1 /* 2 * Copyright 2009 Jerome Glisse. 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 * Authors: 28 * Jerome Glisse <glisse@freedesktop.org> 29 * Thomas Hellstrom <thomas-at-tungstengraphics-dot-com> 30 * Dave Airlie 31 */ 32 33 #include <linux/dma-mapping.h> 34 #include <linux/iommu.h> 35 #include <linux/hmm.h> 36 #include <linux/pagemap.h> 37 #include <linux/sched/task.h> 38 #include <linux/sched/mm.h> 39 #include <linux/seq_file.h> 40 #include <linux/slab.h> 41 #include <linux/swap.h> 42 #include <linux/swiotlb.h> 43 #include <linux/dma-buf.h> 44 #include <linux/sizes.h> 45 46 #include <drm/ttm/ttm_bo_api.h> 47 #include <drm/ttm/ttm_bo_driver.h> 48 #include <drm/ttm/ttm_placement.h> 49 #include <drm/ttm/ttm_module.h> 50 51 #include <drm/drm_debugfs.h> 52 #include <drm/amdgpu_drm.h> 53 54 #include "amdgpu.h" 55 #include "amdgpu_object.h" 56 #include "amdgpu_trace.h" 57 #include "amdgpu_amdkfd.h" 58 #include "amdgpu_sdma.h" 59 #include "amdgpu_ras.h" 60 #include "amdgpu_atomfirmware.h" 61 #include "bif/bif_4_1_d.h" 62 63 #define AMDGPU_TTM_VRAM_MAX_DW_READ (size_t)128 64 65 static int amdgpu_ttm_backend_bind(struct ttm_bo_device *bdev, 66 struct ttm_tt *ttm, 67 struct ttm_resource *bo_mem); 68 static void amdgpu_ttm_backend_unbind(struct ttm_bo_device *bdev, 69 struct ttm_tt *ttm); 70 71 static int amdgpu_ttm_init_on_chip(struct amdgpu_device *adev, 72 unsigned int type, 73 uint64_t size_in_page) 74 { 75 return ttm_range_man_init(&adev->mman.bdev, type, 76 false, size_in_page); 77 } 78 79 /** 80 * amdgpu_evict_flags - Compute placement flags 81 * 82 * @bo: The buffer object to evict 83 * @placement: Possible destination(s) for evicted BO 84 * 85 * Fill in placement data when ttm_bo_evict() is called 86 */ 87 static void amdgpu_evict_flags(struct ttm_buffer_object *bo, 88 struct ttm_placement *placement) 89 { 90 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev); 91 struct amdgpu_bo *abo; 92 static const struct ttm_place placements = { 93 .fpfn = 0, 94 .lpfn = 0, 95 .mem_type = TTM_PL_SYSTEM, 96 .flags = 0 97 }; 98 99 /* Don't handle scatter gather BOs */ 100 if (bo->type == ttm_bo_type_sg) { 101 placement->num_placement = 0; 102 placement->num_busy_placement = 0; 103 return; 104 } 105 106 /* Object isn't an AMDGPU object so ignore */ 107 if (!amdgpu_bo_is_amdgpu_bo(bo)) { 108 placement->placement = &placements; 109 placement->busy_placement = &placements; 110 placement->num_placement = 1; 111 placement->num_busy_placement = 1; 112 return; 113 } 114 115 abo = ttm_to_amdgpu_bo(bo); 116 switch (bo->mem.mem_type) { 117 case AMDGPU_PL_GDS: 118 case AMDGPU_PL_GWS: 119 case AMDGPU_PL_OA: 120 placement->num_placement = 0; 121 placement->num_busy_placement = 0; 122 return; 123 124 case TTM_PL_VRAM: 125 if (!adev->mman.buffer_funcs_enabled) { 126 /* Move to system memory */ 127 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU); 128 } else if (!amdgpu_gmc_vram_full_visible(&adev->gmc) && 129 !(abo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED) && 130 amdgpu_bo_in_cpu_visible_vram(abo)) { 131 132 /* Try evicting to the CPU inaccessible part of VRAM 133 * first, but only set GTT as busy placement, so this 134 * BO will be evicted to GTT rather than causing other 135 * BOs to be evicted from VRAM 136 */ 137 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM | 138 AMDGPU_GEM_DOMAIN_GTT); 139 abo->placements[0].fpfn = adev->gmc.visible_vram_size >> PAGE_SHIFT; 140 abo->placements[0].lpfn = 0; 141 abo->placement.busy_placement = &abo->placements[1]; 142 abo->placement.num_busy_placement = 1; 143 } else { 144 /* Move to GTT memory */ 145 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_GTT); 146 } 147 break; 148 case TTM_PL_TT: 149 default: 150 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU); 151 break; 152 } 153 *placement = abo->placement; 154 } 155 156 /** 157 * amdgpu_verify_access - Verify access for a mmap call 158 * 159 * @bo: The buffer object to map 160 * @filp: The file pointer from the process performing the mmap 161 * 162 * This is called by ttm_bo_mmap() to verify whether a process 163 * has the right to mmap a BO to their process space. 164 */ 165 static int amdgpu_verify_access(struct ttm_buffer_object *bo, struct file *filp) 166 { 167 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo); 168 169 /* 170 * Don't verify access for KFD BOs. They don't have a GEM 171 * object associated with them. 172 */ 173 if (abo->kfd_bo) 174 return 0; 175 176 if (amdgpu_ttm_tt_get_usermm(bo->ttm)) 177 return -EPERM; 178 return drm_vma_node_verify_access(&abo->tbo.base.vma_node, 179 filp->private_data); 180 } 181 182 /** 183 * amdgpu_mm_node_addr - Compute the GPU relative offset of a GTT buffer. 184 * 185 * @bo: The bo to assign the memory to. 186 * @mm_node: Memory manager node for drm allocator. 187 * @mem: The region where the bo resides. 188 * 189 */ 190 static uint64_t amdgpu_mm_node_addr(struct ttm_buffer_object *bo, 191 struct drm_mm_node *mm_node, 192 struct ttm_resource *mem) 193 { 194 uint64_t addr = 0; 195 196 if (mm_node->start != AMDGPU_BO_INVALID_OFFSET) { 197 addr = mm_node->start << PAGE_SHIFT; 198 addr += amdgpu_ttm_domain_start(amdgpu_ttm_adev(bo->bdev), 199 mem->mem_type); 200 } 201 return addr; 202 } 203 204 /** 205 * amdgpu_find_mm_node - Helper function finds the drm_mm_node corresponding to 206 * @offset. It also modifies the offset to be within the drm_mm_node returned 207 * 208 * @mem: The region where the bo resides. 209 * @offset: The offset that drm_mm_node is used for finding. 210 * 211 */ 212 static struct drm_mm_node *amdgpu_find_mm_node(struct ttm_resource *mem, 213 uint64_t *offset) 214 { 215 struct drm_mm_node *mm_node = mem->mm_node; 216 217 while (*offset >= (mm_node->size << PAGE_SHIFT)) { 218 *offset -= (mm_node->size << PAGE_SHIFT); 219 ++mm_node; 220 } 221 return mm_node; 222 } 223 224 /** 225 * amdgpu_ttm_map_buffer - Map memory into the GART windows 226 * @bo: buffer object to map 227 * @mem: memory object to map 228 * @mm_node: drm_mm node object to map 229 * @num_pages: number of pages to map 230 * @offset: offset into @mm_node where to start 231 * @window: which GART window to use 232 * @ring: DMA ring to use for the copy 233 * @tmz: if we should setup a TMZ enabled mapping 234 * @addr: resulting address inside the MC address space 235 * 236 * Setup one of the GART windows to access a specific piece of memory or return 237 * the physical address for local memory. 238 */ 239 static int amdgpu_ttm_map_buffer(struct ttm_buffer_object *bo, 240 struct ttm_resource *mem, 241 struct drm_mm_node *mm_node, 242 unsigned num_pages, uint64_t offset, 243 unsigned window, struct amdgpu_ring *ring, 244 bool tmz, uint64_t *addr) 245 { 246 struct amdgpu_device *adev = ring->adev; 247 struct amdgpu_job *job; 248 unsigned num_dw, num_bytes; 249 struct dma_fence *fence; 250 uint64_t src_addr, dst_addr; 251 void *cpu_addr; 252 uint64_t flags; 253 unsigned int i; 254 int r; 255 256 BUG_ON(adev->mman.buffer_funcs->copy_max_bytes < 257 AMDGPU_GTT_MAX_TRANSFER_SIZE * 8); 258 259 /* Map only what can't be accessed directly */ 260 if (!tmz && mem->start != AMDGPU_BO_INVALID_OFFSET) { 261 *addr = amdgpu_mm_node_addr(bo, mm_node, mem) + offset; 262 return 0; 263 } 264 265 *addr = adev->gmc.gart_start; 266 *addr += (u64)window * AMDGPU_GTT_MAX_TRANSFER_SIZE * 267 AMDGPU_GPU_PAGE_SIZE; 268 *addr += offset & ~PAGE_MASK; 269 270 num_dw = ALIGN(adev->mman.buffer_funcs->copy_num_dw, 8); 271 num_bytes = num_pages * 8; 272 273 r = amdgpu_job_alloc_with_ib(adev, num_dw * 4 + num_bytes, 274 AMDGPU_IB_POOL_DELAYED, &job); 275 if (r) 276 return r; 277 278 src_addr = num_dw * 4; 279 src_addr += job->ibs[0].gpu_addr; 280 281 dst_addr = amdgpu_bo_gpu_offset(adev->gart.bo); 282 dst_addr += window * AMDGPU_GTT_MAX_TRANSFER_SIZE * 8; 283 amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_addr, 284 dst_addr, num_bytes, false); 285 286 amdgpu_ring_pad_ib(ring, &job->ibs[0]); 287 WARN_ON(job->ibs[0].length_dw > num_dw); 288 289 flags = amdgpu_ttm_tt_pte_flags(adev, bo->ttm, mem); 290 if (tmz) 291 flags |= AMDGPU_PTE_TMZ; 292 293 cpu_addr = &job->ibs[0].ptr[num_dw]; 294 295 if (mem->mem_type == TTM_PL_TT) { 296 dma_addr_t *dma_address; 297 298 dma_address = &bo->ttm->dma_address[offset >> PAGE_SHIFT]; 299 r = amdgpu_gart_map(adev, 0, num_pages, dma_address, flags, 300 cpu_addr); 301 if (r) 302 goto error_free; 303 } else { 304 dma_addr_t dma_address; 305 306 dma_address = (mm_node->start << PAGE_SHIFT) + offset; 307 dma_address += adev->vm_manager.vram_base_offset; 308 309 for (i = 0; i < num_pages; ++i) { 310 r = amdgpu_gart_map(adev, i << PAGE_SHIFT, 1, 311 &dma_address, flags, cpu_addr); 312 if (r) 313 goto error_free; 314 315 dma_address += PAGE_SIZE; 316 } 317 } 318 319 r = amdgpu_job_submit(job, &adev->mman.entity, 320 AMDGPU_FENCE_OWNER_UNDEFINED, &fence); 321 if (r) 322 goto error_free; 323 324 dma_fence_put(fence); 325 326 return r; 327 328 error_free: 329 amdgpu_job_free(job); 330 return r; 331 } 332 333 /** 334 * amdgpu_copy_ttm_mem_to_mem - Helper function for copy 335 * @adev: amdgpu device 336 * @src: buffer/address where to read from 337 * @dst: buffer/address where to write to 338 * @size: number of bytes to copy 339 * @tmz: if a secure copy should be used 340 * @resv: resv object to sync to 341 * @f: Returns the last fence if multiple jobs are submitted. 