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 	pr_debug("In func %s\n", __func__);
434 
435 	pr_debug("qpd->queues_list is %s\n",
436 			list_empty(&qpd->queues_list) ? "empty" : "not empty");
437 
438 	retval = 0;
439 	mutex_lock(&dqm->lock);
440 
441 	list_for_each_entry_safe(cur, next, &dqm->queues, list) {
442 		if (qpd == cur->qpd) {
443 			list_del(&cur->list);
444 			kfree(cur);
445 			dqm->processes_count--;
446 			goto out;
447 		}
448 	}
449 	/* qpd not found in dqm list */
450 	retval = 1;
451 out:
452 	mutex_unlock(&dqm->lock);
453 	return retval;
454 }
455 
456 static int
457 set_pasid_vmid_mapping(struct device_queue_manager *dqm, unsigned int pasid,
458 			unsigned int vmid)
459 {
460 	uint32_t pasid_mapping;
461 
462 	pasid_mapping = (pasid == 0) ? 0 :
463 		(uint32_t)pasid |
464 		ATC_VMID_PASID_MAPPING_VALID;
465 
466 	return dqm->dev->kfd2kgd->set_pasid_vmid_mapping(
467 						dqm->dev->kgd, pasid_mapping,
468 						vmid);
469 }
470 
471 int init_pipelines(struct device_queue_manager *dqm,
472 			unsigned int pipes_num, unsigned int first_pipe)
473 {
474 	void *hpdptr;
475 	struct mqd_manager *mqd;
476 	unsigned int i, err, inx;
477 	uint64_t pipe_hpd_addr;
478 
479 	BUG_ON(!dqm || !dqm->dev);
480 
481 	pr_debug("kfd: In func %s\n", __func__);
482 
483 	/*
484 	 * Allocate memory for the HPDs. This is hardware-owned per-pipe data.
485 	 * The driver never accesses this memory after zeroing it.
486 	 * It doesn't even have to be saved/restored on suspend/resume
487 	 * because it contains no data when there are no active queues.
488 	 */
489 
490 	err = kfd_gtt_sa_allocate(dqm->dev, CIK_HPD_EOP_BYTES * pipes_num,
491 					&dqm->pipeline_mem);
492 
493 	if (err) {
494 		pr_err("kfd: error allocate vidmem num pipes: %d\n",
495 			pipes_num);
496 		return -ENOMEM;
497 	}
498 
499 	hpdptr = dqm->pipeline_mem->cpu_ptr;
500 	dqm->pipelines_addr = dqm->pipeline_mem->gpu_addr;
501 
502 	memset(hpdptr, 0, CIK_HPD_EOP_BYTES * pipes_num);
503 
504 	mqd = dqm->ops.get_mqd_manager(dqm, KFD_MQD_TYPE_COMPUTE);
505 	if (mqd == NULL) {
506 		kfd_gtt_sa_free(dqm->dev, dqm->pipeline_mem);
507 		return -ENOMEM;
508 	}
509 
510 	for (i = 0; i < pipes_num; i++) {
511 		inx = i + first_pipe;
512 		/*
513 		 * HPD buffer on GTT is allocated by amdkfd, no need to waste
514 		 * space in GTT for pipelines we don't initialize
515 		 */
516 		pipe_hpd_addr = dqm->pipelines_addr + i * CIK_HPD_EOP_BYTES;
517 		pr_debug("kfd: pipeline address %llX\n", pipe_hpd_addr);
518 		/* = log2(bytes/4)-1 */
519 		dqm->dev->kfd2kgd->init_pipeline(dqm->dev->kgd, inx,
520 				CIK_HPD_EOP_BYTES_LOG2 - 3, pipe_hpd_addr);
521 	}
522 
523 	return 0;
524 }
525 
526 static int init_scheduler(struct device_queue_manager *dqm)
527 {
528 	int retval;
529 
530 	BUG_ON(!dqm);
531 
532 	pr_debug("kfd: In %s\n", __func__);
533 
534 	retval = init_pipelines(dqm, get_pipes_num(dqm), get_first_pipe(dqm));
535 	return retval;
536 }
537 
538 static int initialize_nocpsch(struct device_queue_manager *dqm)
539 {
540 	int i;
541 
542 	BUG_ON(!dqm);
543 
544 	pr_debug("kfd: In func %s num of pipes: %d\n",
545 			__func__, get_pipes_num(dqm));
546 
547 	mutex_init(&dqm->lock);
548 	INIT_LIST_HEAD(&dqm->queues);
549 	dqm->queue_count = dqm->next_pipe_to_allocate = 0;
