xref: /linux/drivers/media/platform/amphion/vpu_v4l2.c (revision 68a052239fc4b351e961f698b824f7654a346091)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright 2020-2021 NXP
4  */
5 
6 #include <linux/init.h>
7 #include <linux/interconnect.h>
8 #include <linux/ioctl.h>
9 #include <linux/list.h>
10 #include <linux/kernel.h>
11 #include <linux/module.h>
12 #include <linux/pm_runtime.h>
13 #include <linux/videodev2.h>
14 #include <media/v4l2-device.h>
15 #include <media/v4l2-event.h>
16 #include <media/v4l2-mem2mem.h>
17 #include <media/v4l2-ioctl.h>
18 #include <media/videobuf2-v4l2.h>
19 #include <media/videobuf2-dma-contig.h>
20 #include <media/videobuf2-vmalloc.h>
21 #include "vpu.h"
22 #include "vpu_core.h"
23 #include "vpu_v4l2.h"
24 #include "vpu_msgs.h"
25 #include "vpu_helpers.h"
26 
27 static char *vpu_type_name(u32 type)
28 {
29 	return V4L2_TYPE_IS_OUTPUT(type) ? "output" : "capture";
30 }
31 
32 void vpu_inst_lock(struct vpu_inst *inst)
33 {
34 	mutex_lock(&inst->lock);
35 }
36 
37 void vpu_inst_unlock(struct vpu_inst *inst)
38 {
39 	mutex_unlock(&inst->lock);
40 }
41 
42 dma_addr_t vpu_get_vb_phy_addr(struct vb2_buffer *vb, u32 plane_no)
43 {
44 	if (plane_no >= vb->num_planes)
45 		return 0;
46 	return vb2_dma_contig_plane_dma_addr(vb, plane_no) +
47 			vb->planes[plane_no].data_offset;
48 }
49 
50 static unsigned int vpu_get_vb_length(struct vb2_buffer *vb, u32 plane_no)
51 {
52 	if (plane_no >= vb->num_planes)
53 		return 0;
54 	return vb2_plane_size(vb, plane_no) - vb->planes[plane_no].data_offset;
55 }
56 
57 void vpu_set_buffer_state(struct vb2_v4l2_buffer *vbuf, unsigned int state)
58 {
59 	struct vpu_vb2_buffer *vpu_buf = to_vpu_vb2_buffer(vbuf);
60 
61 	vpu_buf->state = state;
62 }
63 
64 unsigned int vpu_get_buffer_state(struct vb2_v4l2_buffer *vbuf)
65 {
66 	struct vpu_vb2_buffer *vpu_buf = to_vpu_vb2_buffer(vbuf);
67 
68 	return vpu_buf->state;
69 }
70 
71 void vpu_set_buffer_average_qp(struct vb2_v4l2_buffer *vbuf, u32 qp)
72 {
73 	struct vpu_vb2_buffer *vpu_buf = to_vpu_vb2_buffer(vbuf);
74 
75 	vpu_buf->average_qp = qp;
76 }
77 
78 void vpu_v4l2_set_error(struct vpu_inst *inst)
79 {
80 	vpu_inst_lock(inst);
81 	dev_err(inst->dev, "some error occurs in codec\n");
82 	if (inst->fh.m2m_ctx) {
83 		vb2_queue_error(v4l2_m2m_get_src_vq(inst->fh.m2m_ctx));
84 		vb2_queue_error(v4l2_m2m_get_dst_vq(inst->fh.m2m_ctx));
85 	}
86 	vpu_inst_unlock(inst);
87 }
88 
89 static int vpu_notify_eos(struct vpu_inst *inst)
90 {
91 	static const struct v4l2_event ev = {
92 		.id = 0,
93 		.type = V4L2_EVENT_EOS
94 	};
95 
96 	vpu_trace(inst->dev, "[%d]\n", inst->id);
97 	v4l2_event_queue_fh(&inst->fh, &ev);
98 
99 	return 0;
100 }
101 
102 int vpu_notify_source_change(struct vpu_inst *inst)
103 {
104 	static const struct v4l2_event ev = {
105 		.id = 0,
106 		.type = V4L2_EVENT_SOURCE_CHANGE,
107 		.u.src_change.changes = V4L2_EVENT_SRC_CH_RESOLUTION
108 	};
109 
110 	vpu_trace(inst->dev, "[%d]\n", inst->id);
111 	v4l2_event_queue_fh(&inst->fh, &ev);
112 	return 0;
113 }
114 
115 int vpu_set_last_buffer_dequeued(struct vpu_inst *inst, bool eos)
116 {
117 	struct vb2_queue *q;
118 
119 	if (!inst || !inst->fh.m2m_ctx)
120 		return -EINVAL;
121 
122 	q = v4l2_m2m_get_dst_vq(inst->fh.m2m_ctx);
