[PATCH v2 2/5] powerpc/perf: Implement a global lock to avoid races between trace, core and thread imc events.
Anju T Sudhakar
anju at linux.vnet.ibm.com
Tue Jan 21 21:17:25 AEDT 2020
IMC(In-memory Collection Counters) does performance monitoring in
two different modes, i.e accumulation mode(core-imc and thread-imc events),
and trace mode(trace-imc events). A cpu thread can either be in
accumulation-mode or trace-mode at a time and this is done via the LDBAR
register in POWER architecture. The current design does not address the
races between thread-imc and trace-imc events.
Patch implements a global id and lock to avoid the races between
core, trace and thread imc events. With this global id-lock
implementation, the system can either run core, thread or trace imc
events at a time. i.e. to run any core-imc events, thread/trace imc events
should not be enabled/monitored.
Signed-off-by: Anju T Sudhakar <anju at linux.vnet.ibm.com>
---
arch/powerpc/perf/imc-pmu.c | 177 +++++++++++++++++++++++++++++++-----
1 file changed, 153 insertions(+), 24 deletions(-)
diff --git a/arch/powerpc/perf/imc-pmu.c b/arch/powerpc/perf/imc-pmu.c
index cb50a9e1fd2d..2e220f199530 100644
--- a/arch/powerpc/perf/imc-pmu.c
+++ b/arch/powerpc/perf/imc-pmu.c
@@ -44,6 +44,16 @@ static DEFINE_PER_CPU(u64 *, trace_imc_mem);
static struct imc_pmu_ref *trace_imc_refc;
static int trace_imc_mem_size;
+/*
+ * Global data structure used to avoid races between thread,
+ * core and trace-imc
+ */
+static struct imc_pmu_ref imc_global_refc = {
+ .lock = __MUTEX_INITIALIZER(imc_global_refc.lock),
+ .id = 0,
+ .refc = 0,
+};
+
static struct imc_pmu *imc_event_to_pmu(struct perf_event *event)
{
return container_of(event->pmu, struct imc_pmu, pmu);
@@ -759,6 +769,20 @@ static void core_imc_counters_release(struct perf_event *event)
ref->refc = 0;
}
mutex_unlock(&ref->lock);
+
+ mutex_lock(&imc_global_refc.lock);
+ if (imc_global_refc.id == IMC_DOMAIN_CORE) {
+ imc_global_refc.refc--;
+ /*
+ * If no other thread is running any core-imc
+ * event, set the global id to zero.
+ */
+ if (imc_global_refc.refc <= 0) {
+ imc_global_refc.refc = 0;
+ imc_global_refc.id = 0;
+ }
+ }
+ mutex_unlock(&imc_global_refc.lock);
}
static int core_imc_event_init(struct perf_event *event)
@@ -779,6 +803,22 @@ static int core_imc_event_init(struct perf_event *event)
if (event->cpu < 0)
return -EINVAL;
+ /*
+ * Take the global lock, and make sure
+ * no other thread is running any trace OR thread imc event
+ */
+ mutex_lock(&imc_global_refc.lock);
+ if (imc_global_refc.id == 0) {
+ imc_global_refc.id = IMC_DOMAIN_CORE;
+ imc_global_refc.refc++;
+ } else if (imc_global_refc.id == IMC_DOMAIN_CORE) {
+ imc_global_refc.refc++;
+ } else {
+ mutex_unlock(&imc_global_refc.lock);
+ return -EBUSY;
+ }
+ mutex_unlock(&imc_global_refc.lock);
+
event->hw.idx = -1;
pmu = imc_event_to_pmu(event);
@@ -877,7 +917,16 @@ static int ppc_thread_imc_cpu_online(unsigned int cpu)
static int ppc_thread_imc_cpu_offline(unsigned int cpu)
{
- mtspr(SPRN_LDBAR, 0);
+ /*
+ * Toggle the bit 0 of LDBAR.
+ *
+ * If bit 0 of LDBAR is unset, it will stop posting
+ * the counetr data to memory.
+ * For thread-imc, bit 0 of LDBAR will be set to 1 in the
+ * event_add function. So toggle this bit here, to stop the updates
+ * to memory in the cpu_offline path.
