Loading...
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/mutex.h>
26#include "kfd_device_queue_manager.h"
27#include "kfd_kernel_queue.h"
28#include "kfd_priv.h"
29#include "kfd_pm4_headers.h"
30#include "kfd_pm4_headers_vi.h"
31#include "kfd_pm4_opcodes.h"
32
33static inline void inc_wptr(unsigned int *wptr, unsigned int increment_bytes,
34 unsigned int buffer_size_bytes)
35{
36 unsigned int temp = *wptr + increment_bytes / sizeof(uint32_t);
37
38 BUG_ON((temp * sizeof(uint32_t)) > buffer_size_bytes);
39 *wptr = temp;
40}
41
42static unsigned int build_pm4_header(unsigned int opcode, size_t packet_size)
43{
44 union PM4_MES_TYPE_3_HEADER header;
45
46 header.u32all = 0;
47 header.opcode = opcode;
48 header.count = packet_size/sizeof(uint32_t) - 2;
49 header.type = PM4_TYPE_3;
50
51 return header.u32all;
52}
53
54static void pm_calc_rlib_size(struct packet_manager *pm,
55 unsigned int *rlib_size,
56 bool *over_subscription)
57{
58 unsigned int process_count, queue_count;
59 unsigned int map_queue_size;
60
61 BUG_ON(!pm || !rlib_size || !over_subscription);
62
63 process_count = pm->dqm->processes_count;
64 queue_count = pm->dqm->queue_count;
65
66 /* check if there is over subscription*/
67 *over_subscription = false;
68 if ((process_count > 1) ||
69 queue_count > PIPE_PER_ME_CP_SCHEDULING * QUEUES_PER_PIPE) {
70 *over_subscription = true;
71 pr_debug("kfd: over subscribed runlist\n");
72 }
73
74 map_queue_size =
75 (pm->dqm->dev->device_info->asic_family == CHIP_CARRIZO) ?
76 sizeof(struct pm4_mes_map_queues) :
77 sizeof(struct pm4_map_queues);
78 /* calculate run list ib allocation size */
79 *rlib_size = process_count * sizeof(struct pm4_map_process) +
80 queue_count * map_queue_size;
81
82 /*
83 * Increase the allocation size in case we need a chained run list
84 * when over subscription
85 */
86 if (*over_subscription)
87 *rlib_size += sizeof(struct pm4_runlist);
88
89 pr_debug("kfd: runlist ib size %d\n", *rlib_size);
90}
91
92static int pm_allocate_runlist_ib(struct packet_manager *pm,
93 unsigned int **rl_buffer,
94 uint64_t *rl_gpu_buffer,
95 unsigned int *rl_buffer_size,
96 bool *is_over_subscription)
97{
98 int retval;
99
100 BUG_ON(!pm);
101 BUG_ON(pm->allocated == true);
102 BUG_ON(is_over_subscription == NULL);
103
104 pm_calc_rlib_size(pm, rl_buffer_size, is_over_subscription);
105
106 retval = kfd_gtt_sa_allocate(pm->dqm->dev, *rl_buffer_size,
107 &pm->ib_buffer_obj);
108
109 if (retval != 0) {
110 pr_err("kfd: failed to allocate runlist IB\n");
111 return retval;
112 }
113
114 *(void **)rl_buffer = pm->ib_buffer_obj->cpu_ptr;
115 *rl_gpu_buffer = pm->ib_buffer_obj->gpu_addr;
116
117 memset(*rl_buffer, 0, *rl_buffer_size);
118 pm->allocated = true;
119 return retval;
120}
121
122static int pm_create_runlist(struct packet_manager *pm, uint32_t *buffer,
123 uint64_t ib, size_t ib_size_in_dwords, bool chain)
124{
125 struct pm4_runlist *packet;
126
127 BUG_ON(!pm || !buffer || !ib);
128
129 packet = (struct pm4_runlist *)buffer;
130
131 memset(buffer, 0, sizeof(struct pm4_runlist));
132 packet->header.u32all = build_pm4_header(IT_RUN_LIST,
133 sizeof(struct pm4_runlist));
134
135 packet->bitfields4.ib_size = ib_size_in_dwords;
136 packet->bitfields4.chain = chain ? 1 : 0;
137 packet->bitfields4.offload_polling = 0;
138 packet->bitfields4.valid = 1;
139 packet->ordinal2 = lower_32_bits(ib);
140 packet->bitfields3.ib_base_hi = upper_32_bits(ib);
141
142 return 0;
143}
144
145static int pm_create_map_process(struct packet_manager *pm, uint32_t *buffer,
146 struct qcm_process_device *qpd)
147{
148 struct pm4_map_process *packet;
149 struct queue *cur;
150 uint32_t num_queues;
151
152 BUG_ON(!pm || !buffer || !qpd);
153
154 packet = (struct pm4_map_process *)buffer;
155
156 pr_debug("kfd: In func %s\n", __func__);
