The CL submission had two jobs embedded in an exec struct. When I added TFU support, I had to replicate some of the exec stuff and some of the job stuff. As I went to add CSD, it became clear that actually what was in exec should just be in the two CL jobs, and it would let us share a lot more code between the 4 queues. v2: Fix missing error path in TFU ioctl's bo[] allocation.
Unified · Split
| @@ -67,8 +67,8 @@ struct v3d_dev { · 66 unmodified lines | |||
| 67 | 67 | ||
| 68 | struct work_struct overflow_mem_work; | 68 | struct work_struct overflow_mem_work; |
| 69 | 69 | ||
| 70 | struct v3d_exec_info *bin_job; | 70 | struct v3d_bin_job *bin_job; |
| 71 | struct v3d_exec_info *render_job; | 71 | struct v3d_render_job *render_job; |
| 72 | struct v3d_tfu_job *tfu_job; | 72 | struct v3d_tfu_job *tfu_job; |
| 73 | 73 | ||
| 74 | struct v3d_queue_state queue[V3D_MAX_QUEUES]; | 74 | struct v3d_queue_state queue[V3D_MAX_QUEUES]; |
| @@ -117,7 +117,7 @@ struct v3d_bo { · 42 unmodified lines | |||
| 117 | struct drm_mm_node node; | 117 | struct drm_mm_node node; |
| 118 | 118 | ||
| 119 | /* List entry for the BO's position in | 119 | /* List entry for the BO's position in |
| 120 | * v3d_exec_info->unref_list | 120 | * v3d_render_job->unref_list |
| 121 | */ | 121 | */ |
| 122 | struct list_head unref_head; | 122 | struct list_head unref_head; |
| 123 | }; | 123 | }; |
| @@ -157,7 +157,15 @@ to_v3d_fence(struct dma_fence *fence) · 33 unmodified lines | |||
| 157 | struct v3d_job { | 157 | struct v3d_job { |
| 158 | struct drm_sched_job base; | 158 | struct drm_sched_job base; |
| 159 | 159 | ||
| 160 | struct v3d_exec_info *exec; | 160 | struct kref refcount; |
| 161 | |||
| 162 | struct v3d_dev *v3d; | ||
| 163 | |||
| 164 | /* This is the array of BOs that were looked up at the start | ||
| 165 | * of submission. | ||
| 166 | */ | ||
| 167 | struct drm_gem_object **bo; | ||
| 168 | u32 bo_count; | ||
| 161 | 169 | ||
| 162 | /* An optional fence userspace can pass in for the job to depend on. */ | 170 | /* An optional fence userspace can pass in for the job to depend on. */ |
| 163 | struct dma_fence *in_fence; | 171 | struct dma_fence *in_fence; |
| @@ -165,59 +173,53 @@ struct v3d_job { · 1 unmodified line | |||
| 165 | /* v3d fence to be signaled by IRQ handler when the job is complete. */ | 173 | /* v3d fence to be signaled by IRQ handler when the job is complete. */ |
| 166 | struct dma_fence *irq_fence; | 174 | struct dma_fence *irq_fence; |
| 167 | 175 | ||
| 176 | /* scheduler fence for when the job is considered complete and | ||
| 177 | * the BO reservations can be released. | ||
| 178 | */ | ||
| 179 | struct dma_fence *done_fence; | ||
| 180 | |||
| 181 | /* Callback for the freeing of the job on refcount going to 0. */ | ||
| 182 | void (*free)(struct kref *ref); | ||
| 183 | }; | ||
| 184 | |||
| 185 | struct v3d_bin_job { | ||
| 186 | struct v3d_job base; | ||
| 187 | |||
| 168 | /* GPU virtual addresses of the start/end of the CL job. */ | 188 | /* GPU virtual addresses of the start/end of the CL job. */ |
| 169 | u32 start, end; | 189 | u32 start, end; |
| 170 | 190 | ||
| 171 | u32 timedout_ctca, timedout_ctra; | 191 | u32 timedout_ctca, timedout_ctra; |
| 172 | }; | ||
| 173 | 192 | ||
| 174 | struct v3d_exec_info { | 193 | /* Corresponding render job, for attaching our overflow memory. */ |
| 175 | struct v3d_dev *v3d; | 194 | struct v3d_render_job *render; |
| 195 | |||
| 196 | /* Submitted tile memory allocation start/size, tile state. */ | ||
| 197 | u32 qma, qms, qts; | ||
| 198 | }; | ||
| 176 | 199 | ||
| 177 | struct v3d_job bin, render; | 200 | struct v3d_render_job { |
| 201 | struct v3d_job base; | ||
| 178 | 202 | ||
| 179 | /* Fence for when the scheduler considers the binner to be | 203 | /* Optional fence for the binner, to depend on before starting |
| 180 | * done, for render to depend on. | 204 | * our job. |
| 181 | */ | 205 | */ |
| 182 | struct dma_fence *bin_done_fence; | 206 | struct dma_fence *bin_done_fence; |
| 183 | 207 | ||
| 184 | /* Fence for when the scheduler considers the render to be | 208 | /* GPU virtual addresses of the start/end of the CL job. */ |
| 185 | * done, for when the BOs reservations should be complete. | 209 | u32 start, end; |
| 186 | */ | ||
| 187 | struct dma_fence *render_done_fence; | ||
| 188 | |||
| 189 | struct kref refcount; | ||
| 190 | 210 | ||
| 191 | /* This is the array of BOs that were looked up at the start of exec. */ | 211 | u32 timedout_ctca, timedout_ctra; |
| 192 | struct drm_gem_object **bo; | ||
| 193 | u32 bo_count; | ||
| 194 | 212 | ||
| 195 | /* List of overflow BOs used in the job that need to be | 213 | /* List of overflow BOs used in the job that need to be |
| 196 | * released once the job is complete. | 214 | * released once the job is complete. |
| 197 | */ | 215 | */ |
| 198 | struct list_head unref_list; | 216 | struct list_head unref_list; |
| 199 | |||
| 200 | /* Submitted tile memory allocation start/size, tile state. */ | ||
| 201 | u32 qma, qms, qts; | ||
| 202 | }; | 217 | }; |
| 203 | 218 | ||
| 204 | struct v3d_tfu_job { | 219 | struct v3d_tfu_job { |
| 205 | struct drm_sched_job base; | 220 | struct v3d_job base; |
| 206 | 221 | ||
| 207 | struct drm_v3d_submit_tfu args; | 222 | struct drm_v3d_submit_tfu args; |
| 208 | |||
| 209 | /* An optional fence userspace can pass in for the job to depend on. */ | ||
| 210 | struct dma_fence *in_fence; | ||
| 211 | |||
| 212 | /* v3d fence to be signaled by IRQ handler when the job is complete. */ | ||
| 213 | struct dma_fence *irq_fence; | ||
| 214 | |||
| 215 | struct v3d_dev *v3d; | ||
| 216 | |||
| 217 | struct kref refcount; | ||
| 218 | |||
| 219 | /* This is the array of BOs that were looked up at the start of exec. */ | ||
| 220 | struct drm_gem_object *bo[4]; | ||
| 221 | }; | 223 | }; |
| 222 | 224 | ||
| 223 | /** | 225 | /** |
| @@ -283,8 +285,7 @@ int v3d_submit_tfu_ioctl(struct drm_device *dev, void *data, · 59 unmodified lines | |||
| 283 | struct drm_file *file_priv); | 285 | struct drm_file *file_priv); |
| 284 | int v3d_wait_bo_ioctl(struct drm_device *dev, void *data, | 286 | int v3d_wait_bo_ioctl(struct drm_device *dev, void *data, |
| 285 | struct drm_file *file_priv); | 287 | struct drm_file *file_priv); |
| 286 | void v3d_exec_put(struct v3d_exec_info *exec); | 288 | void v3d_job_put(struct v3d_job *job); |
