drivers/gpu/drm/amd/amdgpu/sdma_v4_0.c

   1 /*
   2  * Copyright 2016 Advanced Micro Devices, Inc.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE SOFTWARE.
  21  *
  22  */
  23
  24 #include <linux/firmware.h>
  25 #include <drm/drmP.h>
  26 #include "amdgpu.h"
  27 #include "amdgpu_ucode.h"
  28 #include "amdgpu_trace.h"
  29
  30 #include "sdma0/sdma0_4_2_offset.h"
  31 #include "sdma0/sdma0_4_2_sh_mask.h"
  32 #include "sdma1/sdma1_4_2_offset.h"
  33 #include "sdma1/sdma1_4_2_sh_mask.h"
  34 #include "hdp/hdp_4_0_offset.h"
  35 #include "sdma0/sdma0_4_1_default.h"
  36
  37 #include "soc15_common.h"
  38 #include "soc15.h"
  39 #include "vega10_sdma_pkt_open.h"
  40
  41 #include "ivsrcid/sdma0/irqsrcs_sdma0_4_0.h"
  42 #include "ivsrcid/sdma1/irqsrcs_sdma1_4_0.h"
  43
  44 MODULE_FIRMWARE("amdgpu/vega10_sdma.bin");
  45 MODULE_FIRMWARE("amdgpu/vega10_sdma1.bin");
  46 MODULE_FIRMWARE("amdgpu/vega12_sdma.bin");
  47 MODULE_FIRMWARE("amdgpu/vega12_sdma1.bin");
  48 MODULE_FIRMWARE("amdgpu/vega20_sdma.bin");
  49 MODULE_FIRMWARE("amdgpu/vega20_sdma1.bin");
  50 MODULE_FIRMWARE("amdgpu/raven_sdma.bin");
  51 MODULE_FIRMWARE("amdgpu/picasso_sdma.bin");
  52 MODULE_FIRMWARE("amdgpu/raven2_sdma.bin");
  53
  54 #define SDMA0_POWER_CNTL__ON_OFF_CONDITION_HOLD_TIME_MASK  0x000000F8L
  55 #define SDMA0_POWER_CNTL__ON_OFF_STATUS_DURATION_TIME_MASK 0xFC000000L
  56
  57 #define WREG32_SDMA(instance, offset, value) \
  58         WREG32(sdma_v4_0_get_reg_offset(adev, (instance), (offset)), value)
  59 #define RREG32_SDMA(instance, offset) \
  60         RREG32(sdma_v4_0_get_reg_offset(adev, (instance), (offset)))
  61
  62 static void sdma_v4_0_set_ring_funcs(struct amdgpu_device *adev);
  63 static void sdma_v4_0_set_buffer_funcs(struct amdgpu_device *adev);
  64 static void sdma_v4_0_set_vm_pte_funcs(struct amdgpu_device *adev);
  65 static void sdma_v4_0_set_irq_funcs(struct amdgpu_device *adev);
  66
  67 static const struct soc15_reg_golden golden_settings_sdma_4[] = {
  68         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_CHICKEN_BITS, 0xfe931f07, 0x02831d07),
  69         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_CLK_CTRL, 0xff000ff0, 0x3f000100),
  70         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GFX_IB_CNTL, 0x800f0100, 0x00000100),
  71         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
  72         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_PAGE_IB_CNTL, 0x800f0100, 0x00000100),
  73         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_PAGE_RB_WPTR_POLL_CNTL, 0x0000fff0, 0x00403000),
  74         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_POWER_CNTL, 0x003ff006, 0x0003c000),
  75         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC0_IB_CNTL, 0x800f0100, 0x00000100),
  76         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC0_RB_WPTR_POLL_CNTL, 0x0000fff0, 0x00403000),
  77         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC1_IB_CNTL, 0x800f0100, 0x00000100),
  78         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC1_RB_WPTR_POLL_CNTL, 0x0000fff0, 0x00403000),
  79         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_UTCL1_PAGE, 0x000003ff, 0x000003c0),
  80         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_UTCL1_WATERMK, 0xfc000000, 0x00000000),
  81         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_CLK_CTRL, 0xffffffff, 0x3f000100),
  82         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GFX_IB_CNTL, 0x800f0100, 0x00000100),
  83         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GFX_RB_WPTR_POLL_CNTL, 0x0000fff0, 0x00403000),
  84         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_PAGE_IB_CNTL, 0x800f0100, 0x00000100),
  85         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_PAGE_RB_WPTR_POLL_CNTL, 0x0000fff0, 0x00403000),
  86         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_POWER_CNTL, 0x003ff000, 0x0003c000),
  87         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC0_IB_CNTL, 0x800f0100, 0x00000100),
  88         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC0_RB_WPTR_POLL_CNTL, 0x0000fff0, 0x00403000),
  89         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC1_IB_CNTL, 0x800f0100, 0x00000100),
  90         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC1_RB_WPTR_POLL_CNTL, 0x0000fff0, 0x00403000),
  91         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_UTCL1_PAGE, 0x000003ff, 0x000003c0),
  92         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_UTCL1_WATERMK, 0xfc000000, 0x00000000)
  93 };
  94
  95 static const struct soc15_reg_golden golden_settings_sdma_vg10[] = {
  96         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG, 0x0018773f, 0x00104002),
  97         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG_READ, 0x0018773f, 0x00104002),
  98         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_CHICKEN_BITS, 0xfe931f07, 0x02831d07),
  99         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GB_ADDR_CONFIG, 0x0018773f, 0x00104002),
 100         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GB_ADDR_CONFIG_READ, 0x0018773f, 0x00104002)
 101 };
 102
 103 static const struct soc15_reg_golden golden_settings_sdma_vg12[] = {
 104         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG, 0x0018773f, 0x00104001),
 105         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG_READ, 0x0018773f, 0x00104001),
 106         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_CHICKEN_BITS, 0xfe931f07, 0x02831d07),
 107         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GB_ADDR_CONFIG, 0x0018773f, 0x00104001),
 108         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GB_ADDR_CONFIG_READ, 0x0018773f, 0x00104001)
 109 };
 110
 111 static const struct soc15_reg_golden golden_settings_sdma_4_1[] = {
 112         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_CHICKEN_BITS, 0xfe931f07, 0x02831d07),
 113         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_CLK_CTRL, 0xffffffff, 0x3f000100),
 114         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GFX_IB_CNTL, 0x800f0111, 0x00000100),
 115         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
 116         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_POWER_CNTL, 0xfc3fffff, 0x40000051),
 117         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC0_IB_CNTL, 0x800f0111, 0x00000100),
 118         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
 119         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC1_IB_CNTL, 0x800f0111, 0x00000100),
 120         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
 121         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_UTCL1_PAGE, 0x000003ff, 0x000003c0),
 122         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_UTCL1_WATERMK, 0xfc000000, 0x00000000)
 123 };
 124
 125 static const struct soc15_reg_golden golden_settings_sdma0_4_2_init[] = {
 126         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff0, 0x00403000),
 127 };
 128
 129 static const struct soc15_reg_golden golden_settings_sdma0_4_2[] =
 130 {
 131         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_CHICKEN_BITS, 0xfe931f07, 0x02831f07),
 132         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_CLK_CTRL, 0xffffffff, 0x3f000100),
 133         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG, 0x0000773f, 0x00004002),
 134         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG_READ, 0x0000773f, 0x00004002),
 135         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GFX_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001),
 136         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
 137         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_PAGE_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001),
 138         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_PAGE_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
 139         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RD_BURST_CNTL, 0x0000000f, 0x00000003),
 140         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC0_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001),
 141         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff0, 0x00403000),
 142         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC1_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001),
 143         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
 144         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC2_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001),
 145         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC2_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
 146         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC3_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001),
 147         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
 148         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC4_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001),
 149         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC4_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
 150         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC5_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001),
 151         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC5_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
 152         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC6_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001),
 153         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC6_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
 154         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC7_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001),
 155         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC7_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
 156         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_UTCL1_PAGE, 0x000003ff, 0x000003c0),
 157         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_UTCL1_WATERMK, 0xFE000000, 0x00000000),
 158 };
 159
 160 static const struct soc15_reg_golden golden_settings_sdma1_4_2[] = {
 161         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_CHICKEN_BITS, 0xfe931f07, 0x02831f07),
 162         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_CLK_CTRL, 0xffffffff, 0x3f000100),
 163         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GB_ADDR_CONFIG, 0x0000773f, 0x00004002),
 164         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GB_ADDR_CONFIG_READ, 0x0000773f, 0x00004002),
 165         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GFX_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001),
 166         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
 167         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_PAGE_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001),
 168         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_PAGE_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
 169         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RD_BURST_CNTL, 0x0000000f, 0x00000003),
 170         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC0_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001),
 171         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff0, 0x00403000),
 172         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC1_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001),
 173         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
 174         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC2_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001),
 175         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC2_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
 176         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC3_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001),
 177         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
 178         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC4_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001),
 179         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC4_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
 180         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC5_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001),
 181         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC5_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
 182         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC6_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001),
 183         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC6_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
 184         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC7_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001),
 185         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC7_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
 186         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_UTCL1_PAGE, 0x000003ff, 0x000003c0),
 187         SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_UTCL1_WATERMK, 0xFE000000, 0x00000000),
 188 };
 189
 190 static const struct soc15_reg_golden golden_settings_sdma_rv1[] =
 191 {
 192         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG, 0x0018773f, 0x00000002),
 193         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG_READ, 0x0018773f, 0x00000002)
 194 };
 195
 196 static const struct soc15_reg_golden golden_settings_sdma_rv2[] =
 197 {
 198         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG, 0x0018773f, 0x00003001),
 199         SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG_READ, 0x0018773f, 0x00003001)
 200 };
 201
 202 static u32 sdma_v4_0_get_reg_offset(struct amdgpu_device *adev,
 203                 u32 instance, u32 offset)
 204 {
 205         return ( 0 == instance ? (adev->reg_offset[SDMA0_HWIP][0][0] + offset) :
 206                         (adev->reg_offset[SDMA1_HWIP][0][0] + offset));
 207 }
 208
 209 static void sdma_v4_0_init_golden_registers(struct amdgpu_device *adev)
 210 {
 211         switch (adev->asic_type) {
 212         case CHIP_VEGA10:
 213                 soc15_program_register_sequence(adev,
 214                                                  golden_settings_sdma_4,
 215                                                  ARRAY_SIZE(golden_settings_sdma_4));
 216                 soc15_program_register_sequence(adev,
 217                                                  golden_settings_sdma_vg10,
 218                                                  ARRAY_SIZE(golden_settings_sdma_vg10));
 219                 break;
 220         case CHIP_VEGA12:
 221                 soc15_program_register_sequence(adev,
 222                                                 golden_settings_sdma_4,
 223                                                 ARRAY_SIZE(golden_settings_sdma_4));
 224                 soc15_program_register_sequence(adev,
 225                                                 golden_settings_sdma_vg12,
 226                                                 ARRAY_SIZE(golden_settings_sdma_vg12));
 227                 break;
 228         case CHIP_VEGA20:
 229                 soc15_program_register_sequence(adev,
 230                                                 golden_settings_sdma0_4_2_init,
 231                                                 ARRAY_SIZE(golden_settings_sdma0_4_2_init));
 232                 soc15_program_register_sequence(adev,
 233                                                 golden_settings_sdma0_4_2,
 234                                                 ARRAY_SIZE(golden_settings_sdma0_4_2));
 235                 soc15_program_register_sequence(adev,
 236                                                 golden_settings_sdma1_4_2,
 237                                                 ARRAY_SIZE(golden_settings_sdma1_4_2));
 238                 break;
 239         case CHIP_RAVEN:
 240                 soc15_program_register_sequence(adev,
 241                                                 golden_settings_sdma_4_1,
 242                                                 ARRAY_SIZE(golden_settings_sdma_4_1));
 243                 if (adev->rev_id >= 8)
 244                         soc15_program_register_sequence(adev,
 245                                                         golden_settings_sdma_rv2,
 246                                                         ARRAY_SIZE(golden_settings_sdma_rv2));
 247                 else
 248                         soc15_program_register_sequence(adev,
 249                                                         golden_settings_sdma_rv1,
 250                                                         ARRAY_SIZE(golden_settings_sdma_rv1));
 251                 break;
 252         default:
 253                 break;
 254         }
 255 }
 256
 257 /**
 258  * sdma_v4_0_init_microcode - load ucode images from disk
 259  *
 260  * @adev: amdgpu_device pointer
 261  *
 262  * Use the firmware interface to load the ucode images into
 263  * the driver (not loaded into hw).
