1
2
3
4
5
6
7
8#include <uapi/misc/habanalabs.h>
9#include "habanalabs.h"
10
11#include <linux/uaccess.h>
12#include <linux/slab.h>
13
14#define HL_CS_FLAGS_SIG_WAIT (HL_CS_FLAGS_SIGNAL | HL_CS_FLAGS_WAIT)
15
16static void job_wq_completion(struct work_struct *work);
17static long _hl_cs_wait_ioctl(struct hl_device *hdev,
18 struct hl_ctx *ctx, u64 timeout_us, u64 seq);
19static void cs_do_release(struct kref *ref);
20
21static void hl_sob_reset(struct kref *ref)
22{
23 struct hl_hw_sob *hw_sob = container_of(ref, struct hl_hw_sob,
24 kref);
25 struct hl_device *hdev = hw_sob->hdev;
26
27 hdev->asic_funcs->reset_sob(hdev, hw_sob);
28}
29
30void hl_sob_reset_error(struct kref *ref)
31{
32 struct hl_hw_sob *hw_sob = container_of(ref, struct hl_hw_sob,
33 kref);
34 struct hl_device *hdev = hw_sob->hdev;
35
36 dev_crit(hdev->dev,
37 "SOB release shouldn't be called here, q_idx: %d, sob_id: %d\n",
38 hw_sob->q_idx, hw_sob->sob_id);
39}
40
41static const char *hl_fence_get_driver_name(struct dma_fence *fence)
42{
43 return "HabanaLabs";
44}
45
46static const char *hl_fence_get_timeline_name(struct dma_fence *fence)
47{
48 struct hl_cs_compl *hl_cs_compl =
49 container_of(fence, struct hl_cs_compl, base_fence);
50
51 return dev_name(hl_cs_compl->hdev->dev);
52}
53
54static bool hl_fence_enable_signaling(struct dma_fence *fence)
55{
56 return true;
57}
58
59static void hl_fence_release(struct dma_fence *fence)
60{
61 struct hl_cs_compl *hl_cs_cmpl =
62 container_of(fence, struct hl_cs_compl, base_fence);
63 struct hl_device *hdev = hl_cs_cmpl->hdev;
64
65
66
67
68 if (fence->error == -EBUSY)
69 goto free;
70
71 if ((hl_cs_cmpl->type == CS_TYPE_SIGNAL) ||
72 (hl_cs_cmpl->type == CS_TYPE_WAIT)) {
73
74 dev_dbg(hdev->dev,
75 "CS 0x%llx type %d finished, sob_id: %d, sob_val: 0x%x\n",
76 hl_cs_cmpl->cs_seq,
77 hl_cs_cmpl->type,
78 hl_cs_cmpl->hw_sob->sob_id,
79 hl_cs_cmpl->sob_val);
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98 kref_put(&hl_cs_cmpl->hw_sob->kref, hl_sob_reset);
99 }
100
101free:
102 kfree_rcu(hl_cs_cmpl, base_fence.rcu);
103}
104
105static const struct dma_fence_ops hl_fence_ops = {
106 .get_driver_name = hl_fence_get_driver_name,
107 .get_timeline_name = hl_fence_get_timeline_name,
108 .enable_signaling = hl_fence_enable_signaling,
109 .release = hl_fence_release
110};
111
112static void cs_get(struct hl_cs *cs)
113{
114 kref_get(&cs->refcount);
115}
116
117static int cs_get_unless_zero(struct hl_cs *cs)
118{
119 return kref_get_unless_zero(&cs->refcount);
120}
121
122static void cs_put(struct hl_cs *cs)
123{
124 kref_put(&cs->refcount, cs_do_release);
125}
126
127static bool is_cb_patched(struct hl_device *hdev, struct hl_cs_job *job)
128{
129
130
131
132
133 return (job->queue_type == QUEUE_TYPE_EXT ||
134 (job->queue_type == QUEUE_TYPE_HW &&
135 job->is_kernel_allocated_cb &&
136 !hdev->mmu_enable));
137}
138
139
140
141
142
143
144
145
146
147
148
149
150static int cs_parser(struct hl_fpriv *hpriv, struct hl_cs_job *job)
151{
152 struct hl_device *hdev = hpriv->hdev;
153 struct hl_cs_parser parser;
154 int rc;
155
156 parser.ctx_id = job->cs->ctx->asid;
157 parser.cs_sequence = job->cs->sequence;
158 parser.job_id = job->id;
159
160 parser.hw_queue_id = job->hw_queue_id;
161 parser.job_userptr_list = &job->userptr_list;
162 parser.patched_cb = NULL;
163 parser.user_cb = job->user_cb;
164 parser.user_cb_size = job->user_cb_size;
165 parser.queue_type = job->queue_type;
166 parser.is_kernel_allocated_cb = job->is_kernel_allocated_cb;
167 job->patched_cb = NULL;
168
169 rc = hdev->asic_funcs->cs_parser(hdev, &parser);
170
171 if (is_cb_patched(hdev, job)) {
172 if (!rc) {
173 job->patched_cb = parser.patched_cb;
