1 /* 2 * Copyright 2016 Advanced Micro Devices, Inc. 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice shall be included in 12 * all copies or substantial portions of the Software. 13 * 14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 20 * OTHER DEALINGS IN THE SOFTWARE. 21 * 22 */ 23 24 #include <linux/firmware.h> 25 #include <linux/pci.h> 26 27 #include <drm/drm_cache.h> 28 29 #include "amdgpu.h" 30 #include "gmc_v9_0.h" 31 #include "amdgpu_atomfirmware.h" 32 #include "amdgpu_gem.h" 33 34 #include "gc/gc_9_0_sh_mask.h" 35 #include "dce/dce_12_0_offset.h" 36 #include "dce/dce_12_0_sh_mask.h" 37 #include "vega10_enum.h" 38 #include "mmhub/mmhub_1_0_offset.h" 39 #include "athub/athub_1_0_sh_mask.h" 40 #include "athub/athub_1_0_offset.h" 41 #include "oss/osssys_4_0_offset.h" 42 43 #include "soc15.h" 44 #include "soc15d.h" 45 #include "soc15_common.h" 46 #include "umc/umc_6_0_sh_mask.h" 47 48 #include "gfxhub_v1_0.h" 49 #include "mmhub_v1_0.h" 50 #include "athub_v1_0.h" 51 #include "gfxhub_v1_1.h" 52 #include "mmhub_v9_4.h" 53 #include "mmhub_v1_7.h" 54 #include "umc_v6_1.h" 55 #include "umc_v6_0.h" 56 #include "umc_v6_7.h" 57 #include "hdp_v4_0.h" 58 #include "mca_v3_0.h" 59 60 #include "ivsrcid/vmc/irqsrcs_vmc_1_0.h" 61 62 #include "amdgpu_ras.h" 63 #include "amdgpu_xgmi.h" 64 65 #include "amdgpu_reset.h" 66 67 /* add these here since we already include dce12 headers and these are for DCN */ 68 #define mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION 0x055d 69 #define mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION_BASE_IDX 2 70 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_WIDTH__SHIFT 0x0 71 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_HEIGHT__SHIFT 0x10 72 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_WIDTH_MASK 0x00003FFFL 73 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_HEIGHT_MASK 0x3FFF0000L 74 #define mmDCHUBBUB_SDPIF_MMIO_CNTRL_0 0x049d 75 #define mmDCHUBBUB_SDPIF_MMIO_CNTRL_0_BASE_IDX 2 76 77 #define mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION_DCN2 0x05ea 78 #define mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION_DCN2_BASE_IDX 2 79 80 81 static const char *gfxhub_client_ids[] = { 82 "CB", 83 "DB", 84 "IA", 85 "WD", 86 "CPF", 87 "CPC", 88 "CPG", 89 "RLC", 90 "TCP", 91 "SQC (inst)", 92 "SQC (data)", 93 "SQG", 94 "PA", 95 }; 96 97 static const char *mmhub_client_ids_raven[][2] = { 98 [0][0] = "MP1", 99 [1][0] = "MP0", 100 [2][0] = "VCN", 101 [3][0] = "VCNU", 102 [4][0] = "HDP", 103 [5][0] = "DCE", 104 [13][0] = "UTCL2", 105 [19][0] = "TLS", 106 [26][0] = "OSS", 107 [27][0] = "SDMA0", 108 [0][1] = "MP1", 109 [1][1] = "MP0", 110 [2][1] = "VCN", 111 [3][1] = "VCNU", 112 [4][1] = "HDP", 113 [5][1] = "XDP", 114 [6][1] = "DBGU0", 115 [7][1] = "DCE", 116 [8][1] = "DCEDWB0", 117 [9][1] = "DCEDWB1", 118 [26][1] = "OSS", 119 [27][1] = "SDMA0", 120 }; 121 122 static const char *mmhub_client_ids_renoir[][2] = { 123 [0][0] = "MP1", 124 [1][0] = "MP0", 125 [2][0] = "HDP", 126 [4][0] = "DCEDMC", 127 [5][0] = "DCEVGA", 128 [13][0] = "UTCL2", 129 [19][0] = "TLS", 130 [26][0] = "OSS", 131 [27][0] = "SDMA0", 132 [28][0] = "VCN", 133 [29][0] = "VCNU", 134 [30][0] = "JPEG", 135 [0][1] = "MP1", 136 [1][1] = "MP0", 137 [2][1] = "HDP", 138 [3][1] = "XDP", 139 [6][1] = "DBGU0", 140 [7][1] = "DCEDMC", 141 [8][1] = "DCEVGA", 142 [9][1] = "DCEDWB", 143 [26][1] = "OSS", 144 [27][1] = "SDMA0", 145 [28][1] = "VCN", 146 [29][1] = "VCNU", 147 [30][1] = "JPEG", 148 }; 149 150 static const char *mmhub_client_ids_vega10[][2] = { 151 [0][0] = "MP0", 152 [1][0] = "UVD", 153 [2][0] = "UVDU", 154 [3][0] = "HDP", 155 [13][0] = "UTCL2", 156 [14][0] = "OSS", 157 [15][0] = "SDMA1", 158 [32+0][0] = "VCE0", 159 [32+1][0] = "VCE0U", 160 [32+2][0] = "XDMA", 161 [32+3][0] = "DCE", 162 [32+4][0] = "MP1", 163 [32+14][0] = "SDMA0", 164 [0][1] = "MP0", 165 [1][1] = "UVD", 166 [2][1] = "UVDU", 167 [3][1] = "DBGU0", 168 [4][1] = "HDP", 169 [5][1] = "XDP", 170 [14][1] = "OSS", 171 [15][1] = "SDMA0", 172 [32+0][1] = "VCE0", 173 [32+1][1] = "VCE0U", 174 [32+2][1] = "XDMA", 175 [32+3][1] = "DCE", 176 [32+4][1] = "DCEDWB", 177 [32+5][1] = "MP1", 178 [32+6][1] = "DBGU1", 179 [32+14][1] = "SDMA1", 180 }; 181 182 static const char *mmhub_client_ids_vega12[][2] = { 183 [0][0] = "MP0", 184 [1][0] = "VCE0", 185 [2][0] = "VCE0U", 186 [3][0] = "HDP", 187 [13][0] = "UTCL2", 188 [14][0] = "OSS", 189 [15][0] = "SDMA1", 190 [32+0][0] = "DCE", 191 [32+1][0] = "XDMA", 192 [32+2][0] = "UVD", 193 [32+3][0] = "UVDU", 194 [32+4][0] = "MP1", 195 [32+15][0] = "SDMA0", 196 [0][1] = "MP0", 197 [1][1] = "VCE0", 198 [2][1] = "VCE0U", 199 [3][1] = "DBGU0", 200 [4][1] = "HDP", 201 [5][1] = "XDP", 202 [14][1] = "OSS", 203 [15][1] = "SDMA0", 204 [32+0][1] = "DCE", 205 [32+1][1] = "DCEDWB", 206 [32+2][1] = "XDMA", 207 [32+3][1] = "UVD", 208 [32+4][1] = "UVDU", 209 [32+5][1] = "MP1", 210 [32+6][1] = "DBGU1", 211 [32+15][1] = "SDMA1", 212 }; 213 214 static const char *mmhub_client_ids_vega20[][2] = { 215 [0][0] = "XDMA", 216 [1][0] = "DCE", 217 [2][0] = "VCE0", 218 [3][0] = "VCE0U", 219 [4][0] = "UVD", 220 [5][0] = "UVD1U", 221 [13][0] = "OSS", 222 [14][0] = "HDP", 223 [15][0] = "SDMA0", 224 [32+0][0] = "UVD", 225 [32+1][0] = "UVDU", 226 [32+2][0] = "MP1", 227 [32+3][0] = "MP0", 228 [32+12][0] = "UTCL2", 229 [32+14][0] = "SDMA1", 230 [0][1] = "XDMA", 231 [1][1] = "DCE", 232 [2][1] = "DCEDWB", 233 [3][1] = "VCE0", 234 [4][1] = "VCE0U", 235 [5][1] = "UVD1", 236 [6][1] = "UVD1U", 237 [7][1] = "DBGU0", 238 [8][1] = "XDP", 239 [13][1] = "OSS", 240 [14][1] = "HDP", 241 [15][1] = "SDMA0", 242 [32+0][1] = "UVD", 243 [32+1][1] = "UVDU", 244 [32+2][1] = "DBGU1", 245 [32+3][1] = "MP1", 246 [32+4][1] = "MP0", 247 [32+14][1] = "SDMA1", 248 }; 249 250 static const char *mmhub_client_ids_arcturus[][2] = { 251 [0][0] = "DBGU1", 252 [1][0] = "XDP", 253 [2][0] = "MP1", 254 [14][0] = "HDP", 255 [171][0] = "JPEG", 256 [172][0] = "VCN", 257 [173][0] = "VCNU", 258 [203][0] = "JPEG1", 259 [204][0] = "VCN1", 260 [205][0] = "VCN1U", 261 [256][0] = "SDMA0", 262 [257][0] = "SDMA1", 263 [258][0] = "SDMA2", 264 [259][0] = "SDMA3", 265 [260][0] = "SDMA4", 266 [261][0] = "SDMA5", 267 [262][0] = "SDMA6", 268 [263][0] = "SDMA7", 269 [384][0] = "OSS", 270 [0][1] = "DBGU1", 271 [1][1] = "XDP", 272 [2][1] = "MP1", 273 [14][1] = "HDP", 274 [171][1] = "JPEG", 275 [172][1] = "VCN", 276 [173][1] = "VCNU", 277 [203][1] = "JPEG1", 278 [204][1] = "VCN1", 279 [205][1] = "VCN1U", 280 [256][1] = "SDMA0", 281 [257][1] = "SDMA1", 282 [258][1] = "SDMA2", 283 [259][1] = "SDMA3", 284 [260][1] = "SDMA4", 285 [261][1] = "SDMA5", 286 [262][1] = "SDMA6", 287 [263][1] = "SDMA7", 288 [384][1] = "OSS", 289 }; 290 291 static const char *mmhub_client_ids_aldebaran[][2] = { 292 [2][0] = "MP1", 293 [3][0] = "MP0", 294 [32+1][0] = "DBGU_IO0", 295 [32+2][0] = "DBGU_IO2", 296 [32+4][0] = "MPIO", 297 [96+11][0] = "JPEG0", 298 [96+12][0] = "VCN0", 299 [96+13][0] = "VCNU0", 300 [128+11][0] = "JPEG1", 301 [128+12][0] = "VCN1", 302 [128+13][0] = "VCNU1", 303 [160+1][0] = "XDP", 304 [160+14][0] = "HDP", 305 [256+0][0] = "SDMA0", 306 [256+1][0] = "SDMA1", 307 [256+2][0] = "SDMA2", 308 [256+3][0] = "SDMA3", 309 [256+4][0] = "SDMA4", 310 [384+0][0] = "OSS", 311 [2][1] = "MP1", 312 [3][1] = "MP0", 313 [32+1][1] = "DBGU_IO0", 314 [32+2][1] = "DBGU_IO2", 315 [32+4][1] = "MPIO", 316 [96+11][1] = "JPEG0", 317 [96+12][1] = "VCN0", 318 [96+13][1] = "VCNU0", 319 [128+11][1] = "JPEG1", 320 [128+12][1] = "VCN1", 321 [128+13][1] = "VCNU1", 322 [160+1][1] = "XDP", 323 [160+14][1] = "HDP", 324 [256+0][1] = "SDMA0", 325 [256+1][1] = "SDMA1", 326 [256+2][1] = "SDMA2", 327 [256+3][1] = "SDMA3", 