drm/amdgpu: add SDMA v4.0 implementation (v2)
authorKen Wang <Qingqing.Wang@amd.com>
Fri, 3 Mar 2017 22:31:51 +0000 (17:31 -0500)
committerAlex Deucher <alexander.deucher@amd.com>
Thu, 30 Mar 2017 03:54:45 +0000 (23:54 -0400)
v2: fix Makefile

Signed-off-by: Ken Wang <Qingqing.Wang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
drivers/gpu/drm/amd/amdgpu/Makefile
drivers/gpu/drm/amd/amdgpu/sdma_v4_0.c [new file with mode: 0644]
drivers/gpu/drm/amd/amdgpu/sdma_v4_0.h [new file with mode: 0644]

index ccbc4c0f5b4276b32b6e7b43260e7b8129f548ab..9c117d03b0e3891e1a625e854e860470ca1a7742 100644 (file)
@@ -69,7 +69,8 @@ amdgpu-y += \
 # add async DMA block
 amdgpu-y += \
        sdma_v2_4.o \
-       sdma_v3_0.o
+       sdma_v3_0.o \
+       sdma_v4_0.o
 
 # add UVD block
 amdgpu-y += \
diff --git a/drivers/gpu/drm/amd/amdgpu/sdma_v4_0.c b/drivers/gpu/drm/amd/amdgpu/sdma_v4_0.c
new file mode 100644 (file)
index 0000000..5bb8438
--- /dev/null
@@ -0,0 +1,1553 @@
+/*
+ * Copyright 2016 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 <drm/drmP.h>
+#include "amdgpu.h"
+#include "amdgpu_ucode.h"
+#include "amdgpu_trace.h"
+
+#include "vega10/soc15ip.h"
+#include "vega10/SDMA0/sdma0_4_0_offset.h"
+#include "vega10/SDMA0/sdma0_4_0_sh_mask.h"
+#include "vega10/SDMA1/sdma1_4_0_offset.h"
+#include "vega10/SDMA1/sdma1_4_0_sh_mask.h"
+#include "vega10/MMHUB/mmhub_1_0_offset.h"
+#include "vega10/MMHUB/mmhub_1_0_sh_mask.h"
+#include "vega10/HDP/hdp_4_0_offset.h"
+
+#include "soc15_common.h"
+#include "soc15.h"
+#include "vega10_sdma_pkt_open.h"
+
+MODULE_FIRMWARE("amdgpu/vega10_sdma.bin");
+MODULE_FIRMWARE("amdgpu/vega10_sdma1.bin");
+
+static void sdma_v4_0_set_ring_funcs(struct amdgpu_device *adev);
+static void sdma_v4_0_set_buffer_funcs(struct amdgpu_device *adev);
+static void sdma_v4_0_set_vm_pte_funcs(struct amdgpu_device *adev);
+static void sdma_v4_0_set_irq_funcs(struct amdgpu_device *adev);
+
+static const u32 golden_settings_sdma_4[] =
+{
+       SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_CHICKEN_BITS), 0xfe931f07, 0x02831f07,
+       SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_CLK_CTRL), 0xff000ff0, 0x3f000100,
+       SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_GFX_IB_CNTL), 0x800f0100, 0x00000100,
+       SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_GFX_RB_WPTR_POLL_CNTL), 0xfffffff7, 0x00403000,
+       SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_PAGE_IB_CNTL), 0x800f0100, 0x00000100,
+       SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_PAGE_RB_WPTR_POLL_CNTL), 0x0000fff0, 0x00403000,
+       SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_POWER_CNTL), 0x003ff006, 0x0003c000,
+       SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_RLC0_IB_CNTL), 0x800f0100, 0x00000100,
+       SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_RLC0_RB_WPTR_POLL_CNTL), 0x0000fff0, 0x00403000,
+       SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_RLC1_IB_CNTL), 0x800f0100, 0x00000100,
+       SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_RLC1_RB_WPTR_POLL_CNTL), 0x0000fff0, 0x00403000,
+       SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_UTCL1_PAGE), 0x000003ff, 0x000003c0,
+       SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_CHICKEN_BITS), 0xfe931f07, 0x02831f07,
+       SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_CLK_CTRL), 0xffffffff, 0x3f000100,
+       SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_GFX_IB_CNTL), 0x800f0100, 0x00000100,
+       SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_GFX_RB_WPTR_POLL_CNTL), 0x0000fff0, 0x00403000,
+       SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_PAGE_IB_CNTL), 0x800f0100, 0x00000100,
+       SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_PAGE_RB_WPTR_POLL_CNTL), 0x0000fff0, 0x00403000,
+       SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_POWER_CNTL), 0x003ff000, 0x0003c000,
+       SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_RLC0_IB_CNTL), 0x800f0100, 0x00000100,
+       SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_RLC0_RB_WPTR_POLL_CNTL), 0x0000fff0, 0x00403000,
+       SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_RLC1_IB_CNTL), 0x800f0100, 0x00000100,
+       SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_RLC1_RB_WPTR_POLL_CNTL), 0x0000fff0, 0x00403000,
+       SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_UTCL1_PAGE), 0x000003ff, 0x000003c0
+};
+
+static const u32 golden_settings_sdma_vg10[] =
+{
+       SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG), 0x0018773f, 0x00104002,
+       SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG_READ), 0x0018773f, 0x00104002,
+       SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_GB_ADDR_CONFIG), 0x0018773f, 0x00104002,
+       SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_GB_ADDR_CONFIG_READ), 0x0018773f, 0x00104002
+};
+
+static u32 sdma_v4_0_get_reg_offset(u32 instance, u32 internal_offset)
+{
+       u32 base = 0;
+       switch (instance) {
+               case 0:
+                       base = SDMA0_BASE.instance[0].segment[0];
+                       break;
+               case 1:
+                       base = SDMA1_BASE.instance[0].segment[0];
+                       break;
+               default:
+                       BUG();
+                       break;
+       }
+
+       return base + internal_offset;
+}
+
+static void sdma_v4_0_init_golden_registers(struct amdgpu_device *adev)
+{
+       switch (adev->asic_type) {
+       case CHIP_VEGA10:
+               amdgpu_program_register_sequence(adev,
+                                                golden_settings_sdma_4,
+                                                (const u32)ARRAY_SIZE(golden_settings_sdma_4));
+               amdgpu_program_register_sequence(adev,
+                                                golden_settings_sdma_vg10,
+                                                (const u32)ARRAY_SIZE(golden_settings_sdma_vg10));
+               break;
+       default:
+               break;
+       }
+}
+
+static void sdma_v4_0_print_ucode_regs(void *handle)
+{
+       int i;
+       struct amdgpu_device *adev = (struct amdgpu_device *)handle;
+
+       dev_info(adev->dev, "VEGA10 SDMA ucode registers\n");
+       for (i = 0; i < adev->sdma.num_instances; i++) {
+               dev_info(adev->dev, "  SDMA%d_UCODE_ADDR=0x%08X\n",
+                        i, RREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_UCODE_ADDR)));
+               dev_info(adev->dev, "  SDMA%d_UCODE_CHECKSUM=0x%08X\n",
+                        i, RREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_UCODE_CHECKSUM)));
+       }
+}
+
+/**
+ * sdma_v4_0_init_microcode - load ucode images from disk
+ *
+ * @adev: amdgpu_device pointer
+ *
+ * Use the firmware interface to load the ucode images into
+ * the driver (not loaded into hw).
+ * Returns 0 on success, error on failure.
