Merge tag 'for-6.6-rc3-tag' of git://git.kernel.org/pub/scm/linux/kernel/git/kdave...
[platform/kernel/linux-rpi.git] / drivers / accel / ivpu / ivpu_job.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2020-2023 Intel Corporation
4  */
5
6 #include <drm/drm_file.h>
7
8 #include <linux/bitfield.h>
9 #include <linux/highmem.h>
10 #include <linux/kthread.h>
11 #include <linux/pci.h>
12 #include <linux/module.h>
13 #include <uapi/drm/ivpu_accel.h>
14
15 #include "ivpu_drv.h"
16 #include "ivpu_hw.h"
17 #include "ivpu_ipc.h"
18 #include "ivpu_job.h"
19 #include "ivpu_jsm_msg.h"
20 #include "ivpu_pm.h"
21
22 #define CMD_BUF_IDX          0
23 #define JOB_ID_JOB_MASK      GENMASK(7, 0)
24 #define JOB_ID_CONTEXT_MASK  GENMASK(31, 8)
25 #define JOB_MAX_BUFFER_COUNT 65535
26
27 static unsigned int ivpu_tdr_timeout_ms;
28 module_param_named(tdr_timeout_ms, ivpu_tdr_timeout_ms, uint, 0644);
29 MODULE_PARM_DESC(tdr_timeout_ms, "Timeout for device hang detection, in milliseconds, 0 - default");
30
31 static void ivpu_cmdq_ring_db(struct ivpu_device *vdev, struct ivpu_cmdq *cmdq)
32 {
33         ivpu_hw_reg_db_set(vdev, cmdq->db_id);
34 }
35
36 static struct ivpu_cmdq *ivpu_cmdq_alloc(struct ivpu_file_priv *file_priv, u16 engine)
37 {
38         struct ivpu_device *vdev = file_priv->vdev;
39         struct vpu_job_queue_header *jobq_header;
40         struct ivpu_cmdq *cmdq;
41
42         cmdq = kzalloc(sizeof(*cmdq), GFP_KERNEL);
43         if (!cmdq)
44                 return NULL;
45
46         cmdq->mem = ivpu_bo_alloc_internal(vdev, 0, SZ_4K, DRM_IVPU_BO_WC);
47         if (!cmdq->mem)
48                 goto cmdq_free;
49
50         cmdq->db_id = file_priv->ctx.id + engine * ivpu_get_context_count(vdev);
51         cmdq->entry_count = (u32)((cmdq->mem->base.size - sizeof(struct vpu_job_queue_header)) /
52                                   sizeof(struct vpu_job_queue_entry));
53
54         cmdq->jobq = (struct vpu_job_queue *)cmdq->mem->kvaddr;
55         jobq_header = &cmdq->jobq->header;
56         jobq_header->engine_idx = engine;
57         jobq_header->head = 0;
58         jobq_header->tail = 0;
59         wmb(); /* Flush WC buffer for jobq->header */
60
61         return cmdq;
62
63 cmdq_free:
64         kfree(cmdq);
65         return NULL;
66 }
67
68 static void ivpu_cmdq_free(struct ivpu_file_priv *file_priv, struct ivpu_cmdq *cmdq)
69 {
70         if (!cmdq)
71                 return;
72
73         ivpu_bo_free_internal(cmdq->mem);
74         kfree(cmdq);
75 }
76
77 static struct ivpu_cmdq *ivpu_cmdq_acquire(struct ivpu_file_priv *file_priv, u16 engine)
78 {
79         struct ivpu_device *vdev = file_priv->vdev;
80         struct ivpu_cmdq *cmdq = file_priv->cmdq[engine];
81         int ret;
82
83         lockdep_assert_held(&file_priv->lock);
84
85         if (!cmdq) {
86                 cmdq = ivpu_cmdq_alloc(file_priv, engine);
87                 if (!cmdq)
88                         return NULL;
89                 file_priv->cmdq[engine] = cmdq;
90         }
91
92         if (cmdq->db_registered)
93                 return cmdq;
94
95         ret = ivpu_jsm_register_db(vdev, file_priv->ctx.id, cmdq->db_id,
96                                    cmdq->mem->vpu_addr, cmdq->mem->base.size);
97         if (ret)
98                 return NULL;
99
100         cmdq->db_registered = true;
101
102         return cmdq;
103 }
104
105 static void ivpu_cmdq_release_locked(struct ivpu_file_priv *file_priv, u16 engine)
106 {
107         struct ivpu_cmdq *cmdq = file_priv->cmdq[engine];
108
