[libvirt] Suspending access to opened/active /dev/nodes during application runtime
Problem: Has anyone thought about a mechanism to limit/remove an access to a device during an application runtime? Meaning we have an application that has an open file descriptor to some /dev/node and depending on *something* it gains or looses the access to it gracefully (with or without a notification, but without any fatal consequences). Example: LXC. Imagine we have 2 separate containers. Both running full operating systems. Specifically with 2 X servers. Both running concurrently of course. Both need the same input devices (e.g. we have just one mouse). This creates a security problem when we want to have completely separate environments. One container is active (being displayed on a monitor and controlled with a mouse) while the other container runs evtest /dev/input/something and grabs the secret password user typed in the other. Solutions: The complete solution would comprise of 2 parts: - a mechanism that would allow to temporally "hide" a device from an open file descriptor. - a mechanism for deciding whether application/process/namespace should have an access to a specific device at a specific moment Let's focus on the first problem only, as it would need to be solved first anyway. I haven't found anything that would allow me to do it. There are a lot mechanisms that make it possible to restrict an access during open(): - DAC - ACL (controlled by hand or with uaccess) - LSM (in general) - device cgroups But all of those can't do a thing when the device is already opened and an application has a file descriptor. I don't see such mechanism in kernel sources either. I do imagine that it would not be possible for every device to handle such a thing (dri comes to mind) without breaking something (graphics card state in dri example). But there is class of simple input/output devices that would handle this without problems. I did implement some proof-of-concept solution for an evdev driver by allowing or disallowing events that go to evdev_client structure using some arbitrary condition. But this is far from a generic solution. My proof-of-concept is somewhat similar to this (I just found it): http://www.spinics.net/lists/linux-input/msg25547.html Though a little bit wider in scope. But neither is flawless nor generic. Has anyone had any thoughts about a similar problem? -- Regards Havner
On Fri, 07.03.14 19:45, Lukasz Pawelczyk (havner@gmail.com) wrote:
Problem: Has anyone thought about a mechanism to limit/remove an access to a device during an application runtime? Meaning we have an application that has an open file descriptor to some /dev/node and depending on *something* it gains or looses the access to it gracefully (with or without a notification, but without any fatal consequences).
logind can mute input devices as sessions are switched, to enable unpriviliged X11 and wayland compositors.
Example: LXC. Imagine we have 2 separate containers. Both running full operating systems. Specifically with 2 X servers. Both running concurrently of
Well, devices are not namespaced on Linux (with the single exception of network devices). An X server needs device access, hence this doesn't fly at all. When you enumerate devices with libudev in a container they will never be marked as "initialized" and you do not get any udev hotplug events in containers, and you don#t have the host's udev db around, nor would it make any sense to you if you had. X11 and friends rely on udev however... Before you think about doing something like this, you need to fix the kernel to provide namespaced devices (good luck!)
course. Both need the same input devices (e.g. we have just one mouse). This creates a security problem when we want to have completely separate environments. One container is active (being displayed on a monitor and controlled with a mouse) while the other container runs evtest /dev/input/something and grabs the secret password user typed in the other.
logind can do this for you between sessions. But such a container setup will never work without proper device namespacing.
Solutions: The complete solution would comprise of 2 parts: - a mechanism that would allow to temporally "hide" a device from an open file descriptor. - a mechanism for deciding whether application/process/namespace should have an access to a specific device at a specific moment
Well, there's no point in inventing any "mechanisms" like this, as long as devices are not namespaced in the kernel, so that userspace in containers can enumerate/probe/identify/... things correctly... Lennart -- Lennart Poettering, Red Hat
On 7 Mar 2014, at 20:24, Lennart Poettering <mzerqung@0pointer.de> wrote:
On Fri, 07.03.14 19:45, Lukasz Pawelczyk (havner@gmail.com) wrote:
Problem: Has anyone thought about a mechanism to limit/remove an access to a device during an application runtime? Meaning we have an application that has an open file descriptor to some /dev/node and depending on *something* it gains or looses the access to it gracefully (with or without a notification, but without any fatal consequences).
logind can mute input devices as sessions are switched, to enable unpriviliged X11 and wayland compositors.
