1 /* 2 * Copyright 2016 Advanced Micro Devices, Inc. 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice shall be included in 12 * all copies or substantial portions of the Software. 13 * 14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 20 * OTHER DEALINGS IN THE SOFTWARE. 21 * 22 */ 23 24 #include <linux/firmware.h> 25 #include <linux/pci.h> 26 27 #include <drm/drm_cache.h> 28 29 #include "amdgpu.h" 30 #include "gmc_v9_0.h" 31 #include "amdgpu_atomfirmware.h" 32 #include "amdgpu_gem.h" 33 34 #include "gc/gc_9_0_sh_mask.h" 35 #include "dce/dce_12_0_offset.h" 36 #include "dce/dce_12_0_sh_mask.h" 37 #include "vega10_enum.h" 38 #include "mmhub/mmhub_1_0_offset.h" 39 #include "athub/athub_1_0_sh_mask.h" 40 #include "athub/athub_1_0_offset.h" 41 #include "oss/osssys_4_0_offset.h" 42 43 #include "soc15.h" 44 #include "soc15d.h" 45 #include "soc15_common.h" 46 #include "umc/umc_6_0_sh_mask.h" 47 48 #include "gfxhub_v1_0.h" 49 #include "mmhub_v1_0.h" 50 #include "athub_v1_0.h" 51 #include "gfxhub_v1_1.h" 52 #include "mmhub_v9_4.h" 53 #include "mmhub_v1_7.h" 54 #include "umc_v6_1.h" 55 #include "umc_v6_0.h" 56 #include "hdp_v4_0.h" 57 58 #include "ivsrcid/vmc/irqsrcs_vmc_1_0.h" 59 60 #include "amdgpu_ras.h" 61 #include "amdgpu_xgmi.h" 62 63 /* add these here since we already include dce12 headers and these are for DCN */ 64 #define mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION 0x055d 65 #define mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION_BASE_IDX 2 66 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_WIDTH__SHIFT 0x0 67 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_HEIGHT__SHIFT 0x10 68 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_WIDTH_MASK 0x00003FFFL 69 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_HEIGHT_MASK 0x3FFF0000L 70 #define mmDCHUBBUB_SDPIF_MMIO_CNTRL_0 0x049d 71 #define mmDCHUBBUB_SDPIF_MMIO_CNTRL_0_BASE_IDX 2 72 73 74 static const char *gfxhub_client_ids[] = { 75 "CB", 76 "DB", 77 "IA", 78 "WD", 79 "CPF", 80 "CPC", 81 "CPG", 82 "RLC", 83 "TCP", 84 "SQC (inst)", 85 "SQC (data)", 86 "SQG", 87 "PA", 88 }; 89 90 static const char *mmhub_client_ids_raven[][2] = { 91 [0][0] = "MP1", 92 [1][0] = "MP0", 93 [2][0] = "VCN", 94 [3][0] = "VCNU", 95 [4][0] = "HDP", 96 [5][0] = "DCE", 97 [13][0] = "UTCL2", 98 [19][0] = "TLS", 99 [26][0] = "OSS", 100 [27][0] = "SDMA0", 101 [0][1] = "MP1", 102 [1][1] = "MP0", 103 [2][1] = "VCN", 104 [3][1] = "VCNU", 105 [4][1] = "HDP", 106 [5][1] = "XDP", 107 [6][1] = "DBGU0", 108 [7][1] = "DCE", 109 [8][1] = "DCEDWB0", 110 [9][1] = "DCEDWB1", 111 [26][1] = "OSS", 112 [27][1] = "SDMA0", 113 }; 114 115 static const char *mmhub_client_ids_renoir[][2] = { 116 [0][0] = "MP1", 117 [1][0] = "MP0", 118 [2][0] = "HDP", 119 [4][0] = "DCEDMC", 120 [5][0] = "DCEVGA", 121 [13][0] = "UTCL2", 122 [19][0] = "TLS", 123 [26][0] = "OSS", 124 [27][0] = "SDMA0", 125 [28][0] = "VCN", 126 [29][0] = "VCNU", 127 [30][0] = "JPEG", 128 [0][1] = "MP1", 129 [1][1] = "MP0", 130 [2][1] = "HDP", 131 [3][1] = "XDP", 132 [6][1] = "DBGU0", 133 [7][1] = "DCEDMC", 134 [8][1] = "DCEVGA", 135 [9][1] = "DCEDWB", 136 [26][1] = "OSS", 137 [27][1] = "SDMA0", 138 [28][1] = "VCN", 139 [29][1] = "VCNU", 140 [30][1] = "JPEG", 141 }; 142 143 static const char *mmhub_client_ids_vega10[][2] = { 144 [0][0] = "MP0", 145 [1][0] = "UVD", 146 [2][0] = "UVDU", 147 [3][0] = "HDP", 148 [13][0] = "UTCL2", 149 [14][0] = "OSS", 150 [15][0] = "SDMA1", 151 [32+0][0] = "VCE0", 152 [32+1][0] = "VCE0U", 153 [32+2][0] = "XDMA", 154 [32+3][0] = "DCE", 155 [32+4][0] = "MP1", 156 [32+14][0] = "SDMA0", 157 [0][1] = "MP0", 158 [1][1] = "UVD", 159 [2][1] = "UVDU", 160 [3][1] = "DBGU0", 161 [4][1] = "HDP", 162 [5][1] = "XDP", 163 [14][1] = "OSS", 164 [15][1] = "SDMA0", 165 [32+0][1] = "VCE0", 166 [32+1][1] = "VCE0U", 167 [32+2][1] = "XDMA", 168 [32+3][1] = "DCE", 169 [32+4][1] = "DCEDWB", 170 [32+5][1] = "MP1", 171 [32+6][1] = "DBGU1", 172 [32+14][1] = "SDMA1", 173 }; 174 175 static const char *mmhub_client_ids_vega12[][2] = { 176 [0][0] = "MP0", 177 [1][0] = "VCE0", 178 [2][0] = "VCE0U", 179 [3][0] = "HDP", 180 [13][0] = "UTCL2", 181 [14][0] = "OSS", 182 [15][0] = "SDMA1", 183 [32+0][0] = "DCE", 184 [32+1][0] = "XDMA", 185 [32+2][0] = "UVD", 186 [32+3][0] = "UVDU", 187 [32+4][0] = "MP1", 188 [32+15][0] = "SDMA0", 189 [0][1] = "MP0", 190 [1][1] = "VCE0", 191 [2][1] = "VCE0U", 192 [3][1] = "DBGU0", 193 [4][1] = "HDP", 194 [5][1] = "XDP", 195 [14][1] = "OSS", 196 [15][1] = "SDMA0", 197 [32+0][1] = "DCE", 198 [32+1][1] = "DCEDWB", 199 [32+2][1] = "XDMA", 200 [32+3][1] = "UVD", 201 [32+4][1] = "UVDU", 202 [32+5][1] = "MP1", 203 [32+6][1] = "DBGU1", 204 [32+15][1] = "SDMA1", 205 }; 206 207 static const char *mmhub_client_ids_vega20[][2] = { 208 [0][0] = "XDMA", 209 [1][0] = "DCE", 210 [2][0] = "VCE0", 211 [3][0] = "VCE0U", 212 [4][0] = "UVD", 213 [5][0] = "UVD1U", 214 [13][0] = "OSS", 215 [14][0] = "HDP", 216 [15][0] = "SDMA0", 217 [32+0][0] = "UVD", 218 [32+1][0] = "UVDU", 219 [32+2][0] = "MP1", 220 [32+3][0] = "MP0", 221 [32+12][0] = "UTCL2", 222 [32+14][0] = "SDMA1", 223 [0][1] = "XDMA", 224 [1][1] = "DCE", 225 [2][1] = "DCEDWB", 226 [3][1] = "VCE0", 227 [4][1] = "VCE0U", 228 [5][1] = "UVD1", 229 [6][1] = "UVD1U", 230 [7][1] = "DBGU0", 231 [8][1] = "XDP", 232 [13][1] = "OSS", 233 [14][1] = "HDP", 234 [15][1] = "SDMA0", 235 [32+0][1] = "UVD", 236 [32+1][1] = "UVDU", 237 [32+2][1] = "DBGU1", 238 [32+3][1] = "MP1", 239 [32+4][1] = "MP0", 240 [32+14][1] = "SDMA1", 241 }; 242 243 static const char *mmhub_client_ids_arcturus[][2] = { 244 [0][0] = "DBGU1", 245 [1][0] = "XDP", 246 [2][0] = "MP1", 247 [14][0] = "HDP", 248 [171][0] = "JPEG", 249 [172][0] = "VCN", 250 [173][0] = "VCNU", 251 [203][0] = "JPEG1", 252 [204][0] = "VCN1", 253 [205][0] = "VCN1U", 254 [256][0] = "SDMA0", 255 [257][0] = "SDMA1", 256 [258][0] = "SDMA2", 257 [259][0] = "SDMA3", 258 [260][0] = "SDMA4", 259 [261][0] = "SDMA5", 260 [262][0] = "SDMA6", 261 [263][0] = "SDMA7", 262 [384][0] = "OSS", 263 [0][1] = "DBGU1", 264 [1][1] = "XDP", 265 [2][1] = "MP1", 266 [14][1] = "HDP", 267 [171][1] = "JPEG", 268 [172][1] = "VCN", 269 [173][1] = "VCNU", 270 [203][1] = "JPEG1", 271 [204][1] = "VCN1", 272 [205][1] = "VCN1U", 273 [256][1] = "SDMA0", 274 [257][1] = "SDMA1", 275 [258][1] = "SDMA2", 276 [259][1] = "SDMA3", 277 [260][1] = "SDMA4", 278 [261][1] = "SDMA5", 279 [262][1] = "SDMA6", 280 [263][1] = "SDMA7", 281 [384][1] = "OSS", 282 }; 283 284 static const char *mmhub_client_ids_aldebaran[][2] = { 285 [2][0] = "MP1", 286 [3][0] = "MP0", 287 [32+1][0] = "DBGU_IO0", 288 [32+2][0] = "DBGU_IO2", 289 [32+4][0] = "MPIO", 290 [96+11][0] = "JPEG0", 291 [96+12][0] = "VCN0", 292 [96+13][0] = "VCNU0", 293 [128+11][0] = "JPEG1", 294 [128+12][0] = "VCN1", 295 [128+13][0] = "VCNU1", 296 [160+1][0] = "XDP", 297 [160+14][0] = "HDP", 298 [256+0][0] = "SDMA0", 299 [256+1][0] = "SDMA1", 300 [256+2][0] = "SDMA2", 301 [256+3][0] = "SDMA3", 302 [256+4][0] = "SDMA4", 303 [384+0][0] = "OSS", 304 [2][1] = "MP1", 305 [3][1] = "MP0", 306 [32+1][1] = "DBGU_IO0", 307 [32+2][1] = "DBGU_IO2", 308 [32+4][1] = "MPIO", 309 [96+11][1] = "JPEG0", 310 [96+12][1] = "VCN0", 311 [96+13][1] = "VCNU0", 312 [128+11][1] = "JPEG1", 313 [128+12][1] = "VCN1", 314 [128+13][1] = "VCNU1", 315 [160+1][1] = "XDP", 316 [160+14][1] = "HDP", 317 [256+0][1] = "SDMA0", 318 [256+1][1] = "SDMA1", 319 [256+2][1] = "SDMA2", 320 [256+3][1] = "SDMA3", 321 [256+4][1] = "SDMA4", 322 [384+0][1] = "OSS", 323 }; 324 325 static const struct soc15_reg_golden golden_settings_mmhub_1_0_0[] = 326 { 327 SOC15_REG_GOLDEN_VALUE(MMHUB, 0, mmDAGB1_WRCLI2, 0x00000007, 0xfe5fe0fa), 328 SOC15_REG_GOLDEN_VALUE(MMHUB, 0, mmMMEA1_DRAM_WR_CLI2GRP_MAP0, 0x00000030, 0x55555565) 329 }; 330 331 static const struct soc15_reg_golden golden_settings_athub_1_0_0[] = 332 { 333 SOC15_REG_GOLDEN_VALUE(ATHUB, 0, mmRPB_ARB_CNTL, 0x0000ff00, 0x00000800), 334 SOC15_REG_GOLDEN_VALUE(ATHUB, 0, mmRPB_ARB_CNTL2, 0x00ff00ff, 0x00080008) 335 }; 336 337 static const uint32_t ecc_umc_mcumc_ctrl_addrs[] = { 338 (0x000143c0 + 0x00000000), 339 (0x000143c0 + 0x00000800), 340 (0x000143c0 + 0x00001000), 341 (0x000143c0 + 0x00001800), 342 (0x000543c0 + 0x00000000), 343 (0x000543c0 + 0x00000800), 344 (0x000543c0 + 0x00001000), 345 (0x000543c0 + 0x00001800), 346 (0x000943c0 + 0x00000000), 347 (0x000943c0 + 0x00000800), 348 (0x000943c0 + 0x00001000), 349 (0x000943c0 + 0x00001800), 350 (0x000d43c0 + 0x00000000), 351 (0x000d43c0 + 0x00000800), 352 (0x000d43c0 + 0x00001000), 353 (0x000d43c0 + 0x00001800), 354 (0x001143c0 + 0x00000000), 355 (0x001143c0 + 0x00000800), 356 (0x001143c0 + 0x00001000), 357 (0x001143c0 + 0x00001800), 358 (0x001543c0 + 0x00000000), 359 (0x001543c0 + 0x00000800), 360 (0x001543c0 + 0x00001000), 361 (0x001543c0 + 0x00001800), 362 (0x001943c0 + 0x00000000), 363 (0x001943c0 + 0x00000800), 364 (0x001943c0 + 0x00001000), 365 (0x001943c0 + 0x00001800), 366 (0x001d43c0 + 0x00000000), 367 (0x001d43c0 + 0x00000800), 368 (0x001d43c0 + 0x00001000), 369 (0x001d43c0 + 0x00001800), 370 }; 371 372 static const uint32_t ecc_umc_mcumc_ctrl_mask_addrs[] = { 373 (0x000143e0 + 0x00000000), 374 (0x000143e0 + 0x00000800), 375 (0x000143e0 + 0x00001000), 376 (0x000143e0 + 0x00001800), 377 (0x000543e0 + 0x00000000), 378 (0x000543e0 + 0x00000800), 379 (0x000543e0 + 0x00001000), 380 (0x000543e0 + 0x00001800), 381 (0x000943e0 + 0x00000000), 382 (0x000943e0 + 0x00000800), 383 (0x000943e0 + 0x00001000), 384 (0x000943e0 + 0x00001800), 385 (0x000d43e0 + 0x00000000), 386 (0x000d43e0 + 0x00000800), 387 (0x000d43e0 + 0x00001000), 388 (0x000d43e0 + 0x00001800), 389 (0x001143e0 + 0x00000000), 390 (0x001143e0 + 0x00000800), 391 (0x001143e0 + 0x00001000), 392 (0x001143e0 + 0x00001800), 393 (0x001543e0 + 0x00000000), 394 (0x001543e0 + 0x00000800), 395 (0x001543e0 + 0x00001000), 396 (0x001543e0 + 0x00001800), 397 (0x001943e0 + 0x00000000), 398 (0x001943e0 + 0x00000800), 399 (0x001943e0 + 0x00001000), 400 (0x001943e0 + 0x00001800), 401 (0x001d43e0 + 0x00000000), 402 (0x001d43e0 + 0x00000800), 403 (0x001d43e0 + 0x00001000), 404 (0x001d43e0 + 0x00001800), 405 }; 406 407 static int gmc_v9_0_ecc_interrupt_state(struct amdgpu_device *adev, 408 struct amdgpu_irq_src *src, 409 unsigned type, 410 enum amdgpu_interrupt_state state) 411 { 412 u32 bits, i, tmp, reg; 413 414 /* Devices newer then VEGA10/12 shall have these programming 415 sequences performed by PSP BL */ 416 if (adev->asic_type >= CHIP_VEGA20) 417 return 0; 418 419 bits = 0x7f; 420 421 switch (state) { 422 case AMDGPU_IRQ_STATE_DISABLE: 423 for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_addrs); i++) { 424 reg = ecc_umc_mcumc_ctrl_addrs[i]; 425 tmp = RREG32(reg); 426 tmp &= ~bits; 427 WREG32(reg, tmp); 428 } 429 for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_mask_addrs); i++) { 430 reg = ecc_umc_mcumc_ctrl_mask_addrs[i]; 431 tmp = RREG32(reg); 432 tmp &= ~bits; 433 WREG32(reg, tmp); 434 } 435 break; 436 case AMDGPU_IRQ_STATE_ENABLE: 437 for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_addrs); i++) { 438 reg = ecc_umc_mcumc_ctrl_addrs[i]; 439 tmp = RREG32(reg); 440 tmp |= bits; 441 WREG32(reg, tmp); 442 } 443 for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_mask_addrs); i++) { 444 reg = ecc_umc_mcumc_ctrl_mask_addrs[i]; 445 tmp = RREG32(reg); 446 tmp |= bits; 447 WREG32(reg, tmp); 448 } 449 break; 450 default: 451 break; 452 } 453 454 return 0; 455 } 456 457 static int gmc_v9_0_vm_fault_interrupt_state(struct amdgpu_device *adev, 458 struct amdgpu_irq_src *src, 459 unsigned type, 460 enum amdgpu_interrupt_state state) 461 { 462 struct amdgpu_vmhub *hub; 463 u32 tmp, reg, bits, i, j; 464 465 bits = VM_CONTEXT1_CNTL__RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 466 VM_CONTEXT1_CNTL__DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 467 VM_CONTEXT1_CNTL__PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 468 VM_CONTEXT1_CNTL__VALID_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 469 VM_CONTEXT1_CNTL__READ_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 470 VM_CONTEXT1_CNTL__WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 471 VM_CONTEXT1_CNTL__EXECUTE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK; 472 473 switch (state) { 474 case AMDGPU_IRQ_STATE_DISABLE: 475 for (j = 0; j < adev->num_vmhubs; j++) { 476 hub = &adev->vmhub[j]; 477 for (i = 0; i < 16; i++) { 478 reg = hub->vm_context0_cntl + i; 479 tmp = RREG32(reg); 480 tmp &= ~bits; 481 WREG32(reg, tmp); 482 } 483 } 484 break; 485 case AMDGPU_IRQ_STATE_ENABLE: 486 for (j = 0; j < adev->num_vmhubs; j++) { 487 hub = &adev->vmhub[j]; 488 for (i = 0; i < 16; i++) { 489 reg = hub->vm_context0_cntl + i; 490 tmp = RREG32(reg); 491 tmp |= bits; 492 WREG32(reg, tmp); 493 } 494 } 495 break; 496 default: 497 break; 498 } 499 500 return 0; 501 } 502 503 static int gmc_v9_0_process_interrupt(struct amdgpu_device *adev, 504 struct amdgpu_irq_src *source, 505 struct amdgpu_iv_entry *entry) 506 { 507 bool retry_fault = !!