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 #ifdef CONFIG_X86 29 #include <asm/hypervisor.h> 30 #endif 31 32 #include "amdgpu.h" 33 #include "amdgpu_gmc.h" 34 #include "amdgpu_ras.h" 35 #include "amdgpu_xgmi.h" 36 37 #include <drm/drm_drv.h> 38 #include <drm/ttm/ttm_tt.h> 39 40 /** 41 * amdgpu_gmc_pdb0_alloc - allocate vram for pdb0 42 * 43 * @adev: amdgpu_device pointer 44 * 45 * Allocate video memory for pdb0 and map it for CPU access 46 * Returns 0 for success, error for failure. 47 */ 48 int amdgpu_gmc_pdb0_alloc(struct amdgpu_device *adev) 49 { 50 int r; 51 struct amdgpu_bo_param bp; 52 u64 vram_size = adev->gmc.xgmi.node_segment_size * adev->gmc.xgmi.num_physical_nodes; 53 uint32_t pde0_page_shift = adev->gmc.vmid0_page_table_block_size + 21; 54 uint32_t npdes = (vram_size + (1ULL << pde0_page_shift) -1) >> pde0_page_shift; 55 56 memset(&bp, 0, sizeof(bp)); 57 bp.size = PAGE_ALIGN((npdes + 1) * 8); 58 bp.byte_align = PAGE_SIZE; 59 bp.domain = AMDGPU_GEM_DOMAIN_VRAM; 60 bp.flags = AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED | 61 AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS; 62 bp.type = ttm_bo_type_kernel; 63 bp.resv = NULL; 64 bp.bo_ptr_size = sizeof(struct amdgpu_bo); 65 66 r = amdgpu_bo_create(adev, &bp, &adev->gmc.pdb0_bo); 67 if (r) 68 return r; 69 70 r = amdgpu_bo_reserve(adev->gmc.pdb0_bo, false); 71 if (unlikely(r != 0)) 72 goto bo_reserve_failure; 73 74 r = amdgpu_bo_pin(adev->gmc.pdb0_bo, AMDGPU_GEM_DOMAIN_VRAM); 75 if (r) 76 goto bo_pin_failure; 77 r = amdgpu_bo_kmap(adev->gmc.pdb0_bo, &adev->gmc.ptr_pdb0); 78 if (r) 79 goto bo_kmap_failure; 80 81 amdgpu_bo_unreserve(adev->gmc.pdb0_bo); 82 return 0; 83 84 bo_kmap_failure: 85 amdgpu_bo_unpin(adev->gmc.pdb0_bo); 86 bo_pin_failure: 87 amdgpu_bo_unreserve(adev->gmc.pdb0_bo); 88 bo_reserve_failure: 89 amdgpu_bo_unref(&adev->gmc.pdb0_bo); 90 return r; 91 } 92 93 /** 94 * amdgpu_gmc_get_pde_for_bo - get the PDE for a BO 95 * 96 * @bo: the BO to get the PDE for 97 * @level: the level in the PD hirarchy 98 * @addr: resulting addr 99 * @flags: resulting flags 100 * 101 * Get the address and flags to be used for a PDE (Page Directory Entry). 102 */ 103 void amdgpu_gmc_get_pde_for_bo(struct amdgpu_bo *bo, int level, 104 uint64_t *addr, uint64_t *flags) 105 { 106 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); 107 108 switch (bo->tbo.resource->mem_type) { 109 case TTM_PL_TT: 110 *addr = bo->tbo.ttm->dma_address[0]; 111 break; 112 case TTM_PL_VRAM: 113 *addr = amdgpu_bo_gpu_offset(bo); 114 break; 115 default: 116 *addr = 0; 117 break; 118 } 119 *flags = amdgpu_ttm_tt_pde_flags(bo->tbo.ttm, bo->tbo.resource); 120 amdgpu_gmc_get_vm_pde(adev, level, addr, flags); 121 } 122 123 /* 124 * amdgpu_gmc_pd_addr - return the address of the root directory 125 */ 126 uint64_t amdgpu_gmc_pd_addr(struct amdgpu_bo *bo) 127 { 128 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); 129 uint64_t pd_addr; 130 131 /* TODO: move that into ASIC specific code */ 132 if (adev->asic_type >= CHIP_VEGA10) { 133 uint64_t flags = AMDGPU_PTE_VALID; 134 135 amdgpu_gmc_get_pde_for_bo(bo, -1, &pd_addr, &flags); 136 pd_addr |= flags; 137 } else { 138 pd_addr = amdgpu_bo_gpu_offset(bo); 139 } 140 return pd_addr; 141 } 142 143 /** 144 * amdgpu_gmc_set_pte_pde - update the page