1 // SPDX-License-Identifier: GPL-2.0 OR MIT 2 /* Copyright 2017-2019 Qiang Yu <yuq825@gmail.com> */ 3 4 #include <linux/mm.h> 5 #include <linux/sync_file.h> 6 #include <linux/pagemap.h> 7 #include <linux/shmem_fs.h> 8 #include <linux/dma-mapping.h> 9 10 #include <drm/drm_file.h> 11 #include <drm/drm_syncobj.h> 12 #include <drm/drm_utils.h> 13 14 #include <drm/lima_drm.h> 15 16 #include "lima_drv.h" 17 #include "lima_gem.h" 18 #include "lima_vm.h" 19 20 int lima_heap_alloc(struct lima_bo *bo, struct lima_vm *vm) 21 { 22 struct page **pages; 23 struct address_space *mapping = bo->base.base.filp->f_mapping; 24 struct device *dev = bo->base.base.dev->dev; 25 size_t old_size = bo->heap_size; 26 size_t new_size = bo->heap_size ? bo->heap_size * 2 : 27 (lima_heap_init_nr_pages << PAGE_SHIFT); 28 struct sg_table sgt; 29 int i, ret; 30 31 if (bo->heap_size >= bo->base.base.size) 32 return -ENOSPC; 33 34 new_size = min(new_size, bo->base.base.size); 35 36 mutex_lock(&bo->base.pages_lock); 37 38 if (bo->base.pages) { 39 pages = bo->base.pages; 40 } else { 41 pages = kvmalloc_array(bo->base.base.size >> PAGE_SHIFT, 42 sizeof(*pages), GFP_KERNEL | __GFP_ZERO); 43 if (!pages) { 44 mutex_unlock(&bo->base.pages_lock); 45 return -ENOMEM; 46 } 47 48 bo->base.pages = pages; 49 bo->base.pages_use_count = 1; 50 51 mapping_set_unevictable(mapping); 52 } 53 54 for (i = old_size >> PAGE_SHIFT; i < new_size >> PAGE_SHIFT; i++) { 55 struct page *page = shmem_read_mapping_page(mapping, i); 56 57 if (IS_ERR(page)) { 58 mutex_unlock(&bo->base.pages_lock); 59 return PTR_ERR(page); 60 } 61 pages[i] = page; 62 } 63 64 mutex_unlock(&bo->base.pages_lock); 65 66 ret = sg_alloc_table_from_pages(&sgt, pages, i, 0, 67 new_size, GFP_KERNEL); 68 if (ret) 69 return ret; 70 71 if (bo->base.sgt) { 72 dma_unmap_sgtable(dev, bo->base.sgt, DMA_BIDIRECTIONAL, 0); 73 sg_free_table(bo->base.sgt); 74 } else { 75 bo->base.sgt = kmalloc(sizeof(*bo->base.sgt), GFP_KERNEL); 76 if (!bo->base.sgt) { 77 sg_free_table(&sgt); 78 return -ENOMEM; 79 } 80 } 81 82 ret = dma_map_sgtable(dev, &sgt, DMA_BIDIRECTIONAL, 0); 83 if (ret) { 84 sg_free_table(&sgt); 85 kfree(bo->base.sgt); 86 bo->base.sgt = NULL; 87 return ret; 88 } 89 90 *bo->base.sgt = sgt; 91 92 if (vm) { 93 ret = lima_vm_map_bo(vm, bo, old_size >> PAGE_SHIFT); 94 if (ret) 95 return ret; 96 } 97 98 bo->heap_size = new_size; 99 return 0; 100 } 101 102 int lima_gem_create_handle(struct drm_device *dev, struct drm_file *file, 103 u32 size, u32 flags, u32 *handle) 104 { 105 int err; 106 gfp_t mask; 107 struct drm_gem_shmem_object *shmem; 108 struct drm_gem_object *obj; 109 struct lima_bo *bo; 110 bool is_heap = flags & LIMA_BO_FLAG_HEAP; 111 112 shmem = drm_gem_shmem_create(dev, size); 113 if (IS_ERR(shmem)) 114 return PTR_ERR(shmem); 115 116 obj = &shmem->base; 117 118 /* Mali Utgard GPU can only support 32bit address space */ 119 mask = mapping_gfp_mask(obj->filp->f_mapping); 120 mask &= ~__GFP_HIGHMEM; 121 mask |= __GFP_DMA32; 122 mapping_set_gfp_mask(obj->filp->f_mapping, mask); 123 124 if (is_heap) { 125 bo = to_lima_bo(obj); 126 err = lima_heap_alloc(bo, NULL); 127 if (err) 128 goto out; 129 } else { 130 struct sg_table *sgt = drm_gem_shmem_get_pages_sgt(obj); 131 132 if (IS_ERR(sgt)) { 133 err = PTR_ERR(sgt); 134 goto out; 135 } 136 } 137 138 err = drm_gem_handle_create(file, obj, handle); 139 140 out: 141 /* drop reference from allocate - handle holds it