One of the xml tests in the test suite was created using a
now-deprecated qemu machine type ("fedora-13", which was only ever
valid for Fedora builds of qemu). Although strictly speaking it's not
necessary to replace it with an actual supported qemu machine type
(since the xml in question is never actually sent to qemu), this patch
changes it to the actually-supported "pc-0.13" just for general
tidiness. (Also, on some Fedora builds which contain a special patch
to rid the world of "fedora-13", having it mentioned in the test suite
will cause make check to fail.)
This patch addresses https://bugzilla.redhat.com/show_bug.cgi?id=760442
When a network has any forward type other than route, nat or none, the
network configuration should be done completely external to libvirt -
libvirt only uses these types to allow configuring guests in a manner
that isn't tied to a specific host (all the host-specific information,
in particular interface names, port profile data, and bandwidth
configuration is in the network definition, and the guest
configuration only references it).
Due to a bug in the bridge network driver, libvirt was adding iptables
rules for networks with forward type='bridge' etc. any time libvirtd
was restarted while one of these networks was active.
This patch eliminates that error by only "reloading" iptables rules if
forward type is route, nat, or none.
In some error situations, the function testDomainRestoreFlags() could
unlock the test driver mutex without first locking it. This patch
moves the lock operation earlier, so that it occurs before any
potential jump down to the unlock call.
I found this problem while auditing the test driver lock usage to
determine the cause of a hang while running the following test:
cd tests; while true; do printf x; ./undefine; done
This patch *does not* solve that problem, but we now understand its
actual source, and danpb is working on a patch.
One of the top questions by libvirt users is how to create a host
bridge device so that guests can be directly on the physical
network. There are several example documents that explain how to do
this manually, but following them often results in confusion and
failure. virt-manager does a good job of creating a bridge based on an
existing network device, but not everyone wants to use virt-manager.
This patch adds a new command, iface-bridge that makes it just about
as simple as possible to create a new bridge device based on an
existing ethernet/vlan/bond device (including associating IP
configuration with the bridge rather than the now-attached device),
and start that new bridge up ready for action, eg:
virsh iface-bridge eth0 br0
For symmetry's sake, it also adds a command to remove a device from a
bridge, restoring the IP config to the now-unattached device:
virsh iface-unbridge br0
(I had a short debate about whether to do "iface-unbridge eth0"
instead, but that would involve searching through all bridge devices
for the one that contained eth0, which seems like a bit too much
trouble).
NOTE: These two commands require that the netcf library be available
on the host. Hopefully this will provide some extra incentive for
people using suse, debian, ubuntu, and other similar systems to polish
up (and push downstream) the ports to those distros recently pushed to
the upstream netcf repo by Dan Berrange. Anyone interested in helping
with that effort in any way should join the netcf-devel mailing list
(subscription info at
https://fedorahosted.org/mailman/listinfo/netcf-devel)
During creation of the bridge, it's possible to specify whether or not
the STP protocol should be started up on the bridge and, if so, how
many seconds the bridge should squelch traffic from newly added
devices while learning new topology (defaults are stp='on' and
delay='0', which seems to usually work best for bridges used in the
context of libvirt guests).
There is also an option to not immediately start the bridge (and a
similar option to not immediately start the un-attached device after
destroying the bridge. Default is to start the new device, because in
the case of iface-unbridge not starting is strongly discouraged as it
will leave the system with no network connectivity on that interface
(because it's necessary to destroy/undefine the bridge device before
the unattached device can be defined), and it seemed better to make
the option for iface-bridge behave consistently.
NOTE TO THOSE TRYING THESE COMMANDS FOR THE FIRST TIME: to guard
against any "unexpected" change to configuration, it is advisable to
issue an "virsh iface-begin" command before starting any interface
config changes, and "virsh iface-commit" only after you've verified
that everything is working as you expect. If something goes wrong,
you can always run "virsh iface-rollback" or reboot the system (which
should automatically do iface-rollback).
Aside from adding the code for these two functions, and the two
entries into the command table, the only other change to virsh.c was
to add the option name to vshCommandOptInterfaceBy(), because the
iface-unbridge command names its interface option as "bridge".
virsh.pod has also been updated with short descriptions of these two
new commands.
In a recent expansion of the documentation on network forward modes, I
incorrectly stated that incoming sessions to guests on routed networks
were blocked. This is true for guests on NATed networks, but not
routed. This patch corrects that error, and adds a pointer to the
nwfilter page for those who do want to restrict incoming sessions to
hosts on routed networks.
A few people have attempted to use the new forwarding modes with older
versions of libvirt. The docs where the modes are described have
always stated the minimum required libvirt version, but the examples
at the end didn't, which I believe is what has caused the confusion.
Similarly, the section on portgroups now has a version tag added at
the beginning.
I also noticed that there was no example of defining a <dns> hostname,
so I added one, as well as making the domain name example more
recognizable (by adding ".com" to the domain).
Previously libvirt's disk device XML only had a single attribute,
error_policy, to control both read and write error policy, but qemu
has separate options for controlling read and write. In one case
(enospc) a policy is allowed for write errors but not read errors.
This patch adds a separate attribute that sets only the read error
policy. If just error_policy is set, it will apply to both read and
write error policy (previous behavior), but if the new rerror_policy
attribute is set, it will override error_policy for read errors only.
Possible values for rerror_policy are "stop", "report", and "ignore"
("report" is the qemu-controlled default for rerror_policy when
error_policy isn't specified).
For consistency, the value "report" has been added to the possible
values for error_policy as well.
commit 12062ab set rerror=ignore when error_policy="enospace" was
selected (since the rerror option in qemu doesn't accept "enospc", as
the werror option does).
After that patch was already pushed, Paolo Bonzini noticed it and
commented that leaving rerror at the default ("report") would be a
better choice. This patch corrects the problem - if error_policy =
"enospace" is given, rerror is left off the qemu commandline,
effectively setting it to "report". For other values, rerror is still
set to match werror.
Additionally, the parsing of error_policy was changed to no longer
erroneously allow "default" as a choice - as with most other
attributes, if you want the default setting, just don't specify an
error_policy.
Finally, two ommissions in the first patch were corrected - a
long-dormant qemuxml2argv test for enospace was enabled, and fixed to
pass, and the argv2xml parser in qemu_command.c was updated to
recognize the different spelling on the qemu commandline.
This resolves:
https://bugzilla.redhat.com/show_bug.cgi?id=730909
When support for setting the qemu disk error policy to "enospc" was
added, it was inadvertently spelled "enospace". This patch corrects
that on the qemu commandline (while retaining the "enospace" spelling
for libvirt's XML).
Also, while examining the qemu source, I found that "enospc" is not
allowed for the read error policy, only for write error policy (makes
sense). Since libvirt currently only has a single error policy
setting, when "enospace" is selected, the read error policy is set to
"ignore".
This patch is a fix for:
https://bugzilla.redhat.com/show_bug.cgi?id=743176
which was discovered by Dan Berrange while making bandwidth
configuration work for LXC guests.
Background: Although virtportprofile data from a network portgroup is
only applicable for direct mode interfaces, the code that copies
bandwidth data from the portgroup was also only being executed in the
case of direct mode interfaces. The result was that interfaces using
traditional virtual networks (forward mode='nat|route|none'), and
those using a host bridge for forwarding, would not pick up bandwidth
data from a portgroup defined in the network.