Would you please elaborate on this? Where is this mechanism? How does it work without kernel space support? Is there some kernel space support I’m not aware of?
Example: LXC. Imagine we have 2 separate containers. Both running full operating systems. Specifically with 2 X servers. Both running concurrently of
Well, devices are not namespaced on Linux (with the single exception of network devices). An X server needs device access, hence this doesn't fly at all.
When you enumerate devices with libudev in a container they will never be marked as "initialized" and you do not get any udev hotplug events in containers, and you don#t have the host's udev db around, nor would it make any sense to you if you had. X11 and friends rely on udev however...
Before you think about doing something like this, you need to fix the kernel to provide namespaced devices (good luck!)
Precisly! That’s the generic idea. I’m not for implementing it though at this moment. I just wanted to know whether anybody actually though about it or maybe someone is interested in starting such a work, etc.
course. Both need the same input devices (e.g. we have just one mouse). This creates a security problem when we want to have completely separate environments. One container is active (being displayed on a monitor and controlled with a mouse) while the other container runs evtest /dev/input/something and grabs the secret password user typed in the other.
logind can do this for you between sessions. But such a container setup will never work without proper device namespacing.
So how can it do it when there is no kernel support? You mean it could be doing this if the support were there?
Solutions: The complete solution would comprise of 2 parts: - a mechanism that would allow to temporally "hide" a device from an open file descriptor. - a mechanism for deciding whether application/process/namespace should have an access to a specific device at a specific moment
Well, there's no point in inventing any "mechanisms" like this, as long as devices are not namespaced in the kernel, so that userspace in containers can enumerate/probe/identify/... things correctly…
True. My point is about kernel space implementation. Like I wrote. I haven’t seen anything like this in kernel source and I’m well away it should be done there. I would just like to know if anybody is interested in this, if anybody started or would like to start such a thing. I do understand that systemd/logind would only provide a mechanism for determining who should have an access and who shouldn’t (or to be more specific it would utilize some kernel space configuration like cgroups). But the work itself has to be done in kernel space. -- Regards, Havner
On Fri, 07.03.14 21:51, Lukasz Pawelczyk (havner@gmail.com) wrote:
Problem: Has anyone thought about a mechanism to limit/remove an access to a device during an application runtime? Meaning we have an application that has an open file descriptor to some /dev/node and depending on *something* it gains or looses the access to it gracefully (with or without a notification, but without any fatal consequences).
logind can mute input devices as sessions are switched, to enable unpriviliged X11 and wayland compositors.
Would you please elaborate on this? Where is this mechanism? How does it work without kernel space support? Is there some kernel space support I’m not aware of?
There's EVIOCREVOKE for input devices and DRM_IOCTL_SET_MASTER/DRM_IOCTL_DROP_MASTER for DRM devices. See logind sources.
Before you think about doing something like this, you need to fix the kernel to provide namespaced devices (good luck!)
Precisly! That’s the generic idea. I’m not for implementing it though at this moment. I just wanted to know whether anybody actually though about it or maybe someone is interested in starting such a work, etc.
It's not just about turning on and turning off access to the event stream. It's mostly about enumeration and probing which doesn't work in containers, and is particularly messy if you intend to share devices between containers.
logind can do this for you between sessions. But such a container setup will never work without proper device namespacing.
So how can it do it when there is no kernel support? You mean it could be doing this if the support were there?
