Some applications want to write to /proc. For instance:
docker run -it centos groupadd foo
Gives: groupadd: failure while writing changes to /etc/group
And strace reveals why:
open("/proc/self/task/13/attr/fscreate", O_RDWR) = -1 EROFS (Read-only file system)
I've looked at what other systems do, and systemd-nspawn makes /proc read-write
and /proc/sys readonly, while lxc allows "proc:mixed" which does the same,
plus it makes /proc/sysrq-trigger also readonly.
The later seems like a prudent idea, so we follows lxc proc:mixed.
Additionally we make /proc/irq and /proc/bus, as these seem to let
you control various hardware things.
Docker-DCO-1.1-Signed-off-by: Alexander Larsson <alexl@redhat.com> (github: alexlarsson)
We don't have the flexibility to do extra things with lxc because it is
a black box and most fo the magic happens before we get a chance to
interact with it in dockerinit.
Docker-DCO-1.1-Signed-off-by: Michael Crosby <michael@crosbymichael.com> (github: crosbymichael)
There is not need for the remount hack, we use aa_change_onexec so the
apparmor profile is not applied until we exec the users app.
Docker-DCO-1.1-Signed-off-by: Michael Crosby <michael@crosbymichael.com> (github: crosbymichael)
This also cleans up some of the left over restriction paths code from
before.
Docker-DCO-1.1-Signed-off-by: Michael Crosby <michael@crosbymichael.com> (github: crosbymichael)
It has been pointed out that some files in /proc and /sys can be used
to break out of containers. However, if those filesystems are mounted
read-only, most of the known exploits are mitigated, since they rely
on writing some file in those filesystems.
This does not replace security modules (like SELinux or AppArmor), it
is just another layer of security. Likewise, it doesn't mean that the
other mitigations (shadowing parts of /proc or /sys with bind mounts)
are useless. Those measures are still useful. As such, the shadowing
of /proc/kcore is still enabled with both LXC and native drivers.
Special care has to be taken with /proc/1/attr, which still needs to
be mounted read-write in order to enable the AppArmor profile. It is
bind-mounted from a private read-write mount of procfs.
All that enforcement is done in dockerinit. The code doing the real
work is in libcontainer. The init function for the LXC driver calls
the function from libcontainer to avoid code duplication.
Docker-DCO-1.1-Signed-off-by: Jérôme Petazzoni <jerome@docker.com> (github: jpetazzo)
This duplicates some of the Exec code but I think it it worth it because
the native driver is more straight forward and does not have the
complexity have handling the type issues for now.
Docker-DCO-1.1-Signed-off-by: Michael Crosby <michael@crosbymichael.com> (github: crosbymichael)
This temp. expands the Exec method's signature but adds a more robust
way to know when the container's process is actually released and begins
to run. The network interfaces are not guaranteed to be up yet but this
provides a more accurate view with a single callback at this time.
Docker-DCO-1.1-Signed-off-by: Michael Crosby <michael@crosbymichael.com> (github: crosbymichael)
Without this patch, containers inherit the open file descriptors of the daemon, so my "exec 42>&2" allows us to "echo >&42 some nasty error with some bad advice" directly into the daemon log. :)
Also, "hack/dind" was already doing this due to issues caused by the inheritance, so I'm removing that hack too since this patch obsoletes it by generalizing it for all containers.
Docker-DCO-1.1-Signed-off-by: Andrew Page <admwiggin@gmail.com> (github: tianon)
This has every container using the docker daemon's pid for the processes
label so it does not work correctly.
Docker-DCO-1.1-Signed-off-by: Michael Crosby <michael@crosbymichael.com> (github: crosbymichael)
These are unnecessary since the user package handles these cases properly already (as evidenced by the LXC backend not having these special cases).
I also updated the errors returned to match the other libcontainer error messages in this same file.
Also, switching from Setresuid to Setuid directly isn't a problem, because the "setuid" system call will automatically do that if our own effective UID is root currently: (from `man 2 setuid`)
setuid() sets the effective user ID of the calling process. If the
effective UID of the caller is root, the real UID and saved set-user-
ID are also set.
Docker-DCO-1.1-Signed-off-by: Andrew Page <admwiggin@gmail.com> (github: tianon)
Victor Marmol
Alexander Larsson
Sridhar Ratnakumar
Bernerd Schaefer
Michael Crosby
Michael Crosby
Rohit Jnagal
Jérôme Petazzoni
Guillaume J. Charmes
Tianon Gravi
Victor Vieux