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_reset.h" 36 #include "amdgpu_xgmi.h" 37 38 #include <drm/drm_drv.h> 39 #include <drm/ttm/ttm_tt.h> 40 41 /** 42 * amdgpu_gmc_pdb0_alloc - allocate vram for pdb0 43 * 44 * @adev: amdgpu_device pointer 45 * 46 * Allocate video memory for pdb0 and map it for CPU access 47 * Returns 0 for success, error for failure. 48 */ 49 int amdgpu_gmc_pdb0_alloc(struct amdgpu_device *adev) 50 { 51 int r; 52 struct amdgpu_bo_param bp; 53 u64 vram_size = adev->gmc.xgmi.node_segment_size * adev->gmc.xgmi.num_physical_nodes; 54 uint32_t pde0_page_shift = adev->gmc.vmid0_page_table_block_size + 21; 55 uint32_t npdes = (vram_size + (1ULL << pde0_page_shift) -1) >> pde0_page_shift; 56 57 memset(&bp, 0, sizeof(bp)); 58 bp.size = PAGE_ALIGN((npdes + 1) * 8); 59 bp.byte_align = PAGE_SIZE; 60 bp.domain = AMDGPU_GEM_DOMAIN_VRAM; 61 bp.flags = AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED | 62 AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS; 63 bp.type = ttm_bo_type_kernel; 64 bp.resv = NULL; 65 bp.bo_ptr_size = sizeof(struct amdgpu_bo); 66 67 r = amdgpu_bo_create(adev, &bp, &adev->gmc.pdb0_bo); 68 if (r) 69 return r; 70 71 r = amdgpu_bo_reserve(adev->gmc.pdb0_bo, false); 72 if (unlikely(r != 0)) 73 goto bo_reserve_failure; 74 75 r = amdgpu_bo_pin(adev->gmc.pdb0_bo, AMDGPU_GEM_DOMAIN_VRAM); 76 if (r) 77 goto bo_pin_failure; 78 r = amdgpu_bo_kmap(adev->gmc.pdb0_bo, &adev->gmc.ptr_pdb0); 79 if (r) 80 goto bo_kmap_failure; 81 82 amdgpu_bo_unreserve(adev->gmc.pdb0_bo); 83 return 0; 84 85 bo_kmap_failure: 86 amdgpu_bo_unpin(adev->gmc.pdb0_bo); 87 bo_pin_failure: 88 amdgpu_bo_unreserve(adev->gmc.pdb0_bo); 89 bo_reserve_failure: 90 amdgpu_bo_unref(&adev->gmc.pdb0_bo); 91 return r; 92 } 93 94 /** 95 * amdgpu_gmc_get_pde_for_bo - get the PDE for a BO 96 * 97 * @bo: the BO to get the PDE for 98 * @level: the level in the PD hirarchy 99 * @addr: resulting addr 100 * @flags: resulting flags 101 * 102 * Get the address and flags to be used for a PDE (Page Directory Entry). 103 */ 104 void amdgpu_gmc_get_pde_for_bo(struct amdgpu_bo *bo, int level, 105 uint64_t *addr, uint64_t *flags) 106 { 107 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); 108 109 switch (bo->tbo.resource->mem_type) { 110 case TTM_PL_TT: 111 *addr = bo->tbo.ttm->dma_address[0]; 112 break; 113 case TTM_PL_VRAM: 114 *addr = amdgpu_bo_gpu_offset(bo); 115 break; 116 default: 117 *addr = 0; 118 break; 119 } 120 *flags = amdgpu_ttm_tt_pde_flags(bo->tbo.ttm, bo->tbo.resource); 121 amdgpu_gmc_get_vm_pde(adev, level, addr, flags); 122 } 123 124 /* 125 * amdgpu_gmc_pd_addr - return the address of the root directory 126 */ 127 uint64_t amdgpu_gmc_pd_addr(struct amdgpu_bo *bo) 128 { 129 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); 130 uint64_t pd_addr; 131 132 /* TODO: move that into ASIC specific code */ 133 if (adev->asic_type >= CHIP_VEGA10) { 134 uint64_t flags = AMDGPU_PTE_VALID; 135 136 amdgpu_gmc_get_pde_for_bo(bo, -1, &pd_addr, &flags); 137 pd_addr |= flags; 138 } else { 139 pd_addr = amdgpu_bo_gpu_offset(bo); 140 } 141 return pd_addr; 142 } 143 144 /** 145 * amdgpu_gmc_set_pte_pde - update the page tables using CPU 146 * 147 * @adev: amdgpu_device pointer 148 * @cpu_pt_addr: cpu address of the page table 149 * @gpu_page_idx: entry in the page table to update 150 * @addr: dst addr to write into pte/pde 151 * @flags: access flags 152 * 153 * Update the page tables using CPU. 154 */ 155 int amdgpu_gmc_set_pte_pde(struct amdgpu_device *adev, void *cpu_pt_addr, 156 uint32_t gpu_page_idx, uint64_t addr, 157 uint64_t flags) 158 { 159 void __iomem *ptr = (void *)cpu_pt_addr; 160 uint64_t value; 161 162 /* 163 * The following is for PTE only. GART does not have PDEs. 164 */ 165 value = addr & 0x0000FFFFFFFFF000ULL; 166 value |= flags; 167 writeq(value, ptr + (gpu_page_idx * 8)); 168 169 return 0; 170 } 171 172 /** 173 * amdgpu_gmc_agp_addr - return the address in the AGP address space 174 * 175 * @bo: TTM BO which needs the address, must be in GTT domain 176 * 177 * Tries to figure out how to access the BO through the AGP aperture. Returns 178 * AMDGPU_BO_INVALID_OFFSET if that is not possible. 