NIC_RX_FILTER_CHANGED is sent by qemu any time a NIC driver in the
guest modified the NIC's RX Filter (for example, if the MAC address of
the NIC is changed by the guest).
This patch doesn't do anything useful with that event; it just sets up
all the plumbing to get news of the event into a worker thread with
all proper locking/reference counting, and provide an easy place to
add in desired functionality.
See src/qemu/EVENTHANDLERS.txt for information/instructions on adding
a libvirt-internal handler for a qemu event (using
NIC_RX_FILTER_CHANGED as an example).
This function can be called at any time to get the current status of a
guest's network device rx-filter. In particular it is useful to call
after libvirt recieves a NIC_RX_FILTER_CHANGED event - this event only
tells you that something has changed in the rx-filter, the details are
retrieved with the query-rx-filter monitor command (only available in
the json monitor). The command sent to the qemu monitor looks like this:
{"execute":"query-rx-filter", "arguments": {"name":"net2"} }'
and the results will look something like this:
{
"return": [
{
"promiscuous": false,
"name": "net2",
"main-mac": "52:54:00:98:2d:e3",
"unicast": "normal",
"vlan": "normal",
"vlan-table": [
42,
0
],
"unicast-table": [
],
"multicast": "normal",
"multicast-overflow": false,
"unicast-overflow": false,
"multicast-table": [
"33:33:ff:98:2d:e3",
"01:80:c2:00:00:21",
"01:00:5e:00:00:fb",
"33:33:ff:98:2d:e2",
"01:00:5e:00:00:01",
"33:33:00:00:00:01"
],
"broadcast-allowed": false
}
],
"id": "libvirt-14"
}
This is all parsed from JSON into a virNetDevRxFilter object for
easier consumption. (unicast-table is usually empty, but is also an
array of mac addresses similar to multicast-table).
(NB: LIBNL_CFLAGS was added to tests/Makefile.am because virnetdev.h
now includes util/virnetlink.h, which includes netlink/msg.h when
appropriate. Without LIBNL_CFLAGS, gcc can't find that file (if
libnl/netlink isn't available, LIBNL_CFLAGS will be empty and
virnetlink.h won't try to include netlink/msg.h anyway).)
While our code gathers block stats via "query-blockstats" some
information need to be gathered via "query-block". Add a helper function
that will update the blockstats structure if requested.
The current block stats code matched up the disk name with the actual
stats by the order in the data returned from qemu. This unfortunately
isn't right as qemu may return the disks in any order. Fix this by
returning a hash of stats and index them by the disk alias.
Currently we only support TCP protocol for native QEMU migration but
this is going to be changed. Let's make the code more general and remove
hardcoded TCP protocol from several places.
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
If the qemu being used doesn't support JSON, then querying for IOThread
data would fail. In that case, ensure the *iothreads is NULL and return 0
as the count of iothreads available.
This patch implements the VIR_DOMAIN_STATS_BLOCK group of statistics.
To do so, a helper function to get the block stats of all the disks of
a domain is added.
Signed-off-by: Francesco Romani <fromani@redhat.com>
Signed-off-by: Peter Krempa <pkrempa@redhat.com>
Coverity complains about the calculation of the buf & len within
the PROBE macro. So to quiet things down, do the calculation prior
to usage in either write() or qemuMonitorIOWriteWithFD() calls and
then have the PROBE use the calculated values - which works.
Upstream qemu 1.4 added some drive-mirror tunables not present
when it was first introduced in 1.3. Management apps may want
to set these in some cases (for example, without tuning
granularity down to sector size, a copy may end up occupying
more bytes than the original because an entire cluster is
copied even when only a sector within the cluster is dirty,
although tuning it down results in more CPU time to do the
copy). I haven't personally needed to use the parameters, but
since they exist, and since the new API supports virTypedParams,
we might as well expose them.
Since the tuning parameters aren't often used, and omitted from
the QMP command when unspecified, I think it is safe to rely on
qemu 1.3 to issue an error about them being unsupported, rather
than trying to create a new capability bit in libvirt.
