drm/msm: Add A6XX device support

Add support for the A6XX family of Adreno GPUs. The biggest addition
is the GMU (Graphics Management Unit) which takes over most of the
power management of the GPU itself but in a ironic twist of fate
needs a goodly amount of management itself. Add support for the
A6XX core code, the GMU and the HFI (hardware firmware interface)
queue that the CPU uses to communicate with the GMU.

Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>

10 files changed, 2828 insertions, 1 deletions

diff --git a/drivers/gpu/drm/msm/Makefile b/drivers/gpu/drm/msm/Makefile
index 7c773e003663..261fa79d456d 100644
--- a/drivers/gpu/drm/msm/Makefile
+++ b/drivers/gpu/drm/msm/Makefile
@@ -11,6 +11,9 @@ msm-y := \
 	adreno/a5xx_gpu.o \
 	adreno/a5xx_power.o \
 	adreno/a5xx_preempt.o \
+	adreno/a6xx_gpu.o \
+	adreno/a6xx_gmu.o \
+	adreno/a6xx_hfi.o \
 	hdmi/hdmi.o \
 	hdmi/hdmi_audio.o \
 	hdmi/hdmi_bridge.o \
diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gmu.c b/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
new file mode 100644
index 000000000000..fbb501986720
--- /dev/null
+++ b/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
@@ -0,0 +1,1207 @@
+// SPDX-License-Identifier: GPL-2.0
+/* Copyright (c) 2017-2018 The Linux Foundation. All rights reserved. */
+
+#include <linux/clk.h>
+#include <linux/iopoll.h>
+#include <linux/pm_opp.h>
+#include <soc/qcom/cmd-db.h>
+
+#include "a6xx_gpu.h"
+#include "a6xx_gmu.xml.h"
+
+static irqreturn_t a6xx_gmu_irq(int irq, void *data)
+{
+	struct a6xx_gmu *gmu = data;
+	u32 status;
+
+	status = gmu_read(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_STATUS);
+	gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_CLR, status);
+
+	if (status & A6XX_GMU_AO_HOST_INTERRUPT_STATUS_WDOG_BITE) {
+		dev_err_ratelimited(gmu->dev, "GMU watchdog expired\n");
+
+		/* Temporary until we can recover safely */
+		BUG();
+	}
+
+	if (status &  A6XX_GMU_AO_HOST_INTERRUPT_STATUS_HOST_AHB_BUS_ERROR)
+		dev_err_ratelimited(gmu->dev, "GMU AHB bus error\n");
+
+	if (status & A6XX_GMU_AO_HOST_INTERRUPT_STATUS_FENCE_ERR)
+		dev_err_ratelimited(gmu->dev, "GMU fence error: 0x%x\n",
+			gmu_read(gmu, REG_A6XX_GMU_AHB_FENCE_STATUS));
+
+	return IRQ_HANDLED;
+}
+
+static irqreturn_t a6xx_hfi_irq(int irq, void *data)
+{
+	struct a6xx_gmu *gmu = data;
+	u32 status;
+
+	status = gmu_read(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO);
+	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR, status);
+
+	if (status & A6XX_GMU_GMU2HOST_INTR_INFO_MSGQ)
+		tasklet_schedule(&gmu->hfi_tasklet);
+
+	if (status & A6XX_GMU_GMU2HOST_INTR_INFO_CM3_FAULT) {
+		dev_err_ratelimited(gmu->dev, "GMU firmware fault\n");
+
+		/* Temporary until we can recover safely */
+		BUG();
+	}
+
+	return IRQ_HANDLED;
+}
+
+/* Check to see if the GX rail is still powered */
+static bool a6xx_gmu_gx_is_on(struct a6xx_gmu *gmu)
+{
+	u32 val = gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS);
+
+	return !(val &
+		(A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_GX_HM_GDSC_POWER_OFF |
+		A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_GX_HM_CLK_OFF));
+}
+
+static int a6xx_gmu_set_freq(struct a6xx_gmu *gmu, int index)
+{
+	gmu_write(gmu, REG_A6XX_GMU_DCVS_ACK_OPTION, 0);
+
+	gmu_write(gmu, REG_A6XX_GMU_DCVS_PERF_SETTING,
+		((index << 24) & 0xff) | (3 & 0xf));
+
+	/*
+	 * Send an invalid index as a vote for the bus bandwidth and let the
+	 * firmware decide on the right vote
+	 */
+	gmu_write(gmu, REG_A6XX_GMU_DCVS_BW_SETTING, 0xff);
+
+	/* Set and clear the OOB for DCVS to trigger the GMU */
+	a6xx_gmu_set_oob(gmu, GMU_OOB_DCVS_SET);
+	a6xx_gmu_clear_oob(gmu, GMU_OOB_DCVS_SET);
+
+	return gmu_read(gmu, REG_A6XX_GMU_DCVS_RETURN);
+}
+
+static bool a6xx_gmu_check_idle_level(struct a6xx_gmu *gmu)
+{
+	u32 val;
+	int local = gmu->idle_level;
+
+	/* SPTP and IFPC both report as IFPC */
+	if (gmu->idle_level == GMU_IDLE_STATE_SPTP)
+		local = GMU_IDLE_STATE_IFPC;
+
+	val = gmu_read(gmu, REG_A6XX_GPU_GMU_CX_GMU_RPMH_POWER_STATE);
+
+	if (val == local) {
+		if (gmu->idle_level != GMU_IDLE_STATE_IFPC ||
+			!a6xx_gmu_gx_is_on(gmu))
+			return true;
+	}
+
+	return false;
+}
+
+/* Wait for the GMU to get to its most idle state */
+int a6xx_gmu_wait_for_idle(struct a6xx_gpu *a6xx_gpu)
+{
+	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
+
+	return spin_until(a6xx_gmu_check_idle_level(gmu));
+}
+
+static int a6xx_gmu_start(struct a6xx_gmu *gmu)
+{
+	int ret;
+	u32 val;
+
+	gmu_write(gmu, REG_A6XX_GMU_CM3_SYSRESET, 1);
+	gmu_write(gmu, REG_A6XX_GMU_CM3_SYSRESET, 0);
+
+	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_CM3_FW_INIT_RESULT, val,
+		val == 0xbabeface, 100, 10000);
+
+	if (ret)
+		dev_err(gmu->dev, "GMU firmware initalization timed out\n");
+
+	return ret;
+}
+
+static int a6xx_gmu_hfi_start(struct a6xx_gmu *gmu)
+{
+	u32 val;
+	int ret;
+
+	gmu_rmw(gmu, REG_A6XX_GMU_GMU2HOST_INTR_MASK,
+		A6XX_GMU_GMU2HOST_INTR_INFO_MSGQ, 0);
+
+	gmu_write(gmu, REG_A6XX_GMU_HFI_CTRL_INIT, 1);
+
+	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_HFI_CTRL_STATUS, val,
+		val & 1, 100, 10000);
+	if (ret)
+		dev_err(gmu->dev, "Unable to start the HFI queues\n");
+
+	return ret;
+}
+
+/* Trigger a OOB (out of band) request to the GMU */
+int a6xx_gmu_set_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state)
+{
+	int ret;
+	u32 val;
+	int request, ack;
+	const char *name;
+
+	switch (state) {
+	case GMU_OOB_GPU_SET:
+		request = GMU_OOB_GPU_SET_REQUEST;
+		ack = GMU_OOB_GPU_SET_ACK;
+		name = "GPU_SET";
+		break;
+	case GMU_OOB_BOOT_SLUMBER:
+		request = GMU_OOB_BOOT_SLUMBER_REQUEST;
+		ack = GMU_OOB_BOOT_SLUMBER_ACK;
+		name = "BOOT_SLUMBER";
+		break;
+	case GMU_OOB_DCVS_SET:
+		request = GMU_OOB_DCVS_REQUEST;
+		ack = GMU_OOB_DCVS_ACK;
+		name = "GPU_DCVS";
+		break;
+	default:
+		return -EINVAL;
+	}
+
+	/* Trigger the equested OOB operation */
+	gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET, 1 << request);
+
+	/* Wait for the acknowledge interrupt */
+	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO, val,
+		val & (1 << ack), 100, 10000);
+
+	if (ret)
+		dev_err(gmu->dev,
+			"Timeout waiting for GMU OOB set %s: 0x%x\n",
+				name,
+				gmu_read(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO));
+
+	/* Clear the acknowledge interrupt */
+	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR, 1 << ack);
+
+	return ret;
+}
+
+/* Clear a pending OOB state in the GMU */
+void a6xx_gmu_clear_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state)
+{
+	switch (state) {
+	case GMU_OOB_GPU_SET:
+		gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET,
+			1 << GMU_OOB_GPU_SET_CLEAR);
+		break;
+	case GMU_OOB_BOOT_SLUMBER:
+		gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET,
+			1 << GMU_OOB_BOOT_SLUMBER_CLEAR);
+		break;
+	case GMU_OOB_DCVS_SET:
+		gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET,
+			1 << GMU_OOB_DCVS_CLEAR);
+		break;
+	}
+}
+
+/* Enable CPU control of SPTP power power collapse */
+static int a6xx_sptprac_enable(struct a6xx_gmu *gmu)
+{
+	int ret;
+	u32 val;
+
+	gmu_write(gmu, REG_A6XX_GMU_GX_SPTPRAC_POWER_CONTROL, 0x778000);
+
+	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS, val,
+		(val & 0x38) == 0x28, 1, 100);
+
+	if (ret) {
+		dev_err(gmu->dev, "Unable to power on SPTPRAC: 0x%x\n",
+			gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS));
+	}
+
+	return 0;
+}
+
+/* Disable CPU control of SPTP power power collapse */
+static void a6xx_sptprac_disable(struct a6xx_gmu *gmu)
+{
+	u32 val;
+	int ret;
+
+	/* Make sure retention is on */
+	gmu_rmw(gmu, REG_A6XX_GPU_CC_GX_GDSCR, 0, (1 << 11));
+
+	gmu_write(gmu, REG_A6XX_GMU_GX_SPTPRAC_POWER_CONTROL, 0x778001);
+
+	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS, val,
+		(val & 0x04), 100, 10000);
+
+	if (ret)
+		dev_err(gmu->dev, "failed to power off SPTPRAC: 0x%x\n",
+			gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS));
+}
+
+/* Let the GMU know we are starting a boot sequence */
+static int a6xx_gmu_gfx_rail_on(struct a6xx_gmu *gmu)
+{
+	u32 vote;
+
+	/* Let the GMU know we are getting ready for boot */
+	gmu_write(gmu, REG_A6XX_GMU_BOOT_SLUMBER_OPTION, 0);
+
+	/* Choose the "default" power level as the highest available */
+	vote = gmu->gx_arc_votes[gmu->nr_gpu_freqs - 1];
+
+	gmu_write(gmu, REG_A6XX_GMU_GX_VOTE_IDX, vote & 0xff);
+	gmu_write(gmu, REG_A6XX_GMU_MX_VOTE_IDX, (vote >> 8) & 0xff);
+
+	/* Let the GMU know the boot sequence has started */
+	return a6xx_gmu_set_oob(gmu, GMU_OOB_BOOT_SLUMBER);
+}
+
+/* Let the GMU know that we are about to go into slumber */
+static int a6xx_gmu_notify_slumber(struct a6xx_gmu *gmu)
+{
+	int ret;
+
+	/* Disable the power counter so the GMU isn't busy */
+	gmu_write(gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_ENABLE, 0);
+
+	/* Disable SPTP_PC if the CPU is responsible for it */
+	if (gmu->idle_level < GMU_IDLE_STATE_SPTP)
+		a6xx_sptprac_disable(gmu);
+
+	/* Tell the GMU to get ready to slumber */
+	gmu_write(gmu, REG_A6XX_GMU_BOOT_SLUMBER_OPTION, 1);
+
+	ret = a6xx_gmu_set_oob(gmu, GMU_OOB_BOOT_SLUMBER);
+	a6xx_gmu_clear_oob(gmu, GMU_OOB_BOOT_SLUMBER);
+
+	if (!ret) {
+		/* Check to see if the GMU really did slumber */
+		if (gmu_read(gmu, REG_A6XX_GPU_GMU_CX_GMU_RPMH_POWER_STATE)
+			!= 0x0f) {
+			dev_err(gmu->dev, "The GMU did not go into slumber\n");
+			ret = -ETIMEDOUT;
+		}
+	}
+
+	/* Put fence into allow mode */
+	gmu_write(gmu, REG_A6XX_GMU_AO_AHB_FENCE_CTRL, 0);
+	return ret;
+}
+
+static int a6xx_rpmh_start(struct a6xx_gmu *gmu)
+{
+	int ret;
+	u32 val;
+
+	gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 1 << 1);
+	/* Wait for the register to finish posting */
+	wmb();
+
+	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_RSCC_CONTROL_ACK, val,
+		val & (1 << 1), 100, 10000);
+	if (ret) {
+		dev_err(gmu->dev, "Unable to power on the GPU RSC\n");
+		return ret;
+	}
+
+	ret = gmu_poll_timeout(gmu, REG_A6XX_RSCC_SEQ_BUSY_DRV0, val,
+		!val, 100, 10000);
+
+	if (!ret) {
+		gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 0);
+
+		/* Re-enable the power counter */
+		gmu_write(gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_ENABLE, 1);
+		return 0;
+	}
+
+	dev_err(gmu->dev, "GPU RSC sequence stuck while waking up the GPU\n");
+	return ret;
+}
+
+static void a6xx_rpmh_stop(struct a6xx_gmu *gmu)
+{
+	int ret;
+	u32 val;
+
+	gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 1);
+
+	ret = gmu_poll_timeout(gmu, REG_A6XX_GPU_RSCC_RSC_STATUS0_DRV0,
+		val, val & (1 << 16), 100, 10000);
+	if (ret)
+		dev_err(gmu->dev, "Unable to power off the GPU RSC\n");
+
+	gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 0);
+}
+
+static void a6xx_gmu_rpmh_init(struct a6xx_gmu *gmu)
+{
+	/* Disable SDE clock gating */
+	gmu_write(gmu, REG_A6XX_GPU_RSCC_RSC_STATUS0_DRV0, BIT(24));
