This set follows what has been done for filters by adding an ioctl
to communicate PMU specific driver configuration to the kernel.
Patch 01 and 02 provide the functionality in user space while patch 04
and 05 take care of the kernel part. Patches 03 and 06 use the newly
added functionality to communicate the selection of CoreSight sink
for each event.
This functionality is needed to identify the CoreSight sink to use when
working with CPU-wide trace scenarios. It also paves the way to support
complex tracer configurations where dozens of trace parameters need to
be communicated to the CoreSight PMU driver.
Applies cleanly on v4.18-rc3.
Best regards,
Mathieu
Mathieu Poirier (6):
perf tools: Use ioctl to communicate driver configuration to kernel
perf tools: Make perf_evsel accessible to PMU driver configuration
code
perf tools: Use ioctl function to send sink configuration to kernel
perf/aux: Make perf_event accessible to setup_aux()
perf/core: Use ioctl to communicate driver configuration to kernel
coresight: Use PMU driver configuration for sink selection
arch/s390/kernel/perf_cpum_sf.c | 4 +-
arch/x86/events/intel/bts.c | 4 +-
arch/x86/events/intel/pt.c | 5 +-
drivers/hwtracing/coresight/coresight-etm-perf.c | 134 ++++++++++++++++++++---
drivers/hwtracing/coresight/coresight-etm-perf.h | 10 ++
drivers/perf/arm_spe_pmu.c | 6 +-
include/linux/perf_event.h | 56 +++++++++-
include/uapi/linux/perf_event.h | 1 +
kernel/events/core.c | 110 +++++++++++++++++++
kernel/events/ring_buffer.c | 2 +-
tools/include/uapi/linux/perf_event.h | 1 +
tools/perf/arch/arm/util/cs-etm.c | 60 +++-------
tools/perf/arch/arm/util/cs-etm.h | 3 +-
tools/perf/util/drv_configs.c | 30 ++---
tools/perf/util/evsel.c | 7 ++
tools/perf/util/evsel.h | 1 +
tools/perf/util/pmu.h | 3 +-
17 files changed, 345 insertions(+), 92 deletions(-)
--
2.7.4
This patch follows what has been done for filters by adding an ioctl()
option to communicate to the kernel arbitrary PMU specific configuration
that don't fit in the conventional struct perf_event_attr to the kernel.
Signed-off-by: Mathieu Poirier <[email protected]>
---
include/linux/perf_event.h | 54 ++++++++++++++++++++++
kernel/events/core.c | 110 +++++++++++++++++++++++++++++++++++++++++++++
2 files changed, 164 insertions(+)
diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
index 4d9c8f30ca6c..6e06b63c262f 100644
--- a/include/linux/perf_event.h
+++ b/include/linux/perf_event.h
@@ -178,6 +178,12 @@ struct hw_perf_event {
/* Last sync'ed generation of filters */
unsigned long addr_filters_gen;
+ /* PMU driver configuration */
+ void *drv_config;
+
+ /* Last sync'ed generation of driver config */
+ unsigned long drv_config_gen;
+
/*
* hw_perf_event::state flags; used to track the PERF_EF_* state.
*/
@@ -447,6 +453,26 @@ struct pmu {
* Filter events for PMU-specific reasons.
*/
int (*filter_match) (struct perf_event *event); /* optional */
+
+ /*
+ * Valiate complex PMU configuration that don't fit in the
+ * perf_event_attr struct. Returns a PMU specific pointer or an error
+ * value < 0.
+ *
+ * As with addr_filters_validate(), runs in the context of the ioctl()
+ * process and is not serialized with the rest of the PMU callbacks.
+ */
+ void *(*drv_config_validate) (struct perf_event *event,
+ char *config_str);
+
+ /* Synchronize PMU driver configuration */
+ void (*drv_config_sync) (struct perf_event *event);
+
+ /*
+ * Release PMU specific configuration acquired by
+ * drv_config_validate()
+ */
+ void (*drv_config_free) (void *drv_data);
};
enum perf_addr_filter_action_t {
@@ -489,6 +515,11 @@ struct perf_addr_filters_head {
unsigned int nr_file_filters;
};
+struct perf_drv_config {
+ void *drv_config;
+ raw_spinlock_t lock;
+};
+
/**
* enum perf_event_state - the states of a event
*/
@@ -668,6 +699,10 @@ struct perf_event {
unsigned long *addr_filters_offs;
unsigned long addr_filters_gen;
+ /* PMU driver specific configuration */
+ struct perf_drv_config drv_config;
+ unsigned long drv_config_gen;
+
void (*destroy)(struct perf_event *);
struct rcu_head rcu_head;
@@ -1234,6 +1269,13 @@ static inline bool has_addr_filter(struct perf_event *event)
return event->pmu->nr_addr_filters;
}
+static inline bool has_drv_config(struct perf_event *event)
+{
+ return event->pmu->drv_config_validate &&
+ event->pmu->drv_config_sync &&
+ event->pmu->drv_config_free;
+}
+
/*
* An inherited event uses parent's filters
*/
@@ -1248,7 +1290,19 @@ perf_event_addr_filters(struct perf_event *event)
return ifh;
}
+static inline struct perf_drv_config *
+perf_event_get_drv_config(struct perf_event *event)
+{
+ struct perf_drv_config *cfg = &event->drv_config;
+
+ if (event->parent)
+ cfg = &event->parent->drv_config;
+
+ return cfg;
+}
+
extern void perf_event_addr_filters_sync(struct perf_event *event);
+extern void perf_event_drv_config_sync(struct perf_event *event);
extern int perf_output_begin(struct perf_output_handle *handle,
struct perf_event *event, unsigned int size);
diff --git a/kernel/events/core.c b/kernel/events/core.c
index 8f0434a9951a..701839866789 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -2829,6 +2829,29 @@ void perf_event_addr_filters_sync(struct perf_event *event)
}
EXPORT_SYMBOL_GPL(perf_event_addr_filters_sync);
+/*
+ * PMU driver configuration works the same way as filter management above,
+ * but without the need to deal with memory mapping. Driver configuration
+ * arrives through the SET_DRV_CONFIG ioctl() where it is validated and applied
+ * to the event. When the PMU is ready it calls perf_event_drv_config_sync() to
+ * bring the configuration information within reach of the PMU.
+ */
+void perf_event_drv_config_sync(struct perf_event *event)
+{
+ struct perf_drv_config *drv_config = perf_event_get_drv_config(event);
+
+ if (!has_drv_config(event))
+ return;
+
+ raw_spin_lock(&drv_config->lock);
+ if (event->drv_config_gen != event->hw.drv_config_gen) {
+ event->pmu->drv_config_sync(event);
+ event->hw.drv_config_gen = event->drv_config_gen;
+ }
+ raw_spin_unlock(&drv_config->lock);
+}
+EXPORT_SYMBOL_GPL(perf_event_drv_config_sync);
+
static int _perf_event_refresh(struct perf_event *event, int refresh)
{
/*
@@ -4410,6 +4433,7 @@ static bool exclusive_event_installable(struct perf_event *event,
static void perf_addr_filters_splice(struct perf_event *event,
struct list_head *head);
+static void perf_drv_config_splice(struct perf_event *event, void *drv_data);
static void _free_event(struct perf_event *event)
{
@@ -4440,6 +4464,7 @@ static void _free_event(struct perf_event *event)
perf_event_free_bpf_prog(event);
perf_addr_filters_splice(event, NULL);
kfree(event->addr_filters_offs);
+ perf_drv_config_splice(event, NULL);
if (event->destroy)
event->destroy(event);
@@ -5002,6 +5027,8 @@ static inline int perf_fget_light(int fd, struct fd *p)
static int perf_event_set_output(struct perf_event *event,
struct perf_event *output_event);
static int perf_event_set_filter(struct perf_event *event, void __user *arg);
+static int perf_event_set_drv_config(struct perf_event *event,
+ void __user *arg);
static int perf_event_set_bpf_prog(struct perf_event *event, u32 prog_fd);
static int perf_copy_attr(struct perf_event_attr __user *uattr,
struct perf_event_attr *attr);
@@ -5088,6 +5115,10 @@ static long _perf_ioctl(struct perf_event *event, unsigned int cmd, unsigned lon
return perf_event_modify_attr(event, &new_attr);
}
+
+ case PERF_EVENT_IOC_SET_DRV_CONFIG:
+ return perf_event_set_drv_config(event, (void __user *)arg);
+
default:
return -ENOTTY;
}
@@ -9086,6 +9117,85 @@ static int perf_event_set_filter(struct perf_event *event, void __user *arg)
return ret;
}
+static void perf_drv_config_splice(struct perf_event *event, void *drv_data)
+{
+ unsigned long flags;
+ void *old_drv_data;
+
+ if (!has_drv_config(event))
+ return;
+
+ /* Children take their configuration from their parent */
+ if (event->parent)
+ return;
+
+ raw_spin_lock_irqsave(&event->drv_config.lock, flags);
+
+ old_drv_data = event->drv_config.drv_config;
+ event->drv_config.drv_config = drv_data;
+
+ raw_spin_unlock_irqrestore(&event->drv_config.lock, flags);
+
+ event->pmu->drv_config_free(old_drv_data);
+}
+
+static void perf_event_drv_config_apply(struct perf_event *event)
+{
+ unsigned long flags;
+ struct perf_drv_config *drv_config = perf_event_get_drv_config(event);
+
+ /* Notify event that a new configuration is available */
+ raw_spin_lock_irqsave(&drv_config->lock, flags);
+ event->drv_config_gen++;
+ raw_spin_unlock_irqrestore(&drv_config->lock, flags);
+}
+
+static int
+perf_event_process_drv_config(struct perf_event *event, char *config_str)
+{
+ int ret = -EINVAL;
+ void *drv_data;
+
+ /* Make sure ctx.mutex it held */
+ lockdep_assert_held(&event->ctx->mutex);
+
+ /* Children take their configuration from their parent */
+ if (WARN_ON_ONCE(event->parent))
+ goto out;
+
+ drv_data = event->pmu->drv_config_validate(event, config_str);
+ if (!drv_data)
+ goto out;
+ if (IS_ERR(drv_data)) {
+ ret = PTR_ERR(drv_data);
+ goto out;
+ }
+
+ perf_drv_config_splice(event, drv_data);
+
+ perf_event_for_each_child(event, perf_event_drv_config_apply);
+
+ ret = 0;
+out:
+ return ret;
+}
+
+static int perf_event_set_drv_config(struct perf_event *event, void __user *arg)
+{
+ int ret = -EINVAL;
+ char *config_str;
+
+ config_str = strndup_user(arg, PAGE_SIZE);
+ if (IS_ERR(config_str))
+ return PTR_ERR(config_str);
+
+ if (has_drv_config(event))
+ ret = perf_event_process_drv_config(event, config_str);
+
+ kfree(config_str);
+ return ret;
+}
+
/*
* hrtimer based swevent callback
*/
--
2.7.4
Make structure perf_evsel available to the PMU driver configuration code.
That way function perf_evsel__apply_drv_config() can be used from within
that code and information pertaining to the 'perf_evsel_config_term' is
still available.
