1 // SPDX-License-Identifier: GPL-2.0 OR MIT
2 /*
3  * Copyright 2014-2022 Advanced Micro Devices, Inc.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice shall be included in
13  * all copies or substantial portions of the Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
19  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
21  * OTHER DEALINGS IN THE SOFTWARE.
22  *
23  */
24 
25 #include <linux/slab.h>
26 #include <linux/list.h>
27 #include "kfd_device_queue_manager.h"
28 #include "kfd_priv.h"
29 #include "kfd_kernel_queue.h"
30 #include "amdgpu_amdkfd.h"
31 #include "amdgpu_reset.h"
32 
33 static inline struct process_queue_node *get_queue_by_qid(
34 			struct process_queue_manager *pqm, unsigned int qid)
35 {
36 	struct process_queue_node *pqn;
37 
38 	list_for_each_entry(pqn, &pqm->queues, process_queue_list) {
39 		if ((pqn->q && pqn->q->properties.queue_id == qid) ||
40 		    (pqn->kq && pqn->kq->queue->properties.queue_id == qid))
41 			return pqn;
42 	}
43 
44 	return NULL;
45 }
46 
47 static int assign_queue_slot_by_qid(struct process_queue_manager *pqm,
48 				    unsigned int qid)
49 {
50 	if (qid >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS)
51 		return -EINVAL;
52 
53 	if (__test_and_set_bit(qid, pqm->queue_slot_bitmap)) {
54 		pr_err("Cannot create new queue because requested qid(%u) is in use\n", qid);
55 		return -ENOSPC;
56 	}
57 
58 	return 0;
59 }
60 
61 static int find_available_queue_slot(struct process_queue_manager *pqm,
62 					unsigned int *qid)
63 {
64 	unsigned long found;
65 
66 	found = find_first_zero_bit(pqm->queue_slot_bitmap,
67 			KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);
68 
69 	pr_debug("The new slot id %lu\n", found);
70 
71 	if (found >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) {
72 		pr_info("Cannot open more queues for process with pasid 0x%x\n",
73 				pqm->process->pasid);
74 		return -ENOMEM;
75 	}
76 
77 	set_bit(found, pqm->queue_slot_bitmap);
78 	*qid = found;
79 
80 	return 0;
81 }
82 
83 void kfd_process_dequeue_from_device(struct kfd_process_device *pdd)
84 {
85 	struct kfd_node *dev = pdd->dev;
86 
87 	if (pdd->already_dequeued)
88 		return;
89 
90 	dev->dqm->ops.process_termination(dev->dqm, &pdd->qpd);
91 	if (dev->kfd->shared_resources.enable_mes &&
92 	    down_read_trylock(&dev->adev->reset_domain->sem)) {
93 		amdgpu_mes_flush_shader_debugger(dev->adev,
94 						 pdd->proc_ctx_gpu_addr);
95 		up_read(&dev->adev->reset_domain->sem);
96 	}
97 	pdd->already_dequeued = true;
98 }
99 
100 int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid,
101 			void *gws)
102 {
103 	struct kfd_node *dev = NULL;
104 	struct process_queue_node *pqn;
105 	struct kfd_process_device *pdd;
106 	struct kgd_mem *mem = NULL;
107 	int ret;
108 
109 	pqn = get_queue_by_qid(pqm, qid);
110 	if (!pqn) {
111 		pr_err("Queue id does not match any known queue\n");
112 		return -EINVAL;
113 	}
114 
115 	if (pqn->q)
116 		dev = pqn->q->device;
117 	if (WARN_ON(!dev))
118 		return -ENODEV;
119 
120 	pdd = kfd_get_process_device_data(dev, pqm->process);
121 	if (!pdd) {
122 		pr_err("Process device data doesn't exist\n");
123 		return -EINVAL;
124 	}
125 
126 	/* Only allow one queue per process can have GWS assigned */
127 	if (gws && pdd->qpd.num_gws)
128 		return -EBUSY;
129 
130 	if (!gws && pdd->qpd.num_gws == 0)
131 		return -EINVAL;
132 
133 	if (KFD_GC_VERSION(dev) != IP_VERSION(9, 4, 3) && !dev->kfd->shared_resources.enable_mes) {
134 		if (gws)
135 			ret = amdgpu_amdkfd_add_gws_to_process(pdd->process->kgd_process_info,
136 				gws, &mem);
137 		else
138 			ret = amdgpu_amdkfd_remove_gws_from_process(pdd->process->kgd_process_info,
139 				pqn->q->gws);
140 		if (unlikely(ret))
141 			return ret;
142 		pqn->q->gws = mem;
143 	} else {
144 		/*
145 		 * Intentionally set GWS to a non-NULL value
146 		 * for devices that do not use GWS for global wave
147 		 * synchronization but require the formality
148 		 * of setting GWS for cooperative groups.
