| 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/debugfs.h> |
| 34 | #include <linux/dma-mapping.h> |
| 35 | #include <linux/pagemap.h> |
| 36 | #include <linux/pci.h> |
| 37 | #include <linux/seq_file.h> |
| 38 | #include <linux/slab.h> |
| 39 | #include <linux/swap.h> |
| 40 | |
| 41 | #include <drm/drm_device.h> |
| 42 | #include <drm/drm_file.h> |
| 43 | #include <drm/drm_prime.h> |
| 44 | #include <drm/radeon_drm.h> |
| 45 | #include <drm/ttm/ttm_bo.h> |
| 46 | #include <drm/ttm/ttm_placement.h> |
| 47 | #include <drm/ttm/ttm_range_manager.h> |
| 48 | #include <drm/ttm/ttm_tt.h> |
| 49 | |
| 50 | #include "radeon_reg.h" |
| 51 | #include "radeon.h" |
| 52 | #include "radeon_ttm.h" |
| 53 | |
| 54 | static void radeon_ttm_debugfs_init(struct radeon_device *rdev); |
| 55 | |
| 56 | static int radeon_ttm_tt_bind(struct ttm_device *bdev, struct ttm_tt *ttm, |
| 57 | struct ttm_resource *bo_mem); |
| 58 | static void radeon_ttm_tt_unbind(struct ttm_device *bdev, struct ttm_tt *ttm); |
| 59 | |
| 60 | struct radeon_device *radeon_get_rdev(struct ttm_device *bdev) |
| 61 | { |
| 62 | struct radeon_mman *mman; |
| 63 | struct radeon_device *rdev; |
| 64 | |
| 65 | mman = container_of(bdev, struct radeon_mman, bdev); |
| 66 | rdev = container_of(mman, struct radeon_device, mman); |
| 67 | return rdev; |
| 68 | } |
| 69 | |
| 70 | static int radeon_ttm_init_vram(struct radeon_device *rdev) |
| 71 | { |
| 72 | return ttm_range_man_init(bdev: &rdev->mman.bdev, TTM_PL_VRAM, |
| 73 | use_tt: false, p_size: rdev->mc.real_vram_size >> PAGE_SHIFT); |
| 74 | } |
| 75 | |
| 76 | static int radeon_ttm_init_gtt(struct radeon_device *rdev) |
| 77 | { |
| 78 | return ttm_range_man_init(bdev: &rdev->mman.bdev, TTM_PL_TT, |
| 79 | use_tt: true, p_size: rdev->mc.gtt_size >> PAGE_SHIFT); |
| 80 | } |
| 81 | |
| 82 | static void radeon_evict_flags(struct ttm_buffer_object *bo, |
| 83 | struct ttm_placement *placement) |
| 84 | { |
| 85 | static const struct ttm_place placements = { |
| 86 | .fpfn = 0, |
| 87 | .lpfn = 0, |
| 88 | .mem_type = TTM_PL_SYSTEM, |
| 89 | .flags = 0 |
| 90 | }; |
| 91 | |
| 92 | struct radeon_bo *rbo; |
| 93 | |
| 94 | if (!radeon_ttm_bo_is_radeon_bo(bo)) { |
| 95 | placement->placement = &placements; |
| 96 | placement->num_placement = 1; |
| 97 | return; |
| 98 | } |
| 99 | rbo = container_of(bo, struct radeon_bo, tbo); |
| 100 | switch (bo->resource->mem_type) { |
| 101 | case TTM_PL_VRAM: |
| 102 | if (rbo->rdev->ring[radeon_copy_ring_index(rbo->rdev)].ready == false) |
| 103 | radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU); |
| 104 | else if (rbo->rdev->mc.visible_vram_size < rbo->rdev->mc.real_vram_size && |
| 105 | bo->resource->start < (rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT)) { |
| 106 | unsigned fpfn = rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT; |
| 107 | int i; |
| 108 | |
| 109 | /* Try evicting to the CPU inaccessible part of VRAM |
| 110 | * first, but only set GTT as busy placement, so this |
| 111 | * BO will be evicted to GTT rather than causing other |
| 112 | * BOs to be evicted from VRAM |
| 113 | */ |
| 114 | radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM | |
| 115 | RADEON_GEM_DOMAIN_GTT); |
| 116 | for (i = 0; i < rbo->placement.num_placement; i++) { |
| 117 | if (rbo->placements[i].mem_type == TTM_PL_VRAM) { |
| 118 | if (rbo->placements[i].fpfn < fpfn) |
| 119 | rbo->placements[i].fpfn = fpfn; |
| 120 | rbo->placements[0].flags |= TTM_PL_FLAG_DESIRED; |
| 121 | } |
| 122 | } |
| 123 | } else |
| 124 | radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT); |
| 125 | break; |
| 126 | case TTM_PL_TT: |
| 127 | default: |
| 128 | radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU); |
| 129 | } |
| 130 | *placement = rbo->placement; |
