Currently, kernel supports up to 8 queues for a multiqueue tap device.
However, if user tries to enter a huge number (e.g. one million) the tap
allocation fails, as expected. But what is not expected is the log full
of warnings:
warning : virFileClose:83 : Tried to close invalid fd 0
The problem is, upon error we iterate over an array of FDs (handlers to
queues) and VIR_FORCE_CLOSE() over each item. However, the array is
pre-filled with zeros. Hence, we repeatedly close stdin. Ouch.
But there's more. The queues allocation is done in virNetDevTapCreate()
which cleans up the FDs in case of error. Then, its caller, the
virNetDevTapCreateInBridgePort() iterates over the FD array and tries to
close them too. And so does qemuNetworkIfaceConnect() and
qemuBuildInterfaceCommandLine().
One of my previous patches 5cfe0d37cd tried to handle the case when
libvirt is a submodule of another project. In that case, the .git is
just a link to the parent .git directory (which the autogen.sh script
didn't count on). The fix was missing 'test' though.
Similarly to qemu_driver.c, we can join often repeating code of looking
up network into one function: networkObjFromNetwork.
Signed-off-by: Michal Privoznik <mprivozn@redhat.com>
Since 785ff34bf8 we are using the outputStr variable in cleanup label.
However, there is a possibility to jump to the label before the variable
has been declared:
virsh-pool.c: In function 'cmdPoolList':
virsh-pool.c:1121:25: error: jump skips variable initialization [-Werror=jump-misses-init]
goto asprintf_failure;
^
virsh-pool.c:1308:1: note: label 'asprintf_failure' defined here
asprintf_failure:
^
virsh-pool.c:1267:11: note: 'outputStr' declared here
char *outputStr = NULL;
If there's no hard_limit set and domain uses VFIO we still must lock
the guest memory (prerequisite from qemu). Hence, we should compute
the amount to be locked from max_balloon.
If user requested multiqueue networking, beside multiple /dev/tap and
/dev/vhost-net openings, we forgot to pass mq=on onto the -device
virtio-net-pci command line. This is advised at:
http://www.linux-kvm.org/page/Multiqueue#Enable_MQ_feature
If there's no hard_limit set and domain uses VFIO we still must lock the
guest memory (prerequisite from qemu). Hence, we should compute the
amount to be locked from max_balloon.
Since 16bcb3 we have a regression. The hard_limit is set
unconditionally. By default the limit is zero. Hence, if user hasn't
configured any, we set the zero in cgroup subsystem making the kernel
kill the corresponding qemu process immediately. The proper fix is to
set hard_limit iff user has configured any.
In one of my previous patches I am removing the hard_limit heuristic to
guess the correct value if none set. However, it turned out, this limit
is hard to guess even for users. We should advise them to not set the
limit as their domains may be OOM killed. Sigh.
This function is to guess the correct limit for maximal memory
usage by qemu for given domain. This can never be guessed
correctly, not to mention all the pains and sleepless nights this
code has caused. Once somebody discovers algorithm to solve the
Halting Problem, we can compute the limit algorithmically. But
till then, this code should never see the light of the release
again.
https://bugzilla.redhat.com/show_bug.cgi?id=979477
Since 1.0.3 we are using the new way to copy non shared storage during
migration (the NBD way). However, whether the new or old way is used is
not controllable by user but unconditionally turned on if both sides of
migration support it. Moreover, the implementation is not complete: the
combination for VIR_MIGRATE_TUNNELLED flag is missing (as we need to
open new port on the destination) in which case we just error out. This
is a deadly combination: not letting users choose their destiny and
erroring out. We should not do that but VIR_WARN and turn the NBD off
instead.
This configuration knob lets user to set the length of queue of
connection requests waiting to be accept()-ed by the daemon. IOW, it
just controls the @backlog passed to listen:
int listen(int sockfd, int backlog);
Currently, even if max_client limit is hit, we accept() incoming
connection request, but close it immediately. This has disadvantage of
not using listen() queue. We should accept() only those clients we
know we can serve and let all other wait in the (limited) queue.
The outbound/@peak is ignored (since QoS was introduced). This is due to
kernel limitation of know allowing ingress filters to have peak just
average rate. However, we should document this limitation to not confuse
users.
