These changes are the raw update to linux-4.4.6-rt14. Kernel sources
[kvmfornfv.git] / kernel / drivers / gpu / drm / amd / amdgpu / fiji_smc.c
diff --git a/kernel/drivers/gpu/drm/amd/amdgpu/fiji_smc.c b/kernel/drivers/gpu/drm/amd/amdgpu/fiji_smc.c
new file mode 100644 (file)
index 0000000..bda1249
--- /dev/null
@@ -0,0 +1,857 @@
+/*
+ * Copyright 2014 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ *
+ */
+
+#include <linux/firmware.h>
+#include "drmP.h"
+#include "amdgpu.h"
+#include "fiji_ppsmc.h"
+#include "fiji_smumgr.h"
+#include "smu_ucode_xfer_vi.h"
+#include "amdgpu_ucode.h"
+
+#include "smu/smu_7_1_3_d.h"
+#include "smu/smu_7_1_3_sh_mask.h"
+
+#define FIJI_SMC_SIZE 0x20000
+
+static int fiji_set_smc_sram_address(struct amdgpu_device *adev, uint32_t smc_address, uint32_t limit)
+{
+       uint32_t val;
+
+       if (smc_address & 3)
+               return -EINVAL;
+
+       if ((smc_address + 3) > limit)
+               return -EINVAL;
+
+       WREG32(mmSMC_IND_INDEX_0, smc_address);
+
+       val = RREG32(mmSMC_IND_ACCESS_CNTL);
+       val = REG_SET_FIELD(val, SMC_IND_ACCESS_CNTL, AUTO_INCREMENT_IND_0, 0);
+       WREG32(mmSMC_IND_ACCESS_CNTL, val);
+
+       return 0;
+}
+
+static int fiji_copy_bytes_to_smc(struct amdgpu_device *adev, uint32_t smc_start_address, const uint8_t *src, uint32_t byte_count, uint32_t limit)
+{
+       uint32_t addr;
+       uint32_t data, orig_data;
+       int result = 0;
+       uint32_t extra_shift;
+       unsigned long flags;
+
+       if (smc_start_address & 3)
+               return -EINVAL;
+
+       if ((smc_start_address + byte_count) > limit)
+               return -EINVAL;
+
+       addr = smc_start_address;
+
+       spin_lock_irqsave(&adev->smc_idx_lock, flags);
+       while (byte_count >= 4) {
+               /* Bytes are written into the SMC addres space with the MSB first */
+               data = (src[0] << 24) + (src[1] << 16) + (src[2] << 8) + src[3];
+
+               result = fiji_set_smc_sram_address(adev, addr, limit);
+
+               if (result)
+                       goto out;
+
+               WREG32(mmSMC_IND_DATA_0, data);
+
+               src += 4;
+               byte_count -= 4;
+               addr += 4;
+       }
+
+       if (0 != byte_count) {
+               /* Now write odd bytes left, do a read modify write cycle */
+               data = 0;
+
+               result = fiji_set_smc_sram_address(adev, addr, limit);
+               if (result)
+                       goto out;
+
+               orig_data = RREG32(mmSMC_IND_DATA_0);
+               extra_shift = 8 * (4 - byte_count);
+
+               while (byte_count > 0) {
+                       data = (data << 8) + *src++;
+                       byte_count--;
+               }
+
+               data <<= extra_shift;
+               data |= (orig_data & ~((~0UL) << extra_shift));
+
+               result = fiji_set_smc_sram_address(adev, addr, limit);
+               if (result)
+                       goto out;
+
+               WREG32(mmSMC_IND_DATA_0, data);
+       }
+
+out:
+       spin_unlock_irqrestore(&adev->smc_idx_lock, flags);
+       return result;
+}
+
+static int fiji_program_jump_on_start(struct amdgpu_device *adev)
+{
+       static unsigned char data[] = {0xE0, 0x00, 0x80, 0x40};
+       fiji_copy_bytes_to_smc(adev, 0x0, data, 4, sizeof(data)+1);
+
+       return 0;
+}
+
+static bool fiji_is_smc_ram_running(struct amdgpu_device *adev)
+{
+       uint32_t val = RREG32_SMC(ixSMC_SYSCON_CLOCK_CNTL_0);
+       val = REG_GET_FIELD(val, SMC_SYSCON_CLOCK_CNTL_0, ck_disable);
