2:38 a.m.
Patchset adds performance stats reporting support for nvdimm.
Added interface includes support for a pmu register function and
callbacks to be used by arch/platform specific drivers.
User could use standard perf tool to access perf events exposed via pmu.
Patchset adds a structure called nvdimm_pmu which can
be used to add platform specific data like supported event list and
callbacks to pmu functions like event_init/add/delete/read.
Patchset includes an implements the to expose IBM pseries platform nmem*
device performance stats using this interface.
Result from power9 pseries lpar with 2 nvdimm device:
command:# perf list nmem
nmem0/cchrhcnt/ [Kernel PMU event]
nmem0/cchwhcnt/ [Kernel PMU event]
nmem0/critrscu/ [Kernel PMU event]
nmem0/ctlresct/ [Kernel PMU event]
nmem0/ctlrestm/ [Kernel PMU event]
nmem0/fastwcnt/ [Kernel PMU event]
nmem0/hostlcnt/ [Kernel PMU event]
nmem0/hostldur/ [Kernel PMU event]
nmem0/hostscnt/ [Kernel PMU event]
nmem0/hostsdur/ [Kernel PMU event]
nmem0/medrcnt/ [Kernel PMU event]
nmem0/medrdur/ [Kernel PMU event]
nmem0/medwcnt/ [Kernel PMU event]
nmem0/medwdur/ [Kernel PMU event]
nmem0/memlife/ [Kernel PMU event]
nmem0/noopstat/ [Kernel PMU event]
nmem0/ponsecs/ [Kernel PMU event]
nmem1/cchrhcnt/ [Kernel PMU event]
nmem1/cchwhcnt/ [Kernel PMU event]
nmem1/critrscu/ [Kernel PMU event]
...
nmem1/noopstat/ [Kernel PMU event]
nmem1/ponsecs/ [Kernel PMU event]
Patch1:
Introduces the nvdimm_pmu structure, common function for pmu
register along with callback routine check.
Pacth2
Add code in arch/powerpc/platform/pseries/papr_scm.c to expose
nmem* pmu. It fills in the nvdimm_pmu structure with event attrs
and event callback functions and then registers the pmu by adding
callback to register_nvdimm_pmu.
Patch3:
Sysfs documentation patch
Patch4:
Adds cpuhotplug support.
Kajol Jain (4):
drivers/nvdimm: Add perf interface to expose nvdimm performance stats
powerpc/papr_scm: Add perf interface support
powerpc/papr_scm: Document papr_scm sysfs event format entries
powerpc/papr_scm: Add cpu hotplug support for nvdimm pmu device
Documentation/ABI/testing/sysfs-bus-papr-pmem | 31 ++
arch/powerpc/include/asm/device.h | 5 +
arch/powerpc/platforms/pseries/papr_scm.c | 346 +++++++++++++++++-
drivers/nvdimm/Makefile | 1 +
drivers/nvdimm/nd_perf.c | 111 ++++++
include/linux/nd.h | 31 ++
6 files changed, 524 insertions(+), 1 deletion(-)
create mode 100644 drivers/nvdimm/nd_perf.c
--
2.27.0
2:38 a.m.
New subject: [RFC 1/4] drivers/nvdimm: Add perf interface to expose nvdimm performance stats
Patch adds performance stats reporting support for nvdimm.
Added interface includes support for a pmu register function and
callbacks to be used by the arch/platform specific drivers.
User could use the standard perf tool to access perf events exposed
via pmu.
A structure is added called nvdimm_pmu which can be used to add
platform specific data like supported events and callbacks to pmu
functions like event_init/add/delete/read. It also adds
unregister_nvdimm_pmu function to handle unregistering of a pmu device.
Signed-off-by: Kajol Jain <kjain(a)linux.ibm.com>
---
drivers/nvdimm/Makefile | 1 +
drivers/nvdimm/nd_perf.c | 111 +++++++++++++++++++++++++++++++++++++++
include/linux/nd.h | 31 +++++++++++
3 files changed, 143 insertions(+)
create mode 100644 drivers/nvdimm/nd_perf.c
diff --git a/drivers/nvdimm/Makefile b/drivers/nvdimm/Makefile
index 29203f3d3069..25dba6095612 100644
--- a/drivers/nvdimm/Makefile
+++ b/drivers/nvdimm/Makefile
@@ -18,6 +18,7 @@ nd_e820-y := e820.o
libnvdimm-y := core.o
libnvdimm-y += bus.o
libnvdimm-y += dimm_devs.o
+libnvdimm-y += nd_perf.o
libnvdimm-y += dimm.o
libnvdimm-y += region_devs.o
libnvdimm-y += region.o
diff --git a/drivers/nvdimm/nd_perf.c b/drivers/nvdimm/nd_perf.c
new file mode 100644
index 000000000000..d28bec2b61a2
--- /dev/null
+++ b/drivers/nvdimm/nd_perf.c
@@ -0,0 +1,111 @@
+// SPDX-License-Identifier: GPL-2.0-or-later
+/*
+ * nd_perf.c: NVDIMM Device Performance Monitoring Unit support
+ *
+ * Perf interface to expose nvdimm performance stats.
