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 #include "amdgpu_amdkfd.h"
37 
38 /* Size of the per-pipe EOP queue */
39 #define CIK_HPD_EOP_BYTES_LOG2 11
40 #define CIK_HPD_EOP_BYTES (1U << CIK_HPD_EOP_BYTES_LOG2)
41 
42 static int set_pasid_vmid_mapping(struct device_queue_manager *dqm,
43 				  u32 pasid, unsigned int vmid);
44 
45 static int execute_queues_cpsch(struct device_queue_manager *dqm,
46 				enum kfd_unmap_queues_filter filter,
47 				uint32_t filter_param);
48 static int unmap_queues_cpsch(struct device_queue_manager *dqm,
49 				enum kfd_unmap_queues_filter filter,
50 				uint32_t filter_param);
51 
52 static int map_queues_cpsch(struct device_queue_manager *dqm);
53 
54 static void deallocate_sdma_queue(struct device_queue_manager *dqm,
55 				struct queue *q);
56 
57 static inline void deallocate_hqd(struct device_queue_manager *dqm,
58 				struct queue *q);
59 static int allocate_hqd(struct device_queue_manager *dqm, struct queue *q);
60 static int allocate_sdma_queue(struct device_queue_manager *dqm,
61 				struct queue *q);
62 static void kfd_process_hw_exception(struct work_struct *work);
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 || type == KFD_QUEUE_TYPE_SDMA_XGMI)
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.cp_queue_bitmap))
82 			return true;
83 	return false;
84 }
85 
86 unsigned int get_cp_queues_num(struct device_queue_manager *dqm)
87 {
88 	return bitmap_weight(dqm->dev->shared_resources.cp_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 static unsigned int get_num_sdma_engines(struct device_queue_manager *dqm)
103 {
104 	return dqm->dev->device_info->num_sdma_engines;
105 }
106 
107 static unsigned int get_num_xgmi_sdma_engines(struct device_queue_manager *dqm)
108 {
109 	return dqm->dev->device_info->num_xgmi_sdma_engines;
110 }
111 
112 static unsigned int get_num_all_sdma_engines(struct device_queue_manager *dqm)
113 {
114 	return get_num_sdma_engines(dqm) + get_num_xgmi_sdma_engines(dqm);
115 }
116 
117 unsigned int get_num_sdma_queues(struct device_queue_manager *dqm)
118 {
119 	return dqm->dev->device_info->num_sdma_engines
120 			* dqm->dev->device_info->num_sdma_queues_per_engine;
121 }
122 
123 unsigned int get_num_xgmi_sdma_queues(struct device_queue_manager *dqm)
124 {
125 	return dqm->dev->device_info->num_xgmi_sdma_engines
126 			* dqm->dev->device_info->num_sdma_queues_per_engine;
127 }
128 
129 void program_sh_mem_settings(struct device_queue_manager *dqm,
130 					struct qcm_process_device *qpd)
131 {
132 	return dqm->dev->kfd2kgd->program_sh_mem_settings(
133 						dqm->dev->kgd, qpd->vmid,
134 						qpd->sh_mem_config,
135 						qpd->sh_mem_ape1_base,
136 						qpd->sh_mem_ape1_limit,
137 						qpd->sh_mem_bases);
138 }
139 
140 static void increment_queue_count(struct device_queue_manager *dqm,
141 			enum kfd_queue_type type)
142 {
143 	dqm->active_queue_count++;
144 	if (type == KFD_QUEUE_TYPE_COMPUTE || type == KFD_QUEUE_TYPE_DIQ)
145 		dqm->active_cp_queue_count++;
146 }
147 
148 static void decrement_queue_count(struct device_queue_manager *dqm,
149 			enum kfd_queue_type type)
150 {
151 	dqm->active_queue_count--;
152 	if (type == KFD_QUEUE_TYPE_COMPUTE || type == KFD_QUEUE_TYPE_DIQ)
153 		dqm->active_cp_queue_count--;
154 }
155 
156 static int allocate_doorbell(struct qcm_process_device *qpd, struct queue *q)
157 {
158 	struct kfd_dev *dev = qpd->dqm->dev;
159 
160 	if (!KFD_IS_SOC15(dev->device_info->asic_family)) {
161 		/* On pre-SOC15 chips we need to use the queue ID to
162 		 * preserve the user mode ABI.
163 		 */
164 		q->doorbell_id = q->properties.queue_id;
165 	} else if (q->properties.type == KFD_QUEUE_TYPE_SDMA ||
166 			q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
167 		/* For SDMA queues on SOC15 with 8-byte doorbell, use static
168 		 * doorbell assignments based on the engine and queue id.
169 		 * The doobell index distance between RLC (2*i) and (2*i+1)
170 		 * for a SDMA engine is 512.
171 		 */
172 		uint32_t *idx_offset =
173 				dev->shared_resources.sdma_doorbell_idx;
174 
175 		q->doorbell_id = idx_offset[q->properties.sdma_engine_id]
176 			+ (q->properties.sdma_queue_id & 1)
177 			* KFD_QUEUE_DOORBELL_MIRROR_OFFSET
178 			+ (q->properties.sdma_queue_id >> 1);
179 	} else {
180 		/* For CP queues on SOC15 reserve a free doorbell ID */
181 		unsigned int found;
182 
183 		found = find_first_zero_bit(qpd->doorbell_bitmap,
184 					    KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);
185 		if (found >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) {
186 			pr_debug("No doorbells available");
187 			return -EBUSY;
188 		}
189 		set_bit(found, qpd->doorbell_bitmap);
190 		q->doorbell_id = found;
191 	}
192 
193 	q->properties.doorbell_off =
194 		kfd_get_doorbell_dw_offset_in_bar(dev, qpd_to_pdd(qpd),
195 					  q->doorbell_id);
196 	return 0;
197 }
198 
199 static void deallocate_doorbell(struct qcm_process_device *qpd,
200 				struct queue *q)
201 {
202 	unsigned int old;
203 	struct kfd_dev *dev = qpd->dqm->dev;
204 
205 	if (!KFD_IS_SOC15(dev->device_info->asic_family) ||
206 	    q->properties.type == KFD_QUEUE_TYPE_SDMA ||
207 	    q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)
208 		return;
209 
210 	old = test_and_clear_bit(q->doorbell_id, qpd->doorbell_bitmap);
211 	WARN_ON(!old);
212 }
213 
214 static int allocate_vmid(struct device_queue_manager *dqm,
215 			struct qcm_process_device *qpd,
216 			struct queue *q)
217 {
218 	int allocated_vmid = -1, i;
219 
220 	for (i = dqm->dev->vm_info.first_vmid_kfd;
221 			i <= dqm->dev->vm_info.last_vmid_kfd; i++) {
222 		if (!dqm->vmid_pasid[i]) {
223 			allocated_vmid = i;
224 			break;
225 		}
226 	}
227 
228 	if (allocated_vmid < 0) {
229 		pr_err("no more vmid to allocate\n");
230 		return -ENOSPC;
231 	}
232 
233 	pr_debug("vmid allocated: %d\n", allocated_vmid);
234 
235 	dqm->vmid_pasid[allocated_vmid] = q->process->pasid;
236 
237 	set_pasid_vmid_mapping(dqm, q->process->pasid, allocated_vmid);
238 
239 	qpd->vmid = allocated_vmid;
240 	q->properties.vmid = allocated_vmid;
241 
242 	program_sh_mem_settings(dqm, qpd);
243 
244 	/* qpd->page_table_base is set earlier when register_process()
245 	 * is called, i.e. when the first queue is created.
246 	 */
247 	dqm->dev->kfd2kgd->set_vm_context_page_table_base(dqm->dev->kgd,
248 			qpd->vmid,
249 			qpd->page_table_base);
250 	/* invalidate the VM context after pasid and vmid mapping is set up */
251 	kfd_flush_tlb(qpd_to_pdd(qpd));
252 
253 	if (dqm->dev->kfd2kgd->set_scratch_backing_va)
254 		dqm->dev->kfd2kgd->set_scratch_backing_va(dqm->dev->kgd,
255 				qpd->sh_hidden_private_base, qpd->vmid);
256 
257 	return 0;
258 }
259 
260 static int flush_texture_cache_nocpsch(struct kfd_dev *kdev,
261 				struct qcm_process_device *qpd)
262 {
263 	const struct packet_manager_funcs *pmf = qpd->dqm->packets.pmf;
264 	int ret;
265 
266 	if (!qpd->ib_kaddr)
267 		return -ENOMEM;
268 
269 	ret = pmf->release_mem(qpd->ib_base, (uint32_t *)qpd->ib_kaddr);
270 	if (ret)
271 		return ret;
272 
273 	return amdgpu_amdkfd_submit_ib(kdev->kgd, KGD_ENGINE_MEC1, qpd->vmid,
274 				qpd->ib_base, (uint32_t *)qpd->ib_kaddr,
275 				pmf->release_mem_size / sizeof(uint32_t));
276 }
277 
278 static void deallocate_vmid(struct device_queue_manager *dqm,
279 				struct qcm_process_device *qpd,
280 				struct queue *q)
281 {
282 	/* On GFX v7, CP doesn't flush TC at dequeue */
283 	if (q->device->device_info->asic_family == CHIP_HAWAII)
284 		if (flush_texture_cache_nocpsch(q->device, qpd))
285 			pr_err("Failed to flush TC\n");
286 
287 	kfd_flush_tlb(qpd_to_pdd(qpd));
288 
289 	/* Release the vmid mapping */
290 	set_pasid_vmid_mapping(dqm, 0, qpd->vmid);
291 	dqm->vmid_pasid[qpd->vmid] = 0;
292 
293 	qpd->vmid = 0;
294 	q->properties.vmid = 0;
295 }
296 
297 static int create_queue_nocpsch(struct device_queue_manager *dqm,
298 				struct queue *q,
299 				struct qcm_process_device *qpd)
300 {
301 	struct mqd_manager *mqd_mgr;
302 	int retval;
303 
304 	dqm_lock(dqm);
305 
306 	if (dqm->total_queue_count >= max_num_of_queues_per_device) {
307 		pr_warn("Can't create new usermode queue because %d queues were already created\n",
308 				dqm->total_queue_count);
309 		retval = -EPERM;
310 		goto out_unlock;
311 	}
312 
313 	if (list_empty(&qpd->queues_list)) {
314 		retval = allocate_vmid(dqm, qpd, q);
315 		if (retval)
316 			goto out_unlock;
317 	}
318 	q->properties.vmid = qpd->vmid;
319 	/*
320 	 * Eviction state logic: mark all queues as evicted, even ones
321 	 * not currently active. Restoring inactive queues later only
322 	 * updates the is_evicted flag but is a no-op otherwise.
