7 @@ static int thunderx_lmc_probe(struct pci_dev *pdev,
}
}
+ lmc->pmu_data = cvm_pmu_probe(pdev, lmc->regs, CVM_PMU_LMC);
return 0;
err_free:
@@ -824,6 +827,8 @@ static void thunderx_lmc_remove(struct pci_dev *pdev)
struct mem_ctl_info *mci = pci_get_drvdata(pdev);
struct thunderx_lmc *lmc = mci->pvt_info;
+ cvm_pmu_remove(pdev, lmc->pmu_data, CVM_PMU_LMC);
+
writeq(LMC_INT_ENA_ALL, lmc->regs + LMC_INT_ENA_W1C);
edac_mc_del_mc(&pdev->dev);
diff --git a/drivers/perf/Kconfig b/drivers/perf/Kconfig
index aa587ed..bd0d89f 100644
--- a/drivers/perf/Kconfig
+++ b/drivers/perf/Kconfig
@@ -42,4 +42,13 @@ config XGENE_PMU
help
Say y if you want to use APM X-Gene SoC performance monitors.
+config CAVIUM_PMU
+ bool "Cavium SOC PMU"
+ default y
+ depends on EDAC_THUNDERX && PERF_EVENTS
+ help
+ Provides access to various performance counters on Caviums
+ ARM64 SOCs. Adds support for memory controller (LMC) and
+ interconnect link (OCX TLK) counters.
+
endmenu
diff --git a/drivers/perf/Makefile b/drivers/perf/Makefile
index 6420bd4..b304646 100644
--- a/drivers/perf/Makefile
+++ b/drivers/perf/Makefile
@@ -3,3 +3,4 @@ obj-$(CONFIG_ARM_PMU_ACPI) += arm_pmu_acpi.o
obj-$(CONFIG_QCOM_L2_PMU) += qcom_l2_pmu.o
obj-$(CONFIG_QCOM_L3_PMU) += qcom_l3_pmu.o
obj-$(CONFIG_XGENE_PMU) += xgene_pmu.o
+obj-$(CONFIG_CAVIUM_PMU) += cavium_pmu.o
diff --git a/drivers/perf/cavium_pmu.c b/drivers/perf/cavium_pmu.c
new file mode 100644
index 0000000..4a85c7fac
--- /dev/null
+++ b/drivers/perf/cavium_pmu.c
@@ -0,0 +1,418 @@
+/*
+ * Cavium ARM SOC "uncore" PMU counters
+ *
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License. See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright Cavium, Inc. 2017
+ * Author(s): Jan Glauber <[email protected]>
+ *
+ */
+#include <linux/cpumask.h>
+#include <linux/export.h>
+#include <linux/list.h>
+#include <linux/mutex.h>
+#include <linux/pci.h>
+#include <linux/perf_event.h>
+#include <linux/slab.h>
+
+#include <linux/perf/cavium_pmu.h>
+
+static bool cvm_pmu_initialized;
+static struct list_head cvm_pmu_lmcs;
+
+/*
+ * Common Cavium PMU stuff
+ *
+ * Shared properties of the different PMU types:
+ * - all counters are 64 bit long
+ * - there are no overflow interrupts
+ * - all devices with PMU counters appear as PCI devices
+ *
+ * Counter control, access and device association depends on the
+ * PMU type.
+ */
+
+#define to_pmu_dev(x) container_of((x), struct cvm_pmu_dev, pmu)
+
+static int cvm_pmu_event_init(struct perf_event *event)
+{
+ struct hw_perf_event *hwc = &event->hw;
+ struct cvm_pmu_dev *pmu_dev;
+ struct perf_event *sibling;
+
+ if (event->attr.type != event->pmu->type)
+ return -ENOENT;
+
+ /* we do not support sampling */
+ if (is_sampling_event(event))
+ return -EINVAL;
+
+ /* PMU counters do not support any these bits */
+ if (event->attr.exclude_user ||
+ event->attr.exclude_kernel ||
+ event->attr.exclude_host ||
+ event->attr.exclude_guest ||
+ event->attr.exclude_hv ||
+ event->attr.exclude_idle)
+ return -EINVAL;
+
+ pmu_dev = to_pmu_dev(event->pmu);
+ if (!pmu_dev->event_valid(event->attr.config))
+ return -EINVAL;
+
+ /*
+ * Forbid groups containing mixed PMUs, 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;
+
+ hwc->config = event->attr.config;
+ hwc->idx = -1;
+ return 0;
+}
+
+static void cvm_pmu_read(struct perf_event *event)
+{
+ struct cvm_pmu_dev *pmu_dev = to_pmu_dev(event->pmu);