+
+       return ((0 == val) && (0x20100 <= RREG32_SMC(ixSMC_PC_C)));
+}
+
+static int wait_smu_response(struct amdgpu_device *adev)
+{
+       int i;
+       uint32_t val;
+
+       for (i = 0; i < adev->usec_timeout; i++) {
+               val = RREG32(mmSMC_RESP_0);
+               if (REG_GET_FIELD(val, SMC_RESP_0, SMC_RESP))
+                       break;
+               udelay(1);
+       }
+
+       if (i == adev->usec_timeout)
+               return -EINVAL;
+
+       return 0;
+}
+
+static int fiji_send_msg_to_smc_offset(struct amdgpu_device *adev)
+{
+       if (wait_smu_response(adev)) {
+               DRM_ERROR("Failed to send previous message\n");
+               return -EINVAL;
+       }
+
+       WREG32(mmSMC_MSG_ARG_0, 0x20000);
+       WREG32(mmSMC_MESSAGE_0, PPSMC_MSG_Test);
+
+       if (wait_smu_response(adev)) {
+               DRM_ERROR("Failed to send message\n");
+               return -EINVAL;
+       }
+
+       return 0;
+}
+
+static int fiji_send_msg_to_smc(struct amdgpu_device *adev, PPSMC_Msg msg)
+{
+       if (!fiji_is_smc_ram_running(adev))
+       {
+               return -EINVAL;;
+       }
+
+       if (wait_smu_response(adev)) {
+               DRM_ERROR("Failed to send previous message\n");
+               return -EINVAL;
+       }
+
+       WREG32(mmSMC_MESSAGE_0, msg);
+
+       if (wait_smu_response(adev)) {
+               DRM_ERROR("Failed to send message\n");
+               return -EINVAL;
+       }
+
+       return 0;
+}
+
+static int fiji_send_msg_to_smc_without_waiting(struct amdgpu_device *adev,
+                                               PPSMC_Msg msg)
+{
+       if (wait_smu_response(adev)) {
+               DRM_ERROR("Failed to send previous message\n");
+               return -EINVAL;
+       }
+
+       WREG32(mmSMC_MESSAGE_0, msg);
+
+       return 0;
+}
+
+static int fiji_send_msg_to_smc_with_parameter(struct amdgpu_device *adev,
+                                               PPSMC_Msg msg,
+                                               uint32_t parameter)
+{
+       if (!fiji_is_smc_ram_running(adev))
+               return -EINVAL;
+
+       if (wait_smu_response(adev)) {
+               DRM_ERROR("Failed to send previous message\n");
+               return -EINVAL;
+       }
+
+       WREG32(mmSMC_MSG_ARG_0, parameter);
+
+       return fiji_send_msg_to_smc(adev, msg);
+}
+
+static int fiji_send_msg_to_smc_with_parameter_without_waiting(
+                                       struct amdgpu_device *adev,
+                                       PPSMC_Msg msg, uint32_t parameter)
+{
+       if (wait_smu_response(adev)) {
+               DRM_ERROR("Failed to send previous message\n");
+               return -EINVAL;
+       }
+
+       WREG32(mmSMC_MSG_ARG_0, parameter);
+
+       return fiji_send_msg_to_smc_without_waiting(adev, msg);
+}
+
+#if 0 /* not used yet */
+static int fiji_wait_for_smc_inactive(struct amdgpu_device *adev)
+{
+       int i;
+       uint32_t val;
+
+       if (!fiji_is_smc_ram_running(adev))
+               return -EINVAL;
+
+       for (i = 0; i < adev->usec_timeout; i++) {
+               val = RREG32_SMC(ixSMC_SYSCON_CLOCK_CNTL_0);
+               if (REG_GET_FIELD(val, SMC_SYSCON_CLOCK_CNTL_0, cken) == 0)
+                       break;
+               udelay(1);
+       }
+
+       if (i == adev->usec_timeout)
+               return -EINVAL;
+
+       return 0;
+}
+#endif
+
+static int fiji_smu_upload_firmware_image(struct amdgpu_device *adev)
+{
+       const struct smc_firmware_header_v1_0 *hdr;
+       uint32_t ucode_size;
