The functions in virCommand() after fork() must be careful with regard
to accessing any mutexes that may have been locked by other threads in
the parent process. It is possible that another thread in the parent
process holds the lock for the virQEMUDriver while fork() is called.
This leads to a deadlock in the child process when
'virQEMUDriverGetConfig(driver)' is called and therefore the handshake
never completes between the child and the parent process. Ultimately
the virDomainObjectPtr will never be unlocked.

It gets much worse if the other thread of the parent process, that
holds the lock for the virQEMUDriver, tries to lock the already locked
virDomainObject. This leads to a completely unresponsive libvirtd.

It's possible to reproduce this case with calling 'virsh start XXX'
and 'virsh managedsave XXX' in a tight loop for multiple domains.

This commit fixes the deadlock in the same way as it is described in
commit 61b52d2e.
Signed-off-by: NMarc Hartmayer <mhartmay@linux.vnet.ibm.com>
Reviewed-by: NBoris Fiuczynski <fiuczy@linux.vnet.ibm.com>

When enabling virgl, qemu opens /dev/dri/render*. So far, we are
not allowing that in devices CGroup nor creating the file in
domain's namespace and thus requiring users to set the paths in
qemu.conf. This, however, is suboptimal as it allows access to
ALL qemu processes even those which don't have virgl configured.
Now that we have a way to specify render node that qemu will use
we can be more cautious and enable just that.
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>

So far, qemuDomainGetHostdevPath has no knowledge of the reasong
it is called and thus reports /dev/vfio/vfio for every VFIO
backed device. This is suboptimal, as we want it to:

a) report /dev/vfio/vfio on every addition or domain startup
b) report /dev/vfio/vfio only on last VFIO device being unplugged

If a domain is being stopped then namespace and CGroup die with
it so no need to worry about that. I mean, even when a domain
that's exiting has more than one VFIO devices assigned to it,
this function does not clean /dev/vfio/vfio in CGroup nor in the
namespace. But that doesn't matter.
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>
Reviewed-by: NMarc-André Lureau <marcandre.lureau@redhat.com>

So far, we are allowing /dev/vfio/vfio in the devices cgroup
unconditionally (and creating it in the namespace too). Even if
domain has no hostdev assignment configured. This is potential
security hole. Therefore, when starting the domain (or
hotplugging a hostdev) create & allow /dev/vfio/vfio too (if
needed).
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>
Reviewed-by: NMarc-André Lureau <marcandre.lureau@redhat.com>

Since these two functions are nearly identical (with
qemuSetupHostdevCgroup actually calling virCgroupAllowDevicePath)
we can have one function call the other and thus de-duplicate
some code.
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>
Reviewed-by: NMarc-André Lureau <marcandre.lureau@redhat.com>

The bare fact that mnt namespace is available is not enough for
us to allow/enable qemu namespaces feature. There are other
requirements: we must copy all the ACL & SELinux labels otherwise
we might grant access that is administratively forbidden or vice
versa.
At the same time, the check for namespace prerequisites is moved
from domain startup time to qemu.conf parser as it doesn't make
much sense to allow users to start misconfigured libvirt just to
find out they can't start a single domain.
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>

mknod() is affected my the current umask, so we're not
guaranteed the newly-created device node will have the
right permissions.

Call chmod(), which is not affected by the current umask,
immediately afterwards to solve the issue.

Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=1421036

This controller only allows up to 15 ports.

https://bugzilla.redhat.com/show_bug.cgi?id=1375417

Just like we need wrappers over other virSecurityManager APIs, we
need one for virSecurityManagerSetImageLabel and
virSecurityManagerRestoreImageLabel. Otherwise we might end up
relabelling device in wrong namespace.
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>

Firstly, instead of checking for next->path the
virStorageSourceIsEmpty() function should be used which also
takes disk type into account.
Secondly, not every disk source passed has the correct type set
(due to our laziness). Therefore, instead of checking for
virStorageSourceIsBlockLocal() and also S_ISBLK() the former can
be refined to just virStorageSourceIsLocalStorage().
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>

Again, one missed bit. This time without this commit there is no
/dev entry  in the namespace of the qemu process when doing disk
snapshots or block-copy.
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>

These functions do not need to see the whole virDomainDiskDef.
Moreover, they are going to be called from places where we don't
have access to the full disk definition. Sticking with
virStorageSource is more than enough.
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>

Since its introduction in 81df2150 this variable was never
used.
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>

There is no need for this. None of the namespace helpers uses it.
Historically it was used when calling secdriver APIs, but we
don't to that anymore.
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>

In order for memory locking to work, the hard limit on memory
locking (and usage) has to be set appropriately by the user.