328 [256+4][1] = "SDMA4", 329 [384+0][1] = "OSS", 330 }; 331 332 static const struct soc15_reg_golden golden_settings_mmhub_1_0_0[] = 333 { 334 SOC15_REG_GOLDEN_VALUE(MMHUB, 0, mmDAGB1_WRCLI2, 0x00000007, 0xfe5fe0fa), 335 SOC15_REG_GOLDEN_VALUE(MMHUB, 0, mmMMEA1_DRAM_WR_CLI2GRP_MAP0, 0x00000030, 0x55555565) 336 }; 337 338 static const struct soc15_reg_golden golden_settings_athub_1_0_0[] = 339 { 340 SOC15_REG_GOLDEN_VALUE(ATHUB, 0, mmRPB_ARB_CNTL, 0x0000ff00, 0x00000800), 341 SOC15_REG_GOLDEN_VALUE(ATHUB, 0, mmRPB_ARB_CNTL2, 0x00ff00ff, 0x00080008) 342 }; 343 344 static const uint32_t ecc_umc_mcumc_ctrl_addrs[] = { 345 (0x000143c0 + 0x00000000), 346 (0x000143c0 + 0x00000800), 347 (0x000143c0 + 0x00001000), 348 (0x000143c0 + 0x00001800), 349 (0x000543c0 + 0x00000000), 350 (0x000543c0 + 0x00000800), 351 (0x000543c0 + 0x00001000), 352 (0x000543c0 + 0x00001800), 353 (0x000943c0 + 0x00000000), 354 (0x000943c0 + 0x00000800), 355 (0x000943c0 + 0x00001000), 356 (0x000943c0 + 0x00001800), 357 (0x000d43c0 + 0x00000000), 358 (0x000d43c0 + 0x00000800), 359 (0x000d43c0 + 0x00001000), 360 (0x000d43c0 + 0x00001800), 361 (0x001143c0 + 0x00000000), 362 (0x001143c0 + 0x00000800), 363 (0x001143c0 + 0x00001000), 364 (0x001143c0 + 0x00001800), 365 (0x001543c0 + 0x00000000), 366 (0x001543c0 + 0x00000800), 367 (0x001543c0 + 0x00001000), 368 (0x001543c0 + 0x00001800), 369 (0x001943c0 + 0x00000000), 370 (0x001943c0 + 0x00000800), 371 (0x001943c0 + 0x00001000), 372 (0x001943c0 + 0x00001800), 373 (0x001d43c0 + 0x00000000), 374 (0x001d43c0 + 0x00000800), 375 (0x001d43c0 + 0x00001000), 376 (0x001d43c0 + 0x00001800), 377 }; 378 379 static const uint32_t ecc_umc_mcumc_ctrl_mask_addrs[] = { 380 (0x000143e0 + 0x00000000), 381 (0x000143e0 + 0x00000800), 382 (0x000143e0 + 0x00001000), 383 (0x000143e0 + 0x00001800), 384 (0x000543e0 + 0x00000000), 385 (0x000543e0 + 0x00000800), 386 (0x000543e0 + 0x00001000), 387 (0x000543e0 + 0x00001800), 388 (0x000943e0 + 0x00000000), 389 (0x000943e0 + 0x00000800), 390 (0x000943e0 + 0x00001000), 391 (0x000943e0 + 0x00001800), 392 (0x000d43e0 + 0x00000000), 393 (0x000d43e0 + 0x00000800), 394 (0x000d43e0 + 0x00001000), 395 (0x000d43e0 + 0x00001800), 396 (0x001143e0 + 0x00000000), 397 (0x001143e0 + 0x00000800), 398 (0x001143e0 + 0x00001000), 399 (0x001143e0 + 0x00001800), 400 (0x001543e0 + 0x00000000), 401 (0x001543e0 + 0x00000800), 402 (0x001543e0 + 0x00001000), 403 (0x001543e0 + 0x00001800), 404 (0x001943e0 + 0x00000000), 405 (0x001943e0 + 0x00000800), 406 (0x001943e0 + 0x00001000), 407 (0x001943e0 + 0x00001800), 408 (0x001d43e0 + 0x00000000), 409 (0x001d43e0 + 0x00000800), 410 (0x001d43e0 + 0x00001000), 411 (0x001d43e0 + 0x00001800), 412 }; 413 414 static int gmc_v9_0_ecc_interrupt_state(struct amdgpu_device *adev, 415 struct amdgpu_irq_src *src, 416 unsigned type, 417 enum amdgpu_interrupt_state state) 418 { 419 u32 bits, i, tmp, reg; 420 421 /* Devices newer then VEGA10/12 shall have these programming 422 sequences performed by PSP BL */ 423 if (adev->asic_type >= CHIP_VEGA20) 424 return 0; 425 426 bits = 0x7f; 427 428 switch (state) { 429 case AMDGPU_IRQ_STATE_DISABLE: 430 for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_addrs); i++) { 431 reg = ecc_umc_mcumc_ctrl_addrs[i]; 432 tmp = RREG32(reg); 433 tmp &= ~bits; 434 WREG32(reg, tmp); 435 } 436 for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_mask_addrs); i++) { 437 reg = ecc_umc_mcumc_ctrl_mask_addrs[i]; 438 tmp = RREG32(reg); 439 tmp &= ~bits; 440 WREG32(reg, tmp); 441 } 442 break; 443 case AMDGPU_IRQ_STATE_ENABLE: 444 for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_addrs); i++) { 445 reg = ecc_umc_mcumc_ctrl_addrs[i]; 446 tmp = RREG32(reg); 447 tmp |= bits; 448 WREG32(reg, tmp); 449 } 450 for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_mask_addrs); i++) { 451 reg = ecc_umc_mcumc_ctrl_mask_addrs[i]; 452 tmp = RREG32(reg); 453 tmp |= bits; 454 WREG32(reg, tmp); 455 } 456 break; 457 default: 458 break; 459 } 460 461 return 0; 462 } 463 464 static int gmc_v9_0_vm_fault_interrupt_state(struct amdgpu_device *adev, 465 struct amdgpu_irq_src *src, 466 unsigned type, 467 enum amdgpu_interrupt_state state) 468 { 469 struct amdgpu_vmhub *hub; 470 u32 tmp, reg, bits, i, j; 471 472 bits = VM_CONTEXT1_CNTL__RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 473 VM_CONTEXT1_CNTL__DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 474 VM_CONTEXT1_CNTL__PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 475 VM_CONTEXT1_CNTL__VALID_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 476 VM_CONTEXT1_CNTL__READ_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 477 VM_CONTEXT1_CNTL__WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 478 VM_CONTEXT1_CNTL__EXECUTE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK; 479 480 switch (state) { 481 case AMDGPU_IRQ_STATE_DISABLE: 482 for (j = 0; j < adev->num_vmhubs; j++) { 483 hub = &adev->vmhub[j]; 484 for (i = 0; i < 16; i++) { 485 reg = hub->vm_context0_cntl + i; 486 487 if (j == AMDGPU_GFXHUB_0) 488 tmp = RREG32_SOC15_IP(GC, reg); 489 else 490 tmp = RREG32_SOC15_IP(MMHUB, reg); 491 492 tmp &= ~bits; 493 494 if (j == AMDGPU_GFXHUB_0) 495 WREG32_SOC15_IP(GC, reg, tmp); 496 else 497 WREG32_SOC15_IP(MMHUB, reg, tmp); 498 } 499 } 500 break; 501 case AMDGPU_IRQ_STATE_ENABLE: 502 for (j = 0; j < adev->num_vmhubs; j++) { 503 hub = &adev->vmhub[j]; 504 for (i = 0; i < 16; i++) { 505 reg = hub->vm_context0_cntl + i; 506 507 if (j == AMDGPU_GFXHUB_0) 508 tmp = RREG32_SOC15_IP(GC, reg); 509 else 510 tmp = RREG32_SOC15_IP(MMHUB, reg); 511 512 tmp |= bits; 513 514 if (j == AMDGPU_GFXHUB_0) 515 WREG32_SOC15_IP(GC, reg, tmp); 516 else 517 WREG32_SOC15_IP(MMHUB, reg, tmp); 518 } 519 } 520 break; 521 default: 522 break; 523 } 524 525 return 0; 526 } 527 528 static int gmc_v9_0_process_interrupt(struct amdgpu_device *adev, 529 struct amdgpu_irq_src *source, 530 struct amdgpu_iv_entry *entry) 531 { 532 bool retry_fault = !!(entry->src_data[1] & 0x80); 533 bool write_fault = !!(entry->src_data[1] & 0x20); 534 uint32_t status = 0, cid = 0, rw = 0; 535 struct amdgpu_task_info task_info; 536 struct amdgpu_vmhub *hub; 537 const char *mmhub_cid; 538 const char *hub_name; 539 u64 addr; 540 541 addr = (u64)entry->src_data[0] << 12; 542 addr |= ((u64)entry->src_data[1] & 0xf) << 44; 543 544 if (retry_fault) { 545 /* Returning 1 here also prevents sending the IV to the KFD */ 546 547 /* Process it onyl if it's the first fault for this address */ 548 if (entry->ih != &adev->irq.ih_soft && 549 amdgpu_gmc_filter_faults(adev, entry->ih, addr, entry->pasid, 550 entry->timestamp)) 551 return 1; 552 553 /* Delegate it to a different ring if the hardware hasn't 554 * already done it. 555 */ 556 if (entry->ih == &adev->irq.ih) { 557 amdgpu_irq_delegate(adev, entry, 8); 558 return 1; 559 } 560 561 /* Try to handle the recoverable page faults by filling page 562 * tables 563 */ 564 if (amdgpu_vm_handle_fault(adev, entry->pasid, addr, write_fault)) 565 return 1; 566 } 567 568 if (!printk_ratelimit()) 569 return 0; 570 571 if (entry->client_id == SOC15_IH_CLIENTID_VMC) { 572 hub_name = "mmhub0"; 573 hub = &adev->vmhub[AMDGPU_MMHUB_0]; 574 } else if (entry->client_id == SOC15_IH_CLIENTID_VMC1) { 575 hub_name = "mmhub1"; 576 hub = &adev->vmhub[AMDGPU_MMHUB_1]; 577 } else { 578 hub_name = "gfxhub0"; 579 hub = &adev->vmhub[AMDGPU_GFXHUB_0]; 580 } 581 582 memset(&task_info, 0, sizeof(struct amdgpu_task_info)); 583 amdgpu_vm_get_task_info(adev, entry->pasid, &task_info); 584 585 dev_err(adev->dev, 586 "[%s] %s page fault (src_id:%u ring:%u vmid:%u " 587 "pasid:%u, for process %s pid %d thread %s pid %d)\n", 588 hub_name, retry_fault ? "retry" : "no-retry", 589 entry->src_id, entry->ring_id, entry->vmid, 590 entry->pasid, task_info.process_name, task_info.tgid, 591 task_info.task_name, task_info.pid); 592 dev_err(adev->dev, " in page starting at address 0x%016llx from IH client 0x%x (%s)\n", 593 addr, entry->client_id, 594 soc15_ih_clientid_name[entry->client_id]); 595 596 if (amdgpu_sriov_vf(adev)) 597 return 0; 598 599 /* 600 * Issue a dummy read to wait for the status register to 601 * be updated to avoid reading an incorrect value due to 602 * the new fast GRBM interface. 