| File: | dev/pci/drm/amd/amdgpu/amdgpu_object.c |
| Warning: | line 1014, column 24 Value stored to 'adev' during its initialization is never read |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 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 | #include <linux/list.h> |
| 33 | #include <linux/slab.h> |
| 34 | #include <linux/dma-buf.h> |
| 35 | |
| 36 | #include <drm/amdgpu_drm.h> |
| 37 | #include <drm/drm_cache.h> |
| 38 | #include "amdgpu.h" |
| 39 | #include "amdgpu_trace.h" |
| 40 | #include "amdgpu_amdkfd.h" |
| 41 | |
| 42 | /** |
| 43 | * DOC: amdgpu_object |
| 44 | * |
| 45 | * This defines the interfaces to operate on an &amdgpu_bo buffer object which |
| 46 | * represents memory used by driver (VRAM, system memory, etc.). The driver |
| 47 | * provides DRM/GEM APIs to userspace. DRM/GEM APIs then use these interfaces |
| 48 | * to create/destroy/set buffer object which are then managed by the kernel TTM |
| 49 | * memory manager. |
| 50 | * The interfaces are also used internally by kernel clients, including gfx, |
| 51 | * uvd, etc. for kernel managed allocations used by the GPU. |
| 52 | * |
| 53 | */ |
| 54 | |
| 55 | /** |
| 56 | * amdgpu_bo_subtract_pin_size - Remove BO from pin_size accounting |
| 57 | * |
| 58 | * @bo: &amdgpu_bo buffer object |
| 59 | * |
| 60 | * This function is called when a BO stops being pinned, and updates the |
| 61 | * &amdgpu_device pin_size values accordingly. |
| 62 | */ |
| 63 | static void amdgpu_bo_subtract_pin_size(struct amdgpu_bo *bo) |
| 64 | { |
| 65 | struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); |
| 66 | |
| 67 | if (bo->tbo.mem.mem_type == TTM_PL_VRAM2) { |
| 68 | atomic64_sub(amdgpu_bo_size(bo), &adev->vram_pin_size)__sync_fetch_and_sub_8(&adev->vram_pin_size, amdgpu_bo_size (bo)); |
| 69 | atomic64_sub(amdgpu_vram_mgr_bo_visible_size(bo),__sync_fetch_and_sub_8(&adev->visible_pin_size, amdgpu_vram_mgr_bo_visible_size (bo)) |
| 70 | &adev->visible_pin_size)__sync_fetch_and_sub_8(&adev->visible_pin_size, amdgpu_vram_mgr_bo_visible_size (bo)); |
| 71 | } else if (bo->tbo.mem.mem_type == TTM_PL_TT1) { |
| 72 | atomic64_sub(amdgpu_bo_size(bo), &adev->gart_pin_size)__sync_fetch_and_sub_8(&adev->gart_pin_size, amdgpu_bo_size (bo)); |
| 73 | } |
| 74 | } |
| 75 | |
| 76 | static void amdgpu_bo_destroy(struct ttm_buffer_object *tbo) |
| 77 | { |
| 78 | struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev); |
| 79 | struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo); |
| 80 | |
| 81 | if (bo->pin_count > 0) |
| 82 | amdgpu_bo_subtract_pin_size(bo); |
| 83 | |
| 84 | amdgpu_bo_kunmap(bo); |
| 85 | |
| 86 | if (bo->tbo.base.import_attach) |
| 87 | drm_prime_gem_destroy(&bo->tbo.base, bo->tbo.sg); |
| 88 | drm_gem_object_release(&bo->tbo.base); |
| 89 | /* in case amdgpu_device_recover_vram got NULL of bo->parent */ |
| 90 | if (!list_empty(&bo->shadow_list)) { |
| 91 | mutex_lock(&adev->shadow_list_lock)rw_enter_write(&adev->shadow_list_lock); |
| 92 | list_del_init(&bo->shadow_list); |
| 93 | mutex_unlock(&adev->shadow_list_lock)rw_exit_write(&adev->shadow_list_lock); |
| 94 | } |
| 95 | amdgpu_bo_unref(&bo->parent); |
| 96 | |
| 97 | kfree(bo->metadata); |
| 98 | pool_put(&adev->ddev.objpl, bo); |
| 99 | } |
| 100 | |
| 101 | /** |
| 102 | * amdgpu_bo_is_amdgpu_bo - check if the buffer object is an &amdgpu_bo |
| 103 | * @bo: buffer object to be checked |
| 104 | * |
| 105 | * Uses destroy function associated with the object to determine if this is |
| 106 | * an &amdgpu_bo. |
| 107 | * |
| 108 | * Returns: |
| 109 | * true if the object belongs to &amdgpu_bo, false if not. |
| 110 | */ |
| 111 | bool_Bool amdgpu_bo_is_amdgpu_bo(struct ttm_buffer_object *bo) |
| 112 | { |
| 113 | if (bo->destroy == &amdgpu_bo_destroy) |
| 114 | return true1; |
| 115 | return false0; |
| 116 | } |
| 117 | |
| 118 | /** |
| 119 | * amdgpu_bo_placement_from_domain - set buffer's placement |
| 120 | * @abo: &amdgpu_bo buffer object whose placement is to be set |
| 121 | * @domain: requested domain |
| 122 | * |
| 123 | * Sets buffer's placement according to requested domain and the buffer's |
| 124 | * flags. |
| 125 | */ |
| 126 | void amdgpu_bo_placement_from_domain(struct amdgpu_bo *abo, u32 domain) |
| 127 | { |
| 128 | struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev); |
| 129 | struct ttm_placement *placement = &abo->placement; |
| 130 | struct ttm_place *places = abo->placements; |
| 131 | u64 flags = abo->flags; |
| 132 | u32 c = 0; |
| 133 | |
| 134 | if (domain & AMDGPU_GEM_DOMAIN_VRAM0x4) { |
| 135 | unsigned visible_pfn = adev->gmc.visible_vram_size >> PAGE_SHIFT12; |
| 136 | |
| 137 | places[c].fpfn = 0; |
| 138 | places[c].lpfn = 0; |
| 139 | places[c].mem_type = TTM_PL_VRAM2; |
| 140 | places[c].flags = TTM_PL_FLAG_WC(1 << 18) | TTM_PL_FLAG_UNCACHED(1 << 17); |
| 141 | |
| 142 | if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED(1 << 0)) |
| 143 | places[c].lpfn = visible_pfn; |
| 144 | else |
| 145 | places[c].flags |= TTM_PL_FLAG_TOPDOWN(1 << 22); |
| 146 | |
| 147 | if (flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS(1 << 5)) |
| 148 | places[c].flags |= TTM_PL_FLAG_CONTIGUOUS(1 << 19); |
| 149 | c++; |
| 150 | } |
| 151 | |
| 152 | if (domain & AMDGPU_GEM_DOMAIN_GTT0x2) { |
| 153 | places[c].fpfn = 0; |
| 154 | places[c].lpfn = 0; |
| 155 | places[c].mem_type = TTM_PL_TT1; |
| 156 | places[c].flags = 0; |
| 157 | if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC(1 << 2)) |
| 158 | places[c].flags |= TTM_PL_FLAG_WC(1 << 18) | |
| 159 | TTM_PL_FLAG_UNCACHED(1 << 17); |
| 160 | else |
| 161 | places[c].flags |= TTM_PL_FLAG_CACHED(1 << 16); |
| 162 | c++; |
| 163 | } |
| 164 | |
| 165 | if (domain & AMDGPU_GEM_DOMAIN_CPU0x1) { |
| 166 | places[c].fpfn = 0; |
| 167 | places[c].lpfn = 0; |
| 168 | places[c].mem_type = TTM_PL_SYSTEM0; |
| 169 | places[c].flags = 0; |
| 170 | if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC(1 << 2)) |
| 171 | places[c].flags |= TTM_PL_FLAG_WC(1 << 18) | |
| 172 | TTM_PL_FLAG_UNCACHED(1 << 17); |
| 173 | else |
| 174 | places[c].flags |= TTM_PL_FLAG_CACHED(1 << 16); |
| 175 | c++; |
| 176 | } |
| 177 | |
| 178 | if (domain & AMDGPU_GEM_DOMAIN_GDS0x8) { |
| 179 | places[c].fpfn = 0; |
| 180 | places[c].lpfn = 0; |
| 181 | places[c].mem_type = AMDGPU_PL_GDS(3 + 0); |
| 182 | places[c].flags = TTM_PL_FLAG_UNCACHED(1 << 17); |
| 183 | c++; |
| 184 | } |
| 185 | |
| 186 | if (domain & AMDGPU_GEM_DOMAIN_GWS0x10) { |
| 187 | places[c].fpfn = 0; |
| 188 | places[c].lpfn = 0; |
| 189 | places[c].mem_type = AMDGPU_PL_GWS(3 + 1); |
| 190 | places[c].flags = TTM_PL_FLAG_UNCACHED(1 << 17); |
| 191 | c++; |
| 192 | } |
| 193 | |
| 194 | if (domain & AMDGPU_GEM_DOMAIN_OA0x20) { |
| 195 | places[c].fpfn = 0; |
| 196 | places[c].lpfn = 0; |
| 197 | places[c].mem_type = AMDGPU_PL_OA(3 + 2); |
| 198 | places[c].flags = TTM_PL_FLAG_UNCACHED(1 << 17); |
| 199 | c++; |
| 200 | } |
| 201 | |
| 202 | if (!c) { |
| 203 | places[c].fpfn = 0; |
| 204 | places[c].lpfn = 0; |
| 205 | places[c].mem_type = TTM_PL_SYSTEM0; |
| 206 | places[c].flags = TTM_PL_MASK_CACHING((1 << 16) | (1 << 17) | (1 << 18)); |
| 207 | c++; |
| 208 | } |
| 209 | |
| 210 | BUG_ON(c > AMDGPU_BO_MAX_PLACEMENTS)((!(c > 3)) ? (void)0 : __assert("diagnostic ", "/usr/src/sys/dev/pci/drm/amd/amdgpu/amdgpu_object.c" , 210, "!(c > 3)")); |
| 211 | |
| 212 | placement->num_placement = c; |
| 213 | placement->placement = places; |
| 214 | |
| 215 | placement->num_busy_placement = c; |
| 216 | placement->busy_placement = places; |
| 217 | } |
| 218 | |
| 219 | /** |
| 220 | * amdgpu_bo_create_reserved - create reserved BO for kernel use |
| 221 | * |
| 222 | * @adev: amdgpu device object |
| 223 | * @size: size for the new BO |
| 224 | * @align: alignment for the new BO |
| 225 | * @domain: where to place it |
| 226 | * @bo_ptr: used to initialize BOs in structures |
| 227 | * @gpu_addr: GPU addr of the pinned BO |
