Hi, Opening tap devices, such as macvtap, that are created in containers is problematic because the interface for opening tap devices is via /dev/tapNN and devtmpfs is not typically mounted inside a container as its not namespace aware. It is possible to do a mknod() in the container, once the tap devices are created, however, since the tap devices are created dynamically its not possible to apriori allow access to certain major/minor numbers, since we don't know what these are going to be. In addition, its desirable to not allow the mknod capability in containers. This behavior, I think is somewhat inconsistent with the tuntap driver where one can create tuntap devices inside a container by first opening /dev/net/tun and then using them by supplying the tuntap device name via the ioctl(TUNSETIFF). And since TUNSETIFF validates the network namespace, one is limited to opening network devices that belong to your current network namespace. Here are some options to this issue, that I wanted to get feedback about, and just wondering if anybody else has run into this. 1) Don't create the tap device, such as macvtap in the container. Instead, create the tap device outside of the container and then move it into the desired container network namespace. In addition, do a mknod() for the corresponding /dev/tapNN device from outside the container before doing chroot(). This solution still doesn't allow tap devices to be created inside the container. Thus, in the case of kubevirt, which runs libvirtd inside of a container, it would mean changing libvirtd to open existing tap devices (as opposed to the current behavior of creating new ones). This would not require any kernel changes, but as mentioned seems inconsistent with the tuntap interface.

2) Add a new kernel interface for tap devices similar to how /dev/net/tun currently works. It might be nice to use TUNSETIFF for tap devices, but because tap devices have different fops they can't be easily switched after open(). So the suggestion is a new ioctl (TUNGETFDBYNAME?), where the tap device name is supplied and a new fd (distinct from the fd returned by the open of /dev/net/tun) is returned as an output field as part of the new ioctl parameter. It may not make sense to have this new ioctl call for /dev/net/tun since its really about opening a tap device, so it may make sense to introduce it as part of a new device, such as /dev/net/tap. This new ioctl could be used for macvtap and ipvtap (or any tap device). I think it might also improve performance for tuntap devices themselves, if they are opened this way since currently all tun operations such as read() and write() take a reference count on the underlying tuntap device, since it can be changed via TUNSETIFF. I tested this interface out, so I can provide the kernel changes if that's helpful for clarification.

On 07/05/2018 12:10 PM, Daniel P. Berrangé wrote: > On Thu, Jul 05, 2018 at 10:20:16AM -0400, Jason Baron wrote: >> Hi, >> >> Opening tap devices, such as macvtap, that are created in containers is >> problematic because the interface for opening tap devices is via >> /dev/tapNN and devtmpfs is not typically mounted inside a container as >> its not namespace aware. It is possible to do a mknod() in the >> container, once the tap devices are created, however, since the tap >> devices are created dynamically its not possible to apriori allow access >> to certain major/minor numbers, since we don't know what these are going >> to be. In addition, its desirable to not allow the mknod capability in >> containers. This behavior, I think is somewhat inconsistent with the >> tuntap driver where one can create tuntap devices inside a container by >> first opening /dev/net/tun and then using them by supplying the tuntap >> device name via the ioctl(TUNSETIFF). And since TUNSETIFF validates the >> network namespace, one is limited to opening network devices that belong >> to your current network namespace. >> >> Here are some options to this issue, that I wanted to get feedback >> about, and just wondering if anybody else has run into this. >> >> 1) >> >> Don't create the tap device, such as macvtap in the container. Instead, >> create the tap device outside of the container and then move it into the >> desired container network namespace. In addition, do a mknod() for the >> corresponding /dev/tapNN device from outside the container before doing >> chroot(). >> >> This solution still doesn't allow tap devices to be created inside the >> container. Thus, in the case of kubevirt, which runs libvirtd inside of >> a container, it would mean changing libvirtd to open existing tap >> devices (as opposed to the current behavior of creating new ones). This >> would not require any kernel changes, but as mentioned seems >> inconsistent with the tuntap interface. > > Presumably the /dev/tapNN device name also changes when you rename > the tap device interface using SIOCSIFNAME ? > I don't think so. the NN is the ifindex of the device- changing the device name does not affect the ifindex.

> eg if it was /dev/tap24 in the host and you called SIOCSIFNAME(eth0) > when moving it into the container, it would be /dev/eth0 inside the > container ? > When moving it into the container the ifindex can change since the ifindex range is per-namespace (not global).

> Anyway, given that this /dev/tapNN approach is what exists today, > libvirt will likely want to implement support for this regardless > in order to support existing kernels. Ok, in this case whatever created the tap device outside of the container would pass the name of the device to libvirt and make sure that the /dev/tapNN device was setup correctly in the container. I believe this differs from how libvirt works today in that libvirt would need to be modified to open an existing device (I think it currently always creates new ones).

> One slight complication with either of the solutions above is that > libvirt won't know whether it is given a TAP or a MACVTAP device. > It'll only be given the device name. So with code today we would > probably have to first try /dev/tapNNN and if that doesn't exist > then try /dev/net/tun with TUNSETIFF. > hmmm. doesn't libvirt make this distinction today?

> If adding a new /dev/net/tap, something could seemlessy accept > either a TAP or MACTAP nic name would be nice. > > I think if we added a new ioctl() as I proposed it could accept either type of nic.