| 131 | } |
| 132 | |
| 133 | static int radeon_move_blit(struct ttm_buffer_object *bo, |
| 134 | bool evict, |
| 135 | struct ttm_resource *new_mem, |
| 136 | struct ttm_resource *old_mem) |
| 137 | { |
| 138 | struct radeon_device *rdev; |
| 139 | uint64_t old_start, new_start; |
| 140 | struct radeon_fence *fence; |
| 141 | unsigned num_pages; |
| 142 | int r, ridx; |
| 143 | |
| 144 | rdev = radeon_get_rdev(bdev: bo->bdev); |
| 145 | ridx = radeon_copy_ring_index(rdev); |
| 146 | old_start = (u64)old_mem->start << PAGE_SHIFT; |
| 147 | new_start = (u64)new_mem->start << PAGE_SHIFT; |
| 148 | |
| 149 | switch (old_mem->mem_type) { |
| 150 | case TTM_PL_VRAM: |
| 151 | old_start += rdev->mc.vram_start; |
| 152 | break; |
| 153 | case TTM_PL_TT: |
| 154 | old_start += rdev->mc.gtt_start; |
| 155 | break; |
| 156 | default: |
| 157 | DRM_ERROR("Unknown placement %d\n" , old_mem->mem_type); |
| 158 | return -EINVAL; |
| 159 | } |
| 160 | switch (new_mem->mem_type) { |
| 161 | case TTM_PL_VRAM: |
| 162 | new_start += rdev->mc.vram_start; |
| 163 | break; |
| 164 | case TTM_PL_TT: |
| 165 | new_start += rdev->mc.gtt_start; |
| 166 | break; |
| 167 | default: |
| 168 | DRM_ERROR("Unknown placement %d\n" , old_mem->mem_type); |
| 169 | return -EINVAL; |
| 170 | } |
| 171 | if (!rdev->ring[ridx].ready) { |
| 172 | DRM_ERROR("Trying to move memory with ring turned off.\n" ); |
| 173 | return -EINVAL; |
| 174 | } |
| 175 | |
| 176 | BUILD_BUG_ON((PAGE_SIZE % RADEON_GPU_PAGE_SIZE) != 0); |
| 177 | |
| 178 | num_pages = PFN_UP(new_mem->size) * (PAGE_SIZE / RADEON_GPU_PAGE_SIZE); |
| 179 | fence = radeon_copy(rdev, old_start, new_start, num_pages, bo->base.resv); |
| 180 | if (IS_ERR(ptr: fence)) |
| 181 | return PTR_ERR(ptr: fence); |
| 182 | |
| 183 | r = ttm_bo_move_accel_cleanup(bo, fence: &fence->base, evict, pipeline: false, new_mem); |
| 184 | radeon_fence_unref(fence: &fence); |
| 185 | return r; |
| 186 | } |
| 187 | |
| 188 | static int radeon_bo_move(struct ttm_buffer_object *bo, bool evict, |
| 189 | struct ttm_operation_ctx *ctx, |
| 190 | struct ttm_resource *new_mem, |
| 191 | struct ttm_place *hop) |
| 192 | { |
| 193 | struct ttm_resource *old_mem = bo->resource; |
| 194 | struct radeon_device *rdev; |
| 195 | int r; |
| 196 | |
| 197 | if (new_mem->mem_type == TTM_PL_TT) { |
| 198 | r = radeon_ttm_tt_bind(bdev: bo->bdev, ttm: bo->ttm, bo_mem: new_mem); |
| 199 | if (r) |
| 200 | return r; |
| 201 | } |
| 202 | |
| 203 | r = ttm_bo_wait_ctx(bo, ctx); |
| 204 | if (r) |
| 205 | return r; |
| 206 | |
| 207 | rdev = radeon_get_rdev(bdev: bo->bdev); |
| 208 | if (!old_mem || (old_mem->mem_type == TTM_PL_SYSTEM && |
| 209 | bo->ttm == NULL)) { |
| 210 | ttm_bo_move_null(bo, new_mem); |
| 211 | goto out; |
| 212 | } |
| 213 | if (old_mem->mem_type == TTM_PL_SYSTEM && |
| 214 | new_mem->mem_type == TTM_PL_TT) { |
| 215 | ttm_bo_move_null(bo, new_mem); |
| 216 | goto out; |
| 217 | } |
| 218 | |
| 219 | if (old_mem->mem_type == TTM_PL_TT && |
| 220 | new_mem->mem_type == TTM_PL_SYSTEM) { |
| 221 | radeon_ttm_tt_unbind(bdev: bo->bdev, ttm: bo->ttm); |
| 222 | ttm_bo_move_null(bo, new_mem); |
| 223 | goto out; |
| 224 | } |
| 225 | if (rdev->ring[radeon_copy_ring_index(rdev)].ready && |
| 226 | rdev->asic->copy.copy != NULL) { |
| 227 | if ((old_mem->mem_type == TTM_PL_SYSTEM && |
| 228 | new_mem->mem_type == TTM_PL_VRAM) || |
| 229 | (old_mem->mem_type == TTM_PL_VRAM && |
| 230 | new_mem->mem_type == TTM_PL_SYSTEM)) { |
| 231 | hop->fpfn = 0; |
| 232 | hop->lpfn = 0; |
| 233 | hop->mem_type = TTM_PL_TT; |
| 234 | hop->flags = 0; |
| 235 | return -EMULTIHOP; |
| 236 | } |
| 237 | |
| 238 | r = radeon_move_blit(bo, evict, new_mem, old_mem); |
| 239 | } else { |
| 240 | r = -ENODEV; |
| 241 | } |
