I noticed that we allow virDomainGetVcpusFlags even for read-only
connections, but that with a flag, it can require guest agent
interaction. It is feasible that a malicious guest could
intentionally abuse the replies it sends over the guest agent
connection to possibly trigger a bug in libvirt's JSON parser,
or withhold an answer so as to prevent the use of the agent
in a later command such as a shutdown request. Although we
don't know of any such exploits now (and therefore don't mind
posting this patch publicly without trying to get a CVE assigned),
it is better to err on the side of caution and explicitly require
full access to any domain where the API requires guest interaction
to operate correctly.
I audited all commands that are marked as conditionally using a
guest agent. Note that at least virDomainFSTrim is documented
as needing a guest agent, but that such use is unconditional
depending on the hypervisor (so the existing domain:fs_trim ACL
should be sufficient there, rather than also requirng domain:write).
But when designing future APIs, such as the plans for obtaining
a domain's IP addresses, we should copy the approach of this patch
in making interaction with the guest be specified via a flag, and
use that flag to also require stricter access checks.
* src/libvirt.c (virDomainGetVcpusFlags): Forbid guest interaction
on read-only connection.
(virDomainShutdownFlags, virDomainReboot): Improve docs on agent
interaction.
* src/remote/remote_protocol.x
(REMOTE_PROC_DOMAIN_SNAPSHOT_CREATE_XML)
(REMOTE_PROC_DOMAIN_SET_VCPUS_FLAGS)
(REMOTE_PROC_DOMAIN_GET_VCPUS_FLAGS, REMOTE_PROC_DOMAIN_REBOOT)
(REMOTE_PROC_DOMAIN_SHUTDOWN_FLAGS): Require domain:write for any
conditional use of a guest agent.
* src/xen/xen_driver.c: Fix clients.
* src/libxl/libxl_driver.c: Likewise.
* src/uml/uml_driver.c: Likewise.
* src/qemu/qemu_driver.c: Likewise.
* src/lxc/lxc_driver.c: Likewise.
Signed-off-by: Eric Blake <eblake@redhat.com>
Libvirtd would crash if a domain contained an empty cdrom drive of
type='volume' as the disk def->srcpool member would be dereferenced. Fix
it by checking if the source pool is present before dereferencing it.
Also alter tests to catch this issue in the future.
Reported by: Kevin Shanahan
Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=1056328
https://bugzilla.redhat.com/show_bug.cgi?id=1034993
SCSI passthrough disks (<disk .. device="lun">) can't be used as backing
for snapshots. Currently with upstream qemu the vm crashes on such
attempt.
This patch adds a early check to catch an attempt to do such a snapshot
and rejects it right away. qemu will fix the issue but this will let us
control the error message.
We shouldn't access the domain definition while we are in the monitor
section as the domain is unlocked. Additionally after we exit from the
monitor we need to check if the VM is still alive. Not doing so resulted
in a crash if qemu exits while attempting to do an external VM snapshot.
spice-server offers an API to disable file transfer messages
on the agent channel between the client and the guest.
This is supported in qemu through the disable-agent-file-xfer option.
This patch exposes this option to libvirt.
Adds a new element 'filetransfer', with one property,
'enable', which accepts a boolean.
Default is enabled, for backward compatibility.
Depends on the capability exported in the first patch of the series.
Signed-off-by: Francesco Romani <fromani@redhat.com>
spice-server offers an API to disable file transfer messages
on the agent channel between the client and the guest.
This is supported in qemu through the disable-agent-file-xfer option.
This patch detects if QEMU supports this option, and add
a capability if does.
Signed-off-by: Francesco Romani <fromani@redhat.com>
https://bugzilla.redhat.com/show_bug.cgi?id=1046919
If none (KVM, VFIO) of the supported PCI passthrough methods is known to
work on a host, it's better to fail right away with a nice error message
rather than letting attachment fail with a more cryptic message such as
Failed to bind PCI device '0000:07:05.0' to vfio-pci: No such device
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
With this patch, user can setup throttle blkio cgroup
through virsh for qemu domain.
