服务质量（英语：Quality of Service，缩写QoS）是一个术语，在分组交换网络领域中指网络满足给定业务合约的几率；或在许多情况下，非正式地指分组在网络中两点间通过的几率。QoS是一种控制机制，它提供了针对不同用户或者不同数据流采用相应不同的优先级，或者是根据应用程序的要求，保证数据流的性能达到一定的水准。QoS的保证对于容量有限的网络来说是十分重要的，特别是对于流多媒体应用，例如VoIP和IPTV等，因为这些应用常常需要固定的传输率，对延迟也比较敏感。
Vdsm defines a concept of “host network”. These networks are configured by the following Vdsm verbs: addNetwork Add the required Linux networking devices for a new network, as well as the configuration files required to re-create these devices on next boot. delNetwork Delete a previously-added network. editNetwork Replace an existing network definition by a new one. setSafeNetConfig Declare network configuration as “safe”, so it persists after host reboot. setupNetworks
# Networking-related functions @api.logged(on="api.network") def setupNetworks(self, networks, bondings, options): """Add a new network to this vds, replacing an old one."""
if not self._cif._networkSemaphore.acquire(blocking=False): self.log.warn('concurrent network verb already executing') return errCode['unavail']
try: self._cif._netConfigDirty = True supervdsm.getProxy().setupNetworks(networks, bondings, options) if options.get('commitOnSuccess'): # This option ensures that persist is called after # setupNetworks self._cif._netConfigDirty = False return {'status': doneCode} except ConfigNetworkError as e: self.log.error('%s', e.msg, exc_info=True) return {'status': {'code': e.errCode, 'message': e.msg}} except exception.HookError as e: return response.error('hookError', 'Hook error: ' + str(e)) finally: self._cif._networkSemaphore.release()
# https://cn.linux-console.net/?p=2621 声明式网络配置工具 # 这段注释说明了 vdsm 的 network 模块如何使用 nmstate 来设置网络， # 并通过事务和回滚机制来确保网络配置的正确性和可用性。这样可以更好地管理网络配置，并提供可靠的网络连接。 def _setup_nmstate(networks, bondings, options, in_rollback): """ Setup the networks using nmstate as the backend provider. nmstate handles the rollback by itself in case of an error during the transaction or in case the resulted state does not include the desired one.
In order to support the connectivity check, the "regular" rollback is used (the Transaction context). """ logging.info('Processing setup through nmstate') desired_state = nmstate.generate_state(networks, bondings) logging.info('Desired state: %s', desired_state) _setup_dynamic_src_routing(networks) nmstate.setup(desired_state, verify_change=not in_rollback) net_info = NetInfo(netinfo_get())
with Transaction(in_rollback=in_rollback, persistent=False) as config: _setup_qos(networks, net_info, config.networks) for net_name, net_attrs in six.viewitems(networks): if net_attrs.get('remove'): config.removeNetwork(net_name) for net_name, net_attrs in six.viewitems(networks): # 如果是 remove 就执行，remove操作，不是remove，就执行set操作，所以实际上就一个删除，一个更新 if not net_attrs.get('remove'): config.setNetwork(net_name, net_attrs) for bond_name, bond_attrs in six.viewitems(bondings): if bond_attrs.get('remove'): config.removeBonding(bond_name) for bond_name, bond_attrs in six.viewitems(bondings): if not bond_attrs.get('remove'): config.setBonding(bond_name, bond_attrs) config.save() link_setup.setup_custom_bridge_opts(networks) connectivity.check(options)
def configure_outbound(qosOutbound, device, vlan_tag): """Adds the qosOutbound configuration to the backing device (be it bond or nic). Adds a class and filter for default traffic if necessary. vlan_tag can be None""" root_qdisc = netinfo_qos.get_root_qdisc(tc.qdiscs(device)) class_id = '%x' % (_NON_VLANNED_ID if vlan_tag is None else vlan_tag) if not root_qdisc or root_qdisc['kind'] != _SHAPING_QDISC_KIND: _fresh_qdisc_conf_out(device, vlan_tag, class_id, qosOutbound) else: _qdisc_conf_out( device, root_qdisc['handle'], vlan_tag, class_id, qosOutbound )
def _qdisc_conf_out(dev, root_qdisc_handle, vlan_tag, class_id, qos): """Adds the traffic class and filtering to the current hfsc qdisc""" flow_id = _ROOT_QDISC_HANDLE + class_id
filters = [ filt for filt in tc._filters(dev, parent=root_qdisc_handle) if flow_id in (filt_flow_id(filt, 'basic'), filt_flow_id(filt, 'u32')) ]
# Clear up any previous filters to the class for filt in filters: try: tc.filter.delete(dev, filt['pref'], parent=root_qdisc_handle) except tc.TrafficControlException as tce: if tce.errCode != errno.EINVAL: # no filters exist -> EINVAL raise
# Clear the class in case it exists try: tc.cls.delete(dev, classid=root_qdisc_handle + class_id) except tc.TrafficControlException as tce: if tce.errCode != errno.ENOENT: raise
_add_hfsc_cls(dev, root_qdisc_handle, class_id, **qos) if class_id == _DEFAULT_CLASSID: _add_non_vlanned_filter(dev, root_qdisc_handle) else: if not _is_explicit_defined_default_class(dev): (default_class,) = [ c['hfsc'] for c in tc.classes(dev) if c['handle'] == _ROOT_QDISC_HANDLE + _DEFAULT_CLASSID ] ls_max_rate = _max_hfsc_ls_rate(dev) default_class['ls']['m2'] = ls_max_rate