(entry->src_data[1] & 0x80); 508 uint32_t status = 0, cid = 0, rw = 0; 509 struct amdgpu_task_info task_info; 510 struct amdgpu_vmhub *hub; 511 const char *mmhub_cid; 512 const char *hub_name; 513 u64 addr; 514 515 addr = (u64)entry->src_data[0] << 12; 516 addr |= ((u64)entry->src_data[1] & 0xf) << 44; 517 518 if (retry_fault) { 519 /* Returning 1 here also prevents sending the IV to the KFD */ 520 521 /* Process it onyl if it's the first fault for this address */ 522 if (entry->ih != &adev->irq.ih_soft && 523 amdgpu_gmc_filter_faults(adev, addr, entry->pasid, 524 entry->timestamp)) 525 return 1; 526 527 /* Delegate it to a different ring if the hardware hasn't 528 * already done it. 529 */ 530 if (entry->ih == &adev->irq.ih) { 531 amdgpu_irq_delegate(adev, entry, 8); 532 return 1; 533 } 534 535 /* Try to handle the recoverable page faults by filling page 536 * tables 537 */ 538 if (amdgpu_vm_handle_fault(adev, entry->pasid, addr)) 539 return 1; 540 } 541 542 if (!printk_ratelimit()) 543 return 0; 544 545 if (entry->client_id == SOC15_IH_CLIENTID_VMC) { 546 hub_name = "mmhub0"; 547 hub = &adev->vmhub[AMDGPU_MMHUB_0]; 548 } else if (entry->client_id == SOC15_IH_CLIENTID_VMC1) { 549 hub_name = "mmhub1"; 550 hub = &adev->vmhub[AMDGPU_MMHUB_1]; 551 } else { 552 hub_name = "gfxhub0"; 553 hub = &adev->vmhub[AMDGPU_GFXHUB_0]; 554 } 555 556 memset(&task_info, 0, sizeof(struct amdgpu_task_info)); 557 amdgpu_vm_get_task_info(adev, entry->pasid, &task_info); 558 559 dev_err(adev->dev, 560 "[%s] %s page fault (src_id:%u ring:%u vmid:%u " 561 "pasid:%u, for process %s pid %d thread %s pid %d)\n", 562 hub_name, retry_fault ? "retry" : "no-retry", 563 entry->src_id, entry->ring_id, entry->vmid, 564 entry->pasid, task_info.process_name, task_info.tgid, 565 task_info.task_name, task_info.pid); 566 dev_err(adev->dev, " in page starting at address 0x%016llx from IH client 0x%x (%s)\n", 567 addr, entry->client_id, 568 soc15_ih_clientid_name[entry->client_id]); 569 570 if (amdgpu_sriov_vf(adev)) 571 return 0; 572 573 /* 574 * Issue a dummy read to wait for the status register to 575 * be updated to avoid reading an incorrect value due to 576 * the new fast GRBM interface. 577 */ 578 if ((entry->vmid_src == AMDGPU_GFXHUB_0) && 579 (adev->asic_type < CHIP_ALDEBARAN)) 580 RREG32(hub->vm_l2_pro_fault_status); 581 582 status = RREG32(hub->vm_l2_pro_fault_status); 583 cid = REG_GET_FIELD(status, VM_L2_PROTECTION_FAULT_STATUS, CID); 584 rw = REG_GET_FIELD(status, VM_L2_PROTECTION_FAULT_STATUS, RW); 585 WREG32_P(hub->vm_l2_pro_fault_cntl, 1, ~1); 586 587 588 dev_err(adev->dev, 589 "VM_L2_PROTECTION_FAULT_STATUS:0x%08X\n", 590 status); 591 if (hub == &adev->vmhub[AMDGPU_GFXHUB_0]) { 592 dev_err(adev->dev, "\t Faulty UTCL2 client ID: %s (0x%x)\n", 593 cid >= ARRAY_SIZE(gfxhub_client_ids) ? "unknown" : 594 gfxhub_client_ids[cid], 595 cid); 596 } else { 597 switch (adev->asic_type) { 598 case CHIP_VEGA10: 599 mmhub_cid = mmhub_client_ids_vega10[cid][rw]; 600 break; 601 case CHIP_VEGA12: 602 mmhub_cid = mmhub_client_ids_vega12[cid][rw]; 603 break; 604 case CHIP_VEGA20: 605 mmhub_cid = mmhub_client_ids_vega20[cid][rw]; 606 break; 607 case CHIP_ARCTURUS: 608 mmhub_cid = mmhub_client_ids_arcturus[cid][rw]; 609 break; 610 case CHIP_RAVEN: 611 mmhub_cid = mmhub_client_ids_raven[cid][rw]; 612 break; 613 case CHIP_RENOIR: 614 mmhub_cid = mmhub_client_ids_renoir[cid][rw]; 615 break; 616 case CHIP_ALDEBARAN: 617 mmhub_cid = mmhub_client_ids_aldebaran[cid][rw]; 618 break; 619 default: 620 mmhub_cid = NULL; 621 break; 622 } 623 dev_err(adev->dev, "\t Faulty UTCL2 client ID: %s (0x%x)\n", 624 mmhub_cid ? mmhub_cid : "unknown", cid); 625 } 626 dev_err(adev->dev, "\t MORE_FAULTS: 0x%lx\n", 627 REG_GET_FIELD(status, 628 VM_L2_PROTECTION_FAULT_STATUS, MORE_FAULTS)); 629 dev_err(adev->dev, "\t WALKER_ERROR: 0x%lx\n", 630 REG_GET_FIELD(status, 631 VM_L2_PROTECTION_FAULT_STATUS, WALKER_ERROR)); 632 dev_err(adev->dev, "\t PERMISSION_FAULTS: 0x%lx\n", 633 REG_GET_FIELD(status, 634 VM_L2_PROTECTION_FAULT_STATUS, PERMISSION_FAULTS)); 635 dev_err(adev->dev, "\t MAPPING_ERROR: 0x%lx\n", 636 REG_GET_FIELD(status, 637 VM_L2_PROTECTION_FAULT_STATUS, MAPPING_ERROR)); 638 dev_err(adev->dev, "\t RW: 0x%x\n", rw); 639 return 0; 640 } 641 642 static const struct amdgpu_irq_src_funcs gmc_v9_0_irq_funcs = { 643 .set = gmc_v9_0_vm_fault_interrupt_state, 644 .process = gmc_v9_0_process_interrupt, 645 }; 646 647 648 static const struct amdgpu_irq_src_funcs gmc_v9_0_ecc_funcs = { 649 .set = gmc_v9_0_ecc_interrupt_state, 650 .process = amdgpu_umc_process_ecc_irq, 651 }; 652 653 static void gmc_v9_0_set_irq_funcs(struct amdgpu_device *adev) 654 { 655 adev->gmc.vm_fault.num_types = 1; 656 adev->gmc.vm_fault.funcs = &gmc_v9_0_irq_funcs; 657 658 if (!amdgpu_sriov_vf(adev) && 659 !adev->gmc.xgmi.connected_to_cpu) { 660 adev->gmc.ecc_irq.num_types = 1; 661 adev->gmc.ecc_irq.funcs = &gmc_v9_0_ecc_funcs; 662 } 663 } 664 665 static uint32_t gmc_v9_0_get_invalidate_req(unsigned int vmid, 666 uint32_t flush_type) 667 { 668 u32 req = 0; 669 670 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, 671 PER_VMID_INVALIDATE_REQ, 1 << vmid); 672 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, FLUSH_TYPE, flush_type); 673 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PTES, 1); 674 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE0, 1); 675 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE1, 1); 676 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE2, 1); 677 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L1_PTES, 1); 678 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, 679 CLEAR_PROTECTION_FAULT_STATUS_ADDR, 0); 680 681 return req; 682 } 683 684 /** 685 * gmc_v9_0_use_invalidate_semaphore - judge whether to use semaphore 686 * 687 * @adev: amdgpu_device pointer 688 * @vmhub: vmhub type 689 * 690 */ 691 static bool gmc_v9_0_use_invalidate_semaphore(struct amdgpu_device *adev, 692 uint32_t vmhub) 693 { 694 if (adev->asic_type == CHIP_ALDEBARAN) 695 return false; 696 697 return ((vmhub == AMDGPU_MMHUB_0 || 698 vmhub == AMDGPU_MMHUB_1) && 699 (!amdgpu_sriov_vf(adev)) && 700 (!(!(adev->apu_flags & AMD_APU_IS_RAVEN2) && 701 (adev->apu_flags & AMD_APU_IS_PICASSO)))); 702 } 703 704 static bool gmc_v9_0_get_atc_vmid_pasid_mapping_info(struct amdgpu_device *adev, 705 uint8_t vmid, uint16_t *p_pasid) 706 { 707 uint32_t value; 708 709 value = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING) 710 + vmid); 711 *p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK; 712 713 return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK); 714 } 715 716 /* 717 * GART 718 * VMID 0 is the physical GPU addresses as used by the kernel. 