[Devel] [TRD] Autofs migration

[Stanislav Kinsburskiy]

1. Feature

Autofs mount points migration via CRIU
https://jira.sw.ru/browse/PSBM-41217

2. Description

CRIU now supports autofs file system migration, including direct, 
indirect and offset mount types.

3. Products

Virtuozzo 7

Packages:
     criu-2.1.0.4.vz7
     libvzctl-7.0.199

4. Testing

4.1 Basics
     ** Install criu and libvzctl rpm packages
     ** Create a container, and check
     ** Check, that autofs is listed in /proc/filesystems in the container
     ** Check, that /dev/autofs is accessible
     ** Install autofs package inside the container
     ** Follow autofs guide to create an autofs _direct_ mount point 
with some file system, mounted on top (tmpfs, for example). Command "man 
autofs" might help
     ** Follow autofs guide to create an autofs _indirect_ mount point 
with some file system, mounted on top (tmpfs, for example).
  ** Follow autofs guide to create an autofs _offset_ mount point with 
some file system, mounted on top (tmpfs, for example).
     ** Suspend and restore container
     ** Check, that autofs mounts and nested were mounts migrated 
successfully (via /proc, for example).

4.2 Systemd autofs services
     ** Start any systemds autofs service (for example, 
proc-sys-fs-binfmt_misc.automount) in the container
     ** Check, that service started successfully
     ** Suspend and restore container
     ** Check, that autofs and nested mount points were migrated 
successfully.
     ** Check, that systemd service has active status
     ** Unmount nested file system manually
     ** Access systemd autofs mount point and check, that nested file 
system is re-mounted again

4.3 Automount expiration
     ** setup autofs mount  with short timeout (10 seconds, for example) 
in a container via any master: automount, systemd or else
     ** Activate autofs mount point (nested mount point should be 
mounted by autofs master)
     ** Migrate (or suspend/resume) the container.
     ** Check, that nested mount point is unmounted after restore within 
timeout.

5. Known issues

Autofs migration has an issue, related to systemd-controlled autofs 
mount points. Systemd saves autofs mount point device number in it's 
internals and compare this number to actual one, taken  from mount 
point, on each autofs request from kernel (mount, umount, expire, etc).
The problem is that after migration all mount points are created 
manually and has _another_ device id, which leads to ignorance of kernel 
requests from systemd side.
This problem can't be solved without some kind of "device namespaces" 
abstraction. However, some of the systemd services like 
proc-sys-fs-binfmt_misc.automount can be painlessly restarted after 
restore, thus illuminating this issue.
Restart of proc-sys-fs-binfmt_misc.automount service is done by CRIU via 
action script, provided by vzctl.

6. What was checked by developer

Both 4.1 and 4.2 test sequences

7. Feature owners

skinsbursky at virtuozzo.com