1 /* 2 * Copyright 2018 Advanced Micro Devices, Inc. 3 * All Rights Reserved. 4 * 5 * Permission is hereby granted, free of charge, to any person obtaining a 6 * copy of this software and associated documentation files (the 7 * "Software"), to deal in the Software without restriction, including 8 * without limitation the rights to use, copy, modify, merge, publish, 9 * distribute, sub license, and/or sell copies of the Software, and to 10 * permit persons to whom the Software is furnished to do so, subject to 11 * the following conditions: 12 * 13 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 14 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 15 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL 16 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, 17 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR 18 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE 19 * USE OR OTHER DEALINGS IN THE SOFTWARE. 20 * 21 * The above copyright notice and this permission notice (including the 22 * next paragraph) shall be included in all copies or substantial portions 23 * of the Software. 24 * 25 */ 26 27 #include <linux/io-64-nonatomic-lo-hi.h> 28 29 #include "amdgpu.h" 30 #include "amdgpu_gmc.h" 31 #include "amdgpu_ras.h" 32 #include "amdgpu_xgmi.h" 33 34 #include <drm/drm_drv.h> 35 36 /** 37 * amdgpu_gmc_pdb0_alloc - allocate vram for pdb0 38 * 39 * @adev: amdgpu_device pointer 40 * 41 * Allocate video memory for pdb0 and map it for CPU access 42 * Returns 0 for success, error for failure. 43 */ 44 int amdgpu_gmc_pdb0_alloc(struct amdgpu_device *adev) 45 { 46 int r; 47 struct amdgpu_bo_param bp; 48 u64 vram_size = adev->gmc.xgmi.node_segment_size * adev->gmc.xgmi.num_physical_nodes; 49 uint32_t pde0_page_shift = adev->gmc.vmid0_page_table_block_size + 21; 50 uint32_t npdes = (vram_size + (1ULL << pde0_page_shift) -1) >> pde0_page_shift; 51 52 memset(&bp, 0, sizeof(bp)); 53 bp.size = PAGE_ALIGN((npdes + 1) * 8); 54 bp.byte_align = PAGE_SIZE; 55 bp.domain = AMDGPU_GEM_DOMAIN_VRAM; 56 bp.flags = AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED | 57 AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS; 58 bp.type = ttm_bo_type_kernel; 59 bp.resv = NULL; 60 bp.bo_ptr_size = sizeof(struct amdgpu_bo); 61 62 r = amdgpu_bo_create(adev, &bp, &adev->gmc.pdb0_bo); 63 if (r) 64 return r; 65 66 r = amdgpu_bo_reserve(adev->gmc.pdb0_bo, false); 67 if (unlikely(r != 0)) 68 goto bo_reserve_failure; 69 70 r = amdgpu_bo_pin(adev->gmc.pdb0_bo, AMDGPU_GEM_DOMAIN_VRAM); 71 if (r) 72 goto bo_pin_failure; 73 r = amdgpu_bo_kmap(adev->gmc.pdb0_bo, &adev->gmc.ptr_pdb0); 74 if (r) 75 goto bo_kmap_failure; 76 77 amdgpu_bo_unreserve(adev->gmc.pdb0_bo); 78 return 0; 79 80 bo_kmap_failure: 81 amdgpu_bo_unpin(adev->gmc.pdb0_bo); 82 bo_pin_failure: 83 amdgpu_bo_unreserve(adev->gmc.pdb0_bo); 84 bo_reserve_failure: 85 amdgpu_bo_unref(&adev->gmc.pdb0_bo); 86 return r; 87 } 88 89 /** 90 * amdgpu_gmc_get_pde_for_bo - get the PDE for a BO 91 * 92 * @bo: the BO to get the PDE for 93 * @level: the level in the PD hirarchy 94 * @addr: resulting addr 95 * @flags: resulting flags 96 * 97 * Get the address and flags to be used for a PDE (Page Directory Entry). 