chrootarchive.Tar returns an io.ReadCloser backed by the tar-producing
process. The checkpoint export path copies from that stream into the
target file, but never closes the reader.
A successful copy reads the stream to EOF, so the producer normally exits.
On early errors, such as failing to create or chmod the target file, or a
copy error, the producer can be left without a consumer.
Close the tar stream after creating it so error paths release the pipe and
allow the producer to exit.
Signed-off-by: Mikhail Dmitrichenko <m.dmitrichenko00@bk.ru>
When checkpointing a container with --leave-running, libpod dumps the
container's memory via the OCI runtime (CRIU) first and only captures
the rootfs diff and named volumes afterwards. CRIU thaws the container
as soon as the memory dump finishes, so the processes inside the
container continue to run between the memory snapshot and the
file-system capture. As a result, the checkpoint can be inconsistent:
have CRIU images and a file system that reflect different points in time.
To fix this, we freeze the container's cgroup before invoking the OCI
runtime and thaw it again only after the checkpoint image/archive has
been written. The OCI runtime calls CRIU with the freezer cgroup and
restores it to its previous state once the dump completes, so a
container that was already frozen stays frozen across the dump and
the file system is captured at the same instant as the CRIU images.
This mirrors the approach other engines (e.g. CRI-O and containerd).
The default (stopping) checkpoint functionality is not affected by this
issue because CRIU leaves the tasks dead after the dump.
This patch also adds a regression test for the consistency of live
(--leave-running) checkpoints. The container runs a workload that
keeps an in-memory counter in sync with a value written to a file
on its root file system, maintaining the invariant that the on-disk
value never gets ahead of the in-memory counter.
Signed-off-by: Radostin Stoyanov <rstoyanov@fedoraproject.org>
Using os.Is{Exist,NotExist,Permission} checks is not recommended in the
new code (see official documentation). While using it in the existing
code is OK, it may still result in a subtle errors later (for a specific
example of that, see [1]).
Replace those with errors.Is.
Generated by:
gofmt -r 'os.IsExist(a) -> errors.Is(a, os.ErrExist)' -w .
gofmt -r 'os.IsNotExist(a) -> errors.Is(a, os.ErrNotExist)' -w .
gofmt -r 'os.IsPermission(a) -> errors.Is(a, os.ErrPermission)' -w .
goimports -w .
git diff vendor test/tools/vendor | patch -p1 -R
[1]: https://github.com/opencontainers/runc/pull/5061
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
Add missing defer Close() calls in four locations:
- libpod/container.go: specFromState() opens the container config file
but never closes it after reading, leaking one fd per call.
- libpod/container_internal_common.go: checkpoint() and restore() each
open the bundle directory for CRIU statistics but never close it,
leaking one fd per checkpoint/restore operation.
- pkg/machine/shim/host.go: Init() opens the playbook file but never
closes it after ReadAll, leaking one fd per machine init.
Signed-off-by: Sebastien Tardif <sebtardif@ncf.ca>
When a container in a pod is stopped, its container name is removed from /etc/hosts. etchosts.Remove() filters for any entry matching the container name or the pod hostname. A pod with additional host entries like --add-host FQDN;pod-hostname:127.0.0.1 is affected by this deletion, too.
Only the container name needs to be removed when a container is stopped.
Signed-off-by: Clemens Klug <git@agp8x.org>
The podman module paths are moving from github.com/containers/podman to
go.podman.io/podman. This will help with future mobility.
Signed-off-by: Brent Baude <bbaude@redhat.com>
This PR reflects the upstream change of moving the buildah module from
github.com/containers/buildah to go.podman.io/buildah.
Signed-off-by: Brent Baude <bbaude@redhat.com>
This was implemented by containers/netavark #1369; this commit
completes the process by wiring it into Podman. We now respect
the CLI order for configured networks - if a user passes
`--net net1,net2` we guarantee that net1 will be configured
before net2.
For containers created before this patch, we don't retain enough
information to configure networks in CLI order, so we use
alphabetical order instead to still guarantee consistency.
No breaking API changes have been made, but we do add a new
field to supplement the existing map to (optionally) provide
ordering information. The Podman CLI will always pass this.
Existing applications that do not will, again, receive]
deterministic ordering based on an alphabetical sort of network
names.
