| 1 | // SPDX-License-Identifier: GPL-2.0-only OR MIT |
| 2 | /* Copyright (c) 2023 Imagination Technologies Ltd. */ |
| 3 | |
| 4 | #include "pvr_device.h" |
| 5 | #include "pvr_gem.h" |
| 6 | #include "pvr_vm.h" |
| 7 | |
| 8 | #include <drm/drm_gem.h> |
| 9 | #include <drm/drm_prime.h> |
| 10 | |
| 11 | #include <linux/compiler.h> |
| 12 | #include <linux/compiler_attributes.h> |
| 13 | #include <linux/dma-buf.h> |
| 14 | #include <linux/dma-direction.h> |
| 15 | #include <linux/dma-mapping.h> |
| 16 | #include <linux/err.h> |
| 17 | #include <linux/gfp.h> |
| 18 | #include <linux/iosys-map.h> |
| 19 | #include <linux/log2.h> |
| 20 | #include <linux/mutex.h> |
| 21 | #include <linux/pagemap.h> |
| 22 | #include <linux/property.h> |
| 23 | #include <linux/refcount.h> |
| 24 | #include <linux/scatterlist.h> |
| 25 | |
| 26 | static void pvr_gem_object_free(struct drm_gem_object *obj) |
| 27 | { |
| 28 | drm_gem_shmem_object_free(obj); |
| 29 | } |
| 30 | |
| 31 | static struct dma_buf *pvr_gem_export(struct drm_gem_object *obj, int flags) |
| 32 | { |
| 33 | struct pvr_gem_object *pvr_obj = gem_to_pvr_gem(obj); |
| 34 | |
| 35 | if (pvr_obj->flags & DRM_PVR_BO_PM_FW_PROTECT) |
| 36 | return ERR_PTR(error: -EPERM); |
| 37 | |
| 38 | return drm_gem_prime_export(obj, flags); |
| 39 | } |
| 40 | |
| 41 | static int pvr_gem_mmap(struct drm_gem_object *gem_obj, struct vm_area_struct *vma) |
| 42 | { |
| 43 | struct pvr_gem_object *pvr_obj = gem_to_pvr_gem(gem_obj); |
| 44 | struct drm_gem_shmem_object *shmem_obj = shmem_gem_from_pvr_gem(pvr_obj); |
| 45 | |
| 46 | if (!(pvr_obj->flags & DRM_PVR_BO_ALLOW_CPU_USERSPACE_ACCESS)) |
| 47 | return -EINVAL; |
| 48 | |
| 49 | return drm_gem_shmem_mmap(shmem: shmem_obj, vma); |
| 50 | } |
| 51 | |
| 52 | static const struct drm_gem_object_funcs pvr_gem_object_funcs = { |
| 53 | .free = pvr_gem_object_free, |
| 54 | .print_info = drm_gem_shmem_object_print_info, |
| 55 | .export = pvr_gem_export, |
| 56 | .pin = drm_gem_shmem_object_pin, |
| 57 | .unpin = drm_gem_shmem_object_unpin, |
| 58 | .get_sg_table = drm_gem_shmem_object_get_sg_table, |
| 59 | .vmap = drm_gem_shmem_object_vmap, |
| 60 | .vunmap = drm_gem_shmem_object_vunmap, |
| 61 | .mmap = pvr_gem_mmap, |
| 62 | .vm_ops = &drm_gem_shmem_vm_ops, |
| 63 | }; |
| 64 | |
| 65 | /** |
| 66 | * pvr_gem_object_flags_validate() - Verify that a collection of PowerVR GEM |
| 67 | * mapping and/or creation flags form a valid combination. |
| 68 | * @flags: PowerVR GEM mapping/creation flags to validate. |
| 69 | * |
| 70 | * This function explicitly allows kernel-only flags. All ioctl entrypoints |
| 71 | * should do their own validation as well as relying on this function. |
| 72 | * |
| 73 | * Return: |
| 74 | * * %true if @flags contains valid mapping and/or creation flags, or |
| 75 | * * %false otherwise. |
| 76 | */ |
| 77 | static bool |
| 78 | pvr_gem_object_flags_validate(u64 flags) |
| 79 | { |
| 80 | static const u64 invalid_combinations[] = { |
| 81 | /* |
| 82 | * Memory flagged as PM/FW-protected cannot be mapped to |
| 83 | * userspace. To make this explicit, we require that the two |
| 84 | * flags allowing each of these respective features are never |
