render / rpms / libvirt

Forked from rpms/libvirt 9 months ago
Clone
Blob Blame History Raw
From e09abaad84630d6cd25a1ecff1811ce070201006 Mon Sep 17 00:00:00 2001
Message-Id: <e09abaad84630d6cd25a1ecff1811ce070201006@dist-git>
From: Laine Stump <laine@laine.org>
Date: Thu, 3 Sep 2015 12:08:25 -0400
Subject: [PATCH] util: don't use netlink to save/set mac for
 macvtap+passthrough+802.1Qbh

Before libvirt sets the MAC address of the physdev (the physical
ethernet device) linked to a macvtap passthrough device, it always
saves the previous MAC address to restore when the guest is finished
(following a "leave nothing behind" policy). For a long time it
accomplished the save/restore with a combination of
ioctl(SIOCGIFHWADDR) and ioctl(SIOCSIFHWADDR), but in commit cbfe38c
(first in libvirt 1.2.15) this was changed to use netlink RTM_GETLINK
and RTM_SETLINK commands sent to the Physical Function (PF) of any
device that was detected to be a Virtual Function (VF).

We later found out that this caused problems with any devices using
the Cisco enic driver (e.g. vmfex cards) because the enic driver
hasn't implemented the function that is called to gather the
information in the IFLA_VFINFO_LIST attribute of RTM_GETLINK
(ndo_get_vf_config() for those keeping score), so we would never get
back a useful response.

In an ideal world, all drivers would implement all functions, but it
turns out that in this case we can work around this omission without
any bad side effects - since all macvtap passthrough <interface>
definitions pointing to a physdev that uses the enic driver *must*
have a <virtualport type='802.1Qbh'>, and since no other type of
ethernet devices use 802.1Qbh, libvirt can change its behavior in this
case to use the old-style.  ioctl(SIOC[GS]IFHWADDR).  That's what this
patch does.

Resolves:  https://bugzilla.redhat.com/show_bug.cgi?id=1257004
(cherry picked from commit 46cf0cefa0168a0d929ca87010f59e1cba6c689b)

Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
---
 src/util/virnetdevmacvlan.c | 26 ++++++++++++++++++++++----
 1 file changed, 22 insertions(+), 4 deletions(-)

diff --git a/src/util/virnetdevmacvlan.c b/src/util/virnetdevmacvlan.c
index 213b8eb..641bd24 100644
--- a/src/util/virnetdevmacvlan.c
+++ b/src/util/virnetdevmacvlan.c
@@ -778,9 +778,22 @@ int virNetDevMacVLanCreateWithVPortProfile(const char *tgifname,
      * This is especially important when using SRIOV capable cards that
      * emulate their switch in firmware.
      */
+
     if (mode == VIR_NETDEV_MACVLAN_MODE_PASSTHRU) {
-        if (virNetDevReplaceNetConfig(linkdev, -1, macaddress, -1, stateDir) < 0)
-            return -1;
+        if (virtPortProfile &&
+            virtPortProfile->virtPortType == VIR_NETDEV_VPORT_PROFILE_8021QBH) {
+            /* The Cisco enic driver (the only card that uses
+             * 802.1Qbh) doesn't support IFLA_VFINFO_LIST, which is
+             * required for virNetDevReplaceNetConfig(), so we must
+             * use this function (which uses ioctl(SIOCGIFHWADDR)
+             * instead or virNetDevReplaceNetConfig()
+             */
+            if (virNetDevReplaceMacAddress(linkdev, macaddress, stateDir) < 0)
+                return -1;
+        } else {
+            if (virNetDevReplaceNetConfig(linkdev, -1, macaddress, -1, stateDir) < 0)
+                return -1;
+        }
     }
 
     if (tgifname) {
@@ -913,8 +926,13 @@ int virNetDevMacVLanDeleteWithVPortProfile(const char *ifname,
     int ret = 0;
     int vf = -1;
 
-    if (mode == VIR_NETDEV_MACVLAN_MODE_PASSTHRU)
-        ignore_value(virNetDevRestoreNetConfig(linkdev, vf, stateDir));
+    if (mode == VIR_NETDEV_MACVLAN_MODE_PASSTHRU) {
+        if (virtPortProfile &&
+             virtPortProfile->virtPortType == VIR_NETDEV_VPORT_PROFILE_8021QBH)
+            ignore_value(virNetDevRestoreMacAddress(linkdev, stateDir));
+        else
+            ignore_value(virNetDevRestoreNetConfig(linkdev, vf, stateDir));
+    }
 
     if (ifname) {
         if (virNetDevVPortProfileDisassociate(ifname,
-- 
2.5.1