At this point it has a limited functionality and is highly
experimental. Supported domain operations are:
* define
* start
* destroy
* dumpxml
* dominfo
It's only possible to have only one disk device and only one
network, which should be of type bridge.
PS2 devices only work on X86 platform, other platforms may need
USB devices instead. Athough it doesn't influence the QEMU command line,
it's not right to add PS2 mouse/keyboard for non-X86 platform.
Signed-off-by: Li Zhang <zhlcindy@linux.vnet.ibm.com>
Signed-off-by: Ján Tomko <jtomko@redhat.com>
aebbcdd didn't change the non-linux definition of the function,
breaking the build on FreeBSD:
../../src/util/virinitctl.c:164: error: conflicting types for
'virInitctlSetRunLevel'
../../src/util/virinitctl.h:40: error: previous declaration of
'virInitctlSetRunLevel' was here
There might be some use cases, where user wants to prepare the host or
its environment prior to starting a network and do some cleanup after
the network has been shut down. Consider all the functionality that
libvirt doesn't currently have as an example what a hook script can
possibly do.
Signed-off-by: Michal Privoznik <mprivozn@redhat.com>
Use helper virProcessRunInMountNamespace in lxcDomainShutdownFlags and
lxcDomainReboot. Otherwise, a malicious guest could use symlinks
to force the host to manipulate the wrong file in the host's namespace.
Idea by Dan Berrange, based on an initial report by Reco
<recoverym4n@gmail.com> at
http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=732394
Signed-off-by: Eric Blake <eblake@redhat.com>
Implement virProcessRunInMountNamespace, which runs callback of type
virProcessNamespaceCallback in a container namespace. This uses a
child process to run the callback, since you can't change the mount
namespace of a thread. This implies that callbacks have to be careful
about what code they run due to async safety rules.
Idea by Dan Berrange, based on an initial report by Reco
<recoverym4n@gmail.com> at
http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=732394
Signed-off-by: Daniel Berrange <berrange@redhat.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
Add a helper function which takes a file path and ensures
that all directory components leading up to the file exist.
IOW, it strips the filename part of the path and passes
the result to virFileMakePath.
Signed-off-by: Daniel P. Berrange <berrange@redhat.com>
Also try to bind on IPv6 to check if the port is occupied.
Change the mocked bind in the test to return EADDRINUSE
for some ports only for the IPv4/IPv6 socket if we're testing
on a host with IPv6 compiled in.
Also mock socket() to make it fail with EAFNOTSUPPORTED
if LIBVIRT_TEST_IPV4ONLY is set in the environment, to
simulate a host without IPv6 support in the kernel. The
tests are repeated again with this variable set.
https://bugzilla.redhat.com/show_bug.cgi?id=1025407
virConf now honours a VIR_CONF_FLAG_LXC_FORMAT flag to handle LXC
configuration files. The differences are that property names can
contain '.' character and values are all strings without any bounding
quotes.
Provide a new virConfWalk function calling a handler on all non-comment
values. This function will be used by the LXC conversion code to loop
over LXC configuration lines.
In order to make a client-only build successful on RHEL4 (yes, you
read that correctly!), commit 3ed2e54 modified src/util/virnetdev.c so
that the functional version of virNetDevGetVLanID() was only compiled
if GET_VLAN_VID_CMD was defined. However, it is *never* defined, but
is only an enum value, so the proper version was no longer compiled
even on platforms that support it. This resulted in the vlan tag not
being properly set for guest traffic on VEPA mode guest macvtap
interfaces that were bound to a vlan interface (that's the only place
that libvirt currently uses virNetDevGetVLanID)
Since there is no way to compile conditionally based on the presence
of an enum value, this patch modifies configure.ac to check for said
enum value with AC_CHECK_DECLS(), which #defines
HAVE_DECL_GET_VLAN_VID_CMD to 1 if it's successful compiling a test
program that uses GET_VLAN_VID_CMD (and still #defines it, but to 0,
if it's not successful). We can then make the compilation of
virNetDevGetVLanID() conditional on the value of
HAVE_DECL_GET_VLAN_VID_CMD.
Coverity complains about "USE_AFTER_FREE" due to how virPCIDeviceSetStubDriver
"could" return either -1, 0, or 1 from the VIR_STRDUP() and then possibly makes
a call to virPCIDeviceDetach().
