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 #include <linux/module.h> 45 46 #include <drm/drm_drv.h> 47 #include <drm/ttm/ttm_bo_api.h> 48 #include <drm/ttm/ttm_bo_driver.h> 49 #include <drm/ttm/ttm_placement.h> 50 #include <drm/ttm/ttm_range_manager.h> 51 52 #include <drm/amdgpu_drm.h> 53 54 #include "amdgpu.h" 55 #include "amdgpu_object.h" 56 #include "amdgpu_trace.h" 57 #include "amdgpu_amdkfd.h" 58 #include "amdgpu_sdma.h" 59 #include "amdgpu_ras.h" 60 #include "amdgpu_atomfirmware.h" 61 #include "amdgpu_res_cursor.h" 62 #include "bif/bif_4_1_d.h" 63 64 MODULE_IMPORT_NS(DMA_BUF); 65 66 #define AMDGPU_TTM_VRAM_MAX_DW_READ (size_t)128 67 68 static int amdgpu_ttm_backend_bind(struct ttm_device *bdev, 69 struct ttm_tt *ttm, 70 struct ttm_resource *bo_mem); 71 static void amdgpu_ttm_backend_unbind(struct ttm_device *bdev, 72 struct ttm_tt *ttm); 73 74 static int amdgpu_ttm_init_on_chip(struct amdgpu_device *adev, 75 unsigned int type, 76 uint64_t size_in_page) 77 { 78 return ttm_range_man_init(&adev->mman.bdev, type, 79 false, size_in_page); 80 } 81 82 /** 83 * amdgpu_evict_flags - Compute placement flags 84 * 85 * @bo: The buffer object to evict 86 * @placement: Possible destination(s) for evicted BO 87 * 88 * Fill in placement data when ttm_bo_evict() is called 89 */ 90 static void amdgpu_evict_flags(struct ttm_buffer_object *bo, 91 struct ttm_placement *placement) 92 { 93 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev); 94 struct amdgpu_bo *abo; 95 static const struct ttm_place placements = { 96 .fpfn = 0, 97 .lpfn = 0, 98 .mem_type = TTM_PL_SYSTEM, 99 .flags = 0 100 }; 101 102 /* Don't handle scatter gather BOs */ 103 if (bo->type == ttm_bo_type_sg) { 104 placement->num_placement = 0; 105 placement->num_busy_placement = 0; 106 return; 107 } 108 109 /* Object isn't an AMDGPU object so ignore */ 110 if (!amdgpu_bo_is_amdgpu_bo(bo)) { 111 placement->placement = &placements; 112 placement->busy_placement = &placements; 113 placement->num_placement = 1; 114 placement->num_busy_placement = 1; 115 return; 116 } 117 118 abo = ttm_to_amdgpu_bo(bo); 119 if (abo->flags & AMDGPU_AMDKFD_CREATE_SVM_BO) { 120 placement->num_placement = 0; 121 placement->num_busy_placement = 0; 122 return; 123 } 124 125 switch (bo->resource->mem_type) { 126 case AMDGPU_PL_GDS: 127 case AMDGPU_PL_GWS: 128 case AMDGPU_PL_OA: 129 placement->num_placement = 0; 130 placement->num_busy_placement = 0; 131 return; 132 133 case TTM_PL_VRAM: 134 if (!adev->mman.buffer_funcs_enabled) { 135 /* Move to system memory */ 136 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU); 137 } else if (!amdgpu_gmc_vram_full_visible(&adev->gmc) && 138 !(abo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED) && 139 amdgpu_bo_in_cpu_visible_vram(abo)) { 140 141 /* Try evicting to the CPU inaccessible part of VRAM 142 * first, but only set GTT as busy placement, so this 143 * BO will be evicted to GTT rather than causing other 144 * BOs to be evicted from VRAM 145 */ 146 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM | 147 AMDGPU_GEM_DOMAIN_GTT | 148 AMDGPU_GEM_DOMAIN_CPU); 149 abo->placements[0].fpfn = adev->gmc.visible_vram_size >> PAGE_SHIFT; 150 abo->placements[0].lpfn = 0; 151 abo->placement.busy_placement = &abo->placements[1]; 152 abo->placement.num_busy_placement = 1; 153 } else { 154 /* Move to GTT memory */ 155 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_GTT | 156 AMDGPU_GEM_DOMAIN_CPU); 157 } 158 break; 159 case TTM_PL_TT: 160 case AMDGPU_PL_PREEMPT: 161 default: 162 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU); 163 break; 164 } 165 *placement = abo->placement; 166 } 167 168 /** 169 * amdgpu_ttm_map_buffer - Map memory into the GART windows 170 * @bo: buffer object to map 171 * @mem: memory object to map 172 * @mm_cur: range to map 173 * @num_pages: number of pages to map 174 * @window: which GART window to use 175 * @ring: DMA ring to use for the copy 176 * @tmz: if we should setup a TMZ enabled mapping 177 * @addr: resulting address inside the MC address space 178 * 179 * Setup one of the GART windows to access a specific piece of memory or return 180 * the physical address for local memory. 181 */ 182 static int amdgpu_ttm_map_buffer(struct ttm_buffer_object *bo, 183 struct ttm_resource *mem, 184 struct amdgpu_res_cursor *mm_cur, 185 unsigned num_pages, unsigned window, 186 struct amdgpu_ring *ring, bool tmz, 187 uint64_t *addr) 188 { 189 struct amdgpu_device *adev = ring->adev; 190 struct amdgpu_job *job; 191 unsigned num_dw, num_bytes; 192 struct dma_fence *fence; 193 uint64_t src_addr, dst_addr; 194 void *cpu_addr; 195 uint64_t flags; 196 unsigned int i; 197 int r; 198 199 BUG_ON(adev->mman.buffer_funcs->copy_max_bytes < 200 AMDGPU_GTT_MAX_TRANSFER_SIZE * 8); 201 BUG_ON(mem->mem_type == AMDGPU_PL_PREEMPT); 202 203 /* Map only what can't be accessed directly */ 204 if (!tmz && mem->start != AMDGPU_BO_INVALID_OFFSET) { 205 *addr = amdgpu_ttm_domain_start(adev, mem->mem_type) + 206 mm_cur->start; 207 return 0; 208 } 209 210 *addr = adev->gmc.gart_start; 211 *addr += (u64)window * AMDGPU_GTT_MAX_TRANSFER_SIZE * 212 AMDGPU_GPU_PAGE_SIZE; 213 *addr += mm_cur->start & ~PAGE_MASK; 214 215 num_dw = ALIGN(adev->mman.buffer_funcs->copy_num_dw, 8); 216 num_bytes = num_pages * 8 * AMDGPU_GPU_PAGES_IN_CPU_PAGE; 217 218 r = amdgpu_job_alloc_with_ib(adev, num_dw * 4 + num_bytes, 219 AMDGPU_IB_POOL_DELAYED, &job); 220 if (r) 221 return r; 222 223 src_addr = num_dw * 4; 224 src_addr += job->ibs[0].gpu_addr; 225 226 dst_addr = amdgpu_bo_gpu_offset(adev->gart.bo); 227 dst_addr += window * AMDGPU_GTT_MAX_TRANSFER_SIZE * 8; 228 amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_addr, 229 dst_addr, num_bytes, false); 230 231 amdgpu_ring_pad_ib(ring, &job->ibs[0]); 232 WARN_ON(job->ibs[0].length_dw > num_dw); 233 234 flags = amdgpu_ttm_tt_pte_flags(adev, bo->ttm, mem); 235 if (tmz) 236 flags |= AMDGPU_PTE_TMZ; 237 238 cpu_addr = &job->ibs[0].ptr[num_dw]; 239 240 if (mem->mem_type == TTM_PL_TT) { 241 dma_addr_t *dma_addr; 242 243 dma_addr = &bo->ttm->dma_address[mm_cur->start >> PAGE_SHIFT]; 244 r = amdgpu_gart_map(adev, 0, num_pages, dma_addr, flags, 245 cpu_addr); 246 if (r) 247 goto error_free; 248 } else { 249 dma_addr_t dma_address; 250 251 dma_address = mm_cur->start; 252 dma_address += adev->vm_manager.vram_base_offset; 253 254 for (i = 0; i < num_pages; ++i) { 255 r = amdgpu_gart_map(adev, i << PAGE_SHIFT, 1, 256 &dma_address, flags, cpu_addr); 257 if (r) 258 goto error_free; 259 260 dma_address += PAGE_SIZE; 261 } 262 } 263 264 r = amdgpu_job_submit(job, &adev->mman.entity, 265 AMDGPU_FENCE_OWNER_UNDEFINED, &fence); 266 if (r) 267 goto error_free; 268 269 dma_fence_put(fence); 270 271 return r; 272 273 error_free: 274 amdgpu_job_free(job); 275 return r; 276 } 277 278 /** 279 * amdgpu_ttm_copy_mem_to_mem - Helper function for copy 280 * @adev: amdgpu device 281 * @src: buffer/address where to read from 282 * @dst: buffer/address where to write to 283 * @size: number of bytes to copy 284 * @tmz: if a secure copy should be used 285 * @resv: resv object to sync to 286 * @f: Returns the last fence if multiple jobs are submitted. 287 * 288 * The function copies @size bytes from {src->mem + src->offset} to 289 * {dst->mem + dst->offset}. src->bo and dst->bo could be same BO for a 290 * move and different for a BO to BO copy. 291 * 292 */ 293 int amdgpu_ttm_copy_mem_to_mem(struct amdgpu_device *adev, 294 const struct amdgpu_copy_mem *src, 295 const struct amdgpu_copy_mem *dst, 296 uint64_t size, bool tmz, 297 struct dma_resv *resv, 298 struct dma_fence **f) 299 { 300 const uint32_t GTT_MAX_BYTES = (AMDGPU_GTT_MAX_TRANSFER_SIZE * 301 AMDGPU_GPU_PAGE_SIZE); 302 303 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring; 304 struct amdgpu_res_cursor src_mm, dst_mm; 305 struct dma_fence *fence = NULL; 306 int r = 0; 307 308 if (!adev->mman.buffer_funcs_enabled) { 309 DRM_ERROR("Trying to move memory with ring turned off.