1 /* SPDX-License-Identifier: MIT */ 2 /* 3 * Copyright © 2021 Intel Corporation 4 */ 5 6 #ifndef _XE_BO_H_ 7 #define _XE_BO_H_ 8 9 #include <drm/ttm/ttm_tt.h> 10 11 #include "xe_bo_types.h" 12 #include "xe_macros.h" 13 #include "xe_vm_types.h" 14 #include "xe_vm.h" 15 16 #define XE_DEFAULT_GTT_SIZE_MB 3072ULL /* 3GB by default */ 17 18 #define XE_BO_FLAG_USER BIT(0) 19 /* The bits below need to be contiguous, or things break */ 20 #define XE_BO_FLAG_SYSTEM BIT(1) 21 #define XE_BO_FLAG_VRAM0 BIT(2) 22 #define XE_BO_FLAG_VRAM1 BIT(3) 23 #define XE_BO_FLAG_VRAM_MASK (XE_BO_FLAG_VRAM0 | XE_BO_FLAG_VRAM1) 24 /* -- */ 25 #define XE_BO_FLAG_STOLEN BIT(4) 26 #define XE_BO_FLAG_VRAM_IF_DGFX(tile) (IS_DGFX(tile_to_xe(tile)) ? \ 27 XE_BO_FLAG_VRAM0 << (tile)->id : \ 28 XE_BO_FLAG_SYSTEM) 29 #define XE_BO_FLAG_GGTT BIT(5) 30 #define XE_BO_FLAG_IGNORE_MIN_PAGE_SIZE BIT(6) 31 #define XE_BO_FLAG_PINNED BIT(7) 32 #define XE_BO_FLAG_NO_RESV_EVICT BIT(8) 33 #define XE_BO_FLAG_DEFER_BACKING BIT(9) 34 #define XE_BO_FLAG_SCANOUT BIT(10) 35 #define XE_BO_FLAG_FIXED_PLACEMENT BIT(11) 36 #define XE_BO_FLAG_PAGETABLE BIT(12) 37 #define XE_BO_FLAG_NEEDS_CPU_ACCESS BIT(13) 38 #define XE_BO_FLAG_NEEDS_UC BIT(14) 39 #define XE_BO_FLAG_NEEDS_64K BIT(15) 40 #define XE_BO_FLAG_NEEDS_2M BIT(16) 41 #define XE_BO_FLAG_GGTT_INVALIDATE BIT(17) 42 #define XE_BO_FLAG_GGTT0 BIT(18) 43 #define XE_BO_FLAG_GGTT1 BIT(19) 44 #define XE_BO_FLAG_GGTT2 BIT(20) 45 #define XE_BO_FLAG_GGTT3 BIT(21) 46 #define XE_BO_FLAG_GGTT_ALL (XE_BO_FLAG_GGTT0 | \ 47 XE_BO_FLAG_GGTT1 | \ 48 XE_BO_FLAG_GGTT2 | \ 49 XE_BO_FLAG_GGTT3) 50 #define XE_BO_FLAG_CPU_ADDR_MIRROR BIT(22) 51 52 /* this one is trigger internally only */ 53 #define XE_BO_FLAG_INTERNAL_TEST BIT(30) 54 #define XE_BO_FLAG_INTERNAL_64K BIT(31) 55 56 #define XE_BO_FLAG_GGTTx(tile) \ 57 (XE_BO_FLAG_GGTT0 << (tile)->id) 58 59 #define XE_PTE_SHIFT 12 60 #define XE_PAGE_SIZE (1 << XE_PTE_SHIFT) 61 #define XE_PTE_MASK (XE_PAGE_SIZE - 1) 62 #define XE_PDE_SHIFT (XE_PTE_SHIFT - 3) 63 #define XE_PDES (1 << XE_PDE_SHIFT) 64 #define XE_PDE_MASK (XE_PDES - 1) 65 66 #define XE_64K_PTE_SHIFT 16 67 #define XE_64K_PAGE_SIZE (1 << XE_64K_PTE_SHIFT) 68 #define XE_64K_PTE_MASK (XE_64K_PAGE_SIZE - 1) 69 #define XE_64K_PDE_MASK (XE_PDE_MASK >> 4) 70 71 #define XE_PL_SYSTEM TTM_PL_SYSTEM 72 #define XE_PL_TT TTM_PL_TT 73 #define XE_PL_VRAM0 TTM_PL_VRAM 74 #define XE_PL_VRAM1 (XE_PL_VRAM0 + 1) 75 #define XE_PL_STOLEN (TTM_NUM_MEM_TYPES - 1) 76 77 #define XE_BO_PROPS_INVALID (-1) 78 79 #define XE_PCI_BARRIER_MMAP_OFFSET (0x50 << XE_PTE_SHIFT) 80 81 struct sg_table; 82 83 struct xe_bo *xe_bo_alloc(void); 84 void xe_bo_free(struct xe_bo *bo); 85 86 struct xe_bo *___xe_bo_create_locked(struct xe_device *xe, struct xe_bo *bo, 87 struct xe_tile *tile, struct