diff --git a/drivers/gpu/nvgpu/common/fifo/engines.c b/drivers/gpu/nvgpu/common/fifo/engines.c
index 7a7eb994d..2da9747ac 100644
--- a/drivers/gpu/nvgpu/common/fifo/engines.c
+++ b/drivers/gpu/nvgpu/common/fifo/engines.c
@@ -26,6 +26,7 @@
 #include <nvgpu/timers.h>
 #include <nvgpu/bitops.h>
 #include <nvgpu/pmu.h>
+#include <nvgpu/pmu/mutex.h>
 #include <nvgpu/runlist.h>
 #include <nvgpu/engines.h>
 #include <nvgpu/engine_status.h>
diff --git a/drivers/gpu/nvgpu/common/fifo/runlist.c b/drivers/gpu/nvgpu/common/fifo/runlist.c
index a243e933d..4eac0c8b6 100644
--- a/drivers/gpu/nvgpu/common/fifo/runlist.c
+++ b/drivers/gpu/nvgpu/common/fifo/runlist.c
@@ -29,6 +29,7 @@
 #include <nvgpu/bug.h>
 #include <nvgpu/dma.h>
 #include <nvgpu/rc.h>
+#include <nvgpu/pmu/mutex.h>
 
 void nvgpu_runlist_lock_active_runlists(struct gk20a *g)
 {
diff --git a/drivers/gpu/nvgpu/common/gr/gr_falcon.c b/drivers/gpu/nvgpu/common/gr/gr_falcon.c
index 80140e750..2c1e60fb1 100644
--- a/drivers/gpu/nvgpu/common/gr/gr_falcon.c
+++ b/drivers/gpu/nvgpu/common/gr/gr_falcon.c
@@ -31,6 +31,7 @@
 #include <nvgpu/mm.h>
 #include <nvgpu/acr.h>
 #include <nvgpu/pmu/lsfm.h>
+#include <nvgpu/pmu/pmu_pg.h>
 #include <nvgpu/sec2/lsfm.h>
 #include <nvgpu/dma.h>
 #include <nvgpu/safe_ops.h>
diff --git a/drivers/gpu/nvgpu/common/pmu/fw/fw.c b/drivers/gpu/nvgpu/common/pmu/fw/fw.c
index fc2529779..79f0d5b43 100644
--- a/drivers/gpu/nvgpu/common/pmu/fw/fw.c
+++ b/drivers/gpu/nvgpu/common/pmu/fw/fw.c
@@ -30,6 +30,7 @@
 #include <nvgpu/gk20a.h>
 #include <nvgpu/pmu/allocator.h>
 #include <nvgpu/pmu/fw.h>
+#include <nvgpu/pmu/pmu_pg.h>
 
 /* PMU NS UCODE IMG */
 #define NVGPU_PMU_NS_UCODE_IMAGE	"gpmu_ucode.bin"
diff --git a/drivers/gpu/nvgpu/common/pmu/fw/fw_ver_ops.c b/drivers/gpu/nvgpu/common/pmu/fw/fw_ver_ops.c
index 544e2ea16..b145ddf93 100644
--- a/drivers/gpu/nvgpu/common/pmu/fw/fw_ver_ops.c
+++ b/drivers/gpu/nvgpu/common/pmu/fw/fw_ver_ops.c
@@ -38,6 +38,7 @@
 #include <nvgpu/pmu/lsfm.h>
 #include <nvgpu/pmu/super_surface.h>
 #include <nvgpu/pmu/fw.h>
+#include <nvgpu/pmu/seq.h>
 
 /* PMU F/W version */
 #define APP_VERSION_TU10X	25622342U
diff --git a/drivers/gpu/nvgpu/common/pmu/ipc/pmu_cmd.c b/drivers/gpu/nvgpu/common/pmu/ipc/pmu_cmd.c
index e9bd20530..9d00345aa 100644
--- a/drivers/gpu/nvgpu/common/pmu/ipc/pmu_cmd.c
+++ b/drivers/gpu/nvgpu/common/pmu/ipc/pmu_cmd.c
@@ -36,6 +36,7 @@
 #include <nvgpu/pmu/msg.h>
 #include <nvgpu/pmu/fw.h>
 #include <nvgpu/pmu/allocator.h>
+#include <nvgpu/pmu/seq.h>
 
 static bool pmu_validate_in_out_payload(struct nvgpu_pmu *pmu, struct pmu_cmd *cmd,
 					struct pmu_in_out_payload_desc *payload)
diff --git a/drivers/gpu/nvgpu/common/pmu/ipc/pmu_msg.c b/drivers/gpu/nvgpu/common/pmu/ipc/pmu_msg.c
index 23f2d2703..9b074d90e 100644
--- a/drivers/gpu/nvgpu/common/pmu/ipc/pmu_msg.c
+++ b/drivers/gpu/nvgpu/common/pmu/ipc/pmu_msg.c
@@ -30,6 +30,7 @@
 #include <nvgpu/pmu/super_surface.h>
 #include <nvgpu/pmu/pmu_perfmon.h>
 #include <nvgpu/pmu/fw.h>
+#include <nvgpu/pmu/seq.h>
 
