Before this patch we would simply rely on QEMU failing to attach the
device. Since we have a flag in the address set telling us which
controllers support hotplug, we can fail the operation sooner.
This also assures that when hotplugging with no provided PCI address,
that we skip any controllers with hotplug='off', and attempt to assign
the device to a controller that not only supports hotplug, but also
has it enabled.
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
The HOTPLUGGABLE flag is set for appropriates buses in a PCI address
set, and thnis patch updates virDomainPCIAddressFlagsCompatible() to
check the HOTPLUGGABLE flag when searching for a suitable bus/slot for
a device. No devices request HOTPLUGGABLE though (yet), so there is no
observable effect.
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
virDomainPCIAddressBusSetModel() is called for each PCI controller
when building an address set prior to assiging PCI addresses to
devices.
This patch adds a new argument, allowHotplug, to that function that
can be set to false if we know for certain that a particular
controller won't support hotplug
The most interesting case is in qemuDomainPCIAddressSetCreate(), where
the config of each existing controller is available while building the
address set, so we can appropriately set allowHotplug = false when the
user has "hotplug='off'" in the config of a controller that normally
would support hotplug. In all other cases, it is set to true or false
in accordance with the capability of the controller model.
So far we aren't doing anything with this bus flag in the address set.
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
Old behavior: If the address was manually provided by config, copy
device AUTOASSIGN flag into the bus flag, and then later on in the
function *always* check for a match of the flags (which will always
match if the address came from config, since we just copied it).
New behavior: Don't mess with the bus flags - just directly check if
the AUTOASSIGN flag matches in bus and dev, but only make the check if
the address didn't come from config (i.e. it was auto-assigned by
libvirt).
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
When the HOTPLUGGABLE flag was originally added, it was set for all
the PCI controllers that accepted hotplugged devices, and requested
for all devices that were auto-assigned to a controller. While we're
still autoassigning to the same list of controllers, those controllers
may or may not support hotplug, so let's use the flag that fits what
we're actually doing.
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
This new flag will be set for any controller that we decide can have
devices assigned to it automatically during PCI device assignment. In
the past PCI_CONNECT_TYPE_HOTPLUGGABLE was used for this purpose, but
that is overloading that flag, and no longer technically correct; what
we *really* want is to auto-assign devices to any pcie-root-port or
pcie-switch-downstream-port regardless of whether or not that
controller happens to have hotplug enabled.
This patch just adds the flag, but doesn't use it at all. Note that
the numbering of all the other flags was changed in order to insert
the new flag near the beginning of the list; that doesn't cause any
problem because the connect flags aren't stored anywhere between runs
of libvirtd.
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
If a pcie-root-port or pcie-downstream-port has hotplug='off' in its
<target> subelement, and if the qemu binary supports the hotplug=false
option, then it will be added to the commandline for the pcie
controller. This controller will then not allow any hotplug/unplug of
devices while the guest is running (and the hotplug capability won't
be advertised to the guest OS, so the guest OS also won't present
unplugging of PCI devices as an option).
<controller type='pci' model='pcie-root-port'>
<target hotplug='off'/>
</controller>
For any PCI controllers other than pcie-downstream-port and
pcie-root-port, of for qemu binaries that don't support the hotplug
commandline option, an error will be logged during validation.
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
a <controller type='pci'...> element can now have a "hotplug"
attribute in the <target> subelement. This is intended to control
whether or not the slot(s) of the controller support
hotplugging/unplugging a device:
<controller type='pci' model='pcie-root-port'>
<target hotplug='off'/>
</controller>
The default value of hotplug is "on".
Since support for configuring such an option is hypervisor-dependent
(and will vary among different types of PCI controllers even on a
single hypervisor), no validation is done in this patch - that
validation will be done in the patch that wires support for the
setting into the hypervisor.
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
This caps flag is set when the qemu binary supports the option
"hotplug" for pcie-root-port, ioh3420 (Intel pcie-root-port) and
xio3130-downstream (Intel pcie-downstream-port). If it's available,
it's possible to disable hotplugging/unplugging devices on a
particular port by adding ",hotplug=off" to the qemu device
commandline. This option first appears in qemu-5.0.0.
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
During the hypervisor-agnostic validation of network devices, verify
that the interface type is either "network" or "bridge", and that if
there is any <virtualport>, that it doesn't have any type associated
with it.
This needs to be done both for the parse-time validation and for
runtime validation (after a port has been acquired from any associated
network), because an interface with type='network' could have an
actual type at runtime of "hostdev" or "direct", neither of which
support isolated='true' (yet). Likewise, if an interface is
type='network', then at runtime a <virtualport> with a type that
doesn't support isolated='yes' (e.g. "openvswitch", "802.1Qbh" -
currently *none* of the available virtualport types support it)
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
This setting can be updating very easily on an already active
interface by just changing it in sysfs. If the bridge used for
connection is also changed, there is no need to separately update it,
because the new setting isf done as a part of connecting to the bridge
anyway.
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
This patch pushes the isolatedPort setting from the <interface> down
all the way to the callers of virNetDevBridgeAddPort(), and sets
BR_ISOLATED on the port (using virNetDevBridgePortSetIsolated()) after
the port has been successfully added to the bridge.
Signed-off-by: Laine Stump <laine@redhat.com>
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
Similar to the way that the <vlan>, <bandwidth>, and <virtualport>
elements and the trustGuestRxFilters attribute in a <network> (or in
the appropriate <portgroup> element of a <network> can be applied to a
port when it is allocated for a domain's network interface, this patch
checks for a configured value of <port isolated="yes|no"/> in
either the domain <interface> or in the network, setting isolatedPort
in the <networkport> to the first one it finds (the setting from the
domain's <interface> is preferred). This, in turn, is passed back to
the domain when a port is allocated, so that the domain will use that
setting.
(One difference from <vlan>, <bandwidth>, <virtualport>, and
trustGuestRxFilters, is that all of those can be set in a <portgroup>
so that they can be applied only to a subset of interfaces connected
to the network. This didn't really make sense for the isolated setting
due to the way that it's implemented in Linux - the BR_ISOLATED flag
will prevent traffic from passing between two ports that both have
BR_ISOLATED set, but traffic can still go between those ports and
other ports that *don't* have BR_ISOLATED. (It would be nice if all
traffic from a BR_ISOLATED port could be blocked except traffic going
to/from a designated egress port or ports, but instead the entire
feature is implemented as a single flag. Because of this, it's really
only useful if all the ports on a network are isolated, so setting it
for a subset has no practical utility.)
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
This is a very simple thing to parse and format, but needs to be done
in 4 places, so two trivial utility functions have been made that can
be called from all the higher level parser/formatters:
<domain><interface>
<domain><interface><actual> (only in domain status)
<network>
<networkport>
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
When this flag is set for an interface attached to a bridge, traffic
to/from the specified interface can only enter/exit the bridge via
another attached interface that *doesn't* have the BR_ISOLATED flag
set. This can be used to permit guests to communicate with the rest of
the network, but not with each other.
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
Not only was the original error code destroyed in the case of
encountering an error during recovery from a failed attach to the
bridge (and then *that* error was destroyed by logging a *second*
error about the failure to recover - virNetDevBridgeAddPort() already
logs an error, so the one about failing to recover was redundant), but
if the recovery was successful, the function would then return success
to the caller even though it had failed.
Fixes: 2711ac8716
(overwritten errors were introduced along with this functionality)
Fixes: 6bde0a1a37
(the wrong return value was introduced by a refactor)
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
This is in the data structure and the parse/format functions, and is
getting passed all around correctly, it just was omitted from the RNG,
which hasn't been noticed because no human is creating <networkport>
XML, and so it's never getting validated against the schema.
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
Aside from itinerant error (actually warning) messages due to an
unrecognized response from qemu, this isn't even necessary - the
migration proceeds successfully to completion anyway.
(I'm not sure where to see this status reported in the API though - do
we need to add an extra state, or recognition of a new event somewhere?)
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Normally a PCI hostdev can't be migrated, so
qemuMigrationSrcIsAllowedHostdev() won't permit it. In the case of a a
hostdev network interface that has <teaming type='transient'/> set,
QEMU will automatically unplug the device prior to migration, and
re-plug a corresponding device on the destination. This patch modifies
qemuMigrationSrcIsAllowedHostdev() to allow domains with those devices
to be migrated.
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
The QEMU driver uses the <teaming type='persistent|transient'
persistent='blah'/> element to setup a "failover" pair of devices -
the persistent device must be a virtio emulated NIC, with the only
extra configuration being the addition of ",failover=on" to the device
commandline, and the transient device must be a hostdev NIC
(<interface type='hostdev'> or <interface type='network'> with a
network that is a pool of SRIOV VFs) where the extra configuration is
the addition of ",failover_pair_id=$aliasOfVirtio" to the device
commandline. These new options are supported in QEMU 4.2.0 and later.
Extra qemu-specific validation is added to ensure that the device
type/model is appropriate and that the qemu binary supports these
commandline options.
