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/pagemap.h> 36 #include <linux/sched/task.h> 37 #include <linux/sched/mm.h> 38 #include <linux/seq_file.h> 39 #include <linux/slab.h> 40 #include <linux/swap.h> 41 #include <linux/swiotlb.h> 42 #include <linux/dma-buf.h> 43 #include <linux/sizes.h> 44 45 #include <drm/ttm/ttm_bo_api.h> 46 #include <drm/ttm/ttm_bo_driver.h> 47 #include <drm/ttm/ttm_placement.h> 48 #include <drm/ttm/ttm_range_manager.h> 49 50 #include <drm/amdgpu_drm.h> 51 52 #include "amdgpu.h" 53 #include "amdgpu_object.h" 54 #include "amdgpu_trace.h" 55 #include "amdgpu_amdkfd.h" 56 #include "amdgpu_sdma.h" 57 #include "amdgpu_ras.h" 58 #include "amdgpu_atomfirmware.h" 59 #include "amdgpu_res_cursor.h" 60 #include "bif/bif_4_1_d.h" 61 62 #define AMDGPU_TTM_VRAM_MAX_DW_READ (size_t)128 63 64 static int amdgpu_ttm_backend_bind(struct ttm_device *bdev, 65 struct ttm_tt *ttm, 66 struct ttm_resource *bo_mem); 67 static void amdgpu_ttm_backend_unbind(struct ttm_device *bdev, 68 struct ttm_tt *ttm); 69 70 static int amdgpu_ttm_init_on_chip(struct amdgpu_device *adev, 71 unsigned int type, 72 uint64_t size_in_page) 73 { 74 return ttm_range_man_init(&adev->mman.bdev, type, 75 false, size_in_page); 76 } 77 78 /** 79 * amdgpu_evict_flags - Compute placement flags 80 * 81 * @bo: The buffer object to evict 82 * @placement: Possible destination(s) for evicted BO 83 * 84 * Fill in placement data when ttm_bo_evict() is called 85 */ 86 static void amdgpu_evict_flags(struct ttm_buffer_object *bo, 87 struct ttm_placement *placement) 88 { 89 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev); 90 struct amdgpu_bo *abo; 91 static const struct ttm_place placements = { 92 .fpfn = 0, 93 .lpfn = 0, 94 .mem_type = TTM_PL_SYSTEM, 95 .flags = 0 96 }; 97 98 /* Don't handle scatter gather BOs */ 99 if (bo->type == ttm_bo_type_sg) { 100 placement->num_placement = 0; 101 placement->num_busy_placement = 0; 102 return; 103 } 104 105 /* Object isn't an AMDGPU object so ignore */ 106 if (!amdgpu_bo_is_amdgpu_bo(bo)) { 107 placement->placement = &placements; 108 placement->busy_placement = &placements; 109 placement->num_placement = 1; 110 placement->num_busy_placement = 1; 111 return; 112 } 113 114 abo = ttm_to_amdgpu_bo(bo); 115 if (abo->flags & AMDGPU_AMDKFD_CREATE_SVM_BO) { 116 struct dma_fence *fence; 117 struct dma_resv *resv = &bo->base._resv; 118 119 rcu_read_lock(); 120 fence = rcu_dereference(resv->fence_excl); 121 if (fence && !fence->ops->signaled) 122 dma_fence_enable_sw_signaling(fence); 123 124 placement->num_placement = 0; 125 placement->num_busy_placement = 0; 126 rcu_read_unlock(); 127 return; 128 } 129 130 switch (bo->resource->mem_type) { 131 case AMDGPU_PL_GDS: 132 case AMDGPU_PL_GWS: 133 case AMDGPU_PL_OA: 134 placement->num_placement = 0; 135 placement->num_busy_placement = 0; 136 return; 137 138 case TTM_PL_VRAM: 139 if (!adev->mman.buffer_funcs_enabled) { 140 /* Move to system memory */ 141 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU); 142 } else if (!amdgpu_gmc_vram_full_visible(&adev->gmc) && 143 !(abo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED) && 144 amdgpu_bo_in_cpu_visible_vram(abo)) { 145 146 /* Try evicting to the CPU inaccessible part of VRAM 147 * first, but only set GTT as busy placement, so this 148 * BO will be evicted to GTT rather than causing other 149 * BOs to be evicted from VRAM 150 */ 151 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM | 152 AMDGPU_GEM_DOMAIN_GTT | 153 AMDGPU_GEM_DOMAIN_CPU); 154 abo->placements[0].fpfn = adev->gmc.visible_vram_size >> PAGE_SHIFT; 155 abo->placements[0].lpfn = 0; 156 abo->placement.busy_placement = &abo->placements[1]; 157 abo->placement.num_busy_placement = 1; 158 } else { 159 /* Move to GTT memory */ 160 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_GTT | 161 AMDGPU_GEM_DOMAIN_CPU); 162 } 163 break; 164 case TTM_PL_TT: 165 case AMDGPU_PL_PREEMPT: 166 default: 167 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU); 168 break; 169 } 170 *placement = abo->placement; 171 } 172 173 /** 174 * amdgpu_ttm_map_buffer - Map memory into the GART windows 175 * @bo: buffer object to map 176 * @mem: memory object to map 177 * @mm_cur: range to map 178 * @num_pages: number of pages to map 179 * @window: which GART window to use 180 * @ring: DMA ring to use for the copy 181 * @tmz: if we should setup a TMZ enabled mapping 182 * @addr: resulting address inside the MC address space 183 * 184 * Setup one of the GART windows to access a specific piece of memory or return 185 * the physical address for local memory. 186 */ 187 static int amdgpu_ttm_map_buffer(struct ttm_buffer_object *bo, 188 struct ttm_resource *mem, 189 struct amdgpu_res_cursor *mm_cur, 190 unsigned num_pages, unsigned window, 191 struct amdgpu_ring *ring, bool tmz, 192 uint64_t *addr) 193 { 194 struct amdgpu_device *adev = ring->adev; 195 struct amdgpu_job *job; 196 unsigned num_dw, num_bytes; 197 struct dma_fence *fence; 198 uint64_t src_addr, dst_addr; 199 void *cpu_addr; 200 uint64_t flags; 201 unsigned int i; 202 int r; 203 204 BUG_ON(adev->mman.buffer_funcs->copy_max_bytes < 205 AMDGPU_GTT_MAX_TRANSFER_SIZE * 8); 206 BUG_ON(mem->mem_type == AMDGPU_PL_PREEMPT); 207 208 /* Map only what can't be accessed directly */ 209 if (!tmz && mem->start != AMDGPU_BO_INVALID_OFFSET) { 210 *addr = amdgpu_ttm_domain_start(adev, mem->mem_type) + 211 mm_cur->start; 212 return 0; 213 } 214 215 *addr = adev->gmc.gart_start; 216 *addr += (u64)window * AMDGPU_GTT_MAX_TRANSFER_SIZE * 217 AMDGPU_GPU_PAGE_SIZE; 218 *addr += mm_cur->start & ~PAGE_MASK; 219 220 num_dw = ALIGN(adev->mman.buffer_funcs->copy_num_dw, 8); 221 num_bytes = num_pages * 8 * AMDGPU_GPU_PAGES_IN_CPU_PAGE; 222 223 r = amdgpu_job_alloc_with_ib(adev, num_dw * 4 + num_bytes, 224 AMDGPU_IB_POOL_DELAYED, &job); 225 if (r) 226 return r; 227 228 src_addr = num_dw * 4; 229 src_addr += job->ibs[0].gpu_addr; 230 231 dst_addr = amdgpu_bo_gpu_offset(adev->gart.bo); 232 dst_addr += window * AMDGPU_GTT_MAX_TRANSFER_SIZE * 8; 233 amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_addr, 234 dst_addr, num_bytes, false); 235 236 amdgpu_ring_pad_ib(ring, &job->ibs[0]); 237 WARN_ON(job->ibs[0].length_dw > num_dw); 238 239 flags = amdgpu_ttm_tt_pte_flags(adev, bo->ttm, mem); 240 if (tmz) 241 flags |= AMDGPU_PTE_TMZ; 242 243 cpu_addr = &job->ibs[0].ptr[num_dw]; 244 245 if (mem->mem_type == TTM_PL_TT) { 246 dma_addr_t *dma_addr; 247 248 dma_addr = &bo->ttm->dma_address[mm_cur->start >> PAGE_SHIFT]; 249 r = amdgpu_gart_map(adev, 0, num_pages, dma_addr, flags, 250 cpu_addr); 251 if (r) 252 goto error_free; 253 } else { 254 dma_addr_t dma_address; 255 256 dma_address = mm_cur->start; 257 dma_address += adev->vm_manager.vram_base_offset; 258 259 for (i = 0; i < num_pages; ++i) { 260 r = amdgpu_gart_map(adev, i << PAGE_SHIFT, 1, 261 &dma_address, flags, cpu_addr); 262 if (r) 263 goto error_free; 264 265 dma_address += PAGE_SIZE; 266 } 267 } 268 269 r = amdgpu_job_submit(job, &adev->mman.entity, 270 AMDGPU_FENCE_OWNER_UNDEFINED, &fence); 271 if (r) 272 goto error_free; 273 274 dma_fence_put(fence); 275 276 return r; 277 278 error_free: 279 amdgpu_job_free(job); 280 return r; 281 } 282 283 /** 284 * amdgpu_ttm_copy_mem_to_mem - Helper function for copy 285 * @adev: amdgpu device 286 * @src: buffer/address where to read from 287 * @dst: buffer/address where to write to 288 * @size: number of bytes to copy 289 * @tmz: if a secure copy should be used 290 * @resv: resv object to sync to 291 * @f: Returns the last fence if multiple jobs are submitted. 292 * 293 * The function copies @size bytes from {src->mem + src->offset} to 294 * {dst->mem + dst->offset}. src->bo and dst->bo could be same BO for a 295 * move and different for a BO to BO copy. 296 * 297 */ 298 int amdgpu_ttm_copy_mem_to_mem(struct amdgpu_device *adev, 299 const struct amdgpu_copy_mem *src, 300 const struct amdgpu_copy_mem *dst, 301 uint64_t size, bool tmz, 302 struct dma_resv *resv, 303 struct dma_fence **f) 304 { 305 const uint32_t GTT_MAX_BYTES = (AMDGPU_GTT_MAX_TRANSFER_SIZE * 306 AMDGPU_GPU_PAGE_SIZE); 307 308 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring; 309 struct amdgpu_res_cursor src_mm, dst_mm; 310 struct dma_fence *fence = NULL; 311 int r = 0; 312 313 if (!adev->mman.buffer_funcs_enabled) { 314 DRM_ERROR("Trying to move memory with ring turned off.