Helmut Grohne:
Hi Simon and Niels,
(fixed Cc address for architecture-properties maintainers)
On Thu, Aug 01, 2024 at 05:26:04PM +0100, Simon McVittie wrote:
[...]
I talked to Niels off list and he was vaguely positive about riding on
architecture-properties. He had some concerns about the bus factor that
this work might impose on architecture-properties. Guillem also
indicated that he wants to give feedback though not right now.
To me this is sufficient to prototype an implementation. I'm attaching a
patch for architecture-properties that hopefully gets most of the
details right (thanks to Simon's design feedback). If you prefer
reviewing it on salsa, go
https://salsa.debian.org/debian/architecture-properties/-/merge_requests/1
I request that we give people at least one week to reply before
uploading this. I appreciate feedback from Niels as to how acceptable he
finds this contribution.
I think I am good to go on this one. I would rely on someone else to
keep up with the "hacks" required to detect emulations if the current
check stops working and I would not use my self directly, so the tests
mentioned below would be relevant or we do consumer based testing.
Those are the limitations I currently see. The code looks like it would
be relatively stable and small enough that I can throw it directly at
the C compiler maintainers if it ever ICEs or miscompiles.
[...]
Writing tests is a bit non-trivial. Whilst Ubuntu has pioneered an
autopkgtest extension for cross architecture testing, this does not seem
to be available in Debian. Testing the native case is rather boring in
my view. As for manual testing, I verified that :amd64 (natively), :i386
(without installing qemu), and :armhf (qemu) work as expected on amd64.
It would be boring to test the native flow, but it would still assert
that the code minimally works. Therefore, I think we should have that
test case. It would at the very least catch miscompiles where the
c-program just SEGVs due to a miscompile.
For the non-native case, I think we could emulate that by playing with
PATH and use that to mock `readlink` into an `echo /usr/bin/qemu` if we
wanted to test the detection. I think that could also be used to trigger
at least one of the error cases by having the mocked `readlink` return
non-zero?
If we could get proper non-native tests as well, that would surely be
better.
In the event that dpkg gains the Provides suggested by Simon, we can
delete both native-architecture and native-architecture-is.
Helmut
Sounds good to me.
Best regards,
Niels