tables using CPU 145 * 146 * @adev: amdgpu_device pointer 147 * @cpu_pt_addr: cpu address of the page table 148 * @gpu_page_idx: entry in the page table to update 149 * @addr: dst addr to write into pte/pde 150 * @flags: access flags 151 * 152 * Update the page tables using CPU. 153 */ 154 int amdgpu_gmc_set_pte_pde(struct amdgpu_device *adev, void *cpu_pt_addr, 155 uint32_t gpu_page_idx, uint64_t addr, 156 uint64_t flags) 157 { 158 void __iomem *ptr = (void *)cpu_pt_addr; 159 uint64_t value; 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 return 0; 169 } 170 171 /** 172 * amdgpu_gmc_agp_addr - return the address in the AGP address space 173 * 174 * @bo: TTM BO which needs the address, must be in GTT domain 175 * 176 * Tries to figure out how to access the BO through the AGP aperture. Returns 177 * AMDGPU_BO_INVALID_OFFSET if that is not possible. 178 */ 179 uint64_t amdgpu_gmc_agp_addr(struct ttm_buffer_object *bo) 180 { 181 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev); 182 183 if (bo->ttm->num_pages != 1 || bo->ttm->caching == ttm_cached) 184 return AMDGPU_BO_INVALID_OFFSET; 185 186 if (bo->ttm->dma_address[0] + PAGE_SIZE >= adev->gmc.agp_size) 187 return AMDGPU_BO_INVALID_OFFSET; 188 189 return adev->gmc.agp_start + bo->ttm->dma_address[0]; 190 } 191 192 /** 193 * amdgpu_gmc_vram_location - try to find VRAM location 194 * 195 * @adev: amdgpu device structure holding all necessary information 196 * @mc: memory controller structure holding memory information 197 * @base: base address at which to put VRAM 198 * 199 * Function will try to place VRAM at base address provided 200 * as parameter. 201 */ 202 void amdgpu_gmc_vram_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc, 203 u64 base) 204 { 205 uint64_t vis_limit = (uint64_t)amdgpu_vis_vram_limit << 20; 206 uint64_t limit = (uint64_t)amdgpu_vram_limit << 20; 207 208 mc->vram_start = base; 209 mc->vram_end = mc->vram_start + mc->mc_vram_size - 1; 210 if (limit < mc->real_vram_size) 211 mc->real_vram_size = limit; 212 213 if (vis_limit && vis_limit < mc->visible_vram_size) 214 mc->visible_vram_size = vis_limit; 215 216 if (mc->real_vram_size < mc->visible_vram_size) 217 mc->visible_vram_size = mc->real_vram_size; 218 219 if (mc->xgmi.num_physical_nodes == 0) { 220 mc->fb_start = mc->vram_start; 221 mc->fb_end = mc->vram_end; 222 } 223 dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n", 224 mc->mc_vram_size >> 20, mc->vram_start, 225 mc->vram_end, mc->real_vram_size >> 20); 226 } 227 228 /** amdgpu_gmc_sysvm_location - place vram and gart in sysvm aperture 229 * 230 * @adev: amdgpu device structure holding all necessary information 231 * @mc: memory controller structure holding memory information 232 * 233 * This function is only used if use GART for FB translation. In such 234 * case, we use sysvm aperture (vmid0 page tables) for both vram 235 * and gart (aka system memory) access. 