...to reduce flakes.
Reason: journald makes no guarantees. Just because a systemd job
has finished, or podman has written+flushed log entries, doesn't
mean that journald will actually know about them:
https://github.com/systemd/systemd/issues/28650
Workaround: wrap some podman-logs tests inside Eventually()
so they will be retried when log == journald
This addresses, but does not close, #18501. That's a firehose,
with many more failures than I can possibly cross-reference.
I will leave it open, then keep monitoring missing-logs flakes
over time, and pick those off as they occur.
Signed-off-by: Ed Santiago <santiago@redhat.com>
There is a problem where our tail code does not handles correctly
partial log lines. This makes podman logs --tail output possibly
incorrect lines when k8s-file is used.
This manifests as flake in CI because partial lines are only sometimes
written, basically always when the output is flushed before writing a
newline.
For our code we must not count partial lines which was already done but
the important thing we must keep reading backwards until the next full
(F) line. This is because all partial (P) lines still must be added to
the full line. See the added tests for details on how the log file looks
like.
While fixing this, I rework the tail logic a bit, there is absolutely no
reason to read the lines in a separate goroutine just to pass the lines
back via channel. We can do this in the same routine.
The logic is very simple, read the lines backwards, append lines to
result and then at the end invert the result slice as tail must return
the lines in the correct order. This more efficient then having to
allocate two different slices or to prepend the line as this would
require a new allocation for each line.
Lastly the readFromLogFile() function wrote the lines back to the log
line channel in the same routine as the log lines we read, this was bad
and causes a deadlock when the returned lines are bigger than the
channel size. There is no reason to allocate a big channel size we can
just write the log lines in a different goroutine, in this case the main
routine were read the logs anyway.
A new system test and unit tests have been added to check corner cases.
Fixes#19545
Signed-off-by: Paul Holzinger <pholzing@redhat.com>
An unhelpful comment doesn't give any clues why this test was originally
skipped on Ubuntu. In any case, now that CI uses Debian SID, re-enable
the test hoping that it now functions.
Signed-off-by: Chris Evich <cevich@redhat.com>
We do not allow volumes and mounts to be placed at the same
location in the container, with create-time checks to ensure this
does not happen. User-added conflicts cannot be resolved (if the
user adds two separate mounts to, say, /myapp, we can't resolve
that contradiction and error), but for many other volume sources,
we can solve the contradiction ourselves via a priority
hierarchy. Image volumes come first, and are overridden by the
`--volumes-from` flag, which are overridden by user-added mounts,
etc, etc. The problem here is that we were not properly handling
volumes-from overriding image volumes. An inherited volume from
--volumes-from would supercede an image volume, but an inherited
mount would not. Solution is fortunately simple - just clear out
the map entry for the other type when adding volumes-from
volumes.
Makes me wish for Rust sum types - conflict resolution would be a
lot simpler if we could use a sum type for volumes and bind
mounts and thus have a single map instead of two maps, one for
each type.
Fixes#19529
Signed-off-by: Matthew Heon <matthew.heon@pm.me>
The `Exists` field of the `RemoteSocket` struct is marshaled to json with the
`omitempty` setting. This has the disadvantage that by default `podman info`
shows a `remotePath` entry (the remote path is set in
`pkg/domain/infra/abi/systems.go`: `(*ContainerEngine).Info`) but not that this
path does not exist:
```
❯ podman info --format json | jq .host.remoteSocket
{
"path": "/run/user/1000/podman/podman.sock"
}
```
By removing the `omitempty`, we ensure that the existence is always shown:
```
❯ bin/podman info --format json | jq .host.remoteSocket
{
"path": "/run/user/1000/podman/podman.sock",
"exists": false
}
```
Signed-off-by: Dan Čermák <dcermak@suse.com>
Compat api for containers/stop should take -1 value
Add support for `podman stop --time -1`
Add support for `podman restart --time -1`
Add support for `podman rm --time -1`
Add support for `podman pod stop --time -1`
Add support for `podman pod rm --time -1`
Add support for `podman volume rm --time -1`
Add support for `podman network rm --time -1`
Fixes: https://github.com/containers/podman/issues/17542
Signed-off-by: Daniel J Walsh <dwalsh@redhat.com>
Remove "HTTP if one registry" test. It is a NOP, has been skipped
for two months, and nobody knows what its original purpose was.
Closes: #18768
Signed-off-by: Ed Santiago <santiago@redhat.com>
First, all the defaults for TERM=xterm were removed from c/common, then accordingly the same will be added if encountered a set tty flag.
Signed-off-by: Chetan Giradkar <cgiradka@redhat.com>
Adds support for --add-compression which accepts multiple compression
formats and when used it will add all instances in a manifest list with
requested compression formats.
Signed-off-by: Aditya R <arajan@redhat.com>
Some people might expect this to work:
systemctl --wait start foo
journalctl -u foo ---> displays output from foo
Well, it does not. Not reliably, anyway:
https://github.com/systemd/systemd/issues/28650
Shrug, okay, deal with it: write value of %T to a tmpfile
instead of relying on journal. I tested with TMPDIR=<many values>
on an SELinux system and, by golly, it works fine.
Signed-off-by: Ed Santiago <santiago@redhat.com>
Forcing users to set --rm when setting --rmi is just bad UI.
If I want the image to be removed, it implies that I want the
container removed that I am creating.
Fixes: https://github.com/containers/podman/issues/15640
Signed-off-by: Daniel J Walsh <dwalsh@redhat.com>
The original SELinux support in Docker and Podman does not follow the
default SELinux rules for how label transitions are supposed to be
handled. Containers always switch their user and role to
system_u:system_r, rather then maintain the collers user and role.
For example
unconfined_u:unconfined_r:container_t:s0:c1,c2
Advanced SELinux administrators want to confine users but still allow
them to create containers from their role, but not allow them to launch
a privileged container like spc_t.
This means if a user running as
container_user_u:container_user_r:container_user_t:s0
Ran a container they would get
container_user_u:container_user_r:container_t:s0:c1,c2
If they run a privileged container they would run it with:
container_user_u:container_user_r:container_user_t:s0
If they want to force the label they would get an error
podman run --security-opt label=type:spc_t ...
Should fail. Because the container_user_r can not run with the spc_t.
SELinux rules would also prevent the user from forcing system_u user and
the sytem_r role.
Signed-off-by: Daniel J Walsh <dwalsh@redhat.com>
Signed-off-by: Chris Evich <cevich@redhat.com>
We do not use any special netns path for the netns=none case, however
callers that inspect that may still wish to join the netns path directly
without extra work to figure out /proc/$pid/ns/net.
Fixes#16716
Signed-off-by: Paul Holzinger <pholzing@redhat.com>
And lo, a miracle occurred. Containerized checkpoint tests are
no longer hanging. Reenable them.
(Followup miracle: tests are still passing, after a year of not
running!)
Closes: #15015
Signed-off-by: Ed Santiago <santiago@redhat.com>
