1 /* $NetBSD: kfd_process_queue_manager.c,v 1.3 2021/12/18 23:44:59 riastradh Exp $ */
2
3 /*
4 * Copyright 2014 Advanced Micro Devices, Inc.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
22 * OTHER DEALINGS IN THE SOFTWARE.
23 *
24 */
25
26 #include <sys/cdefs.h>
27 __KERNEL_RCSID(0, "$NetBSD: kfd_process_queue_manager.c,v 1.3 2021/12/18 23:44:59 riastradh Exp $");
28
29 #include <linux/slab.h>
30 #include <linux/list.h>
31 #include "kfd_device_queue_manager.h"
32 #include "kfd_priv.h"
33 #include "kfd_kernel_queue.h"
34 #include "amdgpu_amdkfd.h"
35
get_queue_by_qid(struct process_queue_manager * pqm,unsigned int qid)36 static inline struct process_queue_node *get_queue_by_qid(
37 struct process_queue_manager *pqm, unsigned int qid)
38 {
39 struct process_queue_node *pqn;
40
41 list_for_each_entry(pqn, &pqm->queues, process_queue_list) {
42 if ((pqn->q && pqn->q->properties.queue_id == qid) ||
43 (pqn->kq && pqn->kq->queue->properties.queue_id == qid))
44 return pqn;
45 }
46
47 return NULL;
48 }
49
find_available_queue_slot(struct process_queue_manager * pqm,unsigned int * qid)50 static int find_available_queue_slot(struct process_queue_manager *pqm,
51 unsigned int *qid)
52 {
53 unsigned long found;
54
55 found = find_first_zero_bit(pqm->queue_slot_bitmap,
56 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);
57
58 pr_debug("The new slot id %lu\n", found);
59
60 if (found >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) {
61 pr_info("Cannot open more queues for process with pasid 0x%x\n",
62 pqm->process->pasid);
63 return -ENOMEM;
64 }
65
66 set_bit(found, pqm->queue_slot_bitmap);
67 *qid = found;
68
69 return 0;
70 }
71
kfd_process_dequeue_from_device(struct kfd_process_device * pdd)72 void kfd_process_dequeue_from_device(struct kfd_process_device *pdd)
73 {
74 struct kfd_dev *dev = pdd->dev;
75
76 if (pdd->already_dequeued)
77 return;
78
79 dev->dqm->ops.process_termination(dev->dqm, &pdd->qpd);
80 pdd->already_dequeued = true;
81 }
82
pqm_set_gws(struct process_queue_manager * pqm,unsigned int qid,void * gws)83 int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid,
84 void *gws)
85 {
86 struct kfd_dev *dev = NULL;
87 struct process_queue_node *pqn;
88 struct kfd_process_device *pdd;
89 struct kgd_mem *mem = NULL;
90 int ret;
91
92 pqn = get_queue_by_qid(pqm, qid);
93 if (!pqn) {
94 pr_err("Queue id does not match any known queue\n");
95 return -EINVAL;
96 }
97
98 if (pqn->q)
99 dev = pqn->q->device;
100 if (WARN_ON(!dev))
101 return -ENODEV;
102
103 pdd = kfd_get_process_device_data(dev, pqm->process);
104 if (!pdd) {
105 pr_err("Process device data doesn't exist\n");
106 return -EINVAL;
107 }
108
109 /* Only allow one queue per process can have GWS assigned */
110 if (gws && pdd->qpd.num_gws)
111 return -EBUSY;
112
113 if (!gws && pdd->qpd.num_gws == 0)
114 return -EINVAL;
115
116 if (gws)
117 ret = amdgpu_amdkfd_add_gws_to_process(pdd->process->kgd_process_info,
118 gws, &mem);
119 else
120 ret = amdgpu_amdkfd_remove_gws_from_process(pdd->process->kgd_process_info,
121 pqn->q->gws);
122 if (unlikely(ret))
123 return ret;