The result of this will be:
1) The virtio device presented to the guest will have an extra bit set
in its PCI capabilities indicating that it can be used as a failover
backup device. The virtio guest driver will need to be equipped to do
something with this information - this is included in the Linux
virtio-net driver in kernel 4.18 and above (and also backported to
some older distro kernels). Unfortunately there is no way for libvirt
to learn whether or not the guest driver supports failover - if it
doesn't then the extra PCI capability will be ignored and the guest OS
will just see two independent devices. (NB: the current virtio guest
driver also requires that the MAC addresses of the two NICs match in
order to pair them into a bond).
2) When a migration is requested, QEMu will automatically unplug the
transient/hostdev NIC from the guest on the source host before
starting migration, and automatically re-plug a similar device after
restarting the guest CPUs on the destination host. While the transient
NIC is unplugged, all network traffic will go through the
persistent/virtio device, but when the hostdev NIC is plugged in, it
will get all the traffic. This means that in normal circumstances the
guest gets the performance advantage of vfio-assigned "real hardware"
networking, but it can still be migrated with the only downside being
a performance penalty (due to using an emulated NIC) during the
migration.
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
The subelement <teaming> of <interface> devices is used to configure a
simple teaming association between two interfaces in a domain. Example:
<interface type='bridge'>
<source bridge='br0'/>
<model type='virtio'/>
<mac address='00:11:22:33:44:55'/>
<alias name='ua-backup0'/>
<teaming type='persistent'/>
</interface>
<interface type='hostdev'>
<source>
<address type='pci' bus='0x02' slot='0x10' function='0x4'/>
</source>
<mac address='00:11:22:33:44:55'/>
<teaming type='transient' persistent='ua-backup0'/>
</interface>
The interface with <teaming type='persistent'/> is assumed to always
be present, while the interface with type='transient' may be be
unplugged and later re-plugged; the persistent='blah' attribute (and
in the one currently available implementation, also the matching MAC
addresses) is what associates the two devices with each other. It is
up to the hypervisor and the guest network drivers to determine what
to do with this information.
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Presence of the virtio-net-pci option called "failover" indicates
support in a qemu binary of a simplistic bonding of a virtio-net
device with another PCI device. This feature allows migration of
guests that have a network device assigned to a guest with VFIO, by
creating a network bond device in the guest consisting of the
VFIO-assigned device and a virtio-net-pci device, then temporarily
(and automatically) unplugging the VFIO net device prior to migration
(and hotplugging an equivalent device on the migration
destination). (The feature is called "failover" because the bond
device uses the vfio-pci netdev for normal guest networking, but
"fails over" to the virtio-net-pci netdev once the vfio-pci device is
unplugged for migration.)
Full functioning of the feature also requires support in the
virtio-net driver in the guest OS (since that is where the bond device
resides), but if the "failover" commandline option is present for the
virtio-net-pci device in qemu, at least the qemu part of the feature
is available, and libvirt can add the proper options to both the
virtio-net-pci and vfio-pci device commandlines to indicate qemu
should attempt doing the failover during migration.
This patch just adds the qemu capabilities flag "virtio-net.failover".
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Prior to commit 55ce656463 (first in libvirt 4.6.0), the XML sent to
virDomainAttachDeviceFlags() was parsed only once, and the results of
that parse were inserted into both the live object of the running
domain and into the persistent config. Thus, if MAC address was
omitted from in XML for a network device (<interface>), both the live
and config object would have the same MAC address.
Commit 55ce656463 changed the code to parse the incoming XML twice -
once for live and once for config. This does eliminate the problem of
PCI (/scsi/sata) address conflicts caused by allocating an address
based on existing devices in live object, but then inserting the
result into the config (which may already have a device using that
address), BUT it also means that when the MAC address of a network
device hasn't been specified in the XML, each copy will get a
different auto-generated MAC address.
This results in the MAC address of the device changing the next time
the domain is shutdown and restarted, which creates havoc with the
guest OS's network config.
There have been several discussions about this in the last > 1 year,
attempting to find the ideal solution to this problem that makes MAC
addresses consistent and accounts for all sorts of corner cases with
PCI/scsi/sata addresses. All of these discussions fizzled out because
every proposal was either too difficult to implement or failed to fix
some esoteric case someone thought up.
So, in the interest of solving the MAC address problem while not
making the "other address" situation any worse than before, this patch
simply adds a qemuDomainAttachDeviceLiveAndConfigHomogenize() function
that (for now) copies the MAC address from the config object to the
live object (if the original xml had <mac address='blah'/> then this
will be an effective NOP (as the macs already match)).
Any downstream libvirt containing upstream commit
55ce656463 should have this patch as well.
https://bugzilla.redhat.com/1783411
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
In the past the network driver was (mistakenly) being called for all
interfaces, not just those of type='network', and so it had a chance
to validate all interface configs after the actual type of the
interface was known.
But since the network driver has been more completely/properly
separated from qemu, the network driver isn't called during the
startup of any interfaces except those with type='network', so this
validation no longer takes place for, e.g. <interface type='bridge'>
(or direct, etc). This in turn meant that a config could erroneously
specify a vlan tag, or bandwidth settings, for a type of interface
that didn't support it, and the domain would start without complaint,
just silently ignoring those settings.
This patch moves those validation checks out of the network driver,
and into virDomainActualNetDefValidate() so they will be done for all
interfaces, not just type='network'.
https://bugzilla.redhat.com/1741121
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Cole Robinson <crobinso@redhat.com>
<interface> devices (virDomainNetDef) are a bit different from other
types of devices in that their actual type may come from a network (in
the form of a port connection), and that doesn't happen until the
domain is started. This means that any validation of an <interface> at
parse time needs to be a bit liberal in what it accepts - when
type='network', you could think that something is/isn't allowed, but
once the domain is started and a port is created by the configured
network, the opposite might be true.
To solve this problem hypervisor drivers need to do an extra
validation step when the domain is being started. I recently (commit
3cff23f7, libvirt 5.7.0) added a function to peform such validation
for all interfaces to the QEMU driver -
qemuDomainValidateActualNetDef() - but while that function is a good
single point to call for the multiple places that need to "start" an
interface (domain startup, device hotplug, device update), it can't be
called by the other hypervisor drivers, since 1) it's in the QEMU
driver, and 2) it contains some checks specific to QEMU. For
validation that applies to network devices on *all* hypervisors, we
need yet another interface validation function that can be called by
any hypervisor driver (not just QEMU) right after its network port has
been created during domain startup or hotplug. This patch adds that
function - virDomainActualNetDefValidate(), in the conf directory,
and calls it in appropriate places in the QEMU, lxc, and libxl
drivers.
This new function is the place to put all network device validation
that 1) is hypervisor agnostic, and 2) can't be done until we know the
"actual type" of an interface.
There is no framework for validation at domain startup as there is for
post-parse validation, but I don't want to create a whole elaborate
system that will only be used by one type of device. For that reason,
I just made a single function that should be called directly from the
hypervisors, when they are initializing interfaces to start a domain,
right after conditionally allocating the network port (and regardless
of whether or not that was actually needed). In the case of the QEMU
driver, qemuDomainValidateActualNetDef() is already called in all the
appropriate places, so we can just call the new function from
there. In the case of the other hypervisors, we search for
virDomainNetAllocateActualDevice() (which is the hypervisor-agnostic
function that calls virNetworkPortCreateXML()), and add the call to our
new function right after that.
The new function itself could be plunked down into many places in the
code, but we already have 3 validation functions for network devices
in 2 different places (not counting any basic validation done in
virDomainNetDefParseXML() itself):
1) post-parse hypervisor-agnostic
(virDomainNetDefValidate() - domain_conf.c:6145)
2) post-parse hypervisor-specific
(qemuDomainDeviceDefValidateNetwork() - qemu_domain.c:5498)
3) domain-start hypervisor-specific
(qemuDomainValidateActualNetDef() - qemu_domain.c:5390)
I placed (3) right next to (2) when I added it, specifically to avoid
spreading validation all over the code. For the same reason, I decided
to put this new function right next to (1) - this way if someone needs
to add validation specific to qemu, they go to one location, and if
they need to add validation applying to everyone, they go to the
other. It looks a bit strange to have a public function in between a
bunch of statics, but I think it's better than the alternative of
further fragmentation. (I'm open to other ideas though, of course.)
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Cole Robinson <crobinso@redhat.com>
These all just return a scalar value, so there's no daisy-chained
fallout from changing them, and they can easily be combined in a
single patch.
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Cole Robinson <crobinso@redhat.com>
This also isn't required (due to the vportprofile being stored in the
NetDef as a pointer rather than being directly contained), but it
seemed dishonest to not mark it as const (and thus permit users to
modify its contents)
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Cole Robinson <crobinso@redhat.com>
In this case, the virNetDevBandwidthPtr that is returned is not to a
region within the virDomainNetDef arg, but points elsewhere (the
NetDef has the pointer, not the entire object), so technically it's
not necessary to make the return value a const, but it's a bit
disingenuous to *not* do it.
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Cole Robinson <crobinso@redhat.com>
This is needed if we want to call the function when the
virDomainNetDef* we have is a const.