Interestingly, we had <code>floor<code> ... </code>outbound</code> which
results in much larger block of text to be written in code style that
intended.
There's a race in lxc driver causing a deadlock. If a domain is
destroyed immediately after started, the deadlock can occur. When domain
is started, the even loop tries to connect to the monitor. If the
connecting succeeds, virLXCProcessMonitorInitNotify() is called with
@mon->client locked. The first thing that callee does, is
virObjectLock(vm). So the order of locking is: 1) @mon->client, 2) @vm.
However, if there's another thread executing virDomainDestroy on the
very same domain, the first thing done here is locking the @vm. Then,
the corresponding libvirt_lxc process is killed and monitor is closed
via calling virLXCMonitorClose(). This callee tries to lock @mon->client
too. So the order is reversed to the first case. This situation results
in deadlock and unresponsive libvirtd (since the eventloop is involved).
The proper solution is to unlock the @vm in virLXCMonitorClose prior
entering virNetClientClose(). See the backtrace as follows:
Thread 25 (Thread 0x7f1b7c9b8700 (LWP 16312)):
0 0x00007f1b80539714 in __lll_lock_wait () from /lib64/libpthread.so.0
1 0x00007f1b8053516c in _L_lock_516 () from /lib64/libpthread.so.0
2 0x00007f1b80534fbb in pthread_mutex_lock () from /lib64/libpthread.so.0
3 0x00007f1b82a637cf in virMutexLock (m=0x7f1b3c0038d0) at util/virthreadpthread.c:85
4 0x00007f1b82a4ccf2 in virObjectLock (anyobj=0x7f1b3c0038c0) at util/virobject.c:320
5 0x00007f1b82b861f6 in virNetClientCloseInternal (client=0x7f1b3c0038c0, reason=3) at rpc/virnetclient.c:696
6 0x00007f1b82b862f5 in virNetClientClose (client=0x7f1b3c0038c0) at rpc/virnetclient.c:721
7 0x00007f1b6ee12500 in virLXCMonitorClose (mon=0x7f1b3c007210) at lxc/lxc_monitor.c:216
8 0x00007f1b6ee129f0 in virLXCProcessCleanup (driver=0x7f1b68100240, vm=0x7f1b680ceb70, reason=VIR_DOMAIN_SHUTOFF_DESTROYED) at lxc/lxc_process.c:174
9 0x00007f1b6ee14106 in virLXCProcessStop (driver=0x7f1b68100240, vm=0x7f1b680ceb70, reason=VIR_DOMAIN_SHUTOFF_DESTROYED) at lxc/lxc_process.c:710
10 0x00007f1b6ee1aa36 in lxcDomainDestroyFlags (dom=0x7f1b5c002560, flags=0) at lxc/lxc_driver.c:1291
11 0x00007f1b6ee1ab1a in lxcDomainDestroy (dom=0x7f1b5c002560) at lxc/lxc_driver.c:1321
12 0x00007f1b82b05be5 in virDomainDestroy (domain=0x7f1b5c002560) at libvirt.c:2303
13 0x00007f1b835a7e85 in remoteDispatchDomainDestroy (server=0x7f1b857419d0, client=0x7f1b8574ae40, msg=0x7f1b8574acf0, rerr=0x7f1b7c9b7c30, args=0x7f1b5c004a50) at remote_dispatch.h:3143
14 0x00007f1b835a7d78 in remoteDispatchDomainDestroyHelper (server=0x7f1b857419d0, client=0x7f1b8574ae40, msg=0x7f1b8574acf0, rerr=0x7f1b7c9b7c30, args=0x7f1b5c004a50, ret=0x7f1b5c0029e0) at remote_dispatch.h:3121
15 0x00007f1b82b93704 in virNetServerProgramDispatchCall (prog=0x7f1b8573af90, server=0x7f1b857419d0, client=0x7f1b8574ae40, msg=0x7f1b8574acf0) at rpc/virnetserverprogram.c:435
16 0x00007f1b82b93263 in virNetServerProgramDispatch (prog=0x7f1b8573af90, server=0x7f1b857419d0, client=0x7f1b8574ae40, msg=0x7f1b8574acf0) at rpc/virnetserverprogram.c:305