179 */ 180 uint64_t amdgpu_gmc_agp_addr(struct ttm_buffer_object *bo) 181 { 182 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev); 183 184 if (!bo->ttm) 185 return AMDGPU_BO_INVALID_OFFSET; 186 187 if (bo->ttm->num_pages != 1 || bo->ttm->caching == ttm_cached) 188 return AMDGPU_BO_INVALID_OFFSET; 189 190 if (bo->ttm->dma_address[0] + PAGE_SIZE >= adev->gmc.agp_size) 191 return AMDGPU_BO_INVALID_OFFSET; 192 193 return adev->gmc.agp_start + bo->ttm->dma_address[0]; 194 } 195 196 /** 197 * amdgpu_gmc_vram_location - try to find VRAM location 198 * 199 * @adev: amdgpu device structure holding all necessary information 200 * @mc: memory controller structure holding memory information 201 * @base: base address at which to put VRAM 202 * 203 * Function will try to place VRAM at base address provided 204 * as parameter. 205 */ 206 void amdgpu_gmc_vram_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc, 207 u64 base) 208 { 209 uint64_t vis_limit = (uint64_t)amdgpu_vis_vram_limit << 20; 210 uint64_t limit = (uint64_t)amdgpu_vram_limit << 20; 211 212 mc->vram_start = base; 213 mc->vram_end = mc->vram_start + mc->mc_vram_size - 1; 214 if (limit < mc->real_vram_size) 215 mc->real_vram_size = limit; 216 217 if (vis_limit && vis_limit < mc->visible_vram_size) 218 mc->visible_vram_size = vis_limit; 219 220 if (mc->real_vram_size < mc->visible_vram_size) 221 mc->visible_vram_size = mc->real_vram_size; 222 223 if (mc->xgmi.num_physical_nodes == 0) { 224 mc->fb_start = mc->vram_start; 225 mc->fb_end = mc->vram_end; 226 } 227 dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n", 228 mc->mc_vram_size >> 20, mc->vram_start, 229 mc->vram_end, mc->real_vram_size >> 20); 230 } 231 232 /** amdgpu_gmc_sysvm_location - place vram and gart in sysvm aperture 233 * 234 * @adev: amdgpu device structure holding all necessary information 235 * @mc: memory controller structure holding memory information 236 * 237 * This function is only used if use GART for FB translation. In such 238 * case, we use sysvm aperture (vmid0 page tables) for both vram 239 * and gart (aka system memory) access. 240 * 241 * GPUVM (and our organization of vmid0 page tables) require sysvm 242 * aperture to be placed at a location aligned with 8 times of native 243 * page size. For example, if vm_context0_cntl.page_table_block_size 244 * is 12, then native page size is 8G (2M*2^12), sysvm should start 245 * with a 64G aligned address. For simplicity, we just put sysvm at 246 * address 0. So vram start at address 0 and gart is right after vram. 247 */ 248 void amdgpu_gmc_sysvm_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc) 249 { 250 u64 hive_vram_start = 0; 251 u64 hive_vram_end = mc->xgmi.node_segment_size * mc->xgmi.num_physical_nodes - 1; 252 mc->vram_start = mc->xgmi.node_segment_size * mc->xgmi.physical_node_id; 253 mc->vram_end = mc->vram_start + mc->xgmi.node_segment_size - 1; 254 mc->gart_start = hive_vram_end + 1; 255 mc->gart_end = mc->gart_start + mc->gart_size - 1; 256 mc->fb_start = hive_vram_start; 257 mc->fb_end = hive_vram_end; 258 dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n", 259 mc->mc_vram_size >> 20, mc->vram_start, 260 mc->vram_end, mc->real_vram_size >> 20); 261 dev_info(adev->dev, "GART: %lluM 0x%016llX - 0x%016llX\n", 262 mc->gart_size >> 20, mc->gart_start, mc->gart_end); 263 } 264 265 /** 266 * amdgpu_gmc_gart_location - try to find GART location 267 * 268 * @adev: amdgpu device structure holding all necessary information 269 * @mc: memory controller structure holding memory information 270 * @gart_placement: GART placement policy with respect to VRAM 271 * 272 * Function will place try to place GART before or after VRAM. 273 * If GART size is bigger than space left then we ajust GART size. 274 * Thus function will never fails. 