98 */ 99 void amdgpu_gmc_get_pde_for_bo(struct amdgpu_bo *bo, int level, 100 uint64_t *addr, uint64_t *flags) 101 { 102 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); 103 104 switch (bo->tbo.resource->mem_type) { 105 case TTM_PL_TT: 106 *addr = bo->tbo.ttm->dma_address[0]; 107 break; 108 case TTM_PL_VRAM: 109 *addr = amdgpu_bo_gpu_offset(bo); 110 break; 111 default: 112 *addr = 0; 113 break; 114 } 115 *flags = amdgpu_ttm_tt_pde_flags(bo->tbo.ttm, bo->tbo.resource); 116 amdgpu_gmc_get_vm_pde(adev, level, addr, flags); 117 } 118 119 /* 120 * amdgpu_gmc_pd_addr - return the address of the root directory 121 */ 122 uint64_t amdgpu_gmc_pd_addr(struct amdgpu_bo *bo) 123 { 124 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); 125 uint64_t pd_addr; 126 127 /* TODO: move that into ASIC specific code */ 128 if (adev->asic_type >= CHIP_VEGA10) { 129 uint64_t flags = AMDGPU_PTE_VALID; 130 131 amdgpu_gmc_get_pde_for_bo(bo, -1, &pd_addr, &flags); 132 pd_addr |= flags; 133 } else { 134 pd_addr = amdgpu_bo_gpu_offset(bo); 135 } 136 return pd_addr; 137 } 138 139 /** 140 * amdgpu_gmc_set_pte_pde - update the page tables using CPU 141 * 142 * @adev: amdgpu_device pointer 143 * @cpu_pt_addr: cpu address of the page table 144 * @gpu_page_idx: entry in the page table to update 145 * @addr: dst addr to write into pte/pde 146 * @flags: access flags 147 * 148 * Update the page tables using CPU. 149 */ 150 int amdgpu_gmc_set_pte_pde(struct amdgpu_device *adev, void *cpu_pt_addr, 151 uint32_t gpu_page_idx, uint64_t addr, 152 uint64_t flags) 153 { 154 void __iomem *ptr = (void *)cpu_pt_addr; 155 uint64_t value; 156 int idx; 157 158 if (!drm_dev_enter(&adev->ddev, &idx)) 159 return 0; 160 161 /* 162 * The following is for PTE only. GART does not have PDEs. 163 */ 164 value = addr & 0x0000FFFFFFFFF000ULL; 165 value |= flags; 166 writeq(value, ptr + (gpu_page_idx * 8)); 167 168 drm_dev_exit(idx); 169 170 return 0; 171 } 172 173 /** 174 * amdgpu_gmc_agp_addr - return the address in the AGP address space 175 * 176 * @bo: TTM BO which needs the address, must be in GTT domain 177 * 178 * Tries to figure out how to access the BO through the AGP aperture. Returns 179 * AMDGPU_BO_INVALID_OFFSET if that is not possible. 180 */ 181 uint64_t amdgpu_gmc_agp_addr(struct ttm_buffer_object *bo) 182 { 183 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev); 184 185 if (bo->ttm->num_pages != 1 || bo->ttm->caching == ttm_cached) 186 return AMDGPU_BO_INVALID_OFFSET; 187 188 if (bo->ttm->dma_address[0] + PAGE_SIZE >= adev->gmc.agp_size) 189 return AMDGPU_BO_INVALID_OFFSET; 190 191 return adev->gmc.agp_start + bo->ttm->dma_address[0]; 192 } 193 194 /** 195 * amdgpu_gmc_vram_location - try to find VRAM location 196 * 197 * @adev: amdgpu device structure holding all necessary information 198 * @mc: memory controller structure holding memory information 199 * @base: base address at which to put VRAM 200 * 201 * Function will try to place VRAM at base address provided 202 * as parameter. 