123 	if (!list_empty(&q->done_list))
124 		return -EINVAL;
125 
126 	if (q->last_buffer_dequeued)
127 		return 0;
128 	vpu_trace(inst->dev, "last buffer dequeued\n");
129 	q->last_buffer_dequeued = true;
130 	wake_up(&q->done_wq);
131 	if (eos)
132 		vpu_notify_eos(inst);
133 	return 0;
134 }
135 
136 bool vpu_is_source_empty(struct vpu_inst *inst)
137 {
138 	struct v4l2_m2m_buffer *buf = NULL;
139 
140 	if (!inst->fh.m2m_ctx)
141 		return true;
142 	v4l2_m2m_for_each_src_buf(inst->fh.m2m_ctx, buf) {
143 		if (vpu_get_buffer_state(&buf->vb) == VPU_BUF_STATE_IDLE)
144 			return false;
145 	}
146 	return true;
147 }
148 
149 static int vpu_init_format(struct vpu_inst *inst, struct vpu_format *fmt)
150 {
151 	const struct vpu_format *info;
152 
153 	info = vpu_helper_find_format(inst, fmt->type, fmt->pixfmt);
154 	if (!info) {
155 		info = vpu_helper_enum_format(inst, fmt->type, 0);
156 		if (!info)
157 			return -EINVAL;
158 	}
159 	memcpy(fmt, info, sizeof(*fmt));
160 
161 	return 0;
162 }
163 
164 static int vpu_calc_fmt_bytesperline(struct v4l2_format *f, struct vpu_format *fmt)
165 {
166 	struct v4l2_pix_format_mplane *pixmp = &f->fmt.pix_mp;
167 	int i;
168 
169 	if (fmt->flags & V4L2_FMT_FLAG_COMPRESSED) {
170 		for (i = 0; i < fmt->comp_planes; i++)
171 			fmt->bytesperline[i] = 0;
172 		return 0;
173 	}
174 	if (pixmp->num_planes == fmt->comp_planes) {
175 		for (i = 0; i < fmt->comp_planes; i++)
176 			fmt->bytesperline[i] = pixmp->plane_fmt[i].bytesperline;
177 		return 0;
178 	}
179 	if (pixmp->num_planes > 1)
180 		return -EINVAL;
181 
182 	/*amphion vpu only support nv12 and nv12 tiled,
183 	 * so the bytesperline of luma and chroma should be same
184 	 */
185 	for (i = 0; i < fmt->comp_planes; i++)
186 		fmt->bytesperline[i] = pixmp->plane_fmt[0].bytesperline;
187 
188 	return 0;
189 }
190 
191 static int vpu_calc_fmt_sizeimage(struct vpu_inst *inst, struct vpu_format *fmt)
192 {
193 	u32 stride = 1;
194 	int i;
195 
196 	if (!(fmt->flags & V4L2_FMT_FLAG_COMPRESSED)) {
197 		const struct vpu_core_resources *res = vpu_get_resource(inst);
198 
199 		if (res)
200 			stride = res->stride;
201 	}
202 
203 	for (i = 0; i < fmt->comp_planes; i++) {
204 		fmt->sizeimage[i] = vpu_helper_get_plane_size(fmt->pixfmt,
205 							      fmt->width,
206 							      fmt->height,
207 							      i,
208 							      stride,
209 							      fmt->field != V4L2_FIELD_NONE ? 1 : 0,
210 							      &fmt->bytesperline[i]);
211 		fmt->sizeimage[i] = max_t(u32, fmt->sizeimage[i], PAGE_SIZE);
212 		if (fmt->flags & V4L2_FMT_FLAG_COMPRESSED) {
213 			fmt->sizeimage[i] = clamp_val(fmt->sizeimage[i], SZ_128K, SZ_8M);
214 			fmt->bytesperline[i] = 0;
215 		}
216 	}
217 
218 	return 0;
219 }
220 
221 u32 vpu_get_fmt_plane_size(struct vpu_format *fmt, u32 plane_no)
222 {
223 	u32 size;
224 	int i;
225 
226 	if (plane_no >= fmt->mem_planes)
227 		return 0;
228 
229 	if (fmt->comp_planes == fmt->mem_planes)
230 		return fmt->sizeimage[plane_no];
231 	if (plane_no < fmt->mem_planes - 1)
232 		return fmt->sizeimage[plane_no];
233 
234 	size = fmt->sizeimage[plane_no];
235 	for (i = fmt->mem_planes; i < fmt->comp_planes; i++)
236 		size += fmt->sizeimage[i];
237 
238 	return size;
239 }
240 
