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/delay.h> 25 #include <linux/firmware.h> 26 #include <linux/module.h> 27 #include <linux/pci.h> 28 29 #include "amdgpu.h" 30 #include "amdgpu_ucode.h" 31 #include "amdgpu_trace.h" 32 33 #include "sdma0/sdma0_4_2_offset.h" 34 #include "sdma0/sdma0_4_2_sh_mask.h" 35 #include "sdma1/sdma1_4_2_offset.h" 36 #include "sdma1/sdma1_4_2_sh_mask.h" 37 #include "sdma2/sdma2_4_2_2_offset.h" 38 #include "sdma2/sdma2_4_2_2_sh_mask.h" 39 #include "sdma3/sdma3_4_2_2_offset.h" 40 #include "sdma3/sdma3_4_2_2_sh_mask.h" 41 #include "sdma4/sdma4_4_2_2_offset.h" 42 #include "sdma4/sdma4_4_2_2_sh_mask.h" 43 #include "sdma5/sdma5_4_2_2_offset.h" 44 #include "sdma5/sdma5_4_2_2_sh_mask.h" 45 #include "sdma6/sdma6_4_2_2_offset.h" 46 #include "sdma6/sdma6_4_2_2_sh_mask.h" 47 #include "sdma7/sdma7_4_2_2_offset.h" 48 #include "sdma7/sdma7_4_2_2_sh_mask.h" 49 #include "sdma0/sdma0_4_1_default.h" 50 51 #include "soc15_common.h" 52 #include "soc15.h" 53 #include "vega10_sdma_pkt_open.h" 54 55 #include "ivsrcid/sdma0/irqsrcs_sdma0_4_0.h" 56 #include "ivsrcid/sdma1/irqsrcs_sdma1_4_0.h" 57 58 #include "amdgpu_ras.h" 59 #include "sdma_v4_4.h" 60 61 MODULE_FIRMWARE("amdgpu/vega10_sdma.bin"); 62 MODULE_FIRMWARE("amdgpu/vega10_sdma1.bin"); 63 MODULE_FIRMWARE("amdgpu/vega12_sdma.bin"); 64 MODULE_FIRMWARE("amdgpu/vega12_sdma1.bin"); 65 MODULE_FIRMWARE("amdgpu/vega20_sdma.bin"); 66 MODULE_FIRMWARE("amdgpu/vega20_sdma1.bin"); 67 MODULE_FIRMWARE("amdgpu/raven_sdma.bin"); 68 MODULE_FIRMWARE("amdgpu/picasso_sdma.bin"); 69 MODULE_FIRMWARE("amdgpu/raven2_sdma.bin"); 70 MODULE_FIRMWARE("amdgpu/arcturus_sdma.bin"); 71 MODULE_FIRMWARE("amdgpu/renoir_sdma.bin"); 72 MODULE_FIRMWARE("amdgpu/green_sardine_sdma.bin"); 73 MODULE_FIRMWARE("amdgpu/aldebaran_sdma.bin"); 74 75 #define SDMA0_POWER_CNTL__ON_OFF_CONDITION_HOLD_TIME_MASK 0x000000F8L 76 #define SDMA0_POWER_CNTL__ON_OFF_STATUS_DURATION_TIME_MASK 0xFC000000L 77 78 #define WREG32_SDMA(instance, offset, value) \ 79 WREG32(sdma_v4_0_get_reg_offset(adev, (instance), (offset)), value) 80 #define RREG32_SDMA(instance, offset) \ 81 RREG32(sdma_v4_0_get_reg_offset(adev, (instance), (offset))) 82 83 static void sdma_v4_0_set_ring_funcs(struct amdgpu_device *adev); 84 static void sdma_v4_0_set_buffer_funcs(struct amdgpu_device *adev); 85 static void sdma_v4_0_set_vm_pte_funcs(struct amdgpu_device *adev); 86 static void sdma_v4_0_set_irq_funcs(struct amdgpu_device *adev); 87 static void sdma_v4_0_set_ras_funcs(struct amdgpu_device *adev); 88 89 static const struct soc15_reg_golden golden_settings_sdma_4[] = { 90 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_CHICKEN_BITS, 0xfe931f07, 0x02831d07), 91 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_CLK_CTRL, 0xff000ff0, 0x3f000100), 92 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GFX_IB_CNTL, 0x800f0100, 0x00000100), 93 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 94 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_PAGE_IB_CNTL, 0x800f0100, 0x00000100), 95 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_PAGE_RB_WPTR_POLL_CNTL, 0x0000fff0, 0x00403000), 96 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_POWER_CNTL, 0x003ff006, 0x0003c000), 97 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC0_IB_CNTL, 0x800f0100, 0x00000100), 98 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC0_RB_WPTR_POLL_CNTL, 0x0000fff0, 0x00403000), 99 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC1_IB_CNTL, 0x800f0100, 0x00000100), 100 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC1_RB_WPTR_POLL_CNTL, 0x0000fff0, 0x00403000), 101 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_UTCL1_PAGE, 0x000003ff, 0x000003c0), 102 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_UTCL1_WATERMK, 0xfc000000, 0x00000000), 103 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_CLK_CTRL, 0xffffffff, 0x3f000100), 104 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GFX_IB_CNTL, 0x800f0100, 0x00000100), 105 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GFX_RB_WPTR_POLL_CNTL, 0x0000fff0, 0x00403000), 106 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_PAGE_IB_CNTL, 0x800f0100, 0x00000100), 107 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_PAGE_RB_WPTR_POLL_CNTL, 0x0000fff0, 0x00403000), 108 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_POWER_CNTL, 0x003ff000, 0x0003c000), 109 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC0_IB_CNTL, 0x800f0100, 0x00000100), 110 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC0_RB_WPTR_POLL_CNTL, 0x0000fff0, 0x00403000), 111 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC1_IB_CNTL, 0x800f0100, 0x00000100), 112 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC1_RB_WPTR_POLL_CNTL, 0x0000fff0, 0x00403000), 113 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_UTCL1_PAGE, 0x000003ff, 0x000003c0), 114 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_UTCL1_WATERMK, 0xfc000000, 0x00000000) 115 }; 116 117 static const struct soc15_reg_golden golden_settings_sdma_vg10[] = { 118 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG, 0x0018773f, 0x00104002), 119 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG_READ, 0x0018773f, 0x00104002), 120 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_UTCL1_TIMEOUT, 0xffffffff, 0x00010001), 121 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_CHICKEN_BITS, 0xfe931f07, 0x02831d07), 122 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GB_ADDR_CONFIG, 0x0018773f, 0x00104002), 123 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GB_ADDR_CONFIG_READ, 0x0018773f, 0x00104002), 124 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_UTCL1_TIMEOUT, 0xffffffff, 0x00010001), 125 }; 126 127 static const struct soc15_reg_golden golden_settings_sdma_vg12[] = { 128 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG, 0x0018773f, 0x00104001), 129 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG_READ, 0x0018773f, 0x00104001), 130 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_UTCL1_TIMEOUT, 0xffffffff, 0x00010001), 131 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_CHICKEN_BITS, 0xfe931f07, 0x02831d07), 132 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GB_ADDR_CONFIG, 0x0018773f, 0x00104001), 133 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GB_ADDR_CONFIG_READ, 0x0018773f, 0x00104001), 134 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_UTCL1_TIMEOUT, 0xffffffff, 0x00010001), 135 }; 136 137 static const struct soc15_reg_golden golden_settings_sdma_4_1[] = { 138 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_CHICKEN_BITS, 0xfe931f07, 0x02831d07), 139 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_CLK_CTRL, 0xffffffff, 0x3f000100), 140 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GFX_IB_CNTL, 0x800f0111, 0x00000100), 141 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 142 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_POWER_CNTL, 0xfc3fffff, 0x40000051), 143 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC0_IB_CNTL, 0x800f0111, 0x00000100), 144 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 145 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC1_IB_CNTL, 0x800f0111, 0x00000100), 146 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 147 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_UTCL1_PAGE, 0x000003ff, 0x000003e0), 148 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_UTCL1_WATERMK, 0xfc000000, 0x00000000) 149 }; 150 151 static const struct soc15_reg_golden golden_settings_sdma0_4_2_init[] = { 152 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff0, 0x00403000), 153 }; 154 155 static const struct soc15_reg_golden golden_settings_sdma0_4_2[] = 156 { 157 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_CHICKEN_BITS, 0xfe931f07, 0x02831f07), 158 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_CLK_CTRL, 0xffffffff, 0x3f000100), 159 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG, 0x0000773f, 0x00004002), 160 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG_READ, 0x0000773f, 0x00004002), 161 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GFX_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 162 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 163 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_PAGE_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 164 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_PAGE_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 165 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RD_BURST_CNTL, 0x0000000f, 0x00000003), 166 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC0_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 167 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff0, 0x00403000), 168 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC1_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 169 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 170 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC2_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 171 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC2_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 172 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC3_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 173 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 174 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC4_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 175 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC4_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 176 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC5_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 177 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC5_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 178 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC6_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 179 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC6_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 180 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC7_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 181 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC7_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 182 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_UTCL1_PAGE, 0x000003ff, 0x000003c0), 183 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_UTCL1_TIMEOUT, 0xffffffff, 0x00010001), 184 }; 185 186 static const struct soc15_reg_golden golden_settings_sdma1_4_2[] = { 187 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_CHICKEN_BITS, 0xfe931f07, 0x02831f07), 188 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_CLK_CTRL, 0xffffffff, 0x3f000100), 189 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GB_ADDR_CONFIG, 0x0000773f, 0x00004002), 190 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GB_ADDR_CONFIG_READ, 0x0000773f, 0x00004002), 191 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GFX_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 192 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 193 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_PAGE_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 194 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_PAGE_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 195 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RD_BURST_CNTL, 0x0000000f, 0x00000003), 196 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC0_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 197 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff0, 0x00403000), 198 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC1_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 199 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 200 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC2_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 