323 	 */
324 	q->properties.is_evicted = !!qpd->evicted;
325 
326 	q->properties.tba_addr = qpd->tba_addr;
327 	q->properties.tma_addr = qpd->tma_addr;
328 
329 	mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
330 			q->properties.type)];
331 	if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE) {
332 		retval = allocate_hqd(dqm, q);
333 		if (retval)
334 			goto deallocate_vmid;
335 		pr_debug("Loading mqd to hqd on pipe %d, queue %d\n",
336 			q->pipe, q->queue);
337 	} else if (q->properties.type == KFD_QUEUE_TYPE_SDMA ||
338 		q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
339 		retval = allocate_sdma_queue(dqm, q);
340 		if (retval)
341 			goto deallocate_vmid;
342 		dqm->asic_ops.init_sdma_vm(dqm, q, qpd);
343 	}
344 
345 	retval = allocate_doorbell(qpd, q);
346 	if (retval)
347 		goto out_deallocate_hqd;
348 
349 	/* Temporarily release dqm lock to avoid a circular lock dependency */
350 	dqm_unlock(dqm);
351 	q->mqd_mem_obj = mqd_mgr->allocate_mqd(mqd_mgr->dev, &q->properties);
352 	dqm_lock(dqm);
353 
354 	if (!q->mqd_mem_obj) {
355 		retval = -ENOMEM;
356 		goto out_deallocate_doorbell;
357 	}
358 	mqd_mgr->init_mqd(mqd_mgr, &q->mqd, q->mqd_mem_obj,
359 				&q->gart_mqd_addr, &q->properties);
360 	if (q->properties.is_active) {
361 		if (!dqm->sched_running) {
362 			WARN_ONCE(1, "Load non-HWS mqd while stopped\n");
363 			goto add_queue_to_list;
364 		}
365 
366 		if (WARN(q->process->mm != current->mm,
367 					"should only run in user thread"))
368 			retval = -EFAULT;
369 		else
370 			retval = mqd_mgr->load_mqd(mqd_mgr, q->mqd, q->pipe,
371 					q->queue, &q->properties, current->mm);
372 		if (retval)
373 			goto out_free_mqd;
374 	}
375 
376 add_queue_to_list:
377 	list_add(&q->list, &qpd->queues_list);
378 	qpd->queue_count++;
379 	if (q->properties.is_active)
380 		increment_queue_count(dqm, q->properties.type);
381 
382 	/*
383 	 * Unconditionally increment this counter, regardless of the queue's
384 	 * type or whether the queue is active.
385 	 */
386 	dqm->total_queue_count++;
387 	pr_debug("Total of %d queues are accountable so far\n",
388 			dqm->total_queue_count);
389 	goto out_unlock;
390 
391 out_free_mqd:
392 	mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
393 out_deallocate_doorbell:
394 	deallocate_doorbell(qpd, q);
395 out_deallocate_hqd:
396 	if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE)
397 		deallocate_hqd(dqm, q);
398 	else if (q->properties.type == KFD_QUEUE_TYPE_SDMA ||
399 		q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)
400 		deallocate_sdma_queue(dqm, q);
401 deallocate_vmid:
402 	if (list_empty(&qpd->queues_list))
403 		deallocate_vmid(dqm, qpd, q);
404 out_unlock:
405 	dqm_unlock(dqm);
406 	return retval;
407 }
408 
409 static int allocate_hqd(struct device_queue_manager *dqm, struct queue *q)
410 {
411 	bool set;
412 	int pipe, bit, i;
413 
414 	set = false;
415 
416 	for (pipe = dqm->next_pipe_to_allocate, i = 0;
417 			i < get_pipes_per_mec(dqm);
418 			pipe = ((pipe + 1) % get_pipes_per_mec(dqm)), ++i) {
419 
420 		if (!is_pipe_enabled(dqm, 0, pipe))
421 			continue;
422 
423 		if (dqm->allocated_queues[pipe] != 0) {
424 			bit = ffs(dqm->allocated_queues[pipe]) - 1;
425 			dqm->allocated_queues[pipe] &= ~(1 << bit);
426 			q->pipe = pipe;
427 			q->queue = bit;
428 			set = true;
429 			break;
430 		}
431 	}
432 
433 	if (!set)
434 		return -EBUSY;
435 
436 	pr_debug("hqd slot - pipe %d, queue %d\n", q->pipe, q->queue);
437 	/* horizontal hqd allocation */
438 	dqm->next_pipe_to_allocate = (pipe + 1) % get_pipes_per_mec(dqm);
439 
440 	return 0;
441 }
442 
443 static inline void deallocate_hqd(struct device_queue_manager *dqm,
444 				struct queue *q)
445 {
446 	dqm->allocated_queues[q->pipe] |= (1 << q->queue);
447 }
448 
449 /* Access to DQM has to be locked before calling destroy_queue_nocpsch_locked
450  * to avoid asynchronized access
451  */
452 static int destroy_queue_nocpsch_locked(struct device_queue_manager *dqm,
453 				struct qcm_process_device *qpd,
454 				struct queue *q)
455 {
456 	int retval;
457 	struct mqd_manager *mqd_mgr;
458 
459 	mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
460 			q->properties.type)];
461 
462 	if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE)
463 		deallocate_hqd(dqm, q);
464 	else if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
465 		deallocate_sdma_queue(dqm, q);
466 	else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)
467 		deallocate_sdma_queue(dqm, q);
468 	else {
469 		pr_debug("q->properties.type %d is invalid\n",
470 				q->properties.type);
471 		return -EINVAL;
472 	}
473 	dqm->total_queue_count--;
474 
475 	deallocate_doorbell(qpd, q);
476 
477 	if (!dqm->sched_running) {
478 		WARN_ONCE(1, "Destroy non-HWS queue while stopped\n");
479 		return 0;
480 	}
481 
482 	retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd,
483 				KFD_PREEMPT_TYPE_WAVEFRONT_RESET,
484 				KFD_UNMAP_LATENCY_MS,
485 				q->pipe, q->queue);
486 	if (retval == -ETIME)
487 		qpd->reset_wavefronts = true;
488 
489 
490 	mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
491 
492 	list_del(&q->list);
493 	if (list_empty(&qpd->queues_list)) {
494 		if (qpd->reset_wavefronts) {
495 			pr_warn("Resetting wave fronts (nocpsch) on dev %p\n",
496 					dqm->dev);
497 			/* dbgdev_wave_reset_wavefronts has to be called before
498 			 * deallocate_vmid(), i.e. when vmid is still in use.
499 			 */
500 			dbgdev_wave_reset_wavefronts(dqm->dev,
501 					qpd->pqm->process);
502 			qpd->reset_wavefronts = false;
503 		}
504 
505 		deallocate_vmid(dqm, qpd, q);
506 	}
507 	qpd->queue_count--;
508 	if (q->properties.is_active) {
509 		decrement_queue_count(dqm, q->properties.type);
510 		if (q->properties.is_gws) {
511 			dqm->gws_queue_count--;
512 			qpd->mapped_gws_queue = false;
513 		}
514 	}
515 
516 	return retval;
517 }
518 
519 static int destroy_queue_nocpsch(struct device_queue_manager *dqm,
520 				struct qcm_process_device *qpd,
521 				struct queue *q)
522 {
523 	int retval;
524 	uint64_t sdma_val = 0;
525 	struct kfd_process_device *pdd = qpd_to_pdd(qpd);
526 
527 	/* Get the SDMA queue stats */
528 	if ((q->properties.type == KFD_QUEUE_TYPE_SDMA) ||
529 	    (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) {
530 		retval = read_sdma_queue_counter((uint64_t __user *)q->properties.read_ptr,
531 							&sdma_val);
532 		if (retval)
533 			pr_err("Failed to read SDMA queue counter for queue: %d\n",
534 				q->properties.queue_id);
535 	}
536 
537 	dqm_lock(dqm);
538 	retval = destroy_queue_nocpsch_locked(dqm, qpd, q);
539 	if (!retval)
540 		pdd->sdma_past_activity_counter += sdma_val;
541 	dqm_unlock(dqm);
542 
543 	return retval;
544 }
545 
546 static int update_queue(struct device_queue_manager *dqm, struct queue *q)
547 {
548 	int retval = 0;
549 	struct mqd_manager *mqd_mgr;
550 	struct kfd_process_device *pdd;
551 	bool prev_active = false;
552 
553 	dqm_lock(dqm);
554 	pdd = kfd_get_process_device_data(q->device, q->process);
555 	if (!pdd) {
556 		retval = -ENODEV;
557 		goto out_unlock;
558 	}
559 	mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
560 			q->properties.type)];
561 
562 	/* Save previous activity state for counters */
563 	prev_active = q->properties.is_active;
564 
565 	/* Make sure the queue is unmapped before updating the MQD */
566 	if (dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) {
567 		retval = unmap_queues_cpsch(dqm,
568 				KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
569 		if (retval) {
570 			pr_err("unmap queue failed\n");
571 			goto out_unlock;
572 		}
573 	} else if (prev_active &&
574 		   (q->properties.type == KFD_QUEUE_TYPE_COMPUTE ||
575 		    q->properties.type == KFD_QUEUE_TYPE_SDMA ||
576 		    q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) {
577 
578 		if (!dqm->sched_running) {
579 			WARN_ONCE(1, "Update non-HWS queue while stopped\n");
580 			goto out_unlock;
581 		}
582 
583 		retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd,
584 				KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN,
585 				KFD_UNMAP_LATENCY_MS, q->pipe, q->queue);
586 		if (retval) {
587 			pr_err("destroy mqd failed\n");
588 			goto out_unlock;
589 		}
590 	}
591 
592 	mqd_mgr->update_mqd(mqd_mgr, q->mqd, &q->properties);
593 
594 	/*
595 	 * check active state vs. the previous state and modify
596 	 * counter accordingly. map_queues_cpsch uses the
597 	 * dqm->active_queue_count to determine whether a new runlist must be
598 	 * uploaded.