dma_resv *resv, 88 struct ttm_lru_bulk_move *bulk, size_t size, 89 u16 cpu_caching, enum ttm_bo_type type, 90 u32 flags); 91 struct xe_bo * 92 xe_bo_create_locked_range(struct xe_device *xe, 93 struct xe_tile *tile, struct xe_vm *vm, 94 size_t size, u64 start, u64 end, 95 enum ttm_bo_type type, u32 flags, u64 alignment); 96 struct xe_bo *xe_bo_create_locked(struct xe_device *xe, struct xe_tile *tile, 97 struct xe_vm *vm, size_t size, 98 enum ttm_bo_type type, u32 flags); 99 struct xe_bo *xe_bo_create(struct xe_device *xe, struct xe_tile *tile, 100 struct xe_vm *vm, size_t size, 101 enum ttm_bo_type type, u32 flags); 102 struct xe_bo *xe_bo_create_user(struct xe_device *xe, struct xe_tile *tile, 103 struct xe_vm *vm, size_t size, 104 u16 cpu_caching, 105 u32 flags); 106 struct xe_bo *xe_bo_create_pin_map(struct xe_device *xe, struct xe_tile *tile, 107 struct xe_vm *vm, size_t size, 108 enum ttm_bo_type type, u32 flags); 109 struct xe_bo *xe_bo_create_pin_map_at(struct xe_device *xe, struct xe_tile *tile, 110 struct xe_vm *vm, size_t size, u64 offset, 111 enum ttm_bo_type type, u32 flags); 112 struct xe_bo *xe_bo_create_pin_map_at_aligned(struct xe_device *xe, 113 struct xe_tile *tile, 114 struct xe_vm *vm, 115 size_t size, u64 offset, 116 enum ttm_bo_type type, u32 flags, 117 u64 alignment); 118 struct xe_bo *xe_bo_create_from_data(struct xe_device *xe, struct xe_tile *tile, 119 const void *data, size_t size, 120 enum ttm_bo_type type, u32 flags); 121 struct xe_bo *xe_managed_bo_create_pin_map(struct xe_device *xe, struct xe_tile *tile, 122 size_t size, u32 flags); 123 struct xe_bo *xe_managed_bo_create_from_data(struct xe_device *xe, struct xe_tile *tile, 124 const void *data, size_t size, u32 flags); 125 int xe_managed_bo_reinit_in_vram(struct xe_device *xe, struct xe_tile *tile, struct xe_bo **src); 126 127 int xe_bo_placement_for_flags(struct xe_device *xe, struct xe_bo *bo, 128 u32 bo_flags); 129 130 static inline struct xe_bo *ttm_to_xe_bo(const struct ttm_buffer_object *bo) 131 { 132 return container_of(bo, struct xe_bo, ttm); 133 } 134 135 static inline struct xe_bo *gem_to_xe_bo(const struct drm_gem_object *obj) 136 { 137 return container_of(obj, struct xe_bo, ttm.base); 138 } 139 140 #define xe_bo_device(bo) ttm_to_xe_device((bo)->ttm.bdev) 141 142 static inline struct xe_bo *xe_bo_get(struct xe_bo *bo) 143 { 144 if (bo) 145 drm_gem_object_get(&bo->ttm.base); 146 147 return bo; 148 } 149 150 void xe_bo_put(struct xe_bo *bo); 151 152 /* 153 * xe_bo_get_unless_zero() - Conditionally obtain a GEM object refcount on an 154 * xe bo 155 * @bo: The bo for which we want to obtain a refcount. 