| 287 | void v3d_tfu_job_put(struct v3d_tfu_job *exec); | ||
| 288 | void v3d_reset(struct v3d_dev *v3d); | 289 | void v3d_reset(struct v3d_dev *v3d); |
| 289 | void v3d_invalidate_caches(struct v3d_dev *v3d); | 290 | void v3d_invalidate_caches(struct v3d_dev *v3d); |
| 290 | 291 | ||
| @@ -193,18 +193,6 @@ v3d_invalidate_caches(struct v3d_dev *v3d) · 192 unmodified lines | |||
| 193 | v3d_invalidate_slices(v3d, 0); | 193 | v3d_invalidate_slices(v3d, 0); |
| 194 | } | 194 | } |
| 195 | 195 | ||
| 196 | static void | ||
| 197 | v3d_attach_object_fences(struct drm_gem_object **bos, int bo_count, | ||
| 198 | struct dma_fence *fence) | ||
| 199 | { | ||
| 200 | int i; | ||
| 201 | |||
| 202 | for (i = 0; i < bo_count; i++) { | ||
| 203 | /* XXX: Use shared fences for read-only objects. */ | ||
| 204 | reservation_object_add_excl_fence(bos[i]->resv, fence); | ||
| 205 | } | ||
| 206 | } | ||
| 207 | |||
| 208 | /* Takes the reservation lock on all the BOs being referenced, so that | 196 | /* Takes the reservation lock on all the BOs being referenced, so that |
| 209 | * at queue submit time we can update the reservations. | 197 | * at queue submit time we can update the reservations. |
| 210 | * | 198 | * |
| @@ -239,11 +227,11 @@ v3d_lock_bo_reservations(struct drm_gem_object **bos, · 28 unmodified lines | |||
| 239 | } | 227 | } |
| 240 | 228 | ||
| 241 | /** | 229 | /** |
| 242 | * v3d_cl_lookup_bos() - Sets up exec->bo[] with the GEM objects | 230 | * v3d_lookup_bos() - Sets up job->bo[] with the GEM objects |
| 243 | * referenced by the job. | 231 | * referenced by the job. |
| 244 | * @dev: DRM device | 232 | * @dev: DRM device |
| 245 | * @file_priv: DRM file for this fd | 233 | * @file_priv: DRM file for this fd |
| 246 | * @exec: V3D job being set up | 234 | * @job: V3D job being set up |
| 247 | * | 235 | * |
| 248 | * The command validator needs to reference BOs by their index within | 236 | * The command validator needs to reference BOs by their index within |
| 249 | * the submitted job's BO list. This does the validation of the job's | 237 | * the submitted job's BO list. This does the validation of the job's |
| @@ -253,18 +241,19 @@ v3d_lock_bo_reservations(struct drm_gem_object **bos, · 3 unmodified lines | |||
| 253 | * failure, because that will happen at v3d_exec_cleanup() time. | 241 | * failure, because that will happen at v3d_exec_cleanup() time. |
| 254 | */ | 242 | */ |
| 255 | static int | 243 | static int |
| 256 | v3d_cl_lookup_bos(struct drm_device *dev, | 244 | v3d_lookup_bos(struct drm_device *dev, |
| 257 | struct drm_file *file_priv, | 245 | struct drm_file *file_priv, |
| 258 | struct drm_v3d_submit_cl *args, | 246 | struct v3d_job *job, |
| 259 | struct v3d_exec_info *exec) | 247 | u64 bo_handles, |
| 248 | u32 bo_count) | ||
| 260 | { | 249 | { |
| 261 | u32 *handles; | 250 | u32 *handles; |
| 262 | int ret = 0; | 251 | int ret = 0; |
| 263 | int i; | 252 | int i; |
| 264 | 253 | ||
| 265 | exec->bo_count = args->bo_handle_count; | 254 | job->bo_count = bo_count; |
| 266 | 255 | ||
| 267 | if (!exec->bo_count) { | 256 | if (!job->bo_count) { |
| 268 | /* See comment on bo_index for why we have to check | 257 | /* See comment on bo_index for why we have to check |
| 269 | * this. | 258 | * this. |
| 270 | */ | 259 | */ |
| @@ -272,15 +261,15 @@ v3d_cl_lookup_bos(struct drm_device *dev, · 1 unmodified line | |||
| 272 | return -EINVAL; | 261 | return -EINVAL; |
| 273 | } | 262 | } |
| 274 | 263 | ||
| 275 | exec->bo = kvmalloc_array(exec->bo_count, | 264 | job->bo = kvmalloc_array(job->bo_count, |
| 276 | sizeof(struct drm_gem_cma_object *), | 265 | sizeof(struct drm_gem_cma_object *), |
| 277 | GFP_KERNEL | __GFP_ZERO); | 266 | GFP_KERNEL | __GFP_ZERO); |
| 278 | if (!exec->bo) { | 267 | if (!job->bo) { |
| 279 | DRM_DEBUG("Failed to allocate validated BO pointers\n"); | 268 | DRM_DEBUG("Failed to allocate validated BO pointers\n"); |
| 280 | return -ENOMEM; | 269 | return -ENOMEM; |
| 281 | } | 270 | } |
| 282 | 271 | ||
| 283 | handles = kvmalloc_array(exec->bo_count, sizeof(u32), GFP_KERNEL); | 272 | handles = kvmalloc_array(job->bo_count, sizeof(u32), GFP_KERNEL); |
| 284 | if (!handles) { | 273 | if (!handles) { |
| 285 | ret = -ENOMEM; | 274 | ret = -ENOMEM; |
| 286 | DRM_DEBUG("Failed to allocate incoming GEM handles\n"); | 275 | DRM_DEBUG("Failed to allocate incoming GEM handles\n"); |
| @@ -288,15 +277,15 @@ v3d_cl_lookup_bos(struct drm_device *dev, · 1 unmodified line | |||
| 288 | } | 277 | } |
| 289 | 278 | ||
| 290 | if (copy_from_user(handles, | 279 | if (copy_from_user(handles, |
| 291 | (void __user *)(uintptr_t)args->bo_handles, | 280 | (void __user *)(uintptr_t)bo_handles, |
| 292 | exec->bo_count * sizeof(u32))) { | 281 | job->bo_count * sizeof(u32))) { |
| 293 | ret = -EFAULT; | 282 | ret = -EFAULT; |
| 294 | DRM_DEBUG("Failed to copy in GEM handles\n"); | 283 | DRM_DEBUG("Failed to copy in GEM handles\n"); |
| 295 | goto fail; | 284 | goto fail; |
| 296 | } | 285 | } |
| 297 | 286 | ||
| 298 | spin_lock(&file_priv->table_lock); | 287 | spin_lock(&file_priv->table_lock); |
| 299 | for (i = 0; i < exec->bo_count; i++) { | 288 | for (i = 0; i < job->bo_count; i++) { |
| 300 | struct drm_gem_object *bo = idr_find(&file_priv->object_idr, | 289 | struct drm_gem_object *bo = idr_find(&file_priv->object_idr, |
| 301 | handles[i]); | 290 | handles[i]); |
| 302 | if (!bo) { | 291 | if (!bo) { |
| @@ -307,7 +296,7 @@ v3d_cl_lookup_bos(struct drm_device *dev, · 4 unmodified lines | |||
| 307 | goto fail; | 296 | goto fail; |
| 308 | } | 297 | } |
| 309 | drm_gem_object_get(bo); | 298 | drm_gem_object_get(bo); |
| 310 | exec->bo[i] = bo; | 299 | job->bo[i] = bo; |
| 311 | } | 300 | } |
| 312 | spin_unlock(&file_priv->table_lock); | 301 | spin_unlock(&file_priv->table_lock); |
| 313 | 302 | ||
| @@ -317,67 +306,44 @@ fail: · 3 unmodified lines | |||
| 317 | } | 306 | } |
| 318 | 307 | ||
| 319 | static void | 308 | static void |
| 320 | v3d_exec_cleanup(struct kref *ref) | 309 | v3d_job_free(struct kref *ref) |
| 321 | { | 310 | { |
| 322 | struct v3d_exec_info *exec = container_of(ref, struct v3d_exec_info, | 311 | struct v3d_job *job = container_of(ref, struct v3d_job, refcount); |
| 323 | refcount); | 312 | int i; |
| 324 | struct v3d_dev *v3d = exec->v3d; | ||
| 325 | unsigned int i; | ||
| 326 | struct v3d_bo *bo, *save; | ||
| 327 | |||
| 328 | dma_fence_put(exec->bin.in_fence); | ||