| 242 | |
| 243 | if (r) { |
| 244 | r = ttm_bo_move_memcpy(bo, ctx, new_mem); |
| 245 | if (r) |
| 246 | return r; |
| 247 | } |
| 248 | |
| 249 | out: |
| 250 | /* update statistics */ |
| 251 | atomic64_add(i: bo->base.size, v: &rdev->num_bytes_moved); |
| 252 | radeon_bo_move_notify(bo); |
| 253 | return 0; |
| 254 | } |
| 255 | |
| 256 | static int radeon_ttm_io_mem_reserve(struct ttm_device *bdev, struct ttm_resource *mem) |
| 257 | { |
| 258 | struct radeon_device *rdev = radeon_get_rdev(bdev); |
| 259 | size_t bus_size = (size_t)mem->size; |
| 260 | |
| 261 | switch (mem->mem_type) { |
| 262 | case TTM_PL_SYSTEM: |
| 263 | /* system memory */ |
| 264 | return 0; |
| 265 | case TTM_PL_TT: |
| 266 | #if IS_ENABLED(CONFIG_AGP) |
| 267 | if (rdev->flags & RADEON_IS_AGP) { |
| 268 | /* RADEON_IS_AGP is set only if AGP is active */ |
| 269 | mem->bus.offset = (mem->start << PAGE_SHIFT) + |
| 270 | rdev->mc.agp_base; |
| 271 | mem->bus.is_iomem = !rdev->agp->cant_use_aperture; |
| 272 | mem->bus.caching = ttm_write_combined; |
| 273 | } |
| 274 | #endif |
| 275 | break; |
| 276 | case TTM_PL_VRAM: |
| 277 | mem->bus.offset = mem->start << PAGE_SHIFT; |
| 278 | /* check if it's visible */ |
| 279 | if ((mem->bus.offset + bus_size) > rdev->mc.visible_vram_size) |
| 280 | return -EINVAL; |
| 281 | mem->bus.offset += rdev->mc.aper_base; |
| 282 | mem->bus.is_iomem = true; |
| 283 | mem->bus.caching = ttm_write_combined; |
| 284 | #ifdef __alpha__ |
| 285 | /* |
| 286 | * Alpha: use bus.addr to hold the ioremap() return, |
| 287 | * so we can modify bus.base below. |
| 288 | */ |
| 289 | mem->bus.addr = ioremap_wc(mem->bus.offset, bus_size); |
| 290 | if (!mem->bus.addr) |
| 291 | return -ENOMEM; |
| 292 | |
| 293 | /* |
| 294 | * Alpha: Use just the bus offset plus |
| 295 | * the hose/domain memory base for bus.base. |
| 296 | * It then can be used to build PTEs for VRAM |
| 297 | * access, as done in ttm_bo_vm_fault(). |
| 298 | */ |
| 299 | mem->bus.offset = (mem->bus.offset & 0x0ffffffffUL) + |
| 300 | rdev->hose->dense_mem_base; |
| 301 | #endif |
| 302 | break; |
| 303 | default: |
| 304 | return -EINVAL; |
| 305 | } |
| 306 | return 0; |
| 307 | } |
| 308 | |
| 309 | /* |
| 310 | * TTM backend functions. |
| 311 | */ |
| 312 | struct radeon_ttm_tt { |
| 313 | struct ttm_tt ttm; |
| 314 | u64 offset; |
| 315 | |
| 316 | uint64_t userptr; |
| 317 | struct mm_struct *usermm; |
| 318 | uint32_t userflags; |
| 319 | bool bound; |
| 320 | }; |
| 321 | |
| 322 | /* prepare the sg table with the user pages */ |
| 323 | static int radeon_ttm_tt_pin_userptr(struct ttm_device *bdev, struct ttm_tt *ttm) |
| 324 | { |
| 325 | struct radeon_device *rdev = radeon_get_rdev(bdev); |
| 326 | struct radeon_ttm_tt *gtt = (void *)ttm; |
| 327 | unsigned pinned = 0; |
| 328 | int r; |
| 329 | |
| 330 | int write = !(gtt->userflags & RADEON_GEM_USERPTR_READONLY); |
| 331 | enum dma_data_direction direction = write ? |
| 332 | DMA_BIDIRECTIONAL : DMA_TO_DEVICE; |
| 333 | |
| 334 | if (current->mm != gtt->usermm) |
| 335 | return -EPERM; |
| 336 | |
| 337 | if (gtt->userflags & RADEON_GEM_USERPTR_ANONONLY) { |
| 338 | /* check that we only pin down anonymous memory |
| 339 | to prevent problems with writeback */ |
| 340 | unsigned long end = gtt->userptr + (u64)ttm->num_pages * PAGE_SIZE; |
| 341 | struct vm_area_struct *vma; |
| 342 | vma = find_vma(mm: gtt->usermm, addr: gtt->userptr); |
| 343 | if (!vma || vma->vm_file || vma->vm_end < end) |
| 344 | return -EPERM; |
| 345 | } |
| 346 | |
| 347 | do { |
| 348 | unsigned num_pages = ttm->num_pages - pinned; |
| 349 | uint64_t userptr = gtt->userptr + pinned * PAGE_SIZE; |
| 350 | struct page **pages = ttm->pages + pinned; |
| 351 | |