+       uint32_t ucode_start_address;
+       const uint8_t *src;
+       uint32_t val;
+       uint32_t byte_count;
+       uint32_t *data;
+       unsigned long flags;
+
+       if (!adev->pm.fw)
+               return -EINVAL;
+
+       hdr = (const struct smc_firmware_header_v1_0 *)adev->pm.fw->data;
+       amdgpu_ucode_print_smc_hdr(&hdr->header);
+
+       adev->pm.fw_version = le32_to_cpu(hdr->header.ucode_version);
+       ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes);
+       ucode_start_address = le32_to_cpu(hdr->ucode_start_addr);
+       src = (const uint8_t *)
+               (adev->pm.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
+
+       if (ucode_size & 3) {
+               DRM_ERROR("SMC ucode is not 4 bytes aligned\n");
+               return -EINVAL;
+       }
+
+       if (ucode_size > FIJI_SMC_SIZE) {
+               DRM_ERROR("SMC address is beyond the SMC RAM area\n");
+               return -EINVAL;
+       }
+
+       spin_lock_irqsave(&adev->smc_idx_lock, flags);
+       WREG32(mmSMC_IND_INDEX_0, ucode_start_address);
+
+       val = RREG32(mmSMC_IND_ACCESS_CNTL);
+       val = REG_SET_FIELD(val, SMC_IND_ACCESS_CNTL, AUTO_INCREMENT_IND_0, 1);
+       WREG32(mmSMC_IND_ACCESS_CNTL, val);
+
+       byte_count = ucode_size;
+       data = (uint32_t *)src;
+       for (; byte_count >= 4; data++, byte_count -= 4)
+               WREG32(mmSMC_IND_DATA_0, data[0]);
+
+       val = RREG32(mmSMC_IND_ACCESS_CNTL);
+       val = REG_SET_FIELD(val, SMC_IND_ACCESS_CNTL, AUTO_INCREMENT_IND_0, 0);
+       WREG32(mmSMC_IND_ACCESS_CNTL, val);
+       spin_unlock_irqrestore(&adev->smc_idx_lock, flags);
+
+       return 0;
+}
+
+#if 0 /* not used yet */
+static int fiji_read_smc_sram_dword(struct amdgpu_device *adev,
+                               uint32_t smc_address,
+                               uint32_t *value,
+                               uint32_t limit)
+{
+       int result;
+       unsigned long flags;
+
+       spin_lock_irqsave(&adev->smc_idx_lock, flags);
+       result = fiji_set_smc_sram_address(adev, smc_address, limit);
+       if (result == 0)
+               *value = RREG32(mmSMC_IND_DATA_0);
+       spin_unlock_irqrestore(&adev->smc_idx_lock, flags);
+       return result;
+}
+
+static int fiji_write_smc_sram_dword(struct amdgpu_device *adev,
+                               uint32_t smc_address,
+                               uint32_t value,
+                               uint32_t limit)
+{
+       int result;
+       unsigned long flags;
+
+       spin_lock_irqsave(&adev->smc_idx_lock, flags);
+       result = fiji_set_smc_sram_address(adev, smc_address, limit);
+       if (result == 0)
+               WREG32(mmSMC_IND_DATA_0, value);
+       spin_unlock_irqrestore(&adev->smc_idx_lock, flags);
+       return result;
+}
+
+static int fiji_smu_stop_smc(struct amdgpu_device *adev)
+{
+       uint32_t val = RREG32_SMC(ixSMC_SYSCON_RESET_CNTL);
+       val = REG_SET_FIELD(val, SMC_SYSCON_RESET_CNTL, rst_reg, 1);
+       WREG32_SMC(ixSMC_SYSCON_RESET_CNTL, val);
+
+       val = RREG32_SMC(ixSMC_SYSCON_CLOCK_CNTL_0);
+       val = REG_SET_FIELD(val, SMC_SYSCON_CLOCK_CNTL_0, ck_disable, 1);
+       WREG32_SMC(ixSMC_SYSCON_CLOCK_CNTL_0, val);
+
+       return 0;
+}
+#endif
+
+static enum AMDGPU_UCODE_ID fiji_convert_fw_type(uint32_t fw_type)
+{
+       switch (fw_type) {
+               case UCODE_ID_SDMA0:
+                       return AMDGPU_UCODE_ID_SDMA0;
+               case UCODE_ID_SDMA1:
+                       return AMDGPU_UCODE_ID_SDMA1;