201 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC2_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 202 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC3_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 203 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 204 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC4_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 205 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC4_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 206 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC5_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 207 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC5_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 208 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC6_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 209 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC6_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 210 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC7_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 211 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC7_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 212 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_UTCL1_PAGE, 0x000003ff, 0x000003c0), 213 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_UTCL1_TIMEOUT, 0xffffffff, 0x00010001), 214 }; 215 216 static const struct soc15_reg_golden golden_settings_sdma_rv1[] = 217 { 218 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG, 0x0018773f, 0x00000002), 219 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG_READ, 0x0018773f, 0x00000002) 220 }; 221 222 static const struct soc15_reg_golden golden_settings_sdma_rv2[] = 223 { 224 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG, 0x0018773f, 0x00003001), 225 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG_READ, 0x0018773f, 0x00003001) 226 }; 227 228 static const struct soc15_reg_golden golden_settings_sdma_arct[] = 229 { 230 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_CHICKEN_BITS, 0xfe931f07, 0x02831f07), 231 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG, 0x0000773f, 0x00004002), 232 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG_READ, 0x0000773f, 0x00004002), 233 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_UTCL1_TIMEOUT, 0xffffffff, 0x00010001), 234 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_CHICKEN_BITS, 0xfe931f07, 0x02831f07), 235 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GB_ADDR_CONFIG, 0x0000773f, 0x00004002), 236 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GB_ADDR_CONFIG_READ, 0x0000773f, 0x00004002), 237 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_UTCL1_TIMEOUT, 0xffffffff, 0x00010001), 238 SOC15_REG_GOLDEN_VALUE(SDMA2, 0, mmSDMA2_CHICKEN_BITS, 0xfe931f07, 0x02831f07), 239 SOC15_REG_GOLDEN_VALUE(SDMA2, 0, mmSDMA2_GB_ADDR_CONFIG, 0x0000773f, 0x00004002), 240 SOC15_REG_GOLDEN_VALUE(SDMA2, 0, mmSDMA2_GB_ADDR_CONFIG_READ, 0x0000773f, 0x00004002), 241 SOC15_REG_GOLDEN_VALUE(SDMA2, 0, mmSDMA2_UTCL1_TIMEOUT, 0xffffffff, 0x00010001), 242 SOC15_REG_GOLDEN_VALUE(SDMA3, 0, mmSDMA3_CHICKEN_BITS, 0xfe931f07, 0x02831f07), 243 SOC15_REG_GOLDEN_VALUE(SDMA3, 0, mmSDMA3_GB_ADDR_CONFIG, 0x0000773f, 0x00004002), 244 SOC15_REG_GOLDEN_VALUE(SDMA3, 0, mmSDMA3_GB_ADDR_CONFIG_READ, 0x0000773f, 0x00004002), 245 SOC15_REG_GOLDEN_VALUE(SDMA3, 0, mmSDMA3_UTCL1_TIMEOUT, 0xffffffff, 0x00010001), 246 SOC15_REG_GOLDEN_VALUE(SDMA4, 0, mmSDMA4_CHICKEN_BITS, 0xfe931f07, 0x02831f07), 247 SOC15_REG_GOLDEN_VALUE(SDMA4, 0, mmSDMA4_GB_ADDR_CONFIG, 0x0000773f, 0x00004002), 248 SOC15_REG_GOLDEN_VALUE(SDMA4, 0, mmSDMA4_GB_ADDR_CONFIG_READ, 0x0000773f, 0x00004002), 249 SOC15_REG_GOLDEN_VALUE(SDMA4, 0, mmSDMA4_UTCL1_TIMEOUT, 0xffffffff, 0x00010001), 250 SOC15_REG_GOLDEN_VALUE(SDMA5, 0, mmSDMA5_CHICKEN_BITS, 0xfe931f07, 0x02831f07), 251 SOC15_REG_GOLDEN_VALUE(SDMA5, 0, mmSDMA5_GB_ADDR_CONFIG, 0x0000773f, 0x00004002), 252 SOC15_REG_GOLDEN_VALUE(SDMA5, 0, mmSDMA5_GB_ADDR_CONFIG_READ, 0x0000773f, 0x00004002), 253 SOC15_REG_GOLDEN_VALUE(SDMA5, 0, mmSDMA5_UTCL1_TIMEOUT, 0xffffffff, 0x00010001), 254 SOC15_REG_GOLDEN_VALUE(SDMA6, 0, mmSDMA6_CHICKEN_BITS, 0xfe931f07, 0x02831f07), 255 SOC15_REG_GOLDEN_VALUE(SDMA6, 0, mmSDMA6_GB_ADDR_CONFIG, 0x0000773f, 0x00004002), 256 SOC15_REG_GOLDEN_VALUE(SDMA6, 0, mmSDMA6_GB_ADDR_CONFIG_READ, 0x0000773f, 0x00004002), 257 SOC15_REG_GOLDEN_VALUE(SDMA6, 0, mmSDMA6_UTCL1_TIMEOUT, 0xffffffff, 0x00010001), 258 SOC15_REG_GOLDEN_VALUE(SDMA7, 0, mmSDMA7_CHICKEN_BITS, 0xfe931f07, 0x02831f07), 259 SOC15_REG_GOLDEN_VALUE(SDMA7, 0, mmSDMA7_GB_ADDR_CONFIG, 0x0000773f, 0x00004002), 260 SOC15_REG_GOLDEN_VALUE(SDMA7, 0, mmSDMA7_GB_ADDR_CONFIG_READ, 0x0000773f, 0x00004002), 261 SOC15_REG_GOLDEN_VALUE(SDMA7, 0, mmSDMA7_UTCL1_TIMEOUT, 0xffffffff, 0x00010001) 262 }; 263 264 static const struct soc15_reg_golden golden_settings_sdma_aldebaran[] = { 265 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG, 0x0018773f, 0x00104002), 266 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG_READ, 0x0018773f, 0x00104002), 267 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_UTCL1_TIMEOUT, 0xffffffff, 0x00010001), 268 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GB_ADDR_CONFIG, 0x0018773f, 0x00104002), 269 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GB_ADDR_CONFIG_READ, 0x0018773f, 0x00104002), 270 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_UTCL1_TIMEOUT, 0xffffffff, 0x00010001), 271 SOC15_REG_GOLDEN_VALUE(SDMA2, 0, mmSDMA2_GB_ADDR_CONFIG, 0x0018773f, 0x00104002), 272 SOC15_REG_GOLDEN_VALUE(SDMA2, 0, mmSDMA2_GB_ADDR_CONFIG_READ, 0x0018773f, 0x00104002), 273 SOC15_REG_GOLDEN_VALUE(SDMA3, 0, mmSDMA2_UTCL1_TIMEOUT, 0xffffffff, 0x00010001), 274 SOC15_REG_GOLDEN_VALUE(SDMA3, 0, mmSDMA3_GB_ADDR_CONFIG, 0x0018773f, 0x00104002), 275 SOC15_REG_GOLDEN_VALUE(SDMA3, 0, mmSDMA3_GB_ADDR_CONFIG_READ, 0x0018773f, 0x00104002), 276 SOC15_REG_GOLDEN_VALUE(SDMA3, 0, mmSDMA3_UTCL1_TIMEOUT, 0xffffffff, 0x00010001), 277 SOC15_REG_GOLDEN_VALUE(SDMA4, 0, mmSDMA4_GB_ADDR_CONFIG, 0x0018773f, 0x00104002), 278 SOC15_REG_GOLDEN_VALUE(SDMA4, 0, mmSDMA4_GB_ADDR_CONFIG_READ, 0x0018773f, 0x00104002), 279 SOC15_REG_GOLDEN_VALUE(SDMA4, 0, mmSDMA4_UTCL1_TIMEOUT, 0xffffffff, 0x00010001), 280 }; 281 282 static const struct soc15_reg_golden golden_settings_sdma_4_3[] = { 283 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_CHICKEN_BITS, 0xfe931f07, 0x02831f07), 284 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_CLK_CTRL, 0xffffffff, 0x3f000100), 285 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG, 0x0018773f, 0x00000002), 286 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG_READ, 0x0018773f, 0x00000002), 287 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 288 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_POWER_CNTL, 0x003fff07, 0x40000051), 289 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 290 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 291 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_UTCL1_PAGE, 0x000003ff, 0x000003e0), 292 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_UTCL1_WATERMK, 0xfc000000, 0x03fbe1fe) 293 }; 294 295 static const struct soc15_ras_field_entry sdma_v4_0_ras_fields[] = { 296 { "SDMA_UCODE_BUF_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 297 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_UCODE_BUF_SED), 298 0, 0, 299 }, 300 { "SDMA_RB_CMD_BUF_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 301 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_RB_CMD_BUF_SED), 302 0, 0, 303 }, 304 { "SDMA_IB_CMD_BUF_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 305 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_IB_CMD_BUF_SED), 306 0, 0, 307 }, 308 { "SDMA_UTCL1_RD_FIFO_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 309 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_UTCL1_RD_FIFO_SED), 310 0, 0, 311 }, 312 { "SDMA_UTCL1_RDBST_FIFO_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 313 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_UTCL1_RDBST_FIFO_SED), 314 0, 0, 315 }, 316 { "SDMA_DATA_LUT_FIFO_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 317 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_DATA_LUT_FIFO_SED), 318 0, 0, 319 }, 320 { "SDMA_MBANK_DATA_BUF0_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 321 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_MBANK_DATA_BUF0_SED), 322 0, 0, 323 }, 324 { "SDMA_MBANK_DATA_BUF1_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 325 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_MBANK_DATA_BUF1_SED), 326 0, 0, 327 }, 328 { "SDMA_MBANK_DATA_BUF2_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 329 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_MBANK_DATA_BUF2_SED), 330 0, 0, 331 }, 332 { "SDMA_MBANK_DATA_BUF3_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 333 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_MBANK_DATA_BUF3_SED), 334 0, 0, 335 }, 336 { "SDMA_MBANK_DATA_BUF4_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 337 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_MBANK_DATA_BUF4_SED), 338 0, 0, 339 }, 340 { "SDMA_MBANK_DATA_BUF5_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 341 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_MBANK_DATA_BUF5_SED), 342 0, 0, 343 }, 344 { "SDMA_MBANK_DATA_BUF6_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 345 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_MBANK_DATA_BUF6_SED), 346 0, 0, 347 }, 348 { "SDMA_MBANK_DATA_BUF7_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 349 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_MBANK_DATA_BUF7_SED), 350 0, 0, 351 }, 352 { "SDMA_MBANK_DATA_BUF8_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 353 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_MBANK_DATA_BUF8_SED), 354 0, 0, 355 }, 356 { "SDMA_MBANK_DATA_BUF9_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 357 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_MBANK_DATA_BUF9_SED), 358 0, 0, 359 }, 360 { "SDMA_MBANK_DATA_BUF10_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 361 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_MBANK_DATA_BUF10_SED), 362 0, 0, 363 }, 364 { "SDMA_MBANK_DATA_BUF11_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 365 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_MBANK_DATA_BUF11_SED), 366 0, 0, 367 }, 368 { "SDMA_MBANK_DATA_BUF12_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 369 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_MBANK_DATA_BUF12_SED), 370 0, 0, 371 }, 372 { "SDMA_MBANK_DATA_BUF13_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 373 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_MBANK_DATA_BUF13_SED), 374 0, 0, 375 }, 376 { "SDMA_MBANK_DATA_BUF14_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 377 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_MBANK_DATA_BUF14_SED), 378 0, 0, 379 }, 380 { "SDMA_MBANK_DATA_BUF15_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 381 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_MBANK_DATA_BUF15_SED), 382 0, 0, 383 }, 384 { "SDMA_SPLIT_DAT_BUF_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 385 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_SPLIT_DAT_BUF_SED), 386 0, 0, 387 }, 388 { "SDMA_MC_WR_ADDR_FIFO_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 389 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_MC_WR_ADDR_FIFO_SED), 390 0, 0, 391 }, 392 }; 393 394 static u32 sdma_v4_0_get_reg_offset(struct amdgpu_device *adev, 395 u32 instance, u32 offset) 396 { 397 switch (instance) { 398 case 0: 399 return (adev->reg_offset[SDMA0_HWIP][0][0] + offset); 400 case 1: 401 return (adev->reg_offset[SDMA1_HWIP][0][0] + offset); 402 case 2: 403 return (adev->reg_offset[SDMA2_HWIP][0][1] + offset); 