2023-04-17 05:49:12

by Yunfei Dong (董云飞)

[permalink] [raw]
Subject: [PATCH 0/6] media: mediatek: vcodec: Fix decoder under flow and plt test fails randomly

1: Getting decoder under flow error randomly when do stress test with youtube;
2: Video display black when do plt test with one night.

patch 1 can't regard getting lat buffer as fail.
patch 2 ~ 3 using decoder status instead of core work count.
patch 4 ~ 6 using empty lat buffer as the last one to flush decoder.
---
Yunfei Dong (6):
media: mediatek: vcodec: can`t regard getting lat buffer fail as error
media: mediatek: vcodec: add the definition of decoder status
media: mediatek: vcodec: using decoder status instead of core work
count
media: mediatek: vcodec: add one empty lat buffer as the last one to
decode
media: mediatek: vcodec: move core context from device to each
instance
media: mediatek: vcodec: using empty lat buffer as the last one

.../mediatek/vcodec/mtk_vcodec_dec_drv.c | 1 -
.../vcodec/mtk_vcodec_dec_stateless.c | 2 +-
.../platform/mediatek/vcodec/mtk_vcodec_drv.h | 2 -
.../vcodec/vdec/vdec_h264_req_multi_if.c | 6 +-
.../vcodec/vdec/vdec_vp9_req_lat_if.c | 4 +-
.../platform/mediatek/vcodec/vdec_msg_queue.c | 105 ++++++++----------
.../platform/mediatek/vcodec/vdec_msg_queue.h | 23 +++-
7 files changed, 69 insertions(+), 74 deletions(-)

--
2.18.0


2023-04-17 05:49:55

by Yunfei Dong (董云飞)

[permalink] [raw]
Subject: [PATCH 2/6] media: mediatek: vcodec: add the definition of decoder status

Adding the status used to separate different decoder period for
core hardware.

Signed-off-by: Yunfei Dong <[email protected]>
---
drivers/media/platform/mediatek/vcodec/vdec_msg_queue.h | 7 +++++++
1 file changed, 7 insertions(+)

diff --git a/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.h b/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.h
index a5d44bc97c16..19508be08566 100644
--- a/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.h
+++ b/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.h
@@ -21,6 +21,13 @@ struct mtk_vcodec_ctx;
struct mtk_vcodec_dev;
typedef int (*core_decode_cb_t)(struct vdec_lat_buf *lat_buf);

+/* current context isn't work */
+#define CONTEXT_LIST_EMPTY (0)
+/* queued to the core work list */
+#define CONTEXT_LIST_QUEUED (1)
+/* context decode done */
+#define CONTEXT_LIST_DEC_DONE (2)
+
/**
* struct vdec_msg_queue_ctx - represents a queue for buffers ready to be processed
* @ready_to_use: ready used queue used to signalize when get a job queue
--
2.18.0

2023-04-17 05:49:57

by Yunfei Dong (董云飞)

[permalink] [raw]
Subject: [PATCH 3/6] media: mediatek: vcodec: using decoder status instead of core work count

core_work_cnt is the number of core work queued to work queue, the control
is very complex, leading to some unreasonable test result.

Using parameter status to indicate whether queue core work to work queue.

Fixes: 2e0ef56d81cb ("media: mediatek: vcodec: making sure queue_work successfully")
Signed-off-by: Yunfei Dong <[email protected]>
---
.../platform/mediatek/vcodec/vdec_msg_queue.c | 33 ++++++++-----------
.../platform/mediatek/vcodec/vdec_msg_queue.h | 4 +--
2 files changed, 16 insertions(+), 21 deletions(-)

diff --git a/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.c b/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.c
index f3073d1e7f42..03f8d7cd8edd 100644
--- a/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.c
+++ b/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.c
@@ -71,7 +71,6 @@ static void vdec_msg_queue_dec(struct vdec_msg_queue *msg_queue, int hardware_in
int vdec_msg_queue_qbuf(struct vdec_msg_queue_ctx *msg_ctx, struct vdec_lat_buf *buf)
{
struct list_head *head;
- int status;

head = vdec_get_buf_list(msg_ctx->hardware_index, buf);
if (!head) {
@@ -87,12 +86,9 @@ int vdec_msg_queue_qbuf(struct vdec_msg_queue_ctx *msg_ctx, struct vdec_lat_buf
if (msg_ctx->hardware_index != MTK_VDEC_CORE) {
wake_up_all(&msg_ctx->ready_to_use);
} else {
- if (buf->ctx->msg_queue.core_work_cnt <
- atomic_read(&buf->ctx->msg_queue.core_list_cnt)) {
- status = queue_work(buf->ctx->dev->core_workqueue,
- &buf->ctx->msg_queue.core_work);
- if (status)
- buf->ctx->msg_queue.core_work_cnt++;
+ if (!(buf->ctx->msg_queue.status & CONTEXT_LIST_QUEUED)) {
+ queue_work(buf->ctx->dev->core_workqueue, &buf->ctx->msg_queue.core_work);
+ buf->ctx->msg_queue.status |= CONTEXT_LIST_QUEUED;
}
}