+
+	/* Setup RSC PDC handshake for sleep and wakeup */
+	gmu_write(gmu, REG_A6XX_RSCC_PDC_SLAVE_ID_DRV0, 1);
+	gmu_write(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA, 0);
+	gmu_write(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR, 0);
+	gmu_write(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA + 2, 0);
+	gmu_write(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR + 2, 0);
+	gmu_write(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA + 4, 0x80000000);
+	gmu_write(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR + 4, 0);
+	gmu_write(gmu, REG_A6XX_RSCC_OVERRIDE_START_ADDR, 0);
+	gmu_write(gmu, REG_A6XX_RSCC_PDC_SEQ_START_ADDR, 0x4520);
+	gmu_write(gmu, REG_A6XX_RSCC_PDC_MATCH_VALUE_LO, 0x4510);
+	gmu_write(gmu, REG_A6XX_RSCC_PDC_MATCH_VALUE_HI, 0x4514);
+
+	/* Load RSC sequencer uCode for sleep and wakeup */
+	gmu_write(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0, 0xa7a506a0);
+	gmu_write(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 1, 0xa1e6a6e7);
+	gmu_write(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 2, 0xa2e081e1);
+	gmu_write(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 3, 0xe9a982e2);
+	gmu_write(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 4, 0x0020e8a8);
+
+	/* Load PDC sequencer uCode for power up and power down sequence */
+	pdc_write(gmu, REG_A6XX_PDC_GPU_SEQ_MEM_0, 0xfebea1e1);
+	pdc_write(gmu, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 1, 0xa5a4a3a2);
+	pdc_write(gmu, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 2, 0x8382a6e0);
+	pdc_write(gmu, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 3, 0xbce3e284);
+	pdc_write(gmu, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 4, 0x002081fc);
+
+	/* Set TCS commands used by PDC sequence for low power modes */
+	pdc_write(gmu, REG_A6XX_PDC_GPU_TCS1_CMD_ENABLE_BANK, 7);
+	pdc_write(gmu, REG_A6XX_PDC_GPU_TCS1_CMD_WAIT_FOR_CMPL_BANK, 0);
+	pdc_write(gmu, REG_A6XX_PDC_GPU_TCS1_CONTROL, 0);
+	pdc_write(gmu, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID, 0x10108);
+	pdc_write(gmu, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR, 0x30010);
+	pdc_write(gmu, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA, 1);
+	pdc_write(gmu, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID + 4, 0x10108);
+	pdc_write(gmu, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR + 4, 0x30000);
+	pdc_write(gmu, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA + 4, 0x0);
+	pdc_write(gmu, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID + 8, 0x10108);
+	pdc_write(gmu, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR + 8, 0x30080);
+	pdc_write(gmu, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA + 8, 0x0);
+	pdc_write(gmu, REG_A6XX_PDC_GPU_TCS3_CMD_ENABLE_BANK, 7);
+	pdc_write(gmu, REG_A6XX_PDC_GPU_TCS3_CMD_WAIT_FOR_CMPL_BANK, 0);
+	pdc_write(gmu, REG_A6XX_PDC_GPU_TCS3_CONTROL, 0);
+	pdc_write(gmu, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID, 0x10108);
+	pdc_write(gmu, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR, 0x30010);
+	pdc_write(gmu, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA, 2);
+	pdc_write(gmu, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID + 4, 0x10108);
+	pdc_write(gmu, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR + 4, 0x30000);
+	pdc_write(gmu, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 4, 0x3);
+	pdc_write(gmu, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID + 8, 0x10108);
+	pdc_write(gmu, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR + 8, 0x30080);
+	pdc_write(gmu, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 8, 0x3);
+
+	/* Setup GPU PDC */
+	pdc_write(gmu, REG_A6XX_PDC_GPU_SEQ_START_ADDR, 0);
+	pdc_write(gmu, REG_A6XX_PDC_GPU_ENABLE_PDC, 0x80000001);
+
+	/* ensure no writes happen before the uCode is fully written */
+	wmb();
+}
+
+/*
+ * The lowest 16 bits of this value are the number of XO clock cycles for main
+ * hysteresis which is set at 0x1680 cycles (300 us).  The higher 16 bits are
+ * for the shorter hysteresis that happens after main - this is 0xa (.5 us)
+ */
+
+#define GMU_PWR_COL_HYST 0x000a1680
+
+/* Set up the idle state for the GMU */
+static void a6xx_gmu_power_config(struct a6xx_gmu *gmu)
+{
+	/* Disable GMU WB/RB buffer */
+	gmu_write(gmu, REG_A6XX_GMU_SYS_BUS_CONFIG, 0x1);
+
+	gmu_write(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0x9c40400);
+
+	switch (gmu->idle_level) {
+	case GMU_IDLE_STATE_IFPC:
+		gmu_write(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_HYST,
+			GMU_PWR_COL_HYST);
+		gmu_rmw(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0,
+			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_IFPC_ENABLE |
+			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_HM_POWER_COLLAPSE_ENABLE);
+		/* Fall through */
+	case GMU_IDLE_STATE_SPTP:
+		gmu_write(gmu, REG_A6XX_GMU_PWR_COL_SPTPRAC_HYST,
+			GMU_PWR_COL_HYST);
+		gmu_rmw(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0,
+			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_IFPC_ENABLE |
+			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_SPTPRAC_POWER_CONTROL_ENABLE);
+	}
+
+	/* Enable RPMh GPU client */
+	gmu_rmw(gmu, REG_A6XX_GMU_RPMH_CTRL, 0,
+		A6XX_GMU_RPMH_CTRL_RPMH_INTERFACE_ENABLE |
+		A6XX_GMU_RPMH_CTRL_LLC_VOTE_ENABLE |
+		A6XX_GMU_RPMH_CTRL_DDR_VOTE_ENABLE |
+		A6XX_GMU_RPMH_CTRL_MX_VOTE_ENABLE |
+		A6XX_GMU_RPMH_CTRL_CX_VOTE_ENABLE |
+		A6XX_GMU_RPMH_CTRL_GFX_VOTE_ENABLE);
+}
+
+static int a6xx_gmu_fw_start(struct a6xx_gmu *gmu, unsigned int state)
+{
+	static bool rpmh_init;
+	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
+	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
+	int i, ret;
+	u32 chipid;
+	u32 *image;
+
+	if (state == GMU_WARM_BOOT) {
+		ret = a6xx_rpmh_start(gmu);
+		if (ret)
+			return ret;
+	} else {
+		if (WARN(!adreno_gpu->fw[ADRENO_FW_GMU],
+			"GMU firmware is not loaded\n"))
+			return -ENOENT;
+
+		/* Sanity check the size of the firmware that was loaded */
+		if (adreno_gpu->fw[ADRENO_FW_GMU]->size > 0x8000) {
+			dev_err(gmu->dev,
+				"GMU firmware is bigger than the available region\n");
+			return -EINVAL;
+		}
+
+		/* Turn on register retention */
+		gmu_write(gmu, REG_A6XX_GMU_GENERAL_7, 1);
+
+		/* We only need to load the RPMh microcode once */
+		if (!rpmh_init) {
+			a6xx_gmu_rpmh_init(gmu);
+			rpmh_init = true;
+		} else if (state != GMU_RESET) {
+			ret = a6xx_rpmh_start(gmu);
+			if (ret)
+				return ret;
+		}
+
+		image = (u32 *) adreno_gpu->fw[ADRENO_FW_GMU]->data;
+
+		for (i = 0; i < adreno_gpu->fw[ADRENO_FW_GMU]->size >> 2; i++)
+			gmu_write(gmu, REG_A6XX_GMU_CM3_ITCM_START + i,
+				image[i]);
+	}
+
+	gmu_write(gmu, REG_A6XX_GMU_CM3_FW_INIT_RESULT, 0);
+	gmu_write(gmu, REG_A6XX_GMU_CM3_BOOT_CONFIG, 0x02);
+
+	/* Write the iova of the HFI table */
+	gmu_write(gmu, REG_A6XX_GMU_HFI_QTBL_ADDR, gmu->hfi->iova);
+	gmu_write(gmu, REG_A6XX_GMU_HFI_QTBL_INFO, 1);
+
+	gmu_write(gmu, REG_A6XX_GMU_AHB_FENCE_RANGE_0,
+		(1 << 31) | (0xa << 18) | (0xa0));
+
+	chipid = adreno_gpu->rev.core << 24;
+	chipid |= adreno_gpu->rev.major << 16;
+	chipid |= adreno_gpu->rev.minor << 12;
+	chipid |= adreno_gpu->rev.patchid << 8;
+
+	gmu_write(gmu, REG_A6XX_GMU_HFI_SFR_ADDR, chipid);
+
+	/* Set up the lowest idle level on the GMU */
+	a6xx_gmu_power_config(gmu);
+
+	ret = a6xx_gmu_start(gmu);
+	if (ret)
+		return ret;
+
+	ret = a6xx_gmu_gfx_rail_on(gmu);
+	if (ret)
+		return ret;
+
+	/* Enable SPTP_PC if the CPU is responsible for it */
+	if (gmu->idle_level < GMU_IDLE_STATE_SPTP) {
+		ret = a6xx_sptprac_enable(gmu);
+		if (ret)
+			return ret;
+	}
+
+	ret = a6xx_gmu_hfi_start(gmu);
+	if (ret)
+		return ret;
+
+	/* FIXME: Do we need this wmb() here? */
+	wmb();
+
+	return 0;
+}
+
+#define A6XX_HFI_IRQ_MASK \
+	(A6XX_GMU_GMU2HOST_INTR_INFO_MSGQ | \
+	 A6XX_GMU_GMU2HOST_INTR_INFO_CM3_FAULT)
+
+#define A6XX_GMU_IRQ_MASK \
+	(A6XX_GMU_AO_HOST_INTERRUPT_STATUS_WDOG_BITE | \
+	 A6XX_GMU_AO_HOST_INTERRUPT_STATUS_HOST_AHB_BUS_ERROR | \
+	 A6XX_GMU_AO_HOST_INTERRUPT_STATUS_FENCE_ERR)
+
+static void a6xx_gmu_irq_enable(struct a6xx_gmu *gmu)
+{
+	gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_CLR, ~0);
+	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR, ~0);
+
+	gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_MASK,
+		~A6XX_GMU_IRQ_MASK);
+	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_MASK,
+		~A6XX_HFI_IRQ_MASK);
+
+	enable_irq(gmu->gmu_irq);
+	enable_irq(gmu->hfi_irq);
+}
+
+static void a6xx_gmu_irq_disable(struct a6xx_gmu *gmu)
+{
+	disable_irq(gmu->gmu_irq);
+	disable_irq(gmu->hfi_irq);
+
+	gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_MASK, ~0);
+	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_MASK, ~0);
+}
+
+int a6xx_gmu_reset(struct a6xx_gpu *a6xx_gpu)
+{
+	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
+	int ret;
+	u32 val;
+
+	/* Flush all the queues */
+	a6xx_hfi_stop(gmu);
+
+	/* Stop the interrupts */
+	a6xx_gmu_irq_disable(gmu);
+
+	/* Force off SPTP in case the GMU is managing it */
+	a6xx_sptprac_disable(gmu);
+
+	/* Make sure there are no outstanding RPMh votes */
+	gmu_poll_timeout(gmu, REG_A6XX_RSCC_TCS0_DRV0_STATUS, val,
+		(val & 1), 100, 10000);
+	gmu_poll_timeout(gmu, REG_A6XX_RSCC_TCS1_DRV0_STATUS, val,
+		(val & 1), 100, 10000);
+	gmu_poll_timeout(gmu, REG_A6XX_RSCC_TCS2_DRV0_STATUS, val,
+		(val & 1), 100, 10000);
+	gmu_poll_timeout(gmu, REG_A6XX_RSCC_TCS3_DRV0_STATUS, val,
+		(val & 1), 100, 1000);
+
+	/* Force off the GX GSDC */
+	regulator_force_disable(gmu->gx);
+
+	/* Disable the resources */
+	clk_bulk_disable_unprepare(gmu->nr_clocks, gmu->clocks);
+	pm_runtime_put_sync(gmu->dev);
+
+	/* Re-enable the resources */
+	pm_runtime_get_sync(gmu->dev);
+
+	/* Use a known rate to bring up the GMU */
+	clk_set_rate(gmu->core_clk, 200000000);
+	ret = clk_bulk_prepare_enable(gmu->nr_clocks, gmu->clocks);
+	if (ret)
+		goto out;
+
+	a6xx_gmu_irq_enable(gmu);
+
+	ret = a6xx_gmu_fw_start(gmu, GMU_RESET);
+	if (!ret)
+		ret = a6xx_hfi_start(gmu, GMU_COLD_BOOT);
+
+	/* Set the GPU back to the highest power frequency */
+	a6xx_gmu_set_freq(gmu, gmu->nr_gpu_freqs - 1);
+
+out:
+	if (ret)
+		a6xx_gmu_clear_oob(gmu, GMU_OOB_BOOT_SLUMBER);
+
+	return ret;
+}
+
+int a6xx_gmu_resume(struct a6xx_gpu *a6xx_gpu)
+{
+	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
+	int status, ret;
+
+	if (WARN(!gmu->mmio, "The GMU is not set up yet\n"))
+		return 0;
+
+	/* Turn on the resources */
+	pm_runtime_get_sync(gmu->dev);
+
+	/* Use a known rate to bring up the GMU */
+	clk_set_rate(gmu->core_clk, 200000000);
+	ret = clk_bulk_prepare_enable(gmu->nr_clocks, gmu->clocks);
+	if (ret)
+		goto out;
+
+	a6xx_gmu_irq_enable(gmu);
+
+	/* Check to see if we are doing a cold or warm boot */
+	status = gmu_read(gmu, REG_A6XX_GMU_GENERAL_7) == 1 ?
