On Wed, Jul 04, 2007 at 09:36:55PM +0200, Wolfgang Denk wrote:
I wouldn't go as far as to say that. Maybe if the one-size-fits-all approach does not work (at least not without hurting), then there is still potential for a _compatible_ second approach? Who says the "first step" *must* be as small as 4 kB, and *must* be in a separately linked image, and *must* not link higher level functions from the "real" U-Boot? Maybe both is possible, so everybody can select the config he likes and be happy?
v2 was designed in a way to make exactly that choice possible, in a quite flexible way.
The thing is that there are many requirements which are pretty complicated and sometimes contradicting:
- systems booting from NAND or dataflash etc. may require a very small primary bootstrap loader
- systems with very limited resources may need a primary bootstrap loader that unpacks the compressed U-Boot image
Should be no problem with Sascha's design.
- dynamic switching of the console device may require early access to the device tree
- allowing for a configurable console baudrate or for software- adjustable CPU clock rates may require early access to the environment
These ones are problematic.
What I would like to acchieve is an open discussion where none of these requirements gets excluded just because it does not fit into the currently preferred design model. Until a few days ago we had only one code base to complain about. Now - thanks to Saschas excellent work - we have two different implementations which we can compare. This is an excellent chance to evaluate positions, to find out what's good in one design or the other, and what should be avoided.
That's the idea, yes.
If any feature is possible in the old design but leads to ugly code or other issues, we should not simply drop it, but instead discuss how it could be adapted for the other esign based on clean and beautiful (TM) code.
Yep, agreed. Please go ahead and make suggestions.
I agree with that. And as far as I remember the previous discussion, nobody ever raised any concerns or whatever.
Yes, we noticed. So either nobody read the code, or it must be really good :-)
Agreed. The printf() thingy is just a problem that needs a solution, it is not a killing point.
Could you elaborate on how these pluggable-ram boards are used in real life? For all of our boards it would surely be enough to have a compile time switch which declares one of the serial ports the early-debug console, plus hardcode a baudrate on it. That would be configurable, so somebody who wants to have it switched on in a production system has the choice, and someone who doesn't need it because a shipped board is never changing again could configure it off.
It is not the only one - as mentioned above, there are other things which I see difficult to implement in the new code,
Let's identify the problematic cases. At the moment I see these:
- pluggable memory -> should be fixable by more intelligent ram probing code
- debug-output-before-sdram
- debug-output-before-sdram-on-console-taken-from-$COMPLICATED_DEVICE
not to mention the huge effort that will be needed just to port a small fraction of the currently actively used boards. But it wouldn't be software if all was easy and just working ;-)
Well, that's easy. Everybody who wants to have his board supported in v2 just sends a patch. Everybody who doesn't care, does not, and it doesn't matter, because if he does not care, why should the community do the work instead? v1 would still be available for reference, if somebody decides later to go the v2 way.
Please note that what you seem to consider corner cases are real problems for some of our customers.
Well, I learned that one of the most important things in this business is to tell customers that if they do braindamaged decisions, it has implications. So if somebody designs a board with pluggable ram but without beeper, it could have the implication that one doesn't see anything. If a board can only initialize it's RAM without loading an FPGA from software, it's doomed anyway. My very personal decision is that there are projects out there who are so painful that it's probably better if somebody else does them, because I don't want to let my team feel the pain and I don't want to leave the customer with the bad feeling that we didn't do a high quality solution for him.
But I surely understand your argument.
Note that, if these special cases are really rare, there's also the possibility that, for features where there's no good solution, the people continue to use v1.
And I have seen more than enough casesmyself when a board stopped after printing "RAM: " (very often!) or "Flash: " (still quie frequently). Please blieve me that I don't want to discuss this to deat by raising artifical arguments. I am serious.
Sorry, our experience is different.
Ok, so let's stop the discussion here, I assume I've understood your arguments. Please let's continue with the "problematic cases" list above and try to find solutions.
Robert