When we get an EOF event on monitor connection, it may be a result of
either crash or graceful shutdown. QEMU which supports async events
(i.e., we are talking to it using JSON monitor) emits SHUTDOWN event on
graceful shutdown. In case we don't get this event by the time monitor
connection is closed, we assume the associated domain crashed.
Commit febc591683 introduced -vga none in
case no video card is included in domain XML. However, old qemu
versions do not support this and such domain cannot be successfully
started.
Two more calls to remote libvirtd have to be surrounded by
qemuDomainObjEnterRemoteWithDriver() and
qemuDomainObjExitRemoteWithDriver() to prevent possible deadlock between
two communicating libvirt daemons.
See commit f0c8e1cb37 for further details.
virDrvSupportsFeature API is allowed to return -1 on error while all but
one uses of VIR_DRV_SUPPORTS_FEATURE only check for (non)zero return
value. Let's make this macro return zero on error, which is what
everyone expects anyway.
This patch allows for using custom scripts instead of /usr/bin/qemu
emulator in domain XML. To do so, one would specify relative path to the
custom script in <emulator/>. The path needs to be relative to
qemuxml2argvdata directory and it will be transparently made absolute in
runtime. The expected command line needs to contain the exact relative
path as was used in domain XML.
The problem is RelaxNG schema for domain XML only allows for absolute
path within <emulator/>. To workaround it, an extra '/' must be added at
the beginning of the path. That is, instead of "./qemu.sh" or
"../emulator/qemu.sh" one would use "/./qemu.sh" or
"/../emulator/qemu.sh". The extra slash is removed before further
processing. I don't like this workaround, it's very ugly but it's the
best option I was able to come up with. Relaxing domain XML schema is
not an option IMO.
The nodeinfo structure includes
nodes : the number of NUMA cell, 1 for uniform mem access
sockets : number of CPU socket per node
cores : number of core per socket
threads : number of threads per core
which does not work well for NUMA topologies where each node does not
consist of integral number of CPU sockets.
We also have VIR_NODEINFO_MAXCPUS macro in public libvirt.h which
computes maximum number of CPUs as (nodes * sockets * cores * threads).
As a result, we can't just change sockets to report total number of
sockets instead of sockets per node. This would probably be the easiest
since I doubt anyone is using the field directly. But because of the
macro, some apps might be using sockets indirectly.
This patch leaves sockets to be the number of CPU sockets per node (and
fixes qemu driver to comply with this) on machines where sockets can be
divided by nodes. If we can't divide sockets by nodes, we behave as if
there was just one NUMA node containing all sockets. Apps interested in
NUMA should consult capabilities XML, which is what they probably do
anyway.
This way, the only case in which apps that care about NUMA may break is
on machines with funky NUMA topology. And there is a chance libvirt
wasn't able to start any guests on those machines anyway (although it
depends on the topology, total number of CPUs and kernel version).
Nothing changes at all for apps that don't care about NUMA.
So far, readonly=on option is used when qemu supports -device. However,
there are qemu versions which support readonly option with -drive
although they don't have support for -device.
Explicitly raising a nice error in the case user tries to migrate a
guest with assigned host devices is much better than waiting for a
mysterious error with no clue for the reason.
When only some host CPUs given to cpuBaseline contain <vendor> element,
baseline CPU should not contain it. Otherwise the result would not be
compatible with the host CPUs without vendor. CPU vendors are still
taken into account when computing baseline CPU, it's just removed from
the result.
Recent CPU models were specified using invalid vendor element
<vendor>NAME</vendor>, which was silently ignored due to a bug in the
code which was parsing it.
Some features provided by the recently added CPU models were mentioned
twice for each model. This was a result of automatic generation of the
XML from qemu's CPU configuration file without noticing this redundancy.
According to API documentation virDomain{At,De}tachDevice calls are
supposed to only work on active guests for device hotplug. For anything
beyond that, their *Flags variants have to be used.
Despite the variant which was acked on libvirt mailing list
(https://www.redhat.com/archives/libvir-list/2010-January/msg00385.html)
commit ed9c14a7ef (by Jim Fehlig)
introduced automagic behavior of these API calls for xen driver. Since
January, these calls always change persistent configuration of a guest
and if the guest is currently active, they also hot(un)plug the device.
