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 %d\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 		uninit_queue(pqn->q);
154 		list_del(&pqn->process_queue_list);
155 		kfree(pqn);
156 	}
157 
158 	kfree(pqm->queue_slot_bitmap);
159 	pqm->queue_slot_bitmap = NULL;
160 }
161 
162 static int create_cp_queue(struct process_queue_manager *pqm,
163 				struct kfd_dev *dev, struct queue **q,
164 				struct queue_properties *q_properties,
165 				struct file *f, unsigned int qid)
166 {
167 	int retval;
168 
169 	/* Doorbell initialized in user space*/
170 	q_properties->doorbell_ptr = NULL;
171 
172 	/* let DQM handle it*/
173 	q_properties->vmid = 0;
174 	q_properties->queue_id = qid;
175 
176 	retval = init_queue(q, q_properties);
177 	if (retval != 0)
178 		return retval;
179 
180 	(*q)->device = dev;
181 	(*q)->process = pqm->process;
182 
183 	pr_debug("PQM After init queue");
184 
185 	return retval;
186 }
187 
188 int pqm_create_queue(struct process_queue_manager *pqm,
189 			    struct kfd_dev *dev,
190 			    struct file *f,
191 			    struct queue_properties *properties,
192 			    unsigned int *qid)
193 {
194 	int retval;
195 	struct kfd_process_device *pdd;
196 	struct queue *q;
197 	struct process_queue_node *pqn;
198 	struct kernel_queue *kq;
199 	enum kfd_queue_type type = properties->type;
200 	unsigned int max_queues = 127; /* HWS limit */
201 
202 	q = NULL;
203 	kq = NULL;
204 
205 	pdd = kfd_get_process_device_data(dev, pqm->process);
206 	if (!pdd) {
207 		pr_err("Process device data doesn't exist\n");
208 		return -1;
209 	}
210 
211 	/*
212 	 * for debug process, verify that it is within the static queues limit
213 	 * currently limit is set to half of the total avail HQD slots
214 	 * If we are just about to create DIQ, the is_debug flag is not set yet
215 	 * Hence we also check the type as well
216 	 */
217 	if ((pdd->qpd.is_debug) || (type == KFD_QUEUE_TYPE_DIQ))
218 		max_queues = dev->device_info->max_no_of_hqd/2;
219 
220 	if (pdd->qpd.queue_count >= max_queues)
221 		return -ENOSPC;
222 
223 	retval = find_available_queue_slot(pqm, qid);
224 	if (retval != 0)
225 		return retval;
226 
227 	if (list_empty(&pdd->qpd.queues_list) &&
228 	    list_empty(&pdd->qpd.priv_queue_list))
229 		dev->dqm->ops.register_process(dev->dqm, &pdd->qpd);
230 
231 	pqn = kzalloc(sizeof(*pqn), GFP_KERNEL);
232 	if (!pqn) {
233 		retval = -ENOMEM;
234 		goto err_allocate_pqn;
235 	}
236 
237 	switch (type) {
238 	case KFD_QUEUE_TYPE_SDMA:
239 	case KFD_QUEUE_TYPE_SDMA_XGMI:
240 		if ((type == KFD_QUEUE_TYPE_SDMA && dev->dqm->sdma_queue_count
241 			>= get_num_sdma_queues(dev->dqm)) ||
242 			(type == KFD_QUEUE_TYPE_SDMA_XGMI &&
243 			dev->dqm->xgmi_sdma_queue_count
244 			>= get_num_xgmi_sdma_queues(dev->dqm))) {
245 			pr_debug("Over-subscription is not allowed for SDMA.\n");
246 			retval = -EPERM;
247 			goto err_create_queue;
248 		}
249 
250 		retval = create_cp_queue(pqm, dev, &q, properties, f, *qid);
251 		if (retval != 0)
252 			goto err_create_queue;
253 		pqn->q = q;
254 		pqn->kq = NULL;
255 		retval = dev->dqm->ops.create_queue(dev->dqm, q, &pdd->qpd);
256 		pr_debug("DQM returned %d for create_queue\n", retval);
257 		print_queue(q);
258 		break;
259 
260 	case KFD_QUEUE_TYPE_COMPUTE:
261 		/* check if there is over subscription */
262 		if ((dev->dqm->sched_policy ==
263 		     KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION) &&
264 		((dev->dqm->processes_count >= dev->vm_info.vmid_num_kfd) ||
265 		(dev->dqm->queue_count >= get_queues_num(dev->dqm)))) {
266 			pr_debug("Over-subscription is not allowed when amdkfd.sched_policy == 1\n");
267 			retval = -EPERM;
268 			goto err_create_queue;
269 		}
270 
271 		retval = create_cp_queue(pqm, dev, &q, properties, f, *qid);
272 		if (retval != 0)
273 			goto err_create_queue;
274 		pqn->q = q;
275 		pqn->kq = NULL;
276 		retval = dev->dqm->ops.create_queue(dev->dqm, q, &pdd->qpd);
277 		pr_debug("DQM returned %d for create_queue\n", retval);
278 		print_queue(q);
279 		break;
280 	case KFD_QUEUE_TYPE_DIQ:
281 		kq = kernel_queue_init(dev, KFD_QUEUE_TYPE_DIQ);
282 		if (!kq) {
283 			retval = -ENOMEM;
284 			goto err_create_queue;
285 		}
286 		kq->queue->properties.queue_id = *qid;
287 		pqn->kq = kq;
288 		pqn->q = NULL;
289 		retval = dev->dqm->ops.create_kernel_queue(dev->dqm,
290 							kq, &pdd->qpd);
291 		break;
292 	default:
293 		WARN(1, "Invalid queue type %d", type);
294 		retval = -EINVAL;
295 	}
296 
297 	if (retval != 0) {
298 		pr_err("Pasid %d DQM create queue %d failed. ret %d\n",
299 			pqm->process->pasid, type, retval);
300 		goto err_create_queue;
301 	}
302 
303 	if (q)
304 		/* Return the doorbell offset within the doorbell page
305 		 * to the caller so it can be passed up to user mode
306 		 * (in bytes).