109         lockdep_assert_held(&file_priv->lock);
110
111         if (cmdq) {
112                 file_priv->cmdq[engine] = NULL;
113                 if (cmdq->db_registered)
114                         ivpu_jsm_unregister_db(file_priv->vdev, cmdq->db_id);
115
116                 ivpu_cmdq_free(file_priv, cmdq);
117         }
118 }
119
120 void ivpu_cmdq_release_all(struct ivpu_file_priv *file_priv)
121 {
122         int i;
123
124         mutex_lock(&file_priv->lock);
125
126         for (i = 0; i < IVPU_NUM_ENGINES; i++)
127                 ivpu_cmdq_release_locked(file_priv, i);
128
129         mutex_unlock(&file_priv->lock);
130 }
131
132 /*
133  * Mark the doorbell as unregistered and reset job queue pointers.
134  * This function needs to be called when the VPU hardware is restarted
135  * and FW looses job queue state. The next time job queue is used it
136  * will be registered again.
137  */
138 static void ivpu_cmdq_reset_locked(struct ivpu_file_priv *file_priv, u16 engine)
139 {
140         struct ivpu_cmdq *cmdq = file_priv->cmdq[engine];
141
142         lockdep_assert_held(&file_priv->lock);
143
144         if (cmdq) {
145                 cmdq->db_registered = false;
146                 cmdq->jobq->header.head = 0;
147                 cmdq->jobq->header.tail = 0;
148                 wmb(); /* Flush WC buffer for jobq header */
149         }
150 }
151
152 static void ivpu_cmdq_reset_all(struct ivpu_file_priv *file_priv)
153 {
154         int i;
155
156         mutex_lock(&file_priv->lock);
157
158         for (i = 0; i < IVPU_NUM_ENGINES; i++)
159                 ivpu_cmdq_reset_locked(file_priv, i);
160
161         mutex_unlock(&file_priv->lock);
162 }
163
164 void ivpu_cmdq_reset_all_contexts(struct ivpu_device *vdev)
165 {
166         struct ivpu_file_priv *file_priv;
167         unsigned long ctx_id;
168
169         xa_for_each(&vdev->context_xa, ctx_id, file_priv) {
170                 file_priv = ivpu_file_priv_get_by_ctx_id(vdev, ctx_id);
171                 if (!file_priv)
172                         continue;
173
174                 ivpu_cmdq_reset_all(file_priv);
175
176                 ivpu_file_priv_put(&file_priv);
177         }
178 }
179
180 static int ivpu_cmdq_push_job(struct ivpu_cmdq *cmdq, struct ivpu_job *job)
181 {
182         struct ivpu_device *vdev = job->vdev;
183         struct vpu_job_queue_header *header = &cmdq->jobq->header;
184         struct vpu_job_queue_entry *entry;
185         u32 tail = READ_ONCE(header->tail);
186         u32 next_entry = (tail + 1) % cmdq->entry_count;
187
188         /* Check if there is space left in job queue */
189         if (next_entry == header->head) {
190                 ivpu_dbg(vdev, JOB, "Job queue full: ctx %d engine %d db %d head %d tail %d\n",
191                          job->file_priv->ctx.id, job->engine_idx, cmdq->db_id, header->head, tail);
192                 return -EBUSY;
193         }
194
195         entry = &cmdq->jobq->job[tail];
196         entry->batch_buf_addr = job->cmd_buf_vpu_addr;
197         entry->job_id = job->job_id;
198         entry->flags = 0;
199         wmb(); /* Ensure that tail is updated after filling entry */
200         header->tail = next_entry;
201         wmb(); /* Flush WC buffer for jobq header */
202
203         return 0;
204 }
205
206 struct ivpu_fence {
207         struct dma_fence base;
208         spinlock_t lock; /* protects base */
209         struct ivpu_device *vdev;
210 };
211
212 static inline struct ivpu_fence *to_vpu_fence(struct dma_fence *fence)
213 {
214         return container_of(fence, struct ivpu_fence, base);
