1152490bbf532c4bf144b3ecea35e60bfe166e03
[linux-2.6-microblaze.git] / drivers / gpu / drm / amd / amdkfd / kfd_process_queue_manager.c
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 init_user_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                             uint32_t *p_doorbell_offset_in_process)
197 {
198         int retval;
199         struct kfd_process_device *pdd;
200         struct queue *q;
201         struct process_queue_node *pqn;
202         struct kernel_queue *kq;
203         enum kfd_queue_type type = properties->type;
204         unsigned int max_queues = 127; /* HWS limit */
205
206         q = NULL;
207         kq = NULL;
208
209         pdd = kfd_get_process_device_data(dev, pqm->process);
210         if (!pdd) {
211                 pr_err("Process device data doesn't exist\n");
212                 return -1;
213         }
214
215         /*
216          * for debug process, verify that it is within the static queues limit
217          * currently limit is set to half of the total avail HQD slots
218          * If we are just about to create DIQ, the is_debug flag is not set yet
219          * Hence we also check the type as well
220          */
221         if ((pdd->qpd.is_debug) || (type == KFD_QUEUE_TYPE_DIQ))
222                 max_queues = dev->device_info->max_no_of_hqd/2;
223
224         if (pdd->qpd.queue_count >= max_queues)
225                 return -ENOSPC;
226
227         retval = find_available_queue_slot(pqm, qid);
228         if (retval != 0)
229                 return retval;
230
231         if (list_empty(&pdd->qpd.queues_list) &&
232             list_empty(&pdd->qpd.priv_queue_list))
233                 dev->dqm->ops.register_process(dev->dqm, &pdd->qpd);
234
235         pqn = kzalloc(sizeof(*pqn), GFP_KERNEL);
236         if (!pqn) {
237                 retval = -ENOMEM;
238                 goto err_allocate_pqn;
239         }
240
241         switch (type) {
242         case KFD_QUEUE_TYPE_SDMA:
243         case KFD_QUEUE_TYPE_SDMA_XGMI:
244                 if ((type == KFD_QUEUE_TYPE_SDMA && dev->dqm->sdma_queue_count
245                         >= get_num_sdma_queues(dev->dqm)) ||
246                         (type == KFD_QUEUE_TYPE_SDMA_XGMI &&
247                         dev->dqm->xgmi_sdma_queue_count
248                         >= get_num_xgmi_sdma_queues(dev->dqm))) {
249                         pr_debug("Over-subscription is not allowed for SDMA.\n");
250                         retval = -EPERM;
251                         goto err_create_queue;
252                 }
253
254                 retval = init_user_queue(pqm, dev, &q, properties, f, *qid);
255                 if (retval != 0)
256                         goto err_create_queue;
257                 pqn->q = q;
258                 pqn->kq = NULL;
259                 retval = dev->dqm->ops.create_queue(dev->dqm, q, &pdd->qpd);
260                 pr_debug("DQM returned %d for create_queue\n", retval);
261                 print_queue(q);
262                 break;
263
264         case KFD_QUEUE_TYPE_COMPUTE:
265                 /* check if there is over subscription */
266                 if ((dev->dqm->sched_policy ==
267                      KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION) &&
268                 ((dev->dqm->processes_count >= dev->vm_info.vmid_num_kfd) ||
269                 (dev->dqm->queue_count >= get_queues_num(dev->dqm)))) {
270                         pr_debug("Over-subscription is not allowed when amdkfd.sched_policy == 1\n");
271                         retval = -EPERM;
272                         goto err_create_queue;
273                 }
274
275                 retval = init_user_queue(pqm, dev, &q, properties, f, *qid);
276                 if (retval != 0)
277                         goto err_create_queue;
278                 pqn->q = q;
279                 pqn->kq = NULL;
280                 retval = dev->dqm->ops.create_queue(dev->dqm, q, &pdd->qpd);
281                 pr_debug("DQM returned %d for create_queue\n", retval);
282                 print_queue(q);
283                 break;
284         case KFD_QUEUE_TYPE_DIQ:
285                 kq = kernel_queue_init(dev, KFD_QUEUE_TYPE_DIQ);
286                 if (!kq) {
287                         retval = -ENOMEM;
288                         goto err_create_queue;
289                 }
290                 kq->queue->properties.queue_id = *qid;
291                 pqn->kq = kq;
292                 pqn->q = NULL;
293                 retval = dev->dqm->ops.create_kernel_queue(dev->dqm,
294                                                         kq, &pdd->qpd);
295                 break;
296         default:
297                 WARN(1, "Invalid queue type %d", type);
298                 retval = -EINVAL;
299         }
300
301         if (retval != 0) {
302                 pr_err("Pasid 0x%x DQM create queue %d failed. ret %d\n",
303                         pqm->process->pasid, type, retval);
304                 goto err_create_queue;
305         }
306
307         if (q && p_doorbell_offset_in_process)
308                 /* Return the doorbell offset within the doorbell page
309                  * to the caller so it can be passed up to user mode
310                  * (in bytes).
