In some cases (spotted with broken connection during tunneled migration)
we were overwriting the original error with worse or even misleading
errors generated when we were cleaning up after failed migration.
This patch modifies the CPU comparrison function to report the
incompatibilities in more detail to ease identification of problems.
* src/cpu/cpu.h:
cpuGuestData(): Add argument to return detailed error message.
* src/cpu/cpu.c:
cpuGuestData(): Add passthrough for error argument.
* src/cpu/cpu_x86.c
x86FeatureNames(): Add function to convert a CPU definition to flag
names.
x86Compute(): - Add error message parameter
- Add macro for reporting detailed error messages.
- Improve error reporting.
- Simplify calculation of forbidden flags.
x86DataIteratorInit():
x86cpuidMatchAny(): Remove functions that are no longer needed.
* src/qemu/qemu_command.c:
qemuBuildCpuArgStr(): - Modify for new function prototype
- Add detailed error reports
- Change error code on incompatible processors
to VIR_ERR_CONFIG_UNSUPPORTED instead of
internal error
* tests/cputest.c:
cpuTestGuestData(): Modify for new function prototype
Most of our errors complaining about an inability to support a
particular action due to qemu limitations used CONFIG_UNSUPPORTED,
but we had a few outliers. Reported by Jiri Denemark.
* src/qemu/qemu_command.c (qemuBuildDriveDevStr): Prefer
CONFIG_UNSUPPORTED.
* src/qemu/qemu_driver.c (qemuDomainReboot)
(qemuDomainBlockJobImpl): Likewise.
* src/qemu/qemu_hotplug.c (qemuDomainAttachPciControllerDevice):
Likewise.
* src/qemu/qemu_monitor.c (qemuMonitorTransaction)
(qemuMonitorBlockJob, qemuMonitorSystemWakeup): Likewise.
A "ide-drive" device can be either a hard disk or a CD-ROM,
if there is ",media=cdrom" specified for the backend, it's
a CD-ROM, otherwise it's a hard disk.
Upstream qemu splitted "ide-drive" into "ide-hd" and "ide-cd"
since commit 1f56e32, and ",media=cdrom" is not required for
ide-cd anymore. "ide-drive" is still supported for backwards
compatibility, but no doubt we should go foward.
A "scsi-disk" device can be either a hard disk or a CD-ROM,
if there is ",media=cdrom" specified for the backend, it's
a CD-ROM, otherwise it's a hard disk.
But upstream qemu splitted "scsi-disk" into "scsi-hd" and
"scsi-cd" since commit b443ae, and ",media=cdrom" is not
required for scsi-cd anymore. "scsi-disk" is still supported
for backwards compatibility, but no doubt we should go
foward.
If console[0] is an alias for serial[0], do not enforce the former to
have a PTY source type. This breaks serial consoles on stdio and makes
no sense.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Currently, we have 3 boolean arguments we have to pass
to qemuProcessStart(). As libvirt grows it is harder and harder
to remember them and their position. Therefore we should
switch to flags instead.
Instead of returning a CPUs list, numad returns NUMA node
list instead, this patch is to convert the node list to
cpumap before affinity setting. Otherwise, the domain
processes will be pinned only to CPU[$numa_cell_num],
which will cause significiant performance losses.
Also because numad will balance the affinity dynamically,
reflecting the cpuset from numad back doesn't make much
sense then, and it may just could produce confusion for
the users. Thus the better way is not to reflect it back
to XML. And in this case, it's better to ignore the cpuset
when parsing XML.
The codes to update the cpuset is removed in this patch
incidentally, and there will be a follow up patch to ignore
the manually specified "cpuset" if "placement" is "auto",
and document will be updated too.
This patch adds a netlink callback when migrating a VEPA enabled
virtual machine. It fixes a Bug where a VM would not request a port
association when it was cleared by lldpad.
This patch requires the latest git version of lldpad to work.
Signed-off-by: D. Herrendoerfer <d.herrendoerfer@herrendoerfer.name>
If dynamic_ownership is off and we are creating a file on NFS
we force chown. This will fail as chown/chmod are not supported
on NFS. However, with no dynamic_ownership we are not required
to do any chown.
In my testing, I was able to provoke an odd block pull failure:
$ virsh blockpull dom vda --bandwidth 10000
error: Requested operation is not valid: No active operation on device: drive-virtio-disk0
merely by using gdb to artifically wait to do the block job set speed
until after the pull had already finished. But in reality, that should
be a success, since the pull finished before we had a chance to set
speed. Furthermore, using a double job lock is not only annoying, but
a bug in itself - if you do parallel virDomainBlockRebase, and hit
the race window just right, the first call grabs the VM job to start
a fast block job, then the second call grabs the VM job to start
a long-running job with unspecified speed, then the first call finally
regrabs the VM job and sets the speed, which ends up running the
second job under the speed from the first call. By consolidating
things into a single job, we avoid opening that race, as well as reduce
the time between starting the job and changing the speed, for less
likelihood of the speed change happening after block job completion
in the first place.
* src/qemu/qemu_monitor.h (BLOCK_JOB_CMD): Add new mode.
* src/qemu/qemu_driver.c (qemuDomainBlockRebase): Move secondary
job call...
(qemuDomainBlockJobImpl): ...here, for fewer locks.
* src/qemu/qemu_monitor_json.c (qemuMonitorJSONBlockJob): Change
return value on new internal mode.
Without the VIR_DOMAIN_BLOCK_JOB_ABORT_ASYNC flag, libvirt will internally
poll using qemu's "query-block-jobs" API and will not return until the
operation has been completed. API users are advised that this operation
is unbounded and further interaction with the domain during this period
may block. Future patches may refactor things to allow other queries in
parallel with this polling. For older qemu, we synthesize the cancellation
event, since qemu won't generate it.
The choice of polling duration copies from the code in qemu_migration.c.
Signed-off-by: Adam Litke <agl@us.ibm.com>
Cc: Stefan Hajnoczi <stefanha@gmail.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
Probably in the noise, but this will let us scale more efficiently
as we learn to recognize even more qemu events.
* src/qemu/qemu_monitor_json.c (eventHandlers): Sort.
(qemuMonitorEventCompare): New helper function.
(qemuMonitorJSONIOProcessEvent): Optimize event lookup.
RHEL 6.2 was released with an early version of block jobs, which only
worked on the qed file format, where the commands were spelled with
underscore (contrary to QMP style), and where 'block_job_cancel' was
synchronous and did not trigger an event.
The upcoming qemu 1.1 release has fixed these short-comings [1][2]:
the commands now work on multiple file types, are spelled with dash,
and 'block-job-cancel' is asynchronous and emits an event upon conclusion.
[1]qemu commit 370521a1d6f5537ea7271c119f3fbb7b0fa57063
[2]https://lists.gnu.org/archive/html/qemu-devel/2012-04/msg01248.html
This patch recognizes the new spellings, and fixes virDomainBlockRebase
to give a graceful error when talking to a too-old qemu on a partial
rebase attempt. Fixes for the new semantics will come later. This
patch also removes a bogus ATTRIBUTE_NONNULL mistakenly added in
commit 10ec36e2.
