While testing the cni uprev by building in a container with
network=none the following error was found:
go: github.com/Microsoft/go-winio@v0.4.11: Get
https://proxy.golang.org/github.com/%21microsoft/go-winio/@v/v0.4.11.mod:
dial tcp: lookup proxy.golang.org on 128.224.144.130:53:
dial udp 128.224.144.130:53: connect: network is unreachable
After some digging through the go documentation it was found that the
'-mod=vendor' is required for 'go build' to use shipped vendor modules
when building modules. This can be confirmed by look at the
'build_linux.sh' script which is found in the plugins repo.
By using '-mod=vendor' and also ensuring things are properly placed in
the GOPATH (ie $B) we can avoid having to create many of the links we
had been previously.
We also put all the build artifacts into $B to avoid mixing source and
build.
Signed-off-by: Mark Asselstine <mark.asselstine@windriver.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@gmail.com>
Both uprev's are listed as 'minor' in the upstream release
notes. Neither introduces an uprev in spec. This fixes issues we
observed while testing the forthcoming cri-o uprev.
NOTE: this commit should only be used with the follow-on commit [cni:
prevent go from downloading stuff in the background] otherwise you
will end up with files not owned by you which will prevent the recipe
being properly cleaned.
Signed-off-by: Mark Asselstine <mark.asselstine@windriver.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@gmail.com>
Another straightforward uprev with one fairly large change in the
changelog. The Open Virtual Network component has now been moved to
its own repo (https://github.com/ovn-org/ovn.git). If you were using
this functionality a new recipe will need to be created.
The ptest results are similar to after the v2.12 uprev
ERROR: 2206 tests were run,
28 failed unexpectedly.
62 tests were skipped.
The failed tests were in the following areas:
checkpatch.at (5)
ovs-ofctl.at (1)
tunnel.at(1)
tunnel-push-pop.at(3)
tunnel-push-pop-ipv6.at(3)
dpif-netdev.at (1)
pmd.at(1)
ofproto-dpif.at (7)
bridge.at (2)
ovsdb-idl.at(1)
mcast-snooping.at(1)
packet-type-aware.at(2)
None of these affect core functionality or usecases and are similar to
the results we see with v1.12. If specific usecases are affected by
these failures we should address them on a need to fix basis.
Signed-off-by: Mark Asselstine <mark.asselstine@windriver.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@gmail.com>
A mostly straightforward uprev. Unfortunately a required patch for
python3 is only available on a non-release branch so we must carry it
in order to build (the discussion on the mailing list was that an
uprev might have avoided this, but this is not the case).
The ptest results are similar to after the v2.11 uprev
ERROR: 2413 tests were run,
23 failed unexpectedly.
383 tests were skipped.
NOTE, however, that they have now marked many tests as 'skipped', such
as the python2 results, so the failed and skipped numbers have
essentially swapped with each other. The failed tests were in the
following areas:
checkpatch.at (5)
ovs-ofctl.at (1)
dpif-netdev.at (1)
ofproto-dpif.at (6)
bridge.at (2)
ovn.at (2)
ovn-controller-vtep.at (6)
Most were issues with the test or expectations that source code would
be available. There might be an issue around packaging of
"/vswitchd/vswitch.ovsschema" but we should be able to overlook this
for now, as we have with previous versions.
Signed-off-by: Mark Asselstine <mark.asselstine@windriver.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@gmail.com>
Use a standard location to store the cni tools and plugins. This
is more in line how other distributions package cni. Keep a symlink
to /opt/cni/bin for backward compatibility.
Signed-off-by: Stefan Agner <stefan.agner@toradex.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@gmail.com>
slirp4netns allows connecting a network namespace to the Internet
in a completely unprivileged way, by connecting a TAP device in a
network namespace to the usermode TCP/IP stack ("slirp")."
Signed-off-by: Stefan Agner <stefan.agner@toradex.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@gmail.com>
The build has broken again on master, even for non-static builds of
netns. The simplest fix is to extend our existing patch to cover this
case as well.
