1 /* 2 * Copyright 2019 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 "gc/gc_10_1_0_offset.h" 34 #include "gc/gc_10_1_0_sh_mask.h" 35 #include "ivsrcid/sdma0/irqsrcs_sdma0_5_0.h" 36 #include "ivsrcid/sdma1/irqsrcs_sdma1_5_0.h" 37 38 #include "soc15_common.h" 39 #include "soc15.h" 40 #include "navi10_sdma_pkt_open.h" 41 #include "nbio_v2_3.h" 42 #include "sdma_common.h" 43 #include "sdma_v5_0.h" 44 45 MODULE_FIRMWARE("amdgpu/navi10_sdma.bin"); 46 MODULE_FIRMWARE("amdgpu/navi10_sdma1.bin"); 47 48 MODULE_FIRMWARE("amdgpu/navi14_sdma.bin"); 49 MODULE_FIRMWARE("amdgpu/navi14_sdma1.bin"); 50 51 MODULE_FIRMWARE("amdgpu/navi12_sdma.bin"); 52 MODULE_FIRMWARE("amdgpu/navi12_sdma1.bin"); 53 54 MODULE_FIRMWARE("amdgpu/cyan_skillfish2_sdma.bin"); 55 MODULE_FIRMWARE("amdgpu/cyan_skillfish2_sdma1.bin"); 56 57 #define SDMA1_REG_OFFSET 0x600 58 #define SDMA0_HYP_DEC_REG_START 0x5880 59 #define SDMA0_HYP_DEC_REG_END 0x5893 60 #define SDMA1_HYP_DEC_REG_OFFSET 0x20 61 62 static const struct amdgpu_hwip_reg_entry sdma_reg_list_5_0[] = { 63 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_STATUS_REG), 64 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_STATUS1_REG), 65 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_STATUS2_REG), 66 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_STATUS3_REG), 67 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_UCODE_CHECKSUM), 68 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RB_RPTR_FETCH_HI), 69 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RB_RPTR_FETCH), 70 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_UTCL1_RD_STATUS), 71 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_UTCL1_WR_STATUS), 72 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_UTCL1_RD_XNACK0), 73 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_UTCL1_RD_XNACK1), 74 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_UTCL1_WR_XNACK0), 75 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_UTCL1_WR_XNACK1), 76 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_RB_CNTL), 77 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_RB_RPTR), 78 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_RB_RPTR_HI), 79 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_RB_WPTR), 80 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_RB_WPTR_HI), 81 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_IB_OFFSET), 82 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_IB_BASE_LO), 83 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_IB_BASE_HI), 84 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_IB_CNTL), 85 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_IB_RPTR), 86 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_IB_SUB_REMAIN), 87 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_DUMMY_REG), 88 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_RB_CNTL), 89 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_RB_RPTR), 90 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_RB_RPTR_HI), 91 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_RB_WPTR), 92 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_RB_WPTR_HI), 93 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_IB_OFFSET), 94 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_IB_BASE_LO), 95 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_IB_BASE_HI), 96 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_DUMMY_REG), 97 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_RB_CNTL), 98 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_RB_RPTR), 99 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_RB_RPTR_HI), 100 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_RB_WPTR), 101 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_RB_WPTR_HI), 102 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_IB_OFFSET), 103 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_IB_BASE_LO), 104 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_IB_BASE_HI), 105 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_DUMMY_REG), 106 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_INT_STATUS), 107 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_VM_CNTL), 108 SOC15_REG_ENTRY_STR(GC, 0, mmGRBM_STATUS2) 109 }; 110 111 static void sdma_v5_0_set_ring_funcs(struct amdgpu_device *adev); 112 static void sdma_v5_0_set_buffer_funcs(struct amdgpu_device *adev); 113 static void sdma_v5_0_set_vm_pte_funcs(struct amdgpu_device *adev); 114 static void sdma_v5_0_set_irq_funcs(struct amdgpu_device *adev); 115 116 static const struct soc15_reg_golden golden_settings_sdma_5[] = { 117 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_CHICKEN_BITS, 0xffbf1f0f, 0x03ab0107), 118 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 119 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_PAGE_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 120 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 121 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 122 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC2_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 123 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 124 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC4_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 125 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC5_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 126 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC6_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 127 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC7_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 128 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_UTCL1_PAGE, 0x00ffffff, 0x000c5c00), 129 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_CHICKEN_BITS, 0xffbf1f0f, 0x03ab0107), 130 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 131 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_PAGE_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 132 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 133 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 134 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC2_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 135 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 136 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC4_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 137 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC5_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 138 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC6_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 139 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC7_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 140 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_UTCL1_PAGE, 0x00ffffff, 0x000c5c00) 141 }; 142 143 static const struct soc15_reg_golden golden_settings_sdma_5_sriov[] = { 144 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 145 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_PAGE_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 146 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 147 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 148 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC2_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 149 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 150 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC4_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 151 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC5_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 152 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC6_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 153 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC7_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 154 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 155 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_PAGE_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 156 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 157 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 158 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC2_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 159 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 160 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC4_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 161 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC5_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 162 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC6_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 163 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC7_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 164 }; 165 166 static const struct soc15_reg_golden golden_settings_sdma_nv10[] = { 167 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC3_RB_WPTR_POLL_CNTL, 0x0000fff0, 0x00403000), 168 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC3_RB_WPTR_POLL_CNTL, 0x0000fff0, 0x00403000), 169 }; 170 171 static const struct soc15_reg_golden golden_settings_sdma_nv14[] = { 172 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 173 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 174 }; 175 176 static const struct soc15_reg_golden golden_settings_sdma_nv12[] = { 177 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 178 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_GB_ADDR_CONFIG, 0x001877ff, 0x00000044), 179 