On 25/04/2023 22.09, Tom Rini wrote:
On Tue, Apr 25, 2023 at 10:02:01PM +0200, Rasmus Villemoes wrote:
On 25/04/2023 21.31, Tom Rini wrote:
On Fri, Mar 17, 2023 at 11:26:39AM +0100, Rasmus Villemoes wrote:
Now, the only way to be really sure is to build the world with/without this patch and check if any .dtb file changes, but I don't have the means to do that.
So, yes, this causes a bunch of fail to builds, as you noted above. The easiest way I think to confirm things before / after would be to make a quick change to tools/buildman/builderthread.py and self.CopyFiles line for keep_outputs to also keep the dtb or some dts files so you can diff before / after to make sure the end result is the same.
Do the builds outright fail, or do they fail in the sense that some machinery detects a change in the binary artifacts? Can you point me at one or two CI builds that show this?
They outright fail to build, mt8516_pumpkin is the one I was testing with.
Gah, of course.
dtb-$(CONFIG_ARCH_MEDIATEK) += \ mt7622-rfb.dtb \ mt7623a-unielec-u7623-02-emmc.dtb \ mt7622-bananapi-bpi-r64.dtb \ mt7623n-bananapi-bpi-r2.dtb \ mt7629-rfb.dtb \ mt7981-rfb.dtb \ mt7981-emmc-rfb.dtb \ mt7981-sd-rfb.dtb \ mt7986a-rfb.dtb \ mt7986b-rfb.dtb \ mt7986a-sd-rfb.dtb \ mt7986b-sd-rfb.dtb \ mt7986a-emmc-rfb.dtb \ mt7986b-emmc-rfb.dtb \ mt8183-pumpkin.dtb \ mt8512-bm1-emmc.dtb \ mt8516-pumpkin.dtb \ mt8518-ap1-emmc.dtb
means that we end up building a million .dtbs that are not actually relevant to the board we're building for, and the value of CONFIG_SYS_BOARD==mt8516 is of course completely inappropriate for mt7622-rfb.dtb, and the nodes mentioned in mt8516-u-boot.dtsi don't exist in mt7622...
[To add insult to injury, it seems that currently mt8516-u-boot.dtsi is not actually being included when building mt8516-pumpkin.dtb, but it seems that the intention very much is that it should be - except that mt8516-u-boot.dtsi has a typo (it refers to a label topckgen_ , but the trailing underscore shouldn't be there) - confirming that it does in fact not get used.]
I wonder why this isn't already a problem, but I guess that in practice we never hit the CONFIG_SYS_CPU or CONFIG_SYS_VENDOR cases.
Not sure what to do. I think it's a little counterproductive to build all these .dtbs when they are not needed, and silly to have to add one's .dtb to some semi-random list - which is why I pushed for 3609e1dc5f4d to get in.
And since we very much allow the .dtbs to depend on various CONFIG_ settings - both because different .configs can cause different -u-boot-dtsi files to get included, and also because we allow direct use of CONFIG_* values (or in #if, #ifdef), there's no guarantee that the mt7629-rfb.dtb built with mt8516_pumpkin_defconfig is identical to the one built with mt7629_rfb_defconfig. So what exactly is the point of building all those irrelevant .dtbs?
So obviously my patch cannot go in as-is. But I do think there are some things that need to be rethought in our build system.
Now that all of CONFIG_OF_LIST gets built automatically, couldn't we delete most if not all of the dtb-$(CONFIG_SOMETHING) += stanzas?
Or to make the build of those extra dtbs a little more deterministic and fix the issue with a random/wrong/inapplicable -u-boot.dtsi being picked up, perhaps change the %.dtb rule so that the the magic *-u-boot.dtsi file (whichever one applies) is only included when $@ is in CONFIG_$(SPL_)OF_LIST?
Rasmus