Consider the following XML snippets:

  $ cat scsicontroller.xml
      <controller type='scsi' model='virtio-scsi' index='0'/>
  $ cat scsihostdev.xml
      <hostdev mode='subsystem' type='scsi'>
        <source>
          <adapter name='scsi_host0'/>
          <address bus='0' target='8' unit='1074151456'/>
        </source>
      </hostdev>

If we create a guest that includes the contents of scsihostdev.xml,
but forget the virtio-scsi controller described in scsicontroller.xml,
one is silently created for us.  The same holds true when attaching
a hostdev before the matching virtio-scsi controller.
(See qemuDomainFindOrCreateSCSIDiskController for context.)

Detaching the hostdev, followed by the controller, works well and the
guest behaves appropriately.

If we detach the virtio-scsi controller device first, any associated
hostdevs are detached for us by the underlying virtio-scsi code (this
is fine, since the connection is broken).  But all is not well, as the
guest is unable to receive new virtio-scsi devices (the attach commands
succeed, but devices never appear within the guest), nor even be
shutdown, after this point.

While this is not libvirt's problem, we can prevent falling into this
scenario by checking if a controller is being used by any hostdev
devices.  The same is already done for disk elements today.

Applying this patch and then using the XML snippets from earlier:

  $ virsh detach-device guest_01 scsicontroller.xml
  error: Failed to detach device from scsicontroller.xml
  error: operation failed: device cannot be detached: device is busy

  $ virsh detach-device guest_01 scsihostdev.xml
  Device detached successfully

  $ virsh detach-device guest_01 scsicontroller.xml
  Device detached successfully
Signed-off-by: NEric Farman <farman@linux.vnet.ibm.com>
Reviewed-by: NBjoern Walk <bwalk@linux.vnet.ibm.com>
Reviewed-by: NBoris Fiuczynski <fiuczy@linux.vnet.ibm.com>

Although nearly all host devices that are assigned to guests using
VFIO ("<hostdev>" devices in libvirt) are physically PCI Express
devices, until now libvirt's PCI address assignment has always
assigned them addresses on legacy PCI controllers in the guest, even
if the guest's machinetype has a PCIe root bus (e.g. q35 and
aarch64/virt).

This patch tries to assign them to an address on a PCIe controller
instead, when appropriate. First we do some preliminary checks that
might allow setting the flags without doing any extra work, and if
those conditions aren't met (and if libvirt is running privileged so
that it has proper permissions), we perform the (relatively) time
consuming task of reading the device's PCI config to see if it is an
Express device. If this is successful, the connect flags are set based
on the result, but if we aren't able to read the PCI config (most
likely due to the device not being present on the system at the time
of the check) we assume it is (or will be) an Express device, since
that is almost always the case anyway.

If libvirtd is running unprivileged, it can open a device's PCI config
data in sysfs, but can only read the first 64 bytes. But as part of
determining whether a device is Express or legacy PCI,
qemuDomainDeviceCalculatePCIConnectFlags() will be updated in a future
patch to call virPCIDeviceIsPCIExpress(), which tries to read beyond
the first 64 bytes of the PCI config data and fails with an error log
if the read is unsuccessful.

In order to avoid creating a parallel "quiet" version of
virPCIDeviceIsPCIExpress(), this patch passes a virQEMUDriverPtr down
through all the call chains that initialize the
qemuDomainFillDevicePCIConnectFlagsIterData, and saves the driver
pointer with the rest of the iterdata so that it can be used by
qemuDomainDeviceCalculatePCIConnectFlags(). This pointer isn't used
yet, but will be used in an upcoming patch (that detects Express vs
legacy PCI for VFIO assigned devices) to examine driver->privileged.

Restarting libvirtd on the source host at the end of migration when a
domain is already running on the destination would cause image labels to
be reset effectively killing the domain. Commit e8d0166e fixed similar
issue on the destination host, but kept the source always resetting the
labels, which was mostly correct except for the specific case handled by
this patch.

https://bugzilla.redhat.com/show_bug.cgi?id=1343858Signed-off-by: NJiri Denemark <jdenemar@redhat.com>

Post-copy migration needs bi-directional communication between the
source and the destination QEMU processes, which is not supported by
tunnelled migration.

https://bugzilla.redhat.com/show_bug.cgi?id=1371358Signed-off-by: NJiri Denemark <jdenemar@redhat.com>

Thanks to the complex capability caching code virQEMUCapsProbeQMP was
never called when we were starting a new qemu VM. On the other hand,
when we are reconnecting to the qemu process we reload the capability
list from the status XML file. This means that the flag preventing the
function being called was not set and thus we partially reprobed some of
the capabilities.

