1 /*
2  * Copyright 2014 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/slab.h>
25 #include <linux/list.h>
26 #include <linux/types.h>
27 #include <linux/printk.h>
28 #include <linux/bitops.h>
29 #include <linux/sched.h>
30 #include "kfd_priv.h"
31 #include "kfd_device_queue_manager.h"
32 #include "kfd_mqd_manager.h"
33 #include "cik_regs.h"
34 #include "kfd_kernel_queue.h"
35 
36 /* Size of the per-pipe EOP queue */
37 #define CIK_HPD_EOP_BYTES_LOG2 11
38 #define CIK_HPD_EOP_BYTES (1U << CIK_HPD_EOP_BYTES_LOG2)
39 
40 static int set_pasid_vmid_mapping(struct device_queue_manager *dqm,
41 					unsigned int pasid, unsigned int vmid);
42 
43 static int create_compute_queue_nocpsch(struct device_queue_manager *dqm,
44 					struct queue *q,
45 					struct qcm_process_device *qpd);
46 
47 static int execute_queues_cpsch(struct device_queue_manager *dqm, bool lock);
48 static int destroy_queues_cpsch(struct device_queue_manager *dqm, bool lock);
49 
50 static int create_sdma_queue_nocpsch(struct device_queue_manager *dqm,
51 					struct queue *q,
52 					struct qcm_process_device *qpd);
53 
54 static void deallocate_sdma_queue(struct device_queue_manager *dqm,
55 				unsigned int sdma_queue_id);
56 
57 static inline
58 enum KFD_MQD_TYPE get_mqd_type_from_queue_type(enum kfd_queue_type type)
59 {
60 	if (type == KFD_QUEUE_TYPE_SDMA)
61 		return KFD_MQD_TYPE_SDMA;
62 	return KFD_MQD_TYPE_CP;
63 }
64 
65 unsigned int get_first_pipe(struct device_queue_manager *dqm)
66 {
67 	BUG_ON(!dqm || !dqm->dev);
68 	return dqm->dev->shared_resources.first_compute_pipe;
69 }
70 
71 unsigned int get_pipes_num(struct device_queue_manager *dqm)
72 {
73 	BUG_ON(!dqm || !dqm->dev);
74 	return dqm->dev->shared_resources.compute_pipe_count;
75 }
76 
77 static inline unsigned int get_pipes_num_cpsch(void)
78 {
79 	return PIPE_PER_ME_CP_SCHEDULING;
80 }
81 
82 void program_sh_mem_settings(struct device_queue_manager *dqm,
83 					struct qcm_process_device *qpd)
84 {
85 	return dqm->dev->kfd2kgd->program_sh_mem_settings(
86 						dqm->dev->kgd, qpd->vmid,
87 						qpd->sh_mem_config,
88 						qpd->sh_mem_ape1_base,
89 						qpd->sh_mem_ape1_limit,
90 						qpd->sh_mem_bases);
91 }
92 
93 static int allocate_vmid(struct device_queue_manager *dqm,
94 			struct qcm_process_device *qpd,
95 			struct queue *q)
96 {
97 	int bit, allocated_vmid;
98 
99 	if (dqm->vmid_bitmap == 0)
100 		return -ENOMEM;
101 
102 	bit = find_first_bit((unsigned long *)&dqm->vmid_bitmap, CIK_VMID_NUM);
103 	clear_bit(bit, (unsigned long *)&dqm->vmid_bitmap);
104 
105 	/* Kaveri kfd vmid's starts from vmid 8 */
106 	allocated_vmid = bit + KFD_VMID_START_OFFSET;
107 	pr_debug("kfd: vmid allocation %d\n", allocated_vmid);
108 	qpd->vmid = allocated_vmid;
109 	q->properties.vmid = allocated_vmid;
110 
111 	set_pasid_vmid_mapping(dqm, q->process->pasid, q->properties.vmid);
112 	program_sh_mem_settings(dqm, qpd);
113 
114 	return 0;
115 }
116 
117 static void deallocate_vmid(struct device_queue_manager *dqm,
118 				struct qcm_process_device *qpd,
119 				struct queue *q)
120 {
121 	int bit = qpd->vmid - KFD_VMID_START_OFFSET;
122 
123 	/* Release the vmid mapping */
124 	set_pasid_vmid_mapping(dqm, 0, qpd->vmid);
125 
126 	set_bit(bit, (unsigned long *)&dqm->vmid_bitmap);
127 	qpd->vmid = 0;
128 	q->properties.vmid = 0;
129 }
130 
131 static int create_queue_nocpsch(struct device_queue_manager *dqm,
132 				struct queue *q,
133 				struct qcm_process_device *qpd,
134 				int *allocated_vmid)
135 {
136 	int retval;
137 
138 	BUG_ON(!dqm || !q || !qpd || !allocated_vmid);
139 
140 	pr_debug("kfd: In func %s\n", __func__);
141 	print_queue(q);
142 
143 	mutex_lock(&dqm->lock);
144 
145 	if (dqm->total_queue_count >= max_num_of_queues_per_device) {
146 		pr_warn("amdkfd: Can't create new usermode queue because %d queues were already created\n",
147 				dqm->total_queue_count);
148 		mutex_unlock(&dqm->lock);
149 		return -EPERM;
150 	}
151 
152 	if (list_empty(&qpd->queues_list)) {
153 		retval = allocate_vmid(dqm, qpd, q);
154 		if (retval != 0) {
155 			mutex_unlock(&dqm->lock);
156 			return retval;
157 		}
158 	}
159 	*allocated_vmid = qpd->vmid;
160 	q->properties.vmid = qpd->vmid;
161 
162 	if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE)
163 		retval = create_compute_queue_nocpsch(dqm, q, qpd);
164 	if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
165 		retval = create_sdma_queue_nocpsch(dqm, q, qpd);
166 
167 	if (retval != 0) {
168 		if (list_empty(&qpd->queues_list)) {
169 			deallocate_vmid(dqm, qpd, q);
170 			*allocated_vmid = 0;
171 		}
172 		mutex_unlock(&dqm->lock);
173 		return retval;
174 	}
175 
176 	list_add(&q->list, &qpd->queues_list);
177 	if (q->properties.is_active)
178 		dqm->queue_count++;
179 
180 	if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
181 		dqm->sdma_queue_count++;
182 
183 	/*
184 	 * Unconditionally increment this counter, regardless of the queue's
185 	 * type or whether the queue is active.