+ struct hw_perf_event *hwc = &event->hw;
+ u64 prev, delta, new;
+
+again:
+ prev = local64_read(&hwc->prev_count);
+ new = readq(hwc->event_base + pmu_dev->map);
+
+ if (local64_cmpxchg(&hwc->prev_count, prev, new) != prev)
+ goto again;
+
+ delta = new - prev;
+ local64_add(delta, &event->count);
+}
+
+static void cvm_pmu_start(struct perf_event *event, int flags)
+{
+ struct cvm_pmu_dev *pmu_dev = to_pmu_dev(event->pmu);
+ struct hw_perf_event *hwc = &event->hw;
+ u64 new;
+
+ if (WARN_ON_ONCE(!(hwc->state & PERF_HES_STOPPED)))
+ return;
+
+ WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
+ hwc->state = 0;
+
+ /* update prev_count always in order support unstoppable counters */
+ new = readq(hwc->event_base + pmu_dev->map);
+ local64_set(&hwc->prev_count, new);
+
+ perf_event_update_userpage(event);
+}
+
+static void cvm_pmu_stop(struct perf_event *event, int flags)
+{
+ struct hw_perf_event *hwc = &event->hw;
+
+ WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
+ hwc->state |= PERF_HES_STOPPED;
+
+ if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
+ cvm_pmu_read(event);
+ hwc->state |= PERF_HES_UPTODATE;
+ }
+}
+
+static int cvm_pmu_add(struct perf_event *event, int flags, u64 config_base,
+ u64 event_base)
+{
+ struct cvm_pmu_dev *pmu_dev = to_pmu_dev(event->pmu);
+ struct hw_perf_event *hwc = &event->hw;
+
+ if (!cmpxchg(&pmu_dev->events[hwc->config], NULL, event))
+ hwc->idx = hwc->config;
+
+ if (hwc->idx == -1)
+ return -EBUSY;
+
+ hwc->config_base = config_base;
+ hwc->event_base = event_base;
+ hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
+
+ if (flags & PERF_EF_START)
+ pmu_dev->pmu.start(event, PERF_EF_RELOAD);
+
+ return 0;
+}
+
+static void cvm_pmu_del(struct perf_event *event, int flags)
+{
+ struct cvm_pmu_dev *pmu_dev = to_pmu_dev(event->pmu);
+ struct hw_perf_event *hwc = &event->hw;
+ int i;
+
+ event->pmu->stop(event, PERF_EF_UPDATE);
+
+ /*
+ * For programmable counters we need to check where we installed it.
+ * To keep this function generic always test the more complicated
+ * case (free running counters won't need the loop).
+ */
+ for (i = 0; i < pmu_dev->num_counters; i++)
+ if (cmpxchg(&pmu_dev->events[i], event, NULL) == event)
+ break;
+
+ perf_event_update_userpage(event);
+ hwc->idx = -1;
+}
+
+static ssize_t cvm_pmu_event_sysfs_show(struct device *dev,
+ struct device_attribute *attr,
+ char *page)
+{
+ struct perf_pmu_events_attr *pmu_attr =
+ container_of(attr, struct perf_pmu_events_attr, attr);
+
+ if (pmu_attr->event_str)
+ return sprintf(page, "%s", pmu_attr->event_str);
+
+ return 0;
+}
+
+/*
+ * The pmu events are independent from CPUs. Provide a cpumask
+ * nevertheless to prevent perf from adding the event per-cpu and just
+ * set the mask to one online CPU. Use the same cpumask for all "uncore"
+ * devices.
+ *
+ * There is a performance penalty for accessing a device from a CPU on
+ * another socket, but we do not care.
+ */
+static int cvm_pmu_offline_cpu(unsigned int old_cpu, struct hlist_node *node)
+{
+ struct cvm_pmu_dev *pmu_dev;
+ int new_cpu;
+
+ pmu_dev = hlist_entry_safe(node, struct cvm_pmu_dev, cpuhp_node);
+ if (!cpumask_test_and_clear_cpu(old_cpu, &pmu_dev->active_mask))
+ return 0;
+
+ new_cpu = cpumask_any_but(cpu_online_mask, old_cpu);
+ if (new_cpu >= nr_cpu_ids)
+ return 0;
+ perf_pmu_migrate_context(&pmu_dev->pmu, old_cpu, new_cpu);
+ cpumask_set_cpu(new_cpu, &pmu_dev->active_mask);