156 * 157 * There is a short window between where the bo's GEM object refcount reaches 158 * zero and where we put the final ttm_bo reference. Code in the eviction- and 159 * shrinking path should therefore attempt to grab a gem object reference before 160 * trying to use members outside of the base class ttm object. This function is 161 * intended for that purpose. On successful return, this function must be paired 162 * with an xe_bo_put(). 163 * 164 * Return: @bo on success, NULL on failure. 165 */ 166 static inline __must_check struct xe_bo *xe_bo_get_unless_zero(struct xe_bo *bo) 167 { 168 if (!bo || !kref_get_unless_zero(&bo->ttm.base.refcount)) 169 return NULL; 170 171 return bo; 172 } 173 174 static inline void __xe_bo_unset_bulk_move(struct xe_bo *bo) 175 { 176 if (bo) 177 ttm_bo_set_bulk_move(&bo->ttm, NULL); 178 } 179 180 static inline void xe_bo_assert_held(struct xe_bo *bo) 181 { 182 if (bo) 183 dma_resv_assert_held((bo)->ttm.base.resv); 184 } 185 186 int xe_bo_lock(struct xe_bo *bo, bool intr); 187 188 void xe_bo_unlock(struct xe_bo *bo); 189 190 static inline void xe_bo_unlock_vm_held(struct xe_bo *bo) 191 { 192 if (bo) { 193 XE_WARN_ON(bo->vm && bo->ttm.base.resv != xe_vm_resv(bo->vm)); 194 if (bo->vm) 195 xe_vm_assert_held(bo->vm); 196 else 197 dma_resv_unlock(bo->ttm.base.resv); 198 } 199 } 200 201 int xe_bo_pin_external(struct xe_bo *bo); 202 int xe_bo_pin(struct xe_bo *bo); 203 void xe_bo_unpin_external(struct xe_bo *bo); 204 void xe_bo_unpin(struct xe_bo *bo); 205 int xe_bo_validate(struct xe_bo *bo, struct xe_vm *vm, bool allow_res_evict); 206 207 static inline bool xe_bo_is_pinned(struct xe_bo *bo) 208 { 209 return bo->ttm.pin_count; 210 } 211 212 static inline bool xe_bo_is_protected(const struct xe_bo *bo) 213 { 214 return bo->pxp_key_instance; 215 } 216 217 static inline void xe_bo_unpin_map_no_vm(struct xe_bo *bo) 218 { 219 if (likely(bo)) { 220 xe_bo_lock(bo, false); 221 xe_bo_unpin(bo); 222 xe_bo_unlock(bo); 223 224 xe_bo_put(bo); 225 } 226 } 227 228 bool xe_bo_is_xe_bo(struct ttm_buffer_object *bo); 229 dma_addr_t __xe_bo_addr(struct xe_bo *bo, u64 offset, size_t page_size); 230 dma_addr_t xe_bo_addr(struct xe_bo *bo, u64 offset, size_t page_size); 231 232 static inline dma_addr_t 233 xe_bo_main_addr(struct xe_bo *bo, size_t page_size) 234 { 235 return xe_bo_addr(bo, 0, page_size); 236 } 237 238 static inline u32 239 __xe_bo_ggtt_addr(struct xe_bo *bo, u8 tile_id) 240 { 241 struct xe_ggtt_node *ggtt_node = bo->ggtt_node[tile_id]; 242 243 if (XE_WARN_ON(!ggtt_node)) 244 return 0; 245 246 XE_WARN_ON(ggtt_node->base.size > bo->size); 247 XE_WARN_ON(ggtt_node->base.start + ggtt_node->base.size > (1ull << 32)); 248 return ggtt_node->base.start; 249 } 250 251 static inline u32 252 xe_bo_ggtt_addr(struct xe_bo *bo) 253 { 254 xe_assert(xe_bo_device(bo), bo->tile); 255 256 return __xe_bo_ggtt_addr(bo, bo->tile->id); 257 } 258 259 int