186 	 */
187 	dqm->total_queue_count++;
188 	pr_debug("Total of %d queues are accountable so far\n",
189 			dqm->total_queue_count);
190 
191 	mutex_unlock(&dqm->lock);
192 	return 0;
193 }
194 
195 static int allocate_hqd(struct device_queue_manager *dqm, struct queue *q)
196 {
197 	bool set;
198 	int pipe, bit, i;
199 
200 	set = false;
201 
202 	for (pipe = dqm->next_pipe_to_allocate, i = 0; i < get_pipes_num(dqm);
203 			pipe = ((pipe + 1) % get_pipes_num(dqm)), ++i) {
204 		if (dqm->allocated_queues[pipe] != 0) {
205 			bit = find_first_bit(
206 				(unsigned long *)&dqm->allocated_queues[pipe],
207 				QUEUES_PER_PIPE);
208 
209 			clear_bit(bit,
210 				(unsigned long *)&dqm->allocated_queues[pipe]);
211 			q->pipe = pipe;
212 			q->queue = bit;
213 			set = true;
214 			break;
215 		}
216 	}
217 
218 	if (set == false)
219 		return -EBUSY;
220 
221 	pr_debug("kfd: DQM %s hqd slot - pipe (%d) queue(%d)\n",
222 				__func__, q->pipe, q->queue);
223 	/* horizontal hqd allocation */
224 	dqm->next_pipe_to_allocate = (pipe + 1) % get_pipes_num(dqm);
225 
226 	return 0;
227 }
228 
229 static inline void deallocate_hqd(struct device_queue_manager *dqm,
230 				struct queue *q)
231 {
232 	set_bit(q->queue, (unsigned long *)&dqm->allocated_queues[q->pipe]);
233 }
234 
235 static int create_compute_queue_nocpsch(struct device_queue_manager *dqm,
236 					struct queue *q,
237 					struct qcm_process_device *qpd)
238 {
239 	int retval;
240 	struct mqd_manager *mqd;
241 
242 	BUG_ON(!dqm || !q || !qpd);
243 
244 	mqd = dqm->ops.get_mqd_manager(dqm, KFD_MQD_TYPE_COMPUTE);
245 	if (mqd == NULL)
246 		return -ENOMEM;
247 
248 	retval = allocate_hqd(dqm, q);
249 	if (retval != 0)
250 		return retval;
251 
252 	retval = mqd->init_mqd(mqd, &q->mqd, &q->mqd_mem_obj,
253 				&q->gart_mqd_addr, &q->properties);
254 	if (retval != 0) {
255 		deallocate_hqd(dqm, q);
256 		return retval;
257 	}
258 
259 	pr_debug("kfd: loading mqd to hqd on pipe (%d) queue (%d)\n",
260 			q->pipe,
261 			q->queue);
262 
263 	retval = mqd->load_mqd(mqd, q->mqd, q->pipe,
264 			q->queue, (uint32_t __user *) q->properties.write_ptr);
265 	if (retval != 0) {
266 		deallocate_hqd(dqm, q);
267 		mqd->uninit_mqd(mqd, q->mqd, q->mqd_mem_obj);
268 		return retval;
269 	}
270 
271 	return 0;
272 }
273 
274 static int destroy_queue_nocpsch(struct device_queue_manager *dqm,
275 				struct qcm_process_device *qpd,
276 				struct queue *q)
277 {
278 	int retval;
279 	struct mqd_manager *mqd;
280 
281 	BUG_ON(!dqm || !q || !q->mqd || !qpd);
282 
283 	retval = 0;
284 
285 	pr_debug("kfd: In Func %s\n", __func__);
286 
287 	mutex_lock(&dqm->lock);
288 
289 	if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE) {
290 		mqd = dqm->ops.get_mqd_manager(dqm, KFD_MQD_TYPE_COMPUTE);
291 		if (mqd == NULL) {
292 			retval = -ENOMEM;
293 			goto out;
294 		}
295 		deallocate_hqd(dqm, q);
296 	} else if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
297 		mqd = dqm->ops.get_mqd_manager(dqm, KFD_MQD_TYPE_SDMA);
298 		if (mqd == NULL) {
299 			retval = -ENOMEM;
300 			goto out;
301 		}
302 		dqm->sdma_queue_count--;
303 		deallocate_sdma_queue(dqm, q->sdma_id);
304 	} else {
305 		pr_debug("q->properties.type is invalid (%d)\n",
306 				q->properties.type);
307 		retval = -EINVAL;
308 		goto out;
309 	}
310 
311 	retval = mqd->destroy_mqd(mqd, q->mqd,
312 				KFD_PREEMPT_TYPE_WAVEFRONT_RESET,
313 				QUEUE_PREEMPT_DEFAULT_TIMEOUT_MS,
314 				q->pipe, q->queue);
315 
316 	if (retval != 0)
317 		goto out;
318 
319 	mqd->uninit_mqd(mqd, q->mqd, q->mqd_mem_obj);
320 
321 	list_del(&q->list);
322 	if (list_empty(&qpd->queues_list))
323 		deallocate_vmid(dqm, qpd, q);
324 	if (q->properties.is_active)
325 		dqm->queue_count--;
326 
327 	/*
328 	 * Unconditionally decrement this counter, regardless of the queue's
329 	 * type
330 	 */
331 	dqm->total_queue_count--;
