projects / linux-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
drm/amdgpu: teach amdgpu how to enable interrupts for any pipe v3
[linux-block.git] / drivers / gpu / drm / amd / amdkfd / kfd_device_queue_manager.c
CommitLineData
64c7f8cf
BG		 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 <linux/types.h>
27#include <linux/printk.h>
28#include <linux/bitops.h>
99331a51 29#include <linux/sched.h>
64c7f8cf
BG		 30#include "kfd_priv.h"
31#include "kfd_device_queue_manager.h"
32#include "kfd_mqd_manager.h"
33#include "cik_regs.h"
34#include "kfd_kernel_queue.h"
64c7f8cf
BG		 35
36/* Size of the per-pipe EOP queue */
37#define CIK_HPD_EOP_BYTES_LOG2 11
38#define CIK_HPD_EOP_BYTES (1U << CIK_HPD_EOP_BYTES_LOG2)
39
64c7f8cf
BG		 40static int set_pasid_vmid_mapping(struct device_queue_manager *dqm,
41 unsigned int pasid, unsigned int vmid);
42
43static int create_compute_queue_nocpsch(struct device_queue_manager *dqm,
44 struct queue *q,
45 struct qcm_process_device *qpd);
bcea3081 46
64c7f8cf 47static int execute_queues_cpsch(struct device_queue_manager *dqm, bool lock);
992839ad
YS		 48static int destroy_queues_cpsch(struct device_queue_manager *dqm,
49 bool preempt_static_queues, bool lock);
64c7f8cf 50
bcea3081
BG		 51static int create_sdma_queue_nocpsch(struct device_queue_manager *dqm,
52 struct queue *q,
53 struct qcm_process_device *qpd);
54
55static void deallocate_sdma_queue(struct device_queue_manager *dqm,
56 unsigned int sdma_queue_id);
64c7f8cf 57
bcea3081
BG		 58static inline
59enum KFD_MQD_TYPE get_mqd_type_from_queue_type(enum kfd_queue_type type)
64c7f8cf 60{
bcea3081 61 if (type == KFD_QUEUE_TYPE_SDMA)
85d258f9
BG		 62 return KFD_MQD_TYPE_SDMA;
63 return KFD_MQD_TYPE_CP;
64c7f8cf
BG		 64}
65
1365aa62 66unsigned int get_first_pipe(struct device_queue_manager *dqm)
64c7f8cf 67{
1365aa62 68 BUG_ON(!dqm || !dqm->dev);
64c7f8cf
BG		 69 return dqm->dev->shared_resources.first_compute_pipe;
70}
71
64ea8f4a
OG		 72unsigned int get_pipes_num(struct device_queue_manager *dqm)
73{
74 BUG_ON(!dqm || !dqm->dev);
75 return dqm->dev->shared_resources.compute_pipe_count;
76}
77
64c7f8cf
BG		 78static inline unsigned int get_pipes_num_cpsch(void)
79{
80 return PIPE_PER_ME_CP_SCHEDULING;
81}
82
a22fc854 83void program_sh_mem_settings(struct device_queue_manager *dqm,
64c7f8cf
BG		 84 struct qcm_process_device *qpd)
85{
cea405b1
XZ		 86 return dqm->dev->kfd2kgd->program_sh_mem_settings(
87 dqm->dev->kgd, qpd->vmid,
64c7f8cf
BG		 88 qpd->sh_mem_config,
89 qpd->sh_mem_ape1_base,
90 qpd->sh_mem_ape1_limit,
91 qpd->sh_mem_bases);
92}
93
94static int allocate_vmid(struct device_queue_manager *dqm,
95 struct qcm_process_device *qpd,
96 struct queue *q)
97{
98 int bit, allocated_vmid;
99
100 if (dqm->vmid_bitmap == 0)
101 return -ENOMEM;
102
103 bit = find_first_bit((unsigned long *)&dqm->vmid_bitmap, CIK_VMID_NUM);
104 clear_bit(bit, (unsigned long *)&dqm->vmid_bitmap);
105
106 /* Kaveri kfd vmid's starts from vmid 8 */
107 allocated_vmid = bit + KFD_VMID_START_OFFSET;
108 pr_debug("kfd: vmid allocation %d\n", allocated_vmid);
109 qpd->vmid = allocated_vmid;
110 q->properties.vmid = allocated_vmid;
111
112 set_pasid_vmid_mapping(dqm, q->process->pasid, q->properties.vmid);
113 program_sh_mem_settings(dqm, qpd);
114
115 return 0;
116}
117
118static void deallocate_vmid(struct device_queue_manager *dqm,
119 struct qcm_process_device *qpd,
120 struct queue *q)
121{
122 int bit = qpd->vmid - KFD_VMID_START_OFFSET;
123
2030664b
BG		 124 /* Release the vmid mapping */
125 set_pasid_vmid_mapping(dqm, 0, qpd->vmid);
126
64c7f8cf
BG		 127 set_bit(bit, (unsigned long *)&dqm->vmid_bitmap);
128 qpd->vmid = 0;
129 q->properties.vmid = 0;
130}
131
132static int create_queue_nocpsch(struct device_queue_manager *dqm,
133 struct queue *q,
134 struct qcm_process_device *qpd,
135 int *allocated_vmid)
136{
137 int retval;
138
139 BUG_ON(!dqm || !q || !qpd || !allocated_vmid);
140
141 pr_debug("kfd: In func %s\n", __func__);
142 print_queue(q);
143
144 mutex_lock(&dqm->lock);
145
b8cbab04
OG		 146 if (dqm->total_queue_count >= max_num_of_queues_per_device) {
147 pr_warn("amdkfd: Can't create new usermode queue because %d queues were already created\n",
148 dqm->total_queue_count);
149 mutex_unlock(&dqm->lock);
150 return -EPERM;
151 }
152
64c7f8cf
BG		 153 if (list_empty(&qpd->queues_list)) {
154 retval = allocate_vmid(dqm, qpd, q);
155 if (retval != 0) {
156 mutex_unlock(&dqm->lock);
157 return retval;
158 }
159 }
160 *allocated_vmid = qpd->vmid;
161 q->properties.vmid = qpd->vmid;
162
bcea3081
BG		 163 if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE)
164 retval = create_compute_queue_nocpsch(dqm, q, qpd);
165 if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
166 retval = create_sdma_queue_nocpsch(dqm, q, qpd);
64c7f8cf
BG		 167
168 if (retval != 0) {
169 if (list_empty(&qpd->queues_list)) {
170 deallocate_vmid(dqm, qpd, q);
171 *allocated_vmid = 0;
172 }
173 mutex_unlock(&dqm->lock);
174 return retval;
175 }
176
177 list_add(&q->list, &qpd->queues_list);
b6819cec
JC		 178 if (q->properties.is_active)
179 dqm->queue_count++;
64c7f8cf 180
bcea3081
BG		 181 if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
182 dqm->sdma_queue_count++;
64c7f8cf 183
b8cbab04
OG		 184 /*
185 * Unconditionally increment this counter, regardless of the queue's
186 * type or whether the queue is active.