\n"); 315 return -EINVAL; 316 } 317 318 amdgpu_res_first(src->mem, src->offset, size, &src_mm); 319 amdgpu_res_first(dst->mem, dst->offset, size, &dst_mm); 320 321 mutex_lock(&adev->mman.gtt_window_lock); 322 while (src_mm.remaining) { 323 uint32_t src_page_offset = src_mm.start & ~PAGE_MASK; 324 uint32_t dst_page_offset = dst_mm.start & ~PAGE_MASK; 325 struct dma_fence *next; 326 uint32_t cur_size; 327 uint64_t from, to; 328 329 /* Copy size cannot exceed GTT_MAX_BYTES. So if src or dst 330 * begins at an offset, then adjust the size accordingly 331 */ 332 cur_size = max(src_page_offset, dst_page_offset); 333 cur_size = min(min3(src_mm.size, dst_mm.size, size), 334 (uint64_t)(GTT_MAX_BYTES - cur_size)); 335 336 /* Map src to window 0 and dst to window 1. */ 337 r = amdgpu_ttm_map_buffer(src->bo, src->mem, &src_mm, 338 PFN_UP(cur_size + src_page_offset), 339 0, ring, tmz, &from); 340 if (r) 341 goto error; 342 343 r = amdgpu_ttm_map_buffer(dst->bo, dst->mem, &dst_mm, 344 PFN_UP(cur_size + dst_page_offset), 345 1, ring, tmz, &to); 346 if (r) 347 goto error; 348 349 r = amdgpu_copy_buffer(ring, from, to, cur_size, 350 resv, &next, false, true, tmz); 351 if (r) 352 goto error; 353 354 dma_fence_put(fence); 355 fence = next; 356 357 amdgpu_res_next(&src_mm, cur_size); 358 amdgpu_res_next(&dst_mm, cur_size); 359 } 360 error: 361 mutex_unlock(&adev->mman.gtt_window_lock); 362 if (f) 363 *f = dma_fence_get(fence); 364 dma_fence_put(fence); 365 return r; 366 } 367 368 /* 369 * amdgpu_move_blit - Copy an entire buffer to another buffer 370 * 371 * This is a helper called by amdgpu_bo_move() and amdgpu_move_vram_ram() to 372 * help move buffers to and from VRAM. 373 */ 374 static int amdgpu_move_blit(struct ttm_buffer_object *bo, 375 bool evict, 376 struct ttm_resource *new_mem, 377 struct ttm_resource *old_mem) 378 { 379 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev); 380 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo); 381 struct amdgpu_copy_mem src, dst; 382 struct dma_fence *fence = NULL; 383 int r; 384 385 src.bo = bo; 386 dst.bo = bo; 387 src.mem = old_mem; 388 dst.mem = new_mem; 389 src.offset = 0; 390 dst.offset = 0; 391 392 r = amdgpu_ttm_copy_mem_to_mem(adev, &src, &dst, 393 new_mem->num_pages << PAGE_SHIFT, 394 amdgpu_bo_encrypted(abo), 395 bo->base.resv, &fence); 396 if (r) 397 goto error; 398 399 /* clear the space being freed */ 400 if (old_mem->mem_type == TTM_PL_VRAM && 401 (abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE)) { 402 struct dma_fence *wipe_fence = NULL; 403 404 r = amdgpu_fill_buffer(ttm_to_amdgpu_bo(bo), AMDGPU_POISON, 405 NULL, &wipe_fence); 406 if (r) { 407 goto error; 408 } else if (wipe_fence) { 409 dma_fence_put(fence); 410 fence = wipe_fence; 411 } 412 } 413 414 /* Always block for VM page tables before committing the new location */ 415 if (bo->type == ttm_bo_type_kernel) 416 r = ttm_bo_move_accel_cleanup(bo, fence, true, false, new_mem); 417 else 418 r = ttm_bo_move_accel_cleanup(bo, fence, evict, true, new_mem); 419 dma_fence_put(fence); 420 return r; 421 422 error: 423 if (fence) 424 dma_fence_wait(fence, false); 425 dma_fence_put(fence); 426 return r; 427 } 428 429 /* 430 * amdgpu_mem_visible - Check that memory can be accessed by ttm_bo_move_memcpy 431 * 432 * Called by amdgpu_bo_move() 433 */ 434 static bool amdgpu_mem_visible(struct amdgpu_device *adev, 435 struct ttm_resource *mem) 436 { 437 uint64_t mem_size = (u64)mem->num_pages << PAGE_SHIFT; 438 struct amdgpu_res_cursor cursor; 439 440 if (mem->mem_type == TTM_PL_SYSTEM || 441 mem->mem_type == TTM_PL_TT) 442 return true; 443 if (mem->mem_type != TTM_PL_VRAM) 444 return false; 445 446 amdgpu_res_first(mem, 0, mem_size, &cursor); 447 448 /* ttm_resource_ioremap only supports contiguous memory */ 449 if (cursor.size != mem_size) 450 return false; 451 452 return cursor.start + cursor.size <= adev->gmc.visible_vram_size; 453 } 454 455 /* 456 * amdgpu_bo_move - Move a buffer object to a new memory location 457 * 458 * Called by ttm_bo_handle_move_mem() 459 */ 460 static int amdgpu_bo_move(struct ttm_buffer_object *bo, bool evict, 461 struct ttm_operation_ctx *ctx, 462 struct ttm_resource *new_mem, 463 struct ttm_place *hop) 464 { 465 struct amdgpu_device *adev; 466 struct amdgpu_bo *abo; 467 struct ttm_resource *old_mem = bo->resource; 468 int r; 469 470 if (new_mem->mem_type == TTM_PL_TT || 471 new_mem->mem_type == AMDGPU_PL_PREEMPT) { 472 r = amdgpu_ttm_backend_bind(bo->bdev, bo->ttm, new_mem); 473 if (r) 474 return r; 475 } 476 477 /* Can't move a pinned BO */ 478 abo = ttm_to_amdgpu_bo(bo); 479 if (WARN_ON_ONCE(abo->tbo.pin_count > 0)) 480 return -EINVAL; 481 482 adev = amdgpu_ttm_adev(bo->bdev); 483 484 if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) { 485 ttm_bo_move_null(bo, new_mem); 486 goto out; 487 } 488 if (old_mem->mem_type == TTM_PL_SYSTEM && 489 (new_mem->mem_type == TTM_PL_TT || 490 new_mem->mem_type == AMDGPU_PL_PREEMPT)) { 491 ttm_bo_move_null(bo, new_mem); 492 goto out; 493 } 494 if ((old_mem->mem_type == TTM_PL_TT || 495 old_mem->mem_type == AMDGPU_PL_PREEMPT) && 496 new_mem->mem_type == TTM_PL_SYSTEM) { 497 r = ttm_bo_wait_ctx(bo, ctx); 498 if (r) 499 return r; 500 501 amdgpu_ttm_backend_unbind(bo->bdev, bo->ttm); 502 ttm_resource_free(bo, &bo->resource); 503 ttm_bo_assign_mem(bo, new_mem); 504 goto out; 505 } 506 507 if (old_mem->mem_type == AMDGPU_PL_GDS || 508 old_mem->mem_type == AMDGPU_PL_GWS || 509 old_mem->mem_type == AMDGPU_PL_OA || 510 new_mem->mem_type == AMDGPU_PL_GDS || 511 new_mem->mem_type == AMDGPU_PL_GWS || 512 new_mem->mem_type == AMDGPU_PL_OA) { 513 /* Nothing to save here */ 514 ttm_bo_move_null(bo, new_mem); 515 goto out; 516 } 517 518 if (adev->mman.buffer_funcs_enabled) { 519 if (((old_mem->mem_type == TTM_PL_SYSTEM && 520 new_mem->mem_type == TTM_PL_VRAM) || 521 (old_mem->mem_type == TTM_PL_VRAM && 522 new_mem->mem_type == TTM_PL_SYSTEM))) { 523 hop->fpfn = 0; 524 hop->lpfn = 0; 525 hop->mem_type = TTM_PL_TT; 526 hop->flags = TTM_PL_FLAG_TEMPORARY; 527 return -EMULTIHOP; 528 } 529 530 r = amdgpu_move_blit(bo, evict, new_mem, old_mem); 531 } else { 532 r = -ENODEV; 533 } 534 535 if (r) { 536 /* Check that all memory is CPU accessible */ 537 if (!amdgpu_mem_visible(adev, old_mem) || 538 !amdgpu_mem_visible(adev, new_mem)) { 539 pr_err("Move buffer fallback to memcpy unavailable\n"); 540 return r; 541 } 542 543 r = ttm_bo_move_memcpy(bo, ctx, new_mem); 544 if (r) 545 return r; 546 } 547 548 if (bo->type == ttm_bo_type_device && 549 new_mem->mem_type == TTM_PL_VRAM && 550 old_mem->mem_type != TTM_PL_VRAM) { 551 /* amdgpu_bo_fault_reserve_notify will re-set this if the CPU 552 * accesses the BO after it's moved. 553 */ 554 abo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED; 555 } 556 557 out: 558 /* update statistics */ 559 atomic64_add(bo->base.size, &adev->num_bytes_moved); 560 amdgpu_bo_move_notify(bo, evict, new_mem); 561 return 0; 562 } 563 564 /* 565 * amdgpu_ttm_io_mem_reserve - Reserve a block of memory during a fault 566 * 567 * Called by ttm_mem_io_reserve() ultimately via ttm_bo_vm_fault() 568 */ 569 static int amdgpu_ttm_io_mem_reserve(struct ttm_device *bdev, 570 struct ttm_resource *mem) 571 { 572 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev); 573 size_t bus_size = (size_t)mem->num_pages << PAGE_SHIFT; 574 575 switch (mem->mem_type) { 576 case TTM_PL_SYSTEM: 577 /* system memory */ 578 return 0; 579 case TTM_PL_TT: 580 case AMDGPU_PL_PREEMPT: 581 break; 582 case TTM_PL_VRAM: 583 mem->bus.offset = mem->start << PAGE_SHIFT; 584 /* check if it's visible */ 585 if ((mem->bus.offset + bus_size) > adev->gmc.visible_vram_size) 586 return -EINVAL; 587 588 if (adev->mman.aper_base_kaddr && 589 mem->placement & TTM_PL_FLAG_CONTIGUOUS) 590 mem->bus.addr = (u8 *)adev->mman.aper_base_kaddr + 591 mem->bus.offset; 592 593 mem->bus.offset += adev->gmc.aper_base; 594 mem->bus.is_iomem = true; 595 break; 596 default: 597 return -EINVAL; 598 } 599 return 0; 600 } 601 602 static unsigned long amdgpu_ttm_io_mem_pfn(struct ttm_buffer_object *bo, 603 unsigned long page_offset) 604 { 605 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev); 606 struct amdgpu_res_cursor cursor; 607 608 amdgpu_res_first(bo->resource, (u64)page_offset << PAGE_SHIFT, 0, 609 &cursor); 610 return (adev->gmc.aper_base + cursor.start) >> PAGE_SHIFT; 611 } 612 613 /** 614 * amdgpu_ttm_domain_start - Returns GPU start address 615 * @adev: amdgpu device object 616 * @type: type of the memory 617 * 618 * Returns: 619 * GPU start address of a memory domain 620 */ 621 622 uint64_t amdgpu_ttm_domain_start(struct amdgpu_device *adev, uint32_t type) 623 { 624 switch (type) { 625 case TTM_PL_TT: 626 return adev->gmc.gart_start; 627 case TTM_PL_VRAM: 628 return adev->gmc.vram_start; 629 } 630 631 return 0; 632 } 633 634 /* 635 * TTM backend functions. 