332 	pr_debug("Total of %d queues are accountable so far\n",
333 			dqm->total_queue_count);
334 
335 out:
336 	mutex_unlock(&dqm->lock);
337 	return retval;
338 }
339 
340 static int update_queue(struct device_queue_manager *dqm, struct queue *q)
341 {
342 	int retval;
343 	struct mqd_manager *mqd;
344 	bool prev_active = false;
345 
346 	BUG_ON(!dqm || !q || !q->mqd);
347 
348 	mutex_lock(&dqm->lock);
349 	mqd = dqm->ops.get_mqd_manager(dqm,
350 			get_mqd_type_from_queue_type(q->properties.type));
351 	if (mqd == NULL) {
352 		mutex_unlock(&dqm->lock);
353 		return -ENOMEM;
354 	}
355 
356 	if (q->properties.is_active == true)
357 		prev_active = true;
358 
359 	/*
360 	 *
361 	 * check active state vs. the previous state
362 	 * and modify counter accordingly
363 	 */
364 	retval = mqd->update_mqd(mqd, q->mqd, &q->properties);
365 	if ((q->properties.is_active == true) && (prev_active == false))
366 		dqm->queue_count++;
367 	else if ((q->properties.is_active == false) && (prev_active == true))
368 		dqm->queue_count--;
369 
370 	if (sched_policy != KFD_SCHED_POLICY_NO_HWS)
371 		retval = execute_queues_cpsch(dqm, false);
372 
373 	mutex_unlock(&dqm->lock);
374 	return retval;
375 }
376 
377 static struct mqd_manager *get_mqd_manager_nocpsch(
378 		struct device_queue_manager *dqm, enum KFD_MQD_TYPE type)
379 {
380 	struct mqd_manager *mqd;
381 
382 	BUG_ON(!dqm || type >= KFD_MQD_TYPE_MAX);
383 
384 	pr_debug("kfd: In func %s mqd type %d\n", __func__, type);
385 
386 	mqd = dqm->mqds[type];
387 	if (!mqd) {
388 		mqd = mqd_manager_init(type, dqm->dev);
389 		if (mqd == NULL)
390 			pr_err("kfd: mqd manager is NULL");
391 		dqm->mqds[type] = mqd;
392 	}
393 
394 	return mqd;
395 }
396 
397 static int register_process_nocpsch(struct device_queue_manager *dqm,
398 					struct qcm_process_device *qpd)
399 {
400 	struct device_process_node *n;
401 	int retval;
402 
403 	BUG_ON(!dqm || !qpd);
404 
405 	pr_debug("kfd: In func %s\n", __func__);
406 
407 	n = kzalloc(sizeof(struct device_process_node), GFP_KERNEL);
408 	if (!n)
409 		return -ENOMEM;
410 
411 	n->qpd = qpd;
412 
413 	mutex_lock(&dqm->lock);
414 	list_add(&n->list, &dqm->queues);
415 
416 	retval = dqm->ops_asic_specific.register_process(dqm, qpd);
417 
418 	dqm->processes_count++;
419 
420 	mutex_unlock(&dqm->lock);
421 
422 	return retval;
423 }
424 
425 static int unregister_process_nocpsch(struct device_queue_manager *dqm,
426 					struct qcm_process_device *qpd)
427 {
428 	int retval;
429 	struct device_process_node *cur, *next;
430 
431 	BUG_ON(!dqm || !qpd);
432 
433 	BUG_ON(!list_empty(&qpd->queues_list));
434 
435 	pr_debug("kfd: In func %s\n", __func__);
436 
437 	retval = 0;
438 	mutex_lock(&dqm->lock);
439 
440 	list_for_each_entry_safe(cur, next, &dqm->queues, list) {
441 		if (qpd == cur->qpd) {
442 			list_del(&cur->list);
443 			kfree(cur);
444 			dqm->processes_count--;
445 			goto out;
446 		}
447 	}
448 	/* qpd not found in dqm list */
449 	retval = 1;
450 out:
451 	mutex_unlock(&dqm->lock);
452 	return retval;
453 }
454 
455 static int
456 set_pasid_vmid_mapping(struct device_queue_manager *dqm, unsigned int pasid,
457 			unsigned int vmid)
458 {
459 	uint32_t pasid_mapping;
460 
461 	pasid_mapping = (pasid == 0) ? 0 :
462 		(uint32_t)pasid |
463 		ATC_VMID_PASID_MAPPING_VALID;
464 
465 	return dqm->dev->kfd2kgd->set_pasid_vmid_mapping(
466 						dqm->dev->kgd, pasid_mapping,
467 						vmid);
468 }
469 
470 int init_pipelines(struct device_queue_manager *dqm,
471 			unsigned int pipes_num, unsigned int first_pipe)
472 {
473 	void *hpdptr;
474 	struct mqd_manager *mqd;
475 	unsigned int i, err, inx;
476 	uint64_t pipe_hpd_addr;
477 
478 	BUG_ON(!dqm || !dqm->dev);
479 
480 	pr_debug("kfd: In func %s\n", __func__);
481 
482 	/*
483 	 * Allocate memory for the HPDs. This is hardware-owned per-pipe data.