157
158 memset(buffer, 0, sizeof(struct pm4_map_process));
159
160 packet->header.u32all = build_pm4_header(IT_MAP_PROCESS,
161 sizeof(struct pm4_map_process));
162 packet->bitfields2.diq_enable = (qpd->is_debug) ? 1 : 0;
163 packet->bitfields2.process_quantum = 1;
164 packet->bitfields2.pasid = qpd->pqm->process->pasid;
165 packet->bitfields3.page_table_base = qpd->page_table_base;
166 packet->bitfields10.gds_size = qpd->gds_size;
167 packet->bitfields10.num_gws = qpd->num_gws;
168 packet->bitfields10.num_oac = qpd->num_oac;
169 num_queues = 0;
170 list_for_each_entry(cur, &qpd->queues_list, list)
171 num_queues++;
172 packet->bitfields10.num_queues = (qpd->is_debug) ? 0 : num_queues;
173
174 packet->sh_mem_config = qpd->sh_mem_config;
175 packet->sh_mem_bases = qpd->sh_mem_bases;
176 packet->sh_mem_ape1_base = qpd->sh_mem_ape1_base;
177 packet->sh_mem_ape1_limit = qpd->sh_mem_ape1_limit;
178
179 packet->gds_addr_lo = lower_32_bits(qpd->gds_context_area);
180 packet->gds_addr_hi = upper_32_bits(qpd->gds_context_area);
181
182 return 0;
183}
184
185static int pm_create_map_queue_vi(struct packet_manager *pm, uint32_t *buffer,
186 struct queue *q, bool is_static)
187{
188 struct pm4_mes_map_queues *packet;
189 bool use_static = is_static;
190
191 BUG_ON(!pm || !buffer || !q);
192
193 pr_debug("kfd: In func %s\n", __func__);
194
195 packet = (struct pm4_mes_map_queues *)buffer;
196 memset(buffer, 0, sizeof(struct pm4_map_queues));
197
198 packet->header.u32all = build_pm4_header(IT_MAP_QUEUES,
199 sizeof(struct pm4_map_queues));
200 packet->bitfields2.alloc_format =
201 alloc_format__mes_map_queues__one_per_pipe_vi;
202 packet->bitfields2.num_queues = 1;
203 packet->bitfields2.queue_sel =
204 queue_sel__mes_map_queues__map_to_hws_determined_queue_slots_vi;
205
206 packet->bitfields2.engine_sel =
207 engine_sel__mes_map_queues__compute_vi;
208 packet->bitfields2.queue_type =
209 queue_type__mes_map_queues__normal_compute_vi;
210
211 switch (q->properties.type) {
212 case KFD_QUEUE_TYPE_COMPUTE:
213 if (use_static)
214 packet->bitfields2.queue_type =
215 queue_type__mes_map_queues__normal_latency_static_queue_vi;
216 break;
217 case KFD_QUEUE_TYPE_DIQ:
218 packet->bitfields2.queue_type =
219 queue_type__mes_map_queues__debug_interface_queue_vi;
220 break;
221 case KFD_QUEUE_TYPE_SDMA:
222 packet->bitfields2.engine_sel =
223 engine_sel__mes_map_queues__sdma0_vi;
224 use_static = false; /* no static queues under SDMA */
225 break;
226 default:
227 pr_err("kfd: in %s queue type %d\n", __func__,
228 q->properties.type);
229 BUG();
230 break;
231 }
232 packet->bitfields3.doorbell_offset =
233 q->properties.doorbell_off;
234
235 packet->mqd_addr_lo =
236 lower_32_bits(q->gart_mqd_addr);
237
238 packet->mqd_addr_hi =
239 upper_32_bits(q->gart_mqd_addr);
240
241 packet->wptr_addr_lo =
242 lower_32_bits((uint64_t)q->properties.write_ptr);
243
244 packet->wptr_addr_hi =
245 upper_32_bits((uint64_t)q->properties.write_ptr);
246
247 return 0;
248}
249
250static int pm_create_map_queue(struct packet_manager *pm, uint32_t *buffer,
251 struct queue *q, bool is_static)
252{
253 struct pm4_map_queues *packet;
254 bool use_static = is_static;
255
256 BUG_ON(!pm || !buffer || !q);
257
258 pr_debug("kfd: In func %s\n", __func__);
259
260 packet = (struct pm4_map_queues *)buffer;
261 memset(buffer, 0, sizeof(struct pm4_map_queues));
262
263 packet->header.u32all = build_pm4_header(IT_MAP_QUEUES,
264 sizeof(struct pm4_map_queues));
265 packet->bitfields2.alloc_format =
266 alloc_format__mes_map_queues__one_per_pipe;
267 packet->bitfields2.num_queues = 1;
268 packet->bitfields2.queue_sel =
269 queue_sel__mes_map_queues__map_to_hws_determined_queue_slots;
270
271 packet->bitfields2.vidmem = (q->properties.is_interop) ?