636 */ 637 struct amdgpu_ttm_tt { 638 struct ttm_tt ttm; 639 struct drm_gem_object *gobj; 640 u64 offset; 641 uint64_t userptr; 642 struct task_struct *usertask; 643 uint32_t userflags; 644 bool bound; 645 #if IS_ENABLED(CONFIG_DRM_AMDGPU_USERPTR) 646 struct hmm_range *range; 647 #endif 648 }; 649 650 #ifdef CONFIG_DRM_AMDGPU_USERPTR 651 /* 652 * amdgpu_ttm_tt_get_user_pages - get device accessible pages that back user 653 * memory and start HMM tracking CPU page table update 654 * 655 * Calling function must call amdgpu_ttm_tt_userptr_range_done() once and only 656 * once afterwards to stop HMM tracking 657 */ 658 int amdgpu_ttm_tt_get_user_pages(struct amdgpu_bo *bo, struct page **pages) 659 { 660 struct ttm_tt *ttm = bo->tbo.ttm; 661 struct amdgpu_ttm_tt *gtt = (void *)ttm; 662 unsigned long start = gtt->userptr; 663 struct vm_area_struct *vma; 664 struct mm_struct *mm; 665 bool readonly; 666 int r = 0; 667 668 mm = bo->notifier.mm; 669 if (unlikely(!mm)) { 670 DRM_DEBUG_DRIVER("BO is not registered?\n"); 671 return -EFAULT; 672 } 673 674 /* Another get_user_pages is running at the same time?? */ 675 if (WARN_ON(gtt->range)) 676 return -EFAULT; 677 678 if (!mmget_not_zero(mm)) /* Happens during process shutdown */ 679 return -ESRCH; 680 681 mmap_read_lock(mm); 682 vma = vma_lookup(mm, start); 683 if (unlikely(!vma)) { 684 r = -EFAULT; 685 goto out_unlock; 686 } 687 if (unlikely((gtt->userflags & AMDGPU_GEM_USERPTR_ANONONLY) && 688 vma->vm_file)) { 689 r = -EPERM; 690 goto out_unlock; 691 } 692 693 readonly = amdgpu_ttm_tt_is_readonly(ttm); 694 r = amdgpu_hmm_range_get_pages(&bo->notifier, mm, pages, start, 695 ttm->num_pages, >t->range, readonly, 696 true, NULL); 697 out_unlock: 698 mmap_read_unlock(mm); 699 mmput(mm); 700 701 return r; 702 } 703 704 /* 705 * amdgpu_ttm_tt_userptr_range_done - stop HMM track the CPU page table change 706 * Check if the pages backing this ttm range have been invalidated 707 * 708 * Returns: true if pages are still valid 709 */ 710 bool amdgpu_ttm_tt_get_user_pages_done(struct ttm_tt *ttm) 711 { 712 struct amdgpu_ttm_tt *gtt = (void *)ttm; 713 bool r = false; 714 715 if (!gtt || !gtt->userptr) 716 return false; 717 718 DRM_DEBUG_DRIVER("user_pages_done 0x%llx pages 0x%x\n", 719 gtt->userptr, ttm->num_pages); 720 721 WARN_ONCE(!gtt->range || !gtt->range->hmm_pfns, 722 "No user pages to check\n"); 723 724 if (gtt->range) { 725 /* 726 * FIXME: Must always hold notifier_lock for this, and must 727 * not ignore the return code. 728 */ 729 r = amdgpu_hmm_range_get_pages_done(gtt->range); 730 gtt->range = NULL; 731 } 732 733 return !r; 734 } 735 #endif 736 737 /* 738 * amdgpu_ttm_tt_set_user_pages - Copy pages in, putting old pages as necessary. 739 * 740 * Called by amdgpu_cs_list_validate(). This creates the page list 741 * that backs user memory and will ultimately be mapped into the device 742 * address space. 743 */ 744 void amdgpu_ttm_tt_set_user_pages(struct ttm_tt *ttm, struct page **pages) 745 { 746 unsigned long i; 747 748 for (i = 0; i < ttm->num_pages; ++i) 749 ttm->pages[i] = pages ? pages[i] : NULL; 750 } 751 752 /* 753 * amdgpu_ttm_tt_pin_userptr - prepare the sg table with the user pages 754 * 755 * Called by amdgpu_ttm_backend_bind() 756 **/ 757 static int amdgpu_ttm_tt_pin_userptr(struct ttm_device *bdev, 758 struct ttm_tt *ttm) 759 { 760 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev); 761 struct amdgpu_ttm_tt *gtt = (void *)ttm; 762 int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY); 763 enum dma_data_direction direction = write ? 764 DMA_BIDIRECTIONAL : DMA_TO_DEVICE; 765 int r; 766 767 /* Allocate an SG array and squash pages into it */ 768 r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0, 769 (u64)ttm->num_pages << PAGE_SHIFT, 770 GFP_KERNEL); 771 if (r) 772 goto release_sg; 773 774 /* Map SG to device */ 775 r = dma_map_sgtable(adev->dev, ttm->sg, direction, 0); 776 if (r) 777 goto release_sg; 778 779 /* convert SG to linear array of pages and dma addresses */ 780 drm_prime_sg_to_dma_addr_array(ttm->sg, gtt->ttm.dma_address, 781 ttm->num_pages); 782 783 return 0; 784 785 release_sg: 786 kfree(ttm->sg); 787 ttm->sg = NULL; 788 return r; 789 } 790 791 /* 792 * amdgpu_ttm_tt_unpin_userptr - Unpin and unmap userptr pages 793 */ 794 static void amdgpu_ttm_tt_unpin_userptr(struct ttm_device *bdev, 795 struct ttm_tt *ttm) 796 { 797 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev); 798 struct amdgpu_ttm_tt *gtt = (void *)ttm; 799 int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY); 800 enum dma_data_direction direction = write ? 801 DMA_BIDIRECTIONAL : DMA_TO_DEVICE; 802 803 /* double check that we don't free the table twice */ 804 if (!ttm->sg || !ttm->sg->sgl) 805 return; 806 807 /* unmap the pages mapped to the device */ 808 dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0); 809 sg_free_table(ttm->sg); 810 811 #if IS_ENABLED(CONFIG_DRM_AMDGPU_USERPTR) 812 if (gtt->range) { 813 unsigned long i; 814 815 for (i = 0; i < ttm->num_pages; i++) { 816 if (ttm->pages[i] != 817 hmm_pfn_to_page(gtt->range->hmm_pfns[i])) 818 break; 819 } 820 821 WARN((i == ttm->num_pages), "Missing get_user_page_done\n"); 822 } 823 #endif 824 } 825 826 static int amdgpu_ttm_gart_bind(struct amdgpu_device *adev, 827 struct ttm_buffer_object *tbo, 828 uint64_t flags) 829 { 830 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(tbo); 831 struct ttm_tt *ttm = tbo->ttm; 832 struct amdgpu_ttm_tt *gtt = (void *)ttm; 833 int r; 834 835 if (amdgpu_bo_encrypted(abo)) 836 flags |= AMDGPU_PTE_TMZ; 837 838 if (abo->flags & AMDGPU_GEM_CREATE_CP_MQD_GFX9) { 839 uint64_t page_idx = 1; 840 841 r = amdgpu_gart_bind(adev, gtt->offset, page_idx, 842 gtt->ttm.dma_address, flags); 843 if (r) 844 goto gart_bind_fail; 845 846 /* The memory type of the first page defaults to UC. Now 847 * modify the memory type to NC from the second page of 848 * the BO onward. 849 */ 850 flags &= ~AMDGPU_PTE_MTYPE_VG10_MASK; 851 flags |= AMDGPU_PTE_MTYPE_VG10(AMDGPU_MTYPE_NC); 852 853 r = amdgpu_gart_bind(adev, 854 gtt->offset + (page_idx << PAGE_SHIFT), 855 ttm->num_pages - page_idx, 856 &(gtt->ttm.dma_address[page_idx]), flags); 857 } else { 858 r = amdgpu_gart_bind(adev, gtt->offset, ttm->num_pages, 859 gtt->ttm.dma_address, flags); 860 } 861 862 gart_bind_fail: 863 if (r) 864 DRM_ERROR("failed to bind %u pages at 0x%08llX\n", 865 ttm->num_pages, gtt->offset); 866 867 return r; 868 } 869 870 /* 871 * amdgpu_ttm_backend_bind - Bind GTT memory 872 * 873 * Called by ttm_tt_bind() on behalf of ttm_bo_handle_move_mem(). 874 * This handles binding GTT memory to the device address space. 875 */ 876 static int amdgpu_ttm_backend_bind(struct ttm_device *bdev, 877 struct ttm_tt *ttm, 878 struct ttm_resource *bo_mem) 879 { 880 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev); 881 struct amdgpu_ttm_tt *gtt = (void*)ttm; 882 uint64_t flags; 883 int r = 0; 884 885 if (!bo_mem) 886 return -EINVAL; 887 888 if (gtt->bound) 889 return 0; 890 891 if (gtt->userptr) { 892 r = amdgpu_ttm_tt_pin_userptr(bdev, ttm); 893 if (r) { 894 DRM_ERROR("failed to pin userptr\n"); 895 return r; 896 } 897 } else if (ttm->page_flags & TTM_PAGE_FLAG_SG) { 898 if (!ttm->sg) { 899 struct dma_buf_attachment *attach; 900 struct sg_table *sgt; 901 902 attach = gtt->gobj->import_attach; 903 sgt = dma_buf_map_attachment(attach, DMA_BIDIRECTIONAL); 904 if (IS_ERR(sgt)) 905 return PTR_ERR(sgt); 906 907 ttm->sg = sgt; 908 } 909 910 drm_prime_sg_to_dma_addr_array(ttm->sg, gtt->ttm.dma_address, 911 ttm->num_pages); 912 } 913 914 if (!ttm->num_pages) { 915 WARN(1, "nothing to bind %u pages for mreg %p back %p!