484 	 * The driver never accesses this memory after zeroing it.
485 	 * It doesn't even have to be saved/restored on suspend/resume
486 	 * because it contains no data when there are no active queues.
487 	 */
488 
489 	err = kfd_gtt_sa_allocate(dqm->dev, CIK_HPD_EOP_BYTES * pipes_num,
490 					&dqm->pipeline_mem);
491 
492 	if (err) {
493 		pr_err("kfd: error allocate vidmem num pipes: %d\n",
494 			pipes_num);
495 		return -ENOMEM;
496 	}
497 
498 	hpdptr = dqm->pipeline_mem->cpu_ptr;
499 	dqm->pipelines_addr = dqm->pipeline_mem->gpu_addr;
500 
501 	memset(hpdptr, 0, CIK_HPD_EOP_BYTES * pipes_num);
502 
503 	mqd = dqm->ops.get_mqd_manager(dqm, KFD_MQD_TYPE_COMPUTE);
504 	if (mqd == NULL) {
505 		kfd_gtt_sa_free(dqm->dev, dqm->pipeline_mem);
506 		return -ENOMEM;
507 	}
508 
509 	for (i = 0; i < pipes_num; i++) {
510 		inx = i + first_pipe;
511 		/*
512 		 * HPD buffer on GTT is allocated by amdkfd, no need to waste
513 		 * space in GTT for pipelines we don't initialize
514 		 */
515 		pipe_hpd_addr = dqm->pipelines_addr + i * CIK_HPD_EOP_BYTES;
516 		pr_debug("kfd: pipeline address %llX\n", pipe_hpd_addr);
517 		/* = log2(bytes/4)-1 */
518 		dqm->dev->kfd2kgd->init_pipeline(dqm->dev->kgd, inx,
519 				CIK_HPD_EOP_BYTES_LOG2 - 3, pipe_hpd_addr);
520 	}
521 
522 	return 0;
523 }
524 
525 static int init_scheduler(struct device_queue_manager *dqm)
526 {
527 	int retval;
528 
529 	BUG_ON(!dqm);
530 
531 	pr_debug("kfd: In %s\n", __func__);
532 
533 	retval = init_pipelines(dqm, get_pipes_num(dqm), get_first_pipe(dqm));
534 	return retval;
535 }
536 
537 static int initialize_nocpsch(struct device_queue_manager *dqm)
538 {
539 	int i;
540 
541 	BUG_ON(!dqm);
542 
543 	pr_debug("kfd: In func %s num of pipes: %d\n",
544 			__func__, get_pipes_num(dqm));
545 
546 	mutex_init(&dqm->lock);
547 	INIT_LIST_HEAD(&dqm->queues);
548 	dqm->queue_count = dqm->next_pipe_to_allocate = 0;
549 	dqm->sdma_queue_count = 0;
550 	dqm->allocated_queues = kcalloc(get_pipes_num(dqm),
551 					sizeof(unsigned int), GFP_KERNEL);
552 	if (!dqm->allocated_queues) {
553 		mutex_destroy(&dqm->lock);
554 		return -ENOMEM;
555 	}
556 
557 	for (i = 0; i < get_pipes_num(dqm); i++)
558 		dqm->allocated_queues[i] = (1 << QUEUES_PER_PIPE) - 1;
559 
560 	dqm->vmid_bitmap = (1 << VMID_PER_DEVICE) - 1;
561 	dqm->sdma_bitmap = (1 << CIK_SDMA_QUEUES) - 1;
562 
563 	init_scheduler(dqm);
564 	return 0;
565 }
566 
567 static void uninitialize_nocpsch(struct device_queue_manager *dqm)
568 {
569 	int i;
570 
571 	BUG_ON(!dqm);
572 
573 	BUG_ON(dqm->queue_count > 0 || dqm->processes_count > 0);
574 
575 	kfree(dqm->allocated_queues);
576 	for (i = 0 ; i < KFD_MQD_TYPE_MAX ; i++)
577 		kfree(dqm->mqds[i]);
578 	mutex_destroy(&dqm->lock);
579 	kfd_gtt_sa_free(dqm->dev, dqm->pipeline_mem);
580 }
581 
582 static int start_nocpsch(struct device_queue_manager *dqm)
583 {
584 	return 0;
585 }
586 
587 static int stop_nocpsch(struct device_queue_manager *dqm)
588 {
589 	return 0;
590 }
591 
592 static int allocate_sdma_queue(struct device_queue_manager *dqm,
593 				unsigned int *sdma_queue_id)
594 {
595 	int bit;
596 
597 	if (dqm->sdma_bitmap == 0)
598 		return -ENOMEM;
599 
600 	bit = find_first_bit((unsigned long *)&dqm->sdma_bitmap,
601 				CIK_SDMA_QUEUES);
602 
603 	clear_bit(bit, (unsigned long *)&dqm->sdma_bitmap);
604 	*sdma_queue_id = bit;
605 
606 	return 0;
607 }
608 
609 static void deallocate_sdma_queue(struct device_queue_manager *dqm,
610 				unsigned int sdma_queue_id)
611 {
612 	if (sdma_queue_id >= CIK_SDMA_QUEUES)
613 		return;
614 	set_bit(sdma_queue_id, (unsigned long *)&dqm->sdma_bitmap);
615 }
616 
617 static void init_sdma_vm(struct device_queue_manager *dqm, struct queue *q,
618 				struct qcm_process_device *qpd)
619 {
620 	uint32_t value = SDMA_ATC;
621 
622 	if (q->process->is_32bit_user_mode)
623 		value |= SDMA_VA_PTR32 | get_sh_mem_bases_32(qpd_to_pdd(qpd));
624 	else
625 		value |= SDMA_VA_SHARED_BASE(get_sh_mem_bases_nybble_64(
626 							qpd_to_pdd(qpd)));
627 	q->properties.sdma_vm_addr = value;
628 }
629 
630 static int create_sdma_queue_nocpsch(struct device_queue_manager *dqm,
631 					struct queue *q,
632 					struct qcm_process_device *qpd)
633 {
634 	struct mqd_manager *mqd;
635 	int retval;
636 
637 	mqd = dqm->ops.get_mqd_manager(dqm, KFD_MQD_TYPE_SDMA);
638 	if (!mqd)
639 		return -ENOMEM;
640 
641 	retval = allocate_sdma_queue(dqm, &q->sdma_id);
642 	if (retval != 0)
643 		return retval;
644 
645 	q->properties.sdma_queue_id = q->sdma_id % CIK_SDMA_QUEUES_PER_ENGINE;
