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 "umc_v12_0.h" 60 #include "hdp_v4_0.h" 61 #include "mca_v3_0.h" 62 63 #include "ivsrcid/vmc/irqsrcs_vmc_1_0.h" 64 65 #include "amdgpu_ras.h" 66 #include "amdgpu_xgmi.h" 67 68 /* add these here since we already include dce12 headers and these are for DCN */ 69 #define mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION 0x055d 70 #define mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION_BASE_IDX 2 71 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_WIDTH__SHIFT 0x0 72 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_HEIGHT__SHIFT 0x10 73 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_WIDTH_MASK 0x00003FFFL 74 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_HEIGHT_MASK 0x3FFF0000L 75 #define mmDCHUBBUB_SDPIF_MMIO_CNTRL_0 0x049d 76 #define mmDCHUBBUB_SDPIF_MMIO_CNTRL_0_BASE_IDX 2 77 78 #define mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION_DCN2 0x05ea 79 #define mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION_DCN2_BASE_IDX 2 80 81 #define MAX_MEM_RANGES 8 82 83 static const char * const 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 SOC15_REG_GOLDEN_VALUE(MMHUB, 0, mmDAGB1_WRCLI2, 0x00000007, 0xfe5fe0fa), 336 SOC15_REG_GOLDEN_VALUE(MMHUB, 0, mmMMEA1_DRAM_WR_CLI2GRP_MAP0, 0x00000030, 0x55555565) 337 }; 338 339 static const struct soc15_reg_golden golden_settings_athub_1_0_0[] = { 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 int 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 */ 424 if (adev->asic_type >= CHIP_VEGA20) 425 return 0; 426 427 bits = 0x7f; 428 429 switch (state) { 430 case AMDGPU_IRQ_STATE_DISABLE: 431 for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_addrs); i++) { 432 reg = ecc_umc_mcumc_ctrl_addrs[i]; 433 tmp = RREG32(reg); 434 tmp &= ~bits; 435 WREG32(reg, tmp); 436 } 437 for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_mask_addrs); i++) { 438 reg = ecc_umc_mcumc_ctrl_mask_addrs[i]; 439 tmp = RREG32(reg); 440 tmp &= ~bits; 441 WREG32(reg, tmp); 442 } 443 break; 444 case AMDGPU_IRQ_STATE_ENABLE: 445 for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_addrs); i++) { 446 reg = ecc_umc_mcumc_ctrl_addrs[i]; 447 tmp = RREG32(reg); 448 tmp |= bits; 449 WREG32(reg, tmp); 450 } 451 for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_mask_addrs); i++) { 452 reg = ecc_umc_mcumc_ctrl_mask_addrs[i]; 453 tmp = RREG32(reg); 454 tmp |= bits; 455 WREG32(reg, tmp); 456 } 457 break; 458 default: 459 break; 460 } 461 462 return 0; 463 } 464 465 static int gmc_v9_0_vm_fault_interrupt_state(struct amdgpu_device *adev, 466 struct amdgpu_irq_src *src, 467 unsigned int type, 468 enum amdgpu_interrupt_state state) 469 { 470 struct amdgpu_vmhub *hub; 471 u32 tmp, reg, bits, i, j; 472 473 bits = VM_CONTEXT1_CNTL__RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 474 VM_CONTEXT1_CNTL__DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 475 VM_CONTEXT1_CNTL__PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 476 VM_CONTEXT1_CNTL__VALID_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 477 VM_CONTEXT1_CNTL__READ_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 478 VM_CONTEXT1_CNTL__WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 479 VM_CONTEXT1_CNTL__EXECUTE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK; 480 481 switch (state) { 482 case AMDGPU_IRQ_STATE_DISABLE: 483 for_each_set_bit(j, adev->vmhubs_mask, AMDGPU_MAX_VMHUBS) { 484 hub = &adev->vmhub[j]; 485 for (i = 0; i < 16; i++) { 486 reg = hub->vm_context0_cntl + i; 487 488 /* This works because this interrupt is only 489 * enabled at init/resume and disabled in 490 * fini/suspend, so the overall state doesn't 491 * change over the course of suspend/resume. 492 */ 493 if (adev->in_s0ix && (j == AMDGPU_GFXHUB(0))) 494 continue; 495 496 if (j >= AMDGPU_MMHUB0(0)) 497 tmp = RREG32_SOC15_IP(MMHUB, reg); 498 else 499 tmp = RREG32_SOC15_IP(GC, reg); 500 501 tmp &= ~bits; 502 503 if (j >= AMDGPU_MMHUB0(0)) 504 WREG32_SOC15_IP(MMHUB, reg, tmp); 505 else 506 WREG32_SOC15_IP(GC, reg, tmp); 507 } 508 } 509 break; 510 case AMDGPU_IRQ_STATE_ENABLE: 511 for_each_set_bit(j, adev->vmhubs_mask, AMDGPU_MAX_VMHUBS) { 512 hub = &adev->vmhub[j]; 513 for (i = 0; i < 16; i++) { 514 reg = hub->vm_context0_cntl + i; 515 516 /* This works because this interrupt is only 517 * enabled at init/resume and disabled in 518 * fini/suspend, so the overall state doesn't 519 * change over the course of suspend/resume. 520 */ 521 if (adev->in_s0ix && (j == AMDGPU_GFXHUB(0))) 522 continue; 523 524 if (j >= AMDGPU_MMHUB0(0)) 525 tmp = RREG32_SOC15_IP(MMHUB, reg); 526 else 527 tmp = RREG32_SOC15_IP(GC, reg); 528 529 tmp |= bits; 530 531 if (j >= AMDGPU_MMHUB0(0)) 532 WREG32_SOC15_IP(MMHUB, reg, tmp); 533 else 534 WREG32_SOC15_IP(GC, reg, tmp); 535 } 536 } 537 break; 538 default: 539 break; 540 } 541 542 return 0; 543 } 544 545 static int gmc_v9_0_process_interrupt(struct amdgpu_device *adev, 546 struct amdgpu_irq_src *source, 547 struct amdgpu_iv_entry *entry) 548 { 549 bool retry_fault = !!(entry->src_data[1] & 0x80); 550 bool write_fault = !!(entry->src_data[1] & 0x20); 551 uint32_t status = 0, cid = 0, rw = 0; 552 struct amdgpu_task_info task_info; 553 struct amdgpu_vmhub *hub; 554 const char *mmhub_cid; 555 const char *hub_name; 556 unsigned int vmhub; 557 u64 addr; 558 uint32_t cam_index = 0; 559 int ret, xcc_id = 0; 560 uint32_t node_id; 561 562 node_id = entry->node_id; 563 564 addr = (u64)entry->src_data[0] << 12; 565 addr |= ((u64)entry->src_data[1] & 0xf) << 44; 566 567 if (entry->client_id == SOC15_IH_CLIENTID_VMC) { 568 hub_name = "mmhub0"; 569 vmhub = AMDGPU_MMHUB0(node_id / 4); 570 } else if (entry->client_id == SOC15_IH_CLIENTID_VMC1) { 571 hub_name = "mmhub1"; 572 vmhub = AMDGPU_MMHUB1(0); 573 } else { 574 hub_name = "gfxhub0"; 575 if (adev->gfx.funcs->ih_node_to_logical_xcc) { 576 xcc_id = adev->gfx.funcs->ih_node_to_logical_xcc(adev, 577 node_id); 578 if (xcc_id < 0) 579 xcc_id = 0; 580 } 581 vmhub = xcc_id; 582 } 583 hub = &adev->vmhub[vmhub]; 584 585 if (retry_fault) { 586 if (adev->irq.retry_cam_enabled) { 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 cam_index = entry->src_data[2] & 0x3ff; 596 597 ret = amdgpu_vm_handle_fault(adev, entry->pasid, entry->vmid, node_id, 598 addr, write_fault); 599 WDOORBELL32(adev->irq.retry_cam_doorbell_index, cam_index); 600 if (ret) 601 return 1; 602 } else { 603 /* Process it onyl if it's the first fault for this address */ 604 if (entry->ih != &adev->irq.ih_soft && 605 amdgpu_gmc_filter_faults(adev, entry->ih, addr, entry->pasid, 606 entry->timestamp)) 607 return 1; 608 609 /* Delegate it to a different ring if the hardware hasn't 610 * already done it. 611 */ 612 if (entry->ih == &adev->irq.ih) { 613 amdgpu_irq_delegate(adev, entry, 8); 614 return 1; 615 } 616 617 /* Try to handle the recoverable page faults by filling page 618 * tables 619 */ 620 if (amdgpu_vm_handle_fault(adev, entry->pasid, entry->vmid, node_id, 621 addr, write_fault)) 622 return 1; 623 } 624 } 625 626 if (!printk_ratelimit()) 627 return 0; 628 629 memset(&task_info, 0, sizeof(struct amdgpu_task_info)); 630 amdgpu_vm_get_task_info(adev, entry->pasid, &task_info); 631 632 dev_err(adev->dev, 633 "[%s] %s page fault (src_id:%u ring:%u vmid:%u pasid:%u, for process %s pid %d thread %s pid %d)\n", 634 hub_name, retry_fault ? "retry" : "no-retry", 635 entry->src_id, entry->ring_id, entry->vmid, 636 entry->pasid, task_info.process_name, task_info.tgid, 637 task_info.task_name, task_info.pid); 638 dev_err(adev->dev, " in page starting at address 0x%016llx from IH client 0x%x (%s)\n", 639 addr, entry->client_id, 640 soc15_ih_clientid_name[entry->client_id]); 641 642 if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3)) 643 dev_err(adev->dev, " cookie node_id %d fault from die %s%d%s\n", 644 node_id, node_id % 4 == 3 ? "RSV" : "AID", node_id / 4, 645 node_id % 4 == 1 ? ".XCD0" : node_id % 4 == 2 ? ".XCD1" : ""); 646 647 if (amdgpu_sriov_vf(adev)) 648 return 0; 649 650 /* 651 * Issue a dummy read to wait for the status register to 652 * be updated to avoid reading an incorrect value due to 653 * the new fast GRBM interface. 