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 "gfxhub_v1_2.h" 53 #include "mmhub_v9_4.h" 54 #include "mmhub_v1_7.h" 55 #include "mmhub_v1_8.h" 56 #include "umc_v6_1.h" 57 #include "umc_v6_0.h" 58 #include "umc_v6_7.h" 59 #include "hdp_v4_0.h" 60 #include "mca_v3_0.h" 61 62 #include "ivsrcid/vmc/irqsrcs_vmc_1_0.h" 63 64 #include "amdgpu_ras.h" 65 #include "amdgpu_xgmi.h" 66 67 #include "amdgpu_reset.h" 68 69 /* add these here since we already include dce12 headers and these are for DCN */ 70 #define mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION 0x055d 71 #define mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION_BASE_IDX 2 72 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_WIDTH__SHIFT 0x0 73 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_HEIGHT__SHIFT 0x10 74 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_WIDTH_MASK 0x00003FFFL 75 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_HEIGHT_MASK 0x3FFF0000L 76 #define mmDCHUBBUB_SDPIF_MMIO_CNTRL_0 0x049d 77 #define mmDCHUBBUB_SDPIF_MMIO_CNTRL_0_BASE_IDX 2 78 79 #define mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION_DCN2 0x05ea 80 #define mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION_DCN2_BASE_IDX 2 81 82 83 static const char *gfxhub_client_ids[] = { 84 "CB", 85 "DB", 86 "IA", 87 "WD", 88 "CPF", 89 "CPC", 90 "CPG", 91 "RLC", 92 "TCP", 93 "SQC (inst)", 94 "SQC (data)", 95 "SQG", 96 "PA", 97 }; 98 99 static const char *mmhub_client_ids_raven[][2] = { 100 [0][0] = "MP1", 101 [1][0] = "MP0", 102 [2][0] = "VCN", 103 [3][0] = "VCNU", 104 [4][0] = "HDP", 105 [5][0] = "DCE", 106 [13][0] = "UTCL2", 107 [19][0] = "TLS", 108 [26][0] = "OSS", 109 [27][0] = "SDMA0", 110 [0][1] = "MP1", 111 [1][1] = "MP0", 112 [2][1] = "VCN", 113 [3][1] = "VCNU", 114 [4][1] = "HDP", 115 [5][1] = "XDP", 116 [6][1] = "DBGU0", 117 [7][1] = "DCE", 118 [8][1] = "DCEDWB0", 119 [9][1] = "DCEDWB1", 120 [26][1] = "OSS", 121 [27][1] = "SDMA0", 122 }; 123 124 static const char *mmhub_client_ids_renoir[][2] = { 125 [0][0] = "MP1", 126 [1][0] = "MP0", 127 [2][0] = "HDP", 128 [4][0] = "DCEDMC", 129 [5][0] = "DCEVGA", 130 [13][0] = "UTCL2", 131 [19][0] = "TLS", 132 [26][0] = "OSS", 133 [27][0] = "SDMA0", 134 [28][0] = "VCN", 135 [29][0] = "VCNU", 136 [30][0] = "JPEG", 137 [0][1] = "MP1", 138 [1][1] = "MP0", 139 [2][1] = "HDP", 140 [3][1] = "XDP", 141 [6][1] = "DBGU0", 142 [7][1] = "DCEDMC", 143 [8][1] = "DCEVGA", 144 [9][1] = "DCEDWB", 145 [26][1] = "OSS", 146 [27][1] = "SDMA0", 147 [28][1] = "VCN", 148 [29][1] = "VCNU", 149 [30][1] = "JPEG", 150 }; 151 152 static const char *mmhub_client_ids_vega10[][2] = { 153 [0][0] = "MP0", 154 [1][0] = "UVD", 155 [2][0] = "UVDU", 156 [3][0] = "HDP", 157 [13][0] = "UTCL2", 158 [14][0] = "OSS", 159 [15][0] = "SDMA1", 160 [32+0][0] = "VCE0", 161 [32+1][0] = "VCE0U", 162 [32+2][0] = "XDMA", 163 [32+3][0] = "DCE", 164 [32+4][0] = "MP1", 165 [32+14][0] = "SDMA0", 166 [0][1] = "MP0", 167 [1][1] = "UVD", 168 [2][1] = "UVDU", 169 [3][1] = "DBGU0", 170 [4][1] = "HDP", 171 [5][1] = "XDP", 172 [14][1] = "OSS", 173 [15][1] = "SDMA0", 174 [32+0][1] = "VCE0", 175 [32+1][1] = "VCE0U", 176 [32+2][1] = "XDMA", 177 [32+3][1] = "DCE", 178 [32+4][1] = "DCEDWB", 179 [32+5][1] = "MP1", 180 [32+6][1] = "DBGU1", 181 [32+14][1] = "SDMA1", 182 }; 183 184 static const char *mmhub_client_ids_vega12[][2] = { 185 [0][0] = "MP0", 186 [1][0] = "VCE0", 187 [2][0] = "VCE0U", 188 [3][0] = "HDP", 189 [13][0] = "UTCL2", 190 [14][0] = "OSS", 191 [15][0] = "SDMA1", 192 [32+0][0] = "DCE", 193 [32+1][0] = "XDMA", 194 [32+2][0] = "UVD", 195 [32+3][0] = "UVDU", 196 [32+4][0] = "MP1", 197 [32+15][0] = "SDMA0", 198 [0][1] = "MP0", 199 [1][1] = "VCE0", 200 [2][1] = "VCE0U", 201 [3][1] = "DBGU0", 202 [4][1] = "HDP", 203 [5][1] = "XDP", 204 [14][1] = "OSS", 205 [15][1] = "SDMA0", 206 [32+0][1] = "DCE", 207 [32+1][1] = "DCEDWB", 208 [32+2][1] = "XDMA", 209 [32+3][1] = "UVD", 210 [32+4][1] = "UVDU", 211 [32+5][1] = "MP1", 212 [32+6][1] = "DBGU1", 213 [32+15][1] = "SDMA1", 214 }; 215 216 static const char *mmhub_client_ids_vega20[][2] = { 217 [0][0] = "XDMA", 218 [1][0] = "DCE", 219 [2][0] = "VCE0", 220 [3][0] = "VCE0U", 221 [4][0] = "UVD", 222 [5][0] = "UVD1U", 223 [13][0] = "OSS", 224 [14][0] = "HDP", 225 [15][0] = "SDMA0", 226 [32+0][0] = "UVD", 227 [32+1][0] = "UVDU", 228 [32+2][0] = "MP1", 229 [32+3][0] = "MP0", 230 [32+12][0] = "UTCL2", 231 [32+14][0] = "SDMA1", 232 [0][1] = "XDMA", 233 [1][1] = "DCE", 234 [2][1] = "DCEDWB", 235 [3][1] = "VCE0", 236 [4][1] = "VCE0U", 237 [5][1] = "UVD1", 238 [6][1] = "UVD1U", 239 [7][1] = "DBGU0", 240 [8][1] = "XDP", 241 [13][1] = "OSS", 242 [14][1] = "HDP", 243 [15][1] = "SDMA0", 244 [32+0][1] = "UVD", 245 [32+1][1] = "UVDU", 246 [32+2][1] = "DBGU1", 247 [32+3][1] = "MP1", 248 [32+4][1] = "MP0", 249 [32+14][1] = "SDMA1", 250 }; 251 252 static const char *mmhub_client_ids_arcturus[][2] = { 253 [0][0] = "DBGU1", 254 [1][0] = "XDP", 255 [2][0] = "MP1", 256 [14][0] = "HDP", 257 [171][0] = "JPEG", 258 [172][0] = "VCN", 259 [173][0] = "VCNU", 260 [203][0] = "JPEG1", 261 [204][0] = "VCN1", 262 [205][0] = "VCN1U", 263 [256][0] = "SDMA0", 264 [257][0] = "SDMA1", 265 [258][0] = "SDMA2", 266 [259][0] = "SDMA3", 267 [260][0] = "SDMA4", 268 [261][0] = "SDMA5", 269 [262][0] = "SDMA6", 270 [263][0] = "SDMA7", 271 [384][0] = "OSS", 272 [0][1] = "DBGU1", 273 [1][1] = "XDP", 274 [2][1] = "MP1", 275 [14][1] = "HDP", 276 [171][1] = "JPEG", 277 [172][1] = "VCN", 278 [173][1] = "VCNU", 279 [203][1] = "JPEG1", 280 [204][1] = "VCN1", 281 [205][1] = "VCN1U", 282 [256][1] = "SDMA0", 283 [257][1] = "SDMA1", 284 [258][1] = "SDMA2", 285 [259][1] = "SDMA3", 286 [260][1] = "SDMA4", 287 [261][1] = "SDMA5", 288 [262][1] = "SDMA6", 289 [263][1] = "SDMA7", 290 [384][1] = "OSS", 291 }; 292 293 static const char *mmhub_client_ids_aldebaran[][2] = { 294 [2][0] = "MP1", 295 [3][0] = "MP0", 296 [32+1][0] = "DBGU_IO0", 297 [32+2][0] = "DBGU_IO2", 298 [32+4][0] = "MPIO", 299 [96+11][0] = "JPEG0", 300 [96+12][0] = "VCN0", 301 [96+13][0] = "VCNU0", 302 [128+11][0] = "JPEG1", 303 [128+12][0] = "VCN1", 304 [128+13][0] = "VCNU1", 305 [160+1][0] = "XDP", 306 [160+14][0] = "HDP", 307 [256+0][0] = "SDMA0", 308 [256+1][0] = "SDMA1", 309 [256+2][0] = "SDMA2", 310 [256+3][0] = "SDMA3", 311 [256+4][0] = "SDMA4", 312 [384+0][0] = "OSS", 313 [2][1] = "MP1", 314 [3][1] = "MP0", 315 [32+1][1] = "DBGU_IO0", 316 [32+2][1] = "DBGU_IO2", 317 [32+4][1] = "MPIO", 318 [96+11][1] = "JPEG0", 319 [96+12][1] = "VCN0", 320 [96+13][1] = "VCNU0", 321 [128+11][1] = "JPEG1", 322 [128+12][1] = "VCN1", 323 [128+13][1] = "VCNU1", 324 [160+1][1] = "XDP", 325 [160+14][1] = "HDP", 326 [256+0][1] = "SDMA0", 327 [256+1][1] = "SDMA1", 328 [256+2][1] = "SDMA2", 329 [256+3][1] = "SDMA3", 330 [256+4][1] = "SDMA4", 331 [384+0][1] = "OSS", 332 }; 333 334 static const struct soc15_reg_golden golden_settings_mmhub_1_0_0[] = 335 { 336 SOC15_REG_GOLDEN_VALUE(MMHUB, 0, mmDAGB1_WRCLI2, 0x00000007, 0xfe5fe0fa), 337 SOC15_REG_GOLDEN_VALUE(MMHUB, 0, mmMMEA1_DRAM_WR_CLI2GRP_MAP0, 0x00000030, 0x55555565) 338 }; 339 340 static const struct soc15_reg_golden golden_settings_athub_1_0_0[] = 341 { 342 SOC15_REG_GOLDEN_VALUE(ATHUB, 0, mmRPB_ARB_CNTL, 0x0000ff00, 0x00000800), 343 SOC15_REG_GOLDEN_VALUE(ATHUB, 0, mmRPB_ARB_CNTL2, 0x00ff00ff, 0x00080008) 344 }; 345 346 static const uint32_t ecc_umc_mcumc_ctrl_addrs[] = { 347 (0x000143c0 + 0x00000000), 348 (0x000143c0 + 0x00000800), 349 (0x000143c0 + 0x00001000), 350 (0x000143c0 + 0x00001800), 351 (0x000543c0 + 0x00000000), 352 (0x000543c0 + 0x00000800), 353 (0x000543c0 + 0x00001000), 354 (0x000543c0 + 0x00001800), 