\n", 916 ttm->num_pages, bo_mem, ttm); 917 } 918 919 if (bo_mem->mem_type == AMDGPU_PL_GDS || 920 bo_mem->mem_type == AMDGPU_PL_GWS || 921 bo_mem->mem_type == AMDGPU_PL_OA) 922 return -EINVAL; 923 924 if (bo_mem->mem_type != TTM_PL_TT || 925 !amdgpu_gtt_mgr_has_gart_addr(bo_mem)) { 926 gtt->offset = AMDGPU_BO_INVALID_OFFSET; 927 return 0; 928 } 929 930 /* compute PTE flags relevant to this BO memory */ 931 flags = amdgpu_ttm_tt_pte_flags(adev, ttm, bo_mem); 932 933 /* bind pages into GART page tables */ 934 gtt->offset = (u64)bo_mem->start << PAGE_SHIFT; 935 r = amdgpu_gart_bind(adev, gtt->offset, ttm->num_pages, 936 gtt->ttm.dma_address, flags); 937 938 if (r) 939 DRM_ERROR("failed to bind %u pages at 0x%08llX\n", 940 ttm->num_pages, gtt->offset); 941 gtt->bound = true; 942 return r; 943 } 944 945 /* 946 * amdgpu_ttm_alloc_gart - Make sure buffer object is accessible either 947 * through AGP or GART aperture. 948 * 949 * If bo is accessible through AGP aperture, then use AGP aperture 950 * to access bo; otherwise allocate logical space in GART aperture 951 * and map bo to GART aperture. 952 */ 953 int amdgpu_ttm_alloc_gart(struct ttm_buffer_object *bo) 954 { 955 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev); 956 struct ttm_operation_ctx ctx = { false, false }; 957 struct amdgpu_ttm_tt *gtt = (void *)bo->ttm; 958 struct ttm_placement placement; 959 struct ttm_place placements; 960 struct ttm_resource *tmp; 961 uint64_t addr, flags; 962 int r; 963 964 if (bo->resource->start != AMDGPU_BO_INVALID_OFFSET) 965 return 0; 966 967 addr = amdgpu_gmc_agp_addr(bo); 968 if (addr != AMDGPU_BO_INVALID_OFFSET) { 969 bo->resource->start = addr >> PAGE_SHIFT; 970 return 0; 971 } 972 973 /* allocate GART space */ 974 placement.num_placement = 1; 975 placement.placement = &placements; 976 placement.num_busy_placement = 1; 977 placement.busy_placement = &placements; 978 placements.fpfn = 0; 979 placements.lpfn = adev->gmc.gart_size >> PAGE_SHIFT; 980 placements.mem_type = TTM_PL_TT; 981 placements.flags = bo->resource->placement; 982 983 r = ttm_bo_mem_space(bo, &placement, &tmp, &ctx); 984 if (unlikely(r)) 985 return r; 986 987 /* compute PTE flags for this buffer object */ 988 flags = amdgpu_ttm_tt_pte_flags(adev, bo->ttm, tmp); 989 990 /* Bind pages */ 991 gtt->offset = (u64)tmp->start << PAGE_SHIFT; 992 r = amdgpu_ttm_gart_bind(adev, bo, flags); 993 if (unlikely(r)) { 994 ttm_resource_free(bo, &tmp); 995 return r; 996 } 997 998 amdgpu_gart_invalidate_tlb(adev); 999 ttm_resource_free(bo, &bo->resource); 1000 ttm_bo_assign_mem(bo, tmp); 1001 1002 return 0; 1003 } 1004 1005 /* 1006 * amdgpu_ttm_recover_gart - Rebind GTT pages 1007 * 1008 * Called by amdgpu_gtt_mgr_recover() from amdgpu_device_reset() to 1009 * rebind GTT pages during a GPU reset. 1010 */ 1011 int amdgpu_ttm_recover_gart(struct ttm_buffer_object *tbo) 1012 { 1013 struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev); 1014 uint64_t flags; 1015 int r; 1016 1017 if (!tbo->ttm) 1018 return 0; 1019 1020 flags = amdgpu_ttm_tt_pte_flags(adev, tbo->ttm, tbo->resource); 1021 r = amdgpu_ttm_gart_bind(adev, tbo, flags); 1022 1023 return r; 1024 } 1025 1026 /* 1027 * amdgpu_ttm_backend_unbind - Unbind GTT mapped pages 1028 * 1029 * Called by ttm_tt_unbind() on behalf of ttm_bo_move_ttm() and 1030 * ttm_tt_destroy(). 1031 */ 1032 static void amdgpu_ttm_backend_unbind(struct ttm_device *bdev, 1033 struct ttm_tt *ttm) 1034 { 1035 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev); 1036 struct amdgpu_ttm_tt *gtt = (void *)ttm; 1037 int r; 1038 1039 /* if the pages have userptr pinning then clear that first */ 1040 if (gtt->userptr) { 1041 amdgpu_ttm_tt_unpin_userptr(bdev, ttm); 1042 } else if (ttm->sg && gtt->gobj->import_attach) { 1043 struct dma_buf_attachment *attach; 1044 1045 attach = gtt->gobj->import_attach; 1046 dma_buf_unmap_attachment(attach, ttm->sg, DMA_BIDIRECTIONAL); 1047 ttm->sg = NULL; 1048 } 1049 1050 if (!gtt->bound) 1051 return; 1052 1053 if (gtt->offset == AMDGPU_BO_INVALID_OFFSET) 1054 return; 1055 1056 /* unbind shouldn't be done for GDS/GWS/OA in ttm_bo_clean_mm */ 1057 r = amdgpu_gart_unbind(adev, gtt->offset, ttm->num_pages); 1058 if (r) 1059 DRM_ERROR("failed to unbind %u pages at 0x%08llX\n", 1060 gtt->ttm.num_pages, gtt->offset); 1061 gtt->bound = false; 1062 } 1063 1064 static void amdgpu_ttm_backend_destroy(struct ttm_device *bdev, 1065 struct ttm_tt *ttm) 1066 { 1067 struct amdgpu_ttm_tt *gtt = (void *)ttm; 1068 1069 amdgpu_ttm_backend_unbind(bdev, ttm); 1070 ttm_tt_destroy_common(bdev, ttm); 1071 if (gtt->usertask) 1072 put_task_struct(gtt->usertask); 1073 1074 ttm_tt_fini(>t->ttm); 1075 kfree(gtt); 1076 } 1077 1078 /** 1079 * amdgpu_ttm_tt_create - Create a ttm_tt object for a given BO 1080 * 1081 * @bo: The buffer object to create a GTT ttm_tt object around 1082 * @page_flags: Page flags to be added to the ttm_tt object 1083 * 1084 * Called by ttm_tt_create(). 1085 */ 1086 static struct ttm_tt *amdgpu_ttm_tt_create(struct ttm_buffer_object *bo, 1087 uint32_t page_flags) 1088 { 1089 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo); 1090 struct amdgpu_ttm_tt *gtt; 1091 enum ttm_caching caching; 1092 1093 gtt = kzalloc(sizeof(struct amdgpu_ttm_tt), GFP_KERNEL); 1094 if (gtt == NULL) { 1095 return NULL; 1096 } 1097 gtt->gobj = &bo->base; 1098 1099 if (abo->flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC) 1100 caching = ttm_write_combined; 1101 else 1102 caching = ttm_cached; 1103 1104 /* allocate space for the uninitialized page entries */ 1105 if (ttm_sg_tt_init(>t->ttm, bo, page_flags, caching)) { 1106 kfree(gtt); 1107 return NULL; 1108 } 1109 return >t->ttm; 1110 } 1111 1112 /* 1113 * amdgpu_ttm_tt_populate - Map GTT pages visible to the device 1114 * 1115 * Map the pages of a ttm_tt object to an address space visible 1116 * to the underlying device. 1117 */ 1118 static int amdgpu_ttm_tt_populate(struct ttm_device *bdev, 1119 struct ttm_tt *ttm, 1120 struct ttm_operation_ctx *ctx) 1121 { 1122 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev); 1123 struct amdgpu_ttm_tt *gtt = (void *)ttm; 1124 1125 /* user pages are bound by amdgpu_ttm_tt_pin_userptr() */ 1126 if (gtt && gtt->userptr) { 1127 ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL); 1128 if (!ttm->sg) 1129 return -ENOMEM; 1130 return 0; 1131 } 1132 1133 if (ttm->page_flags & TTM_PAGE_FLAG_SG) 1134 return 0; 1135 1136 return ttm_pool_alloc(&adev->mman.bdev.pool, ttm, ctx); 1137 } 1138 1139 /* 1140 * amdgpu_ttm_tt_unpopulate - unmap GTT pages and unpopulate page arrays 1141 * 1142 * Unmaps pages of a ttm_tt object from the device address space and 1143 * unpopulates the page array backing it. 1144 */ 1145 static void amdgpu_ttm_tt_unpopulate(struct ttm_device *bdev, 1146 struct ttm_tt *ttm) 1147 { 1148 struct amdgpu_ttm_tt *gtt = (void *)ttm; 1149 struct amdgpu_device *adev; 1150 1151 if (gtt && gtt->userptr) { 1152 amdgpu_ttm_tt_set_user_pages(ttm, NULL); 1153 kfree(ttm->sg); 1154 ttm->sg = NULL; 1155 return; 1156 } 1157 1158 if (ttm->page_flags & TTM_PAGE_FLAG_SG) 1159 return; 1160 1161 adev = amdgpu_ttm_adev(bdev); 1162 return ttm_pool_free(&adev->mman.bdev.pool, ttm); 1163 } 1164 1165 /** 1166 * amdgpu_ttm_tt_set_userptr - Initialize userptr GTT ttm_tt for the current 1167 * task 1168 * 1169 * @bo: The ttm_buffer_object to bind this userptr to 1170 * @addr: The address in the current tasks VM space to use 1171 * @flags: Requirements of userptr object. 1172 * 1173 * Called by amdgpu_gem_userptr_ioctl() to bind userptr pages 1174 * to current task 1175 */ 1176 int amdgpu_ttm_tt_set_userptr(struct ttm_buffer_object *bo, 1177 uint64_t addr, uint32_t flags) 1178 { 1179 struct amdgpu_ttm_tt *gtt; 1180 1181 if (!bo->ttm) { 1182 /* TODO: We want a separate TTM object type for userptrs */ 1183 bo->ttm = amdgpu_ttm_tt_create(bo, 0); 1184 if (bo->ttm == NULL) 1185 return -ENOMEM; 1186 } 1187 1188 /* Set TTM_PAGE_FLAG_SG before populate but after create. */ 1189 bo->ttm->page_flags |= TTM_PAGE_FLAG_SG; 1190 1191 gtt = (void *)bo->ttm; 1192 gtt->userptr = addr; 1193 gtt->userflags = flags; 1194 1195 if (gtt->usertask) 1196 put_task_struct(gtt->usertask); 1197 gtt->usertask = current->group_leader; 1198 get_task_struct(gtt->usertask); 1199 1200 return 0; 1201 } 1202 1203 /* 1204 * amdgpu_ttm_tt_get_usermm - Return memory manager for ttm_tt object 1205 */ 1206 struct mm_struct *amdgpu_ttm_tt_get_usermm(struct ttm_tt *ttm) 1207 { 1208 struct amdgpu_ttm_tt *gtt = (void *)ttm; 1209 1210 if (gtt == NULL) 1211 return NULL; 1212 1213 if (gtt->usertask == NULL) 1214 return NULL; 1215 1216 return gtt->usertask->mm; 1217 } 1218 1219 /* 1220 * amdgpu_ttm_tt_affect_userptr - Determine if a ttm_tt object lays inside an 1221 * address range for the current task. 1222 * 1223 */ 1224 bool amdgpu_ttm_tt_affect_userptr(struct ttm_tt *ttm, unsigned long start, 1225 unsigned long end) 1226 { 1227 struct amdgpu_ttm_tt *gtt = (void *)ttm; 1228 unsigned long size; 1229 1230 if (gtt == NULL || !gtt->userptr) 1231 return false; 1232 1233 /* Return false if no part of the ttm_tt object lies within 1234 * the range 1235 */ 1236 size = (unsigned long)gtt->ttm.num_pages * PAGE_SIZE; 1237 if (gtt->userptr > end || gtt->userptr + size <= start) 1238 return false; 1239 1240 return true; 1241 } 1242 1243 /* 1244 * amdgpu_ttm_tt_is_userptr - Have the pages backing by userptr? 