The recent addition of CPU hotplug clears the
QEMU_CAPS_QUERY_HOTPLUGGABLE_CPUS if the machine does not support it.
The partial re-probe on reconnect results into attempting to call the
unsupported command and then killing the VM.

Remove the partial reprobe and depend on the stored capabilities. If it
will be necessary to reprobe the capabilities in the future, we should
do a full reprobe rather than this partial one.

QEMU 2.8.0 adds support for unavailable-features in
query-cpu-definitions reply. The unavailable-features array lists CPU
features which prevent a corresponding CPU model from being usable on
current host. It can only be used when all the unavailable features are
disabled. Empty array means the CPU model can be used without
modifications.

We can use unavailable-features for providing CPU model usability info
in domain capabilities XML:

    <domainCapabilities>
      ...
      <cpu>
        <mode name='host-passthrough' supported='yes'/>
        <mode name='host-model' supported='yes'>
          <model fallback='allow'>Skylake-Client</model>
          ...
        </mode>
        <mode name='custom' supported='yes'>
          <model usable='yes'>qemu64</model>
          <model usable='yes'>qemu32</model>
          <model usable='no'>phenom</model>
          <model usable='yes'>pentium3</model>
          <model usable='yes'>pentium2</model>
          <model usable='yes'>pentium</model>
          <model usable='yes'>n270</model>
          <model usable='yes'>kvm64</model>
          <model usable='yes'>kvm32</model>
          <model usable='yes'>coreduo</model>
          <model usable='yes'>core2duo</model>
          <model usable='no'>athlon</model>
          <model usable='yes'>Westmere</model>
          <model usable='yes'>Skylake-Client</model>
          ...
        </mode>
      </cpu>
      ...
    </domainCapabilities>
Signed-off-by: NJiri Denemark <jdenemar@redhat.com>

"host" CPU model is supported by a special host-passthrough CPU mode and
users is not allowed to specify this model directly with custom mode.
Thus we should not advertise "host" CPU model in domain capabilities.
This worked well on architectures for which libvirt provides a list of
supported CPU models in cpu_map.xml (since "host" is not in the list).
But we need to explicitly filter "host" model out for all other
architectures.
Signed-off-by: NJiri Denemark <jdenemar@redhat.com>

CPU models (and especially some additional details which we will start
probing for later) differ depending on the accelerator. Thus we need to
call query-cpu-definitions in both KVM and TCG mode to get all data we
want.

Tests in tests/domaincapstest.c are temporarily switched to TCG to avoid
having to squash even more stuff into this single patch. They will all
be switched back later in separate commits.
Signed-off-by: NJiri Denemark <jdenemar@redhat.com>

This patch moves the CPU models formatting code from
virQEMUCapsFormatCache into a separate function.
Signed-off-by: NJiri Denemark <jdenemar@redhat.com>

This patch moves the CPU models parsing code from virQEMUCapsLoadCache
into a separate function.
Signed-off-by: NJiri Denemark <jdenemar@redhat.com>

The function just returned cached capabilities without checking whether
they are still valid. We should check that and refresh the capabilities
to make sure we don't return stale data. In other words, we should do
what all other lookup functions do.
Signed-off-by: NJiri Denemark <jdenemar@redhat.com>

The function is made a little bit more readable and the code which
refreshes cached capabilities if they are not valid any more was moved
into a separate function (virQEMUCapsCacheValidate) so that it can be
reused in other places.
Signed-off-by: NJiri Denemark <jdenemar@redhat.com>

If a user asked for a KVM domain capabilities when KVM is not available,
we would happily return data we got when probing through TCG and
pretended they were relevant for KVM. Let's just report KVM is not
supported to avoid confusion.
Signed-off-by: NJiri Denemark <jdenemar@redhat.com>

When domain capabilities were introduced we did not have enough data to
decide whether KVM works on the host or not and thus working legacy/VFIO
device assignment was used as a witness. Now that we know whether KVM
was enabled when probing QEMU capabilities (and thus we know it's
working), we can use this knowledge to provide better default value for
virttype.
Signed-off-by: NJiri Denemark <jdenemar@redhat.com>

Since some may depend on the accelerator used when probing QEMU the
cache becomes invalid when KVM becomes available or if it is not
available anymore.
Signed-off-by: NJiri Denemark <jdenemar@redhat.com>