215 }
216
217 static const char *ivpu_fence_get_driver_name(struct dma_fence *fence)
218 {
219         return DRIVER_NAME;
220 }
221
222 static const char *ivpu_fence_get_timeline_name(struct dma_fence *fence)
223 {
224         struct ivpu_fence *ivpu_fence = to_vpu_fence(fence);
225
226         return dev_name(ivpu_fence->vdev->drm.dev);
227 }
228
229 static const struct dma_fence_ops ivpu_fence_ops = {
230         .get_driver_name = ivpu_fence_get_driver_name,
231         .get_timeline_name = ivpu_fence_get_timeline_name,
232 };
233
234 static struct dma_fence *ivpu_fence_create(struct ivpu_device *vdev)
235 {
236         struct ivpu_fence *fence;
237
238         fence = kzalloc(sizeof(*fence), GFP_KERNEL);
239         if (!fence)
240                 return NULL;
241
242         fence->vdev = vdev;
243         spin_lock_init(&fence->lock);
244         dma_fence_init(&fence->base, &ivpu_fence_ops, &fence->lock, dma_fence_context_alloc(1), 1);
245
246         return &fence->base;
247 }
248
249 static void job_get(struct ivpu_job *job, struct ivpu_job **link)
250 {
251         struct ivpu_device *vdev = job->vdev;
252
253         kref_get(&job->ref);
254         *link = job;
255
256         ivpu_dbg(vdev, KREF, "Job get: id %u refcount %u\n", job->job_id, kref_read(&job->ref));
257 }
258
259 static void job_release(struct kref *ref)
260 {
261         struct ivpu_job *job = container_of(ref, struct ivpu_job, ref);
262         struct ivpu_device *vdev = job->vdev;
263         u32 i;
264
265         for (i = 0; i < job->bo_count; i++)
266                 if (job->bos[i])
267                         drm_gem_object_put(&job->bos[i]->base);
268
269         dma_fence_put(job->done_fence);
270         ivpu_file_priv_put(&job->file_priv);
271
272         ivpu_dbg(vdev, KREF, "Job released: id %u\n", job->job_id);
273         kfree(job);
274
275         /* Allow the VPU to get suspended, must be called after ivpu_file_priv_put() */
276         ivpu_rpm_put(vdev);
277 }
278
279 static void job_put(struct ivpu_job *job)
280 {
281         struct ivpu_device *vdev = job->vdev;
282
283         ivpu_dbg(vdev, KREF, "Job put: id %u refcount %u\n", job->job_id, kref_read(&job->ref));
284         kref_put(&job->ref, job_release);
285 }
286
287 static struct ivpu_job *
288 ivpu_create_job(struct ivpu_file_priv *file_priv, u32 engine_idx, u32 bo_count)
289 {
290         struct ivpu_device *vdev = file_priv->vdev;
291         struct ivpu_job *job;
292         int ret;
293
294         ret = ivpu_rpm_get(vdev);
295         if (ret < 0)
296                 return NULL;
297
298         job = kzalloc(struct_size(job, bos, bo_count), GFP_KERNEL);
299         if (!job)
300                 goto err_rpm_put;
301
302         kref_init(&job->ref);
303
304         job->vdev = vdev;
305         job->engine_idx = engine_idx;
306         job->bo_count = bo_count;
307         job->done_fence = ivpu_fence_create(vdev);
308         if (!job->done_fence) {
309                 ivpu_warn_ratelimited(vdev, "Failed to create a fence\n");
310                 goto err_free_job;
311         }
312
313         job->file_priv = ivpu_file_priv_get(file_priv);
314
315         ivpu_dbg(vdev, JOB, "Job created: ctx %2d engine %d", file_priv->ctx.id, job->engine_idx);
316
317         return job;
318
319 err_free_job:
320         kfree(job);
321 err_rpm_put:
322         ivpu_rpm_put(vdev);
323         return NULL;
324 }
325
326 static int ivpu_job_done(struct ivpu_device *vdev, u32 job_id, u32 job_status)
327 {
328         struct ivpu_job *job;