342 * 343 * The function copies @size bytes from {src->mem + src->offset} to 344 * {dst->mem + dst->offset}. src->bo and dst->bo could be same BO for a 345 * move and different for a BO to BO copy. 346 * 347 */ 348 int amdgpu_ttm_copy_mem_to_mem(struct amdgpu_device *adev, 349 const struct amdgpu_copy_mem *src, 350 const struct amdgpu_copy_mem *dst, 351 uint64_t size, bool tmz, 352 struct dma_resv *resv, 353 struct dma_fence **f) 354 { 355 const uint32_t GTT_MAX_BYTES = (AMDGPU_GTT_MAX_TRANSFER_SIZE * 356 AMDGPU_GPU_PAGE_SIZE); 357 358 uint64_t src_node_size, dst_node_size, src_offset, dst_offset; 359 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring; 360 struct drm_mm_node *src_mm, *dst_mm; 361 struct dma_fence *fence = NULL; 362 int r = 0; 363 364 if (!adev->mman.buffer_funcs_enabled) { 365 DRM_ERROR("Trying to move memory with ring turned off.\n"); 366 return -EINVAL; 367 } 368 369 src_offset = src->offset; 370 if (src->mem->mm_node) { 371 src_mm = amdgpu_find_mm_node(src->mem, &src_offset); 372 src_node_size = (src_mm->size << PAGE_SHIFT) - src_offset; 373 } else { 374 src_mm = NULL; 375 src_node_size = ULLONG_MAX; 376 } 377 378 dst_offset = dst->offset; 379 if (dst->mem->mm_node) { 380 dst_mm = amdgpu_find_mm_node(dst->mem, &dst_offset); 381 dst_node_size = (dst_mm->size << PAGE_SHIFT) - dst_offset; 382 } else { 383 dst_mm = NULL; 384 dst_node_size = ULLONG_MAX; 385 } 386 387 mutex_lock(&adev->mman.gtt_window_lock); 388 389 while (size) { 390 uint32_t src_page_offset = src_offset & ~PAGE_MASK; 391 uint32_t dst_page_offset = dst_offset & ~PAGE_MASK; 392 struct dma_fence *next; 393 uint32_t cur_size; 394 uint64_t from, to; 395 396 /* Copy size cannot exceed GTT_MAX_BYTES. So if src or dst 397 * begins at an offset, then adjust the size accordingly 398 */ 399 cur_size = max(src_page_offset, dst_page_offset); 400 cur_size = min(min3(src_node_size, dst_node_size, size), 401 (uint64_t)(GTT_MAX_BYTES - cur_size)); 402 403 /* Map src to window 0 and dst to window 1. */ 404 r = amdgpu_ttm_map_buffer(src->bo, src->mem, src_mm, 405 PFN_UP(cur_size + src_page_offset), 406 src_offset, 0, ring, tmz, &from); 407 if (r) 408 goto error; 409 410 r = amdgpu_ttm_map_buffer(dst->bo, dst->mem, dst_mm, 411 PFN_UP(cur_size + dst_page_offset), 412 dst_offset, 1, ring, tmz, &to); 413 if (r) 414 goto error; 415 416 r = amdgpu_copy_buffer(ring, from, to, cur_size, 417 resv, &next, false, true, tmz); 418 if (r) 419 goto error; 420 421 dma_fence_put(fence); 422 fence = next; 423 424 size -= cur_size; 425 if (!size) 426 break; 427 428 src_node_size -= cur_size; 429 if (!src_node_size) { 430 ++src_mm; 431 src_node_size = src_mm->size << PAGE_SHIFT; 432 src_offset = 0; 433 } else { 434 src_offset += cur_size; 435 } 436 437 dst_node_size -= cur_size; 438 if (!dst_node_size) { 439 ++dst_mm; 440 dst_node_size = dst_mm->size << PAGE_SHIFT; 441 dst_offset = 0; 442 } else { 443 dst_offset += cur_size; 444 } 445 } 446 error: 447 mutex_unlock(&adev->mman.gtt_window_lock); 448 if (f) 449 *f = dma_fence_get(fence); 450 dma_fence_put(fence); 451 return r; 452 } 453 454 /** 455 * amdgpu_move_blit - Copy an entire buffer to another buffer 456 * 457 * This is a helper called by amdgpu_bo_move() and amdgpu_move_vram_ram() to 458 * help move buffers to and from VRAM. 459 */ 460 static int amdgpu_move_blit(struct ttm_buffer_object *bo, 461 bool evict, 462 struct ttm_resource *new_mem, 463 struct ttm_resource *old_mem) 464 { 465 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev); 466 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo); 467 struct amdgpu_copy_mem src, dst; 468 struct dma_fence *fence = NULL; 469 int r; 470 471 src.bo = bo; 472 dst.bo = bo; 473 src.mem = old_mem; 474 dst.mem = new_mem; 475 src.offset = 0; 476 dst.offset = 0; 477 478 r = amdgpu_ttm_copy_mem_to_mem(adev, &src, &dst, 479 new_mem->num_pages << PAGE_SHIFT, 480 amdgpu_bo_encrypted(abo), 481 bo->base.resv, &fence); 482 if (r) 483 goto error; 484 485 /* clear the space being freed */ 486 if (old_mem->mem_type == TTM_PL_VRAM && 487 (abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE)) { 488 struct dma_fence *wipe_fence = NULL; 489 490 r = amdgpu_fill_buffer(ttm_to_amdgpu_bo(bo), AMDGPU_POISON, 491 NULL, &wipe_fence); 492 if (r) { 493 goto error; 494 } else if (wipe_fence) { 495 dma_fence_put(fence); 496 fence = wipe_fence; 497 } 498 } 499 500 /* Always block for VM page tables before committing the new location */ 501 if (bo->type == ttm_bo_type_kernel) 502 r = ttm_bo_move_accel_cleanup(bo, fence, true, false, new_mem); 503 else 504 r = ttm_bo_move_accel_cleanup(bo, fence, evict, true, new_mem); 505 dma_fence_put(fence); 506 return r; 507 508 error: 509 if (fence) 510 dma_fence_wait(fence, false); 511 dma_fence_put(fence); 512 return r; 513 } 514 515 /** 516 * amdgpu_mem_visible - Check that memory can be accessed by ttm_bo_move_memcpy 517 * 518 * Called by amdgpu_bo_move() 519 */ 520 static bool amdgpu_mem_visible(struct amdgpu_device *adev, 521 struct ttm_resource *mem) 522 { 523 struct drm_mm_node *nodes = mem->mm_node; 524 525 if (mem->mem_type == TTM_PL_SYSTEM || 526 mem->mem_type == TTM_PL_TT) 527 return true; 528 if (mem->mem_type != TTM_PL_VRAM) 529 return false; 530 531 /* ttm_resource_ioremap only supports contiguous memory */ 532 if (nodes->size != mem->num_pages) 533 return false; 534 535 return ((nodes->start + nodes->size) << PAGE_SHIFT) 536 <= adev->gmc.visible_vram_size; 537 } 538 539 /** 540 * amdgpu_bo_move - Move a buffer object to a new memory location 541 * 542 * Called by ttm_bo_handle_move_mem() 543 */ 544 static int amdgpu_bo_move(struct ttm_buffer_object *bo, bool evict, 545 struct ttm_operation_ctx *ctx, 546 struct ttm_resource *new_mem, 547 struct ttm_place *hop) 548 { 549 struct amdgpu_device *adev; 550 struct amdgpu_bo *abo; 551 struct ttm_resource *old_mem = &bo->mem; 552 int r; 553 554 if (new_mem->mem_type == TTM_PL_TT) { 555 r = amdgpu_ttm_backend_bind(bo->bdev, bo->ttm, new_mem); 556 if (r) 557 return r; 558 } 559 560 /* Can't move a pinned BO */ 561 abo = ttm_to_amdgpu_bo(bo); 562 if (WARN_ON_ONCE(abo->tbo.pin_count > 0)) 563 return -EINVAL; 564 565 adev = amdgpu_ttm_adev(bo->bdev); 566 567 if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) { 568 ttm_bo_move_null(bo, new_mem); 569 goto out; 570 } 571 if (old_mem->mem_type == TTM_PL_SYSTEM && 572 new_mem->mem_type == TTM_PL_TT) { 573 ttm_bo_move_null(bo, new_mem); 574 goto out; 575 } 576 if (old_mem->mem_type == TTM_PL_TT && 577 new_mem->mem_type == TTM_PL_SYSTEM) { 578 r = ttm_bo_wait_ctx(bo, ctx); 579 if (r) 580 return r; 581 582 amdgpu_ttm_backend_unbind(bo->bdev, bo->ttm); 583 ttm_resource_free(bo, &bo->mem); 584 ttm_bo_assign_mem(bo, new_mem); 585 goto out; 586 } 587 588 if (old_mem->mem_type == AMDGPU_PL_GDS || 589 old_mem->mem_type == AMDGPU_PL_GWS || 590 old_mem->mem_type == AMDGPU_PL_OA || 591 new_mem->mem_type == AMDGPU_PL_GDS || 592 new_mem->mem_type == AMDGPU_PL_GWS || 593 new_mem->mem_type == AMDGPU_PL_OA) { 594 /* Nothing to save here */ 595 ttm_bo_move_null(bo, new_mem); 596 goto out; 597 } 598 599 if (adev->mman.buffer_funcs_enabled) { 600 if (((old_mem->mem_type == TTM_PL_SYSTEM && 601 new_mem->mem_type == TTM_PL_VRAM) || 602 (old_mem->mem_type == TTM_PL_VRAM && 603 new_mem->mem_type == TTM_PL_SYSTEM))) { 604 hop->fpfn = 0; 605 hop->lpfn = 0; 606 hop->mem_type = TTM_PL_TT; 607 hop->flags = 0; 608 return -EMULTIHOP; 609 } 610 611 r = amdgpu_move_blit(bo, evict, new_mem, old_mem); 612 } else { 613 r = -ENODEV; 614 } 615 616 if (r) { 617 /* Check that all memory is CPU accessible */ 618 if (!amdgpu_mem_visible(adev, old_mem) || 619 !amdgpu_mem_visible(adev, new_mem)) { 620 pr_err("Move buffer fallback to memcpy unavailable\n"); 621 return r; 622 } 623 624 r = ttm_bo_move_memcpy(bo, ctx, new_mem); 625 if (r) 626 return r; 627 } 628 629 if (bo->type == ttm_bo_type_device && 630 new_mem->mem_type == TTM_PL_VRAM && 631 old_mem->mem_type != TTM_PL_VRAM) { 632 /* amdgpu_bo_fault_reserve_notify will re-set this if the CPU 633 * accesses the BO after it's moved. 634 */ 635 abo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED; 636 } 637 638 out: 639 /* update statistics */ 640 atomic64_add((u64)bo->num_pages << PAGE_SHIFT, &adev->num_bytes_moved); 641 amdgpu_bo_move_notify(bo, evict, new_mem); 642 return 0; 643 } 644 645 /** 646 * amdgpu_ttm_io_mem_reserve - Reserve a block of memory during a fault 647 * 648 * Called by ttm_mem_io_reserve() ultimately via ttm_bo_vm_fault() 649 */ 650 static int amdgpu_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_resource *mem) 651 { 652 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev); 653 struct drm_mm_node *mm_node = mem->mm_node; 654 size_t bus_size = (size_t)mem->num_pages << PAGE_SHIFT; 655 656 switch (mem->mem_type) { 657 case TTM_PL_SYSTEM: 658 /* system memory */ 659 return 0; 660 case TTM_PL_TT: 661 break; 662 case TTM_PL_VRAM: 663 mem->bus.offset = mem->start << PAGE_SHIFT; 664 /* check if it's visible */ 665 if ((mem->bus.offset + bus_size) > adev->gmc.visible_vram_size) 666 return -EINVAL; 667 /* Only physically contiguous buffers apply. In a contiguous 668 * buffer, size of the first mm_node would match the number of 669 * pages in ttm_resource. 