149 		 */
150 		pqn->q->gws = gws ? ERR_PTR(-ENOMEM) : NULL;
151 	}
152 
153 	pdd->qpd.num_gws = gws ? dev->adev->gds.gws_size : 0;
154 
155 	return pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm,
156 							pqn->q, NULL);
157 }
158 
159 void kfd_process_dequeue_from_all_devices(struct kfd_process *p)
160 {
161 	int i;
162 
163 	for (i = 0; i < p->n_pdds; i++)
164 		kfd_process_dequeue_from_device(p->pdds[i]);
165 }
166 
167 int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p)
168 {
169 	INIT_LIST_HEAD(&pqm->queues);
170 	pqm->queue_slot_bitmap = bitmap_zalloc(KFD_MAX_NUM_OF_QUEUES_PER_PROCESS,
171 					       GFP_KERNEL);
172 	if (!pqm->queue_slot_bitmap)
173 		return -ENOMEM;
174 	pqm->process = p;
175 
176 	return 0;
177 }
178 
179 static void pqm_clean_queue_resource(struct process_queue_manager *pqm,
180 				     struct process_queue_node *pqn)
181 {
182 	struct kfd_node *dev;
183 	struct kfd_process_device *pdd;
184 
185 	dev = pqn->q->device;
186 
187 	pdd = kfd_get_process_device_data(dev, pqm->process);
188 	if (!pdd) {
189 		pr_err("Process device data doesn't exist\n");
190 		return;
191 	}
192 
193 	if (pqn->q->gws) {
194 		if (KFD_GC_VERSION(pqn->q->device) != IP_VERSION(9, 4, 3) &&
195 		    !dev->kfd->shared_resources.enable_mes)
196 			amdgpu_amdkfd_remove_gws_from_process(
197 				pqm->process->kgd_process_info, pqn->q->gws);
198 		pdd->qpd.num_gws = 0;
199 	}
200 
201 	if (dev->kfd->shared_resources.enable_mes) {
202 		amdgpu_amdkfd_free_gtt_mem(dev->adev, pqn->q->gang_ctx_bo);
203 		if (pqn->q->wptr_bo)
204 			amdgpu_amdkfd_free_gtt_mem(dev->adev, pqn->q->wptr_bo);
205 	}
206 }
207 
208 void pqm_uninit(struct process_queue_manager *pqm)
209 {
210 	struct process_queue_node *pqn, *next;
211 
212 	list_for_each_entry_safe(pqn, next, &pqm->queues, process_queue_list) {
213 		if (pqn->q)
214 			pqm_clean_queue_resource(pqm, pqn);
215 
216 		kfd_procfs_del_queue(pqn->q);
217 		uninit_queue(pqn->q);
218 		list_del(&pqn->process_queue_list);
219 		kfree(pqn);
220 	}
221 
222 	bitmap_free(pqm->queue_slot_bitmap);
223 	pqm->queue_slot_bitmap = NULL;
224 }
225 
226 static int init_user_queue(struct process_queue_manager *pqm,
227 				struct kfd_node *dev, struct queue **q,
228 				struct queue_properties *q_properties,
229 				struct file *f, struct amdgpu_bo *wptr_bo,
230 				unsigned int qid)
231 {
232 	int retval;
233 
234 	/* Doorbell initialized in user space*/
235 	q_properties->doorbell_ptr = NULL;
236 	q_properties->exception_status = KFD_EC_MASK(EC_QUEUE_NEW);
237 
238 	/* let DQM handle it*/
239 	q_properties->vmid = 0;
240 	q_properties->queue_id = qid;
241 
242 	retval = init_queue(q, q_properties);
243 	if (retval != 0)
244 		return retval;
245 
246 	(*q)->device = dev;
247 	(*q)->process = pqm->process;
248 
249 	if (dev->kfd->shared_resources.enable_mes) {
250 		retval = amdgpu_amdkfd_alloc_gtt_mem(dev->adev,
251 						AMDGPU_MES_GANG_CTX_SIZE,
252 						&(*q)->gang_ctx_bo,
253 						&(*q)->gang_ctx_gpu_addr,
254 						&(*q)->gang_ctx_cpu_ptr,
255 						false);
256 		if (retval) {
257 			pr_err("failed to allocate gang context bo\n");
258 			goto cleanup;
259 		}
260 		memset((*q)->gang_ctx_cpu_ptr, 0, AMDGPU_MES_GANG_CTX_SIZE);
261 		(*q)->wptr_bo = wptr_bo;
262 	}
263 
264 	pr_debug("PQM After init queue");
265 	return 0;
266 
267 cleanup:
268 	uninit_queue(*q);
269 	*q = NULL;
270 	return retval;
271 }
272 
273 int pqm_create_queue(struct process_queue_manager *pqm,
274 			    struct kfd_node *dev,
275 			    struct file *f,
276 			    struct queue_properties *properties,
277 			    unsigned int *qid,
278 			    struct amdgpu_bo *wptr_bo,
279 			    const struct kfd_criu_queue_priv_data *q_data,
280 			    const void *restore_mqd,
281 			    const void *restore_ctl_stack,
282 			    uint32_t *p_doorbell_offset_in_process)
283 {
284 	int retval;
285 	struct kfd_process_device *pdd;
286 	struct queue *q;
287 	struct process_queue_node *pqn;
288 	struct kernel_queue *kq;
289 	enum kfd_queue_type type = properties->type;
290 	unsigned int max_queues = 127; /* HWS limit */
291 
292 	/*
293 	 * On GFX 9.4.3, increase the number of queues that
294 	 * can be created to 255. No HWS limit on GFX 9.4.3.