| 228 | * @cpu_addr: optional CPU address mapping |
| 229 | * |
| 230 | * Allocates and pins a BO for kernel internal use, and returns it still |
| 231 | * reserved. |
| 232 | * |
| 233 | * Note: For bo_ptr new BO is only created if bo_ptr points to NULL. |
| 234 | * |
| 235 | * Returns: |
| 236 | * 0 on success, negative error code otherwise. |
| 237 | */ |
| 238 | int amdgpu_bo_create_reserved(struct amdgpu_device *adev, |
| 239 | unsigned long size, int align, |
| 240 | u32 domain, struct amdgpu_bo **bo_ptr, |
| 241 | u64 *gpu_addr, void **cpu_addr) |
| 242 | { |
| 243 | struct amdgpu_bo_param bp; |
| 244 | bool_Bool free = false0; |
| 245 | int r; |
| 246 | |
| 247 | if (!size) { |
| 248 | amdgpu_bo_unref(bo_ptr); |
| 249 | return 0; |
| 250 | } |
| 251 | |
| 252 | memset(&bp, 0, sizeof(bp))__builtin_memset((&bp), (0), (sizeof(bp))); |
| 253 | bp.size = size; |
| 254 | bp.byte_align = align; |
| 255 | bp.domain = domain; |
| 256 | bp.flags = cpu_addr ? AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED(1 << 0) |
| 257 | : AMDGPU_GEM_CREATE_NO_CPU_ACCESS(1 << 1); |
| 258 | bp.flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS(1 << 5); |
| 259 | bp.type = ttm_bo_type_kernel; |
| 260 | bp.resv = NULL((void *)0); |
| 261 | |
| 262 | if (!*bo_ptr) { |
| 263 | r = amdgpu_bo_create(adev, &bp, bo_ptr); |
| 264 | if (r) { |
| 265 | dev_err(adev->dev, "(%d) failed to allocate kernel bo\n",printf("drm:pid%d:%s *ERROR* " "(%d) failed to allocate kernel bo\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , r) |
| 266 | r)printf("drm:pid%d:%s *ERROR* " "(%d) failed to allocate kernel bo\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , r); |
| 267 | return r; |
| 268 | } |
| 269 | free = true1; |
| 270 | } |
| 271 | |
| 272 | r = amdgpu_bo_reserve(*bo_ptr, false0); |
| 273 | if (r) { |
| 274 | dev_err(adev->dev, "(%d) failed to reserve kernel bo\n", r)printf("drm:pid%d:%s *ERROR* " "(%d) failed to reserve kernel bo\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , r); |
| 275 | goto error_free; |
| 276 | } |
| 277 | |
| 278 | r = amdgpu_bo_pin(*bo_ptr, domain); |
| 279 | if (r) { |
| 280 | dev_err(adev->dev, "(%d) kernel bo pin failed\n", r)printf("drm:pid%d:%s *ERROR* " "(%d) kernel bo pin failed\n", ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , r); |
| 281 | goto error_unreserve; |
| 282 | } |
| 283 | |
| 284 | r = amdgpu_ttm_alloc_gart(&(*bo_ptr)->tbo); |
| 285 | if (r) { |
| 286 | dev_err(adev->dev, "%p bind failed\n", *bo_ptr)printf("drm:pid%d:%s *ERROR* " "%p bind failed\n", ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof (struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p-> ps_pid, __func__ , *bo_ptr); |
| 287 | goto error_unpin; |
| 288 | } |
| 289 | |
| 290 | if (gpu_addr) |
| 291 | *gpu_addr = amdgpu_bo_gpu_offset(*bo_ptr); |
| 292 | |
| 293 | if (cpu_addr) { |
| 294 | r = amdgpu_bo_kmap(*bo_ptr, cpu_addr); |
| 295 | if (r) { |
| 296 | dev_err(adev->dev, "(%d) kernel bo map failed\n", r)printf("drm:pid%d:%s *ERROR* " "(%d) kernel bo map failed\n", ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , r); |
| 297 | goto error_unpin; |
| 298 | } |
| 299 | } |
| 300 | |
| 301 | return 0; |
| 302 | |
| 303 | error_unpin: |
| 304 | amdgpu_bo_unpin(*bo_ptr); |
| 305 | error_unreserve: |
| 306 | amdgpu_bo_unreserve(*bo_ptr); |
| 307 | |
| 308 | error_free: |
| 309 | if (free) |
| 310 | amdgpu_bo_unref(bo_ptr); |
| 311 | |
| 312 | return r; |
| 313 | } |
| 314 | |
| 315 | /** |
| 316 | * amdgpu_bo_create_kernel - create BO for kernel use |
| 317 | * |
| 318 | * @adev: amdgpu device object |
| 319 | * @size: size for the new BO |
| 320 | * @align: alignment for the new BO |
| 321 | * @domain: where to place it |
| 322 | * @bo_ptr: used to initialize BOs in structures |
| 323 | * @gpu_addr: GPU addr of the pinned BO |
| 324 | * @cpu_addr: optional CPU address mapping |
| 325 | * |
| 326 | * Allocates and pins a BO for kernel internal use. |
| 327 | * |
| 328 | * Note: For bo_ptr new BO is only created if bo_ptr points to NULL. |
| 329 | * |
| 330 | * Returns: |
| 331 | * 0 on success, negative error code otherwise. |
| 332 | */ |
| 333 | int amdgpu_bo_create_kernel(struct amdgpu_device *adev, |
| 334 | unsigned long size, int align, |
| 335 | u32 domain, struct amdgpu_bo **bo_ptr, |
| 336 | u64 *gpu_addr, void **cpu_addr) |
| 337 | { |
| 338 | int r; |
| 339 | |
| 340 | r = amdgpu_bo_create_reserved(adev, size, align, domain, bo_ptr, |
| 341 | gpu_addr, cpu_addr); |
| 342 | |
| 343 | if (r) |
| 344 | return r; |
| 345 | |
| 346 | if (*bo_ptr) |
| 347 | amdgpu_bo_unreserve(*bo_ptr); |
| 348 | |
| 349 | return 0; |
| 350 | } |
| 351 | |
| 352 | /** |
| 353 | * amdgpu_bo_create_kernel_at - create BO for kernel use at specific location |
| 354 | * |
| 355 | * @adev: amdgpu device object |
| 356 | * @offset: offset of the BO |
| 357 | * @size: size of the BO |
| 358 | * @domain: where to place it |
| 359 | * @bo_ptr: used to initialize BOs in structures |
| 360 | * @cpu_addr: optional CPU address mapping |
| 361 | * |
| 362 | * Creates a kernel BO at a specific offset in the address space of the domain. |
| 363 | * |
| 364 | * Returns: |
| 365 | * 0 on success, negative error code otherwise. |
| 366 | */ |
| 367 | int amdgpu_bo_create_kernel_at(struct amdgpu_device *adev, |
| 368 | uint64_t offset, uint64_t size, uint32_t domain, |
| 369 | struct amdgpu_bo **bo_ptr, void **cpu_addr) |
| 370 | { |
| 371 | struct ttm_operation_ctx ctx = { false0, false0 }; |
| 372 | unsigned int i; |
| 373 | int r; |
| 374 | |
| 375 | offset &= LINUX_PAGE_MASK(~((1 << 12) - 1)); |
| 376 | size = roundup2(size, PAGE_SIZE)(((size) + (((1 << 12)) - 1)) & (~((__typeof(size)) ((1 << 12)) - 1))); |
| 377 | |
| 378 | r = amdgpu_bo_create_reserved(adev, size, PAGE_SIZE(1 << 12), domain, bo_ptr, |
| 379 | NULL((void *)0), cpu_addr); |
| 380 | if (r) |
| 381 | return r; |
| 382 | |
| 383 | if ((*bo_ptr) == NULL((void *)0)) |
| 384 | return 0; |
| 385 | |
| 386 | /* |
| 387 | * Remove the original mem node and create a new one at the request |
| 388 | * position. |
| 389 | */ |
| 390 | if (cpu_addr) |
| 391 | amdgpu_bo_kunmap(*bo_ptr); |
| 392 | |
| 393 | ttm_resource_free(&(*bo_ptr)->tbo, &(*bo_ptr)->tbo.mem); |
| 394 | |
| 395 | for (i = 0; i < (*bo_ptr)->placement.num_placement; ++i) { |
| 396 | (*bo_ptr)->placements[i].fpfn = offset >> PAGE_SHIFT12; |
| 397 | (*bo_ptr)->placements[i].lpfn = (offset + size) >> PAGE_SHIFT12; |
| 398 | } |
| 399 | r = ttm_bo_mem_space(&(*bo_ptr)->tbo, &(*bo_ptr)->placement, |
| 400 | &(*bo_ptr)->tbo.mem, &ctx); |
| 401 | if (r) |
| 402 | goto error; |
| 403 | |
| 404 | if (cpu_addr) { |
| 405 | r = amdgpu_bo_kmap(*bo_ptr, cpu_addr); |
| 406 | if (r) |
| 407 | goto error; |
| 408 | } |
| 409 | |
| 410 | amdgpu_bo_unreserve(*bo_ptr); |
| 411 | return 0; |
| 412 | |
| 413 | error: |
| 414 | amdgpu_bo_unreserve(*bo_ptr); |
| 415 | amdgpu_bo_unref(bo_ptr); |
| 416 | return r; |
| 417 | } |
| 418 | |
| 419 | /** |
| 420 | * amdgpu_bo_free_kernel - free BO for kernel use |
| 421 | * |
| 422 | * @bo: amdgpu BO to free |
| 423 | * @gpu_addr: pointer to where the BO's GPU memory space address was stored |
| 424 | * @cpu_addr: pointer to where the BO's CPU memory space address was stored |
| 425 | * |
| 426 | * unmaps and unpin a BO for kernel internal use. |
| 427 | */ |
| 428 | void amdgpu_bo_free_kernel(struct amdgpu_bo **bo, u64 *gpu_addr, |
| 429 | void **cpu_addr) |
| 430 | { |
| 431 | if (*bo == NULL((void *)0)) |
| 432 | return; |
| 433 | |
| 434 | if (likely(amdgpu_bo_reserve(*bo, true) == 0)__builtin_expect(!!(amdgpu_bo_reserve(*bo, 1) == 0), 1)) { |
| 435 | if (cpu_addr) |
| 436 | amdgpu_bo_kunmap(*bo); |
| 437 | |
| 438 | amdgpu_bo_unpin(*bo); |
| 439 | amdgpu_bo_unreserve(*bo); |
| 440 | } |
| 441 | amdgpu_bo_unref(bo); |
| 442 | |
| 443 | if (gpu_addr) |
| 444 | *gpu_addr = 0; |
| 445 | |
| 446 | if (cpu_addr) |