174 job->job_cb_size = parser.patched_cb_size;
175 job->contains_dma_pkt = parser.contains_dma_pkt;
176
177 spin_lock(&job->patched_cb->lock);
178 job->patched_cb->cs_cnt++;
179 spin_unlock(&job->patched_cb->lock);
180 }
181
182
183
184
185
186
187 spin_lock(&job->user_cb->lock);
188 job->user_cb->cs_cnt--;
189 spin_unlock(&job->user_cb->lock);
190 hl_cb_put(job->user_cb);
191 job->user_cb = NULL;
192 } else if (!rc) {
193 job->job_cb_size = job->user_cb_size;
194 }
195
196 return rc;
197}
198
199static void free_job(struct hl_device *hdev, struct hl_cs_job *job)
200{
201 struct hl_cs *cs = job->cs;
202
203 if (is_cb_patched(hdev, job)) {
204 hl_userptr_delete_list(hdev, &job->userptr_list);
205
206
207
208
209
210 if (job->patched_cb) {
211 spin_lock(&job->patched_cb->lock);
212 job->patched_cb->cs_cnt--;
213 spin_unlock(&job->patched_cb->lock);
214
215 hl_cb_put(job->patched_cb);
216 }
217 }
218
219
220
221
222
223 if (job->queue_type == QUEUE_TYPE_HW &&
224 job->is_kernel_allocated_cb && hdev->mmu_enable) {
225 spin_lock(&job->user_cb->lock);
226 job->user_cb->cs_cnt--;
227 spin_unlock(&job->user_cb->lock);
228
229 hl_cb_put(job->user_cb);
230 }
231
232
233
234
235
236 spin_lock(&cs->job_lock);
237 list_del(&job->cs_node);
238 spin_unlock(&cs->job_lock);
239
240 hl_debugfs_remove_job(hdev, job);
241
242 if (job->queue_type == QUEUE_TYPE_EXT ||
243 job->queue_type == QUEUE_TYPE_HW)
244 cs_put(cs);
245
246 kfree(job);
247}
248
249static void cs_counters_aggregate(struct hl_device *hdev, struct hl_ctx *ctx)
250{
251 hdev->aggregated_cs_counters.device_in_reset_drop_cnt +=
252 ctx->cs_counters.device_in_reset_drop_cnt;
253 hdev->aggregated_cs_counters.out_of_mem_drop_cnt +=
254 ctx->cs_counters.out_of_mem_drop_cnt;
255 hdev->aggregated_cs_counters.parsing_drop_cnt +=
256 ctx->cs_counters.parsing_drop_cnt;
257 hdev->aggregated_cs_counters.queue_full_drop_cnt +=
258 ctx->cs_counters.queue_full_drop_cnt;
259}
260
261static void cs_do_release(struct kref *ref)
262{
263 struct hl_cs *cs = container_of(ref, struct hl_cs,
264 refcount);
265 struct hl_device *hdev = cs->ctx->hdev;
266 struct hl_cs_job *job, *tmp;
267
268 cs->completed = true;
269
270
271
272
273
274
275
276
277
278 list_for_each_entry_safe(job, tmp, &cs->job_list, cs_node)
279 free_job(hdev, job);
280
281
282 if (cs->submitted) {
283 hdev->asic_funcs->hw_queues_lock(hdev);
284
285 hdev->cs_active_cnt--;
286 if (!hdev->cs_active_cnt) {
287 struct hl_device_idle_busy_ts *ts;
288
289 ts = &hdev->idle_busy_ts_arr[hdev->idle_busy_ts_idx++];
290 ts->busy_to_idle_ts = ktime_get();
291
292 if (hdev->idle_busy_ts_idx == HL_IDLE_BUSY_TS_ARR_SIZE)
293 hdev->idle_busy_ts_idx = 0;
294 } else if (hdev->cs_active_cnt < 0) {
295 dev_crit(hdev->dev, "CS active cnt %d is negative\n",
296 hdev->cs_active_cnt);
297 }
298
299 hdev->asic_funcs->hw_queues_unlock(hdev);
300
301 hl_int_hw_queue_update_ci(cs);
302
303 spin_lock(&hdev->hw_queues_mirror_lock);
304
305 list_del_init(&cs->mirror_node);
306 spin_unlock(&hdev->hw_queues_mirror_lock);
307
308
309
310
311
312 if ((!cs->timedout) &&
313 (hdev->timeout_jiffies != MAX_SCHEDULE_TIMEOUT)) {
314 struct hl_cs *next;
315
316 if (cs->tdr_active)
317 cancel_delayed_work_sync(&cs->work_tdr);
318
319 spin_lock(&hdev->hw_queues_mirror_lock);
320
321
322 next = list_first_entry_or_null(
323 &hdev->hw_queues_mirror_list,
324 struct hl_cs, mirror_node);
325
326 if ((next) && (!next->tdr_active)) {
327 next->tdr_active = true;
328 schedule_delayed_work(&next->work_tdr,
329 hdev->timeout_jiffies);
330 }
331
332 spin_unlock(&hdev->hw_queues_mirror_lock);
333 }
334 } else if (cs->type == CS_TYPE_WAIT) {
335
336
337
338