404 case 3: 405 return (adev->reg_offset[SDMA3_HWIP][0][1] + offset); 406 case 4: 407 return (adev->reg_offset[SDMA4_HWIP][0][1] + offset); 408 case 5: 409 return (adev->reg_offset[SDMA5_HWIP][0][1] + offset); 410 case 6: 411 return (adev->reg_offset[SDMA6_HWIP][0][1] + offset); 412 case 7: 413 return (adev->reg_offset[SDMA7_HWIP][0][1] + offset); 414 default: 415 break; 416 } 417 return 0; 418 } 419 420 static unsigned sdma_v4_0_seq_to_irq_id(int seq_num) 421 { 422 switch (seq_num) { 423 case 0: 424 return SOC15_IH_CLIENTID_SDMA0; 425 case 1: 426 return SOC15_IH_CLIENTID_SDMA1; 427 case 2: 428 return SOC15_IH_CLIENTID_SDMA2; 429 case 3: 430 return SOC15_IH_CLIENTID_SDMA3; 431 case 4: 432 return SOC15_IH_CLIENTID_SDMA4; 433 case 5: 434 return SOC15_IH_CLIENTID_SDMA5; 435 case 6: 436 return SOC15_IH_CLIENTID_SDMA6; 437 case 7: 438 return SOC15_IH_CLIENTID_SDMA7; 439 default: 440 break; 441 } 442 return -EINVAL; 443 } 444 445 static int sdma_v4_0_irq_id_to_seq(unsigned client_id) 446 { 447 switch (client_id) { 448 case SOC15_IH_CLIENTID_SDMA0: 449 return 0; 450 case SOC15_IH_CLIENTID_SDMA1: 451 return 1; 452 case SOC15_IH_CLIENTID_SDMA2: 453 return 2; 454 case SOC15_IH_CLIENTID_SDMA3: 455 return 3; 456 case SOC15_IH_CLIENTID_SDMA4: 457 return 4; 458 case SOC15_IH_CLIENTID_SDMA5: 459 return 5; 460 case SOC15_IH_CLIENTID_SDMA6: 461 return 6; 462 case SOC15_IH_CLIENTID_SDMA7: 463 return 7; 464 default: 465 break; 466 } 467 return -EINVAL; 468 } 469 470 static void sdma_v4_0_init_golden_registers(struct amdgpu_device *adev) 471 { 472 switch (amdgpu_ip_version(adev, SDMA0_HWIP, 0)) { 473 case IP_VERSION(4, 0, 0): 474 soc15_program_register_sequence(adev, 475 golden_settings_sdma_4, 476 ARRAY_SIZE(golden_settings_sdma_4)); 477 soc15_program_register_sequence(adev, 478 golden_settings_sdma_vg10, 479 ARRAY_SIZE(golden_settings_sdma_vg10)); 480 break; 481 case IP_VERSION(4, 0, 1): 482 soc15_program_register_sequence(adev, 483 golden_settings_sdma_4, 484 ARRAY_SIZE(golden_settings_sdma_4)); 485 soc15_program_register_sequence(adev, 486 golden_settings_sdma_vg12, 487 ARRAY_SIZE(golden_settings_sdma_vg12)); 488 break; 489 case IP_VERSION(4, 2, 0): 490 soc15_program_register_sequence(adev, 491 golden_settings_sdma0_4_2_init, 492 ARRAY_SIZE(golden_settings_sdma0_4_2_init)); 493 soc15_program_register_sequence(adev, 494 golden_settings_sdma0_4_2, 495 ARRAY_SIZE(golden_settings_sdma0_4_2)); 496 soc15_program_register_sequence(adev, 497 golden_settings_sdma1_4_2, 498 ARRAY_SIZE(golden_settings_sdma1_4_2)); 499 break; 500 case IP_VERSION(4, 2, 2): 501 soc15_program_register_sequence(adev, 502 golden_settings_sdma_arct, 503 ARRAY_SIZE(golden_settings_sdma_arct)); 504 break; 505 case IP_VERSION(4, 4, 0): 506 soc15_program_register_sequence(adev, 507 golden_settings_sdma_aldebaran, 508 ARRAY_SIZE(golden_settings_sdma_aldebaran)); 509 break; 510 case IP_VERSION(4, 1, 0): 511 case IP_VERSION(4, 1, 1): 512 soc15_program_register_sequence(adev, 513 golden_settings_sdma_4_1, 514 ARRAY_SIZE(golden_settings_sdma_4_1)); 515 if (adev->apu_flags & AMD_APU_IS_RAVEN2) 516 soc15_program_register_sequence(adev, 517 golden_settings_sdma_rv2, 518 ARRAY_SIZE(golden_settings_sdma_rv2)); 519 else 520 soc15_program_register_sequence(adev, 521 golden_settings_sdma_rv1, 522 ARRAY_SIZE(golden_settings_sdma_rv1)); 523 break; 524 case IP_VERSION(4, 1, 2): 525 soc15_program_register_sequence(adev, 526 golden_settings_sdma_4_3, 527 ARRAY_SIZE(golden_settings_sdma_4_3)); 528 break; 529 default: 530 break; 531 } 532 } 533 534 static void sdma_v4_0_setup_ulv(struct amdgpu_device *adev) 535 { 536 int i; 537 538 /* 539 * The only chips with SDMAv4 and ULV are VG10 and VG20. 540 * Server SKUs take a different hysteresis setting from other SKUs. 541 */ 542 switch (amdgpu_ip_version(adev, SDMA0_HWIP, 0)) { 543 case IP_VERSION(4, 0, 0): 544 if (adev->pdev->device == 0x6860) 545 break; 546 return; 547 case IP_VERSION(4, 2, 0): 548 if (adev->pdev->device == 0x66a1) 549 break; 550 return; 551 default: 552 return; 553 } 554 555 for (i = 0; i < adev->sdma.num_instances; i++) { 556 uint32_t temp; 557 558 temp = RREG32_SDMA(i, mmSDMA0_ULV_CNTL); 559 temp = REG_SET_FIELD(temp, SDMA0_ULV_CNTL, HYSTERESIS, 0x0); 560 WREG32_SDMA(i, mmSDMA0_ULV_CNTL, temp); 561 } 562 } 563 564 /** 565 * sdma_v4_0_init_microcode - load ucode images from disk 566 * 567 * @adev: amdgpu_device pointer 568 * 569 * Use the firmware interface to load the ucode images into 570 * the driver (not loaded into hw). 571 * Returns 0 on success, error on failure. 572 */ 573 574 // emulation only, won't work on real chip 575 // vega10 real chip need to use PSP to load firmware 576 static int sdma_v4_0_init_microcode(struct amdgpu_device *adev) 577 { 578 int ret, i; 579 580 for (i = 0; i < adev->sdma.num_instances; i++) { 581 if (amdgpu_ip_version(adev, SDMA0_HWIP, 0) == 582 IP_VERSION(4, 2, 2) || 583 amdgpu_ip_version(adev, SDMA0_HWIP, 0) == 584 IP_VERSION(4, 4, 0)) { 585 /* Acturus & Aldebaran will leverage the same FW memory 586 for every SDMA instance */ 587 ret = amdgpu_sdma_init_microcode(adev, 0, true); 588 break; 589 } else { 590 ret = amdgpu_sdma_init_microcode(adev, i, false); 591 if (ret) 592 return ret; 593 } 594 } 595 596 return ret; 597 } 598 599 /** 600 * sdma_v4_0_ring_get_rptr - get the current read pointer 601 * 602 * @ring: amdgpu ring pointer 603 * 604 * Get the current rptr from the hardware (VEGA10+). 605 */ 606 static uint64_t sdma_v4_0_ring_get_rptr(struct amdgpu_ring *ring) 607 { 608 u64 *rptr; 609 610 /* XXX check if swapping is necessary on BE */ 611 rptr = ((u64 *)ring->rptr_cpu_addr); 612 613 DRM_DEBUG("rptr before shift == 0x%016llx\n", *rptr); 614 return ((*rptr) >> 2); 615 } 616 617 /** 618 * sdma_v4_0_ring_get_wptr - get the current write pointer 619 * 620 * @ring: amdgpu ring pointer 621 * 622 * Get the current wptr from the hardware (VEGA10+). 623 */ 624 static uint64_t sdma_v4_0_ring_get_wptr(struct amdgpu_ring *ring) 625 { 626 struct amdgpu_device *adev = ring->adev; 627 u64 wptr; 628 629 if (ring->use_doorbell) { 630 /* XXX check if swapping is necessary on BE */ 631 wptr = READ_ONCE(*((u64 *)ring->wptr_cpu_addr)); 632 DRM_DEBUG("wptr/doorbell before shift == 0x%016llx\n", wptr); 633 } else { 634 wptr = RREG32_SDMA(ring->me, mmSDMA0_GFX_RB_WPTR_HI); 635 wptr = wptr << 32; 636 wptr |= RREG32_SDMA(ring->me, mmSDMA0_GFX_RB_WPTR); 637 DRM_DEBUG("wptr before shift [%i] wptr == 0x%016llx\n", 638 ring->me, wptr); 639 } 640 641 return wptr >> 2; 642 } 643 644 /** 645 * sdma_v4_0_ring_set_wptr - commit the write pointer 646 * 647 * @ring: amdgpu ring pointer 648 * 649 * Write the wptr back to the hardware (VEGA10+). 650 */ 651 static void sdma_v4_0_ring_set_wptr(struct amdgpu_ring *ring) 652 { 653 struct amdgpu_device *adev = ring->adev; 654 655 DRM_DEBUG("Setting write pointer\n"); 656 if (ring->use_doorbell) { 657 u64 *wb = (u64 *)ring->wptr_cpu_addr; 658 659 DRM_DEBUG("Using doorbell -- " 660 "wptr_offs == 0x%08x " 661 "lower_32_bits(ring->wptr << 2) == 0x%08x " 662 "upper_32_bits(ring->wptr << 2) == 0x%08x\n", 663 ring->wptr_offs, 664 lower_32_bits(ring->wptr << 2), 665 upper_32_bits(ring->wptr << 2)); 666 /* XXX check if swapping is necessary on BE */ 667 WRITE_ONCE(*wb, (ring->wptr << 2)); 668 DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n", 669 ring->doorbell_index, ring->wptr << 2); 670 WDOORBELL64(ring->doorbell_index, ring->wptr << 2); 671 } else { 672 DRM_DEBUG("Not using doorbell -- " 673 "mmSDMA%i_GFX_RB_WPTR == 0x%08x " 674 "mmSDMA%i_GFX_RB_WPTR_HI == 0x%08x\n", 675 ring->me, 676 lower_32_bits(ring->wptr << 2), 677 ring->me, 678 upper_32_bits(ring->wptr << 2)); 679 WREG32_SDMA(ring->me, mmSDMA0_GFX_RB_WPTR, 680 lower_32_bits(ring->wptr << 2)); 681 WREG32_SDMA(ring->me, mmSDMA0_GFX_RB_WPTR_HI, 682 upper_32_bits(ring->wptr << 2)); 683 } 684 } 685 686 /** 687 * sdma_v4_0_page_ring_get_wptr - get the current write pointer 688 * 689 * @ring: amdgpu ring pointer 690 * 691 * Get the current wptr from the hardware (VEGA10+). 692 */ 693 static uint64_t sdma_v4_0_page_ring_get_wptr(struct amdgpu_ring *ring) 694 { 695 struct amdgpu_device *adev = ring->adev; 696 u64 wptr; 697 698 if (ring->use_doorbell) { 699 /* XXX check if swapping is necessary on BE */ 700 wptr = READ_ONCE(*((u64 *)ring->wptr_cpu_addr)); 701 } else { 702 wptr = RREG32_SDMA(ring->me, mmSDMA0_PAGE_RB_WPTR_HI); 703 wptr = wptr << 32; 704 wptr |= RREG32_SDMA(ring->me, mmSDMA0_PAGE_RB_WPTR); 705 } 706 707 return wptr >> 2; 708 } 709 710 /** 711 * sdma_v4_0_page_ring_set_wptr - commit the write pointer 712 * 713 * @ring: amdgpu ring pointer 714 * 715 * Write the wptr back to the hardware (VEGA10+). 716 */ 717 static void sdma_v4_0_page_ring_set_wptr(struct amdgpu_ring *ring) 718 { 719 struct amdgpu_device *adev = ring->adev; 720 721 if (ring->use_doorbell) { 722 u64 *wb = (u64 *)ring->wptr_cpu_addr; 723 724 /* XXX check if swapping is necessary on BE */ 725 WRITE_ONCE(*wb, (ring->wptr << 2)); 726 WDOORBELL64(ring->doorbell_index, ring->wptr << 2); 727 } else { 728 uint64_t wptr = ring->wptr << 2; 729 730 WREG32_SDMA(ring->me, mmSDMA0_PAGE_RB_WPTR, 731 lower_32_bits(wptr)); 732 WREG32_SDMA(ring->me, mmSDMA0_PAGE_RB_WPTR_HI, 733 upper_32_bits(wptr)); 734 } 735 } 736 737 static void sdma_v4_0_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count) 738 { 739 struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring); 740 int i; 741 742 for (i = 0; i < count; i++) 743 if (sdma && sdma->burst_nop && (i == 0)) 744 amdgpu_ring_write(ring, ring->funcs->nop | 745 SDMA_PKT_NOP_HEADER_COUNT(count - 1)); 746 else 747 amdgpu_ring_write(ring, ring->funcs->nop); 748 } 749 750 /** 751 * sdma_v4_0_ring_emit_ib - Schedule an IB on the DMA engine 752 * 753 * @ring: amdgpu ring pointer 754 * @job: job to retrieve vmid from 755 * @ib: IB object to schedule 756 * @flags: unused 757 * 758 * Schedule an IB in the DMA ring (VEGA10). 759 */ 760 static void sdma_v4_0_ring_emit_ib(struct amdgpu_ring *ring, 761 struct amdgpu_job *job, 762 struct amdgpu_ib *ib, 763 uint32_t flags) 764 { 765 unsigned vmid = AMDGPU_JOB_GET_VMID(job); 766 767 /* IB packet must end on a 8 DW boundary */ 768 sdma_v4_0_ring_insert_nop(ring, (2 - lower_32_bits(ring->wptr)) & 7); 769 770 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) | 771 SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf)); 772 /* base must be 32 byte aligned */ 773 amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0); 774 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr)); 775 amdgpu_ring_write(ring, ib->length_dw); 776 amdgpu_ring_write(ring, 0); 777 amdgpu_ring_write(ring, 0); 778 779 } 780 781 static void sdma_v4_0_wait_reg_mem(struct amdgpu_ring *ring, 782 int mem_space, int hdp, 783 uint32_t addr0, uint32_t addr1, 784 uint32_t ref, uint32_t mask, 785 uint32_t inv) 786 { 787 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) | 788 SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(hdp) | 789 SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(mem_space) | 790 SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */ 791 if (mem_space) { 792 /* memory */ 793 amdgpu_ring_write(ring, addr0); 794 amdgpu_ring_write(ring, addr1); 795 } else { 796 /* registers */ 797 amdgpu_ring_write(ring, addr0 << 2); 798 amdgpu_ring_write(ring, addr1 << 2); 799 } 800 amdgpu_ring_write(ring, ref); /* reference */ 801 amdgpu_ring_write(ring, mask); /* mask */ 802 amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) | 803 SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(inv)); /* retry count, poll interval */ 804 } 805 806 /** 807 * sdma_v4_0_ring_emit_hdp_flush - emit an hdp flush on the DMA ring 808 * 809 * @ring: amdgpu ring pointer 810 * 811 * Emit an hdp flush packet on the requested DMA ring. 812 */ 813 static void sdma_v4_0_ring_emit_hdp_flush(struct amdgpu_ring *ring) 814 { 815 struct amdgpu_device *adev = ring->adev; 816 u32 ref_and_mask = 0; 817 const struct nbio_hdp_flush_reg *nbio_hf_reg = adev->nbio.hdp_flush_reg; 818 819 ref_and_mask = nbio_hf_reg->ref_and_mask_sdma0 << ring->me; 820 821 sdma_v4_0_wait_reg_mem(ring, 0, 1, 822 adev->nbio.funcs->get_hdp_flush_done_offset(adev), 823 adev->nbio.funcs->get_hdp_flush_req_offset(adev), 824 ref_and_mask, ref_and_mask, 10); 825 } 826 827 /** 828 * sdma_v4_0_ring_emit_fence - emit a fence on the DMA ring 829 * 830 * @ring: amdgpu ring pointer 831 * @addr: address 832 * @seq: sequence number 833 * @flags: fence related flags 834 * 835 * Add a DMA fence packet to the ring to write 836 * the fence seq number and DMA trap packet to generate 837 * an interrupt if needed (VEGA10). 