Since virDomainNetGetActualVlan returns a pointer to memory that is
within the virDomainNetDefPtr arg, the returned pointer must also be
made const. This leads to a cascade of other virNetDevVlanPtr's that
must be changed to "const virNetDevVlan *".
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Cole Robinson <crobinso@redhat.com>
This makes it easier to understand which interface's config caused the
error.
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Cole Robinson <crobinso@redhat.com>
When libvirt first implemented a stable and configurable MAC address
for the bridges created for libvirt virtual networks (commit
5754dbd56d, in libvirt v0.8.8) most distro stable releases didn't
support explicitly setting the MAC address of a bridge; the bridge
just always assumed the lowest numbered MAC of all attached
interfaces. Because of this, we stabilized the bridge MAC address by
creating a "dummy" tap interface with a MAC address guaranteed to be
lower than any of the guest tap devices' MACs (which all started with
0xFE, so it's not difficult to do) and attached it to the bridge -
this was the inception of the "virbr0-nic" device that has confused so
many people over the years.
Even though the linux kernel had recently gained support for
explicitly setting a bridge MAC, we deemed it unnecessary to set the
MAC that way, because the other (indirect) method worked everywhere.
But recently there have been reports that the bridge MAC address was
not following the setting in the network config, and mismatched the
MAC of the dummy tap device (which was still correct). It turns out
that this is due to a change in systemd-242 that persists whatever MAC
address is set for a bridge when it's initially started. According to
the systemd NEWS file entry for version 242
(https://github.com/systemd/systemd/blob/master/NEWS):
"if a bridge interface is created without any slaves, and gains
a slave later, then now the bridge does not inherit slave's MAC."
This change was the result of:
https://github.com/systemd/systemd/issues/3374
(apparently if there is no MAC saved for a bridge by the name of a
bridge being created, the random MAC generated during creation is
saved, and then that same MAC is used to explicitly set the MAC each
time it is created). Once a bridge has an explicitly set MAC, the "use
the lowest numbered MAC of attached devices" rule is ignored, so our
dummy tap device is like the goggles - it does nothing! (well, almost).
We could whine about changes in default behavior, etc. etc., but
because the change was in response to actual user problems, that seems
likely a fruitless task. Fortunately, time has marched on, and even
distro releases that are old enough that they are no longer supported
by upstream libvirt (e.g. RHEL6) have support for explicitly setting a
bridge device MAC address, either during creation or with a separate
ioctl after creation, so we can now do that.
To enable explicitly setting the mac during bridge creation, we add a
mac arg to virNetDevBridgeCreate(). In the case of platforms where
the bridge is created with a netlink RTM_NEWLINK message, we just add
that mac to the message. For platforms that still use an ioctl (either
SIOCBRADDBR or SIOCIFCREATE2), we make a separate call to
virNetDevSetMAC() after creating the bridge.
(NB: I was unable to test the calling of virNetDevSetMAC() from the
SIOCIFCREATE2 (BSD) version of virNetDevBridgeCreate(); even though I
managed to get a FreeBSD system setup and libvirt built there, when I
tried to start the default network the SIOCIFCREATE2 ioctl itself
failed, so it never even got to the virNetDevSetMAC(). That leaves the
FreeBSD implementation untested.)
This makes the dummy tap pointless for purposes of setting the MAC
address, but it is still useful for IPv6 DAD initialization (which
apparently requires at least one interface to be attached to the
bridge and online), as well as for setting an initial MTU for the
bridge, so it hasn't been removed.
(NB: we can safely *always* call virNetDevBridgeCreate() with
&def->mac from the network driver because, in spite of the existence
of a "mac_specified" bool in the config suggesting that it may not
always be present, in reality a mac address will always be added to
any network that doesn't have one - this is guaranteed in all cases by
commit a47ae7c004)
https://bugzilla.redhat.com/show_bug.cgi?id=1760851
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Jiri Denemark <jdenemar@redhat.com>
Although until now, any use of the extra_args argument (a pointer to a
struct containing extra attributes to add the the RTM_NEWLINK message)
would always have the ifindex and mac set, so the code could assume it
was safe to add both to the message if extra_args != NULL. There is
now a use for setting a MAC address in the RTM_NEWLINK without setting
the ifindex, so we should check each of these separately.
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Jiri Denemark <jdenemar@redhat.com>
Commit 01ca4010d8 (libvirt v5.1.0) moved address reservation for
hotplugged interface devices up to an earlier point in
qemuDomainAttachNetDevice(), because that function calls
qemuDomainSupportsNicdev() (in the case of
VIR_DOMAIN_NET_TYPE_VHOSTUSER), and qemuDomainSupportsNicdev() needs
to know the address type (for ARM machinetypes) and returns incorrect
results when the address type is "none".
This bugfix unfortunately caused a regression, because it also made PCI
address reservation happen before we noticed that the device was a
*hostdev* interface. Those interfaces are hotplugged by just calling
out to qemuDomainAttachHostdevDevice() - that function would then also
attempt to reserve the *same PCI address* that had just been reserved
in qemuDomainAttachNetDevice().
The solution is to move the bit of code that short-circuits out to
virDomainHostdevAttach() up *even earlier* so that no PCI address has
been allocated by the time it's called.
https://bugzilla.redhat.com/show_bug.cgi?id=1744523
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Andrea Bolognani <abologna@redhat.com>
Back in July 2009, in the days before libvirt supported explicitly
assigning a PCI address to every device, code was added to save the
PCI addresses of hotplugged network, disk, and hostdevs in the domain
status with this XML element:
<state devaddr='domain:bus:slot'/>
This was added in commits 4e21a95a, 01654107, in v0.7.0, and 0c5b7b93
in v0.7.1.
Then just a few months later, in November 2009, The code that actually
formatted the "devaddr='blah'" into the status XML was removed by
commit 1b0cce7d3 (which "introduced a standardized data structure for
device addresses"). The code to *parse* the devaddr from the status
was left in for backward compatibility though (it just parses it into
the "standard" PCI address).
At the time the devaddr attribute was added, a few other attributes
already existed in the <state> element for network devices, and these
were removed over time (I haven't checked the exact dates of this),
but 10 years later, in libvirt v5.8.0, we *still* maintain code to
parse <state devaddr='blah'/> from the domain status.
In the meantime, even distros so old that we no longer support them in
upstream libvirt are using a libvirt new enough that it doesn't ever
write <state devaddr='blah'/> to the domain status XML.
Since the only way a current libvirt would ever encounter this element
would be if someone was upgrading directly from libvirt <= v0.7.5 with
running guests, it seems safe to finally remove the code that parses it.
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
Before the refactoring that properly separated the network driver from
the hypervisor driver and forced all interaction to go through public
APIs, all network usage counters were zeroed when the network driver
was initialized, and the network driver's now-deprecated
"semi-private" API networkNotifyActualDevice() was called for every
interface of every domain as each hypervisor "reconnected" its domains
during a libvirtd restart, and this would refresh the usage count for
each network.
Post-driver-split, during libvirtd restart/reconnection of the running
domains, the function virDomainNetNotifyActualDevice() is called by
each hypervisor driver for every interface of every domain restart,
and this function has code to re-register interfaces, but it only
calls into the network driver to re-register those ports that don't
already have a valid portid (ie. one that is not simply all 0),
assuming that those with valid portids are already known (and counted)
by the network driver.
commit 7ab9bdd47 recently modified the network driver so that, in most
cases, it properly resyncs each network's connection count during
libvirtd (or maybe virtnetworkd) restart by iterating through the
network's port list. This doesn't account for the case where a network
is destroyed and restarted while there are running domains that have
active ports on the network. In that case, the entire port list and
connection count for that network is lost, and now even a restart of
libvirtd/virtnetworkd/virtqemud, which in the past would resync the
connection count, doesn't help (the network driver thinks there are no
active ports, while the hypervisor driver knows about all the active
ports, but mistakenly believes that the network driver also knows).
The solution to this is to not just bypass valid portids during the
call to virDomainNetworkNotifyActualDevice(). Instead, we query the
network driver about the portid that was preserved in the domain
status, and if it is not registered, we register it.
(NB: while it would technically be correct to just generate a new
portid for these cases, it makes for less churn in portids (and thus
may make troubleshooting simpler) if we make the small fix to
virDomainNetDefActualToNetworkPort() that preserves existing valid
portids rather than unconditionally generating a new one.)
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
define a VIR_DEFINE_AUTOPTR_FUNC() to autofree virNetworkPortDefs, and
convert all uses of virNetworkPortDefPtr that are appropriate to use
it.
This coincidentally fixes multiple potential memory leaks (in failure
cases) in networkPortCreateXML()
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
A virDomainNetDef object in a domain's nets array might contain a
virDomainHostdevDef, and when this is the case, the domain's hostdevs
array will also have a pointer to this embedded hostdev (this is done
so that internal functions that need to perform some operation on all
hostdevs won't leave out the type='hostdev' network interfaces).
When a network device was updated with virDomainUpdateDeviceFlags(),
we were replacing the entry in the nets array (and free'ing the
original) but forgetting about the pointer in the hostdevs array
(which would then point to the now-free'd hostdev contained in the old
net object.) This often resulted in a libvirtd crash.