This patch moves that code outside the conditional, so that bandwidth
information is *alway* copied from the appropriate portgroup (unless
the <interface> definition itself already has bandwidth information,
which would take precedence over what's in the portgroup anyway).
When support for was added for PCI multifunction cards (in commit
9f8baf, first included in libvirt 0.9.3), it was done by always
turning on the multifunction bit for all PCI devices. Since that time
it has been realized that this is not an ideal solution, and that the
multifunction bit must be selectively turned on. For example, see
https://bugzilla.redhat.com/show_bug.cgi?id=728174
and the discussion before and after
https://www.redhat.com/archives/libvir-list/2011-September/msg01036.html
This patch modifies multifunction support so that the multifunction=on
option is only added to the qemu commandline for a device if its PCI
<address> definition has the attribute "multifunction='on'", e.g.:
<address type='pci' domain='0x0000' bus='0x00'
slot='0x04' function='0x0' multifunction='on'/>
In practice, the multifunction bit should only be turned on if
function='0' AND other functions will be used in the same slot - it
usually isn't needed for functions 1-7 (although there are apparently
some exceptions, e.g. the Intel X53 according to the QEMU source
code), and should never be set if only function 0 will be used in the
slot. The test cases have been changed accordingly to illustrate.
With this patch in place, if a user attempts to assign multiple
functions in a slot without setting the multifunction bit for function
0, libvirt will issue an error when the domain is defined, and the
define operation will fail. In the future, we may decide to detect
this situation and automatically add multifunction=on to avoid the
error; even then it will still be useful to have a manual method of
turning on multifunction since, as stated above, there are some
devices that excpect it to be turned on for all functions in a slot.
A side effect of this patch is that attempts to use the same PCI
address for two different devices will now log an error (previously
this would cause the domain define operation to fail, but there would
be no log message generated). Because the function doing this log was
almost completely rewritten, I didn't think it worthwhile to make a
separate patch for that fix (the entire patch would immediately be
obsoleted).
This patch fixes the regression with using named pipes for qemu serial
devices noted in:
https://bugzilla.redhat.com/show_bug.cgi?id=740478
The problem was that, while new code in libvirt looks for a single
bidirectional fifo of the name given in the config, then relabels that
and continues without looking for / relabelling the two unidirectional
fifos named ${name}.in and ${name}.out, qemu looks in the opposite
order. So if the user had naively created all three fifos, libvirt
would relabel the bidirectional fifo to allow qemu access, but qemu
would attempt to use the two unidirectional fifos and fail (because it
didn't have proper permissions/rights).
This patch changes the order that libvirt looks for the fifos to match
what qemu does - first it looks for the dual fifos, then it looks for
the single bidirectional fifo. If it finds the dual unidirectional
fifos first, it labels/chowns them and ignores any possible
bidirectional fifo.
(Note commit d37c6a3a (which first appeared in libvirt-0.9.2) added
the code that checked for a bidirectional fifo. Prior to that commit,
bidirectional fifos for serial devices didn't work because libvirt
always required the ${name}.(in|out) fifos to exist, and qemu would
always prefer those.
This patch was made in response to:
https://bugzilla.redhat.com/show_bug.cgi?id=738095
In short, qemu's default for the rombar setting (which makes the
firmware ROM of a PCI device visible/not on the guest) was previously
0 (not visible), but they recently changed the default to 1
(visible). Unfortunately, there are some PCI devices that fail in the
guest when rombar is 1, so the setting must be exposed in libvirt to
prevent a regression in behavior (it will still require explicitly
setting <rom bar='off'/> in the guest XML).
rombar is forced on/off by adding:
<rom bar='on|off'/>
inside a <hostdev> element that defines a PCI device. It is currently
ignored for all other types of devices.
At the moment there is no clean method to determine whether or not the
rombar option is supported by QEMU - this patch uses the advice of a
QEMU developer to assume support for qemu-0.12+. There is currently a
patch in the works to put this information in the output of "qemu-kvm
-device pci-assign,?", but of course if we switch to keying off that,
we would lose support for setting rombar on all the versions of qemu
between 0.12 and whatever version gets that patch.
virFileOpenAs takes desired uid:gid as arguments, and not only uses
them for a fork/setuid/setgid when retrying failed open operations,
but additionally always forces the opened file to be owned by the
given uid:gid.
One example of the problems this causes is that, when restoring a
domain from a file that is owned by the qemu user, opening the file
chowns it to root. if dynamic_ownership=1 this is coincidentally
expected, but if dynamic_ownership=0, no existing file should ever
have its ownership changed.
This patch adds an extra check before calling fchown() - it only does
it if O_CREAT was passed to virFileOpenAs() in the openflags.
In some versions of qemu, both virtio-blk-pci and virtio-net-pci
devices can have an event_idx setting that determines some details of
event processing. When it is enabled, it "reduces the number of
interrupts and exits for the guest". qemu will automatically enable
this feature when it is available, but there may be cases where this
new feature could actually make performance worse (NB: no such case
has been found so far).
As a safety switch in case such a situation is encountered in the
field, this patch adds a new attribute "event_idx" to the <driver>
element of both disk and interface devices. event_idx can be set to
"on" (to force event_idx on in case qemu has it disabled by default)
or "off" (for force event_idx off). In the case that event_idx support
isn't present in qemu, the attribute is ignored (this on the advice of
the qemu developer).
docs/formatdomain.html.in: document the new flag (marking it as
"don't mess with this!"
docs/schemas/domain.rng: add event_idx in appropriate places
src/conf/domain_conf.[ch]: add event_idx to parser and formatter
src/libvirt_private.syms: export
virDomainVirtioEventIdx(From|To)String
src/qemu/qemu_capabilities.[ch]: detect and report event_idx in
disk/net
src/qemu/qemu_command.c: add event_idx parameter to qemu commandline
when appropriate.
tests/qemuxml2argvdata/qemuxml2argv-event_idx.args,
tests/qemuxml2argvdata/qemuxml2argv-event_idx.xml,
tests/qemuxml2argvtest.c,
tests/qemuxml2xmltest.c: test cases for event_idx.
While the first encountered dns host record is being parsed, it's
possible for virNetworkDef::hosts to point to memory that has been
allocated, but virNetworkDef::nhosts to still be 0. If there is a
failure during that time, virNetworkDef::hosts will be leaked.
Although this isn't currently the case for virNetworkDef::txtrecords,
it could become that way through future re-factoring, and it hurts
nothing to restructure the freeing of txtrecord data to match that of
hosts data.
Someone in an IRC channel or an email pointed out a few days ago that
the examples of IPv6 addresses in the libvirt documentation were not
in the officially reserved "documentation" range. This addresses their
concern.
This should have been done with the rest of the patch for virtual
switch / network device abstraction. If documents the new elements
(and new usage of existing elements) in the <network> XML to support
libvirt networks that use existing host bridges and macvtap direct
connections, as well as the new <portgroup> element.
This addresses https://bugzilla.redhat.com/show_bug.cgi?id=713728
When "defining" a new network (or one that exists but isn't currently
active) the new definition is stored in network->def, but for a
network that already exists and is active, the new definition is
stored in network->newDef, and then moved over to network->def as soon
as the network is destroyed.
However, the code that writes the dhcp and dns hosts files used by
dnsmasq was always using network->def for its information, even when
the new data was actually in network->newDef, so the hosts files
always lagged one edit behind the definition.
This patch changes the code to keep the pointer to the new definition
after it's been assigned into the network, and use it directly
(regardless of whether it's stored in network->newDef or network->def)
to construct the hosts files.