355 (0x000943c0 + 0x00000000), 356 (0x000943c0 + 0x00000800), 357 (0x000943c0 + 0x00001000), 358 (0x000943c0 + 0x00001800), 359 (0x000d43c0 + 0x00000000), 360 (0x000d43c0 + 0x00000800), 361 (0x000d43c0 + 0x00001000), 362 (0x000d43c0 + 0x00001800), 363 (0x001143c0 + 0x00000000), 364 (0x001143c0 + 0x00000800), 365 (0x001143c0 + 0x00001000), 366 (0x001143c0 + 0x00001800), 367 (0x001543c0 + 0x00000000), 368 (0x001543c0 + 0x00000800), 369 (0x001543c0 + 0x00001000), 370 (0x001543c0 + 0x00001800), 371 (0x001943c0 + 0x00000000), 372 (0x001943c0 + 0x00000800), 373 (0x001943c0 + 0x00001000), 374 (0x001943c0 + 0x00001800), 375 (0x001d43c0 + 0x00000000), 376 (0x001d43c0 + 0x00000800), 377 (0x001d43c0 + 0x00001000), 378 (0x001d43c0 + 0x00001800), 379 }; 380 381 static const uint32_t ecc_umc_mcumc_ctrl_mask_addrs[] = { 382 (0x000143e0 + 0x00000000), 383 (0x000143e0 + 0x00000800), 384 (0x000143e0 + 0x00001000), 385 (0x000143e0 + 0x00001800), 386 (0x000543e0 + 0x00000000), 387 (0x000543e0 + 0x00000800), 388 (0x000543e0 + 0x00001000), 389 (0x000543e0 + 0x00001800), 390 (0x000943e0 + 0x00000000), 391 (0x000943e0 + 0x00000800), 392 (0x000943e0 + 0x00001000), 393 (0x000943e0 + 0x00001800), 394 (0x000d43e0 + 0x00000000), 395 (0x000d43e0 + 0x00000800), 396 (0x000d43e0 + 0x00001000), 397 (0x000d43e0 + 0x00001800), 398 (0x001143e0 + 0x00000000), 399 (0x001143e0 + 0x00000800), 400 (0x001143e0 + 0x00001000), 401 (0x001143e0 + 0x00001800), 402 (0x001543e0 + 0x00000000), 403 (0x001543e0 + 0x00000800), 404 (0x001543e0 + 0x00001000), 405 (0x001543e0 + 0x00001800), 406 (0x001943e0 + 0x00000000), 407 (0x001943e0 + 0x00000800), 408 (0x001943e0 + 0x00001000), 409 (0x001943e0 + 0x00001800), 410 (0x001d43e0 + 0x00000000), 411 (0x001d43e0 + 0x00000800), 412 (0x001d43e0 + 0x00001000), 413 (0x001d43e0 + 0x00001800), 414 }; 415 416 static int gmc_v9_0_ecc_interrupt_state(struct amdgpu_device *adev, 417 struct amdgpu_irq_src *src, 418 unsigned type, 419 enum amdgpu_interrupt_state state) 420 { 421 u32 bits, i, tmp, reg; 422 423 /* Devices newer then VEGA10/12 shall have these programming 424 sequences performed by PSP BL */ 425 if (adev->asic_type >= CHIP_VEGA20) 426 return 0; 427 428 bits = 0x7f; 429 430 switch (state) { 431 case AMDGPU_IRQ_STATE_DISABLE: 432 for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_addrs); i++) { 433 reg = ecc_umc_mcumc_ctrl_addrs[i]; 434 tmp = RREG32(reg); 435 tmp &= ~bits; 436 WREG32(reg, tmp); 437 } 438 for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_mask_addrs); i++) { 439 reg = ecc_umc_mcumc_ctrl_mask_addrs[i]; 440 tmp = RREG32(reg); 441 tmp &= ~bits; 442 WREG32(reg, tmp); 443 } 444 break; 445 case AMDGPU_IRQ_STATE_ENABLE: 446 for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_addrs); i++) { 447 reg = ecc_umc_mcumc_ctrl_addrs[i]; 448 tmp = RREG32(reg); 449 tmp |= bits; 450 WREG32(reg, tmp); 451 } 452 for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_mask_addrs); i++) { 453 reg = ecc_umc_mcumc_ctrl_mask_addrs[i]; 454 tmp = RREG32(reg); 455 tmp |= bits; 456 WREG32(reg, tmp); 457 } 458 break; 459 default: 460 break; 461 } 462 463 return 0; 464 } 465 466 static int gmc_v9_0_vm_fault_interrupt_state(struct amdgpu_device *adev, 467 struct amdgpu_irq_src *src, 468 unsigned type, 469 enum amdgpu_interrupt_state state) 470 { 471 struct amdgpu_vmhub *hub; 472 u32 tmp, reg, bits, i, j; 473 474 bits = VM_CONTEXT1_CNTL__RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 475 VM_CONTEXT1_CNTL__DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 476 VM_CONTEXT1_CNTL__PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 477 VM_CONTEXT1_CNTL__VALID_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 478 VM_CONTEXT1_CNTL__READ_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 479 VM_CONTEXT1_CNTL__WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 480 VM_CONTEXT1_CNTL__EXECUTE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK; 481 482 switch (state) { 483 case AMDGPU_IRQ_STATE_DISABLE: 484 for (j = 0; j < adev->num_vmhubs; j++) { 485 hub = &adev->vmhub[j]; 486 for (i = 0; i < 16; i++) { 487 reg = hub->vm_context0_cntl + i; 488 489 /* This works because this interrupt is only 490 * enabled at init/resume and disabled in 491 * fini/suspend, so the overall state doesn't 492 * change over the course of suspend/resume. 493 */ 494 if (adev->in_s0ix && (j == AMDGPU_GFXHUB_0)) 495 continue; 496 497 if (j == AMDGPU_GFXHUB_0) 498 tmp = RREG32_SOC15_IP(GC, reg); 499 else 500 tmp = RREG32_SOC15_IP(MMHUB, reg); 501 502 tmp &= ~bits; 503 504 if (j == AMDGPU_GFXHUB_0) 505 WREG32_SOC15_IP(GC, reg, tmp); 506 else 507 WREG32_SOC15_IP(MMHUB, reg, tmp); 508 } 509 } 510 break; 511 case AMDGPU_IRQ_STATE_ENABLE: 512 for (j = 0; j < adev->num_vmhubs; j++) { 513 hub = &adev->vmhub[j]; 514 for (i = 0; i < 16; i++) { 515 reg = hub->vm_context0_cntl + i; 516 517 /* This works because this interrupt is only 518 * enabled at init/resume and disabled in 519 * fini/suspend, so the overall state doesn't 520 * change over the course of suspend/resume. 521 */ 522 if (adev->in_s0ix && (j == AMDGPU_GFXHUB_0)) 523 continue; 524 525 if (j == AMDGPU_GFXHUB_0) 526 tmp = RREG32_SOC15_IP(GC, reg); 527 else 528 tmp = RREG32_SOC15_IP(MMHUB, reg); 529 530 tmp |= bits; 531 532 if (j == AMDGPU_GFXHUB_0) 533 WREG32_SOC15_IP(GC, reg, tmp); 534 else 535 WREG32_SOC15_IP(MMHUB, reg, tmp); 536 } 537 } 538 break; 539 default: 540 break; 541 } 542 543 return 0; 544 } 545 546 static int gmc_v9_0_process_interrupt(struct amdgpu_device *adev, 547 struct amdgpu_irq_src *source, 548 struct amdgpu_iv_entry *entry) 549 { 550 bool retry_fault = !!(entry->src_data[1] & 0x80); 551 bool write_fault = !!(entry->src_data[1] & 0x20); 552 uint32_t status = 0, cid = 0, rw = 0; 553 struct amdgpu_task_info task_info; 554 struct amdgpu_vmhub *hub; 555 const char *mmhub_cid; 556 const char *hub_name; 557 u64 addr; 558 uint32_t cam_index = 0; 559 int ret; 560 561 addr = (u64)entry->src_data[0] << 12; 562 addr |= ((u64)entry->src_data[1] & 0xf) << 44; 563 564 if (retry_fault) { 565 if (adev->irq.retry_cam_enabled) { 566 /* Delegate it to a different ring if the hardware hasn't 567 * already done it. 568 */ 569 if (entry->ih == &adev->irq.ih) { 570 amdgpu_irq_delegate(adev, entry, 8); 571 return 1; 572 } 573 574 cam_index = entry->src_data[2] & 0x3ff; 575 576 ret = amdgpu_vm_handle_fault(adev, entry->pasid, addr, write_fault); 577 WDOORBELL32(adev->irq.retry_cam_doorbell_index, cam_index); 578 if (ret) 579 return 1; 580 } else { 581 /* Process it onyl if it's the first fault for this address */ 582 if (entry->ih != &adev->irq.ih_soft && 583 amdgpu_gmc_filter_faults(adev, entry->ih, addr, entry->pasid, 584 entry->timestamp)) 585 return 1; 586 587 /* Delegate it to a different ring if the hardware hasn't 588 * already done it. 589 */ 590 if (entry->ih == &adev->irq.ih) { 591 amdgpu_irq_delegate(adev, entry, 8); 592 return 1; 593 } 594 595 /* Try to handle the recoverable page faults by filling page 596 * tables 597 */ 598 if (amdgpu_vm_handle_fault(adev, entry->pasid, addr, write_fault)) 599 return 1; 600 } 601 } 602 603 if (!printk_ratelimit()) 604 return 0; 605 606 if (entry->client_id == SOC15_IH_CLIENTID_VMC) { 607 hub_name = "mmhub0"; 608 hub = &adev->vmhub[AMDGPU_MMHUB_0]; 609 } else if (entry->client_id == SOC15_IH_CLIENTID_VMC1) { 610 hub_name = "mmhub1"; 611 hub = &adev->vmhub[AMDGPU_MMHUB_1]; 612 } else { 613 hub_name = "gfxhub0"; 614 hub = &adev->vmhub[AMDGPU_GFXHUB_0]; 615 } 616 617 memset(&task_info, 0, sizeof(struct amdgpu_task_info)); 618 amdgpu_vm_get_task_info(adev, entry->pasid, &task_info); 619 620 dev_err(adev->dev, 621 "[%s] %s page fault (src_id:%u ring:%u vmid:%u " 622 "pasid:%u, for process %s pid %d thread %s pid %d)\n", 623 hub_name, retry_fault ? "retry" : "no-retry", 624 entry->src_id, entry->ring_id, entry->vmid, 625 entry->pasid, task_info.process_name, task_info.tgid, 626 task_info.task_name, task_info.pid); 627 dev_err(adev->dev, " in page starting at address 0x%016llx from IH client 0x%x (%s)\n", 628 addr, entry->client_id, 629 soc15_ih_clientid_name[entry->client_id]); 630 631 if (amdgpu_sriov_vf(adev)) 632 return 0; 633 634 /* 635 * Issue a dummy read to wait for the status register to 636 * be updated to avoid reading an incorrect value due to 637 * the new fast GRBM interface. 