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