| 85 | * specified together. |
| 86 | */ |
| 87 | (DRM_PVR_BO_PM_FW_PROTECT | |
| 88 | DRM_PVR_BO_ALLOW_CPU_USERSPACE_ACCESS), |
| 89 | }; |
| 90 | |
| 91 | /* |
| 92 | * Check for bits set in undefined regions. Reserved regions refer to |
| 93 | * options that can only be set by the kernel. These are explicitly |
| 94 | * allowed in most cases, and must be checked specifically in IOCTL |
| 95 | * callback code. |
| 96 | */ |
| 97 | if ((flags & PVR_BO_UNDEFINED_MASK) != 0) |
| 98 | return false; |
| 99 | |
| 100 | /* |
| 101 | * Check for all combinations of flags marked as invalid in the array |
| 102 | * above. |
| 103 | */ |
| 104 | for (int i = 0; i < ARRAY_SIZE(invalid_combinations); ++i) { |
| 105 | u64 combo = invalid_combinations[i]; |
| 106 | |
| 107 | if ((flags & combo) == combo) |
| 108 | return false; |
| 109 | } |
| 110 | |
| 111 | return true; |
| 112 | } |
| 113 | |
| 114 | /** |
| 115 | * pvr_gem_object_into_handle() - Convert a reference to an object into a |
| 116 | * userspace-accessible handle. |
| 117 | * @pvr_obj: [IN] Target PowerVR-specific object. |
| 118 | * @pvr_file: [IN] File to associate the handle with. |
| 119 | * @handle: [OUT] Pointer to store the created handle in. Remains unmodified if |
| 120 | * an error is encountered. |
| 121 | * |
| 122 | * If an error is encountered, ownership of @pvr_obj will not have been |
| 123 | * transferred. If this function succeeds, however, further use of @pvr_obj is |
| 124 | * considered undefined behaviour unless another reference to it is explicitly |
| 125 | * held. |
| 126 | * |
| 127 | * Return: |
| 128 | * * 0 on success, or |
| 129 | * * Any error encountered while attempting to allocate a handle on @pvr_file. |
| 130 | */ |
| 131 | int |
| 132 | pvr_gem_object_into_handle(struct pvr_gem_object *pvr_obj, |
| 133 | struct pvr_file *pvr_file, u32 *handle) |
| 134 | { |
| 135 | struct drm_gem_object *gem_obj = gem_from_pvr_gem(pvr_obj); |
| 136 | struct drm_file *file = from_pvr_file(pvr_file); |
| 137 | |
| 138 | u32 new_handle; |
| 139 | int err; |
| 140 | |
| 141 | err = drm_gem_handle_create(file_priv: file, obj: gem_obj, handlep: &new_handle); |
| 142 | if (err) |
| 143 | return err; |
| 144 | |
| 145 | /* |
| 146 | * Release our reference to @pvr_obj, effectively transferring |
| 147 | * ownership to the handle. |
| 148 | */ |
| 149 | pvr_gem_object_put(pvr_obj); |
| 150 | |
| 151 | /* |
| 152 | * Do not store the new handle in @handle until no more errors can |
| 153 | * occur. |
| 154 | */ |
| 155 | *handle = new_handle; |
| 156 | |
| 157 | return 0; |
| 158 | } |
| 159 | |
| 160 | /** |
| 161 | * pvr_gem_object_from_handle() - Obtain a reference to an object from a |
| 162 | * userspace handle. |
| 163 | * @pvr_file: PowerVR-specific file to which @handle is associated. |
| 164 | * @handle: Userspace handle referencing the target object. |
| 165 | * |
| 166 | * On return, @handle always maintains its reference to the requested object |
| 167 | * (if it had one in the first place). If this function succeeds, the returned |
| 168 | * object will hold an additional reference. When the caller is finished with |
