@@ -159,6 +159,19 @@ u64 riscv_pmu_ctr_get_width_mask(struct perf_event *event)
return GENMASK_ULL(cwidth, 0);
}
+static void riscv_pmu_sched_task(struct perf_event_pmu_context *pmu_ctx,
+ bool sched_in)
+{
+ struct riscv_pmu *pmu;
+
+ if (!pmu_ctx)
+ return;
+
+ pmu = to_riscv_pmu(pmu_ctx->pmu);
+ if (pmu->sched_task)
+ pmu->sched_task(pmu_ctx, sched_in);
+}
+
u64 riscv_pmu_event_update(struct perf_event *event)
{
struct riscv_pmu *rvpmu = to_riscv_pmu(event->pmu);
@@ -406,6 +419,7 @@ struct riscv_pmu *riscv_pmu_alloc(void)
for_each_possible_cpu(cpuid) {
cpuc = per_cpu_ptr(pmu->hw_events, cpuid);
cpuc->n_events = 0;
+ cpuc->ctr_users = 0;
for (i = 0; i < RISCV_MAX_COUNTERS; i++)
cpuc->events[i] = NULL;
}
@@ -419,6 +433,7 @@ struct riscv_pmu *riscv_pmu_alloc(void)
.start = riscv_pmu_start,
.stop = riscv_pmu_stop,
.read = riscv_pmu_read,
+ .sched_task = riscv_pmu_sched_task,
};
return pmu;
@@ -670,6 +670,14 @@ static void rvpmu_sbi_ctr_stop(struct perf_event *event, unsigned long flag)
hwc->idx, sbi_err_map_linux_errno(ret.error));
}
+static void pmu_sched_task(struct perf_event_pmu_context *pmu_ctx,
+ bool sched_in)
+{
+ struct riscv_pmu *pmu = to_riscv_pmu(pmu_ctx->pmu);
+
+ /* Call CTR specific Sched hook. */
+}
+
static int rvpmu_sbi_find_num_ctrs(void)
{
struct sbiret ret;
@@ -1494,6 +1502,7 @@ static int rvpmu_device_probe(struct platform_device *pdev)
pmu->event_mapped = rvpmu_event_mapped;
pmu->event_unmapped = rvpmu_event_unmapped;
pmu->csr_index = rvpmu_csr_index;
+ pmu->sched_task = pmu_sched_task;
ret = cpuhp_state_add_instance(CPUHP_AP_PERF_RISCV_STARTING, &pmu->node);
if (ret)
@@ -33,6 +33,13 @@
#define RISCV_PMU_CYCLE_FIXED_CTR_MASK 0x01
#define RISCV_PMU_INSTRUCTION_FIXED_CTR_MASK 0x04
+#define MAX_BRANCH_RECORDS 256
+
+struct branch_records {
+ struct perf_branch_stack branch_stack;
+ struct perf_branch_entry branch_entries[MAX_BRANCH_RECORDS];
+};
+
struct cpu_hw_events {
/* currently enabled events */
int n_events;
@@ -44,6 +51,12 @@ struct cpu_hw_events {
DECLARE_BITMAP(used_hw_ctrs, RISCV_MAX_COUNTERS);
/* currently enabled firmware counters */
DECLARE_BITMAP(used_fw_ctrs, RISCV_MAX_COUNTERS);
+
+ /* Saved branch records. */
+ struct branch_records *branches;
+
+ /* Active events requesting branch records */
+ int ctr_users;
};
struct riscv_pmu {
@@ -64,10 +77,13 @@ struct riscv_pmu {
void (*event_mapped)(struct perf_event *event, struct mm_struct *mm);
void (*event_unmapped)(struct perf_event *event, struct mm_struct *mm);
uint8_t (*csr_index)(struct perf_event *event);
+ void (*sched_task)(struct perf_event_pmu_context *ctx, bool sched_in);
struct cpu_hw_events __percpu *hw_events;
struct hlist_node node;
struct notifier_block riscv_pm_nb;
+
+ unsigned int ctr_depth;
};
#define to_riscv_pmu(p) (container_of(p, struct riscv_pmu, pmu))
To support Control Transfer Records (CTR) extension, we need to extend the riscv_pmu framework with some basic infrastructure for branch stack sampling. Subsequent patches will use this to add support for CTR in the riscv_pmu_dev driver. With CTR, the branches are stored into a hardware FIFO, which will be sampled by software when perf events overflow. A task may be context- switched between overflows, and to avoid leaking samples we need to clear the last task's records when a task is context-switched In. To do this we will be using the pmu::sched_task() callback added in this patch. Signed-off-by: Rajnesh Kanwal <rkanwal@rivosinc.com> --- drivers/perf/riscv_pmu_common.c | 15 +++++++++++++++ drivers/perf/riscv_pmu_dev.c | 9 +++++++++ include/linux/perf/riscv_pmu.h | 16 ++++++++++++++++ 3 files changed, 40 insertions(+)