On Thu, Jul 5, 2018 at 4:20 PM Jason Baron <jbaron(a)akamai.com&gt; wrote: > Hi, > > Opening tap devices, such as macvtap, that are created in containers is > problematic because the interface for opening tap devices is via > /dev/tapNN and devtmpfs is not typically mounted inside a container as > its not namespace aware. It is possible to do a mknod() in the > container, once the tap devices are created, however, since the tap > devices are created dynamically its not possible to apriori allow access > to certain major/minor numbers, since we don't know what these are going > to be. In addition, its desirable to not allow the mknod capability in > containers. This behavior, I think is somewhat inconsistent with the > tuntap driver where one can create tuntap devices inside a container by > first opening /dev/net/tun and then using them by supplying the tuntap > device name via the ioctl(TUNSETIFF). And since TUNSETIFF validates the > network namespace, one is limited to opening network devices that belong > to your current network namespace. > > Here are some options to this issue, that I wanted to get feedback > about, and just wondering if anybody else has run into this. > > 1) > > Don't create the tap device, such as macvtap in the container. Instead, > create the tap device outside of the container and then move it into the > desired container network namespace. In addition, do a mknod() for the > corresponding /dev/tapNN device from outside the container before doing > chroot(). > > This solution still doesn't allow tap devices to be created inside the > container. Thus, in the case of kubevirt, which runs libvirtd inside of > a container, it would mean changing libvirtd to open existing tap > devices (as opposed to the current behavior of creating new ones). This > would not require any kernel changes, but as mentioned seems > inconsistent with the tuntap interface. > For KubeVirt, apart from how exactly the device ends up in the container, I would want to pursue a way where all network preparations which require privileges happens from a privileged process *outside* of the container. Like CNI solutions do it. They run outside, have privileges and then create devices in the right network/mount namespace or move them there. The final goal for KubeVirt is that our pod with the qemu process is completely unprivileged and privileged setup happens from outside. As a consequence, and depending on which route Dan pursues with the restructured libvirt, I would assume that either a privileged libvirtd-part outside of containers creates the devices by entering the right namespaces, or that libvirt in the container can consume pre-created tun/tap devices, like qemu.

On 07/08/2018 02:01 AM, Martin Kletzander wrote: > On Thu, Jul 05, 2018 at 06:24:20PM +0200, Roman Mohr wrote: >> On Thu, Jul 5, 2018 at 4:20 PM Jason Baron <jbaron(a)akamai.com&gt; wrote: >> >>> Hi, >>> >>> Opening tap devices, such as macvtap, that are created in containers is >>> problematic because the interface for opening tap devices is via >>> /dev/tapNN and devtmpfs is not typically mounted inside a container as >>> its not namespace aware. It is possible to do a mknod() in the >>> container, once the tap devices are created, however, since the tap >>> devices are created dynamically its not possible to apriori allow access >>> to certain major/minor numbers, since we don't know what these are going >>> to be. In addition, its desirable to not allow the mknod capability in >>> containers. This behavior, I think is somewhat inconsistent with the >>> tuntap driver where one can create tuntap devices inside a container by >>> first opening /dev/net/tun and then using them by supplying the tuntap >>> device name via the ioctl(TUNSETIFF). And since TUNSETIFF validates the >>> network namespace, one is limited to opening network devices that belong >>> to your current network namespace. >>> >>> Here are some options to this issue, that I wanted to get feedback >>> about, and just wondering if anybody else has run into this. >>> >>> 1) >>> >>> Don't create the tap device, such as macvtap in the container. Instead, >>> create the tap device outside of the container and then move it into the >>> desired container network namespace. In addition, do a mknod() for the >>> corresponding /dev/tapNN device from outside the container before doing >>> chroot(). >>> >>> This solution still doesn't allow tap devices to be created inside the >>> container. Thus, in the case of kubevirt, which runs libvirtd inside of >>> a container, it would mean changing libvirtd to open existing tap >>> devices (as opposed to the current behavior of creating new ones). This >>> would not require any kernel changes, but as mentioned seems >>> inconsistent with the tuntap interface. >>> >> >> For KubeVirt, apart from how exactly the device ends up in the >> container, I >> would want to pursue a way where all network preparations which require >> privileges happens from a privileged process *outside* of the container. >> Like CNI solutions do it. They run outside, have privileges and then >> create >> devices in the right network/mount namespace or move them there. The >> final >> goal for KubeVirt is that our pod with the qemu process is completely >> unprivileged and privileged setup happens from outside. >> >> As a consequence, and depending on which route Dan pursues with the >> restructured libvirt, I would assume that either a privileged >> libvirtd-part >> outside of containers creates the devices by entering the right >> namespaces, >> or that libvirt in the container can consume pre-created tun/tap devices, >> like qemu. >> > > That would be nice, but as far as I understand there will always be a > need for > some privileges if you want to use a tap device.  It's nice that CNI > does that > and all the containers can run unprivileged, but that's because they do > not open > the tap device and they do not do any privileged operations on it.  But > QEMU > needs to.  So the only way would be passing an opened fd to the > container or > opening the tap device there and making the fd usable for one process in > the > container.  Is this already supported for some type of containers in > some way? > > Martin Hi, So another option here call it #3 is to pass open fds via unix sockets. If there are privileged operations that QEMU is trying to do with the fd though, how will opening it first and then passing it to an unprivileged QEMU address that? Is the opener doing those operations first?

From libvirt's POV, it would be preferrable to be able to open the