Meanwhile, all versions of qemu from 1.4 to 2.1 have a bug where
a bad granularity (such as non-power-of-2) gives a poor message:
error: internal error: unable to execute QEMU command 'drive-mirror': Invalid parameter 'drive-virtio-disk0'
because of abuse of QERR_INVALID_PARAMETER (which is supposed to
name the parameter that was given a bad value, rather than the
value passed to some other parameter). I don't see that a
capability check will help, so we'll just live with it (and it
has since been improved in upstream qemu).
* src/qemu/qemu_monitor.h (qemuMonitorDriveMirror): Add
parameters.
* src/qemu/qemu_monitor.c (qemuMonitorDriveMirror): Likewise.
* src/qemu/qemu_monitor_json.h (qemuMonitorJSONDriveMirror):
Likewise.
* src/qemu/qemu_monitor_json.c (qemuMonitorJSONDriveMirror):
Likewise.
* src/qemu/qemu_driver.c (qemuDomainBlockCopyCommon): Likewise.
(qemuDomainBlockRebase, qemuDomainBlockCopy): Adjust callers.
* src/qemu/qemu_migration.c (qemuMigrationDriveMirror): Likewise.
* tests/qemumonitorjsontest.c (qemuMonitorJSONDriveMirror): Likewise.
Signed-off-by: Eric Blake <eblake@redhat.com>
When QEMU fails during incoming migration after we successfully started
it (i.e., during Perform or Finish phase), we report a rather unhelpful
message
Unable to read from monitor: Connection reset by peer
We already have a code that takes error messages from QEMU's error
output but we disable it once QEMU successfully starts. This patch
postpones this until the end of Finish phase during incoming migration
so that we can report a much better error message:
internal error: early end of file from monitor: possible problem:
Unknown savevm section or instance '0000:00:05.0/virtio-balloon' 0
load of migration failed
https://bugzilla.redhat.com/show_bug.cgi?id=1090093
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
While reviewing the new virDomainBlockCopy API, Peter Krempa
pointed out that our existing design of using MiB/s for block
job bandwidth is rather coarse, especially since qemu tracks
it in bytes/s; so virDomainBlockCopy only accepts bytes/s.
But once the new API is implemented for qemu, we will be in
the situation where it is possible to set a value that cannot
be accurately reflected back to the user, because the existing
virDomainGetBlockJobInfo defaults to the coarser units.
Fortunately, we have an escape hatch; and one that has already
served us well in the past: we can use the flags argument to
specify which scale to use (see virDomainBlockResize for prior
art). This patch fixes the query side of the API; made easier
by previous patches that split the query side out from the
modification code. Later patches will address the virsh
interface, as well retrofitting all other blockjob APIs to
also accept a flag for toggling bandwidth units.
* include/libvirt/libvirt.h.in (_virDomainBlockJobInfo)
(VIR_DOMAIN_BLOCK_COPY_BANDWIDTH): Document sizing issues.
(virDomainBlockJobInfoFlags): New enum.
* src/libvirt.c (virDomainGetBlockJobInfo): Document new flag.
* src/qemu/qemu_monitor.h (qemuMonitorBlockJobInfo): Add parameter.
* src/qemu/qemu_monitor.c (qemuMonitorBlockJobInfo): Likewise.
* src/qemu/qemu_monitor_json.h (qemuMonitorJSONBlockJobInfo):
Likewise.
* src/qemu/qemu_monitor_json.c (qemuMonitorJSONBlockJobInfo)
(qemuMonitorJSONGetBlockJobInfoOne): Likewise. Don't scale here.
* src/qemu/qemu_migration.c (qemuMigrationDriveMirror): Update
callers.
* src/qemu/qemu_driver.c (qemuDomainBlockPivot)
(qemuDomainBlockJobImpl): Likewise.
(qemuDomainGetBlockJobInfo): Likewise, and support new flag.
Signed-off-by: Eric Blake <eblake@redhat.com>
qemu treats blockjob bandwidth as a 64-bit number, in the units
of bytes/second. But we stupidly modeled block job bandwidth
after migration bandwidth, which in turn was an 'unsigned long'
and therefore subject to 32-bit vs. 64-bit interpretations, and
with a scale of MiB/s. Our code already has to convert between
the two scales, and report overflow as appropriate; although
this conversion currently lives in the monitor code. In fact,
our conversion code limited things to 63 bits, because we
checked against LLONG_MAX and reject what would be negative
bandwidth if treated as signed.