1245 */ 1246 bool amdgpu_ttm_tt_is_userptr(struct ttm_tt *ttm) 1247 { 1248 struct amdgpu_ttm_tt *gtt = (void *)ttm; 1249 1250 if (gtt == NULL || !gtt->userptr) 1251 return false; 1252 1253 return true; 1254 } 1255 1256 /* 1257 * amdgpu_ttm_tt_is_readonly - Is the ttm_tt object read only? 1258 */ 1259 bool amdgpu_ttm_tt_is_readonly(struct ttm_tt *ttm) 1260 { 1261 struct amdgpu_ttm_tt *gtt = (void *)ttm; 1262 1263 if (gtt == NULL) 1264 return false; 1265 1266 return !!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY); 1267 } 1268 1269 /** 1270 * amdgpu_ttm_tt_pde_flags - Compute PDE flags for ttm_tt object 1271 * 1272 * @ttm: The ttm_tt object to compute the flags for 1273 * @mem: The memory registry backing this ttm_tt object 1274 * 1275 * Figure out the flags to use for a VM PDE (Page Directory Entry). 1276 */ 1277 uint64_t amdgpu_ttm_tt_pde_flags(struct ttm_tt *ttm, struct ttm_resource *mem) 1278 { 1279 uint64_t flags = 0; 1280 1281 if (mem && mem->mem_type != TTM_PL_SYSTEM) 1282 flags |= AMDGPU_PTE_VALID; 1283 1284 if (mem && (mem->mem_type == TTM_PL_TT || 1285 mem->mem_type == AMDGPU_PL_PREEMPT)) { 1286 flags |= AMDGPU_PTE_SYSTEM; 1287 1288 if (ttm->caching == ttm_cached) 1289 flags |= AMDGPU_PTE_SNOOPED; 1290 } 1291 1292 if (mem && mem->mem_type == TTM_PL_VRAM && 1293 mem->bus.caching == ttm_cached) 1294 flags |= AMDGPU_PTE_SNOOPED; 1295 1296 return flags; 1297 } 1298 1299 /** 1300 * amdgpu_ttm_tt_pte_flags - Compute PTE flags for ttm_tt object 1301 * 1302 * @adev: amdgpu_device pointer 1303 * @ttm: The ttm_tt object to compute the flags for 1304 * @mem: The memory registry backing this ttm_tt object 1305 * 1306 * Figure out the flags to use for a VM PTE (Page Table Entry). 1307 */ 1308 uint64_t amdgpu_ttm_tt_pte_flags(struct amdgpu_device *adev, struct ttm_tt *ttm, 1309 struct ttm_resource *mem) 1310 { 1311 uint64_t flags = amdgpu_ttm_tt_pde_flags(ttm, mem); 1312 1313 flags |= adev->gart.gart_pte_flags; 1314 flags |= AMDGPU_PTE_READABLE; 1315 1316 if (!amdgpu_ttm_tt_is_readonly(ttm)) 1317 flags |= AMDGPU_PTE_WRITEABLE; 1318 1319 return flags; 1320 } 1321 1322 /* 1323 * amdgpu_ttm_bo_eviction_valuable - Check to see if we can evict a buffer 1324 * object. 1325 * 1326 * Return true if eviction is sensible. Called by ttm_mem_evict_first() on 1327 * behalf of ttm_bo_mem_force_space() which tries to evict buffer objects until 1328 * it can find space for a new object and by ttm_bo_force_list_clean() which is 1329 * used to clean out a memory space. 1330 */ 1331 static bool amdgpu_ttm_bo_eviction_valuable(struct ttm_buffer_object *bo, 1332 const struct ttm_place *place) 1333 { 1334 unsigned long num_pages = bo->resource->num_pages; 1335 struct amdgpu_res_cursor cursor; 1336 struct dma_resv_list *flist; 1337 struct dma_fence *f; 1338 int i; 1339 1340 /* Swapout? */ 1341 if (bo->resource->mem_type == TTM_PL_SYSTEM) 1342 return true; 1343 1344 if (bo->type == ttm_bo_type_kernel && 1345 !amdgpu_vm_evictable(ttm_to_amdgpu_bo(bo))) 1346 return false; 1347 1348 /* If bo is a KFD BO, check if the bo belongs to the current process. 1349 * If true, then return false as any KFD process needs all its BOs to 1350 * be resident to run successfully 1351 */ 1352 flist = dma_resv_shared_list(bo->base.resv); 1353 if (flist) { 1354 for (i = 0; i < flist->shared_count; ++i) { 1355 f = rcu_dereference_protected(flist->shared[i], 1356 dma_resv_held(bo->base.resv)); 1357 if (amdkfd_fence_check_mm(f, current->mm)) 1358 return false; 1359 } 1360 } 1361 1362 switch (bo->resource->mem_type) { 1363 case AMDGPU_PL_PREEMPT: 1364 /* Preemptible BOs don't own system resources managed by the 1365 * driver (pages, VRAM, GART space). They point to resources 1366 * owned by someone else (e.g. pageable memory in user mode 1367 * or a DMABuf). They are used in a preemptible context so we 1368 * can guarantee no deadlocks and good QoS in case of MMU 1369 * notifiers or DMABuf move notifiers from the resource owner. 1370 */ 1371 return false; 1372 case TTM_PL_TT: 1373 if (amdgpu_bo_is_amdgpu_bo(bo) && 1374 amdgpu_bo_encrypted(ttm_to_amdgpu_bo(bo))) 1375 return false; 1376 return true; 1377 1378 case TTM_PL_VRAM: 1379 /* Check each drm MM node individually */ 1380 amdgpu_res_first(bo->resource, 0, (u64)num_pages << PAGE_SHIFT, 1381 &cursor); 1382 while (cursor.remaining) { 1383 if (place->fpfn < PFN_DOWN(cursor.start + cursor.size) 1384 && !(place->lpfn && 1385 place->lpfn <= PFN_DOWN(cursor.start))) 1386 return true; 1387 1388 amdgpu_res_next(&cursor, cursor.size); 1389 } 1390 return false; 1391 1392 default: 1393 break; 1394 } 1395 1396 return ttm_bo_eviction_valuable(bo, place); 1397 } 1398 1399 /** 1400 * amdgpu_ttm_access_memory - Read or Write memory that backs a buffer object. 1401 * 1402 * @bo: The buffer object to read/write 1403 * @offset: Offset into buffer object 1404 * @buf: Secondary buffer to write/read from 1405 * @len: Length in bytes of access 1406 * @write: true if writing 1407 * 1408 * This is used to access VRAM that backs a buffer object via MMIO 1409 * access for debugging purposes. 1410 */ 1411 static int amdgpu_ttm_access_memory(struct ttm_buffer_object *bo, 1412 unsigned long offset, void *buf, int len, 1413 int write) 1414 { 1415 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo); 1416 struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev); 1417 struct amdgpu_res_cursor cursor; 1418 unsigned long flags; 1419 uint32_t value = 0; 1420 int ret = 0; 1421 1422 if (bo->resource->mem_type != TTM_PL_VRAM) 1423 return -EIO; 1424 1425 amdgpu_res_first(bo->resource, offset, len, &cursor); 1426 while (cursor.remaining) { 1427 uint64_t aligned_pos = cursor.start & ~(uint64_t)3; 1428 uint64_t bytes = 4 - (cursor.start & 3); 1429 uint32_t shift = (cursor.start & 3) * 8; 1430 uint32_t mask = 0xffffffff << shift; 1431 1432 if (cursor.size < bytes) { 1433 mask &= 0xffffffff >> (bytes - cursor.size) * 8; 1434 bytes = cursor.size; 1435 } 1436 1437 if (mask != 0xffffffff) { 1438 spin_lock_irqsave(&adev->mmio_idx_lock, flags); 1439 WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)aligned_pos) | 0x80000000); 1440 WREG32_NO_KIQ(mmMM_INDEX_HI, aligned_pos >> 31); 1441 value = RREG32_NO_KIQ(mmMM_DATA); 1442 if (write) { 1443 value &= ~mask; 1444 value |= (*(uint32_t *)buf << shift) & mask; 1445 WREG32_NO_KIQ(mmMM_DATA, value); 1446 } 1447 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags); 1448 if (!write) { 1449 value = (value & mask) >> shift; 1450 memcpy(buf, &value, bytes); 1451 } 1452 } else { 1453 bytes = cursor.size & ~0x3ULL; 1454 amdgpu_device_vram_access(adev, cursor.start, 1455 (uint32_t *)buf, bytes, 1456 write); 1457 } 1458 1459 ret += bytes; 1460 buf = (uint8_t *)buf + bytes; 1461 amdgpu_res_next(&cursor, bytes); 1462 } 1463 1464 return ret; 1465 } 1466 1467 static void 1468 amdgpu_bo_delete_mem_notify(struct ttm_buffer_object *bo) 1469 { 1470 amdgpu_bo_move_notify(bo, false, NULL); 1471 } 1472 1473 static struct ttm_device_funcs amdgpu_bo_driver = { 1474 .ttm_tt_create = &amdgpu_ttm_tt_create, 1475 .ttm_tt_populate = &amdgpu_ttm_tt_populate, 1476 .ttm_tt_unpopulate = &amdgpu_ttm_tt_unpopulate, 1477 .ttm_tt_destroy = &amdgpu_ttm_backend_destroy, 1478 .eviction_valuable = amdgpu_ttm_bo_eviction_valuable, 1479 .evict_flags = &amdgpu_evict_flags, 1480 .move = &amdgpu_bo_move, 1481 .delete_mem_notify = &amdgpu_bo_delete_mem_notify, 1482 .release_notify = &amdgpu_bo_release_notify, 1483 .io_mem_reserve = &amdgpu_ttm_io_mem_reserve, 1484 .io_mem_pfn = amdgpu_ttm_io_mem_pfn, 1485 .access_memory = &amdgpu_ttm_access_memory, 1486 .del_from_lru_notify = &amdgpu_vm_del_from_lru_notify 1487 }; 1488 1489 /* 1490 * Firmware Reservation functions 1491 */ 1492 /** 1493 * amdgpu_ttm_fw_reserve_vram_fini - free fw reserved vram 1494 * 1495 * @adev: amdgpu_device pointer 1496 * 1497 * free fw reserved vram if it has been reserved. 