 static int pmu_payload_extract(struct nvgpu_pmu *pmu, struct pmu_sequence *seq)
 {
diff --git a/drivers/gpu/nvgpu/common/pmu/lpwr/lpwr.c b/drivers/gpu/nvgpu/common/pmu/lpwr/lpwr.c
index fe34da20b..c0b1a18b7 100644
--- a/drivers/gpu/nvgpu/common/pmu/lpwr/lpwr.c
+++ b/drivers/gpu/nvgpu/common/pmu/lpwr/lpwr.c
@@ -28,6 +28,7 @@
 #include <nvgpu/pmu/perf.h>
 #include <nvgpu/pmu/cmd.h>
 #include <nvgpu/pmu/pmu_pstate.h>
+#include <nvgpu/pmu/pmu_pg.h>
 
 #include "common/vbios/bios_sw_gp106.h"
 #include "lpwr.h"
diff --git a/drivers/gpu/nvgpu/common/pmu/pg/pg_sw_gm20b.c b/drivers/gpu/nvgpu/common/pmu/pg/pg_sw_gm20b.c
index 11d0b3f86..4e943b47e 100644
--- a/drivers/gpu/nvgpu/common/pmu/pg/pg_sw_gm20b.c
+++ b/drivers/gpu/nvgpu/common/pmu/pg/pg_sw_gm20b.c
@@ -25,6 +25,7 @@
 #include <nvgpu/pmu/lpwr.h>
 #include <nvgpu/pmu/cmd.h>
 #include <nvgpu/bug.h>
+#include <nvgpu/pmu/pmu_pg.h>
 
 #include "pg_sw_gm20b.h"
 
diff --git a/drivers/gpu/nvgpu/common/pmu/pg/pg_sw_gp106.c b/drivers/gpu/nvgpu/common/pmu/pg/pg_sw_gp106.c
index b0a122565..42d601790 100644
--- a/drivers/gpu/nvgpu/common/pmu/pg/pg_sw_gp106.c
+++ b/drivers/gpu/nvgpu/common/pmu/pg/pg_sw_gp106.c
@@ -26,6 +26,7 @@
 #include <nvgpu/bug.h>
 #include <nvgpu/pmu/cmd.h>
 #include <nvgpu/clk_arb.h>
+#include <nvgpu/pmu/pmu_pg.h>
 
 #include "pg_sw_gp106.h"
 
diff --git a/drivers/gpu/nvgpu/common/pmu/pg/pg_sw_gp10b.c b/drivers/gpu/nvgpu/common/pmu/pg/pg_sw_gp10b.c
index b9beafbe4..d9e0540ee 100644
--- a/drivers/gpu/nvgpu/common/pmu/pg/pg_sw_gp10b.c
+++ b/drivers/gpu/nvgpu/common/pmu/pg/pg_sw_gp10b.c
@@ -25,6 +25,7 @@
 #include <nvgpu/gk20a.h>
 #include <nvgpu/bug.h>
 #include <nvgpu/pmu/cmd.h>
+#include <nvgpu/pmu/pmu_pg.h>
 
 #include "pg_sw_gp10b.h"
 #include "pg_sw_gm20b.h"
diff --git a/drivers/gpu/nvgpu/common/pmu/pg/pg_sw_gv11b.c b/drivers/gpu/nvgpu/common/pmu/pg/pg_sw_gv11b.c
index e7b83eda5..358279c01 100644
--- a/drivers/gpu/nvgpu/common/pmu/pg/pg_sw_gv11b.c
+++ b/drivers/gpu/nvgpu/common/pmu/pg/pg_sw_gv11b.c
@@ -24,6 +24,7 @@
 #include <nvgpu/gk20a.h>
 #include <nvgpu/bug.h>
 #include <nvgpu/pmu/cmd.h>
+#include <nvgpu/pmu/pmu_pg.h>
 
 #include "pg_sw_gv11b.h"
 #include "pg_sw_gp106.h"
diff --git a/drivers/gpu/nvgpu/common/pmu/pg/pmu_pg.c b/drivers/gpu/nvgpu/common/pmu/pg/pmu_pg.c
index 6db4a8fab..c0db115e9 100644
--- a/drivers/gpu/nvgpu/common/pmu/pg/pmu_pg.c
+++ b/drivers/gpu/nvgpu/common/pmu/pg/pmu_pg.c
@@ -33,6 +33,7 @@
 #include <nvgpu/dma.h>
 #include <nvgpu/pmu/fw.h>
 #include <nvgpu/pmu/debug.h>
+#include <nvgpu/pmu/pmu_pg.h>
 