The only way this could happen is if NULL were passed as the "driver" name
and virStrdup() returned 0. Since the calling functions check < 0 on the
initial function call, the 0 possibility causes Coverity to complain.
To fix this - enforce that the second parameter is not NULL using
ATTRIBUTE_NONNULL(2) for the function prototype, then in virPCIDeviceDetach
add an sa_assert(dev->stubDriver). This will result in Coverity not complaining
any more.
https://bugzilla.redhat.com/show_bug.cgi?id=1045124
When loading modules, libvirt does not honor the modprobe blacklist.
Use the new virKModLoad() API in order to attempt load with blacklist check.
Use the new virKModIsBlacklisted() API to check if the failure to load
was due to the blacklist
Signed-off-by: John Ferlan <jferlan@redhat.com>
virKModConfig() - Return a buffer containing kernel module configuration
virKModLoad() - Load a specific module into the kernel configuration
virKModUnload() - Unload a specific module from the kernel configuration
virKModIsBlacklisted() - Determine whether a module is blacklisted within
the kernel configuration
There are a number of pthreads impls available on Win32
these days, in particular the mingw64 project has a good
impl. Delete the native windows thread implementation and
rely on using pthreads everywhere.
Signed-off-by: Daniel P. Berrange <berrange@redhat.com>
Commit 10c9ceff6d intended to introduce new argument for the
testing purpose, but it missed the similar changing of the
device's sg_path. The problem was hidden since my laptop has
the /dev/sg0 and /dev/sg1. A later patch will modify the tests
accordingly.
Signed-off-by: Osier Yang <jyang@redhat.com>
Reported-by: Pavel Hrdina <phrdina@redhat.com>
To support passing the path of the test data to the utils, one
more argument is added to virSCSIDeviceGetSgName,
virSCSIDeviceGetDevName, and virSCSIDeviceNew, and the related
code is changed accordingly.
Later tests for the scsi utils will be based on this patch.
Signed-off-by: Osier Yang <jyang@redhat.com>
It doesn't make sense to fail if the SCSI host device is specified
as "shareable" explicitly between domains (NB, it works if and only
if the device is specified as "shareable" for *all* domains,
otherwise it fails).
To fix the problem, this patch introduces an array for virSCSIDevice
struct, which records all the names of domain which are using the
device (note that the recorded domains must specify the device as
shareable). And the change on the data struct brings on many
subsequent changes in the code.
Prior to this patch, the "shareable" tag didn't work as expected,
it actually work like "non-shareable". So this patch also added notes
in formatdomain.html to declare the fact.
* src/util/virscsi.h:
- Remove virSCSIDeviceGetUsedBy
- Change definition of virSCSIDeviceGetUsedBy and virSCSIDeviceListDel
- Add virSCSIDeviceIsAvailable
* src/util/virscsi.c:
- struct virSCSIDevice: Change "used_by" to be an array; Add
"n_used_by" as the array count
- virSCSIDeviceGetUsedBy: Removed
- virSCSIDeviceFree: frees the "used_by" array
- virSCSIDeviceSetUsedBy: Copy the domain name to avoid potential
memory corruption
- virSCSIDeviceIsAvailable: New
- virSCSIDeviceListDel: Change the logic, for device which is already
in the list, just remove the corresponding entry in "used_by". And
since it's only used in one place, we can safely removing the code
to find out the dev in the list first.
- Copyright updating
* src/libvirt_private.sys:
- virSCSIDeviceGetUsedBy: Remove
- virSCSIDeviceIsAvailable: New
* src/qemu/qemu_hostdev.c:
- qemuUpdateActiveScsiHostdevs: Check if the device existing before
adding it to the list;
- qemuPrepareHostdevSCSIDevices: Error out if the not all domains
use the device as "shareable"; Also don't try to add the device
to the activeScsiHostdevs list if it already there; And make
more sensible error w.r.t the current "shareable" value in
driver->activeScsiHostdevs.
- qemuDomainReAttachHostScsiDevices: Change the logic according
to the changes on helpers.
Signed-off-by: Osier Yang <jyang@redhat.com>
This reverts commit 2996e6be19
and some parts of 2636dc8c4d.
The former one tried to implement QoS setting on bridgeless networks.