@@ -261,7 +257,10 @@ static void vdec_msg_queue_core_work(struct work_struct *work)
container_of(msg_queue, struct mtk_vcodec_ctx, msg_queue);
struct mtk_vcodec_dev *dev = ctx->dev;
struct vdec_lat_buf *lat_buf;
- int status;
+
+ spin_lock(&ctx->dev->msg_queue_core_ctx.ready_lock);
+ ctx->msg_queue.status &= ~CONTEXT_LIST_QUEUED;
+ spin_unlock(&ctx->dev->msg_queue_core_ctx.ready_lock);

lat_buf = vdec_msg_queue_dqbuf(&dev->msg_queue_core_ctx);
if (!lat_buf)
@@ -278,17 +277,13 @@ static void vdec_msg_queue_core_work(struct work_struct *work)
vdec_msg_queue_qbuf(&ctx->msg_queue.lat_ctx, lat_buf);

wake_up_all(&ctx->msg_queue.core_dec_done);
- spin_lock(&dev->msg_queue_core_ctx.ready_lock);
- lat_buf->ctx->msg_queue.core_work_cnt--;
-
- if (lat_buf->ctx->msg_queue.core_work_cnt <
- atomic_read(&lat_buf->ctx->msg_queue.core_list_cnt)) {
- status = queue_work(lat_buf->ctx->dev->core_workqueue,
- &lat_buf->ctx->msg_queue.core_work);
- if (status)
- lat_buf->ctx->msg_queue.core_work_cnt++;
+ if (!(ctx->msg_queue.status & CONTEXT_LIST_QUEUED) &&
+ atomic_read(&msg_queue->core_list_cnt)) {
+ spin_lock(&ctx->dev->msg_queue_core_ctx.ready_lock);
+ ctx->msg_queue.status |= CONTEXT_LIST_QUEUED;
+ spin_unlock(&ctx->dev->msg_queue_core_ctx.ready_lock);
+ queue_work(ctx->dev->core_workqueue, &msg_queue->core_work);
}
- spin_unlock(&dev->msg_queue_core_ctx.ready_lock);
}

int vdec_msg_queue_init(struct vdec_msg_queue *msg_queue,
@@ -303,13 +298,13 @@ int vdec_msg_queue_init(struct vdec_msg_queue *msg_queue,
return 0;

msg_queue->ctx = ctx;
- msg_queue->core_work_cnt = 0;
vdec_msg_queue_init_ctx(&msg_queue->lat_ctx, MTK_VDEC_LAT0);
INIT_WORK(&msg_queue->core_work, vdec_msg_queue_core_work);

atomic_set(&msg_queue->lat_list_cnt, 0);
atomic_set(&msg_queue->core_list_cnt, 0);
init_waitqueue_head(&msg_queue->core_dec_done);
+ msg_queue->status = CONTEXT_LIST_EMPTY;

msg_queue->wdma_addr.size =
vde_msg_queue_get_trans_size(ctx->picinfo.buf_w,
diff --git a/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.h b/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.h
index 19508be08566..a21ee3aece5c 100644
--- a/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.h
+++ b/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.h
@@ -84,7 +84,7 @@ struct vdec_lat_buf {
* @lat_list_cnt: used to record each instance lat list count
* @core_list_cnt: used to record each instance core list count
* @core_dec_done: core work queue decode done event
- * @core_work_cnt: the number of core work in work queue
+ * @status: current context decode status for core hardware
*/
struct vdec_msg_queue {
struct vdec_lat_buf lat_buf[NUM_BUFFER_COUNT];
@@ -100,7 +100,7 @@ struct vdec_msg_queue {
atomic_t lat_list_cnt;
atomic_t core_list_cnt;
wait_queue_head_t core_dec_done;
- int core_work_cnt;
+ int status;
};

/**
--
2.18.0

2023-04-17 05:50:07

by Yunfei Dong (董云飞)

[permalink] [raw]
Subject: [PATCH 5/6] media: mediatek: vcodec: move core context from device to each instance

There are so many lat buffer in core context list, some instances
maybe be scheduled for a very long time. Moving the core context to
each instance, it only be used to control lat buffer of each instance.
And the core work queue of each instance is scheduled by system.