| 329 | dma_fence_put(exec->render.in_fence); | ||
| 330 | |||
| 331 | dma_fence_put(exec->bin.irq_fence); | ||
| 332 | dma_fence_put(exec->render.irq_fence); | ||
| 333 | |||
| 334 | dma_fence_put(exec->bin_done_fence); | ||
| 335 | dma_fence_put(exec->render_done_fence); | ||
| 336 | |||
| 337 | for (i = 0; i < exec->bo_count; i++) | ||
| 338 | drm_gem_object_put_unlocked(exec->bo[i]); | ||
| 339 | kvfree(exec->bo); | ||
| 340 | 313 | ||
| 341 | list_for_each_entry_safe(bo, save, &exec->unref_list, unref_head) { | 314 | for (i = 0; i < job->bo_count; i++) { |
| 342 | drm_gem_object_put_unlocked(&bo->base.base); | 315 | if (job->bo[i]) |
| 316 | drm_gem_object_put_unlocked(job->bo[i]); | ||
| 343 | } | 317 | } |
| 318 | kvfree(job->bo); | ||
| 344 | 319 | ||
| 345 | pm_runtime_mark_last_busy(v3d->dev); | 320 | dma_fence_put(job->in_fence); |
| 346 | pm_runtime_put_autosuspend(v3d->dev); | 321 | dma_fence_put(job->irq_fence); |
| 322 | dma_fence_put(job->done_fence); | ||
| 347 | 323 | ||
| 348 | kfree(exec); | 324 | pm_runtime_mark_last_busy(job->v3d->dev); |
| 349 | } | 325 | pm_runtime_put_autosuspend(job->v3d->dev); |
| 350 | 326 | ||
| 351 | void v3d_exec_put(struct v3d_exec_info *exec) | 327 | kfree(job); |
| 352 | { | ||
| 353 | kref_put(&exec->refcount, v3d_exec_cleanup); | ||
| 354 | } | 328 | } |
| 355 | 329 | ||
| 356 | static void | 330 | static void |
| 357 | v3d_tfu_job_cleanup(struct kref *ref) | 331 | v3d_render_job_free(struct kref *ref) |
| 358 | { | 332 | { |
| 359 | struct v3d_tfu_job *job = container_of(ref, struct v3d_tfu_job, | 333 | struct v3d_render_job *job = container_of(ref, struct v3d_render_job, |
| 360 | refcount); | 334 | base.refcount); |
| 361 | struct v3d_dev *v3d = job->v3d; | 335 | struct v3d_bo *bo, *save; |
| 362 | unsigned int i; | ||
| 363 | |||
| 364 | dma_fence_put(job->in_fence); | ||
| 365 | dma_fence_put(job->irq_fence); | ||
| 366 | 336 | ||
| 367 | for (i = 0; i < ARRAY_SIZE(job->bo); i++) { | 337 | list_for_each_entry_safe(bo, save, &job->unref_list, unref_head) { |
| 368 | if (job->bo[i]) | 338 | drm_gem_object_put_unlocked(&bo->base.base); |
| 369 | drm_gem_object_put_unlocked(job->bo[i]); | ||
| 370 | } | 339 | } |
| 371 | 340 | ||
| 372 | pm_runtime_mark_last_busy(v3d->dev); | 341 | v3d_job_free(ref); |
| 373 | pm_runtime_put_autosuspend(v3d->dev); | ||
| 374 | |||
| 375 | kfree(job); | ||
| 376 | } | 342 | } |
| 377 | 343 | ||
| 378 | void v3d_tfu_job_put(struct v3d_tfu_job *job) | 344 | void v3d_job_put(struct v3d_job *job) |
| 379 | { | 345 | { |
| 380 | kref_put(&job->refcount, v3d_tfu_job_cleanup); | 346 | kref_put(&job->refcount, job->free); |
| 381 | } | 347 | } |
| 382 | 348 | ||
| 383 | int | 349 | int |
| @@ -413,6 +379,77 @@ v3d_wait_bo_ioctl(struct drm_device *dev, void *data, · 29 unmodified lines | |||
| 413 | return ret; | 379 | return ret; |
| 414 | } | 380 | } |
| 415 | 381 | ||
| 382 | static int | ||
| 383 | v3d_job_init(struct v3d_dev *v3d, struct drm_file *file_priv, | ||
| 384 | struct v3d_job *job, void (*free)(struct kref *ref), | ||
| 385 | u32 in_sync) | ||
| 386 | { | ||
| 387 | int ret; | ||
| 388 | |||
| 389 | job->v3d = v3d; | ||
| 390 | job->free = free; | ||
| 391 | |||
| 392 | ret = pm_runtime_get_sync(v3d->dev); | ||
| 393 | if (ret < 0) | ||
| 394 | return ret; | ||
| 395 | |||
| 396 | ret = drm_syncobj_find_fence(file_priv, in_sync, 0, 0, &job->in_fence); | ||
| 397 | if (ret == -EINVAL) { | ||
| 398 | pm_runtime_put_autosuspend(v3d->dev); | ||
| 399 | return ret; | ||
| 400 | } | ||
| 401 | |||
| 402 | kref_init(&job->refcount); | ||
| 403 | |||
| 404 | return 0; | ||
| 405 | } | ||
| 406 | |||
| 407 | static int | ||
| 408 | v3d_push_job(struct v3d_file_priv *v3d_priv, | ||
| 409 | struct v3d_job *job, enum v3d_queue queue) | ||
| 410 | { | ||
| 411 | int ret; | ||
| 412 | |||
| 413 | ret = drm_sched_job_init(&job->base, &v3d_priv->sched_entity[queue], | ||
| 414 | v3d_priv); | ||
| 415 | if (ret) | ||
| 416 | return ret; | ||
| 417 | |||
| 418 | job->done_fence = dma_fence_get(&job->base.s_fence->finished); | ||
| 419 | |||
| 420 | /* put by scheduler job completion */ | ||
| 421 | kref_get(&job->refcount); | ||
| 422 | |||
| 423 | drm_sched_entity_push_job(&job->base, &v3d_priv->sched_entity[queue]); | ||
| 424 | |||
| 425 | return 0; | ||
| 426 | } | ||
| 427 | |||
| 428 | static void | ||
| 429 | v3d_attach_fences_and_unlock_reservation(struct drm_file *file_priv, | ||
| 430 | struct v3d_job *job, | ||
| 431 | struct ww_acquire_ctx *acquire_ctx, | ||
| 432 | u32 out_sync) | ||
| 433 | { | ||
| 434 | struct drm_syncobj *sync_out; | ||
| 435 | int i; | ||
| 436 | |||
| 437 | for (i = 0; i < job->bo_count; i++) { | ||
| 438 | /* XXX: Use shared fences for read-only objects. */ | ||
| 439 | reservation_object_add_excl_fence(job->bo[i]->resv, | ||
| 440 | job->done_fence); | ||
| 441 | } | ||
| 442 | |||
| 443 | drm_gem_unlock_reservations(job->bo, job->bo_count, acquire_ctx); | ||
| 444 | |||
| 445 | /* Update the return sync object for the job */ | ||
| 446 | sync_out = drm_syncobj_find(file_priv, out_sync); | ||
| 447 | if (sync_out) { | ||
| 448 | drm_syncobj_replace_fence(sync_out, job->done_fence); | ||
| 449 | drm_syncobj_put(sync_out); | ||
| 450 | } | ||
| 451 | } | ||
| 452 | |||
| 416 | /** | 453 | /** |
| 417 | * v3d_submit_cl_ioctl() - Submits a job (frame) to the V3D. | 454 | * v3d_submit_cl_ioctl() - Submits a job (frame) to the V3D. |
| 418 | * @dev: DRM device | 455 | * @dev: DRM device |
| @@ -432,9 +469,9 @@ v3d_submit_cl_ioctl(struct drm_device *dev, void *data, · 13 unmodified lines | |||
| 432 | struct v3d_dev *v3d = to_v3d_dev(dev); | 469 | struct v3d_dev *v3d = to_v3d_dev(dev); |
| 433 | struct v3d_file_priv *v3d_priv = file_priv->driver_priv; | 470 | struct v3d_file_priv *v3d_priv = file_priv->driver_priv; |
| 434 | struct drm_v3d_submit_cl *args = data; | 471 | struct drm_v3d_submit_cl *args = data; |
| 435 | struct v3d_exec_info *exec; | 472 | struct v3d_bin_job *bin = NULL; |
| 473 | struct v3d_render_job *render; | ||
| 436 | struct ww_acquire_ctx acquire_ctx; | 474 | struct ww_acquire_ctx acquire_ctx; |
| 437 | struct drm_syncobj *sync_out; | ||
| 438 | int ret = 0; | 475 | int ret = 0; |
| 439 | 476 | ||
| 440 | trace_v3d_submit_cl_ioctl(&v3d->drm, args->rcl_start, args->rcl_end); | 477 | trace_v3d_submit_cl_ioctl(&v3d->drm, args->rcl_start, args->rcl_end); |