550 	dqm->sdma_queue_count = 0;
551 	dqm->allocated_queues = kcalloc(get_pipes_num(dqm),
552 					sizeof(unsigned int), GFP_KERNEL);
553 	if (!dqm->allocated_queues) {
554 		mutex_destroy(&dqm->lock);
555 		return -ENOMEM;
556 	}
557 
558 	for (i = 0; i < get_pipes_num(dqm); i++)
559 		dqm->allocated_queues[i] = (1 << QUEUES_PER_PIPE) - 1;
560 
561 	dqm->vmid_bitmap = (1 << VMID_PER_DEVICE) - 1;
562 	dqm->sdma_bitmap = (1 << CIK_SDMA_QUEUES) - 1;
563 
564 	init_scheduler(dqm);
565 	return 0;
566 }
567 
568 static void uninitialize_nocpsch(struct device_queue_manager *dqm)
569 {
570 	int i;
571 
572 	BUG_ON(!dqm);
573 
574 	BUG_ON(dqm->queue_count > 0 || dqm->processes_count > 0);
575 
576 	kfree(dqm->allocated_queues);
577 	for (i = 0 ; i < KFD_MQD_TYPE_MAX ; i++)
578 		kfree(dqm->mqds[i]);
579 	mutex_destroy(&dqm->lock);
580 	kfd_gtt_sa_free(dqm->dev, dqm->pipeline_mem);
581 }
582 
583 static int start_nocpsch(struct device_queue_manager *dqm)
584 {
585 	return 0;
586 }
587 
588 static int stop_nocpsch(struct device_queue_manager *dqm)
589 {
590 	return 0;
591 }
592 
593 static int allocate_sdma_queue(struct device_queue_manager *dqm,
594 				unsigned int *sdma_queue_id)
595 {
596 	int bit;
597 
598 	if (dqm->sdma_bitmap == 0)
599 		return -ENOMEM;
600 
601 	bit = find_first_bit((unsigned long *)&dqm->sdma_bitmap,
602 				CIK_SDMA_QUEUES);
603 
604 	clear_bit(bit, (unsigned long *)&dqm->sdma_bitmap);
605 	*sdma_queue_id = bit;
606 
607 	return 0;
608 }
609 
610 static void deallocate_sdma_queue(struct device_queue_manager *dqm,
611 				unsigned int sdma_queue_id)
612 {
613 	if (sdma_queue_id >= CIK_SDMA_QUEUES)
614 		return;
615 	set_bit(sdma_queue_id, (unsigned long *)&dqm->sdma_bitmap);
616 }
617 
618 static void init_sdma_vm(struct device_queue_manager *dqm, struct queue *q,
619 				struct qcm_process_device *qpd)
620 {
621 	uint32_t value = SDMA_ATC;
622 
623 	if (q->process->is_32bit_user_mode)
624 		value |= SDMA_VA_PTR32 | get_sh_mem_bases_32(qpd_to_pdd(qpd));
625 	else
626 		value |= SDMA_VA_SHARED_BASE(get_sh_mem_bases_nybble_64(
627 							qpd_to_pdd(qpd)));
628 	q->properties.sdma_vm_addr = value;
629 }
630 
631 static int create_sdma_queue_nocpsch(struct device_queue_manager *dqm,
632 					struct queue *q,
633 					struct qcm_process_device *qpd)
634 {
635 	struct mqd_manager *mqd;
636 	int retval;
637 
638 	mqd = dqm->ops.get_mqd_manager(dqm, KFD_MQD_TYPE_SDMA);
639 	if (!mqd)
640 		return -ENOMEM;
641 
642 	retval = allocate_sdma_queue(dqm, &q->sdma_id);
643 	if (retval != 0)
644 		return retval;
645 
646 	q->properties.sdma_queue_id = q->sdma_id % CIK_SDMA_QUEUES_PER_ENGINE;
647 	q->properties.sdma_engine_id = q->sdma_id / CIK_SDMA_ENGINE_NUM;
648 
649 	pr_debug("kfd: sdma id is:    %d\n", q->sdma_id);
650 	pr_debug("     sdma queue id: %d\n", q->properties.sdma_queue_id);
651 	pr_debug("     sdma engine id: %d\n", q->properties.sdma_engine_id);
652 
653 	init_sdma_vm(dqm, q, qpd);
654 	retval = mqd->init_mqd(mqd, &q->mqd, &q->mqd_mem_obj,
655 				&q->gart_mqd_addr, &q->properties);
656 	if (retval != 0) {
657 		deallocate_sdma_queue(dqm, q->sdma_id);
658 		return retval;
659 	}
660 
661 	retval = mqd->load_mqd(mqd, q->mqd, 0,
662 				0, NULL);
663 	if (retval != 0) {