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_GB_ADDR_CONFIG_READ, 0x001877ff, 0x00000044), 180 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_GB_ADDR_CONFIG, 0x001877ff, 0x00000044), 181 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_GB_ADDR_CONFIG_READ, 0x001877ff, 0x00000044), 182 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 183 }; 184 185 static const struct soc15_reg_golden golden_settings_sdma_cyan_skillfish[] = { 186 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_CHICKEN_BITS, 0xffbf1f0f, 0x03ab0107), 187 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_GB_ADDR_CONFIG, 0x001877ff, 0x00000044), 188 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_GB_ADDR_CONFIG_READ, 0x001877ff, 0x00000044), 189 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 190 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_PAGE_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 191 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 192 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 193 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC2_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 194 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 195 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC4_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 196 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC5_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 197 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC6_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 198 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC7_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 199 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_UTCL1_PAGE, 0x007fffff, 0x004c5c00), 200 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_CHICKEN_BITS, 0xffbf1f0f, 0x03ab0107), 201 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_GB_ADDR_CONFIG, 0x001877ff, 0x00000044), 202 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_GB_ADDR_CONFIG_READ, 0x001877ff, 0x00000044), 203 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 204 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_PAGE_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 205 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 206 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 207 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC2_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 208 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 209 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC4_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 210 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC5_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 211 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC6_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 212 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC7_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 213 SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_UTCL1_PAGE, 0x007fffff, 0x004c5c00) 214 }; 215 216 static u32 sdma_v5_0_get_reg_offset(struct amdgpu_device *adev, u32 instance, u32 internal_offset) 217 { 218 u32 base; 219 220 if (internal_offset >= SDMA0_HYP_DEC_REG_START && 221 internal_offset <= SDMA0_HYP_DEC_REG_END) { 222 base = adev->reg_offset[GC_HWIP][0][1]; 223 if (instance == 1) 224 internal_offset += SDMA1_HYP_DEC_REG_OFFSET; 225 } else { 226 base = adev->reg_offset[GC_HWIP][0][0]; 227 if (instance == 1) 228 internal_offset += SDMA1_REG_OFFSET; 229 } 230 231 return base + internal_offset; 232 } 233 234 static void sdma_v5_0_init_golden_registers(struct amdgpu_device *adev) 235 { 236 switch (amdgpu_ip_version(adev, SDMA0_HWIP, 0)) { 237 case IP_VERSION(5, 0, 0): 238 soc15_program_register_sequence(adev, 239 golden_settings_sdma_5, 240 (const u32)ARRAY_SIZE(golden_settings_sdma_5)); 241 soc15_program_register_sequence(adev, 242 golden_settings_sdma_nv10, 243 (const u32)ARRAY_SIZE(golden_settings_sdma_nv10)); 244 break; 245 case IP_VERSION(5, 0, 2): 246 soc15_program_register_sequence(adev, 247 golden_settings_sdma_5, 248 (const u32)ARRAY_SIZE(golden_settings_sdma_5)); 249 soc15_program_register_sequence(adev, 250 golden_settings_sdma_nv14, 251 (const u32)ARRAY_SIZE(golden_settings_sdma_nv14)); 252 break; 253 case IP_VERSION(5, 0, 5): 254 if (amdgpu_sriov_vf(adev)) 255 soc15_program_register_sequence(adev, 256 golden_settings_sdma_5_sriov, 257 (const u32)ARRAY_SIZE(golden_settings_sdma_5_sriov)); 258 else 259 soc15_program_register_sequence(adev, 260 golden_settings_sdma_5, 261 (const u32)ARRAY_SIZE(golden_settings_sdma_5)); 262 soc15_program_register_sequence(adev, 263 golden_settings_sdma_nv12, 264 (const u32)ARRAY_SIZE(golden_settings_sdma_nv12)); 265 break; 266 case IP_VERSION(5, 0, 1): 267 soc15_program_register_sequence(adev, 268 golden_settings_sdma_cyan_skillfish, 269 (const u32)ARRAY_SIZE(golden_settings_sdma_cyan_skillfish)); 270 break; 271 default: 272 break; 273 } 274 } 275 276 /** 277 * sdma_v5_0_init_microcode - load ucode images from disk 278 * 279 * @adev: amdgpu_device pointer 280 * 281 * Use the firmware interface to load the ucode images into 282 * the driver (not loaded into hw). 283 * Returns 0 on success, error on failure. 284 */ 285 286 // emulation only, won't work on real chip 287 // navi10 real chip need to use PSP to load firmware 288 static int sdma_v5_0_init_microcode(struct amdgpu_device *adev) 289 { 290 int ret, i; 291 292 for (i = 0; i < adev->sdma.num_instances; i++) { 293 ret = amdgpu_sdma_init_microcode(adev, i, false); 294 if (ret) 295 return ret; 296 } 297 298 return ret; 299 } 300 301 static unsigned sdma_v5_0_ring_init_cond_exec(struct amdgpu_ring *ring, 302 uint64_t addr) 303 { 304 unsigned ret; 305 306 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_COND_EXE)); 307 amdgpu_ring_write(ring, lower_32_bits(addr)); 308 amdgpu_ring_write(ring, upper_32_bits(addr)); 309 amdgpu_ring_write(ring, 1); 310 /* this is the offset we need patch later */ 311 ret = ring->wptr & ring->buf_mask; 312 /* insert dummy here and patch it later */ 313 amdgpu_ring_write(ring, 0); 314 315 return ret; 316 } 317 318 /** 319 * sdma_v5_0_ring_get_rptr - get the current read pointer 320 * 321 * @ring: amdgpu ring pointer 322 * 323 * Get the current rptr from the hardware (NAVI10+). 324 */ 325 static uint64_t sdma_v5_0_ring_get_rptr(struct amdgpu_ring *ring) 326 { 327 u64 *rptr; 328 329 /* XXX check if swapping is necessary on BE */ 330 rptr = (u64 *)ring->rptr_cpu_addr; 331 332 DRM_DEBUG("rptr before shift == 0x%016llx\n", *rptr); 333 return ((*rptr) >> 2); 334 } 335 336 /** 337 * sdma_v5_0_ring_get_wptr - get the current write pointer 338 * 339 * @ring: amdgpu ring pointer 340 * 341 * Get the current wptr from the hardware (NAVI10+). 342 */ 343 static uint64_t sdma_v5_0_ring_get_wptr(struct amdgpu_ring *ring) 344 { 345 struct amdgpu_device *adev = ring->adev; 346 u64 wptr; 347 348 if (ring->use_doorbell) { 349 /* XXX check if swapping is necessary on BE */ 350 wptr = READ_ONCE(*((u64 *)ring->wptr_cpu_addr)); 351 DRM_DEBUG("wptr/doorbell before shift == 0x%016llx\n", wptr); 352 } else { 353 wptr = RREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI)); 354 wptr = wptr << 32; 355 wptr |= RREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR)); 356 DRM_DEBUG("wptr before shift [%i] wptr == 0x%016llx\n", ring->me, wptr); 357 } 358 359 return wptr >> 2; 360 } 361 362 /** 363 * sdma_v5_0_ring_set_wptr - commit the write pointer 364 * 365 * @ring: amdgpu ring pointer 366 * 367 * Write the wptr back to the hardware (NAVI10+). 368 */ 369 static void sdma_v5_0_ring_set_wptr(struct amdgpu_ring *ring) 370 { 371 struct amdgpu_device *adev = ring->adev; 372 uint32_t *wptr_saved; 373 uint32_t *is_queue_unmap; 374 uint64_t aggregated_db_index; 375 uint32_t mqd_size = adev->mqds[AMDGPU_HW_IP_DMA].mqd_size; 376 377 DRM_DEBUG("Setting write pointer\n"); 378 if (ring->is_mes_queue) { 379 wptr_saved = (uint32_t *)(ring->mqd_ptr + mqd_size); 380 is_queue_unmap = (uint32_t *)(ring->mqd_ptr + mqd_size + 381 sizeof(uint32_t)); 382 aggregated_db_index = 383 amdgpu_mes_get_aggregated_doorbell_index(adev, 384 AMDGPU_MES_PRIORITY_LEVEL_NORMAL); 385 386 atomic64_set((atomic64_t *)ring->wptr_cpu_addr, 387 ring->wptr << 2); 388 *wptr_saved = ring->wptr << 2; 389 if (*is_queue_unmap) { 390 WDOORBELL64(aggregated_db_index, ring->wptr << 2); 391 DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n", 392 ring->doorbell_index, ring->wptr << 2); 393 WDOORBELL64(ring->doorbell_index, ring->wptr << 2); 394 } else { 395 DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n", 396 ring->doorbell_index, ring->wptr << 2); 397 WDOORBELL64(ring->doorbell_index, ring->wptr << 2); 398 399 if (*is_queue_unmap) 400 WDOORBELL64(aggregated_db_index, 401 ring->wptr << 2); 402 } 403 } else { 404 if (ring->use_doorbell) { 405 DRM_DEBUG("Using doorbell -- " 406 "wptr_offs == 0x%08x " 407 "lower_32_bits(ring->wptr) << 2 == 0x%08x " 408 "upper_32_bits(ring->wptr) << 2 == 0x%08x\n", 409 ring->wptr_offs, 410 lower_32_bits(ring->wptr << 2), 411 upper_32_bits(ring->wptr << 2)); 412 /* XXX check if swapping is necessary on BE */ 413 atomic64_set((atomic64_t *)ring->wptr_cpu_addr, 414 ring->wptr << 2); 415 DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n", 416 ring->doorbell_index, ring->wptr << 2); 417 WDOORBELL64(ring->doorbell_index, ring->wptr << 2); 418 } else { 419 DRM_DEBUG("Not using doorbell -- " 420 "mmSDMA%i_GFX_RB_WPTR == 0x%08x " 421 "mmSDMA%i_GFX_RB_WPTR_HI == 0x%08x\n", 422 ring->me, 423 lower_32_bits(ring->wptr << 2), 424 ring->me, 425 upper_32_bits(ring->wptr << 2)); 426 WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, 427 ring->me, mmSDMA0_GFX_RB_WPTR), 428 lower_32_bits(ring->wptr << 2)); 429 WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, 430 ring->me, mmSDMA0_GFX_RB_WPTR_HI), 431 upper_32_bits(ring->wptr << 2)); 432 } 433 } 434 } 435 436 static void sdma_v5_0_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count) 437 { 438 struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring); 439 int i; 440 441 for (i = 0; i < count; i++) 442 if (sdma && sdma->burst_nop && (i == 0)) 443 amdgpu_ring_write(ring, ring->funcs->nop | 444 SDMA_PKT_NOP_HEADER_COUNT(count - 1)); 445 else 446 amdgpu_ring_write(ring, ring->funcs->nop); 447 } 448 449 /** 450 * sdma_v5_0_ring_emit_ib - Schedule an IB on the DMA engine 451 * 452 * @ring: amdgpu ring pointer 453 * @job: job to retrieve vmid from 454 * @ib: IB object to schedule 455 * @flags: unused 456 * 457 * Schedule an IB in the DMA ring (NAVI10). 