On the bright side, our use of MiB/s means that even with a
32-bit unsigned long, we still have no problem representing a
bandwidth of 2GiB/s, which is starting to be more feasible as
10-gigabit or even faster interfaces are used. And once you
get past the physical speeds of existing interfaces, any larger
bandwidth number behaves the same - effectively unlimited.
But on the low side, the granularity of 1MiB/s tuning is rather
coarse. So the new virDomainBlockJob API decided to go with
a direct 64-bit bytes/sec number instead of the scaled number
that prior blockjob APIs had used. But there is no point in
rounding this number to MiB/s just to scale it back to bytes/s
for handing to qemu.
In order to make future code sharing possible between the old
virDomainBlockRebase and the new virDomainBlockCopy, this patch
moves the scaling and overflow detection into the driver code.
Several of the block job calls that can set speed are fed
through a common interface, so it was easier to adjust all block
jobs at once, for consistency. This patch is just code motion;
there should be no user-visible change in behavior.
* src/qemu/qemu_monitor.h (qemuMonitorBlockJob)
(qemuMonitorBlockCommit, qemuMonitorDriveMirror): Change
parameter type and scale.
* src/qemu/qemu_monitor.c (qemuMonitorBlockJob)
(qemuMonitorBlockCommit, qemuMonitorDriveMirror): Move scaling
and overflow detection...
* src/qemu/qemu_driver.c (qemuDomainBlockJobImpl)
(qemuDomainBlockRebase, qemuDomainBlockCommit): ...here.
(qemuDomainBlockCopy): Use bytes/sec.
Signed-off-by: Eric Blake <eblake@redhat.com>
Another layer of overly-multiplexed code that deserves to be
split into obviously separate paths for query vs. modify.
This continues the cleanup started in commit cefe0ba.
In the process, make some tweaks to simplify the logic when
parsing the JSON reply. There should be no user-visible
semantic changes.
* src/qemu/qemu_monitor.h (qemuMonitorBlockJob): Drop parameter.
(qemuMonitorBlockJobInfo): New prototype.
(BLOCK_JOB_INFO): Drop enum.
* src/qemu/qemu_monitor_json.h (qemuMonitorJSONBlockJob)
(qemuMonitorJSONBlockJobInfo): Likewise.
* src/qemu/qemu_monitor.c (qemuMonitorBlockJob): Split...
(qemuMonitorBlockJobInfo): ...into second function.
* src/qemu/qemu_monitor_json.c (qemuMonitorJSONBlockJob): Move
block info portions...
(qemuMonitorJSONGetBlockJobInfo): ...here, and rename...
(qemuMonitorJSONBlockJobInfo): ...and export.
(qemuMonitorJSONGetBlockJobInfoOne): Alter return semantics.
* src/qemu/qemu_driver.c (qemuDomainBlockPivot)
(qemuDomainBlockJobImpl, qemuDomainGetBlockJobInfo): Adjust
callers.
* src/qemu/qemu_migration.c (qemuMigrationDriveMirror)
(qemuMigrationCancelDriveMirror): Likewise.
Signed-off-by: Eric Blake <eblake@redhat.com>
https://bugzilla.redhat.com/show_bug.cgi?id=1103245
An advice appeared there on the qemu-devel list [1]. When a domain is
suspended and then resumed guest kernel is not aware of this. So we've
introduced virDomainSetTime API that resets the time within guest
using qemu-ga. On the other hand, qemu itself is trying to make RTC
beat faster to catch the difference. But if we don't tell qemu that
guest's time was reset via the other method, both mechanisms are
applied resulting in again wrong guest time. In order to avoid summing
both corrections we need to tell qemu that it should not use the RTC
injection if the guest time is set via guest agent.
1: http://www.mail-archive.com/qemu-devel@nongnu.org/msg236435.html
Signed-off-by: Michal Privoznik <mprivozn@redhat.com>
That can be lately achieved with by having .param == NULL in the
virQEMUCapsCommandLineProps struct.
Signed-off-by: Martin Kletzander <mkletzan@redhat.com>
To allow changing the name that is recorded in the top of the current
image chain used in a block pull/rebase operation, we need to specify
the backing name to qemu. This is done via the "backing-file" attribute
to the block-stream commad.
