The OCICNI port format has one big problem: It does not support ranges.
So if a users forwards a range of 1k ports with podman run -p 1001-2000
we have to store each of the thousand ports individually as array element.
This bloats the db and makes the JSON encoding and decoding much slower.
In many places we already use a better port struct type which supports
ranges, e.g. `pkg/specgen` or the new network interface.
Because of this we have to do many runtime conversions between the two
port formats. If everything uses the new format we can skip the runtime
conversions.
This commit adds logic to replace all occurrences of the old format
with the new one. The database will automatically migrate the ports
to new format when the container config is read for the first time
after the update.
The `ParsePortMapping` function is `pkg/specgen/generate` has been
reworked to better work with the new format. The new logic is able
to deduplicate the given ports. This is necessary the ensure we
store them efficiently in the DB. The new code should also be more
performant than the old one.
To prove that the code is fast enough I added go benchmarks. Parsing
1 million ports took less than 0.5 seconds on my laptop.
Benchmark normalize PortMappings in specgen:
Please note that the 1 million ports are actually 20x 50k ranges
because we cannot have bigger ranges than 65535 ports.
```
$ go test -bench=. -benchmem ./pkg/specgen/generate/
goos: linux
goarch: amd64
pkg: github.com/containers/podman/v3/pkg/specgen/generate
cpu: Intel(R) Core(TM) i7-10850H CPU @ 2.70GHz
BenchmarkParsePortMappingNoPorts-12 480821532 2.230 ns/op 0 B/op 0 allocs/op
BenchmarkParsePortMapping1-12 38972 30183 ns/op 131584 B/op 9 allocs/op
BenchmarkParsePortMapping100-12 18752 60688 ns/op 141088 B/op 315 allocs/op
BenchmarkParsePortMapping1k-12 3104 331719 ns/op 223840 B/op 3018 allocs/op
BenchmarkParsePortMapping10k-12 376 3122930 ns/op 1223650 B/op 30027 allocs/op
BenchmarkParsePortMapping1m-12 3 390869926 ns/op 124593840 B/op 4000624 allocs/op
BenchmarkParsePortMappingReverse100-12 18940 63414 ns/op 141088 B/op 315 allocs/op
BenchmarkParsePortMappingReverse1k-12 3015 362500 ns/op 223841 B/op 3018 allocs/op
BenchmarkParsePortMappingReverse10k-12 343 3318135 ns/op 1223650 B/op 30027 allocs/op
BenchmarkParsePortMappingReverse1m-12 3 403392469 ns/op 124593840 B/op 4000624 allocs/op
BenchmarkParsePortMappingRange1-12 37635 28756 ns/op 131584 B/op 9 allocs/op
BenchmarkParsePortMappingRange100-12 39604 28935 ns/op 131584 B/op 9 allocs/op
BenchmarkParsePortMappingRange1k-12 38384 29921 ns/op 131584 B/op 9 allocs/op
BenchmarkParsePortMappingRange10k-12 29479 40381 ns/op 131584 B/op 9 allocs/op
BenchmarkParsePortMappingRange1m-12 927 1279369 ns/op 143022 B/op 164 allocs/op
PASS
ok github.com/containers/podman/v3/pkg/specgen/generate 25.492s
```
Benchmark convert old port format to new one:
```
go test -bench=. -benchmem ./libpod/
goos: linux
goarch: amd64
pkg: github.com/containers/podman/v3/libpod
cpu: Intel(R) Core(TM) i7-10850H CPU @ 2.70GHz
Benchmark_ocicniPortsToNetTypesPortsNoPorts-12 663526126 1.663 ns/op 0 B/op 0 allocs/op
Benchmark_ocicniPortsToNetTypesPorts1-12 7858082 141.9 ns/op 72 B/op 2 allocs/op
Benchmark_ocicniPortsToNetTypesPorts10-12 2065347 571.0 ns/op 536 B/op 4 allocs/op
Benchmark_ocicniPortsToNetTypesPorts100-12 138478 8641 ns/op 4216 B/op 4 allocs/op
Benchmark_ocicniPortsToNetTypesPorts1k-12 9414 120964 ns/op 41080 B/op 4 allocs/op
Benchmark_ocicniPortsToNetTypesPorts10k-12 781 1490526 ns/op 401528 B/op 4 allocs/op
Benchmark_ocicniPortsToNetTypesPorts1m-12 4 250579010 ns/op 40001656 B/op 4 allocs/op
PASS
ok github.com/containers/podman/v3/libpod 11.727s
```
Signed-off-by: Paul Holzinger <pholzing@redhat.com>
Trying to restore a container that was started with '--ipc host' fails
with:
Error: error creating container storage: ProcessLabel and Mountlabel must either not be specified or both specified
We already fixed this exact same error message for containers started
with '--privileged'. The previous fix was to check if the to be restored
container is a privileged container (c.config.Privileged). Unfortunately
this does not work for containers started with '--ipc host'.