+		GMU_WARM_BOOT : GMU_COLD_BOOT;
+
+	ret = a6xx_gmu_fw_start(gmu, status);
+	if (ret)
+		goto out;
+
+	ret = a6xx_hfi_start(gmu, status);
+
+	/* Set the GPU to the highest power frequency */
+	a6xx_gmu_set_freq(gmu, gmu->nr_gpu_freqs - 1);
+
+out:
+	/* Make sure to turn off the boot OOB request on error */
+	if (ret)
+		a6xx_gmu_clear_oob(gmu, GMU_OOB_BOOT_SLUMBER);
+
+	return ret;
+}
+
+bool a6xx_gmu_isidle(struct a6xx_gmu *gmu)
+{
+	u32 reg;
+
+	if (!gmu->mmio)
+		return true;
+
+	reg = gmu_read(gmu, REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS);
+
+	if (reg &  A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS_GPUBUSYIGNAHB)
+		return false;
+
+	return true;
+}
+
+int a6xx_gmu_stop(struct a6xx_gpu *a6xx_gpu)
+{
+	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
+	u32 val;
+
+	/*
+	 * The GMU may still be in slumber unless the GPU started so check and
+	 * skip putting it back into slumber if so
+	 */
+	val = gmu_read(gmu, REG_A6XX_GPU_GMU_CX_GMU_RPMH_POWER_STATE);
+
+	if (val != 0xf) {
+		int ret = a6xx_gmu_wait_for_idle(a6xx_gpu);
+
+		/* Temporary until we can recover safely */
+		BUG_ON(ret);
+
+		/* tell the GMU we want to slumber */
+		a6xx_gmu_notify_slumber(gmu);
+
+		ret = gmu_poll_timeout(gmu,
+			REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS, val,
+			!(val & A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS_GPUBUSYIGNAHB),
+			100, 10000);
+
+		/*
+		 * Let the user know we failed to slumber but don't worry too
+		 * much because we are powering down anyway
+		 */
+
+		if (ret)
+			dev_err(gmu->dev,
+				"Unable to slumber GMU: status = 0%x/0%x\n",
+				gmu_read(gmu,
+					REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS),
+				gmu_read(gmu,
+					REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS2));
+	}
+
+	/* Turn off HFI */
+	a6xx_hfi_stop(gmu);
+
+	/* Stop the interrupts and mask the hardware */
+	a6xx_gmu_irq_disable(gmu);
+
+	/* Tell RPMh to power off the GPU */
+	a6xx_rpmh_stop(gmu);
+
+	clk_bulk_disable_unprepare(gmu->nr_clocks, gmu->clocks);
+
+	pm_runtime_put_sync(gmu->dev);
+
+	return 0;
+}
+
+static void a6xx_gmu_memory_free(struct a6xx_gmu *gmu, struct a6xx_gmu_bo *bo)
+{
+	int count, i;
+	u64 iova;
+
+	if (IS_ERR_OR_NULL(bo))
+		return;
+
+	count = bo->size >> PAGE_SHIFT;
+	iova = bo->iova;
+
+	for (i = 0; i < count; i++, iova += PAGE_SIZE) {
+		iommu_unmap(gmu->domain, iova, PAGE_SIZE);
+		__free_pages(bo->pages[i], 0);
+	}
+
+	kfree(bo->pages);
+	kfree(bo);
+}
+
+static struct a6xx_gmu_bo *a6xx_gmu_memory_alloc(struct a6xx_gmu *gmu,
+		size_t size)
+{
+	struct a6xx_gmu_bo *bo;
+	int ret, count, i;
+
+	bo = kzalloc(sizeof(*bo), GFP_KERNEL);
+	if (!bo)
+		return ERR_PTR(-ENOMEM);
+
+	bo->size = PAGE_ALIGN(size);
+
+	count = bo->size >> PAGE_SHIFT;
+
+	bo->pages = kcalloc(count, sizeof(struct page *), GFP_KERNEL);
+	if (!bo->pages) {
+		kfree(bo);
+		return ERR_PTR(-ENOMEM);
+	}
+
+	for (i = 0; i < count; i++) {
+		bo->pages[i] = alloc_page(GFP_KERNEL);
+		if (!bo->pages[i])
+			goto err;
+	}
+
+	bo->iova = gmu->uncached_iova_base;
+
+	for (i = 0; i < count; i++) {
+		ret = iommu_map(gmu->domain,
+			bo->iova + (PAGE_SIZE * i),
+			page_to_phys(bo->pages[i]), PAGE_SIZE,
+			IOMMU_READ | IOMMU_WRITE);
+
+		if (ret) {
+			dev_err(gmu->dev, "Unable to map GMU buffer object\n");
+
+			for (i = i - 1 ; i >= 0; i--)
+				iommu_unmap(gmu->domain,
+					bo->iova + (PAGE_SIZE * i),
+					PAGE_SIZE);
+
+			goto err;
+		}
+	}
+
+	bo->virt = vmap(bo->pages, count, VM_IOREMAP,
+		pgprot_writecombine(PAGE_KERNEL));
+	if (!bo->virt)
+		goto err;
+
+	/* Align future IOVA addresses on 1MB boundaries */
+	gmu->uncached_iova_base += ALIGN(size, SZ_1M);
+
+	return bo;
+
+err:
+	for (i = 0; i < count; i++) {
+		if (bo->pages[i])
+			__free_pages(bo->pages[i], 0);
+	}
+
+	kfree(bo->pages);
+	kfree(bo);
+
+	return ERR_PTR(-ENOMEM);
+}
+
+static int a6xx_gmu_memory_probe(struct a6xx_gmu *gmu)
+{
+	int ret;
+
+	/*
+	 * The GMU address space is hardcoded to treat the range
+	 * 0x60000000 - 0x80000000 as un-cached memory. All buffers shared
+	 * between the GMU and the CPU will live in this space
+	 */
+	gmu->uncached_iova_base = 0x60000000;
+
+
+	gmu->domain = iommu_domain_alloc(&platform_bus_type);
+	if (!gmu->domain)
+		return -ENODEV;
+
+	ret = iommu_attach_device(gmu->domain, gmu->dev);
+
+	if (ret) {
+		iommu_domain_free(gmu->domain);
+		gmu->domain = NULL;
+	}
+
+	return ret;
+}
+
+/* Get the list of RPMh voltage levels from cmd-db */
+static int a6xx_gmu_rpmh_arc_cmds(const char *id, void *vals, int size)
+{
+	u32 len = cmd_db_read_aux_data_len(id);
+
+	if (!len)
+		return 0;
+
+	if (WARN_ON(len > size))
+		return -EINVAL;
+
+	cmd_db_read_aux_data(id, vals, len);
+
+	/*
+	 * The data comes back as an array of unsigned shorts so adjust the
+	 * count accordingly
+	 */
+	return len >> 1;
+}
+
+/* Return the 'arc-level' for the given frequency */
+static u32 a6xx_gmu_get_arc_level(struct device *dev, unsigned long freq)
+{
+	struct dev_pm_opp *opp;
+	struct device_node *np;
+	u32 val = 0;
+
+	if (!freq)
+		return 0;
+
+	opp  = dev_pm_opp_find_freq_exact(dev, freq, true);
+	if (IS_ERR(opp))
+		return 0;
+
+	np = dev_pm_opp_get_of_node(opp);
+
+	if (np) {
+		of_property_read_u32(np, "qcom,level", &val);
+		of_node_put(np);
+	}
+
+	dev_pm_opp_put(opp);
+
+	return val;
+}
+
+static int a6xx_gmu_rpmh_arc_votes_init(struct device *dev, u32 *votes,
+		unsigned long *freqs, int freqs_count,
+		u16 *pri, int pri_count,
+		u16 *sec, int sec_count)
+{
+	int i, j;
+
+	/* Construct a vote for each frequency */
+	for (i = 0; i < freqs_count; i++) {
+		u8 pindex = 0, sindex = 0;
+		u32 level = a6xx_gmu_get_arc_level(dev, freqs[i]);
+
+		/* Get the primary index that matches the arc level */
+		for (j = 0; j < pri_count; j++) {
+			if (pri[j] >= level) {
+				pindex = j;
+				break;
+			}
+		}
+
+		if (j == pri_count) {
+			dev_err(dev,
+				"Level %u not found in in the RPMh list\n",
+					level);
+			dev_err(dev, "Available levels:\n");
+			for (j = 0; j < pri_count; j++)
+				dev_err(dev, "  %u\n", pri[j]);
+
+			return -EINVAL;
+		}
+
+		/*
+		 * Look for a level in in the secondary list that matches. If
+		 * nothing fits, use the maximum non zero vote
+		 */
+
+		for (j = 0; j < sec_count; j++) {
+			if (sec[j] >= level) {
+				sindex = j;
+				break;
+			} else if (sec[j]) {
+				sindex = j;
+			}
+		}
+
+		/* Construct the vote */
+		votes[i] = ((pri[pindex] & 0xffff) << 16) |
+			(sindex << 8) | pindex;
+	}
+
+	return 0;
+}
+
+/*
+ * The GMU votes with the RPMh for itself and on behalf of the GPU but we need
+ * to construct the list of votes on the CPU and send it over. Query the RPMh
+ * voltage levels and build the votes
+ */
+
+static int a6xx_gmu_rpmh_votes_init(struct a6xx_gmu *gmu)
+{
+	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
+	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
+	struct msm_gpu *gpu = &adreno_gpu->base;
+
+	u16 gx[16], cx[16], mx[16];
+	u32 gxcount, cxcount, mxcount;
+	int ret;
+
+	/* Get the list of available voltage levels for each component */
+	gxcount = a6xx_gmu_rpmh_arc_cmds("gfx.lvl", gx, sizeof(gx));
+	cxcount = a6xx_gmu_rpmh_arc_cmds("cx.lvl", cx, sizeof(cx));
+	mxcount = a6xx_gmu_rpmh_arc_cmds("mx.lvl", mx, sizeof(mx));
+
+	/* Build the GX votes */
+	ret = a6xx_gmu_rpmh_arc_votes_init(&gpu->pdev->dev, gmu->gx_arc_votes,
+		gmu->gpu_freqs, gmu->nr_gpu_freqs,
+		gx, gxcount, mx, mxcount);
+
+	/* Build the CX votes */
+	ret |= a6xx_gmu_rpmh_arc_votes_init(gmu->dev, gmu->cx_arc_votes,
+		gmu->gmu_freqs, gmu->nr_gmu_freqs,
+		cx, cxcount, mx, mxcount);
+
+	return ret;
+}
+
+static int a6xx_gmu_build_freq_table(struct device *dev, unsigned long *freqs,
+		u32 size)
+{
+	int count = dev_pm_opp_get_opp_count(dev);
+	struct dev_pm_opp *opp;
+	int i, index = 0;
+	unsigned long freq = 1;
+
+	/*
+	 * The OPP table doesn't contain the "off" frequency level so we need to
+	 * add 1 to the table size to account for it
+	 */
+
+	if (WARN(count + 1 > size,
+		"The GMU frequency table is being truncated\n"))
+		count = size - 1;
+
+	/* Set the "off" frequency */
+	freqs[index++] = 0;
+
+	for (i = 0; i < count; i++) {
+		opp = dev_pm_opp_find_freq_ceil(dev, &freq);
+		if (IS_ERR(opp))
+			break;
+
+		dev_pm_opp_put(opp);
+		freqs[index++] = freq++;
+	}
+
+	return index;
+}
+
+static int a6xx_gmu_pwrlevels_probe(struct a6xx_gmu *gmu)
+{
+	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
+	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
+	struct msm_gpu *gpu = &adreno_gpu->base;
+
+	int ret = 0;
+
+	/*
+	 * The GMU handles its own frequency switching so build a list of
+	 * available frequencies to send during initalization
+	 */
+	ret = dev_pm_opp_of_add_table(gmu->dev);
+	if (ret) {
+		dev_err(gmu->dev, "Unable to set the OPP table for the GMU\n");
+		return ret;
+	}
+
+	gmu->nr_gmu_freqs = a6xx_gmu_build_freq_table(gmu->dev,
+		gmu->gmu_freqs, ARRAY_SIZE(gmu->gmu_freqs));
+
+	/*
+	 * The GMU also handles GPU frequency switching so build a list
+	 * from the GPU OPP table
+	 */
+	gmu->nr_gpu_freqs = a6xx_gmu_build_freq_table(&gpu->pdev->dev,
+		gmu->gpu_freqs, ARRAY_SIZE(gmu->gpu_freqs));
+
+	/* Build the list of RPMh votes that we'll send to the GMU */
+	return a6xx_gmu_rpmh_votes_init(gmu);
+}
+
+static int a6xx_gmu_clocks_probe(struct a6xx_gmu *gmu)
+{
+	int ret = msm_clk_bulk_get(gmu->dev, &gmu->clocks);
+
+	if (ret < 1)
+		return ret;
+
+	gmu->nr_clocks = ret;
+
+	gmu->core_clk = msm_clk_bulk_get_clock(gmu->clocks,
+		gmu->nr_clocks, "gmu");
+
+	return 0;
+}
+
+static void __iomem *a6xx_gmu_get_mmio(struct platform_device *pdev,
+		const char *name)
+{
+	void __iomem *ret;
+	struct resource *res = platform_get_resource_byname(pdev,
+			IORESOURCE_MEM, name);
+
+	if (!res) {
+		dev_err(&pdev->dev, "Unable to find the %s registers\n", name);
+		return ERR_PTR(-EINVAL);
+	}
+
+	ret = devm_ioremap(&pdev->dev, res->start, resource_size(res));
+	if (!ret) {
+		dev_err(&pdev->dev, "Unable to map the %s registers\n", name);
+		return ERR_PTR(-EINVAL);
+	}
+
+	return ret;
+}
+
+static int a6xx_gmu_get_irq(struct a6xx_gmu *gmu, struct platform_device *pdev,
+		const char *name, irq_handler_t handler)
+{
+	int irq, ret;
+
+	irq = platform_get_irq_byname(pdev, name);
+
+	ret = devm_request_irq(&pdev->dev, irq, handler, IRQF_TRIGGER_HIGH,
+		name, gmu);
+	if (ret) {
+		dev_err(&pdev->dev, "Unable to get interrupt %s\n", name);
+		return ret;
+	}
+
+	disable_irq(irq);
+
+	return irq;
+}
+
+void a6xx_gmu_remove(struct a6xx_gpu *a6xx_gpu)
+{
+	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
+
+	if (IS_ERR_OR_NULL(gmu->mmio))
+		return;
+
+	pm_runtime_disable(gmu->dev);
+	a6xx_gmu_stop(a6xx_gpu);
+
+	a6xx_gmu_irq_disable(gmu);
+	a6xx_gmu_memory_free(gmu, gmu->hfi);
+
+	iommu_detach_device(gmu->domain, gmu->dev);
+
+	iommu_domain_free(gmu->domain);
+}
+
+int a6xx_gmu_probe(struct a6xx_gpu *a6xx_gpu, struct device_node *node)
+{
+	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
+	struct platform_device *pdev = of_find_device_by_node(node);
+	int ret;
+
+	if (!pdev)
+		return -ENODEV;
+
+	gmu->dev = &pdev->dev;
+
+	of_dma_configure(gmu->dev, node, false);
+
+	/* Fow now, don't do anything fancy until we get our feet under us */
+	gmu->idle_level = GMU_IDLE_STATE_ACTIVE;
+
+	pm_runtime_enable(gmu->dev);
+	gmu->gx = devm_regulator_get(gmu->dev, "vdd");
+
+	/* Get the list of clocks */
+	ret = a6xx_gmu_clocks_probe(gmu);
+	if (ret)
+		return ret;
+
+	/* Set up the IOMMU context bank */
+	ret = a6xx_gmu_memory_probe(gmu);
+	if (ret)
+		return ret;
+
+	/* Allocate memory for for the HFI queues */
+	gmu->hfi = a6xx_gmu_memory_alloc(gmu, SZ_16K);
+	if (IS_ERR(gmu->hfi))
+		goto err;
+
+	/* Allocate memory for the GMU debug region */
+	gmu->debug = a6xx_gmu_memory_alloc(gmu, SZ_16K);
+	if (IS_ERR(gmu->debug))
+		goto err;
+
+	/* Map the GMU registers */
+	gmu->mmio = a6xx_gmu_get_mmio(pdev, "gmu");
+
+	/* Map the GPU power domain controller registers */
+	gmu->pdc_mmio = a6xx_gmu_get_mmio(pdev, "gmu_pdc");
+
+	if (IS_ERR(gmu->mmio) || IS_ERR(gmu->pdc_mmio))
+		goto err;
+
+	/* Get the HFI and GMU interrupts */
+	gmu->hfi_irq = a6xx_gmu_get_irq(gmu, pdev, "hfi", a6xx_hfi_irq);
+	gmu->gmu_irq = a6xx_gmu_get_irq(gmu, pdev, "gmu", a6xx_gmu_irq);
+
+	if (gmu->hfi_irq < 0 || gmu->gmu_irq < 0)
+		goto err;
+
+	/* Set up a tasklet to handle GMU HFI responses */
+	tasklet_init(&gmu->hfi_tasklet, a6xx_hfi_task, (unsigned long) gmu);
+
+	/* Get the power levels for the GMU and GPU */
+	a6xx_gmu_pwrlevels_probe(gmu);
+
+	/* Set up the HFI queues */
+	a6xx_hfi_init(gmu);
+
+	return 0;
+err:
+	a6xx_gmu_memory_free(gmu, gmu->hfi);
+
+	if (gmu->domain) {
+		iommu_detach_device(gmu->domain, gmu->dev);
+
+		iommu_domain_free(gmu->domain);
+	}
+
+	return -ENODEV;
+}
diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gmu.h b/drivers/gpu/drm/msm/adreno/a6xx_gmu.h
new file mode 100644
index 000000000000..d9a386c18799
--- /dev/null
+++ b/drivers/gpu/drm/msm/adreno/a6xx_gmu.h
@@ -0,0 +1,162 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/* Copyright (c) 2017 The Linux Foundation. All rights reserved. */
+
+#ifndef _A6XX_GMU_H_
+#define _A6XX_GMU_H_
+
+#include <linux/interrupt.h>
+#include "msm_drv.h"
+#include "a6xx_hfi.h"
+
+struct a6xx_gmu_bo {
+	void *virt;
+	size_t size;
+	u64 iova;
+	struct page **pages;
+};
+
+/*
+ * These define the different GMU wake up options - these define how both the
+ * CPU and the GMU bring up the hardware
+ */
+
+/* THe GMU has already been booted and the rentention registers are active */
+#define GMU_WARM_BOOT 0
+
+/* the GMU is coming up for the first time or back from a power collapse */
+#define GMU_COLD_BOOT 1
+
+/* The GMU is being soft reset after a fault */
+#define GMU_RESET 2
+
+/*
+ * These define the level of control that the GMU has - the higher the number
+ * the more things that the GMU hardware controls on its own.