To allow changing the name that is recorded in the overlay of the TOP
image used in a block commit operation, we need to specify the backing
name to qemu. This is done via the "backing-file" attribute to the
block-commit command.
We are about to turn on support for active block commit. Although
qemu 2.0 was the first version to mostly support it, that version
mis-handles 0-length files, and doesn't have anything available for
easy probing. But qemu 2.1 fixed bugs, and made life simpler by
letting the 'top' argument be optional. Unless someone begs for
active commit with qemu 2.0, for now we are just going to enable
it only by probing for qemu 2.1 behavior (anyone backporting active
commit can also backport the optional argument behavior). This
requires qemu.git commit 7676e2c597000eff3a7233b40cca768b358f9bc9.
Although all our actual uses of block-commit supply arguments for
both base and top, we can omit both arguments and use a bogus
device string to trigger an interesting behavior in qemu. All QMP
commands first do argument validation, failing with GenericError
if a mandatory argument is missing. Once that passes, the code
in the specific command gets to do further checking, and the qemu
developers made sure that if device is the only supplied argument,
then the block-commit code will look up the device first, with a
failure of DeviceNotFound, before attempting any further argument
validation (most other validations fail with GenericError). Thus,
the category of error class can reliably be used to decipher
whether the top argument was optional, which in turn implies a
working active commit. Since we expect our bogus device string to
trigger an error either way, the code is written to return a
distinct return value without spamming the logs.
* src/qemu/qemu_monitor.h (qemuMonitorSupportsActiveCommit): New
prototype.
* src/qemu/qemu_monitor.c (qemuMonitorSupportsActiveCommit):
Implement it.
* src/qemu/qemu_monitor_json.h (qemuMonitorJSONBlockCommit):
Allow NULL for top and base, for probing purposes.
* src/qemu/qemu_monitor_json.c (qemuMonitorJSONBlockCommit):
Likewise, implementing the probe.
* tests/qemumonitorjsontest.c (mymain): Enable...
(testQemuMonitorJSONqemuMonitorSupportsActiveCommit): ...a new test.
Signed-off-by: Eric Blake <eblake@redhat.com>
Replace:
if (virBufferError(&buf)) {
virBufferFreeAndReset(&buf);
virReportOOMError();
...
}
with:
if (virBufferCheckError(&buf) < 0)
...
This should not be a functional change (unless some callers
misused the virBuffer APIs - a different error would be reported
then)
In "src/conf/domain_conf.h" there are many enum declarations. The
cleanup in this header filer was started, but it wasn't enough and
there are many other files that has enum variables declared. So, the
commit was starting to be big. This commit finish the cleanup in this
header file and in other files that has enum variables, parameters,
or functions declared.
Signed-off-by: Julio Faracco <jcfaracco@gmail.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
When passing migration bandwidth to QEMU, we multiply it by 1024 * 1024
to convert the speed to B/s and the result still needs to fit in
int64_t.
https://bugzilla.redhat.com/show_bug.cgi?id=1083483
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
I almost wrote a hash value free function that just called
VIR_FREE, then realized I couldn't be the first person to
do that. Sure enough, it was worth factoring into a common
helper routine.
* src/util/virhash.h (virHashValueFree): New function.
* src/util/virhash.c (virHashValueFree): Implement it.
* src/util/virobject.h (virObjectFreeHashData): New function.
* src/libvirt_private.syms (virhash.h, virobject.h): Export them.
* src/nwfilter/nwfilter_learnipaddr.c (virNWFilterLearnInit): Use
common function.
* src/qemu/qemu_capabilities.c (virQEMUCapsCacheNew): Likewise.
* src/qemu/qemu_command.c (qemuDomainCCWAddressSetCreate):
Likewise.
* src/qemu/qemu_monitor.c (qemuMonitorGetBlockInfo): Likewise.
* src/qemu/qemu_process.c (qemuProcessWaitForMonitor): Likewise.
* src/util/virclosecallbacks.c (virCloseCallbacksNew): Likewise.
* src/util/virkeyfile.c (virKeyFileParseGroup): Likewise.
* tests/qemumonitorjsontest.c
(testQemuMonitorJSONqemuMonitorJSONGetBlockInfo): Likewise.