| 447 | *cpu_addr = NULL((void *)0); |
| 448 | } |
| 449 | |
| 450 | /* Validate bo size is bit bigger then the request domain */ |
| 451 | static bool_Bool amdgpu_bo_validate_size(struct amdgpu_device *adev, |
| 452 | unsigned long size, u32 domain) |
| 453 | { |
| 454 | struct ttm_resource_manager *man = NULL((void *)0); |
| 455 | |
| 456 | /* |
| 457 | * If GTT is part of requested domains the check must succeed to |
| 458 | * allow fall back to GTT |
| 459 | */ |
| 460 | if (domain & AMDGPU_GEM_DOMAIN_GTT0x2) { |
| 461 | man = ttm_manager_type(&adev->mman.bdev, TTM_PL_TT1); |
| 462 | |
| 463 | if (size < (man->size << PAGE_SHIFT12)) |
| 464 | return true1; |
| 465 | else |
| 466 | goto fail; |
| 467 | } |
| 468 | |
| 469 | if (domain & AMDGPU_GEM_DOMAIN_VRAM0x4) { |
| 470 | man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM2); |
| 471 | |
| 472 | if (size < (man->size << PAGE_SHIFT12)) |
| 473 | return true1; |
| 474 | else |
| 475 | goto fail; |
| 476 | } |
| 477 | |
| 478 | |
| 479 | /* TODO add more domains checks, such as AMDGPU_GEM_DOMAIN_CPU */ |
| 480 | return true1; |
| 481 | |
| 482 | fail: |
| 483 | DRM_DEBUG("BO size %lu > total memory in domain: %llu\n", size,__drm_dbg(DRM_UT_CORE, "BO size %lu > total memory in domain: %llu\n" , size, man->size << 12) |
| 484 | man->size << PAGE_SHIFT)__drm_dbg(DRM_UT_CORE, "BO size %lu > total memory in domain: %llu\n" , size, man->size << 12); |
| 485 | return false0; |
| 486 | } |
| 487 | |
| 488 | bool_Bool amdgpu_bo_support_uswc(u64 bo_flags) |
| 489 | { |
| 490 | |
| 491 | #ifdef CONFIG_X86_32 |
| 492 | /* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit |
| 493 | * See https://bugs.freedesktop.org/show_bug.cgi?id=84627 |
| 494 | */ |
| 495 | return false0; |
| 496 | #elif defined(CONFIG_X861) && !defined(CONFIG_X86_PAT1) |
| 497 | /* Don't try to enable write-combining when it can't work, or things |
| 498 | * may be slow |
| 499 | * See https://bugs.freedesktop.org/show_bug.cgi?id=88758 |
| 500 | */ |
| 501 | |
| 502 | #ifndef CONFIG_COMPILE_TEST |
| 503 | #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT1 for better performance \ |
| 504 | thanks to write-combining |
| 505 | #endif |
| 506 | |
| 507 | if (bo_flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC(1 << 2)) |
| 508 | DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "({ static int __warned; if (!__warned) { printk("\0016" "[" "drm" "] " "Please enable CONFIG_MTRR and CONFIG_X86_PAT for " "better performance thanks to write-combining\n" ); __warned = 1; } }) |
| 509 | "better performance thanks to write-combining\n")({ static int __warned; if (!__warned) { printk("\0016" "[" "drm" "] " "Please enable CONFIG_MTRR and CONFIG_X86_PAT for " "better performance thanks to write-combining\n" ); __warned = 1; } }); |
| 510 | return false0; |
| 511 | #else |
| 512 | /* For architectures that don't support WC memory, |
| 513 | * mask out the WC flag from the BO |
| 514 | */ |
| 515 | if (!drm_arch_can_wc_memory()) |
| 516 | return false0; |
| 517 | |
| 518 | return true1; |
| 519 | #endif |
| 520 | } |
| 521 | |
| 522 | static int amdgpu_bo_do_create(struct amdgpu_device *adev, |
| 523 | struct amdgpu_bo_param *bp, |
| 524 | struct amdgpu_bo **bo_ptr) |
| 525 | { |
| 526 | struct ttm_operation_ctx ctx = { |
| 527 | .interruptible = (bp->type != ttm_bo_type_kernel), |
| 528 | .no_wait_gpu = bp->no_wait_gpu, |
| 529 | .resv = bp->resv, |
| 530 | .flags = bp->type != ttm_bo_type_kernel ? |
| 531 | TTM_OPT_FLAG_ALLOW_RES_EVICT0x1 : 0 |
| 532 | }; |
| 533 | struct amdgpu_bo *bo; |
| 534 | unsigned long page_align, size = bp->size; |
| 535 | size_t acc_size; |
| 536 | int r; |
| 537 | |
| 538 | /* Note that GDS/GWS/OA allocates 1 page per byte/resource. */ |
| 539 | if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS0x10 | AMDGPU_GEM_DOMAIN_OA0x20)) { |
| 540 | /* GWS and OA don't need any alignment. */ |
| 541 | page_align = bp->byte_align; |
| 542 | size <<= PAGE_SHIFT12; |
| 543 | } else if (bp->domain & AMDGPU_GEM_DOMAIN_GDS0x8) { |
| 544 | /* Both size and alignment must be a multiple of 4. */ |
| 545 | page_align = roundup2(bp->byte_align, 4)(((bp->byte_align) + ((4) - 1)) & (~((__typeof(bp-> byte_align))(4) - 1))); |
| 546 | size = roundup2(size, 4)(((size) + ((4) - 1)) & (~((__typeof(size))(4) - 1))) << PAGE_SHIFT12; |
| 547 | } else { |
| 548 | /* Memory should be aligned at least to a page size. */ |
| 549 | page_align = roundup2(bp->byte_align, PAGE_SIZE)(((bp->byte_align) + (((1 << 12)) - 1)) & (~((__typeof (bp->byte_align))((1 << 12)) - 1))) >> PAGE_SHIFT12; |
| 550 | size = roundup2(size, PAGE_SIZE)(((size) + (((1 << 12)) - 1)) & (~((__typeof(size)) ((1 << 12)) - 1))); |
| 551 | } |
| 552 | |
| 553 | if (!amdgpu_bo_validate_size(adev, size, bp->domain)) |
| 554 | return -ENOMEM12; |
| 555 | |
| 556 | *bo_ptr = NULL((void *)0); |
| 557 | |
| 558 | acc_size = ttm_bo_dma_acc_size(&adev->mman.bdev, size, |
| 559 | sizeof(struct amdgpu_bo)); |
| 560 | |
| 561 | bo = pool_get(&adev->ddev.objpl, PR_WAITOK0x0001 | PR_ZERO0x0008); |
| 562 | if (bo == NULL((void *)0)) |
| 563 | return -ENOMEM12; |
| 564 | drm_gem_private_object_init(adev_to_drm(adev), &bo->tbo.base, size); |
| 565 | bo->adev = adev; |
| 566 | INIT_LIST_HEAD(&bo->shadow_list); |
| 567 | bo->vm_bo = NULL((void *)0); |
| 568 | bo->preferred_domains = bp->preferred_domain ? bp->preferred_domain : |
| 569 | bp->domain; |
| 570 | bo->allowed_domains = bo->preferred_domains; |
| 571 | if (bp->type != ttm_bo_type_kernel && |
| 572 | bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM0x4) |
| 573 | bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT0x2; |
| 574 | |
| 575 | bo->flags = bp->flags; |
| 576 | |
| 577 | if (!amdgpu_bo_support_uswc(bo->flags)) |
| 578 | bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC(1 << 2); |
| 579 | |
| 580 | bo->tbo.bdev = &adev->mman.bdev; |
| 581 | if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS0x10 | AMDGPU_GEM_DOMAIN_OA0x20 | |
| 582 | AMDGPU_GEM_DOMAIN_GDS0x8)) |
| 583 | amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU0x1); |
| 584 | else |
| 585 | amdgpu_bo_placement_from_domain(bo, bp->domain); |
| 586 | if (bp->type == ttm_bo_type_kernel) |
| 587 | bo->tbo.priority = 1; |
| 588 | |
| 589 | r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, size, bp->type, |
| 590 | &bo->placement, page_align, &ctx, acc_size, |
| 591 | NULL((void *)0), bp->resv, &amdgpu_bo_destroy); |
| 592 | if (unlikely(r != 0)__builtin_expect(!!(r != 0), 0)) |
| 593 | return r; |
| 594 | |
| 595 | if (!amdgpu_gmc_vram_full_visible(&adev->gmc) && |
| 596 | bo->tbo.mem.mem_type == TTM_PL_VRAM2 && |
| 597 | bo->tbo.mem.start < adev->gmc.visible_vram_size >> PAGE_SHIFT12) |
| 598 | amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, |
| 599 | ctx.bytes_moved); |
| 600 | else |
| 601 | amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0); |
| 602 | |
| 603 | if (bp->flags & AMDGPU_GEM_CREATE_VRAM_CLEARED(1 << 3) && |
| 604 | bo->tbo.mem.mem_type == TTM_PL_VRAM2) { |
| 605 | struct dma_fence *fence; |
| 606 | |
| 607 | r = amdgpu_fill_buffer(bo, 0, bo->tbo.base.resv, &fence); |
| 608 | if (unlikely(r)__builtin_expect(!!(r), 0)) |
| 609 | goto fail_unreserve; |
| 610 | |
| 611 | amdgpu_bo_fence(bo, fence, false0); |
| 612 | dma_fence_put(bo->tbo.moving); |
| 613 | bo->tbo.moving = dma_fence_get(fence); |
| 614 | dma_fence_put(fence); |
| 615 | } |
| 616 | if (!bp->resv) |
| 617 | amdgpu_bo_unreserve(bo); |
| 618 | *bo_ptr = bo; |
| 619 | |
| 620 | trace_amdgpu_bo_create(bo); |
| 621 | |
| 622 | /* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */ |
| 623 | if (bp->type == ttm_bo_type_device) |
| 624 | bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED(1 << 0); |
| 625 | |
| 626 | return 0; |
| 627 | |
| 628 | fail_unreserve: |
| 629 | if (!bp->resv) |
| 630 | dma_resv_unlock(bo->tbo.base.resv); |
| 631 | amdgpu_bo_unref(&bo); |
| 632 | return r; |