339 dma_fence_put(cs->signal_fence);
340 }
341
342
343
344
345
346 hl_debugfs_remove_cs(cs);
347
348 hl_ctx_put(cs->ctx);
349
350
351
352
353
354 if (cs->timedout)
355 dma_fence_set_error(cs->fence, -ETIMEDOUT);
356 else if (cs->aborted)
357 dma_fence_set_error(cs->fence, -EIO);
358 else if (!cs->submitted)
359 dma_fence_set_error(cs->fence, -EBUSY);
360
361 dma_fence_signal(cs->fence);
362 dma_fence_put(cs->fence);
363
364 cs_counters_aggregate(hdev, cs->ctx);
365
366 kfree(cs->jobs_in_queue_cnt);
367 kfree(cs);
368}
369
370static void cs_timedout(struct work_struct *work)
371{
372 struct hl_device *hdev;
373 int rc;
374 struct hl_cs *cs = container_of(work, struct hl_cs,
375 work_tdr.work);
376 rc = cs_get_unless_zero(cs);
377 if (!rc)
378 return;
379
380 if ((!cs->submitted) || (cs->completed)) {
381 cs_put(cs);
382 return;
383 }
384
385
386 cs->timedout = true;
387
388 hdev = cs->ctx->hdev;
389
390 dev_err(hdev->dev,
391 "Command submission %llu has not finished in time!\n",
392 cs->sequence);
393
394 cs_put(cs);
395
396 if (hdev->reset_on_lockup)
397 hl_device_reset(hdev, false, false);
398}
399
400static int allocate_cs(struct hl_device *hdev, struct hl_ctx *ctx,
401 enum hl_cs_type cs_type, struct hl_cs **cs_new)
402{
403 struct hl_cs_compl *cs_cmpl;
404 struct dma_fence *other = NULL;
405 struct hl_cs *cs;
406 int rc;
407
408 cs = kzalloc(sizeof(*cs), GFP_ATOMIC);
409 if (!cs)
410 return -ENOMEM;
411
412 cs->ctx = ctx;
413 cs->submitted = false;
414 cs->completed = false;
415 cs->type = cs_type;
416 INIT_LIST_HEAD(&cs->job_list);
417 INIT_DELAYED_WORK(&cs->work_tdr, cs_timedout);
418 kref_init(&cs->refcount);
419 spin_lock_init(&cs->job_lock);
420
421 cs_cmpl = kmalloc(sizeof(*cs_cmpl), GFP_ATOMIC);
422 if (!cs_cmpl) {
423 rc = -ENOMEM;
424 goto free_cs;
425 }
426
427 cs_cmpl->hdev = hdev;
428 cs_cmpl->type = cs->type;
429 spin_lock_init(&cs_cmpl->lock);
430 cs->fence = &cs_cmpl->base_fence;
431
432 spin_lock(&ctx->cs_lock);
433
434 cs_cmpl->cs_seq = ctx->cs_sequence;
435 other = ctx->cs_pending[cs_cmpl->cs_seq &
436 (hdev->asic_prop.max_pending_cs - 1)];
437 if ((other) && (!dma_fence_is_signaled(other))) {
438 dev_dbg(hdev->dev,
439 "Rejecting CS because of too many in-flights CS\n");
440 rc = -EAGAIN;
441 goto free_fence;
442 }
443
444 cs->jobs_in_queue_cnt = kcalloc(hdev->asic_prop.max_queues,
445 sizeof(*cs->jobs_in_queue_cnt), GFP_ATOMIC);
446 if (!cs->jobs_in_queue_cnt) {
447 rc = -ENOMEM;
448 goto free_fence;
449 }
450
451 dma_fence_init(&cs_cmpl->base_fence, &hl_fence_ops, &cs_cmpl->lock,
452 ctx->asid, ctx->cs_sequence);
453
454 cs->sequence = cs_cmpl->cs_seq;
455
456 ctx->cs_pending[cs_cmpl->cs_seq &
457 (hdev->asic_prop.max_pending_cs - 1)] =
458 &cs_cmpl->base_fence;
459 ctx->cs_sequence++;
460
461 dma_fence_get(&cs_cmpl->base_fence);
462
463 dma_fence_put(other);
464
465 spin_unlock(&ctx->cs_lock);
466
467 *cs_new = cs;
468
469 return 0;
470
471free_fence:
472 spin_unlock(&ctx->cs_lock);
473 kfree(cs_cmpl);
474free_cs:
475 kfree(cs);
476 return rc;
477}
478
479static void cs_rollback(struct hl_device *hdev, struct hl_cs *cs)
480{
481 struct hl_cs_job *job, *tmp;
482
483 list_for_each_entry_safe(job, tmp, &cs->job_list, cs_node)
484 free_job(hdev, job);
485}
486
487void hl_cs_rollback_all(struct hl_device *hdev)
488{
489 int i;
490 struct hl_cs *cs, *tmp;
491
492
493 for (i = 0 ; i < hdev->asic_prop.completion_queues_count ; i++)
494 flush_workqueue(hdev->cq_wq[i]);
495
496
497 list_for_each_entry_safe(cs, tmp, &hdev->hw_queues_mirror_list,
498 mirror_node) {
499 cs_get(cs);
500 cs->aborted = true;
501 dev_warn_ratelimited(hdev->dev, "Killing CS %d.%llu\n",
502 cs->ctx->asid, cs->sequence);
