"Instead of developing one CPU with 12 cores, the Magny Cours is
actually two 6 core “Bulldozer” CPUs combined in to one package"
I.e, each package has two NUMA nodes, and the two numa nodes share
the same core ID set (0-6), which means parsing the cores number
from sysfs doesn't work in this case.
And the wrong CPU number could cause three problems for libvirt:
1) performance lost
A domain without "cpuset" or "placement='auto'" (to drive numad)
specified will be only pinned to part of the CPUs.
2) domain can be started
If a domain uses numad, and the advisory nodeset returned from
numad contains node which exceeds the range of wrong total CPU
number. The domain will fail to start, as the bitmask passed to
sched_setaffinity could be fully filled with zero.
3) wrong CPU number affects lots of stuffs.
E.g. for command "virsh vcpuinfo", "virsh vcpupin", it will always
output with the truncated CPU list.
For more details:
https://www.redhat.com/archives/libvir-list/2012-May/msg00607.html
This patch is to fix the problem by parsing /proc/cpuinfo to get
the value of field "cpu cores", and use it as nodeinfo->cores if
it's greater than the cores number from sysfs.
Like for 'static' placement, when the memory policy mode is
'strict', set the memory policy by writing the advisory nodeset
returned from numad to cgroup file cpuset.mems,
On some of the NUMA platforms, the CPU index in each NUMA node
grows non-consecutive. While on other platforms, it can be inconsecutive,
E.g.
% numactl --hardware
available: 4 nodes (0-3)
node 0 cpus: 0 4 8 12 16 20 24 28
node 0 size: 131058 MB
node 0 free: 86531 MB
node 1 cpus: 1 5 9 13 17 21 25 29
node 1 size: 131072 MB
node 1 free: 127070 MB
node 2 cpus: 2 6 10 14 18 22 26 30
node 2 size: 131072 MB
node 2 free: 127758 MB
node 3 cpus: 3 7 11 15 19 23 27 31
node 3 size: 131072 MB
node 3 free: 127226 MB
node distances:
node 0 1 2 3
0: 10 20 20 20
1: 20 10 20 20
2: 20 20 10 20
3: 20 20 20 10
This patch is to fix the problem by using the CPU index in
caps->host.numaCell[i]->cpus[i] to set the bitmask instead of
assuming the CPU index of the NUMA nodes are always sequential.
For pseries guest, the default controller model is
ibmvscsi controller, this controller only can work
on spapr-vio address.
This patch is to assign spapr-vio address type to
ibmvscsi controller and correct vscsi test case.
Signed-off-by: Li Zhang <zhlcindy@linux.vnet.ibm.com>
Test 2 data grabbed from a 2-core 1-node laptop.
Test 3 data grabbed from a 48-cpu AMD Magny Cours box.
* tests/nodeinfodata/linux-nodeinfo-sysfs-test-2*: New test data.
* tests/nodeinfodata/linux-nodeinfo-sysfs-test-3*: Likewise.
* tests/nodeinfotest.c (mymain): Run them.
* cfg.mk
(exclude_file_name_regexp--sc_prohibit_empty_lines_at_EOF): Exempt
new test files.
We had previously weakened our nodeinfotest in order to ignore parsed
node values, because the parse function was mistakenly relying on
host files. A better fix is to avoid using the numactl library, but
to instead parse the same files that numactl would read, all while
allowing the files to be relative to our choice of directory.
* src/nodeinfo.c (CPU_SYS_PATH, NODE_SYS_PATH): Replace with...
(SYSFS_SYSTEM_PATH): ...parent directory.
(linuxNodeInfoCPUPopulate): Check NUMA nodes from requested
directory (by inlining numactl code).
(nodeGetCPUmap, nodeGetMemoryStats): Adjust macro use.
* tests/nodeinfotest.c (linuxTestCompareFiles, linuxTestNodeInfo):
Update test to match.
We were wasting time to malloc a copy of a constant string, then
copy it into static storage, for every call to nodeGetInfo. At
least we were lucky that it was a constant source, and thus not
subject to even worse issues with one thread clobbering the static
storage while another was using it. This gets rid of the waste,
by passing the string through the stack instead, as well as renaming
internal functions to better match our conventions.
* src/nodeinfo.c (sysfs_path): Delete.
(get_cpu_value, count_thread_siblings, parse_socket): Add
parameter, and rename...
(virNodeGetCpuValue, virNodeCountThreadSiblings)
(virNodeParseSocket): ... into a common namespace.
(cpu_online, parse_core): Inline into callers.
(linuxNodeInfoCPUPopulate): Update caller.
(nodeGetInfo): Drop a useless malloc.
