1 /* 2 * Copyright 2022 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 #include "amdgpu.h" 24 #include "amdgpu_xcp.h" 25 #include "gfxhub_v1_2.h" 26 #include "gfxhub_v1_1.h" 27 28 #include "gc/gc_9_4_3_offset.h" 29 #include "gc/gc_9_4_3_sh_mask.h" 30 #include "vega10_enum.h" 31 32 #include "soc15_common.h" 33 34 #define regVM_L2_CNTL3_DEFAULT 0x80100007 35 #define regVM_L2_CNTL4_DEFAULT 0x000000c1 36 37 static u64 gfxhub_v1_2_get_mc_fb_offset(struct amdgpu_device *adev) 38 { 39 return (u64)RREG32_SOC15(GC, GET_INST(GC, 0), regMC_VM_FB_OFFSET) << 24; 40 } 41 42 static void gfxhub_v1_2_xcc_setup_vm_pt_regs(struct amdgpu_device *adev, 43 uint32_t vmid, 44 uint64_t page_table_base, 45 uint32_t xcc_mask) 46 { 47 struct amdgpu_vmhub *hub; 48 int i; 49 50 for_each_inst(i, xcc_mask) { 51 hub = &adev->vmhub[AMDGPU_GFXHUB(i)]; 52 WREG32_SOC15_OFFSET(GC, GET_INST(GC, i), 53 regVM_CONTEXT0_PAGE_TABLE_BASE_ADDR_LO32, 54 hub->ctx_addr_distance * vmid, 55 lower_32_bits(page_table_base)); 56 57 WREG32_SOC15_OFFSET(GC, GET_INST(GC, i), 58 regVM_CONTEXT0_PAGE_TABLE_BASE_ADDR_HI32, 59 hub->ctx_addr_distance * vmid, 60 upper_32_bits(page_table_base)); 61 } 62 } 63 64 static void gfxhub_v1_2_setup_vm_pt_regs(struct amdgpu_device *adev, 65 uint32_t vmid, 66 uint64_t page_table_base) 67 { 68 uint32_t xcc_mask; 69 70 xcc_mask = GENMASK(NUM_XCC(adev->gfx.xcc_mask) - 1, 0); 71 gfxhub_v1_2_xcc_setup_vm_pt_regs(adev, vmid, page_table_base, xcc_mask); 72 } 73 74 static void gfxhub_v1_2_xcc_init_gart_aperture_regs(struct amdgpu_device *adev, 75 uint32_t xcc_mask) 76 { 77 uint64_t gart_start = amdgpu_virt_xgmi_migrate_enabled(adev) ? 78 adev->gmc.vram_start : adev->gmc.fb_start; 79 uint64_t pt_base; 80 int i; 81 82 if (adev->gmc.pdb0_bo) 83 pt_base = amdgpu_gmc_pd_addr(adev->gmc.pdb0_bo); 84 else 85 pt_base = amdgpu_gmc_pd_addr(adev->gart.bo); 86 87 gfxhub_v1_2_xcc_setup_vm_pt_regs(adev, 0, pt_base, xcc_mask); 88 89 /* If use GART for FB translation, vmid0 page table covers both 90 * vram and system memory (gart) 91 */ 92 for_each_inst(i, xcc_mask) { 93 if (adev->gmc.pdb0_bo) { 94 WREG32_SOC15(GC, GET_INST(GC, i), 95 regVM_CONTEXT0_PAGE_TABLE_START_ADDR_LO32, 96 (u32)(gart_start >> 12)); 97 WREG32_SOC15(GC, GET_INST(GC, i), 98 regVM_CONTEXT0_PAGE_TABLE_START_ADDR_HI32, 99 (u32)(gart_start >> 44)); 100 101 WREG32_SOC15(GC, GET_INST(GC, i), 102 regVM_CONTEXT0_PAGE_TABLE_END_ADDR_LO32, 103 (u32)(adev->gmc.gart_end >> 12)); 104 WREG32_SOC15(GC, GET_INST(GC, i), 105 regVM_CONTEXT0_PAGE_TABLE_END_ADDR_HI32, 106 (u32)(adev->gmc.gart_end >> 44)); 107 } else { 108 WREG32_SOC15(GC, GET_INST(GC, i), 109 regVM_CONTEXT0_PAGE_TABLE_START_ADDR_LO32, 110 (u32)(adev->gmc.gart_start >> 12)); 111 WREG32_SOC15(GC, GET_INST(GC, i), 112 regVM_CONTEXT0_PAGE_TABLE_START_ADDR_HI32, 113 (u32)(adev->gmc.gart_start >> 44)); 114 115 WREG32_SOC15(GC, GET_INST(GC, i), 