599 	 */
600 	if (q->properties.is_active && !prev_active)
601 		increment_queue_count(dqm, q->properties.type);
602 	else if (!q->properties.is_active && prev_active)
603 		decrement_queue_count(dqm, q->properties.type);
604 
605 	if (q->gws && !q->properties.is_gws) {
606 		if (q->properties.is_active) {
607 			dqm->gws_queue_count++;
608 			pdd->qpd.mapped_gws_queue = true;
609 		}
610 		q->properties.is_gws = true;
611 	} else if (!q->gws && q->properties.is_gws) {
612 		if (q->properties.is_active) {
613 			dqm->gws_queue_count--;
614 			pdd->qpd.mapped_gws_queue = false;
615 		}
616 		q->properties.is_gws = false;
617 	}
618 
619 	if (dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS)
620 		retval = map_queues_cpsch(dqm);
621 	else if (q->properties.is_active &&
622 		 (q->properties.type == KFD_QUEUE_TYPE_COMPUTE ||
623 		  q->properties.type == KFD_QUEUE_TYPE_SDMA ||
624 		  q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) {
625 		if (WARN(q->process->mm != current->mm,
626 			 "should only run in user thread"))
627 			retval = -EFAULT;
628 		else
629 			retval = mqd_mgr->load_mqd(mqd_mgr, q->mqd,
630 						   q->pipe, q->queue,
631 						   &q->properties, current->mm);
632 	}
633 
634 out_unlock:
635 	dqm_unlock(dqm);
636 	return retval;
637 }
638 
639 static int evict_process_queues_nocpsch(struct device_queue_manager *dqm,
640 					struct qcm_process_device *qpd)
641 {
642 	struct queue *q;
643 	struct mqd_manager *mqd_mgr;
644 	struct kfd_process_device *pdd;
645 	int retval, ret = 0;
646 
647 	dqm_lock(dqm);
648 	if (qpd->evicted++ > 0) /* already evicted, do nothing */
649 		goto out;
650 
651 	pdd = qpd_to_pdd(qpd);
652 	pr_debug_ratelimited("Evicting PASID 0x%x queues\n",
653 			    pdd->process->pasid);
654 
655 	pdd->last_evict_timestamp = get_jiffies_64();
656 	/* Mark all queues as evicted. Deactivate all active queues on
657 	 * the qpd.
658 	 */
659 	list_for_each_entry(q, &qpd->queues_list, list) {
660 		q->properties.is_evicted = true;
661 		if (!q->properties.is_active)
662 			continue;
663 
664 		mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
665 				q->properties.type)];
666 		q->properties.is_active = false;
667 		decrement_queue_count(dqm, q->properties.type);
668 		if (q->properties.is_gws) {
669 			dqm->gws_queue_count--;
670 			qpd->mapped_gws_queue = false;
671 		}
672 
673 		if (WARN_ONCE(!dqm->sched_running, "Evict when stopped\n"))
674 			continue;
675 
676 		retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd,
677 				KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN,
678 				KFD_UNMAP_LATENCY_MS, q->pipe, q->queue);
679 		if (retval && !ret)
680 			/* Return the first error, but keep going to
681 			 * maintain a consistent eviction state
682 			 */
683 			ret = retval;
684 	}
685 
686 out:
687 	dqm_unlock(dqm);
688 	return ret;
689 }
690 
691 static int evict_process_queues_cpsch(struct device_queue_manager *dqm,
692 				      struct qcm_process_device *qpd)
693 {
694 	struct queue *q;
695 	struct kfd_process_device *pdd;
696 	int retval = 0;
697 
698 	dqm_lock(dqm);
699 	if (qpd->evicted++ > 0) /* already evicted, do nothing */
700 		goto out;
701 
702 	pdd = qpd_to_pdd(qpd);
703 	pr_debug_ratelimited("Evicting PASID 0x%x queues\n",
704 			    pdd->process->pasid);
705 
706 	/* Mark all queues as evicted. Deactivate all active queues on
707 	 * the qpd.
708 	 */
709 	list_for_each_entry(q, &qpd->queues_list, list) {
710 		q->properties.is_evicted = true;
711 		if (!q->properties.is_active)
712 			continue;
713 
714 		q->properties.is_active = false;
715 		decrement_queue_count(dqm, q->properties.type);
716 	}
717 	pdd->last_evict_timestamp = get_jiffies_64();
718 	retval = execute_queues_cpsch(dqm,
719 				qpd->is_debug ?
720 				KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES :
721 				KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
722 
723 out:
724 	dqm_unlock(dqm);
725 	return retval;
726 }
727 
728 static int restore_process_queues_nocpsch(struct device_queue_manager *dqm,
729 					  struct qcm_process_device *qpd)
730 {
731 	struct mm_struct *mm = NULL;
732 	struct queue *q;
733 	struct mqd_manager *mqd_mgr;
734 	struct kfd_process_device *pdd;
735 	uint64_t pd_base;
736 	uint64_t eviction_duration;
737 	int retval, ret = 0;
738 
739 	pdd = qpd_to_pdd(qpd);
740 	/* Retrieve PD base */
741 	pd_base = amdgpu_amdkfd_gpuvm_get_process_page_dir(pdd->drm_priv);
742 
743 	dqm_lock(dqm);
744 	if (WARN_ON_ONCE(!qpd->evicted)) /* already restored, do nothing */
745 		goto out;
746 	if (qpd->evicted > 1) { /* ref count still > 0, decrement & quit */
747 		qpd->evicted--;
748 		goto out;
749 	}
750 
751 	pr_debug_ratelimited("Restoring PASID 0x%x queues\n",
752 			    pdd->process->pasid);
753 
754 	/* Update PD Base in QPD */
755 	qpd->page_table_base = pd_base;
756 	pr_debug("Updated PD address to 0x%llx\n", pd_base);
757 
758 	if (!list_empty(&qpd->queues_list)) {
759 		dqm->dev->kfd2kgd->set_vm_context_page_table_base(
760 				dqm->dev->kgd,
761 				qpd->vmid,
762 				qpd->page_table_base);
763 		kfd_flush_tlb(pdd);
764 	}
765 
766 	/* Take a safe reference to the mm_struct, which may otherwise
767 	 * disappear even while the kfd_process is still referenced.
768 	 */
769 	mm = get_task_mm(pdd->process->lead_thread);
770 	if (!mm) {
771 		ret = -EFAULT;
772 		goto out;
773 	}
774 
775 	/* Remove the eviction flags. Activate queues that are not
776 	 * inactive for other reasons.
777 	 */
778 	list_for_each_entry(q, &qpd->queues_list, list) {
779 		q->properties.is_evicted = false;
780 		if (!QUEUE_IS_ACTIVE(q->properties))
781 			continue;
782 
783 		mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
784 				q->properties.type)];
785 		q->properties.is_active = true;
786 		increment_queue_count(dqm, q->properties.type);
787 		if (q->properties.is_gws) {
788 			dqm->gws_queue_count++;
789 			qpd->mapped_gws_queue = true;
790 		}
791 
792 		if (WARN_ONCE(!dqm->sched_running, "Restore when stopped\n"))
793 			continue;
794 
795 		retval = mqd_mgr->load_mqd(mqd_mgr, q->mqd, q->pipe,
796 				       q->queue, &q->properties, mm);
797 		if (retval && !ret)
798 			/* Return the first error, but keep going to
799 			 * maintain a consistent eviction state
800 			 */
801 			ret = retval;
802 	}
803 	qpd->evicted = 0;
804 	eviction_duration = get_jiffies_64() - pdd->last_evict_timestamp;
805 	atomic64_add(eviction_duration, &pdd->evict_duration_counter);
806 out:
807 	if (mm)
808 		mmput(mm);
809 	dqm_unlock(dqm);
810 	return ret;
811 }
812 
813 static int restore_process_queues_cpsch(struct device_queue_manager *dqm,
814 					struct qcm_process_device *qpd)
815 {
816 	struct queue *q;
817 	struct kfd_process_device *pdd;
818 	uint64_t pd_base;
819 	uint64_t eviction_duration;
820 	int retval = 0;
821 
822 	pdd = qpd_to_pdd(qpd);
823 	/* Retrieve PD base */
824 	pd_base = amdgpu_amdkfd_gpuvm_get_process_page_dir(pdd->drm_priv);
825 
826 	dqm_lock(dqm);
827 	if (WARN_ON_ONCE(!qpd->evicted)) /* already restored, do nothing */
828 		goto out;
829 	if (qpd->evicted > 1) { /* ref count still > 0, decrement & quit */
830 		qpd->evicted--;
831 		goto out;
832 	}
833 
834 	pr_debug_ratelimited("Restoring PASID 0x%x queues\n",
835 			    pdd->process->pasid);
836 
837 	/* Update PD Base in QPD */
838 	qpd->page_table_base = pd_base;
839 	pr_debug("Updated PD address to 0x%llx\n", pd_base);
840 
841 	/* activate all active queues on the qpd */
842 	list_for_each_entry(q, &qpd->queues_list, list) {
843 		q->properties.is_evicted = false;
844 		if (!QUEUE_IS_ACTIVE(q->properties))
845 			continue;
846 
847 		q->properties.is_active = true;
848 		increment_queue_count(dqm, q->properties.type);
849 	}
850 	retval = execute_queues_cpsch(dqm,
851 				KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
852 	qpd->evicted = 0;
853 	eviction_duration = get_jiffies_64() - pdd->last_evict_timestamp;
854 	atomic64_add(eviction_duration, &pdd->evict_duration_counter);
855 out:
856 	dqm_unlock(dqm);
857 	return retval;
858 }
859 
860 static int register_process(struct device_queue_manager *dqm,
861 					struct qcm_process_device *qpd)
862 {
863 	struct device_process_node *n;
864 	struct kfd_process_device *pdd;
865 	uint64_t pd_base;
866 	int retval;
867 
868 	n = kzalloc(sizeof(*n), GFP_KERNEL);
869 	if (!n)
870 		return -ENOMEM;
871 
872 	n->qpd = qpd;
873 
874 	pdd = qpd_to_pdd(qpd);
875 	/* Retrieve PD base */
876 	pd_base = amdgpu_amdkfd_gpuvm_get_process_page_dir(pdd->drm_priv);
877 
878 	dqm_lock(dqm);
879 	list_add(&n->list, &dqm->queues);
880 
881 	/* Update PD Base in QPD */
882 	qpd->page_table_base = pd_base;
883 	pr_debug("Updated PD address to 0x%llx\n", pd_base);
884 
885 	retval = dqm->asic_ops.update_qpd(dqm, qpd);
886 
887 	dqm->processes_count++;
888 
889 	dqm_unlock(dqm);
890 
891 	/* Outside the DQM lock because under the DQM lock we can't do
892 	 * reclaim or take other locks that others hold while reclaiming.