17 0x00007f1b82b8c0f6 in virNetServerProcessMsg (srv=0x7f1b857419d0, client=0x7f1b8574ae40, prog=0x7f1b8573af90, msg=0x7f1b8574acf0) at rpc/virnetserver.c:163
18 0x00007f1b82b8c1da in virNetServerHandleJob (jobOpaque=0x7f1b8574dca0, opaque=0x7f1b857419d0) at rpc/virnetserver.c:184
19 0x00007f1b82a64158 in virThreadPoolWorker (opaque=0x7f1b8573cb10) at util/virthreadpool.c:144
20 0x00007f1b82a63ae5 in virThreadHelper (data=0x7f1b8574b9f0) at util/virthreadpthread.c:161
21 0x00007f1b80532f4a in start_thread () from /lib64/libpthread.so.0
22 0x00007f1b7fc4f20d in clone () from /lib64/libc.so.6
Thread 1 (Thread 0x7f1b83546740 (LWP 16297)):
0 0x00007f1b80539714 in __lll_lock_wait () from /lib64/libpthread.so.0
1 0x00007f1b8053516c in _L_lock_516 () from /lib64/libpthread.so.0
2 0x00007f1b80534fbb in pthread_mutex_lock () from /lib64/libpthread.so.0
3 0x00007f1b82a637cf in virMutexLock (m=0x7f1b680ceb80) at util/virthreadpthread.c:85
4 0x00007f1b82a4ccf2 in virObjectLock (anyobj=0x7f1b680ceb70) at util/virobject.c:320
5 0x00007f1b6ee13bd7 in virLXCProcessMonitorInitNotify (mon=0x7f1b3c007210, initpid=4832, vm=0x7f1b680ceb70) at lxc/lxc_process.c:601
6 0x00007f1b6ee11fd3 in virLXCMonitorHandleEventInit (prog=0x7f1b3c001f10, client=0x7f1b3c0038c0, evdata=0x7f1b8574a7d0, opaque=0x7f1b3c007210) at lxc/lxc_monitor.c:109
7 0x00007f1b82b8a196 in virNetClientProgramDispatch (prog=0x7f1b3c001f10, client=0x7f1b3c0038c0, msg=0x7f1b3c003928) at rpc/virnetclientprogram.c:259
8 0x00007f1b82b87030 in virNetClientCallDispatchMessage (client=0x7f1b3c0038c0) at rpc/virnetclient.c:1019
9 0x00007f1b82b876bb in virNetClientCallDispatch (client=0x7f1b3c0038c0) at rpc/virnetclient.c:1140
10 0x00007f1b82b87d41 in virNetClientIOHandleInput (client=0x7f1b3c0038c0) at rpc/virnetclient.c:1312
11 0x00007f1b82b88f51 in virNetClientIncomingEvent (sock=0x7f1b3c0044e0, events=1, opaque=0x7f1b3c0038c0) at rpc/virnetclient.c:1832
12 0x00007f1b82b9e1c8 in virNetSocketEventHandle (watch=3321, fd=54, events=1, opaque=0x7f1b3c0044e0) at rpc/virnetsocket.c:1695
13 0x00007f1b82a272cf in virEventPollDispatchHandles (nfds=21, fds=0x7f1b8574ded0) at util/vireventpoll.c:498
14 0x00007f1b82a27af2 in virEventPollRunOnce () at util/vireventpoll.c:645
15 0x00007f1b82a25a61 in virEventRunDefaultImpl () at util/virevent.c:273
16 0x00007f1b82b8e97e in virNetServerRun (srv=0x7f1b857419d0) at rpc/virnetserver.c:1097
17 0x00007f1b8359db6b in main (argc=2, argv=0x7ffff98dbaa8) at libvirtd.c:1512
Currently, in the autogen.sh script we check whether .git is an existing
directory in which case bootstrap is run. However, if libvirt is a
submodule, then the .git is just a file (with reference to the topmost
.git directory). However, our submodule routines work well. So there's
no real reason why we should prohibit users to build libvirt from
submodule.
With the majority of fields in the virLXCDriverPtr struct
now immutable or self-locking, there is no need for practically
any methods to be using the LXC driver lock. Only a handful
of helper APIs now need it.
The activeUsbHostdevs item in LXCDriver are lockable, but the lock has
to be called explicitly. Call the virObject(Un)Lock() in order to
achieve mutual exclusion once lxcDriverLock is removed.