Signed-off-by: Mathieu Poirier <[email protected]>
---
tools/perf/arch/arm/util/cs-etm.c | 22 +++++++++++++++++++++-
tools/perf/arch/arm/util/cs-etm.h | 3 ++-
tools/perf/util/drv_configs.c | 30 +++++++-----------------------
tools/perf/util/pmu.h | 3 ++-
4 files changed, 32 insertions(+), 26 deletions(-)
diff --git a/tools/perf/arch/arm/util/cs-etm.c b/tools/perf/arch/arm/util/cs-etm.c
index 2f595cd73da6..d8081c2e6d44 100644
--- a/tools/perf/arch/arm/util/cs-etm.c
+++ b/tools/perf/arch/arm/util/cs-etm.c
@@ -635,7 +635,7 @@ static int __printf(2, 3) cs_device__print_file(const char *name, const char *fm
return ret;
}
-int cs_etm_set_drv_config(struct perf_evsel_config_term *term)
+static int cs_etm_set_drv_config_term(struct perf_evsel_config_term *term)
{
int ret;
char enable_sink[ENABLE_SINK_MAX];
@@ -649,3 +649,23 @@ int cs_etm_set_drv_config(struct perf_evsel_config_term *term)
return 0;
}
+
+int cs_etm_set_drv_config(struct perf_evsel *evsel,
+ struct perf_evsel_config_term **err_term)
+{
+ int err = 0;
+ struct perf_evsel_config_term *term;
+
+ list_for_each_entry(term, &evsel->config_terms, list) {
+ if (term->type != PERF_EVSEL__CONFIG_TERM_DRV_CFG)
+ continue;
+
+ err = cs_etm_set_drv_config_term(term);
+ if (err) {
+ *err_term = term;
+ break;
+ }
+ }
+
+ return err;
+}
diff --git a/tools/perf/arch/arm/util/cs-etm.h b/tools/perf/arch/arm/util/cs-etm.h
index 1a12e64f5127..a3f8dde6ccef 100644
--- a/tools/perf/arch/arm/util/cs-etm.h
+++ b/tools/perf/arch/arm/util/cs-etm.h
@@ -10,6 +10,7 @@
#include "../../util/evsel.h"
struct auxtrace_record *cs_etm_record_init(int *err);
-int cs_etm_set_drv_config(struct perf_evsel_config_term *term);
+int cs_etm_set_drv_config(struct perf_evsel *evsel,
+ struct perf_evsel_config_term **err_term);
#endif
diff --git a/tools/perf/util/drv_configs.c b/tools/perf/util/drv_configs.c
index eec754243f4d..f7c1bcf08549 100644
--- a/tools/perf/util/drv_configs.c
+++ b/tools/perf/util/drv_configs.c
@@ -25,7 +25,6 @@ perf_evsel__apply_drv_configs(struct perf_evsel *evsel,
{
bool found = false;
int err = 0;
- struct perf_evsel_config_term *term;
struct perf_pmu *pmu = NULL;
while ((pmu = perf_pmu__scan(pmu)) != NULL)
@@ -34,29 +33,14 @@ perf_evsel__apply_drv_configs(struct perf_evsel *evsel,
break;
}
- list_for_each_entry(term, &evsel->config_terms, list) {
- if (term->type != PERF_EVSEL__CONFIG_TERM_DRV_CFG)
- continue;
+ /*
+ * No need to continue if we didn't get a match or if there is no
+ * driver configuration function for this PMU.
+ */
+ if (!found || !pmu->set_drv_config)
+ return err;
- /*
- * We have a configuration term, report an error if we
- * can't find the PMU or if the PMU driver doesn't support
- * cmd line driver configuration.
- */
- if (!found || !pmu->set_drv_config) {
- err = -EINVAL;
- *err_term = term;
- break;
- }
-
- err = pmu->set_drv_config(term);
- if (err) {
- *err_term = term;
- break;
- }
- }
-
- return err;
+ return pmu->set_drv_config(evsel, err_term);
}
int perf_evlist__apply_drv_configs(struct perf_evlist *evlist,
diff --git a/tools/perf/util/pmu.h b/tools/perf/util/pmu.h
index 76fecec7b3f9..47f44394042b 100644
--- a/tools/perf/util/pmu.h
+++ b/tools/perf/util/pmu.h
@@ -29,7 +29,8 @@ struct perf_pmu {
struct list_head format; /* HEAD struct perf_pmu_format -> list */
struct list_head aliases; /* HEAD struct perf_pmu_alias -> list */
struct list_head list; /* ELEM */
- int (*set_drv_config) (struct perf_evsel_config_term *term);
+ int (*set_drv_config) (struct perf_evsel *evsel,
+ struct perf_evsel_config_term **err_term);
};
struct perf_pmu_info {
--
2.7.4
It can be advantagous to have access to all the information conveyed by
a perf_event when setting up the AUX buffer, as it is the case when
dealing with PMU specific driver configuration communicated to the kernel
using an ioctl() call.
As such simply replace the cpu information by the complete perf_event
structure and change all affected customers.
Signed-off-by: Mathieu Poirier <[email protected]>
---
arch/s390/kernel/perf_cpum_sf.c | 4 ++--
arch/x86/events/intel/bts.c | 4 +++-
arch/x86/events/intel/pt.c | 5 +++--
drivers/hwtracing/coresight/coresight-etm-perf.c | 6 +++---
drivers/perf/arm_spe_pmu.c | 6 +++---
include/linux/perf_event.h | 2 +-
kernel/events/ring_buffer.c | 2 +-
7 files changed, 16 insertions(+), 13 deletions(-)
diff --git a/arch/s390/kernel/perf_cpum_sf.c b/arch/s390/kernel/perf_cpum_sf.c
index 0292d68e7dde..e06daaa08894 100644
--- a/arch/s390/kernel/perf_cpum_sf.c
+++ b/arch/s390/kernel/perf_cpum_sf.c
@@ -1601,8 +1601,8 @@ static void aux_buffer_free(void *data)
*
* Return the private AUX buffer structure if success or NULL if fails.
*/
-static void *aux_buffer_setup(int cpu, void **pages, int nr_pages,
- bool snapshot)
+static void *aux_buffer_setup(struct perf_event *event, void **pages,
+ int nr_pages, bool snapshot);
{
struct sf_buffer *sfb;
struct aux_buffer *aux;
diff --git a/arch/x86/events/intel/bts.c b/arch/x86/events/intel/bts.c
index 24ffa1e88cf9..7139f6bf27ad 100644
--- a/arch/x86/events/intel/bts.c
+++ b/arch/x86/events/intel/bts.c
@@ -77,10 +77,12 @@ static size_t buf_size(struct page *page)
}
static void *
-bts_buffer_setup_aux(int cpu, void **pages, int nr_pages, bool overwrite)
+bts_buffer_setup_aux(struct perf_event *event, void **pages,
+ int nr_pages, bool overwrite)
{
struct bts_buffer *buf;
struct page *page;
+ int cpu = event->cpu;
int node = (cpu == -1) ? cpu : cpu_to_node(cpu);
unsigned long offset;
size_t size = nr_pages << PAGE_SHIFT;
diff --git a/arch/x86/events/intel/pt.c b/arch/x86/events/intel/pt.c
index 8d016ce5b80d..8f4c98fdd03c 100644
--- a/arch/x86/events/intel/pt.c
+++ b/arch/x86/events/intel/pt.c
@@ -1104,10 +1104,11 @@ static int pt_buffer_init_topa(struct pt_buffer *buf, unsigned long nr_pages,
* Return: Our private PT buffer structure.
*/
static void *
-pt_buffer_setup_aux(int cpu, void **pages, int nr_pages, bool snapshot)
+pt_buffer_setup_aux(struct perf_event *event, void **pages,
+ int nr_pages, bool snapshot)
{
struct pt_buffer *buf;
- int node, ret;
+ int node, ret, cpu = event->cpu;
if (!nr_pages)
return NULL;
diff --git a/drivers/hwtracing/coresight/coresight-etm-perf.c b/drivers/hwtracing/coresight/coresight-etm-perf.c
index 677695635211..0f5e03e4df22 100644
--- a/drivers/hwtracing/coresight/coresight-etm-perf.c
+++ b/drivers/hwtracing/coresight/coresight-etm-perf.c
@@ -181,15 +181,15 @@ static void etm_free_aux(void *data)
schedule_work(&event_data->work);
}
-static void *etm_setup_aux(int event_cpu, void **pages,
+static void *etm_setup_aux(struct perf_event *event, void **pages,
int nr_pages, bool overwrite)
{
- int cpu;
+ int cpu = event->cpu;
cpumask_t *mask;
struct coresight_device *sink;
struct etm_event_data *event_data = NULL;
- event_data = alloc_event_data(event_cpu);
+ event_data = alloc_event_data(cpu);
if (!event_data)
return NULL;
INIT_WORK(&event_data->work, free_event_data);
diff --git a/drivers/perf/arm_spe_pmu.c b/drivers/perf/arm_spe_pmu.c
index 54ec278d2fc4..4dcd7bf14dcc 100644
--- a/drivers/perf/arm_spe_pmu.c
+++ b/drivers/perf/arm_spe_pmu.c
@@ -824,10 +824,10 @@ static void arm_spe_pmu_read(struct perf_event *event)
{
}
-static void *arm_spe_pmu_setup_aux(int cpu, void **pages, int nr_pages,
- bool snapshot)
+static void *arm_spe_pmu_setup_aux(struct perf_event *event, void **pages,
+ int nr_pages, bool snapshot)
{
- int i;
+ int i, cpu = event->cpu;
struct page **pglist;
struct arm_spe_pmu_buf *buf;
diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
index 1fa12887ec02..4d9c8f30ca6c 100644
--- a/include/linux/perf_event.h
+++ b/include/linux/perf_event.h
@@ -409,7 +409,7 @@ struct pmu {
/*
* Set up pmu-private data structures for an AUX area
*/
- void *(*setup_aux) (int cpu, void **pages,
+ void *(*setup_aux) (struct perf_event *event, void **pages,
int nr_pages, bool overwrite);
/* optional */
diff --git a/kernel/events/ring_buffer.c b/kernel/events/ring_buffer.c
index 5d3cf407e374..4c96c7575224 100644
--- a/kernel/events/ring_buffer.c
+++ b/kernel/events/ring_buffer.c
@@ -648,7 +648,7 @@ int rb_alloc_aux(struct ring_buffer *rb, struct perf_event *event,
goto out;
}
- rb->aux_priv = event->pmu->setup_aux(event->cpu, rb->aux_pages, nr_pages,
+ rb->aux_priv = event->pmu->setup_aux(event, rb->aux_pages, nr_pages,
overwrite);
if (!rb->aux_priv)
goto out;
--
2.7.4
Following in the footsteps of what was done for filters, adding the
necessary mechanic needed to push down driver specific configuration
to the kernel using an ioctl. By proceeding this way PMU specific
configuration can be communicated to each event.
Signed-off-by: Mathieu Poirier <[email protected]>
---
include/uapi/linux/perf_event.h | 1 +
tools/include/uapi/linux/perf_event.h | 1 +
tools/perf/util/evsel.c | 7 +++++++
tools/perf/util/evsel.h | 1 +
4 files changed, 10 insertions(+)
diff --git a/include/uapi/linux/perf_event.h b/include/uapi/linux/perf_event.h
index b8e288a1f740..b5b3241877df 100644
--- a/include/uapi/linux/perf_event.h
+++ b/include/uapi/linux/perf_event.h
@@ -460,6 +460,7 @@ struct perf_event_query_bpf {
#define PERF_EVENT_IOC_PAUSE_OUTPUT _IOW('$', 9, __u32)
#define PERF_EVENT_IOC_QUERY_BPF _IOWR('$', 10, struct perf_event_query_bpf *)
#define PERF_EVENT_IOC_MODIFY_ATTRIBUTES _IOW('$', 11, struct perf_event_attr *)
+#define PERF_EVENT_IOC_SET_DRV_CONFIG _IOW('$', 12, char *)
enum perf_event_ioc_flags {
PERF_IOC_FLAG_GROUP = 1U << 0,
diff --git a/tools/include/uapi/linux/perf_event.h b/tools/include/uapi/linux/perf_event.h
index b8e288a1f740..b5b3241877df 100644
--- a/tools/include/uapi/linux/perf_event.h
+++ b/tools/include/uapi/linux/perf_event.h
@@ -460,6 +460,7 @@ struct perf_event_query_bpf {
#define PERF_EVENT_IOC_PAUSE_OUTPUT _IOW('$', 9, __u32)
#define PERF_EVENT_IOC_QUERY_BPF _IOWR('$', 10, struct perf_event_query_bpf *)
#define PERF_EVENT_IOC_MODIFY_ATTRIBUTES _IOW('$', 11, struct perf_event_attr *)
+#define PERF_EVENT_IOC_SET_DRV_CONFIG _IOW('$', 12, char *)
enum perf_event_ioc_flags {
PERF_IOC_FLAG_GROUP = 1U << 0,
diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
index 94fce4f537e9..534aca4c642c 100644
--- a/tools/perf/util/evsel.c
+++ b/tools/perf/util/evsel.c
@@ -1162,6 +1162,13 @@ static int perf_evsel__append_filter(struct perf_evsel *evsel,
return -1;
}
+int perf_evsel__apply_drv_config(struct perf_evsel *evsel, const char *config)
+{
+ return perf_evsel__run_ioctl(evsel,
+ PERF_EVENT_IOC_SET_DRV_CONFIG,
+ (void *)config);
+}
+
int perf_evsel__append_tp_filter(struct perf_evsel *evsel, const char *filter)
{
return perf_evsel__append_filter(evsel, "(%s) && (%s)", filter);
diff --git a/tools/perf/util/evsel.h b/tools/perf/util/evsel.h
index d277930b19a1..0f671bd2a988 100644
--- a/tools/perf/util/evsel.h
+++ b/tools/perf/util/evsel.h
@@ -271,6 +271,7 @@ int perf_evsel__append_tp_filter(struct perf_evsel *evsel, const char *filter);
int perf_evsel__append_addr_filter(struct perf_evsel *evsel,
const char *filter);
int perf_evsel__apply_filter(struct perf_evsel *evsel, const char *filter);
+int perf_evsel__apply_drv_config(struct perf_evsel *evsel, const char *config);
int perf_evsel__enable(struct perf_evsel *evsel);
int perf_evsel__disable(struct perf_evsel *evsel);
--
2.7.4
Using sysFS to communicate sink information for a trace session doesn't
work when more than one CPU is involved in the scenario. As such
communicate the sink information to each event by using the SET_DRV_CONFIG
ioctl command.