+ return 0;
+}
+
+static ssize_t cvm_pmu_attr_show_cpumask(struct device *dev,
+ struct device_attribute *attr,
+ char *buf)
+{
+ struct pmu *pmu = dev_get_drvdata(dev);
+ struct cvm_pmu_dev *pmu_dev = container_of(pmu, struct cvm_pmu_dev, pmu);
+
+ return cpumap_print_to_pagebuf(true, buf, &pmu_dev->active_mask);
+}
+
+static DEVICE_ATTR(cpumask, S_IRUGO, cvm_pmu_attr_show_cpumask, NULL);
+
+static struct attribute *cvm_pmu_attrs[] = {
+ &dev_attr_cpumask.attr,
+ NULL,
+};
+
+static struct attribute_group cvm_pmu_attr_group = {
+ .attrs = cvm_pmu_attrs,
+};
+
+/*
+ * LMC (memory controller) counters:
+ * - not stoppable, always on, read-only
+ * - one PCI device per memory controller
+ */
+#define LMC_CONFIG_OFFSET 0x188
+#define LMC_CONFIG_RESET_BIT BIT(17)
+
+/* LMC events */
+#define LMC_EVENT_IFB_CNT 0x1d0
+#define LMC_EVENT_OPS_CNT 0x1d8
+#define LMC_EVENT_DCLK_CNT 0x1e0
+#define LMC_EVENT_BANK_CONFLICT1 0x360
+#define LMC_EVENT_BANK_CONFLICT2 0x368
+
+#define CVM_PMU_LMC_EVENT_ATTR(_name, _id) \
+ &((struct perf_pmu_events_attr[]) { \
+ { \
+ __ATTR(_name, S_IRUGO, cvm_pmu_event_sysfs_show, NULL), \
+ _id, \
+ "lmc_event=" __stringify(_id), \
+ } \
+ })[0].attr.attr
+
+/* map counter numbers to register offsets */
+static int lmc_events[] = {
+ LMC_EVENT_IFB_CNT,
+ LMC_EVENT_OPS_CNT,
+ LMC_EVENT_DCLK_CNT,
+ LMC_EVENT_BANK_CONFLICT1,
+ LMC_EVENT_BANK_CONFLICT2,
+};
+
+static int cvm_pmu_lmc_add(struct perf_event *event, int flags)
+{
+ struct hw_perf_event *hwc = &event->hw;
+
+ return cvm_pmu_add(event, flags, LMC_CONFIG_OFFSET,
+ lmc_events[hwc->config]);
+}
+
+PMU_FORMAT_ATTR(lmc_event, "config:0-2");
+
+static struct attribute *cvm_pmu_lmc_format_attr[] = {
+ &format_attr_lmc_event.attr,
+ NULL,
+};
+
+static struct attribute_group cvm_pmu_lmc_format_group = {
+ .name = "format",
+ .attrs = cvm_pmu_lmc_format_attr,
+};
+
+static struct attribute *cvm_pmu_lmc_events_attr[] = {
+ CVM_PMU_LMC_EVENT_ATTR(ifb_cnt, 0),
+ CVM_PMU_LMC_EVENT_ATTR(ops_cnt, 1),
+ CVM_PMU_LMC_EVENT_ATTR(dclk_cnt, 2),
+ CVM_PMU_LMC_EVENT_ATTR(bank_conflict1, 3),
+ CVM_PMU_LMC_EVENT_ATTR(bank_conflict2, 4),
+ NULL,
+};
+
+static struct attribute_group cvm_pmu_lmc_events_group = {
+ .name = "events",
+ .attrs = cvm_pmu_lmc_events_attr,
+};
+
+static const struct attribute_group *cvm_pmu_lmc_attr_groups[] = {
+ &cvm_pmu_attr_group,
+ &cvm_pmu_lmc_format_group,
+ &cvm_pmu_lmc_events_group,
+ NULL,
+};
+
+static bool cvm_pmu_lmc_event_valid(u64 config)
+{
+ return (config < ARRAY_SIZE(lmc_events));
+}
+
+static void *cvm_pmu_lmc_probe(struct pci_dev *pdev, void __iomem *regs)
+{
+ struct cvm_pmu_dev *next, *lmc;
+ int nr = 0, ret = -ENOMEM;
+
+ lmc = kzalloc(sizeof(*lmc), GFP_KERNEL);
+ if (!lmc)
+ goto fail_nomem;
+
+ list_for_each_entry(next, &cvm_pmu_lmcs, entry)
+ nr++;
+ lmc->pmu_name = kasprintf(GFP_KERNEL, "lmc%d", nr);
+
+ lmc->pdev = pdev;
+ lmc->map = regs;
+ lmc->num_counters = ARRAY_SIZE(lmc_events);
+ lmc->pmu = (struct pmu) {
+ .task_ctx_nr = perf_invalid_context,
+ .event_init = cvm_pmu_event_init,
+ .add = cvm_pmu_lmc_add,
+ .del = cvm_pmu_del,
+ .start = cvm_pmu_start,
+ .stop = cvm_pmu_stop,
+ .read = cvm_pmu_read,
+ .attr_groups = cvm_pmu_lmc_attr_groups,
+ };
+
+ cpuhp_state_add_instance_nocalls(CPUHP_AP_PERF_ARM_CVM_ONLINE,
+ &lmc->cpuhp_node);
+
+ /*
+ * perf PMU is CPU dependent so pick a random CPU and migrate away
+ * if it goes offline.