272 vidmem__mes_map_queues__uses_video_memory :
273 vidmem__mes_map_queues__uses_no_video_memory;
274
275 switch (q->properties.type) {
276 case KFD_QUEUE_TYPE_COMPUTE:
277 case KFD_QUEUE_TYPE_DIQ:
278 packet->bitfields2.engine_sel =
279 engine_sel__mes_map_queues__compute;
280 break;
281 case KFD_QUEUE_TYPE_SDMA:
282 packet->bitfields2.engine_sel =
283 engine_sel__mes_map_queues__sdma0;
284 use_static = false; /* no static queues under SDMA */
285 break;
286 default:
287 BUG();
288 break;
289 }
290
291 packet->mes_map_queues_ordinals[0].bitfields3.doorbell_offset =
292 q->properties.doorbell_off;
293
294 packet->mes_map_queues_ordinals[0].bitfields3.is_static =
295 (use_static == true) ? 1 : 0;
296
297 packet->mes_map_queues_ordinals[0].mqd_addr_lo =
298 lower_32_bits(q->gart_mqd_addr);
299
300 packet->mes_map_queues_ordinals[0].mqd_addr_hi =
301 upper_32_bits(q->gart_mqd_addr);
302
303 packet->mes_map_queues_ordinals[0].wptr_addr_lo =
304 lower_32_bits((uint64_t)q->properties.write_ptr);
305
306 packet->mes_map_queues_ordinals[0].wptr_addr_hi =
307 upper_32_bits((uint64_t)q->properties.write_ptr);
308
309 return 0;
310}
311
312static int pm_create_runlist_ib(struct packet_manager *pm,
313 struct list_head *queues,
314 uint64_t *rl_gpu_addr,
315 size_t *rl_size_bytes)
316{
317 unsigned int alloc_size_bytes;
318 unsigned int *rl_buffer, rl_wptr, i;
319 int retval, proccesses_mapped;
320 struct device_process_node *cur;
321 struct qcm_process_device *qpd;
322 struct queue *q;
323 struct kernel_queue *kq;
324 bool is_over_subscription;
325
326 BUG_ON(!pm || !queues || !rl_size_bytes || !rl_gpu_addr);
327
328 rl_wptr = retval = proccesses_mapped = 0;
329
330 retval = pm_allocate_runlist_ib(pm, &rl_buffer, rl_gpu_addr,
331 &alloc_size_bytes, &is_over_subscription);
332 if (retval != 0)
333 return retval;
334
335 *rl_size_bytes = alloc_size_bytes;
336
337 pr_debug("kfd: In func %s\n", __func__);
338 pr_debug("kfd: building runlist ib process count: %d queues count %d\n",
339 pm->dqm->processes_count, pm->dqm->queue_count);
340
341 /* build the run list ib packet */
342 list_for_each_entry(cur, queues, list) {
343 qpd = cur->qpd;
344 /* build map process packet */
345 if (proccesses_mapped >= pm->dqm->processes_count) {
346 pr_debug("kfd: not enough space left in runlist IB\n");
347 pm_release_ib(pm);
348 return -ENOMEM;
349 }
350
351 retval = pm_create_map_process(pm, &rl_buffer[rl_wptr], qpd);
352 if (retval != 0)
353 return retval;
354
355 proccesses_mapped++;
356 inc_wptr(&rl_wptr, sizeof(struct pm4_map_process),
357 alloc_size_bytes);
358
359 list_for_each_entry(kq, &qpd->priv_queue_list, list) {
360 if (kq->queue->properties.is_active != true)
361 continue;
362
363 pr_debug("kfd: static_queue, mapping kernel q %d, is debug status %d\n",
364 kq->queue->queue, qpd->is_debug);
365
366 if (pm->dqm->dev->device_info->asic_family ==
367 CHIP_CARRIZO)
368 retval = pm_create_map_queue_vi(pm,
369 &rl_buffer[rl_wptr],
370 kq->queue,
371 qpd->is_debug);
372 else
373 retval = pm_create_map_queue(pm,
374 &rl_buffer[rl_wptr],
375 kq->queue,
376 qpd->is_debug);
377 if (retval != 0)
378 return retval;
379
380 inc_wptr(&rl_wptr,
381 sizeof(struct pm4_map_queues),
382 alloc_size_bytes);
383 }
384
385 list_for_each_entry(q, &qpd->queues_list, list) {
386 if (q->properties.is_active != true)
387 continue;
388
389 pr_debug("kfd: static_queue, mapping user queue %d, is debug status %d\n",
390 q->queue, qpd->is_debug);
391
392 if (pm->dqm->dev->device_info->asic_family ==
393 CHIP_CARRIZO)
394 retval = pm_create_map_queue_vi(pm,
395 &rl_buffer[rl_wptr],
396 q,
397 qpd->is_debug);
398 else
399 retval = pm_create_map_queue(pm,
400 &rl_buffer[rl_wptr],
401 q,
402 qpd->is_debug);
403
404 if (retval != 0)
405 return retval;
406
407 inc_wptr(&rl_wptr,
408 sizeof(struct pm4_map_queues),
409 alloc_size_bytes);
410 }
411 }
412