+ */
+
+// emulation only, won't work on real chip
+// vega10 real chip need to use PSP to load firmware
+static int sdma_v4_0_init_microcode(struct amdgpu_device *adev)
+{
+       const char *chip_name;
+       char fw_name[30];
+       int err = 0, i;
+       struct amdgpu_firmware_info *info = NULL;
+       const struct common_firmware_header *header = NULL;
+       const struct sdma_firmware_header_v1_0 *hdr;
+
+       DRM_DEBUG("\n");
+
+       switch (adev->asic_type) {
+       case CHIP_VEGA10:
+               chip_name = "vega10";
+               break;
+       default: BUG();
+       }
+
+       for (i = 0; i < adev->sdma.num_instances; i++) {
+               if (i == 0)
+                       snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma.bin", chip_name);
+               else
+                       snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma1.bin", chip_name);
+               err = request_firmware(&adev->sdma.instance[i].fw, fw_name, adev->dev);
+               if (err)
+                       goto out;
+               err = amdgpu_ucode_validate(adev->sdma.instance[i].fw);
+               if (err)
+                       goto out;
+               hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
+               adev->sdma.instance[i].fw_version = le32_to_cpu(hdr->header.ucode_version);
+               adev->sdma.instance[i].feature_version = le32_to_cpu(hdr->ucode_feature_version);
+               if (adev->sdma.instance[i].feature_version >= 20)
+                       adev->sdma.instance[i].burst_nop = true;
+               DRM_DEBUG("psp_load == '%s'\n",
+                               adev->firmware.load_type == AMDGPU_FW_LOAD_PSP? "true": "false");
+
+               if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
+                       info = &adev->firmware.ucode[AMDGPU_UCODE_ID_SDMA0 + i];
+                       info->ucode_id = AMDGPU_UCODE_ID_SDMA0 + i;
+                       info->fw = adev->sdma.instance[i].fw;
+                       header = (const struct common_firmware_header *)info->fw->data;
+                       adev->firmware.fw_size +=
+                               ALIGN(le32_to_cpu(header->ucode_size_bytes), PAGE_SIZE);
+               }
+       }
+out:
+       if (err) {
+               printk(KERN_ERR
+                      "sdma_v4_0: Failed to load firmware \"%s\"\n",
+                      fw_name);
+               for (i = 0; i < adev->sdma.num_instances; i++) {
+                       release_firmware(adev->sdma.instance[i].fw);
+                       adev->sdma.instance[i].fw = NULL;
+               }
+       }
+       return err;
+}
+
+/**
+ * sdma_v4_0_ring_get_rptr - get the current read pointer
+ *
+ * @ring: amdgpu ring pointer
+ *
+ * Get the current rptr from the hardware (VEGA10+).
+ */
+static uint64_t sdma_v4_0_ring_get_rptr(struct amdgpu_ring *ring)
+{
+       u64* rptr;
+
+       /* XXX check if swapping is necessary on BE */
+       rptr =((u64*)&ring->adev->wb.wb[ring->rptr_offs]);
+
+       DRM_DEBUG("rptr before shift == 0x%016llx\n", *rptr);
+       return ((*rptr) >> 2);
+}
+
+/**
+ * sdma_v4_0_ring_get_wptr - get the current write pointer
+ *
+ * @ring: amdgpu ring pointer
+ *
+ * Get the current wptr from the hardware (VEGA10+).
+ */
+static uint64_t sdma_v4_0_ring_get_wptr(struct amdgpu_ring *ring)
+{
+       struct amdgpu_device *adev = ring->adev;
+       u64* wptr = NULL;
+       uint64_t local_wptr=0;
+
+       if (ring->use_doorbell) {
+               /* XXX check if swapping is necessary on BE */
+               wptr = ((u64*)&adev->wb.wb[ring->wptr_offs]);
+               DRM_DEBUG("wptr/doorbell before shift == 0x%016llx\n", *wptr);
+               *wptr = (*wptr) >> 2;
+               DRM_DEBUG("wptr/doorbell after shift == 0x%016llx\n", *wptr);
+       } else {
+               u32 lowbit, highbit;
+               int me = (ring == &adev->sdma.instance[0].ring) ? 0 : 1;
+               wptr=&local_wptr;
+               lowbit = RREG32(sdma_v4_0_get_reg_offset(me, mmSDMA0_GFX_RB_WPTR)) >> 2;
+               highbit = RREG32(sdma_v4_0_get_reg_offset(me, mmSDMA0_GFX_RB_WPTR_HI)) >> 2;
+
+               DRM_DEBUG("wptr [%i]high== 0x%08x low==0x%08x\n",
+                               me, highbit, lowbit);
+               *wptr = highbit;
+               *wptr = (*wptr) << 32;
+               *wptr |= lowbit;
+       }
+
+       return *wptr;
+}
+
+/**
+ * sdma_v4_0_ring_set_wptr - commit the write pointer
+ *
+ * @ring: amdgpu ring pointer
+ *
+ * Write the wptr back to the hardware (VEGA10+).
+ */
+static void sdma_v4_0_ring_set_wptr(struct amdgpu_ring *ring)
+{
+       struct amdgpu_device *adev = ring->adev;
+
+       DRM_DEBUG("Setting write pointer\n");
+       if (ring->use_doorbell) {
+               DRM_DEBUG("Using doorbell -- "
+                               "wptr_offs == 0x%08x "
+                               "lower_32_bits(ring->wptr) << 2 == 0x%08x "
+                               "upper_32_bits(ring->wptr) << 2 == 0x%08x\n",
+                               ring->wptr_offs,
+                               lower_32_bits(ring->wptr << 2),
+                               upper_32_bits(ring->wptr << 2));
+               /* XXX check if swapping is necessary on BE */
+               adev->wb.wb[ring->wptr_offs] = lower_32_bits(ring->wptr << 2);
+               adev->wb.wb[ring->wptr_offs + 1] = upper_32_bits(ring->wptr << 2);
+               DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n",
+                               ring->doorbell_index, ring->wptr << 2);
+               WDOORBELL64(ring->doorbell_index, ring->wptr << 2);
+       } else {
+               int me = (ring == &ring->adev->sdma.instance[0].ring) ? 0 : 1;
+               DRM_DEBUG("Not using doorbell -- "
+                               "mmSDMA%i_GFX_RB_WPTR == 0x%08x "
+                               "mmSDMA%i_GFX_RB_WPTR_HI == 0x%08x \n",
+                               me,
+                               me,
+                               lower_32_bits(ring->wptr << 2),
+                               upper_32_bits(ring->wptr << 2));
+               WREG32(sdma_v4_0_get_reg_offset(me, mmSDMA0_GFX_RB_WPTR), lower_32_bits(ring->wptr << 2));
+               WREG32(sdma_v4_0_get_reg_offset(me, mmSDMA0_GFX_RB_WPTR_HI), upper_32_bits(ring->wptr << 2));
+       }
+}
+
+static void sdma_v4_0_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
+{
+       struct amdgpu_sdma_instance *sdma = amdgpu_get_sdma_instance(ring);
+       int i;
+
+       for (i = 0; i < count; i++)
+               if (sdma && sdma->burst_nop && (i == 0))
+                       amdgpu_ring_write(ring, ring->funcs->nop |
+                               SDMA_PKT_NOP_HEADER_COUNT(count - 1));
+               else
+                       amdgpu_ring_write(ring, ring->funcs->nop);
+}
+
+/**
+ * sdma_v4_0_ring_emit_ib - Schedule an IB on the DMA engine
+ *
+ * @ring: amdgpu ring pointer
+ * @ib: IB object to schedule
+ *
+ * Schedule an IB in the DMA ring (VEGA10).