The 'driver->caps' pointer can be changed on the fly. Accessing
it currently requires the global driver lock. Isolate this
access in a single helper, so a future patch can relax the
locking constraints.
Currently the virLXCDriverPtr struct contains an wide variety
of data with varying access needs. Move all the static config
data into a dedicated virLXCDriverConfigPtr object. The only
locking requirement is to hold the driver lock, while obtaining
an instance of virLXCDriverConfigPtr. Once a reference is held
on the config object, it can be used completely lockless since
it is immutable.
NB, not all APIs correctly hold the driver lock while getting
a reference to the config object in this patch. This is safe
for now since the config is never updated on the fly. Later
patches will address this fully.
If testQemuHotplugAttach succeeds, the vm->def steals the dev pointer.
However, not the envelope, which needs to be freed. In addition,
driver.config is allocated, but never freed.
Commit 75c1256 states that virGetGroupList must not be called
between fork and exec, then commit ee777e99 promptly violated
that for lxc.
Patch originally posted by Eric Blake <eblake@redhat.com>.
While generating seclabels, we check the seclabel stack if required
driver is in the stack. If not, an error is returned. However, it is
possible for a seclabel to not have any model set (happens with LXC
domains that have just <seclabel type='none'>). If that's the case,
we should just skip the iteration instead of calling STREQ(NULL, ...)
and SIGSEGV-ing subsequently.
Currently, if the primary security driver is 'none', we skip
initializing caps->host.secModels. This means, later, when LXC domain
XML is parsed and <seclabel type='none'/> is found (see
virSecurityLabelDefsParseXML), the model name is not copied to the
seclabel. This leads to subsequent crash in virSecurityManagerGenLabel
where we call STREQ() over the model (note, that we are expecting model
to be !NULL).
The test is currently testing just device update function. However,
chardev hotplug is implemented just for device attach and detach. This
fact means, the test needs to be rewritten (the majority of the code is
still shared). Moreover, we are now able to pass VM among multiple test
runs. So for instance, while we add a device in the first run, we can
remove it in the second run.
There are two levels on which a device may be hotplugged: config
and live. The config level requires just an insert or remove from
internal domain definition structure, which is exactly what this
patch does. There is currently no implementation for a chardev
update action, as there's not much to be updated. But more
importantly, the only thing that can be updated is path or socket
address by which chardevs are distinguished. So the update action
is currently not supported.
Now that we have callbacks, we should auto fill in omitted pieces of
information. It's important for chardev hotplug to fill in the correct
/{serial,parallel,console,channel}/target/@port if no value has been
provided by user.
The function being introduced is responsible for creating command
line argument for '-device' for given character device. Based on
the chardev type, it calls appropriate qemuBuild.*ChrDeviceStr(),
e.g. qemuBuildSerialChrDeviceStr() for serial chardev and so on.
The chardev alias assignment is going to be needed in a separate
places, so it should be moved into a separate function rather
than copying code randomly around.
The function being introduced is responsible for preparing and
executing 'chardev-add' qemu monitor command. Moreover, in case
of PTY chardev, the corresponding pty path is updated.
Currently, we are building InetSocketAddress qemu json type
within the qemuMonitorJSONNBDServerStart function. However, other
future functions may profit from the code as well. So it should
be moved into a static function.
For now, only these three helpers are needed:
virDomainChrFind - to find a duplicate chardev within VM def
virDomainChrInsert - wrapper for inserting a new chardev into VM def
virDomainChrRemove - wrapper for removing chardev from VM def
There is, however, one internal helper as well:
virDomainChrGetDomainPtrs which sets given pointers to one of
vmdef->{parallels,serials,consoles,channels} based on passed
chardev type.
With current code, error reporting for unsupported devices for hot plug,
unplug and update is total mess. The VIR_ERR_CONFIG_UNSUPPORTED error
code is reported instead of VIR_ERR_OPERATION_UNSUPPORTED. Moreover, the
error messages are not helping to find the root cause (lack of
implementation).
When adding a new domain device, it is fairly easy to forget to add
corresponding piece into virDomainDeviceDefParse. However, if the
internal structure is changed to one bit switch() the compiler will warn
about not handled enum item.