+               case UCODE_ID_CP_CE:
+                       return AMDGPU_UCODE_ID_CP_CE;
+               case UCODE_ID_CP_PFP:
+                       return AMDGPU_UCODE_ID_CP_PFP;
+               case UCODE_ID_CP_ME:
+                       return AMDGPU_UCODE_ID_CP_ME;
+               case UCODE_ID_CP_MEC:
+               case UCODE_ID_CP_MEC_JT1:
+               case UCODE_ID_CP_MEC_JT2:
+                       return AMDGPU_UCODE_ID_CP_MEC1;
+               case UCODE_ID_RLC_G:
+                       return AMDGPU_UCODE_ID_RLC_G;
+               default:
+                       DRM_ERROR("ucode type is out of range!\n");
+                       return AMDGPU_UCODE_ID_MAXIMUM;
+       }
+}
+
+static int fiji_smu_populate_single_firmware_entry(struct amdgpu_device *adev,
+                                               uint32_t fw_type,
+                                               struct SMU_Entry *entry)
+{
+       enum AMDGPU_UCODE_ID id = fiji_convert_fw_type(fw_type);
+       struct amdgpu_firmware_info *ucode = &adev->firmware.ucode[id];
+       const struct gfx_firmware_header_v1_0 *header = NULL;
+       uint64_t gpu_addr;
+       uint32_t data_size;
+
+       if (ucode->fw == NULL)
+               return -EINVAL;
+       gpu_addr  = ucode->mc_addr;
+       header = (const struct gfx_firmware_header_v1_0 *)ucode->fw->data;
+       data_size = le32_to_cpu(header->header.ucode_size_bytes);
+
+       if ((fw_type == UCODE_ID_CP_MEC_JT1) ||
+               (fw_type == UCODE_ID_CP_MEC_JT2)) {
+               gpu_addr += le32_to_cpu(header->jt_offset) << 2;
+               data_size = le32_to_cpu(header->jt_size) << 2;
+       }
+
+       entry->version = (uint16_t)le32_to_cpu(header->header.ucode_version);
+       entry->id = (uint16_t)fw_type;
+       entry->image_addr_high = upper_32_bits(gpu_addr);
+       entry->image_addr_low = lower_32_bits(gpu_addr);
+       entry->meta_data_addr_high = 0;
+       entry->meta_data_addr_low = 0;
+       entry->data_size_byte = data_size;
+       entry->num_register_entries = 0;
+
+       if (fw_type == UCODE_ID_RLC_G)
+               entry->flags = 1;
+       else
+               entry->flags = 0;
+
+       return 0;
+}
+
+static int fiji_smu_request_load_fw(struct amdgpu_device *adev)
+{
+       struct fiji_smu_private_data *private = (struct fiji_smu_private_data *)adev->smu.priv;
+       struct SMU_DRAMData_TOC *toc;
+       uint32_t fw_to_load;
+
+       WREG32_SMC(ixSOFT_REGISTERS_TABLE_28, 0);
+
+       fiji_send_msg_to_smc_with_parameter(adev, PPSMC_MSG_SMU_DRAM_ADDR_HI, private->smu_buffer_addr_high);
+       fiji_send_msg_to_smc_with_parameter(adev, PPSMC_MSG_SMU_DRAM_ADDR_LO, private->smu_buffer_addr_low);
+
+       toc = (struct SMU_DRAMData_TOC *)private->header;
+       toc->num_entries = 0;
+       toc->structure_version = 1;
+
+       if (!adev->firmware.smu_load)
+               return 0;
+
+       if (fiji_smu_populate_single_firmware_entry(adev, UCODE_ID_RLC_G,
+                       &toc->entry[toc->num_entries++])) {
+               DRM_ERROR("Failed to get firmware entry for RLC\n");
+               return -EINVAL;
+       }
+
+       if (fiji_smu_populate_single_firmware_entry(adev, UCODE_ID_CP_CE,
+                       &toc->entry[toc->num_entries++])) {
+               DRM_ERROR("Failed to get firmware entry for CE\n");
+               return -EINVAL;
+       }
+
+       if (fiji_smu_populate_single_firmware_entry(adev, UCODE_ID_CP_PFP,
+                       &toc->entry[toc->num_entries++])) {
+               DRM_ERROR("Failed to get firmware entry for PFP\n");