+ */
+static void sdma_v4_0_ring_emit_ib(struct amdgpu_ring *ring,
+                                   struct amdgpu_ib *ib,
+                                   unsigned vm_id, bool ctx_switch)
+{
+        u32 vmid = vm_id & 0xf;
+
+        /* IB packet must end on a 8 DW boundary */
+        sdma_v4_0_ring_insert_nop(ring, (10 - (lower_32_bits(ring->wptr) & 7)) % 8);
+
+        amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) |
+                          SDMA_PKT_INDIRECT_HEADER_VMID(vmid));
+        /* base must be 32 byte aligned */
+        amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
+        amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
+        amdgpu_ring_write(ring, ib->length_dw);
+        amdgpu_ring_write(ring, 0);
+        amdgpu_ring_write(ring, 0);
+
+}
+
+/**
+ * sdma_v4_0_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
+ *
+ * @ring: amdgpu ring pointer
+ *
+ * Emit an hdp flush packet on the requested DMA ring.
+ */
+static void sdma_v4_0_ring_emit_hdp_flush(struct amdgpu_ring *ring)
+{
+       u32 ref_and_mask = 0;
+       struct nbio_hdp_flush_reg *nbio_hf_reg;
+
+       if (ring->adev->asic_type == CHIP_VEGA10)
+               nbio_hf_reg = &nbio_v6_1_hdp_flush_reg;
+
+       if (ring == &ring->adev->sdma.instance[0].ring)
+               ref_and_mask = nbio_hf_reg->ref_and_mask_sdma0;
+       else
+               ref_and_mask = nbio_hf_reg->ref_and_mask_sdma1;
+
+       amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
+                         SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
+                         SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
+       amdgpu_ring_write(ring, nbio_hf_reg->hdp_flush_done_offset << 2);
+       amdgpu_ring_write(ring, nbio_hf_reg->hdp_flush_req_offset << 2);
+       amdgpu_ring_write(ring, ref_and_mask); /* reference */
+       amdgpu_ring_write(ring, ref_and_mask); /* mask */
+       amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
+                         SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
+}
+
+static void sdma_v4_0_ring_emit_hdp_invalidate(struct amdgpu_ring *ring)
+{
+       amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
+                         SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
+       amdgpu_ring_write(ring, SOC15_REG_OFFSET(HDP, 0, mmHDP_DEBUG0));
+       amdgpu_ring_write(ring, 1);
+}
+
+/**
+ * sdma_v4_0_ring_emit_fence - emit a fence on the DMA ring
+ *
+ * @ring: amdgpu ring pointer
+ * @fence: amdgpu fence object
+ *
+ * Add a DMA fence packet to the ring to write
+ * the fence seq number and DMA trap packet to generate
+ * an interrupt if needed (VEGA10).
+ */
+static void sdma_v4_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
+                                     unsigned flags)
+{
+       bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
+       /* write the fence */
+       amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE));
+       /* zero in first two bits */
+       BUG_ON(addr & 0x3);
+       amdgpu_ring_write(ring, lower_32_bits(addr));
+       amdgpu_ring_write(ring, upper_32_bits(addr));
+       amdgpu_ring_write(ring, lower_32_bits(seq));
+
+       /* optionally write high bits as well */
+       if (write64bit) {
+               addr += 4;
+               amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE));
+               /* zero in first two bits */
+               BUG_ON(addr & 0x3);
+               amdgpu_ring_write(ring, lower_32_bits(addr));
+               amdgpu_ring_write(ring, upper_32_bits(addr));
+               amdgpu_ring_write(ring, upper_32_bits(seq));
+       }
+
+       /* generate an interrupt */
+       amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP));
+       amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(0));
+}
+
+
+/**
+ * sdma_v4_0_gfx_stop - stop the gfx async dma engines
+ *
+ * @adev: amdgpu_device pointer
+ *
+ * Stop the gfx async dma ring buffers (VEGA10).
+ */
+static void sdma_v4_0_gfx_stop(struct amdgpu_device *adev)
+{
+       struct amdgpu_ring *sdma0 = &adev->sdma.instance[0].ring;
+       struct amdgpu_ring *sdma1 = &adev->sdma.instance[1].ring;
+       u32 rb_cntl, ib_cntl;
+       int i;
+
+       if ((adev->mman.buffer_funcs_ring == sdma0) ||
+           (adev->mman.buffer_funcs_ring == sdma1))
+               amdgpu_ttm_set_active_vram_size(adev, adev->mc.visible_vram_size);
+
+       for (i = 0; i < adev->sdma.num_instances; i++) {
+               rb_cntl = RREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_GFX_RB_CNTL));
+               rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0);
+               WREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
+               ib_cntl = RREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_GFX_IB_CNTL));
+               ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0);
+               WREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_GFX_IB_CNTL), ib_cntl);
+       }
+
+       sdma0->ready = false;
+       sdma1->ready = false;
+}
+
+/**
+ * sdma_v4_0_rlc_stop - stop the compute async dma engines
+ *
+ * @adev: amdgpu_device pointer
+ *
+ * Stop the compute async dma queues (VEGA10).
+ */
+static void sdma_v4_0_rlc_stop(struct amdgpu_device *adev)
+{
+       /* XXX todo */
+}
+
+/**
+ * sdma_v_0_ctx_switch_enable - stop the async dma engines context switch
+ *
+ * @adev: amdgpu_device pointer
+ * @enable: enable/disable the DMA MEs context switch.
+ *
+ * Halt or unhalt the async dma engines context switch (VEGA10).
+ */
+static void sdma_v4_0_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
+{
+       u32 f32_cntl;
+       int i;
+
+       for (i = 0; i < adev->sdma.num_instances; i++) {
+               f32_cntl = RREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_CNTL));
+               f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
+                               AUTO_CTXSW_ENABLE, enable ? 1 : 0);
+               WREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_CNTL), f32_cntl);
+       }
+
+}
+
+/**
+ * sdma_v4_0_enable - stop the async dma engines
+ *
+ * @adev: amdgpu_device pointer
+ * @enable: enable/disable the DMA MEs.
+ *
+ * Halt or unhalt the async dma engines (VEGA10).
+ */
+static void sdma_v4_0_enable(struct amdgpu_device *adev, bool enable)
+{
+       u32 f32_cntl;
+       int i;
+
+       if (enable == false) {
+               sdma_v4_0_gfx_stop(adev);
+               sdma_v4_0_rlc_stop(adev);
+       }
+
+       for (i = 0; i < adev->sdma.num_instances; i++) {
+               f32_cntl = RREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_F32_CNTL));
+               f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, enable ? 0 : 1);
+               WREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_F32_CNTL), f32_cntl);
+       }
+}
+
+/**
+ * sdma_v4_0_gfx_resume - setup and start the async dma engines
+ *
+ * @adev: amdgpu_device pointer
+ *
+ * Set up the gfx DMA ring buffers and enable them (VEGA10).
+ * Returns 0 for success, error for failure.