241 int vpu_try_fmt_common(struct vpu_inst *inst, struct v4l2_format *f, struct vpu_format *fmt)
242 {
243 	struct v4l2_pix_format_mplane *pixmp = &f->fmt.pix_mp;
244 	int i;
245 	int ret;
246 
247 	fmt->pixfmt = pixmp->pixelformat;
248 	fmt->type = f->type;
249 	ret = vpu_init_format(inst, fmt);
250 	if (ret < 0)
251 		return ret;
252 
253 	fmt->width = pixmp->width;
254 	fmt->height = pixmp->height;
255 	if (fmt->width)
256 		fmt->width = vpu_helper_valid_frame_width(inst, fmt->width);
257 	if (fmt->height)
258 		fmt->height = vpu_helper_valid_frame_height(inst, fmt->height);
259 	fmt->field = pixmp->field == V4L2_FIELD_ANY ? V4L2_FIELD_NONE : pixmp->field;
260 	vpu_calc_fmt_bytesperline(f, fmt);
261 	vpu_calc_fmt_sizeimage(inst, fmt);
262 	if ((fmt->flags & V4L2_FMT_FLAG_COMPRESSED) && pixmp->plane_fmt[0].sizeimage)
263 		fmt->sizeimage[0] = clamp_val(pixmp->plane_fmt[0].sizeimage, SZ_128K, SZ_8M);
264 
265 	pixmp->pixelformat = fmt->pixfmt;
266 	pixmp->width = fmt->width;
267 	pixmp->height = fmt->height;
268 	pixmp->flags = fmt->flags;
269 	pixmp->num_planes = fmt->mem_planes;
270 	pixmp->field = fmt->field;
271 	memset(pixmp->reserved, 0, sizeof(pixmp->reserved));
272 	for (i = 0; i < pixmp->num_planes; i++) {
273 		pixmp->plane_fmt[i].bytesperline = fmt->bytesperline[i];
274 		pixmp->plane_fmt[i].sizeimage = vpu_get_fmt_plane_size(fmt, i);
275 		memset(pixmp->plane_fmt[i].reserved, 0, sizeof(pixmp->plane_fmt[i].reserved));
276 	}
277 
278 	return 0;
279 }
280 
281 static bool vpu_check_ready(struct vpu_inst *inst, u32 type)
282 {
283 	if (!inst)
284 		return false;
285 	if (inst->state == VPU_CODEC_STATE_DEINIT || inst->id < 0)
286 		return false;
287 	if (!inst->ops->check_ready)
288 		return true;
289 	return call_vop(inst, check_ready, type);
290 }
291 
292 int vpu_process_output_buffer(struct vpu_inst *inst)
293 {
294 	struct v4l2_m2m_buffer *buf = NULL;
295 	struct vb2_v4l2_buffer *vbuf = NULL;
296 
297 	if (!inst || !inst->fh.m2m_ctx)
298 		return -EINVAL;
299 
300 	if (!vpu_check_ready(inst, inst->out_format.type))
301 		return -EINVAL;
302 
303 	v4l2_m2m_for_each_src_buf(inst->fh.m2m_ctx, buf) {
304 		vbuf = &buf->vb;
305 		if (vpu_get_buffer_state(vbuf) == VPU_BUF_STATE_IDLE)
306 			break;
307 		vbuf = NULL;
308 	}
309 
310 	if (!vbuf)
311 		return -EINVAL;
312 
313 	dev_dbg(inst->dev, "[%d]frame id = %d / %d\n",
314 		inst->id, vbuf->sequence, inst->sequence);
315 	return call_vop(inst, process_output, &vbuf->vb2_buf);
316 }
317 
318 int vpu_process_capture_buffer(struct vpu_inst *inst)
319 {
320 	struct v4l2_m2m_buffer *buf = NULL;
321 	struct vb2_v4l2_buffer *vbuf = NULL;
322 
323 	if (!inst || !inst->fh.m2m_ctx)
324 		return -EINVAL;
325 
326 	if (!vpu_check_ready(inst, inst->cap_format.type))
327 		return -EINVAL;
328 
329 	v4l2_m2m_for_each_dst_buf(inst->fh.m2m_ctx, buf) {
330 		vbuf = &buf->vb;
331 		if (vpu_get_buffer_state(vbuf) == VPU_BUF_STATE_IDLE)
332 			break;
333 		vbuf = NULL;
334 	}
335 	if (!vbuf)
336 		return -EINVAL;
337 
338 	return call_vop(inst, process_capture, &vbuf->vb2_buf);
339 }
340 
341 struct vb2_v4l2_buffer *vpu_next_src_buf(struct vpu_inst *inst)
342 {
343 	struct vb2_v4l2_buffer *src_buf = NULL;
344 
345 	if (!inst->fh.m2m_ctx)
346 		return NULL;
347 