| 633 | } |
| 634 | |
| 635 | static int amdgpu_bo_create_shadow(struct amdgpu_device *adev, |
| 636 | unsigned long size, |
| 637 | struct amdgpu_bo *bo) |
| 638 | { |
| 639 | struct amdgpu_bo_param bp; |
| 640 | int r; |
| 641 | |
| 642 | if (bo->shadow) |
| 643 | return 0; |
| 644 | |
| 645 | memset(&bp, 0, sizeof(bp))__builtin_memset((&bp), (0), (sizeof(bp))); |
| 646 | bp.size = size; |
| 647 | bp.domain = AMDGPU_GEM_DOMAIN_GTT0x2; |
| 648 | bp.flags = AMDGPU_GEM_CREATE_CPU_GTT_USWC(1 << 2) | |
| 649 | AMDGPU_GEM_CREATE_SHADOW(1 << 4); |
| 650 | bp.type = ttm_bo_type_kernel; |
| 651 | bp.resv = bo->tbo.base.resv; |
| 652 | |
| 653 | r = amdgpu_bo_do_create(adev, &bp, &bo->shadow); |
| 654 | if (!r) { |
| 655 | bo->shadow->parent = amdgpu_bo_ref(bo); |
| 656 | mutex_lock(&adev->shadow_list_lock)rw_enter_write(&adev->shadow_list_lock); |
| 657 | list_add_tail(&bo->shadow->shadow_list, &adev->shadow_list); |
| 658 | mutex_unlock(&adev->shadow_list_lock)rw_exit_write(&adev->shadow_list_lock); |
| 659 | } |
| 660 | |
| 661 | return r; |
| 662 | } |
| 663 | |
| 664 | /** |
| 665 | * amdgpu_bo_create - create an &amdgpu_bo buffer object |
| 666 | * @adev: amdgpu device object |
| 667 | * @bp: parameters to be used for the buffer object |
| 668 | * @bo_ptr: pointer to the buffer object pointer |
| 669 | * |
| 670 | * Creates an &amdgpu_bo buffer object; and if requested, also creates a |
| 671 | * shadow object. |
| 672 | * Shadow object is used to backup the original buffer object, and is always |
| 673 | * in GTT. |
| 674 | * |
| 675 | * Returns: |
| 676 | * 0 for success or a negative error code on failure. |
| 677 | */ |
| 678 | int amdgpu_bo_create(struct amdgpu_device *adev, |
| 679 | struct amdgpu_bo_param *bp, |
| 680 | struct amdgpu_bo **bo_ptr) |
| 681 | { |
| 682 | u64 flags = bp->flags; |
| 683 | int r; |
| 684 | |
| 685 | bp->flags = bp->flags & ~AMDGPU_GEM_CREATE_SHADOW(1 << 4); |
| 686 | r = amdgpu_bo_do_create(adev, bp, bo_ptr); |
| 687 | if (r) |
| 688 | return r; |
| 689 | |
| 690 | if ((flags & AMDGPU_GEM_CREATE_SHADOW(1 << 4)) && !(adev->flags & AMD_IS_APU)) { |
| 691 | if (!bp->resv) |
| 692 | WARN_ON(dma_resv_lock((*bo_ptr)->tbo.base.resv,({ int __ret = !!(dma_resv_lock((*bo_ptr)->tbo.base.resv, ( (void *)0))); if (__ret) printf("WARNING %s failed at %s:%d\n" , "dma_resv_lock((*bo_ptr)->tbo.base.resv, ((void *)0))", "/usr/src/sys/dev/pci/drm/amd/amdgpu/amdgpu_object.c" , 693); __builtin_expect(!!(__ret), 0); }) |
| 693 | NULL))({ int __ret = !!(dma_resv_lock((*bo_ptr)->tbo.base.resv, ( (void *)0))); if (__ret) printf("WARNING %s failed at %s:%d\n" , "dma_resv_lock((*bo_ptr)->tbo.base.resv, ((void *)0))", "/usr/src/sys/dev/pci/drm/amd/amdgpu/amdgpu_object.c" , 693); __builtin_expect(!!(__ret), 0); }); |
| 694 | |
| 695 | r = amdgpu_bo_create_shadow(adev, bp->size, *bo_ptr); |
| 696 | |
| 697 | if (!bp->resv) |
| 698 | dma_resv_unlock((*bo_ptr)->tbo.base.resv); |
| 699 | |
| 700 | if (r) |
| 701 | amdgpu_bo_unref(bo_ptr); |
| 702 | } |
| 703 | |
| 704 | return r; |
| 705 | } |
| 706 | |
| 707 | /** |
| 708 | * amdgpu_bo_validate - validate an &amdgpu_bo buffer object |
| 709 | * @bo: pointer to the buffer object |
| 710 | * |
| 711 | * Sets placement according to domain; and changes placement and caching |
| 712 | * policy of the buffer object according to the placement. |
| 713 | * This is used for validating shadow bos. It calls ttm_bo_validate() to |
| 714 | * make sure the buffer is resident where it needs to be. |
| 715 | * |
| 716 | * Returns: |
| 717 | * 0 for success or a negative error code on failure. |
| 718 | */ |
| 719 | int amdgpu_bo_validate(struct amdgpu_bo *bo) |
| 720 | { |
| 721 | struct ttm_operation_ctx ctx = { false0, false0 }; |
| 722 | uint32_t domain; |
| 723 | int r; |
| 724 | |
| 725 | if (bo->pin_count) |
| 726 | return 0; |
| 727 | |
| 728 | domain = bo->preferred_domains; |
| 729 | |
| 730 | retry: |
| 731 | amdgpu_bo_placement_from_domain(bo, domain); |
| 732 | r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); |
| 733 | if (unlikely(r == -ENOMEM)__builtin_expect(!!(r == -12), 0) && domain != bo->allowed_domains) { |
| 734 | domain = bo->allowed_domains; |
| 735 | goto retry; |
| 736 | } |
| 737 | |
| 738 | return r; |
| 739 | } |
| 740 | |
| 741 | /** |
| 742 | * amdgpu_bo_restore_shadow - restore an &amdgpu_bo shadow |
| 743 | * |
| 744 | * @shadow: &amdgpu_bo shadow to be restored |
| 745 | * @fence: dma_fence associated with the operation |
| 746 | * |
| 747 | * Copies a buffer object's shadow content back to the object. |
| 748 | * This is used for recovering a buffer from its shadow in case of a gpu |
| 749 | * reset where vram context may be lost. |
| 750 | * |
| 751 | * Returns: |
| 752 | * 0 for success or a negative error code on failure. |
| 753 | */ |
| 754 | int amdgpu_bo_restore_shadow(struct amdgpu_bo *shadow, struct dma_fence **fence) |
| 755 | |
| 756 | { |
| 757 | struct amdgpu_device *adev = amdgpu_ttm_adev(shadow->tbo.bdev); |
| 758 | struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring; |
| 759 | uint64_t shadow_addr, parent_addr; |
| 760 | |
| 761 | shadow_addr = amdgpu_bo_gpu_offset(shadow); |
| 762 | parent_addr = amdgpu_bo_gpu_offset(shadow->parent); |
| 763 | |
| 764 | return amdgpu_copy_buffer(ring, shadow_addr, parent_addr, |
| 765 | amdgpu_bo_size(shadow), NULL((void *)0), fence, |
| 766 | true1, false0, false0); |
| 767 | } |
| 768 | |
| 769 | /** |
| 770 | * amdgpu_bo_kmap - map an &amdgpu_bo buffer object |
| 771 | * @bo: &amdgpu_bo buffer object to be mapped |
| 772 | * @ptr: kernel virtual address to be returned |
| 773 | * |
| 774 | * Calls ttm_bo_kmap() to set up the kernel virtual mapping; calls |
| 775 | * amdgpu_bo_kptr() to get the kernel virtual address. |
| 776 | * |
| 777 | * Returns: |
| 778 | * 0 for success or a negative error code on failure. |
| 779 | */ |
| 780 | int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr) |
| 781 | { |
| 782 | void *kptr; |
| 783 | long r; |
| 784 | |
| 785 | if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS(1 << 1)) |
| 786 | return -EPERM1; |
| 787 | |
| 788 | kptr = amdgpu_bo_kptr(bo); |
| 789 | if (kptr) { |
| 790 | if (ptr) |
| 791 | *ptr = kptr; |
| 792 | return 0; |
| 793 | } |
| 794 | |
| 795 | r = dma_resv_wait_timeout_rcu(bo->tbo.base.resv, false0, false0, |
| 796 | MAX_SCHEDULE_TIMEOUT(0x7fffffff)); |
| 797 | if (r < 0) |
| 798 | return r; |
| 799 | |
| 800 | r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap); |
| 801 | if (r) |
| 802 | return r; |
| 803 | |
| 804 | if (ptr) |
| 805 | *ptr = amdgpu_bo_kptr(bo); |
| 806 | |
| 807 | return 0; |
| 808 | } |
| 809 | |
| 810 | /** |
| 811 | * amdgpu_bo_kptr - returns a kernel virtual address of the buffer object |
| 812 | * @bo: &amdgpu_bo buffer object |
| 813 | * |
| 814 | * Calls ttm_kmap_obj_virtual() to get the kernel virtual address |
| 815 | * |
| 816 | * Returns: |
| 817 | * the virtual address of a buffer object area. |
| 818 | */ |
| 819 | void *amdgpu_bo_kptr(struct amdgpu_bo *bo) |
| 820 | { |
| 821 | bool_Bool is_iomem; |
| 822 | |
| 823 | return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem); |
| 824 | } |
| 825 | |
| 826 | /** |
| 827 | * amdgpu_bo_kunmap - unmap an &amdgpu_bo buffer object |
| 828 | * @bo: &amdgpu_bo buffer object to be unmapped |
| 829 | * |
| 830 | * Unmaps a kernel map set up by amdgpu_bo_kmap(). |
| 831 | */ |
| 832 | void amdgpu_bo_kunmap(struct amdgpu_bo *bo) |
| 833 | { |
| 834 | if (bo->kmap.bo) |
| 835 | ttm_bo_kunmap(&bo->kmap); |
| 836 | } |
| 837 | |
| 838 | /** |
| 839 | * amdgpu_bo_ref - reference an &amdgpu_bo buffer object |
| 840 | * @bo: &amdgpu_bo buffer object |
| 841 | * |
| 842 | * References the contained &ttm_buffer_object. |
| 843 | * |
| 844 | * Returns: |
| 845 | * a refcounted pointer to the &amdgpu_bo buffer object. |
| 846 | */ |
| 847 | struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo) |