664 		deallocate_sdma_queue(dqm, q->sdma_id);
665 		mqd->uninit_mqd(mqd, q->mqd, q->mqd_mem_obj);
666 		return retval;
667 	}
668 
669 	return 0;
670 }
671 
672 /*
673  * Device Queue Manager implementation for cp scheduler
674  */
675 
676 static int set_sched_resources(struct device_queue_manager *dqm)
677 {
678 	struct scheduling_resources res;
679 	unsigned int queue_num, queue_mask;
680 
681 	BUG_ON(!dqm);
682 
683 	pr_debug("kfd: In func %s\n", __func__);
684 
685 	queue_num = get_pipes_num_cpsch() * QUEUES_PER_PIPE;
686 	queue_mask = (1 << queue_num) - 1;
687 	res.vmid_mask = (1 << VMID_PER_DEVICE) - 1;
688 	res.vmid_mask <<= KFD_VMID_START_OFFSET;
689 	res.queue_mask = queue_mask << (get_first_pipe(dqm) * QUEUES_PER_PIPE);
690 	res.gws_mask = res.oac_mask = res.gds_heap_base =
691 						res.gds_heap_size = 0;
692 
693 	pr_debug("kfd: scheduling resources:\n"
694 			"      vmid mask: 0x%8X\n"
695 			"      queue mask: 0x%8llX\n",
696 			res.vmid_mask, res.queue_mask);
697 
698 	return pm_send_set_resources(&dqm->packets, &res);
699 }
700 
701 static int initialize_cpsch(struct device_queue_manager *dqm)
702 {
703 	int retval;
704 
705 	BUG_ON(!dqm);
706 
707 	pr_debug("kfd: In func %s num of pipes: %d\n",
708 			__func__, get_pipes_num_cpsch());
709 
710 	mutex_init(&dqm->lock);
711 	INIT_LIST_HEAD(&dqm->queues);
712 	dqm->queue_count = dqm->processes_count = 0;
713 	dqm->sdma_queue_count = 0;
714 	dqm->active_runlist = false;
715 	retval = dqm->ops_asic_specific.initialize(dqm);
716 	if (retval != 0)
717 		goto fail_init_pipelines;
718 
719 	return 0;
720 
721 fail_init_pipelines:
722 	mutex_destroy(&dqm->lock);
723 	return retval;
724 }
725 
726 static int start_cpsch(struct device_queue_manager *dqm)
727 {
728 	struct device_process_node *node;
729 	int retval;
730 
731 	BUG_ON(!dqm);
732 
733 	retval = 0;
734 
735 	retval = pm_init(&dqm->packets, dqm);
736 	if (retval != 0)
737 		goto fail_packet_manager_init;
738 
739 	retval = set_sched_resources(dqm);
740 	if (retval != 0)
741 		goto fail_set_sched_resources;
742 
743 	pr_debug("kfd: allocating fence memory\n");
744 
745 	/* allocate fence memory on the gart */
746 	retval = kfd_gtt_sa_allocate(dqm->dev, sizeof(*dqm->fence_addr),
747 					&dqm->fence_mem);
748 
749 	if (retval != 0)
750 		goto fail_allocate_vidmem;
751 
752 	dqm->fence_addr = dqm->fence_mem->cpu_ptr;
753 	dqm->fence_gpu_addr = dqm->fence_mem->gpu_addr;
754 	list_for_each_entry(node, &dqm->queues, list)
755 		if (node->qpd->pqm->process && dqm->dev)
756 			kfd_bind_process_to_device(dqm->dev,
757 						node->qpd->pqm->process);
758 
759 	execute_queues_cpsch(dqm, true);
760 
761 	return 0;
762 fail_allocate_vidmem:
763 fail_set_sched_resources:
764 	pm_uninit(&dqm->packets);
765 fail_packet_manager_init:
766 	return retval;
767 }
768 
769 static int stop_cpsch(struct device_queue_manager *dqm)
770 {
771 	struct device_process_node *node;
772 	struct kfd_process_device *pdd;
773 
774 	BUG_ON(!dqm);
775 
776 	destroy_queues_cpsch(dqm, true);
777 
778 	list_for_each_entry(node, &dqm->queues, list) {
779 		pdd = qpd_to_pdd(node->qpd);
780 		pdd->bound = false;
781 	}
782 	kfd_gtt_sa_free(dqm->dev, dqm->fence_mem);
783 	pm_uninit(&dqm->packets);
784 
785 	return 0;
786 }
787 