+               return -EINVAL;
+       }
+
+       if (fiji_smu_populate_single_firmware_entry(adev, UCODE_ID_CP_ME,
+                       &toc->entry[toc->num_entries++])) {
+               DRM_ERROR("Failed to get firmware entry for ME\n");
+               return -EINVAL;
+       }
+
+       if (fiji_smu_populate_single_firmware_entry(adev, UCODE_ID_CP_MEC,
+                       &toc->entry[toc->num_entries++])) {
+               DRM_ERROR("Failed to get firmware entry for MEC\n");
+               return -EINVAL;
+       }
+
+       if (fiji_smu_populate_single_firmware_entry(adev, UCODE_ID_CP_MEC_JT1,
+                       &toc->entry[toc->num_entries++])) {
+               DRM_ERROR("Failed to get firmware entry for MEC_JT1\n");
+               return -EINVAL;
+       }
+
+       if (fiji_smu_populate_single_firmware_entry(adev, UCODE_ID_CP_MEC_JT2,
+                       &toc->entry[toc->num_entries++])) {
+               DRM_ERROR("Failed to get firmware entry for MEC_JT2\n");
+               return -EINVAL;
+       }
+
+       if (fiji_smu_populate_single_firmware_entry(adev, UCODE_ID_SDMA0,
+                       &toc->entry[toc->num_entries++])) {
+               DRM_ERROR("Failed to get firmware entry for SDMA0\n");
+               return -EINVAL;
+       }
+
+       if (fiji_smu_populate_single_firmware_entry(adev, UCODE_ID_SDMA1,
+                       &toc->entry[toc->num_entries++])) {
+               DRM_ERROR("Failed to get firmware entry for SDMA1\n");
+               return -EINVAL;
+       }
+
+       fiji_send_msg_to_smc_with_parameter(adev, PPSMC_MSG_DRV_DRAM_ADDR_HI, private->header_addr_high);
+       fiji_send_msg_to_smc_with_parameter(adev, PPSMC_MSG_DRV_DRAM_ADDR_LO, private->header_addr_low);
+
+       fw_to_load = UCODE_ID_RLC_G_MASK |
+                       UCODE_ID_SDMA0_MASK |
+                       UCODE_ID_SDMA1_MASK |
+                       UCODE_ID_CP_CE_MASK |
+                       UCODE_ID_CP_ME_MASK |
+                       UCODE_ID_CP_PFP_MASK |
+                       UCODE_ID_CP_MEC_MASK;
+
+       if (fiji_send_msg_to_smc_with_parameter_without_waiting(adev, PPSMC_MSG_LoadUcodes, fw_to_load)) {
+               DRM_ERROR("Fail to request SMU load ucode\n");
+               return -EINVAL;
+       }
+
+       return 0;
+}
+
+static uint32_t fiji_smu_get_mask_for_fw_type(uint32_t fw_type)
+{
+       switch (fw_type) {
+               case AMDGPU_UCODE_ID_SDMA0:
+                       return UCODE_ID_SDMA0_MASK;
+               case AMDGPU_UCODE_ID_SDMA1:
+                       return UCODE_ID_SDMA1_MASK;
+               case AMDGPU_UCODE_ID_CP_CE:
+                       return UCODE_ID_CP_CE_MASK;
+               case AMDGPU_UCODE_ID_CP_PFP:
+                       return UCODE_ID_CP_PFP_MASK;
+               case AMDGPU_UCODE_ID_CP_ME:
+                       return UCODE_ID_CP_ME_MASK;
+               case AMDGPU_UCODE_ID_CP_MEC1:
+                       return UCODE_ID_CP_MEC_MASK;
+               case AMDGPU_UCODE_ID_CP_MEC2:
+                       return UCODE_ID_CP_MEC_MASK;
+               case AMDGPU_UCODE_ID_RLC_G:
+                       return UCODE_ID_RLC_G_MASK;
+               default:
+                       DRM_ERROR("ucode type is out of range!\n");
+                       return 0;
+       }
+}
+
+static int fiji_smu_check_fw_load_finish(struct amdgpu_device *adev,
+                                       uint32_t fw_type)
+{
+       uint32_t fw_mask = fiji_smu_get_mask_for_fw_type(fw_type);
+       int i;
+
+       for (i = 0; i < adev->usec_timeout; i++) {