\n"); 310 return -EINVAL; 311 } 312 313 amdgpu_res_first(src->mem, src->offset, size, &src_mm); 314 amdgpu_res_first(dst->mem, dst->offset, size, &dst_mm); 315 316 mutex_lock(&adev->mman.gtt_window_lock); 317 while (src_mm.remaining) { 318 uint32_t src_page_offset = src_mm.start & ~PAGE_MASK; 319 uint32_t dst_page_offset = dst_mm.start & ~PAGE_MASK; 320 struct dma_fence *next; 321 uint32_t cur_size; 322 uint64_t from, to; 323 324 /* Copy size cannot exceed GTT_MAX_BYTES. So if src or dst 325 * begins at an offset, then adjust the size accordingly 326 */ 327 cur_size = max(src_page_offset, dst_page_offset); 328 cur_size = min(min3(src_mm.size, dst_mm.size, size), 329 (uint64_t)(GTT_MAX_BYTES - cur_size)); 330 331 /* Map src to window 0 and dst to window 1. */ 332 r = amdgpu_ttm_map_buffer(src->bo, src->mem, &src_mm, 333 PFN_UP(cur_size + src_page_offset), 334 0, ring, tmz, &from); 335 if (r) 336 goto error; 337 338 r = amdgpu_ttm_map_buffer(dst->bo, dst->mem, &dst_mm, 339 PFN_UP(cur_size + dst_page_offset), 340 1, ring, tmz, &to); 341 if (r) 342 goto error; 343 344 r = amdgpu_copy_buffer(ring, from, to, cur_size, 345 resv, &next, false, true, tmz); 346 if (r) 347 goto error; 348 349 dma_fence_put(fence); 350 fence = next; 351 352 amdgpu_res_next(&src_mm, cur_size); 353 amdgpu_res_next(&dst_mm, cur_size); 354 } 355 error: 356 mutex_unlock(&adev->mman.gtt_window_lock); 357 if (f) 358 *f = dma_fence_get(fence); 359 dma_fence_put(fence); 360 return r; 361 } 362 363 /* 364 * amdgpu_move_blit - Copy an entire buffer to another buffer 365 * 366 * This is a helper called by amdgpu_bo_move() and amdgpu_move_vram_ram() to 367 * help move buffers to and from VRAM. 368 */ 369 static int amdgpu_move_blit(struct ttm_buffer_object *bo, 370 bool evict, 371 struct ttm_resource *new_mem, 372 struct ttm_resource *old_mem) 373 { 374 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev); 375 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo); 376 struct amdgpu_copy_mem src, dst; 377 struct dma_fence *fence = NULL; 378 int r; 379 380 src.bo = bo; 381 dst.bo = bo; 382 src.mem = old_mem; 383 dst.mem = new_mem; 384 src.offset = 0; 385 dst.offset = 0; 386 387 r = amdgpu_ttm_copy_mem_to_mem(adev, &src, &dst, 388 new_mem->num_pages << PAGE_SHIFT, 389 amdgpu_bo_encrypted(abo), 390 bo->base.resv, &fence); 391 if (r) 392 goto error; 393 394 /* clear the space being freed */ 395 if (old_mem->mem_type == TTM_PL_VRAM && 396 (abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE)) { 397 struct dma_fence *wipe_fence = NULL; 398 399 r = amdgpu_fill_buffer(ttm_to_amdgpu_bo(bo), AMDGPU_POISON, 400 NULL, &wipe_fence); 401 if (r) { 402 goto error; 403 } else if (wipe_fence) { 404 dma_fence_put(fence); 405 fence = wipe_fence; 406 } 407 } 408 409 /* Always block for VM page tables before committing the new location */ 410 if (bo->type == ttm_bo_type_kernel) 411 r = ttm_bo_move_accel_cleanup(bo, fence, true, false, new_mem); 412 else 413 r = ttm_bo_move_accel_cleanup(bo, fence, evict, true, new_mem); 414 dma_fence_put(fence); 415 return r; 416 417 error: 418 if (fence) 419 dma_fence_wait(fence, false); 420 dma_fence_put(fence); 421 return r; 422 } 423 424 /* 425 * amdgpu_mem_visible - Check that memory can be accessed by ttm_bo_move_memcpy 426 * 427 * Called by amdgpu_bo_move() 428 */ 429 static bool amdgpu_mem_visible(struct amdgpu_device *adev, 430 struct ttm_resource *mem) 431 { 432 uint64_t mem_size = (u64)mem->num_pages << PAGE_SHIFT; 433 struct amdgpu_res_cursor cursor; 434 435 if (mem->mem_type == TTM_PL_SYSTEM || 436 mem->mem_type == TTM_PL_TT) 437 return true; 438 if (mem->mem_type != TTM_PL_VRAM) 439 return false; 440 441 amdgpu_res_first(mem, 0, mem_size, &cursor); 442 443 /* ttm_resource_ioremap only supports contiguous memory */ 444 if (cursor.size != mem_size) 445 return false; 446 447 return cursor.start + cursor.size <= adev->gmc.visible_vram_size; 448 } 449 450 /* 451 * amdgpu_bo_move - Move a buffer object to a new memory location 452 * 453 * Called by ttm_bo_handle_move_mem() 454 */ 455 static int amdgpu_bo_move(struct ttm_buffer_object *bo, bool evict, 456 struct ttm_operation_ctx *ctx, 457 struct ttm_resource *new_mem, 458 struct ttm_place *hop) 459 { 460 struct amdgpu_device *adev; 461 struct amdgpu_bo *abo; 462 struct ttm_resource *old_mem = bo->resource; 463 int r; 464 465 if (new_mem->mem_type == TTM_PL_TT || 466 new_mem->mem_type == AMDGPU_PL_PREEMPT) { 467 r = amdgpu_ttm_backend_bind(bo->bdev, bo->ttm, new_mem); 468 if (r) 469 return r; 470 } 471 472 /* Can't move a pinned BO */ 473 abo = ttm_to_amdgpu_bo(bo); 474 if (WARN_ON_ONCE(abo->tbo.pin_count > 0)) 475 return -EINVAL; 476 477 adev = amdgpu_ttm_adev(bo->bdev); 478 479 if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) { 480 ttm_bo_move_null(bo, new_mem); 481 goto out; 482 } 483 if (old_mem->mem_type == TTM_PL_SYSTEM && 484 (new_mem->mem_type == TTM_PL_TT || 485 new_mem->mem_type == AMDGPU_PL_PREEMPT)) { 486 ttm_bo_move_null(bo, new_mem); 487 goto out; 488 } 489 if ((old_mem->mem_type == TTM_PL_TT || 490 old_mem->mem_type == AMDGPU_PL_PREEMPT) && 491 new_mem->mem_type == TTM_PL_SYSTEM) { 492 r = ttm_bo_wait_ctx(bo, ctx); 493 if (r) 494 return r; 495 496 amdgpu_ttm_backend_unbind(bo->bdev, bo->ttm); 497 ttm_resource_free(bo, &bo->resource); 498 ttm_bo_assign_mem(bo, new_mem); 499 goto out; 500 } 501 502 if (old_mem->mem_type == AMDGPU_PL_GDS || 503 old_mem->mem_type == AMDGPU_PL_GWS || 504 old_mem->mem_type == AMDGPU_PL_OA || 505 new_mem->mem_type == AMDGPU_PL_GDS || 506 new_mem->mem_type == AMDGPU_PL_GWS || 507 new_mem->mem_type == AMDGPU_PL_OA) { 508 /* Nothing to save here */ 509 ttm_bo_move_null(bo, new_mem); 510 goto out; 511 } 512 513 if (bo->type == ttm_bo_type_device && 514 new_mem->mem_type == TTM_PL_VRAM && 515 old_mem->mem_type != TTM_PL_VRAM) { 516 /* amdgpu_bo_fault_reserve_notify will re-set this if the CPU 517 * accesses the BO after it's moved. 518 */ 519 abo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED; 520 } 521 522 if (adev->mman.buffer_funcs_enabled) { 523 if (((old_mem->mem_type == TTM_PL_SYSTEM && 524 new_mem->mem_type == TTM_PL_VRAM) || 525 (old_mem->mem_type == TTM_PL_VRAM && 526 new_mem->mem_type == TTM_PL_SYSTEM))) { 527 hop->fpfn = 0; 528 hop->lpfn = 0; 529 hop->mem_type = TTM_PL_TT; 530 hop->flags = TTM_PL_FLAG_TEMPORARY; 531 return -EMULTIHOP; 532 } 533 534 r = amdgpu_move_blit(bo, evict, new_mem, old_mem); 535 } else { 536 r = -ENODEV; 537 } 538 539 if (r) { 540 /* Check that all memory is CPU accessible */ 541 if (!amdgpu_mem_visible(adev, old_mem) || 542 !amdgpu_mem_visible(adev, new_mem)) { 543 pr_err("Move buffer fallback to memcpy unavailable\n"); 544 return r; 545 } 546 547 r = ttm_bo_move_memcpy(bo, ctx, new_mem); 548 if (r) 549 return r; 550 } 551 552 out: 553 /* update statistics */ 554 atomic64_add(bo->base.size, &adev->num_bytes_moved); 555 amdgpu_bo_move_notify(bo, evict, new_mem); 556 return 0; 557 } 558 559 /* 560 * amdgpu_ttm_io_mem_reserve - Reserve a block of memory during a fault 561 * 562 * Called by ttm_mem_io_reserve() ultimately via ttm_bo_vm_fault() 563 */ 564 static int amdgpu_ttm_io_mem_reserve(struct ttm_device *bdev, 565 struct ttm_resource *mem) 566 { 567 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev); 568 size_t bus_size = (size_t)mem->num_pages << PAGE_SHIFT; 569 570 switch (mem->mem_type) { 571 case TTM_PL_SYSTEM: 572 /* system memory */ 573 return 0; 574 case TTM_PL_TT: 575 