[pacman-dev] architecture warn/check

[Dan McGee]

On Mon, Jul 20, 2009 at 5:28 PM, Aaron Griffin<aaronmgriffin at gmail.com> wrote:
> On Mon, Jul 20, 2009 at 5:24 PM, Dan McGee<dpmcgee at gmail.com> wrote:
>> On Mon, Jul 20, 2009 at 5:12 PM, Aaron Griffin<aaronmgriffin at gmail.com> wrote:
>>> On Mon, Jul 20, 2009 at 4:33 PM, Dan McGee<dpmcgee at gmail.com> wrote:
>>>> On Mon, Jul 20, 2009 at 3:38 PM, Steven Blatchford<sblatchford at gmail.com> wrote:
>>>>> Hi Dan,
>>>>>
>>>>> I'm sure this has been brought up in the pacman ML but I couldn't find
>>>>> it quickly.  Do you think it would be useful to check the architecture
>>>>> of the machine (eg the output of 'uname -m') against the binary pacman
>>>>> is downloading?  Twice I've sync'd the file /etc/pacman.d/mirrorlist via
>>>>> unison to my slicehost server from my i686 network.  The latest bash4.0
>>>>> upgrade hurt... like there were tears... and henceforth it's now known
>>>>> in my house as "Grumpy Sunday".
>>>>>
>>>>> I have no trouble creating a wrapper script, I just thought I'd toss it
>>>>> out there.
>>>>>
>>>>> Lastly, if you suggest I go the wrapper script method, besides trying to
>>>>> parse the mirrorlist file, is there a nice way to get the architecture
>>>>> of a file from pacman before it downloads it? /installs it?
>>>>
>>>> Would you mind sending this to the pacman-dev ML or filing a bug
>>>> report instead next time? Unfortunately it will just get buried in my
>>>> personal email inbox. I'm copying the list on this response.
>>>>
>>>> With that said, I think we could perhaps take some precautions for
>>>> such things, such as adding a pacman.conf option to verify the
>>>> architecture. Something such as:
>>>>
>>>> RootDir = /
>>>> DBPath = /var/lib/pacman
>>>> Architecture = x86_64
>>>>
>>>> Where the accepted options would be something like:
>>>>
>>>> Architecture = { i686, x86_64, ppc, etc... } or "auto", which would
>>>> make a uname system call, check the machine[] field, and use that
>>>> instead of a value being hardcoded?
>>>>
>>>> What does the rest of the list think? This wouldn't be too hard, and
>>>> of course a package coded with architecture "any" would get a free
>>>> pass.
>>>
>>> Yeah, I definitely don't think using "uname -m" by default should be
>>> done - what happens if I booted and i686 livecd to I could recover
>>> something borked on my x86_64 machine? "Can't install package, wrong
>>> arch" Grrr. Sure, you could use "linux64" in this case, but if you're
>>> already chrooted to a live system that's nicely configured, this extra
>>> step shouldn't be needed.
>>>
>>> I don't think "auto" should be a setting though - I think it should
>>> only be used if Architecture isn't found in pacman.conf and should
>>> output a warning saying "Architecture not set in pacman.conf, using
>>> <blah>"
>>
>> I'm going to disagree with this- my default was going to be "don't
>> check" if left unset. However, I could go either way as long as both
>> "auto" and "nocheck" are somehow accommodated.
>
> Ah, I wasn't actually thinking about the ability to circumvent the
> check (why would you want to?) - could you explain when this would be
> useful? Most cases where uname != (what you want) are probably in a
> chroot and you have a pacman.conf there anyway.

1. Making something impossible is never good (this is mostly the
"unforeseen difficulties" excuse)
2. "probably" leaves a lot of wiggle room
3. If your name is Allan McRae (or anyone else) and you run an x86_64
kernel in an i686 userspace
4. If your name is <whoever> and you run random i686 package on a
mostly-x86_64 machine

Probably more, and some of these are weak, but there is enough of a
reason to allow it that I think it would be silly to lay down the law
for people that may need to circumvent the check.

-Dan