gpu_scheduler.c 14.3 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109
/*
 * Copyright 2015 Advanced Micro Devices, Inc.
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
 * and/or sell copies of the Software, and to permit persons to whom the
 * Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice shall be included in
 * all copies or substantial portions of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
 * OTHER DEALINGS IN THE SOFTWARE.
 *
 *
 */
#include <linux/kthread.h>
#include <linux/wait.h>
#include <linux/sched.h>
#include <drm/drmP.h>
#include "gpu_scheduler.h"

/* Initialize a given run queue struct */
static void init_rq(struct amd_run_queue *rq)
{
	INIT_LIST_HEAD(&rq->head.list);
	rq->head.belongto_rq = rq;
	mutex_init(&rq->lock);
	atomic_set(&rq->nr_entity, 0);
	rq->current_entity = &rq->head;
}

/* Note: caller must hold the lock or in a atomic context */
static void rq_remove_entity(struct amd_run_queue *rq,
			     struct amd_sched_entity *entity)
{
	if (rq->current_entity == entity)
		rq->current_entity = list_entry(entity->list.prev,
						typeof(*entity), list);
	list_del_init(&entity->list);
	atomic_dec(&rq->nr_entity);
}

static void rq_add_entity(struct amd_run_queue *rq,
			  struct amd_sched_entity *entity)
{
	list_add_tail(&entity->list, &rq->head.list);
	atomic_inc(&rq->nr_entity);
}

/**
 * Select next entity from a specified run queue with round robin policy.
 * It could return the same entity as current one if current is the only
 * available one in the queue. Return NULL if nothing available.
 */
static struct amd_sched_entity *rq_select_entity(struct amd_run_queue *rq)
{
	struct amd_sched_entity *p = rq->current_entity;
	int i = atomic_read(&rq->nr_entity) + 1; /*real count + dummy head*/
	while (i) {
		p = list_entry(p->list.next, typeof(*p), list);
		if (!rq->check_entity_status(p)) {
			rq->current_entity = p;
			break;
		}
		i--;
	}
	return i ? p : NULL;
}

static bool context_entity_is_waiting(struct amd_context_entity *entity)
{
	/* TODO: sync obj for multi-ring synchronization */
	return false;
}

static int gpu_entity_check_status(struct amd_sched_entity *entity)
{
	struct amd_context_entity *tmp = NULL;

	if (entity == &entity->belongto_rq->head)
		return -1;

	tmp = container_of(entity, typeof(*tmp), generic_entity);
	if (kfifo_is_empty(&tmp->job_queue) ||
	    context_entity_is_waiting(tmp))
		return -1;

	return 0;
}

/**
 * Note: This function should only been called inside scheduler main
 * function for thread safety, there is no other protection here.
 * return ture if scheduler has something ready to run.
 *
 * For active_hw_rq, there is only one producer(scheduler thread) and
 * one consumer(ISR). It should be safe to use this function in scheduler
 * main thread to decide whether to continue emit more IBs.
*/
static bool is_scheduler_ready(struct amd_gpu_scheduler *sched)
{
110 111 112 113 114 115 116 117
	unsigned long flags;
	bool full;
	spin_lock_irqsave(&sched->queue_lock, flags);
	full = atomic64_read(&sched->hw_rq_count) <
		sched->hw_submission_limit ? true : false;
	spin_unlock_irqrestore(&sched->queue_lock, flags);

	return full;
118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181
}

/**
 * Select next entity from the kernel run queue, if not available,
 * return null.
*/
static struct amd_context_entity *kernel_rq_select_context(
	struct amd_gpu_scheduler *sched)
{
	struct amd_sched_entity *sched_entity = NULL;
	struct amd_context_entity *tmp = NULL;
	struct amd_run_queue *rq = &sched->kernel_rq;

	mutex_lock(&rq->lock);
	sched_entity = rq_select_entity(rq);
	if (sched_entity)
		tmp = container_of(sched_entity,
				   typeof(*tmp),
				   generic_entity);
	mutex_unlock(&rq->lock);
	return tmp;
}

/**
 * Select next entity containing real IB submissions
*/
static struct amd_context_entity *select_context(
	struct amd_gpu_scheduler *sched)
{
	struct amd_context_entity *wake_entity = NULL;
	struct amd_context_entity *tmp;
	struct amd_run_queue *rq;

	if (!is_scheduler_ready(sched))
		return NULL;