+               if (fw_mask == (RREG32_SMC(ixSOFT_REGISTERS_TABLE_28) & fw_mask))
+                       break;
+               udelay(1);
+       }
+
+       if (i == adev->usec_timeout) {
+               DRM_ERROR("check firmware loading failed\n");
+               return -EINVAL;
+       }
+
+       return 0;
+}
+
+static int fiji_smu_start_in_protection_mode(struct amdgpu_device *adev)
+{
+       int result;
+       uint32_t val;
+       int i;
+
+       /* Assert reset */
+       val = RREG32_SMC(ixSMC_SYSCON_RESET_CNTL);
+       val = REG_SET_FIELD(val, SMC_SYSCON_RESET_CNTL, rst_reg, 1);
+       WREG32_SMC(ixSMC_SYSCON_RESET_CNTL, val);
+
+       result = fiji_smu_upload_firmware_image(adev);
+       if (result)
+               return result;
+
+       /* Clear status */
+       WREG32_SMC(ixSMU_STATUS, 0);
+
+       /* Enable clock */
+       val = RREG32_SMC(ixSMC_SYSCON_CLOCK_CNTL_0);
+       val = REG_SET_FIELD(val, SMC_SYSCON_CLOCK_CNTL_0, ck_disable, 0);
+       WREG32_SMC(ixSMC_SYSCON_CLOCK_CNTL_0, val);
+
+       /* De-assert reset */
+       val = RREG32_SMC(ixSMC_SYSCON_RESET_CNTL);
+       val = REG_SET_FIELD(val, SMC_SYSCON_RESET_CNTL, rst_reg, 0);
+       WREG32_SMC(ixSMC_SYSCON_RESET_CNTL, val);
+
+       /* Set SMU Auto Start */
+       val = RREG32_SMC(ixSMU_INPUT_DATA);
+       val = REG_SET_FIELD(val, SMU_INPUT_DATA, AUTO_START, 1);
+       WREG32_SMC(ixSMU_INPUT_DATA, val);
+
+       /* Clear firmware interrupt enable flag */
+       WREG32_SMC(ixFIRMWARE_FLAGS, 0);
+
+       for (i = 0; i < adev->usec_timeout; i++) {
+               val = RREG32_SMC(ixRCU_UC_EVENTS);
+               if (REG_GET_FIELD(val, RCU_UC_EVENTS, INTERRUPTS_ENABLED))
+                       break;
+               udelay(1);
+       }
+
+       if (i == adev->usec_timeout) {
+               DRM_ERROR("Interrupt is not enabled by firmware\n");
+               return -EINVAL;
+       }
+
+       /* Call Test SMU message with 0x20000 offset
+        * to trigger SMU start
+        */
+       fiji_send_msg_to_smc_offset(adev);
+       DRM_INFO("[FM]try triger smu start\n");
+       /* Wait for done bit to be set */
+       for (i = 0; i < adev->usec_timeout; i++) {
+               val = RREG32_SMC(ixSMU_STATUS);
+               if (REG_GET_FIELD(val, SMU_STATUS, SMU_DONE))
+                       break;
+               udelay(1);
+       }
+
+       if (i == adev->usec_timeout) {
+               DRM_ERROR("Timeout for SMU start\n");
+               return -EINVAL;
+       }
+
+       /* Check pass/failed indicator */
+       val = RREG32_SMC(ixSMU_STATUS);
+       if (!REG_GET_FIELD(val, SMU_STATUS, SMU_PASS)) {
+               DRM_ERROR("SMU Firmware start failed\n");
+               return -EINVAL;
+       }
+       DRM_INFO("[FM]smu started\n");
+       /* Wait for firmware to initialize */
+       for (i = 0; i < adev->usec_timeout; i++) {
+               val = RREG32_SMC(ixFIRMWARE_FLAGS);
+               if(REG_GET_FIELD(val, FIRMWARE_FLAGS, INTERRUPTS_ENABLED))
+                       break;
+               udelay(1);
+       }
+
+       if (i == adev->usec_timeout) {
+               DRM_ERROR("SMU firmware initialization failed\n");
+               return -EINVAL;
+       }
+       DRM_INFO("[FM]smu initialized\n");
+
+       return 0;
+}
+
+static int fiji_smu_start_in_non_protection_mode(struct amdgpu_device *adev)
+{
+       int i, result;
+       uint32_t val;
+
+       /* wait for smc boot up */
+       for (i = 0; i < adev->usec_timeout; i++) {
+               val = RREG32_SMC(ixRCU_UC_EVENTS);