The documentation mentions the requirement already: with this
patch, it's going to be enforced by runtime checks as well,
by forbidding a non-compliant guest from being defined as well
as edited and started.

Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=1316774

When working with symlinks it is fairly easy to get into a loop.
Don't.
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>

Similarly to one of the previous commits, we need to deal
properly with symlinks in hotplug case too.
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>

When working with symlinks it is fairly easy to get into a loop.
Don't.
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>

Imagine you have a disk with the following source set up:

/dev/disk/by-uuid/$uuid (symlink to) -> /dev/sda

After cbc45525 the transitive end of the symlink chain is
created (/dev/sda), but we need to create any item in chain too.
Others might rely on that.
In this case, /dev/disk/by-uuid/$uuid comes from domain XML thus
it is this path that secdriver tries to relabel. Not the resolved
one.
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>

After previous commit this has become redundant step.
Also setting up devices in namespace and setting their label
later on are two different steps and should be not done at once.
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>

Not only we should set the MTU on the host end of the device but
also let qemu know what MTU did we set.
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>

So far we allow to set MTU for libvirt networks. However, not all
domain interfaces have to be plugged into a libvirt network and
even if they are, they might want to have a different MTU (e.g.
for testing purposes).
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>

We lacked of timestamp in tainting of guests log,
which bring troubles for finding guest issues:
such as whether a guest powerdown caused by qemu-monitor-command
or others issues inside guests.
If we had timestamp in tainting of guests log,
it would be helpful when checking guest's /var/log/messages.
Signed-off-by: NChen Hanxiao <chenhanxiao@gmail.com>

Based on work of Mehdi Abaakouk <sileht@sileht.net>.

When parsing vhost-user interface XML and no ifname is found we
can try to fill it in in post parse callback. The way this works
is we try to make up interface name from given socket path and
then ask openvswitch whether it knows the interface.
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>

https://bugzilla.redhat.com/show_bug.cgi?id=1413922

While all the code that deals with qemu namespaces correctly
detects whether we are running as root (and turn into NO-OP for
qemu:///session) the actual unshare() call is not guarded with
such check. Therefore any attempt to start a domain under
qemu:///session shall fail as unshare() is reserved for root.

The fix consists of moving unshare() call (for which we have a
wrapper called virProcessSetupPrivateMountNS) into
qemuDomainBuildNamespace() where the proper check is performed.
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>
Tested-by: NRichard W.M. Jones <rjones@redhat.com>

When creating new /dev/* for qemu, we do chown() and copy ACLs to
create the exact copy from the original /dev. I though that
copying SELinux labels is not necessary as SELinux will chose the
sane defaults. Surprisingly, it does not leaving namespace with
the following labels:

crw-rw-rw-. root root system_u:object_r:tmpfs_t:s0     random
crw-------. root root system_u:object_r:tmpfs_t:s0     rtc0
drwxrwxrwt. root root system_u:object_r:tmpfs_t:s0     shm
crw-rw-rw-. root root system_u:object_r:tmpfs_t:s0     urandom

As a result, domain is unable to start:

error: internal error: process exited while connecting to monitor: Error in GnuTLS initialization: Failed to acquire random data.
qemu-kvm: cannot initialize crypto: Unable to initialize GNUTLS library: Failed to acquire random data.