275 */ 276 void amdgpu_gmc_gart_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc, 277 enum amdgpu_gart_placement gart_placement) 278 { 279 const uint64_t four_gb = 0x100000000ULL; 280 u64 size_af, size_bf; 281 /*To avoid the hole, limit the max mc address to AMDGPU_GMC_HOLE_START*/ 282 u64 max_mc_address = min(adev->gmc.mc_mask, AMDGPU_GMC_HOLE_START - 1); 283 284 /* VCE doesn't like it when BOs cross a 4GB segment, so align 285 * the GART base on a 4GB boundary as well. 286 */ 287 size_bf = mc->fb_start; 288 size_af = max_mc_address + 1 - ALIGN(mc->fb_end + 1, four_gb); 289 290 if (mc->gart_size > max(size_bf, size_af)) { 291 dev_warn(adev->dev, "limiting GART\n"); 292 mc->gart_size = max(size_bf, size_af); 293 } 294 295 switch (gart_placement) { 296 case AMDGPU_GART_PLACEMENT_HIGH: 297 mc->gart_start = max_mc_address - mc->gart_size + 1; 298 break; 299 case AMDGPU_GART_PLACEMENT_LOW: 300 mc->gart_start = 0; 301 break; 302 case AMDGPU_GART_PLACEMENT_BEST_FIT: 303 default: 304 if ((size_bf >= mc->gart_size && size_bf < size_af) || 305 (size_af < mc->gart_size)) 306 mc->gart_start = 0; 307 else 308 mc->gart_start = max_mc_address - mc->gart_size + 1; 309 break; 310 } 311 312 mc->gart_start &= ~(four_gb - 1); 313 mc->gart_end = mc->gart_start + mc->gart_size - 1; 314 dev_info(adev->dev, "GART: %lluM 0x%016llX - 0x%016llX\n", 315 mc->gart_size >> 20, mc->gart_start, mc->gart_end); 316 } 317 318 /** 319 * amdgpu_gmc_agp_location - try to find AGP location 320 * @adev: amdgpu device structure holding all necessary information 321 * @mc: memory controller structure holding memory information 322 * 323 * Function will place try to find a place for the AGP BAR in the MC address 324 * space. 325 * 326 * AGP BAR will be assigned the largest available hole in the address space. 327 * Should be called after VRAM and GART locations are setup. 328 */ 329 void amdgpu_gmc_agp_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc) 330 { 331 const uint64_t sixteen_gb = 1ULL << 34; 332 const uint64_t sixteen_gb_mask = ~(sixteen_gb - 1); 333 u64 size_af, size_bf; 334 335 if (mc->fb_start > mc->gart_start) { 336 size_bf = (mc->fb_start & sixteen_gb_mask) - 337 ALIGN(mc->gart_end + 1, sixteen_gb); 338 size_af = mc->mc_mask + 1 - ALIGN(mc->fb_end + 1, sixteen_gb); 339 } else { 340 size_bf = mc->fb_start & sixteen_gb_mask; 341 size_af = (mc->gart_start & sixteen_gb_mask) - 342 ALIGN(mc->fb_end + 1, sixteen_gb); 343 } 344 345 if (size_bf > size_af) { 346 mc->agp_start = (mc->fb_start - size_bf) & sixteen_gb_mask; 347 mc->agp_size = size_bf; 348 } else { 349 mc->agp_start = ALIGN(mc->fb_end + 1, sixteen_gb); 350 mc->agp_size = size_af; 351 } 352 353 mc->agp_end = mc->agp_start + mc->agp_size - 1; 354 dev_info(adev->dev, "AGP: %lluM 0x%016llX - 0x%016llX\n", 355 mc->agp_size >> 20, mc->agp_start, mc->agp_end); 356 } 357 358 /** 359 * amdgpu_gmc_set_agp_default - Set the default AGP aperture value. 360 * @adev: amdgpu device structure holding all necessary information 361 * @mc: memory controller structure holding memory information 362 * 363 * To disable the AGP aperture, you need to set the start to a larger 364 * value than the end. This function sets the default value which 365 * can then be overridden using amdgpu_gmc_agp_location() if you want 366 * to enable the AGP aperture on a specific chip. 367 * 368 */ 369 void amdgpu_gmc_set_agp_default(struct amdgpu_device *adev, 370 struct amdgpu_gmc *mc) 371 { 372 mc->agp_start = 0xffffffffffff; 373 mc->agp_end = 0; 374 mc->agp_size = 0; 375 } 376 377 /** 378 * amdgpu_gmc_fault_key - get hask key from vm fault address and pasid 379 * 380 * @addr: 48 bit physical address, page aligned (36 significant bits) 381 * @pasid: 16 bit process address space identifier 382 */ 383 static inline uint64_t amdgpu_gmc_fault_key(uint64_t addr, uint16_t pasid) 384 { 385 return addr << 4 | pasid; 386 } 387 388 /** 389 * amdgpu_gmc_filter_faults - filter VM faults 390 * 391 * @adev: amdgpu device structure 392 * @ih: interrupt ring that the fault received from 393 * @addr: address of the VM fault 394 * @pasid: PASID of the process causing the fault 395 * @timestamp: timestamp of the fault 396 * 397 * Returns: 398 * True if the fault was filtered and should not be processed further. 399 * False if the fault is a new one and needs to be handled. 