The current code makes a poor effort at updating the device arrays after
hot-unplug. Fix that and combine the two code paths into one.
* src/qemu_driver.c: fix list updating in qemudDomainDetachNetDevice(),
qemudDomainDetachPciDiskDevice() and qemudDomainDetachHostPciDevice()
Implement a compressed save image format for qemu. While ideally
we would have the choice between compressed/non-compressed
available to the libvirt API, unfortunately there is no "flags"
parameter to the virDomainSave() API. Therefore, implement this
as a qemu.conf option. gzip, bzip2, and lzma are implemented, and
it should be very easy to implement additional compression
methods.
One open question is if/how we should detect the compression
binaries. One way to do it is to do compile-time setting of the
paths (via configure.in), but that doesn't seem like a great thing
to do. My preferred solution is not to detect at all;
when we go to run the commands that need them, if they
aren't available, or aren't available in one of the standard paths,
then we'll fail. That's also the solution implemented in this patch.
In the future, we'll have a more robust (managed) save/restore API,
at which time we can expose this functionality properly in the API.
V2: get rid of redundant dd command and just use >> to append data.
V3: Add back the missing pieces for the enum and bumping the save version.
V4: Make the compressed field in the save_header an int.
Implement LZMA compression.
Signed-off-by: Chris Lalancette <clalance@redhat.com>
If a PCI device reset causes other devices to be reset, allow it so long
as those other devices are note assigned to another active domain.
Note, we need to take the driver lock qemudNodeDeviceReset() because the
check function will iterate over the domain list.
* src/qemu_conf.c: add qemuCheckPciHostDevice() to iterate over active
domains checking whether the affected device is assigned
* src/pci.[ch]: add pciDeviceEquals() helper
When using a Secondary Bus Reset, all devices on the bus are reset.
Extend the pciResetDevice() API so that a 'check' callback can be
supplied which will verify that it is safe to reset the other devices
on the bus.
The virDomainObjPtr parameter is needed so that when the check function
iterates over the domain list, it can avoid double locking.
* src/pci.[ch]: add a 'check' callback to pciResetDevice(), re-work
pciIterDevices() to pass the check function to the iter functions,
use the check function in the bus iterator, return the first unsafe
device from pciBusCheckOtherDevices() and include its details in
the bus reset error message.
* src/qemu_driver.c, src/xen_uninified.c: just pass NULL as the
check function for now
Currently, if we are unable to reset a PCI device we return a fairly
generic 'No PCI reset capability available' error message.
Fix that by returning an error from the individual reset messages and
using that error to construct the higher level error mesage.
* src/pci.c: set errors in pciTryPowerManagementReset() and
pciTrySecondaryBusReset() on failure; use those error messages
in pciResetDevice(), or explain that no reset support is available
When the guest shuts down, we should attempt to restore all PCI host
devices to a sane state.
In the case of managed hostdevs, we should reset and re-attach the
devices. In the case of unmanaged hostdevs, we should just reset them.
Note, KVM will already reset assigned devices when the guest shuts
down using whatever means it can, so we are only doing it to cover the
cases the kernel can't handle.
* src/qemu_driver.c: add qemuDomainReAttachHostDevices() and call
it from qemudShutdownVMDaemon()
Attaching a host PCI device to a qemu guest is done with a
straightforward 'pci_add auto host host=XX:XX.X' command.
Like with NIC and disk hotplug, we need to retain the guest PCI address
assigned by qemu so that we can use it for hot-unplug.
Identifying a device for detach is done using the host PCI address.
Managed mode is handled by detaching/resetting the device before
attaching it to the guest and re-attaching it after detaching it from
the guest.
* src/qemu_driver.c: add qemudDomainAttachHostPciDevice() and
qemudDomainDetachHostPciDevice()
* src/domain_conf.h: add somewhere to store the guest PCI address
* src/domain_conf.c: handle formatting and parsing the guest PCI
address
Re-factor the hostdev hotplug code so that we can easily add PCI
hostdev hotplug to qemudDomainAttachHostDevice().