1498 */ 1499 static void amdgpu_ttm_fw_reserve_vram_fini(struct amdgpu_device *adev) 1500 { 1501 amdgpu_bo_free_kernel(&adev->mman.fw_vram_usage_reserved_bo, 1502 NULL, &adev->mman.fw_vram_usage_va); 1503 } 1504 1505 /** 1506 * amdgpu_ttm_fw_reserve_vram_init - create bo vram reservation from fw 1507 * 1508 * @adev: amdgpu_device pointer 1509 * 1510 * create bo vram reservation from fw. 1511 */ 1512 static int amdgpu_ttm_fw_reserve_vram_init(struct amdgpu_device *adev) 1513 { 1514 uint64_t vram_size = adev->gmc.visible_vram_size; 1515 1516 adev->mman.fw_vram_usage_va = NULL; 1517 adev->mman.fw_vram_usage_reserved_bo = NULL; 1518 1519 if (adev->mman.fw_vram_usage_size == 0 || 1520 adev->mman.fw_vram_usage_size > vram_size) 1521 return 0; 1522 1523 return amdgpu_bo_create_kernel_at(adev, 1524 adev->mman.fw_vram_usage_start_offset, 1525 adev->mman.fw_vram_usage_size, 1526 AMDGPU_GEM_DOMAIN_VRAM, 1527 &adev->mman.fw_vram_usage_reserved_bo, 1528 &adev->mman.fw_vram_usage_va); 1529 } 1530 1531 /* 1532 * Memoy training reservation functions 1533 */ 1534 1535 /** 1536 * amdgpu_ttm_training_reserve_vram_fini - free memory training reserved vram 1537 * 1538 * @adev: amdgpu_device pointer 1539 * 1540 * free memory training reserved vram if it has been reserved. 1541 */ 1542 static int amdgpu_ttm_training_reserve_vram_fini(struct amdgpu_device *adev) 1543 { 1544 struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx; 1545 1546 ctx->init = PSP_MEM_TRAIN_NOT_SUPPORT; 1547 amdgpu_bo_free_kernel(&ctx->c2p_bo, NULL, NULL); 1548 ctx->c2p_bo = NULL; 1549 1550 return 0; 1551 } 1552 1553 static void amdgpu_ttm_training_data_block_init(struct amdgpu_device *adev) 1554 { 1555 struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx; 1556 1557 memset(ctx, 0, sizeof(*ctx)); 1558 1559 ctx->c2p_train_data_offset = 1560 ALIGN((adev->gmc.mc_vram_size - adev->mman.discovery_tmr_size - SZ_1M), SZ_1M); 1561 ctx->p2c_train_data_offset = 1562 (adev->gmc.mc_vram_size - GDDR6_MEM_TRAINING_OFFSET); 1563 ctx->train_data_size = 1564 GDDR6_MEM_TRAINING_DATA_SIZE_IN_BYTES; 1565 1566 DRM_DEBUG("train_data_size:%llx,p2c_train_data_offset:%llx,c2p_train_data_offset:%llx.\n", 1567 ctx->train_data_size, 1568 ctx->p2c_train_data_offset, 1569 ctx->c2p_train_data_offset); 1570 } 1571 1572 /* 1573 * reserve TMR memory at the top of VRAM which holds 1574 * IP Discovery data and is protected by PSP. 1575 */ 1576 static int amdgpu_ttm_reserve_tmr(struct amdgpu_device *adev) 1577 { 1578 int ret; 1579 struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx; 1580 bool mem_train_support = false; 1581 1582 if (!amdgpu_sriov_vf(adev)) { 1583 if (amdgpu_atomfirmware_mem_training_supported(adev)) 1584 mem_train_support = true; 1585 else 1586 DRM_DEBUG("memory training does not support!\n"); 1587 } 1588 1589 /* 1590 * Query reserved tmr size through atom firmwareinfo for Sienna_Cichlid and onwards for all 1591 * the use cases (IP discovery/G6 memory training/profiling/diagnostic data.etc) 1592 * 1593 * Otherwise, fallback to legacy approach to check and reserve tmr block for ip 1594 * discovery data and G6 memory training data respectively 1595 */ 1596 adev->mman.discovery_tmr_size = 1597 amdgpu_atomfirmware_get_fw_reserved_fb_size(adev); 1598 if (!adev->mman.discovery_tmr_size) 1599 adev->mman.discovery_tmr_size = DISCOVERY_TMR_OFFSET; 1600 1601 if (mem_train_support) { 1602 /* reserve vram for mem train according to TMR location */ 1603 amdgpu_ttm_training_data_block_init(adev); 1604 ret = amdgpu_bo_create_kernel_at(adev, 1605 ctx->c2p_train_data_offset, 1606 ctx->train_data_size, 1607 AMDGPU_GEM_DOMAIN_VRAM, 1608 &ctx->c2p_bo, 1609 NULL); 1610 if (ret) { 1611 DRM_ERROR("alloc c2p_bo failed(%d)!\n", ret); 1612 amdgpu_ttm_training_reserve_vram_fini(adev); 1613 return ret; 1614 } 1615 ctx->init = PSP_MEM_TRAIN_RESERVE_SUCCESS; 1616 } 1617 1618 ret = amdgpu_bo_create_kernel_at(adev, 1619 adev->gmc.real_vram_size - adev->mman.discovery_tmr_size, 1620 adev->mman.discovery_tmr_size, 1621 AMDGPU_GEM_DOMAIN_VRAM, 1622 &adev->mman.discovery_memory, 1623 NULL); 1624 if (ret) { 1625 DRM_ERROR("alloc tmr failed(%d)!\n", ret); 1626 amdgpu_bo_free_kernel(&adev->mman.discovery_memory, NULL, NULL); 1627 return ret; 1628 } 1629 1630 return 0; 1631 } 1632 1633 /* 1634 * amdgpu_ttm_init - Init the memory management (ttm) as well as various 1635 * gtt/vram related fields. 1636 * 1637 * This initializes all of the memory space pools that the TTM layer 1638 * will need such as the GTT space (system memory mapped to the device), 1639 * VRAM (on-board memory), and on-chip memories (GDS, GWS, OA) which 1640 * can be mapped per VMID. 1641 */ 1642 int amdgpu_ttm_init(struct amdgpu_device *adev) 1643 { 1644 uint64_t gtt_size; 1645 int r; 1646 u64 vis_vram_limit; 1647 1648 mutex_init(&adev->mman.gtt_window_lock); 1649 1650 /* No others user of address space so set it to 0 */ 1651 r = ttm_device_init(&adev->mman.bdev, &amdgpu_bo_driver, adev->dev, 1652 adev_to_drm(adev)->anon_inode->i_mapping, 1653 adev_to_drm(adev)->vma_offset_manager, 1654 adev->need_swiotlb, 1655 dma_addressing_limited(adev->dev)); 1656 if (r) { 1657 DRM_ERROR("failed initializing buffer object driver(%d).\n", r); 1658 return r; 1659 } 1660 adev->mman.initialized = true; 1661 1662 /* Initialize VRAM pool with all of VRAM divided into pages */ 1663 r = amdgpu_vram_mgr_init(adev); 1664 if (r) { 1665 DRM_ERROR("Failed initializing VRAM heap.\n"); 1666 return r; 1667 } 1668 1669 /* Reduce size of CPU-visible VRAM if requested */ 1670 vis_vram_limit = (u64)amdgpu_vis_vram_limit * 1024 * 1024; 1671 if (amdgpu_vis_vram_limit > 0 && 1672 vis_vram_limit <= adev->gmc.visible_vram_size) 1673 adev->gmc.visible_vram_size = vis_vram_limit; 1674 1675 /* Change the size here instead of the init above so only lpfn is affected */ 1676 amdgpu_ttm_set_buffer_funcs_status(adev, false); 1677 #ifdef CONFIG_64BIT 1678 #ifdef CONFIG_X86 1679 if (adev->gmc.xgmi.connected_to_cpu) 1680 adev->mman.aper_base_kaddr = ioremap_cache(adev->gmc.aper_base, 1681 adev->gmc.visible_vram_size); 1682 1683 else 1684 #endif 1685 adev->mman.aper_base_kaddr = ioremap_wc(adev->gmc.aper_base, 1686 adev->gmc.visible_vram_size); 1687 #endif 1688 1689 /* 1690 *The reserved vram for firmware must be pinned to the specified 1691 *place on the VRAM, so reserve it early. 1692 */ 1693 r = amdgpu_ttm_fw_reserve_vram_init(adev); 1694 if (r) { 1695 return r; 1696 } 1697 1698 /* 1699 * only NAVI10 and onwards ASIC support for IP discovery. 1700 * If IP discovery enabled, a block of memory should be 1701 * reserved for IP discovey. 1702 */ 1703 if (adev->mman.discovery_bin) { 1704 r = amdgpu_ttm_reserve_tmr(adev); 1705 if (r) 1706 return r; 1707 } 1708 1709 /* allocate memory as required for VGA 1710 * This is used for VGA emulation and pre-OS scanout buffers to 1711 * avoid display artifacts while transitioning between pre-OS 1712 * and driver. */ 1713 r = amdgpu_bo_create_kernel_at(adev, 0, adev->mman.stolen_vga_size, 1714 AMDGPU_GEM_DOMAIN_VRAM, 1715 &adev->mman.stolen_vga_memory, 1716 NULL); 1717 if (r) 1718 return r; 1719 r = amdgpu_bo_create_kernel_at(adev, adev->mman.stolen_vga_size, 1720 adev->mman.stolen_extended_size, 1721 AMDGPU_GEM_DOMAIN_VRAM, 1722 &adev->mman.stolen_extended_memory, 1723 NULL); 1724 if (r) 1725 return r; 1726 r = amdgpu_bo_create_kernel_at(adev, adev->mman.stolen_reserved_offset, 1727 adev->mman.stolen_reserved_size, 1728 AMDGPU_GEM_DOMAIN_VRAM, 1729 &adev->mman.stolen_reserved_memory, 1730 NULL); 1731 if (r) 1732 return r; 1733 1734 DRM_INFO("amdgpu: %uM of VRAM memory ready\n", 1735 (unsigned) (adev->gmc.real_vram_size / (1024 * 1024))); 1736 1737 /* Compute GTT size, either bsaed on 3/4th the size of RAM size 1738 * or whatever the user passed on module init */ 1739 if (amdgpu_gtt_size == -1) { 1740 struct sysinfo si; 1741 1742 si_meminfo(&si); 1743 gtt_size = min(max((AMDGPU_DEFAULT_GTT_SIZE_MB << 20), 1744 adev->gmc.mc_vram_size), 1745 ((uint64_t)si.totalram * si.mem_unit * 3/4)); 1746 } 1747 else 1748 gtt_size = (uint64_t)amdgpu_gtt_size << 20; 1749 1750 /* Initialize GTT memory pool */ 1751 r = amdgpu_gtt_mgr_init(adev, gtt_size); 1752 if (r) { 1753 DRM_ERROR("Failed initializing GTT heap.