295 	 */
296 	if (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 3))
297 		max_queues = 255;
298 
299 	q = NULL;
300 	kq = NULL;
301 
302 	pdd = kfd_get_process_device_data(dev, pqm->process);
303 	if (!pdd) {
304 		pr_err("Process device data doesn't exist\n");
305 		return -1;
306 	}
307 
308 	/*
309 	 * for debug process, verify that it is within the static queues limit
310 	 * currently limit is set to half of the total avail HQD slots
311 	 * If we are just about to create DIQ, the is_debug flag is not set yet
312 	 * Hence we also check the type as well
313 	 */
314 	if ((pdd->qpd.is_debug) || (type == KFD_QUEUE_TYPE_DIQ))
315 		max_queues = dev->kfd->device_info.max_no_of_hqd/2;
316 
317 	if (pdd->qpd.queue_count >= max_queues)
318 		return -ENOSPC;
319 
320 	if (q_data) {
321 		retval = assign_queue_slot_by_qid(pqm, q_data->q_id);
322 		*qid = q_data->q_id;
323 	} else
324 		retval = find_available_queue_slot(pqm, qid);
325 
326 	if (retval != 0)
327 		return retval;
328 
329 	if (list_empty(&pdd->qpd.queues_list) &&
330 	    list_empty(&pdd->qpd.priv_queue_list))
331 		dev->dqm->ops.register_process(dev->dqm, &pdd->qpd);
332 
333 	pqn = kzalloc(sizeof(*pqn), GFP_KERNEL);
334 	if (!pqn) {
335 		retval = -ENOMEM;
336 		goto err_allocate_pqn;
337 	}
338 
339 	switch (type) {
340 	case KFD_QUEUE_TYPE_SDMA:
341 	case KFD_QUEUE_TYPE_SDMA_XGMI:
342 		/* SDMA queues are always allocated statically no matter
343 		 * which scheduler mode is used. We also do not need to
344 		 * check whether a SDMA queue can be allocated here, because
345 		 * allocate_sdma_queue() in create_queue() has the
346 		 * corresponding check logic.
347 		 */
348 		retval = init_user_queue(pqm, dev, &q, properties, f, wptr_bo, *qid);
349 		if (retval != 0)
350 			goto err_create_queue;
351 		pqn->q = q;
352 		pqn->kq = NULL;
353 		retval = dev->dqm->ops.create_queue(dev->dqm, q, &pdd->qpd, q_data,
354 						    restore_mqd, restore_ctl_stack);
355 		print_queue(q);
356 		break;
357 
358 	case KFD_QUEUE_TYPE_COMPUTE:
359 		/* check if there is over subscription */
360 		if ((dev->dqm->sched_policy ==
361 		     KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION) &&
362 		((dev->dqm->processes_count >= dev->vm_info.vmid_num_kfd) ||
363 		(dev->dqm->active_queue_count >= get_cp_queues_num(dev->dqm)))) {
364 			pr_debug("Over-subscription is not allowed when amdkfd.sched_policy == 1\n");
365 			retval = -EPERM;
366 			goto err_create_queue;
367 		}
368 
369 		retval = init_user_queue(pqm, dev, &q, properties, f, wptr_bo, *qid);
370 		if (retval != 0)
371 			goto err_create_queue;
372 		pqn->q = q;
373 		pqn->kq = NULL;
374 		retval = dev->dqm->ops.create_queue(dev->dqm, q, &pdd->qpd, q_data,
375 						    restore_mqd, restore_ctl_stack);
376 		print_queue(q);
377 		break;
378 	case KFD_QUEUE_TYPE_DIQ:
379 		kq = kernel_queue_init(dev, KFD_QUEUE_TYPE_DIQ);
380 		if (!kq) {
381 			retval = -ENOMEM;
382 			goto err_create_queue;
383 		}
384 		kq->queue->properties.queue_id = *qid;
385 		pqn->kq = kq;
386 		pqn->q = NULL;
387 		retval = kfd_process_drain_interrupts(pdd);
388 		if (retval)
389 			break;
390 
391 		retval = dev->dqm->ops.create_kernel_queue(dev->dqm,
392 							kq, &pdd->qpd);
393 		break;
394 	default:
395 		WARN(1, "Invalid queue type %d", type);
396 		retval = -EINVAL;
397 	}
398 
399 	if (retval != 0) {
400 		pr_err("Pasid 0x%x DQM create queue type %d failed. ret %d\n",
401 			pqm->process->pasid, type, retval);
402 		goto err_create_queue;
403 	}
404 
405 	if (q && p_doorbell_offset_in_process) {
406 		/* Return the doorbell offset within the doorbell page
407 		 * to the caller so it can be passed up to user mode
408 		 * (in bytes).
409 		 * relative doorbell index = Absolute doorbell index -
410 		 * absolute index of first doorbell in the page.