670 */ 671 if (adev->mman.aper_base_kaddr && 672 (mm_node->size == mem->num_pages)) 673 mem->bus.addr = (u8 *)adev->mman.aper_base_kaddr + 674 mem->bus.offset; 675 676 mem->bus.offset += adev->gmc.aper_base; 677 mem->bus.is_iomem = true; 678 mem->bus.caching = ttm_write_combined; 679 break; 680 default: 681 return -EINVAL; 682 } 683 return 0; 684 } 685 686 static unsigned long amdgpu_ttm_io_mem_pfn(struct ttm_buffer_object *bo, 687 unsigned long page_offset) 688 { 689 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev); 690 uint64_t offset = (page_offset << PAGE_SHIFT); 691 struct drm_mm_node *mm; 692 693 mm = amdgpu_find_mm_node(&bo->mem, &offset); 694 offset += adev->gmc.aper_base; 695 return mm->start + (offset >> PAGE_SHIFT); 696 } 697 698 /** 699 * amdgpu_ttm_domain_start - Returns GPU start address 700 * @adev: amdgpu device object 701 * @type: type of the memory 702 * 703 * Returns: 704 * GPU start address of a memory domain 705 */ 706 707 uint64_t amdgpu_ttm_domain_start(struct amdgpu_device *adev, uint32_t type) 708 { 709 switch (type) { 710 case TTM_PL_TT: 711 return adev->gmc.gart_start; 712 case TTM_PL_VRAM: 713 return adev->gmc.vram_start; 714 } 715 716 return 0; 717 } 718 719 /* 720 * TTM backend functions. 721 */ 722 struct amdgpu_ttm_tt { 723 struct ttm_tt ttm; 724 struct drm_gem_object *gobj; 725 u64 offset; 726 uint64_t userptr; 727 struct task_struct *usertask; 728 uint32_t userflags; 729 bool bound; 730 #if IS_ENABLED(CONFIG_DRM_AMDGPU_USERPTR) 731 struct hmm_range *range; 732 #endif 733 }; 734 735 #ifdef CONFIG_DRM_AMDGPU_USERPTR 736 /** 737 * amdgpu_ttm_tt_get_user_pages - get device accessible pages that back user 738 * memory and start HMM tracking CPU page table update 739 * 740 * Calling function must call amdgpu_ttm_tt_userptr_range_done() once and only 741 * once afterwards to stop HMM tracking 742 */ 743 int amdgpu_ttm_tt_get_user_pages(struct amdgpu_bo *bo, struct page **pages) 744 { 745 struct ttm_tt *ttm = bo->tbo.ttm; 746 struct amdgpu_ttm_tt *gtt = (void *)ttm; 747 unsigned long start = gtt->userptr; 748 struct vm_area_struct *vma; 749 struct hmm_range *range; 750 unsigned long timeout; 751 struct mm_struct *mm; 752 unsigned long i; 753 int r = 0; 754 755 mm = bo->notifier.mm; 756 if (unlikely(!mm)) { 757 DRM_DEBUG_DRIVER("BO is not registered?\n"); 758 return -EFAULT; 759 } 760 761 /* Another get_user_pages is running at the same time?? */ 762 if (WARN_ON(gtt->range)) 763 return -EFAULT; 764 765 if (!mmget_not_zero(mm)) /* Happens during process shutdown */ 766 return -ESRCH; 767 768 range = kzalloc(sizeof(*range), GFP_KERNEL); 769 if (unlikely(!range)) { 770 r = -ENOMEM; 771 goto out; 772 } 773 range->notifier = &bo->notifier; 774 range->start = bo->notifier.interval_tree.start; 775 range->end = bo->notifier.interval_tree.last + 1; 776 range->default_flags = HMM_PFN_REQ_FAULT; 777 if (!amdgpu_ttm_tt_is_readonly(ttm)) 778 range->default_flags |= HMM_PFN_REQ_WRITE; 779 780 range->hmm_pfns = kvmalloc_array(ttm->num_pages, 781 sizeof(*range->hmm_pfns), GFP_KERNEL); 782 if (unlikely(!range->hmm_pfns)) { 783 r = -ENOMEM; 784 goto out_free_ranges; 785 } 786 787 mmap_read_lock(mm); 788 vma = find_vma(mm, start); 789 if (unlikely(!vma || start < vma->vm_start)) { 790 r = -EFAULT; 791 goto out_unlock; 792 } 793 if (unlikely((gtt->userflags & AMDGPU_GEM_USERPTR_ANONONLY) && 794 vma->vm_file)) { 795 r = -EPERM; 796 goto out_unlock; 797 } 798 mmap_read_unlock(mm); 799 timeout = jiffies + msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT); 800 801 retry: 802 range->notifier_seq = mmu_interval_read_begin(&bo->notifier); 803 804 mmap_read_lock(mm); 805 r = hmm_range_fault(range); 806 mmap_read_unlock(mm); 807 if (unlikely(r)) { 808 /* 809 * FIXME: This timeout should encompass the retry from 810 * mmu_interval_read_retry() as well. 811 */ 812 if (r == -EBUSY && !time_after(jiffies, timeout)) 813 goto retry; 814 goto out_free_pfns; 815 } 816 817 /* 818 * Due to default_flags, all pages are HMM_PFN_VALID or 819 * hmm_range_fault() fails. FIXME: The pages cannot be touched outside 820 * the notifier_lock, and mmu_interval_read_retry() must be done first. 821 */ 822 for (i = 0; i < ttm->num_pages; i++) 823 pages[i] = hmm_pfn_to_page(range->hmm_pfns[i]); 824 825 gtt->range = range; 826 mmput(mm); 827 828 return 0; 829 830 out_unlock: 831 mmap_read_unlock(mm); 832 out_free_pfns: 833 kvfree(range->hmm_pfns); 834 out_free_ranges: 835 kfree(range); 836 out: 837 mmput(mm); 838 return r; 839 } 840 841 /** 842 * amdgpu_ttm_tt_userptr_range_done - stop HMM track the CPU page table change 843 * Check if the pages backing this ttm range have been invalidated 844 * 845 * Returns: true if pages are still valid 846 */ 847 bool amdgpu_ttm_tt_get_user_pages_done(struct ttm_tt *ttm) 848 { 849 struct amdgpu_ttm_tt *gtt = (void *)ttm; 850 bool r = false; 851 852 if (!gtt || !gtt->userptr) 853 return false; 854 855 DRM_DEBUG_DRIVER("user_pages_done 0x%llx pages 0x%x\n", 856 gtt->userptr, ttm->num_pages); 857 858 WARN_ONCE(!gtt->range || !gtt->range->hmm_pfns, 859 "No user pages to check\n"); 860 861 if (gtt->range) { 862 /* 863 * FIXME: Must always hold notifier_lock for this, and must 864 * not ignore the return code. 865 */ 866 r = mmu_interval_read_retry(gtt->range->notifier, 867 gtt->range->notifier_seq); 868 kvfree(gtt->range->hmm_pfns); 869 kfree(gtt->range); 870 gtt->range = NULL; 871 } 872 873 return !r; 874 } 875 #endif 876 877 /** 878 * amdgpu_ttm_tt_set_user_pages - Copy pages in, putting old pages as necessary. 879 * 880 * Called by amdgpu_cs_list_validate(). This creates the page list 881 * that backs user memory and will ultimately be mapped into the device 882 * address space. 883 */ 884 void amdgpu_ttm_tt_set_user_pages(struct ttm_tt *ttm, struct page **pages) 885 { 886 unsigned long i; 887 888 for (i = 0; i < ttm->num_pages; ++i) 889 ttm->pages[i] = pages ? pages[i] : NULL; 890 } 891 892 /** 893 * amdgpu_ttm_tt_pin_userptr - prepare the sg table with the user pages 894 * 895 * Called by amdgpu_ttm_backend_bind() 896 **/ 897 static int amdgpu_ttm_tt_pin_userptr(struct ttm_bo_device *bdev, 898 struct ttm_tt *ttm) 899 { 900 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev); 901 struct amdgpu_ttm_tt *gtt = (void *)ttm; 902 int r; 903 904 int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY); 905 enum dma_data_direction direction = write ? 906 DMA_BIDIRECTIONAL : DMA_TO_DEVICE; 907 908 /* Allocate an SG array and squash pages into it */ 909 r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0, 910 ttm->num_pages << PAGE_SHIFT, 911 GFP_KERNEL); 912 if (r) 913 goto release_sg; 914 915 /* Map SG to device */ 916 r = dma_map_sgtable(adev->dev, ttm->sg, direction, 0); 917 if (r) 918 goto release_sg; 919 920 /* convert SG to linear array of pages and dma addresses */ 921 drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages, 922 gtt->ttm.dma_address, ttm->num_pages); 923 924 return 0; 925 926 release_sg: 927 kfree(ttm->sg); 928 ttm->sg = NULL; 929 return r; 930 } 931 932 /** 933 * amdgpu_ttm_tt_unpin_userptr - Unpin and unmap userptr pages 934 */ 935 static void amdgpu_ttm_tt_unpin_userptr(struct ttm_bo_device *bdev, 936 struct ttm_tt *ttm) 937 { 938 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev); 939 struct amdgpu_ttm_tt *gtt = (void *)ttm; 940 941 int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY); 942 enum dma_data_direction direction = write ? 943 DMA_BIDIRECTIONAL : DMA_TO_DEVICE; 944 945 /* double check that we don't free the table twice */ 946 if (!ttm->sg->sgl) 947 return; 948 949 /* unmap the pages mapped to the device */ 950 dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0); 951 sg_free_table(ttm->sg); 952 953 #if IS_ENABLED(CONFIG_DRM_AMDGPU_USERPTR) 954 if (gtt->range) { 955 unsigned long i; 956 957 for (i = 0; i < ttm->num_pages; i++) { 958 if (ttm->pages[i] != 959 hmm_pfn_to_page(gtt->range->hmm_pfns[i])) 960 break; 961 } 962 963 WARN((i == ttm->num_pages), "Missing get_user_page_done\n"); 964 } 965 #endif 966 } 967 968 static int amdgpu_ttm_gart_bind(struct amdgpu_device *adev, 969 struct ttm_buffer_object *tbo, 970 uint64_t flags) 971 { 972 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(tbo); 973 struct ttm_tt *ttm = tbo->ttm; 974 struct amdgpu_ttm_tt *gtt = (void *)ttm; 975 int r; 976 977 if (amdgpu_bo_encrypted(abo)) 978 flags |= AMDGPU_PTE_TMZ; 979 980 if (abo->flags & AMDGPU_GEM_CREATE_CP_MQD_GFX9) { 981 uint64_t page_idx = 1; 982 983 r = amdgpu_gart_bind(adev, gtt->offset, page_idx, 984 ttm->pages, gtt->ttm.dma_address, flags); 985 if (r) 986 goto gart_bind_fail; 987 988 /* The memory type of the first page defaults to UC. Now 989 * modify the memory type to NC from the second page of 990 * the BO onward. 