203 */ 204 void amdgpu_gmc_vram_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc, 205 u64 base) 206 { 207 uint64_t limit = (uint64_t)amdgpu_vram_limit << 20; 208 209 mc->vram_start = base; 210 mc->vram_end = mc->vram_start + mc->mc_vram_size - 1; 211 if (limit && limit < mc->real_vram_size) 212 mc->real_vram_size = limit; 213 214 if (mc->xgmi.num_physical_nodes == 0) { 215 mc->fb_start = mc->vram_start; 216 mc->fb_end = mc->vram_end; 217 } 218 dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n", 219 mc->mc_vram_size >> 20, mc->vram_start, 220 mc->vram_end, mc->real_vram_size >> 20); 221 } 222 223 /** amdgpu_gmc_sysvm_location - place vram and gart in sysvm aperture 224 * 225 * @adev: amdgpu device structure holding all necessary information 226 * @mc: memory controller structure holding memory information 227 * 228 * This function is only used if use GART for FB translation. In such 229 * case, we use sysvm aperture (vmid0 page tables) for both vram 230 * and gart (aka system memory) access. 231 * 232 * GPUVM (and our organization of vmid0 page tables) require sysvm 233 * aperture to be placed at a location aligned with 8 times of native 234 * page size. For example, if vm_context0_cntl.page_table_block_size 235 * is 12, then native page size is 8G (2M*2^12), sysvm should start 236 * with a 64G aligned address. For simplicity, we just put sysvm at 237 * address 0. So vram start at address 0 and gart is right after vram. 238 */ 239 void amdgpu_gmc_sysvm_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc) 240 { 241 u64 hive_vram_start = 0; 242 u64 hive_vram_end = mc->xgmi.node_segment_size * mc->xgmi.num_physical_nodes - 1; 243 mc->vram_start = mc->xgmi.node_segment_size * mc->xgmi.physical_node_id; 244 mc->vram_end = mc->vram_start + mc->xgmi.node_segment_size - 1; 245 mc->gart_start = hive_vram_end + 1; 246 mc->gart_end = mc->gart_start + mc->gart_size - 1; 247 mc->fb_start = hive_vram_start; 248 mc->fb_end = hive_vram_end; 249 dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n", 250 mc->mc_vram_size >> 20, mc->vram_start, 251 mc->vram_end, mc->real_vram_size >> 20); 252 dev_info(adev->dev, "GART: %lluM 0x%016llX - 0x%016llX\n", 253 mc->gart_size >> 20, mc->gart_start, mc->gart_end); 254 } 255 256 /** 257 * amdgpu_gmc_gart_location - try to find GART location 258 * 259 * @adev: amdgpu device structure holding all necessary information 260 * @mc: memory controller structure holding memory information 261 * 262 * Function will place try to place GART before or after VRAM. 263 * If GART size is bigger than space left then we ajust GART size. 264 * Thus function will never fails. 265 */ 266 void amdgpu_gmc_gart_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc) 267 { 268 const uint64_t four_gb = 0x100000000ULL; 269 u64 size_af, size_bf; 270 /*To avoid the hole, limit the max mc address to AMDGPU_GMC_HOLE_START*/ 271 u64 max_mc_address = min(adev->gmc.mc_mask, AMDGPU_GMC_HOLE_START - 1); 272 273 /* VCE doesn't like it when BOs cross a 4GB segment, so align 274 * the GART base on a 4GB boundary as well. 275 */ 276 size_bf = mc->fb_start; 277 size_af = max_mc_address + 1 - ALIGN(mc->fb_end + 1, four_gb); 278 279 if (mc->gart_size > max(size_bf, size_af)) { 280 