Skopeo CI tests run under podman; hence the registries
run in the tests will be podman-in-podman. This requires
complex muckery to make work:
- install bats, jq, and podman in the test image
- add new test-system Make target. It runs podman
with /var/lib/containers bind-mounted to a tmpdir
and with other necessary options; and invokes a
test script that hack-edits /etc/containers/storage.conf
before running podman for the first time.
- add --cgroup-manager=cgroupfs option to podman
invocations in BATS: without this, podman-in-podman
fails with:
systemd cgroup flag passed, but systemd support for managing cgroups is not available
Also: gpg --pinentry-mode option is not available on all
our test platforms. Check for it before using.
Signed-off-by: Ed Santiago <santiago@redhat.com>
This script is meant to be used in CI after a `make vendor` run. It's
sole purpose is to execute a `git status --porcelain` and fail with the
list of files reported by it.
Signed-off-by: Valentin Rothberg <rothberg@redhat.com>
Before we use "go get" in CI, run "go version" so that we can be sure of
which version of the toolchain we're using.
Signed-off-by: Nalin Dahyabhai <nalin@redhat.com>
Replace the occurrences of `github.com/projectatomic` with
`github.com/containers` to ensure clean clones of the project are
building, travis badges on the README work as expected and other minor
things.
Signed-off-by: Flavio Castelli <fcastelli@suse.com>
The goal is to include the c/image documentation in a skopeo release,
so that RPMs and other distribution mechanisms can ship the c/image
documentation without having to create a separate package for c/image
(which would not otherwise be needed because it is vendored in users).
So, unify the updates of the "vendor" subdirectory as (make vendor),
and document it in README.md. Also drop hack/vendor.sh, we neither
use nor document it, so updating it as well seems pointless.
Apparently, it was never documented to use (go vet $somefile.go)
(but (go tool vet $somefile.go) was).
go 1.10 seems to do more checks within packages, and $somefile.go
is interpreted as a package with only that file (even if other files
from that package are in the same directory), leading to spurious
"undefined: $symbol" errors.
So, just run (go vet) on ./... (explicitly excluding skopeo/vendor for the
benefit of Go 1.8). We only have three subpackages, so the savings, if any,
from running (go vet) only on the modified subpackages would be small.
More importantly, on a toolchain update, ./... allows us to see the newly
detected issues all at once, instead of randomly waiting for a commit that
changes one of the affected files for the failure to show up.
vndr is almost exactly the same as our old good hack/vendor.sh. Except
it's cleaner and it allows to re-vendor just one dependency if needed
(which we do a lot for containers/image).
Signed-off-by: Antonio Murdaca <runcom@redhat.com>
Vendor containers/storage, and its dependencies github.com/pborman/uuid
and github.com/mistifyio/go-zfs, which we didn't already use.
Update the build Dockerfile to install their dependencies.
Add scriptlets that try to detect whether or not we need to use the
"libdm_no_deferred_remove" and/or "btrfs_noversion" build tags.
Signed-off-by: Nalin Dahyabhai <nalin@redhat.com>
(skopeo copy) will soon ALWAYS require a present policy file. So,
install one by (make install), and ensure that integration tests do so
as well.
Also simplifies the usage of install(1) a bit.
So that people don't need to install all dependencies just to build.
Make it so that "make binary" does nothing if nothing changed.
Remove ${DEST}
Signed-off-by: Doug Davis <dug@us.ibm.com>
Note that this assumes that both (docker login) and (oc login) has
happened, the credentials can be read from the usual config files,
and that the default OpenShift instance should be used.
This includes copy&pasted/modified/simplified code from OpenShift
and Kubernetes, primarily for config file parsing and setting up
TLS and HTTP authentication.
This is much smaller than linking to the upstream OpenShift client
libraries, which via various abstractions and registration drag in much
(dozens of megabytes) more code.
The primary loss from this simplification is automatic conversions
between various versions of the API objects, both for the REST API and
for local configuration storage.
This does not contain downloading/uploading signatures, which depends on
server-side support.
Set GOPATH to start with ./vendor so that we use the dependencies in our
vendored versions instead of dependencies in whatever other version is
elsewhere in GOPATH.
And then undo it when trying to list the non-vendor subpackages in the
current directory.
github.com/coreos/etcd as of v2.2.5 uses a Godeps subdirectory, and
imports packages by including the Godeps path fragments directly in the
package name; so we can't just remove the subdirectory and vendor the
included package directly. So, add a flag to clone() to surpress
removing the vendor subdirectories.
Instead of only checking dependencies of the "main" packages, include
also test dependencies of all subpackages of the project, and their
transitive dependencies.
Otherwise the "clean" step of hack/vendor.sh would drop most .go files
from vendor/ as unused.
Also commits refreshed versions of a few of the vendored packages.
Validating only committed files is not useful in the natural
$test_everything_passes; commit; push
workflow; the failures will not be caught locally, only by Travis later
(and only if PRs are used instead of direct commits to master).
So, use the working directory state instead of last commit for
validations; and remove misleading comments in checks which already use
the working directory state.