case AMDGPU_PL_PREEMPT: 576 break; 577 case TTM_PL_VRAM: 578 mem->bus.offset = mem->start << PAGE_SHIFT; 579 /* check if it's visible */ 580 if ((mem->bus.offset + bus_size) > adev->gmc.visible_vram_size) 581 return -EINVAL; 582 583 if (adev->mman.aper_base_kaddr && 584 mem->placement & TTM_PL_FLAG_CONTIGUOUS) 585 mem->bus.addr = (u8 *)adev->mman.aper_base_kaddr + 586 mem->bus.offset; 587 588 mem->bus.offset += adev->gmc.aper_base; 589 mem->bus.is_iomem = true; 590 break; 591 default: 592 return -EINVAL; 593 } 594 return 0; 595 } 596 597 static unsigned long amdgpu_ttm_io_mem_pfn(struct ttm_buffer_object *bo, 598 unsigned long page_offset) 599 { 600 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev); 601 struct amdgpu_res_cursor cursor; 602 603 amdgpu_res_first(bo->resource, (u64)page_offset << PAGE_SHIFT, 0, 604 &cursor); 605 return (adev->gmc.aper_base + cursor.start) >> PAGE_SHIFT; 606 } 607 608 /** 609 * amdgpu_ttm_domain_start - Returns GPU start address 610 * @adev: amdgpu device object 611 * @type: type of the memory 612 * 613 * Returns: 614 * GPU start address of a memory domain 615 */ 616 617 uint64_t amdgpu_ttm_domain_start(struct amdgpu_device *adev, uint32_t type) 618 { 619 switch (type) { 620 case TTM_PL_TT: 621 return adev->gmc.gart_start; 622 case TTM_PL_VRAM: 623 return adev->gmc.vram_start; 624 } 625 626 return 0; 627 } 628 629 /* 630 * TTM backend functions. 631 */ 632 struct amdgpu_ttm_tt { 633 struct ttm_tt ttm; 634 struct drm_gem_object *gobj; 635 u64 offset; 636 uint64_t userptr; 637 struct task_struct *usertask; 638 uint32_t userflags; 639 bool bound; 640 #if IS_ENABLED(CONFIG_DRM_AMDGPU_USERPTR) 641 struct hmm_range *range; 642 #endif 643 }; 644 645 #ifdef CONFIG_DRM_AMDGPU_USERPTR 646 /* 647 * amdgpu_ttm_tt_get_user_pages - get device accessible pages that back user 648 * memory and start HMM tracking CPU page table update 649 * 650 * Calling function must call amdgpu_ttm_tt_userptr_range_done() once and only 651 * once afterwards to stop HMM tracking 652 */ 653 int amdgpu_ttm_tt_get_user_pages(struct amdgpu_bo *bo, struct page **pages) 654 { 655 struct ttm_tt *ttm = bo->tbo.ttm; 656 struct amdgpu_ttm_tt *gtt = (void *)ttm; 657 unsigned long start = gtt->userptr; 658 struct vm_area_struct *vma; 659 struct mm_struct *mm; 660 bool readonly; 661 int r = 0; 662 663 mm = bo->notifier.mm; 664 if (unlikely(!mm)) { 665 DRM_DEBUG_DRIVER("BO is not registered?\n"); 666 return -EFAULT; 667 } 668 669 /* Another get_user_pages is running at the same time?? */ 670 if (WARN_ON(gtt->range)) 671 return -EFAULT; 672 673 if (!mmget_not_zero(mm)) /* Happens during process shutdown */ 674 return -ESRCH; 675 676 mmap_read_lock(mm); 677 vma = vma_lookup(mm, start); 678 if (unlikely(!vma)) { 679 r = -EFAULT; 680 goto out_unlock; 681 } 682 if (unlikely((gtt->userflags & AMDGPU_GEM_USERPTR_ANONONLY) && 683 vma->vm_file)) { 684 r = -EPERM; 685 goto out_unlock; 686 } 687 688 readonly = amdgpu_ttm_tt_is_readonly(ttm); 689 r = amdgpu_hmm_range_get_pages(&bo->notifier, mm, pages, start, 690 ttm->num_pages, >t->range, readonly, 691 true, NULL); 692 out_unlock: 693 mmap_read_unlock(mm); 694 if (r) 695 pr_debug("failed %d to get user pages 0x%lx\n", r, start); 696 697 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_TT_FLAG_EXTERNAL) { 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 != TTM_PL_TT || 918 !amdgpu_gtt_mgr_has_gart_addr(bo_mem)) { 919 gtt->offset = AMDGPU_BO_INVALID_OFFSET; 920 return 0; 921 } 922 923 /* compute PTE flags relevant to this BO memory */ 924 flags = amdgpu_ttm_tt_pte_flags(adev, ttm, bo_mem); 925 926 /* bind pages into GART page tables */ 927 gtt->offset = (u64)bo_mem->start << PAGE_SHIFT; 928 r = amdgpu_gart_bind(adev, gtt->offset, ttm->num_pages, 929 gtt->ttm.dma_address, flags); 930 931 if (r) 932 DRM_ERROR("failed to bind %u pages at 0x%08llX\n", 933 ttm->num_pages, gtt->offset); 934 gtt->bound = true; 935 return r; 936 } 937 938 /* 939 * amdgpu_ttm_alloc_gart - Make sure buffer object is accessible either 940 * through AGP or GART aperture. 941 * 942 * If bo is accessible through AGP aperture, then use AGP aperture 943 * to access bo; otherwise allocate logical space in GART aperture 944 * and map bo to GART aperture. 945 */ 946 int amdgpu_ttm_alloc_gart(struct ttm_buffer_object *bo) 947 { 948 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev); 949 struct ttm_operation_ctx ctx = { false, false }; 950 struct amdgpu_ttm_tt *gtt = (void *)bo->ttm; 951 struct ttm_placement placement; 952 struct ttm_place placements; 953 struct ttm_resource *tmp; 954 uint64_t addr, flags; 955 int r; 956 957 if (bo->resource->start != AMDGPU_BO_INVALID_OFFSET) 958 return 0; 959 960 addr = amdgpu_gmc_agp_addr(bo); 961 if (addr != AMDGPU_BO_INVALID_OFFSET) { 962 bo->resource->start = addr >> PAGE_SHIFT; 963 return 0; 964 } 965 966 /* allocate GART space */ 967 placement.num_placement = 1; 968 placement.placement = &placements; 969 placement.num_busy_placement = 1; 970 placement.busy_placement = &placements; 971 placements.fpfn = 0; 972 placements.lpfn = adev->gmc.gart_size >> PAGE_SHIFT; 973 placements.mem_type = TTM_PL_TT; 974 placements.flags = bo->resource->placement; 975 976 r = ttm_bo_mem_space(bo, &placement, &tmp, &ctx); 977 if (unlikely(r)) 978 return r; 979 980 /* compute PTE flags for this buffer object */ 981 flags = amdgpu_ttm_tt_pte_flags(adev, bo->ttm, tmp); 982 983 /* Bind pages */ 984 gtt->offset = (u64)tmp->start << PAGE_SHIFT; 985 r = amdgpu_ttm_gart_bind(adev, bo, flags); 986 if (unlikely(r)) { 987 ttm_resource_free(bo, &tmp); 988 return r; 989 } 990 991 amdgpu_gart_invalidate_tlb(adev); 992 ttm_resource_free(bo, &bo->resource); 993 ttm_bo_assign_mem(bo, tmp); 994 995 return 0; 996 } 997 998 /* 999 * amdgpu_ttm_recover_gart - Rebind GTT pages 1000 * 1001 * Called by amdgpu_gtt_mgr_recover() from amdgpu_device_reset() to 1002 * rebind GTT pages during a GPU reset. 1003 */ 1004 int amdgpu_ttm_recover_gart(struct ttm_buffer_object *tbo) 1005 { 1006 struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev); 1007 uint64_t flags; 1008 int r; 1009 1010 if (!tbo->ttm) 1011 return 0; 1012 1013 flags = amdgpu_ttm_tt_pte_flags(adev, tbo->ttm, tbo->resource); 1014 r = amdgpu_ttm_gart_bind(adev, tbo, flags); 1015 1016 return r; 1017 } 1018 1019 /* 1020 * amdgpu_ttm_backend_unbind - Unbind GTT mapped pages 1021 * 1022 * Called by ttm_tt_unbind() on behalf of ttm_bo_move_ttm() and 1023 * ttm_tt_destroy(). 1024 */ 1025 static void amdgpu_ttm_backend_unbind(struct ttm_device *bdev, 1026 struct ttm_tt *ttm) 1027 { 1028 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev); 1029 struct amdgpu_ttm_tt *gtt = (void *)ttm; 1030 int r; 1031 1032 /* if the pages have userptr pinning then clear that first */ 1033 if (gtt->userptr) { 1034 amdgpu_ttm_tt_unpin_userptr(bdev, ttm); 1035 } else if (ttm->sg && gtt->gobj->import_attach) { 1036 struct dma_buf_attachment *attach; 1037 1038 attach = gtt->gobj->import_attach; 1039 dma_buf_unmap_attachment(attach, ttm->sg, DMA_BIDIRECTIONAL); 1040 ttm->sg = NULL; 1041 } 1042 1043 if (!gtt->bound) 1044 return; 1045 1046 if (gtt->offset == AMDGPU_BO_INVALID_OFFSET) 1047 return; 1048 1049 /* unbind shouldn't be done for GDS/GWS/OA in ttm_bo_clean_mm */ 1050 r = amdgpu_gart_unbind(adev, gtt->offset, ttm->num_pages); 1051 if (r) 1052 DRM_ERROR("failed to unbind %u pages at 0x%08llX\n", 1053 gtt->ttm.num_pages, gtt->offset); 1054 gtt->bound = false; 1055 } 1056 1057 static void amdgpu_ttm_backend_destroy(struct ttm_device *bdev, 1058 struct ttm_tt *ttm) 1059 { 1060 struct amdgpu_ttm_tt *gtt = (void *)ttm; 1061 1062 if (gtt->usertask) 1063 put_task_struct(gtt->usertask); 1064 1065 ttm_tt_fini(>t->ttm); 1066 kfree(gtt); 1067 } 1068 1069 /** 1070 * amdgpu_ttm_tt_create - Create a ttm_tt object for a given BO 1071 * 1072 * @bo: The buffer object to create a GTT ttm_tt object around 1073 * @page_flags: Page flags to be added to the ttm_tt object 1074 * 1075 * Called by ttm_tt_create(). 