now */ 142 drm_gem_object_put(obj); 143 144 return err; 145 } 146 147 static void lima_gem_free_object(struct drm_gem_object *obj) 148 { 149 struct lima_bo *bo = to_lima_bo(obj); 150 151 if (!list_empty(&bo->va)) 152 dev_err(obj->dev->dev, "lima gem free bo still has va\n"); 153 154 drm_gem_shmem_free_object(obj); 155 } 156 157 static int lima_gem_object_open(struct drm_gem_object *obj, struct drm_file *file) 158 { 159 struct lima_bo *bo = to_lima_bo(obj); 160 struct lima_drm_priv *priv = to_lima_drm_priv(file); 161 struct lima_vm *vm = priv->vm; 162 163 return lima_vm_bo_add(vm, bo, true); 164 } 165 166 static void lima_gem_object_close(struct drm_gem_object *obj, struct drm_file *file) 167 { 168 struct lima_bo *bo = to_lima_bo(obj); 169 struct lima_drm_priv *priv = to_lima_drm_priv(file); 170 struct lima_vm *vm = priv->vm; 171 172 lima_vm_bo_del(vm, bo); 173 } 174 175 static int lima_gem_pin(struct drm_gem_object *obj) 176 { 177 struct lima_bo *bo = to_lima_bo(obj); 178 179 if (bo->heap_size) 180 return -EINVAL; 181 182 return drm_gem_shmem_pin(obj); 183 } 184 185 static int lima_gem_vmap(struct drm_gem_object *obj, struct dma_buf_map *map) 186 { 187 struct lima_bo *bo = to_lima_bo(obj); 188 189 if (bo->heap_size) 190 return -EINVAL; 191 192 return drm_gem_shmem_vmap(obj, map); 193 } 194 195 static int lima_gem_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma) 196 { 197 struct lima_bo *bo = to_lima_bo(obj); 198 199 if (bo->heap_size) 200 return -EINVAL; 201 202 return drm_gem_shmem_mmap(obj, vma); 203 } 204 205 static const struct drm_gem_object_funcs lima_gem_funcs = { 206 .free = lima_gem_free_object, 207 .open = lima_gem_object_open, 208 .close = lima_gem_object_close, 209 .print_info = drm_gem_shmem_print_info, 210 .pin = lima_gem_pin, 211 .unpin = drm_gem_shmem_unpin, 212 .get_sg_table = drm_gem_shmem_get_sg_table, 213 .vmap = lima_gem_vmap, 214 .vunmap = drm_gem_shmem_vunmap, 215 .mmap = lima_gem_mmap, 216 }; 217 218 struct drm_gem_object *lima_gem_create_object(struct drm_device *dev, size_t size) 219 { 220 struct lima_bo *bo; 221 222 bo = kzalloc(sizeof(*bo), GFP_KERNEL); 223 if (!bo) 224 return NULL; 225 226 mutex_init(&bo->lock); 227 INIT_LIST_HEAD(&bo->va); 228 229 bo->base.base.funcs = &lima_gem_funcs; 230 231 return &bo->base.base; 232 } 233 234 int lima_gem_get_info(struct drm_file *file, u32 handle, u32 *va, u64 *offset) 235 { 236 struct drm_gem_object *obj; 237 struct lima_bo *bo; 238 struct lima_drm_priv *priv = to_lima_drm_priv(file); 239 struct lima_vm *vm = priv->vm; 240 241 obj = drm_gem_object_lookup(file, handle); 242 if (!obj) 243 return -ENOENT; 244 245 bo = to_lima_bo(obj); 246 247 *va = lima_vm_get_va(vm, bo); 248 249 *offset = drm_vma_node_offset_addr(&obj->vma_node); 250 251 drm_gem_object_put(obj); 252 return 0; 253 } 254 255 static int lima_gem_sync_bo(struct lima_sched_task *task, struct lima_bo *bo, 256 bool write, bool explicit) 257 { 258 int err = 0; 259 260 if (!write) { 261 err = dma_resv_reserve_shared(lima_bo_resv(bo), 1); 262 if (err) 263 return err; 264 } 265 266 /* explicit sync use user passed dep fence */ 267 if (explicit) 268 return 0; 269 270 return drm_gem_fence_array_add_implicit(&task->deps, &bo->base.base, write); 271 } 272 273 static int lima_gem_add_deps(struct drm_file *file, struct lima_submit *submit) 274 { 275 int i, err; 276 277 for (i = 0; i < ARRAY_SIZE(submit->in_sync); i++) { 278 struct