That change didn't follow API documentation and also broke device
hot(un)plug for older xend implementations which do not support changing
persistent configuration of a guest and hot(un)plugging in one step.
This patch should not break anything for active guests. On the other
hand, changing inactive guests is not supported any more.
When a user calls to virDomain{Attach,Detach,Update}DeviceFlags() with
flags == VIR_DOMAIN_DEVICE_MODIFY_LIVE on an inactive guest running on
an old Xen hypervisor (such as RHEL-5) xend_internal driver reports:
Xend version does not support modifying persistent config
which is pretty confusing since no-one requested to modify persistent
config.
The xm internal xen driver only supports disk and network devices to be
added to a guest. On an attempt to attach any other device the xm driver
used VIR_ERR_XML_ERROR which resulted in a completely bogus error
message:
error: Failed to attach device from pci.xml
error: XML description for unknown device is not well formed or invalid
libvirt-guests init script should be started as late as possible during
host startup and stopped as early as possible during host shutdown to
make sure required services are already/still up and running at the time
libvirt-guests runs.
cmdAttachInterface and cmdAttachDisk still used vshRealloc and sprintf
for generating XML, which is hardly maintainable. Let's get rid of this
old code.
Unless --driver tap|file option was given to attach-disk, virsh would
generate <disk type='block'> XML which might be fine for Xen but not for
other hypervisors. This patch introduces a new option --sourcetype which
can be used to explicitly set the type of disk source. The option
accepts either "file" or "block" types.
For static-only DHCP, i.e. with no <range> but at least one <host>
element within <dhcp> element, we have to add "--dhcp-range IP,static"
option to dnsmasq to actually enable the service. Without this option,
dnsmasq will not respond to DHCP requests.
Since libvirt-guests init script and its configuration do not require
libvirtd to be running/installed, it was a bad idea to put them into
daemon directory. libvirt.spec even includes these files in
libvirt-client subpackage, which may result in build failure for
client-only builds when the whole daemon directory is just skipped.
RHEL-5 Xen doesn't support the old style vnc configuration. In sexpr, we
can't really check it with rhel5-api turned on. However, for XM
configuration files it's sufficient to use cfg version 1 instead of 2.
The most common cause of errors with rhel5-api turn on was missing
"(type ioemu)" in sexpr or its equivalent in XM configuration file. This
happens because the presence of that part in sexpr (or cfg) depends on
xen version the host is running. Let's avoid it by explicitly specifying
interface model which ensures "type ioemu" will always be emitted.
This patch adds
<model type='e1000'/>
withing the interface element in all affected xml files. And
(model 'e1000')
to all corresponding sexpr files with similar fix to cfg files. Such
configuration works regardless on Xen version.
We already filled the PCI address structure when we checked whether it's
free or not, so let's just use the structure here instead of filling it
again.
When reconnecting to existing VMs, we re-reserved only those PCI
addresses which were explicitly mentioned in domain XML. Since some
addresses are always reserved (e.g., 0:0:0 and 0:0:1), we need to handle
those too.
Also all this should only be done if device flag is supported by qemu.
--dhcp-no-override description from dnsmasq man page:
Disable re-use of the DHCP servername and filename fields as
extra option space. If it can, dnsmasq moves the boot server and
filename information (from dhcp-boot) out of their dedicated
fields into DHCP options. This make extra space available in the
DHCP packet for options but can, rarely, confuse old or broken
clients. This flag forces "simple and safe" behaviour to avoid
problems in such a case.
It seems some virtual network card ROMs are this old/buggy so let's add
--dhcp-no-override as a workaround for them. We don't use extra DHCP
options so this should be safe. The option was added in dnsmasq-2.41,
which becomes the minimum required version.
We add --dhcp-lease-max=xxx argument when network->def->nranges > 0 but
we only allocate space for in the opposite case :-) I guess we are lucky
enough to miscount somewhere else so that we actually allocate more
space than we need since no-one has hit this bug so far.
device_del command is not synchronous for PCI devices, it merely asks
the guest to release the device and returns. If the host wants to use
that device before the guest actually releases it, we are in big
trouble. To avoid this, we already added a loop which waits up to 10
seconds until the device is actually released before we do anything else
with that device. But we only added this loop for managed PCI devices
before we try reattach them back to the host.
