What cannot be checkpointed

CRIU cannot dump all the possible states applications can live in. This article summarizes our experience in that area.

External resources

By default CRIU allows to dump the set of processes and their resources if this set has no connections outside. However, in some situations it makes sense to ignore this external connection on dump and recreate one on restore. So, what this "connection outside" is? For example:

• UNIX socket: Application may have a UNIX socket connected to some other app and the latter one is not dumped. This is called external UNIX socket.
• TTY, group and session: If you start a program from shell, the tty, process group and session of the new program can be shared with the shell itself. People often meet this when they try CRIU for the first tome on a simple loop.
• TCP connection: This socket is literally an external connection, so CRIU should be explicitly allowed to dump one.

File locks

A file lock is an object, that belongs to some filesystem. On dump it's impossible to find out whether this lock help by one task can be used by some other. Thus, CRIU doesn't dump tasks with held locks. The --file-locks CLI option tells CRIU to dump the lock.

Devices

If a task has opened or mapped any character or block device, this typically means, it wants some connection to the hardware. In this case dump (and restore) is impossible. The exception is virtual devices like null, zero, etc. and TUN network device (used by OpenVPN).

The explanation why device file cannot be dumped and restored in a generic way is two fold. First of all, application can be dumped for live migration and on restore there can be no such device. But that's easy. Other than this, we don't know how all the devices work. App might have loaded some state into it, and in order to dump it properly we need to fetch that state. This is not something that can be done in a generic manner.

Tasks with debugger attached

CRIU uses the same API as debuggers do to get some tasks' state and this API (the ptrace one) doesn't allow for multiple debuggers to explore a task. Thus tasks under gdb or strace cannot be dumped.

Task from different user (for non root)

For security reasons, if CRIU is requested by non-root to dump some other task, it doesn't do it unless the dumpee belongs to the same user.

UNIX sockets with relative path

For bound unix sockets kernel only provides the path with this the bind() syscall was called. If the path is relative we have no reliable way to find the exact socket location out and just refuse to dump the task.

TimerFD

Sockets other than TCP, UDP, UNIX, packet and netlink

Packetized pipes (created with O_DIRECT)

Cork-ed UDP sockets