458 */ 459 static void sdma_v5_0_ring_emit_ib(struct amdgpu_ring *ring, 460 struct amdgpu_job *job, 461 struct amdgpu_ib *ib, 462 uint32_t flags) 463 { 464 unsigned vmid = AMDGPU_JOB_GET_VMID(job); 465 uint64_t csa_mc_addr = amdgpu_sdma_get_csa_mc_addr(ring, vmid); 466 467 /* An IB packet must end on a 8 DW boundary--the next dword 468 * must be on a 8-dword boundary. Our IB packet below is 6 469 * dwords long, thus add x number of NOPs, such that, in 470 * modular arithmetic, 471 * wptr + 6 + x = 8k, k >= 0, which in C is, 472 * (wptr + 6 + x) % 8 = 0. 473 * The expression below, is a solution of x. 474 */ 475 sdma_v5_0_ring_insert_nop(ring, (2 - lower_32_bits(ring->wptr)) & 7); 476 477 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) | 478 SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf)); 479 /* base must be 32 byte aligned */ 480 amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0); 481 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr)); 482 amdgpu_ring_write(ring, ib->length_dw); 483 amdgpu_ring_write(ring, lower_32_bits(csa_mc_addr)); 484 amdgpu_ring_write(ring, upper_32_bits(csa_mc_addr)); 485 } 486 487 /** 488 * sdma_v5_0_ring_emit_mem_sync - flush the IB by graphics cache rinse 489 * 490 * @ring: amdgpu ring pointer 491 * 492 * flush the IB by graphics cache rinse. 493 */ 494 static void sdma_v5_0_ring_emit_mem_sync(struct amdgpu_ring *ring) 495 { 496 uint32_t gcr_cntl = SDMA_GCR_GL2_INV | SDMA_GCR_GL2_WB | SDMA_GCR_GLM_INV | 497 SDMA_GCR_GL1_INV | SDMA_GCR_GLV_INV | SDMA_GCR_GLK_INV | 498 SDMA_GCR_GLI_INV(1); 499 500 /* flush entire cache L0/L1/L2, this can be optimized by performance requirement */ 501 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_GCR_REQ)); 502 amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD1_BASE_VA_31_7(0)); 503 amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD2_GCR_CONTROL_15_0(gcr_cntl) | 504 SDMA_PKT_GCR_REQ_PAYLOAD2_BASE_VA_47_32(0)); 505 amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD3_LIMIT_VA_31_7(0) | 506 SDMA_PKT_GCR_REQ_PAYLOAD3_GCR_CONTROL_18_16(gcr_cntl >> 16)); 507 amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD4_LIMIT_VA_47_32(0) | 508 SDMA_PKT_GCR_REQ_PAYLOAD4_VMID(0)); 509 } 510 511 /** 512 * sdma_v5_0_ring_emit_hdp_flush - emit an hdp flush on the DMA ring 513 * 514 * @ring: amdgpu ring pointer 515 * 516 * Emit an hdp flush packet on the requested DMA ring. 517 */ 518 static void sdma_v5_0_ring_emit_hdp_flush(struct amdgpu_ring *ring) 519 { 520 struct amdgpu_device *adev = ring->adev; 521 u32 ref_and_mask = 0; 522 const struct nbio_hdp_flush_reg *nbio_hf_reg = adev->nbio.hdp_flush_reg; 523 524 if (ring->me == 0) 525 ref_and_mask = nbio_hf_reg->ref_and_mask_sdma0; 526 else 527 ref_and_mask = nbio_hf_reg->ref_and_mask_sdma1; 528 529 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) | 530 SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) | 531 SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */ 532 amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_done_offset(adev)) << 2); 533 amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_req_offset(adev)) << 2); 534 amdgpu_ring_write(ring, ref_and_mask); /* reference */ 535 amdgpu_ring_write(ring, ref_and_mask); /* mask */ 536 amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) | 537 SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */ 538 } 539 540 /** 541 * sdma_v5_0_ring_emit_fence - emit a fence on the DMA ring 542 * 543 * @ring: amdgpu ring pointer 544 * @addr: address 545 * @seq: sequence number 546 * @flags: fence related flags 547 * 548 * Add a DMA fence packet to the ring to write 549 * the fence seq number and DMA trap packet to generate 550 * an interrupt if needed (NAVI10). 551 */ 552 static void sdma_v5_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq, 553 unsigned flags) 554 { 555 bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT; 556 /* write the fence */ 557 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) | 558 SDMA_PKT_FENCE_HEADER_MTYPE(0x3)); /* Ucached(UC) */ 559 /* zero in first two bits */ 560 BUG_ON(addr & 0x3); 561 amdgpu_ring_write(ring, lower_32_bits(addr)); 562 amdgpu_ring_write(ring, upper_32_bits(addr)); 563 amdgpu_ring_write(ring, lower_32_bits(seq)); 564 565 /* optionally write high bits as well */ 566 if (write64bit) { 567 addr += 4; 568 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) | 569 SDMA_PKT_FENCE_HEADER_MTYPE(0x3)); 570 /* zero in first two bits */ 571 BUG_ON(addr & 0x3); 572 amdgpu_ring_write(ring, lower_32_bits(addr)); 573 amdgpu_ring_write(ring, upper_32_bits(addr)); 574 amdgpu_ring_write(ring, upper_32_bits(seq)); 575 } 576 577 if (flags & AMDGPU_FENCE_FLAG_INT) { 578 uint32_t ctx = ring->is_mes_queue ? 579 (ring->hw_queue_id | AMDGPU_FENCE_MES_QUEUE_FLAG) : 0; 580 /* generate an interrupt */ 581 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP)); 582 amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(ctx)); 583 } 584 } 585 586 587 /** 588 * sdma_v5_0_gfx_stop - stop the gfx async dma engines 589 * 590 * @adev: amdgpu_device pointer 591 * 592 * Stop the gfx async dma ring buffers (NAVI10). 593 */ 594 static void sdma_v5_0_gfx_stop(struct amdgpu_device *adev) 595 { 596 u32 rb_cntl, ib_cntl; 597 int i; 598 599 for (i = 0; i < adev->sdma.num_instances; i++) { 600 rb_cntl = RREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL)); 601 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0); 602 WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl); 603 ib_cntl = RREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL)); 604 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0); 605 WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl); 606 } 607 } 608 609 /** 610 * sdma_v5_0_rlc_stop - stop the compute async dma engines 611 * 612 * @adev: amdgpu_device pointer 613 * 614 * Stop the compute async dma queues (NAVI10). 615 */ 616 static void sdma_v5_0_rlc_stop(struct amdgpu_device *adev) 617 { 618 /* XXX todo */ 619 } 620 621 /** 622 * sdma_v5_0_ctx_switch_enable - stop the async dma engines context switch 623 * 624 * @adev: amdgpu_device pointer 625 * @enable: enable/disable the DMA MEs context switch. 626 * 627 * Halt or unhalt the async dma engines context switch (NAVI10). 628 */ 629 static void sdma_v5_0_ctx_switch_enable(struct amdgpu_device *adev, bool enable) 630 { 631 u32 f32_cntl = 0, phase_quantum = 0; 632 int i; 633 634 if (amdgpu_sdma_phase_quantum) { 635 unsigned value = amdgpu_sdma_phase_quantum; 636 unsigned unit = 0; 637 638 while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >> 639 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) { 640 value = (value + 1) >> 1; 641 unit++; 642 } 643 if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >> 644 SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) { 645 value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >> 646 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT); 647 unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >> 648 SDMA0_PHASE0_QUANTUM__UNIT__SHIFT); 649 WARN_ONCE(1, 650 "clamping sdma_phase_quantum to %uK clock cycles\n", 651 value << unit); 652 } 653 phase_quantum = 654 value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT | 655 unit << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT; 656 } 657 658 for (i = 0; i < adev->sdma.num_instances; i++) { 659 if (!amdgpu_sriov_vf(adev)) { 660 f32_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CNTL)); 661 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL, 662 AUTO_CTXSW_ENABLE, enable ? 1 : 0); 663 } 664 665 if (enable && amdgpu_sdma_phase_quantum) { 666 WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_PHASE0_QUANTUM), 667 phase_quantum); 668 WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_PHASE1_QUANTUM), 669 phase_quantum); 670 WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_PHASE2_QUANTUM), 671 phase_quantum); 672 } 673 if (!amdgpu_sriov_vf(adev)) 674 WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl); 675 } 676 677 } 678 679 /** 680 * sdma_v5_0_enable - stop the async dma engines 681 * 682 * @adev: amdgpu_device pointer 683 * @enable: enable/disable the DMA MEs. 684 * 685 * Halt or unhalt the async dma engines (NAVI10). 