In the past, generic SCSI commands issued from a guest to a virtio
disk were always passed through to the underlying disk by qemu, and
the kernel would also pass them on.
As a result of CVE-2011-4127 (see:
http://seclists.org/oss-sec/2011/q4/536), qemu now honors its
scsi=on|off device option for virtio-blk-pci (which enables/disables
passthrough of generic SCSI commands), and the kernel will only allow
the commands for physical devices (not for partitions or logical
volumes). The default behavior of qemu is still to allow sending
generic SCSI commands to physical disks that are presented to a guest
as virtio-blk-pci devices, but libvirt prefers to disable those
commands in the standard virtio block devices, enabling it only when
specifically requested (hopefully indicating that the requester
understands what they're asking for). For this purpose, a new libvirt
disk device type (device='lun') has been created.
device='lun' is identical to the default device='disk', except that:
1) It is only allowed if bus='virtio', type='block', and the qemu
version is "new enough" to support it ("new enough" == qemu 0.11 or
better), otherwise the domain will fail to start and a
CONFIG_UNSUPPORTED error will be logged).
2) The option "scsi=on" will be added to the -device arg to allow
SG_IO commands (if device !='lun', "scsi=off" will be added to the
-device arg so that SG_IO commands are specifically forbidden).
Guests which continue to use disk device='disk' (the default) will no
longer be able to use SG_IO commands on the disk; those that have
their disk device changed to device='lun' will still be able to use SG_IO
commands.
*docs/formatdomain.html.in - document the new device attribute value.
*docs/schemas/domaincommon.rng - allow it in the RNG
*tests/* - update the args of several existing tests to add scsi=off, and
add one new test that will test scsi=on.
*src/conf/domain_conf.c - update domain XML parser and formatter
*src/qemu/qemu_(command|driver|hotplug).c - treat
VIR_DOMAIN_DISK_DEVICE_LUN *almost* identically to
VIR_DOMAIN_DISK_DEVICE_DISK, except as indicated above.
Note that no support for this new device value was added to any
hypervisor drivers other than qemu, because it's unclear what it might
mean (if anything) to those drivers.
This patch adds two capabilities flags to deal with various aspects
of supporting SG_IO commands on virtio-blk-pci devices:
QEMU_CAPS_VIRTIO_BLK_SCSI
set if -device virtio-blk-pci accepts the scsi="on|off" option
When present, this is on by default, but can be set to off to disable
SG_IO functions.
QEMU_CAPS_VIRTIO_BLK_SG_IO
set if SG_IO commands are supported in the virtio-blk-pci driver
(present since qemu 0.11 according to a qemu developer, if I
understood correctly)
This fixes https://bugzilla.redhat.com/show_bug.cgi?id=638633
Although scripts are not used by interfaces of type other than
"ethernet" in qemu, due to the fact that the parser stores the script
name in a union that is only valid when type is ethernet or bridge,
there is no way for anyone except the parser itself to catch the
problem of specifying an interface script for an inappropriate
interface type (by the time the parsed data gets back to the code that
called the parser, all evidence that a script was specified is
forgotten).
Since the parser itself should be agnostic to which type of interface
allows scripts (an example of why: a script specified for an interface
of type bridge is valid for xen domains, but not for qemu domains),
the solution here is to move the script out of the union(s) in the
DomainNetDef, always populate it when specified (regardless of
interface type), and let the driver decide whether or not it is
appropriate.
Currently the qemu, xen, libxml, and uml drivers recognize the script
parameter and do something with it (the uml driver only to report that
it isn't supported). Those drivers have been updated to log a
CONFIG_UNSUPPORTED error when a script is specified for an interface
type that's inappropriate for that particular hypervisor.
(NB: There was earlier discussion of solving this problem by adding a
VALIDATE flag to all libvirt APIs that accept XML, which would cause
the XML to be validated against the RNG files. One statement during
that discussion was that the RNG shouldn't contain hypervisor-specific
things, though, and a proper solution to this problem would require
that (again, because a script for an interface of type "bridge" is
accepted by xen, but not by qemu).
Commit ae523427 missed one pair of functions that could use
the helper routine.