Signed-off-by: Paul Barker <paul@betafive.co.uk>
Signed-off-by: Bruce Ashfield <bruce.ashfield@gmail.com>
openembedded-core commit fb064356af61 [Remove LSB support] dropped the
'lsb' recipe which caused openvswitch to throw a dependency fail for
both sysvinit and systemd builds. LSB init functions for
log_begin_msg, log_end_msg and others were being used. We now use the
functions from ovs-lib which are part of OVS and supply the remaining
ones directly. This allows us to regain the functionality and drop the
dependency on 'lsb'.
Signed-off-by: Mark Asselstine <mark.asselstine@windriver.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@gmail.com>
Install ovs-systemd-reload as openvswitch.service uses it.
openvswitch.service:ExecReload=/usr/share/openvswitch/scripts/ovs-systemd-reload
Signed-off-by: Chen Qi <Qi.Chen@windriver.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@gmail.com>
Also build plugins under ipam. Other distros default to contain
these plugins.
Signed-off-by: Chen Qi <Qi.Chen@windriver.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@gmail.com>
When setting up k8s, the following error appeared, failing
to setup coredns pod.
NetworkPlugin cni failed to set up pod "coredns-5f7fc64c95-4nnfq_kube-system" network: fork/exec /opt/cni/bin/loopback: no such file or directory
This is because the binary, /opt/cni/bin/loopback, has wrong interpreter.
So fix do_compile to use ${GO} instead of go to fix this problem.
Signed-off-by: Chen Qi <Qi.Chen@windriver.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@gmail.com>
These settings are either in go.bbclass or duplicate, thus
removing them.
Keep only CGO_ENABLED and GOPATH to easier manipulation while
upgrading cni.
Signed-off-by: Chen Qi <Qi.Chen@windriver.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@gmail.com>
When the go-lang container recipes were first created there were issues
with strip and the resulting binaries. As such, strip was inhibited for
the various packages.
This variable is now set in the default classes, and tests show that
strip works on the binaries (saving up to 2M on disk for some binaries)
with no runtime issues found.
So we drop our explicit set of the inhibit and let the build proceed
by the defaults.
If issues are found, we can re-enable the setting or bbappends can
turn it back on for builds showing issues.
Signed-off-by: Bruce Ashfield <bruce.ashfield@gmail.com>
Drop obsolete patches and forward-port the remaining required patch. We
also need to fix up permissions after the build so that we can clean the
build directory without errors if needed.
Signed-off-by: Paul Barker <paul@betafive.co.uk>
Signed-off-by: Bruce Ashfield <bruce.ashfield@gmail.com>
Attempting to build for x86-64 will result in an error:
cannot find package runtime/cgo (using -importcfg)
...go/pkg/tool/linux_amd64/link: cannot open file : open : no such file or directory
As a workaround we switch the build target to 'build' which results in
a slightly smaller executable but does mean that several dynamic
libraries are required to run the generated executable. This shouldn't
be a problem as the build system will ensure the libraries are present
on images where the executable will be included. This change only
means the executable can't be copied as a single entity to a different
machine and run.
We do also patch the Makefile to allow the 'static' target to build
successfully and in the patch commit log we describe the issue in full
(at least in as much detail as we currently have on the issue). But
since the issue isn't fully understood we don't recommend the 'static'
build to be used at this time, thus default to the 'build' target.
Signed-off-by: Mark Asselstine <mark.asselstine@windriver.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@gmail.com>
The v2.11 version fixed a bug as follow.
Error info:
ovs|00002|db_ctl_base|ERR|external-ids:hostname=: argument does not
end in "=" followed by a value.
The result of ptest between v2.11 and v2.10.1 is similar.
v2.11:
ERROR: 2765 tests were run,
317 failed (1 expected failure).
85 tests were skipped.
v2.10.1:
ERROR: 2662 tests were run,
311 failed (1 expected failure).
85 tests were skipped.
I checked the detailed result. The failed tests were mostly related
to python2 as the image only use python3.
Signed-off-by: Hongzhi.Song <hongzhi.song@windriver.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@gmail.com>
This change reduces the length of ${PV} for several recipes and gives us
auto-incrementing version numbers.