 #include "pg_sw_gm20b.h"
 #include "pg_sw_gv11b.h"
diff --git a/drivers/gpu/nvgpu/common/pmu/pmu.c b/drivers/gpu/nvgpu/common/pmu/pmu.c
index 87c547309..970297024 100644
--- a/drivers/gpu/nvgpu/common/pmu/pmu.c
+++ b/drivers/gpu/nvgpu/common/pmu/pmu.c
@@ -22,19 +22,13 @@
 
 #include <nvgpu/pmu.h>
 #include <nvgpu/pmu/pmu_pg.h>
-#include <nvgpu/dma.h>
 #include <nvgpu/pmu/mutex.h>
 #include <nvgpu/pmu/seq.h>
 #include <nvgpu/log.h>
-#include <nvgpu/pmu/pmuif/nvgpu_cmdif.h>
 #include <nvgpu/enabled.h>
-#include <nvgpu/engine_queue.h>
-#include <nvgpu/barrier.h>
-#include <nvgpu/timers.h>
 #include <nvgpu/bug.h>
 #include <nvgpu/utils.h>
 #include <nvgpu/gk20a.h>
-#include <nvgpu/string.h>
 #include <nvgpu/power_features/cg.h>
 #include <nvgpu/nvgpu_err.h>
 #include <nvgpu/pmu/lsfm.h>
@@ -43,7 +37,6 @@
 #include <nvgpu/pmu/pmu_perfmon.h>
 #include <nvgpu/pmu/pmu_pg.h>
 #include <nvgpu/pmu/fw.h>
-#include <nvgpu/firmware.h>
 #include <nvgpu/pmu/debug.h>
 #include <nvgpu/boardobj.h>
 #include <nvgpu/boardobjgrp.h>
diff --git a/drivers/gpu/nvgpu/common/pmu/pmu_debug.c b/drivers/gpu/nvgpu/common/pmu/pmu_debug.c
index 4f64e88d9..7711a4753 100644
--- a/drivers/gpu/nvgpu/common/pmu/pmu_debug.c
+++ b/drivers/gpu/nvgpu/common/pmu/pmu_debug.c
@@ -24,6 +24,7 @@
 #include <nvgpu/gk20a.h>
 #include <nvgpu/pmu/fw.h>
 #include <nvgpu/dma.h>
+#include <nvgpu/pmu/pmu_pg.h>
 #include <nvgpu/pmu/debug.h>
 
 bool nvgpu_find_hex_in_string(char *strings, struct gk20a *g, u32 *hex_pos)
diff --git a/drivers/gpu/nvgpu/common/power_features/pg/pg.c b/drivers/gpu/nvgpu/common/power_features/pg/pg.c
index ad7592eb6..1d14457d7 100644
--- a/drivers/gpu/nvgpu/common/power_features/pg/pg.c
+++ b/drivers/gpu/nvgpu/common/power_features/pg/pg.c
@@ -22,6 +22,7 @@
 
 #include <nvgpu/gk20a.h>
 #include <nvgpu/pmu.h>
+#include <nvgpu/pmu/pmu_pg.h>
 #include <nvgpu/power_features/pg.h>
 
 bool nvgpu_pg_elpg_is_enabled(struct gk20a *g)
diff --git a/drivers/gpu/nvgpu/hal/fifo/preempt_gk20a.c b/drivers/gpu/nvgpu/hal/fifo/preempt_gk20a.c
index ab494d63c..1e4d3bd50 100644
--- a/drivers/gpu/nvgpu/hal/fifo/preempt_gk20a.c
+++ b/drivers/gpu/nvgpu/hal/fifo/preempt_gk20a.c
@@ -32,6 +32,7 @@
 #include <nvgpu/channel.h>
 #include <nvgpu/tsg.h>
 #include <nvgpu/preempt.h>
+#include <nvgpu/pmu/mutex.h>
 
 #include "preempt_gk20a.h"
 
diff --git a/drivers/gpu/nvgpu/hal/fifo/preempt_gv11b.c b/drivers/gpu/nvgpu/hal/fifo/preempt_gv11b.c
index 59d18918f..8162c3c8a 100644
--- a/drivers/gpu/nvgpu/hal/fifo/preempt_gv11b.c
+++ b/drivers/gpu/nvgpu/hal/fifo/preempt_gv11b.c
@@ -37,6 +37,7 @@
 #include <nvgpu/engine_status.h>
 #include <nvgpu/preempt.h>
 #include <nvgpu/nvgpu_err.h>
+#include <nvgpu/pmu/mutex.h>
 