116 regVM_CONTEXT0_PAGE_TABLE_END_ADDR_LO32, 117 (u32)(adev->gmc.gart_end >> 12)); 118 WREG32_SOC15(GC, GET_INST(GC, i), 119 regVM_CONTEXT0_PAGE_TABLE_END_ADDR_HI32, 120 (u32)(adev->gmc.gart_end >> 44)); 121 } 122 } 123 } 124 125 static void 126 gfxhub_v1_2_xcc_init_system_aperture_regs(struct amdgpu_device *adev, 127 uint32_t xcc_mask) 128 { 129 uint64_t value; 130 uint32_t tmp; 131 int i; 132 133 for_each_inst(i, xcc_mask) { 134 /* Program the AGP BAR */ 135 WREG32_SOC15_RLC(GC, GET_INST(GC, i), regMC_VM_AGP_BASE, 0); 136 WREG32_SOC15_RLC(GC, GET_INST(GC, i), regMC_VM_AGP_BOT, adev->gmc.agp_start >> 24); 137 WREG32_SOC15_RLC(GC, GET_INST(GC, i), regMC_VM_AGP_TOP, adev->gmc.agp_end >> 24); 138 139 if (!amdgpu_sriov_vf(adev) || adev->asic_type <= CHIP_VEGA10) { 140 /* Program the system aperture low logical page number. */ 141 WREG32_SOC15_RLC(GC, GET_INST(GC, i), regMC_VM_SYSTEM_APERTURE_LOW_ADDR, 142 min(adev->gmc.fb_start, adev->gmc.agp_start) >> 18); 143 144 if (adev->apu_flags & (AMD_APU_IS_RAVEN2 | 145 AMD_APU_IS_RENOIR | 146 AMD_APU_IS_GREEN_SARDINE)) 147 /* 148 * Raven2 has a HW issue that it is unable to use the 149 * vram which is out of MC_VM_SYSTEM_APERTURE_HIGH_ADDR. 150 * So here is the workaround that increase system 151 * aperture high address (add 1) to get rid of the VM 152 * fault and hardware hang. 153 */ 154 WREG32_SOC15_RLC(GC, GET_INST(GC, i), 155 regMC_VM_SYSTEM_APERTURE_HIGH_ADDR, 156 max((adev->gmc.fb_end >> 18) + 0x1, 157 adev->gmc.agp_end >> 18)); 158 else 159 WREG32_SOC15_RLC(GC, GET_INST(GC, i), 160 regMC_VM_SYSTEM_APERTURE_HIGH_ADDR, 161 max(adev->gmc.fb_end, adev->gmc.agp_end) >> 18); 162 163 /* Set default page address. */ 164 value = amdgpu_gmc_vram_mc2pa(adev, adev->mem_scratch.gpu_addr); 165 WREG32_SOC15(GC, GET_INST(GC, i), regMC_VM_SYSTEM_APERTURE_DEFAULT_ADDR_LSB, 166 (u32)(value >> 12)); 167 WREG32_SOC15(GC, GET_INST(GC, i), regMC_VM_SYSTEM_APERTURE_DEFAULT_ADDR_MSB, 168 (u32)(value >> 44)); 169 170 /* Program "protection fault". */ 171 WREG32_SOC15(GC, GET_INST(GC, i), regVM_L2_PROTECTION_FAULT_DEFAULT_ADDR_LO32, 172 (u32)(adev->dummy_page_addr >> 12)); 173 WREG32_SOC15(GC, GET_INST(GC, i), regVM_L2_PROTECTION_FAULT_DEFAULT_ADDR_HI32, 174 (u32)((u64)adev->dummy_page_addr >> 44)); 175 176 tmp = RREG32_SOC15(GC, GET_INST(GC, i), regVM_L2_PROTECTION_FAULT_CNTL2); 177 tmp = REG_SET_FIELD(tmp, VM_L2_PROTECTION_FAULT_CNTL2, 178 ACTIVE_PAGE_MIGRATION_PTE_READ_RETRY, 1); 179 WREG32_SOC15(GC, GET_INST(GC, i), regVM_L2_PROTECTION_FAULT_CNTL2, tmp); 180 } 181 182 /* In the case squeezing vram into GART aperture, we don't use 183 * FB aperture and AGP aperture. Disable them. 184 */ 185 if (adev->gmc.pdb0_bo && adev->gmc.xgmi.connected_to_cpu) { 186 WREG32_SOC15(GC, GET_INST(GC, i), regMC_VM_FB_LOCATION_TOP, 0); 187 WREG32_SOC15(GC, GET_INST(GC, i), regMC_VM_FB_LOCATION_BASE, 