838 */ 839 static void sdma_v4_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq, 840 unsigned flags) 841 { 842 bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT; 843 /* write the fence */ 844 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE)); 845 /* zero in first two bits */ 846 BUG_ON(addr & 0x3); 847 amdgpu_ring_write(ring, lower_32_bits(addr)); 848 amdgpu_ring_write(ring, upper_32_bits(addr)); 849 amdgpu_ring_write(ring, lower_32_bits(seq)); 850 851 /* optionally write high bits as well */ 852 if (write64bit) { 853 addr += 4; 854 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE)); 855 /* zero in first two bits */ 856 BUG_ON(addr & 0x3); 857 amdgpu_ring_write(ring, lower_32_bits(addr)); 858 amdgpu_ring_write(ring, upper_32_bits(addr)); 859 amdgpu_ring_write(ring, upper_32_bits(seq)); 860 } 861 862 /* generate an interrupt */ 863 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP)); 864 amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(0)); 865 } 866 867 868 /** 869 * sdma_v4_0_gfx_enable - enable the gfx async dma engines 870 * 871 * @adev: amdgpu_device pointer 872 * @enable: enable SDMA RB/IB 873 * control the gfx async dma ring buffers (VEGA10). 874 */ 875 static void sdma_v4_0_gfx_enable(struct amdgpu_device *adev, bool enable) 876 { 877 u32 rb_cntl, ib_cntl; 878 int i; 879 880 for (i = 0; i < adev->sdma.num_instances; i++) { 881 rb_cntl = RREG32_SDMA(i, mmSDMA0_GFX_RB_CNTL); 882 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, enable ? 1 : 0); 883 WREG32_SDMA(i, mmSDMA0_GFX_RB_CNTL, rb_cntl); 884 ib_cntl = RREG32_SDMA(i, mmSDMA0_GFX_IB_CNTL); 885 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, enable ? 1 : 0); 886 WREG32_SDMA(i, mmSDMA0_GFX_IB_CNTL, ib_cntl); 887 } 888 } 889 890 /** 891 * sdma_v4_0_rlc_stop - stop the compute async dma engines 892 * 893 * @adev: amdgpu_device pointer 894 * 895 * Stop the compute async dma queues (VEGA10). 896 */ 897 static void sdma_v4_0_rlc_stop(struct amdgpu_device *adev) 898 { 899 /* XXX todo */ 900 } 901 902 /** 903 * sdma_v4_0_page_stop - stop the page async dma engines 904 * 905 * @adev: amdgpu_device pointer 906 * 907 * Stop the page async dma ring buffers (VEGA10). 908 */ 909 static void sdma_v4_0_page_stop(struct amdgpu_device *adev) 910 { 911 u32 rb_cntl, ib_cntl; 912 int i; 913 914 for (i = 0; i < adev->sdma.num_instances; i++) { 915 rb_cntl = RREG32_SDMA(i, mmSDMA0_PAGE_RB_CNTL); 916 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_PAGE_RB_CNTL, 917 RB_ENABLE, 0); 918 WREG32_SDMA(i, mmSDMA0_PAGE_RB_CNTL, rb_cntl); 919 ib_cntl = RREG32_SDMA(i, mmSDMA0_PAGE_IB_CNTL); 920 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_PAGE_IB_CNTL, 921 IB_ENABLE, 0); 922 WREG32_SDMA(i, mmSDMA0_PAGE_IB_CNTL, ib_cntl); 923 } 924 } 925 926 /** 927 * sdma_v4_0_ctx_switch_enable - stop the async dma engines context switch 928 * 929 * @adev: amdgpu_device pointer 930 * @enable: enable/disable the DMA MEs context switch. 931 * 932 * Halt or unhalt the async dma engines context switch (VEGA10). 933 */ 934 static void sdma_v4_0_ctx_switch_enable(struct amdgpu_device *adev, bool enable) 935 { 936 u32 f32_cntl, phase_quantum = 0; 937 int i; 938 939 if (amdgpu_sdma_phase_quantum) { 940 unsigned value = amdgpu_sdma_phase_quantum; 941 unsigned unit = 0; 942 943 while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >> 944 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) { 945 value = (value + 1) >> 1; 946 unit++; 947 } 948 if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >> 949 SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) { 950 value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >> 951 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT); 952 unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >> 953 SDMA0_PHASE0_QUANTUM__UNIT__SHIFT); 954 WARN_ONCE(1, 955 "clamping sdma_phase_quantum to %uK clock cycles\n", 956 value << unit); 957 } 958 phase_quantum = 959 value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT | 960 unit << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT; 961 } 962 963 for (i = 0; i < adev->sdma.num_instances; i++) { 964 f32_cntl = RREG32_SDMA(i, mmSDMA0_CNTL); 965 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL, 966 AUTO_CTXSW_ENABLE, enable ? 1 : 0); 967 if (enable && amdgpu_sdma_phase_quantum) { 968 WREG32_SDMA(i, mmSDMA0_PHASE0_QUANTUM, phase_quantum); 969 WREG32_SDMA(i, mmSDMA0_PHASE1_QUANTUM, phase_quantum); 970 WREG32_SDMA(i, mmSDMA0_PHASE2_QUANTUM, phase_quantum); 971 } 972 WREG32_SDMA(i, mmSDMA0_CNTL, f32_cntl); 973 974 /* 975 * Enable SDMA utilization. Its only supported on 976 * Arcturus for the moment and firmware version 14 977 * and above. 978 */ 979 if (amdgpu_ip_version(adev, SDMA0_HWIP, 0) == 980 IP_VERSION(4, 2, 2) && 981 adev->sdma.instance[i].fw_version >= 14) 982 WREG32_SDMA(i, mmSDMA0_PUB_DUMMY_REG2, enable); 983 /* Extend page fault timeout to avoid interrupt storm */ 984 WREG32_SDMA(i, mmSDMA0_UTCL1_TIMEOUT, 0x00800080); 985 } 986 987 } 988 989 /** 990 * sdma_v4_0_enable - stop the async dma engines 991 * 992 * @adev: amdgpu_device pointer 993 * @enable: enable/disable the DMA MEs. 994 * 995 * Halt or unhalt the async dma engines (VEGA10). 996 */ 997 static void sdma_v4_0_enable(struct amdgpu_device *adev, bool enable) 998 { 999 u32 f32_cntl; 1000 int i; 1001 1002 if (!enable) { 1003 sdma_v4_0_gfx_enable(adev, enable); 1004 sdma_v4_0_rlc_stop(adev); 1005 if (adev->sdma.has_page_queue) 1006 sdma_v4_0_page_stop(adev); 1007 } 1008 1009 for (i = 0; i < adev->sdma.num_instances; i++) { 1010 f32_cntl = RREG32_SDMA(i, mmSDMA0_F32_CNTL); 1011 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, enable ? 0 : 1); 1012 WREG32_SDMA(i, mmSDMA0_F32_CNTL, f32_cntl); 1013 } 1014 } 1015 1016 /* 1017 * sdma_v4_0_rb_cntl - get parameters for rb_cntl 1018 */ 1019 static uint32_t sdma_v4_0_rb_cntl(struct amdgpu_ring *ring, uint32_t rb_cntl) 1020 { 1021 /* Set ring buffer size in dwords */ 1022 uint32_t rb_bufsz = order_base_2(ring->ring_size / 4); 1023 1024 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz); 1025 #ifdef __BIG_ENDIAN 1026 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1); 1027 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, 1028 RPTR_WRITEBACK_SWAP_ENABLE, 1); 1029 #endif 1030 return rb_cntl; 1031 } 1032 1033 /** 1034 * sdma_v4_0_gfx_resume - setup and start the async dma engines 1035 * 1036 * @adev: amdgpu_device pointer 1037 * @i: instance to resume 1038 * 1039 * Set up the gfx DMA ring buffers and enable them (VEGA10). 1040 * Returns 0 for success, error for failure. 1041 */ 1042 static void sdma_v4_0_gfx_resume(struct amdgpu_device *adev, unsigned int i) 1043 { 1044 struct amdgpu_ring *ring = &adev->sdma.instance[i].ring; 1045 u32 rb_cntl, ib_cntl, wptr_poll_cntl; 1046 u32 doorbell; 1047 u32 doorbell_offset; 1048 u64 wptr_gpu_addr; 1049 1050 rb_cntl = RREG32_SDMA(i, mmSDMA0_GFX_RB_CNTL); 1051 rb_cntl = sdma_v4_0_rb_cntl(ring, rb_cntl); 1052 WREG32_SDMA(i, mmSDMA0_GFX_RB_CNTL, rb_cntl); 1053 1054 /* Initialize the ring buffer's read and write pointers */ 1055 WREG32_SDMA(i, mmSDMA0_GFX_RB_RPTR, 0); 1056 WREG32_SDMA(i, mmSDMA0_GFX_RB_RPTR_HI, 0); 1057 WREG32_SDMA(i, mmSDMA0_GFX_RB_WPTR, 0); 1058 WREG32_SDMA(i, mmSDMA0_GFX_RB_WPTR_HI, 0); 1059 1060 /* set the wb address whether it's enabled or not */ 1061 WREG32_SDMA(i, mmSDMA0_GFX_RB_RPTR_ADDR_HI, 1062 upper_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFF); 1063 WREG32_SDMA(i, mmSDMA0_GFX_RB_RPTR_ADDR_LO, 1064 lower_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFC); 1065 1066 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, 1067 RPTR_WRITEBACK_ENABLE, 1); 1068 1069 WREG32_SDMA(i, mmSDMA0_GFX_RB_BASE, ring->gpu_addr >> 8); 1070 WREG32_SDMA(i, mmSDMA0_GFX_RB_BASE_HI, ring->gpu_addr >> 40); 1071 1072 ring->wptr = 0; 1073 1074 /* before programing wptr to a less value, need set minor_ptr_update first */ 1075 WREG32_SDMA(i, mmSDMA0_GFX_MINOR_PTR_UPDATE, 1); 1076 1077 doorbell = RREG32_SDMA(i, mmSDMA0_GFX_DOORBELL); 1078 doorbell_offset = RREG32_SDMA(i, mmSDMA0_GFX_DOORBELL_OFFSET); 1079 1080 doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 1081 ring->use_doorbell); 1082 doorbell_offset = REG_SET_FIELD(doorbell_offset, 1083 SDMA0_GFX_DOORBELL_OFFSET, 1084 OFFSET, ring->doorbell_index); 1085 WREG32_SDMA(i, mmSDMA0_GFX_DOORBELL, doorbell); 1086 WREG32_SDMA(i, mmSDMA0_GFX_DOORBELL_OFFSET, doorbell_offset); 1087 1088 sdma_v4_0_ring_set_wptr(ring); 1089 1090 /* set minor_ptr_update to 0 after wptr programed */ 1091 WREG32_SDMA(i, mmSDMA0_GFX_MINOR_PTR_UPDATE, 0); 1092 1093 /* setup the wptr shadow polling */ 1094 wptr_gpu_addr = ring->wptr_gpu_addr; 1095 WREG32_SDMA(i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_LO, 1096 lower_32_bits(wptr_gpu_addr)); 1097 WREG32_SDMA(i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_HI, 1098 upper_32_bits(wptr_gpu_addr)); 1099 wptr_poll_cntl = RREG32_SDMA(i, mmSDMA0_GFX_RB_WPTR_POLL_CNTL); 1100 wptr_poll_cntl = REG_SET_FIELD(wptr_poll_cntl, 1101 SDMA0_GFX_RB_WPTR_POLL_CNTL, 1102 F32_POLL_ENABLE, amdgpu_sriov_vf(adev)? 1 : 0); 1103 WREG32_SDMA(i, mmSDMA0_GFX_RB_WPTR_POLL_CNTL, wptr_poll_cntl); 1104 1105 /* enable DMA RB */ 1106 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1); 1107 WREG32_SDMA(i, mmSDMA0_GFX_RB_CNTL, rb_cntl); 1108 1109 ib_cntl = RREG32_SDMA(i, mmSDMA0_GFX_IB_CNTL); 1110 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1); 1111 #ifdef __BIG_ENDIAN 1112 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1); 1113 #endif 1114 /* enable DMA IBs */ 1115 WREG32_SDMA(i, mmSDMA0_GFX_IB_CNTL, ib_cntl); 1116 } 1117 1118 /** 1119 * sdma_v4_0_page_resume - setup and start the async dma engines 1120 * 1121 * @adev: amdgpu_device pointer 1122 * @i: instance to resume 1123 * 1124 * Set up the page DMA ring buffers and enable them (VEGA10). 1125 * Returns 0 for success, error for failure. 1126 */ 1127 static void sdma_v4_0_page_resume(struct amdgpu_device *adev, unsigned int i) 1128 { 1129 struct amdgpu_ring *ring = &adev->sdma.instance[i].page; 1130 u32 rb_cntl, ib_cntl, wptr_poll_cntl; 1131 u32 doorbell; 1132 u32 doorbell_offset; 1133 u64 wptr_gpu_addr; 1134 1135 rb_cntl = RREG32_SDMA(i, mmSDMA0_PAGE_RB_CNTL); 1136 rb_cntl = sdma_v4_0_rb_cntl(ring, rb_cntl); 1137 WREG32_SDMA(i, mmSDMA0_PAGE_RB_CNTL, rb_cntl); 1138 1139 /* Initialize the ring buffer's read and write pointers */ 1140 WREG32_SDMA(i, mmSDMA0_PAGE_RB_RPTR, 0); 1141 WREG32_SDMA(i, mmSDMA0_PAGE_RB_RPTR_HI, 0); 1142 WREG32_SDMA(i, mmSDMA0_PAGE_RB_WPTR, 0); 1143 WREG32_SDMA(i, mmSDMA0_PAGE_RB_WPTR_HI, 0); 1144 1145 /* set the wb address whether it's enabled or not */ 1146 WREG32_SDMA(i, mmSDMA0_PAGE_RB_RPTR_ADDR_HI, 1147 upper_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFF); 1148 WREG32_SDMA(i, mmSDMA0_PAGE_RB_RPTR_ADDR_LO, 1149 lower_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFC); 1150 1151 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_PAGE_RB_CNTL, 1152 RPTR_WRITEBACK_ENABLE, 1); 1153 1154 WREG32_SDMA(i, mmSDMA0_PAGE_RB_BASE, ring->gpu_addr >> 8); 1155 WREG32_SDMA(i, mmSDMA0_PAGE_RB_BASE_HI, ring->gpu_addr >> 40); 1156 1157 ring->wptr = 0; 1158 1159 /* before programing wptr to a less value, need set minor_ptr_update first */ 1160 WREG32_SDMA(i, mmSDMA0_PAGE_MINOR_PTR_UPDATE, 1); 1161 1162 doorbell = RREG32_SDMA(i, mmSDMA0_PAGE_DOORBELL); 1163 doorbell_offset = RREG32_SDMA(i, mmSDMA0_PAGE_DOORBELL_OFFSET); 1164 1165 doorbell = REG_SET_FIELD(doorbell, SDMA0_PAGE_DOORBELL, ENABLE, 1166 ring->use_doorbell); 1167 doorbell_offset = REG_SET_FIELD(doorbell_offset, 1168 SDMA0_PAGE_DOORBELL_OFFSET, 1169 OFFSET, ring->doorbell_index); 1170 WREG32_SDMA(i, mmSDMA0_PAGE_DOORBELL, doorbell); 1171 WREG32_SDMA(i, mmSDMA0_PAGE_DOORBELL_OFFSET, doorbell_offset); 1172 1173 /* paging queue doorbell range is setup at sdma_v4_0_gfx_resume */ 1174 sdma_v4_0_page_ring_set_wptr(ring); 1175 1176 /* set minor_ptr_update to 0 after wptr programed */ 1177 WREG32_SDMA(i, mmSDMA0_PAGE_MINOR_PTR_UPDATE, 0); 1178 1179 /* setup the wptr shadow polling */ 1180 wptr_gpu_addr = ring->wptr_gpu_addr; 1181 WREG32_SDMA(i, mmSDMA0_PAGE_RB_WPTR_POLL_ADDR_LO, 1182 lower_32_bits(wptr_gpu_addr)); 1183 WREG32_SDMA(i, mmSDMA0_PAGE_RB_WPTR_POLL_ADDR_HI, 1184 upper_32_bits(wptr_gpu_addr)); 1185 wptr_poll_cntl = RREG32_SDMA(i, mmSDMA0_PAGE_RB_WPTR_POLL_CNTL); 1186 wptr_poll_cntl = REG_SET_FIELD(wptr_poll_cntl, 1187 SDMA0_PAGE_RB_WPTR_POLL_CNTL, 1188 F32_POLL_ENABLE, amdgpu_sriov_vf(adev)? 