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.
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().
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.
No thanks to 64-bit windows, with 64-bit pid_t, we have to avoid
constructs like 'int pid'. Our API in libvirt-qemu cannot be
changed without breaking ABI; but then again, libvirt-qemu can
only be used on systems that support UNIX sockets, which rules
out Windows (even if qemu could be compiled there) - so for all
points on the call chain that interact with this API decision,
we require a different variable name to make it clear that we
audited the use for safety.
Adding a syntax-check rule only solves half the battle; anywhere
that uses printf on a pid_t still needs to be converted, but that
will be a separate patch.
* cfg.mk (sc_correct_id_types): New syntax check.
* src/libvirt-qemu.c (virDomainQemuAttach): Document why we didn't
use pid_t for pid, and validate for overflow.
* include/libvirt/libvirt-qemu.h (virDomainQemuAttach): Tweak name
for syntax check.
* src/vmware/vmware_conf.c (vmwareExtractPid): Likewise.
* src/driver.h (virDrvDomainQemuAttach): Likewise.
* tools/virsh.c (cmdQemuAttach): Likewise.
* src/remote/qemu_protocol.x (qemu_domain_attach_args): Likewise.
* src/qemu_protocol-structs (qemu_domain_attach_args): Likewise.
* src/util/cgroup.c (virCgroupPidCode, virCgroupKillInternal):
Likewise.
* src/qemu/qemu_command.c(qemuParseProcFileStrings): Likewise.
(qemuParseCommandLinePid): Use pid_t for pid.
* daemon/libvirtd.c (daemonForkIntoBackground): Likewise.
* src/conf/domain_conf.h (_virDomainObj): Likewise.
* src/probes.d (rpc_socket_new): Likewise.
* src/qemu/qemu_command.h (qemuParseCommandLinePid): Likewise.
* src/qemu/qemu_driver.c (qemudGetProcessInfo, qemuDomainAttach):
Likewise.
* src/qemu/qemu_process.c (qemuProcessAttach): Likewise.
* src/qemu/qemu_process.h (qemuProcessAttach): Likewise.
* src/uml/uml_driver.c (umlGetProcessInfo): Likewise.
* src/util/virnetdev.h (virNetDevSetNamespace): Likewise.
* src/util/virnetdev.c (virNetDevSetNamespace): Likewise.
* tests/testutils.c (virtTestCaptureProgramOutput): Likewise.
* src/conf/storage_conf.h (_virStoragePerms): Use mode_t, uid_t,
and gid_t rather than int.
* src/security/security_dac.c (virSecurityDACSetOwnership): Likewise.
* src/conf/storage_conf.c (virStorageDefParsePerms): Avoid
compiler warning.
This patch is based on an earlier patch by Eric Blake which was never
committed:
https://www.redhat.com/archives/libvir-list/2011-November/msg00243.html
Aside from rebasing, this patch only drops the driver lock once (prior
to the first time the function sleeps), then leaves it dropped until
it returns (Eric's patch would drop and re-acquire the lock around
each call to sleep).
At the time Eric sent his patch, the response (from Dan Berrange) was
that, while it wasn't a good thing to be holding the driver lock while
sleeping, we really need to rethink locking wrt the driver object,
switching to a finer-grained approach that locks individual items
within the driver object separately to allow for greater concurrency.
This is a good plan, and at the time it made sense to not apply the
patch because there was no known bug related to the driver lock being
held in this function.
However, we now know that the length of the wait in qemuProcessKill is
sometimes too short to allow the qemu process to fully flush its disk
cache before SIGKILL is sent, so we need to lengthen the timeout (in
order to improve the situation with management applications until they
can be updated to use the new VIR_DOMAIN_DESTROY_GRACEFUL flag added
in commit 72f8a7f197). But, if we
lengthen the timeout, we also lengthen the amount of time that all
other threads in libvirtd are essentially blocked from doing anything
(since just about everything needs to acquire the driver lock, if only
for long enough to get a pointer to a domain).
The solution is to modify qemuProcessKill to drop the driver lock
while sleeping, as proposed in Eric's patch. Then we can increase the
timeout with a clear conscience, and thus at least lower the chances
that someone running with existing management software will suffer the
consequence's of qemu's disk cache not being flushed.
In the meantime, we still should work on Dan's proposal to make
locking within the driver object more fine grained.
(NB: although I couldn't find any instance where qemuProcessKill() was
called with no jobs active for the domain (or some other guarantee
that the current thread had at least one refcount on the domain
object), this patch still follows Eric's method of temporarily adding
a ref prior to unlocking the domain object, because I couldn't
convince myself 100% that this was the case.)
When libvirt's virDomainDestroy API is shutting down the qemu process,
it first sends SIGTERM, then waits for 1.6 seconds and, if it sees the
process still there, sends a SIGKILL.