411 		 */
412 		uint32_t first_db_index = amdgpu_doorbell_index_on_bar(pdd->dev->adev,
413 								       pdd->qpd.proc_doorbells,
414 								       0,
415 								       pdd->dev->kfd->device_info.doorbell_size);
416 
417 		*p_doorbell_offset_in_process = (q->properties.doorbell_off
418 						- first_db_index) * sizeof(uint32_t);
419 	}
420 
421 	pr_debug("PQM After DQM create queue\n");
422 
423 	list_add(&pqn->process_queue_list, &pqm->queues);
424 
425 	if (q) {
426 		pr_debug("PQM done creating queue\n");
427 		kfd_procfs_add_queue(q);
428 		print_queue_properties(&q->properties);
429 	}
430 
431 	return retval;
432 
433 err_create_queue:
434 	uninit_queue(q);
435 	if (kq)
436 		kernel_queue_uninit(kq, false);
437 	kfree(pqn);
438 err_allocate_pqn:
439 	/* check if queues list is empty unregister process from device */
440 	clear_bit(*qid, pqm->queue_slot_bitmap);
441 	if (list_empty(&pdd->qpd.queues_list) &&
442 	    list_empty(&pdd->qpd.priv_queue_list))
443 		dev->dqm->ops.unregister_process(dev->dqm, &pdd->qpd);
444 	return retval;
445 }
446 
447 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid)
448 {
449 	struct process_queue_node *pqn;
450 	struct kfd_process_device *pdd;
451 	struct device_queue_manager *dqm;
452 	struct kfd_node *dev;
453 	int retval;
454 
455 	dqm = NULL;
456 
457 	retval = 0;
458 
459 	pqn = get_queue_by_qid(pqm, qid);
460 	if (!pqn) {
461 		pr_err("Queue id does not match any known queue\n");
462 		return -EINVAL;
463 	}
464 
465 	dev = NULL;
466 	if (pqn->kq)
467 		dev = pqn->kq->dev;
468 	if (pqn->q)
469 		dev = pqn->q->device;
470 	if (WARN_ON(!dev))
471 		return -ENODEV;
472 
473 	pdd = kfd_get_process_device_data(dev, pqm->process);
474 	if (!pdd) {
475 		pr_err("Process device data doesn't exist\n");
476 		return -1;
477 	}
478 
479 	if (pqn->kq) {
480 		/* destroy kernel queue (DIQ) */
481 		dqm = pqn->kq->dev->dqm;
482 		dqm->ops.destroy_kernel_queue(dqm, pqn->kq, &pdd->qpd);
483 		kernel_queue_uninit(pqn->kq, false);
484 	}
485 
486 	if (pqn->q) {
487 		kfd_procfs_del_queue(pqn->q);
488 		dqm = pqn->q->device->dqm;
489 		retval = dqm->ops.destroy_queue(dqm, &pdd->qpd, pqn->q);
490 		if (retval) {
491 			pr_err("Pasid 0x%x destroy queue %d failed, ret %d\n",
492 				pqm->process->pasid,
493 				pqn->q->properties.queue_id, retval);
494 			if (retval != -ETIME)
495 				goto err_destroy_queue;
496 		}
497 
498 		pqm_clean_queue_resource(pqm, pqn);
499 		uninit_queue(pqn->q);
500 	}
501 
502 	list_del(&pqn->process_queue_list);
503 	kfree(pqn);
504 	clear_bit(qid, pqm->queue_slot_bitmap);
505 
506 	if (list_empty(&pdd->qpd.queues_list) &&
507 	    list_empty(&pdd->qpd.priv_queue_list))
508 		dqm->ops.unregister_process(dqm, &pdd->qpd);
509 
510 err_destroy_queue:
511 	return retval;
512 }
513 
514 int pqm_update_queue_properties(struct process_queue_manager *pqm,
515 				unsigned int qid, struct queue_properties *p)
516 {
517 	int retval;
518 	struct process_queue_node *pqn;
519 
520 	pqn = get_queue_by_qid(pqm, qid);
521 	if (!pqn) {
522 		pr_debug("No queue %d exists for update operation\n", qid);
523 		return -EFAULT;
524 	}
525 
526 	pqn->q->properties.queue_address = p->queue_address;
527 	pqn->q->properties.queue_size = p->queue_size;
528 	pqn->q->properties.queue_percent = p->queue_percent;
529 	pqn->q->properties.priority = p->priority;
530 	pqn->q->properties.pm4_target_xcc = p->pm4_target_xcc;
531 
532 	retval = pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm,
533 							pqn->q, NULL);
534 	if (retval != 0)
535 		return retval;
536 
537 	return 0;
538 }
539 
540 int pqm_update_mqd(struct process_queue_manager *pqm,
541 				unsigned int qid, struct mqd_update_info *minfo)
542 {
543 	int retval;
544 	struct process_queue_node *pqn;
545 
546 	pqn = get_queue_by_qid(pqm, qid);
547 	if (!pqn) {
548 		pr_debug("No queue %d exists for update operation\n", qid);
549 		return -EFAULT;