| @@ -444,100 +481,83 @@ v3d_submit_cl_ioctl(struct drm_device *dev, void *data, · 3 unmodified lines | |||
| 444 | return -EINVAL; | 481 | return -EINVAL; |
| 445 | } | 482 | } |
| 446 | 483 | ||
| 447 | exec = kcalloc(1, sizeof(*exec), GFP_KERNEL); | 484 | render = kcalloc(1, sizeof(*render), GFP_KERNEL); |
| 448 | if (!exec) | 485 | if (!render) |
| 449 | return -ENOMEM; | 486 | return -ENOMEM; |
| 450 | 487 | ||
| 451 | ret = pm_runtime_get_sync(v3d->dev); | 488 | render->start = args->rcl_start; |
| 452 | if (ret < 0) { | 489 | render->end = args->rcl_end; |
| 453 | kfree(exec); | 490 | INIT_LIST_HEAD(&render->unref_list); |
| 491 | |||
| 492 | ret = v3d_job_init(v3d, file_priv, &render->base, | ||
| 493 | v3d_render_job_free, args->in_sync_rcl); | ||
| 494 | if (ret) { | ||
| 495 | kfree(render); | ||
| 454 | return ret; | 496 | return ret; |
| 455 | } | 497 | } |
| 456 | 498 | ||
| 457 | kref_init(&exec->refcount); | 499 | if (args->bcl_start != args->bcl_end) { |
| 500 | bin = kcalloc(1, sizeof(*bin), GFP_KERNEL); | ||
| 501 | if (!bin) | ||
| 502 | return -ENOMEM; | ||
| 458 | 503 | ||
| 459 | ret = drm_syncobj_find_fence(file_priv, args->in_sync_bcl, | 504 | ret = v3d_job_init(v3d, file_priv, &bin->base, |
| 460 | 0, 0, &exec->bin.in_fence); | 505 | v3d_job_free, args->in_sync_bcl); |
| 461 | if (ret == -EINVAL) | 506 | if (ret) { |
| 462 | goto fail; | 507 | v3d_job_put(&render->base); |
| 508 | return ret; | ||
| 509 | } | ||
| 463 | 510 | ||
| 464 | ret = drm_syncobj_find_fence(file_priv, args->in_sync_rcl, | 511 | bin->start = args->bcl_start; |
| 465 | 0, 0, &exec->render.in_fence); | 512 | bin->end = args->bcl_end; |
| 466 | if (ret == -EINVAL) | 513 | bin->qma = args->qma; |
| 467 | goto fail; | 514 | bin->qms = args->qms; |
| 515 | bin->qts = args->qts; | ||
| 516 | bin->render = render; | ||
| 517 | } | ||
| 468 | 518 | ||
| 469 | exec->qma = args->qma; | 519 | ret = v3d_lookup_bos(dev, file_priv, &render->base, |
| 470 | exec->qms = args->qms; | 520 | args->bo_handles, args->bo_handle_count); |
| 471 | exec->qts = args->qts; | ||
| 472 | exec->bin.exec = exec; | ||
| 473 | exec->bin.start = args->bcl_start; | ||
| 474 | exec->bin.end = args->bcl_end; | ||
| 475 | exec->render.exec = exec; | ||
| 476 | exec->render.start = args->rcl_start; | ||
| 477 | exec->render.end = args->rcl_end; | ||
| 478 | exec->v3d = v3d; | ||
| 479 | INIT_LIST_HEAD(&exec->unref_list); | ||
| 480 | |||
| 481 | ret = v3d_cl_lookup_bos(dev, file_priv, args, exec); | ||
| 482 | if (ret) | 521 | if (ret) |
| 483 | goto fail; | 522 | goto fail; |
| 484 | 523 | ||
| 485 | ret = v3d_lock_bo_reservations(exec->bo, exec->bo_count, | 524 | ret = v3d_lock_bo_reservations(render->base.bo, render->base.bo_count, |
| 486 | &acquire_ctx); | 525 | &acquire_ctx); |
| 487 | if (ret) | 526 | if (ret) |
| 488 | goto fail; | 527 | goto fail; |
| 489 | 528 | ||
| 490 | mutex_lock(&v3d->sched_lock); | 529 | mutex_lock(&v3d->sched_lock); |
| 491 | if (exec->bin.start != exec->bin.end) { | 530 | if (bin) { |
| 492 | ret = drm_sched_job_init(&exec->bin.base, | 531 | ret = v3d_push_job(v3d_priv, &bin->base, V3D_BIN); |
| 493 | &v3d_priv->sched_entity[V3D_BIN], | ||
| 494 | v3d_priv); | ||
| 495 | if (ret) | 532 | if (ret) |
| 496 | goto fail_unreserve; | 533 | goto fail_unreserve; |
| 497 | 534 | ||
| 498 | exec->bin_done_fence = | 535 | render->bin_done_fence = dma_fence_get(bin->base.done_fence); |
| 499 | dma_fence_get(&exec->bin.base.s_fence->finished); | ||
| 500 | |||
| 501 | kref_get(&exec->refcount); /* put by scheduler job completion */ | ||
| 502 | drm_sched_entity_push_job(&exec->bin.base, | ||
| 503 | &v3d_priv->sched_entity[V3D_BIN]); | ||
| 504 | } | 536 | } |
| 505 | 537 | ||
| 506 | ret = drm_sched_job_init(&exec->render.base, | 538 | ret = v3d_push_job(v3d_priv, &render->base, V3D_RENDER); |
| 507 | &v3d_priv->sched_entity[V3D_RENDER], | ||
| 508 | v3d_priv); | ||
| 509 | if (ret) | 539 | if (ret) |
| 510 | goto fail_unreserve; | 540 | goto fail_unreserve; |
| 511 | |||
| 512 | exec->render_done_fence = | ||
| 513 | dma_fence_get(&exec->render.base.s_fence->finished); | ||
| 514 | |||
| 515 | kref_get(&exec->refcount); /* put by scheduler job completion */ | ||
| 516 | drm_sched_entity_push_job(&exec->render.base, | ||
| 517 | &v3d_priv->sched_entity[V3D_RENDER]); | ||
| 518 | mutex_unlock(&v3d->sched_lock); | 541 | mutex_unlock(&v3d->sched_lock); |
| 519 | 542 | ||
| 520 | v3d_attach_object_fences(exec->bo, exec->bo_count, | 543 | v3d_attach_fences_and_unlock_reservation(file_priv, |
| 521 | exec->render_done_fence); | 544 | &render->base, &acquire_ctx, |
| 522 | 545 | args->out_sync); | |
| 523 | drm_gem_unlock_reservations(exec->bo, exec->bo_count, &acquire_ctx); | ||
| 524 | |||
| 525 | /* Update the return sync object for the */ | ||
| 526 | sync_out = drm_syncobj_find(file_priv, args->out_sync); | ||
| 527 | if (sync_out) { | ||
| 528 | drm_syncobj_replace_fence(sync_out, exec->render_done_fence); | ||
| 529 | drm_syncobj_put(sync_out); | ||
| 530 | } | ||
| 531 | 546 | ||
| 532 | v3d_exec_put(exec); | 547 | if (bin) |
| 548 | v3d_job_put(&bin->base); | ||
| 549 | v3d_job_put(&render->base); | ||
| 533 | 550 | ||
| 534 | return 0; | 551 | return 0; |
| 535 | 552 | ||
| 536 | fail_unreserve: | 553 | fail_unreserve: |
| 537 | mutex_unlock(&v3d->sched_lock); | 554 | mutex_unlock(&v3d->sched_lock); |
| 538 | drm_gem_unlock_reservations(exec->bo, exec->bo_count, &acquire_ctx); | 555 | drm_gem_unlock_reservations(render->base.bo, |
| 556 | render->base.bo_count, &acquire_ctx); | ||
| 539 | fail: | 557 | fail: |
| 540 | v3d_exec_put(exec); | 558 | if (bin) |
| 559 | v3d_job_put(&bin->base); | ||
| 560 | v3d_job_put(&render->base); | ||
| 541 | 561 | ||
| 542 | return ret; | 562 | return ret; |
| 543 | } | 563 | } |
| @@ -560,10 +580,7 @@ v3d_submit_tfu_ioctl(struct drm_device *dev, void *data, · 16 unmodified lines | |||
| 560 | struct drm_v3d_submit_tfu *args = data; | 580 | struct drm_v3d_submit_tfu *args = data; |
| 561 | struct v3d_tfu_job *job; | 581 | struct v3d_tfu_job *job; |
| 562 | struct ww_acquire_ctx acquire_ctx; | 582 | struct ww_acquire_ctx acquire_ctx; |
| 563 | struct drm_syncobj *sync_out; | ||
| 564 | struct dma_fence *sched_done_fence; | ||
| 565 | int ret = 0; | 583 | int ret = 0; |
| 566 | int bo_count; | ||
| 567 | 584 | ||
| 568 | trace_v3d_submit_tfu_ioctl(&v3d->drm, args->iia); | 585 | trace_v3d_submit_tfu_ioctl(&v3d->drm, args->iia); |