+ *
+ * Copyright (C) 2021 IBM Corporation
+ */
+
+#define pr_fmt(fmt) "nvdimm_pmu: " fmt
+
+#include <linux/nd.h>
+
+#define to_nvdimm_pmu(_pmu) container_of(_pmu, struct nvdimm_pmu, pmu)
+
+static int nvdimm_pmu_event_init(struct perf_event *event)
+{
+ struct nvdimm_pmu *nd_pmu = to_nvdimm_pmu(event->pmu);
+
+ /* test the event attr type for PMU enumeration */
+ if (event->attr.type != event->pmu->type)
+ return -ENOENT;
+
+ /* it does not support event sampling mode */
+ if (is_sampling_event(event))
+ return -EINVAL;
+
+ /* no branch sampling */
+ if (has_branch_stack(event))
+ return -EOPNOTSUPP;
+
+ /* jump to arch/platform specific callbacks if any */
+ if (nd_pmu && nd_pmu->event_init)
+ return nd_pmu->event_init(event, nd_pmu->dev);
+
+ return 0;
+}
+
+static void nvdimm_pmu_read(struct perf_event *event)
+{
+ struct nvdimm_pmu *nd_pmu = to_nvdimm_pmu(event->pmu);
+
+ /* jump to arch/platform specific callbacks if any */
+ if (nd_pmu && nd_pmu->read)
+ nd_pmu->read(event, nd_pmu->dev);
+}
+
+static void nvdimm_pmu_del(struct perf_event *event, int flags)
+{
+ struct nvdimm_pmu *nd_pmu = to_nvdimm_pmu(event->pmu);
+
+ /* jump to arch/platform specific callbacks if any */
+ if (nd_pmu && nd_pmu->del)
+ nd_pmu->del(event, flags, nd_pmu->dev);
+}
+
+static int nvdimm_pmu_add(struct perf_event *event, int flags)
+{
+ struct nvdimm_pmu *nd_pmu = to_nvdimm_pmu(event->pmu);
+
+ if (flags & PERF_EF_START)
+ /* jump to arch/platform specific callbacks if any */
+ if (nd_pmu && nd_pmu->add)
+ return nd_pmu->add(event, flags, nd_pmu->dev);
+ return 0;
+}
+
+int register_nvdimm_pmu(struct nvdimm_pmu *nd_pmu, struct platform_device *pdev)
+{
+ int rc;
+
+ if (!nd_pmu || !pdev)
+ return -EINVAL;
+
+ nd_pmu->pmu.task_ctx_nr = perf_invalid_context;
+ nd_pmu->pmu.event_init = nvdimm_pmu_event_init;
+ nd_pmu->pmu.add = nvdimm_pmu_add;
+ nd_pmu->pmu.del = nvdimm_pmu_del;
+ nd_pmu->pmu.read = nvdimm_pmu_read;
+ nd_pmu->pmu.name = nd_pmu->name;
+ nd_pmu->pmu.attr_groups = nd_pmu->attr_groups;
+ nd_pmu->pmu.capabilities = PERF_PMU_CAP_NO_INTERRUPT |
+ PERF_PMU_CAP_NO_EXCLUDE;
+
+ /*
+ * Adding platform_device->dev pointer to nvdimm_pmu, so that we can
+ * access that device data in PMU callbacks and also pass it to
+ * arch/platform specific code.
+ */
+ nd_pmu->dev = &pdev->dev;
+
+ rc = perf_pmu_register(&nd_pmu->pmu, nd_pmu->name, -1);
+ if (rc)
+ return rc;
+
+ pr_info("%s NVDIMM performance monitor support registered\n",
+ nd_pmu->name);
+
+ return 0;
+}
+EXPORT_SYMBOL_GPL(register_nvdimm_pmu);
+
+void unregister_nvdimm_pmu(struct pmu *nd_pmu)
+{
+ /*
+ * nd_pmu will get free in arch/platform specific code once
+ * corresponding pmu get unregistered.
+ */
+ perf_pmu_unregister(nd_pmu);
+}
+EXPORT_SYMBOL_GPL(unregister_nvdimm_pmu);
diff --git a/include/linux/nd.h b/include/linux/nd.h
index ee9ad76afbba..fa6e60b2b368 100644
--- a/include/linux/nd.h
+++ b/include/linux/nd.h
@@ -8,6 +8,8 @@
#include <linux/ndctl.h>
#include <linux/device.h>
#include <linux/badblocks.h>
+#include <linux/platform_device.h>
+#include <linux/perf_event.h>
enum nvdimm_event {
NVDIMM_REVALIDATE_POISON,
@@ -23,6 +25,35 @@ enum nvdimm_claim_class {
NVDIMM_CCLASS_UNKNOWN,
};
+/**
+ * struct nvdimm_pmu - data structure for nvdimm perf driver
+ *
+ * @name: name of the nvdimm pmu device.