550 	}
551 
552 	/* CUs are masked for debugger requirements so deny user mask  */
553 	if (pqn->q->properties.is_dbg_wa && minfo && minfo->cu_mask.ptr)
554 		return -EBUSY;
555 
556 	/* ASICs that have WGPs must enforce pairwise enabled mask checks. */
557 	if (minfo && minfo->cu_mask.ptr &&
558 			KFD_GC_VERSION(pqn->q->device) >= IP_VERSION(10, 0, 0)) {
559 		int i;
560 
561 		for (i = 0; i < minfo->cu_mask.count; i += 2) {
562 			uint32_t cu_pair = (minfo->cu_mask.ptr[i / 32] >> (i % 32)) & 0x3;
563 
564 			if (cu_pair && cu_pair != 0x3) {
565 				pr_debug("CUs must be adjacent pairwise enabled.\n");
566 				return -EINVAL;
567 			}
568 		}
569 	}
570 
571 	retval = pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm,
572 							pqn->q, minfo);
573 	if (retval != 0)
574 		return retval;
575 
576 	if (minfo && minfo->cu_mask.ptr)
577 		pqn->q->properties.is_user_cu_masked = true;
578 
579 	return 0;
580 }
581 
582 struct kernel_queue *pqm_get_kernel_queue(
583 					struct process_queue_manager *pqm,
584 					unsigned int qid)
585 {
586 	struct process_queue_node *pqn;
587 
588 	pqn = get_queue_by_qid(pqm, qid);
589 	if (pqn && pqn->kq)
590 		return pqn->kq;
591 
592 	return NULL;
593 }
594 
595 struct queue *pqm_get_user_queue(struct process_queue_manager *pqm,
596 					unsigned int qid)
597 {
598 	struct process_queue_node *pqn;
599 
600 	pqn = get_queue_by_qid(pqm, qid);
601 	return pqn ? pqn->q : NULL;
602 }
603 
604 int pqm_get_wave_state(struct process_queue_manager *pqm,
605 		       unsigned int qid,
606 		       void __user *ctl_stack,
607 		       u32 *ctl_stack_used_size,
608 		       u32 *save_area_used_size)
609 {
610 	struct process_queue_node *pqn;
611 
612 	pqn = get_queue_by_qid(pqm, qid);
613 	if (!pqn) {
614 		pr_debug("amdkfd: No queue %d exists for operation\n",
615 			 qid);
616 		return -EFAULT;
617 	}
618 
619 	return pqn->q->device->dqm->ops.get_wave_state(pqn->q->device->dqm,
620 						       pqn->q,
621 						       ctl_stack,
622 						       ctl_stack_used_size,
623 						       save_area_used_size);
624 }
625 
626 int pqm_get_queue_snapshot(struct process_queue_manager *pqm,
627 			   uint64_t exception_clear_mask,
628 			   void __user *buf,
629 			   int *num_qss_entries,
630 			   uint32_t *entry_size)
631 {
632 	struct process_queue_node *pqn;
633 	struct kfd_queue_snapshot_entry src;
634 	uint32_t tmp_entry_size = *entry_size, tmp_qss_entries = *num_qss_entries;
635 	int r = 0;
636 
637 	*num_qss_entries = 0;
638 	if (!(*entry_size))
639 		return -EINVAL;
640 
641 	*entry_size = min_t(size_t, *entry_size, sizeof(struct kfd_queue_snapshot_entry));
642 	mutex_lock(&pqm->process->event_mutex);
643 
644 	memset(&src, 0, sizeof(src));
645 
646 	list_for_each_entry(pqn, &pqm->queues, process_queue_list) {
647 		if (!pqn->q)
648 			continue;
649 
650 		if (*num_qss_entries < tmp_qss_entries) {
651 			set_queue_snapshot_entry(pqn->q, exception_clear_mask, &src);
652 
653 			if (copy_to_user(buf, &src, *entry_size)) {
654 				r = -EFAULT;
655 				break;
656 			}
657 			buf += tmp_entry_size;
658 		}
659 		*num_qss_entries += 1;
660 	}
661 
662 	mutex_unlock(&pqm->process->event_mutex);
663 	return r;
664 }
665 
666 static int get_queue_data_sizes(struct kfd_process_device *pdd,
667 				struct queue *q,
668 				uint32_t *mqd_size,
669 				uint32_t *ctl_stack_size)
670 {
671 	int ret;
672 
673 	ret = pqm_get_queue_checkpoint_info(&pdd->process->pqm,
674 					    q->properties.queue_id,
675 					    mqd_size,
676 					    ctl_stack_size);
677 	if (ret)
678 		pr_err("Failed to get queue dump info (%d)\n", ret);
679 
680 	return ret;
681 }
682 
683 int kfd_process_get_queue_info(struct kfd_process *p,
684 			       uint32_t *num_queues,
685 			       uint64_t *priv_data_sizes)
686 {
687 	uint32_t extra_data_sizes = 0;
688 	struct queue *q;
689 	int i;
690 	int ret;
691 
692 	*num_queues = 0;
693 
694 	/* Run over all PDDs of the process */