xe_bo_vmap(struct xe_bo *bo); 260 void xe_bo_vunmap(struct xe_bo *bo); 261 int xe_bo_read(struct xe_bo *bo, u64 offset, void *dst, int size); 262 263 bool mem_type_is_vram(u32 mem_type); 264 bool xe_bo_is_vram(struct xe_bo *bo); 265 bool xe_bo_is_stolen(struct xe_bo *bo); 266 bool xe_bo_is_stolen_devmem(struct xe_bo *bo); 267 bool xe_bo_is_vm_bound(struct xe_bo *bo); 268 bool xe_bo_has_single_placement(struct xe_bo *bo); 269 uint64_t vram_region_gpu_offset(struct ttm_resource *res); 270 271 bool xe_bo_can_migrate(struct xe_bo *bo, u32 mem_type); 272 273 int xe_bo_migrate(struct xe_bo *bo, u32 mem_type); 274 int xe_bo_evict(struct xe_bo *bo, bool force_alloc); 275 276 int xe_bo_evict_pinned(struct xe_bo *bo); 277 int xe_bo_restore_pinned(struct xe_bo *bo); 278 279 extern const struct ttm_device_funcs xe_ttm_funcs; 280 extern const char *const xe_mem_type_to_name[]; 281 282 int xe_gem_create_ioctl(struct drm_device *dev, void *data, 283 struct drm_file *file); 284 int xe_gem_mmap_offset_ioctl(struct drm_device *dev, void *data, 285 struct drm_file *file); 286 void xe_bo_runtime_pm_release_mmap_offset(struct xe_bo *bo); 287 288 int xe_bo_dumb_create(struct drm_file *file_priv, 289 struct drm_device *dev, 290 struct drm_mode_create_dumb *args); 291 292 bool xe_bo_needs_ccs_pages(struct xe_bo *bo); 293 294 static inline size_t xe_bo_ccs_pages_start(struct xe_bo *bo) 295 { 296 return PAGE_ALIGN(bo->ttm.base.size); 297 } 298 299 static inline bool xe_bo_has_pages(struct xe_bo *bo) 300 { 301 if ((bo->ttm.ttm && ttm_tt_is_populated(bo->ttm.ttm)) || 302 xe_bo_is_vram(bo)) 303 return true; 304 305 return false; 306 } 307 308 void __xe_bo_release_dummy(struct kref *kref); 309 310 /** 311 * xe_bo_put_deferred() - Put a buffer object with delayed final freeing 312 * @bo: The bo to put. 313 * @deferred: List to which to add the buffer object if we cannot put, or 314 * NULL if the function is to put unconditionally. 315 * 316 * Since the final freeing of an object includes both sleeping and (!) 317 * memory allocation in the dma_resv individualization, it's not ok 318 * to put an object from atomic context nor from within a held lock 319 * tainted by reclaim. In such situations we want to defer the final 320 * freeing until we've exited the restricting context, or in the worst 321 * case to a workqueue. 322 * This function either puts the object if possible without the refcount 323 * reaching zero, or adds it to the @deferred list if that was not possible. 324 * The caller needs to follow up with a call to xe_bo_put_commit() to actually 325 * put the bo iff this function returns true. It's safe to always 326 * follow up with a call to xe_bo_put_commit(). 