| 169 | * the returned object, they should call pvr_gem_object_put() on it to release |
| 170 | * this reference. |
| 171 | * |
| 172 | * Return: |
| 173 | * * A pointer to the requested PowerVR-specific object on success, or |
| 174 | * * %NULL otherwise. |
| 175 | */ |
| 176 | struct pvr_gem_object * |
| 177 | pvr_gem_object_from_handle(struct pvr_file *pvr_file, u32 handle) |
| 178 | { |
| 179 | struct drm_file *file = from_pvr_file(pvr_file); |
| 180 | struct drm_gem_object *gem_obj; |
| 181 | |
| 182 | gem_obj = drm_gem_object_lookup(filp: file, handle); |
| 183 | if (!gem_obj) |
| 184 | return NULL; |
| 185 | |
| 186 | return gem_to_pvr_gem(gem_obj); |
| 187 | } |
| 188 | |
| 189 | /** |
| 190 | * pvr_gem_object_vmap() - Map a PowerVR GEM object into CPU virtual address |
| 191 | * space. |
| 192 | * @pvr_obj: Target PowerVR GEM object. |
| 193 | * |
| 194 | * Once the caller is finished with the CPU mapping, they must call |
| 195 | * pvr_gem_object_vunmap() on @pvr_obj. |
| 196 | * |
| 197 | * If @pvr_obj is CPU-cached, dma_sync_sgtable_for_cpu() is called to make |
| 198 | * sure the CPU mapping is consistent. |
| 199 | * |
| 200 | * Return: |
| 201 | * * A pointer to the CPU mapping on success, |
| 202 | * * -%ENOMEM if the mapping fails, or |
| 203 | * * Any error encountered while attempting to acquire a reference to the |
| 204 | * backing pages for @pvr_obj. |
| 205 | */ |
| 206 | void * |
| 207 | pvr_gem_object_vmap(struct pvr_gem_object *pvr_obj) |
| 208 | { |
| 209 | struct drm_gem_shmem_object *shmem_obj = shmem_gem_from_pvr_gem(pvr_obj); |
| 210 | struct drm_gem_object *obj = gem_from_pvr_gem(pvr_obj); |
| 211 | struct iosys_map map; |
| 212 | int err; |
| 213 | |
| 214 | dma_resv_lock(obj: obj->resv, NULL); |
| 215 | |
| 216 | err = drm_gem_shmem_vmap_locked(shmem: shmem_obj, map: &map); |
| 217 | if (err) |
| 218 | goto err_unlock; |
| 219 | |
| 220 | if (pvr_obj->flags & PVR_BO_CPU_CACHED) { |
| 221 | struct device *dev = shmem_obj->base.dev->dev; |
| 222 | |
| 223 | /* If shmem_obj->sgt is NULL, that means the buffer hasn't been mapped |
| 224 | * in GPU space yet. |
| 225 | */ |
| 226 | if (shmem_obj->sgt) |
| 227 | dma_sync_sgtable_for_cpu(dev, sgt: shmem_obj->sgt, dir: DMA_BIDIRECTIONAL); |
| 228 | } |
| 229 | |
| 230 | dma_resv_unlock(obj: obj->resv); |
| 231 | |
| 232 | return map.vaddr; |
| 233 | |
| 234 | err_unlock: |
| 235 | dma_resv_unlock(obj: obj->resv); |
| 236 | |
| 237 | return ERR_PTR(error: err); |
| 238 | } |
| 239 | |
| 240 | /** |
| 241 | * pvr_gem_object_vunmap() - Unmap a PowerVR memory object from CPU virtual |
| 242 | * address space. |
| 243 | * @pvr_obj: Target PowerVR GEM object. |
| 244 | * |
| 245 | * If @pvr_obj is CPU-cached, dma_sync_sgtable_for_device() is called to make |
| 246 | * sure the GPU mapping is consistent. |
| 247 | */ |
| 248 | void |
| 249 | pvr_gem_object_vunmap(struct pvr_gem_object *pvr_obj) |
| 250 | { |
| 251 | struct drm_gem_shmem_object *shmem_obj = shmem_gem_from_pvr_gem(pvr_obj); |
| 252 | struct iosys_map map = IOSYS_MAP_INIT_VADDR(shmem_obj->vaddr); |
| 253 | struct drm_gem_object *obj = gem_from_pvr_gem(pvr_obj); |
| 254 | |
| 255 | if (WARN_ON(!map.vaddr)) |
| 256 | return; |
| 257 | |