329
330         job = xa_erase(&vdev->submitted_jobs_xa, job_id);
331         if (!job)
332                 return -ENOENT;
333
334         if (job->file_priv->has_mmu_faults)
335                 job_status = VPU_JSM_STATUS_ABORTED;
336
337         job->bos[CMD_BUF_IDX]->job_status = job_status;
338         dma_fence_signal(job->done_fence);
339
340         ivpu_dbg(vdev, JOB, "Job complete:  id %3u ctx %2d engine %d status 0x%x\n",
341                  job->job_id, job->file_priv->ctx.id, job->engine_idx, job_status);
342
343         job_put(job);
344         return 0;
345 }
346
347 static void ivpu_job_done_message(struct ivpu_device *vdev, void *msg)
348 {
349         struct vpu_ipc_msg_payload_job_done *payload;
350         struct vpu_jsm_msg *job_ret_msg = msg;
351         int ret;
352
353         payload = (struct vpu_ipc_msg_payload_job_done *)&job_ret_msg->payload;
354
355         ret = ivpu_job_done(vdev, payload->job_id, payload->job_status);
356         if (ret)
357                 ivpu_err(vdev, "Failed to finish job %d: %d\n", payload->job_id, ret);
358 }
359
360 void ivpu_jobs_abort_all(struct ivpu_device *vdev)
361 {
362         struct ivpu_job *job;
363         unsigned long id;
364
365         xa_for_each(&vdev->submitted_jobs_xa, id, job)
366                 ivpu_job_done(vdev, id, VPU_JSM_STATUS_ABORTED);
367 }
368
369 static int ivpu_direct_job_submission(struct ivpu_job *job)
370 {
371         struct ivpu_file_priv *file_priv = job->file_priv;
372         struct ivpu_device *vdev = job->vdev;
373         struct xa_limit job_id_range;
374         struct ivpu_cmdq *cmdq;
375         int ret;
376
377         mutex_lock(&file_priv->lock);
378
379         cmdq = ivpu_cmdq_acquire(job->file_priv, job->engine_idx);
380         if (!cmdq) {
381                 ivpu_warn(vdev, "Failed get job queue, ctx %d engine %d\n",
382                           file_priv->ctx.id, job->engine_idx);
383                 ret = -EINVAL;
384                 goto err_unlock;
385         }
386
387         job_id_range.min = FIELD_PREP(JOB_ID_CONTEXT_MASK, (file_priv->ctx.id - 1));
388         job_id_range.max = job_id_range.min | JOB_ID_JOB_MASK;
389
390         job_get(job, &job);
391         ret = xa_alloc(&vdev->submitted_jobs_xa, &job->job_id, job, job_id_range, GFP_KERNEL);
392         if (ret) {
393                 ivpu_warn_ratelimited(vdev, "Failed to allocate job id: %d\n", ret);
394                 goto err_job_put;
395         }
396
397         ret = ivpu_cmdq_push_job(cmdq, job);
398         if (ret)
399                 goto err_xa_erase;
400
401         ivpu_dbg(vdev, JOB, "Job submitted: id %3u addr 0x%llx ctx %2d engine %d next %d\n",
402                  job->job_id, job->cmd_buf_vpu_addr, file_priv->ctx.id,
403                  job->engine_idx, cmdq->jobq->header.tail);
404
405         if (ivpu_test_mode == IVPU_TEST_MODE_NULL_HW) {
406                 ivpu_job_done(vdev, job->job_id, VPU_JSM_STATUS_SUCCESS);
407                 cmdq->jobq->header.head = cmdq->jobq->header.tail;
408                 wmb(); /* Flush WC buffer for jobq header */
409         } else {
410                 ivpu_cmdq_ring_db(vdev, cmdq);
411         }
412
413         mutex_unlock(&file_priv->lock);
414         return 0;
415
416 err_xa_erase:
417         xa_erase(&vdev->submitted_jobs_xa, job->job_id);
418 err_job_put:
419         job_put(job);
420 err_unlock:
421         mutex_unlock(&file_priv->lock);
422         return ret;
423 }
424
425 static int
426 ivpu_job_prepare_bos_for_submit(struct drm_file *file, struct ivpu_job *job, u32 *buf_handles,