236 * 237 * GPUVM (and our organization of vmid0 page tables) require sysvm 238 * aperture to be placed at a location aligned with 8 times of native 239 * page size. For example, if vm_context0_cntl.page_table_block_size 240 * is 12, then native page size is 8G (2M*2^12), sysvm should start 241 * with a 64G aligned address. For simplicity, we just put sysvm at 242 * address 0. So vram start at address 0 and gart is right after vram. 243 */ 244 void amdgpu_gmc_sysvm_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc) 245 { 246 u64 hive_vram_start = 0; 247 u64 hive_vram_end = mc->xgmi.node_segment_size * mc->xgmi.num_physical_nodes - 1; 248 mc->vram_start = mc->xgmi.node_segment_size * mc->xgmi.physical_node_id; 249 mc->vram_end = mc->vram_start + mc->xgmi.node_segment_size - 1; 250 mc->gart_start = hive_vram_end + 1; 251 mc->gart_end = mc->gart_start + mc->gart_size - 1; 252 mc->fb_start = hive_vram_start; 253 mc->fb_end = hive_vram_end; 254 dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n", 255 mc->mc_vram_size >> 20, mc->vram_start, 256 mc->vram_end, mc->real_vram_size >> 20); 257 dev_info(adev->dev, "GART: %lluM 0x%016llX - 0x%016llX\n", 258 mc->gart_size >> 20, mc->gart_start, mc->gart_end); 259 } 260 261 /** 262 * amdgpu_gmc_gart_location - try to find GART location 263 * 264 * @adev: amdgpu device structure holding all necessary information 265 * @mc: memory controller structure holding memory information 266 * 267 * Function will place try to place GART before or after VRAM. 268 * If GART size is bigger than space left then we ajust GART size. 269 * Thus function will never fails. 270 */ 271 void amdgpu_gmc_gart_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc) 272 { 273 const uint64_t four_gb = 0x100000000ULL; 274 u64 size_af, size_bf; 275 /*To avoid the hole, limit the max mc address to AMDGPU_GMC_HOLE_START*/ 276 u64 max_mc_address = min(adev->gmc.mc_mask, AMDGPU_GMC_HOLE_START - 1); 277 278 /* VCE doesn't like it when BOs cross a 4GB segment, so align 279 * the GART base on a 4GB boundary as well. 280 */ 281 size_bf = mc->fb_start; 282 size_af = max_mc_address + 1 - ALIGN(mc->fb_end + 1, four_gb); 283 284 if (mc->gart_size > max(size_bf, size_af)) { 285 dev_warn(adev->dev, "limiting GART\n"); 286 mc->gart_size = max(size_bf, size_af); 287 } 288 289 if ((size_bf >= mc->gart_size && size_bf < size_af) || 290 (size_af < mc->gart_size)) 291 mc->gart_start = 0; 292 else 293 mc->gart_start = max_mc_address - mc->gart_size + 1; 294 295 mc->gart_start &= ~(four_gb - 1); 296 mc->gart_end = mc->gart_start + mc->gart_size - 1; 297 dev_info(adev->dev, "GART: %lluM 0x%016llX - 0x%016llX\n", 298 mc->gart_size >> 20, mc->gart_start, mc->gart_end); 299 } 300 301 /** 302 * amdgpu_gmc_agp_location - try to find AGP location 303 * @adev: amdgpu device structure holding all necessary information 304 * @mc: memory controller structure holding memory information 305 * 306 * Function will place try to find a place for the AGP BAR in the MC address 307 * space. 308 * 309 * AGP BAR will be assigned the largest available hole in the address space. 310 * Should be called after VRAM and GART locations are setup. 311 */ 312 void amdgpu_gmc_agp_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc) 313 { 314 const uint64_t sixteen_gb = 1ULL << 34; 315 const uint64_t sixteen_gb_mask = ~(sixteen_gb - 1); 316 u64 size_af, size_bf; 317 318 if (amdgpu_sriov_vf(adev)) { 319 mc->agp_start = 0xffffffffffff; 320 mc->agp_end = 0x0; 321 mc->agp_size = 0; 322 323 return; 324 } 325 326 if (mc->fb_start > mc->gart_start) { 327 size_bf = (mc->fb_start & sixteen_gb_mask) - 328 ALIGN(mc->gart_end + 1, sixteen_gb); 329 size_af = mc->mc_mask + 1 - ALIGN(mc->fb_end + 1, sixteen_gb); 330 } else { 331 size_bf = mc->fb_start & sixteen_gb_mask; 332 size_af = (mc->gart_start & sixteen_gb_mask) - 333 ALIGN(mc->fb_end + 1, sixteen_gb); 334 } 335 336 if (size_bf > size_af) { 337 mc->agp_start = (mc->fb_start - size_bf) & sixteen_gb_mask; 338 mc->agp_size = size_bf; 339 } else { 340 mc->agp_start = ALIGN(mc->fb_end + 1, sixteen_gb); 341 mc->agp_size = size_af; 342 } 343 344 mc->agp_end = mc->agp_start + mc->agp_size - 1; 345 dev_info(adev->dev, "AGP: %lluM 0x%016llX - 0x%016llX\n", 346 mc->agp_size >> 20, mc->agp_start, mc->agp_end); 347 } 348 349 /** 350 * amdgpu_gmc_fault_key - get hask key from vm fault address and pasid 351 * 352 * @addr: 48 bit physical address, page aligned (36 significant bits) 353 * @pasid: 16 bit process address space identifier 354 */ 355 static inline uint64_t amdgpu_gmc_fault_key(uint64_t addr, uint16_t pasid) 356 { 357 return addr << 4 | pasid; 358 } 359 360 /** 361 * amdgpu_gmc_filter_faults - filter VM faults 362 * 363 * @adev: amdgpu device structure 364 * @ih: interrupt ring that the fault received from 365 * @addr: address of the VM fault 366 * @pasid: PASID of the process causing the fault 367 * @timestamp: timestamp of the fault 368 * 369 * Returns: 370 * True if the fault was filtered and should not be processed further. 371 * False if the fault is a new one and needs to be handled. 372 */ 373 bool amdgpu_gmc_filter_faults(struct amdgpu_device *adev, 374 struct amdgpu_ih_ring *ih, uint64_t addr, 375 uint16_t pasid, uint64_t timestamp) 376 { 377 struct amdgpu_gmc *gmc = &adev->gmc; 378 uint64_t stamp, key = amdgpu_gmc_fault_key(addr, pasid); 379 struct amdgpu_gmc_fault *fault; 380 uint32_t hash; 381 382 /* Stale retry fault if timestamp goes backward */ 383 if (amdgpu_ih_ts_after(timestamp, ih->processed_timestamp)) 384 return true; 385 386 /* If we don't have space left in the ring buffer return immediately */ 387 stamp = max(timestamp, AMDGPU_GMC_FAULT_TIMEOUT + 1) - 388 AMDGPU_GMC_FAULT_TIMEOUT; 389 if (gmc->fault_ring[gmc->last_fault].timestamp >= stamp) 390 return true; 391 392 /* Try to find the fault in the hash */ 393 hash = hash_64(key, AMDGPU_GMC_FAULT_HASH_ORDER); 394 fault = &gmc->fault_ring[gmc->fault_hash[hash].idx]; 395 while (fault->timestamp >= stamp) { 396 uint64_t tmp; 397 398 if (atomic64_read(&fault->key) == key) { 399 /* 400 * if we get a fault which is already present in 401 * the fault_ring and the timestamp of 402 * the fault is after the expired timestamp, 403 * then this is a new fault that needs to be added 404 * into the fault ring. 405 */ 406 if (fault->timestamp_expiry != 0 && 407 amdgpu_ih_ts_after(fault->timestamp_expiry, 408 timestamp)) 409 break; 410 else 411 return true; 412 } 413 414 tmp = fault->timestamp; 415 fault = &gmc->fault_ring[fault->next]; 416 417 /* Check if the entry was reused */ 418 if (fault->timestamp >= tmp) 419 break; 420 } 421 422 /* Add the fault to the ring */ 423 fault = &gmc->fault_ring[gmc->last_fault]; 424 atomic64_set(&fault->key, key); 425 fault->timestamp = timestamp; 426 427 /* And update the hash */ 428 fault->next = gmc->fault_hash[hash].idx; 429 gmc->fault_hash[hash].idx = gmc->last_fault++; 430 return false; 431 } 432 433 /** 434 * amdgpu_gmc_filter_faults_remove - remove address from VM faults filter 435 * 436 * @adev: amdgpu device structure 437 * @addr: address of the VM fault 438 * @pasid: PASID of the process causing the fault 439 * 440 * Remove the address from fault filter, then future vm fault on this address 441 * will pass to retry fault handler to recover. 