307 		 */
308 		properties->doorbell_off =
309 			(q->properties.doorbell_off * sizeof(uint32_t)) &
310 			(kfd_doorbell_process_slice(dev) - 1);
311 
312 	pr_debug("PQM After DQM create queue\n");
313 
314 	list_add(&pqn->process_queue_list, &pqm->queues);
315 
316 	if (q) {
317 		pr_debug("PQM done creating queue\n");
318 		print_queue_properties(&q->properties);
319 	}
320 
321 	return retval;
322 
323 err_create_queue:
324 	kfree(pqn);
325 err_allocate_pqn:
326 	/* check if queues list is empty unregister process from device */
327 	clear_bit(*qid, pqm->queue_slot_bitmap);
328 	if (list_empty(&pdd->qpd.queues_list) &&
329 	    list_empty(&pdd->qpd.priv_queue_list))
330 		dev->dqm->ops.unregister_process(dev->dqm, &pdd->qpd);
331 	return retval;
332 }
333 
334 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid)
335 {
336 	struct process_queue_node *pqn;
337 	struct kfd_process_device *pdd;
338 	struct device_queue_manager *dqm;
339 	struct kfd_dev *dev;
340 	int retval;
341 
342 	dqm = NULL;
343 
344 	retval = 0;
345 
346 	pqn = get_queue_by_qid(pqm, qid);
347 	if (!pqn) {
348 		pr_err("Queue id does not match any known queue\n");
349 		return -EINVAL;
350 	}
351 
352 	dev = NULL;
353 	if (pqn->kq)
354 		dev = pqn->kq->dev;
355 	if (pqn->q)
356 		dev = pqn->q->device;
357 	if (WARN_ON(!dev))
358 		return -ENODEV;
359 
360 	pdd = kfd_get_process_device_data(dev, pqm->process);
361 	if (!pdd) {
362 		pr_err("Process device data doesn't exist\n");
363 		return -1;
364 	}
365 
366 	if (pqn->kq) {
367 		/* destroy kernel queue (DIQ) */
368 		dqm = pqn->kq->dev->dqm;
369 		dqm->ops.destroy_kernel_queue(dqm, pqn->kq, &pdd->qpd);
370 		kernel_queue_uninit(pqn->kq);
371 	}
372 
373 	if (pqn->q) {
374 		dqm = pqn->q->device->dqm;
375 		retval = dqm->ops.destroy_queue(dqm, &pdd->qpd, pqn->q);
376 		if (retval) {
377 			pr_err("Pasid %d destroy queue %d failed, ret %d\n",
378 				pqm->process->pasid,
379 				pqn->q->properties.queue_id, retval);
380 			if (retval != -ETIME)
381 				goto err_destroy_queue;
382 		}
383 
384 		if (pqn->q->gws) {
385 			amdgpu_amdkfd_remove_gws_from_process(pqm->process->kgd_process_info,
386 				pqn->q->gws);
387 			pdd->qpd.num_gws = 0;
388 		}
389 
390 		kfree(pqn->q->properties.cu_mask);
391 		pqn->q->properties.cu_mask = NULL;
392 		uninit_queue(pqn->q);
393 	}
394 
395 	list_del(&pqn->process_queue_list);
396 	kfree(pqn);
397 	clear_bit(qid, pqm->queue_slot_bitmap);
398 
399 	if (list_empty(&pdd->qpd.queues_list) &&
400 	    list_empty(&pdd->qpd.priv_queue_list))
401 		dqm->ops.unregister_process(dqm, &pdd->qpd);
402 
403 err_destroy_queue:
404 	return retval;
405 }
406 
407 int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid,
408 			struct queue_properties *p)
409 {
410 	int retval;
411 	struct process_queue_node *pqn;
412 
413 	pqn = get_queue_by_qid(pqm, qid);
414 	if (!pqn) {
415 		pr_debug("No queue %d exists for update operation\n", qid);
416 		return -EFAULT;
417 	}
418 
419 	pqn->q->properties.queue_address = p->queue_address;
420 	pqn->q->properties.queue_size = p->queue_size;
421 	pqn->q->properties.queue_percent = p->queue_percent;
422 	pqn->q->properties.priority = p->priority;
423 
424 	retval = pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm,
425 							pqn->q);
426 	if (retval != 0)
427 		return retval;
428 
429 	return 0;
430 }
431 
432 int pqm_set_cu_mask(struct process_queue_manager *pqm, unsigned int qid,
433 			struct queue_properties *p)
434 {
435 	int retval;
436 	struct process_queue_node *pqn;
437 
438 	pqn = get_queue_by_qid(pqm, qid);
439 	if (!pqn) {
440 		pr_debug("No queue %d exists for update operation\n", qid);
441 		return -EFAULT;
442 	}
443 
444 	/* Free the old CU mask memory if it is already allocated, then
445 	 * allocate memory for the new CU mask.