| 352 | r = get_user_pages(start: userptr, nr_pages: num_pages, gup_flags: write ? FOLL_WRITE : 0, |
| 353 | pages); |
| 354 | if (r < 0) |
| 355 | goto release_pages; |
| 356 | |
| 357 | pinned += r; |
| 358 | |
| 359 | } while (pinned < ttm->num_pages); |
| 360 | |
| 361 | r = sg_alloc_table_from_pages(sgt: ttm->sg, pages: ttm->pages, n_pages: ttm->num_pages, offset: 0, |
| 362 | size: (u64)ttm->num_pages << PAGE_SHIFT, |
| 363 | GFP_KERNEL); |
| 364 | if (r) |
| 365 | goto release_sg; |
| 366 | |
| 367 | r = dma_map_sgtable(dev: rdev->dev, sgt: ttm->sg, dir: direction, attrs: 0); |
| 368 | if (r) |
| 369 | goto release_sg; |
| 370 | |
| 371 | drm_prime_sg_to_dma_addr_array(sgt: ttm->sg, addrs: gtt->ttm.dma_address, |
| 372 | max_pages: ttm->num_pages); |
| 373 | |
| 374 | return 0; |
| 375 | |
| 376 | release_sg: |
| 377 | kfree(objp: ttm->sg); |
| 378 | |
| 379 | release_pages: |
| 380 | release_pages(ttm->pages, nr: pinned); |
| 381 | return r; |
| 382 | } |
| 383 | |
| 384 | static void radeon_ttm_tt_unpin_userptr(struct ttm_device *bdev, struct ttm_tt *ttm) |
| 385 | { |
| 386 | struct radeon_device *rdev = radeon_get_rdev(bdev); |
| 387 | struct radeon_ttm_tt *gtt = (void *)ttm; |
| 388 | struct sg_page_iter sg_iter; |
| 389 | |
| 390 | int write = !(gtt->userflags & RADEON_GEM_USERPTR_READONLY); |
| 391 | enum dma_data_direction direction = write ? |
| 392 | DMA_BIDIRECTIONAL : DMA_TO_DEVICE; |
| 393 | |
| 394 | /* double check that we don't free the table twice */ |
| 395 | if (!ttm->sg || !ttm->sg->sgl) |
| 396 | return; |
| 397 | |
| 398 | /* free the sg table and pages again */ |
| 399 | dma_unmap_sgtable(dev: rdev->dev, sgt: ttm->sg, dir: direction, attrs: 0); |
| 400 | |
| 401 | for_each_sgtable_page(ttm->sg, &sg_iter, 0) { |
| 402 | struct page *page = sg_page_iter_page(piter: &sg_iter); |
| 403 | if (!(gtt->userflags & RADEON_GEM_USERPTR_READONLY)) |
| 404 | set_page_dirty(page); |
| 405 | |
| 406 | mark_page_accessed(page); |
| 407 | put_page(page); |
| 408 | } |
| 409 | |
| 410 | sg_free_table(ttm->sg); |
| 411 | } |
| 412 | |
| 413 | static bool radeon_ttm_backend_is_bound(struct ttm_tt *ttm) |
| 414 | { |
| 415 | struct radeon_ttm_tt *gtt = (void*)ttm; |
| 416 | |
| 417 | return (gtt->bound); |
| 418 | } |
| 419 | |
| 420 | static int radeon_ttm_backend_bind(struct ttm_device *bdev, |
| 421 | struct ttm_tt *ttm, |
| 422 | struct ttm_resource *bo_mem) |
| 423 | { |
| 424 | struct radeon_ttm_tt *gtt = (void*)ttm; |
| 425 | struct radeon_device *rdev = radeon_get_rdev(bdev); |
| 426 | uint32_t flags = RADEON_GART_PAGE_VALID | RADEON_GART_PAGE_READ | |
| 427 | RADEON_GART_PAGE_WRITE; |
| 428 | int r; |
| 429 | |
| 430 | if (gtt->bound) |
| 431 | return 0; |
| 432 | |
| 433 | if (gtt->userptr) { |
| 434 | radeon_ttm_tt_pin_userptr(bdev, ttm); |
| 435 | flags &= ~RADEON_GART_PAGE_WRITE; |
| 436 | } |
| 437 | |
| 438 | gtt->offset = (unsigned long)(bo_mem->start << PAGE_SHIFT); |
| 439 | if (!ttm->num_pages) { |
| 440 | WARN(1, "nothing to bind %u pages for mreg %p back %p!\n" , |
| 441 | ttm->num_pages, bo_mem, ttm); |
| 442 | } |
| 443 | if (ttm->caching == ttm_cached) |
| 444 | flags |= RADEON_GART_PAGE_SNOOP; |
| 445 | r = radeon_gart_bind(rdev, offset: gtt->offset, pages: ttm->num_pages, |
| 446 | pagelist: ttm->pages, dma_addr: gtt->ttm.dma_address, flags); |
| 447 | if (r) { |
| 448 | DRM_ERROR("failed to bind %u pages at 0x%08X\n" , |
| 449 | ttm->num_pages, (unsigned)gtt->offset); |
| 450 | return r; |
| 451 | } |
| 452 | gtt->bound = true; |
| 453 | return 0; |
| 454 | } |
| 455 | |
| 456 | static void radeon_ttm_backend_unbind(struct ttm_device *bdev, struct ttm_tt *ttm) |
| 457 | { |
| 458 | struct radeon_ttm_tt *gtt = (void *)ttm; |
| 459 | struct radeon_device *rdev = radeon_get_rdev(bdev); |
| 460 | |
| 461 | if (gtt->userptr) |