* src/qemu/qemu_capabilities.h (QEMU_CAPS_BLOCKJOB_SYNC)
(QEMU_CAPS_BLOCKJOB_ASYNC): New bits.
* src/qemu/qemu_capabilities.c (qemuCaps): Name them.
* src/qemu/qemu_monitor_json.c (qemuMonitorJSONCheckCommands): Set
them.
(qemuMonitorJSONBlockJob): Manage both command names.
(qemuMonitorJSONDiskSnapshot): Minor formatting fix.
* src/qemu/qemu_monitor.h (qemuMonitorBlockJob): Alter signature.
* src/qemu/qemu_monitor_json.h (qemuMonitorJSONBlockJob): Likewise.
* src/qemu/qemu_monitor.c (qemuMonitorBlockJob): Pass through
capability bit.
* src/qemu/qemu_driver.c (qemuDomainBlockJobImpl): Update callers.
The new safe console handling introduced a possibility to deadlock the
qemu driver when a new console connection forcibly disconnects a
previous console stream that belongs to an already closed connection.
The virStreamFree function calls subsequently a the virReleaseConnect
function that tries to lock the driver while discarding the connection,
but the driver was already locked in qemuDomainOpenConsole.
Backtrace of the deadlocked thread:
0 0x00007f66e5aa7f14 in __lll_lock_wait () from /lib64/libpthread.so.0
1 0x00007f66e5aa3411 in _L_lock_500 () from /lib64/libpthread.so.0
2 0x00007f66e5aa322a in pthread_mutex_lock () from/lib64/libpthread.so.0
3 0x0000000000462bbd in qemudClose ()
4 0x00007f66e6e178eb in virReleaseConnect () from/usr/lib64/libvirt.so.0
5 0x00007f66e6e19c8c in virUnrefStream () from /usr/lib64/libvirt.so.0
6 0x00007f66e6e3d1de in virStreamFree () from /usr/lib64/libvirt.so.0
7 0x00007f66e6e09a5d in virConsoleHashEntryFree () from/usr/lib64/libvirt.so.0
8 0x00007f66e6db7282 in virHashRemoveEntry () from/usr/lib64/libvirt.so.0
9 0x00007f66e6e09c4e in virConsoleOpen () from /usr/lib64/libvirt.so.0
10 0x00000000004526e9 in qemuDomainOpenConsole ()
11 0x00007f66e6e421f1 in virDomainOpenConsole () from/usr/lib64/libvirt.so.0
12 0x00000000004361e4 in remoteDispatchDomainOpenConsoleHelper ()
13 0x00007f66e6e80375 in virNetServerProgramDispatch () from/usr/lib64/libvirt.so.0
14 0x00007f66e6e7ae11 in virNetServerHandleJob () from/usr/lib64/libvirt.so.0
15 0x00007f66e6da897d in virThreadPoolWorker () from/usr/lib64/libvirt.so.0
16 0x00007f66e6da7ff6 in virThreadHelper () from/usr/lib64/libvirt.so.0
17 0x00007f66e5aa0c5c in start_thread () from /lib64/libpthread.so.0
18 0x00007f66e57e7fcd in clone () from /lib64/libc.so.6
* src/qemu/qemu_driver.c: qemuDomainOpenConsole()
-- unlock the qemu driver right after acquiring the domain
object
In case an API fails with "cannot acquire state change lock", searching
for the API that possibly forgot to end its job is not always easy.
Let's keep track of the job owner and print it out for easier
identification.
As reported by Daniel Berrangé, we have a huge performance regression
for virDomainGetInfo() due to the change which makes virDomainEndJob()
save the XML status file every time it is called. Previous to that
change, 2000 calls to virDomainGetInfo() took ~2.5 seconds. After that
change, 2000 calls to virDomainGetInfo() take 2 *minutes* 45 secs.
We made the change to be able to recover from libvirtd restart in the
middle of a job. However, only destroy and async jobs are taken care of.
Thus it makes more sense to only save domain state XML when these jobs
are started/stopped.
* src/qemu/qemu_command.c: Wire up -bios with <loader>
* tests/qemuxml2argvdata/qemuxml2argv-bios.args,
tests/qemuxml2argvdata/qemuxml2argv-bios.xml: Expand
existing BIOS test case to cover <loader>
Leak introduced in commit 0436d32. If we allocate an actions array,
but fail early enough to never consume it with the qemu monitor
transaction call, we leaked memory.
But our semantics of making the transaction command free the caller's
memory is awkward; avoiding the memory leak requires making every
intermediate function in the call chain check for error. It is much
easier to fix things so that the function that allocates also frees,
while the call chain leaves the caller's data intact. To do that,
I had to hack our JSON data structure to make it easy to protect a
portion of an arbitrary JSON tree from being freed.
* src/util/json.h (virJSONType): Name the enum.
(_virJSONValue): New field.
* src/util/json.c (virJSONValueFree): Use it to protect a portion
of an array.
* src/qemu/qemu_monitor_json.c (qemuMonitorJSONTransaction): Avoid
freeing caller's data.
* src/qemu/qemu_driver.c (qemuDomainSnapshotCreateDiskActive):
Free actions array on failure.
We can tell qemuDomainSnapshotFSThaw if we want it to report errors or
not. However, if we don't want to and an error has been already set by
previous qemuReportError() we must keep copy of that error not just a
pointer to it. Otherwise, it get overwritten if FSThaw reports an error.
If the daemon is restarted it will lose list of active
USB devices assigned to active domains. Therefore we need
to rebuild this list on qemuProcessReconnect().
To prevent assigning one USB device to two domains,
we keep a list of assigned USB devices. On domain
startup - qemuProcessStart() - we insert devices
used by domain into the list but remove them only
on detach-device. Devices are, however, released
on qemuProcessStop() as well.
Originally, qemuDomainCheckEjectableMedia was entering monitor with qemu
driver lock. Commit 2067e31bf9, which I
made to fix that, revealed another issue we had (but didn't notice it
since the driver was locked): we didn't set nested job when
qemuDomainCheckEjectableMedia is called during migration. Thus the
original fix I made was wrong.
Since Xen 3.1 the clock=variable semantic is supported. In addition to
qemu/kvm Xen also knows about a variant where the offset is relative to
'localtime' instead of 'utc'.
Extends the libvirt structure with a flag 'basis' to specify, if the
offset is relative to 'localtime' or 'utc'.
Extends the libvirt structure with a flag 'reset' to force the reset
behaviour of 'localtime' and 'utc'; this is needed for backward
compatibility with previous versions of libvirt, since they report
incorrect XML.
Adapt the only user 'qemu' to the new name.
Extend the RelaxNG schema accordingly.
Document the new 'basis' attribute in the HTML documentation.
Adapt test for the new attribute.