The solution is to add a function, virDomainNetUpdate(), called by
qemuDomainUpdateDeviceConfig(), that updates the hostdevs array
appropriately along with the nets array.
Resolves: https://bugzilla.redhat.com/1558934
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
When the bridge re-attach handling was moved out of the network driver
and into the hypervisor driver (commit b806a60e) as a part of the
refactor to split the network driver into a separate daemon, the check
was accidentally changed to only check for type='bridge'. The check for
type in this case needs to check for type='network' as well.
(at the time we thought that the two types could be conflated for
interface actual type, but this turned out to be too problematic to
do).
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
qemuDomainAttachNetDevice() (hotplug) previously had some of the
validation that is in qemuDomainValidateActualNetDef(), but it was
incomplete. qemuDomainChangeNet() had none of that validation, but it
is all appropriate in both cases.
This is the final piece of a previously partial resolution to
https://bugzilla.redhat.com/1502754
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
The same validation should be done for both static network devices and
hotplugged devices, but they are currently inconsistent. Move all the
relevant validation from qemuBuildInterfaceCommandLine() into the new
function qemuDomainValidateActualNetDef() and call the latter from
the former.
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
It needs to be used by a function that only has a const pointer to
virDomainNetDef.
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
Traditionally, macvtap devices are supported using <interface
type='direct'>, but that type requires specifying a source device name
and macvtap mode which can't be altered after the initial device
creation (and may not even be available to the management software
that's creating the XML config to feed to libvirt).
But the attributes in the <source> are essentially describing how the
device will be connected to the network, and if libvirt is to be
supplied with the name of a macvtap device that has already been
created, that device will also already be connected to the network
(and the connection can't be changed). Thus it seems more appropriate
to use type='ethernet', which was created explicitly for this purpose
- for devices that have already been (or will be) connected to the
external network by someone/something outside of libvirt. The fact
that it is a *macv*tap rather than a contentional tap device is just a
detail.
This patch supports using an existing macvtap device with <interface
type='ethernet'> by checking the supplied target dev name to see if it
is a macvtap device and, when this is the case, calling
virNetDevMacVLanTapOpen() instead of virNetDevTapCreate(). For
consistency, this is only done when target managed='no'.
Resolves: https://bugzilla.redhat.com/1723367 (partially)
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
If managed='no', then the tap device must already exist, and setting
of MAC address and online status (IFF_UP) is skipped.
NB: we still set IFF_VNET_HDR and IFF_MULTI_QUEUE as appropriate,
because those bits must be properly set in the TUNSETIFF we use to set
the tap device name of the handle we've opened - if IFF_VNET_HDR has
not been set and we set it the request will be honored even when
running libvirtd unprivileged; if IFF_MULTI_QUEUE is requested to be
different than how it was created, that will result in an error from
the kernel. This means that you don't need to pay attention to
IFF_VNET_HDR when creating the tap devices, but you *do* need to set
IFF_MULTI_QUEUE if you're going to use multiple queues for your tap
device.
NB2: /dev/vhost-net normally has permissions 600, so it can't be
opened by an unprivileged process. This would normally cause a warning
message when using a virtio net device from an unprivileged
libvirtd. I've found that setting the permissions for /dev/vhost-net
permits unprivileged libvirtd to use vhost-net for virtio devices, but
have no idea what sort of security implications that has. I haven't
changed libvrit's code to avoid *attempting* to open /dev/vhost-net -
if you are concerned about the security of opening up permissions of
/dev/vhost-net (probably a good idea at least until we ask someone who
knows about the code) then add <driver name='qemu'/> to the interface
definition and you'll avoid the warning message.
Note that virNetDevTapCreate() is the correct function to call in the
case of an existing device, because the same ioctl() that creates a
new tap device will also open an existing tap device.
Resolves: https://bugzilla.redhat.com/1723367 (partially)
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Although <interface type='ethernet'> has always been able to use an
existing tap device, this is just a coincidence due to the fact that
the same ioctl is used to create a new tap device or get a handle to
an existing device.
Even then, once we have the handle to the device, we still insist on
doing extra setup to it (setting the MAC address and IFF_UP). That
*might* be okay if libvirtd is running as a privileged process, but if
libvirtd is running as an unprivileged user, those attempted
modifications to the tap device will fail (yes, even if the tap is set
to be owned by the user running libvirtd). We could avoid this if we
knew that the device already existed, but as stated above, an existing
device and new device are both accessed in the same manner, and
anyway, we need to preserve existing behavior for those who are
already using pre-existing devices with privileged libvirtd (and
allowing/expecting libvirt to configure the pre-existing device).
In order to cleanly support the idea of using a pre-existing and
pre-configured tap device, this patch introduces a new optional
attribute "managed" for the interface <target> element. This
attribute is only valid for <interface type='ethernet'> (since all
other interface types have mandatory config that doesn't apply in the
case where we expect the tap device to be setup before we
get it). The syntax would look something like this:
<interface type='ethernet'>
<target dev='mytap0' managed='no'/>
...
</interface>
This patch just adds managed to the grammar and parser for <target>,
but has no functionality behind it.
(NB: when managed='no' (the default when not specified is 'yes'), the
target dev is always a name explicitly provided, so we don't
auto-remove it from the config just because it starts with "vnet"
(VIR_NET_GENERATED_TAP_PREFIX); this makes it possible to use the
same pattern of names that libvirt itself uses when it automatically
creates the tap devices.)
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
This will simplify addition of another attribute to the <target> element
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
This just moves around a few things in qemuInterfaceConnect() with no
functional difference (except that a few failures that would have
previously resulted in a "success" audit log will now properly produce
a "fail" audit). The change is so that adding support for unmanaged
tap/macvtap devices will be more easily reviewable.
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
In virNetDevMacVLanOpen(), The "retries" arg has been removed and the
value hardcoded as 10, since previously the function was only called
from one place, so it was always 10.
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
This function returns T if the given name is a macvtap device. This is
determined by 1) getting the ifindex of the device with that name (if
there is one), and 2) checking for existence of /dev/tapXX, where "XX"
is the ifindex learned in (1).
It's also possible to learn this by getting a netlink dump of the
interface and parsing through it to look for some attributes, but that
is complicated to figure out, takes longer to execute, and I'm lazy.
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
virErrorPreserveLast()/virErrorRestore() (added in commit 8333e7455
back in 2017), do a better better job of saving and restoring the last
libvirt error than virSaveLastError()/virErrorRestore() (they're
simpler, and they also save/restore the system errno).
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
During networkPortCreateXML, if networkAllocatePort() failed,
networkReleasePort() would be called, which would (in the case of
network pools of macvtap passthrough devices) attempt to find the
allocated device by comparing port->plug.direct.linkdev to each device
in the pool. Since port->plug.direct.linkdev was still NULL, the
attempted strcmp would result in a SEGV.
Calling networkReleasePort() during error cleanup is something that
should only be done if networkAllocatePort() has already succeeded. It
turns out there is one other possible error exit from
networkPortCreateXML() that happens after networkAllocatePort() has
succeeded, so the code to call networkReleasePort() was just moved
down to there.
Resolves: https://bugzilla.redhat.com/1741390
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
Commit e69444e17 (first appeared in libvirt-5.5.0) added the new value
"VIR_ACCESS_PERM_NETWORK_SEARCH_PORTS" to the virAccessPerNetwork
enum, and also the string "search_ports" to the VIR_ENUM_IMPL() macro
for that enum. Unfortunately, the enum value was added in the middle
of the list, while the string was added to the end of the
VIR_ENUM_IMPL().
This patch corrects that error by moving the new value to the end of
the enum definition, so that the order matches that of the string
list.
Resolves: https://bugzilla.redhat.com/1741428
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
Back in July 2010, commit 6ea90b84 (meant to resolve
https://bugzilla.redhat.com/571991 ) added code to set the MAC address
of any tap device to the associated guest interface's MAC, but with
the first byte replaced with 0xFE. This was done in order to assure
that
1) the tap MAC and guest interface MAC were different (otherwise L2
forwarding through the tap would not work, and the kernel would
repeatedly issue a warning stating as much).
2) any bridge device that had one of these taps attached would *not*
take on the MAC of the tap (leading to network instability as
guests started and stopped)
A couple years later, https://bugzilla.redhat.com/798467 was filed,
complaining that a user could configure a tap-based guest interface to
have a MAC address that itself had a first byte of 0xFE, silently
(other than the kernel warning messages) resulting in a non-working
configuration. This was fixed by commit 5d571045, which logged an
error and failed the guest start / interface attach if the MAC's first
byte was 0xFE.
Although this restriction only reduces the potential pool of MAC
addresses from 2^46 (last two bits of byte 1 must be set to 10) by
2^32 (still 4 orders of magnitude larger than the entire IPv4 address
space), it also means that management software that autogenerates MAC
addresses must have special code to avoid an 0xFE prefix. Now after 7
years, someone has noticed this restriction and requested that we
remove it.