* src/qemu_driver.c: rename qemudDomainAttachHostDevice() to
qemudDomainAttachHostUsbDevice(); make qemudDomainAttachHostDevice()
handle all hostdev types
* src/libvirt_private.syms: export a couple of hostdev related
ToString() functions
As of qemu 0.10.6, qemu now honors the -S flag on incoming migration.
That means that when the migration completes, we have to issue a
'cont' command to get the VM running again. We do it unconditionally
since it won't hurt on older qemu.
Signed-off-by: Chris Lalancette <clalance@redhat.com>
When doing a restore, we were forgetting to update the state file
for the VM. That means that if you do a save/restore, then shut
down libvirtd, then start it back up, you'll see the state of the
guest as "paused", even though it is really running. We were
just forgetting a "virDomainSaveStatus" call in the restor path.
Signed-off-by: Chris Lalancette <clalance@redhat.com>
If we're running qemu unprivileged, we need to chown any supplied kernel
or initrd before spawning it.
* src/qemu_driver.c: rename qemuDomainSetDiskOwnership() to
qemuDomainSetFileOwnership(), pass it a path string instead of a disk
definition and use it for chowning the kernel/initrd in
qemuDomainSetAllDeviceOwnership()
Fix up qemudDomainMigratePrepare2 to use virGetHostname instead of
gethostname. Besides the fact that virGetHostname is far more clever,
there was a latent bug in the handling that could cause a buffer overflow
on a very long hostname.
Signed-off-by: Chris Lalancette <clalance@redhat.com>
* src/util.c: Don't drop capabilities until after the PID file has
been written. Kill off child if writing the PID file fails
* src/qemu_driver.c: Remove bogus trailing '/' in state dir
* src/qemu_driver.c: fix qemudOpenMonitorUnix() to retry on ENOENT
instead of EACCES which is the error one receive when the socket
error hasn't shown up yet
In order to hotplug a network/bridge backed NIC, we need to first create
the tap file descriptor, add the tap interface to the bridge and then
pass the file descriptor to the qemu process using the 'getfd' monitor
command.
Once the tapfd has been accepted, we create the network backend using
host_net_add, supplying the name assigned to the tapfd. If this fails,
we need to close the tapfd in qemu using the 'closefd' monitor command.
If the version of qemu does not support the getfd/closefd monitor
commands we detect "unknown command" in the getfd reply and fail the
attach operation.
* src/qemu_driver.c: add support for tapfd based hotplug in
qemudDomainAttachNetDevice()
Add qemudMonitorCommandWithFd() which allows a file descriptor to be
sent to qemu over a unix monitor socket using SCM_RIGHTS. See the
unix(7) and cmsg(3) man pages.
* src/qemu_conf.c: add a scm_fd param to qemudMonitorCommandExtra(),
add qemudMonitorCommandWithFd(), implement SCM_RIGHTS support in
qemudMonitorSendUnix()
Switch from using write() to using sendmsg() on QEMU's monitor socket
so that we can add support for SCM_RIGHTS.
* src/qemu_driver.c: add sendmsg() based qemudMonitorSendUnix() and use
it when the monitor fd is a unix socket
Add a little helper function to write the monitor command followed by
carriage return in a single write.
This doesn't make any real difference, but allows us to more easily
switch to using sendmsg() when using the monitor over a unix socket.
* src/qemu_conf.c: split qemudMonitorSend() out
In subsequent patches we're going to have a file descriptor to close
too, so centralize the error handling cleanups to make things easier.
* src/qemu_conf.c: in qemudDomainAttachNetDevice() consolidate the
error handling cleanups together
With hotplug, we're going to want to pass a tapfd name rather than an
actual file descriptor, so prepare the way by passing a string tapfd to
qemuBuildHostNetStr().