348 	src_buf = v4l2_m2m_next_src_buf(inst->fh.m2m_ctx);
349 	if (!src_buf || vpu_get_buffer_state(src_buf) == VPU_BUF_STATE_IDLE)
350 		return NULL;
351 
352 	while (vpu_vb_is_codecconfig(src_buf)) {
353 		v4l2_m2m_src_buf_remove(inst->fh.m2m_ctx);
354 		vpu_set_buffer_state(src_buf, VPU_BUF_STATE_IDLE);
355 		v4l2_m2m_buf_done(src_buf, VB2_BUF_STATE_DONE);
356 
357 		src_buf = v4l2_m2m_next_src_buf(inst->fh.m2m_ctx);
358 		if (!src_buf || vpu_get_buffer_state(src_buf) == VPU_BUF_STATE_IDLE)
359 			return NULL;
360 	}
361 
362 	return src_buf;
363 }
364 
365 void vpu_skip_frame(struct vpu_inst *inst, int count)
366 {
367 	struct vb2_v4l2_buffer *src_buf;
368 	enum vb2_buffer_state state;
369 	int i = 0;
370 
371 	if (count <= 0 || !inst->fh.m2m_ctx)
372 		return;
373 
374 	while (i < count) {
375 		src_buf = v4l2_m2m_src_buf_remove(inst->fh.m2m_ctx);
376 		if (!src_buf || vpu_get_buffer_state(src_buf) == VPU_BUF_STATE_IDLE)
377 			return;
378 		if (vpu_get_buffer_state(src_buf) == VPU_BUF_STATE_DECODED)
379 			state = VB2_BUF_STATE_DONE;
380 		else
381 			state = VB2_BUF_STATE_ERROR;
382 		i++;
383 		vpu_set_buffer_state(src_buf, VPU_BUF_STATE_IDLE);
384 		v4l2_m2m_buf_done(src_buf, state);
385 	}
386 }
387 
388 struct vb2_v4l2_buffer *vpu_find_buf_by_sequence(struct vpu_inst *inst, u32 type, u32 sequence)
389 {
390 	struct v4l2_m2m_buffer *buf = NULL;
391 	struct vb2_v4l2_buffer *vbuf = NULL;
392 
393 	if (!inst || !inst->fh.m2m_ctx)
394 		return NULL;
395 
396 	if (V4L2_TYPE_IS_OUTPUT(type)) {
397 		v4l2_m2m_for_each_src_buf(inst->fh.m2m_ctx, buf) {
398 			vbuf = &buf->vb;
399 			if (vbuf->sequence == sequence)
400 				break;
401 			vbuf = NULL;
402 		}
403 	} else {
404 		v4l2_m2m_for_each_dst_buf(inst->fh.m2m_ctx, buf) {
405 			vbuf = &buf->vb;
406 			if (vbuf->sequence == sequence)
407 				break;
408 			vbuf = NULL;
409 		}
410 	}
411 
412 	return vbuf;
413 }
414 
415 struct vb2_v4l2_buffer *vpu_find_buf_by_idx(struct vpu_inst *inst, u32 type, u32 idx)
416 {
417 	struct v4l2_m2m_buffer *buf = NULL;
418 	struct vb2_v4l2_buffer *vbuf = NULL;
419 
420 	if (!inst || !inst->fh.m2m_ctx)
421 		return NULL;
422 
423 	if (V4L2_TYPE_IS_OUTPUT(type)) {
424 		v4l2_m2m_for_each_src_buf(inst->fh.m2m_ctx, buf) {
425 			vbuf = &buf->vb;
426 			if (vbuf->vb2_buf.index == idx)
427 				break;
428 			vbuf = NULL;
429 		}
430 	} else {
431 		v4l2_m2m_for_each_dst_buf(inst->fh.m2m_ctx, buf) {
432 			vbuf = &buf->vb;
433 			if (vbuf->vb2_buf.index == idx)
434 				break;
435 			vbuf = NULL;
436 		}
437 	}
438 
439 	return vbuf;
440 }
441 
442 int vpu_get_num_buffers(struct vpu_inst *inst, u32 type)
443 {
444 	struct vb2_queue *q;
445 
446 	if (!inst || !inst->fh.m2m_ctx)
447 		return -EINVAL;
448 
449 	if (V4L2_TYPE_IS_OUTPUT(type))
450 		q = v4l2_m2m_get_src_vq(inst->fh.m2m_ctx);
451 	else
452 		q = v4l2_m2m_get_dst_vq(inst->fh.m2m_ctx);
453 
454 	return vb2_get_num_buffers(q);
455 }
456 
457 static void vpu_m2m_device_run(void *priv)
458 {
459 }
460 
461 static void vpu_m2m_job_abort(void *priv)
462 {
463 	struct vpu_inst *inst = priv;
464 	struct v4l2_m2m_ctx *m2m_ctx = inst->fh.m2m_ctx;
465 
466 	v4l2_m2m_job_finish(m2m_ctx->m2m_dev, m2m_ctx);
467 }
468 
469 static const struct v4l2_m2m_ops vpu_m2m_ops = {