dev_warn(adev->dev, "limiting GART\n"); 281 mc->gart_size = max(size_bf, size_af); 282 } 283 284 if ((size_bf >= mc->gart_size && size_bf < size_af) || 285 (size_af < mc->gart_size)) 286 mc->gart_start = 0; 287 else 288 mc->gart_start = max_mc_address - mc->gart_size + 1; 289 290 mc->gart_start &= ~(four_gb - 1); 291 mc->gart_end = mc->gart_start + mc->gart_size - 1; 292 dev_info(adev->dev, "GART: %lluM 0x%016llX - 0x%016llX\n", 293 mc->gart_size >> 20, mc->gart_start, mc->gart_end); 294 } 295 296 /** 297 * amdgpu_gmc_agp_location - try to find AGP location 298 * @adev: amdgpu device structure holding all necessary information 299 * @mc: memory controller structure holding memory information 300 * 301 * Function will place try to find a place for the AGP BAR in the MC address 302 * space. 303 * 304 * AGP BAR will be assigned the largest available hole in the address space. 305 * Should be called after VRAM and GART locations are setup. 306 */ 307 void amdgpu_gmc_agp_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc) 308 { 309 const uint64_t sixteen_gb = 1ULL << 34; 310 const uint64_t sixteen_gb_mask = ~(sixteen_gb - 1); 311 u64 size_af, size_bf; 312 313 if (amdgpu_sriov_vf(adev)) { 314 mc->agp_start = 0xffffffffffff; 315 mc->agp_end = 0x0; 316 mc->agp_size = 0; 317 318 return; 319 } 320 321 if (mc->fb_start > mc->gart_start) { 322 size_bf = (mc->fb_start & sixteen_gb_mask) - 323 ALIGN(mc->gart_end + 1, sixteen_gb); 324 size_af = mc->mc_mask + 1 - ALIGN(mc->fb_end + 1, sixteen_gb); 325 } else { 326 size_bf = mc->fb_start & sixteen_gb_mask; 327 size_af = (mc->gart_start & sixteen_gb_mask) - 328 ALIGN(mc->fb_end + 1, sixteen_gb); 329 } 330 331 if (size_bf > size_af) { 332 mc->agp_start = (mc->fb_start - size_bf) & sixteen_gb_mask; 333 mc->agp_size = size_bf; 334 } else { 335 mc->agp_start = ALIGN(mc->fb_end + 1, sixteen_gb); 336 mc->agp_size = size_af; 337 } 338 339 mc->agp_end = mc->agp_start + mc->agp_size - 1; 340 dev_info(adev->dev, "AGP: %lluM 0x%016llX - 0x%016llX\n", 341 mc->agp_size >> 20, mc->agp_start, mc->agp_end); 342 } 343 344 /** 345 * amdgpu_gmc_fault_key - get hask key from vm fault address and pasid 346 * 347 * @addr: 48 bit physical address, page aligned (36 significant bits) 348 * @pasid: 16 bit process address space identifier 349 */ 350 static inline uint64_t amdgpu_gmc_fault_key(uint64_t addr, uint16_t pasid) 351 { 352 return addr << 4 | pasid; 353 } 354 355 /** 356 * amdgpu_gmc_filter_faults - filter VM faults 357 * 358 * @adev: amdgpu device structure 359 * @addr: address of the VM fault 360 * @pasid: PASID of the process causing the fault 361 * @timestamp: timestamp of the fault 362 * 363 * Returns: 364 * True if the fault was filtered and should not be processed further. 365 * False if the fault is a new one and needs to be handled. 