[libcamera-devel,v3,3/6] libcamera: buffer: Initialise status
diff mbox series

Message ID 20210324150125.1318325-4-kieran.bingham@ideasonboard.com
State Rejected
Delegated to: Kieran Bingham
Headers show
Series
  • IPU3 Stability
Related show

Commit Message

Kieran Bingham March 24, 2021, 3:01 p.m. UTC
Buffers queued to a pipeline handler may not be yet queued to a device
when the request is cancelled.

This can lead to the FrameMetadata having never been explicitly set by
an underlying V4L2 device.

The status field on this is used to check the state of the buffer to
determine if it was correctly filled, or if it was cancelled.

In the event that the buffer is not used, it must be marked as Cancelled
as the metadata associated with that frame will not be valid.

Initialise the FrameMetadata to FrameCancelled to prevent uninitialised access.
Furthermore, re-initialise the metadata to this state when the buffers
are re-used, or added to a Request to ensure that they are in the
correct state in the event that they do not get used at all.

Signed-off-by: Kieran Bingham <kieran.bingham@ideasonboard.com>

---
v3:
 - Initialise as FrameCancelled
 - Re-initialise to FrameCancelled when buffers are re-used
 - Do not re-order the enum

 include/libcamera/buffer.h | 2 +-
 src/libcamera/request.cpp  | 9 +++++++--
 2 files changed, 8 insertions(+), 3 deletions(-)

Comments

Laurent Pinchart March 24, 2021, 3:38 p.m. UTC | #1
Hi Kieran,

Thank you for the patch.

On Wed, Mar 24, 2021 at 03:01:22PM +0000, Kieran Bingham wrote:
> Buffers queued to a pipeline handler may not be yet queued to a device
> when the request is cancelled.
> 
> This can lead to the FrameMetadata having never been explicitly set by
> an underlying V4L2 device.
> 
> The status field on this is used to check the state of the buffer to
> determine if it was correctly filled, or if it was cancelled.
> 
> In the event that the buffer is not used, it must be marked as Cancelled
> as the metadata associated with that frame will not be valid.
> 
> Initialise the FrameMetadata to FrameCancelled to prevent uninitialised access.
> Furthermore, re-initialise the metadata to this state when the buffers
> are re-used, or added to a Request to ensure that they are in the
> correct state in the event that they do not get used at all.

I wonder if this scheme could be a bit fragile. We're now essentially
telling pipeline handlers that they don't need to take care of marking
the buffer status for cancelled requests if the buffers haven't been
queued to the device. This means they could possibly overlook some other
important processing needed on buffer. There's no such thing at the
moment, time will tell if there will be in the future, in which case we
may want to reconsider this.

> Signed-off-by: Kieran Bingham <kieran.bingham@ideasonboard.com>
> 
> ---
> v3:
>  - Initialise as FrameCancelled
>  - Re-initialise to FrameCancelled when buffers are re-used
>  - Do not re-order the enum
> 
>  include/libcamera/buffer.h | 2 +-
>  src/libcamera/request.cpp  | 9 +++++++--
>  2 files changed, 8 insertions(+), 3 deletions(-)
> 
> diff --git a/include/libcamera/buffer.h b/include/libcamera/buffer.h
> index 302fe3d3e86b..f6673e2f6eda 100644
> --- a/include/libcamera/buffer.h
> +++ b/include/libcamera/buffer.h
> @@ -28,7 +28,7 @@ struct FrameMetadata {
>  		unsigned int bytesused;
>  	};
>  
> -	Status status;
> +	Status status = FrameCancelled;

This isn't strictly needed given the code below, but it's nice to be
consistent. Maybe we should rename it to FrameInvalid or something along
those lines later, to reflect the fact that it's not just about
cancelled requests now.

Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>

>  	unsigned int sequence;
>  	uint64_t timestamp;
>  	std::vector<Plane> planes;
> diff --git a/src/libcamera/request.cpp b/src/libcamera/request.cpp
> index 0071667e4a2c..9ea6ed09446b 100644
> --- a/src/libcamera/request.cpp
> +++ b/src/libcamera/request.cpp
> @@ -118,8 +118,12 @@ void Request::reuse(ReuseFlag flags)
>  	pending_.clear();
>  	if (flags & ReuseBuffers) {
>  		for (auto pair : bufferMap_) {
> -			pair.second->request_ = this;
> -			pending_.insert(pair.second);
> +			FrameBuffer *buffer = pair.second;
> +
> +			buffer->metadata_.status = FrameMetadata::Status::FrameCancelled;
> +			buffer->request_ = this;
> +
> +			pending_.insert(buffer);
>  		}
>  	} else {
>  		bufferMap_.clear();
> @@ -188,6 +192,7 @@ int Request::addBuffer(const Stream *stream, FrameBuffer *buffer)
>  	}
>  
>  	buffer->request_ = this;
> +	buffer->metadata_.status = FrameMetadata::Status::FrameCancelled;
>  	pending_.insert(buffer);
>  	bufferMap_[stream] = buffer;
>
Kieran Bingham March 24, 2021, 3:54 p.m. UTC | #2
Hi Laurent,

On 24/03/2021 15:38, Laurent Pinchart wrote:
> Hi Kieran,
> 
> Thank you for the patch.
> 
> On Wed, Mar 24, 2021 at 03:01:22PM +0000, Kieran Bingham wrote:
>> Buffers queued to a pipeline handler may not be yet queued to a device
>> when the request is cancelled.
>>
>> This can lead to the FrameMetadata having never been explicitly set by
>> an underlying V4L2 device.
>>
>> The status field on this is used to check the state of the buffer to
>> determine if it was correctly filled, or if it was cancelled.
>>
>> In the event that the buffer is not used, it must be marked as Cancelled
>> as the metadata associated with that frame will not be valid.
>>
>> Initialise the FrameMetadata to FrameCancelled to prevent uninitialised access.
>> Furthermore, re-initialise the metadata to this state when the buffers
>> are re-used, or added to a Request to ensure that they are in the
>> correct state in the event that they do not get used at all.
> 
> I wonder if this scheme could be a bit fragile. We're now essentially
> telling pipeline handlers that they don't need to take care of marking
> the buffer status for cancelled requests if the buffers haven't been
> queued to the device. This means they could possibly overlook some other
> important processing needed on buffer. There's no such thing at the
> moment, time will tell if there will be in the future, in which case we
> may want to reconsider this.
> 
>> Signed-off-by: Kieran Bingham <kieran.bingham@ideasonboard.com>
>>
>> ---
>> v3:
>>  - Initialise as FrameCancelled
>>  - Re-initialise to FrameCancelled when buffers are re-used
>>  - Do not re-order the enum
>>
>>  include/libcamera/buffer.h | 2 +-
>>  src/libcamera/request.cpp  | 9 +++++++--
>>  2 files changed, 8 insertions(+), 3 deletions(-)
>>
>> diff --git a/include/libcamera/buffer.h b/include/libcamera/buffer.h
>> index 302fe3d3e86b..f6673e2f6eda 100644
>> --- a/include/libcamera/buffer.h
>> +++ b/include/libcamera/buffer.h
>> @@ -28,7 +28,7 @@ struct FrameMetadata {
>>  		unsigned int bytesused;
>>  	};
>>  
>> -	Status status;
>> +	Status status = FrameCancelled;
> 
> This isn't strictly needed given the code below, but it's nice to be
> consistent. Maybe we should rename it to FrameInvalid or something along
> those lines later, to reflect the fact that it's not just about
> cancelled requests now.

Indeed, now that it's set during addBuffer() this becomes at least a
little redundant.

I agree on the naming doesn't feel right any more either.
FrameInvalid might be appropriate, and suitable for both an unused, and
cancelled frame...


> Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
> 
>>  	unsigned int sequence;
>>  	uint64_t timestamp;
>>  	std::vector<Plane> planes;
>> diff --git a/src/libcamera/request.cpp b/src/libcamera/request.cpp
>> index 0071667e4a2c..9ea6ed09446b 100644
>> --- a/src/libcamera/request.cpp
>> +++ b/src/libcamera/request.cpp
>> @@ -118,8 +118,12 @@ void Request::reuse(ReuseFlag flags)
>>  	pending_.clear();
>>  	if (flags & ReuseBuffers) {
>>  		for (auto pair : bufferMap_) {
>> -			pair.second->request_ = this;
>> -			pending_.insert(pair.second);
>> +			FrameBuffer *buffer = pair.second;
>> +
>> +			buffer->metadata_.status = FrameMetadata::Status::FrameCancelled;
>> +			buffer->request_ = this;
>> +
>> +			pending_.insert(buffer);
>>  		}
>>  	} else {
>>  		bufferMap_.clear();
>> @@ -188,6 +192,7 @@ int Request::addBuffer(const Stream *stream, FrameBuffer *buffer)
>>  	}
>>  
>>  	buffer->request_ = this;
>> +	buffer->metadata_.status = FrameMetadata::Status::FrameCancelled;
>>  	pending_.insert(buffer);
>>  	bufferMap_[stream] = buffer;
>>  
>
Kieran Bingham March 24, 2021, 4:25 p.m. UTC | #3
Hi Laurent,

On 24/03/2021 15:54, Kieran Bingham wrote:
> Hi Laurent,
> 
> On 24/03/2021 15:38, Laurent Pinchart wrote:
>> Hi Kieran,
>>
>> Thank you for the patch.
>>
>> On Wed, Mar 24, 2021 at 03:01:22PM +0000, Kieran Bingham wrote:
>>> Buffers queued to a pipeline handler may not be yet queued to a device
>>> when the request is cancelled.
>>>
>>> This can lead to the FrameMetadata having never been explicitly set by
>>> an underlying V4L2 device.
>>>
>>> The status field on this is used to check the state of the buffer to
>>> determine if it was correctly filled, or if it was cancelled.
>>>
>>> In the event that the buffer is not used, it must be marked as Cancelled
>>> as the metadata associated with that frame will not be valid.
>>>
>>> Initialise the FrameMetadata to FrameCancelled to prevent uninitialised access.
>>> Furthermore, re-initialise the metadata to this state when the buffers
>>> are re-used, or added to a Request to ensure that they are in the
>>> correct state in the event that they do not get used at all.
>>
>> I wonder if this scheme could be a bit fragile. We're now essentially
>> telling pipeline handlers that they don't need to take care of marking
>> the buffer status for cancelled requests if the buffers haven't been
>> queued to the device. This means they could possibly overlook some other
>> important processing needed on buffer. There's no such thing at the
>> moment, time will tell if there will be in the future, in which case we
>> may want to reconsider this.

Indeed, I was trying to ensure that /all/ unused buffers were
initialised to an invalid state so that they weren't incorrectly
identified as valid (or that the unintialised status flag may be read by
an applciation, which could be set to 0 meaning it's a good buffer).

Perhaps we should also do this:

> From d7e9bba52e6395c4669a69bd09450f99c842c078 Mon Sep 17 00:00:00 2001
> From: Kieran Bingham <kieran.bingham@ideasonboard.com>
> Date: Wed, 24 Mar 2021 16:21:46 +0000
> Subject: [PATCH] libcamera: pipeline: ipu3: Cancel unused buffers
> 
> When the CIO2 returns a cancelled buffer, we will not queue buffers
> to the IMGU.
> 
> These buffers should be explicitly marked as cancelled to ensure
> the application knows there is no valid metadata or frame data
> provided in the buffer.
> 
> Signed-off-by: Kieran Bingham <kieran.bingham@ideasonboard.com>
> ---
>  src/libcamera/pipeline/ipu3/ipu3.cpp | 7 +++++--
>  1 file changed, 5 insertions(+), 2 deletions(-)
> 
> diff --git a/src/libcamera/pipeline/ipu3/ipu3.cpp b/src/libcamera/pipeline/ipu3/ipu3.cpp
> index af05cb2940c8..b495ea0a5b98 100644
> --- a/src/libcamera/pipeline/ipu3/ipu3.cpp
> +++ b/src/libcamera/pipeline/ipu3/ipu3.cpp
> @@ -1265,8 +1265,11 @@ void IPU3CameraData::cio2BufferReady(FrameBuffer *buffer)
>  
>  	/* If the buffer is cancelled force a complete of the whole request. */
>  	if (buffer->metadata().status == FrameMetadata::FrameCancelled) {
> -		for (auto it : request->buffers())
> -			pipe_->completeBuffer(request, it.second);
> +		for (auto it : request->buffers()) {
> +			FrameBufffer *b = it.second;
> +			b->metadata_.status = FrameMetadata::Status::FrameCancelled;
> +			pipe_->completeBuffer(request, b);
> +		}
>  
>  		frameInfos_.remove(info);
>  		pipe_->completeRequest(request);
> -- 
> 2.25.1


But I still think the state should be initialised to something to make
sure we don't find ourselves making choices on an unitialised state
variable.