Signed-off-by: Philipp Hahn <hahn@univention.de>
If we round up a user's memory request, we should update the XML
to reflect the actual value in use by the VM, rather than giving
an artificially small value back to the user.
* src/qemu/qemu_command.c (qemuBuildNumaArgStr)
(qemuBuildCommandLine): Reflect rounding back to XML.
This patch was created to resolve this upstream bug:
https://bugzilla.redhat.com/show_bug.cgi?id=784767
and is at least a partial solution to this RHEL RFE:
https://bugzilla.redhat.com/show_bug.cgi?id=805071
Previously the only attribute of a network device that could be
modified by virUpdateDeviceFlags() ("virsh update-device") was the
link state; attempts to change any other attribute would log an error
and fail.
This patch adds recognition of a change in bridge device name, and
supports reconnecting the guest's interface to the new device.
Standard audit logs for detaching and attaching a network device are
also generated. Although the current auditing function doesn't log the
bridge being attached to, this will later be changed in a separate
patch.
qemuBuildHostNetStr had a switch-within-a-switch where both were
looking at the same variable. This was apparently to take advantage of
code common to three different cases (while also taking care of some
code that was different). However, there were only 2 lines common to
all, one of those can be eliminated by merging it into the
virAsprintfs that are in each case. On top of that, all the extra
empty cases cause Coverity complaints (because they are unreachable),
but absence of the empty cases causes a compile error due to
"enumeration value not handled in switch".
The solution is to just make each toplevel case independent, folding
in the common code to each.
commit b0e2bb33 set a default value for the SPICE agent channel by
inserting it during parsing of the channel XML. That method of setting
a default is problematic because it makes a format/parse roundtrip
unclean, and experience with setting other values as a side effect of
parsing has led to headaches (e.g. automatically setting a MAC address
in the parser when one isn't specified in the input XML).
This patch does not revert commit b0e2bb33 (it will be reverted in a
separate patch) but adds the alternate implementation of simply
inserting the default value in the appropriate place on the qemu
commandline when no value is provided.
If we issue guest command and GA is not running, the issuing thread
will block endlessly. We can check for GA presence by issuing
guest-sync with unique ID (timestamp). We don't want to issue real
command as even if GA is not running, once it is started, it process
all commands written to GA socket.
The code is splattered with a mix of
sizeof foo
sizeof (foo)
sizeof(foo)
Standardize on sizeof(foo) and add a syntax check rule to
enforce it
Signed-off-by: Daniel P. Berrange <berrange@redhat.com>
When qemu cannot start, we may call qemuProcessStop() twice.
We have check whether the vm is running at the beginning of
qemuProcessStop() to avoid libvirt deadlock. We call
qemuProcessStop() with driver and vm locked. It seems that
we can avoid libvirt deadlock. But unfortunately we may
unlock driver and vm in the function qemuProcessKill() while
vm->def->id is not -1. So qemuProcessStop() will be run twice,
and monitor will be freed unexpectedly. So we should set
vm->def->id to -1 at the beginning of qemuProcessStop().
In the current V3 migration protocol, Libvirt does not
check the result of the function
qemuMigrationVPAssociatePortProfiles
This means that it is possible for a migration to complete
successfully even when the VM loses network connectivity on
the destination host.
With this change libvirt aborts the migration
(during the "finish" step) when the above function fails, that
is to say when at least one of the port profile associations fails.
Signed-off by: Christian Benvenuti <benve@cisco.com>
Commit d42a2ff caused a regression in creating a disk-only snapshot
of a qcow2 disk; by passing the wrong variable to the monitor call,
libvirt ended up creating JSON that looked like "format":null instead
of the intended "format":"qcow2".
To make it easier to diagnose this in the future, make JSON creation
error out if "s:arg" is paired with NULL (it is still possible to
use "n:arg" in the rare cases where qemu will accept a null).
* src/qemu/qemu_driver.c
(qemuDomainSnapshotCreateSingleDiskActive): Pass correct value.
* src/qemu/qemu_monitor_json.c (qemuMonitorJSONMakeCommandRaw):
Improve error message.
When libvirtd is restarted, also restart the netlink event
message callbacks for existing VEPA connections and send
a message to lldpad for these existing links, so it learns
the new libvirtd pid.
Signed-off-by: D. Herrendoerfer <d.herrendoerfer@herrendoerfer.name>
This avoids possible deadlock of the qemu driver in case a domain is
begin migrated (in Begin phase) and unrelated connection to qemu driver
is closed at the right time.
I checked all callers of qemuDomainCheckEjectableMedia() and they are
calling this function with qemu driver locked.
Found when attempting to build on Fedora 17 alpha with:
./autogen.sh --system --enable-compile-warnings=error
(this same build command works without problem on Fedora 16). Since
the consumer of the qemuProcessReconnectData doesn't assume that the
other fields of the struct are initialized (although it uses them
internally), the simpler solution is to just switch to C99-style
struct initialization (which doesn't require specification of all
fields).
libvirt always adds -Werror-frame-larger-than=4096 to the flags when
it builds. When building on Fedora 17, two functions with multiple
1024 buffers declared inside if {} blocks would generate frame size
errors; apparently the version of gcc on Fedora 16 will merge these
multiple buffers into a single buffer even when optimization is off,
but Fedora 17 won't.
The fix is to declare a single 1024 buffer at the top of the two
offending functions, and reuse the single buffer throughout the
functions.
Return statements with parameter enclosed in parentheses were modified
and parentheses were removed. The whole change was scripted, here is how:
List of files was obtained using this command:
git grep -l -e '\<return\s*([^()]*\(([^()]*)[^()]*\)*)\s*;' | \
grep -e '\.[ch]$' -e '\.py$'
Found files were modified with this command:
sed -i -e \
's_^\(.*\<return\)\s*(\(\([^()]*([^()]*)[^()]*\)*\))\s*\(;.*$\)_\1 \2\4_' \
-e 's_^\(.*\<return\)\s*(\([^()]*\))\s*\(;.*$\)_\1 \2\3_'
Then checked for nonsense.
The whole command looks like this:
git grep -l -e '\<return\s*([^()]*\(([^()]*)[^()]*\)*)\s*;' | \
grep -e '\.[ch]$' -e '\.py$' | xargs sed -i -e \
's_^\(.*\<return\)\s*(\(\([^()]*([^()]*)[^()]*\)*\))\s*\(;.*$\)_\1 \2\4_' \
-e 's_^\(.*\<return\)\s*(\([^()]*\))\s*\(;.*$\)_\1 \2\3_'
The oVirt developers have stated that the real reasons they want
to have qemu reuse existing volumes when creating a snapshot are:
1. the management framework is set up so that creation has to be
done from a central node for proper resource tracking, and having
libvirt and/or qemu create things violates the framework, and
2. qemu defaults to creating snapshots with an absolute path to
the backing file, but oVirt wants to manage a backing chain that
uses just relative names, to allow for easier migration of a chain
across storage locations.