442 */ 443 void amdgpu_gmc_filter_faults_remove(struct amdgpu_device *adev, uint64_t addr, 444 uint16_t pasid) 445 { 446 struct amdgpu_gmc *gmc = &adev->gmc; 447 uint64_t key = amdgpu_gmc_fault_key(addr, pasid); 448 struct amdgpu_ih_ring *ih; 449 struct amdgpu_gmc_fault *fault; 450 uint32_t last_wptr; 451 uint64_t last_ts; 452 uint32_t hash; 453 uint64_t tmp; 454 455 ih = adev->irq.retry_cam_enabled ? &adev->irq.ih_soft : &adev->irq.ih1; 456 /* Get the WPTR of the last entry in IH ring */ 457 last_wptr = amdgpu_ih_get_wptr(adev, ih); 458 /* Order wptr with ring data. */ 459 rmb(); 460 /* Get the timetamp of the last entry in IH ring */ 461 last_ts = amdgpu_ih_decode_iv_ts(adev, ih, last_wptr, -1); 462 463 hash = hash_64(key, AMDGPU_GMC_FAULT_HASH_ORDER); 464 fault = &gmc->fault_ring[gmc->fault_hash[hash].idx]; 465 do { 466 if (atomic64_read(&fault->key) == key) { 467 /* 468 * Update the timestamp when this fault 469 * expired. 470 */ 471 fault->timestamp_expiry = last_ts; 472 break; 473 } 474 475 tmp = fault->timestamp; 476 fault = &gmc->fault_ring[fault->next]; 477 } while (fault->timestamp < tmp); 478 } 479 480 int amdgpu_gmc_ras_sw_init(struct amdgpu_device *adev) 481 { 482 int r; 483 484 /* umc ras block */ 485 r = amdgpu_umc_ras_sw_init(adev); 486 if (r) 487 return r; 488 489 /* mmhub ras block */ 490 r = amdgpu_mmhub_ras_sw_init(adev); 491 if (r) 492 return r; 493 494 /* hdp ras block */ 495 r = amdgpu_hdp_ras_sw_init(adev); 496 if (r) 497 return r; 498 499 /* mca.x ras block */ 500 r = amdgpu_mca_mp0_ras_sw_init(adev); 501 if (r) 502 return r; 503 504 r = amdgpu_mca_mp1_ras_sw_init(adev); 505 if (r) 506 return r; 507 508 r = amdgpu_mca_mpio_ras_sw_init(adev); 509 if (r) 510 return r; 511 512 /* xgmi ras block */ 513 r = amdgpu_xgmi_ras_sw_init(adev); 514 if (r) 515 return r; 516 517 return 0; 518 } 519 520 int amdgpu_gmc_ras_late_init(struct amdgpu_device *adev) 521 { 522 return 0; 523 } 524 525 void amdgpu_gmc_ras_fini(struct amdgpu_device *adev) 526 { 527 528 } 529 530 /* 531 * The latest engine allocation on gfx9/10 is: 532 * Engine 2, 3: firmware 533 * Engine 0, 1, 4~16: amdgpu ring, 534 * subject to change when ring number changes 535 * Engine 17: Gart flushes 536 */ 537 #define GFXHUB_FREE_VM_INV_ENGS_BITMAP 0x1FFF3 538 #define MMHUB_FREE_VM_INV_ENGS_BITMAP 0x1FFF3 539 540 int amdgpu_gmc_allocate_vm_inv_eng(struct amdgpu_device *adev) 541 { 542 struct amdgpu_ring *ring; 543 unsigned vm_inv_engs[AMDGPU_MAX_VMHUBS] = 544 {GFXHUB_FREE_VM_INV_ENGS_BITMAP, MMHUB_FREE_VM_INV_ENGS_BITMAP, 545 GFXHUB_FREE_VM_INV_ENGS_BITMAP}; 546 unsigned i; 547 unsigned vmhub, inv_eng; 548 549 if (adev->enable_mes) { 550 /* reserve engine 5 for firmware */ 551 for (vmhub = 0; vmhub < AMDGPU_MAX_VMHUBS; vmhub++) 552 vm_inv_engs[vmhub] &= ~(1 << 5); 553 } 554 555 for (i = 0; i < adev->num_rings; ++i) { 556 ring = adev->rings[i]; 557 vmhub = ring->vm_hub; 558 559 if (ring == &adev->mes.ring) 560 continue; 561 562 inv_eng = ffs(vm_inv_engs[vmhub]); 563 if (!inv_eng) { 