+ * @pmu: pmu data structure for nvdimm performance stats.
+ * @cpu: designated cpu for counter access.
+ * @dev: nvdimm device pointer.
+ * @functions(event_init/add/del/read): platform specific callbacks.
+ * @attr_groups: data structure for events/formats/cpumask.
+ * @node: node for cpu hotplug notifier link.
+ * @cpuhp_state: state for cpu hotplug notification.
+ */
+struct nvdimm_pmu {
+ const char *name;
+ struct pmu pmu;
+ int cpu;
+ struct device *dev;
+ int (*event_init)(struct perf_event *event, struct device *dev);
+ int (*add)(struct perf_event *event, int flags, struct device *dev);
+ void (*del)(struct perf_event *event, int flags, struct device *dev);
+ void (*read)(struct perf_event *event, struct device *dev);
+ const struct attribute_group **attr_groups;
+ struct hlist_node node;
+ enum cpuhp_state cpuhp_state;
+};
+
+int register_nvdimm_pmu(struct nvdimm_pmu *nvdimm, struct platform_device *pdev);
+void unregister_nvdimm_pmu(struct pmu *pmu);
+
struct nd_device_driver {
struct device_driver drv;
unsigned long type;
--
2.27.0
2:38 a.m.
New subject: [RFC 2/4] powerpc/papr_scm: Add perf interface support
This patch adds support for performance monitoring of papr
nvdimm devices via perf interface. It adds callbacks functions
like add/del/read/event_init for nvdimm_pmu structure.
Patch adds a new parameter 'priv' in pdev_archdata structure to save
nvdimm_pmu device pointer, to handle the unregistering of pmu device.
papr_scm_pmu_register function populates the nvdimm_pmu structure
with events, attribute groups along with event handling functions.
Event handling functions internally uses hcall to get events and
counter data. Finally the populated nvdimm_pmu structure is passed
to register the pmu device.
Result in power9 machine with 2 nvdimm device:
Ex: List all event by perf list
command:# perf list nmem
nmem0/cchrhcnt/ [Kernel PMU event]
nmem0/cchwhcnt/ [Kernel PMU event]
nmem0/critrscu/ [Kernel PMU event]
nmem0/ctlresct/ [Kernel PMU event]
nmem0/ctlrestm/ [Kernel PMU event]
nmem0/fastwcnt/ [Kernel PMU event]
nmem0/hostlcnt/ [Kernel PMU event]
nmem0/hostldur/ [Kernel PMU event]
nmem0/hostscnt/ [Kernel PMU event]
nmem0/hostsdur/ [Kernel PMU event]
nmem0/medrcnt/ [Kernel PMU event]
nmem0/medrdur/ [Kernel PMU event]
nmem0/medwcnt/ [Kernel PMU event]
nmem0/medwdur/ [Kernel PMU event]
nmem0/memlife/ [Kernel PMU event]
nmem0/noopstat/ [Kernel PMU event]
nmem0/ponsecs/ [Kernel PMU event]
nmem1/cchrhcnt/ [Kernel PMU event]
nmem1/cchwhcnt/ [Kernel PMU event]
nmem1/critrscu/ [Kernel PMU event]
...
nmem1/noopstat/ [Kernel PMU event]
nmem1/ponsecs/ [Kernel PMU event]
Signed-off-by: Kajol Jain <kjain(a)linux.ibm.com>
---
arch/powerpc/include/asm/device.h | 5 +
arch/powerpc/platforms/pseries/papr_scm.c | 284 +++++++++++++++++++++-
2 files changed, 288 insertions(+), 1 deletion(-)
diff --git a/arch/powerpc/include/asm/device.h b/arch/powerpc/include/asm/device.h
index 219559d65864..47ed639f3b8f 100644
--- a/arch/powerpc/include/asm/device.h
+++ b/arch/powerpc/include/asm/device.h
@@ -48,6 +48,11 @@ struct dev_archdata {
struct pdev_archdata {
u64 dma_mask;
+ /*
+ * Pointer to nvdimm_pmu structure, to handle the unregistering
+ * of pmu device
+ */
+ void *priv;
};
#endif /* _ASM_POWERPC_DEVICE_H */
diff --git a/arch/powerpc/platforms/pseries/papr_scm.c
b/arch/powerpc/platforms/pseries/papr_scm.c
index ef26fe40efb0..997d379094d0 100644