This commit changes the check for a privileged container to check if
both the ProcessLabel and the MountLabel is actually set and only then
re-uses those labels.
Signed-off-by: Adrian Reber <areber@redhat.com>
Make sure to create the mounts for containers with an overlay root FS in
the runtime dir (e.g., /run/user/1000/...) to guarantee that we can
actually overlay mount on the specific path which is not the case for
the graph root.
[NO NEW TESTS NEEDED] since it is not a user-facing change.
Signed-off-by: Valentin Rothberg <rothberg@redhat.com>
Allow chainging ownership of mountpoint created on top external overlay
rootfs to support use-cases when custom --uidmap and --gidmap are
specified.
Signed-off-by: Aditya Rajan <arajan@redhat.com>
Made changes so that if the pod contains all exited containers and only infra is running, remove the pod.
resolves#11713
Signed-off-by: cdoern <cdoern@redhat.com>
Following commit ensures not dandling mounts are left behind when we are
creating an overlay on top of external rootfs.
Co-authored-by: Valentin Rothberg <rothberg@redhat.com>
Signed-off-by: Aditya Rajan <arajan@redhat.com>
Access the container's spec field directly inside of libpod instead of
calling Spec() which in turn creates expensive JSON deep copies.
Accessing the field directly drops memory consumption of a simple
podman run --rm busybox true from ~700kB to ~600kB.
[NO TESTS NEEDED]
Signed-off-by: Valentin Rothberg <rothberg@redhat.com>
Access the container's config field directly inside of libpod instead of
calling `Config()` which in turn creates expensive JSON deep copies.
Accessing the field directly drops memory consumption of a simple
`podman run --rm busybox true` from 1245kB to 410kB.
[NO TESTS NEEDED]
Signed-off-by: Valentin Rothberg <rothberg@redhat.com>
Make use of the new network interface in libpod.
This commit contains several breaking changes:
- podman network create only outputs the new network name and not file
path.
- podman network ls shows the network driver instead of the cni version
and plugins.
- podman network inspect outputs the new network struct and not the cni
conflist.
- The bindings and libpod api endpoints have been changed to use the new
network structure.
The container network status is stored in a new field in the state. The
status should be received with the new `c.getNetworkStatus`. This will
migrate the old status to the new format. Therefore old containers should
contine to work correctly in all cases even when network connect/
disconnect is used.
New features:
- podman network reload keeps the ip and mac for more than one network.
- podman container restore keeps the ip and mac for more than one
network.
- The network create compat endpoint can now use more than one ipam
config.
The man pages and the swagger doc are updated to reflect the latest
changes.
Signed-off-by: Paul Holzinger <pholzing@redhat.com>
Allows users to specify a readonly rootfs with :O, in exchange podman will create a writable overlay.
bump builah to v1.22.1-0.20210823173221-da2b428c56ce
[NO TESTS NEEDED]
Signed-off-by: flouthoc <flouthoc.git@gmail.com>
When a container is automatically restarted due its restart policy and
the container uses rootless cni networking with ports forwarded we have
to start a new rootlessport process since it exits with conmon.
Signed-off-by: Paul Holzinger <pholzing@redhat.com>
When inspecting a container, we now report whether the container
was stopped by a `podman checkpoint` operation via a new bool in
the State portion of inspected, `Checkpointed`.
Signed-off-by: Matthew Heon <mheon@redhat.com>
There are use-cases where users would want to use overlay-mounts as
workdir. For such cases workdir should be resolved after all the mounts
are completed during the container init process.
Signed-off-by: Aditya Rajan <arajan@redhat.com>
InfraContainer should go through the same creation process as regular containers. This change was from the cmd level
down, involving new container CLI opts and specgen creating functions. What now happens is that both container and pod
cli options are populated in cmd and used to create a podSpecgen and a containerSpecgen. The process then goes as follows
FillOutSpecGen (infra) -> MapSpec (podOpts -> infraOpts) -> PodCreate -> MakePod -> createPodOptions -> NewPod -> CompleteSpec (infra) -> MakeContainer -> NewContainer -> newContainer -> AddInfra (to pod state)
Signed-off-by: cdoern <cdoern@redhat.com>
Add the --userns flag to podman pod create and keep
track of the userns setting that pod was created with
so that all containers created within the pod will inherit
that userns setting.
Specifically we need to be able to launch a pod with
--userns=keep-id
Signed-off-by: Daniel J Walsh <dwalsh@redhat.com>
Signed-off-by: Urvashi Mohnani <umohnani@redhat.com>
Currently we override the SELinux labels specified by the user
if the container is runing a kata container or systemd container.