788 static int create_kernel_queue_cpsch(struct device_queue_manager *dqm,
789 					struct kernel_queue *kq,
790 					struct qcm_process_device *qpd)
791 {
792 	BUG_ON(!dqm || !kq || !qpd);
793 
794 	pr_debug("kfd: In func %s\n", __func__);
795 
796 	mutex_lock(&dqm->lock);
797 	if (dqm->total_queue_count >= max_num_of_queues_per_device) {
798 		pr_warn("amdkfd: Can't create new kernel queue because %d queues were already created\n",
799 				dqm->total_queue_count);
800 		mutex_unlock(&dqm->lock);
801 		return -EPERM;
802 	}
803 
804 	/*
805 	 * Unconditionally increment this counter, regardless of the queue's
806 	 * type or whether the queue is active.
807 	 */
808 	dqm->total_queue_count++;
809 	pr_debug("Total of %d queues are accountable so far\n",
810 			dqm->total_queue_count);
811 
812 	list_add(&kq->list, &qpd->priv_queue_list);
813 	dqm->queue_count++;
814 	qpd->is_debug = true;
815 	execute_queues_cpsch(dqm, false);
816 	mutex_unlock(&dqm->lock);
817 
818 	return 0;
819 }
820 
821 static void destroy_kernel_queue_cpsch(struct device_queue_manager *dqm,
822 					struct kernel_queue *kq,
823 					struct qcm_process_device *qpd)
824 {
825 	BUG_ON(!dqm || !kq);
826 
827 	pr_debug("kfd: In %s\n", __func__);
828 
829 	mutex_lock(&dqm->lock);
830 	destroy_queues_cpsch(dqm, false);
831 	list_del(&kq->list);
832 	dqm->queue_count--;
833 	qpd->is_debug = false;
834 	execute_queues_cpsch(dqm, false);
835 	/*
836 	 * Unconditionally decrement this counter, regardless of the queue's
837 	 * type.
838 	 */
839 	dqm->total_queue_count--;
840 	pr_debug("Total of %d queues are accountable so far\n",
841 			dqm->total_queue_count);
842 	mutex_unlock(&dqm->lock);
843 }
844 
845 static void select_sdma_engine_id(struct queue *q)
846 {
847 	static int sdma_id;
848 
849 	q->sdma_id = sdma_id;
850 	sdma_id = (sdma_id + 1) % 2;
851 }
852 
853 static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q,
854 			struct qcm_process_device *qpd, int *allocate_vmid)
855 {
856 	int retval;
857 	struct mqd_manager *mqd;
858 
859 	BUG_ON(!dqm || !q || !qpd);
860 
861 	retval = 0;
862 
863 	if (allocate_vmid)
864 		*allocate_vmid = 0;
865 
866 	mutex_lock(&dqm->lock);
867 
868 	if (dqm->total_queue_count >= max_num_of_queues_per_device) {
869 		pr_warn("amdkfd: Can't create new usermode queue because %d queues were already created\n",
870 				dqm->total_queue_count);
871 		retval = -EPERM;
872 		goto out;
873 	}
874 
875 	if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
876 		select_sdma_engine_id(q);
877 
878 	mqd = dqm->ops.get_mqd_manager(dqm,
879 			get_mqd_type_from_queue_type(q->properties.type));
880 
881 	if (mqd == NULL) {
882 		mutex_unlock(&dqm->lock);
883 		return -ENOMEM;
884 	}
885 
886 	init_sdma_vm(dqm, q, qpd);
887 
888 	retval = mqd->init_mqd(mqd, &q->mqd, &q->mqd_mem_obj,
889 				&q->gart_mqd_addr, &q->properties);
890 	if (retval != 0)
891 		goto out;
892 
893 	list_add(&q->list, &qpd->queues_list);
894 	if (q->properties.is_active) {
895 		dqm->queue_count++;
896 		retval = execute_queues_cpsch(dqm, false);
897 	}
898 
899 	if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
900 			dqm->sdma_queue_count++;
901 	/*
902 	 * Unconditionally increment this counter, regardless of the queue's
903 	 * type or whether the queue is active.