>>
>>> Signed-off-by: Kieran Bingham <kieran.bingham@ideasonboard.com>
>>>
>>> ---
>>> v3:
>>>  - Initialise as FrameCancelled
>>>  - Re-initialise to FrameCancelled when buffers are re-used
>>>  - Do not re-order the enum
>>>
>>>  include/libcamera/buffer.h | 2 +-
>>>  src/libcamera/request.cpp  | 9 +++++++--
>>>  2 files changed, 8 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/include/libcamera/buffer.h b/include/libcamera/buffer.h
>>> index 302fe3d3e86b..f6673e2f6eda 100644
>>> --- a/include/libcamera/buffer.h
>>> +++ b/include/libcamera/buffer.h
>>> @@ -28,7 +28,7 @@ struct FrameMetadata {
>>>  		unsigned int bytesused;
>>>  	};
>>>  
>>> -	Status status;
>>> +	Status status = FrameCancelled;
>>
>> This isn't strictly needed given the code below, but it's nice to be
>> consistent. Maybe we should rename it to FrameInvalid or something along
>> those lines later, to reflect the fact that it's not just about
>> cancelled requests now.
> 
> Indeed, now that it's set during addBuffer() this becomes at least a
> little redundant.
> 
> I agree on the naming doesn't feel right any more either.
> FrameInvalid might be appropriate, and suitable for both an unused, and
> cancelled frame...
> 
> 
>> Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
>>
>>>  	unsigned int sequence;
>>>  	uint64_t timestamp;
>>>  	std::vector<Plane> planes;
>>> diff --git a/src/libcamera/request.cpp b/src/libcamera/request.cpp
>>> index 0071667e4a2c..9ea6ed09446b 100644
>>> --- a/src/libcamera/request.cpp
>>> +++ b/src/libcamera/request.cpp
>>> @@ -118,8 +118,12 @@ void Request::reuse(ReuseFlag flags)
>>>  	pending_.clear();
>>>  	if (flags & ReuseBuffers) {
>>>  		for (auto pair : bufferMap_) {
>>> -			pair.second->request_ = this;
>>> -			pending_.insert(pair.second);
>>> +			FrameBuffer *buffer = pair.second;
>>> +
>>> +			buffer->metadata_.status = FrameMetadata::Status::FrameCancelled;
>>> +			buffer->request_ = this;
>>> +
>>> +			pending_.insert(buffer);
>>>  		}
>>>  	} else {
>>>  		bufferMap_.clear();
>>> @@ -188,6 +192,7 @@ int Request::addBuffer(const Stream *stream, FrameBuffer *buffer)
>>>  	}
>>>  
>>>  	buffer->request_ = this;
>>> +	buffer->metadata_.status = FrameMetadata::Status::FrameCancelled;
>>>  	pending_.insert(buffer);
>>>  	bufferMap_[stream] = buffer;
>>>  
>>
>
Laurent Pinchart March 24, 2021, 4:33 p.m. UTC | #4
Hi Kieran,

