Hi Sakari, On Wed, Apr 17, 2019 at 01:41:48PM +0300, Sakari Ailus wrote: > Hi Jacopo, > > I promised to write a bit about the routing problem I attempted to write on > #v4l. It's a bit late, but here it is. Let me know if you share my > understanding (and especially do so if you don't :)). Thanks for doing this, and sorry it took me a while to reply! > > A route is identified by the source and sink pads as well as the stream ID > on the source and the sink pads. Its flags allow enabling or disabling a > route. > > Most devices that function as transmitters (such as sensors) have a fixed > number of routes they can support. These's no need to change them; enabling > and disabling them will be enough for the user. > Fine so far. We indeed define routes as: (sink pad - sink stream) -> (source pad - source stream) > For receivers this is different. What needs to be supported on the receiver > side is directing any incoming stream (a 32-bit unsigned integer) to any > source pad. > > This is because pads are not alike --- one may be connected to a block that > further processes the image whereas the others may be connected to a DMA > engine, just writing the data to memory. > > The receivers also may not make assumptions beyond the sub-device API: the > stream is a 32-bit unsigned integer, there is currently no API requirement > to have the stream IDs on a particular low integer range. In principle we > could define that range, but I'd rather try to find better solutions than > that: it's hard to come up with a number as it depends on the hardware. > Some kind of an upper limit could be the number of CSI-2 channels > multiplied by CSI-2 data types. That would be enough for CSI-2. 16 or 32 > virtual channels and up to 64 data types would mean up to 2048 routes > between a demultiplexer's sink pad and *each* of its source pads. And this > comes with an assumption that the source pads only support a single > stream... I lost you here. My mental model was far more (too?) simple: - Routing tables inside an entity might have an arbitrary size, as how that configuration depends on the device and the driver implementation. As an example, the adv748x accepts 7 analogue inputs to chose from and route to a CSI-2 source pad. depending on the implementation, those 7 inputs could be modeled as 7 pads with one stream each, or a single pad with 7 channels or whatever else, and I agree the 'stream_id' values range is totally up to the driver implementation -inside an entity-. - The cross-entity (sorry, I lack a better term here) multiplexing happens on physical bus that allows so, and I can only think of CSI-2 at the moment. Sure, you could share the lines of a parallel bus playing with enables/disables of the transmitters, but this is a custom hack that does not play well in this model. Each CSI-2 source pad has up to 4 streams (one per VC) and the content of those streams is retrieved from the transmitter by the receiver through the remote frame_desc operation [*], as receivers might need to be setup using to filter on particular VC/DT combinations to receive the stream. - Each CSI-2 receiver sink pad supports 4 streams (the CSI-2 VCs) and any of those streams can be directed to any of its source pads, to implement what you have described (one pad connected to an ISP-like device, on other to the DMA engine directly etc) - DT negotiation is still a bit vague. The example we have on the series (adv748x and r-car csi2) configure the receiver's accepted DT inspecting the remote frame_desc. Again, as per [*] this migh be limitied to 1 DT for VC, which might not be enough and would require re-thinking the operation used for the negotiation. - Configuring a multiplexed source pad image format is today not possible, the format is always propagated from the sink pad to which a route is enabled to the source pad. Might this be a limitation on how we control which DTs are multiplexed inside a VC and won't allow to model any format conversion that might happen at the source pad output. I was toying myself with the idea of a stream-aware set format operation for multiplexed source pads, not sure it might work though. Can I ask you why: - you mention 16 or 32 VCs ? Each CSI-2 link supports up to 4. - you put DT in the routing mix, and I suspect it is here where our disconnection happens. I always assume DT configuration as a result of a format configuration operation, which currently has limitations as noted here above. [*] I would later like to talk about if this is the most appropriate operation to handle this negotiation, as I'm not sure we can handle DT negotiation properly with that, but that's for later. > > CSI-2 receivers support a number of simultaneous streams, and as the stream > is demultiplexed there, this means there will be as many source pads as > there are supported simultaneous streams. This heavily depe`nds on the > hardware, but the number could be e.g. 1, 4 or 8. Much smaller than 2048 in > any case. > Why 8? Is this related to DT multiplexing again? > Another option could be creating no routes at all at device init time, and > letting the user create them. We could add a new flag for routes telling > that a route is dynamic: disabling a dynamic route would simply delete it. > Likewise, a SUBDEV_G_ROUTING returning no routes (but no error either) > would also tell the user only dynamic routes are supported. > > We should document this and change the driver implementation accordingly. > No API changes are needed (apart from adding the dynamic flag --- or rather > call it "volatile", to avoid confusing with routes that can be enabled or > disabled while streaming). > I have v5 of the series ready. Do you want me to post it to progress discussion, or should we clarify this first? Thanks j > -- > Best regards, > > Sakari Ailus > sakari.ailus@linux.intel.com