1 : 0); 1189 WREG32_SDMA(i, mmSDMA0_PAGE_RB_WPTR_POLL_CNTL, wptr_poll_cntl); 1190 1191 /* enable DMA RB */ 1192 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_PAGE_RB_CNTL, RB_ENABLE, 1); 1193 WREG32_SDMA(i, mmSDMA0_PAGE_RB_CNTL, rb_cntl); 1194 1195 ib_cntl = RREG32_SDMA(i, mmSDMA0_PAGE_IB_CNTL); 1196 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_PAGE_IB_CNTL, IB_ENABLE, 1); 1197 #ifdef __BIG_ENDIAN 1198 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_PAGE_IB_CNTL, IB_SWAP_ENABLE, 1); 1199 #endif 1200 /* enable DMA IBs */ 1201 WREG32_SDMA(i, mmSDMA0_PAGE_IB_CNTL, ib_cntl); 1202 } 1203 1204 static void 1205 sdma_v4_1_update_power_gating(struct amdgpu_device *adev, bool enable) 1206 { 1207 uint32_t def, data; 1208 1209 if (enable && (adev->pg_flags & AMD_PG_SUPPORT_SDMA)) { 1210 /* enable idle interrupt */ 1211 def = data = RREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_CNTL)); 1212 data |= SDMA0_CNTL__CTXEMPTY_INT_ENABLE_MASK; 1213 1214 if (data != def) 1215 WREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_CNTL), data); 1216 } else { 1217 /* disable idle interrupt */ 1218 def = data = RREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_CNTL)); 1219 data &= ~SDMA0_CNTL__CTXEMPTY_INT_ENABLE_MASK; 1220 if (data != def) 1221 WREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_CNTL), data); 1222 } 1223 } 1224 1225 static void sdma_v4_1_init_power_gating(struct amdgpu_device *adev) 1226 { 1227 uint32_t def, data; 1228 1229 /* Enable HW based PG. */ 1230 def = data = RREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_POWER_CNTL)); 1231 data |= SDMA0_POWER_CNTL__PG_CNTL_ENABLE_MASK; 1232 if (data != def) 1233 WREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_POWER_CNTL), data); 1234 1235 /* enable interrupt */ 1236 def = data = RREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_CNTL)); 1237 data |= SDMA0_CNTL__CTXEMPTY_INT_ENABLE_MASK; 1238 if (data != def) 1239 WREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_CNTL), data); 1240 1241 /* Configure hold time to filter in-valid power on/off request. Use default right now */ 1242 def = data = RREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_POWER_CNTL)); 1243 data &= ~SDMA0_POWER_CNTL__ON_OFF_CONDITION_HOLD_TIME_MASK; 1244 data |= (mmSDMA0_POWER_CNTL_DEFAULT & SDMA0_POWER_CNTL__ON_OFF_CONDITION_HOLD_TIME_MASK); 1245 /* Configure switch time for hysteresis purpose. Use default right now */ 1246 data &= ~SDMA0_POWER_CNTL__ON_OFF_STATUS_DURATION_TIME_MASK; 1247 data |= (mmSDMA0_POWER_CNTL_DEFAULT & SDMA0_POWER_CNTL__ON_OFF_STATUS_DURATION_TIME_MASK); 1248 if(data != def) 1249 WREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_POWER_CNTL), data); 1250 } 1251 1252 static void sdma_v4_0_init_pg(struct amdgpu_device *adev) 1253 { 1254 if (!(adev->pg_flags & AMD_PG_SUPPORT_SDMA)) 1255 return; 1256 1257 switch (amdgpu_ip_version(adev, SDMA0_HWIP, 0)) { 1258 case IP_VERSION(4, 1, 0): 1259 case IP_VERSION(4, 1, 1): 1260 case IP_VERSION(4, 1, 2): 1261 sdma_v4_1_init_power_gating(adev); 1262 sdma_v4_1_update_power_gating(adev, true); 1263 break; 1264 default: 1265 break; 1266 } 1267 } 1268 1269 /** 1270 * sdma_v4_0_rlc_resume - setup and start the async dma engines 1271 * 1272 * @adev: amdgpu_device pointer 1273 * 1274 * Set up the compute DMA queues and enable them (VEGA10). 1275 * Returns 0 for success, error for failure. 1276 */ 1277 static int sdma_v4_0_rlc_resume(struct amdgpu_device *adev) 1278 { 1279 sdma_v4_0_init_pg(adev); 1280 1281 return 0; 1282 } 1283 1284 /** 1285 * sdma_v4_0_load_microcode - load the sDMA ME ucode 1286 * 1287 * @adev: amdgpu_device pointer 1288 * 1289 * Loads the sDMA0/1 ucode. 1290 * Returns 0 for success, -EINVAL if the ucode is not available. 1291 */ 1292 static int sdma_v4_0_load_microcode(struct amdgpu_device *adev) 1293 { 1294 const struct sdma_firmware_header_v1_0 *hdr; 1295 const __le32 *fw_data; 1296 u32 fw_size; 1297 int i, j; 1298 1299 /* halt the MEs */ 1300 sdma_v4_0_enable(adev, false); 1301 1302 for (i = 0; i < adev->sdma.num_instances; i++) { 1303 if (!adev->sdma.instance[i].fw) 1304 return -EINVAL; 1305 1306 hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data; 1307 amdgpu_ucode_print_sdma_hdr(&hdr->header); 1308 fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4; 1309 1310 fw_data = (const __le32 *) 1311 (adev->sdma.instance[i].fw->data + 1312 le32_to_cpu(hdr->header.ucode_array_offset_bytes)); 1313 1314 WREG32_SDMA(i, mmSDMA0_UCODE_ADDR, 0); 1315 1316 for (j = 0; j < fw_size; j++) 1317 WREG32_SDMA(i, mmSDMA0_UCODE_DATA, 1318 le32_to_cpup(fw_data++)); 1319 1320 WREG32_SDMA(i, mmSDMA0_UCODE_ADDR, 1321 adev->sdma.instance[i].fw_version); 1322 } 1323 1324 return 0; 1325 } 1326 1327 /** 1328 * sdma_v4_0_start - setup and start the async dma engines 1329 * 1330 * @adev: amdgpu_device pointer 1331 * 1332 * Set up the DMA engines and enable them (VEGA10). 1333 * Returns 0 for success, error for failure. 1334 */ 1335 static int sdma_v4_0_start(struct amdgpu_device *adev) 1336 { 1337 struct amdgpu_ring *ring; 1338 int i, r = 0; 1339 1340 if (amdgpu_sriov_vf(adev)) { 1341 sdma_v4_0_ctx_switch_enable(adev, false); 1342 sdma_v4_0_enable(adev, false); 1343 } else { 1344 1345 if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP) { 1346 r = sdma_v4_0_load_microcode(adev); 1347 if (r) 1348 return r; 1349 } 1350 1351 /* unhalt the MEs */ 1352 sdma_v4_0_enable(adev, true); 1353 /* enable sdma ring preemption */ 1354 sdma_v4_0_ctx_switch_enable(adev, true); 1355 } 1356 1357 /* start the gfx rings and rlc compute queues */ 1358 for (i = 0; i < adev->sdma.num_instances; i++) { 1359 uint32_t temp; 1360 1361 WREG32_SDMA(i, mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL, 0); 1362 sdma_v4_0_gfx_resume(adev, i); 1363 if (adev->sdma.has_page_queue) 1364 sdma_v4_0_page_resume(adev, i); 1365 1366 /* set utc l1 enable flag always to 1 */ 1367 temp = RREG32_SDMA(i, mmSDMA0_CNTL); 1368 temp = REG_SET_FIELD(temp, SDMA0_CNTL, UTC_L1_ENABLE, 1); 1369 WREG32_SDMA(i, mmSDMA0_CNTL, temp); 1370 1371 if (!amdgpu_sriov_vf(adev)) { 1372 /* unhalt engine */ 1373 temp = RREG32_SDMA(i, mmSDMA0_F32_CNTL); 1374 temp = REG_SET_FIELD(temp, SDMA0_F32_CNTL, HALT, 0); 1375 WREG32_SDMA(i, mmSDMA0_F32_CNTL, temp); 1376 } 1377 } 1378 1379 if (amdgpu_sriov_vf(adev)) { 1380 sdma_v4_0_ctx_switch_enable(adev, true); 1381 sdma_v4_0_enable(adev, true); 1382 } else { 1383 r = sdma_v4_0_rlc_resume(adev); 1384 if (r) 1385 return r; 1386 } 1387 1388 for (i = 0; i < adev->sdma.num_instances; i++) { 1389 ring = &adev->sdma.instance[i].ring; 1390 1391 r = amdgpu_ring_test_helper(ring); 1392 if (r) 1393 return r; 1394 1395 if (adev->sdma.has_page_queue) { 1396 struct amdgpu_ring *page = &adev->sdma.instance[i].page; 1397 1398 r = amdgpu_ring_test_helper(page); 1399 if (r) 1400 return r; 1401 } 1402 } 1403 1404 return r; 1405 } 1406 1407 /** 1408 * sdma_v4_0_ring_test_ring - simple async dma engine test 1409 * 1410 * @ring: amdgpu_ring structure holding ring information 1411 * 1412 * Test the DMA engine by writing using it to write an 1413 * value to memory. (VEGA10). 1414 * Returns 0 for success, error for failure. 1415 */ 1416 static int sdma_v4_0_ring_test_ring(struct amdgpu_ring *ring) 1417 { 1418 struct amdgpu_device *adev = ring->adev; 1419 unsigned i; 1420 unsigned index; 1421 int r; 1422 u32 tmp; 1423 u64 gpu_addr; 1424 1425 r = amdgpu_device_wb_get(adev, &index); 1426 if (r) 1427 return r; 1428 1429 gpu_addr = adev->wb.gpu_addr + (index * 4); 1430 tmp = 0xCAFEDEAD; 1431 adev->wb.wb[index] = cpu_to_le32(tmp); 1432 1433 r = amdgpu_ring_alloc(ring, 5); 1434 if (r) 1435 goto error_free_wb; 1436 1437 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) | 1438 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR)); 1439 amdgpu_ring_write(ring, lower_32_bits(gpu_addr)); 1440 amdgpu_ring_write(ring, upper_32_bits(gpu_addr)); 1441 amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0)); 1442 amdgpu_ring_write(ring, 0xDEADBEEF); 1443 amdgpu_ring_commit(ring); 1444 1445 for (i = 0; i < adev->usec_timeout; i++) { 1446 tmp = le32_to_cpu(adev->wb.wb[index]); 1447 if (tmp == 0xDEADBEEF) 1448 break; 1449 udelay(1); 1450 } 1451 1452 if (i >= adev->usec_timeout) 1453 r = -ETIMEDOUT; 1454 1455 error_free_wb: 1456 amdgpu_device_wb_free(adev, index); 1457 return r; 1458 } 1459 1460 /** 1461 * sdma_v4_0_ring_test_ib - test an IB on the DMA engine 1462 * 1463 * @ring: amdgpu_ring structure holding ring information 1464 * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT 1465 * 1466 * Test a simple IB in the DMA ring (VEGA10). 1467 * Returns 0 on success, error on failure. 1468 */ 1469 static int sdma_v4_0_ring_test_ib(struct amdgpu_ring *ring, long timeout) 1470 { 1471 struct amdgpu_device *adev = ring->adev; 1472 struct amdgpu_ib ib; 1473 struct dma_fence *f = NULL; 1474 unsigned index; 1475 long r; 1476 u32 tmp = 0; 1477 u64 gpu_addr; 1478 1479 r = amdgpu_device_wb_get(adev, &index); 1480 if (r) 1481 return r; 1482 1483 gpu_addr = adev->wb.gpu_addr + (index * 4); 1484 tmp = 0xCAFEDEAD; 1485 adev->wb.wb[index] = cpu_to_le32(tmp); 1486 memset(&ib, 0, sizeof(ib)); 1487 r = amdgpu_ib_get(adev, NULL, 256, 1488 AMDGPU_IB_POOL_DIRECT, &ib); 1489 if (r) 1490 goto err0; 1491 1492 ib.ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) | 1493 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR); 1494 ib.ptr[1] = lower_32_bits(gpu_addr); 1495 ib.ptr[2] = upper_32_bits(gpu_addr); 1496 ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0); 1497 ib.ptr[4] = 0xDEADBEEF; 1498 ib.ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP); 1499 ib.ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP); 1500 ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP); 1501 ib.length_dw = 8; 1502 1503 r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f); 1504 if (r) 1505 goto err1; 1506 1507 r = dma_fence_wait_timeout(f, false, timeout); 1508 if (r == 0) { 1509 r = -ETIMEDOUT; 1510 goto err1; 1511 } else if (r < 0) { 1512 goto err1; 1513 } 1514 tmp = le32_to_cpu(adev->wb.wb[index]); 1515 if (tmp == 0xDEADBEEF) 1516 r = 0; 1517 else 1518 r = -EINVAL; 1519 1520 err1: 1521 amdgpu_ib_free(adev, &ib, NULL); 1522 dma_fence_put(f); 1523 err0: 1524 amdgpu_device_wb_free(adev, index); 1525 return r; 1526 } 1527 1528 1529 /** 1530 * sdma_v4_0_vm_copy_pte - update PTEs by copying them from the GART 1531 * 1532 * @ib: indirect buffer to fill with commands 1533 * @pe: addr of the page entry 1534 * @src: src addr to copy from 1535 * @count: number of page entries to update 1536 * 1537 * Update PTEs by copying them from the GART using sDMA (VEGA10). 