187 */
188 dqm->total_queue_count++;
189 pr_debug("Total of %d queues are accountable so far\n",
190 dqm->total_queue_count);
191
64c7f8cf
BG		 192 mutex_unlock(&dqm->lock);
193 return 0;
194}
195
196static int allocate_hqd(struct device_queue_manager *dqm, struct queue *q)
197{
198 bool set;
f0ec5b99 199 int pipe, bit, i;
64c7f8cf
BG		 200
201 set = false;
202
f0ec5b99
BG		 203 for (pipe = dqm->next_pipe_to_allocate, i = 0; i < get_pipes_num(dqm);
204 pipe = ((pipe + 1) % get_pipes_num(dqm)), ++i) {
64c7f8cf
BG		 205 if (dqm->allocated_queues[pipe] != 0) {
206 bit = find_first_bit(
207 (unsigned long *)&dqm->allocated_queues[pipe],
208 QUEUES_PER_PIPE);
209
210 clear_bit(bit,
211 (unsigned long *)&dqm->allocated_queues[pipe]);
212 q->pipe = pipe;
213 q->queue = bit;
214 set = true;
215 break;
216 }
217 }
218
991ca8ee 219 if (!set)
64c7f8cf
BG		 220 return -EBUSY;
221
222 pr_debug("kfd: DQM %s hqd slot - pipe (%d) queue(%d)\n",
223 __func__, q->pipe, q->queue);
224 /* horizontal hqd allocation */
225 dqm->next_pipe_to_allocate = (pipe + 1) % get_pipes_num(dqm);
226
227 return 0;
228}
229
230static inline void deallocate_hqd(struct device_queue_manager *dqm,
231 struct queue *q)
232{
233 set_bit(q->queue, (unsigned long *)&dqm->allocated_queues[q->pipe]);
234}
235
236static int create_compute_queue_nocpsch(struct device_queue_manager *dqm,
237 struct queue *q,
238 struct qcm_process_device *qpd)
239{
240 int retval;
241 struct mqd_manager *mqd;
242
243 BUG_ON(!dqm || !q || !qpd);
244
45c9a5e4 245 mqd = dqm->ops.get_mqd_manager(dqm, KFD_MQD_TYPE_COMPUTE);
64c7f8cf
BG		 246 if (mqd == NULL)
247 return -ENOMEM;
248
249 retval = allocate_hqd(dqm, q);
250 if (retval != 0)
251 return retval;
252
253 retval = mqd->init_mqd(mqd, &q->mqd, &q->mqd_mem_obj,
254 &q->gart_mqd_addr, &q->properties);
255 if (retval != 0) {
256 deallocate_hqd(dqm, q);
257 return retval;
258 }
259
030e416b
BG		 260 pr_debug("kfd: loading mqd to hqd on pipe (%d) queue (%d)\n",
261 q->pipe,
262 q->queue);
263
264 retval = mqd->load_mqd(mqd, q->mqd, q->pipe,
8dfe58b2 265 q->queue, (uint32_t __user *) q->properties.write_ptr);
030e416b
BG		 266 if (retval != 0) {
267 deallocate_hqd(dqm, q);
268 mqd->uninit_mqd(mqd, q->mqd, q->mqd_mem_obj);
269 return retval;
270 }
271
64c7f8cf
BG		 272 return 0;
273}
274
275static int destroy_queue_nocpsch(struct device_queue_manager *dqm,
276 struct qcm_process_device *qpd,
277 struct queue *q)
278{
279 int retval;
280 struct mqd_manager *mqd;
281
282 BUG_ON(!dqm || !q || !q->mqd || !qpd);
283
284 retval = 0;
285
286 pr_debug("kfd: In Func %s\n", __func__);
287
288 mutex_lock(&dqm->lock);
64c7f8cf 289
c2e1b3a4 290 if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE) {
45c9a5e4 291 mqd = dqm->ops.get_mqd_manager(dqm, KFD_MQD_TYPE_COMPUTE);
c2e1b3a4
BG		 292 if (mqd == NULL) {
293 retval = -ENOMEM;
294 goto out;
295 }
296 deallocate_hqd(dqm, q);
297 } else if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
45c9a5e4 298 mqd = dqm->ops.get_mqd_manager(dqm, KFD_MQD_TYPE_SDMA);
c2e1b3a4
BG		 299 if (mqd == NULL) {
300 retval = -ENOMEM;
301 goto out;
302 }
303 dqm->sdma_queue_count--;
304 deallocate_sdma_queue(dqm, q->sdma_id);
7113cd65
OG		 305 } else {
306 pr_debug("q->properties.type is invalid (%d)\n",
307 q->properties.type);
308 retval = -EINVAL;
64c7f8cf
BG		 309 goto out;
310 }
311
312 retval = mqd->destroy_mqd(mqd, q->mqd,
c2e1b3a4 313 KFD_PREEMPT_TYPE_WAVEFRONT_RESET,
64c7f8cf
BG		 314 QUEUE_PREEMPT_DEFAULT_TIMEOUT_MS,
315 q->pipe, q->queue);
316
317 if (retval != 0)
318 goto out;
319
64c7f8cf
BG		 320 mqd->uninit_mqd(mqd, q->mqd, q->mqd_mem_obj);
321
322 list_del(&q->list);
323 if (list_empty(&qpd->queues_list))
324 deallocate_vmid(dqm, qpd, q);
b6819cec
JC		 325 if (q->properties.is_active)
326 dqm->queue_count--;
b8cbab04
OG		 327
328 /*
329 * Unconditionally decrement this counter, regardless of the queue's
330 * type