* src/qemu_conf.h: qemuBuildHostNetStr() takes a string tapfd now
* src/qemu_conf.c: pass qemuBuildHostNetStr() a string rather than an
actual file descriptor
* src/qemu_driver.c: update qemudDomainAttachNetDevice() for change
By probing for qemu machine types, we increased the time of a
GetCapabilities call from 100us to a whopping 60ms.
This patch takes the approach of only probing for machine types
when the mtime of the emulator binary changed since the last time
the capabilities were generated.
* src/capabilities.h: cache the emulator binary mtime
* src/qemu_conf.c: add qemudGetOldMachines() to copy the machine
types from the old caps struct if the mtime for the binary hasn't
changed
* src/qemu_conf.h, src/qemu_driver.c: pass the old caps pointer to
qemudCapsInit()
Not all possible emulators are actually in the capabilities, so if we
don't find the supplied emulator we should probe it directly for machine
types.
* src/qemu_driver.c: add qemudCanonicalizeMachineDirect() to directly
probe an emulator for the canonical machine type
In qemu-0.11 there is a 'pc-0.10' machine type which allows you to run
guests with a machine which is compatible with the pc machine in
qemu-0.10 - e.g. using the original PCI class for virtio-blk and
virtio-console and disabling MSI support in virtio-net. The idea here
is that we don't want to suprise guests by changing the hardware when
qemu is updated.
I've just posted some patches for qemu-0.11 which allows libvirt to
canonicalize the 'pc' machine alias to the latest machine version.
This patches makes us use that so that when a guest is configured to
use the 'pc' machine type, we resolve that to 'pc-0.11' machine and
save that in the guest XML.
See also:
https://fedoraproject.org/wiki/Features/KVM_Stable_Guest_ABI
* src/qemu_conf.c: add qemudCanonicalizeMachine() to canonicalize
the machine type according to the machine aliases in capabilities
* src/qemu_driver.c: parse aliases in qemudParseMachineTypesStr()
Do the check in libvirt.c, to save drivers from the burden. This changes
behavior slightly in the qemu driver: we no longer explictly error if
passed an empty string. An error will still be thrown when the device
lookup fails.
* qemud/libvirtd_qemu.aug, qemud/test_libvirtd_qemu.aug,
src/qemu.conf: Add 'cgroups_controllers' and 'cgroups_device_acl'
parameters
* src/qemu_conf.h, src/qemu_conf.c: Load & parse configuration params
for cgroups
* src/qemu_driver.c: Only use cgroups controllers that are activated,
and use configured device whitelist instead of default, if set.
* src/qemu_driver.c: Set a restrictive block device whitelist for
all QEMU guests. Update whitelist when hotplugging disks.
* src/cgroup.h, src/cgroup.c: Add some more convenience methods
for dealing with block device whitelists.
* src/qemu_driver.c: Add driver methods qemuGetSchedulerType,
qemuGetSchedulerParameters, qemuSetSchedulerParameters
* src/lxc_driver.c: Fix to use unsigned long long consistently
for schedular parameters
* src/cgroup.h, src/cgroup.c: Fix cpu_shares to take unsigned
long long
* src/util.c, src/util.h, src/libvirt_private.syms: Add a
virStrToDouble helper
* src/virsh.c: Fix handling of --set arg to schedinfo command
to honour the designated data type of each schedular tunable
as declared by the driver
* src/datatypes.c src/domain_conf.c src/interface_conf.c
src/lxc_driver.c src/qemu_driver.c src/storage_backend.c src/virsh.c:
add bare %s format string to printf-derivatives called with no format
string
* include/libvirt/libvirt.h include/libvirt/libvirt.h.in: adds the new
flag VIR_MEMORY_PHYSICAL for virDomainMemoryPeek
* src/libvirt.c: update the front-end checking
* src/qemu_driver.c: extend the QEmu driver
qemu network devices are hot-unplugged in two stages - first the PCI NIC
is removed using 'pci_del <pci_addr>' and then the backend is removed
using 'host_net_remove <vlan> <name>'.