719 * VMIDs 1-15 are used for userspace clients and are handled 720 * by the amdgpu vm/hsa code. 721 */ 722 723 /** 724 * gmc_v9_0_flush_gpu_tlb - tlb flush with certain type 725 * 726 * @adev: amdgpu_device pointer 727 * @vmid: vm instance to flush 728 * @vmhub: which hub to flush 729 * @flush_type: the flush type 730 * 731 * Flush the TLB for the requested page table using certain type. 732 */ 733 static void gmc_v9_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid, 734 uint32_t vmhub, uint32_t flush_type) 735 { 736 bool use_semaphore = gmc_v9_0_use_invalidate_semaphore(adev, vmhub); 737 const unsigned eng = 17; 738 u32 j, inv_req, inv_req2, tmp; 739 struct amdgpu_vmhub *hub; 740 741 BUG_ON(vmhub >= adev->num_vmhubs); 742 743 hub = &adev->vmhub[vmhub]; 744 if (adev->gmc.xgmi.num_physical_nodes && 745 adev->asic_type == CHIP_VEGA20) { 746 /* Vega20+XGMI caches PTEs in TC and TLB. Add a 747 * heavy-weight TLB flush (type 2), which flushes 748 * both. Due to a race condition with concurrent 749 * memory accesses using the same TLB cache line, we 750 * still need a second TLB flush after this. 751 */ 752 inv_req = gmc_v9_0_get_invalidate_req(vmid, 2); 753 inv_req2 = gmc_v9_0_get_invalidate_req(vmid, flush_type); 754 } else { 755 inv_req = gmc_v9_0_get_invalidate_req(vmid, flush_type); 756 inv_req2 = 0; 757 } 758 759 /* This is necessary for a HW workaround under SRIOV as well 760 * as GFXOFF under bare metal 761 */ 762 if (adev->gfx.kiq.ring.sched.ready && 763 (amdgpu_sriov_runtime(adev) || !amdgpu_sriov_vf(adev)) && 764 down_read_trylock(&adev->reset_sem)) { 765 uint32_t req = hub->vm_inv_eng0_req + hub->eng_distance * eng; 766 uint32_t ack = hub->vm_inv_eng0_ack + hub->eng_distance * eng; 767 768 amdgpu_virt_kiq_reg_write_reg_wait(adev, req, ack, inv_req, 769 1 << vmid); 770 up_read(&adev->reset_sem); 771 return; 772 } 773 774 spin_lock(&adev->gmc.invalidate_lock); 775 776 /* 777 * It may lose gpuvm invalidate acknowldege state across power-gating 778 * off cycle, add semaphore acquire before invalidation and semaphore 779 * release after invalidation to avoid entering power gated state 780 * to WA the Issue 781 */ 782 783 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 784 if (use_semaphore) { 785 for (j = 0; j < adev->usec_timeout; j++) { 786 /* a read return value of 1 means semaphore acuqire */ 787 tmp = RREG32_NO_KIQ(hub->vm_inv_eng0_sem + 788 hub->eng_distance * eng); 789 if (tmp & 0x1) 790 break; 791 udelay(1); 792 } 793 794 if (j >= adev->usec_timeout) 795 DRM_ERROR("Timeout waiting for sem acquire in VM flush!\n"); 796 } 797 798 do { 799 WREG32_NO_KIQ(hub->vm_inv_eng0_req + 800 hub->eng_distance * eng, inv_req); 801 802 /* 803 * Issue a dummy read to wait for the ACK register to 804 * be cleared to avoid a false ACK due to the new fast 805 * GRBM interface. 806 */ 807 if ((vmhub == AMDGPU_GFXHUB_0) && 808 (adev->asic_type < CHIP_ALDEBARAN)) 809 RREG32_NO_KIQ(hub->vm_inv_eng0_req + 810 hub->eng_distance * eng); 811 812 for (j = 0; j < adev->usec_timeout; j++) { 813 tmp = RREG32_NO_KIQ(hub->vm_inv_eng0_ack + 814 hub->eng_distance * eng); 815 if (tmp & (1 << vmid)) 816 break; 817 udelay(1); 818 } 819 820 inv_req = inv_req2; 821 inv_req2 = 0; 822 } while (inv_req); 823 824 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 825 if (use_semaphore) 826 /* 827 * add semaphore release after invalidation, 828 * write with 0 means semaphore release 829 */ 830 WREG32_NO_KIQ(hub->vm_inv_eng0_sem + 831 hub->eng_distance * eng, 0); 832 833 spin_unlock(&adev->gmc.invalidate_lock); 834 835 if (j < adev->usec_timeout) 836 return; 837 838 DRM_ERROR("Timeout waiting for VM flush ACK!\n"); 839 } 840 841 /** 842 * gmc_v9_0_flush_gpu_tlb_pasid - tlb flush via pasid 843 * 844 * @adev: amdgpu_device pointer 845 * @pasid: pasid to be flush 846 * @flush_type: the flush type 847 * @all_hub: flush all hubs 848 * 849 * Flush the TLB for the requested pasid. 850 */ 851 static int gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev, 852 uint16_t pasid, uint32_t flush_type, 853 bool all_hub) 854 { 855 int vmid, i; 856 signed long r; 857 uint32_t seq; 858 uint16_t queried_pasid; 859 bool ret; 860 struct amdgpu_ring *ring = &adev->gfx.kiq.ring; 861 struct amdgpu_kiq *kiq = &adev->gfx.kiq; 862 863 if (amdgpu_in_reset(adev)) 864 return -EIO; 865 866 if (ring->sched.ready && down_read_trylock(&adev->reset_sem)) { 867 /* Vega20+XGMI caches PTEs in TC and TLB. Add a 868 * heavy-weight TLB flush (type 2), which flushes 869 * both. Due to a race condition with concurrent 870 * memory accesses using the same TLB cache line, we 871 * still need a second TLB flush after this. 872 */ 873 bool vega20_xgmi_wa = (adev->gmc.xgmi.num_physical_nodes && 874 adev->asic_type == CHIP_VEGA20); 875 /* 2 dwords flush + 8 dwords fence */ 876 unsigned int ndw = kiq->pmf->invalidate_tlbs_size + 8; 877 878 if (vega20_xgmi_wa) 879 ndw += kiq->pmf->invalidate_tlbs_size; 880 881 spin_lock(&adev->gfx.kiq.ring_lock); 882 /* 2 dwords flush + 8 dwords fence */ 883 amdgpu_ring_alloc(ring, ndw); 884 if (vega20_xgmi_wa) 885 kiq->pmf->kiq_invalidate_tlbs(ring, 886 pasid, 2, all_hub); 887 kiq->pmf->kiq_invalidate_tlbs(ring, 888 pasid, flush_type, all_hub); 889 r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT); 890 if (r) { 891 amdgpu_ring_undo(ring); 892 spin_unlock(&adev->gfx.kiq.ring_lock); 893 up_read(&adev->reset_sem); 894 return -ETIME; 895 } 896 897 amdgpu_ring_commit(ring); 898 spin_unlock(&adev->gfx.kiq.ring_lock); 899 r = amdgpu_fence_wait_polling(ring, seq, adev->usec_timeout); 900 if (r < 1) { 901 dev_err(adev->dev, "wait for kiq fence error: %ld.\n", r); 902 up_read(&adev->reset_sem); 903 return -ETIME; 904 } 905 up_read(&adev->reset_sem); 906 return 0; 907 } 908 909 for (vmid = 1; vmid < 16; vmid++) { 910 911 ret = gmc_v9_0_get_atc_vmid_pasid_mapping_info(adev, vmid, 912 &queried_pasid); 913 if (ret && queried_pasid == pasid) { 914 if (all_hub) { 915 for (i = 0; i < adev->num_vmhubs; i++) 916 gmc_v9_0_flush_gpu_tlb(adev, vmid, 917 i, flush_type); 918 } else { 919 gmc_v9_0_flush_gpu_tlb(adev, vmid, 920 AMDGPU_GFXHUB_0, flush_type); 921 } 922 break; 923 } 924 } 925 926 return 0; 927 928 } 929 930 static uint64_t gmc_v9_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring, 931 unsigned vmid, uint64_t pd_addr) 932 { 933 bool use_semaphore = gmc_v9_0_use_invalidate_semaphore(ring->adev, ring->funcs->vmhub); 934 struct amdgpu_device *adev = ring->adev; 935 struct amdgpu_vmhub *hub = &adev->vmhub[ring->funcs->vmhub]; 936 uint32_t req = gmc_v9_0_get_invalidate_req(vmid, 0); 937 unsigned eng = ring->vm_inv_eng; 938 939 /* 940 * It may lose gpuvm invalidate acknowldege state across power-gating 941 * off cycle, add semaphore acquire before invalidation and semaphore 942 * release after invalidation to avoid entering power gated state 943 * to WA the Issue 944 */ 945 946 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 947 if (use_semaphore) 948 /* a read return value of 1 means semaphore