	/* Kernel run queue has higher priority than normal run queue*/
	tmp = kernel_rq_select_context(sched);
	if (tmp != NULL)
		goto exit;

	WARN_ON(offsetof(struct amd_context_entity, generic_entity) != 0);

	rq = &sched->sched_rq;
	mutex_lock(&rq->lock);
	tmp = container_of(rq_select_entity(rq),
			   typeof(*tmp), generic_entity);
	mutex_unlock(&rq->lock);
exit:
	if (sched->current_entity && (sched->current_entity != tmp))
		wake_entity = sched->current_entity;
	sched->current_entity = tmp;
	if (wake_entity)
		wake_up(&wake_entity->wait_queue);
	return tmp;
}

/**
 * Init a context entity used by scheduler when submit to HW ring.
 *
 * @sched	The pointer to the scheduler
 * @entity	The pointer to a valid amd_context_entity
 * @parent	The parent entity of this amd_context_entity
 * @rq		The run queue this entity belongs
182
 * @kernel	If this is an entity for the kernel
183
 * @jobs	The max number of jobs in the job queue
184 185 186 187 188 189 190
 *
 * return 0 if succeed. negative error code on failure
*/
int amd_context_entity_init(struct amd_gpu_scheduler *sched,
			    struct amd_context_entity *entity,
			    struct amd_sched_entity *parent,
			    struct amd_run_queue *rq,
191
			    uint32_t jobs)
192 193 194 195 196 197 198 199 200 201 202 203 204 205 206
{
	uint64_t seq_ring = 0;

	if (!(sched && entity && rq))
		return -EINVAL;

	memset(entity, 0, sizeof(struct amd_context_entity));
	seq_ring = ((uint64_t)sched->ring_id) << 60;
	spin_lock_init(&entity->lock);
	entity->generic_entity.belongto_rq = rq;
	entity->generic_entity.parent = parent;
	entity->scheduler = sched;
	init_waitqueue_head(&entity->wait_queue);
	init_waitqueue_head(&entity->wait_emit);
	if(kfifo_alloc(&entity->job_queue,
207
		       jobs * sizeof(void *),
208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280
		       GFP_KERNEL))
		return -EINVAL;

	spin_lock_init(&entity->queue_lock);
	atomic64_set(&entity->last_emitted_v_seq, seq_ring);
	atomic64_set(&entity->last_queued_v_seq, seq_ring);

	/* Add the entity to the run queue */
	mutex_lock(&rq->lock);
	rq_add_entity(rq, &entity->generic_entity);
	mutex_unlock(&rq->lock);
	return 0;
}

/**
 * Query if entity is initialized
 *
 * @sched       Pointer to scheduler instance
 * @entity	The pointer to a valid scheduler entity
 *
 * return true if entity is initialized, false otherwise
*/
static bool is_context_entity_initialized(struct amd_gpu_scheduler *sched,
					  struct amd_context_entity *entity)
{
	return entity->scheduler == sched &&
		entity->generic_entity.belongto_rq != NULL;
}

static bool is_context_entity_idle(struct amd_gpu_scheduler *sched,
				   struct amd_context_entity *entity)
{
	/**
	 * Idle means no pending IBs, and the entity is not
	 * currently being used.
	*/
	barrier();
	if ((sched->current_entity != entity) &&
	    kfifo_is_empty(&entity->job_queue))
		return true;

	return false;
}

/**
 * Destroy a context entity
 *
 * @sched       Pointer to scheduler instance
 * @entity	The pointer to a valid scheduler entity
 *
 * return 0 if succeed. negative error code on failure
 */
int amd_context_entity_fini(struct amd_gpu_scheduler *sched,
			    struct amd_context_entity *entity)
{
	int r = 0;
	struct amd_run_queue *rq = entity->generic_entity.belongto_rq;

	if (!is_context_entity_initialized(sched, entity))
		return 0;

	/**
	 * The client will not queue more IBs during this fini, consume existing
	 * queued IBs
	*/
	r = wait_event_timeout(
		entity->wait_queue,
		is_context_entity_idle(sched, entity),
		msecs_to_jiffies(AMD_GPU_WAIT_IDLE_TIMEOUT_IN_MS)
		) ?  0 : -1;

	if (r) {
		if (entity->is_pending)
281
			DRM_INFO("Entity %p is in waiting state during fini,\
282
				all pending ibs will be canceled.\n",
283
				 entity);
284 285 286 287 288 289 290 291 292 293 294 295 296 297 298
	}

	mutex_lock(&rq->lock);
	rq_remove_entity(rq, &entity->generic_entity);
	mutex_unlock(&rq->lock);
	kfifo_free(&entity->job_queue);
	return r;
}