Signed-off-by: Guan Qiang <hzguanqiang@corp.netease.com>
Signed-off-by: Gao feng <gaofeng@cn.fujitsu.com>
This patch introduces virCgroupSetBlkioDeviceReadIops,
virCgroupSetBlkioDeviceWriteIops,
virCgroupSetBlkioDeviceReadBps and
virCgroupSetBlkioDeviceWriteBps,
we can use these interfaces to set up throttle
blkio cgroup for domain.
This patch also adds the new throttle blkio cgroup
elements to the test xml.
Signed-off-by: Guan Qiang <hzguanqiang@corp.netease.com>
Signed-off-by: Gao feng <gaofeng@cn.fujitsu.com>
While auditing the error reporting, I noticed that migration
had some issues. Some of the static helper functions tried
to call virDispatchError(), even though their caller will also
report the error. Also, if a migration is cancelled early
because a uri was not set, we did not guarantee that the finish
stage would not overwrite the first error message.
* src/qemu/qemu_migration.c (doPeer2PeerMigrate2)
(doPeer2PeerMigrate3): Preserve first error when cancelling.
* src/libvirt.c (virDomainMigrateVersion3Full): Likewise.
(virDomainMigrateVersion1, virDomainMigrateVersion2)
(virDomainMigrateDirect): Avoid redundant error dispatch.
(virDomainMigrateFinish2, virDomainMigrateFinish3)
(virDomainMigrateFinish3Params): Don't report error on cleanup
path.
Signed-off-by: Eric Blake <eblake@redhat.com>
The public virConnectRef and virConnectClose API are just thin
wrappers around virObjectRef/virObjectRef, with added object
validation and an error reset. Within our backend drivers, use
of the object validation is just an inefficiency since we always
pass valid objects. More important to think about is what
happens with the error reset; our uses of virConnectRef happened
to be safe (since we hadn't encountered any earlier errors), but
in several cases the use of virConnectClose could lose a real
error.
Ideally, we should also avoid calling virConnectOpen() from
within backend drivers - but that is a known situation that
needs much more design work.
* src/qemu/qemu_process.c (qemuProcessReconnectHelper)
(qemuProcessReconnect): Avoid nested public API call.
* src/qemu/qemu_driver.c (qemuAutostartDomains)
(qemuStateInitialize, qemuStateStop): Likewise.
* src/qemu/qemu_migration.c (doPeer2PeerMigrate): Likewise.
* src/storage/storage_driver.c (storageDriverAutostart):
Likewise.
* src/uml/uml_driver.c (umlAutostartConfigs): Likewise.
* src/lxc/lxc_process.c (virLXCProcessAutostartAll): Likewise.
(virLXCProcessReboot): Likewise, and avoid leaking conn on error.
Signed-off-by: Eric Blake <eblake@redhat.com>
There is a number of reported issues when we fail starting a domain.
Turns out that, in some scenarios like high load, 3 second timeout is
not enough for qemu to start up to the phase where the socket is
created. Since there is no downside of waiting longer, raise the
timeout right to 30 seconds.
Signed-off-by: Martin Kletzander <mkletzan@redhat.com>
Ever since ACL filtering was added in commit 7639736 (v1.1.1), a
user could still use event registration to obtain access to a
domain that they could not normally access via virDomainLookup*
or virConnectListAllDomains and friends. We already have the
framework in the RPC generator for creating the filter, and
previous cleanup patches got us to the point that we can now
wire the filter through the entire object event stack.
Furthermore, whether or not domain:getattr is honored, use of
global events is a form of obtaining a list of networks, which
is covered by connect:search_domains added in a93cd08 (v1.1.0).
Ideally, we'd have a way to enforce connect:search_domains when
doing global registrations while omitting that check on a
per-domain registration. But this patch just unconditionally
requires connect:search_domains, even when no list could be
obtained, based on the following observations:
1. Administrators are unlikely to grant domain:getattr for one
or all domains while still denying connect:search_domains - a
user that is able to manage domains will want to be able to
manage them efficiently, but efficient management includes being
able to list the domains they can access. The idea of denying
connect:search_domains while still granting access to individual
domains is therefore not adding any real security, but just
serves as a layer of obscurity to annoy the end user.
2. In the current implementation, domain events are filtered
on the client; the server has no idea if a domain filter was
requested, and must therefore assume that all domain event
requests are global. Even if we fix the RPC protocol to
allow for server-side filtering for newer client/server combos,
making the connect:serach_domains ACL check conditional on
whether the domain argument was NULL won't benefit older clients.
