drivers/gpu/drm/amd/scheduler/gpu_scheduler.c

   1 /*
   2  * Copyright 2015 Advanced Micro Devices, Inc.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE SOFTWARE.
  21  *
  22  *
  23  */
  24 #include <linux/kthread.h>
  25 #include <linux/wait.h>
  26 #include <linux/sched.h>
  27 #include <drm/drmP.h>
  28 #include "gpu_scheduler.h"
  29
  30 /* Initialize a given run queue struct */
  31 static void amd_sched_rq_init(struct amd_sched_rq *rq)
  32 {
  33         spin_lock_init(&rq->lock);
  34         INIT_LIST_HEAD(&rq->entities);
  35         rq->current_entity = NULL;
  36 }
  37
  38 static void amd_sched_rq_add_entity(struct amd_sched_rq *rq,
  39                                     struct amd_sched_entity *entity)
  40 {
  41         spin_lock(&rq->lock);
  42         list_add_tail(&entity->list, &rq->entities);
  43         spin_unlock(&rq->lock);
  44 }
  45
  46 static void amd_sched_rq_remove_entity(struct amd_sched_rq *rq,
  47                                        struct amd_sched_entity *entity)
  48 {
  49         spin_lock(&rq->lock);
  50         list_del_init(&entity->list);
  51         if (rq->current_entity == entity)
  52                 rq->current_entity = NULL;
  53         spin_unlock(&rq->lock);
  54 }
  55
  56 /**
  57  * Select next entity from a specified run queue with round robin policy.
  58  * It could return the same entity as current one if current is the only
  59  * available one in the queue. Return NULL if nothing available.
  60  */
  61 static struct amd_sched_entity *
  62 amd_sched_rq_select_entity(struct amd_sched_rq *rq)
  63 {
  64         struct amd_sched_entity *entity;
  65
  66         spin_lock(&rq->lock);
  67
  68         entity = rq->current_entity;
  69         if (entity) {
  70                 list_for_each_entry_continue(entity, &rq->entities, list) {
  71                         if (!kfifo_is_empty(&entity->job_queue)) {
  72                                 rq->current_entity = entity;
  73                                 spin_unlock(&rq->lock);
  74                                 return rq->current_entity;
  75                         }
  76                 }
  77         }
  78
  79         list_for_each_entry(entity, &rq->entities, list) {
  80
  81                 if (!kfifo_is_empty(&entity->job_queue)) {
  82                         rq->current_entity = entity;
  83                         spin_unlock(&rq->lock);
  84                         return rq->current_entity;
  85                 }
  86
  87                 if (entity == rq->current_entity)
  88                         break;
  89         }
  90
  91         spin_unlock(&rq->lock);
  92
  93         return NULL;
  94 }
  95
  96 /**
  97  * Return ture if we can push more jobs to the hw.
  98  */
  99 static bool amd_sched_ready(struct amd_gpu_scheduler *sched)
 100 {
 101         return atomic_read(&sched->hw_rq_count) <
 102                 sched->hw_submission_limit;
 103 }
 104
 105 /**
 106  * Select next entity containing real IB submissions
 107 */
 108 static struct amd_sched_entity *
 109 select_context(struct amd_gpu_scheduler *sched)
 110 {
 111         struct amd_sched_entity *wake_entity = NULL;
 112         struct amd_sched_entity *tmp;
 113
 114         if (!amd_sched_ready(sched))
 115                 return NULL;
 116
 117         /* Kernel run queue has higher priority than normal run queue*/
 118         tmp = amd_sched_rq_select_entity(&sched->kernel_rq);
 119         if (tmp == NULL)
 120                 tmp = amd_sched_rq_select_entity(&sched->sched_rq);
 121
 122         if (sched->current_entity && (sched->current_entity != tmp))
 123                 wake_entity = sched->current_entity;
 124         sched->current_entity = tmp;
 125         if (wake_entity && wake_entity->need_wakeup)
 126                 wake_up(&wake_entity->wait_queue);
 127         return tmp;
 128 }
 129
 130 /**
 131  * Init a context entity used by scheduler when submit to HW ring.
