1 // SPDX-License-Identifier: MIT 2 3 #include <linux/slab.h> 4 #include <drm/gpu_scheduler.h> 5 #include <drm/drm_syncobj.h> 6 7 #include "nouveau_drv.h" 8 #include "nouveau_gem.h" 9 #include "nouveau_mem.h" 10 #include "nouveau_dma.h" 11 #include "nouveau_exec.h" 12 #include "nouveau_abi16.h" 13 #include "nouveau_sched.h" 14 15 /* FIXME 16 * 17 * We want to make sure that jobs currently executing can't be deferred by 18 * other jobs competing for the hardware. Otherwise we might end up with job 19 * timeouts just because of too many clients submitting too many jobs. We don't 20 * want jobs to time out because of system load, but because of the job being 21 * too bulky. 22 * 23 * For now allow for up to 16 concurrent jobs in flight until we know how many 24 * rings the hardware can process in parallel. 25 */ 26 #define NOUVEAU_SCHED_HW_SUBMISSIONS 16 27 #define NOUVEAU_SCHED_JOB_TIMEOUT_MS 10000 28 29 int 30 nouveau_job_init(struct nouveau_job *job, 31 struct nouveau_job_args *args) 32 { 33 struct nouveau_sched_entity *entity = args->sched_entity; 34 int ret; 35 36 job->file_priv = args->file_priv; 37 job->cli = nouveau_cli(args->file_priv); 38 job->entity = entity; 39 40 job->sync = args->sync; 41 job->resv_usage = args->resv_usage; 42 43 job->ops = args->ops; 44 45 job->in_sync.count = args->in_sync.count; 46 if (job->in_sync.count) { 47 if (job->sync) 48 return -EINVAL; 49 50 job->in_sync.data = kmemdup(args->in_sync.s, 51 sizeof(*args->in_sync.s) * 52 args->in_sync.count, 53 GFP_KERNEL); 54 if (!job->in_sync.data) 55 return -ENOMEM; 56 } 57 58 job->out_sync.count = args->out_sync.count; 59 if (job->out_sync.count) { 60 if (job->sync) { 61 ret = -EINVAL; 62 goto err_free_in_sync; 63 } 64 65 job->out_sync.data = kmemdup(args->out_sync.s, 66 sizeof(*args->out_sync.s) * 67 args->out_sync.count, 68 GFP_KERNEL); 69 if (!job->out_sync.data) { 70 ret = -ENOMEM; 71 goto err_free_in_sync; 72 } 73 74 job->out_sync.objs = kcalloc(job->out_sync.count, 75 sizeof(*job->out_sync.objs), 76 GFP_KERNEL); 77 if (!job->out_sync.objs) { 78 ret = -ENOMEM; 79 goto err_free_out_sync; 80 } 81 82 job->out_sync.chains = kcalloc(job->out_sync.count, 83 sizeof(*job->out_sync.chains), 84 GFP_KERNEL); 85 if (!job->out_sync.chains) { 86 ret = -ENOMEM; 87 goto err_free_objs; 88 } 89 90 } 91 92 ret = drm_sched_job_init(&job->base, &entity->base, NULL); 93 if (ret) 94 goto err_free_chains; 95 96 job->state = NOUVEAU_JOB_INITIALIZED; 97 98 return 0; 99 100 err_free_chains: 101 kfree(job->out_sync.chains); 102 err_free_objs: 103 kfree(job->out_sync.objs); 104 err_free_out_sync: 105 kfree(job->out_sync.data); 106 err_free_in_sync: 107 kfree(job->in_sync.data); 108 return ret; 109 } 110 111 void 112 nouveau_job_free(struct nouveau_job *job) 113 { 114 kfree(job->in_sync.data); 115 kfree(job->out_sync.data); 116 kfree(job->out_sync.objs); 117 kfree(job->out_sync.chains); 118 } 119 120 void nouveau_job_fini(struct nouveau_job *job) 121 { 122 dma_fence_put(job->done_fence); 123 drm_sched_job_cleanup(&job->base); 124 job->ops->free(job); 125 } 126 127 static int 128 sync_find_fence(struct nouveau_job *job, 129 struct drm_nouveau_sync *sync, 130 struct dma_fence **fence) 131 { 132 u32 stype = sync->flags & DRM_NOUVEAU_SYNC_TYPE_MASK; 133 u64 point = 0; 134 int ret; 135 136 if (stype != DRM_NOUVEAU_SYNC_SYNCOBJ && 137 stype != DRM_NOUVEAU_SYNC_TIMELINE_SYNCOBJ) 138 return -EOPNOTSUPP; 139 140 if (stype == DRM_NOUVEAU_SYNC_TIMELINE_SYNCOBJ) 141 point = sync->timeline_value; 142 143 ret = drm_syncobj_find_fence(job->file_priv, 144 sync->handle, point, 145 0 /* flags */, fence); 146 if (ret) 147 return ret; 148 149 return 0; 150 } 151 152 static int 153 nouveau_job_add_deps(struct nouveau_job *job) 154 { 155 struct dma_fence *in_fence = NULL; 156 int ret, i; 157 158 for (i = 0; i < job->in_sync.count; i++) { 159 struct drm_nouveau_sync *sync = &job->in_sync.data[i]; 160 161 ret = sync_find_fence(job, sync, &in_fence); 162 if (ret) { 163 NV_PRINTK(warn, job->cli, 164 "Failed to find syncobj (-> in): handle=%d\n", 165 sync->handle); 166 return ret; 167 } 168 169 ret = drm_sched_job_add_dependency(&job->base, in_fence); 170 if (ret) 171 return ret; 172 } 173 174 return 0; 175 } 176 177 static void 178 nouveau_job_fence_attach_cleanup(struct nouveau_job *job) 179 { 180 int i; 181 182 for (i = 0; i < job->out_sync.count; i++) { 183 struct drm_syncobj *obj = job->out_sync.objs[i]; 184 struct dma_fence_chain *chain = job->out_sync.chains[i]; 185 186 if (obj) 187 drm_syncobj_put(obj); 188 189 if (chain) 190 dma_fence_chain_free(chain); 191 } 192 } 193 194 static int 195 nouveau_job_fence_attach_prepare(struct nouveau_job *job) 196 { 197 int i, ret; 198 199 for (i = 0; i < job->out_sync.count; i++) { 200 struct drm_nouveau_sync *sync = &job->out_sync.data[i]; 201 struct drm_syncobj **pobj = &job->out_sync.objs[i]; 202 struct dma_fence_chain **pchain = &job->out_sync.chains[i]; 203 u32 stype = sync->flags & DRM_NOUVEAU_SYNC_TYPE_MASK; 204 205 if (stype != DRM_NOUVEAU_SYNC_SYNCOBJ && 206 stype != DRM_NOUVEAU_SYNC_TIMELINE_SYNCOBJ) { 207 ret = -EINVAL; 208 goto err_sync_cleanup; 209 } 210 211 *pobj = drm_syncobj_find(job->file_priv, sync->handle); 212 if (!*pobj) { 213 NV_PRINTK(warn, job->cli, 214 "Failed to find syncobj (-> out): handle=%d\n", 215 sync->handle); 216 ret = -ENOENT; 217 goto err_sync_cleanup; 218 } 219 220 if (stype == DRM_NOUVEAU_SYNC_TIMELINE_SYNCOBJ) { 221 *pchain = dma_fence_chain_alloc(); 222 if (!*pchain) { 223 ret = -ENOMEM; 224 goto err_sync_cleanup; 225 } 226 } 227 } 228 229 return 0; 230 231 err_sync_cleanup: 232 nouveau_job_fence_attach_cleanup(job); 233 return ret; 234 } 235 236 static void 237 nouveau_job_fence_attach(struct nouveau_job *job) 238 { 239 struct dma_fence *fence = job->done_fence; 240 int i; 241 242 for (i = 0; i < job->out_sync.count; i++) { 243 struct drm_nouveau_sync *sync = &job->out_sync.data[i]; 244 struct drm_syncobj **pobj = &job->out_sync.objs[i]; 245 struct dma_fence_chain **pchain = &job->out_sync.chains[i]; 246 u32 stype = sync->flags & DRM_NOUVEAU_SYNC_TYPE_MASK; 247 248 if (stype == DRM_NOUVEAU_SYNC_TIMELINE_SYNCOBJ) { 249 drm_syncobj_add_point(*pobj, *pchain, fence, 250 sync->timeline_value); 251 } else { 252 drm_syncobj_replace_fence(*pobj, fence); 253 } 254 255 drm_syncobj_put(*pobj); 256 *pobj = NULL; 257 *pchain = NULL; 258 } 259 } 260 261 int 262 nouveau_job_submit(struct nouveau_job *job) 263 { 264 struct nouveau_sched_entity *entity = to_nouveau_sched_entity(job->base.entity); 265 struct dma_fence *done_fence = NULL; 266 int ret; 267 268 ret = nouveau_job_add_deps(job); 269 if (ret) 270 goto err; 271 272 ret = nouveau_job_fence_attach_prepare(job); 273 if (ret) 274 goto err; 275 276 /* Make sure the job appears on the sched_entity's queue in the same 277 * order as it was submitted. 278 */ 279 mutex_lock(&entity->mutex); 280 281 /* Guarantee we won't fail after the submit() callback returned 282 * successfully. 283 */ 284 if (job->ops->submit) { 285 ret = job->ops->submit(job); 286 if (ret) 287 goto err_cleanup; 288 } 289 290 drm_sched_job_arm(&job->base); 291 job->done_fence = dma_fence_get(&job->base.s_fence->finished); 292 if (job->sync) 293 done_fence = dma_fence_get(job->done_fence); 294 295 /* If a sched job depends on a dma-fence from a job from the same GPU 296 * scheduler instance, but a different scheduler entity, the GPU 297 * scheduler does only wait for the particular job to be scheduled, 298 * rather than for the job to fully complete. This is due to the GPU 299 * scheduler assuming that there is a scheduler instance per ring. 300 * However, the current implementation, in order to avoid arbitrary 301 * amounts of kthreads, has a single scheduler instance while scheduler 302 * entities represent rings. 