638 */ 639 if ((entry->vmid_src == AMDGPU_GFXHUB_0) && 640 (adev->ip_versions[GC_HWIP][0] < IP_VERSION(9, 4, 2))) 641 RREG32(hub->vm_l2_pro_fault_status); 642 643 status = RREG32(hub->vm_l2_pro_fault_status); 644 cid = REG_GET_FIELD(status, VM_L2_PROTECTION_FAULT_STATUS, CID); 645 rw = REG_GET_FIELD(status, VM_L2_PROTECTION_FAULT_STATUS, RW); 646 WREG32_P(hub->vm_l2_pro_fault_cntl, 1, ~1); 647 648 649 dev_err(adev->dev, 650 "VM_L2_PROTECTION_FAULT_STATUS:0x%08X\n", 651 status); 652 if (hub == &adev->vmhub[AMDGPU_GFXHUB_0]) { 653 dev_err(adev->dev, "\t Faulty UTCL2 client ID: %s (0x%x)\n", 654 cid >= ARRAY_SIZE(gfxhub_client_ids) ? "unknown" : 655 gfxhub_client_ids[cid], 656 cid); 657 } else { 658 switch (adev->ip_versions[MMHUB_HWIP][0]) { 659 case IP_VERSION(9, 0, 0): 660 mmhub_cid = mmhub_client_ids_vega10[cid][rw]; 661 break; 662 case IP_VERSION(9, 3, 0): 663 mmhub_cid = mmhub_client_ids_vega12[cid][rw]; 664 break; 665 case IP_VERSION(9, 4, 0): 666 mmhub_cid = mmhub_client_ids_vega20[cid][rw]; 667 break; 668 case IP_VERSION(9, 4, 1): 669 mmhub_cid = mmhub_client_ids_arcturus[cid][rw]; 670 break; 671 case IP_VERSION(9, 1, 0): 672 case IP_VERSION(9, 2, 0): 673 mmhub_cid = mmhub_client_ids_raven[cid][rw]; 674 break; 675 case IP_VERSION(1, 5, 0): 676 case IP_VERSION(2, 4, 0): 677 mmhub_cid = mmhub_client_ids_renoir[cid][rw]; 678 break; 679 case IP_VERSION(1, 8, 0): 680 case IP_VERSION(9, 4, 2): 681 mmhub_cid = mmhub_client_ids_aldebaran[cid][rw]; 682 break; 683 default: 684 mmhub_cid = NULL; 685 break; 686 } 687 dev_err(adev->dev, "\t Faulty UTCL2 client ID: %s (0x%x)\n", 688 mmhub_cid ? mmhub_cid : "unknown", cid); 689 } 690 dev_err(adev->dev, "\t MORE_FAULTS: 0x%lx\n", 691 REG_GET_FIELD(status, 692 VM_L2_PROTECTION_FAULT_STATUS, MORE_FAULTS)); 693 dev_err(adev->dev, "\t WALKER_ERROR: 0x%lx\n", 694 REG_GET_FIELD(status, 695 VM_L2_PROTECTION_FAULT_STATUS, WALKER_ERROR)); 696 dev_err(adev->dev, "\t PERMISSION_FAULTS: 0x%lx\n", 697 REG_GET_FIELD(status, 698 VM_L2_PROTECTION_FAULT_STATUS, PERMISSION_FAULTS)); 699 dev_err(adev->dev, "\t MAPPING_ERROR: 0x%lx\n", 700 REG_GET_FIELD(status, 701 VM_L2_PROTECTION_FAULT_STATUS, MAPPING_ERROR)); 702 dev_err(adev->dev, "\t RW: 0x%x\n", rw); 703 return 0; 704 } 705 706 static const struct amdgpu_irq_src_funcs gmc_v9_0_irq_funcs = { 707 .set = gmc_v9_0_vm_fault_interrupt_state, 708 .process = gmc_v9_0_process_interrupt, 709 }; 710 711 712 static const struct amdgpu_irq_src_funcs gmc_v9_0_ecc_funcs = { 713 .set = gmc_v9_0_ecc_interrupt_state, 714 .process = amdgpu_umc_process_ecc_irq, 715 }; 716 717 static void gmc_v9_0_set_irq_funcs(struct amdgpu_device *adev) 718 { 719 adev->gmc.vm_fault.num_types = 1; 720 adev->gmc.vm_fault.funcs = &gmc_v9_0_irq_funcs; 721 722 if (!amdgpu_sriov_vf(adev) && 723 !adev->gmc.xgmi.connected_to_cpu) { 724 adev->gmc.ecc_irq.num_types = 1; 725 adev->gmc.ecc_irq.funcs = &gmc_v9_0_ecc_funcs; 726 } 727 } 728 729 static uint32_t gmc_v9_0_get_invalidate_req(unsigned int vmid, 730 uint32_t flush_type) 731 { 732 u32 req = 0; 733 734 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, 735 PER_VMID_INVALIDATE_REQ, 1 << vmid); 736 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, FLUSH_TYPE, flush_type); 737 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PTES, 1); 738 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE0, 1); 739 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE1, 1); 740 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE2, 1); 741 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L1_PTES, 1); 742 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, 743 CLEAR_PROTECTION_FAULT_STATUS_ADDR, 0); 744 745 return req; 746 } 747 748 /** 749 * gmc_v9_0_use_invalidate_semaphore - judge whether to use semaphore 750 * 751 * @adev: amdgpu_device pointer 752 * @vmhub: vmhub type 753 * 754 */ 755 static bool gmc_v9_0_use_invalidate_semaphore(struct amdgpu_device *adev, 756 uint32_t vmhub) 757 { 758 if (adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 2) || 759 adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 3)) 760 return false; 761 762 return ((vmhub == AMDGPU_MMHUB_0 || 763 vmhub == AMDGPU_MMHUB_1) && 764 (!amdgpu_sriov_vf(adev)) && 765 (!(!(adev->apu_flags & AMD_APU_IS_RAVEN2) && 766 (adev->apu_flags & AMD_APU_IS_PICASSO)))); 767 } 768 769 static bool gmc_v9_0_get_atc_vmid_pasid_mapping_info(struct amdgpu_device *adev, 770 uint8_t vmid, uint16_t *p_pasid) 771 { 772 uint32_t value; 773 774 value = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING) 775 + vmid); 776 *p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK; 777 778 return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK); 779 } 780 781 /* 782 * GART 783 * VMID 0 is the physical GPU addresses as used by the kernel. 784 * VMIDs 1-15 are used for userspace clients and are handled 785 * by the amdgpu vm/hsa code. 786 */ 787 788 /** 789 * gmc_v9_0_flush_gpu_tlb - tlb flush with certain type 790 * 791 * @adev: amdgpu_device pointer 792 * @vmid: vm instance to flush 793 * @vmhub: which hub to flush 794 * @flush_type: the flush type 795 * 796 * Flush the TLB for the requested page table using certain type. 797 */ 798 static void gmc_v9_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid, 799 uint32_t vmhub, uint32_t flush_type) 800 { 801 bool use_semaphore = gmc_v9_0_use_invalidate_semaphore(adev, vmhub); 802 const unsigned eng = 17; 803 u32 j, inv_req, inv_req2, tmp; 804 struct amdgpu_vmhub *hub; 805 806 BUG_ON(vmhub >= adev->num_vmhubs); 807 808 hub = &adev->vmhub[vmhub]; 809 if (adev->gmc.xgmi.num_physical_nodes && 810 adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 0)) { 811 /* Vega20+XGMI caches PTEs in TC and TLB. Add a 812 * heavy-weight TLB flush (type 2), which flushes 813 * both. Due to a race condition with concurrent 814 * memory accesses using the same TLB cache line, we 815 * still need a second TLB flush after this. 816 */ 817 inv_req = gmc_v9_0_get_invalidate_req(vmid, 2); 818 inv_req2 = gmc_v9_0_get_invalidate_req(vmid, flush_type); 819 } else { 820 inv_req = gmc_v9_0_get_invalidate_req(vmid, flush_type); 821 inv_req2 = 0; 822 } 823 824 /* This is necessary for a HW workaround under SRIOV as well 825 * as GFXOFF under bare metal 826 */ 827 if (adev->gfx.kiq.ring.sched.ready && 828 (amdgpu_sriov_runtime(adev) || !amdgpu_sriov_vf(adev)) && 829 down_read_trylock(&adev->reset_domain->sem)) { 830 uint32_t req = hub->vm_inv_eng0_req + hub->eng_distance * eng; 831 uint32_t ack = hub->vm_inv_eng0_ack + hub->eng_distance * eng; 832 833 amdgpu_virt_kiq_reg_write_reg_wait(adev, req, ack, inv_req, 834 1 << vmid); 835 up_read(&adev->reset_domain->sem); 836 return; 837 } 838 839 spin_lock(&adev->gmc.invalidate_lock); 840 841 /* 842 * It may lose gpuvm invalidate acknowldege state across power-gating 843 * off cycle, add semaphore acquire before invalidation and semaphore 844 * release after invalidation to avoid entering power gated state 845 * to WA the Issue 846 */ 847 848 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 849 if (use_semaphore) { 850 for (j = 0; j < adev->usec_timeout; j++) { 851 /* a read return value of 1 means semaphore acquire */ 852 if (vmhub == AMDGPU_GFXHUB_0) 853 tmp = RREG32_SOC15_IP_NO_KIQ(GC, hub->vm_inv_eng0_sem + hub->eng_distance * eng); 854 else 855 tmp = RREG32_SOC15_IP_NO_KIQ(MMHUB, hub->vm_inv_eng0_sem + hub->eng_distance * eng); 856 857 if (tmp & 0x1) 858 break; 859 udelay(1); 860 } 861 862 if (j >= adev->usec_timeout) 863 DRM_ERROR("Timeout waiting for sem acquire in VM flush!