[OE-core] [PATCH 1/1] opkg 0.1.8: respect to the arch when choose the alternatives

[Koen Kooi]

Op 4 jun. 2012, om 12:39 heeft Martin Jansa het volgende geschreven:

> On Mon, Jun 04, 2012 at 05:31:26PM +0800, Robert Yang wrote:
>> 
>> 
>> On 06/01/2012 06:35 PM, Koen Kooi wrote:
>>> 
>>> Op 1 jun. 2012, om 12:02 heeft Richard Purdie het volgende geschreven:
>>> 
>>>> On Fri, 2012-06-01 at 11:04 +0200, Koen Kooi wrote:
>>>>> Op 1 jun. 2012, om 10:17 heeft Richard Purdie het volgende geschreven:
>>>>> 
>>>>>> On Thu, 2012-05-31 at 17:01 +0200, Koen Kooi wrote:
>>>>>>> Op 31 mei 2012, om 16:13 heeft Robert Yang het volgende geschreven:
>>>>>>> 
>>>>>>>> There is a bug if we:
>>>>>>>> 1) bitbake core-image-sato-sdk MACHINE=qemux86
>>>>>>>> 2) bitbake core-image-sato with MACHINE=crownbay
>>>>>>>> 
>>>>>>>> Then several pkgs in deploy/ipk/i586 would be installed to crownbay's
>>>>>>>> image even if there is one in deploy/ipk/core2 and we have set the
>>>>>>>> core2's priority higher than i586, when the version in deploy/ipk/i586 is
>>>>>>>> higher. This doesn't work for us, for example, what the crownbay need is
>>>>>>>> xserver-xorg-1.9.3, but it installs xserver-xorg-1.11.2.
>>>>>>> 
>>>>>>> And this is working exactly as intended. Don't break opkg because your
>>>>>>> hardware driver situation sucks.
>>>>>>> 
>>>>>>> So: NAK on this patch.
>>>>>> 
>>>>>> I think we do have a problem here. For example, the system is ignoring a
>>>>>> PREFERRED_VERSION directive here
>>>>> 
>>>>> A PREFERRED_VERSION set in a machine.conf, which is the wrong place.
>>>> 
>>>> Its strongly not recommended. You can do it and if things are setup
>>>> correctly, we do support machine specific alterations however...
>>>> 
>>>>> Let's change the above build sequence to this:
>>>>> 
>>>>> 1) MACHINE=qemux86 bitbake xserver-xorg
>>>>> 2) MACHINE=othercore2machine bitbake xserver-xorg
>>>>> 3) MACHINE=crownbay bitbake xserver-xorg
>>>>> 
>>>>> Oh look, I get 1.11.2 on crownbay regardless of this patch.
>>>> 
>>>> I've been assuming this xserver is marked as machine specific. If its
>>>> not, that is a bug and we can fix that.
>>> 
>>> the X server not machine specific and that is NOT a bug. The recipe for the DDX only works with a specific X version. The real problem here is that xf86-video-something (again, not machine specific, think pci cards) has a strict requirement on the x server version. This is now purely a DISTRO problem. A few possible solutions are:
>>> 
>>> 1) Lock X to 1.9.3 for all compatible x86 archs in your DISTRO
>>> 2) Lock X to 1.11.2 for all compatible x86 archs in your DISTRO, live with the crownbay breakage
>>> 3) Only use a package manager with no notion of compatible archs (dpkg) or with strong version pinning (rpm)
>>> 4) Be really careful with the build sequence and structure your feed configs to not have compatible archs in their feed URIs
>>> 5) remove 'x11' from DISTRO_FEATURES
>>> 
>>> This is a point where you cannot make everybody happy. But as you say below, let's decouple the above problem from the arch vs version discussion.
>>> 
>>>>>> by building one thing and then
>>>>>> installing another. We're also inconsistent between the dpkg/rpm and
>>>>>> opkg backends. There is therefore definitely some kind of user
>>>>>> experience issue at stake here since this behaviour is not obvious,
>>>>>> expected or particularly correct.
>>>>>> 
>>>>>> The fact the example is hardware related is not particularly relevant,
>>>>>> its the bigger picture I worry about
>>>>> 
>>>>> I also worry about the bigger picture, especially about what Martin
>>>>> said about this breaking feeds.
>>>> 
>>>> As far as I understood Martin's comments, this actually helps avoid some
>>>> of the issues he's been experiencing with feeds?
>>> 
>>> No, it allows you to add a package-of-death. Example:
>>> 
>>> broken-app_1.0_machine.ipk is in the feeds, being machine specific
>>> broken-apps author fixes the machine specific bits properly
>>> broken-app_2.0_i586.ipk hits the feeds and is not picked up.
>> 
>> I think that the broken-app_2.0_i586.ipk came unexpectedly,
>> there was a broken-app_1.0_machine.ipk (maybe also broken-app_1.0_i586.ipk,
>> but it didn't matter), if it has been fixed to version 2.0, then there should
>> be also a broken-app_2.0_machine.ipk (and it would be picked up).
> 
> If broken-app developers are now detecting machine capabilities (or
> whatever) since 2.0 in runtime, then we don't need to build broken-app
> for every single machine we support, so changing PACKAGE_ARCH from
> MACHINE_ARCH to TUNE_PKGARCH is right thing to do to save compile time
> and feed disk space.

Exactly

> Same use-case broken by this patch reported here:
> http://lists.linuxtogo.org/pipermail/openembedded-core/2012-May/023154.html

I borrowed heavily from your post since both Richard and Robert don't seem to see the damage this will do to feeds when a package/recipe changes archs.

regards,

Koen