| 462 | radeon_ttm_tt_unpin_userptr(bdev, ttm); |
| 463 | |
| 464 | if (!gtt->bound) |
| 465 | return; |
| 466 | |
| 467 | radeon_gart_unbind(rdev, offset: gtt->offset, pages: ttm->num_pages); |
| 468 | |
| 469 | gtt->bound = false; |
| 470 | } |
| 471 | |
| 472 | static void radeon_ttm_backend_destroy(struct ttm_device *bdev, struct ttm_tt *ttm) |
| 473 | { |
| 474 | struct radeon_ttm_tt *gtt = (void *)ttm; |
| 475 | |
| 476 | ttm_tt_fini(ttm: >t->ttm); |
| 477 | kfree(objp: gtt); |
| 478 | } |
| 479 | |
| 480 | static struct ttm_tt *radeon_ttm_tt_create(struct ttm_buffer_object *bo, |
| 481 | uint32_t page_flags) |
| 482 | { |
| 483 | struct radeon_ttm_tt *gtt; |
| 484 | enum ttm_caching caching; |
| 485 | struct radeon_bo *rbo; |
| 486 | #if IS_ENABLED(CONFIG_AGP) |
| 487 | struct radeon_device *rdev = radeon_get_rdev(bdev: bo->bdev); |
| 488 | |
| 489 | if (rdev->flags & RADEON_IS_AGP) { |
| 490 | return ttm_agp_tt_create(bo, bridge: rdev->agp->bridge, page_flags); |
| 491 | } |
| 492 | #endif |
| 493 | rbo = container_of(bo, struct radeon_bo, tbo); |
| 494 | |
| 495 | gtt = kzalloc(sizeof(struct radeon_ttm_tt), GFP_KERNEL); |
| 496 | if (gtt == NULL) { |
| 497 | return NULL; |
| 498 | } |
| 499 | |
| 500 | if (rbo->flags & RADEON_GEM_GTT_UC) |
| 501 | caching = ttm_uncached; |
| 502 | else if (rbo->flags & RADEON_GEM_GTT_WC) |
| 503 | caching = ttm_write_combined; |
| 504 | else |
| 505 | caching = ttm_cached; |
| 506 | |
| 507 | if (ttm_sg_tt_init(ttm_dma: >t->ttm, bo, page_flags, caching)) { |
| 508 | kfree(objp: gtt); |
| 509 | return NULL; |
| 510 | } |
| 511 | return >t->ttm; |
| 512 | } |
| 513 | |
| 514 | static struct radeon_ttm_tt *radeon_ttm_tt_to_gtt(struct radeon_device *rdev, |
| 515 | struct ttm_tt *ttm) |
| 516 | { |
| 517 | #if IS_ENABLED(CONFIG_AGP) |
| 518 | if (rdev->flags & RADEON_IS_AGP) |
| 519 | return NULL; |
| 520 | #endif |
| 521 | |
| 522 | if (!ttm) |
| 523 | return NULL; |
| 524 | return container_of(ttm, struct radeon_ttm_tt, ttm); |
| 525 | } |
| 526 | |
| 527 | static int radeon_ttm_tt_populate(struct ttm_device *bdev, |
| 528 | struct ttm_tt *ttm, |
| 529 | struct ttm_operation_ctx *ctx) |
| 530 | { |
| 531 | struct radeon_device *rdev = radeon_get_rdev(bdev); |
| 532 | struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(rdev, ttm); |
| 533 | bool slave = !!(ttm->page_flags & TTM_TT_FLAG_EXTERNAL); |
| 534 | |
| 535 | if (gtt && gtt->userptr) { |
| 536 | ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL); |
| 537 | if (!ttm->sg) |
| 538 | return -ENOMEM; |
| 539 | |
| 540 | ttm->page_flags |= TTM_TT_FLAG_EXTERNAL; |
| 541 | return 0; |
| 542 | } |
| 543 | |
| 544 | if (slave && ttm->sg) { |
| 545 | drm_prime_sg_to_dma_addr_array(sgt: ttm->sg, addrs: gtt->ttm.dma_address, |
| 546 | max_pages: ttm->num_pages); |
| 547 | return 0; |
| 548 | } |
| 549 | |
| 550 | return ttm_pool_alloc(pool: &rdev->mman.bdev.pool, tt: ttm, ctx); |
| 551 | } |
| 552 | |
| 553 | static void radeon_ttm_tt_unpopulate(struct ttm_device *bdev, struct ttm_tt *ttm) |
| 554 | { |
| 555 | struct radeon_device *rdev = radeon_get_rdev(bdev); |
| 556 | struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(rdev, ttm); |
| 557 | bool slave = !!(ttm->page_flags & TTM_TT_FLAG_EXTERNAL); |
| 558 | |
| 559 | radeon_ttm_tt_unbind(bdev, ttm); |
| 560 | |
| 561 | if (gtt && gtt->userptr) { |
| 562 | kfree(objp: ttm->sg); |
| 563 | ttm->page_flags &= ~TTM_TT_FLAG_EXTERNAL; |
| 564 | return; |
| 565 | } |
| 566 | |
| 567 | if (slave) |
| 568 | return; |
| 569 | |
| 570 | return ttm_pool_free(pool: &rdev->mman.bdev.pool, tt: ttm); |
| 571 | } |
| 572 | |
| 573 | int radeon_ttm_tt_set_userptr(struct radeon_device *rdev, |
| 574 | struct ttm_tt *ttm, uint64_t addr, |
| 575 | uint32_t flags) |
| 576 | { |