+ */
+
+/* The GMU does not do any idle state management */
+#define GMU_IDLE_STATE_ACTIVE 0
+
+/* The GMU manages SPTP power collapse */
+#define GMU_IDLE_STATE_SPTP 2
+
+/* The GMU does automatic IFPC (intra-frame power collapse) */
+#define GMU_IDLE_STATE_IFPC 3
+
+struct a6xx_gmu {
+	struct device *dev;
+
+	void * __iomem mmio;
+	void * __iomem pdc_mmio;
+
+	int hfi_irq;
+	int gmu_irq;
+
+	struct regulator *gx;
+
+	struct iommu_domain *domain;
+	u64 uncached_iova_base;
+
+	int idle_level;
+
+	struct a6xx_gmu_bo *hfi;
+	struct a6xx_gmu_bo *debug;
+
+	int nr_clocks;
+	struct clk_bulk_data *clocks;
+	struct clk *core_clk;
+
+	int nr_gpu_freqs;
+	unsigned long gpu_freqs[16];
+	u32 gx_arc_votes[16];
+
+	int nr_gmu_freqs;
+	unsigned long gmu_freqs[4];
+	u32 cx_arc_votes[4];
+
+	struct a6xx_hfi_queue queues[2];
+
+	struct tasklet_struct hfi_tasklet;
+};
+
+static inline u32 gmu_read(struct a6xx_gmu *gmu, u32 offset)
+{
+	return msm_readl(gmu->mmio + (offset << 2));
+}
+
+static inline void gmu_write(struct a6xx_gmu *gmu, u32 offset, u32 value)
+{
+	return msm_writel(value, gmu->mmio + (offset << 2));
+}
+
+static inline void pdc_write(struct a6xx_gmu *gmu, u32 offset, u32 value)
+{
+	return msm_writel(value, gmu->pdc_mmio + (offset << 2));
+}
+
+static inline void gmu_rmw(struct a6xx_gmu *gmu, u32 reg, u32 mask, u32 or)
+{
+	u32 val = gmu_read(gmu, reg);
+
+	val &= ~mask;
+
+	gmu_write(gmu, reg, val | or);
+}
+
+#define gmu_poll_timeout(gmu, addr, val, cond, interval, timeout) \
+	readl_poll_timeout((gmu)->mmio + ((addr) << 2), val, cond, \
+		interval, timeout)
+
+/*
+ * These are the available OOB (out of band requests) to the GMU where "out of
+ * band" means that the CPU talks to the GMU directly and not through HFI.
+ * Normally this works by writing a ITCM/DTCM register and then triggering a
+ * interrupt (the "request" bit) and waiting for an acknowledgment (the "ack"
+ * bit). The state is cleared by writing the "clear' bit to the GMU interrupt.
+ *
+ * These are used to force the GMU/GPU to stay on during a critical sequence or
+ * for hardware workarounds.
+ */
+
+enum a6xx_gmu_oob_state {
+	GMU_OOB_BOOT_SLUMBER = 0,
+	GMU_OOB_GPU_SET,
+	GMU_OOB_DCVS_SET,
+};
+
+/* These are the interrupt / ack bits for each OOB request that are set
+ * in a6xx_gmu_set_oob and a6xx_clear_oob
+ */
+
+/*
+ * Let the GMU know that a boot or slumber operation has started. The value in
+ * REG_A6XX_GMU_BOOT_SLUMBER_OPTION lets the GMU know which operation we are
+ * doing
+ */
+#define GMU_OOB_BOOT_SLUMBER_REQUEST	22
+#define GMU_OOB_BOOT_SLUMBER_ACK	30
+#define GMU_OOB_BOOT_SLUMBER_CLEAR	30
+
+/*
+ * Set a new power level for the GPU when the CPU is doing frequency scaling
+ */
+#define GMU_OOB_DCVS_REQUEST	23
+#define GMU_OOB_DCVS_ACK	31
+#define GMU_OOB_DCVS_CLEAR	31
+
+/*
+ * Let the GMU know to not turn off any GPU registers while the CPU is in a
+ * critical section
+ */
+#define GMU_OOB_GPU_SET_REQUEST	16
+#define GMU_OOB_GPU_SET_ACK	24
+#define GMU_OOB_GPU_SET_CLEAR	24
+
+
+void a6xx_hfi_init(struct a6xx_gmu *gmu);
+int a6xx_hfi_start(struct a6xx_gmu *gmu, int boot_state);
+void a6xx_hfi_stop(struct a6xx_gmu *gmu);
+
+void a6xx_hfi_task(unsigned long data);
+
+#endif
diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
new file mode 100644
index 000000000000..c629f742a1d1
--- /dev/null
+++ b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
@@ -0,0 +1,818 @@
+// SPDX-License-Identifier: GPL-2.0
+/* Copyright (c) 2017-2018 The Linux Foundation. All rights reserved. */
+
+
+#include "msm_gem.h"
+#include "msm_mmu.h"
+#include "a6xx_gpu.h"
+#include "a6xx_gmu.xml.h"
+
+static inline bool _a6xx_check_idle(struct msm_gpu *gpu)
+{
+	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
+	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
+
+	/* Check that the GMU is idle */
+	if (!a6xx_gmu_isidle(&a6xx_gpu->gmu))
+		return false;
+
+	/* Check tha the CX master is idle */
+	if (gpu_read(gpu, REG_A6XX_RBBM_STATUS) &
+			~A6XX_RBBM_STATUS_CP_AHB_BUSY_CX_MASTER)
+		return false;
+
+	return !(gpu_read(gpu, REG_A6XX_RBBM_INT_0_STATUS) &
+		A6XX_RBBM_INT_0_MASK_RBBM_HANG_DETECT);
+}
+
+bool a6xx_idle(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
+{
+	/* wait for CP to drain ringbuffer: */
+	if (!adreno_idle(gpu, ring))
+		return false;
+
+	if (spin_until(_a6xx_check_idle(gpu))) {
+		DRM_ERROR("%s: %ps: timeout waiting for GPU to idle: status %8.8X irq %8.8X rptr/wptr %d/%d\n",
+			gpu->name, __builtin_return_address(0),
+			gpu_read(gpu, REG_A6XX_RBBM_STATUS),
+			gpu_read(gpu, REG_A6XX_RBBM_INT_0_STATUS),
+			gpu_read(gpu, REG_A6XX_CP_RB_RPTR),
+			gpu_read(gpu, REG_A6XX_CP_RB_WPTR));
+		return false;
+	}
+
+	return true;
+}
+
+static void a6xx_flush(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
+{
+	uint32_t wptr;
+	unsigned long flags;
+
+	spin_lock_irqsave(&ring->lock, flags);
+
+	/* Copy the shadow to the actual register */
+	ring->cur = ring->next;
+
+	/* Make sure to wrap wptr if we need to */
+	wptr = get_wptr(ring);
+
+	spin_unlock_irqrestore(&ring->lock, flags);
+
+	/* Make sure everything is posted before making a decision */
+	mb();
+
+	gpu_write(gpu, REG_A6XX_CP_RB_WPTR, wptr);
+}
+
+static void a6xx_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit,
+	struct msm_file_private *ctx)
+{
+	struct msm_drm_private *priv = gpu->dev->dev_private;
+	struct msm_ringbuffer *ring = submit->ring;
+	unsigned int i;
+
+	/* Invalidate CCU depth and color */
+	OUT_PKT7(ring, CP_EVENT_WRITE, 1);
+	OUT_RING(ring, PC_CCU_INVALIDATE_DEPTH);
+
+	OUT_PKT7(ring, CP_EVENT_WRITE, 1);
+	OUT_RING(ring, PC_CCU_INVALIDATE_COLOR);
+
+	/* Submit the commands */
+	for (i = 0; i < submit->nr_cmds; i++) {
+		switch (submit->cmd[i].type) {
+		case MSM_SUBMIT_CMD_IB_TARGET_BUF:
+			break;
+		case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
+			if (priv->lastctx == ctx)
+				break;
+		case MSM_SUBMIT_CMD_BUF:
+			OUT_PKT7(ring, CP_INDIRECT_BUFFER_PFE, 3);
+			OUT_RING(ring, lower_32_bits(submit->cmd[i].iova));
+			OUT_RING(ring, upper_32_bits(submit->cmd[i].iova));
+			OUT_RING(ring, submit->cmd[i].size);
+			break;
+		}
+	}
+
+	/* Write the fence to the scratch register */
+	OUT_PKT4(ring, REG_A6XX_CP_SCRATCH_REG(2), 1);
+	OUT_RING(ring, submit->seqno);
+
+	/*
+	 * Execute a CACHE_FLUSH_TS event. This will ensure that the
+	 * timestamp is written to the memory and then triggers the interrupt
+	 */
+	OUT_PKT7(ring, CP_EVENT_WRITE, 4);
+	OUT_RING(ring, CACHE_FLUSH_TS | (1 << 31));
+	OUT_RING(ring, lower_32_bits(rbmemptr(ring, fence)));
+	OUT_RING(ring, upper_32_bits(rbmemptr(ring, fence)));
+	OUT_RING(ring, submit->seqno);
+
+	a6xx_flush(gpu, ring);
+}
+
+static const struct {
+	u32 offset;
+	u32 value;
+} a6xx_hwcg[] = {
+	{REG_A6XX_RBBM_CLOCK_CNTL_SP0, 0x22222222},
+	{REG_A6XX_RBBM_CLOCK_CNTL_SP1, 0x22222222},
+	{REG_A6XX_RBBM_CLOCK_CNTL_SP2, 0x22222222},
+	{REG_A6XX_RBBM_CLOCK_CNTL_SP3, 0x22222222},
+	{REG_A6XX_RBBM_CLOCK_CNTL2_SP0, 0x02022220},
+	{REG_A6XX_RBBM_CLOCK_CNTL2_SP1, 0x02022220},
+	{REG_A6XX_RBBM_CLOCK_CNTL2_SP2, 0x02022220},
+	{REG_A6XX_RBBM_CLOCK_CNTL2_SP3, 0x02022220},
+	{REG_A6XX_RBBM_CLOCK_DELAY_SP0, 0x00000080},
+	{REG_A6XX_RBBM_CLOCK_DELAY_SP1, 0x00000080},
+	{REG_A6XX_RBBM_CLOCK_DELAY_SP2, 0x00000080},
+	{REG_A6XX_RBBM_CLOCK_DELAY_SP3, 0x00000080},
+	{REG_A6XX_RBBM_CLOCK_HYST_SP0, 0x0000f3cf},
+	{REG_A6XX_RBBM_CLOCK_HYST_SP1, 0x0000f3cf},
+	{REG_A6XX_RBBM_CLOCK_HYST_SP2, 0x0000f3cf},
+	{REG_A6XX_RBBM_CLOCK_HYST_SP3, 0x0000f3cf},
+	{REG_A6XX_RBBM_CLOCK_CNTL_TP0, 0x02222222},
+	{REG_A6XX_RBBM_CLOCK_CNTL_TP1, 0x02222222},
+	{REG_A6XX_RBBM_CLOCK_CNTL_TP2, 0x02222222},
+	{REG_A6XX_RBBM_CLOCK_CNTL_TP3, 0x02222222},
+	{REG_A6XX_RBBM_CLOCK_CNTL2_TP0, 0x22222222},
+	{REG_A6XX_RBBM_CLOCK_CNTL2_TP1, 0x22222222},
+	{REG_A6XX_RBBM_CLOCK_CNTL2_TP2, 0x22222222},
+	{REG_A6XX_RBBM_CLOCK_CNTL2_TP3, 0x22222222},
+	{REG_A6XX_RBBM_CLOCK_CNTL3_TP0, 0x22222222},
+	{REG_A6XX_RBBM_CLOCK_CNTL3_TP1, 0x22222222},
+	{REG_A6XX_RBBM_CLOCK_CNTL3_TP2, 0x22222222},
+	{REG_A6XX_RBBM_CLOCK_CNTL3_TP3, 0x22222222},
+	{REG_A6XX_RBBM_CLOCK_CNTL4_TP0, 0x00022222},
+	{REG_A6XX_RBBM_CLOCK_CNTL4_TP1, 0x00022222},
+	{REG_A6XX_RBBM_CLOCK_CNTL4_TP2, 0x00022222},
+	{REG_A6XX_RBBM_CLOCK_CNTL4_TP3, 0x00022222},
+	{REG_A6XX_RBBM_CLOCK_HYST_TP0, 0x77777777},
+	{REG_A6XX_RBBM_CLOCK_HYST_TP1, 0x77777777},
+	{REG_A6XX_RBBM_CLOCK_HYST_TP2, 0x77777777},
+	{REG_A6XX_RBBM_CLOCK_HYST_TP3, 0x77777777},
+	{REG_A6XX_RBBM_CLOCK_HYST2_TP0, 0x77777777},
+	{REG_A6XX_RBBM_CLOCK_HYST2_TP1, 0x77777777},
+	{REG_A6XX_RBBM_CLOCK_HYST2_TP2, 0x77777777},
+	{REG_A6XX_RBBM_CLOCK_HYST2_TP3, 0x77777777},
+	{REG_A6XX_RBBM_CLOCK_HYST3_TP0, 0x77777777},
+	{REG_A6XX_RBBM_CLOCK_HYST3_TP1, 0x77777777},
+	{REG_A6XX_RBBM_CLOCK_HYST3_TP2, 0x77777777},
+	{REG_A6XX_RBBM_CLOCK_HYST3_TP3, 0x77777777},
+	{REG_A6XX_RBBM_CLOCK_HYST4_TP0, 0x00077777},
+	{REG_A6XX_RBBM_CLOCK_HYST4_TP1, 0x00077777},
+	{REG_A6XX_RBBM_CLOCK_HYST4_TP2, 0x00077777},
+	{REG_A6XX_RBBM_CLOCK_HYST4_TP3, 0x00077777},
+	{REG_A6XX_RBBM_CLOCK_DELAY_TP0, 0x11111111},
+	{REG_A6XX_RBBM_CLOCK_DELAY_TP1, 0x11111111},
+	{REG_A6XX_RBBM_CLOCK_DELAY_TP2, 0x11111111},
+	{REG_A6XX_RBBM_CLOCK_DELAY_TP3, 0x11111111},
+	{REG_A6XX_RBBM_CLOCK_DELAY2_TP0, 0x11111111},
+	{REG_A6XX_RBBM_CLOCK_DELAY2_TP1, 0x11111111},
+	{REG_A6XX_RBBM_CLOCK_DELAY2_TP2, 0x11111111},
+	{REG_A6XX_RBBM_CLOCK_DELAY2_TP3, 0x11111111},
+	{REG_A6XX_RBBM_CLOCK_DELAY3_TP0, 0x11111111},
+	{REG_A6XX_RBBM_CLOCK_DELAY3_TP1, 0x11111111},
+	{REG_A6XX_RBBM_CLOCK_DELAY3_TP2, 0x11111111},
+	{REG_A6XX_RBBM_CLOCK_DELAY3_TP3, 0x11111111},
+	{REG_A6XX_RBBM_CLOCK_DELAY4_TP0, 0x00011111},
+	{REG_A6XX_RBBM_CLOCK_DELAY4_TP1, 0x00011111},
+	{REG_A6XX_RBBM_CLOCK_DELAY4_TP2, 0x00011111},
+	{REG_A6XX_RBBM_CLOCK_DELAY4_TP3, 0x00011111},
+	{REG_A6XX_RBBM_CLOCK_CNTL_UCHE, 0x22222222},
+	{REG_A6XX_RBBM_CLOCK_CNTL2_UCHE, 0x22222222},
+	{REG_A6XX_RBBM_CLOCK_CNTL3_UCHE, 0x22222222},
+	{REG_A6XX_RBBM_CLOCK_CNTL4_UCHE, 0x00222222},
+	{REG_A6XX_RBBM_CLOCK_HYST_UCHE, 0x00000004},
+	{REG_A6XX_RBBM_CLOCK_DELAY_UCHE, 0x00000002},
+	{REG_A6XX_RBBM_CLOCK_CNTL_RB0, 0x22222222},
+	{REG_A6XX_RBBM_CLOCK_CNTL_RB1, 0x22222222},
+	{REG_A6XX_RBBM_CLOCK_CNTL_RB2, 0x22222222},
+	{REG_A6XX_RBBM_CLOCK_CNTL_RB3, 0x22222222},