Fixes: 2cfca6c1bf80 ("media: mediatek: vcodec: move lat_buf to the top of core list")
Signed-off-by: Yunfei Dong <[email protected]>
---
.../mediatek/vcodec/mtk_vcodec_dec_drv.c | 1 -
.../platform/mediatek/vcodec/mtk_vcodec_drv.h | 2 -
.../vcodec/vdec/vdec_h264_req_multi_if.c | 4 +-
.../vcodec/vdec/vdec_vp9_req_lat_if.c | 2 +-
.../platform/mediatek/vcodec/vdec_msg_queue.c | 53 +++++++------------
.../platform/mediatek/vcodec/vdec_msg_queue.h | 6 ++-
6 files changed, 25 insertions(+), 43 deletions(-)

diff --git a/drivers/media/platform/mediatek/vcodec/mtk_vcodec_dec_drv.c b/drivers/media/platform/mediatek/vcodec/mtk_vcodec_dec_drv.c
index 3755acc61d7d..d2543fb27ba7 100644
--- a/drivers/media/platform/mediatek/vcodec/mtk_vcodec_dec_drv.c
+++ b/drivers/media/platform/mediatek/vcodec/mtk_vcodec_dec_drv.c
@@ -310,7 +310,6 @@ static int mtk_vcodec_probe(struct platform_device *pdev)
}

if (IS_VDEC_LAT_ARCH(dev->vdec_pdata->hw_arch)) {
- vdec_msg_queue_init_ctx(&dev->msg_queue_core_ctx, MTK_VDEC_CORE);
dev->core_workqueue =
alloc_ordered_workqueue("core-decoder",
WQ_MEM_RECLAIM | WQ_FREEZABLE);
diff --git a/drivers/media/platform/mediatek/vcodec/mtk_vcodec_drv.h b/drivers/media/platform/mediatek/vcodec/mtk_vcodec_drv.h
index 9acab54fd650..15d2cb171b89 100644
--- a/drivers/media/platform/mediatek/vcodec/mtk_vcodec_drv.h
+++ b/drivers/media/platform/mediatek/vcodec/mtk_vcodec_drv.h
@@ -461,7 +461,6 @@ struct mtk_vcodec_enc_pdata {
* @enc_capability: used to identify encode capability
*
* @core_workqueue: queue used for core hardware decode
- * @msg_queue_core_ctx: msg queue context used for core workqueue
*
* @subdev_dev: subdev hardware device
* @subdev_prob_done: check whether all used hw device is prob done
@@ -510,7 +509,6 @@ struct mtk_vcodec_dev {
unsigned int enc_capability;

struct workqueue_struct *core_workqueue;
- struct vdec_msg_queue_ctx msg_queue_core_ctx;

void *subdev_dev[MTK_VDEC_HW_MAX];
int (*subdev_prob_done)(struct mtk_vcodec_dev *vdec_dev);
diff --git a/drivers/media/platform/mediatek/vcodec/vdec/vdec_h264_req_multi_if.c b/drivers/media/platform/mediatek/vcodec/vdec/vdec_h264_req_multi_if.c
index 181cc52e0847..a7e8e3257b7f 100644
--- a/drivers/media/platform/mediatek/vcodec/vdec/vdec_h264_req_multi_if.c
+++ b/drivers/media/platform/mediatek/vcodec/vdec/vdec_h264_req_multi_if.c
@@ -672,7 +672,7 @@ static int vdec_h264_slice_lat_decode(void *h_vdec, struct mtk_vcodec_mem *bs,
if (IS_VDEC_INNER_RACING(inst->ctx->dev->dec_capability)) {
memcpy(&share_info->h264_slice_params, &inst->vsi->h264_slice_params,
sizeof(share_info->h264_slice_params));
- vdec_msg_queue_qbuf(&inst->ctx->dev->msg_queue_core_ctx, lat_buf);
+ vdec_msg_queue_qbuf(&inst->ctx->msg_queue.core_ctx, lat_buf);
}