| 577 | struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(rdev, ttm); |
| 578 | |
| 579 | if (gtt == NULL) |
| 580 | return -EINVAL; |
| 581 | |
| 582 | gtt->userptr = addr; |
| 583 | gtt->usermm = current->mm; |
| 584 | gtt->userflags = flags; |
| 585 | return 0; |
| 586 | } |
| 587 | |
| 588 | bool radeon_ttm_tt_is_bound(struct ttm_device *bdev, |
| 589 | struct ttm_tt *ttm) |
| 590 | { |
| 591 | #if IS_ENABLED(CONFIG_AGP) |
| 592 | struct radeon_device *rdev = radeon_get_rdev(bdev); |
| 593 | if (rdev->flags & RADEON_IS_AGP) |
| 594 | return ttm_agp_is_bound(ttm); |
| 595 | #endif |
| 596 | return radeon_ttm_backend_is_bound(ttm); |
| 597 | } |
| 598 | |
| 599 | static int radeon_ttm_tt_bind(struct ttm_device *bdev, |
| 600 | struct ttm_tt *ttm, |
| 601 | struct ttm_resource *bo_mem) |
| 602 | { |
| 603 | #if IS_ENABLED(CONFIG_AGP) |
| 604 | struct radeon_device *rdev = radeon_get_rdev(bdev); |
| 605 | #endif |
| 606 | |
| 607 | if (!bo_mem) |
| 608 | return -EINVAL; |
| 609 | #if IS_ENABLED(CONFIG_AGP) |
| 610 | if (rdev->flags & RADEON_IS_AGP) |
| 611 | return ttm_agp_bind(ttm, bo_mem); |
| 612 | #endif |
| 613 | |
| 614 | return radeon_ttm_backend_bind(bdev, ttm, bo_mem); |
| 615 | } |
| 616 | |
| 617 | static void radeon_ttm_tt_unbind(struct ttm_device *bdev, |
| 618 | struct ttm_tt *ttm) |
| 619 | { |
| 620 | #if IS_ENABLED(CONFIG_AGP) |
| 621 | struct radeon_device *rdev = radeon_get_rdev(bdev); |
| 622 | |
| 623 | if (rdev->flags & RADEON_IS_AGP) { |
| 624 | ttm_agp_unbind(ttm); |
| 625 | return; |
| 626 | } |
| 627 | #endif |
| 628 | radeon_ttm_backend_unbind(bdev, ttm); |
| 629 | } |
| 630 | |
| 631 | static void radeon_ttm_tt_destroy(struct ttm_device *bdev, |
| 632 | struct ttm_tt *ttm) |
| 633 | { |
| 634 | #if IS_ENABLED(CONFIG_AGP) |
| 635 | struct radeon_device *rdev = radeon_get_rdev(bdev); |
| 636 | |
| 637 | if (rdev->flags & RADEON_IS_AGP) { |
| 638 | ttm_agp_destroy(ttm); |
| 639 | return; |
| 640 | } |
| 641 | #endif |
| 642 | radeon_ttm_backend_destroy(bdev, ttm); |
| 643 | } |
| 644 | |
| 645 | bool radeon_ttm_tt_has_userptr(struct radeon_device *rdev, |
| 646 | struct ttm_tt *ttm) |
| 647 | { |
| 648 | struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(rdev, ttm); |
| 649 | |
| 650 | if (gtt == NULL) |
| 651 | return false; |
| 652 | |
| 653 | return !!gtt->userptr; |
| 654 | } |
| 655 | |
| 656 | bool radeon_ttm_tt_is_readonly(struct radeon_device *rdev, |
| 657 | struct ttm_tt *ttm) |
| 658 | { |
| 659 | struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(rdev, ttm); |
| 660 | |
| 661 | if (gtt == NULL) |
| 662 | return false; |
| 663 | |
| 664 | return !!(gtt->userflags & RADEON_GEM_USERPTR_READONLY); |
| 665 | } |
| 666 | |
| 667 | static struct ttm_device_funcs radeon_bo_driver = { |
| 668 | .ttm_tt_create = &radeon_ttm_tt_create, |
| 669 | .ttm_tt_populate = &radeon_ttm_tt_populate, |
| 670 | .ttm_tt_unpopulate = &radeon_ttm_tt_unpopulate, |
| 671 | .ttm_tt_destroy = &radeon_ttm_tt_destroy, |
| 672 | .eviction_valuable = ttm_bo_eviction_valuable, |
| 673 | .evict_flags = &radeon_evict_flags, |
| 674 | .move = &radeon_bo_move, |
| 675 | .io_mem_reserve = &radeon_ttm_io_mem_reserve, |
| 676 | }; |
| 677 | |
| 678 | int radeon_ttm_init(struct radeon_device *rdev) |
| 679 | { |
| 680 | int r; |
| 681 | |
| 682 | /* No others user of address space so set it to 0 */ |
| 683 | r = ttm_device_init(bdev: &rdev->mman.bdev, funcs: &radeon_bo_driver, dev: rdev->dev, |
| 684 | mapping: rdev_to_drm(rdev)->anon_inode->i_mapping, |
| 685 | vma_manager: rdev_to_drm(rdev)->vma_offset_manager, |
| 686 | alloc_flags: (rdev->need_swiotlb ? |
| 687 | TTM_ALLOCATION_POOL_USE_DMA_ALLOC : 0) | |
| 688 | (dma_addressing_limited(dev: &rdev->pdev->dev) ? |
| 689 | TTM_ALLOCATION_POOL_USE_DMA32 : 0)); |
