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.