991 */ 992 flags &= ~AMDGPU_PTE_MTYPE_VG10_MASK; 993 flags |= AMDGPU_PTE_MTYPE_VG10(AMDGPU_MTYPE_NC); 994 995 r = amdgpu_gart_bind(adev, 996 gtt->offset + (page_idx << PAGE_SHIFT), 997 ttm->num_pages - page_idx, 998 &ttm->pages[page_idx], 999 &(gtt->ttm.dma_address[page_idx]), flags); 1000 } else { 1001 r = amdgpu_gart_bind(adev, gtt->offset, ttm->num_pages, 1002 ttm->pages, gtt->ttm.dma_address, flags); 1003 } 1004 1005 gart_bind_fail: 1006 if (r) 1007 DRM_ERROR("failed to bind %u pages at 0x%08llX\n", 1008 ttm->num_pages, gtt->offset); 1009 1010 return r; 1011 } 1012 1013 /** 1014 * amdgpu_ttm_backend_bind - Bind GTT memory 1015 * 1016 * Called by ttm_tt_bind() on behalf of ttm_bo_handle_move_mem(). 1017 * This handles binding GTT memory to the device address space. 1018 */ 1019 static int amdgpu_ttm_backend_bind(struct ttm_bo_device *bdev, 1020 struct ttm_tt *ttm, 1021 struct ttm_resource *bo_mem) 1022 { 1023 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev); 1024 struct amdgpu_ttm_tt *gtt = (void*)ttm; 1025 uint64_t flags; 1026 int r = 0; 1027 1028 if (!bo_mem) 1029 return -EINVAL; 1030 1031 if (gtt->bound) 1032 return 0; 1033 1034 if (gtt->userptr) { 1035 r = amdgpu_ttm_tt_pin_userptr(bdev, ttm); 1036 if (r) { 1037 DRM_ERROR("failed to pin userptr\n"); 1038 return r; 1039 } 1040 } 1041 if (!ttm->num_pages) { 1042 WARN(1, "nothing to bind %u pages for mreg %p back %p!\n", 1043 ttm->num_pages, bo_mem, ttm); 1044 } 1045 1046 if (bo_mem->mem_type == AMDGPU_PL_GDS || 1047 bo_mem->mem_type == AMDGPU_PL_GWS || 1048 bo_mem->mem_type == AMDGPU_PL_OA) 1049 return -EINVAL; 1050 1051 if (!amdgpu_gtt_mgr_has_gart_addr(bo_mem)) { 1052 gtt->offset = AMDGPU_BO_INVALID_OFFSET; 1053 return 0; 1054 } 1055 1056 /* compute PTE flags relevant to this BO memory */ 1057 flags = amdgpu_ttm_tt_pte_flags(adev, ttm, bo_mem); 1058 1059 /* bind pages into GART page tables */ 1060 gtt->offset = (u64)bo_mem->start << PAGE_SHIFT; 1061 r = amdgpu_gart_bind(adev, gtt->offset, ttm->num_pages, 1062 ttm->pages, gtt->ttm.dma_address, flags); 1063 1064 if (r) 1065 DRM_ERROR("failed to bind %u pages at 0x%08llX\n", 1066 ttm->num_pages, gtt->offset); 1067 gtt->bound = true; 1068 return r; 1069 } 1070 1071 /** 1072 * amdgpu_ttm_alloc_gart - Make sure buffer object is accessible either 1073 * through AGP or GART aperture. 1074 * 1075 * If bo is accessible through AGP aperture, then use AGP aperture 1076 * to access bo; otherwise allocate logical space in GART aperture 1077 * and map bo to GART aperture. 1078 */ 1079 int amdgpu_ttm_alloc_gart(struct ttm_buffer_object *bo) 1080 { 1081 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev); 1082 struct ttm_operation_ctx ctx = { false, false }; 1083 struct amdgpu_ttm_tt *gtt = (void *)bo->ttm; 1084 struct ttm_resource tmp; 1085 struct ttm_placement placement; 1086 struct ttm_place placements; 1087 uint64_t addr, flags; 1088 int r; 1089 1090 if (bo->mem.start != AMDGPU_BO_INVALID_OFFSET) 1091 return 0; 1092 1093 addr = amdgpu_gmc_agp_addr(bo); 1094 if (addr != AMDGPU_BO_INVALID_OFFSET) { 1095 bo->mem.start = addr >> PAGE_SHIFT; 1096 } else { 1097 1098 /* allocate GART space */ 1099 tmp = bo->mem; 1100 tmp.mm_node = NULL; 1101 placement.num_placement = 1; 1102 placement.placement = &placements; 1103 placement.num_busy_placement = 1; 1104 placement.busy_placement = &placements; 1105 placements.fpfn = 0; 1106 placements.lpfn = adev->gmc.gart_size >> PAGE_SHIFT; 1107 placements.mem_type = TTM_PL_TT; 1108 placements.flags = bo->mem.placement; 1109 1110 r = ttm_bo_mem_space(bo, &placement, &tmp, &ctx); 1111 if (unlikely(r)) 1112 return r; 1113 1114 /* compute PTE flags for this buffer object */ 1115 flags = amdgpu_ttm_tt_pte_flags(adev, bo->ttm, &tmp); 1116 1117 /* Bind pages */ 1118 gtt->offset = (u64)tmp.start << PAGE_SHIFT; 1119 r = amdgpu_ttm_gart_bind(adev, bo, flags); 1120 if (unlikely(r)) { 1121 ttm_resource_free(bo, &tmp); 1122 return r; 1123 } 1124 1125 ttm_resource_free(bo, &bo->mem); 1126 bo->mem = tmp; 1127 } 1128 1129 return 0; 1130 } 1131 1132 /** 1133 * amdgpu_ttm_recover_gart - Rebind GTT pages 1134 * 1135 * Called by amdgpu_gtt_mgr_recover() from amdgpu_device_reset() to 1136 * rebind GTT pages during a GPU reset. 1137 */ 1138 int amdgpu_ttm_recover_gart(struct ttm_buffer_object *tbo) 1139 { 1140 struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev); 1141 uint64_t flags; 1142 int r; 1143 1144 if (!tbo->ttm) 1145 return 0; 1146 1147 flags = amdgpu_ttm_tt_pte_flags(adev, tbo->ttm, &tbo->mem); 1148 r = amdgpu_ttm_gart_bind(adev, tbo, flags); 1149 1150 return r; 1151 } 1152 1153 /** 1154 * amdgpu_ttm_backend_unbind - Unbind GTT mapped pages 1155 * 1156 * Called by ttm_tt_unbind() on behalf of ttm_bo_move_ttm() and 1157 * ttm_tt_destroy(). 1158 */ 1159 static void amdgpu_ttm_backend_unbind(struct ttm_bo_device *bdev, 1160 struct ttm_tt *ttm) 1161 { 1162 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev); 1163 struct amdgpu_ttm_tt *gtt = (void *)ttm; 1164 int r; 1165 1166 if (!gtt->bound) 1167 return; 1168 1169 /* if the pages have userptr pinning then clear that first */ 1170 if (gtt->userptr) 1171 amdgpu_ttm_tt_unpin_userptr(bdev, ttm); 1172 1173 if (gtt->offset == AMDGPU_BO_INVALID_OFFSET) 1174 return; 1175 1176 /* unbind shouldn't be done for GDS/GWS/OA in ttm_bo_clean_mm */ 1177 r = amdgpu_gart_unbind(adev, gtt->offset, ttm->num_pages); 1178 if (r) 1179 DRM_ERROR("failed to unbind %u pages at 0x%08llX\n", 1180 gtt->ttm.num_pages, gtt->offset); 1181 gtt->bound = false; 1182 } 1183 1184 static void amdgpu_ttm_backend_destroy(struct ttm_bo_device *bdev, 1185 struct ttm_tt *ttm) 1186 { 1187 struct amdgpu_ttm_tt *gtt = (void *)ttm; 1188 1189 amdgpu_ttm_backend_unbind(bdev, ttm); 1190 ttm_tt_destroy_common(bdev, ttm); 1191 if (gtt->usertask) 1192 put_task_struct(gtt->usertask); 1193 1194 ttm_tt_fini(>t->ttm); 1195 kfree(gtt); 1196 } 1197 1198 /** 1199 * amdgpu_ttm_tt_create - Create a ttm_tt object for a given BO 1200 * 1201 * @bo: The buffer object to create a GTT ttm_tt object around 1202 * 1203 * Called by ttm_tt_create(). 1204 */ 1205 static struct ttm_tt *amdgpu_ttm_tt_create(struct ttm_buffer_object *bo, 1206 uint32_t page_flags) 1207 { 1208 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo); 1209 struct amdgpu_ttm_tt *gtt; 1210 enum ttm_caching caching; 1211 1212 gtt = kzalloc(sizeof(struct amdgpu_ttm_tt), GFP_KERNEL); 1213 if (gtt == NULL) { 1214 return NULL; 1215 } 1216 gtt->gobj = &bo->base; 1217 1218 if (abo->flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC) 1219 caching = ttm_write_combined; 1220 else 1221 caching = ttm_cached; 1222 1223 /* allocate space for the uninitialized page entries */ 1224 if (ttm_sg_tt_init(>t->ttm, bo, page_flags, caching)) { 1225 kfree(gtt); 1226 return NULL; 1227 } 1228 return >t->ttm; 1229 } 1230 1231 /** 1232 * amdgpu_ttm_tt_populate - Map GTT pages visible to the device 1233 * 1234 * Map the pages of a ttm_tt object to an address space visible 1235 * to the underlying device. 1236 */ 1237 static int amdgpu_ttm_tt_populate(struct ttm_bo_device *bdev, 1238 struct ttm_tt *ttm, 1239 struct ttm_operation_ctx *ctx) 1240 { 1241 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev); 1242 struct amdgpu_ttm_tt *gtt = (void *)ttm; 1243 1244 /* user pages are bound by amdgpu_ttm_tt_pin_userptr() */ 1245 if (gtt && gtt->userptr) { 1246 ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL); 1247 if (!ttm->sg) 1248 return -ENOMEM; 1249 1250 ttm->page_flags |= TTM_PAGE_FLAG_SG; 1251 return 0; 1252 } 1253 1254 if (ttm->page_flags & TTM_PAGE_FLAG_SG) { 1255 if (!ttm->sg) { 1256 struct dma_buf_attachment *attach; 1257 struct sg_table *sgt; 1258 1259 attach = gtt->gobj->import_attach; 1260 sgt = dma_buf_map_attachment(attach, DMA_BIDIRECTIONAL); 1261 if (IS_ERR(sgt)) 1262 return PTR_ERR(sgt); 1263 1264 ttm->sg = sgt; 1265 } 1266 1267 drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages, 1268 gtt->ttm.dma_address, 1269 ttm->num_pages); 1270 return 0; 1271 } 1272 1273 return ttm_pool_alloc(&adev->mman.bdev.pool, ttm, ctx); 1274 } 1275 1276 /** 1277 * amdgpu_ttm_tt_unpopulate - unmap GTT pages and unpopulate page arrays 1278 * 1279 * Unmaps pages of a ttm_tt object from the device address space and 1280 * unpopulates the page array backing it. 1281 */ 1282 static void amdgpu_ttm_tt_unpopulate(struct ttm_bo_device *bdev, 1283 struct ttm_tt *ttm) 1284 { 1285 struct amdgpu_ttm_tt *gtt = (void *)ttm; 1286 struct amdgpu_device *adev; 1287 1288 if (gtt && gtt->userptr) { 1289 amdgpu_ttm_tt_set_user_pages(ttm, NULL); 1290 kfree(ttm->sg); 1291 ttm->page_flags &= ~TTM_PAGE_FLAG_SG; 1292 return; 1293 } 1294 1295 if (ttm->sg && gtt->gobj->import_attach) { 1296 struct dma_buf_attachment *attach; 1297 1298 attach = gtt->gobj->import_attach; 1299 dma_buf_unmap_attachment(attach, ttm->sg, DMA_BIDIRECTIONAL); 1300 ttm->sg = NULL; 1301 return; 1302 } 1303 1304 if (ttm->page_flags & TTM_PAGE_FLAG_SG) 1305 return; 1306 1307 adev = amdgpu_ttm_adev(bdev); 1308 return ttm_pool_free(&adev->mman.bdev.pool, ttm); 1309 } 1310 1311 /** 1312 * amdgpu_ttm_tt_set_userptr - Initialize userptr GTT ttm_tt for the current 1313 * task 1314 * 1315 * @bo: The ttm_buffer_object to bind this userptr to 1316 * @addr: The address in the current tasks VM space to use 1317 * @flags: Requirements of userptr object. 