686 */ 687 static void sdma_v5_0_enable(struct amdgpu_device *adev, bool enable) 688 { 689 u32 f32_cntl; 690 int i; 691 692 if (!enable) { 693 sdma_v5_0_gfx_stop(adev); 694 sdma_v5_0_rlc_stop(adev); 695 } 696 697 if (amdgpu_sriov_vf(adev)) 698 return; 699 700 for (i = 0; i < adev->sdma.num_instances; i++) { 701 f32_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_F32_CNTL)); 702 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, enable ? 0 : 1); 703 WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl); 704 } 705 } 706 707 /** 708 * sdma_v5_0_gfx_resume_instance - start/restart a certain sdma engine 709 * 710 * @adev: amdgpu_device pointer 711 * @i: instance 712 * @restore: used to restore wptr when restart 713 * 714 * Set up the gfx DMA ring buffers and enable them. On restart, we will restore wptr and rptr. 715 * Return 0 for success. 716 */ 717 static int sdma_v5_0_gfx_resume_instance(struct amdgpu_device *adev, int i, bool restore) 718 { 719 struct amdgpu_ring *ring; 720 u32 rb_cntl, ib_cntl; 721 u32 rb_bufsz; 722 u32 doorbell; 723 u32 doorbell_offset; 724 u32 temp; 725 u32 wptr_poll_cntl; 726 u64 wptr_gpu_addr; 727 728 ring = &adev->sdma.instance[i].ring; 729 730 if (!amdgpu_sriov_vf(adev)) 731 WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL), 0); 732 733 /* Set ring buffer size in dwords */ 734 rb_bufsz = order_base_2(ring->ring_size / 4); 735 rb_cntl = RREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL)); 736 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz); 737 #ifdef __BIG_ENDIAN 738 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1); 739 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, 740 RPTR_WRITEBACK_SWAP_ENABLE, 1); 741 #endif 742 WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl); 743 744 /* Initialize the ring buffer's read and write pointers */ 745 if (restore) { 746 WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR), lower_32_bits(ring->wptr << 2)); 747 WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_HI), upper_32_bits(ring->wptr << 2)); 748 WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), lower_32_bits(ring->wptr << 2)); 749 WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), upper_32_bits(ring->wptr << 2)); 750 } else { 751 WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR), 0); 752 WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_HI), 0); 753 WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), 0); 754 WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), 0); 755 } 756 /* setup the wptr shadow polling */ 757 wptr_gpu_addr = ring->wptr_gpu_addr; 758 WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_LO), 759 lower_32_bits(wptr_gpu_addr)); 760 WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_HI), 761 upper_32_bits(wptr_gpu_addr)); 762 wptr_poll_cntl = RREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, 763 mmSDMA0_GFX_RB_WPTR_POLL_CNTL)); 764 wptr_poll_cntl = REG_SET_FIELD(wptr_poll_cntl, 765 SDMA0_GFX_RB_WPTR_POLL_CNTL, 766 F32_POLL_ENABLE, 1); 767 WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_CNTL), 768 wptr_poll_cntl); 769 770 /* set the wb address whether it's enabled or not */ 771 WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_HI), 772 upper_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFF); 773 WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_LO), 774 lower_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFC); 775 776 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1); 777 778 WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE), 779 ring->gpu_addr >> 8); 780 WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE_HI), 781 ring->gpu_addr >> 40); 782 783 if (!restore) 784 ring->wptr = 0; 785 786 /* before programing wptr to a less value, need set minor_ptr_update first */ 787 WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 1); 788 789 if (!amdgpu_sriov_vf(adev)) { /* only bare-metal use register write for wptr */ 790 WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), 791 lower_32_bits(ring->wptr << 2)); 792 WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), 793 upper_32_bits(ring->wptr << 2)); 794 } 795 796 doorbell = RREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL)); 797 doorbell_offset = RREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, 798 mmSDMA0_GFX_DOORBELL_OFFSET)); 799 800 if (ring->use_doorbell) { 801 doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 1); 802 doorbell_offset = REG_SET_FIELD(doorbell_offset, SDMA0_GFX_DOORBELL_OFFSET, 803 OFFSET, ring->doorbell_index); 804 } else { 805 doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 0); 806 } 807 WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL), doorbell); 808 WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL_OFFSET), 809 doorbell_offset); 810 811 adev->nbio.funcs->sdma_doorbell_range(adev, i, ring->use_doorbell, 812 ring->doorbell_index, 20); 813 814 if (amdgpu_sriov_vf(adev)) 815 sdma_v5_0_ring_set_wptr(ring); 816 817 /* set minor_ptr_update to 0 after wptr programed */ 818 WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 0); 819 820 if (!amdgpu_sriov_vf(adev)) { 821 /* set utc l1 enable flag always to 1 */ 822 temp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CNTL)); 823 temp = REG_SET_FIELD(temp, SDMA0_CNTL, UTC_L1_ENABLE, 1); 824 825 /* enable MCBP */ 826 temp = REG_SET_FIELD(temp, SDMA0_CNTL, MIDCMD_PREEMPT_ENABLE, 1); 827 WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CNTL), temp); 828 829 /* Set up RESP_MODE to non-copy addresses */ 830 temp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL)); 831 temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, RESP_MODE, 3); 832 temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, REDO_DELAY, 9); 833 WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL), temp); 834 835 /* program default cache read and write policy */ 836 temp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE)); 837 /* clean read policy and write policy bits */ 838 temp &= 0xFF0FFF; 839 temp |= ((CACHE_READ_POLICY_L2__DEFAULT << 12) | (CACHE_WRITE_POLICY_L2__DEFAULT << 14)); 840 WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE), temp); 841 } 842 843 if (!amdgpu_sriov_vf(adev)) { 844 /* unhalt engine */ 845 temp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_F32_CNTL)); 846 temp = REG_SET_FIELD(temp, SDMA0_F32_CNTL, HALT, 0); 847 WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), temp); 848 } 849 850 /* enable DMA RB */ 851 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1); 852 WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl); 853 854 ib_cntl = RREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL)); 855 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1); 856 #ifdef __BIG_ENDIAN 857 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1); 858 #endif 859 /* enable DMA IBs */ 860 WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl); 861 862 if (amdgpu_sriov_vf(adev)) { /* bare-metal sequence doesn't need below to lines */ 863 sdma_v5_0_ctx_switch_enable(adev, true); 864 sdma_v5_0_enable(adev, true); 865 } 866 867 return amdgpu_ring_test_helper(ring); 868 } 869 870 /** 871 * sdma_v5_0_gfx_resume - setup and start the async dma engines 872 * 873 * @adev: amdgpu_device pointer 874 * 875 * Set up the gfx DMA ring buffers and enable them (NAVI10). 876 * Returns 0 for success, error for failure. 877 */ 878 static int sdma_v5_0_gfx_resume(struct amdgpu_device *adev) 879 { 880 int i, r; 881 882 for (i = 0; i < adev->sdma.num_instances; i++) { 883 r = sdma_v5_0_gfx_resume_instance(adev, i, false); 884 if (r) 885 return r; 886 } 887 888 return 0; 889 } 890 891 /** 892 * sdma_v5_0_rlc_resume - setup and start the async dma engines 893 * 894 * @adev: amdgpu_device pointer 895 * 896 * Set up the compute DMA queues and enable them (NAVI10). 897 * Returns 0 for success, error for failure. 898 */ 899 static int sdma_v5_0_rlc_resume(struct amdgpu_device *adev) 900 { 901 return 0; 902 } 903 904 /** 905 * sdma_v5_0_load_microcode - load the sDMA ME ucode 906 * 907 * @adev: amdgpu_device pointer 908 * 909 * Loads the sDMA0/1 ucode. 910 * Returns 0 for success, -EINVAL if the ucode is not available. 911 */ 912 static int sdma_v5_0_load_microcode(struct amdgpu_device *adev) 913 { 914 const struct sdma_firmware_header_v1_0 *hdr; 915 const __le32 *fw_data; 916 u32 fw_size; 917 int i, j; 918 919 /* halt the MEs */ 920 sdma_v5_0_enable(adev, false); 921 922 for (i = 0; i < adev->sdma.num_instances; i++) { 923 if (!adev->sdma.instance[i].fw) 924 return -EINVAL; 925 926 hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data; 927 amdgpu_ucode_print_sdma_hdr(&hdr->header); 928 fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4; 929 930 fw_data = (const __le32 *) 931 (adev->sdma.instance[i].fw->data + 932 le32_to_cpu(hdr->header.ucode_array_offset_bytes)); 933 934 WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), 0); 935 936 for (j = 0; j < fw_size; j++) { 937 if (amdgpu_emu_mode == 1 && j % 500 == 0) 938 msleep(1); 939 WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UCODE_DATA), le32_to_cpup(fw_data++)); 940 } 941 942 WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), adev->sdma.instance[i].fw_version); 943 } 944 945 return 0; 946 } 947 948 /** 949 * sdma_v5_0_start - setup and start the async dma engines 950 * 951 * @adev: amdgpu_device pointer 952 * 953 * Set up the DMA engines and enable them (NAVI10). 954 * Returns 0 for success, error for failure. 