413 pr_debug("kfd: finished map process and queues to runlist\n");
414
415 if (is_over_subscription)
416 pm_create_runlist(pm, &rl_buffer[rl_wptr], *rl_gpu_addr,
417 alloc_size_bytes / sizeof(uint32_t), true);
418
419 for (i = 0; i < alloc_size_bytes / sizeof(uint32_t); i++)
420 pr_debug("0x%2X ", rl_buffer[i]);
421 pr_debug("\n");
422
423 return 0;
424}
425
426int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm)
427{
428 BUG_ON(!dqm);
429
430 pm->dqm = dqm;
431 mutex_init(&pm->lock);
432 pm->priv_queue = kernel_queue_init(dqm->dev, KFD_QUEUE_TYPE_HIQ);
433 if (pm->priv_queue == NULL) {
434 mutex_destroy(&pm->lock);
435 return -ENOMEM;
436 }
437 pm->allocated = false;
438
439 return 0;
440}
441
442void pm_uninit(struct packet_manager *pm)
443{
444 BUG_ON(!pm);
445
446 mutex_destroy(&pm->lock);
447 kernel_queue_uninit(pm->priv_queue);
448}
449
450int pm_send_set_resources(struct packet_manager *pm,
451 struct scheduling_resources *res)
452{
453 struct pm4_set_resources *packet;
454
455 BUG_ON(!pm || !res);
456
457 pr_debug("kfd: In func %s\n", __func__);
458
459 mutex_lock(&pm->lock);
460 pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue,
461 sizeof(*packet) / sizeof(uint32_t),
462 (unsigned int **)&packet);
463 if (packet == NULL) {
464 mutex_unlock(&pm->lock);
465 pr_err("kfd: failed to allocate buffer on kernel queue\n");
466 return -ENOMEM;
467 }
468
469 memset(packet, 0, sizeof(struct pm4_set_resources));
470 packet->header.u32all = build_pm4_header(IT_SET_RESOURCES,
471 sizeof(struct pm4_set_resources));
472
473 packet->bitfields2.queue_type =
474 queue_type__mes_set_resources__hsa_interface_queue_hiq;
475 packet->bitfields2.vmid_mask = res->vmid_mask;
476 packet->bitfields2.unmap_latency = KFD_UNMAP_LATENCY;
477 packet->bitfields7.oac_mask = res->oac_mask;
478 packet->bitfields8.gds_heap_base = res->gds_heap_base;
479 packet->bitfields8.gds_heap_size = res->gds_heap_size;
480
481 packet->gws_mask_lo = lower_32_bits(res->gws_mask);
482 packet->gws_mask_hi = upper_32_bits(res->gws_mask);
483
484 packet->queue_mask_lo = lower_32_bits(res->queue_mask);
485 packet->queue_mask_hi = upper_32_bits(res->queue_mask);
486
487 pm->priv_queue->ops.submit_packet(pm->priv_queue);
488
489 mutex_unlock(&pm->lock);
490
491 return 0;
492}
493
494int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues)
495{
496 uint64_t rl_gpu_ib_addr;
497 uint32_t *rl_buffer;
498 size_t rl_ib_size, packet_size_dwords;
499 int retval;
500
501 BUG_ON(!pm || !dqm_queues);
502
503 retval = pm_create_runlist_ib(pm, dqm_queues, &rl_gpu_ib_addr,
504 &rl_ib_size);
505 if (retval != 0)
506 goto fail_create_runlist_ib;
507
508 pr_debug("kfd: runlist IB address: 0x%llX\n", rl_gpu_ib_addr);
509
510 packet_size_dwords = sizeof(struct pm4_runlist) / sizeof(uint32_t);
511 mutex_lock(&pm->lock);
512
513 retval = pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue,
514 packet_size_dwords, &rl_buffer);
515 if (retval != 0)
516 goto fail_acquire_packet_buffer;
517
518 retval = pm_create_runlist(pm, rl_buffer, rl_gpu_ib_addr,
519 rl_ib_size / sizeof(uint32_t), false);
520 if (retval != 0)
521 goto fail_create_runlist;
522
523 pm->priv_queue->ops.submit_packet(pm->priv_queue);
524
525 mutex_unlock(&pm->lock);
526
527 return retval;
528
529fail_create_runlist:
530 pm->priv_queue->ops.rollback_packet(pm->priv_queue);
531fail_acquire_packet_buffer:
532 mutex_unlock(&pm->lock);
533fail_create_runlist_ib:
534 if (pm->allocated == true)
535 pm_release_ib(pm);
536 return retval;
537}
538
539int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
540 uint32_t fence_value)
541{
542 int retval;
543 struct pm4_query_status *packet;
544
545 BUG_ON(!pm || !fence_address);
546
547 mutex_lock(&pm->lock);
548 retval = pm->priv_queue->ops.acquire_packet_buffer(
549 pm->priv_queue,
550 sizeof(struct pm4_query_status) / sizeof(uint32_t),
551 (unsigned int **)&packet);
552 if (retval != 0)
