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.
So far, when CPUID data were converted into CPU model and features, the
features can only be added to the model. As a result, when a guest asked
for something like "qemu64,-svm" it would get a qemu32 plus a bunch of
additional features instead.
This patch adds support for removing feature from the base model.
Selection algorithm remains the same: the best CPU model is the model
which requires lowest number of features to be added/removed from it.
Qemu committed a patch which list some CPU names in [] when asked for
supported CPUs (qemu -cpu ?). Yet, it needs such CPUs to be passed
without those square braces. When probing for supported CPU models, we
can just strip the square braces and pretend we have never seen them.
First, inital VCPU pinning is set correctly but then it is reset by
assigning qemu process to a new cgroup (which contains all CPUs). It's
easily fixed by swapping these two actions.
The idea is that every API implementation in driver which has flags
parameter should first call virCheckFlags() macro to check the function
was called with supported flags:
virCheckFlags(VIR_SUPPORTED_FLAG_1 |
VIR_SUPPORTED_FLAG_2 |
VIR_ANOTHER_SUPPORTED_FLAG, -1);
The error massage which is printed when unsupported flags are passed
looks like:
invalid argument in virFooBar: unsupported flags (0x2)
Where the unsupported flags part only prints those flags which were
passed but are not supported rather than all flags passed.
When comparing a CPU to host CPU, the result would be
VIR_CPU_COMPARE_SUPERSET (or even VIR_CPU_COMPARE_INCOMPATIBLE if strict
match was required) even though the two CPUs were identical.
When qemu libvirt driver doesn't support guest CPU selection with given
qemu binary, guests requiring specific CPU should fail to start instead
of being silently supplied with a default CPU.
This flag is used in migration prepare step to send updated XML
definition of a guest.
Also ``virsh dumpxml --update-cpu [--inactive] guest'' command can be
used to see the updated CPU requirements.
Useful mainly for migration. cpuUpdate changes guest CPU requirements in
the following way:
- match == "strict" || match == "exact"
- optional features which are supported by host CPU are changed into
required features
- optional features which are not supported by host CPU are disabled
- all other features remain untouched
- match == "minimum"
- match is changed into "exact"
- optional features and all features not mentioned in guest CPU
specification which are supported by host CPU become required
features
- other optional features are disabled
- all other features remain untouched
This ensures that no feature will suddenly disappear from the guest
after migration.
When a domain is defined on host1, migrated to host2 and then migrated
back to host1, its current configuration would overwrite the libvirtd's
in-memory copy of persistent configuration of that domain. This is not
desired as we want to preserve the persistent configuration untouched.
This patch introduces new 'live' parameter to virDomainAssignDef.
Passing 'true' for 'live' means the configuration passed to
virDomainAssignDef describes a configuration of live instance of the
domain. This applies for saved domains which are being restored or for
incoming domains during migration.
All callers have been changed to pass the appropriate value.
Currently no command can be sent to a qemu process while another job is
active. This patch adds support for signaling long-running jobs (such as
migration) so that other threads may request predefined operations to be
done during such jobs. Two signals are defined so far:
- QEMU_JOB_SIGNAL_CANCEL
- QEMU_JOB_SIGNAL_SUSPEND
The first one is used by qemuDomainAbortJob.
The second one is used by qemudDomainSuspend for suspending a domain
during migration, which allows for changing live migration into offline
migration. However, there is a small issue in the way qemudDomainSuspend
is currently implemented for migrating domains. The API calls returns
immediately after signaling migration job which means it is asynchronous
in this specific case.
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
The nodeinfo structure wasn't initialized in qemu driver and with the
recent change in CPU topology parsing, old value of nodeinfo->sockets
could be used and incremented giving totally bogus results.
Let's just wipe the structure completely.
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
Various safezero() implementations used either -1, errno or -errno
return values. This patch fixes them all to return -1 and set errno
appropriately.
There was also a bug in size parameter passed to safewrite() which could
result in an attempt to write gigabytes out of a megabyte buffer.
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
Recently we introduced O_DSYNC flag when creating raw storage files to
avoid filling all disk cache with dirty pages. However, the patch got
lost when virStorageBackendCreateRaw was reworked using
virFileOperation. Let's use O_DSYNC again.
All other libvirt functions use array first and then number of elements
in that array. Let's make cpuDecode follow this rule.