There have been reports that this behavior can lead to data loss
because the guest running in qemu doesn't have time to flush its disk
cache buffers before it's unceremoniously whacked.
This patch maintains that default behavior, but provides a new flag
VIR_DOMAIN_DESTROY_GRACEFUL to alter the behavior. If this flag is set
in the call to virDomainDestroyFlags, SIGKILL will never be sent to
the qemu process; instead, if the timeout is reached and the qemu
process still exists, virDomainDestroy will return an error.
Once this patch is in, the recommended method for applications to call
virDomainDestroyFlags will be with VIR_DOMAIN_DESTROY_GRACEFUL
included. If that fails, then the application can decide if and when
to call virDomainDestroyFlags again without
VIR_DOMAIN_DESTROY_GRACEFUL (to force the issue with SIGKILL).
(Note that this does not address the issue of existing applications
that have not yet been modified to use VIR_DOMAIN_DESTROY_GRACEFUL.
That is a separate patch.)
In preparation for code re-organization, rename the Macvtap
management APIs to have the following patterns
virNetDevMacVLanXXXXX - macvlan/macvtap interface management
virNetDevVPortProfileXXXX - virtual port profile management
* src/util/macvtap.c, src/util/macvtap.h: Rename APIs
* src/conf/domain_conf.c, src/network/bridge_driver.c,
src/qemu/qemu_command.c, src/qemu/qemu_command.h,
src/qemu/qemu_driver.c, src/qemu/qemu_hotplug.c,
src/qemu/qemu_migration.c, src/qemu/qemu_process.c,
src/qemu/qemu_process.h: Update for renamed APIs
Ever since we introduced fake reboot, we call qemuProcessKill as a
reaction to SHUTDOWN event. Unfortunately, qemu doesn't guarantee it
flushed all internal buffers before sending SHUTDOWN, in which case
killing the process forcibly may result in (virtual) disk corruption.
By sending just SIGTERM without SIGKILL we give qemu time to to flush
all buffers and exit. Once qemu exits, we will see an EOF on monitor
connection and tear down the domain. In case qemu ignores SIGTERM or
just hangs there, the process stays running but that's not any different
from a possible hang anytime during the shutdown process so I think it's
just fine.
Also qemu (since 0.14 until it's fixed) has a bug in SIGTERM processing
which causes it not to exit but instead send new SHUTDOWN event and keep
waiting. I think the best we can do is to ignore duplicate SHUTDOWN
events to avoid a SHUTDOWN-SIGTERM loop and leave the domain in paused
state.
Changing the current vm, and writing that change to the file
system, all before a new qemu starts, is risky; it's hard to
roll back if starting the new qemu fails for some reason.
Instead of abusing vm->current_snapshot and making the command
line generator decide whether the current snapshot warrants
using -loadvm, it is better to just directly pass a snapshot all
the way through the call chain if it is to be loaded.
This frees up the last use of snapshot->def->active for qemu's
use, so the next patch can repurpose that field for tracking
which snapshot is current.
* src/qemu/qemu_command.c (qemuBuildCommandLine): Don't use active
field of snapshot.
* src/qemu/qemu_process.c (qemuProcessStart): Add a parameter.
* src/qemu/qemu_process.h (qemuProcessStart): Update prototype.
* src/qemu/qemu_migration.c (qemuMigrationPrepareAny): Update
callers.
* src/qemu/qemu_driver.c (qemudDomainCreate)
(qemuDomainSaveImageStartVM, qemuDomainObjStart)
(qemuDomainRevertToSnapshot): Likewise.
(qemuDomainSnapshotSetCurrentActive)
(qemuDomainSnapshotSetCurrentInactive): Delete unused functions.
Currently, we attempt to run sync job and async job at the same time. It
means that the monitor commands for two jobs can be run in any order.
In the function qemuDomainObjEnterMonitorInternal():
if (priv->job.active == QEMU_JOB_NONE && priv->job.asyncJob) {
if (qemuDomainObjBeginNestedJob(driver, obj) < 0)
We check whether the caller is an async job by priv->job.active and
priv->job.asynJob. But when an async job is running, and a sync job is
also running at the time of the check, then priv->job.active is not
QEMU_JOB_NONE. So we cannot check whether the caller is an async job
in the function qemuDomainObjEnterMonitorInternal(), and must instead
put the burden on the caller to tell us when an async command wants
to do a nested job.
Once the burden is on the caller, then only async monitor enters need
to worry about whether the VM is still running; for sync monitor enter,
the internal return is always 0, so lots of ignore_value can be dropped.
* src/qemu/THREADS.txt: Reflect new rules.
* src/qemu/qemu_domain.h (qemuDomainObjEnterMonitorAsync): New
prototype.
* src/qemu/qemu_process.h (qemuProcessStartCPUs)
(qemuProcessStopCPUs): Add parameter.