| 569 | 586 | ||
| @@ -571,81 +588,71 @@ v3d_submit_tfu_ioctl(struct drm_device *dev, void *data, · 1 unmodified line | |||
| 571 | if (!job) | 588 | if (!job) |
| 572 | return -ENOMEM; | 589 | return -ENOMEM; |
| 573 | 590 | ||
| 574 | ret = pm_runtime_get_sync(v3d->dev); | 591 | ret = v3d_job_init(v3d, file_priv, &job->base, |
| 575 | if (ret < 0) { | 592 | v3d_job_free, args->in_sync); |
| 593 | if (ret) { | ||
| 576 | kfree(job); | 594 | kfree(job); |
| 577 | return ret; | 595 | return ret; |
| 578 | } | 596 | } |
| 579 | 597 | ||
| 580 | kref_init(&job->refcount); | 598 | job->base.bo = kcalloc(ARRAY_SIZE(args->bo_handles), |
| 581 | 599 | sizeof(*job->base.bo), GFP_KERNEL); | |
| 582 | ret = drm_syncobj_find_fence(file_priv, args->in_sync, | 600 | if (!job->base.bo) { |
| 583 | 0, 0, &job->in_fence); | 601 | v3d_job_put(&job->base); |
| 584 | if (ret == -EINVAL) | 602 | return -ENOMEM; |
| 585 | goto fail; | 603 | } |
| 586 | 604 | ||
| 587 | job->args = *args; | 605 | job->args = *args; |
| 588 | job->v3d = v3d; | ||
| 589 | 606 | ||
| 590 | spin_lock(&file_priv->table_lock); | 607 | spin_lock(&file_priv->table_lock); |
| 591 | for (bo_count = 0; bo_count < ARRAY_SIZE(job->bo); bo_count++) { | 608 | for (job->base.bo_count = 0; |
| 609 | job->base.bo_count < ARRAY_SIZE(args->bo_handles); | ||
| 610 | job->base.bo_count++) { | ||
| 592 | struct drm_gem_object *bo; | 611 | struct drm_gem_object *bo; |
| 593 | 612 | ||
| 594 | if (!args->bo_handles[bo_count]) | 613 | if (!args->bo_handles[job->base.bo_count]) |
| 595 | break; | 614 | break; |
| 596 | 615 | ||
| 597 | bo = idr_find(&file_priv->object_idr, | 616 | bo = idr_find(&file_priv->object_idr, |
| 598 | args->bo_handles[bo_count]); | 617 | args->bo_handles[job->base.bo_count]); |
| 599 | if (!bo) { | 618 | if (!bo) { |
| 600 | DRM_DEBUG("Failed to look up GEM BO %d: %d\n", | 619 | DRM_DEBUG("Failed to look up GEM BO %d: %d\n", |
| 601 | bo_count, args->bo_handles[bo_count]); | 620 | job->base.bo_count, |
| 621 | args->bo_handles[job->base.bo_count]); | ||
| 602 | ret = -ENOENT; | 622 | ret = -ENOENT; |
| 603 | spin_unlock(&file_priv->table_lock); | 623 | spin_unlock(&file_priv->table_lock); |
| 604 | goto fail; | 624 | goto fail; |
| 605 | } | 625 | } |
| 606 | drm_gem_object_get(bo); | 626 | drm_gem_object_get(bo); |
| 607 | job->bo[bo_count] = bo; | 627 | job->base.bo[job->base.bo_count] = bo; |
| 608 | } | 628 | } |
| 609 | spin_unlock(&file_priv->table_lock); | 629 | spin_unlock(&file_priv->table_lock); |
| 610 | 630 | ||
| 611 | ret = v3d_lock_bo_reservations(job->bo, bo_count, &acquire_ctx); | 631 | ret = v3d_lock_bo_reservations(job->base.bo, job->base.bo_count, |
| 632 | &acquire_ctx); | ||
| 612 | if (ret) | 633 | if (ret) |
| 613 | goto fail; | 634 | goto fail; |
| 614 | 635 | ||
| 615 | mutex_lock(&v3d->sched_lock); | 636 | mutex_lock(&v3d->sched_lock); |
| 616 | ret = drm_sched_job_init(&job->base, | 637 | ret = v3d_push_job(v3d_priv, &job->base, V3D_TFU); |
| 617 | &v3d_priv->sched_entity[V3D_TFU], | ||
| 618 | v3d_priv); | ||
| 619 | if (ret) | 638 | if (ret) |
| 620 | goto fail_unreserve; | 639 | goto fail_unreserve; |
| 621 | |||
| 622 | sched_done_fence = dma_fence_get(&job->base.s_fence->finished); | ||
| 623 | |||
| 624 | kref_get(&job->refcount); /* put by scheduler job completion */ | ||
| 625 | drm_sched_entity_push_job(&job->base, &v3d_priv->sched_entity[V3D_TFU]); | ||
| 626 | mutex_unlock(&v3d->sched_lock); | 640 | mutex_unlock(&v3d->sched_lock); |
| 627 | 641 | ||
| 628 | v3d_attach_object_fences(job->bo, bo_count, sched_done_fence); | 642 | v3d_attach_fences_and_unlock_reservation(file_priv, |
| 629 | 643 | &job->base, &acquire_ctx, | |
| 630 | drm_gem_unlock_reservations(job->bo, bo_count, &acquire_ctx); | 644 | args->out_sync); |
| 631 | |||
| 632 | /* Update the return sync object */ | ||
| 633 | sync_out = drm_syncobj_find(file_priv, args->out_sync); | ||
| 634 | if (sync_out) { | ||
| 635 | drm_syncobj_replace_fence(sync_out, sched_done_fence); | ||
| 636 | drm_syncobj_put(sync_out); | ||
| 637 | } | ||
| 638 | dma_fence_put(sched_done_fence); | ||
| 639 | 645 | ||
| 640 | v3d_tfu_job_put(job); | 646 | v3d_job_put(&job->base); |
| 641 | 647 | ||
| 642 | return 0; | 648 | return 0; |
| 643 | 649 | ||
| 644 | fail_unreserve: | 650 | fail_unreserve: |
| 645 | mutex_unlock(&v3d->sched_lock); | 651 | mutex_unlock(&v3d->sched_lock); |
| 646 | drm_gem_unlock_reservations(job->bo, bo_count, &acquire_ctx); | 652 | drm_gem_unlock_reservations(job->base.bo, job->base.bo_count, |
| 653 | &acquire_ctx); | ||
| 647 | fail: | 654 | fail: |
| 648 | v3d_tfu_job_put(job); | 655 | v3d_job_put(&job->base); |
| 649 | 656 | ||
| 650 | return ret; | 657 | return ret; |
| 651 | } | 658 | } |
| @@ -703,7 +710,7 @@ v3d_gem_destroy(struct drm_device *dev) · 51 unmodified lines | |||
| 703 | 710 | ||
| 704 | v3d_sched_fini(v3d); | 711 | v3d_sched_fini(v3d); |
| 705 | 712 | ||
| 706 | /* Waiting for exec to finish would need to be done before | 713 | /* Waiting for jobs to finish would need to be done before |
| 707 | * unregistering V3D. | 714 | * unregistering V3D. |
| 708 | */ | 715 | */ |
| 709 | WARN_ON(v3d->bin_job); | 716 | WARN_ON(v3d->bin_job); |
| @@ -62,7 +62,7 @@ v3d_overflow_mem_work(struct work_struct *work) · 61 unmodified lines | |||
| 62 | } | 62 | } |
| 63 | 63 | ||
| 64 | drm_gem_object_get(obj); | 64 | drm_gem_object_get(obj); |
| 65 | list_add_tail(&bo->unref_head, &v3d->bin_job->unref_list); | 65 | list_add_tail(&bo->unref_head, &v3d->bin_job->render->unref_list); |
| 66 | spin_unlock_irqrestore(&v3d->job_lock, irqflags); | 66 | spin_unlock_irqrestore(&v3d->job_lock, irqflags); |
| 67 | 67 | ||
| 68 | V3D_CORE_WRITE(0, V3D_PTB_BPOA, bo->node.start << PAGE_SHIFT); | 68 | V3D_CORE_WRITE(0, V3D_PTB_BPOA, bo->node.start << PAGE_SHIFT); |
| @@ -96,7 +96,7 @@ v3d_irq(int irq, void *arg) · 27 unmodified lines | |||
| 96 | 96 | ||
| 97 | if (intsts & V3D_INT_FLDONE) { | 97 | if (intsts & V3D_INT_FLDONE) { |
| 98 | struct v3d_fence *fence = | 98 | struct v3d_fence *fence = |
| 99 | to_v3d_fence(v3d->bin_job->bin.irq_fence); | 99 | to_v3d_fence(v3d->bin_job->base.irq_fence); |
| 100 | 100 | ||
| 101 | trace_v3d_bcl_irq(&v3d->drm, fence->seqno); | 101 | trace_v3d_bcl_irq(&v3d->drm, fence->seqno); |
| 102 | dma_fence_signal(&fence->base); | 102 | dma_fence_signal(&fence->base); |
| @@ -105,7 +105,7 @@ v3d_irq(int irq, void *arg) · 2 unmodified lines | |||