695 	for (i = 0; i < p->n_pdds; i++) {
696 		struct kfd_process_device *pdd = p->pdds[i];
697 
698 		list_for_each_entry(q, &pdd->qpd.queues_list, list) {
699 			if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE ||
700 				q->properties.type == KFD_QUEUE_TYPE_SDMA ||
701 				q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
702 				uint32_t mqd_size, ctl_stack_size;
703 
704 				*num_queues = *num_queues + 1;
705 
706 				ret = get_queue_data_sizes(pdd, q, &mqd_size, &ctl_stack_size);
707 				if (ret)
708 					return ret;
709 
710 				extra_data_sizes += mqd_size + ctl_stack_size;
711 			} else {
712 				pr_err("Unsupported queue type (%d)\n", q->properties.type);
713 				return -EOPNOTSUPP;
714 			}
715 		}
716 	}
717 	*priv_data_sizes = extra_data_sizes +
718 				(*num_queues * sizeof(struct kfd_criu_queue_priv_data));
719 
720 	return 0;
721 }
722 
723 static int pqm_checkpoint_mqd(struct process_queue_manager *pqm,
724 			      unsigned int qid,
725 			      void *mqd,
726 			      void *ctl_stack)
727 {
728 	struct process_queue_node *pqn;
729 
730 	pqn = get_queue_by_qid(pqm, qid);
731 	if (!pqn) {
732 		pr_debug("amdkfd: No queue %d exists for operation\n", qid);
733 		return -EFAULT;
734 	}
735 
736 	if (!pqn->q->device->dqm->ops.checkpoint_mqd) {
737 		pr_err("amdkfd: queue dumping not supported on this device\n");
738 		return -EOPNOTSUPP;
739 	}
740 
741 	return pqn->q->device->dqm->ops.checkpoint_mqd(pqn->q->device->dqm,
742 						       pqn->q, mqd, ctl_stack);
743 }
744 
745 static int criu_checkpoint_queue(struct kfd_process_device *pdd,
746 			   struct queue *q,
747 			   struct kfd_criu_queue_priv_data *q_data)
748 {
749 	uint8_t *mqd, *ctl_stack;
750 	int ret;
751 
752 	mqd = (void *)(q_data + 1);
753 	ctl_stack = mqd + q_data->mqd_size;
754 
755 	q_data->gpu_id = pdd->user_gpu_id;
756 	q_data->type = q->properties.type;
757 	q_data->format = q->properties.format;
758 	q_data->q_id =  q->properties.queue_id;
759 	q_data->q_address = q->properties.queue_address;
760 	q_data->q_size = q->properties.queue_size;
761 	q_data->priority = q->properties.priority;
762 	q_data->q_percent = q->properties.queue_percent;
763 	q_data->read_ptr_addr = (uint64_t)q->properties.read_ptr;
764 	q_data->write_ptr_addr = (uint64_t)q->properties.write_ptr;
765 	q_data->doorbell_id = q->doorbell_id;
766 
767 	q_data->sdma_id = q->sdma_id;
768 
769 	q_data->eop_ring_buffer_address =
770 		q->properties.eop_ring_buffer_address;
771 
772 	q_data->eop_ring_buffer_size = q->properties.eop_ring_buffer_size;
773 
774 	q_data->ctx_save_restore_area_address =
775 		q->properties.ctx_save_restore_area_address;
776 
777 	q_data->ctx_save_restore_area_size =
778 		q->properties.ctx_save_restore_area_size;
779 
780 	q_data->gws = !!q->gws;
781 
782 	ret = pqm_checkpoint_mqd(&pdd->process->pqm, q->properties.queue_id, mqd, ctl_stack);
783 	if (ret) {
784 		pr_err("Failed checkpoint queue_mqd (%d)\n", ret);
785 		return ret;
786 	}
787 
788 	pr_debug("Dumping Queue: gpu_id:%x queue_id:%u\n", q_data->gpu_id, q_data->q_id);
789 	return ret;
790 }
791 
792 static int criu_checkpoint_queues_device(struct kfd_process_device *pdd,
793 				   uint8_t __user *user_priv,
794 				   unsigned int *q_index,
795 				   uint64_t *queues_priv_data_offset)
796 {
797 	unsigned int q_private_data_size = 0;
798 	uint8_t *q_private_data = NULL; /* Local buffer to store individual queue private data */
799 	struct queue *q;
800 	int ret = 0;
801 
802 	list_for_each_entry(q, &pdd->qpd.queues_list, list) {
803 		struct kfd_criu_queue_priv_data *q_data;
804 		uint64_t q_data_size;
805 		uint32_t mqd_size;
806 		uint32_t ctl_stack_size;
807 
808 		if (q->properties.type != KFD_QUEUE_TYPE_COMPUTE &&
809 			q->properties.type != KFD_QUEUE_TYPE_SDMA &&
810 			q->properties.type != KFD_QUEUE_TYPE_SDMA_XGMI) {
811 
812 			pr_err("Unsupported queue type (%d)\n", q->properties.type);