553 goto fail_acquire_packet_buffer;
554
555 packet->header.u32all = build_pm4_header(IT_QUERY_STATUS,
556 sizeof(struct pm4_query_status));
557
558 packet->bitfields2.context_id = 0;
559 packet->bitfields2.interrupt_sel =
560 interrupt_sel__mes_query_status__completion_status;
561 packet->bitfields2.command =
562 command__mes_query_status__fence_only_after_write_ack;
563
564 packet->addr_hi = upper_32_bits((uint64_t)fence_address);
565 packet->addr_lo = lower_32_bits((uint64_t)fence_address);
566 packet->data_hi = upper_32_bits((uint64_t)fence_value);
567 packet->data_lo = lower_32_bits((uint64_t)fence_value);
568
569 pm->priv_queue->ops.submit_packet(pm->priv_queue);
570 mutex_unlock(&pm->lock);
571
572 return 0;
573
574fail_acquire_packet_buffer:
575 mutex_unlock(&pm->lock);
576 return retval;
577}
578
579int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type,
580 enum kfd_preempt_type_filter mode,
581 uint32_t filter_param, bool reset,
582 unsigned int sdma_engine)
583{
584 int retval;
585 uint32_t *buffer;
586 struct pm4_unmap_queues *packet;
587
588 BUG_ON(!pm);
589
590 mutex_lock(&pm->lock);
591 retval = pm->priv_queue->ops.acquire_packet_buffer(
592 pm->priv_queue,
593 sizeof(struct pm4_unmap_queues) / sizeof(uint32_t),
594 &buffer);
595 if (retval != 0)
596 goto err_acquire_packet_buffer;
597
598 packet = (struct pm4_unmap_queues *)buffer;
599 memset(buffer, 0, sizeof(struct pm4_unmap_queues));
600 pr_debug("kfd: static_queue: unmapping queues: mode is %d , reset is %d , type is %d\n",
601 mode, reset, type);
602 packet->header.u32all = build_pm4_header(IT_UNMAP_QUEUES,
603 sizeof(struct pm4_unmap_queues));
604 switch (type) {
605 case KFD_QUEUE_TYPE_COMPUTE:
606 case KFD_QUEUE_TYPE_DIQ:
607 packet->bitfields2.engine_sel =
608 engine_sel__mes_unmap_queues__compute;
609 break;
610 case KFD_QUEUE_TYPE_SDMA:
611 packet->bitfields2.engine_sel =
612 engine_sel__mes_unmap_queues__sdma0 + sdma_engine;
613 break;
614 default:
615 BUG();
616 break;
617 }
618
619 if (reset)
620 packet->bitfields2.action =
621 action__mes_unmap_queues__reset_queues;
622 else
623 packet->bitfields2.action =
624 action__mes_unmap_queues__preempt_queues;
625
626 switch (mode) {
627 case KFD_PREEMPT_TYPE_FILTER_SINGLE_QUEUE:
628 packet->bitfields2.queue_sel =
629 queue_sel__mes_unmap_queues__perform_request_on_specified_queues;
630 packet->bitfields2.num_queues = 1;
631 packet->bitfields3b.doorbell_offset0 = filter_param;
632 break;
633 case KFD_PREEMPT_TYPE_FILTER_BY_PASID:
634 packet->bitfields2.queue_sel =
635 queue_sel__mes_unmap_queues__perform_request_on_pasid_queues;
636 packet->bitfields3a.pasid = filter_param;
637 break;
638 case KFD_PREEMPT_TYPE_FILTER_ALL_QUEUES:
639 packet->bitfields2.queue_sel =
640 queue_sel__mes_unmap_queues__perform_request_on_all_active_queues;
641 break;
642 case KFD_PREEMPT_TYPE_FILTER_DYNAMIC_QUEUES:
643 /* in this case, we do not preempt static queues */
644 packet->bitfields2.queue_sel =
645 queue_sel__mes_unmap_queues__perform_request_on_dynamic_queues_only;
646 break;
647 default:
648 BUG();
649 break;
650 };
651
652 pm->priv_queue->ops.submit_packet(pm->priv_queue);
653
654 mutex_unlock(&pm->lock);
655 return 0;
656
657err_acquire_packet_buffer:
658 mutex_unlock(&pm->lock);
659 return retval;
660}
661
662void pm_release_ib(struct packet_manager *pm)
663{
664 BUG_ON(!pm);
665
666 mutex_lock(&pm->lock);
667 if (pm->allocated) {
668 kfd_gtt_sa_free(pm->dqm->dev, pm->ib_buffer_obj);
669 pm->allocated = false;
670 }
671 mutex_unlock(&pm->lock);
672}
1// SPDX-License-Identifier: GPL-2.0 OR MIT
2/*
3 * Copyright 2014-2022 Advanced Micro Devices, Inc.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice shall be included in
13 * all copies or substantial portions of the Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
19 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
21 * OTHER DEALINGS IN THE SOFTWARE.