331 */
332 dqm->total_queue_count--;
333 pr_debug("Total of %d queues are accountable so far\n",
334 dqm->total_queue_count);
335
64c7f8cf
BG		 336out:
337 mutex_unlock(&dqm->lock);
338 return retval;
339}
340
341static int update_queue(struct device_queue_manager *dqm, struct queue *q)
342{
343 int retval;
344 struct mqd_manager *mqd;
b6ffbab8 345 bool prev_active = false;
64c7f8cf
BG		 346
347 BUG_ON(!dqm || !q || !q->mqd);
348
349 mutex_lock(&dqm->lock);
0b3674ae
OG		 350 mqd = dqm->ops.get_mqd_manager(dqm,
351 get_mqd_type_from_queue_type(q->properties.type));
64c7f8cf
BG		 352 if (mqd == NULL) {
353 mutex_unlock(&dqm->lock);
354 return -ENOMEM;
355 }
356
991ca8ee 357 if (q->properties.is_active)
b6ffbab8
OG		 358 prev_active = true;
359
360 /*
361 *
362 * check active state vs. the previous state
363 * and modify counter accordingly
364 */
365 retval = mqd->update_mqd(mqd, q->mqd, &q->properties);
991ca8ee 366 if ((q->properties.is_active) && (!prev_active))
64c7f8cf 367 dqm->queue_count++;
991ca8ee 368 else if ((!q->properties.is_active) && (prev_active))
64c7f8cf
BG		 369 dqm->queue_count--;
370
371 if (sched_policy != KFD_SCHED_POLICY_NO_HWS)
372 retval = execute_queues_cpsch(dqm, false);
373
374 mutex_unlock(&dqm->lock);
375 return retval;
376}
377
378static struct mqd_manager *get_mqd_manager_nocpsch(
379 struct device_queue_manager *dqm, enum KFD_MQD_TYPE type)
380{
381 struct mqd_manager *mqd;
382
383 BUG_ON(!dqm || type >= KFD_MQD_TYPE_MAX);
384
385 pr_debug("kfd: In func %s mqd type %d\n", __func__, type);
386
387 mqd = dqm->mqds[type];
388 if (!mqd) {
389 mqd = mqd_manager_init(type, dqm->dev);
390 if (mqd == NULL)
391 pr_err("kfd: mqd manager is NULL");
392 dqm->mqds[type] = mqd;
393 }
394
395 return mqd;
396}
397
398static int register_process_nocpsch(struct device_queue_manager *dqm,
399 struct qcm_process_device *qpd)
400{
401 struct device_process_node *n;
a22fc854 402 int retval;
64c7f8cf
BG		 403
404 BUG_ON(!dqm || !qpd);
405
406 pr_debug("kfd: In func %s\n", __func__);
407
408 n = kzalloc(sizeof(struct device_process_node), GFP_KERNEL);
409 if (!n)
410 return -ENOMEM;
411
412 n->qpd = qpd;
413
414 mutex_lock(&dqm->lock);
415 list_add(&n->list, &dqm->queues);
416
a22fc854
BG		 417 retval = dqm->ops_asic_specific.register_process(dqm, qpd);
418
64c7f8cf
BG		 419 dqm->processes_count++;
420
421 mutex_unlock(&dqm->lock);
422
a22fc854 423 return retval;
64c7f8cf
BG		 424}
425
426static int unregister_process_nocpsch(struct device_queue_manager *dqm,
427 struct qcm_process_device *qpd)
428{
429 int retval;
430 struct device_process_node *cur, *next;
431
432 BUG_ON(!dqm || !qpd);
433
1e5ec956 434 pr_debug("In func %s\n", __func__);
64c7f8cf 435
1e5ec956
OG		 436 pr_debug("qpd->queues_list is %s\n",
437 list_empty(&qpd->queues_list) ? "empty" : "not empty");
64c7f8cf
BG		 438
439 retval = 0;
440 mutex_lock(&dqm->lock);
441
442 list_for_each_entry_safe(cur, next, &dqm->queues, list) {
443 if (qpd == cur->qpd) {
444 list_del(&cur->list);
f5d896bb 445 kfree(cur);
64c7f8cf
BG		 446 dqm->processes_count--;
447 goto out;
448 }
449 }
450 /* qpd not found in dqm list */
451 retval = 1;
452out:
453 mutex_unlock(&dqm->lock);
454 return retval;
455}
456
457static int
458set_pasid_vmid_mapping(struct device_queue_manager *dqm, unsigned int pasid,
459 unsigned int vmid)
460{
461 uint32_t pasid_mapping;
462
cea405b1
XZ		 463 pasid_mapping = (pasid == 0) ? 0 :
464 (uint32_t)pasid |
465 ATC_VMID_PASID_MAPPING_VALID;
466
467 return dqm->dev->kfd2kgd->set_pasid_vmid_mapping(
468 dqm->dev->kgd, pasid_mapping,
64c7f8cf
BG		 469 vmid);
470}
471
a22fc854 472int init_pipelines(struct device_queue_manager *dqm,
64c7f8cf
BG		 473 unsigned int pipes_num, unsigned int first_pipe)
474{
64c7f8cf
BG		 475 BUG_ON(!dqm || !dqm->dev);
476
477 pr_debug("kfd: In func %s\n", __func__);
478
64c7f8cf
BG		 479 return 0;
480}
481
2249d558