When 0.9.10 added VIR_DOMAIN_SNAPSHOT_CREATE_REUSE_EXT (commit
4e9953a4), it only addressed point 1, but libvirt was still using
O_TRUNC which violates point 2. Meanwhile, the new qemu
'transaction' monitor command includes a new optional mode argument
that will force qemu to reuse the metadata of the file it just
opened (with the burden on the caller to have valid metadata there
in the first place). So, this tweaks the meaning of the flag to
cover both points as intended for use by oVirt. It is not strictly
backward-compatible to 0.9.10 behavior, but it can be argued that
the O_TRUNC of 0.9.10 was a bug.
Note that this flag is all-or-nothing, and only selects between
'existing' and the default 'absolute-paths'. A more flexible
approach that would allow per-disk selections, as well as adding
support for the 'no-backing-file' mode, would be possible by
extending the <domainsnapshot> xml to have a per-disk mode, but
until we have a management application expressing a need for that
additional complexity, it is not worth doing.
* src/libvirt.c (virDomainSnapshotCreateXML): Tweak documentation.
* src/qemu/qemu_monitor.h (qemuMonitorDiskSnapshot): Add
parameters.
* src/qemu/qemu_monitor_json.h (qemuMonitorJSONDiskSnapshot):
Likewise.
* src/qemu/qemu_monitor.c (qemuMonitorDiskSnapshot): Pass them
through.
* src/qemu/qemu_monitor_json.c (qemuMonitorJSONDiskSnapshot): Use
new monitor command arguments.
* src/qemu/qemu_driver.c (qemuDomainSnapshotCreateDiskActive)
(qemuDomainSnapshotCreateSingleDiskActive): Adjust callers.
(qemuDomainSnapshotDiskPrepare): Allow qed, modify rules on reuse.
The hardest part about adding transactions is not using the new
monitor command, but undoing the partial changes we made prior
to a failed transaction.
* src/qemu/qemu_driver.c (qemuDomainSnapshotCreateDiskActive): Use
transaction when available.
(qemuDomainSnapshotUndoSingleDiskActive): New function.
(qemuDomainSnapshotCreateSingleDiskActive): Pass through actions.
(qemuDomainSnapshotCreateXML): Adjust caller.
QEmu 1.1 is adding a 'transaction' command to the JSON monitor.
Each element of a transaction corresponds to a top-level command,
with the additional guarantee that the transaction flushes all
pending I/O, then guarantees that all actions will be successful
as a group or that failure will roll back the state to what it
was before the monitor command. The difference between a
top-level command:
{ "execute": "blockdev-snapshot-sync", "arguments":
{ "device": "virtio0", ... } }
and a transaction:
{ "execute": "transaction", "arguments":
{ "actions": [
{ "type": "blockdev-snapshot-sync", "data":
{ "device": "virtio0", ... } } ] } }
is just a couple of changed key names and nesting the shorter
command inside a JSON array to the longer command. This patch
just adds the framework; the next patch will actually use a
transaction.
* src/qemu/qemu_monitor_json.c (qemuMonitorJSONMakeCommand): Move
guts...
(qemuMonitorJSONMakeCommandRaw): ...into new helper. Add support
for array element.
(qemuMonitorJSONTransaction): New command.
(qemuMonitorJSONDiskSnapshot): Support use in a transaction.
* src/qemu/qemu_monitor_json.h (qemuMonitorJSONDiskSnapshot): Add
argument.
(qemuMonitorJSONTransaction): New declaration.
* src/qemu/qemu_monitor.h (qemuMonitorTransaction): Likewise.
(qemuMonitorDiskSnapshot): Add argument.
* src/qemu/qemu_monitor.c (qemuMonitorTransaction): New wrapper.
(qemuMonitorDiskSnapshot): Pass argument on.
* src/qemu/qemu_driver.c
(qemuDomainSnapshotCreateSingleDiskActive): Update caller.
Taking an external snapshot of just one disk is atomic, without having
to pause and resume the VM. This also paves the way for later patches
to interact with the new qemu 'transaction' monitor command.
The various scenarios when requesting atomic are:
online, 1 disk, old qemu - safe, allowed by this patch
online, more than 1 disk, old qemu - failure, this patch
offline snapshot - safe, once a future patch implements offline disk snapshot
online, 1 or more disks, new qemu - safe, once future patch uses transaction
Taking an online system checkpoint snapshot is atomic, since it is
done via a single 'savevm' monitor command. Taking an offline system
checkpoint snapshot is atomic, thanks to the previous patch.
* src/qemu/qemu_driver.c (qemuDomainSnapshotCreateXML): Support
new flag for single-disk setups.
(qemuDomainSnapshotDiskPrepare): Check for atomic here.
(qemuDomainSnapshotCreateDiskActive): Skip pausing the VM when
atomic supported.
(qemuDomainSnapshotIsAllowed): Use bool instead of int.
Offline internal snapshots can be rolled back with just a little
bit of refactoring, meaning that we are now automatically atomic.
* src/qemu/qemu_domain.c (qemuDomainSnapshotForEachQcow2): Move
guts...
(qemuDomainSnapshotForEachQcow2Raw): ...to new helper, to allow
rollbacks.
We need a capability bit to gracefully error out if some of the
additions in future patches can't be implemented by the running qemu.
* src/qemu/qemu_capabilities.h (QEMU_CAPS_TRANSACTION): New cap.
* src/qemu/qemu_capabilities.c (qemuCaps): Name it.
* src/qemu/qemu_monitor_json.c (qemuMonitorJSONCheckCommands): Set
it.
This introduces a new running reason VIR_DOMAIN_RUNNING_WAKEUP,
and new suspend event type VIR_DOMAIN_EVENT_STARTED_WAKEUP.
While a wakeup event is emitted, the domain which entered into
VIR_DOMAIN_PMSUSPENDED will be transferred to "running"
with reason VIR_DOMAIN_RUNNING_WAKEUP, and a new domain lifecycle
event emitted with type VIR_DOMAIN_EVENT_STARTED_WAKEUP.
This patch introduces a new event type for the QMP event
SUSPEND:
VIR_DOMAIN_EVENT_ID_PMSUSPEND
The event doesn't take any data, but considering there might
be reason for wakeup in future, the callback definition is:
typedef void
(*virConnectDomainEventSuspendCallback)(virConnectPtr conn,
virDomainPtr dom,
int reason,
void *opaque);
"reason" is unused currently, always passes "0".
This patch introduces a new event type for the QMP event
WAKEUP:
VIR_DOMAIN_EVENT_ID_PMWAKEUP
The event doesn't take any data, but considering there might
be reason for wakeup in future, the callback definition is:
typedef void
(*virConnectDomainEventWakeupCallback)(virConnectPtr conn,
virDomainPtr dom,
int reason,
void *opaque);
"reason" is unused currently, always passes "0".
This is similiar with physical world, one will be surprised if the
box starts with medium exists while the tray is open.