0x00FFFFFF); 188 WREG32_SOC15(GC, GET_INST(GC, i), regMC_VM_AGP_TOP, 0); 189 WREG32_SOC15(GC, GET_INST(GC, i), regMC_VM_AGP_BOT, 0xFFFFFF); 190 WREG32_SOC15(GC, GET_INST(GC, i), regMC_VM_SYSTEM_APERTURE_LOW_ADDR, 0x3FFFFFFF); 191 WREG32_SOC15(GC, GET_INST(GC, i), regMC_VM_SYSTEM_APERTURE_HIGH_ADDR, 0); 192 } 193 } 194 } 195 196 static void gfxhub_v1_2_xcc_init_tlb_regs(struct amdgpu_device *adev, 197 uint32_t xcc_mask) 198 { 199 uint32_t tmp; 200 int i; 201 202 for_each_inst(i, xcc_mask) { 203 /* Setup TLB control */ 204 tmp = RREG32_SOC15(GC, GET_INST(GC, i), regMC_VM_MX_L1_TLB_CNTL); 205 206 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, 207 ENABLE_L1_TLB, 1); 208 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, 209 SYSTEM_ACCESS_MODE, 3); 210 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, 211 ENABLE_ADVANCED_DRIVER_MODEL, 1); 212 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, 213 SYSTEM_APERTURE_UNMAPPED_ACCESS, 0); 214 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, 215 MTYPE, MTYPE_UC);/* XXX for emulation. */ 216 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ATC_EN, 1); 217 218 WREG32_SOC15_RLC(GC, GET_INST(GC, i), regMC_VM_MX_L1_TLB_CNTL, tmp); 219 } 220 } 221 222 static void gfxhub_v1_2_xcc_init_cache_regs(struct amdgpu_device *adev, 223 uint32_t xcc_mask) 224 { 225 uint32_t tmp; 226 int i; 227 228 for_each_inst(i, xcc_mask) { 229 /* Setup L2 cache */ 230 tmp = RREG32_SOC15(GC, GET_INST(GC, i), regVM_L2_CNTL); 231 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_CACHE, 1); 232 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_FRAGMENT_PROCESSING, 1); 233 /* XXX for emulation, Refer to closed source code.*/ 234 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, L2_PDE0_CACHE_TAG_GENERATION_MODE, 235 0); 236 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, PDE_FAULT_CLASSIFICATION, 0); 237 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, CONTEXT1_IDENTITY_ACCESS_MODE, 1); 238 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, IDENTITY_MODE_FRAGMENT_SIZE, 0); 239 WREG32_SOC15_RLC(GC, GET_INST(GC, i), regVM_L2_CNTL, tmp); 240 241 tmp = RREG32_SOC15(GC, GET_INST(GC, i), regVM_L2_CNTL2); 242 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL2, INVALIDATE_ALL_L1_TLBS, 1); 243 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL2, INVALIDATE_L2_CACHE, 1); 244 WREG32_SOC15_RLC(GC, GET_INST(GC, i), regVM_L2_CNTL2, tmp); 245 246 tmp = regVM_L2_CNTL3_DEFAULT; 247 if (adev->gmc.translate_further) { 248 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, BANK_SELECT, 12); 249 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, 250 L2_CACHE_BIGK_FRAGMENT_SIZE, 9); 251 } else { 252 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, BANK_SELECT, 9); 253 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, 254 L2_CACHE_BIGK_FRAGMENT_SIZE, 