* src/qemu/qemu_driver.c (qemuSetSchedulerParametersFlags)
(qemuGetSchedulerParametersFlags): Simplify.
Detected by valgrind. Leak introduced in commit 5745dc1.
* src/qemu/qemu_command.c: fix memory leak on failure and successful path.
* How to reproduce?
% valgrind -v --leak-check=full ./qemuargv2xmltest
* Actual result:
==2196== 80 bytes in 1 blocks are definitely lost in loss record 3 of 4
==2196== at 0x4A05FDE: malloc (vg_replace_malloc.c:236)
==2196== by 0x39CF07F6E1: strdup (in /lib64/libc-2.12.so)
==2196== by 0x419823: qemuParseRBDString (qemu_command.c:1657)
==2196== by 0x4221ED: qemuParseCommandLine (qemu_command.c:5934)
==2196== by 0x422AFB: qemuParseCommandLineString (qemu_command.c:7561)
==2196== by 0x416864: testCompareXMLToArgvHelper (qemuargv2xmltest.c:48)
==2196== by 0x417DB1: virtTestRun (testutils.c:141)
==2196== by 0x415CAF: mymain (qemuargv2xmltest.c:175)
==2196== by 0x4174A7: virtTestMain (testutils.c:696)
==2196== by 0x39CF01ECDC: (below main) (in /lib64/libc-2.12.so)
==2196==
==2196== LEAK SUMMARY:
==2196== definitely lost: 80 bytes in 1 blocks
Signed-off-by: Alex Jia <ajia@redhat.com>
When setting numa nodeset for a domain which has no nodeset set
before, libvirtd crashes by dereferencing the pointer to the old
nodemask which is null in that case.
Commit baade4d fixed a memory leak on failure, but in the process,
introduced a use-after-free on success, which can be triggered with:
1. set bandwidth with --live
2. query bandwidth
3. set bandwidth with --live
* src/qemu/qemu_driver.c (qemuDomainSetInterfaceParameters): Don't
free newBandwidth on success.
Reported by Hu Tao.
Most severe here is a latent (but currently untriggered) memory leak
if any hypervisor ever adds a string interface property; the
remainder are mainly cosmetic.
* include/libvirt/libvirt.h.in (VIR_DOMAIN_BANDWIDTH_*): Move
macros closer to interface that uses them, and document type.
* src/libvirt.c (virDomainSetInterfaceParameters)
(virDomainGetInterfaceParameters): Formatting tweaks.
* daemon/remote.c (remoteDispatchDomainGetInterfaceParameters):
Avoid memory leak.
* src/libvirt_public.syms (LIBVIRT_0.9.9): Sort lines.
* src/libvirt_private.syms (domain_conf.h): Likewise.
* src/qemu/qemu_driver.c (qemuDomainSetInterfaceParameters): Fix
comments, break long lines.
Leak detected by Coverity, and introduced in commit 93ab585.
Reported by Alex Jia.
* src/qemu/qemu_driver.c (qemuDomainSetBlkioParameters): Free
devices array on error.
https://bugzilla.redhat.com/show_bug.cgi?id=770520
We had two nested loops both trying to use 'i' as the iteration
variable, which can result in an infinite loop when the inner
loop interferes with the outer loop. Introduced in commit 93ab585.
* src/qemu/qemu_driver.c (qemuDomainSetBlkioParameters): Don't
reuse iteration variable across two loops.
This patch adds max_files option to qemu.conf which can be used to
override system default limit on number of opened files that are
allowed for qemu user.
Add logic to assign addresses for devices with spapr-vio addresses.
We also do validation of addresses specified by the user, ie. ensuring
that there are not duplicate addresses on the bus.
Signed-off-by: Michael Ellerman <michael@ellerman.id.au>
For QEMU PPC64 we have a machine type ("pseries") which has a virtual
bus called "spapr-vio". We need to be able to create devices on this
bus, and as such need a way to specify the address for those devices.
This patch adds a new address type "spapr-vio", which achieves this.
The addressing is specified with a "reg" property in the address
definition. The reg is optional, if it is not specified QEMU will
auto-assign an address for the device.