955 */ 956 static int sdma_v5_0_start(struct amdgpu_device *adev) 957 { 958 int r = 0; 959 960 if (amdgpu_sriov_vf(adev)) { 961 sdma_v5_0_ctx_switch_enable(adev, false); 962 sdma_v5_0_enable(adev, false); 963 964 /* set RB registers */ 965 r = sdma_v5_0_gfx_resume(adev); 966 return r; 967 } 968 969 if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) { 970 r = sdma_v5_0_load_microcode(adev); 971 if (r) 972 return r; 973 } 974 975 /* unhalt the MEs */ 976 sdma_v5_0_enable(adev, true); 977 /* enable sdma ring preemption */ 978 sdma_v5_0_ctx_switch_enable(adev, true); 979 980 /* start the gfx rings and rlc compute queues */ 981 r = sdma_v5_0_gfx_resume(adev); 982 if (r) 983 return r; 984 r = sdma_v5_0_rlc_resume(adev); 985 986 return r; 987 } 988 989 static int sdma_v5_0_mqd_init(struct amdgpu_device *adev, void *mqd, 990 struct amdgpu_mqd_prop *prop) 991 { 992 struct v10_sdma_mqd *m = mqd; 993 uint64_t wb_gpu_addr; 994 995 m->sdmax_rlcx_rb_cntl = 996 order_base_2(prop->queue_size / 4) << SDMA0_RLC0_RB_CNTL__RB_SIZE__SHIFT | 997 1 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_ENABLE__SHIFT | 998 6 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_TIMER__SHIFT | 999 1 << SDMA0_RLC0_RB_CNTL__RB_PRIV__SHIFT; 1000 1001 m->sdmax_rlcx_rb_base = lower_32_bits(prop->hqd_base_gpu_addr >> 8); 1002 m->sdmax_rlcx_rb_base_hi = upper_32_bits(prop->hqd_base_gpu_addr >> 8); 1003 1004 m->sdmax_rlcx_rb_wptr_poll_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, 0, 1005 mmSDMA0_GFX_RB_WPTR_POLL_CNTL)); 1006 1007 wb_gpu_addr = prop->wptr_gpu_addr; 1008 m->sdmax_rlcx_rb_wptr_poll_addr_lo = lower_32_bits(wb_gpu_addr); 1009 m->sdmax_rlcx_rb_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr); 1010 1011 wb_gpu_addr = prop->rptr_gpu_addr; 1012 m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits(wb_gpu_addr); 1013 m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits(wb_gpu_addr); 1014 1015 m->sdmax_rlcx_ib_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, 0, 1016 mmSDMA0_GFX_IB_CNTL)); 1017 1018 m->sdmax_rlcx_doorbell_offset = 1019 prop->doorbell_index << SDMA0_RLC0_DOORBELL_OFFSET__OFFSET__SHIFT; 1020 1021 m->sdmax_rlcx_doorbell = REG_SET_FIELD(0, SDMA0_RLC0_DOORBELL, ENABLE, 1); 1022 1023 return 0; 1024 } 1025 1026 static void sdma_v5_0_set_mqd_funcs(struct amdgpu_device *adev) 1027 { 1028 adev->mqds[AMDGPU_HW_IP_DMA].mqd_size = sizeof(struct v10_sdma_mqd); 1029 adev->mqds[AMDGPU_HW_IP_DMA].init_mqd = sdma_v5_0_mqd_init; 1030 } 1031 1032 /** 1033 * sdma_v5_0_ring_test_ring - simple async dma engine test 1034 * 1035 * @ring: amdgpu_ring structure holding ring information 1036 * 1037 * Test the DMA engine by writing using it to write an 1038 * value to memory. (NAVI10). 1039 * Returns 0 for success, error for failure. 1040 */ 1041 static int sdma_v5_0_ring_test_ring(struct amdgpu_ring *ring) 1042 { 1043 struct amdgpu_device *adev = ring->adev; 1044 unsigned i; 1045 unsigned index; 1046 int r; 1047 u32 tmp; 1048 u64 gpu_addr; 1049 volatile uint32_t *cpu_ptr = NULL; 1050 1051 tmp = 0xCAFEDEAD; 1052 1053 if (ring->is_mes_queue) { 1054 uint32_t offset = 0; 1055 offset = amdgpu_mes_ctx_get_offs(ring, 1056 AMDGPU_MES_CTX_PADDING_OFFS); 1057 gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset); 1058 cpu_ptr = amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset); 1059 *cpu_ptr = tmp; 1060 } else { 1061 r = amdgpu_device_wb_get(adev, &index); 1062 if (r) { 1063 dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r); 1064 return r; 1065 } 1066 1067 gpu_addr = adev->wb.gpu_addr + (index * 4); 1068 adev->wb.wb[index] = cpu_to_le32(tmp); 1069 } 1070 1071 r = amdgpu_ring_alloc(ring, 20); 1072 if (r) { 1073 DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r); 1074 if (!ring->is_mes_queue) 1075 amdgpu_device_wb_free(adev, index); 1076 return r; 1077 } 1078 1079 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) | 1080 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR)); 1081 amdgpu_ring_write(ring, lower_32_bits(gpu_addr)); 1082 amdgpu_ring_write(ring, upper_32_bits(gpu_addr)); 1083 amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0)); 1084 amdgpu_ring_write(ring, 0xDEADBEEF); 1085 amdgpu_ring_commit(ring); 1086 1087 for (i = 0; i < adev->usec_timeout; i++) { 1088 if (ring->is_mes_queue) 1089 tmp = le32_to_cpu(*cpu_ptr); 1090 else 1091 tmp = le32_to_cpu(adev->wb.wb[index]); 1092 if (tmp == 0xDEADBEEF) 1093 break; 1094 if (amdgpu_emu_mode == 1) 1095 msleep(1); 1096 else 1097 udelay(1); 1098 } 1099 1100 if (i >= adev->usec_timeout) 1101 r = -ETIMEDOUT; 1102 1103 if (!ring->is_mes_queue) 1104 amdgpu_device_wb_free(adev, index); 1105 1106 return r; 1107 } 1108 1109 /** 1110 * sdma_v5_0_ring_test_ib - test an IB on the DMA engine 1111 * 1112 * @ring: amdgpu_ring structure holding ring information 1113 * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT 1114 * 1115 * Test a simple IB in the DMA ring (NAVI10). 1116 * Returns 0 on success, error on failure. 1117 */ 1118 static int sdma_v5_0_ring_test_ib(struct amdgpu_ring *ring, long timeout) 1119 { 1120 struct amdgpu_device *adev = ring->adev; 1121 struct amdgpu_ib ib; 1122 struct dma_fence *f = NULL; 1123 unsigned index; 1124 long r; 1125 u32 tmp = 0; 1126 u64 gpu_addr; 1127 volatile uint32_t *cpu_ptr = NULL; 1128 1129 tmp = 0xCAFEDEAD; 1130 memset(&ib, 0, sizeof(ib)); 1131 1132 if (ring->is_mes_queue) { 1133 uint32_t offset = 0; 1134 offset = amdgpu_mes_ctx_get_offs(ring, AMDGPU_MES_CTX_IB_OFFS); 1135 ib.gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset); 1136 ib.ptr = (void *)amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset); 1137 1138 offset = amdgpu_mes_ctx_get_offs(ring, 1139 AMDGPU_MES_CTX_PADDING_OFFS); 1140 gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset); 1141 cpu_ptr = amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset); 1142 *cpu_ptr = tmp; 1143 } else { 1144 r = amdgpu_device_wb_get(adev, &index); 1145 if (r) { 1146 dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r); 1147 return r; 1148 } 1149 1150 gpu_addr = adev->wb.gpu_addr + (index * 4); 1151 adev->wb.wb[index] = cpu_to_le32(tmp); 1152 1153 r = amdgpu_ib_get(adev, NULL, 256, 1154 AMDGPU_IB_POOL_DIRECT, &ib); 1155 if (r) { 1156 DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r); 1157 goto err0; 1158 } 1159 } 1160 1161 ib.ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) | 1162 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR); 1163 ib.ptr[1] = lower_32_bits(gpu_addr); 1164 ib.ptr[2] = upper_32_bits(gpu_addr); 1165 ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0); 1166 ib.ptr[4] = 0xDEADBEEF; 1167 ib.ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP); 1168 ib.ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP); 1169 ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP); 1170 ib.length_dw = 8; 1171 1172 r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f); 1173 if (r) 1174 goto err1; 1175 1176 r = dma_fence_wait_timeout(f, false, timeout); 1177 if (r == 0) { 1178 DRM_ERROR("amdgpu: IB test timed out\n"); 1179 r = -ETIMEDOUT; 1180 goto err1; 1181 } else if (r < 0) { 1182 DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r); 1183 goto err1; 1184 } 1185 1186 if (ring->is_mes_queue) 1187 tmp = le32_to_cpu(*cpu_ptr); 1188 else 1189 tmp = le32_to_cpu(adev->wb.wb[index]); 1190 1191 if (tmp == 0xDEADBEEF) 1192 r = 0; 1193 else 1194 r = -EINVAL; 1195 1196 err1: 1197 amdgpu_ib_free(adev, &ib, NULL); 1198 dma_fence_put(f); 1199 err0: 1200 if (!ring->is_mes_queue) 1201 amdgpu_device_wb_free(adev, index); 1202 return r; 1203 } 1204 1205 1206 /** 1207 * sdma_v5_0_vm_copy_pte - update PTEs by copying them from the GART 1208 * 1209 * @ib: indirect buffer to fill with commands 1210 * @pe: addr of the page entry 1211 * @src: src addr to copy from 1212 * @count: number of page entries to update 1213 * 1214 * Update PTEs by copying them from the GART using sDMA (NAVI10). 1215 */ 1216 static void sdma_v5_0_vm_copy_pte(struct amdgpu_ib *ib, 1217 uint64_t pe, uint64_t src, 1218 unsigned count) 1219 { 1220 unsigned bytes = count * 8; 1221 1222 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) | 1223 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR); 1224 ib->ptr[ib->length_dw++] = bytes - 1; 1225 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */ 1226 ib->ptr[ib->length_dw++] = lower_32_bits(src); 1227 ib->ptr[ib->length_dw++] = upper_32_bits(src); 1228 ib->ptr[ib->length_dw++] = lower_32_bits(pe); 1229 ib->ptr[ib->length_dw++] = upper_32_bits(pe); 1230 1231 } 1232 1233 /** 1234 * sdma_v5_0_vm_write_pte - update PTEs by writing them manually 1235 * 1236 * @ib: indirect buffer to fill with commands 1237 * @pe: addr of the page entry 1238 * @value: dst addr to write into pe 1239 * @count: number of page entries to update 1240 * @incr: increase next addr by incr bytes 1241 * 1242 * Update PTEs by writing them manually using sDMA (NAVI10). 1243 */ 1244 static void sdma_v5_0_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe, 1245 uint64_t value, unsigned count, 1246 uint32_t incr) 1247 { 1248 unsigned ndw = count * 2; 1249 1250 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) | 1251 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR); 1252 ib->ptr[ib->length_dw++] = lower_32_bits(pe); 1253 ib->ptr[ib->length_dw++] = upper_32_bits(pe); 1254 ib->ptr[ib->length_dw++] = ndw - 1; 1255 for (; ndw > 0; ndw -= 2) { 1256 ib->ptr[ib->length_dw++] = lower_32_bits(value); 1257 ib->ptr[ib->length_dw++] = upper_32_bits(value); 1258 value += incr; 1259 } 1260 } 1261 1262 /** 1263 * sdma_v5_0_vm_set_pte_pde - update the page tables using sDMA 1264 * 1265 * @ib: indirect buffer to fill with commands 1266 * @pe: addr of the page entry 1267 * @addr: dst addr to write into pe 1268 * @count: number of page entries to update 1269 * @incr: increase next addr by incr bytes 1270 * @flags: access flags 1271 * 1272 * Update the page tables using sDMA (NAVI10). 