+ */
+static int sdma_v4_0_gfx_resume(struct amdgpu_device *adev)
+{
+       struct amdgpu_ring *ring;
+       u32 rb_cntl, ib_cntl;
+       u32 rb_bufsz;
+       u32 wb_offset;
+       u32 doorbell;
+       u32 doorbell_offset;
+       int i,r;
+
+       for (i = 0; i < adev->sdma.num_instances; i++) {
+               ring = &adev->sdma.instance[i].ring;
+               wb_offset = (ring->rptr_offs * 4);
+
+               WREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL), 0);
+
+               /* Set ring buffer size in dwords */
+               rb_bufsz = order_base_2(ring->ring_size / 4);
+               rb_cntl = RREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_GFX_RB_CNTL));
+               rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz);
+#ifdef __BIG_ENDIAN
+               rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1);
+               rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL,
+                                       RPTR_WRITEBACK_SWAP_ENABLE, 1);
+#endif
+               WREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
+
+               /* Initialize the ring buffer's read and write pointers */
+               WREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_GFX_RB_RPTR), 0);
+               WREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_GFX_RB_RPTR_HI), 0);
+               WREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_GFX_RB_WPTR), 0);
+               WREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_GFX_RB_WPTR_HI), 0);
+
+               /* set the wb address whether it's enabled or not */
+               WREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_GFX_RB_RPTR_ADDR_HI),
+                      upper_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFF);
+               WREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_GFX_RB_RPTR_ADDR_LO),
+                      lower_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC);
+
+               rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
+
+               WREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_GFX_RB_BASE), ring->gpu_addr >> 8);
+               WREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_GFX_RB_BASE_HI), ring->gpu_addr >> 40);
+
+               ring->wptr = 0;
+               WREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_GFX_RB_WPTR), lower_32_bits(ring->wptr) << 2);
+               WREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_GFX_RB_WPTR_HI), upper_32_bits(ring->wptr) << 2);
+
+               doorbell = RREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_GFX_DOORBELL));
+               doorbell_offset = RREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_GFX_DOORBELL_OFFSET));
+
+               if (ring->use_doorbell){
+                       doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 1);
+                       doorbell_offset = REG_SET_FIELD(doorbell_offset, SDMA0_GFX_DOORBELL_OFFSET,
+                                       OFFSET, ring->doorbell_index);
+               } else {
+                       doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 0);
+               }
+               WREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_GFX_DOORBELL), doorbell);
+               WREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_GFX_DOORBELL_OFFSET), doorbell_offset);
+               nbio_v6_1_sdma_doorbell_range(adev, i, ring->use_doorbell, ring->doorbell_index);
+
+               /* enable DMA RB */
+               rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1);
+               WREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
+
+               ib_cntl = RREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_GFX_IB_CNTL));
+               ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1);
+#ifdef __BIG_ENDIAN
+               ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1);
+#endif
+               /* enable DMA IBs */
+               WREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_GFX_IB_CNTL), ib_cntl);
+
+               ring->ready = true;
+
+               r = amdgpu_ring_test_ring(ring);
+               if (r) {
+                       ring->ready = false;
+                       return r;
+               }
+
+               if (adev->mman.buffer_funcs_ring == ring)
+                       amdgpu_ttm_set_active_vram_size(adev, adev->mc.real_vram_size);
+       }
+
+       return 0;
+}
+
+/**
+ * sdma_v4_0_rlc_resume - setup and start the async dma engines
+ *
+ * @adev: amdgpu_device pointer
+ *
+ * Set up the compute DMA queues and enable them (VEGA10).
+ * Returns 0 for success, error for failure.
+ */
+static int sdma_v4_0_rlc_resume(struct amdgpu_device *adev)
+{
+       /* XXX todo */
+       return 0;
+}
+
+/**
+ * sdma_v4_0_load_microcode - load the sDMA ME ucode
+ *
+ * @adev: amdgpu_device pointer
+ *
+ * Loads the sDMA0/1 ucode.
+ * Returns 0 for success, -EINVAL if the ucode is not available.
+ */
+static int sdma_v4_0_load_microcode(struct amdgpu_device *adev)
+{
+       const struct sdma_firmware_header_v1_0 *hdr;
+       const __le32 *fw_data;
+       u32 fw_size;
+       u32 digest_size = 0;
+       int i, j;
+
+       /* halt the MEs */
+       sdma_v4_0_enable(adev, false);
+
+       for (i = 0; i < adev->sdma.num_instances; i++) {
+               uint16_t version_major;
+               uint16_t version_minor;
+               if (!adev->sdma.instance[i].fw)
+                       return -EINVAL;
+
+               hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
+               amdgpu_ucode_print_sdma_hdr(&hdr->header);
+               fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
+
+               version_major = le16_to_cpu(hdr->header.header_version_major);
+               version_minor = le16_to_cpu(hdr->header.header_version_minor);
+
+               if (version_major == 1 && version_minor >= 1) {
+                       const struct sdma_firmware_header_v1_1 *sdma_v1_1_hdr = (const struct sdma_firmware_header_v1_1 *) hdr;
+                       digest_size = le32_to_cpu(sdma_v1_1_hdr->digest_size);
+               }
+
+               fw_size -= digest_size;
+
+               fw_data = (const __le32 *)
+                       (adev->sdma.instance[i].fw->data +
+                               le32_to_cpu(hdr->header.ucode_array_offset_bytes));
+
+               sdma_v4_0_print_ucode_regs(adev);
+
+               WREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_UCODE_ADDR), 0);
+
+
+               for (j = 0; j < fw_size; j++)
+               {
+                       WREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_UCODE_DATA), le32_to_cpup(fw_data++));
+               }
+
+               WREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_UCODE_ADDR), adev->sdma.instance[i].fw_version);
+       }
+
+       sdma_v4_0_print_ucode_regs(adev);
+
+       return 0;
+}
+
+/**
+ * sdma_v4_0_start - setup and start the async dma engines
+ *
+ * @adev: amdgpu_device pointer
+ *
+ * Set up the DMA engines and enable them (VEGA10).
+ * Returns 0 for success, error for failure.
+ */
+static int sdma_v4_0_start(struct amdgpu_device *adev)
+{
+       int r;
+
+       if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP) {
+               DRM_INFO("Loading via direct write\n");
+               r = sdma_v4_0_load_microcode(adev);
+               if (r)
+                       return r;
+       }
+
+       /* unhalt the MEs */
+       sdma_v4_0_enable(adev, true);
+       /* enable sdma ring preemption */
+       sdma_v4_0_ctx_switch_enable(adev, true);
+
+       /* start the gfx rings and rlc compute queues */
+       r = sdma_v4_0_gfx_resume(adev);
+       if (r)
+               return r;
+       r = sdma_v4_0_rlc_resume(adev);
+       if (r)
+               return r;
+
+       return 0;
+}
+
+/**
+ * sdma_v4_0_ring_test_ring - simple async dma engine test
+ *
+ * @ring: amdgpu_ring structure holding ring information
+ *
+ * Test the DMA engine by writing using it to write an
+ * value to memory. (VEGA10).
+ * Returns 0 for success, error for failure.
+ */
+static int sdma_v4_0_ring_test_ring(struct amdgpu_ring *ring)
+{
+       struct amdgpu_device *adev = ring->adev;
+       unsigned i;
+       unsigned index;
+       int r;
+       u32 tmp;
+       u64 gpu_addr;
+
+       DRM_INFO("In Ring test func\n");
+
+       r = amdgpu_wb_get(adev, &index);
+       if (r) {
+               dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
+               return r;
+       }
+
+       gpu_addr = adev->wb.gpu_addr + (index * 4);
+       tmp = 0xCAFEDEAD;
+       adev->wb.wb[index] = cpu_to_le32(tmp);
+
+       r = amdgpu_ring_alloc(ring, 5);
+       if (r) {
+               DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
+               amdgpu_wb_free(adev, index);
+               return r;
+       }
+
+       amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
+                         SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
+       amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
+       amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
+       amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0));
+       amdgpu_ring_write(ring, 0xDEADBEEF);
+       amdgpu_ring_commit(ring);
+
+       for (i = 0; i < adev->usec_timeout; i++) {
+               tmp = le32_to_cpu(adev->wb.wb[index]);
+               if (tmp == 0xDEADBEEF) {
+                       break;
+               }
+               DRM_UDELAY(1);
+       }
+
+       if (i < adev->usec_timeout) {
+               DRM_INFO("ring test on %d succeeded in %d usecs\n", ring->idx, i);
+       } else {
+               DRM_ERROR("amdgpu: ring %d test failed (0x%08X)\n",
+                         ring->idx, tmp);
+               r = -EINVAL;
+       }
+       amdgpu_wb_free(adev, index);
+
+       return r;
+}
+
+/**
+ * sdma_v4_0_ring_test_ib - test an IB on the DMA engine
+ *
+ * @ring: amdgpu_ring structure holding ring information
+ *
+ * Test a simple IB in the DMA ring (VEGA10).