654 */ 655 if ((entry->vmid_src == AMDGPU_GFXHUB(0)) && 656 (amdgpu_ip_version(adev, GC_HWIP, 0) < IP_VERSION(9, 4, 2))) 657 RREG32(hub->vm_l2_pro_fault_status); 658 659 status = RREG32(hub->vm_l2_pro_fault_status); 660 cid = REG_GET_FIELD(status, VM_L2_PROTECTION_FAULT_STATUS, CID); 661 rw = REG_GET_FIELD(status, VM_L2_PROTECTION_FAULT_STATUS, RW); 662 WREG32_P(hub->vm_l2_pro_fault_cntl, 1, ~1); 663 664 amdgpu_vm_update_fault_cache(adev, entry->pasid, addr, status, vmhub); 665 666 dev_err(adev->dev, 667 "VM_L2_PROTECTION_FAULT_STATUS:0x%08X\n", 668 status); 669 if (entry->vmid_src == AMDGPU_GFXHUB(0)) { 670 dev_err(adev->dev, "\t Faulty UTCL2 client ID: %s (0x%x)\n", 671 cid >= ARRAY_SIZE(gfxhub_client_ids) ? "unknown" : 672 gfxhub_client_ids[cid], 673 cid); 674 } else { 675 switch (amdgpu_ip_version(adev, MMHUB_HWIP, 0)) { 676 case IP_VERSION(9, 0, 0): 677 mmhub_cid = mmhub_client_ids_vega10[cid][rw]; 678 break; 679 case IP_VERSION(9, 3, 0): 680 mmhub_cid = mmhub_client_ids_vega12[cid][rw]; 681 break; 682 case IP_VERSION(9, 4, 0): 683 mmhub_cid = mmhub_client_ids_vega20[cid][rw]; 684 break; 685 case IP_VERSION(9, 4, 1): 686 mmhub_cid = mmhub_client_ids_arcturus[cid][rw]; 687 break; 688 case IP_VERSION(9, 1, 0): 689 case IP_VERSION(9, 2, 0): 690 mmhub_cid = mmhub_client_ids_raven[cid][rw]; 691 break; 692 case IP_VERSION(1, 5, 0): 693 case IP_VERSION(2, 4, 0): 694 mmhub_cid = mmhub_client_ids_renoir[cid][rw]; 695 break; 696 case IP_VERSION(1, 8, 0): 697 case IP_VERSION(9, 4, 2): 698 mmhub_cid = mmhub_client_ids_aldebaran[cid][rw]; 699 break; 700 default: 701 mmhub_cid = NULL; 702 break; 703 } 704 dev_err(adev->dev, "\t Faulty UTCL2 client ID: %s (0x%x)\n", 705 mmhub_cid ? mmhub_cid : "unknown", cid); 706 } 707 dev_err(adev->dev, "\t MORE_FAULTS: 0x%lx\n", 708 REG_GET_FIELD(status, 709 VM_L2_PROTECTION_FAULT_STATUS, MORE_FAULTS)); 710 dev_err(adev->dev, "\t WALKER_ERROR: 0x%lx\n", 711 REG_GET_FIELD(status, 712 VM_L2_PROTECTION_FAULT_STATUS, WALKER_ERROR)); 713 dev_err(adev->dev, "\t PERMISSION_FAULTS: 0x%lx\n", 714 REG_GET_FIELD(status, 715 VM_L2_PROTECTION_FAULT_STATUS, PERMISSION_FAULTS)); 716 dev_err(adev->dev, "\t MAPPING_ERROR: 0x%lx\n", 717 REG_GET_FIELD(status, 718 VM_L2_PROTECTION_FAULT_STATUS, MAPPING_ERROR)); 719 dev_err(adev->dev, "\t RW: 0x%x\n", rw); 720 return 0; 721 } 722 723 static const struct amdgpu_irq_src_funcs gmc_v9_0_irq_funcs = { 724 .set = gmc_v9_0_vm_fault_interrupt_state, 725 .process = gmc_v9_0_process_interrupt, 726 }; 727 728 729 static const struct amdgpu_irq_src_funcs gmc_v9_0_ecc_funcs = { 730 .set = gmc_v9_0_ecc_interrupt_state, 731 .process = amdgpu_umc_process_ecc_irq, 732 }; 733 734 static void gmc_v9_0_set_irq_funcs(struct amdgpu_device *adev) 735 { 736 adev->gmc.vm_fault.num_types = 1; 737 adev->gmc.vm_fault.funcs = &gmc_v9_0_irq_funcs; 738 739 if (!amdgpu_sriov_vf(adev) && 740 !adev->gmc.xgmi.connected_to_cpu && 741 !adev->gmc.is_app_apu) { 742 adev->gmc.ecc_irq.num_types = 1; 743 adev->gmc.ecc_irq.funcs = &gmc_v9_0_ecc_funcs; 744 } 745 } 746 747 static uint32_t gmc_v9_0_get_invalidate_req(unsigned int vmid, 748 uint32_t flush_type) 749 { 750 u32 req = 0; 751 752 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, 753 PER_VMID_INVALIDATE_REQ, 1 << vmid); 754 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, FLUSH_TYPE, flush_type); 755 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PTES, 1); 756 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE0, 1); 757 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE1, 1); 758 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE2, 1); 759 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L1_PTES, 1); 760 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, 761 CLEAR_PROTECTION_FAULT_STATUS_ADDR, 0); 762 763 return req; 764 } 765 766 /** 767 * gmc_v9_0_use_invalidate_semaphore - judge whether to use semaphore 768 * 769 * @adev: amdgpu_device pointer 770 * @vmhub: vmhub type 771 * 772 */ 773 static bool gmc_v9_0_use_invalidate_semaphore(struct amdgpu_device *adev, 774 uint32_t vmhub) 775 { 776 if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 2) || 777 amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3)) 778 return false; 779 780 return ((vmhub == AMDGPU_MMHUB0(0) || 781 vmhub == AMDGPU_MMHUB1(0)) && 782 (!amdgpu_sriov_vf(adev)) && 783 (!(!(adev->apu_flags & AMD_APU_IS_RAVEN2) && 784 (adev->apu_flags & AMD_APU_IS_PICASSO)))); 785 } 786 787 static bool gmc_v9_0_get_atc_vmid_pasid_mapping_info(struct amdgpu_device *adev, 788 uint8_t vmid, uint16_t *p_pasid) 789 { 790 uint32_t value; 791 792 value = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING) 793 + vmid); 794 *p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK; 795 796 return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK); 797 } 798 799 /* 800 * GART 801 * VMID 0 is the physical GPU addresses as used by the kernel. 802 * VMIDs 1-15 are used for userspace clients and are handled 803 * by the amdgpu vm/hsa code. 804 */ 805 806 /** 807 * gmc_v9_0_flush_gpu_tlb - tlb flush with certain type 808 * 809 * @adev: amdgpu_device pointer 810 * @vmid: vm instance to flush 811 * @vmhub: which hub to flush 812 * @flush_type: the flush type 813 * 814 * Flush the TLB for the requested page table using certain type. 815 */ 816 static void gmc_v9_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid, 817 uint32_t vmhub, uint32_t flush_type) 818 { 819 bool use_semaphore = gmc_v9_0_use_invalidate_semaphore(adev, vmhub); 820 u32 j, inv_req, tmp, sem, req, ack, inst; 821 const unsigned int eng = 17; 822 struct amdgpu_vmhub *hub; 823 824 BUG_ON(vmhub >= AMDGPU_MAX_VMHUBS); 825 826 hub = &adev->vmhub[vmhub]; 827 inv_req = gmc_v9_0_get_invalidate_req(vmid, flush_type); 828 sem = hub->vm_inv_eng0_sem + hub->eng_distance * eng; 829 req = hub->vm_inv_eng0_req + hub->eng_distance * eng; 830 ack = hub->vm_inv_eng0_ack + hub->eng_distance * eng; 831 832 /* This is necessary for a HW workaround under SRIOV as well 833 * as GFXOFF under bare metal 834 */ 835 if (vmhub >= AMDGPU_MMHUB0(0)) 836 inst = GET_INST(GC, 0); 837 else 838 inst = vmhub; 839 if (adev->gfx.kiq[inst].ring.sched.ready && 840 (amdgpu_sriov_runtime(adev) || !amdgpu_sriov_vf(adev))) { 841 uint32_t req = hub->vm_inv_eng0_req + hub->eng_distance * eng; 842 uint32_t ack = hub->vm_inv_eng0_ack + hub->eng_distance * eng; 843 844 amdgpu_virt_kiq_reg_write_reg_wait(adev, req, ack, inv_req, 845 1 << vmid, inst); 846 return; 847 } 848 849 spin_lock(&adev->gmc.invalidate_lock); 850 851 /* 852 * It may lose gpuvm invalidate acknowldege state across power-gating 853 * off cycle, add semaphore acquire before invalidation and semaphore 854 * release after invalidation to avoid entering power gated state 855 * to WA the Issue 856 */ 857 858 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 859 if (use_semaphore) { 860 for (j = 0; j < adev->usec_timeout; j++) { 861 /* a read return value of 1 means semaphore acquire */ 862 if (vmhub >= AMDGPU_MMHUB0(0)) 863 tmp = RREG32_SOC15_IP_NO_KIQ(MMHUB, sem, inst); 864 else 865 tmp = RREG32_SOC15_IP_NO_KIQ(GC, sem, inst); 866 if (tmp & 0x1) 867 break; 868 udelay(1); 869 } 870 871 if (j >= adev->usec_timeout) 872 DRM_ERROR("Timeout waiting for sem acquire in VM flush!\n"); 873 } 874 875 if (vmhub >= AMDGPU_MMHUB0(0)) 876 WREG32_SOC15_IP_NO_KIQ(MMHUB, req, inv_req, inst); 877 else 878 WREG32_SOC15_IP_NO_KIQ(GC, req, inv_req, inst); 879 880 /* 881 * Issue a dummy read to wait for the ACK register to 882 * be cleared to avoid a false ACK due to the new fast 883 * GRBM interface. 