This PR fixes to use the label specified by the user.
Fixes: https://github.com/containers/podman/issues/11100
Signed-off-by: Daniel J Walsh <dwalsh@redhat.com>
The global flag will work in either location, and this flag just breaks
users expectations, and is basically a noop.
Also fix global storage-opt so that podman-remote can use it.
[NO TESTS NEEDED] Since it would be difficult to test in ci/cd.
Fixes: https://github.com/containers/podman/issues/10264
Signed-off-by: Daniel J Walsh <dwalsh@redhat.com>
Create the /etc and /etc/mtab directories with the
correct ownership based on what the UID and GID is
for the container. This was causing issue when starting
the infra container with userns as the /etc directory
wasn't being created with the correct ownership.
Signed-off-by: Urvashi Mohnani <umohnani@redhat.com>
Unfortunately --pre-checkpointing never worked as intended and recent
changes to runc have shown that it is broken.
To create a pre-checkpoint CRIU expects the paths between the
pre-checkpoints to be a relative path. If having a previous checkpoint
it needs the be referenced like this: --prev-images-dir ../parent
Unfortunately Podman was giving runc (and CRIU) an absolute path.
Unfortunately, again, until March 2021 CRIU silently ignored if
the path was not relative and switch back to normal checkpointing.
This has been now fixed in CRIU and runc and running pre-checkpoint
with the latest runc fails, because runc already sees that the path is
absolute and returns an error.
This commit fixes this by giving runc a relative path.
This commit also fixes a second pre-checkpointing error which was just
recently introduced.
So summarizing: pre-checkpointing never worked correctly because CRIU
ignored wrong parameters and recent changes broke it even more.
Now both errors should be fixed.
[NO TESTS NEEDED]
Signed-off-by: Adrian Reber <areber@redhat.com>
Signed-off-by: Adrian Reber <adrian@lisas.de>
Checkpointed containers started with --privileged fail during restore
with:
Error: error creating container storage: ProcessLabel and Mountlabel must either not be specified or both specified
This commit fixes it by not setting the labels when restoring a
privileged container.
[NO TESTS NEEDED]
Signed-off-by: Adrian Reber <areber@redhat.com>
Support UID, GID, Mode options for mount type secrets. Also, change
default secret permissions to 444 so all users can read secret.
Signed-off-by: Ashley Cui <acui@redhat.com>
We should create the /etc/mtab->/proc/mountinfo link
so that mount command will work within the container.
Docker does this by default.
Fixes: https://github.com/containers/podman/issues/10263
Signed-off-by: Daniel J Walsh <dwalsh@redhat.com>
When a container is automatically restarted due its restart policy and
the container used the slirp4netns netmode, the slirp4netns process
died. This caused the container to lose network connectivity.
To fix this we have to start a new slirp4netns process.
Fixes#8047
Signed-off-by: Paul Holzinger <paul.holzinger@web.de>
volatile containers are a storage optimization that disables *sync()
syscalls for the container rootfs.
If a container is created with --rm, then automatically set the
volatile storage flag as anyway the container won't persist after a
reboot or machine crash.
[NO TESTS NEEDED]
Signed-off-by: Giuseppe Scrivano <gscrivan@redhat.com>
Migrate the Podman code base over to `common/libimage` which replaces
`libpod/image` and a lot of glue code entirely.
Note that I tried to leave bread crumbs for changed tests.
Miscellaneous changes:
* Some errors yield different messages which required to alter some
tests.
* I fixed some pre-existing issues in the code. Others were marked as
`//TODO`s to prevent the PR from exploding.
* The `NamesHistory` of an image is returned as is from the storage.
Previously, we did some filtering which I think is undesirable.
Instead we should return the data as stored in the storage.
* Touched handlers use the ABI interfaces where possible.
* Local image resolution: previously Podman would match "foo" on
"myfoo". This behaviour has been changed and Podman will now
only match on repository boundaries such that "foo" would match
"my/foo" but not "myfoo". I consider the old behaviour to be a
bug, at the very least an exotic corner case.
* Futhermore, "foo:none" does *not* resolve to a local image "foo"
without tag anymore. It's a hill I am (almost) willing to die on.
* `image prune` prints the IDs of pruned images. Previously, in some
cases, the names were printed instead. The API clearly states ID,
so we should stick to it.
* Compat endpoint image removal with _force_ deletes the entire not
only the specified tag.
Signed-off-by: Valentin Rothberg <rothberg@redhat.com>
Some packages used by the remote client imported the libpod package.
This is not wanted because it adds unnecessary bloat to the client and
also causes problems with platform specific code(linux only), see #9710.