In order to perform these operations we need to have retained the
PCI address, backend name and vlan number.
* src/qemu_driver.c: add qemudDomainDetachNetDevice()
When we pci_add a NIC, we need to retain the PCI address assigned by
qemu for using during detach.
* src/qemu_driver.c: use qemudParsePciAddReply() to pull the PCI
address from the pci_add reply
* src/domain_conf.c: handle storing and parsing the PCI address in the
domain state XML file
The current code for parsing pci_add replies ignores the the domain and
bus numbers. Re-write the code to rectify that.
Also, since pci_add is used for NIC hotplug as well ask disk hotplug,
re-factor the code into a separate function.
* src/qemu_driver.c: add qemudParsePciAddReply() function which can
handle parsing domain and bus numbers
If we fail to pci_add a NIC, we should remove the network backend and
leave things the way we found them. To do that, we pre-allocate a
host_net_remove monitor command and issue that if the pci_add fails.
If the remove fails, we just log a warning.
We can only do this if we have a name for the network backend and
we know the vlan number its associated with.
* src/qemu_driver.c: host_net_remove the network backend if the
pci_add fails
Implement basic NIC hotplug support using the 'host_net_add' and
'pci_add' qemu monitor commands.
For now, we don't support 'bridge' or 'network' types.
Also, if pci_add fails, we currently fail to remove the backend
which we added.
Finally, NIC hot-unplug support is missing.
* src/qemu_driver.c: add qemudDomainAttachNetDevice()
* src/qemu_conf.[ch]: export qemuBuildNicStr(), qemuBuildHostNetStr()
and qemuAssignNames()
* src/libvirt_private.syms: export virDomainNetTypeToString()
qemudDomainChangeEjectableMedia() currently extracts the qemu command
line flags, but other device attaching code might need it, so move
the qemudExtractVersionInfo() call up a frame.
* src/qemu_driver.c: move the qemudExtractVersionInfo() call from
qemudDomainChangeEjectableMedia() to qemudDomainAttachDevice()
When we hot-plug a disk device into a qemu guest, we need to retain its
PCI address so that it can be removed again later. Currently, we do
retain the slot number, but not across libvirtd restarts.
Add <state devaddr="xxxx:xx:xx"/> to the disk device XML config when the
VIR_DOMAIN_XML_INTERNAL_STATUS flag is used. We still don't parse the
domain and bus number, but the format allows us to do that in future.
* src/domain_conf.h: replace slotnum with pci_addr struct, add helper
for testing whether the address is valid
* src/domain_conf.c: handle formatting and parsing the address
* src/qemu_driver.c: store the parsed slot number as a full PCI address,
and use this address with the pci_del monitor command
* src/vbox/vbox_tmpl.c: we're debug printing slotnum here even though
it can never be set, just delete it
Unlike the pty monitor (which we know exists since we scrape its path from
stdout), we have no way of knowing that the unix monitor socket should exist/
be initialized. As a result, some of my KVM guests randomly fail to start on
F10 host.
Try to open the unix socket in a 3 second timeout loop. Ignore EACCES (path
does not exist if a first time run) and ECONNREFUSED (leftover socket from
a previous run hasn't been removed yet). Fixes things for me.
* configure.in: Add --with-qemu-user and --with-qemu-group args
* libvirt.spec.in: use 'qemu' for user/group for Fedora >= 12
* qemud/libvirtd_qemu.arg, qemud/test_libvirtd_qemu.aug,
src/qemu.conf: Add 'user' and 'group' args for configuration
* src/Makefile.am: Create %localstatedir/cache/libvirt/qemu
* src/qemu_conf.c, src/qemu_conf.h: Load user/group from config
* src/qemu_driver.c: Change user ID/group ID when launching QEMU
guests. Change user/group ownership on disks/usb/pci devs.
Put memory dumps in %localstatedir/cache/libvirt/qemu
* src/util.c, src/util.h: Add convenient APIs for converting
username/groupname to user ID / group ID