Docker

From CRIU
Jump to: navigation, search

This HOWTO page describes how to checkpoint and restore a Docker container.

Introduction[edit]

Docker wants to manage the full lifecycle of processes running inside one if its containers, which makes it important for CRIU and Docker to work closely together when trying to checkpoint and restore a container. This is being achieved by adding the ability to checkpoint and restore directly into Docker itself, powered under the hood by CRIU. This integration is a work in progress, and its status will be outlined below.

Docker 1.10[edit]

The easiest way to try CRIU and Docker together is to install this pre-compiled version of Docker. It's based on Docker 1.10, and built with the DOCKER_EXPERIMENTAL build tag.

To install, download the docker-1.10.0-dev binary to your system. You'll need to start a docker daemon from this binary, and then you can use the same binary to communicate with that daemon. To start a docker daemon, run a command something like this:

$ docker-1.10.0-dev daemon -D --graph=/var/lib/docker-dev --host unix:///var/run/docker-dev.sock

The graph and host options will prevent colliding with an existing installation of Docker, but you can replace your existing docker if desired. In another shell, you can then connect to that daemon:

$ docker-1.10.0-dev --host unix:///var/run/docker-dev.sock run -d busybox top

Dependencies[edit]

In addition to downloading the binary above (or compiling one yourself), you need CRIU installed on your system, with at least version 2.0. You also need some shared libraries on your system. The most likely things you'll need to install are libprotobuf-c and libnl-3. Here's an output of ldd on my system:

$ ldd `which criu`
   	linux-vdso.so.1 =>  (0x00007ffc09fda000)
   	libpthread.so.0 => /lib/x86_64-linux-gnu/libpthread.so.0 (0x00007fd28b2c7000)
   	libprotobuf-c.so.0 => /usr/lib/x86_64-linux-gnu/libprotobuf-c.so.0 (0x00007fd28b0b7000)
   	libdl.so.2 => /lib/x86_64-linux-gnu/libdl.so.2 (0x00007fd28aeb2000)
   	libnl-3.so.200 => /lib/x86_64-linux-gnu/libnl-3.so.200 (0x00007fd28ac98000)
   	libc.so.6 => /lib/x86_64-linux-gnu/libc.so.6 (0x00007fd28a8d3000)
   	/lib64/ld-linux-x86-64.so.2 (0x000056386bb38000)
   	libm.so.6 => /lib/x86_64-linux-gnu/libm.so.6 (0x00007fd28a5cc000)

checkpoint[edit]

Creating a checkpoint is a top level Docker command with this new version of Docker. Here's an example that simply logs an integer in a loop. From this point forward, commands are show using docker instead of docker-dev-1.10, but if you have not installed this version globally you can use the latter.

First, we create container:

$ docker run -d --name looper --security-opt seccomp:unconfined busybox  \
         /bin/sh -c 'i=0; while true; do echo $i; i=$(expr $i + 1); sleep 1; done'

You can verify the container is running by printings its logs:

$ docker logs looper

If you do this a few times you'll notice the integer increasing. Now, we checkpoint the container:

$ docker checkpoint looper

You should see that the process is no longer running, and if you print the logs a few times no new logs will be printed.

restore[edit]

Like checkpoint, restore is a top level command in this version of Docker. Continuing our example, let's restore the same container:

$ docker restore looper

If we then print the logs, you should see they start from where we left off and continue to increase.

Restoring into a new container[edit]

Beyond the straightforward case of checkpointing and restoring the same container, it's also possible to checkpoint one container, and then restore the checkpoint into a completely different container. Right now that is done with the --force option, in conjunction with the --image-dir option. Here's a slightly revised example from before:

$ docker run -d --name looper2 --security-opt seccomp:unconfined busybox \
         /bin/sh -c 'i=0; while true; do echo $i; i=$(expr $i + 1); sleep 1; done'

# wait a few seconds to give the container an opportunity to print a few lines, then
$ docker checkpoint --image-dir=/tmp/checkpoint1 looper2