893 	 */
894 	kfd_inc_compute_active(dqm->dev);
895 
896 	return retval;
897 }
898 
899 static int unregister_process(struct device_queue_manager *dqm,
900 					struct qcm_process_device *qpd)
901 {
902 	int retval;
903 	struct device_process_node *cur, *next;
904 
905 	pr_debug("qpd->queues_list is %s\n",
906 			list_empty(&qpd->queues_list) ? "empty" : "not empty");
907 
908 	retval = 0;
909 	dqm_lock(dqm);
910 
911 	list_for_each_entry_safe(cur, next, &dqm->queues, list) {
912 		if (qpd == cur->qpd) {
913 			list_del(&cur->list);
914 			kfree(cur);
915 			dqm->processes_count--;
916 			goto out;
917 		}
918 	}
919 	/* qpd not found in dqm list */
920 	retval = 1;
921 out:
922 	dqm_unlock(dqm);
923 
924 	/* Outside the DQM lock because under the DQM lock we can't do
925 	 * reclaim or take other locks that others hold while reclaiming.
926 	 */
927 	if (!retval)
928 		kfd_dec_compute_active(dqm->dev);
929 
930 	return retval;
931 }
932 
933 static int
934 set_pasid_vmid_mapping(struct device_queue_manager *dqm, u32 pasid,
935 			unsigned int vmid)
936 {
937 	return dqm->dev->kfd2kgd->set_pasid_vmid_mapping(
938 						dqm->dev->kgd, pasid, vmid);
939 }
940 
941 static void init_interrupts(struct device_queue_manager *dqm)
942 {
943 	unsigned int i;
944 
945 	for (i = 0 ; i < get_pipes_per_mec(dqm) ; i++)
946 		if (is_pipe_enabled(dqm, 0, i))
947 			dqm->dev->kfd2kgd->init_interrupts(dqm->dev->kgd, i);
948 }
949 
950 static int initialize_nocpsch(struct device_queue_manager *dqm)
951 {
952 	int pipe, queue;
953 
954 	pr_debug("num of pipes: %d\n", get_pipes_per_mec(dqm));
955 
956 	dqm->allocated_queues = kcalloc(get_pipes_per_mec(dqm),
957 					sizeof(unsigned int), GFP_KERNEL);
958 	if (!dqm->allocated_queues)
959 		return -ENOMEM;
960 
961 	mutex_init(&dqm->lock_hidden);
962 	INIT_LIST_HEAD(&dqm->queues);
963 	dqm->active_queue_count = dqm->next_pipe_to_allocate = 0;
964 	dqm->active_cp_queue_count = 0;
965 	dqm->gws_queue_count = 0;
966 
967 	for (pipe = 0; pipe < get_pipes_per_mec(dqm); pipe++) {
968 		int pipe_offset = pipe * get_queues_per_pipe(dqm);
969 
970 		for (queue = 0; queue < get_queues_per_pipe(dqm); queue++)
971 			if (test_bit(pipe_offset + queue,
972 				     dqm->dev->shared_resources.cp_queue_bitmap))
973 				dqm->allocated_queues[pipe] |= 1 << queue;
974 	}
975 
976 	memset(dqm->vmid_pasid, 0, sizeof(dqm->vmid_pasid));
977 
978 	dqm->sdma_bitmap = ~0ULL >> (64 - get_num_sdma_queues(dqm));
979 	dqm->xgmi_sdma_bitmap = ~0ULL >> (64 - get_num_xgmi_sdma_queues(dqm));
980 
981 	return 0;
982 }
983 
984 static void uninitialize(struct device_queue_manager *dqm)
985 {
986 	int i;
987 
988 	WARN_ON(dqm->active_queue_count > 0 || dqm->processes_count > 0);
989 
990 	kfree(dqm->allocated_queues);
991 	for (i = 0 ; i < KFD_MQD_TYPE_MAX ; i++)
992 		kfree(dqm->mqd_mgrs[i]);
993 	mutex_destroy(&dqm->lock_hidden);
994 }
995 
996 static int start_nocpsch(struct device_queue_manager *dqm)
997 {
998 	pr_info("SW scheduler is used");
999 	init_interrupts(dqm);
1000 
1001 	if (dqm->dev->device_info->asic_family == CHIP_HAWAII)
1002 		return pm_init(&dqm->packets, dqm);
1003 	dqm->sched_running = true;
1004 
1005 	return 0;
1006 }
1007 
1008 static int stop_nocpsch(struct device_queue_manager *dqm)
1009 {
1010 	if (dqm->dev->device_info->asic_family == CHIP_HAWAII)
1011 		pm_uninit(&dqm->packets, false);
1012 	dqm->sched_running = false;
1013 
1014 	return 0;
1015 }
1016 
1017 static void pre_reset(struct device_queue_manager *dqm)
1018 {
1019 	dqm_lock(dqm);
1020 	dqm->is_resetting = true;
1021 	dqm_unlock(dqm);
1022 }
1023 
1024 static int allocate_sdma_queue(struct device_queue_manager *dqm,
1025 				struct queue *q)
1026 {
1027 	int bit;
1028 
1029 	if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
1030 		if (dqm->sdma_bitmap == 0) {
1031 			pr_err("No more SDMA queue to allocate\n");
1032 			return -ENOMEM;
1033 		}
1034 
1035 		bit = __ffs64(dqm->sdma_bitmap);
1036 		dqm->sdma_bitmap &= ~(1ULL << bit);
1037 		q->sdma_id = bit;
1038 		q->properties.sdma_engine_id = q->sdma_id %
1039 				get_num_sdma_engines(dqm);
1040 		q->properties.sdma_queue_id = q->sdma_id /
1041 				get_num_sdma_engines(dqm);
1042 	} else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
1043 		if (dqm->xgmi_sdma_bitmap == 0) {
1044 			pr_err("No more XGMI SDMA queue to allocate\n");
1045 			return -ENOMEM;
1046 		}
1047 		bit = __ffs64(dqm->xgmi_sdma_bitmap);
1048 		dqm->xgmi_sdma_bitmap &= ~(1ULL << bit);
1049 		q->sdma_id = bit;
1050 		/* sdma_engine_id is sdma id including
1051 		 * both PCIe-optimized SDMAs and XGMI-
1052 		 * optimized SDMAs. The calculation below
1053 		 * assumes the first N engines are always
1054 		 * PCIe-optimized ones
1055 		 */
1056 		q->properties.sdma_engine_id = get_num_sdma_engines(dqm) +
1057 				q->sdma_id % get_num_xgmi_sdma_engines(dqm);
1058 		q->properties.sdma_queue_id = q->sdma_id /
1059 				get_num_xgmi_sdma_engines(dqm);
1060 	}
1061 
1062 	pr_debug("SDMA engine id: %d\n", q->properties.sdma_engine_id);
1063 	pr_debug("SDMA queue id: %d\n", q->properties.sdma_queue_id);
1064 
1065 	return 0;
1066 }
1067 
1068 static void deallocate_sdma_queue(struct device_queue_manager *dqm,
1069 				struct queue *q)
1070 {
1071 	if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
1072 		if (q->sdma_id >= get_num_sdma_queues(dqm))
1073 			return;
1074 		dqm->sdma_bitmap |= (1ULL << q->sdma_id);
1075 	} else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
1076 		if (q->sdma_id >= get_num_xgmi_sdma_queues(dqm))
1077 			return;
1078 		dqm->xgmi_sdma_bitmap |= (1ULL << q->sdma_id);
1079 	}
1080 }
1081 
1082 /*
1083  * Device Queue Manager implementation for cp scheduler
1084  */
1085 
1086 static int set_sched_resources(struct device_queue_manager *dqm)
1087 {
1088 	int i, mec;
1089 	struct scheduling_resources res;
1090 
1091 	res.vmid_mask = dqm->dev->shared_resources.compute_vmid_bitmap;
1092 
1093 	res.queue_mask = 0;
1094 	for (i = 0; i < KGD_MAX_QUEUES; ++i) {
1095 		mec = (i / dqm->dev->shared_resources.num_queue_per_pipe)
1096 			/ dqm->dev->shared_resources.num_pipe_per_mec;
1097 
1098 		if (!test_bit(i, dqm->dev->shared_resources.cp_queue_bitmap))
1099 			continue;
1100 
1101 		/* only acquire queues from the first MEC */
1102 		if (mec > 0)
1103 			continue;
1104 
1105 		/* This situation may be hit in the future if a new HW
1106 		 * generation exposes more than 64 queues. If so, the
1107 		 * definition of res.queue_mask needs updating
1108 		 */
1109 		if (WARN_ON(i >= (sizeof(res.queue_mask)*8))) {
1110 			pr_err("Invalid queue enabled by amdgpu: %d\n", i);
1111 			break;
1112 		}
1113 
1114 		res.queue_mask |= 1ull
1115 			<< amdgpu_queue_mask_bit_to_set_resource_bit(
1116 				(struct amdgpu_device *)dqm->dev->kgd, i);
1117 	}
1118 	res.gws_mask = ~0ull;
1119 	res.oac_mask = res.gds_heap_base = res.gds_heap_size = 0;
1120 
1121 	pr_debug("Scheduling resources:\n"
1122 			"vmid mask: 0x%8X\n"
1123 			"queue mask: 0x%8llX\n",
1124 			res.vmid_mask, res.queue_mask);
1125 
1126 	return pm_send_set_resources(&dqm->packets, &res);
1127 }
1128 
1129 static int initialize_cpsch(struct device_queue_manager *dqm)
1130 {
1131 	uint64_t num_sdma_queues;
1132 	uint64_t num_xgmi_sdma_queues;
1133 
1134 	pr_debug("num of pipes: %d\n", get_pipes_per_mec(dqm));
1135 