22 *
23 */
24
25#include <linux/slab.h>
26#include <linux/mutex.h>
27#include "kfd_device_queue_manager.h"
28#include "kfd_kernel_queue.h"
29#include "kfd_priv.h"
30
31static inline void inc_wptr(unsigned int *wptr, unsigned int increment_bytes,
32 unsigned int buffer_size_bytes)
33{
34 unsigned int temp = *wptr + increment_bytes / sizeof(uint32_t);
35
36 WARN((temp * sizeof(uint32_t)) > buffer_size_bytes,
37 "Runlist IB overflow");
38 *wptr = temp;
39}
40
41static void pm_calc_rlib_size(struct packet_manager *pm,
42 unsigned int *rlib_size,
43 bool *over_subscription)
44{
45 unsigned int process_count, queue_count, compute_queue_count, gws_queue_count;
46 unsigned int map_queue_size;
47 unsigned int max_proc_per_quantum = 1;
48 struct kfd_dev *dev = pm->dqm->dev;
49
50 process_count = pm->dqm->processes_count;
51 queue_count = pm->dqm->active_queue_count;
52 compute_queue_count = pm->dqm->active_cp_queue_count;
53 gws_queue_count = pm->dqm->gws_queue_count;
54
55 /* check if there is over subscription
56 * Note: the arbitration between the number of VMIDs and
57 * hws_max_conc_proc has been done in
58 * kgd2kfd_device_init().
59 */
60 *over_subscription = false;
61
62 if (dev->max_proc_per_quantum > 1)
63 max_proc_per_quantum = dev->max_proc_per_quantum;
64
65 if ((process_count > max_proc_per_quantum) ||
66 compute_queue_count > get_cp_queues_num(pm->dqm) ||
67 gws_queue_count > 1) {
68 *over_subscription = true;
69 pr_debug("Over subscribed runlist\n");
70 }
71
72 map_queue_size = pm->pmf->map_queues_size;
73 /* calculate run list ib allocation size */
74 *rlib_size = process_count * pm->pmf->map_process_size +
75 queue_count * map_queue_size;
76
77 /*
78 * Increase the allocation size in case we need a chained run list
79 * when over subscription
80 */
81 if (*over_subscription)
82 *rlib_size += pm->pmf->runlist_size;
83
84 pr_debug("runlist ib size %d\n", *rlib_size);
85}
86
87static int pm_allocate_runlist_ib(struct packet_manager *pm,
88 unsigned int **rl_buffer,
89 uint64_t *rl_gpu_buffer,
90 unsigned int *rl_buffer_size,
91 bool *is_over_subscription)
92{
93 int retval;
94
95 if (WARN_ON(pm->allocated))
96 return -EINVAL;
97
98 pm_calc_rlib_size(pm, rl_buffer_size, is_over_subscription);
99
100 mutex_lock(&pm->lock);
101
102 retval = kfd_gtt_sa_allocate(pm->dqm->dev, *rl_buffer_size,
103 &pm->ib_buffer_obj);
104
105 if (retval) {
106 pr_err("Failed to allocate runlist IB\n");
107 goto out;
108 }
109
110 *(void **)rl_buffer = pm->ib_buffer_obj->cpu_ptr;
111 *rl_gpu_buffer = pm->ib_buffer_obj->gpu_addr;
112
113 memset(*rl_buffer, 0, *rl_buffer_size);
114 pm->allocated = true;
115
116out:
117 mutex_unlock(&pm->lock);
118 return retval;
119}
120
121static int pm_create_runlist_ib(struct packet_manager *pm,
122 struct list_head *queues,
123 uint64_t *rl_gpu_addr,
124 size_t *rl_size_bytes)
125{
126 unsigned int alloc_size_bytes;
127 unsigned int *rl_buffer, rl_wptr, i;
128 int retval, processes_mapped;
129 struct device_process_node *cur;
130 struct qcm_process_device *qpd;
131 struct queue *q;
132 struct kernel_queue *kq;
133 bool is_over_subscription;
134
135 rl_wptr = retval = processes_mapped = 0;
136
137 retval = pm_allocate_runlist_ib(pm, &rl_buffer, rl_gpu_addr,
138 &alloc_size_bytes, &is_over_subscription);
139 if (retval)
140 return retval;
141
142 *rl_size_bytes = alloc_size_bytes;
143 pm->ib_size_bytes = alloc_size_bytes;
144
145 pr_debug("Building runlist ib process count: %d queues count %d\n",
146 pm->dqm->processes_count, pm->dqm->active_queue_count);
147
148 /* build the run list ib packet */
149 list_for_each_entry(cur, queues, list) {
150 qpd = cur->qpd;
151 /* build map process packet */
152 if (processes_mapped >= pm->dqm->processes_count) {
153 pr_debug("Not enough space left in runlist IB\n");