124
125 pqn->q->gws = mem;
126 pdd->qpd.num_gws = gws ? amdgpu_amdkfd_get_num_gws(dev->kgd) : 0;
127
128 return pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm,
129 pqn->q);
130 }
131
kfd_process_dequeue_from_all_devices(struct kfd_process * p)132 void kfd_process_dequeue_from_all_devices(struct kfd_process *p)
133 {
134 struct kfd_process_device *pdd;
135
136 list_for_each_entry(pdd, &p->per_device_data, per_device_list)
137 kfd_process_dequeue_from_device(pdd);
138 }
139
pqm_init(struct process_queue_manager * pqm,struct kfd_process * p)140 int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p)
141 {
142 INIT_LIST_HEAD(&pqm->queues);
143 pqm->queue_slot_bitmap =
144 kzalloc(DIV_ROUND_UP(KFD_MAX_NUM_OF_QUEUES_PER_PROCESS,
145 BITS_PER_BYTE), GFP_KERNEL);
146 if (!pqm->queue_slot_bitmap)
147 return -ENOMEM;
148 pqm->process = p;
149
150 return 0;
151 }
152
pqm_uninit(struct process_queue_manager * pqm)153 void pqm_uninit(struct process_queue_manager *pqm)
154 {
155 struct process_queue_node *pqn, *next;
156
157 list_for_each_entry_safe(pqn, next, &pqm->queues, process_queue_list) {
158 if (pqn->q && pqn->q->gws)
159 amdgpu_amdkfd_remove_gws_from_process(pqm->process->kgd_process_info,
160 pqn->q->gws);
161 uninit_queue(pqn->q);
162 list_del(&pqn->process_queue_list);
163 kfree(pqn);
164 }
165
166 kfree(pqm->queue_slot_bitmap);
167 pqm->queue_slot_bitmap = NULL;
168 }
169
init_user_queue(struct process_queue_manager * pqm,struct kfd_dev * dev,struct queue ** q,struct queue_properties * q_properties,struct file * f,unsigned int qid)170 static int init_user_queue(struct process_queue_manager *pqm,
171 struct kfd_dev *dev, struct queue **q,
172 struct queue_properties *q_properties,
173 struct file *f, unsigned int qid)
174 {
175 int retval;
176
177 /* Doorbell initialized in user space*/
178 q_properties->doorbell_ptr = NULL;
179
180 /* let DQM handle it*/
181 q_properties->vmid = 0;
182 q_properties->queue_id = qid;
183
184 retval = init_queue(q, q_properties);
185 if (retval != 0)
186 return retval;
187
188 (*q)->device = dev;
189 (*q)->process = pqm->process;
190
191 pr_debug("PQM After init queue");
192
193 return retval;
194 }
195
pqm_create_queue(struct process_queue_manager * pqm,struct kfd_dev * dev,struct file * f,struct queue_properties * properties,unsigned int * qid,uint32_t * p_doorbell_offset_in_process)196 int pqm_create_queue(struct process_queue_manager *pqm,
197 struct kfd_dev *dev,
198 struct file *f,
199 struct queue_properties *properties,
200 unsigned int *qid,
201 uint32_t *p_doorbell_offset_in_process)
202 {
203 int retval;
204 struct kfd_process_device *pdd;
205 struct queue *q;
206 struct process_queue_node *pqn;
207 struct kernel_queue *kq;
208 enum kfd_queue_type type = properties->type;
209 unsigned int max_queues = 127; /* HWS limit */
210
211 q = NULL;
212 kq = NULL;
213
214 pdd = kfd_get_process_device_data(dev, pqm->process);
215 if (!pdd) {
216 pr_err("Process device data doesn't exist\n");
217 return -1;
218 }
219
220 /*
221 * for debug process, verify that it is within the static queues limit