+ */
+ cpumask_set_cpu(smp_processor_id(), &lmc->active_mask);
+
+ ret = perf_pmu_register(&lmc->pmu, lmc->pmu_name, -1);
+ if (ret)
+ goto fail_hp;
+
+ list_add(&lmc->entry, &cvm_pmu_lmcs);
+
+ lmc->event_valid = cvm_pmu_lmc_event_valid;
+ dev_info(&pdev->dev, "Enabled %s PMU with %d counters\n",
+ lmc->pmu_name, lmc->num_counters);
+ return lmc;
+
+fail_hp:
+ kfree(lmc->pmu_name);
+ cpuhp_state_remove_instance(CPUHP_AP_PERF_ARM_CVM_ONLINE,
+ &lmc->cpuhp_node);
+ kfree(lmc);
+fail_nomem:
+ return ERR_PTR(ret);
+}
+EXPORT_SYMBOL_GPL(cvm_pmu_lmc_probe);
+
+static void cvm_pmu_lmc_remove(struct pci_dev *pdev, void *pmu_data)
+{
+ struct cvm_pmu_dev *lmc = (struct cvm_pmu_dev *) pmu_data;
+
+ cpuhp_state_remove_instance(CPUHP_AP_PERF_ARM_CVM_ONLINE,
+ &lmc->cpuhp_node);
+ list_del(&lmc->entry);
+ perf_pmu_unregister(&lmc->pmu);
+ kfree(lmc->pmu_name);
+ kfree(lmc);
+}
+EXPORT_SYMBOL_GPL(cvm_pmu_lmc_remove);
+
+static int cvm_pmu_init(void)
+{
+ if (cvm_pmu_initialized)
+ return 0;
+
+ INIT_LIST_HEAD(&cvm_pmu_lmcs);
+
+ return cpuhp_setup_state_multi(CPUHP_AP_PERF_ARM_CVM_ONLINE,
+ "perf/arm/cvm:online", NULL,
+ cvm_pmu_offline_cpu);
+}
+
+void *cvm_pmu_probe(struct pci_dev *pdev, void __iomem *regs, int type)
+{
+ if (!cvm_pmu_initialized) {
+ cvm_pmu_init();
+ cvm_pmu_initialized = true;
+ }
+
+ switch (type) {
+ case CVM_PMU_LMC:
+ return cvm_pmu_lmc_probe(pdev, regs);
+ }
+ return NULL;
+}
+EXPORT_SYMBOL_GPL(cvm_pmu_probe);
+
+void cvm_pmu_remove(struct pci_dev *pdev, void *pmu_data, int type)
+{
+ switch (type) {
+ case CVM_PMU_LMC:
+ return cvm_pmu_lmc_remove(pdev, pmu_data);
+ }
+}
+EXPORT_SYMBOL_GPL(cvm_pmu_remove);
diff --git a/include/linux/cpuhotplug.h b/include/linux/cpuhotplug.h
index 0f2a803..2fe906c 100644
--- a/include/linux/cpuhotplug.h
+++ b/include/linux/cpuhotplug.h
@@ -141,6 +141,7 @@ enum cpuhp_state {
CPUHP_AP_PERF_ARM_QCOM_L3_ONLINE,
CPUHP_AP_WORKQUEUE_ONLINE,
CPUHP_AP_RCUTREE_ONLINE,
+ CPUHP_AP_PERF_ARM_CVM_ONLINE,
CPUHP_AP_ONLINE_DYN,
CPUHP_AP_ONLINE_DYN_END = CPUHP_AP_ONLINE_DYN + 30,
CPUHP_AP_X86_HPET_ONLINE,
diff --git a/include/linux/perf/cavium_pmu.h b/include/linux/perf/cavium_pmu.h
new file mode 100644
index 0000000..6ddc1c6
--- /dev/null
+++ b/include/linux/perf/cavium_pmu.h
@@ -0,0 +1,53 @@
+#ifndef _CAVIUM_PMU_H
+#define _CAVIUM_PMU_H
+
+#include <linux/io.h>
+#include <linux/pci.h>
+
+enum cvm_pmu_type {
+ CVM_PMU_LMC,
+};
+
+#ifdef CONFIG_CAVIUM_PMU
+
+#include <linux/cpumask.h>
+#include <linux/list.h>
+#include <linux/perf_event.h>
+
+/* maximum number of parallel hardware counters for all pmu types */
+#define CVM_PMU_MAX_COUNTERS 64
+
+/* generic struct to cover the different pmu types */
+struct cvm_pmu_dev {
+ struct pmu pmu;
+ const char *pmu_name;
+ bool (*event_valid)(u64);
+ void __iomem *map;
+ struct pci_dev *pdev;
+ int num_counters;
+ struct perf_event *events[CVM_PMU_MAX_COUNTERS];
+ struct list_head entry;
+ struct hlist_node cpuhp_node;
+ cpumask_t active_mask;
+};
+
+
+/* PMU interface used by EDAC driver */
+void *cvm_pmu_probe(struct pci_dev *pdev, void __iomem *regs, int type);
+void cvm_pmu_remove(struct pci_dev *pdev, void *pmu_data, int type);
+
+#else
+static inline void *cvm_pmu_probe(struct pci_dev *pdev, void __iomem *regs,
+ int type)
+{
+ return NULL;
+}
+
+static inline void cvm_pmu_remove(struct pci_dev *pdev, void *pmu_data,
+ int type)
+{
+}
+
+#endif
+
+#endif
--
2.9.0.rc0.21.g7777322
Document Cavium SoC PMUs.