884 */ 885 if ((vmhub == AMDGPU_GFXHUB(0)) && 886 (adev->ip_versions[GC_HWIP][0] < IP_VERSION(9, 4, 2))) 887 RREG32_NO_KIQ(req); 888 889 for (j = 0; j < adev->usec_timeout; j++) { 890 if (vmhub >= AMDGPU_MMHUB0(0)) 891 tmp = RREG32_SOC15_IP_NO_KIQ(MMHUB, ack, inst); 892 else 893 tmp = RREG32_SOC15_IP_NO_KIQ(GC, ack, inst); 894 if (tmp & (1 << vmid)) 895 break; 896 udelay(1); 897 } 898 899 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 900 if (use_semaphore) { 901 /* 902 * add semaphore release after invalidation, 903 * write with 0 means semaphore release 904 */ 905 if (vmhub >= AMDGPU_MMHUB0(0)) 906 WREG32_SOC15_IP_NO_KIQ(MMHUB, sem, 0, inst); 907 else 908 WREG32_SOC15_IP_NO_KIQ(GC, sem, 0, inst); 909 } 910 911 spin_unlock(&adev->gmc.invalidate_lock); 912 913 if (j < adev->usec_timeout) 914 return; 915 916 DRM_ERROR("Timeout waiting for VM flush ACK!\n"); 917 } 918 919 /** 920 * gmc_v9_0_flush_gpu_tlb_pasid - tlb flush via pasid 921 * 922 * @adev: amdgpu_device pointer 923 * @pasid: pasid to be flush 924 * @flush_type: the flush type 925 * @all_hub: flush all hubs 926 * @inst: is used to select which instance of KIQ to use for the invalidation 927 * 928 * Flush the TLB for the requested pasid. 929 */ 930 static void gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev, 931 uint16_t pasid, uint32_t flush_type, 932 bool all_hub, uint32_t inst) 933 { 934 uint16_t queried; 935 int i, vmid; 936 937 for (vmid = 1; vmid < 16; vmid++) { 938 bool valid; 939 940 valid = gmc_v9_0_get_atc_vmid_pasid_mapping_info(adev, vmid, 941 &queried); 942 if (!valid || queried != pasid) 943 continue; 944 945 if (all_hub) { 946 for_each_set_bit(i, adev->vmhubs_mask, 947 AMDGPU_MAX_VMHUBS) 948 gmc_v9_0_flush_gpu_tlb(adev, vmid, i, 949 flush_type); 950 } else { 951 gmc_v9_0_flush_gpu_tlb(adev, vmid, 952 AMDGPU_GFXHUB(0), 953 flush_type); 954 } 955 } 956 } 957 958 static uint64_t gmc_v9_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring, 959 unsigned int vmid, uint64_t pd_addr) 960 { 961 bool use_semaphore = gmc_v9_0_use_invalidate_semaphore(ring->adev, ring->vm_hub); 962 struct amdgpu_device *adev = ring->adev; 963 struct amdgpu_vmhub *hub = &adev->vmhub[ring->vm_hub]; 964 uint32_t req = gmc_v9_0_get_invalidate_req(vmid, 0); 965 unsigned int eng = ring->vm_inv_eng; 966 967 /* 968 * It may lose gpuvm invalidate acknowldege state across power-gating 969 * off cycle, add semaphore acquire before invalidation and semaphore 970 * release after invalidation to avoid entering power gated state 971 * to WA the Issue 972 */ 973 974 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 975 if (use_semaphore) 976 /* a read return value of 1 means semaphore acuqire */ 977 amdgpu_ring_emit_reg_wait(ring, 978 hub->vm_inv_eng0_sem + 979 hub->eng_distance * eng, 0x1, 0x1); 980 981 amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_lo32 + 982 (hub->ctx_addr_distance * vmid), 983 lower_32_bits(pd_addr)); 984 985 amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 + 986 (hub->ctx_addr_distance * vmid), 987 upper_32_bits(pd_addr)); 988 989 amdgpu_ring_emit_reg_write_reg_wait(ring, hub->vm_inv_eng0_req + 990 hub->eng_distance * eng, 991 hub->vm_inv_eng0_ack + 992 hub->eng_distance * eng, 993 req, 1 << vmid); 994 995 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 996 if (use_semaphore) 997 /* 998 * add semaphore release after invalidation, 999 * write with 0 means semaphore release 1000 */ 1001 amdgpu_ring_emit_wreg(ring, hub->vm_inv_eng0_sem + 1002 hub->eng_distance * eng, 0); 1003 1004 return pd_addr; 1005 } 1006 1007 static void gmc_v9_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned int vmid, 1008 unsigned int pasid) 1009 { 1010 struct amdgpu_device *adev = ring->adev; 1011 uint32_t reg; 1012 1013 /* Do nothing because there's no lut register for mmhub1. */ 1014 if (ring->vm_hub == AMDGPU_MMHUB1(0)) 1015 return; 1016 1017 if (ring->vm_hub == AMDGPU_GFXHUB(0)) 1018 reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid; 1019 else 1020 reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT_MM) + vmid; 1021 1022 amdgpu_ring_emit_wreg(ring, reg, pasid); 1023 } 1024 1025 /* 1026 * PTE format on VEGA 10: 1027 * 63:59 reserved 1028 * 58:57 mtype 1029 * 56 F 1030 * 55 L 1031 * 54 P 1032 * 53 SW 1033 * 52 T 1034 * 50:48 reserved 1035 * 47:12 4k physical page base address 1036 * 11:7 fragment 1037 * 6 write 1038 * 5 read 1039 * 4 exe 1040 * 3 Z 1041 * 2 snooped 1042 * 1 system 1043 * 0 valid 1044 * 1045 * PDE format on VEGA 10: 1046 * 63:59 block fragment size 1047 * 58:55 reserved 1048 * 54 P 1049 * 53:48 reserved 1050 * 47:6 physical base address of PD or PTE 1051 * 5:3 reserved 1052 * 2 C 1053 * 1 system 1054 * 0 valid 1055 */ 1056 1057 static uint64_t gmc_v9_0_map_mtype(struct amdgpu_device *adev, uint32_t flags) 1058 1059 { 1060 switch (flags) { 1061 case AMDGPU_VM_MTYPE_DEFAULT: 1062 return AMDGPU_PTE_MTYPE_VG10(MTYPE_NC); 1063 case AMDGPU_VM_MTYPE_NC: 1064 return AMDGPU_PTE_MTYPE_VG10(MTYPE_NC); 1065 case AMDGPU_VM_MTYPE_WC: 1066 return AMDGPU_PTE_MTYPE_VG10(MTYPE_WC); 1067 case AMDGPU_VM_MTYPE_RW: 1068 return AMDGPU_PTE_MTYPE_VG10(MTYPE_RW); 1069 case AMDGPU_VM_MTYPE_CC: 1070 return AMDGPU_PTE_MTYPE_VG10(MTYPE_CC); 1071 case AMDGPU_VM_MTYPE_UC: 1072 return AMDGPU_PTE_MTYPE_VG10(MTYPE_UC); 1073 default: 1074 return AMDGPU_PTE_MTYPE_VG10(MTYPE_NC); 1075 } 1076 } 1077 1078 static void gmc_v9_0_get_vm_pde(struct amdgpu_device *adev, int level, 1079 uint64_t *addr, uint64_t *flags) 1080 { 1081 if (!(*flags & AMDGPU_PDE_PTE) && !(*flags & AMDGPU_PTE_SYSTEM)) 1082 *addr = amdgpu_gmc_vram_mc2pa(adev, *addr); 1083 BUG_ON(*addr & 0xFFFF00000000003FULL); 1084 1085 if (!adev->gmc.translate_further) 1086 return; 1087 1088 if (level == AMDGPU_VM_PDB1) { 1089 /* Set the block fragment size */ 1090 if (!(*flags & AMDGPU_PDE_PTE)) 1091 *flags |= AMDGPU_PDE_BFS(0x9); 1092 1093 } else if (level == AMDGPU_VM_PDB0) { 1094 if (*flags & AMDGPU_PDE_PTE) { 1095 *flags &= ~AMDGPU_PDE_PTE; 1096 if (!(*flags & AMDGPU_PTE_VALID)) 1097 *addr |= 1 << PAGE_SHIFT; 1098 } else { 1099 *flags |= AMDGPU_PTE_TF; 1100 } 1101 } 1102 } 1103 1104 static void gmc_v9_0_get_coherence_flags(struct amdgpu_device *adev, 1105 struct amdgpu_bo *bo, 1106 struct amdgpu_bo_va_mapping *mapping, 1107 uint64_t *flags) 1108 { 1109 struct amdgpu_device *bo_adev = amdgpu_ttm_adev(bo->tbo.bdev); 1110 bool is_vram = bo->tbo.resource->mem_type == TTM_PL_VRAM; 1111 bool coherent = bo->flags & (AMDGPU_GEM_CREATE_COHERENT | AMDGPU_GEM_CREATE_EXT_COHERENT); 1112 bool ext_coherent = bo->flags & AMDGPU_GEM_CREATE_EXT_COHERENT; 1113 bool uncached = bo->flags & AMDGPU_GEM_CREATE_UNCACHED; 1114 struct amdgpu_vm *vm = mapping->bo_va->base.vm; 1115 unsigned int mtype_local, mtype; 1116 bool snoop = false; 1117 bool is_local; 1118 1119 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 1120 case IP_VERSION(9, 4, 1): 1121 case IP_VERSION(9, 4, 2): 1122 if (is_vram) { 1123 if (bo_adev == adev) { 1124 if (uncached) 1125 mtype = MTYPE_UC; 1126 else if (coherent) 1127 mtype = MTYPE_CC; 1128 else 1129 mtype = MTYPE_RW; 1130 /* FIXME: is this still needed? Or does 1131 * amdgpu_ttm_tt_pde_flags already handle this? 1132 */ 1133 if ((amdgpu_ip_version(adev, GC_HWIP, 0) == 1134 IP_VERSION(9, 4, 2) || 1135 amdgpu_ip_version(adev, GC_HWIP, 0) == 1136 IP_VERSION(9, 4, 3)) && 1137 adev->gmc.xgmi.connected_to_cpu) 1138 snoop = true; 1139 } else { 1140 if (uncached || coherent) 1141 mtype = MTYPE_UC; 1142 else 1143 mtype = MTYPE_NC; 1144 if (mapping->bo_va->is_xgmi) 1145 snoop = true; 1146 } 1147 } else { 1148 if (uncached || coherent) 1149 mtype = MTYPE_UC; 1150 else 1151 mtype = MTYPE_NC; 1152 /* FIXME: is this still needed? Or does 1153 * amdgpu_ttm_tt_pde_flags already handle this? 1154 */ 1155 snoop = true; 1156 } 1157 break; 1158 case IP_VERSION(9, 4, 3): 1159 /* Only local VRAM BOs or system memory on non-NUMA APUs 1160 * can be assumed to be local in their entirety. Choose 1161 * MTYPE_NC as safe fallback for all system memory BOs on 1162 * NUMA systems. Their MTYPE can be overridden per-page in 1163 * gmc_v9_0_override_vm_pte_flags. 