Commit cdce2f42d tried to silence a compiler warning on 32-bit builds,
but the gcc shipped with RHEL 5 is old enough that the type conversion
via multiplication by 1 was insufficient for the task.
* src/qemu/qemu_monitor.c (qemuMonitorBlockJob): Previous attempt
didn't get past all gcc versions.
As defined in:
http://standards.freedesktop.org/basedir-spec/basedir-spec-latest.html
This offers a number of advantages:
* Allows sharing a home directory between different machines, or
sessions (eg. using NFS)
* Cleanly separates cache, runtime (eg. sockets), or app data from
user settings
* Supports performing smart or selective migration of settings
between different OS versions
* Supports reseting settings without breaking things
* Makes it possible to clear cache data to make room when the disk
is filling up
* Allows us to write a robust and efficient backup solution
* Allows an admin flexibility to change where data and settings are stored
* Dramatically reduces the complexity and incoherence of the
system for administrators
Appending an item to a list transfers ownership of that item to the
list owner. But an error can occur in between item allocation and
appending it to the list. In this case the item has to be freed
explicitly. This was not done in some special cases resulting in
possible memory leaks.
Reported by Coverity.
This patch lifts the limit of calling thread detection code only on KVM
guests. With upstream qemu the thread mappings are reported also on
non-KVM machines.
QEMU adopted the thread_id information from the kvm branch.
To remain compatible with older upstream versions of qemu the check is
attempted but the failure to detect threads (or even run the monitor
command - on older versions without SMP support) is treated non-fatal
and the code reports one vCPU with pid of the hypervisor (in same
fashion this was done on non-KVM guests).
After a cpu hotplug the qemu driver did not refresh information about
virtual processors used by qemu and their corresponding threads. This
patch forces a re-detection as is done on start of QEMU.
This ensures that correct information is reported by the
virDomainGetVcpus API and "virsh vcpuinfo".
A failure to obtain the thread<->vcpu mapping is treated non-fatal and
the mapping is not updated in a case of failure as not all versions of
QEMU report this in the info cpus command.
This patch changes a switch statement into ifs when handling live vs.
configuration modifications getting rid of redundant code in case when
both live and persistent configuration gets changed.
Ever since commit c964b6a, make was trying to find the timestamp
of '""./apibuild.py".stamp"', but only touching 'apibuild.py.stamp',
and thus always rebuilding. Reported by Daniel P. Berrange.
* docs/Makefile.am (APIBUILD, APIBUILD_STAMP): Omit bogus quotes.
when failing to attach another usb device to a domain for some reason
which has one use device attached before, the libvirtd crashed.
The crash is caused by null-pointer dereference error in invoking
usbDeviceListSteal passed in NULL value usb variable.
commit 05abd1507d introduces the bug.
Based on a report by Seth Vidal. Just because _you_ can use virsh
to connect to both source and destinations does not mean that libvirtd
on the source (aka _root_) can likewise connect to the destination;
this matters when setting up a peer-to-peer migration instead of a
native one.
* docs/migration.html.in: Mention that in peer-to-peer, the owner
of the source libvirtd (usually root) must be able to connect to
the destination.
Detected by valgrind. Leaks are introduced in commit 122fa379.
src/conf/storage_conf.c: fix memory leaks.
How to reproduce?
$ make && make -C tests check TESTS=storagepoolxml2xmltest
$ cd tests && valgrind -v --leak-check=full ./storagepoolxml2xmltest
actual result:
==28571== LEAK SUMMARY:
==28571== definitely lost: 40 bytes in 5 blocks
==28571== indirectly lost: 0 bytes in 0 blocks
==28571== possibly lost: 0 bytes in 0 blocks
==28571== still reachable: 1,054 bytes in 21 blocks
==28571== suppressed: 0 bytes in 0 blocks
Signed-off-by: Alex Jia <ajia@redhat.com>
No useful error was being reported when an invalid character device
target type is specified in the domainXML. E.g.
...
<console type="pty">
<source path="/dev/pts/2"/>
<target type="kvm" port="0"/>
</console>
...
resulted in
error: Failed to define domain from x.xml
error: An error occurred, but the cause is unknown
With this small patch, the error is more helpful
error: Failed to define domain from x.xml
error: XML error: unknown target type 'kvm' specified for character device
Since now we pre-set memory policy using libnuma to fully
drive numad, it needs to check numactl-devel if "with_numad"
is "yes".
configure with groups "--with-numad=yes --with-numactl=yes",
"--with-numad=no --with-numactl=yes", "--with-numad=yes
--with-numactl=yes" works fine after the change.
<vcpu> is not an optional node. The value for its 'placement'
actually always defaults to 'static' in the underlying codes.
(Even no 'cpuset' and 'placement' is specified, the domain
process will be pinned to all the available pCPUs).