+               val = REG_GET_FIELD(val, RCU_UC_EVENTS, boot_seq_done);
+               if (val)
+                       break;
+               udelay(1);
+       }
+
+       if (i == adev->usec_timeout) {
+               DRM_ERROR("SMC boot sequence is not completed\n");
+               return -EINVAL;
+       }
+
+       /* Clear firmware interrupt enable flag */
+       WREG32_SMC(ixFIRMWARE_FLAGS, 0);
+
+       /* Assert reset */
+       val = RREG32_SMC(ixSMC_SYSCON_RESET_CNTL);
+       val = REG_SET_FIELD(val, SMC_SYSCON_RESET_CNTL, rst_reg, 1);
+       WREG32_SMC(ixSMC_SYSCON_RESET_CNTL, val);
+
+       result = fiji_smu_upload_firmware_image(adev);
+       if (result)
+               return result;
+
+       /* Set smc instruct start point at 0x0 */
+       fiji_program_jump_on_start(adev);
+
+       /* Enable clock */
+       val = RREG32_SMC(ixSMC_SYSCON_CLOCK_CNTL_0);
+       val = REG_SET_FIELD(val, SMC_SYSCON_CLOCK_CNTL_0, ck_disable, 0);
+       WREG32_SMC(ixSMC_SYSCON_CLOCK_CNTL_0, val);
+
+       /* De-assert reset */
+       val = RREG32_SMC(ixSMC_SYSCON_RESET_CNTL);
+       val = REG_SET_FIELD(val, SMC_SYSCON_RESET_CNTL, rst_reg, 0);
+       WREG32_SMC(ixSMC_SYSCON_RESET_CNTL, val);
+
+       /* Wait for firmware to initialize */
+       for (i = 0; i < adev->usec_timeout; i++) {
+               val = RREG32_SMC(ixFIRMWARE_FLAGS);
+               if (REG_GET_FIELD(val, FIRMWARE_FLAGS, INTERRUPTS_ENABLED))
+                       break;
+               udelay(1);
+       }
+
+       if (i == adev->usec_timeout) {
+               DRM_ERROR("Timeout for SMC firmware initialization\n");
+               return -EINVAL;
+       }
+
+       return 0;
+}
+
+int fiji_smu_start(struct amdgpu_device *adev)
+{
+       int result;
+       uint32_t val;
+
+       if (!fiji_is_smc_ram_running(adev)) {
+               val = RREG32_SMC(ixSMU_FIRMWARE);
+               if (!REG_GET_FIELD(val, SMU_FIRMWARE, SMU_MODE)) {
+                       DRM_INFO("[FM]start smu in nonprotection mode\n");
+                       result = fiji_smu_start_in_non_protection_mode(adev);
+                       if (result)
+                               return result;
+               } else {
+                       DRM_INFO("[FM]start smu in protection mode\n");
+                       result = fiji_smu_start_in_protection_mode(adev);
+                       if (result)
+                               return result;
+               }
+       }
+
+       return fiji_smu_request_load_fw(adev);
+}
+
+static const struct amdgpu_smumgr_funcs fiji_smumgr_funcs = {
+       .check_fw_load_finish = fiji_smu_check_fw_load_finish,
+       .request_smu_load_fw = NULL,
+       .request_smu_specific_fw = NULL,
+};
+
+int fiji_smu_init(struct amdgpu_device *adev)
+{
+       struct fiji_smu_private_data *private;
+       uint32_t image_size = ((sizeof(struct SMU_DRAMData_TOC) / 4096) + 1) * 4096;
+       uint32_t smu_internal_buffer_size = 200*4096;
+       struct amdgpu_bo **toc_buf = &adev->smu.toc_buf;
+       struct amdgpu_bo **smu_buf = &adev->smu.smu_buf;
+       uint64_t mc_addr;
+       void *toc_buf_ptr;
+       void *smu_buf_ptr;
+       int ret;
+
+       private = kzalloc(sizeof(struct fiji_smu_private_data), GFP_KERNEL);
+       if (NULL == private)
+               return -ENOMEM;
+
+       /* allocate firmware buffers */
+       if (adev->firmware.smu_load)
+               amdgpu_ucode_init_bo(adev);
+
+       adev->smu.priv = private;
+       adev->smu.fw_flags = 0;
+
+       /* Allocate FW image data structure and header buffer */
