xref: /openbmc/linux/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c (revision f79e4d5f92a129a1159c973735007d4ddc8541f3)
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/ratelimit.h>
25 #include <linux/printk.h>
26 #include <linux/slab.h>
27 #include <linux/list.h>
28 #include <linux/types.h>
29 #include <linux/bitops.h>
30 #include <linux/sched.h>
31 #include "kfd_priv.h"
32 #include "kfd_device_queue_manager.h"
33 #include "kfd_mqd_manager.h"
34 #include "cik_regs.h"
35 #include "kfd_kernel_queue.h"
36 
37 /* Size of the per-pipe EOP queue */
38 #define CIK_HPD_EOP_BYTES_LOG2 11
39 #define CIK_HPD_EOP_BYTES (1U << CIK_HPD_EOP_BYTES_LOG2)
40 
41 static int set_pasid_vmid_mapping(struct device_queue_manager *dqm,
42 					unsigned int pasid, unsigned int vmid);
43 
44 static int create_compute_queue_nocpsch(struct device_queue_manager *dqm,
45 					struct queue *q,
46 					struct qcm_process_device *qpd);
47 
48 static int execute_queues_cpsch(struct device_queue_manager *dqm,
49 				enum kfd_unmap_queues_filter filter,
50 				uint32_t filter_param);
51 static int unmap_queues_cpsch(struct device_queue_manager *dqm,
52 				enum kfd_unmap_queues_filter filter,
53 				uint32_t filter_param);
54 
55 static int map_queues_cpsch(struct device_queue_manager *dqm);
56 
57 static int create_sdma_queue_nocpsch(struct device_queue_manager *dqm,
58 					struct queue *q,
59 					struct qcm_process_device *qpd);
60 
61 static void deallocate_sdma_queue(struct device_queue_manager *dqm,
62 				unsigned int sdma_queue_id);
63 
64 static inline
65 enum KFD_MQD_TYPE get_mqd_type_from_queue_type(enum kfd_queue_type type)
66 {
67 	if (type == KFD_QUEUE_TYPE_SDMA)
68 		return KFD_MQD_TYPE_SDMA;
69 	return KFD_MQD_TYPE_CP;
70 }
71 
72 static bool is_pipe_enabled(struct device_queue_manager *dqm, int mec, int pipe)
73 {
74 	int i;
75 	int pipe_offset = mec * dqm->dev->shared_resources.num_pipe_per_mec
76 		+ pipe * dqm->dev->shared_resources.num_queue_per_pipe;
77 
78 	/* queue is available for KFD usage if bit is 1 */
79 	for (i = 0; i <  dqm->dev->shared_resources.num_queue_per_pipe; ++i)
80 		if (test_bit(pipe_offset + i,
81 			      dqm->dev->shared_resources.queue_bitmap))
82 			return true;
83 	return false;
84 }
85 
86 unsigned int get_queues_num(struct device_queue_manager *dqm)
87 {
88 	return bitmap_weight(dqm->dev->shared_resources.queue_bitmap,
89 				KGD_MAX_QUEUES);
90 }
91 
92 unsigned int get_queues_per_pipe(struct device_queue_manager *dqm)
93 {
94 	return dqm->dev->shared_resources.num_queue_per_pipe;
95 }
96 
97 unsigned int get_pipes_per_mec(struct device_queue_manager *dqm)
98 {
99 	return dqm->dev->shared_resources.num_pipe_per_mec;
100 }
101 
102 void program_sh_mem_settings(struct device_queue_manager *dqm,
103 					struct qcm_process_device *qpd)
104 {
105 	return dqm->dev->kfd2kgd->program_sh_mem_settings(
106 						dqm->dev->kgd, qpd->vmid,
107 						qpd->sh_mem_config,
108 						qpd->sh_mem_ape1_base,
109 						qpd->sh_mem_ape1_limit,
110 						qpd->sh_mem_bases);
111 }
112 
113 static int allocate_doorbell(struct qcm_process_device *qpd, struct queue *q)
114 {
115 	struct kfd_dev *dev = qpd->dqm->dev;
116 
117 	if (!KFD_IS_SOC15(dev->device_info->asic_family)) {
118 		/* On pre-SOC15 chips we need to use the queue ID to
119 		 * preserve the user mode ABI.
120 		 */
121 		q->doorbell_id = q->properties.queue_id;
122 	} else if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
123 		/* For SDMA queues on SOC15, use static doorbell
124 		 * assignments based on the engine and queue.
125 		 */
126 		q->doorbell_id = dev->shared_resources.sdma_doorbell
127 			[q->properties.sdma_engine_id]
128 			[q->properties.sdma_queue_id];
129 	} else {
130 		/* For CP queues on SOC15 reserve a free doorbell ID */
131 		unsigned int found;
132 
133 		found = find_first_zero_bit(qpd->doorbell_bitmap,
134 					    KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);
135 		if (found >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) {
136 			pr_debug("No doorbells available");
137 			return -EBUSY;
138 		}
139 		set_bit(found, qpd->doorbell_bitmap);
140 		q->doorbell_id = found;
141 	}
142 
143 	q->properties.doorbell_off =
144 		kfd_doorbell_id_to_offset(dev, q->process,
145 					  q->doorbell_id);
146 
147 	return 0;
148 }
149 
150 static void deallocate_doorbell(struct qcm_process_device *qpd,
151 				struct queue *q)
152 {
153 	unsigned int old;
154 	struct kfd_dev *dev = qpd->dqm->dev;
155 
156 	if (!KFD_IS_SOC15(dev->device_info->asic_family) ||
157 	    q->properties.type == KFD_QUEUE_TYPE_SDMA)
158 		return;
159 
160 	old = test_and_clear_bit(q->doorbell_id, qpd->doorbell_bitmap);
161 	WARN_ON(!old);
162 }
163 
164 static int allocate_vmid(struct device_queue_manager *dqm,
165 			struct qcm_process_device *qpd,
166 			struct queue *q)
167 {
168 	int bit, allocated_vmid;
169 
170 	if (dqm->vmid_bitmap == 0)
171 		return -ENOMEM;
172 
173 	bit = ffs(dqm->vmid_bitmap) - 1;
174 	dqm->vmid_bitmap &= ~(1 << bit);
175 
176 	allocated_vmid = bit + dqm->dev->vm_info.first_vmid_kfd;
177 	pr_debug("vmid allocation %d\n", allocated_vmid);
178 	qpd->vmid = allocated_vmid;
179 	q->properties.vmid = allocated_vmid;
180 
181 	set_pasid_vmid_mapping(dqm, q->process->pasid, q->properties.vmid);
182 	program_sh_mem_settings(dqm, qpd);
183 
184 	/* qpd->page_table_base is set earlier when register_process()
185 	 * is called, i.e. when the first queue is created.