+	{REG_A6XX_RBBM_CLOCK_CNTL2_RB0, 0x00002222},
+	{REG_A6XX_RBBM_CLOCK_CNTL2_RB1, 0x00002222},
+	{REG_A6XX_RBBM_CLOCK_CNTL2_RB2, 0x00002222},
+	{REG_A6XX_RBBM_CLOCK_CNTL2_RB3, 0x00002222},
+	{REG_A6XX_RBBM_CLOCK_CNTL_CCU0, 0x00002220},
+	{REG_A6XX_RBBM_CLOCK_CNTL_CCU1, 0x00002220},
+	{REG_A6XX_RBBM_CLOCK_CNTL_CCU2, 0x00002220},
+	{REG_A6XX_RBBM_CLOCK_CNTL_CCU3, 0x00002220},
+	{REG_A6XX_RBBM_CLOCK_HYST_RB_CCU0, 0x00040f00},
+	{REG_A6XX_RBBM_CLOCK_HYST_RB_CCU1, 0x00040f00},
+	{REG_A6XX_RBBM_CLOCK_HYST_RB_CCU2, 0x00040f00},
+	{REG_A6XX_RBBM_CLOCK_HYST_RB_CCU3, 0x00040f00},
+	{REG_A6XX_RBBM_CLOCK_CNTL_RAC, 0x05022022},
+	{REG_A6XX_RBBM_CLOCK_CNTL2_RAC, 0x00005555},
+	{REG_A6XX_RBBM_CLOCK_DELAY_RAC, 0x00000011},
+	{REG_A6XX_RBBM_CLOCK_HYST_RAC, 0x00445044},
+	{REG_A6XX_RBBM_CLOCK_CNTL_TSE_RAS_RBBM, 0x04222222},
+	{REG_A6XX_RBBM_CLOCK_MODE_GPC, 0x00222222},
+	{REG_A6XX_RBBM_CLOCK_MODE_VFD, 0x00002222},
+	{REG_A6XX_RBBM_CLOCK_HYST_TSE_RAS_RBBM, 0x00000000},
+	{REG_A6XX_RBBM_CLOCK_HYST_GPC, 0x04104004},
+	{REG_A6XX_RBBM_CLOCK_HYST_VFD, 0x00000000},
+	{REG_A6XX_RBBM_CLOCK_DELAY_HLSQ, 0x00000000},
+	{REG_A6XX_RBBM_CLOCK_DELAY_TSE_RAS_RBBM, 0x00004000},
+	{REG_A6XX_RBBM_CLOCK_DELAY_GPC, 0x00000200},
+	{REG_A6XX_RBBM_CLOCK_DELAY_VFD, 0x00002222},
+	{REG_A6XX_RBBM_CLOCK_DELAY_HLSQ_2, 0x00000002},
+	{REG_A6XX_RBBM_CLOCK_MODE_HLSQ, 0x00002222},
+	{REG_A6XX_RBBM_CLOCK_CNTL_GMU_GX, 0x00000222},
+	{REG_A6XX_RBBM_CLOCK_DELAY_GMU_GX, 0x00000111},
+	{REG_A6XX_RBBM_CLOCK_HYST_GMU_GX, 0x00000555}
+};
+
+static void a6xx_set_hwcg(struct msm_gpu *gpu, bool state)
+{
+	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
+	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
+	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
+	unsigned int i;
+	u32 val;
+
+	val = gpu_read(gpu, REG_A6XX_RBBM_CLOCK_CNTL);
+
+	/* Don't re-program the registers if they are already correct */
+	if ((!state && !val) || (state && (val == 0x8aa8aa02)))
+		return;
+
+	/* Disable SP clock before programming HWCG registers */
+	gmu_rmw(gmu, REG_A6XX_GPU_GMU_GX_SPTPRAC_CLOCK_CONTROL, 1, 0);
+
+	for (i = 0; i < ARRAY_SIZE(a6xx_hwcg); i++)
+		gpu_write(gpu, a6xx_hwcg[i].offset,
+			state ? a6xx_hwcg[i].value : 0);
+
+	/* Enable SP clock */
+	gmu_rmw(gmu, REG_A6XX_GPU_GMU_GX_SPTPRAC_CLOCK_CONTROL, 0, 1);
+
+	gpu_write(gpu, REG_A6XX_RBBM_CLOCK_CNTL, state ? 0x8aa8aa02 : 0);
+}
+
+static int a6xx_cp_init(struct msm_gpu *gpu)
+{
+	struct msm_ringbuffer *ring = gpu->rb[0];
+
+	OUT_PKT7(ring, CP_ME_INIT, 8);
+
+	OUT_RING(ring, 0x0000002f);
+
+	/* Enable multiple hardware contexts */
+	OUT_RING(ring, 0x00000003);
+
+	/* Enable error detection */
+	OUT_RING(ring, 0x20000000);
+
+	/* Don't enable header dump */
+	OUT_RING(ring, 0x00000000);
+	OUT_RING(ring, 0x00000000);
+
+	/* No workarounds enabled */
+	OUT_RING(ring, 0x00000000);
+
+	/* Pad rest of the cmds with 0's */
+	OUT_RING(ring, 0x00000000);
+	OUT_RING(ring, 0x00000000);
+
+	a6xx_flush(gpu, ring);
+	return a6xx_idle(gpu, ring) ? 0 : -EINVAL;
+}
+
+static int a6xx_ucode_init(struct msm_gpu *gpu)
+{
+	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
+	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
+
+	if (!a6xx_gpu->sqe_bo) {
+		a6xx_gpu->sqe_bo = adreno_fw_create_bo(gpu,
+			adreno_gpu->fw[ADRENO_FW_SQE], &a6xx_gpu->sqe_iova);
+
+		if (IS_ERR(a6xx_gpu->sqe_bo)) {
+			int ret = PTR_ERR(a6xx_gpu->sqe_bo);
+
+			a6xx_gpu->sqe_bo = NULL;
+			DRM_DEV_ERROR(&gpu->pdev->dev,
+				"Could not allocate SQE ucode: %d\n", ret);
+
+			return ret;
+		}
+	}
+
+	gpu_write64(gpu, REG_A6XX_CP_SQE_INSTR_BASE_LO,
+		REG_A6XX_CP_SQE_INSTR_BASE_HI, a6xx_gpu->sqe_iova);
+
+	return 0;
+}
+
+#define A6XX_INT_MASK (A6XX_RBBM_INT_0_MASK_CP_AHB_ERROR | \
+	  A6XX_RBBM_INT_0_MASK_RBBM_ATB_ASYNCFIFO_OVERFLOW | \
+	  A6XX_RBBM_INT_0_MASK_CP_HW_ERROR | \
+	  A6XX_RBBM_INT_0_MASK_CP_IB2 | \
+	  A6XX_RBBM_INT_0_MASK_CP_IB1 | \
+	  A6XX_RBBM_INT_0_MASK_CP_RB | \
+	  A6XX_RBBM_INT_0_MASK_CP_CACHE_FLUSH_TS | \
+	  A6XX_RBBM_INT_0_MASK_RBBM_ATB_BUS_OVERFLOW | \
+	  A6XX_RBBM_INT_0_MASK_RBBM_HANG_DETECT | \
+	  A6XX_RBBM_INT_0_MASK_UCHE_OOB_ACCESS | \
+	  A6XX_RBBM_INT_0_MASK_UCHE_TRAP_INTR)
+
+static int a6xx_hw_init(struct msm_gpu *gpu)
+{
+	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
+	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
+	int ret;
+
+	/* Make sure the GMU keeps the GPU on while we set it up */
+	a6xx_gmu_set_oob(&a6xx_gpu->gmu, GMU_OOB_GPU_SET);
+
+	gpu_write(gpu, REG_A6XX_RBBM_SECVID_TSB_CNTL, 0);
+
+	/*
+	 * Disable the trusted memory range - we don't actually supported secure
+	 * memory rendering at this point in time and we don't want to block off
+	 * part of the virtual memory space.
+	 */
+	gpu_write64(gpu, REG_A6XX_RBBM_SECVID_TSB_TRUSTED_BASE_LO,
+		REG_A6XX_RBBM_SECVID_TSB_TRUSTED_BASE_HI, 0x00000000);
+	gpu_write(gpu, REG_A6XX_RBBM_SECVID_TSB_TRUSTED_SIZE, 0x00000000);
+
+	/* enable hardware clockgating */
+	a6xx_set_hwcg(gpu, true);
+
+	/* VBIF start */
+	gpu_write(gpu, REG_A6XX_VBIF_GATE_OFF_WRREQ_EN, 0x00000009);
+	gpu_write(gpu, REG_A6XX_RBBM_VBIF_CLIENT_QOS_CNTL, 0x3);
+
+	/* Make all blocks contribute to the GPU BUSY perf counter */
+	gpu_write(gpu, REG_A6XX_RBBM_PERFCTR_GPU_BUSY_MASKED, 0xffffffff);
+
+	/* Disable L2 bypass in the UCHE */
+	gpu_write(gpu, REG_A6XX_UCHE_WRITE_RANGE_MAX_LO, 0xffffffc0);
+	gpu_write(gpu, REG_A6XX_UCHE_WRITE_RANGE_MAX_HI, 0x0001ffff);
+	gpu_write(gpu, REG_A6XX_UCHE_TRAP_BASE_LO, 0xfffff000);
+	gpu_write(gpu, REG_A6XX_UCHE_TRAP_BASE_HI, 0x0001ffff);
+	gpu_write(gpu, REG_A6XX_UCHE_WRITE_THRU_BASE_LO, 0xfffff000);
+	gpu_write(gpu, REG_A6XX_UCHE_WRITE_THRU_BASE_HI, 0x0001ffff);
+
+	/* Set the GMEM VA range [0x100000:0x100000 + gpu->gmem - 1] */
+	gpu_write64(gpu, REG_A6XX_UCHE_GMEM_RANGE_MIN_LO,
+		REG_A6XX_UCHE_GMEM_RANGE_MIN_HI, 0x00100000);
+
+	gpu_write64(gpu, REG_A6XX_UCHE_GMEM_RANGE_MAX_LO,
+		REG_A6XX_UCHE_GMEM_RANGE_MAX_HI,
+		0x00100000 + adreno_gpu->gmem - 1);
+
+	gpu_write(gpu, REG_A6XX_UCHE_FILTER_CNTL, 0x804);
+	gpu_write(gpu, REG_A6XX_UCHE_CACHE_WAYS, 0x4);
+
+	gpu_write(gpu, REG_A6XX_CP_ROQ_THRESHOLDS_2, 0x010000c0);
+	gpu_write(gpu, REG_A6XX_CP_ROQ_THRESHOLDS_1, 0x8040362c);
+
+	/* Setting the mem pool size */
+	gpu_write(gpu, REG_A6XX_CP_MEM_POOL_SIZE, 128);
+
+	/* Setting the primFifo thresholds default values */
+	gpu_write(gpu, REG_A6XX_PC_DBG_ECO_CNTL, (0x300 << 11));
+
+	/* Set the AHB default slave response to "ERROR" */
+	gpu_write(gpu, REG_A6XX_CP_AHB_CNTL, 0x1);
+
+	/* Turn on performance counters */
+	gpu_write(gpu, REG_A6XX_RBBM_PERFCTR_CNTL, 0x1);
+
+	/* Select CP0 to always count cycles */
+	gpu_write(gpu, REG_A6XX_CP_PERFCTR_CP_SEL_0, PERF_CP_ALWAYS_COUNT);
+
+	/* FIXME: not sure if this should live here or in a6xx_gmu.c */
+	gmu_write(&a6xx_gpu->gmu,  REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_MASK,
+		0xff000000);
+	gmu_rmw(&a6xx_gpu->gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_SELECT_0,
+		0xff, 0x20);
+	gmu_write(&a6xx_gpu->gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_ENABLE,
+		0x01);
+
+	gpu_write(gpu, REG_A6XX_RB_NC_MODE_CNTL, 2 << 1);
+	gpu_write(gpu, REG_A6XX_TPL1_NC_MODE_CNTL, 2 << 1);
+	gpu_write(gpu, REG_A6XX_SP_NC_MODE_CNTL, 2 << 1);
+	gpu_write(gpu, REG_A6XX_UCHE_MODE_CNTL, 2 << 21);
+
+	/* Enable fault detection */
+	gpu_write(gpu, REG_A6XX_RBBM_INTERFACE_HANG_INT_CNTL,
+		(1 << 30) | 0x1fffff);
+
+	gpu_write(gpu, REG_A6XX_UCHE_CLIENT_PF, 1);
+
+	/* Protect registers from the CP */
+	gpu_write(gpu, REG_A6XX_CP_PROTECT_CNTL, 0x00000003);
+
+	gpu_write(gpu, REG_A6XX_CP_PROTECT(0),
+		A6XX_PROTECT_RDONLY(0x600, 0x51));
+	gpu_write(gpu, REG_A6XX_CP_PROTECT(1), A6XX_PROTECT_RW(0xae50, 0x2));
+	gpu_write(gpu, REG_A6XX_CP_PROTECT(2), A6XX_PROTECT_RW(0x9624, 0x13));
+	gpu_write(gpu, REG_A6XX_CP_PROTECT(3), A6XX_PROTECT_RW(0x8630, 0x8));
+	gpu_write(gpu, REG_A6XX_CP_PROTECT(4), A6XX_PROTECT_RW(0x9e70, 0x1));
+	gpu_write(gpu, REG_A6XX_CP_PROTECT(5), A6XX_PROTECT_RW(0x9e78, 0x187));
+	gpu_write(gpu, REG_A6XX_CP_PROTECT(6), A6XX_PROTECT_RW(0xf000, 0x810));
+	gpu_write(gpu, REG_A6XX_CP_PROTECT(7),
+		A6XX_PROTECT_RDONLY(0xfc00, 0x3));
+	gpu_write(gpu, REG_A6XX_CP_PROTECT(8), A6XX_PROTECT_RW(0x50e, 0x0));
+	gpu_write(gpu, REG_A6XX_CP_PROTECT(9), A6XX_PROTECT_RDONLY(0x50f, 0x0));
+	gpu_write(gpu, REG_A6XX_CP_PROTECT(10), A6XX_PROTECT_RW(0x510, 0x0));
+	gpu_write(gpu, REG_A6XX_CP_PROTECT(11),
+		A6XX_PROTECT_RDONLY(0x0, 0x4f9));
+	gpu_write(gpu, REG_A6XX_CP_PROTECT(12),
+		A6XX_PROTECT_RDONLY(0x501, 0xa));
+	gpu_write(gpu, REG_A6XX_CP_PROTECT(13),
+		A6XX_PROTECT_RDONLY(0x511, 0x44));
+	gpu_write(gpu, REG_A6XX_CP_PROTECT(14), A6XX_PROTECT_RW(0xe00, 0xe));
+	gpu_write(gpu, REG_A6XX_CP_PROTECT(15), A6XX_PROTECT_RW(0x8e00, 0x0));
+	gpu_write(gpu, REG_A6XX_CP_PROTECT(16), A6XX_PROTECT_RW(0x8e50, 0xf));
+	gpu_write(gpu, REG_A6XX_CP_PROTECT(17), A6XX_PROTECT_RW(0xbe02, 0x0));
+	gpu_write(gpu, REG_A6XX_CP_PROTECT(18),
+		A6XX_PROTECT_RW(0xbe20, 0x11f3));
+	gpu_write(gpu, REG_A6XX_CP_PROTECT(19), A6XX_PROTECT_RW(0x800, 0x82));
+	gpu_write(gpu, REG_A6XX_CP_PROTECT(20), A6XX_PROTECT_RW(0x8a0, 0x8));
+	gpu_write(gpu, REG_A6XX_CP_PROTECT(21), A6XX_PROTECT_RW(0x8ab, 0x19));
+	gpu_write(gpu, REG_A6XX_CP_PROTECT(22), A6XX_PROTECT_RW(0x900, 0x4d));
+	gpu_write(gpu, REG_A6XX_CP_PROTECT(23), A6XX_PROTECT_RW(0x98d, 0x76));
+	gpu_write(gpu, REG_A6XX_CP_PROTECT(24),
+			A6XX_PROTECT_RDONLY(0x8d0, 0x23));
+	gpu_write(gpu, REG_A6XX_CP_PROTECT(25),
+			A6XX_PROTECT_RDONLY(0x980, 0x4));
+	gpu_write(gpu, REG_A6XX_CP_PROTECT(26), A6XX_PROTECT_RW(0xa630, 0x0));
+
+	/* Enable interrupts */
+	gpu_write(gpu, REG_A6XX_RBBM_INT_0_MASK, A6XX_INT_MASK);
+
+	ret = adreno_hw_init(gpu);
+	if (ret)
+		goto out;
+
+	ret = a6xx_ucode_init(gpu);
+	if (ret)
+		goto out;
+
+	/* Always come up on rb 0 */
+	a6xx_gpu->cur_ring = gpu->rb[0];
+
+	/* Enable the SQE_to start the CP engine */
+	gpu_write(gpu, REG_A6XX_CP_SQE_CNTL, 1);
+
+	ret = a6xx_cp_init(gpu);
+	if (ret)
+		goto out;
+
+	gpu_write(gpu, REG_A6XX_RBBM_SECVID_TRUST_CNTL, 0x0);
+
+out:
+	/*
+	 * Tell the GMU that we are done touching the GPU and it can start power
+	 * management
+	 */
+	a6xx_gmu_clear_oob(&a6xx_gpu->gmu, GMU_OOB_GPU_SET);
+
+	/* Take the GMU out of its special boot mode */