 264  * Returns 0 on success, error on failure.
 265  */
 266
 267 // emulation only, won't work on real chip
 268 // vega10 real chip need to use PSP to load firmware
 269 static int sdma_v4_0_init_microcode(struct amdgpu_device *adev)
 270 {
 271         const char *chip_name;
 272         char fw_name[30];
 273         int err = 0, i;
 274         struct amdgpu_firmware_info *info = NULL;
 275         const struct common_firmware_header *header = NULL;
 276         const struct sdma_firmware_header_v1_0 *hdr;
 277
 278         DRM_DEBUG("\n");
 279
 280         switch (adev->asic_type) {
 281         case CHIP_VEGA10:
 282                 chip_name = "vega10";
 283                 break;
 284         case CHIP_VEGA12:
 285                 chip_name = "vega12";
 286                 break;
 287         case CHIP_VEGA20:
 288                 chip_name = "vega20";
 289                 break;
 290         case CHIP_RAVEN:
 291                 if (adev->rev_id >= 8)
 292                         chip_name = "raven2";
 293                 else if (adev->pdev->device == 0x15d8)
 294                         chip_name = "picasso";
 295                 else
 296                         chip_name = "raven";
 297                 break;
 298         default:
 299                 BUG();
 300         }
 301
 302         for (i = 0; i < adev->sdma.num_instances; i++) {
 303                 if (i == 0)
 304                         snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma.bin", chip_name);
 305                 else
 306                         snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma1.bin", chip_name);
 307                 err = request_firmware(&adev->sdma.instance[i].fw, fw_name, adev->dev);
 308                 if (err)
 309                         goto out;
 310                 err = amdgpu_ucode_validate(adev->sdma.instance[i].fw);
 311                 if (err)
 312                         goto out;
 313                 hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
 314                 adev->sdma.instance[i].fw_version = le32_to_cpu(hdr->header.ucode_version);
 315                 adev->sdma.instance[i].feature_version = le32_to_cpu(hdr->ucode_feature_version);
 316                 if (adev->sdma.instance[i].feature_version >= 20)
 317                         adev->sdma.instance[i].burst_nop = true;
 318                 DRM_DEBUG("psp_load == '%s'\n",
 319                                 adev->firmware.load_type == AMDGPU_FW_LOAD_PSP ? "true" : "false");
 320
 321                 if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
 322                         info = &adev->firmware.ucode[AMDGPU_UCODE_ID_SDMA0 + i];
 323                         info->ucode_id = AMDGPU_UCODE_ID_SDMA0 + i;
 324                         info->fw = adev->sdma.instance[i].fw;
 325                         header = (const struct common_firmware_header *)info->fw->data;
 326                         adev->firmware.fw_size +=
 327                                 ALIGN(le32_to_cpu(header->ucode_size_bytes), PAGE_SIZE);
 328                 }
 329         }
 330 out:
 331         if (err) {
 332                 DRM_ERROR("sdma_v4_0: Failed to load firmware \"%s\"\n", fw_name);
 333                 for (i = 0; i < adev->sdma.num_instances; i++) {
 334                         release_firmware(adev->sdma.instance[i].fw);
 335                         adev->sdma.instance[i].fw = NULL;
 336                 }
 337         }
 338         return err;
 339 }
 340
 341 /**
 342  * sdma_v4_0_ring_get_rptr - get the current read pointer
 343  *
 344  * @ring: amdgpu ring pointer
 345  *
 346  * Get the current rptr from the hardware (VEGA10+).
 347  */
 348 static uint64_t sdma_v4_0_ring_get_rptr(struct amdgpu_ring *ring)
 349 {
 350         u64 *rptr;
 351
 352         /* XXX check if swapping is necessary on BE */
 353         rptr = ((u64 *)&ring->adev->wb.wb[ring->rptr_offs]);
 354
 355         DRM_DEBUG("rptr before shift == 0x%016llx\n", *rptr);
 356         return ((*rptr) >> 2);
 357 }
 358
 359 /**
 360  * sdma_v4_0_ring_get_wptr - get the current write pointer
 361  *
 362  * @ring: amdgpu ring pointer
 363  *
 364  * Get the current wptr from the hardware (VEGA10+).
 365  */
 366 static uint64_t sdma_v4_0_ring_get_wptr(struct amdgpu_ring *ring)
 367 {
 368         struct amdgpu_device *adev = ring->adev;
 369         u64 wptr;
 370
 371         if (ring->use_doorbell) {
 372                 /* XXX check if swapping is necessary on BE */
 373                 wptr = READ_ONCE(*((u64 *)&adev->wb.wb[ring->wptr_offs]));
 374                 DRM_DEBUG("wptr/doorbell before shift == 0x%016llx\n", wptr);
 375         } else {
 376                 wptr = RREG32_SDMA(ring->me, mmSDMA0_GFX_RB_WPTR_HI);
 377                 wptr = wptr << 32;
 378                 wptr |= RREG32_SDMA(ring->me, mmSDMA0_GFX_RB_WPTR);
 379                 DRM_DEBUG("wptr before shift [%i] wptr == 0x%016llx\n",
 380                                 ring->me, wptr);
 381         }
 382
 383         return wptr >> 2;
 384 }
 385
 386 /**
 387  * sdma_v4_0_ring_set_wptr - commit the write pointer
 388  *
 389  * @ring: amdgpu ring pointer
 390  *
 391  * Write the wptr back to the hardware (VEGA10+).
 392  */
 393 static void sdma_v4_0_ring_set_wptr(struct amdgpu_ring *ring)
 394 {
 395         struct amdgpu_device *adev = ring->adev;
 396
 397         DRM_DEBUG("Setting write pointer\n");
 398         if (ring->use_doorbell) {
 399                 u64 *wb = (u64 *)&adev->wb.wb[ring->wptr_offs];
 400
 401                 DRM_DEBUG("Using doorbell -- "
 402                                 "wptr_offs == 0x%08x "
 403                                 "lower_32_bits(ring->wptr) << 2 == 0x%08x "
 404                                 "upper_32_bits(ring->wptr) << 2 == 0x%08x\n",
 405                                 ring->wptr_offs,
 406                                 lower_32_bits(ring->wptr << 2),
 407                                 upper_32_bits(ring->wptr << 2));
 408                 /* XXX check if swapping is necessary on BE */
 409                 WRITE_ONCE(*wb, (ring->wptr << 2));
 410                 DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n",
 411                                 ring->doorbell_index, ring->wptr << 2);
 412                 WDOORBELL64(ring->doorbell_index, ring->wptr << 2);
 413         } else {
 414                 DRM_DEBUG("Not using doorbell -- "
 415                                 "mmSDMA%i_GFX_RB_WPTR == 0x%08x "
 416                                 "mmSDMA%i_GFX_RB_WPTR_HI == 0x%08x\n",
 417                                 ring->me,
 418                                 lower_32_bits(ring->wptr << 2),
 419                                 ring->me,
 420                                 upper_32_bits(ring->wptr << 2));
 421                 WREG32_SDMA(ring->me, mmSDMA0_GFX_RB_WPTR,
 422                             lower_32_bits(ring->wptr << 2));
 423                 WREG32_SDMA(ring->me, mmSDMA0_GFX_RB_WPTR_HI,
 424                             upper_32_bits(ring->wptr << 2));
 425         }
 426 }
 427
 428 /**
 429  * sdma_v4_0_page_ring_get_wptr - get the current write pointer
 430  *
 431  * @ring: amdgpu ring pointer
 432  *
 433  * Get the current wptr from the hardware (VEGA10+).
 434  */
 435 static uint64_t sdma_v4_0_page_ring_get_wptr(struct amdgpu_ring *ring)
 436 {
 437         struct amdgpu_device *adev = ring->adev;
 438         u64 wptr;
 439
 440         if (ring->use_doorbell) {
 441                 /* XXX check if swapping is necessary on BE */
 442                 wptr = READ_ONCE(*((u64 *)&adev->wb.wb[ring->wptr_offs]));
 443         } else {
 444                 wptr = RREG32_SDMA(ring->me, mmSDMA0_PAGE_RB_WPTR_HI);
 445                 wptr = wptr << 32;
 446                 wptr |= RREG32_SDMA(ring->me, mmSDMA0_PAGE_RB_WPTR);
 447         }
 448
 449         return wptr >> 2;
 450 }
 451
 452 /**
 453  * sdma_v4_0_ring_set_wptr - commit the write pointer
 454  *
 455  * @ring: amdgpu ring pointer
 456  *
 457  * Write the wptr back to the hardware (VEGA10+).
 458  */
 459 static void sdma_v4_0_page_ring_set_wptr(struct amdgpu_ring *ring)
 460 {
 461         struct amdgpu_device *adev = ring->adev;
 462
 463         if (ring->use_doorbell) {
 464                 u64 *wb = (u64 *)&adev->wb.wb[ring->wptr_offs];
 465
 466                 /* XXX check if swapping is necessary on BE */
 467                 WRITE_ONCE(*wb, (ring->wptr << 2));
 468                 WDOORBELL64(ring->doorbell_index, ring->wptr << 2);
 469         } else {
 470                 uint64_t wptr = ring->wptr << 2;
 471
 472                 WREG32_SDMA(ring->me, mmSDMA0_PAGE_RB_WPTR,
 473                             lower_32_bits(wptr));
 474                 WREG32_SDMA(ring->me, mmSDMA0_PAGE_RB_WPTR_HI,
 475                             upper_32_bits(wptr));
 476         }
 477 }
 478
 479 static void sdma_v4_0_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
 480 {
 481         struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
 482         int i;
 483
 484         for (i = 0; i < count; i++)
 485                 if (sdma && sdma->burst_nop && (i == 0))
 486                         amdgpu_ring_write(ring, ring->funcs->nop |
 487                                 SDMA_PKT_NOP_HEADER_COUNT(count - 1));
 488                 else
 489                         amdgpu_ring_write(ring, ring->funcs->nop);
 490 }
 491
 492 /**
 493  * sdma_v4_0_ring_emit_ib - Schedule an IB on the DMA engine
 494  *
 495  * @ring: amdgpu ring pointer
 496  * @ib: IB object to schedule
 497  *
 498  * Schedule an IB in the DMA ring (VEGA10).