EVIOCREVOKE and the DRM ioctls are pretty real... Lennart -- Lennart Poettering, Red Hat
On Fri, Mar 7, 2014 at 3:51 PM, Lukasz Pawelczyk <havner@gmail.com> wrote:
On 7 Mar 2014, at 20:24, Lennart Poettering <mzerqung@0pointer.de> wrote:
On Fri, 07.03.14 19:45, Lukasz Pawelczyk (havner@gmail.com) wrote:
Problem: Has anyone thought about a mechanism to limit/remove an access to a device during an application runtime? Meaning we have an application that has an open file descriptor to some /dev/node and depending on *something* it gains or looses the access to it gracefully (with or without a notification, but without any fatal consequences).
logind can mute input devices as sessions are switched, to enable unpriviliged X11 and wayland compositors.
Would you please elaborate on this? Where is this mechanism? How does it work without kernel space support? Is there some kernel space support I'm not aware of?
Example: LXC. Imagine we have 2 separate containers. Both running full operating systems. Specifically with 2 X servers. Both running concurrently of
Well, devices are not namespaced on Linux (with the single exception of network devices). An X server needs device access, hence this doesn't fly at all.
When you enumerate devices with libudev in a container they will never be marked as "initialized" and you do not get any udev hotplug events in containers, and you don#t have the host's udev db around, nor would it make any sense to you if you had. X11 and friends rely on udev however...
Before you think about doing something like this, you need to fix the kernel to provide namespaced devices (good luck!)
Precisly! That's the generic idea. I'm not for implementing it though at this moment. I just wanted to know whether anybody actually though about it or maybe someone is interested in starting such a work, etc.
Yes, we have started such a thing. Here is the link to the wiki: https://github.com/Cellrox/devns-patches/wiki [...] Oren.
Hi On Fri, Mar 7, 2014 at 7:45 PM, Lukasz Pawelczyk <havner@gmail.com> wrote:
Problem: Has anyone thought about a mechanism to limit/remove an access to a device during an application runtime? Meaning we have an application that has an open file descriptor to some /dev/node and depending on *something* it gains or looses the access to it gracefully (with or without a notification, but without any fatal consequences).
Example: LXC. Imagine we have 2 separate containers. Both running full operating systems. Specifically with 2 X servers. Both running concurrently of course. Both need the same input devices (e.g. we have just one mouse). This creates a security problem when we want to have completely separate environments. One container is active (being displayed on a monitor and controlled with a mouse) while the other container runs evtest /dev/input/something and grabs the secret password user typed in the other.
Solutions: The complete solution would comprise of 2 parts: - a mechanism that would allow to temporally "hide" a device from an open file descriptor. - a mechanism for deciding whether application/process/namespace should have an access to a specific device at a specific moment
Let's focus on the first problem only, as it would need to be solved first anyway. I haven't found anything that would allow me to do it. There are a lot mechanisms that make it possible to restrict an access during open(): - DAC - ACL (controlled by hand or with uaccess) - LSM (in general) - device cgroups But all of those can't do a thing when the device is already opened and an application has a file descriptor. I don't see such mechanism in kernel sources either.
I do imagine that it would not be possible for every device to handle such a thing (dri comes to mind) without breaking something (graphics card state in dri example). But there is class of simple input/output devices that would handle this without problems.
I did implement some proof-of-concept solution for an evdev driver by allowing or disallowing events that go to evdev_client structure using some arbitrary condition. But this is far from a generic solution.
My proof-of-concept is somewhat similar to this (I just found it): http://www.spinics.net/lists/linux-input/msg25547.html Though a little bit wider in scope. But neither is flawless nor generic.
Has anyone had any thoughts about a similar problem?
Lennart and Greg have already answered most of this, few notes from me: * EVIOCREVOKE and DRM_SET_MASTER/DROP_MASTER are real. We use them. They solve your problem for gfx and input devices. * EVIOCMUTE is *bad*. It is a privileged ioctl compared to EVIOCREVOKE, so we've never merged it. It neither has major advantages over revoke. So use EVIOCREVOKE. * A generic frevoke() syscall would solve all is, but is unlikely to ever appear upstream. Cheers David
participants (4)
-
David Herrmann -
Lennart Poettering -
Lukasz Pawelczyk -
Oren Laadan