This is in response to:
https://bugzilla.redhat.com/show_bug.cgi?id=723862
which points out that a guest on an "isolated" network could
potentially exploit the DNS forwarding provided by dnsmasq to create a
communication channel to the outside.
This patch eliminates that possibility by adding the "--no-resolv"
argument to the dnsmasq commandline, which tells dnsmasq to not
forward on any requests that it can't resolve itself (by looking at
its own static hosts files and runtime list of dhcp clients), but to
instead return a failure for those requests.
This shouldn't cause any undesirable change from current
behavior, even in the case where a guest is currently configured with
multiple interfaces, one of them being connected to an isolated
network, and another to a network that does have connectivity to the
outside. If the isolated network's DNS server is queried for a name
it doesn't know, it will return "Refused" rather than "Unknown", which
indicates to the guest that it should query other servers, so it then
queries the connected DNS server, and gets the desired response.
This failure was introduced by commit dacee3d, which removed
listenAddr from the unions in virDomainGraphicsDef in favor of putting
it in the address attribute of virDomainGraphicsListenDef.
The domain XML now understands the <listen> subelement of its
<graphics> element (including when listen type='network'), and the
network driver has an internal API that will turn a network name into
an IP address, so the final logical step is to put the glue into the
qemu driver so that when it is starting up a domain, if it finds
<listen type='network' network='xyz'/> in the XML, it will call the
network driver to get an IPv4 address associated with network xyz, and
tell qemu to listen for vnc (or spice) on that address rather than the
default address (localhost).
The motivation for this is that a large installation may want the
guests' VNC servers listening on physical interfaces rather than
localhost, so that users can connect directly from the outside; this
requires sending qemu the appropriate IP address to listen on. But
this address will of course be different for each host, and if a guest
might be migrated around from one host to another, it's important that
the guest's config not have any information embedded in it that is
specific to one particular host. <listen type='network.../> can solve
this problem in the following manner:
1) on each host, define a libvirt network of the same name,
associated with the interface on that host that should be used
for listening (for example, a simple macvtap network: <forward
mode='bridge' dev='eth0'/>, or host bridge network: <forward
mode='bridge'/> <bridge name='br0'/>
2) in the <graphics> element of each guest's domain xml, tell vnc to
listen on the network name used in step 1:
<graphics type='vnc' port='5922'>
<listen type='network'network='example-net'/>
</graphics>
(all the above also applies for graphics type='spice').
Once it's plugged in, the <listen> element will be an optional
replacement for the "listen" attribute that graphics elements already
have. If the <listen> element is type='address', it will have an
attribute called 'address' which will contain an IP address or dns
name that the guest's display server should listen on. If, however,
type='network', the <listen> element should have an attribute called
'network' that will be set to the name of a network configuration to
get the IP address from.
* docs/schemas/domain.rng: updated to allow the <listen> element
* docs/formatdomain.html.in: document the <listen> element and its
attributes.
* src/conf/domain_conf.[hc]:
1) The domain parser, formatter, and data structure are modified to
support 0 or more <listen> subelements to each <graphics>
element. The old style "legacy" listen attribute is also still
accepted, and will be stored internally just as if it were a
separate <listen> element. On output (i.e. format), the address
attribute of the first <listen> element of type 'address' will be
duplicated in the legacy "listen" attribute of the <graphic>
element.
2) The "listenAddr" attribute has been removed from the unions in
virDomainGRaphicsDef for graphics types vnc, rdp, and spice.
This attribute is now in the <listen> subelement (aka
virDomainGraphicsListenDef)
3) Helper functions were written to provide simple access
(both Get and Set) to the listen elements and their attributes.
* src/libvirt_private.syms: export the listen helper functions
* src/qemu/qemu_command.c, src/qemu/qemu_hotplug.c,
src/qemu/qemu_migration.c, src/vbox/vbox_tmpl.c,
src/vmx/vmx.c, src/xenxs/xen_sxpr.c, src/xenxs/xen_xm.c
Modify all these files to use the listen helper functions rather
than directly referencing the (now missing) listenAddr
attribute. There can be multiple <listen> elements to a single
<graphics>, but the drivers all currently only support one, so all
replacements of direct access with a helper function indicate index
"0".
* tests/* - only 3 of these are new files added explicitly to test the
new <listen> element. All the others have been modified to reflect
the fact that any legacy "listen" attributes passed in to the domain
parse will be saved in a <listen> element (i.e. one of the
virDomainGraphicsListenDefs), and during the domain format function,
both the <listen> element as well as the legacy attributes will be
output.
Although most functions in libvirt return 0 on success and < 0 on
failure, there are a few functions lingering around that return errno
(a positive value) on failure, and sometimes code calling those
functions incorrectly assumes the <0 standard. I noticed one of these
the other day when auditing networkStartDhcpDaemon after Guido Gunther
found a place where success was improperly returned on failure (that
patch has been acked and is pending a push). The problem was that it
expected the return value from virFileReadPid to be < 0 on failure,
but it was actually positive (it was also neglected to set the return
code in this case, similar to the bug found by Guido).
This all led to the fact that *all* of the virFile*Pid functions in
util.c are returning errno on failure. This patch remedies that
problem by changing them all to return -errno on failure, and makes
any necessary changes to callers of the functions. (In the meantime, I
also properly set the return code on failure of virFileReadPid in
networkStartDhcpDaemon).
The new listenNetwork attribute needs to learn an IP address based on a
named network. This patch provides a function networkGetNetworkAddress
which provides that.
Some networks have an IP address explicitly in their configuration
(ie, those with a forward type of "none", "route", or "nat"). For
those, we can just return the IP address from the config.
The rest will have a physical device associated with them (either via
<bridge name='...'/>, <forward ... dev='...'/>, or possibly via a pool
of interfaces inside the network's <forward> element) and we will need
to ask the kernel for a current IP address of that device (via the
newly added ifaceGetIPAddress)
If networkGetNetworkAddress encounters an error while trying to learn
the address for a network, it will return -1. In the case that libvirt
has been compiled without the network driver, the call is a macro
which reduces to -2. This allows differentiating between a failure of
the network driver, and its complete absence.
This function uses ioctl(SIOCGIFADDR), which limits it to returning
the first IPv4 address of an interface, but that's what we want right
now (the place we're going to use the address only accepts one).
All of the functions in util/interface.c were returning 0 on success,
but some returned -1 on error, and some returned a positive value
(usually the value of errno, but sometimes just 1). Libvirt's standard
is to return < 0 on error (in the case of functions that need to
return errno, -errno is returned.
This patch modifies all functions in interface.c to consistently
return < 0 on error, and makes changes to callers of those functions
where necessary.
The network driver needs to assign physical devices for use by modes
that use macvtap, keeping track of which physical devices are in use
(and how many instances, when the devices can be shared). Three calls
are added:
networkAllocateActualDevice - finds a physical device for use by the
domain, and sets up the virDomainActualNetDef accordingly.
networkNotifyActualDevice - assumes that the domain was already
running, but libvirtd was restarted, and needs to be notified by each
already-running domain about what interfaces they are using.
networkReleaseActualDevice - decrements the usage count of the
allocated physical device, and frees the virDomainActualNetDef to
avoid later accidentally using the device.
bridge_driver.[hc] - the new APIs. When WITH_NETWORK is false, these
functions are all #defined to be "0" in the .h file (effectively
becoming a NOP) to prevent link errors.
qemu_(command|driver|hotplug|process).c - add calls to the above APIs
in the appropriate places.
tests/Makefile.am - we need to include libvirt_driver_network.la
whenever libvirt_driver_qemu.la is linked, to avoid unreferenced
symbols (in functions that are never called by the test
programs...)