6); 255 } 256 WREG32_SOC15_RLC(GC, GET_INST(GC, i), regVM_L2_CNTL3, tmp); 257 258 tmp = regVM_L2_CNTL4_DEFAULT; 259 /* For AMD APP APUs setup WC memory */ 260 if (adev->gmc.xgmi.connected_to_cpu || adev->gmc.is_app_apu) { 261 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_PDE_REQUEST_PHYSICAL, 1); 262 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_PTE_REQUEST_PHYSICAL, 1); 263 } else { 264 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_PDE_REQUEST_PHYSICAL, 0); 265 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_PTE_REQUEST_PHYSICAL, 0); 266 } 267 WREG32_SOC15_RLC(GC, GET_INST(GC, i), regVM_L2_CNTL4, tmp); 268 } 269 } 270 271 static void gfxhub_v1_2_xcc_enable_system_domain(struct amdgpu_device *adev, 272 uint32_t xcc_mask) 273 { 274 uint32_t tmp; 275 int i; 276 277 for_each_inst(i, xcc_mask) { 278 tmp = RREG32_SOC15(GC, GET_INST(GC, i), regVM_CONTEXT0_CNTL); 279 tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, ENABLE_CONTEXT, 1); 280 tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, PAGE_TABLE_DEPTH, 281 adev->gmc.vmid0_page_table_depth); 282 tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, PAGE_TABLE_BLOCK_SIZE, 283 adev->gmc.vmid0_page_table_block_size); 284 tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, 285 RETRY_PERMISSION_OR_INVALID_PAGE_FAULT, 0); 286 WREG32_SOC15(GC, GET_INST(GC, i), regVM_CONTEXT0_CNTL, tmp); 287 } 288 } 289 290 static void 291 gfxhub_v1_2_xcc_disable_identity_aperture(struct amdgpu_device *adev, 292 uint32_t xcc_mask) 293 { 294 int i; 295 296 for_each_inst(i, xcc_mask) { 297 WREG32_SOC15(GC, GET_INST(GC, i), 298 regVM_L2_CONTEXT1_IDENTITY_APERTURE_LOW_ADDR_LO32, 299 0XFFFFFFFF); 300 WREG32_SOC15(GC, GET_INST(GC, i), 301 regVM_L2_CONTEXT1_IDENTITY_APERTURE_LOW_ADDR_HI32, 302 0x0000000F); 303 304 WREG32_SOC15(GC, GET_INST(GC, i), 305 regVM_L2_CONTEXT1_IDENTITY_APERTURE_HIGH_ADDR_LO32, 306 0); 307 WREG32_SOC15(GC, GET_INST(GC, i), 308 regVM_L2_CONTEXT1_IDENTITY_APERTURE_HIGH_ADDR_HI32, 309 0); 310 311 WREG32_SOC15(GC, GET_INST(GC, i), 312 regVM_L2_CONTEXT_IDENTITY_PHYSICAL_OFFSET_LO32, 0); 313 WREG32_SOC15(GC, GET_INST(GC, i), 314 regVM_L2_CONTEXT_IDENTITY_PHYSICAL_OFFSET_HI32, 0); 315 } 316 } 317 318 static inline bool 319 gfxhub_v1_2_per_process_xnack_support(struct amdgpu_device *adev) 320 { 321 /* 322 * TODO: Check if this function is really needed, so far only 9.4.3 323 * variants use GFXHUB 1.2 324 */ 325 return !!adev->aid_mask; 326 } 327 328 static void gfxhub_v1_2_xcc_setup_vmid_config(struct amdgpu_device *adev, 329 uint32_t xcc_mask) 330 { 331 struct amdgpu_vmhub *hub; 332 unsigned int num_level, block_size; 333 uint32_t tmp; 334 int i, j; 335 336 num_level = adev->vm_manager.num_level; 337 block_size = adev->vm_manager.block_size; 338 if (adev->gmc.translate_further) 339 num_level -= 1; 340 else 341 