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
The current schema is more permissive than the XML parsing code in
libvirt. Precisely, 'match' attribute is optional in schema while in
reality its use is bound to <model> element:
- <cpu> element without 'match' attribute is allowed only if <topology>
element is the only child element of <cpu>
- <cpu> element with 'match' attribute requires <model> element to be
present; <topology> and <feature> elements are optional
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
When creating preallocated large raw files opening them with O_DSYNC
prevents long delays in reading because cache pages can be immediately
reused without writing them on a disk first.
I noticed some debug messages are printed with an empty lines after
them. This patch removes these empty lines from all invocations of the
following macros:
VIR_DEBUG
VIR_DEBUG0
VIR_ERROR
VIR_ERROR0
VIR_INFO
VIR_WARN
VIR_WARN0
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
New pciDeviceIsAssignable() function for checking whether a given PCI
device can be assigned to a guest was added. Currently it only checks
for ACS being enabled on all PCIe switches between root and the PCI
device. In the future, it could be the right place to check whether a
device is unbound or bound to a stub driver.
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
Currently CPU topology may only be specified together with CPU model:
<cpu match='exact'>
<model>name</model>
<topology sockets='1' cores='2' threads='3'/>
</cpu>
This patch allows for CPU topology specification without the need for
also specifying CPU model:
<cpu>
<topology sockets='1' cores='2' threads='3'/>
</cpu>
'match' attribute and 'model' element are made optional with the
restriction that 'match' attribute has to be set when 'model' is
present.
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
When comparing x86 CPUs, features with 'disabled' policy were mistakenly
required to be supported by the host CPU.
Likewise, features with 'force' policy which were supported by host CPU
would make CPUs incompatible if 'strict' match was used by guest CPU.
This patch fixes both issues.
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
QEMU's command line equivalent for the following domain XML fragment
<vcpus>2</vcpus>
<cpu ...>
...
<topology sockets='1' cores='2', threads='1'/>
</cpu>
is
-smp 2,sockets=1,cores=2,threads=1
This syntax was introduced in QEMU-0.12.
Version 2 changes:
- -smp argument build split into a separate function
- always add ",sockets=S,cores=C,threads=T" to -smp if qemu supports it
- use qemuParseCommandLineKeywords for command line parsing
Version 3 changes:
- ADD_ARG_LIT => ADD_ARG and line reordering in qemudBuildCommandLine
- rebased
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
Current version expects name=value,... list and when an incorrect string
such as "a,b,c=d" would be parsed as "a,b,c" keyword with "d" value
without reporting any error, which is probably not the expected
behavior.
This patch adds an extra argument called allowEmptyValue, which if
non-zero will permit keywords with no value; "a,b=c,,d=" will be parsed
as follows:
keyword value
"a" NULL
"b" "c"
"" NULL
"d" ""
In case allowEmptyValue is zero, the string is required to contain
name=value pairs only; retvalues is guaranteed to contain non-NULL
pointers. Now, "a,b,c=d" will result in an error.
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
Current implementation of x86Decode() used for CPUID -> model+features
translation does not always select the closest CPU model. When walking
through all models from cpu_map.xml the function considers a new
candidate as a better choice than a previously selected candidate only
if the new one is a superset of the old one. In case the new candidate
is closer to host CPU but lacks some feature comparing to the old
candidate, the function does not choose well.
This patch changes the algorithm so that the closest model is always
selected. That is, the model which requires the lowest number of
additional features to describe host CPU.
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
Each driver supporting CPU selection must fill in host CPU capabilities.
When filling them, drivers for hypervisors running on the same node as
libvirtd can use cpuNodeData() to obtain raw CPU data. Other drivers,
such as VMware, need to implement their own way of getting such data.
Raw data can be decoded into virCPUDefPtr using cpuDecode() function.
When implementing virConnectCompareCPU(), a hypervisor driver can just
call cpuCompareXML() function with host CPU capabilities.
For each guest for which a driver supports selecting CPU models, it must
set the appropriate feature in guest's capabilities:
virCapabilitiesAddGuestFeature(guest, "cpuselection", 1, 0)
Actions needed when a domain is being created depend on whether the
hypervisor understands raw CPU data (currently CPUID for i686, x86_64
architectures) or symbolic names has to be used.