This is the one function outside of domain_conf.c that plays around
with (even modifying) the internals of the virDomainNetDef, and thus
can't be fixed up simply by replacing direct accesses to the fields of
the struct with the GetActual*() access functions.
In this case, we need to check if the defined type is "network", and
if it is *then* check the actual type; if the actual type is "bridge",
then we can at least put the bridgename in a place where it can be
used; otherwise (if type isn't "bridge"), we behave exactly as we used
to - just null out *everything*.
The qemu driver accesses fields in the virDomainNetDef directly, but
with the advent of the virDomainActualNetDef, some pieces of
information may be found in a different place (the ActualNetDef) if
the network connection is of type='network' and that network is of
forward type='bridge|private|vepa|passthrough'. The previous patch
added functions to mask this difference from callers - they hide the
decision making process and just pick the value from the proper place.
This patch uses those functions in the qemu driver as a first step in
making qemu work with the new network types. At this point, the
virDomainActualNetDef is guaranteed always NULL, so the GetActualX()
function will return exactly what the def->X that's being replaced
would have returned (ie bisecting is not compromised).
There is one place (in qemu_driver.c) where the internal details of
the NetDef are directly manipulated by the code, so the GetActual
functions cannot be used there without extra additional code; that
file will be treated in a separate patch.
Previously all networks were composed of bridge devices created and
managed by libvirt, and the same operations needed to be done for all
of them when they were started and stopped (create and start the
bridge device, configure its MAC address and IP address, add iptables
rules). The new network types are (for now at least) managed outside
of libvirt, and the network object is used only to contain information
about the network, which is then used as each individual guest
connects itself.
This means that when starting/stopping one of these new networks, we
really want to do nothing, aside from marking the network as
active/inactive.
This has been setup as toplevel Start/Shutdown functions that do the
small bit of common stuff, then have a switch statement to execute
network type-specific start/shutdown code, then do a bit more common
code. The type-specific functions called for the new host bridge and
macvtap based types are currently empty.
In the future these functions may actually do something, and we will
surely add more functions that are similarly patterned. Once
everything has settled, we can make a table of "sub-driver" function
pointers for each network type, and store a pointer to that table in
the network object, then we can replace the switch statements with
calls to functions in the table.
The final step in this will be to add a new table (and corresponding
new functions) for new network types as they are added.
The network XML is updated in the following ways:
1) The <forward> element can now contain a list of forward interfaces:
<forward .... >
<interface dev='eth10'/>
<interface dev='eth11'/>
<interface dev='eth12'/>
<interface dev='eth13'/>
</forward>
The first of these takes the place of the dev attribute that is
normally in <forward> - when defining a network you can specify
either one, and on output both will be present. If you specify
both on input, they must match.
2) In addition to forward modes of 'nat' and 'route', these new modes
are supported:
private, passthrough, vepa - when this network is referenced by a
domain's interface, it will have the same effect as if the
interface had been defined as type='direct', e.g.:
<interface type='direct'>
<source mode='${mode}' dev='${dev}>
...
</interface>
where ${mode} is one of the three new modes, and ${dev} is an interface
selected from the list given in <forward>.
bridge - if a <forward> dev (or multiple devs) is defined, and
forward mode is 'bridge' this is just like the modes 'private',
'passthrough', and 'vepa' above. If there is no forward dev
specified but a bridge name is given (e.g. "<bridge
name='br0'/>"), then guest interfaces using this network will use
libvirt's "host bridge" mode, equivalent to this:
<interface type='bridge'>
<source bridge='${bridge-name}'/>
...
</interface>
3) A network can have multiple <portgroup> elements, which may be
selected by the guest interface definition (by adding
"portgroup='${name}'" in the <source> element along with the
network name). Currently a portgroup can only contain a
virtportprofile, but the intent is that other configuration items
may be put there int the future (e.g. bandwidth config). When
building a guest's interface, if the <interface> XML itself has no
virtportprofile, and if the requested network has a portgroup with
a name matching the name given in the <interface> (or if one of the
network's portgroups is marked with the "default='yes'" attribute),
the virtportprofile from that portgroup will be used by the
interface.
4) A network can have a virtportprofile defined at the top level,
which will be used by a guest interface when connecting in one of
the 'direct' modes if the guest interface XML itself hasn't
specified any virtportprofile, and if there are also no matching
portgroups on the network.
the domain XML <interface> element is updated in the following ways:
1) <virtualportprofile> can be specified when source type='network'
(previously it was only valid for source type='direct')
2) A new attribute "portgroup" has been added to the <source>
element. When source type='network' (the only time portgroup is
recognized), extra configuration information will be taken from the
<portgroup> element of the given name in the network definition.
3) Each virDomainNetDef now also potentially has a
virDomainActualNetDef which is a private object (never
exported/imported via the public API, and not defined in the RNG) that
is used to maintain information about the physical device that was
actually used for a NetDef of type VIR_DOMAIN_NET_TYPE_NETWORK.
The virDomainActualNetDef will only be parsed/formatted if the
parse/format function is called with the
VIR_DOMAIN_XML_INTERNAL_ACTUAL_NET flag set (which is only needed when
saving/loading a running domain's state info to the stateDir).
The virtPortProfile in the domain interface struct is now a separately
allocated object *pointed to by* (rather than contained in) the main
virDomainNetDef object. This is done to make it easier to figure out
when a virtualPortProfile has/hasn't been specified in a particular
config.
virtPortProfiles are currently only used in the domain XML, but will
soon also be used in the network XML. To prepare for that change, this
patch moves the structure definition into util/network.h and the parse
and format functions into util/network.c (I decided that this was a
better choice than macvtap.h/c for something that needed to always be
available on all platforms).
domain.rng, network.rng, and interface.rng already use a few of the
same types (or in some cases *should* but don't), and an upcoming code
change will have them sharing even more. To prepare for that, this
patch takes those common data type definitions and moves them into
basictypes.rng.
This may break some rule about the need to RNG files to be autonomous
or something, but I saw that storageencryption.rng is used in this
way, so I figured it must not be completely against the law...
Although most functions with flags check to verify no application is
passing in flag bits that are currently undefined, for some reason
this function wasn't.
* Change all flags args from int to unsigned int
* Allow passing flags in virDomainObjParseFile (and propogate those
flags all the way down the call chain). Previously the flags were
hardcoded (to VIR_DOMAIN_XML_INTERNAL_STATUS) several layers down
the chain. Pass that value in at the one place that is currently
calling virDomainObjParseFile.
If a domain name is defined for a network, add the --expand-hosts
option to the dnsmasq commandline. This results in the domain being
added to any hostname that is defined in a dns <host> element and
contains no '.' characters (i.e. it is an "unqualified"
hostname). Since PTR records are automatically created for any name
defined in <host>, the result of a PTR request will change from the
unqualified name to the qualified name.
This also has the same effect on any hostnames that dnsmasq reads
from the host's /etc/hosts file.
(In the case of guest hostnames that were learned by dnsmasq via DHCP
requests, they were already getting the domain name added on, even
without --expand-hosts).
The sub-elements of <ip> had been placed at the same level of
indentation as ip itself, implying that they were really elements of
<network>. Within that, sub-elements of ip/dhcp were also at that same
level. These have been double-indented.