 499  */
 500 static void sdma_v4_0_ring_emit_ib(struct amdgpu_ring *ring,
 501                                    struct amdgpu_job *job,
 502                                    struct amdgpu_ib *ib,
 503                                    bool ctx_switch)
 504 {
 505         unsigned vmid = AMDGPU_JOB_GET_VMID(job);
 506
 507         /* IB packet must end on a 8 DW boundary */
 508         sdma_v4_0_ring_insert_nop(ring, (10 - (lower_32_bits(ring->wptr) & 7)) % 8);
 509
 510         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) |
 511                           SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf));
 512         /* base must be 32 byte aligned */
 513         amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
 514         amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
 515         amdgpu_ring_write(ring, ib->length_dw);
 516         amdgpu_ring_write(ring, 0);
 517         amdgpu_ring_write(ring, 0);
 518
 519 }
 520
 521 static void sdma_v4_0_wait_reg_mem(struct amdgpu_ring *ring,
 522                                    int mem_space, int hdp,
 523                                    uint32_t addr0, uint32_t addr1,
 524                                    uint32_t ref, uint32_t mask,
 525                                    uint32_t inv)
 526 {
 527         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
 528                           SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(hdp) |
 529                           SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(mem_space) |
 530                           SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
 531         if (mem_space) {
 532                 /* memory */
 533                 amdgpu_ring_write(ring, addr0);
 534                 amdgpu_ring_write(ring, addr1);
 535         } else {
 536                 /* registers */
 537                 amdgpu_ring_write(ring, addr0 << 2);
 538                 amdgpu_ring_write(ring, addr1 << 2);
 539         }
 540         amdgpu_ring_write(ring, ref); /* reference */
 541         amdgpu_ring_write(ring, mask); /* mask */
 542         amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
 543                           SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(inv)); /* retry count, poll interval */
 544 }
 545
 546 /**
 547  * sdma_v4_0_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
 548  *
 549  * @ring: amdgpu ring pointer
 550  *
 551  * Emit an hdp flush packet on the requested DMA ring.
 552  */
 553 static void sdma_v4_0_ring_emit_hdp_flush(struct amdgpu_ring *ring)
 554 {
 555         struct amdgpu_device *adev = ring->adev;
 556         u32 ref_and_mask = 0;
 557         const struct nbio_hdp_flush_reg *nbio_hf_reg = adev->nbio_funcs->hdp_flush_reg;
 558
 559         if (ring->me == 0)
 560                 ref_and_mask = nbio_hf_reg->ref_and_mask_sdma0;
 561         else
 562                 ref_and_mask = nbio_hf_reg->ref_and_mask_sdma1;
 563
 564         sdma_v4_0_wait_reg_mem(ring, 0, 1,
 565                                adev->nbio_funcs->get_hdp_flush_done_offset(adev),
 566                                adev->nbio_funcs->get_hdp_flush_req_offset(adev),
 567                                ref_and_mask, ref_and_mask, 10);
 568 }
 569
 570 /**
 571  * sdma_v4_0_ring_emit_fence - emit a fence on the DMA ring
 572  *
 573  * @ring: amdgpu ring pointer
 574  * @fence: amdgpu fence object
 575  *
 576  * Add a DMA fence packet to the ring to write
 577  * the fence seq number and DMA trap packet to generate
 578  * an interrupt if needed (VEGA10).
 579  */
 580 static void sdma_v4_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
 581                                       unsigned flags)
 582 {
 583         bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
 584         /* write the fence */
 585         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE));
 586         /* zero in first two bits */
 587         BUG_ON(addr & 0x3);
 588         amdgpu_ring_write(ring, lower_32_bits(addr));
 589         amdgpu_ring_write(ring, upper_32_bits(addr));
 590         amdgpu_ring_write(ring, lower_32_bits(seq));
 591
 592         /* optionally write high bits as well */
 593         if (write64bit) {
 594                 addr += 4;
 595                 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE));
 596                 /* zero in first two bits */
 597                 BUG_ON(addr & 0x3);
 598                 amdgpu_ring_write(ring, lower_32_bits(addr));
 599                 amdgpu_ring_write(ring, upper_32_bits(addr));
 600                 amdgpu_ring_write(ring, upper_32_bits(seq));
 601         }
 602
 603         /* generate an interrupt */
 604         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP));
 605         amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(0));
 606 }
 607
 608
 609 /**
 610  * sdma_v4_0_gfx_stop - stop the gfx async dma engines
 611  *
 612  * @adev: amdgpu_device pointer
 613  *
 614  * Stop the gfx async dma ring buffers (VEGA10).
 615  */
 616 static void sdma_v4_0_gfx_stop(struct amdgpu_device *adev)
 617 {
 618         struct amdgpu_ring *sdma0 = &adev->sdma.instance[0].ring;
 619         struct amdgpu_ring *sdma1 = &adev->sdma.instance[1].ring;
 620         u32 rb_cntl, ib_cntl;
 621         int i;
 622
 623         if ((adev->mman.buffer_funcs_ring == sdma0) ||
 624             (adev->mman.buffer_funcs_ring == sdma1))
 625                         amdgpu_ttm_set_buffer_funcs_status(adev, false);
 626
 627         for (i = 0; i < adev->sdma.num_instances; i++) {
 628                 rb_cntl = RREG32_SDMA(i, mmSDMA0_GFX_RB_CNTL);
 629                 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0);
 630                 WREG32_SDMA(i, mmSDMA0_GFX_RB_CNTL, rb_cntl);
 631                 ib_cntl = RREG32_SDMA(i, mmSDMA0_GFX_IB_CNTL);
 632                 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0);
 633                 WREG32_SDMA(i, mmSDMA0_GFX_IB_CNTL, ib_cntl);
 634         }
 635
 636         sdma0->sched.ready = false;
 637         sdma1->sched.ready = false;
 638 }
 639
 640 /**
 641  * sdma_v4_0_rlc_stop - stop the compute async dma engines
 642  *
 643  * @adev: amdgpu_device pointer
 644  *
 645  * Stop the compute async dma queues (VEGA10).
 646  */
 647 static void sdma_v4_0_rlc_stop(struct amdgpu_device *adev)
 648 {
 649         /* XXX todo */
 650 }
 651
 652 /**
 653  * sdma_v4_0_page_stop - stop the page async dma engines
 654  *
 655  * @adev: amdgpu_device pointer
 656  *
 657  * Stop the page async dma ring buffers (VEGA10).
 658  */
 659 static void sdma_v4_0_page_stop(struct amdgpu_device *adev)
 660 {
 661         struct amdgpu_ring *sdma0 = &adev->sdma.instance[0].page;
 662         struct amdgpu_ring *sdma1 = &adev->sdma.instance[1].page;
 663         u32 rb_cntl, ib_cntl;
 664         int i;
 665
 666         if ((adev->mman.buffer_funcs_ring == sdma0) ||
 667             (adev->mman.buffer_funcs_ring == sdma1))
 668                 amdgpu_ttm_set_buffer_funcs_status(adev, false);
 669
 670         for (i = 0; i < adev->sdma.num_instances; i++) {
 671                 rb_cntl = RREG32_SDMA(i, mmSDMA0_PAGE_RB_CNTL);
 672                 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_PAGE_RB_CNTL,
 673                                         RB_ENABLE, 0);
 674                 WREG32_SDMA(i, mmSDMA0_PAGE_RB_CNTL, rb_cntl);
 675                 ib_cntl = RREG32_SDMA(i, mmSDMA0_PAGE_IB_CNTL);
 676                 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_PAGE_IB_CNTL,
 677                                         IB_ENABLE, 0);
 678                 WREG32_SDMA(i, mmSDMA0_PAGE_IB_CNTL, ib_cntl);
 679         }
 680
 681         sdma0->sched.ready = false;
 682         sdma1->sched.ready = false;
 683 }
 684
 685 /**
 686  * sdma_v_0_ctx_switch_enable - stop the async dma engines context switch
 687  *
 688  * @adev: amdgpu_device pointer
 689  * @enable: enable/disable the DMA MEs context switch.
 690  *
 691  * Halt or unhalt the async dma engines context switch (VEGA10).
 692  */
 693 static void sdma_v4_0_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
 694 {
 695         u32 f32_cntl, phase_quantum = 0;
 696         int i;
 697
 698         if (amdgpu_sdma_phase_quantum) {
 699                 unsigned value = amdgpu_sdma_phase_quantum;
 700                 unsigned unit = 0;
 701
 702                 while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
 703                                 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) {
 704                         value = (value + 1) >> 1;
 705                         unit++;
 706                 }
 707                 if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
 708                             SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) {
 709                         value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
 710                                  SDMA0_PHASE0_QUANTUM__VALUE__SHIFT);
 711                         unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
 712                                 SDMA0_PHASE0_QUANTUM__UNIT__SHIFT);
 713                         WARN_ONCE(1,
 714                         "clamping sdma_phase_quantum to %uK clock cycles\n",
 715                                   value << unit);
 716                 }
 717                 phase_quantum =
 718                         value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
 719                         unit  << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
 720         }
 721
 722         for (i = 0; i < adev->sdma.num_instances; i++) {
 723                 f32_cntl = RREG32_SDMA(i, mmSDMA0_CNTL);
 724                 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
 725                                 AUTO_CTXSW_ENABLE, enable ? 1 : 0);
 726                 if (enable && amdgpu_sdma_phase_quantum) {
 727                         WREG32_SDMA(i, mmSDMA0_PHASE0_QUANTUM, phase_quantum);
 728                         WREG32_SDMA(i, mmSDMA0_PHASE1_QUANTUM, phase_quantum);
 729                         WREG32_SDMA(i, mmSDMA0_PHASE2_QUANTUM, phase_quantum);
 730                 }
 731                 WREG32_SDMA(i, mmSDMA0_CNTL, f32_cntl);
 732         }
 733
 734 }
 735
 736 /**
 737  * sdma_v4_0_enable - stop the async dma engines
 738  *
 739  * @adev: amdgpu_device pointer
 740  * @enable: enable/disable the DMA MEs.
 741  *
 742  * Halt or unhalt the async dma engines (VEGA10).
 743  */
 744 static void sdma_v4_0_enable(struct amdgpu_device *adev, bool enable)
 745 {
 746         u32 f32_cntl;
 747         int i;
 748
 749         if (enable == false) {
 750                 sdma_v4_0_gfx_stop(adev);
 751                 sdma_v4_0_rlc_stop(adev);
 752                 if (adev->sdma.has_page_queue)
 753                         sdma_v4_0_page_stop(adev);
 754         }
 755
 756         for (i = 0; i < adev->sdma.num_instances; i++) {
 757                 f32_cntl = RREG32_SDMA(i, mmSDMA0_F32_CNTL);
 758                 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, enable ? 0 : 1);
 759                 WREG32_SDMA(i, mmSDMA0_F32_CNTL, f32_cntl);
 760         }
 761 }
 762
 763 /**
 764  * sdma_v4_0_rb_cntl - get parameters for rb_cntl
 765  */
 766 static uint32_t sdma_v4_0_rb_cntl(struct amdgpu_ring *ring, uint32_t rb_cntl)
 767 {
 768         /* Set ring buffer size in dwords */
 769         uint32_t rb_bufsz = order_base_2(ring->ring_size / 4);
 770
 771         rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz);
 772 #ifdef __BIG_ENDIAN
 773         rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1);
 774         rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL,
 775                                 RPTR_WRITEBACK_SWAP_ENABLE, 1);
 776 #endif
 777         return rb_cntl;
 778 }
 779
 780 /**
 781  * sdma_v4_0_gfx_resume - setup and start the async dma engines
 782  *
 783  * @adev: amdgpu_device pointer
 784  * @i: instance to resume
 785  *
 786  * Set up the gfx DMA ring buffers and enable them (VEGA10).