| 848 | { |
| 849 | if (bo == NULL((void *)0)) |
| 850 | return NULL((void *)0); |
| 851 | |
| 852 | ttm_bo_get(&bo->tbo); |
| 853 | return bo; |
| 854 | } |
| 855 | |
| 856 | /** |
| 857 | * amdgpu_bo_unref - unreference an &amdgpu_bo buffer object |
| 858 | * @bo: &amdgpu_bo buffer object |
| 859 | * |
| 860 | * Unreferences the contained &ttm_buffer_object and clear the pointer |
| 861 | */ |
| 862 | void amdgpu_bo_unref(struct amdgpu_bo **bo) |
| 863 | { |
| 864 | struct ttm_buffer_object *tbo; |
| 865 | |
| 866 | if ((*bo) == NULL((void *)0)) |
| 867 | return; |
| 868 | |
| 869 | tbo = &((*bo)->tbo); |
| 870 | ttm_bo_put(tbo); |
| 871 | *bo = NULL((void *)0); |
| 872 | } |
| 873 | |
| 874 | /** |
| 875 | * amdgpu_bo_pin_restricted - pin an &amdgpu_bo buffer object |
| 876 | * @bo: &amdgpu_bo buffer object to be pinned |
| 877 | * @domain: domain to be pinned to |
| 878 | * @min_offset: the start of requested address range |
| 879 | * @max_offset: the end of requested address range |
| 880 | * |
| 881 | * Pins the buffer object according to requested domain and address range. If |
| 882 | * the memory is unbound gart memory, binds the pages into gart table. Adjusts |
| 883 | * pin_count and pin_size accordingly. |
| 884 | * |
| 885 | * Pinning means to lock pages in memory along with keeping them at a fixed |
| 886 | * offset. It is required when a buffer can not be moved, for example, when |
| 887 | * a display buffer is being scanned out. |
| 888 | * |
| 889 | * Compared with amdgpu_bo_pin(), this function gives more flexibility on |
| 890 | * where to pin a buffer if there are specific restrictions on where a buffer |
| 891 | * must be located. |
| 892 | * |
| 893 | * Returns: |
| 894 | * 0 for success or a negative error code on failure. |
| 895 | */ |
| 896 | int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain, |
| 897 | u64 min_offset, u64 max_offset) |
| 898 | { |
| 899 | struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); |
| 900 | struct ttm_operation_ctx ctx = { false0, false0 }; |
| 901 | int r, i; |
| 902 | |
| 903 | if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) |
| 904 | return -EPERM1; |
| 905 | |
| 906 | if (WARN_ON_ONCE(min_offset > max_offset)({ static int __warned; int __ret = !!(min_offset > max_offset ); if (__ret && !__warned) { printf("WARNING %s failed at %s:%d\n" , "min_offset > max_offset", "/usr/src/sys/dev/pci/drm/amd/amdgpu/amdgpu_object.c" , 906); __warned = 1; } __builtin_expect(!!(__ret), 0); })) |
| 907 | return -EINVAL22; |
| 908 | |
| 909 | /* A shared bo cannot be migrated to VRAM */ |
| 910 | if (bo->prime_shared_count) { |
| 911 | if (domain & AMDGPU_GEM_DOMAIN_GTT0x2) |
| 912 | domain = AMDGPU_GEM_DOMAIN_GTT0x2; |
| 913 | else |
| 914 | return -EINVAL22; |
| 915 | } |
| 916 | |
| 917 | /* This assumes only APU display buffers are pinned with (VRAM|GTT). |
| 918 | * See function amdgpu_display_supported_domains() |
| 919 | */ |
| 920 | domain = amdgpu_bo_get_preferred_pin_domain(adev, domain); |
| 921 | |
| 922 | if (bo->pin_count) { |
| 923 | uint32_t mem_type = bo->tbo.mem.mem_type; |
| 924 | |
| 925 | if (!(domain & amdgpu_mem_type_to_domain(mem_type))) |
| 926 | return -EINVAL22; |
| 927 | |
| 928 | bo->pin_count++; |
| 929 | |
| 930 | if (max_offset != 0) { |
| 931 | u64 domain_start = amdgpu_ttm_domain_start(adev, |
| 932 | mem_type); |
| 933 | WARN_ON_ONCE(max_offset <({ static int __warned; int __ret = !!(max_offset < (amdgpu_bo_gpu_offset (bo) - domain_start)); if (__ret && !__warned) { printf ("WARNING %s failed at %s:%d\n", "max_offset < (amdgpu_bo_gpu_offset(bo) - domain_start)" , "/usr/src/sys/dev/pci/drm/amd/amdgpu/amdgpu_object.c", 934) ; __warned = 1; } __builtin_expect(!!(__ret), 0); }) |
| 934 | (amdgpu_bo_gpu_offset(bo) - domain_start))({ static int __warned; int __ret = !!(max_offset < (amdgpu_bo_gpu_offset (bo) - domain_start)); if (__ret && !__warned) { printf ("WARNING %s failed at %s:%d\n", "max_offset < (amdgpu_bo_gpu_offset(bo) - domain_start)" , "/usr/src/sys/dev/pci/drm/amd/amdgpu/amdgpu_object.c", 934) ; __warned = 1; } __builtin_expect(!!(__ret), 0); }); |
| 935 | } |
| 936 | |
| 937 | return 0; |
| 938 | } |
| 939 | |
| 940 | #ifdef notyet |
| 941 | if (bo->tbo.base.import_attach) |
| 942 | dma_buf_pin(bo->tbo.base.import_attach); |
| 943 | #endif |
| 944 | |
| 945 | bo->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS(1 << 5); |
| 946 | /* force to pin into visible video ram */ |
| 947 | if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS(1 << 1))) |
| 948 | bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED(1 << 0); |
| 949 | amdgpu_bo_placement_from_domain(bo, domain); |
| 950 | for (i = 0; i < bo->placement.num_placement; i++) { |
| 951 | unsigned fpfn, lpfn; |
| 952 | |
| 953 | fpfn = min_offset >> PAGE_SHIFT12; |
| 954 | lpfn = max_offset >> PAGE_SHIFT12; |
| 955 | |
| 956 | if (fpfn > bo->placements[i].fpfn) |
| 957 | bo->placements[i].fpfn = fpfn; |
| 958 | if (!bo->placements[i].lpfn || |
| 959 | (lpfn && lpfn < bo->placements[i].lpfn)) |
| 960 | bo->placements[i].lpfn = lpfn; |
| 961 | bo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT(1 << 21); |
| 962 | } |
| 963 | |
| 964 | r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); |
| 965 | if (unlikely(r)__builtin_expect(!!(r), 0)) { |
| 966 | dev_err(adev->dev, "%p pin failed\n", bo)printf("drm:pid%d:%s *ERROR* " "%p pin failed\n", ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof (struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p-> ps_pid, __func__ , bo); |
| 967 | goto error; |
| 968 | } |
| 969 | |
| 970 | bo->pin_count = 1; |
| 971 | |
| 972 | domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type); |
| 973 | if (domain == AMDGPU_GEM_DOMAIN_VRAM0x4) { |
| 974 | atomic64_add(amdgpu_bo_size(bo), &adev->vram_pin_size)__sync_fetch_and_add_8(&adev->vram_pin_size, amdgpu_bo_size (bo)); |
| 975 | atomic64_add(amdgpu_vram_mgr_bo_visible_size(bo),__sync_fetch_and_add_8(&adev->visible_pin_size, amdgpu_vram_mgr_bo_visible_size (bo)) |
| 976 | &adev->visible_pin_size)__sync_fetch_and_add_8(&adev->visible_pin_size, amdgpu_vram_mgr_bo_visible_size (bo)); |
| 977 | } else if (domain == AMDGPU_GEM_DOMAIN_GTT0x2) { |
| 978 | atomic64_add(amdgpu_bo_size(bo), &adev->gart_pin_size)__sync_fetch_and_add_8(&adev->gart_pin_size, amdgpu_bo_size (bo)); |
| 979 | } |
| 980 | |
| 981 | error: |
| 982 | return r; |
| 983 | } |
| 984 | |
| 985 | /** |
| 986 | * amdgpu_bo_pin - pin an &amdgpu_bo buffer object |
| 987 | * @bo: &amdgpu_bo buffer object to be pinned |
| 988 | * @domain: domain to be pinned to |
| 989 | * |
| 990 | * A simple wrapper to amdgpu_bo_pin_restricted(). |
| 991 | * Provides a simpler API for buffers that do not have any strict restrictions |
| 992 | * on where a buffer must be located. |
| 993 | * |
| 994 | * Returns: |
| 995 | * 0 for success or a negative error code on failure. |
| 996 | */ |
| 997 | int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain) |
| 998 | { |
| 999 | return amdgpu_bo_pin_restricted(bo, domain, 0, 0); |
| 1000 | } |
| 1001 | |
| 1002 | /** |
| 1003 | * amdgpu_bo_unpin - unpin an &amdgpu_bo buffer object |
| 1004 | * @bo: &amdgpu_bo buffer object to be unpinned |
| 1005 | * |
| 1006 | * Decreases the pin_count, and clears the flags if pin_count reaches 0. |
| 1007 | * Changes placement and pin size accordingly. |
| 1008 | * |
| 1009 | * Returns: |
| 1010 | * 0 for success or a negative error code on failure. |
| 1011 | */ |
| 1012 | int amdgpu_bo_unpin(struct amdgpu_bo *bo) |
| 1013 | { |
| 1014 | struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); |
Value stored to 'adev' during its initialization is never read | |
| 1015 | struct ttm_operation_ctx ctx = { false0, false0 }; |
| 1016 | int r, i; |
| 1017 | |