acuqire */ 949 amdgpu_ring_emit_reg_wait(ring, 950 hub->vm_inv_eng0_sem + 951 hub->eng_distance * eng, 0x1, 0x1); 952 953 amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_lo32 + 954 (hub->ctx_addr_distance * vmid), 955 lower_32_bits(pd_addr)); 956 957 amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 + 958 (hub->ctx_addr_distance * vmid), 959 upper_32_bits(pd_addr)); 960 961 amdgpu_ring_emit_reg_write_reg_wait(ring, hub->vm_inv_eng0_req + 962 hub->eng_distance * eng, 963 hub->vm_inv_eng0_ack + 964 hub->eng_distance * eng, 965 req, 1 << vmid); 966 967 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 968 if (use_semaphore) 969 /* 970 * add semaphore release after invalidation, 971 * write with 0 means semaphore release 972 */ 973 amdgpu_ring_emit_wreg(ring, hub->vm_inv_eng0_sem + 974 hub->eng_distance * eng, 0); 975 976 return pd_addr; 977 } 978 979 static void gmc_v9_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned vmid, 980 unsigned pasid) 981 { 982 struct amdgpu_device *adev = ring->adev; 983 uint32_t reg; 984 985 /* Do nothing because there's no lut register for mmhub1. */ 986 if (ring->funcs->vmhub == AMDGPU_MMHUB_1) 987 return; 988 989 if (ring->funcs->vmhub == AMDGPU_GFXHUB_0) 990 reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid; 991 else 992 reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT_MM) + vmid; 993 994 amdgpu_ring_emit_wreg(ring, reg, pasid); 995 } 996 997 /* 998 * PTE format on VEGA 10: 999 * 63:59 reserved 1000 * 58:57 mtype 1001 * 56 F 1002 * 55 L 1003 * 54 P 1004 * 53 SW 1005 * 52 T 1006 * 50:48 reserved 1007 * 47:12 4k physical page base address 1008 * 11:7 fragment 1009 * 6 write 1010 * 5 read 1011 * 4 exe 1012 * 3 Z 1013 * 2 snooped 1014 * 1 system 1015 * 0 valid 1016 * 1017 * PDE format on VEGA 10: 1018 * 63:59 block fragment size 1019 * 58:55 reserved 1020 * 54 P 1021 * 53:48 reserved 1022 * 47:6 physical base address of PD or PTE 1023 * 5:3 reserved 1024 * 2 C 1025 * 1 system 1026 * 0 valid 1027 */ 1028 1029 static uint64_t gmc_v9_0_map_mtype(struct amdgpu_device *adev, uint32_t flags) 1030 1031 { 1032 switch (flags) { 1033 case AMDGPU_VM_MTYPE_DEFAULT: 1034 return AMDGPU_PTE_MTYPE_VG10(MTYPE_NC); 1035 case AMDGPU_VM_MTYPE_NC: 1036 return AMDGPU_PTE_MTYPE_VG10(MTYPE_NC); 1037 case AMDGPU_VM_MTYPE_WC: 1038 return AMDGPU_PTE_MTYPE_VG10(MTYPE_WC); 1039 case AMDGPU_VM_MTYPE_RW: 1040 return AMDGPU_PTE_MTYPE_VG10(MTYPE_RW); 1041 case AMDGPU_VM_MTYPE_CC: 1042 return AMDGPU_PTE_MTYPE_VG10(MTYPE_CC); 1043 case AMDGPU_VM_MTYPE_UC: 1044 return AMDGPU_PTE_MTYPE_VG10(MTYPE_UC); 1045 default: 1046 return AMDGPU_PTE_MTYPE_VG10(MTYPE_NC); 1047 } 1048 } 1049 1050 static void gmc_v9_0_get_vm_pde(struct amdgpu_device *adev, int level, 1051 uint64_t *addr, uint64_t *flags) 1052 { 1053 if (!(*flags & AMDGPU_PDE_PTE) && !(*flags & AMDGPU_PTE_SYSTEM)) 1054 *addr = amdgpu_gmc_vram_mc2pa(adev, *addr); 1055 BUG_ON(*addr & 0xFFFF00000000003FULL); 1056 1057 if (!adev->gmc.translate_further) 1058 return; 1059 1060 if (level == AMDGPU_VM_PDB1) { 1061 /* Set the block fragment size */ 1062 if (!(*flags & AMDGPU_PDE_PTE)) 1063 *flags |= AMDGPU_PDE_BFS(0x9); 1064 1065 } else if (level == AMDGPU_VM_PDB0) { 1066 if (*flags & AMDGPU_PDE_PTE) 1067 *flags &= ~AMDGPU_PDE_PTE; 1068 else 1069 *flags |= AMDGPU_PTE_TF; 1070 } 1071 } 1072 1073 static void gmc_v9_0_get_vm_pte(struct amdgpu_device *adev, 1074 struct amdgpu_bo_va_mapping *mapping, 1075 uint64_t *flags) 1076 { 1077 *flags &= ~AMDGPU_PTE_EXECUTABLE; 1078 *flags |= mapping->flags & AMDGPU_PTE_EXECUTABLE; 1079 1080 *flags &= ~AMDGPU_PTE_MTYPE_VG10_MASK; 1081 *flags |= mapping->flags & AMDGPU_PTE_MTYPE_VG10_MASK; 1082 1083 if (mapping->flags & AMDGPU_PTE_PRT) { 1084 *flags |= AMDGPU_PTE_PRT; 1085 *flags &= ~AMDGPU_PTE_VALID; 1086 } 1087 1088 if ((adev->asic_type == CHIP_ARCTURUS || 1089 adev->asic_type == CHIP_ALDEBARAN) && 1090 !(*flags & AMDGPU_PTE_SYSTEM) && 1091 mapping->bo_va->is_xgmi) 1092 *flags |= AMDGPU_PTE_SNOOPED; 1093 1094 if (adev->asic_type == CHIP_ALDEBARAN) 1095 *flags |= mapping->flags & AMDGPU_PTE_SNOOPED; 1096 } 1097 1098 static unsigned gmc_v9_0_get_vbios_fb_size(struct amdgpu_device *adev) 1099 { 1100 u32 d1vga_control = RREG32_SOC15(DCE, 0, mmD1VGA_CONTROL); 1101 unsigned size; 1102 1103 if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) { 1104 size = AMDGPU_VBIOS_VGA_ALLOCATION; 1105 } else { 1106 u32 viewport; 1107 1108 switch (adev->asic_type) { 1109 case CHIP_RAVEN: 1110 case CHIP_RENOIR: 1111 viewport = RREG32_SOC15(DCE, 0, mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION); 1112 size = (REG_GET_FIELD(viewport, 1113 HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_HEIGHT) * 1114 REG_GET_FIELD(viewport, 1115 HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_WIDTH) * 1116 4); 1117 break; 1118 case CHIP_VEGA10: 1119 case CHIP_VEGA12: 1120 case CHIP_VEGA20: 1121 default: 1122 viewport = RREG32_SOC15(DCE, 0, mmSCL0_VIEWPORT_SIZE); 1123 size = (REG_GET_FIELD(viewport, SCL0_VIEWPORT_SIZE, VIEWPORT_HEIGHT) * 1124 REG_GET_FIELD(viewport, SCL0_VIEWPORT_SIZE, VIEWPORT_WIDTH) * 1125 4); 1126 break; 1127 } 1128 } 1129 1130 return size; 1131 } 1132 1133 static const struct amdgpu_gmc_funcs gmc_v9_0_gmc_funcs = { 1134 .flush_gpu_tlb = gmc_v9_0_flush_gpu_tlb, 1135 .flush_gpu_tlb_pasid = gmc_v9_0_flush_gpu_tlb_pasid, 1136 .emit_flush_gpu_tlb = gmc_v9_0_emit_flush_gpu_tlb, 1137 .emit_pasid_mapping = gmc_v9_0_emit_pasid_mapping, 1138 .map_mtype = gmc_v9_0_map_mtype, 1139 .get_vm_pde = gmc_v9_0_get_vm_pde, 1140 .get_vm_pte = gmc_v9_0_get_vm_pte, 1141 .get_vbios_fb_size = gmc_v9_0_get_vbios_fb_size, 1142 }; 1143 1144 static void gmc_v9_0_set_gmc_funcs(struct amdgpu_device *adev) 1145 { 1146 adev->gmc.gmc_funcs = &gmc_v9_0_gmc_funcs; 1147 } 1148 1149 static void gmc_v9_0_set_umc_funcs(struct amdgpu_device *adev) 1150 { 1151 switch (adev->asic_type) { 1152 case CHIP_VEGA10: 1153 adev->umc.funcs = &umc_v6_0_funcs; 1154 break; 1155 case CHIP_VEGA20: 1156 adev->umc.max_ras_err_cnt_per_query = UMC_V6_1_TOTAL_CHANNEL_NUM; 1157 adev->umc.channel_inst_num = UMC_V6_1_CHANNEL_INSTANCE_NUM; 1158 adev->umc.umc_inst_num = UMC_V6_1_UMC_INSTANCE_NUM; 1159 adev->umc.channel_offs = UMC_V6_1_PER_CHANNEL_OFFSET_VG20; 1160 adev->umc.channel_idx_tbl = &umc_v6_1_channel_idx_tbl[0][0]; 1161 adev->umc.ras_funcs = &umc_v6_1_ras_funcs; 1162 break; 1163 case CHIP_ARCTURUS: 1164 adev->umc.max_ras_err_cnt_per_query = UMC_V6_1_TOTAL_CHANNEL_NUM; 1165 adev->umc.channel_inst_num = UMC_V6_1_CHANNEL_INSTANCE_NUM; 1166 adev->umc.umc_inst_num = UMC_V6_1_UMC_INSTANCE_NUM; 1167 adev->umc.channel_offs = UMC_V6_1_PER_CHANNEL_OFFSET_ARCT; 1168 adev->umc.channel_idx_tbl = &umc_v6_1_channel_idx_tbl[0][0]; 1169 adev->umc.ras_funcs = &umc_v6_1_ras_funcs; 1170 break; 1171 default: 1172 break; 1173 } 1174 } 1175 1176 static void gmc_v9_0_set_mmhub_funcs(struct amdgpu_device *adev) 1177 { 1178 switch (adev->asic_type) { 1179 case CHIP_ARCTURUS: 1180 adev->mmhub.funcs = &mmhub_v9_4_funcs; 1181 break; 1182 case CHIP_ALDEBARAN: 1183 adev->mmhub.funcs = &mmhub_v1_7_funcs; 1184 break; 1185 default: 1186 adev->mmhub.funcs = &mmhub_v1_0_funcs; 1187 break; 1188 } 1189 } 1190 1191 static void gmc_v9_0_set_mmhub_ras_funcs(struct amdgpu_device *adev) 1192 { 1193 switch (adev->asic_type) { 1194 case CHIP_VEGA20: 1195 adev->mmhub.ras_funcs = &mmhub_v1_0_ras_funcs; 1196 break; 1197 case CHIP_ARCTURUS: 1198 adev->mmhub.ras_funcs = &mmhub_v9_4_ras_funcs; 1199 break; 1200 case CHIP_ALDEBARAN: 1201 adev->mmhub.ras_funcs = &mmhub_v1_7_ras_funcs; 1202 break; 1203 default: 1204 /* mmhub ras is not available */ 1205 break; 1206 } 1207 } 1208 1209 static void gmc_v9_0_set_gfxhub_funcs(struct amdgpu_device *adev) 1210 { 1211 adev->gfxhub.funcs = &gfxhub_v1_0_funcs; 1212 } 1213 1214 static void gmc_v9_0_set_hdp_ras_funcs(struct amdgpu_device *adev) 1215 { 1216 adev->hdp.ras_funcs = &hdp_v4_0_ras_funcs; 1217 } 1218 1219 static int gmc_v9_0_early_init(void *handle) 1220 { 1221 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1222 1223 if (adev->asic_type == CHIP_VEGA20 || 1224 adev->asic_type == CHIP_ARCTURUS) 1225 adev->gmc.xgmi.supported = true; 1226 1227 if (adev->asic_type == CHIP_ALDEBARAN) { 1228 adev->gmc.xgmi.supported = true; 1229 adev->gmc.xgmi.connected_to_cpu = 1230 adev->smuio.funcs->is_host_gpu_xgmi_supported(adev); 1231 } 1232 1233 gmc_v9_0_set_gmc_funcs(adev); 1234 gmc_v9_0_set_irq_funcs(adev); 1235 gmc_v9_0_set_umc_funcs(adev); 1236 gmc_v9_0_set_mmhub_funcs(adev); 1237 gmc_v9_0_set_mmhub_ras_funcs(adev); 1238 gmc_v9_0_set_gfxhub_funcs(adev); 1239 gmc_v9_0_set_hdp_ras_funcs(adev); 1240 1241 adev->gmc.shared_aperture_start = 0x2000000000000000ULL; 1242 adev->gmc.shared_aperture_end = 1243 adev->gmc.shared_aperture_start + (4ULL << 30) - 1; 1244 adev->gmc.private_aperture_start = 0x1000000000000000ULL; 1245 adev->gmc.private_aperture_end = 1246 adev->gmc.private_aperture_start + (4ULL << 30) - 1; 1247 1248 return 0; 1249 } 1250 1251 static int gmc_v9_0_late_init(void *handle) 1252 { 1253 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1254 int r; 1255 1256 r = amdgpu_gmc_allocate_vm_inv_eng(adev); 1257 if (r) 1258 return r; 1259 1260 /* 1261 * Workaround performance drop issue with VBIOS enables partial 1262 * writes, while disables HBM ECC for vega10. 1263 */ 1264 if (!amdgpu_sriov_vf(adev) && (adev->asic_type == CHIP_VEGA10)) { 1265 if (!(adev->ras_enabled & (1 << AMDGPU_RAS_BLOCK__UMC))) { 1266 if (adev->df.funcs->enable_ecc_force_par_wr_rmw) 1267 adev->df.funcs->enable_ecc_force_par_wr_rmw(adev, false); 1268 } 1269 } 1270 1271 if (!amdgpu_persistent_edc_harvesting_supported(adev)) { 1272 if (adev->mmhub.ras_funcs && 1273 adev->mmhub.ras_funcs->reset_ras_error_count) 1274 adev->mmhub.ras_funcs->reset_ras_error_count(adev); 1275 1276 if (adev->hdp.ras_funcs && 1277 adev->hdp.ras_funcs->reset_ras_error_count) 1278 adev->hdp.ras_funcs->reset_ras_error_count(adev); 1279 } 1280 1281 r = amdgpu_gmc_ras_late_init(adev); 1282 if (r) 1283 return r; 1284 1285 return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0); 1286 } 1287 1288 static void gmc_v9_0_vram_gtt_location(struct amdgpu_device *adev, 1289 struct amdgpu_gmc *mc) 1290 { 1291 u64 base = 0; 1292 1293 if (!amdgpu_sriov_vf(adev)) 1294 base = adev->mmhub.funcs->get_fb_location(adev); 1295 1296 /* add the xgmi offset of the physical node */ 1297 base += adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size; 1298 if (adev->gmc.xgmi.connected_to_cpu) { 1299 amdgpu_gmc_sysvm_location(adev, mc); 1300 } else { 1301 amdgpu_gmc_vram_location(adev, mc, base); 1302 amdgpu_gmc_gart_location(adev, mc); 1303 amdgpu_gmc_agp_location(adev, mc); 1304 } 1305 /* base offset of vram pages */ 1306 adev->vm_manager.vram_base_offset = adev->gfxhub.funcs->get_mc_fb_offset(adev); 1307 1308 /* XXX: add the xgmi offset of the physical node? */ 1309 adev->vm_manager.vram_base_offset += 1310 adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size; 1311 } 1312 1313 /** 1314 * gmc_v9_0_mc_init - initialize the memory controller driver params 1315 * 1316 * @adev: amdgpu_device pointer 1317 * 1318 * Look up the amount of vram, vram width, and decide how to place 1319 * vram and gart within the GPU's physical address space. 1320 * Returns 0 for success. 1321 */ 1322 static int gmc_v9_0_mc_init(struct amdgpu_device *adev) 1323 { 1324 int r; 1325 1326 /* size in MB on si */ 1327 adev->gmc.mc_vram_size = 1328 adev->nbio.funcs->get_memsize(adev) * 1024ULL * 1024ULL; 1329 adev->gmc.real_vram_size = adev->gmc.mc_vram_size; 1330 1331 if (!(adev->flags & AMD_IS_APU) && 1332 !adev->gmc.xgmi.connected_to_cpu) { 1333 r = amdgpu_device_resize_fb_bar(adev); 1334 if (r) 1335 return r; 1336 } 1337 adev->gmc.aper_base = pci_resource_start(adev->pdev, 0); 1338 adev->gmc.aper_size = pci_resource_len(adev->pdev, 0); 1339 1340 #ifdef CONFIG_X86_64 1341 /* 1342 * AMD Accelerated Processing Platform (APP) supporting GPU-HOST xgmi 1343 * interface can use VRAM through here as it appears system reserved 1344 * memory in host address space. 1345 * 1346 * For APUs, VRAM is just the stolen system memory and can be accessed 1347 * directly. 1348 * 1349 * Otherwise, use the legacy Host Data Path (HDP) through PCIe BAR. 1350 */ 1351 1352 /* check whether both host-gpu and gpu-gpu xgmi links exist */ 1353 if ((adev->flags & AMD_IS_APU) || 1354 (adev->gmc.xgmi.supported && 1355 adev->gmc.xgmi.connected_to_cpu)) { 1356 adev->gmc.aper_base = 1357 adev->gfxhub.funcs->get_mc_fb_offset(adev) + 1358 adev->gmc.xgmi.physical_node_id * 1359 adev->gmc.xgmi.node_segment_size; 1360 adev->gmc.aper_size = adev->gmc.real_vram_size; 1361 } 1362 1363 #endif 1364 /* In case the PCI BAR is larger than the actual amount of vram */ 1365 adev->gmc.visible_vram_size = adev->gmc.aper_size; 1366 if (adev->gmc.visible_vram_size > adev->gmc.real_vram_size) 1367 adev->gmc.visible_vram_size = adev->gmc.real_vram_size; 1368 1369 /* set the gart size */ 1370 if (amdgpu_gart_size == -1) { 1371 switch (adev->asic_type) { 1372 case CHIP_VEGA10: /* all engines support GPUVM */ 1373 case CHIP_VEGA12: /* all engines support GPUVM */ 1374 case CHIP_VEGA20: 1375 case CHIP_ARCTURUS: 1376 case CHIP_ALDEBARAN: 1377 default: 1378 adev->gmc.gart_size = 512ULL << 20; 1379 break; 1380 case CHIP_RAVEN: /* DCE SG support */ 1381 case CHIP_RENOIR: 1382 adev->gmc.gart_size = 1024ULL << 20; 1383 break; 1384 } 1385 } else { 1386 adev->gmc.gart_size = (u64)amdgpu_gart_size << 20; 1387 } 1388 1389 adev->gmc.gart_size += adev->pm.smu_prv_buffer_size; 1390 1391 gmc_v9_0_vram_gtt_location(adev, &adev->gmc); 1392 1393 return 0; 1394 } 1395 1396 static int gmc_v9_0_gart_init(struct amdgpu_device *adev) 1397 { 1398 int r; 1399 1400 if (adev->gart.bo) { 1401 WARN(1, "VEGA10 PCIE GART already initialized\n"); 1402 return 0; 1403 } 1404 1405 if (adev->gmc.xgmi.connected_to_cpu) { 1406 adev->gmc.vmid0_page_table_depth = 1; 1407 adev->gmc.vmid0_page_table_block_size = 12; 1408 } else { 1409 adev->gmc.vmid0_page_table_depth = 0; 1410 adev->gmc.vmid0_page_table_block_size = 0; 1411 } 1412 1413 /* Initialize common gart structure */ 1414 r = amdgpu_gart_init(adev); 1415 if (r) 1416 return r; 1417 adev->gart.table_size = adev->gart.num_gpu_pages * 8; 1418 adev->gart.gart_pte_flags = AMDGPU_PTE_MTYPE_VG10(MTYPE_UC) | 1419 AMDGPU_PTE_EXECUTABLE; 1420 1421 r = amdgpu_gart_table_vram_alloc(adev); 1422 if (r) 1423 return r; 1424 1425 if (adev->gmc.xgmi.connected_to_cpu) { 1426 r = amdgpu_gmc_pdb0_alloc(adev); 1427 } 1428 1429 return r; 1430 } 1431 1432 /** 1433 * gmc_v9_0_save_registers - saves regs 1434 * 1435 * @adev: amdgpu_device pointer 1436 * 1437 * This saves potential register values that should be 1438 * restored upon resume 1439 */ 1440 static void gmc_v9_0_save_registers(struct amdgpu_device *adev) 1441 { 1442 if (adev->asic_type == CHIP_RAVEN) 1443 adev->gmc.sdpif_register = RREG32_SOC15(DCE, 0, mmDCHUBBUB_SDPIF_MMIO_CNTRL_0); 1444 } 1445 1446 static int gmc_v9_0_sw_init(void *handle) 1447 { 1448 int r, vram_width = 0, vram_type = 0, vram_vendor = 0; 1449 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1450 1451 adev->gfxhub.funcs->init(adev); 1452 1453 adev->mmhub.funcs->init(adev); 1454 1455 spin_lock_init(&adev->gmc.invalidate_lock); 1456 1457 r = amdgpu_atomfirmware_get_vram_info(adev, 1458 &vram_width, &vram_type, &vram_vendor); 1459 if (amdgpu_sriov_vf(adev)) 1460 /* For Vega10 SR-IOV, vram_width can't be read from ATOM as RAVEN, 1461 * and DF related registers is not readable, seems hardcord is the 1462 * only way to set the correct vram_width 1463 */ 1464 adev->gmc.vram_width = 2048; 1465 else if (amdgpu_emu_mode != 1) 1466 adev->gmc.vram_width = vram_width; 1467 1468 if (!adev->gmc.vram_width) { 1469 int chansize, numchan; 1470 1471 /* hbm memory channel size */ 1472 if (adev->flags & AMD_IS_APU) 1473 chansize = 64; 1474 else 1475 chansize = 128; 1476 1477 numchan = adev->df.funcs->get_hbm_channel_number(adev); 1478 adev->gmc.vram_width = numchan * chansize; 1479 } 1480 1481 adev->gmc.vram_type = vram_type; 1482 adev->gmc.vram_vendor = vram_vendor; 1483 switch (adev->asic_type) { 1484 case CHIP_RAVEN: 1485 adev->num_vmhubs = 2; 1486 1487 if (adev->rev_id == 0x0 || adev->rev_id == 0x1) { 1488 amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48); 1489 } else { 1490 /* vm_size is 128TB + 512GB for legacy 3-level page support */ 1491 amdgpu_vm_adjust_size(adev, 128 * 1024 + 512, 9, 2, 48); 1492 adev->gmc.translate_further = 1493 adev->vm_manager.num_level > 1; 1494 } 1495 break; 1496 case CHIP_VEGA10: 1497 case CHIP_VEGA12: 1498 case CHIP_VEGA20: 1499 case CHIP_RENOIR: 1500 case CHIP_ALDEBARAN: 1501 adev->num_vmhubs = 2; 1502 1503 1504 /* 1505 * To fulfill 4-level page support, 1506 * vm size is 256TB (48bit), maximum size of Vega10, 1507 * block size 512 (9bit) 1508 */ 1509 /* sriov restrict max_pfn below AMDGPU_GMC_HOLE */ 1510 if (amdgpu_sriov_vf(adev)) 1511 amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 47); 1512 else 1513 amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48); 1514 break; 1515 case CHIP_ARCTURUS: 1516 adev->num_vmhubs = 3; 1517 1518 /* Keep the vm size same with Vega20 */ 1519 amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48); 1520 break; 1521 default: 1522 break; 1523 } 1524 1525 /* This interrupt is VMC page fault.*/ 1526 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC, VMC_1_0__SRCID__VM_FAULT, 1527 &adev->gmc.vm_fault); 1528 if (r) 1529 return r; 1530 1531 if (adev->asic_type == CHIP_ARCTURUS) { 1532 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC1, VMC_1_0__SRCID__VM_FAULT, 1533 &adev->gmc.vm_fault); 1534 if (r) 1535 return r; 1536 } 1537 1538 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_UTCL2, UTCL2_1_0__SRCID__FAULT, 1539 &adev->gmc.vm_fault); 1540 1541 if (r) 1542 return r; 1543 1544 if (!amdgpu_sriov_vf(adev) && 1545 !adev->gmc.xgmi.connected_to_cpu) { 1546 /* interrupt sent to DF. */ 1547 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_DF, 0, 1548 &adev->gmc.ecc_irq); 1549 if (r) 1550 return r; 1551 } 1552 1553 /* Set the internal MC address mask 1554 * This is the max address of the GPU's 1555 * internal address space. 1556 */ 1557 adev->gmc.mc_mask = 0xffffffffffffULL; /* 48 bit MC */ 1558 1559 r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(44)); 1560 if (r) { 1561 printk(KERN_WARNING "amdgpu: No suitable DMA available.\n"); 1562 return r; 1563 } 1564 adev->need_swiotlb = drm_need_swiotlb(44); 1565 1566 if (adev->gmc.xgmi.supported) { 1567 r = adev->gfxhub.funcs->get_xgmi_info(adev); 1568 if (r) 1569 return r; 1570 } 1571 1572 r = gmc_v9_0_mc_init(adev); 1573 if (r) 1574 return r; 1575 1576 amdgpu_gmc_get_vbios_allocations(adev); 1577 1578 /* Memory manager */ 1579 r = amdgpu_bo_init(adev); 1580 if (r) 1581 return r; 1582 1583 r = gmc_v9_0_gart_init(adev); 1584 if (r) 1585 return r; 1586 1587 /* 1588 * number of VMs 1589 * VMID 0 is reserved for System 1590 * amdgpu graphics/compute will use VMIDs 1..n-1 1591 * amdkfd will use VMIDs n..15 1592 * 1593 * The first KFD VMID is 8 for GPUs with graphics, 3 for 1594 * compute-only GPUs. On compute-only GPUs that leaves 2 VMIDs 1595 * for video processing. 1596 */ 1597 adev->vm_manager.first_kfd_vmid = 1598 (adev->asic_type == CHIP_ARCTURUS || 1599 adev->asic_type == CHIP_ALDEBARAN) ? 