New tests are added, tests disk-{cdrom,floppy}-tray are for the qemu
supports "-device" flag, and disk-{cdrom,floppy}-no-device-cap are
for old qemu, i.e. which doesn't support "-device" flag.
This patch introduces a new event type for the QMP event
DEVICE_TRAY_MOVED, which occurs when the tray of a removable
disk is moved (i.e opened or closed):
VIR_DOMAIN_EVENT_ID_TRAY_CHANGE
The event's data includes the device alias and the reason
for tray status' changing, which indicates why the tray
status was changed. Thus the callback definition for the event
is:
enum {
VIR_DOMAIN_EVENT_TRAY_CHANGE_OPEN = 0,
VIR_DOMAIN_EVENT_TRAY_CHANGE_CLOSE,
\#ifdef VIR_ENUM_SENTINELS
VIR_DOMAIN_EVENT_TRAY_CHANGE_LAST
\#endif
} virDomainEventTrayChangeReason;
typedef void
(*virConnectDomainEventTrayChangeCallback)(virConnectPtr conn,
virDomainPtr dom,
const char *devAlias,
int reason,
void *opaque);
Since we defined a custom virURIPtr type, we should use a
virURIFree method instead of assuming it will always be
a typedef for xmlURIPtr
* src/util/viruri.c, src/util/viruri.h, src/libvirt_private.syms:
Add a virURIFree method
* src/datatypes.c, src/esx/esx_driver.c, src/libvirt.c,
src/qemu/qemu_migration.c, src/vmx/vmx.c, src/xen/xend_internal.c,
tests/viruritest.c: s/xmlFreeURI/virURIFree/
Signed-off-by: Daniel P. Berrange <berrange@redhat.com>
When a client which started non-p2p migration dies in a bad time, the
source libvirtd never clears the migration job and almost nothing can be
done with the domain without restarting the daemon. This patch makes use
of connection close callbacks and ensures that migration job is properly
discarded when the client disconnects.
Destination daemon should not rely on the client or source daemon
(depending on the type of migration) to call Finish when migration
fails, because the client may crash before it can do so. The domain
prepared for incoming migration is set to be destroyed (and migration
job cleaned up) when connection with the client closes but this is not
enough. If the associated qemu process crashes after Prepare step and
the domain is cleaned up before the connection gets closed, autodestroy
is not called for the domain and migration jobs remains set. In case the
domain is defined on destination host (i.e., it is not completely
removed once destroyed) we keep the job set for ever. To fix this, we
register a cleanup callback which is responsible to clean migration-in
job when a domain dies anywhere between Prepare and Finish steps. Note
that we can't blindly clean any job when spotting EOF on monitor since
normally an API is running at that time.
This reverts commit 61f2b6ba5f and most of
commit d8916dc8e2, which effectively
brings back commit ef1065cf5a written by
Jim Fehlig:
The qemu migration speed default is 32MiB/s as defined in migration.c
/* Migration speed throttling */
static int64_t max_throttle = (32 << 20);
There's no need to throttle migration when targeting a file, so set
migration speed to unlimited prior to migration, and restore to libvirt
default value after migration.
Default units is MB for migrate_set_speed monitor command, so
(INT64_MAX / (1024 * 1024)) is used for unlimited migration speed.
This was reverted because migration to file could not be canceled and
even monitored since qemu was not processing any monitor commands until
the migration finished. This is now different as we make sure the
file descriptor we pass to qemu is able to properly report EAGAIN.
Recent qemu changes might have helped as well.
I tested managedsave with this patch in and indeed, it is 10x faster
while I can still monitor its progress.
If a guest is paused, we were silently ignoring the quiesce flag,
which results in unclean snapshots, contrary to the intent of the
flag. Since we can't quiesce without guest agent support, we should
instead fail if the guest is not running.
Meanwhile, if we attempt a quiesce command, but the guest agent
doesn't respond, and we time out, we may have left the command
pending on the guest's queue, and when the guest resumes parsing
commands, it will freeze even though our command is no longer
around to issue a thaw. To be safe, we must _always_ pair every
quiesce call with a counterpart thaw, even if the quiesce call
failed due to a timeout, so that if a guest wakes up and starts
processing a command backlog, it will not get stuck in a frozen
state.
* src/qemu/qemu_driver.c (qemuDomainSnapshotCreateDiskActive):
Always issue thaw after a quiesce, even if quiesce failed.
(qemuDomainSnapshotFSThaw): Add a parameter.
A common coding pattern for changing blkio parameters is
1. virDomainGetBlkioParameters
2. change one or more params
3. virDomainSetBlkioParameters
For this to work, it must be possible to roundtrip through
the methods without error. Unfortunately virDomainGetBlkioParameters
will return "" for the deviceWeight parameter for guests by default,
which virDomainSetBlkioParameters will then reject as invalid.
This fixes the handling of "" to be a no-op, and also improves the
error message to tell you what was invalid
Wire up the domain graphics event notifications for SPICE. Adapted
from a RHEL-only patch written by Dan Berrange that used custom
__com.redhat_SPICE events - equivalent events are now available in
upstream QEMU (including a SPICE_CONNECTED event, which was missing in
the __COM.redhat_SPICE version).
* src/qemu/qemu_monitor_json.c: Wire up SPICE graphics events
numad is an user-level daemon that monitors NUMA topology and
processes resource consumption to facilitate good NUMA resource
alignment of applications/virtual machines to improve performance
and minimize cost of remote memory latencies. It provides a
pre-placement advisory interface, so significant processes can
be pre-bound to nodes with sufficient available resources.
More details: http://fedoraproject.org/wiki/Features/numad
"numad -w ncpus:memory_amount" is the advisory interface numad
provides currently.
This patch add the support by introducing a new XML attribute
for <vcpu>. e.g.
<vcpu placement="auto">4</vcpu>
<vcpu placement="static" cpuset="1-10^6">4</vcpu>
The returned advisory nodeset from numad will be printed
in domain's dumped XML. e.g.
<vcpu placement="auto" cpuset="1-10^6">4</vcpu>
If placement is "auto", the number of vcpus and the current
memory amount specified in domain XML will be used for numad
command line (numad uses MB for memory amount):
numad -w $num_of_vcpus:$current_memory_amount / 1024
The advisory nodeset returned from numad will be used to set
domain process CPU affinity then. (e.g. qemuProcessInitCpuAffinity).
If the user specifies both CPU affinity policy (e.g.
(<vcpu cpuset="1-10,^7,^8">4</vcpu>) and placement == "auto"
the specified CPU affinity will be overridden.
Only QEMU/KVM drivers support it now.
See docs update in patch for more details.
With current code, we pass true iff domain is cold booting. However,
if disk is inaccessible and startupPolicy for that disk is set to
'requisite' we have to fail iff cold booting.
Even though we say in documentation setting (tls-)port to -1 is legacy
compat style for enabling autoport, we're roughly doing this for VNC.
