xref: /openbmc/linux/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c (revision b240b419db5d624ce7a5a397d6f62a1a686009ec)
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 
30 static inline struct process_queue_node *get_queue_by_qid(
31 			struct process_queue_manager *pqm, unsigned int qid)
32 {
33 	struct process_queue_node *pqn;
34 
35 	list_for_each_entry(pqn, &pqm->queues, process_queue_list) {
36 		if ((pqn->q && pqn->q->properties.queue_id == qid) ||
37 		    (pqn->kq && pqn->kq->queue->properties.queue_id == qid))
38 			return pqn;
39 	}
40 
41 	return NULL;
42 }
43 
44 static int find_available_queue_slot(struct process_queue_manager *pqm,
45 					unsigned int *qid)
46 {
47 	unsigned long found;
48 
49 	found = find_first_zero_bit(pqm->queue_slot_bitmap,
50 			KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);
51 
52 	pr_debug("The new slot id %lu\n", found);
53 
54 	if (found >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) {
55 		pr_info("Cannot open more queues for process with pasid %d\n",
56 				pqm->process->pasid);
57 		return -ENOMEM;
58 	}
59 
60 	set_bit(found, pqm->queue_slot_bitmap);
61 	*qid = found;
62 
63 	return 0;
64 }
65 
66 void kfd_process_dequeue_from_device(struct kfd_process_device *pdd)
67 {
68 	struct kfd_dev *dev = pdd->dev;
69 
70 	if (pdd->already_dequeued)
71 		return;
72 
73 	dev->dqm->ops.process_termination(dev->dqm, &pdd->qpd);
74 	pdd->already_dequeued = true;
75 }
76 
77 void kfd_process_dequeue_from_all_devices(struct kfd_process *p)
78 {
79 	struct kfd_process_device *pdd;
80 
81 	list_for_each_entry(pdd, &p->per_device_data, per_device_list)
82 		kfd_process_dequeue_from_device(pdd);
83 }
84 
85 int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p)
86 {
87 	INIT_LIST_HEAD(&pqm->queues);
88 	pqm->queue_slot_bitmap =
89 			kzalloc(DIV_ROUND_UP(KFD_MAX_NUM_OF_QUEUES_PER_PROCESS,
90 					BITS_PER_BYTE), GFP_KERNEL);
91 	if (!pqm->queue_slot_bitmap)
92 		return -ENOMEM;
93 	pqm->process = p;
94 
95 	return 0;
96 }
97 
98 void pqm_uninit(struct process_queue_manager *pqm)
99 {
100 	struct process_queue_node *pqn, *next;
101 
102 	list_for_each_entry_safe(pqn, next, &pqm->queues, process_queue_list) {
103 		uninit_queue(pqn->q);
104 		list_del(&pqn->process_queue_list);
105 		kfree(pqn);
106 	}
107 
108 	kfree(pqm->queue_slot_bitmap);
109 	pqm->queue_slot_bitmap = NULL;
110 }
111 
112 static int create_cp_queue(struct process_queue_manager *pqm,
113 				struct kfd_dev *dev, struct queue **q,
114 				struct queue_properties *q_properties,
115 				struct file *f, unsigned int qid)
116 {
117 	int retval;
118 
119 	/* Doorbell initialized in user space*/
120 	q_properties->doorbell_ptr = NULL;
121 
122 	q_properties->doorbell_off =
123 			kfd_queue_id_to_doorbell(dev, pqm->process, qid);
124 
125 	/* let DQM handle it*/
126 	q_properties->vmid = 0;
127 	q_properties->queue_id = qid;
128 
129 	retval = init_queue(q, q_properties);
130 	if (retval != 0)
131 		return retval;
132 
133 	(*q)->device = dev;
134 	(*q)->process = pqm->process;
135 
136 	pr_debug("PQM After init queue");
137 
138 	return retval;
139 }
140 
141 int pqm_create_queue(struct process_queue_manager *pqm,
142 			    struct kfd_dev *dev,
143 			    struct file *f,
144 			    struct queue_properties *properties,
145 			    unsigned int *qid)
146 {
147 	int retval;
148 	struct kfd_process_device *pdd;
149 	struct queue *q;
150 	struct process_queue_node *pqn;
151 	struct kernel_queue *kq;
152 	enum kfd_queue_type type = properties->type;
153 	unsigned int max_queues = 127; /* HWS limit */
154 
155 	q = NULL;
156 	kq = NULL;
157 
158 	pdd = kfd_get_process_device_data(dev, pqm->process);
159 	if (!pdd) {
160 		pr_err("Process device data doesn't exist\n");
161 		return -1;
162 	}
163 
164 	/*
165 	 * for debug process, verify that it is within the static queues limit