/* wait decoder done interrupt */
@@ -698,7 +698,7 @@ static int vdec_h264_slice_lat_decode(void *h_vdec, struct mtk_vcodec_mem *bs,
if (!IS_VDEC_INNER_RACING(inst->ctx->dev->dec_capability)) {
memcpy(&share_info->h264_slice_params, &inst->vsi->h264_slice_params,
sizeof(share_info->h264_slice_params));
- vdec_msg_queue_qbuf(&inst->ctx->dev->msg_queue_core_ctx, lat_buf);
+ vdec_msg_queue_qbuf(&inst->ctx->msg_queue.core_ctx, lat_buf);
}
mtk_vcodec_debug(inst, "dec num: %d lat crc: 0x%x 0x%x 0x%x", inst->slice_dec_num,
inst->vsi->dec.crc[0], inst->vsi->dec.crc[1], inst->vsi->dec.crc[2]);
diff --git a/drivers/media/platform/mediatek/vcodec/vdec/vdec_vp9_req_lat_if.c b/drivers/media/platform/mediatek/vcodec/vdec/vdec_vp9_req_lat_if.c
index 6d981d7341d2..c2f90848f498 100644
--- a/drivers/media/platform/mediatek/vcodec/vdec/vdec_vp9_req_lat_if.c
+++ b/drivers/media/platform/mediatek/vcodec/vdec/vdec_vp9_req_lat_if.c
@@ -2119,7 +2119,7 @@ static int vdec_vp9_slice_lat_decode(void *h_vdec, struct mtk_vcodec_mem *bs,
vdec_msg_queue_update_ube_wptr(&ctx->msg_queue,
vsi->trans.dma_addr_end +
ctx->msg_queue.wdma_addr.dma_addr);
- vdec_msg_queue_qbuf(&ctx->dev->msg_queue_core_ctx, lat_buf);
+ vdec_msg_queue_qbuf(&ctx->msg_queue.core_ctx, lat_buf);

return 0;
err_free_fb_out:
diff --git a/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.c b/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.c
index ef141f9d7bb1..a74bd839d7e2 100644
--- a/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.c
+++ b/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.c
@@ -177,26 +177,17 @@ void vdec_msg_queue_update_ube_wptr(struct vdec_msg_queue *msg_queue, uint64_t u

bool vdec_msg_queue_wait_lat_buf_full(struct vdec_msg_queue *msg_queue)
{
- struct vdec_lat_buf *buf, *tmp;
- struct list_head *list_core[3];
- struct vdec_msg_queue_ctx *core_ctx;
- int ret, i, in_core_count = 0, count = 0;
+ int ret;
long timeout_jiff;

- core_ctx = &msg_queue->ctx->dev->msg_queue_core_ctx;
- spin_lock(&core_ctx->ready_lock);
- list_for_each_entry_safe(buf, tmp, &core_ctx->ready_queue, core_list) {
- if (buf && buf->ctx == msg_queue->ctx) {
- list_core[in_core_count++] = &buf->core_list;
- list_del(&buf->core_list);
- }
- }
-
- for (i = 0; i < in_core_count; i++) {
- list_add(list_core[in_core_count - (1 + i)], &core_ctx->ready_queue);
- queue_work(msg_queue->ctx->dev->core_workqueue, &msg_queue->core_work);
+ if (atomic_read(&msg_queue->lat_list_cnt) == NUM_BUFFER_COUNT) {
+ mtk_v4l2_debug(3, "wait buf full: list(%d %d) ready_num:%d status:%d",
+ atomic_read(&msg_queue->lat_list_cnt),
+ atomic_read(&msg_queue->core_list_cnt),
+ msg_queue->lat_ctx.ready_num,
+ msg_queue->status);
+ return true;
}
- spin_unlock(&core_ctx->ready_lock);

timeout_jiff = msecs_to_jiffies(1000 * (NUM_BUFFER_COUNT + 2));
ret = wait_event_timeout(msg_queue->ctx->msg_queue.core_dec_done,
@@ -208,18 +199,9 @@ bool vdec_msg_queue_wait_lat_buf_full(struct vdec_msg_queue *msg_queue)
return true;
}

- spin_lock(&core_ctx->ready_lock);
- list_for_each_entry_safe(buf, tmp, &core_ctx->ready_queue, core_list) {
- if (buf && buf->ctx == msg_queue->ctx) {
- count++;
- list_del(&buf->core_list);
- }
- }
- spin_unlock(&core_ctx->ready_lock);
-
- mtk_v4l2_err("failed with lat buf isn't full: list(%d %d) count:%d",
+ mtk_v4l2_err("failed with lat buf isn't full: list(%d %d)",
atomic_read(&msg_queue->lat_list_cnt),
- atomic_read(&msg_queue->core_list_cnt), count);
+ atomic_read(&msg_queue->core_list_cnt));

return false;
}
@@ -247,6 +229,8 @@ void vdec_msg_queue_deinit(struct vdec_msg_queue *msg_queue,

kfree(lat_buf->private_data);
}
+
+ cancel_work_sync(&msg_queue->core_work);
}

static void vdec_msg_queue_core_work(struct work_struct *work)
@@ -258,11 +242,11 @@ static void vdec_msg_queue_core_work(struct work_struct *work)
struct mtk_vcodec_dev *dev = ctx->dev;
struct vdec_lat_buf *lat_buf;