1164 */ 1165 mtype_local = MTYPE_RW; 1166 if (amdgpu_mtype_local == 1) { 1167 DRM_INFO_ONCE("Using MTYPE_NC for local memory\n"); 1168 mtype_local = MTYPE_NC; 1169 } else if (amdgpu_mtype_local == 2) { 1170 DRM_INFO_ONCE("Using MTYPE_CC for local memory\n"); 1171 mtype_local = MTYPE_CC; 1172 } else { 1173 DRM_INFO_ONCE("Using MTYPE_RW for local memory\n"); 1174 } 1175 is_local = (!is_vram && (adev->flags & AMD_IS_APU) && 1176 num_possible_nodes() <= 1) || 1177 (is_vram && adev == bo_adev && 1178 KFD_XCP_MEM_ID(adev, bo->xcp_id) == vm->mem_id); 1179 snoop = true; 1180 if (uncached) { 1181 mtype = MTYPE_UC; 1182 } else if (ext_coherent) { 1183 if (adev->rev_id) 1184 mtype = is_local ? MTYPE_CC : MTYPE_UC; 1185 else 1186 mtype = MTYPE_UC; 1187 } else if (adev->flags & AMD_IS_APU) { 1188 mtype = is_local ? mtype_local : MTYPE_NC; 1189 } else { 1190 /* dGPU */ 1191 if (is_local) 1192 mtype = mtype_local; 1193 else if (is_vram) 1194 mtype = MTYPE_NC; 1195 else 1196 mtype = MTYPE_UC; 1197 } 1198 1199 break; 1200 default: 1201 if (uncached || coherent) 1202 mtype = MTYPE_UC; 1203 else 1204 mtype = MTYPE_NC; 1205 1206 /* FIXME: is this still needed? Or does 1207 * amdgpu_ttm_tt_pde_flags already handle this? 1208 */ 1209 if (!is_vram) 1210 snoop = true; 1211 } 1212 1213 if (mtype != MTYPE_NC) 1214 *flags = (*flags & ~AMDGPU_PTE_MTYPE_VG10_MASK) | 1215 AMDGPU_PTE_MTYPE_VG10(mtype); 1216 *flags |= snoop ? AMDGPU_PTE_SNOOPED : 0; 1217 } 1218 1219 static void gmc_v9_0_get_vm_pte(struct amdgpu_device *adev, 1220 struct amdgpu_bo_va_mapping *mapping, 1221 uint64_t *flags) 1222 { 1223 struct amdgpu_bo *bo = mapping->bo_va->base.bo; 1224 1225 *flags &= ~AMDGPU_PTE_EXECUTABLE; 1226 *flags |= mapping->flags & AMDGPU_PTE_EXECUTABLE; 1227 1228 *flags &= ~AMDGPU_PTE_MTYPE_VG10_MASK; 1229 *flags |= mapping->flags & AMDGPU_PTE_MTYPE_VG10_MASK; 1230 1231 if (mapping->flags & AMDGPU_PTE_PRT) { 1232 *flags |= AMDGPU_PTE_PRT; 1233 *flags &= ~AMDGPU_PTE_VALID; 1234 } 1235 1236 if (bo && bo->tbo.resource) 1237 gmc_v9_0_get_coherence_flags(adev, mapping->bo_va->base.bo, 1238 mapping, flags); 1239 } 1240 1241 static void gmc_v9_0_override_vm_pte_flags(struct amdgpu_device *adev, 1242 struct amdgpu_vm *vm, 1243 uint64_t addr, uint64_t *flags) 1244 { 1245 int local_node, nid; 1246 1247 /* Only GFX 9.4.3 APUs associate GPUs with NUMA nodes. Local system 1248 * memory can use more efficient MTYPEs. 1249 */ 1250 if (amdgpu_ip_version(adev, GC_HWIP, 0) != IP_VERSION(9, 4, 3)) 1251 return; 1252 1253 /* Only direct-mapped memory allows us to determine the NUMA node from 1254 * the DMA address. 1255 */ 1256 if (!adev->ram_is_direct_mapped) { 1257 dev_dbg_ratelimited(adev->dev, "RAM is not direct mapped\n"); 1258 return; 1259 } 1260 1261 /* MTYPE_NC is the same default and can be overridden. 1262 * MTYPE_UC will be present if the memory is extended-coherent 1263 * and can also be overridden. 1264 */ 1265 if ((*flags & AMDGPU_PTE_MTYPE_VG10_MASK) != 1266 AMDGPU_PTE_MTYPE_VG10(MTYPE_NC) && 1267 (*flags & AMDGPU_PTE_MTYPE_VG10_MASK) != 1268 AMDGPU_PTE_MTYPE_VG10(MTYPE_UC)) { 1269 dev_dbg_ratelimited(adev->dev, "MTYPE is not NC or UC\n"); 1270 return; 1271 } 1272 1273 /* FIXME: Only supported on native mode for now. For carve-out, the 1274 * NUMA affinity of the GPU/VM needs to come from the PCI info because 1275 * memory partitions are not associated with different NUMA nodes. 1276 */ 1277 if (adev->gmc.is_app_apu && vm->mem_id >= 0) { 1278 local_node = adev->gmc.mem_partitions[vm->mem_id].numa.node; 1279 } else { 1280 dev_dbg_ratelimited(adev->dev, "Only native mode APU is supported.\n"); 1281 return; 1282 } 1283 1284 /* Only handle real RAM. Mappings of PCIe resources don't have struct 1285 * page or NUMA nodes. 1286 */ 1287 if (!page_is_ram(addr >> PAGE_SHIFT)) { 1288 dev_dbg_ratelimited(adev->dev, "Page is not RAM.\n"); 1289 return; 1290 } 1291 nid = pfn_to_nid(addr >> PAGE_SHIFT); 1292 dev_dbg_ratelimited(adev->dev, "vm->mem_id=%d, local_node=%d, nid=%d\n", 1293 vm->mem_id, local_node, nid); 1294 if (nid == local_node) { 1295 uint64_t old_flags = *flags; 1296 if ((*flags & AMDGPU_PTE_MTYPE_VG10_MASK) == 1297 AMDGPU_PTE_MTYPE_VG10(MTYPE_NC)) { 1298 unsigned int mtype_local = MTYPE_RW; 1299 1300 if (amdgpu_mtype_local == 1) 1301 mtype_local = MTYPE_NC; 1302 else if (amdgpu_mtype_local == 2) 1303 mtype_local = MTYPE_CC; 1304 1305 *flags = (*flags & ~AMDGPU_PTE_MTYPE_VG10_MASK) | 1306 AMDGPU_PTE_MTYPE_VG10(mtype_local); 1307 } else if (adev->rev_id) { 1308 /* MTYPE_UC case */ 1309 *flags = (*flags & ~AMDGPU_PTE_MTYPE_VG10_MASK) | 1310 AMDGPU_PTE_MTYPE_VG10(MTYPE_CC); 1311 } 1312 1313 dev_dbg_ratelimited(adev->dev, "flags updated from %llx to %llx\n", 1314 old_flags, *flags); 1315 } 1316 } 1317 1318 static unsigned int gmc_v9_0_get_vbios_fb_size(struct amdgpu_device *adev) 1319 { 1320 u32 d1vga_control = RREG32_SOC15(DCE, 0, mmD1VGA_CONTROL); 1321 unsigned int size; 1322 1323 /* TODO move to DC so GMC doesn't need to hard-code DCN registers */ 1324 1325 if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) { 1326 size = AMDGPU_VBIOS_VGA_ALLOCATION; 1327 } else { 1328 u32 viewport; 1329 1330 switch (amdgpu_ip_version(adev, DCE_HWIP, 0)) { 1331 case IP_VERSION(1, 0, 0): 1332 case IP_VERSION(1, 0, 1): 1333 viewport = RREG32_SOC15(DCE, 0, mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION); 1334 size = (REG_GET_FIELD(viewport, 1335 HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_HEIGHT) * 1336 REG_GET_FIELD(viewport, 1337 HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_WIDTH) * 1338 4); 1339 break; 1340 case IP_VERSION(2, 1, 0): 1341 viewport = RREG32_SOC15(DCE, 0, mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION_DCN2); 1342 size = (REG_GET_FIELD(viewport, 1343 HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_HEIGHT) * 1344 REG_GET_FIELD(viewport, 1345 HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_WIDTH) * 1346 4); 1347 break; 1348 default: 1349 viewport = RREG32_SOC15(DCE, 0, mmSCL0_VIEWPORT_SIZE); 1350 size = (REG_GET_FIELD(viewport, SCL0_VIEWPORT_SIZE, VIEWPORT_HEIGHT) * 1351 REG_GET_FIELD(viewport, SCL0_VIEWPORT_SIZE, VIEWPORT_WIDTH) * 1352 4); 1353 break; 1354 } 1355 } 1356 1357 return size; 1358 } 1359 1360 static enum amdgpu_memory_partition 1361 gmc_v9_0_get_memory_partition(struct amdgpu_device *adev, u32 *supp_modes) 1362 { 1363 enum amdgpu_memory_partition mode = UNKNOWN_MEMORY_PARTITION_MODE; 1364 1365 if (adev->nbio.funcs->get_memory_partition_mode) 1366 mode = adev->nbio.funcs->get_memory_partition_mode(adev, 1367 supp_modes); 1368 1369 return mode; 1370 } 1371 1372 static enum amdgpu_memory_partition 1373 gmc_v9_0_query_memory_partition(struct amdgpu_device *adev) 1374 { 1375 if (amdgpu_sriov_vf(adev)) 1376 return AMDGPU_NPS1_PARTITION_MODE; 1377 1378 return gmc_v9_0_get_memory_partition(adev, NULL); 1379 } 1380 1381 static const struct amdgpu_gmc_funcs gmc_v9_0_gmc_funcs = { 1382 .flush_gpu_tlb = gmc_v9_0_flush_gpu_tlb, 1383 .flush_gpu_tlb_pasid = gmc_v9_0_flush_gpu_tlb_pasid, 1384 .emit_flush_gpu_tlb = gmc_v9_0_emit_flush_gpu_tlb, 1385 .emit_pasid_mapping = gmc_v9_0_emit_pasid_mapping, 1386 .map_mtype = gmc_v9_0_map_mtype, 1387 .get_vm_pde = gmc_v9_0_get_vm_pde, 1388 .get_vm_pte = gmc_v9_0_get_vm_pte, 1389 .override_vm_pte_flags = gmc_v9_0_override_vm_pte_flags, 1390 .get_vbios_fb_size = gmc_v9_0_get_vbios_fb_size, 1391 .query_mem_partition_mode = &gmc_v9_0_query_memory_partition, 1392 }; 1393 1394 static void gmc_v9_0_set_gmc_funcs(struct amdgpu_device *adev) 1395 { 1396 adev->gmc.gmc_funcs = &gmc_v9_0_gmc_funcs; 1397 } 1398 1399 static void gmc_v9_0_set_umc_funcs(struct amdgpu_device *adev) 1400 { 1401 switch (amdgpu_ip_version(adev, UMC_HWIP, 0)) { 1402 case IP_VERSION(6, 0, 0): 1403 adev->umc.funcs = &umc_v6_0_funcs; 1404 break; 1405 case IP_VERSION(6, 1, 1): 1406 adev->umc.max_ras_err_cnt_per_query = UMC_V6_1_TOTAL_CHANNEL_NUM; 1407 adev->umc.channel_inst_num = UMC_V6_1_CHANNEL_INSTANCE_NUM; 1408 adev->umc.umc_inst_num = UMC_V6_1_UMC_INSTANCE_NUM; 1409 adev->umc.channel_offs = UMC_V6_1_PER_CHANNEL_OFFSET_VG20; 1410 adev->umc.retire_unit = 1; 1411 adev->umc.channel_idx_tbl = &umc_v6_1_channel_idx_tbl[0][0]; 1412 adev->umc.ras = &umc_v6_1_ras; 1413 break; 1414 case IP_VERSION(6, 1, 2): 1415 adev->umc.max_ras_err_cnt_per_query = UMC_V6_1_TOTAL_CHANNEL_NUM; 1416 adev->umc.channel_inst_num = UMC_V6_1_CHANNEL_INSTANCE_NUM; 1417 adev->umc.umc_inst_num = UMC_V6_1_UMC_INSTANCE_NUM; 1418 