813 			ret = -EOPNOTSUPP;
814 			break;
815 		}
816 
817 		ret = get_queue_data_sizes(pdd, q, &mqd_size, &ctl_stack_size);
818 		if (ret)
819 			break;
820 
821 		q_data_size = sizeof(*q_data) + mqd_size + ctl_stack_size;
822 
823 		/* Increase local buffer space if needed */
824 		if (q_private_data_size < q_data_size) {
825 			kfree(q_private_data);
826 
827 			q_private_data = kzalloc(q_data_size, GFP_KERNEL);
828 			if (!q_private_data) {
829 				ret = -ENOMEM;
830 				break;
831 			}
832 			q_private_data_size = q_data_size;
833 		}
834 
835 		q_data = (struct kfd_criu_queue_priv_data *)q_private_data;
836 
837 		/* data stored in this order: priv_data, mqd, ctl_stack */
838 		q_data->mqd_size = mqd_size;
839 		q_data->ctl_stack_size = ctl_stack_size;
840 
841 		ret = criu_checkpoint_queue(pdd, q, q_data);
842 		if (ret)
843 			break;
844 
845 		q_data->object_type = KFD_CRIU_OBJECT_TYPE_QUEUE;
846 
847 		ret = copy_to_user(user_priv + *queues_priv_data_offset,
848 				q_data, q_data_size);
849 		if (ret) {
850 			ret = -EFAULT;
851 			break;
852 		}
853 		*queues_priv_data_offset += q_data_size;
854 		*q_index = *q_index + 1;
855 	}
856 
857 	kfree(q_private_data);
858 
859 	return ret;
860 }
861 
862 int kfd_criu_checkpoint_queues(struct kfd_process *p,
863 			 uint8_t __user *user_priv_data,
864 			 uint64_t *priv_data_offset)
865 {
866 	int ret = 0, pdd_index, q_index = 0;
867 
868 	for (pdd_index = 0; pdd_index < p->n_pdds; pdd_index++) {
869 		struct kfd_process_device *pdd = p->pdds[pdd_index];
870 
871 		/*
872 		 * criu_checkpoint_queues_device will copy data to user and update q_index and
873 		 * queues_priv_data_offset
874 		 */
875 		ret = criu_checkpoint_queues_device(pdd, user_priv_data, &q_index,
876 					      priv_data_offset);
877 
878 		if (ret)
879 			break;
880 	}
881 
882 	return ret;
883 }
884 
885 static void set_queue_properties_from_criu(struct queue_properties *qp,
886 					  struct kfd_criu_queue_priv_data *q_data)
887 {
888 	qp->is_interop = false;
889 	qp->queue_percent = q_data->q_percent;
890 	qp->priority = q_data->priority;
891 	qp->queue_address = q_data->q_address;
892 	qp->queue_size = q_data->q_size;
893 	qp->read_ptr = (uint32_t *) q_data->read_ptr_addr;
894 	qp->write_ptr = (uint32_t *) q_data->write_ptr_addr;
895 	qp->eop_ring_buffer_address = q_data->eop_ring_buffer_address;
896 	qp->eop_ring_buffer_size = q_data->eop_ring_buffer_size;
897 	qp->ctx_save_restore_area_address = q_data->ctx_save_restore_area_address;
898 	qp->ctx_save_restore_area_size = q_data->ctx_save_restore_area_size;
899 	qp->ctl_stack_size = q_data->ctl_stack_size;
900 	qp->type = q_data->type;
901 	qp->format = q_data->format;
902 }
903 
904 int kfd_criu_restore_queue(struct kfd_process *p,
905 			   uint8_t __user *user_priv_ptr,
906 			   uint64_t *priv_data_offset,
907 			   uint64_t max_priv_data_size)
908 {
909 	uint8_t *mqd, *ctl_stack, *q_extra_data = NULL;
910 	struct kfd_criu_queue_priv_data *q_data;
911 	struct kfd_process_device *pdd;
912 	uint64_t q_extra_data_size;
913 	struct queue_properties qp;
914 	unsigned int queue_id;
915 	int ret = 0;
916 
917 	if (*priv_data_offset + sizeof(*q_data) > max_priv_data_size)
918 		return -EINVAL;
919 
920 	q_data = kmalloc(sizeof(*q_data), GFP_KERNEL);
921 	if (!q_data)
922 		return -ENOMEM;
923 
924 	ret = copy_from_user(q_data, user_priv_ptr + *priv_data_offset, sizeof(*q_data));
925 	if (ret) {
926 		ret = -EFAULT;
927 		goto exit;
928 	}
929 
930 	*priv_data_offset += sizeof(*q_data);
931 	q_extra_data_size = (uint64_t)q_data->ctl_stack_size + q_data->mqd_size;
932 
933 	if (*priv_data_offset + q_extra_data_size > max_priv_data_size) {
934 		ret = -EINVAL;
935 		goto exit;
936 	}
937 
938 	q_extra_data = kmalloc(q_extra_data_size, GFP_KERNEL);
939 	if (!q_extra_data) {
940 		ret = -ENOMEM;
941 		goto exit;
942 	}
943 
