diff --git a/drivers/gpu/msm/adreno.c b/drivers/gpu/msm/adreno.c
index 89564ba66705..231f6900e610 100644
--- a/drivers/gpu/msm/adreno.c
+++ b/drivers/gpu/msm/adreno.c
@@ -324,6 +324,7 @@ void adreno_fault_detect_stop(struct adreno_device *adreno_dev)
 /* Send an NMI to the GMU */
 void adreno_gmu_send_nmi(struct adreno_device *adreno_dev)
 {
+	u32 val;
 	/* Mask so there's no interrupt caused by NMI */
 	adreno_write_gmureg(adreno_dev,
 			ADRENO_REG_GMU_GMU2HOST_INTR_MASK, 0xFFFFFFFF);
@@ -333,9 +334,10 @@ void adreno_gmu_send_nmi(struct adreno_device *adreno_dev)
 	if (ADRENO_QUIRK(adreno_dev, ADRENO_QUIRK_HFI_USE_REG))
 		adreno_write_gmureg(adreno_dev,
 				ADRENO_REG_GMU_NMI_CONTROL_STATUS, 0);
-	adreno_write_gmureg(adreno_dev,
-			ADRENO_REG_GMU_CM3_CFG,
-			(1 << GMU_CM3_CFG_NONMASKINTR_SHIFT));
+
+	adreno_read_gmureg(adreno_dev, ADRENO_REG_GMU_CM3_CFG, &val);
+	val |= 1 << GMU_CM3_CFG_NONMASKINTR_SHIFT;
+	adreno_write_gmureg(adreno_dev, ADRENO_REG_GMU_CM3_CFG, val);
 
 	/* Make sure the NMI is invoked before we proceed*/
 	wmb();
diff --git a/drivers/gpu/msm/adreno_a6xx_gmu.c b/drivers/gpu/msm/adreno_a6xx_gmu.c
index 4ff6f1c1cc5e..3b2951f2be1d 100644
--- a/drivers/gpu/msm/adreno_a6xx_gmu.c
+++ b/drivers/gpu/msm/adreno_a6xx_gmu.c
@@ -357,6 +357,10 @@ static int a6xx_gmu_start(struct kgsl_device *device)
 
 	/* Bring GMU out of reset */
 	gmu_core_regwrite(device, A6XX_GMU_CM3_SYSRESET, 0);
+
+	/* Make sure the write is posted before moving ahead */
+	wmb();
+
 	if (timed_poll_check(device,
 			A6XX_GMU_CM3_FW_INIT_RESULT,
 			val, GMU_START_TIMEOUT, mask)) {
@@ -1058,6 +1062,13 @@ static int a6xx_gmu_fw_start(struct kgsl_device *device,
 	gmu_core_regwrite(device, A6XX_GMU_AHB_FENCE_RANGE_0,
 			GMU_FENCE_RANGE_MASK);
 
+	/*
+	 * Make sure that CM3 state is at reset value. Snapshot is changing
+	 * NMI bit and if we boot up GMU with NMI bit set GMU will boot
+	 * straight in to NMI handler without executing __main code
+	 */
+	gmu_core_regwrite(device, A6XX_GMU_CM3_CFG, 0x4052);
+
 	/* Pass chipid to GMU FW, must happen before starting GMU */
 
 	/* Keep Core and Major bitfields unchanged */