603 */ 604 if ((entry->vmid_src == AMDGPU_GFXHUB_0) && 605 (adev->ip_versions[GC_HWIP][0] < IP_VERSION(9, 4, 2))) 606 RREG32(hub->vm_l2_pro_fault_status); 607 608 status = RREG32(hub->vm_l2_pro_fault_status); 609 cid = REG_GET_FIELD(status, VM_L2_PROTECTION_FAULT_STATUS, CID); 610 rw = REG_GET_FIELD(status, VM_L2_PROTECTION_FAULT_STATUS, RW); 611 WREG32_P(hub->vm_l2_pro_fault_cntl, 1, ~1); 612 613 614 dev_err(adev->dev, 615 "VM_L2_PROTECTION_FAULT_STATUS:0x%08X\n", 616 status); 617 if (hub == &adev->vmhub[AMDGPU_GFXHUB_0]) { 618 dev_err(adev->dev, "\t Faulty UTCL2 client ID: %s (0x%x)\n", 619 cid >= ARRAY_SIZE(gfxhub_client_ids) ? "unknown" : 620 gfxhub_client_ids[cid], 621 cid); 622 } else { 623 switch (adev->ip_versions[MMHUB_HWIP][0]) { 624 case IP_VERSION(9, 0, 0): 625 mmhub_cid = mmhub_client_ids_vega10[cid][rw]; 626 break; 627 case IP_VERSION(9, 3, 0): 628 mmhub_cid = mmhub_client_ids_vega12[cid][rw]; 629 break; 630 case IP_VERSION(9, 4, 0): 631 mmhub_cid = mmhub_client_ids_vega20[cid][rw]; 632 break; 633 case IP_VERSION(9, 4, 1): 634 mmhub_cid = mmhub_client_ids_arcturus[cid][rw]; 635 break; 636 case IP_VERSION(9, 1, 0): 637 case IP_VERSION(9, 2, 0): 638 mmhub_cid = mmhub_client_ids_raven[cid][rw]; 639 break; 640 case IP_VERSION(1, 5, 0): 641 case IP_VERSION(2, 4, 0): 642 mmhub_cid = mmhub_client_ids_renoir[cid][rw]; 643 break; 644 case IP_VERSION(9, 4, 2): 645 mmhub_cid = mmhub_client_ids_aldebaran[cid][rw]; 646 break; 647 default: 648 mmhub_cid = NULL; 649 break; 650 } 651 dev_err(adev->dev, "\t Faulty UTCL2 client ID: %s (0x%x)\n", 652 mmhub_cid ? mmhub_cid : "unknown", cid); 653 } 654 dev_err(adev->dev, "\t MORE_FAULTS: 0x%lx\n", 655 REG_GET_FIELD(status, 656 VM_L2_PROTECTION_FAULT_STATUS, MORE_FAULTS)); 657 dev_err(adev->dev, "\t WALKER_ERROR: 0x%lx\n", 658 REG_GET_FIELD(status, 659 VM_L2_PROTECTION_FAULT_STATUS, WALKER_ERROR)); 660 dev_err(adev->dev, "\t PERMISSION_FAULTS: 0x%lx\n", 661 REG_GET_FIELD(status, 662 VM_L2_PROTECTION_FAULT_STATUS, PERMISSION_FAULTS)); 663 dev_err(adev->dev, "\t MAPPING_ERROR: 0x%lx\n", 664 REG_GET_FIELD(status, 665 VM_L2_PROTECTION_FAULT_STATUS, MAPPING_ERROR)); 666 dev_err(adev->dev, "\t RW: 0x%x\n", rw); 667 return 0; 668 } 669 670 static const struct amdgpu_irq_src_funcs gmc_v9_0_irq_funcs = { 671 .set = gmc_v9_0_vm_fault_interrupt_state, 672 .process = gmc_v9_0_process_interrupt, 673 }; 674 675 676 static const struct amdgpu_irq_src_funcs gmc_v9_0_ecc_funcs = { 677 .set = gmc_v9_0_ecc_interrupt_state, 678 .process = amdgpu_umc_process_ecc_irq, 679 }; 680 681 static void gmc_v9_0_set_irq_funcs(struct amdgpu_device *adev) 682 { 683 adev->gmc.vm_fault.num_types = 1; 684 adev->gmc.vm_fault.funcs = &gmc_v9_0_irq_funcs; 685 686 if (!amdgpu_sriov_vf(adev) && 687 !adev->gmc.xgmi.connected_to_cpu) { 688 adev->gmc.ecc_irq.num_types = 1; 689 adev->gmc.ecc_irq.funcs = &gmc_v9_0_ecc_funcs; 690 } 691 } 692 693 static uint32_t gmc_v9_0_get_invalidate_req(unsigned int vmid, 694 uint32_t flush_type) 695 { 696 u32 req = 0; 697 698 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, 699 PER_VMID_INVALIDATE_REQ, 1 << vmid); 700 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, FLUSH_TYPE, flush_type); 701 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PTES, 1); 702 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE0, 1); 703 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE1, 1); 704 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE2, 1); 705 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L1_PTES, 1); 706 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, 707 CLEAR_PROTECTION_FAULT_STATUS_ADDR, 0); 708 709 return req; 710 } 711 712 /** 713 * gmc_v9_0_use_invalidate_semaphore - judge whether to use semaphore 714 * 715 * @adev: amdgpu_device pointer 716 * @vmhub: vmhub type 717 * 718 */ 719 static bool gmc_v9_0_use_invalidate_semaphore(struct amdgpu_device *adev, 720 uint32_t vmhub) 721 { 722 if (adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 2)) 723 return false; 724 725 return ((vmhub == AMDGPU_MMHUB_0 || 726 vmhub == AMDGPU_MMHUB_1) && 727 (!amdgpu_sriov_vf(adev)) && 728 (!(!(adev->apu_flags & AMD_APU_IS_RAVEN2) && 729 (adev->apu_flags & AMD_APU_IS_PICASSO)))); 730 } 731 732 static bool gmc_v9_0_get_atc_vmid_pasid_mapping_info(struct amdgpu_device *adev, 733 uint8_t vmid, uint16_t *p_pasid) 734 { 735 uint32_t value; 736 737 value = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING) 738 + vmid); 739 *p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK; 740 741 return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK); 742 } 743 744 /* 745 * GART 746 * VMID 0 is the physical GPU addresses as used by the kernel. 747 * VMIDs 1-15 are used for userspace clients and are handled 748 * by the amdgpu vm/hsa code. 749 */ 750 751 /** 752 * gmc_v9_0_flush_gpu_tlb - tlb flush with certain type 753 * 754 * @adev: amdgpu_device pointer 755 * @vmid: vm instance to flush 756 * @vmhub: which hub to flush 757 * @flush_type: the flush type 758 * 759 * Flush the TLB for the requested page table using certain type. 760 */ 761 static void gmc_v9_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid, 762 uint32_t vmhub, uint32_t flush_type) 763 { 764 bool use_semaphore = gmc_v9_0_use_invalidate_semaphore(adev, vmhub); 765 const unsigned eng = 17; 766 u32 j, inv_req, inv_req2, tmp; 767 struct amdgpu_vmhub *hub; 768 769 BUG_ON(vmhub >= adev->num_vmhubs); 770 771 hub = &adev->vmhub[vmhub]; 772 if (adev->gmc.xgmi.num_physical_nodes && 773 adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 0)) { 774 /* Vega20+XGMI caches PTEs in TC and TLB. Add a 775 * heavy-weight TLB flush (type 2), which flushes 776 * both. Due to a race condition with concurrent 777 * memory accesses using the same TLB cache line, we 778 * still need a second TLB flush after this. 779 */ 780 inv_req = gmc_v9_0_get_invalidate_req(vmid, 2); 781 inv_req2 = gmc_v9_0_get_invalidate_req(vmid, flush_type); 782 } else { 783 inv_req = gmc_v9_0_get_invalidate_req(vmid, flush_type); 784 inv_req2 = 0; 785 } 786 787 /* This is necessary for a HW workaround under SRIOV as well 788 * as GFXOFF under bare metal 789 */ 790 if (adev->gfx.kiq.ring.sched.ready && 791 (amdgpu_sriov_runtime(adev) || !amdgpu_sriov_vf(adev)) && 792 down_read_trylock(&adev->reset_domain->sem)) { 793 uint32_t req = hub->vm_inv_eng0_req + hub->eng_distance * eng; 794 uint32_t ack = hub->vm_inv_eng0_ack + hub->eng_distance * eng; 795 796 amdgpu_virt_kiq_reg_write_reg_wait(adev, req, ack, inv_req, 797 1 << vmid); 798 up_read(&adev->reset_domain->sem); 799 return; 800 } 801 802 spin_lock(&adev->gmc.invalidate_lock); 803 804 /* 805 * It may lose gpuvm invalidate acknowldege state across power-gating 806 * off cycle, add semaphore acquire before invalidation and semaphore 807 * release after invalidation to avoid entering power gated state 808 * to WA the Issue 809 */ 810 811 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 812 if (use_semaphore) { 813 for (j = 0; j < adev->usec_timeout; j++) { 814 /* a read return value of 1 means semaphore acquire */ 815 if (vmhub == AMDGPU_GFXHUB_0) 816 tmp = RREG32_SOC15_IP_NO_KIQ(GC, hub->vm_inv_eng0_sem + hub->eng_distance * eng); 817 else 818 tmp = RREG32_SOC15_IP_NO_KIQ(MMHUB, hub->vm_inv_eng0_sem + hub->eng_distance * eng); 819 820 if (tmp & 0x1) 821 break; 822 udelay(1); 823 } 824 825 if (j >= adev->usec_timeout) 826 DRM_ERROR("Timeout waiting for sem acquire in VM flush!