/**
 * Submit a normal job to the job queue
 *
 * @sched	The pointer to the scheduler
 * @c_entity    The pointer to amd_context_entity
 * @job		The pointer to job required to submit
299 300 301 302
 * return 0 if succeed. -1 if failed.
 *        -2 indicate queue is full for this client, client should wait untill
 *	     scheduler consum some queued command.
 *	  -1 other fail.
303
*/
304
int amd_sched_push_job(struct amd_gpu_scheduler *sched,
305 306 307 308 309 310 311 312 313 314 315 316 317
		       struct amd_context_entity *c_entity,
		       void *job)
{
	while (kfifo_in_spinlocked(&c_entity->job_queue, &job, sizeof(void *),
				   &c_entity->queue_lock) != sizeof(void *)) {
		/**
		 * Current context used up all its IB slots
		 * wait here, or need to check whether GPU is hung
		*/
		schedule();
	}

	wake_up_interruptible(&sched->wait_queue);
318
	return 0;
319 320 321
}

/**
322
 * Wait for a virtual sequence number to be emitted.
323 324 325 326 327 328 329 330 331
 *
 * @c_entity	The pointer to a valid context entity
 * @seq         The virtual sequence number to wait
 * @intr	Interruptible or not
 * @timeout	Timeout in ms, wait infinitely if <0
 * @emit        wait for emit or signal
 *
 * return =0 signaled ,  <0 failed
*/
332 333 334 335
int amd_sched_wait_emit(struct amd_context_entity *c_entity,
			uint64_t seq,
			bool intr,
			long timeout)
336
{
337 338
	atomic64_t *v_seq = &c_entity->last_emitted_v_seq;
	wait_queue_head_t *wait_queue = &c_entity->wait_emit;
339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378

	if (intr && (timeout < 0)) {
		wait_event_interruptible(
			*wait_queue,
			seq <= atomic64_read(v_seq));
		return 0;
	} else if (intr && (timeout >= 0)) {
		wait_event_interruptible_timeout(
			*wait_queue,
			seq <= atomic64_read(v_seq),
			msecs_to_jiffies(timeout));
		return (seq <= atomic64_read(v_seq)) ?
			0 : -1;
	} else if (!intr && (timeout < 0)) {
		wait_event(
			*wait_queue,
			seq <= atomic64_read(v_seq));
		return 0;
	} else if (!intr && (timeout >= 0)) {
		wait_event_timeout(
			*wait_queue,
			seq <= atomic64_read(v_seq),
			msecs_to_jiffies(timeout));
		return (seq <= atomic64_read(v_seq)) ?
			0 : -1;
	}
	return 0;
}

static int amd_sched_main(void *param)
{
	int r;
	void *job;
	struct sched_param sparam = {.sched_priority = 1};
	struct amd_context_entity *c_entity = NULL;
	struct amd_gpu_scheduler *sched = (struct amd_gpu_scheduler *)param;

	sched_setscheduler(current, SCHED_FIFO, &sparam);

	while (!kthread_should_stop()) {
379
		struct amd_sched_job *sched_job = NULL;
380 381 382 383 384 385 386
		wait_event_interruptible(sched->wait_queue,
					 is_scheduler_ready(sched) &&
					 (c_entity = select_context(sched)));
		r = kfifo_out(&c_entity->job_queue, &job, sizeof(void *));
		if (r != sizeof(void *))
			continue;
		r = sched->ops->prepare_job(sched, c_entity, job);
387 388 389 390 391 392 393 394 395 396 397 398 399 400 401
		if (!r) {
			unsigned long flags;
			sched_job = kzalloc(sizeof(struct amd_sched_job),
					    GFP_KERNEL);
			if (!sched_job) {
				WARN(true, "No memory to allocate\n");
				continue;
			}
			sched_job->job = job;
			sched_job->sched = sched;
			spin_lock_irqsave(&sched->queue_lock, flags);
			list_add_tail(&sched_job->list, &sched->active_hw_rq);
			atomic64_inc(&sched->hw_rq_count);
			spin_unlock_irqrestore(&sched->queue_lock, flags);
		}
402
		mutex_lock(&sched->sched_lock);
403
		sched->ops->run_job(sched, c_entity, sched_job);
404 405 406 407 408 409 410 411 412 413 414
		mutex_unlock(&sched->sched_lock);
	}
	return 0;
}