\n"); 864 } 865 866 do { 867 if (vmhub == AMDGPU_GFXHUB_0) 868 WREG32_SOC15_IP_NO_KIQ(GC, hub->vm_inv_eng0_req + hub->eng_distance * eng, inv_req); 869 else 870 WREG32_SOC15_IP_NO_KIQ(MMHUB, hub->vm_inv_eng0_req + hub->eng_distance * eng, inv_req); 871 872 /* 873 * Issue a dummy read to wait for the ACK register to 874 * be cleared to avoid a false ACK due to the new fast 875 * GRBM interface. 876 */ 877 if ((vmhub == AMDGPU_GFXHUB_0) && 878 (adev->ip_versions[GC_HWIP][0] < IP_VERSION(9, 4, 2))) 879 RREG32_NO_KIQ(hub->vm_inv_eng0_req + 880 hub->eng_distance * eng); 881 882 for (j = 0; j < adev->usec_timeout; j++) { 883 if (vmhub == AMDGPU_GFXHUB_0) 884 tmp = RREG32_SOC15_IP_NO_KIQ(GC, hub->vm_inv_eng0_ack + hub->eng_distance * eng); 885 else 886 tmp = RREG32_SOC15_IP_NO_KIQ(MMHUB, hub->vm_inv_eng0_ack + hub->eng_distance * eng); 887 888 if (tmp & (1 << vmid)) 889 break; 890 udelay(1); 891 } 892 893 inv_req = inv_req2; 894 inv_req2 = 0; 895 } while (inv_req); 896 897 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 898 if (use_semaphore) { 899 /* 900 * add semaphore release after invalidation, 901 * write with 0 means semaphore release 902 */ 903 if (vmhub == AMDGPU_GFXHUB_0) 904 WREG32_SOC15_IP_NO_KIQ(GC, hub->vm_inv_eng0_sem + hub->eng_distance * eng, 0); 905 else 906 WREG32_SOC15_IP_NO_KIQ(MMHUB, hub->vm_inv_eng0_sem + hub->eng_distance * eng, 0); 907 } 908 909 spin_unlock(&adev->gmc.invalidate_lock); 910 911 if (j < adev->usec_timeout) 912 return; 913 914 DRM_ERROR("Timeout waiting for VM flush ACK!\n"); 915 } 916 917 /** 918 * gmc_v9_0_flush_gpu_tlb_pasid - tlb flush via pasid 919 * 920 * @adev: amdgpu_device pointer 921 * @pasid: pasid to be flush 922 * @flush_type: the flush type 923 * @all_hub: flush all hubs 924 * 925 * Flush the TLB for the requested pasid. 926 */ 927 static int gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev, 928 uint16_t pasid, uint32_t flush_type, 929 bool all_hub) 930 { 931 int vmid, i; 932 signed long r; 933 uint32_t seq; 934 uint16_t queried_pasid; 935 bool ret; 936 u32 usec_timeout = amdgpu_sriov_vf(adev) ? SRIOV_USEC_TIMEOUT : adev->usec_timeout; 937 struct amdgpu_ring *ring = &adev->gfx.kiq.ring; 938 struct amdgpu_kiq *kiq = &adev->gfx.kiq; 939 940 if (amdgpu_in_reset(adev)) 941 return -EIO; 942 943 if (ring->sched.ready && down_read_trylock(&adev->reset_domain->sem)) { 944 /* Vega20+XGMI caches PTEs in TC and TLB. Add a 945 * heavy-weight TLB flush (type 2), which flushes 946 * both. Due to a race condition with concurrent 947 * memory accesses using the same TLB cache line, we 948 * still need a second TLB flush after this. 949 */ 950 bool vega20_xgmi_wa = (adev->gmc.xgmi.num_physical_nodes && 951 adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 0)); 952 /* 2 dwords flush + 8 dwords fence */ 953 unsigned int ndw = kiq->pmf->invalidate_tlbs_size + 8; 954 955 if (vega20_xgmi_wa) 956 ndw += kiq->pmf->invalidate_tlbs_size; 957 958 spin_lock(&adev->gfx.kiq.ring_lock); 959 /* 2 dwords flush + 8 dwords fence */ 960 amdgpu_ring_alloc(ring, ndw); 961 if (vega20_xgmi_wa) 962 kiq->pmf->kiq_invalidate_tlbs(ring, 963 pasid, 2, all_hub); 964 kiq->pmf->kiq_invalidate_tlbs(ring, 965 pasid, flush_type, all_hub); 966 r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT); 967 if (r) { 968 amdgpu_ring_undo(ring); 969 spin_unlock(&adev->gfx.kiq.ring_lock); 970 up_read(&adev->reset_domain->sem); 971 return -ETIME; 972 } 973 974 amdgpu_ring_commit(ring); 975 spin_unlock(&adev->gfx.kiq.ring_lock); 976 r = amdgpu_fence_wait_polling(ring, seq, usec_timeout); 977 if (r < 1) { 978 dev_err(adev->dev, "wait for kiq fence error: %ld.\n", r); 979 up_read(&adev->reset_domain->sem); 980 return -ETIME; 981 } 982 up_read(&adev->reset_domain->sem); 983 return 0; 984 } 985 986 for (vmid = 1; vmid < 16; vmid++) { 987 988 ret = gmc_v9_0_get_atc_vmid_pasid_mapping_info(adev, vmid, 989 &queried_pasid); 990 if (ret && queried_pasid == pasid) { 991 if (all_hub) { 992 for (i = 0; i < adev->num_vmhubs; i++) 993 gmc_v9_0_flush_gpu_tlb(adev, vmid, 994 i, flush_type); 995 } else { 996 gmc_v9_0_flush_gpu_tlb(adev, vmid, 997 AMDGPU_GFXHUB_0, flush_type); 998 } 999 break; 1000 } 1001 } 1002 1003 return 0; 1004 1005 } 1006 1007 static uint64_t gmc_v9_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring, 1008 unsigned vmid, uint64_t pd_addr) 1009 { 1010 bool use_semaphore = gmc_v9_0_use_invalidate_semaphore(ring->adev, ring->vm_hub); 1011 struct amdgpu_device *adev = ring->adev; 1012 struct amdgpu_vmhub *hub = &adev->vmhub[ring->vm_hub]; 1013 uint32_t req = gmc_v9_0_get_invalidate_req(vmid, 0); 1014 unsigned eng = ring->vm_inv_eng; 1015 1016 /* 1017 * It may lose gpuvm invalidate acknowldege state across power-gating 1018 * off cycle, add semaphore acquire before invalidation and semaphore 1019 * release after invalidation to avoid entering power gated state 1020 * to WA the Issue 1021 */ 1022 1023 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 1024 if (use_semaphore) 1025 /* a read return value of 1 means semaphore acuqire */ 1026 amdgpu_ring_emit_reg_wait(ring, 1027 hub->vm_inv_eng0_sem + 1028 hub->eng_distance * eng, 0x1, 0x1); 1029 1030 amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_lo32 + 1031 (hub->ctx_addr_distance * vmid), 1032 lower_32_bits(pd_addr)); 1033 1034 amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 + 1035 (hub->ctx_addr_distance * vmid), 1036 upper_32_bits(pd_addr)); 1037 1038 amdgpu_ring_emit_reg_write_reg_wait(ring, hub->vm_inv_eng0_req + 1039 hub->eng_distance * eng, 1040 hub->vm_inv_eng0_ack + 1041 hub->eng_distance * eng, 1042 req, 1 << vmid); 1043 1044 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 1045 if (use_semaphore) 1046 /* 1047 * add semaphore release after invalidation, 1048 * write with 0 means semaphore release 1049 */ 1050 amdgpu_ring_emit_wreg(ring, hub->vm_inv_eng0_sem + 1051 hub->eng_distance * eng, 0); 1052 1053 return pd_addr; 1054 } 1055 1056 static void gmc_v9_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned vmid, 1057 unsigned pasid) 1058 { 1059 struct amdgpu_device *adev = ring->adev; 1060 uint32_t reg; 1061 1062 /* Do nothing because there's no lut register for mmhub1. */ 1063 if (ring->vm_hub == AMDGPU_MMHUB_1) 1064 return; 1065 1066 if (ring->vm_hub == AMDGPU_GFXHUB_0) 1067 reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid; 1068 else 1069 reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT_MM) + vmid; 1070 1071 amdgpu_ring_emit_wreg(ring, reg, pasid); 1072 } 1073 1074 /* 1075 * PTE format on VEGA 10: 1076 * 63:59 reserved 1077 * 58:57 mtype 1078 * 56 F 1079 * 55 L 1080 * 54 P 1081 * 53 SW 1082 * 52 T 1083 * 50:48 reserved 1084 * 47:12 4k physical page base address 1085 * 11:7 fragment 1086 * 6 write 1087 * 5 read 1088 * 4 exe 1089 * 3 Z 1090 * 2 snooped 1091 * 1 system 1092 * 0 valid 1093 * 1094 * PDE format on VEGA 10: 1095 * 63:59 block fragment size 1096 * 58:55 reserved 1097 * 54 P 1098 * 53:48 reserved 1099 * 47:6 physical base address of PD or PTE 1100 * 5:3 reserved 1101 * 2 C 1102 * 1 system 1103 * 0 valid 1104 */ 1105 1106 static uint64_t gmc_v9_0_map_mtype(struct amdgpu_device *adev, uint32_t flags) 1107 1108 { 1109 switch (flags) { 1110 case AMDGPU_VM_MTYPE_DEFAULT: 1111 return AMDGPU_PTE_MTYPE_VG10(MTYPE_NC); 1112 case AMDGPU_VM_MTYPE_NC: 1113 return AMDGPU_PTE_MTYPE_VG10(MTYPE_NC); 1114 case AMDGPU_VM_MTYPE_WC: 1115 return AMDGPU_PTE_MTYPE_VG10(MTYPE_WC); 1116 case AMDGPU_VM_MTYPE_RW: 1117 return AMDGPU_PTE_MTYPE_VG10(MTYPE_RW); 1118 case AMDGPU_VM_MTYPE_CC: 1119 return AMDGPU_PTE_MTYPE_VG10(MTYPE_CC); 1120 case AMDGPU_VM_MTYPE_UC: 1121 return AMDGPU_PTE_MTYPE_VG10(MTYPE_UC); 1122 default: 1123 return AMDGPU_PTE_MTYPE_VG10(MTYPE_NC); 1124 } 1125 } 1126 1127 static void gmc_v9_0_get_vm_pde(struct amdgpu_device *adev, int level, 1128 uint64_t *addr, uint64_t *flags) 1129 { 1130 if (!