366 */ 367 bool amdgpu_gmc_filter_faults(struct amdgpu_device *adev, uint64_t addr, 368 uint16_t pasid, uint64_t timestamp) 369 { 370 struct amdgpu_gmc *gmc = &adev->gmc; 371 uint64_t stamp, key = amdgpu_gmc_fault_key(addr, pasid); 372 struct amdgpu_gmc_fault *fault; 373 uint32_t hash; 374 375 /* If we don't have space left in the ring buffer return immediately */ 376 stamp = max(timestamp, AMDGPU_GMC_FAULT_TIMEOUT + 1) - 377 AMDGPU_GMC_FAULT_TIMEOUT; 378 if (gmc->fault_ring[gmc->last_fault].timestamp >= stamp) 379 return true; 380 381 /* Try to find the fault in the hash */ 382 hash = hash_64(key, AMDGPU_GMC_FAULT_HASH_ORDER); 383 fault = &gmc->fault_ring[gmc->fault_hash[hash].idx]; 384 while (fault->timestamp >= stamp) { 385 uint64_t tmp; 386 387 if (atomic64_read(&fault->key) == key) 388 return true; 389 390 tmp = fault->timestamp; 391 fault = &gmc->fault_ring[fault->next]; 392 393 /* Check if the entry was reused */ 394 if (fault->timestamp >= tmp) 395 break; 396 } 397 398 /* Add the fault to the ring */ 399 fault = &gmc->fault_ring[gmc->last_fault]; 400 atomic64_set(&fault->key, key); 401 fault->timestamp = timestamp; 402 403 /* And update the hash */ 404 fault->next = gmc->fault_hash[hash].idx; 405 gmc->fault_hash[hash].idx = gmc->last_fault++; 406 return false; 407 } 408 409 /** 410 * amdgpu_gmc_filter_faults_remove - remove address from VM faults filter 411 * 412 * @adev: amdgpu device structure 413 * @addr: address of the VM fault 414 * @pasid: PASID of the process causing the fault 415 * 416 * Remove the address from fault filter, then future vm fault on this address 417 * will pass to retry fault handler to recover. 418 */ 419 void amdgpu_gmc_filter_faults_remove(struct amdgpu_device *adev, uint64_t addr, 420 uint16_t pasid) 421 { 422 struct amdgpu_gmc *gmc = &adev->gmc; 423 uint64_t key = amdgpu_gmc_fault_key(addr, pasid); 424 struct amdgpu_gmc_fault *fault; 425 uint32_t hash; 426 uint64_t tmp; 427 428 hash = hash_64(key, AMDGPU_GMC_FAULT_HASH_ORDER); 429 fault = &gmc->fault_ring[gmc->fault_hash[hash].idx]; 430 do { 431 if (atomic64_cmpxchg(&fault->key, key, 0) == key) 432 break; 433 434 tmp = fault->timestamp; 435 fault = &gmc->fault_ring[fault->next]; 436 } while (fault->timestamp < tmp); 437 } 438 439 int amdgpu_gmc_ras_late_init(struct amdgpu_device *adev) 440 { 441 int r; 442 443 if (adev->umc.ras_funcs && 444 adev->umc.ras_funcs->ras_late_init) { 445 r = adev->umc.ras_funcs->ras_late_init(adev); 446 if (r) 447 return r; 448 } 449 450 if (adev->mmhub.ras_funcs && 451 adev->mmhub.ras_funcs->ras_late_init) { 452 r = adev->mmhub.ras_funcs->ras_late_init(adev); 453 if (r) 454 return r; 455 } 456 457 if (!adev->gmc.xgmi.connected_to_cpu) 458 adev->gmc.xgmi.ras_funcs = &xgmi_ras_funcs; 459 460 if (adev->gmc.xgmi.ras_funcs && 461 adev->gmc.xgmi.ras_funcs->ras_late_init) { 462 r = adev->gmc.xgmi.ras_funcs->ras_late_init(adev); 463 if (r) 464 return r; 465 } 466 467 if (adev->hdp.ras_funcs && 468 adev->hdp.ras_funcs->ras_late_init) { 469 r = adev->hdp.ras_funcs->ras_late_init(adev); 470 if (r) 471 return r; 472 } 473 474 return 0; 475 } 476 477 void amdgpu_gmc_ras_fini(struct amdgpu_device *adev) 478 { 479 if (adev->umc.ras_funcs && 480 adev->umc.ras_funcs->ras_fini) 481 adev->umc.ras_funcs->ras_fini(adev); 482 483 if (adev->mmhub.ras_funcs && 484 adev->mmhub.ras_funcs->ras_fini) 485 adev->mmhub.ras_funcs->ras_fini(adev); 486 487 if (adev->gmc.xgmi.ras_funcs && 488 adev->gmc.xgmi.ras_funcs->ras_fini) 489 