Re: [CFT][PATCH 00/10] Making new mounts of proc and sysfs as safe as bind mounts (take 2)

On Thu, May 28, 2015 at 8:03 AM, Eric W. Biederman
[off-list ref] wrote:

Serge Hallyn [off-list ref] writes:

quoted

Quoting Andy Lutomirski (luto@amacapital.net):

quoted

On Fri, May 22, 2015 at 10:39 AM, Eric W. Biederman
[off-list ref] wrote:

quoted

I had hoped to get some Tested-By's on that patch series.

Sorry, I've been totally swamped.

I suspect that Sandstorm is okay, but I haven't had a chance to test
it for real.  Sandstorm makes only limited use of proc and sysfs in
containers, but I'll see if I can test it for real this weekend.

Testing this with unprivileged containers, I get

lxc-start: conf.c: lxc_mount_auto_mounts: 808 Operation not permitted
- error mounting sysfs on
/usr/lib/x86_64-linux-gnu/lxc/sys/devices/virtual/net flags 0

Grr..  I was afraid this would break something. :(

Looking at my system I see that sysfs is currently mounted
"nosuid,nodev,noexec"

Looking at the lxc-start code I don't see it as including any of those
mount options.  In practice for sysfs I think those options are
meaningless (as there should be no devices and nothing executable in
sysfs) but I can understand the past concerns with chmod on virtual
filesystems that would incline people to use them, so I think the
failure is reporting a legitimate security issue in the lxc userspace
code where the the unprivileged code is currently attempting to give
greater access to sysfs than was given by the original mount of sysfs.

As nosuid,nodev,noexec should not impair the operation of sysfs
operation it looks like you can always specify those options and just
make this concern go away.

Linus is pretty strict about not breaking the ABI, and this definitely
counts as breaking the ABI.  There's an exception for security issues,
but is there really a security issue here?  That is, do we lose
anything important if we just drop the offending part of the patch
set?  As you've said, there shouldn't be sensitive device nodes,
executables, or setuid files in proc or sysfs in the first place.

--Andy