Signed-off-by: Eric Blake <eblake@redhat.com>
This patch adds qemuMonitorGetDumpGuestMemoryCapability, which is used to check
whether the specified dump-guest-memory format is supported by qemu.
Signed-off-by: Qiao Nuohan <qiaonuohan@cn.fujitsu.com>
Busy enterprise workloads hosted on large sized VM's tend to dirty
memory faster than the transfer rate achieved via live guest migration.
Despite some good recent improvements (& using dedicated 10Gig NICs
between hosts) the live migration may NOT converge.
Recently support was added in qemu (version 1.6) to allow a user to
choose if they wish to force convergence of their migration via a
new migration capability : "auto-converge". This feature allows for qemu
to auto-detect lack of convergence and trigger a throttle-down of the
VCPUs.
This patch includes the libvirt support needed to trigger this
feature. (Testing is in progress)
Signed-off-by: Chegu Vinod <chegu_vinod@hp.com>
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
https://bugzilla.redhat.com/show_bug.cgi?id=844378
When qemu dies early after connecting to its monitor but before we
actually try to read something from the monitor, we would just fail
domain start with useless message:
"An error occurred, but the cause is unknown"
This is because the real error gets reported in a monitor EOF handler
executing within libvirt's event loop.
The fix is to take any error set in qemuMonitor structure and propagate
it into the thread-local error when qemuMonitorClose is called and no
thread-local error is set.
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
Wire up all the pieces to send arbitrary qemu events to a
client using libvirt-qemu.so. If the extra bookkeeping of
generating event objects even when no one is listening turns
out to be noticeable, we can try to further optimize things
by adding a counter for how many connections are using events,
and only dump events when the counter is non-zero; but for
now, I didn't think it was worth the code complexity.
* src/qemu/qemu_driver.c
(qemuConnectDomainQemuMonitorEventRegister)
(qemuConnectDomainQemuMonitorEventDeregister): New functions.
* src/qemu/qemu_monitor.h (qemuMonitorEmitEvent): New prototype.
(qemuMonitorDomainEventCallback): New typedef.
* src/qemu/qemu_monitor_json.c (qemuMonitorJSONIOProcessEvent):
Report events.
* src/qemu/qemu_monitor.c (qemuMonitorEmitEvent): New function, to
pass events through.
* src/qemu/qemu_process.c (qemuProcessHandleEvent): Likewise.
Signed-off-by: Eric Blake <eblake@redhat.com>
Any source file which calls the logging APIs now needs
to have a VIR_LOG_INIT("source.name") declaration at
the start of the file. This provides a static variable
of the virLogSource type.
Signed-off-by: Daniel P. Berrange <berrange@redhat.com>
The dtrace probe macros rely on the logging API. We can't make
the internal.h header include the virlog.h header though since
that'd be a circular include. Instead simply split the dtrace
probes into their own header file, since there's no compelling
reason for them to be in the main internal.h header.
Signed-off-by: Daniel P. Berrange <berrange@redhat.com>
While investigating https://bugzilla.redhat.com/show_bug.cgi?id=1061827
I noticed that we pass user input unscathed for block-pull, but
always pass a canonical absolute name through for block-commit.
[Note that we probably _ought_ to validate that the user's request
for block-pull actually matches the backing chain, the way we already
do for block-commit - but that's a separate issue. Further note that
the ability to pass user input through unscathed allows backdoors
such as specifying a backing image that is a network URI such as
a gluster disk, instead of forcing things to the local file system;
which is an area still under active investigation on whether libvirt
needs to behave differently for network disks.]
Since qemu may write the name that the user passed in as the backing
file, a user may have a reason to want a relative file name passed
through to qemu, and always munging things to absolute prevents that.
Put another way, if you have the backing chain:
[A] <- [B(back=./A)] <- [C(back=./B)]
and commit B into A (virsh blockcommit $dom vda --base A --top B),
the metadata of C will have to be re-written. But should it be
rewritten as [C(back=./A)] or as [C(back=/path/to/A)]? Still up in
the air is whether qemu's decision should be based on whether B
and/or C had relative paths, or on whether the --base and/or
--top arguments to the command were relative paths; but if we always
pass a canonical name, we've prevented the spelling of the command
arguments from being part of the hueristics that qemu uses.
I also audited the code, and verified that we never call
qemuMonitorBlockCommit() with a NULL base, either before or after
the change to qemu_driver.c.