1076 */ 1077 static struct ttm_tt *amdgpu_ttm_tt_create(struct ttm_buffer_object *bo, 1078 uint32_t page_flags) 1079 { 1080 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo); 1081 struct amdgpu_ttm_tt *gtt; 1082 enum ttm_caching caching; 1083 1084 gtt = kzalloc(sizeof(struct amdgpu_ttm_tt), GFP_KERNEL); 1085 if (gtt == NULL) { 1086 return NULL; 1087 } 1088 gtt->gobj = &bo->base; 1089 1090 if (abo->flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC) 1091 caching = ttm_write_combined; 1092 else 1093 caching = ttm_cached; 1094 1095 /* allocate space for the uninitialized page entries */ 1096 if (ttm_sg_tt_init(>t->ttm, bo, page_flags, caching)) { 1097 kfree(gtt); 1098 return NULL; 1099 } 1100 return >t->ttm; 1101 } 1102 1103 /* 1104 * amdgpu_ttm_tt_populate - Map GTT pages visible to the device 1105 * 1106 * Map the pages of a ttm_tt object to an address space visible 1107 * to the underlying device. 1108 */ 1109 static int amdgpu_ttm_tt_populate(struct ttm_device *bdev, 1110 struct ttm_tt *ttm, 1111 struct ttm_operation_ctx *ctx) 1112 { 1113 struct amdgpu_device *adev = amdgpu_ttm_adev(bdev); 1114 struct amdgpu_ttm_tt *gtt = (void *)ttm; 1115 pgoff_t i; 1116 int ret; 1117 1118 /* user pages are bound by amdgpu_ttm_tt_pin_userptr() */ 1119 if (gtt->userptr) { 1120 ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL); 1121 if (!ttm->sg) 1122 return -ENOMEM; 1123 return 0; 1124 } 1125 1126 if (ttm->page_flags & TTM_TT_FLAG_EXTERNAL) 1127 return 0; 1128 1129 ret = ttm_pool_alloc(&adev->mman.bdev.pool, ttm, ctx); 1130 if (ret) 1131 return ret; 1132 1133 for (i = 0; i < ttm->num_pages; ++i) 1134 ttm->pages[i]->mapping = bdev->dev_mapping; 1135 1136 return 0; 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 pgoff_t i; 1151 1152 amdgpu_ttm_backend_unbind(bdev, ttm); 1153 1154 if (gtt->userptr) { 1155 amdgpu_ttm_tt_set_user_pages(ttm, NULL); 1156 kfree(ttm->sg); 1157 ttm->sg = NULL; 1158 return; 1159 } 1160 1161 if (ttm->page_flags & TTM_TT_FLAG_EXTERNAL) 1162 return; 1163 1164 for (i = 0; i < ttm->num_pages; ++i) 1165 ttm->pages[i]->mapping = NULL; 1166 1167 adev = amdgpu_ttm_adev(bdev); 1168 return ttm_pool_free(&adev->mman.bdev.pool, ttm); 1169 } 1170 1171 /** 1172 * amdgpu_ttm_tt_set_userptr - Initialize userptr GTT ttm_tt for the current 1173 * task 1174 * 1175 * @bo: The ttm_buffer_object to bind this userptr to 1176 * @addr: The address in the current tasks VM space to use 1177 * @flags: Requirements of userptr object. 1178 * 1179 * Called by amdgpu_gem_userptr_ioctl() to bind userptr pages 1180 * to current task 1181 */ 1182 int amdgpu_ttm_tt_set_userptr(struct ttm_buffer_object *bo, 1183 uint64_t addr, uint32_t flags) 1184 { 1185 struct amdgpu_ttm_tt *gtt; 1186 1187 if (!bo->ttm) { 1188 /* TODO: We want a separate TTM object type for userptrs */ 1189 bo->ttm = amdgpu_ttm_tt_create(bo, 0); 1190 if (bo->ttm == NULL) 1191 return -ENOMEM; 1192 } 1193 1194 /* Set TTM_TT_FLAG_EXTERNAL before populate but after create. */ 1195 bo->ttm->page_flags |= TTM_TT_FLAG_EXTERNAL; 1196 1197 gtt = (void *)bo->ttm; 1198 gtt->userptr = addr; 1199 gtt->userflags = flags; 1200 1201 if (gtt->usertask) 1202 put_task_struct(gtt->usertask); 1203 gtt->usertask = current->group_leader; 1204 get_task_struct(gtt->usertask); 1205 1206 return 0; 1207 } 1208 1209 /* 1210 * amdgpu_ttm_tt_get_usermm - Return memory manager for ttm_tt object 1211 */ 1212 struct mm_struct *amdgpu_ttm_tt_get_usermm(struct ttm_tt *ttm) 1213 { 1214 struct amdgpu_ttm_tt *gtt = (void *)ttm; 1215 1216 if (gtt == NULL) 1217 return NULL; 1218 1219 if (gtt->usertask == NULL) 1220 return NULL; 1221 1222 return gtt->usertask->mm; 1223 } 1224 1225 /* 1226 * amdgpu_ttm_tt_affect_userptr - Determine if a ttm_tt object lays inside an 1227 * address range for the current task. 1228 * 1229 */ 1230 bool amdgpu_ttm_tt_affect_userptr(struct ttm_tt *ttm, unsigned long start, 1231 unsigned long end, unsigned long *userptr) 1232 { 1233 struct amdgpu_ttm_tt *gtt = (void *)ttm; 1234 unsigned long size; 1235 1236 if (gtt == NULL || !gtt->userptr) 1237 return false; 1238 1239 /* Return false if no part of the ttm_tt object lies within 1240 * the range 1241 */ 1242 size = (unsigned long)gtt->ttm.num_pages * PAGE_SIZE; 1243 if (gtt->userptr > end || gtt->userptr + size <= start) 1244 return false; 1245 1246 if (userptr) 1247 *userptr = gtt->userptr; 1248 return true; 1249 } 1250 1251 /* 1252 * amdgpu_ttm_tt_is_userptr - Have the pages backing by userptr? 1253 */ 1254 bool amdgpu_ttm_tt_is_userptr(struct ttm_tt *ttm) 1255 { 1256 struct amdgpu_ttm_tt *gtt = (void *)ttm; 1257 1258 if (gtt == NULL || !gtt->userptr) 1259 return false; 1260 1261 return true; 1262 } 1263 1264 /* 1265 * amdgpu_ttm_tt_is_readonly - Is the ttm_tt object read only? 1266 */ 1267 bool amdgpu_ttm_tt_is_readonly(struct ttm_tt *ttm) 1268 { 1269 struct amdgpu_ttm_tt *gtt = (void *)ttm; 1270 1271 if (gtt == NULL) 1272 return false; 1273 1274 return !!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY); 1275 } 1276 1277 /** 1278 * amdgpu_ttm_tt_pde_flags - Compute PDE flags for ttm_tt object 1279 * 1280 * @ttm: The ttm_tt object to compute the flags for 1281 * @mem: The memory registry backing this ttm_tt object 1282 * 1283 * Figure out the flags to use for a VM PDE (Page Directory Entry). 1284 */ 1285 uint64_t amdgpu_ttm_tt_pde_flags(struct ttm_tt *ttm, struct ttm_resource *mem) 1286 { 1287 uint64_t flags = 0; 1288 1289 if (mem && mem->mem_type != TTM_PL_SYSTEM) 1290 flags |= AMDGPU_PTE_VALID; 1291 1292 if (mem && (mem->mem_type == TTM_PL_TT || 1293 mem->mem_type == AMDGPU_PL_PREEMPT)) { 1294 flags |= AMDGPU_PTE_SYSTEM; 1295 1296 if (ttm->caching == ttm_cached) 1297 flags |= AMDGPU_PTE_SNOOPED; 1298 } 1299 1300 if (mem && mem->mem_type == TTM_PL_VRAM && 1301 mem->bus.caching == ttm_cached) 1302 flags |= AMDGPU_PTE_SNOOPED; 1303 1304 return flags; 1305 } 1306 1307 /** 1308 * amdgpu_ttm_tt_pte_flags - Compute PTE flags for ttm_tt object 1309 * 1310 * @adev: amdgpu_device pointer 1311 * @ttm: The ttm_tt object to compute the flags for 1312 * @mem: The memory registry backing this ttm_tt object 1313 * 1314 * Figure out the flags to use for a VM PTE (Page Table Entry). 1315 */ 1316 uint64_t amdgpu_ttm_tt_pte_flags(struct amdgpu_device *adev, struct ttm_tt *ttm, 1317 struct ttm_resource *mem) 1318 { 1319 uint64_t flags = amdgpu_ttm_tt_pde_flags(ttm, mem); 1320 1321 flags |= adev->gart.gart_pte_flags; 1322 flags |= AMDGPU_PTE_READABLE; 1323 1324 if (!amdgpu_ttm_tt_is_readonly(ttm)) 1325 flags |= AMDGPU_PTE_WRITEABLE; 1326 1327 return flags; 1328 } 1329 1330 /* 1331 * amdgpu_ttm_bo_eviction_valuable - Check to see if we can evict a buffer 1332 * object. 1333 * 1334 * Return true if eviction is sensible. Called by ttm_mem_evict_first() on 1335 * behalf of ttm_bo_mem_force_space() which tries to evict buffer objects until 1336 * it can find space for a new object and by ttm_bo_force_list_clean() which is 1337 * used to clean out a memory space. 1338 */ 1339 static bool amdgpu_ttm_bo_eviction_valuable(struct ttm_buffer_object *bo, 1340 const struct ttm_place *place) 1341 { 1342 unsigned long num_pages = bo->resource->num_pages; 1343 struct dma_resv_iter resv_cursor; 1344 struct amdgpu_res_cursor cursor; 1345 struct dma_fence *f; 1346 1347 /* Swapout? */ 1348 if (bo->resource->mem_type == TTM_PL_SYSTEM) 1349 return true; 1350 1351 if (bo->type == ttm_bo_type_kernel && 1352 !amdgpu_vm_evictable(ttm_to_amdgpu_bo(bo))) 1353 return false; 1354 1355 /* If bo is a KFD BO, check if the bo belongs to the current process. 