470 	.device_run = vpu_m2m_device_run,
471 	.job_abort = vpu_m2m_job_abort
472 };
473 
474 static int vpu_vb2_queue_setup(struct vb2_queue *vq,
475 			       unsigned int *buf_count,
476 			       unsigned int *plane_count,
477 			       unsigned int psize[],
478 			       struct device *allocators[])
479 {
480 	struct vpu_inst *inst = vb2_get_drv_priv(vq);
481 	struct vpu_format *cur_fmt;
482 	int i;
483 
484 	cur_fmt = vpu_get_format(inst, vq->type);
485 
486 	if (*plane_count) {
487 		if (*plane_count != cur_fmt->mem_planes)
488 			return -EINVAL;
489 		for (i = 0; i < cur_fmt->mem_planes; i++) {
490 			if (psize[i] < vpu_get_fmt_plane_size(cur_fmt, i))
491 				return -EINVAL;
492 		}
493 		return 0;
494 	}
495 
496 	if (V4L2_TYPE_IS_OUTPUT(vq->type))
497 		*buf_count = max_t(unsigned int, *buf_count, inst->min_buffer_out);
498 	else
499 		*buf_count = max_t(unsigned int, *buf_count, inst->min_buffer_cap);
500 	*plane_count = cur_fmt->mem_planes;
501 	for (i = 0; i < cur_fmt->mem_planes; i++)
502 		psize[i] = vpu_get_fmt_plane_size(cur_fmt, i);
503 
504 	if (V4L2_TYPE_IS_OUTPUT(vq->type) && inst->state == VPU_CODEC_STATE_SEEK) {
505 		vpu_trace(inst->dev, "reinit when VIDIOC_REQBUFS(OUTPUT, 0)\n");
506 		call_void_vop(inst, release);
507 	}
508 
509 	if (V4L2_TYPE_IS_CAPTURE(vq->type))
510 		call_void_vop(inst, reset_frame_store);
511 
512 	return 0;
513 }
514 
515 static int vpu_vb2_buf_init(struct vb2_buffer *vb)
516 {
517 	struct vb2_v4l2_buffer *vbuf = to_vb2_v4l2_buffer(vb);
518 	struct vpu_vb2_buffer *vpu_buf = to_vpu_vb2_buffer(vbuf);
519 	struct vpu_inst *inst = vb2_get_drv_priv(vb->vb2_queue);
520 
521 	vpu_buf->fs_id = -1;
522 	vpu_set_buffer_state(vbuf, VPU_BUF_STATE_IDLE);
523 
524 	if (!inst->ops->attach_frame_store || V4L2_TYPE_IS_OUTPUT(vb->type))
525 		return 0;
526 
527 	call_void_vop(inst, attach_frame_store, vb);
528 	return 0;
529 }
530 
531 static int vpu_vb2_buf_out_validate(struct vb2_buffer *vb)
532 {
533 	struct vb2_v4l2_buffer *vbuf = to_vb2_v4l2_buffer(vb);
534 
535 	vbuf->field = V4L2_FIELD_NONE;
536 
537 	return 0;
538 }
539 
540 static int vpu_vb2_buf_prepare(struct vb2_buffer *vb)
541 {
542 	struct vpu_inst *inst = vb2_get_drv_priv(vb->vb2_queue);
543 	struct vb2_v4l2_buffer *vbuf = to_vb2_v4l2_buffer(vb);
544 	struct vpu_format *cur_fmt;
545 	u32 i;
546 
547 	cur_fmt = vpu_get_format(inst, vb->type);
548 	for (i = 0; i < cur_fmt->mem_planes; i++) {
549 		if (vpu_get_vb_length(vb, i) < vpu_get_fmt_plane_size(cur_fmt, i)) {
550 			dev_dbg(inst->dev, "[%d] %s buf[%d] is invalid\n",
551 				inst->id, vpu_type_name(vb->type), vb->index);
552 			vpu_set_buffer_state(vbuf, VPU_BUF_STATE_ERROR);
553 		}
554 	}
555 
556 	return 0;
557 }
558 
559 static void vpu_vb2_buf_finish(struct vb2_buffer *vb)
560 {
561 	struct vb2_v4l2_buffer *vbuf = to_vb2_v4l2_buffer(vb);
562 	struct vpu_inst *inst = vb2_get_drv_priv(vb->vb2_queue);
563 	struct vb2_queue *q = vb->vb2_queue;
564 
565 	if (V4L2_TYPE_IS_CAPTURE(vb->type)) {
566 		struct vpu_vb2_buffer *vpu_buf = to_vpu_vb2_buffer(vbuf);
567 		struct v4l2_ctrl *ctrl = v4l2_ctrl_find(&inst->ctrl_handler,
568 							V4L2_CID_MPEG_VIDEO_AVERAGE_QP);
569 
570 		if (ctrl)
571 			v4l2_ctrl_s_ctrl(ctrl, vpu_buf->average_qp);
572 	}
573 