\n"); 1754 return r; 1755 } 1756 DRM_INFO("amdgpu: %uM of GTT memory ready.\n", 1757 (unsigned)(gtt_size / (1024 * 1024))); 1758 1759 /* Initialize preemptible memory pool */ 1760 r = amdgpu_preempt_mgr_init(adev); 1761 if (r) { 1762 DRM_ERROR("Failed initializing PREEMPT heap.\n"); 1763 return r; 1764 } 1765 1766 /* Initialize various on-chip memory pools */ 1767 r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_GDS, adev->gds.gds_size); 1768 if (r) { 1769 DRM_ERROR("Failed initializing GDS heap.\n"); 1770 return r; 1771 } 1772 1773 r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_GWS, adev->gds.gws_size); 1774 if (r) { 1775 DRM_ERROR("Failed initializing gws heap.\n"); 1776 return r; 1777 } 1778 1779 r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_OA, adev->gds.oa_size); 1780 if (r) { 1781 DRM_ERROR("Failed initializing oa heap.\n"); 1782 return r; 1783 } 1784 1785 return 0; 1786 } 1787 1788 /* 1789 * amdgpu_ttm_fini - De-initialize the TTM memory pools 1790 */ 1791 void amdgpu_ttm_fini(struct amdgpu_device *adev) 1792 { 1793 if (!adev->mman.initialized) 1794 return; 1795 1796 amdgpu_ttm_training_reserve_vram_fini(adev); 1797 /* return the stolen vga memory back to VRAM */ 1798 amdgpu_bo_free_kernel(&adev->mman.stolen_vga_memory, NULL, NULL); 1799 amdgpu_bo_free_kernel(&adev->mman.stolen_extended_memory, NULL, NULL); 1800 /* return the IP Discovery TMR memory back to VRAM */ 1801 amdgpu_bo_free_kernel(&adev->mman.discovery_memory, NULL, NULL); 1802 if (adev->mman.stolen_reserved_size) 1803 amdgpu_bo_free_kernel(&adev->mman.stolen_reserved_memory, 1804 NULL, NULL); 1805 amdgpu_ttm_fw_reserve_vram_fini(adev); 1806 1807 amdgpu_vram_mgr_fini(adev); 1808 amdgpu_gtt_mgr_fini(adev); 1809 amdgpu_preempt_mgr_fini(adev); 1810 ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_GDS); 1811 ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_GWS); 1812 ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_OA); 1813 ttm_device_fini(&adev->mman.bdev); 1814 adev->mman.initialized = false; 1815 DRM_INFO("amdgpu: ttm finalized\n"); 1816 } 1817 1818 /** 1819 * amdgpu_ttm_set_buffer_funcs_status - enable/disable use of buffer functions 1820 * 1821 * @adev: amdgpu_device pointer 1822 * @enable: true when we can use buffer functions. 1823 * 1824 * Enable/disable use of buffer functions during suspend/resume. This should 1825 * only be called at bootup or when userspace isn't running. 1826 */ 1827 void amdgpu_ttm_set_buffer_funcs_status(struct amdgpu_device *adev, bool enable) 1828 { 1829 struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM); 1830 uint64_t size; 1831 int r; 1832 1833 if (!adev->mman.initialized || amdgpu_in_reset(adev) || 1834 adev->mman.buffer_funcs_enabled == enable) 1835 return; 1836 1837 if (enable) { 1838 struct amdgpu_ring *ring; 1839 struct drm_gpu_scheduler *sched; 1840 1841 ring = adev->mman.buffer_funcs_ring; 1842 sched = &ring->sched; 1843 r = drm_sched_entity_init(&adev->mman.entity, 1844 DRM_SCHED_PRIORITY_KERNEL, &sched, 1845 1, NULL); 1846 if (r) { 1847 DRM_ERROR("Failed setting up TTM BO move entity (%d)\n", 1848 r); 1849 return; 1850 } 1851 } else { 1852 drm_sched_entity_destroy(&adev->mman.entity); 1853 dma_fence_put(man->move); 1854 man->move = NULL; 1855 } 1856 1857 /* this just adjusts TTM size idea, which sets lpfn to the correct value */ 1858 if (enable) 1859 size = adev->gmc.real_vram_size; 1860 else 1861 size = adev->gmc.visible_vram_size; 1862 man->size = size >> PAGE_SHIFT; 1863 adev->mman.buffer_funcs_enabled = enable; 1864 } 1865 1866 int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset, 1867 uint64_t dst_offset, uint32_t byte_count, 1868 struct dma_resv *resv, 1869 struct dma_fence **fence, bool direct_submit, 1870 bool vm_needs_flush, bool tmz) 1871 { 1872 enum amdgpu_ib_pool_type pool = direct_submit ? AMDGPU_IB_POOL_DIRECT : 1873 AMDGPU_IB_POOL_DELAYED; 1874 struct amdgpu_device *adev = ring->adev; 1875 struct amdgpu_job *job; 1876 1877 uint32_t max_bytes; 1878 unsigned num_loops, num_dw; 1879 unsigned i; 1880 int r; 1881 1882 if (direct_submit && !ring->sched.ready) { 1883 DRM_ERROR("Trying to move memory with ring turned off.\n"); 1884 return -EINVAL; 1885 } 1886 1887 max_bytes = adev->mman.buffer_funcs->copy_max_bytes; 1888 num_loops = DIV_ROUND_UP(byte_count, max_bytes); 1889 num_dw = ALIGN(num_loops * adev->mman.buffer_funcs->copy_num_dw, 8); 1890 1891 r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, pool, &job); 1892 if (r) 1893 return r; 1894 1895 if (vm_needs_flush) { 1896 job->vm_pd_addr = amdgpu_gmc_pd_addr(adev->gmc.pdb0_bo ? 1897 adev->gmc.pdb0_bo : adev->gart.bo); 1898 job->vm_needs_flush = true; 1899 } 1900 if (resv) { 1901 r = amdgpu_sync_resv(adev, &job->sync, resv, 1902 AMDGPU_SYNC_ALWAYS, 1903 AMDGPU_FENCE_OWNER_UNDEFINED); 1904 if (r) { 1905 DRM_ERROR("sync failed (%d).\n", r); 1906 goto error_free; 1907 } 1908 } 1909 1910 for (i = 0; i < num_loops; i++) { 1911 uint32_t cur_size_in_bytes = min(byte_count, max_bytes); 1912 1913 amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_offset, 1914 dst_offset, cur_size_in_bytes, tmz); 1915 1916 src_offset += cur_size_in_bytes; 1917 dst_offset += cur_size_in_bytes; 1918 byte_count -= cur_size_in_bytes; 1919 } 1920 1921 amdgpu_ring_pad_ib(ring, &job->ibs[0]); 1922 WARN_ON(job->ibs[0].length_dw > num_dw); 1923 if (direct_submit) 1924 r = amdgpu_job_submit_direct(job, ring, fence); 1925 else 1926 r = amdgpu_job_submit(job, &adev->mman.entity, 1927 AMDGPU_FENCE_OWNER_UNDEFINED, fence); 1928 if (r) 1929 goto error_free; 1930 1931 return r; 1932 1933 error_free: 1934 amdgpu_job_free(job); 1935 DRM_ERROR("Error scheduling IBs (%d)\n", r); 1936 return r; 1937 } 1938 1939 int amdgpu_fill_buffer(struct amdgpu_bo *bo, 1940 uint32_t src_data, 1941 struct dma_resv *resv, 1942 struct dma_fence **fence) 1943 { 1944 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); 1945 uint32_t max_bytes = adev->mman.buffer_funcs->fill_max_bytes; 1946 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring; 1947 1948 struct amdgpu_res_cursor cursor; 1949 unsigned int num_loops, num_dw; 1950 uint64_t num_bytes; 1951 1952 struct amdgpu_job *job; 1953 int r; 1954 1955 if (!adev->mman.buffer_funcs_enabled) { 1956 DRM_ERROR("Trying to clear memory with ring turned off.\n"); 1957 return -EINVAL; 1958 } 1959 1960 if (bo->tbo.resource->mem_type == AMDGPU_PL_PREEMPT) { 1961 DRM_ERROR("Trying to clear preemptible memory.\n"); 1962 return -EINVAL; 1963 } 1964 1965 if (bo->tbo.resource->mem_type == TTM_PL_TT) { 1966 r = amdgpu_ttm_alloc_gart(&bo->tbo); 1967 if (r) 1968 return r; 1969 } 1970 1971 num_bytes = bo->tbo.resource->num_pages << PAGE_SHIFT; 1972 num_loops = 0; 1973 1974 amdgpu_res_first(bo->tbo.resource, 0, num_bytes, &cursor); 1975 while (cursor.remaining) { 1976 num_loops += DIV_ROUND_UP_ULL(cursor.size, max_bytes); 1977 amdgpu_res_next(&cursor, cursor.size); 1978 } 1979 num_dw = num_loops * adev->mman.buffer_funcs->fill_num_dw; 1980 1981 /* for IB padding */ 1982 num_dw += 64; 1983 1984 r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, AMDGPU_IB_POOL_DELAYED, 1985 &job); 1986 if (r) 1987 return r; 1988 1989 if (resv) { 1990 r = amdgpu_sync_resv(adev, &job->sync, resv, 1991 AMDGPU_SYNC_ALWAYS, 1992 AMDGPU_FENCE_OWNER_UNDEFINED); 1993 if (r) { 1994 DRM_ERROR("sync failed (%d).