+ * Returns 0 on success, error on failure.
+ */
+static int sdma_v4_0_ring_test_ib(struct amdgpu_ring *ring, long timeout)
+{
+       struct amdgpu_device *adev = ring->adev;
+       struct amdgpu_ib ib;
+       struct dma_fence *f = NULL;
+       unsigned index;
+       long r;
+       u32 tmp = 0;
+       u64 gpu_addr;
+
+       r = amdgpu_wb_get(adev, &index);
+       if (r) {
+               dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r);
+               return r;
+       }
+
+       gpu_addr = adev->wb.gpu_addr + (index * 4);
+       tmp = 0xCAFEDEAD;
+       adev->wb.wb[index] = cpu_to_le32(tmp);
+       memset(&ib, 0, sizeof(ib));
+       r = amdgpu_ib_get(adev, NULL, 256, &ib);
+       if (r) {
+               DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r);
+               goto err0;
+       }
+
+       ib.ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
+               SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
+       ib.ptr[1] = lower_32_bits(gpu_addr);
+       ib.ptr[2] = upper_32_bits(gpu_addr);
+       ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0);
+       ib.ptr[4] = 0xDEADBEEF;
+       ib.ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
+       ib.ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
+       ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
+       ib.length_dw = 8;
+
+       r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
+       if (r)
+               goto err1;
+
+        r = dma_fence_wait_timeout(f, false, timeout);
+        if (r == 0) {
+                DRM_ERROR("amdgpu: IB test timed out\n");
+                r = -ETIMEDOUT;
+                goto err1;
+        } else if (r < 0) {
+                DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
+                goto err1;
+        }
+        tmp = le32_to_cpu(adev->wb.wb[index]);
+        if (tmp == 0xDEADBEEF) {
+                DRM_INFO("ib test on ring %d succeeded\n", ring->idx);
+                r = 0;
+        } else {
+                DRM_ERROR("amdgpu: ib test failed (0x%08X)\n", tmp);
+                r = -EINVAL;
+        }
+err1:
+        amdgpu_ib_free(adev, &ib, NULL);
+        dma_fence_put(f);
+err0:
+        amdgpu_wb_free(adev, index);
+        return r;
+}
+
+
+/**
+ * sdma_v4_0_vm_copy_pte - update PTEs by copying them from the GART
+ *
+ * @ib: indirect buffer to fill with commands
+ * @pe: addr of the page entry
+ * @src: src addr to copy from
+ * @count: number of page entries to update
+ *
+ * Update PTEs by copying them from the GART using sDMA (VEGA10).
+ */
+static void sdma_v4_0_vm_copy_pte(struct amdgpu_ib *ib,
+                                 uint64_t pe, uint64_t src,
+                                 unsigned count)
+{
+       unsigned bytes = count * 8;
+
+       ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
+               SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
+       ib->ptr[ib->length_dw++] = bytes - 1;
+       ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
+       ib->ptr[ib->length_dw++] = lower_32_bits(src);
+       ib->ptr[ib->length_dw++] = upper_32_bits(src);
+       ib->ptr[ib->length_dw++] = lower_32_bits(pe);
+       ib->ptr[ib->length_dw++] = upper_32_bits(pe);
+
+}
+
+/**
+ * sdma_v4_0_vm_write_pte - update PTEs by writing them manually
+ *
+ * @ib: indirect buffer to fill with commands
+ * @pe: addr of the page entry
+ * @addr: dst addr to write into pe
+ * @count: number of page entries to update
+ * @incr: increase next addr by incr bytes
+ * @flags: access flags
+ *
+ * Update PTEs by writing them manually using sDMA (VEGA10).
+ */
+static void sdma_v4_0_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
+                                  uint64_t value, unsigned count,
+                                  uint32_t incr)
+{
+       unsigned ndw = count * 2;
+
+       ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
+               SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
+       ib->ptr[ib->length_dw++] = lower_32_bits(pe);
+       ib->ptr[ib->length_dw++] = upper_32_bits(pe);
+       ib->ptr[ib->length_dw++] = ndw - 1;
+       for (; ndw > 0; ndw -= 2) {
+               ib->ptr[ib->length_dw++] = lower_32_bits(value);
+               ib->ptr[ib->length_dw++] = upper_32_bits(value);
+               value += incr;
+       }
+}
+
+/**
+ * sdma_v4_0_vm_set_pte_pde - update the page tables using sDMA
+ *
+ * @ib: indirect buffer to fill with commands
+ * @pe: addr of the page entry
+ * @addr: dst addr to write into pe
+ * @count: number of page entries to update
+ * @incr: increase next addr by incr bytes
+ * @flags: access flags
+ *
+ * Update the page tables using sDMA (VEGA10).
+ */
+static void sdma_v4_0_vm_set_pte_pde(struct amdgpu_ib *ib,
+                                    uint64_t pe,
+                                    uint64_t addr, unsigned count,
+                                    uint32_t incr, uint64_t flags)
+{
+       /* for physically contiguous pages (vram) */
+       ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_PTEPDE);
+       ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
+       ib->ptr[ib->length_dw++] = upper_32_bits(pe);
+       ib->ptr[ib->length_dw++] = flags; /* mask */
+       ib->ptr[ib->length_dw++] = 0;
+       ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
+       ib->ptr[ib->length_dw++] = upper_32_bits(addr);
+       ib->ptr[ib->length_dw++] = incr; /* increment size */
+       ib->ptr[ib->length_dw++] = 0;
+       ib->ptr[ib->length_dw++] = count - 1; /* number of entries */
+}
+
+/**
+ * sdma_v4_0_ring_pad_ib - pad the IB to the required number of dw
+ *
+ * @ib: indirect buffer to fill with padding
+ *
+ */
+static void sdma_v4_0_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
+{
+       struct amdgpu_sdma_instance *sdma = amdgpu_get_sdma_instance(ring);
+       u32 pad_count;
+       int i;
+
+       pad_count = (8 - (ib->length_dw & 0x7)) % 8;
+       for (i = 0; i < pad_count; i++)
+               if (sdma && sdma->burst_nop && (i == 0))
+                       ib->ptr[ib->length_dw++] =
+                               SDMA_PKT_HEADER_OP(SDMA_OP_NOP) |
+                               SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
+               else
+                       ib->ptr[ib->length_dw++] =
+                               SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
+}
+
+
+/**
+ * sdma_v4_0_ring_emit_pipeline_sync - sync the pipeline
+ *
+ * @ring: amdgpu_ring pointer
+ *
+ * Make sure all previous operations are completed (CIK).