574 	if (vbuf->flags & V4L2_BUF_FLAG_LAST)
575 		vpu_notify_eos(inst);
576 
577 	if (list_empty(&q->done_list))
578 		call_void_vop(inst, on_queue_empty, q->type);
579 }
580 
581 static void vpu_vb2_buffers_return(struct vpu_inst *inst, unsigned int type,
582 				   enum vb2_buffer_state state)
583 {
584 	struct vb2_v4l2_buffer *buf;
585 
586 	if (V4L2_TYPE_IS_OUTPUT(type)) {
587 		while ((buf = v4l2_m2m_src_buf_remove(inst->fh.m2m_ctx))) {
588 			vpu_set_buffer_state(buf, VPU_BUF_STATE_IDLE);
589 			v4l2_m2m_buf_done(buf, state);
590 		}
591 	} else {
592 		while ((buf = v4l2_m2m_dst_buf_remove(inst->fh.m2m_ctx))) {
593 			vpu_set_buffer_state(buf, VPU_BUF_STATE_IDLE);
594 			v4l2_m2m_buf_done(buf, state);
595 		}
596 	}
597 }
598 
599 static int vpu_vb2_start_streaming(struct vb2_queue *q, unsigned int count)
600 {
601 	struct vpu_inst *inst = vb2_get_drv_priv(q);
602 	struct vpu_format *fmt = vpu_get_format(inst, q->type);
603 	int ret;
604 
605 	vpu_inst_unlock(inst);
606 	ret = vpu_inst_register(inst);
607 	vpu_inst_lock(inst);
608 	if (ret) {
609 		vpu_vb2_buffers_return(inst, q->type, VB2_BUF_STATE_QUEUED);
610 		return ret;
611 	}
612 
613 	vpu_trace(inst->dev, "[%d] %s %c%c%c%c %dx%d %u(%u) %u(%u) %u(%u) %d\n",
614 		  inst->id, vpu_type_name(q->type),
615 		  fmt->pixfmt,
616 		  fmt->pixfmt >> 8,
617 		  fmt->pixfmt >> 16,
618 		  fmt->pixfmt >> 24,
619 		  fmt->width, fmt->height,
620 		  fmt->sizeimage[0], fmt->bytesperline[0],
621 		  fmt->sizeimage[1], fmt->bytesperline[1],
622 		  fmt->sizeimage[2], fmt->bytesperline[2],
623 		  vb2_get_num_buffers(q));
624 	vb2_clear_last_buffer_dequeued(q);
625 	ret = call_vop(inst, start, q->type);
626 	if (ret)
627 		vpu_vb2_buffers_return(inst, q->type, VB2_BUF_STATE_QUEUED);
628 
629 	return ret;
630 }
631 
632 static void vpu_vb2_stop_streaming(struct vb2_queue *q)
633 {
634 	struct vpu_inst *inst = vb2_get_drv_priv(q);
635 
636 	vpu_trace(inst->dev, "[%d] %s\n", inst->id, vpu_type_name(q->type));
637 
638 	call_void_vop(inst, stop, q->type);
639 	vpu_vb2_buffers_return(inst, q->type, VB2_BUF_STATE_ERROR);
640 	if (V4L2_TYPE_IS_OUTPUT(q->type))
641 		inst->sequence = 0;
642 }
643 
644 static void vpu_vb2_buf_queue(struct vb2_buffer *vb)
645 {
646 	struct vb2_v4l2_buffer *vbuf = to_vb2_v4l2_buffer(vb);
647 	struct vpu_inst *inst = vb2_get_drv_priv(vb->vb2_queue);
648 
649 	if (V4L2_TYPE_IS_OUTPUT(vb->type))
650 		vbuf->sequence = inst->sequence++;
651 
652 	v4l2_m2m_buf_queue(inst->fh.m2m_ctx, vbuf);
653 	vpu_process_output_buffer(inst);
654 	vpu_process_capture_buffer(inst);
655 }
656 
657 static const struct vb2_ops vpu_vb2_ops = {
658 	.queue_setup        = vpu_vb2_queue_setup,
659 	.buf_init           = vpu_vb2_buf_init,
660 	.buf_out_validate   = vpu_vb2_buf_out_validate,
661 	.buf_prepare        = vpu_vb2_buf_prepare,
662 	.buf_finish         = vpu_vb2_buf_finish,
663 	.start_streaming    = vpu_vb2_start_streaming,
664 	.stop_streaming     = vpu_vb2_stop_streaming,
665 	.buf_queue          = vpu_vb2_buf_queue,
666 };
667 
668 static int vpu_m2m_queue_init(void *priv, struct vb2_queue *src_vq, struct vb2_queue *dst_vq)
669 {
670 	struct vpu_inst *inst = priv;
671 	int ret;
672 
673 	src_vq->type = V4L2_BUF_TYPE_VIDEO_OUTPUT_MPLANE;