400 */ 401 bool amdgpu_gmc_filter_faults(struct amdgpu_device *adev, 402 struct amdgpu_ih_ring *ih, uint64_t addr, 403 uint16_t pasid, uint64_t timestamp) 404 { 405 struct amdgpu_gmc *gmc = &adev->gmc; 406 uint64_t stamp, key = amdgpu_gmc_fault_key(addr, pasid); 407 struct amdgpu_gmc_fault *fault; 408 uint32_t hash; 409 410 /* Stale retry fault if timestamp goes backward */ 411 if (amdgpu_ih_ts_after(timestamp, ih->processed_timestamp)) 412 return true; 413 414 /* If we don't have space left in the ring buffer return immediately */ 415 stamp = max(timestamp, AMDGPU_GMC_FAULT_TIMEOUT + 1) - 416 AMDGPU_GMC_FAULT_TIMEOUT; 417 if (gmc->fault_ring[gmc->last_fault].timestamp >= stamp) 418 return true; 419 420 /* Try to find the fault in the hash */ 421 hash = hash_64(key, AMDGPU_GMC_FAULT_HASH_ORDER); 422 fault = &gmc->fault_ring[gmc->fault_hash[hash].idx]; 423 while (fault->timestamp >= stamp) { 424 uint64_t tmp; 425 426 if (atomic64_read(&fault->key) == key) { 427 /* 428 * if we get a fault which is already present in 429 * the fault_ring and the timestamp of 430 * the fault is after the expired timestamp, 431 * then this is a new fault that needs to be added 432 * into the fault ring. 433 */ 434 if (fault->timestamp_expiry != 0 && 435 amdgpu_ih_ts_after(fault->timestamp_expiry, 436 timestamp)) 437 break; 438 else 439 return true; 440 } 441 442 tmp = fault->timestamp; 443 fault = &gmc->fault_ring[fault->next]; 444 445 /* Check if the entry was reused */ 446 if (fault->timestamp >= tmp) 447 break; 448 } 449 450 /* Add the fault to the ring */ 451 fault = &gmc->fault_ring[gmc->last_fault]; 452 atomic64_set(&fault->key, key); 453 fault->timestamp = timestamp; 454 455 /* And update the hash */ 456 fault->next = gmc->fault_hash[hash].idx; 457 gmc->fault_hash[hash].idx = gmc->last_fault++; 458 return false; 459 } 460 461 /** 462 * amdgpu_gmc_filter_faults_remove - remove address from VM faults filter 463 * 464 * @adev: amdgpu device structure 465 * @addr: address of the VM fault 466 * @pasid: PASID of the process causing the fault 467 * 468 * Remove the address from fault filter, then future vm fault on this address 469 * will pass to retry fault handler to recover. 470 */ 471 void amdgpu_gmc_filter_faults_remove(struct amdgpu_device *adev, uint64_t addr, 472 uint16_t pasid) 473 { 474 struct amdgpu_gmc *gmc = &adev->gmc; 475 uint64_t key = amdgpu_gmc_fault_key(addr, pasid); 476 struct amdgpu_ih_ring *ih; 477 struct amdgpu_gmc_fault *fault; 478 uint32_t last_wptr; 479 uint64_t last_ts; 480 uint32_t hash; 481 uint64_t tmp; 482 483 if (adev->irq.retry_cam_enabled) 484 return; 485 486 ih = &adev->irq.ih1; 487 /* Get the WPTR of the last entry in IH ring */ 488 last_wptr = amdgpu_ih_get_wptr(adev, ih); 489 /* Order wptr with ring data. */ 490 rmb(); 491 /* Get the timetamp of the last entry in IH ring */ 492 last_ts = amdgpu_ih_decode_iv_ts(adev, ih, last_wptr, -1); 493 494 hash = hash_64(key, AMDGPU_GMC_FAULT_HASH_ORDER); 495 fault = &gmc->fault_ring[gmc->fault_hash[hash].idx]; 496 do { 497 if (atomic64_read(&fault->key) == key) { 498 /* 499 * Update the timestamp when this fault 500 * expired. 501 */ 502 fault->timestamp_expiry = last_ts; 503 break; 504 } 505 506 tmp = fault->timestamp; 507 fault = &gmc->fault_ring[fault->next]; 508 } while (fault->timestamp < tmp); 509 } 510 511 int amdgpu_gmc_ras_sw_init(struct amdgpu_device *adev) 512 { 513 int r; 514 515 /* umc ras block */ 516 r = amdgpu_umc_ras_sw_init(adev); 517 if (r) 518 return r; 519 520 /* mmhub ras block */ 521 r = amdgpu_mmhub_ras_sw_init(adev); 522 if (r) 523 return r; 524 525 /* hdp ras block */ 526 r = amdgpu_hdp_ras_sw_init(adev); 527 if (r) 528 return r; 529 530 /* mca.x ras block */ 531 r = amdgpu_mca_mp0_ras_sw_init(adev); 532 if (r) 533 return r; 534 535 r = amdgpu_mca_mp1_ras_sw_init(adev); 536 if (r) 537 return r; 538 539 r = amdgpu_mca_mpio_ras_sw_init(adev); 540 if (r) 541 return r; 542 543 /* xgmi ras block */ 544 r = amdgpu_xgmi_ras_sw_init(adev); 545 if (r) 546 return r; 547 548 return 0; 549 } 550 551 int amdgpu_gmc_ras_late_init(struct amdgpu_device *adev) 552 { 553 return 0; 554 } 555 556 void amdgpu_gmc_ras_fini(struct amdgpu_device *adev) 557 { 558 559 } 560 561 /* 562 * The latest engine allocation on gfx9/10 is: 563 * Engine 2, 3: firmware 564 * Engine 0, 1, 4~16: amdgpu ring, 565 * subject