- spin_lock(&ctx->dev->msg_queue_core_ctx.ready_lock);
+ spin_lock(&msg_queue->core_ctx.ready_lock);
ctx->msg_queue.status &= ~CONTEXT_LIST_QUEUED;
- spin_unlock(&ctx->dev->msg_queue_core_ctx.ready_lock);
+ spin_unlock(&msg_queue->core_ctx.ready_lock);

- lat_buf = vdec_msg_queue_dqbuf(&dev->msg_queue_core_ctx);
+ lat_buf = vdec_msg_queue_dqbuf(&msg_queue->core_ctx);
if (!lat_buf)
return;

@@ -276,12 +260,11 @@ static void vdec_msg_queue_core_work(struct work_struct *work)
mtk_vcodec_dec_disable_hardware(ctx, MTK_VDEC_CORE);
vdec_msg_queue_qbuf(&ctx->msg_queue.lat_ctx, lat_buf);

- wake_up_all(&ctx->msg_queue.core_dec_done);
if (!(ctx->msg_queue.status & CONTEXT_LIST_QUEUED) &&
atomic_read(&msg_queue->core_list_cnt)) {
- spin_lock(&ctx->dev->msg_queue_core_ctx.ready_lock);
+ spin_lock(&msg_queue->core_ctx.ready_lock);
ctx->msg_queue.status |= CONTEXT_LIST_QUEUED;
- spin_unlock(&ctx->dev->msg_queue_core_ctx.ready_lock);
+ spin_unlock(&msg_queue->core_ctx.ready_lock);
queue_work(ctx->dev->core_workqueue, &msg_queue->core_work);
}
}
@@ -297,8 +280,8 @@ int vdec_msg_queue_init(struct vdec_msg_queue *msg_queue,
if (msg_queue->wdma_addr.size)
return 0;

- msg_queue->ctx = ctx;
vdec_msg_queue_init_ctx(&msg_queue->lat_ctx, MTK_VDEC_LAT0);
+ vdec_msg_queue_init_ctx(&msg_queue->core_ctx, MTK_VDEC_CORE);
INIT_WORK(&msg_queue->core_work, vdec_msg_queue_core_work);

atomic_set(&msg_queue->lat_list_cnt, 0);
diff --git a/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.h b/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.h
index a80b9853cec9..ae37d020a1bd 100644
--- a/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.h
+++ b/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.h
@@ -83,10 +83,11 @@ struct vdec_lat_buf {
* @wdma_wptr_addr: ube write point
* @core_work: core hardware work
* @lat_ctx: used to store lat buffer list
- * @ctx: point to mtk_vcodec_ctx
+ * @core_ctx: used to store core buffer list
*
* @lat_list_cnt: used to record each instance lat list count
* @core_list_cnt: used to record each instance core list count
+ * @flush_done: core flush done status
* @empty_lat_buf: the last lat buf used to flush decode
* @core_dec_done: core work queue decode done event
* @status: current context decode status for core hardware
@@ -100,10 +101,11 @@ struct vdec_msg_queue {

struct work_struct core_work;
struct vdec_msg_queue_ctx lat_ctx;
- struct mtk_vcodec_ctx *ctx;
+ struct vdec_msg_queue_ctx core_ctx;

atomic_t lat_list_cnt;
atomic_t core_list_cnt;
+ bool flush_done;
struct vdec_lat_buf empty_lat_buf;
wait_queue_head_t core_dec_done;
int status;
--
2.18.0

2023-04-17 05:50:12

by Yunfei Dong (董云飞)

[permalink] [raw]
Subject: [PATCH 4/6] media: mediatek: vcodec: add one empty lat buffer as the last one to decode

Adding one empty lat buffer with the parameter 'is_empty_flag = true' used
to flush core work queue decode.

Signed-off-by: Yunfei Dong <[email protected]>
---
drivers/media/platform/mediatek/vcodec/vdec_msg_queue.c | 5 +++++
drivers/media/platform/mediatek/vcodec/vdec_msg_queue.h | 6 ++++++
2 files changed, 11 insertions(+)

diff --git a/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.c b/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.c
index 03f8d7cd8edd..ef141f9d7bb1 100644
--- a/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.c
+++ b/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.c
@@ -317,6 +317,10 @@ int vdec_msg_queue_init(struct vdec_msg_queue *msg_queue,
msg_queue->wdma_rptr_addr = msg_queue->wdma_addr.dma_addr;
msg_queue->wdma_wptr_addr = msg_queue->wdma_addr.dma_addr;