\n"); 827 } 828 829 do { 830 if (vmhub == AMDGPU_GFXHUB_0) 831 WREG32_SOC15_IP_NO_KIQ(GC, hub->vm_inv_eng0_req + hub->eng_distance * eng, inv_req); 832 else 833 WREG32_SOC15_IP_NO_KIQ(MMHUB, hub->vm_inv_eng0_req + hub->eng_distance * eng, inv_req); 834 835 /* 836 * Issue a dummy read to wait for the ACK register to 837 * be cleared to avoid a false ACK due to the new fast 838 * GRBM interface. 839 */ 840 if ((vmhub == AMDGPU_GFXHUB_0) && 841 (adev->ip_versions[GC_HWIP][0] < IP_VERSION(9, 4, 2))) 842 RREG32_NO_KIQ(hub->vm_inv_eng0_req + 843 hub->eng_distance * eng); 844 845 for (j = 0; j < adev->usec_timeout; j++) { 846 if (vmhub == AMDGPU_GFXHUB_0) 847 tmp = RREG32_SOC15_IP_NO_KIQ(GC, hub->vm_inv_eng0_ack + hub->eng_distance * eng); 848 else 849 tmp = RREG32_SOC15_IP_NO_KIQ(MMHUB, hub->vm_inv_eng0_ack + hub->eng_distance * eng); 850 851 if (tmp & (1 << vmid)) 852 break; 853 udelay(1); 854 } 855 856 inv_req = inv_req2; 857 inv_req2 = 0; 858 } while (inv_req); 859 860 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 861 if (use_semaphore) { 862 /* 863 * add semaphore release after invalidation, 864 * write with 0 means semaphore release 865 */ 866 if (vmhub == AMDGPU_GFXHUB_0) 867 WREG32_SOC15_IP_NO_KIQ(GC, hub->vm_inv_eng0_sem + hub->eng_distance * eng, 0); 868 else 869 WREG32_SOC15_IP_NO_KIQ(MMHUB, hub->vm_inv_eng0_sem + hub->eng_distance * eng, 0); 870 } 871 872 spin_unlock(&adev->gmc.invalidate_lock); 873 874 if (j < adev->usec_timeout) 875 return; 876 877 DRM_ERROR("Timeout waiting for VM flush ACK!\n"); 878 } 879 880 /** 881 * gmc_v9_0_flush_gpu_tlb_pasid - tlb flush via pasid 882 * 883 * @adev: amdgpu_device pointer 884 * @pasid: pasid to be flush 885 * @flush_type: the flush type 886 * @all_hub: flush all hubs 887 * 888 * Flush the TLB for the requested pasid. 889 */ 890 static int gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev, 891 uint16_t pasid, uint32_t flush_type, 892 bool all_hub) 893 { 894 int vmid, i; 895 signed long r; 896 uint32_t seq; 897 uint16_t queried_pasid; 898 bool ret; 899 u32 usec_timeout = amdgpu_sriov_vf(adev) ? SRIOV_USEC_TIMEOUT : adev->usec_timeout; 900 struct amdgpu_ring *ring = &adev->gfx.kiq.ring; 901 struct amdgpu_kiq *kiq = &adev->gfx.kiq; 902 903 if (amdgpu_in_reset(adev)) 904 return -EIO; 905 906 if (ring->sched.ready && down_read_trylock(&adev->reset_domain->sem)) { 907 /* Vega20+XGMI caches PTEs in TC and TLB. Add a 908 * heavy-weight TLB flush (type 2), which flushes 909 * both. Due to a race condition with concurrent 910 * memory accesses using the same TLB cache line, we 911 * still need a second TLB flush after this. 912 */ 913 bool vega20_xgmi_wa = (adev->gmc.xgmi.num_physical_nodes && 914 adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 0)); 915 /* 2 dwords flush + 8 dwords fence */ 916 unsigned int ndw = kiq->pmf->invalidate_tlbs_size + 8; 917 918 if (vega20_xgmi_wa) 919 ndw += kiq->pmf->invalidate_tlbs_size; 920 921 spin_lock(&adev->gfx.kiq.ring_lock); 922 /* 2 dwords flush + 8 dwords fence */ 923 amdgpu_ring_alloc(ring, ndw); 924 if (vega20_xgmi_wa) 925 kiq->pmf->kiq_invalidate_tlbs(ring, 926 pasid, 2, all_hub); 927 kiq->pmf->kiq_invalidate_tlbs(ring, 928 pasid, flush_type, all_hub); 929 r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT); 930 if (r) { 931 amdgpu_ring_undo(ring); 932 spin_unlock(&adev->gfx.kiq.ring_lock); 933 up_read(&adev->reset_domain->sem); 934 return -ETIME; 935 } 936 937 amdgpu_ring_commit(ring); 938 spin_unlock(&adev->gfx.kiq.ring_lock); 939 r = amdgpu_fence_wait_polling(ring, seq, usec_timeout); 940 if (r < 1) { 941 dev_err(adev->dev, "wait for kiq fence error: %ld.\n", r); 942 up_read(&adev->reset_domain->sem); 943 return -ETIME; 944 } 945 up_read(&adev->reset_domain->sem); 946 return 0; 947 } 948 949 for (vmid = 1; vmid < 16; vmid++) { 950 951 ret = gmc_v9_0_get_atc_vmid_pasid_mapping_info(adev, vmid, 952 &queried_pasid); 953 if (ret && queried_pasid == pasid) { 954 if (all_hub) { 955 for (i = 0; i < adev->num_vmhubs; i++) 956 gmc_v9_0_flush_gpu_tlb(adev, vmid, 957 i, flush_type); 958 } else { 959 gmc_v9_0_flush_gpu_tlb(adev, vmid, 960 AMDGPU_GFXHUB_0, flush_type); 961 } 962 break; 963 } 964 } 965 966 return 0; 967 968 } 969 970 static uint64_t gmc_v9_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring, 971 unsigned vmid, uint64_t pd_addr) 972 { 973 bool use_semaphore = gmc_v9_0_use_invalidate_semaphore(ring->adev, ring->funcs->vmhub); 974 struct amdgpu_device *adev = ring->adev; 975 struct amdgpu_vmhub *hub = &adev->vmhub[ring->funcs->vmhub]; 976 uint32_t req = gmc_v9_0_get_invalidate_req(vmid, 0); 977 unsigned eng = ring->vm_inv_eng; 978 979 /* 980 * It may lose gpuvm invalidate acknowldege state across power-gating 981 * off cycle, add semaphore acquire before invalidation and semaphore 982 * release after invalidation to avoid entering power gated state 983 * to WA the Issue 984 */ 985 986 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 987 if (use_semaphore) 988 /* a read return value of 1 means semaphore acuqire */ 989 amdgpu_ring_emit_reg_wait(ring, 990 hub->vm_inv_eng0_sem + 991 hub->eng_distance * eng, 0x1, 0x1); 992 993 amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_lo32 + 994 (hub->ctx_addr_distance * vmid), 995 lower_32_bits(pd_addr)); 996 997 amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 + 998 (hub->ctx_addr_distance * vmid), 999 upper_32_bits(pd_addr)); 1000 1001 amdgpu_ring_emit_reg_write_reg_wait(ring, hub->vm_inv_eng0_req + 1002 hub->eng_distance * eng, 1003 hub->vm_inv_eng0_ack + 1004 hub->eng_distance * eng, 1005 req, 1 << vmid); 1006 1007 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 1008 if (use_semaphore) 1009 /* 1010 * add semaphore release after invalidation, 1011 * write with 0 means semaphore release 1012 */ 1013 amdgpu_ring_emit_wreg(ring, hub->vm_inv_eng0_sem + 1014 hub->eng_distance * eng, 0); 1015 1016 return pd_addr; 1017 } 1018 1019 static void gmc_v9_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned vmid, 1020 unsigned pasid) 1021 { 1022 struct amdgpu_device *adev = ring->adev; 1023 uint32_t reg; 1024 1025 /* Do nothing because there's no lut register for mmhub1. */ 1026 if (ring->funcs->vmhub == AMDGPU_MMHUB_1) 1027 return; 1028 1029 if (ring->funcs->vmhub == AMDGPU_GFXHUB_0) 1030 reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid; 1031 else 1032 reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT_MM) + vmid; 1033 1034 amdgpu_ring_emit_wreg(ring, reg, pasid); 1035 } 1036 1037 /* 1038 * PTE format on VEGA 10: 1039 * 63:59 reserved 1040 * 58:57 mtype 1041 * 56 F 1042 * 55 L 1043 * 54 P 1044 * 53 SW 1045 * 52 T 1046 * 50:48 reserved 1047 * 47:12 4k physical page base address 1048 * 11:7 fragment 1049 * 6 write 1050 * 5 read 1051 * 4 exe 1052 * 3 Z 1053 * 2 snooped 1054 * 1 system 1055 * 0 valid 1056 * 1057 * PDE format on VEGA 10: 1058 * 63:59 block fragment size 1059 * 58:55 reserved 1060 * 54 P 1061 * 53:48 reserved 1062 * 47:6 physical base address of PD or PTE 1063 * 5:3 reserved 1064 * 2 C 1065 * 1 system 1066 * 0 valid 1067 */ 1068 1069 static uint64_t gmc_v9_0_map_mtype(struct amdgpu_device *adev, uint32_t flags) 1070 1071 { 1072 switch (flags) { 1073 case AMDGPU_VM_MTYPE_DEFAULT: 1074 return AMDGPU_PTE_MTYPE_VG10(MTYPE_NC); 1075 case AMDGPU_VM_MTYPE_NC: 1076 return AMDGPU_PTE_MTYPE_VG10(MTYPE_NC); 1077 case AMDGPU_VM_MTYPE_WC: 1078 return AMDGPU_PTE_MTYPE_VG10(MTYPE_WC); 1079 case AMDGPU_VM_MTYPE_RW: 1080 return AMDGPU_PTE_MTYPE_VG10(MTYPE_RW); 1081 case AMDGPU_VM_MTYPE_CC: 1082 return AMDGPU_PTE_MTYPE_VG10(MTYPE_CC); 1083 case AMDGPU_VM_MTYPE_UC: 1084 return AMDGPU_PTE_MTYPE_VG10(MTYPE_UC); 1085 default: 1086 return AMDGPU_PTE_MTYPE_VG10(MTYPE_NC); 1087 } 1088 } 1089 1090 static void gmc_v9_0_get_vm_pde(struct amdgpu_device *adev, int level, 1091 uint64_t *addr, uint64_t *flags) 1092 { 1093 if (!