+ */
+static void sdma_v4_0_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
+{
+       uint32_t seq = ring->fence_drv.sync_seq;
+       uint64_t addr = ring->fence_drv.gpu_addr;
+
+       /* wait for idle */
+       amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
+                         SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
+                         SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
+                         SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1));
+       amdgpu_ring_write(ring, addr & 0xfffffffc);
+       amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
+       amdgpu_ring_write(ring, seq); /* reference */
+       amdgpu_ring_write(ring, 0xfffffff); /* mask */
+       amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
+                         SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */
+}
+
+
+/**
+ * sdma_v4_0_ring_emit_vm_flush - vm flush using sDMA
+ *
+ * @ring: amdgpu_ring pointer
+ * @vm: amdgpu_vm pointer
+ *
+ * Update the page table base and flush the VM TLB
+ * using sDMA (VEGA10).
+ */
+static void sdma_v4_0_ring_emit_vm_flush(struct amdgpu_ring *ring,
+                                        unsigned vm_id, uint64_t pd_addr)
+{
+       unsigned eng = ring->idx;
+       unsigned i;
+
+       pd_addr = pd_addr | 0x1; /* valid bit */
+       /* now only use physical base address of PDE and valid */
+       BUG_ON(pd_addr & 0xFFFF00000000003EULL);
+
+       for (i = 0; i < AMDGPU_MAX_VMHUBS; ++i) {
+               struct amdgpu_vmhub *hub = &ring->adev->vmhub[i];
+               uint32_t req = hub->get_invalidate_req(vm_id);
+
+               amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
+                                 SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
+               amdgpu_ring_write(ring, hub->ctx0_ptb_addr_lo32 + vm_id * 2);
+               amdgpu_ring_write(ring, lower_32_bits(pd_addr));
+
+               amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
+                                 SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
+               amdgpu_ring_write(ring, hub->ctx0_ptb_addr_hi32 + vm_id * 2);
+               amdgpu_ring_write(ring, upper_32_bits(pd_addr));
+
+               /* flush TLB */
+               amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
+                                 SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
+               amdgpu_ring_write(ring, hub->vm_inv_eng0_req + eng);
+               amdgpu_ring_write(ring, req);
+
+               /* wait for flush */
+               amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
+                                 SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
+                                 SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* equal */
+               amdgpu_ring_write(ring, (hub->vm_inv_eng0_ack + eng) << 2);
+               amdgpu_ring_write(ring, 0);
+               amdgpu_ring_write(ring, 1 << vm_id); /* reference */
+               amdgpu_ring_write(ring, 1 << vm_id); /* mask */
+               amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
+                                 SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10));
+       }
+}
+
+static int sdma_v4_0_early_init(void *handle)
+{
+       struct amdgpu_device *adev = (struct amdgpu_device *)handle;
+
+       adev->sdma.num_instances = 2;
+
+       sdma_v4_0_set_ring_funcs(adev);
+       sdma_v4_0_set_buffer_funcs(adev);
+       sdma_v4_0_set_vm_pte_funcs(adev);
+       sdma_v4_0_set_irq_funcs(adev);
+
+       return 0;
+}
+
+
+static int sdma_v4_0_sw_init(void *handle)
+{
+       struct amdgpu_ring *ring;
+       int r, i;
+       struct amdgpu_device *adev = (struct amdgpu_device *)handle;
+
+       /* SDMA trap event */
+       r = amdgpu_irq_add_id(adev, AMDGPU_IH_CLIENTID_SDMA0, 224,
+                             &adev->sdma.trap_irq);
+       if (r)
+               return r;
+
+       /* SDMA trap event */
+       r = amdgpu_irq_add_id(adev, AMDGPU_IH_CLIENTID_SDMA1, 224,
+                             &adev->sdma.trap_irq);
+       if (r)
+               return r;
+
+       r = sdma_v4_0_init_microcode(adev);
+       if (r) {
+               DRM_ERROR("Failed to load sdma firmware!\n");
+               return r;
+       }
+
+       for (i = 0; i < adev->sdma.num_instances; i++) {
+               ring = &adev->sdma.instance[i].ring;
+               ring->ring_obj = NULL;
+               ring->use_doorbell = true;
+
+               DRM_INFO("use_doorbell being set to: [%s]\n",
+                               ring->use_doorbell?"true":"false");
+
+               ring->doorbell_index = (i == 0) ?
+                       (AMDGPU_DOORBELL64_sDMA_ENGINE0 << 1) //get DWORD offset
+                       : (AMDGPU_DOORBELL64_sDMA_ENGINE1 << 1); // get DWORD offset
+
+               sprintf(ring->name, "sdma%d", i);
+               r = amdgpu_ring_init(adev, ring, 1024,
+                                    &adev->sdma.trap_irq,
+                                    (i == 0) ?
+                                    AMDGPU_SDMA_IRQ_TRAP0 :
+                                    AMDGPU_SDMA_IRQ_TRAP1);
+               if (r)
+                       return r;
+       }
+
+       return r;
+}
+
+static int sdma_v4_0_sw_fini(void *handle)
+{
+       struct amdgpu_device *adev = (struct amdgpu_device *)handle;
+       int i;
+
+       for (i = 0; i < adev->sdma.num_instances; i++)
+               amdgpu_ring_fini(&adev->sdma.instance[i].ring);
+
+       return 0;
+}
+
+static int sdma_v4_0_hw_init(void *handle)
+{
+       int r;
+       struct amdgpu_device *adev = (struct amdgpu_device *)handle;
+
+       sdma_v4_0_init_golden_registers(adev);
+
+       r = sdma_v4_0_start(adev);
+       if (r)
+               return r;
+
+       return r;
+}
+
+static int sdma_v4_0_hw_fini(void *handle)
+{
+       struct amdgpu_device *adev = (struct amdgpu_device *)handle;
+
+       sdma_v4_0_ctx_switch_enable(adev, false);
+       sdma_v4_0_enable(adev, false);
+
+       return 0;
+}
+
+static int sdma_v4_0_suspend(void *handle)
+{
+       struct amdgpu_device *adev = (struct amdgpu_device *)handle;
+
+       return sdma_v4_0_hw_fini(adev);
+}
+
+static int sdma_v4_0_resume(void *handle)
+{
+       struct amdgpu_device *adev = (struct amdgpu_device *)handle;
+
+       return sdma_v4_0_hw_init(adev);
+}
+
+static bool sdma_v4_0_is_idle(void *handle)
+{
+       struct amdgpu_device *adev = (struct amdgpu_device *)handle;
+       u32 i;
+       for (i = 0; i < adev->sdma.num_instances; i++) {
+               u32 tmp = RREG32(sdma_v4_0_get_reg_offset(i, mmSDMA0_STATUS_REG));
+               if (!(tmp & SDMA0_STATUS_REG__IDLE_MASK))
+                        return false;
+       }
+
+       return true;
+}
+
+static int sdma_v4_0_wait_for_idle(void *handle)
+{
+       unsigned i;
+       u32 sdma0,sdma1;
+       struct amdgpu_device *adev = (struct amdgpu_device *)handle;
+       for (i = 0; i < adev->usec_timeout; i++) {
+               sdma0 = RREG32(sdma_v4_0_get_reg_offset(0, mmSDMA0_STATUS_REG));
+               sdma1 = RREG32(sdma_v4_0_get_reg_offset(1, mmSDMA0_STATUS_REG));
+
+               if (sdma0 & sdma1 & SDMA0_STATUS_REG__IDLE_MASK)
+                       return 0;
+               udelay(1);
+       }
+       return -ETIMEDOUT;
+}
+
+static int sdma_v4_0_soft_reset(void *handle)
+{
+       /* todo */
+
+       return 0;
+}
+
+static int sdma_v4_0_set_trap_irq_state(struct amdgpu_device *adev,
+                                       struct amdgpu_irq_src *source,
+                                       unsigned type,
+                                       enum amdgpu_interrupt_state state)
+{
+       u32 sdma_cntl;
+
+       u32 reg_offset = (type == AMDGPU_SDMA_IRQ_TRAP0) ?