Signed-off-by: Paul Barker <paul@betafive.co.uk>
Signed-off-by: Bruce Ashfield <bruce.ashfield@windriver.com>
If we don't explicitly set the GO variable as an argument to make then
the native go toolchain is invoked instead of the cross-toolchain we
actually want. This can result in a netns binary which is built for the
wrong architecture.
Signed-off-by: Paul Barker <paul@betafive.co.uk>
Signed-off-by: Bruce Ashfield <bruce.ashfield@windriver.com>
Updating the cni plugins to match k8s and cri-o. We also filter out
the windows plugins, since they use a different build infrastructure
than the other variants .. and are not currently needed. This can
be revisted in the future as needed.
Signed-off-by: Bruce Ashfield <bruce.ashfield@windriver.com>
Openvswitch recipe depends on the kernel source, which is machine specific.
It can then use this machine specific resource to configure and build the
recipe. This mix causes the hash values of this package to change based on
the machine, so this package must also be a machine_arch package.
Signed-off-by: Mark Hatle <mark.hatle@windriver.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@windriver.com>
The openssl library is an optional component for the openvswitch. The
problem with it enabled by default is that it will consume system
entropy to try to initialize the openssl library even though we are not
using it by default. With the 4.16 kernel and up there is not always
enough entropy available at the early boot time which can cause a
lengthy stall, while waiting to initialize the openvswitch.
If ssl is needed, it can of course be turned on with the package
config option "ssl".
Signed-off-by: Jason Wessel <jason.wessel@windriver.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@windriver.com>
Since `229f4e9 package.bbclass: add support for
pkg_postinst_ontarget()' applied in oe-core, use
pkg_postinst_ontarget to run postinst at first boot.
Signed-off-by: Hongxu Jia <hongxu.jia@windriver.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@windriver.com>
The service currently fails to run since the runtime directories
aren't being created. Create the runtime directories and fixup the
path to echo to get things working again.
Signed-off-by: Mark Asselstine <mark.asselstine@windriver.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@windriver.com>
The current version (v2.9.3) started to fail to build due to changes
in openssl. Upstream must have run into similar issues as this was
identified and addressed in v2.10.0 by including a copy of
'dhparams.c' instead of relying on it being generated. (see commit
"dhparams: Add pregenerated .c file to the repository.")
Additionally v2.10.0 is better aligned for our kernel and dpdk
versions:
kernel
===
2.10.x 3.10 to 4.17
dpdk
===
2.10.x 17.11.3
The changes in this commit allow for the building of v2.10.0. A second
commit will be necessary to deal with a few runtime changes which
prevent ovsdb-server from starting correctly.
Signed-off-by: Mark Asselstine <mark.asselstine@windriver.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@windriver.com>
Updating the latest plugins.
We have to make some minor build adjustments, but otherwise, this is
a simple update.
We also drop the ptest build, since it is not used and is causing
build errors.
Signed-off-by: Bruce Ashfield <bruce.ashfield@windriver.com>
commit 1fd1ff3720 [openvswitch: uprev to v2.9.2] left a bunch of
unused patches in place. Remove unreferenced patches as they are no
longer needed.
Signed-off-by: Mark Asselstine <mark.asselstine@windriver.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@windriver.com>
This brings us up to date with the latest upstream release. I suspect
there will be an upcoming release to add compatibility with DPDK
v18.05 (the DPDK release is due out any day now) but getting this
fairly large release bump out of the way first will facilitate any
upcoming uprev.
We are able to drop they python3 patches as they have been merged
upstream. Some scripts which needed to be updated to use python3
disappeared, new ones appeared so the 'use python3' patches are
updated accordingly. Beyond this the biggest change is related to the
systemd unit files, the ovsdb-server has been updated upstream to be
generated on the fly via the spec file, we mimic this in the
install_prepend. We also add the various configuration files which the
unit files source before launching the services.
As usual this was tested against out typical usecases including usage
in meta-overc. As well the ptests have been run and the results are no
better or worse.