1538 */ 1539 static void sdma_v4_0_vm_copy_pte(struct amdgpu_ib *ib, 1540 uint64_t pe, uint64_t src, 1541 unsigned count) 1542 { 1543 unsigned bytes = count * 8; 1544 1545 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) | 1546 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR); 1547 ib->ptr[ib->length_dw++] = bytes - 1; 1548 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */ 1549 ib->ptr[ib->length_dw++] = lower_32_bits(src); 1550 ib->ptr[ib->length_dw++] = upper_32_bits(src); 1551 ib->ptr[ib->length_dw++] = lower_32_bits(pe); 1552 ib->ptr[ib->length_dw++] = upper_32_bits(pe); 1553 1554 } 1555 1556 /** 1557 * sdma_v4_0_vm_write_pte - update PTEs by writing them manually 1558 * 1559 * @ib: indirect buffer to fill with commands 1560 * @pe: addr of the page entry 1561 * @value: dst addr to write into pe 1562 * @count: number of page entries to update 1563 * @incr: increase next addr by incr bytes 1564 * 1565 * Update PTEs by writing them manually using sDMA (VEGA10). 1566 */ 1567 static void sdma_v4_0_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe, 1568 uint64_t value, unsigned count, 1569 uint32_t incr) 1570 { 1571 unsigned ndw = count * 2; 1572 1573 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) | 1574 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR); 1575 ib->ptr[ib->length_dw++] = lower_32_bits(pe); 1576 ib->ptr[ib->length_dw++] = upper_32_bits(pe); 1577 ib->ptr[ib->length_dw++] = ndw - 1; 1578 for (; ndw > 0; ndw -= 2) { 1579 ib->ptr[ib->length_dw++] = lower_32_bits(value); 1580 ib->ptr[ib->length_dw++] = upper_32_bits(value); 1581 value += incr; 1582 } 1583 } 1584 1585 /** 1586 * sdma_v4_0_vm_set_pte_pde - update the page tables using sDMA 1587 * 1588 * @ib: indirect buffer to fill with commands 1589 * @pe: addr of the page entry 1590 * @addr: dst addr to write into pe 1591 * @count: number of page entries to update 1592 * @incr: increase next addr by incr bytes 1593 * @flags: access flags 1594 * 1595 * Update the page tables using sDMA (VEGA10). 1596 */ 1597 static void sdma_v4_0_vm_set_pte_pde(struct amdgpu_ib *ib, 1598 uint64_t pe, 1599 uint64_t addr, unsigned count, 1600 uint32_t incr, uint64_t flags) 1601 { 1602 /* for physically contiguous pages (vram) */ 1603 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_PTEPDE); 1604 ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */ 1605 ib->ptr[ib->length_dw++] = upper_32_bits(pe); 1606 ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */ 1607 ib->ptr[ib->length_dw++] = upper_32_bits(flags); 1608 ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */ 1609 ib->ptr[ib->length_dw++] = upper_32_bits(addr); 1610 ib->ptr[ib->length_dw++] = incr; /* increment size */ 1611 ib->ptr[ib->length_dw++] = 0; 1612 ib->ptr[ib->length_dw++] = count - 1; /* number of entries */ 1613 } 1614 1615 /** 1616 * sdma_v4_0_ring_pad_ib - pad the IB to the required number of dw 1617 * 1618 * @ring: amdgpu_ring structure holding ring information 1619 * @ib: indirect buffer to fill with padding 1620 */ 1621 static void sdma_v4_0_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib) 1622 { 1623 struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring); 1624 u32 pad_count; 1625 int i; 1626 1627 pad_count = (-ib->length_dw) & 7; 1628 for (i = 0; i < pad_count; i++) 1629 if (sdma && sdma->burst_nop && (i == 0)) 1630 ib->ptr[ib->length_dw++] = 1631 SDMA_PKT_HEADER_OP(SDMA_OP_NOP) | 1632 SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1); 1633 else 1634 ib->ptr[ib->length_dw++] = 1635 SDMA_PKT_HEADER_OP(SDMA_OP_NOP); 1636 } 1637 1638 1639 /** 1640 * sdma_v4_0_ring_emit_pipeline_sync - sync the pipeline 1641 * 1642 * @ring: amdgpu_ring pointer 1643 * 1644 * Make sure all previous operations are completed (CIK). 1645 */ 1646 static void sdma_v4_0_ring_emit_pipeline_sync(struct amdgpu_ring *ring) 1647 { 1648 uint32_t seq = ring->fence_drv.sync_seq; 1649 uint64_t addr = ring->fence_drv.gpu_addr; 1650 1651 /* wait for idle */ 1652 sdma_v4_0_wait_reg_mem(ring, 1, 0, 1653 addr & 0xfffffffc, 1654 upper_32_bits(addr) & 0xffffffff, 1655 seq, 0xffffffff, 4); 1656 } 1657 1658 1659 /** 1660 * sdma_v4_0_ring_emit_vm_flush - vm flush using sDMA 1661 * 1662 * @ring: amdgpu_ring pointer 1663 * @vmid: vmid number to use 1664 * @pd_addr: address 1665 * 1666 * Update the page table base and flush the VM TLB 1667 * using sDMA (VEGA10). 1668 */ 1669 static void sdma_v4_0_ring_emit_vm_flush(struct amdgpu_ring *ring, 1670 unsigned vmid, uint64_t pd_addr) 1671 { 1672 amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr); 1673 } 1674 1675 static void sdma_v4_0_ring_emit_wreg(struct amdgpu_ring *ring, 1676 uint32_t reg, uint32_t val) 1677 { 1678 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) | 1679 SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf)); 1680 amdgpu_ring_write(ring, reg); 1681 amdgpu_ring_write(ring, val); 1682 } 1683 1684 static void sdma_v4_0_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg, 1685 uint32_t val, uint32_t mask) 1686 { 1687 sdma_v4_0_wait_reg_mem(ring, 0, 0, reg, 0, val, mask, 10); 1688 } 1689 1690 static bool sdma_v4_0_fw_support_paging_queue(struct amdgpu_device *adev) 1691 { 1692 uint fw_version = adev->sdma.instance[0].fw_version; 1693 1694 switch (amdgpu_ip_version(adev, SDMA0_HWIP, 0)) { 1695 case IP_VERSION(4, 0, 0): 1696 return fw_version >= 430; 1697 case IP_VERSION(4, 0, 1): 1698 /*return fw_version >= 31;*/ 1699 return false; 1700 case IP_VERSION(4, 2, 0): 1701 return fw_version >= 123; 1702 default: 1703 return false; 1704 } 1705 } 1706 1707 static int sdma_v4_0_early_init(void *handle) 1708 { 1709 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1710 int r; 1711 1712 r = sdma_v4_0_init_microcode(adev); 1713 if (r) 1714 return r; 1715 1716 /* TODO: Page queue breaks driver reload under SRIOV */ 1717 if ((amdgpu_ip_version(adev, SDMA0_HWIP, 0) == IP_VERSION(4, 0, 0)) && 1718 amdgpu_sriov_vf((adev))) 1719 adev->sdma.has_page_queue = false; 1720 else if (sdma_v4_0_fw_support_paging_queue(adev)) 1721 adev->sdma.has_page_queue = true; 1722 1723 sdma_v4_0_set_ring_funcs(adev); 1724 sdma_v4_0_set_buffer_funcs(adev); 1725 sdma_v4_0_set_vm_pte_funcs(adev); 1726 sdma_v4_0_set_irq_funcs(adev); 1727 sdma_v4_0_set_ras_funcs(adev); 1728 1729 return 0; 1730 } 1731 1732 static int sdma_v4_0_process_ras_data_cb(struct amdgpu_device *adev, 1733 void *err_data, 1734 struct amdgpu_iv_entry *entry); 1735 1736 static int sdma_v4_0_late_init(void *handle) 1737 { 1738 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1739 1740 sdma_v4_0_setup_ulv(adev); 1741 1742 if (!amdgpu_persistent_edc_harvesting_supported(adev)) 1743 amdgpu_ras_reset_error_count(adev, AMDGPU_RAS_BLOCK__SDMA); 1744 1745 return 0; 1746 } 1747 1748 static int sdma_v4_0_sw_init(void *handle) 1749 { 1750 struct amdgpu_ring *ring; 1751 int r, i; 1752 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1753 1754 /* SDMA trap event */ 1755 for (i = 0; i < adev->sdma.num_instances; i++) { 1756 r = amdgpu_irq_add_id(adev, sdma_v4_0_seq_to_irq_id(i), 1757 SDMA0_4_0__SRCID__SDMA_TRAP, 1758 &adev->sdma.trap_irq); 1759 if (r) 1760 return r; 1761 } 1762 1763 /* SDMA SRAM ECC event */ 1764 for (i = 0; i < adev->sdma.num_instances; i++) { 1765 r = amdgpu_irq_add_id(adev, sdma_v4_0_seq_to_irq_id(i), 1766 SDMA0_4_0__SRCID__SDMA_SRAM_ECC, 1767 &adev->sdma.ecc_irq); 1768 if (r) 1769 return r; 1770 } 1771 1772 /* SDMA VM_HOLE/DOORBELL_INV/POLL_TIMEOUT/SRBM_WRITE_PROTECTION event*/ 1773 for (i = 0; i < adev->sdma.num_instances; i++) { 1774 r = amdgpu_irq_add_id(adev, sdma_v4_0_seq_to_irq_id(i), 1775 SDMA0_4_0__SRCID__SDMA_VM_HOLE, 1776 &adev->sdma.vm_hole_irq); 1777 if (r) 1778 return r; 1779 1780 r = amdgpu_irq_add_id(adev, sdma_v4_0_seq_to_irq_id(i), 1781 SDMA0_4_0__SRCID__SDMA_DOORBELL_INVALID, 1782 &adev->sdma.doorbell_invalid_irq); 1783 if (r) 1784 return r; 1785 1786 r = amdgpu_irq_add_id(adev, sdma_v4_0_seq_to_irq_id(i), 1787 SDMA0_4_0__SRCID__SDMA_POLL_TIMEOUT, 1788 &adev->sdma.pool_timeout_irq); 1789 if (r) 1790 return r; 1791 1792 r = amdgpu_irq_add_id(adev, sdma_v4_0_seq_to_irq_id(i), 1793 SDMA0_4_0__SRCID__SDMA_SRBMWRITE, 1794 &adev->sdma.srbm_write_irq); 1795 if (r) 1796 return r; 1797 } 1798 1799 for (i = 0; i < adev->sdma.num_instances; i++) { 1800 ring = &adev->sdma.instance[i].ring; 1801 ring->ring_obj = NULL; 1802 ring->use_doorbell = true; 1803 1804 DRM_DEBUG("SDMA %d use_doorbell being set to: [%s]\n", i, 1805 ring->use_doorbell?"true":"false"); 1806 1807 /* doorbell size is 2 dwords, get DWORD offset */ 1808 ring->doorbell_index = adev->doorbell_index.sdma_engine[i] << 1; 1809 1810 /* 1811 * On Arcturus, SDMA instance 5~7 has a different vmhub 1812 * type(AMDGPU_MMHUB1). 1813 */ 1814 if (amdgpu_ip_version(adev, SDMA0_HWIP, 0) == 1815 IP_VERSION(4, 2, 2) && 1816 i >= 5) 1817 ring->vm_hub = AMDGPU_MMHUB1(0); 1818 else 1819 ring->vm_hub = AMDGPU_MMHUB0(0); 1820 1821 sprintf(ring->name, "sdma%d", i); 1822 r = amdgpu_ring_init(adev, ring, 1024, &adev->sdma.trap_irq, 1823 AMDGPU_SDMA_IRQ_INSTANCE0 + i, 1824 AMDGPU_RING_PRIO_DEFAULT, NULL); 1825 if (r) 1826 return r; 1827 1828 if (adev->sdma.has_page_queue) { 1829 ring = &adev->sdma.instance[i].page; 1830 ring->ring_obj = NULL; 1831 ring->use_doorbell = true; 1832 1833 /* paging queue use same doorbell index/routing as gfx queue 1834 * with 0x400 (4096 dwords) offset on second doorbell page 1835 */ 1836 if (amdgpu_ip_version(adev, SDMA0_HWIP, 0) >= 1837 IP_VERSION(4, 0, 0) && 1838 amdgpu_ip_version(adev, SDMA0_HWIP, 0) < 1839 IP_VERSION(4, 2, 0)) { 1840 ring->doorbell_index = 1841 adev->doorbell_index.sdma_engine[i] << 1; 1842 ring->doorbell_index += 0x400; 1843 } else { 1844 /* From vega20, the sdma_doorbell_range in 1st 1845 * doorbell page is reserved for page queue. 1846 */ 1847 ring->doorbell_index = 1848 (adev->doorbell_index.sdma_engine[i] + 1) << 1; 1849 } 1850 1851 if (amdgpu_ip_version(adev, SDMA0_HWIP, 0) == 1852 IP_VERSION(4, 2, 2) && 1853 i >= 5) 1854 ring->vm_hub = AMDGPU_MMHUB1(0); 1855 else 1856 ring->vm_hub = AMDGPU_MMHUB0(0); 1857 1858 sprintf(ring->name, "page%d", i); 1859 r = amdgpu_ring_init(adev, ring, 1024, 1860 &adev->sdma.trap_irq, 1861 AMDGPU_SDMA_IRQ_INSTANCE0 + i, 1862 AMDGPU_RING_PRIO_DEFAULT, NULL); 1863 if (r) 1864 return r; 1865 } 1866 } 1867 1868 if (amdgpu_sdma_ras_sw_init(adev)) { 1869 dev_err(adev->dev, "Failed to initialize sdma ras block!\n"); 1870 return -EINVAL; 1871 } 1872 1873 return r; 1874 } 1875 1876 static int sdma_v4_0_sw_fini(void *handle) 1877 { 1878 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1879 int i; 1880 1881 for (i = 0; i < adev->sdma.num_instances; i++) { 1882 amdgpu_ring_fini(&adev->sdma.instance[i].ring); 1883 if (adev->sdma.has_page_queue) 1884 amdgpu_ring_fini(&adev->sdma.instance[i].page); 1885 } 1886 1887 if (amdgpu_ip_version(adev, SDMA0_HWIP, 0) == IP_VERSION(4, 2, 2) || 1888 amdgpu_ip_version(adev, SDMA0_HWIP, 0) == IP_VERSION(4, 4, 0)) 1889 amdgpu_sdma_destroy_inst_ctx(adev, true); 1890 else 1891 amdgpu_sdma_destroy_inst_ctx(adev, false); 1892 1893 return 0; 1894 } 1895 1896 static int sdma_v4_0_hw_init(void *handle) 1897 { 1898 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1899 1900 if (adev->flags & AMD_IS_APU) 1901 amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_SDMA, false); 1902 1903 if (!amdgpu_sriov_vf(adev)) 1904 sdma_v4_0_init_golden_registers(adev); 1905 1906 return sdma_v4_0_start(adev); 1907 } 1908 1909 static int sdma_v4_0_hw_fini(void *handle) 1910 { 1911 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1912 int i; 1913 1914 if (amdgpu_sriov_vf(adev)) 1915 return 0; 1916 1917 if (amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__SDMA)) { 1918 for (i = 0; i < adev->sdma.num_instances; i++) { 1919 amdgpu_irq_put(adev, &adev->sdma.ecc_irq, 1920 AMDGPU_SDMA_IRQ_INSTANCE0 + i); 1921 } 1922 } 1923 1924 sdma_v4_0_ctx_switch_enable(adev, false); 1925 sdma_v4_0_enable(adev, false); 1926 1927 if (adev->flags & AMD_IS_APU) 1928 amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_SDMA, true); 1929 1930 return 0; 1931 } 1932 1933 static int sdma_v4_0_suspend(void *handle) 1934 { 1935 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1936 1937 /* SMU saves SDMA state for us */ 1938 if (adev->in_s0ix) { 1939 sdma_v4_0_gfx_enable(adev, false); 1940 return 0; 1941 } 1942 1943 return sdma_v4_0_hw_fini(adev); 1944 } 1945 1946 static int sdma_v4_0_resume(void *handle) 1947 { 1948 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1949 1950 /* SMU restores SDMA state for us */ 1951 if (adev->in_s0ix) { 1952 sdma_v4_0_enable(adev, true); 1953 sdma_v4_0_gfx_enable(adev, true); 1954 return 0; 1955 } 1956 1957 return sdma_v4_0_hw_init(adev); 1958 } 1959 1960 static bool sdma_v4_0_is_idle(void *handle) 1961 { 1962 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1963 u32 i; 1964 1965 for (i = 0; i < adev->sdma.num_instances; i++) { 1966 u32 tmp = RREG32_SDMA(i, mmSDMA0_STATUS_REG); 1967 1968 if (!