AL		 482static void init_interrupts(struct device_queue_manager *dqm)
483{
484 unsigned int i;
485
486 BUG_ON(dqm == NULL);
487
488 for (i = 0 ; i < get_pipes_num(dqm) ; i++)
489 dqm->dev->kfd2kgd->init_interrupts(dqm->dev->kgd,
490 i + get_first_pipe(dqm));
491}
492
64c7f8cf
BG		 493static int init_scheduler(struct device_queue_manager *dqm)
494{
495 int retval;
496
497 BUG_ON(!dqm);
498
499 pr_debug("kfd: In %s\n", __func__);
500
9fa843e7 501 retval = init_pipelines(dqm, get_pipes_num(dqm), get_first_pipe(dqm));
64c7f8cf
BG		 502 return retval;
503}
504
505static int initialize_nocpsch(struct device_queue_manager *dqm)
506{
507 int i;
508
509 BUG_ON(!dqm);
510
511 pr_debug("kfd: In func %s num of pipes: %d\n",
512 __func__, get_pipes_num(dqm));
513
514 mutex_init(&dqm->lock);
515 INIT_LIST_HEAD(&dqm->queues);
516 dqm->queue_count = dqm->next_pipe_to_allocate = 0;
bcea3081 517 dqm->sdma_queue_count = 0;
64c7f8cf
BG		 518 dqm->allocated_queues = kcalloc(get_pipes_num(dqm),
519 sizeof(unsigned int), GFP_KERNEL);
520 if (!dqm->allocated_queues) {
521 mutex_destroy(&dqm->lock);
522 return -ENOMEM;
523 }
524
525 for (i = 0; i < get_pipes_num(dqm); i++)
526 dqm->allocated_queues[i] = (1 << QUEUES_PER_PIPE) - 1;
527
528 dqm->vmid_bitmap = (1 << VMID_PER_DEVICE) - 1;
bcea3081 529 dqm->sdma_bitmap = (1 << CIK_SDMA_QUEUES) - 1;
64c7f8cf
BG		 530
531 init_scheduler(dqm);
532 return 0;
533}
534
535static void uninitialize_nocpsch(struct device_queue_manager *dqm)
536{
6f9d54fd
OG		 537 int i;
538
64c7f8cf
BG		 539 BUG_ON(!dqm);
540
541 BUG_ON(dqm->queue_count > 0 || dqm->processes_count > 0);
542
543 kfree(dqm->allocated_queues);
6f9d54fd
OG		 544 for (i = 0 ; i < KFD_MQD_TYPE_MAX ; i++)
545 kfree(dqm->mqds[i]);
64c7f8cf 546 mutex_destroy(&dqm->lock);
a86aa3ca 547 kfd_gtt_sa_free(dqm->dev, dqm->pipeline_mem);
64c7f8cf
BG		 548}
549
550static int start_nocpsch(struct device_queue_manager *dqm)
551{
2249d558 552 init_interrupts(dqm);
64c7f8cf
BG		 553 return 0;
554}
555
556static int stop_nocpsch(struct device_queue_manager *dqm)
557{
558 return 0;
559}
560
bcea3081
BG		 561static int allocate_sdma_queue(struct device_queue_manager *dqm,
562 unsigned int *sdma_queue_id)
563{
564 int bit;
565
566 if (dqm->sdma_bitmap == 0)
567 return -ENOMEM;
568
569 bit = find_first_bit((unsigned long *)&dqm->sdma_bitmap,
570 CIK_SDMA_QUEUES);
571
572 clear_bit(bit, (unsigned long *)&dqm->sdma_bitmap);
573 *sdma_queue_id = bit;
574
575 return 0;
576}
577
578static void deallocate_sdma_queue(struct device_queue_manager *dqm,
579 unsigned int sdma_queue_id)
580{
010b82e7 581 if (sdma_queue_id >= CIK_SDMA_QUEUES)
bcea3081
BG		 582 return;
583 set_bit(sdma_queue_id, (unsigned long *)&dqm->sdma_bitmap);
584}
585
bcea3081
BG		 586static int create_sdma_queue_nocpsch(struct device_queue_manager *dqm,
587 struct queue *q,
588 struct qcm_process_device *qpd)
589{
590 struct mqd_manager *mqd;
591 int retval;
592
45c9a5e4 593 mqd = dqm->ops.get_mqd_manager(dqm, KFD_MQD_TYPE_SDMA);
bcea3081
BG		 594 if (!mqd)
595 return -ENOMEM;
596
597 retval = allocate_sdma_queue(dqm, &q->sdma_id);
598 if (retval != 0)
599 return retval;
600
601 q->properties.sdma_queue_id = q->sdma_id % CIK_SDMA_QUEUES_PER_ENGINE;
602 q->properties.sdma_engine_id = q->sdma_id / CIK_SDMA_ENGINE_NUM;
603
604 pr_debug("kfd: sdma id is: %d\n", q->sdma_id);
605 pr_debug(" sdma queue id: %d\n", q->properties.sdma_queue_id);
606 pr_debug(" sdma engine id: %d\n", q->properties.sdma_engine_id);
607
3e3f6e1a 608 dqm->ops_asic_specific.init_sdma_vm(dqm, q, qpd);
bcea3081
BG		 609 retval = mqd->init_mqd(mqd, &q->mqd, &q->mqd_mem_obj,
610 &q->gart_mqd_addr, &q->properties);
611 if (retval != 0) {
612 deallocate_sdma_queue(dqm, q->sdma_id);
613 return retval;
614 }
615
4fadf6b6
BG		 616 retval = mqd->load_mqd(mqd, q->mqd, 0,
617 0, NULL);
618 if (retval != 0) {
619 deallocate_sdma_queue(dqm, q->sdma_id);