503 cs_rollback(hdev, cs);
504 cs_put(cs);
505 }
506}
507
508static void job_wq_completion(struct work_struct *work)
509{
510 struct hl_cs_job *job = container_of(work, struct hl_cs_job,
511 finish_work);
512 struct hl_cs *cs = job->cs;
513 struct hl_device *hdev = cs->ctx->hdev;
514
515
516 free_job(hdev, job);
517}
518
519static int validate_queue_index(struct hl_device *hdev,
520 struct hl_cs_chunk *chunk,
521 enum hl_queue_type *queue_type,
522 bool *is_kernel_allocated_cb)
523{
524 struct asic_fixed_properties *asic = &hdev->asic_prop;
525 struct hw_queue_properties *hw_queue_prop;
526
527
528
529
530 if (chunk->queue_index >= asic->max_queues) {
531 dev_err(hdev->dev, "Queue index %d is invalid\n",
532 chunk->queue_index);
533 return -EINVAL;
534 }
535
536 hw_queue_prop = &asic->hw_queues_props[chunk->queue_index];
537
538 if (hw_queue_prop->type == QUEUE_TYPE_NA) {
539 dev_err(hdev->dev, "Queue index %d is invalid\n",
540 chunk->queue_index);
541 return -EINVAL;
542 }
543
544 if (hw_queue_prop->driver_only) {
545 dev_err(hdev->dev,
546 "Queue index %d is restricted for the kernel driver\n",
547 chunk->queue_index);
548 return -EINVAL;
549 }
550
551 *queue_type = hw_queue_prop->type;
552 *is_kernel_allocated_cb = !!hw_queue_prop->requires_kernel_cb;
553
554 return 0;
555}
556
557static struct hl_cb *get_cb_from_cs_chunk(struct hl_device *hdev,
558 struct hl_cb_mgr *cb_mgr,
559 struct hl_cs_chunk *chunk)
560{
561 struct hl_cb *cb;
562 u32 cb_handle;
563
564 cb_handle = (u32) (chunk->cb_handle >> PAGE_SHIFT);
565
566 cb = hl_cb_get(hdev, cb_mgr, cb_handle);
567 if (!cb) {
568 dev_err(hdev->dev, "CB handle 0x%x invalid\n", cb_handle);
569 return NULL;
570 }
571
572 if ((chunk->cb_size < 8) || (chunk->cb_size > cb->size)) {
573 dev_err(hdev->dev, "CB size %u invalid\n", chunk->cb_size);
574 goto release_cb;
575 }
576
577 spin_lock(&cb->lock);
578 cb->cs_cnt++;
579 spin_unlock(&cb->lock);
580
581 return cb;
582
583release_cb:
584 hl_cb_put(cb);
585 return NULL;
586}
587
588struct hl_cs_job *hl_cs_allocate_job(struct hl_device *hdev,
589 enum hl_queue_type queue_type, bool is_kernel_allocated_cb)
590{
591 struct hl_cs_job *job;
592
593 job = kzalloc(sizeof(*job), GFP_ATOMIC);
594 if (!job)
595 return NULL;
596
597 job->queue_type = queue_type;
598 job->is_kernel_allocated_cb = is_kernel_allocated_cb;
599
600 if (is_cb_patched(hdev, job))
601 INIT_LIST_HEAD(&job->userptr_list);
602
603 if (job->queue_type == QUEUE_TYPE_EXT)
604 INIT_WORK(&job->finish_work, job_wq_completion);
605
606 return job;
607}
608
609static int cs_ioctl_default(struct hl_fpriv *hpriv, void __user *chunks,
610 u32 num_chunks, u64 *cs_seq)
611{
612 struct hl_device *hdev = hpriv->hdev;
613 struct hl_cs_chunk *cs_chunk_array;
614 struct hl_cs_job *job;
615 struct hl_cs *cs;
616 struct hl_cb *cb;
617 bool int_queues_only = true;
618 u32 size_to_copy;
619 int rc, i;
620
621 *cs_seq = ULLONG_MAX;
622
623 if (num_chunks > HL_MAX_JOBS_PER_CS) {
624 dev_err(hdev->dev,
625 "Number of chunks can NOT be larger than %d\n",
626 HL_MAX_JOBS_PER_CS);
627 rc = -EINVAL;
628 goto out;
629 }
630
631 cs_chunk_array = kmalloc_array(num_chunks, sizeof(*cs_chunk_array),
632 GFP_ATOMIC);
633 if (!cs_chunk_array) {
634 rc = -ENOMEM;
635 goto out;
636 }
637
638 size_to_copy = num_chunks * sizeof(struct hl_cs_chunk);
639 if (copy_from_user(cs_chunk_array, chunks, size_to_copy)) {
640 dev_err(hdev->dev, "Failed to copy cs chunk array from user\n");
641 rc = -EFAULT;
642 goto free_cs_chunk_array;
643 }
644
645
646 hl_ctx_get(hdev, hpriv->ctx);
647
648 rc = allocate_cs(hdev, hpriv->ctx, CS_TYPE_DEFAULT, &cs);
649 if (rc) {
650 hl_ctx_put(hpriv->ctx);
651 goto free_cs_chunk_array;