adev->gmc.xgmi.ras_funcs->ras_fini(adev); 490 491 if (adev->hdp.ras_funcs && 492 adev->hdp.ras_funcs->ras_fini) 493 adev->hdp.ras_funcs->ras_fini(adev); 494 } 495 496 /* 497 * The latest engine allocation on gfx9/10 is: 498 * Engine 2, 3: firmware 499 * Engine 0, 1, 4~16: amdgpu ring, 500 * subject to change when ring number changes 501 * Engine 17: Gart flushes 502 */ 503 #define GFXHUB_FREE_VM_INV_ENGS_BITMAP 0x1FFF3 504 #define MMHUB_FREE_VM_INV_ENGS_BITMAP 0x1FFF3 505 506 int amdgpu_gmc_allocate_vm_inv_eng(struct amdgpu_device *adev) 507 { 508 struct amdgpu_ring *ring; 509 unsigned vm_inv_engs[AMDGPU_MAX_VMHUBS] = 510 {GFXHUB_FREE_VM_INV_ENGS_BITMAP, MMHUB_FREE_VM_INV_ENGS_BITMAP, 511 GFXHUB_FREE_VM_INV_ENGS_BITMAP}; 512 unsigned i; 513 unsigned vmhub, inv_eng; 514 515 for (i = 0; i < adev->num_rings; ++i) { 516 ring = adev->rings[i]; 517 vmhub = ring->funcs->vmhub; 518 519 if (ring == &adev->mes.ring) 520 continue; 521 522 inv_eng = ffs(vm_inv_engs[vmhub]); 523 if (!inv_eng) { 524 dev_err(adev->dev, "no VM inv eng for ring %s\n", 525 ring->name); 526 return -EINVAL; 527 } 528 529 ring->vm_inv_eng = inv_eng - 1; 530 vm_inv_engs[vmhub] &= ~(1 << ring->vm_inv_eng); 531 532 dev_info(adev->dev, "ring %s uses VM inv eng %u on hub %u\n", 533 ring->name, ring->vm_inv_eng, ring->funcs->vmhub); 534 } 535 536 return 0; 537 } 538 539 /** 540 * amdgpu_gmc_tmz_set -- check and set if a device supports TMZ 541 * @adev: amdgpu_device pointer 542 * 543 * Check and set if an the device @adev supports Trusted Memory 544 * Zones (TMZ). 545 */ 546 void amdgpu_gmc_tmz_set(struct amdgpu_device *adev) 547 { 548 switch (adev->asic_type) { 549 case CHIP_RAVEN: 550 case CHIP_RENOIR: 551 if (amdgpu_tmz == 0) { 552 adev->gmc.tmz_enabled = false; 553 dev_info(adev->dev, 554 "Trusted Memory Zone (TMZ) feature disabled (cmd line)\n"); 555 } else { 556 adev->gmc.tmz_enabled = true; 557 dev_info(adev->dev, 558 "Trusted Memory Zone (TMZ) feature enabled\n"); 559 } 560 break; 561 case CHIP_NAVI10: 562 case CHIP_NAVI14: 563 case CHIP_NAVI12: 564 case CHIP_VANGOGH: 565 case CHIP_YELLOW_CARP: 566 /* Don't enable it by default yet. 567 */ 568 if (amdgpu_tmz < 1) { 569 adev->gmc.tmz_enabled = false; 570 dev_info(adev->dev, 571 "Trusted Memory Zone (TMZ) feature disabled as experimental (default)\n"); 572 } else { 573 adev->gmc.tmz_enabled = true; 574 dev_info(adev->dev, 575 "Trusted Memory Zone (TMZ) feature enabled as experimental (cmd line)\n"); 576 } 577 break; 578 default: 579 adev->gmc.tmz_enabled = false; 580 dev_warn(adev->dev, 581 "Trusted Memory Zone (TMZ) feature not supported\n"); 582 break; 583 } 584 } 585 586 /** 587 * amdgpu_gmc_noretry_set -- set per asic noretry defaults 588 * @adev: amdgpu_device pointer 589 * 590 * Set a per asic default for the no-retry parameter. 591 * 592 */ 593 void amdgpu_gmc_noretry_set(struct amdgpu_device *adev) 594 { 595 struct amdgpu_gmc *gmc = &adev->gmc; 596 597 switch (adev->asic_type) { 598 case CHIP_VEGA10: 599 case CHIP_VEGA20: 600 case CHIP_ARCTURUS: 601 case CHIP_ALDEBARAN: 602 /* 603 * noretry = 0 will cause kfd page fault tests fail 604 * for some ASICs, so set default to 1 for these ASICs. 