On Wed, Mar 24, 2021 at 04:25:07PM +0000, Kieran Bingham wrote:
> On 24/03/2021 15:54, Kieran Bingham wrote:
> > On 24/03/2021 15:38, Laurent Pinchart wrote:
> >> On Wed, Mar 24, 2021 at 03:01:22PM +0000, Kieran Bingham wrote:
> >>> Buffers queued to a pipeline handler may not be yet queued to a device
> >>> when the request is cancelled.
> >>>
> >>> This can lead to the FrameMetadata having never been explicitly set by
> >>> an underlying V4L2 device.
> >>>
> >>> The status field on this is used to check the state of the buffer to
> >>> determine if it was correctly filled, or if it was cancelled.
> >>>
> >>> In the event that the buffer is not used, it must be marked as Cancelled
> >>> as the metadata associated with that frame will not be valid.
> >>>
> >>> Initialise the FrameMetadata to FrameCancelled to prevent uninitialised access.
> >>> Furthermore, re-initialise the metadata to this state when the buffers
> >>> are re-used, or added to a Request to ensure that they are in the
> >>> correct state in the event that they do not get used at all.
> >>
> >> I wonder if this scheme could be a bit fragile. We're now essentially
> >> telling pipeline handlers that they don't need to take care of marking
> >> the buffer status for cancelled requests if the buffers haven't been
> >> queued to the device. This means they could possibly overlook some other
> >> important processing needed on buffer. There's no such thing at the
> >> moment, time will tell if there will be in the future, in which case we
> >> may want to reconsider this.
> 
> Indeed, I was trying to ensure that /all/ unused buffers were
> initialised to an invalid state so that they weren't incorrectly
> identified as valid (or that the unintialised status flag may be read by
> an applciation, which could be set to 0 meaning it's a good buffer).
> 
> Perhaps we should also do this:
> 
> > From d7e9bba52e6395c4669a69bd09450f99c842c078 Mon Sep 17 00:00:00 2001
> > From: Kieran Bingham <kieran.bingham@ideasonboard.com>
> > Date: Wed, 24 Mar 2021 16:21:46 +0000
> > Subject: [PATCH] libcamera: pipeline: ipu3: Cancel unused buffers
> > 
> > When the CIO2 returns a cancelled buffer, we will not queue buffers
> > to the IMGU.
> > 
> > These buffers should be explicitly marked as cancelled to ensure
> > the application knows there is no valid metadata or frame data
> > provided in the buffer.
> > 
> > Signed-off-by: Kieran Bingham <kieran.bingham@ideasonboard.com>
> > ---
> >  src/libcamera/pipeline/ipu3/ipu3.cpp | 7 +++++--
> >  1 file changed, 5 insertions(+), 2 deletions(-)
> > 
> > diff --git a/src/libcamera/pipeline/ipu3/ipu3.cpp b/src/libcamera/pipeline/ipu3/ipu3.cpp
> > index af05cb2940c8..b495ea0a5b98 100644
> > --- a/src/libcamera/pipeline/ipu3/ipu3.cpp
> > +++ b/src/libcamera/pipeline/ipu3/ipu3.cpp
> > @@ -1265,8 +1265,11 @@ void IPU3CameraData::cio2BufferReady(FrameBuffer *buffer)
> >  
> >  	/* If the buffer is cancelled force a complete of the whole request. */
> >  	if (buffer->metadata().status == FrameMetadata::FrameCancelled) {
> > -		for (auto it : request->buffers())
> > -			pipe_->completeBuffer(request, it.second);
> > +		for (auto it : request->buffers()) {
> > +			FrameBufffer *b = it.second;
> > +			b->metadata_.status = FrameMetadata::Status::FrameCancelled;
> > +			pipe_->completeBuffer(request, b);
> > +		}
> >  
> >  		frameInfos_.remove(info);
> >  		pipe_->completeRequest(request);

This looks good to me too. It has the advantage of not making us feel
safe erronously that the libcamera core will always provide a solution
for us.

Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>

But I'm also OK with the original patch for the time being, you can pick
the one you like best (or the one selected by your favourite random
number generator :-)).

> But I still think the state should be initialised to something to make
> sure we don't find ourselves making choices on an unitialised state
> variable.

Fully agreed.