At the same time, I realized that the documentation for the new <dns>
element had been placed right in the middle of the description of the
sub-elements of <ip>. I moved it up out of the way.
This fixes:
https://bugzilla.redhat.com/show_bug.cgi?id=702044https://bugzilla.redhat.com/show_bug.cgi?id=709454
Both of these complain of a failure to use an image file that resides
on a read-only NFS volume. The function in the DAC security driver
that chowns image files to the qemu user:group before using them
already has special cases to ignore failure of chown on read-only file
systems, and in a few other cases, but it hadn't been checking for
EINVAL, which is what is returned if the qemu user doesn't even exist
on the NFS server.
Since the explanation of EINVAL in the chown man page almost exactly
matches the log message already present for the case of EOPNOTSUPP,
I've just added EINVAL to that same conditional.
It was suggested during review of a different patch that the libvirt
interface driver API's should have "netcf:" in their log
messages. This patch eliminates that from all interface driver API
functions, and also eliminates the extra " - " in the case that netcf
returns no details in its error info (which *never* happens at
present, but could happen sometime in the future.
NB: the enum that uses the string vnet-host (now changed to vhost-net)
is used in XML, but fortunately that hasn't been in an official
release yet, so it can still be fixed.
This was discussed in:
https://www.redhat.com/archives/libvir-list/2011-May/msg01370.html
The capabilities code only sets the flag to allow use of vhost-net if
kvm is detected (set if the help string contains "(qemu-kvm-" or
"(kvm-"), but actually vhost-net is available in some qemu builds that
don't have kvm in their name, so just checking for ",vhost=" is enough.
The lone caller to hostsFileWrite (and the callers for at least 3
levels up the return stack) assume that the return value will be < 0
on failure. However, hostsFileWrite returns 0 on success, and a
positive errno on failure. This patch changes hostsFileWrite to return
-errno on failure.
This patch addresses:
https://bugzilla.redhat.com/show_bug.cgi?id=694382
In order to give each libvirt-created bridge a fixed MAC address,
commit 5754dbd56d, added code to create
a dummy tap device with guaranteed lowest MAC address and attach it to
the bridge. This tap device was given the name "${bridgename}-nic".
However, an interface device name must be IFNAMSIZ (15) characters or
less, so a valid ${bridgename} such as "verylongname123" (15
characters) would lead to an invalid tap device name
("verylongname123-nic" - 19 characters), and that in turn led to a
failure to bring up the network.
The solution is to shorten the part of the original name used to
generate the tap device name. However, simply truncating it is
insufficient, because the last few characters of an interface name are
often a number used to indicate one of a list of several similar
devices (for example, "verylongname123", "verylongname124", etc) and
simple truncation would lead to duplicate names (eg "verlongnam-nic"
and "verylongnam-nic"). So instead we take the first 8 characters of
$bridgename ("verylong" in the example), add on the final 3 bytes
("123"), then add "-nic" (so "verylong123-nic"). Not pretty, but it
is much more likely to generate a unique name, and is reproducible
(unlike, say, a random number).
It was just pointed out that, although I added documentation for the
IPv6 additions to the network XML, I neglected to use those additions
in the examples. This patch adds an IPv6 address to each of the
examples except for the "default" network, since that is a faithful
reproduction of the default network config that's automatically
installed, which doesn't include any IPv6 address (for good reason -
because there is no such thing as IPv6 NAT, there is no one IPv6
address that would work for all installations).
This is detailed in:
https://bugzilla.redhat.com/show_bug.cgi?id=688957
Since radvd is executed by daemonizing it, the attempt to exec the
radvd binary doesn't happen until after libvirtd has already received
an exit code from the intermediate forked process, so no error is
detected or logged by __virExec().
We can't require radvd as a prerequisite for the libvirt package (many
installations don't use IPv6, so they don't need it), so instead we
add in a check to verify there is an executable radvd binary prior to
trying to exec it.
A bug in libnl (see https://bugzilla.redhat.com/show_bug.cgi?id=677724
and https://bugzilla.redhat.com/show_bug.cgi?id=677725) makes it very
easy to create a failure to connect to the netlink socket when trying
to open a macvtap network device ("type='direct'" in domain interface
XML). When that error occurred (during a call to libnl's nl_connect()
from libvirt's nlComm(), there was no log message, leading virsh (for
example) to report "unknown error".
There were two other cases in nlComm where an error in a libnl
function might return with failure but no error reported. In all three
cases, this patch logs a message which will hopefully be more useful.
Note that more detailed information about the failure might be
available from libnl's nl_geterror() function, but it calls
strerror(), which is not threadsafe, so we can't use it.
The newly added call to qemuAuditNetDevice in qemuPhysIfaceConnect was
assuming that res_ifname (the name of the macvtap device) was always
valid, but this isn't the case. If openMacvtapTap fails, it always
returns NULL, which would result in a segv.
Since the audit log only needs a record of devices that are actually
sent to qemu, and a failure to open the macvtap device means that no
device will be sent to qemu, we can solve this problem by only doing
the audit if openMacvtapTap is successful (in which case res_ifname is
guaranteed valid).
Normally dnsmasq will send a default route (the address of the host in
the network definition) to any client requesting an address via
DHCP. On an isolated network this makes no sense, as we have iptables
to prevent any traffic going out via that interface, so anything sent
that way would be dropped anyway.
This extra/unusable default route becomes problematic if you have
setup a guest with multiple network interfaces, with one connected to
an isolated network and another that provides connectivity to the
outside (example - one interface directly connecting to a physical
interface via macvtap, with a second connected to an isolated network
so that the host and guest can communicate (macvtap doesn't support
guest<->host communication without an external switch that supports
vepa, or reflecting all traffic back)). In this case, if the guest
chooses the default route of the isolated network, the guest will not
be able to get network traffic beyond the host.
To prevent dnsmasq from sending a default route, you can tell it to
send 0 bytes of data for the default route option (option number 3)
with --dhcp-option=3 (normally the data to send for the option would
follow the option number; no extra data means "don't send this option").
I have checked on RHEL5 (a good representative of the oldest supported
libvirt platforms) and its version of dnsmasq (2.45) does support
--dhcp-option, so this shouldn't create any compatibility problems.
By default, all dnsmasq processes share the same leases file. libvirt
also uses the --dhcp-lease-max option to control the maximum number of
leases allowed. The problem is that libvirt puts in a number equal to
the number of addresses in the range for the one network handled by a
single instance of dnsmasq, but dnsmasq checks the total number of
leases in the file (which could potentially contain many more).
The solution is to tell each instance of dnsmasq to create and use its
own leases file. (/var/lib/libvirt/network/<net-name>.leases).
This file is created by dnsmasq when it starts, but not deleted when
it exists. This is fine when the network is just being stopped, but if
the leases file was left around when a network was undefined, we could
end up with an ever-increasing number of dead files - instead, we
explicitly unlink the leases file when a network is undefined.
Note that Ubuntu carries a patch against an older version of libvirt for this:
hhttps://bugs.launchpad.net/ubuntu/+source/libvirt/+bug/713071
ttp://bazaar.launchpad.net/~serge-hallyn/ubuntu/maverick/libvirt/bugall/revision/109
I was certain I'd also seen discussion of this on libvir-list or
libvirt-users, but couldn't find it.