1318 * 1319 * Called by amdgpu_gem_userptr_ioctl() to bind userptr pages 1320 * to current task 1321 */ 1322 int amdgpu_ttm_tt_set_userptr(struct ttm_buffer_object *bo, 1323 uint64_t addr, uint32_t flags) 1324 { 1325 struct amdgpu_ttm_tt *gtt; 1326 1327 if (!bo->ttm) { 1328 /* TODO: We want a separate TTM object type for userptrs */ 1329 bo->ttm = amdgpu_ttm_tt_create(bo, 0); 1330 if (bo->ttm == NULL) 1331 return -ENOMEM; 1332 } 1333 1334 gtt = (void *)bo->ttm; 1335 gtt->userptr = addr; 1336 gtt->userflags = flags; 1337 1338 if (gtt->usertask) 1339 put_task_struct(gtt->usertask); 1340 gtt->usertask = current->group_leader; 1341 get_task_struct(gtt->usertask); 1342 1343 return 0; 1344 } 1345 1346 /** 1347 * amdgpu_ttm_tt_get_usermm - Return memory manager for ttm_tt object 1348 */ 1349 struct mm_struct *amdgpu_ttm_tt_get_usermm(struct ttm_tt *ttm) 1350 { 1351 struct amdgpu_ttm_tt *gtt = (void *)ttm; 1352 1353 if (gtt == NULL) 1354 return NULL; 1355 1356 if (gtt->usertask == NULL) 1357 return NULL; 1358 1359 return gtt->usertask->mm; 1360 } 1361 1362 /** 1363 * amdgpu_ttm_tt_affect_userptr - Determine if a ttm_tt object lays inside an 1364 * address range for the current task. 1365 * 1366 */ 1367 bool amdgpu_ttm_tt_affect_userptr(struct ttm_tt *ttm, unsigned long start, 1368 unsigned long end) 1369 { 1370 struct amdgpu_ttm_tt *gtt = (void *)ttm; 1371 unsigned long size; 1372 1373 if (gtt == NULL || !gtt->userptr) 1374 return false; 1375 1376 /* Return false if no part of the ttm_tt object lies within 1377 * the range 1378 */ 1379 size = (unsigned long)gtt->ttm.num_pages * PAGE_SIZE; 1380 if (gtt->userptr > end || gtt->userptr + size <= start) 1381 return false; 1382 1383 return true; 1384 } 1385 1386 /** 1387 * amdgpu_ttm_tt_is_userptr - Have the pages backing by userptr? 1388 */ 1389 bool amdgpu_ttm_tt_is_userptr(struct ttm_tt *ttm) 1390 { 1391 struct amdgpu_ttm_tt *gtt = (void *)ttm; 1392 1393 if (gtt == NULL || !gtt->userptr) 1394 return false; 1395 1396 return true; 1397 } 1398 1399 /** 1400 * amdgpu_ttm_tt_is_readonly - Is the ttm_tt object read only? 1401 */ 1402 bool amdgpu_ttm_tt_is_readonly(struct ttm_tt *ttm) 1403 { 1404 struct amdgpu_ttm_tt *gtt = (void *)ttm; 1405 1406 if (gtt == NULL) 1407 return false; 1408 1409 return !!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY); 1410 } 1411 1412 /** 1413 * amdgpu_ttm_tt_pde_flags - Compute PDE flags for ttm_tt object 1414 * 1415 * @ttm: The ttm_tt object to compute the flags for 1416 * @mem: The memory registry backing this ttm_tt object 1417 * 1418 * Figure out the flags to use for a VM PDE (Page Directory Entry). 1419 */ 1420 uint64_t amdgpu_ttm_tt_pde_flags(struct ttm_tt *ttm, struct ttm_resource *mem) 1421 { 1422 uint64_t flags = 0; 1423 1424 if (mem && mem->mem_type != TTM_PL_SYSTEM) 1425 flags |= AMDGPU_PTE_VALID; 1426 1427 if (mem && mem->mem_type == TTM_PL_TT) { 1428 flags |= AMDGPU_PTE_SYSTEM; 1429 1430 if (ttm->caching == ttm_cached) 1431 flags |= AMDGPU_PTE_SNOOPED; 1432 } 1433 1434 return flags; 1435 } 1436 1437 /** 1438 * amdgpu_ttm_tt_pte_flags - Compute PTE flags for ttm_tt object 1439 * 1440 * @ttm: The ttm_tt object to compute the flags for 1441 * @mem: The memory registry backing this ttm_tt object 1442 1443 * Figure out the flags to use for a VM PTE (Page Table Entry). 1444 */ 1445 uint64_t amdgpu_ttm_tt_pte_flags(struct amdgpu_device *adev, struct ttm_tt *ttm, 1446 struct ttm_resource *mem) 1447 { 1448 uint64_t flags = amdgpu_ttm_tt_pde_flags(ttm, mem); 1449 1450 flags |= adev->gart.gart_pte_flags; 1451 flags |= AMDGPU_PTE_READABLE; 1452 1453 if (!amdgpu_ttm_tt_is_readonly(ttm)) 1454 flags |= AMDGPU_PTE_WRITEABLE; 1455 1456 return flags; 1457 } 1458 1459 /** 1460 * amdgpu_ttm_bo_eviction_valuable - Check to see if we can evict a buffer 1461 * object. 1462 * 1463 * Return true if eviction is sensible. Called by ttm_mem_evict_first() on 1464 * behalf of ttm_bo_mem_force_space() which tries to evict buffer objects until 1465 * it can find space for a new object and by ttm_bo_force_list_clean() which is 1466 * used to clean out a memory space. 1467 */ 1468 static bool amdgpu_ttm_bo_eviction_valuable(struct ttm_buffer_object *bo, 1469 const struct ttm_place *place) 1470 { 1471 unsigned long num_pages = bo->mem.num_pages; 1472 struct drm_mm_node *node = bo->mem.mm_node; 1473 struct dma_resv_list *flist; 1474 struct dma_fence *f; 1475 int i; 1476 1477 if (bo->type == ttm_bo_type_kernel && 1478 !amdgpu_vm_evictable(ttm_to_amdgpu_bo(bo))) 1479 return false; 1480 1481 /* If bo is a KFD BO, check if the bo belongs to the current process. 1482 * If true, then return false as any KFD process needs all its BOs to 1483 * be resident to run successfully 1484 */ 1485 flist = dma_resv_get_list(bo->base.resv); 1486 if (flist) { 1487 for (i = 0; i < flist->shared_count; ++i) { 1488 f = rcu_dereference_protected(flist->shared[i], 1489 dma_resv_held(bo->base.resv)); 1490 if (amdkfd_fence_check_mm(f, current->mm)) 1491 return false; 1492 } 1493 } 1494 1495 switch (bo->mem.mem_type) { 1496 case TTM_PL_TT: 1497 if (amdgpu_bo_is_amdgpu_bo(bo) && 1498 amdgpu_bo_encrypted(ttm_to_amdgpu_bo(bo))) 1499 return false; 1500 return true; 1501 1502 case TTM_PL_VRAM: 1503 /* Check each drm MM node individually */ 1504 while (num_pages) { 1505 if (place->fpfn < (node->start + node->size) && 1506 !(place->lpfn && place->lpfn <= node->start)) 1507 return true; 1508 1509 num_pages -= node->size; 1510 ++node; 1511 } 1512 return false; 1513 1514 default: 1515 break; 1516 } 1517 1518 return ttm_bo_eviction_valuable(bo, place); 1519 } 1520 1521 /** 1522 * amdgpu_ttm_access_memory - Read or Write memory that backs a buffer object. 1523 * 1524 * @bo: The buffer object to read/write 1525 * @offset: Offset into buffer object 1526 * @buf: Secondary buffer to write/read from 1527 * @len: Length in bytes of access 1528 * @write: true if writing 1529 * 1530 * This is used to access VRAM that backs a buffer object via MMIO 1531 * access for debugging purposes. 1532 */ 1533 static int amdgpu_ttm_access_memory(struct ttm_buffer_object *bo, 1534 unsigned long offset, 1535 void *buf, int len, int write) 1536 { 1537 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo); 1538 struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev); 1539 struct drm_mm_node *nodes; 1540 uint32_t value = 0; 1541 int ret = 0; 1542 uint64_t pos; 1543 unsigned long flags; 1544 1545 if (bo->mem.mem_type != TTM_PL_VRAM) 1546 return -EIO; 1547 1548 pos = offset; 1549 nodes = amdgpu_find_mm_node(&abo->tbo.mem, &pos); 1550 pos += (nodes->start << PAGE_SHIFT); 1551 1552 while (len && pos < adev->gmc.mc_vram_size) { 1553 uint64_t aligned_pos = pos & ~(uint64_t)3; 1554 uint64_t bytes = 4 - (pos & 3); 1555 uint32_t shift = (pos & 3) * 8; 1556 uint32_t mask = 0xffffffff << shift; 1557 1558 if (len < bytes) { 1559 mask &= 0xffffffff >> (bytes - len) * 8; 1560 bytes = len; 1561 } 1562 1563 if (mask != 0xffffffff) { 1564 spin_lock_irqsave(&adev->mmio_idx_lock, flags); 1565 WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)aligned_pos) | 0x80000000); 1566 WREG32_NO_KIQ(mmMM_INDEX_HI, aligned_pos >> 31); 1567 if (!write || mask != 0xffffffff) 1568 value = RREG32_NO_KIQ(mmMM_DATA); 1569 if (write) { 1570 value &= ~mask; 1571 value |= (*(uint32_t *)buf << shift) & mask; 1572 WREG32_NO_KIQ(mmMM_DATA, value); 1573 } 1574 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags); 1575 if (!write) { 1576 value = (value & mask) >> shift; 1577 memcpy(buf, &value, bytes); 1578 } 1579 } else { 1580 bytes = (nodes->start + nodes->size) << PAGE_SHIFT; 1581 bytes = min(bytes - pos, (uint64_t)len & ~0x3ull); 1582 1583 amdgpu_device_vram_access(adev, pos, (uint32_t *)buf, 1584 bytes, write); 1585 } 1586 1587 ret += bytes; 1588 buf = (uint8_t *)buf + bytes; 1589 pos += bytes; 1590 len -= bytes; 1591 if (pos >= (nodes->start + nodes->size) << PAGE_SHIFT) { 1592 ++nodes; 1593 pos = (nodes->start << PAGE_SHIFT); 1594 } 1595 } 1596 1597 return ret; 1598 } 1599 1600 static void 1601 amdgpu_bo_delete_mem_notify(struct ttm_buffer_object *bo) 1602 { 1603 amdgpu_bo_move_notify(bo, false, NULL); 1604 } 1605 1606 static struct ttm_bo_driver amdgpu_bo_driver = { 1607 .ttm_tt_create = &amdgpu_ttm_tt_create, 1608 .ttm_tt_populate = &amdgpu_ttm_tt_populate, 1609 .ttm_tt_unpopulate = &amdgpu_ttm_tt_unpopulate, 1610 .ttm_tt_destroy = &amdgpu_ttm_backend_destroy, 1611 .eviction_valuable = amdgpu_ttm_bo_eviction_valuable, 1612 .evict_flags = &amdgpu_evict_flags, 1613 .move = &amdgpu_bo_move, 1614 .verify_access = &amdgpu_verify_access, 1615 .delete_mem_notify = &amdgpu_bo_delete_mem_notify, 1616 .release_notify = &amdgpu_bo_release_notify, 1617 .io_mem_reserve = &amdgpu_ttm_io_mem_reserve, 1618 .io_mem_pfn = amdgpu_ttm_io_mem_pfn, 1619 .access_memory = &amdgpu_ttm_access_memory, 1620 .del_from_lru_notify = &amdgpu_vm_del_from_lru_notify 1621 }; 1622 1623 /* 1624 * Firmware Reservation functions 1625 */ 1626 /** 1627 * amdgpu_ttm_fw_reserve_vram_fini - free fw reserved vram 1628 * 1629 * @adev: amdgpu_device pointer 1630 * 1631 * free fw reserved vram if it has been reserved. 