DORA RFC: FILESOVERIDES in dora
We have an issue in the dora branch at the moment. When I merged in the set of bug fixes post release, I added:
http://git.yoctoproject.org/cgit.cgi/poky/commit/?h=dora&id=0bd63125c3b44a656e44f2a76cc5f832c9db4bbd
which wasn't entirely intentional. It changes the behaviour of FILESOVERRIDES. The bug there is real with "arm" overrides being preferred to "armv7a" for example and other strange/incorrect ordering. It was in fact impossible to override certain things which people needed to in some circumstances and defeated the purpose of OVERRIDES.
I don't doubt this change is correct in master, the question is whether it should be in dora. The issue is that some layers don't work correctly before/after the change and we now have a bit of a confused situation.
Our options are either:
a) Revert the change in the branch. Anyone who adapted to the correct behaviour on the branch would need to revert those changes
b) Keep this change in there and adapt layers to fix any incorrect behaviour.
I'm open to either option but I think this needs discussion before we do anything. I'd appreciate other people's views, particularly those of layer maintainers.
Cheers,
Richard
Quoted 2 lines · Richard Purdie
a) Revert the change in the branch. Anyone who adapted to the correct behaviour on the branch would need to revert those changes
As there hasn't been a .1 of Dora yet this only impacts people who have been tracking git, so I think it's probably best to apologise for this merge and revert it, on the grounds that it's a serious behaviour change.
Ross
Quoted 9 lines · Ross
Quoted 2 lines · Richard Purdie
a) Revert the change in the branch. Anyone who adapted to the correct behaviour on the branch would need to revert those changes
As there hasn't been a .1 of Dora yet this only impacts people who have been tracking git, so I think it's probably best to apologise for this merge and revert it, on the grounds that it's a serious behaviour change.
I agree with Ross here.
Signature · 3 lines
Otavio Salvador O.S. Systems http://www.ossystems.com.br http://code.ossystems.com.br Mobile: +55 (53) 9981-7854 Mobile: +1 (347) 903-9750
Quoted 10 lines · Ross
Quoted 2 lines · Richard Purdie
a) Revert the change in the branch. Anyone who adapted to the correct behaviour on the branch would need to revert those changes
As there hasn't been a .1 of Dora yet this only impacts people who have been tracking git, so I think it's probably best to apologise for this merge and revert it, on the grounds that it's a serious behaviour change.
I'm afraid that the problem is it doesn't work well after revert the patch, either, as the patch says:
Currently the overrides are being applied backwards. This means something which is platform specific is overriding something which is machine specific which is clearly not intended.
I don't know how many layers are affected, I've tested the current dora branch in the local autobuilder, didin't find any errors.
// Robert
Quoted 3 lines
Ross
Quoted 2 lines · Robert Yang
I'm afraid that the problem is it doesn't work well after revert the patch, either, as the patch says:
Sure, but Dora's been released and tested: changing the behaviour there in a branch means huge amounts of validation will need to be done. Just how long has the order been wrong?
Ross
Quoted 9 lines · Ross
Quoted 2 lines · Richard Purdie
a) Revert the change in the branch. Anyone who adapted to the correct behaviour on the branch would need to revert those changes
As there hasn't been a .1 of Dora yet this only impacts people who have been tracking git, so I think it's probably best to apologise for this merge and revert it, on the grounds that it's a serious behaviour change.
I agree with Ross, it would be really bad to ask people if they use dora as 1.5, 1.5.1 or what exact revision between them when debugging some FILESPATH related issue.
BTW: https://wiki.yoctoproject.org/wiki/Releases still says Dora is in development, that note should have been dropped when links to 1.6. were added.
FWIW: I'm not sure if I like reversed for loops, commit says "give the result the user would expect", maybe I'm just used to old behavior, but I expect e.g. version specific patches to be preferred over DISTRO specific patch in "files" directory.
Imagine hypothetical case:
conman_123.bb:file://static-configuration.patch connman/qemuall/static-configuration.patch # something qemu specific to keep NFS root mounted connman/static-configuration.patch
someone adds new connman with different configuration format connman_345.bbconnman-345/static-configuration.patch # new format required for 345 and we don't need qemuall to override it anymore
With for loops reversed connman_345 will use old connman/qemuall/static-configuration.patch
Someone can show similar case where the reversed order makes sense.
But I think that people will usually notice version change and if needed update distro/machine specific files to provide their own variant for new recipe version.
E.g. if connman gets broken between 2 images I'll easily see in installed-packages.txt that connman was upgraded end quickly find that our specific .patch file wasn't used anymore. But I won't notice that someone added new file to qemuall which caused only EXTENDPRAUTO change and rebuild with different file.
Is it right thing to change behavior people are used to? Especially if it doesn't fix any real-world issue (Not talking about changed order of overrides which makes sense to me).
Regards,
Signature · 1 line
Martin 'JaMa' Jansa jabber: [email protected]