311                  * There are always 1024 doorbells per process, so in case
312                  * of 8-byte doorbells, there are two doorbell pages per
313                  * process.
314                  */
315                 *p_doorbell_offset_in_process =
316                         (q->properties.doorbell_off * sizeof(uint32_t)) &
317                         (kfd_doorbell_process_slice(dev) - 1);
318
319         pr_debug("PQM After DQM create queue\n");
320
321         list_add(&pqn->process_queue_list, &pqm->queues);
322
323         if (q) {
324                 pr_debug("PQM done creating queue\n");
325                 print_queue_properties(&q->properties);
326         }
327
328         return retval;
329
330 err_create_queue:
331         kfree(pqn);
332 err_allocate_pqn:
333         /* check if queues list is empty unregister process from device */
334         clear_bit(*qid, pqm->queue_slot_bitmap);
335         if (list_empty(&pdd->qpd.queues_list) &&
336             list_empty(&pdd->qpd.priv_queue_list))
337                 dev->dqm->ops.unregister_process(dev->dqm, &pdd->qpd);
338         return retval;
339 }
340
341 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid)
342 {
343         struct process_queue_node *pqn;
344         struct kfd_process_device *pdd;
345         struct device_queue_manager *dqm;
346         struct kfd_dev *dev;
347         int retval;
348
349         dqm = NULL;
350
351         retval = 0;
352
353         pqn = get_queue_by_qid(pqm, qid);
354         if (!pqn) {
355                 pr_err("Queue id does not match any known queue\n");
356                 return -EINVAL;
357         }
358
359         dev = NULL;
360         if (pqn->kq)
361                 dev = pqn->kq->dev;
362         if (pqn->q)
363                 dev = pqn->q->device;
364         if (WARN_ON(!dev))
365                 return -ENODEV;
366
367         pdd = kfd_get_process_device_data(dev, pqm->process);
368         if (!pdd) {
369                 pr_err("Process device data doesn't exist\n");
370                 return -1;
371         }
372
373         if (pqn->kq) {
374                 /* destroy kernel queue (DIQ) */
375                 dqm = pqn->kq->dev->dqm;
376                 dqm->ops.destroy_kernel_queue(dqm, pqn->kq, &pdd->qpd);
377                 kernel_queue_uninit(pqn->kq);
378         }
379
380         if (pqn->q) {
381                 dqm = pqn->q->device->dqm;
382                 retval = dqm->ops.destroy_queue(dqm, &pdd->qpd, pqn->q);
383                 if (retval) {
384                         pr_err("Pasid 0x%x destroy queue %d failed, ret %d\n",
385                                 pqm->process->pasid,
386                                 pqn->q->properties.queue_id, retval);
387                         if (retval != -ETIME)
388                                 goto err_destroy_queue;
389                 }
390
391                 if (pqn->q->gws) {
392                         amdgpu_amdkfd_remove_gws_from_process(pqm->process->kgd_process_info,
393                                 pqn->q->gws);
394                         pdd->qpd.num_gws = 0;
395                 }
396
397                 kfree(pqn->q->properties.cu_mask);
398                 pqn->q->properties.cu_mask = NULL;
399                 uninit_queue(pqn->q);
400         }
401
402         list_del(&pqn->process_queue_list);
403         kfree(pqn);
404         clear_bit(qid, pqm->queue_slot_bitmap);
405
406         if (list_empty(&pdd->qpd.queues_list) &&
407             list_empty(&pdd->qpd.priv_queue_list))
408                 dqm->ops.unregister_process(dqm, &pdd->qpd);
409
410 err_destroy_queue:
411         return retval;
412 }
413
414 int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid,
415                         struct queue_properties *p)
416 {
417         int retval;
418         struct process_queue_node *pqn;
419
420         pqn = get_queue_by_qid(pqm, qid);
421         if (!pqn) {
422                 pr_debug("No queue %d exists for update operation\n", qid);
423                 return -EFAULT;
424         }
425
426         pqn->q->properties.queue_address = p->queue_address;
427         pqn->q->properties.queue_size = p->queue_size;
428         pqn->q->properties.queue_percent = p->queue_percent;
429         pqn->q->properties.priority = p->priority;
430
431         retval = pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm,
432                                                         pqn->q);
433         if (retval != 0)
434                 return retval;
435
436         return 0;
437 }
438
439 int pqm_set_cu_mask(struct process_queue_manager *pqm, unsigned int qid,
440                         struct queue_properties *p)
441 {
442         int retval;
443         struct process_queue_node *pqn;
444
445         pqn = get_queue_by_qid(pqm, qid);
446         if (!pqn) {
447                 pr_debug("No queue %d exists for update operation\n", qid);
448                 return -EFAULT;
449         }
450
451         /* Free the old CU mask memory if it is already allocated, then
452          * allocate memory for the new CU mask.
