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