to change when ring number changes 566 * Engine 17: Gart flushes 567 */ 568 #define AMDGPU_VMHUB_INV_ENG_BITMAP 0x1FFF3 569 570 int amdgpu_gmc_allocate_vm_inv_eng(struct amdgpu_device *adev) 571 { 572 struct amdgpu_ring *ring; 573 unsigned vm_inv_engs[AMDGPU_MAX_VMHUBS] = {0}; 574 unsigned i; 575 unsigned vmhub, inv_eng; 576 577 /* init the vm inv eng for all vmhubs */ 578 for_each_set_bit(i, adev->vmhubs_mask, AMDGPU_MAX_VMHUBS) { 579 vm_inv_engs[i] = AMDGPU_VMHUB_INV_ENG_BITMAP; 580 /* reserve engine 5 for firmware */ 581 if (adev->enable_mes) 582 vm_inv_engs[i] &= ~(1 << 5); 583 /* reserve mmhub engine 3 for firmware */ 584 if (adev->enable_umsch_mm) 585 vm_inv_engs[i] &= ~(1 << 3); 586 } 587 588 for (i = 0; i < adev->num_rings; ++i) { 589 ring = adev->rings[i]; 590 vmhub = ring->vm_hub; 591 592 if (ring == &adev->mes.ring || 593 ring == &adev->umsch_mm.ring) 594 continue; 595 596 inv_eng = ffs(vm_inv_engs[vmhub]); 597 if (!inv_eng) { 598 dev_err(adev->dev, "no VM inv eng for ring %s\n", 599 ring->name); 600 return -EINVAL; 601 } 602 603 ring->vm_inv_eng = inv_eng - 1; 604 vm_inv_engs[vmhub] &= ~(1 << ring->vm_inv_eng); 605 606 dev_info(adev->dev, "ring %s uses VM inv eng %u on hub %u\n", 607 ring->name, ring->vm_inv_eng, ring->vm_hub); 608 } 609 610 return 0; 611 } 612 613 void amdgpu_gmc_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid, 614 uint32_t vmhub, uint32_t flush_type) 615 { 616 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring; 617 struct amdgpu_vmhub *hub = &adev->vmhub[vmhub]; 618 struct dma_fence *fence; 619 struct amdgpu_job *job; 620 int r; 621 622 if (!hub->sdma_invalidation_workaround || vmid || 623 !adev->mman.buffer_funcs_enabled || 624 !adev->ib_pool_ready || amdgpu_in_reset(adev) || 625 !ring->sched.ready) { 626 627 /* 628 * A GPU reset should flush all TLBs anyway, so no need to do 629 * this while one is ongoing. 630 */ 631 if (!down_read_trylock(&adev->reset_domain->sem)) 632 return; 633 634 if (adev->gmc.flush_tlb_needs_extra_type_2) 635 adev->gmc.gmc_funcs->flush_gpu_tlb(adev, vmid, 636 vmhub, 2); 637 638 if (adev->gmc.flush_tlb_needs_extra_type_0 && flush_type == 2) 639 adev->gmc.gmc_funcs->flush_gpu_tlb(adev, vmid, 640 vmhub, 0); 641 642 adev->gmc.gmc_funcs->flush_gpu_tlb(adev, vmid, vmhub, 643 flush_type); 644 up_read(&adev->reset_domain->sem); 645 return; 646 } 647 648 /* The SDMA on Navi 1x has a bug which can theoretically result in memory 649 * corruption if an invalidation happens at the same time as an VA 650 * translation. Avoid this by doing the invalidation from the SDMA 651 * itself at least for GART. 652 */ 653 mutex_lock(&adev->mman.gtt_window_lock); 654 r = amdgpu_job_alloc_with_ib(ring->adev, &adev->mman.high_pr, 655 AMDGPU_FENCE_OWNER_UNDEFINED, 656 16 * 4, AMDGPU_IB_POOL_IMMEDIATE, 657 &job); 658 if (r) 659 goto error_alloc; 660 661 job->vm_pd_addr = amdgpu_gmc_pd_addr(adev->gart.bo); 662 job->vm_needs_flush = true; 663 job->ibs->ptr[job->ibs->length_dw++] = ring->funcs->nop; 664 amdgpu_ring_pad_ib(ring, &job->ibs[0]); 665 fence = amdgpu_job_submit(job); 666 mutex_unlock(&adev->mman.gtt_window_lock); 667 668 dma_fence_wait(fence, false); 669 dma_fence_put(fence); 670 671 return; 672 673 error_alloc: 674 mutex_unlock(&adev->mman.gtt_window_lock); 675 dev_err(adev->dev, "Error flushing GPU TLB using the SDMA (%d)!\n", r); 676 } 677 678 int amdgpu_gmc_flush_gpu_tlb_pasid(struct amdgpu_device *adev, uint16_t pasid, 679 uint32_t flush_type, bool all_hub, 680 uint32_t inst) 681 { 682 u32 usec_timeout = amdgpu_sriov_vf(adev) ? SRIOV_USEC_TIMEOUT : 683 adev->usec_timeout; 684 struct amdgpu_ring *ring = &adev->gfx.kiq[inst].ring; 685 struct amdgpu_kiq *kiq = &adev->gfx.kiq[inst]; 686 unsigned int ndw; 687 signed long r; 688 uint32_t seq; 689 690 if (!adev->gmc.flush_pasid_uses_kiq || !ring->sched.ready || 691 !down_read_trylock(&adev->reset_domain->sem)) { 692 693 if (adev->gmc.flush_tlb_needs_extra_type_2) 694 adev->gmc.gmc_funcs->flush_gpu_tlb_pasid(adev, pasid, 695 2, all_hub, 696 inst); 697 698 if (adev->gmc.flush_tlb_needs_extra_type_0 && flush_type == 2) 699 adev->gmc.gmc_funcs->flush_gpu_tlb_pasid(adev, pasid, 700 0, all_hub, 