1356 * If true, then return false as any KFD process needs all its BOs to 1357 * be resident to run successfully 1358 */ 1359 dma_resv_for_each_fence(&resv_cursor, bo->base.resv, true, f) { 1360 if (amdkfd_fence_check_mm(f, current->mm)) 1361 return false; 1362 } 1363 1364 switch (bo->resource->mem_type) { 1365 case AMDGPU_PL_PREEMPT: 1366 /* Preemptible BOs don't own system resources managed by the 1367 * driver (pages, VRAM, GART space). They point to resources 1368 * owned by someone else (e.g. pageable memory in user mode 1369 * or a DMABuf). They are used in a preemptible context so we 1370 * can guarantee no deadlocks and good QoS in case of MMU 1371 * notifiers or DMABuf move notifiers from the resource owner. 1372 */ 1373 return false; 1374 case TTM_PL_TT: 1375 if (amdgpu_bo_is_amdgpu_bo(bo) && 1376 amdgpu_bo_encrypted(ttm_to_amdgpu_bo(bo))) 1377 return false; 1378 return true; 1379 1380 case TTM_PL_VRAM: 1381 /* Check each drm MM node individually */ 1382 amdgpu_res_first(bo->resource, 0, (u64)num_pages << PAGE_SHIFT, 1383 &cursor); 1384 while (cursor.remaining) { 1385 if (place->fpfn < PFN_DOWN(cursor.start + cursor.size) 1386 && !(place->lpfn && 1387 place->lpfn <= PFN_DOWN(cursor.start))) 1388 return true; 1389 1390 amdgpu_res_next(&cursor, cursor.size); 1391 } 1392 return false; 1393 1394 default: 1395 break; 1396 } 1397 1398 return ttm_bo_eviction_valuable(bo, place); 1399 } 1400 1401 static void amdgpu_ttm_vram_mm_access(struct amdgpu_device *adev, loff_t pos, 1402 void *buf, size_t size, bool write) 1403 { 1404 while (size) { 1405 uint64_t aligned_pos = ALIGN_DOWN(pos, 4); 1406 uint64_t bytes = 4 - (pos & 0x3); 1407 uint32_t shift = (pos & 0x3) * 8; 1408 uint32_t mask = 0xffffffff << shift; 1409 uint32_t value = 0; 1410 1411 if (size < bytes) { 1412 mask &= 0xffffffff >> (bytes - size) * 8; 1413 bytes = size; 1414 } 1415 1416 if (mask != 0xffffffff) { 1417 amdgpu_device_mm_access(adev, aligned_pos, &value, 4, false); 1418 if (write) { 1419 value &= ~mask; 1420 value |= (*(uint32_t *)buf << shift) & mask; 1421 amdgpu_device_mm_access(adev, aligned_pos, &value, 4, true); 1422 } else { 1423 value = (value & mask) >> shift; 1424 memcpy(buf, &value, bytes); 1425 } 1426 } else { 1427 amdgpu_device_mm_access(adev, aligned_pos, buf, 4, write); 1428 } 1429 1430 pos += bytes; 1431 buf += bytes; 1432 size -= bytes; 1433 } 1434 } 1435 1436 /** 1437 * amdgpu_ttm_access_memory - Read or Write memory that backs a buffer object. 1438 * 1439 * @bo: The buffer object to read/write 1440 * @offset: Offset into buffer object 1441 * @buf: Secondary buffer to write/read from 1442 * @len: Length in bytes of access 1443 * @write: true if writing 1444 * 1445 * This is used to access VRAM that backs a buffer object via MMIO 1446 * access for debugging purposes. 1447 */ 1448 static int amdgpu_ttm_access_memory(struct ttm_buffer_object *bo, 1449 unsigned long offset, void *buf, int len, 1450 int write) 1451 { 1452 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo); 1453 struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev); 1454 struct amdgpu_res_cursor cursor; 1455 int ret = 0; 1456 1457 if (bo->resource->mem_type != TTM_PL_VRAM) 1458 return -EIO; 1459 1460 amdgpu_res_first(bo->resource, offset, len, &cursor); 1461 while (cursor.remaining) { 1462 size_t count, size = cursor.size; 1463 loff_t pos = cursor.start; 1464 1465 count = amdgpu_device_aper_access(adev, pos, buf, size, write); 1466 size -= count; 1467 if (size) { 1468 /* using MM to access rest vram and handle un-aligned address */ 1469 pos += count; 1470 buf += count; 1471 amdgpu_ttm_vram_mm_access(adev, pos, buf, size, write); 1472 } 1473 1474 ret += cursor.size; 1475 buf += cursor.size; 1476 amdgpu_res_next(&cursor, cursor.size); 1477 } 1478 1479 return ret; 1480 } 1481 1482 static void 1483 amdgpu_bo_delete_mem_notify(struct ttm_buffer_object *bo) 1484 { 1485 amdgpu_bo_move_notify(bo, false, NULL); 1486 } 1487 1488 static struct ttm_device_funcs amdgpu_bo_driver = { 1489 .ttm_tt_create = &amdgpu_ttm_tt_create, 1490 .ttm_tt_populate = &amdgpu_ttm_tt_populate, 1491 .ttm_tt_unpopulate = &amdgpu_ttm_tt_unpopulate, 1492 .ttm_tt_destroy = &amdgpu_ttm_backend_destroy, 1493 .eviction_valuable = amdgpu_ttm_bo_eviction_valuable, 1494 .evict_flags = &amdgpu_evict_flags, 1495 .move = &amdgpu_bo_move, 1496 .delete_mem_notify = &amdgpu_bo_delete_mem_notify, 1497 .release_notify = &amdgpu_bo_release_notify, 1498 .io_mem_reserve = &amdgpu_ttm_io_mem_reserve, 1499 .io_mem_pfn = amdgpu_ttm_io_mem_pfn, 1500 .access_memory = &amdgpu_ttm_access_memory, 1501 .del_from_lru_notify = &amdgpu_vm_del_from_lru_notify 1502 }; 1503 1504 /* 1505 * Firmware Reservation functions 1506 */ 1507 /** 1508 * amdgpu_ttm_fw_reserve_vram_fini - free fw reserved vram 1509 * 1510 * @adev: amdgpu_device pointer 1511 * 1512 * free fw reserved vram if it has been reserved. 1513 */ 1514 static void amdgpu_ttm_fw_reserve_vram_fini(struct amdgpu_device *adev) 1515 { 1516 amdgpu_bo_free_kernel(&adev->mman.fw_vram_usage_reserved_bo, 1517 NULL, &adev->mman.fw_vram_usage_va); 1518 } 1519 1520 /** 1521 * amdgpu_ttm_fw_reserve_vram_init - create bo vram reservation from fw 1522 * 1523 * @adev: amdgpu_device pointer 1524 * 1525 * create bo vram reservation from fw. 1526 */ 1527 static int amdgpu_ttm_fw_reserve_vram_init(struct amdgpu_device *adev) 1528 { 1529 uint64_t vram_size = adev->gmc.visible_vram_size; 1530 1531 adev->mman.fw_vram_usage_va = NULL; 1532 adev->mman.fw_vram_usage_reserved_bo = NULL; 1533 1534 if (adev->mman.fw_vram_usage_size == 0 || 1535 adev->mman.fw_vram_usage_size > vram_size) 1536 return 0; 1537 1538 return amdgpu_bo_create_kernel_at(adev, 1539 adev->mman.fw_vram_usage_start_offset, 1540 adev->mman.fw_vram_usage_size, 1541 AMDGPU_GEM_DOMAIN_VRAM, 1542 &adev->mman.fw_vram_usage_reserved_bo, 1543 &adev->mman.fw_vram_usage_va); 1544 } 1545 1546 /* 1547 * Memoy training reservation functions 1548 */ 1549 1550 /** 1551 * amdgpu_ttm_training_reserve_vram_fini - free memory training reserved vram 1552 * 1553 * @adev: amdgpu_device pointer 1554 * 1555 * free memory training reserved vram if it has been reserved. 1556 */ 1557 static int amdgpu_ttm_training_reserve_vram_fini(struct amdgpu_device *adev) 1558 { 1559 struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx; 1560 1561 ctx->init = PSP_MEM_TRAIN_NOT_SUPPORT; 1562 amdgpu_bo_free_kernel(&ctx->c2p_bo, NULL, NULL); 1563 ctx->c2p_bo = NULL; 1564 1565 return 0; 1566 } 1567 1568 static void amdgpu_ttm_training_data_block_init(struct amdgpu_device *adev) 1569 { 1570 struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx; 1571 1572 memset(ctx, 0, sizeof(*ctx)); 1573 1574 ctx->c2p_train_data_offset = 1575 ALIGN((adev->gmc.mc_vram_size - adev->mman.discovery_tmr_size - SZ_1M), SZ_1M); 1576 ctx->p2c_train_data_offset = 1577 (adev->gmc.mc_vram_size - GDDR6_MEM_TRAINING_OFFSET); 1578 ctx->train_data_size = 1579 GDDR6_MEM_TRAINING_DATA_SIZE_IN_BYTES; 1580 1581 DRM_DEBUG("train_data_size:%llx,p2c_train_data_offset:%llx,c2p_train_data_offset:%llx.\n", 1582 ctx->train_data_size, 1583 ctx->p2c_train_data_offset, 1584 ctx->c2p_train_data_offset); 1585 } 1586 1587 /* 1588 * reserve TMR memory at the top of VRAM which holds 1589 * IP Discovery data and is protected by PSP. 1590 */ 1591 static int amdgpu_ttm_reserve_tmr(struct amdgpu_device *adev) 1592 { 1593 int ret; 1594 struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx; 1595 bool mem_train_support = false; 1596 1597 if (!amdgpu_sriov_vf(adev)) { 1598 if (amdgpu_atomfirmware_mem_training_supported(adev)) 1599 mem_train_support = true; 1600 else 1601 DRM_DEBUG("memory training does not support!