1 /*
2  * Copyright 2014 Advanced Micro Devices, Inc.
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice shall be included in
12  * all copies or substantial portions of the Software.
13  *
14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20  * OTHER DEALINGS IN THE SOFTWARE.
21  *
22  */
23 
24 #include <linux/slab.h>
25 #include <linux/list.h>
26 #include "kfd_device_queue_manager.h"
27 #include "kfd_priv.h"
28 #include "kfd_kernel_queue.h"
29 #include "amdgpu_amdkfd.h"
30 
31 static inline struct process_queue_node *get_queue_by_qid(
32 			struct process_queue_manager *pqm, unsigned int qid)
33 {
34 	struct process_queue_node *pqn;
35 
36 	list_for_each_entry(pqn, &pqm->queues, process_queue_list) {
37 		if ((pqn->q && pqn->q->properties.queue_id == qid) ||
38 		    (pqn->kq && pqn->kq->queue->properties.queue_id == qid))
39 			return pqn;
40 	}
41 
42 	return NULL;
43 }
44 
45 static int find_available_queue_slot(struct process_queue_manager *pqm,
46 					unsigned int *qid)
47 {
48 	unsigned long found;
49 
50 	found = find_first_zero_bit(pqm->queue_slot_bitmap,
51 			KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);
52 
53 	pr_debug("The new slot id %lu\n", found);
54 
55 	if (found >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) {
56 		pr_info("Cannot open more queues for process with pasid 0x%x\n",
57 				pqm->process->pasid);
58 		return -ENOMEM;
59 	}
60 
61 	set_bit(found, pqm->queue_slot_bitmap);
62 	*qid = found;
63 
64 	return 0;
65 }
66 
67 void kfd_process_dequeue_from_device(struct kfd_process_device *pdd)
68 {
69 	struct kfd_dev *dev = pdd->dev;
70 
71 	if (pdd->already_dequeued)
72 		return;
73 
74 	dev->dqm->ops.process_termination(dev->dqm, &pdd->qpd);
75 	pdd->already_dequeued = true;
76 }
77 
78 int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid,
79 			void *gws)
80 {
81 	struct kfd_dev *dev = NULL;
82 	struct process_queue_node *pqn;
83 	struct kfd_process_device *pdd;
84 	struct kgd_mem *mem = NULL;
85 	int ret;
86 
87 	pqn = get_queue_by_qid(pqm, qid);
88 	if (!pqn) {
89 		pr_err("Queue id does not match any known queue\n");
90 		return -EINVAL;
91 	}
92 
93 	if (pqn->q)
94 		dev = pqn->q->device;
95 	if (WARN_ON(!dev))
96 		return -ENODEV;
97 
98 	pdd = kfd_get_process_device_data(dev, pqm->process);
99 	if (!pdd) {
100 		pr_err("Process device data doesn't exist\n");
101 		return -EINVAL;
102 	}
103 
104 	/* Only allow one queue per process can have GWS assigned */
105 	if (gws && pdd->qpd.num_gws)
106 		return -EBUSY;
107 
108 	if (!gws && pdd->qpd.num_gws == 0)
109 		return -EINVAL;
110 
111 	if (gws)
112 		ret = amdgpu_amdkfd_add_gws_to_process(pdd->process->kgd_process_info,
113 			gws, &mem);
114 	else
115 		ret = amdgpu_amdkfd_remove_gws_from_process(pdd->process->kgd_process_info,
116 			pqn->q->gws);
117 	if (unlikely(ret))
118 		return ret;
119 
120 	pqn->q->gws = mem;