Signed-off-by: Mathieu Poirier <[email protected]>
---
tools/perf/arch/arm/util/cs-etm.c | 54 ++++-----------------------------------
1 file changed, 5 insertions(+), 49 deletions(-)
diff --git a/tools/perf/arch/arm/util/cs-etm.c b/tools/perf/arch/arm/util/cs-etm.c
index d8081c2e6d44..cb0978eb7181 100644
--- a/tools/perf/arch/arm/util/cs-etm.c
+++ b/tools/perf/arch/arm/util/cs-etm.c
@@ -599,55 +599,10 @@ struct auxtrace_record *cs_etm_record_init(int *err)
return NULL;
}
-static FILE *cs_device__open_file(const char *name)
+static int cs_etm_set_drv_config_term(struct perf_evsel *evsel,
+ const char *term)
{
- struct stat st;
- char path[PATH_MAX];
- const char *sysfs;
-
- sysfs = sysfs__mountpoint();
- if (!sysfs)
- return NULL;
-
- snprintf(path, PATH_MAX,
- "%s" CS_BUS_DEVICE_PATH "%s", sysfs, name);
-
- if (stat(path, &st) < 0)
- return NULL;
-
- return fopen(path, "w");
-
-}
-
-static int __printf(2, 3) cs_device__print_file(const char *name, const char *fmt, ...)
-{
- va_list args;
- FILE *file;
- int ret = -EINVAL;
-
- va_start(args, fmt);
- file = cs_device__open_file(name);
- if (file) {
- ret = vfprintf(file, fmt, args);
- fclose(file);
- }
- va_end(args);
- return ret;
-}
-
-static int cs_etm_set_drv_config_term(struct perf_evsel_config_term *term)
-{
- int ret;
- char enable_sink[ENABLE_SINK_MAX];
-
- snprintf(enable_sink, ENABLE_SINK_MAX, "%s/%s",
- term->val.drv_cfg, "enable_sink");
-
- ret = cs_device__print_file(enable_sink, "%d", 1);
- if (ret < 0)
- return ret;
-
- return 0;
+ return perf_evsel__apply_drv_config(evsel, term);
}
int cs_etm_set_drv_config(struct perf_evsel *evsel,
@@ -660,7 +615,8 @@ int cs_etm_set_drv_config(struct perf_evsel *evsel,
if (term->type != PERF_EVSEL__CONFIG_TERM_DRV_CFG)
continue;
- err = cs_etm_set_drv_config_term(term);
+ err = cs_etm_set_drv_config_term(evsel,
+ term->val.drv_cfg);
if (err) {
*err_term = term;
break;
--
2.7.4
This patch uses the PMU driver configuration API to select a sink
for each event that is created (the old sysFS way of working is kept
around for backward compatibility).
By proceeding in this way a sink can be used by multiple sessions
without having to play games with entries in sysFS.
Signed-off-by: Mathieu Poirier <[email protected]>
---
drivers/hwtracing/coresight/coresight-etm-perf.c | 128 ++++++++++++++++++++---
drivers/hwtracing/coresight/coresight-etm-perf.h | 10 ++
2 files changed, 126 insertions(+), 12 deletions(-)
diff --git a/drivers/hwtracing/coresight/coresight-etm-perf.c b/drivers/hwtracing/coresight/coresight-etm-perf.c
index 0f5e03e4df22..37c8b97d0449 100644
--- a/drivers/hwtracing/coresight/coresight-etm-perf.c
+++ b/drivers/hwtracing/coresight/coresight-etm-perf.c
@@ -11,6 +11,7 @@
#include <linux/list.h>
#include <linux/mm.h>
#include <linux/init.h>
+#include <linux/parser.h>
#include <linux/perf_event.h>
#include <linux/slab.h>
#include <linux/types.h>
@@ -22,6 +23,8 @@
static struct pmu etm_pmu;
static bool etm_perf_up;
+#define CORESIGHT_DEVICE_MAX_NAME_LEN 256
+
/**
* struct etm_event_data - Coresight specifics associated to an event
* @work: Handle to free allocated memory outside IRQ context.
@@ -81,10 +84,29 @@ static int etm_addr_filters_alloc(struct perf_event *event)
return 0;
}
+static int etm_drv_config_alloc(struct perf_event *event)
+{
+ char *sink;
+ int node = event->cpu == -1 ? -1 : cpu_to_node(event->cpu);
+
+ sink = kzalloc_node(CORESIGHT_DEVICE_MAX_NAME_LEN, GFP_KERNEL, node);
+ if (!sink)
+ return -ENOMEM;
+
+ if (event->parent)
+ memcpy(sink, event->parent->hw.addr_filters,
+ CORESIGHT_DEVICE_MAX_NAME_LEN);
+
+ event->hw.drv_config = sink;
+
+ return 0;
+}
+
static void etm_event_destroy(struct perf_event *event)
{
kfree(event->hw.addr_filters);
event->hw.addr_filters = NULL;
+ kfree(event->hw.drv_config);
}
static int etm_event_init(struct perf_event *event)
@@ -100,6 +122,10 @@ static int etm_event_init(struct perf_event *event)
if (ret)
goto out;
+ ret = etm_drv_config_alloc(event);
+ if (ret)
+ goto out;
+
event->destroy = etm_event_destroy;
out:
return ret;
@@ -181,6 +207,38 @@ static void etm_free_aux(void *data)
schedule_work(&event_data->work);
}
+static struct coresight_device *etm_event_get_sink(struct perf_event *event)
+{
+ /*
+ * Try the preferred method first, i.e passing the sink information
+ * using the ioctl() method.
+ */
+ if (event->hw.drv_config) {
+ struct device *dev;
+ struct coresight_device *sink;
+
+ dev = bus_find_device_by_name(&coresight_bustype, NULL,
+ (char *)event->hw.drv_config);
+ if (dev) {
+ sink = to_coresight_device(dev);
+ /* Put reference from 'bus_find_device()' */
+ put_device(dev);
+ return sink;
+ }
+ }
+
+ /*
+ * No luck with the above method, so we are working with an older user
+ * space. See if a sink has been set using sysFS. If this is the case
+ * CPU-wide session will only be able to use a single sink.
+ *
+ * When operated from sysFS users are responsible to enable the sink
+ * while from perf, the perf tools will do it based on the choice made
+ * on the cmd line. As such the "enable_sink" flag in sysFS is reset.
+ */
+ return coresight_get_enabled_sink(true);
+}
+
static void *etm_setup_aux(struct perf_event *event, void **pages,
int nr_pages, bool overwrite)
{
@@ -194,18 +252,10 @@ static void *etm_setup_aux(struct perf_event *event, void **pages,
return NULL;
INIT_WORK(&event_data->work, free_event_data);
- /*
- * In theory nothing prevent tracers in a trace session from being
- * associated with different sinks, nor having a sink per tracer. But
- * until we have HW with this kind of topology we need to assume tracers
- * in a trace session are using the same sink. Therefore go through
- * the coresight bus and pick the first enabled sink.
- *
- * When operated from sysFS users are responsible to enable the sink
- * while from perf, the perf tools will do it based on the choice made
- * on the cmd line. As such the "enable_sink" flag in sysFS is reset.
- */
- sink = coresight_get_enabled_sink(true);
+ perf_event_drv_config_sync(event);
+
+ /* First get the sink to use for this event */
+ sink = etm_event_get_sink(event);
if (!sink)
goto err;
@@ -442,6 +492,57 @@ static void etm_addr_filters_sync(struct perf_event *event)
filters->nr_filters = i;
}
+static const match_table_t config_tokens = {
+ { ETM_CFG_SINK, "%u.%s" },
+ { ETM_CFG_NONE, NULL },
+};
+
+static void *etm_drv_config_validate(struct perf_event *event, char *cstr)
+{
+ char *str, *to_parse, *sink = NULL;
+ int token, ret = -EINVAL;
+ substring_t args[MAX_OPT_ARGS];
+
+ to_parse = kstrdup(cstr, GFP_KERNEL);
+ if (!to_parse)
+ return ERR_PTR(-ENOMEM);
+
+ while ((str = strsep(&to_parse, " ,\n")) != NULL) {
+ if (!*str)
+ continue;
+
+ token = match_token(str, config_tokens, args);
+ switch (token) {
+ case ETM_CFG_SINK:
+ sink = kstrdup(str, GFP_KERNEL);
+ if (!sink) {
+ ret = -ENOMEM;
+ goto out;
+ }
+ break;
+ default:
+ goto out;
+ }
+ }
+
+out:
+ kfree(to_parse);
+ return sink ? sink : ERR_PTR(ret);
+}
+
+static void etm_drv_config_sync(struct perf_event *event)
+{
+ struct perf_drv_config *drv_config = perf_event_get_drv_config(event);
+
+ memcpy(event->hw.drv_config, drv_config->drv_config,
+ CORESIGHT_DEVICE_MAX_NAME_LEN);
+}
+
+static void etm_drv_config_free(void *drv_data)
+{
+ kfree(drv_data);
+}
+
int etm_perf_symlink(struct coresight_device *csdev, bool link)
{
char entry[sizeof("cpu9999999")];
@@ -486,6 +587,9 @@ static int __init etm_perf_init(void)
etm_pmu.addr_filters_sync = etm_addr_filters_sync;
etm_pmu.addr_filters_validate = etm_addr_filters_validate;
etm_pmu.nr_addr_filters = ETM_ADDR_CMP_MAX;
+ etm_pmu.drv_config_validate = etm_drv_config_validate;
+ etm_pmu.drv_config_sync = etm_drv_config_sync;
+ etm_pmu.drv_config_free = etm_drv_config_free;
ret = perf_pmu_register(&etm_pmu, CORESIGHT_ETM_PMU_NAME, -1);
if (ret == 0)
diff --git a/drivers/hwtracing/coresight/coresight-etm-perf.h b/drivers/hwtracing/coresight/coresight-etm-perf.h
index 4197df4faf5e..175895c8d635 100644
--- a/drivers/hwtracing/coresight/coresight-etm-perf.h
+++ b/drivers/hwtracing/coresight/coresight-etm-perf.h
@@ -42,6 +42,16 @@ struct etm_filters {
bool ssstatus;
};
+enum etm_config_elem_type {
+ ETM_CFG_NONE = -1,
+ ETM_CFG_SINK,
+};
+
+struct etm_config_elem {
+ struct list_head entry;
+ enum etm_config_elem_type type;
+ void *config;
+};
#ifdef CONFIG_CORESIGHT
int etm_perf_symlink(struct coresight_device *csdev, bool link);
--
2.7.4
Hi Mathieu,
I love your patch! Yet something to improve:
[auto build test ERROR on tip/perf/core]
[also build test ERROR on v4.18-rc3 next-20180702]
[if your patch is applied to the wrong git tree, please drop us a note to help improve the system]
url: https://github.com/0day-ci/linux/commits/Mathieu-Poirier/perf-Add-ioctl-for-PMU-driver-configuration/20180703-064327
config: s390-defconfig (attached as .config)
compiler: s390x-linux-gnu-gcc (Debian 7.2.0-11) 7.2.0
reproduce:
wget https://raw.githubusercontent.com/intel/lkp-tests/master/sbin/make.cross -O ~/bin/make.cross
chmod +x ~/bin/make.cross
# save the attached .config to linux build tree
GCC_VERSION=7.2.0 make.cross ARCH=s390
All error/warnings (new ones prefixed by >>):
>> arch/s390/kernel/perf_cpum_sf.c:1606:1: error: expected identifier or '(' before '{' token
{
^
>> arch/s390/kernel/perf_cpum_sf.c:1604:14: warning: 'aux_buffer_setup' used but never defined
static void *aux_buffer_setup(struct perf_event *event, void **pages,
^~~~~~~~~~~~~~~~
vim +1606 arch/s390/kernel/perf_cpum_sf.c
ca5955cd Pu Hou 2016-11-11 1589
ca5955cd Pu Hou 2016-11-11 1590 /*
ca5955cd Pu Hou 2016-11-11 1591 * aux_buffer_setup() - Setup AUX buffer for diagnostic mode sampling
ca5955cd Pu Hou 2016-11-11 1592 * @cpu: On which to allocate, -1 means current
ca5955cd Pu Hou 2016-11-11 1593 * @pages: Array of pointers to buffer pages passed from perf core
ca5955cd Pu Hou 2016-11-11 1594 * @nr_pages: Total pages
ca5955cd Pu Hou 2016-11-11 1595 * @snapshot: Flag for snapshot mode
ca5955cd Pu Hou 2016-11-11 1596 *
ca5955cd Pu Hou 2016-11-11 1597 * This is the callback when setup an event using AUX buffer. Perf tool can
ca5955cd Pu Hou 2016-11-11 1598 * trigger this by an additional mmap() call on the event. Unlike the buffer
ca5955cd Pu Hou 2016-11-11 1599 * for basic samples, AUX buffer belongs to the event. It is scheduled with
ca5955cd Pu Hou 2016-11-11 1600 * the task among online cpus when it is a per-thread event.