646 	q->properties.sdma_engine_id = q->sdma_id / CIK_SDMA_ENGINE_NUM;
647 
648 	pr_debug("kfd: sdma id is:    %d\n", q->sdma_id);
649 	pr_debug("     sdma queue id: %d\n", q->properties.sdma_queue_id);
650 	pr_debug("     sdma engine id: %d\n", q->properties.sdma_engine_id);
651 
652 	init_sdma_vm(dqm, q, qpd);
653 	retval = mqd->init_mqd(mqd, &q->mqd, &q->mqd_mem_obj,
654 				&q->gart_mqd_addr, &q->properties);
655 	if (retval != 0) {
656 		deallocate_sdma_queue(dqm, q->sdma_id);
657 		return retval;
658 	}
659 
660 	retval = mqd->load_mqd(mqd, q->mqd, 0,
661 				0, NULL);
662 	if (retval != 0) {
663 		deallocate_sdma_queue(dqm, q->sdma_id);
664 		mqd->uninit_mqd(mqd, q->mqd, q->mqd_mem_obj);
665 		return retval;
666 	}
667 
668 	return 0;
669 }
670 
671 /*
672  * Device Queue Manager implementation for cp scheduler
673  */
674 
675 static int set_sched_resources(struct device_queue_manager *dqm)
676 {
677 	struct scheduling_resources res;
678 	unsigned int queue_num, queue_mask;
679 
680 	BUG_ON(!dqm);
681 
682 	pr_debug("kfd: In func %s\n", __func__);
683 
684 	queue_num = get_pipes_num_cpsch() * QUEUES_PER_PIPE;
685 	queue_mask = (1 << queue_num) - 1;
686 	res.vmid_mask = (1 << VMID_PER_DEVICE) - 1;
687 	res.vmid_mask <<= KFD_VMID_START_OFFSET;
688 	res.queue_mask = queue_mask << (get_first_pipe(dqm) * QUEUES_PER_PIPE);
689 	res.gws_mask = res.oac_mask = res.gds_heap_base =
690 						res.gds_heap_size = 0;
691 
692 	pr_debug("kfd: scheduling resources:\n"
693 			"      vmid mask: 0x%8X\n"
694 			"      queue mask: 0x%8llX\n",
695 			res.vmid_mask, res.queue_mask);
696 
697 	return pm_send_set_resources(&dqm->packets, &res);
698 }
699 
700 static int initialize_cpsch(struct device_queue_manager *dqm)
701 {
702 	int retval;
703 
704 	BUG_ON(!dqm);
705 
706 	pr_debug("kfd: In func %s num of pipes: %d\n",
707 			__func__, get_pipes_num_cpsch());
708 
709 	mutex_init(&dqm->lock);
710 	INIT_LIST_HEAD(&dqm->queues);
711 	dqm->queue_count = dqm->processes_count = 0;
712 	dqm->sdma_queue_count = 0;
713 	dqm->active_runlist = false;
714 	retval = dqm->ops_asic_specific.initialize(dqm);
715 	if (retval != 0)
716 		goto fail_init_pipelines;
717 
718 	return 0;
719 
720 fail_init_pipelines:
721 	mutex_destroy(&dqm->lock);
722 	return retval;
723 }
724 
725 static int start_cpsch(struct device_queue_manager *dqm)
726 {
727 	struct device_process_node *node;
728 	int retval;
729 
730 	BUG_ON(!dqm);
731 
732 	retval = 0;
733 
734 	retval = pm_init(&dqm->packets, dqm);
735 	if (retval != 0)
736 		goto fail_packet_manager_init;
737 
738 	retval = set_sched_resources(dqm);
739 	if (retval != 0)
740 		goto fail_set_sched_resources;
741 
742 	pr_debug("kfd: allocating fence memory\n");
743 
744 	/* allocate fence memory on the gart */
745 	retval = kfd_gtt_sa_allocate(dqm->dev, sizeof(*dqm->fence_addr),
746 					&dqm->fence_mem);
747 
748 	if (retval != 0)
749 		goto fail_allocate_vidmem;
750 
751 	dqm->fence_addr = dqm->fence_mem->cpu_ptr;
752 	dqm->fence_gpu_addr = dqm->fence_mem->gpu_addr;
753 	list_for_each_entry(node, &dqm->queues, list)
754 		if (node->qpd->pqm->process && dqm->dev)
755 			kfd_bind_process_to_device(dqm->dev,
756 						node->qpd->pqm->process);
757 
758 	execute_queues_cpsch(dqm, true);
759 
760 	return 0;
761 fail_allocate_vidmem:
762 fail_set_sched_resources:
763 	pm_uninit(&dqm->packets);
764 fail_packet_manager_init:
765 	return retval;
766 }
767 
768 static int stop_cpsch(struct device_queue_manager *dqm)
769 {
770 	struct device_process_node *node;
771 	struct kfd_process_device *pdd;
772 
773 	BUG_ON(!dqm);
774 
775 	destroy_queues_cpsch(dqm, true);
776 
777 	list_for_each_entry(node, &dqm->queues, list) {
778 		pdd = qpd_to_pdd(node->qpd);
779 		pdd->bound = false;
780 	}
781 	kfd_gtt_sa_free(dqm->dev, dqm->fence_mem);
782 	pm_uninit(&dqm->packets);
783 
784 	return 0;
785 }
786 
787 static int create_kernel_queue_cpsch(struct device_queue_manager *dqm,
788 					struct kernel_queue *kq,
789 					struct qcm_process_device *qpd)
790 {
791 	BUG_ON(!dqm || !kq || !qpd);
792 
793 	pr_debug("kfd: In func %s\n", __func__);
794 
795 	mutex_lock(&dqm->lock);
796 	if (dqm->total_queue_count >= max_num_of_queues_per_device) {
797 		pr_warn("amdkfd: Can't create new kernel queue because %d queues were already created\n",
798 				dqm->total_queue_count);
799 		mutex_unlock(&dqm->lock);
800 		return -EPERM;
801 	}
802 
803 	/*
804 	 * Unconditionally increment this counter, regardless of the queue's
805 	 * type or whether the queue is active.