However, as discussed upstream [1], the patch is far away from being
useful in even a single case. The whole idea of network QoS is to have
aggregated limits over several interfaces. This patch is doing
completely the opposite when merging two QoS settings (from the network
and the domain interface) into one which is then set at the domain
interface itself, not the network.
The latter one is the test for the previous one. Now none of them makes
sense.
1: https://www.redhat.com/archives/libvir-list/2014-January/msg01441.html
Conflicts:
tests/virnetdevbandwidthtest.c: New test has been introduced since
then.
There are some units within libvirt that utilize virCommand API to run
some commands and deserve own unit testing. These units are, however,
not desired to be rewritten to dig virCommand API usage out. As a great
example virNetDevBandwidth could be used. The problem with the bandwidth
unit is: it uses virCommand API heavily. Therefore we need a mechanism
to not really run a command, but rather see its string representation
after which we can decide if the unit construct the correct sequence of
commands or not.
Signed-off-by: Michal Privoznik <mprivozn@redhat.com>
Implement virProcess{Get,Set}Affinity() using cpuset_getaffinity()
and cpuset_setaffinity() calls. Quick search showed that they are
only available on FreeBSD, so placed it inside existing #ifdef
blocks for FreeBSD instead of adding configure checks.
https://bugzilla.redhat.com/show_bug.cgi?id=1055484
Currently, libvirt's XML schema of network allows QoS to be defined for
every network even though it has no bridge. For instance:
<network>
<name>vdsm-no-bridge</name>
<forward mode='passthrough'>
<interface dev='em1.10'/>
</forward>
<bandwidth>
<inbound average='1000' peak='5000' burst='1024'/>
<outbound average='1000' burst='1024'/>
</bandwidth>
</network>
The bandwidth limitations can be, however, applied even on such
networks. In fact, they are going to be applied on the interface that
will be connected to the network on a domain startup. This approach,
however, has one limitation. With bridged networks, there are two points
where QoS can be set: bridge and domain interface. The lower limit of
the two is enforced then. For instance, if the interface has 10Mbps
average, but the network only 1Mbps, there's no way for interface to
transmit packets faster than the 1Mbps limit. With two points this is
enforced by kernel. With only one point, we must combine both QoS
settings into one which is set afterwards. Look at
virNetDevBandwidthMinimal() and you'll understand immediately what I
mean.
Signed-off-by: Michal Privoznik <mprivozn@redhat.com>
Unlike the host devices of other types, SCSI host device XML supports
"shareable" tag. This patch introduces it for the virSCSIDevice struct
for a later patch use (to detect if the SCSI device is shareable when
preparing the SCSI host device in QEMU driver).
There are 2 issues here: First we shouldn't add "1" to the return
value of numa_max_node(), since the semanteme of the error message
was changed, it's not saying about the number of total NUMA nodes
anymore. Second, the value of "bit" is the position of the first
bit which exceeds either numa_max_node() or NUMA_NUM_NODES, it can
be any number in the range, so saying "bigger than $bit" is quite
confused now. For example, assuming there is a NUMA machine which
has 10 NUMA nodes, and one specifies the "nodeset" as "0,5,88",
the error message will be like:
Nodeset is out of range, host cannot support NUMA node bigger than 88
It sounds like all NUMA node number less than 88 is fine, but
actually the maximum NUMA node number the machine supports is 9.
This patch fixes the issues by removing the addition with "1" and
simplifies the error message as "NUMA node $bit is out of range".
Also simplifies the comparision in the while loop by getting the
smaller one of numa_max_node() and NUMA_NUM_NODES up front.
This is useful in certain circumstances, for example when
libvirtd is being executed by FreeBSD rc script, it cannot find
dmidecode installed from FreeBSD ports because it doesn't have
/usr/local (default prefix for ports) in PATH.
https://bugzilla.redhat.com/show_bug.cgi?id=1046919
Since commit v0.9.0-47-g4e8969e (released in 0.9.1) some failures during
device detach were reported to callers of virPCIDeviceBindToStub as
success. For example, even though a device seemed to be detached
virsh # nodedev-detach pci_0000_07_05_0 --driver vfio
Device pci_0000_07_05_0 detached
one could find similar message in libvirt logs:
Failed to bind PCI device '0000:07:05.0' to vfio-pci: No such device
This patch fixes these paths and also avoids overwriting real errors
with errors encountered during a cleanup phase.
https://bugzilla.redhat.com/show_bug.cgi?id=1046919
When a PCI device is not bound to any driver, reattach should just
trigger driver probe rather than failing with
Invalid device 0000:00:19.0 driver file
/sys/bus/pci/devices/0000:00:19.0/driver is not a symlink
While virPCIDeviceGetDriverPathAndName was documented to return success
and NULL driver and path when a device is not attached to any driver but
didn't do so. Thus callers could not distinguish unbound devices from
failures.