+               sdma_v4_0_get_reg_offset(0, mmSDMA0_CNTL) :
+               sdma_v4_0_get_reg_offset(1, mmSDMA0_CNTL);
+
+       sdma_cntl = RREG32(reg_offset);
+       sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE,
+                      state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
+       WREG32(reg_offset, sdma_cntl);
+
+       return 0;
+}
+
+static int sdma_v4_0_process_trap_irq(struct amdgpu_device *adev,
+                                     struct amdgpu_irq_src *source,
+                                     struct amdgpu_iv_entry *entry)
+{
+       DRM_DEBUG("IH: SDMA trap\n");
+       switch (entry->client_id) {
+       case AMDGPU_IH_CLIENTID_SDMA0:
+               switch (entry->ring_id) {
+               case 0:
+                       amdgpu_fence_process(&adev->sdma.instance[0].ring);
+                       break;
+               case 1:
+                       /* XXX compute */
+                       break;
+               case 2:
+                       /* XXX compute */
+                       break;
+               case 3:
+                       /* XXX page queue*/
+                       break;
+               }
+               break;
+       case AMDGPU_IH_CLIENTID_SDMA1:
+               switch (entry->ring_id) {
+               case 0:
+                       amdgpu_fence_process(&adev->sdma.instance[1].ring);
+                       break;
+               case 1:
+                       /* XXX compute */
+                       break;
+               case 2:
+                       /* XXX compute */
+                       break;
+               case 3:
+                       /* XXX page queue*/
+                       break;
+               }
+               break;
+       }
+       return 0;
+}
+
+static int sdma_v4_0_process_illegal_inst_irq(struct amdgpu_device *adev,
+                                             struct amdgpu_irq_src *source,
+                                             struct amdgpu_iv_entry *entry)
+{
+       DRM_ERROR("Illegal instruction in SDMA command stream\n");
+       schedule_work(&adev->reset_work);
+       return 0;
+}
+
+
+static void sdma_v4_0_update_medium_grain_clock_gating(
+               struct amdgpu_device *adev,
+               bool enable)
+{
+       uint32_t data, def;
+
+       if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
+               /* enable sdma0 clock gating */
+               def = data = RREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_CLK_CTRL));
+               data &= ~(SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK |
+                         SDMA0_CLK_CTRL__SOFT_OVERRIDE6_MASK |
+                         SDMA0_CLK_CTRL__SOFT_OVERRIDE5_MASK |
+                         SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
+                         SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
+                         SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
+                         SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
+                         SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK);
+               if (def != data)
+                       WREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_CLK_CTRL), data);
+
+               if (adev->asic_type == CHIP_VEGA10) {
+                       def = data = RREG32(SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_CLK_CTRL));
+                       data &= ~(SDMA1_CLK_CTRL__SOFT_OVERRIDE7_MASK |
+                                 SDMA1_CLK_CTRL__SOFT_OVERRIDE6_MASK |
+                                 SDMA1_CLK_CTRL__SOFT_OVERRIDE5_MASK |
+                                 SDMA1_CLK_CTRL__SOFT_OVERRIDE4_MASK |
+                                 SDMA1_CLK_CTRL__SOFT_OVERRIDE3_MASK |
+                                 SDMA1_CLK_CTRL__SOFT_OVERRIDE2_MASK |
+                                 SDMA1_CLK_CTRL__SOFT_OVERRIDE1_MASK |
+                                 SDMA1_CLK_CTRL__SOFT_OVERRIDE0_MASK);
+                       if(def != data)
+                               WREG32(SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_CLK_CTRL), data);
+               }
+       } else {
+               /* disable sdma0 clock gating */
+               def = data = RREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_CLK_CTRL));
+               data |= (SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK |
+                        SDMA0_CLK_CTRL__SOFT_OVERRIDE6_MASK |
+                        SDMA0_CLK_CTRL__SOFT_OVERRIDE5_MASK |
+                        SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
+                        SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
+                        SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
+                        SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
+                        SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK);
+
+               if (def != data)
+                       WREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_CLK_CTRL), data);
+
+               if (adev->asic_type == CHIP_VEGA10) {
+                       def = data = RREG32(SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_CLK_CTRL));
+                       data |= (SDMA1_CLK_CTRL__SOFT_OVERRIDE7_MASK |
+                                SDMA1_CLK_CTRL__SOFT_OVERRIDE6_MASK |
+                                SDMA1_CLK_CTRL__SOFT_OVERRIDE5_MASK |
+                                SDMA1_CLK_CTRL__SOFT_OVERRIDE4_MASK |
+                                SDMA1_CLK_CTRL__SOFT_OVERRIDE3_MASK |
+                                SDMA1_CLK_CTRL__SOFT_OVERRIDE2_MASK |
+                                SDMA1_CLK_CTRL__SOFT_OVERRIDE1_MASK |
+                                SDMA1_CLK_CTRL__SOFT_OVERRIDE0_MASK);
+                       if (def != data)
+                               WREG32(SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_CLK_CTRL), data);
+               }
+       }
+}
+
+
+static void sdma_v4_0_update_medium_grain_light_sleep(
+               struct amdgpu_device *adev,
+               bool enable)
+{
+       uint32_t data, def;
+
+       if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) {
+               /* 1-not override: enable sdma0 mem light sleep */
+               def = data = RREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_POWER_CNTL));
+               data |= SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
+               if (def != data)
+                       WREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_POWER_CNTL), data);
+
+               /* 1-not override: enable sdma1 mem light sleep */
+               if (adev->asic_type == CHIP_VEGA10) {
+                        def = data = RREG32(SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_POWER_CNTL));
+                        data |= SDMA1_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
+                        if (def != data)
+                                WREG32(SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_POWER_CNTL), data);
+               }
+       } else {
+               /* 0-override:disable sdma0 mem light sleep */
+               def = data = RREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_POWER_CNTL));
+               data &= ~SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
+               if (def != data)
+                      WREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_POWER_CNTL), data);
+
+               /* 0-override:disable sdma1 mem light sleep */
+               if (adev->asic_type == CHIP_VEGA10) {
+                       def = data = RREG32(SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_POWER_CNTL));