| 258 | dma_resv_lock(obj: obj->resv, NULL); |
| 259 | |
| 260 | if (pvr_obj->flags & PVR_BO_CPU_CACHED) { |
| 261 | struct device *dev = shmem_obj->base.dev->dev; |
| 262 | |
| 263 | /* If shmem_obj->sgt is NULL, that means the buffer hasn't been mapped |
| 264 | * in GPU space yet. |
| 265 | */ |
| 266 | if (shmem_obj->sgt) |
| 267 | dma_sync_sgtable_for_device(dev, sgt: shmem_obj->sgt, dir: DMA_BIDIRECTIONAL); |
| 268 | } |
| 269 | |
| 270 | drm_gem_shmem_vunmap_locked(shmem: shmem_obj, map: &map); |
| 271 | |
| 272 | dma_resv_unlock(obj: obj->resv); |
| 273 | } |
| 274 | |
| 275 | /** |
| 276 | * pvr_gem_object_zero() - Zeroes the physical memory behind an object. |
| 277 | * @pvr_obj: Target PowerVR GEM object. |
| 278 | * |
| 279 | * Return: |
| 280 | * * 0 on success, or |
| 281 | * * Any error encountered while attempting to map @pvr_obj to the CPU (see |
| 282 | * pvr_gem_object_vmap()). |
| 283 | */ |
| 284 | static int |
| 285 | pvr_gem_object_zero(struct pvr_gem_object *pvr_obj) |
| 286 | { |
| 287 | void *cpu_ptr; |
| 288 | |
| 289 | cpu_ptr = pvr_gem_object_vmap(pvr_obj); |
| 290 | if (IS_ERR(ptr: cpu_ptr)) |
| 291 | return PTR_ERR(ptr: cpu_ptr); |
| 292 | |
| 293 | memset(cpu_ptr, 0, pvr_gem_object_size(pvr_obj)); |
| 294 | |
| 295 | /* Make sure the zero-ing is done before vumap-ing the object. */ |
| 296 | wmb(); |
| 297 | |
| 298 | pvr_gem_object_vunmap(pvr_obj); |
| 299 | |
| 300 | return 0; |
| 301 | } |
| 302 | |
| 303 | /** |
| 304 | * pvr_gem_create_object() - Allocate and pre-initializes a pvr_gem_object |
| 305 | * @drm_dev: DRM device creating this object. |
| 306 | * @size: Size of the object to allocate in bytes. |
| 307 | * |
| 308 | * Return: |
| 309 | * * The new pre-initialized GEM object on success, |
| 310 | * * -ENOMEM if the allocation failed. |
| 311 | */ |
| 312 | struct drm_gem_object *pvr_gem_create_object(struct drm_device *drm_dev, size_t size) |
| 313 | { |
| 314 | struct drm_gem_object *gem_obj; |
| 315 | struct pvr_gem_object *pvr_obj; |
| 316 | |
| 317 | pvr_obj = kzalloc(sizeof(*pvr_obj), GFP_KERNEL); |
| 318 | if (!pvr_obj) |
| 319 | return ERR_PTR(error: -ENOMEM); |
| 320 | |
| 321 | gem_obj = gem_from_pvr_gem(pvr_obj); |
| 322 | gem_obj->funcs = &pvr_gem_object_funcs; |
| 323 | |
| 324 | return gem_obj; |
| 325 | } |
| 326 | |
| 327 | /** |
| 328 | * pvr_gem_object_create() - Creates a PowerVR-specific buffer object. |
| 329 | * @pvr_dev: Target PowerVR device. |
| 330 | * @size: Size of the object to allocate in bytes. Must be greater than zero. |
| 331 | * Any value which is not an exact multiple of the system page size will be |
| 332 | * rounded up to satisfy this condition. |
| 333 | * @flags: Options which affect both this operation and future mapping |
| 334 | * operations performed on the returned object. Must be a combination of |
| 335 | * DRM_PVR_BO_* and/or PVR_BO_* flags. |
| 336 | * |
| 337 | * The created object may be larger than @size, but can never be smaller. To |
| 338 | * get the exact size, call pvr_gem_object_size() on the returned pointer. |
| 339 | * |
| 340 | * Return: |
| 341 | * * The newly-minted PowerVR-specific buffer object on success, |
| 342 | * * -%EINVAL if @size is zero or @flags is not valid, |