$ docker create --name looper-force --security-opt seccomp:unconfined busybox \
         /bin/sh -c 'i=0; while true; do echo $i; i=$(expr $i + 1); sleep 1; done'

$ docker restore --force=true --image-dir=/tmp/checkpoint1 looper-force


You should be able to print the logs from looper-force and see that they start from wherever the logs of looper end.

usage[edit]

# docker checkpoint --help
Usage:	docker checkpoint [OPTIONS] CONTAINER

Checkpoint one or more running containers

    --help             Print usage
    --image-dir        directory for storing checkpoint image files
    --leave-running    leave the container running after checkpoint
    --work-dir         directory for storing log file
# docker restore --help
Usage:	docker restore [OPTIONS] CONTAINER

Restore one or more checkpointed containers

    --force            bypass checks for current container state
    --help             Print usage
    --image-dir        directory to restore image files from
    --work-dir         directory for restore log

Docker 1.12[edit]

More detailed instructions on running checkpoint/restore with Docker in version 1.12 will be coming in the future, but in the meantime, you must build the version of Docker available in the docker-checkpoint-restore branch of Boucher's fork of Docker, available here. Make sure to build with the env DOCKER_EXPERIMENTAL=1.

The command line interface has changed from the 1.10 version. docker checkpoint is now an umbrella command for a few checkpoint operations. To create a checkpoint, use the docker checkpoint create command, which takes container_id and checkpoint_id as non-optional arguments. Example:

   docker checkpoint create my_container my_first_checkpoint

Restoring a container is now performed just as an option to docker start. Although typically you may create and start a container in a single step using docker run, under the hood this is actually two steps: docker create followed by docker start. You can also call start on a container that was previously running and has since been stopped or killed. That looks something like this:

   docker start --checkpoint my_first_checkpoint my_container

Integration Status[edit]

CRIU has already been integrated into the lower level components that power Docker, namely runc and containerd. The final step in the process is to integrate with Docker itself. You can track the status of that process in this pull request.

Compatibility Notes[edit]

The latest versions of the Docker integration require at least version 2.0 of CRIU in order to work correctly. Additionally, depending on the storage driver being used by Docker, and other factors, there may be other compatibility issues that will attempt to be listed here.

TTY[edit]

Checkpointing an interactive container is currently not supported.

Seccomp[edit]

You'll notice that all of the above examples disable Docker's default seccomp support. In order to use seccomp, you'll need a newer version of the Kernel. **Update Needed with Exact Version**

OverlayFS[edit]

There is a bug in OverlayFS that reports the wrong mnt_id in /proc/<pid>/fdinfo/<fd> and the wrong symlink target path for /proc/<pid>/<fd>. Fortunately, these bugs have been fixed in the kernel v4.2-rc2. The following small kernel patches fix the mount id and symlink target path issue:

Assuming that you are running Ubuntu Vivid (Linux kernel 3.19), here is how you can patch your kernel:

git clone  git://kernel.ubuntu.com/ubuntu/ubuntu-vivid.git
cd ubuntu-vivid
git remote add torvalds  git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
git remote update

git cherry-pick 155e35d4da
git cherry-pick df1a085af1
git cherry-pick f25801ee46
git cherry-pick 4bacc9c923
git cherry-pick 9391dd00d1

cp /boot/config-$(uname -r) .config
make olddefconfig
make -j 8 bzImage modules
sudo make install modules_install
sudo reboot

Async IO[edit]

If you are using a kernel older than 3.19 and your container uses AIO, you need the following AIO kernel patches from 3.19:


External Checkpoint Restore[edit]

Note.svg Note: External C/R was done as proof-of-concept. Its use is discouraged and the helper script mentioned below will be deprecated in the near future.

This approach is called external because it's happening external to the Docker daemon. After checkpoint, the Docker daemon thinks that the container has exited. After restore, the Docker daemon doesn't know that the container is running again. Therefore, commands such as docker ps, stop, kill and logs will not work correctly.