154 pm_release_ib(pm);
155 return -ENOMEM;
156 }
157
158 retval = pm->pmf->map_process(pm, &rl_buffer[rl_wptr], qpd);
159 if (retval)
160 return retval;
161
162 processes_mapped++;
163 inc_wptr(&rl_wptr, pm->pmf->map_process_size,
164 alloc_size_bytes);
165
166 list_for_each_entry(kq, &qpd->priv_queue_list, list) {
167 if (!kq->queue->properties.is_active)
168 continue;
169
170 pr_debug("static_queue, mapping kernel q %d, is debug status %d\n",
171 kq->queue->queue, qpd->is_debug);
172
173 retval = pm->pmf->map_queues(pm,
174 &rl_buffer[rl_wptr],
175 kq->queue,
176 qpd->is_debug);
177 if (retval)
178 return retval;
179
180 inc_wptr(&rl_wptr,
181 pm->pmf->map_queues_size,
182 alloc_size_bytes);
183 }
184
185 list_for_each_entry(q, &qpd->queues_list, list) {
186 if (!q->properties.is_active)
187 continue;
188
189 pr_debug("static_queue, mapping user queue %d, is debug status %d\n",
190 q->queue, qpd->is_debug);
191
192 retval = pm->pmf->map_queues(pm,
193 &rl_buffer[rl_wptr],
194 q,
195 qpd->is_debug);
196
197 if (retval)
198 return retval;
199
200 inc_wptr(&rl_wptr,
201 pm->pmf->map_queues_size,
202 alloc_size_bytes);
203 }
204 }
205
206 pr_debug("Finished map process and queues to runlist\n");
207
208 if (is_over_subscription) {
209 if (!pm->is_over_subscription)
210 pr_warn("Runlist is getting oversubscribed. Expect reduced ROCm performance.\n");
211 retval = pm->pmf->runlist(pm, &rl_buffer[rl_wptr],
212 *rl_gpu_addr,
213 alloc_size_bytes / sizeof(uint32_t),
214 true);
215 }
216 pm->is_over_subscription = is_over_subscription;
217
218 for (i = 0; i < alloc_size_bytes / sizeof(uint32_t); i++)
219 pr_debug("0x%2X ", rl_buffer[i]);
220 pr_debug("\n");
221
222 return retval;
223}
224
225int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm)
226{
227 switch (dqm->dev->adev->asic_type) {
228 case CHIP_KAVERI:
229 case CHIP_HAWAII:
230 /* PM4 packet structures on CIK are the same as on VI */
231 case CHIP_CARRIZO:
232 case CHIP_TONGA:
233 case CHIP_FIJI:
234 case CHIP_POLARIS10:
235 case CHIP_POLARIS11:
236 case CHIP_POLARIS12:
237 case CHIP_VEGAM:
238 pm->pmf = &kfd_vi_pm_funcs;
239 break;
240 default:
241 if (KFD_GC_VERSION(dqm->dev) == IP_VERSION(9, 4, 2))
242 pm->pmf = &kfd_aldebaran_pm_funcs;
243 else if (KFD_GC_VERSION(dqm->dev) >= IP_VERSION(9, 0, 1))
244 pm->pmf = &kfd_v9_pm_funcs;
245 else {
246 WARN(1, "Unexpected ASIC family %u",
247 dqm->dev->adev->asic_type);
248 return -EINVAL;
249 }
250 }
251
252 pm->dqm = dqm;
253 mutex_init(&pm->lock);
254 pm->priv_queue = kernel_queue_init(dqm->dev, KFD_QUEUE_TYPE_HIQ);
255 if (!pm->priv_queue) {
256 mutex_destroy(&pm->lock);
257 return -ENOMEM;
258 }
259 pm->allocated = false;
260
261 return 0;
262}
263
264void pm_uninit(struct packet_manager *pm, bool hanging)
265{
266 mutex_destroy(&pm->lock);
267 kernel_queue_uninit(pm->priv_queue, hanging);
268 pm->priv_queue = NULL;
269}
270
271int pm_send_set_resources(struct packet_manager *pm,
272 struct scheduling_resources *res)
273{
274 uint32_t *buffer, size;
275 int retval = 0;
276
277 size = pm->pmf->set_resources_size;
278 mutex_lock(&pm->lock);
279 kq_acquire_packet_buffer(pm->priv_queue,
280 size / sizeof(uint32_t),
281 (unsigned int **)&buffer);
282 if (!buffer) {
283 pr_err("Failed to allocate buffer on kernel queue\n");
284 retval = -ENOMEM;
285 goto out;
286 }
287
288 retval = pm->pmf->set_resources(pm, buffer, res);
289 if (!retval)
290 kq_submit_packet(pm->priv_queue);
291 else
292 kq_rollback_packet(pm->priv_queue);
293
294out:
295 mutex_unlock(&pm->lock);
296
297 return retval;
298}
299
300int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues)
301{
302 uint64_t rl_gpu_ib_addr;
303 uint32_t *rl_buffer;
304 size_t rl_ib_size, packet_size_dwords;
305 int retval;
306
