This patch sets or unsets the IFF_VNET_HDR flag depending on what device
is used in the VM. The manipulation of the flag is done in the open
function and is only fatal if the IFF_VNET_HDR flag could not be cleared
although it has to be (or if an ioctl generally fails). In that case the
macvtap tap is closed again and the macvtap interface torn.
* src/qemu/qemu_conf.c src/qemu/qemu_conf.h: pass qemuCmdFlags to
qemudPhysIfaceConnect()
* src/util/macvtap.c src/util/macvtap.h: add vnet_hdr boolean to
openMacvtapTap(), and private function configMacvtapTap()
* src/qemu/qemu_driver.c: add extra qemuCmdFlags when calling
qemudPhysIfaceConnect()
Rework and simplification of teardown of the macvtap device.
Basically all devices with the same MAC address and link device are kept
alive and not attempted to be torn down. If a macvtap device linked to a
physical interface with a certain MAC address 'M' is to be created it
will automatically fail if the interface is 'up'ed and another macvtap
with the same properties (MAC addr 'M', link dev) happens to be 'up'.
This will prevent the VM from starting or the device from being attached
to a running VM. Stale interfaces are assumed to be there for some
reason and not stem from libvirt.
In the VM shutdown path, it's assuming that an interface name is always
available so that if the device type is DIRECT it can be torn down
using its name.
* src/util/macvtap.h src/libvirt_macvtap.syms: change of deleting routine
* src/util/macvtap.c: cleanups and change of deleting routine
* src/qemu/qemu_driver.c: change cleanup on shutdown
* src/qemu/qemu_conf.c: don't delete Macvtap in qemudPhysIfaceConnect()
Similar to the Set*Mem commands, this implementation was bogus and
misleading. Make it clear this is a hotplug only operation, and that the
hotplug piece isn't even implemented.
Also drop the overkill maxvcpus validation: we don't perform this check
at XML define time so clearly no one is missing it, and there is
always the risk that our info will be out of date, possibly preventing
legitimate CPU values.
Signed-off-by: Cole Robinson <crobinso@redhat.com>
SetMem and SetMaxMem are hotplug only APIs, any persistent config
changes are supposed to go via XML definition. The original implementation
of these calls were incorrect and had the nasty side effect of making
a psuedo persistent change that would be lost after libvirtd restart
(I didn't know any better).
Fix these APIs to rightly reject non running domains.
Signed-off-by: Cole Robinson <crobinso@redhat.com>
When in JSON mode, QEMU requires that 'qmp_capabilities' is run as
the first command in the monitor. This is a no-op when run in the
text mode monitor
* src/qemu/qemu_driver.c: Run capabilities negotiation when
connecting to the monitor
* src/qemu/qemu_monitor.c, src/qemu/qemu_monitor.h,
src/qemu/qemu_monitor_json.c, src/qemu/qemu_monitor_json.h: Add
support for the 'qmp_capabilities' command, no-op in text mode.
This part adds support for qemu making a macvtap tap device available
via file descriptor passed to qemu command line. This also attempts to
tear down the macvtap device when a VM terminates. This includes support
for attachment and detachment to/from running VM.
* src/qemu/qemu_conf.[ch] src/qemu/qemu_driver.c: add support in the
QEmu driver
Current PCI addresses are allocated at time of VM startup.
To make them truely persistent, it is neccessary to do this
at time of virDomainDefine/virDomainCreate. The code in
qemuStartVMDaemon still remains in order to cope with upgrades
from older libvirt releases
* src/qemu/qemu_driver.c: Rename existing qemuAssignPCIAddresses
to qemuDetectPCIAddresses. Add new qemuAssignPCIAddresses which
does auto-allocation upfront. Call qemuAssignPCIAddresses from
qemuDomainDefine and qemuDomainCreate to assign PCI addresses that
can then be persisted. Don't clear PCI addresses at shutdown if
they are intended to be persistent
The old text mode monitor prompts for a password when disks are
encrypted. This interactive approach doesn't work for JSON mode
monitor. Thus there is a new 'block_passwd' command that can be
used.