This fixes a regression introduced in commit ad48df, and reported on
the libvirt-users list:
https://www.redhat.com/archives/libvirt-users/2011-March/msg00018.html
The problem in that commit was that we began searching a list of ip
address definitions (rather than just having one) to look for a dhcp
range or static host; when we didn't find any, our pointer (ipdef) was
left at NULL, and when ipdef was NULL, we returned without starting up
dnsmasq.
Previously dnsmasq was started even without any dhcp ranges or static
entries, because it's still useful for DNS services.
Another problem I noticed while investigating was that, if there are
IPv6 addresses, but no IPv4 addresses of any kind, we would jump out
at an ever higher level in the call chain.
This patch does the following:
1) networkBuildDnsmasqArgv() = all uses of ipdef are protected from
NULL dereference. (this patch doesn't change indentation, to make
review easier. The next patch will change just the
indentation). ipdef is intended to point to the first IPv4 address
with DHCP info (or the first IPv4 address if none of them have any
dhcp info).
2) networkStartDhcpDaemon() = if the loop looking for an ipdef with
DHCP info comes up empty, we then grab the first IPv4 def from the
list. Also, instead of returning if there are no IPv4 defs, we just
return if there are no IP defs at all (either v4 or v6). This way a
network that is IPv6-only will still get dnsmasq listening for DNS
queries.
3) in networkStartNetworkDaemon() - we will startup dhcp not just if there
are any IPv4 addresses, but also if there are any IPv6 addresses.
This was also found while investigating
https://bugzilla.redhat.com/show_bug.cgi?id=670848
An EOF on a domain's monitor socket results in an event being queued
to handle the EOF. The handler calls qemuProcessHandleMonitorEOF. If
it is a transient domain, this leads to a call to
virDomainRemoveInactive, which removes the domain from the driver's
hashtable and unref's it. Nowhere in this code is the qemu driver lock
acquired.
However, all modifications to the driver's domain hashtable *must* be
done while holding the driver lock, otherwise the hashtable can become
corrupt, and (even more likely) another thread could call a different
hashtable function and acquire a pointer to the domain that is in the
process of being destroyed.
To prevent such a disaster, qemuProcessHandleMonitorEOF must get the
qemu driver lock *before* it gets the DomainObj's lock, and hold it
until it is finished with the DomainObj. This guarantees that nobody
else modifies the hashtable at the same time, and that anyone who had
already gotten the DomainObj from the hashtable prior to this call has
finished with it before we remove/destroy it.
This was found while researching the root cause of:
https://bugzilla.redhat.com/show_bug.cgi?id=670848
virDomainUnref should only be called with the lock held for the
virDomainObj in question. However, when a transient qemu domain gets
EOF on its monitor socket, it queues an event which frees the monitor,
which unref's the virDomainObj without first locking it. If another
thread has already locked the virDomainObj, the modification of the
refcount could potentially be corrupted. In an extreme case, it could
also be potentially unlocked by virDomainObjFree, thus left open to
modification by anyone else who would have otherwise waited for the
lock (not to mention the fact that they would be accessing freed
data!).
The solution is to have qemuMonitorFree lock the domain object right
before unrefing it. Since the caller to qemuMonitorFree doesn't expect
this lock to be held, if the refcount doesn't go all the way to 0,
qemuMonitorFree must unlock it after the unref.
This fixes https://bugzilla.redhat.com/show_bug.cgi?id=609463
The problem was that, since a bridge always acquires the MAC address
of the connected interface with the numerically lowest MAC, as guests
are started and stopped, it was possible for the MAC address to change
over time, and this change in the network was being detected by
Windows 7 (it sees the MAC of the default route change), so on each
reboot it would bring up a dialog box asking about this "new network".
The solution is to create a dummy tap interface with a MAC guaranteed
to be lower than any guest interface's MAC, and attach that tap to the
bridge as soon as it's created. Since all guest MAC addresses start
with 0xFE, we can just generate a MAC with the standard "0x52, 0x54,
0" prefix, and it's guaranteed to always win (physical interfaces are
never connected to these bridges, so we don't need to worry about
competing numerically with them).
Note that the dummy tap is never set to IFF_UP state - that's not
necessary in order for the bridge to take its MAC, and not setting it
to UP eliminates the clutter of having an (eg) "virbr0-nic" displayed
in the output of the ifconfig command.
I chose to not auto-generate the MAC address in the network XML
parser, as there are likely to be consumers of that API that don't
need or want to have a MAC address associated with the
bridge.
Instead, in bridge_driver.c when the network is being defined, if
there is no MAC, one is generated. To account for virtual network
configs that already exist when upgrading from an older version of
libvirt, I've added a %post script to the specfile that searches for
all network definitions in both the config directory
(/etc/libvirt/qemu/networks) and the state directory
(/var/lib/libvirt/network) that are missing a mac address, generates a
random address, and adds it to the config (and a matching address to
the state file, if there is one).
docs/formatnetwork.html.in: document <mac address.../>
docs/schemas/network.rng: add nac address to schema
libvirt.spec.in: %post script to update existing networks
src/conf/network_conf.[ch]: parse and format <mac address.../>
src/libvirt_private.syms: export a couple private symbols we need
src/network/bridge_driver.c:
auto-generate mac address when needed,
create dummy interface if mac address is present.
tests/networkxml2xmlin/isolated-network.xml
tests/networkxml2xmlin/routed-network.xml
tests/networkxml2xmlout/isolated-network.xml
tests/networkxml2xmlout/routed-network.xml: add mac address to some tests
An upcoming patch has a use for a tap device to be created that
doesn't need to be actually put into the "up" state, and keeping it
"down" keeps the output of ifconfig from being unnecessarily cluttered
(ifconfig won't show down interfaces unless you add "-a").
bridge.[ch]: add "up" as an arg to brAddTap()
uml_conf.c, qemu_command.c: add "up" (set to "true") to brAddTap() call.
This is in response to:
https://bugzilla.redhat.com/show_bug.cgi?id=629662
Explanation
qemu's virtio-net-pci driver allows setting the algorithm used for tx
packets to either "bh" or "timer". This is done by adding ",tx=bh" or
",tx=timer" to the "-device virtio-net-pci" commandline option.
'bh' stands for 'bottom half'; when this is set, packet tx is all done
in an iothread in the bottom half of the driver. (In libvirt, this
option is called the more descriptive "iothread".)
'timer' means that tx work is done in qemu, and if there is more tx
data than can be sent at the present time, a timer is set before qemu
moves on to do other things; when the timer fires, another attempt is
made to send more data. (libvirt retains the name "timer" for this
option.)
The resulting difference, according to the qemu developer who added
the option is:
bh makes tx more asynchronous and reduces latency, but potentially
causes more processor bandwidth contention since the cpu doing the
tx isn't necessarily the cpu where the guest generated the
packets.
Solution
This patch provides a libvirt domain xml knob to change the option on
the qemu commandline, by adding a new attribute "txmode" to the
<driver> element that can be placed inside any <interface> element in
a domain definition. It's use would be something like this:
<interface ...>
...
<model type='virtio'/>
<driver txmode='iothread'/>
...
</interface>
I chose to put this setting as an attribute to <driver> rather than as
a sub-element to <tune> because it is specific to the virtio-net
driver, not something that is generally usable by all network drivers.
(note that this is the same placement as the "driver name=..."
attribute used to choose kernel vs. userland backend for the
virtio-net driver.)
Actually adding the tx=xxx option to the qemu commandline is only done
if the version of qemu being used advertises it in the output of
qemu -device virtio-net-pci,?