Signed-off-by: Michael Ellerman <michael@ellerman.id.au>
Currently non-x86 guests must have <acpi/> defined in <features> to
prevent libvirt from running qemu with -no-acpi. Although it works, it
is a hack.
Instead add a capability flag which indicates whether qemu understands
the -no-acpi option. Use it to control whether libvirt emits -no-acpi.
Current versions of qemu always display -no-acpi in their help output,
so this patch has no effect. However the development version of qemu
has been modified such that -no-acpi is only displayed when it is
actually supported.
Signed-off-by: Michael Ellerman <michael@ellerman.id.au>
The lifetime of the virDomainEventState object is tied to
the lifetime of the driver, which in stateless drivers is
tied to the lifetime of the virConnectPtr.
If we add & remove a timer when allocating/freeing the
virDomainEventState object, we can get a situation where
the timer still triggers once after virDomainEventState
has been freed. The timeout callback can't keep a ref
on the event state though, since that would be a circular
reference.
The trick is to only register the timer when a callback
is registered with the event state & remove the timer
when the callback is unregistered.
The demo for the bug is to run
while true ; do date ; ../tools/virsh -q -c test:///default 'shutdown test; undefine test; dominfo test' ; done
prior to this fix, it will frequently hang and / or
crash, or corrupt memory
Currently all drivers using domain events need to provide a callback
for handling a timer to dispatch events in a clean stack. There is
no technical reason for dispatch to go via driver specific code. It
could trivially be dispatched directly from the domain event code,
thus removing tedious boilerplate code from all drivers
Also fix the libxl & xen drivers to pass 'true' when creating the
virDomainEventState, since they run inside the daemon & thus always
expect events to be present.
* src/conf/domain_event.c, src/conf/domain_event.h: Internalize
dispatch of events from timer callback
* src/libxl/libxl_driver.c, src/lxc/lxc_driver.c,
src/qemu/qemu_domain.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: Remove all timer dispatch functions
When registering a callback for a particular event some callers
need to know how many callbacks already exist for that event.
While it is possible to ask for a count, this is not free from
race conditions when threaded. Thus the API for registering
callbacks should return the count of callbacks. Also rename
virDomainEventStateDeregisterAny to virDomainEventStateDeregisterID
* src/conf/domain_event.c, src/conf/domain_event.h,
src/libvirt_private.syms: Return count of callbacks when
registering callbacks
* src/libxl/libxl_driver.c, src/libxl/libxl_driver.c,
src/qemu/qemu_driver.c, src/remote/remote_driver.c,
src/remote/remote_driver.c, src/uml/uml_driver.c,
src/vbox/vbox_tmpl.c, src/xen/xen_driver.c: Update
for change in APIs
A generic error code was returned, if the user aborted a migration job.
This made it hard to distinguish between a user requested abort and an
error that might have occured. This patch introduces a new error code,
which is returned in the specific case of a user abort, while leaving
all other failures with their existing code. This makes it easier to
distinguish between failure while mirgrating and an user requested
abort.
* include/libvirt/virterror.h: - add new error code
* src/util/virterror.c: - add message for the new error code
* src/qemu/qemu_migration.h: - Emit operation aborted error instead of
operation failed, on migration abort
If managed save fails at the right point in time, then the save
image can end up with 0 bytes in length (no valid header), and
our attempts in commit 55d88def to detect and skip invalid save
files missed this case.
* src/qemu/qemu_driver.c (qemuDomainSaveImageOpen): Also unlink
empty file as corrupt. Reported by Dennis Householder.
Currently, on device detach, we parse given XML, find the device
in domain object, free it and try to restore security labels.
However, in some cases (e.g. usb hostdev) parsed XML contains
less information than freed device. In usb case it is bus & device
IDs. These are needed during label restoring as a symlink into
/dev/bus is generated from them. Therefore don't drop device
configuration until security labels are restored.
In commit 6f84e110 I mistakenly set default migration speed to
33554432 Mb! The units of migMaxBandwidth is Mb, with conversion
handled in qemuMonitor{JSON,Text}SetMigrationSpeed().
Also, remove definition of QEMU_DOMAIN_FILE_MIG_BANDWIDTH_MAX since
it is no longer used after reverting commit ef1065cf.