 787  * Returns 0 for success, error for failure.
 788  */
 789 static void sdma_v4_0_gfx_resume(struct amdgpu_device *adev, unsigned int i)
 790 {
 791         struct amdgpu_ring *ring = &adev->sdma.instance[i].ring;
 792         u32 rb_cntl, ib_cntl, wptr_poll_cntl;
 793         u32 wb_offset;
 794         u32 doorbell;
 795         u32 doorbell_offset;
 796         u64 wptr_gpu_addr;
 797
 798         wb_offset = (ring->rptr_offs * 4);
 799
 800         rb_cntl = RREG32_SDMA(i, mmSDMA0_GFX_RB_CNTL);
 801         rb_cntl = sdma_v4_0_rb_cntl(ring, rb_cntl);
 802         WREG32_SDMA(i, mmSDMA0_GFX_RB_CNTL, rb_cntl);
 803
 804         /* Initialize the ring buffer's read and write pointers */
 805         WREG32_SDMA(i, mmSDMA0_GFX_RB_RPTR, 0);
 806         WREG32_SDMA(i, mmSDMA0_GFX_RB_RPTR_HI, 0);
 807         WREG32_SDMA(i, mmSDMA0_GFX_RB_WPTR, 0);
 808         WREG32_SDMA(i, mmSDMA0_GFX_RB_WPTR_HI, 0);
 809
 810         /* set the wb address whether it's enabled or not */
 811         WREG32_SDMA(i, mmSDMA0_GFX_RB_RPTR_ADDR_HI,
 812                upper_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFF);
 813         WREG32_SDMA(i, mmSDMA0_GFX_RB_RPTR_ADDR_LO,
 814                lower_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC);
 815
 816         rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL,
 817                                 RPTR_WRITEBACK_ENABLE, 1);
 818
 819         WREG32_SDMA(i, mmSDMA0_GFX_RB_BASE, ring->gpu_addr >> 8);
 820         WREG32_SDMA(i, mmSDMA0_GFX_RB_BASE_HI, ring->gpu_addr >> 40);
 821
 822         ring->wptr = 0;
 823
 824         /* before programing wptr to a less value, need set minor_ptr_update first */
 825         WREG32_SDMA(i, mmSDMA0_GFX_MINOR_PTR_UPDATE, 1);
 826
 827         doorbell = RREG32_SDMA(i, mmSDMA0_GFX_DOORBELL);
 828         doorbell_offset = RREG32_SDMA(i, mmSDMA0_GFX_DOORBELL_OFFSET);
 829
 830         doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE,
 831                                  ring->use_doorbell);
 832         doorbell_offset = REG_SET_FIELD(doorbell_offset,
 833                                         SDMA0_GFX_DOORBELL_OFFSET,
 834                                         OFFSET, ring->doorbell_index);
 835         WREG32_SDMA(i, mmSDMA0_GFX_DOORBELL, doorbell);
 836         WREG32_SDMA(i, mmSDMA0_GFX_DOORBELL_OFFSET, doorbell_offset);
 837         adev->nbio_funcs->sdma_doorbell_range(adev, i, ring->use_doorbell,
 838                                               ring->doorbell_index);
 839
 840         sdma_v4_0_ring_set_wptr(ring);
 841
 842         /* set minor_ptr_update to 0 after wptr programed */
 843         WREG32_SDMA(i, mmSDMA0_GFX_MINOR_PTR_UPDATE, 0);
 844
 845         /* setup the wptr shadow polling */
 846         wptr_gpu_addr = adev->wb.gpu_addr + (ring->wptr_offs * 4);
 847         WREG32_SDMA(i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_LO,
 848                     lower_32_bits(wptr_gpu_addr));
 849         WREG32_SDMA(i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_HI,
 850                     upper_32_bits(wptr_gpu_addr));
 851         wptr_poll_cntl = RREG32_SDMA(i, mmSDMA0_GFX_RB_WPTR_POLL_CNTL);
 852         wptr_poll_cntl = REG_SET_FIELD(wptr_poll_cntl,
 853                                        SDMA0_GFX_RB_WPTR_POLL_CNTL,
 854                                        F32_POLL_ENABLE, amdgpu_sriov_vf(adev));
 855         WREG32_SDMA(i, mmSDMA0_GFX_RB_WPTR_POLL_CNTL, wptr_poll_cntl);
 856
 857         /* enable DMA RB */
 858         rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1);
 859         WREG32_SDMA(i, mmSDMA0_GFX_RB_CNTL, rb_cntl);
 860
 861         ib_cntl = RREG32_SDMA(i, mmSDMA0_GFX_IB_CNTL);
 862         ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1);
 863 #ifdef __BIG_ENDIAN
 864         ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1);
 865 #endif
 866         /* enable DMA IBs */
 867         WREG32_SDMA(i, mmSDMA0_GFX_IB_CNTL, ib_cntl);
 868
 869         ring->sched.ready = true;
 870 }
 871
 872 /**
 873  * sdma_v4_0_page_resume - setup and start the async dma engines
 874  *
 875  * @adev: amdgpu_device pointer
 876  * @i: instance to resume
 877  *
 878  * Set up the page DMA ring buffers and enable them (VEGA10).
 879  * Returns 0 for success, error for failure.
 880  */
 881 static void sdma_v4_0_page_resume(struct amdgpu_device *adev, unsigned int i)
 882 {
 883         struct amdgpu_ring *ring = &adev->sdma.instance[i].page;
 884         u32 rb_cntl, ib_cntl, wptr_poll_cntl;
 885         u32 wb_offset;
 886         u32 doorbell;
 887         u32 doorbell_offset;
 888         u64 wptr_gpu_addr;
 889
 890         wb_offset = (ring->rptr_offs * 4);
 891
 892         rb_cntl = RREG32_SDMA(i, mmSDMA0_PAGE_RB_CNTL);
 893         rb_cntl = sdma_v4_0_rb_cntl(ring, rb_cntl);
 894         WREG32_SDMA(i, mmSDMA0_PAGE_RB_CNTL, rb_cntl);
 895
 896         /* Initialize the ring buffer's read and write pointers */
 897         WREG32_SDMA(i, mmSDMA0_PAGE_RB_RPTR, 0);
 898         WREG32_SDMA(i, mmSDMA0_PAGE_RB_RPTR_HI, 0);
 899         WREG32_SDMA(i, mmSDMA0_PAGE_RB_WPTR, 0);
 900         WREG32_SDMA(i, mmSDMA0_PAGE_RB_WPTR_HI, 0);
 901
 902         /* set the wb address whether it's enabled or not */
 903         WREG32_SDMA(i, mmSDMA0_PAGE_RB_RPTR_ADDR_HI,
 904                upper_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFF);
 905         WREG32_SDMA(i, mmSDMA0_PAGE_RB_RPTR_ADDR_LO,
 906                lower_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC);
 907
 908         rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_PAGE_RB_CNTL,
 909                                 RPTR_WRITEBACK_ENABLE, 1);
 910
 911         WREG32_SDMA(i, mmSDMA0_PAGE_RB_BASE, ring->gpu_addr >> 8);
 912         WREG32_SDMA(i, mmSDMA0_PAGE_RB_BASE_HI, ring->gpu_addr >> 40);
 913
 914         ring->wptr = 0;
 915
 916         /* before programing wptr to a less value, need set minor_ptr_update first */
 917         WREG32_SDMA(i, mmSDMA0_PAGE_MINOR_PTR_UPDATE, 1);
 918
 919         doorbell = RREG32_SDMA(i, mmSDMA0_PAGE_DOORBELL);
 920         doorbell_offset = RREG32_SDMA(i, mmSDMA0_PAGE_DOORBELL_OFFSET);
 921
 922         doorbell = REG_SET_FIELD(doorbell, SDMA0_PAGE_DOORBELL, ENABLE,
 923                                  ring->use_doorbell);
 924         doorbell_offset = REG_SET_FIELD(doorbell_offset,
 925                                         SDMA0_PAGE_DOORBELL_OFFSET,
 926                                         OFFSET, ring->doorbell_index);
 927         WREG32_SDMA(i, mmSDMA0_PAGE_DOORBELL, doorbell);
 928         WREG32_SDMA(i, mmSDMA0_PAGE_DOORBELL_OFFSET, doorbell_offset);
 929
 930         /* paging queue doorbell range is setup at sdma_v4_0_gfx_resume */
 931         sdma_v4_0_page_ring_set_wptr(ring);
 932
 933         /* set minor_ptr_update to 0 after wptr programed */
 934         WREG32_SDMA(i, mmSDMA0_PAGE_MINOR_PTR_UPDATE, 0);
 935
 936         /* setup the wptr shadow polling */
 937         wptr_gpu_addr = adev->wb.gpu_addr + (ring->wptr_offs * 4);
 938         WREG32_SDMA(i, mmSDMA0_PAGE_RB_WPTR_POLL_ADDR_LO,
 939                     lower_32_bits(wptr_gpu_addr));
 940         WREG32_SDMA(i, mmSDMA0_PAGE_RB_WPTR_POLL_ADDR_HI,
 941                     upper_32_bits(wptr_gpu_addr));
 942         wptr_poll_cntl = RREG32_SDMA(i, mmSDMA0_PAGE_RB_WPTR_POLL_CNTL);
 943         wptr_poll_cntl = REG_SET_FIELD(wptr_poll_cntl,
 944                                        SDMA0_PAGE_RB_WPTR_POLL_CNTL,
 945                                        F32_POLL_ENABLE, amdgpu_sriov_vf(adev));
 946         WREG32_SDMA(i, mmSDMA0_PAGE_RB_WPTR_POLL_CNTL, wptr_poll_cntl);
 947
 948         /* enable DMA RB */
 949         rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_PAGE_RB_CNTL, RB_ENABLE, 1);
 950         WREG32_SDMA(i, mmSDMA0_PAGE_RB_CNTL, rb_cntl);
 951
 952         ib_cntl = RREG32_SDMA(i, mmSDMA0_PAGE_IB_CNTL);
 953         ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_PAGE_IB_CNTL, IB_ENABLE, 1);
 954 #ifdef __BIG_ENDIAN
 955         ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_PAGE_IB_CNTL, IB_SWAP_ENABLE, 1);
 956 #endif
 957         /* enable DMA IBs */
 958         WREG32_SDMA(i, mmSDMA0_PAGE_IB_CNTL, ib_cntl);
 959
 960         ring->sched.ready = true;
 961 }
 962
 963 static void
 964 sdma_v4_1_update_power_gating(struct amdgpu_device *adev, bool enable)
 965 {
 966         uint32_t def, data;
 967
 968         if (enable && (adev->pg_flags & AMD_PG_SUPPORT_SDMA)) {
 969                 /* enable idle interrupt */
 970                 def = data = RREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_CNTL));
 971                 data |= SDMA0_CNTL__CTXEMPTY_INT_ENABLE_MASK;
 972
 973                 if (data != def)
 974                         WREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_CNTL), data);
 975         } else {
 976                 /* disable idle interrupt */
 977                 def = data = RREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_CNTL));
 978                 data &= ~SDMA0_CNTL__CTXEMPTY_INT_ENABLE_MASK;
 979                 if (data != def)
 980                         WREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_CNTL), data);
 981         }
 982 }
 983
 984 static void sdma_v4_1_init_power_gating(struct amdgpu_device *adev)
 985 {
 986         uint32_t def, data;
 987
 988         /* Enable HW based PG. */
 989         def = data = RREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_POWER_CNTL));
 990         data |= SDMA0_POWER_CNTL__PG_CNTL_ENABLE_MASK;
 991         if (data != def)
 992                 WREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_POWER_CNTL), data);
 993
 994         /* enable interrupt */
 995         def = data = RREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_CNTL));
 996         data |= SDMA0_CNTL__CTXEMPTY_INT_ENABLE_MASK;
 997         if (data != def)
 998                 WREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_CNTL), data);
 999
1000         /* Configure hold time to filter in-valid power on/off request. Use default right now */
1001         def = data = RREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_POWER_CNTL));
1002         data &= ~SDMA0_POWER_CNTL__ON_OFF_CONDITION_HOLD_TIME_MASK;
1003         data |= (mmSDMA0_POWER_CNTL_DEFAULT & SDMA0_POWER_CNTL__ON_OFF_CONDITION_HOLD_TIME_MASK);
1004         /* Configure switch time for hysteresis purpose. Use default right now */
1005         data &= ~SDMA0_POWER_CNTL__ON_OFF_STATUS_DURATION_TIME_MASK;
1006         data |= (mmSDMA0_POWER_CNTL_DEFAULT & SDMA0_POWER_CNTL__ON_OFF_STATUS_DURATION_TIME_MASK);
1007         if(data != def)
1008                 WREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_POWER_CNTL), data);
1009 }
1010
1011 static void sdma_v4_0_init_pg(struct amdgpu_device *adev)
1012 {
1013         if (!(adev->pg_flags & AMD_PG_SUPPORT_SDMA))
1014                 return;
1015
1016         switch (adev->asic_type) {
1017         case CHIP_RAVEN:
1018                 sdma_v4_1_init_power_gating(adev);
1019                 sdma_v4_1_update_power_gating(adev, true);
1020                 break;
1021         default:
1022                 break;
1023         }
1024 }
1025
1026 /**
1027  * sdma_v4_0_rlc_resume - setup and start the async dma engines
1028  *
1029  * @adev: amdgpu_device pointer
1030  *
1031  * Set up the compute DMA queues and enable them (VEGA10).
