Skip to contentSkip to conversation

liburing packaging issues

DISCUSSION Guillem Jover 2 participants 3 messages Feb 8, 2020 – Feb 15, 2020 io_uring

lore
Share & tools
lore
#0 lore raw

Hi!

Stefan Metzmacher asked whether I could package and upload liburing for Debian (and as a side effect Ubuntu). And here are some of the things I've noticed that would be nice to get fixed or improved before I upload it there.

* The README states that the license is LGPL (I assume 2.1+ in contrast to 2.1?) and MIT, but there's at least one header that's GPL-2.0 + Linux-syscall-note, which I assume is due to it coming from the kernel headers. Would be nice to have every file with an explicit license grant, say at least with an SPDX tag, and update the README.

* From the RPM spec file and the debian packaging in the repo, I assume there is no actual release tarball (didn't see on in kernel.dk nor kernel.org)? It would be nice to have one with a detached OpenPGP signature, so that we can include it in the Debian source package, to chain and verify the provenance from upstream.

* The test suite fails for me on the following unit tests:

read-write accept-link poll-many poll-v-poll short-read send_recv

while running on Linux 5.5.0-rc0. I read from the README it is not supposed to work on old kernels, and might even crash them. But it would still be nice if these tests would get SKIPPED, so that I can enable them unconditionally to catch possible regressions and so they do not make the package fail to build from source on the Debian build daemons due to too old kernels, which in most cases will be one from a Debian stable release (Linux 4.19.x or so).

There are a couple of issues with the build system, for which I'll be sending out a couple of patches later.

Thanks,
Guillem

maintainer
#1 lore raw
Quoted 7 lines · Guillem Jover

Hi!

Stefan Metzmacher asked whether I could package and upload liburing for Debian (and as a side effect Ubuntu). And here are some of the things I've noticed that would be nice to get fixed or improved before I upload it there.

Quoted 6 lines

* The README states that the license is LGPL (I assume 2.1+ in contrast to 2.1?) and MIT, but there's at least one header that's GPL-2.0 + Linux-syscall-note, which I assume is due to it coming from the kernel headers. Would be nice to have every file with an explicit license grant, say at least with an SPDX tag, and update the README.

OK, I can clarify that and add SPDX headers.

Quoted 6 lines

* From the RPM spec file and the debian packaging in the repo, I assume there is no actual release tarball (didn't see on in kernel.dk nor kernel.org)? It would be nice to have one with a detached OpenPGP signature, so that we can include it in the Debian source package, to chain and verify the provenance from upstream.

I do generate release tar balls with a detached signature, see:

https://brick.kernel.dk/snaps/

Quoted 11 lines

* The test suite fails for me on the following unit tests:

read-write accept-link poll-many poll-v-poll short-read send_recv

while running on Linux 5.5.0-rc0. I read from the README it is not supposed to work on old kernels, and might even crash them. But it would still be nice if these tests would get SKIPPED, so that I can enable them unconditionally to catch possible regressions and so they do not make the package fail to build from source on the Debian build daemons due to too old kernels, which in most cases will be one from a Debian stable release (Linux 4.19.x or so).

The tests should build fine on any system, they just won't pass on any system. So I don't think this is a packaging issue, don't run them as part of building the package.

Quoted 2 lines

There are a couple of issues with the build system, for which I'll be sending out a couple of patches later.

Thanks!

Signature · 1 line
Jens Axboe
#2 lore raw
Quoted 6 lines · Jens Axboe
Quoted 5 lines · Guillem Jover

Stefan Metzmacher asked whether I could package and upload liburing for Debian (and as a side effect Ubuntu). And here are some of the things I've noticed that would be nice to get fixed or improved before I upload it there.

Quoted 8 lines
Quoted 6 lines

* The README states that the license is LGPL (I assume 2.1+ in contrast to 2.1?) and MIT, but there's at least one header that's GPL-2.0 + Linux-syscall-note, which I assume is due to it coming from the kernel headers. Would be nice to have every file with an explicit license grant, say at least with an SPDX tag, and update the README.

OK, I can clarify that and add SPDX headers.

Thanks for those.

I see now almost all code files are marked as MIT, the man pages as LGPL-2.1 (not clear whether >= or not), and one header with GPL-2.0 with Linux-syscall-note or MIT. But there's no COPYING for the GPL, only for the LGPL, and the README does not reflect the above. Could you either add the missing COPYING if necessary, clarify the license for the files, or the README and RPM spec so that they are consistent (I guess you might want to do that on the debian/ directory too, but it's not a problem for the Debian packaging as we'll be overriding it completely anyway).

I'm sorry to be a pain on this, but this is going to be the thing most scrutinized by the Debian ftp-masters on the first upload for manual acceptance, which tends to take a while, so I'd like to get any inconsistencies sorted out before that. :)

Quoted 10 lines
Quoted 6 lines

* From the RPM spec file and the debian packaging in the repo, I assume there is no actual release tarball (didn't see on in kernel.dk nor kernel.org)? It would be nice to have one with a detached OpenPGP signature, so that we can include it in the Debian source package, to chain and verify the provenance from upstream.

I do generate release tar balls with a detached signature, see:

https://brick.kernel.dk/snaps/

Ah great, thanks. It was not obvious before. (Perhaps add a mention to the README too? And another for places where patches can be sent to, say the list and github? :)

Quoted 12 lines
Quoted 11 lines

* The test suite fails for me on the following unit tests:

read-write accept-link poll-many poll-v-poll short-read send_recv

while running on Linux 5.5.0-rc0. I read from the README it is not supposed to work on old kernels, and might even crash them. But it would still be nice if these tests would get SKIPPED, so that I can enable them unconditionally to catch possible regressions and so they do not make the package fail to build from source on the Debian build daemons due to too old kernels, which in most cases will be one from a Debian stable release (Linux 4.19.x or so).

Quoted 3 lines

The tests should build fine on any system, they just won't pass on any system. So I don't think this is a packaging issue, don't run them as part of building the package.

I'd really like to run these both at build-time, and at "install-time" as part of our CI infra (https://ci.debian.net/), so that we can catch possible regressions or similar. Even if they still fail for now, I'd probably still run them at build-time but as non-fatal (like I'm doing with libaio), so that we can have visibility over all our ports.

After the merge requests I sent I'm now only seeing two problems on the above mentioned Linux version (from Debian experimental):

- short-read always fails with "Read failed: 0".
- there seems to be some kind of resource leak or similar, because after running the tests multiple times, almost if not all of them start to fail with -ENOMEM from io_uring_setup().

Thanks,
Guillem