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
This patch introduces virCgroupSetBlkioDeviceReadIops,
virCgroupSetBlkioDeviceWriteIops,
virCgroupSetBlkioDeviceReadBps and
virCgroupSetBlkioDeviceWriteBps,
we can use these interfaces to set up throttle
blkio cgroup for domain.
This patch also adds the new throttle blkio cgroup
elements to the test xml.
Signed-off-by: Guan Qiang <hzguanqiang@corp.netease.com>
Signed-off-by: Gao feng <gaofeng@cn.fujitsu.com>
A "xmlstr" string may not be assigned into a "doc" pointer and it
could cause memory leak. To fix it if the "doc" pointer is NULL and
the "xmlstr" string is not assigned we should free it.
This has been found by coverity.
Signed-off-by: Pavel Hrdina <phrdina@redhat.com>
On my Fedora 20 box with mingw cross-compiler, the build failed with:
../../src/rpc/virnetclient.c: In function 'virNetClientSetTLSSession':
../../src/rpc/virnetclient.c:745:14: error: unused variable 'oldmask' [-Werror=unused-variable]
sigset_t oldmask, blockedsigs;
^
I traced it to the fact that mingw64-winpthreads installs a header
that does #define pthread_sigmask(...) 0, which means any argument
only ever passed to pthread_sigmask is reported as unused. This
patch works around the compilation failure, with behavior no worse
than what mingw already gives us regarding the function being a
no-op.
* configure.ac (pthread_sigmask): Probe for broken mingw macro.
* src/util/virutil.h (pthread_sigmask): Rewrite to something that
avoids unused variables.
Signed-off-by: Eric Blake <eblake@redhat.com>
Like commit 94a26c7e from Eric Blake, the old fuzzy code should
be replaced by the new array management macros now.
And the type of scsi->count should be changed into "size_t", and
thus virSCSIDeviceListCount should return size_t instead, similar
for vir{PCI,USB}DeviceListCount.
When the host is configured with very restrictive firewall (default policy
is DROP for all chains, including OUTPUT), the bridge driver for Linux
adds netfilter entries to allow DHCP and DNS requests to go from the VM
to the dnsmasq of the host.
The issue that this commit fixes is the fact that a DROP policy on the OUTPUT
chain blocks the DHCP replies from the host’s dnsmasq to the VM.
As DHCP replies are sent in UDP, they are not caught by any --ctstate ESTABLISHED
rule and so, need to be explicitly allowed.
Signed-off-by: Lénaïc Huard <lenaic@lhuard.fr.eu.org>
When determining if a device is behind a PCI bridge, the PCI device
class is checked by reading the config space. However, there are some
devices which have the wrong class on the config space, but the class is
initialized by Linux correctly as a PCI BRIDGE. This class can be read
by the sysfs file '/sys/bus/pci/devices/xxxx:xx:xx.x/class'.
One example of such bridge is IBM PCI Bridge 1014:03b9, which is
identified as a Host Bridge when reading the config space.
Signed-off-by: Thadeu Lima de Souza Cascardo <cascardo@linux.vnet.ibm.com>
Some of our operation denied messages are outright stupid; for
example, if virIdentitySetAttr fails:
error: operation Identity attribute is already set forbidden for read only access
This patch fixes things to a saner:
error: operation forbidden: Identity attribute is already set
It also consolidates the most common usage pattern for operation
denied errors: read-only connections preventing a public API. In
this case, 'virsh -r -c test:///default destroy test' changes from:
error: operation virDomainDestroy forbidden for read only access
to:
error: operation forbidden: read only access prevents virDomainDestroy
Note that we were previously inconsistent on which APIs used
VIR_FROM_DOM (such as virDomainDestroy) vs. VIR_FROM_NONE (such as
virDomainPMSuspendForDuration). After this patch, all uses
consistently use VIR_FROM_NONE, on the grounds that it is unlikely
that a caller learning that a call is denied can do anything in
particular with extra knowledge which error domain the call belongs
to (similar to what we did in commit baa7244).