(*flags & AMDGPU_PDE_PTE) && !(*flags & AMDGPU_PTE_SYSTEM)) 1094 *addr = amdgpu_gmc_vram_mc2pa(adev, *addr); 1095 BUG_ON(*addr & 0xFFFF00000000003FULL); 1096 1097 if (!adev->gmc.translate_further) 1098 return; 1099 1100 if (level == AMDGPU_VM_PDB1) { 1101 /* Set the block fragment size */ 1102 if (!(*flags & AMDGPU_PDE_PTE)) 1103 *flags |= AMDGPU_PDE_BFS(0x9); 1104 1105 } else if (level == AMDGPU_VM_PDB0) { 1106 if (*flags & AMDGPU_PDE_PTE) { 1107 *flags &= ~AMDGPU_PDE_PTE; 1108 if (!(*flags & AMDGPU_PTE_VALID)) 1109 *addr |= 1 << PAGE_SHIFT; 1110 } else { 1111 *flags |= AMDGPU_PTE_TF; 1112 } 1113 } 1114 } 1115 1116 static void gmc_v9_0_get_coherence_flags(struct amdgpu_device *adev, 1117 struct amdgpu_bo *bo, 1118 struct amdgpu_bo_va_mapping *mapping, 1119 uint64_t *flags) 1120 { 1121 struct amdgpu_device *bo_adev = amdgpu_ttm_adev(bo->tbo.bdev); 1122 bool is_vram = bo->tbo.resource->mem_type == TTM_PL_VRAM; 1123 bool coherent = bo->flags & AMDGPU_GEM_CREATE_COHERENT; 1124 bool uncached = bo->flags & AMDGPU_GEM_CREATE_UNCACHED; 1125 unsigned int mtype; 1126 bool snoop = false; 1127 1128 switch (adev->ip_versions[GC_HWIP][0]) { 1129 case IP_VERSION(9, 4, 1): 1130 case IP_VERSION(9, 4, 2): 1131 if (is_vram) { 1132 if (bo_adev == adev) { 1133 if (uncached) 1134 mtype = MTYPE_UC; 1135 else if (coherent) 1136 mtype = MTYPE_CC; 1137 else 1138 mtype = MTYPE_RW; 1139 /* FIXME: is this still needed? Or does 1140 * amdgpu_ttm_tt_pde_flags already handle this? 1141 */ 1142 if (adev->ip_versions[GC_HWIP][0] == 1143 IP_VERSION(9, 4, 2) && 1144 adev->gmc.xgmi.connected_to_cpu) 1145 snoop = true; 1146 } else { 1147 if (uncached || coherent) 1148 mtype = MTYPE_UC; 1149 else 1150 mtype = MTYPE_NC; 1151 if (mapping->bo_va->is_xgmi) 1152 snoop = true; 1153 } 1154 } else { 1155 if (uncached || coherent) 1156 mtype = MTYPE_UC; 1157 else 1158 mtype = MTYPE_NC; 1159 /* FIXME: is this still needed? Or does 1160 * amdgpu_ttm_tt_pde_flags already handle this? 1161 */ 1162 snoop = true; 1163 } 1164 break; 1165 default: 1166 if (uncached || coherent) 1167 mtype = MTYPE_UC; 1168 else 1169 mtype = MTYPE_NC; 1170 1171 /* FIXME: is this still needed? Or does 1172 * amdgpu_ttm_tt_pde_flags already handle this? 1173 */ 1174 if (!is_vram) 1175 snoop = true; 1176 } 1177 1178 if (mtype != MTYPE_NC) 1179 *flags = (*flags & ~AMDGPU_PTE_MTYPE_VG10_MASK) | 1180 AMDGPU_PTE_MTYPE_VG10(mtype); 1181 *flags |= snoop ? AMDGPU_PTE_SNOOPED : 0; 1182 } 1183 1184 static void gmc_v9_0_get_vm_pte(struct amdgpu_device *adev, 1185 struct amdgpu_bo_va_mapping *mapping, 1186 uint64_t *flags) 1187 { 1188 struct amdgpu_bo *bo = mapping->bo_va->base.bo; 1189 1190 *flags &= ~AMDGPU_PTE_EXECUTABLE; 1191 *flags |= mapping->flags & AMDGPU_PTE_EXECUTABLE; 1192 1193 *flags &= ~AMDGPU_PTE_MTYPE_VG10_MASK; 1194 *flags |= mapping->flags & AMDGPU_PTE_MTYPE_VG10_MASK; 1195 1196 if (mapping->flags & AMDGPU_PTE_PRT) { 1197 *flags |= AMDGPU_PTE_PRT; 1198 *flags &= ~AMDGPU_PTE_VALID; 1199 } 1200 1201 if (bo && bo->tbo.resource) 1202 gmc_v9_0_get_coherence_flags(adev, mapping->bo_va->base.bo, 1203 mapping, flags); 1204 } 1205 1206 static unsigned gmc_v9_0_get_vbios_fb_size(struct amdgpu_device *adev) 1207 { 1208 u32 d1vga_control = RREG32_SOC15(DCE, 0, mmD1VGA_CONTROL); 1209 unsigned size; 1210 1211 /* TODO move to DC so GMC doesn't need to hard-code DCN registers */ 1212 1213 if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) { 1214 size = AMDGPU_VBIOS_VGA_ALLOCATION; 1215 } else { 1216 u32 viewport; 1217 1218 switch (adev->ip_versions[DCE_HWIP][0]) { 1219 case IP_VERSION(1, 0, 0): 1220 case IP_VERSION(1, 0, 1): 1221 viewport = RREG32_SOC15(DCE, 0, mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION); 1222 size = (REG_GET_FIELD(viewport, 1223 HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_HEIGHT) * 1224 REG_GET_FIELD(viewport, 1225 HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_WIDTH) * 1226 4); 1227 break; 1228 case IP_VERSION(2, 1, 0): 1229 viewport = RREG32_SOC15(DCE, 0, mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION_DCN2); 1230 size = (REG_GET_FIELD(viewport, 1231 HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_HEIGHT) * 1232 REG_GET_FIELD(viewport, 1233 HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_WIDTH) * 1234 4); 1235 break; 1236 default: 1237 viewport = RREG32_SOC15(DCE, 0, mmSCL0_VIEWPORT_SIZE); 1238 size = (REG_GET_FIELD(viewport, SCL0_VIEWPORT_SIZE, VIEWPORT_HEIGHT) * 1239 REG_GET_FIELD(viewport, SCL0_VIEWPORT_SIZE, VIEWPORT_WIDTH) * 1240 4); 1241 break; 1242 } 1243 } 1244 1245 return size; 1246 } 1247 1248 static const struct amdgpu_gmc_funcs gmc_v9_0_gmc_funcs = { 1249 .flush_gpu_tlb = gmc_v9_0_flush_gpu_tlb, 1250 .flush_gpu_tlb_pasid = gmc_v9_0_flush_gpu_tlb_pasid, 1251 .emit_flush_gpu_tlb = gmc_v9_0_emit_flush_gpu_tlb, 1252 .emit_pasid_mapping = gmc_v9_0_emit_pasid_mapping, 1253 .map_mtype = gmc_v9_0_map_mtype, 1254 .get_vm_pde = gmc_v9_0_get_vm_pde, 1255 .get_vm_pte = gmc_v9_0_get_vm_pte, 1256 .get_vbios_fb_size = gmc_v9_0_get_vbios_fb_size, 1257 }; 1258 1259 static void gmc_v9_0_set_gmc_funcs(struct amdgpu_device *adev) 1260 { 1261 adev->gmc.gmc_funcs = &gmc_v9_0_gmc_funcs; 1262 } 1263 1264 static void gmc_v9_0_set_umc_funcs(struct amdgpu_device *adev) 1265 { 1266 switch (adev->ip_versions[UMC_HWIP][0]) { 1267 case IP_VERSION(6, 0, 0): 1268 adev->umc.funcs = &umc_v6_0_funcs; 1269 break; 1270 case IP_VERSION(6, 1, 1): 1271 adev->umc.max_ras_err_cnt_per_query = UMC_V6_1_TOTAL_CHANNEL_NUM; 1272 adev->umc.channel_inst_num = UMC_V6_1_CHANNEL_INSTANCE_NUM; 1273 adev->umc.umc_inst_num = UMC_V6_1_UMC_INSTANCE_NUM; 1274 adev->umc.channel_offs = UMC_V6_1_PER_CHANNEL_OFFSET_VG20; 1275 adev->umc.channel_idx_tbl = &umc_v6_1_channel_idx_tbl[0][0]; 1276 adev->umc.ras = &umc_v6_1_ras; 1277 break; 1278 case IP_VERSION(6, 1, 2): 1279 adev->umc.max_ras_err_cnt_per_query = UMC_V6_1_TOTAL_CHANNEL_NUM; 1280 adev->umc.channel_inst_num = UMC_V6_1_CHANNEL_INSTANCE_NUM; 1281 adev->umc.umc_inst_num = UMC_V6_1_UMC_INSTANCE_NUM; 1282 adev->umc.channel_offs = UMC_V6_1_PER_CHANNEL_OFFSET_ARCT; 1283 adev->umc.channel_idx_tbl = &umc_v6_1_channel_idx_tbl[0][0]; 1284 adev->umc.ras = &umc_v6_1_ras; 1285 break; 1286 case IP_VERSION(6, 7, 0): 1287 adev->umc.max_ras_err_cnt_per_query = 1288 UMC_V6_7_TOTAL_CHANNEL_NUM * UMC_V6_7_BAD_PAGE_NUM_PER_CHANNEL; 1289 adev->umc.channel_inst_num = UMC_V6_7_CHANNEL_INSTANCE_NUM; 1290 adev->umc.umc_inst_num = UMC_V6_7_UMC_INSTANCE_NUM; 1291 adev->umc.channel_offs = UMC_V6_7_PER_CHANNEL_OFFSET; 1292 if (!adev->gmc.xgmi.connected_to_cpu) 1293 adev->umc.ras = &umc_v6_7_ras; 1294 if (1 & adev->smuio.funcs->get_die_id(adev)) 1295 adev->umc.channel_idx_tbl = &umc_v6_7_channel_idx_tbl_first[0][0]; 1296 else 1297 adev->umc.channel_idx_tbl = &umc_v6_7_channel_idx_tbl_second[0][0]; 1298 break; 1299 default: 1300 break; 1301 } 1302 1303 if (adev->umc.ras) { 1304 amdgpu_ras_register_ras_block(adev, &adev->umc.ras->ras_block); 1305 1306 strcpy(adev->umc.ras->ras_block.ras_comm.name, "umc"); 1307 adev->umc.ras->ras_block.ras_comm.block = AMDGPU_RAS_BLOCK__UMC; 1308 adev->umc.ras->ras_block.ras_comm.type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE; 1309 adev->umc.ras_if = &adev->umc.ras->ras_block.ras_comm; 1310 1311 /* If don't define special ras_late_init function, use default ras_late_init */ 1312 if (!adev->umc.ras->ras_block.ras_late_init) 1313 adev->umc.ras->ras_block.ras_late_init = amdgpu_umc_ras_late_init; 1314 1315 /* If not defined special ras_cb function, use default ras_cb */ 1316 if (!adev->umc.ras->ras_block.ras_cb) 1317 adev->umc.ras->ras_block.ras_cb = amdgpu_umc_process_ras_data_cb; 1318 } 1319 } 1320 1321 static void gmc_v9_0_set_mmhub_funcs(struct amdgpu_device *adev) 1322 { 1323 switch (adev->ip_versions[MMHUB_HWIP][0]) { 1324 case IP_VERSION(9, 4, 1): 1325 adev->mmhub.funcs = &mmhub_v9_4_funcs; 1326 break; 1327 case IP_VERSION(9, 4, 2): 1328 adev->mmhub.funcs = &mmhub_v1_7_funcs; 1329 break; 1330 default: 1331 adev->mmhub.funcs = &mmhub_v1_0_funcs; 1332 break; 1333 } 1334 } 1335 1336 static void gmc_v9_0_set_mmhub_ras_funcs(struct amdgpu_device *adev) 1337 { 1338 switch (adev->ip_versions[MMHUB_HWIP][0]) { 1339 case IP_VERSION(9, 4, 0): 1340 adev->mmhub.ras = &mmhub_v1_0_ras; 1341 break; 1342 case IP_VERSION(9, 4, 1): 1343 adev->mmhub.ras = &mmhub_v9_4_ras; 1344 break; 1345 case IP_VERSION(9, 4, 2): 1346 adev->mmhub.ras = &mmhub_v1_7_ras; 1347 break; 1348 default: 1349 /* mmhub ras is not available */ 1350 break; 1351 } 1352 1353 if (adev->mmhub.ras) { 1354 amdgpu_ras_register_ras_block(adev, &adev->mmhub.ras->ras_block); 1355 1356 strcpy(adev->mmhub.ras->ras_block.ras_comm.name, "mmhub"); 1357 adev->mmhub.ras->ras_block.ras_comm.block = AMDGPU_RAS_BLOCK__MMHUB; 1358 adev->mmhub.ras->ras_block.ras_comm.type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE; 1359 adev->mmhub.ras_if = &adev->mmhub.ras->ras_block.ras_comm; 1360 } 1361 } 1362 1363 static void gmc_v9_0_set_gfxhub_funcs(struct amdgpu_device *adev) 1364 { 1365 adev->gfxhub.funcs = &gfxhub_v1_0_funcs; 1366 } 1367 1368 static void gmc_v9_0_set_hdp_ras_funcs(struct amdgpu_device *adev) 1369 { 1370 adev->hdp.ras = &hdp_v4_0_ras; 1371 amdgpu_ras_register_ras_block(adev, &adev->hdp.ras->ras_block); 1372 adev->hdp.ras_if = &adev->hdp.ras->ras_block.ras_comm; 1373 } 1374 1375 static void gmc_v9_0_set_mca_funcs(struct amdgpu_device *adev) 1376 { 1377 /* is UMC the right IP to check for MCA? Maybe DF? */ 1378 switch (adev->ip_versions[UMC_HWIP][0]) { 1379 case IP_VERSION(6, 7, 0): 1380 if (!adev->gmc.xgmi.connected_to_cpu) 1381 adev->mca.funcs = &mca_v3_0_funcs; 1382 break; 1383 default: 1384 break; 1385 } 1386 } 1387 1388 static int gmc_v9_0_early_init(void *handle) 1389 { 1390 int r; 1391 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1392 1393 /* ARCT and VEGA20 don't have XGMI defined in their IP discovery tables */ 1394 if (adev->asic_type == CHIP_VEGA20 || 1395 adev->asic_type == CHIP_ARCTURUS) 1396 adev->gmc.xgmi.supported = true; 1397 1398 if (adev->ip_versions[XGMI_HWIP][0] == IP_VERSION(6, 1, 0)) { 1399 adev->gmc.xgmi.supported = true; 1400 adev->gmc.xgmi.connected_to_cpu = 1401 adev->smuio.funcs->is_host_gpu_xgmi_supported(adev); 1402 } 1403 1404 gmc_v9_0_set_gmc_funcs(adev); 1405 gmc_v9_0_set_irq_funcs(adev); 1406 gmc_v9_0_set_umc_funcs(adev); 1407 gmc_v9_0_set_mmhub_funcs(adev); 1408 gmc_v9_0_set_mmhub_ras_funcs(adev); 1409 gmc_v9_0_set_gfxhub_funcs(adev); 1410 gmc_v9_0_set_hdp_ras_funcs(adev); 1411 gmc_v9_0_set_mca_funcs(adev); 1412 1413 adev->gmc.shared_aperture_start = 0x2000000000000000ULL; 1414 adev->gmc.shared_aperture_end = 1415 adev->gmc.shared_aperture_start + (4ULL << 30) - 1; 1416 adev->gmc.private_aperture_start = 0x1000000000000000ULL; 1417 adev->gmc.private_aperture_end = 1418 adev->gmc.private_aperture_start + (4ULL << 30) - 1; 1419 1420 r = amdgpu_gmc_ras_early_init(adev); 1421 if (r) 1422 return r; 1423 1424 return 0; 1425 } 1426 1427 static int gmc_v9_0_late_init(void *handle) 1428 { 1429 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1430 int r; 1431 1432 r = amdgpu_gmc_allocate_vm_inv_eng(adev); 1433 if (r) 1434 return r; 1435 1436 /* 1437 * Workaround performance drop issue with VBIOS enables partial 1438 * writes, while disables HBM ECC for vega10. 1439 */ 1440 if (!amdgpu_sriov_vf(adev) && 1441 (adev->ip_versions[UMC_HWIP][0] == IP_VERSION(6, 0, 0))) { 1442 if (!(adev->ras_enabled & (1 << AMDGPU_RAS_BLOCK__UMC))) { 1443 if (adev->df.funcs && 1444 adev->df.funcs->enable_ecc_force_par_wr_rmw) 1445 adev->df.funcs->enable_ecc_force_par_wr_rmw(adev, false); 1446 } 1447 } 1448 1449 if (!amdgpu_persistent_edc_harvesting_supported(adev)) { 1450 if (adev->mmhub.ras && adev->mmhub.ras->ras_block.hw_ops && 1451 adev->mmhub.ras->ras_block.hw_ops->reset_ras_error_count) 1452 adev->mmhub.ras->ras_block.hw_ops->reset_ras_error_count(adev); 1453 1454 if (adev->hdp.ras && adev->hdp.ras->ras_block.hw_ops && 1455 adev->hdp.ras->ras_block.hw_ops->reset_ras_error_count) 1456 adev->hdp.ras->ras_block.hw_ops->reset_ras_error_count(adev); 1457 } 1458 1459 r = amdgpu_gmc_ras_late_init(adev); 1460 if (r) 1461 return r; 1462 1463 return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0); 1464 } 1465 1466 static void gmc_v9_0_vram_gtt_location(struct amdgpu_device *adev, 1467 struct amdgpu_gmc *mc) 1468 { 1469 u64 base = adev->mmhub.funcs->get_fb_location(adev); 1470 1471 /* add the xgmi offset of the physical node */ 1472 base += adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size; 1473 if (adev->gmc.xgmi.connected_to_cpu) { 1474 amdgpu_gmc_sysvm_location(adev, mc); 1475 } else { 1476 amdgpu_gmc_vram_location(adev, mc, base); 1477 amdgpu_gmc_gart_location(adev, mc); 1478 amdgpu_gmc_agp_location(adev, mc); 1479 } 1480 /* base offset of vram pages */ 1481 adev->vm_manager.vram_base_offset = adev->gfxhub.funcs->get_mc_fb_offset(adev); 1482 1483 /* XXX: add the xgmi offset of the physical node? */ 1484 adev->vm_manager.vram_base_offset += 1485 adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size; 1486 } 1487 1488 /** 1489 * gmc_v9_0_mc_init - initialize the memory controller driver params 1490 * 1491 * @adev: amdgpu_device pointer 1492 * 1493 * Look up the amount of vram, vram width, and decide how to place 1494 * vram and gart within the GPU's physical address space. 1495 * Returns 0 for success. 1496 */ 1497 static int gmc_v9_0_mc_init(struct amdgpu_device *adev) 1498 { 1499 int r; 1500 1501 /* size in MB on si */ 1502 adev->gmc.mc_vram_size = 1503 adev->nbio.funcs->get_memsize(adev) * 1024ULL * 1024ULL; 1504 adev->gmc.real_vram_size = adev->gmc.mc_vram_size; 1505 1506 if (!(adev->flags & AMD_IS_APU) && 1507 !adev->gmc.xgmi.connected_to_cpu) { 1508 r = amdgpu_device_resize_fb_bar(adev); 1509 if (r) 1510 return r; 1511 } 1512 adev->gmc.aper_base = pci_resource_start(adev->pdev, 0); 1513 adev->gmc.aper_size = pci_resource_len(adev->pdev, 0); 1514 1515 #ifdef CONFIG_X86_64 1516 /* 1517 * AMD Accelerated Processing Platform (APP) supporting GPU-HOST xgmi 1518 * interface can use VRAM through here as it appears system reserved 1519 * memory in host address space. 1520 * 1521 * For APUs, VRAM is just the stolen system memory and can be accessed 1522 * directly. 