Not all device types are currently parsed in virDomainDeviceDefParse,
Since all needed functions do exist, nothing holds us back to make the
implementation complete. Similarly, the virDomainDeviceDefFree needs to
be updated as well.
Actually, I'm turning this function into a macro as filename,
function name and line number needs to be passed. The new
function virAsprintfInternal is introduced with the extended set
of arguments.
When removing a TAP device, the associated bandwidth settings are
removed. Currently, the /sbin/tc is used for that. It is spawned
several times. Moreover, we use the same @cmd variable to
construct the command and its arguments. That means we need to
virCommandFree(cmd); prior to each virCommandNew(TC); which
wasn't done.
After abf75aea24 the compiler screams:
qemu/qemu_driver.c: In function 'qemuNodeDeviceDetachFlags':
qemu/qemu_driver.c:10693:9: error: 'domain' may be used uninitialized in this function [-Werror=maybe-uninitialized]
pci = virPCIDeviceNew(domain, bus, slot, function);
^
qemu/qemu_driver.c:10693:9: error: 'bus' may be used uninitialized in this function [-Werror=maybe-uninitialized]
qemu/qemu_driver.c:10693:9: error: 'slot' may be used uninitialized in this function [-Werror=maybe-uninitialized]
qemu/qemu_driver.c:10693:9: error: 'function' may be used uninitialized in this function [-Werror=maybe-uninitialized]
Since the other functions qemuNodeDeviceReAttach and qemuNodeDeviceReset
looks exactly the same, I've initialized the variables there as well.
However, I am still wondering why those functions don't matter to gcc
while the first one does.
As my punishment for the break in 7f15ebc7 (fixed in 752596b5dd) I'm
introducing this test to make sure it won't happen again. Currently,
only test for <graphics/> is supported.
Currently, we have a bug when updating a graphics device. A graphics device can
have a listen address set. This address is either defined by user (in which case
it's type is VIR_DOMAIN_GRAPHICS_LISTEN_TYPE_ADDRESS) or it can be inherited
from a network (in which case it's type is
VIR_DOMAIN_GRAPHICS_LISTEN_TYPE_NETWORK). However, in both cases we have a
listen address to process (e.g. during migration, as I've tried to fix in
7f15ebc7).
Later, when a user tries to update the graphics device (e.g. set a password),
we check if listen addresses match the original as qemu doesn't know how to
change listen address yet. Hence, users are required to not change the listen
address. The implementation then just dumps listen addresses and compare them.
Previously, while dumping the listen addresses, NULL was returned for NETWORK.
After my patch, this is no longer true, and we get a listen address for olddev
even if it is a type of NETWORK. So we have a real string on one side, the NULL
from user's XML on the other side and hence we think user wants to change the
listen address and we refuse it.
Therefore, we must take the type of listen address into account as well.
Currently, we wait for SPICE to migrate in the very same loop where we
wait for qemu to migrate. This has a disadvantage of slowing seamless
migration down. One one hand, we should not kill the domain until all
SPICE data has been migrated. On the other hand, there is no need to
wait in the very same loop and hence slowing down 'cont' on the
destination. For instance, if users are watching a movie, they can
experience the movie to be stopped for a couple of seconds, as
processors are not running nor on src nor on dst as libvirt waits for
SPICE to migrate. We should move the waiting phase to migration CONFIRM
phase.
https://bugzilla.redhat.com/show_bug.cgi?id=971485
As of d7f9d82753 we copy the listen
address from the qemu.conf config file in case none has been provided
via XML. But later, when migrating, we should not include such listen
address in the migratable XML as it is something autogenerated, not
requested by user. Moreover, the binding to the listen address will
likely fail, unless the address is '0.0.0.0' or its IPv6 equivalent.
This patch introduces a new boolean attribute to virDomainGraphicsListenDef
to distinguish autofilled listen addresses. However, we must keep the
attribute over libvirtd restarts, so it must be kept within status XML.
Currently, there's a path to use the ncpuinfo variable uninitialized,
which leads to a compiler warning:
qemu/qemu_driver.c: In function 'qemuDomainGetVcpusFlags':
qemu/qemu_driver.c:4573:9: error: 'ncpuinfo' may be used
uninitialized in this function [-Werror=maybe-uninitialized]
for (i = 0; i < ncpuinfo; i++) {
^
Michal Privoznik