--- a/arch/powerpc/platforms/pseries/papr_scm.c
+++ b/arch/powerpc/platforms/pseries/papr_scm.c
@@ -18,6 +18,8 @@
#include <asm/plpar_wrappers.h>
#include <asm/papr_pdsm.h>
#include <asm/mce.h>
+#include <linux/perf_event.h>
+#include <linux/ctype.h>
#define BIND_ANY_ADDR (~0ul)
@@ -116,6 +118,9 @@ struct papr_scm_priv {
/* length of the stat buffer as expected by phyp */
size_t stat_buffer_len;
+
+ /* array to have event_code and stat_id mappings */
+ char **nvdimm_events_map;
};
static int papr_scm_pmem_flush(struct nd_region *nd_region,
@@ -329,6 +334,271 @@ static ssize_t drc_pmem_query_stats(struct papr_scm_priv *p,
return 0;
}
+static struct attribute_group nvdimm_pmu_events_group = {
+ .name = "events",
+ /* .attrs is set in papr_scm_pmu_check_events function */
+};
+
+PMU_FORMAT_ATTR(event, "config:0-37");
+
+static struct attribute *nvdimm_pmu_format_attr[] = {
+ &format_attr_event.attr,
+ NULL,
+};
+
+static struct attribute_group nvdimm_pmu_format_group = {
+ .name = "format",
+ .attrs = nvdimm_pmu_format_attr,
+};
+
+static const struct attribute_group *nvdimm_pmu_attr_groups[] = {
+ &nvdimm_pmu_format_group,
+ &nvdimm_pmu_events_group,
+ NULL,
+};
+
+static void papr_scm_pmu_get_value(struct perf_event *event, struct device *dev, u64
*count)
+{
+ struct papr_scm_perf_stat *stat;
+ struct papr_scm_perf_stats *stats;
+ struct papr_scm_priv *p = (struct papr_scm_priv *)dev->driver_data;
+ int rc, size;
+ u64 statval;
+
+ /* Allocate buffer to hold single performance stat */
+ size = sizeof(struct papr_scm_perf_stats) +
+ sizeof(struct papr_scm_perf_stat);
+
+ if (!p->nvdimm_events_map)
+ return;
+
+ stats = kzalloc(size, GFP_KERNEL);
+ if (!stats)
+ return;
+
+ stat = &stats->scm_statistic[0];
+ memcpy(&stat->stat_id,
+ p->nvdimm_events_map[event->attr.config - 1],
+ sizeof(stat->stat_id));
+ stat->stat_val = 0;
+
+ rc = drc_pmem_query_stats(p, stats, 1);
+ if (rc < 0) {
+ kfree(stats);
+ return;
+ }
+
+ statval = be64_to_cpu(stat->stat_val);
+ *count = statval;
+ kfree(stats);
+}
+
+static int papr_scm_pmu_add(struct perf_event *event, int flags, struct device *dev)
+{
+ u64 count = 0;
+
+ papr_scm_pmu_get_value(event, dev, &count);
+ local64_set(&event->hw.prev_count, count);
+ return 0;
+}
+
+static void papr_scm_pmu_read(struct perf_event *event, struct device *dev)
+{
+ u64 prev, now = 0;
+
+ papr_scm_pmu_get_value(event, dev, &now);
+ prev = local64_xchg(&event->hw.prev_count, now);
+
+ if (now - prev >= 0)
+ local64_add(now - prev, &event->count);
+}
+
+static void papr_scm_pmu_del(struct perf_event *event, int flags, struct device *dev)
+{
+ papr_scm_pmu_read(event, dev);
+}
+
+static void nvdimm_pmu_uinit(struct nvdimm_pmu *nd_pmu)
+{
+ unregister_nvdimm_pmu(&nd_pmu->pmu);
+ kfree(nd_pmu);
+}
+
+static int papr_scm_pmu_register(struct papr_scm_priv *p)
+{
+ struct nvdimm_pmu *papr_scm_pmu;
+ int rc;
+
+ papr_scm_pmu = devm_kzalloc(&p->pdev->dev, sizeof(*papr_scm_pmu),
GFP_KERNEL);
+ if (!papr_scm_pmu)
+ return -ENOMEM;
+
+ papr_scm_pmu->name = nvdimm_name(p->nvdimm);
+ papr_scm_pmu->read = papr_scm_pmu_read;
+ papr_scm_pmu->add = papr_scm_pmu_add;
+ papr_scm_pmu->del = papr_scm_pmu_del;
+ papr_scm_pmu->attr_groups = nvdimm_pmu_attr_groups;
+
+ rc = register_nvdimm_pmu(papr_scm_pmu, p->pdev);
+ if (rc)
+ goto pmu_register_err;
+
+ /*
+ * Set archdata.priv value to nvdimm_pmu structure, to handle the
+ * unregistering of pmu device.