| 690 | if (r) { |
| 691 | DRM_ERROR("failed initializing buffer object driver(%d).\n" , r); |
| 692 | return r; |
| 693 | } |
| 694 | rdev->mman.initialized = true; |
| 695 | |
| 696 | r = radeon_ttm_init_vram(rdev); |
| 697 | if (r) { |
| 698 | DRM_ERROR("Failed initializing VRAM heap.\n" ); |
| 699 | return r; |
| 700 | } |
| 701 | /* Change the size here instead of the init above so only lpfn is affected */ |
| 702 | radeon_ttm_set_active_vram_size(rdev, size: rdev->mc.visible_vram_size); |
| 703 | |
| 704 | r = radeon_bo_create(rdev, size: 256 * 1024, PAGE_SIZE, kernel: true, |
| 705 | RADEON_GEM_DOMAIN_VRAM, flags: 0, NULL, |
| 706 | NULL, bo_ptr: &rdev->stolen_vga_memory); |
| 707 | if (r) { |
| 708 | return r; |
| 709 | } |
| 710 | r = radeon_bo_reserve(bo: rdev->stolen_vga_memory, no_intr: false); |
| 711 | if (r) |
| 712 | return r; |
| 713 | r = radeon_bo_pin(bo: rdev->stolen_vga_memory, RADEON_GEM_DOMAIN_VRAM, NULL); |
| 714 | radeon_bo_unreserve(bo: rdev->stolen_vga_memory); |
| 715 | if (r) { |
| 716 | radeon_bo_unref(bo: &rdev->stolen_vga_memory); |
| 717 | return r; |
| 718 | } |
| 719 | DRM_INFO("radeon: %uM of VRAM memory ready\n" , |
| 720 | (unsigned) (rdev->mc.real_vram_size / (1024 * 1024))); |
| 721 | |
| 722 | r = radeon_ttm_init_gtt(rdev); |
| 723 | if (r) { |
| 724 | DRM_ERROR("Failed initializing GTT heap.\n" ); |
| 725 | return r; |
| 726 | } |
| 727 | DRM_INFO("radeon: %uM of GTT memory ready.\n" , |
| 728 | (unsigned)(rdev->mc.gtt_size / (1024 * 1024))); |
| 729 | |
| 730 | radeon_ttm_debugfs_init(rdev); |
| 731 | |
| 732 | return 0; |
| 733 | } |
| 734 | |
| 735 | void radeon_ttm_fini(struct radeon_device *rdev) |
| 736 | { |
| 737 | int r; |
| 738 | |
| 739 | if (!rdev->mman.initialized) |
| 740 | return; |
| 741 | |
| 742 | if (rdev->stolen_vga_memory) { |
| 743 | r = radeon_bo_reserve(bo: rdev->stolen_vga_memory, no_intr: false); |
| 744 | if (r == 0) { |
| 745 | radeon_bo_unpin(bo: rdev->stolen_vga_memory); |
| 746 | radeon_bo_unreserve(bo: rdev->stolen_vga_memory); |
| 747 | } |
| 748 | radeon_bo_unref(bo: &rdev->stolen_vga_memory); |
| 749 | } |
| 750 | ttm_range_man_fini(bdev: &rdev->mman.bdev, TTM_PL_VRAM); |
| 751 | ttm_range_man_fini(bdev: &rdev->mman.bdev, TTM_PL_TT); |
| 752 | ttm_device_fini(bdev: &rdev->mman.bdev); |
| 753 | radeon_gart_fini(rdev); |
| 754 | rdev->mman.initialized = false; |
| 755 | DRM_INFO("radeon: ttm finalized\n" ); |
| 756 | } |
| 757 | |
| 758 | /* this should only be called at bootup or when userspace |
| 759 | * isn't running */ |
| 760 | void radeon_ttm_set_active_vram_size(struct radeon_device *rdev, u64 size) |
| 761 | { |
| 762 | struct ttm_resource_manager *man; |
| 763 | |
| 764 | if (!rdev->mman.initialized) |
| 765 | return; |
| 766 | |
| 767 | man = ttm_manager_type(bdev: &rdev->mman.bdev, TTM_PL_VRAM); |
| 768 | /* this just adjusts TTM size idea, which sets lpfn to the correct value */ |
| 769 | man->size = size >> PAGE_SHIFT; |
| 770 | } |
| 771 | |
| 772 | #if defined(CONFIG_DEBUG_FS) |
| 773 | |
| 774 | static int radeon_ttm_page_pool_show(struct seq_file *m, void *data) |
| 775 | { |
| 776 | struct radeon_device *rdev = m->private; |
| 777 | |
| 778 | return ttm_pool_debugfs(pool: &rdev->mman.bdev.pool, m); |
| 779 | } |
| 780 | |
| 781 | DEFINE_SHOW_ATTRIBUTE(radeon_ttm_page_pool); |
| 782 | |
| 783 | static int radeon_ttm_vram_open(struct inode *inode, struct file *filep) |
| 784 | { |
| 785 | struct radeon_device *rdev = inode->i_private; |
| 786 | i_size_write(inode, i_size: rdev->mc.mc_vram_size); |
| 787 | filep->private_data = inode->i_private; |
| 788 | return 0; |
| 789 | } |
| 790 | |
| 791 | static ssize_t radeon_ttm_vram_read(struct file *f, char __user *buf, |
| 792 | size_t size, loff_t *pos) |