+	a6xx_gmu_clear_oob(&a6xx_gpu->gmu, GMU_OOB_BOOT_SLUMBER);
+
+	return ret;
+}
+
+static void a6xx_dump(struct msm_gpu *gpu)
+{
+	dev_info(&gpu->pdev->dev, "status:   %08x\n",
+			gpu_read(gpu, REG_A6XX_RBBM_STATUS));
+	adreno_dump(gpu);
+}
+
+#define VBIF_RESET_ACK_TIMEOUT	100
+#define VBIF_RESET_ACK_MASK	0x00f0
+
+static void a6xx_recover(struct msm_gpu *gpu)
+{
+	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
+	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
+	int i;
+
+	adreno_dump_info(gpu);
+
+	for (i = 0; i < 8; i++)
+		dev_info(&gpu->pdev->dev, "CP_SCRATCH_REG%d: %u\n", i,
+			gpu_read(gpu, REG_A6XX_CP_SCRATCH_REG(i)));
+
+	if (hang_debug)
+		a6xx_dump(gpu);
+
+	/*
+	 * Turn off keep alive that might have been enabled by the hang
+	 * interrupt
+	 */
+	gmu_write(&a6xx_gpu->gmu, REG_A6XX_GMU_GMU_PWR_COL_KEEPALIVE, 0);
+
+	gpu->funcs->pm_suspend(gpu);
+	gpu->funcs->pm_resume(gpu);
+
+	msm_gpu_hw_init(gpu);
+}
+
+static int a6xx_fault_handler(void *arg, unsigned long iova, int flags)
+{
+	struct msm_gpu *gpu = arg;
+
+	pr_warn_ratelimited("*** gpu fault: iova=%08lx, flags=%d (%u,%u,%u,%u)\n",
+			iova, flags,
+			gpu_read(gpu, REG_A6XX_CP_SCRATCH_REG(4)),
+			gpu_read(gpu, REG_A6XX_CP_SCRATCH_REG(5)),
+			gpu_read(gpu, REG_A6XX_CP_SCRATCH_REG(6)),
+			gpu_read(gpu, REG_A6XX_CP_SCRATCH_REG(7)));
+
+	return -EFAULT;
+}
+
+static void a6xx_cp_hw_err_irq(struct msm_gpu *gpu)
+{
+	u32 status = gpu_read(gpu, REG_A6XX_CP_INTERRUPT_STATUS);
+
+	if (status & A6XX_CP_INT_CP_OPCODE_ERROR) {
+		u32 val;
+
+		gpu_write(gpu, REG_A6XX_CP_SQE_STAT_ADDR, 1);
+		val = gpu_read(gpu, REG_A6XX_CP_SQE_STAT_DATA);
+		dev_err_ratelimited(&gpu->pdev->dev,
+			"CP | opcode error | possible opcode=0x%8.8X\n",
+			val);
+	}
+
+	if (status & A6XX_CP_INT_CP_UCODE_ERROR)
+		dev_err_ratelimited(&gpu->pdev->dev,
+			"CP ucode error interrupt\n");
+
+	if (status & A6XX_CP_INT_CP_HW_FAULT_ERROR)
+		dev_err_ratelimited(&gpu->pdev->dev, "CP | HW fault | status=0x%8.8X\n",
+			gpu_read(gpu, REG_A6XX_CP_HW_FAULT));
+
+	if (status & A6XX_CP_INT_CP_REGISTER_PROTECTION_ERROR) {
+		u32 val = gpu_read(gpu, REG_A6XX_CP_PROTECT_STATUS);
+
+		dev_err_ratelimited(&gpu->pdev->dev,
+			"CP | protected mode error | %s | addr=0x%8.8X | status=0x%8.8X\n",
+			val & (1 << 20) ? "READ" : "WRITE",
+			(val & 0x3ffff), val);
+	}
+
+	if (status & A6XX_CP_INT_CP_AHB_ERROR)
+		dev_err_ratelimited(&gpu->pdev->dev, "CP AHB error interrupt\n");
+
+	if (status & A6XX_CP_INT_CP_VSD_PARITY_ERROR)
+		dev_err_ratelimited(&gpu->pdev->dev, "CP VSD decoder parity error\n");
+
+	if (status & A6XX_CP_INT_CP_ILLEGAL_INSTR_ERROR)
+		dev_err_ratelimited(&gpu->pdev->dev, "CP illegal instruction error\n");
+
+}
+
+static void a6xx_fault_detect_irq(struct msm_gpu *gpu)
+{
+	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
+	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
+	struct drm_device *dev = gpu->dev;
+	struct msm_drm_private *priv = dev->dev_private;
+	struct msm_ringbuffer *ring = gpu->funcs->active_ring(gpu);
+
+	/*
+	 * Force the GPU to stay on until after we finish
+	 * collecting information
+	 */
+	gmu_write(&a6xx_gpu->gmu, REG_A6XX_GMU_GMU_PWR_COL_KEEPALIVE, 1);
+
+	DRM_DEV_ERROR(&gpu->pdev->dev,
+		"gpu fault ring %d fence %x status %8.8X rb %4.4x/%4.4x ib1 %16.16llX/%4.4x ib2 %16.16llX/%4.4x\n",
+		ring ? ring->id : -1, ring ? ring->seqno : 0,
+		gpu_read(gpu, REG_A6XX_RBBM_STATUS),
+		gpu_read(gpu, REG_A6XX_CP_RB_RPTR),
+		gpu_read(gpu, REG_A6XX_CP_RB_WPTR),
+		gpu_read64(gpu, REG_A6XX_CP_IB1_BASE, REG_A6XX_CP_IB1_BASE_HI),
+		gpu_read(gpu, REG_A6XX_CP_IB1_REM_SIZE),
+		gpu_read64(gpu, REG_A6XX_CP_IB2_BASE, REG_A6XX_CP_IB2_BASE_HI),
+		gpu_read(gpu, REG_A6XX_CP_IB2_REM_SIZE));
+
+	/* Turn off the hangcheck timer to keep it from bothering us */
+	del_timer(&gpu->hangcheck_timer);
+
+	queue_work(priv->wq, &gpu->recover_work);
+}
+
+static irqreturn_t a6xx_irq(struct msm_gpu *gpu)
+{
+	u32 status = gpu_read(gpu, REG_A6XX_RBBM_INT_0_STATUS);
+
+	gpu_write(gpu, REG_A6XX_RBBM_INT_CLEAR_CMD, status);
+
+	if (status & A6XX_RBBM_INT_0_MASK_RBBM_HANG_DETECT)
+		a6xx_fault_detect_irq(gpu);
+
+	if (status & A6XX_RBBM_INT_0_MASK_CP_AHB_ERROR)
+		dev_err_ratelimited(&gpu->pdev->dev, "CP | AHB bus error\n");
+
+	if (status & A6XX_RBBM_INT_0_MASK_CP_HW_ERROR)
+		a6xx_cp_hw_err_irq(gpu);
+
+	if (status & A6XX_RBBM_INT_0_MASK_RBBM_ATB_ASYNCFIFO_OVERFLOW)
+		dev_err_ratelimited(&gpu->pdev->dev, "RBBM | ATB ASYNC overflow\n");
+
+	if (status & A6XX_RBBM_INT_0_MASK_RBBM_ATB_BUS_OVERFLOW)
+		dev_err_ratelimited(&gpu->pdev->dev, "RBBM | ATB bus overflow\n");
+
+	if (status & A6XX_RBBM_INT_0_MASK_UCHE_OOB_ACCESS)
+		dev_err_ratelimited(&gpu->pdev->dev, "UCHE | Out of bounds access\n");
+
+	if (status & A6XX_RBBM_INT_0_MASK_CP_CACHE_FLUSH_TS)
+		msm_gpu_retire(gpu);
+
+	return IRQ_HANDLED;
+}
+
+static const u32 a6xx_register_offsets[REG_ADRENO_REGISTER_MAX] = {
+	REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_BASE, REG_A6XX_CP_RB_BASE),
+	REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_BASE_HI, REG_A6XX_CP_RB_BASE_HI),
+	REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_RPTR_ADDR,
+		REG_A6XX_CP_RB_RPTR_ADDR_LO),
+	REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_RPTR_ADDR_HI,
+		REG_A6XX_CP_RB_RPTR_ADDR_HI),
+	REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_RPTR, REG_A6XX_CP_RB_RPTR),
+	REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_WPTR, REG_A6XX_CP_RB_WPTR),
+	REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_CNTL, REG_A6XX_CP_RB_CNTL),
+};
+
+static const u32 a6xx_registers[] = {
+	0x0000, 0x0002, 0x0010, 0x0010, 0x0012, 0x0012, 0x0018, 0x001b,
+	0x001e, 0x0032, 0x0038, 0x003c, 0x0042, 0x0042, 0x0044, 0x0044,
+	0x0047, 0x0047, 0x0056, 0x0056, 0x00ad, 0x00ae, 0x00b0, 0x00fb,
+	0x0100, 0x011d, 0x0200, 0x020d, 0x0210, 0x0213, 0x0218, 0x023d,
+	0x0400, 0x04f9, 0x0500, 0x0500, 0x0505, 0x050b, 0x050e, 0x0511,
+	0x0533, 0x0533, 0x0540, 0x0555, 0x0800, 0x0808, 0x0810, 0x0813,
+	0x0820, 0x0821, 0x0823, 0x0827, 0x0830, 0x0833, 0x0840, 0x0843,
+	0x084f, 0x086f, 0x0880, 0x088a, 0x08a0, 0x08ab, 0x08c0, 0x08c4,
+	0x08d0, 0x08dd, 0x08f0, 0x08f3, 0x0900, 0x0903, 0x0908, 0x0911,
+	0x0928, 0x093e, 0x0942, 0x094d, 0x0980, 0x0984, 0x098d, 0x0996,
+	0x0998, 0x099e, 0x09a0, 0x09a6, 0x09a8, 0x09ae, 0x09b0, 0x09b1,
+	0x09c2, 0x09c8, 0x0a00, 0x0a03, 0x0c00, 0x0c04, 0x0c06, 0x0c06,
+	0x0c10, 0x0cd9, 0x0e00, 0x0e0e, 0x0e10, 0x0e13, 0x0e17, 0x0e19,
+	0x0e1c, 0x0e2b, 0x0e30, 0x0e32, 0x0e38, 0x0e39, 0x8600, 0x8601,
+	0x8610, 0x861b, 0x8620, 0x8620, 0x8628, 0x862b, 0x8630, 0x8637,
+	0x8e01, 0x8e01, 0x8e04, 0x8e05, 0x8e07, 0x8e08, 0x8e0c, 0x8e0c,
+	0x8e10, 0x8e1c, 0x8e20, 0x8e25, 0x8e28, 0x8e28, 0x8e2c, 0x8e2f,
+	0x8e3b, 0x8e3e, 0x8e40, 0x8e43, 0x8e50, 0x8e5e, 0x8e70, 0x8e77,
+	0x9600, 0x9604, 0x9624, 0x9637, 0x9e00, 0x9e01, 0x9e03, 0x9e0e,
+	0x9e11, 0x9e16, 0x9e19, 0x9e19, 0x9e1c, 0x9e1c, 0x9e20, 0x9e23,
+	0x9e30, 0x9e31, 0x9e34, 0x9e34, 0x9e70, 0x9e72, 0x9e78, 0x9e79,
+	0x9e80, 0x9fff, 0xa600, 0xa601, 0xa603, 0xa603, 0xa60a, 0xa60a,
+	0xa610, 0xa617, 0xa630, 0xa630,
+	~0
+};
+
+static int a6xx_pm_resume(struct msm_gpu *gpu)
+{
+	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
+	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
+	int ret;
+
+	ret = a6xx_gmu_resume(a6xx_gpu);
+
+	gpu->needs_hw_init = true;
+
+	return ret;
+}
+
+static int a6xx_pm_suspend(struct msm_gpu *gpu)
+{
+	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
+	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
+
+	/*
+	 * Make sure the GMU is idle before continuing (because some transitions
+	 * may use VBIF
+	 */
+	a6xx_gmu_wait_for_idle(a6xx_gpu);
+
+	/* Clear the VBIF pipe before shutting down */
+	/* FIXME: This accesses the GPU - do we need to make sure it is on? */
+	gpu_write(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL0, 0xf);
+	spin_until((gpu_read(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL1) & 0xf) == 0xf);
+	gpu_write(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL0, 0);
+
+	return a6xx_gmu_stop(a6xx_gpu);
+}
+
+static int a6xx_get_timestamp(struct msm_gpu *gpu, uint64_t *value)
+{
+	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
+	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
+
+	/* Force the GPU power on so we can read this register */
+	a6xx_gmu_set_oob(&a6xx_gpu->gmu, GMU_OOB_GPU_SET);
+
+	*value = gpu_read64(gpu, REG_A6XX_RBBM_PERFCTR_CP_0_LO,
+		REG_A6XX_RBBM_PERFCTR_CP_0_HI);
+
+	a6xx_gmu_clear_oob(&a6xx_gpu->gmu, GMU_OOB_GPU_SET);
+	return 0;
+}
+
+#if defined(CONFIG_DEBUG_FS) || defined(CONFIG_DEV_COREDUMP)
+static void a6xx_show(struct msm_gpu *gpu, struct msm_gpu_state *state,
+		struct drm_printer *p)
+{
+	adreno_show(gpu, state, p);
+}
+#endif
+
+static struct msm_ringbuffer *a6xx_active_ring(struct msm_gpu *gpu)
+{
+	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
+	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
+
+	return a6xx_gpu->cur_ring;
+}
+
+static void a6xx_destroy(struct msm_gpu *gpu)
+{
+	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
+	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
+
+	if (a6xx_gpu->sqe_bo) {
+		if (a6xx_gpu->sqe_iova)
+			msm_gem_put_iova(a6xx_gpu->sqe_bo, gpu->aspace);
+		drm_gem_object_unreference_unlocked(a6xx_gpu->sqe_bo);
+	}
+
+	a6xx_gmu_remove(a6xx_gpu);
+
+	adreno_gpu_cleanup(adreno_gpu);
+	kfree(a6xx_gpu);
+}
+
+static const struct adreno_gpu_funcs funcs = {
+	.base = {
+		.get_param = adreno_get_param,
+		.hw_init = a6xx_hw_init,
+		.pm_suspend = a6xx_pm_suspend,
+		.pm_resume = a6xx_pm_resume,
+		.recover = a6xx_recover,
+		.submit = a6xx_submit,
+		.flush = a6xx_flush,
+		.active_ring = a6xx_active_ring,
+		.irq = a6xx_irq,
+		.destroy = a6xx_destroy,
+#if defined(CONFIG_DEBUG_FS) || defined(CONFIG_DEV_COREDUMP)
+		.show = a6xx_show,
+#endif
+	},
+	.get_timestamp = a6xx_get_timestamp,
+};
+
+struct msm_gpu *a6xx_gpu_init(struct drm_device *dev)
+{
+	struct msm_drm_private *priv = dev->dev_private;
+	struct platform_device *pdev = priv->gpu_pdev;
+	struct device_node *node;
+	struct a6xx_gpu *a6xx_gpu;
+	struct adreno_gpu *adreno_gpu;
+	struct msm_gpu *gpu;
+	int ret;
+
+	a6xx_gpu = kzalloc(sizeof(*a6xx_gpu), GFP_KERNEL);
+	if (!a6xx_gpu)
+		return ERR_PTR(-ENOMEM);
+