+ */
+ p->pdev->archdata.priv = papr_scm_pmu;
+ return 0;
+
+pmu_register_err:
+ kfree(papr_scm_pmu);
+ return rc;
+}
+
+static ssize_t device_show_string(struct device *dev, struct device_attribute *attr,
+ char *buf)
+{
+ struct perf_pmu_events_attr *d;
+
+ d = container_of(attr, struct perf_pmu_events_attr, attr);
+
+ return sysfs_emit(buf, "%s\n", (char *)d->event_str);
+}
+
+static char *strtolower(char *updated_name)
+{
+ int i = 0;
+
+ while (updated_name[i]) {
+ if (isupper(updated_name[i]))
+ updated_name[i] = tolower(updated_name[i]);
+ i++;
+ }
+ updated_name[i] = '\0';
+ return strim(updated_name);
+}
+
+/* device_str_attr_create : Populate event "name" and string "str" in
attribute */
+static struct attribute *device_str_attr_create_(char *name, char *str)
+{
+ struct perf_pmu_events_attr *attr;
+
+ attr = kzalloc(sizeof(*attr), GFP_KERNEL);
+
+ if (!attr)
+ return NULL;
+
+ sysfs_attr_init(&attr->attr.attr);
+ attr->event_str = str;
+ attr->attr.attr.name = strtolower(name);
+ attr->attr.attr.mode = 0444;
+ attr->attr.show = device_show_string;
+
+ return &attr->attr.attr;
+}
+
+static int papr_scm_pmu_check_events(struct papr_scm_priv *p)
+{
+ struct papr_scm_perf_stat *stat;
+ struct papr_scm_perf_stats *stats, *single_stats;
+ int index, size, rc, attrs;
+ u32 total_events;
+ struct attribute **events;
+ char *eventcode, *eventname, *statid;
+
+ if (!p->stat_buffer_len)
+ return -ENOENT;
+
+ total_events = (p->stat_buffer_len - sizeof(struct papr_scm_perf_stats))
+ / sizeof(struct papr_scm_perf_stat);
+
+ /* Allocate the buffer for phyp where stats are written */
+ stats = kzalloc(p->stat_buffer_len, GFP_KERNEL);
+ if (!stats)
+ return -ENOMEM;
+
+ /* Allocate memory to nvdimm_event_map */
+ p->nvdimm_events_map = kcalloc(total_events, sizeof(char *), GFP_KERNEL);
+ if (!p->nvdimm_events_map) {
+ rc = -ENOMEM;
+ goto out_stats;
+ }
+
+ /* Called to get list of events supported */
+ rc = drc_pmem_query_stats(p, stats, 0);
+ if (rc)
+ goto out_nvdimm_events_map;
+
+ /* Allocate buffer to hold single performance stat */
+ size = sizeof(struct papr_scm_perf_stats) + sizeof(struct papr_scm_perf_stat);
+
+ single_stats = kzalloc(size, GFP_KERNEL);
+ if (!single_stats) {
+ rc = -ENOMEM;
+ goto out_nvdimm_events_map;
+ }
+
+ events = kzalloc(total_events * sizeof(struct attribute *), GFP_KERNEL);
+ if (!events) {
+ rc = -ENOMEM;
+ goto out_single_stats;
+ }
+
+ for (index = 0, stat = stats->scm_statistic, attrs = 0;
+ index < total_events; index++, ++stat) {
+
+ single_stats->scm_statistic[0] = *stat;
+ rc = drc_pmem_query_stats(p, single_stats, 1);
+
+ if (rc < 0) {
+ pr_info("Event not supported %s for device %s\n",
+ stat->stat_id, nvdimm_name(p->nvdimm));
+ } else {
+ eventcode = kasprintf(GFP_KERNEL, "event=0x%x", attrs + 1);
+ eventname = kzalloc(strlen(stat->stat_id) + 1, GFP_KERNEL);
+ statid = kzalloc(strlen(stat->stat_id) + 1, GFP_KERNEL);
+
+ if (!eventname || !statid || !eventcode)
+ goto out;
+
+ strcpy(eventname, stat->stat_id);
+ events[attrs] = device_str_attr_create_(eventname,
+ eventcode);
+ if (!events[attrs])
+ goto out;
+
+ strcpy(statid, stat->stat_id);
+ p->nvdimm_events_map[attrs] = statid;
+ attrs++;
+ continue;
+out:
+ kfree(eventcode);
+ kfree(eventname);
+ kfree(statid);
+ }
+ }
+ events[attrs] = NULL;
+ p->nvdimm_events_map[attrs] = NULL;
+
+ if (!attrs)
+ goto out_events;
+
+ nvdimm_pmu_events_group.attrs = events;
+ kfree(single_stats);
+ kfree(stats);
+ return 0;
+
+out_events:
+ kfree(events);
+out_single_stats:
+ kfree(single_stats);
+out_nvdimm_events_map:
+ kfree(p->nvdimm_events_map);
+out_stats:
+ kfree(stats);
+ return rc;
+}
+
/*
* Issue hcall to retrieve dimm health info and populate papr_scm_priv with the
* health information.