222 * currently limit is set to half of the total avail HQD slots
223 * If we are just about to create DIQ, the is_debug flag is not set yet
224 * Hence we also check the type as well
225 */
226 if ((pdd->qpd.is_debug) || (type == KFD_QUEUE_TYPE_DIQ))
227 max_queues = dev->device_info->max_no_of_hqd/2;
228
229 if (pdd->qpd.queue_count >= max_queues)
230 return -ENOSPC;
231
232 retval = find_available_queue_slot(pqm, qid);
233 if (retval != 0)
234 return retval;
235
236 if (list_empty(&pdd->qpd.queues_list) &&
237 list_empty(&pdd->qpd.priv_queue_list))
238 dev->dqm->ops.register_process(dev->dqm, &pdd->qpd);
239
240 pqn = kzalloc(sizeof(*pqn), GFP_KERNEL);
241 if (!pqn) {
242 retval = -ENOMEM;
243 goto err_allocate_pqn;
244 }
245
246 switch (type) {
247 case KFD_QUEUE_TYPE_SDMA:
248 case KFD_QUEUE_TYPE_SDMA_XGMI:
249 if ((type == KFD_QUEUE_TYPE_SDMA && dev->dqm->sdma_queue_count
250 >= get_num_sdma_queues(dev->dqm)) ||
251 (type == KFD_QUEUE_TYPE_SDMA_XGMI &&
252 dev->dqm->xgmi_sdma_queue_count
253 >= get_num_xgmi_sdma_queues(dev->dqm))) {
254 pr_debug("Over-subscription is not allowed for SDMA.\n");
255 retval = -EPERM;
256 goto err_create_queue;
257 }
258
259 retval = init_user_queue(pqm, dev, &q, properties, f, *qid);
260 if (retval != 0)
261 goto err_create_queue;
262 pqn->q = q;
263 pqn->kq = NULL;
264 retval = dev->dqm->ops.create_queue(dev->dqm, q, &pdd->qpd);
265 pr_debug("DQM returned %d for create_queue\n", retval);
266 print_queue(q);
267 break;
268
269 case KFD_QUEUE_TYPE_COMPUTE:
270 /* check if there is over subscription */
271 if ((dev->dqm->sched_policy ==
272 KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION) &&
273 ((dev->dqm->processes_count >= dev->vm_info.vmid_num_kfd) ||
274 (dev->dqm->queue_count >= get_queues_num(dev->dqm)))) {
275 pr_debug("Over-subscription is not allowed when amdkfd.sched_policy == 1\n");
276 retval = -EPERM;
277 goto err_create_queue;
278 }
279
280 retval = init_user_queue(pqm, dev, &q, properties, f, *qid);
281 if (retval != 0)
282 goto err_create_queue;
283 pqn->q = q;
284 pqn->kq = NULL;
285 retval = dev->dqm->ops.create_queue(dev->dqm, q, &pdd->qpd);
286 pr_debug("DQM returned %d for create_queue\n", retval);
287 print_queue(q);
288 break;
289 case KFD_QUEUE_TYPE_DIQ:
290 kq = kernel_queue_init(dev, KFD_QUEUE_TYPE_DIQ);
291 if (!kq) {
292 retval = -ENOMEM;
293 goto err_create_queue;
294 }
295 kq->queue->properties.queue_id = *qid;
296 pqn->kq = kq;
297 pqn->q = NULL;
298 retval = dev->dqm->ops.create_kernel_queue(dev->dqm,
299 kq, &pdd->qpd);
300 break;
301 default:
302 WARN(1, "Invalid queue type %d", type);
303 retval = -EINVAL;
304 }
305
306 if (retval != 0) {
307 pr_err("Pasid 0x%x DQM create queue %d failed. ret %d\n",
308 pqm->process->pasid, type, retval);
309 goto err_create_queue;
310 }
311
312 if (q && p_doorbell_offset_in_process)
313 /* Return the doorbell offset within the doorbell page
314 * to the caller so it can be passed up to user mode
315 * (in bytes).
316 * There are always 1024 doorbells per process, so in case
317 * of 8-byte doorbells, there are two doorbell pages per
318 * process.