Signed-off-by: Jan Glauber <[email protected]>
---
Documentation/perf/cavium-pmu.txt | 74 +++++++++++++++++++++++++++++++++++++++
1 file changed, 74 insertions(+)
create mode 100644 Documentation/perf/cavium-pmu.txt
diff --git a/Documentation/perf/cavium-pmu.txt b/Documentation/perf/cavium-pmu.txt
new file mode 100644
index 0000000..37669b9
--- /dev/null
+++ b/Documentation/perf/cavium-pmu.txt
@@ -0,0 +1,74 @@
+Cavium ThunderX and OcteonTx Performance Monitoring Unit (PMU)
+==============================================================
+
+Cavium SoCs contain various system devices such as L2 caches, processor
+interconnect and memory controllers. Unfortunately the PMU counters
+are not following a common design so each device has a slightly different
+approach how to control and use the PMU counters.
+
+Common properties of all devices carrying PMU counters:
+- The devices are PCI devices and the counters are embedded somewhere
+ in the PCI register space.
+- All counters are 64 bit wide.
+- There are no overflow interrupts (unnecessary because of the 64 bit wide counters).
+
+Properties depending on the device type:
+- How to start/stop the counters
+- Programmable vs. fixed purpose counters
+- Stoppable vs. always running counters
+- Independent vs. grouped counters
+- Read-only vs. writable counters
+- PCI device to PMU group relationship
+
+
+Devices with PMU counters
+-------------------------
+
+Memory controller (LMC):
+- one PCI device per LMC
+- fixed-purpose counters
+- always running counters without start/stop/reset control
+- read-only counters
+
+CCPI interface controller (OCX) Transmit link (TLK) counters:
+- writable counters
+- only one PCI device exposes multiple TLK units (3 units on T88)
+- start/stop control per unit
+- only present on multi-socket systems
+
+PMU (perf) driver
+-----------------
+
+The cavium-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/<lmcX/ocx_tlkX>/.
+
+The "format" directory describes format of the config (event ID),
+The "events" directory shows the names of the events and provides configuration
+templates for all supported event types that can be used with perf tool. For
+example, "lmc0/dclk_cnt/" is an equivalent of "lmc0/config=2/".
+
+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 lmc
+ lmc0/bank_conflict1/ [Kernel PMU event]
+ lmc0/bank_conflict2/ [Kernel PMU event]
+ lmc0/dclk_cnt/ [Kernel PMU event]
+ lmc0/ifb_cnt/ [Kernel PMU event]
+ lmc0/ops_cnt/ [Kernel PMU event]
+
+ / # perf stat -a -e lmc0/ops_cnt/,lmc0/dclk_cnt/ -- sleep 1
+
+ Performance counter stats for 'system wide':
+
+ 176,133 lmc0/ops_cnt/
+ 670,243,653 lmc0/dclk_cnt/
+
+ 1.005479295 seconds time elapsed
+
+The driver does not support sampling, therefore "perf record" will
+not work. System wide mode ("-a") must be used as per-task (without "-a")
+perf sessions are not supported.
--
2.9.0.rc0.21.g7777322
Add support for the transmit-link (OCX TLK) PMU counters found
on Caviums SOCs with a processor interconnect.