CPU related capabilities may differ depending on accelerator used when
probing. Let's use KVM if available when probing QEMU and fall back to
TCG. The created capabilities already contain all we need to distinguish
whether KVM or TCG was used:

    - KVM was used when probing capabilities:
        QEMU_CAPS_KVM is set
        QEMU_CAPS_ENABLE_KVM is not set

    - TCG was used and QEMU supports KVM, but it failed (e.g., missing
      kernel module or wrong /dev/kvm permissions)
        QEMU_CAPS_KVM is not set
        QEMU_CAPS_ENABLE_KVM is set

    - KVM was not used and QEMU does not support it
        QEMU_CAPS_KVM is not set
        QEMU_CAPS_ENABLE_KVM is not set
Signed-off-by: NJiri Denemark <jdenemar@redhat.com>

Let's set QEMU_CAPS_KVM and QEMU_CAPS_ENABLE_KVM early so that the rest
of the probing code can use these capabilities to handle KVM/TCG replies
differently.
Signed-off-by: NJiri Denemark <jdenemar@redhat.com>

Using -machine instead of -M for QMP probing is safe because any QEMU
binary which is capable of QMP probing supports -machine.
Signed-off-by: NJiri Denemark <jdenemar@redhat.com>

The code that runs a new QEMU process to be used for probing
capabilities is separated into four reusable functions so that any code
that wants to probe a QEMU process may just follow a few simple steps:

    cmd = virQEMUCapsInitQMPCommandNew(...);
    virQEMUCapsInitQMPCommandRun(cmd);

    /* talk to the running QEMU process using its QMP monitor */

    if (reprobeIsRequired) {
        virQEMUCapsInitQMPCommandAbort(cmd, ...);
        virQEMUCapsInitQMPCommandRun(cmd);

        /* talk to the running QEMU process again */
    }

    virQEMUCapsInitQMPCommandFree(cmd);
Signed-off-by: NJiri Denemark <jdenemar@redhat.com>

We have couple of functions that operate over NULL terminated
lits of strings. However, our naming sucks:

virStringJoin
virStringFreeList
virStringFreeListCount
virStringArrayHasString
virStringGetFirstWithPrefix

We can do better:

virStringListJoin
virStringListFree
virStringListFreeCount
virStringListHasString
virStringListGetFirstWithPrefix
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>

If libvirt is compiled without NUMACTL support starting libvirtd
reports a libvirt internal error "NUMA isn't available on this host"
without checking if NUMA support is compiled into the libvirt binaries.
This patch adds the missing NUMA support check to prevent the internal error.
It also includes a check if the cgroup controller cpuset is available before
using it.

The error was noticed when libvirtd was restarted with running domains and
on libvirtd start the qemuConnectCgroup gets called during qemuProcessReconnect.
Signed-off-by: NBoris Fiuczynski <fiuczy@linux.vnet.ibm.com>
Reviewed-by: NBjoern Walk <bwalk@linux.vnet.ibm.com>

Adjust the device string that is built for vhost-scsi devices so that it
can be invoked from hotplug.

From the QEMU command line, the file descriptors are expect to be numeric only.
However, for hotplug, the file descriptors are expected to begin with at least
one alphabetic character else this error occurs:

  # virsh attach-device guest_0001 ~/vhost.xml
  error: Failed to attach device from /root/vhost.xml
  error: internal error: unable to execute QEMU command 'getfd':
  Parameter 'fdname' expects a name not starting with a digit

We also close the file descriptor in this case, so that shutting down the
guest cleans up the host cgroup entries and allows future guests to use
vhost-scsi devices.  (Otherwise the guest will silently end.)
Signed-off-by: NEric Farman <farman@linux.vnet.ibm.com>

Open /dev/vhost-scsi, and record the resulting file descriptor, so that
the guest has access to the host device outside of the libvirt daemon.
Pass this information, along with data parsed from the XML file, to build
a device string for the qemu command line.  That device string will be
for either a vhost-scsi-ccw device in the case of an s390 machine, or
vhost-scsi-pci for any others.
Signed-off-by: NEric Farman <farman@linux.vnet.ibm.com>

We already have a "scsi" hostdev subsys type, which refers to a single
LUN that is passed through to a guest.  But what of things where
multiple LUNs are passed through via a single SCSI HBA, such as with
the vhost-scsi target?  Create a new hostdev subsys type that will
carry this.
Signed-off-by: NEric Farman <farman@linux.vnet.ibm.com>