dma_fence *fence = NULL; 279 280 if (!submit->in_sync[i]) 281 continue; 282 283 err = drm_syncobj_find_fence(file, submit->in_sync[i], 284 0, 0, &fence); 285 if (err) 286 return err; 287 288 err = drm_gem_fence_array_add(&submit->task->deps, fence); 289 if (err) { 290 dma_fence_put(fence); 291 return err; 292 } 293 } 294 295 return 0; 296 } 297 298 int lima_gem_submit(struct drm_file *file, struct lima_submit *submit) 299 { 300 int i, err = 0; 301 struct ww_acquire_ctx ctx; 302 struct lima_drm_priv *priv = to_lima_drm_priv(file); 303 struct lima_vm *vm = priv->vm; 304 struct drm_syncobj *out_sync = NULL; 305 struct dma_fence *fence; 306 struct lima_bo **bos = submit->lbos; 307 308 if (submit->out_sync) { 309 out_sync = drm_syncobj_find(file, submit->out_sync); 310 if (!out_sync) 311 return -ENOENT; 312 } 313 314 for (i = 0; i < submit->nr_bos; i++) { 315 struct drm_gem_object *obj; 316 struct lima_bo *bo; 317 318 obj = drm_gem_object_lookup(file, submit->bos[i].handle); 319 if (!obj) { 320 err = -ENOENT; 321 goto err_out0; 322 } 323 324 bo = to_lima_bo(obj); 325 326 /* increase refcnt of gpu va map to prevent unmapped when executing, 327 * will be decreased when task done 328 */ 329 err = lima_vm_bo_add(vm, bo, false); 330 if (err) { 331 drm_gem_object_put(obj); 332 goto err_out0; 333 } 334 335 bos[i] = bo; 336 } 337 338 err = drm_gem_lock_reservations((struct drm_gem_object **)bos, 339 submit->nr_bos, &ctx); 340 if (err) 341 goto err_out0; 342 343 err = lima_sched_task_init( 344 submit->task, submit->ctx->context + submit->pipe, 345 bos, submit->nr_bos, vm); 346 if (err) 347 goto err_out1; 348 349 err = lima_gem_add_deps(file, submit); 350 if (err) 351 goto err_out2; 352 353 for (i = 0; i < submit->nr_bos; i++) { 354 err = lima_gem_sync_bo( 355 submit->task, bos[i], 356 submit->bos[i].flags & LIMA_SUBMIT_BO_WRITE, 357 submit->flags & LIMA_SUBMIT_FLAG_EXPLICIT_FENCE); 358 if (err) 359 goto err_out2; 360 } 361 362 fence = lima_sched_context_queue_task( 363 submit->ctx->context + submit->pipe, submit->task); 364 365 for (i = 0; i < submit->nr_bos; i++) { 366 if (submit->bos[i].flags & LIMA_SUBMIT_BO_WRITE) 367 dma_resv_add_excl_fence(lima_bo_resv(bos[i]), fence); 368 else 369 dma_resv_add_shared_fence(lima_bo_resv(bos[i]), fence); 370 } 371 372 drm_gem_unlock_reservations((struct drm_gem_object **)bos, 373 submit->nr_bos, &ctx); 374 375 for (i = 0; i < submit->nr_bos; i++) 376 drm_gem_object_put(&bos[i]->base.base); 377 378 if (out_sync) { 379 drm_syncobj_replace_fence(out_sync, fence); 380 drm_syncobj_put(out_sync); 381 } 382 383 dma_fence_put(fence); 384 385 return 0; 386 387 err_out2: 388 lima_sched_task_fini(submit->task); 389 err_out1: 390 drm_gem_unlock_reservations((struct drm_gem_object **)bos, 391 submit->nr_bos, &ctx); 392 err_out0: 393 for (i = 0; i < submit->nr_bos; i++) { 394 if (!bos[i]) 395 break; 396 lima_vm_bo_del(vm, bos[i]); 397 drm_gem_object_put(&bos[i]->base.base); 398 } 399 if (out_sync) 400 drm_syncobj_put(out_sync); 401 return err; 402 } 403 404 int lima_gem_wait(struct drm_file *file, u32 handle, u32 op, s64 timeout_ns) 405 { 406 bool write = op & LIMA_GEM_WAIT_WRITE; 407 long ret, timeout; 408 409 if (!op) 410 return 0; 411 412 timeout = drm_timeout_abs_to_jiffies(timeout_ns); 413 414 ret = drm_gem_dma_resv_wait(file, handle, write, timeout); 415 if (ret == -ETIME) 416 ret = timeout ? -ETIMEDOUT : -EBUSY; 417 418 return ret; 419 } 420