Therefore, we choose to document that connect:search_domains
is a pre-requisite to any domain event management.
Network events need the same treatment, with the obvious
change of using connect:search_networks and network:getattr.
* src/access/viraccessperm.h
(VIR_ACCESS_PERM_CONNECT_SEARCH_DOMAINS)
(VIR_ACCESS_PERM_CONNECT_SEARCH_NETWORKS): Document additional
effect of the permission.
* src/conf/domain_event.h (virDomainEventStateRegister)
(virDomainEventStateRegisterID): Add new parameter.
* src/conf/network_event.h (virNetworkEventStateRegisterID):
Likewise.
* src/conf/object_event_private.h (virObjectEventStateRegisterID):
Likewise.
* src/conf/object_event.c (_virObjectEventCallback): Track a filter.
(virObjectEventDispatchMatchCallback): Use filter.
(virObjectEventCallbackListAddID): Register filter.
* src/conf/domain_event.c (virDomainEventFilter): New function.
(virDomainEventStateRegister, virDomainEventStateRegisterID):
Adjust callers.
* src/conf/network_event.c (virNetworkEventFilter): New function.
(virNetworkEventStateRegisterID): Adjust caller.
* src/remote/remote_protocol.x
(REMOTE_PROC_CONNECT_DOMAIN_EVENT_REGISTER)
(REMOTE_PROC_CONNECT_DOMAIN_EVENT_REGISTER_ANY)
(REMOTE_PROC_CONNECT_NETWORK_EVENT_REGISTER_ANY): Generate a
filter, and require connect:search_domains instead of weaker
connect:read.
* src/test/test_driver.c (testConnectDomainEventRegister)
(testConnectDomainEventRegisterAny)
(testConnectNetworkEventRegisterAny): Update callers.
* src/remote/remote_driver.c (remoteConnectDomainEventRegister)
(remoteConnectDomainEventRegisterAny): Likewise.
* src/xen/xen_driver.c (xenUnifiedConnectDomainEventRegister)
(xenUnifiedConnectDomainEventRegisterAny): Likewise.
* src/vbox/vbox_tmpl.c (vboxDomainGetXMLDesc): Likewise.
* src/libxl/libxl_driver.c (libxlConnectDomainEventRegister)
(libxlConnectDomainEventRegisterAny): Likewise.
* src/qemu/qemu_driver.c (qemuConnectDomainEventRegister)
(qemuConnectDomainEventRegisterAny): Likewise.
* src/uml/uml_driver.c (umlConnectDomainEventRegister)
(umlConnectDomainEventRegisterAny): Likewise.
* src/network/bridge_driver.c
(networkConnectNetworkEventRegisterAny): Likewise.
* src/lxc/lxc_driver.c (lxcConnectDomainEventRegister)
(lxcConnectDomainEventRegisterAny): Likewise.
Signed-off-by: Eric Blake <eblake@redhat.com>
https://bugzilla.redhat.com/show_bug.cgi?id=1047659
If a VM dies very early during an attempted connect to the guest agent
while the locks are down the domain monitor object will be freed. The
object is then accessed later as any failure during guest agent startup
isn't considered fatal.
In the current upstream version this doesn't lead to a crash as
virObjectLock called when entering the monitor in
qemuProcessDetectVcpuPIDs checks the pointer before attempting to
dereference (lock) it. The NULL pointer is then caught in the monitor
helper code.
Before the introduction of virObjectLockable - observed on 0.10.2 - the
pointer is locked directly via virMutexLock leading to a crash.
To avoid this problem we need to differentiate between the guest agent
not being present and the VM quitting when the locks were down. The fix
reorganizes the code in qemuConnectAgent to add the check and then adds
special handling to the callers.
CVE-2013-6458
Generally, every API that is going to begin a job should do that before
fetching data from vm->def. However, qemuDomainGetBlockInfo does not
know whether it will have to start a job or not before checking vm->def.
To avoid using disk alias that might have been freed while we were
waiting for a job, we use its copy. In case the disk was removed in the
meantime, we will fail with "cannot find statistics for device '...'"
error message.
