This patch is based on an earlier patch by Eric Blake which was never
committed:
https://www.redhat.com/archives/libvir-list/2011-November/msg00243.html
Aside from rebasing, this patch only drops the driver lock once (prior
to the first time the function sleeps), then leaves it dropped until
it returns (Eric's patch would drop and re-acquire the lock around
each call to sleep).
At the time Eric sent his patch, the response (from Dan Berrange) was
that, while it wasn't a good thing to be holding the driver lock while
sleeping, we really need to rethink locking wrt the driver object,
switching to a finer-grained approach that locks individual items
within the driver object separately to allow for greater concurrency.
This is a good plan, and at the time it made sense to not apply the
patch because there was no known bug related to the driver lock being
held in this function.
However, we now know that the length of the wait in qemuProcessKill is
sometimes too short to allow the qemu process to fully flush its disk
cache before SIGKILL is sent, so we need to lengthen the timeout (in
order to improve the situation with management applications until they
can be updated to use the new VIR_DOMAIN_DESTROY_GRACEFUL flag added
in commit 72f8a7f19753506ed957b78ad800c0f3892c9304). But, if we
lengthen the timeout, we also lengthen the amount of time that all
other threads in libvirtd are essentially blocked from doing anything
(since just about everything needs to acquire the driver lock, if only
for long enough to get a pointer to a domain).
The solution is to modify qemuProcessKill to drop the driver lock
while sleeping, as proposed in Eric's patch. Then we can increase the
timeout with a clear conscience, and thus at least lower the chances
that someone running with existing management software will suffer the
consequence's of qemu's disk cache not being flushed.
In the meantime, we still should work on Dan's proposal to make
locking within the driver object more fine grained.
(NB: although I couldn't find any instance where qemuProcessKill() was
called with no jobs active for the domain (or some other guarantee
that the current thread had at least one refcount on the domain
object), this patch still follows Eric's method of temporarily adding
a ref prior to unlocking the domain object, because I couldn't
convince myself 100% that this was the case.)
In the future (my next patch in fact) we may want to make
decisions depending on qemu having a monitor command or not.
Therefore, we want to set qemuCaps flag instead of querying
on the monitor each time we are about to make that decision.
When blkdeviotune was first committed in 0.9.8, we had the limitation
that setting one value reset all others. But bytes and iops should
be relatively independent. Furthermore, setting tuning values on
a live domain followed by dumpxml did not output the new settings.
* src/qemu/qemu_driver.c (qemuDiskPathToAlias): Add parameter, and
update callers.
(qemuDomainSetBlockIoTune): Don't lose previous unrelated
settings. Make live changes reflect to dumpxml output.
* tools/virsh.pod (blkdeviotune): Update documentation.
The auto-generated WWN comply with the new addressing schema of WWN:
<quote>
the first nibble is either hex 5 or 6 followed by a 3-byte vendor
identifier and 36 bits for a vendor-specified serial number.
</quote>
We choose hex 5 for the first nibble. And for the 3-bytes vendor ID,
we uses the OUI according to underlying hypervisor type, (invoking
virConnectGetType to get the virt type). e.g. If virConnectGetType
returns "QEMU", we use Qumranet's OUI (00:1A:4A), if returns
ESX|VMWARE, we use VMWARE's OUI (00:05:69). Currently it only
supports qemu|xen|libxl|xenapi|hyperv|esx|vmware drivers. The last
36 bits are auto-generated.
Some tools, such as virt-manager, prefers having the default USB
controller explicit in the XML document. This patch makes sure there
is one. With this patch, it is now possible to switch from USB1 to
USB2 from the release 0.9.1 of virt-manager.
Fix tests to pass with this change.
virsh blkiotune dom --device-weights /dev/sda,400 --config
wasn't working correctly.
* src/qemu/qemu_driver.c (qemuDomainSetBlkioParameters): Use
correct definition.
Reported by Alex Jia:
==21503== 112 (32 direct, 80 indirect) bytes in 1 blocks are
definitely lost in loss record 37 of 40
==21503== at 0x4A04A28: calloc (vg_replace_malloc.c:467)
==21503== by 0x4A8991: virAlloc (memory.c:101)
==21503== by 0x505A6C: x86DataCopy (cpu_x86.c:247)
==21503== by 0x507B34: x86Compute (cpu_x86.c:1225)
==21503== by 0x43103C: qemuBuildCommandLine (qemu_command.c:3561)
==21503== by 0x41C9F7: testCompareXMLToArgvHelper
(qemuxml2argvtest.c:183)
==21503== by 0x41E10D: virtTestRun (testutils.c:141)
==21503== by 0x41B942: mymain (qemuxml2argvtest.c:705)
==21503== by 0x41D7E7: virtTestMain (testutils.c:696)
Qemu uses non-blocking I/O which doesn't play nice with regular file
descriptors. We need to pass a pipe to qemu instead, which can easily be
done using iohelper.
virFileDirectFd was used for accessing files opened with O_DIRECT using
libvirt_iohelper. We will want to use the helper for accessing files
regardless on O_DIRECT and thus virFileDirectFd was generalized and
renamed to virFileWrapperFd.