Previous version:
ERROR: 2332 tests were run,
21 failed unexpectedly.
3 tests were skipped.
New version:
ERROR: 2527 tests were run,
29 failed unexpectedly.
3 tests were skipped.
Signed-off-by: Mark Asselstine <mark.asselstine@windriver.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@windriver.com>
The netns project has been moved into the 'genuinetools' organisation and so
URLs have been updated. The copyright line in the license file has been updated
to reference "The Genuinetools Authors".
The patch name has been updated to make it more suitable for submission
upstream.
Signed-off-by: Paul Barker <pbarker@toganlabs.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@windriver.com>
Refresh patches with devtool command to fix do_patch warning.
Drop CVE-2017-9263.patch since it had been fixed upstream.
Signed-off-by: Yi Zhao <yi.zhao@windriver.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@windriver.com>
Several python3 packages have been removed as the functionality has
been moved into python3 core. As such we can no longer RDEPENDS on
these packages, nor is there a need to.
See similar commits in meta-openembedded:
df997e0747de - Remove deprecated python3-subprocess from the RDEPENDS
90b1f996af78 - Remove deprecated python3-argparse from the RDEPENDS
ef52e9980433 - Remove deprecated python3-lang from the RDEPENDS
This is fallout from openembedded core commit:
54ac820b8a63 - python3: Restructure python3 packaging and replace it
with autopackaging
Signed-off-by: Mark Asselstine <mark.asselstine@windriver.com>
Eventually we will have to move to OVS 2.8.x or 2.9.x to support the
version of DPDK in meta-dpdk but before we make a version jump(s)
let's uprev to the latest 2.7.x stable release.
Signed-off-by: Mark Asselstine <mark.asselstine@windriver.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@windriver.com>
The makefile for netns needs to be patched to ensure that we use the binaries
from go-cross and not go-native.
Signed-off-by: Paul Barker <pbarker@toganlabs.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@windriver.com>
Many distros make a 'kubernetes-cni' package available. This is no
more than a binary extracted set of 'cni' plugins. We already have
'cni' available in the build system, so we make sure it rprovides
kubernetes-cni to keep conflicts and confusion at a minimum.
Signed-off-by: Bruce Ashfield <bruce.ashfield@windriver.com>
Introducing the CNI framework to allow platforms such as
kubernetes + cri-o to use thm for networking configuration.
The networking plugins are installed, but are not configured
by default to create any specific networking interfaces, which
are the responsibility of the image and distribution.
Signed-off-by: Bruce Ashfield <bruce.ashfield@windriver.com>
Recently in the oe-core the go.bbclass changed and requires the
defition of the GO_IMPORT variable. This was intended to simplify how
the compilation works with go packages and it is still a work in
progress.
This patch set makes the recipes compatible to generate the same end
result as before using the new go.bbclass from oe-core.
Any patches that were included in the recipes had to have the paths
adjusted because the new go.bbclass manipulates the notion of S to be
S + "src" + "$GO_IMPORT" internally for the purpose of unpack, patch
and compile.
Signed-off-by: Jason Wessel <jason.wessel@windriver.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@windriver.com>
If you have in your $CC from the compiler template any kind of
directive which contains a comma character it is going to cause ovs to
not build due to the macro expansion in the Makefile:
CC = $(if $(C),env REAL_CC="'"$CC"'" CHECK="$(SPARSE) -I $(top_srcdir)/include/sparse $(SPARSEFLAGS) $(SPARSE_EXTRA_INCLUDES) " cgcc $(CGCCFLAGS),'"$CC"
Put more simply:
CC = $(if $(C), CONFIGURE_CC_DEFINITION, cgcc args CONFIGURE_CC_DEFINITION)
The CONFIGURE_CC_DEFINITION for the i585-nlp definition had
"-Wa,-momit-lock-prefix=yes". The comma there caused a strings to
"break" in the wrong place in the $(if) macro and the build of recipe
will fail with the error shown below:
/bin/sh: -c: line 1: unexpected EOF while looking for matching `"'
/bin/sh: -c: line 3: syntax error: unexpected end of file
The solution is to transform the "-Wa," -> "-Xassembler ".