652 }
653
654 *cs_seq = cs->sequence;
655
656 hl_debugfs_add_cs(cs);
657
658
659 for (i = 0 ; i < num_chunks ; i++) {
660 struct hl_cs_chunk *chunk = &cs_chunk_array[i];
661 enum hl_queue_type queue_type;
662 bool is_kernel_allocated_cb;
663
664 rc = validate_queue_index(hdev, chunk, &queue_type,
665 &is_kernel_allocated_cb);
666 if (rc) {
667 hpriv->ctx->cs_counters.parsing_drop_cnt++;
668 goto free_cs_object;
669 }
670
671 if (is_kernel_allocated_cb) {
672 cb = get_cb_from_cs_chunk(hdev, &hpriv->cb_mgr, chunk);
673 if (!cb) {
674 hpriv->ctx->cs_counters.parsing_drop_cnt++;
675 rc = -EINVAL;
676 goto free_cs_object;
677 }
678 } else {
679 cb = (struct hl_cb *) (uintptr_t) chunk->cb_handle;
680 }
681
682 if (queue_type == QUEUE_TYPE_EXT || queue_type == QUEUE_TYPE_HW)
683 int_queues_only = false;
684
685 job = hl_cs_allocate_job(hdev, queue_type,
686 is_kernel_allocated_cb);
687 if (!job) {
688 hpriv->ctx->cs_counters.out_of_mem_drop_cnt++;
689 dev_err(hdev->dev, "Failed to allocate a new job\n");
690 rc = -ENOMEM;
691 if (is_kernel_allocated_cb)
692 goto release_cb;
693 else
694 goto free_cs_object;
695 }
696
697 job->id = i + 1;
698 job->cs = cs;
699 job->user_cb = cb;
700 job->user_cb_size = chunk->cb_size;
701 job->hw_queue_id = chunk->queue_index;
702
703 cs->jobs_in_queue_cnt[job->hw_queue_id]++;
704
705 list_add_tail(&job->cs_node, &cs->job_list);
706
707
708
709
710
711
712
713 if (job->queue_type == QUEUE_TYPE_EXT ||
714 job->queue_type == QUEUE_TYPE_HW)
715 cs_get(cs);
716
717 hl_debugfs_add_job(hdev, job);
718
719 rc = cs_parser(hpriv, job);
720 if (rc) {
721 hpriv->ctx->cs_counters.parsing_drop_cnt++;
722 dev_err(hdev->dev,
723 "Failed to parse JOB %d.%llu.%d, err %d, rejecting the CS\n",
724 cs->ctx->asid, cs->sequence, job->id, rc);
725 goto free_cs_object;
726 }
727 }
728
729 if (int_queues_only) {
730 hpriv->ctx->cs_counters.parsing_drop_cnt++;
731 dev_err(hdev->dev,
732 "Reject CS %d.%llu because only internal queues jobs are present\n",
733 cs->ctx->asid, cs->sequence);
734 rc = -EINVAL;
735 goto free_cs_object;
736 }
737
738 rc = hl_hw_queue_schedule_cs(cs);
739 if (rc) {
740 if (rc != -EAGAIN)
741 dev_err(hdev->dev,
742 "Failed to submit CS %d.%llu to H/W queues, error %d\n",
743 cs->ctx->asid, cs->sequence, rc);
744 goto free_cs_object;
745 }
746
747 rc = HL_CS_STATUS_SUCCESS;
748 goto put_cs;
749
750release_cb:
751 spin_lock(&cb->lock);
752 cb->cs_cnt--;
753 spin_unlock(&cb->lock);
754 hl_cb_put(cb);
755free_cs_object:
756 cs_rollback(hdev, cs);
757 *cs_seq = ULLONG_MAX;
758
759put_cs:
760
761 cs_put(cs);
762free_cs_chunk_array:
763 kfree(cs_chunk_array);
764out:
765 return rc;
766}
767
768static int cs_ioctl_signal_wait(struct hl_fpriv *hpriv, enum hl_cs_type cs_type,
769 void __user *chunks, u32 num_chunks,
770 u64 *cs_seq)
771{
772 struct hl_device *hdev = hpriv->hdev;
773 struct hl_ctx *ctx = hpriv->ctx;
774 struct hl_cs_chunk *cs_chunk_array, *chunk;
775 struct hw_queue_properties *hw_queue_prop;
776 struct dma_fence *sig_fence = NULL;
777 struct hl_cs_job *job;
778 struct hl_cs *cs;
779 struct hl_cb *cb;
780 enum hl_queue_type q_type;
781 u64 *signal_seq_arr = NULL, signal_seq;
782 u32 size_to_copy, q_idx, signal_seq_arr_len, cb_size;
783 int rc;
784
785 *cs_seq = ULLONG_MAX;
786
787 if (num_chunks > HL_MAX_JOBS_PER_CS) {
788 dev_err(hdev->dev,
789 "Number of chunks can NOT be larger than %d\n",
790 HL_MAX_JOBS_PER_CS);
791 rc = -EINVAL;
792 goto out;
793 }
794
795 cs_chunk_array = kmalloc_array(num_chunks, sizeof(*cs_chunk_array),
796 GFP_ATOMIC);
797 if (!cs_chunk_array) {
798 rc = -ENOMEM;
799 goto out;
800 }
801
802 size_to_copy = num_chunks * sizeof(struct hl_cs_chunk);