1032  * Returns 0 for success, error for failure.
1033  */
1034 static int sdma_v4_0_rlc_resume(struct amdgpu_device *adev)
1035 {
1036         sdma_v4_0_init_pg(adev);
1037
1038         return 0;
1039 }
1040
1041 /**
1042  * sdma_v4_0_load_microcode - load the sDMA ME ucode
1043  *
1044  * @adev: amdgpu_device pointer
1045  *
1046  * Loads the sDMA0/1 ucode.
1047  * Returns 0 for success, -EINVAL if the ucode is not available.
1048  */
1049 static int sdma_v4_0_load_microcode(struct amdgpu_device *adev)
1050 {
1051         const struct sdma_firmware_header_v1_0 *hdr;
1052         const __le32 *fw_data;
1053         u32 fw_size;
1054         int i, j;
1055
1056         /* halt the MEs */
1057         sdma_v4_0_enable(adev, false);
1058
1059         for (i = 0; i < adev->sdma.num_instances; i++) {
1060                 if (!adev->sdma.instance[i].fw)
1061                         return -EINVAL;
1062
1063                 hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
1064                 amdgpu_ucode_print_sdma_hdr(&hdr->header);
1065                 fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
1066
1067                 fw_data = (const __le32 *)
1068                         (adev->sdma.instance[i].fw->data +
1069                                 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1070
1071                 WREG32_SDMA(i, mmSDMA0_UCODE_ADDR, 0);
1072
1073                 for (j = 0; j < fw_size; j++)
1074                         WREG32_SDMA(i, mmSDMA0_UCODE_DATA,
1075                                     le32_to_cpup(fw_data++));
1076
1077                 WREG32_SDMA(i, mmSDMA0_UCODE_ADDR,
1078                             adev->sdma.instance[i].fw_version);
1079         }
1080
1081         return 0;
1082 }
1083
1084 /**
1085  * sdma_v4_0_start - setup and start the async dma engines
1086  *
1087  * @adev: amdgpu_device pointer
1088  *
1089  * Set up the DMA engines and enable them (VEGA10).
1090  * Returns 0 for success, error for failure.
1091  */
1092 static int sdma_v4_0_start(struct amdgpu_device *adev)
1093 {
1094         struct amdgpu_ring *ring;
1095         int i, r;
1096
1097         if (amdgpu_sriov_vf(adev)) {
1098                 sdma_v4_0_ctx_switch_enable(adev, false);
1099                 sdma_v4_0_enable(adev, false);
1100         } else {
1101
1102                 if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP) {
1103                         r = sdma_v4_0_load_microcode(adev);
1104                         if (r)
1105                                 return r;
1106                 }
1107
1108                 /* unhalt the MEs */
1109                 sdma_v4_0_enable(adev, true);
1110                 /* enable sdma ring preemption */
1111                 sdma_v4_0_ctx_switch_enable(adev, true);
1112         }
1113
1114         /* start the gfx rings and rlc compute queues */
1115         for (i = 0; i < adev->sdma.num_instances; i++) {
1116                 uint32_t temp;
1117
1118                 WREG32_SDMA(i, mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL, 0);
1119                 sdma_v4_0_gfx_resume(adev, i);
1120                 if (adev->sdma.has_page_queue)
1121                         sdma_v4_0_page_resume(adev, i);
1122
1123                 /* set utc l1 enable flag always to 1 */
1124                 temp = RREG32_SDMA(i, mmSDMA0_CNTL);
1125                 temp = REG_SET_FIELD(temp, SDMA0_CNTL, UTC_L1_ENABLE, 1);
1126                 WREG32_SDMA(i, mmSDMA0_CNTL, temp);
1127
1128                 if (!amdgpu_sriov_vf(adev)) {
1129                         /* unhalt engine */
1130                         temp = RREG32_SDMA(i, mmSDMA0_F32_CNTL);
1131                         temp = REG_SET_FIELD(temp, SDMA0_F32_CNTL, HALT, 0);
1132                         WREG32_SDMA(i, mmSDMA0_F32_CNTL, temp);
1133                 }
1134         }
1135
1136         if (amdgpu_sriov_vf(adev)) {
1137                 sdma_v4_0_ctx_switch_enable(adev, true);
1138                 sdma_v4_0_enable(adev, true);
1139         } else {
1140                 r = sdma_v4_0_rlc_resume(adev);
1141                 if (r)
1142                         return r;
1143         }
1144
1145         for (i = 0; i < adev->sdma.num_instances; i++) {
1146                 ring = &adev->sdma.instance[i].ring;
1147
1148                 r = amdgpu_ring_test_helper(ring);
1149                 if (r)
1150                         return r;
1151
1152                 if (adev->sdma.has_page_queue) {
1153                         struct amdgpu_ring *page = &adev->sdma.instance[i].page;
1154
1155                         r = amdgpu_ring_test_helper(page);
1156                         if (r)
1157                                 return r;
1158
1159                         if (adev->mman.buffer_funcs_ring == page)
1160                                 amdgpu_ttm_set_buffer_funcs_status(adev, true);
1161                 }
1162
1163                 if (adev->mman.buffer_funcs_ring == ring)
1164                         amdgpu_ttm_set_buffer_funcs_status(adev, true);
1165         }
1166
1167         return r;
1168 }
1169
1170 /**
1171  * sdma_v4_0_ring_test_ring - simple async dma engine test
1172  *
1173  * @ring: amdgpu_ring structure holding ring information
1174  *
1175  * Test the DMA engine by writing using it to write an
1176  * value to memory. (VEGA10).
1177  * Returns 0 for success, error for failure.
1178  */
1179 static int sdma_v4_0_ring_test_ring(struct amdgpu_ring *ring)
1180 {
1181         struct amdgpu_device *adev = ring->adev;
1182         unsigned i;
1183         unsigned index;
1184         int r;
1185         u32 tmp;
1186         u64 gpu_addr;
1187
1188         r = amdgpu_device_wb_get(adev, &index);
1189         if (r)
1190                 return r;
1191
1192         gpu_addr = adev->wb.gpu_addr + (index * 4);
1193         tmp = 0xCAFEDEAD;
1194         adev->wb.wb[index] = cpu_to_le32(tmp);
1195
1196         r = amdgpu_ring_alloc(ring, 5);
1197         if (r)
1198                 goto error_free_wb;
1199
1200         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
1201                           SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
1202         amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
1203         amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
1204         amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0));
1205         amdgpu_ring_write(ring, 0xDEADBEEF);
1206         amdgpu_ring_commit(ring);
1207
1208         for (i = 0; i < adev->usec_timeout; i++) {
1209                 tmp = le32_to_cpu(adev->wb.wb[index]);
1210                 if (tmp == 0xDEADBEEF)
1211                         break;
1212                 DRM_UDELAY(1);
1213         }
1214
1215         if (i >= adev->usec_timeout)
1216                 r = -ETIMEDOUT;
1217
1218 error_free_wb:
1219         amdgpu_device_wb_free(adev, index);
1220         return r;
1221 }
1222
1223 /**
1224  * sdma_v4_0_ring_test_ib - test an IB on the DMA engine
1225  *
1226  * @ring: amdgpu_ring structure holding ring information
1227  *
1228  * Test a simple IB in the DMA ring (VEGA10).
1229  * Returns 0 on success, error on failure.