+                       data &= ~SDMA1_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
+                       if (def != data)
+                               WREG32(SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_POWER_CNTL), data);
+               }
+       }
+}
+
+static int sdma_v4_0_set_clockgating_state(void *handle,
+                                         enum amd_clockgating_state state)
+{
+       struct amdgpu_device *adev = (struct amdgpu_device *)handle;
+
+       switch (adev->asic_type) {
+       case CHIP_VEGA10:
+               sdma_v4_0_update_medium_grain_clock_gating(adev,
+                               state == AMD_CG_STATE_GATE ? true : false);
+               sdma_v4_0_update_medium_grain_light_sleep(adev,
+                               state == AMD_CG_STATE_GATE ? true : false);
+               break;
+       default:
+               break;
+       }
+       return 0;
+}
+
+static int sdma_v4_0_set_powergating_state(void *handle,
+                                         enum amd_powergating_state state)
+{
+       return 0;
+}
+
+const struct amd_ip_funcs sdma_v4_0_ip_funcs = {
+       .name = "sdma_v4_0",
+       .early_init = sdma_v4_0_early_init,
+       .late_init = NULL,
+       .sw_init = sdma_v4_0_sw_init,
+       .sw_fini = sdma_v4_0_sw_fini,
+       .hw_init = sdma_v4_0_hw_init,
+       .hw_fini = sdma_v4_0_hw_fini,
+       .suspend = sdma_v4_0_suspend,
+       .resume = sdma_v4_0_resume,
+       .is_idle = sdma_v4_0_is_idle,
+       .wait_for_idle = sdma_v4_0_wait_for_idle,
+       .soft_reset = sdma_v4_0_soft_reset,
+       .set_clockgating_state = sdma_v4_0_set_clockgating_state,
+       .set_powergating_state = sdma_v4_0_set_powergating_state,
+};
+
+static const struct amdgpu_ring_funcs sdma_v4_0_ring_funcs = {
+       .type = AMDGPU_RING_TYPE_SDMA,
+       .align_mask = 0xf,
+       .nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP),
+       .support_64bit_ptrs = true,
+       .get_rptr = sdma_v4_0_ring_get_rptr,
+       .get_wptr = sdma_v4_0_ring_get_wptr,
+       .set_wptr = sdma_v4_0_ring_set_wptr,
+       .emit_frame_size =
+               6 + /* sdma_v4_0_ring_emit_hdp_flush */
+               3 + /* sdma_v4_0_ring_emit_hdp_invalidate */
+               6 + /* sdma_v4_0_ring_emit_pipeline_sync */
+               36 + /* sdma_v4_0_ring_emit_vm_flush */
+               10 + 10 + 10, /* sdma_v4_0_ring_emit_fence x3 for user fence, vm fence */
+       .emit_ib_size = 7 + 6, /* sdma_v4_0_ring_emit_ib */
+       .emit_ib = sdma_v4_0_ring_emit_ib,
+       .emit_fence = sdma_v4_0_ring_emit_fence,
+       .emit_pipeline_sync = sdma_v4_0_ring_emit_pipeline_sync,
+       .emit_vm_flush = sdma_v4_0_ring_emit_vm_flush,
+       .emit_hdp_flush = sdma_v4_0_ring_emit_hdp_flush,
+       .emit_hdp_invalidate = sdma_v4_0_ring_emit_hdp_invalidate,
+       .test_ring = sdma_v4_0_ring_test_ring,
+       .test_ib = sdma_v4_0_ring_test_ib,
+       .insert_nop = sdma_v4_0_ring_insert_nop,
+       .pad_ib = sdma_v4_0_ring_pad_ib,
+};
+
+static void sdma_v4_0_set_ring_funcs(struct amdgpu_device *adev)
+{
+       int i;
+
+       for (i = 0; i < adev->sdma.num_instances; i++)
+               adev->sdma.instance[i].ring.funcs = &sdma_v4_0_ring_funcs;
+}
+
+static const struct amdgpu_irq_src_funcs sdma_v4_0_trap_irq_funcs = {
+       .set = sdma_v4_0_set_trap_irq_state,
+       .process = sdma_v4_0_process_trap_irq,
+};
+
+static const struct amdgpu_irq_src_funcs sdma_v4_0_illegal_inst_irq_funcs = {
+       .process = sdma_v4_0_process_illegal_inst_irq,
+};
+
+static void sdma_v4_0_set_irq_funcs(struct amdgpu_device *adev)
+{
+       adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
+       adev->sdma.trap_irq.funcs = &sdma_v4_0_trap_irq_funcs;
+       adev->sdma.illegal_inst_irq.funcs = &sdma_v4_0_illegal_inst_irq_funcs;
+}
+
+/**
+ * sdma_v4_0_emit_copy_buffer - copy buffer using the sDMA engine
+ *
+ * @ring: amdgpu_ring structure holding ring information
+ * @src_offset: src GPU address
+ * @dst_offset: dst GPU address
+ * @byte_count: number of bytes to xfer
+ *
+ * Copy GPU buffers using the DMA engine (VEGA10).
+ * Used by the amdgpu ttm implementation to move pages if
+ * registered as the asic copy callback.
+ */
+static void sdma_v4_0_emit_copy_buffer(struct amdgpu_ib *ib,
+                                      uint64_t src_offset,
+                                      uint64_t dst_offset,
+                                      uint32_t byte_count)
+{
+       ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
+               SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
+       ib->ptr[ib->length_dw++] = byte_count - 1;
+       ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
+       ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
+       ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
+       ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
+       ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
+}
+
+/**
+ * sdma_v4_0_emit_fill_buffer - fill buffer using the sDMA engine
+ *
+ * @ring: amdgpu_ring structure holding ring information
+ * @src_data: value to write to buffer
+ * @dst_offset: dst GPU address
+ * @byte_count: number of bytes to xfer
+ *
+ * Fill GPU buffers using the DMA engine (VEGA10).
+ */
+static void sdma_v4_0_emit_fill_buffer(struct amdgpu_ib *ib,
+                                      uint32_t src_data,
+                                      uint64_t dst_offset,
+                                      uint32_t byte_count)
+{
+       ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL);
+       ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
+       ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
+       ib->ptr[ib->length_dw++] = src_data;
+       ib->ptr[ib->length_dw++] = byte_count - 1;
+}
+
+static const struct amdgpu_buffer_funcs sdma_v4_0_buffer_funcs = {
+       .copy_max_bytes = 0x400000,
+       .copy_num_dw = 7,
+       .emit_copy_buffer = sdma_v4_0_emit_copy_buffer,
+
+       .fill_max_bytes = 0x400000,
+       .fill_num_dw = 5,
+       .emit_fill_buffer = sdma_v4_0_emit_fill_buffer,
+};
+
+static void sdma_v4_0_set_buffer_funcs(struct amdgpu_device *adev)
+{
+       if (adev->mman.buffer_funcs == NULL) {
+               adev->mman.buffer_funcs = &sdma_v4_0_buffer_funcs;
+               adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
+       }
+}
+
+static const struct amdgpu_vm_pte_funcs sdma_v4_0_vm_pte_funcs = {
+       .copy_pte = sdma_v4_0_vm_copy_pte,
+       .write_pte = sdma_v4_0_vm_write_pte,
+       .set_pte_pde = sdma_v4_0_vm_set_pte_pde,
+};
+
+static void sdma_v4_0_set_vm_pte_funcs(struct amdgpu_device *adev)
+{
+       unsigned i;
+
+       if (adev->vm_manager.vm_pte_funcs == NULL) {
+               adev->vm_manager.vm_pte_funcs = &sdma_v4_0_vm_pte_funcs;
+               for (i = 0; i < adev->sdma.num_instances; i++)
+                       adev->vm_manager.vm_pte_rings[i] =
+                               &adev->sdma.instance[i].ring;
+
+               adev->vm_manager.vm_pte_num_rings = adev->sdma.num_instances;
+       }
+}
+
+const struct amdgpu_ip_block_version sdma_v4_0_ip_block =
+{
+       .type = AMD_IP_BLOCK_TYPE_SDMA,
+       .major = 4,
+       .minor = 0,
+       .rev = 0,
+       .funcs = &sdma_v4_0_ip_funcs,
+};
diff --git a/drivers/gpu/drm/amd/amdgpu/sdma_v4_0.h b/drivers/gpu/drm/amd/amdgpu/sdma_v4_0.h
new file mode 100644 (file)
index 0000000..5c5a747
--- /dev/null
@@ -0,0 +1,30 @@
+/*
+ * Copyright 2016 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.
+ *
+ */
+
+#ifndef __SDMA_V4_0_H__
+#define __SDMA_V4_0_H__
+
+extern const struct amd_ip_funcs sdma_v4_0_ip_funcs;
+extern const struct amdgpu_ip_block_version sdma_v4_0_ip_block;
+
+#endif