803 if (copy_from_user(cs_chunk_array, chunks, size_to_copy)) {
804 dev_err(hdev->dev, "Failed to copy cs chunk array from user\n");
805 rc = -EFAULT;
806 goto free_cs_chunk_array;
807 }
808
809
810 chunk = &cs_chunk_array[0];
811
812 if (chunk->queue_index >= hdev->asic_prop.max_queues) {
813 dev_err(hdev->dev, "Queue index %d is invalid\n",
814 chunk->queue_index);
815 rc = -EINVAL;
816 goto free_cs_chunk_array;
817 }
818
819 q_idx = chunk->queue_index;
820 hw_queue_prop = &hdev->asic_prop.hw_queues_props[q_idx];
821 q_type = hw_queue_prop->type;
822
823 if ((q_idx >= hdev->asic_prop.max_queues) ||
824 (!hw_queue_prop->supports_sync_stream)) {
825 dev_err(hdev->dev, "Queue index %d is invalid\n", q_idx);
826 rc = -EINVAL;
827 goto free_cs_chunk_array;
828 }
829
830 if (cs_type == CS_TYPE_WAIT) {
831 struct hl_cs_compl *sig_waitcs_cmpl;
832
833 signal_seq_arr_len = chunk->num_signal_seq_arr;
834
835
836 if (signal_seq_arr_len != 1) {
837 dev_err(hdev->dev,
838 "Wait for signal CS supports only one signal CS seq\n");
839 rc = -EINVAL;
840 goto free_cs_chunk_array;
841 }
842
843 signal_seq_arr = kmalloc_array(signal_seq_arr_len,
844 sizeof(*signal_seq_arr),
845 GFP_ATOMIC);
846 if (!signal_seq_arr) {
847 rc = -ENOMEM;
848 goto free_cs_chunk_array;
849 }
850
851 size_to_copy = chunk->num_signal_seq_arr *
852 sizeof(*signal_seq_arr);
853 if (copy_from_user(signal_seq_arr,
854 u64_to_user_ptr(chunk->signal_seq_arr),
855 size_to_copy)) {
856 dev_err(hdev->dev,
857 "Failed to copy signal seq array from user\n");
858 rc = -EFAULT;
859 goto free_signal_seq_array;
860 }
861
862
863 signal_seq = signal_seq_arr[0];
864 sig_fence = hl_ctx_get_fence(ctx, signal_seq);
865 if (IS_ERR(sig_fence)) {
866 dev_err(hdev->dev,
867 "Failed to get signal CS with seq 0x%llx\n",
868 signal_seq);
869 rc = PTR_ERR(sig_fence);
870 goto free_signal_seq_array;
871 }
872
873 if (!sig_fence) {
874
875 rc = 0;
876 goto free_signal_seq_array;
877 }
878
879 sig_waitcs_cmpl =
880 container_of(sig_fence, struct hl_cs_compl, base_fence);
881
882 if (sig_waitcs_cmpl->type != CS_TYPE_SIGNAL) {
883 dev_err(hdev->dev,
884 "CS seq 0x%llx is not of a signal CS\n",
885 signal_seq);
886 dma_fence_put(sig_fence);
887 rc = -EINVAL;
888 goto free_signal_seq_array;
889 }
890
891 if (dma_fence_is_signaled(sig_fence)) {
892
893 dma_fence_put(sig_fence);
894 rc = 0;
895 goto free_signal_seq_array;
896 }
897 }
898
899
900 hl_ctx_get(hdev, ctx);
901
902 rc = allocate_cs(hdev, ctx, cs_type, &cs);
903 if (rc) {
904 if (cs_type == CS_TYPE_WAIT)
905 dma_fence_put(sig_fence);
906 hl_ctx_put(ctx);
907 goto free_signal_seq_array;
908 }
909
910
911
912
913
914 if (cs->type == CS_TYPE_WAIT)
915 cs->signal_fence = sig_fence;
916
917 hl_debugfs_add_cs(cs);
918
919 *cs_seq = cs->sequence;
920
921 job = hl_cs_allocate_job(hdev, q_type, true);
922 if (!job) {
923 ctx->cs_counters.out_of_mem_drop_cnt++;
924 dev_err(hdev->dev, "Failed to allocate a new job\n");
925 rc = -ENOMEM;
926 goto put_cs;
927 }
928
929 if (cs->type == CS_TYPE_WAIT)
930 cb_size = hdev->asic_funcs->get_wait_cb_size(hdev);
931 else
932 cb_size = hdev->asic_funcs->get_signal_cb_size(hdev);
933
934 cb = hl_cb_kernel_create(hdev, cb_size,
935 q_type == QUEUE_TYPE_HW && hdev->mmu_enable);
936 if (!cb) {
937 ctx->cs_counters.out_of_mem_drop_cnt++;
938 kfree(job);
939 rc = -EFAULT;
940 goto put_cs;
941 }
942
943 job->id = 0;
944 job->cs = cs;
945 job->user_cb = cb;
946 job->user_cb->cs_cnt++;
947 job->user_cb_size = cb_size;
948 job->hw_queue_id = q_idx;
949
950
951
952
953
954
955
956 job->patched_cb = job->user_cb;
957 job->job_cb_size = job->user_cb_size;
958 hl_cb_destroy(hdev, &hdev->kernel_cb_mgr, cb->id << PAGE_SHIFT);
959