427                                 u32 buf_count, u32 commands_offset)
428 {
429         struct ivpu_file_priv *file_priv = file->driver_priv;
430         struct ivpu_device *vdev = file_priv->vdev;
431         struct ww_acquire_ctx acquire_ctx;
432         enum dma_resv_usage usage;
433         struct ivpu_bo *bo;
434         int ret;
435         u32 i;
436
437         for (i = 0; i < buf_count; i++) {
438                 struct drm_gem_object *obj = drm_gem_object_lookup(file, buf_handles[i]);
439
440                 if (!obj)
441                         return -ENOENT;
442
443                 job->bos[i] = to_ivpu_bo(obj);
444
445                 ret = ivpu_bo_pin(job->bos[i]);
446                 if (ret)
447                         return ret;
448         }
449
450         bo = job->bos[CMD_BUF_IDX];
451         if (!dma_resv_test_signaled(bo->base.resv, DMA_RESV_USAGE_READ)) {
452                 ivpu_warn(vdev, "Buffer is already in use\n");
453                 return -EBUSY;
454         }
455
456         if (commands_offset >= bo->base.size) {
457                 ivpu_warn(vdev, "Invalid command buffer offset %u\n", commands_offset);
458                 return -EINVAL;
459         }
460
461         job->cmd_buf_vpu_addr = bo->vpu_addr + commands_offset;
462
463         ret = drm_gem_lock_reservations((struct drm_gem_object **)job->bos, buf_count,
464                                         &acquire_ctx);
465         if (ret) {
466                 ivpu_warn(vdev, "Failed to lock reservations: %d\n", ret);
467                 return ret;
468         }
469
470         for (i = 0; i < buf_count; i++) {
471                 ret = dma_resv_reserve_fences(job->bos[i]->base.resv, 1);
472                 if (ret) {
473                         ivpu_warn(vdev, "Failed to reserve fences: %d\n", ret);
474                         goto unlock_reservations;
475                 }
476         }
477
478         for (i = 0; i < buf_count; i++) {
479                 usage = (i == CMD_BUF_IDX) ? DMA_RESV_USAGE_WRITE : DMA_RESV_USAGE_BOOKKEEP;
480                 dma_resv_add_fence(job->bos[i]->base.resv, job->done_fence, usage);
481         }
482
483 unlock_reservations:
484         drm_gem_unlock_reservations((struct drm_gem_object **)job->bos, buf_count, &acquire_ctx);
485
486         wmb(); /* Flush write combining buffers */
487
488         return ret;
489 }
490
491 int ivpu_submit_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
492 {
493         struct ivpu_file_priv *file_priv = file->driver_priv;
494         struct ivpu_device *vdev = file_priv->vdev;
495         struct drm_ivpu_submit *params = data;
496         struct ivpu_job *job;
497         u32 *buf_handles;
498         int idx, ret;
499
500         if (params->engine > DRM_IVPU_ENGINE_COPY)
501                 return -EINVAL;
502
503         if (params->buffer_count == 0 || params->buffer_count > JOB_MAX_BUFFER_COUNT)
504                 return -EINVAL;
505
506         if (!IS_ALIGNED(params->commands_offset, 8))
507                 return -EINVAL;
508
509         if (!file_priv->ctx.id)
510                 return -EINVAL;
511
512         if (file_priv->has_mmu_faults)
513                 return -EBADFD;
514
515         buf_handles = kcalloc(params->buffer_count, sizeof(u32), GFP_KERNEL);
516         if (!buf_handles)
517                 return -ENOMEM;
518
519         ret = copy_from_user(buf_handles,
520                              (void __user *)params->buffers_ptr,