701 inst); 702 703 adev->gmc.gmc_funcs->flush_gpu_tlb_pasid(adev, pasid, 704 flush_type, all_hub, 705 inst); 706 return 0; 707 } 708 709 /* 2 dwords flush + 8 dwords fence */ 710 ndw = kiq->pmf->invalidate_tlbs_size + 8; 711 712 if (adev->gmc.flush_tlb_needs_extra_type_2) 713 ndw += kiq->pmf->invalidate_tlbs_size; 714 715 if (adev->gmc.flush_tlb_needs_extra_type_0) 716 ndw += kiq->pmf->invalidate_tlbs_size; 717 718 spin_lock(&adev->gfx.kiq[inst].ring_lock); 719 amdgpu_ring_alloc(ring, ndw); 720 if (adev->gmc.flush_tlb_needs_extra_type_2) 721 kiq->pmf->kiq_invalidate_tlbs(ring, pasid, 2, all_hub); 722 723 if (flush_type == 2 && adev->gmc.flush_tlb_needs_extra_type_0) 724 kiq->pmf->kiq_invalidate_tlbs(ring, pasid, 0, all_hub); 725 726 kiq->pmf->kiq_invalidate_tlbs(ring, pasid, flush_type, all_hub); 727 r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT); 728 if (r) { 729 amdgpu_ring_undo(ring); 730 spin_unlock(&adev->gfx.kiq[inst].ring_lock); 731 goto error_unlock_reset; 732 } 733 734 amdgpu_ring_commit(ring); 735 spin_unlock(&adev->gfx.kiq[inst].ring_lock); 736 r = amdgpu_fence_wait_polling(ring, seq, usec_timeout); 737 if (r < 1) { 738 dev_err(adev->dev, "wait for kiq fence error: %ld.\n", r); 739 r = -ETIME; 740 goto error_unlock_reset; 741 } 742 r = 0; 743 744 error_unlock_reset: 745 up_read(&adev->reset_domain->sem); 746 return r; 747 } 748 749 /** 750 * amdgpu_gmc_tmz_set -- check and set if a device supports TMZ 751 * @adev: amdgpu_device pointer 752 * 753 * Check and set if an the device @adev supports Trusted Memory 754 * Zones (TMZ). 755 */ 756 void amdgpu_gmc_tmz_set(struct amdgpu_device *adev) 757 { 758 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 759 /* RAVEN */ 760 case IP_VERSION(9, 2, 2): 761 case IP_VERSION(9, 1, 0): 762 /* RENOIR looks like RAVEN */ 763 case IP_VERSION(9, 3, 0): 764 /* GC 10.3.7 */ 765 case IP_VERSION(10, 3, 7): 766 /* GC 11.0.1 */ 767 case IP_VERSION(11, 0, 1): 768 if (amdgpu_tmz == 0) { 769 adev->gmc.tmz_enabled = false; 770 dev_info(adev->dev, 771 "Trusted Memory Zone (TMZ) feature disabled (cmd line)\n"); 772 } else { 773 adev->gmc.tmz_enabled = true; 774 dev_info(adev->dev, 775 "Trusted Memory Zone (TMZ) feature enabled\n"); 776 } 777 break; 778 case IP_VERSION(10, 1, 10): 779 case IP_VERSION(10, 1, 1): 780 case IP_VERSION(10, 1, 2): 781 case IP_VERSION(10, 1, 3): 782 case IP_VERSION(10, 3, 0): 783 case IP_VERSION(10, 3, 2): 784 case IP_VERSION(10, 3, 4): 785 case IP_VERSION(10, 3, 5): 786 case IP_VERSION(10, 3, 6): 787 /* VANGOGH */ 788 case IP_VERSION(10, 3, 1): 789 /* YELLOW_CARP*/ 790 case IP_VERSION(10, 3, 3): 791 case IP_VERSION(11, 0, 4): 792 case IP_VERSION(11, 5, 0): 793 /* Don't enable it by default yet. 794 */ 795 if (amdgpu_tmz < 1) { 796 adev->gmc.tmz_enabled = false; 797 dev_info(adev->dev, 798 "Trusted Memory Zone (TMZ) feature disabled as experimental (default)\n"); 799 } else { 800 adev->gmc.tmz_enabled = true; 801 dev_info(adev->dev, 802 "Trusted Memory Zone (TMZ) feature enabled as experimental (cmd line)\n"); 803 } 804 break; 805 default: 806 adev->gmc.tmz_enabled = false; 807 dev_info(adev->dev, 808 "Trusted Memory Zone (TMZ) feature not supported\n"); 809 break; 810 } 811 } 812 813 /** 814 * amdgpu_gmc_noretry_set -- set per asic noretry defaults 815 * @adev: amdgpu_device pointer 816 * 817 * Set a per asic default for the no-retry parameter. 818 * 819 */ 820 void amdgpu_gmc_noretry_set(struct amdgpu_device *adev) 821 { 822 struct amdgpu_gmc *gmc = &adev->gmc; 823 uint32_t gc_ver = amdgpu_ip_version(adev, GC_HWIP, 0); 824 bool noretry_default = (gc_ver == IP_VERSION(9, 0, 1) || 825 gc_ver == IP_VERSION(9, 3, 0) || 826 gc_ver == IP_VERSION(9, 4, 0) || 827 gc_ver == IP_VERSION(9, 4, 1) || 828 gc_ver == IP_VERSION(9, 4, 2) || 829 gc_ver == IP_VERSION(9, 4, 3) || 830 gc_ver >= IP_VERSION(10, 3, 0)); 831 832 if (!amdgpu_sriov_xnack_support(adev)) 833 gmc->noretry = 1; 834 else 835 gmc->noretry = (amdgpu_noretry == -1) ? noretry_default : amdgpu_noretry; 836 } 837 838 void amdgpu_gmc_set_vm_fault_masks(struct amdgpu_device *adev, int hub_type, 839 bool enable) 840 { 841 struct amdgpu_vmhub *hub; 842 u32 tmp, reg, i; 843 844 hub = &adev->vmhub[hub_type]; 845 for (i = 0; i < 16; i++) { 846 reg = hub->vm_context0_cntl + hub->ctx_distance * i; 847 848 tmp = (hub_type == AMDGPU_GFXHUB(0)) ? 