166 	 * currently limit is set to half of the total avail HQD slots
167 	 * If we are just about to create DIQ, the is_debug flag is not set yet
168 	 * Hence we also check the type as well
169 	 */
170 	if ((pdd->qpd.is_debug) || (type == KFD_QUEUE_TYPE_DIQ))
171 		max_queues = dev->device_info->max_no_of_hqd/2;
172 
173 	if (pdd->qpd.queue_count >= max_queues)
174 		return -ENOSPC;
175 
176 	retval = find_available_queue_slot(pqm, qid);
177 	if (retval != 0)
178 		return retval;
179 
180 	if (list_empty(&pdd->qpd.queues_list) &&
181 	    list_empty(&pdd->qpd.priv_queue_list))
182 		dev->dqm->ops.register_process(dev->dqm, &pdd->qpd);
183 
184 	pqn = kzalloc(sizeof(*pqn), GFP_KERNEL);
185 	if (!pqn) {
186 		retval = -ENOMEM;
187 		goto err_allocate_pqn;
188 	}
189 
190 	switch (type) {
191 	case KFD_QUEUE_TYPE_SDMA:
192 		if (dev->dqm->queue_count >=
193 			CIK_SDMA_QUEUES_PER_ENGINE * CIK_SDMA_ENGINE_NUM) {
194 			pr_err("Over-subscription is not allowed for SDMA.\n");
195 			retval = -EPERM;
196 			goto err_create_queue;
197 		}
198 
199 		retval = create_cp_queue(pqm, dev, &q, properties, f, *qid);
200 		if (retval != 0)
201 			goto err_create_queue;
202 		pqn->q = q;
203 		pqn->kq = NULL;
204 		retval = dev->dqm->ops.create_queue(dev->dqm, q, &pdd->qpd);
205 		pr_debug("DQM returned %d for create_queue\n", retval);
206 		print_queue(q);
207 		break;
208 
209 	case KFD_QUEUE_TYPE_COMPUTE:
210 		/* check if there is over subscription */
211 		if ((dev->dqm->sched_policy ==
212 		     KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION) &&
213 		((dev->dqm->processes_count >= dev->vm_info.vmid_num_kfd) ||
214 		(dev->dqm->queue_count >= get_queues_num(dev->dqm)))) {
215 			pr_err("Over-subscription is not allowed in radeon_kfd.sched_policy == 1\n");
216 			retval = -EPERM;
217 			goto err_create_queue;
218 		}
219 
220 		retval = create_cp_queue(pqm, dev, &q, properties, f, *qid);
221 		if (retval != 0)
222 			goto err_create_queue;
223 		pqn->q = q;
224 		pqn->kq = NULL;
225 		retval = dev->dqm->ops.create_queue(dev->dqm, q, &pdd->qpd);
226 		pr_debug("DQM returned %d for create_queue\n", retval);
227 		print_queue(q);
228 		break;
229 	case KFD_QUEUE_TYPE_DIQ:
230 		kq = kernel_queue_init(dev, KFD_QUEUE_TYPE_DIQ);
231 		if (!kq) {
232 			retval = -ENOMEM;
233 			goto err_create_queue;
234 		}
235 		kq->queue->properties.queue_id = *qid;
236 		pqn->kq = kq;
237 		pqn->q = NULL;
238 		retval = dev->dqm->ops.create_kernel_queue(dev->dqm,
239 							kq, &pdd->qpd);
240 		break;
241 	default:
242 		WARN(1, "Invalid queue type %d", type);
243 		retval = -EINVAL;
244 	}
245 
246 	if (retval != 0) {
247 		pr_err("DQM create queue failed\n");
248 		goto err_create_queue;
249 	}
250 
251 	pr_debug("PQM After DQM create queue\n");
252 
253 	list_add(&pqn->process_queue_list, &pqm->queues);
254 
255 	if (q) {
256 		pr_debug("PQM done creating queue\n");
257 		print_queue_properties(&q->properties);
258 	}
259 
260 	return retval;
261 
262 err_create_queue:
263 	kfree(pqn);
264 err_allocate_pqn:
265 	/* check if queues list is empty unregister process from device */
266 	clear_bit(*qid, pqm->queue_slot_bitmap);
267 	if (list_empty(&pdd->qpd.queues_list) &&
268 	    list_empty(&pdd->qpd.priv_queue_list))
269 		dev->dqm->ops.unregister_process(dev->dqm, &pdd->qpd);
270 	return retval;
271 }
272 
273 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid)
274 {
275 	struct process_queue_node *pqn;
276 	struct kfd_process_device *pdd;
277 	struct device_queue_manager *dqm;
278 	struct kfd_dev *dev;
279 	int retval;
280 
281 	dqm = NULL;
282 
283 	retval = 0;
284 
285 	pqn = get_queue_by_qid(pqm, qid);
286 	if (!pqn) {
287 		pr_err("Queue id does not match any known queue\n");
288 		return -EINVAL;
289 	}
290 
291 	dev = NULL;
292 	if (pqn->kq)