620 mqd->uninit_mqd(mqd, q->mqd, q->mqd_mem_obj);
621 return retval;
622 }
623
bcea3081
BG		 624 return 0;
625}
626
64c7f8cf
BG		 627/*
628 * Device Queue Manager implementation for cp scheduler
629 */
630
631static int set_sched_resources(struct device_queue_manager *dqm)
632{
633 struct scheduling_resources res;
634 unsigned int queue_num, queue_mask;
635
636 BUG_ON(!dqm);
637
638 pr_debug("kfd: In func %s\n", __func__);
639
640 queue_num = get_pipes_num_cpsch() * QUEUES_PER_PIPE;
641 queue_mask = (1 << queue_num) - 1;
642 res.vmid_mask = (1 << VMID_PER_DEVICE) - 1;
643 res.vmid_mask <<= KFD_VMID_START_OFFSET;
644 res.queue_mask = queue_mask << (get_first_pipe(dqm) * QUEUES_PER_PIPE);
645 res.gws_mask = res.oac_mask = res.gds_heap_base =
646 res.gds_heap_size = 0;
647
648 pr_debug("kfd: scheduling resources:\n"
649 " vmid mask: 0x%8X\n"
650 " queue mask: 0x%8llX\n",
651 res.vmid_mask, res.queue_mask);
652
653 return pm_send_set_resources(&dqm->packets, &res);
654}
655
656static int initialize_cpsch(struct device_queue_manager *dqm)
657{
658 int retval;
659
660 BUG_ON(!dqm);
661
662 pr_debug("kfd: In func %s num of pipes: %d\n",
663 __func__, get_pipes_num_cpsch());
664
665 mutex_init(&dqm->lock);
666 INIT_LIST_HEAD(&dqm->queues);
667 dqm->queue_count = dqm->processes_count = 0;
bcea3081 668 dqm->sdma_queue_count = 0;
64c7f8cf 669 dqm->active_runlist = false;
a22fc854 670 retval = dqm->ops_asic_specific.initialize(dqm);
64c7f8cf
BG		 671 if (retval != 0)
672 goto fail_init_pipelines;
673
674 return 0;
675
676fail_init_pipelines:
677 mutex_destroy(&dqm->lock);
678 return retval;
679}
680
681static int start_cpsch(struct device_queue_manager *dqm)
682{
683 struct device_process_node *node;
684 int retval;
685
686 BUG_ON(!dqm);
687
688 retval = 0;
689
690 retval = pm_init(&dqm->packets, dqm);
691 if (retval != 0)
692 goto fail_packet_manager_init;
693
694 retval = set_sched_resources(dqm);
695 if (retval != 0)
696 goto fail_set_sched_resources;
697
698 pr_debug("kfd: allocating fence memory\n");
699
700 /* allocate fence memory on the gart */
a86aa3ca
OG		 701 retval = kfd_gtt_sa_allocate(dqm->dev, sizeof(*dqm->fence_addr),
702 &dqm->fence_mem);
64c7f8cf
BG		 703
704 if (retval != 0)
705 goto fail_allocate_vidmem;
706
707 dqm->fence_addr = dqm->fence_mem->cpu_ptr;
708 dqm->fence_gpu_addr = dqm->fence_mem->gpu_addr;
2249d558
AL		 709
710 init_interrupts(dqm);
711
64c7f8cf
BG		 712 list_for_each_entry(node, &dqm->queues, list)
713 if (node->qpd->pqm->process && dqm->dev)
714 kfd_bind_process_to_device(dqm->dev,
715 node->qpd->pqm->process);
716
717 execute_queues_cpsch(dqm, true);
718
719 return 0;
720fail_allocate_vidmem:
721fail_set_sched_resources:
722 pm_uninit(&dqm->packets);
723fail_packet_manager_init:
724 return retval;
725}
726
727static int stop_cpsch(struct device_queue_manager *dqm)
728{
729 struct device_process_node *node;
730 struct kfd_process_device *pdd;
731
732 BUG_ON(!dqm);
733
992839ad 734 destroy_queues_cpsch(dqm, true, true);
64c7f8cf
BG		 735
736 list_for_each_entry(node, &dqm->queues, list) {
52a5fdce 737 pdd = qpd_to_pdd(node->qpd);
64c7f8cf
BG		 738 pdd->bound = false;
739 }
a86aa3ca 740 kfd_gtt_sa_free(dqm->dev, dqm->fence_mem);
64c7f8cf
BG		 741 pm_uninit(&dqm->packets);
742
743 return 0;
744}
745
746static int create_kernel_queue_cpsch(struct device_queue_manager *dqm,
747 struct kernel_queue *kq,
748 struct qcm_process_device *qpd)
749{
750 BUG_ON(!dqm || !kq || !qpd);
751
752 pr_debug("kfd: In func %s\n", __func__);
753
754 mutex_lock(&dqm->lock);
b8cbab04
OG		 755 if (dqm->total_queue_count >= max_num_of_queues_per_device) {
756 pr_warn("amdkfd: Can't create new kernel queue because %d queues were already created\n",
757 dqm->total_queue_count);
758 mutex_unlock(&dqm->lock);
759 return -EPERM;
760 }
761
762 /*
763 * Unconditionally increment this counter, regardless of the queue's
764 * type or whether the queue is active.