186 	 */
187 	dqm->dev->kfd2kgd->set_vm_context_page_table_base(dqm->dev->kgd,
188 			qpd->vmid,
189 			qpd->page_table_base);
190 	/* invalidate the VM context after pasid and vmid mapping is set up */
191 	kfd_flush_tlb(qpd_to_pdd(qpd));
192 
193 	return 0;
194 }
195 
196 static int flush_texture_cache_nocpsch(struct kfd_dev *kdev,
197 				struct qcm_process_device *qpd)
198 {
199 	const struct packet_manager_funcs *pmf = qpd->dqm->packets.pmf;
200 	int ret;
201 
202 	if (!qpd->ib_kaddr)
203 		return -ENOMEM;
204 
205 	ret = pmf->release_mem(qpd->ib_base, (uint32_t *)qpd->ib_kaddr);
206 	if (ret)
207 		return ret;
208 
209 	return kdev->kfd2kgd->submit_ib(kdev->kgd, KGD_ENGINE_MEC1, qpd->vmid,
210 				qpd->ib_base, (uint32_t *)qpd->ib_kaddr,
211 				pmf->release_mem_size / sizeof(uint32_t));
212 }
213 
214 static void deallocate_vmid(struct device_queue_manager *dqm,
215 				struct qcm_process_device *qpd,
216 				struct queue *q)
217 {
218 	int bit = qpd->vmid - dqm->dev->vm_info.first_vmid_kfd;
219 
220 	/* On GFX v7, CP doesn't flush TC at dequeue */
221 	if (q->device->device_info->asic_family == CHIP_HAWAII)
222 		if (flush_texture_cache_nocpsch(q->device, qpd))
223 			pr_err("Failed to flush TC\n");
224 
225 	kfd_flush_tlb(qpd_to_pdd(qpd));
226 
227 	/* Release the vmid mapping */
228 	set_pasid_vmid_mapping(dqm, 0, qpd->vmid);
229 
230 	dqm->vmid_bitmap |= (1 << bit);
231 	qpd->vmid = 0;
232 	q->properties.vmid = 0;
233 }
234 
235 static int create_queue_nocpsch(struct device_queue_manager *dqm,
236 				struct queue *q,
237 				struct qcm_process_device *qpd)
238 {
239 	int retval;
240 
241 	print_queue(q);
242 
243 	mutex_lock(&dqm->lock);
244 
245 	if (dqm->total_queue_count >= max_num_of_queues_per_device) {
246 		pr_warn("Can't create new usermode queue because %d queues were already created\n",
247 				dqm->total_queue_count);
248 		retval = -EPERM;
249 		goto out_unlock;
250 	}
251 
252 	if (list_empty(&qpd->queues_list)) {
253 		retval = allocate_vmid(dqm, qpd, q);
254 		if (retval)
255 			goto out_unlock;
256 	}
257 	q->properties.vmid = qpd->vmid;
258 	/*
259 	 * Eviction state logic: we only mark active queues as evicted
260 	 * to avoid the overhead of restoring inactive queues later
261 	 */
262 	if (qpd->evicted)
263 		q->properties.is_evicted = (q->properties.queue_size > 0 &&
264 					    q->properties.queue_percent > 0 &&
265 					    q->properties.queue_address != 0);
266 
267 	q->properties.tba_addr = qpd->tba_addr;
268 	q->properties.tma_addr = qpd->tma_addr;
269 
270 	if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE)
271 		retval = create_compute_queue_nocpsch(dqm, q, qpd);
272 	else if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
273 		retval = create_sdma_queue_nocpsch(dqm, q, qpd);
274 	else
275 		retval = -EINVAL;
276 
277 	if (retval) {
278 		if (list_empty(&qpd->queues_list))
279 			deallocate_vmid(dqm, qpd, q);
280 		goto out_unlock;
281 	}
282 
283 	list_add(&q->list, &qpd->queues_list);
284 	qpd->queue_count++;
285 	if (q->properties.is_active)
286 		dqm->queue_count++;
287 
288 	if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
289 		dqm->sdma_queue_count++;
290 
291 	/*
292 	 * Unconditionally increment this counter, regardless of the queue's
293 	 * type or whether the queue is active.
294 	 */
295 	dqm->total_queue_count++;
296 	pr_debug("Total of %d queues are accountable so far\n",
297 			dqm->total_queue_count);
298 
299 out_unlock:
300 	mutex_unlock(&dqm->lock);
301 	return retval;
302 }
303 
304 static int allocate_hqd(struct device_queue_manager *dqm, struct queue *q)
305 {
306 	bool set;
307 	int pipe, bit, i;
308 
309 	set = false;
310 
311 	for (pipe = dqm->next_pipe_to_allocate, i = 0;
312 			i < get_pipes_per_mec(dqm);
313 			pipe = ((pipe + 1) % get_pipes_per_mec(dqm)), ++i) {
314 
315 		if (!is_pipe_enabled(dqm, 0, pipe))
316 			continue;
317 
318 		if (dqm->allocated_queues[pipe] != 0) {
319 			bit = ffs(dqm->allocated_queues[pipe]) - 1;
320 			dqm->allocated_queues[pipe] &= ~(1 << bit);
321 			q->pipe = pipe;
322 			q->queue = bit;
323 			set = true;
324 			break;
325 		}
326 	}
327 
328 	if (!set)
329 		return -EBUSY;
330 
331 	pr_debug("hqd slot - pipe %d, queue %d\n", q->pipe, q->queue);
332 	/* horizontal hqd allocation */
333 	dqm->next_pipe_to_allocate = (pipe + 1) % get_pipes_per_mec(dqm);
334 
335 	return 0;
336 }
337 
338 static inline void deallocate_hqd(struct device_queue_manager *dqm,
339 				struct queue *q)
340 {
341 	dqm->allocated_queues[q->pipe] |= (1 << q->queue);
342 }
343 
344 static int create_compute_queue_nocpsch(struct device_queue_manager *dqm,
345 					struct queue *q,
346 					struct qcm_process_device *qpd)
347 {
348 	int retval;
349 	struct mqd_manager *mqd;
350 
351 	mqd = dqm->ops.get_mqd_manager(dqm, KFD_MQD_TYPE_COMPUTE);
352 	if (!mqd)
353 		return -ENOMEM;
354 
355 	retval = allocate_hqd(dqm, q);
356 	if (retval)
357 		return retval;
358 
359 	retval = allocate_doorbell(qpd, q);
360 	if (retval)
361 		goto out_deallocate_hqd;
362 
363 	retval = mqd->init_mqd(mqd, &q->mqd, &q->mqd_mem_obj,
364 				&q->gart_mqd_addr, &q->properties);
365 	if (retval)
366 		goto out_deallocate_doorbell;
367 
368 	pr_debug("Loading mqd to hqd on pipe %d, queue %d\n",
369 			q->pipe, q->queue);
370 
371 	dqm->dev->kfd2kgd->set_scratch_backing_va(
372 			dqm->dev->kgd, qpd->sh_hidden_private_base, qpd->vmid);
373 
374 	if (!q->properties.is_active)
375 		return 0;
376 
377 	retval = mqd->load_mqd(mqd, q->mqd, q->pipe, q->queue, &q->properties,
378 			       q->process->mm);
379 	if (retval)
380 		goto out_uninit_mqd;
381 
382 	return 0;
383 
384 out_uninit_mqd:
385 	mqd->uninit_mqd(mqd, q->mqd, q->mqd_mem_obj);
386 out_deallocate_doorbell:
387 	deallocate_doorbell(qpd, q);
388 out_deallocate_hqd:
389 	deallocate_hqd(dqm, q);
390 
391 	return retval;
392 }
393 
394 /* Access to DQM has to be locked before calling destroy_queue_nocpsch_locked
395  * to avoid asynchronized access
396  */
397 static int destroy_queue_nocpsch_locked(struct device_queue_manager *dqm,
398 				struct qcm_process_device *qpd,
399 				struct queue *q)
400 {
401 	int retval;
402 	struct mqd_manager *mqd;
403 
404 	mqd = dqm->ops.get_mqd_manager(dqm,
405 		get_mqd_type_from_queue_type(q->properties.type));
406 	if (!mqd)
407 		return -ENOMEM;
408 
409 	if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE) {
410 		deallocate_hqd(dqm, q);
411 	} else if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
412 		dqm->sdma_queue_count--;
413 		deallocate_sdma_queue(dqm, q->sdma_id);
414 	} else {
415 		pr_debug("q->properties.type %d is invalid\n",
416 				q->properties.type);
417 		return -EINVAL;
418 	}
419 	dqm->total_queue_count--;
420 
421 	deallocate_doorbell(qpd, q);
422 
423 	retval = mqd->destroy_mqd(mqd, q->mqd,
424 				KFD_PREEMPT_TYPE_WAVEFRONT_RESET,
425 				KFD_UNMAP_LATENCY_MS,
426 				q->pipe, q->queue);
427 	if (retval == -ETIME)
428 		qpd->reset_wavefronts = true;
429 
430 	mqd->uninit_mqd(mqd, q->mqd, q->mqd_mem_obj);
431 
432 	list_del(&q->list);
433 	if (list_empty(&qpd->queues_list)) {
434 		if (qpd->reset_wavefronts) {
435 			pr_warn("Resetting wave fronts (nocpsch) on dev %p\n",
436 					dqm->dev);
437 			/* dbgdev_wave_reset_wavefronts has to be called before
438 			 * deallocate_vmid(), i.e. when vmid is still in use.