| 1018 | if (WARN_ON_ONCE(!bo->pin_count)({ static int __warned; int __ret = !!(!bo->pin_count); if (__ret && !__warned) { printf("WARNING %s failed at %s:%d\n" , "!bo->pin_count", "/usr/src/sys/dev/pci/drm/amd/amdgpu/amdgpu_object.c" , 1018); __warned = 1; } __builtin_expect(!!(__ret), 0); })) { |
| 1019 | dev_warn(adev->dev, "%p unpin not necessary\n", bo)printf("drm:pid%d:%s *WARNING* " "%p unpin not necessary\n", ( {struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , bo); |
| 1020 | return 0; |
| 1021 | } |
| 1022 | bo->pin_count--; |
| 1023 | if (bo->pin_count) |
| 1024 | return 0; |
| 1025 | |
| 1026 | amdgpu_bo_subtract_pin_size(bo); |
| 1027 | |
| 1028 | #ifdef notyet |
| 1029 | if (bo->tbo.base.import_attach) |
| 1030 | dma_buf_unpin(bo->tbo.base.import_attach); |
| 1031 | #endif |
| 1032 | |
| 1033 | for (i = 0; i < bo->placement.num_placement; i++) { |
| 1034 | bo->placements[i].lpfn = 0; |
| 1035 | bo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT(1 << 21); |
| 1036 | } |
| 1037 | r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); |
| 1038 | if (unlikely(r)__builtin_expect(!!(r), 0)) |
| 1039 | dev_err(adev->dev, "%p validate failed for unpin\n", bo)printf("drm:pid%d:%s *ERROR* " "%p validate failed for unpin\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , bo); |
| 1040 | |
| 1041 | return r; |
| 1042 | } |
| 1043 | |
| 1044 | /** |
| 1045 | * amdgpu_bo_evict_vram - evict VRAM buffers |
| 1046 | * @adev: amdgpu device object |
| 1047 | * |
| 1048 | * Evicts all VRAM buffers on the lru list of the memory type. |
| 1049 | * Mainly used for evicting vram at suspend time. |
| 1050 | * |
| 1051 | * Returns: |
| 1052 | * 0 for success or a negative error code on failure. |
| 1053 | */ |
| 1054 | int amdgpu_bo_evict_vram(struct amdgpu_device *adev) |
| 1055 | { |
| 1056 | /* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */ |
| 1057 | #ifndef CONFIG_HIBERNATION |
| 1058 | if (adev->flags & AMD_IS_APU) { |
| 1059 | /* Useless to evict on IGP chips */ |
| 1060 | return 0; |
| 1061 | } |
| 1062 | #endif |
| 1063 | return ttm_bo_evict_mm(&adev->mman.bdev, TTM_PL_VRAM2); |
| 1064 | } |
| 1065 | |
| 1066 | static const char *amdgpu_vram_names[] = { |
| 1067 | "UNKNOWN", |
| 1068 | "GDDR1", |
| 1069 | "DDR2", |
| 1070 | "GDDR3", |
| 1071 | "GDDR4", |
| 1072 | "GDDR5", |
| 1073 | "HBM", |
| 1074 | "DDR3", |
| 1075 | "DDR4", |
| 1076 | "GDDR6", |
| 1077 | }; |
| 1078 | |
| 1079 | /** |
| 1080 | * amdgpu_bo_init - initialize memory manager |
| 1081 | * @adev: amdgpu device object |
| 1082 | * |
| 1083 | * Calls amdgpu_ttm_init() to initialize amdgpu memory manager. |
| 1084 | * |
| 1085 | * Returns: |
| 1086 | * 0 for success or a negative error code on failure. |
| 1087 | */ |
| 1088 | int amdgpu_bo_init(struct amdgpu_device *adev) |
| 1089 | { |
| 1090 | paddr_t start, end; |
| 1091 | |
| 1092 | #ifdef __linux__ |
| 1093 | /* reserve PAT memory space to WC for VRAM */ |
| 1094 | arch_io_reserve_memtype_wc(adev->gmc.aper_base, |
| 1095 | adev->gmc.aper_size); |
| 1096 | |
| 1097 | /* Add an MTRR for the VRAM */ |
| 1098 | adev->gmc.vram_mtrr = arch_phys_wc_add(adev->gmc.aper_base, |
| 1099 | adev->gmc.aper_size); |
| 1100 | #else |
| 1101 | drm_mtrr_add(adev->gmc.aper_base, adev->gmc.aper_size, DRM_MTRR_WC(1<<1)); |
| 1102 | |
| 1103 | start = atop(bus_space_mmap(adev->memt, adev->gmc.aper_base, 0, 0, 0))((((adev->memt)->mmap((adev->gmc.aper_base), (0), (0 ), (0)))) >> 12); |
| 1104 | end = start + atop(adev->gmc.aper_size)((adev->gmc.aper_size) >> 12); |
| 1105 | uvm_page_physload(start, end, start, end, PHYSLOAD_DEVICE0x01); |
| 1106 | #endif |
| 1107 | DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",printk("\0016" "[" "drm" "] " "Detected VRAM RAM=%lluM, BAR=%lluM\n" , adev->gmc.mc_vram_size >> 20, (unsigned long long) adev->gmc.aper_size >> 20) |
| 1108 | adev->gmc.mc_vram_size >> 20,printk("\0016" "[" "drm" "] " "Detected VRAM RAM=%lluM, BAR=%lluM\n" , adev->gmc.mc_vram_size >> 20, (unsigned long long) adev->gmc.aper_size >> 20) |
| 1109 | (unsigned long long)adev->gmc.aper_size >> 20)printk("\0016" "[" "drm" "] " "Detected VRAM RAM=%lluM, BAR=%lluM\n" , adev->gmc.mc_vram_size >> 20, (unsigned long long) adev->gmc.aper_size >> 20); |
| 1110 | DRM_INFO("RAM width %dbits %s\n",printk("\0016" "[" "drm" "] " "RAM width %dbits %s\n", adev-> gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]) |
| 1111 | adev->gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type])printk("\0016" "[" "drm" "] " "RAM width %dbits %s\n", adev-> gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]); |
| 1112 | return amdgpu_ttm_init(adev); |
| 1113 | } |
| 1114 | |
| 1115 | /** |
| 1116 | * amdgpu_bo_late_init - late init |
| 1117 | * @adev: amdgpu device object |
| 1118 | * |
| 1119 | * Calls amdgpu_ttm_late_init() to free resources used earlier during |
| 1120 | * initialization. |
| 1121 | * |
| 1122 | * Returns: |
| 1123 | * 0 for success or a negative error code on failure. |
| 1124 | */ |
| 1125 | int amdgpu_bo_late_init(struct amdgpu_device *adev) |
| 1126 | { |
| 1127 | amdgpu_ttm_late_init(adev); |
| 1128 | |
| 1129 | return 0; |
| 1130 | } |
| 1131 | |
| 1132 | /** |
| 1133 | * amdgpu_bo_fini - tear down memory manager |
| 1134 | * @adev: amdgpu device object |
| 1135 | * |
| 1136 | * Reverses amdgpu_bo_init() to tear down memory manager. |
| 1137 | */ |
| 1138 | void amdgpu_bo_fini(struct amdgpu_device *adev) |
| 1139 | { |
| 1140 | amdgpu_ttm_fini(adev); |
| 1141 | #ifdef __linux__ |
| 1142 | arch_phys_wc_del(adev->gmc.vram_mtrr); |
| 1143 | arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size); |
| 1144 | #else |
| 1145 | drm_mtrr_del(0, adev->gmc.aper_base, adev->gmc.aper_size, DRM_MTRR_WC(1<<1)); |
| 1146 | #endif |
| 1147 | } |
| 1148 | |
| 1149 | #ifdef notyet |
| 1150 | /** |
| 1151 | * amdgpu_bo_fbdev_mmap - mmap fbdev memory |
| 1152 | * @bo: &amdgpu_bo buffer object |
| 1153 | * @vma: vma as input from the fbdev mmap method |
| 1154 | * |
| 1155 | * Calls ttm_fbdev_mmap() to mmap fbdev memory if it is backed by a bo. |
| 1156 | * |
| 1157 | * Returns: |
| 1158 | * 0 for success or a negative error code on failure. |
| 1159 | */ |
| 1160 | int amdgpu_bo_fbdev_mmap(struct amdgpu_bo *bo, |
| 1161 | struct vm_area_struct *vma) |
| 1162 | { |
| 1163 | if (vma->vm_pgoff != 0) |
| 1164 | return -EACCES13; |
| 1165 | |
| 1166 | return ttm_bo_mmap_obj(vma, &bo->tbo); |
| 1167 | } |
| 1168 | #endif |
| 1169 | |
| 1170 | /** |
| 1171 | * amdgpu_bo_set_tiling_flags - set tiling flags |
| 1172 | * @bo: &amdgpu_bo buffer object |
| 1173 | * @tiling_flags: new flags |
| 1174 | * |
| 1175 | * Sets buffer object's tiling flags with the new one. Used by GEM ioctl or |
| 1176 | * kernel driver to set the tiling flags on a buffer. |
| 1177 | * |
| 1178 | * Returns: |
| 1179 | * 0 for success or a negative error code on failure. |
| 1180 | */ |
| 1181 | int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags) |
| 1182 | { |
| 1183 | struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); |
| 1184 | |
| 1185 | if (adev->family <= AMDGPU_FAMILY_CZ135 && |
| 1186 | AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT)(((__u64)(tiling_flags) >> 9) & 0x7) > 6) |
| 1187 | return -EINVAL22; |
| 1188 | |
| 1189 | bo->tiling_flags = tiling_flags; |
| 1190 | return 0; |
| 1191 | } |
| 1192 | |
| 1193 | /** |
| 1194 | * amdgpu_bo_get_tiling_flags - get tiling flags |
| 1195 | * @bo: &amdgpu_bo buffer object |
| 1196 | * @tiling_flags: returned flags |
| 1197 | * |
| 1198 | * Gets buffer object's tiling flags. Used by GEM ioctl or kernel driver to |
| 1199 | * set the tiling flags on a buffer. |
| 1200 | */ |
| 1201 | void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags) |
| 1202 | { |
| 1203 | dma_resv_assert_held(bo->tbo.base.resv)do { (void)(&(bo->tbo.base.resv)->lock.base); } while (0); |
| 1204 | |
| 1205 | if (tiling_flags) |
| 1206 | *tiling_flags = bo->tiling_flags; |
| 1207 | } |
| 1208 | |
| 1209 | /** |
| 1210 | * amdgpu_bo_set_metadata - set metadata |
| 1211 | * @bo: &amdgpu_bo buffer object |