944 	ret = copy_from_user(q_extra_data, user_priv_ptr + *priv_data_offset, q_extra_data_size);
945 	if (ret) {
946 		ret = -EFAULT;
947 		goto exit;
948 	}
949 
950 	*priv_data_offset += q_extra_data_size;
951 
952 	pdd = kfd_process_device_data_by_id(p, q_data->gpu_id);
953 	if (!pdd) {
954 		pr_err("Failed to get pdd\n");
955 		ret = -EINVAL;
956 		goto exit;
957 	}
958 
959 	/* data stored in this order: mqd, ctl_stack */
960 	mqd = q_extra_data;
961 	ctl_stack = mqd + q_data->mqd_size;
962 
963 	memset(&qp, 0, sizeof(qp));
964 	set_queue_properties_from_criu(&qp, q_data);
965 
966 	print_queue_properties(&qp);
967 
968 	ret = pqm_create_queue(&p->pqm, pdd->dev, NULL, &qp, &queue_id, NULL, q_data, mqd, ctl_stack,
969 				NULL);
970 	if (ret) {
971 		pr_err("Failed to create new queue err:%d\n", ret);
972 		goto exit;
973 	}
974 
975 	if (q_data->gws)
976 		ret = pqm_set_gws(&p->pqm, q_data->q_id, pdd->dev->gws);
977 
978 exit:
979 	if (ret)
980 		pr_err("Failed to restore queue (%d)\n", ret);
981 	else
982 		pr_debug("Queue id %d was restored successfully\n", queue_id);
983 
984 	kfree(q_data);
985 
986 	return ret;
987 }
988 
989 int pqm_get_queue_checkpoint_info(struct process_queue_manager *pqm,
990 				  unsigned int qid,
991 				  uint32_t *mqd_size,
992 				  uint32_t *ctl_stack_size)
993 {
994 	struct process_queue_node *pqn;
995 
996 	pqn = get_queue_by_qid(pqm, qid);
997 	if (!pqn) {
998 		pr_debug("amdkfd: No queue %d exists for operation\n", qid);
999 		return -EFAULT;
1000 	}
1001 
1002 	if (!pqn->q->device->dqm->ops.get_queue_checkpoint_info) {
1003 		pr_err("amdkfd: queue dumping not supported on this device\n");
1004 		return -EOPNOTSUPP;
1005 	}
1006 
1007 	pqn->q->device->dqm->ops.get_queue_checkpoint_info(pqn->q->device->dqm,
1008 						       pqn->q, mqd_size,
1009 						       ctl_stack_size);
1010 	return 0;
1011 }
1012 
1013 #if defined(CONFIG_DEBUG_FS)
1014 
1015 int pqm_debugfs_mqds(struct seq_file *m, void *data)
1016 {
1017 	struct process_queue_manager *pqm = data;
1018 	struct process_queue_node *pqn;
1019 	struct queue *q;
1020 	enum KFD_MQD_TYPE mqd_type;
1021 	struct mqd_manager *mqd_mgr;
1022 	int r = 0, xcc, num_xccs = 1;
1023 	void *mqd;
1024 	uint64_t size = 0;
1025 
1026 	list_for_each_entry(pqn, &pqm->queues, process_queue_list) {
1027 		if (pqn->q) {
1028 			q = pqn->q;
1029 			switch (q->properties.type) {
1030 			case KFD_QUEUE_TYPE_SDMA:
1031 			case KFD_QUEUE_TYPE_SDMA_XGMI:
1032 				seq_printf(m, "  SDMA queue on device %x\n",
1033 					   q->device->id);
1034 				mqd_type = KFD_MQD_TYPE_SDMA;
1035 				break;
1036 			case KFD_QUEUE_TYPE_COMPUTE:
1037 				seq_printf(m, "  Compute queue on device %x\n",
1038 					   q->device->id);
1039 				mqd_type = KFD_MQD_TYPE_CP;
1040 				num_xccs = NUM_XCC(q->device->xcc_mask);
1041 				break;
1042 			default:
1043 				seq_printf(m,
1044 				"  Bad user queue type %d on device %x\n",
1045 					   q->properties.type, q->device->id);
1046 				continue;
1047 			}
1048 			mqd_mgr = q->device->dqm->mqd_mgrs[mqd_type];
1049 			size = mqd_mgr->mqd_stride(mqd_mgr,
1050 							&q->properties);
1051 		} else if (pqn->kq) {
1052 			q = pqn->kq->queue;
1053 			mqd_mgr = pqn->kq->mqd_mgr;
1054 			switch (q->properties.type) {
1055 			case KFD_QUEUE_TYPE_DIQ:
1056 				seq_printf(m, "  DIQ on device %x\n",
1057 					   pqn->kq->dev->id);
1058 				break;
1059 			default:
1060 				seq_printf(m,
1061 				"  Bad kernel queue type %d on device %x\n",
1062 					   q->properties.type,
1063 					   pqn->kq->dev->id);
1064 				continue;
1065 			}
1066 		} else {
1067 			seq_printf(m,
1068 		"  Weird: Queue node with neither kernel nor user queue\n");
1069 			continue;
1070 		}
1071 
1072 		for (xcc = 0; xcc < num_xccs; xcc++) {
1073 			mqd = q->mqd + size * xcc;
1074 			r = mqd_mgr->debugfs_show_mqd(m, mqd);
1075 			if (r != 0)
1076 				break;
1077 		}
1078 	}
1079 
1080 	return r;
1081 }
1082 
1083 #endif
1084