So instead of failing when 0xFE is found as the first byte, this patch
removes the restriction by just replacing the first byte in the tap
device MAC with 0xFA if the first byte in the guest interface is
0xFE. 0xFA is the next-highest value that still has 10 as the lowest
two bits, and still
2) meets the requirement of "tap MAC must be different from guest
interface MAC", and
3) is high enough that there should never be an issue of the attached
bridge device taking on the MAC of the tap.
The result is that *any* MAC can be chosen by management software
(although it would still not work correctly if a multicast MAC (lowest
bit of first byte set to 1) was chosen), but that's a different
issue).
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Michal Privoznik <mprivozn@redhat.com
The network driver used to reload the firewall rules whenever a dbus
NameOwnerChanged message for org.fedoraproject.FirewallD1 was
received. Presumably at some point in the past this was successful at
reloading our rules after a firewalld restart. Recently though I
noticed that once firewalld was restarted, libvirt's logs would get this
message:
The name org.fedoraproject.FirewallD1 was not provided by any .service files
After this point, no networks could be started until libvirtd itself
was restarted.
The problem is that the NameOwnerChanged message is sent twice during
a firewalld restart - once when the old firewalld is stopped, and
again when the new firewalld is started. If we try to reload at the
point the old firewalld is stopped, none of the firewalld dbus calls
will succeed.
The solution is to check the new_owner field of the message - we
should reload our firewall rules only if new_owner is non-empty (it is
set to "" when firewalld is stopped, and some sort of epoch number
when it is again started).
Signed-off-by: Laine Stump <laine@laine.org>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
When virDBusMessageRead() and virDBusMessageDecode were first added in
commit 834c9c94, they were identical except that virDBusMessageRead()
would unref the message after decoding it.
This difference was eliminated later in commit dc7f3ffc after it
became apparent that unref-ing the message so soon was never the right
thing to do. The two identical functions remained though, with the
tests and virDBus library itself calling the Decode variant, and all
other users calling the Read variant.
This patch eliminates the duplication, switching all users to
virDBusMessageDecode (and moving the nice API documentation comment
from the Read function up to the Decode function).
Signed-off-by: Laine Stump <laine@laine.org>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
The firewalld package in Fedora 30 didn't get support for rich rule
priorities, which is required by the libvirt zonefile that's installed
when the build is configured with --with-firewalld-zone, so we need to
set --without-firewalld-zone for that version of Fedora. The needed
feature is already upstream in firewalld, so it just needs another
upstream release to be there. Let's be optimistic and assume that will
happen prior to F31.
Resolves: https://bugzilla.redhat.com/1699051
Signed-off-by: Laine Stump <laine@laine.org>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Acked-by: Eric Garver <eric@garver.life>
commit edaf13565 modified the stats retrieval for OVS interfaces to
not fail when one of the fields was unrecognized by the ovs-vsctl
command, but ovs-vsctl was still returning an error, and libvirt was
cluttering the logs with these inconsequential error messages.
This patch modifies the GET_STAT macro to add "--if-exists" to the
ovs-vsctl command, which causes it to return an empty string (and exit
with success) if the requested statistic isn't in its database, thus
eliminating the ugly error messages from the log.
Resolves: https://bugzilla.redhat.com/1683175
Signed-off-by: Laine Stump <laine@laine.org>
For [some unknown reason, possibly/probably pure chance], Net devices
have been taken offline and their bandwidth tc rules cleared as the
very first operation when detaching the device. This is contrary to
every other type of device, where all hostside teardown is delayed
until we receive the DEVICE_DELETED event back from qemu, indicating
that the guest has finished with the device.
This patch delays these two operations until receipt of
DEVICE_DELETED, which removes an ugly wart from
qemuDomainDetachDeviceLive(), and also seems to be a more correct
sequence of events.
Signed-off-by: Laine Stump <laine@laine.org>
ACKed-by: Peter Krempa <pkrempa@redhat.com>
The VIR_DOMAIN_EVENT_ID_DEVICE_REMOVED event is sent after qemu has
responded to a device_del command with a DEVICE_DELETED event. Before
queuing the event, *some* of the final teardown of the device's
trappings in libvirt is done, but not *all* of it. As a result, an
application may receive and process the DEVICE_REMOVED event before
libvirt has really finished with it.
Usually this doesn't cause a problem, but it can - in the case of the
bug report referenced below, vdsm is assigning a PCI device to a guest
with managed='no', using livirt's virNodeDeviceDetachFlags() and
virNodeDeviceReAttach() APIs. Immediately after receiving a
DEVICE_REMOVED event from libvirt signalling that the device had been
successfully unplugged, vdsm would cal virNodeDeviceReAttach() to
unbind the device from vfio-pci and rebind it to the host driverm but
because the event was received before libvirt had completely finished
processing the removal, that device was still on the "activeDevs"
list, and so virNodeDeviceReAttach() failed.
Experimentation with additional debug logs proved that libvirt would
always end up dispatching the DEVICE_REMOVED event before it had
removed the device from activeDevs (with a *much* greater difference
with managed='yes', since in that case the re-binding of the device
occurred after queuing the device).
Although the case of hostdev devices is the most extreme (since there
is so much involved in tearing down the device), *all* device types
suffer from the same problem - the DEVICE_REMOVED event is queued very
early in the qemuDomainRemove*Device() function for all of them,
resulting in a possibility of any application receiving the event
before libvirt has really finished with the device.
The solution is to save the device's alias (which is the only piece of
info from the device object that is needed for the event) at the
beginning of processing the device removal, and then queue the event
as a final act before returning. Since all of the
qemuDomainRemove*Device() functions (except
qemuDomainRemoveChrDevice()) are now called exclusively from
qemuDomainRemoveDevice() (which selects which of the subordinates to
call in a switch statement based on the type of device), the shortest
route to a solution is to doing the saving of alias, and later
queueing of the event, in the higher level qemuDomainRemoveDevice(),
and just completely remove the event-related code from all the
subordinate functions.
The single exception to this, as mentioned before, is
qemuDomainRemoveChrDevice(), which is still called from somewhere
other than qemuDomainRemoveDevice() (and has a separate arg used to
trigger different behavior when the chr device has targetType ==
GUESTFWD), so it must keep its original behavior intact, and must be
treated differently by qemuDomainRemoveDevice() (similar to the way
that qemuDomainDetachDeviceLive() treats chr and lease devices
differently from all the others).
Resolves: https://bugzilla.redhat.com/1658198
Signed-off-by: Laine Stump <laine@laine.org>
ACKed-by: Peter Krempa <pkrempa@redhat.com>
Now that all the qemuDomainDetachPrep*() functions look nearly
identical at the end, we can put one copy of that identical code in
qemuDomainDetachDeviceLive() at the point after the individual prep
functions have been called, and remove the duplicated code from all
the prep functions. The code to locate the target "detach" device
based on the "match" device remains, as do all device-type-specific
validations.
Unfortunately there are a few things going on at once in this patch,
which makes it a bit more difficult to follow than the others; it was
just impossible to do the changes in stages and still have a
buildable/testable tree at each step.
The other changes of note:
* The individual prep functions no longer need their driver or async
args, so those are removed, as are the local "ret" variables, since
in all cases the functions just directly return -1 or 0.
* Some of the prep functions were checking for a valid alias and/or
for attempts to detach a multifunction PCI device, but not all. In
fact, both checks are valid (or at least harmless) for *all* device
types, so they are removed from the prep functions, and done a
single time in the common function.
(any attempts to *create* an alias when there isn't one has been
removed, since that is doomed to failure anyway; the only way the
device wouldn't have an alias is if 1) the domain was created by
calling virsh qemu-attach to attach an existing qemu process to
libvirt, and 2) the qemu command that started said process used "old
style" arguments for creating devices that didn't have any device
ids. Even if we constructed a device id for one of these devices,
qemu wouldn't recognize it in the device_del command anyway, so we
may as well fail earlier with "device missing alias" rather than
failing later with "couldn't delete device net0".)
* Only one type of device has shutdown code that must not be called
until after *all* validation of the device is done (including
checking for multifunction PCI and valid alias, which is done in the
toplevel common code). For this reason, the Net function has been
split in two, with the 2nd half (qemuDomainDetachShutdownNet())
called from the common function, right before sending the delete
command to qemu.
Signed-off-by: Laine Stump <laine@laine.org>
ACKed-by: Peter Krempa <pkrempa@redhat.com>
Although all hotpluggable devices other than lease, controller,
watchdof, and vsock can be audited, and *are* audited when an unplug
is successful, only disk, net, and hostdev were actually being audited
on failure.
This patch corrects that omission.
Signed-off-by: Laine Stump <laine@laine.org>
ACKed-by: Peter Krempa <pkrempa@redhat.com>
This function can be called with a virDomainDevicePtr and whether or
not the removal was successful, and it will call the appropriate
virDomainAudit*() function with the appropriate args for whatever type
of device it's given (or do nothing, if that's appropriate). This
permits generalizing some code that currently has a separate copy for
each type of device.
NB: Although the function initially will be called only with
success=false, that has been made an argument so that in the future
(when the qemuDomainRemove*Device() functions have had their common
functionality consolidated into qemuDomainRemoveDevice()), this new
common code can call qemuDomainRemoveAuditDevice() for all types.