293 		dev = pqn->kq->dev;
294 	if (pqn->q)
295 		dev = pqn->q->device;
296 	if (WARN_ON(!dev))
297 		return -ENODEV;
298 
299 	pdd = kfd_get_process_device_data(dev, pqm->process);
300 	if (!pdd) {
301 		pr_err("Process device data doesn't exist\n");
302 		return -1;
303 	}
304 
305 	if (pqn->kq) {
306 		/* destroy kernel queue (DIQ) */
307 		dqm = pqn->kq->dev->dqm;
308 		dqm->ops.destroy_kernel_queue(dqm, pqn->kq, &pdd->qpd);
309 		kernel_queue_uninit(pqn->kq);
310 	}
311 
312 	if (pqn->q) {
313 		dqm = pqn->q->device->dqm;
314 		retval = dqm->ops.destroy_queue(dqm, &pdd->qpd, pqn->q);
315 		if (retval) {
316 			pr_debug("Destroy queue failed, returned %d\n", retval);
317 			goto err_destroy_queue;
318 		}
319 		uninit_queue(pqn->q);
320 	}
321 
322 	list_del(&pqn->process_queue_list);
323 	kfree(pqn);
324 	clear_bit(qid, pqm->queue_slot_bitmap);
325 
326 	if (list_empty(&pdd->qpd.queues_list) &&
327 	    list_empty(&pdd->qpd.priv_queue_list))
328 		dqm->ops.unregister_process(dqm, &pdd->qpd);
329 
330 err_destroy_queue:
331 	return retval;
332 }
333 
334 int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid,
335 			struct queue_properties *p)
336 {
337 	int retval;
338 	struct process_queue_node *pqn;
339 
340 	pqn = get_queue_by_qid(pqm, qid);
341 	if (!pqn) {
342 		pr_debug("No queue %d exists for update operation\n", qid);
343 		return -EFAULT;
344 	}
345 
346 	pqn->q->properties.queue_address = p->queue_address;
347 	pqn->q->properties.queue_size = p->queue_size;
348 	pqn->q->properties.queue_percent = p->queue_percent;
349 	pqn->q->properties.priority = p->priority;
350 
351 	retval = pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm,
352 							pqn->q);
353 	if (retval != 0)
354 		return retval;
355 
356 	return 0;
357 }
358 
359 struct kernel_queue *pqm_get_kernel_queue(
360 					struct process_queue_manager *pqm,
361 					unsigned int qid)
362 {
363 	struct process_queue_node *pqn;
364 
365 	pqn = get_queue_by_qid(pqm, qid);
366 	if (pqn && pqn->kq)
367 		return pqn->kq;
368 
369 	return NULL;
370 }
371 
372 #if defined(CONFIG_DEBUG_FS)
373 
374 int pqm_debugfs_mqds(struct seq_file *m, void *data)
375 {
376 	struct process_queue_manager *pqm = data;
377 	struct process_queue_node *pqn;
378 	struct queue *q;
379 	enum KFD_MQD_TYPE mqd_type;
380 	struct mqd_manager *mqd_manager;
381 	int r = 0;
382 
383 	list_for_each_entry(pqn, &pqm->queues, process_queue_list) {
384 		if (pqn->q) {
385 			q = pqn->q;
386 			switch (q->properties.type) {
387 			case KFD_QUEUE_TYPE_SDMA:
388 				seq_printf(m, "  SDMA queue on device %x\n",
389 					   q->device->id);
390 				mqd_type = KFD_MQD_TYPE_SDMA;
391 				break;
392 			case KFD_QUEUE_TYPE_COMPUTE:
393 				seq_printf(m, "  Compute queue on device %x\n",
394 					   q->device->id);
395 				mqd_type = KFD_MQD_TYPE_CP;
396 				break;
397 			default:
398 				seq_printf(m,
399 				"  Bad user queue type %d on device %x\n",
400 					   q->properties.type, q->device->id);
401 				continue;
402 			}
403 			mqd_manager = q->device->dqm->ops.get_mqd_manager(
404 				q->device->dqm, mqd_type);
405 		} else if (pqn->kq) {
406 			q = pqn->kq->queue;
407 			mqd_manager = pqn->kq->mqd;
408 			switch (q->properties.type) {
409 			case KFD_QUEUE_TYPE_DIQ:
410 				seq_printf(m, "  DIQ on device %x\n",
411 					   pqn->kq->dev->id);
412 				mqd_type = KFD_MQD_TYPE_HIQ;
413 				break;
414 			default:
415 				seq_printf(m,
416 				"  Bad kernel queue type %d on device %x\n",
417 					   q->properties.type,
418 					   pqn->kq->dev->id);
419 				continue;
420 			}
421 		} else {
422 			seq_printf(m,
423 		"  Weird: Queue node with neither kernel nor user queue\n");
424 			continue;
425 		}
426 
427 		r = mqd_manager->debugfs_show_mqd(m, q->mqd);
428 		if (r != 0)
429 			break;
430 	}
431 
432 	return r;
433 }
434 
435 #endif
436