+ msg_queue->empty_lat_buf.ctx = ctx;
+ msg_queue->empty_lat_buf.core_decode = NULL;
+ msg_queue->empty_lat_buf.is_last_frame = true;
+
for (i = 0; i < NUM_BUFFER_COUNT; i++) {
lat_buf = &msg_queue->lat_buf[i];

@@ -342,6 +346,7 @@ int vdec_msg_queue_init(struct vdec_msg_queue *msg_queue,

lat_buf->ctx = ctx;
lat_buf->core_decode = core_decode;
+ lat_buf->is_last_frame = false;
err = vdec_msg_queue_qbuf(&msg_queue->lat_ctx, lat_buf);
if (err) {
mtk_v4l2_err("failed to qbuf buf[%d]", i);
diff --git a/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.h b/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.h
index a21ee3aece5c..a80b9853cec9 100644
--- a/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.h
+++ b/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.h
@@ -57,6 +57,8 @@ struct vdec_msg_queue_ctx {
* @core_decode: different codec use different decode callback function
* @lat_list: add lat buffer to lat head list
* @core_list: add lat buffer to core head list
+ *
+ * @is_last_frame: meaning this buffer is the last frame
*/
struct vdec_lat_buf {
struct mtk_vcodec_mem wdma_err_addr;
@@ -69,6 +71,8 @@ struct vdec_lat_buf {
core_decode_cb_t core_decode;
struct list_head lat_list;
struct list_head core_list;
+
+ bool is_last_frame;
};

/**
@@ -83,6 +87,7 @@ struct vdec_lat_buf {
*
* @lat_list_cnt: used to record each instance lat list count
* @core_list_cnt: used to record each instance core list count
+ * @empty_lat_buf: the last lat buf used to flush decode
* @core_dec_done: core work queue decode done event
* @status: current context decode status for core hardware
*/
@@ -99,6 +104,7 @@ struct vdec_msg_queue {

atomic_t lat_list_cnt;
atomic_t core_list_cnt;
+ struct vdec_lat_buf empty_lat_buf;
wait_queue_head_t core_dec_done;
int status;
};
--
2.18.0

2023-04-17 05:50:19

by Yunfei Dong (董云飞)

[permalink] [raw]
Subject: [PATCH 6/6] media: mediatek: vcodec: using empty lat buffer as the last one

Queue the empty lat buffer to core list when driver need to flush decode.
It's mean core already decode all existed lat buffer when get empty lat
buffer, then wake up core decode done event, the driver will exit when getting
core dec done event.

Fixes: d227af847ac2 ("media: mediatek: vcodec: add core decode done event")
Signed-off-by: Yunfei Dong <[email protected]>
---
.../platform/mediatek/vcodec/vdec_msg_queue.c | 30 +++++++++----------
1 file changed, 15 insertions(+), 15 deletions(-)

diff --git a/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.c b/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.c
index a74bd839d7e2..d2e48eefc958 100644
--- a/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.c
+++ b/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.c
@@ -177,9 +177,6 @@ void vdec_msg_queue_update_ube_wptr(struct vdec_msg_queue *msg_queue, uint64_t u

bool vdec_msg_queue_wait_lat_buf_full(struct vdec_msg_queue *msg_queue)
{
- int ret;
- long timeout_jiff;
-
if (atomic_read(&msg_queue->lat_list_cnt) == NUM_BUFFER_COUNT) {
mtk_v4l2_debug(3, "wait buf full: list(%d %d) ready_num:%d status:%d",
atomic_read(&msg_queue->lat_list_cnt),
@@ -189,19 +186,14 @@ bool vdec_msg_queue_wait_lat_buf_full(struct vdec_msg_queue *msg_queue)
return true;
}

- timeout_jiff = msecs_to_jiffies(1000 * (NUM_BUFFER_COUNT + 2));
- ret = wait_event_timeout(msg_queue->ctx->msg_queue.core_dec_done,
- msg_queue->lat_ctx.ready_num == NUM_BUFFER_COUNT,
- timeout_jiff);
- if (ret) {
- mtk_v4l2_debug(3, "success to get lat buf: %d",
- msg_queue->lat_ctx.ready_num);
- return true;
- }
+ msg_queue->flush_done = false;
+ vdec_msg_queue_qbuf(&msg_queue->core_ctx, &msg_queue->empty_lat_buf);
+ wait_event(msg_queue->core_dec_done, msg_queue->flush_done);