1230  */
1231 static int sdma_v4_0_ring_test_ib(struct amdgpu_ring *ring, long timeout)
1232 {
1233         struct amdgpu_device *adev = ring->adev;
1234         struct amdgpu_ib ib;
1235         struct dma_fence *f = NULL;
1236         unsigned index;
1237         long r;
1238         u32 tmp = 0;
1239         u64 gpu_addr;
1240
1241         r = amdgpu_device_wb_get(adev, &index);
1242         if (r)
1243                 return r;
1244
1245         gpu_addr = adev->wb.gpu_addr + (index * 4);
1246         tmp = 0xCAFEDEAD;
1247         adev->wb.wb[index] = cpu_to_le32(tmp);
1248         memset(&ib, 0, sizeof(ib));
1249         r = amdgpu_ib_get(adev, NULL, 256, &ib);
1250         if (r)
1251                 goto err0;
1252
1253         ib.ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
1254                 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1255         ib.ptr[1] = lower_32_bits(gpu_addr);
1256         ib.ptr[2] = upper_32_bits(gpu_addr);
1257         ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0);
1258         ib.ptr[4] = 0xDEADBEEF;
1259         ib.ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1260         ib.ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1261         ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1262         ib.length_dw = 8;
1263
1264         r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
1265         if (r)
1266                 goto err1;
1267
1268         r = dma_fence_wait_timeout(f, false, timeout);
1269         if (r == 0) {
1270                 r = -ETIMEDOUT;
1271                 goto err1;
1272         } else if (r < 0) {
1273                 goto err1;
1274         }
1275         tmp = le32_to_cpu(adev->wb.wb[index]);
1276         if (tmp == 0xDEADBEEF)
1277                 r = 0;
1278         else
1279                 r = -EINVAL;
1280
1281 err1:
1282         amdgpu_ib_free(adev, &ib, NULL);
1283         dma_fence_put(f);
1284 err0:
1285         amdgpu_device_wb_free(adev, index);
1286         return r;
1287 }
1288
1289
1290 /**
1291  * sdma_v4_0_vm_copy_pte - update PTEs by copying them from the GART
1292  *
1293  * @ib: indirect buffer to fill with commands
1294  * @pe: addr of the page entry
1295  * @src: src addr to copy from
1296  * @count: number of page entries to update
1297  *
1298  * Update PTEs by copying them from the GART using sDMA (VEGA10).
1299  */
1300 static void sdma_v4_0_vm_copy_pte(struct amdgpu_ib *ib,
1301                                   uint64_t pe, uint64_t src,
1302                                   unsigned count)
1303 {
1304         unsigned bytes = count * 8;
1305
1306         ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1307                 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
1308         ib->ptr[ib->length_dw++] = bytes - 1;
1309         ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1310         ib->ptr[ib->length_dw++] = lower_32_bits(src);
1311         ib->ptr[ib->length_dw++] = upper_32_bits(src);
1312         ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1313         ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1314
1315 }
1316
1317 /**
1318  * sdma_v4_0_vm_write_pte - update PTEs by writing them manually
1319  *
1320  * @ib: indirect buffer to fill with commands
1321  * @pe: addr of the page entry
1322  * @addr: dst addr to write into pe
1323  * @count: number of page entries to update
1324  * @incr: increase next addr by incr bytes
1325  * @flags: access flags
1326  *
1327  * Update PTEs by writing them manually using sDMA (VEGA10).
1328  */
1329 static void sdma_v4_0_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
1330                                    uint64_t value, unsigned count,
1331                                    uint32_t incr)
1332 {
1333         unsigned ndw = count * 2;
1334
1335         ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
1336                 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1337         ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1338         ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1339         ib->ptr[ib->length_dw++] = ndw - 1;
1340         for (; ndw > 0; ndw -= 2) {
1341                 ib->ptr[ib->length_dw++] = lower_32_bits(value);
1342                 ib->ptr[ib->length_dw++] = upper_32_bits(value);
1343                 value += incr;
1344         }
1345 }
1346
1347 /**
1348  * sdma_v4_0_vm_set_pte_pde - update the page tables using sDMA
1349  *
1350  * @ib: indirect buffer to fill with commands
1351  * @pe: addr of the page entry
1352  * @addr: dst addr to write into pe
1353  * @count: number of page entries to update
1354  * @incr: increase next addr by incr bytes
1355  * @flags: access flags
1356  *
1357  * Update the page tables using sDMA (VEGA10).
1358  */
1359 static void sdma_v4_0_vm_set_pte_pde(struct amdgpu_ib *ib,
1360                                      uint64_t pe,
1361                                      uint64_t addr, unsigned count,
1362                                      uint32_t incr, uint64_t flags)
1363 {
1364         /* for physically contiguous pages (vram) */
1365         ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_PTEPDE);
1366         ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
1367         ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1368         ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
1369         ib->ptr[ib->length_dw++] = upper_32_bits(flags);
1370         ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
1371         ib->ptr[ib->length_dw++] = upper_32_bits(addr);
1372         ib->ptr[ib->length_dw++] = incr; /* increment size */
1373         ib->ptr[ib->length_dw++] = 0;
1374         ib->ptr[ib->length_dw++] = count - 1; /* number of entries */
1375 }
1376
1377 /**
1378  * sdma_v4_0_ring_pad_ib - pad the IB to the required number of dw
1379  *
1380  * @ib: indirect buffer to fill with padding
1381  *
1382  */
1383 static void sdma_v4_0_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
1384 {
1385         struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
1386         u32 pad_count;
1387         int i;
1388
1389         pad_count = (8 - (ib->length_dw & 0x7)) % 8;
1390         for (i = 0; i < pad_count; i++)
1391                 if (sdma && sdma->burst_nop && (i == 0))
1392                         ib->ptr[ib->length_dw++] =
1393                                 SDMA_PKT_HEADER_OP(SDMA_OP_NOP) |
1394                                 SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
1395                 else
1396                         ib->ptr[ib->length_dw++] =
1397                                 SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
1398 }
1399
1400
1401 /**
1402  * sdma_v4_0_ring_emit_pipeline_sync - sync the pipeline
1403  *
1404  * @ring: amdgpu_ring pointer
1405  *
1406  * Make sure all previous operations are completed (CIK).
1407  */
1408 static void sdma_v4_0_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
1409 {
1410         uint32_t seq = ring->fence_drv.sync_seq;
1411         uint64_t addr = ring->fence_drv.gpu_addr;
1412
1413         /* wait for idle */
1414         sdma_v4_0_wait_reg_mem(ring, 1, 0,
1415                                addr & 0xfffffffc,
1416                                upper_32_bits(addr) & 0xffffffff,
1417                                seq, 0xffffffff, 4);
1418 }
1419
1420
1421 /**
1422  * sdma_v4_0_ring_emit_vm_flush - vm flush using sDMA
1423  *
1424  * @ring: amdgpu_ring pointer
1425  * @vm: amdgpu_vm pointer
1426  *
1427  * Update the page table base and flush the VM TLB
1428  * using sDMA (VEGA10).
1429  */
1430 static void sdma_v4_0_ring_emit_vm_flush(struct amdgpu_ring *ring,
1431                                          unsigned vmid, uint64_t pd_addr)
1432 {
1433         amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
1434 }
1435
1436 static void sdma_v4_0_ring_emit_wreg(struct amdgpu_ring *ring,
1437                                      uint32_t reg, uint32_t val)
1438 {
1439         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
1440                           SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
1441         amdgpu_ring_write(ring, reg);
1442         amdgpu_ring_write(ring, val);
1443 }
1444
1445 static void sdma_v4_0_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg,
1446                                          uint32_t val, uint32_t mask)
1447 {
1448         sdma_v4_0_wait_reg_mem(ring, 0, 0, reg, 0, val, mask, 10);
1449 }
1450
1451 static bool sdma_v4_0_fw_support_paging_queue(struct amdgpu_device *adev)
1452 {
1453         uint fw_version = adev->sdma.instance[0].fw_version;
1454
1455         switch (adev->asic_type) {
1456         case CHIP_VEGA10:
1457                 return fw_version >= 430;
1458         case CHIP_VEGA12:
1459                 /*return fw_version >= 31;*/
1460                 return false;
1461         case CHIP_VEGA20:
1462                 return fw_version >= 123;
1463         default:
1464                 return false;
1465         }
1466 }
1467
1468 static int sdma_v4_0_early_init(void *handle)
1469 {
1470         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1471         int r;
1472
1473         if (adev->asic_type == CHIP_RAVEN)
1474                 adev->sdma.num_instances = 1;
1475         else
1476                 adev->sdma.num_instances = 2;
1477
1478         r = sdma_v4_0_init_microcode(adev);
1479         if (r) {
1480                 DRM_ERROR("Failed to load sdma firmware!\n");
1481                 return r;
1482         }
1483
1484         /* TODO: Page queue breaks driver reload under SRIOV */
1485         if ((adev->asic_type == CHIP_VEGA10) && amdgpu_sriov_vf((adev)))
1486                 adev->sdma.has_page_queue = false;
1487         else if (sdma_v4_0_fw_support_paging_queue(adev))
1488                 adev->sdma.has_page_queue = true;
1489
1490         sdma_v4_0_set_ring_funcs(adev);
1491         sdma_v4_0_set_buffer_funcs(adev);
1492         sdma_v4_0_set_vm_pte_funcs(adev);
1493         sdma_v4_0_set_irq_funcs(adev);
1494
1495         return 0;
1496 }
1497
1498 static int sdma_v4_0_sw_init(void *handle)
1499 {
1500         struct amdgpu_ring *ring;
1501         int r, i;
1502         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1503
1504         /* SDMA trap event */
1505         r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_SDMA0, SDMA0_4_0__SRCID__SDMA_TRAP,
1506                               &adev->sdma.trap_irq);
1507         if (r)
1508                 return r;
1509
1510         /* SDMA trap event */
1511         r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_SDMA1, SDMA1_4_0__SRCID__SDMA_TRAP,
1512                               &adev->sdma.trap_irq);
1513         if (r)
1514                 return r;
1515
1516         for (i = 0; i < adev->sdma.num_instances; i++) {
1517                 ring = &adev->sdma.instance[i].ring;
1518                 ring->ring_obj = NULL;
1519                 ring->use_doorbell = true;
1520
1521                 DRM_INFO("use_doorbell being set to: [%s]\n",
1522                                 ring->use_doorbell?"true":"false");
1523
1524                 /* doorbell size is 2 dwords, get DWORD offset */
1525                 ring->doorbell_index = (i == 0) ?
1526                         (adev->doorbell_index.sdma_engine0 << 1)
1527                         : (adev->doorbell_index.sdma_engine1 << 1);
1528
1529                 sprintf(ring->name, "sdma%d", i);
1530                 r = amdgpu_ring_init(adev, ring, 1024,
1531                                      &adev->sdma.trap_irq,
1532                                      (i == 0) ?
1533                                      AMDGPU_SDMA_IRQ_TRAP0 :
1534                                      AMDGPU_SDMA_IRQ_TRAP1);
1535                 if (r)
1536                         return r;
1537
1538                 if (adev->sdma.has_page_queue) {
1539                         ring = &adev->sdma.instance[i].page;
1540                         ring->ring_obj = NULL;
1541                         ring->use_doorbell = true;
1542
1543                         /* paging queue use same doorbell index/routing as gfx queue
1544                          * with 0x400 (4096 dwords) offset on second doorbell page
1545                          */
1546                         ring->doorbell_index = (i == 0) ?
1547                                 (adev->doorbell_index.sdma_engine0 << 1)
1548                                 : (adev->doorbell_index.sdma_engine1 << 1);
1549                         ring->doorbell_index += 0x400;
1550
1551                         sprintf(ring->name, "page%d", i);
1552                         r = amdgpu_ring_init(adev, ring, 1024,
1553                                              &adev->sdma.trap_irq,
1554                                              (i == 0) ?