* src/qemu/qemu_driver.c: Split out code for looking up a disk
secret from findVolumeQcowPassphrase, into a new method
getVolumeQcowPassphrase. Enhance qemuInitPasswords() to also
set the disk encryption password via the monitor
* src/qemu/qemu_monitor.c, src/qemu/qemu_monitor.h,
src/qemu/qemu_monitor_json.c, src/qemu/qemu_monitor_json.h,
src/qemu/qemu_monitor_text.c, src/qemu/qemu_monitor_text.h: Add
support for the 'block_passwd' monitor command.
Since c26cb9234f4b9fa46d7caa3385ae36704167c53f, the dname
parameter has been ignored by these two functions. Use it.
* src/qemu/qemu_driver.c (qemudDomainMigratePrepareTunnel): Honor dname
parameter once again.
(qemudDomainMigratePrepare2): Likewise.
Currently the timeout for reading startup output is 3 seconds. If the
host is under any sort of load, we can easily trigger this. Lets bump
it to 30 seconds.
Since the polling loop checks to see if the process has died, we shouldn't
erroneously hit this timeout if qemu bombs (only if it is stuck in some
infinite loop).
The virConnectPtr is no longer required for error reporting since
that is recorded in a thread local. Remove use of virConnectPtr
from all APIs in cpu_conf.{h,c} and update all callers to
match
The virConnectPtr is no longer required for error reporting since
that is recorded in a thread local. Remove use of virConnectPtr
from all APIs in node_device_conf.{h,c} and update all callers to
match
The QEMU flags are commonly stored as a signed or unsigned int,
allowing only 31 flags. This limit is rather close, so to aid
future patches, change it to a 64-bit int
* src/qemu/qemu_conf.c, src/qemu/qemu_conf.h, src/qemu/qemu_driver.c,
tests/qemuargv2xmltest.c, tests/qemuhelptest.c, tests/qemuxml2argvtest.c:
Use 'unsigned long long' for QEMU flags
The virConnectPtr is no longer required for error reporting since
that is recorded in a thread local. Remove use of virConnectPtr
from all APIs in security_driver.{h,c} and update all callers to
match
The security driver was mistakenly initialized before the QEMU
config file was loaded. This prevents it being turned off again.
The capabilities XML was also getting the wrong security driver
name, due to the stacked driver arrangement.
* src/qemu/qemu_driver.c: Fix initialization order and capabilities
model name
If the primary security driver (SELinux/AppArmour) was disabled
then the secondary QEMU DAC security driver was also disabled.
This is mistaken, because the latter must be active at all times
* src/qemu/qemu_driver.c: Ensure DAC driver is always active
To allow devices to be hot(un-)plugged it is neccessary to ensure
they all have a unique device aliases. This fixes the hotplug
methods to assign device aliases before invoking the monitor
commands which need them
* src/qemu/qemu_conf.c, src/qemu/qemu_conf.h: Expose methods
for assigning device aliases for disks, host devices and
controllers
* src/qemu/qemu_driver.c: Assign device aliases when hotplugging
all types of device
* tests/qemuxml2argvdata/qemuxml2argv-hostdev-pci-address-device.args,
tests/qemuxml2argvdata/qemuxml2argv-hostdev-usb-address-device.args:
Update for changed hostdev naming scheme
This patch re-arranges the QEMU device alias assignment code to
make it easier to call into the same codeblock when performing
device hotplug. The new code has the ability to skip over already
assigned names to facilitate hotplug
* src/qemu/qemu_driver.c: Call qemuAssignDeviceNetAlias()
instead of qemuAssignNetNames
* src/qemu/qemu_conf.h: Export qemuAssignDeviceNetAlias()
instead of qemuAssignNetNames
* src/qemu/qemu_driver.c: Merge the legacy disk/network alias
assignment code into the main methods
The current way of assigning names to the host network backend and
NIC device in QEMU was over complicated, by varying naming scheme
based on the NIC model and backend type. This simplifies the naming
to simply be 'net0' and 'hostnet0', allowing code to easily determine
the host network name and vlan based off the primary device alias
name 'net0'. This in turn allows removal of alot of QEMU specific
code from the XML parser, and makes it easier to assign new unique
names for NICs that are hotplugged
* src/conf/domain_conf.c, src/conf/domain_conf.h: Remove hostnet_name
and vlan fields from virNetworkDefPtr
* src/qemu/qemu_conf.c, src/qemu/qemu_conf.h, src/qemu/qemu_driver.c:
Use a single network alias naming scheme regardless of NIC type
or backend type. Determine VLANs from the alias name.