@@ -923,7 +1193,7 @@ static int papr_scm_nvdimm_init(struct papr_scm_priv *p)
struct nd_mapping_desc mapping;
struct nd_region_desc ndr_desc;
unsigned long dimm_flags;
- int target_nid, online_nid;
+ int target_nid, online_nid, rc;
ssize_t stat_size;
p->bus_desc.ndctl = papr_scm_ndctl;
@@ -1015,6 +1285,15 @@ static int papr_scm_nvdimm_init(struct papr_scm_priv *p)
p->stat_buffer_len = stat_size;
dev_dbg(&p->pdev->dev, "Max perf-stat size %lu-bytes\n",
p->stat_buffer_len);
+
+ rc = papr_scm_pmu_check_events(p);
+ if (rc) {
+ dev_info(&p->pdev->dev, "nvdimm pmu check events failed, rc=%d\n",
rc);
+ } else {
+ rc = papr_scm_pmu_register(p);
+ if (rc)
+ dev_info(&p->pdev->dev, "nvdimm pmu didn't register rc=%d\n",
rc);
+ }
} else {
dev_info(&p->pdev->dev, "Dimm performance stats unavailable\n");
}
@@ -1195,7 +1474,10 @@ static int papr_scm_remove(struct platform_device *pdev)
nvdimm_bus_unregister(p->bus);
drc_pmem_unbind(p);
+ nvdimm_pmu_uinit(pdev->archdata.priv);
+ pdev->archdata.priv = NULL;
kfree(p->bus_desc.provider_name);
+ kfree(p->nvdimm_events_map);
kfree(p);
return 0;
--
2.27.0
2:38 a.m.
New subject: [RFC 3/4] powerpc/papr_scm: Document papr_scm sysfs event format entries
This patch add event format and events details in ABI
documentation
Signed-off-by: Kajol Jain <kjain(a)linux.ibm.com>
---
Documentation/ABI/testing/sysfs-bus-papr-pmem | 25 +++++++++++++++++++
1 file changed, 25 insertions(+)
diff --git a/Documentation/ABI/testing/sysfs-bus-papr-pmem
b/Documentation/ABI/testing/sysfs-bus-papr-pmem
index 8316c33862a0..216f70deca7e 100644
--- a/Documentation/ABI/testing/sysfs-bus-papr-pmem
+++ b/Documentation/ABI/testing/sysfs-bus-papr-pmem
@@ -59,3 +59,28 @@ Description:
* "CchRHCnt" : Cache Read Hit Count
* "CchWHCnt" : Cache Write Hit Count
* "FastWCnt" : Fast Write Count
+
+What: /sys/devices/nmemX/format
+Date: May 2021
+Contact: linuxppc-dev <linuxppc-dev(a)lists.ozlabs.org>, linux-nvdimm(a)lists.01.org,
+Description: (RO) Attribute group to describe the magic bits
+ that go into perf_event_attr.config for a particular pmu.
+ (See ABI/testing/sysfs-bus-event_source-devices-format).
+
+ Each attribute under this group defines a bit range of the
+ perf_event_attr.config. Supported attributes is listed
+ below::
+
+ event = "config:0-37" - event ID
+
+ For example::
+ noopstat = "event=0x1"
+
+What: /sys/devices/nmemX/events
+Date: May 2021
+Contact: linuxppc-dev <linuxppc-dev(a)lists.ozlabs.org>, linux-nvdimm(a)lists.01.org,
+Description: (RO) Attribute group to describe performance monitoring
+ events specific to papr_pmem. Each attribute in this group describes
+ a single performance monitoring event supported by this nvdimm pmu.
+ The name of the file is the name of the event.
+ (See ABI/testing/sysfs-bus-event_source-devices-events).
--
2.27.0
2:38 a.m.
New subject: [RFC 4/4] powerpc/papr_scm: Add cpu hotplug support for nvdimm pmu device
Patch here adds cpu hotplug functions to nvdimm pmu.
It adds cpumask to designate a cpu to make HCALL to
collect the counter data for the nvdimm device and
update ABI documentation accordingly.
Result in power9 lpar system:
command:# cat /sys/devices/nmem0/cpumask
0
Signed-off-by: Kajol Jain <kjain(a)linux.ibm.com>
---
Documentation/ABI/testing/sysfs-bus-papr-pmem | 6 ++
arch/powerpc/platforms/pseries/papr_scm.c | 62 +++++++++++++++++++
2 files changed, 68 insertions(+)
diff --git a/Documentation/ABI/testing/sysfs-bus-papr-pmem
b/Documentation/ABI/testing/sysfs-bus-papr-pmem
index 216f70deca7e..a40fbec683a8 100644
--- a/Documentation/ABI/testing/sysfs-bus-papr-pmem
+++ b/Documentation/ABI/testing/sysfs-bus-papr-pmem
@@ -76,6 +76,12 @@ Description: (RO) Attribute group to describe the magic bits
For example::
noopstat = "event=0x1"
+What: /sys/devices/nmemX/cpumask
+Date: May 2021
+Contact: linuxppc-dev <linuxppc-dev(a)lists.ozlabs.org>, linux-nvdimm(a)lists.01.org,
+Description: (RO) This sysfs file exposes the cpumask which is designated to make
+ HCALLs to retrieve nvdimm pmu event counter data.