This requires the latest version of Netavark to work properly.
Signed-off-by: Matthew Heon <matthew.heon@pm.me>
Remove all slirp4netns-specific setup and helper functions while
preserving the RootlessRLK port mapping functions that are still
used by pasta and bridge networking.
Rename networking_slirp4netns.go to networking_rootlessport.go.
Convert the slirp4netns compose test to use pasta instead.
Remove rootlessSlirpSyncR/W fields from container struct and
getSlirp4netnsIP stub from FreeBSD networking.
Signed-off-by: Lokesh Mandvekar <lsm5@redhat.com>
Just as additional hardening.
Note chrootarchive does not work on macos/windows, in that case it still
falls back to the regular pkg/archive.
Signed-off-by: Paul Holzinger <pholzing@redhat.com>
- Don't allocate it on-demand, it will always be created in the end.
- Embed the SystemContext directly, without using a pointer,
to make it clear it always exists.
Should not change behavior.
Signed-off-by: Miloslav Trmač <mitr@redhat.com>
Remove CNI-specific conditional logic and update comments throughout
the libpod networking code:
- Simplified DNS configuration logic in container_internal_common.go
to always use netavark behavior (removed backend checks)
- Removed CNI-specific iptables chain error regex pattern
- Updated all comments referencing 'CNI' to use 'netavark' or
'network backend'
- Renamed variable 'cniNet' to 'netInfo' for clarity
- Updated field and type documentation to remove CNI references
All networking code now assumes netavark as the sole backend.
Signed-off-by: Lokesh Mandvekar <lsm5@redhat.com>
The code checks for isPathOnVolume and isPathOnMount so we can just use
the SecureJoin here directly to check for path existance.
Then instead of walking symlinks and trying to guess if they are on a
mount just assume if it is a link (path is different from the normal
joined one) then don't error out early and let the OCI runtime deal with
it. The runtime does produce a less readable error but it still fails
and we have much less fragile code.
Signed-off-by: Paul Holzinger <pholzing@redhat.com>
MkdirAll can fail with EEXIST when the path is a symlink and the target
doesn't exist. As such we should ignore the error.
Note there is something fundemantal wrong here with the path access as
it is following the symlink to the host, however it is only for a
stat() so it is not an security issue here.
Fixes: 637c264e2e ("fix issues found by nilness")
Signed-off-by: Paul Holzinger <pholzing@redhat.com>
Tremendous amount of changes in here, but all should amount to
the same thing: changing Go import paths from v5 to v6.
Also bumped go.mod to github.com/containers/podman/v6 and updated
version to v6.0.0-dev.
Signed-off-by: Matt Heon <mheon@redhat.com>
Problem: While removing cgroupsv1 code, I noticed my neovim Go config
automatically changed fileperms to the new octal format and I didn't
want that polluting my diffs.
Decision: I thought it best to switch to the new octal format in a dedicated PR.
Action:
- Cursor switched to new octal format for all fileperm ocurrences in Go
source and test files.
- vendor/, docs/ and non-Go files were ignored.
- Reviewed manually.
Ref: https://go.dev/ref/spec#Go_1.13
Signed-off-by: Lokesh Mandvekar <lsm5@redhat.com>
pkg/libartifact has been moved to common and as such needs to be removed
from podman and the new common vendored in along with required deps.
https://issues.redhat.com/browse/RUN-3618
Signed-off-by: Brent Baude <bbaude@redhat.com>
Using golang.org/x/tools/gopls/internal/analysis/modernize/cmd/modernize
+ some manual cleanup in libpod/lock/shm/shm_lock_test.go as it
generated an unused variable
+ restored one removed comment
Signed-off-by: Paul Holzinger <pholzing@redhat.com>
Starting with runc 1.3.0 it errors when we pass unknown mount options to
the runtime, the copy/nocopy options are specific to podman when we
mount the volume and are not valid mount options for the runtime.
Fixes: #26938
Signed-off-by: Paul Holzinger <pholzing@redhat.com>
An artifact without the title annoation just gets the digest as name
which is less than ideal. While it is a decent default to avoid
conflicts users would like to configure the name.
With the name=abc option we will call the file abc in case of a signle
artifact and otherwise we use abc-x where x is the layer index starting
at 0 to avoid conflicts.