Starting with CRIU 1.3, it is possible to checkpoint and restore a process tree running inside a Docker container. However, it's important to note that Docker needs native support for checkpoint and restore in order to maintain its parent-child relationship and to correctly keep track of container states. In other words, while CRIU can C/R a process tree, the restored tree will not become a child of Docker and, from Docker's point of view, the container's state will remain "Exited" (even after successful restore).

It's important to re-emphasize that by checkpointing and restoring a Docker container, we mean C/R of a process tree running inside a container, excluding the Docker daemon itself. As CRIU currently does not support nested PID namespaces, the C/R process tree cannot include the Docker daemon which runs in the global PID namespace.

Command Line Options[edit]

In addition to the usual CRIU command line options used when checkpointing and restoring a process tree, the following command line options are needed for Docker containers.

--root[edit]

This option has been used in the past only for restore operations that wanted to change the root of the mount namespace. It was not used for checkpoint operations.

However, because Docker by default uses the AUFS graph driver and the AUFS module in the kernel reveals branch pathnames in /proc/pid/map_files, option --root is used to specify the root of the mount namespace. Once the kernel AUFS module is fixed, it won't be necessary to specify this option anymore.

--ext-mount-map[edit]

This option is used to specify the path of the external bind mounts. Docker sets up /etc/{hostname,hosts,resolv.conf} as targets with source files outside the container's mount namespace. Older versions of Docker also bind mount /.dockerinit.

For example, assuming the default Docker configuration, /etc/hostname in the container's mount namespace is bind mounted from the source at /var/lib/docker/containers/container_id/hostname.

--manage-cgroups[edit]

When a process tree exits after a checkpoint operation, the cgroups that Docker had created for the container are removed. This option is needed during restore to move the process tree into its cgroups, re-creating them if necessary.

--evasive-devices[edit]

Docker bind mounts /dev/null on /dev/stdin for detached containers (i.e., docker run -d ...). Since earlier versions of Docker used /dev/null in the global namespace, this option tells CRIU to treat the global /dev/null and the container /dev/null as the same device.

--inherit-fd[edit]

For native C/R support, this option tells CRIU to let the restored process "inherit" its specified file descriptor (instead of restoring from checkpoint).

Restore Prework for External C/R[edit]

Docker supports many storage drivers (AKA graph drivers) including AUFS, Btrfs, ZFS, DeviceMapper, OverlayFS, and VFS. The user can specify his/her desired storage driver via the DOCKER_DRIVER environment variable or the -s (--storage-driver) command line option.

Currently C/R can only be done on containers using either AUFS, OverlayFS, or VFS. In the following example, we assume AUFS.

When Docker notices that the container has exited (due to CRIU dump), it dismantles the container's filesystem. We need to set up the container's filesystem again before attempting to restore.

An External C/R Example[edit]

Below is an example to show C/R operations for a shell script that continuously appends a number to a file. You can use tail -f to see the process in action.

As you will see below, after restore, the process's parent is PID 1 (init), not Docker. Also, although the process has been successfully restored, Docker still thinks that the container has exited.

To set up the container's AUFS filesystem before restore, its branch information should be saved before checkpointing the container. For convenience, however, AUFS branch information is saved in the dump.log file. So we can examine dump.log to set up the filesystem again.

For brevity, the 64-character long container ID is replaced by the string <container_id> in the following lines.

