This directory contains various tools for managing the lifetime of VM instances and for providing any services those VMs may need while they are running.
vm_launcher is responsible for starting up a new VM instance. It first allocates some pooled resources for the VM (like a mac address, IPv4 subnet, unique identifier, and virtual socket host id) and then invokes the userspace hypervisor (crosvm or lkvm) to actually start running the VM.
maitred is the agent running inside the VM responsible for managing the VM instance. It acts as the init system, starting up system services, mounting file systems, and launching the container with the actual application that the user wants to run. It is responsible for shutting down the VM once the user's application exits or if requested by a process on the host.
See docs/init.md for more details on the duties maitred carries out as pid 1.
vm_syslog is the syslog daemon that runs inside the VM. It is automatically started by maitred and provides a socket at /dev/log for applications to send it log records. vm_syslog aggregates the log records and then forwards them outside the VM to the logging service running on the host. The logging service tags the records it receives with the unique identifier for the VM from which the logs originated and then forwards them on to the host syslog service. This ensures that the VM logs are captured in any feedback reports that are uploaded to Google's servers.
Additionally, vm_syslog reads kernel logs from /dev/kmsg (inside the VM) and forwards those to the logging service running on the host.
See docs/logging.md for more details on log handling.
crash_collector is responsible for collecting crash reports of applications running inside the VM and forwarding them out to the crash collector service provided by vm_launcher. When maitred first starts up it configures /proc/sys/kernel/core_pattern to start the crash_collector program and send the core dump over a pipe to that program. crash_collector then parses the core dump and converts it to a minidump before sending it out to vm_launcher. vm_launcher passes the report on to crash-reporter, which takes care of uploading it to Google servers.
All communication between vm_launcher and the applications inside the VM happen over a vsock transport. The actual RPC communication uses the gRPC framework. Every maitred instance listens on a known port in the vsock namespace (port 8888). vm_launcher instances allocate a port number via a shared pool and communicate that port number to the VM via the linux kernel command line.
See docs/vsock.md for more details about vsock.
Since each maitred instance listens on a known port number, it is possible for an application inside a VM to send a message to maitred over a loopback interface. To prevent this we block all loopback connections over vsock.
It is not possible for processes in different VMs to send messages to each other over vsock. This is blocked by the host kernel driver that manages data transfer.
gRPC uses protocol buffers as the serialization format for messages sent over the vsock transport. The proto directory holds the definitions for all the messages sent and services provided between the host and the VM.