- mtk_v4l2_err("failed with lat buf isn't full: list(%d %d)",
- atomic_read(&msg_queue->lat_list_cnt),
- atomic_read(&msg_queue->core_list_cnt));
+ mtk_v4l2_debug("flush done => ready_num:%d status:%d list(%d %d)",
+ msg_queue->lat_ctx.ready_num, msg_queue->status,
+ atomic_read(&msg_queue->lat_list_cnt),
+ atomic_read(&msg_queue->core_list_cnt));

return false;
}
@@ -250,6 +242,14 @@ static void vdec_msg_queue_core_work(struct work_struct *work)
if (!lat_buf)
return;

+ if (lat_buf->is_last_frame) {
+ ctx->msg_queue.status = CONTEXT_LIST_DEC_DONE;
+ msg_queue->flush_done = true;
+ wake_up(&ctx->msg_queue.core_dec_done);
+
+ return;
+ }
+
ctx = lat_buf->ctx;
mtk_vcodec_dec_enable_hardware(ctx, MTK_VDEC_CORE);
mtk_vcodec_set_curr_ctx(dev, ctx, MTK_VDEC_CORE);
--
2.18.0

Subject: Re: [PATCH 2/6] media: mediatek: vcodec: add the definition of decoder status

Il 17/04/23 07:48, Yunfei Dong ha scritto:
> Adding the status used to separate different decoder period for
> core hardware.
>
> Signed-off-by: Yunfei Dong <[email protected]>
> ---
> drivers/media/platform/mediatek/vcodec/vdec_msg_queue.h | 7 +++++++
> 1 file changed, 7 insertions(+)
>
> diff --git a/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.h b/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.h
> index a5d44bc97c16..19508be08566 100644
> --- a/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.h
> +++ b/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.h
> @@ -21,6 +21,13 @@ struct mtk_vcodec_ctx;
> struct mtk_vcodec_dev;
> typedef int (*core_decode_cb_t)(struct vdec_lat_buf *lat_buf);
>
> +/* current context isn't work */
> +#define CONTEXT_LIST_EMPTY (0)
> +/* queued to the core work list */
> +#define CONTEXT_LIST_QUEUED (1)
> +/* context decode done */
> +#define CONTEXT_LIST_DEC_DONE (2)

I would prefer an enumeration instead; you can keep the documentation for those
status signals with kerneldoc on the new enum.

/**
* enum core_ctx_status - Context decode status for core hardware
* @CONTEXT_LIST_EMPTY: No buffer queued on Core HW (must always be 0)
* @CONTEXT_LIST_QUEUED: Buffer queued to the core work list
* @CONTEXT_LIST_DEC_DONE: Context decode done
*/
enum core_ctx_status {
CONTEXT_LIST_EMPTY = 0,
CONTEXT_LIST_QUEUED,
CONTEXT_LIST_DEC_DONE,
}

Moreover, since this is a rather simple addition, please squash this commit with
patch [3/6] where you actually also introduce the actual usage of the new enum.

Cheers,
Angelo

Subject: Re: [PATCH 5/6] media: mediatek: vcodec: move core context from device to each instance

Il 17/04/23 07:48, Yunfei Dong ha scritto:
> There are so many lat buffer in core context list, some instances
> maybe be scheduled for a very long time. Moving the core context to
> each instance, it only be used to control lat buffer of each instance.
> And the core work queue of each instance is scheduled by system.
>
> Fixes: 2cfca6c1bf80 ("media: mediatek: vcodec: move lat_buf to the top of core list")
> Signed-off-by: Yunfei Dong <[email protected]>
> ---
> .../mediatek/vcodec/mtk_vcodec_dec_drv.c | 1 -
> .../platform/mediatek/vcodec/mtk_vcodec_drv.h | 2 -
> .../vcodec/vdec/vdec_h264_req_multi_if.c | 4 +-
> .../vcodec/vdec/vdec_vp9_req_lat_if.c | 2 +-
> .../platform/mediatek/vcodec/vdec_msg_queue.c | 53 +++++++------------
> .../platform/mediatek/vcodec/vdec_msg_queue.h | 6 ++-
> 6 files changed, 25 insertions(+), 43 deletions(-)
>

..snip..

> diff --git a/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.h b/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.h
> index a80b9853cec9..ae37d020a1bd 100644
> --- a/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.h
> +++ b/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.h
> @@ -83,10 +83,11 @@ struct vdec_lat_buf {
> * @wdma_wptr_addr: ube write point
> * @core_work: core hardware work
> * @lat_ctx: used to store lat buffer list
> - * @ctx: point to mtk_vcodec_ctx
> + * @core_ctx: used to store core buffer list
> *
> * @lat_list_cnt: used to record each instance lat list count
> * @core_list_cnt: used to record each instance core list count
> + * @flush_done: core flush done status
> * @empty_lat_buf: the last lat buf used to flush decode
> * @core_dec_done: core work queue decode done event
> * @status: current context decode status for core hardware
> @@ -100,10 +101,11 @@ struct vdec_msg_queue {
>
> struct work_struct core_work;
> struct vdec_msg_queue_ctx lat_ctx;
> - struct mtk_vcodec_ctx *ctx;
> + struct vdec_msg_queue_ctx core_ctx;
>
> atomic_t lat_list_cnt;
> atomic_t core_list_cnt;
> + bool flush_done;

flush_done is used in patch [6/6]: this does not belong to this patch,
please move the addition of this member in the same patch where you use it.