If a particular txmode is requested in the XML, and the option isn't
listed in that help output, an UNSUPPORTED_CONFIG error is logged, and
the domain fails to start.
When the <driver> element (and its "name" attribute) was added to the
domain XML's interface element, a "backend" enum was simply added to
the toplevel of the virDomainNetDef struct.
Ignoring the naming inconsistency ("name" vs. "backend"), this is fine
when there's only a single item contained in the driver element of the
XML, but doesn't scale well as we add more attributes that apply to
the backend of the virtio-net driver, or add attributes applicable to
other drivers.
This patch changes virDomainNetDef in two ways:
1) Rename the item in the struct from "backend" to "name", so that
it's the same in the XML and in the struct, hopefully avoiding
confusion for someone unfamiliar with the function of the
attribute.
2) Create a "driver" union within virDomainNetDef, and a "virtio"
struct in that struct, which contains the "name" enum value.
3) Move around the virDomainNetParse and virDomainNetFormat functions
to allow for simple plugin of new attributes without disturbing
existing code. (you'll note that this results in a seemingly
redundant if() in the format function, but that will no longer be
the case as soon as a 2nd attribute is added).
In the future, new attributes for the virtio driver backend can be
added to the "virtio" struct, and any other network device backend that
needs an attribute will have its own struct added to the "driver"
union.
If a guest image is saved in compressed format, and the restore fails
in some way after the intermediate process used to uncompress the
image has been started, but before qemu has been started to hook up to
the uncompressor, libvirt will endlessly wait for the uncompressor to
finish, but it never will because it's still waiting to have something
hooked up to drain its output.
The solution is to close the pipes on both sides of the uncompressor,
then send a SIGTERM before calling waitpid on it (only if the restore
has failed, of course).
This patch is a partial resolution to the following bug:
https://bugzilla.redhat.com/show_bug.cgi?id=667756
(to complete the fix, an updated selinux-policy package is required,
to add the policy that allows libvirt to set the context of a fifo,
which was previously not allowed).
Explanation : When an incoming migration is over a pipe (for example,
if the image was compressed and is being fed through gzip, or was on a
root-squash nfs server, so needed to be opened by a child process
running as a different uid), qemu cannot read it unless the selinux
context label for the pipe has been set properly.
The solution is to check the fd used as the source of the migration
just before passing it to qemu; if it's a fifo (implying that it's a
pipe), we call the newly added virSecurityManagerSetFDLabel() function
to set the context properly.
A need was found to set the SELinux context label on an open fd (a
pipe, as a matter of fact). This patch adds a function to the security
driver API that will set the label on an open fd to secdef.label. For
all drivers other than the SELinux driver, it's a NOP. For the SElinux
driver, it calls fsetfilecon().
If the return is a failure, it only returns error up to the caller if
1) the desired label is different from the existing label, 2) the
destination fd is of a type that supports setting the selinux context,
and 3) selinux is in enforcing mode. Otherwise it will return
success. This follows the pattern of the existing function
SELinuxSetFilecon().
This is in response to a request in:
https://bugzilla.redhat.com/show_bug.cgi?id=665293
In short, under heavy load, it's possible for qemu's networking to
lock up due to the tap device's default 1MB sndbuf being
inadequate. adding "sndbuf=0" to the qemu commandline -netdevice
option will alleviate this problem (sndbuf=0 actually sets it to
0xffffffff).
Because we must be able to explicitly specify "0" as a value, the
standard practice of "0 means not specified" won't work here. Instead,
virDomainNetDef also has a sndbuf_specified, which defaults to 0, but
is set to 1 if some value was given.
The sndbuf value is put inside a <tune> element of each <interface> in
the domain. The intent is that further tunable settings will also be
placed inside this element.
<interface type='network'>
...
<tune>
<sndbuf>0</sndbuf>
...
</tune>
</interface>
This patch is in response to
https://bugzilla.redhat.com/show_bug.cgi?id=643050
The existing libvirt support for the vhost-net backend to the virtio
network driver happens automatically - if the vhost-net device is
available, it is always enabled, otherwise the standard userland
virtio backend is used.
This patch makes it possible to force whether or not vhost-net is used
with a bit of XML. Adding a <driver> element to the interface XML, eg:
<interface type="network">
<model type="virtio"/>
<driver name="vhost"/>
will force use of vhost-net (if it's not available, the domain will
fail to start). if driver name="qemu", vhost-net will not be used even
if it is available.
If there is no <driver name='xxx'/> in the config, libvirt will revert
to the pre-existing automatic behavior - use vhost-net if it's
available, and userland backend if vhost-net isn't available.
When dynamic_ownership=0, saved images must be owned by the same uid
as is used to run the qemu process, otherwise restore won't work. To
accomplish this, qemuSecurityDACRestoreSavedStateLabel() needs to
simply return when it's called.
This fix is in response to:
https://bugzilla.redhat.com/show_bug.cgi?id=661720
Although the upper-layer code protected against it, it was possible to
call iptablesForwardMasquerade() with an IPv6 address and have it
attempt to add a rule to the MASQUERADE chain of ip6tables (which
doesn't exist).
This patch changes that function to check the protocol of the given
address, generate an error log if it's not IPv4 (AF_INET), and finally
hardcodes all the family parameters sent down to lower-level functions.
This is partially in response to
https://bugzilla.redhat.com/show_bug.cgi?id=653300
The crash in that report was coincidentally fixed when we switched
from using inet_pton() to using virSocketParseAddr(), but the absence
of an ip address in a dhcp static host definition was still silently
ignored (and that entry discarded from the saved XML). This patch
turns that into a logged failure; likewise if the entry has neither a
mac address nor a name attribute (the entry is useless without at
least one of those, plus an ip address).
Since the network name is now pulled into this function in order for
those error logs to be more informative, the other error messages in
the function have also been changed to take advantage.
The original version of these functions would modify the address sent
in, meaning that the caller would usually need to copy the address
first. This change makes the original a const, and puts the resulting
masked address into a new arg (which could point to the same
virSocketAddr as the original, if the caller really wants to modify
it).
This also makes the API consistent with virSocketAddrBroadcast[ByPrefix].
Previously we used ioctl() to set the IP address and netmask of the
bridges used for virtual networks, and apparently the SIOCSIFNETMASK
ioctl implicitly set the broadcast address for the interface. The new
method of using the "ip" command requires broadcast address to be
explicitly specified though.
These functions work only for IPv4, becasue IPv6 doesn't have the same
concept of "broadcast address" as IPv4. They merely OR the inverse of
the netmask with the given host address, thus turning on all the host
bits.
This patch fixes https://bugzilla.redhat.com/show_bug.cgi?id=664406
If qemu is run as a different uid, it has been unable to access mode
0660 files that are owned by a different user, but with a group that
the qemu is a member of (aside from the one group listed in the passwd
file), because initgroups() is not being called prior to the
exec. initgroups will change the group membership of the process (and
its children) to match the new uid.
To make this happen, the setregid()/setreuid() code in
qemuSecurityDACSetProcessLabel has been replaced with a call to
virSetUIDGID(), which does both of those, plus calls initgroups.
Similar, but not identical, code in qemudOpenAsUID() has been replaced
with virSetUIDGID(). This not only consolidates the functionality to a
single location, but also potentially fixes some as-yet unreported
bugs.
virSetUIDGID() sets both the real and effective group and user of the
process, and additionally calls initgroups() to assure that the
process joins all the auxiliary groups that the given uid is a member
of.