However, in case of SPICE auto enable autoport iff both port & tlsPort
are equal -1 as documentation says autoport plays with both.
In qemuDomainDetachNetDevice, detach was being used before it had been
validated. If no matching device was found, this resulted in a
dereference of a NULL pointer.
This behavior was a regression introduced in commit
cf90342be0, so it has not been a part of
any official libvirt release.
When host-model and host-passthrouh CPU modes were introduced, qemu
driver was properly modify to update guest CPU definition during
migration so that we use the right CPU at the destination. However,
similar treatment is needed for (managed)save and snapshots since they
need to save the exact CPU so that a domain can be properly restored.
To avoid repetition of such situation, all places that need live XML
share the code which generates it.
As a side effect, this patch fixes error reporting from
qemuDomainSnapshotWriteMetadata().
Thanks to cgroups, providing user vs. system time of the overall
guest is easy to add to our existing API.
* include/libvirt/libvirt.h.in (VIR_DOMAIN_CPU_STATS_USERTIME)
(VIR_DOMAIN_CPU_STATS_SYSTEMTIME): New constants.
* src/util/virtypedparam.h (virTypedParameterArrayValidate)
(virTypedParameterAssign): Enforce checking the result.
* src/qemu/qemu_driver.c (qemuDomainGetPercpuStats): Fix offender.
(qemuDomainGetTotalcpuStats): Implement new parameters.
* tools/virsh.c (cmdCPUStats): Tweak output accordingly.
If there is a disk file with a comma in the name, QEmu expects a double
comma instead of a single one (e.g., the file "virtual,disk.img" needs
to be specified as "virtual,,disk.img" in QEmu's command line). This
patch fixes libvirt to work with that feature. Fix RHBZ #801036.
Based on an initial patch by Crístian Viana.
* src/util/buf.h (virBufferEscape): Alter signature.
* src/util/buf.c (virBufferEscape): Add parameter.
(virBufferEscapeSexpr): Fix caller.
* src/qemu/qemu_command.c (qemuBuildRBDString): Likewise. Also
escape commas in file names.
(qemuBuildDriveStr): Escape commas in file names.
* docs/schemas/basictypes.rng (absFilePath): Relax RNG to allow
commas in input file names.
* tests/qemuxml2argvdata/*-disk-drive-network-sheepdog.*: Update
test.
Signed-off-by: Eric Blake <eblake@redhat.com>
If user hasn't supplied any tlsPort we default to setting it
to zero in our internal structure. However, when building command
line we test it against -1 which is obviously wrong.
This function was freeing a virDomainNetDef with
VIR_FREE(). virDomainNetDef is a complex structure with many pointers
to other dynamically allocated data; to properly free it
virDomainNetDefFree() must be called instead, otherwise several
strings (and potentially other things) will be leaked.
For some reason, although live hotplug of <hostdev> devices is
supported, persistent hotplug is not. This patch adds the proper
VIR_DOMAIN_DEVICE_HOSTDEV cases to the switches in
qemuDomainAttachDeviceConfig and qemuDomainDetachDeviceConfig.
There are several functions in domain_conf.c that remove a device
object from the domain's list of that object type, but don't free the
object or return it to the caller to free. In many cases this isn't a
problem because the caller already had a pointer to the object and
frees it afterward, but in several cases the removed object was just
left floating around with no references to it.
In particular, the function qemuDomainDetachDeviceConfig() calls
functions to locate and remove net (virDomainNetRemoveByMac), disk
(virDomainDiskRemoveByName()), and lease (virDomainLeaseRemove())
devices, but neither it nor its caller qemuDomainModifyDeviceConfig()
ever obtain a pointer to the device being removed, much less free it.
This patch modifies the following "remove" functions to return a
pointer to the device object being removed from the domain device
arrays, to give the caller the option of freeing the device object
using that pointer if needed. In places where the object was
previously leaked, it is now freed:
virDomainDiskRemove
virDomainDiskRemoveByName
virDomainNetRemove
virDomainNetRemoveByMac
virDomainHostdevRemove
virDomainLeaseRemove
virDomainLeaseRemoveAt
The functions that had been leaking:
libxlDomainDetachConfig - leaked a virDomainDiskDef
qemuDomainDetachDeviceConfig - could leak a virDomainDiskDef,
a virDomainNetDef, or a
virDomainLeaseDef
qemuDomainDetachLease - leaked a virDomainLeaseDef
There were certain paths through the hostdev detach code that could
lead to the lower level function failing (and not removing the object
from the domain's hostdevs list), but the higher level function
free'ing the hostdev object anyway. This would leave a stale
hostdevdef pointer in the list, which would surely cause a problem
eventually.
This patch relocates virDomainHostdevRemove from the lower level
functions qemuDomainDetachThisHostDevice and
qemuDomainDetachHostPciDevice, to their caller
qemuDomainDetachThisHostDevice, placing it just before the call to
virDomainHostdevDefFree. This makes it easy to verify that either both
operations are done, or neither.
NB: The "dangling pointer" part of this problem was introduced in
commit 13d5a6, so it is not present in libvirt versions prior to
0.9.9. Earlier versions would return failure in certain cases even
though the the device object was removed/deleted, but the removal and
deletion operations would always both happen or neither.
This patch will allow OpenFlow controllers to identify which interface
belongs to a particular VM by using the Domain UUID.
ovs-vsctl get Interface vnet0 external_ids
{attached-mac="52:54:00:8C:55:2C", iface-id="83ce45d6-3639-096e-ab3c-21f66a05f7fa", iface-status=active, vm-id="142a90a7-0acc-ab92-511c-586f12da8851"}
V2 changes:
Replaced vm-uuid with vm-id. There was a discussion in Open vSwitch
mailinglist that we should stick with the same DB key postfixes for the
sake of consistency (e.g iface-id, vm-id ...).
Some members are generated during XML parse (e.g. MAC address of
an interface); However, with current implementation, if we
are plugging a device both to persistent and live config,
we parse given XML twice: first time for live, second for config.
This is wrong then as the second time we are not guaranteed
to generate same values as we did for the first time.
To prevent that we need to create a copy of DeviceDefPtr;
This is done through format/parse process instead of writing
functions for deep copy as it is easier to maintain:
adding new field to any virDomain*DefPtr doesn't require change
of copying function.
Currently, startupPolicy='requisite' was determining cold boot
by migrateFrom != NULL. That means, if domain was started up
with migrateFrom set we didn't require disk source path and allowed
it to be dropped. However, on snapshot-revert domain wasn't migrated
but according to documentation, requisite should drop disk source
as well.
Using 'unsigned long' for memory values is risky on 32-bit platforms,
as a PAE guest can have more than 4GiB memory. Our API is
(unfortunately) locked at 'unsigned long' and a scale of 1024, but
the rest of our system should consistently use 64-bit values,
especially since the previous patch centralized overflow checking.
* src/conf/domain_conf.h (_virDomainDef): Always use 64-bit values
for memory. Change hugepage_backed to a bool.