1273 */ 1274 static void sdma_v5_0_vm_set_pte_pde(struct amdgpu_ib *ib, 1275 uint64_t pe, 1276 uint64_t addr, unsigned count, 1277 uint32_t incr, uint64_t flags) 1278 { 1279 /* for physically contiguous pages (vram) */ 1280 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_PTEPDE); 1281 ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */ 1282 ib->ptr[ib->length_dw++] = upper_32_bits(pe); 1283 ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */ 1284 ib->ptr[ib->length_dw++] = upper_32_bits(flags); 1285 ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */ 1286 ib->ptr[ib->length_dw++] = upper_32_bits(addr); 1287 ib->ptr[ib->length_dw++] = incr; /* increment size */ 1288 ib->ptr[ib->length_dw++] = 0; 1289 ib->ptr[ib->length_dw++] = count - 1; /* number of entries */ 1290 } 1291 1292 /** 1293 * sdma_v5_0_ring_pad_ib - pad the IB 1294 * @ring: amdgpu_ring structure holding ring information 1295 * @ib: indirect buffer to fill with padding 1296 * 1297 * Pad the IB with NOPs to a boundary multiple of 8. 1298 */ 1299 static void sdma_v5_0_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib) 1300 { 1301 struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring); 1302 u32 pad_count; 1303 int i; 1304 1305 pad_count = (-ib->length_dw) & 0x7; 1306 for (i = 0; i < pad_count; i++) 1307 if (sdma && sdma->burst_nop && (i == 0)) 1308 ib->ptr[ib->length_dw++] = 1309 SDMA_PKT_HEADER_OP(SDMA_OP_NOP) | 1310 SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1); 1311 else 1312 ib->ptr[ib->length_dw++] = 1313 SDMA_PKT_HEADER_OP(SDMA_OP_NOP); 1314 } 1315 1316 1317 /** 1318 * sdma_v5_0_ring_emit_pipeline_sync - sync the pipeline 1319 * 1320 * @ring: amdgpu_ring pointer 1321 * 1322 * Make sure all previous operations are completed (CIK). 1323 */ 1324 static void sdma_v5_0_ring_emit_pipeline_sync(struct amdgpu_ring *ring) 1325 { 1326 uint32_t seq = ring->fence_drv.sync_seq; 1327 uint64_t addr = ring->fence_drv.gpu_addr; 1328 1329 /* wait for idle */ 1330 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) | 1331 SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) | 1332 SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */ 1333 SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1)); 1334 amdgpu_ring_write(ring, addr & 0xfffffffc); 1335 amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff); 1336 amdgpu_ring_write(ring, seq); /* reference */ 1337 amdgpu_ring_write(ring, 0xffffffff); /* mask */ 1338 amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) | 1339 SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */ 1340 } 1341 1342 1343 /** 1344 * sdma_v5_0_ring_emit_vm_flush - vm flush using sDMA 1345 * 1346 * @ring: amdgpu_ring pointer 1347 * @vmid: vmid number to use 1348 * @pd_addr: address 1349 * 1350 * Update the page table base and flush the VM TLB 1351 * using sDMA (NAVI10). 1352 */ 1353 static void sdma_v5_0_ring_emit_vm_flush(struct amdgpu_ring *ring, 1354 unsigned vmid, uint64_t pd_addr) 1355 { 1356 amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr); 1357 } 1358 1359 static void sdma_v5_0_ring_emit_wreg(struct amdgpu_ring *ring, 1360 uint32_t reg, uint32_t val) 1361 { 1362 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) | 1363 SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf)); 1364 amdgpu_ring_write(ring, reg); 1365 amdgpu_ring_write(ring, val); 1366 } 1367 1368 static void sdma_v5_0_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg, 1369 uint32_t val, uint32_t mask) 1370 { 1371 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) | 1372 SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) | 1373 SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* equal */ 1374 amdgpu_ring_write(ring, reg << 2); 1375 amdgpu_ring_write(ring, 0); 1376 amdgpu_ring_write(ring, val); /* reference */ 1377 amdgpu_ring_write(ring, mask); /* mask */ 1378 amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) | 1379 SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); 1380 } 1381 1382 static void sdma_v5_0_ring_emit_reg_write_reg_wait(struct amdgpu_ring *ring, 1383 uint32_t reg0, uint32_t reg1, 1384 uint32_t ref, uint32_t mask) 1385 { 1386 amdgpu_ring_emit_wreg(ring, reg0, ref); 1387 /* wait for a cycle to reset vm_inv_eng*_ack */ 1388 amdgpu_ring_emit_reg_wait(ring, reg0, 0, 0); 1389 amdgpu_ring_emit_reg_wait(ring, reg1, mask, mask); 1390 } 1391 1392 static int sdma_v5_0_early_init(struct amdgpu_ip_block *ip_block) 1393 { 1394 struct amdgpu_device *adev = ip_block->adev; 1395 int r; 1396 1397 r = sdma_v5_0_init_microcode(adev); 1398 if (r) 1399 return r; 1400 1401 sdma_v5_0_set_ring_funcs(adev); 1402 sdma_v5_0_set_buffer_funcs(adev); 1403 sdma_v5_0_set_vm_pte_funcs(adev); 1404 sdma_v5_0_set_irq_funcs(adev); 1405 sdma_v5_0_set_mqd_funcs(adev); 1406 1407 return 0; 1408 } 1409 1410 1411 static int sdma_v5_0_sw_init(struct amdgpu_ip_block *ip_block) 1412 { 1413 struct amdgpu_ring *ring; 1414 int r, i; 1415 struct amdgpu_device *adev = ip_block->adev; 1416 uint32_t reg_count = ARRAY_SIZE(sdma_reg_list_5_0); 1417 uint32_t *ptr; 1418 1419 /* SDMA trap event */ 1420 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_SDMA0, 1421 SDMA0_5_0__SRCID__SDMA_TRAP, 1422 &adev->sdma.trap_irq); 1423 if (r) 1424 return r; 1425 1426 /* SDMA trap event */ 1427 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_SDMA1, 1428 SDMA1_5_0__SRCID__SDMA_TRAP, 1429 &adev->sdma.trap_irq); 1430 if (r) 1431 return r; 1432 1433 for (i = 0; i < adev->sdma.num_instances; i++) { 1434 ring = &adev->sdma.instance[i].ring; 1435 ring->ring_obj = NULL; 1436 ring->use_doorbell = true; 1437 1438 DRM_DEBUG("SDMA %d use_doorbell being set to: [%s]\n", i, 1439 ring->use_doorbell?"true":"false"); 1440 1441 ring->doorbell_index = (i == 0) ? 1442 (adev->doorbell_index.sdma_engine[0] << 1) //get DWORD offset 1443 : (adev->doorbell_index.sdma_engine[1] << 1); // get DWORD offset 1444 1445 ring->vm_hub = AMDGPU_GFXHUB(0); 1446 sprintf(ring->name, "sdma%d", i); 1447 r = amdgpu_ring_init(adev, ring, 1024, &adev->sdma.trap_irq, 1448 (i == 0) ? AMDGPU_SDMA_IRQ_INSTANCE0 : 1449 AMDGPU_SDMA_IRQ_INSTANCE1, 1450 AMDGPU_RING_PRIO_DEFAULT, NULL); 1451 if (r) 1452 return r; 1453 } 1454 1455 adev->sdma.supported_reset = 1456 amdgpu_get_soft_full_reset_mask(&adev->sdma.instance[0].ring); 1457 switch (amdgpu_ip_version(adev, SDMA0_HWIP, 0)) { 1458 case IP_VERSION(5, 0, 0): 1459 case IP_VERSION(5, 0, 2): 1460 case IP_VERSION(5, 0, 5): 1461 if (adev->sdma.instance[0].fw_version >= 35) 1462 adev->sdma.supported_reset |= AMDGPU_RESET_TYPE_PER_QUEUE; 1463 break; 1464 default: 1465 break; 1466 } 1467 1468 /* Allocate memory for SDMA IP Dump buffer */ 1469 ptr = kcalloc(adev->sdma.num_instances * reg_count, sizeof(uint32_t), GFP_KERNEL); 1470 if (ptr) 1471 adev->sdma.ip_dump = ptr; 1472 else 1473 DRM_ERROR("Failed to allocated memory for SDMA IP Dump\n"); 1474 1475 r = amdgpu_sdma_sysfs_reset_mask_init(adev); 1476 if (r) 1477 return r; 1478 1479 return r; 1480 } 1481 1482 static int sdma_v5_0_sw_fini(struct amdgpu_ip_block *ip_block) 1483 { 1484 struct amdgpu_device *adev = ip_block->adev; 1485 int i; 1486 1487 for (i = 0; i < adev->sdma.num_instances; i++) 1488 amdgpu_ring_fini(&adev->sdma.instance[i].ring); 1489 1490 amdgpu_sdma_sysfs_reset_mask_fini(adev); 1491 amdgpu_sdma_destroy_inst_ctx(adev, false); 1492 1493 kfree(adev->sdma.ip_dump); 1494 1495 return 0; 1496 } 1497 1498 static int sdma_v5_0_hw_init(struct amdgpu_ip_block *ip_block) 1499 { 1500 int r; 1501 struct amdgpu_device *adev = ip_block->adev; 1502 1503 sdma_v5_0_init_golden_registers(adev); 1504 1505 r = sdma_v5_0_start(adev); 1506 1507 return r; 1508 } 1509 1510 static int sdma_v5_0_hw_fini(struct amdgpu_ip_block *ip_block) 1511 { 1512 struct amdgpu_device *adev = ip_block->adev; 1513 1514 if (amdgpu_sriov_vf(adev)) 1515 return 0; 1516 1517 sdma_v5_0_ctx_switch_enable(adev, false); 1518 sdma_v5_0_enable(adev, false); 1519 1520 return 0; 1521 } 1522 1523 static int sdma_v5_0_suspend(struct amdgpu_ip_block *ip_block) 1524 { 1525 return sdma_v5_0_hw_fini(ip_block); 1526 } 1527 1528 static int sdma_v5_0_resume(struct amdgpu_ip_block *ip_block) 1529 { 1530 return sdma_v5_0_hw_init(ip_block); 1531 } 1532 1533 static bool sdma_v5_0_is_idle(void *handle) 1534 { 1535 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1536 u32 i; 1537 1538 for (i = 0; i < adev->sdma.num_instances; i++) { 1539 u32 tmp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_STATUS_REG)); 1540 1541 if (!