765 */
766 dqm->total_queue_count++;
767 pr_debug("Total of %d queues are accountable so far\n",
768 dqm->total_queue_count);
769
64c7f8cf
BG		 770 list_add(&kq->list, &qpd->priv_queue_list);
771 dqm->queue_count++;
772 qpd->is_debug = true;
773 execute_queues_cpsch(dqm, false);
774 mutex_unlock(&dqm->lock);
775
776 return 0;
777}
778
779static void destroy_kernel_queue_cpsch(struct device_queue_manager *dqm,
780 struct kernel_queue *kq,
781 struct qcm_process_device *qpd)
782{
783 BUG_ON(!dqm || !kq);
784
785 pr_debug("kfd: In %s\n", __func__);
786
787 mutex_lock(&dqm->lock);
992839ad
YS		 788 /* here we actually preempt the DIQ */
789 destroy_queues_cpsch(dqm, true, false);
64c7f8cf
BG		 790 list_del(&kq->list);
791 dqm->queue_count--;
792 qpd->is_debug = false;
793 execute_queues_cpsch(dqm, false);
b8cbab04
OG		 794 /*
795 * Unconditionally decrement this counter, regardless of the queue's
796 * type.
797 */
8b58f261 798 dqm->total_queue_count--;
b8cbab04
OG		 799 pr_debug("Total of %d queues are accountable so far\n",
800 dqm->total_queue_count);
64c7f8cf
BG		 801 mutex_unlock(&dqm->lock);
802}
803
bcea3081
BG		 804static void select_sdma_engine_id(struct queue *q)
805{
806 static int sdma_id;
807
808 q->sdma_id = sdma_id;
809 sdma_id = (sdma_id + 1) % 2;
810}
811
64c7f8cf
BG		 812static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q,
813 struct qcm_process_device *qpd, int *allocate_vmid)
814{
815 int retval;
816 struct mqd_manager *mqd;
817
818 BUG_ON(!dqm || !q || !qpd);
819
820 retval = 0;
821
822 if (allocate_vmid)
823 *allocate_vmid = 0;
824
825 mutex_lock(&dqm->lock);
826
b8cbab04
OG		 827 if (dqm->total_queue_count >= max_num_of_queues_per_device) {
828 pr_warn("amdkfd: Can't create new usermode queue because %d queues were already created\n",
829 dqm->total_queue_count);
830 retval = -EPERM;
831 goto out;
832 }
833
bcea3081
BG		 834 if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
835 select_sdma_engine_id(q);
836
45c9a5e4 837 mqd = dqm->ops.get_mqd_manager(dqm,
bcea3081
BG		 838 get_mqd_type_from_queue_type(q->properties.type));
839
64c7f8cf
BG		 840 if (mqd == NULL) {
841 mutex_unlock(&dqm->lock);
842 return -ENOMEM;
843 }
844
bdcddf95 845 dqm->ops_asic_specific.init_sdma_vm(dqm, q, qpd);
64c7f8cf
BG		 846 retval = mqd->init_mqd(mqd, &q->mqd, &q->mqd_mem_obj,
847 &q->gart_mqd_addr, &q->properties);
848 if (retval != 0)
849 goto out;
850
851 list_add(&q->list, &qpd->queues_list);
852 if (q->properties.is_active) {
853 dqm->queue_count++;
854 retval = execute_queues_cpsch(dqm, false);
855 }
856
bcea3081
BG		 857 if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
858 dqm->sdma_queue_count++;
b8cbab04
OG		 859 /*
860 * Unconditionally increment this counter, regardless of the queue's
861 * type or whether the queue is active.
862 */
863 dqm->total_queue_count++;
864
865 pr_debug("Total of %d queues are accountable so far\n",
866 dqm->total_queue_count);
867
64c7f8cf
BG		 868out:
869 mutex_unlock(&dqm->lock);
870 return retval;
871}
872
788bf83d 873int amdkfd_fence_wait_timeout(unsigned int *fence_addr,
d80d19bd
OG		 874 unsigned int fence_value,
875 unsigned long timeout)
64c7f8cf
BG		 876{
877 BUG_ON(!fence_addr);
878 timeout += jiffies;
879
880 while (*fence_addr != fence_value) {
881 if (time_after(jiffies, timeout)) {
882 pr_err("kfd: qcm fence wait loop timeout expired\n");
883 return -ETIME;
884 }
99331a51 885 schedule();
64c7f8cf
BG		 886 }
887
888 return 0;
889}
890
bcea3081
BG		 891static int destroy_sdma_queues(struct device_queue_manager *dqm,
892 unsigned int sdma_engine)
893{
894 return pm_send_unmap_queue(&dqm->packets, KFD_QUEUE_TYPE_SDMA,
992839ad 895 KFD_PREEMPT_TYPE_FILTER_DYNAMIC_QUEUES, 0, false,
bcea3081
BG		 896 sdma_engine);
897}
898
992839ad
YS		 899static int destroy_queues_cpsch(struct device_queue_manager *dqm,
900 bool preempt_static_queues, bool lock)
64c7f8cf
BG		 901{
902 int retval;
992839ad 903 enum kfd_preempt_type_filter preempt_type;
a82918f1 904 struct kfd_process_device *pdd;
64c7f8cf
BG		 905
906 BUG_ON(!dqm);
907
908 retval = 0;
909
910 if (lock)
911 mutex_lock(&dqm->lock);
991ca8ee 912 if (!dqm->active_runlist)
64c7f8cf 913 goto out;
bcea3081
BG		 914
915 pr_debug("kfd: Before destroying queues, sdma queue count is : %u\n",
916 dqm->sdma_queue_count);
917
918 if (dqm->sdma_queue_count > 0) {
919 destroy_sdma_queues(dqm, 0);
920 destroy_sdma_queues(dqm, 1);
921 }
922
992839ad
YS		 923 preempt_type = preempt_static_queues ?