* src/conf/domain_conf.c (virDomainDefParseXML)
(virDomainDefCheckABIStability, virDomainDefFormatInternal): Fix
clients.
* src/vmx/vmx.c (virVMXFormatConfig): Likewise.
* src/xenxs/xen_sxpr.c (xenParseSxpr, xenFormatSxpr): Likewise.
* src/xenxs/xen_xm.c (xenXMConfigGetULongLong): New function.
(xenXMConfigGetULong, xenXMConfigSetInt): Avoid truncation.
(xenParseXM, xenFormatXM): Fix clients.
* src/phyp/phyp_driver.c (phypBuildLpar): Likewise.
* src/openvz/openvz_driver.c (openvzDomainSetMemoryInternal):
Likewise.
* src/vbox/vbox_tmpl.c (vboxDomainDefineXML): Likewise.
* src/qemu/qemu_command.c (qemuBuildCommandLine): Likewise.
* src/qemu/qemu_process.c (qemuProcessStart): Likewise.
* src/qemu/qemu_monitor.h (qemuMonitorGetBalloonInfo): Likewise.
* src/qemu/qemu_monitor_text.h (qemuMonitorTextGetBalloonInfo):
Likewise.
* src/qemu/qemu_monitor_text.c (qemuMonitorTextGetBalloonInfo):
Likewise.
* src/qemu/qemu_monitor_json.h (qemuMonitorJSONGetBalloonInfo):
Likewise.
* src/qemu/qemu_monitor_json.c (qemuMonitorJSONGetBalloonInfo):
Likewise.
* src/qemu/qemu_driver.c (qemudDomainGetInfo)
(qemuDomainGetXMLDesc): Likewise.
* src/uml/uml_conf.c (umlBuildCommandLine): Likewise.
On 64-bit platforms, unsigned long and unsigned long long are
identical, so we don't have to worry about overflow checks.
On 32-bit platforms, anywhere we narrow unsigned long long back
to unsigned long, we have to worry about overflow; it's easier
to do this in one place by having most of the code use the same
or wider types, and only doing the narrowing at the last minute.
Therefore, the memory set commands remain unsigned long, and
the memory get command now centralizes the overflow check into
libvirt.c, so that drivers don't have to repeat the work.
This also fixes a bug where xen returned the wrong value on
failure (most APIs return -1 on failure, but getMaxMemory
must return 0 on failure).
* src/driver.h (virDrvDomainGetMaxMemory): Use long long.
* src/libvirt.c (virDomainGetMaxMemory): Raise overflow.
* src/test/test_driver.c (testGetMaxMemory): Fix driver.
* src/rpc/gendispatch.pl (name_to_ProcName): Likewise.
* src/xen/xen_hypervisor.c (xenHypervisorGetMaxMemory): Likewise.
* src/xen/xen_driver.c (xenUnifiedDomainGetMaxMemory): Likewise.
* src/xen/xend_internal.c (xenDaemonDomainGetMaxMemory):
Likewise.
* src/xen/xend_internal.h (xenDaemonDomainGetMaxMemory):
Likewise.
* src/xen/xm_internal.c (xenXMDomainGetMaxMemory): Likewise.
* src/xen/xm_internal.h (xenXMDomainGetMaxMemory): Likewise.
* src/xen/xs_internal.c (xenStoreDomainGetMaxMemory): Likewise.
* src/xen/xs_internal.h (xenStoreDomainGetMaxMemory): Likewise.
* src/xenapi/xenapi_driver.c (xenapiDomainGetMaxMemory):
Likewise.
* src/esx/esx_driver.c (esxDomainGetMaxMemory): Likewise.
* src/libxl/libxl_driver.c (libxlDomainGetMaxMemory): Likewise.
* src/qemu/qemu_driver.c (qemudDomainGetMaxMemory): Likewise.
* src/lxc/lxc_driver.c (lxcDomainGetMaxMemory): Likewise.
* src/uml/uml_driver.c (umlDomainGetMaxMemory): Likewise.
Overflow can be user-induced, so it deserves more than being called
an internal error. Note that in general, 32-bit platforms have
far more places to trigger this error (anywhere the public API
used 'unsigned long' but the other side of the connection is a
64-bit server); but some are possible on 64-bit platforms (where
the public API computes the product of two numbers).
* include/libvirt/virterror.h (VIR_ERR_OVERFLOW): New error.
* src/util/virterror.c (virErrorMsg): Translate it.
* src/libvirt.c (virDomainSetVcpusFlags, virDomainGetVcpuPinInfo)
(virDomainGetVcpus, virDomainGetCPUStats): Use it.
* daemon/remote.c (HYPER_TO_TYPE): Likewise.
* src/qemu/qemu_driver.c (qemuDomainBlockResize): Likewise.
The RPC code assumed that the array returned by the driver would be
fully populated; that is, ncpus on entry resulted in ncpus * return
value on exit. However, while we don't support holes in the middle
of ncpus, we do want to permit the case of ncpus on entry being
longer than the array returned by the driver (that is, it should be
safe for the caller to pass ncpus=128 on entry, and the driver will
stop populating the array when it hits max_id).
Additionally, a successful return implies that the caller will then
use virTypedParamArrayClear on the entire array; for this to not
free uninitialized memory, the driver must ensure that all skipped
entries are explicitly zeroed (the RPC driver did this, but not
the qemu driver).
There are now three cases:
server 0.9.10 and client 0.9.10 or newer: No impact - there were no
hypervisor drivers that supported cpu stats
server 0.9.11 or newer and client 0.9.10: if the client calls with
ncpus beyond the max, then the rpc call will fail on the client side
and disconnect the client, but the server is no worse for the wear
server 0.9.11 or newer and client 0.9.11: the server can return a
truncated array and the client will do just fine
I reproduced the problem by using a host with 2 CPUs, and doing:
virsh cpu-stats $dom --start 1 --count 2
* daemon/remote.c (remoteDispatchDomainGetCPUStats): Allow driver
to omit tail of array.
* src/remote/remote_driver.c (remoteDomainGetCPUStats):
Accommodate driver that omits tail of array.
* src/libvirt.c (virDomainGetCPUStats): Document this.
* src/qemu/qemu_driver.c (qemuDomainGetPercpuStats): Clear all
unpopulated entries.
* For now, only "cpu_time" is supported.
* cpuacct cgroup is used for providing percpu cputime information.
* src/qemu/qemu.conf - take care of cpuacct cgroup.
* src/qemu/qemu_conf.c - take care of cpuacct cgroup.
* src/qemu/qemu_driver.c - added an interface
* src/util/cgroup.c/h - added interface for getting percpu cputime
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
These changes are applied only if the hostdev has a parent net device
(i.e. if it was defined as "<interface type='hostdev'>" rather than
just "<hostdev>"). If the parent netdevice has virtual port
information, the original virtualport associate functions are called
(these set and restore both mac and port profile on an
interface). Otherwise, only mac address is set on the device.