block_size -= 9; 342 343 for_each_inst(j, xcc_mask) { 344 hub = &adev->vmhub[AMDGPU_GFXHUB(j)]; 345 for (i = 0; i <= 14; i++) { 346 tmp = RREG32_SOC15_OFFSET(GC, GET_INST(GC, j), regVM_CONTEXT1_CNTL, 347 i * hub->ctx_distance); 348 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, ENABLE_CONTEXT, 1); 349 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PAGE_TABLE_DEPTH, 350 num_level); 351 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 352 RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, 1); 353 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 354 DUMMY_PAGE_PROTECTION_FAULT_ENABLE_DEFAULT, 355 1); 356 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 357 PDE0_PROTECTION_FAULT_ENABLE_DEFAULT, 1); 358 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 359 VALID_PROTECTION_FAULT_ENABLE_DEFAULT, 1); 360 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 361 READ_PROTECTION_FAULT_ENABLE_DEFAULT, 1); 362 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 363 WRITE_PROTECTION_FAULT_ENABLE_DEFAULT, 1); 364 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 365 EXECUTE_PROTECTION_FAULT_ENABLE_DEFAULT, 1); 366 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 367 PAGE_TABLE_BLOCK_SIZE, 368 block_size); 369 /* Send no-retry XNACK on fault to suppress VM fault storm. 370 * On 9.4.3 variants, XNACK can be enabled in 371 * the SQ per-process. 372 * Retry faults need to be enabled for that to work. 373 */ 374 tmp = REG_SET_FIELD( 375 tmp, VM_CONTEXT1_CNTL, 376 RETRY_PERMISSION_OR_INVALID_PAGE_FAULT, 377 !adev->gmc.noretry || 378 gfxhub_v1_2_per_process_xnack_support( 379 adev)); 380 WREG32_SOC15_OFFSET(GC, GET_INST(GC, j), regVM_CONTEXT1_CNTL, 381 i * hub->ctx_distance, tmp); 382 WREG32_SOC15_OFFSET(GC, GET_INST(GC, j), 383 regVM_CONTEXT1_PAGE_TABLE_START_ADDR_LO32, 384 i * hub->ctx_addr_distance, 0); 385 WREG32_SOC15_OFFSET(GC, GET_INST(GC, j), 386 regVM_CONTEXT1_PAGE_TABLE_START_ADDR_HI32, 387 i * hub->ctx_addr_distance, 0); 388 WREG32_SOC15_OFFSET(GC, GET_INST(GC, j), 389 regVM_CONTEXT1_PAGE_TABLE_END_ADDR_LO32, 390 i * hub->ctx_addr_distance, 391 lower_32_bits(adev->vm_manager.max_pfn - 1)); 392 WREG32_SOC15_OFFSET(GC, GET_INST(GC, j), 393 regVM_CONTEXT1_PAGE_TABLE_END_ADDR_HI32, 394 i * hub->ctx_addr_distance, 395 upper_32_bits(adev->vm_manager.max_pfn - 1)); 396 } 397 } 398 } 399 400 static void gfxhub_v1_2_xcc_program_invalidation(struct amdgpu_device *adev, 401 uint32_t xcc_mask) 402 { 403 struct amdgpu_vmhub *hub; 404 unsigned int i, j; 405 406 for_each_inst(j, xcc_mask) { 407 hub = &adev->vmhub[AMDGPU_GFXHUB(j)]; 408 409 for (i = 0 ; i < 18; ++i) { 410 WREG32_SOC15_OFFSET(GC, GET_INST(GC, j), regVM_INVALIDATE_ENG0_ADDR_RANGE_LO32, 411 i * hub->eng_addr_distance, 0xffffffff); 412 WREG32_SOC15_OFFSET(GC, GET_INST(GC, j), regVM_INVALIDATE_ENG0_ADDR_RANGE_HI32, 413 i * hub->eng_addr_distance, 0x1f); 414 } 415 } 416 } 417 418 static int gfxhub_v1_2_xcc_gart_enable(struct amdgpu_device *adev, 419 uint32_t xcc_mask) 420 { 421 /* GART Enable. */ 422 gfxhub_v1_2_xcc_init_gart_aperture_regs(adev, xcc_mask); 423 gfxhub_v1_2_xcc_init_system_aperture_regs(adev, xcc_mask); 424 gfxhub_v1_2_xcc_init_tlb_regs(adev, xcc_mask); 425 if (!amdgpu_sriov_vf(adev)) 426 gfxhub_v1_2_xcc_init_cache_regs(adev, xcc_mask); 427 428 gfxhub_v1_2_xcc_enable_system_domain(adev, xcc_mask); 429 if (!amdgpu_sriov_vf(adev)) 430 gfxhub_v1_2_xcc_disable_identity_aperture(adev, xcc_mask); 431 gfxhub_v1_2_xcc_setup_vmid_config(adev, xcc_mask); 432 gfxhub_v1_2_xcc_program_invalidation(adev, xcc_mask); 433 434 return 0; 435 } 436 437 static int gfxhub_v1_2_gart_enable(struct amdgpu_device *adev) 438 { 439 uint32_t xcc_mask; 440 441 xcc_mask = GENMASK(NUM_XCC(adev->gfx.xcc_mask) - 1, 0); 442 return gfxhub_v1_2_xcc_gart_enable(adev, xcc_mask); 443 } 444 445 static void gfxhub_v1_2_xcc_gart_disable(struct amdgpu_device *adev, 446 uint32_t xcc_mask) 447 { 448 struct amdgpu_vmhub *hub; 449 u32 tmp; 450 u32 i, j; 451 452 for_each_inst(j, xcc_mask) { 453 hub = &adev->vmhub[AMDGPU_GFXHUB(j)]; 454 /* Disable all tables */ 455 for (i = 0; i < 16; i++) 456 WREG32_SOC15_OFFSET(GC, GET_INST(GC, j), regVM_CONTEXT0_CNTL, 457 i * hub->ctx_distance, 0); 458 459 /* Setup TLB control */ 460 tmp = RREG32_SOC15(GC, GET_INST(GC, j), regMC_VM_MX_L1_TLB_CNTL); 461 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_TLB, 0); 462 tmp = REG_SET_FIELD(tmp, 463 MC_VM_MX_L1_TLB_CNTL, 464 ENABLE_ADVANCED_DRIVER_MODEL, 465 0); 466 WREG32_SOC15_RLC(GC, GET_INST(GC, j), regMC_VM_MX_L1_TLB_CNTL, tmp); 467 468 /* Setup L2 cache */ 469 if (!amdgpu_sriov_vf(adev)) { 470 tmp = RREG32_SOC15(GC, GET_INST(GC, j), regVM_L2_CNTL); 471 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_CACHE, 0); 472 WREG32_SOC15(GC, GET_INST(GC, j), regVM_L2_CNTL, tmp); 473 WREG32_SOC15(GC, GET_INST(GC, j), regVM_L2_CNTL3, 0); 474 } 475 } 476 } 477 478 static void gfxhub_v1_2_gart_disable(struct amdgpu_device *adev) 479 { 480 uint32_t xcc_mask; 481 482 xcc_mask = GENMASK(NUM_XCC(adev->gfx.xcc_mask) - 1, 0); 483 gfxhub_v1_2_xcc_gart_disable(adev, xcc_mask); 484 } 485 486 static void gfxhub_v1_2_xcc_set_fault_enable_default(struct amdgpu_device *adev, 487 bool value, 488 uint32_t xcc_mask) 489 { 490 u32 tmp; 491 int i; 492 493 for_each_inst(i, xcc_mask) { 494 tmp = RREG32_SOC15(GC, GET_INST(GC, i), regVM_L2_PROTECTION_FAULT_CNTL); 495 tmp = REG_SET_FIELD(tmp, VM_L2_PROTECTION_FAULT_CNTL, 496 RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, value); 497 tmp = REG_SET_FIELD(tmp, VM_L2_PROTECTION_FAULT_CNTL, 498 PDE0_PROTECTION_FAULT_ENABLE_DEFAULT, value); 499 tmp = REG_SET_FIELD(tmp, VM_L2_PROTECTION_FAULT_CNTL, 500 PDE1_PROTECTION_FAULT_ENABLE_DEFAULT, value); 501 tmp = REG_SET_FIELD(tmp, VM_L2_PROTECTION_FAULT_CNTL, 502 PDE2_PROTECTION_FAULT_ENABLE_DEFAULT, value); 503 tmp = REG_SET_FIELD(tmp, 504 VM_L2_PROTECTION_FAULT_CNTL, 505 TRANSLATE_FURTHER_PROTECTION_FAULT_ENABLE_DEFAULT, 506 value); 507 tmp = REG_SET_FIELD(tmp, VM_L2_PROTECTION_FAULT_CNTL, 508 NACK_PROTECTION_FAULT_ENABLE_DEFAULT, value); 509 tmp = REG_SET_FIELD(tmp, VM_L2_PROTECTION_FAULT_CNTL, 510 DUMMY_PAGE_PROTECTION_FAULT_ENABLE_DEFAULT, value); 511 tmp = REG_SET_FIELD(tmp, VM_L2_PROTECTION_FAULT_CNTL, 512 VALID_PROTECTION_FAULT_ENABLE_DEFAULT, value); 513 tmp = REG_SET_FIELD(tmp, VM_L2_PROTECTION_FAULT_CNTL, 514 READ_PROTECTION_FAULT_ENABLE_DEFAULT, value); 515 tmp = REG_SET_FIELD(tmp, VM_L2_PROTECTION_FAULT_CNTL, 516 WRITE_PROTECTION_FAULT_ENABLE_DEFAULT, value); 517 tmp = REG_SET_FIELD(tmp, VM_L2_PROTECTION_FAULT_CNTL, 518 EXECUTE_PROTECTION_FAULT_ENABLE_DEFAULT, value); 519 if (!value) { 520 tmp = REG_SET_FIELD(tmp, VM_L2_PROTECTION_FAULT_CNTL, 521 CRASH_ON_NO_RETRY_FAULT, 1); 522 tmp = REG_SET_FIELD(tmp, VM_L2_PROTECTION_FAULT_CNTL, 523 CRASH_ON_RETRY_FAULT, 1); 524 } 525 WREG32_SOC15(GC, GET_INST(GC, i), regVM_L2_PROTECTION_FAULT_CNTL, tmp); 526 } 527 } 528 529 /** 530 * gfxhub_v1_2_set_fault_enable_default - update GART/VM fault handling 531 * 532 * @adev: amdgpu_device pointer 533 * @value: true redirects VM faults to the default page 534 */ 535 static void gfxhub_v1_2_set_fault_enable_default(struct amdgpu_device *adev, 536 bool value) 537 { 538 uint32_t xcc_mask; 539 540 xcc_mask = GENMASK(NUM_XCC(adev->gfx.xcc_mask) - 1, 0); 541 gfxhub_v1_2_xcc_set_fault_enable_default(adev, value, xcc_mask); 542 } 543 544 static void gfxhub_v1_2_xcc_init(struct amdgpu_device *adev, uint32_t xcc_mask) 545 { 546 struct amdgpu_vmhub *hub; 547 int i; 548 549 for_each_inst(i, xcc_mask) { 550 hub = &adev->vmhub[AMDGPU_GFXHUB(i)]; 551 552 hub->ctx0_ptb_addr_lo32 = 553 SOC15_REG_OFFSET(GC, GET_INST(GC, i), 554 regVM_CONTEXT0_PAGE_TABLE_BASE_ADDR_LO32); 555 hub->ctx0_ptb_addr_hi32 = 556 SOC15_REG_OFFSET(GC, GET_INST(GC, i), 557 regVM_CONTEXT0_PAGE_TABLE_BASE_ADDR_HI32); 558 hub->vm_inv_eng0_sem = 559 SOC15_REG_OFFSET(GC, GET_INST(GC, i), regVM_INVALIDATE_ENG0_SEM); 560 hub->vm_inv_eng0_req = 561 SOC15_REG_OFFSET(GC, GET_INST(GC, i), regVM_INVALIDATE_ENG0_REQ); 562 hub->vm_inv_eng0_ack = 563 SOC15_REG_OFFSET(GC, GET_INST(GC, i), regVM_INVALIDATE_ENG0_ACK); 564 hub->vm_context0_cntl = 565 SOC15_REG_OFFSET(GC, GET_INST(GC, i), regVM_CONTEXT0_CNTL); 566 hub->vm_l2_pro_fault_status = 567 SOC15_REG_OFFSET(GC, GET_INST(GC, i), 568 regVM_L2_PROTECTION_FAULT_STATUS); 569 hub->vm_l2_pro_fault_cntl = 570 SOC15_REG_OFFSET(GC, GET_INST(GC, i), regVM_L2_PROTECTION_FAULT_CNTL); 571 572 hub->ctx_distance = regVM_CONTEXT1_CNTL - 573 regVM_CONTEXT0_CNTL; 574 hub->ctx_addr_distance = 575 regVM_CONTEXT1_PAGE_TABLE_BASE_ADDR_LO32 - 576 