| 793 | { |
| 794 | struct radeon_device *rdev = f->private_data; |
| 795 | ssize_t result = 0; |
| 796 | int r; |
| 797 | |
| 798 | if (size & 0x3 || *pos & 0x3) |
| 799 | return -EINVAL; |
| 800 | |
| 801 | while (size) { |
| 802 | unsigned long flags; |
| 803 | uint32_t value; |
| 804 | |
| 805 | if (*pos >= rdev->mc.mc_vram_size) |
| 806 | return result; |
| 807 | |
| 808 | spin_lock_irqsave(&rdev->mmio_idx_lock, flags); |
| 809 | WREG32(RADEON_MM_INDEX, ((uint32_t)*pos) | 0x80000000); |
| 810 | if (rdev->family >= CHIP_CEDAR) |
| 811 | WREG32(EVERGREEN_MM_INDEX_HI, *pos >> 31); |
| 812 | value = RREG32(RADEON_MM_DATA); |
| 813 | spin_unlock_irqrestore(lock: &rdev->mmio_idx_lock, flags); |
| 814 | |
| 815 | r = put_user(value, (uint32_t __user *)buf); |
| 816 | if (r) |
| 817 | return r; |
| 818 | |
| 819 | result += 4; |
| 820 | buf += 4; |
| 821 | *pos += 4; |
| 822 | size -= 4; |
| 823 | } |
| 824 | |
| 825 | return result; |
| 826 | } |
| 827 | |
| 828 | static const struct file_operations radeon_ttm_vram_fops = { |
| 829 | .owner = THIS_MODULE, |
| 830 | .open = radeon_ttm_vram_open, |
| 831 | .read = radeon_ttm_vram_read, |
| 832 | .llseek = default_llseek |
| 833 | }; |
| 834 | |
| 835 | static int radeon_ttm_gtt_open(struct inode *inode, struct file *filep) |
| 836 | { |
| 837 | struct radeon_device *rdev = inode->i_private; |
| 838 | i_size_write(inode, i_size: rdev->mc.gtt_size); |
| 839 | filep->private_data = inode->i_private; |
| 840 | return 0; |
| 841 | } |
| 842 | |
| 843 | static ssize_t radeon_ttm_gtt_read(struct file *f, char __user *buf, |
| 844 | size_t size, loff_t *pos) |
| 845 | { |
| 846 | struct radeon_device *rdev = f->private_data; |
| 847 | ssize_t result = 0; |
| 848 | int r; |
| 849 | |
| 850 | while (size) { |
| 851 | loff_t p = *pos / PAGE_SIZE; |
| 852 | unsigned off = *pos & ~PAGE_MASK; |
| 853 | size_t cur_size = min_t(size_t, size, PAGE_SIZE - off); |
| 854 | struct page *page; |
| 855 | void *ptr; |
| 856 | |
| 857 | if (p >= rdev->gart.num_cpu_pages) |
| 858 | return result; |
| 859 | |
| 860 | page = rdev->gart.pages[p]; |
| 861 | if (page) { |
| 862 | ptr = kmap_local_page(page); |
| 863 | ptr += off; |
| 864 | |
| 865 | r = copy_to_user(to: buf, from: ptr, n: cur_size); |
| 866 | kunmap_local(ptr); |
| 867 | } else |
| 868 | r = clear_user(to: buf, n: cur_size); |
| 869 | |
| 870 | if (r) |
| 871 | return -EFAULT; |
| 872 | |
| 873 | result += cur_size; |
| 874 | buf += cur_size; |
| 875 | *pos += cur_size; |
| 876 | size -= cur_size; |
| 877 | } |
| 878 | |
| 879 | return result; |
| 880 | } |
| 881 | |
| 882 | static const struct file_operations radeon_ttm_gtt_fops = { |
| 883 | .owner = THIS_MODULE, |
| 884 | .open = radeon_ttm_gtt_open, |
| 885 | .read = radeon_ttm_gtt_read, |
| 886 | .llseek = default_llseek |
| 887 | }; |
| 888 | |
| 889 | #endif |
| 890 | |
| 891 | static void radeon_ttm_debugfs_init(struct radeon_device *rdev) |
| 892 | { |
| 893 | #if defined(CONFIG_DEBUG_FS) |
| 894 | struct drm_minor *minor = rdev_to_drm(rdev)->primary; |
| 895 | struct dentry *root = minor->debugfs_root; |
| 896 | |
| 897 | debugfs_create_file("radeon_vram" , 0444, root, rdev, |
| 898 | &radeon_ttm_vram_fops); |
| 899 | debugfs_create_file("radeon_gtt" , 0444, root, rdev, |
| 900 | &radeon_ttm_gtt_fops); |
| 901 | debugfs_create_file("ttm_page_pool" , 0444, root, rdev, |
| 902 | &radeon_ttm_page_pool_fops); |
| 903 | ttm_resource_manager_create_debugfs(man: ttm_manager_type(bdev: &rdev->mman.bdev, |
| 904 | TTM_PL_VRAM), |
| 905 | parent: root, name: "radeon_vram_mm" ); |
| 906 | ttm_resource_manager_create_debugfs(man: ttm_manager_type(bdev: &rdev->mman.bdev, |
| 907 | TTM_PL_TT), |
| 908 | parent: root, name: "radeon_gtt_mm" ); |
| 909 | #endif |
| 910 | } |
| 911 | |