3 : 8; 1600 1601 amdgpu_vm_manager_init(adev); 1602 1603 gmc_v9_0_save_registers(adev); 1604 1605 return 0; 1606 } 1607 1608 static int gmc_v9_0_sw_fini(void *handle) 1609 { 1610 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1611 1612 amdgpu_gmc_ras_fini(adev); 1613 amdgpu_gem_force_release(adev); 1614 amdgpu_vm_manager_fini(adev); 1615 amdgpu_gart_table_vram_free(adev); 1616 amdgpu_bo_unref(&adev->gmc.pdb0_bo); 1617 amdgpu_bo_fini(adev); 1618 amdgpu_gart_fini(adev); 1619 1620 return 0; 1621 } 1622 1623 static void gmc_v9_0_init_golden_registers(struct amdgpu_device *adev) 1624 { 1625 1626 switch (adev->asic_type) { 1627 case CHIP_VEGA10: 1628 if (amdgpu_sriov_vf(adev)) 1629 break; 1630 fallthrough; 1631 case CHIP_VEGA20: 1632 soc15_program_register_sequence(adev, 1633 golden_settings_mmhub_1_0_0, 1634 ARRAY_SIZE(golden_settings_mmhub_1_0_0)); 1635 soc15_program_register_sequence(adev, 1636 golden_settings_athub_1_0_0, 1637 ARRAY_SIZE(golden_settings_athub_1_0_0)); 1638 break; 1639 case CHIP_VEGA12: 1640 break; 1641 case CHIP_RAVEN: 1642 /* TODO for renoir */ 1643 soc15_program_register_sequence(adev, 1644 golden_settings_athub_1_0_0, 1645 ARRAY_SIZE(golden_settings_athub_1_0_0)); 1646 break; 1647 default: 1648 break; 1649 } 1650 } 1651 1652 /** 1653 * gmc_v9_0_restore_registers - restores regs 1654 * 1655 * @adev: amdgpu_device pointer 1656 * 1657 * This restores register values, saved at suspend. 1658 */ 1659 void gmc_v9_0_restore_registers(struct amdgpu_device *adev) 1660 { 1661 if (adev->asic_type == CHIP_RAVEN) { 1662 WREG32_SOC15(DCE, 0, mmDCHUBBUB_SDPIF_MMIO_CNTRL_0, adev->gmc.sdpif_register); 1663 WARN_ON(adev->gmc.sdpif_register != 1664 RREG32_SOC15(DCE, 0, mmDCHUBBUB_SDPIF_MMIO_CNTRL_0)); 1665 } 1666 } 1667 1668 /** 1669 * gmc_v9_0_gart_enable - gart enable 1670 * 1671 * @adev: amdgpu_device pointer 1672 */ 1673 static int gmc_v9_0_gart_enable(struct amdgpu_device *adev) 1674 { 1675 int r; 1676 1677 if (adev->gmc.xgmi.connected_to_cpu) 1678 amdgpu_gmc_init_pdb0(adev); 1679 1680 if (adev->gart.bo == NULL) { 1681 dev_err(adev->dev, "No VRAM object for PCIE GART.\n"); 1682 return -EINVAL; 1683 } 1684 1685 r = amdgpu_gart_table_vram_pin(adev); 1686 if (r) 1687 return r; 1688 1689 r = adev->gfxhub.funcs->gart_enable(adev); 1690 if (r) 1691 return r; 1692 1693 r = adev->mmhub.funcs->gart_enable(adev); 1694 if (r) 1695 return r; 1696 1697 DRM_INFO("PCIE GART of %uM enabled.\n", 1698 (unsigned)(adev->gmc.gart_size >> 20)); 1699 if (adev->gmc.pdb0_bo) 1700 DRM_INFO("PDB0 located at 0x%016llX\n", 1701 (unsigned long long)amdgpu_bo_gpu_offset(adev->gmc.pdb0_bo)); 1702 DRM_INFO("PTB located at 0x%016llX\n", 1703 (unsigned long long)amdgpu_bo_gpu_offset(adev->gart.bo)); 1704 1705 adev->gart.ready = true; 1706 return 0; 1707 } 1708 1709 static int gmc_v9_0_hw_init(void *handle) 1710 { 1711 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1712 bool value; 1713 int r, i; 1714 1715 /* The sequence of these two function calls matters.*/ 1716 gmc_v9_0_init_golden_registers(adev); 1717 1718 if (adev->mode_info.num_crtc) { 1719 /* Lockout access through VGA aperture*/ 1720 WREG32_FIELD15(DCE, 0, VGA_HDP_CONTROL, VGA_MEMORY_DISABLE, 1); 1721 /* disable VGA render */ 1722 WREG32_FIELD15(DCE, 0, VGA_RENDER_CONTROL, VGA_VSTATUS_CNTL, 0); 1723 } 1724 1725 if (adev->mmhub.funcs->update_power_gating) 1726 adev->mmhub.funcs->update_power_gating(adev, true); 1727 1728 adev->hdp.funcs->init_registers(adev); 1729 1730 /* After HDP is initialized, flush HDP.*/ 1731 adev->hdp.funcs->flush_hdp(adev, NULL); 1732 1733 if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS) 1734 value = false; 1735 else 1736 value = true; 1737 1738 if (!amdgpu_sriov_vf(adev)) { 1739 adev->gfxhub.funcs->set_fault_enable_default(adev, value); 1740 adev->mmhub.funcs->set_fault_enable_default(adev, value); 1741 } 1742 for (i = 0; i < adev->num_vmhubs; ++i) 1743 gmc_v9_0_flush_gpu_tlb(adev, 0, i, 0); 1744 1745 if (adev->umc.funcs && adev->umc.funcs->init_registers) 1746 adev->umc.funcs->init_registers(adev); 1747 1748 r = gmc_v9_0_gart_enable(adev); 1749 1750 return r; 1751 } 1752 1753 /** 1754 * gmc_v9_0_gart_disable - gart disable 1755 * 1756 * @adev: amdgpu_device pointer 1757 * 1758 * This disables all VM page table. 1759 */ 1760 static void gmc_v9_0_gart_disable(struct amdgpu_device *adev) 1761 { 1762 adev->gfxhub.funcs->gart_disable(adev); 1763 adev->mmhub.funcs->gart_disable(adev); 1764 amdgpu_gart_table_vram_unpin(adev); 1765 } 1766 1767 static int gmc_v9_0_hw_fini(void *handle) 1768 { 1769 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1770 1771 if (amdgpu_sriov_vf(adev)) { 1772 /* full access mode, so don't touch any GMC register */ 1773 DRM_DEBUG("For SRIOV client, shouldn't do anything.\n"); 1774 return 0; 1775 } 1776 1777 amdgpu_irq_put(adev, &adev->gmc.ecc_irq, 0); 1778 amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0); 1779 gmc_v9_0_gart_disable(adev); 1780 1781 return 0; 1782 } 1783 1784 static int gmc_v9_0_suspend(void *handle) 1785 { 1786 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1787 1788 return gmc_v9_0_hw_fini(adev); 1789 } 1790 1791 static int gmc_v9_0_resume(void *handle) 1792 { 1793 int r; 1794 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1795 1796 r = gmc_v9_0_hw_init(adev); 1797 if (r) 1798 return r; 1799 1800 amdgpu_vmid_reset_all(adev); 1801 1802 return 0; 1803 } 1804 1805 static bool gmc_v9_0_is_idle(void *handle) 1806 { 1807 /* MC is always ready in GMC v9.*/ 1808 return true; 1809 } 1810 1811 static int gmc_v9_0_wait_for_idle(void *handle) 1812 { 1813 /* There is no need to wait for MC idle in GMC v9.*/ 1814 return 0; 1815 } 1816 1817 static int gmc_v9_0_soft_reset(void *handle) 1818 { 1819 /* XXX for emulation.*/ 1820 return 0; 1821 } 1822 1823 static int gmc_v9_0_set_clockgating_state(void *handle, 1824 enum amd_clockgating_state state) 1825 { 1826 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1827 1828 adev->mmhub.funcs->set_clockgating(adev, state); 1829 1830 athub_v1_0_set_clockgating(adev, state); 1831 1832 return 0; 1833 } 1834 1835 static void gmc_v9_0_get_clockgating_state(void *handle, u32 *flags) 1836 { 1837 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1838 1839 adev->mmhub.funcs->get_clockgating(adev, flags); 1840 1841 athub_v1_0_get_clockgating(adev, flags); 1842 } 1843 1844 static int gmc_v9_0_set_powergating_state(void *handle, 1845 enum amd_powergating_state state) 1846 { 1847 return 0; 1848 } 1849 1850 const struct amd_ip_funcs gmc_v9_0_ip_funcs = { 1851 .name = "gmc_v9_0", 1852 .early_init = gmc_v9_0_early_init, 1853 .late_init = gmc_v9_0_late_init, 1854 .sw_init = gmc_v9_0_sw_init, 1855 .sw_fini = gmc_v9_0_sw_fini, 1856 .hw_init = gmc_v9_0_hw_init, 1857 .hw_fini = gmc_v9_0_hw_fini, 1858 .suspend = gmc_v9_0_suspend, 1859 .resume = gmc_v9_0_resume, 1860 .is_idle = gmc_v9_0_is_idle, 1861 .wait_for_idle = gmc_v9_0_wait_for_idle, 1862 .soft_reset = gmc_v9_0_soft_reset, 1863 .set_clockgating_state = gmc_v9_0_set_clockgating_state, 1864 .set_powergating_state = gmc_v9_0_set_powergating_state, 1865 .get_clockgating_state = gmc_v9_0_get_clockgating_state, 1866 }; 1867 1868 const struct amdgpu_ip_block_version gmc_v9_0_ip_block = 1869 { 1870 .type = AMD_IP_BLOCK_TYPE_GMC, 1871 .major = 9, 1872 .minor = 0, 1873 .rev = 0, 1874 .funcs = &gmc_v9_0_ip_funcs, 1875 }; 1876