Though numad will manage the memory allocation of task dynamically,
it wants management application (libvirt) to pre-set the memory
policy according to the advisory nodeset returned from querying numad,
(just like pre-bind CPU nodeset for domain process), and thus the
performance could benefit much more from it.
This patch introduces new XML tag 'placement', value 'auto' indicates
whether to set the memory policy with the advisory nodeset from numad,
and its value defaults to the value of <vcpu> placement, or 'static'
if 'nodeset' is specified. Example of the new XML tag's usage:
<numatune>
<memory placement='auto' mode='interleave'/>
</numatune>
Just like what current "numatune" does, the 'auto' numa memory policy
setting uses libnuma's API too.
If <vcpu> "placement" is "auto", and <numatune> is not specified
explicitly, a default <numatume> will be added with "placement"
set as "auto", and "mode" set as "strict".
The following XML can now fully drive numad:
1) <vcpu> placement is 'auto', no <numatune> is specified.
<vcpu placement='auto'>10</vcpu>
2) <vcpu> placement is 'auto', no 'placement' is specified for
<numatune>.
<vcpu placement='auto'>10</vcpu>
<numatune>
<memory mode='interleave'/>
</numatune>
And it's also able to control the CPU placement and memory policy
independently. e.g.
1) <vcpu> placement is 'auto', and <numatune> placement is 'static'
<vcpu placement='auto'>10</vcpu>
<numatune>
<memory mode='strict' nodeset='0-10,^7'/>
</numatune>
2) <vcpu> placement is 'static', and <numatune> placement is 'auto'
<vcpu placement='static' cpuset='0-24,^12'>10</vcpu>
<numatune>
<memory mode='interleave' placement='auto'/>
</numatume>
A follow up patch will change the XML formatting codes to always output
'placement' for <vcpu>, even it's 'static'.
It turns out that when cgroups are enabled, the use of a block device
for a snapshot target was failing with EPERM due to libvirt failing
to add the block device to the cgroup whitelist. See also
https://bugzilla.redhat.com/show_bug.cgi?id=810200
* src/qemu/qemu_driver.c
(qemuDomainSnapshotCreateSingleDiskActive)
(qemuDomainSnapshotUndoSingleDiskActive): Account for cgroup.
(qemuDomainSnapshotCreateDiskActive): Update caller.
Alon tried './qemuxml2argvtest --help' to figure out a test failure,
but it didn't help. The information is in HACKING, but it doesn't
hurt to make the tests also provide their own help.
Signed-off-by: Alon Levy <alevy@redhat.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
qemu's behavior in this case is to change the spice server behavior to
require secure connection to any channel not otherwise specified as
being in plaintext mode. libvirt doesn't currently allow requesting this
(via plaintext-channel=<channel name>).
RHBZ: 819499
Signed-off-by: Alon Levy <alevy@redhat.com>
We only know -lpthread exists on platforms where we build
threads-pthread.c; but when we build threads-win32.c, LIB_PTHREAD
is empty.
* tests/Makefile.am (shunloadtest_LDADD): Use correct library.
* tools/virsh.c (vshParseSnapshotDiskspec): Fix off-by-3 memmove
that would corrupt heap when parsing escaped --diskspec comma.
Bug introduced via commit v0.9.4-260-g35d52b5.
Until now, the nl_pid of the source address of every message sent by
virNetlinkCommand has been set to the value of getpid(). Most of the
time this doesn't matter, and in the one case where it does
(communication with lldpad), it previously was the proper thing to do,
because the netlink event service (which listens on a netlink socket
for unsolicited messages from lldpad) coincidentally always happened
to bind with a local nl_pid == getpid().
With the fix for:
https://bugzilla.redhat.com/show_bug.cgi?id=816465
that particular nl_pid is now effectively a reserved value, so the
netlink event service will always bind to something else
(coincidentally "getpid() + (1 << 22)", but it really could be
anything). The result is that communication between lldpad and
libvirtd is broken (lldpad gets a "disconnected" error when it tries
to send a directed message).
The solution to this problem caused by a solution, is to query the
netlink event service's nlhandle for its "local_port", and send that
as the source nl_pid (but only when sending to lldpad, of course - in
other cases we maintain the old behavior of sending getpid()).
There are two cases where a message is being directed at lldpad - one
in virNetDevLinkDump, and one in virNetDevVPortProfileOpSetLink.
The case of virNetDevVPortProfileOpSetLink is simplest to explain -
only if !nltarget_kernel, i.e. the message isn't targetted for the
kernel, is the dst_pid set (by calling
virNetDevVPortProfileGetLldpadPid()), so only in that case do we call
virNetlinkEventServiceLocalPid() to set src_pid.