564 dev_err(adev->dev, "no VM inv eng for ring %s\n", 565 ring->name); 566 return -EINVAL; 567 } 568 569 ring->vm_inv_eng = inv_eng - 1; 570 vm_inv_engs[vmhub] &= ~(1 << ring->vm_inv_eng); 571 572 dev_info(adev->dev, "ring %s uses VM inv eng %u on hub %u\n", 573 ring->name, ring->vm_inv_eng, ring->vm_hub); 574 } 575 576 return 0; 577 } 578 579 /** 580 * amdgpu_gmc_tmz_set -- check and set if a device supports TMZ 581 * @adev: amdgpu_device pointer 582 * 583 * Check and set if an the device @adev supports Trusted Memory 584 * Zones (TMZ). 585 */ 586 void amdgpu_gmc_tmz_set(struct amdgpu_device *adev) 587 { 588 switch (adev->ip_versions[GC_HWIP][0]) { 589 /* RAVEN */ 590 case IP_VERSION(9, 2, 2): 591 case IP_VERSION(9, 1, 0): 592 /* RENOIR looks like RAVEN */ 593 case IP_VERSION(9, 3, 0): 594 /* GC 10.3.7 */ 595 case IP_VERSION(10, 3, 7): 596 if (amdgpu_tmz == 0) { 597 adev->gmc.tmz_enabled = false; 598 dev_info(adev->dev, 599 "Trusted Memory Zone (TMZ) feature disabled (cmd line)\n"); 600 } else { 601 adev->gmc.tmz_enabled = true; 602 dev_info(adev->dev, 603 "Trusted Memory Zone (TMZ) feature enabled\n"); 604 } 605 break; 606 case IP_VERSION(10, 1, 10): 607 case IP_VERSION(10, 1, 1): 608 case IP_VERSION(10, 1, 2): 609 case IP_VERSION(10, 1, 3): 610 case IP_VERSION(10, 3, 0): 611 case IP_VERSION(10, 3, 2): 612 case IP_VERSION(10, 3, 4): 613 case IP_VERSION(10, 3, 5): 614 case IP_VERSION(10, 3, 6): 615 /* VANGOGH */ 616 case IP_VERSION(10, 3, 1): 617 /* YELLOW_CARP*/ 618 case IP_VERSION(10, 3, 3): 619 case IP_VERSION(11, 0, 1): 620 case IP_VERSION(11, 0, 4): 621 /* Don't enable it by default yet. 622 */ 623 if (amdgpu_tmz < 1) { 624 adev->gmc.tmz_enabled = false; 625 dev_info(adev->dev, 626 "Trusted Memory Zone (TMZ) feature disabled as experimental (default)\n"); 627 } else { 628 adev->gmc.tmz_enabled = true; 629 dev_info(adev->dev, 630 "Trusted Memory Zone (TMZ) feature enabled as experimental (cmd line)\n"); 631 } 632 break; 633 default: 634 adev->gmc.tmz_enabled = false; 635 dev_info(adev->dev, 636 "Trusted Memory Zone (TMZ) feature not supported\n"); 637 break; 638 } 639 } 640 641 /** 642 * amdgpu_gmc_noretry_set -- set per asic noretry defaults 643 * @adev: amdgpu_device pointer 644 * 645 * Set a per asic default for the no-retry parameter. 646 * 647 */ 648 void amdgpu_gmc_noretry_set(struct amdgpu_device *adev) 649 { 650 struct amdgpu_gmc *gmc = &adev->gmc; 651 uint32_t gc_ver = adev->ip_versions[GC_HWIP][0]; 652 bool noretry_default = (gc_ver == IP_VERSION(9, 0, 1) || 653 gc_ver == IP_VERSION(9, 3, 0) || 654 gc_ver == IP_VERSION(9, 4, 0) || 655 gc_ver == IP_VERSION(9, 4, 1) || 656 gc_ver == IP_VERSION(9, 4, 2) || 657 gc_ver == IP_VERSION(9, 4, 3) || 658 gc_ver >= IP_VERSION(10, 3, 0)); 659 660 gmc->noretry = (amdgpu_noretry == -1) ? noretry_default : amdgpu_noretry; 661 } 662 663 void amdgpu_gmc_set_vm_fault_masks(struct amdgpu_device *adev, int hub_type, 664 bool enable) 665 { 666 struct amdgpu_vmhub *hub; 667 u32 tmp, reg, i; 668 669 hub = &adev->vmhub[hub_type]; 670 for (i = 0; i < 16; i++) { 671 reg = hub->vm_context0_cntl + hub->ctx_distance * i; 672 673 tmp = (hub_type == AMDGPU_GFXHUB_0) ? 674 RREG32_SOC15_IP(GC, reg) : 675 RREG32_SOC15_IP(MMHUB, reg); 676 677 if (enable) 678 tmp |= hub->vm_cntx_cntl_vm_fault; 679 else 680 tmp &= ~hub->vm_cntx_cntl_vm_fault; 681 682 (hub_type == AMDGPU_GFXHUB_0) ? 