 132  *
 133  * @sched       The pointer to the scheduler
 134  * @entity      The pointer to a valid amd_sched_entity
 135  * @rq          The run queue this entity belongs
 136  * @kernel      If this is an entity for the kernel
 137  * @jobs        The max number of jobs in the job queue
 138  *
 139  * return 0 if succeed. negative error code on failure
 140 */
 141 int amd_sched_entity_init(struct amd_gpu_scheduler *sched,
 142                           struct amd_sched_entity *entity,
 143                           struct amd_sched_rq *rq,
 144                           uint32_t jobs)
 145 {
 146         char name[20];
 147
 148         if (!(sched && entity && rq))
 149                 return -EINVAL;
 150
 151         memset(entity, 0, sizeof(struct amd_sched_entity));
 152         spin_lock_init(&entity->lock);
 153         entity->belongto_rq = rq;
 154         entity->scheduler = sched;
 155         init_waitqueue_head(&entity->wait_queue);
 156         init_waitqueue_head(&entity->wait_emit);
 157         entity->fence_context = fence_context_alloc(1);
 158         snprintf(name, sizeof(name), "c_entity[%llu]", entity->fence_context);
 159         memcpy(entity->name, name, 20);
 160         entity->need_wakeup = false;
 161         if(kfifo_alloc(&entity->job_queue,
 162                        jobs * sizeof(void *),
 163                        GFP_KERNEL))
 164                 return -EINVAL;
 165
 166         spin_lock_init(&entity->queue_lock);
 167         atomic_set(&entity->fence_seq, 0);
 168
 169         /* Add the entity to the run queue */
 170         amd_sched_rq_add_entity(rq, entity);
 171         return 0;
 172 }
 173
 174 /**
 175  * Query if entity is initialized
 176  *
 177  * @sched       Pointer to scheduler instance
 178  * @entity      The pointer to a valid scheduler entity
 179  *
 180  * return true if entity is initialized, false otherwise
 181 */
 182 static bool is_context_entity_initialized(struct amd_gpu_scheduler *sched,
 183                                           struct amd_sched_entity *entity)
 184 {
 185         return entity->scheduler == sched &&
 186                 entity->belongto_rq != NULL;
 187 }
 188
 189 static bool is_context_entity_idle(struct amd_gpu_scheduler *sched,
 190                                    struct amd_sched_entity *entity)
 191 {
 192         /**
 193          * Idle means no pending IBs, and the entity is not
 194          * currently being used.
 195         */
 196         barrier();
 197         if ((sched->current_entity != entity) &&
 198             kfifo_is_empty(&entity->job_queue))
 199                 return true;
 200
 201         return false;
 202 }
 203
 204 /**
 205  * Destroy a context entity
 206  *
 207  * @sched       Pointer to scheduler instance
 208  * @entity      The pointer to a valid scheduler entity
 209  *
 210  * return 0 if succeed. negative error code on failure
 211  */
 212 int amd_sched_entity_fini(struct amd_gpu_scheduler *sched,
 213                             struct amd_sched_entity *entity)
 214 {
 215         int r = 0;
 216         struct amd_sched_rq *rq = entity->belongto_rq;
 217
 218         if (!is_context_entity_initialized(sched, entity))
 219                 return 0;
 220         entity->need_wakeup = true;
 221         /**
 222          * The client will not queue more IBs during this fini, consume existing
 223          * queued IBs
 224         */
 225         r = wait_event_timeout(
 226                 entity->wait_queue,
 227                 is_context_entity_idle(sched, entity),
 228                 msecs_to_jiffies(AMD_GPU_WAIT_IDLE_TIMEOUT_IN_MS)
 229                 ) ?  0 : -1;
 230
 231         if (r) {
 232                 if (entity->is_pending)
 233                         DRM_INFO("Entity %p is in waiting state during fini,\
 234                                 all pending ibs will be canceled.\n",
 235                                  entity);
 236         }
 237
 238         amd_sched_rq_remove_entity(rq, entity);
 239         kfifo_free(&entity->job_queue);
 240         return r;
 241 }
 242
 243 /**
 244  * Submit a normal job to the job queue
 245  *
 246  * @sched       The pointer to the scheduler
 247  * @c_entity    The pointer to amd_sched_entity
 248  * @job         The pointer to job required to submit
 249  * return 0 if succeed. -1 if failed.
 250  *        -2 indicate queue is full for this client, client should wait untill
 251  *           scheduler consum some queued command.
 252  *        -1 other fail.