121 	pdd->qpd.num_gws = gws ? amdgpu_amdkfd_get_num_gws(dev->kgd) : 0;
122 
123 	return pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm,
124 							pqn->q);
125 }
126 
127 void kfd_process_dequeue_from_all_devices(struct kfd_process *p)
128 {
129 	struct kfd_process_device *pdd;
130 
131 	list_for_each_entry(pdd, &p->per_device_data, per_device_list)
132 		kfd_process_dequeue_from_device(pdd);
133 }
134 
135 int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p)
136 {
137 	INIT_LIST_HEAD(&pqm->queues);
138 	pqm->queue_slot_bitmap =
139 			kzalloc(DIV_ROUND_UP(KFD_MAX_NUM_OF_QUEUES_PER_PROCESS,
140 					BITS_PER_BYTE), GFP_KERNEL);
141 	if (!pqm->queue_slot_bitmap)
142 		return -ENOMEM;
143 	pqm->process = p;
144 
145 	return 0;
146 }
147 
148 void pqm_uninit(struct process_queue_manager *pqm)
149 {
150 	struct process_queue_node *pqn, *next;
151 
152 	list_for_each_entry_safe(pqn, next, &pqm->queues, process_queue_list) {
153 		if (pqn->q && pqn->q->gws)
154 			amdgpu_amdkfd_remove_gws_from_process(pqm->process->kgd_process_info,
155 				pqn->q->gws);
156 		uninit_queue(pqn->q);
157 		list_del(&pqn->process_queue_list);
158 		kfree(pqn);
159 	}
160 
161 	kfree(pqm->queue_slot_bitmap);
162 	pqm->queue_slot_bitmap = NULL;
163 }
164 
165 static int create_cp_queue(struct process_queue_manager *pqm,
166 				struct kfd_dev *dev, struct queue **q,
167 				struct queue_properties *q_properties,
168 				struct file *f, unsigned int qid)
169 {
170 	int retval;
171 
172 	/* Doorbell initialized in user space*/
173 	q_properties->doorbell_ptr = NULL;
174 
175 	/* let DQM handle it*/
176 	q_properties->vmid = 0;
177 	q_properties->queue_id = qid;
178 
179 	retval = init_queue(q, q_properties);
180 	if (retval != 0)
181 		return retval;
182 
183 	(*q)->device = dev;
184 	(*q)->process = pqm->process;
185 
186 	pr_debug("PQM After init queue");
187 
188 	return retval;
189 }
190 
191 int pqm_create_queue(struct process_queue_manager *pqm,
192 			    struct kfd_dev *dev,
193 			    struct file *f,
194 			    struct queue_properties *properties,
195 			    unsigned int *qid)
196 {
197 	int retval;
198 	struct kfd_process_device *pdd;
199 	struct queue *q;
200 	struct process_queue_node *pqn;
201 	struct kernel_queue *kq;
202 	enum kfd_queue_type type = properties->type;
203 	unsigned int max_queues = 127; /* HWS limit */
204 
205 	q = NULL;
206 	kq = NULL;
207 
208 	pdd = kfd_get_process_device_data(dev, pqm->process);
209 	if (!pdd) {
210 		pr_err("Process device data doesn't exist\n");
211 		return -1;
212 	}
213 
214 	/*
215 	 * for debug process, verify that it is within the static queues limit
216 	 * currently limit is set to half of the total avail HQD slots
217 	 * If we are just about to create DIQ, the is_debug flag is not set yet
218 	 * Hence we also check the type as well
219 	 */
220 	if ((pdd->qpd.is_debug) || (type == KFD_QUEUE_TYPE_DIQ))
