<div class="zcontentRow"> On 06/14/2017 04:29 AM, ZhiPeng Lu wrote: > Before libvirt that calls virNetDevMacVLanCreateWithVPortProfile sets mac address > or vlan of a Virtual Function(VF) linked to a macvtap passthrough device of migration > destination host in migration start step. If we ping the migrating vm, > we get the network does not pass. Because VFs of migration source and destination > have the same MAC address. The patch later calling qemuMigrationVPAssociatePortProfiles > sets mac address of VF in migration finish step instead of start step. > The patch aims to reduce packet loss rate. >I missed this patch when you initially sent it, and see that nobody else >responded... >What you're doing here shouldn't be needed. During migration, a macvtap >device should be created with ~IFF_UP, and not brought online until >qemuInterfaceStartDevice() is called just prior to starting the guest >CPUs on the destination; by this time the guest CPUs on the source have >already been stopped (and the guest's macvtap interfaces on the source >have been set offline). Until a device is IFF_UP, it shouldn't send out >any traffic with its MAC address. ----Probably not understanding what I mean when the network packet loss happens. This patch describes the time before the guest CPU on the source has been stopped and the macvtap device was IFF_UP.if we send network data from the destination, the network packet loss. >If you look at the macvtap device on the destination and it has IFF_UP >set prior to the completion of migration, then there is a bug in libvirt >that we need to fix. Otherwise, I think any packet loss you're seeing is >coming from something else. > > Signed-off-by: ZhiPeng Lu <lu.zhipeng@zte.com.cn> > --- > src/qemu/qemu_migration.c | 18 ++++++++++++++++-- > src/util/virnetdevmacvlan.c | 17 +++++++++++------ > 2 files changed, 27 insertions(+), 8 deletions(-) > > diff --git a/src/qemu/qemu_migration.c b/src/qemu/qemu_migration.c > index 09adb04..795ed71 100644 > --- a/src/qemu/qemu_migration.c > +++ b/src/qemu/qemu_migration.c > @@ -5004,7 +5004,7 @@ qemuMigrationPerform(virQEMUDriverPtr driver, > } > > static int > -qemuMigrationVPAssociatePortProfiles(virDomainDefPtr def) > +qemuMigrationVPAssociatePortProfiles(virDomainDefPtr def, const char *stateDir) <div class="zMailSign"><div><div><div><div> 芦志朋 luzhipeng IT开发工程师 IT Development Engineer 操作系统产品部/中心研究院/系统产品 OS Product Dept./Central R＆D Institute/System Product <table style="color: rgb(0, 0, 0); font-family: 宋体; widows: 1;"><tbody><tr class="firstRow"><td valign="top" width="100"><img id="sign-icon" src="cid:9ae3e214c17d49ed935d87c674ba3ee2" width="130" height="120"></td><td valign="top" width="500"><img id="sign-logo" src="cid:24242e5637af428891c4db731e7765ad" width="115" height="38"> 深圳市南山区科技南路55号中兴通讯研发大楼33楼 33/F, R&D Building, ZTE Corporation Hi-tech Road South, Hi-tech Industrial Park Nanshan District, Shenzhen, P.R.China, 518057 T: +86 755 xxxxxxxx F:+86 755 xxxxxxxx M: +86 xxxxxxxxxxx E: <a href="mailto:lu.zhipeng@zte.com.cn">lu.zhipeng@zte.com.cn</a> <a href="http://www.zte.com.cn/" target="_blank">www.zte.com.cn</a></td></tr></tbody></table></div></div></div></div></div><div><div class="zhistoryRow" style="display:block"><div class="zhistoryDes" style="width: 100%; height: 28px; line-height: 28px; background-color: #E0E5E9; color: #1388FF; text-align: center;" language-data="HistoryOrgTxt">原始邮件</div><div id="zwriteHistoryContainer"><div class="control-group zhistoryPanel"><div class="zhistoryHeader" style="padding: 8px; background-color: #F5F6F8;"><div>发件人： <laine@laine.org>;</div><div>收件人： <libvir-list@redhat.com>;</div><div>抄送人：芦志朋10108272;</div><div>日期：2017年07月17日 05:01</div><div>主题：Re: [libvirt] [v2 RESEND PATCH] qemu: reduce packet loss rate for vmwith macvtap passthrough mode in migration</div></div> <div>On 06/14/2017 04:29 AM, ZhiPeng Lu wrote: > Before libvirt that calls virNetDevMacVLanCreateWithVPortProfile sets mac address > or vlan of a Virtual Function(VF) linked to a macvtap passthrough device of migration > destination host in migration