960 cs->jobs_in_queue_cnt[job->hw_queue_id]++;
961
962 list_add_tail(&job->cs_node, &cs->job_list);
963
964
965 cs_get(cs);
966
967 hl_debugfs_add_job(hdev, job);
968
969 rc = hl_hw_queue_schedule_cs(cs);
970 if (rc) {
971 if (rc != -EAGAIN)
972 dev_err(hdev->dev,
973 "Failed to submit CS %d.%llu to H/W queues, error %d\n",
974 ctx->asid, cs->sequence, rc);
975 goto free_cs_object;
976 }
977
978 rc = HL_CS_STATUS_SUCCESS;
979 goto put_cs;
980
981free_cs_object:
982 cs_rollback(hdev, cs);
983 *cs_seq = ULLONG_MAX;
984
985put_cs:
986
987 cs_put(cs);
988free_signal_seq_array:
989 if (cs_type == CS_TYPE_WAIT)
990 kfree(signal_seq_arr);
991free_cs_chunk_array:
992 kfree(cs_chunk_array);
993out:
994 return rc;
995}
996
997int hl_cs_ioctl(struct hl_fpriv *hpriv, void *data)
998{
999 struct hl_device *hdev = hpriv->hdev;
1000 union hl_cs_args *args = data;
1001 struct hl_ctx *ctx = hpriv->ctx;
1002 void __user *chunks_execute, *chunks_restore;
1003 enum hl_cs_type cs_type;
1004 u32 num_chunks_execute, num_chunks_restore, sig_wait_flags;
1005 u64 cs_seq = ULONG_MAX;
1006 int rc, do_ctx_switch;
1007 bool need_soft_reset = false;
1008
1009 if (hl_device_disabled_or_in_reset(hdev)) {
1010 dev_warn_ratelimited(hdev->dev,
1011 "Device is %s. Can't submit new CS\n",
1012 atomic_read(&hdev->in_reset) ? "in_reset" : "disabled");
1013 rc = -EBUSY;
1014 goto out;
1015 }
1016
1017 sig_wait_flags = args->in.cs_flags & HL_CS_FLAGS_SIG_WAIT;
1018
1019 if (unlikely(sig_wait_flags == HL_CS_FLAGS_SIG_WAIT)) {
1020 dev_err(hdev->dev,
1021 "Signal and wait CS flags are mutually exclusive, context %d\n",
1022 ctx->asid);
1023 rc = -EINVAL;
1024 goto out;
1025 }
1026
1027 if (unlikely((sig_wait_flags & HL_CS_FLAGS_SIG_WAIT) &&
1028 (!hdev->supports_sync_stream))) {
1029 dev_err(hdev->dev, "Sync stream CS is not supported\n");
1030 rc = -EINVAL;
1031 goto out;
1032 }
1033
1034 if (args->in.cs_flags & HL_CS_FLAGS_SIGNAL)
1035 cs_type = CS_TYPE_SIGNAL;
1036 else if (args->in.cs_flags & HL_CS_FLAGS_WAIT)
1037 cs_type = CS_TYPE_WAIT;
1038 else
1039 cs_type = CS_TYPE_DEFAULT;
1040
1041 chunks_execute = (void __user *) (uintptr_t) args->in.chunks_execute;
1042 num_chunks_execute = args->in.num_chunks_execute;
1043
1044 if (cs_type == CS_TYPE_DEFAULT) {
1045 if (!num_chunks_execute) {
1046 dev_err(hdev->dev,
1047 "Got execute CS with 0 chunks, context %d\n",
1048 ctx->asid);
1049 rc = -EINVAL;
1050 goto out;
1051 }
1052 } else if (num_chunks_execute != 1) {
1053 dev_err(hdev->dev,
1054 "Sync stream CS mandates one chunk only, context %d\n",
1055 ctx->asid);
1056 rc = -EINVAL;
1057 goto out;
1058 }
1059
1060 do_ctx_switch = atomic_cmpxchg(&ctx->thread_ctx_switch_token, 1, 0);
1061
1062 if (do_ctx_switch || (args->in.cs_flags & HL_CS_FLAGS_FORCE_RESTORE)) {
1063 long ret;
1064
1065 chunks_restore =
1066 (void __user *) (uintptr_t) args->in.chunks_restore;
1067 num_chunks_restore = args->in.num_chunks_restore;
1068
1069 mutex_lock(&hpriv->restore_phase_mutex);
1070
1071 if (do_ctx_switch) {
1072 rc = hdev->asic_funcs->context_switch(hdev, ctx->asid);
1073 if (rc) {
1074 dev_err_ratelimited(hdev->dev,
1075 "Failed to switch to context %d, rejecting CS! %d\n",
1076 ctx->asid, rc);
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086 if ((rc == -ETIMEDOUT) || (rc == -EBUSY))
1087 need_soft_reset = true;
1088 mutex_unlock(&hpriv->restore_phase_mutex);
1089 goto out;
1090 }
1091 }
1092
1093 hdev->asic_funcs->restore_phase_topology(hdev);
1094
1095 if (!num_chunks_restore) {
1096 dev_dbg(hdev->dev,
1097 "Need to run restore phase but restore CS is empty\n");
1098 rc = 0;
1099 } else {
1100 rc = cs_ioctl_default(hpriv, chunks_restore,
1101 num_chunks_restore, &cs_seq);