ca5955cd Pu Hou 2016-11-11 1601 *
ca5955cd Pu Hou 2016-11-11 1602 * Return the private AUX buffer structure if success or NULL if fails.
ca5955cd Pu Hou 2016-11-11 1603 */
ceb39bf0 Mathieu Poirier 2018-07-02 @1604 static void *aux_buffer_setup(struct perf_event *event, void **pages,
ceb39bf0 Mathieu Poirier 2018-07-02 1605 int nr_pages, bool snapshot);
ca5955cd Pu Hou 2016-11-11 @1606 {
ca5955cd Pu Hou 2016-11-11 1607 struct sf_buffer *sfb;
ca5955cd Pu Hou 2016-11-11 1608 struct aux_buffer *aux;
ca5955cd Pu Hou 2016-11-11 1609 unsigned long *new, *tail;
ca5955cd Pu Hou 2016-11-11 1610 int i, n_sdbt;
ca5955cd Pu Hou 2016-11-11 1611
ca5955cd Pu Hou 2016-11-11 1612 if (!nr_pages || !pages)
ca5955cd Pu Hou 2016-11-11 1613 return NULL;
ca5955cd Pu Hou 2016-11-11 1614
ca5955cd Pu Hou 2016-11-11 1615 if (nr_pages > CPUM_SF_MAX_SDB * CPUM_SF_SDB_DIAG_FACTOR) {
ca5955cd Pu Hou 2016-11-11 1616 pr_err("AUX buffer size (%i pages) is larger than the "
ca5955cd Pu Hou 2016-11-11 1617 "maximum sampling buffer limit\n",
ca5955cd Pu Hou 2016-11-11 1618 nr_pages);
ca5955cd Pu Hou 2016-11-11 1619 return NULL;
ca5955cd Pu Hou 2016-11-11 1620 } else if (nr_pages < CPUM_SF_MIN_SDB * CPUM_SF_SDB_DIAG_FACTOR) {
ca5955cd Pu Hou 2016-11-11 1621 pr_err("AUX buffer size (%i pages) is less than the "
ca5955cd Pu Hou 2016-11-11 1622 "minimum sampling buffer limit\n",
ca5955cd Pu Hou 2016-11-11 1623 nr_pages);
ca5955cd Pu Hou 2016-11-11 1624 return NULL;
ca5955cd Pu Hou 2016-11-11 1625 }
ca5955cd Pu Hou 2016-11-11 1626
ca5955cd Pu Hou 2016-11-11 1627 /* Allocate aux_buffer struct for the event */
ca5955cd Pu Hou 2016-11-11 1628 aux = kmalloc(sizeof(struct aux_buffer), GFP_KERNEL);
ca5955cd Pu Hou 2016-11-11 1629 if (!aux)
ca5955cd Pu Hou 2016-11-11 1630 goto no_aux;
ca5955cd Pu Hou 2016-11-11 1631 sfb = &aux->sfb;
ca5955cd Pu Hou 2016-11-11 1632
ca5955cd Pu Hou 2016-11-11 1633 /* Allocate sdbt_index for fast reference */
ca5955cd Pu Hou 2016-11-11 1634 n_sdbt = (nr_pages + CPUM_SF_SDB_PER_TABLE - 1) / CPUM_SF_SDB_PER_TABLE;
ca5955cd Pu Hou 2016-11-11 1635 aux->sdbt_index = kmalloc_array(n_sdbt, sizeof(void *), GFP_KERNEL);
ca5955cd Pu Hou 2016-11-11 1636 if (!aux->sdbt_index)
ca5955cd Pu Hou 2016-11-11 1637 goto no_sdbt_index;
ca5955cd Pu Hou 2016-11-11 1638
ca5955cd Pu Hou 2016-11-11 1639 /* Allocate sdb_index for fast reference */
ca5955cd Pu Hou 2016-11-11 1640 aux->sdb_index = kmalloc_array(nr_pages, sizeof(void *), GFP_KERNEL);
ca5955cd Pu Hou 2016-11-11 1641 if (!aux->sdb_index)
ca5955cd Pu Hou 2016-11-11 1642 goto no_sdb_index;
ca5955cd Pu Hou 2016-11-11 1643
ca5955cd Pu Hou 2016-11-11 1644 /* Allocate the first SDBT */
ca5955cd Pu Hou 2016-11-11 1645 sfb->num_sdbt = 0;
ca5955cd Pu Hou 2016-11-11 1646 sfb->sdbt = (unsigned long *) get_zeroed_page(GFP_KERNEL);
ca5955cd Pu Hou 2016-11-11 1647 if (!sfb->sdbt)
ca5955cd Pu Hou 2016-11-11 1648 goto no_sdbt;
ca5955cd Pu Hou 2016-11-11 1649 aux->sdbt_index[sfb->num_sdbt++] = (unsigned long)sfb->sdbt;
ca5955cd Pu Hou 2016-11-11 1650 tail = sfb->tail = sfb->sdbt;
ca5955cd Pu Hou 2016-11-11 1651
ca5955cd Pu Hou 2016-11-11 1652 /*
ca5955cd Pu Hou 2016-11-11 1653 * Link the provided pages of AUX buffer to SDBT.
ca5955cd Pu Hou 2016-11-11 1654 * Allocate SDBT if needed.
ca5955cd Pu Hou 2016-11-11 1655 */
ca5955cd Pu Hou 2016-11-11 1656 for (i = 0; i < nr_pages; i++, tail++) {
ca5955cd Pu Hou 2016-11-11 1657 if (require_table_link(tail)) {
ca5955cd Pu Hou 2016-11-11 1658 new = (unsigned long *) get_zeroed_page(GFP_KERNEL);
ca5955cd Pu Hou 2016-11-11 1659 if (!new)
ca5955cd Pu Hou 2016-11-11 1660 goto no_sdbt;
ca5955cd Pu Hou 2016-11-11 1661 aux->sdbt_index[sfb->num_sdbt++] = (unsigned long)new;
ca5955cd Pu Hou 2016-11-11 1662 /* Link current page to tail of chain */
ca5955cd Pu Hou 2016-11-11 1663 *tail = (unsigned long)(void *) new + 1;
ca5955cd Pu Hou 2016-11-11 1664 tail = new;
ca5955cd Pu Hou 2016-11-11 1665 }
ca5955cd Pu Hou 2016-11-11 1666 /* Tail is the entry in a SDBT */
ca5955cd Pu Hou 2016-11-11 1667 *tail = (unsigned long)pages[i];
ca5955cd Pu Hou 2016-11-11 1668 aux->sdb_index[i] = (unsigned long)pages[i];
ca5955cd Pu Hou 2016-11-11 1669 }
ca5955cd Pu Hou 2016-11-11 1670 sfb->num_sdb = nr_pages;
ca5955cd Pu Hou 2016-11-11 1671
ca5955cd Pu Hou 2016-11-11 1672 /* Link the last entry in the SDBT to the first SDBT */
ca5955cd Pu Hou 2016-11-11 1673 *tail = (unsigned long) sfb->sdbt + 1;
ca5955cd Pu Hou 2016-11-11 1674 sfb->tail = tail;
ca5955cd Pu Hou 2016-11-11 1675
ca5955cd Pu Hou 2016-11-11 1676 /*
ca5955cd Pu Hou 2016-11-11 1677 * Initial all SDBs are zeroed. Mark it as empty.
ca5955cd Pu Hou 2016-11-11 1678 * So there is no need to clear the full indicator
ca5955cd Pu Hou 2016-11-11 1679 * when this event is first added.
ca5955cd Pu Hou 2016-11-11 1680 */
ca5955cd Pu Hou 2016-11-11 1681 aux->empty_mark = sfb->num_sdb - 1;
ca5955cd Pu Hou 2016-11-11 1682
ca5955cd Pu Hou 2016-11-11 1683 debug_sprintf_event(sfdbg, 4, "aux_buffer_setup: setup %lu SDBTs"
ca5955cd Pu Hou 2016-11-11 1684 " and %lu SDBs\n",
ca5955cd Pu Hou 2016-11-11 1685 sfb->num_sdbt, sfb->num_sdb);
ca5955cd Pu Hou 2016-11-11 1686
ca5955cd Pu Hou 2016-11-11 1687 return aux;
ca5955cd Pu Hou 2016-11-11 1688
ca5955cd Pu Hou 2016-11-11 1689 no_sdbt:
ca5955cd Pu Hou 2016-11-11 1690 /* SDBs (AUX buffer pages) are freed by caller */
ca5955cd Pu Hou 2016-11-11 1691 for (i = 0; i < sfb->num_sdbt; i++)
ca5955cd Pu Hou 2016-11-11 1692 free_page(aux->sdbt_index[i]);
ca5955cd Pu Hou 2016-11-11 1693 kfree(aux->sdb_index);
ca5955cd Pu Hou 2016-11-11 1694 no_sdb_index:
ca5955cd Pu Hou 2016-11-11 1695 kfree(aux->sdbt_index);
ca5955cd Pu Hou 2016-11-11 1696 no_sdbt_index:
ca5955cd Pu Hou 2016-11-11 1697 kfree(aux);
ca5955cd Pu Hou 2016-11-11 1698 no_aux:
ca5955cd Pu Hou 2016-11-11 1699 return NULL;
ca5955cd Pu Hou 2016-11-11 1700 }
ca5955cd Pu Hou 2016-11-11 1701
:::::: The code at line 1606 was first introduced by commit
:::::: ca5955cdeae744edd3dcc65d677e833fc29658c2 s390/cpumf: introduce AUX buffer for dump diagnostic sample data
:::::: TO: Pu Hou <[email protected]>
:::::: CC: Martin Schwidefsky <[email protected]>
---
0-DAY kernel test infrastructure Open Source Technology Center
https://lists.01.org/pipermail/kbuild-all Intel Corporation
On Mon, Jul 02, 2018 at 04:33:28PM -0600, Mathieu Poirier wrote:
> It can be advantagous to have access to all the information conveyed by
> a perf_event when setting up the AUX buffer, as it is the case when
> dealing with PMU specific driver configuration communicated to the kernel
> using an ioctl() call.
>
> As such simply replace the cpu information by the complete perf_event
> structure and change all affected customers.