1523 * 1524 * Otherwise, use the legacy Host Data Path (HDP) through PCIe BAR. 1525 */ 1526 1527 /* check whether both host-gpu and gpu-gpu xgmi links exist */ 1528 if (((adev->flags & AMD_IS_APU) && !amdgpu_passthrough(adev)) || 1529 (adev->gmc.xgmi.supported && 1530 adev->gmc.xgmi.connected_to_cpu)) { 1531 adev->gmc.aper_base = 1532 adev->gfxhub.funcs->get_mc_fb_offset(adev) + 1533 adev->gmc.xgmi.physical_node_id * 1534 adev->gmc.xgmi.node_segment_size; 1535 adev->gmc.aper_size = adev->gmc.real_vram_size; 1536 } 1537 1538 #endif 1539 /* In case the PCI BAR is larger than the actual amount of vram */ 1540 adev->gmc.visible_vram_size = adev->gmc.aper_size; 1541 if (adev->gmc.visible_vram_size > adev->gmc.real_vram_size) 1542 adev->gmc.visible_vram_size = adev->gmc.real_vram_size; 1543 1544 /* set the gart size */ 1545 if (amdgpu_gart_size == -1) { 1546 switch (adev->ip_versions[GC_HWIP][0]) { 1547 case IP_VERSION(9, 0, 1): /* all engines support GPUVM */ 1548 case IP_VERSION(9, 2, 1): /* all engines support GPUVM */ 1549 case IP_VERSION(9, 4, 0): 1550 case IP_VERSION(9, 4, 1): 1551 case IP_VERSION(9, 4, 2): 1552 default: 1553 adev->gmc.gart_size = 512ULL << 20; 1554 break; 1555 case IP_VERSION(9, 1, 0): /* DCE SG support */ 1556 case IP_VERSION(9, 2, 2): /* DCE SG support */ 1557 case IP_VERSION(9, 3, 0): 1558 adev->gmc.gart_size = 1024ULL << 20; 1559 break; 1560 } 1561 } else { 1562 adev->gmc.gart_size = (u64)amdgpu_gart_size << 20; 1563 } 1564 1565 adev->gmc.gart_size += adev->pm.smu_prv_buffer_size; 1566 1567 gmc_v9_0_vram_gtt_location(adev, &adev->gmc); 1568 1569 return 0; 1570 } 1571 1572 static int gmc_v9_0_gart_init(struct amdgpu_device *adev) 1573 { 1574 int r; 1575 1576 if (adev->gart.bo) { 1577 WARN(1, "VEGA10 PCIE GART already initialized\n"); 1578 return 0; 1579 } 1580 1581 if (adev->gmc.xgmi.connected_to_cpu) { 1582 adev->gmc.vmid0_page_table_depth = 1; 1583 adev->gmc.vmid0_page_table_block_size = 12; 1584 } else { 1585 adev->gmc.vmid0_page_table_depth = 0; 1586 adev->gmc.vmid0_page_table_block_size = 0; 1587 } 1588 1589 /* Initialize common gart structure */ 1590 r = amdgpu_gart_init(adev); 1591 if (r) 1592 return r; 1593 adev->gart.table_size = adev->gart.num_gpu_pages * 8; 1594 adev->gart.gart_pte_flags = AMDGPU_PTE_MTYPE_VG10(MTYPE_UC) | 1595 AMDGPU_PTE_EXECUTABLE; 1596 1597 r = amdgpu_gart_table_vram_alloc(adev); 1598 if (r) 1599 return r; 1600 1601 if (adev->gmc.xgmi.connected_to_cpu) { 1602 r = amdgpu_gmc_pdb0_alloc(adev); 1603 } 1604 1605 return r; 1606 } 1607 1608 /** 1609 * gmc_v9_0_save_registers - saves regs 1610 * 1611 * @adev: amdgpu_device pointer 1612 * 1613 * This saves potential register values that should be 1614 * restored upon resume 1615 */ 1616 static void gmc_v9_0_save_registers(struct amdgpu_device *adev) 1617 { 1618 if ((adev->ip_versions[DCE_HWIP][0] == IP_VERSION(1, 0, 0)) || 1619 (adev->ip_versions[DCE_HWIP][0] == IP_VERSION(1, 0, 1))) 1620 adev->gmc.sdpif_register = RREG32_SOC15(DCE, 0, mmDCHUBBUB_SDPIF_MMIO_CNTRL_0); 1621 } 1622 1623 static int gmc_v9_0_sw_init(void *handle) 1624 { 1625 int r, vram_width = 0, vram_type = 0, vram_vendor = 0, dma_addr_bits; 1626 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1627 1628 adev->gfxhub.funcs->init(adev); 1629 1630 adev->mmhub.funcs->init(adev); 1631 if (adev->mca.funcs) 1632 adev->mca.funcs->init(adev); 1633 1634 spin_lock_init(&adev->gmc.invalidate_lock); 1635 1636 r = amdgpu_atomfirmware_get_vram_info(adev, 1637 &vram_width, &vram_type, &vram_vendor); 1638 if (amdgpu_sriov_vf(adev)) 1639 /* For Vega10 SR-IOV, vram_width can't be read from ATOM as RAVEN, 1640 * and DF related registers is not readable, seems hardcord is the 1641 * only way to set the correct vram_width 1642 */ 1643 adev->gmc.vram_width = 2048; 1644 else if (amdgpu_emu_mode != 1) 1645 adev->gmc.vram_width = vram_width; 1646 1647 if (!adev->gmc.vram_width) { 1648 int chansize, numchan; 1649 1650 /* hbm memory channel size */ 1651 if (adev->flags & AMD_IS_APU) 1652 chansize = 64; 1653 else 1654 chansize = 128; 1655 if (adev->df.funcs && 1656 adev->df.funcs->get_hbm_channel_number) { 1657 numchan = adev->df.funcs->get_hbm_channel_number(adev); 1658 adev->gmc.vram_width = numchan * chansize; 1659 } 1660 } 1661 1662 adev->gmc.vram_type = vram_type; 1663 adev->gmc.vram_vendor = vram_vendor; 1664 switch (adev->ip_versions[GC_HWIP][0]) { 1665 case IP_VERSION(9, 1, 0): 1666 case IP_VERSION(9, 2, 2): 1667 adev->num_vmhubs = 2; 1668 1669 if (adev->rev_id == 0x0 || adev->rev_id == 0x1) { 1670 amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48); 1671 } else { 1672 /* vm_size is 128TB + 512GB for legacy 3-level page support */ 1673 amdgpu_vm_adjust_size(adev, 128 * 1024 + 512, 9, 2, 48); 1674 adev->gmc.translate_further = 1675 adev->vm_manager.num_level > 1; 1676 } 1677 break; 1678 case IP_VERSION(9, 0, 1): 1679 case IP_VERSION(9, 2, 1): 1680 case IP_VERSION(9, 4, 0): 1681 case IP_VERSION(9, 3, 0): 1682 case IP_VERSION(9, 4, 2): 1683 adev->num_vmhubs = 2; 1684 1685 1686 /* 1687 * To fulfill 4-level page support, 1688 * vm size is 256TB (48bit), maximum size of Vega10, 1689 * block size 512 (9bit) 1690 */ 1691 /* sriov restrict max_pfn below AMDGPU_GMC_HOLE */ 1692 if (amdgpu_sriov_vf(adev)) 1693 amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 47); 1694 else 1695 amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48); 1696 if (adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 2)) 1697 adev->gmc.translate_further = adev->vm_manager.num_level > 1; 1698 break; 1699 case IP_VERSION(9, 4, 1): 1700 adev->num_vmhubs = 3; 1701 1702 /* Keep the vm size same with Vega20 */ 1703 amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48); 1704 adev->gmc.translate_further = adev->vm_manager.num_level > 1; 1705 break; 1706 default: 1707 break; 1708 } 1709 1710 /* This interrupt is VMC page fault.*/ 1711 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC, VMC_1_0__SRCID__VM_FAULT, 1712 &adev->gmc.vm_fault); 1713 if (r) 1714 return r; 1715 1716 if (adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 1)) { 1717 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC1, VMC_1_0__SRCID__VM_FAULT, 1718 &adev->gmc.vm_fault); 1719 if (r) 1720 return r; 1721 } 1722 1723 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_UTCL2, UTCL2_1_0__SRCID__FAULT, 1724 &adev->gmc.vm_fault); 1725 1726 if (r) 1727 return r; 1728 1729 if (!amdgpu_sriov_vf(adev) && 1730 !adev->gmc.xgmi.connected_to_cpu) { 1731 /* interrupt sent to DF. */ 1732 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_DF, 0, 1733 &adev->gmc.ecc_irq); 1734 if (r) 1735 return r; 1736 } 1737 1738 /* Set the internal MC address mask 1739 * This is the max address of the GPU's 1740 * internal address space. 1741 */ 1742 adev->gmc.mc_mask = 0xffffffffffffULL; /* 48 bit MC */ 1743 1744 dma_addr_bits = adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 2) ? 48:44; 1745 r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(dma_addr_bits)); 1746 if (r) { 1747 printk(KERN_WARNING "amdgpu: No suitable DMA available.\n"); 1748 return r; 1749 } 1750 adev->need_swiotlb = drm_need_swiotlb(dma_addr_bits); 1751 1752 r = gmc_v9_0_mc_init(adev); 1753 if (r) 1754 return r; 1755 1756 amdgpu_gmc_get_vbios_allocations(adev); 1757 1758 /* Memory manager */ 1759 r = amdgpu_bo_init(adev); 1760 if (r) 1761 return r; 1762 1763 r = gmc_v9_0_gart_init(adev); 1764 if (r) 1765 return r; 1766 1767 /* 1768 * number of VMs 1769 * VMID 0 is reserved for System 1770 * amdgpu graphics/compute will use VMIDs 1..n-1 1771 * amdkfd will use VMIDs n..15 1772 * 1773 * The first KFD VMID is 8 for GPUs with graphics, 3 for 1774 * compute-only GPUs. On compute-only GPUs that leaves 2 VMIDs 1775 * for video processing. 1776 */ 1777 adev->vm_manager.first_kfd_vmid = 1778 (adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 1) || 1779 adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 2)) ? 