By default, pfifo_fast queueing discipline (qdisc) is set on
newly created interfaces (including TAPs). This qdisc has three
queues and packets that want to be sent through given NIC are
placed into one of the queues based on TOS field. Queues are then
emptied based on their priority allowing interactive sessions
stay interactive whilst something else is downloading a large
file.
Obviously, this means that kernel has to be involved and some
locking has to happen (when placing packets into queues). If
virtualization is taken into account then the above algorithm
happens twice - once in the guest and the second time in the
host.
This is arguably not optimal as it burns host CPU cycles
needlessly. Guest already made it choice and sent packets in the
order it wants.
To resolve this, Linux kernel offers 'noqueue' qdisc which can be
applied on virtual interfaces and in fact for 'lo' it is by
default:
lo: <LOOPBACK,UP,LOWER_UP> mtu 65536 qdisc noqueue
Set it for other TAP devices we create for domains too. With this
change I was able to squeeze 1Mbps more from a macvtap attached
to a guest and to my 1Gbps LAN (as measured by iperf3).
Resolves:
https://bugzilla.redhat.com/show_bug.cgi?id=1329644
Signed-off-by: Michal Privoznik <mprivozn(a)redhat.com>
---
src/qemu/qemu_command.c | 2 ++
src/qemu/qemu_domain.c | 36 ++++++++++++++++++++++++++++++++++++
src/qemu/qemu_domain.h | 4 ++++
src/qemu/qemu_hotplug.c | 2 ++
4 files changed, 44 insertions(+)
diff --git a/src/qemu/qemu_command.c b/src/qemu/qemu_command.c
index 9519861e92..eec860382c 100644
--- a/src/qemu/qemu_command.c
+++ b/src/qemu/qemu_command.c
@@ -8267,6 +8267,8 @@ qemuBuildInterfaceCommandLine(virQEMUDriverPtr driver,
}
}
+ qemuDomainInterfaceSetDefaultQDisc(driver, net);
+
if (net->mtu &&
virNetDevSetMTU(net->ifname, net->mtu) < 0)
goto cleanup;
diff --git a/src/qemu/qemu_domain.c b/src/qemu/qemu_domain.c
index 9623123d3c..72da28666f 100644
--- a/src/qemu/qemu_domain.c
+++ b/src/qemu/qemu_domain.c
@@ -11048,3 +11048,39 @@ qemuDomainFileWrapperFDClose(virDomainObjPtr vm,
}
return ret;
}
+
+
+/**
+ * qemuDomainInterfaceSetDefaultQDisc:
+ * @driver: QEMU driver
+ * @net: domain interface
+ *
+ * Set the noqueue qdisc on @net if running as privileged. The
+ * noqueue qdisc is a lockless transmit and thus faster than the
+ * default pfifo_fast (at least in theory). But we can modify
+ * root qdisc only if we have CAP_NET_ADMIN.
+ *
+ * Returns: 0 on success,
+ * -1 otherwise.
+ */
+int
+qemuDomainInterfaceSetDefaultQDisc(virQEMUDriverPtr driver,
+ virDomainNetDefPtr net)
+{
+ virDomainNetType actualType = virDomainNetGetActualType(net);
+
+ if (!driver->privileged || !net->ifname)
+ return 0;
+
+ /* We want only those types which are represented as TAP
+ * devices in the host. */
+ if (actualType == VIR_DOMAIN_NET_TYPE_ETHERNET ||
+ actualType == VIR_DOMAIN_NET_TYPE_NETWORK ||
+ actualType == VIR_DOMAIN_NET_TYPE_BRIDGE ||
+ actualType == VIR_DOMAIN_NET_TYPE_DIRECT) {
+ if (virNetDevSetRootQDisc(net->ifname, "noqueue") < 0)
+ return -1;
+ }
+
+ return 0;
+}
diff --git a/src/qemu/qemu_domain.h b/src/qemu/qemu_domain.h
index 9bf32e16c9..91b3b67cb6 100644
--- a/src/qemu/qemu_domain.h
+++ b/src/qemu/qemu_domain.h
@@ -1046,3 +1046,7 @@ qemuDomainOpenFile(virQEMUDriverPtr driver,
int
qemuDomainFileWrapperFDClose(virDomainObjPtr vm,
virFileWrapperFdPtr fd);
+
+int
+qemuDomainInterfaceSetDefaultQDisc(virQEMUDriverPtr driver,
+ virDomainNetDefPtr net);
diff --git a/src/qemu/qemu_hotplug.c b/src/qemu/qemu_hotplug.c
index 7a54fcb221..2c184b9ba0 100644
--- a/src/qemu/qemu_hotplug.c
+++ b/src/qemu/qemu_hotplug.c
@@ -1368,6 +1368,8 @@ qemuDomainAttachNetDevice(virQEMUDriverPtr driver,
virNetDevSetMTU(net->ifname, net->mtu) < 0)
goto cleanup;
+ qemuDomainInterfaceSetDefaultQDisc(driver, net);
+
for (i = 0; i < tapfdSize; i++) {
if (qemuSecuritySetTapFDLabel(driver->securityManager,
vm->def, tapfd[i]) < 0)
--
2.26.2