Properties of the OCX TLK counters:
- per-unit control
- fixed purpose
- writable
- one PCI device with multiple TLK units
Signed-off-by: Jan Glauber <[email protected]>
---
drivers/edac/thunderx_edac.c | 6 ++
drivers/perf/cavium_pmu.c | 225 ++++++++++++++++++++++++++++++++++++++++
include/linux/perf/cavium_pmu.h | 1 +
3 files changed, 232 insertions(+)
diff --git a/drivers/edac/thunderx_edac.c b/drivers/edac/thunderx_edac.c
index 884499a..6b7a43e 100644
--- a/drivers/edac/thunderx_edac.c
+++ b/drivers/edac/thunderx_edac.c
@@ -1094,6 +1094,8 @@ struct thunderx_ocx {
unsigned long link_ring_head;
unsigned long link_ring_tail;
+
+ void *pmu_data;
};
#define OCX_MESSAGE_SIZE SZ_1K
@@ -1491,6 +1493,8 @@ static int thunderx_ocx_probe(struct pci_dev *pdev,
writeq(OCX_COM_INT_ENA_ALL, ocx->regs + OCX_COM_INT_ENA_W1S);
+ ocx->pmu_data = cvm_pmu_probe(pdev, ocx->regs, CVM_PMU_TLK);
+
return 0;
err_free:
edac_device_free_ctl_info(edac_dev);
@@ -1504,6 +1508,8 @@ static void thunderx_ocx_remove(struct pci_dev *pdev)
struct thunderx_ocx *ocx = edac_dev->pvt_info;
int i;
+ cvm_pmu_remove(pdev, ocx->pmu_data, CVM_PMU_TLK);
+
writeq(OCX_COM_INT_ENA_ALL, ocx->regs + OCX_COM_INT_ENA_W1C);
for (i = 0; i < OCX_INTS; i++) {
diff --git a/drivers/perf/cavium_pmu.c b/drivers/perf/cavium_pmu.c
index 4a85c7fac..ce5648f 100644
--- a/drivers/perf/cavium_pmu.c
+++ b/drivers/perf/cavium_pmu.c
@@ -21,6 +21,7 @@
static bool cvm_pmu_initialized;
static struct list_head cvm_pmu_lmcs;
+static struct list_head cvm_pmu_tlks;
/*
* Common Cavium PMU stuff
@@ -381,12 +382,232 @@ static void cvm_pmu_lmc_remove(struct pci_dev *pdev, void *pmu_data)
}
EXPORT_SYMBOL_GPL(cvm_pmu_lmc_remove);
+/*
+ * CCPI interface controller (OCX) Transmit link (TLK) counters:
+ * - per-unit control
+ * - writable
+ * - one PCI device with multiple TLK units
+ */
+
+#define TLK_NR_UNITS 3
+#define TLK_UNIT_OFFSET 0x2000
+#define TLK_START_ADDR 0x10000
+#define TLK_STAT_CTL_OFFSET 0x40
+#define TLK_STAT_OFFSET 0x400
+
+#define TLK_STAT_ENABLE_BIT BIT(0)
+#define TLK_STAT_RESET_BIT BIT(1)
+
+#define CVM_PMU_TLK_EVENT_ATTR(_name, _id) \
+ &((struct perf_pmu_events_attr[]) { \
+ { \
+ __ATTR(_name, S_IRUGO, cvm_pmu_event_sysfs_show, NULL), \
+ _id, \
+ "tlk_event=" __stringify(_id), \
+ } \
+ })[0].attr.attr
+
+static void cvm_pmu_tlk_enable_pmu(struct pmu *pmu)
+{
+ struct cvm_pmu_dev *pmu_dev = container_of(pmu, struct cvm_pmu_dev, pmu);
+
+ /* enable all counters */
+ writeb(TLK_STAT_ENABLE_BIT, pmu_dev->map + TLK_STAT_CTL_OFFSET);
+}
+
+static void cvm_pmu_tlk_disable_pmu(struct pmu *pmu)
+{
+ struct cvm_pmu_dev *pmu_dev = container_of(pmu, struct cvm_pmu_dev, pmu);
+
+ /* disable all counters */
+ writeb(0, pmu_dev->map + TLK_STAT_CTL_OFFSET);
+}
+
+static int cvm_pmu_tlk_add(struct perf_event *event, int flags)
+{
+ struct hw_perf_event *hwc = &event->hw;
+
+ return cvm_pmu_add(event, flags, TLK_STAT_CTL_OFFSET,
+ TLK_STAT_OFFSET + hwc->config * 8);
+}
+
+PMU_FORMAT_ATTR(tlk_event, "config:0-5");
+
+static struct attribute *cvm_pmu_tlk_format_attr[] = {
+ &format_attr_tlk_event.attr,
+ NULL,
+};
+
+static struct attribute_group cvm_pmu_tlk_format_group = {
+ .name = "format",
+ .attrs = cvm_pmu_tlk_format_attr,
+};
+
+static struct attribute *cvm_pmu_tlk_events_attr[] = {
+ CVM_PMU_TLK_EVENT_ATTR(idle_cnt, 0x00),