806 	 */
807 	dqm->total_queue_count++;
808 	pr_debug("Total of %d queues are accountable so far\n",
809 			dqm->total_queue_count);
810 
811 	list_add(&kq->list, &qpd->priv_queue_list);
812 	dqm->queue_count++;
813 	qpd->is_debug = true;
814 	execute_queues_cpsch(dqm, false);
815 	mutex_unlock(&dqm->lock);
816 
817 	return 0;
818 }
819 
820 static void destroy_kernel_queue_cpsch(struct device_queue_manager *dqm,
821 					struct kernel_queue *kq,
822 					struct qcm_process_device *qpd)
823 {
824 	BUG_ON(!dqm || !kq);
825 
826 	pr_debug("kfd: In %s\n", __func__);
827 
828 	mutex_lock(&dqm->lock);
829 	destroy_queues_cpsch(dqm, false);
830 	list_del(&kq->list);
831 	dqm->queue_count--;
832 	qpd->is_debug = false;
833 	execute_queues_cpsch(dqm, false);
834 	/*
835 	 * Unconditionally decrement this counter, regardless of the queue's
836 	 * type.
837 	 */
838 	dqm->total_queue_count--;
839 	pr_debug("Total of %d queues are accountable so far\n",
840 			dqm->total_queue_count);
841 	mutex_unlock(&dqm->lock);
842 }
843 
844 static void select_sdma_engine_id(struct queue *q)
845 {
846 	static int sdma_id;
847 
848 	q->sdma_id = sdma_id;
849 	sdma_id = (sdma_id + 1) % 2;
850 }
851 
852 static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q,
853 			struct qcm_process_device *qpd, int *allocate_vmid)
854 {
855 	int retval;
856 	struct mqd_manager *mqd;
857 
858 	BUG_ON(!dqm || !q || !qpd);
859 
860 	retval = 0;
861 
862 	if (allocate_vmid)
863 		*allocate_vmid = 0;
864 
865 	mutex_lock(&dqm->lock);
866 
867 	if (dqm->total_queue_count >= max_num_of_queues_per_device) {
868 		pr_warn("amdkfd: Can't create new usermode queue because %d queues were already created\n",
869 				dqm->total_queue_count);
870 		retval = -EPERM;
871 		goto out;
872 	}
873 
874 	if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
875 		select_sdma_engine_id(q);
876 
877 	mqd = dqm->ops.get_mqd_manager(dqm,
878 			get_mqd_type_from_queue_type(q->properties.type));
879 
880 	if (mqd == NULL) {
881 		mutex_unlock(&dqm->lock);
882 		return -ENOMEM;
883 	}
884 
885 	retval = mqd->init_mqd(mqd, &q->mqd, &q->mqd_mem_obj,
886 				&q->gart_mqd_addr, &q->properties);
887 	if (retval != 0)
888 		goto out;
889 
890 	list_add(&q->list, &qpd->queues_list);
891 	if (q->properties.is_active) {
892 		dqm->queue_count++;
893 		retval = execute_queues_cpsch(dqm, false);
894 	}
895 
896 	if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
897 			dqm->sdma_queue_count++;
898 	/*
899 	 * Unconditionally increment this counter, regardless of the queue's
900 	 * type or whether the queue is active.