CVE-2013-6458
https://bugzilla.redhat.com/show_bug.cgi?id=1043069
When virDomainDetachDeviceFlags is called concurrently to
virDomainBlockStats: libvirtd may crash because qemuDomainBlockStats
finds a disk in vm->def before getting a job on a domain and uses the
disk pointer after getting the job. However, the domain in unlocked
while waiting on a job condition and thus data behind the disk pointer
may disappear. This happens when thread 1 runs
virDomainDetachDeviceFlags and enters monitor to actually remove the
disk. Then another thread starts running virDomainBlockStats, finds the
disk in vm->def, and while it's waiting on the job condition (owned by
the first thread), the first thread finishes the disk removal. When the
second thread gets the job, the memory pointed to be the disk pointer is
already gone.
That said, every API that is going to begin a job should do that before
fetching data from vm->def.
This patch fixes a segmentation fault when creating new virtual machines using QEMU.
The segmentation fault is caused by commit f41830680e
and commit cbb6ec42e2.
In virQEMUCapsProbeQMPMachineTypes, when copying machines to qemuCaps, "none" is skipped.
Therefore, the value of i and "qemuCaps->nmachineTypes - 1" do not always match.
However, defIdx value (used to call virQEMUCapsSetDefaultMachine) is set using the value in i
when the array elements are in qemuCaps->nmachineTypes - 1.
So, when libvirt tries to create virtual machines using the default machine type,
qemuCaps->machineTypes[defIdx] is accessed and since the defIdx is NULL, it results in segmentation fault.
Signed-off-by: Yudai Yamagishi <yummy@sfc.wide.ad.jp>
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
Currently, the qemuProcessStop tries to open the domain log file
and saves the original error afterwards. Then all the cleanup is
done after which the error is restored back. This has however one
flaw: if opening of the log file fails an error is reported,
which results in previous error being overwritten (the useful
one, e.g. "PCI device XXXX:XXXX could not be found"). Hence, user
sees something like:
error: failed to create logfile /var/log/libvirt/qemu/ovirt_usb.log: No such file or directory
instead of:
error: internal error: Did not find USB device 8644:8003
Signed-off-by: Michal Privoznik <mprivozn@redhat.com>
Reported-by: Zhou Yimin <zhouyimin@huawei.com>
@listenAddress and @cookiein arguments, should be exchanged,
because the order of the caller and the callee does not match.
This results in the listen address being ignored for peer-to-peer
migration and the cookie being ignored for v2 migration.
Introduced by c4ac7ef (v1.1.4-rc1~141).
https://bugzilla.redhat.com/show_bug.cgi?id=1049338
Signed-off-by: Minoru Usui <usui@mxm.nes.nec.co.jp>
Signed-off-by: Ján Tomko <jtomko@redhat.com>
AArch64 qemu has similar behavior as armv7l, like use of mmio etc.
This patch adds similar bypass checks what we have for armv7l to aarch64.
E.g. we are enabling mmio transport for Nicdev.
Making addDefaultUSB and addDefaultMemballoon to false etc.
V3:
- Adding missing domain rng schema for aarcg64 and test case in
testutilsqemu.c which was causing test suite failure
while running make check.
V2:
- Added testcase to qemuxml2argvtest as suggested
during review comments of V1.
V1:
- Initial patch.
Signed-off-by: Anup Patel <anup.patel@linaro.org>
Signed-off-by: Pranavkumar Sawargaonkar <pranavkumar@linaro.org>
https://bugzilla.redhat.com/show_bug.cgi?id=1047234
Add a range check for supported numa memory placement modes provided by
the user before setting them in the domain definition. Without the check
the user is able to provide a (yet) unknown mode which is then stored in
the domain definition. This potentially causes a NULL dereference when
the defintion is formatted into the XML.
To reproduce run:
virsh numatune DOMNAME --mode 6 --nodeset 0
The XML will then contain:
<numatune>
<memory mode='(null)' nodeset='0'/>
</numatune>
With this fix, the command fails:
error: Unable to change numa parameters
error: invalid argument: unsupported numa_mode: '6'
Add whitespace to separate logical code blocks, reformat error messages
and clean up code flow.
This patch changes error handling in some cases where the the loop would
be continued to jump to cleanup instead and error out rather than modify
the domain any further.