+
What: /sys/devices/nmemX/events
Date: May 2021
Contact: linuxppc-dev <linuxppc-dev(a)lists.ozlabs.org>, linux-nvdimm(a)lists.01.org,
diff --git a/arch/powerpc/platforms/pseries/papr_scm.c
b/arch/powerpc/platforms/pseries/papr_scm.c
index 997d379094d0..6d94c2f260aa 100644
--- a/arch/powerpc/platforms/pseries/papr_scm.c
+++ b/arch/powerpc/platforms/pseries/papr_scm.c
@@ -334,6 +334,28 @@ static ssize_t drc_pmem_query_stats(struct papr_scm_priv *p,
return 0;
}
+static ssize_t cpumask_show(struct device *dev,
+ struct device_attribute *attr, char *buf)
+{
+ struct pmu *pmu = dev_get_drvdata(dev);
+ struct nvdimm_pmu *nd_pmu;
+
+ nd_pmu = container_of(pmu, struct nvdimm_pmu, pmu);
+
+ return cpumap_print_to_pagebuf(true, buf, cpumask_of(nd_pmu->cpu));
+}
+
+static DEVICE_ATTR_RO(cpumask);
+
+static struct attribute *nvdimm_cpumask_attrs[] = {
+ &dev_attr_cpumask.attr,
+ NULL,
+};
+
+static const struct attribute_group nvdimm_pmu_cpumask_group = {
+ .attrs = nvdimm_cpumask_attrs,
+};
+
static struct attribute_group nvdimm_pmu_events_group = {
.name = "events",
/* .attrs is set in papr_scm_pmu_check_events function */
@@ -354,6 +376,7 @@ static struct attribute_group nvdimm_pmu_format_group = {
static const struct attribute_group *nvdimm_pmu_attr_groups[] = {
&nvdimm_pmu_format_group,
&nvdimm_pmu_events_group,
+ &nvdimm_pmu_cpumask_group,
NULL,
};
@@ -418,10 +441,30 @@ static void papr_scm_pmu_del(struct perf_event *event, int flags,
struct device
papr_scm_pmu_read(event, dev);
}
+static int nvdimm_pmu_offline_cpu(unsigned int cpu, struct hlist_node *node)
+{
+ struct nvdimm_pmu *pmu;
+ int target;
+
+ pmu = hlist_entry_safe(node, struct nvdimm_pmu, node);
+
+ if (cpu != pmu->cpu)
+ return 0;
+
+ target = cpumask_last(cpu_active_mask);
+ if (target < 0 || target >= nr_cpu_ids)
+ return -1;
+
+ pmu->cpu = target;
+ return 0;
+}
+
static void nvdimm_pmu_uinit(struct nvdimm_pmu *nd_pmu)
{
unregister_nvdimm_pmu(&nd_pmu->pmu);
kfree(nd_pmu);
+ cpuhp_state_remove_instance_nocalls(nd_pmu->cpuhp_state, &nd_pmu->node);
+ cpuhp_remove_multi_state(nd_pmu->cpuhp_state);
}
static int papr_scm_pmu_register(struct papr_scm_priv *p)
@@ -438,6 +481,22 @@ static int papr_scm_pmu_register(struct papr_scm_priv *p)
papr_scm_pmu->add = papr_scm_pmu_add;
papr_scm_pmu->del = papr_scm_pmu_del;
papr_scm_pmu->attr_groups = nvdimm_pmu_attr_groups;
+ papr_scm_pmu->cpu = raw_smp_processor_id();
+
+ rc = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN,
+ "perf/nvdimm:online",
+ NULL, nvdimm_pmu_offline_cpu);
+ if (rc < 0) {
+ kfree(papr_scm_pmu);
+ return rc;
+ }
+
+ papr_scm_pmu->cpuhp_state = rc;
+
+ /* Register the pmu instance for cpu hotplug */
+ rc = cpuhp_state_add_instance_nocalls(papr_scm_pmu->cpuhp_state,
&papr_scm_pmu->node);
+ if (rc)
+ goto cpuhp_instance_err;
rc = register_nvdimm_pmu(papr_scm_pmu, p->pdev);
if (rc)
@@ -451,6 +510,9 @@ static int papr_scm_pmu_register(struct papr_scm_priv *p)
return 0;
pmu_register_err:
+ cpuhp_state_remove_instance_nocalls(papr_scm_pmu->cpuhp_state,
&papr_scm_pmu->node);
+cpuhp_instance_err:
+ cpuhp_remove_multi_state(papr_scm_pmu->cpuhp_state);
kfree(papr_scm_pmu);
return rc;
}
--
2.27.0
3:27 a.m.