319 */
320 *p_doorbell_offset_in_process =
321 (q->properties.doorbell_off * sizeof(uint32_t)) &
322 (kfd_doorbell_process_slice(dev) - 1);
323
324 pr_debug("PQM After DQM create queue\n");
325
326 list_add(&pqn->process_queue_list, &pqm->queues);
327
328 if (q) {
329 pr_debug("PQM done creating queue\n");
330 print_queue_properties(&q->properties);
331 }
332
333 return retval;
334
335 err_create_queue:
336 kfree(pqn);
337 err_allocate_pqn:
338 /* check if queues list is empty unregister process from device */
339 clear_bit(*qid, pqm->queue_slot_bitmap);
340 if (list_empty(&pdd->qpd.queues_list) &&
341 list_empty(&pdd->qpd.priv_queue_list))
342 dev->dqm->ops.unregister_process(dev->dqm, &pdd->qpd);
343 return retval;
344 }
345
pqm_destroy_queue(struct process_queue_manager * pqm,unsigned int qid)346 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid)
347 {
348 struct process_queue_node *pqn;
349 struct kfd_process_device *pdd;
350 struct device_queue_manager *dqm;
351 struct kfd_dev *dev;
352 int retval;
353
354 dqm = NULL;
355
356 retval = 0;
357
358 pqn = get_queue_by_qid(pqm, qid);
359 if (!pqn) {
360 pr_err("Queue id does not match any known queue\n");
361 return -EINVAL;
362 }
363
364 dev = NULL;
365 if (pqn->kq)
366 dev = pqn->kq->dev;
367 if (pqn->q)
368 dev = pqn->q->device;
369 if (WARN_ON(!dev))
370 return -ENODEV;
371
372 pdd = kfd_get_process_device_data(dev, pqm->process);
373 if (!pdd) {
374 pr_err("Process device data doesn't exist\n");
375 return -1;
376 }
377
378 if (pqn->kq) {
379 /* destroy kernel queue (DIQ) */
380 dqm = pqn->kq->dev->dqm;
381 dqm->ops.destroy_kernel_queue(dqm, pqn->kq, &pdd->qpd);
382 kernel_queue_uninit(pqn->kq, false);
383 }
384
385 if (pqn->q) {
386 dqm = pqn->q->device->dqm;
387 retval = dqm->ops.destroy_queue(dqm, &pdd->qpd, pqn->q);
388 if (retval) {
389 pr_err("Pasid 0x%x destroy queue %d failed, ret %d\n",
390 pqm->process->pasid,
391 pqn->q->properties.queue_id, retval);
392 if (retval != -ETIME)
393 goto err_destroy_queue;
394 }
395
396 if (pqn->q->gws) {
397 amdgpu_amdkfd_remove_gws_from_process(pqm->process->kgd_process_info,
398 pqn->q->gws);
399 pdd->qpd.num_gws = 0;
400 }
401
402 kfree(pqn->q->properties.cu_mask);
403 pqn->q->properties.cu_mask = NULL;
404 uninit_queue(pqn->q);
405 }
406
407 list_del(&pqn->process_queue_list);
408 kfree(pqn);
409 clear_bit(qid, pqm->queue_slot_bitmap);
410
411 if (list_empty(&pdd->qpd.queues_list) &&
412 list_empty(&pdd->qpd.priv_queue_list))
413 dqm->ops.unregister_process(dqm, &pdd->qpd);
414
415 err_destroy_queue:
416 return retval;
417 }
418
pqm_update_queue(struct process_queue_manager * pqm,unsigned int qid,struct queue_properties * p)419 int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid,
420 struct queue_properties *p)
421 {
422 int retval;
423 struct process_queue_node *pqn;
424
425 pqn = get_queue_by_qid(pqm, qid);
426 if (!pqn) {
427 pr_debug("No queue %d exists for update operation\n", qid);
428 return -EFAULT;
429 }
430
431 pqn->q->properties.queue_address = p->queue_address;
432 pqn->q->properties.queue_size = p->queue_size;
433 pqn->q->properties.queue_percent = p->queue_percent;
434 pqn->q->properties.priority = p->priority;
435
436 retval = pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm,
437 pqn->q);
438 if (retval != 0)
439 return retval;
440
441 return 0;
442 }
443
pqm_set_cu_mask(struct process_queue_manager * pqm,unsigned int qid,struct queue_properties * p)444 int pqm_set_cu_mask(struct process_queue_manager *pqm, unsigned int qid,
445 struct queue_properties *p)
446 {
447 int retval;
448 struct process_queue_node *pqn;
449
450 pqn = get_queue_by_qid(pqm, qid);
451 if (!pqn) {
452 pr_debug("No queue %d exists for update operation\n", qid);
453 return -EFAULT;
454 }
455
456 /* Free the old CU mask memory if it is already allocated, then
457 * allocate memory for the new CU mask.
