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