| 1212 | * @metadata: new metadata |
| 1213 | * @metadata_size: size of the new metadata |
| 1214 | * @flags: flags of the new metadata |
| 1215 | * |
| 1216 | * Sets buffer object's metadata, its size and flags. |
| 1217 | * Used via GEM ioctl. |
| 1218 | * |
| 1219 | * Returns: |
| 1220 | * 0 for success or a negative error code on failure. |
| 1221 | */ |
| 1222 | int amdgpu_bo_set_metadata (struct amdgpu_bo *bo, void *metadata, |
| 1223 | uint32_t metadata_size, uint64_t flags) |
| 1224 | { |
| 1225 | void *buffer; |
| 1226 | |
| 1227 | if (!metadata_size) { |
| 1228 | if (bo->metadata_size) { |
| 1229 | kfree(bo->metadata); |
| 1230 | bo->metadata = NULL((void *)0); |
| 1231 | bo->metadata_size = 0; |
| 1232 | } |
| 1233 | return 0; |
| 1234 | } |
| 1235 | |
| 1236 | if (metadata == NULL((void *)0)) |
| 1237 | return -EINVAL22; |
| 1238 | |
| 1239 | buffer = kmemdup(metadata, metadata_size, GFP_KERNEL(0x0001 | 0x0004)); |
| 1240 | if (buffer == NULL((void *)0)) |
| 1241 | return -ENOMEM12; |
| 1242 | |
| 1243 | kfree(bo->metadata); |
| 1244 | bo->metadata_flags = flags; |
| 1245 | bo->metadata = buffer; |
| 1246 | bo->metadata_size = metadata_size; |
| 1247 | |
| 1248 | return 0; |
| 1249 | } |
| 1250 | |
| 1251 | /** |
| 1252 | * amdgpu_bo_get_metadata - get metadata |
| 1253 | * @bo: &amdgpu_bo buffer object |
| 1254 | * @buffer: returned metadata |
| 1255 | * @buffer_size: size of the buffer |
| 1256 | * @metadata_size: size of the returned metadata |
| 1257 | * @flags: flags of the returned metadata |
| 1258 | * |
| 1259 | * Gets buffer object's metadata, its size and flags. buffer_size shall not be |
| 1260 | * less than metadata_size. |
| 1261 | * Used via GEM ioctl. |
| 1262 | * |
| 1263 | * Returns: |
| 1264 | * 0 for success or a negative error code on failure. |
| 1265 | */ |
| 1266 | int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer, |
| 1267 | size_t buffer_size, uint32_t *metadata_size, |
| 1268 | uint64_t *flags) |
| 1269 | { |
| 1270 | if (!buffer && !metadata_size) |
| 1271 | return -EINVAL22; |
| 1272 | |
| 1273 | if (buffer) { |
| 1274 | if (buffer_size < bo->metadata_size) |
| 1275 | return -EINVAL22; |
| 1276 | |
| 1277 | if (bo->metadata_size) |
| 1278 | memcpy(buffer, bo->metadata, bo->metadata_size)__builtin_memcpy((buffer), (bo->metadata), (bo->metadata_size )); |
| 1279 | } |
| 1280 | |
| 1281 | if (metadata_size) |
| 1282 | *metadata_size = bo->metadata_size; |
| 1283 | if (flags) |
| 1284 | *flags = bo->metadata_flags; |
| 1285 | |
| 1286 | return 0; |
| 1287 | } |
| 1288 | |
| 1289 | /** |
| 1290 | * amdgpu_bo_move_notify - notification about a memory move |
| 1291 | * @bo: pointer to a buffer object |
| 1292 | * @evict: if this move is evicting the buffer from the graphics address space |
| 1293 | * @new_mem: new information of the bufer object |
| 1294 | * |
| 1295 | * Marks the corresponding &amdgpu_bo buffer object as invalid, also performs |
| 1296 | * bookkeeping. |
| 1297 | * TTM driver callback which is called when ttm moves a buffer. |
| 1298 | */ |
| 1299 | void amdgpu_bo_move_notify(struct ttm_buffer_object *bo, |
| 1300 | bool_Bool evict, |
| 1301 | struct ttm_resource *new_mem) |
| 1302 | { |
| 1303 | struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev); |
| 1304 | struct amdgpu_bo *abo; |
| 1305 | struct ttm_resource *old_mem = &bo->mem; |
| 1306 | |
| 1307 | if (!amdgpu_bo_is_amdgpu_bo(bo)) |
| 1308 | return; |
| 1309 | |
| 1310 | abo = ttm_to_amdgpu_bo(bo); |
| 1311 | amdgpu_vm_bo_invalidate(adev, abo, evict); |
| 1312 | |
| 1313 | amdgpu_bo_kunmap(abo); |
| 1314 | |
| 1315 | #ifdef notyet |
| 1316 | if (abo->tbo.base.dma_buf && !abo->tbo.base.import_attach && |
| 1317 | bo->mem.mem_type != TTM_PL_SYSTEM0) |
| 1318 | dma_buf_move_notify(abo->tbo.base.dma_buf); |
| 1319 | #endif |
| 1320 | |
| 1321 | /* remember the eviction */ |
| 1322 | if (evict) |
| 1323 | atomic64_inc(&adev->num_evictions)__sync_fetch_and_add_8(&adev->num_evictions, 1); |
| 1324 | |
| 1325 | /* update statistics */ |
| 1326 | if (!new_mem) |
| 1327 | return; |
| 1328 | |
| 1329 | /* move_notify is called before move happens */ |
| 1330 | trace_amdgpu_bo_move(abo, new_mem->mem_type, old_mem->mem_type); |
| 1331 | } |
| 1332 | |
| 1333 | /** |
| 1334 | * amdgpu_bo_release_notify - notification about a BO being released |
| 1335 | * @bo: pointer to a buffer object |
| 1336 | * |
| 1337 | * Wipes VRAM buffers whose contents should not be leaked before the |
| 1338 | * memory is released. |
| 1339 | */ |
| 1340 | void amdgpu_bo_release_notify(struct ttm_buffer_object *bo) |
| 1341 | { |
| 1342 | struct dma_fence *fence = NULL((void *)0); |
| 1343 | struct amdgpu_bo *abo; |
| 1344 | int r; |
| 1345 | |
| 1346 | if (!amdgpu_bo_is_amdgpu_bo(bo)) |
| 1347 | return; |
| 1348 | |
| 1349 | abo = ttm_to_amdgpu_bo(bo); |
| 1350 | |
| 1351 | if (abo->kfd_bo) |
| 1352 | amdgpu_amdkfd_unreserve_memory_limit(abo); |
| 1353 | |
| 1354 | /* We only remove the fence if the resv has individualized. */ |
| 1355 | WARN_ON_ONCE(bo->type == ttm_bo_type_kernel({ static int __warned; int __ret = !!(bo->type == ttm_bo_type_kernel && bo->base.resv != &bo->base._resv); if ( __ret && !__warned) { printf("WARNING %s failed at %s:%d\n" , "bo->type == ttm_bo_type_kernel && bo->base.resv != &bo->base._resv" , "/usr/src/sys/dev/pci/drm/amd/amdgpu/amdgpu_object.c", 1356 ); __warned = 1; } __builtin_expect(!!(__ret), 0); }) |
| 1356 | && bo->base.resv != &bo->base._resv)({ static int __warned; int __ret = !!(bo->type == ttm_bo_type_kernel && bo->base.resv != &bo->base._resv); if ( __ret && !__warned) { printf("WARNING %s failed at %s:%d\n" , "bo->type == ttm_bo_type_kernel && bo->base.resv != &bo->base._resv" , "/usr/src/sys/dev/pci/drm/amd/amdgpu/amdgpu_object.c", 1356 ); __warned = 1; } __builtin_expect(!!(__ret), 0); }); |
| 1357 | if (bo->base.resv == &bo->base._resv) |
| 1358 | amdgpu_amdkfd_remove_fence_on_pt_pd_bos(abo); |
| 1359 | |
| 1360 | if (bo->mem.mem_type != TTM_PL_VRAM2 || !bo->mem.mm_node || |
| 1361 | !(abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE(1 << 9))) |
| 1362 | return; |
| 1363 | |
| 1364 | dma_resv_lock(bo->base.resv, NULL((void *)0)); |
| 1365 | |
| 1366 | r = amdgpu_fill_buffer(abo, AMDGPU_POISON0xd0bed0be, bo->base.resv, &fence); |
| 1367 | if (!WARN_ON(r)({ int __ret = !!(r); if (__ret) printf("WARNING %s failed at %s:%d\n" , "r", "/usr/src/sys/dev/pci/drm/amd/amdgpu/amdgpu_object.c", 1367); __builtin_expect(!!(__ret), 0); })) { |
| 1368 | amdgpu_bo_fence(abo, fence, false0); |
| 1369 | dma_fence_put(fence); |
| 1370 | } |
| 1371 | |
| 1372 | dma_resv_unlock(bo->base.resv); |
| 1373 | } |
| 1374 | |
| 1375 | /** |
| 1376 | * amdgpu_bo_fault_reserve_notify - notification about a memory fault |
| 1377 | * @bo: pointer to a buffer object |
| 1378 | * |
| 1379 | * Notifies the driver we are taking a fault on this BO and have reserved it, |
| 1380 | * also performs bookkeeping. |
| 1381 | * TTM driver callback for dealing with vm faults. |
| 1382 | * |
| 1383 | * Returns: |
| 1384 | * 0 for success or a negative error code on failure. |
| 1385 | */ |
| 1386 | int amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo) |
| 1387 | { |
| 1388 | struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev); |
| 1389 | struct ttm_operation_ctx ctx = { false0, false0 }; |
| 1390 | struct amdgpu_bo *abo; |
| 1391 | unsigned long offset, size; |
| 1392 | int r; |
| 1393 | |
| 1394 | if (!amdgpu_bo_is_amdgpu_bo(bo)) |
| 1395 | return 0; |
| 1396 | |
| 1397 | abo = ttm_to_amdgpu_bo(bo); |
| 1398 | |
| 1399 | /* Remember that this BO was accessed by the CPU */ |
| 1400 | abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED(1 << 0); |
| 1401 | |
| 1402 | if (bo->mem.mem_type != TTM_PL_VRAM2) |
| 1403 | return 0; |
| 1404 | |
| 1405 | size = bo->mem.num_pages << PAGE_SHIFT12; |
| 1406 | offset = bo->mem.start << PAGE_SHIFT12; |