 #include "preempt_gv11b.h"
 
diff --git a/drivers/gpu/nvgpu/hal/pmu/pmu_gk20a.c b/drivers/gpu/nvgpu/hal/pmu/pmu_gk20a.c
index f4b629bcf..74ca49979 100644
--- a/drivers/gpu/nvgpu/hal/pmu/pmu_gk20a.c
+++ b/drivers/gpu/nvgpu/hal/pmu/pmu_gk20a.c
@@ -29,6 +29,7 @@
 #include <nvgpu/pmu/mutex.h>
 #include <nvgpu/pmu/fw.h>
 #include <nvgpu/pmu/debug.h>
+#include <nvgpu/pmu/pmu_pg.h>
 
 #include <nvgpu/hw/gk20a/hw_pwr_gk20a.h>
 #include <nvgpu/hw/gk20a/hw_mc_gk20a.h>
diff --git a/drivers/gpu/nvgpu/hal/rc/rc_gv11b.c b/drivers/gpu/nvgpu/hal/rc/rc_gv11b.c
index 4c6aef5d9..682e57976 100644
--- a/drivers/gpu/nvgpu/hal/rc/rc_gv11b.c
+++ b/drivers/gpu/nvgpu/hal/rc/rc_gv11b.c
@@ -38,6 +38,7 @@
 #include <nvgpu/power_features/power_features.h>
 #include <nvgpu/gr/fecs_trace.h>
 #include <nvgpu/preempt.h>
+#include <nvgpu/pmu/mutex.h>
 
 #include "rc_gv11b.h"
 
diff --git a/drivers/gpu/nvgpu/include/nvgpu/gk20a.h b/drivers/gpu/nvgpu/include/nvgpu/gk20a.h
index 4a66152c6..6453b9a1c 100644
--- a/drivers/gpu/nvgpu/include/nvgpu/gk20a.h
+++ b/drivers/gpu/nvgpu/include/nvgpu/gk20a.h
@@ -89,6 +89,7 @@ struct nvgpu_fecs_ecc_status;
 struct _resmgr_context;
 struct nvgpu_gpfifo_entry;
 struct vm_gk20a_mapping_batch;
+struct pmu_pg_stats_data;
 
 enum nvgpu_unit;
 enum nvgpu_flush_op;
diff --git a/drivers/gpu/nvgpu/include/nvgpu/pmu.h b/drivers/gpu/nvgpu/include/nvgpu/pmu.h
index be8417273..e80377b49 100644
--- a/drivers/gpu/nvgpu/include/nvgpu/pmu.h
+++ b/drivers/gpu/nvgpu/include/nvgpu/pmu.h
@@ -27,13 +27,10 @@
 #include <nvgpu/allocator.h>
 #include <nvgpu/lock.h>
 #include <nvgpu/nvgpu_common.h>
+#include <nvgpu/nvgpu_mem.h>
 #include <nvgpu/flcnif_cmn.h>
-#include <nvgpu/pmu/pmuif/nvgpu_cmdif.h>
 #include <nvgpu/falcon.h>
 #include <nvgpu/timers.h>
-#include <nvgpu/pmu/pmu_pg.h>
-#include <nvgpu/pmu/seq.h>
-#include <nvgpu/pmu/mutex.h>
 #include <nvgpu/pmu/queue.h>
 #include <nvgpu/pmu/msg.h>
 #include <nvgpu/pmu/fw.h>
diff --git a/drivers/gpu/nvgpu/os/linux/debug_pmu.c b/drivers/gpu/nvgpu/os/linux/debug_pmu.c
index 34e4fefd7..4b55389bc 100644
--- a/drivers/gpu/nvgpu/os/linux/debug_pmu.c
+++ b/drivers/gpu/nvgpu/os/linux/debug_pmu.c
@@ -15,6 +15,7 @@
 #include <nvgpu/enabled.h>
 #include <nvgpu/pmu/pmu_perfmon.h>
 #include <nvgpu/pmu/debug.h>
+#include <nvgpu/pmu/pmu_pg.h>
 
 #include "debug_pmu.h"
 #include "os_linux.h"
diff --git a/drivers/gpu/nvgpu/os/linux/sysfs.c b/drivers/gpu/nvgpu/os/linux/sysfs.c
index 868f08f55..8a7004728 100644
--- a/drivers/gpu/nvgpu/os/linux/sysfs.c
+++ b/drivers/gpu/nvgpu/os/linux/sysfs.c
@@ -32,6 +32,7 @@
 #include <nvgpu/power_features/pg.h>
 #include <nvgpu/pmu/pmu_perfmon.h>
 #include <nvgpu/pmu/fw.h>
+#include <nvgpu/pmu/pmu_pg.h>
 
 #include "os_linux.h"
 #include "sysfs.h"