Hi Simon,
This is reading a lot of DT stuff very early, which may be slow. It is
also reading from the DT in the bind() step which we sometimes have to
do, but try to avoid.
Could we set up the livetree pre-bind? What about MMU? On armv8 at least
this would have a huge impact on performance. I've done some
measurements and there is at least 1 order of magnitude difference
between parsing FDT with no caches vs parsing livetree with, it's huge.
That seems like a great idea to me, in general. The fact that SPL sets
up the MMU on armv8 makes it more practical.
Well, on qcom we don't use SPL (yet?), we did have a cyclical dependency
since we rely on DTB for the memory layout, although I have some patches
to do all the memory parsing in board_fdt_blob_setup() since that's
needed for multi-dtb FIT. So I guess we could enable caches at the same
time.
Yes...it seems that enabling cache in SPL has become common on armv8.
As to the memory layout, I'm not sure what is happening there, but it
seems that the DT does not describe it in general (at least not until
U-Boot adds the nodes).
I suppose this depends on the platform. On Qualcomm we use DT as the
source of truth as it lets us support many platforms (with totally
different memory maps) with a single U-Boot binary, at least for
development this is quite nice.
But for this series I believe we are going to have to define what
happens in what phase. We have power_init_board() which is the old way
of doing this...but perhaps we could use that as a way to start up
regulators which are needed.
As to my question about whether this happens in SPL / pre-reloc /
proper, I forgot that we have the bootph tags for that, so it should
be fine. The main issue is that in U-Boot proper we will re-init the
regulators even though that has already been done. Probably that can
be handled by Kconfig or a flag in SPL handoff.
Ensuring that it isn't done multiple times sounds like the right
approach to me.
OK...I wonder how we can solve this. Needs some though.
Also this seems to happen in SPL and again pre-reloc and again in
U-Boot post-reloc?
Should we have a step in the init sequence where we set up the
regulators, by calling regulators_enable_boot_on() ?
Regards,
Simon
--
// Caleb (they/them)