* tests/qemuxml2argvdata/qemuxml2argv-net-eth-names.args,
tests/qemuxml2argvdata/qemuxml2argv-net-virtio-device.args,
tests/qemuxml2argvdata/qemuxml2argv-net-virtio-netdev.args: Update
for new simpler naming scheme
PCI disk, disk controllers, net devices and host devices need to
have PCI addresses assigned before they are hot-plugged
* src/qemu/qemu_conf.c: Add APIs for ensuring a device has an
address and releasing unused addresses
* src/qemu/qemu_driver.c: Ensure all devices have addresses
when hotplugging.
The current QEMU code allocates PCI addresses incrementally starting
at 4. This is not satisfactory because the user may have given some
addresses in their XML config, which need to be skipped over when
allocating addresses to remaining devices.
It is thus neccessary to maintain a list of already allocated PCI
addresses and then only allocate ones that remain unused. This is
also required for domain device hotplug to work properly later.
* src/qemu/qemu_conf.c, src/qemu/qemu_conf.h: Add APIs for creating
list of existing PCI addresses, and allocating new addresses.
Refactor address assignment to use this code
* src/qemu/qemu_driver.c: Pull PCI address assignment up into the
qemuStartVMDaemon() method, as a prelude to moving it into the
'define' method. Update list of allocated addresses when connecting
to a running VM at daemon startup.
* tests/qemuxml2argvtest.c, tests/qemuargv2xmltest.c,
tests/qemuxml2xmltest.c: Remove USB product test since all
passthrough is done based on address
* tests/qemuxml2argvdata/qemuxml2argv-hostdev-usb-product.args,
tests/qemuxml2argvdata/qemuxml2argv-hostdev-usb-product.xml: Kil
unused data files
Since QEMU startup uses the new -device argument, the hotplug
code needs todo the same. This converts disk, network and
host device hotplug to use the device_add command
* src/qemu/qemu_driver.c: Use new device_add monitor APIs
whereever possible
All the helper functions for building command line arguments
now return a 'char *', instead of acepting a 'char **' or
virBufferPtr argument
* qemu/qemu_conf.c: Standardize syntax for building args
* qemu/qemu_conf.h: Export all functions for building args
* qemu/qemu_driver.c: Update for changed syntax for building
NIC/hostnet args
Similar to the race fixed by
be34c3c7efbb1ea8999530f98b99c5dde3793f84, make sure
to wait around for KVM to release the resources from
a hot-detached PCI device before attempting to
rebind that device to the host driver.
Signed-off-by: Chris Lalancette <clalance@redhat.com>
If you shutdown libvirtd while a domain with PCI
devices is running, then try to restart libvirtd,
libvirtd will crash.
This happens because qemuUpdateActivePciHostdevs() is calling
pciDeviceListSteal() with a dev of 0x0 (NULL), and then trying
to dereference it. This patch fixes it up so that
qemuUpdateActivePciHostdevs() steals the devices after first
Get()'ting them, avoiding the crash.
Signed-off-by: Chris Lalancette <clalance@redhat.com>
Certain hypervisors (like qemu/kvm) map the PCI bar(s) on
the host when doing device passthrough. This can lead to a race
condition where the hypervisor is still cleaning up the device while
libvirt is trying to re-attach it to the host device driver. To avoid
this situation, we look through /proc/iomem, and if the hypervisor is
still holding onto the bar (denoted by the string in the matcher variable),
then we can wait around a bit for that to clear up.
v2: Thanks to review by DV, make sure we wait the full timeout per-device
Signed-off-by: Chris Lalancette <clalance@redhat.com>
The loop looking for the controller associated with a SCI drive had
an off by one, causing it to miss the last controller.
* src/qemu/qemu_driver.c: Fix off-by-1 in searching for SCSI
drive hotplug
The hotplug code in QEMU was leaking memory because although the
inner device object was being moved into the main virDomainDefPtr
config object, the outer container virDomainDeviceDefPtr was not.
* src/qemu/qemu_driver.c: Clarify code to show that the inner
device object is owned by the main domain config upon
successfull attach.