(tmp & SDMA0_STATUS_REG__IDLE_MASK)) 1969 return false; 1970 } 1971 1972 return true; 1973 } 1974 1975 static int sdma_v4_0_wait_for_idle(void *handle) 1976 { 1977 unsigned i, j; 1978 u32 sdma[AMDGPU_MAX_SDMA_INSTANCES]; 1979 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1980 1981 for (i = 0; i < adev->usec_timeout; i++) { 1982 for (j = 0; j < adev->sdma.num_instances; j++) { 1983 sdma[j] = RREG32_SDMA(j, mmSDMA0_STATUS_REG); 1984 if (!(sdma[j] & SDMA0_STATUS_REG__IDLE_MASK)) 1985 break; 1986 } 1987 if (j == adev->sdma.num_instances) 1988 return 0; 1989 udelay(1); 1990 } 1991 return -ETIMEDOUT; 1992 } 1993 1994 static int sdma_v4_0_soft_reset(void *handle) 1995 { 1996 /* todo */ 1997 1998 return 0; 1999 } 2000 2001 static int sdma_v4_0_set_trap_irq_state(struct amdgpu_device *adev, 2002 struct amdgpu_irq_src *source, 2003 unsigned type, 2004 enum amdgpu_interrupt_state state) 2005 { 2006 u32 sdma_cntl; 2007 2008 sdma_cntl = RREG32_SDMA(type, mmSDMA0_CNTL); 2009 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 2010 state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0); 2011 WREG32_SDMA(type, mmSDMA0_CNTL, sdma_cntl); 2012 2013 return 0; 2014 } 2015 2016 static int sdma_v4_0_process_trap_irq(struct amdgpu_device *adev, 2017 struct amdgpu_irq_src *source, 2018 struct amdgpu_iv_entry *entry) 2019 { 2020 uint32_t instance; 2021 2022 DRM_DEBUG("IH: SDMA trap\n"); 2023 instance = sdma_v4_0_irq_id_to_seq(entry->client_id); 2024 switch (entry->ring_id) { 2025 case 0: 2026 amdgpu_fence_process(&adev->sdma.instance[instance].ring); 2027 break; 2028 case 1: 2029 if (amdgpu_ip_version(adev, SDMA0_HWIP, 0) == 2030 IP_VERSION(4, 2, 0)) 2031 amdgpu_fence_process(&adev->sdma.instance[instance].page); 2032 break; 2033 case 2: 2034 /* XXX compute */ 2035 break; 2036 case 3: 2037 if (amdgpu_ip_version(adev, SDMA0_HWIP, 0) != 2038 IP_VERSION(4, 2, 0)) 2039 amdgpu_fence_process(&adev->sdma.instance[instance].page); 2040 break; 2041 } 2042 return 0; 2043 } 2044 2045 static int sdma_v4_0_process_ras_data_cb(struct amdgpu_device *adev, 2046 void *err_data, 2047 struct amdgpu_iv_entry *entry) 2048 { 2049 int instance; 2050 2051 /* When “Full RAS” is enabled, the per-IP interrupt sources should 2052 * be disabled and the driver should only look for the aggregated 2053 * interrupt via sync flood 2054 */ 2055 if (amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__GFX)) 2056 goto out; 2057 2058 instance = sdma_v4_0_irq_id_to_seq(entry->client_id); 2059 if (instance < 0) 2060 goto out; 2061 2062 amdgpu_sdma_process_ras_data_cb(adev, err_data, entry); 2063 2064 out: 2065 return AMDGPU_RAS_SUCCESS; 2066 } 2067 2068 static int sdma_v4_0_process_illegal_inst_irq(struct amdgpu_device *adev, 2069 struct amdgpu_irq_src *source, 2070 struct amdgpu_iv_entry *entry) 2071 { 2072 int instance; 2073 2074 DRM_ERROR("Illegal instruction in SDMA command stream\n"); 2075 2076 instance = sdma_v4_0_irq_id_to_seq(entry->client_id); 2077 if (instance < 0) 2078 return 0; 2079 2080 switch (entry->ring_id) { 2081 case 0: 2082 drm_sched_fault(&adev->sdma.instance[instance].ring.sched); 2083 break; 2084 } 2085 return 0; 2086 } 2087 2088 static int sdma_v4_0_set_ecc_irq_state(struct amdgpu_device *adev, 2089 struct amdgpu_irq_src *source, 2090 unsigned type, 2091 enum amdgpu_interrupt_state state) 2092 { 2093 u32 sdma_edc_config; 2094 2095 sdma_edc_config = RREG32_SDMA(type, mmSDMA0_EDC_CONFIG); 2096 sdma_edc_config = REG_SET_FIELD(sdma_edc_config, SDMA0_EDC_CONFIG, ECC_INT_ENABLE, 2097 state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0); 2098 WREG32_SDMA(type, mmSDMA0_EDC_CONFIG, sdma_edc_config); 2099 2100 return 0; 2101 } 2102 2103 static int sdma_v4_0_print_iv_entry(struct amdgpu_device *adev, 2104 struct amdgpu_iv_entry *entry) 2105 { 2106 int instance; 2107 struct amdgpu_task_info *task_info; 2108 u64 addr; 2109 2110 instance = sdma_v4_0_irq_id_to_seq(entry->client_id); 2111 if (instance < 0 || instance >= adev->sdma.num_instances) { 2112 dev_err(adev->dev, "sdma instance invalid %d\n", instance); 2113 return -EINVAL; 2114 } 2115 2116 addr = (u64)entry->src_data[0] << 12; 2117 addr |= ((u64)entry->src_data[1] & 0xf) << 44; 2118 2119 dev_dbg_ratelimited(adev->dev, 2120 "[sdma%d] address:0x%016llx src_id:%u ring:%u vmid:%u pasid:%u\n", 2121 instance, addr, entry->src_id, entry->ring_id, entry->vmid, 2122 entry->pasid); 2123 2124 task_info = amdgpu_vm_get_task_info_pasid(adev, entry->pasid); 2125 if (task_info) { 2126 dev_dbg_ratelimited(adev->dev, 2127 " for process %s pid %d thread %s pid %d\n", 2128 task_info->process_name, task_info->tgid, 2129 task_info->task_name, task_info->pid); 2130 amdgpu_vm_put_task_info(task_info); 2131 } 2132 2133 return 0; 2134 } 2135 2136 static int sdma_v4_0_process_vm_hole_irq(struct amdgpu_device *adev, 2137 struct amdgpu_irq_src *source, 2138 struct amdgpu_iv_entry *entry) 2139 { 2140 dev_dbg_ratelimited(adev->dev, "MC or SEM address in VM hole\n"); 2141 sdma_v4_0_print_iv_entry(adev, entry); 2142 return 0; 2143 } 2144 2145 static int sdma_v4_0_process_doorbell_invalid_irq(struct amdgpu_device *adev, 2146 struct amdgpu_irq_src *source, 2147 struct amdgpu_iv_entry *entry) 2148 { 2149 dev_dbg_ratelimited(adev->dev, "SDMA received a doorbell from BIF with byte_enable !=0xff\n"); 2150 sdma_v4_0_print_iv_entry(adev, entry); 2151 return 0; 2152 } 2153 2154 static int sdma_v4_0_process_pool_timeout_irq(struct amdgpu_device *adev, 2155 struct amdgpu_irq_src *source, 2156 struct amdgpu_iv_entry *entry) 2157 { 2158 dev_dbg_ratelimited(adev->dev, 2159 "Polling register/memory timeout executing POLL_REG/MEM with finite timer\n"); 2160 sdma_v4_0_print_iv_entry(adev, entry); 2161 return 0; 2162 } 2163 2164 static int sdma_v4_0_process_srbm_write_irq(struct amdgpu_device *adev, 2165 struct amdgpu_irq_src *source, 2166 struct amdgpu_iv_entry *entry) 2167 { 2168 dev_dbg_ratelimited(adev->dev, 2169 "SDMA gets an Register Write SRBM_WRITE command in non-privilege command buffer\n"); 2170 sdma_v4_0_print_iv_entry(adev, entry); 2171 return 0; 2172 } 2173 2174 static void sdma_v4_0_update_medium_grain_clock_gating( 2175 struct amdgpu_device *adev, 2176 bool enable) 2177 { 2178 uint32_t data, def; 2179 int i; 2180 2181 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) { 2182 for (i = 0; i < adev->sdma.num_instances; i++) { 2183 def = data = RREG32_SDMA(i, mmSDMA0_CLK_CTRL); 2184 data &= ~(SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK | 2185 SDMA0_CLK_CTRL__SOFT_OVERRIDE6_MASK | 2186 SDMA0_CLK_CTRL__SOFT_OVERRIDE5_MASK | 2187 SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK | 2188 SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK | 2189 SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK | 2190 SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK | 2191 SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK); 2192 if (def != data) 2193 WREG32_SDMA(i, mmSDMA0_CLK_CTRL, data); 2194 } 2195 } else { 2196 for (i = 0; i < adev->sdma.num_instances; i++) { 2197 def = data = RREG32_SDMA(i, mmSDMA0_CLK_CTRL); 2198 data |= (SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK | 2199 SDMA0_CLK_CTRL__SOFT_OVERRIDE6_MASK | 2200 SDMA0_CLK_CTRL__SOFT_OVERRIDE5_MASK | 2201 SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK | 2202 SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK | 2203 SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK | 2204 SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK | 2205 SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK); 2206 if (def != data) 2207 WREG32_SDMA(i, mmSDMA0_CLK_CTRL, data); 2208 } 2209 } 2210 } 2211 2212 2213 static void sdma_v4_0_update_medium_grain_light_sleep( 2214 struct amdgpu_device *adev, 2215 bool enable) 2216 { 2217 uint32_t data, def; 2218 int i; 2219 2220 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) { 2221 for (i = 0; i < adev->sdma.num_instances; i++) { 2222 /* 1-not override: enable sdma mem light sleep */ 2223 def = data = RREG32_SDMA(0, mmSDMA0_POWER_CNTL); 2224 data |= SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK; 2225 if (def != data) 2226 WREG32_SDMA(0, mmSDMA0_POWER_CNTL, data); 2227 } 2228 } else { 2229 for (i = 0; i < adev->sdma.num_instances; i++) { 2230 /* 0-override:disable sdma mem light sleep */ 2231 def = data = RREG32_SDMA(0, mmSDMA0_POWER_CNTL); 2232 data &= ~SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK; 2233 if (def != data) 2234 WREG32_SDMA(0, mmSDMA0_POWER_CNTL, data); 2235 } 2236 } 2237 } 2238 2239 static int sdma_v4_0_set_clockgating_state(void *handle, 2240 enum amd_clockgating_state state) 2241 { 2242 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 2243 2244 if (amdgpu_sriov_vf(adev)) 2245 return 0; 2246 2247 sdma_v4_0_update_medium_grain_clock_gating(adev, 2248 state == AMD_CG_STATE_GATE); 2249 sdma_v4_0_update_medium_grain_light_sleep(adev, 2250 state == AMD_CG_STATE_GATE); 2251 return 0; 2252 } 2253 2254 static int sdma_v4_0_set_powergating_state(void *handle, 2255 enum amd_powergating_state state) 2256 { 2257 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 2258 2259 switch (amdgpu_ip_version(adev, SDMA0_HWIP, 0)) { 2260 case IP_VERSION(4, 1, 0): 2261 case IP_VERSION(4, 1, 1): 2262 case IP_VERSION(4, 1, 2): 2263 sdma_v4_1_update_power_gating(adev, 2264 state == AMD_PG_STATE_GATE); 2265 break; 2266 default: 2267 break; 2268 } 2269 2270 return 0; 2271 } 2272 2273 static void sdma_v4_0_get_clockgating_state(void *handle, u64 *flags) 2274 { 2275 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 2276 int data; 2277 2278 if (amdgpu_sriov_vf(adev)) 2279 *flags = 0; 2280 2281 /* AMD_CG_SUPPORT_SDMA_MGCG */ 2282 data = RREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_CLK_CTRL)); 2283 if (!