Signed-off-by: Paul Holzinger <pholzing@redhat.com>
If the artifact has a single blob then use the dst path directly as
mount in case it does not exist.
Signed-off-by: Paul Holzinger <pholzing@redhat.com>
We like to append the host servers in that case so that we do not only
force dns.podman.
Fixes: #24713
Fixes: https://issues.redhat.com/browse/RHEL-83787
Signed-off-by: Paul Holzinger <pholzing@redhat.com>
After the system reboot, the Rootfs for infra-container can
be removed. This can happen when it is stored on tmpfs.
This commit recreates the infra-container directory which is
used for Rootfs for infra-container before mounting it.
Fixes: #26190
Signed-off-by: Jan Kaluza <jkaluza@redhat.com>
The conditions are always true so they can be removed. And in the case
of exportCheckpoint() the scope means addToTarFiles was overwritten and
thus when it looped over it later the slice was always empty.
Signed-off-by: Paul Holzinger <pholzing@redhat.com>
This commit removes the code to build a local pause
image from the Containerfile. It is replaced with
code to find the catatonit binary and include it in
the Rootfs.
This removes the need to build a local pause container
image.
The same logic is also applied to createServiceContainer
which is originally also based on the pause image.
Fixes: #23292
Signed-off-by: Jan Kaluza <jkaluza@redhat.com>
create the /etc/passwd and /etc/group files before any user/group
lookup so that the entries added dynamically are found by --user.
As a side effect, do not automatically create the group with same
value as the uid when not specified, since it is expected to run with
gid=0.
Closes: https://github.com/containers/podman/issues/25805
Signed-off-by: Giuseppe Scrivano <gscrivan@redhat.com>
These:
> libpod/container_copy_common.go:34:16: QF1011: could omit type bool from declaration; it will be inferred from the right-hand side (staticcheck)
> locked bool = true
> ^
> libpod/container_internal_common.go:793:3: QF1006: could lift into loop condition (staticcheck)
> if maxSymLinks > 40 {
> ^
> libpod/oci_conmon_linux.go:170:2: QF1007: could merge conditional assignment into variable declaration (staticcheck)
> mustCreateCgroup := true
> ^
Should not result in any change of logic.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
This commit adds new --cdi-spec-dir global option. This
option is used to add additional CDI spec paths.
Signed-off-by: Micah Chambers (eos) <mchambers@anduril.com>
Signed-off-by: Jan Kaluza <jkaluza@redhat.com>
Have one function without a `defer lock.unlock()` as one of the
commands in it calls a function that also takes the same lock,
so the unlock has to happen prior to function completion.
Unfortunately, this is prone to errors, like the one here: I
missed a case, and we could return without unlocking, causing a
deadlock later in the cleanup code as we tried to take the same
lock again.
Refactor the command to use `defer unlock()` to simplify and
avoid any further errors of this type.
Introduced by e66b788a51 - this
should be included in any backports of that commit.
Fixes#25585
Signed-off-by: Matt Heon <mheon@redhat.com>
Add a new option to allow for mounting artifacts in the container, the
syntax is added to the existing --mount option:
type=artifact,src=$artifactName,dest=/path[,digest=x][,title=x]
This works very similar to image mounts. The name is passed down into
the container config and then on each start we lookup the artifact and
the figure out which blobs to mount. There is no protaction against a
user removing the artifact while still being used in a container. When
the container is running the bind mounted files will stay there (as the
kernel keeps the mounts active even if the bind source was deleted).
On the next start it will fail to start as if it does not find the
artifact. The good thing is that this technically allows someone to
update the artifact with the new file by creating a new artifact with
the same name.
Signed-off-by: Paul Holzinger <pholzing@redhat.com>
I'm not sure if there is an equivalent to CAP_SYS_RESOURCE on FreeBSD
but for now, I have added a no-op stub which returns false.
Signed-off-by: Doug Rabson <dfr@rabson.org>
Fixes: https://github.com/containers/podman/issues/25002
Also add the ability to inspect containers for
UseImageHosts and UseImageHostname.
Finally fixed some bugs in handling of --no-hosts for Pods,
which I descovered.
Signed-off-by: Daniel J Walsh <dwalsh@redhat.com>