adev->umc.channel_offs = UMC_V6_1_PER_CHANNEL_OFFSET_ARCT; 1419 adev->umc.retire_unit = 1; 1420 adev->umc.channel_idx_tbl = &umc_v6_1_channel_idx_tbl[0][0]; 1421 adev->umc.ras = &umc_v6_1_ras; 1422 break; 1423 case IP_VERSION(6, 7, 0): 1424 adev->umc.max_ras_err_cnt_per_query = 1425 UMC_V6_7_TOTAL_CHANNEL_NUM * UMC_V6_7_BAD_PAGE_NUM_PER_CHANNEL; 1426 adev->umc.channel_inst_num = UMC_V6_7_CHANNEL_INSTANCE_NUM; 1427 adev->umc.umc_inst_num = UMC_V6_7_UMC_INSTANCE_NUM; 1428 adev->umc.channel_offs = UMC_V6_7_PER_CHANNEL_OFFSET; 1429 adev->umc.retire_unit = (UMC_V6_7_NA_MAP_PA_NUM * 2); 1430 if (!adev->gmc.xgmi.connected_to_cpu) 1431 adev->umc.ras = &umc_v6_7_ras; 1432 if (1 & adev->smuio.funcs->get_die_id(adev)) 1433 adev->umc.channel_idx_tbl = &umc_v6_7_channel_idx_tbl_first[0][0]; 1434 else 1435 adev->umc.channel_idx_tbl = &umc_v6_7_channel_idx_tbl_second[0][0]; 1436 break; 1437 case IP_VERSION(12, 0, 0): 1438 adev->umc.max_ras_err_cnt_per_query = 1439 UMC_V12_0_TOTAL_CHANNEL_NUM(adev) * UMC_V12_0_BAD_PAGE_NUM_PER_CHANNEL; 1440 adev->umc.channel_inst_num = UMC_V12_0_CHANNEL_INSTANCE_NUM; 1441 adev->umc.umc_inst_num = UMC_V12_0_UMC_INSTANCE_NUM; 1442 adev->umc.node_inst_num /= UMC_V12_0_UMC_INSTANCE_NUM; 1443 adev->umc.channel_offs = UMC_V12_0_PER_CHANNEL_OFFSET; 1444 adev->umc.active_mask = adev->aid_mask; 1445 adev->umc.retire_unit = UMC_V12_0_BAD_PAGE_NUM_PER_CHANNEL; 1446 adev->umc.channel_idx_tbl = &umc_v12_0_channel_idx_tbl[0][0][0]; 1447 if (!adev->gmc.xgmi.connected_to_cpu && !adev->gmc.is_app_apu) 1448 adev->umc.ras = &umc_v12_0_ras; 1449 break; 1450 default: 1451 break; 1452 } 1453 } 1454 1455 static void gmc_v9_0_set_mmhub_funcs(struct amdgpu_device *adev) 1456 { 1457 switch (amdgpu_ip_version(adev, MMHUB_HWIP, 0)) { 1458 case IP_VERSION(9, 4, 1): 1459 adev->mmhub.funcs = &mmhub_v9_4_funcs; 1460 break; 1461 case IP_VERSION(9, 4, 2): 1462 adev->mmhub.funcs = &mmhub_v1_7_funcs; 1463 break; 1464 case IP_VERSION(1, 8, 0): 1465 adev->mmhub.funcs = &mmhub_v1_8_funcs; 1466 break; 1467 default: 1468 adev->mmhub.funcs = &mmhub_v1_0_funcs; 1469 break; 1470 } 1471 } 1472 1473 static void gmc_v9_0_set_mmhub_ras_funcs(struct amdgpu_device *adev) 1474 { 1475 switch (amdgpu_ip_version(adev, MMHUB_HWIP, 0)) { 1476 case IP_VERSION(9, 4, 0): 1477 adev->mmhub.ras = &mmhub_v1_0_ras; 1478 break; 1479 case IP_VERSION(9, 4, 1): 1480 adev->mmhub.ras = &mmhub_v9_4_ras; 1481 break; 1482 case IP_VERSION(9, 4, 2): 1483 adev->mmhub.ras = &mmhub_v1_7_ras; 1484 break; 1485 case IP_VERSION(1, 8, 0): 1486 adev->mmhub.ras = &mmhub_v1_8_ras; 1487 break; 1488 default: 1489 /* mmhub ras is not available */ 1490 break; 1491 } 1492 } 1493 1494 static void gmc_v9_0_set_gfxhub_funcs(struct amdgpu_device *adev) 1495 { 1496 if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3)) 1497 adev->gfxhub.funcs = &gfxhub_v1_2_funcs; 1498 else 1499 adev->gfxhub.funcs = &gfxhub_v1_0_funcs; 1500 } 1501 1502 static void gmc_v9_0_set_hdp_ras_funcs(struct amdgpu_device *adev) 1503 { 1504 adev->hdp.ras = &hdp_v4_0_ras; 1505 } 1506 1507 static void gmc_v9_0_set_mca_ras_funcs(struct amdgpu_device *adev) 1508 { 1509 struct amdgpu_mca *mca = &adev->mca; 1510 1511 /* is UMC the right IP to check for MCA? Maybe DF? */ 1512 switch (amdgpu_ip_version(adev, UMC_HWIP, 0)) { 1513 case IP_VERSION(6, 7, 0): 1514 if (!adev->gmc.xgmi.connected_to_cpu) { 1515 mca->mp0.ras = &mca_v3_0_mp0_ras; 1516 mca->mp1.ras = &mca_v3_0_mp1_ras; 1517 mca->mpio.ras = &mca_v3_0_mpio_ras; 1518 } 1519 break; 1520 default: 1521 break; 1522 } 1523 } 1524 1525 static void gmc_v9_0_set_xgmi_ras_funcs(struct amdgpu_device *adev) 1526 { 1527 if (!adev->gmc.xgmi.connected_to_cpu) 1528 adev->gmc.xgmi.ras = &xgmi_ras; 1529 } 1530 1531 static int gmc_v9_0_early_init(void *handle) 1532 { 1533 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1534 1535 /* 1536 * 9.4.0, 9.4.1 and 9.4.3 don't have XGMI defined 1537 * in their IP discovery tables 1538 */ 1539 if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 0) || 1540 amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 1) || 1541 amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3)) 1542 adev->gmc.xgmi.supported = true; 1543 1544 if (amdgpu_ip_version(adev, XGMI_HWIP, 0) == IP_VERSION(6, 1, 0)) { 1545 adev->gmc.xgmi.supported = true; 1546 adev->gmc.xgmi.connected_to_cpu = 1547 adev->smuio.funcs->is_host_gpu_xgmi_supported(adev); 1548 } 1549 1550 if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3)) { 1551 enum amdgpu_pkg_type pkg_type = 1552 adev->smuio.funcs->get_pkg_type(adev); 1553 /* On GFXIP 9.4.3. APU, there is no physical VRAM domain present 1554 * and the APU, can be in used two possible modes: 1555 * - carveout mode 1556 * - native APU mode 1557 * "is_app_apu" can be used to identify the APU in the native 1558 * mode. 1559 */ 1560 adev->gmc.is_app_apu = (pkg_type == AMDGPU_PKG_TYPE_APU && 1561 !pci_resource_len(adev->pdev, 0)); 1562 } 1563 1564 gmc_v9_0_set_gmc_funcs(adev); 1565 gmc_v9_0_set_irq_funcs(adev); 1566 gmc_v9_0_set_umc_funcs(adev); 1567 gmc_v9_0_set_mmhub_funcs(adev); 1568 gmc_v9_0_set_mmhub_ras_funcs(adev); 1569 gmc_v9_0_set_gfxhub_funcs(adev); 1570 gmc_v9_0_set_hdp_ras_funcs(adev); 1571 gmc_v9_0_set_mca_ras_funcs(adev); 1572 gmc_v9_0_set_xgmi_ras_funcs(adev); 1573 1574 adev->gmc.shared_aperture_start = 0x2000000000000000ULL; 1575 adev->gmc.shared_aperture_end = 1576 adev->gmc.shared_aperture_start + (4ULL << 30) - 1; 1577 adev->gmc.private_aperture_start = 0x1000000000000000ULL; 1578 adev->gmc.private_aperture_end = 1579 adev->gmc.private_aperture_start + (4ULL << 30) - 1; 1580 adev->gmc.noretry_flags = AMDGPU_VM_NORETRY_FLAGS_TF; 1581 1582 return 0; 1583 } 1584 1585 static int gmc_v9_0_late_init(void *handle) 1586 { 1587 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1588 int r; 1589 1590 r = amdgpu_gmc_allocate_vm_inv_eng(adev); 1591 if (r) 1592 return r; 1593 1594 /* 1595 * Workaround performance drop issue with VBIOS enables partial 1596 * writes, while disables HBM ECC for vega10. 1597 */ 1598 if (!amdgpu_sriov_vf(adev) && 1599 (amdgpu_ip_version(adev, UMC_HWIP, 0) == IP_VERSION(6, 0, 0))) { 1600 if (!(adev->ras_enabled & (1 << AMDGPU_RAS_BLOCK__UMC))) { 1601 if (adev->df.funcs && 1602 adev->df.funcs->enable_ecc_force_par_wr_rmw) 1603 adev->df.funcs->enable_ecc_force_par_wr_rmw(adev, false); 1604 } 1605 } 1606 1607 if (!amdgpu_persistent_edc_harvesting_supported(adev)) { 1608 amdgpu_ras_reset_error_count(adev, AMDGPU_RAS_BLOCK__MMHUB); 1609 amdgpu_ras_reset_error_count(adev, AMDGPU_RAS_BLOCK__HDP); 1610 } 1611 1612 r = amdgpu_gmc_ras_late_init(adev); 1613 if (r) 1614 return r; 1615 1616 return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0); 1617 } 1618 1619 static void gmc_v9_0_vram_gtt_location(struct amdgpu_device *adev, 1620 struct amdgpu_gmc *mc) 1621 { 1622 u64 base = adev->mmhub.funcs->get_fb_location(adev); 1623 1624 amdgpu_gmc_set_agp_default(adev, mc); 1625 1626 /* add the xgmi offset of the physical node */ 1627 base += adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size; 1628 if (adev->gmc.xgmi.connected_to_cpu) { 1629 amdgpu_gmc_sysvm_location(adev, mc); 1630 } else { 1631 amdgpu_gmc_vram_location(adev, mc, base); 1632 amdgpu_gmc_gart_location(adev, mc, AMDGPU_GART_PLACEMENT_BEST_FIT); 1633 if (!amdgpu_sriov_vf(adev)) 1634 amdgpu_gmc_agp_location(adev, mc); 1635 } 1636 /* base offset of vram pages */ 1637 adev->vm_manager.vram_base_offset = adev->gfxhub.funcs->get_mc_fb_offset(adev); 1638 1639 /* XXX: add the xgmi offset of the physical node? */ 1640 adev->vm_manager.vram_base_offset += 1641 adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size; 1642 } 1643 1644 /** 1645 * gmc_v9_0_mc_init - initialize the memory controller driver params 1646 * 1647 * @adev: amdgpu_device pointer 1648 * 1649 * Look up the amount of vram, vram width, and decide how to place 1650 * vram and gart within the GPU's physical address space. 1651 * Returns 0 for success. 