(*flags & AMDGPU_PDE_PTE) && !(*flags & AMDGPU_PTE_SYSTEM)) 1131 *addr = amdgpu_gmc_vram_mc2pa(adev, *addr); 1132 BUG_ON(*addr & 0xFFFF00000000003FULL); 1133 1134 if (!adev->gmc.translate_further) 1135 return; 1136 1137 if (level == AMDGPU_VM_PDB1) { 1138 /* Set the block fragment size */ 1139 if (!(*flags & AMDGPU_PDE_PTE)) 1140 *flags |= AMDGPU_PDE_BFS(0x9); 1141 1142 } else if (level == AMDGPU_VM_PDB0) { 1143 if (*flags & AMDGPU_PDE_PTE) { 1144 *flags &= ~AMDGPU_PDE_PTE; 1145 if (!(*flags & AMDGPU_PTE_VALID)) 1146 *addr |= 1 << PAGE_SHIFT; 1147 } else { 1148 *flags |= AMDGPU_PTE_TF; 1149 } 1150 } 1151 } 1152 1153 static void gmc_v9_0_get_coherence_flags(struct amdgpu_device *adev, 1154 struct amdgpu_bo *bo, 1155 struct amdgpu_bo_va_mapping *mapping, 1156 uint64_t *flags) 1157 { 1158 struct amdgpu_device *bo_adev = amdgpu_ttm_adev(bo->tbo.bdev); 1159 bool is_vram = bo->tbo.resource->mem_type == TTM_PL_VRAM; 1160 bool coherent = bo->flags & AMDGPU_GEM_CREATE_COHERENT; 1161 bool uncached = bo->flags & AMDGPU_GEM_CREATE_UNCACHED; 1162 unsigned int mtype; 1163 bool snoop = false; 1164 1165 switch (adev->ip_versions[GC_HWIP][0]) { 1166 case IP_VERSION(9, 4, 1): 1167 case IP_VERSION(9, 4, 2): 1168 case IP_VERSION(9, 4, 3): 1169 if (is_vram) { 1170 if (bo_adev == adev) { 1171 if (uncached) 1172 mtype = MTYPE_UC; 1173 else if (coherent) 1174 mtype = MTYPE_CC; 1175 else 1176 mtype = MTYPE_RW; 1177 /* FIXME: is this still needed? Or does 1178 * amdgpu_ttm_tt_pde_flags already handle this? 1179 */ 1180 if ((adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 2) || 1181 adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 3)) && 1182 adev->gmc.xgmi.connected_to_cpu) 1183 snoop = true; 1184 } else { 1185 if (uncached || coherent) 1186 mtype = MTYPE_UC; 1187 else 1188 mtype = MTYPE_NC; 1189 if (mapping->bo_va->is_xgmi) 1190 snoop = true; 1191 } 1192 } else { 1193 if (uncached || coherent) 1194 mtype = MTYPE_UC; 1195 else 1196 mtype = MTYPE_NC; 1197 /* FIXME: is this still needed? Or does 1198 * amdgpu_ttm_tt_pde_flags already handle this? 1199 */ 1200 snoop = true; 1201 } 1202 break; 1203 default: 1204 if (uncached || coherent) 1205 mtype = MTYPE_UC; 1206 else 1207 mtype = MTYPE_NC; 1208 1209 /* FIXME: is this still needed? Or does 1210 * amdgpu_ttm_tt_pde_flags already handle this? 1211 */ 1212 if (!is_vram) 1213 snoop = true; 1214 } 1215 1216 if (mtype != MTYPE_NC) 1217 *flags = (*flags & ~AMDGPU_PTE_MTYPE_VG10_MASK) | 1218 AMDGPU_PTE_MTYPE_VG10(mtype); 1219 *flags |= snoop ? AMDGPU_PTE_SNOOPED : 0; 1220 } 1221 1222 static void gmc_v9_0_get_vm_pte(struct amdgpu_device *adev, 1223 struct amdgpu_bo_va_mapping *mapping, 1224 uint64_t *flags) 1225 { 1226 struct amdgpu_bo *bo = mapping->bo_va->base.bo; 1227 1228 *flags &= ~AMDGPU_PTE_EXECUTABLE; 1229 *flags |= mapping->flags & AMDGPU_PTE_EXECUTABLE; 1230 1231 *flags &= ~AMDGPU_PTE_MTYPE_VG10_MASK; 1232 *flags |= mapping->flags & AMDGPU_PTE_MTYPE_VG10_MASK; 1233 1234 if (mapping->flags & AMDGPU_PTE_PRT) { 1235 *flags |= AMDGPU_PTE_PRT; 1236 *flags &= ~AMDGPU_PTE_VALID; 1237 } 1238 1239 if (bo && bo->tbo.resource) 1240 gmc_v9_0_get_coherence_flags(adev, mapping->bo_va->base.bo, 1241 mapping, flags); 1242 } 1243 1244 static unsigned gmc_v9_0_get_vbios_fb_size(struct amdgpu_device *adev) 1245 { 1246 u32 d1vga_control = RREG32_SOC15(DCE, 0, mmD1VGA_CONTROL); 1247 unsigned size; 1248 1249 /* TODO move to DC so GMC doesn't need to hard-code DCN registers */ 1250 1251 if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) { 1252 size = AMDGPU_VBIOS_VGA_ALLOCATION; 1253 } else { 1254 u32 viewport; 1255 1256 switch (adev->ip_versions[DCE_HWIP][0]) { 1257 case IP_VERSION(1, 0, 0): 1258 case IP_VERSION(1, 0, 1): 1259 viewport = RREG32_SOC15(DCE, 0, mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION); 1260 size = (REG_GET_FIELD(viewport, 1261 HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_HEIGHT) * 1262 REG_GET_FIELD(viewport, 1263 HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_WIDTH) * 1264 4); 1265 break; 1266 case IP_VERSION(2, 1, 0): 1267 viewport = RREG32_SOC15(DCE, 0, mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION_DCN2); 1268 size = (REG_GET_FIELD(viewport, 1269 HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_HEIGHT) * 1270 REG_GET_FIELD(viewport, 1271 HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_WIDTH) * 1272 4); 1273 break; 1274 default: 1275 viewport = RREG32_SOC15(DCE, 0, mmSCL0_VIEWPORT_SIZE); 1276 size = (REG_GET_FIELD(viewport, SCL0_VIEWPORT_SIZE, VIEWPORT_HEIGHT) * 1277 REG_GET_FIELD(viewport, SCL0_VIEWPORT_SIZE, VIEWPORT_WIDTH) * 1278 4); 1279 break; 1280 } 1281 } 1282 1283 return size; 1284 } 1285 1286 static const struct amdgpu_gmc_funcs gmc_v9_0_gmc_funcs = { 1287 .flush_gpu_tlb = gmc_v9_0_flush_gpu_tlb, 1288 .flush_gpu_tlb_pasid = gmc_v9_0_flush_gpu_tlb_pasid, 1289 .emit_flush_gpu_tlb = gmc_v9_0_emit_flush_gpu_tlb, 1290 .emit_pasid_mapping = gmc_v9_0_emit_pasid_mapping, 1291 .map_mtype = gmc_v9_0_map_mtype, 1292 .get_vm_pde = gmc_v9_0_get_vm_pde, 1293 .get_vm_pte = gmc_v9_0_get_vm_pte, 1294 .get_vbios_fb_size = gmc_v9_0_get_vbios_fb_size, 1295 }; 1296 1297 static void gmc_v9_0_set_gmc_funcs(struct amdgpu_device *adev) 1298 { 1299 adev->gmc.gmc_funcs = &gmc_v9_0_gmc_funcs; 1300 } 1301 1302 static void gmc_v9_0_set_umc_funcs(struct amdgpu_device *adev) 1303 { 1304 switch (adev->ip_versions[UMC_HWIP][0]) { 1305 case IP_VERSION(6, 0, 0): 1306 adev->umc.funcs = &umc_v6_0_funcs; 1307 break; 1308 case IP_VERSION(6, 1, 1): 1309 adev->umc.max_ras_err_cnt_per_query = UMC_V6_1_TOTAL_CHANNEL_NUM; 1310 adev->umc.channel_inst_num = UMC_V6_1_CHANNEL_INSTANCE_NUM; 1311 adev->umc.umc_inst_num = UMC_V6_1_UMC_INSTANCE_NUM; 1312 adev->umc.channel_offs = UMC_V6_1_PER_CHANNEL_OFFSET_VG20; 1313 adev->umc.retire_unit = 1; 1314 adev->umc.channel_idx_tbl = &umc_v6_1_channel_idx_tbl[0][0]; 1315 adev->umc.ras = &umc_v6_1_ras; 1316 break; 1317 case IP_VERSION(6, 1, 2): 1318 adev->umc.max_ras_err_cnt_per_query = UMC_V6_1_TOTAL_CHANNEL_NUM; 1319 adev->umc.channel_inst_num = UMC_V6_1_CHANNEL_INSTANCE_NUM; 1320 adev->umc.umc_inst_num = UMC_V6_1_UMC_INSTANCE_NUM; 1321 adev->umc.channel_offs = UMC_V6_1_PER_CHANNEL_OFFSET_ARCT; 1322 adev->umc.retire_unit = 1; 1323 adev->umc.channel_idx_tbl = &umc_v6_1_channel_idx_tbl[0][0]; 1324 adev->umc.ras = &umc_v6_1_ras; 1325 break; 1326 case IP_VERSION(6, 7, 0): 1327 adev->umc.max_ras_err_cnt_per_query = 1328 UMC_V6_7_TOTAL_CHANNEL_NUM * UMC_V6_7_BAD_PAGE_NUM_PER_CHANNEL; 1329 adev->umc.channel_inst_num = UMC_V6_7_CHANNEL_INSTANCE_NUM; 1330 adev->umc.umc_inst_num = UMC_V6_7_UMC_INSTANCE_NUM; 1331 adev->umc.channel_offs = UMC_V6_7_PER_CHANNEL_OFFSET; 