Typical use by hypervisors which prefer CPUID (such as VMware and Xen):
- convert guest CPU configuration from domain's XML into a set of raw
data structures each representing one of the feature policies:
cpuEncode(conn, architecture, guest_cpu_config,
&forced_data, &required_data, &optional_data,
&disabled_data, &forbidden_data)
- create a mask or whatever the hypervisor expects to see and pass it
to the hypervisor
Typical use by hypervisors with symbolic model names (such as QEMU):
- get raw CPU data for a computed guest CPU:
cpuGuestData(conn, host_cpu, guest_cpu_config, &data)
- decode raw data into virCPUDefPtr with a possible restriction on
allowed model names:
cpuDecode(conn, guest, data, n_allowed_models, allowed_models)
- pass guest->model and guest->features to the hypervisor
* src/cpu/cpu.c src/cpu/cpu.h src/cpu/cpu_generic.c
src/cpu/cpu_generic.h src/cpu/cpu_map.c src/cpu/cpu_map.h
src/cpu/cpu_x86.c src/cpu/cpu_x86.h src/cpu/cpu_x86_data.h
* configure.in: check for CPUID instruction
* src/Makefile.am: glue the new files in
* src/libvirt_private.syms: add new private symbols
* po/POTFILES.in: add new cpu files containing translatable strings
* src/remote/remote_protocol.x: update with new entry point
* daemon/remote.c: add the new server dispatcher
* daemon/remote_dispatch_args.h daemon/remote_dispatch_prototypes.h
daemon/remote_dispatch_ret.h daemon/remote_dispatch_table.h
src/remote/remote_protocol.c src/remote/remote_protocol.h: regenerated
* src/driver.h: add an extra entry point in the structure
* src/esx/esx_driver.c src/lxc/lxc_driver.c src/opennebula/one_driver.c
src/openvz/openvz_driver.c src/phyp/phyp_driver.c src/qemu/qemu_driver.c
src/remote/remote_driver.c src/test/test_driver.c src/uml/uml_driver.c
src/vbox/vbox_tmpl.c src/xen/xen_driver.c: add NULL entry points for
all drivers
* include/libvirt/virterror.h src/util/virterror.c: add new domain
VIR_FROM_CPU for errors
* src/conf/cpu_conf.c src/conf/cpu_conf.h: new parsing module
* src/Makefile.am proxy/Makefile.am: include new files
* src/conf/capabilities.[ch] src/conf/domain_conf.[ch]: reference
new code
* src/libvirt_private.syms: private export of new entry points
XML schema for CPU flags
Firstly, CPU topology and model with optional features have to be
advertised in host capabilities:
<host>
<cpu>
<arch>ARCHITECTURE</arch>
<features>
<!-- old-style features are here -->
</features>
<model>NAME</model>
<topology sockets="S" cores="C" threads="T"/>
<feature name="NAME"/>
</cpu>
...
</host>
Secondly, drivers which support detailed CPU specification have to
advertise
it in guest capabilities:
<guest>
...
<features>
<cpuselection/>
</features>
</guest>
And finally, CPU may be configured in domain XML configuration:
<domain>
...
<cpu match="MATCH">
<model>NAME</model>
<topology sockets="S" cores="C" threads="T"/>
<feature policy="POLICY" name="NAME"/>
</cpu>
</domain>
Where MATCH can be one of:
- 'minimum' specified CPU is the minimum requested CPU
- 'exact' disable all additional features provided by host CPU
- 'strict' fail if host CPU doesn't exactly match
POLICY can be one of:
- 'force' turn on the feature, even if host doesn't have it
- 'require' fail if host doesn't have the feature
- 'optional' match host
- 'disable' turn off the feature, even if host has it
- 'forbid' fail if host has the feature
'force' and 'disable' policies turn on/off the feature regardless of its
availability on host. 'force' is unlikely to be used but its there for
completeness since Xen and VMWare allow it.
'require' and 'forbid' policies prevent a guest from being started on a host
which doesn't/does have the feature. 'forbid' is for cases where you disable
the feature but a guest may still try to access it anyway and you don't want
it to succeed.
'optional' policy sets the feature according to its availability on host.
When a guest is booted on a host that has the feature and then migrated to
another host, the policy changes to 'require' as we can't take the feature
away from a running guest.
Default policy for features provided by host CPU but not specified in domain
configuration is set using match attribute of cpu tag. If 'minimum' match is
requested, additional features will be treated as if they were specified
with 'optional' policy. 'exact' match implies 'disable' policy and 'strict'
match stands for 'forbid' policy.
* docs/schemas/capability.rng docs/schemas/domain.rng: extend the
RelaxNG schemas to add CPU flags support
src/node_device/node_device_udev.c was using a function available only
on the daemon code, fix this and use the function available globally
* src/node_device/node_device_udev.c: replace use of virEventAddHandleImpl
by virEventAddHandle
Some of the very useful calls for XML parsing provided by util/xml.[ch]
were not exported as private symbols. This patch fixes this.
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
Jiri Denemark