904 	 */
905 	dqm->total_queue_count++;
906 
907 	pr_debug("Total of %d queues are accountable so far\n",
908 			dqm->total_queue_count);
909 
910 out:
911 	mutex_unlock(&dqm->lock);
912 	return retval;
913 }
914 
915 static int amdkfd_fence_wait_timeout(unsigned int *fence_addr,
916 				unsigned int fence_value,
917 				unsigned long timeout)
918 {
919 	BUG_ON(!fence_addr);
920 	timeout += jiffies;
921 
922 	while (*fence_addr != fence_value) {
923 		if (time_after(jiffies, timeout)) {
924 			pr_err("kfd: qcm fence wait loop timeout expired\n");
925 			return -ETIME;
926 		}
927 		schedule();
928 	}
929 
930 	return 0;
931 }
932 
933 static int destroy_sdma_queues(struct device_queue_manager *dqm,
934 				unsigned int sdma_engine)
935 {
936 	return pm_send_unmap_queue(&dqm->packets, KFD_QUEUE_TYPE_SDMA,
937 			KFD_PREEMPT_TYPE_FILTER_ALL_QUEUES, 0, false,
938 			sdma_engine);
939 }
940 
941 static int destroy_queues_cpsch(struct device_queue_manager *dqm, bool lock)
942 {
943 	int retval;
944 
945 	BUG_ON(!dqm);
946 
947 	retval = 0;
948 
949 	if (lock)
950 		mutex_lock(&dqm->lock);
951 	if (dqm->active_runlist == false)
952 		goto out;
953 
954 	pr_debug("kfd: Before destroying queues, sdma queue count is : %u\n",
955 		dqm->sdma_queue_count);
956 
957 	if (dqm->sdma_queue_count > 0) {
958 		destroy_sdma_queues(dqm, 0);
959 		destroy_sdma_queues(dqm, 1);
960 	}
961 
962 	retval = pm_send_unmap_queue(&dqm->packets, KFD_QUEUE_TYPE_COMPUTE,
963 			KFD_PREEMPT_TYPE_FILTER_ALL_QUEUES, 0, false, 0);
964 	if (retval != 0)
965 		goto out;
966 
967 	*dqm->fence_addr = KFD_FENCE_INIT;
968 	pm_send_query_status(&dqm->packets, dqm->fence_gpu_addr,
969 				KFD_FENCE_COMPLETED);
970 	/* should be timed out */
971 	amdkfd_fence_wait_timeout(dqm->fence_addr, KFD_FENCE_COMPLETED,
972 				QUEUE_PREEMPT_DEFAULT_TIMEOUT_MS);
973 	pm_release_ib(&dqm->packets);
974 	dqm->active_runlist = false;
975 
976 out:
977 	if (lock)
978 		mutex_unlock(&dqm->lock);
979 	return retval;
980 }
981 
982 static int execute_queues_cpsch(struct device_queue_manager *dqm, bool lock)
983 {
984 	int retval;
985 
986 	BUG_ON(!dqm);
987 
988 	if (lock)
989 		mutex_lock(&dqm->lock);
990 
991 	retval = destroy_queues_cpsch(dqm, false);
992 	if (retval != 0) {
993 		pr_err("kfd: the cp might be in an unrecoverable state due to an unsuccessful queues preemption");
994 		goto out;
995 	}
996 
997 	if (dqm->queue_count <= 0 || dqm->processes_count <= 0) {
998 		retval = 0;
999 		goto out;
1000 	}
1001 
1002 	if (dqm->active_runlist) {
1003 		retval = 0;
1004 		goto out;
1005 	}
1006 
1007 	retval = pm_send_runlist(&dqm->packets, &dqm->queues);
1008 	if (retval != 0) {