* src/qemu/qemu_migration.h (qemuMigrationToFile): Likewise.
(qemuMigrationWaitForCompletion): Make static.
* src/qemu/qemu_domain.c (qemuDomainObjEnterMonitorInternal): Add
parameter.
(qemuDomainObjEnterMonitorAsync): New function.
(qemuDomainObjEnterMonitor, qemuDomainObjEnterMonitorWithDriver):
Update callers.
* src/qemu/qemu_driver.c (qemuDomainSaveInternal)
(qemudDomainCoreDump, doCoreDump, processWatchdogEvent)
(qemudDomainSuspend, qemudDomainResume, qemuDomainSaveImageStartVM)
(qemuDomainSnapshotCreateActive, qemuDomainRevertToSnapshot):
Likewise.
* src/qemu/qemu_process.c (qemuProcessStopCPUs)
(qemuProcessFakeReboot, qemuProcessRecoverMigration)
(qemuProcessRecoverJob, qemuProcessStart): Likewise.
* src/qemu/qemu_migration.c (qemuMigrationToFile)
(qemuMigrationWaitForCompletion, qemuMigrationUpdateJobStatus)
(qemuMigrationJobStart, qemuDomainMigrateGraphicsRelocate)
(doNativeMigrate, doTunnelMigrate, qemuMigrationPerformJob)
(qemuMigrationPerformPhase, qemuMigrationFinish)
(qemuMigrationConfirm): Likewise.
* src/qemu/qemu_hotplug.c: Drop unneeded ignore_value.
Given a PID, the QEMU driver reads /proc/$PID/cmdline and
/proc/$PID/environ to get the configuration. This is fed
into the ARGV->XML convertor to build an XML configuration
for the process.
/proc/$PID/exe is resolved to identify the full command
binary path
After checking for name/uuid uniqueness, an attempt is
made to connect to the monitor socket. If successful
then 'info status' and 'info kvm' are issued to determine
whether the CPUs are running and if KVM is enabled.
* src/qemu/qemu_driver.c: Implement virDomainQemuAttach
* src/qemu/qemu_process.h, src/qemu/qemu_process.c: Add
qemuProcessAttach to connect to the monitor of an
existing QEMU process
If an application is using libvirt + KVM as a piece of its
internal infrastructure to perform a specific task, it can
be desirable to guarentee the VM dies when the virConnectPtr
disconnects from libvirtd. This ensures the app can't leak
any VMs it was using. Adding VIR_DOMAIN_START_AUTOKILL as
a flag when starting guests enables this to be done.
* include/libvirt/libvirt.h.in: All VIR_DOMAIN_START_AUTOKILL
* src/qemu/qemu_driver.c: Support automatic killing of guests
upon connection close
* tools/virsh.c: Add --autokill flag to 'start' and 'create'
commands
Sometimes it is useful to be able to automatically destroy a guest when
a connection is closed. For example, kill an incoming migration if
the client managing the migration dies. This introduces a map between
guest 'uuid' strings and virConnectPtr objects. When a connection is
closed, any associated guests are killed off.
* src/qemu/qemu_conf.h: Add autokill hash table to qemu driver
* src/qemu/qemu_process.c, src/qemu/qemu_process.h: Add APIs
for performing autokill of guests associated with a connection
* src/qemu/qemu_driver.c: Initialize autodestroy map
Only in drivers which use virDomainObj, drivers that query hypervisor
for domain status need to be updated separately in case their hypervisor
supports this functionality.
The reason is also saved into domain state XML so if a domain is not
running (i.e., no state XML exists) the reason will be lost by libvirtd
restart. I think this is an acceptable limitation.
Introduce a virProcessKill function that can be safely called
even when the job mutex is held. This allows virDomainDestroy
to kill any VM even if it is asleep in a monitor job. The PID
will die and the thread asleep on the monitor will then wake
up releasing the job mutex.
* src/qemu/qemu_driver.c: Kill process before using qemuProcessStop
to ensure job is released
* src/qemu/qemu_process.c: Add virProcessKill for killing off
QEMU processes
Move the qemudStartVMDaemon and qemudShutdownVMDaemon
methods into a separate file, renaming them to
qemuProcessStart, qemuProcessStop. All helper methods
called by these are also moved & renamed to match
* src/Makefile.am: Add qemu_process.c/.h
* src/qemu/qemu_command.c: Add qemuDomainAssignPCIAddresses
* src/qemu/qemu_command.h: Add VNC port min/max
* src/qemu/qemu_domain.c, src/qemu/qemu_domain.h: Add
domain event queue helpers
* src/qemu/qemu_driver.c, src/qemu/qemu_driver.h: Remove
all QEMU process startup/shutdown functions
* src/qemu/qemu_process.c, src/qemu/qemu_process.h: Add
all QEMU process startup/shutdown functions