439 			 */
440 			dbgdev_wave_reset_wavefronts(dqm->dev,
441 					qpd->pqm->process);
442 			qpd->reset_wavefronts = false;
443 		}
444 
445 		deallocate_vmid(dqm, qpd, q);
446 	}
447 	qpd->queue_count--;
448 	if (q->properties.is_active)
449 		dqm->queue_count--;
450 
451 	return retval;
452 }
453 
454 static int destroy_queue_nocpsch(struct device_queue_manager *dqm,
455 				struct qcm_process_device *qpd,
456 				struct queue *q)
457 {
458 	int retval;
459 
460 	mutex_lock(&dqm->lock);
461 	retval = destroy_queue_nocpsch_locked(dqm, qpd, q);
462 	mutex_unlock(&dqm->lock);
463 
464 	return retval;
465 }
466 
467 static int update_queue(struct device_queue_manager *dqm, struct queue *q)
468 {
469 	int retval;
470 	struct mqd_manager *mqd;
471 	struct kfd_process_device *pdd;
472 	bool prev_active = false;
473 
474 	mutex_lock(&dqm->lock);
475 	pdd = kfd_get_process_device_data(q->device, q->process);
476 	if (!pdd) {
477 		retval = -ENODEV;
478 		goto out_unlock;
479 	}
480 	mqd = dqm->ops.get_mqd_manager(dqm,
481 			get_mqd_type_from_queue_type(q->properties.type));
482 	if (!mqd) {
483 		retval = -ENOMEM;
484 		goto out_unlock;
485 	}
486 	/*
487 	 * Eviction state logic: we only mark active queues as evicted
488 	 * to avoid the overhead of restoring inactive queues later
489 	 */
490 	if (pdd->qpd.evicted)
491 		q->properties.is_evicted = (q->properties.queue_size > 0 &&
492 					    q->properties.queue_percent > 0 &&
493 					    q->properties.queue_address != 0);
494 
495 	/* Save previous activity state for counters */
496 	prev_active = q->properties.is_active;
497 
498 	/* Make sure the queue is unmapped before updating the MQD */
499 	if (dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) {
500 		retval = unmap_queues_cpsch(dqm,
501 				KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
502 		if (retval) {
503 			pr_err("unmap queue failed\n");
504 			goto out_unlock;
505 		}
506 	} else if (prev_active &&
507 		   (q->properties.type == KFD_QUEUE_TYPE_COMPUTE ||
508 		    q->properties.type == KFD_QUEUE_TYPE_SDMA)) {
509 		retval = mqd->destroy_mqd(mqd, q->mqd,
510 				KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN,
511 				KFD_UNMAP_LATENCY_MS, q->pipe, q->queue);
512 		if (retval) {
513 			pr_err("destroy mqd failed\n");
514 			goto out_unlock;
515 		}
516 	}
517 
518 	retval = mqd->update_mqd(mqd, q->mqd, &q->properties);
519 
520 	/*
521 	 * check active state vs. the previous state and modify
522 	 * counter accordingly. map_queues_cpsch uses the
523 	 * dqm->queue_count to determine whether a new runlist must be
524 	 * uploaded.
525 	 */
526 	if (q->properties.is_active && !prev_active)
527 		dqm->queue_count++;
528 	else if (!q->properties.is_active && prev_active)
529 		dqm->queue_count--;
530 
531 	if (dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS)
532 		retval = map_queues_cpsch(dqm);
533 	else if (q->properties.is_active &&
534 		 (q->properties.type == KFD_QUEUE_TYPE_COMPUTE ||
535 		  q->properties.type == KFD_QUEUE_TYPE_SDMA))
536 		retval = mqd->load_mqd(mqd, q->mqd, q->pipe, q->queue,
537 				       &q->properties, q->process->mm);
538 
539 out_unlock:
540 	mutex_unlock(&dqm->lock);
541 	return retval;
542 }
543 
544 static struct mqd_manager *get_mqd_manager(
545 		struct device_queue_manager *dqm, enum KFD_MQD_TYPE type)
546 {
547 	struct mqd_manager *mqd;
548 
549 	if (WARN_ON(type >= KFD_MQD_TYPE_MAX))
550 		return NULL;
551 
552 	pr_debug("mqd type %d\n", type);
553 
554 	mqd = dqm->mqds[type];
555 	if (!mqd) {
556 		mqd = mqd_manager_init(type, dqm->dev);
557 		if (!mqd)
558 			pr_err("mqd manager is NULL");
559 		dqm->mqds[type] = mqd;
560 	}
561 
562 	return mqd;
563 }
564 
565 static int evict_process_queues_nocpsch(struct device_queue_manager *dqm,
566 					struct qcm_process_device *qpd)
567 {
568 	struct queue *q;
569 	struct mqd_manager *mqd;
570 	struct kfd_process_device *pdd;
571 	int retval = 0;
572 
573 	mutex_lock(&dqm->lock);
574 	if (qpd->evicted++ > 0) /* already evicted, do nothing */
575 		goto out;
576 
577 	pdd = qpd_to_pdd(qpd);
578 	pr_info_ratelimited("Evicting PASID %u queues\n",
579 			    pdd->process->pasid);
580 
581 	/* unactivate all active queues on the qpd */
582 	list_for_each_entry(q, &qpd->queues_list, list) {
583 		if (!q->properties.is_active)
584 			continue;
585 		mqd = dqm->ops.get_mqd_manager(dqm,
586 			get_mqd_type_from_queue_type(q->properties.type));
587 		if (!mqd) { /* should not be here */
588 			pr_err("Cannot evict queue, mqd mgr is NULL\n");
589 			retval = -ENOMEM;
590 			goto out;
591 		}
592 		q->properties.is_evicted = true;
593 		q->properties.is_active = false;
594 		retval = mqd->destroy_mqd(mqd, q->mqd,
595 				KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN,
596 				KFD_UNMAP_LATENCY_MS, q->pipe, q->queue);
597 		if (retval)
598 			goto out;
599 		dqm->queue_count--;
600 	}
601 
602 out:
603 	mutex_unlock(&dqm->lock);
604 	return retval;
605 }
606 
607 static int evict_process_queues_cpsch(struct device_queue_manager *dqm,
608 				      struct qcm_process_device *qpd)
609 {
610 	struct queue *q;
611 	struct kfd_process_device *pdd;
612 	int retval = 0;
613 
614 	mutex_lock(&dqm->lock);
615 	if (qpd->evicted++ > 0) /* already evicted, do nothing */
616 		goto out;
617 
618 	pdd = qpd_to_pdd(qpd);
619 	pr_info_ratelimited("Evicting PASID %u queues\n",
620 			    pdd->process->pasid);
621 
622 	/* unactivate all active queues on the qpd */
623 	list_for_each_entry(q, &qpd->queues_list, list) {
624 		if (!q->properties.is_active)
625 			continue;
626 		q->properties.is_evicted = true;
627 		q->properties.is_active = false;
628 		dqm->queue_count--;
629 	}
630 	retval = execute_queues_cpsch(dqm,
631 				qpd->is_debug ?