327 * TODO: It's TTM that is the villain here. Perhaps TTM should add an 328 * interface like this. 329 * 330 * Return: true if @bo was the first object put on the @freed list, 331 * false otherwise. 332 */ 333 static inline bool 334 xe_bo_put_deferred(struct xe_bo *bo, struct llist_head *deferred) 335 { 336 if (!deferred) { 337 xe_bo_put(bo); 338 return false; 339 } 340 341 if (!kref_put(&bo->ttm.base.refcount, __xe_bo_release_dummy)) 342 return false; 343 344 return llist_add(&bo->freed, deferred); 345 } 346 347 void xe_bo_put_commit(struct llist_head *deferred); 348 349 /** 350 * xe_bo_put_async() - Put BO async 351 * @bo: The bo to put. 352 * 353 * Put BO async, the final put is deferred to a worker to exit an IRQ context. 354 */ 355 static inline void 356 xe_bo_put_async(struct xe_bo *bo) 357 { 358 struct xe_bo_dev *bo_device = &xe_bo_device(bo)->bo_device; 359 360 if (xe_bo_put_deferred(bo, &bo_device->async_list)) 361 schedule_work(&bo_device->async_free); 362 } 363 364 void xe_bo_dev_init(struct xe_bo_dev *bo_device); 365 366 void xe_bo_dev_fini(struct xe_bo_dev *bo_device); 367 368 struct sg_table *xe_bo_sg(struct xe_bo *bo); 369 370 /* 371 * xe_sg_segment_size() - Provides upper limit for sg segment size. 372 * @dev: device pointer 373 * 374 * Returns the maximum segment size for the 'struct scatterlist' 375 * elements. 376 */ 377 static inline unsigned int xe_sg_segment_size(struct device *dev) 378 { 379 struct scatterlist __maybe_unused sg; 380 size_t max = BIT_ULL(sizeof(sg.length) * 8) - 1; 381 382 max = min_t(size_t, max, dma_max_mapping_size(dev)); 383 384 /* 385 * The iommu_dma_map_sg() function ensures iova allocation doesn't 386 * cross dma segment boundary. It does so by padding some sg elements. 387 * This can cause overflow, ending up with sg->length being set to 0. 388 * Avoid this by ensuring maximum segment size is half of 'max' 389 * rounded down to PAGE_SIZE. 390 */ 391 return round_down(max / 2, PAGE_SIZE); 392 } 393 394 /** 395 * struct xe_bo_shrink_flags - flags governing the shrink behaviour. 396 * @purge: Only purging allowed. Don't shrink if bo not purgeable. 397 * @writeback: Attempt to immediately move content to swap. 398 */ 399 struct xe_bo_shrink_flags { 400 u32 purge : 1; 401 u32 writeback : 1; 402 }; 403 404 long xe_bo_shrink(struct ttm_operation_ctx *ctx, struct ttm_buffer_object *bo, 405 const struct xe_bo_shrink_flags flags, 406 unsigned long *scanned); 407 408 /** 409 * xe_bo_is_mem_type - Whether the bo currently resides in the given 410 * TTM memory type 411 * @bo: The bo to check. 412 * @mem_type: The TTM memory type. 413 * 414 * Return: true iff the bo resides in @mem_type, false otherwise. 415 */ 416 static inline bool xe_bo_is_mem_type(struct xe_bo *bo, u32 mem_type) 417 { 418 xe_bo_assert_held(bo); 419 return bo->ttm.resource->mem_type == mem_type; 420 } 421 #endif 422