1332 adev->umc.retire_unit = (UMC_V6_7_NA_MAP_PA_NUM * 2); 1333 if (!adev->gmc.xgmi.connected_to_cpu) 1334 adev->umc.ras = &umc_v6_7_ras; 1335 if (1 & adev->smuio.funcs->get_die_id(adev)) 1336 adev->umc.channel_idx_tbl = &umc_v6_7_channel_idx_tbl_first[0][0]; 1337 else 1338 adev->umc.channel_idx_tbl = &umc_v6_7_channel_idx_tbl_second[0][0]; 1339 break; 1340 default: 1341 break; 1342 } 1343 } 1344 1345 static void gmc_v9_0_set_mmhub_funcs(struct amdgpu_device *adev) 1346 { 1347 switch (adev->ip_versions[MMHUB_HWIP][0]) { 1348 case IP_VERSION(9, 4, 1): 1349 adev->mmhub.funcs = &mmhub_v9_4_funcs; 1350 break; 1351 case IP_VERSION(9, 4, 2): 1352 adev->mmhub.funcs = &mmhub_v1_7_funcs; 1353 break; 1354 case IP_VERSION(1, 8, 0): 1355 adev->mmhub.funcs = &mmhub_v1_8_funcs; 1356 break; 1357 default: 1358 adev->mmhub.funcs = &mmhub_v1_0_funcs; 1359 break; 1360 } 1361 } 1362 1363 static void gmc_v9_0_set_mmhub_ras_funcs(struct amdgpu_device *adev) 1364 { 1365 switch (adev->ip_versions[MMHUB_HWIP][0]) { 1366 case IP_VERSION(9, 4, 0): 1367 adev->mmhub.ras = &mmhub_v1_0_ras; 1368 break; 1369 case IP_VERSION(9, 4, 1): 1370 adev->mmhub.ras = &mmhub_v9_4_ras; 1371 break; 1372 case IP_VERSION(9, 4, 2): 1373 adev->mmhub.ras = &mmhub_v1_7_ras; 1374 break; 1375 default: 1376 /* mmhub ras is not available */ 1377 break; 1378 } 1379 } 1380 1381 static void gmc_v9_0_set_gfxhub_funcs(struct amdgpu_device *adev) 1382 { 1383 if (adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 3)) 1384 adev->gfxhub.funcs = &gfxhub_v1_2_funcs; 1385 else 1386 adev->gfxhub.funcs = &gfxhub_v1_0_funcs; 1387 } 1388 1389 static void gmc_v9_0_set_hdp_ras_funcs(struct amdgpu_device *adev) 1390 { 1391 adev->hdp.ras = &hdp_v4_0_ras; 1392 } 1393 1394 static void gmc_v9_0_set_mca_ras_funcs(struct amdgpu_device *adev) 1395 { 1396 struct amdgpu_mca *mca = &adev->mca; 1397 1398 /* is UMC the right IP to check for MCA? Maybe DF? */ 1399 switch (adev->ip_versions[UMC_HWIP][0]) { 1400 case IP_VERSION(6, 7, 0): 1401 if (!adev->gmc.xgmi.connected_to_cpu) { 1402 mca->mp0.ras = &mca_v3_0_mp0_ras; 1403 mca->mp1.ras = &mca_v3_0_mp1_ras; 1404 mca->mpio.ras = &mca_v3_0_mpio_ras; 1405 } 1406 break; 1407 default: 1408 break; 1409 } 1410 } 1411 1412 static void gmc_v9_0_set_xgmi_ras_funcs(struct amdgpu_device *adev) 1413 { 1414 if (!adev->gmc.xgmi.connected_to_cpu) 1415 adev->gmc.xgmi.ras = &xgmi_ras; 1416 } 1417 1418 static int gmc_v9_0_early_init(void *handle) 1419 { 1420 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1421 1422 /* ARCT and VEGA20 don't have XGMI defined in their IP discovery tables */ 1423 if (adev->asic_type == CHIP_VEGA20 || 1424 adev->asic_type == CHIP_ARCTURUS) 1425 adev->gmc.xgmi.supported = true; 1426 1427 if (adev->ip_versions[XGMI_HWIP][0] == IP_VERSION(6, 1, 0)) { 1428 adev->gmc.xgmi.supported = true; 1429 adev->gmc.xgmi.connected_to_cpu = 1430 adev->smuio.funcs->is_host_gpu_xgmi_supported(adev); 1431 } 1432 1433 gmc_v9_0_set_gmc_funcs(adev); 1434 gmc_v9_0_set_irq_funcs(adev); 1435 gmc_v9_0_set_umc_funcs(adev); 1436 gmc_v9_0_set_mmhub_funcs(adev); 1437 gmc_v9_0_set_mmhub_ras_funcs(adev); 1438 gmc_v9_0_set_gfxhub_funcs(adev); 1439 gmc_v9_0_set_hdp_ras_funcs(adev); 1440 gmc_v9_0_set_mca_ras_funcs(adev); 1441 gmc_v9_0_set_xgmi_ras_funcs(adev); 1442 1443 adev->gmc.shared_aperture_start = 0x2000000000000000ULL; 1444 adev->gmc.shared_aperture_end = 1445 adev->gmc.shared_aperture_start + (4ULL << 30) - 1; 1446 adev->gmc.private_aperture_start = 0x1000000000000000ULL; 1447 adev->gmc.private_aperture_end = 1448 adev->gmc.private_aperture_start + (4ULL << 30) - 1; 1449 1450 return 0; 1451 } 1452 1453 static int gmc_v9_0_late_init(void *handle) 1454 { 1455 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1456 int r; 1457 1458 r = amdgpu_gmc_allocate_vm_inv_eng(adev); 1459 if (r) 1460 return r; 1461 1462 /* 1463 * Workaround performance drop issue with VBIOS enables partial 1464 * writes, while disables HBM ECC for vega10. 1465 */ 1466 if (!amdgpu_sriov_vf(adev) && 1467 (adev->ip_versions[UMC_HWIP][0] == IP_VERSION(6, 0, 0))) { 1468 if (!(adev->ras_enabled & (1 << AMDGPU_RAS_BLOCK__UMC))) { 1469 if (adev->df.funcs && 1470 adev->df.funcs->enable_ecc_force_par_wr_rmw) 1471 adev->df.funcs->enable_ecc_force_par_wr_rmw(adev, false); 1472 } 1473 } 1474 1475 if (!amdgpu_persistent_edc_harvesting_supported(adev)) { 1476 if (adev->mmhub.ras && adev->mmhub.ras->ras_block.hw_ops && 1477 adev->mmhub.ras->ras_block.hw_ops->reset_ras_error_count) 1478 adev->mmhub.ras->ras_block.hw_ops->reset_ras_error_count(adev); 1479 1480 if (adev->hdp.ras && adev->hdp.ras->ras_block.hw_ops && 1481 adev->hdp.ras->ras_block.hw_ops->reset_ras_error_count) 1482 adev->hdp.ras->ras_block.hw_ops->reset_ras_error_count(adev); 1483 } 1484 1485 r = amdgpu_gmc_ras_late_init(adev); 1486 if (r) 1487 return r; 1488 1489 return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0); 1490 } 1491 1492 static void gmc_v9_0_vram_gtt_location(struct amdgpu_device *adev, 1493 struct amdgpu_gmc *mc) 1494 { 1495 u64 base = adev->mmhub.funcs->get_fb_location(adev); 1496 1497 /* add the xgmi offset of the physical node */ 1498 base += adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size; 1499 if (adev->gmc.xgmi.connected_to_cpu) { 1500 amdgpu_gmc_sysvm_location(adev, mc); 1501 } else { 1502 amdgpu_gmc_vram_location(adev, mc, base); 1503 amdgpu_gmc_gart_location(adev, mc); 1504 amdgpu_gmc_agp_location(adev, mc); 1505 } 1506 /* base offset of vram pages */ 1507 adev->vm_manager.vram_base_offset = adev->gfxhub.funcs->get_mc_fb_offset(adev); 1508 1509 /* XXX: add the xgmi offset of the physical node? */ 1510 adev->vm_manager.vram_base_offset += 1511 adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size; 1512 } 1513 1514 /** 1515 * gmc_v9_0_mc_init - initialize the memory controller driver params 1516 * 1517 * @adev: amdgpu_device pointer 1518 * 1519 * Look up the amount of vram, vram width, and decide how to place 1520 * vram and gart within the GPU's physical address space. 1521 * Returns 0 for success. 1522 */ 1523 static int gmc_v9_0_mc_init(struct amdgpu_device *adev) 1524 { 1525 int r; 1526 1527 /* size in MB on si */ 1528 adev->gmc.mc_vram_size = 1529 adev->nbio.funcs->get_memsize(adev) * 1024ULL * 1024ULL; 1530 adev->gmc.real_vram_size = adev->gmc.mc_vram_size; 1531 1532 if (!(adev->flags & AMD_IS_APU) && 1533 !adev->gmc.xgmi.connected_to_cpu) { 1534 r = amdgpu_device_resize_fb_bar(adev); 1535 if (r) 1536 return r; 1537 } 1538 adev->gmc.aper_base = pci_resource_start(adev->pdev, 0); 1539 adev->gmc.aper_size = pci_resource_len(adev->pdev, 0); 1540 1541 #ifdef CONFIG_X86_64 1542 /* 1543 * AMD Accelerated Processing Platform (APP) supporting GPU-HOST xgmi 1544 * interface can use VRAM through here as it appears system reserved 1545 * memory in host address space. 1546 * 1547 * For APUs, VRAM is just the stolen system memory and can be accessed 1548 * directly. 