[OpenWrt-Devel] [PATCH 01/13] build metadata: Allow to build a subset of profiles in a single build

Felix Fietkau nbd at openwrt.org
Tue Jan 19 06:45:34 EST 2016


On 2016-01-19 12:21, Daniel Dickinson wrote:
>>> +# $(1) macro
>>> +# $(2) macro parameter
>>> +define BuildForProfiles
>>> +  $(foreach profile,$(PROFILES_BUILD),$(call $(1),$(2),$(profile)))
>>> +endef
>> Please don't call all the legacy Image/* code once per profile.
>> This will generate unnecessary duplicate files and extra build time.
>> If you rework targets to just stop using $(PROFILE) and check the
>> profile enabled/disabled variables directly, you don't need this at all.
> 
> I think doing that, particularly on targets such as ar71xx, turns this 
> into a much broader reworking of profile/image system than I had planned 
> on undertaking, or thought belonged in a single patch series.  See below.
I don't think it takes that much reworking, especially since targets
like ar71xx push so many build aspects through a small number of templates.
Since selecting "Default" already builds all of those images in one run,
I don't see any reason for adding multiple Image/Build/* calls per
profile. It's just a matter of reworking the target specific templates
in the right way.

>> Also, the types and flags are rather confusing here. From the metadata
>> processing point of view, there should be nothing special about profiles
>> named 'Default' and 'Minimal' aside from a few simple distinctions:
>> There should be one profile that is enable by default
>> There should be a distinction between device profiles and profiles that
>> build all (or most) devices.
>> If you write the code to clearly handle the above, this should make the
>> rest of your patch less confusing.
>>
>> Now that I read your patches again, I get the sense that I disagree with
>> part of your changes on a conceptual level. I think giving the
>> impression that selecting multiple entries in the profile list will
>> 'Build multiple profiles' seems wrong to me, making the description
>> 'Build all profiles' even more wrong. The code is not doing that,
>> because that would imply having a different package selection for every
>> profile build, and the build system can't do that.
>>
>> I think a cleaner description would be building for all devices or a
>> subset of the devices on the list, and just not feed the impression that
>> it has that much to do with 'profiles' at all.
>>
>> Essentially, this is nothing but a simple way to make an image that is
>> compatible with a specific set of devices, and we should present it to
>> the user that way.
> 
> I the confusion stems from that fact that in the past selecting a 
> profile was how you picked what images to build, and so I thought of a 
> profile as an 'image-selector' rather as in terms of the package sets 
> the profiles select.
> 
> In fact as I pointed out in another email the profiles on the targets 
> that have 'Default' or 'Generic' targets that build images for all 
> profiles for the arch_subtarget end up all using the same rootfs, but do 
> still generate distinct images, or in the case of brcm47xx one has a few 
> sets of generic profiles that build multiple images each.
> 
> There are actually other moving parts to the 'profile/image' equation 
> (which is my on ar71xx most profiles correspond to a single device's 
> images), which include the per-image kernel and bootloader 'magic' (and 
> I think if the correct terminology is used, the kernel and 'magic' to 
> get images accepted by the bootloader is a per-image (or image family) 
> piece rather than a per-profile piece.
> 
> There perhaps should really be an additional tier in the build selection 
> so that instead of only talking about selecting profiles, one selects 
> from generic profiles and then makes a sub-selection of images to build 
> for a particular profile, and to get rid of the number of board-specific 
> 'profiles' that are each the same as the other except they invoke the 
> magic for a specific board.
I think that makes sense, but that kind of rework doesn't have to be
done as part of the patch series we're discussing here. Let's just defer
that piece of the discussion, since it would lead to much more intrusive
changes.

> Some care needs to be take with platforms such as sunxi because of e.g. 
> board-specific u-boot.  There is no reason the uboots can't all be 
> built, but one probably doesn't want to build uboot for boards one is 
> not using.
Then let's just leave out multi-profile selection for sunxi for now.
We should start simple and mostly stick to the platforms that matter
most, e.g. ar71xx, ramips, etc.
I think multi-profile selection should simply be opt-in per target.

- Felix
_______________________________________________
openwrt-devel mailing list
openwrt-devel at lists.openwrt.org
https://lists.openwrt.org/cgi-bin/mailman/listinfo/openwrt-devel



More information about the openwrt-devel mailing list