The memory device alias needs to be treated as machine ABI as qemu is
using it in the migration stream for section labels. To simplify this
generate the alias from the slot number unless an existing broken
configuration is detected.
With this patch the aliases are predictable and even certain
configurations which would not be migratable previously are fixed.
Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=1359135
As with other devices assign the slot number right away when adding the
device. This will make the slot numbers static as we do with other
addressing elements and it will ultimately simplify allocation of the
alias in a static way which does not break with qemu.
Detect on reconnect to a running qemu VM whether the alias of a
hotpluggable memory device (dimm) does not match the dimm slot number
where it's connected to. This is necessary as qemu is actually
considering the alias as machine ABI used to connect the backend object
to the dimm device.
This will require us to keep them consistent so that we can reliably
restore them on migration. In some situations it was currently possible
to create a mismatched configuration and qemu would refuse to restore
the migration stream.
To avoid breaking existing VMs we'll need to keep the old algorithm
though.
Simplify handling of the 'dimm' address element by allowing to specify
the slot number only. This will allow libvirt to allocate slot numbers
before starting qemu.
Allow detecting capabilities according to the qemu QMP schema. This is
necessary as sometimes the availability of certain options depends on
the presence of a field in the schema.
This patch adds support for loading the QMP schema when detecting qemu
capabilities and adds a very simple query language to allow traversing
the schema and selecting a certain element from it.
The infrastructure in this patch uses a query path to set a specific
capability flag according to the availability of the given element in
the schema.
Remove the command from the supported commands list temporarily so that
QMP introspection code can be added without breaking tests and having to
tweak the test data in the same commit.
This will be later reverted and test data will be added. The aarch64
code is special as it calls additional commands and thus the test data
can't be added upfront.
qemuBuildSmbiosBiosStr and qemuBuildSmbiosSystemStr return NULL if
there's nothing to format on the commandline. Reporting errors from
buffer creation doesn't make sense since it would be ignored.
If the VM is offline virsh attempted to at least report the pinning
information for the VM. This would not work properly now that the vcpus
can be sparse. Fix it by getting the vcpu states from the XML.
Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=1375920
The fallback code used if virDomainGetVcpusFlags is not supported used
wrong XPath queries and basically did not work at all. Fix them to point
to the <domain> <vcpu> element instead of <vcpus> which was not present
until lately.
Make sure that the topology results into a sane number of cpus (up to
UINT_MAX) so that it can be sanely compared to the vcpu count of the VM.
Additionally the helper added in this patch allows to fetch the total
number the topology results to so that it does not have to be
reimplemented later.
Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=1378290
Since introduction of chardev hotplug the code was wrong for the UDP
case and basically created a TCP socket instead. Use proper objects and
type for UDP.
Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=1377602
Until now the test was rather useless since it didn't check the
arguments formatted and didn't use properly configured chardev objects.
Add the expected arguments and instrument the test to validate them.
Modify some test cases to actually add valid data.
Note that the UDP test data is currently wrong due to a bug.
The chardev attach test would do all the tests in one virTestRun
instance. If one sub-test failed then the test would report failure
improperly and the error would be hard to debug since the error pointer
was overwritten.
For compatibility reasons virDomainSetVcpus needs to add vcpus as non
hotpluggable which means that the users will not be able to unplug it
after the VM has started. Add a flag that will allow to tell the API
that the unpluggable vcpus are okay.
If attaching to a qemu process fails after opening the monitor socket
libvirt does not clean up the monitor. As the monitor also holds a
reference to the domain object the qemu attach API basically leaks it.
QEMU also does not interact on a second monitor connection and thus a
further attempt to attach to it would lock up.
Prevent libvirt from leaking the monitor by explicitly closing it.
Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=1378401
Attaching to a existing qemu process allows to get us into a situation
when qemu is new enough to have JSON monitor and new vCPU hotplug but
the json monitor is not used. The vCPU detection code would require it
though. This broke attaching to qemu processes.
Make the condition less strict and just skip the vCPU hotplug detection
if JSON monitor is not available.
Resolves one of the symptoms in:
https://bugzilla.redhat.com/show_bug.cgi?id=1378401
Certain operations may make the vcpu order information invalid. Since
the order is primarily used to ensure migration compatibility and has
basically no other user benefits, clear the order prior to certain
operations and document that it may be cleared.
All the operations that would clear the order can still be properly
executed by defining a new domain configuration rather than using the
helper APIs.
Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=1370357
virDomainDefSetVcpus was not designed to handle coldplug of vcpus now
that we can set state of vcpus individually.
Introduce qemuDomainSetVcpusConfig that properly handles state changes
of vcpus when coldplugging so that invalid configurations are not
created.
Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=1375939
The current code that validates duplicate vcpu order would not work
properly if the order would exceed def->maxvcpus. Limit the order to the
interval described.
The bitmap indexes for the order duplicate check are shifted to 0 since
vcpu order 0 is not allowed. The error message doesn't need such
treating though.
Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=1370360
Calling virDomainGetEmulatorPinInfo on a live VM with automatic NUMA
pinning and VIR_DOMAIN_AFFECT_CONFIG would return the automatic pinning
data in some cases which is bogus. Use the autoCpuset property only when
called on a live definition.
Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=1365779
Calling virDomainGetVcpuPinInfo on a live VM with automatic NUMA pinning
and VIR_DOMAIN_AFFECT_CONFIG would return the automatic pinning data
in some cases which is bogus. Use the autoCpuset property only when
called on a live definition.
Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=1365779
Sometimes adding a separate variable to access vm->privateData is not
necessary. Add a macro that will do the typecasting rather than having
to add a temp variable to force the compiler to typecast it.
Use the state information (online, hotpluggable) provided by the monitor
code rather than trying to infer it. This fixes an issue where on
architectures that require hotplug of multiple threads at once the
sub-cores would get updated as offline on daemon restart thus creating
an invalid configuration.
Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=1375783
Return whether a vcpu entry is hotpluggable or online so that upper
layers don't have to infer the information from other data.
Advantage is that this code can be tested by unit tests.
The algorithm that matches data from query-cpus and
query-hotpluggable-cpus is quite complex. Start using descriptive
iterator names to avoid confusion.
Name it virNumaGetHostMemoryNodeset and return only NUMA nodes which
have memory installed. This is necessary as the kernel is not very happy
to set the memory cgroup setting for nodes which do not have any memory.
This would break vcpu hotplug with following message on such
configruation:
Invalid value '0,8' for 'cpuset.mems': Invalid argument
Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=1375268
Qemu always opens the tray if forced to. Skip the waiting step in such
case.
This also helps if qemu does not report the tray change event when
opening the cdrom forcibly (the documentation says that the event will
not be sent although qemu in fact does trigger it even if @force is
selceted).
This is a workaround for a qemu issue where qemu does not send the tray
change event in some cases (after migration with empty closed locked
drive) and thus renders the cdrom useless from libvirt's point of view.
Partially resolves: https://bugzilla.redhat.com/show_bug.cgi?id=1368368
When a source image is dropped when missing due to startup policy the
policy needs to be cleared since it was relevant only for the given
storage source. New sources need to update it if needed.
Similarly to vcpu hotplug the emulator thread cgroup numa mapping needs
to be relaxed while hot-adding vcpus so that the threads can allocate
data in the DMA zone.
Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=1370084
When hot-adding vcpus qemu needs to allocate some structures in the DMA
zone which may be outside of the numa pinning. Extract the code doing
this in a set of helpers so that it can be reused.
Setting vcpu count when cpu topology is specified may result into an
invalid configuration. Since the topology can't be modified, reject the
setting if it doesn't match the requested topology. This will allow
fixing the topology in case it was broken.
Partially fixes: https://bugzilla.redhat.com/show_bug.cgi?id=1370066
Validating the vcpu count is more intricate and doing it in the XML
parser will make previously valid configs (with older qemus) vanish.
Now that we have a very similar check in the qemu domain validation
callback we can do it in a more appropriate place.
This basically reverts commit b54de0830a.
Partially resolves: https://bugzilla.redhat.com/show_bug.cgi?id=1370066
ce43cca0e refactored the helper to prepare it for sparse topologies but
forgot to fix the iterator used to fill the structures. This would
result into a weirdly sparse populated array and possible out of bounds
access and crash once sparse vcpu topologies were allowed.
Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=1369988
virVcpuInfo contains the vcpu number that the data refers to. Report
what's returned by the daemon rather than the sequence number as with
sparse vcpu topologies they won't match.
This patch removes the old vcpu unplug code completely and replaces it
with the new code using device_del. The old hotplug code basically never
worked with any recent qemu and thus is useless.
As the new code is using device_del all the implications of using it
are present. Contrary to the device deletion code, the vcpu deletion
code fails if the unplug request is not executed in time.
To allow unplugging the vcpus, hotplugging of vcpus on platforms which
require to plug multiple logical vcpus at once or plugging them in an
arbitrary order it's necessary to use the new device_add interface for
vcpu hotplug.
This patch adds support for the device_add interface using the old
setvcpus API by implementing an algorithm to select the appropriate
entities to plug in.
Add support for using the new approach to hotplug vcpus using device_add
during startup of qemu to allow sparse vcpu topologies.
There are a few limitations imposed by qemu on the supported
configuration:
- vcpu0 needs to be always present and not hotpluggable
- non-hotpluggable cpus need to be ordered at the beginning
- order of the vcpus needs to be unique for every single hotpluggable
entity
Qemu also doesn't really allow to query the information necessary to
start a VM with the vcpus directly on the commandline. Fortunately they
can be hotplugged during startup.
The new hotplug code uses the following approach:
- non-hotpluggable vcpus are counted and put to the -smp option
- qemu is started
- qemu is queried for the necessary information
- the configuration is checked
- the hotpluggable vcpus are hotplugged
- vcpus are started
This patch adds a lot of checking code and enables the support to
specify the individual vcpu element with qemu.
The vcpu order information is extracted only for hotpluggable entities,
while vcpu definitions belonging to the same hotpluggable entity need
to all share the order information.