521                              params->buffer_count * sizeof(u32));
522         if (ret) {
523                 ret = -EFAULT;
524                 goto free_handles;
525         }
526
527         if (!drm_dev_enter(&vdev->drm, &idx)) {
528                 ret = -ENODEV;
529                 goto free_handles;
530         }
531
532         ivpu_dbg(vdev, JOB, "Submit ioctl: ctx %u buf_count %u\n",
533                  file_priv->ctx.id, params->buffer_count);
534
535         job = ivpu_create_job(file_priv, params->engine, params->buffer_count);
536         if (!job) {
537                 ivpu_err(vdev, "Failed to create job\n");
538                 ret = -ENOMEM;
539                 goto dev_exit;
540         }
541
542         ret = ivpu_job_prepare_bos_for_submit(file, job, buf_handles, params->buffer_count,
543                                               params->commands_offset);
544         if (ret) {
545                 ivpu_err(vdev, "Failed to prepare job, ret %d\n", ret);
546                 goto job_put;
547         }
548
549         ret = ivpu_direct_job_submission(job);
550         if (ret) {
551                 dma_fence_signal(job->done_fence);
552                 ivpu_err(vdev, "Failed to submit job to the HW, ret %d\n", ret);
553         }
554
555 job_put:
556         job_put(job);
557 dev_exit:
558         drm_dev_exit(idx);
559 free_handles:
560         kfree(buf_handles);
561
562         return ret;
563 }
564
565 static int ivpu_job_done_thread(void *arg)
566 {
567         struct ivpu_device *vdev = (struct ivpu_device *)arg;
568         struct ivpu_ipc_consumer cons;
569         struct vpu_jsm_msg jsm_msg;
570         bool jobs_submitted;
571         unsigned int timeout;
572         int ret;
573
574         ivpu_dbg(vdev, JOB, "Started %s\n", __func__);
575
576         ivpu_ipc_consumer_add(vdev, &cons, VPU_IPC_CHAN_JOB_RET);
577
578         while (!kthread_should_stop()) {
579                 timeout = ivpu_tdr_timeout_ms ? ivpu_tdr_timeout_ms : vdev->timeout.tdr;
580                 jobs_submitted = !xa_empty(&vdev->submitted_jobs_xa);
581                 ret = ivpu_ipc_receive(vdev, &cons, NULL, &jsm_msg, timeout);
582                 if (!ret) {
583                         ivpu_job_done_message(vdev, &jsm_msg);
584                 } else if (ret == -ETIMEDOUT) {
585                         if (jobs_submitted && !xa_empty(&vdev->submitted_jobs_xa)) {
586                                 ivpu_err(vdev, "TDR detected, timeout %d ms", timeout);
587                                 ivpu_hw_diagnose_failure(vdev);
588                                 ivpu_pm_schedule_recovery(vdev);
589                         }
590                 }
591         }
592
593         ivpu_ipc_consumer_del(vdev, &cons);
594
595         ivpu_jobs_abort_all(vdev);
596
597         ivpu_dbg(vdev, JOB, "Stopped %s\n", __func__);
598         return 0;
599 }
600
601 int ivpu_job_done_thread_init(struct ivpu_device *vdev)
602 {
603         struct task_struct *thread;
604
605         thread = kthread_run(&ivpu_job_done_thread, (void *)vdev, "ivpu_job_done_thread");
606         if (IS_ERR(thread)) {
607                 ivpu_err(vdev, "Failed to start job completion thread\n");
608                 return -EIO;
609         }
610
611         get_task_struct(thread);
612         wake_up_process(thread);
613
614         vdev->job_done_thread = thread;
615
616         return 0;
617 }
618
619 void ivpu_job_done_thread_fini(struct ivpu_device *vdev)
620 {
621         kthread_stop(vdev->job_done_thread);
622         put_task_struct(vdev->job_done_thread);
623 }