1555                                              AMDGPU_SDMA_IRQ_TRAP0 :
1556                                              AMDGPU_SDMA_IRQ_TRAP1);
1557                         if (r)
1558                                 return r;
1559                 }
1560         }
1561
1562         return r;
1563 }
1564
1565 static int sdma_v4_0_sw_fini(void *handle)
1566 {
1567         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1568         int i;
1569
1570         for (i = 0; i < adev->sdma.num_instances; i++) {
1571                 amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1572                 if (adev->sdma.has_page_queue)
1573                         amdgpu_ring_fini(&adev->sdma.instance[i].page);
1574         }
1575
1576         for (i = 0; i < adev->sdma.num_instances; i++) {
1577                 release_firmware(adev->sdma.instance[i].fw);
1578                 adev->sdma.instance[i].fw = NULL;
1579         }
1580
1581         return 0;
1582 }
1583
1584 static int sdma_v4_0_hw_init(void *handle)
1585 {
1586         int r;
1587         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1588
1589         if (adev->asic_type == CHIP_RAVEN && adev->powerplay.pp_funcs &&
1590                         adev->powerplay.pp_funcs->set_powergating_by_smu)
1591                 amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_SDMA, false);
1592
1593         sdma_v4_0_init_golden_registers(adev);
1594
1595         r = sdma_v4_0_start(adev);
1596
1597         return r;
1598 }
1599
1600 static int sdma_v4_0_hw_fini(void *handle)
1601 {
1602         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1603
1604         if (amdgpu_sriov_vf(adev))
1605                 return 0;
1606
1607         sdma_v4_0_ctx_switch_enable(adev, false);
1608         sdma_v4_0_enable(adev, false);
1609
1610         if (adev->asic_type == CHIP_RAVEN && adev->powerplay.pp_funcs
1611                         && adev->powerplay.pp_funcs->set_powergating_by_smu)
1612                 amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_SDMA, true);
1613
1614         return 0;
1615 }
1616
1617 static int sdma_v4_0_suspend(void *handle)
1618 {
1619         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1620
1621         return sdma_v4_0_hw_fini(adev);
1622 }
1623
1624 static int sdma_v4_0_resume(void *handle)
1625 {
1626         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1627
1628         return sdma_v4_0_hw_init(adev);
1629 }
1630
1631 static bool sdma_v4_0_is_idle(void *handle)
1632 {
1633         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1634         u32 i;
1635
1636         for (i = 0; i < adev->sdma.num_instances; i++) {
1637                 u32 tmp = RREG32_SDMA(i, mmSDMA0_STATUS_REG);
1638
1639                 if (!(tmp & SDMA0_STATUS_REG__IDLE_MASK))
1640                         return false;
1641         }
1642
1643         return true;
1644 }
1645
1646 static int sdma_v4_0_wait_for_idle(void *handle)
1647 {
1648         unsigned i;
1649         u32 sdma0, sdma1;
1650         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1651
1652         for (i = 0; i < adev->usec_timeout; i++) {
1653                 sdma0 = RREG32_SDMA(0, mmSDMA0_STATUS_REG);
1654                 sdma1 = RREG32_SDMA(1, mmSDMA0_STATUS_REG);
1655
1656                 if (sdma0 & sdma1 & SDMA0_STATUS_REG__IDLE_MASK)
1657                         return 0;
1658                 udelay(1);
1659         }
1660         return -ETIMEDOUT;
1661 }
1662
1663 static int sdma_v4_0_soft_reset(void *handle)
1664 {
1665         /* todo */
1666
1667         return 0;
1668 }
1669
1670 static int sdma_v4_0_set_trap_irq_state(struct amdgpu_device *adev,
1671                                         struct amdgpu_irq_src *source,
1672                                         unsigned type,
1673                                         enum amdgpu_interrupt_state state)
1674 {
1675         unsigned int instance = (type == AMDGPU_SDMA_IRQ_TRAP0) ? 0 : 1;
1676         u32 sdma_cntl;
1677
1678         sdma_cntl = RREG32_SDMA(instance, mmSDMA0_CNTL);
1679         sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE,
1680                        state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
1681         WREG32_SDMA(instance, mmSDMA0_CNTL, sdma_cntl);
1682
1683         return 0;
1684 }
1685
1686 static int sdma_v4_0_process_trap_irq(struct amdgpu_device *adev,
1687                                       struct amdgpu_irq_src *source,
1688                                       struct amdgpu_iv_entry *entry)
1689 {
1690         uint32_t instance;
1691
1692         DRM_DEBUG("IH: SDMA trap\n");
1693         switch (entry->client_id) {
1694         case SOC15_IH_CLIENTID_SDMA0:
1695                 instance = 0;
1696                 break;
1697         case SOC15_IH_CLIENTID_SDMA1:
1698                 instance = 1;
1699                 break;
1700         default:
1701                 return 0;
1702         }
1703
1704         switch (entry->ring_id) {
1705         case 0:
1706                 amdgpu_fence_process(&adev->sdma.instance[instance].ring);
1707                 break;
1708         case 1:
1709                 if (adev->asic_type == CHIP_VEGA20)
1710                         amdgpu_fence_process(&adev->sdma.instance[instance].page);
1711                 break;
1712         case 2:
1713                 /* XXX compute */
1714                 break;
1715         case 3:
1716                 if (adev->asic_type != CHIP_VEGA20)
1717                         amdgpu_fence_process(&adev->sdma.instance[instance].page);
1718                 break;
1719         }
1720         return 0;
1721 }
1722
1723 static int sdma_v4_0_process_illegal_inst_irq(struct amdgpu_device *adev,
1724                                               struct amdgpu_irq_src *source,
1725                                               struct amdgpu_iv_entry *entry)
1726 {
1727         int instance;
1728
1729         DRM_ERROR("Illegal instruction in SDMA command stream\n");
1730
1731         switch (entry->client_id) {
1732         case SOC15_IH_CLIENTID_SDMA0:
1733                 instance = 0;
1734                 break;
1735         case SOC15_IH_CLIENTID_SDMA1:
1736                 instance = 1;
1737                 break;
1738         default:
1739                 return 0;
1740         }
1741
1742         switch (entry->ring_id) {
1743         case 0:
1744                 drm_sched_fault(&adev->sdma.instance[instance].ring.sched);
1745                 break;
1746         }
1747         return 0;
1748 }
1749
1750 static void sdma_v4_0_update_medium_grain_clock_gating(
1751                 struct amdgpu_device *adev,
1752                 bool enable)
1753 {
1754         uint32_t data, def;
1755
1756         if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
1757                 /* enable sdma0 clock gating */
1758                 def = data = RREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_CLK_CTRL));
1759                 data &= ~(SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK |
1760                           SDMA0_CLK_CTRL__SOFT_OVERRIDE6_MASK |
1761                           SDMA0_CLK_CTRL__SOFT_OVERRIDE5_MASK |
1762                           SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1763                           SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1764                           SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1765                           SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1766                           SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK);
1767                 if (def != data)
1768                         WREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_CLK_CTRL), data);
1769
1770                 if (adev->sdma.num_instances > 1) {
1771                         def = data = RREG32(SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_CLK_CTRL));
1772                         data &= ~(SDMA1_CLK_CTRL__SOFT_OVERRIDE7_MASK |
1773                                   SDMA1_CLK_CTRL__SOFT_OVERRIDE6_MASK |
1774                                   SDMA1_CLK_CTRL__SOFT_OVERRIDE5_MASK |
1775                                   SDMA1_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1776                                   SDMA1_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1777                                   SDMA1_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1778                                   SDMA1_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1779                                   SDMA1_CLK_CTRL__SOFT_OVERRIDE0_MASK);
1780                         if (def != data)
1781                                 WREG32(SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_CLK_CTRL), data);
1782                 }
1783         } else {
1784                 /* disable sdma0 clock gating */
1785                 def = data = RREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_CLK_CTRL));
1786                 data |= (SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK |
1787                          SDMA0_CLK_CTRL__SOFT_OVERRIDE6_MASK |
1788                          SDMA0_CLK_CTRL__SOFT_OVERRIDE5_MASK |
1789                          SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1790                          SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1791                          SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1792                          SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1793                          SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK);
1794
1795                 if (def != data)
1796                         WREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_CLK_CTRL), data);
1797
1798                 if (adev->sdma.num_instances > 1) {
1799                         def = data = RREG32(SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_CLK_CTRL));
1800                         data |= (SDMA1_CLK_CTRL__SOFT_OVERRIDE7_MASK |
1801                                  SDMA1_CLK_CTRL__SOFT_OVERRIDE6_MASK |
1802                                  SDMA1_CLK_CTRL__SOFT_OVERRIDE5_MASK |
1803                                  SDMA1_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1804                                  SDMA1_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1805                                  SDMA1_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1806                                  SDMA1_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1807                                  SDMA1_CLK_CTRL__SOFT_OVERRIDE0_MASK);
1808                         if (def != data)
1809                                 WREG32(SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_CLK_CTRL), data);
1810                 }
1811         }
1812 }
1813
1814
1815 static void sdma_v4_0_update_medium_grain_light_sleep(
1816                 struct amdgpu_device *adev,
1817                 bool enable)
1818 {
1819         uint32_t data, def;
1820
1821         if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) {
1822                 /* 1-not override: enable sdma0 mem light sleep */
1823                 def = data = RREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_POWER_CNTL));
1824                 data |= SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1825                 if (def != data)
1826                         WREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_POWER_CNTL), data);
1827
1828                 /* 1-not override: enable sdma1 mem light sleep */
1829                 if (adev->sdma.num_instances > 1) {
1830                         def = data = RREG32(SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_POWER_CNTL));
1831                         data |= SDMA1_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1832                         if (def != data)
1833                                 WREG32(SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_POWER_CNTL), data);
1834                 }
1835         } else {
1836                 /* 0-override:disable sdma0 mem light sleep */
1837                 def = data = RREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_POWER_CNTL));
1838                 data &= ~SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1839                 if (def != data)
1840                         WREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_POWER_CNTL), data);