1632 */ 1633 static void amdgpu_ttm_fw_reserve_vram_fini(struct amdgpu_device *adev) 1634 { 1635 amdgpu_bo_free_kernel(&adev->mman.fw_vram_usage_reserved_bo, 1636 NULL, &adev->mman.fw_vram_usage_va); 1637 } 1638 1639 /** 1640 * amdgpu_ttm_fw_reserve_vram_init - create bo vram reservation from fw 1641 * 1642 * @adev: amdgpu_device pointer 1643 * 1644 * create bo vram reservation from fw. 1645 */ 1646 static int amdgpu_ttm_fw_reserve_vram_init(struct amdgpu_device *adev) 1647 { 1648 uint64_t vram_size = adev->gmc.visible_vram_size; 1649 1650 adev->mman.fw_vram_usage_va = NULL; 1651 adev->mman.fw_vram_usage_reserved_bo = NULL; 1652 1653 if (adev->mman.fw_vram_usage_size == 0 || 1654 adev->mman.fw_vram_usage_size > vram_size) 1655 return 0; 1656 1657 return amdgpu_bo_create_kernel_at(adev, 1658 adev->mman.fw_vram_usage_start_offset, 1659 adev->mman.fw_vram_usage_size, 1660 AMDGPU_GEM_DOMAIN_VRAM, 1661 &adev->mman.fw_vram_usage_reserved_bo, 1662 &adev->mman.fw_vram_usage_va); 1663 } 1664 1665 /* 1666 * Memoy training reservation functions 1667 */ 1668 1669 /** 1670 * amdgpu_ttm_training_reserve_vram_fini - free memory training reserved vram 1671 * 1672 * @adev: amdgpu_device pointer 1673 * 1674 * free memory training reserved vram if it has been reserved. 1675 */ 1676 static int amdgpu_ttm_training_reserve_vram_fini(struct amdgpu_device *adev) 1677 { 1678 struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx; 1679 1680 ctx->init = PSP_MEM_TRAIN_NOT_SUPPORT; 1681 amdgpu_bo_free_kernel(&ctx->c2p_bo, NULL, NULL); 1682 ctx->c2p_bo = NULL; 1683 1684 return 0; 1685 } 1686 1687 static void amdgpu_ttm_training_data_block_init(struct amdgpu_device *adev) 1688 { 1689 struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx; 1690 1691 memset(ctx, 0, sizeof(*ctx)); 1692 1693 ctx->c2p_train_data_offset = 1694 ALIGN((adev->gmc.mc_vram_size - adev->mman.discovery_tmr_size - SZ_1M), SZ_1M); 1695 ctx->p2c_train_data_offset = 1696 (adev->gmc.mc_vram_size - GDDR6_MEM_TRAINING_OFFSET); 1697 ctx->train_data_size = 1698 GDDR6_MEM_TRAINING_DATA_SIZE_IN_BYTES; 1699 1700 DRM_DEBUG("train_data_size:%llx,p2c_train_data_offset:%llx,c2p_train_data_offset:%llx.\n", 1701 ctx->train_data_size, 1702 ctx->p2c_train_data_offset, 1703 ctx->c2p_train_data_offset); 1704 } 1705 1706 /* 1707 * reserve TMR memory at the top of VRAM which holds 1708 * IP Discovery data and is protected by PSP. 1709 */ 1710 static int amdgpu_ttm_reserve_tmr(struct amdgpu_device *adev) 1711 { 1712 int ret; 1713 struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx; 1714 bool mem_train_support = false; 1715 1716 if (!amdgpu_sriov_vf(adev)) { 1717 ret = amdgpu_mem_train_support(adev); 1718 if (ret == 1) 1719 mem_train_support = true; 1720 else if (ret == -1) 1721 return -EINVAL; 1722 else 1723 DRM_DEBUG("memory training does not support!\n"); 1724 } 1725 1726 /* 1727 * Query reserved tmr size through atom firmwareinfo for Sienna_Cichlid and onwards for all 1728 * the use cases (IP discovery/G6 memory training/profiling/diagnostic data.etc) 1729 * 1730 * Otherwise, fallback to legacy approach to check and reserve tmr block for ip 1731 * discovery data and G6 memory training data respectively 1732 */ 1733 adev->mman.discovery_tmr_size = 1734 amdgpu_atomfirmware_get_fw_reserved_fb_size(adev); 1735 if (!adev->mman.discovery_tmr_size) 1736 adev->mman.discovery_tmr_size = DISCOVERY_TMR_OFFSET; 1737 1738 if (mem_train_support) { 1739 /* reserve vram for mem train according to TMR location */ 1740 amdgpu_ttm_training_data_block_init(adev); 1741 ret = amdgpu_bo_create_kernel_at(adev, 1742 ctx->c2p_train_data_offset, 1743 ctx->train_data_size, 1744 AMDGPU_GEM_DOMAIN_VRAM, 1745 &ctx->c2p_bo, 1746 NULL); 1747 if (ret) { 1748 DRM_ERROR("alloc c2p_bo failed(%d)!\n", ret); 1749 amdgpu_ttm_training_reserve_vram_fini(adev); 1750 return ret; 1751 } 1752 ctx->init = PSP_MEM_TRAIN_RESERVE_SUCCESS; 1753 } 1754 1755 ret = amdgpu_bo_create_kernel_at(adev, 1756 adev->gmc.real_vram_size - adev->mman.discovery_tmr_size, 1757 adev->mman.discovery_tmr_size, 1758 AMDGPU_GEM_DOMAIN_VRAM, 1759 &adev->mman.discovery_memory, 1760 NULL); 1761 if (ret) { 1762 DRM_ERROR("alloc tmr failed(%d)!\n", ret); 1763 amdgpu_bo_free_kernel(&adev->mman.discovery_memory, NULL, NULL); 1764 return ret; 1765 } 1766 1767 return 0; 1768 } 1769 1770 /** 1771 * amdgpu_ttm_init - Init the memory management (ttm) as well as various 1772 * gtt/vram related fields. 1773 * 1774 * This initializes all of the memory space pools that the TTM layer 1775 * will need such as the GTT space (system memory mapped to the device), 1776 * VRAM (on-board memory), and on-chip memories (GDS, GWS, OA) which 1777 * can be mapped per VMID. 1778 */ 1779 int amdgpu_ttm_init(struct amdgpu_device *adev) 1780 { 1781 uint64_t gtt_size; 1782 int r; 1783 u64 vis_vram_limit; 1784 1785 mutex_init(&adev->mman.gtt_window_lock); 1786 1787 /* No others user of address space so set it to 0 */ 1788 r = ttm_bo_device_init(&adev->mman.bdev, &amdgpu_bo_driver, adev->dev, 1789 adev_to_drm(adev)->anon_inode->i_mapping, 1790 adev_to_drm(adev)->vma_offset_manager, 1791 adev->need_swiotlb, 1792 dma_addressing_limited(adev->dev)); 1793 if (r) { 1794 DRM_ERROR("failed initializing buffer object driver(%d).\n", r); 1795 return r; 1796 } 1797 adev->mman.initialized = true; 1798 1799 /* Initialize VRAM pool with all of VRAM divided into pages */ 1800 r = amdgpu_vram_mgr_init(adev); 1801 if (r) { 1802 DRM_ERROR("Failed initializing VRAM heap.\n"); 1803 return r; 1804 } 1805 1806 /* Reduce size of CPU-visible VRAM if requested */ 1807 vis_vram_limit = (u64)amdgpu_vis_vram_limit * 1024 * 1024; 1808 if (amdgpu_vis_vram_limit > 0 && 1809 vis_vram_limit <= adev->gmc.visible_vram_size) 1810 adev->gmc.visible_vram_size = vis_vram_limit; 1811 1812 /* Change the size here instead of the init above so only lpfn is affected */ 1813 amdgpu_ttm_set_buffer_funcs_status(adev, false); 1814 #ifdef CONFIG_64BIT 1815 adev->mman.aper_base_kaddr = ioremap_wc(adev->gmc.aper_base, 1816 adev->gmc.visible_vram_size); 1817 #endif 1818 1819 /* 1820 *The reserved vram for firmware must be pinned to the specified 1821 *place on the VRAM, so reserve it early. 1822 */ 1823 r = amdgpu_ttm_fw_reserve_vram_init(adev); 1824 if (r) { 1825 return r; 1826 } 1827 1828 /* 1829 * only NAVI10 and onwards ASIC support for IP discovery. 1830 * If IP discovery enabled, a block of memory should be 1831 * reserved for IP discovey. 1832 */ 1833 if (adev->mman.discovery_bin) { 1834 r = amdgpu_ttm_reserve_tmr(adev); 1835 if (r) 1836 return r; 1837 } 1838 1839 /* allocate memory as required for VGA 1840 * This is used for VGA emulation and pre-OS scanout buffers to 1841 * avoid display artifacts while transitioning between pre-OS 1842 * and driver. */ 1843 r = amdgpu_bo_create_kernel_at(adev, 0, adev->mman.stolen_vga_size, 1844 AMDGPU_GEM_DOMAIN_VRAM, 1845 &adev->mman.stolen_vga_memory, 1846 NULL); 1847 if (r) 1848 return r; 1849 r = amdgpu_bo_create_kernel_at(adev, adev->mman.stolen_vga_size, 1850 adev->mman.stolen_extended_size, 1851 AMDGPU_GEM_DOMAIN_VRAM, 1852 &adev->mman.stolen_extended_memory, 1853 NULL); 1854 if (r) 1855 return r; 1856 1857 DRM_INFO("amdgpu: %uM of VRAM memory ready\n", 1858 (unsigned) (adev->gmc.real_vram_size / (1024 * 1024))); 1859 1860 /* Compute GTT size, either bsaed on 3/4th the size of RAM size 1861 * or whatever the user passed on module init */ 1862 if (amdgpu_gtt_size == -1) { 1863 struct sysinfo si; 1864 1865 si_meminfo(&si); 1866 gtt_size = min(max((AMDGPU_DEFAULT_GTT_SIZE_MB << 20), 1867 adev->gmc.mc_vram_size), 1868 ((uint64_t)si.totalram * si.mem_unit * 3/4)); 1869 } 1870 else 1871 gtt_size = (uint64_t)amdgpu_gtt_size << 20; 1872 1873 /* Initialize GTT memory pool */ 1874 r = amdgpu_gtt_mgr_init(adev, gtt_size); 1875 if (r) { 1876 DRM_ERROR("Failed initializing GTT heap.\n"); 1877 return r; 1878 } 1879 DRM_INFO("amdgpu: %uM of GTT memory ready.\n", 1880 (unsigned)(gtt_size / (1024 * 1024))); 1881 1882 /* Initialize various on-chip memory pools */ 1883 r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_GDS, adev->gds.gds_size); 1884 if (r) { 1885 DRM_ERROR("Failed initializing GDS heap.\n"); 1886 return r; 1887 } 1888 1889 r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_GWS, adev->gds.gws_size); 1890 if (r) { 1891 DRM_ERROR("Failed initializing gws heap.\n"); 1892 return r; 1893 } 1894 1895 r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_OA, adev->gds.oa_size); 1896 if (r) { 1897 DRM_ERROR("Failed initializing oa heap.