221 		max_queues = dev->device_info->max_no_of_hqd/2;
222 
223 	if (pdd->qpd.queue_count >= max_queues)
224 		return -ENOSPC;
225 
226 	retval = find_available_queue_slot(pqm, qid);
227 	if (retval != 0)
228 		return retval;
229 
230 	if (list_empty(&pdd->qpd.queues_list) &&
231 	    list_empty(&pdd->qpd.priv_queue_list))
232 		dev->dqm->ops.register_process(dev->dqm, &pdd->qpd);
233 
234 	pqn = kzalloc(sizeof(*pqn), GFP_KERNEL);
235 	if (!pqn) {
236 		retval = -ENOMEM;
237 		goto err_allocate_pqn;
238 	}
239 
240 	switch (type) {
241 	case KFD_QUEUE_TYPE_SDMA:
242 	case KFD_QUEUE_TYPE_SDMA_XGMI:
243 		if ((type == KFD_QUEUE_TYPE_SDMA && dev->dqm->sdma_queue_count
244 			>= get_num_sdma_queues(dev->dqm)) ||
245 			(type == KFD_QUEUE_TYPE_SDMA_XGMI &&
246 			dev->dqm->xgmi_sdma_queue_count
247 			>= get_num_xgmi_sdma_queues(dev->dqm))) {
248 			pr_debug("Over-subscription is not allowed for SDMA.\n");
249 			retval = -EPERM;
250 			goto err_create_queue;
251 		}
252 
253 		retval = create_cp_queue(pqm, dev, &q, properties, f, *qid);
254 		if (retval != 0)
255 			goto err_create_queue;
256 		pqn->q = q;
257 		pqn->kq = NULL;
258 		retval = dev->dqm->ops.create_queue(dev->dqm, q, &pdd->qpd);
259 		pr_debug("DQM returned %d for create_queue\n", retval);
260 		print_queue(q);
261 		break;
262 
263 	case KFD_QUEUE_TYPE_COMPUTE:
264 		/* check if there is over subscription */
265 		if ((dev->dqm->sched_policy ==
266 		     KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION) &&
267 		((dev->dqm->processes_count >= dev->vm_info.vmid_num_kfd) ||
268 		(dev->dqm->queue_count >= get_queues_num(dev->dqm)))) {
269 			pr_debug("Over-subscription is not allowed when amdkfd.sched_policy == 1\n");
270 			retval = -EPERM;
271 			goto err_create_queue;
272 		}
273 
274 		retval = create_cp_queue(pqm, dev, &q, properties, f, *qid);
275 		if (retval != 0)
276 			goto err_create_queue;
277 		pqn->q = q;
278 		pqn->kq = NULL;
279 		retval = dev->dqm->ops.create_queue(dev->dqm, q, &pdd->qpd);
280 		pr_debug("DQM returned %d for create_queue\n", retval);
281 		print_queue(q);
282 		break;
283 	case KFD_QUEUE_TYPE_DIQ:
284 		kq = kernel_queue_init(dev, KFD_QUEUE_TYPE_DIQ);
285 		if (!kq) {
286 			retval = -ENOMEM;
287 			goto err_create_queue;
288 		}
289 		kq->queue->properties.queue_id = *qid;
290 		pqn->kq = kq;
291 		pqn->q = NULL;
292 		retval = dev->dqm->ops.create_kernel_queue(dev->dqm,
293 							kq, &pdd->qpd);
294 		break;
295 	default:
296 		WARN(1, "Invalid queue type %d", type);
297 		retval = -EINVAL;
298 	}
299 
300 	if (retval != 0) {
301 		pr_err("Pasid 0x%x DQM create queue %d failed. ret %d\n",
302 			pqm->process->pasid, type, retval);
303 		goto err_create_queue;
304 	}
305 
306 	if (q)
307 		/* Return the doorbell offset within the doorbell page
308 		 * to the caller so it can be passed up to user mode
309 		 * (in bytes).