849 RREG32_SOC15_IP(GC, reg) : 850 RREG32_SOC15_IP(MMHUB, reg); 851 852 if (enable) 853 tmp |= hub->vm_cntx_cntl_vm_fault; 854 else 855 tmp &= ~hub->vm_cntx_cntl_vm_fault; 856 857 (hub_type == AMDGPU_GFXHUB(0)) ? 858 WREG32_SOC15_IP(GC, reg, tmp) : 859 WREG32_SOC15_IP(MMHUB, reg, tmp); 860 } 861 } 862 863 void amdgpu_gmc_get_vbios_allocations(struct amdgpu_device *adev) 864 { 865 unsigned size; 866 867 /* 868 * Some ASICs need to reserve a region of video memory to avoid access 869 * from driver 870 */ 871 adev->mman.stolen_reserved_offset = 0; 872 adev->mman.stolen_reserved_size = 0; 873 874 /* 875 * TODO: 876 * Currently there is a bug where some memory client outside 877 * of the driver writes to first 8M of VRAM on S3 resume, 878 * this overrides GART which by default gets placed in first 8M and 879 * causes VM_FAULTS once GTT is accessed. 880 * Keep the stolen memory reservation until the while this is not solved. 881 */ 882 switch (adev->asic_type) { 883 case CHIP_VEGA10: 884 adev->mman.keep_stolen_vga_memory = true; 885 /* 886 * VEGA10 SRIOV VF with MS_HYPERV host needs some firmware reserved area. 887 */ 888 #ifdef CONFIG_X86 889 if (amdgpu_sriov_vf(adev) && hypervisor_is_type(X86_HYPER_MS_HYPERV)) { 890 adev->mman.stolen_reserved_offset = 0x500000; 891 adev->mman.stolen_reserved_size = 0x200000; 892 } 893 #endif 894 break; 895 case CHIP_RAVEN: 896 case CHIP_RENOIR: 897 adev->mman.keep_stolen_vga_memory = true; 898 break; 899 default: 900 adev->mman.keep_stolen_vga_memory = false; 901 break; 902 } 903 904 if (amdgpu_sriov_vf(adev) || 905 !amdgpu_device_has_display_hardware(adev)) { 906 size = 0; 907 } else { 908 size = amdgpu_gmc_get_vbios_fb_size(adev); 909 910 if (adev->mman.keep_stolen_vga_memory) 911 size = max(size, (unsigned)AMDGPU_VBIOS_VGA_ALLOCATION); 912 } 913 914 /* set to 0 if the pre-OS buffer uses up most of vram */ 915 if ((adev->gmc.real_vram_size - size) < (8 * 1024 * 1024)) 916 size = 0; 917 918 if (size > AMDGPU_VBIOS_VGA_ALLOCATION) { 919 adev->mman.stolen_vga_size = AMDGPU_VBIOS_VGA_ALLOCATION; 920 adev->mman.stolen_extended_size = size - adev->mman.stolen_vga_size; 921 } else { 922 adev->mman.stolen_vga_size = size; 923 adev->mman.stolen_extended_size = 0; 924 } 925 } 926 927 /** 928 * amdgpu_gmc_init_pdb0 - initialize PDB0 929 * 930 * @adev: amdgpu_device pointer 931 * 932 * This function is only used when GART page table is used 933 * for FB address translatioin. In such a case, we construct 934 * a 2-level system VM page table: PDB0->PTB, to cover both 935 * VRAM of the hive and system memory. 936 * 937 * PDB0 is static, initialized once on driver initialization. 938 * The first n entries of PDB0 are used as PTE by setting 939 * P bit to 1, pointing to VRAM. The n+1'th entry points 940 * to a big PTB covering system memory. 941 * 942 */ 943 void amdgpu_gmc_init_pdb0(struct amdgpu_device *adev) 944 { 945 int i; 946 uint64_t flags = adev->gart.gart_pte_flags; //TODO it is UC. explore NC/RW? 947 /* Each PDE0 (used as PTE) covers (2^vmid0_page_table_block_size)*2M 948 */ 949 u64 vram_size = adev->gmc.xgmi.node_segment_size * adev->gmc.xgmi.num_physical_nodes; 950 u64 pde0_page_size = (1ULL<<adev->gmc.vmid0_page_table_block_size)<<21; 951 u64 vram_addr = adev->vm_manager.vram_base_offset - 952 adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size; 953 u64 vram_end = vram_addr + vram_size; 954 u64 gart_ptb_gpu_pa = amdgpu_gmc_vram_pa(adev, adev->gart.bo); 955 int idx; 956 957 if (!drm_dev_enter(adev_to_drm(adev), &idx)) 958 return; 959 960 flags |= AMDGPU_PTE_VALID | AMDGPU_PTE_READABLE; 961 flags |= AMDGPU_PTE_WRITEABLE; 962 flags |= AMDGPU_PTE_SNOOPED; 963 flags |= AMDGPU_PTE_FRAG((adev->gmc.vmid0_page_table_block_size + 9*1)); 964 flags |= AMDGPU_PDE_PTE; 965 966 /* The first n PDE0 entries are used as PTE, 967 * pointing to vram 968 */ 969 for (i = 0; vram_addr < vram_end; i++, vram_addr += pde0_page_size) 970 amdgpu_gmc_set_pte_pde(adev, adev->gmc.ptr_pdb0, i, vram_addr, flags); 