diff mbox series

[v1,3/5] media: venus: do not destroy video session during queue setup

Message ID 1538222432-25894-4-git-send-email-sgorle@codeaurora.org (mailing list archive)
State New, archived
Headers show
Series Venus - Decode reconfig sequence | expand

Commit Message

Srinu Gorle Sept. 29, 2018, noon UTC
- open and close video sessions for plane properties is incorrect.
- add check to ensure, same instance persist from driver open to close.

Signed-off-by: Srinu Gorle <sgorle@codeaurora.org>
---
 drivers/media/platform/qcom/venus/hfi.c  | 3 +++
 drivers/media/platform/qcom/venus/vdec.c | 2 ++
 2 files changed, 5 insertions(+)

Comments

Stanimir Varbanov Nov. 9, 2018, 9:59 a.m. UTC | #1
Hi Srinu,

On 9/29/18 3:00 PM, Srinu Gorle wrote:
> - open and close video sessions for plane properties is incorrect.

Could you rephrase this statement? I really don't understand what you mean.

> - add check to ensure, same instance persist from driver open to close.

This assumption is wrong. The v4l client can change the codec by SFMT
without close the device node, in that case we have to destroy and
create a new session with new codec.

> 
> Signed-off-by: Srinu Gorle <sgorle@codeaurora.org>
> ---
>  drivers/media/platform/qcom/venus/hfi.c  | 3 +++
>  drivers/media/platform/qcom/venus/vdec.c | 2 ++
>  2 files changed, 5 insertions(+)
> 
> diff --git a/drivers/media/platform/qcom/venus/hfi.c b/drivers/media/platform/qcom/venus/hfi.c
> index 36a4784..59c34ba 100644
> --- a/drivers/media/platform/qcom/venus/hfi.c
> +++ b/drivers/media/platform/qcom/venus/hfi.c
> @@ -207,6 +207,9 @@ int hfi_session_init(struct venus_inst *inst, u32 pixfmt)
>  	const struct hfi_ops *ops = core->ops;
>  	int ret;
>  
> +	if (inst->state >= INST_INIT && inst->state < INST_STOP)
> +		return 0;

In fact you want to be able to call session_init multiple times but
deinit the session only once? The hfi.c layer is designed to follow the
states as they are expected by the firmware side, if you want to call
session_init multiple times just make a wrapper in the vdec.c with
reference counting.

> +
>  	inst->hfi_codec = to_codec_type(pixfmt);
>  	reinit_completion(&inst->done);
>  
> diff --git a/drivers/media/platform/qcom/venus/vdec.c b/drivers/media/platform/qcom/venus/vdec.c
> index afe3b36..0035cf2 100644
> --- a/drivers/media/platform/qcom/venus/vdec.c
> +++ b/drivers/media/platform/qcom/venus/vdec.c
> @@ -700,6 +700,8 @@ static int vdec_num_buffers(struct venus_inst *inst, unsigned int *in_num,
>  
>  	*out_num = HFI_BUFREQ_COUNT_MIN(&bufreq, ver);
>  
> +	return 0;

OK in present implementation I decided that the codec is settled when
streamon on both queues is called (i.e. the final session_init is made
in streamon). IMO the correct one is to init the session in
reqbuf(output) and deinit session in reqbuf(output, count=0)?

The problem I see when you skip session_deinit is that the codec cannot
be changed without closing the video node.

> +
>  deinit:
>  	hfi_session_deinit(inst);
>  
>
Tomasz Figa Nov. 14, 2018, 11:02 a.m. UTC | #2
On Fri, Nov 9, 2018 at 7:00 PM Stanimir Varbanov
<stanimir.varbanov@linaro.org> wrote:
>
> Hi Srinu,
>
> On 9/29/18 3:00 PM, Srinu Gorle wrote:
> > - open and close video sessions for plane properties is incorrect.
>
> Could you rephrase this statement? I really don't understand what you mean.
>
> > - add check to ensure, same instance persist from driver open to close.
>
> This assumption is wrong. The v4l client can change the codec by SFMT
> without close the device node, in that case we have to destroy and
> create a new session with new codec.

Right.

[snip]
> > +
> >       inst->hfi_codec = to_codec_type(pixfmt);
> >       reinit_completion(&inst->done);
> >
> > diff --git a/drivers/media/platform/qcom/venus/vdec.c b/drivers/media/platform/qcom/venus/vdec.c
> > index afe3b36..0035cf2 100644
> > --- a/drivers/media/platform/qcom/venus/vdec.c
> > +++ b/drivers/media/platform/qcom/venus/vdec.c
> > @@ -700,6 +700,8 @@ static int vdec_num_buffers(struct venus_inst *inst, unsigned int *in_num,
> >
> >       *out_num = HFI_BUFREQ_COUNT_MIN(&bufreq, ver);
> >
> > +     return 0;
>
> OK in present implementation I decided that the codec is settled when
> streamon on both queues is called (i.e. the final session_init is made
> in streamon). IMO the correct one is to init the session in
> reqbuf(output) and deinit session in reqbuf(output, count=0)?
>

Depending on what you mean by "settled" that could be fine.

Generally for initialization, one would typically use REQBUFS(OUTPUT,
>0) (since it's not possible to change the format after the
allocation) or STREAMON(OUTPUT) (to defer the hardware operations
until really necessary).

Since STREAMOFF(OUTPUT) is expected to just trigger a seek,
termination should be done in REQBUFS(OUTPUT, 0) indeed and after that
one should be able to switch to another codec (OUTPUT format).

Best regards,
Tomasz
diff mbox series

Patch

diff --git a/drivers/media/platform/qcom/venus/hfi.c b/drivers/media/platform/qcom/venus/hfi.c
index 36a4784..59c34ba 100644
--- a/drivers/media/platform/qcom/venus/hfi.c
+++ b/drivers/media/platform/qcom/venus/hfi.c
@@ -207,6 +207,9 @@  int hfi_session_init(struct venus_inst *inst, u32 pixfmt)
 	const struct hfi_ops *ops = core->ops;
 	int ret;
 
+	if (inst->state >= INST_INIT && inst->state < INST_STOP)
+		return 0;
+
 	inst->hfi_codec = to_codec_type(pixfmt);
 	reinit_completion(&inst->done);
 
diff --git a/drivers/media/platform/qcom/venus/vdec.c b/drivers/media/platform/qcom/venus/vdec.c
index afe3b36..0035cf2 100644
--- a/drivers/media/platform/qcom/venus/vdec.c
+++ b/drivers/media/platform/qcom/venus/vdec.c
@@ -700,6 +700,8 @@  static int vdec_num_buffers(struct venus_inst *inst, unsigned int *in_num,
 
 	*out_num = HFI_BUFREQ_COUNT_MIN(&bufreq, ver);
 
+	return 0;
+
 deinit:
 	hfi_session_deinit(inst);