>
> Signed-off-by: Mathieu Poirier <[email protected]>
> ---
> arch/s390/kernel/perf_cpum_sf.c | 4 ++--
> arch/x86/events/intel/bts.c | 4 +++-
> arch/x86/events/intel/pt.c | 5 +++--
> drivers/hwtracing/coresight/coresight-etm-perf.c | 6 +++---
> drivers/perf/arm_spe_pmu.c | 6 +++---
> include/linux/perf_event.h | 2 +-
> kernel/events/ring_buffer.c | 2 +-
> 7 files changed, 16 insertions(+), 13 deletions(-)
>
> diff --git a/arch/s390/kernel/perf_cpum_sf.c b/arch/s390/kernel/perf_cpum_sf.c
> index 0292d68e7dde..e06daaa08894 100644
> --- a/arch/s390/kernel/perf_cpum_sf.c
> +++ b/arch/s390/kernel/perf_cpum_sf.c
> @@ -1601,8 +1601,8 @@ static void aux_buffer_free(void *data)
> *
> * Return the private AUX buffer structure if success or NULL if fails.
> */
> -static void *aux_buffer_setup(int cpu, void **pages, int nr_pages,
> - bool snapshot)
> +static void *aux_buffer_setup(struct perf_event *event, void **pages,
> + int nr_pages, bool snapshot);
Please remove the trailing semi-colon (;) in the function definition causing
the kbuild error. Also, it would be great if you also could update the
function comment and replace the @cpu by the @event.
Many thanks.
On Mon, Jul 02, 2018 at 04:33:28PM -0600, Mathieu Poirier wrote:
> It can be advantagous to have access to all the information conveyed by
> a perf_event when setting up the AUX buffer, as it is the case when
> dealing with PMU specific driver configuration communicated to the kernel
> using an ioctl() call.
>
> As such simply replace the cpu information by the complete perf_event
> structure and change all affected customers.
Ok, but can you provide details about which parts of perf_event you need
and how they are used? Is it the attribute? What's the ioctl command in
question? Also, should we worry that the PMU specific configuration isn't
part of the attribute? Or, if not, why? I think we talked about this
before, but I forgot and this patch needs to explain it anyway.
Thanks!
--
Alex
On Mon, Jul 02, 2018 at 04:33:29PM -0600, Mathieu Poirier wrote:
> This patch follows what has been done for filters by adding an ioctl()
> option to communicate to the kernel arbitrary PMU specific configuration
> that don't fit in the conventional struct perf_event_attr to the kernel.
Ok, so what *is* the PMU specific configuration that doesn't fit in the
attribute and needs to be re-configured by the driver using the generation
tracking?
> Signed-off-by: Mathieu Poirier <[email protected]>
> ---
> include/linux/perf_event.h | 54 ++++++++++++++++++++++
> kernel/events/core.c | 110 +++++++++++++++++++++++++++++++++++++++++++++
> 2 files changed, 164 insertions(+)
>
> diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
> index 4d9c8f30ca6c..6e06b63c262f 100644
> --- a/include/linux/perf_event.h
> +++ b/include/linux/perf_event.h
> @@ -178,6 +178,12 @@ struct hw_perf_event {
> /* Last sync'ed generation of filters */
> unsigned long addr_filters_gen;
>
> + /* PMU driver configuration */
> + void *drv_config;
> +
> + /* Last sync'ed generation of driver config */
> + unsigned long drv_config_gen;
> +
> /*
> * hw_perf_event::state flags; used to track the PERF_EF_* state.
> */
> @@ -447,6 +453,26 @@ struct pmu {
> * Filter events for PMU-specific reasons.
> */
> int (*filter_match) (struct perf_event *event); /* optional */
> +
> + /*
> + * Valiate complex PMU configuration that don't fit in the
> + * perf_event_attr struct. Returns a PMU specific pointer or an error
> + * value < 0.
> + *
> + * As with addr_filters_validate(), runs in the context of the ioctl()
> + * process and is not serialized with the rest of the PMU callbacks.
Yes, but what is it? I get it that it's probably in one of the other
patches, but we still need to mention it somewhere here.
> + */
> + void *(*drv_config_validate) (struct perf_event *event,
> + char *config_str);
> +
> + /* Synchronize PMU driver configuration */
> + void (*drv_config_sync) (struct perf_event *event);
> +
> + /*
> + * Release PMU specific configuration acquired by
> + * drv_config_validate()
> + */
> + void (*drv_config_free) (void *drv_data);
> };
>
> enum perf_addr_filter_action_t {
> @@ -489,6 +515,11 @@ struct perf_addr_filters_head {
> unsigned int nr_file_filters;
> };
>
> +struct perf_drv_config {
> + void *drv_config;
> + raw_spinlock_t lock;
> +};
> +
> /**
> * enum perf_event_state - the states of a event
> */
> @@ -668,6 +699,10 @@ struct perf_event {
> unsigned long *addr_filters_offs;
> unsigned long addr_filters_gen;
>
> + /* PMU driver specific configuration */
> + struct perf_drv_config drv_config;
> + unsigned long drv_config_gen;
> +
> void (*destroy)(struct perf_event *);
> struct rcu_head rcu_head;
>
> @@ -1234,6 +1269,13 @@ static inline bool has_addr_filter(struct perf_event *event)
> return event->pmu->nr_addr_filters;
> }
>
> +static inline bool has_drv_config(struct perf_event *event)
> +{
> + return event->pmu->drv_config_validate &&
> + event->pmu->drv_config_sync &&
> + event->pmu->drv_config_free;
> +}
> +
> /*
> * An inherited event uses parent's filters
> */
> @@ -1248,7 +1290,19 @@ perf_event_addr_filters(struct perf_event *event)
> return ifh;
> }
>
> +static inline struct perf_drv_config *
> +perf_event_get_drv_config(struct perf_event *event)
> +{
> + struct perf_drv_config *cfg = &event->drv_config;
> +
> + if (event->parent)
> + cfg = &event->parent->drv_config;
> +
> + return cfg;
> +}
> +
> extern void perf_event_addr_filters_sync(struct perf_event *event);
> +extern void perf_event_drv_config_sync(struct perf_event *event);
>
> extern int perf_output_begin(struct perf_output_handle *handle,
> struct perf_event *event, unsigned int size);
> diff --git a/kernel/events/core.c b/kernel/events/core.c
> index 8f0434a9951a..701839866789 100644
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -2829,6 +2829,29 @@ void perf_event_addr_filters_sync(struct perf_event *event)
> }
> EXPORT_SYMBOL_GPL(perf_event_addr_filters_sync);
>
> +/*
> + * PMU driver configuration works the same way as filter management above,
> + * but without the need to deal with memory mapping. Driver configuration
> + * arrives through the SET_DRV_CONFIG ioctl() where it is validated and applied
> + * to the event. When the PMU is ready it calls perf_event_drv_config_sync() to
> + * bring the configuration information within reach of the PMU.
Wait a second. The reason why we dance around with the generations of filters
is the locking order of ctx::mutex vs mmap_sem. In an mmap path, where we're
notified about mapping changes, we're called under the latter, and we'd need
to grab the former to update the event configuration. In your case, the
update comes in via perf_ioctl(), where we're already holding the ctx::mutex,
so you can just kick the PMU right there, via an event_function_call() or
perf_event_stop(restart=1). In the latter case, your pmu::start() would just
grab the new configuration. Should also be about 90% less code. :)
Would this work for you or am I misunderstanding something about your
requirements?
> + */
> +void perf_event_drv_config_sync(struct perf_event *event)
> +{
> + struct perf_drv_config *drv_config = perf_event_get_drv_config(event);
> +
> + if (!has_drv_config(event))
> + return;
> +
> + raw_spin_lock(&drv_config->lock);
> + if (event->drv_config_gen != event->hw.drv_config_gen) {
> + event->pmu->drv_config_sync(event);
> + event->hw.drv_config_gen = event->drv_config_gen;
> + }
> + raw_spin_unlock(&drv_config->lock);
> +}
> +EXPORT_SYMBOL_GPL(perf_event_drv_config_sync);
> +
> static int _perf_event_refresh(struct perf_event *event, int refresh)
> {
> /*
> @@ -4410,6 +4433,7 @@ static bool exclusive_event_installable(struct perf_event *event,
>
> static void perf_addr_filters_splice(struct perf_event *event,
> struct list_head *head);
> +static void perf_drv_config_splice(struct perf_event *event, void *drv_data);
>
> static void _free_event(struct perf_event *event)
> {
> @@ -4440,6 +4464,7 @@ static void _free_event(struct perf_event *event)
> perf_event_free_bpf_prog(event);
> perf_addr_filters_splice(event, NULL);
> kfree(event->addr_filters_offs);
> + perf_drv_config_splice(event, NULL);
>
> if (event->destroy)
> event->destroy(event);
> @@ -5002,6 +5027,8 @@ static inline int perf_fget_light(int fd, struct fd *p)
> static int perf_event_set_output(struct perf_event *event,
> struct perf_event *output_event);
> static int perf_event_set_filter(struct perf_event *event, void __user *arg);
> +static int perf_event_set_drv_config(struct perf_event *event,
> + void __user *arg);
> static int perf_event_set_bpf_prog(struct perf_event *event, u32 prog_fd);
> static int perf_copy_attr(struct perf_event_attr __user *uattr,
> struct perf_event_attr *attr);
> @@ -5088,6 +5115,10 @@ static long _perf_ioctl(struct perf_event *event, unsigned int cmd, unsigned lon
>
> return perf_event_modify_attr(event, &new_attr);
> }
> +
> + case PERF_EVENT_IOC_SET_DRV_CONFIG:
> + return perf_event_set_drv_config(event, (void __user *)arg);
> +
> default:
> return -ENOTTY;
> }
> @@ -9086,6 +9117,85 @@ static int perf_event_set_filter(struct perf_event *event, void __user *arg)
> return ret;
> }
>
> +static void perf_drv_config_splice(struct perf_event *event, void *drv_data)
I think the address filter counterpart is called "splice" because it takes
a list_head as a parameter and splices that list into the list of filters.
I'd suggest that this one is more like "replace", but up to you.
> +{
> + unsigned long flags;
> + void *old_drv_data;
> +
> + if (!has_drv_config(event))
> + return;
> +
> + /* Children take their configuration from their parent */
> + if (event->parent)
> + return;
> +
> + raw_spin_lock_irqsave(&event->drv_config.lock, flags);
> +
> + old_drv_data = event->drv_config.drv_config;
> + event->drv_config.drv_config = drv_data;
Now I'm thinking: should we reset the generation here (and also in the
address filters bit)? At least, it deserves a comment.
> +
> + raw_spin_unlock_irqrestore(&event->drv_config.lock, flags);
> +
> + event->pmu->drv_config_free(old_drv_data);
> +}
> +
> +static void perf_event_drv_config_apply(struct perf_event *event)
> +{
> + unsigned long flags;
> + struct perf_drv_config *drv_config = perf_event_get_drv_config(event);
> +
> + /* Notify event that a new configuration is available */
> + raw_spin_lock_irqsave(&drv_config->lock, flags);
> + event->drv_config_gen++;
> + raw_spin_unlock_irqrestore(&drv_config->lock, flags);
Should we also mention how this new locks fits into the existing locking
order?
Regards,
--
Alex
On Tue, Jul 03, 2018 at 01:03:48PM +0300, Alexander Shishkin wrote:
> On Mon, Jul 02, 2018 at 04:33:29PM -0600, Mathieu Poirier wrote:
> > +/*
> > + * PMU driver configuration works the same way as filter management above,
> > + * but without the need to deal with memory mapping. Driver configuration
> > + * arrives through the SET_DRV_CONFIG ioctl() where it is validated and applied
> > + * to the event. When the PMU is ready it calls perf_event_drv_config_sync() to
> > + * bring the configuration information within reach of the PMU.
>
> Wait a second. The reason why we dance around with the generations of filters
> is the locking order of ctx::mutex vs mmap_sem. In an mmap path, where we're
> notified about mapping changes, we're called under the latter, and we'd need
> to grab the former to update the event configuration. In your case, the
> update comes in via perf_ioctl(), where we're already holding the ctx::mutex,
> so you can just kick the PMU right there, via an event_function_call() or
> perf_event_stop(restart=1). In the latter case, your pmu::start() would just
> grab the new configuration. Should also be about 90% less code. :)
Also, since it affects the AUX buffer configuration, it is probably a one
time ioctl command that you issue before you mmap the buffer. If that's
the case, you don't even have to worry about stopping the event, as it
shouldn't be running, because without the buffer perf_aux_output_begin()
should fail and so should the pmu::add() iirc.
