AM5728 video display query

Prasad Mallya said:

With kernel v3.14, though drmModeSetPlane() call was async, using the ioctl DRM_IOCTL_WAIT_VBLANK, we were able to know when a VBLANK interrupt has occurred, hence we were able to know when the work is done. Thus, we have achieved full fps with multiple plane/display use cases without any screen tearing with this mechanism.

That works sometimes, but is not guaranteed to work. When the call to drmModeSetPlane() returns, you don't know if the new configuration has been committed to the HW. The next vblank can happen before the new configuration is written to the HW. Or, if you've called drmModePageFlip() before drmModeSetPlane(), most likely the work for SetPlane will be delayed until after the next vblank, and, again, waiting for vblank will return too early.

Prasad Mallya said:

Is the libdrm provided as part of PSDK V02.00.01 up to date or else when can we expect a release with the updated libdrm? Can you provide us the list of changes to be done in the libdrm?
To use the IOCTLs directly, is there any reference, which we can refer, to understand how the ioctl DRM_IOCTL_MODE_SETPLANE has to be used?

I'm not that familiar with the PSDK versions, but we're just about to update the libdrm for our next internal release, so anything out there at the moment does not have updated libdrm.

libdrm needs to have support for atomic modesetting. I think that support was added in libdrm v2.4.62.

If you want to use the ioctls directly, I think the best reference is to look at a more recent libdrm, and how it deals with it.

With atomic modesetting, DRM_IOCTL_MODE_SETPLANE is no longer used. Instead, the page flipping and configuring is done via atomic modesetting ioctls. The related libdrm functions are drmModeAtomicAlloc, drmModeAtomicFree, drmModeAtomicAddProperty and drmModeAtomicCommit.

Prasad Mallya said:

Our understanding from browsing through the kernel source code of PSDK 02.00.00.00 ( linux kernel 4.1.6 gittag: 52c4aa7cdb-g52c4aa7) is that setplane_internal which is called by drm_mode_setplane() function locks the drm device and then submits and waits for a set plane request to complete. Since this lock is holding the entire drm resource, will it be possible to use drmModeSetPlane() call in synchronous mode and still achieve full fps with multiple plane/display use cases?

No, but that's not possible with 3.14 either. It's true that with 3.14, if carefully timed, you can get full fps in some cases. But even then the app can't be sure if the framebuffers are currently being shown on the screen or not, so even with full fps you may get tearing.

Without atomic modesetting, I'd recommend only using drmModePageFlip(), i.e. not using the planes at all. The drmModeSetPlane() API is, in my opinion, "broken".

Atomic modesetting API fixes all these issues, and all planes can be used on multiple displays with full fps.

Now, as discussed, unfortunately we couldn't get the libdrm updated, so until a PSDK release comes out that contains updated libdrm, so extra work is needed by the user to use atomic modesetting.

I think there are a few options for that:

• Update the libdrm yourself to a more recent version. I believe this will make SGX not to work, so this is not an option if SGX is required.
• Copy the atomic modesetting code from more recent libdrm to your app, which will call the ioctls. It'll be easy to drop that code later and use libdrm, when libdrm is updated.
• Call the ioctls yourself.