1549 * 1550 * Otherwise, use the legacy Host Data Path (HDP) through PCIe BAR. 1551 */ 1552 1553 /* check whether both host-gpu and gpu-gpu xgmi links exist */ 1554 if (((adev->flags & AMD_IS_APU) && !amdgpu_passthrough(adev)) || 1555 (adev->gmc.xgmi.supported && 1556 adev->gmc.xgmi.connected_to_cpu)) { 1557 adev->gmc.aper_base = 1558 adev->gfxhub.funcs->get_mc_fb_offset(adev) + 1559 adev->gmc.xgmi.physical_node_id * 1560 adev->gmc.xgmi.node_segment_size; 1561 adev->gmc.aper_size = adev->gmc.real_vram_size; 1562 } 1563 1564 #endif 1565 adev->gmc.visible_vram_size = adev->gmc.aper_size; 1566 1567 /* set the gart size */ 1568 if (amdgpu_gart_size == -1) { 1569 switch (adev->ip_versions[GC_HWIP][0]) { 1570 case IP_VERSION(9, 0, 1): /* all engines support GPUVM */ 1571 case IP_VERSION(9, 2, 1): /* all engines support GPUVM */ 1572 case IP_VERSION(9, 4, 0): 1573 case IP_VERSION(9, 4, 1): 1574 case IP_VERSION(9, 4, 2): 1575 case IP_VERSION(9, 4, 3): 1576 default: 1577 adev->gmc.gart_size = 512ULL << 20; 1578 break; 1579 case IP_VERSION(9, 1, 0): /* DCE SG support */ 1580 case IP_VERSION(9, 2, 2): /* DCE SG support */ 1581 case IP_VERSION(9, 3, 0): 1582 adev->gmc.gart_size = 1024ULL << 20; 1583 break; 1584 } 1585 } else { 1586 adev->gmc.gart_size = (u64)amdgpu_gart_size << 20; 1587 } 1588 1589 adev->gmc.gart_size += adev->pm.smu_prv_buffer_size; 1590 1591 gmc_v9_0_vram_gtt_location(adev, &adev->gmc); 1592 1593 return 0; 1594 } 1595 1596 static int gmc_v9_0_gart_init(struct amdgpu_device *adev) 1597 { 1598 int r; 1599 1600 if (adev->gart.bo) { 1601 WARN(1, "VEGA10 PCIE GART already initialized\n"); 1602 return 0; 1603 } 1604 1605 if (adev->gmc.xgmi.connected_to_cpu) { 1606 adev->gmc.vmid0_page_table_depth = 1; 1607 adev->gmc.vmid0_page_table_block_size = 12; 1608 } else { 1609 adev->gmc.vmid0_page_table_depth = 0; 1610 adev->gmc.vmid0_page_table_block_size = 0; 1611 } 1612 1613 /* Initialize common gart structure */ 1614 r = amdgpu_gart_init(adev); 1615 if (r) 1616 return r; 1617 adev->gart.table_size = adev->gart.num_gpu_pages * 8; 1618 adev->gart.gart_pte_flags = AMDGPU_PTE_MTYPE_VG10(MTYPE_UC) | 1619 AMDGPU_PTE_EXECUTABLE; 1620 1621 r = amdgpu_gart_table_vram_alloc(adev); 1622 if (r) 1623 return r; 1624 1625 if (adev->gmc.xgmi.connected_to_cpu) { 1626 r = amdgpu_gmc_pdb0_alloc(adev); 1627 } 1628 1629 return r; 1630 } 1631 1632 /** 1633 * gmc_v9_0_save_registers - saves regs 1634 * 1635 * @adev: amdgpu_device pointer 1636 * 1637 * This saves potential register values that should be 1638 * restored upon resume 1639 */ 1640 static void gmc_v9_0_save_registers(struct amdgpu_device *adev) 1641 { 1642 if ((adev->ip_versions[DCE_HWIP][0] == IP_VERSION(1, 0, 0)) || 1643 (adev->ip_versions[DCE_HWIP][0] == IP_VERSION(1, 0, 1))) 1644 adev->gmc.sdpif_register = RREG32_SOC15(DCE, 0, mmDCHUBBUB_SDPIF_MMIO_CNTRL_0); 1645 } 1646 1647 static int gmc_v9_0_sw_init(void *handle) 1648 { 1649 int r, vram_width = 0, vram_type = 0, vram_vendor = 0, dma_addr_bits; 1650 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1651 1652 adev->gfxhub.funcs->init(adev); 1653 1654 adev->mmhub.funcs->init(adev); 1655 1656 spin_lock_init(&adev->gmc.invalidate_lock); 1657 1658 r = amdgpu_atomfirmware_get_vram_info(adev, 1659 &vram_width, &vram_type, &vram_vendor); 1660 if (amdgpu_sriov_vf(adev)) 1661 /* For Vega10 SR-IOV, vram_width can't be read from ATOM as RAVEN, 1662 * and DF related registers is not readable, seems hardcord is the 1663 * only way to set the correct vram_width 1664 */ 1665 adev->gmc.vram_width = 2048; 1666 else if (amdgpu_emu_mode != 1) 1667 adev->gmc.vram_width = vram_width; 1668 1669 if (!adev->gmc.vram_width) { 1670 int chansize, numchan; 1671 1672 /* hbm memory channel size */ 1673 if (adev->flags & AMD_IS_APU) 1674 chansize = 64; 1675 else 1676 chansize = 128; 1677 if (adev->df.funcs && 1678 adev->df.funcs->get_hbm_channel_number) { 1679 numchan = adev->df.funcs->get_hbm_channel_number(adev); 1680 adev->gmc.vram_width = numchan * chansize; 1681 } 1682 } 1683 1684 adev->gmc.vram_type = vram_type; 1685 adev->gmc.vram_vendor = vram_vendor; 1686 switch (adev->ip_versions[GC_HWIP][0]) { 1687 case IP_VERSION(9, 1, 0): 1688 case IP_VERSION(9, 2, 2): 1689 adev->num_vmhubs = 2; 1690 1691 if (adev->rev_id == 0x0 || adev->rev_id == 0x1) { 1692 amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48); 1693 } else { 1694 /* vm_size is 128TB + 512GB for legacy 3-level page support */ 1695 amdgpu_vm_adjust_size(adev, 128 * 1024 + 512, 9, 2, 48); 1696 adev->gmc.translate_further = 1697 adev->vm_manager.num_level > 1; 1698 } 1699 break; 1700 case IP_VERSION(9, 0, 1): 1701 case IP_VERSION(9, 2, 1): 1702 case IP_VERSION(9, 4, 0): 1703 case IP_VERSION(9, 3, 0): 1704 case IP_VERSION(9, 4, 2): 1705 case IP_VERSION(9, 4, 3): 1706 adev->num_vmhubs = 2; 1707 1708 1709 /* 1710 * To fulfill 4-level page support, 1711 * vm size is 256TB (48bit), maximum size of Vega10, 1712 * block size 512 (9bit) 1713 */ 1714 /* sriov restrict max_pfn below AMDGPU_GMC_HOLE */ 1715 if (amdgpu_sriov_vf(adev)) 1716 amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 47); 1717 else 1718 amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48); 1719 if (adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 2)) 1720 adev->gmc.translate_further = adev->vm_manager.num_level > 1; 1721 break; 1722 case IP_VERSION(9, 4, 1): 1723 adev->num_vmhubs = 3; 1724 1725 /* Keep the vm size same with Vega20 */ 1726 amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48); 1727 adev->gmc.translate_further = adev->vm_manager.num_level > 1; 1728 break; 1729 default: 1730 break; 1731 } 1732 1733 /* This interrupt is VMC page fault.*/ 1734 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC, VMC_1_0__SRCID__VM_FAULT, 1735 &adev->gmc.vm_fault); 1736 if (r) 1737 return r; 1738 1739 if (adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 1)) { 1740 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC1, VMC_1_0__SRCID__VM_FAULT, 1741 &adev->gmc.vm_fault); 1742 if (r) 1743 return r; 1744 } 1745 1746 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_UTCL2, UTCL2_1_0__SRCID__FAULT, 1747 &adev->gmc.vm_fault); 1748 1749 if (r) 1750 return r; 1751 1752 if (!amdgpu_sriov_vf(adev) && 1753 !adev->gmc.xgmi.connected_to_cpu) { 1754 /* interrupt sent to DF. */ 1755 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_DF, 0, 1756 &adev->gmc.ecc_irq); 1757 if (r) 1758 return r; 1759 } 1760 1761 /* Set the internal MC address mask 1762 * This is the max address of the GPU's 1763 * internal address space. 1764 */ 1765 adev->gmc.mc_mask = 0xffffffffffffULL; /* 48 bit MC */ 1766 1767 dma_addr_bits = adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 2) ? 48:44; 1768 r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(dma_addr_bits)); 1769 if (r) { 1770 printk(KERN_WARNING "amdgpu: No suitable DMA available.\n"); 1771 return r; 1772 } 1773 adev->need_swiotlb = drm_need_swiotlb(dma_addr_bits); 1774 1775 r = gmc_v9_0_mc_init(adev); 1776 if (r) 1777 return r; 1778 1779 amdgpu_gmc_get_vbios_allocations(adev); 1780 1781 /* Memory manager */ 1782 r = amdgpu_bo_init(adev); 1783 if (r) 1784 return r; 1785 1786 r = gmc_v9_0_gart_init(adev); 1787 if (r) 1788 return r; 1789 1790 /* 1791 * number of VMs 1792 * VMID 0 is reserved for System 1793 * amdgpu graphics/compute will use VMIDs 1..n-1 1794 * amdkfd will use VMIDs n..15 1795 * 1796 * The first KFD VMID is 8 for GPUs with graphics, 3 for 1797 * compute-only GPUs. On compute-only GPUs that leaves 2 VMIDs 1798 * for video processing. 1799 */ 1800 adev->vm_manager.first_kfd_vmid = 1801 (adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 1) || 1802 adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 2) || 1803 adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 3)) ? 