\n"); 1898 return r; 1899 } 1900 1901 return 0; 1902 } 1903 1904 /** 1905 * amdgpu_ttm_late_init - Handle any late initialization for amdgpu_ttm 1906 */ 1907 void amdgpu_ttm_late_init(struct amdgpu_device *adev) 1908 { 1909 /* return the VGA stolen memory (if any) back to VRAM */ 1910 if (!adev->mman.keep_stolen_vga_memory) 1911 amdgpu_bo_free_kernel(&adev->mman.stolen_vga_memory, NULL, NULL); 1912 amdgpu_bo_free_kernel(&adev->mman.stolen_extended_memory, NULL, NULL); 1913 } 1914 1915 /** 1916 * amdgpu_ttm_fini - De-initialize the TTM memory pools 1917 */ 1918 void amdgpu_ttm_fini(struct amdgpu_device *adev) 1919 { 1920 if (!adev->mman.initialized) 1921 return; 1922 1923 amdgpu_ttm_training_reserve_vram_fini(adev); 1924 /* return the stolen vga memory back to VRAM */ 1925 if (adev->mman.keep_stolen_vga_memory) 1926 amdgpu_bo_free_kernel(&adev->mman.stolen_vga_memory, NULL, NULL); 1927 /* return the IP Discovery TMR memory back to VRAM */ 1928 amdgpu_bo_free_kernel(&adev->mman.discovery_memory, NULL, NULL); 1929 amdgpu_ttm_fw_reserve_vram_fini(adev); 1930 1931 if (adev->mman.aper_base_kaddr) 1932 iounmap(adev->mman.aper_base_kaddr); 1933 adev->mman.aper_base_kaddr = NULL; 1934 1935 amdgpu_vram_mgr_fini(adev); 1936 amdgpu_gtt_mgr_fini(adev); 1937 ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_GDS); 1938 ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_GWS); 1939 ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_OA); 1940 ttm_bo_device_release(&adev->mman.bdev); 1941 adev->mman.initialized = false; 1942 DRM_INFO("amdgpu: ttm finalized\n"); 1943 } 1944 1945 /** 1946 * amdgpu_ttm_set_buffer_funcs_status - enable/disable use of buffer functions 1947 * 1948 * @adev: amdgpu_device pointer 1949 * @enable: true when we can use buffer functions. 1950 * 1951 * Enable/disable use of buffer functions during suspend/resume. This should 1952 * only be called at bootup or when userspace isn't running. 1953 */ 1954 void amdgpu_ttm_set_buffer_funcs_status(struct amdgpu_device *adev, bool enable) 1955 { 1956 struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM); 1957 uint64_t size; 1958 int r; 1959 1960 if (!adev->mman.initialized || amdgpu_in_reset(adev) || 1961 adev->mman.buffer_funcs_enabled == enable) 1962 return; 1963 1964 if (enable) { 1965 struct amdgpu_ring *ring; 1966 struct drm_gpu_scheduler *sched; 1967 1968 ring = adev->mman.buffer_funcs_ring; 1969 sched = &ring->sched; 1970 r = drm_sched_entity_init(&adev->mman.entity, 1971 DRM_SCHED_PRIORITY_KERNEL, &sched, 1972 1, NULL); 1973 if (r) { 1974 DRM_ERROR("Failed setting up TTM BO move entity (%d)\n", 1975 r); 1976 return; 1977 } 1978 } else { 1979 drm_sched_entity_destroy(&adev->mman.entity); 1980 dma_fence_put(man->move); 1981 man->move = NULL; 1982 } 1983 1984 /* this just adjusts TTM size idea, which sets lpfn to the correct value */ 1985 if (enable) 1986 size = adev->gmc.real_vram_size; 1987 else 1988 size = adev->gmc.visible_vram_size; 1989 man->size = size >> PAGE_SHIFT; 1990 adev->mman.buffer_funcs_enabled = enable; 1991 } 1992 1993 static vm_fault_t amdgpu_ttm_fault(struct vm_fault *vmf) 1994 { 1995 struct ttm_buffer_object *bo = vmf->vma->vm_private_data; 1996 vm_fault_t ret; 1997 1998 ret = ttm_bo_vm_reserve(bo, vmf); 1999 if (ret) 2000 return ret; 2001 2002 ret = amdgpu_bo_fault_reserve_notify(bo); 2003 if (ret) 2004 goto unlock; 2005 2006 ret = ttm_bo_vm_fault_reserved(vmf, vmf->vma->vm_page_prot, 2007 TTM_BO_VM_NUM_PREFAULT, 1); 2008 if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) 2009 return ret; 2010 2011 unlock: 2012 dma_resv_unlock(bo->base.resv); 2013 return ret; 2014 } 2015 2016 static struct vm_operations_struct amdgpu_ttm_vm_ops = { 2017 .fault = amdgpu_ttm_fault, 2018 .open = ttm_bo_vm_open, 2019 .close = ttm_bo_vm_close, 2020 .access = ttm_bo_vm_access 2021 }; 2022 2023 int amdgpu_mmap(struct file *filp, struct vm_area_struct *vma) 2024 { 2025 struct drm_file *file_priv = filp->private_data; 2026 struct amdgpu_device *adev = drm_to_adev(file_priv->minor->dev); 2027 int r; 2028 2029 r = ttm_bo_mmap(filp, vma, &adev->mman.bdev); 2030 if (unlikely(r != 0)) 2031 return r; 2032 2033 vma->vm_ops = &amdgpu_ttm_vm_ops; 2034 return 0; 2035 } 2036 2037 int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset, 2038 uint64_t dst_offset, uint32_t byte_count, 2039 struct dma_resv *resv, 2040 struct dma_fence **fence, bool direct_submit, 2041 bool vm_needs_flush, bool tmz) 2042 { 2043 enum amdgpu_ib_pool_type pool = direct_submit ? AMDGPU_IB_POOL_DIRECT : 2044 AMDGPU_IB_POOL_DELAYED; 2045 struct amdgpu_device *adev = ring->adev; 2046 struct amdgpu_job *job; 2047 2048 uint32_t max_bytes; 2049 unsigned num_loops, num_dw; 2050 unsigned i; 2051 int r; 2052 2053 if (direct_submit && !ring->sched.ready) { 2054 DRM_ERROR("Trying to move memory with ring turned off.\n"); 2055 return -EINVAL; 2056 } 2057 2058 max_bytes = adev->mman.buffer_funcs->copy_max_bytes; 2059 num_loops = DIV_ROUND_UP(byte_count, max_bytes); 2060 num_dw = ALIGN(num_loops * adev->mman.buffer_funcs->copy_num_dw, 8); 2061 2062 r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, pool, &job); 2063 if (r) 2064 return r; 2065 2066 if (vm_needs_flush) { 2067 job->vm_pd_addr = amdgpu_gmc_pd_addr(adev->gart.bo); 2068 job->vm_needs_flush = true; 2069 } 2070 if (resv) { 2071 r = amdgpu_sync_resv(adev, &job->sync, resv, 2072 AMDGPU_SYNC_ALWAYS, 2073 AMDGPU_FENCE_OWNER_UNDEFINED); 2074 if (r) { 2075 DRM_ERROR("sync failed (%d).\n", r); 2076 goto error_free; 2077 } 2078 } 2079 2080 for (i = 0; i < num_loops; i++) { 2081 uint32_t cur_size_in_bytes = min(byte_count, max_bytes); 2082 2083 amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_offset, 2084 dst_offset, cur_size_in_bytes, tmz); 2085 2086 src_offset += cur_size_in_bytes; 2087 dst_offset += cur_size_in_bytes; 2088 byte_count -= cur_size_in_bytes; 2089 } 2090 2091 amdgpu_ring_pad_ib(ring, &job->ibs[0]); 2092 WARN_ON(job->ibs[0].length_dw > num_dw); 2093 if (direct_submit) 2094 r = amdgpu_job_submit_direct(job, ring, fence); 2095 else 2096 r = amdgpu_job_submit(job, &adev->mman.entity, 2097 AMDGPU_FENCE_OWNER_UNDEFINED, fence); 2098 if (r) 2099 goto error_free; 2100 2101 return r; 2102 2103 error_free: 2104 amdgpu_job_free(job); 2105 DRM_ERROR("Error scheduling IBs (%d)\n", r); 2106 return r; 2107 } 2108 2109 int amdgpu_fill_buffer(struct amdgpu_bo *bo, 2110 uint32_t src_data, 2111 struct dma_resv *resv, 2112 struct dma_fence **fence) 2113 { 2114 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); 2115 uint32_t max_bytes = adev->mman.buffer_funcs->fill_max_bytes; 2116 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring; 2117 2118 struct drm_mm_node *mm_node; 2119 unsigned long num_pages; 2120 unsigned int num_loops, num_dw; 2121 2122 struct amdgpu_job *job; 2123 int r; 2124 2125 if (!adev->mman.buffer_funcs_enabled) { 2126 DRM_ERROR("Trying to clear memory with ring turned off.\n"); 2127 return -EINVAL; 2128 } 2129 2130 if (bo->tbo.mem.mem_type == TTM_PL_TT) { 2131 r = amdgpu_ttm_alloc_gart(&bo->tbo); 2132 if (r) 2133 return r; 2134 } 2135 2136 num_pages = bo->tbo.num_pages; 2137 mm_node = bo->tbo.mem.mm_node; 2138 num_loops = 0; 2139 while (num_pages) { 2140 uint64_t byte_count = mm_node->size << PAGE_SHIFT; 2141 2142 num_loops += DIV_ROUND_UP_ULL(byte_count, max_bytes); 2143 num_pages -= mm_node->size; 2144 ++mm_node; 2145 } 2146 num_dw = num_loops * adev->mman.buffer_funcs->fill_num_dw; 2147 2148 /* for IB padding */ 2149 num_dw += 64; 2150 2151 r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, AMDGPU_IB_POOL_DELAYED, 2152 &job); 2153 if (r) 2154 return r; 2155 2156 if (resv) { 2157 r = amdgpu_sync_resv(adev, &job->sync, resv, 2158 AMDGPU_SYNC_ALWAYS, 2159 AMDGPU_FENCE_OWNER_UNDEFINED); 2160 if (r) { 2161 DRM_ERROR("sync failed (%d).