310 		 */
311 		properties->doorbell_off =
312 			(q->properties.doorbell_off * sizeof(uint32_t)) &
313 			(kfd_doorbell_process_slice(dev) - 1);
314 
315 	pr_debug("PQM After DQM create queue\n");
316 
317 	list_add(&pqn->process_queue_list, &pqm->queues);
318 
319 	if (q) {
320 		pr_debug("PQM done creating queue\n");
321 		print_queue_properties(&q->properties);
322 	}
323 
324 	return retval;
325 
326 err_create_queue:
327 	kfree(pqn);
328 err_allocate_pqn:
329 	/* check if queues list is empty unregister process from device */
330 	clear_bit(*qid, pqm->queue_slot_bitmap);
331 	if (list_empty(&pdd->qpd.queues_list) &&
332 	    list_empty(&pdd->qpd.priv_queue_list))
333 		dev->dqm->ops.unregister_process(dev->dqm, &pdd->qpd);
334 	return retval;
335 }
336 
337 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid)
338 {
339 	struct process_queue_node *pqn;
340 	struct kfd_process_device *pdd;
341 	struct device_queue_manager *dqm;
342 	struct kfd_dev *dev;
343 	int retval;
344 
345 	dqm = NULL;
346 
347 	retval = 0;
348 
349 	pqn = get_queue_by_qid(pqm, qid);
350 	if (!pqn) {
351 		pr_err("Queue id does not match any known queue\n");
352 		return -EINVAL;
353 	}
354 
355 	dev = NULL;
356 	if (pqn->kq)
357 		dev = pqn->kq->dev;
358 	if (pqn->q)
359 		dev = pqn->q->device;
360 	if (WARN_ON(!dev))
361 		return -ENODEV;
362 
363 	pdd = kfd_get_process_device_data(dev, pqm->process);
364 	if (!pdd) {
365 		pr_err("Process device data doesn't exist\n");
366 		return -1;
367 	}
368 
369 	if (pqn->kq) {
370 		/* destroy kernel queue (DIQ) */
371 		dqm = pqn->kq->dev->dqm;
372 		dqm->ops.destroy_kernel_queue(dqm, pqn->kq, &pdd->qpd);
373 		kernel_queue_uninit(pqn->kq);
374 	}
375 
376 	if (pqn->q) {
377 		dqm = pqn->q->device->dqm;
378 		retval = dqm->ops.destroy_queue(dqm, &pdd->qpd, pqn->q);
379 		if (retval) {
380 			pr_err("Pasid 0x%x destroy queue %d failed, ret %d\n",
381 				pqm->process->pasid,
382 				pqn->q->properties.queue_id, retval);
383 			if (retval != -ETIME)
384 				goto err_destroy_queue;
385 		}
386 
387 		if (pqn->q->gws) {
388 			amdgpu_amdkfd_remove_gws_from_process(pqm->process->kgd_process_info,
389 				pqn->q->gws);
390 			pdd->qpd.num_gws = 0;
391 		}
392 
393 		kfree(pqn->q->properties.cu_mask);
394 		pqn->q->properties.cu_mask = NULL;
395 		uninit_queue(pqn->q);
396 	}
397 
398 	list_del(&pqn->process_queue_list);
399 	kfree(pqn);
400 	clear_bit(qid, pqm->queue_slot_bitmap);
401 
402 	if (list_empty(&pdd->qpd.queues_list) &&
403 	    list_empty(&pdd->qpd.priv_queue_list))
404 		dqm->ops.unregister_process(dqm, &pdd->qpd);
405 
406 err_destroy_queue:
407 	return retval;
408 }
409 
410 int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid,
411 			struct queue_properties *p)
412 {
413 	int retval;
414 	struct process_queue_node *pqn;
415 
416 	pqn = get_queue_by_qid(pqm, qid);
417 	if (!pqn) {
418 		pr_debug("No queue %d exists for update operation\n", qid);
419 		return -EFAULT;
420 	}
421 
422 	pqn->q->properties.queue_address = p->queue_address;
423 	pqn->q->properties.queue_size = p->queue_size;
424 	pqn->q->properties.queue_percent = p->queue_percent;
425 	pqn->q->properties.priority = p->priority;
426 
427 	retval = pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm,
428 							pqn->q);
429 	if (retval != 0)
430 		return retval;
431 
432 	return 0;
433 }
434 
435 int pqm_set_cu_mask(struct process_queue_manager *pqm, unsigned int qid,
436 			struct queue_properties *p)
437 {
438 	int retval;
439 	struct process_queue_node *pqn;
440 
441 	pqn = get_queue_by_qid(pqm, qid);
442 	if (!pqn) {
443 		pr_debug("No queue %d exists for update operation\n", qid);
444 		return -EFAULT;
445 	}
446 
447 	/* Free the old CU mask memory if it is already allocated, then
448 	 * allocate memory for the new CU mask.