/**
 * ISR to handle EOP inetrrupts
 *
 * @sched: gpu scheduler
 *
*/
415
void amd_sched_process_job(struct amd_sched_job *sched_job)
416
{
417 418 419 420 421 422 423 424 425 426 427 428
	unsigned long flags;
	struct amd_gpu_scheduler *sched;
	if (!sched_job)
		return;
	sched = sched_job->sched;
	spin_lock_irqsave(&sched->queue_lock, flags);
	list_del(&sched_job->list);
	atomic64_dec(&sched->hw_rq_count);
	spin_unlock_irqrestore(&sched->queue_lock, flags);

	sched->ops->process_job(sched, sched_job->job);
	kfree(sched_job);
429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446
	wake_up_interruptible(&sched->wait_queue);
}

/**
 * Create a gpu scheduler
 *
 * @device	The device context for this scheduler
 * @ops		The backend operations for this scheduler.
 * @id	        The scheduler is per ring, here is ring id.
 * @granularity	The minumum ms unit the scheduler will scheduled.
 * @preemption  Indicate whether this ring support preemption, 0 is no.
 *
 * return the pointer to scheduler for success, otherwise return NULL
*/
struct amd_gpu_scheduler *amd_sched_create(void *device,
					   struct amd_sched_backend_ops *ops,
					   unsigned ring,
					   unsigned granularity,
447 448
					   unsigned preemption,
					   unsigned hw_submission)
449 450 451 452 453 454 455 456 457 458 459 460 461
{
	struct amd_gpu_scheduler *sched;
	char name[20] = "gpu_sched[0]";

	sched = kzalloc(sizeof(struct amd_gpu_scheduler), GFP_KERNEL);
	if (!sched)
		return NULL;

	sched->device = device;
	sched->ops = ops;
	sched->granularity = granularity;
	sched->ring_id = ring;
	sched->preemption = preemption;
462
	sched->hw_submission_limit = hw_submission;
463 464 465 466 467 468 469 470 471 472
	snprintf(name, sizeof(name), "gpu_sched[%d]", ring);
	mutex_init(&sched->sched_lock);
	spin_lock_init(&sched->queue_lock);
	init_rq(&sched->sched_rq);
	sched->sched_rq.check_entity_status = gpu_entity_check_status;

	init_rq(&sched->kernel_rq);
	sched->kernel_rq.check_entity_status = gpu_entity_check_status;

	init_waitqueue_head(&sched->wait_queue);
473 474
	INIT_LIST_HEAD(&sched->active_hw_rq);
	atomic64_set(&sched->hw_rq_count, 0);
475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500
	/* Each scheduler will run on a seperate kernel thread */
	sched->thread = kthread_create(amd_sched_main, sched, name);
	if (sched->thread) {
		wake_up_process(sched->thread);
		return sched;
	}

	DRM_ERROR("Failed to create scheduler for id %d.\n", ring);
	kfree(sched);
	return NULL;
}

/**
 * Destroy a gpu scheduler
 *
 * @sched	The pointer to the scheduler
 *
 * return 0 if succeed. -1 if failed.
 */
int amd_sched_destroy(struct amd_gpu_scheduler *sched)
{
	kthread_stop(sched->thread);
	kfree(sched);
	return  0;
}

501 502 503 504 505 506 507 508 509 510 511 512
/**
 * Update emitted sequence and wake up the waiters, called by run_job
 * in driver side
 *
 * @entity The context entity
 * @seq The sequence number for the latest emitted job
*/
void amd_sched_emit(struct amd_context_entity *c_entity, uint64_t seq)
{
	atomic64_set(&c_entity->last_emitted_v_seq, seq);
	wake_up_all(&c_entity->wait_emit);
}
513 514 515 516 517 518 519 520 521 522 523 524

/**
 * Get next queued sequence number
 *
 * @entity The context entity
 *
 * return the next queued sequence number
*/
uint64_t amd_sched_next_queued_seq(struct amd_context_entity *c_entity)
{
	return atomic64_read(&c_entity->last_queued_v_seq) + 1;
}