632 				KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES :
633 				KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
634 
635 out:
636 	mutex_unlock(&dqm->lock);
637 	return retval;
638 }
639 
640 static int restore_process_queues_nocpsch(struct device_queue_manager *dqm,
641 					  struct qcm_process_device *qpd)
642 {
643 	struct queue *q;
644 	struct mqd_manager *mqd;
645 	struct kfd_process_device *pdd;
646 	uint32_t pd_base;
647 	int retval = 0;
648 
649 	pdd = qpd_to_pdd(qpd);
650 	/* Retrieve PD base */
651 	pd_base = dqm->dev->kfd2kgd->get_process_page_dir(pdd->vm);
652 
653 	mutex_lock(&dqm->lock);
654 	if (WARN_ON_ONCE(!qpd->evicted)) /* already restored, do nothing */
655 		goto out;
656 	if (qpd->evicted > 1) { /* ref count still > 0, decrement & quit */
657 		qpd->evicted--;
658 		goto out;
659 	}
660 
661 	pr_info_ratelimited("Restoring PASID %u queues\n",
662 			    pdd->process->pasid);
663 
664 	/* Update PD Base in QPD */
665 	qpd->page_table_base = pd_base;
666 	pr_debug("Updated PD address to 0x%08x\n", pd_base);
667 
668 	if (!list_empty(&qpd->queues_list)) {
669 		dqm->dev->kfd2kgd->set_vm_context_page_table_base(
670 				dqm->dev->kgd,
671 				qpd->vmid,
672 				qpd->page_table_base);
673 		kfd_flush_tlb(pdd);
674 	}
675 
676 	/* activate all active queues on the qpd */
677 	list_for_each_entry(q, &qpd->queues_list, list) {
678 		if (!q->properties.is_evicted)
679 			continue;
680 		mqd = dqm->ops.get_mqd_manager(dqm,
681 			get_mqd_type_from_queue_type(q->properties.type));
682 		if (!mqd) { /* should not be here */
683 			pr_err("Cannot restore queue, mqd mgr is NULL\n");
684 			retval = -ENOMEM;
685 			goto out;
686 		}
687 		q->properties.is_evicted = false;
688 		q->properties.is_active = true;
689 		retval = mqd->load_mqd(mqd, q->mqd, q->pipe,
690 				       q->queue, &q->properties,
691 				       q->process->mm);
692 		if (retval)
693 			goto out;
694 		dqm->queue_count++;
695 	}
696 	qpd->evicted = 0;
697 out:
698 	mutex_unlock(&dqm->lock);
699 	return retval;
700 }
701 
702 static int restore_process_queues_cpsch(struct device_queue_manager *dqm,
703 					struct qcm_process_device *qpd)
704 {
705 	struct queue *q;
706 	struct kfd_process_device *pdd;
707 	uint32_t pd_base;
708 	int retval = 0;
709 
710 	pdd = qpd_to_pdd(qpd);
711 	/* Retrieve PD base */
712 	pd_base = dqm->dev->kfd2kgd->get_process_page_dir(pdd->vm);
713 
714 	mutex_lock(&dqm->lock);
715 	if (WARN_ON_ONCE(!qpd->evicted)) /* already restored, do nothing */
716 		goto out;
717 	if (qpd->evicted > 1) { /* ref count still > 0, decrement & quit */
718 		qpd->evicted--;
719 		goto out;
720 	}
721 
722 	pr_info_ratelimited("Restoring PASID %u queues\n",
723 			    pdd->process->pasid);
724 
725 	/* Update PD Base in QPD */
726 	qpd->page_table_base = pd_base;
727 	pr_debug("Updated PD address to 0x%08x\n", pd_base);
728 
729 	/* activate all active queues on the qpd */
730 	list_for_each_entry(q, &qpd->queues_list, list) {
731 		if (!q->properties.is_evicted)
732 			continue;
733 		q->properties.is_evicted = false;
734 		q->properties.is_active = true;
735 		dqm->queue_count++;
736 	}
737 	retval = execute_queues_cpsch(dqm,
738 				KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
739 	if (!retval)
740 		qpd->evicted = 0;
741 out:
742 	mutex_unlock(&dqm->lock);
743 	return retval;
744 }
745 
746 static int register_process(struct device_queue_manager *dqm,
747 					struct qcm_process_device *qpd)
748 {
749 	struct device_process_node *n;
750 	struct kfd_process_device *pdd;
751 	uint32_t pd_base;
752 	int retval;
753 
754 	n = kzalloc(sizeof(*n), GFP_KERNEL);
755 	if (!n)
756 		return -ENOMEM;
757 
758 	n->qpd = qpd;
759 
760 	pdd = qpd_to_pdd(qpd);
761 	/* Retrieve PD base */
762 	pd_base = dqm->dev->kfd2kgd->get_process_page_dir(pdd->vm);
763 
764 	mutex_lock(&dqm->lock);
765 	list_add(&n->list, &dqm->queues);
766 
767 	/* Update PD Base in QPD */
768 	qpd->page_table_base = pd_base;
769 
770 	retval = dqm->asic_ops.update_qpd(dqm, qpd);
771 
772 	dqm->processes_count++;
773 
774 	mutex_unlock(&dqm->lock);
775 
776 	return retval;
777 }
778 
779 static int unregister_process(struct device_queue_manager *dqm,
780 					struct qcm_process_device *qpd)
781 {
782 	int retval;
783 	struct device_process_node *cur, *next;
784 
785 	pr_debug("qpd->queues_list is %s\n",
786 			list_empty(&qpd->queues_list) ? "empty" : "not empty");
787 
788 	retval = 0;
789 	mutex_lock(&dqm->lock);
790 
791 	list_for_each_entry_safe(cur, next, &dqm->queues, list) {
792 		if (qpd == cur->qpd) {
793 			list_del(&cur->list);
794 			kfree(cur);
795 			dqm->processes_count--;
796 			goto out;
797 		}
798 	}
799 	/* qpd not found in dqm list */
800 	retval = 1;
801 out:
802 	mutex_unlock(&dqm->lock);
803 	return retval;
804 }
805 
806 static int
807 set_pasid_vmid_mapping(struct device_queue_manager *dqm, unsigned int pasid,
808 			unsigned int vmid)
809 {
810 	uint32_t pasid_mapping;
811 
812 	pasid_mapping = (pasid == 0) ? 0 :
813 		(uint32_t)pasid |
814 		ATC_VMID_PASID_MAPPING_VALID;
815 
816 	return dqm->dev->kfd2kgd->set_pasid_vmid_mapping(
817 						dqm->dev->kgd, pasid_mapping,
818 						vmid);
819 }
820 
821 static void init_interrupts(struct device_queue_manager *dqm)
822 {
823 	unsigned int i;
824 
825 	for (i = 0 ; i < get_pipes_per_mec(dqm) ; i++)
826 		if (is_pipe_enabled(dqm, 0, i))
827 			dqm->dev->kfd2kgd->init_interrupts(dqm->dev->kgd, i);
828 }
829 
830 static int initialize_nocpsch(struct device_queue_manager *dqm)
831 {
832 	int pipe, queue;
833 
834 	pr_debug("num of pipes: %d\n", get_pipes_per_mec(dqm));
835 
836 	dqm->allocated_queues = kcalloc(get_pipes_per_mec(dqm),
837 					sizeof(unsigned int), GFP_KERNEL);
838 	if (!dqm->allocated_queues)
839 		return -ENOMEM;
840 
841 	mutex_init(&dqm->lock);
842 	INIT_LIST_HEAD(&dqm->queues);
843 	dqm->queue_count = dqm->next_pipe_to_allocate = 0;
844 	dqm->sdma_queue_count = 0;
845 
846 	for (pipe = 0; pipe < get_pipes_per_mec(dqm); pipe++) {
847 		int pipe_offset = pipe * get_queues_per_pipe(dqm);
848 
849 		for (queue = 0; queue < get_queues_per_pipe(dqm); queue++)
850 			if (test_bit(pipe_offset + queue,
851 				     dqm->dev->shared_resources.queue_bitmap))
852 				dqm->allocated_queues[pipe] |= 1 << queue;
853 	}
854 
855 	dqm->vmid_bitmap = (1 << dqm->dev->vm_info.vmid_num_kfd) - 1;
856 	dqm->sdma_bitmap = (1 << CIK_SDMA_QUEUES) - 1;
857 
858 	return 0;
859 }
860 
861 static void uninitialize(struct device_queue_manager *dqm)
862 {
863 	int i;
864 
865 	WARN_ON(dqm->queue_count > 0 || dqm->processes_count > 0);
866 
867 	kfree(dqm->allocated_queues);
868 	for (i = 0 ; i < KFD_MQD_TYPE_MAX ; i++)
869 		kfree(dqm->mqds[i]);
870 	mutex_destroy(&dqm->lock);
871 	kfd_gtt_sa_free(dqm->dev, dqm->pipeline_mem);