Do not leave the PCI address of the primary video card set
to the legacy default (0000:00:02.0) if we're doing two-pass
allocation.
Since QEMU 1.6 (QEMU_CAPS_VIDEO_PRIMARY) we allow the primary
video card to be on other slots than 0000:00:02.0 (as we use
-device instead of -vga).
However we fail to assign it an address if:
* another device explicitly uses 0000:00:02.0 and
* the primary video device has no address specified
On the first pass, we have set the address to default, then checked
if it's available, leaving it set even if it wasn't. This address
got picked up by the second pass, resulting in a conflict:
XML error: Attempted double use of PCI slot 0000:00:02.0
(may need "multifunction='on'" for device on function 0)
Also fix the test that was supposed to catch this.
This eliminates the misleading error message that was being logged
when a vfio hostdev hotplug failed:
error: unable to set user and group to '107:107' on '/dev/vfio/22':
No such file or directory
as documented in:
https://bugzilla.redhat.com/show_bug.cgi?id=1035490
Commit ee414b5d (pushed as a fix for Bug 1016511 and part of Bug
1025108) replaced the single call to
virSecurityManagerSetHostdevLabel() in qemuDomainAttachHostDevice()
with individual calls to that same function in each
device-type-specific attach function (for PCI, USB, and SCSI). It also
added a corresponding call to virSecurityManagerRestoreHostdevLabel()
in the error handling of the device-type-specific functions, but
forgot to remove the common call to that from
qemuDomainAttachHostDevice() - this resulted in a duplicate call to
virSecurityManagerRestoreHostdevLabel(), with the second occurrence
being after (e.g.) a PCI device has already been re-attached to the
host driver, thus destroying some of the device nodes / links that we
then attempted to re-label (e.f. /dev/vfio/22) and generating an error
log that obscured the original error.
This resolves:
https://bugzilla.redhat.com/show_bug.cgi?id=1035490
virProcessSetMaxMemLock() (which is a wrapper over prlimit(3)) expects
the memory size in bytes, but libvirt's domain definition (which was
being used by qemuDomainAttachHostPciDevice()) stores all memory
tuning parameters in KiB. This was being accounted for when setting
MaxMemLock at domain startup time (so cold-plugged devices would
work), but not for hotplug.
This patch simplifies the few lines that call
virProcessSetMemMaxLock(), and multiply the amount * 1024 so that
we're locking the correct amount of memory.
What remains a mystery to me is why hot-plug of a managed='no' device
would succeed (at least on my system) while managed='yes' would
fail. I guess in one case the memory was coincidentally already
resident and in the other it wasn't.
On a system that is enforcing FIPS, most libraries honor the
current mode by default. Qemu, on the other hand, refused to
honor FIPS mode unless you add the '-enable-fips' command
line option; worse, this option is not discoverable via QMP,
and is only present on binaries built for Linux. So, if we
detect FIPS mode, then we unconditionally ask for FIPS; either
qemu is new enough to have the option and then correctly
cripple insecure VNC passwords, or it is so old that we are
correctly avoiding a FIPS violation by preventing qemu from
starting. Meanwhile, if we don't detect FIPS mode, then
omitting the argument is safe whether the qemu has the option
(but it would do nothing because FIPS is disabled) or whether
qemu lacks the option (including in the case where we are not
running on Linux).
The testsuite was a bit interesting: we don't want our test
to depend on whether it is being run in FIPS mode, so I had
to tweak things to set the capability bit outside of our
normal interaction with capability parsing.
This fixes https://bugzilla.redhat.com/show_bug.cgi?id=1035474
* src/qemu/qemu_capabilities.h (QEMU_CAPS_ENABLE_FIPS): New bit.
* src/qemu/qemu_capabilities.c (virQEMUCapsInitQMP): Conditionally
set capability according to detection of FIPS mode.
* src/qemu/qemu_command.c (qemuBuildCommandLine): Use it.
* tests/qemucapabilitiestest.c (testQemuCaps): Conditionally set
capability to test expected output.
* tests/qemucapabilitiesdata/caps_1.2.2-1.caps: Update list.
* tests/qemucapabilitiesdata/caps_1.6.0-1.caps: Likewise.