+ CVM_PMU_TLK_EVENT_ATTR(data_cnt, 0x01),
+ CVM_PMU_TLK_EVENT_ATTR(sync_cnt, 0x02),
+ CVM_PMU_TLK_EVENT_ATTR(retry_cnt, 0x03),
+ CVM_PMU_TLK_EVENT_ATTR(err_cnt, 0x04),
+ CVM_PMU_TLK_EVENT_ATTR(mat0_cnt, 0x08),
+ CVM_PMU_TLK_EVENT_ATTR(mat1_cnt, 0x09),
+ CVM_PMU_TLK_EVENT_ATTR(mat2_cnt, 0x0a),
+ CVM_PMU_TLK_EVENT_ATTR(mat3_cnt, 0x0b),
+ CVM_PMU_TLK_EVENT_ATTR(vc0_cmd, 0x10),
+ CVM_PMU_TLK_EVENT_ATTR(vc1_cmd, 0x11),
+ CVM_PMU_TLK_EVENT_ATTR(vc2_cmd, 0x12),
+ CVM_PMU_TLK_EVENT_ATTR(vc3_cmd, 0x13),
+ CVM_PMU_TLK_EVENT_ATTR(vc4_cmd, 0x14),
+ CVM_PMU_TLK_EVENT_ATTR(vc5_cmd, 0x15),
+ CVM_PMU_TLK_EVENT_ATTR(vc0_pkt, 0x20),
+ CVM_PMU_TLK_EVENT_ATTR(vc1_pkt, 0x21),
+ CVM_PMU_TLK_EVENT_ATTR(vc2_pkt, 0x22),
+ CVM_PMU_TLK_EVENT_ATTR(vc3_pkt, 0x23),
+ CVM_PMU_TLK_EVENT_ATTR(vc4_pkt, 0x24),
+ CVM_PMU_TLK_EVENT_ATTR(vc5_pkt, 0x25),
+ CVM_PMU_TLK_EVENT_ATTR(vc6_pkt, 0x26),
+ CVM_PMU_TLK_EVENT_ATTR(vc7_pkt, 0x27),
+ CVM_PMU_TLK_EVENT_ATTR(vc8_pkt, 0x28),
+ CVM_PMU_TLK_EVENT_ATTR(vc9_pkt, 0x29),
+ CVM_PMU_TLK_EVENT_ATTR(vc10_pkt, 0x2a),
+ CVM_PMU_TLK_EVENT_ATTR(vc11_pkt, 0x2b),
+ CVM_PMU_TLK_EVENT_ATTR(vc12_pkt, 0x2c),
+ CVM_PMU_TLK_EVENT_ATTR(vc13_pkt, 0x2d),
+ CVM_PMU_TLK_EVENT_ATTR(vc0_con, 0x30),
+ CVM_PMU_TLK_EVENT_ATTR(vc1_con, 0x31),
+ CVM_PMU_TLK_EVENT_ATTR(vc2_con, 0x32),
+ CVM_PMU_TLK_EVENT_ATTR(vc3_con, 0x33),
+ CVM_PMU_TLK_EVENT_ATTR(vc4_con, 0x34),
+ CVM_PMU_TLK_EVENT_ATTR(vc5_con, 0x35),
+ CVM_PMU_TLK_EVENT_ATTR(vc6_con, 0x36),
+ CVM_PMU_TLK_EVENT_ATTR(vc7_con, 0x37),
+ CVM_PMU_TLK_EVENT_ATTR(vc8_con, 0x38),
+ CVM_PMU_TLK_EVENT_ATTR(vc9_con, 0x39),
+ CVM_PMU_TLK_EVENT_ATTR(vc10_con, 0x3a),
+ CVM_PMU_TLK_EVENT_ATTR(vc11_con, 0x3b),
+ CVM_PMU_TLK_EVENT_ATTR(vc12_con, 0x3c),
+ CVM_PMU_TLK_EVENT_ATTR(vc13_con, 0x3d),
+ NULL,
+};
+
+static struct attribute_group cvm_pmu_tlk_events_group = {
+ .name = "events",
+ .attrs = cvm_pmu_tlk_events_attr,
+};
+static const struct attribute_group *cvm_pmu_tlk_attr_groups[] = {
+ &cvm_pmu_attr_group,
+ &cvm_pmu_tlk_format_group,
+ &cvm_pmu_tlk_events_group,
+ NULL,
+};
+
+static bool cvm_pmu_tlk_event_valid(u64 config)
+{
+ struct perf_pmu_events_attr *attr;
+ int i;
+
+ for (i = 0; i < ARRAY_SIZE(cvm_pmu_tlk_events_attr) - 1; i++) {
+ attr = (struct perf_pmu_events_attr *)cvm_pmu_tlk_events_attr[i];
+ if (attr->id == config)
+ return true;
+ }
+ return false;
+}
+
+static void *cvm_pmu_tlk_probe_unit(struct pci_dev *pdev, void __iomem *regs,
+ int nr)
+{
+ struct cvm_pmu_dev *tlk;
+ int ret = -ENOMEM;
+
+ tlk = kzalloc(sizeof(*tlk), GFP_KERNEL);
+ if (!tlk)
+ goto fail_nomem;
+
+ tlk->pmu_name = kasprintf(GFP_KERNEL, "ocx_tlk%d", nr);
+
+ tlk->pdev = pdev;
+ tlk->map = regs + TLK_START_ADDR + nr * TLK_UNIT_OFFSET;
+ tlk->num_counters = ARRAY_SIZE(cvm_pmu_tlk_events_attr) - 1;
+ tlk->pmu = (struct pmu) {
+ .task_ctx_nr = perf_invalid_context,
+ .pmu_enable = cvm_pmu_tlk_enable_pmu,
+ .pmu_disable = cvm_pmu_tlk_disable_pmu,
+ .event_init = cvm_pmu_event_init,
+ .add = cvm_pmu_tlk_add,
+ .del = cvm_pmu_del,
+ .start = cvm_pmu_start,
+ .stop = cvm_pmu_stop,
+ .read = cvm_pmu_read,
+ .attr_groups = cvm_pmu_tlk_attr_groups,
+ };
+
+ cpuhp_state_add_instance_nocalls(CPUHP_AP_PERF_ARM_CVM_ONLINE,
+ &tlk->cpuhp_node);
+
+ /*
+ * perf PMU is CPU dependent so pick a random CPU and migrate away
+ * if it goes offline.