872 }
873 
874 static int start_nocpsch(struct device_queue_manager *dqm)
875 {
876 	init_interrupts(dqm);
877 	return pm_init(&dqm->packets, dqm);
878 }
879 
880 static int stop_nocpsch(struct device_queue_manager *dqm)
881 {
882 	pm_uninit(&dqm->packets);
883 	return 0;
884 }
885 
886 static int allocate_sdma_queue(struct device_queue_manager *dqm,
887 				unsigned int *sdma_queue_id)
888 {
889 	int bit;
890 
891 	if (dqm->sdma_bitmap == 0)
892 		return -ENOMEM;
893 
894 	bit = ffs(dqm->sdma_bitmap) - 1;
895 	dqm->sdma_bitmap &= ~(1 << bit);
896 	*sdma_queue_id = bit;
897 
898 	return 0;
899 }
900 
901 static void deallocate_sdma_queue(struct device_queue_manager *dqm,
902 				unsigned int sdma_queue_id)
903 {
904 	if (sdma_queue_id >= CIK_SDMA_QUEUES)
905 		return;
906 	dqm->sdma_bitmap |= (1 << sdma_queue_id);
907 }
908 
909 static int create_sdma_queue_nocpsch(struct device_queue_manager *dqm,
910 					struct queue *q,
911 					struct qcm_process_device *qpd)
912 {
913 	struct mqd_manager *mqd;
914 	int retval;
915 
916 	mqd = dqm->ops.get_mqd_manager(dqm, KFD_MQD_TYPE_SDMA);
917 	if (!mqd)
918 		return -ENOMEM;
919 
920 	retval = allocate_sdma_queue(dqm, &q->sdma_id);
921 	if (retval)
922 		return retval;
923 
924 	q->properties.sdma_queue_id = q->sdma_id / CIK_SDMA_QUEUES_PER_ENGINE;
925 	q->properties.sdma_engine_id = q->sdma_id % CIK_SDMA_QUEUES_PER_ENGINE;
926 
927 	retval = allocate_doorbell(qpd, q);
928 	if (retval)
929 		goto out_deallocate_sdma_queue;
930 
931 	pr_debug("SDMA id is:    %d\n", q->sdma_id);
932 	pr_debug("SDMA queue id: %d\n", q->properties.sdma_queue_id);
933 	pr_debug("SDMA engine id: %d\n", q->properties.sdma_engine_id);
934 
935 	dqm->asic_ops.init_sdma_vm(dqm, q, qpd);
936 	retval = mqd->init_mqd(mqd, &q->mqd, &q->mqd_mem_obj,
937 				&q->gart_mqd_addr, &q->properties);
938 	if (retval)
939 		goto out_deallocate_doorbell;
940 
941 	retval = mqd->load_mqd(mqd, q->mqd, 0, 0, &q->properties, NULL);
942 	if (retval)
943 		goto out_uninit_mqd;
944 
945 	return 0;
946 
947 out_uninit_mqd:
948 	mqd->uninit_mqd(mqd, q->mqd, q->mqd_mem_obj);
949 out_deallocate_doorbell:
950 	deallocate_doorbell(qpd, q);
951 out_deallocate_sdma_queue:
952 	deallocate_sdma_queue(dqm, q->sdma_id);
953 
954 	return retval;
955 }
956 
957 /*
958  * Device Queue Manager implementation for cp scheduler
959  */
960 
961 static int set_sched_resources(struct device_queue_manager *dqm)
962 {
963 	int i, mec;
964 	struct scheduling_resources res;
965 
966 	res.vmid_mask = dqm->dev->shared_resources.compute_vmid_bitmap;
967 
968 	res.queue_mask = 0;
969 	for (i = 0; i < KGD_MAX_QUEUES; ++i) {
970 		mec = (i / dqm->dev->shared_resources.num_queue_per_pipe)
971 			/ dqm->dev->shared_resources.num_pipe_per_mec;
972 
973 		if (!test_bit(i, dqm->dev->shared_resources.queue_bitmap))
974 			continue;
975 
976 		/* only acquire queues from the first MEC */
977 		if (mec > 0)
978 			continue;
979 
980 		/* This situation may be hit in the future if a new HW
981 		 * generation exposes more than 64 queues. If so, the
982 		 * definition of res.queue_mask needs updating
983 		 */
984 		if (WARN_ON(i >= (sizeof(res.queue_mask)*8))) {
985 			pr_err("Invalid queue enabled by amdgpu: %d\n", i);
986 			break;
987 		}
988 
989 		res.queue_mask |= (1ull << i);
990 	}
991 	res.gws_mask = res.oac_mask = res.gds_heap_base =
992 						res.gds_heap_size = 0;
993 
994 	pr_debug("Scheduling resources:\n"
995 			"vmid mask: 0x%8X\n"
996 			"queue mask: 0x%8llX\n",
997 			res.vmid_mask, res.queue_mask);
998 
999 	return pm_send_set_resources(&dqm->packets, &res);
1000 }
1001 
1002 static int initialize_cpsch(struct device_queue_manager *dqm)
1003 {
1004 	pr_debug("num of pipes: %d\n", get_pipes_per_mec(dqm));
1005 
1006 	mutex_init(&dqm->lock);
1007 	INIT_LIST_HEAD(&dqm->queues);
1008 	dqm->queue_count = dqm->processes_count = 0;
1009 	dqm->sdma_queue_count = 0;
1010 	dqm->active_runlist = false;
1011 	dqm->sdma_bitmap = (1 << CIK_SDMA_QUEUES) - 1;
1012 
1013 	return 0;
1014 }
1015 
1016 static int start_cpsch(struct device_queue_manager *dqm)
1017 {
1018 	int retval;
1019 
1020 	retval = 0;
1021 
1022 	retval = pm_init(&dqm->packets, dqm);
1023 	if (retval)
1024 		goto fail_packet_manager_init;
1025 
1026 	retval = set_sched_resources(dqm);
1027 	if (retval)
1028 		goto fail_set_sched_resources;
1029 
1030 	pr_debug("Allocating fence memory\n");
1031 
1032 	/* allocate fence memory on the gart */
1033 	retval = kfd_gtt_sa_allocate(dqm->dev, sizeof(*dqm->fence_addr),
1034 					&dqm->fence_mem);
1035 
1036 	if (retval)
1037 		goto fail_allocate_vidmem;
1038 
1039 	dqm->fence_addr = dqm->fence_mem->cpu_ptr;
1040 	dqm->fence_gpu_addr = dqm->fence_mem->gpu_addr;
1041 
1042 	init_interrupts(dqm);
1043 
1044 	mutex_lock(&dqm->lock);
1045 	execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
1046 	mutex_unlock(&dqm->lock);
1047 
1048 	return 0;
1049 fail_allocate_vidmem:
1050 fail_set_sched_resources:
1051 	pm_uninit(&dqm->packets);
1052 fail_packet_manager_init:
1053 	return retval;
1054 }
1055 
1056 static int stop_cpsch(struct device_queue_manager *dqm)
1057 {
1058 	mutex_lock(&dqm->lock);
1059 	unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0);
1060 	mutex_unlock(&dqm->lock);
1061 
1062 	kfd_gtt_sa_free(dqm->dev, dqm->fence_mem);
1063 	pm_uninit(&dqm->packets);
1064 
1065 	return 0;
1066 }
1067 
1068 static int create_kernel_queue_cpsch(struct device_queue_manager *dqm,
1069 					struct kernel_queue *kq,
1070 					struct qcm_process_device *qpd)
1071 {
1072 	mutex_lock(&dqm->lock);
1073 	if (dqm->total_queue_count >= max_num_of_queues_per_device) {
1074 		pr_warn("Can't create new kernel queue because %d queues were already created\n",
1075 				dqm->total_queue_count);
1076 		mutex_unlock(&dqm->lock);
1077 		return -EPERM;
1078 	}
1079 
1080 	/*
1081 	 * Unconditionally increment this counter, regardless of the queue's
1082 	 * type or whether the queue is active.
1083 	 */
1084 	dqm->total_queue_count++;
1085 	pr_debug("Total of %d queues are accountable so far\n",
1086 			dqm->total_queue_count);
1087 
1088 	list_add(&kq->list, &qpd->priv_queue_list);
1089 	dqm->queue_count++;
1090 	qpd->is_debug = true;
1091 	execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
1092 	mutex_unlock(&dqm->lock);
1093 
1094 	return 0;
1095 }
1096 
1097 static void destroy_kernel_queue_cpsch(struct device_queue_manager *dqm,
1098 					struct kernel_queue *kq,
1099 					struct qcm_process_device *qpd)
1100 {
1101 	mutex_lock(&dqm->lock);
1102 	list_del(&kq->list);
1103 	dqm->queue_count--;
1104 	qpd->is_debug = false;
1105 	execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0);
1106 	/*
1107 	 * Unconditionally decrement this counter, regardless of the queue's
1108 	 * type.