+       ret = amdgpu_bo_create(adev, image_size, PAGE_SIZE,
+                              true, AMDGPU_GEM_DOMAIN_VRAM,
+                              AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
+                              NULL, NULL, toc_buf);
+       if (ret) {
+               DRM_ERROR("Failed to allocate memory for TOC buffer\n");
+               return -ENOMEM;
+       }
+
+       /* Allocate buffer for SMU internal buffer */
+       ret = amdgpu_bo_create(adev, smu_internal_buffer_size, PAGE_SIZE,
+                              true, AMDGPU_GEM_DOMAIN_VRAM,
+                              AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
+                              NULL, NULL, smu_buf);
+       if (ret) {
+               DRM_ERROR("Failed to allocate memory for SMU internal buffer\n");
+               return -ENOMEM;
+       }
+
+       /* Retrieve GPU address for header buffer and internal buffer */
+       ret = amdgpu_bo_reserve(adev->smu.toc_buf, false);
+       if (ret) {
+               amdgpu_bo_unref(&adev->smu.toc_buf);
+               DRM_ERROR("Failed to reserve the TOC buffer\n");
+               return -EINVAL;
+       }
+
+       ret = amdgpu_bo_pin(adev->smu.toc_buf, AMDGPU_GEM_DOMAIN_VRAM, &mc_addr);
+       if (ret) {
+               amdgpu_bo_unreserve(adev->smu.toc_buf);
+               amdgpu_bo_unref(&adev->smu.toc_buf);
+               DRM_ERROR("Failed to pin the TOC buffer\n");
+               return -EINVAL;
+       }
+
+       ret = amdgpu_bo_kmap(*toc_buf, &toc_buf_ptr);
+       if (ret) {
+               amdgpu_bo_unreserve(adev->smu.toc_buf);
+               amdgpu_bo_unref(&adev->smu.toc_buf);
+               DRM_ERROR("Failed to map the TOC buffer\n");
+               return -EINVAL;
+       }
+
+       amdgpu_bo_unreserve(adev->smu.toc_buf);
+       private->header_addr_low = lower_32_bits(mc_addr);
+       private->header_addr_high = upper_32_bits(mc_addr);
+       private->header = toc_buf_ptr;
+
+       ret = amdgpu_bo_reserve(adev->smu.smu_buf, false);
+       if (ret) {
+               amdgpu_bo_unref(&adev->smu.smu_buf);
+               amdgpu_bo_unref(&adev->smu.toc_buf);
+               DRM_ERROR("Failed to reserve the SMU internal buffer\n");
+               return -EINVAL;
+       }
+
+       ret = amdgpu_bo_pin(adev->smu.smu_buf, AMDGPU_GEM_DOMAIN_VRAM, &mc_addr);
+       if (ret) {
+               amdgpu_bo_unreserve(adev->smu.smu_buf);
+               amdgpu_bo_unref(&adev->smu.smu_buf);
+               amdgpu_bo_unref(&adev->smu.toc_buf);
+               DRM_ERROR("Failed to pin the SMU internal buffer\n");
+               return -EINVAL;
+       }
+
+       ret = amdgpu_bo_kmap(*smu_buf, &smu_buf_ptr);
+       if (ret) {
+               amdgpu_bo_unreserve(adev->smu.smu_buf);
+               amdgpu_bo_unref(&adev->smu.smu_buf);
+               amdgpu_bo_unref(&adev->smu.toc_buf);
+               DRM_ERROR("Failed to map the SMU internal buffer\n");
+               return -EINVAL;
+       }
+
+       amdgpu_bo_unreserve(adev->smu.smu_buf);
+       private->smu_buffer_addr_low = lower_32_bits(mc_addr);
+       private->smu_buffer_addr_high = upper_32_bits(mc_addr);
+
+       adev->smu.smumgr_funcs = &fiji_smumgr_funcs;
+
+       return 0;
+}
+
+int fiji_smu_fini(struct amdgpu_device *adev)
+{
+       amdgpu_bo_unref(&adev->smu.toc_buf);
+       amdgpu_bo_unref(&adev->smu.smu_buf);
+       kfree(adev->smu.priv);
+       adev->smu.priv = NULL;
+       if (adev->firmware.fw_buf)
+               amdgpu_ucode_fini_bo(adev);
+
+       return 0;
+}