674 	inst->out_format.type = src_vq->type;
675 	src_vq->io_modes = VB2_MMAP | VB2_DMABUF;
676 	src_vq->timestamp_flags = V4L2_BUF_FLAG_TIMESTAMP_COPY;
677 	src_vq->ops = &vpu_vb2_ops;
678 	src_vq->mem_ops = &vb2_dma_contig_memops;
679 	if (inst->type == VPU_CORE_TYPE_DEC && inst->use_stream_buffer)
680 		src_vq->mem_ops = &vb2_vmalloc_memops;
681 	src_vq->drv_priv = inst;
682 	src_vq->buf_struct_size = sizeof(struct vpu_vb2_buffer);
683 	src_vq->min_queued_buffers = 1;
684 	src_vq->dev = inst->vpu->dev;
685 	src_vq->lock = &inst->lock;
686 	ret = vb2_queue_init(src_vq);
687 	if (ret)
688 		return ret;
689 
690 	dst_vq->type = V4L2_BUF_TYPE_VIDEO_CAPTURE_MPLANE;
691 	inst->cap_format.type = dst_vq->type;
692 	dst_vq->io_modes = VB2_MMAP | VB2_DMABUF;
693 	dst_vq->timestamp_flags = V4L2_BUF_FLAG_TIMESTAMP_COPY;
694 	dst_vq->ops = &vpu_vb2_ops;
695 	dst_vq->mem_ops = &vb2_dma_contig_memops;
696 	if (inst->type == VPU_CORE_TYPE_ENC && inst->use_stream_buffer)
697 		dst_vq->mem_ops = &vb2_vmalloc_memops;
698 	dst_vq->drv_priv = inst;
699 	dst_vq->buf_struct_size = sizeof(struct vpu_vb2_buffer);
700 	dst_vq->min_queued_buffers = 1;
701 	dst_vq->dev = inst->vpu->dev;
702 	dst_vq->lock = &inst->lock;
703 	ret = vb2_queue_init(dst_vq);
704 	if (ret) {
705 		vb2_queue_release(src_vq);
706 		return ret;
707 	}
708 
709 	return 0;
710 }
711 
712 static int vpu_v4l2_release(struct vpu_inst *inst)
713 {
714 	vpu_trace(inst->vpu->dev, "%p\n", inst);
715 
716 	vpu_release_core(inst->core);
717 	put_device(inst->dev);
718 
719 	if (inst->workqueue) {
720 		cancel_work_sync(&inst->msg_work);
721 		destroy_workqueue(inst->workqueue);
722 		inst->workqueue = NULL;
723 	}
724 
725 	v4l2_ctrl_handler_free(&inst->ctrl_handler);
726 	mutex_destroy(&inst->lock);
727 
728 	call_void_vop(inst, cleanup);
729 
730 	return 0;
731 }
732 
733 int vpu_v4l2_open(struct file *file, struct vpu_inst *inst)
734 {
735 	struct vpu_dev *vpu = video_drvdata(file);
736 	struct vpu_func *func;
737 	int ret = 0;
738 
739 	if (!inst || !inst->ops)
740 		return -EINVAL;
741 
742 	if (inst->type == VPU_CORE_TYPE_ENC)
743 		func = &vpu->encoder;
744 	else
745 		func = &vpu->decoder;
746 
747 	atomic_set(&inst->ref_count, 0);
748 	atomic_long_set(&inst->last_response_cmd, 0);
749 	vpu_inst_get(inst);
750 	inst->vpu = vpu;
751 	inst->core = vpu_request_core(vpu, inst->type);
752 	if (inst->core)
753 		inst->dev = get_device(inst->core->dev);
754 	mutex_init(&inst->lock);
755 	INIT_LIST_HEAD(&inst->cmd_q);
756 	inst->id = VPU_INST_NULL_ID;
757 	inst->release = vpu_v4l2_release;
758 	inst->pid = current->pid;
759 	inst->tgid = current->tgid;
760 	inst->min_buffer_cap = 2;
761 	inst->min_buffer_out = 2;
762 	v4l2_fh_init(&inst->fh, func->vfd);
763 	v4l2_fh_add(&inst->fh, file);
764 
765 	ret = call_vop(inst, ctrl_init);
766 	if (ret)
767 		goto error;
768 
769 	inst->fh.m2m_ctx = v4l2_m2m_ctx_init(func->m2m_dev, inst, vpu_m2m_queue_init);
770 	if (IS_ERR(inst->fh.m2m_ctx)) {
771 		dev_err(vpu->dev, "v4l2_m2m_ctx_init fail\n");
772 		ret = PTR_ERR(inst->fh.m2m_ctx);
773 		goto error;
774 	}
775 
776 	inst->fh.ctrl_handler = &inst->ctrl_handler;
777 	inst->state = VPU_CODEC_STATE_DEINIT;
778 	inst->workqueue = alloc_ordered_workqueue("vpu_inst", WQ_MEM_RECLAIM);