1109 	 */
1110 	dqm->total_queue_count--;
1111 	pr_debug("Total of %d queues are accountable so far\n",
1112 			dqm->total_queue_count);
1113 	mutex_unlock(&dqm->lock);
1114 }
1115 
1116 static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q,
1117 			struct qcm_process_device *qpd)
1118 {
1119 	int retval;
1120 	struct mqd_manager *mqd;
1121 
1122 	retval = 0;
1123 
1124 	mutex_lock(&dqm->lock);
1125 
1126 	if (dqm->total_queue_count >= max_num_of_queues_per_device) {
1127 		pr_warn("Can't create new usermode queue because %d queues were already created\n",
1128 				dqm->total_queue_count);
1129 		retval = -EPERM;
1130 		goto out_unlock;
1131 	}
1132 
1133 	if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
1134 		retval = allocate_sdma_queue(dqm, &q->sdma_id);
1135 		if (retval)
1136 			goto out_unlock;
1137 		q->properties.sdma_queue_id =
1138 			q->sdma_id / CIK_SDMA_QUEUES_PER_ENGINE;
1139 		q->properties.sdma_engine_id =
1140 			q->sdma_id % CIK_SDMA_QUEUES_PER_ENGINE;
1141 	}
1142 
1143 	retval = allocate_doorbell(qpd, q);
1144 	if (retval)
1145 		goto out_deallocate_sdma_queue;
1146 
1147 	mqd = dqm->ops.get_mqd_manager(dqm,
1148 			get_mqd_type_from_queue_type(q->properties.type));
1149 
1150 	if (!mqd) {
1151 		retval = -ENOMEM;
1152 		goto out_deallocate_doorbell;
1153 	}
1154 	/*
1155 	 * Eviction state logic: we only mark active queues as evicted
1156 	 * to avoid the overhead of restoring inactive queues later
1157 	 */
1158 	if (qpd->evicted)
1159 		q->properties.is_evicted = (q->properties.queue_size > 0 &&
1160 					    q->properties.queue_percent > 0 &&
1161 					    q->properties.queue_address != 0);
1162 
1163 	dqm->asic_ops.init_sdma_vm(dqm, q, qpd);
1164 
1165 	q->properties.tba_addr = qpd->tba_addr;
1166 	q->properties.tma_addr = qpd->tma_addr;
1167 	retval = mqd->init_mqd(mqd, &q->mqd, &q->mqd_mem_obj,
1168 				&q->gart_mqd_addr, &q->properties);
1169 	if (retval)
1170 		goto out_deallocate_doorbell;
1171 
1172 	list_add(&q->list, &qpd->queues_list);
1173 	qpd->queue_count++;
1174 	if (q->properties.is_active) {
1175 		dqm->queue_count++;
1176 		retval = execute_queues_cpsch(dqm,
1177 				KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
1178 	}
1179 
1180 	if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
1181 		dqm->sdma_queue_count++;
1182 	/*
1183 	 * Unconditionally increment this counter, regardless of the queue's
1184 	 * type or whether the queue is active.
1185 	 */
1186 	dqm->total_queue_count++;
1187 
1188 	pr_debug("Total of %d queues are accountable so far\n",
1189 			dqm->total_queue_count);
1190 
1191 	mutex_unlock(&dqm->lock);
1192 	return retval;
1193 
1194 out_deallocate_doorbell:
1195 	deallocate_doorbell(qpd, q);
1196 out_deallocate_sdma_queue:
1197 	if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
1198 		deallocate_sdma_queue(dqm, q->sdma_id);
1199 out_unlock:
1200 	mutex_unlock(&dqm->lock);
1201 	return retval;
1202 }
1203 
1204 int amdkfd_fence_wait_timeout(unsigned int *fence_addr,
1205 				unsigned int fence_value,
1206 				unsigned int timeout_ms)
1207 {
1208 	unsigned long end_jiffies = msecs_to_jiffies(timeout_ms) + jiffies;
1209 
1210 	while (*fence_addr != fence_value) {
1211 		if (time_after(jiffies, end_jiffies)) {
1212 			pr_err("qcm fence wait loop timeout expired\n");
1213 			return -ETIME;
1214 		}
1215 		schedule();
1216 	}
1217 
1218 	return 0;
1219 }
1220 
1221 static int unmap_sdma_queues(struct device_queue_manager *dqm,
1222 				unsigned int sdma_engine)
1223 {
1224 	return pm_send_unmap_queue(&dqm->packets, KFD_QUEUE_TYPE_SDMA,
1225 			KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, false,
1226 			sdma_engine);
1227 }
1228 
1229 /* dqm->lock mutex has to be locked before calling this function */
1230 static int map_queues_cpsch(struct device_queue_manager *dqm)
1231 {
1232 	int retval;
1233 
1234 	if (dqm->queue_count <= 0 || dqm->processes_count <= 0)
1235 		return 0;
1236 
1237 	if (dqm->active_runlist)
1238 		return 0;
1239 
1240 	retval = pm_send_runlist(&dqm->packets, &dqm->queues);
1241 	if (retval) {
1242 		pr_err("failed to execute runlist\n");
1243 		return retval;
1244 	}
1245 	dqm->active_runlist = true;
1246 
1247 	return retval;
1248 }
1249 
1250 /* dqm->lock mutex has to be locked before calling this function */
1251 static int unmap_queues_cpsch(struct device_queue_manager *dqm,
1252 				enum kfd_unmap_queues_filter filter,
1253 				uint32_t filter_param)
1254 {
1255 	int retval = 0;
1256 
1257 	if (!dqm->active_runlist)
1258 		return retval;
1259 
1260 	pr_debug("Before destroying queues, sdma queue count is : %u\n",
1261 		dqm->sdma_queue_count);
1262 
1263 	if (dqm->sdma_queue_count > 0) {
1264 		unmap_sdma_queues(dqm, 0);
1265 		unmap_sdma_queues(dqm, 1);
1266 	}
1267 
1268 	retval = pm_send_unmap_queue(&dqm->packets, KFD_QUEUE_TYPE_COMPUTE,
1269 			filter, filter_param, false, 0);
1270 	if (retval)
1271 		return retval;
1272 
1273 	*dqm->fence_addr = KFD_FENCE_INIT;
1274 	pm_send_query_status(&dqm->packets, dqm->fence_gpu_addr,
1275 				KFD_FENCE_COMPLETED);
1276 	/* should be timed out */
1277 	retval = amdkfd_fence_wait_timeout(dqm->fence_addr, KFD_FENCE_COMPLETED,
1278 				QUEUE_PREEMPT_DEFAULT_TIMEOUT_MS);
1279 	if (retval)
1280 		return retval;
1281 
1282 	pm_release_ib(&dqm->packets);
1283 	dqm->active_runlist = false;
1284 
1285 	return retval;
1286 }
1287 
1288 /* dqm->lock mutex has to be locked before calling this function */
1289 static int execute_queues_cpsch(struct device_queue_manager *dqm,
1290 				enum kfd_unmap_queues_filter filter,
1291 				uint32_t filter_param)
1292 {
1293 	int retval;
1294 
1295 	retval = unmap_queues_cpsch(dqm, filter, filter_param);
1296 	if (retval) {
1297 		pr_err("The cp might be in an unrecoverable state due to an unsuccessful queues preemption\n");
1298 		return retval;
1299 	}
1300 
1301 	return map_queues_cpsch(dqm);
1302 }
1303 
1304 static int destroy_queue_cpsch(struct device_queue_manager *dqm,
1305 				struct qcm_process_device *qpd,
1306 				struct queue *q)
1307 {
1308 	int retval;
1309 	struct mqd_manager *mqd;
1310 	bool preempt_all_queues;
1311 
1312 	preempt_all_queues = false;
1313 
1314 	retval = 0;
1315 
1316 	/* remove queue from list to prevent rescheduling after preemption */
1317 	mutex_lock(&dqm->lock);
1318 
1319 	if (qpd->is_debug) {
1320 		/*
1321 		 * error, currently we do not allow to destroy a queue
1322 		 * of a currently debugged process
1323 		 */
1324 		retval = -EBUSY;
1325 		goto failed_try_destroy_debugged_queue;
1326 
1327 	}
1328 
1329 	mqd = dqm->ops.get_mqd_manager(dqm,
1330 			get_mqd_type_from_queue_type(q->properties.type));
1331 	if (!mqd) {
1332 		retval = -ENOMEM;
1333 		goto failed;
1334 	}
1335 
1336 	deallocate_doorbell(qpd, q);
1337 
1338 	if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
1339 		dqm->sdma_queue_count--;
1340 		deallocate_sdma_queue(dqm, q->sdma_id);
1341 	}
1342 
1343 	list_del(&q->list);
1344 	qpd->queue_count--;
1345 	if (q->properties.is_active) {
1346 		dqm->queue_count--;
1347 		retval = execute_queues_cpsch(dqm,
1348 				KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
1349 		if (retval == -ETIME)
1350 			qpd->reset_wavefronts = true;
1351 	}
1352 
1353 	mqd->uninit_mqd(mqd, q->mqd, q->mqd_mem_obj);
1354 
1355 	/*
1356 	 * Unconditionally decrement this counter, regardless of the queue's
1357 	 * type
1358 	 */
1359 	dqm->total_queue_count--;
1360 	pr_debug("Total of %d queues are accountable so far\n",
1361 			dqm->total_queue_count);
1362 
1363 	mutex_unlock(&dqm->lock);
1364 
1365 	return retval;
1366 
1367 failed:
1368 failed_try_destroy_debugged_queue:
1369 
1370 	mutex_unlock(&dqm->lock);
1371 	return retval;
1372 }
1373 
1374 /*
1375  * Low bits must be 0000/FFFF as required by HW, high bits must be 0 to
1376  * stay in user mode.