Do all the stuff for the vhost-scsi capability in QEMU,
so it's in place for our checks later.
Signed-off-by: NEric Farman <farman@linux.vnet.ibm.com>
Reviewed-by: NBoris Fiuczynski <fiuczy@linux.vnet.ibm.com>

Removed the comment 'Set the migration source' as it isn't valid anymore
and 'start it up' isn't useful as qemuProcessStart() is already a
speaking name.
Signed-off-by: NMarc Hartmayer <mhartmay@linux.vnet.ibm.com>

Just like in the previous commit, we are not updating CGroups on
chardev hot(un-)plug and thus leaving qemu unable to access any
non-default device users are trying to hotplug.
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>

If users try to hotplug RNG device with a backend different to
/dev/random or /dev/urandom the whole operation fails as qemu is
unable to access the device. The problem is we don't update
device CGroups during the operation.
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>

qemuDomainObjExitAgent is unsafe.

First it accesses domain object without domain lock.
Second it uses outdated logic that goes back to commit 79533da1 of
year 2009 when code was quite different. (unref function
instead of unreferencing only unlocked and disposed object
in case of last reference and leaved unlocking to the caller otherwise).
Nowadays this logic may lead to disposing locked object
i guess.

Another problem is that the callers of qemuDomainObjEnterAgent
use domain object again (namely priv->agent) without domain lock.

This patch address these two problems.

qemuDomainGetAgent is dropped as unused.

Sometimes after domain restart agent is unavailabe even
if it is up and running in guest. Diagnostic message is
"QEMU guest agent is not available due to an error"
that is 'priv->agentError' is set. Investiagion shows that
'priv->agent' is not NULL, so error flag is set probably
during domain shutdown process and not cleaned up eventually.

The patch is quite simple - just clean up error flag unconditionally
upon domain stop.

Other hunks address other cases when error flag is not cleaned up.

1. processSerialChangedEvent. We need to clean error flag
unconditionally here too. For example if upon first 'connected' event we
fail to connect and set error flag and then connect on second
'connected' event then error flag will remain set erroneously
and make agent unavailable.

2. qemuProcessHandleAgentEOF. If error flag is set and we get
EOF we need to change state (and diagnostic) from 'error' to
'not connected'.

qemuConnectAgent return -1 or -2 in case of different errors.
A. -1 is a case of unsuccessuful connection to guest agent.
B. -2 is a case of destoyed domain during connection attempt.

All qemuConnectAgent callers handle the first error the same way
so let's move this logic into qemuConnectAgent itself. Patched
function returns 0 in case A and -1 in case B.

Use the util function virHostdevIsSCSIDevice() to simplify if
statements.
Signed-off-by: NMarc Hartmayer <mhartmay@linux.vnet.ibm.com>
Reviewed-by: NBjoern Walk <bwalk@linux.vnet.ibm.com>
Reviewed-by: NBoris Fiuczynski <fiuczy@linux.vnet.ibm.com>

Add missing checks if a hostdev is a subsystem/SCSI device before access
the union member 'subsys'/'scsi'.  Also fix indentation and simplify
qemuDomainObjCheckHostdevTaint().
Signed-off-by: NMarc Hartmayer <mhartmay@linux.vnet.ibm.com>
Reviewed-by: NBjoern Walk <bwalk@linux.vnet.ibm.com>
Reviewed-by: NBoris Fiuczynski <fiuczy@linux.vnet.ibm.com>

New line character in name of domain is now forbidden because it
mess virsh output and can be confusing for users.
Validation of name is done in drivers, after parsing XML to avoid
problems with dissappeared domains which was already created with
new-line char in name.
Signed-off-by: NMichal Privoznik <mprivozn@redhat.com>

Commit 3f71c797 added 'qemu_id' field to track the id of the cpu
as reported by query-cpus. The patch did not include changes necessary
to propagate the id through the functions matching the data to the
libvirt cpu structures and thus all vcpus had id 0.

Don't use qemuMonitorGetCPUInfo which does a lot of matching to get the
full picture which is not necessary and would be mostly discarded.

Refresh only the vcpu halted state using data from query-cpus.

We don't need to call qemuMonitorGetCPUInfo which is very inefficient to
get data required to update the vcpu 'halted' state.

Add a monitor helper that will retrieve the halted state and return it
in a bitmap so that it can be indexed easily.