779 	if (inst->workqueue) {
780 		INIT_WORK(&inst->msg_work, vpu_inst_run_work);
781 		ret = kfifo_init(&inst->msg_fifo,
782 				 inst->msg_buffer,
783 				 rounddown_pow_of_two(sizeof(inst->msg_buffer)));
784 		if (ret) {
785 			destroy_workqueue(inst->workqueue);
786 			inst->workqueue = NULL;
787 		}
788 	}
789 	vpu_trace(vpu->dev, "tgid = %d, pid = %d, type = %s, inst = %p\n",
790 		  inst->tgid, inst->pid, vpu_core_type_desc(inst->type), inst);
791 
792 	return 0;
793 error:
794 	v4l2_fh_del(&inst->fh, file);
795 	v4l2_fh_exit(&inst->fh);
796 	vpu_inst_put(inst);
797 	return ret;
798 }
799 
800 int vpu_v4l2_close(struct file *file)
801 {
802 	struct vpu_dev *vpu = video_drvdata(file);
803 	struct vpu_inst *inst = to_inst(file);
804 
805 	vpu_trace(vpu->dev, "tgid = %d, pid = %d, inst = %p\n", inst->tgid, inst->pid, inst);
806 
807 	vpu_inst_lock(inst);
808 	if (inst->fh.m2m_ctx) {
809 		v4l2_m2m_ctx_release(inst->fh.m2m_ctx);
810 		inst->fh.m2m_ctx = NULL;
811 	}
812 	call_void_vop(inst, release);
813 	vpu_inst_unlock(inst);
814 
815 	v4l2_fh_del(&inst->fh, file);
816 	v4l2_fh_exit(&inst->fh);
817 
818 	vpu_inst_unregister(inst);
819 	vpu_inst_put(inst);
820 
821 	return 0;
822 }
823 
824 int vpu_add_func(struct vpu_dev *vpu, struct vpu_func *func)
825 {
826 	struct video_device *vfd;
827 	int ret;
828 
829 	if (!vpu || !func)
830 		return -EINVAL;
831 
832 	if (func->vfd)
833 		return 0;
834 
835 	func->m2m_dev = v4l2_m2m_init(&vpu_m2m_ops);
836 	if (IS_ERR(func->m2m_dev)) {
837 		dev_err(vpu->dev, "v4l2_m2m_init fail\n");
838 		func->vfd = NULL;
839 		return PTR_ERR(func->m2m_dev);
840 	}
841 
842 	vfd = video_device_alloc();
843 	if (!vfd) {
844 		v4l2_m2m_release(func->m2m_dev);
845 		dev_err(vpu->dev, "alloc vpu decoder video device fail\n");
846 		return -ENOMEM;
847 	}
848 	vfd->release = video_device_release;
849 	vfd->vfl_dir = VFL_DIR_M2M;
850 	vfd->v4l2_dev = &vpu->v4l2_dev;
851 	vfd->device_caps = V4L2_CAP_VIDEO_M2M_MPLANE | V4L2_CAP_STREAMING;
852 	if (func->type == VPU_CORE_TYPE_ENC) {
853 		strscpy(vfd->name, "amphion-vpu-encoder", sizeof(vfd->name));
854 		vfd->fops = venc_get_fops();
855 		vfd->ioctl_ops = venc_get_ioctl_ops();
856 	} else {
857 		strscpy(vfd->name, "amphion-vpu-decoder", sizeof(vfd->name));
858 		vfd->fops = vdec_get_fops();
859 		vfd->ioctl_ops = vdec_get_ioctl_ops();
860 	}
861 	video_set_drvdata(vfd, vpu);
862 
863 	ret = video_register_device(vfd, VFL_TYPE_VIDEO, -1);
864 	if (ret) {
865 		video_device_release(vfd);
866 		v4l2_m2m_release(func->m2m_dev);
867 		return ret;
868 	}
869 	func->vfd = vfd;
870 
871 	ret = v4l2_m2m_register_media_controller(func->m2m_dev, func->vfd, func->function);
872 	if (ret) {
873 		v4l2_m2m_release(func->m2m_dev);
874 		func->m2m_dev = NULL;
875 		video_unregister_device(func->vfd);
876 		func->vfd = NULL;
877 		return ret;
878 	}
879 
880 	return 0;
881 }
882 
883 void vpu_remove_func(struct vpu_func *func)
884 {
885 	if (!func)
886 		return;
887 
888 	if (func->m2m_dev) {
889 		v4l2_m2m_unregister_media_controller(func->m2m_dev);
890 		v4l2_m2m_release(func->m2m_dev);
891 		func->m2m_dev = NULL;
892 	}
893 	if (func->vfd) {
894 		video_unregister_device(func->vfd);
895 		func->vfd = NULL;
896 	}
897 }
898