Signed-off-by: Laine Stump <laine@laine.org>
ACKed-by: Peter Krempa <pkrempa@redhat.com>
qemuDomainDetachDeviceChr and qemuDomainDetachDeviceLease are more
consistent with each other.
Signed-off-by: Laine Stump <laine@laine.org>
ACKed-by: Peter Krempa <pkrempa@redhat.com>
Most of these functions will soon contain only some setup for
detaching the device, not the detach code proper (since that code is
identical for these devices). Their device specific functions are all
being renamed to qemuDomainDetachPrep*(), where * is the
name of that device's data member in the virDomainDeviceDef
object.
Since there will be other code in qemuDomainDetachDeviceLive() after
the calls to qemuDomainDetachPrep*() that could still fail, we no
longer directly set "ret" with the return code from
qemuDomainDetachPrep*() functions, but simply return -1 on
failure, and wait until the end of qemuDomainDetachDeviceLive() to set
ret = 0.
Along with the rename, qemuDomainDetachPrep*() functions are also
given similar arglists, including an arg called "match" that points to
the proto-object of the device we want to delete, and another arg
"detach" that is used to return a pointer to the actual object that
will be (for now *has been*) detached. To make sure these new args
aren't confused with existing local pointers that sometimes had the
same name (detach), the local pointer to the device is now named after
the device type ("controller", "disk", etc). These point to the same
place as (*detach)->data.blah, it's just easier on the eyes to have,
e.g., "disk->dst" rather than "(*detach)->data.disk-dst".
Signed-off-by: Laine Stump <laine@laine.org>
ACKed-by: Peter Krempa <pkrempa@redhat.com>
The Chr and Lease devices have detach code that is too different from
the other device types to handle with common functionality (which will
soon be added at the end of qemuDomainDetachDeviceLive(). In order to
make this difference obvious, move the cases for those two device
types to the top of the switch statement in
qemuDomainDetachDeviceLive(), have the cases return immediately so the
future common code at the end of the function will be skipped, and
also include some hopefully helpful comments to remind future
maintainers why these two device types are treated differently.
Any attempt to detach an unsupported device type should also skip the
future common code at the end of the function, so the case for
unsupported types is similarly changed from a simple break to a return
-1.
Signed-off-by: Laine Stump <laine@laine.org>
ACKed-by: Peter Krempa <pkrempa@redhat.com>
I'm about to add a second virDomainDeviceDef to this function that
will point to the actual device in the domain object. while this is
just a partially filled-in example of what to look for. Naming it
match will make the code easier to follow.
Signed-off-by: Laine Stump <laine@laine.org>
ACKed-by: Peter Krempa <pkrempa@redhat.com>
These are no longer called from qemu_driver.c, since the function that
called them (qemuDomainDetachDeviceLive()) has been moved to
qemu_hotplug.c, and they are no longer called from testqemuhotplug.c
because it now just called qemuDomainDetachDeviceLive() instead of all
the subordinate functions.
Signed-off-by: Laine Stump <laine@laine.org>
ACKed-by: Peter Krempa <pkrempa@redhat.com>
The individual qemuDomainDetach*Device() functions will soon be "less
functional", since some of the code that is duplicated in 10 of the 12
detach functions is going to be moved into the common
qemuDomainDetachDeviceLive(), which calls them all.
qemuhotplugtest.c is the only place any of these individual functions
is called other than qemuDomainDetachDeviceLive() itself. Fortunately,
qemuDomainDetachDeviceLive() provides exactly the functionality needed
by the test driver (except that it supports detach of more device
types than the test driver has tests for).
This patch replaces the calls to
qemuDomainDetach(Chr|Shmen|Watchdog|Disk)Device with a single call to
the higher level function, allowing us to shift functionality between
the lower level functions without breaking the tests.
Signed-off-by: Laine Stump <laine@laine.org>
ACKed-by: Peter Krempa <pkrempa@redhat.com>
qemuDomainDetachDeviceLive() is called from two places in
qemu_driver.c, and qemuDomainUpdateDeviceList() is called from the
end of qemuDomainDetachDeviceLive(), which is now in qemu_hotplug.c
This patch replaces the single call to qemuDomainUpdateDeviceList()
with two calls to it immediately after return from
qemuDomainDetachDeviceLive(). This is only done if the return from
that function is exactly 0, in order to exactly preserve previous
behavior.
Removing that one call from qemuDomainDetachDeviceList() will permit
us to call it from the test driver hotplug test, replacing the
separate calls to qemuDomainDetachDeviceDiskLive(),
qemuDomainDetachChrDevice(), qemuDomainDetachShmemDevice() and
qemuDomainDetachWatchdog(). We want to do this so that part of the
common functionality of those three functions (and the rest of the
device-specific Detach functions) can be pulled up into
qemuDomainDetachDeviceLive() without breaking the test. (This is done
in the next patch).
NB: Almost certainly this is "not the best place" to call
qemuDomainUpdateDeviceList() (actually, it is provably the *wrong*
place), since it's purpose is to retrieve an "up to date" list of
aliases for all devices from qemu, and if the guest OS hasn't yet
processed the detach request, the now-being-removed device may still
be on that list. It would arguably be better to instead call
qemuDomainUpdateDevicesList() later during the response to the
DEVICE_DELETED event for the device. But removing the call from the
current point in the detach could have some unforeseen ill effect due
to changed timing, so the change to move it into
qemuDomainRemove*Device() will be done in a separate patch (in order
to make it easily revertible in case it causes a regression).
Signed-off-by: Laine Stump <laine@laine.org>
ACKed-by: Peter Krempa <pkrempa@redhat.com>
qemuDomainDetachDeviceControllerLive() just checks if the controller
type is SCSI, and then either returns failure, or calls
qemuDomainDetachControllerDevice().
Instead, lets just check for type != SCSI at the top of the latter
function, and call it directly.
Signed-off-by: Laine Stump <laine@laine.org>
ACKed-by: Peter Krempa <pkrempa@redhat.com>
This function is going to take on some of the functionality of its
subordinate functions, which all live in qemu_hotplug.c.
qemuDomainDetachDeviceControllerLive() is only called from
qemuDomainDetachDeviceLive() (and will soon be merged into
qemuDomainDetachControllerDevice(), which is in qemu_hotplug.c), so
it is also moved.
Signed-off-by: Laine Stump <laine@laine.org>
ACKed-by: Peter Krempa <pkrempa@redhat.com>
The Attach and Detach Lease functions were together in the middle of
the Detach functions. Put them at the end of their respective
sections, since they behave differently from the other attach/detach
functions (DetachLease doesn't use qemuDomainDeleteDevice(), and is
always synchronous).
Signed-off-by: Laine Stump <laine@laine.org>
ACKed-by: Peter Krempa <pkrempa@redhat.com>
There were two outliers at the end of the file beyond the Vcpu
functions.
Signed-off-by: Laine Stump <laine@laine.org>
ACKed-by: Peter Krempa <pkrempa@redhat.com>
It was sitting down in the middle of all the qemuDomainDetach*()
functions. Move it up with the rest of the qemuDomain*Graphics*()
functions.
Signed-off-by: Laine Stump <laine@laine.org>
ACKed-by: Peter Krempa <pkrempa@redhat.com>
It's now only called from one place, and combining the two functions
highlights the similarity with Detach functions for other device
types.
Signed-off-by: Laine Stump <laine@laine.org>
ACKed-by: Peter Krempa <pkrempa@redhat.com>
Back in the bad old days different device types required a different
qemu monitor call to detach them, and so an <interface type='hostdev'>
needed to call the function for detaching hostdevs, while other
<interface> types could be deleted as netdevs.
Times have changed, and *all* device types are detached by calling the
common function qemuDomainDeleteDevice(vm, alias), so we don't need to
differentiate between hostdev interfaces and the others for that
reason.
There are a few other netdev-specific functions called during
qemuDomainDetachNetDevice() (clearing bandwidth limits, stopping the
interface), but those turn into NOPs when type=hostdev, so they're
safe to call for type=hostdev.
The only thing that is different + not a NOP is the call to
virDomainAudit*() when qemuDomainDeleteDevice() fails, so if we add a
conditional for that small bit of code, we can eliminate the callout
from qemuDomainDetachNetDevice() to qemuDomainDetachThisDevice(),
which makes this function fit the desired pattern for merging with the
other detach functions, and paves the way to simplifying
qemuDomainDetachHostDevice() too.
Signed-off-by: Laine Stump <laine@laine.org>
ACKed-by: Peter Krempa <pkrempa@redhat.com>
qemuDomainDetachDiskDevice() is only called from one place. Moving the
contents of the function to that place makes
qemuDomainDetachDiskLive() more similar to the other Detach functions
called by the toplevel qemuDomainDetachDevice().
The goal is to make each of the device-type-specific functions do this:
1) find the exact device
2) do any device-specific validation
3) do general validation
4) do device-specific shutdown (only needed for net devices)
5) do the common block of code to send device_del to qemu, then
optionally wait for a corresponding DEVICE_DELETED event from
qemu.
with the final aim being that only items 1 & 2 will remain in each
device-type-specific function, while 3 & 5 (which are the same for
almost every type) will be de-duplicated and moved to the toplevel
function that calls all of these (qemuDomainDetachDeviceLive(), which
will also contain a callout to the one instance of (4) (netdev).
