gpu: nvgpu: using pmu counters for load estimate

PMU counters #0 and #4 are used to count total cycles and busy cycles.
These counts are used by podgov to estimate GPU load.

PMU idle intr status register is used to monitor overflow. Overflow
rarely occurs because frequency governor reads and resets the counters
at a high cadence. When overflow occurs, 100% work load is reported to
frequency governor.

Bug 1963732

Change-Id: I046480ebde162e6eda24577932b96cfd91b77c69
Signed-off-by: Peng Liu <pengliu@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1939547
(cherry picked from commit 34df0035194e0203f68f679acdd84e5533a48149)
Reviewed-on: https://git-master.nvidia.com/r/1979495
Reviewed-by: Aaron Tian <atian@nvidia.com>
Tested-by: Aaron Tian <atian@nvidia.com>
Reviewed-by: Rajkumar Kasirajan <rkasirajan@nvidia.com>
Tested-by: Rajkumar Kasirajan <rkasirajan@nvidia.com>
Reviewed-by: Bibek Basu <bbasu@nvidia.com>
GVS: Gerrit_Virtual_Submit
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>

1 files changed, 42 insertions, 0 deletions

diff --git a/drivers/gpu/nvgpu/common/pmu/pmu_perfmon.c b/drivers/gpu/nvgpu/common/pmu/pmu_perfmon.c
index 57a4ea40..bf07bd79 100644
--- a/drivers/gpu/nvgpu/common/pmu/pmu_perfmon.c
+++ b/drivers/gpu/nvgpu/common/pmu/pmu_perfmon.c
@@ -236,6 +236,48 @@ int nvgpu_pmu_load_update(struct gk20a *g)
         return 0;
 }
 
+int nvgpu_pmu_busy_cycles_norm(struct gk20a *g, u32 *norm)
+{
+        u64 busy_cycles, total_cycles;
+        u32 intr_status;
+
+        gk20a_busy_noresume(g);
+        if (!g->power_on) {
+                *norm = 0;
+                goto exit;
+        }
+
+        if (g->ops.pmu.pmu_read_idle_counter == NULL ||
+            g->ops.pmu.pmu_reset_idle_counter == NULL ||
+            g->ops.pmu.pmu_read_idle_intr_status == NULL ||
+            g->ops.pmu.pmu_clear_idle_intr_status == NULL) {
+                *norm = PMU_BUSY_CYCLES_NORM_MAX;
+                goto exit;
+        }
+
+        busy_cycles = g->ops.pmu.pmu_read_idle_counter(g, 4);
+        total_cycles = g->ops.pmu.pmu_read_idle_counter(g, 0);
+        intr_status = g->ops.pmu.pmu_read_idle_intr_status(g);
+
+        g->ops.pmu.pmu_reset_idle_counter(g, 4);
+        g->ops.pmu.pmu_reset_idle_counter(g, 0);
+
+        if (intr_status != 0UL) {
+                *norm = PMU_BUSY_CYCLES_NORM_MAX;
+                g->ops.pmu.pmu_clear_idle_intr_status(g);
+        } else if (total_cycles == 0ULL || busy_cycles > total_cycles) {
+                *norm = PMU_BUSY_CYCLES_NORM_MAX;
+        } else {
+                *norm = (u32)(busy_cycles * PMU_BUSY_CYCLES_NORM_MAX
+                              / total_cycles);
+        }
+
+exit:
+        gk20a_idle_nosuspend(g);
+
+        return 0;
+}
+
 void nvgpu_pmu_get_load_counters(struct gk20a *g, u32 *busy_cycles,
                                  u32 *total_cycles)
 {