605 */ 606 if (amdgpu_noretry == -1) 607 gmc->noretry = 1; 608 else 609 gmc->noretry = amdgpu_noretry; 610 break; 611 case CHIP_RAVEN: 612 default: 613 /* Raven currently has issues with noretry 614 * regardless of what we decide for other 615 * asics, we should leave raven with 616 * noretry = 0 until we root cause the 617 * issues. 618 * 619 * default this to 0 for now, but we may want 620 * to change this in the future for certain 621 * GPUs as it can increase performance in 622 * certain cases. 623 */ 624 if (amdgpu_noretry == -1) 625 gmc->noretry = 0; 626 else 627 gmc->noretry = amdgpu_noretry; 628 break; 629 } 630 } 631 632 void amdgpu_gmc_set_vm_fault_masks(struct amdgpu_device *adev, int hub_type, 633 bool enable) 634 { 635 struct amdgpu_vmhub *hub; 636 u32 tmp, reg, i; 637 638 hub = &adev->vmhub[hub_type]; 639 for (i = 0; i < 16; i++) { 640 reg = hub->vm_context0_cntl + hub->ctx_distance * i; 641 642 tmp = (hub_type == AMDGPU_GFXHUB_0) ? 643 RREG32_SOC15_IP(GC, reg) : 644 RREG32_SOC15_IP(MMHUB, reg); 645 646 if (enable) 647 tmp |= hub->vm_cntx_cntl_vm_fault; 648 else 649 tmp &= ~hub->vm_cntx_cntl_vm_fault; 650 651 (hub_type == AMDGPU_GFXHUB_0) ? 652 WREG32_SOC15_IP(GC, reg, tmp) : 653 WREG32_SOC15_IP(MMHUB, reg, tmp); 654 } 655 } 656 657 void amdgpu_gmc_get_vbios_allocations(struct amdgpu_device *adev) 658 { 659 unsigned size; 660 661 /* 662 * TODO: 663 * Currently there is a bug where some memory client outside 664 * of the driver writes to first 8M of VRAM on S3 resume, 665 * this overrides GART which by default gets placed in first 8M and 666 * causes VM_FAULTS once GTT is accessed. 667 * Keep the stolen memory reservation until the while this is not solved. 668 */ 669 switch (adev->asic_type) { 670 case CHIP_VEGA10: 671 case CHIP_RAVEN: 672 case CHIP_RENOIR: 673 adev->mman.keep_stolen_vga_memory = true; 674 break; 675 default: 676 adev->mman.keep_stolen_vga_memory = false; 677 break; 678 } 679 680 if (amdgpu_sriov_vf(adev) || 681 !amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_DCE)) { 682 size = 0; 683 } else { 684 size = amdgpu_gmc_get_vbios_fb_size(adev); 685 686 if (adev->mman.keep_stolen_vga_memory) 687 size = max(size, (unsigned)AMDGPU_VBIOS_VGA_ALLOCATION); 688 } 689 690 /* set to 0 if the pre-OS buffer uses up most of vram */ 691 if ((adev->gmc.real_vram_size - size) < (8 * 1024 * 1024)) 692 size = 0; 693 694 if (size > AMDGPU_VBIOS_VGA_ALLOCATION) { 695 adev->mman.stolen_vga_size = AMDGPU_VBIOS_VGA_ALLOCATION; 696 adev->mman.stolen_extended_size = size - adev->mman.stolen_vga_size; 697 } else { 698 adev->mman.stolen_vga_size = size; 699 adev->mman.stolen_extended_size = 0; 700 } 701 } 702 703 /** 704 * amdgpu_gmc_init_pdb0 - initialize PDB0 705 * 706 * @adev: amdgpu_device pointer 707 * 708 * This function is only used when GART page table is used 709 * for FB address translatioin. In such a case, we construct 710 * a 2-level system VM page table: PDB0->PTB, to cover both 711 * VRAM of the hive and system memory. 712 * 713 * PDB0 is static, initialized once on driver initialization. 