Signed-off-by: Laine Stump <laine@laine.org>
ACKed-by: Peter Krempa <pkrempa@redhat.com>
qemuDomainDetachHostDevice() has a check at the end that calls
qemuDomainDetachNetDevice() in the case that the hostdev is actually a
Net device of type='hostdev'. A long time ago when device removal was
(supposedly but not actually) synchronous, this would cause some extra
code to be run prior to removing the device (e.g. restoring the original MAC
address of the device, undoing some sort of virtual port profile, etc).
For quite awhile now the device removal has been asynchronous, so that
"extra teardown" isn't handled by the detach function, but instead is
handled by the Remove function called at a later time. The result is
that when we call qemuDomainDetachNetDevice() from
qemuDomainDetachHostDevice(), it ends up just calling
qemuDomainDetachThisHostDevice() and returning, which is exactly what
we do for all other hostdevs anyway.
Based on that, remove the behavioral difference when parent.type ==
VIR_DOMAIN_DEVICE_NET, and just call qemuDomainDetachThisHostDevice()
for all hostdevs.
Signed-off-by: Laine Stump <laine@laine.org>
ACKed-by: Peter Krempa <pkrempa@redhat.com>
There are separate Detach functions for PCI, USB, SCSI, Vhost, and
Mediated hostdevs, but the functions are all 100% the same code,
except that the PCI function checks for the guest side of the device
being a PCI Multifunction device, while the other 4 check that the
device's alias != NULL.
The check for multifunction PCI devices should be done for *all*
devices that are connected to the PCI bus in the guest, not just PCI
hostdevs, and qemuIsMultiFunctionDevice() conveniently returns false
if the queried device doesn't connect with PCI, so it is safe to make
this check for all hostdev devices. (It also needs to be done for many
other device types, but that will be addressed in a future patch).
Likewise, since all hostdevs are detached by calling
qemuDomainDeleteDevice(), which requires the device's alias, checking
for a valid alias is a reasonable thing for PCI hostdevs too.
Signed-off-by: Laine Stump <laine@laine.org>
ACKed-by: Peter Krempa <pkrempa@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
Having an InfoPtr named "dev" made my brain hurt. Renaming it to
"info" gives one less thing to confuse when looking at the code.
Signed-off-by: Laine Stump <laine@laine.org>
ACKed-by: Peter Krempa <pkrempa@redhat.com>
When support for hotplug/unplug of SCSI controllers was added way back
in December 2009 (commit da9d937b), unplug was handled by calling the
now-extinct function qemuMonitorRemovePCIDevice(), which required a
PCI address as an argument. At the same time, the idea of every device
in the config having a PCI address apparently was not yet fully
implemented, because the author of the patch including a check for a
valid PCI address in the device object.
These days, all PCI devices are guaranteed to have a valid PCI
address. But more important than that, we no longer detach devices by
PCI address, but instead use qemuDomainDeleteDevice(), which
identifies the device by its alias. So checking for a valid PCI
address is just pointless extra code that obscures the high level of
similarity between all the individual qemuDomainDetach*Device()
functions.
Signed-off-by: Laine Stump <laine@laine.org>
ACKed-by: Peter Krempa <pkrempa@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
qemuDomainRemoveRNGDevice() calls qemuDomainDetachExtensionDevice().
According to commit 1d1e264f1 that added this code, it should not be
necessary to explicitly remove the zPCI extension device for a PCI
device during unplug, because "QEMU implements an unplug callback
which will unplug both PCI and zPCI device in a cascaded way". In
fact, no other devices call qemuDomainDetachExtensionDevice() during
their qemuDomainRemove*Device() function, so it should be removed from
qemuDomainRemoveRNGDevice as well.
Signed-off-by: Laine Stump <laine@laine.org>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
Reviewed-by: Boris Fiuczynski <fiuczy@linux.ibm.com>
qemuDomainDetachControllerDevice() calls
qemuDomainDetachExtensionDevice() when the controller type is
PCI. This is incorrect in multiple ways:
* Any code that tears down a device should be in the
qemuDomainRemove*Device() function (which is called after libvirt
gets a DEVICE_DELETED event from qemu indicating that the guest is
finished with the device on its end. The qemuDomainDetach*Device()
functions should only contain code that ensures the requested
operation is valid, and sends the command to qemu to initiate the
unplug.
* qemuDomainDetachExtensionDevice() is a function that applies to
devices that plug into a PCI slot, *not* necessarily PCI controllers
(which is what's being checked in the offending code). The proper
way to check for this would be to see if the DeviceInfo for the
controller device had a PCI address, not to check if the controller
is a PCI controller (the code being removed was doing the latter).
* According to commit 1d1e264f1 that added this code (and other
support for hotplugging zPCI devices on s390), it's not necessary to
explicitly detach the zPCI device when unplugging a PCI device. To
quote:
There's no need to implement hot unplug for zPCI as QEMU
implements an unplug callback which will unplug both PCI and
zPCI device in a cascaded way.
and the evidence bears this out - all the other uses of
qemuDomainDetachExtensionDevice() (except one, which I believe is
also in error, and is being removed in a separate patch) are only to
remove the zPCI extension device in cases where it was successfully
added, but there was some other failure later in the hotplug process
(so there was no regular PCI device to remove and trigger removal of
the zPCI extension device).
* PCI controllers are not hot pluggable, so this is dead code
anyway. (The only controllers that can currently be
hotplugged/unplugged are SCSI controllers).
Signed-off-by: Laine Stump <laine@laine.org>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
Reviewed-by: Boris Fiuczynski <fiuczy@linux.ibm.com>
dnsmasq documentation says that the *IPv4* prefix/network
address/broadcast address sent to dhcp clients will be automatically
determined by dnsmasq by looking at the interface it's listening on,
so the original libvirt code did not add a netmask to the dnsmasq
commandline (or later, the dnsmasq conf file).
For *IPv6* however, dnsmasq apparently cannot automatically determine
the prefix (functionally the same as a netmask), and it must be
explicitly provided in the conf file (as a part of the dhcp-range
option). So many years after IPv4 DHCP support had been added, when
IPv6 dhcp support was added the prefix was included at the end of the
dhcp-range setting, but only for IPv6.
A user had reported a bug on a host where one of the interfaces was a
superset of the libvirt network where dhcp is needed (e.g., the host's
ethernet is 10.0.0.20/8, and the libvirt network is 10.10.0.1/24). For
some reason dnsmasq was supplying the netmask for the /8 network to
clients requesting an address on the /24 interface.
This seems like a bug in dnsmasq, but even if/when it gets fixed
there, it looks like there is no harm in just always adding the
netmask to all IPv4 dhcp-range options similar to how prefix is added
to all IPv6 dhcp-range options.
Signed-off-by: Laine Stump <laine@laine.org>
Reviewed-by: John Ferlan <jferlan@redhat.com>
This fixes a bug that has been present since the original version of
the function was pushed in commit 1ab80f3 on Nov. 26 2010 (by me). The
virSocketAddr::len was not being set.
Apparently until now we were always calling
virSocketAddrPrefixToNetmask with virSocketAddr object that was
already (coincidentally) initialized for the proper address family,
but the bug became apparent when trying to use it to fill in an
otherwise uninitialized object.
Signed-off-by: Laine Stump <laine@laine.org>
Reviewed-by: John Ferlan <jferlan@redhat.com>
The libvirt zonefile for firewalld (added in commit 3b71f2e4) does the
following:
1) lists specific services it wants to allow, then
2) uses a lower priority <reject/> rule to block all other services to
the host, and then finally,
3) relies on the zone's default "accept" policy to, accept all
forwarded traffic (since forwarded traffic is ignored by the
slightly higher priority <reject/> rule in (2)).
I had assumed that icmp traffic was either being allowed at the top of
the rules, or that it would be ignored by the <reject/> rule and
passed by the default accept policy (similar to forwarded traffic),
but this assumption was incorrect; the <reject/> rule does block icmp
traffic. This became apparent when DHCPv6 which requires ICMPv6 in
addition to udp/dhcpv6) failed to work.
This all means that in order to achieve our original goal of "similar
behavior to a default reject policy, but also allowing forwarded
traffic", we need to add rules to allow all icmp and icmpv6 traffic to
the libvirt zone, and that's what this patch does.
This is a further refinement of the resolution to
https://bugzilla.redhat.com/1650320
Signed-off-by: Laine Stump <laine@laine.org>
Acked-by: Eric Garver <eric@garver.life>
commit 3bba4825 added the new function virFirewallDInterfaceSetZone()
which calledsends virDBUSCallMethod a DBusMessage** for the reply
message, but doesn't use the reply, and also doesn't free it. Since
this arg is allowed to be NULL, this patch simply sets it to NULL so
we don't have to deal with it.
