From: Rob Clark rob@ti.com
I've been working for the better part of the week on solving some of the omapdss vs kms mismatches, which is one of the bigger remaining issues in the TODO before moving omapdrm out of staging.
The biggest place that this shows up is in GO bit handling. Basically, some of the scanout related registers in DSS hw block are only shadow registers, and if the GO bit is set during the vblank the hw copies into the actual registers (not accessible to CPU) and clears the GO bit. When the GO bit is set, there is no safe way to update these registers without undefined results. So omapdss tries to be friendly and abstract this, by buffering up the updated state, and applying it on the next vblank once the GO bit is cleared. But this causes all sorts of mayhem at the omapdrm layer, which would like to unpin the previous scanout buffer(s) on the next vblank (or endwin) irq. Due to the buffering in omapdss, we have no way to know on a vblank if we have switched to the scanout buffer or not. Basically it works ok as long as userspace is only ever updating on layer (either crtc or drm plane) at a time. But throw together hw mouse cursor (drm plane) plus a window manager like compiz which does page flips, or wayland (weston drm compositor) with hw composition (drm plane), and things start to fail in a big way.
I've tried a few approaches to preserve the omapdss more or less as it is, by adding callbacks for when GO bit is cleared, etc. But the sequencing of setting up connector/encoder/crtc is not really what omapdss expects, and would generally end up confusing the "apply" layer in omapdss (it would end up not programming various registers because various dirty flags would get cleared, for example mgr updated before overlay connected, etc).
Finally, in frustration, this afternoon I hit upon an idea. Why not just use the dispc code in omapdss, which is basically a stateless layer of helper functions, and bypass the stateful layer of omapdss. Since KMS helper functions already give us the correct sequence for setting up the hardware, this turned out to be rather easy. And fit it quite nicely with my mechanism to queue up updates when the GO bit is not clear. And, unlike my previous attempts, it actually worked.. not only that, but it worked on the first boot!
So I am pretty happy about how this is shaping up. Not only is it simpler that my previous attepmts, and solves a few tricky buffer unpin related issues. But it also makes it very easy to wire in the missing userspace vblank event handling without resorting to duct- tape.
Obviously there is stuff still missing, and some hacks. This is really just a proof of concept at this stage. But I wanted to send an RFC so we could start discussing how to move forward. Ie. could we reasonably add support to build dispc as a library of stateless helper functions, sharing it and the panel drivers between omapdrm and the legacy omapdss based drivers. Or is there no clean way to do that, in which case we should just copy the code we need into omapdrm, and leave the deprecated omapdss as it is for legacy drivers.
Rob Clark (3): OMAPDSS: expose dispc for use from omapdrm omap2+: use dss_dispc hwmod for omapdrm drm/omap: use dispc directly
arch/arm/mach-omap2/drm.c | 24 +- drivers/staging/omapdrm/Makefile | 1 + drivers/staging/omapdrm/omap_connector.c | 20 +- drivers/staging/omapdrm/omap_crtc.c | 42 ++-- drivers/staging/omapdrm/omap_drv.c | 70 +----- drivers/staging/omapdrm/omap_drv.h | 64 ++++- drivers/staging/omapdrm/omap_encoder.c | 213 ++++++++++++++--- drivers/staging/omapdrm/omap_irq.c | 133 +++++++++++ drivers/staging/omapdrm/omap_plane.c | 382 +++++++++--------------------- drivers/video/omap2/dss/apply.c | 4 +- drivers/video/omap2/dss/dispc.c | 76 ++++-- drivers/video/omap2/dss/dss.h | 2 + drivers/video/omap2/dss/hdmi.c | 18 +- 13 files changed, 628 insertions(+), 421 deletions(-) create mode 100644 drivers/staging/omapdrm/omap_irq.c