303 * 304 * As a workaround, set the DRM_SCHED_FENCE_DONT_PIPELINE for all 305 * out-fences in order to force the scheduler to wait for full job 306 * completion for dependent jobs from different entities and same 307 * scheduler instance. 308 * 309 * There is some work in progress [1] to address the issues of firmware 310 * schedulers; once it is in-tree the scheduler topology in Nouveau 311 * should be re-worked accordingly. 312 * 313 * [1] https://lore.kernel.org/dri-devel/20230801205103.627779-1-matthew.brost@intel.com/ 314 */ 315 set_bit(DRM_SCHED_FENCE_DONT_PIPELINE, &job->done_fence->flags); 316 317 if (job->ops->armed_submit) 318 job->ops->armed_submit(job); 319 320 nouveau_job_fence_attach(job); 321 322 /* Set job state before pushing the job to the scheduler, 323 * such that we do not overwrite the job state set in run(). 324 */ 325 job->state = NOUVEAU_JOB_SUBMIT_SUCCESS; 326 327 drm_sched_entity_push_job(&job->base); 328 329 mutex_unlock(&entity->mutex); 330 331 if (done_fence) { 332 dma_fence_wait(done_fence, true); 333 dma_fence_put(done_fence); 334 } 335 336 return 0; 337 338 err_cleanup: 339 mutex_unlock(&entity->mutex); 340 nouveau_job_fence_attach_cleanup(job); 341 err: 342 job->state = NOUVEAU_JOB_SUBMIT_FAILED; 343 return ret; 344 } 345 346 bool 347 nouveau_sched_entity_qwork(struct nouveau_sched_entity *entity, 348 struct work_struct *work) 349 { 350 return queue_work(entity->sched_wq, work); 351 } 352 353 static struct dma_fence * 354 nouveau_job_run(struct nouveau_job *job) 355 { 356 struct dma_fence *fence; 357 358 fence = job->ops->run(job); 359 if (IS_ERR(fence)) 360 job->state = NOUVEAU_JOB_RUN_FAILED; 361 else 362 job->state = NOUVEAU_JOB_RUN_SUCCESS; 363 364 return fence; 365 } 366 367 static struct dma_fence * 368 nouveau_sched_run_job(struct drm_sched_job *sched_job) 369 { 370 struct nouveau_job *job = to_nouveau_job(sched_job); 371 372 return nouveau_job_run(job); 373 } 374 375 static enum drm_gpu_sched_stat 376 nouveau_sched_timedout_job(struct drm_sched_job *sched_job) 377 { 378 struct nouveau_job *job = to_nouveau_job(sched_job); 379 380 NV_PRINTK(warn, job->cli, "Job timed out.\n"); 381 382 if (job->ops->timeout) 383 return job->ops->timeout(job); 384 385 return DRM_GPU_SCHED_STAT_ENODEV; 386 } 387 388 static void 389 nouveau_sched_free_job(struct drm_sched_job *sched_job) 390 { 391 struct nouveau_job *job = to_nouveau_job(sched_job); 392 393 nouveau_job_fini(job); 394 } 395 396 int nouveau_sched_entity_init(struct nouveau_sched_entity *entity, 397 struct drm_gpu_scheduler *sched, 398 struct workqueue_struct *sched_wq) 399 { 400 mutex_init(&entity->mutex); 401 spin_lock_init(&entity->job.list.lock); 402 INIT_LIST_HEAD(&entity->job.list.head); 403 init_waitqueue_head(&entity->job.wq); 404 405 entity->sched_wq = sched_wq; 406 return drm_sched_entity_init(&entity->base, 407 DRM_SCHED_PRIORITY_NORMAL, 408 &sched, 1, NULL); 409 } 410 411 void 412 nouveau_sched_entity_fini(struct nouveau_sched_entity *entity) 413 { 414 drm_sched_entity_destroy(&entity->base); 415 } 416 417 static const struct drm_sched_backend_ops nouveau_sched_ops = { 418 .run_job = nouveau_sched_run_job, 419 .timedout_job = nouveau_sched_timedout_job, 420 .free_job = nouveau_sched_free_job, 421 }; 422 423 int nouveau_sched_init(struct nouveau_drm *drm) 424 { 425 struct drm_gpu_scheduler *sched = &drm->sched; 426 long job_hang_limit = msecs_to_jiffies(NOUVEAU_SCHED_JOB_TIMEOUT_MS); 427 428 drm->sched_wq = create_singlethread_workqueue("nouveau_sched_wq"); 429 if (!drm->sched_wq) 430 return -ENOMEM; 431 432 return drm_sched_init(sched, &nouveau_sched_ops, 433 NOUVEAU_SCHED_HW_SUBMISSIONS, 0, job_hang_limit, 434 NULL, NULL, "nouveau_sched", drm->dev->dev); 435 } 436 437 void nouveau_sched_fini(struct nouveau_drm *drm) 438 { 439 destroy_workqueue(drm->sched_wq); 440 drm_sched_fini(&drm->sched); 441 } 442