1377  */
1378 #define APE1_FIXED_BITS_MASK 0xFFFF80000000FFFFULL
1379 /* APE1 limit is inclusive and 64K aligned. */
1380 #define APE1_LIMIT_ALIGNMENT 0xFFFF
1381 
1382 static bool set_cache_memory_policy(struct device_queue_manager *dqm,
1383 				   struct qcm_process_device *qpd,
1384 				   enum cache_policy default_policy,
1385 				   enum cache_policy alternate_policy,
1386 				   void __user *alternate_aperture_base,
1387 				   uint64_t alternate_aperture_size)
1388 {
1389 	bool retval = true;
1390 
1391 	if (!dqm->asic_ops.set_cache_memory_policy)
1392 		return retval;
1393 
1394 	mutex_lock(&dqm->lock);
1395 
1396 	if (alternate_aperture_size == 0) {
1397 		/* base > limit disables APE1 */
1398 		qpd->sh_mem_ape1_base = 1;
1399 		qpd->sh_mem_ape1_limit = 0;
1400 	} else {
1401 		/*
1402 		 * In FSA64, APE1_Base[63:0] = { 16{SH_MEM_APE1_BASE[31]},
1403 		 *			SH_MEM_APE1_BASE[31:0], 0x0000 }
1404 		 * APE1_Limit[63:0] = { 16{SH_MEM_APE1_LIMIT[31]},
1405 		 *			SH_MEM_APE1_LIMIT[31:0], 0xFFFF }
1406 		 * Verify that the base and size parameters can be
1407 		 * represented in this format and convert them.
1408 		 * Additionally restrict APE1 to user-mode addresses.
1409 		 */
1410 
1411 		uint64_t base = (uintptr_t)alternate_aperture_base;
1412 		uint64_t limit = base + alternate_aperture_size - 1;
1413 
1414 		if (limit <= base || (base & APE1_FIXED_BITS_MASK) != 0 ||
1415 		   (limit & APE1_FIXED_BITS_MASK) != APE1_LIMIT_ALIGNMENT) {
1416 			retval = false;
1417 			goto out;
1418 		}
1419 
1420 		qpd->sh_mem_ape1_base = base >> 16;
1421 		qpd->sh_mem_ape1_limit = limit >> 16;
1422 	}
1423 
1424 	retval = dqm->asic_ops.set_cache_memory_policy(
1425 			dqm,
1426 			qpd,
1427 			default_policy,
1428 			alternate_policy,
1429 			alternate_aperture_base,
1430 			alternate_aperture_size);
1431 
1432 	if ((dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) && (qpd->vmid != 0))
1433 		program_sh_mem_settings(dqm, qpd);
1434 
1435 	pr_debug("sh_mem_config: 0x%x, ape1_base: 0x%x, ape1_limit: 0x%x\n",
1436 		qpd->sh_mem_config, qpd->sh_mem_ape1_base,
1437 		qpd->sh_mem_ape1_limit);
1438 
1439 out:
1440 	mutex_unlock(&dqm->lock);
1441 	return retval;
1442 }
1443 
1444 static int set_trap_handler(struct device_queue_manager *dqm,
1445 				struct qcm_process_device *qpd,
1446 				uint64_t tba_addr,
1447 				uint64_t tma_addr)
1448 {
1449 	uint64_t *tma;
1450 
1451 	if (dqm->dev->cwsr_enabled) {
1452 		/* Jump from CWSR trap handler to user trap */
1453 		tma = (uint64_t *)(qpd->cwsr_kaddr + KFD_CWSR_TMA_OFFSET);
1454 		tma[0] = tba_addr;
1455 		tma[1] = tma_addr;
1456 	} else {
1457 		qpd->tba_addr = tba_addr;
1458 		qpd->tma_addr = tma_addr;
1459 	}
1460 
1461 	return 0;
1462 }
1463 
1464 static int process_termination_nocpsch(struct device_queue_manager *dqm,
1465 		struct qcm_process_device *qpd)
1466 {
1467 	struct queue *q, *next;
1468 	struct device_process_node *cur, *next_dpn;
1469 	int retval = 0;
1470 
1471 	mutex_lock(&dqm->lock);
1472 
1473 	/* Clear all user mode queues */
1474 	list_for_each_entry_safe(q, next, &qpd->queues_list, list) {
1475 		int ret;
1476 
1477 		ret = destroy_queue_nocpsch_locked(dqm, qpd, q);
1478 		if (ret)
1479 			retval = ret;
1480 	}
1481 
1482 	/* Unregister process */
1483 	list_for_each_entry_safe(cur, next_dpn, &dqm->queues, list) {
1484 		if (qpd == cur->qpd) {
1485 			list_del(&cur->list);
1486 			kfree(cur);
1487 			dqm->processes_count--;
1488 			break;
1489 		}
1490 	}
1491 
1492 	mutex_unlock(&dqm->lock);
1493 	return retval;
1494 }
1495 
1496 
1497 static int process_termination_cpsch(struct device_queue_manager *dqm,
1498 		struct qcm_process_device *qpd)
1499 {
1500 	int retval;
1501 	struct queue *q, *next;
1502 	struct kernel_queue *kq, *kq_next;
1503 	struct mqd_manager *mqd;
1504 	struct device_process_node *cur, *next_dpn;
1505 	enum kfd_unmap_queues_filter filter =
1506 		KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES;
1507 
1508 	retval = 0;
1509 
1510 	mutex_lock(&dqm->lock);
1511 
1512 	/* Clean all kernel queues */
1513 	list_for_each_entry_safe(kq, kq_next, &qpd->priv_queue_list, list) {
1514 		list_del(&kq->list);
1515 		dqm->queue_count--;
1516 		qpd->is_debug = false;
1517 		dqm->total_queue_count--;
1518 		filter = KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES;
1519 	}
1520 
1521 	/* Clear all user mode queues */
1522 	list_for_each_entry(q, &qpd->queues_list, list) {
1523 		if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
1524 			dqm->sdma_queue_count--;
1525 			deallocate_sdma_queue(dqm, q->sdma_id);
1526 		}
1527 
1528 		if (q->properties.is_active)
1529 			dqm->queue_count--;
1530 
1531 		dqm->total_queue_count--;
1532 	}
1533 
1534 	/* Unregister process */
1535 	list_for_each_entry_safe(cur, next_dpn, &dqm->queues, list) {
1536 		if (qpd == cur->qpd) {
1537 			list_del(&cur->list);
1538 			kfree(cur);
1539 			dqm->processes_count--;
1540 			break;
1541 		}
1542 	}
1543 
1544 	retval = execute_queues_cpsch(dqm, filter, 0);
1545 	if (retval || qpd->reset_wavefronts) {
1546 		pr_warn("Resetting wave fronts (cpsch) on dev %p\n", dqm->dev);
1547 		dbgdev_wave_reset_wavefronts(dqm->dev, qpd->pqm->process);
1548 		qpd->reset_wavefronts = false;
1549 	}
1550 
1551 	/* lastly, free mqd resources */
1552 	list_for_each_entry_safe(q, next, &qpd->queues_list, list) {
1553 		mqd = dqm->ops.get_mqd_manager(dqm,
1554 			get_mqd_type_from_queue_type(q->properties.type));
1555 		if (!mqd) {
1556 			retval = -ENOMEM;
1557 			goto out;
1558 		}
1559 		list_del(&q->list);
1560 		qpd->queue_count--;
1561 		mqd->uninit_mqd(mqd, q->mqd, q->mqd_mem_obj);
1562 	}
1563 
1564 out:
1565 	mutex_unlock(&dqm->lock);
1566 	return retval;
1567 }
1568 
1569 struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev)
1570 {
1571 	struct device_queue_manager *dqm;
1572 
1573 	pr_debug("Loading device queue manager\n");
1574 
1575 	dqm = kzalloc(sizeof(*dqm), GFP_KERNEL);
1576 	if (!dqm)
1577 		return NULL;
1578 
1579 	switch (dev->device_info->asic_family) {
1580 	/* HWS is not available on Hawaii. */
1581 	case CHIP_HAWAII:
1582 	/* HWS depends on CWSR for timely dequeue. CWSR is not
1583 	 * available on Tonga.