446 	 */
447 	kfree(pqn->q->properties.cu_mask);
448 
449 	pqn->q->properties.cu_mask_count = p->cu_mask_count;
450 	pqn->q->properties.cu_mask = p->cu_mask;
451 
452 	retval = pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm,
453 							pqn->q);
454 	if (retval != 0)
455 		return retval;
456 
457 	return 0;
458 }
459 
460 struct kernel_queue *pqm_get_kernel_queue(
461 					struct process_queue_manager *pqm,
462 					unsigned int qid)
463 {
464 	struct process_queue_node *pqn;
465 
466 	pqn = get_queue_by_qid(pqm, qid);
467 	if (pqn && pqn->kq)
468 		return pqn->kq;
469 
470 	return NULL;
471 }
472 
473 int pqm_get_wave_state(struct process_queue_manager *pqm,
474 		       unsigned int qid,
475 		       void __user *ctl_stack,
476 		       u32 *ctl_stack_used_size,
477 		       u32 *save_area_used_size)
478 {
479 	struct process_queue_node *pqn;
480 
481 	pqn = get_queue_by_qid(pqm, qid);
482 	if (!pqn) {
483 		pr_debug("amdkfd: No queue %d exists for operation\n",
484 			 qid);
485 		return -EFAULT;
486 	}
487 
488 	return pqn->q->device->dqm->ops.get_wave_state(pqn->q->device->dqm,
489 						       pqn->q,
490 						       ctl_stack,
491 						       ctl_stack_used_size,
492 						       save_area_used_size);
493 }
494 
495 #if defined(CONFIG_DEBUG_FS)
496 
497 int pqm_debugfs_mqds(struct seq_file *m, void *data)
498 {
499 	struct process_queue_manager *pqm = data;
500 	struct process_queue_node *pqn;
501 	struct queue *q;
502 	enum KFD_MQD_TYPE mqd_type;
503 	struct mqd_manager *mqd_mgr;
504 	int r = 0;
505 
506 	list_for_each_entry(pqn, &pqm->queues, process_queue_list) {
507 		if (pqn->q) {
508 			q = pqn->q;
509 			switch (q->properties.type) {
510 			case KFD_QUEUE_TYPE_SDMA:
511 			case KFD_QUEUE_TYPE_SDMA_XGMI:
512 				seq_printf(m, "  SDMA queue on device %x\n",
513 					   q->device->id);
514 				mqd_type = KFD_MQD_TYPE_SDMA;
515 				break;
516 			case KFD_QUEUE_TYPE_COMPUTE:
517 				seq_printf(m, "  Compute queue on device %x\n",
518 					   q->device->id);
519 				mqd_type = KFD_MQD_TYPE_CP;
520 				break;
521 			default:
522 				seq_printf(m,
523 				"  Bad user queue type %d on device %x\n",
524 					   q->properties.type, q->device->id);
525 				continue;
526 			}
527 			mqd_mgr = q->device->dqm->mqd_mgrs[mqd_type];
528 		} else if (pqn->kq) {
529 			q = pqn->kq->queue;
530 			mqd_mgr = pqn->kq->mqd_mgr;
531 			switch (q->properties.type) {
532 			case KFD_QUEUE_TYPE_DIQ:
533 				seq_printf(m, "  DIQ on device %x\n",
534 					   pqn->kq->dev->id);
535 				break;
536 			default:
537 				seq_printf(m,
538 				"  Bad kernel queue type %d on device %x\n",
539 					   q->properties.type,
540 					   pqn->kq->dev->id);
541 				continue;
542 			}
543 		} else {
544 			seq_printf(m,
545 		"  Weird: Queue node with neither kernel nor user queue\n");
546 			continue;
547 		}
548 
549 		r = mqd_mgr->debugfs_show_mqd(m, q->mqd);
550 		if (r != 0)
551 			break;
552 	}
553 
554 	return r;
555 }
556 
557 #endif
558