* src/util/virerror.c (virErrorMsg): Rework OPERATION_DENIED error
message.
* src/internal.h (virCheckReadOnlyGoto): New macro.
* src/util/virerror.h (virReportRestrictedError): New macro.
* src/libvirt-lxc.c: Use new macros.
* src/libvirt-qemu.c: Likewise.
* src/libvirt.c: Likewise.
* src/locking/lock_daemon.c (virLockDaemonClientNew): Likewise.
Signed-off-by: Eric Blake <eblake@redhat.com>
We weren't very consistent in our use of VIR_ERR_NO_SUPPORT; many
users just passed __FUNCTION__ on, while others passed "%s" to
silence over-eager compilers that warn about __FUNCTION__ not
containing any %. It's nicer to route all these uses through
a single macro, so that if we ever need to change the reporting,
we can do it in one place.
I verified that 'virsh -c test:///default qemu-monitor-command test foo'
gives the same error message before and after this patch:
error: this function is not supported by the connection driver: virDomainQemuMonitorCommand
Note that in libvirt.c, we were inconsistent on whether virDomain*
API used virLibConnError() (with VIR_FROM_NONE) or virLibDomainError()
(with VIR_FROM_DOMAIN); this patch unifies these errors to all use
VIR_FROM_NONE, on the grounds that it is unlikely that a caller
learning that a call is unimplemented can do anything in particular
with extra knowledge of which error domain it belongs to.
One particular change to note is virDomainOpenGraphics which was
trying to fail with VIR_ERR_NO_SUPPORT after a failed
VIR_DRV_SUPPORTS_FEATURE check; all other places that fail a
feature check report VIR_ERR_ARGUMENT_UNSUPPORTED.
* src/util/virerror.h (virReportUnsupportedError): New macro.
* src/libvirt-qemu.c: Use new macro.
* src/libvirt-lxc.c: Likewise.
* src/lxc/lxc_driver.c: Likewise.
* src/security/security_manager.c: Likewise.
* src/util/virinitctl.c: Likewise.
* src/libvirt.c: Likewise.
(virDomainOpenGraphics): Use correct error for unsupported feature.
Signed-off-by: Eric Blake <eblake@redhat.com>
Having one API call into another is generally not good; among
other issues, it gives confusing logs, and is not quite as
efficient.
This fixes several instances, but not all: we still have instances
in both libvirt.c and in backend hypervisors (lxc and qemu) calling
the public virTypedParamsGetString and friends, which dispatch
errors immediately. I'm not sure if it is worth trying to clean
that up in a separate patch (such a cleanup may be easiest by
separating the public function into a wrapper around the internal,
then tweaking internal.h so that internal users directly use the
internal function).
* src/libvirt.c (virDomainGetUUIDString, virNetworkGetUUIDString)
(virStoragePoolGetUUIDString, virSecretGetUUIDString)
(virNWFilterGetUUIDString): Avoid nested public API call.
* src/util/virtypedparam.c (virTypedParamsReplaceString): Don't
dispatch errors here.
(virTypedParamsGet): No need to reset errors.
(virTypedParamsGetBoolean): Use consistent ordering.
Signed-off-by: Eric Blake <eblake@redhat.com>
I noticed that the virDomainQemuMonitorCommand debug output wasn't
telling me the name of the domain it was working on. While it was
easy enough to determine which pointer matches the domain based on
other log messages, it is nicer to be consistent.
* src/util/viruuid.h (VIR_UUID_DEBUG): Moved here from...
* src/libvirt.c (VIR_UUID_DEBUG): ...here.
(VIR_ARG15, VIR_HAS_COMMA, VIR_DOMAIN_DEBUG_EXPAND)
(VIR_DOMAIN_DEBUG_PASTE, VIR_DOMAIN_DEBUG_0, VIR_DOMAIN_DEBUG_1)
(VIR_DOMAIN_DEBUG_2, VIR_DOMAIN_DEBUG): Move...