Signed-off-by: Jason Wessel <jason.wessel@windriver.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@windriver.com>
Openvswitch has been created to search for and use both py2 and py3,
with a preference for py2. There is no configure option to force the
use of py3 only, and many of the scripts only use the value of PYTHON
and not PYTHON3 (again there is a general preference for py2). In
order to force this issue we set PYTHON to 'python3'. This isn't ideal
but allows for the desired outcome without us having to carry changes
against upstream OVS.
During our py3 work with OVS this change was identified but for some
reason was not included as part of the already merged changes.
Signed-off-by: Mark Asselstine <mark.asselstine@windriver.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@windriver.com>
Add PACKAGECONFIG for libcap-ng, and default to enable it.
Signed-off-by: Chen Qi <Qi.Chen@windriver.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@windriver.com>
Add dpdk as a runtime dependency when 'dpdk' PACKAGECONFIG is enabled.
Also, introduce a variable DPDK_INSTALL_DIR, and default it to /opt/dpdk.
This is because users might install DPDK into different locations,
e.g. /usr. And it would easy for users to override this variable
to make openvswitch build with dpdk correctly.
Signed-off-by: Chen Qi <Qi.Chen@windriver.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@windriver.com>
While attempting to get ovs to be built and run with py3 (completely
free of py2) host contamination was found (builds on hosts without
python-six installed would fail). It was also determined that pyc
files were still being generated with py2 and not py3. This resulted
in more work being done to achieve the desired results. This work was
sent upstream and subsequently merged. Unfortunately this didn't make
v2.7.1 and may not be available until the next major release, so here
we backport these commits and adjust the recipe to get a clean py3
only build.
Signed-off-by: Mark Asselstine <mark.asselstine@windriver.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@windriver.com>
Released at the end of June 2017 this is a bugfix release with no
major functional changes. Surprisingly the patch for CVE-2017-9263 is
still not applied upstream so we continue to carry this change.
The remaining patches were audited and cleaned up as needed.
Signed-off-by: Mark Asselstine <mark.asselstine@windriver.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@windriver.com>
As far as I can tell openvswitch has been updated to work with python3
as well as with python(2). Switch to depend on python3 and use python3
for associated scripts. For the most part openvswitch will bind at
runtime to either py2 or py3 regardless of these changes, with these
changes we just do a better job of setting up the dependencies to
facilitate py3 bindings. The openvswitch autotests results are mostly
identical before and after this switch (failures move from python3 to
python(2) test cases as expected, with some exceptions see below).
When running the autotests/ptest with python(2) vs python3 we see a
slightly higher failure rate (334 failures vs. 284 failures). I do not
believe this higher fail rate reflects actual errors in the runtime,
rather the tests are not adapted to python3. At any rate like the rest
of openvswitch it is fairly straightforward to hack the logic for
autotests to be run using py2 as long as it is available in the image,
so these changes don't prevent falling back to py2 for autotests. This
should facilitate any debugging we need to do based on us switching to
favor py3.
Signed-off-by: Mark Asselstine <mark.asselstine@windriver.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@windriver.com>
On a minimally provisioned rootfs the tests are currently getting
'stuck' after running for a few minutes because 'find' and 'which' are
not being found. Add these to the ptest package RDEPENDS to ensure
they are present in the rootfs.
Signed-off-by: Mark Asselstine <mark.asselstine@windriver.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@windriver.com>
Continue work to use go infra in oe-core instead of the support for go
previously found in meta-virt. This is a 1:1 drop in replacement and
removes one more go piece from meta-virt in favor of the common
support found in oe-core.
Signed-off-by: Mark Asselstine <mark.asselstine@windriver.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@windriver.com>
There have been a couple of fixes and new features upstream, let's bring
them in. The upstream URL has also changed slightly.
Signed-off-by: Paul Barker <pbarker@toganlabs.com>
Signed-off-by: Bruce Ashfield <bruce.ashfield@windriver.com>