+	adreno_gpu = &a6xx_gpu->base;
+	gpu = &adreno_gpu->base;
+
+	adreno_gpu->registers = a6xx_registers;
+	adreno_gpu->reg_offsets = a6xx_register_offsets;
+
+	ret = adreno_gpu_init(dev, pdev, adreno_gpu, &funcs, 1);
+	if (ret) {
+		a6xx_destroy(&(a6xx_gpu->base.base));
+		return ERR_PTR(ret);
+	}
+
+	/* Check if there is a GMU phandle and set it up */
+	node = of_parse_phandle(pdev->dev.of_node, "gmu", 0);
+
+	/* FIXME: How do we gracefully handle this? */
+	BUG_ON(!node);
+
+	ret = a6xx_gmu_probe(a6xx_gpu, node);
+	if (ret) {
+		a6xx_destroy(&(a6xx_gpu->base.base));
+		return ERR_PTR(ret);
+	}
+
+	if (gpu->aspace)
+		msm_mmu_set_fault_handler(gpu->aspace->mmu, gpu,
+				a6xx_fault_handler);
+
+	return gpu;
+}
diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gpu.h b/drivers/gpu/drm/msm/adreno/a6xx_gpu.h
new file mode 100644
index 000000000000..dd69e5b0e692
--- /dev/null
+++ b/drivers/gpu/drm/msm/adreno/a6xx_gpu.h
@@ -0,0 +1,60 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/* Copyright (c) 2017 The Linux Foundation. All rights reserved. */
+
+#ifndef __A6XX_GPU_H__
+#define __A6XX_GPU_H__
+
+
+#include "adreno_gpu.h"
+#include "a6xx.xml.h"
+
+#include "a6xx_gmu.h"
+
+extern bool hang_debug;
+
+struct a6xx_gpu {
+	struct adreno_gpu base;
+
+	struct drm_gem_object *sqe_bo;
+	uint64_t sqe_iova;
+
+	struct msm_ringbuffer *cur_ring;
+
+	struct a6xx_gmu gmu;
+};
+
+#define to_a6xx_gpu(x) container_of(x, struct a6xx_gpu, base)
+
+/*
+ * Given a register and a count, return a value to program into
+ * REG_CP_PROTECT_REG(n) - this will block both reads and writes for _len
+ * registers starting at _reg.
+ */
+#define A6XX_PROTECT_RW(_reg, _len) \
+	((1 << 31) | \
+	(((_len) & 0x3FFF) << 18) | ((_reg) & 0x3FFFF))
+
+/*
+ * Same as above, but allow reads over the range. For areas of mixed use (such
+ * as performance counters) this allows us to protect a much larger range with a
+ * single register
+ */
+#define A6XX_PROTECT_RDONLY(_reg, _len) \
+	((((_len) & 0x3FFF) << 18) | ((_reg) & 0x3FFFF))
+
+
+int a6xx_gmu_resume(struct a6xx_gpu *gpu);
+int a6xx_gmu_stop(struct a6xx_gpu *gpu);
+
+int a6xx_gmu_wait_for_idle(struct a6xx_gpu *gpu);
+
+int a6xx_gmu_reset(struct a6xx_gpu *a6xx_gpu);
+bool a6xx_gmu_isidle(struct a6xx_gmu *gmu);
+
+int a6xx_gmu_set_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state);
+void a6xx_gmu_clear_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state);
+
+int a6xx_gmu_probe(struct a6xx_gpu *a6xx_gpu, struct device_node *node);
+void a6xx_gmu_remove(struct a6xx_gpu *a6xx_gpu);
+
+#endif /* __A6XX_GPU_H__ */
diff --git a/drivers/gpu/drm/msm/adreno/a6xx_hfi.c b/drivers/gpu/drm/msm/adreno/a6xx_hfi.c
new file mode 100644
index 000000000000..f19ef4cb6ea4
--- /dev/null
+++ b/drivers/gpu/drm/msm/adreno/a6xx_hfi.c
@@ -0,0 +1,435 @@
+// SPDX-License-Identifier: GPL-2.0
+/* Copyright (c) 2017-2018 The Linux Foundation. All rights reserved. */
+
+#include <linux/completion.h>
+#include <linux/circ_buf.h>
+#include <linux/list.h>
+
+#include "a6xx_gmu.h"
+#include "a6xx_gmu.xml.h"
+
+#define HFI_MSG_ID(val) [val] = #val
+
+static const char * const a6xx_hfi_msg_id[] = {
+	HFI_MSG_ID(HFI_H2F_MSG_INIT),
+	HFI_MSG_ID(HFI_H2F_MSG_FW_VERSION),
+	HFI_MSG_ID(HFI_H2F_MSG_BW_TABLE),
+	HFI_MSG_ID(HFI_H2F_MSG_PERF_TABLE),
+	HFI_MSG_ID(HFI_H2F_MSG_TEST),
+};
+
+static int a6xx_hfi_queue_read(struct a6xx_hfi_queue *queue, u32 *data,
+		u32 dwords)
+{
+	struct a6xx_hfi_queue_header *header = queue->header;
+	u32 i, hdr, index = header->read_index;
+
+	if (header->read_index == header->write_index) {
+		header->rx_request = 1;
+		return 0;
+	}
+
+	hdr = queue->data[index];
+
+	/*
+	 * If we are to assume that the GMU firmware is in fact a rational actor
+	 * and is programmed to not send us a larger response than we expect
+	 * then we can also assume that if the header size is unexpectedly large
+	 * that it is due to memory corruption and/or hardware failure. In this
+	 * case the only reasonable course of action is to BUG() to help harden
+	 * the failure.
+	 */
+
+	BUG_ON(HFI_HEADER_SIZE(hdr) > dwords);
+
+	for (i = 0; i < HFI_HEADER_SIZE(hdr); i++) {
+		data[i] = queue->data[index];
+		index = (index + 1) % header->size;
+	}
+
+	header->read_index = index;
+	return HFI_HEADER_SIZE(hdr);
+}
+
+static int a6xx_hfi_queue_write(struct a6xx_gmu *gmu,
+	struct a6xx_hfi_queue *queue, u32 *data, u32 dwords)
+{
+	struct a6xx_hfi_queue_header *header = queue->header;
+	u32 i, space, index = header->write_index;
+
+	spin_lock(&queue->lock);
+
+	space = CIRC_SPACE(header->write_index, header->read_index,
+		header->size);
+	if (space < dwords) {
+		header->dropped++;
+		spin_unlock(&queue->lock);
+		return -ENOSPC;
+	}
+
+	for (i = 0; i < dwords; i++) {
+		queue->data[index] = data[i];
+		index = (index + 1) % header->size;
+	}
+
+	header->write_index = index;
+	spin_unlock(&queue->lock);
+
+	gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET, 0x01);
+	return 0;
+}
+
+struct a6xx_hfi_response {
+	u32 id;
+	u32 seqnum;
+	struct list_head node;
+	struct completion complete;
+
+	u32 error;
+	u32 payload[16];
+};
+
+/*
+ * Incoming HFI ack messages can come in out of order so we need to store all
+ * the pending messages on a list until they are handled.
+ */
+static spinlock_t hfi_ack_lock = __SPIN_LOCK_UNLOCKED(message_lock);
+static LIST_HEAD(hfi_ack_list);
+
+static void a6xx_hfi_handle_ack(struct a6xx_gmu *gmu,
+		struct a6xx_hfi_msg_response *msg)
+{
+	struct a6xx_hfi_response *resp;
+	u32 id, seqnum;
+
+	/* msg->ret_header contains the header of the message being acked */
+	id = HFI_HEADER_ID(msg->ret_header);
+	seqnum = HFI_HEADER_SEQNUM(msg->ret_header);
+
+	spin_lock(&hfi_ack_lock);
+	list_for_each_entry(resp, &hfi_ack_list, node) {
+		if (resp->id == id && resp->seqnum == seqnum) {
+			resp->error = msg->error;
+			memcpy(resp->payload, msg->payload,
+				sizeof(resp->payload));
+
+			complete(&resp->complete);
+			spin_unlock(&hfi_ack_lock);
+			return;
+		}
+	}
+	spin_unlock(&hfi_ack_lock);
+
+	dev_err(gmu->dev, "Nobody was waiting for HFI message %d\n", seqnum);
+}
+
+static void a6xx_hfi_handle_error(struct a6xx_gmu *gmu,
+		struct a6xx_hfi_msg_response *msg)
+{
+	struct a6xx_hfi_msg_error *error = (struct a6xx_hfi_msg_error *) msg;
+
+	dev_err(gmu->dev, "GMU firmware error %d\n", error->code);
+}
+
+void a6xx_hfi_task(unsigned long data)
+{
+	struct a6xx_gmu *gmu = (struct a6xx_gmu *) data;
+	struct a6xx_hfi_queue *queue = &gmu->queues[HFI_RESPONSE_QUEUE];
+	struct a6xx_hfi_msg_response resp;
+
+	for (;;) {
+		u32 id;
+		int ret = a6xx_hfi_queue_read(queue, (u32 *) &resp,
+			sizeof(resp) >> 2);
+
+		/* Returns the number of bytes copied or negative on error */
+		if (ret <= 0) {
+			if (ret < 0)
+				dev_err(gmu->dev,
+					"Unable to read the HFI message queue\n");
+			break;
+		}
+
+		id = HFI_HEADER_ID(resp.header);
+
+		if (id == HFI_F2H_MSG_ACK)
+			a6xx_hfi_handle_ack(gmu, &resp);
+		else if (id == HFI_F2H_MSG_ERROR)
+			a6xx_hfi_handle_error(gmu, &resp);
+	}
+}
+
+static int a6xx_hfi_send_msg(struct a6xx_gmu *gmu, int id,
+		void *data, u32 size, u32 *payload, u32 payload_size)
+{
+	struct a6xx_hfi_queue *queue = &gmu->queues[HFI_COMMAND_QUEUE];
+	struct a6xx_hfi_response resp = { 0 };
+	int ret, dwords = size >> 2;
+	u32 seqnum;
+
+	seqnum = atomic_inc_return(&queue->seqnum) % 0xfff;
+
+	/* First dword of the message is the message header - fill it in */
+	*((u32 *) data) = (seqnum << 20) | (HFI_MSG_CMD << 16) |
+		(dwords << 8) | id;
+
+	init_completion(&resp.complete);
+	resp.id = id;
+	resp.seqnum = seqnum;
+
+	spin_lock_bh(&hfi_ack_lock);
+	list_add_tail(&resp.node, &hfi_ack_list);
+	spin_unlock_bh(&hfi_ack_lock);
+
+	ret = a6xx_hfi_queue_write(gmu, queue, data, dwords);
+	if (ret) {
+		dev_err(gmu->dev, "Unable to send message %s id %d\n",
+			a6xx_hfi_msg_id[id], seqnum);
+		goto out;
+	}
+
+	/* Wait up to 5 seconds for the response */
+	ret = wait_for_completion_timeout(&resp.complete,
+		msecs_to_jiffies(5000));
+	if (!ret) {
+		dev_err(gmu->dev,
+			"Message %s id %d timed out waiting for response\n",
+			a6xx_hfi_msg_id[id], seqnum);
+		ret = -ETIMEDOUT;
+	} else
+		ret = 0;
+
+out:
+	spin_lock_bh(&hfi_ack_lock);
+	list_del(&resp.node);
+	spin_unlock_bh(&hfi_ack_lock);
+
+	if (ret)
+		return ret;
+
+	if (resp.error) {
+		dev_err(gmu->dev, "Message %s id %d returned error %d\n",
+			a6xx_hfi_msg_id[id], seqnum, resp.error);
+		return -EINVAL;
+	}
+
+	if (payload && payload_size) {
+		int copy = min_t(u32, payload_size, sizeof(resp.payload));
+
+		memcpy(payload, resp.payload, copy);
+	}
+
+	return 0;
+}
+
+static int a6xx_hfi_send_gmu_init(struct a6xx_gmu *gmu, int boot_state)
+{
+	struct a6xx_hfi_msg_gmu_init_cmd msg = { 0 };
+
+	msg.dbg_buffer_addr = (u32) gmu->debug->iova;
+	msg.dbg_buffer_size = (u32) gmu->debug->size;
+	msg.boot_state = boot_state;
+
+	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_INIT, &msg, sizeof(msg),
+		NULL, 0);
+}
+
+static int a6xx_hfi_get_fw_version(struct a6xx_gmu *gmu, u32 *version)
+{
+	struct a6xx_hfi_msg_fw_version msg = { 0 };
+
+	/* Currently supporting version 1.1 */
+	msg.supported_version = (1 << 28) | (1 << 16);
+
+	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_FW_VERSION, &msg, sizeof(msg),
+		version, sizeof(*version));
+}
+
+static int a6xx_hfi_send_perf_table(struct a6xx_gmu *gmu)
+{
+	struct a6xx_hfi_msg_perf_table msg = { 0 };
+	int i;
+
+	msg.num_gpu_levels = gmu->nr_gpu_freqs;
+	msg.num_gmu_levels = gmu->nr_gmu_freqs;
+
+	for (i = 0; i < gmu->nr_gpu_freqs; i++) {
+		msg.gx_votes[i].vote = gmu->gx_arc_votes[i];
+		msg.gx_votes[i].freq = gmu->gpu_freqs[i] / 1000;
+	}
+
+	for (i = 0; i < gmu->nr_gmu_freqs; i++) {
+		msg.cx_votes[i].vote = gmu->cx_arc_votes[i];
+		msg.cx_votes[i].freq = gmu->gmu_freqs[i] / 1000;
+	}
+
+	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_PERF_TABLE, &msg, sizeof(msg),
+		NULL, 0);
+}
+
+static int a6xx_hfi_send_bw_table(struct a6xx_gmu *gmu)
+{
+	struct a6xx_hfi_msg_bw_table msg = { 0 };
+
+	/*
+	 * The sdm845 GMU doesn't do bus frequency scaling on its own but it
+	 * does need at least one entry in the list because it might be accessed
+	 * when the GMU is shutting down. Send a single "off" entry.