1136 	mutex_init(&dqm->lock_hidden);
1137 	INIT_LIST_HEAD(&dqm->queues);
1138 	dqm->active_queue_count = dqm->processes_count = 0;
1139 	dqm->active_cp_queue_count = 0;
1140 	dqm->gws_queue_count = 0;
1141 	dqm->active_runlist = false;
1142 
1143 	num_sdma_queues = get_num_sdma_queues(dqm);
1144 	if (num_sdma_queues >= BITS_PER_TYPE(dqm->sdma_bitmap))
1145 		dqm->sdma_bitmap = ULLONG_MAX;
1146 	else
1147 		dqm->sdma_bitmap = (BIT_ULL(num_sdma_queues) - 1);
1148 
1149 	num_xgmi_sdma_queues = get_num_xgmi_sdma_queues(dqm);
1150 	if (num_xgmi_sdma_queues >= BITS_PER_TYPE(dqm->xgmi_sdma_bitmap))
1151 		dqm->xgmi_sdma_bitmap = ULLONG_MAX;
1152 	else
1153 		dqm->xgmi_sdma_bitmap = (BIT_ULL(num_xgmi_sdma_queues) - 1);
1154 
1155 	INIT_WORK(&dqm->hw_exception_work, kfd_process_hw_exception);
1156 
1157 	return 0;
1158 }
1159 
1160 static int start_cpsch(struct device_queue_manager *dqm)
1161 {
1162 	int retval;
1163 
1164 	retval = 0;
1165 
1166 	retval = pm_init(&dqm->packets, dqm);
1167 	if (retval)
1168 		goto fail_packet_manager_init;
1169 
1170 	retval = set_sched_resources(dqm);
1171 	if (retval)
1172 		goto fail_set_sched_resources;
1173 
1174 	pr_debug("Allocating fence memory\n");
1175 
1176 	/* allocate fence memory on the gart */
1177 	retval = kfd_gtt_sa_allocate(dqm->dev, sizeof(*dqm->fence_addr),
1178 					&dqm->fence_mem);
1179 
1180 	if (retval)
1181 		goto fail_allocate_vidmem;
1182 
1183 	dqm->fence_addr = (uint64_t *)dqm->fence_mem->cpu_ptr;
1184 	dqm->fence_gpu_addr = dqm->fence_mem->gpu_addr;
1185 
1186 	init_interrupts(dqm);
1187 
1188 	dqm_lock(dqm);
1189 	/* clear hang status when driver try to start the hw scheduler */
1190 	dqm->is_hws_hang = false;
1191 	dqm->is_resetting = false;
1192 	dqm->sched_running = true;
1193 	execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
1194 	dqm_unlock(dqm);
1195 
1196 	return 0;
1197 fail_allocate_vidmem:
1198 fail_set_sched_resources:
1199 	pm_uninit(&dqm->packets, false);
1200 fail_packet_manager_init:
1201 	return retval;
1202 }
1203 
1204 static int stop_cpsch(struct device_queue_manager *dqm)
1205 {
1206 	bool hanging;
1207 
1208 	dqm_lock(dqm);
1209 	if (!dqm->is_hws_hang)
1210 		unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0);
1211 	hanging = dqm->is_hws_hang || dqm->is_resetting;
1212 	dqm->sched_running = false;
1213 	dqm_unlock(dqm);
1214 
1215 	pm_release_ib(&dqm->packets);
1216 
1217 	kfd_gtt_sa_free(dqm->dev, dqm->fence_mem);
1218 	pm_uninit(&dqm->packets, hanging);
1219 
1220 	return 0;
1221 }
1222 
1223 static int create_kernel_queue_cpsch(struct device_queue_manager *dqm,
1224 					struct kernel_queue *kq,
1225 					struct qcm_process_device *qpd)
1226 {
1227 	dqm_lock(dqm);
1228 	if (dqm->total_queue_count >= max_num_of_queues_per_device) {
1229 		pr_warn("Can't create new kernel queue because %d queues were already created\n",
1230 				dqm->total_queue_count);
1231 		dqm_unlock(dqm);
1232 		return -EPERM;
1233 	}
1234 
1235 	/*
1236 	 * Unconditionally increment this counter, regardless of the queue's
1237 	 * type or whether the queue is active.
1238 	 */
1239 	dqm->total_queue_count++;
1240 	pr_debug("Total of %d queues are accountable so far\n",
1241 			dqm->total_queue_count);
1242 
1243 	list_add(&kq->list, &qpd->priv_queue_list);
1244 	increment_queue_count(dqm, kq->queue->properties.type);
1245 	qpd->is_debug = true;
1246 	execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
1247 	dqm_unlock(dqm);
1248 
1249 	return 0;
1250 }
1251 
1252 static void destroy_kernel_queue_cpsch(struct device_queue_manager *dqm,
1253 					struct kernel_queue *kq,
1254 					struct qcm_process_device *qpd)
1255 {
1256 	dqm_lock(dqm);
1257 	list_del(&kq->list);
1258 	decrement_queue_count(dqm, kq->queue->properties.type);
1259 	qpd->is_debug = false;
1260 	execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0);
1261 	/*
1262 	 * Unconditionally decrement this counter, regardless of the queue's
1263 	 * type.
1264 	 */
1265 	dqm->total_queue_count--;
1266 	pr_debug("Total of %d queues are accountable so far\n",
1267 			dqm->total_queue_count);
1268 	dqm_unlock(dqm);
1269 }
1270 
1271 static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q,
1272 			struct qcm_process_device *qpd)
1273 {
1274 	int retval;
1275 	struct mqd_manager *mqd_mgr;
1276 
1277 	if (dqm->total_queue_count >= max_num_of_queues_per_device) {
1278 		pr_warn("Can't create new usermode queue because %d queues were already created\n",
1279 				dqm->total_queue_count);
1280 		retval = -EPERM;
1281 		goto out;
1282 	}
1283 
1284 	if (q->properties.type == KFD_QUEUE_TYPE_SDMA ||
1285 		q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
1286 		dqm_lock(dqm);
1287 		retval = allocate_sdma_queue(dqm, q);
1288 		dqm_unlock(dqm);
1289 		if (retval)
1290 			goto out;
1291 	}
1292 
1293 	retval = allocate_doorbell(qpd, q);
1294 	if (retval)
1295 		goto out_deallocate_sdma_queue;
1296 
1297 	mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
1298 			q->properties.type)];
1299 
1300 	if (q->properties.type == KFD_QUEUE_TYPE_SDMA ||
1301 		q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)
1302 		dqm->asic_ops.init_sdma_vm(dqm, q, qpd);
1303 	q->properties.tba_addr = qpd->tba_addr;
1304 	q->properties.tma_addr = qpd->tma_addr;
1305 	q->mqd_mem_obj = mqd_mgr->allocate_mqd(mqd_mgr->dev, &q->properties);
1306 	if (!q->mqd_mem_obj) {
1307 		retval = -ENOMEM;
1308 		goto out_deallocate_doorbell;
1309 	}
1310 
1311 	dqm_lock(dqm);
1312 	/*
1313 	 * Eviction state logic: mark all queues as evicted, even ones
1314 	 * not currently active. Restoring inactive queues later only
1315 	 * updates the is_evicted flag but is a no-op otherwise.
1316 	 */
1317 	q->properties.is_evicted = !!qpd->evicted;
1318 	mqd_mgr->init_mqd(mqd_mgr, &q->mqd, q->mqd_mem_obj,
1319 				&q->gart_mqd_addr, &q->properties);
1320 
1321 	list_add(&q->list, &qpd->queues_list);
1322 	qpd->queue_count++;
1323 
1324 	if (q->properties.is_active) {
1325 		increment_queue_count(dqm, q->properties.type);
1326 
1327 		execute_queues_cpsch(dqm,
1328 				KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
1329 	}
1330 
1331 	/*
1332 	 * Unconditionally increment this counter, regardless of the queue's
1333 	 * type or whether the queue is active.
1334 	 */
1335 	dqm->total_queue_count++;
1336 
1337 	pr_debug("Total of %d queues are accountable so far\n",
1338 			dqm->total_queue_count);
1339 
1340 	dqm_unlock(dqm);
1341 	return retval;
1342 
1343 out_deallocate_doorbell:
1344 	deallocate_doorbell(qpd, q);
1345 out_deallocate_sdma_queue:
1346 	if (q->properties.type == KFD_QUEUE_TYPE_SDMA ||
1347 		q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
1348 		dqm_lock(dqm);
1349 		deallocate_sdma_queue(dqm, q);
1350 		dqm_unlock(dqm);
1351 	}
1352 out:
1353 	return retval;
1354 }
1355 
1356 int amdkfd_fence_wait_timeout(uint64_t *fence_addr,
1357 				uint64_t fence_value,
1358 				unsigned int timeout_ms)
1359 {
1360 	unsigned long end_jiffies = msecs_to_jiffies(timeout_ms) + jiffies;
1361 
1362 	while (*fence_addr != fence_value) {
1363 		if (time_after(jiffies, end_jiffies)) {
1364 			pr_err("qcm fence wait loop timeout expired\n");
1365 			/* In HWS case, this is used to halt the driver thread
1366 			 * in order not to mess up CP states before doing
1367 			 * scandumps for FW debugging.