If an async job run on a domain will stop the domain at the end of the
job, a concurrently run query job can hang in qemu monitor and nothing
can be done with that domain from this point on. An attempt to start
such domain results in "Timed out during operation: cannot acquire state
change lock" error.
However, quite a few things have to happen at the right time... There
must be an async job running which stops a domain at the end. This race
was reported with dump --crash but other similar jobs, such as
(managed)save and migration, should be able to trigger this bug as well.
While this async job is processing its last monitor command, that is a
query-migrate to which qemu replies with status "completed", a new
libvirt API that results in a query job must arrive and stay waiting
until the query-migrate command finishes. Once query-migrate is done but
before the async job closes qemu monitor while stopping the domain, the
other thread needs to wake up and call qemuMonitorSend to send its
command to qemu. Before qemu gets a chance to respond to this command,
the async job needs to close the monitor. At this point, the query job
thread is waiting for a condition that no-one will ever signal so it
never finishes the job.
* src/qemu/qemu_hostdev.c (qemuDomainReAttachHostdevDevices):
pciDeviceListFree(pcidevs) in the end free()s the device even if
it's in use by other domain, which can cause a race.
How to reproduce:
<script>
virsh nodedev-dettach pci_0000_00_19_0
virsh start test
virsh attach-device test hostdev.xml
virsh start test2
for i in {1..5}; do
echo "[ -- ${i}th time --]"
virsh nodedev-reattach pci_0000_00_19_0
done
echo "clean up"
virsh destroy test
virsh nodedev-reattach pci_0000_00_19_0
</script>
Device pci_0000_00_19_0 dettached
Domain test started
Device attached successfully
error: Failed to start domain test2
error: Requested operation is not valid: PCI device 0000:00:19.0 is in use by domain test
[ -- 1th time --]
Device pci_0000_00_19_0 re-attached
[ -- 2th time --]
Device pci_0000_00_19_0 re-attached
[ -- 3th time --]
Device pci_0000_00_19_0 re-attached
[ -- 4th time --]
Device pci_0000_00_19_0 re-attached
[ -- 5th time --]
Device pci_0000_00_19_0 re-attached
clean up
Domain test destroyed
Device pci_0000_00_19_0 re-attached
The patch also fixes another problem, there won't be error like
"qemuDomainReAttachHostdevDevices: Not reattaching active
device 0000:00:19.0" in daemon log if some device is in active.
As pciResetDevice and pciReattachDevice won't be called for
the device anymore. This is sensible as we already reported
error when preparing the device if it's active. Blindly trying
to pciResetDevice & pciReattachDevice on the device and getting
an error is just redundant.
This patch fixes two problems:
1) The device will be reattached to host even if it's not
managed, as there is a "pciDeviceSetManaged".
2) The device won't be reattached to host with original
driver properly. As it doesn't honor the device original
properties which are maintained by driver->activePciHostdevs.
This chunk of code below repeated in several functions, factor it into
a helper method virDomainLiveConfigHelperMethod to eliminate duplicated code
based on Eric and Adam's suggestion. I have tested it for all the
relevant APIs changed.
Signed-off-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Lei Li <lilei@linux.vnet.ibm.com>
When destroying a domain qemuDomainDestroy kills its qemu process and
starts a new job, which means it unlocks the domain object and locks it
again after some time. Although the object is usually unlocked for a
pretty short time, chances are another thread processing an EOF event on
qemu monitor is able to lock the object first and does all the cleanup
by itself. This leads to wrong shutoff reason and lifecycle event detail
and virDomainDestroy API incorrectly reporting failure to destroy an
inactive domain.
Reported by Charlie Smurthwaite.
Currently qemuDomainAssignPCIAddresses() is called to assign addresses
to PCI devices.
We need to do something similar for devices with spapr-vio addresses.
So create one place where address assignment will be done, that is
qemuDomainAssignAddresses().
Signed-off-by: Michael Ellerman <michael@ellerman.id.au>
For the PPC64 pseries machine type we need to add address information
for the spapr-vty device.
Signed-off-by: Michael Ellerman <michael@ellerman.id.au>
When parsing ppc64 models on an x86 host an out-of-memory error message is displayed due
to it checking for retcpus being NULL. Fix this by removing the check whether retcpus is NULL
since we will realloc into this variable.