+ */
+ mtspr(SPRN_LDBAR, (mfspr(SPRN_LDBAR) ^ (1UL << 63)));
return 0;
}
@@ -889,6 +938,24 @@ static int thread_imc_cpu_init(void)
ppc_thread_imc_cpu_offline);
}
+static void thread_imc_counters_release(struct perf_event *event)
+{
+
+ mutex_lock(&imc_global_refc.lock);
+ if (imc_global_refc.id == IMC_DOMAIN_THREAD) {
+ imc_global_refc.refc--;
+ /*
+ * If no other thread is running any thread-imc
+ * event, set the global id to zero.
+ */
+ if (imc_global_refc.refc <= 0) {
+ imc_global_refc.refc = 0;
+ imc_global_refc.id = 0;
+ }
+ }
+ mutex_unlock(&imc_global_refc.lock);
+}
+
static int thread_imc_event_init(struct perf_event *event)
{
u32 config = event->attr.config;
@@ -905,6 +972,27 @@ static int thread_imc_event_init(struct perf_event *event)
if (event->hw.sample_period)
return -EINVAL;
+ mutex_lock(&imc_global_refc.lock);
+ /*
+ * Check if any other thread is running
+ * core-engine, if not set the global id to
+ * thread-imc.
+ */
+ if (imc_global_refc.id == 0) {
+ imc_global_refc.id = IMC_DOMAIN_THREAD;
+ imc_global_refc.refc++;
+ } else if (imc_global_refc.id == IMC_DOMAIN_THREAD) {
+ /*
+ * Increase the ref count if the global id is
+ * set to thread-imc.
+ */
+ imc_global_refc.refc++;
+ } else {
+ mutex_unlock(&imc_global_refc.lock);
+ return -EBUSY;
+ }
+ mutex_unlock(&imc_global_refc.lock);
+
event->hw.idx = -1;
pmu = imc_event_to_pmu(event);
@@ -917,6 +1005,7 @@ static int thread_imc_event_init(struct perf_event *event)
return -EINVAL;
event->pmu->task_ctx_nr = perf_sw_context;
+ event->destroy = thread_imc_counters_release;
return 0;
}
@@ -1063,10 +1152,12 @@ static void thread_imc_event_del(struct perf_event *event, int flags)
int core_id;
struct imc_pmu_ref *ref;
- mtspr(SPRN_LDBAR, 0);
-
core_id = smp_processor_id() / threads_per_core;
ref = &core_imc_refc[core_id];
+ if (!ref) {
+ pr_debug("imc: Failed to get event reference count\n");
+ return;
+ }
mutex_lock(&ref->lock);
ref->refc--;
@@ -1082,6 +1173,10 @@ static void thread_imc_event_del(struct perf_event *event, int flags)
ref->refc = 0;
}
mutex_unlock(&ref->lock);
+
+ /* Toggle bit 0 of LDBAR */
+ mtspr(SPRN_LDBAR, (mfspr(SPRN_LDBAR) ^ (1UL << 63)));
+
/*
* Take a snapshot and calculate the delta and update
* the event counter values.