> >>> Signed-off-by: Kieran Bingham <kieran.bingham@ideasonboard.com>
> >>>
> >>> ---
> >>> v3:
> >>>  - Initialise as FrameCancelled
> >>>  - Re-initialise to FrameCancelled when buffers are re-used
> >>>  - Do not re-order the enum
> >>>
> >>>  include/libcamera/buffer.h | 2 +-
> >>>  src/libcamera/request.cpp  | 9 +++++++--
> >>>  2 files changed, 8 insertions(+), 3 deletions(-)
> >>>
> >>> diff --git a/include/libcamera/buffer.h b/include/libcamera/buffer.h
> >>> index 302fe3d3e86b..f6673e2f6eda 100644
> >>> --- a/include/libcamera/buffer.h
> >>> +++ b/include/libcamera/buffer.h
> >>> @@ -28,7 +28,7 @@ struct FrameMetadata {
> >>>  		unsigned int bytesused;
> >>>  	};
> >>>  
> >>> -	Status status;
> >>> +	Status status = FrameCancelled;
> >>
> >> This isn't strictly needed given the code below, but it's nice to be
> >> consistent. Maybe we should rename it to FrameInvalid or something along
> >> those lines later, to reflect the fact that it's not just about
> >> cancelled requests now.
> > 
> > Indeed, now that it's set during addBuffer() this becomes at least a
> > little redundant.
> > 
> > I agree on the naming doesn't feel right any more either.
> > FrameInvalid might be appropriate, and suitable for both an unused, and
> > cancelled frame...
> > 
> >> Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
> >>
> >>>  	unsigned int sequence;
> >>>  	uint64_t timestamp;
> >>>  	std::vector<Plane> planes;
> >>> diff --git a/src/libcamera/request.cpp b/src/libcamera/request.cpp
> >>> index 0071667e4a2c..9ea6ed09446b 100644
> >>> --- a/src/libcamera/request.cpp
> >>> +++ b/src/libcamera/request.cpp
> >>> @@ -118,8 +118,12 @@ void Request::reuse(ReuseFlag flags)
> >>>  	pending_.clear();
> >>>  	if (flags & ReuseBuffers) {
> >>>  		for (auto pair : bufferMap_) {
> >>> -			pair.second->request_ = this;
> >>> -			pending_.insert(pair.second);
> >>> +			FrameBuffer *buffer = pair.second;
> >>> +
> >>> +			buffer->metadata_.status = FrameMetadata::Status::FrameCancelled;
> >>> +			buffer->request_ = this;
> >>> +
> >>> +			pending_.insert(buffer);
> >>>  		}
> >>>  	} else {
> >>>  		bufferMap_.clear();
> >>> @@ -188,6 +192,7 @@ int Request::addBuffer(const Stream *stream, FrameBuffer *buffer)
> >>>  	}
> >>>  
> >>>  	buffer->request_ = this;
> >>> +	buffer->metadata_.status = FrameMetadata::Status::FrameCancelled;
> >>>  	pending_.insert(buffer);
> >>>  	bufferMap_[stream] = buffer;
> >>>

Patch
diff mbox series

diff --git a/include/libcamera/buffer.h b/include/libcamera/buffer.h
index 302fe3d3e86b..f6673e2f6eda 100644
--- a/include/libcamera/buffer.h
+++ b/include/libcamera/buffer.h
@@ -28,7 +28,7 @@  struct FrameMetadata {
 		unsigned int bytesused;
 	};
 
-	Status status;
+	Status status = FrameCancelled;
 	unsigned int sequence;
 	uint64_t timestamp;
 	std::vector<Plane> planes;
diff --git a/src/libcamera/request.cpp b/src/libcamera/request.cpp
index 0071667e4a2c..9ea6ed09446b 100644
--- a/src/libcamera/request.cpp
+++ b/src/libcamera/request.cpp
@@ -118,8 +118,12 @@  void Request::reuse(ReuseFlag flags)
 	pending_.clear();
 	if (flags & ReuseBuffers) {
 		for (auto pair : bufferMap_) {
-			pair.second->request_ = this;
-			pending_.insert(pair.second);
+			FrameBuffer *buffer = pair.second;
+
+			buffer->metadata_.status = FrameMetadata::Status::FrameCancelled;
+			buffer->request_ = this;
+
+			pending_.insert(buffer);
 		}
 	} else {
 		bufferMap_.clear();
@@ -188,6 +192,7 @@  int Request::addBuffer(const Stream *stream, FrameBuffer *buffer)
 	}
 
 	buffer->request_ = this;
+	buffer->metadata_.status = FrameMetadata::Status::FrameCancelled;
 	pending_.insert(buffer);
 	bufferMap_[stream] = buffer;