| 105 | 105 | ||
| 106 | if (intsts & V3D_INT_FRDONE) { | 106 | if (intsts & V3D_INT_FRDONE) { |
| 107 | struct v3d_fence *fence = | 107 | struct v3d_fence *fence = |
| 108 | to_v3d_fence(v3d->render_job->render.irq_fence); | 108 | to_v3d_fence(v3d->render_job->base.irq_fence); |
| 109 | 109 | ||
| 110 | trace_v3d_rcl_irq(&v3d->drm, fence->seqno); | 110 | trace_v3d_rcl_irq(&v3d->drm, fence->seqno); |
| 111 | dma_fence_signal(&fence->base); | 111 | dma_fence_signal(&fence->base); |
| @@ -141,7 +141,7 @@ v3d_hub_irq(int irq, void *arg) · 29 unmodified lines | |||
| 141 | 141 | ||
| 142 | if (intsts & V3D_HUB_INT_TFUC) { | 142 | if (intsts & V3D_HUB_INT_TFUC) { |
| 143 | struct v3d_fence *fence = | 143 | struct v3d_fence *fence = |
| 144 | to_v3d_fence(v3d->tfu_job->irq_fence); | 144 | to_v3d_fence(v3d->tfu_job->base.irq_fence); |
| 145 | 145 | ||
| 146 | trace_v3d_tfu_irq(&v3d->drm, fence->seqno); | 146 | trace_v3d_tfu_irq(&v3d->drm, fence->seqno); |
| 147 | dma_fence_signal(&fence->base); | 147 | dma_fence_signal(&fence->base); |
| @@ -30,43 +30,44 @@ to_v3d_job(struct drm_sched_job *sched_job) · 29 unmodified lines | |||
| 30 | return container_of(sched_job, struct v3d_job, base); | 30 | return container_of(sched_job, struct v3d_job, base); |
| 31 | } | 31 | } |
| 32 | 32 | ||
| 33 | static struct v3d_tfu_job * | 33 | static struct v3d_bin_job * |
| 34 | to_tfu_job(struct drm_sched_job *sched_job) | 34 | to_bin_job(struct drm_sched_job *sched_job) |
| 35 | { | 35 | { |
| 36 | return container_of(sched_job, struct v3d_tfu_job, base); | 36 | return container_of(sched_job, struct v3d_bin_job, base.base); |
| 37 | } | 37 | } |
| 38 | 38 | ||
| 39 | static void | 39 | static struct v3d_render_job * |
| 40 | v3d_job_free(struct drm_sched_job *sched_job) | 40 | to_render_job(struct drm_sched_job *sched_job) |
| 41 | { | 41 | { |
| 42 | struct v3d_job *job = to_v3d_job(sched_job); | 42 | return container_of(sched_job, struct v3d_render_job, base.base); |
| 43 | 43 | } | |
| 44 | drm_sched_job_cleanup(sched_job); | ||
| 45 | 44 | ||
| 46 | v3d_exec_put(job->exec); | 45 | static struct v3d_tfu_job * |
| 46 | to_tfu_job(struct drm_sched_job *sched_job) | ||
| 47 | { | ||
| 48 | return container_of(sched_job, struct v3d_tfu_job, base.base); | ||
| 47 | } | 49 | } |
| 48 | 50 | ||
| 49 | static void | 51 | static void |
| 50 | v3d_tfu_job_free(struct drm_sched_job *sched_job) | 52 | v3d_job_free(struct drm_sched_job *sched_job) |
| 51 | { | 53 | { |
| 52 | struct v3d_tfu_job *job = to_tfu_job(sched_job); | 54 | struct v3d_job *job = to_v3d_job(sched_job); |
| 53 | 55 | ||
| 54 | drm_sched_job_cleanup(sched_job); | 56 | drm_sched_job_cleanup(sched_job); |
| 55 | 57 | v3d_job_put(job); | |
| 56 | v3d_tfu_job_put(job); | ||
| 57 | } | 58 | } |
| 58 | 59 | ||
| 59 | /** | 60 | /** |
| 60 | * Returns the fences that the bin or render job depends on, one by one. | 61 | * Returns the fences that the job depends on, one by one. |
| 61 | * v3d_job_run() won't be called until all of them have been signaled. | 62 | * |
| 63 | * If placed in the scheduler's .dependency method, the corresponding | ||
| 64 | * .run_job won't be called until all of them have been signaled. | ||
| 62 | */ | 65 | */ |
| 63 | static struct dma_fence * | 66 | static struct dma_fence * |
| 64 | v3d_job_dependency(struct drm_sched_job *sched_job, | 67 | v3d_job_dependency(struct drm_sched_job *sched_job, |
| 65 | struct drm_sched_entity *s_entity) | 68 | struct drm_sched_entity *s_entity) |
| 66 | { | 69 | { |
| 67 | struct v3d_job *job = to_v3d_job(sched_job); | 70 | struct v3d_job *job = to_v3d_job(sched_job); |
| 68 | struct v3d_exec_info *exec = job->exec; | ||
| 69 | enum v3d_queue q = job == &exec->bin ? V3D_BIN : V3D_RENDER; | ||
| 70 | struct dma_fence *fence; | 71 | struct dma_fence *fence; |
| 71 | 72 | ||
| 72 | fence = job->in_fence; | 73 | fence = job->in_fence; |
| @@ -75,113 +76,132 @@ v3d_job_dependency(struct drm_sched_job *sched_job, · 2 unmodified lines | |||
| 75 | return fence; | 76 | return fence; |
| 76 | } | 77 | } |
| 77 | 78 | ||
| 78 | if (q == V3D_RENDER) { | ||
| 79 | /* If we had a bin job, the render job definitely depends on | ||
| 80 | * it. We first have to wait for bin to be scheduled, so that | ||
| 81 | * its done_fence is created. | ||
| 82 | */ | ||
| 83 | fence = exec->bin_done_fence; | ||
| 84 | if (fence) { | ||
| 85 | exec->bin_done_fence = NULL; | ||
| 86 | return fence; | ||
| 87 | } | ||
| 88 | } | ||
| 89 | |||
| 90 | /* XXX: Wait on a fence for switching the GMP if necessary, | 79 | /* XXX: Wait on a fence for switching the GMP if necessary, |
| 91 | * and then do so. | 80 | * and then do so. |
| 92 | */ | 81 | */ |
| 93 | 82 | ||
| 94 | return fence; | 83 | return NULL; |
| 95 | } | 84 | } |
| 96 | 85 | ||
| 97 | /** | 86 | /** |
| 98 | * Returns the fences that the TFU job depends on, one by one. | 87 | * Returns the fences that the render job depends on, one by one. |
| 99 | * v3d_tfu_job_run() won't be called until all of them have been | 88 | * v3d_job_run() won't be called until all of them have been signaled. |
| 100 | * signaled. | ||
| 101 | */ | 89 | */ |
| 102 | static struct dma_fence * | 90 | static struct dma_fence * |
| 103 | v3d_tfu_job_dependency(struct drm_sched_job *sched_job, | 91 | v3d_render_job_dependency(struct drm_sched_job *sched_job, |
| 104 | struct drm_sched_entity *s_entity) | 92 | struct drm_sched_entity *s_entity) |
| 105 | { | 93 | { |
| 106 | struct v3d_tfu_job *job = to_tfu_job(sched_job); | 94 | struct v3d_render_job *job = to_render_job(sched_job); |
| 107 | struct dma_fence *fence; | 95 | struct dma_fence *fence; |
| 108 | 96 | ||
| 109 | fence = job->in_fence; | 97 | fence = v3d_job_dependency(sched_job, s_entity); |
| 98 | if (fence) | ||
| 99 | return fence; | ||
| 100 | |||
| 101 | /* If we had a bin job, the render job definitely depends on | ||
| 102 | * it. We first have to wait for bin to be scheduled, so that | ||
| 103 | * its done_fence is created. | ||
| 104 | */ | ||
| 105 | fence = job->bin_done_fence; | ||
| 110 | if (fence) { | 106 | if (fence) { |
| 111 | job->in_fence = NULL; | 107 | job->bin_done_fence = NULL; |
| 112 | return fence; | 108 | return fence; |
| 113 | } | 109 | } |
| 114 | 110 | ||
| 115 | return NULL; | 111 | return fence; |
| 116 | } | 112 | } |
| 117 | 113 | ||