+ */
+ cpumask_set_cpu(smp_processor_id(), &tlk->active_mask);
+
+ ret = perf_pmu_register(&tlk->pmu, tlk->pmu_name, -1);
+ if (ret)
+ goto fail_hp;
+
+ list_add(&tlk->entry, &cvm_pmu_tlks);
+
+ tlk->event_valid = cvm_pmu_tlk_event_valid;
+ return tlk;
+
+fail_hp:
+ kfree(tlk->pmu_name);
+ cpuhp_state_remove_instance(CPUHP_AP_PERF_ARM_CVM_ONLINE,
+ &tlk->cpuhp_node);
+ kfree(tlk);
+fail_nomem:
+ return ERR_PTR(ret);
+}
+
+static void *cvm_pmu_tlk_probe(struct pci_dev *pdev, void __iomem *regs)
+{
+ struct cvm_pmu_dev *tlk;
+ int i;
+
+ for (i = 0; i < TLK_NR_UNITS; i++) {
+ tlk = cvm_pmu_tlk_probe_unit(pdev, regs, i);
+ if (PTR_ERR(tlk))
+ continue;
+ dev_info(&pdev->dev, "Enabled %s PMU with %d counters\n",
+ tlk->pmu_name, tlk->num_counters);
+ }
+ return &cvm_pmu_tlks;
+}
+
+static void cvm_pmu_tlk_remove(struct pci_dev *pdev, void *pmu_data)
+{
+ struct list_head *l, *tmp;
+ struct cvm_pmu_dev *tlk;
+
+ list_for_each_safe(l, tmp, &cvm_pmu_tlks) {
+ tlk = list_entry(l, struct cvm_pmu_dev, entry);
+
+ list_del(&tlk->entry);
+ cpuhp_state_remove_instance(CPUHP_AP_PERF_ARM_CVM_ONLINE,
+ &tlk->cpuhp_node);
+ perf_pmu_unregister(&tlk->pmu);
+ kfree(tlk->pmu_name);
+ kfree(tlk);
+ }
+}
+
static int cvm_pmu_init(void)
{
if (cvm_pmu_initialized)
return 0;
INIT_LIST_HEAD(&cvm_pmu_lmcs);
+ INIT_LIST_HEAD(&cvm_pmu_tlks);
return cpuhp_setup_state_multi(CPUHP_AP_PERF_ARM_CVM_ONLINE,
"perf/arm/cvm:online", NULL,
@@ -403,6 +624,8 @@ void *cvm_pmu_probe(struct pci_dev *pdev, void __iomem *regs, int type)
switch (type) {
case CVM_PMU_LMC:
return cvm_pmu_lmc_probe(pdev, regs);
+ case CVM_PMU_TLK:
+ return cvm_pmu_tlk_probe(pdev, regs);
}
return NULL;
}
@@ -413,6 +636,8 @@ void cvm_pmu_remove(struct pci_dev *pdev, void *pmu_data, int type)
switch (type) {
case CVM_PMU_LMC:
return cvm_pmu_lmc_remove(pdev, pmu_data);
+ case CVM_PMU_TLK:
+ return cvm_pmu_tlk_remove(pdev, pmu_data);
}
}
EXPORT_SYMBOL_GPL(cvm_pmu_remove);
diff --git a/include/linux/perf/cavium_pmu.h b/include/linux/perf/cavium_pmu.h
index 6ddc1c6..d15b435 100644
--- a/include/linux/perf/cavium_pmu.h
+++ b/include/linux/perf/cavium_pmu.h
@@ -6,6 +6,7 @@
enum cvm_pmu_type {
CVM_PMU_LMC,
+ CVM_PMU_TLK,
};
#ifdef CONFIG_CAVIUM_PMU
--
2.9.0.rc0.21.g7777322
On Fri, Jun 23, 2017 at 03:01:25PM +0200, Jan Glauber wrote:
> Add support for various PMU counters found on the Cavium ThunderX and
> OcteonTx SoC.
>
> The driver provides common "uncore" functions to avoid code duplication and
> support adding more device PMUs (like L2 cache) in the future.
>
> Probe and removal of the PMUs is done by hooking into the ThunderX EDAC
> driver as this driver owns all the PCI devices containing the PMU counters.
... which is a very bad idea, of course. EDAC doesn't have anything to
do with performance measurements.
And I don't see why the EDAC driver should own any PCI devices -
EDAC simply uses them like anything else in the kernel so the PMU
functionality can simply use them too, independent from EDAC.
--
Regards/Gruss,
Boris.
Good mailing practices for 400: avoid top-posting and trim the reply.