1368 			 */
1369 			while (halt_if_hws_hang)
1370 				schedule();
1371 
1372 			return -ETIME;
1373 		}
1374 		schedule();
1375 	}
1376 
1377 	return 0;
1378 }
1379 
1380 /* dqm->lock mutex has to be locked before calling this function */
1381 static int map_queues_cpsch(struct device_queue_manager *dqm)
1382 {
1383 	int retval;
1384 
1385 	if (!dqm->sched_running)
1386 		return 0;
1387 	if (dqm->active_queue_count <= 0 || dqm->processes_count <= 0)
1388 		return 0;
1389 	if (dqm->active_runlist)
1390 		return 0;
1391 
1392 	retval = pm_send_runlist(&dqm->packets, &dqm->queues);
1393 	pr_debug("%s sent runlist\n", __func__);
1394 	if (retval) {
1395 		pr_err("failed to execute runlist\n");
1396 		return retval;
1397 	}
1398 	dqm->active_runlist = true;
1399 
1400 	return retval;
1401 }
1402 
1403 /* dqm->lock mutex has to be locked before calling this function */
1404 static int unmap_queues_cpsch(struct device_queue_manager *dqm,
1405 				enum kfd_unmap_queues_filter filter,
1406 				uint32_t filter_param)
1407 {
1408 	int retval = 0;
1409 	struct mqd_manager *mqd_mgr;
1410 
1411 	if (!dqm->sched_running)
1412 		return 0;
1413 	if (dqm->is_hws_hang)
1414 		return -EIO;
1415 	if (!dqm->active_runlist)
1416 		return retval;
1417 
1418 	retval = pm_send_unmap_queue(&dqm->packets, KFD_QUEUE_TYPE_COMPUTE,
1419 			filter, filter_param, false, 0);
1420 	if (retval)
1421 		return retval;
1422 
1423 	*dqm->fence_addr = KFD_FENCE_INIT;
1424 	pm_send_query_status(&dqm->packets, dqm->fence_gpu_addr,
1425 				KFD_FENCE_COMPLETED);
1426 	/* should be timed out */
1427 	retval = amdkfd_fence_wait_timeout(dqm->fence_addr, KFD_FENCE_COMPLETED,
1428 				queue_preemption_timeout_ms);
1429 	if (retval) {
1430 		pr_err("The cp might be in an unrecoverable state due to an unsuccessful queues preemption\n");
1431 		dqm->is_hws_hang = true;
1432 		/* It's possible we're detecting a HWS hang in the
1433 		 * middle of a GPU reset. No need to schedule another
1434 		 * reset in this case.
1435 		 */
1436 		if (!dqm->is_resetting)
1437 			schedule_work(&dqm->hw_exception_work);
1438 		return retval;
1439 	}
1440 
1441 	/* In the current MEC firmware implementation, if compute queue
1442 	 * doesn't response to the preemption request in time, HIQ will
1443 	 * abandon the unmap request without returning any timeout error
1444 	 * to driver. Instead, MEC firmware will log the doorbell of the
1445 	 * unresponding compute queue to HIQ.MQD.queue_doorbell_id fields.
1446 	 * To make sure the queue unmap was successful, driver need to
1447 	 * check those fields
1448 	 */
1449 	mqd_mgr = dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ];
1450 	if (mqd_mgr->read_doorbell_id(dqm->packets.priv_queue->queue->mqd)) {
1451 		pr_err("HIQ MQD's queue_doorbell_id0 is not 0, Queue preemption time out\n");
1452 		while (halt_if_hws_hang)
1453 			schedule();
1454 		return -ETIME;
1455 	}
1456 
1457 	pm_release_ib(&dqm->packets);
1458 	dqm->active_runlist = false;
1459 
1460 	return retval;
1461 }
1462 
1463 /* dqm->lock mutex has to be locked before calling this function */
1464 static int execute_queues_cpsch(struct device_queue_manager *dqm,
1465 				enum kfd_unmap_queues_filter filter,
1466 				uint32_t filter_param)
1467 {
1468 	int retval;
1469 
1470 	if (dqm->is_hws_hang)
1471 		return -EIO;
1472 	retval = unmap_queues_cpsch(dqm, filter, filter_param);
1473 	if (retval)
1474 		return retval;
1475 
1476 	return map_queues_cpsch(dqm);
1477 }
1478 
1479 static int destroy_queue_cpsch(struct device_queue_manager *dqm,
1480 				struct qcm_process_device *qpd,
1481 				struct queue *q)
1482 {
1483 	int retval;
1484 	struct mqd_manager *mqd_mgr;
1485 	uint64_t sdma_val = 0;
1486 	struct kfd_process_device *pdd = qpd_to_pdd(qpd);
1487 
1488 	/* Get the SDMA queue stats */
1489 	if ((q->properties.type == KFD_QUEUE_TYPE_SDMA) ||
1490 	    (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) {
1491 		retval = read_sdma_queue_counter((uint64_t __user *)q->properties.read_ptr,
1492 							&sdma_val);
1493 		if (retval)
1494 			pr_err("Failed to read SDMA queue counter for queue: %d\n",
1495 				q->properties.queue_id);
1496 	}
1497 
1498 	retval = 0;
1499 
1500 	/* remove queue from list to prevent rescheduling after preemption */
1501 	dqm_lock(dqm);
1502 
1503 	if (qpd->is_debug) {
1504 		/*
1505 		 * error, currently we do not allow to destroy a queue
1506 		 * of a currently debugged process
1507 		 */
1508 		retval = -EBUSY;
1509 		goto failed_try_destroy_debugged_queue;
1510 
1511 	}
1512 
1513 	mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
1514 			q->properties.type)];
1515 
1516 	deallocate_doorbell(qpd, q);
1517 
1518 	if ((q->properties.type == KFD_QUEUE_TYPE_SDMA) ||
1519 	    (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) {
1520 		deallocate_sdma_queue(dqm, q);
1521 		pdd->sdma_past_activity_counter += sdma_val;
1522 	}
1523 
1524 	list_del(&q->list);
1525 	qpd->queue_count--;
1526 	if (q->properties.is_active) {
1527 		decrement_queue_count(dqm, q->properties.type);
1528 		retval = execute_queues_cpsch(dqm,
1529 				KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
1530 		if (retval == -ETIME)
1531 			qpd->reset_wavefronts = true;
1532 		if (q->properties.is_gws) {
1533 			dqm->gws_queue_count--;
1534 			qpd->mapped_gws_queue = false;
1535 		}
1536 	}
1537 
1538 	/*
1539 	 * Unconditionally decrement this counter, regardless of the queue's
1540 	 * type
1541 	 */
1542 	dqm->total_queue_count--;
1543 	pr_debug("Total of %d queues are accountable so far\n",
1544 			dqm->total_queue_count);
1545 
1546 	dqm_unlock(dqm);
1547 
1548 	/* Do free_mqd after dqm_unlock(dqm) to avoid circular locking */
1549 	mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
1550 
1551 	return retval;
1552 
1553 failed_try_destroy_debugged_queue:
1554 
1555 	dqm_unlock(dqm);
1556 	return retval;
1557 }
1558 
1559 /*
1560  * Low bits must be 0000/FFFF as required by HW, high bits must be 0 to
1561  * stay in user mode.
1562  */
1563 #define APE1_FIXED_BITS_MASK 0xFFFF80000000FFFFULL
1564 /* APE1 limit is inclusive and 64K aligned. */
1565 #define APE1_LIMIT_ALIGNMENT 0xFFFF
1566 
1567 static bool set_cache_memory_policy(struct device_queue_manager *dqm,
1568 				   struct qcm_process_device *qpd,
1569 				   enum cache_policy default_policy,
1570 				   enum cache_policy alternate_policy,
1571 				   void __user *alternate_aperture_base,
1572 				   uint64_t alternate_aperture_size)
1573 {
1574 	bool retval = true;
1575 
1576 	if (!dqm->asic_ops.set_cache_memory_policy)
1577 		return retval;
1578 
1579 	dqm_lock(dqm);
1580 
1581 	if (alternate_aperture_size == 0) {
1582 		/* base > limit disables APE1 */
1583 		qpd->sh_mem_ape1_base = 1;
1584 		qpd->sh_mem_ape1_limit = 0;
1585 	} else {
1586 		/*
1587 		 * In FSA64, APE1_Base[63:0] = { 16{SH_MEM_APE1_BASE[31]},
1588 		 *			SH_MEM_APE1_BASE[31:0], 0x0000 }
1589 		 * APE1_Limit[63:0] = { 16{SH_MEM_APE1_LIMIT[31]},
1590 		 *			SH_MEM_APE1_LIMIT[31:0], 0xFFFF }
1591 		 * Verify that the base and size parameters can be
1592 		 * represented in this format and convert them.
1593 		 * Additionally restrict APE1 to user-mode addresses.
