Hi Masahiro,
On 15 January 2015 at 12:10, Masahiro YAMADA yamada.m@jp.panasonic.com wrote:
Hi
2015-01-15 23:46 GMT+09:00 Simon Glass sjg@chromium.org:
Hi,
On 14 January 2015 at 01:18, Alexey Brodkin Alexey.Brodkin@synopsys.com wrote:
Hi Simon, Masahiro-san,
On Tue, 2015-01-13 at 20:18 -0800, Simon Glass wrote:
Probably I'm missing details of our Kconfig migration plan if one exists. Then I'd like to get a reference to the plan so I'm not attempting to do things that are already scheduled and could be even in a process of implementation. Otherwise if there's no current plan for Kconfig migration we may start discussion on how to deal with "commands" in particular.
I proposed the following way, but no big conversion movement has happened yet. http://thread.gmane.org/gmane.comp.boot-loaders.u-boot/199965/focus=200089
If you have an idea, please propose it.
Unfortunately I cannot open this link - I see "Connection to 80.91.229.5 failed". Do you mind to provide a link for the same message in http://lists.denx.de ?
Kconfig conversion is a too big task to be done by a single indivisual. Any suggestion, any form of contribution is very appreciated.
Indeed and that stops me from starting this work as well. So my point is:
- anybody may do updates to Kconfig like add more U-Boot commands,
some settings (CONFIG_SYS_xxx) 2) board maintainers are responsible for updates in defconfigs and corresponding headers in "include/configs/" 3) that would be helpful if at (1) board maintainers are notified via direct emails on the change that has been done
I personally keep away from any global changes until non-generic boards are dumped.
As Tom said in the following mail, http://lists.denx.de/pipermail/u-boot/2015-January/201032.html we are going to remove lots of boards.
I do not want to make extra efforts on non-generic boards.
That makes sense but why don't we start implementation of changes in good boards we know already converted to "generic"? There're even entire arches that switched to "generic boards".
The point is commands are low-hanging fruits in terms of Kconfig migration - there're no extra options and tweaks, once all commands are added in Kconfig (essentially with dependencies etc) we may clean all board headers. The only real problem here is amount of work - lots of headers/defconfigs to patch. But still this is doable.
I realized one problem when I was doing this task for my boards in commit 25e274e20208.
The defconfigs of my boards are almost the same: (configs/ph1_ld4_defconfig, configs/ph1_pro4_defconfig, configs/ph1_sld8_defconfig)
What is inconvenient as for defconfig is that it does not support "include" directive like C headers.
People generally like to add CONFIG_CMD_* to a common header file such as include/configs/tegra-common.h
On the other hand, on defconfig, we must touch each defconfig of the board family.
I have not been able to decide right direction to solve this issue.
I think we need to discuss about how to live with lots of defconfigs.
How about this:
- Support an 'include' option in defconfig to include another file.
This could take the form of CONFIG_DEFCONFIG_INCLUDES (comma-separated list of filenames). Don't allow nested includes - this CONFIG provides a complete list of includes. 2. When you use menuconfig to edit the config, it ignores the include and just writes out everything. Then it runs a simple script which parses the include files, and adjusts the file so that things in the includes are not repeated in the defconfig.
I think this scheme would be good enough, but can probably be improved. At least it would be better than what we have.
Or we could instead define particular includes for different purposes:
- Arch
- SOC
- SOC Vendor
- Board vendor
- Board family
We could then require that standard filenames are used, and search for the correct file based on the SOC/board/vendor setting, etc.
As an example, for trimslice, we could search for arm.h, tegra.h, nvidia.h, compulab.h, some_family_name.h.
This is less flexible though.
Frankly I don't like this approach with post-processing steps. It will inevitably end-up with messed up configs.
Why don't we just use default values in Kconfig for ARCH/SOC/Board? It's pretty obvious that 1 board may have N flavors but then have a baseline options selected in "board/vendor/board_name/Kconfig" and only put options that differ between boards in your defconfigs.
This way "savedefconfig" will automatically strip down all extra lines for a particular board.
This is how things work in Linux kernel and Buildroot Kconfig-based build systems. Probably I'm missing something here because U-Boot differs from mentioned projects in some aspects - if so please correct me.
I started with this approach and Masahiro was not very keen on it. I'm OK with it, particularly as it is already supported, but I wonder whether we can do better.
Honestly, I do not like baseline options in board-Kconfig very much. The advantage is that it works without any change.
What I suggested before was to use scripts/kconfig/merge_config.sh.
For example, put the SoC baseline options into tegra30_defconfig. Put the difference from that into tegra30_(board).config For example,
make tegra30_defconfig tegra30_seaboard.config make tegra30_defconfig tegra30_trimslice.configThe disadvantage is that we would have to input more for the configuration and has some impact on other projects such as BuildRoot...
Is there any way that tegra30_seaboard could include tegra30 automatically?
Another option is "multi-image". Actually barebox adopted this solution to solve the increasing defconfig problem.
In barebox, for example, all the Tegra boards share a single defconfig, "tegra_v7_defconfig"
tegra_v7_defconfig creates images for beaver, jetson, colibri, ac100, ...
It takes advantage of Device Tree configuration and garbage collection. So, it generates multipule images without increasing memory foot-print.
I've been keen on that approach and have taken some steps for Tegra and Exynos. Exynos is actually not to far away, but Tegra 124 pinmux approach makes this hard.
Still I think this is the best solution - fewer boards to build also.
I notice that some kernel distributions have a script which pulls out common kernel configs in a separate file (e.g. for ARM, or for Tegra) for checking into source control, then combine them again for the kernel. This suggests that others have this problem too. It does seem like a convenient feature to be able to have some sort of hierarchy of config.
Re messed-up configs, what will mess up? We can make it part of writing the config, perhaps.
The hierarchical config strategy makes "savedefconfig" difficult , I think.
I never said it would be easy. Kconfig wasn't easy.
That's a separate topic. I do agree that CMD_FPGA makes not much sense for most of boards, as well as some others.
And I think during migration process of commands to Kconfig it's a good time to reconsider default commands.
Agree. We should reconsider the default. In my opition, most of the ones in config_cmd_default.h are not default commands.
Right but if we expect to switch to Kconfig solely for commands selection we may not care much about contents of "config_cmd_default.h" assuming it will go away sometime soon.
Probably we may just focus on:
- Add all existing commands in Kconfig
- Select very few of them to be globally default
(1) is pretty simple and anybody may do it if time permits (2) requires wide discussion
But what's really nice with Kconfig is its "savedefconfig". So we may not set any commands as globally default and let every architecture, board or defconfig to select really required options.
Then once any option is marked as global (or even arch or board) default in a matter of simple script that does "make XXX_defconfig && make savedefconfig && mv defconfig configs/XXX_defconfig" all relevant defconfigs will be correctly updated (default options will be stripped out).
Yes, it would be great to get all command configs into Kconfig.
I wonder if we need a script which automates the conversion? The header file includes in the include/configs/... files makes this non-trivial. But if we had a script which could take a CONFIG as a parameter, remove it from all config headers, and add it to relevant defconfigs, it would be very useful.
I have a Python script that moves each CONFIG from header files to defconfig files.
I wrote it for my local use, so it might not be very clean.
If necessary, I can share it on ML.
Yes please!
- Simon