* src/qemu/qemu_driver.c (qemuDomainBlockCommit): Preserve user's
spelling, since absolute vs. relative matters to qemu.
* src/qemu/qemu_monitor.h (qemuMonitorBlockCommit): Base is never
null.
* src/qemu/qemu_monitor.c (qemuMonitorBlockCommit): Likewise.
* src/qemu/qemu_monitor_json.h (qemuMonitorJSONBlockCommit):
Likewise.
* src/qemu/qemu_monitor_json.c (qemuMonitorJSONBlockCommit):
Likewise.
Signed-off-by: Eric Blake <eblake@redhat.com>
If virDomainMemoryStats was run on a domain with virtio balloon driver
running on an old qemu which supports QMP but does not support qom-list
QMP command, libvirtd would crash. The reason is we did not check if
qemuMonitorJSONGetObjectListPaths failed and moreover we even stored its
result in an unsigned integer type.
There is a number of reported issues when we fail starting a domain.
Turns out that, in some scenarios like high load, 3 second timeout is
not enough for qemu to start up to the phase where the socket is
created. Since there is no downside of waiting longer, raise the
timeout right to 30 seconds.
Signed-off-by: Martin Kletzander <mkletzan@redhat.com>
Recent changes to events (commit 8a29ffcf) resulted in new compile
failures on some targets (such as ARM OMAP5):
conf/domain_event.c: In function 'virDomainEventDispatchDefaultFunc':
conf/domain_event.c:1198:30: error: cast increases required alignment of
target type [-Werror=cast-align]
conf/domain_event.c:1314:34: error: cast increases required alignment of
target type [-Werror=cast-align]
cc1: all warnings being treated as errors
The error is due to alignment; the base class is merely aligned
to the worst of 'int' and 'void*', while the child class must
be aligned to a 'long long'. The solution is to include a
'long long' (and for good measure, a function pointer) in the
base class to ensure correct alignment regardless of what a
child class may add, but to wrap the inclusion in a union so
as to not incur any wasted space. On a typical x86_64 platform,
the base class remains 16 bytes; on i686, the base class remains
12 bytes; and on the impacted ARM platform, the base class grows
from 12 bytes to 16 bytes due to the increase of alignment from
4 to 8 bytes.
Reported by Michele Paolino and others.
* src/util/virobject.h (_virObject): Use a union to ensure that
subclasses never have stricter alignment than the parent.
* src/util/virobject.c (virObjectNew, virObjectUnref)
(virObjectRef): Adjust clients.
* src/libvirt.c (virConnectRef, virDomainRef, virNetworkRef)
(virInterfaceRef, virStoragePoolRef, virStorageVolRef)
(virNodeDeviceRef, virSecretRef, virStreamRef, virNWFilterRef)
(virDomainSnapshotRef): Likewise.
* src/qemu/qemu_monitor.c (qemuMonitorOpenInternal)
(qemuMonitorClose): Likewise.
Signed-off-by: Eric Blake <eblake@redhat.com>
Most of our code base uses space after comma but not before;
fix the remaining uses before adding a syntax check.
* src/qemu/qemu_cgroup.c: Consistently use commas.
* src/qemu/qemu_command.c: Likewise.
* src/qemu/qemu_conf.c: Likewise.
* src/qemu/qemu_driver.c: Likewise.
* src/qemu/qemu_monitor.c: Likewise.
Signed-off-by: Eric Blake <eblake@redhat.com>
QEMU 1.6.0 introduced new migration status: setup
Libvirt does not expect such string in QMP and refuses to migrate with error
"unexpected migration status in setup"
This patch fixes it.