 253 */
 254 int amd_sched_push_job(struct amd_sched_job *sched_job)
 255 {
 256         struct amd_sched_fence  *fence =
 257                 amd_sched_fence_create(sched_job->s_entity);
 258         if (!fence)
 259                 return -EINVAL;
 260         fence_get(&fence->base);
 261         sched_job->s_fence = fence;
 262         while (kfifo_in_spinlocked(&sched_job->s_entity->job_queue,
 263                                    &sched_job, sizeof(void *),
 264                                    &sched_job->s_entity->queue_lock) !=
 265                sizeof(void *)) {
 266                 /**
 267                  * Current context used up all its IB slots
 268                  * wait here, or need to check whether GPU is hung
 269                 */
 270                 schedule();
 271         }
 272         /* first job wake up scheduler */
 273         if ((kfifo_len(&sched_job->s_entity->job_queue) / sizeof(void *)) == 1)
 274                 wake_up_interruptible(&sched_job->sched->wait_queue);
 275         return 0;
 276 }
 277
 278 static void amd_sched_process_job(struct fence *f, struct fence_cb *cb)
 279 {
 280         struct amd_sched_job *sched_job =
 281                 container_of(cb, struct amd_sched_job, cb);
 282         struct amd_gpu_scheduler *sched;
 283
 284         sched = sched_job->sched;
 285         amd_sched_fence_signal(sched_job->s_fence);
 286         atomic_dec(&sched->hw_rq_count);
 287         fence_put(&sched_job->s_fence->base);
 288         sched->ops->process_job(sched, sched_job);
 289         wake_up_interruptible(&sched->wait_queue);
 290 }
 291
 292 static int amd_sched_main(void *param)
 293 {
 294         int r;
 295         struct amd_sched_job *job;
 296         struct sched_param sparam = {.sched_priority = 1};
 297         struct amd_sched_entity *c_entity = NULL;
 298         struct amd_gpu_scheduler *sched = (struct amd_gpu_scheduler *)param;
 299
 300         sched_setscheduler(current, SCHED_FIFO, &sparam);
 301
 302         while (!kthread_should_stop()) {
 303                 struct fence *fence;
 304
 305                 wait_event_interruptible(sched->wait_queue,
 306                                          amd_sched_ready(sched) &&
 307                                          (c_entity = select_context(sched)));
 308                 r = kfifo_out(&c_entity->job_queue, &job, sizeof(void *));
 309                 if (r != sizeof(void *))
 310                         continue;
 311                 r = 0;
 312                 if (sched->ops->prepare_job)
 313                         r = sched->ops->prepare_job(sched, c_entity, job);
 314                 if (!r) {
 315                         atomic_inc(&sched->hw_rq_count);
 316                 }
 317                 mutex_lock(&sched->sched_lock);
 318                 fence = sched->ops->run_job(sched, c_entity, job);
 319                 if (fence) {
 320                         r = fence_add_callback(fence, &job->cb,
 321                                                amd_sched_process_job);
 322                         if (r == -ENOENT)
 323                                 amd_sched_process_job(fence, &job->cb);
 324                         else if (r)
 325                                 DRM_ERROR("fence add callback failed (%d)\n", r);
 326                         fence_put(fence);
 327                 }
 328                 mutex_unlock(&sched->sched_lock);
 329         }
 330         return 0;
 331 }
 332
 333 /**
 334  * Create a gpu scheduler
 335  *
 336  * @device      The device context for this scheduler
 337  * @ops         The backend operations for this scheduler.
 338  * @id          The scheduler is per ring, here is ring id.
 339  * @granularity The minumum ms unit the scheduler will scheduled.
 340  * @preemption  Indicate whether this ring support preemption, 0 is no.
 341  *
 342  * return the pointer to scheduler for success, otherwise return NULL
 343 */
 344 struct amd_gpu_scheduler *amd_sched_create(void *device,
 345                                            struct amd_sched_backend_ops *ops,
 346                                            unsigned ring,
 347                                            unsigned granularity,
 348                                            unsigned preemption,
 349                                            unsigned hw_submission)
 350 {
 351         struct amd_gpu_scheduler *sched;
 352         char name[20];
 353
 354         sched = kzalloc(sizeof(struct amd_gpu_scheduler), GFP_KERNEL);
 355         if (!sched)
 356                 return NULL;
 357
 358         sched->device = device;
 359         sched->ops = ops;
 360         sched->granularity = granularity;
 361         sched->ring_id = ring;
 362         sched->preemption = preemption;
 363         sched->hw_submission_limit = hw_submission;
 364         snprintf(name, sizeof(name), "gpu_sched[%d]", ring);
 365         mutex_init(&sched->sched_lock);
 366         amd_sched_rq_init(&sched->sched_rq);
 367         amd_sched_rq_init(&sched->kernel_rq);
 368
 369         init_waitqueue_head(&sched->wait_queue);
 370         atomic_set(&sched->hw_rq_count, 0);
 371         /* Each scheduler will run on a seperate kernel thread */
 372         sched->thread = kthread_create(amd_sched_main, sched, name);
 373         if (sched->thread) {
 374                 wake_up_process(sched->thread);
 375                 return sched;
 376         }
 377
 378         DRM_ERROR("Failed to create scheduler for id %d.\n", ring);
 379         kfree(sched);
 380         return NULL;
 381 }
 382
 383 /**
 384  * Destroy a gpu scheduler
 385  *
 386  * @sched       The pointer to the scheduler
 387  *
 388  * return 0 if succeed. -1 if failed.
 389  */
 390 int amd_sched_destroy(struct amd_gpu_scheduler *sched)
 391 {
 392         kthread_stop(sched->thread);
 393         kfree(sched);
 394         return  0;
 395 }