Regards,
--
Alex
On Mon, Jul 02, 2018 at 04:33:29PM -0600, Mathieu Poirier wrote:
SNIP
> +static int
> +perf_event_process_drv_config(struct perf_event *event, char *config_str)
> +{
> + int ret = -EINVAL;
> + void *drv_data;
> +
> + /* Make sure ctx.mutex it held */
> + lockdep_assert_held(&event->ctx->mutex);
> +
> + /* Children take their configuration from their parent */
> + if (WARN_ON_ONCE(event->parent))
> + goto out;
> +
> + drv_data = event->pmu->drv_config_validate(event, config_str);
> + if (!drv_data)
> + goto out;
what's this check for? doc does not mention that,
also the coresight callback does not return NULL
jirka
On Mon, Jul 02, 2018 at 04:33:29PM -0600, Mathieu Poirier wrote:
SNIP
> + ret = 0;
> +out:
> + return ret;
> +}
> +
> +static int perf_event_set_drv_config(struct perf_event *event, void __user *arg)
> +{
> + int ret = -EINVAL;
> + char *config_str;
> +
> + config_str = strndup_user(arg, PAGE_SIZE);
> + if (IS_ERR(config_str))
> + return PTR_ERR(config_str);
> +
> + if (has_drv_config(event))
would it be better to make this check before
the alloc/copy data from user takes place?
jirka
On Tue, 3 Jul 2018 at 01:31, Hendrik Brueckner <[email protected]> wrote:
>
> On Mon, Jul 02, 2018 at 04:33:28PM -0600, Mathieu Poirier wrote:
> > It can be advantagous to have access to all the information conveyed by
> > a perf_event when setting up the AUX buffer, as it is the case when
> > dealing with PMU specific driver configuration communicated to the kernel
> > using an ioctl() call.
> >
> > As such simply replace the cpu information by the complete perf_event
> > structure and change all affected customers.
> >
> > Signed-off-by: Mathieu Poirier <[email protected]>
> > ---
> > arch/s390/kernel/perf_cpum_sf.c | 4 ++--
> > arch/x86/events/intel/bts.c | 4 +++-
> > arch/x86/events/intel/pt.c | 5 +++--
> > drivers/hwtracing/coresight/coresight-etm-perf.c | 6 +++---
> > drivers/perf/arm_spe_pmu.c | 6 +++---
> > include/linux/perf_event.h | 2 +-
> > kernel/events/ring_buffer.c | 2 +-
> > 7 files changed, 16 insertions(+), 13 deletions(-)
> >
> > diff --git a/arch/s390/kernel/perf_cpum_sf.c b/arch/s390/kernel/perf_cpum_sf.c
> > index 0292d68e7dde..e06daaa08894 100644
> > --- a/arch/s390/kernel/perf_cpum_sf.c
> > +++ b/arch/s390/kernel/perf_cpum_sf.c
> > @@ -1601,8 +1601,8 @@ static void aux_buffer_free(void *data)
> > *
> > * Return the private AUX buffer structure if success or NULL if fails.
> > */
> > -static void *aux_buffer_setup(int cpu, void **pages, int nr_pages,
> > - bool snapshot)
> > +static void *aux_buffer_setup(struct perf_event *event, void **pages,
> > + int nr_pages, bool snapshot);
>
> Please remove the trailing semi-colon (;) in the function definition causing
> the kbuild error. Also, it would be great if you also could update the
> function comment and replace the @cpu by the @event.
Had I realised it was just as easy to compile for s390 as it is for
ARM64, I would have done so ahead of time. Well done on that front.
All fixed now.
Thanks,
Mathieu
>
> Many thanks.
>
On Tue, 3 Jul 2018 at 07:40, Jiri Olsa <[email protected]> wrote:
>
> On Mon, Jul 02, 2018 at 04:33:29PM -0600, Mathieu Poirier wrote:
>
> SNIP
>
> > +static int
> > +perf_event_process_drv_config(struct perf_event *event, char *config_str)
> > +{
> > + int ret = -EINVAL;
> > + void *drv_data;
> > +
> > + /* Make sure ctx.mutex it held */
> > + lockdep_assert_held(&event->ctx->mutex);
> > +
> > + /* Children take their configuration from their parent */
> > + if (WARN_ON_ONCE(event->parent))
> > + goto out;
> > +
> > + drv_data = event->pmu->drv_config_validate(event, config_str);
> > + if (!drv_data)
> > + goto out;
>
> what's this check for? doc does not mention that,
> also the coresight callback does not return NULL
>
Very true - thanks for pointing this out.
Mathieu
> jirka
Hi Alex,
On Tue, 3 Jul 2018 at 04:03, Alexander Shishkin
<[email protected]> wrote:
>
> On Mon, Jul 02, 2018 at 04:33:29PM -0600, Mathieu Poirier wrote:
> > This patch follows what has been done for filters by adding an ioctl()
> > option to communicate to the kernel arbitrary PMU specific configuration
> > that don't fit in the conventional struct perf_event_attr to the kernel.
>
> Ok, so what *is* the PMU specific configuration that doesn't fit in the
> attribute and needs to be re-configured by the driver using the generation
> tracking?
>
In this patchset I'm am after the specification of sink information
for each event, i.e what sink a CPU is supposed to use for the
session. I simply don't see putting something that PMU specific in
the generic perf_event_attr structure. I also intend to use the same
ioctl mechanism to communicate complex tracer configuration for
sequencers, counters and input events. I don't see a nice way of
doing that from the perf_event_attr, and that is even without thinking
about the different flavours of tracers out there, all with their own
features.
I've looked around and the only clean way I found to support this is
via an ioctl(). That way each tracer can easily identify the sink it
should be using without smearing the perf_event_attr structure. I
would be happy to explore a different avenue should you think of
something.
Thanks,
Mathieu
> > Signed-off-by: Mathieu Poirier <[email protected]>
> > ---
> > include/linux/perf_event.h | 54 ++++++++++++++++++++++
> > kernel/events/core.c | 110 +++++++++++++++++++++++++++++++++++++++++++++
> > 2 files changed, 164 insertions(+)
> >
> > diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
> > index 4d9c8f30ca6c..6e06b63c262f 100644
> > --- a/include/linux/perf_event.h
> > +++ b/include/linux/perf_event.h
> > @@ -178,6 +178,12 @@ struct hw_perf_event {
> > /* Last sync'ed generation of filters */
> > unsigned long addr_filters_gen;
> >
> > + /* PMU driver configuration */
> > + void *drv_config;
> > +
> > + /* Last sync'ed generation of driver config */
> > + unsigned long drv_config_gen;
> > +
> > /*
> > * hw_perf_event::state flags; used to track the PERF_EF_* state.
> > */
> > @@ -447,6 +453,26 @@ struct pmu {
> > * Filter events for PMU-specific reasons.
> > */
> > int (*filter_match) (struct perf_event *event); /* optional */
> > +
> > + /*
> > + * Valiate complex PMU configuration that don't fit in the
> > + * perf_event_attr struct. Returns a PMU specific pointer or an error
> > + * value < 0.
> > + *
> > + * As with addr_filters_validate(), runs in the context of the ioctl()
> > + * process and is not serialized with the rest of the PMU callbacks.
>
> Yes, but what is it? I get it that it's probably in one of the other
> patches, but we still need to mention it somewhere here.
>
> > + */
> > + void *(*drv_config_validate) (struct perf_event *event,
> > + char *config_str);
> > +
> > + /* Synchronize PMU driver configuration */
> > + void (*drv_config_sync) (struct perf_event *event);
> > +
> > + /*
> > + * Release PMU specific configuration acquired by
> > + * drv_config_validate()
> > + */
> > + void (*drv_config_free) (void *drv_data);
> > };
> >
> > enum perf_addr_filter_action_t {
> > @@ -489,6 +515,11 @@ struct perf_addr_filters_head {
> > unsigned int nr_file_filters;
> > };
> >
> > +struct perf_drv_config {
> > + void *drv_config;
> > + raw_spinlock_t lock;
> > +};
> > +
> > /**
> > * enum perf_event_state - the states of a event
> > */
> > @@ -668,6 +699,10 @@ struct perf_event {
> > unsigned long *addr_filters_offs;
> > unsigned long addr_filters_gen;
> >
> > + /* PMU driver specific configuration */
> > + struct perf_drv_config drv_config;
> > + unsigned long drv_config_gen;
> > +
> > void (*destroy)(struct perf_event *);
> > struct rcu_head rcu_head;
> >
> > @@ -1234,6 +1269,13 @@ static inline bool has_addr_filter(struct perf_event *event)
> > return event->pmu->nr_addr_filters;
> > }
> >
> > +static inline bool has_drv_config(struct perf_event *event)
> > +{
> > + return event->pmu->drv_config_validate &&
> > + event->pmu->drv_config_sync &&
> > + event->pmu->drv_config_free;
> > +}
> > +
> > /*
> > * An inherited event uses parent's filters
> > */
> > @@ -1248,7 +1290,19 @@ perf_event_addr_filters(struct perf_event *event)
> > return ifh;
> > }
> >
> > +static inline struct perf_drv_config *
> > +perf_event_get_drv_config(struct perf_event *event)
> > +{
> > + struct perf_drv_config *cfg = &event->drv_config;
> > +
> > + if (event->parent)
> > + cfg = &event->parent->drv_config;
> > +
> > + return cfg;
> > +}
> > +
> > extern void perf_event_addr_filters_sync(struct perf_event *event);
> > +extern void perf_event_drv_config_sync(struct perf_event *event);
> >
> > extern int perf_output_begin(struct perf_output_handle *handle,
> > struct perf_event *event, unsigned int size);
> > diff --git a/kernel/events/core.c b/kernel/events/core.c
> > index 8f0434a9951a..701839866789 100644
> > --- a/kernel/events/core.c
> > +++ b/kernel/events/core.c
> > @@ -2829,6 +2829,29 @@ void perf_event_addr_filters_sync(struct perf_event *event)
> > }
> > EXPORT_SYMBOL_GPL(perf_event_addr_filters_sync);
> >
> > +/*
> > + * PMU driver configuration works the same way as filter management above,
> > + * but without the need to deal with memory mapping. Driver configuration
> > + * arrives through the SET_DRV_CONFIG ioctl() where it is validated and applied
> > + * to the event. When the PMU is ready it calls perf_event_drv_config_sync() to
> > + * bring the configuration information within reach of the PMU.
>
> Wait a second. The reason why we dance around with the generations of filters
> is the locking order of ctx::mutex vs mmap_sem. In an mmap path, where we're
> notified about mapping changes, we're called under the latter, and we'd need
> to grab the former to update the event configuration. In your case, the
> update comes in via perf_ioctl(), where we're already holding the ctx::mutex,
> so you can just kick the PMU right there, via an event_function_call() or
> perf_event_stop(restart=1). In the latter case, your pmu::start() would just
> grab the new configuration. Should also be about 90% less code. :)
>
> Would this work for you or am I misunderstanding something about your
> requirements?