However, we need to wait even for non-managed devices. We don't reattach
them automatically, but we still want to prevent the host from using it.
This was revealed thanks to sVirt: when we relabel sysfs files
corresponding to the PCI device before the guest finished releasing the
device, qemu is no longer allowed to access those files and if it wants
(as a result of guest's request) to write anything to them, it just
exits, which kills the guest.
This is not a proper fix and needs some further work both on libvirt and
qemu side in the future.
Patch version revamped by Eric Blake <eblake@redhat.com> of Jiri
Denemark <jdenemar@redhat.com> original patch
When attaching a PCI device which doesn't explicitly set its PCI
address, libvirt allocates the address automatically. The problem is
that when checking which PCI address is unused, we only check for those
with slot number higher than the highest slot number ever used.
Thus attaching/detaching such device several times in a row (31 is the
theoretical limit, less then 30 tries are enough in practise) makes any
further device attachment fail. Furthermore, attaching a device with
predefined PCI address to 0:0:31 immediately forbids attachment of any
PCI device without explicit address.
This patch changes the logic so that we always check all PCI addresses
before we say there is no PCI address available.
Modifications from v1: revert back to remembering the last slot
reserved, but allow wraparound to not be limited by the end.
In this way, slots are still assigned in the same order as
before the patch, rather than filling in the gaps closest to
0 and risking making windows guests mad.
* src/qemu/qemu_conf.c: fix pci reservation code to do a round-robbin
check of all available PCI splot availability before failing.
When only client parts of libvirt are installed (i.e., no libvirtd
daemon), libvirt-guests init script in its default configuration would
throw seriously looking errors during host shutdown:
Running guests on default URI: error: unable to connect to
'/var/run/libvirt/libvirt-sock', libvirtd may need to be started: No
such file or directory
error: failed to connect to the hypervisor
This patch changes the script to print rather harmless message in that
situation:
Running guests on default URI: libvirtd not installed; skipping this
URI.
When comparing a CPU without <model> element, such as
<cpu>
<topology sockets='1' cores='1' threads='1'/>
</cpu>
libvirt would happily crash without warning.
When autodetecting whether XML describes guest or host CPU, the presence
of <arch> element is checked. If it's present, we treat the XML as host
CPU definition. Which is right, since guest CPU definitions do not
contain <arch> element. However, if at the same time the root <cpu>
element contains `match' attribute, we would silently ignore it and
still treat the XML as host CPU. We should rather refuse such invalid
XML.
When a CPU to be compared with host CPU describes a host CPU instead of
a guest CPU, the result is incorrect. This is because instead of
treating additional features in host CPU description as required, they
were treated as if they were mentioned with all possible policies at the
same time.
In case qemu supports -nodefconfig, libvirt adds uses it when launching
new guests. Since this option may affect CPU models supported by qemu,
we need to use it when probing for available models.
By specifying <vendor> element in CPU requirements a guest can be
restricted to run only on CPUs by a given vendor. Host CPU vendor is
also specified in capabilities XML.
The vendor is checked when migrating a guest but it's not forced, i.e.,
guests configured without <vendor> element can be freely migrated.
All features in the baseline CPU definition were always created with
policy='require' even though an arch driver returned them with different
policy settings.
Running virsh while having /var/lib/libvirt/libvirt-guests file open
makes SELinux emit messages about preventing virsh from reading that
file. Since virsh doesn't really want to read anything, it's better to
run it with /dev/null on stdin to prevent those messages.
- Fix documentation for virGetStorageVol: it has 'key' argument instead
of 'uuid'.
- Remove TODO comment from virReleaseStorageVol: we use volume key as an
identifier instead of UUID.
- Print human-readable UUID string in debug message in virReleaseSecret.
Per-connection hashes for domains, networks, storage pools and network
filter pools were indexed by names which was not the best choice. UUIDs
are better identifiers, so lets use them.
Firstly, the init script has to touch its file under /var/lock/subsys
when started, otherwise the system would think it's not running and
won't stop it during shutdown.
Secondly, for some reason there is a policy to automatically enable
init scripts when they are installed, so let the specfile do this. We
also need to start the init script to ensure it will be stopped during
the first shutdown after installing the package.
Also $LISTFILE should be enclosed by quotes everywhere as suggested by
Eric.