1009 		pr_err("kfd: failed to execute runlist");
1010 		goto out;
1011 	}
1012 	dqm->active_runlist = true;
1013 
1014 out:
1015 	if (lock)
1016 		mutex_unlock(&dqm->lock);
1017 	return retval;
1018 }
1019 
1020 static int destroy_queue_cpsch(struct device_queue_manager *dqm,
1021 				struct qcm_process_device *qpd,
1022 				struct queue *q)
1023 {
1024 	int retval;
1025 	struct mqd_manager *mqd;
1026 
1027 	BUG_ON(!dqm || !qpd || !q);
1028 
1029 	retval = 0;
1030 
1031 	/* remove queue from list to prevent rescheduling after preemption */
1032 	mutex_lock(&dqm->lock);
1033 	mqd = dqm->ops.get_mqd_manager(dqm,
1034 			get_mqd_type_from_queue_type(q->properties.type));
1035 	if (!mqd) {
1036 		retval = -ENOMEM;
1037 		goto failed;
1038 	}
1039 
1040 	if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
1041 		dqm->sdma_queue_count--;
1042 
1043 	list_del(&q->list);
1044 	if (q->properties.is_active)
1045 		dqm->queue_count--;
1046 
1047 	execute_queues_cpsch(dqm, false);
1048 
1049 	mqd->uninit_mqd(mqd, q->mqd, q->mqd_mem_obj);
1050 
1051 	/*
1052 	 * Unconditionally decrement this counter, regardless of the queue's
1053 	 * type
1054 	 */
1055 	dqm->total_queue_count--;
1056 	pr_debug("Total of %d queues are accountable so far\n",
1057 			dqm->total_queue_count);
1058 
1059 	mutex_unlock(&dqm->lock);
1060 
1061 	return 0;
1062 
1063 failed:
1064 	mutex_unlock(&dqm->lock);
1065 	return retval;
1066 }
1067 
1068 /*
1069  * Low bits must be 0000/FFFF as required by HW, high bits must be 0 to
1070  * stay in user mode.
1071  */
1072 #define APE1_FIXED_BITS_MASK 0xFFFF80000000FFFFULL
1073 /* APE1 limit is inclusive and 64K aligned. */
1074 #define APE1_LIMIT_ALIGNMENT 0xFFFF
1075 
1076 static bool set_cache_memory_policy(struct device_queue_manager *dqm,
1077 				   struct qcm_process_device *qpd,
1078 				   enum cache_policy default_policy,
1079 				   enum cache_policy alternate_policy,
1080 				   void __user *alternate_aperture_base,
1081 				   uint64_t alternate_aperture_size)
1082 {
1083 	bool retval;
1084 
1085 	pr_debug("kfd: In func %s\n", __func__);
1086 
1087 	mutex_lock(&dqm->lock);
1088 
1089 	if (alternate_aperture_size == 0) {
1090 		/* base > limit disables APE1 */
1091 		qpd->sh_mem_ape1_base = 1;
1092 		qpd->sh_mem_ape1_limit = 0;
1093 	} else {
1094 		/*
1095 		 * In FSA64, APE1_Base[63:0] = { 16{SH_MEM_APE1_BASE[31]},
1096 		 *			SH_MEM_APE1_BASE[31:0], 0x0000 }
1097 		 * APE1_Limit[63:0] = { 16{SH_MEM_APE1_LIMIT[31]},
1098 		 *			SH_MEM_APE1_LIMIT[31:0], 0xFFFF }
1099 		 * Verify that the base and size parameters can be
1100 		 * represented in this format and convert them.
1101 		 * Additionally restrict APE1 to user-mode addresses.