>
> > + */
> > +void perf_event_drv_config_sync(struct perf_event *event)
> > +{
> > + struct perf_drv_config *drv_config = perf_event_get_drv_config(event);
> > +
> > + if (!has_drv_config(event))
> > + return;
> > +
> > + raw_spin_lock(&drv_config->lock);
> > + if (event->drv_config_gen != event->hw.drv_config_gen) {
> > + event->pmu->drv_config_sync(event);
> > + event->hw.drv_config_gen = event->drv_config_gen;
> > + }
> > + raw_spin_unlock(&drv_config->lock);
> > +}
> > +EXPORT_SYMBOL_GPL(perf_event_drv_config_sync);
> > +
> > static int _perf_event_refresh(struct perf_event *event, int refresh)
> > {
> > /*
> > @@ -4410,6 +4433,7 @@ static bool exclusive_event_installable(struct perf_event *event,
> >
> > static void perf_addr_filters_splice(struct perf_event *event,
> > struct list_head *head);
> > +static void perf_drv_config_splice(struct perf_event *event, void *drv_data);
> >
> > static void _free_event(struct perf_event *event)
> > {
> > @@ -4440,6 +4464,7 @@ static void _free_event(struct perf_event *event)
> > perf_event_free_bpf_prog(event);
> > perf_addr_filters_splice(event, NULL);
> > kfree(event->addr_filters_offs);
> > + perf_drv_config_splice(event, NULL);
> >
> > if (event->destroy)
> > event->destroy(event);
> > @@ -5002,6 +5027,8 @@ static inline int perf_fget_light(int fd, struct fd *p)
> > static int perf_event_set_output(struct perf_event *event,
> > struct perf_event *output_event);
> > static int perf_event_set_filter(struct perf_event *event, void __user *arg);
> > +static int perf_event_set_drv_config(struct perf_event *event,
> > + void __user *arg);
> > static int perf_event_set_bpf_prog(struct perf_event *event, u32 prog_fd);
> > static int perf_copy_attr(struct perf_event_attr __user *uattr,
> > struct perf_event_attr *attr);
> > @@ -5088,6 +5115,10 @@ static long _perf_ioctl(struct perf_event *event, unsigned int cmd, unsigned lon
> >
> > return perf_event_modify_attr(event, &new_attr);
> > }
> > +
> > + case PERF_EVENT_IOC_SET_DRV_CONFIG:
> > + return perf_event_set_drv_config(event, (void __user *)arg);
> > +
> > default:
> > return -ENOTTY;
> > }
> > @@ -9086,6 +9117,85 @@ static int perf_event_set_filter(struct perf_event *event, void __user *arg)
> > return ret;
> > }
> >
> > +static void perf_drv_config_splice(struct perf_event *event, void *drv_data)
>
> I think the address filter counterpart is called "splice" because it takes
> a list_head as a parameter and splices that list into the list of filters.
> I'd suggest that this one is more like "replace", but up to you.
>
> > +{
> > + unsigned long flags;
> > + void *old_drv_data;
> > +
> > + if (!has_drv_config(event))
> > + return;
> > +
> > + /* Children take their configuration from their parent */
> > + if (event->parent)
> > + return;
> > +
> > + raw_spin_lock_irqsave(&event->drv_config.lock, flags);
> > +
> > + old_drv_data = event->drv_config.drv_config;
> > + event->drv_config.drv_config = drv_data;
>
> Now I'm thinking: should we reset the generation here (and also in the
> address filters bit)? At least, it deserves a comment.
>
> > +
> > + raw_spin_unlock_irqrestore(&event->drv_config.lock, flags);
> > +
> > + event->pmu->drv_config_free(old_drv_data);
> > +}
> > +
> > +static void perf_event_drv_config_apply(struct perf_event *event)
> > +{
> > + unsigned long flags;
> > + struct perf_drv_config *drv_config = perf_event_get_drv_config(event);
> > +
> > + /* Notify event that a new configuration is available */
> > + raw_spin_lock_irqsave(&drv_config->lock, flags);
> > + event->drv_config_gen++;
> > + raw_spin_unlock_irqrestore(&drv_config->lock, flags);
>
> Should we also mention how this new locks fits into the existing locking
> order?
>
> Regards,
> --
> Alex
On Tue, 3 Jul 2018 at 04:03, Alexander Shishkin
<[email protected]> wrote:
>
> On Mon, Jul 02, 2018 at 04:33:29PM -0600, Mathieu Poirier wrote:
> > This patch follows what has been done for filters by adding an ioctl()
> > option to communicate to the kernel arbitrary PMU specific configuration
> > that don't fit in the conventional struct perf_event_attr to the kernel.
>
> Ok, so what *is* the PMU specific configuration that doesn't fit in the
> attribute and needs to be re-configured by the driver using the generation
> tracking?
>
> > Signed-off-by: Mathieu Poirier <[email protected]>
> > ---
> > include/linux/perf_event.h | 54 ++++++++++++++++++++++
> > kernel/events/core.c | 110 +++++++++++++++++++++++++++++++++++++++++++++
> > 2 files changed, 164 insertions(+)
> >
> > diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
> > index 4d9c8f30ca6c..6e06b63c262f 100644
> > --- a/include/linux/perf_event.h
> > +++ b/include/linux/perf_event.h
> > @@ -178,6 +178,12 @@ struct hw_perf_event {
> > /* Last sync'ed generation of filters */
> > unsigned long addr_filters_gen;
> >
> > + /* PMU driver configuration */
> > + void *drv_config;
> > +
> > + /* Last sync'ed generation of driver config */
> > + unsigned long drv_config_gen;
> > +
> > /*
> > * hw_perf_event::state flags; used to track the PERF_EF_* state.
> > */
> > @@ -447,6 +453,26 @@ struct pmu {
> > * Filter events for PMU-specific reasons.
> > */
> > int (*filter_match) (struct perf_event *event); /* optional */
> > +
> > + /*
> > + * Valiate complex PMU configuration that don't fit in the
> > + * perf_event_attr struct. Returns a PMU specific pointer or an error
> > + * value < 0.
> > + *
> > + * As with addr_filters_validate(), runs in the context of the ioctl()
> > + * process and is not serialized with the rest of the PMU callbacks.
>
> Yes, but what is it? I get it that it's probably in one of the other
> patches, but we still need to mention it somewhere here.
I could write a more detailed description here, something about the
specification of sink and configuration of coresight specific features
(sequencers, counters, input/output) but I decided to keep things
generic. Rather than doing that I thought it best to leave things
generic and let people look at the code for more details should they
want to. Let me know what you think is best.
>
> > + */
> > + void *(*drv_config_validate) (struct perf_event *event,
> > + char *config_str);
> > +
> > + /* Synchronize PMU driver configuration */
> > + void (*drv_config_sync) (struct perf_event *event);
> > +
> > + /*
> > + * Release PMU specific configuration acquired by
> > + * drv_config_validate()
> > + */
> > + void (*drv_config_free) (void *drv_data);
> > };
> >
> > enum perf_addr_filter_action_t {
> > @@ -489,6 +515,11 @@ struct perf_addr_filters_head {
> > unsigned int nr_file_filters;
> > };
> >
> > +struct perf_drv_config {
> > + void *drv_config;
> > + raw_spinlock_t lock;
> > +};
> > +
> > /**
> > * enum perf_event_state - the states of a event
> > */
> > @@ -668,6 +699,10 @@ struct perf_event {
> > unsigned long *addr_filters_offs;
> > unsigned long addr_filters_gen;
> >
> > + /* PMU driver specific configuration */
> > + struct perf_drv_config drv_config;
> > + unsigned long drv_config_gen;
> > +
> > void (*destroy)(struct perf_event *);
> > struct rcu_head rcu_head;
> >
> > @@ -1234,6 +1269,13 @@ static inline bool has_addr_filter(struct perf_event *event)
> > return event->pmu->nr_addr_filters;
> > }
> >
> > +static inline bool has_drv_config(struct perf_event *event)
> > +{
> > + return event->pmu->drv_config_validate &&
> > + event->pmu->drv_config_sync &&
> > + event->pmu->drv_config_free;
> > +}
> > +
> > /*
> > * An inherited event uses parent's filters
> > */
> > @@ -1248,7 +1290,19 @@ perf_event_addr_filters(struct perf_event *event)
> > return ifh;
> > }
> >
> > +static inline struct perf_drv_config *
> > +perf_event_get_drv_config(struct perf_event *event)
> > +{
> > + struct perf_drv_config *cfg = &event->drv_config;
> > +
> > + if (event->parent)
> > + cfg = &event->parent->drv_config;
> > +
> > + return cfg;
> > +}
> > +
> > extern void perf_event_addr_filters_sync(struct perf_event *event);
> > +extern void perf_event_drv_config_sync(struct perf_event *event);
> >
> > extern int perf_output_begin(struct perf_output_handle *handle,
> > struct perf_event *event, unsigned int size);
> > diff --git a/kernel/events/core.c b/kernel/events/core.c
> > index 8f0434a9951a..701839866789 100644
> > --- a/kernel/events/core.c
> > +++ b/kernel/events/core.c
> > @@ -2829,6 +2829,29 @@ void perf_event_addr_filters_sync(struct perf_event *event)
> > }
> > EXPORT_SYMBOL_GPL(perf_event_addr_filters_sync);
> >
> > +/*
> > + * PMU driver configuration works the same way as filter management above,
> > + * but without the need to deal with memory mapping. Driver configuration
> > + * arrives through the SET_DRV_CONFIG ioctl() where it is validated and applied
> > + * to the event. When the PMU is ready it calls perf_event_drv_config_sync() to
> > + * bring the configuration information within reach of the PMU.
>
> Wait a second. The reason why we dance around with the generations of filters
> is the locking order of ctx::mutex vs mmap_sem. In an mmap path, where we're
> notified about mapping changes, we're called under the latter, and we'd need
> to grab the former to update the event configuration. In your case, the
> update comes in via perf_ioctl(), where we're already holding the ctx::mutex,
> so you can just kick the PMU right there, via an event_function_call() or
> perf_event_stop(restart=1). In the latter case, your pmu::start() would just
> grab the new configuration. Should also be about 90% less code. :)
>
> Would this work for you or am I misunderstanding something about your
> requirements?
>
> > + */
> > +void perf_event_drv_config_sync(struct perf_event *event)
> > +{
> > + struct perf_drv_config *drv_config = perf_event_get_drv_config(event);
> > +
> > + if (!has_drv_config(event))
> > + return;
> > +
> > + raw_spin_lock(&drv_config->lock);
> > + if (event->drv_config_gen != event->hw.drv_config_gen) {
> > + event->pmu->drv_config_sync(event);
> > + event->hw.drv_config_gen = event->drv_config_gen;
> > + }
> > + raw_spin_unlock(&drv_config->lock);
> > +}
> > +EXPORT_SYMBOL_GPL(perf_event_drv_config_sync);
> > +
> > static int _perf_event_refresh(struct perf_event *event, int refresh)
> > {
> > /*
> > @@ -4410,6 +4433,7 @@ static bool exclusive_event_installable(struct perf_event *event,
> >
> > static void perf_addr_filters_splice(struct perf_event *event,
> > struct list_head *head);
> > +static void perf_drv_config_splice(struct perf_event *event, void *drv_data);
> >
> > static void _free_event(struct perf_event *event)
> > {
> > @@ -4440,6 +4464,7 @@ static void _free_event(struct perf_event *event)
> > perf_event_free_bpf_prog(event);
> > perf_addr_filters_splice(event, NULL);
> > kfree(event->addr_filters_offs);
> > + perf_drv_config_splice(event, NULL);
> >
> > if (event->destroy)
> > event->destroy(event);
> > @@ -5002,6 +5027,8 @@ static inline int perf_fget_light(int fd, struct fd *p)
> > static int perf_event_set_output(struct perf_event *event,
> > struct perf_event *output_event);
> > static int perf_event_set_filter(struct perf_event *event, void __user *arg);
> > +static int perf_event_set_drv_config(struct perf_event *event,
> > + void __user *arg);
> > static int perf_event_set_bpf_prog(struct perf_event *event, u32 prog_fd);
> > static int perf_copy_attr(struct perf_event_attr __user *uattr,
> > struct perf_event_attr *attr);
> > @@ -5088,6 +5115,10 @@ static long _perf_ioctl(struct perf_event *event, unsigned int cmd, unsigned lon
> >
> > return perf_event_modify_attr(event, &new_attr);
> > }
> > +
> > + case PERF_EVENT_IOC_SET_DRV_CONFIG:
> > + return perf_event_set_drv_config(event, (void __user *)arg);
> > +
> > default:
> > return -ENOTTY;
> > }
> > @@ -9086,6 +9117,85 @@ static int perf_event_set_filter(struct perf_event *event, void __user *arg)
> > return ret;
> > }
> >
> > +static void perf_drv_config_splice(struct perf_event *event, void *drv_data)
>
> I think the address filter counterpart is called "splice" because it takes
> a list_head as a parameter and splices that list into the list of filters.
> I'd suggest that this one is more like "replace", but up to you.
I was on the fence about the naming convention... I wanted the
drv_config mechanism to be as close as possible to the filters, that
way if someone understands the filters they'll easily understand the
drv_config. But it may be more confusing than anything else - I'll
change it.