| 1407 | if ((offset + size) <= adev->gmc.visible_vram_size) |
| 1408 | return 0; |
| 1409 | |
| 1410 | /* Can't move a pinned BO to visible VRAM */ |
| 1411 | if (abo->pin_count > 0) |
| 1412 | return -EINVAL22; |
| 1413 | |
| 1414 | /* hurrah the memory is not visible ! */ |
| 1415 | atomic64_inc(&adev->num_vram_cpu_page_faults)__sync_fetch_and_add_8(&adev->num_vram_cpu_page_faults , 1); |
| 1416 | amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM0x4 | |
| 1417 | AMDGPU_GEM_DOMAIN_GTT0x2); |
| 1418 | |
| 1419 | /* Avoid costly evictions; only set GTT as a busy placement */ |
| 1420 | abo->placement.num_busy_placement = 1; |
| 1421 | abo->placement.busy_placement = &abo->placements[1]; |
| 1422 | |
| 1423 | r = ttm_bo_validate(bo, &abo->placement, &ctx); |
| 1424 | if (unlikely(r != 0)__builtin_expect(!!(r != 0), 0)) |
| 1425 | return r; |
| 1426 | |
| 1427 | offset = bo->mem.start << PAGE_SHIFT12; |
| 1428 | /* this should never happen */ |
| 1429 | if (bo->mem.mem_type == TTM_PL_VRAM2 && |
| 1430 | (offset + size) > adev->gmc.visible_vram_size) |
| 1431 | return -EINVAL22; |
| 1432 | |
| 1433 | return 0; |
| 1434 | } |
| 1435 | |
| 1436 | /** |
| 1437 | * amdgpu_bo_fence - add fence to buffer object |
| 1438 | * |
| 1439 | * @bo: buffer object in question |
| 1440 | * @fence: fence to add |
| 1441 | * @shared: true if fence should be added shared |
| 1442 | * |
| 1443 | */ |
| 1444 | void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence, |
| 1445 | bool_Bool shared) |
| 1446 | { |
| 1447 | struct dma_resv *resv = bo->tbo.base.resv; |
| 1448 | |
| 1449 | if (shared) |
| 1450 | dma_resv_add_shared_fence(resv, fence); |
| 1451 | else |
| 1452 | dma_resv_add_excl_fence(resv, fence); |
| 1453 | } |
| 1454 | |
| 1455 | /** |
| 1456 | * amdgpu_bo_sync_wait_resv - Wait for BO reservation fences |
| 1457 | * |
| 1458 | * @adev: amdgpu device pointer |
| 1459 | * @resv: reservation object to sync to |
| 1460 | * @sync_mode: synchronization mode |
| 1461 | * @owner: fence owner |
| 1462 | * @intr: Whether the wait is interruptible |
| 1463 | * |
| 1464 | * Extract the fences from the reservation object and waits for them to finish. |
| 1465 | * |
| 1466 | * Returns: |
| 1467 | * 0 on success, errno otherwise. |
| 1468 | */ |
| 1469 | int amdgpu_bo_sync_wait_resv(struct amdgpu_device *adev, struct dma_resv *resv, |
| 1470 | enum amdgpu_sync_mode sync_mode, void *owner, |
| 1471 | bool_Bool intr) |
| 1472 | { |
| 1473 | struct amdgpu_sync sync; |
| 1474 | int r; |
| 1475 | |
| 1476 | amdgpu_sync_create(&sync); |
| 1477 | amdgpu_sync_resv(adev, &sync, resv, sync_mode, owner); |
| 1478 | r = amdgpu_sync_wait(&sync, intr); |
| 1479 | amdgpu_sync_free(&sync); |
| 1480 | return r; |
| 1481 | } |
| 1482 | |
| 1483 | /** |
| 1484 | * amdgpu_bo_sync_wait - Wrapper for amdgpu_bo_sync_wait_resv |
| 1485 | * @bo: buffer object to wait for |
| 1486 | * @owner: fence owner |
| 1487 | * @intr: Whether the wait is interruptible |
| 1488 | * |
| 1489 | * Wrapper to wait for fences in a BO. |
| 1490 | * Returns: |
| 1491 | * 0 on success, errno otherwise. |
| 1492 | */ |
| 1493 | int amdgpu_bo_sync_wait(struct amdgpu_bo *bo, void *owner, bool_Bool intr) |
| 1494 | { |
| 1495 | struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); |
| 1496 | |
| 1497 | return amdgpu_bo_sync_wait_resv(adev, bo->tbo.base.resv, |
| 1498 | AMDGPU_SYNC_NE_OWNER, owner, intr); |
| 1499 | } |
| 1500 | |
| 1501 | /** |
| 1502 | * amdgpu_bo_gpu_offset - return GPU offset of bo |
| 1503 | * @bo: amdgpu object for which we query the offset |
| 1504 | * |
| 1505 | * Note: object should either be pinned or reserved when calling this |
| 1506 | * function, it might be useful to add check for this for debugging. |
| 1507 | * |
| 1508 | * Returns: |
| 1509 | * current GPU offset of the object. |
| 1510 | */ |
| 1511 | u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo) |
| 1512 | { |
| 1513 | WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_SYSTEM)({ static int __warned; int __ret = !!(bo->tbo.mem.mem_type == 0); if (__ret && !__warned) { printf("WARNING %s failed at %s:%d\n" , "bo->tbo.mem.mem_type == 0", "/usr/src/sys/dev/pci/drm/amd/amdgpu/amdgpu_object.c" , 1513); __warned = 1; } __builtin_expect(!!(__ret), 0); }); |
| 1514 | WARN_ON_ONCE(!dma_resv_is_locked(bo->tbo.base.resv) &&({ static int __warned; int __ret = !!(!dma_resv_is_locked(bo ->tbo.base.resv) && !bo->pin_count && bo ->tbo.type != ttm_bo_type_kernel); if (__ret && !__warned ) { printf("WARNING %s failed at %s:%d\n", "!dma_resv_is_locked(bo->tbo.base.resv) && !bo->pin_count && bo->tbo.type != ttm_bo_type_kernel" , "/usr/src/sys/dev/pci/drm/amd/amdgpu/amdgpu_object.c", 1515 ); __warned = 1; } __builtin_expect(!!(__ret), 0); }) |
| 1515 | !bo->pin_count && bo->tbo.type != ttm_bo_type_kernel)({ static int __warned; int __ret = !!(!dma_resv_is_locked(bo ->tbo.base.resv) && !bo->pin_count && bo ->tbo.type != ttm_bo_type_kernel); if (__ret && !__warned ) { printf("WARNING %s failed at %s:%d\n", "!dma_resv_is_locked(bo->tbo.base.resv) && !bo->pin_count && bo->tbo.type != ttm_bo_type_kernel" , "/usr/src/sys/dev/pci/drm/amd/amdgpu/amdgpu_object.c", 1515 ); __warned = 1; } __builtin_expect(!!(__ret), 0); }); |
| 1516 | WARN_ON_ONCE(bo->tbo.mem.start == AMDGPU_BO_INVALID_OFFSET)({ static int __warned; int __ret = !!(bo->tbo.mem.start == 0x7fffffffffffffffL); if (__ret && !__warned) { printf ("WARNING %s failed at %s:%d\n", "bo->tbo.mem.start == 0x7fffffffffffffffL" , "/usr/src/sys/dev/pci/drm/amd/amdgpu/amdgpu_object.c", 1516 ); __warned = 1; } __builtin_expect(!!(__ret), 0); }); |
| 1517 | WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_VRAM &&({ static int __warned; int __ret = !!(bo->tbo.mem.mem_type == 2 && !(bo->flags & (1 << 5))); if (__ret && !__warned) { printf("WARNING %s failed at %s:%d\n" , "bo->tbo.mem.mem_type == 2 && !(bo->flags & (1 << 5))" , "/usr/src/sys/dev/pci/drm/amd/amdgpu/amdgpu_object.c", 1518 ); __warned = 1; } __builtin_expect(!!(__ret), 0); }) |
| 1518 | !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS))({ static int __warned; int __ret = !!(bo->tbo.mem.mem_type == 2 && !(bo->flags & (1 << 5))); if (__ret && !__warned) { printf("WARNING %s failed at %s:%d\n" , "bo->tbo.mem.mem_type == 2 && !(bo->flags & (1 << 5))" , "/usr/src/sys/dev/pci/drm/amd/amdgpu/amdgpu_object.c", 1518 ); __warned = 1; } __builtin_expect(!!(__ret), 0); }); |
| 1519 | |
| 1520 | return amdgpu_bo_gpu_offset_no_check(bo); |
| 1521 | } |
| 1522 | |
| 1523 | /** |
| 1524 | * amdgpu_bo_gpu_offset_no_check - return GPU offset of bo |
| 1525 | * @bo: amdgpu object for which we query the offset |
| 1526 | * |
| 1527 | * Returns: |
| 1528 | * current GPU offset of the object without raising warnings. |
| 1529 | */ |
| 1530 | u64 amdgpu_bo_gpu_offset_no_check(struct amdgpu_bo *bo) |
| 1531 | { |
| 1532 | struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); |
| 1533 | uint64_t offset; |
| 1534 | |
| 1535 | offset = (bo->tbo.mem.start << PAGE_SHIFT12) + |
| 1536 | amdgpu_ttm_domain_start(adev, bo->tbo.mem.mem_type); |
| 1537 | |
| 1538 | return amdgpu_gmc_sign_extend(offset); |
| 1539 | } |
| 1540 | |
| 1541 | /** |
| 1542 | * amdgpu_bo_get_preferred_pin_domain - get preferred domain for scanout |
| 1543 | * @adev: amdgpu device object |
| 1544 | * @domain: allowed :ref:`memory domains <amdgpu_memory_domains>` |
| 1545 | * |
| 1546 | * Returns: |
| 1547 | * Which of the allowed domains is preferred for pinning the BO for scanout. |
| 1548 | */ |
| 1549 | uint32_t amdgpu_bo_get_preferred_pin_domain(struct amdgpu_device *adev, |
| 1550 | uint32_t domain) |
| 1551 | { |
| 1552 | if (domain == (AMDGPU_GEM_DOMAIN_VRAM0x4 | AMDGPU_GEM_DOMAIN_GTT0x2)) { |
| 1553 | domain = AMDGPU_GEM_DOMAIN_VRAM0x4; |
| 1554 | if (adev->gmc.real_vram_size <= AMDGPU_SG_THRESHOLD(256*1024*1024)) |
| 1555 | domain = AMDGPU_GEM_DOMAIN_GTT0x2; |
| 1556 | } |
| 1557 | return domain; |
| 1558 | } |