Signed-off-by: Laine Stump <laine@laine.org>
Reviewed-by: Andrea Bolognani <abologna@redhat.com>
Just before pushing the series containing commit 3bba4825 I had added
a "return true" to the top of virFirewallDZoneExists() to measure the
impact of calling that function once per network during startup. I
found that the effect was minimal, but forgot to remove the "return
true" before pushing. This unfortunately causes a failure to start
networks on systems that have a firewalld version that doesn't support
our libvirt zone file (i.e. pretty much everyone).
This patch removes the unintended line.
Signed-off-by: Laine Stump <laine@laine.org>
Since we're setting the zone anyway, it will be useful to allow
setting a different (custom) zone for each network. This will be done
by adding a "zone" attribute to the "bridge" element, e.g.:
...
<bridge name='virbr0' zone='myzone'/>
...
If a zone is specified in the config and it can't be honored, this
will be an error.
Signed-off-by: Laine Stump <laine@laine.org>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
This patch restores broken guest network connectivity after a host
firewalld is switched to using an nftables backend. It does this by
adding libvirt networks' bridge interfaces to the new "libvirt" zone
in firewalld.
After this patch, the bridge interface of any network created by
libvirt (when firewalld is active) will be added to the firewalld
zone called "libvirt" if it exists (regardless of the firewalld
backend setting). This behavior does *not* depend on whether or not
libvirt has installed the libvirt zone file (set with
"--with[out]-firewalld-zone" during the configure phase of the package
build).
If the libvirt zone doesn't exist (either because the package was
configured to not install it, or possibly it was installed, but
firewalld doesn't support rule priorities, resulting in a parse
error), the bridge will remain in firewalld's default zone, which
could be innocuous (in the case that the firewalld backend is
iptables, guest networking will still function properly with the
bridge in the default zone), or it could be disastrous (if the
firewalld backend is nftables, we can be assured that guest networking
will fail). In order to be unobtrusive in the former case, and
informative in the latter, when the libvirt zone doesn't exist we
then check the firewalld version to see if it's new enough to support
the nftables backend, and then if the backend is actually set to
nftables, before logging an error (and failing the net-start
operation, since the network couldn't possibly work anyway).
When the libvirt zone is used, network behavior is *slightly*
different from behavior of previous libvirt. In the past, libvirt
network behavior would be affected by the configuration of firewalld's
default zone (usually "public"), but now it is affected only by the
"libvirt" zone), and thus almost surely warrants a release note for
any distro upgrading to libvirt 5.1 or above. Although it's
unfortunate that we have to deal with a mandatory behavior change, the
architecture of multiple hooks makes it impossible to *not* change
behavior in some way, and the new behavior is arguably better (since
it will now be possible to manage access to the host from virtual
machines vs from public interfaces separately).
Creates-and-Resolves: https://bugzilla.redhat.com/1650320
Resolves: https://bugzilla.redhat.com/1638342
Signed-off-by: Laine Stump <laine@laine.org>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
In the past (when both libvirt and firewalld used iptables), if either
libvirt's rules *OR* firewalld's rules accepted a packet, it would
be accepted. This was because libvirt and firewalld rules were
processed during the same kernel hook, and a single ACCEPT result
would terminate the rule traversal and cause the packet to be
accepted.
But now firewalld can use nftables for its backend, while libvirt's
firewall rules are still using iptables; iptables rules are still
processed, but at a different time during packet processing
(i.e. during a different hook) than the firewalld nftables rules. The
result is that a packet must be accepted by *BOTH* the libvirt
iptables rules *AND* the firewalld nftable rules in order to be
accepted.
This causes pain because
1) libvirt always adds rules to permit DNS and DHCP (and sometimes
TFTP) from guests to the host network's bridge interface. But
libvirt's bridges are in firewalld's "default" zone (which is usually
the zone called "public"). The public zone allows ssh, but doesn't
allow DNS, DHCP, or TFTP. So even though libvirt's rules allow the
DHCP and DNS traffic, the firewalld rules (now processed during a
different hook) dont, thus guests connected to libvirt's bridges can't
acquire an IP address from DHCP, nor can they make DNS queries to the
DNS server libvirt has setup on the host. (This could be solved by
modifying the default firewalld zone to allow DNS and DHCP, but that
would open *all* interfaces in the default zone to those services,
which is most likely not what the host's admin wants.)
2) Even though libvirt adds iptables rules to allow forwarded traffic
to pass the iptables hook, firewalld's higher level "rich rules" don't
yet have the ability to configure the acceptance of forwarded traffic
(traffic that is going somewhere beyond the host), so any traffic that
needs to be forwarded from guests to the network beyond the host is
rejected during the nftables hook by the default zone's "default
reject" policy (which rejects all traffic in the zone not specifically
allowed by the rules in the zone, whether that traffic is destined to
be forwarded or locally received by the host).
libvirt can't send "direct" nftables rules (firewalld only supports
direct/passthrough rules for iptables), so we can't solve this problem
by just sending explicit nftables rules instead of explicit iptables
rules (which, if it could be done, would place libvirt's rules in the
same hook as firewalld's native rules, and thus eliminate the need for
packets to be accepted by both libvirt's and firewalld's own rules).
However, we can take advantage of a quirk in firewalld zones that have
a default policy of "accept" (meaning any packet that doesn't match a
specific rule in the zone will be *accepted*) - this default accept will
also accept forwarded traffic (not just traffic destined for the host).
Of course we don't want to modify firewalld's default zone in that
way, because that would affect the filtering of traffic coming into
the host from other interfaces using that zone. Instead, we will
create a new zone called "libvirt". The libvirt zone will have a
default policy of accept so that forwarded traffic can pass and list
specific services that will be allowed into the host from guests (DNS,
DHCP, SSH, and TFTP).
But the same default accept policy that fixes forwarded traffic also
causes *all* traffic from guest to host to be accepted. To close this
new hole, the libvirt zone can take advantage of a new feature in
firewalld (currently slated for firewalld-0.7.0) - priorities for rich
rules - to add a low priority rule that rejects all local traffic (but
leaves alone all forwarded traffic).
So, our new zone will start with a list of services that are allowed
(dhcp, dns, tftp, and ssh to start, but configurable via any firewalld
management application, or direct editing of the zone file in
/etc/firewalld/zones/libvirt.xml), followed by a low priority
<reject/> rule (to reject all other traffic from guest to host), and
finally with a default policy of accept (to allow forwarded traffic).
This patch only creates the zonefile for the new zone, and implements
a configure.ac option to selectively enable/disable installation of
the new zone. A separate patch contains the necessary code to actually
place bridge interfaces in the libvirt zone.
Why do we need a configure option to disable installation of the new
libvirt zone? It uses a new firewalld attribute that sets the priority
of a rich rule; this feature first appears in firewalld-0.7.0 (unless
it has been backported to am earlier firewalld by a downstream
maintainer). If the file were installed on a system with firewalld
that didn't support rule priorities, firewalld would log an error
every time it restarted, causing confusion and lots of extra bug
reports.
So we add two new configure.ac switches to avoid polluting the system
logs with this error on systems that don't support rule priorities -
"--with-firewalld-zone" and "--without-firewalld-zone". A package
builder can use these to include/exclude the libvirt zone file in the
installation. If firewalld is enabled (--with-firewalld), the default
is --with-firewalld-zone, but it can be disabled during configure
(using --without-firewalld-zone). Targets that are using a firewalld
version too old to support the rule priority setting in the libvirt
zone file can simply add --without-firewalld-zone to their configure
commandline.
These switches only affect whether or not the libvirt zone file is
*installed* in /usr/lib/firewalld/zones, but have no effect on whether
or not libvirt looks for a zone called libvirt and tries to use it.
NB: firewalld zones can only be added to the permanent config of
firewalld, and won't be loaded/enabled until firewalld is restarted,
so at package install/upgrade time we have to restart firewalld. For
rpm-based distros, this is done in the libvirt.spec file by calling
the %firewalld_restart rpm macro, which is a part of the
firewalld-filesystem package. (For distros that don't use rpm
packages, the command "firewalld-cmd --reload" will have the same
effect).
Signed-off-by: Laine Stump <laine@laine.org>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
virFirewallDGetBackend() reports whether firewalld is currently using
an iptables or an nftables backend.
virFirewallDGetVersion() learns the version of the firewalld running
on this system and returns it as 1000000*major + 1000*minor + micro.
virFirewallDGetZones() gets a list of all currently active firewalld
zones.
virFirewallDInterfaceSetZone() sets the firewalld zone of the given
interface.
virFirewallDZoneExists() can be used to learn whether or not a
particular zone is present and active in firewalld.
Signed-off-by: Laine Stump <laine@laine.org>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
In preparation for adding several other firewalld-specific functions,
separate the code that's unique to firewalld from the more-generic
"firewall" file.
Signed-off-by: Laine Stump <laine@laine.org>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Support for firewalld is a feature that can be selectively enabled or
disabled (using --with-firewalld/--without-firewalld), not merely
something that must be accounted for in the code if it is present with
no exceptions. It is more consistent with other usage in libvirt to
use WITH_FIREWALLD rather than HAVE_FIREWALLD.
Signed-off-by: Laine Stump <laine@laine.org>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Laine Stump
Laine Stump