449 	 */
450 	kfree(pqn->q->properties.cu_mask);
451 
452 	pqn->q->properties.cu_mask_count = p->cu_mask_count;
453 	pqn->q->properties.cu_mask = p->cu_mask;
454 
455 	retval = pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm,
456 							pqn->q);
457 	if (retval != 0)
458 		return retval;
459 
460 	return 0;
461 }
462 
463 struct kernel_queue *pqm_get_kernel_queue(
464 					struct process_queue_manager *pqm,
465 					unsigned int qid)
466 {
467 	struct process_queue_node *pqn;
468 
469 	pqn = get_queue_by_qid(pqm, qid);
470 	if (pqn && pqn->kq)
471 		return pqn->kq;
472 
473 	return NULL;
474 }
475 
476 int pqm_get_wave_state(struct process_queue_manager *pqm,
477 		       unsigned int qid,
478 		       void __user *ctl_stack,
479 		       u32 *ctl_stack_used_size,
480 		       u32 *save_area_used_size)
481 {
482 	struct process_queue_node *pqn;
483 
484 	pqn = get_queue_by_qid(pqm, qid);
485 	if (!pqn) {
486 		pr_debug("amdkfd: No queue %d exists for operation\n",
487 			 qid);
488 		return -EFAULT;
489 	}
490 
491 	return pqn->q->device->dqm->ops.get_wave_state(pqn->q->device->dqm,
492 						       pqn->q,
493 						       ctl_stack,
494 						       ctl_stack_used_size,
495 						       save_area_used_size);
496 }
497 
498 #if defined(CONFIG_DEBUG_FS)
499 
500 int pqm_debugfs_mqds(struct seq_file *m, void *data)
501 {
502 	struct process_queue_manager *pqm = data;
503 	struct process_queue_node *pqn;
504 	struct queue *q;
505 	enum KFD_MQD_TYPE mqd_type;
506 	struct mqd_manager *mqd_mgr;
507 	int r = 0;
508 
509 	list_for_each_entry(pqn, &pqm->queues, process_queue_list) {
510 		if (pqn->q) {
511 			q = pqn->q;
512 			switch (q->properties.type) {
513 			case KFD_QUEUE_TYPE_SDMA:
514 			case KFD_QUEUE_TYPE_SDMA_XGMI:
515 				seq_printf(m, "  SDMA queue on device %x\n",
516 					   q->device->id);
517 				mqd_type = KFD_MQD_TYPE_SDMA;
518 				break;
519 			case KFD_QUEUE_TYPE_COMPUTE:
520 				seq_printf(m, "  Compute queue on device %x\n",
521 					   q->device->id);
522 				mqd_type = KFD_MQD_TYPE_CP;
523 				break;
524 			default:
525 				seq_printf(m,
526 				"  Bad user queue type %d on device %x\n",
527 					   q->properties.type, q->device->id);
528 				continue;
529 			}
530 			mqd_mgr = q->device->dqm->mqd_mgrs[mqd_type];
531 		} else if (pqn->kq) {
532 			q = pqn->kq->queue;
533 			mqd_mgr = pqn->kq->mqd_mgr;
534 			switch (q->properties.type) {
535 			case KFD_QUEUE_TYPE_DIQ:
536 				seq_printf(m, "  DIQ on device %x\n",
537 					   pqn->kq->dev->id);
538 				break;
539 			default:
540 				seq_printf(m,
541 				"  Bad kernel queue type %d on device %x\n",
542 					   q->properties.type,
543 					   pqn->kq->dev->id);
544 				continue;
545 			}
546 		} else {
547 			seq_printf(m,
548 		"  Weird: Queue node with neither kernel nor user queue\n");
549 			continue;
550 		}
551 
552 		r = mqd_mgr->debugfs_show_mqd(m, q->mqd);
553 		if (r != 0)
554 			break;
555 	}
556 
557 	return r;
558 }
559 
560 #endif
561