1584 	 *
1585 	 * FIXME: This argument also applies to Kaveri.
1586 	 */
1587 	case CHIP_TONGA:
1588 		dqm->sched_policy = KFD_SCHED_POLICY_NO_HWS;
1589 		break;
1590 	default:
1591 		dqm->sched_policy = sched_policy;
1592 		break;
1593 	}
1594 
1595 	dqm->dev = dev;
1596 	switch (dqm->sched_policy) {
1597 	case KFD_SCHED_POLICY_HWS:
1598 	case KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION:
1599 		/* initialize dqm for cp scheduling */
1600 		dqm->ops.create_queue = create_queue_cpsch;
1601 		dqm->ops.initialize = initialize_cpsch;
1602 		dqm->ops.start = start_cpsch;
1603 		dqm->ops.stop = stop_cpsch;
1604 		dqm->ops.destroy_queue = destroy_queue_cpsch;
1605 		dqm->ops.update_queue = update_queue;
1606 		dqm->ops.get_mqd_manager = get_mqd_manager;
1607 		dqm->ops.register_process = register_process;
1608 		dqm->ops.unregister_process = unregister_process;
1609 		dqm->ops.uninitialize = uninitialize;
1610 		dqm->ops.create_kernel_queue = create_kernel_queue_cpsch;
1611 		dqm->ops.destroy_kernel_queue = destroy_kernel_queue_cpsch;
1612 		dqm->ops.set_cache_memory_policy = set_cache_memory_policy;
1613 		dqm->ops.set_trap_handler = set_trap_handler;
1614 		dqm->ops.process_termination = process_termination_cpsch;
1615 		dqm->ops.evict_process_queues = evict_process_queues_cpsch;
1616 		dqm->ops.restore_process_queues = restore_process_queues_cpsch;
1617 		break;
1618 	case KFD_SCHED_POLICY_NO_HWS:
1619 		/* initialize dqm for no cp scheduling */
1620 		dqm->ops.start = start_nocpsch;
1621 		dqm->ops.stop = stop_nocpsch;
1622 		dqm->ops.create_queue = create_queue_nocpsch;
1623 		dqm->ops.destroy_queue = destroy_queue_nocpsch;
1624 		dqm->ops.update_queue = update_queue;
1625 		dqm->ops.get_mqd_manager = get_mqd_manager;
1626 		dqm->ops.register_process = register_process;
1627 		dqm->ops.unregister_process = unregister_process;
1628 		dqm->ops.initialize = initialize_nocpsch;
1629 		dqm->ops.uninitialize = uninitialize;
1630 		dqm->ops.set_cache_memory_policy = set_cache_memory_policy;
1631 		dqm->ops.set_trap_handler = set_trap_handler;
1632 		dqm->ops.process_termination = process_termination_nocpsch;
1633 		dqm->ops.evict_process_queues = evict_process_queues_nocpsch;
1634 		dqm->ops.restore_process_queues =
1635 			restore_process_queues_nocpsch;
1636 		break;
1637 	default:
1638 		pr_err("Invalid scheduling policy %d\n", dqm->sched_policy);
1639 		goto out_free;
1640 	}
1641 
1642 	switch (dev->device_info->asic_family) {
1643 	case CHIP_CARRIZO:
1644 		device_queue_manager_init_vi(&dqm->asic_ops);
1645 		break;
1646 
1647 	case CHIP_KAVERI:
1648 		device_queue_manager_init_cik(&dqm->asic_ops);
1649 		break;
1650 
1651 	case CHIP_HAWAII:
1652 		device_queue_manager_init_cik_hawaii(&dqm->asic_ops);
1653 		break;
1654 
1655 	case CHIP_TONGA:
1656 	case CHIP_FIJI:
1657 	case CHIP_POLARIS10:
1658 	case CHIP_POLARIS11:
1659 		device_queue_manager_init_vi_tonga(&dqm->asic_ops);
1660 		break;
1661 
1662 	case CHIP_VEGA10:
1663 	case CHIP_RAVEN:
1664 		device_queue_manager_init_v9(&dqm->asic_ops);
1665 		break;
1666 	default:
1667 		WARN(1, "Unexpected ASIC family %u",
1668 		     dev->device_info->asic_family);
1669 		goto out_free;
1670 	}
1671 
1672 	if (!dqm->ops.initialize(dqm))
1673 		return dqm;
1674 
1675 out_free:
1676 	kfree(dqm);
1677 	return NULL;
1678 }
1679 
1680 void device_queue_manager_uninit(struct device_queue_manager *dqm)
1681 {
1682 	dqm->ops.uninitialize(dqm);
1683 	kfree(dqm);
1684 }
1685 
1686 #if defined(CONFIG_DEBUG_FS)
1687 
1688 static void seq_reg_dump(struct seq_file *m,
1689 			 uint32_t (*dump)[2], uint32_t n_regs)
1690 {
1691 	uint32_t i, count;
1692 
1693 	for (i = 0, count = 0; i < n_regs; i++) {
1694 		if (count == 0 ||
1695 		    dump[i-1][0] + sizeof(uint32_t) != dump[i][0]) {
1696 			seq_printf(m, "%s    %08x: %08x",
1697 				   i ? "\n" : "",
1698 				   dump[i][0], dump[i][1]);
1699 			count = 7;
1700 		} else {
1701 			seq_printf(m, " %08x", dump[i][1]);
1702 			count--;
1703 		}
1704 	}
1705 
1706 	seq_puts(m, "\n");
1707 }
1708 
1709 int dqm_debugfs_hqds(struct seq_file *m, void *data)
1710 {
1711 	struct device_queue_manager *dqm = data;
1712 	uint32_t (*dump)[2], n_regs;
1713 	int pipe, queue;
1714 	int r = 0;
1715 
1716 	r = dqm->dev->kfd2kgd->hqd_dump(dqm->dev->kgd,
1717 		KFD_CIK_HIQ_PIPE, KFD_CIK_HIQ_QUEUE, &dump, &n_regs);
1718 	if (!r) {
1719 		seq_printf(m, "  HIQ on MEC %d Pipe %d Queue %d\n",
1720 				KFD_CIK_HIQ_PIPE/get_pipes_per_mec(dqm)+1,
1721 				KFD_CIK_HIQ_PIPE%get_pipes_per_mec(dqm),
1722 				KFD_CIK_HIQ_QUEUE);
1723 		seq_reg_dump(m, dump, n_regs);
1724 
1725 		kfree(dump);
1726 	}
1727 
1728 	for (pipe = 0; pipe < get_pipes_per_mec(dqm); pipe++) {
1729 		int pipe_offset = pipe * get_queues_per_pipe(dqm);
1730 
1731 		for (queue = 0; queue < get_queues_per_pipe(dqm); queue++) {
1732 			if (!test_bit(pipe_offset + queue,
1733 				      dqm->dev->shared_resources.queue_bitmap))
1734 				continue;
1735 
1736 			r = dqm->dev->kfd2kgd->hqd_dump(
1737 				dqm->dev->kgd, pipe, queue, &dump, &n_regs);
1738 			if (r)
1739 				break;
1740 
1741 			seq_printf(m, "  CP Pipe %d, Queue %d\n",
1742 				  pipe, queue);
1743 			seq_reg_dump(m, dump, n_regs);
1744 
1745 			kfree(dump);
1746 		}
1747 	}
1748 
1749 	for (pipe = 0; pipe < CIK_SDMA_ENGINE_NUM; pipe++) {
1750 		for (queue = 0; queue < CIK_SDMA_QUEUES_PER_ENGINE; queue++) {
1751 			r = dqm->dev->kfd2kgd->hqd_sdma_dump(
1752 				dqm->dev->kgd, pipe, queue, &dump, &n_regs);
1753 			if (r)
1754 				break;
1755 
1756 			seq_printf(m, "  SDMA Engine %d, RLC %d\n",
1757 				  pipe, queue);
1758 			seq_reg_dump(m, dump, n_regs);
1759 
1760 			kfree(dump);
1761 		}
1762 	}
1763 
1764 	return r;
1765 }
1766 
1767 #endif
1768