(tmp & SDMA0_STATUS_REG__IDLE_MASK)) 1542 return false; 1543 } 1544 1545 return true; 1546 } 1547 1548 static int sdma_v5_0_wait_for_idle(struct amdgpu_ip_block *ip_block) 1549 { 1550 unsigned i; 1551 u32 sdma0, sdma1; 1552 struct amdgpu_device *adev = ip_block->adev; 1553 1554 for (i = 0; i < adev->usec_timeout; i++) { 1555 sdma0 = RREG32(sdma_v5_0_get_reg_offset(adev, 0, mmSDMA0_STATUS_REG)); 1556 sdma1 = RREG32(sdma_v5_0_get_reg_offset(adev, 1, mmSDMA0_STATUS_REG)); 1557 1558 if (sdma0 & sdma1 & SDMA0_STATUS_REG__IDLE_MASK) 1559 return 0; 1560 udelay(1); 1561 } 1562 return -ETIMEDOUT; 1563 } 1564 1565 static int sdma_v5_0_soft_reset(struct amdgpu_ip_block *ip_block) 1566 { 1567 /* todo */ 1568 1569 return 0; 1570 } 1571 1572 static int sdma_v5_0_reset_queue(struct amdgpu_ring *ring, unsigned int vmid) 1573 { 1574 struct amdgpu_device *adev = ring->adev; 1575 int i, j, r; 1576 u32 rb_cntl, ib_cntl, f32_cntl, freeze, cntl, preempt, soft_reset, stat1_reg; 1577 1578 if (amdgpu_sriov_vf(adev)) 1579 return -EINVAL; 1580 1581 for (i = 0; i < adev->sdma.num_instances; i++) { 1582 if (ring == &adev->sdma.instance[i].ring) 1583 break; 1584 } 1585 1586 if (i == adev->sdma.num_instances) { 1587 DRM_ERROR("sdma instance not found\n"); 1588 return -EINVAL; 1589 } 1590 1591 amdgpu_gfx_rlc_enter_safe_mode(adev, 0); 1592 1593 /* stop queue */ 1594 ib_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL)); 1595 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0); 1596 WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl); 1597 1598 rb_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL)); 1599 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0); 1600 WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl); 1601 1602 /* engine stop SDMA1_F32_CNTL.HALT to 1 and SDMAx_FREEZE freeze bit to 1 */ 1603 freeze = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_FREEZE)); 1604 freeze = REG_SET_FIELD(freeze, SDMA0_FREEZE, FREEZE, 1); 1605 WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_FREEZE), freeze); 1606 1607 for (j = 0; j < adev->usec_timeout; j++) { 1608 freeze = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_FREEZE)); 1609 if (REG_GET_FIELD(freeze, SDMA0_FREEZE, FROZEN) & 1) 1610 break; 1611 udelay(1); 1612 } 1613 1614 /* check sdma copy engine all idle if frozen not received*/ 1615 if (j == adev->usec_timeout) { 1616 stat1_reg = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_STATUS1_REG)); 1617 if ((stat1_reg & 0x3FF) != 0x3FF) { 1618 DRM_ERROR("cannot soft reset as sdma not idle\n"); 1619 r = -ETIMEDOUT; 1620 goto err0; 1621 } 1622 } 1623 1624 f32_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_F32_CNTL)); 1625 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 1); 1626 WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl); 1627 1628 cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CNTL)); 1629 cntl = REG_SET_FIELD(cntl, SDMA0_CNTL, UTC_L1_ENABLE, 0); 1630 WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CNTL), cntl); 1631 1632 /* soft reset SDMA_GFX_PREEMPT.IB_PREEMPT = 0 mmGRBM_SOFT_RESET.SOFT_RESET_SDMA0/1 = 1 */ 1633 preempt = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_PREEMPT)); 1634 preempt = REG_SET_FIELD(preempt, SDMA0_GFX_PREEMPT, IB_PREEMPT, 0); 1635 WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_PREEMPT), preempt); 1636 1637 soft_reset = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET); 1638 soft_reset |= 1 << GRBM_SOFT_RESET__SOFT_RESET_SDMA0__SHIFT << i; 1639 1640 WREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET, soft_reset); 1641 1642 udelay(50); 1643 1644 soft_reset &= ~(1 << GRBM_SOFT_RESET__SOFT_RESET_SDMA0__SHIFT << i); 1645 WREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET, soft_reset); 1646 1647 /* unfreeze*/ 1648 freeze = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_FREEZE)); 1649 freeze = REG_SET_FIELD(freeze, SDMA0_FREEZE, FREEZE, 0); 1650 WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_FREEZE), freeze); 1651 1652 r = sdma_v5_0_gfx_resume_instance(adev, i, true); 1653 1654 err0: 1655 amdgpu_gfx_rlc_exit_safe_mode(adev, 0); 1656 return r; 1657 } 1658 1659 static int sdma_v5_0_ring_preempt_ib(struct amdgpu_ring *ring) 1660 { 1661 int i, r = 0; 1662 struct amdgpu_device *adev = ring->adev; 1663 u32 index = 0; 1664 u64 sdma_gfx_preempt; 1665 1666 amdgpu_sdma_get_index_from_ring(ring, &index); 1667 if (index == 0) 1668 sdma_gfx_preempt = mmSDMA0_GFX_PREEMPT; 1669 else 1670 sdma_gfx_preempt = mmSDMA1_GFX_PREEMPT; 1671 1672 /* assert preemption condition */ 1673 amdgpu_ring_set_preempt_cond_exec(ring, false); 1674 1675 /* emit the trailing fence */ 1676 ring->trail_seq += 1; 1677 amdgpu_ring_alloc(ring, 10); 1678 sdma_v5_0_ring_emit_fence(ring, ring->trail_fence_gpu_addr, 1679 ring->trail_seq, 0); 1680 amdgpu_ring_commit(ring); 1681 1682 /* assert IB preemption */ 1683 WREG32(sdma_gfx_preempt, 1); 1684 1685 /* poll the trailing fence */ 1686 for (i = 0; i < adev->usec_timeout; i++) { 1687 if (ring->trail_seq == 1688 le32_to_cpu(*(ring->trail_fence_cpu_addr))) 1689 break; 1690 udelay(1); 1691 } 1692 1693 if (i >= adev->usec_timeout) { 1694 r = -EINVAL; 1695 DRM_ERROR("ring %d failed to be preempted\n", ring->idx); 1696 } 1697 1698 /* deassert IB preemption */ 1699 WREG32(sdma_gfx_preempt, 0); 1700 1701 /* deassert the preemption condition */ 1702 amdgpu_ring_set_preempt_cond_exec(ring, true); 1703 return r; 1704 } 1705 1706 static int sdma_v5_0_set_trap_irq_state(struct amdgpu_device *adev, 1707 struct amdgpu_irq_src *source, 1708 unsigned type, 1709 enum amdgpu_interrupt_state state) 1710 { 1711 u32 sdma_cntl; 1712 1713 if (!amdgpu_sriov_vf(adev)) { 1714 u32 reg_offset = (type == AMDGPU_SDMA_IRQ_INSTANCE0) ? 1715 sdma_v5_0_get_reg_offset(adev, 0, mmSDMA0_CNTL) : 1716 sdma_v5_0_get_reg_offset(adev, 1, mmSDMA0_CNTL); 1717 1718 sdma_cntl = RREG32(reg_offset); 1719 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 1720 state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0); 1721 WREG32(reg_offset, sdma_cntl); 1722 } 1723 1724 return 0; 1725 } 1726 1727 static int sdma_v5_0_process_trap_irq(struct amdgpu_device *adev, 1728 struct amdgpu_irq_src *source, 1729 struct amdgpu_iv_entry *entry) 1730 { 1731 uint32_t mes_queue_id = entry->src_data[0]; 1732 1733 DRM_DEBUG("IH: SDMA trap\n"); 1734 1735 if (adev->enable_mes && (mes_queue_id & AMDGPU_FENCE_MES_QUEUE_FLAG)) { 1736 struct amdgpu_mes_queue *queue; 1737 1738 mes_queue_id &= AMDGPU_FENCE_MES_QUEUE_ID_MASK; 1739 1740 spin_lock(&adev->mes.queue_id_lock); 1741 queue = idr_find(&adev->mes.queue_id_idr, mes_queue_id); 1742 if (queue) { 1743 DRM_DEBUG("process smda queue id = %d\n", mes_queue_id); 1744 amdgpu_fence_process(queue->ring); 1745 } 1746 spin_unlock(&adev->mes.queue_id_lock); 1747 return 0; 1748 } 1749 1750 switch (entry->client_id) { 1751 case SOC15_IH_CLIENTID_SDMA0: 1752 switch (entry->ring_id) { 1753 case 0: 1754 amdgpu_fence_process(&adev->sdma.instance[0].ring); 1755 break; 1756 case 1: 1757 /* XXX compute */ 1758 break; 1759 case 2: 1760 /* XXX compute */ 1761 break; 1762 case 3: 1763 /* XXX page queue*/ 1764 break; 1765 } 1766 break; 1767 case SOC15_IH_CLIENTID_SDMA1: 1768 switch (entry->ring_id) { 1769 case 0: 1770 amdgpu_fence_process(&adev->sdma.instance[1].ring); 1771 break; 1772 case 1: 1773 /* XXX compute */ 1774 break; 1775 case 2: 1776 /* XXX compute */ 1777 break; 1778 case 3: 1779 /* XXX page queue*/ 1780 break; 1781 } 1782 break; 1783 } 1784 return 0; 1785 } 1786 1787 static int sdma_v5_0_process_illegal_inst_irq(struct amdgpu_device *adev, 1788 struct amdgpu_irq_src *source, 1789 struct amdgpu_iv_entry *entry) 1790 { 1791 return 0; 1792 } 1793 1794 static void sdma_v5_0_update_medium_grain_clock_gating(struct amdgpu_device *adev, 1795 bool enable) 1796 { 1797 uint32_t data, def; 1798 int i; 1799 1800 for (i = 0; i < adev->sdma.num_instances; i++) { 1801 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) { 1802 /* Enable sdma clock gating */ 1803 def = data = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL)); 1804 data &= ~(SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK | 1805 SDMA0_CLK_CTRL__SOFT_OVERRIDE6_MASK | 1806 SDMA0_CLK_CTRL__SOFT_OVERRIDE5_MASK | 1807 SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK | 1808 SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK | 1809 SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK | 1810 SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK | 1811 SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK); 1812 if (def != data) 1813 WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data); 1814 } else { 1815 /* Disable sdma clock gating */ 1816 def = data = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL)); 1817 data |= (SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK | 1818 SDMA0_CLK_CTRL__SOFT_OVERRIDE6_MASK | 1819 SDMA0_CLK_CTRL__SOFT_OVERRIDE5_MASK | 1820 SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK | 1821 SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK | 1822 SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK | 1823 SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK | 1824 SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK); 1825 if (def != data) 1826 WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data); 1827 } 1828 } 1829 } 1830 1831 static void sdma_v5_0_update_medium_grain_light_sleep(struct amdgpu_device *adev, 1832 bool enable) 1833 { 1834 uint32_t data, def; 1835 int i; 1836 1837 for (i = 0; i < adev->sdma.num_instances; i++) { 1838 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) { 1839 /* Enable sdma mem light sleep */ 1840 def = data = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL)); 1841 data |= SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK; 1842 if (def != data) 1843 WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data); 1844 1845 } else { 1846 /* Disable sdma mem light sleep */ 1847 def = data = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL)); 1848 data &= ~SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK; 1849 if (def != data) 1850 WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data); 1851 1852 } 1853 } 1854 } 1855 1856 static int sdma_v5_0_set_clockgating_state(void *handle, 1857 enum amd_clockgating_state state) 1858 { 1859 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1860 1861 if (amdgpu_sriov_vf(adev)) 1862 return 0; 1863 1864 switch (amdgpu_ip_version(adev, SDMA0_HWIP, 0)) { 1865 case IP_VERSION(5, 0, 0): 1866 case IP_VERSION(5, 0, 2): 1867 case IP_VERSION(5, 0, 5): 1868 sdma_v5_0_update_medium_grain_clock_gating(adev, 1869 state == AMD_CG_STATE_GATE); 1870 sdma_v5_0_update_medium_grain_light_sleep(adev, 1871 state == AMD_CG_STATE_GATE); 1872 break; 1873 default: 1874 break; 1875 } 1876 1877 return 0; 1878 } 1879 1880 static int sdma_v5_0_set_powergating_state(void *handle, 1881 enum amd_powergating_state state) 1882 { 1883 return 0; 1884 } 1885 1886 static void sdma_v5_0_get_clockgating_state(void *handle, u64 *flags) 1887 { 1888 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1889 int data; 1890 1891 if (amdgpu_sriov_vf(adev)) 1892 *flags = 0; 1893 1894 /* AMD_CG_SUPPORT_SDMA_MGCG */ 1895 data = RREG32(sdma_v5_0_get_reg_offset(adev, 0, mmSDMA0_CLK_CTRL)); 1896 if (!