1652 */ 1653 static int gmc_v9_0_mc_init(struct amdgpu_device *adev) 1654 { 1655 int r; 1656 1657 /* size in MB on si */ 1658 if (!adev->gmc.is_app_apu) { 1659 adev->gmc.mc_vram_size = 1660 adev->nbio.funcs->get_memsize(adev) * 1024ULL * 1024ULL; 1661 } else { 1662 DRM_DEBUG("Set mc_vram_size = 0 for APP APU\n"); 1663 adev->gmc.mc_vram_size = 0; 1664 } 1665 adev->gmc.real_vram_size = adev->gmc.mc_vram_size; 1666 1667 if (!(adev->flags & AMD_IS_APU) && 1668 !adev->gmc.xgmi.connected_to_cpu) { 1669 r = amdgpu_device_resize_fb_bar(adev); 1670 if (r) 1671 return r; 1672 } 1673 adev->gmc.aper_base = pci_resource_start(adev->pdev, 0); 1674 adev->gmc.aper_size = pci_resource_len(adev->pdev, 0); 1675 1676 #ifdef CONFIG_X86_64 1677 /* 1678 * AMD Accelerated Processing Platform (APP) supporting GPU-HOST xgmi 1679 * interface can use VRAM through here as it appears system reserved 1680 * memory in host address space. 1681 * 1682 * For APUs, VRAM is just the stolen system memory and can be accessed 1683 * directly. 1684 * 1685 * Otherwise, use the legacy Host Data Path (HDP) through PCIe BAR. 1686 */ 1687 1688 /* check whether both host-gpu and gpu-gpu xgmi links exist */ 1689 if ((!amdgpu_sriov_vf(adev) && 1690 (adev->flags & AMD_IS_APU) && !amdgpu_passthrough(adev)) || 1691 (adev->gmc.xgmi.supported && 1692 adev->gmc.xgmi.connected_to_cpu)) { 1693 adev->gmc.aper_base = 1694 adev->gfxhub.funcs->get_mc_fb_offset(adev) + 1695 adev->gmc.xgmi.physical_node_id * 1696 adev->gmc.xgmi.node_segment_size; 1697 adev->gmc.aper_size = adev->gmc.real_vram_size; 1698 } 1699 1700 #endif 1701 adev->gmc.visible_vram_size = adev->gmc.aper_size; 1702 1703 /* set the gart size */ 1704 if (amdgpu_gart_size == -1) { 1705 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 1706 case IP_VERSION(9, 0, 1): /* all engines support GPUVM */ 1707 case IP_VERSION(9, 2, 1): /* all engines support GPUVM */ 1708 case IP_VERSION(9, 4, 0): 1709 case IP_VERSION(9, 4, 1): 1710 case IP_VERSION(9, 4, 2): 1711 case IP_VERSION(9, 4, 3): 1712 default: 1713 adev->gmc.gart_size = 512ULL << 20; 1714 break; 1715 case IP_VERSION(9, 1, 0): /* DCE SG support */ 1716 case IP_VERSION(9, 2, 2): /* DCE SG support */ 1717 case IP_VERSION(9, 3, 0): 1718 adev->gmc.gart_size = 1024ULL << 20; 1719 break; 1720 } 1721 } else { 1722 adev->gmc.gart_size = (u64)amdgpu_gart_size << 20; 1723 } 1724 1725 adev->gmc.gart_size += adev->pm.smu_prv_buffer_size; 1726 1727 gmc_v9_0_vram_gtt_location(adev, &adev->gmc); 1728 1729 return 0; 1730 } 1731 1732 static int gmc_v9_0_gart_init(struct amdgpu_device *adev) 1733 { 1734 int r; 1735 1736 if (adev->gart.bo) { 1737 WARN(1, "VEGA10 PCIE GART already initialized\n"); 1738 return 0; 1739 } 1740 1741 if (adev->gmc.xgmi.connected_to_cpu) { 1742 adev->gmc.vmid0_page_table_depth = 1; 1743 adev->gmc.vmid0_page_table_block_size = 12; 1744 } else { 1745 adev->gmc.vmid0_page_table_depth = 0; 1746 adev->gmc.vmid0_page_table_block_size = 0; 1747 } 1748 1749 /* Initialize common gart structure */ 1750 r = amdgpu_gart_init(adev); 1751 if (r) 1752 return r; 1753 adev->gart.table_size = adev->gart.num_gpu_pages * 8; 1754 adev->gart.gart_pte_flags = AMDGPU_PTE_MTYPE_VG10(MTYPE_UC) | 1755 AMDGPU_PTE_EXECUTABLE; 1756 1757 if (!adev->gmc.real_vram_size) { 1758 dev_info(adev->dev, "Put GART in system memory for APU\n"); 1759 r = amdgpu_gart_table_ram_alloc(adev); 1760 if (r) 1761 dev_err(adev->dev, "Failed to allocate GART in system memory\n"); 1762 } else { 1763 r = amdgpu_gart_table_vram_alloc(adev); 1764 if (r) 1765 return r; 1766 1767 if (adev->gmc.xgmi.connected_to_cpu) 1768 r = amdgpu_gmc_pdb0_alloc(adev); 1769 } 1770 1771 return r; 1772 } 1773 1774 /** 1775 * gmc_v9_0_save_registers - saves regs 1776 * 1777 * @adev: amdgpu_device pointer 1778 * 1779 * This saves potential register values that should be 1780 * restored upon resume 1781 */ 1782 static void gmc_v9_0_save_registers(struct amdgpu_device *adev) 1783 { 1784 if ((amdgpu_ip_version(adev, DCE_HWIP, 0) == IP_VERSION(1, 0, 0)) || 1785 (amdgpu_ip_version(adev, DCE_HWIP, 0) == IP_VERSION(1, 0, 1))) 1786 adev->gmc.sdpif_register = RREG32_SOC15(DCE, 0, mmDCHUBBUB_SDPIF_MMIO_CNTRL_0); 1787 } 1788 1789 static bool gmc_v9_0_validate_partition_info(struct amdgpu_device *adev) 1790 { 1791 enum amdgpu_memory_partition mode; 1792 u32 supp_modes; 1793 bool valid; 1794 1795 mode = gmc_v9_0_get_memory_partition(adev, &supp_modes); 1796 1797 /* Mode detected by hardware not present in supported modes */ 1798 if ((mode != UNKNOWN_MEMORY_PARTITION_MODE) && 1799 !(BIT(mode - 1) & supp_modes)) 1800 return false; 1801 1802 switch (mode) { 1803 case UNKNOWN_MEMORY_PARTITION_MODE: 1804 case AMDGPU_NPS1_PARTITION_MODE: 1805 valid = (adev->gmc.num_mem_partitions == 1); 1806 break; 1807 case AMDGPU_NPS2_PARTITION_MODE: 1808 valid = (adev->gmc.num_mem_partitions == 2); 1809 break; 1810 case AMDGPU_NPS4_PARTITION_MODE: 1811 valid = (adev->gmc.num_mem_partitions == 3 || 1812 adev->gmc.num_mem_partitions == 4); 1813 break; 1814 default: 1815 valid = false; 1816 } 1817 1818 return valid; 1819 } 1820 1821 static bool gmc_v9_0_is_node_present(int *node_ids, int num_ids, int nid) 1822 { 1823 int i; 1824 1825 /* Check if node with id 'nid' is present in 'node_ids' array */ 1826 for (i = 0; i < num_ids; ++i) 1827 if (node_ids[i] == nid) 1828 return true; 1829 1830 return false; 1831 } 1832 1833 static void 1834 gmc_v9_0_init_acpi_mem_ranges(struct amdgpu_device *adev, 1835 struct amdgpu_mem_partition_info *mem_ranges) 1836 { 1837 struct amdgpu_numa_info numa_info; 1838 int node_ids[MAX_MEM_RANGES]; 1839 int num_ranges = 0, ret; 1840 int num_xcc, xcc_id; 1841 uint32_t xcc_mask; 1842 1843 num_xcc = NUM_XCC(adev->gfx.xcc_mask); 1844 xcc_mask = (1U << num_xcc) - 1; 1845 1846 for_each_inst(xcc_id, xcc_mask) { 1847 ret = amdgpu_acpi_get_mem_info(adev, xcc_id, &numa_info); 1848 if (ret) 1849 continue; 1850 1851 if (numa_info.nid == NUMA_NO_NODE) { 1852 mem_ranges[0].size = numa_info.size; 1853 mem_ranges[0].numa.node = numa_info.nid; 1854 num_ranges = 1; 1855 break; 1856 } 1857 1858 if (gmc_v9_0_is_node_present(node_ids, num_ranges, 1859 numa_info.nid)) 1860 continue; 1861 1862 node_ids[num_ranges] = numa_info.nid; 1863 mem_ranges[num_ranges].numa.node = numa_info.nid; 1864 mem_ranges[num_ranges].size = numa_info.size; 1865 ++num_ranges; 1866 } 1867 1868 adev->gmc.num_mem_partitions = num_ranges; 1869 } 1870 1871 static void 1872 gmc_v9_0_init_sw_mem_ranges(struct amdgpu_device *adev, 1873 struct amdgpu_mem_partition_info *mem_ranges) 1874 { 1875 enum amdgpu_memory_partition mode; 1876 u32 start_addr = 0, size; 1877 int i; 1878 1879 mode = gmc_v9_0_query_memory_partition(adev); 1880 1881 switch (mode) { 1882 case UNKNOWN_MEMORY_PARTITION_MODE: 1883 case AMDGPU_NPS1_PARTITION_MODE: 1884 adev->gmc.num_mem_partitions = 1; 1885 break; 1886 case AMDGPU_NPS2_PARTITION_MODE: 1887 adev->gmc.num_mem_partitions = 2; 1888 break; 1889 case AMDGPU_NPS4_PARTITION_MODE: 1890 if (adev->flags & AMD_IS_APU) 1891 adev->gmc.num_mem_partitions = 3; 1892 else 1893 adev->gmc.num_mem_partitions = 4; 1894 break; 1895 default: 1896 adev->gmc.num_mem_partitions = 1; 1897 break; 1898 } 1899 1900 size = adev->gmc.real_vram_size >> AMDGPU_GPU_PAGE_SHIFT; 1901 size /= adev->gmc.num_mem_partitions; 1902 1903 for (i = 0; i < adev->gmc.num_mem_partitions; ++i) { 1904 mem_ranges[i].range.fpfn = start_addr; 1905 mem_ranges[i].size = ((u64)size << AMDGPU_GPU_PAGE_SHIFT); 1906 mem_ranges[i].range.lpfn = start_addr + size - 1; 1907 start_addr += size; 1908 } 1909 1910 /* Adjust the last one */ 1911 mem_ranges[adev->gmc.num_mem_partitions - 1].range.lpfn = 1912 (adev->gmc.real_vram_size >> AMDGPU_GPU_PAGE_SHIFT) - 1; 1913 mem_ranges[adev->gmc.num_mem_partitions - 1].size = 1914 adev->gmc.real_vram_size - 1915 ((u64)mem_ranges[adev->gmc.num_mem_partitions - 1].range.fpfn 1916 << AMDGPU_GPU_PAGE_SHIFT); 1917 } 1918 1919 static int gmc_v9_0_init_mem_ranges(struct amdgpu_device *adev) 1920 { 1921 bool valid; 1922 1923 adev->gmc.mem_partitions = kcalloc(MAX_MEM_RANGES, 1924 sizeof(struct amdgpu_mem_partition_info), 1925 GFP_KERNEL); 1926 if (!adev->gmc.mem_partitions) 1927 return -ENOMEM; 1928 1929 /* TODO : Get the range from PSP/Discovery for dGPU */ 1930 if (adev->gmc.is_app_apu) 1931 gmc_v9_0_init_acpi_mem_ranges(adev, adev->gmc.mem_partitions); 1932 else 1933 gmc_v9_0_init_sw_mem_ranges(adev, adev->gmc.mem_partitions); 1934 1935 if (amdgpu_sriov_vf(adev)) 1936 valid = true; 1937 else 1938 valid = gmc_v9_0_validate_partition_info(adev); 1939 if (!valid) { 1940 /* TODO: handle invalid case */ 1941 dev_WARN(adev->dev, 1942 "Mem ranges not matching with hardware config"); 1943 } 1944 1945 return 0; 1946 } 1947 1948 static void gmc_v9_4_3_init_vram_info(struct amdgpu_device *adev) 1949 { 1950 static const u32 regBIF_BIOS_SCRATCH_4 = 0x50; 1951 u32 vram_info; 1952 1953 if (!amdgpu_sriov_vf(adev)) { 1954 vram_info = RREG32(regBIF_BIOS_SCRATCH_4); 1955 adev->gmc.vram_vendor = vram_info & 0xF; 1956 } 1957 adev->gmc.vram_type = AMDGPU_VRAM_TYPE_HBM; 1958 adev->gmc.vram_width = 128 * 64; 1959 } 1960 1961 static int gmc_v9_0_sw_init(void *handle) 1962 { 1963 int r, vram_width = 0, vram_type = 0, vram_vendor = 0, dma_addr_bits; 1964 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1965 unsigned long inst_mask = adev->aid_mask; 1966 1967 adev->gfxhub.funcs->init(adev); 1968 1969 adev->mmhub.funcs->init(adev); 1970 1971 spin_lock_init(&adev->gmc.invalidate_lock); 1972 1973 if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3)) { 1974 gmc_v9_4_3_init_vram_info(adev); 1975 } else if (!adev->bios) { 1976 if (adev->flags & AMD_IS_APU) { 1977 adev->gmc.vram_type = AMDGPU_VRAM_TYPE_DDR4; 1978 adev->gmc.vram_width = 64 * 64; 1979 } else { 1980 adev->gmc.vram_type = AMDGPU_VRAM_TYPE_HBM; 1981 adev->gmc.vram_width = 128 * 64; 1982 } 1983 } else { 1984 r = amdgpu_atomfirmware_get_vram_info(adev, 1985 &vram_width, &vram_type, &vram_vendor); 1986 if (amdgpu_sriov_vf(adev)) 1987 /* For Vega10 SR-IOV, vram_width can't be read from ATOM as RAVEN, 1988 * and DF related registers is not readable, seems hardcord is the 1989 * only way to set the correct vram_width 1990 */ 1991 adev->gmc.vram_width = 2048; 1992 else if (amdgpu_emu_mode != 1) 1993 adev->gmc.vram_width = vram_width; 1994 1995 if (!adev->gmc.vram_width) { 1996 int chansize, numchan; 1997 1998 /* hbm memory channel size */ 1999 if (adev->flags & AMD_IS_APU) 2000 chansize = 64; 2001 else 2002 chansize = 128; 2003 if (adev->df.funcs && 2004 adev->df.funcs->get_hbm_channel_number) { 2005 numchan = adev->df.funcs->get_hbm_channel_number(adev); 2006 adev->gmc.vram_width = numchan * chansize; 2007 } 2008 } 2009 2010 adev->gmc.vram_type = vram_type; 2011 adev->gmc.vram_vendor = vram_vendor; 2012 } 2013 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 2014 case IP_VERSION(9, 1, 0): 2015 case IP_VERSION(9, 2, 2): 2016 set_bit(AMDGPU_GFXHUB(0), adev->vmhubs_mask); 2017 set_bit(AMDGPU_MMHUB0(0), adev->vmhubs_mask); 2018 2019 if (adev->rev_id == 0x0 || adev->rev_id == 0x1) { 2020 amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48); 2021 } else { 2022 /* vm_size is 128TB + 512GB for legacy 3-level page support */ 2023 amdgpu_vm_adjust_size(adev, 128 * 1024 + 512, 9, 2, 48); 2024 adev->gmc.translate_further = 2025 adev->vm_manager.num_level > 1; 2026 } 2027 break; 2028 case IP_VERSION(9, 0, 1): 2029 case IP_VERSION(9, 2, 1): 2030 case IP_VERSION(9, 4, 0): 2031 case IP_VERSION(9, 3, 0): 2032 case IP_VERSION(9, 4, 2): 2033 set_bit(AMDGPU_GFXHUB(0), adev->vmhubs_mask); 2034 set_bit(AMDGPU_MMHUB0(0), adev->vmhubs_mask); 2035 2036 /* 2037 * To fulfill 4-level page support, 2038 * vm size is 256TB (48bit), maximum size of Vega10, 2039 * block size 512 (9bit) 2040 */ 2041 2042 amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48); 2043 if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 2)) 2044 adev->gmc.translate_further = adev->vm_manager.num_level > 1; 2045 break; 2046 case IP_VERSION(9, 4, 1): 2047 set_bit(AMDGPU_GFXHUB(0), adev->vmhubs_mask); 2048 set_bit(AMDGPU_MMHUB0(0), adev->vmhubs_mask); 2049 set_bit(AMDGPU_MMHUB1(0), adev->vmhubs_mask); 2050 2051 /* Keep the vm size same with Vega20 */ 2052 amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48); 2053 adev->gmc.translate_further = adev->vm_manager.num_level > 1; 2054 break; 2055 case IP_VERSION(9, 4, 3): 2056 bitmap_set(adev->vmhubs_mask, AMDGPU_GFXHUB(0), 2057 NUM_XCC(adev->gfx.xcc_mask)); 2058 2059 inst_mask <<= AMDGPU_MMHUB0(0); 2060 bitmap_or(adev->vmhubs_mask, adev->vmhubs_mask, &inst_mask, 32); 2061 2062 amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48); 2063 adev->gmc.translate_further = adev->vm_manager.num_level > 1; 2064 break; 2065 default: 2066 break; 2067 } 2068 2069 /* This interrupt is VMC page fault.*/ 2070 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC, VMC_1_0__SRCID__VM_FAULT, 2071 &adev->gmc.vm_fault); 2072 if (r) 2073 return r; 2074 2075 if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 1)) { 2076 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC1, VMC_1_0__SRCID__VM_FAULT, 2077 &adev->gmc.vm_fault); 2078 if (r) 2079 return r; 2080 } 2081 2082 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_UTCL2, UTCL2_1_0__SRCID__FAULT, 2083 &adev->gmc.vm_fault); 2084 2085 if (r) 2086 return r; 2087 2088 if (!amdgpu_sriov_vf(adev) && 2089 !adev->gmc.xgmi.connected_to_cpu && 2090 !adev->gmc.is_app_apu) { 2091 /* interrupt sent to DF. */ 2092 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_DF, 0, 2093 &adev->gmc.ecc_irq); 2094 if (r) 2095 return r; 2096 } 2097 2098 /* Set the internal MC address mask 2099 * This is the max address of the GPU's 2100 * internal address space. 2101 */ 2102 adev->gmc.mc_mask = 0xffffffffffffULL; /* 48 bit MC */ 2103 2104 dma_addr_bits = amdgpu_ip_version(adev, GC_HWIP, 0) >= 2105 IP_VERSION(9, 4, 2) ? 2106 48 : 2107 44; 2108 r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(dma_addr_bits)); 2109 if (r) { 2110 dev_warn(adev->dev, "amdgpu: No suitable DMA available.\n"); 2111 return r; 2112 } 2113 adev->need_swiotlb = drm_need_swiotlb(dma_addr_bits); 2114 2115 r = gmc_v9_0_mc_init(adev); 2116 if (r) 2117 return r; 2118 2119 amdgpu_gmc_get_vbios_allocations(adev); 2120 2121 if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3)) { 2122 r = gmc_v9_0_init_mem_ranges(adev); 2123 if (r) 2124 return r; 2125 } 2126 2127 /* Memory manager */ 2128 r = amdgpu_bo_init(adev); 2129 if (r) 2130 return r; 2131 2132 r = gmc_v9_0_gart_init(adev); 2133 if (r) 2134 return r; 2135 2136 /* 2137 * number of VMs 2138 * VMID 0 is reserved for System 2139 * amdgpu graphics/compute will use VMIDs 1..n-1 2140 * amdkfd will use VMIDs n..15 2141 * 2142 * The first KFD VMID is 8 for GPUs with graphics, 3 for 2143 * compute-only GPUs. On compute-only GPUs that leaves 2 VMIDs 2144 * for video processing. 2145 */ 2146 adev->vm_manager.first_kfd_vmid = 2147 (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 1) || 2148 amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 2) || 2149 amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3)) ? 2150 3 : 2151 8; 2152 2153 amdgpu_vm_manager_init(adev); 2154 2155 gmc_v9_0_save_registers(adev); 2156 2157 r = amdgpu_gmc_ras_sw_init(adev); 2158 if (r) 2159 return r; 2160 2161 if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3)) 2162 amdgpu_gmc_sysfs_init(adev); 2163 2164 return 0; 2165 } 2166 2167 static int gmc_v9_0_sw_fini(void *handle) 2168 { 2169 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 2170 2171 if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3)) 2172 amdgpu_gmc_sysfs_fini(adev); 2173 adev->gmc.num_mem_partitions = 0; 2174 kfree(adev->gmc.mem_partitions); 2175 2176 amdgpu_gmc_ras_fini(adev); 2177 amdgpu_gem_force_release(adev); 2178 amdgpu_vm_manager_fini(adev); 2179 if (!adev->gmc.real_vram_size) { 2180 dev_info(adev->dev, "Put GART in system memory for APU free\n"); 2181 amdgpu_gart_table_ram_free(adev); 2182 } else { 2183 amdgpu_gart_table_vram_free(adev); 2184 } 2185 amdgpu_bo_free_kernel(&adev->gmc.pdb0_bo, NULL, &adev->gmc.ptr_pdb0); 2186 amdgpu_bo_fini(adev); 2187 2188 return 0; 2189 } 2190 2191 static void gmc_v9_0_init_golden_registers(struct amdgpu_device *adev) 2192 { 2193 switch (amdgpu_ip_version(adev, MMHUB_HWIP, 0)) { 2194 case IP_VERSION(9, 0, 0): 2195 if (amdgpu_sriov_vf(adev)) 2196 break; 2197 fallthrough; 2198 case IP_VERSION(9, 4, 0): 2199 soc15_program_register_sequence(adev, 2200 golden_settings_mmhub_1_0_0, 2201 ARRAY_SIZE(golden_settings_mmhub_1_0_0)); 2202 soc15_program_register_sequence(adev, 2203 golden_settings_athub_1_0_0, 2204 ARRAY_SIZE(golden_settings_athub_1_0_0)); 2205 break; 2206 case IP_VERSION(9, 1, 0): 2207 case IP_VERSION(9, 2, 0): 2208 /* TODO for renoir */ 2209 soc15_program_register_sequence(adev, 2210 golden_settings_athub_1_0_0, 2211 ARRAY_SIZE(golden_settings_athub_1_0_0)); 2212 break; 2213 default: 2214 break; 2215 } 2216 } 2217 2218 /** 2219 * gmc_v9_0_restore_registers - restores regs 2220 * 2221 * @adev: amdgpu_device pointer 2222 * 2223 * This restores register values, saved at suspend. 