+	 */
+
+	msg.bw_level_num = 1;
+
+	msg.ddr_cmds_num = 3;
+	msg.ddr_wait_bitmask = 0x07;
+
+	msg.ddr_cmds_addrs[0] = 0x50000;
+	msg.ddr_cmds_addrs[1] = 0x5005c;
+	msg.ddr_cmds_addrs[2] = 0x5000c;
+
+	msg.ddr_cmds_data[0][0] =  0x40000000;
+	msg.ddr_cmds_data[0][1] =  0x40000000;
+	msg.ddr_cmds_data[0][2] =  0x40000000;
+
+	/*
+	 * These are the CX (CNOC) votes.  This is used but the values for the
+	 * sdm845 GMU are known and fixed so we can hard code them.
+	 */
+
+	msg.cnoc_cmds_num = 3;
+	msg.cnoc_wait_bitmask = 0x05;
+
+	msg.cnoc_cmds_addrs[0] = 0x50034;
+	msg.cnoc_cmds_addrs[1] = 0x5007c;
+	msg.cnoc_cmds_addrs[2] = 0x5004c;
+
+	msg.cnoc_cmds_data[0][0] =  0x40000000;
+	msg.cnoc_cmds_data[0][1] =  0x00000000;
+	msg.cnoc_cmds_data[0][2] =  0x40000000;
+
+	msg.cnoc_cmds_data[1][0] =  0x60000001;
+	msg.cnoc_cmds_data[1][1] =  0x20000001;
+	msg.cnoc_cmds_data[1][2] =  0x60000001;
+
+	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_BW_TABLE, &msg, sizeof(msg),
+		NULL, 0);
+}
+
+static int a6xx_hfi_send_test(struct a6xx_gmu *gmu)
+{
+	struct a6xx_hfi_msg_test msg = { 0 };
+
+	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_TEST, &msg, sizeof(msg),
+		NULL, 0);
+}
+
+int a6xx_hfi_start(struct a6xx_gmu *gmu, int boot_state)
+{
+	int ret;
+
+	ret = a6xx_hfi_send_gmu_init(gmu, boot_state);
+	if (ret)
+		return ret;
+
+	ret = a6xx_hfi_get_fw_version(gmu, NULL);
+	if (ret)
+		return ret;
+
+	/*
+	 * We have to get exchange version numbers per the sequence but at this
+	 * point th kernel driver doesn't need to know the exact version of
+	 * the GMU firmware
+	 */
+
+	ret = a6xx_hfi_send_perf_table(gmu);
+	if (ret)
+		return ret;
+
+	ret = a6xx_hfi_send_bw_table(gmu);
+	if (ret)
+		return ret;
+
+	/*
+	 * Let the GMU know that there won't be any more HFI messages until next
+	 * boot
+	 */
+	a6xx_hfi_send_test(gmu);
+
+	return 0;
+}
+
+void a6xx_hfi_stop(struct a6xx_gmu *gmu)
+{
+	int i;
+
+	for (i = 0; i < ARRAY_SIZE(gmu->queues); i++) {
+		struct a6xx_hfi_queue *queue = &gmu->queues[i];
+
+		if (!queue->header)
+			continue;
+
+		if (queue->header->read_index != queue->header->write_index)
+			dev_err(gmu->dev, "HFI queue %d is not empty\n", i);
+
+		queue->header->read_index = 0;
+		queue->header->write_index = 0;
+	}
+}
+
+static void a6xx_hfi_queue_init(struct a6xx_hfi_queue *queue,
+		struct a6xx_hfi_queue_header *header, void *virt, u64 iova,
+		u32 id)
+{
+	spin_lock_init(&queue->lock);
+	queue->header = header;
+	queue->data = virt;
+	atomic_set(&queue->seqnum, 0);
+
+	/* Set up the shared memory header */
+	header->iova = iova;
+	header->type =  10 << 8 | id;
+	header->status = 1;
+	header->size = SZ_4K >> 2;
+	header->msg_size = 0;
+	header->dropped = 0;
+	header->rx_watermark = 1;
+	header->tx_watermark = 1;
+	header->rx_request = 1;
+	header->tx_request = 0;
+	header->read_index = 0;
+	header->write_index = 0;
+}
+
+void a6xx_hfi_init(struct a6xx_gmu *gmu)
+{
+	struct a6xx_gmu_bo *hfi = gmu->hfi;
+	struct a6xx_hfi_queue_table_header *table = hfi->virt;
+	struct a6xx_hfi_queue_header *headers = hfi->virt + sizeof(*table);
+	u64 offset;
+	int table_size;
+
+	/*
+	 * The table size is the size of the table header plus all of the queue
+	 * headers
+	 */
+	table_size = sizeof(*table);
+	table_size += (ARRAY_SIZE(gmu->queues) *
+		sizeof(struct a6xx_hfi_queue_header));
+
+	table->version = 0;
+	table->size = table_size;
+	/* First queue header is located immediately after the table header */
+	table->qhdr0_offset = sizeof(*table) >> 2;
+	table->qhdr_size = sizeof(struct a6xx_hfi_queue_header) >> 2;
+	table->num_queues = ARRAY_SIZE(gmu->queues);
+	table->active_queues = ARRAY_SIZE(gmu->queues);
+
+	/* Command queue */
+	offset = SZ_4K;
+	a6xx_hfi_queue_init(&gmu->queues[0], &headers[0], hfi->virt + offset,
+		hfi->iova + offset, 0);
+
+	/* GMU response queue */
+	offset += SZ_4K;
+	a6xx_hfi_queue_init(&gmu->queues[1], &headers[1], hfi->virt + offset,
+		hfi->iova + offset, 4);
+}
diff --git a/drivers/gpu/drm/msm/adreno/a6xx_hfi.h b/drivers/gpu/drm/msm/adreno/a6xx_hfi.h
new file mode 100644
index 000000000000..60d1319fa44f
--- /dev/null
+++ b/drivers/gpu/drm/msm/adreno/a6xx_hfi.h
@@ -0,0 +1,127 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/* Copyright (c) 2017 The Linux Foundation. All rights reserved. */
+
+#ifndef _A6XX_HFI_H_
+#define _A6XX_HFI_H_
+
+struct a6xx_hfi_queue_table_header {
+	u32 version;
+	u32 size;		/* Size of the queue table in dwords */
+	u32 qhdr0_offset;	/* Offset of the first queue header */
+	u32 qhdr_size;		/* Size of the queue headers */
+	u32 num_queues;		/* Number of total queues */
+	u32 active_queues;	/* Number of active queues */
+};
+
+struct a6xx_hfi_queue_header {
+	u32 status;
+	u32 iova;
+	u32 type;
+	u32 size;
+	u32 msg_size;
+	u32 dropped;
+	u32 rx_watermark;
+	u32 tx_watermark;
+	u32 rx_request;
+	u32 tx_request;
+	u32 read_index;
+	u32 write_index;
+};
+
+struct a6xx_hfi_queue {
+	struct a6xx_hfi_queue_header *header;
+	spinlock_t lock;
+	u32 *data;
+	atomic_t seqnum;
+};
+
+/* This is the outgoing queue to the GMU */
+#define HFI_COMMAND_QUEUE 0
+
+/* THis is the incoming response queue from the GMU */
+#define HFI_RESPONSE_QUEUE 1
+
+#define HFI_HEADER_ID(msg) ((msg) & 0xff)
+#define HFI_HEADER_SIZE(msg) (((msg) >> 8) & 0xff)
+#define HFI_HEADER_SEQNUM(msg) (((msg) >> 20) & 0xfff)
+
+/* FIXME: Do we need this or can we use ARRAY_SIZE? */
+#define HFI_RESPONSE_PAYLOAD_SIZE 16
+
+/* HFI message types */
+
+#define HFI_MSG_CMD 0
+#define HFI_MSG_ACK 2
+
+#define HFI_F2H_MSG_ACK 126
+
+struct a6xx_hfi_msg_response {
+	u32 header;
+	u32 ret_header;
+	u32 error;
+	u32 payload[HFI_RESPONSE_PAYLOAD_SIZE];
+};
+
+#define HFI_F2H_MSG_ERROR 100
+
+struct a6xx_hfi_msg_error {
+	u32 header;
+	u32 code;
+	u32 payload[2];
+};
+
+#define HFI_H2F_MSG_INIT 0
+
+struct a6xx_hfi_msg_gmu_init_cmd {
+	u32 header;
+	u32 seg_id;
+	u32 dbg_buffer_addr;
+	u32 dbg_buffer_size;
+	u32 boot_state;
+};
+
+#define HFI_H2F_MSG_FW_VERSION 1
+
+struct a6xx_hfi_msg_fw_version {
+	u32 header;
+	u32 supported_version;
+};
+
+#define HFI_H2F_MSG_PERF_TABLE 4
+
+struct perf_level {
+	u32 vote;
+	u32 freq;
+};
+
+struct a6xx_hfi_msg_perf_table {
+	u32 header;
+	u32 num_gpu_levels;
+	u32 num_gmu_levels;
+
+	struct perf_level gx_votes[16];
+	struct perf_level cx_votes[4];
+};
+
+#define HFI_H2F_MSG_BW_TABLE 3
+
+struct a6xx_hfi_msg_bw_table {
+	u32 header;
+	u32 bw_level_num;
+	u32 cnoc_cmds_num;
+	u32 ddr_cmds_num;
+	u32 cnoc_wait_bitmask;
+	u32 ddr_wait_bitmask;
+	u32 cnoc_cmds_addrs[6];
+	u32 cnoc_cmds_data[2][6];
+	u32 ddr_cmds_addrs[8];
+	u32 ddr_cmds_data[16][8];
+};
+
+#define HFI_H2F_MSG_TEST 5
+
+struct a6xx_hfi_msg_test {
+	u32 header;
+};
+
+#endif
diff --git a/drivers/gpu/drm/msm/adreno/adreno_device.c b/drivers/gpu/drm/msm/adreno/adreno_device.c
index 37746f1d54cf..7d3e9a129ac7 100644
--- a/drivers/gpu/drm/msm/adreno/adreno_device.c
+++ b/drivers/gpu/drm/msm/adreno/adreno_device.c
@@ -111,6 +111,16 @@ static const struct adreno_info gpulist[] = {
 			ADRENO_QUIRK_FAULT_DETECT_MASK,
 		.init = a5xx_gpu_init,
 		.zapfw = "a530_zap.mdt",
+	}, {
+		.rev = ADRENO_REV(6, 3, 0, ANY_ID),
+		.revn = 630,
+		.name = "A630",
+		.fw = {
+			[ADRENO_FW_SQE] = "a630_sqe.fw",
+			[ADRENO_FW_GMU] = "a630_gmu.bin",
+		},
+		.gmem = SZ_1M,
+		.init = a6xx_gpu_init,
 	},
 };
 
@@ -127,6 +137,8 @@ MODULE_FIRMWARE("qcom/a530_zap.mdt");
 MODULE_FIRMWARE("qcom/a530_zap.b00");
 MODULE_FIRMWARE("qcom/a530_zap.b01");
 MODULE_FIRMWARE("qcom/a530_zap.b02");
+MODULE_FIRMWARE("qcom/a630_sqe.fw");
+MODULE_FIRMWARE("qcom/a630_gmu.bin");
 
 static inline bool _rev_match(uint8_t entry, uint8_t id)
 {
diff --git a/drivers/gpu/drm/msm/adreno/adreno_gpu.h b/drivers/gpu/drm/msm/adreno/adreno_gpu.h
index d391ff377612..de6e6ee42fba 100644
--- a/drivers/gpu/drm/msm/adreno/adreno_gpu.h
+++ b/drivers/gpu/drm/msm/adreno/adreno_gpu.h
@@ -50,7 +50,9 @@ enum adreno_regs {
 
 enum {
 	ADRENO_FW_PM4 = 0,
+	ADRENO_FW_SQE = 0, /* a6xx */
 	ADRENO_FW_PFP = 1,
+	ADRENO_FW_GMU = 1, /* a6xx */
 	ADRENO_FW_GPMU = 2,
 	ADRENO_FW_MAX,
 };
@@ -335,6 +337,7 @@ static inline void adreno_gpu_write(struct adreno_gpu *gpu,
 struct msm_gpu *a3xx_gpu_init(struct drm_device *dev);
 struct msm_gpu *a4xx_gpu_init(struct drm_device *dev);
 struct msm_gpu *a5xx_gpu_init(struct drm_device *dev);
+struct msm_gpu *a6xx_gpu_init(struct drm_device *dev);
 
 static inline void adreno_gpu_write64(struct adreno_gpu *gpu,
 		enum adreno_regs lo, enum adreno_regs hi, u64 data)
diff --git a/drivers/gpu/drm/msm/msm_gpu.c b/drivers/gpu/drm/msm/msm_gpu.c
index ca368490b3ee..5e808cfec345 100644
--- a/drivers/gpu/drm/msm/msm_gpu.c
+++ b/drivers/gpu/drm/msm/msm_gpu.c
@@ -88,7 +88,7 @@ static struct devfreq_dev_profile msm_devfreq_profile = {
 static void msm_devfreq_init(struct msm_gpu *gpu)
 {
 	/* We need target support to do devfreq */
-	if (!gpu->funcs->gpu_busy)
+	if (!gpu->funcs->gpu_busy || !gpu->core_clk)
 		return;
 
 	msm_devfreq_profile.initial_freq = gpu->fast_rate;