\n"); 1602 } 1603 1604 /* 1605 * Query reserved tmr size through atom firmwareinfo for Sienna_Cichlid and onwards for all 1606 * the use cases (IP discovery/G6 memory training/profiling/diagnostic data.etc) 1607 * 1608 * Otherwise, fallback to legacy approach to check and reserve tmr block for ip 1609 * discovery data and G6 memory training data respectively 1610 */ 1611 adev->mman.discovery_tmr_size = 1612 amdgpu_atomfirmware_get_fw_reserved_fb_size(adev); 1613 if (!adev->mman.discovery_tmr_size) 1614 adev->mman.discovery_tmr_size = DISCOVERY_TMR_OFFSET; 1615 1616 if (mem_train_support) { 1617 /* reserve vram for mem train according to TMR location */ 1618 amdgpu_ttm_training_data_block_init(adev); 1619 ret = amdgpu_bo_create_kernel_at(adev, 1620 ctx->c2p_train_data_offset, 1621 ctx->train_data_size, 1622 AMDGPU_GEM_DOMAIN_VRAM, 1623 &ctx->c2p_bo, 1624 NULL); 1625 if (ret) { 1626 DRM_ERROR("alloc c2p_bo failed(%d)!\n", ret); 1627 amdgpu_ttm_training_reserve_vram_fini(adev); 1628 return ret; 1629 } 1630 ctx->init = PSP_MEM_TRAIN_RESERVE_SUCCESS; 1631 } 1632 1633 ret = amdgpu_bo_create_kernel_at(adev, 1634 adev->gmc.real_vram_size - adev->mman.discovery_tmr_size, 1635 adev->mman.discovery_tmr_size, 1636 AMDGPU_GEM_DOMAIN_VRAM, 1637 &adev->mman.discovery_memory, 1638 NULL); 1639 if (ret) { 1640 DRM_ERROR("alloc tmr failed(%d)!\n", ret); 1641 amdgpu_bo_free_kernel(&adev->mman.discovery_memory, NULL, NULL); 1642 return ret; 1643 } 1644 1645 return 0; 1646 } 1647 1648 /* 1649 * amdgpu_ttm_init - Init the memory management (ttm) as well as various 1650 * gtt/vram related fields. 1651 * 1652 * This initializes all of the memory space pools that the TTM layer 1653 * will need such as the GTT space (system memory mapped to the device), 1654 * VRAM (on-board memory), and on-chip memories (GDS, GWS, OA) which 1655 * can be mapped per VMID. 1656 */ 1657 int amdgpu_ttm_init(struct amdgpu_device *adev) 1658 { 1659 uint64_t gtt_size; 1660 int r; 1661 u64 vis_vram_limit; 1662 1663 mutex_init(&adev->mman.gtt_window_lock); 1664 1665 /* No others user of address space so set it to 0 */ 1666 r = ttm_device_init(&adev->mman.bdev, &amdgpu_bo_driver, adev->dev, 1667 adev_to_drm(adev)->anon_inode->i_mapping, 1668 adev_to_drm(adev)->vma_offset_manager, 1669 adev->need_swiotlb, 1670 dma_addressing_limited(adev->dev)); 1671 if (r) { 1672 DRM_ERROR("failed initializing buffer object driver(%d).\n", r); 1673 return r; 1674 } 1675 adev->mman.initialized = true; 1676 1677 /* Initialize VRAM pool with all of VRAM divided into pages */ 1678 r = amdgpu_vram_mgr_init(adev); 1679 if (r) { 1680 DRM_ERROR("Failed initializing VRAM heap.\n"); 1681 return r; 1682 } 1683 1684 /* Reduce size of CPU-visible VRAM if requested */ 1685 vis_vram_limit = (u64)amdgpu_vis_vram_limit * 1024 * 1024; 1686 if (amdgpu_vis_vram_limit > 0 && 1687 vis_vram_limit <= adev->gmc.visible_vram_size) 1688 adev->gmc.visible_vram_size = vis_vram_limit; 1689 1690 /* Change the size here instead of the init above so only lpfn is affected */ 1691 amdgpu_ttm_set_buffer_funcs_status(adev, false); 1692 #ifdef CONFIG_64BIT 1693 #ifdef CONFIG_X86 1694 if (adev->gmc.xgmi.connected_to_cpu) 1695 adev->mman.aper_base_kaddr = ioremap_cache(adev->gmc.aper_base, 1696 adev->gmc.visible_vram_size); 1697 1698 else 1699 #endif 1700 adev->mman.aper_base_kaddr = ioremap_wc(adev->gmc.aper_base, 1701 adev->gmc.visible_vram_size); 1702 #endif 1703 1704 /* 1705 *The reserved vram for firmware must be pinned to the specified 1706 *place on the VRAM, so reserve it early. 1707 */ 1708 r = amdgpu_ttm_fw_reserve_vram_init(adev); 1709 if (r) { 1710 return r; 1711 } 1712 1713 /* 1714 * only NAVI10 and onwards ASIC support for IP discovery. 1715 * If IP discovery enabled, a block of memory should be 1716 * reserved for IP discovey. 1717 */ 1718 if (adev->mman.discovery_bin) { 1719 r = amdgpu_ttm_reserve_tmr(adev); 1720 if (r) 1721 return r; 1722 } 1723 1724 /* allocate memory as required for VGA 1725 * This is used for VGA emulation and pre-OS scanout buffers to 1726 * avoid display artifacts while transitioning between pre-OS 1727 * and driver. */ 1728 r = amdgpu_bo_create_kernel_at(adev, 0, adev->mman.stolen_vga_size, 1729 AMDGPU_GEM_DOMAIN_VRAM, 1730 &adev->mman.stolen_vga_memory, 1731 NULL); 1732 if (r) 1733 return r; 1734 r = amdgpu_bo_create_kernel_at(adev, adev->mman.stolen_vga_size, 1735 adev->mman.stolen_extended_size, 1736 AMDGPU_GEM_DOMAIN_VRAM, 1737 &adev->mman.stolen_extended_memory, 1738 NULL); 1739 if (r) 1740 return r; 1741 r = amdgpu_bo_create_kernel_at(adev, adev->mman.stolen_reserved_offset, 1742 adev->mman.stolen_reserved_size, 1743 AMDGPU_GEM_DOMAIN_VRAM, 1744 &adev->mman.stolen_reserved_memory, 1745 NULL); 1746 if (r) 1747 return r; 1748 1749 DRM_INFO("amdgpu: %uM of VRAM memory ready\n", 1750 (unsigned) (adev->gmc.real_vram_size / (1024 * 1024))); 1751 1752 /* Compute GTT size, either bsaed on 3/4th the size of RAM size 1753 * or whatever the user passed on module init */ 1754 if (amdgpu_gtt_size == -1) { 1755 struct sysinfo si; 1756 1757 si_meminfo(&si); 1758 gtt_size = min(max((AMDGPU_DEFAULT_GTT_SIZE_MB << 20), 1759 adev->gmc.mc_vram_size), 1760 ((uint64_t)si.totalram * si.mem_unit * 3/4)); 1761 } 1762 else 1763 gtt_size = (uint64_t)amdgpu_gtt_size << 20; 1764 1765 /* Initialize GTT memory pool */ 1766 r = amdgpu_gtt_mgr_init(adev, gtt_size); 1767 if (r) { 1768 DRM_ERROR("Failed initializing GTT heap.\n"); 1769 return r; 1770 } 1771 DRM_INFO("amdgpu: %uM of GTT memory ready.\n", 1772 (unsigned)(gtt_size / (1024 * 1024))); 1773 1774 /* Initialize preemptible memory pool */ 1775 r = amdgpu_preempt_mgr_init(adev); 1776 if (r) { 1777 DRM_ERROR("Failed initializing PREEMPT heap.\n"); 1778 return r; 1779 } 1780 1781 /* Initialize various on-chip memory pools */ 1782 r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_GDS, adev->gds.gds_size); 1783 if (r) { 1784 DRM_ERROR("Failed initializing GDS heap.\n"); 1785 return r; 1786 } 1787 1788 r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_GWS, adev->gds.gws_size); 1789 if (r) { 1790 DRM_ERROR("Failed initializing gws heap.\n"); 1791 return r; 1792 } 1793 1794 r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_OA, adev->gds.oa_size); 1795 if (r) { 1796 DRM_ERROR("Failed initializing oa heap.\n"); 1797 return r; 1798 } 1799 1800 return 0; 1801 } 1802 1803 /* 1804 * amdgpu_ttm_fini - De-initialize the TTM memory pools 1805 */ 1806 void amdgpu_ttm_fini(struct amdgpu_device *adev) 1807 { 1808 int idx; 1809 if (!adev->mman.initialized) 1810 return; 1811 1812 amdgpu_ttm_training_reserve_vram_fini(adev); 1813 /* return the stolen vga memory back to VRAM */ 1814 amdgpu_bo_free_kernel(&adev->mman.stolen_vga_memory, NULL, NULL); 1815 amdgpu_bo_free_kernel(&adev->mman.stolen_extended_memory, NULL, NULL); 1816 /* return the IP Discovery TMR memory back to VRAM */ 1817 amdgpu_bo_free_kernel(&adev->mman.discovery_memory, NULL, NULL); 1818 if (adev->mman.stolen_reserved_size) 1819 amdgpu_bo_free_kernel(&adev->mman.stolen_reserved_memory, 1820 NULL, NULL); 1821 amdgpu_ttm_fw_reserve_vram_fini(adev); 1822 1823 if (drm_dev_enter(adev_to_drm(adev), &idx)) { 1824 1825 if (adev->mman.aper_base_kaddr) 1826 iounmap(adev->mman.aper_base_kaddr); 1827 adev->mman.aper_base_kaddr = NULL; 1828 1829 drm_dev_exit(idx); 1830 } 1831 1832 amdgpu_vram_mgr_fini(adev); 1833 amdgpu_gtt_mgr_fini(adev); 1834 amdgpu_preempt_mgr_fini(adev); 1835 ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_GDS); 1836 ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_GWS); 1837 ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_OA); 1838 ttm_device_fini(&adev->mman.bdev); 1839 adev->mman.initialized = false; 1840 DRM_INFO("amdgpu: ttm finalized\n"); 1841 } 1842 1843 /** 1844 * amdgpu_ttm_set_buffer_funcs_status - enable/disable use of buffer functions 1845 * 1846 * @adev: amdgpu_device pointer 1847 * @enable: true when we can use buffer functions. 