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
https://bugzilla.redhat.com/show_bug.cgi?id=1018267
The aim of virObject refing and urefing is to tell where the object is
to be used and when is no longer needed. Hence any object shouldn't be
used after it has been unrefed, as we might be the last to hold the
reference. The better way is to call virObjectUnref() *after* the last
object usage. In this specific case, the monitor EOF handler was called
after the qemuMonitorIO called virObjectUnref. Not only that @mon was
disposed (which is not used in the handler anyway) but the @mon->vm
which is causing a SIGSEGV:
2013-11-15 10:17:54.425+0000: 20110: error : qemuMonitorIO:688 : internal error: early end of file from monitor: possible problem:
qemu-kvm: -incoming tcp:01.01.01.0:49152: Failed to bind socket: Cannot assign requested address
Program received signal SIGSEGV, Segmentation fault.
qemuProcessHandleMonitorEOF (mon=<optimized out>, vm=0x7fb728004170) at qemu/qemu_process.c:299
299 if (priv->beingDestroyed) {
(gdb) p *priv
Cannot access memory at address 0x0
(gdb) p vm
$1 = (virDomainObj *) 0x7fb728004170
(gdb) p *vm
$2 = {parent = {parent = {magic = 3735928559, refs = 0, klass = 0xdeadbeef}, lock = {lock = {__data = {__lock = 2, __count = 0, __owner = 20110, __nusers = 1, __kind = 0, __spins = 0, __list = {__prev = 0x0,
__next = 0x0}}, __size = "\002\000\000\000\000\000\000\000\216N\000\000\001", '\000' <repeats 26 times>, __align = 2}}}, pid = 0, state = {state = 0, reason = 0}, autostart = 0, persistent = 0,
updated = 0, def = 0x0, newDef = 0x0, snapshots = 0x0, current_snapshot = 0x0, hasManagedSave = false, privateData = 0x0, privateDataFreeFunc = 0x0, taint = 304}
Signed-off-by: Michal Privoznik <mprivozn@redhat.com>
The qemu monitor supports retrieval of actual CPUID bits presented to
the guest using QMP monitor. Add APIs to extract these information and
tests for them.
Signed-off-by: Peter Krempa <pkrempa@redhat.com>
Since the 90139a62 commit the error is copied into mon->lastError but
it's never freed from there.
==31989== 395 bytes in 1 blocks are definitely lost in loss record 877 of 978
==31989== at 0x4A06C2B: malloc (in /usr/lib64/valgrind/vgpreload_memcheck-amd64-linux.so)
==31989== by 0x7EAF129: strdup (in /lib64/libc-2.15.so)
==31989== by 0x50D586C: virStrdup (virstring.c:554)
==31989== by 0x50976C1: virCopyError (virerror.c:191)
==31989== by 0x5097A35: virCopyLastError (virerror.c:312)
==31989== by 0x114909A9: qemuMonitorIO (qemu_monitor.c:690)
==31989== by 0x509BEDE: virEventPollDispatchHandles (vireventpoll.c:501)
==31989== by 0x509C701: virEventPollRunOnce (vireventpoll.c:648)
==31989== by 0x509A620: virEventRunDefaultImpl (virevent.c:274)
==31989== by 0x520D21C: virNetServerRun (virnetserver.c:1112)
==31989== by 0x11F368: main (libvirtd.c:1513)
Signed-off-by: Michal Privoznik <mprivozn@redhat.com>
I've noticed a SIGSEGV-ing libvirtd on the destination when the qemu
died too quickly = in Prepare phase. What is happening here is:
1) [Thread 3493] We are in qemuMigrationPrepareAny() and calling
qemuProcessStart() which subsequently calls qemuProcessWaitForMonitor()
and qemuConnectMonitor(). So far so good. The qemuMonitorOpen()
succeeds, however switching monitor to QMP mode fails as qemu died
meanwhile. That is qemuMonitorSetCapabilities() returns -1.
2013-10-08 15:54:10.629+0000: 3493: debug : qemuMonitorSetCapabilities:1356 : mon=0x14a53da0
2013-10-08 15:54:10.630+0000: 3493: debug : qemuMonitorJSONCommandWithFd:262 : Send command '{"execute":"qmp_capabilities","id":"libvirt-1"}' for write with FD -1
2013-10-08 15:54:10.630+0000: 3493: debug : virEventPollUpdateHandle:147 : EVENT_POLL_UPDATE_HANDLE: watch=17 events=13
...
2013-10-08 15:54:10.631+0000: 3493: debug : qemuMonitorSend:956 : QEMU_MONITOR_SEND_MSG: mon=0x14a53da0 msg={"execute":"qmp_capabilities","id":"libvirt-1"}
fd=-1
2013-10-08 15:54:10.631+0000: 3262: debug : virEventPollRunOnce:641 : Poll got 1 event(s)
2) [Thread 3262] The event loop is trying to do the talking to monitor.
