On Sat, Apr 08, 2023 at 09:35:48AM +1200, Simon Glass wrote:
Hi Tom,
On Sat, 8 Apr 2023 at 08:22, Tom Rini trini@konsulko.com wrote:
On Sat, Apr 08, 2023 at 07:53:10AM +1200, Simon Glass wrote:
Hi Tom,
On Sat, 8 Apr 2023 at 07:39, Tom Rini trini@konsulko.com wrote:
On Fri, Apr 07, 2023 at 10:36:38PM +1200, Simon Glass wrote:
This series converts rockchip boards over to use standard boot. It also fixes various problems which have come up recently, showing differences between the current implementation and the distroboot scripts.
This should get us closer to being able to turn down the scripts.
Alright, so I grabbed a few parts of this series to investigate the points I'm trying to grasp better, and I think this is going the wrong track. We should start off by dropping "default y" from BOOTSTD, and then start adding "default y if" for SoCs as we convert them. The end goal should be that we get to the point where we can "default y if ARM || RISCV || X86" or perhaps "default y !(PPC || M68K || ...)" as it's just a few architectures that haven't ended up being converted. But today, there's too much churn on platforms that aren't making any use of this. And I don't think this is going to be functionally worse than all of the places we "imply DISTRO_DEFAULTS" today, as functionally we can replace that with "imply BOOTSTD" as they get migrated.
That would really be a backward step. I'm not sure what to say at this point. I've put a lot of effort in trying to get this over the line, but the only way we get feedback is when it is applied.
Having bootstd enabled and not functional (because boot_targets aren't set) isn't helping the migration happen. And the hard part of the migration isn't knowing it's possible, or enabling 1-2 options, it's testing it and also it really just being 1-2 options.
Standard boot does not need boot_targets to be set. It works fine without it. It just goes through the boot devices in a pre-defined order, from fastest to slowest. It matches what most boards do anyway. The main reason we kept it is for compatibility with distro boot.
What most boards do today is just sit at the prompt and wait for input, which this changes, which is part of the big source of size churn here.
I don't think testing in advance is a feasible approach in general. See for example the rpi series which hasn't got any comment. It likely won't until it is applied. That's how we get feedback. We have months to resolve issues and I believe that the code is fundamentally sound.
We need some spot testing here and there to see how things react and how people use it. You found a lot of things with just rk3399, and now the rest of rockchip looks to be fairly direct. You found more things doing x86. When you can convert some other SoC and the change is just dropping distro_bootcmd and calling bootflow scan instead (or that + setting the order again), that'll be good.
But I'm not sure that changing the platforms that don't today opt-in to distro_bootcmd (which has been a thing for a long time now) to force opt-in to this is the right call. It might be in some cases (mediatek maybe? Or maybe no, everyone does android of some flavour so a different bootstd option) but not others (those *_evm_r5_defconfig boards).
What churn are you seeing? Do you mean:
disable BOOTSTD for boards with custom commands? You asked for that patch disabling BOOTSTD_DEFAULTS? You asked for that patch enable BOOTSTD_DEFAULTS by default? We can drop it if you like
We need all 3 of those patches because without the 3rd you don't get a good experience when you do enable bootstd for a platform. The problem is that unless you have distro_bootcmd today you're getting between 63kB (a lot of mediatek platforms) and 5k (silk, as a semi-random example) of growth because bootstd is on and now is the default bootcmd, when before they had nothing. And probably had board docs saying "now do ... to boot". And that's largely setting aside the *_r5_* platforms that I know are just doing something else, and could disable it.
Er, I thought you wanted it to default on if the boards has no bootcmd? If not, we can disable it for those as well. If you don't want any increase we can disable it for boards without DISTRO_DEFAULTS too. After all, presumably those boards are doing something custom anyway.
I think I might have originally, but now that I'm looking at the results it was too optimistic. Every branch I build I look at the per-board breakdown, not just the summary. And it's too much on all of these platforms that had no default bootcmd today.
We want to convert everyone doing distro_bootcmd over to this, that's good. The problem is we don't have a symbol today that means "we want distro_bootcmd" and also isn't overloaded (DISTRO_DEFAULTS is overloaded in this sense).
The wrong direction part of this series is that for platforms that aren't in the middle of converting we're increasing their size between somewhat and very very much, and we haven't tested that it'll work. And yes, there's some automatic guessing logic, which hasn't been tested on these platforms either, so we don't actually know if going from no bootcmd (and so drops to prompt) to attempts to autoboot something is an improvement.
So, the wrong direction comes from the last three patches. Is that right?
The wrong direction comes from enabling bootstd (and so, a bootcmd and distro boot) on platforms that just sit at the prompt today. We are not making a good heuristic guess at what the should be doing. Reaching out to the maintainers to get them to do the conversion, especially once it's just a matter for most of them of just enabling bootstd and then distro, if they want that or once bootmeth android gets done, for those platforms) is how we get them moved smoothly.
Fundamentally the problem I have is that I know where I would like this to head, which is everything using standard boot and turning down the scripts. But it feels like every time I touch bootstd we have to have the EFI discussion again. You can imagine how I feel about disabling BOOTSTD by default...it would basically kill it.
Well, we enabled bootstd by default too quickly perhaps then, and just like we narrowed down EFI_LOADER defaults, we need to narrow this down until it's easily convertable.
This is not really an arch-specific thing, nor an SoC-specific thing. The underlying logic is the same for everything. The reason I think we need to do a few cases before we enable it everywhere is that we need to find the little tweaks needed in that logic.
It's a generic framework to a board specific thing.
How about we apply the first patches in this series, skipping the last three, then apply the rpi series as well. That should get people actually using it and we can iron out the problems. It also keeps things moving. We have months before the release.
Enabling by default can come later once we decide what we want to do about size increases, boards that don't use DISTRO_DEFAULTS and boards that don't have a boot command.
How about disabling it by default and imply'ing it for everything that implies/selects DISTRO_DEFAULTS today, since the part of bootstd that is done is the distro bootmeth?