1102 }
1103
1104 mutex_unlock(&hpriv->restore_phase_mutex);
1105
1106 if (rc) {
1107 dev_err(hdev->dev,
1108 "Failed to submit restore CS for context %d (%d)\n",
1109 ctx->asid, rc);
1110 goto out;
1111 }
1112
1113
1114 if (num_chunks_restore) {
1115 ret = _hl_cs_wait_ioctl(hdev, ctx,
1116 jiffies_to_usecs(hdev->timeout_jiffies),
1117 cs_seq);
1118 if (ret <= 0) {
1119 dev_err(hdev->dev,
1120 "Restore CS for context %d failed to complete %ld\n",
1121 ctx->asid, ret);
1122 rc = -ENOEXEC;
1123 goto out;
1124 }
1125 }
1126
1127 ctx->thread_ctx_switch_wait_token = 1;
1128 } else if (!ctx->thread_ctx_switch_wait_token) {
1129 u32 tmp;
1130
1131 rc = hl_poll_timeout_memory(hdev,
1132 &ctx->thread_ctx_switch_wait_token, tmp, (tmp == 1),
1133 100, jiffies_to_usecs(hdev->timeout_jiffies), false);
1134
1135 if (rc == -ETIMEDOUT) {
1136 dev_err(hdev->dev,
1137 "context switch phase timeout (%d)\n", tmp);
1138 goto out;
1139 }
1140 }
1141
1142 if (cs_type == CS_TYPE_DEFAULT)
1143 rc = cs_ioctl_default(hpriv, chunks_execute, num_chunks_execute,
1144 &cs_seq);
1145 else
1146 rc = cs_ioctl_signal_wait(hpriv, cs_type, chunks_execute,
1147 num_chunks_execute, &cs_seq);
1148
1149out:
1150 if (rc != -EAGAIN) {
1151 memset(args, 0, sizeof(*args));
1152 args->out.status = rc;
1153 args->out.seq = cs_seq;
1154 }
1155
1156 if (((rc == -ETIMEDOUT) || (rc == -EBUSY)) && (need_soft_reset))
1157 hl_device_reset(hdev, false, false);
1158
1159 return rc;
1160}
1161
1162static long _hl_cs_wait_ioctl(struct hl_device *hdev,
1163 struct hl_ctx *ctx, u64 timeout_us, u64 seq)
1164{
1165 struct dma_fence *fence;
1166 unsigned long timeout;
1167 long rc;
1168
1169 if (timeout_us == MAX_SCHEDULE_TIMEOUT)
1170 timeout = timeout_us;
1171 else
1172 timeout = usecs_to_jiffies(timeout_us);
1173
1174 hl_ctx_get(hdev, ctx);
1175
1176 fence = hl_ctx_get_fence(ctx, seq);
1177 if (IS_ERR(fence)) {
1178 rc = PTR_ERR(fence);
1179 if (rc == -EINVAL)
1180 dev_notice_ratelimited(hdev->dev,
1181 "Can't wait on CS %llu because current CS is at seq %llu\n",
1182 seq, ctx->cs_sequence);
1183 } else if (fence) {
1184 rc = dma_fence_wait_timeout(fence, true, timeout);
1185 if (fence->error == -ETIMEDOUT)
1186 rc = -ETIMEDOUT;
1187 else if (fence->error == -EIO)
1188 rc = -EIO;
1189 dma_fence_put(fence);
1190 } else {
1191 dev_dbg(hdev->dev,
1192 "Can't wait on seq %llu because current CS is at seq %llu (Fence is gone)\n",
1193 seq, ctx->cs_sequence);
1194 rc = 1;
1195 }
1196
1197 hl_ctx_put(ctx);
1198
1199 return rc;
1200}
1201
1202int hl_cs_wait_ioctl(struct hl_fpriv *hpriv, void *data)
1203{
1204 struct hl_device *hdev = hpriv->hdev;
1205 union hl_wait_cs_args *args = data;
1206 u64 seq = args->in.seq;
1207 long rc;
1208
1209 rc = _hl_cs_wait_ioctl(hdev, hpriv->ctx, args->in.timeout_us, seq);
1210
1211 memset(args, 0, sizeof(*args));
1212
1213 if (rc < 0) {
1214 if (rc == -ERESTARTSYS) {
1215 dev_err_ratelimited(hdev->dev,
1216 "user process got signal while waiting for CS handle %llu\n",
1217 seq);
1218 args->out.status = HL_WAIT_CS_STATUS_INTERRUPTED;
1219 rc = -EINTR;
1220 } else if (rc == -ETIMEDOUT) {
1221 dev_err_ratelimited(hdev->dev,
1222 "CS %llu has timed-out while user process is waiting for it\n",
1223 seq);
1224 args->out.status = HL_WAIT_CS_STATUS_TIMEDOUT;
1225 } else if (rc == -EIO) {
1226 dev_err_ratelimited(hdev->dev,
1227 "CS %llu has been aborted while user process is waiting for it\n",
1228 seq);
1229 args->out.status = HL_WAIT_CS_STATUS_ABORTED;
1230 }
1231 return rc;
1232 }
1233
1234 if (rc == 0)
1235 args->out.status = HL_WAIT_CS_STATUS_BUSY;
1236 else
1237 args->out.status = HL_WAIT_CS_STATUS_COMPLETED;
1238
1239 return 0;
1240}
1241