>
> > +{
> > + unsigned long flags;
> > + void *old_drv_data;
> > +
> > + if (!has_drv_config(event))
> > + return;
> > +
> > + /* Children take their configuration from their parent */
> > + if (event->parent)
> > + return;
> > +
> > + raw_spin_lock_irqsave(&event->drv_config.lock, flags);
> > +
> > + old_drv_data = event->drv_config.drv_config;
> > + event->drv_config.drv_config = drv_data;
>
> Now I'm thinking: should we reset the generation here (and also in the
> address filters bit)? At least, it deserves a comment.
I thought your way of doing things was quite nice, hence doing the
same... xyz_splice() does exactly that, it replaces the value but
downstream won't see it for as long as xyz_gen isn't updated - and
that is exactly what xyz_apply() does. I can add a comment to clarify
how things work but I think we should keep the current scheme.
>
> > +
> > + raw_spin_unlock_irqrestore(&event->drv_config.lock, flags);
> > +
> > + event->pmu->drv_config_free(old_drv_data);
> > +}
> > +
> > +static void perf_event_drv_config_apply(struct perf_event *event)
> > +{
> > + unsigned long flags;
> > + struct perf_drv_config *drv_config = perf_event_get_drv_config(event);
> > +
> > + /* Notify event that a new configuration is available */
> > + raw_spin_lock_irqsave(&drv_config->lock, flags);
> > + event->drv_config_gen++;
> > + raw_spin_unlock_irqrestore(&drv_config->lock, flags);
>
> Should we also mention how this new locks fits into the existing locking
> order?
>
> Regards,
> --
> Alex
On Tue, 3 Jul 2018 at 04:57, Alexander Shishkin
<[email protected]> wrote:
>
> On Tue, Jul 03, 2018 at 01:03:48PM +0300, Alexander Shishkin wrote:
> > On Mon, Jul 02, 2018 at 04:33:29PM -0600, Mathieu Poirier wrote:
> > > +/*
> > > + * PMU driver configuration works the same way as filter management above,
> > > + * but without the need to deal with memory mapping. Driver configuration
> > > + * arrives through the SET_DRV_CONFIG ioctl() where it is validated and applied
> > > + * to the event. When the PMU is ready it calls perf_event_drv_config_sync() to
> > > + * bring the configuration information within reach of the PMU.
> >
> > Wait a second. The reason why we dance around with the generations of filters
> > is the locking order of ctx::mutex vs mmap_sem. In an mmap path, where we're
> > notified about mapping changes, we're called under the latter, and we'd need
> > to grab the former to update the event configuration. In your case, the
> > update comes in via perf_ioctl(), where we're already holding the ctx::mutex,
> > so you can just kick the PMU right there, via an event_function_call() or
> > perf_event_stop(restart=1). In the latter case, your pmu::start() would just
> > grab the new configuration. Should also be about 90% less code. :)
>
> Also, since it affects the AUX buffer configuration, it is probably a one
> time ioctl command that you issue before you mmap the buffer. If that's
> the case, you don't even have to worry about stopping the event, as it
> shouldn't be running, because without the buffer perf_aux_output_begin()
> should fail and so should the pmu::add() iirc.
>
The idea behind the current approach was to make the SET_DRV_CONFIG
ioctl() usable by other drivers where multiple ioctl() calls could be
performed while a session in ongoing. I also opted to introduce a
_sync() function to let the PMU refresh its configuration at the time
of its own choosing rather than having to interrupt the session.
But all I need for coresight is to have available the sink information
and PMU configuration (in an upcoming patchset) by the time
setup_aux() is called. You are correct, this is a one time
configuration and since the event isn't running there is no need for
locking - I should be able to access the PMU when the ioctl is called.
If you are fine with this bare-bone scenario and don't care much about
usability in different situation, I'll do a respin with minimal
functionality that cover my needs.
Thanks for the review,
Mathieu
> Regards,
> --
> Alex
Mathieu Poirier <[email protected]> writes:
> On Tue, 3 Jul 2018 at 04:57, Alexander Shishkin
> <[email protected]> wrote:
>>
>> On Tue, Jul 03, 2018 at 01:03:48PM +0300, Alexander Shishkin wrote:
>> > On Mon, Jul 02, 2018 at 04:33:29PM -0600, Mathieu Poirier wrote:
>> > > +/*
>> > > + * PMU driver configuration works the same way as filter management above,
>> > > + * but without the need to deal with memory mapping. Driver configuration
>> > > + * arrives through the SET_DRV_CONFIG ioctl() where it is validated and applied
>> > > + * to the event. When the PMU is ready it calls perf_event_drv_config_sync() to
>> > > + * bring the configuration information within reach of the PMU.
>> >
>> > Wait a second. The reason why we dance around with the generations of filters
>> > is the locking order of ctx::mutex vs mmap_sem. In an mmap path, where we're
>> > notified about mapping changes, we're called under the latter, and we'd need
>> > to grab the former to update the event configuration. In your case, the
>> > update comes in via perf_ioctl(), where we're already holding the ctx::mutex,
>> > so you can just kick the PMU right there, via an event_function_call() or
>> > perf_event_stop(restart=1). In the latter case, your pmu::start() would just
>> > grab the new configuration. Should also be about 90% less code. :)
>>
>> Also, since it affects the AUX buffer configuration, it is probably a one
>> time ioctl command that you issue before you mmap the buffer. If that's
>> the case, you don't even have to worry about stopping the event, as it
>> shouldn't be running, because without the buffer perf_aux_output_begin()
>> should fail and so should the pmu::add() iirc.
>>
>
> The idea behind the current approach was to make the SET_DRV_CONFIG
> ioctl() usable by other drivers where multiple ioctl() calls could be
> performed while a session in ongoing. I also opted to introduce a
> _sync() function to let the PMU refresh its configuration at the time
> of its own choosing rather than having to interrupt the session.
Yes, but the times of PMU's own choosing would still be more or less
limited to ->start()/->stop().
You can also do an event_function_call(), which would call
->config_sync(), which would be free to decide what to do with the new
information, up to and including doing a ->stop()/->start() sequence. My
guess is that you'd want to do either of the following:
* decide to apply the new configuration immediately, and do the
start-stop thing,
* decide to defer the new configuration until the next ->start().
Both should work via cross call directly from the ioctl() call.
> But all I need for coresight is to have available the sink information
> and PMU configuration (in an upcoming patchset) by the time
> setup_aux() is called. You are correct, this is a one time
> configuration and since the event isn't running there is no need for
> locking - I should be able to access the PMU when the ioctl is called.
>
> If you are fine with this bare-bone scenario and don't care much about
> usability in different situation, I'll do a respin with minimal
> functionality that cover my needs.
It doesn't have to be bare-bones, what I'm saying is that you shouldn't
need the event->drv_config, as you can directly call
pmu::config(new_config) (or config_sync(), but I'm guessing the _sync
part is redundant if you don't keep the configuration in 2 parts) from
the ioctl() and it should cover all your bases.
Regards,
--
Alex
Mathieu Poirier <[email protected]> writes:
> Hi Alex,
Hi Mathieu,
> On Tue, 3 Jul 2018 at 04:03, Alexander Shishkin
> <[email protected]> wrote:
>>
>> On Mon, Jul 02, 2018 at 04:33:29PM -0600, Mathieu Poirier wrote:
>> > This patch follows what has been done for filters by adding an ioctl()
>> > option to communicate to the kernel arbitrary PMU specific configuration
>> > that don't fit in the conventional struct perf_event_attr to the kernel.
>>
>> Ok, so what *is* the PMU specific configuration that doesn't fit in the
>> attribute and needs to be re-configured by the driver using the generation
>> tracking?
>>
>
> In this patchset I'm am after the specification of sink information
> for each event, i.e what sink a CPU is supposed to use for the
> session. I simply don't see putting something that PMU specific in
> the generic perf_event_attr structure. I also intend to use the same
> ioctl mechanism to communicate complex tracer configuration for
> sequencers, counters and input events. I don't see a nice way of
> doing that from the perf_event_attr, and that is even without thinking
> about the different flavours of tracers out there, all with their own
> features.
Yes, the sequencers and counters seem tricky. Here's a wild idea: can
the sequencer/counter configuration be expressed as an eBPF program? Or,
can an eBPF program be used to program those?
> I've looked around and the only clean way I found to support this is
> via an ioctl(). That way each tracer can easily identify the sink it
> should be using without smearing the perf_event_attr structure. I
> would be happy to explore a different avenue should you think of
> something.
Yes, I also have something similar on my todo list and I was previously
thinking along the lines of pipe()/splice(). As in, you take the AUX
event file descriptor and feed it to the sink, at which point the trace
path is configured. I need to dig up the notes that I made back in the
day to continue this conversation in more concrete terms.
Regards,
--
Alex
On Wed, 4 Jul 2018 at 04:35, Alexander Shishkin
<[email protected]> wrote:
>
> Mathieu Poirier <[email protected]> writes:
>
> > On Tue, 3 Jul 2018 at 04:57, Alexander Shishkin
> > <[email protected]> wrote:
> >>
> >> On Tue, Jul 03, 2018 at 01:03:48PM +0300, Alexander Shishkin wrote:
> >> > On Mon, Jul 02, 2018 at 04:33:29PM -0600, Mathieu Poirier wrote:
> >> > > +/*
> >> > > + * PMU driver configuration works the same way as filter management above,
> >> > > + * but without the need to deal with memory mapping. Driver configuration
> >> > > + * arrives through the SET_DRV_CONFIG ioctl() where it is validated and applied
> >> > > + * to the event. When the PMU is ready it calls perf_event_drv_config_sync() to
> >> > > + * bring the configuration information within reach of the PMU.
> >> >
> >> > Wait a second. The reason why we dance around with the generations of filters
> >> > is the locking order of ctx::mutex vs mmap_sem. In an mmap path, where we're
> >> > notified about mapping changes, we're called under the latter, and we'd need
> >> > to grab the former to update the event configuration. In your case, the
> >> > update comes in via perf_ioctl(), where we're already holding the ctx::mutex,
> >> > so you can just kick the PMU right there, via an event_function_call() or
> >> > perf_event_stop(restart=1). In the latter case, your pmu::start() would just
> >> > grab the new configuration. Should also be about 90% less code. :)
> >>
> >> Also, since it affects the AUX buffer configuration, it is probably a one
> >> time ioctl command that you issue before you mmap the buffer. If that's
> >> the case, you don't even have to worry about stopping the event, as it
> >> shouldn't be running, because without the buffer perf_aux_output_begin()
> >> should fail and so should the pmu::add() iirc.
> >>
> >
> > The idea behind the current approach was to make the SET_DRV_CONFIG
> > ioctl() usable by other drivers where multiple ioctl() calls could be
> > performed while a session in ongoing. I also opted to introduce a
> > _sync() function to let the PMU refresh its configuration at the time
> > of its own choosing rather than having to interrupt the session.
>
> Yes, but the times of PMU's own choosing would still be more or less
> limited to ->start()/->stop().
>
> You can also do an event_function_call(), which would call
> ->config_sync(), which would be free to decide what to do with the new
> information, up to and including doing a ->stop()/->start() sequence. My
> guess is that you'd want to do either of the following:
> * decide to apply the new configuration immediately, and do the
> start-stop thing,
> * decide to defer the new configuration until the next ->start().
>
> Both should work via cross call directly from the ioctl() call.
Other than doing a cross call, do you see any advantage of using
event_function_call()? For what I currently need the cross call is
not necessary.
>
> > But all I need for coresight is to have available the sink information
> > and PMU configuration (in an upcoming patchset) by the time
> > setup_aux() is called. You are correct, this is a one time
> > configuration and since the event isn't running there is no need for
> > locking - I should be able to access the PMU when the ioctl is called.
> >
> > If you are fine with this bare-bone scenario and don't care much about
> > usability in different situation, I'll do a respin with minimal
> > functionality that cover my needs.
>
> It doesn't have to be bare-bones, what I'm saying is that you shouldn't
> need the event->drv_config, as you can directly call
> pmu::config(new_config) (or config_sync(), but I'm guessing the _sync
> part is redundant if you don't keep the configuration in 2 parts) from
> the ioctl() and it should cover all your bases.
I think I get your point - I will do another respin and we can go from there.
Thanks,
Mathieu
>
> Regards,
> --
> Alex