2224 */ 2225 void gmc_v9_0_restore_registers(struct amdgpu_device *adev) 2226 { 2227 if ((amdgpu_ip_version(adev, DCE_HWIP, 0) == IP_VERSION(1, 0, 0)) || 2228 (amdgpu_ip_version(adev, DCE_HWIP, 0) == IP_VERSION(1, 0, 1))) { 2229 WREG32_SOC15(DCE, 0, mmDCHUBBUB_SDPIF_MMIO_CNTRL_0, adev->gmc.sdpif_register); 2230 WARN_ON(adev->gmc.sdpif_register != 2231 RREG32_SOC15(DCE, 0, mmDCHUBBUB_SDPIF_MMIO_CNTRL_0)); 2232 } 2233 } 2234 2235 /** 2236 * gmc_v9_0_gart_enable - gart enable 2237 * 2238 * @adev: amdgpu_device pointer 2239 */ 2240 static int gmc_v9_0_gart_enable(struct amdgpu_device *adev) 2241 { 2242 int r; 2243 2244 if (adev->gmc.xgmi.connected_to_cpu) 2245 amdgpu_gmc_init_pdb0(adev); 2246 2247 if (adev->gart.bo == NULL) { 2248 dev_err(adev->dev, "No VRAM object for PCIE GART.\n"); 2249 return -EINVAL; 2250 } 2251 2252 amdgpu_gtt_mgr_recover(&adev->mman.gtt_mgr); 2253 2254 if (!adev->in_s0ix) { 2255 r = adev->gfxhub.funcs->gart_enable(adev); 2256 if (r) 2257 return r; 2258 } 2259 2260 r = adev->mmhub.funcs->gart_enable(adev); 2261 if (r) 2262 return r; 2263 2264 DRM_INFO("PCIE GART of %uM enabled.\n", 2265 (unsigned int)(adev->gmc.gart_size >> 20)); 2266 if (adev->gmc.pdb0_bo) 2267 DRM_INFO("PDB0 located at 0x%016llX\n", 2268 (unsigned long long)amdgpu_bo_gpu_offset(adev->gmc.pdb0_bo)); 2269 DRM_INFO("PTB located at 0x%016llX\n", 2270 (unsigned long long)amdgpu_bo_gpu_offset(adev->gart.bo)); 2271 2272 return 0; 2273 } 2274 2275 static int gmc_v9_0_hw_init(void *handle) 2276 { 2277 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 2278 bool value; 2279 int i, r; 2280 2281 adev->gmc.flush_pasid_uses_kiq = true; 2282 2283 /* Vega20+XGMI caches PTEs in TC and TLB. Add a heavy-weight TLB flush 2284 * (type 2), which flushes both. Due to a race condition with 2285 * concurrent memory accesses using the same TLB cache line, we still 2286 * need a second TLB flush after this. 2287 */ 2288 adev->gmc.flush_tlb_needs_extra_type_2 = 2289 amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 0) && 2290 adev->gmc.xgmi.num_physical_nodes; 2291 /* 2292 * TODO: This workaround is badly documented and had a buggy 2293 * implementation. We should probably verify what we do here. 2294 */ 2295 adev->gmc.flush_tlb_needs_extra_type_0 = 2296 amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3) && 2297 adev->rev_id == 0; 2298 2299 /* The sequence of these two function calls matters.*/ 2300 gmc_v9_0_init_golden_registers(adev); 2301 2302 if (adev->mode_info.num_crtc) { 2303 /* Lockout access through VGA aperture*/ 2304 WREG32_FIELD15(DCE, 0, VGA_HDP_CONTROL, VGA_MEMORY_DISABLE, 1); 2305 /* disable VGA render */ 2306 WREG32_FIELD15(DCE, 0, VGA_RENDER_CONTROL, VGA_VSTATUS_CNTL, 0); 2307 } 2308 2309 if (adev->mmhub.funcs->update_power_gating) 2310 adev->mmhub.funcs->update_power_gating(adev, true); 2311 2312 adev->hdp.funcs->init_registers(adev); 2313 2314 /* After HDP is initialized, flush HDP.*/ 2315 adev->hdp.funcs->flush_hdp(adev, NULL); 2316 2317 if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS) 2318 value = false; 2319 else 2320 value = true; 2321 2322 if (!amdgpu_sriov_vf(adev)) { 2323 if (!adev->in_s0ix) 2324 adev->gfxhub.funcs->set_fault_enable_default(adev, value); 2325 adev->mmhub.funcs->set_fault_enable_default(adev, value); 2326 } 2327 for_each_set_bit(i, adev->vmhubs_mask, AMDGPU_MAX_VMHUBS) { 2328 if (adev->in_s0ix && (i == AMDGPU_GFXHUB(0))) 2329 continue; 2330 gmc_v9_0_flush_gpu_tlb(adev, 0, i, 0); 2331 } 2332 2333 if (adev->umc.funcs && adev->umc.funcs->init_registers) 2334 adev->umc.funcs->init_registers(adev); 2335 2336 r = gmc_v9_0_gart_enable(adev); 2337 if (r) 2338 return r; 2339 2340 if (amdgpu_emu_mode == 1) 2341 return amdgpu_gmc_vram_checking(adev); 2342 else 2343 return r; 2344 } 2345 2346 /** 2347 * gmc_v9_0_gart_disable - gart disable 2348 * 2349 * @adev: amdgpu_device pointer 2350 * 2351 * This disables all VM page table. 2352 */ 2353 static void gmc_v9_0_gart_disable(struct amdgpu_device *adev) 2354 { 2355 if (!adev->in_s0ix) 2356 adev->gfxhub.funcs->gart_disable(adev); 2357 adev->mmhub.funcs->gart_disable(adev); 2358 } 2359 2360 static int gmc_v9_0_hw_fini(void *handle) 2361 { 2362 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 2363 2364 gmc_v9_0_gart_disable(adev); 2365 2366 if (amdgpu_sriov_vf(adev)) { 2367 /* full access mode, so don't touch any GMC register */ 2368 DRM_DEBUG("For SRIOV client, shouldn't do anything.\n"); 2369 return 0; 2370 } 2371 2372 /* 2373 * Pair the operations did in gmc_v9_0_hw_init and thus maintain 2374 * a correct cached state for GMC. Otherwise, the "gate" again 2375 * operation on S3 resuming will fail due to wrong cached state. 2376 */ 2377 if (adev->mmhub.funcs->update_power_gating) 2378 adev->mmhub.funcs->update_power_gating(adev, false); 2379 2380 amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0); 2381 2382 return 0; 2383 } 2384 2385 static int gmc_v9_0_suspend(void *handle) 2386 { 2387 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 2388 2389 return gmc_v9_0_hw_fini(adev); 2390 } 2391 2392 static int gmc_v9_0_resume(void *handle) 2393 { 2394 int r; 2395 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 2396 2397 r = gmc_v9_0_hw_init(adev); 2398 if (r) 2399 return r; 2400 2401 amdgpu_vmid_reset_all(adev); 2402 2403 return 0; 2404 } 2405 2406 static bool gmc_v9_0_is_idle(void *handle) 2407 { 2408 /* MC is always ready in GMC v9.*/ 2409 return true; 2410 } 2411 2412 static int gmc_v9_0_wait_for_idle(void *handle) 2413 { 2414 /* There is no need to wait for MC idle in GMC v9.*/ 2415 return 0; 2416 } 2417 2418 static int gmc_v9_0_soft_reset(void *handle) 2419 { 2420 /* XXX for emulation.*/ 2421 return 0; 2422 } 2423 2424 static int gmc_v9_0_set_clockgating_state(void *handle, 2425 enum amd_clockgating_state state) 2426 { 2427 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 2428 2429 adev->mmhub.funcs->set_clockgating(adev, state); 2430 2431 athub_v1_0_set_clockgating(adev, state); 2432 2433 return 0; 2434 } 2435 2436 static void gmc_v9_0_get_clockgating_state(void *handle, u64 *flags) 2437 { 2438 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 2439 2440 adev->mmhub.funcs->get_clockgating(adev, flags); 2441 2442 athub_v1_0_get_clockgating(adev, flags); 2443 } 2444 2445 static int gmc_v9_0_set_powergating_state(void *handle, 2446 enum amd_powergating_state state) 2447 { 2448 return 0; 2449 } 2450 2451 const struct amd_ip_funcs gmc_v9_0_ip_funcs = { 2452 .name = "gmc_v9_0", 2453 .early_init = gmc_v9_0_early_init, 2454 .late_init = gmc_v9_0_late_init, 2455 .sw_init = gmc_v9_0_sw_init, 2456 .sw_fini = gmc_v9_0_sw_fini, 2457 .hw_init = gmc_v9_0_hw_init, 2458 .hw_fini = gmc_v9_0_hw_fini, 2459 .suspend = gmc_v9_0_suspend, 2460 .resume = gmc_v9_0_resume, 2461 .is_idle = gmc_v9_0_is_idle, 2462 .wait_for_idle = gmc_v9_0_wait_for_idle, 2463 .soft_reset = gmc_v9_0_soft_reset, 2464 .set_clockgating_state = gmc_v9_0_set_clockgating_state, 2465 .set_powergating_state = gmc_v9_0_set_powergating_state, 2466 .get_clockgating_state = gmc_v9_0_get_clockgating_state, 2467 }; 2468 2469 const struct amdgpu_ip_block_version gmc_v9_0_ip_block = { 2470 .type = AMD_IP_BLOCK_TYPE_GMC, 2471 .major = 9, 2472 .minor = 0, 2473 .rev = 0, 2474 .funcs = &gmc_v9_0_ip_funcs, 2475 }; 2476