Also in the X86 model parser display the OOM error at the location where it happens.
A preparatory patch for DHCP snooping where we want to be able to
differentiate between a VM's interface using the tuple of
<VM UUID, Interface MAC address>. We assume that MAC addresses could
possibly be re-used between different networks (VLANs) thus do not only
want to rely on the MAC address to identify an interface.
At the current 'final destination' in virNWFilterInstantiate I am leaving
the vmuuid parameter as ATTRIBUTE_UNUSED until the DHCP snooping patches arrive.
(we may not post the DHCP snooping patches for 0.9.9, though)
Mostly this is a pretty trivial patch. On the lowest layers, in lxc_driver
and uml_conf, I am passing the virDomainDefPtr around until I am passing
only the VM's uuid into the NWFilter calls.
This patch cleans up return codes in the nwfilter subsystem.
Some functions in nwfilter_conf.c (validators and formatters) are
keeping their bool return for now and I am converting their return
code to true/false.
All other functions now have failure return codes of -1 and success
of 0.
[I searched for all occurences of ' 1;' and checked all 'if ' and
adapted where needed. After that I did a grep for 'NWFilter' in the source
tree.]
assumptions from generic code.
This implements the minimal set of changes needed in libvirt to launch a
PowerPC-KVM based guest.
It removes x86-specific assumptions about choice of serial driver backend
from generic qemu guest commandline generation code.
It also restricts the ACPI capability to be available for an x86 or
x86_64 domain.
This is not a complete solution -- it still does not guarantee libvirt
the capability to flag non-supported options in guest XML. (Eg, an ACPI
specification in a PowerPC guest XML will still get processed, even
though qemu-system-ppc64 does not support it while qemu-system-x86_64 does.)
This drawback exists because libvirt falls back on qemu to query supported
features, and qemu '-h' blindly lists all capabilities -- irrespective
of whether they are available while emulating a given architecture or not.
The long-term solution would be for qemu to list out capabilities based
on architecture and platform -- so that libvirt can cleanly make out what
devices are supported on an arch (say 'ppc64') and platform (say, 'mac99').
Signed-off-by: Prerna Saxena <prerna@linux.vnet.ibm.com>
This enables libvirt to select the correct qemu binary (qemu-system-ppc64)
for a guest vm based on arch 'ppc64'.
Also, libvirt is enabled to correctly parse the list of supported PowerPC
CPUs, generated by running 'qemu-system-ppc64 -cpu ?'
Signed-off-by: Prerna Saxena <prerna@linux.vnet.ibm.com>
Acked-by: Stefan Berger <stefanb@linux.vnet.ibm.com>
With security_driver set to "none" in /etc/libvirt/qemu.conf,
libvirtd would crash when attempted to attach to an existing
qemu process. Only copy the security model if it actually exists.
During virDomainDestroy, QEMU may emit SHUTDOWN event as a response to
SIGTERM and since domain object is still locked, the event is processed
after the domain is destroyed. We need to ignore this event in such case
to avoid changing domain state from shutoff to shutdown.
When QEMU guest finishes its shutdown sequence, qemu stops virtual CPUs
and when started with -no-shutdown waits for us to kill it using
SGITERM. Since QEMU is flushing its internal buffers, some time may pass
before QEMU actually dies. We mistakenly used "paused" state (and
events) for this which is quite confusing since users may see a domain
going to pause while they expect it to shutdown. Since we already have
"shutdown" state with "the domain is being shut down" semantics, we
should use it for this state.
However, the state didn't have a corresponding event so I created one
and called its detail as VIR_DOMAIN_EVENT_SHUTDOWN_FINISHED (guest OS
finished its shutdown sequence) with the intent to add
VIR_DOMAIN_EVENT_SHUTDOWN_STARTED in the future if we have a
sufficiently capable guest agent that can notify us when guest OS starts
to shutdown.
Fix a logic error, the initial value of ret = -1, if just set --config,
it will goto endjob directly without doing its really job here.
Signed-off-by: Lei Li <lilei@linux.vnet.ibm.com>