Calling qemuDomainMigrateGraphicsRelocate notifies spice clients to
connect to destination qemu so that they can seamlessly switch streams
once migration is done. Unfortunately, current qemu is not able to
accept any connections while incoming migration connection is open.
Thus, we need to delay opening the migration connection to the point
spice client is already connected to the destination qemu.
This eliminates the warning message reported in:
https://bugzilla.redhat.com/show_bug.cgi?id=624447
It was caused by a failure to open an image file that is not
accessible by root (the uid libvirtd is running as) because it's on a
root-squash NFS share, owned by a different user, with permissions of
660 (or maybe 600).
The solution is to use virFileOpenAs() rather than open(). The
codepath that generates the error is during qemuSetupDiskCGroup(), but
the actual open() is in a lower-level generic function called from
many places (virDomainDiskDefForeachPath), so some other pieces of the
code were touched just to add dummy (or possibly useful) uid and gid
arguments.
Eliminating this warning message has the nice side effect that the
requested operation may even succeed (which in this case isn't
necessary, but shouldn't hurt anything either).
virFileOpenAs previously would only try opening a file as the current
user, or as a different user, but wouldn't try both methods in a
single call. This made it cumbersome to use as a replacement for
open(2). Additionally, it had a lot of historical baggage that led to
it being difficult to understand.
This patch refactors virFileOpenAs in the following ways:
* reorganize the code so that everything dealing with both the parent
and child sides of the "fork+setuid+setgid+open" method are in a
separate function. This makes the public function easier to understand.
* Allow a single call to virFileOpenAs() to first attempt the open as
the current user, and if that fails to automatically re-try after
doing fork+setuid (if deemed appropriate, i.e. errno indicates it
would now be successful, and the file is on a networkFS). This makes
it possible (in many, but possibly not all, cases) to drop-in
virFileOpenAs() as a replacement for open(2).
(NB: currently qemuOpenFile() calls virFileOpenAs() twice, once
without forking, then again with forking. That unfortunately can't
be changed without at least some discussion of the ramifications,
because the requested file permissions are different in each case,
which is something that a single call to virFileOpenAs() can't deal
with.)
* Add a flag so that any fchown() of the file to a different uid:gid
is explicitly requested when the function is called, rather than it
being implied by the presence of the O_CREAT flag. This just makes
for less subtle surprises to consumers. (Commit
b1643dc15c5de886fefe56ad18608d65f1325a2c added the check for O_CREAT
before forcing ownership. This patch just makes that restriction
more explicit.)
* If either the uid or gid is specified as "-1", virFileOpenAs will
interpret this to mean "the current [gu]id".
All current consumers of virFileOpenAs should retain their present
behavior (after a few minor changes to their setup code and
arguments).
When libvirt's virDomainDestroy API is shutting down the qemu process,
it first sends SIGTERM, then waits for 1.6 seconds and, if it sees the
process still there, sends a SIGKILL.
There have been reports that this behavior can lead to data loss
because the guest running in qemu doesn't have time to flush its disk
cache buffers before it's unceremoniously whacked.
This patch maintains that default behavior, but provides a new flag
VIR_DOMAIN_DESTROY_GRACEFUL to alter the behavior. If this flag is set
in the call to virDomainDestroyFlags, SIGKILL will never be sent to
the qemu process; instead, if the timeout is reached and the qemu
process still exists, virDomainDestroy will return an error.
Once this patch is in, the recommended method for applications to call
virDomainDestroyFlags will be with VIR_DOMAIN_DESTROY_GRACEFUL
included. If that fails, then the application can decide if and when
to call virDomainDestroyFlags again without
VIR_DOMAIN_DESTROY_GRACEFUL (to force the issue with SIGKILL).
(Note that this does not address the issue of existing applications
that have not yet been modified to use VIR_DOMAIN_DESTROY_GRACEFUL.
That is a separate patch.)
Curently security labels can be of type 'dynamic' or 'static'.
If no security label is given, then 'dynamic' is assumed. The
current code takes advantage of this default, and avoids even
saving <seclabel> elements with type='dynamic' to disk. This
means if you temporarily change security driver, the guests
can all still start.
With the introduction of sVirt to LXC though, there needs to be
a new default of 'none' to allow unconfined LXC containers.