The hotplug code was not correctly invoking the security driver
in error paths. If a hotplug attempt failed, the device would
be left with VM permissions applied, rather than restored to the
original permissions. Also, a CDROM media that is ejected was
not restored to original permissions. Finally there was a bogus
call to set hostdev permissions in the hostdev unplug code
* qemu/qemu_driver.c: Fix security driver usage in hotplug/unplug
If there is a problem with VM startup, PCI devices may be left
assigned to pci-stub / pci-back. Adding a call to reattach
host devices in the cleanup path is required.
* qemu/qemu_driver.c: qemuDomainReAttachHostDevices() when
VM startup fails
Remove all the QEMU driver calls for setting file ownership and
process uid/gid. Instead wire in the QEMU DAC security driver,
stacking it ontop of the primary SELinux/AppArmour driver.
* qemu/qemu_driver.c: Switch over to new DAC security driver
Pulling the disk labelling code out of the exec hook, and into
libvirtd will allow it to access shared state in the daemon. It
will also make debugging & error reporting easier / more reliable.
* qemu/qemu_driver.c: Move initial disk labelling calls up into
libvirtd. Add cleanup of disk labels upon failure
If a VM fails to start, we can't simply free the security label
strings, we must call the domainReleaseSecurityLabel() method
otherwise the reserved 'mcs' level will be leaked in SElinux
* src/qemu/qemu_driver.c: Invoke domainReleaseSecurityLabel()
when domain fails to start
The current security driver architecture has the following
split of logic
* domainGenSecurityLabel
Allocate the unique label for the domain about to be started
* domainGetSecurityLabel
Retrieve the current live security label for a process
* domainSetSecurityLabel
Apply the previously allocated label to the current process
Setup all disk image / device labelling
* domainRestoreSecurityLabel
Restore the original disk image / device labelling.
Release the unique label for the domain
The 'domainSetSecurityLabel' method is special because it runs
in the context of the child process between the fork + exec.
This is require in order to set the process label. It is not
required in order to label disks/devices though. Having the
disk labelling code run in the child process limits what it
can do.
In particularly libvirtd would like to remember the current
disk image label, and only change shared image labels for the
first VM to start. This requires use & update of global state
in the libvirtd daemon, and thus cannot run in the child
process context.
The solution is to split domainSetSecurityLabel into two parts,
one applies process label, and the other handles disk image
labelling. At the same time domainRestoreSecurityLabel is
similarly split, just so that it matches the style. Thus the
previous 4 methods are replaced by the following 6 new methods
* domainGenSecurityLabel
Allocate the unique label for the domain about to be started
No actual change here.
* domainReleaseSecurityLabel
Release the unique label for the domain
* domainGetSecurityProcessLabel
Retrieve the current live security label for a process
Merely renamed for clarity.
* domainSetSecurityProcessLabel
Apply the previously allocated label to the current process
* domainRestoreSecurityAllLabel
Restore the original disk image / device labelling.
* domainSetSecurityAllLabel
Setup all disk image / device labelling
The SELinux and AppArmour drivers are then updated to comply with
this new spec. Notice that the AppArmour driver was actually a
little different. It was creating its profile for the disk image
and device labels in the 'domainGenSecurityLabel' method, where as
the SELinux driver did it in 'domainSetSecurityLabel'. With the
new method split, we can have consistency, with both drivers doing
that in the domainSetSecurityAllLabel method.
NB, the AppArmour changes here haven't been compiled so may not
build.
The QEMU driver is doing 90% of the calls to check for static vs
dynamic labelling. Except it is forgetting todo so in many places,
in particular hotplug is mistakenly assigning disk labels. Move
all this logic into the security drivers themselves, so the HV
drivers don't have to think about it.
* src/security/security_driver.h: Add virDomainObjPtr parameter
to virSecurityDomainRestoreHostdevLabel and to
virSecurityDomainRestoreSavedStateLabel
* src/security/security_selinux.c, src/security/security_apparmor.c:
Add explicit checks for VIR_DOMAIN_SECLABEL_STATIC and skip all
chcon() code in those cases
* src/qemu/qemu_driver.c: Remove all checks for VIR_DOMAIN_SECLABEL_STATIC
or VIR_DOMAIN_SECLABEL_DYNAMIC. Add missing checks for possibly NULL
driver entry points.