(data & SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK)) 1897 *flags |= AMD_CG_SUPPORT_SDMA_MGCG; 1898 1899 /* AMD_CG_SUPPORT_SDMA_LS */ 1900 data = RREG32(sdma_v5_0_get_reg_offset(adev, 0, mmSDMA0_POWER_CNTL)); 1901 if (data & SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK) 1902 *flags |= AMD_CG_SUPPORT_SDMA_LS; 1903 } 1904 1905 static void sdma_v5_0_print_ip_state(struct amdgpu_ip_block *ip_block, struct drm_printer *p) 1906 { 1907 struct amdgpu_device *adev = ip_block->adev; 1908 int i, j; 1909 uint32_t reg_count = ARRAY_SIZE(sdma_reg_list_5_0); 1910 uint32_t instance_offset; 1911 1912 if (!adev->sdma.ip_dump) 1913 return; 1914 1915 drm_printf(p, "num_instances:%d\n", adev->sdma.num_instances); 1916 for (i = 0; i < adev->sdma.num_instances; i++) { 1917 instance_offset = i * reg_count; 1918 drm_printf(p, "\nInstance:%d\n", i); 1919 1920 for (j = 0; j < reg_count; j++) 1921 drm_printf(p, "%-50s \t 0x%08x\n", sdma_reg_list_5_0[j].reg_name, 1922 adev->sdma.ip_dump[instance_offset + j]); 1923 } 1924 } 1925 1926 static void sdma_v5_0_dump_ip_state(struct amdgpu_ip_block *ip_block) 1927 { 1928 struct amdgpu_device *adev = ip_block->adev; 1929 int i, j; 1930 uint32_t instance_offset; 1931 uint32_t reg_count = ARRAY_SIZE(sdma_reg_list_5_0); 1932 1933 if (!adev->sdma.ip_dump) 1934 return; 1935 1936 amdgpu_gfx_off_ctrl(adev, false); 1937 for (i = 0; i < adev->sdma.num_instances; i++) { 1938 instance_offset = i * reg_count; 1939 for (j = 0; j < reg_count; j++) 1940 adev->sdma.ip_dump[instance_offset + j] = 1941 RREG32(sdma_v5_0_get_reg_offset(adev, i, 1942 sdma_reg_list_5_0[j].reg_offset)); 1943 } 1944 amdgpu_gfx_off_ctrl(adev, true); 1945 } 1946 1947 static const struct amd_ip_funcs sdma_v5_0_ip_funcs = { 1948 .name = "sdma_v5_0", 1949 .early_init = sdma_v5_0_early_init, 1950 .sw_init = sdma_v5_0_sw_init, 1951 .sw_fini = sdma_v5_0_sw_fini, 1952 .hw_init = sdma_v5_0_hw_init, 1953 .hw_fini = sdma_v5_0_hw_fini, 1954 .suspend = sdma_v5_0_suspend, 1955 .resume = sdma_v5_0_resume, 1956 .is_idle = sdma_v5_0_is_idle, 1957 .wait_for_idle = sdma_v5_0_wait_for_idle, 1958 .soft_reset = sdma_v5_0_soft_reset, 1959 .set_clockgating_state = sdma_v5_0_set_clockgating_state, 1960 .set_powergating_state = sdma_v5_0_set_powergating_state, 1961 .get_clockgating_state = sdma_v5_0_get_clockgating_state, 1962 .dump_ip_state = sdma_v5_0_dump_ip_state, 1963 .print_ip_state = sdma_v5_0_print_ip_state, 1964 }; 1965 1966 static const struct amdgpu_ring_funcs sdma_v5_0_ring_funcs = { 1967 .type = AMDGPU_RING_TYPE_SDMA, 1968 .align_mask = 0xf, 1969 .nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP), 1970 .support_64bit_ptrs = true, 1971 .secure_submission_supported = true, 1972 .get_rptr = sdma_v5_0_ring_get_rptr, 1973 .get_wptr = sdma_v5_0_ring_get_wptr, 1974 .set_wptr = sdma_v5_0_ring_set_wptr, 1975 .emit_frame_size = 1976 5 + /* sdma_v5_0_ring_init_cond_exec */ 1977 6 + /* sdma_v5_0_ring_emit_hdp_flush */ 1978 3 + /* hdp_invalidate */ 1979 6 + /* sdma_v5_0_ring_emit_pipeline_sync */ 1980 /* sdma_v5_0_ring_emit_vm_flush */ 1981 SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 + 1982 SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6 * 2 + 1983 10 + 10 + 10, /* sdma_v5_0_ring_emit_fence x3 for user fence, vm fence */ 1984 .emit_ib_size = 5 + 7 + 6, /* sdma_v5_0_ring_emit_ib */ 1985 .emit_ib = sdma_v5_0_ring_emit_ib, 1986 .emit_mem_sync = sdma_v5_0_ring_emit_mem_sync, 1987 .emit_fence = sdma_v5_0_ring_emit_fence, 1988 .emit_pipeline_sync = sdma_v5_0_ring_emit_pipeline_sync, 1989 .emit_vm_flush = sdma_v5_0_ring_emit_vm_flush, 1990 .emit_hdp_flush = sdma_v5_0_ring_emit_hdp_flush, 1991 .test_ring = sdma_v5_0_ring_test_ring, 1992 .test_ib = sdma_v5_0_ring_test_ib, 1993 .insert_nop = sdma_v5_0_ring_insert_nop, 1994 .pad_ib = sdma_v5_0_ring_pad_ib, 1995 .emit_wreg = sdma_v5_0_ring_emit_wreg, 1996 .emit_reg_wait = sdma_v5_0_ring_emit_reg_wait, 1997 .emit_reg_write_reg_wait = sdma_v5_0_ring_emit_reg_write_reg_wait, 1998 .init_cond_exec = sdma_v5_0_ring_init_cond_exec, 1999 .preempt_ib = sdma_v5_0_ring_preempt_ib, 2000 .reset = sdma_v5_0_reset_queue, 2001 }; 2002 2003 static void sdma_v5_0_set_ring_funcs(struct amdgpu_device *adev) 2004 { 2005 int i; 2006 2007 for (i = 0; i < adev->sdma.num_instances; i++) { 2008 adev->sdma.instance[i].ring.funcs = &sdma_v5_0_ring_funcs; 2009 adev->sdma.instance[i].ring.me = i; 2010 } 2011 } 2012 2013 static const struct amdgpu_irq_src_funcs sdma_v5_0_trap_irq_funcs = { 2014 .set = sdma_v5_0_set_trap_irq_state, 2015 .process = sdma_v5_0_process_trap_irq, 2016 }; 2017 2018 static const struct amdgpu_irq_src_funcs sdma_v5_0_illegal_inst_irq_funcs = { 2019 .process = sdma_v5_0_process_illegal_inst_irq, 2020 }; 2021 2022 static void sdma_v5_0_set_irq_funcs(struct amdgpu_device *adev) 2023 { 2024 adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_INSTANCE0 + 2025 adev->sdma.num_instances; 2026 adev->sdma.trap_irq.funcs = &sdma_v5_0_trap_irq_funcs; 2027 adev->sdma.illegal_inst_irq.funcs = &sdma_v5_0_illegal_inst_irq_funcs; 2028 } 2029 2030 /** 2031 * sdma_v5_0_emit_copy_buffer - copy buffer using the sDMA engine 2032 * 2033 * @ib: indirect buffer to copy to 2034 * @src_offset: src GPU address 2035 * @dst_offset: dst GPU address 2036 * @byte_count: number of bytes to xfer 2037 * @copy_flags: copy flags for the buffers 2038 * 2039 * Copy GPU buffers using the DMA engine (NAVI10). 2040 * Used by the amdgpu ttm implementation to move pages if 2041 * registered as the asic copy callback. 2042 */ 2043 static void sdma_v5_0_emit_copy_buffer(struct amdgpu_ib *ib, 2044 uint64_t src_offset, 2045 uint64_t dst_offset, 2046 uint32_t byte_count, 2047 uint32_t copy_flags) 2048 { 2049 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) | 2050 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR) | 2051 SDMA_PKT_COPY_LINEAR_HEADER_TMZ((copy_flags & AMDGPU_COPY_FLAGS_TMZ) ? 1 : 0); 2052 ib->ptr[ib->length_dw++] = byte_count - 1; 2053 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */ 2054 ib->ptr[ib->length_dw++] = lower_32_bits(src_offset); 2055 ib->ptr[ib->length_dw++] = upper_32_bits(src_offset); 2056 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset); 2057 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset); 2058 } 2059 2060 /** 2061 * sdma_v5_0_emit_fill_buffer - fill buffer using the sDMA engine 2062 * 2063 * @ib: indirect buffer to fill 2064 * @src_data: value to write to buffer 2065 * @dst_offset: dst GPU address 2066 * @byte_count: number of bytes to xfer 2067 * 2068 * Fill GPU buffers using the DMA engine (NAVI10). 2069 */ 2070 static void sdma_v5_0_emit_fill_buffer(struct amdgpu_ib *ib, 2071 uint32_t src_data, 2072 uint64_t dst_offset, 2073 uint32_t byte_count) 2074 { 2075 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL); 2076 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset); 2077 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset); 2078 ib->ptr[ib->length_dw++] = src_data; 2079 ib->ptr[ib->length_dw++] = byte_count - 1; 2080 } 2081 2082 static const struct amdgpu_buffer_funcs sdma_v5_0_buffer_funcs = { 2083 .copy_max_bytes = 0x400000, 2084 .copy_num_dw = 7, 2085 .emit_copy_buffer = sdma_v5_0_emit_copy_buffer, 2086 2087 .fill_max_bytes = 0x400000, 2088 .fill_num_dw = 5, 2089 .emit_fill_buffer = sdma_v5_0_emit_fill_buffer, 2090 }; 2091 2092 static void sdma_v5_0_set_buffer_funcs(struct amdgpu_device *adev) 2093 { 2094 if (adev->mman.buffer_funcs == NULL) { 2095 adev->mman.buffer_funcs = &sdma_v5_0_buffer_funcs; 2096 adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring; 2097 } 2098 } 2099 2100 static const struct amdgpu_vm_pte_funcs sdma_v5_0_vm_pte_funcs = { 2101 .copy_pte_num_dw = 7, 2102 .copy_pte = sdma_v5_0_vm_copy_pte, 2103 .write_pte = sdma_v5_0_vm_write_pte, 2104 .set_pte_pde = sdma_v5_0_vm_set_pte_pde, 2105 }; 2106 2107 static void sdma_v5_0_set_vm_pte_funcs(struct amdgpu_device *adev) 2108 { 2109 unsigned i; 2110 2111 if (adev->vm_manager.vm_pte_funcs == NULL) { 2112 adev->vm_manager.vm_pte_funcs = &sdma_v5_0_vm_pte_funcs; 2113 for (i = 0; i < adev->sdma.num_instances; i++) { 2114 adev->vm_manager.vm_pte_scheds[i] = 2115 &adev->sdma.instance[i].ring.sched; 2116 } 2117 adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances; 2118 } 2119 } 2120 2121 const struct amdgpu_ip_block_version sdma_v5_0_ip_block = { 2122 .type = AMD_IP_BLOCK_TYPE_SDMA, 2123 .major = 5, 2124 .minor = 0, 2125 .rev = 0, 2126 .funcs = &sdma_v5_0_ip_funcs, 2127 }; 2128