This patch introduces two new security label types
- default: the host configuration decides whether to run the
guest with type 'none' or 'dynamic' at guest start
- none: the guest will run unconfined by security policy
The 'none' label type will obviously be undesirable for some
deployments, so a new qemu.conf option allows a host admin to
mandate confined guests. It is also possible to turn off default
confinement
security_default_confined = 1|0 (default == 1)
security_require_confined = 1|0 (default == 0)
* src/conf/domain_conf.c, src/conf/domain_conf.h: Add new
seclabel types
* src/security/security_manager.c, src/security/security_manager.h:
Set default sec label types
* src/security/security_selinux.c: Handle 'none' seclabel type
* src/qemu/qemu.conf, src/qemu/qemu_conf.c, src/qemu/qemu_conf.h,
src/qemu/libvirtd_qemu.aug: New security config options
* src/qemu/qemu_driver.c: Tell security driver about default
config
This is a trivial implementation, which works with the current
released qemu 1.0 with backports of preliminary block pull but
no partial rebase. Future patches will update the monitor handling
to support an optional parameter for partial rebase; but as qemu
1.1 is unreleased, it can be in later patches, designed to be
backported on top of the supported API.
* src/qemu/qemu_driver.c (qemuDomainBlockJobImpl): Add parameter,
and adjust callers. Drop redundant check.
(qemuDomainBlockPull): Move guts...
(qemuDomainBlockRebase): ...to new function.
This patch adds support for the new api into the qemu driver to support
modification and retrieval of domain description and title. This patch
does not add support for modifying the <metadata> element.
In qemuDomainShutdownFlags if we try to use guest agent,
which has error or is not configured, we jump go endjob
label even if we haven't started any job yet. This may
lead to the daemon crash:
1) virsh shutdown --mode agent on a domain without agent configured
2) wait until domain quits
3) virsh edit
Fix my typo at
commit 74e034964c32edb1732d0ff7642f3977f3587d72
"disk->rawio == -1" indicates that this value is not
specified. So in case of this, domain must not
be tainted.
Signed-off-by: Taku Izumi <izumi.taku@jp.fujitsu.com>
This patch revises qemuProcessStart() function for qemu
processes to retain CAP_SYS_RAWIO if needed.
And in case of that, add taint flag to domain.
Signed-off-by: Taku Izumi <izumi.taku@jp.fujitsu.com>
Signed-off-by: Shota Hirae <m11g1401@hibikino.ne.jp>
This patch addresses: https://bugzilla.redhat.com/show_bug.cgi?id=781562
Along with the "rombar" option that controls whether or not a boot rom
is made visible to the guest, qemu also has a "romfile" option that
allows specifying a binary file to present as the ROM BIOS of any
emulated or passthrough PCI device. This patch adds support for
specifying romfile to both passthrough PCI devices, and emulated
network devices that attach to the guest's PCI bus (just about
everything other than ne2k_isa).
One example of the usefulness of this option is described in the
bugzilla report: 82576 sriov network adapters don't provide a ROM BIOS
for the cards virtual functions (VF), but an image of such a ROM is
available, and with this ROM visible to the guest, it can PXE boot.
In libvirt's xml, the new option is configured like this:
<hostdev>
...
<rom file='/etc/fake/boot.bin'/>
...
</hostdev
(similarly for <interface>).
When support for the rombar option was added, it was only added for
PCI passthrough devices, configured with <hostdev>. The same option is
available for any network device that is attached to the guest's PCI
bus. This patch allows setting rombar for any PCI network device type.
After adding cases to test this to qemuxml2argv-hostdev-pci-rombar.*,
I decided to rename those files (to qemuxml2argv-pci-rom.*) to more
accurately reflect the additional tests, and also noticed that up to
now we've only been performing a domainschematest for that case, so I
added the "pci-rom" test to both qemuxml2argv and qemuxml2xml (and in
the process found some bugs whose fixes I squashed into previous
commits of this series).
To help consolidate the commonality between virDomainHostdevDef and
virDomainNetDef into as few members as possible (and because I
think it makes sense), this patch moves the rombar and bootIndex
members into the "info" member that is common to both (and to all the
other structs that use them).
It's a bit problematic that this gives rombar and bootIndex to many
device types that don't use them, but this is already the case for the
master and mastertype members of virDomainDeviceInfo, and is properly
commented as such in the definition.
Note that this opens the door to supporting rombar for other devices
that are attached to the guest PCI bus - virtio-blk-pci,
virtio-net-pci, various other network adapters - which which have that
capability in qemu, but previously had no support in libvirt.
If yajl was not compiled in, we end up freeing an incoming
parameter, which leads to a bogus free later on. Regression
introduced in commit 6e769eb.
* src/qemu/qemu_capabilities.c (qemuCapsParseHelpStr): Avoid alloc
on failure path, which in turn fixes bogus free.
Reported by Cole Robinson.