714 * The first n entries of PDB0 are used as PTE by setting 715 * P bit to 1, pointing to VRAM. The n+1'th entry points 716 * to a big PTB covering system memory. 717 * 718 */ 719 void amdgpu_gmc_init_pdb0(struct amdgpu_device *adev) 720 { 721 int i; 722 uint64_t flags = adev->gart.gart_pte_flags; //TODO it is UC. explore NC/RW? 723 /* Each PDE0 (used as PTE) covers (2^vmid0_page_table_block_size)*2M 724 */ 725 u64 vram_size = adev->gmc.xgmi.node_segment_size * adev->gmc.xgmi.num_physical_nodes; 726 u64 pde0_page_size = (1ULL<<adev->gmc.vmid0_page_table_block_size)<<21; 727 u64 vram_addr = adev->vm_manager.vram_base_offset - 728 adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size; 729 u64 vram_end = vram_addr + vram_size; 730 u64 gart_ptb_gpu_pa = amdgpu_gmc_vram_pa(adev, adev->gart.bo); 731 732 flags |= AMDGPU_PTE_VALID | AMDGPU_PTE_READABLE; 733 flags |= AMDGPU_PTE_WRITEABLE; 734 flags |= AMDGPU_PTE_SNOOPED; 735 flags |= AMDGPU_PTE_FRAG((adev->gmc.vmid0_page_table_block_size + 9*1)); 736 flags |= AMDGPU_PDE_PTE; 737 738 /* The first n PDE0 entries are used as PTE, 739 * pointing to vram 740 */ 741 for (i = 0; vram_addr < vram_end; i++, vram_addr += pde0_page_size) 742 amdgpu_gmc_set_pte_pde(adev, adev->gmc.ptr_pdb0, i, vram_addr, flags); 743 744 /* The n+1'th PDE0 entry points to a huge 745 * PTB who has more than 512 entries each 746 * pointing to a 4K system page 747 */ 748 flags = AMDGPU_PTE_VALID; 749 flags |= AMDGPU_PDE_BFS(0) | AMDGPU_PTE_SNOOPED; 750 /* Requires gart_ptb_gpu_pa to be 4K aligned */ 751 amdgpu_gmc_set_pte_pde(adev, adev->gmc.ptr_pdb0, i, gart_ptb_gpu_pa, flags); 752 } 753 754 /** 755 * amdgpu_gmc_vram_mc2pa - calculate vram buffer's physical address from MC 756 * address 757 * 758 * @adev: amdgpu_device pointer 759 * @mc_addr: MC address of buffer 760 */ 761 uint64_t amdgpu_gmc_vram_mc2pa(struct amdgpu_device *adev, uint64_t mc_addr) 762 { 763 return mc_addr - adev->gmc.vram_start + adev->vm_manager.vram_base_offset; 764 } 765 766 /** 767 * amdgpu_gmc_vram_pa - calculate vram buffer object's physical address from 768 * GPU's view 769 * 770 * @adev: amdgpu_device pointer 771 * @bo: amdgpu buffer object 772 */ 773 uint64_t amdgpu_gmc_vram_pa(struct amdgpu_device *adev, struct amdgpu_bo *bo) 774 { 775 return amdgpu_gmc_vram_mc2pa(adev, amdgpu_bo_gpu_offset(bo)); 776 } 777 778 /** 779 * amdgpu_gmc_vram_cpu_pa - calculate vram buffer object's physical address 780 * from CPU's view 781 * 782 * @adev: amdgpu_device pointer 783 * @bo: amdgpu buffer object 784 */ 785 uint64_t amdgpu_gmc_vram_cpu_pa(struct amdgpu_device *adev, struct amdgpu_bo *bo) 786 { 787 return amdgpu_bo_gpu_offset(bo) - adev->gmc.vram_start + adev->gmc.aper_base; 788 } 789 790 void amdgpu_gmc_get_reserved_allocation(struct amdgpu_device *adev) 791 { 792 /* Some ASICs need to reserve a region of video memory to avoid access 793 * from driver */ 794 adev->mman.stolen_reserved_offset = 0; 795 adev->mman.stolen_reserved_size = 0; 796 797 switch (adev->asic_type) { 798 case CHIP_YELLOW_CARP: 799 if (amdgpu_discovery == 0) { 800 adev->mman.stolen_reserved_offset = 0x1ffb0000; 801 adev->mman.stolen_reserved_size = 64 * PAGE_SIZE; 802 } 803 break; 804 default: 805 break; 806 } 807 } 808