453          */
454         kfree(pqn->q->properties.cu_mask);
455
456         pqn->q->properties.cu_mask_count = p->cu_mask_count;
457         pqn->q->properties.cu_mask = p->cu_mask;
458
459         retval = pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm,
460                                                         pqn->q);
461         if (retval != 0)
462                 return retval;
463
464         return 0;
465 }
466
467 struct kernel_queue *pqm_get_kernel_queue(
468                                         struct process_queue_manager *pqm,
469                                         unsigned int qid)
470 {
471         struct process_queue_node *pqn;
472
473         pqn = get_queue_by_qid(pqm, qid);
474         if (pqn && pqn->kq)
475                 return pqn->kq;
476
477         return NULL;
478 }
479
480 int pqm_get_wave_state(struct process_queue_manager *pqm,
481                        unsigned int qid,
482                        void __user *ctl_stack,
483                        u32 *ctl_stack_used_size,
484                        u32 *save_area_used_size)
485 {
486         struct process_queue_node *pqn;
487
488         pqn = get_queue_by_qid(pqm, qid);
489         if (!pqn) {
490                 pr_debug("amdkfd: No queue %d exists for operation\n",
491                          qid);
492                 return -EFAULT;
493         }
494
495         return pqn->q->device->dqm->ops.get_wave_state(pqn->q->device->dqm,
496                                                        pqn->q,
497                                                        ctl_stack,
498                                                        ctl_stack_used_size,
499                                                        save_area_used_size);
500 }
501
502 #if defined(CONFIG_DEBUG_FS)
503
504 int pqm_debugfs_mqds(struct seq_file *m, void *data)
505 {
506         struct process_queue_manager *pqm = data;
507         struct process_queue_node *pqn;
508         struct queue *q;
509         enum KFD_MQD_TYPE mqd_type;
510         struct mqd_manager *mqd_mgr;
511         int r = 0;
512
513         list_for_each_entry(pqn, &pqm->queues, process_queue_list) {
514                 if (pqn->q) {
515                         q = pqn->q;
516                         switch (q->properties.type) {
517                         case KFD_QUEUE_TYPE_SDMA:
518                         case KFD_QUEUE_TYPE_SDMA_XGMI:
519                                 seq_printf(m, "  SDMA queue on device %x\n",
520                                            q->device->id);
521                                 mqd_type = KFD_MQD_TYPE_SDMA;
522                                 break;
523                         case KFD_QUEUE_TYPE_COMPUTE:
524                                 seq_printf(m, "  Compute queue on device %x\n",
525                                            q->device->id);
526                                 mqd_type = KFD_MQD_TYPE_CP;
527                                 break;
528                         default:
529                                 seq_printf(m,
530                                 "  Bad user queue type %d on device %x\n",
531                                            q->properties.type, q->device->id);
532                                 continue;
533                         }
534                         mqd_mgr = q->device->dqm->mqd_mgrs[mqd_type];
535                 } else if (pqn->kq) {
536                         q = pqn->kq->queue;
537                         mqd_mgr = pqn->kq->mqd_mgr;
538                         switch (q->properties.type) {
539                         case KFD_QUEUE_TYPE_DIQ:
540                                 seq_printf(m, "  DIQ on device %x\n",
541                                            pqn->kq->dev->id);
542                                 break;
543                         default:
544                                 seq_printf(m,
545                                 "  Bad kernel queue type %d on device %x\n",
546                                            q->properties.type,
547                                            pqn->kq->dev->id);
548                                 continue;
549                         }
550                 } else {
551                         seq_printf(m,
552                 "  Weird: Queue node with neither kernel nor user queue\n");
553                         continue;
554                 }
555
556                 r = mqd_mgr->debugfs_show_mqd(m, q->mqd);
557                 if (r != 0)
558                         break;
559         }
560
561         return r;
562 }
563
564 #endif