683 WREG32_SOC15_IP(GC, reg, tmp) : 684 WREG32_SOC15_IP(MMHUB, reg, tmp); 685 } 686 } 687 688 void amdgpu_gmc_get_vbios_allocations(struct amdgpu_device *adev) 689 { 690 unsigned size; 691 692 /* 693 * Some ASICs need to reserve a region of video memory to avoid access 694 * from driver 695 */ 696 adev->mman.stolen_reserved_offset = 0; 697 adev->mman.stolen_reserved_size = 0; 698 699 /* 700 * TODO: 701 * Currently there is a bug where some memory client outside 702 * of the driver writes to first 8M of VRAM on S3 resume, 703 * this overrides GART which by default gets placed in first 8M and 704 * causes VM_FAULTS once GTT is accessed. 705 * Keep the stolen memory reservation until the while this is not solved. 706 */ 707 switch (adev->asic_type) { 708 case CHIP_VEGA10: 709 adev->mman.keep_stolen_vga_memory = true; 710 /* 711 * VEGA10 SRIOV VF with MS_HYPERV host needs some firmware reserved area. 712 */ 713 #ifdef CONFIG_X86 714 if (amdgpu_sriov_vf(adev) && hypervisor_is_type(X86_HYPER_MS_HYPERV)) { 715 adev->mman.stolen_reserved_offset = 0x500000; 716 adev->mman.stolen_reserved_size = 0x200000; 717 } 718 #endif 719 break; 720 case CHIP_RAVEN: 721 case CHIP_RENOIR: 722 adev->mman.keep_stolen_vga_memory = true; 723 break; 724 case CHIP_YELLOW_CARP: 725 if (amdgpu_discovery == 0) { 726 adev->mman.stolen_reserved_offset = 0x1ffb0000; 727 adev->mman.stolen_reserved_size = 64 * PAGE_SIZE; 728 } 729 break; 730 default: 731 adev->mman.keep_stolen_vga_memory = false; 732 break; 733 } 734 735 if (amdgpu_sriov_vf(adev) || 736 !amdgpu_device_has_display_hardware(adev)) { 737 size = 0; 738 } else { 739 size = amdgpu_gmc_get_vbios_fb_size(adev); 740 741 if (adev->mman.keep_stolen_vga_memory) 742 size = max(size, (unsigned)AMDGPU_VBIOS_VGA_ALLOCATION); 743 } 744 745 /* set to 0 if the pre-OS buffer uses up most of vram */ 746 if ((adev->gmc.real_vram_size - size) < (8 * 1024 * 1024)) 747 size = 0; 748 749 if (size > AMDGPU_VBIOS_VGA_ALLOCATION) { 750 adev->mman.stolen_vga_size = AMDGPU_VBIOS_VGA_ALLOCATION; 751 adev->mman.stolen_extended_size = size - adev->mman.stolen_vga_size; 752 } else { 753 adev->mman.stolen_vga_size = size; 754 adev->mman.stolen_extended_size = 0; 755 } 756 } 757 758 /** 759 * amdgpu_gmc_init_pdb0 - initialize PDB0 760 * 761 * @adev: amdgpu_device pointer 762 * 763 * This function is only used when GART page table is used 764 * for FB address translatioin. In such a case, we construct 765 * a 2-level system VM page table: PDB0->PTB, to cover both 766 * VRAM of the hive and system memory. 767 * 768 * PDB0 is static, initialized once on driver initialization. 769 * The first n entries of PDB0 are used as PTE by setting 770 * P bit to 1, pointing to VRAM. The n+1'th entry points 771 * to a big PTB covering system memory. 