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 		pdd->qpd.pqm = pqm;
183 		dev->dqm->ops.register_process(dev->dqm, &pdd->qpd);
184 	}
185 
186 	pqn = kzalloc(sizeof(*pqn), GFP_KERNEL);
187 	if (!pqn) {
188 		retval = -ENOMEM;
189 		goto err_allocate_pqn;
190 	}
191 
192 	switch (type) {
193 	case KFD_QUEUE_TYPE_SDMA:
194 		if (dev->dqm->queue_count >=
195 			CIK_SDMA_QUEUES_PER_ENGINE * CIK_SDMA_ENGINE_NUM) {
196 			pr_err("Over-subscription is not allowed for SDMA.\n");
197 			retval = -EPERM;
198 			goto err_create_queue;
199 		}
200 
201 		retval = create_cp_queue(pqm, dev, &q, properties, f, *qid);
202 		if (retval != 0)
203 			goto err_create_queue;
204 		pqn->q = q;
205 		pqn->kq = NULL;
206 		retval = dev->dqm->ops.create_queue(dev->dqm, q, &pdd->qpd,
207 						&q->properties.vmid);
208 		pr_debug("DQM returned %d for create_queue\n", retval);
209 		print_queue(q);
210 		break;
211 
212 	case KFD_QUEUE_TYPE_COMPUTE:
213 		/* check if there is over subscription */
214 		if ((sched_policy == KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION) &&
215 		((dev->dqm->processes_count >= dev->vm_info.vmid_num_kfd) ||
216 		(dev->dqm->queue_count >= get_queues_num(dev->dqm)))) {
217 			pr_err("Over-subscription is not allowed in radeon_kfd.sched_policy == 1\n");
218 			retval = -EPERM;
219 			goto err_create_queue;
220 		}
221 
222 		retval = create_cp_queue(pqm, dev, &q, properties, f, *qid);
223 		if (retval != 0)
224 			goto err_create_queue;
225 		pqn->q = q;
226 		pqn->kq = NULL;
227 		retval = dev->dqm->ops.create_queue(dev->dqm, q, &pdd->qpd,
228 						&q->properties.vmid);
229 		pr_debug("DQM returned %d for create_queue\n", retval);
230 		print_queue(q);
231 		break;
232 	case KFD_QUEUE_TYPE_DIQ:
233 		kq = kernel_queue_init(dev, KFD_QUEUE_TYPE_DIQ);
234 		if (!kq) {
235 			retval = -ENOMEM;
236 			goto err_create_queue;
237 		}
238 		kq->queue->properties.queue_id = *qid;
239 		pqn->kq = kq;
240 		pqn->q = NULL;
241 		retval = dev->dqm->ops.create_kernel_queue(dev->dqm,
242 							kq, &pdd->qpd);
243 		break;
244 	default:
245 		WARN(1, "Invalid queue type %d", type);
246 		retval = -EINVAL;
247 	}
248 
249 	if (retval != 0) {
250 		pr_err("DQM create queue failed\n");
251 		goto err_create_queue;
252 	}
253 
254 	pr_debug("PQM After DQM create queue\n");
255 
256 	list_add(&pqn->process_queue_list, &pqm->queues);
257 
258 	if (q) {
259 		pr_debug("PQM done creating queue\n");
260 		print_queue_properties(&q->properties);
261 	}
262 
263 	return retval;
264 
265 err_create_queue:
266 	kfree(pqn);
267 err_allocate_pqn:
268 	/* check if queues list is empty unregister process from device */
269 	clear_bit(*qid, pqm->queue_slot_bitmap);
270 	if (list_empty(&pdd->qpd.queues_list) &&
271 	    list_empty(&pdd->qpd.priv_queue_list))
272 		dev->dqm->ops.unregister_process(dev->dqm, &pdd->qpd);
273 	return retval;
274 }
275 
276 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid)
277 {
278 	struct process_queue_node *pqn;
279 	struct kfd_process_device *pdd;
280 	struct device_queue_manager *dqm;
281 	struct kfd_dev *dev;
282 	int retval;
283 
284 	dqm = NULL;
285 
286 	retval = 0;
287 
288 	pqn = get_queue_by_qid(pqm, qid);
289 	if (!pqn) {
290 		pr_err("Queue id does not match any known queue\n");
291 		return -EINVAL;
292 	}
293 
294 	dev = NULL;
295 	if (pqn->kq)
296 		dev = pqn->kq->dev;
297 	if (pqn->q)
298 		dev = pqn->q->device;
299 	if (WARN_ON(!dev))
300 		return -ENODEV;
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 	if (pqn->kq) {
309 		/* destroy kernel queue (DIQ) */
310 		dqm = pqn->kq->dev->dqm;
311 		dqm->ops.destroy_kernel_queue(dqm, pqn->kq, &pdd->qpd);
312 		kernel_queue_uninit(pqn->kq);
313 	}
314 
315 	if (pqn->q) {
316 		dqm = pqn->q->device->dqm;
317 		retval = dqm->ops.destroy_queue(dqm, &pdd->qpd, pqn->q);
318 		uninit_queue(pqn->q);
319 	}
320 
321 	list_del(&pqn->process_queue_list);
322 	kfree(pqn);
323 	clear_bit(qid, pqm->queue_slot_bitmap);
324 
325 	if (list_empty(&pdd->qpd.queues_list) &&
326 	    list_empty(&pdd->qpd.priv_queue_list))
327 		dqm->ops.unregister_process(dqm, &pdd->qpd);
328 
329 	return retval;
330 }
331 
332 int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid,
333 			struct queue_properties *p)
334 {
335 	int retval;
336 	struct process_queue_node *pqn;
337 
338 	pqn = get_queue_by_qid(pqm, qid);
339 	if (!pqn) {
340 		pr_debug("No queue %d exists for update operation\n", qid);
341 		return -EFAULT;
342 	}
343 
344 	pqn->q->properties.queue_address = p->queue_address;
345 	pqn->q->properties.queue_size = p->queue_size;
346 	pqn->q->properties.queue_percent = p->queue_percent;
347 	pqn->q->properties.priority = p->priority;
348 
349 	retval = pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm,
350 							pqn->q);
351 	if (retval != 0)
352 		return retval;
353 
354 	return 0;
355 }
356 
357 struct kernel_queue *pqm_get_kernel_queue(
358 					struct process_queue_manager *pqm,
359 					unsigned int qid)
360 {
361 	struct process_queue_node *pqn;
362 
363 	pqn = get_queue_by_qid(pqm, qid);
364 	if (pqn && pqn->kq)
365 		return pqn->kq;
366 
367 	return NULL;
368 }
369 
370 
371