* src/datatypes.h: ...here.
* src/libvirt-qemu.c (virDomainQemuMonitorCommand)
(virDomainQemuAgentCommand): Better debug messages.
* src/libvirt-lxc.c (virDomainLxcOpenNamespace): Likewise.
Signed-off-by: Eric Blake <eblake@redhat.com>
Since libvirt 0.9.3, the entire virevent.c file has been a public
API, so improve the documentation in this file. Also, fix a
potential core dump - it could only be triggered by bogus use of
the API and would only affect the caller (not libvirtd), but we
might as well be nice.
* src/libvirt.c (virConnectSetKeepAlive)
(virConnectDomainEventRegister, virConnectDomainEventRegisterAny)
(virConnectNetworkEventRegisterAny): Document event loop requirement.
* src/util/virevent.c (virEventAddHandle, virEventRemoveHandle)
(virEventAddTimeout, virEventRemoveTimeout): Likewise.
(virEventUpdateHandle, virEventUpdateTimeout): Likewise, and avoid
core dump if caller didn't register handler.
(virEventRunDefaultImpl): Expand example, and set up code block in
html docs.
(virEventRegisterImpl, virEventRegisterDefaultImpl): Document more
on the use of the event loop.
Signed-off-by: Eric Blake <eblake@redhat.com>
https://bugzilla.redhat.com/show_bug.cgi?id=1044806
Currently, sending the ANSI_A keycode from os_x codepage doesn't work as
it has a special value of 0x0. Our internal code handles that no
different to other not defined keycodes. Hence, in order to allow it we
must change all the undefined keycodes from 0 to -1 and adapt some code
too.
# virsh send-key guestname --codeset os_x ANSI_A
error: invalid keycode: 'ANSI_A'
# virsh send-key guestname --codeset os_x ANSI_B
# virsh send-key guestname --codeset os_x ANSI_C
Signed-off-by: Michal Privoznik <mprivozn@redhat.com>
Currently the virDBusAddWatch does
virEventAddHandle(fd, flags,
virDBusWatchCallback,
watch, NULL);
dbus_watch_set_data(watch, info, virDBusWatchFree);
Unfortunately this is racy - since the event loop is in a
different thread, the virDBusWatchCallback method may be
run before we get to calling dbus_watch_set_data. We must
reverse the order of these calls
See https://bugzilla.redhat.com/show_bug.cgi?id=885445
Signed-off-by: Daniel P. Berrange <berrange@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>
Since kernel 3.12 (commit 34ff8dc08956098563989d8599840b130be81252 in
linux-stable.git in particular) the value for 'unlimited' in cgroup
memory limits changed from LLONG_MAX to ULLONG_MAX. Due to rather
unfortunate choice of our VIR_DOMAIN_MEMORY_PARAM_UNLIMITED constant
(which we transfer as an unsigned long long in Kibibytes), we ended up
with the situation described below (applies to x86_64):
- 2^64-1 (ULLONG_MAX) -- "unlimited" in kernel = 3.12
- 2^63-1 (LLONG_MAX) -- "unlimited" in kernel < 3.12
- 2^63-1024 -- our PARAM_UNLIMITED scaled to Bytes
- 2^53-1 -- our PARAM_UNLIMITED unscaled (in Kibibytes)
This means that when any number within (2^63-1, 2^64-1] is read from
memory cgroup, we are transferring that number instead of "unlimited".
Unfortunately, changing VIR_DOMAIN_MEMORY_PARAM_UNLIMITED would break
ABI compatibility and thus we have to resort to a different solution.
With this patch every value greater than PARAM_UNLIMITED means
"unlimited". Even though this may seem misleading, we are already in
such unclear situation when running 3.12 kernel with memory limits set
to 2^63.
One example showing most of the problems at once (with kernel 3.12.2):
# virsh memtune asdf --hard-limit 9007199254740991 --swap-hard-limit -1
# echo 12345678901234567890 >\
/sys/fs/cgroup/memory/machine/asdf.libvirt-qemu/memory.soft_limit_in_bytes
# virsh memtune asdf
hard_limit : 18014398509481983
soft_limit : 12056327051986884
swap_hard_limit: 18014398509481983
Signed-off-by: Martin Kletzander <mkletzan@redhat.com>