New subject: [RFC 1/4] drivers/nvdimm: Add perf interface to expose nvdimm
performance stats
On Wed, May 12, 2021 at 10:08:21PM +0530, Kajol Jain wrote:
+static void nvdimm_pmu_read(struct perf_event *event)
+{
+ struct nvdimm_pmu *nd_pmu = to_nvdimm_pmu(event->pmu);
+
+ /* jump to arch/platform specific callbacks if any */
+ if (nd_pmu && nd_pmu->read)
+ nd_pmu->read(event, nd_pmu->dev);
+}
+
+static void nvdimm_pmu_del(struct perf_event *event, int flags)
+{
+ struct nvdimm_pmu *nd_pmu = to_nvdimm_pmu(event->pmu);
+
+ /* jump to arch/platform specific callbacks if any */
+ if (nd_pmu && nd_pmu->del)
+ nd_pmu->del(event, flags, nd_pmu->dev);
+}
+
+static int nvdimm_pmu_add(struct perf_event *event, int flags)
+{
+ struct nvdimm_pmu *nd_pmu = to_nvdimm_pmu(event->pmu);
+
+ if (flags & PERF_EF_START)
+ /* jump to arch/platform specific callbacks if any */
+ if (nd_pmu && nd_pmu->add)
+ return nd_pmu->add(event, flags, nd_pmu->dev);
+ return 0;
+}
What's the value add here? Why can't you directly set driver pointers? I
also don't really believe ->{add,del,read} can be optional and still
have a sane driver.
10:26 p.m.
New subject: [RFC 1/4] drivers/nvdimm: Add perf interface to expose nvdimm
performance stats
On 5/12/21 10:57 PM, Peter Zijlstra wrote:
On Wed, May 12, 2021 at 10:08:21PM +0530, Kajol Jain wrote:
> +static void nvdimm_pmu_read(struct perf_event *event)
> +{
> + struct nvdimm_pmu *nd_pmu = to_nvdimm_pmu(event->pmu);
> +
> + /* jump to arch/platform specific callbacks if any */
> + if (nd_pmu && nd_pmu->read)
> + nd_pmu->read(event, nd_pmu->dev);
> +}
> +
> +static void nvdimm_pmu_del(struct perf_event *event, int flags)
> +{
> + struct nvdimm_pmu *nd_pmu = to_nvdimm_pmu(event->pmu);
> +
> + /* jump to arch/platform specific callbacks if any */
> + if (nd_pmu && nd_pmu->del)
> + nd_pmu->del(event, flags, nd_pmu->dev);
> +}
> +
> +static int nvdimm_pmu_add(struct perf_event *event, int flags)
> +{
> + struct nvdimm_pmu *nd_pmu = to_nvdimm_pmu(event->pmu);
> +
> + if (flags & PERF_EF_START)
> + /* jump to arch/platform specific callbacks if any */
> + if (nd_pmu && nd_pmu->add)
> + return nd_pmu->add(event, flags, nd_pmu->dev);
> + return 0;
> +}
What's the value add here? Why can't you directly set driver pointers? I
also don't really believe ->{add,del,read} can be optional and still
have a sane driver.
Hi Peter,
The intend for adding these callbacks is to give flexibility to the
arch/platform specific driver code to use its own routine for getting
counter data or specific checks/operations. Arch/platform driver code
would have different method to get the counter data like IBM pseries
nmem* device which uses a hypervisor call(hcall).
But yes the current read/add/del functions are not adding value. We
could add an arch/platform specific function which could handle the
capturing of the counter data and do the rest of the operation here,
is this approach better?
Thanks,
Kajol Jain
9:47 p.m.
New subject: [RFC 1/4] drivers/nvdimm: Add perf interface to expose nvdimm
performance stats
On Thu, May 13, 2021 at 05:56:14PM +0530, kajoljain wrote:
But yes the current read/add/del functions are not adding value. We
could add an arch/platform specific function which could handle the
capturing of the counter data and do the rest of the operation here,
is this approach better?
Right; have your register_nvdimm_pmu() set pmu->{add,del,read} to
nd_pmu->{add,del,read} directly, don't bother with these intermediates.
Also you can WARN_ON_ONCE() if any of them are NULL and fail
registration at that point.
4:43 p.m.
New subject: [RFC 1/4] drivers/nvdimm: Add perf interface to expose nvdimm
performance stats
On 5/14/21 5:17 PM, Peter Zijlstra wrote:
On Thu, May 13, 2021 at 05:56:14PM +0530, kajoljain wrote:
> But yes the current read/add/del functions are not adding value. We
> could add an arch/platform specific function which could handle the
> capturing of the counter data and do the rest of the operation here,
> is this approach better?
Right; have your register_nvdimm_pmu() set pmu->{add,del,read} to
nd_pmu->{add,del,read} directly, don't bother with these intermediates.
Also you can WARN_ON_ONCE() if any of them are NULL and fail
registration at that point.
Hi Peter,
I will make all required changes and send next version of this patchset soon.
Thanks,
Kajol Jain
489
days inactive
494
days old
8 comments
3 participants
participants (3)
-
Kajol Jain -
kajoljain
-
Peter Zijlstra