regVM_CONTEXT0_PAGE_TABLE_BASE_ADDR_LO32; 577 hub->eng_distance = regVM_INVALIDATE_ENG1_REQ - 578 regVM_INVALIDATE_ENG0_REQ; 579 hub->eng_addr_distance = 580 regVM_INVALIDATE_ENG1_ADDR_RANGE_LO32 - 581 regVM_INVALIDATE_ENG0_ADDR_RANGE_LO32; 582 } 583 } 584 585 static void gfxhub_v1_2_init(struct amdgpu_device *adev) 586 { 587 uint32_t xcc_mask; 588 589 xcc_mask = GENMASK(NUM_XCC(adev->gfx.xcc_mask) - 1, 0); 590 gfxhub_v1_2_xcc_init(adev, xcc_mask); 591 } 592 593 static int gfxhub_v1_2_get_xgmi_info(struct amdgpu_device *adev) 594 { 595 u32 max_num_physical_nodes; 596 u32 max_physical_node_id; 597 u32 xgmi_lfb_cntl; 598 u32 max_region; 599 u64 seg_size; 600 601 xgmi_lfb_cntl = RREG32_SOC15(GC, GET_INST(GC, 0), regMC_VM_XGMI_LFB_CNTL); 602 seg_size = REG_GET_FIELD( 603 RREG32_SOC15(GC, GET_INST(GC, 0), regMC_VM_XGMI_LFB_SIZE), 604 MC_VM_XGMI_LFB_SIZE, PF_LFB_SIZE) << 24; 605 max_region = 606 REG_GET_FIELD(xgmi_lfb_cntl, MC_VM_XGMI_LFB_CNTL, PF_MAX_REGION); 607 608 609 610 max_num_physical_nodes = 8; 611 max_physical_node_id = 7; 612 613 /* PF_MAX_REGION=0 means xgmi is disabled */ 614 if (max_region || adev->gmc.xgmi.connected_to_cpu) { 615 adev->gmc.xgmi.num_physical_nodes = max_region + 1; 616 617 if (adev->gmc.xgmi.num_physical_nodes > max_num_physical_nodes) 618 return -EINVAL; 619 620 adev->gmc.xgmi.physical_node_id = 621 REG_GET_FIELD(xgmi_lfb_cntl, MC_VM_XGMI_LFB_CNTL, 622 PF_LFB_REGION); 623 624 if (adev->gmc.xgmi.physical_node_id > max_physical_node_id) 625 return -EINVAL; 626 627 adev->gmc.xgmi.node_segment_size = seg_size; 628 } 629 630 return 0; 631 } 632 633 const struct amdgpu_gfxhub_funcs gfxhub_v1_2_funcs = { 634 .get_mc_fb_offset = gfxhub_v1_2_get_mc_fb_offset, 635 .setup_vm_pt_regs = gfxhub_v1_2_setup_vm_pt_regs, 636 .gart_enable = gfxhub_v1_2_gart_enable, 637 .gart_disable = gfxhub_v1_2_gart_disable, 638 .set_fault_enable_default = gfxhub_v1_2_set_fault_enable_default, 639 .init = gfxhub_v1_2_init, 640 .get_xgmi_info = gfxhub_v1_2_get_xgmi_info, 641 }; 642 643 static int gfxhub_v1_2_xcp_resume(void *handle, uint32_t inst_mask) 644 { 645 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 646 bool value; 647 648 if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS) 649 value = false; 650 else 651 value = true; 652 653 gfxhub_v1_2_xcc_set_fault_enable_default(adev, value, inst_mask); 654 655 if (!amdgpu_sriov_vf(adev)) 656 return gfxhub_v1_2_xcc_gart_enable(adev, inst_mask); 657 658 return 0; 659 } 660 661 static int gfxhub_v1_2_xcp_suspend(void *handle, uint32_t inst_mask) 662 { 663 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 664 665 if (!amdgpu_sriov_vf(adev)) 666 gfxhub_v1_2_xcc_gart_disable(adev, inst_mask); 667 668 return 0; 669 } 670 671 struct amdgpu_xcp_ip_funcs gfxhub_v1_2_xcp_funcs = { 672 .suspend = &gfxhub_v1_2_xcp_suspend, 673 .resume = &gfxhub_v1_2_xcp_resume 674 }; 675