Re: [PATCH 2/2] media: cedrus: Allow using the current dst buffer as reference

From: Hans Verkuil
Date: Wed Jan 09 2019 - 10:40:36 EST


On 01/09/19 15:42, Paul Kocialkowski wrote:
> Hi,
>
> On Wed, 2019-01-09 at 15:29 +0100, Hans Verkuil wrote:
>> On 01/09/19 15:19, Paul Kocialkowski wrote:
>>> It was reported that some cases of interleaved video decoding require
>>> using the current destination buffer as a reference. However, this is
>>> no longer possible after the move to vb2_find_timestamp because only
>>> dequeued and done buffers are considered.
>>>
>>> Add a helper in our driver that also considers the current destination
>>> buffer before resorting to vb2_find_timestamp and use it in MPEG-2.
>>
>> This patch looks good, but can you also add checks to handle the case
>> when no buffer with the given timestamp was found? Probably should be done
>> in a third patch.
>>
>> I suspect the driver will crash if an unknown timestamp is passed on to the
>> driver. I would really like to see that corner case fixed.
>
> You're totally right and I think that more generarlly, the current code
> flow is rather fragile whenever something wrong happens in our setup()
> codec op. I think we should at least have that op return an error code
> and properly deal with it when that occurs.
>
> I am planning on making a cleanup series to fix that.
>
> Before using timestamps, reference buffer index validation was done in
> std_validate and now it's fully up to the driver. I wonder if it would
> be feasible to bring something back in there since all stateless
> drivers will face the same issue. The conditions set in
> cedrus_reference_index_find seem generic enough for that, but we should
> check that std_validate is not called too early, when the queue is in a
> different state (and the buffer might not be dequeud or done yet).
>
> What do you think?

I don't think you can do that. Say you queue a buffer that refers to a
ref frame F, and F is a buffer that's been dequeued. When std_validate is
called, this is fine and valid. Now later (but before the request is
processed) buffer F is queued by the application. Now F is no longer valid.

So when the driver processes the request, it won't be able to find F anymore.

A more interesting question is what you should do if a reference frame isn't
found. And should that be documented in the stateless decoder spec? (Or perhaps
it's there already, I'm not sure).

Regards,

Hans

>
> Cheers,
>
> Paul
>
>>> Signed-off-by: Paul Kocialkowski <paul.kocialkowski@xxxxxxxxxxx>
>>> ---
>>> drivers/staging/media/sunxi/cedrus/cedrus_dec.c | 13 +++++++++++++
>>> drivers/staging/media/sunxi/cedrus/cedrus_dec.h | 2 ++
>>> drivers/staging/media/sunxi/cedrus/cedrus_mpeg2.c | 10 ++++++----
>>> 3 files changed, 21 insertions(+), 4 deletions(-)
>>>
>>> diff --git a/drivers/staging/media/sunxi/cedrus/cedrus_dec.c b/drivers/staging/media/sunxi/cedrus/cedrus_dec.c
>>> index 443fb037e1cf..2c295286766c 100644
>>> --- a/drivers/staging/media/sunxi/cedrus/cedrus_dec.c
>>> +++ b/drivers/staging/media/sunxi/cedrus/cedrus_dec.c
>>> @@ -22,6 +22,19 @@
>>> #include "cedrus_dec.h"
>>> #include "cedrus_hw.h"
>>>
>>> +int cedrus_reference_index_find(struct vb2_queue *queue,
>>> + struct vb2_buffer *vb2_buf, u64 timestamp)
>>> +{
>>> + /*
>>> + * Allow using the current capture buffer as reference, which can occur
>>> + * for field-coded pictures.
>>> + */
>>> + if (vb2_buf->timestamp == timestamp)
>>> + return vb2_buf->index;
>>> + else
>>> + return vb2_find_timestamp(queue, timestamp, 0);
>>> +}
>>> +
>>> void cedrus_device_run(void *priv)
>>> {
>>> struct cedrus_ctx *ctx = priv;
>>> diff --git a/drivers/staging/media/sunxi/cedrus/cedrus_dec.h b/drivers/staging/media/sunxi/cedrus/cedrus_dec.h
>>> index d1ae7903677b..8d0fc248220f 100644
>>> --- a/drivers/staging/media/sunxi/cedrus/cedrus_dec.h
>>> +++ b/drivers/staging/media/sunxi/cedrus/cedrus_dec.h
>>> @@ -16,6 +16,8 @@
>>> #ifndef _CEDRUS_DEC_H_
>>> #define _CEDRUS_DEC_H_
>>>
>>> +int cedrus_reference_index_find(struct vb2_queue *queue,
>>> + struct vb2_buffer *vb2_buf, u64 timestamp);
>>> void cedrus_device_run(void *priv);
>>>
>>> #endif
>>> diff --git a/drivers/staging/media/sunxi/cedrus/cedrus_mpeg2.c b/drivers/staging/media/sunxi/cedrus/cedrus_mpeg2.c
>>> index cb45fda9aaeb..81c66a8aa1ac 100644
>>> --- a/drivers/staging/media/sunxi/cedrus/cedrus_mpeg2.c
>>> +++ b/drivers/staging/media/sunxi/cedrus/cedrus_mpeg2.c
>>> @@ -10,6 +10,7 @@
>>> #include <media/videobuf2-dma-contig.h>
>>>
>>> #include "cedrus.h"
>>> +#include "cedrus_dec.h"
>>> #include "cedrus_hw.h"
>>> #include "cedrus_regs.h"
>>>
>>> @@ -159,8 +160,8 @@ static void cedrus_mpeg2_setup(struct cedrus_ctx *ctx, struct cedrus_run *run)
>>> cedrus_write(dev, VE_DEC_MPEG_PICBOUNDSIZE, reg);
>>>
>>> /* Forward and backward prediction reference buffers. */
>>> - forward_idx = vb2_find_timestamp(cap_q,
>>> - slice_params->forward_ref_ts, 0);
>>> + forward_idx = cedrus_reference_index_find(cap_q, &run->dst->vb2_buf,
>>> + slice_params->forward_ref_ts);
>>>
>>> fwd_luma_addr = cedrus_dst_buf_addr(ctx, forward_idx, 0);
>>> fwd_chroma_addr = cedrus_dst_buf_addr(ctx, forward_idx, 1);
>>> @@ -168,8 +169,9 @@ static void cedrus_mpeg2_setup(struct cedrus_ctx *ctx, struct cedrus_run *run)
>>> cedrus_write(dev, VE_DEC_MPEG_FWD_REF_LUMA_ADDR, fwd_luma_addr);
>>> cedrus_write(dev, VE_DEC_MPEG_FWD_REF_CHROMA_ADDR, fwd_chroma_addr);
>>>
>>> - backward_idx = vb2_find_timestamp(cap_q,
>>> - slice_params->backward_ref_ts, 0);
>>> + backward_idx = cedrus_reference_index_find(cap_q, &run->dst->vb2_buf,
>>> + slice_params->backward_ref_ts);
>>> +
>>> bwd_luma_addr = cedrus_dst_buf_addr(ctx, backward_idx, 0);
>>> bwd_chroma_addr = cedrus_dst_buf_addr(ctx, backward_idx, 1);
>>>
>>>