$ docker run -d busybox:latest /bin/sh -c 'i=0; while true; do echo $i >> /foo; i=$(expr $i + 1); sleep 3; done'
<container_id>
$ 
$ docker ps
CONTAINER ID  IMAGE           COMMAND           CREATED        STATUS
168aefb8881b  busybox:latest  "/bin/sh -c 'i=0; 6 seconds ago  Up 4 seconds
$ 
$ sudo criu dump -o dump.log -v4 -t 17810 \
	-D /tmp/img/<container_id> \
	--root /var/lib/docker/aufs/mnt/<container_id> \
	--ext-mount-map /etc/resolv.conf:/etc/resolv.conf \
	--ext-mount-map /etc/hosts:/etc/hosts \
	--ext-mount-map /etc/hostname:/etc/hostname \
	--ext-mount-map /.dockerinit:/.dockerinit \
	--manage-cgroups \
	--evasive-devices
$
$ sudo grep successful /tmp/img/<container_id>/dump.log
(00.020103) Dumping finished successfully
$
$ docker ps -a
CONTAINER ID  IMAGE           COMMAND           CREATED        STATUS
168aefb8881b  busybox:latest  "/bin/sh -c 'i=0; 6 minutes ago  Exited (-1) 4 minutes ago
$
$ sudo mount -t aufs -o br=\
/var/lib/docker/aufs/diff/<container_id>:\
/var/lib/docker/aufs/diff/<container_id>-init:\
/var/lib/docker/aufs/diff/a9eb172552348a9a49180694790b33a1097f546456d041b6e82e4d7716ddb721:\
/var/lib/docker/aufs/diff/120e218dd395ec314e7b6249f39d2853911b3d6def6ea164ae05722649f34b16:\
/var/lib/docker/aufs/diff/42eed7f1bf2ac3f1610c5e616d2ab1ee9c7290234240388d6297bc0f32c34229:\
/var/lib/docker/aufs/diff/511136ea3c5a64f264b78b5433614aec563103b4d4702f3ba7d4d2698e22c158:\
none /var/lib/docker/aufs/mnt/<container_id>
$
$ sudo criu restore -o restore.log -v4 -d
	-D /tmp/img/<container_id> \
	--root /var/lib/docker/aufs/mnt/<container_id> \
	--ext-mount-map /etc/resolv.conf:/var/lib/docker/containers/<container_id>/resolv.conf \
	--ext-mount-map /etc/hosts:/var/lib/docker/containers/<container_id>/hosts \
	--ext-mount-map /etc/hostname:/var/lib/docker/containers/<container_id>/hostname \
	--ext-mount-map /.dockerinit:/var/lib/docker/init/dockerinit-1.0.0 \
	--manage-cgroups \
	--evasive-devices
$
$ sudo grep successful /tmp/img/<container_id>/restore.log
(00.424428) Restore finished successfully. Resuming tasks.
$
$ ps -ef | grep /bin/sh
root     18580     1  0 12:38 ?        00:00:00 /bin/sh -c i=0; while true; do echo $i >> /foo; i=$(expr $i + 1); sleep 3; done
$
$ docker ps -a
CONTAINER ID  IMAGE           COMMAND           CREATED        STATUS
168aefb8881b  busybox:latest  "/bin/sh -c 'i=0; 7 minutes ago  Exited (-1) 5 minutes ago
$

External C/R Helper Script[edit]

As seen in the above examples, the CRIU command line for checkpointing and restoring a Docker container is pretty long. For restore, there is also an additional step to set up the root filesystem before invoking CRIU.

To automate the C/R process, there is a helper script in the contrib subdirectory of CRIU sources, called docker_cr.sh. In addition to invoking CRIU, this helper script sets up the root filesystem for AUFS, UnionFS, and VFS for restore.

With docker_cr.sh, all you have to provide is the container ID. If you don't specify a container ID, docker_cr.sh will list all running containers and prompt you to choose one. Also, as shown in the help output below, by setting the appropriate environment variable, it's possible to tell docker_cr.sh which Docker and CRIU binaries to use, where Docker's home directory is, and where CRIU should save and look for its image files.

# docker_cr.sh --help
Usage:
	docker_cr.sh -c|-r [-hv] [<container_id>]
	-c, --checkpoint	checkpoint container
	-h, --help		print help message
	-r, --restore		restore container
	-v, --verbose		enable verbose mode

Environment:
	DOCKER_HOME		(default /var/lib/docker)
	CRIU_IMG_DIR		(default /var/lib/docker/criu_img)
	DOCKER_BINARY		(default docker)
	CRIU_BINARY		(default criu)

Below is an example to checkpoint and restore Docker container 4397:

# docker_cr.sh -c 4397
dump successful
# docker_cr.sh -r 4397
restore successful

Optionally, you can specify -v to see the commands that docker_cr.sh executes. For example:

# docker_cr.sh -c -v 40d3
docker binary: docker
criu binary: criu
image directory: /var/lib/docker/criu_img/40d363f564e00a2f893579fa012a200e475dcf8df47f2a22b7dd0860ffc3d7bf
container root directory: /var/lib/docker/aufs/mnt/40d363f564e00a2f893579fa012a200e475dcf8df47f2a22b7dd0860ffc3d7bf

criu dump -v4 -D /var/lib/docker/criu_img/40d363f564e00a2f893579fa012a200e475dcf8df47f2a22b7dd0860ffc3d7bf -o dump.log \
     --manage-cgroups --evasive-devices \
     --ext-mount-map /etc/resolv.conf:/etc/resolv.conf \
     --ext-mount-map /etc/hosts:/etc/hosts \
     --ext-mount-map /etc/hostname:/etc/hostname \
     --ext-mount-map /.dockerinit:/.dockerinit \
     -t 5991 --root /var/lib/docker/aufs/mnt/40d363f564e00a2f893579fa012a200e475dcf8df47f2a22b7dd0860ffc3d7bf

dump successful
(00.020827) Dumping finished successfully

# docker_cr.sh -r -v 40d3
docker binary: docker
criu binary: criu
image directory: /var/lib/docker/criu_img/40d363f564e00a2f893579fa012a200e475dcf8df47f2a22b7dd0860ffc3d7bf
container root directory: /var/lib/docker/aufs/mnt/40d363f564e00a2f893579fa012a200e475dcf8df47f2a22b7dd0860ffc3d7bf

mount -t aufs -o
/var/lib/docker/aufs/diff/40d363f564e00a2f893579fa012a200e475dcf8df47f2a22b7dd0860ffc3d7bf
/var/lib/docker/aufs/diff/40d363f564e00a2f893579fa012a200e475dcf8df47f2a22b7dd0860ffc3d7bf-init
/var/lib/docker/aufs/diff/a9eb172552348a9a49180694790b33a1097f546456d041b6e82e4d7716ddb721
/var/lib/docker/aufs/diff/120e218dd395ec314e7b6249f39d2853911b3d6def6ea164ae05722649f34b16
/var/lib/docker/aufs/diff/42eed7f1bf2ac3f1610c5e616d2ab1ee9c7290234240388d6297bc0f32c34229
/var/lib/docker/aufs/diff/511136ea3c5a64f264b78b5433614aec563103b4d4702f3ba7d4d2698e22c158
none
/var/lib/docker/aufs/mnt/40d363f564e00a2f893579fa012a200e475dcf8df47f2a22b7dd0860ffc3d7bf

criu restore -v4 -D /var/lib/docker/criu_img/40d363f564e00a2f893579fa012a200e475dcf8df47f2a22b7dd0860ffc3d7bf \
     -o restore.log --manage-cgroups --evasive-devices \
     --ext-mount-map /etc/resolv.conf:/var/lib/docker/containers/40d363f564e00a2f893579fa012a200e475dcf8df47f2a22b7dd0860ffc3d7bf/resolv.conf \
     --ext-mount-map /etc/hosts:/var/lib/docker/containers/40d363f564e00a2f893579fa012a200e475dcf8df47f2a22b7dd0860ffc3d7bf/hosts \
     --ext-mount-map /etc/hostname:/var/lib/docker/containers/40d363f564e00a2f893579fa012a200e475dcf8df47f2a22b7dd0860ffc3d7bf/hostname \
     --ext-mount-map /.dockerinit:/var/lib/docker/init/dockerinit-1.0.0 \
     -d --root /var/lib/docker/aufs/mnt/40d363f564e00a2f893579fa012a200e475dcf8df47f2a22b7dd0860ffc3d7bf \
     --pidfile /var/lib/docker/criu_img/40d363f564e00a2f893579fa012a200e475dcf8df47f2a22b7dd0860ffc3d7bf/restore.pid

restore successful
(00.408807) Restore finished successfully. Resuming tasks.

root      6206     1  1 10:49 ?        00:00:00 /bin/sh -c i=0; while true; do echo $i >> /foo; i=$(expr $i + 1); sleep 3; done