Signed-off-by: Eric Blake <eblake@redhat.com>
The support for <boot rebootTimeout="12345"/> was added before we were
checking for qemu command line options in QMP, so we haven't properly
adapted virQEMUCaps when using it and thus we report unsupported
option with new enough qemu.
Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=1042690
Signed-off-by: Martin Kletzander <mkletzan@redhat.com>
Recent changes to events (commit 8a29ffcf) resulted in new compile
failures on some targets (such as ARM OMAP5):
conf/domain_event.c: In function 'virDomainEventDispatchDefaultFunc':
conf/domain_event.c:1198:30: error: cast increases required alignment of
target type [-Werror=cast-align]
conf/domain_event.c:1314:34: error: cast increases required alignment of
target type [-Werror=cast-align]
cc1: all warnings being treated as errors
The error is due to alignment; the base class is merely aligned
to the worst of 'int' and 'void*', while the child class must
be aligned to a 'long long'. The solution is to include a
'long long' (and for good measure, a function pointer) in the
base class to ensure correct alignment regardless of what a
child class may add, but to wrap the inclusion in a union so
as to not incur any wasted space. On a typical x86_64 platform,
the base class remains 16 bytes; on i686, the base class remains
12 bytes; and on the impacted ARM platform, the base class grows
from 12 bytes to 16 bytes due to the increase of alignment from
4 to 8 bytes.
Reported by Michele Paolino and others.
* src/util/virobject.h (_virObject): Use a union to ensure that
subclasses never have stricter alignment than the parent.
* src/util/virobject.c (virObjectNew, virObjectUnref)
(virObjectRef): Adjust clients.
* src/libvirt.c (virConnectRef, virDomainRef, virNetworkRef)
(virInterfaceRef, virStoragePoolRef, virStorageVolRef)
(virNodeDeviceRef, virSecretRef, virStreamRef, virNWFilterRef)
(virDomainSnapshotRef): Likewise.
* src/qemu/qemu_monitor.c (qemuMonitorOpenInternal)
(qemuMonitorClose): Likewise.
Signed-off-by: Eric Blake <eblake@redhat.com>
Map the new <panic> device in XML to the '-device pvpanic' command
line of qemu. Clients can then couple the <panic> device and the
<on_crash> directive to control behavior when the guest reports
a panic to qemu.
Signed-off-by: Hu Tao <hutao@cn.fujitsu.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
https://bugzilla.redhat.com/show_bug.cgi?id=1035955
There's a window when starting a qemu process between fork() and exec()
during which we are doing things that may fail but not tunnelling the
error to the daemon. This is basically all within qemuProcessHook().
So whenever we fail in something, e.g. placing a process onto numa node,
users are left with:
error: Child quit during startup handshake: Input/output error
while the original error is thrown into the domain log:
libvirt: error : internal error: NUMA memory tuning in 'preferred'
mode only supports single node
Hence, we should read the log file and search for the error message and
report it to users.
Signed-off-by: Michal Privoznik <mprivozn@redhat.com>
For dead domains that have no memtune limits, we return 0 instead of
"unlimited", this patch fixes it to return PARAM_UNLIMITED.
Signed-off-by: Martin Kletzander <mkletzan@redhat.com>
We were unconditionally removing the device from the host list, when it
should only be done on error.
This fixes USB collision detection when hotplugging the same device to
two guests.
If we hit a collision, we free the USB device while it is still part
of our temporary USBDeviceList. When the list is unref'd, the device
is free'd again.
Make the initial device freeing dependent on whether it is present
in the temporary list or not.
Similar to what Jiri did for cgroup setup/teardown in 05e149f94, push
it all into the device handler functions so we can do the necessary prep
work before claiming the device.
This also fixes hotplugging USB devices by product/vendor (virt-manager's
default behavior):
https://bugzilla.redhat.com/show_bug.cgi?id=1016511
https://bugzilla.redhat.com/show_bug.cgi?id=1035108
When attempting to enable more vCPUs in the guest than is currently
enabled in the guest but less than the maximum count for the VM we
currently reported an unhelpful message:
error: internal error: guest agent reports less cpu than requested
This patch changes it to:
error: invalid argument: requested vcpu count is greater than the count
of enabled vcpus in the domain: 3 > 2