772 * 773 */ 774 void amdgpu_gmc_init_pdb0(struct amdgpu_device *adev) 775 { 776 int i; 777 uint64_t flags = adev->gart.gart_pte_flags; //TODO it is UC. explore NC/RW? 778 /* Each PDE0 (used as PTE) covers (2^vmid0_page_table_block_size)*2M 779 */ 780 u64 vram_size = adev->gmc.xgmi.node_segment_size * adev->gmc.xgmi.num_physical_nodes; 781 u64 pde0_page_size = (1ULL<<adev->gmc.vmid0_page_table_block_size)<<21; 782 u64 vram_addr = adev->vm_manager.vram_base_offset - 783 adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size; 784 u64 vram_end = vram_addr + vram_size; 785 u64 gart_ptb_gpu_pa = amdgpu_gmc_vram_pa(adev, adev->gart.bo); 786 int idx; 787 788 if (!drm_dev_enter(adev_to_drm(adev), &idx)) 789 return; 790 791 flags |= AMDGPU_PTE_VALID | AMDGPU_PTE_READABLE; 792 flags |= AMDGPU_PTE_WRITEABLE; 793 flags |= AMDGPU_PTE_SNOOPED; 794 flags |= AMDGPU_PTE_FRAG((adev->gmc.vmid0_page_table_block_size + 9*1)); 795 flags |= AMDGPU_PDE_PTE; 796 797 /* The first n PDE0 entries are used as PTE, 798 * pointing to vram 799 */ 800 for (i = 0; vram_addr < vram_end; i++, vram_addr += pde0_page_size) 801 amdgpu_gmc_set_pte_pde(adev, adev->gmc.ptr_pdb0, i, vram_addr, flags); 802 803 /* The n+1'th PDE0 entry points to a huge 804 * PTB who has more than 512 entries each 805 * pointing to a 4K system page 806 */ 807 flags = AMDGPU_PTE_VALID; 808 flags |= AMDGPU_PDE_BFS(0) | AMDGPU_PTE_SNOOPED; 809 /* Requires gart_ptb_gpu_pa to be 4K aligned */ 810 amdgpu_gmc_set_pte_pde(adev, adev->gmc.ptr_pdb0, i, gart_ptb_gpu_pa, flags); 811 drm_dev_exit(idx); 812 } 813 814 /** 815 * amdgpu_gmc_vram_mc2pa - calculate vram buffer's physical address from MC 816 * address 817 * 818 * @adev: amdgpu_device pointer 819 * @mc_addr: MC address of buffer 820 */ 821 uint64_t amdgpu_gmc_vram_mc2pa(struct amdgpu_device *adev, uint64_t mc_addr) 822 { 823 return mc_addr - adev->gmc.vram_start + adev->vm_manager.vram_base_offset; 824 } 825 826 /** 827 * amdgpu_gmc_vram_pa - calculate vram buffer object's physical address from 828 * GPU's view 829 * 830 * @adev: amdgpu_device pointer 831 * @bo: amdgpu buffer object 832 */ 833 uint64_t amdgpu_gmc_vram_pa(struct amdgpu_device *adev, struct amdgpu_bo *bo) 834 { 835 return amdgpu_gmc_vram_mc2pa(adev, amdgpu_bo_gpu_offset(bo)); 836 } 837 838 /** 839 * amdgpu_gmc_vram_cpu_pa - calculate vram buffer object's physical address 840 * from CPU's view 841 * 842 * @adev: amdgpu_device pointer 843 * @bo: amdgpu buffer object 844 */ 845 uint64_t amdgpu_gmc_vram_cpu_pa(struct amdgpu_device *adev, struct amdgpu_bo *bo) 846 { 847 return amdgpu_bo_gpu_offset(bo) - adev->gmc.vram_start + adev->gmc.aper_base; 848 } 849 850 int amdgpu_gmc_vram_checking(struct amdgpu_device *adev) 851 { 852 struct amdgpu_bo *vram_bo = NULL; 853 uint64_t vram_gpu = 0; 854 void *vram_ptr = NULL; 855 856 int ret, size = 0x100000; 857 uint8_t cptr[10]; 858 859 ret = amdgpu_bo_create_kernel(adev, size, PAGE_SIZE, 860 AMDGPU_GEM_DOMAIN_VRAM, 861 &vram_bo, 862 &vram_gpu, 863 &vram_ptr); 864 if (ret) 865 return ret; 866 867 memset(vram_ptr, 0x86, size); 868 memset(cptr, 0x86, 10); 869 870 /** 871 * Check the start, the mid, and the end of the memory if the content of 872 * each byte is the pattern "0x86". If yes, we suppose the vram bo is 873 * workable. 874 * 875 * Note: If check the each byte of whole 1M bo, it will cost too many 876 * seconds, so here, we just pick up three parts for emulation. 877 */ 878 ret = memcmp(vram_ptr, cptr, 10); 879 if (ret) 880 return ret; 881 882 ret = memcmp(vram_ptr + (size / 2), cptr, 10); 883 if (ret) 884 return ret; 885 886 ret = memcmp(vram_ptr + size - 10, cptr, 10); 887 if (ret) 888 return ret; 889 890 amdgpu_bo_free_kernel(&vram_bo, &vram_gpu, 891 &vram_ptr); 892 893 return 0; 894 } 895