For virNetDevLinkDump, it's a bit more complicated. The call to
virNetDevVPortProfileGetLldpadPid() was effectively up one level (in
virNetDevVPortProfileOpCommon), although obscured by an unnecessary
passing of a function pointer. This patch removes the function
pointer, and calls virNetDevVPortProfileGetLldpadPid() directly in
virNetDevVPortProfileOpCommon - if it's doing this, it knows that it
should also call virNetlinkEventServiceLocalPid() to set src_pid too;
then it just passes src_pid and dst_pid down to
virNetDevLinkDump. Since (src_pid == 0 && dst_pid == 0) implies that
the kernel is the destination, there is no longer any need to send
nltarget_kernel as an arg to virNetDevLinkDump, so it's been removed.
The disparity between src_pid being int and dst_pid being uint32_t may
be a bit disconcerting to some, but I didn't want to complicate
virNetlinkEventServiceLocalPid() by having status returned separately
from the value.
This value will be needed to set the src_pid when sending netlink
messages to lldpad. It is part of the solution to:
https://bugzilla.redhat.com/show_bug.cgi?id=816465
Note that libnl's port generation algorithm guarantees that the
nl_socket_get_local_port() will always be > 0 (since it is "getpid() +
(n << 22>" where n is always < 1024), so it is okay to cast the
uint32_t to int (thus allowing us to use -1 as an error sentinel).
Until now, virNetlinkCommand has assumed that the nl_pid in the source
address of outgoing netlink messages should always be the return value
of getpid(). In most cases it actually doesn't matter, but in the case
of communication with lldpad, lldpad saves this info and later uses it
to send netlink messages back to libvirt. A recent patch to fix Bug
816465 changed the order of the universe such that the netlink event
service socket is no longer bound with nl_pid == getpid(), so lldpad
could no longer send unsolicited messages to libvirtd. Adding src_pid
as an argument to virNetlinkCommand() is the first step in notifying
lldpad of the proper address of the netlink event service socket.
src/qemu/qemu_hostdev.c:
refactor qemuPrepareHostdevUSBDevices function, make it focus on
adding usb device to activeUsbHostdevs after check. After that,
the usb hotplug function qemuDomainAttachHostDevice also could use
it.
expand qemuPrepareHostUSBDevices to perform the usb search,
rollback on failure.
src/qemu/qemu_hotplug.c:
If there are multiple usb devices available with same vendorID and productID,
but with different value of "bus, device", we give an error to let user
use <address> to specify the desired one.
usbFindDevice():get usb device according to
idVendor, idProduct, bus, device
it is the exact match of the four parameters
usbFindDeviceByBus():get usb device according to bus, device
it returns only one usb device same as usbFindDevice
usbFindDeviceByVendor():get usb device according to idVendor,idProduct
it probably returns multiple usb devices.
usbDeviceSearch(): a helper function to do the actual search
When we added the default USB controller into domain XML, we efficiently
broke migration to older versions of libvirt that didn't support USB
controllers at all (0.9.4 and earlier) even for domains that don't use
anything that the older libvirt can't provide. We still want to present
the default USB controller in any XML seen by a user/app but we can
safely remove it from the domain XML used during migration. If we are
migrating to a new enough libvirt, it will add the controller XML back,
while older libvirt won't be confused with it although it will still
tell qemu to create the controller.
Similar approach can be used in the future whenever we find out we
always enabled some kind of device without properly advertising it in
domain XML.
Commit 4c82f09e added a capability check for qemu per-device io
throttling, but only applied it to domain startup. As mentioned
in the previous commit (98cec05), the user can still get an 'internal
error' message during a hotplug attempt, when the monitor command
doesn't exist. It is confusing to allow tuning on inactive domains
only to then be rejected when starting the domain.
* src/qemu/qemu_driver.c (qemuDomainSetBlockIoTune): Reject
offline tuning if online can't match it.
If you have a qemu build that lacks the blockio tune monitor command,
then this command:
$ virsh blkdeviotune rhel6u2 hda --total_bytes_sec 1000
error: Unable to change block I/O throttle
error: internal error Unexpected error
fails as expected (well, the error message is lousy), but the next
dumpxml shows that the domain was modified anyway. Worse, that means
if you save the domain then restore it, the restore will likely fail
due to throttling being unsupported, even though no throttling should
even be active because the monitor command failed in the first place.
* src/qemu/qemu_driver.c (qemuDomainSetBlockIoTune): Check for
error before making modification permanent.
These two functions are called from main() on all platforms, and
always return success on platforms that don't support libnl. They
still log an error message, though, which doesn't make sense - they
should just be NOPs on those platforms. (Per a suggestion during
review, I've turned the logs into debug messages rather than removing
them completely).