| 118 | static struct dma_fence *v3d_job_run(struct drm_sched_job *sched_job) | 114 | static struct dma_fence *v3d_bin_job_run(struct drm_sched_job *sched_job) |
| 119 | { | 115 | { |
| 120 | struct v3d_job *job = to_v3d_job(sched_job); | 116 | struct v3d_bin_job *job = to_bin_job(sched_job); |
| 121 | struct v3d_exec_info *exec = job->exec; | 117 | struct v3d_dev *v3d = job->base.v3d; |
| 122 | enum v3d_queue q = job == &exec->bin ? V3D_BIN : V3D_RENDER; | ||
| 123 | struct v3d_dev *v3d = exec->v3d; | ||
| 124 | struct drm_device *dev = &v3d->drm; | 118 | struct drm_device *dev = &v3d->drm; |
| 125 | struct dma_fence *fence; | 119 | struct dma_fence *fence; |
| 126 | unsigned long irqflags; | 120 | unsigned long irqflags; |
| 127 | 121 | ||
| 128 | if (unlikely(job->base.s_fence->finished.error)) | 122 | if (unlikely(job->base.base.s_fence->finished.error)) |
| 129 | return NULL; | 123 | return NULL; |
| 130 | 124 | ||
| 131 | /* Lock required around bin_job update vs | 125 | /* Lock required around bin_job update vs |
| 132 | * v3d_overflow_mem_work(). | 126 | * v3d_overflow_mem_work(). |
| 133 | */ | 127 | */ |
| 134 | spin_lock_irqsave(&v3d->job_lock, irqflags); | 128 | spin_lock_irqsave(&v3d->job_lock, irqflags); |
| 135 | if (q == V3D_BIN) { | 129 | v3d->bin_job = job; |
| 136 | v3d->bin_job = job->exec; | 130 | /* Clear out the overflow allocation, so we don't |
| 137 | 131 | * reuse the overflow attached to a previous job. | |
| 138 | /* Clear out the overflow allocation, so we don't | 132 | */ |
| 139 | * reuse the overflow attached to a previous job. | 133 | V3D_CORE_WRITE(0, V3D_PTB_BPOS, 0); |
| 140 | */ | ||
| 141 | V3D_CORE_WRITE(0, V3D_PTB_BPOS, 0); | ||
| 142 | } else { | ||
| 143 | v3d->render_job = job->exec; | ||
| 144 | } | ||
| 145 | spin_unlock_irqrestore(&v3d->job_lock, irqflags); | 134 | spin_unlock_irqrestore(&v3d->job_lock, irqflags); |
| 146 | 135 | ||
| 147 | /* Can we avoid this flush when q==RENDER? We need to be | 136 | v3d_invalidate_caches(v3d); |
| 148 | * careful of scheduling, though -- imagine job0 rendering to | 137 | |
| 149 | * texture and job1 reading, and them being executed as bin0, | 138 | fence = v3d_fence_create(v3d, V3D_BIN); |
| 150 | * bin1, render0, render1, so that render1's flush at bin time | 139 | if (IS_ERR(fence)) |
| 140 | return NULL; | ||
| 141 | |||
| 142 | if (job->base.irq_fence) | ||
| 143 | dma_fence_put(job->base.irq_fence); | ||
| 144 | job->base.irq_fence = dma_fence_get(fence); | ||
| 145 | |||
| 146 | trace_v3d_submit_cl(dev, false, to_v3d_fence(fence)->seqno, | ||
| 147 | job->start, job->end); | ||
| 148 | |||
| 149 | /* Set the current and end address of the control list. | ||
| 150 | * Writing the end register is what starts the job. | ||
| 151 | */ | ||
| 152 | if (job->qma) { | ||
| 153 | V3D_CORE_WRITE(0, V3D_CLE_CT0QMA, job->qma); | ||
| 154 | V3D_CORE_WRITE(0, V3D_CLE_CT0QMS, job->qms); | ||
| 155 | } | ||
| 156 | if (job->qts) { | ||
| 157 | V3D_CORE_WRITE(0, V3D_CLE_CT0QTS, | ||
| 158 | V3D_CLE_CT0QTS_ENABLE | | ||
| 159 | job->qts); | ||
| 160 | } | ||
| 161 | V3D_CORE_WRITE(0, V3D_CLE_CT0QBA, job->start); | ||
| 162 | V3D_CORE_WRITE(0, V3D_CLE_CT0QEA, job->end); | ||
| 163 | |||
| 164 | return fence; | ||
| 165 | } | ||
| 166 | |||
| 167 | static struct dma_fence *v3d_render_job_run(struct drm_sched_job *sched_job) | ||
| 168 | { | ||
| 169 | struct v3d_render_job *job = to_render_job(sched_job); | ||
| 170 | struct v3d_dev *v3d = job->base.v3d; | ||
| 171 | struct drm_device *dev = &v3d->drm; | ||
| 172 | struct dma_fence *fence; | ||
| 173 | |||
| 174 | if (unlikely(job->base.base.s_fence->finished.error)) | ||
| 175 | return NULL; | ||
| 176 | |||
| 177 | v3d->render_job = job; | ||
| 178 | |||
| 179 | /* Can we avoid this flush? We need to be careful of | ||
| 180 | * scheduling, though -- imagine job0 rendering to texture and | ||
| 181 | * job1 reading, and them being executed as bin0, bin1, | ||
| 182 | * render0, render1, so that render1's flush at bin time | ||
| 151 | * wasn't enough. | 183 | * wasn't enough. |
| 152 | */ | 184 | */ |
| 153 | v3d_invalidate_caches(v3d); | 185 | v3d_invalidate_caches(v3d); |
| 154 | 186 | ||
| 155 | fence = v3d_fence_create(v3d, q); | 187 | fence = v3d_fence_create(v3d, V3D_RENDER); |
| 156 | if (IS_ERR(fence)) | 188 | if (IS_ERR(fence)) |
| 157 | return NULL; | 189 | return NULL; |
| 158 | 190 | ||
| 159 | if (job->irq_fence) | 191 | if (job->base.irq_fence) |
| 160 | dma_fence_put(job->irq_fence); | 192 | dma_fence_put(job->base.irq_fence); |
| 161 | job->irq_fence = dma_fence_get(fence); | 193 | job->base.irq_fence = dma_fence_get(fence); |
| 162 | 194 | ||
| 163 | trace_v3d_submit_cl(dev, q == V3D_RENDER, to_v3d_fence(fence)->seqno, | 195 | trace_v3d_submit_cl(dev, true, to_v3d_fence(fence)->seqno, |
| 164 | job->start, job->end); | 196 | job->start, job->end); |
| 165 | 197 | ||
| 166 | if (q == V3D_BIN) { | 198 | /* XXX: Set the QCFG */ |
| 167 | if (exec->qma) { | ||
| 168 | V3D_CORE_WRITE(0, V3D_CLE_CT0QMA, exec->qma); | ||
| 169 | V3D_CORE_WRITE(0, V3D_CLE_CT0QMS, exec->qms); | ||
| 170 | } | ||
| 171 | if (exec->qts) { | ||
| 172 | V3D_CORE_WRITE(0, V3D_CLE_CT0QTS, | ||
| 173 | V3D_CLE_CT0QTS_ENABLE | | ||
| 174 | exec->qts); | ||
| 175 | } | ||
| 176 | } else { | ||
| 177 | /* XXX: Set the QCFG */ | ||
| 178 | } | ||
| 179 | 199 | ||
| 180 | /* Set the current and end address of the control list. | 200 | /* Set the current and end address of the control list. |
| 181 | * Writing the end register is what starts the job. | 201 | * Writing the end register is what starts the job. |
| 182 | */ | 202 | */ |
| 183 | V3D_CORE_WRITE(0, V3D_CLE_CTNQBA(q), job->start); | 203 | V3D_CORE_WRITE(0, V3D_CLE_CT1QBA, job->start); |
| 184 | V3D_CORE_WRITE(0, V3D_CLE_CTNQEA(q), job->end); | 204 | V3D_CORE_WRITE(0, V3D_CLE_CT1QEA, job->end); |
| 185 | 205 | ||
| 186 | return fence; | 206 | return fence; |
| 187 | } | 207 | } |
| @@ -190,7 +210,7 @@ static struct dma_fence * · 2 unmodified lines | |||
| 190 | v3d_tfu_job_run(struct drm_sched_job *sched_job) | 210 | v3d_tfu_job_run(struct drm_sched_job *sched_job) |
| 191 | { | 211 | { |
| 192 | struct v3d_tfu_job *job = to_tfu_job(sched_job); | 212 | struct v3d_tfu_job *job = to_tfu_job(sched_job); |
| 193 | struct v3d_dev *v3d = job->v3d; | 213 | struct v3d_dev *v3d = job->base.v3d; |
| 194 | struct drm_device *dev = &v3d->drm; | 214 | struct drm_device *dev = &v3d->drm; |