| 343 | * * -%ENOMEM if sufficient physical memory cannot be allocated, or |
| 344 | * * Any other error returned by drm_gem_create_mmap_offset(). |
| 345 | */ |
| 346 | struct pvr_gem_object * |
| 347 | pvr_gem_object_create(struct pvr_device *pvr_dev, size_t size, u64 flags) |
| 348 | { |
| 349 | struct drm_device *drm_dev = from_pvr_device(pvr_dev); |
| 350 | struct drm_gem_shmem_object *shmem_obj; |
| 351 | struct pvr_gem_object *pvr_obj; |
| 352 | struct sg_table *sgt; |
| 353 | int err; |
| 354 | |
| 355 | /* Verify @size and @flags before continuing. */ |
| 356 | if (size == 0 || !pvr_gem_object_flags_validate(flags)) |
| 357 | return ERR_PTR(error: -EINVAL); |
| 358 | |
| 359 | if (device_get_dma_attr(dev: drm_dev->dev) == DEV_DMA_COHERENT) |
| 360 | flags |= PVR_BO_CPU_CACHED; |
| 361 | |
| 362 | shmem_obj = drm_gem_shmem_create(dev: drm_dev, size); |
| 363 | if (IS_ERR(ptr: shmem_obj)) |
| 364 | return ERR_CAST(ptr: shmem_obj); |
| 365 | |
| 366 | shmem_obj->pages_mark_dirty_on_put = true; |
| 367 | shmem_obj->map_wc = !(flags & PVR_BO_CPU_CACHED); |
| 368 | pvr_obj = shmem_gem_to_pvr_gem(shmem_obj); |
| 369 | pvr_obj->flags = flags; |
| 370 | |
| 371 | sgt = drm_gem_shmem_get_pages_sgt(shmem: shmem_obj); |
| 372 | if (IS_ERR(ptr: sgt)) { |
| 373 | err = PTR_ERR(ptr: sgt); |
| 374 | goto err_shmem_object_free; |
| 375 | } |
| 376 | |
| 377 | dma_sync_sgtable_for_device(dev: drm_dev->dev, sgt, dir: DMA_BIDIRECTIONAL); |
| 378 | |
| 379 | /* |
| 380 | * Do this last because pvr_gem_object_zero() requires a fully |
| 381 | * configured instance of struct pvr_gem_object. |
| 382 | */ |
| 383 | pvr_gem_object_zero(pvr_obj); |
| 384 | |
| 385 | return pvr_obj; |
| 386 | |
| 387 | err_shmem_object_free: |
| 388 | drm_gem_shmem_free(shmem: shmem_obj); |
| 389 | |
| 390 | return ERR_PTR(error: err); |
| 391 | } |
| 392 | |
| 393 | /** |
| 394 | * pvr_gem_get_dma_addr() - Get DMA address for given offset in object |
| 395 | * @pvr_obj: Pointer to object to lookup address in. |
| 396 | * @offset: Offset within object to lookup address at. |
| 397 | * @dma_addr_out: Pointer to location to store DMA address. |
| 398 | * |
| 399 | * Returns: |
| 400 | * * 0 on success, or |
| 401 | * * -%EINVAL if object is not currently backed, or if @offset is out of valid |
| 402 | * range for this object. |
| 403 | */ |
| 404 | int |
| 405 | pvr_gem_get_dma_addr(struct pvr_gem_object *pvr_obj, u32 offset, |
| 406 | dma_addr_t *dma_addr_out) |
| 407 | { |
| 408 | struct drm_gem_shmem_object *shmem_obj = shmem_gem_from_pvr_gem(pvr_obj); |
| 409 | u32 accumulated_offset = 0; |
| 410 | struct scatterlist *sgl; |
| 411 | unsigned int sgt_idx; |
| 412 | |
| 413 | WARN_ON(!shmem_obj->sgt); |
| 414 | for_each_sgtable_dma_sg(shmem_obj->sgt, sgl, sgt_idx) { |
| 415 | u32 new_offset = accumulated_offset + sg_dma_len(sgl); |
| 416 | |
| 417 | if (offset >= accumulated_offset && offset < new_offset) { |
| 418 | *dma_addr_out = sg_dma_address(sgl) + |
| 419 | (offset - accumulated_offset); |
| 420 | return 0; |
| 421 | } |
| 422 | |
| 423 | accumulated_offset = new_offset; |
| 424 | } |
| 425 | |
| 426 | return -EINVAL; |
| 427 | } |
| 428 | |