QEMU supports a bunch of CPUID features that are tied to the kvm CPUID
nodes rather than the processor's. They are "kvmclock",
"kvm_nopiodelay", "kvm_mmu", "kvm_asyncpf". These are not known to
libvirt and their CPUID leaf might move if (for example) the Hyper-V
extensions are enabled. Hence their handling would anyway require some
special-casing.
However, among these the most useful is kvmclock; an additional
"property" of this feature is that a <timer> element is a better model
than a CPUID feature. Although, creating part of the -cpu command-line
from something other than the <cpu> XML element introduces some
ugliness.
Reviewed-by: Jiri Denemark <jdenemar@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Avoid creating an empty <cpu> element when the QEMU command-line simply
specifies the default "-cpu qemu32" or "-cpu qemu64".
This requires the previous patch, which lets us represent "-cpu qemu32"
as <os arch='i686'> in the generated XML.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The qemu32 CPU model is chosen based on the <os arch=...> name when
creating the QEMU command line for a 64-bit host. For the opposite
transformation we can test the guest CPU model for the "lm" feature.
If it is absent, def->os.arch needs to be corrected.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
When running under KVM, the arch is usually set to i686 because
the name of the emulator is not qemu-system-x86_64. Use the host
arch instead.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The qemu developers have made it clear that modern qemu will no
longer guarantee human monitor command stability; furthermore,
some features, such as async events, are only supported via qmp.
If we are compiled without support for handling JSON, we cannot
expect to sanely interact with modern qemu.
However, things must continue to build on RHEL 5, where qemu
is stuck at 0.10, and where yajl is not available.
Another benefit of this patch: future additions of new monitor
commands need only focus on qemu_monitor_json.c, instead of
also wasting time with qemu_monitor_text.c.
* src/qemu/qemu_capabilities.c (qemuCapsComputeCmdFlags): Report
error if yajl is missing but qemu requires qmp.
(qemuCapsParseHelpStr): Propagate error.
(qemuCapsExtractVersionInfo): Update caller.
* tests/qemuhelptest.c (testHelpStrParsing): Likewise.
I'm getting tired of remembering to backport RHEL-specific
patches when building upstream libvirt on RHEL 6.x or CentOS.
All the affected versions of RHEL qemu-kvm have backported
enough patches to a) make JSON useful, and b) modify the
-help text to mention libvirt as the preferred interface;
which means this string in the help output is a reliable
indicator that we can outsmart a strict version check,
even when upstream qemu 0.12 lacked the needed features.
* src/qemu/qemu_capabilities.c (qemuCapsComputeCmdFlags):
Recognize particular help string present when enough features were
backported to be worth using JSON.
* tests/qemuhelptest.c (mymain): Update tests accordingly.
QEMU always sends details about all available block devices as an answer
for "info block"/"query-block" command. On the other hand, our
qemuMonitorGetBlockInfo was made for a single block devices queries
only. Thus, when asking for multiple devices, we asked qemu multiple
times to always get the same answer from which different parts were
filtered. This patch makes qemuMonitorGetBlockInfo return a hash table
of all block devices, which may later be used for getting details about
specific devices.
In preparation for the patch to include Murmurhash3, which
introduces a virhashcode.h and virhashcode.c files, rename
the existing hash.h and hash.c to virhash.h and virhash.c
respectively.
Direct boot (using kernel, initrd, and command line) is used by
virt-install/virt-manager for network install. While any bootindex has
no direct effect since -kernel is always first, we need it as a hint for
SeaBIOS to present disks in the same order as they will be presented
during normal boot.
This makes use of the QEMU guest agent to implement the
virDomainShutdownFlags and virDomainReboot APIs. With
no flags specified, it will prefer to use the agent, but
fallback to ACPI. Explicit choice can be made by using
a suitable flag
* src/qemu/qemu_driver.c: Wire up use of agent
There is now a standard QEMU guest agent that can be installed
and given a virtio serial channel
<channel type='unix'>
<source mode='bind' path='/var/lib/libvirt/qemu/f16x86_64.agent'/>
<target type='virtio' name='org.qemu.guest_agent.0'/>
</channel>
The protocol that runs over the guest agent is JSON based and
very similar to the JSON monitor. We can't use exactly the same
code because there are some odd differences in the way messages
and errors are structured. The qemu_agent.c file is based on
a combination and simplification of qemu_monitor.c and
qemu_monitor_json.c
* src/qemu/qemu_agent.c, src/qemu/qemu_agent.h: Support for
talking to the agent for shutdown
* src/qemu/qemu_domain.c, src/qemu/qemu_domain.h: Add thread
helpers for talking to the agent
* src/qemu/qemu_process.c: Connect to agent whenever starting
a guest
* src/qemu/qemu_monitor_json.c: Make variable static