307 retval = pm_create_runlist_ib(pm, dqm_queues, &rl_gpu_ib_addr,
308 &rl_ib_size);
309 if (retval)
310 goto fail_create_runlist_ib;
311
312 pr_debug("runlist IB address: 0x%llX\n", rl_gpu_ib_addr);
313
314 packet_size_dwords = pm->pmf->runlist_size / sizeof(uint32_t);
315 mutex_lock(&pm->lock);
316
317 retval = kq_acquire_packet_buffer(pm->priv_queue,
318 packet_size_dwords, &rl_buffer);
319 if (retval)
320 goto fail_acquire_packet_buffer;
321
322 retval = pm->pmf->runlist(pm, rl_buffer, rl_gpu_ib_addr,
323 rl_ib_size / sizeof(uint32_t), false);
324 if (retval)
325 goto fail_create_runlist;
326
327 kq_submit_packet(pm->priv_queue);
328
329 mutex_unlock(&pm->lock);
330
331 return retval;
332
333fail_create_runlist:
334 kq_rollback_packet(pm->priv_queue);
335fail_acquire_packet_buffer:
336 mutex_unlock(&pm->lock);
337fail_create_runlist_ib:
338 pm_release_ib(pm);
339 return retval;
340}
341
342int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
343 uint64_t fence_value)
344{
345 uint32_t *buffer, size;
346 int retval = 0;
347
348 if (WARN_ON(!fence_address))
349 return -EFAULT;
350
351 size = pm->pmf->query_status_size;
352 mutex_lock(&pm->lock);
353 kq_acquire_packet_buffer(pm->priv_queue,
354 size / sizeof(uint32_t), (unsigned int **)&buffer);
355 if (!buffer) {
356 pr_err("Failed to allocate buffer on kernel queue\n");
357 retval = -ENOMEM;
358 goto out;
359 }
360
361 retval = pm->pmf->query_status(pm, buffer, fence_address, fence_value);
362 if (!retval)
363 kq_submit_packet(pm->priv_queue);
364 else
365 kq_rollback_packet(pm->priv_queue);
366
367out:
368 mutex_unlock(&pm->lock);
369 return retval;
370}
371
372int pm_send_unmap_queue(struct packet_manager *pm,
373 enum kfd_unmap_queues_filter filter,
374 uint32_t filter_param, bool reset)
375{
376 uint32_t *buffer, size;
377 int retval = 0;
378
379 size = pm->pmf->unmap_queues_size;
380 mutex_lock(&pm->lock);
381 kq_acquire_packet_buffer(pm->priv_queue,
382 size / sizeof(uint32_t), (unsigned int **)&buffer);
383 if (!buffer) {
384 pr_err("Failed to allocate buffer on kernel queue\n");
385 retval = -ENOMEM;
386 goto out;
387 }
388
389 retval = pm->pmf->unmap_queues(pm, buffer, filter, filter_param, reset);
390 if (!retval)
391 kq_submit_packet(pm->priv_queue);
392 else
393 kq_rollback_packet(pm->priv_queue);
394
395out:
396 mutex_unlock(&pm->lock);
397 return retval;
398}
399
400void pm_release_ib(struct packet_manager *pm)
401{
402 mutex_lock(&pm->lock);
403 if (pm->allocated) {
404 kfd_gtt_sa_free(pm->dqm->dev, pm->ib_buffer_obj);
405 pm->allocated = false;
406 }
407 mutex_unlock(&pm->lock);
408}
409
410#if defined(CONFIG_DEBUG_FS)
411
412int pm_debugfs_runlist(struct seq_file *m, void *data)
413{
414 struct packet_manager *pm = data;
415
416 mutex_lock(&pm->lock);
417
418 if (!pm->allocated) {
419 seq_puts(m, " No active runlist\n");
420 goto out;
421 }
422
423 seq_hex_dump(m, " ", DUMP_PREFIX_OFFSET, 32, 4,
424 pm->ib_buffer_obj->cpu_ptr, pm->ib_size_bytes, false);
425
426out:
427 mutex_unlock(&pm->lock);
428 return 0;
429}
430
431int pm_debugfs_hang_hws(struct packet_manager *pm)
432{
433 uint32_t *buffer, size;
434 int r = 0;
435
436 if (!pm->priv_queue)
437 return -EAGAIN;
438
439 size = pm->pmf->query_status_size;
440 mutex_lock(&pm->lock);
441 kq_acquire_packet_buffer(pm->priv_queue,
442 size / sizeof(uint32_t), (unsigned int **)&buffer);
443 if (!buffer) {
444 pr_err("Failed to allocate buffer on kernel queue\n");
445 r = -ENOMEM;
446 goto out;
447 }
448 memset(buffer, 0x55, size);
449 kq_submit_packet(pm->priv_queue);
450
451 pr_info("Submitting %x %x %x %x %x %x %x to HIQ to hang the HWS.",
452 buffer[0], buffer[1], buffer[2], buffer[3],
453 buffer[4], buffer[5], buffer[6]);
454out:
455 mutex_unlock(&pm->lock);
456 return r;
457}
458
459
460#endif