1594 		 */
1595 
1596 		uint64_t base = (uintptr_t)alternate_aperture_base;
1597 		uint64_t limit = base + alternate_aperture_size - 1;
1598 
1599 		if (limit <= base || (base & APE1_FIXED_BITS_MASK) != 0 ||
1600 		   (limit & APE1_FIXED_BITS_MASK) != APE1_LIMIT_ALIGNMENT) {
1601 			retval = false;
1602 			goto out;
1603 		}
1604 
1605 		qpd->sh_mem_ape1_base = base >> 16;
1606 		qpd->sh_mem_ape1_limit = limit >> 16;
1607 	}
1608 
1609 	retval = dqm->asic_ops.set_cache_memory_policy(
1610 			dqm,
1611 			qpd,
1612 			default_policy,
1613 			alternate_policy,
1614 			alternate_aperture_base,
1615 			alternate_aperture_size);
1616 
1617 	if ((dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) && (qpd->vmid != 0))
1618 		program_sh_mem_settings(dqm, qpd);
1619 
1620 	pr_debug("sh_mem_config: 0x%x, ape1_base: 0x%x, ape1_limit: 0x%x\n",
1621 		qpd->sh_mem_config, qpd->sh_mem_ape1_base,
1622 		qpd->sh_mem_ape1_limit);
1623 
1624 out:
1625 	dqm_unlock(dqm);
1626 	return retval;
1627 }
1628 
1629 static int process_termination_nocpsch(struct device_queue_manager *dqm,
1630 		struct qcm_process_device *qpd)
1631 {
1632 	struct queue *q, *next;
1633 	struct device_process_node *cur, *next_dpn;
1634 	int retval = 0;
1635 	bool found = false;
1636 
1637 	dqm_lock(dqm);
1638 
1639 	/* Clear all user mode queues */
1640 	list_for_each_entry_safe(q, next, &qpd->queues_list, list) {
1641 		int ret;
1642 
1643 		ret = destroy_queue_nocpsch_locked(dqm, qpd, q);
1644 		if (ret)
1645 			retval = ret;
1646 	}
1647 
1648 	/* Unregister process */
1649 	list_for_each_entry_safe(cur, next_dpn, &dqm->queues, list) {
1650 		if (qpd == cur->qpd) {
1651 			list_del(&cur->list);
1652 			kfree(cur);
1653 			dqm->processes_count--;
1654 			found = true;
1655 			break;
1656 		}
1657 	}
1658 
1659 	dqm_unlock(dqm);
1660 
1661 	/* Outside the DQM lock because under the DQM lock we can't do
1662 	 * reclaim or take other locks that others hold while reclaiming.
1663 	 */
1664 	if (found)
1665 		kfd_dec_compute_active(dqm->dev);
1666 
1667 	return retval;
1668 }
1669 
1670 static int get_wave_state(struct device_queue_manager *dqm,
1671 			  struct queue *q,
1672 			  void __user *ctl_stack,
1673 			  u32 *ctl_stack_used_size,
1674 			  u32 *save_area_used_size)
1675 {
1676 	struct mqd_manager *mqd_mgr;
1677 	int r;
1678 
1679 	dqm_lock(dqm);
1680 
1681 	if (q->properties.type != KFD_QUEUE_TYPE_COMPUTE ||
1682 	    q->properties.is_active || !q->device->cwsr_enabled) {
1683 		r = -EINVAL;
1684 		goto dqm_unlock;
1685 	}
1686 
1687 	mqd_mgr = dqm->mqd_mgrs[KFD_MQD_TYPE_CP];
1688 
1689 	if (!mqd_mgr->get_wave_state) {
1690 		r = -EINVAL;
1691 		goto dqm_unlock;
1692 	}
1693 
1694 	r = mqd_mgr->get_wave_state(mqd_mgr, q->mqd, ctl_stack,
1695 			ctl_stack_used_size, save_area_used_size);
1696 
1697 dqm_unlock:
1698 	dqm_unlock(dqm);
1699 	return r;
1700 }
1701 
1702 static int process_termination_cpsch(struct device_queue_manager *dqm,
1703 		struct qcm_process_device *qpd)
1704 {
1705 	int retval;
1706 	struct queue *q, *next;
1707 	struct kernel_queue *kq, *kq_next;
1708 	struct mqd_manager *mqd_mgr;
1709 	struct device_process_node *cur, *next_dpn;
1710 	enum kfd_unmap_queues_filter filter =
1711 		KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES;
1712 	bool found = false;
1713 
1714 	retval = 0;
1715 
1716 	dqm_lock(dqm);
1717 
1718 	/* Clean all kernel queues */
1719 	list_for_each_entry_safe(kq, kq_next, &qpd->priv_queue_list, list) {
1720 		list_del(&kq->list);
1721 		decrement_queue_count(dqm, kq->queue->properties.type);
1722 		qpd->is_debug = false;
1723 		dqm->total_queue_count--;
1724 		filter = KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES;
1725 	}
1726 
1727 	/* Clear all user mode queues */
1728 	list_for_each_entry(q, &qpd->queues_list, list) {
1729 		if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
1730 			deallocate_sdma_queue(dqm, q);
1731 		else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)
1732 			deallocate_sdma_queue(dqm, q);
1733 
1734 		if (q->properties.is_active) {
1735 			decrement_queue_count(dqm, q->properties.type);
1736 			if (q->properties.is_gws) {
1737 				dqm->gws_queue_count--;
1738 				qpd->mapped_gws_queue = false;
1739 			}
1740 		}
1741 
1742 		dqm->total_queue_count--;
1743 	}
1744 
1745 	/* Unregister process */
1746 	list_for_each_entry_safe(cur, next_dpn, &dqm->queues, list) {
1747 		if (qpd == cur->qpd) {
1748 			list_del(&cur->list);
1749 			kfree(cur);
1750 			dqm->processes_count--;
1751 			found = true;
1752 			break;
1753 		}
1754 	}
1755 
1756 	retval = execute_queues_cpsch(dqm, filter, 0);
1757 	if ((!dqm->is_hws_hang) && (retval || qpd->reset_wavefronts)) {
1758 		pr_warn("Resetting wave fronts (cpsch) on dev %p\n", dqm->dev);
1759 		dbgdev_wave_reset_wavefronts(dqm->dev, qpd->pqm->process);
1760 		qpd->reset_wavefronts = false;
1761 	}
1762 
1763 	dqm_unlock(dqm);
1764 
1765 	/* Outside the DQM lock because under the DQM lock we can't do
1766 	 * reclaim or take other locks that others hold while reclaiming.
1767 	 */
1768 	if (found)
1769 		kfd_dec_compute_active(dqm->dev);
1770 
1771 	/* Lastly, free mqd resources.
1772 	 * Do free_mqd() after dqm_unlock to avoid circular locking.
1773 	 */
1774 	list_for_each_entry_safe(q, next, &qpd->queues_list, list) {
1775 		mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
1776 				q->properties.type)];
1777 		list_del(&q->list);
1778 		qpd->queue_count--;
1779 		mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
1780 	}
1781 
1782 	return retval;
1783 }
1784 
1785 static int init_mqd_managers(struct device_queue_manager *dqm)
1786 {
1787 	int i, j;
1788 	struct mqd_manager *mqd_mgr;
1789 
1790 	for (i = 0; i < KFD_MQD_TYPE_MAX; i++) {
1791 		mqd_mgr = dqm->asic_ops.mqd_manager_init(i, dqm->dev);
1792 		if (!mqd_mgr) {
1793 			pr_err("mqd manager [%d] initialization failed\n", i);
1794 			goto out_free;
1795 		}
1796 		dqm->mqd_mgrs[i] = mqd_mgr;
1797 	}
1798 
1799 	return 0;
1800 
1801 out_free:
1802 	for (j = 0; j < i; j++) {
1803 		kfree(dqm->mqd_mgrs[j]);
1804 		dqm->mqd_mgrs[j] = NULL;
1805 	}
1806 
1807 	return -ENOMEM;
1808 }
1809 
1810 /* Allocate one hiq mqd (HWS) and all SDMA mqd in a continuous trunk*/
1811 static int allocate_hiq_sdma_mqd(struct device_queue_manager *dqm)
1812 {
1813 	int retval;
1814 	struct kfd_dev *dev = dqm->dev;
1815 	struct kfd_mem_obj *mem_obj = &dqm->hiq_sdma_mqd;
1816 	uint32_t size = dqm->mqd_mgrs[KFD_MQD_TYPE_SDMA]->mqd_size *
1817 		get_num_all_sdma_engines(dqm) *
1818 		dev->device_info->num_sdma_queues_per_engine +
1819 		dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ]->mqd_size;
1820 
1821 	retval = amdgpu_amdkfd_alloc_gtt_mem(dev->kgd, size,
1822 		&(mem_obj->gtt_mem), &(mem_obj->gpu_addr),
1823 		(void *)&(mem_obj->cpu_ptr), false);
1824 
1825 	return retval;
1826 }
1827 
1828 struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev)
1829 {
1830 	struct device_queue_manager *dqm;
1831 
1832 	pr_debug("Loading device queue manager\n");
1833 
1834 	dqm = kzalloc(sizeof(*dqm), GFP_KERNEL);
1835 	if (!dqm)
1836 		return NULL;
1837 
1838 	switch (dev->device_info->asic_family) {
1839 	/* HWS is not available on Hawaii. */
1840 	case CHIP_HAWAII:
1841 	/* HWS depends on CWSR for timely dequeue. CWSR is not
1842 	 * available on Tonga.
1843 	 *
1844 	 * FIXME: This argument also applies to Kaveri.