1102 		 */
1103 
1104 		uint64_t base = (uintptr_t)alternate_aperture_base;
1105 		uint64_t limit = base + alternate_aperture_size - 1;
1106 
1107 		if (limit <= base)
1108 			goto out;
1109 
1110 		if ((base & APE1_FIXED_BITS_MASK) != 0)
1111 			goto out;
1112 
1113 		if ((limit & APE1_FIXED_BITS_MASK) != APE1_LIMIT_ALIGNMENT)
1114 			goto out;
1115 
1116 		qpd->sh_mem_ape1_base = base >> 16;
1117 		qpd->sh_mem_ape1_limit = limit >> 16;
1118 	}
1119 
1120 	retval = dqm->ops_asic_specific.set_cache_memory_policy(
1121 			dqm,
1122 			qpd,
1123 			default_policy,
1124 			alternate_policy,
1125 			alternate_aperture_base,
1126 			alternate_aperture_size);
1127 
1128 	if ((sched_policy == KFD_SCHED_POLICY_NO_HWS) && (qpd->vmid != 0))
1129 		program_sh_mem_settings(dqm, qpd);
1130 
1131 	pr_debug("kfd: sh_mem_config: 0x%x, ape1_base: 0x%x, ape1_limit: 0x%x\n",
1132 		qpd->sh_mem_config, qpd->sh_mem_ape1_base,
1133 		qpd->sh_mem_ape1_limit);
1134 
1135 	mutex_unlock(&dqm->lock);
1136 	return retval;
1137 
1138 out:
1139 	mutex_unlock(&dqm->lock);
1140 	return false;
1141 }
1142 
1143 struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev)
1144 {
1145 	struct device_queue_manager *dqm;
1146 
1147 	BUG_ON(!dev);
1148 
1149 	pr_debug("kfd: loading device queue manager\n");
1150 
1151 	dqm = kzalloc(sizeof(struct device_queue_manager), GFP_KERNEL);
1152 	if (!dqm)
1153 		return NULL;
1154 
1155 	dqm->dev = dev;
1156 	switch (sched_policy) {
1157 	case KFD_SCHED_POLICY_HWS:
1158 	case KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION:
1159 		/* initialize dqm for cp scheduling */
1160 		dqm->ops.create_queue = create_queue_cpsch;
1161 		dqm->ops.initialize = initialize_cpsch;
1162 		dqm->ops.start = start_cpsch;
1163 		dqm->ops.stop = stop_cpsch;
1164 		dqm->ops.destroy_queue = destroy_queue_cpsch;
1165 		dqm->ops.update_queue = update_queue;
1166 		dqm->ops.get_mqd_manager = get_mqd_manager_nocpsch;
1167 		dqm->ops.register_process = register_process_nocpsch;
1168 		dqm->ops.unregister_process = unregister_process_nocpsch;
1169 		dqm->ops.uninitialize = uninitialize_nocpsch;
1170 		dqm->ops.create_kernel_queue = create_kernel_queue_cpsch;
1171 		dqm->ops.destroy_kernel_queue = destroy_kernel_queue_cpsch;
1172 		dqm->ops.set_cache_memory_policy = set_cache_memory_policy;
1173 		break;
1174 	case KFD_SCHED_POLICY_NO_HWS:
1175 		/* initialize dqm for no cp scheduling */
1176 		dqm->ops.start = start_nocpsch;
1177 		dqm->ops.stop = stop_nocpsch;
1178 		dqm->ops.create_queue = create_queue_nocpsch;
1179 		dqm->ops.destroy_queue = destroy_queue_nocpsch;
1180 		dqm->ops.update_queue = update_queue;
1181 		dqm->ops.get_mqd_manager = get_mqd_manager_nocpsch;
1182 		dqm->ops.register_process = register_process_nocpsch;
1183 		dqm->ops.unregister_process = unregister_process_nocpsch;
1184 		dqm->ops.initialize = initialize_nocpsch;
1185 		dqm->ops.uninitialize = uninitialize_nocpsch;
1186 		dqm->ops.set_cache_memory_policy = set_cache_memory_policy;
1187 		break;
1188 	default:
1189 		BUG();
1190 		break;
1191 	}
1192 
1193 	switch (dev->device_info->asic_family) {
1194 	case CHIP_CARRIZO:
1195 		device_queue_manager_init_vi(&dqm->ops_asic_specific);
1196 		break;
1197 
1198 	case CHIP_KAVERI:
1199 		device_queue_manager_init_cik(&dqm->ops_asic_specific);
1200 		break;
1201 	}
1202 
1203 	if (dqm->ops.initialize(dqm) != 0) {
1204 		kfree(dqm);
1205 		return NULL;
1206 	}
1207 
1208 	return dqm;
1209 }
1210 
1211 void device_queue_manager_uninit(struct device_queue_manager *dqm)
1212 {
1213 	BUG_ON(!dqm);
1214 
1215 	dqm->ops.uninitialize(dqm);
1216 	kfree(dqm);
1217 }
1218