There are cases when we want log an error message, and possibly free
some memory as part of the cleanup, while still preserving errno for a
caller, but the functions that log errors, and virFree (VIR_FREE) make
system calls that will clear errno. This patch preserves errno during
those most basic functions (corresponding to virReportSystemError(),
virReportOOMError(), networkReportError(), etc, as well as
virStrError()). It does *not preserve errno across calls to higher
level items such as virDispatchError(), as it's assumed the caller is
all finished with any need for errno by the time it dispatches the
error.
Running an instance of the router advertisement daemon (radvd) allows
guests using the virtual network to automatically acquire an IPv6
address and default route. Note that acquiring an address only works
for networks with a prefix length of exactly 64 - radvd is still run
in other circumstances, and still advertises routes, but autoconf will
not work because it requires exactly 64 bits of address info from the
network prefix.
This patch avoids a race condition with the pidfile by manually
daemonizing radvd rather than allowing it to daemonize itself, then
creating our own pidfile (in addition to radvd's own file, which is
unnecessary, but there is no way to tell radvd to not create it). This
is accomplished by exec'ing it with "--debug 1" in the commandline,
and using virCommand's features to fork, create a pidfile, and detach
from the newly forked process.
At this point everything is already in place to make IPv6 happen, we just
need to add a few rules, remove some checks for IPv4-only, and document
the changes to the XML on the website.
All of the iptables functions eventually call down to a single
bottom-level function, and fortunately, ip6tables syntax (for all the
args that we use) is identical to iptables format (except the
addresses), so all we need to do is:
1) Get an address family down to the lowest level function in each
case, either implied through an address, or explicitly when no
address is in the parameter list, and
2) At the lowest level, just decide whether to call "iptables" or
"ip6tables" based on the family.
The location of the ip6tables binary is determined at build time by
autoconf. If a particular target system happens to not have ip6tables
installed, any attempts to run it will generate an error, but that
won't happen unless someone tries to define an IPv6 address for a
network. This is identical behavior to IPv4 addresses and iptables.
This patch reorganizes the code in bridge_driver.c to account for the
concept of a single network with multiple IP addresses, without adding
in the extra variable of IPv6. A small bit of code has been
temporarily added that checks all given addresses to verify they are
IPv4 - this will be removed when full IPv6 support is turned on.
This commit adds support for IPv6 parsing and formatting to the
virtual network XML parser, including moving around data definitions
to allow for multiple <ip> elements on a single network, but only
changes the consumers of this API to accommodate for the changes in
API/structure, not to add any actual IPv6 functionality. That will
come in a later patch - this patch attempts to maintain the same final
functionality in both drivers that use the network XML parser - vbox
and "bridge" (the Linux bridge-based driver used by the qemu
hypervisor driver).
* src/libvirt_private.syms: Add new private API functions.
* src/conf/network_conf.[ch]: Change C data structure and
parsing/formatting.
* src/network/bridge_driver.c: Update to use new parser/formatter.
* src/vbox/vbox_tmpl.c: update to use new parser/formatter
* docs/schemas/network.rng: changes to the schema -
* there can now be more than one <ip> element.
* ip address is now an ip-addr (ipv4 or ipv6) rather than ipv4-addr
* new optional "prefix" attribute that can be used in place of "netmask"
* new optional "family" attribute - "ipv4" or "ipv6"
(will default to ipv4)
* define data types for the above
* tests/networkxml2xml(in|out)/nat-network.xml: add multiple <ip> elements
(including IPv6) to a single network definition to verify they are being
correctly parsed and formatted.
In practice this has always been optional, but the RNG has shown it as
mandatory, and since all the examples for make check had it, it was
never noticed. One of the existing test cases has been changed to
check for this.
I also noticed that the dhcp/host/ip was still defined as <text/>,
but should really be <ref name='ipv4-addr'/>
brSetInetAddress can only set a single IP address on the bridge, and
uses a method (ioctl(SIOCSETIFADDR)) that only works for IPv4. Replace
it and brSetInetNetmask with a single function that uses the external
"ip addr add" command to add an address/prefix to the interface - this
supports IPv6, and allows adding multiple addresses to the interface.
Although it isn't currently used in the code, we also add a
brDelInetAddress for completeness' sake.
Also, while we're modifying bridge.c, we change brSetForwardDelay and
brSetEnableSTP to use the new virCommand API rather than the
deprecated virRun, and also log an error message in bridge_driver.c if
either of those fail (previously the failure would be completely
silent).
When a netmask isn't specified for an IPv4 address, one can be implied
based on what network class range the address is in. The
virNetworkDefPrefix function does this for us, so netmask isn't
required.
IPv6 will use prefix exclusively, and IPv4 will also optionally be
able to use it, and the iptables functions really need a prefix
anyway, so use the new virNetworkDefPrefix() function to send prefixes
into iptables functions instead of netmasks.
Also, in a couple places where a netmask is actually needed, use the
new private API function for it rather than getting it directly. This
will allow for cases where no netmask or prefix is specified (it
returns the default for the current class of network.)
Some functions in this file were returning 1 on success and 0 on
failure, and others were returning 0 on success and -1 on
failure. Switch them all to return the libvirt-preferred 0/-1.
The functions in iptables.c all return -1 on failure, but all their
callers (which all happen to be in bridge_driver.c) assume that they
are returning an errno, and the logging is done accordingly. This
patch fixes all the error checking and logging to assume < 0 is an
error, and nothing else.
Later patches will add the possibility to define a network's netmask
as a prefix (0-32, or 0-128 in the case of IPv6). To make it easier to
deal with definition of both kinds (prefix or netmask), add two new
functions:
virNetworkDefNetmask: return a copy of the netmask into a
virSocketAddr. If no netmask was specified in the XML, create a
default netmask based on the network class of the virNetworkDef's IP
address.
virNetworkDefPrefix: return the netmask as numeric prefix (or the
default prefix for the network class of the virNetworkDef's IP
address, if no netmask was specified in the XML)
virSocketPrefixToNetmask: Given a 'prefix', which is the number of 1
bits in a netmask, fill in a virSocketAddr object with a netmask as an
IP address (IPv6 or IPv4).
virSocketAddrMask: Mask off the host bits in one virSocketAddr
according to the netmask in another virSocketAddr.
virSocketAddrMaskByPrefix, Mask off the host bits in a virSocketAddr
according to a prefix (number of 1 bits in netmask).
VIR_SOCKET_FAMILY: return the family of a virSocketAddr
This is pretty straightforward - even though dnsmasq gets daemonized
and uses a pid file, those things are both handled by the dnsmasq
binary itself. And libvirt doesn't need any of the output of the
dnsmasq command either, so we just setup the args and call
virRun(). Mainly it was just a (mostly) mechanical job of replacing
the APPEND_ARG() macro (and some other *printfs()) with
virCommandAddArg*().
IP addresses and MAC addresses had been defined in the RNG simply as
<text/> meaning that, according to the RNG, any string could go in
there. Of course the C parsing code does a much better job of
validating, but we may as well have this describing the contents
accurately (even though it's currently only used during "make check").
All the other RNG files in libvirt are enclosed within <grammar>. This
commit makes the syntactical changes necessary to make network.rng fit
that pattern. (This is the first step in adding some data type
definitions to network.rng for more exact validation of IP and MAC
addresses).
Formatting changes (indentation) will be done in a subsequent commit,
so that actual changes to the code won't be obscured by whitespace.
Laine Stump
Laine Stump