Note that This is only supported for SR-IOV Virtual Functions (not for
standard PCI or USB netdevs), and virtualport association is only
supported for 802.1Qbh. For all other types of cards and types of
virtualport, a "Config Unsupported" error is returned and the
operation fails.
Signed-off-by: Roopa Prabhu <roprabhu@cisco.com>
This patch includes the following changes to virnetdevmacvlan.c and
virnetdevvportprofile.c:
- removes some netlink functions which are now available in
virnetdev.c
- Adds a vf argument to all port profile functions.
For 802.1Qbh devices, the port profile calls can use a vf argument if
passed by the caller. If the vf argument is -1 it will try to derive the vf
if the device passed is a virtual function.
For 802.1Qbg devices, This patch introduces a null check for the device
argument because during port profile assignment on a hostdev, this argument
can be null.
Signed-off-by: Roopa Prabhu <roprabhu@cisco.com>
qemuDomainAttachNetDevice
- re-ordered some things at start of function because
networkAllocateActualDevice should always be run and a slot
in def->nets always allocated, but host_net_add isn't needed
if the actual type is hostdev.
- if actual type is hostdev, defer to
qemuDomainAttachHostDevice (which will reach up to the NetDef
for things like MAC address when necessary). After return
from qemuDomainAttachHostDevice, slip directly to cleanup,
since the rest of the function is specific to emulated net
devices.
- put assignment of new NetDef into expanded def->nets down
below cleanup: (but only on success) since it is also needed
for emulated and hostdev net devices.
qemuDomainDetachHostDevice
- after locating the exact device to detach, check if it's a
network device and, if so, use toplevel
qemuDomainDetachNetDevice instead so that the def->nets list
is properly updated, and 'actual device' properly returned to
network pool if appropriate. Otherwise, for normal hostdevs,
call the lower level qemuDomainDetachThisDevice.
qemuDomainDetachNetDevice
- This is where it gets a bit tricky. After locating the device
on the def->nets list, if the network device type == hostdev,
call the *lower level* qemuDomainDetachThisDevice (which will
reach back up to the parent net device for MAC address /
virtualport when appropriate, then clear the device out of
def->hostdevs) before skipping past all the emulated
net-device-specific code to cleanup:, where the network
device is removed from def->nets, and the network device
object is freed.
In short, any time a hostdev-type network device is detached, we must
go through the toplevel virDomaineDetachNetDevice function first and
last, to make sure 1) the def->nnets list is properly managed, and 2)
any device allocated with networkAllocateActualDevice is properly
freed. At the same time, in the middle we need to go through the
lower-level vidDomainDetach*This*HostDevice to be sure that 1) the
def->hostdevs list is properly managed, 2) the PCI device is properly
detached from the guest and reattached to the host (if appropriate),
and 3) any higher level teardown is called at the appropriate time, by
reaching back up to the NetDef config (part (3) will be covered in a
separate patch).
This patch makes sure that each network device ("interface") of
type='hostdev' appears on both the hostdevs list and the nets list of
the virDomainDef, and it modifies the qemu driver startup code so that
these devices will be presented to qemu on the commandline as hostdevs
rather than as network devices.
It does not add support for hotplug of these type of devices, or code
to honor the <mac address> or <virtualport> given in the config (both
of those will be done in separate patches).
Once each device is placed on both lists, much of what this patch does
is modify places in the code that traverse all the device lists so
that these hybrid devices are only acted on once - either along with
the other hostdevs, or along with the other network interfaces. (In
many cases, only one of the lists is traversed / a specific operation
is performed on only one type of device. In those instances, the code
can remain unchanged.)
There is one special case - when building the commandline, interfaces
are allowed to proceed all the way through
networkAllocateActualDevice() before deciding to skip the rest of
netdev-specific processing - this is so that (once we have support for
networks with pools of hostdev devices) we can get the actual device
allocated, then rely on the loop processing all hostdevs to generate
the correct commandline.
(NB: <interface type='hostdev'> is only supported for PCI network
devices that are SR-IOV Virtual Functions (VF). Standard PCI[e] and
USB devices, and even the Physical Functions (PF) of SR-IOV devices
can only be assigned to a guest using the more basic <hostdev> device
entry. This limitation is mostly due to the fact that non-SR-IOV
ethernet devices tend to lose mac address configuration whenever the
card is reset, which happens when a card is assigned to a guest;
SR-IOV VFs fortunately don't suffer the same problem.)
This is the new interface type that sets up an SR-IOV PCI network
device to be assigned to the guest with PCI passthrough after
initializing some network device-specific things from the config
(e.g. MAC address, virtualport profile parameters). Here is an example
of the syntax:
<interface type='hostdev' managed='yes'>
<source>
<address type='pci' domain='0' bus='0' slot='4' function='3'/>
</source>
<mac address='00:11:22:33:44:55'/>
<address type='pci' domain='0' bus='0' slot='7' function='0'/>
</interface>
This would assign the PCI card from bus 0 slot 4 function 3 on the
host, to bus 0 slot 7 function 0 on the guest, but would first set the
MAC address of the card to 00:11:22:33:44:55.
NB: The parser and formatter don't care if the PCI card being
specified is a standard single function network adapter, or a virtual
function (VF) of an SR-IOV capable network adapter, but the upcoming
code that implements the back end of this config will work *only* with
SR-IOV VFs. This is because modifying the mac address of a standard
network adapter prior to assigning it to a guest is pointless - part
of the device reset that occurs during that process will reset the MAC
address to the value programmed into the card's firmware.
Although it's not supported by any of libvirt's hypervisor drivers,
usb network hostdevs are also supported in the parser and formatter
for completeness and consistency. <source> syntax is identical to that
for plain <hostdev> devices, except that the <address> element should
have "type='usb'" added if bus/device are specified:
<interface type='hostdev'>
<source>
<address type='usb' bus='0' device='4'/>
</source>
<mac address='00:11:22:33:44:55'/>
</interface>
If the vendor/product form of usb specification is used, type='usb'
is implied:
<interface type='hostdev'>
<source>
<vendor id='0x0012'/>
<product id='0x24dd'/>
</source>
<mac address='00:11:22:33:44:55'/>
</interface>
Again, the upcoming patch to fill in the backend of this functionality
will log an error and fail with "Unsupported Config" if you actually
try to assign a USB network adapter to a guest using <interface
type='hostdev'> - just use a standard <hostdev> entry in that case
(and also for single-port PCI adapters).
This refactoring is necessary to support hotplug detach of
type=hostdev network devices, but needs to be in a separate patch to
make potential debugging of regressions more practical.
Rather than the lowest level functions searching for a matching
device, the search is now done in the toplevel function, and an
intermediate-level function (qemuDomainDetachThisHostDevice()), which
expects that the device's entry is already found, is called (this
intermediate function will be called by qemuDomainDetachNetDevice() in
order to support detach of type=hostdev net devices)
This patch should result in 0 differences in functionality.