3 : 8; 1780 1781 amdgpu_vm_manager_init(adev); 1782 1783 gmc_v9_0_save_registers(adev); 1784 1785 return 0; 1786 } 1787 1788 static int gmc_v9_0_sw_fini(void *handle) 1789 { 1790 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1791 1792 amdgpu_gmc_ras_fini(adev); 1793 amdgpu_gem_force_release(adev); 1794 amdgpu_vm_manager_fini(adev); 1795 amdgpu_gart_table_vram_free(adev); 1796 amdgpu_bo_free_kernel(&adev->gmc.pdb0_bo, NULL, &adev->gmc.ptr_pdb0); 1797 amdgpu_bo_fini(adev); 1798 1799 return 0; 1800 } 1801 1802 static void gmc_v9_0_init_golden_registers(struct amdgpu_device *adev) 1803 { 1804 1805 switch (adev->ip_versions[MMHUB_HWIP][0]) { 1806 case IP_VERSION(9, 0, 0): 1807 if (amdgpu_sriov_vf(adev)) 1808 break; 1809 fallthrough; 1810 case IP_VERSION(9, 4, 0): 1811 soc15_program_register_sequence(adev, 1812 golden_settings_mmhub_1_0_0, 1813 ARRAY_SIZE(golden_settings_mmhub_1_0_0)); 1814 soc15_program_register_sequence(adev, 1815 golden_settings_athub_1_0_0, 1816 ARRAY_SIZE(golden_settings_athub_1_0_0)); 1817 break; 1818 case IP_VERSION(9, 1, 0): 1819 case IP_VERSION(9, 2, 0): 1820 /* TODO for renoir */ 1821 soc15_program_register_sequence(adev, 1822 golden_settings_athub_1_0_0, 1823 ARRAY_SIZE(golden_settings_athub_1_0_0)); 1824 break; 1825 default: 1826 break; 1827 } 1828 } 1829 1830 /** 1831 * gmc_v9_0_restore_registers - restores regs 1832 * 1833 * @adev: amdgpu_device pointer 1834 * 1835 * This restores register values, saved at suspend. 1836 */ 1837 void gmc_v9_0_restore_registers(struct amdgpu_device *adev) 1838 { 1839 if ((adev->ip_versions[DCE_HWIP][0] == IP_VERSION(1, 0, 0)) || 1840 (adev->ip_versions[DCE_HWIP][0] == IP_VERSION(1, 0, 1))) { 1841 WREG32_SOC15(DCE, 0, mmDCHUBBUB_SDPIF_MMIO_CNTRL_0, adev->gmc.sdpif_register); 1842 WARN_ON(adev->gmc.sdpif_register != 1843 RREG32_SOC15(DCE, 0, mmDCHUBBUB_SDPIF_MMIO_CNTRL_0)); 1844 } 1845 } 1846 1847 /** 1848 * gmc_v9_0_gart_enable - gart enable 1849 * 1850 * @adev: amdgpu_device pointer 1851 */ 1852 static int gmc_v9_0_gart_enable(struct amdgpu_device *adev) 1853 { 1854 int r; 1855 1856 if (adev->gmc.xgmi.connected_to_cpu) 1857 amdgpu_gmc_init_pdb0(adev); 1858 1859 if (adev->gart.bo == NULL) { 1860 dev_err(adev->dev, "No VRAM object for PCIE GART.\n"); 1861 return -EINVAL; 1862 } 1863 1864 amdgpu_gtt_mgr_recover(&adev->mman.gtt_mgr); 1865 r = adev->gfxhub.funcs->gart_enable(adev); 1866 if (r) 1867 return r; 1868 1869 r = adev->mmhub.funcs->gart_enable(adev); 1870 if (r) 1871 return r; 1872 1873 DRM_INFO("PCIE GART of %uM enabled.\n", 1874 (unsigned)(adev->gmc.gart_size >> 20)); 1875 if (adev->gmc.pdb0_bo) 1876 DRM_INFO("PDB0 located at 0x%016llX\n", 1877 (unsigned long long)amdgpu_bo_gpu_offset(adev->gmc.pdb0_bo)); 1878 DRM_INFO("PTB located at 0x%016llX\n", 1879 (unsigned long long)amdgpu_bo_gpu_offset(adev->gart.bo)); 1880 1881 return 0; 1882 } 1883 1884 static int gmc_v9_0_hw_init(void *handle) 1885 { 1886 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1887 bool value; 1888 int i, r; 1889 1890 /* The sequence of these two function calls matters.*/ 1891 gmc_v9_0_init_golden_registers(adev); 1892 1893 if (adev->mode_info.num_crtc) { 1894 /* Lockout access through VGA aperture*/ 1895 WREG32_FIELD15(DCE, 0, VGA_HDP_CONTROL, VGA_MEMORY_DISABLE, 1); 1896 /* disable VGA render */ 1897 WREG32_FIELD15(DCE, 0, VGA_RENDER_CONTROL, VGA_VSTATUS_CNTL, 0); 1898 } 1899 1900 if (adev->mmhub.funcs->update_power_gating) 1901 adev->mmhub.funcs->update_power_gating(adev, true); 1902 1903 adev->hdp.funcs->init_registers(adev); 1904 1905 /* After HDP is initialized, flush HDP.*/ 1906 adev->hdp.funcs->flush_hdp(adev, NULL); 1907 1908 if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS) 1909 value = false; 1910 else 1911 value = true; 1912 1913 if (!amdgpu_sriov_vf(adev)) { 1914 adev->gfxhub.funcs->set_fault_enable_default(adev, value); 1915 adev->mmhub.funcs->set_fault_enable_default(adev, value); 1916 } 1917 for (i = 0; i < adev->num_vmhubs; ++i) 1918 gmc_v9_0_flush_gpu_tlb(adev, 0, i, 0); 1919 1920 if (adev->umc.funcs && adev->umc.funcs->init_registers) 1921 adev->umc.funcs->init_registers(adev); 1922 1923 r = gmc_v9_0_gart_enable(adev); 1924 if (r) 1925 return r; 1926 1927 if (amdgpu_emu_mode == 1) 1928 return amdgpu_gmc_vram_checking(adev); 1929 else 1930 return r; 1931 } 1932 1933 /** 1934 * gmc_v9_0_gart_disable - gart disable 1935 * 1936 * @adev: amdgpu_device pointer 1937 * 1938 * This disables all VM page table. 1939 */ 1940 static void gmc_v9_0_gart_disable(struct amdgpu_device *adev) 1941 { 1942 adev->gfxhub.funcs->gart_disable(adev); 1943 adev->mmhub.funcs->gart_disable(adev); 1944 } 1945 1946 static int gmc_v9_0_hw_fini(void *handle) 1947 { 1948 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1949 1950 gmc_v9_0_gart_disable(adev); 1951 1952 if (amdgpu_sriov_vf(adev)) { 1953 /* full access mode, so don't touch any GMC register */ 1954 DRM_DEBUG("For SRIOV client, shouldn't do anything.\n"); 1955 return 0; 1956 } 1957 1958 /* 1959 * Pair the operations did in gmc_v9_0_hw_init and thus maintain 1960 * a correct cached state for GMC. Otherwise, the "gate" again 1961 * operation on S3 resuming will fail due to wrong cached state. 1962 */ 1963 if (adev->mmhub.funcs->update_power_gating) 1964 adev->mmhub.funcs->update_power_gating(adev, false); 1965 1966 amdgpu_irq_put(adev, &adev->gmc.ecc_irq, 0); 1967 amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0); 1968 1969 return 0; 1970 } 1971 1972 static int gmc_v9_0_suspend(void *handle) 1973 { 1974 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1975 1976 return gmc_v9_0_hw_fini(adev); 1977 } 1978 1979 static int gmc_v9_0_resume(void *handle) 1980 { 1981 int r; 1982 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1983 1984 r = gmc_v9_0_hw_init(adev); 1985 if (r) 1986 return r; 1987 1988 amdgpu_vmid_reset_all(adev); 1989 1990 return 0; 1991 } 1992 1993 static bool gmc_v9_0_is_idle(void *handle) 1994 { 1995 /* MC is always ready in GMC v9.*/ 1996 return true; 1997 } 1998 1999 static int gmc_v9_0_wait_for_idle(void *handle) 2000 { 2001 /* There is no need to wait for MC idle in GMC v9.*/ 2002 return 0; 2003 } 2004 2005 static int gmc_v9_0_soft_reset(void *handle) 2006 { 2007 /* XXX for emulation.*/ 2008 return 0; 2009 } 2010 2011 static int gmc_v9_0_set_clockgating_state(void *handle, 2012 enum amd_clockgating_state state) 2013 { 2014 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 2015 2016 adev->mmhub.funcs->set_clockgating(adev, state); 2017 2018 athub_v1_0_set_clockgating(adev, state); 2019 2020 return 0; 2021 } 2022 2023 static void gmc_v9_0_get_clockgating_state(void *handle, u64 *flags) 2024 { 2025 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 2026 2027 adev->mmhub.funcs->get_clockgating(adev, flags); 2028 2029 athub_v1_0_get_clockgating(adev, flags); 2030 } 2031 2032 static int gmc_v9_0_set_powergating_state(void *handle, 2033 enum amd_powergating_state state) 2034 { 2035 return 0; 2036 } 2037 2038 const struct amd_ip_funcs gmc_v9_0_ip_funcs = { 2039 .name = "gmc_v9_0", 2040 .early_init = gmc_v9_0_early_init, 2041 .late_init = gmc_v9_0_late_init, 2042 .sw_init = gmc_v9_0_sw_init, 2043 .sw_fini = gmc_v9_0_sw_fini, 2044 .hw_init = gmc_v9_0_hw_init, 2045 .hw_fini = gmc_v9_0_hw_fini, 2046 .suspend = gmc_v9_0_suspend, 2047 .resume = gmc_v9_0_resume, 2048 .is_idle = gmc_v9_0_is_idle, 2049 .wait_for_idle = gmc_v9_0_wait_for_idle, 2050 .soft_reset = gmc_v9_0_soft_reset, 2051 .set_clockgating_state = gmc_v9_0_set_clockgating_state, 2052 .set_powergating_state = gmc_v9_0_set_powergating_state, 2053 .get_clockgating_state = gmc_v9_0_get_clockgating_state, 2054 }; 2055 2056 const struct amdgpu_ip_block_version gmc_v9_0_ip_block = 2057 { 2058 .type = AMD_IP_BLOCK_TYPE_GMC, 2059 .major = 9, 2060 .minor = 0, 2061 .rev = 0, 2062 .funcs = &gmc_v9_0_ip_funcs, 2063 }; 2064