However, qemu is dead already, remember?
2013-10-08 15:54:13.436+0000: 3262: error : qemuMonitorIORead:551 : Unable to read from monitor: Connection reset by peer
2013-10-08 15:54:13.516+0000: 3262: debug : virFileClose:90 : Closed fd 25
...
2013-10-08 15:54:13.533+0000: 3493: debug : qemuMonitorSend:968 : Send command resulted in error internal error: early end of file from monitor: possible problem:
3) [Thread 3493] qemuProcessStart() failed. No big deal. Go to the
'endjob' label and subsequently to the 'cleanup'. Since the domain is
not persistent and ret is -1, the qemuDomainRemoveInactive() is called.
This has an (unpleasant) effect of virObjectUnref()-in the @vm object.
Unpleasant because the event loop which is about to trigger EOF callback
still holds a pointer to the @vm (not the reference). See the valgrind
output below.
4) [Thread 3262] So the event loop starts triggering EOF:
2013-10-08 15:54:13.542+0000: 3262: debug : qemuMonitorIO:729 : Triggering EOF callback
2013-10-08 15:54:13.543+0000: 3262: debug : qemuProcessHandleMonitorEOF:294 : Received EOF on 0x14549110 'migt10'
And the monitor is cleaned up. This results in calling
qemuProcessHandleMonitorEOF with the @vm pointer passed. The pointer is
kept in qemuMonitor struct.
==3262== Thread 1:
==3262== Invalid read of size 4
==3262== at 0x77ECCAA: pthread_mutex_lock (in /lib64/libpthread-2.15.so)
==3262== by 0x52FAA06: virMutexLock (virthreadpthread.c:85)
==3262== by 0x52E3891: virObjectLock (virobject.c:320)
==3262== by 0x11626743: qemuProcessHandleMonitorEOF (qemu_process.c:296)
==3262== by 0x11642593: qemuMonitorIO (qemu_monitor.c:730)
==3262== by 0x52BD526: virEventPollDispatchHandles (vireventpoll.c:501)
==3262== by 0x52BDD49: virEventPollRunOnce (vireventpoll.c:648)
==3262== by 0x52BBC68: virEventRunDefaultImpl (virevent.c:274)
==3262== by 0x542D3D9: virNetServerRun (virnetserver.c:1112)
==3262== by 0x11F368: main (libvirtd.c:1513)
==3262== Address 0x14549128 is 24 bytes inside a block of size 136 free'd
==3262== at 0x4C2AF5C: free (in /usr/lib64/valgrind/vgpreload_memcheck-amd64-linux.so)
==3262== by 0x529B1FF: virFree (viralloc.c:580)
==3262== by 0x52E3703: virObjectUnref (virobject.c:270)
==3262== by 0x531557E: virDomainObjListRemove (domain_conf.c:2355)
==3262== by 0x1160E899: qemuDomainRemoveInactive (qemu_domain.c:2061)
==3262== by 0x1163A0C6: qemuMigrationPrepareAny (qemu_migration.c:2450)
==3262== by 0x1163A923: qemuMigrationPrepareDirect (qemu_migration.c:2626)
==3262== by 0x11682D71: qemuDomainMigratePrepare3Params (qemu_driver.c:10309)
==3262== by 0x53B0976: virDomainMigratePrepare3Params (libvirt.c:7266)
==3262== by 0x1502D3: remoteDispatchDomainMigratePrepare3Params (remote.c:4797)
==3262== by 0x12DECA: remoteDispatchDomainMigratePrepare3ParamsHelper (remote_dispatch.h:5741)
==3262== by 0x54322EB: virNetServerProgramDispatchCall (virnetserverprogram.c:435)
The mon->vm is set in qemuMonitorOpenInternal() which is the correct
place to increase @vm ref counter. The correct place to decrease the ref
counter is then qemuMonitorDispose().
Signed-off-by: Michal Privoznik <mprivozn@redhat.com>
Change the monitor error code to add the ability to access the qemu log
file using a file descriptor so that we can dig in it for a more useful
error message. The error is now logged on monitor hangups and overwrites
a possible lesser error. A hangup on the monitor usualy means that qemu
has crashed and there's a significant chance it produced a useful error
message.
The functionality will be latent until the next patch.