The solution is to move the used functions/variables into extra packages
which do not import libpod.
This change shrinks the remote client size more than 6MB compared to the
current master.
[NO TESTS NEEDED]
I have no idea how to test this properly but with #9710 the cross
compile should fail.
Signed-off-by: Paul Holzinger <paul.holzinger@web.de>
The --trace has helped in early stages analyze Podman code. However,
it's contributing to dependency and binary bloat. The standard go
tooling can also help in profiling, so let's turn `--trace` into a NOP.
[NO TESTS NEEDED]
Signed-off-by: Valentin Rothberg <rothberg@redhat.com>
Unlocking an already unlocked lock is a panic. As such, we have
to make sure that the deferred c.lock.Unlock() in
c.StopWithTimeout() always runs on a locked container. There was
a case in c.stop() where we could return an error after we unlock
the container to stop it, but before we re-lock it - thus
allowing for a double-unlock to occur. Fix the error return to
not happen until after the lock has been re-acquired.
Fixes#9615
Signed-off-by: Matthew Heon <mheon@redhat.com>
Traditionally, the path resolution for containers has been resolved on
the *host*; relative to the container's mount point or relative to
specified bind mounts or volumes.
While this works nicely for non-running containers, it poses a problem
for running ones. In that case, certain kinds of mounts (e.g., tmpfs)
will not resolve correctly. A tmpfs is held in memory and hence cannot
be resolved relatively to the container's mount point. A copy operation
will succeed but the data will not show up inside the container.
To support these kinds of mounts, we need to join the *running*
container's mount namespace (and PID namespace) when copying.
Note that this change implies moving the copy and stat logic into
`libpod` since we need to keep the container locked to avoid race
conditions. The immediate benefit is that all logic is now inside
`libpod`; the code isn't scattered anymore.
Further note that Docker does not support copying to tmpfs mounts.
Tests have been extended to cover *both* path resolutions for running
and created containers. New tests have been added to exercise the
tmpfs-mount case.
For the record: Some tests could be improved by using `start -a` instead
of a start-exec sequence. Unfortunately, `start -a` is flaky in the CI
which forced me to use the more expensive start-exec option.
Signed-off-by: Valentin Rothberg <rothberg@redhat.com>
We missed bumping the go module, so let's do it now :)
* Automated go code with github.com/sirkon/go-imports-rename
* Manually via `vgrep podman/v2` the rest
Signed-off-by: Valentin Rothberg <rothberg@redhat.com>
Instead of using the container's mountpoint as the base of the
chroot and indexing from there by the volume directory, instead
use the full path of what we want to copy as the base of the
chroot and copy everything in it. This resolves the bug, ends up
being a bit simpler code-wise (no string concatenation, as we
already have the full path calculated for other checks), and
seems more understandable than trying to resolve things on the
destination side of the copy-up.
Fixes#9354
Signed-off-by: Matthew Heon <matthew.heon@pm.me>
This one is rather bizarre because it triggers only on some
systems. I've included a CI test, for example, but I'm 99% sure
we use images in CI that have volumes over empty directories, and
the earlier patch to change copy-up implementation passed CI
without complaint.
I can reproduce this on a stock F33 VM, but that's the only place
I have been able to see it.
Regardless, the issue: under certain as-yet-unidentified
environmental conditions, the copier.Get method will return an
ENOENT attempting to stream a directory that is empty. Work
around this by avoiding the copy altogether in this case.
Signed-off-by: Matthew Heon <mheon@redhat.com>
The old copy-up implementation was very unhappy with symlinks,
which could cause containers to fail to start for unclear reasons
when a directory we wanted to copy-up contained one. Rewrite to
use the Buildah Copier, which is more recent and should be both
safer and less likely to blow up over links.
At the same time, fix a deadlock in copy-up for volumes requiring
mounting - the Mountpoint() function tried to take the
already-acquired volume lock.
Fixes#6003
Signed-off-by: Matthew Heon <mheon@redhat.com>
Implement podman secret create, inspect, ls, rm
Implement podman run/create --secret
Secrets are blobs of data that are sensitive.
Currently, the only secret driver supported is filedriver, which means creating a secret stores it in base64 unencrypted in a file.
After creating a secret, a user can use the --secret flag to expose the secret inside the container at /run/secrets/[secretname]
This secret will not be commited to an image on a podman commit
Signed-off-by: Ashley Cui <acui@redhat.com>
Paul Holzinger
OpenShift Merge Robot
Adrian Reber
Valentin Rothberg
Aditya Rajan
Daniel J Walsh
cdoern
flouthoc
Matthew Heon
Urvashi Mohnani
Ashley Cui
Paul Holzinger
Giuseppe Scrivano
Peter Hunt
Matthew Heon