start step. If we ping the migrating vm, > we get the network does not pass. Because VFs of migration source and destination > have the same MAC address. The patch later calling qemuMigrationVPAssociatePortProfiles > sets mac address of VF in migration finish step instead of start step. > The patch aims to reduce packet loss rate. I missed this patch when you initially sent it, and see that nobody else responded... What you're doing here shouldn't be needed. During migration, a macvtap device should be created with ~IFF_UP, and not brought online until qemuInterfaceStartDevice() is called just prior to starting the guest CPUs on the destination; by this time the guest CPUs on the source have already been stopped (and the guest's macvtap interfaces on the source have been set offline). Until a device is IFF_UP, it shouldn't send out any traffic with its MAC address. If you look at the macvtap device on the destination and it has IFF_UP set prior to the completion of migration, then there is a bug in libvirt that we need to fix. Otherwise, I think any packet loss you're seeing is coming from something else. > > Signed-off-by: ZhiPeng Lu <lu.zhipeng@zte.com.cn> > --- > src/qemu/qemu_migration.c | 18 ++++++++++++++++-- > src/util/virnetdevmacvlan.c | 17 +++++++++++------ > 2 files changed, 27 insertions(+), 8 deletions(-) > > diff --git a/src/qemu/qemu_migration.c b/src/qemu/qemu_migration.c > index 09adb04..795ed71 100644 > --- a/src/qemu/qemu_migration.c > +++ b/src/qemu/qemu_migration.c > @@ -5004,7 +5004,7 @@ qemuMigrationPerform(virQEMUDriverPtr driver, > } > > static int > -qemuMigrationVPAssociatePortProfiles(virDomainDefPtr def) > +qemuMigrationVPAssociatePortProfiles(virDomainDefPtr def, const char *stateDir) > { > size_t i; > int last_good_net = -1; > @@ -5013,6 +5013,20 @@ qemuMigrationVPAssociatePortProfiles(virDomainDefPtr def) > for (i = 0; i < def->nnets; i++) { > net = def->nets[i]; > if (virDomainNetGetActualType(net) == VIR_DOMAIN_NET_TYPE_DIRECT) { > + if ((!virDomainNetGetActualVirtPortProfile(net) || (virDomainNetGetActualVirtPortProfile(net) && > + virDomainNetGetActualVirtPortProfile(net)->virtPortType != VIR_NETDEV_VPORT_PROFILE_8021QBG && > + virDomainNetGetActualVirtPortProfile(net)->virtPortType != VIR_NETDEV_VPORT_PROFILE_8021QBH)) && > + virDomainNetGetActualDirectMode(net) == > + VIR_NETDEV_MACVLAN_MODE_PASSTHRU) { > + if (virNetDevSaveNetConfig(virDomainNetGetActualDirectDev(net), > + -1, stateDir, false) < 0) { > + goto err_exit; > + } > + if (virNetDevSetNetConfig(virDomainNetGetActualDirectDev(net), > + -1, NULL, virDomainNetGetActualVlan(net), &net->mac, false) < 0) { > + goto err_exit; > + } > + } > if (virNetDevVPortProfileAssociate(net->ifname, > virDomainNetGetActualVirtPortProfile(net), > &net->mac, > @@ -5187,7 +5201,7 @@ qemuMigrationFinish(virQEMUDriverPtr driver, > goto endjob; > } > > - if (qemuMigrationVPAssociatePortProfiles(vm->def) < 0) > + if (qemuMigrationVPAssociatePortProfiles(vm->def, cfg->stateDir) < 0) > goto endjob; > > if (mig->network && qemuDomainMigrateOPDRelocate(driver, vm, mig) < 0) > diff --git a/src/util/virnetdevmacvlan.c b/src/util/virnetdevmacvlan.c > index 7222b0f..682dcd1 100644 > --- a/src/util/virnetdevmacvlan.c > +++ b/src/util/virnetdevmacvlan.c > @@ -1020,12 +1020,17 @@ virNetDevMacVLanCreateWithVPortProfile(const char *ifnameRequested, > */ > setVlan = false; > } > - > - if (virNetDevSaveNetConfig(linkdev, -1, stateDir, setVlan) < 0) > - return -1; > - > - if (virNetDevSetNetConfig(linkdev, -1, NULL, vlan, macaddress, setVlan) < 0) > - return -1; > + if (vmOp != VIR_NETDEV_VPORT_PROFILE_OP_MIGRATE_IN_START && > + virtPortProfile && (virtPortProfile->virtPortType == VIR_NETDEV_VPORT_PROFILE_8021QBH || > + virtPortProfile->virtPortType == VIR_NETDEV_VPORT_PROFILE_8021QBG)) { > + if (virNetDevSaveNetConfig(linkdev, -1, stateDir, setVlan) < 0) { > + return -1; > + } > + if (virNetDevSetNetConfig(linkdev, -1, NULL, vlan, macaddress, > + setVlan) < 0) { > + return -1; > + } > + } > } > > if (ifnameRequested) { </div> </div></div></div></div> </div>