Example output during shutdown:
Running guests on default URI: console, rhel6-1, rhel5-64
Running guests on lxc:/// URI: lxc-shell
Running guests on xen:/// URI: error: no hypervisor driver available for xen:///
error: failed to connect to the hypervisor
Running guests on vbox+tcp://orkuz/system URI: no running guests.
Suspending guests on default URI...
Suspending console: done
Suspending rhel6-1: done
Suspending rhel5-64: done
Suspending guests on lxc:/// URI...
Suspending lxc-shell: error: Failed to save domain 9cba8bfb-56f4-6589-2d12-8a58c886dd3b state
error: this function is not supported by the hypervisor: virDomainManagedSave
Note, the "Suspending $guest: " shows progress during the suspend phase
if domjobinfo gives meaningful output.
Example output during boot:
Resuming guests on default URI...
Resuming guest rhel6-1: done
Resuming guest rhel5-64: done
Resuming guest console: done
Resuming guests on lxc:/// URI...
Resuming guest lxc-shell: already active
Configuration used for generating the examples above:
URIS='default lxc:/// xen:/// vbox+tcp://orkuz/system'
The script uses /var/lib/libvirt/libvirt-guests files to note all active
guest it should try to resume on next boot. It's content looks like:
default 7f8b9d93-30e1-f0b9-47a7-cb408482654b 085b4c95-5da2-e8e1-712f-6ea6a4156af2 fb4d8360-5305-df3a-2da1-07d682891b8c
lxc:/// 9cba8bfb-56f4-6589-2d12-8a58c886dd3b
We need a common internal function for starting managed domains to be
used during autostart. This patch factors out relevant code from
qemudDomainStart into qemudDomainObjStart and makes it use the
refactored code for domain restore instead of calling qemudDomainRestore
API directly.
We need to be able to assign new def to an existing virDomainObj which
is already locked. This patch factors out the relevant code from
virDomainAssignDef into virDomainObjAssignDef.
We need to be able to restore a domain which we already locked and
started a job for it without undoing these steps. This patch factors
out internals of qemudDomainRestore into separate functions which work
for locked objects.
When closing a monitor using qemuMonitorClose(), we are aware of
the possibility the monitor is still being used somewhere:
/* NB: ordinarily one might immediately set mon->watch to -1
* and mon->fd to -1, but there may be a callback active
* that is still relying on these fields being valid. So
* we merely close them, but not clear their values and
* use this explicit 'closed' flag to track this state */
but since we call virEventAddHandle() on that monitor without increasing
its ref counter, the monitor is still freed which makes possible users
of it quite unhappy. The unhappiness can lead to a hang if qemuMonitorIO
tries to lock mutex which no longer exists.
First calling REMOTE_PROC_CLOSE and then removing watches might lead to
a hang as HANGUP event can be triggered before the watches are actually
removed but after virConnectPtr is already freed. As a result of that
remoteDomainEventFired() would try to lock uninitialized mutex, which
would hang for ever.
With JSON qemu monitor, we get a STOP event from qemu whenever qemu
stops guests CPUs. The downside of it is that vm->state is changed to
PAUSED and a new generic paused event is send to applications. However,
when we ask qemu to stop the CPUs we are not really interested in qemu
event and we usually want to issue a more specific event.
By setting vm->status to PAUSED before actually sending the request to
qemu (and resetting it back if the request fails) we can ignore the
event since the event handler does nothing when the guest is already
paused. This solution is quite hacky but unfortunately it's the best
solution which I was able to come up with and it doesn't introduce a
race condition.
The hang fix in d376b7d63e was incomplete
since it left quite a few {Enter,Exit}Monitor calls which require driver
to be unlocked. Since the driver is locked throughout the whole
function, {Enter,Exit}MonitorWithDriver need to be used instead to
ensure driver is not locked when issuing monitor commands.
The comment in qemuDomainWaitForMigrationComplete says we are polling
every 50ms but the code sleeps only for 50us. This was already discussed
during review but apparently forgotten when the series was pushed.
Adds ability to provide a preferred CPU model for CPUID data decoding.
Such model would be considered as the best possible model (if it's
supported by hypervisor) regardless on number of features which have to
be added or removed for describing required CPU.
Jiri Denemark