1841
1842                 /* 0-override:disable sdma1 mem light sleep */
1843                 if (adev->sdma.num_instances > 1) {
1844                         def = data = RREG32(SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_POWER_CNTL));
1845                         data &= ~SDMA1_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1846                         if (def != data)
1847                                 WREG32(SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_POWER_CNTL), data);
1848                 }
1849         }
1850 }
1851
1852 static int sdma_v4_0_set_clockgating_state(void *handle,
1853                                           enum amd_clockgating_state state)
1854 {
1855         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1856
1857         if (amdgpu_sriov_vf(adev))
1858                 return 0;
1859
1860         switch (adev->asic_type) {
1861         case CHIP_VEGA10:
1862         case CHIP_VEGA12:
1863         case CHIP_VEGA20:
1864         case CHIP_RAVEN:
1865                 sdma_v4_0_update_medium_grain_clock_gating(adev,
1866                                 state == AMD_CG_STATE_GATE ? true : false);
1867                 sdma_v4_0_update_medium_grain_light_sleep(adev,
1868                                 state == AMD_CG_STATE_GATE ? true : false);
1869                 break;
1870         default:
1871                 break;
1872         }
1873         return 0;
1874 }
1875
1876 static int sdma_v4_0_set_powergating_state(void *handle,
1877                                           enum amd_powergating_state state)
1878 {
1879         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1880
1881         switch (adev->asic_type) {
1882         case CHIP_RAVEN:
1883                 sdma_v4_1_update_power_gating(adev,
1884                                 state == AMD_PG_STATE_GATE ? true : false);
1885                 break;
1886         default:
1887                 break;
1888         }
1889
1890         return 0;
1891 }
1892
1893 static void sdma_v4_0_get_clockgating_state(void *handle, u32 *flags)
1894 {
1895         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1896         int data;
1897
1898         if (amdgpu_sriov_vf(adev))
1899                 *flags = 0;
1900
1901         /* AMD_CG_SUPPORT_SDMA_MGCG */
1902         data = RREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_CLK_CTRL));
1903         if (!(data & SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK))
1904                 *flags |= AMD_CG_SUPPORT_SDMA_MGCG;
1905
1906         /* AMD_CG_SUPPORT_SDMA_LS */
1907         data = RREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_POWER_CNTL));
1908         if (data & SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK)
1909                 *flags |= AMD_CG_SUPPORT_SDMA_LS;
1910 }
1911
1912 const struct amd_ip_funcs sdma_v4_0_ip_funcs = {
1913         .name = "sdma_v4_0",
1914         .early_init = sdma_v4_0_early_init,
1915         .late_init = NULL,
1916         .sw_init = sdma_v4_0_sw_init,
1917         .sw_fini = sdma_v4_0_sw_fini,
1918         .hw_init = sdma_v4_0_hw_init,
1919         .hw_fini = sdma_v4_0_hw_fini,
1920         .suspend = sdma_v4_0_suspend,
1921         .resume = sdma_v4_0_resume,
1922         .is_idle = sdma_v4_0_is_idle,
1923         .wait_for_idle = sdma_v4_0_wait_for_idle,
1924         .soft_reset = sdma_v4_0_soft_reset,
1925         .set_clockgating_state = sdma_v4_0_set_clockgating_state,
1926         .set_powergating_state = sdma_v4_0_set_powergating_state,
1927         .get_clockgating_state = sdma_v4_0_get_clockgating_state,
1928 };
1929
1930 static const struct amdgpu_ring_funcs sdma_v4_0_ring_funcs = {
1931         .type = AMDGPU_RING_TYPE_SDMA,
1932         .align_mask = 0xf,
1933         .nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP),
1934         .support_64bit_ptrs = true,
1935         .vmhub = AMDGPU_MMHUB,
1936         .get_rptr = sdma_v4_0_ring_get_rptr,
1937         .get_wptr = sdma_v4_0_ring_get_wptr,
1938         .set_wptr = sdma_v4_0_ring_set_wptr,
1939         .emit_frame_size =
1940                 6 + /* sdma_v4_0_ring_emit_hdp_flush */
1941                 3 + /* hdp invalidate */
1942                 6 + /* sdma_v4_0_ring_emit_pipeline_sync */
1943                 /* sdma_v4_0_ring_emit_vm_flush */
1944                 SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 +
1945                 SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6 +
1946                 10 + 10 + 10, /* sdma_v4_0_ring_emit_fence x3 for user fence, vm fence */
1947         .emit_ib_size = 7 + 6, /* sdma_v4_0_ring_emit_ib */
1948         .emit_ib = sdma_v4_0_ring_emit_ib,
1949         .emit_fence = sdma_v4_0_ring_emit_fence,
1950         .emit_pipeline_sync = sdma_v4_0_ring_emit_pipeline_sync,
1951         .emit_vm_flush = sdma_v4_0_ring_emit_vm_flush,
1952         .emit_hdp_flush = sdma_v4_0_ring_emit_hdp_flush,
1953         .test_ring = sdma_v4_0_ring_test_ring,
1954         .test_ib = sdma_v4_0_ring_test_ib,
1955         .insert_nop = sdma_v4_0_ring_insert_nop,
1956         .pad_ib = sdma_v4_0_ring_pad_ib,
1957         .emit_wreg = sdma_v4_0_ring_emit_wreg,
1958         .emit_reg_wait = sdma_v4_0_ring_emit_reg_wait,
1959         .emit_reg_write_reg_wait = amdgpu_ring_emit_reg_write_reg_wait_helper,
1960 };
1961
1962 static const struct amdgpu_ring_funcs sdma_v4_0_page_ring_funcs = {
1963         .type = AMDGPU_RING_TYPE_SDMA,
1964         .align_mask = 0xf,
1965         .nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP),
1966         .support_64bit_ptrs = true,
1967         .vmhub = AMDGPU_MMHUB,
1968         .get_rptr = sdma_v4_0_ring_get_rptr,
1969         .get_wptr = sdma_v4_0_page_ring_get_wptr,
1970         .set_wptr = sdma_v4_0_page_ring_set_wptr,
1971         .emit_frame_size =
1972                 6 + /* sdma_v4_0_ring_emit_hdp_flush */
1973                 3 + /* hdp invalidate */
1974                 6 + /* sdma_v4_0_ring_emit_pipeline_sync */
1975                 /* sdma_v4_0_ring_emit_vm_flush */
1976                 SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 +
1977                 SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6 +
1978                 10 + 10 + 10, /* sdma_v4_0_ring_emit_fence x3 for user fence, vm fence */
1979         .emit_ib_size = 7 + 6, /* sdma_v4_0_ring_emit_ib */
1980         .emit_ib = sdma_v4_0_ring_emit_ib,
1981         .emit_fence = sdma_v4_0_ring_emit_fence,
1982         .emit_pipeline_sync = sdma_v4_0_ring_emit_pipeline_sync,
1983         .emit_vm_flush = sdma_v4_0_ring_emit_vm_flush,
1984         .emit_hdp_flush = sdma_v4_0_ring_emit_hdp_flush,
1985         .test_ring = sdma_v4_0_ring_test_ring,
1986         .test_ib = sdma_v4_0_ring_test_ib,
1987         .insert_nop = sdma_v4_0_ring_insert_nop,
1988         .pad_ib = sdma_v4_0_ring_pad_ib,
1989         .emit_wreg = sdma_v4_0_ring_emit_wreg,
1990         .emit_reg_wait = sdma_v4_0_ring_emit_reg_wait,
1991         .emit_reg_write_reg_wait = amdgpu_ring_emit_reg_write_reg_wait_helper,
1992 };
1993
1994 static void sdma_v4_0_set_ring_funcs(struct amdgpu_device *adev)
1995 {
1996         int i;
1997
1998         for (i = 0; i < adev->sdma.num_instances; i++) {
1999                 adev->sdma.instance[i].ring.funcs = &sdma_v4_0_ring_funcs;
2000                 adev->sdma.instance[i].ring.me = i;
2001                 if (adev->sdma.has_page_queue) {
2002                         adev->sdma.instance[i].page.funcs = &sdma_v4_0_page_ring_funcs;
2003                         adev->sdma.instance[i].page.me = i;
2004                 }
2005         }
2006 }
2007
2008 static const struct amdgpu_irq_src_funcs sdma_v4_0_trap_irq_funcs = {
2009         .set = sdma_v4_0_set_trap_irq_state,
2010         .process = sdma_v4_0_process_trap_irq,
2011 };
2012
2013 static const struct amdgpu_irq_src_funcs sdma_v4_0_illegal_inst_irq_funcs = {
2014         .process = sdma_v4_0_process_illegal_inst_irq,
2015 };
2016
2017 static void sdma_v4_0_set_irq_funcs(struct amdgpu_device *adev)
2018 {
2019         adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
2020         adev->sdma.trap_irq.funcs = &sdma_v4_0_trap_irq_funcs;
2021         adev->sdma.illegal_inst_irq.funcs = &sdma_v4_0_illegal_inst_irq_funcs;
2022 }
2023
2024 /**
2025  * sdma_v4_0_emit_copy_buffer - copy buffer using the sDMA engine
2026  *
2027  * @ring: amdgpu_ring structure holding ring information
2028  * @src_offset: src GPU address
2029  * @dst_offset: dst GPU address
2030  * @byte_count: number of bytes to xfer
2031  *
2032  * Copy GPU buffers using the DMA engine (VEGA10/12).
2033  * Used by the amdgpu ttm implementation to move pages if
2034  * registered as the asic copy callback.
2035  */
2036 static void sdma_v4_0_emit_copy_buffer(struct amdgpu_ib *ib,
2037                                        uint64_t src_offset,
2038                                        uint64_t dst_offset,
2039                                        uint32_t byte_count)
2040 {
2041         ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
2042                 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
2043         ib->ptr[ib->length_dw++] = byte_count - 1;
2044         ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
2045         ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
2046         ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
2047         ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
2048         ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
2049 }
2050
2051 /**
2052  * sdma_v4_0_emit_fill_buffer - fill buffer using the sDMA engine
2053  *
2054  * @ring: amdgpu_ring structure holding ring information
2055  * @src_data: value to write to buffer
2056  * @dst_offset: dst GPU address
2057  * @byte_count: number of bytes to xfer
2058  *
2059  * Fill GPU buffers using the DMA engine (VEGA10/12).
2060  */
2061 static void sdma_v4_0_emit_fill_buffer(struct amdgpu_ib *ib,
2062                                        uint32_t src_data,
2063                                        uint64_t dst_offset,
2064                                        uint32_t byte_count)
2065 {
2066         ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL);
2067         ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
2068         ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
2069         ib->ptr[ib->length_dw++] = src_data;
2070         ib->ptr[ib->length_dw++] = byte_count - 1;
2071 }
2072
2073 static const struct amdgpu_buffer_funcs sdma_v4_0_buffer_funcs = {
2074         .copy_max_bytes = 0x400000,
2075         .copy_num_dw = 7,
2076         .emit_copy_buffer = sdma_v4_0_emit_copy_buffer,
2077
2078         .fill_max_bytes = 0x400000,
2079         .fill_num_dw = 5,
2080         .emit_fill_buffer = sdma_v4_0_emit_fill_buffer,
2081 };
2082
2083 static void sdma_v4_0_set_buffer_funcs(struct amdgpu_device *adev)
2084 {
2085         adev->mman.buffer_funcs = &sdma_v4_0_buffer_funcs;
2086         if (adev->sdma.has_page_queue)
2087                 adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].page;
2088         else
2089                 adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
2090 }
2091
2092 static const struct amdgpu_vm_pte_funcs sdma_v4_0_vm_pte_funcs = {
2093         .copy_pte_num_dw = 7,
2094         .copy_pte = sdma_v4_0_vm_copy_pte,
2095
2096         .write_pte = sdma_v4_0_vm_write_pte,
2097         .set_pte_pde = sdma_v4_0_vm_set_pte_pde,
2098 };
2099
2100 static void sdma_v4_0_set_vm_pte_funcs(struct amdgpu_device *adev)
2101 {
2102         struct drm_gpu_scheduler *sched;
2103         unsigned i;
2104
2105         adev->vm_manager.vm_pte_funcs = &sdma_v4_0_vm_pte_funcs;
2106         for (i = 0; i < adev->sdma.num_instances; i++) {
2107                 if (adev->sdma.has_page_queue)
2108                         sched = &adev->sdma.instance[i].page.sched;
2109                 else
2110                         sched = &adev->sdma.instance[i].ring.sched;
2111                 adev->vm_manager.vm_pte_rqs[i] =
2112                         &sched->sched_rq[DRM_SCHED_PRIORITY_KERNEL];
2113         }
2114         adev->vm_manager.vm_pte_num_rqs = adev->sdma.num_instances;
2115 }
2116
2117 const struct amdgpu_ip_block_version sdma_v4_0_ip_block = {
2118         .type = AMD_IP_BLOCK_TYPE_SDMA,
2119         .major = 4,
2120         .minor = 0,
2121         .rev = 0,
2122         .funcs = &sdma_v4_0_ip_funcs,
2123 };