458 */
459 kfree(pqn->q->properties.cu_mask);
460
461 pqn->q->properties.cu_mask_count = p->cu_mask_count;
462 pqn->q->properties.cu_mask = p->cu_mask;
463
464 retval = pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm,
465 pqn->q);
466 if (retval != 0)
467 return retval;
468
469 return 0;
470 }
471
pqm_get_kernel_queue(struct process_queue_manager * pqm,unsigned int qid)472 struct kernel_queue *pqm_get_kernel_queue(
473 struct process_queue_manager *pqm,
474 unsigned int qid)
475 {
476 struct process_queue_node *pqn;
477
478 pqn = get_queue_by_qid(pqm, qid);
479 if (pqn && pqn->kq)
480 return pqn->kq;
481
482 return NULL;
483 }
484
pqm_get_wave_state(struct process_queue_manager * pqm,unsigned int qid,void __user * ctl_stack,u32 * ctl_stack_used_size,u32 * save_area_used_size)485 int pqm_get_wave_state(struct process_queue_manager *pqm,
486 unsigned int qid,
487 void __user *ctl_stack,
488 u32 *ctl_stack_used_size,
489 u32 *save_area_used_size)
490 {
491 struct process_queue_node *pqn;
492
493 pqn = get_queue_by_qid(pqm, qid);
494 if (!pqn) {
495 pr_debug("amdkfd: No queue %d exists for operation\n",
496 qid);
497 return -EFAULT;
498 }
499
500 return pqn->q->device->dqm->ops.get_wave_state(pqn->q->device->dqm,
501 pqn->q,
502 ctl_stack,
503 ctl_stack_used_size,
504 save_area_used_size);
505 }
506
507 #if defined(CONFIG_DEBUG_FS)
508
pqm_debugfs_mqds(struct seq_file * m,void * data)509 int pqm_debugfs_mqds(struct seq_file *m, void *data)
510 {
511 struct process_queue_manager *pqm = data;
512 struct process_queue_node *pqn;
513 struct queue *q;
514 enum KFD_MQD_TYPE mqd_type;
515 struct mqd_manager *mqd_mgr;
516 int r = 0;
517
518 list_for_each_entry(pqn, &pqm->queues, process_queue_list) {
519 if (pqn->q) {
520 q = pqn->q;
521 switch (q->properties.type) {
522 case KFD_QUEUE_TYPE_SDMA:
523 case KFD_QUEUE_TYPE_SDMA_XGMI:
524 seq_printf(m, " SDMA queue on device %x\n",
525 q->device->id);
526 mqd_type = KFD_MQD_TYPE_SDMA;
527 break;
528 case KFD_QUEUE_TYPE_COMPUTE:
529 seq_printf(m, " Compute queue on device %x\n",
530 q->device->id);
531 mqd_type = KFD_MQD_TYPE_CP;
532 break;
533 default:
534 seq_printf(m,
535 " Bad user queue type %d on device %x\n",
536 q->properties.type, q->device->id);
537 continue;
538 }
539 mqd_mgr = q->device->dqm->mqd_mgrs[mqd_type];
540 } else if (pqn->kq) {
541 q = pqn->kq->queue;
542 mqd_mgr = pqn->kq->mqd_mgr;
543 switch (q->properties.type) {
544 case KFD_QUEUE_TYPE_DIQ:
545 seq_printf(m, " DIQ on device %x\n",
546 pqn->kq->dev->id);
547 break;
548 default:
549 seq_printf(m,
550 " Bad kernel queue type %d on device %x\n",
551 q->properties.type,
552 pqn->kq->dev->id);
553 continue;
554 }
555 } else {
556 seq_printf(m,
557 " Weird: Queue node with neither kernel nor user queue\n");
558 continue;
559 }
560
561 r = mqd_mgr->debugfs_show_mqd(m, q->mqd);
562 if (r != 0)
563 break;
564 }
565
566 return r;
567 }
568
569 #endif
570