@@ -1133,7 +1228,8 @@ static int ppc_trace_imc_cpu_online(unsigned int cpu)
static int ppc_trace_imc_cpu_offline(unsigned int cpu)
{
- mtspr(SPRN_LDBAR, 0);
+ /* Toggle bit 0 of LDBAR. */
+ mtspr(SPRN_LDBAR, (mfspr(SPRN_LDBAR) ^ (1UL << 63)));
return 0;
}
@@ -1226,15 +1322,14 @@ static int trace_imc_event_add(struct perf_event *event, int flags)
local_mem = get_trace_imc_event_base_addr();
ldbar_value = ((u64)local_mem & THREAD_IMC_LDBAR_MASK) | TRACE_IMC_ENABLE;
- if (core_imc_refc)
- ref = &core_imc_refc[core_id];
+ /* trace-imc reference count */
+ if (trace_imc_refc)
+ ref = &trace_imc_refc[core_id];
if (!ref) {
- /* If core-imc is not enabled, use trace-imc reference count */
- if (trace_imc_refc)
- ref = &trace_imc_refc[core_id];
- if (!ref)
- return -EINVAL;
+ pr_debug("imc: Failed to get the event reference count\n");
+ return -EINVAL;
}
+
mtspr(SPRN_LDBAR, ldbar_value);
mutex_lock(&ref->lock);
if (ref->refc == 0) {
@@ -1242,13 +1337,11 @@ static int trace_imc_event_add(struct perf_event *event, int flags)
get_hard_smp_processor_id(smp_processor_id()))) {
mutex_unlock(&ref->lock);
pr_err("trace-imc: Unable to start the counters for core %d\n", core_id);
- mtspr(SPRN_LDBAR, 0);
return -EINVAL;
}
}
++ref->refc;
mutex_unlock(&ref->lock);
-
return 0;
}
@@ -1274,16 +1367,13 @@ static void trace_imc_event_del(struct perf_event *event, int flags)
int core_id = smp_processor_id() / threads_per_core;
struct imc_pmu_ref *ref = NULL;
- if (core_imc_refc)
- ref = &core_imc_refc[core_id];
+ if (trace_imc_refc)
+ ref = &trace_imc_refc[core_id];
if (!ref) {
- /* If core-imc is not enabled, use trace-imc reference count */
- if (trace_imc_refc)
- ref = &trace_imc_refc[core_id];
- if (!ref)
- return;
+ pr_debug("imc: Failed to get event reference count\n");
+ return;
}
- mtspr(SPRN_LDBAR, 0);
+
mutex_lock(&ref->lock);
ref->refc--;
if (ref->refc == 0) {
@@ -1297,9 +1387,30 @@ static void trace_imc_event_del(struct perf_event *event, int flags)
ref->refc = 0;
}
mutex_unlock(&ref->lock);
+
+ /* Toggle bit 0 of LDBAR */
+ mtspr(SPRN_LDBAR, (mfspr(SPRN_LDBAR) ^ (1UL << 63)));
+
trace_imc_event_stop(event, flags);
}
+static void trace_imc_counters_release(struct perf_event *event)
+{
+ mutex_lock(&imc_global_refc.lock);
+ if (imc_global_refc.id == IMC_DOMAIN_TRACE) {
+ imc_global_refc.refc--;
+ /*
+ * If no other thread is running any trace-imc
+ * event, set the global id to zero.
+ */
+ if (imc_global_refc.refc <= 0) {
+ imc_global_refc.refc = 0;
+ imc_global_refc.id = 0;
+ }
+ }
+ mutex_unlock(&imc_global_refc.lock);
+}
+
static int trace_imc_event_init(struct perf_event *event)
{
struct task_struct *target;
@@ -1314,10 +1425,28 @@ static int trace_imc_event_init(struct perf_event *event)
if (event->attr.sample_period == 0)
return -ENOENT;
+ /*
+ * Take the global lock, and make sure
+ * no other thread is running any core/thread imc
+ * event
+ */
+ mutex_lock(&imc_global_refc.lock);
+ if (imc_global_refc.id == 0) {
+ imc_global_refc.id = IMC_DOMAIN_TRACE;
+ imc_global_refc.refc++;
+ } else if (imc_global_refc.id == IMC_DOMAIN_TRACE) {
+ imc_global_refc.refc++;
+ } else {
+ mutex_unlock(&imc_global_refc.lock);
+ return -EBUSY;
+ }
+ mutex_unlock(&imc_global_refc.lock);
+
event->hw.idx = -1;
target = event->hw.target;
event->pmu->task_ctx_nr = perf_hw_context;
+ event->destroy = trace_imc_counters_release;
return 0;
}
@@ -1429,10 +1558,10 @@ static void cleanup_all_core_imc_memory(void)
static void thread_imc_ldbar_disable(void *dummy)
{
/*
- * By Zeroing LDBAR, we disable thread-imc
- * updates.
+ * By toggling 0th bit of LDBAR, we disable thread-imc
+ * updates to memory.
*/
- mtspr(SPRN_LDBAR, 0);
+ mtspr(SPRN_LDBAR, (mfspr(SPRN_LDBAR) ^ (1UL << 63)));
}
void thread_imc_disable(void)
--
2.20.1
More information about the Linuxppc-dev
mailing list