@api.guard(_not_migrating) def hotplugNic(self, params): xml = params['xml'] nic = vmdevices.common.dev_from_xml(self, xml) dom = xmlutils.fromstring(xml) dom_devices = vmxml.find_first(dom, 'devices') nic_dom = next(iter(dom_devices)) nicXml = xmlutils.tostring(nic_dom) nicXml = hooks.before_nic_hotplug( nicXml, self._custom, params=nic.custom ) nic._deviceXML = nicXml # TODO: this is debug information. For 3.6.x we still need to # see the XML even with 'info' as default level. self.log.info("Hotplug NIC xml: %s", nicXml)
try: nic.setup() self._dom.attachDevice(nicXml) except libvirt.libvirtError as e: self.log.exception("Hotplug failed") nicXml = hooks.after_nic_hotplug_fail( nicXml, self._custom, params=nic.custom) if e.get_error_code() == libvirt.VIR_ERR_NO_DOMAIN: raise exception.NoSuchVM() return response.error('hotplugNic', str(e)) else: # FIXME! We may have a problem here if vdsm dies right after # we sent command to libvirt and before save conf. In this case # we will gather almost all needed info about this NIC from # the libvirt during recovery process. device_conf = self._devices[hwclass.NIC] device_conf.append(nic) self._hotplug_device_metadata(hwclass.NIC, nic) self._updateDomainDescriptor() vmdevices.network.Interface.update_device_info(self, device_conf) hooks.after_nic_hotplug(nicXml, self._custom, params=nic.custom)
mirroredNetworks = [] try: # pylint: disable=no-member for network in nic.portMirroring: supervdsm.getProxy().setPortMirroring(network, nic.name) mirroredNetworks.append(network) # The better way would be catch the proper exception. # One of such exceptions is TrafficControlException, but # I am not sure that we'll get it for all traffic control errors. # In any case we need below rollback for all kind of failures. except Exception as e: self.log.exception("setPortMirroring for network %s failed", network) nic_element = xmlutils.fromstring(nicXml) vmxml.replace_first_child(dom_devices, nic_element) hotunplug_params = {'xml': xmlutils.tostring(dom)} self.hotunplugNic(hotunplug_params, port_mirroring=mirroredNetworks) return response.error('hotplugNic', str(e))
def _run(self): self.log.info("VM wrapper has started") if not self.recovering and \ self._altered_state.origin != _MIGRATION_ORIGIN: self._remove_domain_artifacts()
if not self.recovering and not self._altered_state.origin: # We need to define the domain in order to save device metadata in # _make_devices(). It'll get redefined with the final version # later. domxml = libvirtxml.make_placeholder_domain_xml(self) dom = self._connection.defineXML(domxml) self._dom = virdomain.Defined(self.id, dom)
VM Qos（cpu limits和 IO limits）
vdsm 更新 cpu limits 和 IO limits(磁盘 iops) 是通过调用 libvirt 去更新 vm 的 xml 实现的，更新 vm 的方法有多个。其中一个是 updateVmPolicy，在修改已有的 qos 时，会调用这个方法。 还有一个是 hotplugDisk，这是在附加磁盘时会调用，也是去更新 vm 的 xml。 在网页上设置了 cpu qos 和 磁盘 qos以后，需要附加到虚拟机上，才会调用 updateVmPolicy 方法，使得 qos 生效。没附加到虚拟机之前，这些修改不会在 vdsm 的日志中体现。 updateVmPolicy：
vm 探活时执行：探活过程中有判断，判断数据有变化(可能是根据时间戳对比)时调用，探活频率是5秒一次。页面上点击更新是直接触发。
手动移除 qos 时会执行，移除 storage 和 cpu 的 qos 时会触发。 hotplugDisk：
@api.guard(_not_migrating) def updateVmPolicy(self, params): """ Update the QoS policy settings for VMs.
The params argument contains the actual properties we are about to set. It must not be empty.
Supported properties are:
vcpuLimit - the CPU usage hard limit ioTune - the IO limits
In the case not all properties are provided, the missing properties' setting will be left intact.
If there is an error during the processing, this function immediately stops and returns. Remaining properties are not processed.
:param params: dictionary mapping property name to its value :type params: dict[str] -> anything
:return: standard vdsm result structure """ if not params: self.log.error("updateVmPolicy got an empty policy.") return response.error('MissParam', 'updateVmPolicy got an empty policy.')
# # Get the current QoS block metadata_modified = False qos = self._getVmPolicy() if qos is None: return response.error('updateVmPolicyErr')
# # Process provided properties, remove property after it is processed
if 'vcpuLimit' in params: # Remove old value vcpuLimit = vmxml.find_first(qos, "vcpuLimit", None) if vcpuLimit is not None: vmxml.remove_child(qos, vcpuLimit)
except LookupError as e: return response.error('updateVmPolicyErr', str(e))
if ioTuneParams: io_tunes = []
io_tune_element = vmxml.find_first(qos, "ioTune", None) if io_tune_element is not None: io_tunes = vmtune.io_tune_dom_all_to_list(io_tune_element) vmxml.remove_child(qos, io_tune_element)