\n", r); 2162 goto error_free; 2163 } 2164 } 2165 2166 num_pages = bo->tbo.num_pages; 2167 mm_node = bo->tbo.mem.mm_node; 2168 2169 while (num_pages) { 2170 uint64_t byte_count = mm_node->size << PAGE_SHIFT; 2171 uint64_t dst_addr; 2172 2173 dst_addr = amdgpu_mm_node_addr(&bo->tbo, mm_node, &bo->tbo.mem); 2174 while (byte_count) { 2175 uint32_t cur_size_in_bytes = min_t(uint64_t, byte_count, 2176 max_bytes); 2177 2178 amdgpu_emit_fill_buffer(adev, &job->ibs[0], src_data, 2179 dst_addr, cur_size_in_bytes); 2180 2181 dst_addr += cur_size_in_bytes; 2182 byte_count -= cur_size_in_bytes; 2183 } 2184 2185 num_pages -= mm_node->size; 2186 ++mm_node; 2187 } 2188 2189 amdgpu_ring_pad_ib(ring, &job->ibs[0]); 2190 WARN_ON(job->ibs[0].length_dw > num_dw); 2191 r = amdgpu_job_submit(job, &adev->mman.entity, 2192 AMDGPU_FENCE_OWNER_UNDEFINED, fence); 2193 if (r) 2194 goto error_free; 2195 2196 return 0; 2197 2198 error_free: 2199 amdgpu_job_free(job); 2200 return r; 2201 } 2202 2203 #if defined(CONFIG_DEBUG_FS) 2204 2205 static int amdgpu_mm_dump_table(struct seq_file *m, void *data) 2206 { 2207 struct drm_info_node *node = (struct drm_info_node *)m->private; 2208 unsigned ttm_pl = (uintptr_t)node->info_ent->data; 2209 struct drm_device *dev = node->minor->dev; 2210 struct amdgpu_device *adev = drm_to_adev(dev); 2211 struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev, ttm_pl); 2212 struct drm_printer p = drm_seq_file_printer(m); 2213 2214 man->func->debug(man, &p); 2215 return 0; 2216 } 2217 2218 static int amdgpu_ttm_pool_debugfs(struct seq_file *m, void *data) 2219 { 2220 struct drm_info_node *node = (struct drm_info_node *)m->private; 2221 struct drm_device *dev = node->minor->dev; 2222 struct amdgpu_device *adev = drm_to_adev(dev); 2223 2224 return ttm_pool_debugfs(&adev->mman.bdev.pool, m); 2225 } 2226 2227 static const struct drm_info_list amdgpu_ttm_debugfs_list[] = { 2228 {"amdgpu_vram_mm", amdgpu_mm_dump_table, 0, (void *)TTM_PL_VRAM}, 2229 {"amdgpu_gtt_mm", amdgpu_mm_dump_table, 0, (void *)TTM_PL_TT}, 2230 {"amdgpu_gds_mm", amdgpu_mm_dump_table, 0, (void *)AMDGPU_PL_GDS}, 2231 {"amdgpu_gws_mm", amdgpu_mm_dump_table, 0, (void *)AMDGPU_PL_GWS}, 2232 {"amdgpu_oa_mm", amdgpu_mm_dump_table, 0, (void *)AMDGPU_PL_OA}, 2233 {"ttm_page_pool", amdgpu_ttm_pool_debugfs, 0, NULL}, 2234 }; 2235 2236 /** 2237 * amdgpu_ttm_vram_read - Linear read access to VRAM 2238 * 2239 * Accesses VRAM via MMIO for debugging purposes. 2240 */ 2241 static ssize_t amdgpu_ttm_vram_read(struct file *f, char __user *buf, 2242 size_t size, loff_t *pos) 2243 { 2244 struct amdgpu_device *adev = file_inode(f)->i_private; 2245 ssize_t result = 0; 2246 2247 if (size & 0x3 || *pos & 0x3) 2248 return -EINVAL; 2249 2250 if (*pos >= adev->gmc.mc_vram_size) 2251 return -ENXIO; 2252 2253 size = min(size, (size_t)(adev->gmc.mc_vram_size - *pos)); 2254 while (size) { 2255 size_t bytes = min(size, AMDGPU_TTM_VRAM_MAX_DW_READ * 4); 2256 uint32_t value[AMDGPU_TTM_VRAM_MAX_DW_READ]; 2257 2258 amdgpu_device_vram_access(adev, *pos, value, bytes, false); 2259 if (copy_to_user(buf, value, bytes)) 2260 return -EFAULT; 2261 2262 result += bytes; 2263 buf += bytes; 2264 *pos += bytes; 2265 size -= bytes; 2266 } 2267 2268 return result; 2269 } 2270 2271 /** 2272 * amdgpu_ttm_vram_write - Linear write access to VRAM 2273 * 2274 * Accesses VRAM via MMIO for debugging purposes. 2275 */ 2276 static ssize_t amdgpu_ttm_vram_write(struct file *f, const char __user *buf, 2277 size_t size, loff_t *pos) 2278 { 2279 struct amdgpu_device *adev = file_inode(f)->i_private; 2280 ssize_t result = 0; 2281 int r; 2282 2283 if (size & 0x3 || *pos & 0x3) 2284 return -EINVAL; 2285 2286 if (*pos >= adev->gmc.mc_vram_size) 2287 return -ENXIO; 2288 2289 while (size) { 2290 unsigned long flags; 2291 uint32_t value; 2292 2293 if (*pos >= adev->gmc.mc_vram_size) 2294 return result; 2295 2296 r = get_user(value, (uint32_t *)buf); 2297 if (r) 2298 return r; 2299 2300 spin_lock_irqsave(&adev->mmio_idx_lock, flags); 2301 WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)*pos) | 0x80000000); 2302 WREG32_NO_KIQ(mmMM_INDEX_HI, *pos >> 31); 2303 WREG32_NO_KIQ(mmMM_DATA, value); 2304 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags); 2305 2306 result += 4; 2307 buf += 4; 2308 *pos += 4; 2309 size -= 4; 2310 } 2311 2312 return result; 2313 } 2314 2315 static const struct file_operations amdgpu_ttm_vram_fops = { 2316 .owner = THIS_MODULE, 2317 .read = amdgpu_ttm_vram_read, 2318 .write = amdgpu_ttm_vram_write, 2319 .llseek = default_llseek, 2320 }; 2321 2322 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS 2323 2324 /** 2325 * amdgpu_ttm_gtt_read - Linear read access to GTT memory 2326 */ 2327 static ssize_t amdgpu_ttm_gtt_read(struct file *f, char __user *buf, 2328 size_t size, loff_t *pos) 2329 { 2330 struct amdgpu_device *adev = file_inode(f)->i_private; 2331 ssize_t result = 0; 2332 int r; 2333 2334 while (size) { 2335 loff_t p = *pos / PAGE_SIZE; 2336 unsigned off = *pos & ~PAGE_MASK; 2337 size_t cur_size = min_t(size_t, size, PAGE_SIZE - off); 2338 struct page *page; 2339 void *ptr; 2340 2341 if (p >= adev->gart.num_cpu_pages) 2342 return result; 2343 2344 page = adev->gart.pages[p]; 2345 if (page) { 2346 ptr = kmap(page); 2347 ptr += off; 2348 2349 r = copy_to_user(buf, ptr, cur_size); 2350 kunmap(adev->gart.pages[p]); 2351 } else 2352 r = clear_user(buf, cur_size); 2353 2354 if (r) 2355 return -EFAULT; 2356 2357 result += cur_size; 2358 buf += cur_size; 2359 *pos += cur_size; 2360 size -= cur_size; 2361 } 2362 2363 return result; 2364 } 2365 2366 static const struct file_operations amdgpu_ttm_gtt_fops = { 2367 .owner = THIS_MODULE, 2368 .read = amdgpu_ttm_gtt_read, 2369 .llseek = default_llseek 2370 }; 2371 2372 #endif 2373 2374 /** 2375 * amdgpu_iomem_read - Virtual read access to GPU mapped memory 2376 * 2377 * This function is used to read memory that has been mapped to the 2378 * GPU and the known addresses are not physical addresses but instead 2379 * bus addresses (e.g., what you'd put in an IB or ring buffer). 2380 */ 2381 static ssize_t amdgpu_iomem_read(struct file *f, char __user *buf, 2382 size_t size, loff_t *pos) 2383 { 2384 struct amdgpu_device *adev = file_inode(f)->i_private; 2385 struct iommu_domain *dom; 2386 ssize_t result = 0; 2387 int r; 2388 2389 /* retrieve the IOMMU domain if any for this device */ 2390 dom = iommu_get_domain_for_dev(adev->dev); 2391 2392 while (size) { 2393 phys_addr_t addr = *pos & PAGE_MASK; 2394 loff_t off = *pos & ~PAGE_MASK; 2395 size_t bytes = PAGE_SIZE - off; 2396 unsigned long pfn; 2397 struct page *p; 2398 void *ptr; 2399 2400 bytes = bytes < size ? bytes : size; 2401 2402 /* Translate the bus address to a physical address. If 2403 * the domain is NULL it means there is no IOMMU active 2404 * and the address translation is the identity 2405 */ 2406 addr = dom ? iommu_iova_to_phys(dom, addr) : addr; 2407 2408 pfn = addr >> PAGE_SHIFT; 2409 if (!pfn_valid(pfn)) 2410 return -EPERM; 2411 2412 p = pfn_to_page(pfn); 2413 if (p->mapping != adev->mman.bdev.dev_mapping) 2414 return -EPERM; 2415 2416 ptr = kmap(p); 2417 r = copy_to_user(buf, ptr + off, bytes); 2418 kunmap(p); 2419 if (r) 2420 return -EFAULT; 2421 2422 size -= bytes; 2423 *pos += bytes; 2424 result += bytes; 2425 } 2426 2427 return result; 2428 } 2429 2430 /** 2431 * amdgpu_iomem_write - Virtual write access to GPU mapped memory 2432 * 2433 * This function is used to write memory that has been mapped to the 2434 * GPU and the known addresses are not physical addresses but instead 2435 * bus addresses (e.g., what you'd put in an IB or ring buffer). 2436 */ 2437 static ssize_t amdgpu_iomem_write(struct file *f, const char __user *buf, 2438 size_t size, loff_t *pos) 2439 { 2440 struct amdgpu_device *adev = file_inode(f)->i_private; 2441 struct iommu_domain *dom; 2442 ssize_t result = 0; 2443 int r; 2444 2445 dom = iommu_get_domain_for_dev(adev->dev); 2446 2447 while (size) { 2448 phys_addr_t addr = *pos & PAGE_MASK; 2449 loff_t off = *pos & ~PAGE_MASK; 2450 size_t bytes = PAGE_SIZE - off; 2451 unsigned long pfn; 2452 struct page *p; 2453 void *ptr; 2454 2455 bytes = bytes < size ? bytes : size; 2456 2457 addr = dom ? iommu_iova_to_phys(dom, addr) : addr; 2458 2459 pfn = addr >> PAGE_SHIFT; 2460 if (!pfn_valid(pfn)) 2461 return -EPERM; 2462 2463 p = pfn_to_page(pfn); 2464 if (p->mapping != adev->mman.bdev.dev_mapping) 2465 return -EPERM; 2466 2467 ptr = kmap(p); 2468 r = copy_from_user(ptr + off, buf, bytes); 2469 kunmap(p); 2470 if (r) 2471 return -EFAULT; 2472 2473 size -= bytes; 2474 *pos += bytes; 2475 result += bytes; 2476 } 2477 2478 return result; 2479 } 2480 2481 static const struct file_operations amdgpu_ttm_iomem_fops = { 2482 .owner = THIS_MODULE, 2483 .read = amdgpu_iomem_read, 2484 .write = amdgpu_iomem_write, 2485 .llseek = default_llseek 2486 }; 2487 2488 static const struct { 2489 char *name; 2490 const struct file_operations *fops; 2491 int domain; 2492 } ttm_debugfs_entries[] = { 2493 { "amdgpu_vram", &amdgpu_ttm_vram_fops, TTM_PL_VRAM }, 2494 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS 2495 { "amdgpu_gtt", &amdgpu_ttm_gtt_fops, TTM_PL_TT }, 2496 #endif 2497 { "amdgpu_iomem", &amdgpu_ttm_iomem_fops, TTM_PL_SYSTEM }, 2498 }; 2499 2500 #endif 2501 2502 int amdgpu_ttm_debugfs_init(struct amdgpu_device *adev) 2503 { 2504 #if defined(CONFIG_DEBUG_FS) 2505 unsigned count; 2506 2507 struct drm_minor *minor = adev_to_drm(adev)->primary; 2508 struct dentry *ent, *root = minor->debugfs_root; 2509 2510 for (count = 0; count < ARRAY_SIZE(ttm_debugfs_entries); count++) { 2511 ent = debugfs_create_file( 2512 ttm_debugfs_entries[count].name, 2513 S_IFREG | S_IRUGO, root, 2514 adev, 2515 ttm_debugfs_entries[count].fops); 2516 if (IS_ERR(ent)) 2517 return PTR_ERR(ent); 2518 if (ttm_debugfs_entries[count].domain == TTM_PL_VRAM) 2519 i_size_write(ent->d_inode, adev->gmc.mc_vram_size); 2520 else if (ttm_debugfs_entries[count].domain == TTM_PL_TT) 2521 i_size_write(ent->d_inode, adev->gmc.gart_size); 2522 adev->mman.debugfs_entries[count] = ent; 2523 } 2524 2525 count = ARRAY_SIZE(amdgpu_ttm_debugfs_list); 2526 return amdgpu_debugfs_add_files(adev, amdgpu_ttm_debugfs_list, count); 2527 #else 2528 return 0; 2529 #endif 2530 } 2531