3 : 8; 1804 1805 amdgpu_vm_manager_init(adev); 1806 1807 gmc_v9_0_save_registers(adev); 1808 1809 r = amdgpu_gmc_ras_sw_init(adev); 1810 if (r) 1811 return r; 1812 1813 return 0; 1814 } 1815 1816 static int gmc_v9_0_sw_fini(void *handle) 1817 { 1818 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1819 1820 amdgpu_gmc_ras_fini(adev); 1821 amdgpu_gem_force_release(adev); 1822 amdgpu_vm_manager_fini(adev); 1823 amdgpu_gart_table_vram_free(adev); 1824 amdgpu_bo_free_kernel(&adev->gmc.pdb0_bo, NULL, &adev->gmc.ptr_pdb0); 1825 amdgpu_bo_fini(adev); 1826 1827 return 0; 1828 } 1829 1830 static void gmc_v9_0_init_golden_registers(struct amdgpu_device *adev) 1831 { 1832 1833 switch (adev->ip_versions[MMHUB_HWIP][0]) { 1834 case IP_VERSION(9, 0, 0): 1835 if (amdgpu_sriov_vf(adev)) 1836 break; 1837 fallthrough; 1838 case IP_VERSION(9, 4, 0): 1839 soc15_program_register_sequence(adev, 1840 golden_settings_mmhub_1_0_0, 1841 ARRAY_SIZE(golden_settings_mmhub_1_0_0)); 1842 soc15_program_register_sequence(adev, 1843 golden_settings_athub_1_0_0, 1844 ARRAY_SIZE(golden_settings_athub_1_0_0)); 1845 break; 1846 case IP_VERSION(9, 1, 0): 1847 case IP_VERSION(9, 2, 0): 1848 /* TODO for renoir */ 1849 soc15_program_register_sequence(adev, 1850 golden_settings_athub_1_0_0, 1851 ARRAY_SIZE(golden_settings_athub_1_0_0)); 1852 break; 1853 default: 1854 break; 1855 } 1856 } 1857 1858 /** 1859 * gmc_v9_0_restore_registers - restores regs 1860 * 1861 * @adev: amdgpu_device pointer 1862 * 1863 * This restores register values, saved at suspend. 1864 */ 1865 void gmc_v9_0_restore_registers(struct amdgpu_device *adev) 1866 { 1867 if ((adev->ip_versions[DCE_HWIP][0] == IP_VERSION(1, 0, 0)) || 1868 (adev->ip_versions[DCE_HWIP][0] == IP_VERSION(1, 0, 1))) { 1869 WREG32_SOC15(DCE, 0, mmDCHUBBUB_SDPIF_MMIO_CNTRL_0, adev->gmc.sdpif_register); 1870 WARN_ON(adev->gmc.sdpif_register != 1871 RREG32_SOC15(DCE, 0, mmDCHUBBUB_SDPIF_MMIO_CNTRL_0)); 1872 } 1873 } 1874 1875 /** 1876 * gmc_v9_0_gart_enable - gart enable 1877 * 1878 * @adev: amdgpu_device pointer 1879 */ 1880 static int gmc_v9_0_gart_enable(struct amdgpu_device *adev) 1881 { 1882 int r; 1883 1884 if (adev->gmc.xgmi.connected_to_cpu) 1885 amdgpu_gmc_init_pdb0(adev); 1886 1887 if (adev->gart.bo == NULL) { 1888 dev_err(adev->dev, "No VRAM object for PCIE GART.\n"); 1889 return -EINVAL; 1890 } 1891 1892 amdgpu_gtt_mgr_recover(&adev->mman.gtt_mgr); 1893 1894 if (!adev->in_s0ix) { 1895 r = adev->gfxhub.funcs->gart_enable(adev); 1896 if (r) 1897 return r; 1898 } 1899 1900 r = adev->mmhub.funcs->gart_enable(adev); 1901 if (r) 1902 return r; 1903 1904 DRM_INFO("PCIE GART of %uM enabled.\n", 1905 (unsigned)(adev->gmc.gart_size >> 20)); 1906 if (adev->gmc.pdb0_bo) 1907 DRM_INFO("PDB0 located at 0x%016llX\n", 1908 (unsigned long long)amdgpu_bo_gpu_offset(adev->gmc.pdb0_bo)); 1909 DRM_INFO("PTB located at 0x%016llX\n", 1910 (unsigned long long)amdgpu_bo_gpu_offset(adev->gart.bo)); 1911 1912 return 0; 1913 } 1914 1915 static int gmc_v9_0_hw_init(void *handle) 1916 { 1917 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1918 bool value; 1919 int i, r; 1920 1921 /* The sequence of these two function calls matters.*/ 1922 gmc_v9_0_init_golden_registers(adev); 1923 1924 if (adev->mode_info.num_crtc) { 1925 /* Lockout access through VGA aperture*/ 1926 WREG32_FIELD15(DCE, 0, VGA_HDP_CONTROL, VGA_MEMORY_DISABLE, 1); 1927 /* disable VGA render */ 1928 WREG32_FIELD15(DCE, 0, VGA_RENDER_CONTROL, VGA_VSTATUS_CNTL, 0); 1929 } 1930 1931 if (adev->mmhub.funcs->update_power_gating) 1932 adev->mmhub.funcs->update_power_gating(adev, true); 1933 1934 adev->hdp.funcs->init_registers(adev); 1935 1936 /* After HDP is initialized, flush HDP.*/ 1937 adev->hdp.funcs->flush_hdp(adev, NULL); 1938 1939 if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS) 1940 value = false; 1941 else 1942 value = true; 1943 1944 if (!amdgpu_sriov_vf(adev)) { 1945 if (!adev->in_s0ix) 1946 adev->gfxhub.funcs->set_fault_enable_default(adev, value); 1947 adev->mmhub.funcs->set_fault_enable_default(adev, value); 1948 } 1949 for (i = 0; i < adev->num_vmhubs; ++i) { 1950 if (adev->in_s0ix && (i == AMDGPU_GFXHUB_0)) 1951 continue; 1952 gmc_v9_0_flush_gpu_tlb(adev, 0, i, 0); 1953 } 1954 1955 if (adev->umc.funcs && adev->umc.funcs->init_registers) 1956 adev->umc.funcs->init_registers(adev); 1957 1958 r = gmc_v9_0_gart_enable(adev); 1959 if (r) 1960 return r; 1961 1962 if (amdgpu_emu_mode == 1) 1963 return amdgpu_gmc_vram_checking(adev); 1964 else 1965 return r; 1966 } 1967 1968 /** 1969 * gmc_v9_0_gart_disable - gart disable 1970 * 1971 * @adev: amdgpu_device pointer 1972 * 1973 * This disables all VM page table. 1974 */ 1975 static void gmc_v9_0_gart_disable(struct amdgpu_device *adev) 1976 { 1977 if (!adev->in_s0ix) 1978 adev->gfxhub.funcs->gart_disable(adev); 1979 adev->mmhub.funcs->gart_disable(adev); 1980 } 1981 1982 static int gmc_v9_0_hw_fini(void *handle) 1983 { 1984 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1985 1986 gmc_v9_0_gart_disable(adev); 1987 1988 if (amdgpu_sriov_vf(adev)) { 1989 /* full access mode, so don't touch any GMC register */ 1990 DRM_DEBUG("For SRIOV client, shouldn't do anything.\n"); 1991 return 0; 1992 } 1993 1994 /* 1995 * Pair the operations did in gmc_v9_0_hw_init and thus maintain 1996 * a correct cached state for GMC. Otherwise, the "gate" again 1997 * operation on S3 resuming will fail due to wrong cached state. 1998 */ 1999 if (adev->mmhub.funcs->update_power_gating) 2000 adev->mmhub.funcs->update_power_gating(adev, false); 2001 2002 amdgpu_irq_put(adev, &adev->gmc.ecc_irq, 0); 2003 amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0); 2004 2005 return 0; 2006 } 2007 2008 static int gmc_v9_0_suspend(void *handle) 2009 { 2010 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 2011 2012 return gmc_v9_0_hw_fini(adev); 2013 } 2014 2015 static int gmc_v9_0_resume(void *handle) 2016 { 2017 int r; 2018 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 2019 2020 r = gmc_v9_0_hw_init(adev); 2021 if (r) 2022 return r; 2023 2024 amdgpu_vmid_reset_all(adev); 2025 2026 return 0; 2027 } 2028 2029 static bool gmc_v9_0_is_idle(void *handle) 2030 { 2031 /* MC is always ready in GMC v9.*/ 2032 return true; 2033 } 2034 2035 static int gmc_v9_0_wait_for_idle(void *handle) 2036 { 2037 /* There is no need to wait for MC idle in GMC v9.*/ 2038 return 0; 2039 } 2040 2041 static int gmc_v9_0_soft_reset(void *handle) 2042 { 2043 /* XXX for emulation.*/ 2044 return 0; 2045 } 2046 2047 static int gmc_v9_0_set_clockgating_state(void *handle, 2048 enum amd_clockgating_state state) 2049 { 2050 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 2051 2052 adev->mmhub.funcs->set_clockgating(adev, state); 2053 2054 athub_v1_0_set_clockgating(adev, state); 2055 2056 return 0; 2057 } 2058 2059 static void gmc_v9_0_get_clockgating_state(void *handle, u64 *flags) 2060 { 2061 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 2062 2063 adev->mmhub.funcs->get_clockgating(adev, flags); 2064 2065 athub_v1_0_get_clockgating(adev, flags); 2066 } 2067 2068 static int gmc_v9_0_set_powergating_state(void *handle, 2069 enum amd_powergating_state state) 2070 { 2071 return 0; 2072 } 2073 2074 const struct amd_ip_funcs gmc_v9_0_ip_funcs = { 2075 .name = "gmc_v9_0", 2076 .early_init = gmc_v9_0_early_init, 2077 .late_init = gmc_v9_0_late_init, 2078 .sw_init = gmc_v9_0_sw_init, 2079 .sw_fini = gmc_v9_0_sw_fini, 2080 .hw_init = gmc_v9_0_hw_init, 2081 .hw_fini = gmc_v9_0_hw_fini, 2082 .suspend = gmc_v9_0_suspend, 2083 .resume = gmc_v9_0_resume, 2084 .is_idle = gmc_v9_0_is_idle, 2085 .wait_for_idle = gmc_v9_0_wait_for_idle, 2086 .soft_reset = gmc_v9_0_soft_reset, 2087 .set_clockgating_state = gmc_v9_0_set_clockgating_state, 2088 .set_powergating_state = gmc_v9_0_set_powergating_state, 2089 .get_clockgating_state = gmc_v9_0_get_clockgating_state, 2090 }; 2091 2092 const struct amdgpu_ip_block_version gmc_v9_0_ip_block = 2093 { 2094 .type = AMD_IP_BLOCK_TYPE_GMC, 2095 .major = 9, 2096 .minor = 0, 2097 .rev = 0, 2098 .funcs = &gmc_v9_0_ip_funcs, 2099 }; 2100