901 	 */
902 	dqm->total_queue_count++;
903 
904 	pr_debug("Total of %d queues are accountable so far\n",
905 			dqm->total_queue_count);
906 
907 out:
908 	mutex_unlock(&dqm->lock);
909 	return retval;
910 }
911 
912 static int amdkfd_fence_wait_timeout(unsigned int *fence_addr,
913 				unsigned int fence_value,
914 				unsigned long timeout)
915 {
916 	BUG_ON(!fence_addr);
917 	timeout += jiffies;
918 
919 	while (*fence_addr != fence_value) {
920 		if (time_after(jiffies, timeout)) {
921 			pr_err("kfd: qcm fence wait loop timeout expired\n");
922 			return -ETIME;
923 		}
924 		schedule();
925 	}
926 
927 	return 0;
928 }
929 
930 static int destroy_sdma_queues(struct device_queue_manager *dqm,
931 				unsigned int sdma_engine)
932 {
933 	return pm_send_unmap_queue(&dqm->packets, KFD_QUEUE_TYPE_SDMA,
934 			KFD_PREEMPT_TYPE_FILTER_ALL_QUEUES, 0, false,
935 			sdma_engine);
936 }
937 
938 static int destroy_queues_cpsch(struct device_queue_manager *dqm, bool lock)
939 {
940 	int retval;
941 
942 	BUG_ON(!dqm);
943 
944 	retval = 0;
945 
946 	if (lock)
947 		mutex_lock(&dqm->lock);
948 	if (dqm->active_runlist == false)
949 		goto out;
950 
951 	pr_debug("kfd: Before destroying queues, sdma queue count is : %u\n",
952 		dqm->sdma_queue_count);
953 
954 	if (dqm->sdma_queue_count > 0) {
955 		destroy_sdma_queues(dqm, 0);
956 		destroy_sdma_queues(dqm, 1);
957 	}
958 
959 	retval = pm_send_unmap_queue(&dqm->packets, KFD_QUEUE_TYPE_COMPUTE,
960 			KFD_PREEMPT_TYPE_FILTER_ALL_QUEUES, 0, false, 0);
961 	if (retval != 0)
962 		goto out;
963 
964 	*dqm->fence_addr = KFD_FENCE_INIT;
965 	pm_send_query_status(&dqm->packets, dqm->fence_gpu_addr,
966 				KFD_FENCE_COMPLETED);
967 	/* should be timed out */
968 	amdkfd_fence_wait_timeout(dqm->fence_addr, KFD_FENCE_COMPLETED,
969 				QUEUE_PREEMPT_DEFAULT_TIMEOUT_MS);
970 	pm_release_ib(&dqm->packets);
971 	dqm->active_runlist = false;
972 
973 out:
974 	if (lock)
975 		mutex_unlock(&dqm->lock);
976 	return retval;
977 }
978 
979 static int execute_queues_cpsch(struct device_queue_manager *dqm, bool lock)
980 {
981 	int retval;
982 
983 	BUG_ON(!dqm);
984 
985 	if (lock)
986 		mutex_lock(&dqm->lock);
987 
988 	retval = destroy_queues_cpsch(dqm, false);
989 	if (retval != 0) {
990 		pr_err("kfd: the cp might be in an unrecoverable state due to an unsuccessful queues preemption");
991 		goto out;
992 	}
993 
994 	if (dqm->queue_count <= 0 || dqm->processes_count <= 0) {
995 		retval = 0;
996 		goto out;
997 	}
998 
999 	if (dqm->active_runlist) {
1000 		retval = 0;
1001 		goto out;
1002 	}
1003 
1004 	retval = pm_send_runlist(&dqm->packets, &dqm->queues);
1005 	if (retval != 0) {
1006 		pr_err("kfd: failed to execute runlist");
1007 		goto out;
1008 	}
1009 	dqm->active_runlist = true;
1010 
1011 out:
1012 	if (lock)
1013 		mutex_unlock(&dqm->lock);
1014 	return retval;
1015 }
1016 
1017 static int destroy_queue_cpsch(struct device_queue_manager *dqm,
1018 				struct qcm_process_device *qpd,
1019 				struct queue *q)
1020 {
1021 	int retval;
1022 	struct mqd_manager *mqd;
1023 
1024 	BUG_ON(!dqm || !qpd || !q);
1025 
1026 	retval = 0;
1027 
1028 	/* remove queue from list to prevent rescheduling after preemption */
1029 	mutex_lock(&dqm->lock);
1030 	mqd = dqm->ops.get_mqd_manager(dqm,
1031 			get_mqd_type_from_queue_type(q->properties.type));
1032 	if (!mqd) {
1033 		retval = -ENOMEM;
1034 		goto failed;
1035 	}
1036 
1037 	if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
1038 		dqm->sdma_queue_count--;
1039 
1040 	list_del(&q->list);
1041 	if (q->properties.is_active)
1042 		dqm->queue_count--;
1043 
1044 	execute_queues_cpsch(dqm, false);
1045 
1046 	mqd->uninit_mqd(mqd, q->mqd, q->mqd_mem_obj);
1047 
1048 	/*
1049 	 * Unconditionally decrement this counter, regardless of the queue's
1050 	 * type
1051 	 */
1052 	dqm->total_queue_count--;
1053 	pr_debug("Total of %d queues are accountable so far\n",
1054 			dqm->total_queue_count);
1055 
1056 	mutex_unlock(&dqm->lock);
1057 
1058 	return 0;
1059 
1060 failed:
1061 	mutex_unlock(&dqm->lock);
1062 	return retval;
1063 }
1064 
1065 /*
1066  * Low bits must be 0000/FFFF as required by HW, high bits must be 0 to
1067  * stay in user mode.
1068  */
1069 #define APE1_FIXED_BITS_MASK 0xFFFF80000000FFFFULL
1070 /* APE1 limit is inclusive and 64K aligned. */
1071 #define APE1_LIMIT_ALIGNMENT 0xFFFF
1072 
1073 static bool set_cache_memory_policy(struct device_queue_manager *dqm,
1074 				   struct qcm_process_device *qpd,
1075 				   enum cache_policy default_policy,
1076 				   enum cache_policy alternate_policy,
1077 				   void __user *alternate_aperture_base,
1078 				   uint64_t alternate_aperture_size)
1079 {
1080 	bool retval;
1081 
1082 	pr_debug("kfd: In func %s\n", __func__);
1083 
1084 	mutex_lock(&dqm->lock);
1085 
1086 	if (alternate_aperture_size == 0) {
1087 		/* base > limit disables APE1 */
1088 		qpd->sh_mem_ape1_base = 1;
1089 		qpd->sh_mem_ape1_limit = 0;
1090 	} else {
1091 		/*
1092 		 * In FSA64, APE1_Base[63:0] = { 16{SH_MEM_APE1_BASE[31]},
1093 		 *			SH_MEM_APE1_BASE[31:0], 0x0000 }
1094 		 * APE1_Limit[63:0] = { 16{SH_MEM_APE1_LIMIT[31]},
1095 		 *			SH_MEM_APE1_LIMIT[31:0], 0xFFFF }
1096 		 * Verify that the base and size parameters can be
1097 		 * represented in this format and convert them.