| 195 | struct dma_fence *fence; | 215 | struct dma_fence *fence; |
| 196 | 216 | ||
| @@ -199,9 +219,9 @@ v3d_tfu_job_run(struct drm_sched_job *sched_job) · 2 unmodified lines | |||
| 199 | return NULL; | 219 | return NULL; |
| 200 | 220 | ||
| 201 | v3d->tfu_job = job; | 221 | v3d->tfu_job = job; |
| 202 | if (job->irq_fence) | 222 | if (job->base.irq_fence) |
| 203 | dma_fence_put(job->irq_fence); | 223 | dma_fence_put(job->base.irq_fence); |
| 204 | job->irq_fence = dma_fence_get(fence); | 224 | job->base.irq_fence = dma_fence_get(fence); |
| 205 | 225 | ||
| 206 | trace_v3d_submit_tfu(dev, to_v3d_fence(fence)->seqno); | 226 | trace_v3d_submit_tfu(dev, to_v3d_fence(fence)->seqno); |
| 207 | 227 | ||
| @@ -251,51 +271,74 @@ v3d_gpu_reset_for_timeout(struct v3d_dev *v3d, struct drm_sched_job *sched_job) · 43 unmodified lines | |||
| 251 | mutex_unlock(&v3d->reset_lock); | 271 | mutex_unlock(&v3d->reset_lock); |
| 252 | } | 272 | } |
| 253 | 273 | ||
| 274 | /* If the current address or return address have changed, then the GPU | ||
| 275 | * has probably made progress and we should delay the reset. This | ||
| 276 | * could fail if the GPU got in an infinite loop in the CL, but that | ||
| 277 | * is pretty unlikely outside of an i-g-t testcase. | ||
| 278 | */ | ||
| 254 | static void | 279 | static void |
| 255 | v3d_job_timedout(struct drm_sched_job *sched_job) | 280 | v3d_cl_job_timedout(struct drm_sched_job *sched_job, enum v3d_queue q, |
| 281 | u32 *timedout_ctca, u32 *timedout_ctra) | ||
| 256 | { | 282 | { |
| 257 | struct v3d_job *job = to_v3d_job(sched_job); | 283 | struct v3d_job *job = to_v3d_job(sched_job); |
| 258 | struct v3d_exec_info *exec = job->exec; | 284 | struct v3d_dev *v3d = job->v3d; |
| 259 | struct v3d_dev *v3d = exec->v3d; | 285 | u32 ctca = V3D_CORE_READ(0, V3D_CLE_CTNCA(q)); |
| 260 | enum v3d_queue job_q = job == &exec->bin ? V3D_BIN : V3D_RENDER; | 286 | u32 ctra = V3D_CORE_READ(0, V3D_CLE_CTNRA(q)); |
| 261 | u32 ctca = V3D_CORE_READ(0, V3D_CLE_CTNCA(job_q)); | 287 | |
| 262 | u32 ctra = V3D_CORE_READ(0, V3D_CLE_CTNRA(job_q)); | 288 | if (*timedout_ctca != ctca || *timedout_ctra != ctra) { |
| 263 | 289 | *timedout_ctca = ctca; | |
| 264 | /* If the current address or return address have changed, then | 290 | *timedout_ctra = ctra; |
| 265 | * the GPU has probably made progress and we should delay the | ||
| 266 | * reset. This could fail if the GPU got in an infinite loop | ||
| 267 | * in the CL, but that is pretty unlikely outside of an i-g-t | ||
| 268 | * testcase. | ||
| 269 | */ | ||
| 270 | if (job->timedout_ctca != ctca || job->timedout_ctra != ctra) { | ||
| 271 | job->timedout_ctca = ctca; | ||
| 272 | job->timedout_ctra = ctra; | ||
| 273 | return; | 291 | return; |
| 274 | } | 292 | } |
| 275 | 293 | ||
| 276 | v3d_gpu_reset_for_timeout(v3d, sched_job); | 294 | v3d_gpu_reset_for_timeout(v3d, sched_job); |
| 277 | } | 295 | } |
| 278 | 296 | ||
| 297 | static void | ||
| 298 | v3d_bin_job_timedout(struct drm_sched_job *sched_job) | ||
| 299 | { | ||
| 300 | struct v3d_bin_job *job = to_bin_job(sched_job); | ||
| 301 | |||
| 302 | v3d_cl_job_timedout(sched_job, V3D_BIN, | ||
| 303 | &job->timedout_ctca, &job->timedout_ctra); | ||
| 304 | } | ||
| 305 | |||
| 306 | static void | ||
| 307 | v3d_render_job_timedout(struct drm_sched_job *sched_job) | ||
| 308 | { | ||
| 309 | struct v3d_render_job *job = to_render_job(sched_job); | ||
| 310 | |||
| 311 | v3d_cl_job_timedout(sched_job, V3D_RENDER, | ||
| 312 | &job->timedout_ctca, &job->timedout_ctra); | ||
| 313 | } | ||
| 314 | |||
| 279 | static void | 315 | static void |
| 280 | v3d_tfu_job_timedout(struct drm_sched_job *sched_job) | 316 | v3d_tfu_job_timedout(struct drm_sched_job *sched_job) |
| 281 | { | 317 | { |
| 282 | struct v3d_tfu_job *job = to_tfu_job(sched_job); | 318 | struct v3d_job *job = to_v3d_job(sched_job); |
| 283 | 319 | ||
| 284 | v3d_gpu_reset_for_timeout(job->v3d, sched_job); | 320 | v3d_gpu_reset_for_timeout(job->v3d, sched_job); |
| 285 | } | 321 | } |
| 286 | 322 | ||
| 287 | static const struct drm_sched_backend_ops v3d_sched_ops = { | 323 | static const struct drm_sched_backend_ops v3d_bin_sched_ops = { |
| 288 | .dependency = v3d_job_dependency, | 324 | .dependency = v3d_job_dependency, |
| 289 | .run_job = v3d_job_run, | 325 | .run_job = v3d_bin_job_run, |
| 290 | .timedout_job = v3d_job_timedout, | 326 | .timedout_job = v3d_bin_job_timedout, |
| 291 | .free_job = v3d_job_free | 327 | .free_job = v3d_job_free, |
| 328 | }; | ||
| 329 | |||
| 330 | static const struct drm_sched_backend_ops v3d_render_sched_ops = { | ||
| 331 | .dependency = v3d_render_job_dependency, | ||
| 332 | .run_job = v3d_render_job_run, | ||
| 333 | .timedout_job = v3d_render_job_timedout, | ||
| 334 | .free_job = v3d_job_free, | ||
| 292 | }; | 335 | }; |
| 293 | 336 | ||
| 294 | static const struct drm_sched_backend_ops v3d_tfu_sched_ops = { | 337 | static const struct drm_sched_backend_ops v3d_tfu_sched_ops = { |
| 295 | .dependency = v3d_tfu_job_dependency, | 338 | .dependency = v3d_job_dependency, |
| 296 | .run_job = v3d_tfu_job_run, | 339 | .run_job = v3d_tfu_job_run, |
| 297 | .timedout_job = v3d_tfu_job_timedout, | 340 | .timedout_job = v3d_tfu_job_timedout, |
| 298 | .free_job = v3d_tfu_job_free | 341 | .free_job = v3d_job_free, |
| 299 | }; | 342 | }; |
| 300 | 343 | ||
| 301 | int | 344 | int |
| @@ -307,7 +350,7 @@ v3d_sched_init(struct v3d_dev *v3d) · 5 unmodified lines | |||
| 307 | int ret; | 350 | int ret; |
| 308 | 351 | ||
| 309 | ret = drm_sched_init(&v3d->queue[V3D_BIN].sched, | 352 | ret = drm_sched_init(&v3d->queue[V3D_BIN].sched, |
| 310 | &v3d_sched_ops, | 353 | &v3d_bin_sched_ops, |
| 311 | hw_jobs_limit, job_hang_limit, | 354 | hw_jobs_limit, job_hang_limit, |
| 312 | msecs_to_jiffies(hang_limit_ms), | 355 | msecs_to_jiffies(hang_limit_ms), |
| 313 | "v3d_bin"); | 356 | "v3d_bin"); |
| @@ -317,7 +360,7 @@ v3d_sched_init(struct v3d_dev *v3d) · 3 unmodified lines | |||
| 317 | } | 360 | } |
| 318 | 361 | ||
| 319 | ret = drm_sched_init(&v3d->queue[V3D_RENDER].sched, | 362 | ret = drm_sched_init(&v3d->queue[V3D_RENDER].sched, |
| 320 | &v3d_sched_ops, | 363 | &v3d_render_sched_ops, |
| 321 | hw_jobs_limit, job_hang_limit, | 364 | hw_jobs_limit, job_hang_limit, |
| 322 | msecs_to_jiffies(hang_limit_ms), | 365 | msecs_to_jiffies(hang_limit_ms), |
| 323 | "v3d_render"); | 366 | "v3d_render"); |