924 KFD_PREEMPT_TYPE_FILTER_ALL_QUEUES :
925 KFD_PREEMPT_TYPE_FILTER_DYNAMIC_QUEUES;
926
64c7f8cf 927 retval = pm_send_unmap_queue(&dqm->packets, KFD_QUEUE_TYPE_COMPUTE,
992839ad 928 preempt_type, 0, false, 0);
64c7f8cf
BG		 929 if (retval != 0)
930 goto out;
931
932 *dqm->fence_addr = KFD_FENCE_INIT;
933 pm_send_query_status(&dqm->packets, dqm->fence_gpu_addr,
934 KFD_FENCE_COMPLETED);
935 /* should be timed out */
c3447e81 936 retval = amdkfd_fence_wait_timeout(dqm->fence_addr, KFD_FENCE_COMPLETED,
64c7f8cf 937 QUEUE_PREEMPT_DEFAULT_TIMEOUT_MS);
c3447e81 938 if (retval != 0) {
a82918f1
BG		 939 pdd = kfd_get_process_device_data(dqm->dev,
940 kfd_get_process(current));
941 pdd->reset_wavefronts = true;
c3447e81
BG		 942 goto out;
943 }
64c7f8cf
BG		 944 pm_release_ib(&dqm->packets);
945 dqm->active_runlist = false;
946
947out:
948 if (lock)
949 mutex_unlock(&dqm->lock);
950 return retval;
951}
952
953static int execute_queues_cpsch(struct device_queue_manager *dqm, bool lock)
954{
955 int retval;
956
957 BUG_ON(!dqm);
958
959 if (lock)
960 mutex_lock(&dqm->lock);
961
992839ad 962 retval = destroy_queues_cpsch(dqm, false, false);
64c7f8cf
BG		 963 if (retval != 0) {
964 pr_err("kfd: the cp might be in an unrecoverable state due to an unsuccessful queues preemption");
965 goto out;
966 }
967
968 if (dqm->queue_count <= 0 || dqm->processes_count <= 0) {
969 retval = 0;
970 goto out;
971 }
972
973 if (dqm->active_runlist) {
974 retval = 0;
975 goto out;
976 }
977
978 retval = pm_send_runlist(&dqm->packets, &dqm->queues);
979 if (retval != 0) {
980 pr_err("kfd: failed to execute runlist");
981 goto out;
982 }
983 dqm->active_runlist = true;
984
985out:
986 if (lock)
987 mutex_unlock(&dqm->lock);
988 return retval;
989}
990
991static int destroy_queue_cpsch(struct device_queue_manager *dqm,
992 struct qcm_process_device *qpd,
993 struct queue *q)
994{
995 int retval;
996 struct mqd_manager *mqd;
992839ad 997 bool preempt_all_queues;
64c7f8cf
BG		 998
999 BUG_ON(!dqm || !qpd || !q);
1000
992839ad
YS		 1001 preempt_all_queues = false;
1002
64c7f8cf
BG		 1003 retval = 0;
1004
1005 /* remove queue from list to prevent rescheduling after preemption */
1006 mutex_lock(&dqm->lock);
992839ad
YS		 1007
1008 if (qpd->is_debug) {
1009 /*
1010 * error, currently we do not allow to destroy a queue
1011 * of a currently debugged process
1012 */
1013 retval = -EBUSY;
1014 goto failed_try_destroy_debugged_queue;
1015
1016 }
1017
45c9a5e4 1018 mqd = dqm->ops.get_mqd_manager(dqm,
bcea3081 1019 get_mqd_type_from_queue_type(q->properties.type));
64c7f8cf
BG		 1020 if (!mqd) {
1021 retval = -ENOMEM;
1022 goto failed;
1023 }
1024
bcea3081
BG		 1025 if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
1026 dqm->sdma_queue_count--;
1027
64c7f8cf 1028 list_del(&q->list);
b6819cec
JC		 1029 if (q->properties.is_active)
1030 dqm->queue_count--;
64c7f8cf
BG		 1031
1032 execute_queues_cpsch(dqm, false);
1033
1034 mqd->uninit_mqd(mqd, q->mqd, q->mqd_mem_obj);
b8cbab04
OG		 1035
1036 /*
1037 * Unconditionally decrement this counter, regardless of the queue's
1038 * type
1039 */
1040 dqm->total_queue_count--;
1041 pr_debug("Total of %d queues are accountable so far\n",
1042 dqm->total_queue_count);
64c7f8cf
BG		 1043
1044 mutex_unlock(&dqm->lock);
1045
1046 return 0;
1047
1048failed:
992839ad
YS		 1049failed_try_destroy_debugged_queue:
1050
64c7f8cf
BG		 1051 mutex_unlock(&dqm->lock);
1052 return retval;
1053}
1054
1055/*
1056 * Low bits must be 0000/FFFF as required by HW, high bits must be 0 to
1057 * stay in user mode.
1058 */
1059#define APE1_FIXED_BITS_MASK 0xFFFF80000000FFFFULL
1060/* APE1 limit is inclusive and 64K aligned. */
1061#define APE1_LIMIT_ALIGNMENT 0xFFFF
1062
1063static bool set_cache_memory_policy(struct device_queue_manager *dqm,
1064 struct qcm_process_device *qpd,
1065 enum cache_policy default_policy,
1066 enum cache_policy alternate_policy,
1067 void __user *alternate_aperture_base,
1068 uint64_t alternate_aperture_size)
1069{
a22fc854 1070 bool retval;
64c7f8cf
BG		 1071
1072 pr_debug("kfd: In func %s\n", __func__);
1073
1074 mutex_lock(&dqm->lock);
1075
1076 if (alternate_aperture_size == 0) {
1077 /* base > limit disables APE1 */
1078 qpd->sh_mem_ape1_base = 1;
1079 qpd->sh_mem_ape1_limit = 0;
1080 } else {
1081 /*
1082 * In FSA64, APE1_Base[63:0] = { 16{SH_MEM_APE1_BASE[31]},
1083 * SH_MEM_APE1_BASE[31:0], 0x0000 }
1084 * APE1_Limit[63:0] = { 16{SH_MEM_APE1_LIMIT[31]},
1085 * SH_MEM_APE1_LIMIT[31:0], 0xFFFF }
1086 * Verify that the base and size parameters can be
1087 * represented in this format and convert them.