\n", r); 1995 goto error_free; 1996 } 1997 } 1998 1999 amdgpu_res_first(bo->tbo.resource, 0, num_bytes, &cursor); 2000 while (cursor.remaining) { 2001 uint32_t cur_size = min_t(uint64_t, cursor.size, max_bytes); 2002 uint64_t dst_addr = cursor.start; 2003 2004 dst_addr += amdgpu_ttm_domain_start(adev, 2005 bo->tbo.resource->mem_type); 2006 amdgpu_emit_fill_buffer(adev, &job->ibs[0], src_data, dst_addr, 2007 cur_size); 2008 2009 amdgpu_res_next(&cursor, cur_size); 2010 } 2011 2012 amdgpu_ring_pad_ib(ring, &job->ibs[0]); 2013 WARN_ON(job->ibs[0].length_dw > num_dw); 2014 r = amdgpu_job_submit(job, &adev->mman.entity, 2015 AMDGPU_FENCE_OWNER_UNDEFINED, fence); 2016 if (r) 2017 goto error_free; 2018 2019 return 0; 2020 2021 error_free: 2022 amdgpu_job_free(job); 2023 return r; 2024 } 2025 2026 #if defined(CONFIG_DEBUG_FS) 2027 2028 static int amdgpu_mm_vram_table_show(struct seq_file *m, void *unused) 2029 { 2030 struct amdgpu_device *adev = (struct amdgpu_device *)m->private; 2031 struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev, 2032 TTM_PL_VRAM); 2033 struct drm_printer p = drm_seq_file_printer(m); 2034 2035 man->func->debug(man, &p); 2036 return 0; 2037 } 2038 2039 static int amdgpu_ttm_page_pool_show(struct seq_file *m, void *unused) 2040 { 2041 struct amdgpu_device *adev = (struct amdgpu_device *)m->private; 2042 2043 return ttm_pool_debugfs(&adev->mman.bdev.pool, m); 2044 } 2045 2046 static int amdgpu_mm_tt_table_show(struct seq_file *m, void *unused) 2047 { 2048 struct amdgpu_device *adev = (struct amdgpu_device *)m->private; 2049 struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev, 2050 TTM_PL_TT); 2051 struct drm_printer p = drm_seq_file_printer(m); 2052 2053 man->func->debug(man, &p); 2054 return 0; 2055 } 2056 2057 static int amdgpu_mm_gds_table_show(struct seq_file *m, void *unused) 2058 { 2059 struct amdgpu_device *adev = (struct amdgpu_device *)m->private; 2060 struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev, 2061 AMDGPU_PL_GDS); 2062 struct drm_printer p = drm_seq_file_printer(m); 2063 2064 man->func->debug(man, &p); 2065 return 0; 2066 } 2067 2068 static int amdgpu_mm_gws_table_show(struct seq_file *m, void *unused) 2069 { 2070 struct amdgpu_device *adev = (struct amdgpu_device *)m->private; 2071 struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev, 2072 AMDGPU_PL_GWS); 2073 struct drm_printer p = drm_seq_file_printer(m); 2074 2075 man->func->debug(man, &p); 2076 return 0; 2077 } 2078 2079 static int amdgpu_mm_oa_table_show(struct seq_file *m, void *unused) 2080 { 2081 struct amdgpu_device *adev = (struct amdgpu_device *)m->private; 2082 struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev, 2083 AMDGPU_PL_OA); 2084 struct drm_printer p = drm_seq_file_printer(m); 2085 2086 man->func->debug(man, &p); 2087 return 0; 2088 } 2089 2090 DEFINE_SHOW_ATTRIBUTE(amdgpu_mm_vram_table); 2091 DEFINE_SHOW_ATTRIBUTE(amdgpu_mm_tt_table); 2092 DEFINE_SHOW_ATTRIBUTE(amdgpu_mm_gds_table); 2093 DEFINE_SHOW_ATTRIBUTE(amdgpu_mm_gws_table); 2094 DEFINE_SHOW_ATTRIBUTE(amdgpu_mm_oa_table); 2095 DEFINE_SHOW_ATTRIBUTE(amdgpu_ttm_page_pool); 2096 2097 /* 2098 * amdgpu_ttm_vram_read - Linear read access to VRAM 2099 * 2100 * Accesses VRAM via MMIO for debugging purposes. 2101 */ 2102 static ssize_t amdgpu_ttm_vram_read(struct file *f, char __user *buf, 2103 size_t size, loff_t *pos) 2104 { 2105 struct amdgpu_device *adev = file_inode(f)->i_private; 2106 ssize_t result = 0; 2107 2108 if (size & 0x3 || *pos & 0x3) 2109 return -EINVAL; 2110 2111 if (*pos >= adev->gmc.mc_vram_size) 2112 return -ENXIO; 2113 2114 size = min(size, (size_t)(adev->gmc.mc_vram_size - *pos)); 2115 while (size) { 2116 size_t bytes = min(size, AMDGPU_TTM_VRAM_MAX_DW_READ * 4); 2117 uint32_t value[AMDGPU_TTM_VRAM_MAX_DW_READ]; 2118 2119 amdgpu_device_vram_access(adev, *pos, value, bytes, false); 2120 if (copy_to_user(buf, value, bytes)) 2121 return -EFAULT; 2122 2123 result += bytes; 2124 buf += bytes; 2125 *pos += bytes; 2126 size -= bytes; 2127 } 2128 2129 return result; 2130 } 2131 2132 /* 2133 * amdgpu_ttm_vram_write - Linear write access to VRAM 2134 * 2135 * Accesses VRAM via MMIO for debugging purposes. 2136 */ 2137 static ssize_t amdgpu_ttm_vram_write(struct file *f, const char __user *buf, 2138 size_t size, loff_t *pos) 2139 { 2140 struct amdgpu_device *adev = file_inode(f)->i_private; 2141 ssize_t result = 0; 2142 int r; 2143 2144 if (size & 0x3 || *pos & 0x3) 2145 return -EINVAL; 2146 2147 if (*pos >= adev->gmc.mc_vram_size) 2148 return -ENXIO; 2149 2150 while (size) { 2151 unsigned long flags; 2152 uint32_t value; 2153 2154 if (*pos >= adev->gmc.mc_vram_size) 2155 return result; 2156 2157 r = get_user(value, (uint32_t *)buf); 2158 if (r) 2159 return r; 2160 2161 spin_lock_irqsave(&adev->mmio_idx_lock, flags); 2162 WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)*pos) | 0x80000000); 2163 WREG32_NO_KIQ(mmMM_INDEX_HI, *pos >> 31); 2164 WREG32_NO_KIQ(mmMM_DATA, value); 2165 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags); 2166 2167 result += 4; 2168 buf += 4; 2169 *pos += 4; 2170 size -= 4; 2171 } 2172 2173 return result; 2174 } 2175 2176 static const struct file_operations amdgpu_ttm_vram_fops = { 2177 .owner = THIS_MODULE, 2178 .read = amdgpu_ttm_vram_read, 2179 .write = amdgpu_ttm_vram_write, 2180 .llseek = default_llseek, 2181 }; 2182 2183 /* 2184 * amdgpu_iomem_read - Virtual read access to GPU mapped memory 2185 * 2186 * This function is used to read memory that has been mapped to the 2187 * GPU and the known addresses are not physical addresses but instead 2188 * bus addresses (e.g., what you'd put in an IB or ring buffer). 2189 */ 2190 static ssize_t amdgpu_iomem_read(struct file *f, char __user *buf, 2191 size_t size, loff_t *pos) 2192 { 2193 struct amdgpu_device *adev = file_inode(f)->i_private; 2194 struct iommu_domain *dom; 2195 ssize_t result = 0; 2196 int r; 2197 2198 /* retrieve the IOMMU domain if any for this device */ 2199 dom = iommu_get_domain_for_dev(adev->dev); 2200 2201 while (size) { 2202 phys_addr_t addr = *pos & PAGE_MASK; 2203 loff_t off = *pos & ~PAGE_MASK; 2204 size_t bytes = PAGE_SIZE - off; 2205 unsigned long pfn; 2206 struct page *p; 2207 void *ptr; 2208 2209 bytes = bytes < size ? bytes : size; 2210 2211 /* Translate the bus address to a physical address. If 2212 * the domain is NULL it means there is no IOMMU active 2213 * and the address translation is the identity 2214 */ 2215 addr = dom ? iommu_iova_to_phys(dom, addr) : addr; 2216 2217 pfn = addr >> PAGE_SHIFT; 2218 if (!pfn_valid(pfn)) 2219 return -EPERM; 2220 2221 p = pfn_to_page(pfn); 2222 if (p->mapping != adev->mman.bdev.dev_mapping) 2223 return -EPERM; 2224 2225 ptr = kmap(p); 2226 r = copy_to_user(buf, ptr + off, bytes); 2227 kunmap(p); 2228 if (r) 2229 return -EFAULT; 2230 2231 size -= bytes; 2232 *pos += bytes; 2233 result += bytes; 2234 } 2235 2236 return result; 2237 } 2238 2239 /* 2240 * amdgpu_iomem_write - Virtual write access to GPU mapped memory 2241 * 2242 * This function is used to write memory that has been mapped to the 2243 * GPU and the known addresses are not physical addresses but instead 2244 * bus addresses (e.g., what you'd put in an IB or ring buffer). 2245 */ 2246 static ssize_t amdgpu_iomem_write(struct file *f, const char __user *buf, 2247 size_t size, loff_t *pos) 2248 { 2249 struct amdgpu_device *adev = file_inode(f)->i_private; 2250 struct iommu_domain *dom; 2251 ssize_t result = 0; 2252 int r; 2253 2254 dom = iommu_get_domain_for_dev(adev->dev); 2255 2256 while (size) { 2257 phys_addr_t addr = *pos & PAGE_MASK; 2258 loff_t off = *pos & ~PAGE_MASK; 2259 size_t bytes = PAGE_SIZE - off; 2260 unsigned long pfn; 2261 struct page *p; 2262 void *ptr; 2263 2264 bytes = bytes < size ? bytes : size; 2265 2266 addr = dom ? iommu_iova_to_phys(dom, addr) : addr; 2267 2268 pfn = addr >> PAGE_SHIFT; 2269 if (!pfn_valid(pfn)) 2270 return -EPERM; 2271 2272 p = pfn_to_page(pfn); 2273 if (p->mapping != adev->mman.bdev.dev_mapping) 2274 return -EPERM; 2275 2276 ptr = kmap(p); 2277 r = copy_from_user(ptr + off, buf, bytes); 2278 kunmap(p); 2279 if (r) 2280 return -EFAULT; 2281 2282 size -= bytes; 2283 *pos += bytes; 2284 result += bytes; 2285 } 2286 2287 return result; 2288 } 2289 2290 static const struct file_operations amdgpu_ttm_iomem_fops = { 2291 .owner = THIS_MODULE, 2292 .read = amdgpu_iomem_read, 2293 .write = amdgpu_iomem_write, 2294 .llseek = default_llseek 2295 }; 2296 2297 #endif 2298 2299 void amdgpu_ttm_debugfs_init(struct amdgpu_device *adev) 2300 { 2301 #if defined(CONFIG_DEBUG_FS) 2302 struct drm_minor *minor = adev_to_drm(adev)->primary; 2303 struct dentry *root = minor->debugfs_root; 2304 2305 debugfs_create_file_size("amdgpu_vram", 0444, root, adev, 2306 &amdgpu_ttm_vram_fops, adev->gmc.mc_vram_size); 2307 debugfs_create_file("amdgpu_iomem", 0444, root, adev, 2308 &amdgpu_ttm_iomem_fops); 2309 debugfs_create_file("amdgpu_vram_mm", 0444, root, adev, 2310 &amdgpu_mm_vram_table_fops); 2311 debugfs_create_file("amdgpu_gtt_mm", 0444, root, adev, 2312 &amdgpu_mm_tt_table_fops); 2313 debugfs_create_file("amdgpu_gds_mm", 0444, root, adev, 2314 &amdgpu_mm_gds_table_fops); 2315 debugfs_create_file("amdgpu_gws_mm", 0444, root, adev, 2316 &amdgpu_mm_gws_table_fops); 2317 debugfs_create_file("amdgpu_oa_mm", 0444, root, adev, 2318 &amdgpu_mm_oa_table_fops); 2319 debugfs_create_file("ttm_page_pool", 0444, root, adev, 2320 &amdgpu_ttm_page_pool_fops); 2321 #endif 2322 } 2323