1848 * 1849 * Enable/disable use of buffer functions during suspend/resume. This should 1850 * only be called at bootup or when userspace isn't running. 1851 */ 1852 void amdgpu_ttm_set_buffer_funcs_status(struct amdgpu_device *adev, bool enable) 1853 { 1854 struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM); 1855 uint64_t size; 1856 int r; 1857 1858 if (!adev->mman.initialized || amdgpu_in_reset(adev) || 1859 adev->mman.buffer_funcs_enabled == enable) 1860 return; 1861 1862 if (enable) { 1863 struct amdgpu_ring *ring; 1864 struct drm_gpu_scheduler *sched; 1865 1866 ring = adev->mman.buffer_funcs_ring; 1867 sched = &ring->sched; 1868 r = drm_sched_entity_init(&adev->mman.entity, 1869 DRM_SCHED_PRIORITY_KERNEL, &sched, 1870 1, NULL); 1871 if (r) { 1872 DRM_ERROR("Failed setting up TTM BO move entity (%d)\n", 1873 r); 1874 return; 1875 } 1876 } else { 1877 drm_sched_entity_destroy(&adev->mman.entity); 1878 dma_fence_put(man->move); 1879 man->move = NULL; 1880 } 1881 1882 /* this just adjusts TTM size idea, which sets lpfn to the correct value */ 1883 if (enable) 1884 size = adev->gmc.real_vram_size; 1885 else 1886 size = adev->gmc.visible_vram_size; 1887 man->size = size >> PAGE_SHIFT; 1888 adev->mman.buffer_funcs_enabled = enable; 1889 } 1890 1891 int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset, 1892 uint64_t dst_offset, uint32_t byte_count, 1893 struct dma_resv *resv, 1894 struct dma_fence **fence, bool direct_submit, 1895 bool vm_needs_flush, bool tmz) 1896 { 1897 enum amdgpu_ib_pool_type pool = direct_submit ? AMDGPU_IB_POOL_DIRECT : 1898 AMDGPU_IB_POOL_DELAYED; 1899 struct amdgpu_device *adev = ring->adev; 1900 struct amdgpu_job *job; 1901 1902 uint32_t max_bytes; 1903 unsigned num_loops, num_dw; 1904 unsigned i; 1905 int r; 1906 1907 if (!direct_submit && !ring->sched.ready) { 1908 DRM_ERROR("Trying to move memory with ring turned off.\n"); 1909 return -EINVAL; 1910 } 1911 1912 max_bytes = adev->mman.buffer_funcs->copy_max_bytes; 1913 num_loops = DIV_ROUND_UP(byte_count, max_bytes); 1914 num_dw = ALIGN(num_loops * adev->mman.buffer_funcs->copy_num_dw, 8); 1915 1916 r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, pool, &job); 1917 if (r) 1918 return r; 1919 1920 if (vm_needs_flush) { 1921 job->vm_pd_addr = amdgpu_gmc_pd_addr(adev->gmc.pdb0_bo ? 1922 adev->gmc.pdb0_bo : adev->gart.bo); 1923 job->vm_needs_flush = true; 1924 } 1925 if (resv) { 1926 r = amdgpu_sync_resv(adev, &job->sync, resv, 1927 AMDGPU_SYNC_ALWAYS, 1928 AMDGPU_FENCE_OWNER_UNDEFINED); 1929 if (r) { 1930 DRM_ERROR("sync failed (%d).\n", r); 1931 goto error_free; 1932 } 1933 } 1934 1935 for (i = 0; i < num_loops; i++) { 1936 uint32_t cur_size_in_bytes = min(byte_count, max_bytes); 1937 1938 amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_offset, 1939 dst_offset, cur_size_in_bytes, tmz); 1940 1941 src_offset += cur_size_in_bytes; 1942 dst_offset += cur_size_in_bytes; 1943 byte_count -= cur_size_in_bytes; 1944 } 1945 1946 amdgpu_ring_pad_ib(ring, &job->ibs[0]); 1947 WARN_ON(job->ibs[0].length_dw > num_dw); 1948 if (direct_submit) 1949 r = amdgpu_job_submit_direct(job, ring, fence); 1950 else 1951 r = amdgpu_job_submit(job, &adev->mman.entity, 1952 AMDGPU_FENCE_OWNER_UNDEFINED, fence); 1953 if (r) 1954 goto error_free; 1955 1956 return r; 1957 1958 error_free: 1959 amdgpu_job_free(job); 1960 DRM_ERROR("Error scheduling IBs (%d)\n", r); 1961 return r; 1962 } 1963 1964 int amdgpu_fill_buffer(struct amdgpu_bo *bo, 1965 uint32_t src_data, 1966 struct dma_resv *resv, 1967 struct dma_fence **fence) 1968 { 1969 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); 1970 uint32_t max_bytes = adev->mman.buffer_funcs->fill_max_bytes; 1971 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring; 1972 1973 struct amdgpu_res_cursor cursor; 1974 unsigned int num_loops, num_dw; 1975 uint64_t num_bytes; 1976 1977 struct amdgpu_job *job; 1978 int r; 1979 1980 if (!adev->mman.buffer_funcs_enabled) { 1981 DRM_ERROR("Trying to clear memory with ring turned off.\n"); 1982 return -EINVAL; 1983 } 1984 1985 if (bo->tbo.resource->mem_type == AMDGPU_PL_PREEMPT) { 1986 DRM_ERROR("Trying to clear preemptible memory.\n"); 1987 return -EINVAL; 1988 } 1989 1990 if (bo->tbo.resource->mem_type == TTM_PL_TT) { 1991 r = amdgpu_ttm_alloc_gart(&bo->tbo); 1992 if (r) 1993 return r; 1994 } 1995 1996 num_bytes = bo->tbo.resource->num_pages << PAGE_SHIFT; 1997 num_loops = 0; 1998 1999 amdgpu_res_first(bo->tbo.resource, 0, num_bytes, &cursor); 2000 while (cursor.remaining) { 2001 num_loops += DIV_ROUND_UP_ULL(cursor.size, max_bytes); 2002 amdgpu_res_next(&cursor, cursor.size); 2003 } 2004 num_dw = num_loops * adev->mman.buffer_funcs->fill_num_dw; 2005 2006 /* for IB padding */ 2007 num_dw += 64; 2008 2009 r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, AMDGPU_IB_POOL_DELAYED, 2010 &job); 2011 if (r) 2012 return r; 2013 2014 if (resv) { 2015 r = amdgpu_sync_resv(adev, &job->sync, resv, 2016 AMDGPU_SYNC_ALWAYS, 2017 AMDGPU_FENCE_OWNER_UNDEFINED); 2018 if (r) { 2019 DRM_ERROR("sync failed (%d).\n", r); 2020 goto error_free; 2021 } 2022 } 2023 2024 amdgpu_res_first(bo->tbo.resource, 0, num_bytes, &cursor); 2025 while (cursor.remaining) { 2026 uint32_t cur_size = min_t(uint64_t, cursor.size, max_bytes); 2027 uint64_t dst_addr = cursor.start; 2028 2029 dst_addr += amdgpu_ttm_domain_start(adev, 2030 bo->tbo.resource->mem_type); 2031 amdgpu_emit_fill_buffer(adev, &job->ibs[0], src_data, dst_addr, 2032 cur_size); 2033 2034 amdgpu_res_next(&cursor, cur_size); 2035 } 2036 2037 amdgpu_ring_pad_ib(ring, &job->ibs[0]); 2038 WARN_ON(job->ibs[0].length_dw > num_dw); 2039 r = amdgpu_job_submit(job, &adev->mman.entity, 2040 AMDGPU_FENCE_OWNER_UNDEFINED, fence); 2041 if (r) 2042 goto error_free; 2043 2044 return 0; 2045 2046 error_free: 2047 amdgpu_job_free(job); 2048 return r; 2049 } 2050 2051 /** 2052 * amdgpu_ttm_evict_resources - evict memory buffers 2053 * @adev: amdgpu device object 2054 * @mem_type: evicted BO's memory type 2055 * 2056 * Evicts all @mem_type buffers on the lru list of the memory type. 2057 * 2058 * Returns: 2059 * 0 for success or a negative error code on failure. 2060 */ 2061 int amdgpu_ttm_evict_resources(struct amdgpu_device *adev, int mem_type) 2062 { 2063 struct ttm_resource_manager *man; 2064 2065 switch (mem_type) { 2066 case TTM_PL_VRAM: 2067 case TTM_PL_TT: 2068 case AMDGPU_PL_GWS: 2069 case AMDGPU_PL_GDS: 2070 case AMDGPU_PL_OA: 2071 man = ttm_manager_type(&adev->mman.bdev, mem_type); 2072 break; 2073 default: 2074 DRM_ERROR("Trying to evict invalid memory type\n"); 2075 return -EINVAL; 2076 } 2077 2078 return ttm_resource_manager_evict_all(&adev->mman.bdev, man); 2079 } 2080 2081 #if defined(CONFIG_DEBUG_FS) 2082 2083 static int amdgpu_mm_vram_table_show(struct seq_file *m, void *unused) 2084 { 2085 struct amdgpu_device *adev = (struct amdgpu_device *)m->private; 2086 struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev, 2087 TTM_PL_VRAM); 2088 struct drm_printer p = drm_seq_file_printer(m); 2089 2090 man->func->debug(man, &p); 2091 return 0; 2092 } 2093 2094 static int amdgpu_ttm_page_pool_show(struct seq_file *m, void *unused) 2095 { 2096 struct amdgpu_device *adev = (struct amdgpu_device *)m->private; 2097 2098 return ttm_pool_debugfs(&adev->mman.bdev.pool, m); 2099 } 2100 2101 static int amdgpu_mm_tt_table_show(struct seq_file *m, void *unused) 2102 { 2103 struct amdgpu_device *adev = (struct amdgpu_device *)m->private; 2104 struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev, 2105 TTM_PL_TT); 2106 struct drm_printer p = drm_seq_file_printer(m); 2107 2108 man->func->debug(man, &p); 2109 return 0; 2110 } 2111 2112 static int amdgpu_mm_gds_table_show(struct seq_file *m, void *unused) 2113 { 2114 struct amdgpu_device *adev = (struct amdgpu_device *)m->private; 2115 struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev, 2116 AMDGPU_PL_GDS); 2117 struct drm_printer p = drm_seq_file_printer(m); 2118 2119 man->func->debug(man, &p); 2120 return 0; 2121 } 2122 2123 static int amdgpu_mm_gws_table_show(struct seq_file *m, void *unused) 2124 { 2125 struct amdgpu_device *adev = (struct amdgpu_device *)m->private; 2126 struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev, 2127 AMDGPU_PL_GWS); 2128 struct drm_printer p = drm_seq_file_printer(m); 2129 2130 man->func->debug(man, &p); 2131 return 0; 2132 } 2133 2134 static int amdgpu_mm_oa_table_show(struct seq_file *m, void *unused) 2135 { 2136 struct amdgpu_device *adev = (struct amdgpu_device *)m->private; 2137 struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev, 2138 AMDGPU_PL_OA); 2139 struct drm_printer p = drm_seq_file_printer(m); 2140 2141 man->func->debug(man, &p); 2142 return 0; 2143 } 2144 2145 DEFINE_SHOW_ATTRIBUTE(amdgpu_mm_vram_table); 2146 DEFINE_SHOW_ATTRIBUTE(amdgpu_mm_tt_table); 2147 DEFINE_SHOW_ATTRIBUTE(amdgpu_mm_gds_table); 2148 DEFINE_SHOW_ATTRIBUTE(amdgpu_mm_gws_table); 2149 DEFINE_SHOW_ATTRIBUTE(amdgpu_mm_oa_table); 2150 DEFINE_SHOW_ATTRIBUTE(amdgpu_ttm_page_pool); 2151 2152 /* 2153 * amdgpu_ttm_vram_read - Linear read access to VRAM 2154 * 2155 * Accesses VRAM via MMIO for debugging purposes. 2156 */ 2157 static ssize_t amdgpu_ttm_vram_read(struct file *f, char __user *buf, 2158 size_t size, loff_t *pos) 2159 { 2160 struct amdgpu_device *adev = file_inode(f)->i_private; 2161 ssize_t result = 0; 2162 2163 if (size & 0x3 || *pos & 0x3) 2164 return -EINVAL; 2165 2166 if (*pos >= adev->gmc.mc_vram_size) 2167 return -ENXIO; 2168 2169 size = min(size, (size_t)(adev->gmc.mc_vram_size - *pos)); 2170 while (size) { 2171 size_t bytes = min(size, AMDGPU_TTM_VRAM_MAX_DW_READ * 4); 2172 uint32_t value[AMDGPU_TTM_VRAM_MAX_DW_READ]; 2173 2174 amdgpu_device_vram_access(adev, *pos, value, bytes, false); 2175 if (copy_to_user(buf, value, bytes)) 2176 return -EFAULT; 2177 2178 result += bytes; 2179 buf += bytes; 2180 *pos += bytes; 2181 size -= bytes; 2182 } 2183 2184 return result; 2185 } 2186 2187 /* 2188 * amdgpu_ttm_vram_write - Linear write access to VRAM 2189 * 2190 * Accesses VRAM via MMIO for debugging purposes. 2191 */ 2192 static ssize_t amdgpu_ttm_vram_write(struct file *f, const char __user *buf, 2193 size_t size, loff_t *pos) 2194 { 2195 struct amdgpu_device *adev = file_inode(f)->i_private; 2196 ssize_t result = 0; 2197 int r; 2198 2199 if (size & 0x3 || *pos & 0x3) 2200 return -EINVAL; 2201 2202 if (*pos >= adev->gmc.mc_vram_size) 2203 return -ENXIO; 2204 2205 while (size) { 2206 uint32_t value; 2207 2208 if (*pos >= adev->gmc.mc_vram_size) 2209 return result; 2210 2211 r = get_user(value, (uint32_t *)buf); 2212 if (r) 2213 return r; 2214 2215 amdgpu_device_mm_access(adev, *pos, &value, 4, true); 2216 2217 result += 4; 2218 buf += 4; 2219 *pos += 4; 2220 size -= 4; 2221 } 2222 2223 return result; 2224 } 2225 2226 static const struct file_operations amdgpu_ttm_vram_fops = { 2227 .owner = THIS_MODULE, 2228 .read = amdgpu_ttm_vram_read, 2229 .write = amdgpu_ttm_vram_write, 2230 .llseek = default_llseek, 2231 }; 2232 2233 /* 2234 * amdgpu_iomem_read - Virtual read access to GPU mapped memory 2235 * 2236 * This function is used to read memory that has been mapped to the 2237 * GPU and the known addresses are not physical addresses but instead 2238 * bus addresses (e.g., what you'd put in an IB or ring buffer). 2239 */ 2240 static ssize_t amdgpu_iomem_read(struct file *f, char __user *buf, 2241 size_t size, loff_t *pos) 2242 { 2243 struct amdgpu_device *adev = file_inode(f)->i_private; 2244 struct iommu_domain *dom; 2245 ssize_t result = 0; 2246 int r; 2247 2248 /* retrieve the IOMMU domain if any for this device */ 2249 dom = iommu_get_domain_for_dev(adev->dev); 2250 2251 while (size) { 2252 phys_addr_t addr = *pos & PAGE_MASK; 2253 loff_t off = *pos & ~PAGE_MASK; 2254 size_t bytes = PAGE_SIZE - off; 2255 unsigned long pfn; 2256 struct page *p; 2257 void *ptr; 2258 2259 bytes = bytes < size ? bytes : size; 2260 2261 /* Translate the bus address to a physical address. If 2262 * the domain is NULL it means there is no IOMMU active 2263 * and the address translation is the identity 2264 */ 2265 addr = dom ? iommu_iova_to_phys(dom, addr) : addr; 2266 2267 pfn = addr >> PAGE_SHIFT; 2268 if (!pfn_valid(pfn)) 2269 return -EPERM; 2270 2271 p = pfn_to_page(pfn); 2272 if (p->mapping != adev->mman.bdev.dev_mapping) 2273 return -EPERM; 2274 2275 ptr = kmap(p); 2276 r = copy_to_user(buf, ptr + off, bytes); 2277 kunmap(p); 2278 if (r) 2279 return -EFAULT; 2280 2281 size -= bytes; 2282 *pos += bytes; 2283 result += bytes; 2284 } 2285 2286 return result; 2287 } 2288 2289 /* 2290 * amdgpu_iomem_write - Virtual write access to GPU mapped memory 2291 * 2292 * This function is used to write memory that has been mapped to the 2293 * GPU and the known addresses are not physical addresses but instead 2294 * bus addresses (e.g., what you'd put in an IB or ring buffer). 2295 */ 2296 static ssize_t amdgpu_iomem_write(struct file *f, const char __user *buf, 2297 size_t size, loff_t *pos) 2298 { 2299 struct amdgpu_device *adev = file_inode(f)->i_private; 2300 struct iommu_domain *dom; 2301 ssize_t result = 0; 2302 int r; 2303 2304 dom = iommu_get_domain_for_dev(adev->dev); 2305 2306 while (size) { 2307 phys_addr_t addr = *pos & PAGE_MASK; 2308 loff_t off = *pos & ~PAGE_MASK; 2309 size_t bytes = PAGE_SIZE - off; 2310 unsigned long pfn; 2311 struct page *p; 2312 void *ptr; 2313 2314 bytes = bytes < size ? bytes : size; 2315 2316 addr = dom ? iommu_iova_to_phys(dom, addr) : addr; 2317 2318 pfn = addr >> PAGE_SHIFT; 2319 if (!pfn_valid(pfn)) 2320 return -EPERM; 2321 2322 p = pfn_to_page(pfn); 2323 if (p->mapping != adev->mman.bdev.dev_mapping) 2324 return -EPERM; 2325 2326 ptr = kmap(p); 2327 r = copy_from_user(ptr + off, buf, bytes); 2328 kunmap(p); 2329 if (r) 2330 return -EFAULT; 2331 2332 size -= bytes; 2333 *pos += bytes; 2334 result += bytes; 2335 } 2336 2337 return result; 2338 } 2339 2340 static const struct file_operations amdgpu_ttm_iomem_fops = { 2341 .owner = THIS_MODULE, 2342 .read = amdgpu_iomem_read, 2343 .write = amdgpu_iomem_write, 2344 .llseek = default_llseek 2345 }; 2346 2347 #endif 2348 2349 void amdgpu_ttm_debugfs_init(struct amdgpu_device *adev) 2350 { 2351 #if defined(CONFIG_DEBUG_FS) 2352 struct drm_minor *minor = adev_to_drm(adev)->primary; 2353 struct dentry *root = minor->debugfs_root; 2354 2355 debugfs_create_file_size("amdgpu_vram", 0444, root, adev, 2356 &amdgpu_ttm_vram_fops, adev->gmc.mc_vram_size); 2357 debugfs_create_file("amdgpu_iomem", 0444, root, adev, 2358 &amdgpu_ttm_iomem_fops); 2359 debugfs_create_file("amdgpu_vram_mm", 0444, root, adev, 2360 &amdgpu_mm_vram_table_fops); 2361 debugfs_create_file("amdgpu_gtt_mm", 0444, root, adev, 2362 &amdgpu_mm_tt_table_fops); 2363 debugfs_create_file("amdgpu_gds_mm", 0444, root, adev, 2364 &amdgpu_mm_gds_table_fops); 2365 debugfs_create_file("amdgpu_gws_mm", 0444, root, adev, 2366 &amdgpu_mm_gws_table_fops); 2367 debugfs_create_file("amdgpu_oa_mm", 0444, root, adev, 2368 &amdgpu_mm_oa_table_fops); 2369 debugfs_create_file("ttm_page_pool", 0444, root, adev, 2370 &amdgpu_ttm_page_pool_fops); 2371 #endif 2372 } 2373