1098 		 * Additionally restrict APE1 to user-mode addresses.
1099 		 */
1100 
1101 		uint64_t base = (uintptr_t)alternate_aperture_base;
1102 		uint64_t limit = base + alternate_aperture_size - 1;
1103 
1104 		if (limit <= base)
1105 			goto out;
1106 
1107 		if ((base & APE1_FIXED_BITS_MASK) != 0)
1108 			goto out;
1109 
1110 		if ((limit & APE1_FIXED_BITS_MASK) != APE1_LIMIT_ALIGNMENT)
1111 			goto out;
1112 
1113 		qpd->sh_mem_ape1_base = base >> 16;
1114 		qpd->sh_mem_ape1_limit = limit >> 16;
1115 	}
1116 
1117 	retval = dqm->ops_asic_specific.set_cache_memory_policy(
1118 			dqm,
1119 			qpd,
1120 			default_policy,
1121 			alternate_policy,
1122 			alternate_aperture_base,
1123 			alternate_aperture_size);
1124 
1125 	if ((sched_policy == KFD_SCHED_POLICY_NO_HWS) && (qpd->vmid != 0))
1126 		program_sh_mem_settings(dqm, qpd);
1127 
1128 	pr_debug("kfd: sh_mem_config: 0x%x, ape1_base: 0x%x, ape1_limit: 0x%x\n",
1129 		qpd->sh_mem_config, qpd->sh_mem_ape1_base,
1130 		qpd->sh_mem_ape1_limit);
1131 
1132 	mutex_unlock(&dqm->lock);
1133 	return retval;
1134 
1135 out:
1136 	mutex_unlock(&dqm->lock);
1137 	return false;
1138 }
1139 
1140 struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev)
1141 {
1142 	struct device_queue_manager *dqm;
1143 
1144 	BUG_ON(!dev);
1145 
1146 	pr_debug("kfd: loading device queue manager\n");
1147 
1148 	dqm = kzalloc(sizeof(struct device_queue_manager), GFP_KERNEL);
1149 	if (!dqm)
1150 		return NULL;
1151 
1152 	dqm->dev = dev;
1153 	switch (sched_policy) {
1154 	case KFD_SCHED_POLICY_HWS:
1155 	case KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION:
1156 		/* initialize dqm for cp scheduling */
1157 		dqm->ops.create_queue = create_queue_cpsch;
1158 		dqm->ops.initialize = initialize_cpsch;
1159 		dqm->ops.start = start_cpsch;
1160 		dqm->ops.stop = stop_cpsch;
1161 		dqm->ops.destroy_queue = destroy_queue_cpsch;
1162 		dqm->ops.update_queue = update_queue;
1163 		dqm->ops.get_mqd_manager = get_mqd_manager_nocpsch;
1164 		dqm->ops.register_process = register_process_nocpsch;
1165 		dqm->ops.unregister_process = unregister_process_nocpsch;
1166 		dqm->ops.uninitialize = uninitialize_nocpsch;
1167 		dqm->ops.create_kernel_queue = create_kernel_queue_cpsch;
1168 		dqm->ops.destroy_kernel_queue = destroy_kernel_queue_cpsch;
1169 		dqm->ops.set_cache_memory_policy = set_cache_memory_policy;
1170 		break;
1171 	case KFD_SCHED_POLICY_NO_HWS:
1172 		/* initialize dqm for no cp scheduling */
1173 		dqm->ops.start = start_nocpsch;
1174 		dqm->ops.stop = stop_nocpsch;
1175 		dqm->ops.create_queue = create_queue_nocpsch;
1176 		dqm->ops.destroy_queue = destroy_queue_nocpsch;
1177 		dqm->ops.update_queue = update_queue;
1178 		dqm->ops.get_mqd_manager = get_mqd_manager_nocpsch;
1179 		dqm->ops.register_process = register_process_nocpsch;
1180 		dqm->ops.unregister_process = unregister_process_nocpsch;
1181 		dqm->ops.initialize = initialize_nocpsch;
1182 		dqm->ops.uninitialize = uninitialize_nocpsch;
1183 		dqm->ops.set_cache_memory_policy = set_cache_memory_policy;
1184 		break;
1185 	default:
1186 		BUG();
1187 		break;
1188 	}
1189 
1190 	switch (dev->device_info->asic_family) {
1191 	case CHIP_CARRIZO:
1192 		device_queue_manager_init_vi(&dqm->ops_asic_specific);
1193 		break;
1194 
1195 	case CHIP_KAVERI:
1196 		device_queue_manager_init_cik(&dqm->ops_asic_specific);
1197 		break;
1198 	}
1199 
1200 	if (dqm->ops.initialize(dqm) != 0) {
1201 		kfree(dqm);
1202 		return NULL;
1203 	}
1204 
1205 	return dqm;
1206 }
1207 
1208 void device_queue_manager_uninit(struct device_queue_manager *dqm)
1209 {
1210 	BUG_ON(!dqm);
1211 
1212 	dqm->ops.uninitialize(dqm);
1213 	kfree(dqm);
1214 }
1215