The solution is to copy the SELinux labels as well.
Reported-by: NAndrea Bolognani <abologna@redhat.com>
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>

So far the decision whether /dev/* entry is created in the qemu
namespace is really simple: does the path starts with "/dev/"?
This can be easily fooled by providing path like the following
(for any considered device like disk, rng, chardev, ..):

  /dev/../var/lib/libvirt/images/disk.qcow2

Therefore, before making the decision the path should be
canonicalized.
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>

So far the namespaces were turned on by default unconditionally.
For all non-Linux platforms we provided stub functions that just
ignored whatever namespaces setting there was in qemu.conf and
returned 0 to indicate success. Moreover, we didn't really check
if namespaces are available on the host kernel.

This is suboptimal as we might have ignored user setting.
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>

This is a simple wrapper over mount(). However, not every system
out there is capable of moving a mount point. Therefore, instead
of having to deal with this fact in all the places of our code we
can have a simple wrapper and deal with this fact at just one
place.
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>

Due to a copy-paste error, the debug message reads:

  Setting up disks

It should have been:

  Setting up inputs.
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>

This function is used only from code compiled on Linux. Therefore
on non-Linux platforms it triggers compilation error:

../../src/qemu/qemu_domain.c:209:1: error: unused function 'qemuDomainGetPreservedMounts' [-Werror,-Wunused-function]
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>

After previous commits, this function is no longer needed.
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>

Again, there is no need to create /var/lib/libvirt/$domain.*
directories in CreateNamespace(). It is sufficient to create them
as soon as we need them which is in BuildNamespace. This way we
don't leave them around for the whole lifetime of domain.
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>

The c1140eb9 got me thinking. We don't want to special case /dev
in qemuDomainGetPreservedMounts(), but in all other places in the
code we special case it anyway. I mean,
/var/run/libvirt/$domain.dev path is constructed separately just
so that it is not constructed here. It makes only a little sense
(if any at all).
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>

If something goes wrong in this function we try a rollback. That
is unlink all the directories we created earlier. For some weird
reason unlink() was called instead of rmdir().
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>

Just so it doesn't bite us in the future, even though it's unlikely.

And fix the comment above it as well.  Commit e08ee7cd took the
info from the function it's calling, but that was lie itself in the
first place.
Signed-off-by: NMartin Kletzander <mkletzan@redhat.com>

With my namespace patches, we are spawning qemu in its own
namespace so that we can manage /dev entries ourselves. However,
some filesystems mounted under /dev needs to be preserved in
order to be shared with the parent namespace (e.g. /dev/pts).
Currently, the list of mount points to preserve is hardcoded
which ain't right - on some systems there might be less or more
items under real /dev that on our list. The solution is to parse
/proc/mounts and fetch the list from there.
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>

Again, not something that I'd hit, but there is a chance in
theory that this might bite us. Currently the way we decide
whether or not to create /dev entry for a device is by marching
first four characters of path with "/dev". This might be not
enough. Just imagine somebody has a disk image stored under
"/devil/path/to/disk". We ought to be matching against "/dev/".
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>

Not that I'd encounter any bug here, but the code doesn't look
100% correct. Imagine, somebody is trying to attach a device to a
domain, and the device's /dev entry already exists in the qemu
namespace. This is handled gracefully and the control continues
with setting up ACLs and calling security manager to set up
labels. Now, if any of these steps fail, control jump on the
'cleanup' label and unlink() the file straight away. Even when it
was not us who created the file in the first place. This can be
possibly dangerous.
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>

https://bugzilla.redhat.com/show_bug.cgi?id=1406837

Imagine you have a domain configured in such way that you are
assigning two PCI devices that fall into the same IOMMU group.
With mount namespace enabled what happens is that for the first
PCI device corresponding /dev/vfio/X entry is created and when
the code tries to do the same for the second mknod() fails as
/dev/vfio/X already exists:

2016-12-21 14:40:45.648+0000: 24681: error :
qemuProcessReportLogError:1792 : internal error: Process exited
prior to exec: libvirt: QEMU Driver error : Failed to make device
/var/run/libvirt/qemu/windoze.dev//vfio/22: File exists

Worse, by default there are some devices that are created in the
namespace regardless of domain configuration (e.g. /dev/null,
/dev/urandom, etc.). If one of them is set as backend for some
guest device (e.g. rng, chardev, etc.) it's the same story as
described above.

Weirdly, in attach code this is already handled.
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>