We also can't overwrite it right away in the vcpu info detection code as
the order is necessary to add the hotpluggable vcpus enabled on boot in
the correct order.
The helper will store the order information in places where we are
certain that it's necessary.
Introduce a new migration cookie flag that will be used for any
configurations that are not compatible with libvirt that would not
support the specific vcpu hotplug approach. This will make sure that old
libvirt does not fail to reproduce the configuration correctly.
Individual vCPU hotplug requires us to track the state of any vCPU. To
allow this add the following XML:
<domain>
...
<vcpu current='2'>3</vcpu>
<vcpus>
<vcpu id='0' enabled='yes' hotpluggable='no' order='1'/>
<vcpu id='1' enabled='yes' hotpluggable='yes' order='2'/>
<vcpu id='1' enabled='no' hotpluggable='yes'/>
</vcpus>
...
The 'enabled' attribute allows to control the state of the vcpu.
'hotpluggable' controls whether given vcpu can be hotplugged and 'order'
allows to specify the order to add the vcpus.
Similarly to devices the guest may allow unplug of the VCPU if libvirt
is down. To avoid problems, refresh the vcpu state on reconnect. Don't
mess with the vcpu state otherwise.
Now that the monitor code gathers all the data we can extract it to
relevant places either in the definition or the private data of a vcpu.
As only thread id is broken for TCG guests we may extract the rest of
the data and just skip assigning of the thread id. In case where qemu
would allow cpu hotplug in TCG mode this will make it work eventually.
Power 8 platform's basic hotpluggable unit is a core rather than a
thread for x86_64 family. This introduces most of the complexity of the
matching code and thus needs to be tested.
The test data contain data captured from in-order cpu hotplug and
unplug operations.
During review it was reported that adding at least 11 vcpus creates a
collision of prefixes in the monitor matching algorithm. Add a test case
to verify that the problem won't happen.
As the combination algorithm is rather complex and ugly it's necessary
to make sure it works properly. Add test suite infrastructure for
testing it along with a basic test based on x86_64 platform.
For hotplug purposes it's necessary to retrieve data using
query-hotpluggable-cpus while the old query-cpus API report thread IDs
and order of hotplug.
This patch adds code that merges the data using a rather non-trivial
algorithm and fills the data to the qemuMonitorCPUInfo structure for
adding to appropriate place in the domain definition.
Add support for retrieving information regarding hotpluggable cpu units
supported by qemu. Data returned by the command carries information
needed to figure out the granularity of hotplug, the necessary cpu type
name and the topology information.
Note that qemu doesn't specify any particular order of the entries thus
it's necessary sort them by socket_id, core_id and thread_id to the
order libvirt expects.
To allow matching up the data returned by query-cpus to entries in the
query-hotpluggable-cpus reply for CPU hotplug it's necessary to extract
the QOM path as it's the only link between the two.
QEMU reports whether 'query-hotpluggable-cpus' is supported for a given
machine type. Extract and cache the information using the capability
cache.
When copying the capabilities for a new start of qemu, mask out the
presence of QEMU_CAPS_QUERY_HOTPLUGGABLE_CPUS if the machine type
doesn't support hotpluggable cpus.
As of qemu commit:
commit a32ef3bfc12c8d0588f43f74dcc5280885bbdb30
Author: Thomas Huth <thuth@redhat.com>
Date: Wed Jul 22 15:59:50 2015 +0200
vl: Add another sanity check to smp_parse() function
v2.4.0-952-ga32ef3b
configuration where the maximum CPU count doesn't match the topology is
rejected. Prior to that only configurations where the topology would
contain more cpus than the maximum count would be rejected.
Use QEMU_CAPS_QUERY_HOTPLUGGABLE_CPUS as a relevant recent enough
witness to avoid breaking old configs.
Prepare to extract more data by returning an array of structs rather than
just an array of thread ids. Additionally report fatal errors separately
from qemu not being able to produce data.
All other modes of qemuDomainSetVcpusFlags have helpers so finish the
work by splitting the regular code into a new function.
This patch also touches up the coding (spacing) style.
Mention whether it was the live or persistent definition which caused an
error reported and explicitly error out in case when attempting to set
maximum vcpu count for a live domain.
Call the vcpu thread info validation separately to decrease complexity
of returned values by qemuDomainRefreshVcpuInfo.
This function now returns 0 on success and -1 on error. Certain
failures of qemu to report data are still considered as success. Any
error reported now is fatal.
Validate the presence of the thread id according to state of the vCPU
rather than just checking the vCPU count. Additionally put the new
validation code into a separate function so that the information
retrieval can be split from the validation.
The prettified JSON string already contains a newline so don't print
another one. This allows to pipe the json output (in conjunction with
the --quiet option) to files without having to truncate them afterwards.
Failure to parse the schema file would not trigger a test suite failure.
In addition to making the test fail it's necessary to split up the
parsing of the schema file into a separate test.
This is necessary as the XML validator uses libvirt errors to report
problems parsing of the actual schema RNG needs to be split out into a
separate function and called via virTestRun which has the
infrastructure to report them.
Peter Krempa