On Tuesday 12 March 2013 07:59 PM, Tomi Valkeinen wrote: > On 2013-03-12 16:01, Archit Taneja wrote: >> On Tuesday 12 March 2013 07:07 PM, Tomi Valkeinen wrote: > >>> So, I don't disagree with you. But I don't quite understand why we could >>> not use the fixed channels for now? They should work in all the boards >>> we have, right? Or is there something with DRM that forces the driver to >>> select the channel dynamically? >> >> I think we can use fixed channels, but if the number of different fixed >> channels crosses the number of crtcs the kernel wants, then we would >> need to atleast change the channels of some of the outputs. >> >> For example, suppose omapdrm is asked to use only 2 crtcs, and it picks >> up LCD2 and TV managers. Now if there is some panel which says it's >> recommended channel is LCD, then things won't work. > > Are you saying omapdrm picks the managers for the crtcs before knowing > what panels there are? That can't work right... We need to know what > outputs are to be used before we can select the managers. Or, we always > need crtcs for all the managers.
That's how it is right now. > > If we do know the panels, and thus outputs, then the managers to be used > are found easily from output->dispc_channel. Yes, my patch tries to do the same, but it could assign a manger which isn't the recommended channel. It can pick one from the list of supported channels. I've explained it a bit more below. > > But, of course, the crtc to manager mapping could be changed (if omapdrm > supports this). If omapdrm is asked to use only 1 crtc, but there are > two panels, then only one panel can be used at a time, and the manager > for the crtc needs to be changed when the panel to be used is changed. > But even in this case used manager is clear, it comes from > output->dispc_channel. This is something I don't know that can be done or not, or if it can be done easily. A crtc isn't purely an overlay manager. It also needs to have one plane associated to it. So, if we want to change the overlay manager tied to a crtc on the fly, we should make sure that it's still connected to a plane pointing to the same buffer. This needs a better understanding of drm internals. I guess Rob could answer this better. > >> At the moment, omapdrm maps a crtc with a manger using a function called >> pipe2chan() which just selects a manager with the biggest channel no. So >> if the kernel is configured to have num_crtcs as 1. The single crtc will >> be mapped to LCD2. This method is wrong, as it doesn't even look at the >> type of panels at all. For an omap5 panda, the most suitable manager to >> map to the crtc would be TV(for hdmi). >> >> I think what we probably need to do is to combine both the methods. I.e, >> make each output connectible to only one channel, and also iterate >> through the panels in omapdrm to find the most suitable channels. So in >> my patch, instead of looking at all the supported managers for an >> output(checking with dss_feat_get_supported_outputs() on each manager), >> I just look at the recommended channel, and try to map that manager. > > I don't know, I feel like I'm not understanding something here =). :) I think what I said above is equivalent to what you said here: "If we do know the panels, and thus outputs, then the managers to be used are found easily from output->dispc_channel." Instead of using output->dispc_channel, the patch I made tried to get the first supported channel(in increasing order of channel index) for an output. If that channel was already taken/reserved by a previous output, it tries to take the next supported channel if there are enough crtcs left. My patch would break things if it chooses a manager for an output for which we haven't written the necessary code yet(switching clock sources etc). So, what I'm saying is that we should stick to output->dispc_channel. We iterate through all the panels, and by using output->dispc_channel, we get the manager for an output, and map that manager to a crtc, and make sure the number of unique managers we finally use is equal to NUM_CRTC. Does that sound good? Archit