1845 	 */
1846 	case CHIP_TONGA:
1847 		dqm->sched_policy = KFD_SCHED_POLICY_NO_HWS;
1848 		break;
1849 	default:
1850 		dqm->sched_policy = sched_policy;
1851 		break;
1852 	}
1853 
1854 	dqm->dev = dev;
1855 	switch (dqm->sched_policy) {
1856 	case KFD_SCHED_POLICY_HWS:
1857 	case KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION:
1858 		/* initialize dqm for cp scheduling */
1859 		dqm->ops.create_queue = create_queue_cpsch;
1860 		dqm->ops.initialize = initialize_cpsch;
1861 		dqm->ops.start = start_cpsch;
1862 		dqm->ops.stop = stop_cpsch;
1863 		dqm->ops.pre_reset = pre_reset;
1864 		dqm->ops.destroy_queue = destroy_queue_cpsch;
1865 		dqm->ops.update_queue = update_queue;
1866 		dqm->ops.register_process = register_process;
1867 		dqm->ops.unregister_process = unregister_process;
1868 		dqm->ops.uninitialize = uninitialize;
1869 		dqm->ops.create_kernel_queue = create_kernel_queue_cpsch;
1870 		dqm->ops.destroy_kernel_queue = destroy_kernel_queue_cpsch;
1871 		dqm->ops.set_cache_memory_policy = set_cache_memory_policy;
1872 		dqm->ops.process_termination = process_termination_cpsch;
1873 		dqm->ops.evict_process_queues = evict_process_queues_cpsch;
1874 		dqm->ops.restore_process_queues = restore_process_queues_cpsch;
1875 		dqm->ops.get_wave_state = get_wave_state;
1876 		break;
1877 	case KFD_SCHED_POLICY_NO_HWS:
1878 		/* initialize dqm for no cp scheduling */
1879 		dqm->ops.start = start_nocpsch;
1880 		dqm->ops.stop = stop_nocpsch;
1881 		dqm->ops.pre_reset = pre_reset;
1882 		dqm->ops.create_queue = create_queue_nocpsch;
1883 		dqm->ops.destroy_queue = destroy_queue_nocpsch;
1884 		dqm->ops.update_queue = update_queue;
1885 		dqm->ops.register_process = register_process;
1886 		dqm->ops.unregister_process = unregister_process;
1887 		dqm->ops.initialize = initialize_nocpsch;
1888 		dqm->ops.uninitialize = uninitialize;
1889 		dqm->ops.set_cache_memory_policy = set_cache_memory_policy;
1890 		dqm->ops.process_termination = process_termination_nocpsch;
1891 		dqm->ops.evict_process_queues = evict_process_queues_nocpsch;
1892 		dqm->ops.restore_process_queues =
1893 			restore_process_queues_nocpsch;
1894 		dqm->ops.get_wave_state = get_wave_state;
1895 		break;
1896 	default:
1897 		pr_err("Invalid scheduling policy %d\n", dqm->sched_policy);
1898 		goto out_free;
1899 	}
1900 
1901 	switch (dev->device_info->asic_family) {
1902 	case CHIP_CARRIZO:
1903 		device_queue_manager_init_vi(&dqm->asic_ops);
1904 		break;
1905 
1906 	case CHIP_KAVERI:
1907 		device_queue_manager_init_cik(&dqm->asic_ops);
1908 		break;
1909 
1910 	case CHIP_HAWAII:
1911 		device_queue_manager_init_cik_hawaii(&dqm->asic_ops);
1912 		break;
1913 
1914 	case CHIP_TONGA:
1915 	case CHIP_FIJI:
1916 	case CHIP_POLARIS10:
1917 	case CHIP_POLARIS11:
1918 	case CHIP_POLARIS12:
1919 	case CHIP_VEGAM:
1920 		device_queue_manager_init_vi_tonga(&dqm->asic_ops);
1921 		break;
1922 
1923 	case CHIP_VEGA10:
1924 	case CHIP_VEGA12:
1925 	case CHIP_VEGA20:
1926 	case CHIP_RAVEN:
1927 	case CHIP_RENOIR:
1928 	case CHIP_ARCTURUS:
1929 	case CHIP_ALDEBARAN:
1930 		device_queue_manager_init_v9(&dqm->asic_ops);
1931 		break;
1932 	case CHIP_NAVI10:
1933 	case CHIP_NAVI12:
1934 	case CHIP_NAVI14:
1935 	case CHIP_SIENNA_CICHLID:
1936 	case CHIP_NAVY_FLOUNDER:
1937 	case CHIP_VANGOGH:
1938 	case CHIP_DIMGREY_CAVEFISH:
1939 	case CHIP_BEIGE_GOBY:
1940 		device_queue_manager_init_v10_navi10(&dqm->asic_ops);
1941 		break;
1942 	default:
1943 		WARN(1, "Unexpected ASIC family %u",
1944 		     dev->device_info->asic_family);
1945 		goto out_free;
1946 	}
1947 
1948 	if (init_mqd_managers(dqm))
1949 		goto out_free;
1950 
1951 	if (allocate_hiq_sdma_mqd(dqm)) {
1952 		pr_err("Failed to allocate hiq sdma mqd trunk buffer\n");
1953 		goto out_free;
1954 	}
1955 
1956 	if (!dqm->ops.initialize(dqm))
1957 		return dqm;
1958 
1959 out_free:
1960 	kfree(dqm);
1961 	return NULL;
1962 }
1963 
1964 static void deallocate_hiq_sdma_mqd(struct kfd_dev *dev,
1965 				    struct kfd_mem_obj *mqd)
1966 {
1967 	WARN(!mqd, "No hiq sdma mqd trunk to free");
1968 
1969 	amdgpu_amdkfd_free_gtt_mem(dev->kgd, mqd->gtt_mem);
1970 }
1971 
1972 void device_queue_manager_uninit(struct device_queue_manager *dqm)
1973 {
1974 	dqm->ops.uninitialize(dqm);
1975 	deallocate_hiq_sdma_mqd(dqm->dev, &dqm->hiq_sdma_mqd);
1976 	kfree(dqm);
1977 }
1978 
1979 int kfd_process_vm_fault(struct device_queue_manager *dqm, u32 pasid)
1980 {
1981 	struct kfd_process_device *pdd;
1982 	struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
1983 	int ret = 0;
1984 
1985 	if (!p)
1986 		return -EINVAL;
1987 	WARN(debug_evictions, "Evicting pid %d", p->lead_thread->pid);
1988 	pdd = kfd_get_process_device_data(dqm->dev, p);
1989 	if (pdd)
1990 		ret = dqm->ops.evict_process_queues(dqm, &pdd->qpd);
1991 	kfd_unref_process(p);
1992 
1993 	return ret;
1994 }
1995 
1996 static void kfd_process_hw_exception(struct work_struct *work)
1997 {
1998 	struct device_queue_manager *dqm = container_of(work,
1999 			struct device_queue_manager, hw_exception_work);
2000 	amdgpu_amdkfd_gpu_reset(dqm->dev->kgd);
2001 }
2002 
2003 #if defined(CONFIG_DEBUG_FS)
2004 
2005 static void seq_reg_dump(struct seq_file *m,
2006 			 uint32_t (*dump)[2], uint32_t n_regs)
2007 {
2008 	uint32_t i, count;
2009 
2010 	for (i = 0, count = 0; i < n_regs; i++) {
2011 		if (count == 0 ||
2012 		    dump[i-1][0] + sizeof(uint32_t) != dump[i][0]) {
2013 			seq_printf(m, "%s    %08x: %08x",
2014 				   i ? "\n" : "",
2015 				   dump[i][0], dump[i][1]);
2016 			count = 7;
2017 		} else {
2018 			seq_printf(m, " %08x", dump[i][1]);
2019 			count--;
2020 		}
2021 	}
2022 
2023 	seq_puts(m, "\n");
2024 }
2025 
2026 int dqm_debugfs_hqds(struct seq_file *m, void *data)
2027 {
2028 	struct device_queue_manager *dqm = data;
2029 	uint32_t (*dump)[2], n_regs;
2030 	int pipe, queue;
2031 	int r = 0;
2032 
2033 	if (!dqm->sched_running) {
2034 		seq_printf(m, " Device is stopped\n");
2035 
2036 		return 0;
2037 	}
2038 
2039 	r = dqm->dev->kfd2kgd->hqd_dump(dqm->dev->kgd,
2040 					KFD_CIK_HIQ_PIPE, KFD_CIK_HIQ_QUEUE,
2041 					&dump, &n_regs);
2042 	if (!r) {
2043 		seq_printf(m, "  HIQ on MEC %d Pipe %d Queue %d\n",
2044 			   KFD_CIK_HIQ_PIPE/get_pipes_per_mec(dqm)+1,
2045 			   KFD_CIK_HIQ_PIPE%get_pipes_per_mec(dqm),
2046 			   KFD_CIK_HIQ_QUEUE);
2047 		seq_reg_dump(m, dump, n_regs);
2048 
2049 		kfree(dump);
2050 	}
2051 
2052 	for (pipe = 0; pipe < get_pipes_per_mec(dqm); pipe++) {
2053 		int pipe_offset = pipe * get_queues_per_pipe(dqm);
2054 
2055 		for (queue = 0; queue < get_queues_per_pipe(dqm); queue++) {
2056 			if (!test_bit(pipe_offset + queue,
2057 				      dqm->dev->shared_resources.cp_queue_bitmap))
2058 				continue;
2059 
2060 			r = dqm->dev->kfd2kgd->hqd_dump(
2061 				dqm->dev->kgd, pipe, queue, &dump, &n_regs);
2062 			if (r)
2063 				break;
2064 
2065 			seq_printf(m, "  CP Pipe %d, Queue %d\n",
2066 				  pipe, queue);
2067 			seq_reg_dump(m, dump, n_regs);
2068 
2069 			kfree(dump);
2070 		}
2071 	}
2072 
2073 	for (pipe = 0; pipe < get_num_all_sdma_engines(dqm); pipe++) {
2074 		for (queue = 0;
2075 		     queue < dqm->dev->device_info->num_sdma_queues_per_engine;
2076 		     queue++) {
2077 			r = dqm->dev->kfd2kgd->hqd_sdma_dump(
2078 				dqm->dev->kgd, pipe, queue, &dump, &n_regs);
2079 			if (r)
2080 				break;
2081 
2082 			seq_printf(m, "  SDMA Engine %d, RLC %d\n",
2083 				  pipe, queue);
2084 			seq_reg_dump(m, dump, n_regs);
2085 
2086 			kfree(dump);
2087 		}
2088 	}
2089 
2090 	return r;
2091 }
2092 
2093 int dqm_debugfs_execute_queues(struct device_queue_manager *dqm)
2094 {
2095 	int r = 0;
2096 
2097 	dqm_lock(dqm);
2098 	dqm->active_runlist = true;
2099 	r = execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0);
2100 	dqm_unlock(dqm);
2101 
2102 	return r;
2103 }
2104 
2105 #endif
2106