Regards,
Angelo

Subject: Re: [PATCH 4/6] media: mediatek: vcodec: add one empty lat buffer as the last one to decode

Il 17/04/23 07:48, Yunfei Dong ha scritto:
> Adding one empty lat buffer with the parameter 'is_empty_flag = true' used
> to flush core work queue decode.
>
> Signed-off-by: Yunfei Dong <[email protected]>

Since commit [6/6] depends on this one, you should either squash this with [6/6]
or add the same Fixes tag to this.

I think that the most sensible option is to squash it.


Regards,
Angelo

2023-04-17 12:14:10

by Yunfei Dong (董云飞)

[permalink] [raw]
Subject: Re: [PATCH 2/6] media: mediatek: vcodec: add the definition of decoder status

Hi AngeloGioacchino,

Thanks for your suggestion.

On Mon, 2023-04-17 at 11:26 +0200, AngeloGioacchino Del Regno wrote:
> External email : Please do not click links or open attachments until
> you have verified the sender or the content.
>
>
> Il 17/04/23 07:48, Yunfei Dong ha scritto:
> > Adding the status used to separate different decoder period for
> > core hardware.
> >
> > Signed-off-by: Yunfei Dong <[email protected]>
> > ---
> > drivers/media/platform/mediatek/vcodec/vdec_msg_queue.h | 7
> > +++++++
> > 1 file changed, 7 insertions(+)
> >
> > diff --git
> > a/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.h
> > b/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.h
> > index a5d44bc97c16..19508be08566 100644
> > --- a/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.h
> > +++ b/drivers/media/platform/mediatek/vcodec/vdec_msg_queue.h
> > @@ -21,6 +21,13 @@ struct mtk_vcodec_ctx;
> > struct mtk_vcodec_dev;
> > typedef int (*core_decode_cb_t)(struct vdec_lat_buf *lat_buf);
> >
> > +/* current context isn't work */
> > +#define CONTEXT_LIST_EMPTY (0)
> > +/* queued to the core work list */
> > +#define CONTEXT_LIST_QUEUED (1)
> > +/* context decode done */
> > +#define CONTEXT_LIST_DEC_DONE (2)
>
> I would prefer an enumeration instead; you can keep the documentation
> for those
> status signals with kerneldoc on the new enum.
>
> /**
> * enum core_ctx_status - Context decode status for core hardware
> * @CONTEXT_LIST_EMPTY: No buffer queued on Core HW (must always
> be 0)
> * @CONTEXT_LIST_QUEUED: Buffer queued to the core work list
> * @CONTEXT_LIST_DEC_DONE: Context decode done
> */
> enum core_ctx_status {
> CONTEXT_LIST_EMPTY = 0,
> CONTEXT_LIST_QUEUED,
> CONTEXT_LIST_DEC_DONE,
> }
>
> Moreover, since this is a rather simple addition, please squash this
> commit with
> patch [3/6] where you actually also introduce the actual usage of the
> new enum.
>
Fixed in patch v2.

Best Regards,
Yunfei Dong
> Cheers,
> Angelo
>

2023-04-17 12:15:11

by Yunfei Dong (董云飞)

[permalink] [raw]
Subject: Re: [PATCH 4/6] media: mediatek: vcodec: add one empty lat buffer as the last one to decode

Hi AngeloGioacchino,

Thanks for your suggestion.
On Mon, 2023-04-17 at 11:35 +0200, AngeloGioacchino Del Regno wrote:
> External email : Please do not click links or open attachments until
> you have verified the sender or the content.
>
>
> Il 17/04/23 07:48, Yunfei Dong ha scritto:
> > Adding one empty lat buffer with the parameter 'is_empty_flag =
> > true' used
> > to flush core work queue decode.
> >
> > Signed-off-by: Yunfei Dong <[email protected]>
>
> Since commit [6/6] depends on this one, you should either squash this
> with [6/6]
> or add the same Fixes tag to this.
>
> I think that the most sensible option is to squash it.
>
Fixed in patch v2.

Best Regards,
Yunfei Dong
>
> Regards,
> Angelo