1088 * Additionally restrict APE1 to user-mode addresses.
1089 */
1090
1091 uint64_t base = (uintptr_t)alternate_aperture_base;
1092 uint64_t limit = base + alternate_aperture_size - 1;
1093
1094 if (limit <= base)
1095 goto out;
1096
1097 if ((base & APE1_FIXED_BITS_MASK) != 0)
1098 goto out;
1099
1100 if ((limit & APE1_FIXED_BITS_MASK) != APE1_LIMIT_ALIGNMENT)
1101 goto out;
1102
1103 qpd->sh_mem_ape1_base = base >> 16;
1104 qpd->sh_mem_ape1_limit = limit >> 16;
1105 }
1106
a22fc854
BG		 1107 retval = dqm->ops_asic_specific.set_cache_memory_policy(
1108 dqm,
1109 qpd,
1110 default_policy,
1111 alternate_policy,
1112 alternate_aperture_base,
1113 alternate_aperture_size);
64c7f8cf
BG		 1114
1115 if ((sched_policy == KFD_SCHED_POLICY_NO_HWS) && (qpd->vmid != 0))
1116 program_sh_mem_settings(dqm, qpd);
1117
1118 pr_debug("kfd: sh_mem_config: 0x%x, ape1_base: 0x%x, ape1_limit: 0x%x\n",
1119 qpd->sh_mem_config, qpd->sh_mem_ape1_base,
1120 qpd->sh_mem_ape1_limit);
1121
1122 mutex_unlock(&dqm->lock);
a22fc854 1123 return retval;
64c7f8cf
BG		 1124
1125out:
1126 mutex_unlock(&dqm->lock);
1127 return false;
1128}
1129
1130struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev)
1131{
1132 struct device_queue_manager *dqm;
1133
1134 BUG_ON(!dev);
1135
a22fc854
BG		 1136 pr_debug("kfd: loading device queue manager\n");
1137
64c7f8cf
BG		 1138 dqm = kzalloc(sizeof(struct device_queue_manager), GFP_KERNEL);
1139 if (!dqm)
1140 return NULL;
1141
1142 dqm->dev = dev;
1143 switch (sched_policy) {
1144 case KFD_SCHED_POLICY_HWS:
1145 case KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION:
1146 /* initialize dqm for cp scheduling */
45c9a5e4
OG		 1147 dqm->ops.create_queue = create_queue_cpsch;
1148 dqm->ops.initialize = initialize_cpsch;
1149 dqm->ops.start = start_cpsch;
1150 dqm->ops.stop = stop_cpsch;
1151 dqm->ops.destroy_queue = destroy_queue_cpsch;
1152 dqm->ops.update_queue = update_queue;
1153 dqm->ops.get_mqd_manager = get_mqd_manager_nocpsch;
1154 dqm->ops.register_process = register_process_nocpsch;
1155 dqm->ops.unregister_process = unregister_process_nocpsch;
1156 dqm->ops.uninitialize = uninitialize_nocpsch;
1157 dqm->ops.create_kernel_queue = create_kernel_queue_cpsch;
1158 dqm->ops.destroy_kernel_queue = destroy_kernel_queue_cpsch;
1159 dqm->ops.set_cache_memory_policy = set_cache_memory_policy;
64c7f8cf
BG		 1160 break;
1161 case KFD_SCHED_POLICY_NO_HWS:
1162 /* initialize dqm for no cp scheduling */
45c9a5e4
OG		 1163 dqm->ops.start = start_nocpsch;
1164 dqm->ops.stop = stop_nocpsch;
1165 dqm->ops.create_queue = create_queue_nocpsch;
1166 dqm->ops.destroy_queue = destroy_queue_nocpsch;
1167 dqm->ops.update_queue = update_queue;
1168 dqm->ops.get_mqd_manager = get_mqd_manager_nocpsch;
1169 dqm->ops.register_process = register_process_nocpsch;
1170 dqm->ops.unregister_process = unregister_process_nocpsch;
1171 dqm->ops.initialize = initialize_nocpsch;
1172 dqm->ops.uninitialize = uninitialize_nocpsch;
1173 dqm->ops.set_cache_memory_policy = set_cache_memory_policy;
64c7f8cf
BG		 1174 break;
1175 default:
1176 BUG();
1177 break;
1178 }
1179
a22fc854
BG		 1180 switch (dev->device_info->asic_family) {
1181 case CHIP_CARRIZO:
1182 device_queue_manager_init_vi(&dqm->ops_asic_specific);
300dec95
OG		 1183 break;
1184
a22fc854
BG		 1185 case CHIP_KAVERI:
1186 device_queue_manager_init_cik(&dqm->ops_asic_specific);
300dec95 1187 break;
a22fc854
BG		 1188 }
1189
45c9a5e4 1190 if (dqm->ops.initialize(dqm) != 0) {
64c7f8cf
BG		 1191 kfree(dqm);
1192 return NULL;
1193 }
1194
1195 return dqm;
1196}
1197
1198void device_queue_manager_uninit(struct device_queue_manager *dqm)
1199{
1200 BUG_ON(!dqm);
1201
45c9a5e4 1202 dqm->ops.uninitialize(dqm);
64c7f8cf
BG		 1203 kfree(dqm);
1204}
Linux 6.x block layer and io_uring tree(s)