(data & SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK)) 2284 *flags |= AMD_CG_SUPPORT_SDMA_MGCG; 2285 2286 /* AMD_CG_SUPPORT_SDMA_LS */ 2287 data = RREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_POWER_CNTL)); 2288 if (data & SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK) 2289 *flags |= AMD_CG_SUPPORT_SDMA_LS; 2290 } 2291 2292 const struct amd_ip_funcs sdma_v4_0_ip_funcs = { 2293 .name = "sdma_v4_0", 2294 .early_init = sdma_v4_0_early_init, 2295 .late_init = sdma_v4_0_late_init, 2296 .sw_init = sdma_v4_0_sw_init, 2297 .sw_fini = sdma_v4_0_sw_fini, 2298 .hw_init = sdma_v4_0_hw_init, 2299 .hw_fini = sdma_v4_0_hw_fini, 2300 .suspend = sdma_v4_0_suspend, 2301 .resume = sdma_v4_0_resume, 2302 .is_idle = sdma_v4_0_is_idle, 2303 .wait_for_idle = sdma_v4_0_wait_for_idle, 2304 .soft_reset = sdma_v4_0_soft_reset, 2305 .set_clockgating_state = sdma_v4_0_set_clockgating_state, 2306 .set_powergating_state = sdma_v4_0_set_powergating_state, 2307 .get_clockgating_state = sdma_v4_0_get_clockgating_state, 2308 }; 2309 2310 static const struct amdgpu_ring_funcs sdma_v4_0_ring_funcs = { 2311 .type = AMDGPU_RING_TYPE_SDMA, 2312 .align_mask = 0xff, 2313 .nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP), 2314 .support_64bit_ptrs = true, 2315 .secure_submission_supported = true, 2316 .get_rptr = sdma_v4_0_ring_get_rptr, 2317 .get_wptr = sdma_v4_0_ring_get_wptr, 2318 .set_wptr = sdma_v4_0_ring_set_wptr, 2319 .emit_frame_size = 2320 6 + /* sdma_v4_0_ring_emit_hdp_flush */ 2321 3 + /* hdp invalidate */ 2322 6 + /* sdma_v4_0_ring_emit_pipeline_sync */ 2323 /* sdma_v4_0_ring_emit_vm_flush */ 2324 SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 + 2325 SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6 + 2326 10 + 10 + 10, /* sdma_v4_0_ring_emit_fence x3 for user fence, vm fence */ 2327 .emit_ib_size = 7 + 6, /* sdma_v4_0_ring_emit_ib */ 2328 .emit_ib = sdma_v4_0_ring_emit_ib, 2329 .emit_fence = sdma_v4_0_ring_emit_fence, 2330 .emit_pipeline_sync = sdma_v4_0_ring_emit_pipeline_sync, 2331 .emit_vm_flush = sdma_v4_0_ring_emit_vm_flush, 2332 .emit_hdp_flush = sdma_v4_0_ring_emit_hdp_flush, 2333 .test_ring = sdma_v4_0_ring_test_ring, 2334 .test_ib = sdma_v4_0_ring_test_ib, 2335 .insert_nop = sdma_v4_0_ring_insert_nop, 2336 .pad_ib = sdma_v4_0_ring_pad_ib, 2337 .emit_wreg = sdma_v4_0_ring_emit_wreg, 2338 .emit_reg_wait = sdma_v4_0_ring_emit_reg_wait, 2339 .emit_reg_write_reg_wait = amdgpu_ring_emit_reg_write_reg_wait_helper, 2340 }; 2341 2342 static const struct amdgpu_ring_funcs sdma_v4_0_page_ring_funcs = { 2343 .type = AMDGPU_RING_TYPE_SDMA, 2344 .align_mask = 0xff, 2345 .nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP), 2346 .support_64bit_ptrs = true, 2347 .secure_submission_supported = true, 2348 .get_rptr = sdma_v4_0_ring_get_rptr, 2349 .get_wptr = sdma_v4_0_page_ring_get_wptr, 2350 .set_wptr = sdma_v4_0_page_ring_set_wptr, 2351 .emit_frame_size = 2352 6 + /* sdma_v4_0_ring_emit_hdp_flush */ 2353 3 + /* hdp invalidate */ 2354 6 + /* sdma_v4_0_ring_emit_pipeline_sync */ 2355 /* sdma_v4_0_ring_emit_vm_flush */ 2356 SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 + 2357 SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6 + 2358 10 + 10 + 10, /* sdma_v4_0_ring_emit_fence x3 for user fence, vm fence */ 2359 .emit_ib_size = 7 + 6, /* sdma_v4_0_ring_emit_ib */ 2360 .emit_ib = sdma_v4_0_ring_emit_ib, 2361 .emit_fence = sdma_v4_0_ring_emit_fence, 2362 .emit_pipeline_sync = sdma_v4_0_ring_emit_pipeline_sync, 2363 .emit_vm_flush = sdma_v4_0_ring_emit_vm_flush, 2364 .emit_hdp_flush = sdma_v4_0_ring_emit_hdp_flush, 2365 .test_ring = sdma_v4_0_ring_test_ring, 2366 .test_ib = sdma_v4_0_ring_test_ib, 2367 .insert_nop = sdma_v4_0_ring_insert_nop, 2368 .pad_ib = sdma_v4_0_ring_pad_ib, 2369 .emit_wreg = sdma_v4_0_ring_emit_wreg, 2370 .emit_reg_wait = sdma_v4_0_ring_emit_reg_wait, 2371 .emit_reg_write_reg_wait = amdgpu_ring_emit_reg_write_reg_wait_helper, 2372 }; 2373 2374 static void sdma_v4_0_set_ring_funcs(struct amdgpu_device *adev) 2375 { 2376 int i; 2377 2378 for (i = 0; i < adev->sdma.num_instances; i++) { 2379 adev->sdma.instance[i].ring.funcs = &sdma_v4_0_ring_funcs; 2380 adev->sdma.instance[i].ring.me = i; 2381 if (adev->sdma.has_page_queue) { 2382 adev->sdma.instance[i].page.funcs = 2383 &sdma_v4_0_page_ring_funcs; 2384 adev->sdma.instance[i].page.me = i; 2385 } 2386 } 2387 } 2388 2389 static const struct amdgpu_irq_src_funcs sdma_v4_0_trap_irq_funcs = { 2390 .set = sdma_v4_0_set_trap_irq_state, 2391 .process = sdma_v4_0_process_trap_irq, 2392 }; 2393 2394 static const struct amdgpu_irq_src_funcs sdma_v4_0_illegal_inst_irq_funcs = { 2395 .process = sdma_v4_0_process_illegal_inst_irq, 2396 }; 2397 2398 static const struct amdgpu_irq_src_funcs sdma_v4_0_ecc_irq_funcs = { 2399 .set = sdma_v4_0_set_ecc_irq_state, 2400 .process = amdgpu_sdma_process_ecc_irq, 2401 }; 2402 2403 static const struct amdgpu_irq_src_funcs sdma_v4_0_vm_hole_irq_funcs = { 2404 .process = sdma_v4_0_process_vm_hole_irq, 2405 }; 2406 2407 static const struct amdgpu_irq_src_funcs sdma_v4_0_doorbell_invalid_irq_funcs = { 2408 .process = sdma_v4_0_process_doorbell_invalid_irq, 2409 }; 2410 2411 static const struct amdgpu_irq_src_funcs sdma_v4_0_pool_timeout_irq_funcs = { 2412 .process = sdma_v4_0_process_pool_timeout_irq, 2413 }; 2414 2415 static const struct amdgpu_irq_src_funcs sdma_v4_0_srbm_write_irq_funcs = { 2416 .process = sdma_v4_0_process_srbm_write_irq, 2417 }; 2418 2419 static void sdma_v4_0_set_irq_funcs(struct amdgpu_device *adev) 2420 { 2421 adev->sdma.trap_irq.num_types = adev->sdma.num_instances; 2422 adev->sdma.ecc_irq.num_types = adev->sdma.num_instances; 2423 /*For Arcturus and Aldebaran, add another 4 irq handler*/ 2424 switch (adev->sdma.num_instances) { 2425 case 5: 2426 case 8: 2427 adev->sdma.vm_hole_irq.num_types = adev->sdma.num_instances; 2428 adev->sdma.doorbell_invalid_irq.num_types = adev->sdma.num_instances; 2429 adev->sdma.pool_timeout_irq.num_types = adev->sdma.num_instances; 2430 adev->sdma.srbm_write_irq.num_types = adev->sdma.num_instances; 2431 break; 2432 default: 2433 break; 2434 } 2435 adev->sdma.trap_irq.funcs = &sdma_v4_0_trap_irq_funcs; 2436 adev->sdma.illegal_inst_irq.funcs = &sdma_v4_0_illegal_inst_irq_funcs; 2437 adev->sdma.ecc_irq.funcs = &sdma_v4_0_ecc_irq_funcs; 2438 adev->sdma.vm_hole_irq.funcs = &sdma_v4_0_vm_hole_irq_funcs; 2439 adev->sdma.doorbell_invalid_irq.funcs = &sdma_v4_0_doorbell_invalid_irq_funcs; 2440 adev->sdma.pool_timeout_irq.funcs = &sdma_v4_0_pool_timeout_irq_funcs; 2441 adev->sdma.srbm_write_irq.funcs = &sdma_v4_0_srbm_write_irq_funcs; 2442 } 2443 2444 /** 2445 * sdma_v4_0_emit_copy_buffer - copy buffer using the sDMA engine 2446 * 2447 * @ib: indirect buffer to copy to 2448 * @src_offset: src GPU address 2449 * @dst_offset: dst GPU address 2450 * @byte_count: number of bytes to xfer 2451 * @tmz: if a secure copy should be used 2452 * 2453 * Copy GPU buffers using the DMA engine (VEGA10/12). 2454 * Used by the amdgpu ttm implementation to move pages if 2455 * registered as the asic copy callback. 2456 */ 2457 static void sdma_v4_0_emit_copy_buffer(struct amdgpu_ib *ib, 2458 uint64_t src_offset, 2459 uint64_t dst_offset, 2460 uint32_t byte_count, 2461 bool tmz) 2462 { 2463 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) | 2464 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR) | 2465 SDMA_PKT_COPY_LINEAR_HEADER_TMZ(tmz ? 1 : 0); 2466 ib->ptr[ib->length_dw++] = byte_count - 1; 2467 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */ 2468 ib->ptr[ib->length_dw++] = lower_32_bits(src_offset); 2469 ib->ptr[ib->length_dw++] = upper_32_bits(src_offset); 2470 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset); 2471 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset); 2472 } 2473 2474 /** 2475 * sdma_v4_0_emit_fill_buffer - fill buffer using the sDMA engine 2476 * 2477 * @ib: indirect buffer to copy to 2478 * @src_data: value to write to buffer 2479 * @dst_offset: dst GPU address 2480 * @byte_count: number of bytes to xfer 2481 * 2482 * Fill GPU buffers using the DMA engine (VEGA10/12). 2483 */ 2484 static void sdma_v4_0_emit_fill_buffer(struct amdgpu_ib *ib, 2485 uint32_t src_data, 2486 uint64_t dst_offset, 2487 uint32_t byte_count) 2488 { 2489 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL); 2490 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset); 2491 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset); 2492 ib->ptr[ib->length_dw++] = src_data; 2493 ib->ptr[ib->length_dw++] = byte_count - 1; 2494 } 2495 2496 static const struct amdgpu_buffer_funcs sdma_v4_0_buffer_funcs = { 2497 .copy_max_bytes = 0x400000, 2498 .copy_num_dw = 7, 2499 .emit_copy_buffer = sdma_v4_0_emit_copy_buffer, 2500 2501 .fill_max_bytes = 0x400000, 2502 .fill_num_dw = 5, 2503 .emit_fill_buffer = sdma_v4_0_emit_fill_buffer, 2504 }; 2505 2506 static void sdma_v4_0_set_buffer_funcs(struct amdgpu_device *adev) 2507 { 2508 adev->mman.buffer_funcs = &sdma_v4_0_buffer_funcs; 2509 if (adev->sdma.has_page_queue) 2510 adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].page; 2511 else 2512 adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring; 2513 } 2514 2515 static const struct amdgpu_vm_pte_funcs sdma_v4_0_vm_pte_funcs = { 2516 .copy_pte_num_dw = 7, 2517 .copy_pte = sdma_v4_0_vm_copy_pte, 2518 2519 .write_pte = sdma_v4_0_vm_write_pte, 2520 .set_pte_pde = sdma_v4_0_vm_set_pte_pde, 2521 }; 2522 2523 static void sdma_v4_0_set_vm_pte_funcs(struct amdgpu_device *adev) 2524 { 2525 struct drm_gpu_scheduler *sched; 2526 unsigned i; 2527 2528 adev->vm_manager.vm_pte_funcs = &sdma_v4_0_vm_pte_funcs; 2529 for (i = 0; i < adev->sdma.num_instances; i++) { 2530 if (adev->sdma.has_page_queue) 2531 sched = &adev->sdma.instance[i].page.sched; 2532 else 2533 sched = &adev->sdma.instance[i].ring.sched; 2534 adev->vm_manager.vm_pte_scheds[i] = sched; 2535 } 2536 adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances; 2537 } 2538 2539 static void sdma_v4_0_get_ras_error_count(uint32_t value, 2540 uint32_t instance, 2541 uint32_t *sec_count) 2542 { 2543 uint32_t i; 2544 uint32_t sec_cnt; 2545 2546 /* double bits error (multiple bits) error detection is not supported */ 2547 for (i = 0; i < ARRAY_SIZE(sdma_v4_0_ras_fields); i++) { 2548 /* the SDMA_EDC_COUNTER register in each sdma instance 2549 * shares the same sed shift_mask 2550 * */ 2551 sec_cnt = (value & 2552 sdma_v4_0_ras_fields[i].sec_count_mask) >> 2553 sdma_v4_0_ras_fields[i].sec_count_shift; 2554 if (sec_cnt) { 2555 DRM_INFO("Detected %s in SDMA%d, SED %d\n", 2556 sdma_v4_0_ras_fields[i].name, 2557 instance, sec_cnt); 2558 *sec_count += sec_cnt; 2559 } 2560 } 2561 } 2562 2563 static int sdma_v4_0_query_ras_error_count_by_instance(struct amdgpu_device *adev, 2564 uint32_t instance, void *ras_error_status) 2565 { 2566 struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status; 2567 uint32_t sec_count = 0; 2568 uint32_t reg_value = 0; 2569 2570 reg_value = RREG32_SDMA(instance, mmSDMA0_EDC_COUNTER); 2571 /* double bit error is not supported */ 2572 if (reg_value) 2573 sdma_v4_0_get_ras_error_count(reg_value, 2574 instance, &sec_count); 2575 /* err_data->ce_count should be initialized to 0 2576 * before calling into this function */ 2577 err_data->ce_count += sec_count; 2578 /* double bit error is not supported 2579 * set ue count to 0 */ 2580 err_data->ue_count = 0; 2581 2582 return 0; 2583 }; 2584 2585 static void sdma_v4_0_query_ras_error_count(struct amdgpu_device *adev, void *ras_error_status) 2586 { 2587 int i = 0; 2588 2589 for (i = 0; i < adev->sdma.num_instances; i++) { 2590 if (sdma_v4_0_query_ras_error_count_by_instance(adev, i, ras_error_status)) { 2591 dev_err(adev->dev, "Query ras error count failed in SDMA%d\n", i); 2592 return; 2593 } 2594 } 2595 } 2596 2597 static void sdma_v4_0_reset_ras_error_count(struct amdgpu_device *adev) 2598 { 2599 int i; 2600 2601 /* read back edc counter registers to clear the counters */ 2602 if (amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__SDMA)) { 2603 for (i = 0; i < adev->sdma.num_instances; i++) 2604 RREG32_SDMA(i, mmSDMA0_EDC_COUNTER); 2605 } 2606 } 2607 2608 const struct amdgpu_ras_block_hw_ops sdma_v4_0_ras_hw_ops = { 2609 .query_ras_error_count = sdma_v4_0_query_ras_error_count, 2610 .reset_ras_error_count = sdma_v4_0_reset_ras_error_count, 2611 }; 2612 2613 static struct amdgpu_sdma_ras sdma_v4_0_ras = { 2614 .ras_block = { 2615 .hw_ops = &sdma_v4_0_ras_hw_ops, 2616 .ras_cb = sdma_v4_0_process_ras_data_cb, 2617 }, 2618 }; 2619 2620 static void sdma_v4_0_set_ras_funcs(struct amdgpu_device *adev) 2621 { 2622 switch (amdgpu_ip_version(adev, SDMA0_HWIP, 0)) { 2623 case IP_VERSION(4, 2, 0): 2624 case IP_VERSION(4, 2, 2): 2625 adev->sdma.ras = &sdma_v4_0_ras; 2626 break; 2627 case IP_VERSION(4, 4, 0): 2628 adev->sdma.ras = &sdma_v4_4_ras; 2629 break; 2630 default: 2631 break; 2632 } 2633 } 2634 2635 const struct amdgpu_ip_block_version sdma_v4_0_ip_block = { 2636 .type = AMD_IP_BLOCK_TYPE_SDMA, 2637 .major = 4, 2638 .minor = 0, 2639 .rev = 0, 2640 .funcs = &sdma_v4_0_ip_funcs, 2641 }; 2642