971 972 /* The n+1'th PDE0 entry points to a huge 973 * PTB who has more than 512 entries each 974 * pointing to a 4K system page 975 */ 976 flags = AMDGPU_PTE_VALID; 977 flags |= AMDGPU_PDE_BFS(0) | AMDGPU_PTE_SNOOPED; 978 /* Requires gart_ptb_gpu_pa to be 4K aligned */ 979 amdgpu_gmc_set_pte_pde(adev, adev->gmc.ptr_pdb0, i, gart_ptb_gpu_pa, flags); 980 drm_dev_exit(idx); 981 } 982 983 /** 984 * amdgpu_gmc_vram_mc2pa - calculate vram buffer's physical address from MC 985 * address 986 * 987 * @adev: amdgpu_device pointer 988 * @mc_addr: MC address of buffer 989 */ 990 uint64_t amdgpu_gmc_vram_mc2pa(struct amdgpu_device *adev, uint64_t mc_addr) 991 { 992 return mc_addr - adev->gmc.vram_start + adev->vm_manager.vram_base_offset; 993 } 994 995 /** 996 * amdgpu_gmc_vram_pa - calculate vram buffer object's physical address from 997 * GPU's view 998 * 999 * @adev: amdgpu_device pointer 1000 * @bo: amdgpu buffer object 1001 */ 1002 uint64_t amdgpu_gmc_vram_pa(struct amdgpu_device *adev, struct amdgpu_bo *bo) 1003 { 1004 return amdgpu_gmc_vram_mc2pa(adev, amdgpu_bo_gpu_offset(bo)); 1005 } 1006 1007 /** 1008 * amdgpu_gmc_vram_cpu_pa - calculate vram buffer object's physical address 1009 * from CPU's view 1010 * 1011 * @adev: amdgpu_device pointer 1012 * @bo: amdgpu buffer object 1013 */ 1014 uint64_t amdgpu_gmc_vram_cpu_pa(struct amdgpu_device *adev, struct amdgpu_bo *bo) 1015 { 1016 return amdgpu_bo_gpu_offset(bo) - adev->gmc.vram_start + adev->gmc.aper_base; 1017 } 1018 1019 int amdgpu_gmc_vram_checking(struct amdgpu_device *adev) 1020 { 1021 struct amdgpu_bo *vram_bo = NULL; 1022 uint64_t vram_gpu = 0; 1023 void *vram_ptr = NULL; 1024 1025 int ret, size = 0x100000; 1026 uint8_t cptr[10]; 1027 1028 ret = amdgpu_bo_create_kernel(adev, size, PAGE_SIZE, 1029 AMDGPU_GEM_DOMAIN_VRAM, 1030 &vram_bo, 1031 &vram_gpu, 1032 &vram_ptr); 1033 if (ret) 1034 return ret; 1035 1036 memset(vram_ptr, 0x86, size); 1037 memset(cptr, 0x86, 10); 1038 1039 /** 1040 * Check the start, the mid, and the end of the memory if the content of 1041 * each byte is the pattern "0x86". If yes, we suppose the vram bo is 1042 * workable. 1043 * 1044 * Note: If check the each byte of whole 1M bo, it will cost too many 1045 * seconds, so here, we just pick up three parts for emulation. 1046 */ 1047 ret = memcmp(vram_ptr, cptr, 10); 1048 if (ret) { 1049 ret = -EIO; 1050 goto release_buffer; 1051 } 1052 1053 ret = memcmp(vram_ptr + (size / 2), cptr, 10); 1054 if (ret) { 1055 ret = -EIO; 1056 goto release_buffer; 1057 } 1058 1059 ret = memcmp(vram_ptr + size - 10, cptr, 10); 1060 if (ret) { 1061 ret = -EIO; 1062 goto release_buffer; 1063 } 1064 1065 release_buffer: 1066 amdgpu_bo_free_kernel(&vram_bo, &vram_gpu, 1067 &vram_ptr); 1068 1069 return ret; 1070 } 1071 1072 static ssize_t current_memory_partition_show( 1073 struct device *dev, struct device_attribute *addr, char *buf) 1074 { 1075 struct drm_device *ddev = dev_get_drvdata(dev); 1076 struct amdgpu_device *adev = drm_to_adev(ddev); 1077 enum amdgpu_memory_partition mode; 1078 1079 mode = adev->gmc.gmc_funcs->query_mem_partition_mode(adev); 1080 switch (mode) { 1081 case AMDGPU_NPS1_PARTITION_MODE: 1082 return sysfs_emit(buf, "NPS1\n"); 1083 case AMDGPU_NPS2_PARTITION_MODE: 1084 return sysfs_emit(buf, "NPS2\n"); 1085 case AMDGPU_NPS3_PARTITION_MODE: 1086 return sysfs_emit(buf, "NPS3\n"); 1087 case AMDGPU_NPS4_PARTITION_MODE: 1088 return sysfs_emit(buf, "NPS4\n"); 1089 case AMDGPU_NPS6_PARTITION_MODE: 1090 return sysfs_emit(buf, "NPS6\n"); 1091 case AMDGPU_NPS8_PARTITION_MODE: 1092 return sysfs_emit(buf, "NPS8\n"); 1093 default: 1094 return sysfs_emit(buf, "UNKNOWN\n"); 1095 } 1096 1097 return sysfs_emit(buf, "UNKNOWN\n"); 1098 } 1099 1100 static DEVICE_ATTR_RO(current_memory_partition); 1101 1102 int amdgpu_gmc_sysfs_init(struct amdgpu_device *adev) 1103 { 1104 if (!adev->gmc.gmc_funcs->query_mem_partition_mode) 1105 return 0; 1106 1107 return device_create_file(adev->dev, 1108 &dev_attr_current_memory_partition); 1109 } 1110 1111 void amdgpu_gmc_sysfs_fini(struct amdgpu_device *adev) 1112 { 1113 device_remove_file(adev->dev, &dev_attr_current_memory_partition); 1114 } 1115