Attachment 520373 Details for Bug 733651 – forward port of the RHEL5 patch

[patch] forward port of the RHEL5 patch

xen-netfront-jumbo-frames.patch (text/plain), 6.72 KB, created by Paolo Bonzini on 2011-08-29 12:18:44 UTC

(hide)

Description:

Filename:

MIME Type:

Creator: Paolo Bonzini

Created: 2011-08-29 12:18:44 UTC

Size: 6.72 KB

patch

obsolete

>From 00fb21e9435005a01fdda65bd689e675d81523f6 Mon Sep 17 00:00:00 2001
>From: Paolo Bonzini <pbonzini@redhat.com>
>Date: Fri, 26 Aug 2011 14:04:58 +0200
>Subject: [RHEL6.3 PATCH 0/2] xen-netfront: fix MTU reset after migration
>
>Bugzilla: 733651
>
>Upstream status: Upstream did the same using the hw_features field
>    (and ndo_fix_features/ndo_set_features) that is not in RHEL6, so
>    I'm instead forward porting the RHEL5 changes.  See commits
>    fb50793 and later.
>
>Brew build: http://brewweb.devel.redhat.com/brew/taskinfo?taskID=3585817
>
>The networking backend patches enabled jumbo frame support in the Xen
>networking backend.  However, the front-end would still revert to a
>1500-byte MTU after migration.
>
>The reason for this is that after migration the front-end has to
>renegotiate offloading features with the back-end, and thus it
>calls xennet_set_features.  To ensure the feature set matches the
>backend's, xennet_set_features will temporarily disable scatter/gather.
>Unfortunately this drops the MTU to 1500.
>
>The patches fix the problem by removing the double flip.
>
>(The behavior actually occurs even on the first connection, but it is not
>visible because the MTU is set to 9000 later in the networking scripts.
>Migration however should be transparent to the upper layers, so nobody
>bothers about resetting the MTU).
>
>Paolo Bonzini (2):
>  xen-netfront: force-disable offloading features unsupported by the backend
>  xen-netfront: do not temporarily disable s/g when negotiation features
>
> drivers/net/xen-netfront.c |   64 ++++++++++++++++++++++++++-----------------
> 1 files changed, 39 insertions(+), 25 deletions(-)
>
>-- 
>1.7.6
>
>From 2085842fbd3792de457d75c5de74789c427845d3 Mon Sep 17 00:00:00 2001
>From: Paolo Bonzini <pbonzini@redhat.com>
>Date: Fri, 26 Aug 2011 13:49:16 +0200
>Subject: [RHEL6.3 PATCH 1/2] xen-netfront: force-disable offloading features unsupported by the backend
>
>Even though the Xen backend has support for jumbo frames, currently
>the front-end will still revert to a 1500-byte MTU after migration.
>
>The reason for this is that after migration the front-end has
>to renegotiate offloading features with the back-end, and thus it
>calls xennet_set_features.  To ensure the feature set matches
>the backend's, xennet_set_features will temporarily disable
>scatter/gather.  Unfortunately this drops the MTU to 1500.
>
>As part of fixing this, this patch changes xennet_set_sg and
>xennet_set_tso to always invoke the underlying ethtool_op_set_*
>functions.  If scatter/gather is not supported by the backend, for
>example, xennet_set_sg will force-disable it even if it fails
>with -ENOSYS.  This is fine because it may only happen when
>xennet_set_{sg,tso} are called internally by xennet_set_features.
>
>Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
>---
> drivers/net/xen-netfront.c |   35 +++++++++++++++++++++--------------
> 1 files changed, 21 insertions(+), 14 deletions(-)
>
>diff --git a/drivers/net/xen-netfront.c b/drivers/net/xen-netfront.c
>index 4d839e4..15d0f79 100644
>--- a/drivers/net/xen-netfront.c
>+++ b/drivers/net/xen-netfront.c
>@@ -1526,35 +1526,42 @@ again:
> 
> static int xennet_set_sg(struct net_device *dev, u32 data)
> {
>+	int val, rc;
>+
> 	if (data) {
> 		struct netfront_info *np = netdev_priv(dev);
>-		int val;
>-
> 		if (xenbus_scanf(XBT_NIL, np->xbdev->otherend, "feature-sg",
> 				 "%d", &val) < 0)
> 			val = 0;
>-		if (!val)
>-			return -ENOSYS;
>-	} else if (dev->mtu > ETH_DATA_LEN)
>-		dev->mtu = ETH_DATA_LEN;
>-
>-	return ethtool_op_set_sg(dev, data);
>+	} else
>+		val = 0;
>+
>+	rc = ethtool_op_set_sg(dev, val);
>+	if (rc == 0 && !val) {
>+		if (dev->mtu > ETH_DATA_LEN)
>+			dev->mtu = ETH_DATA_LEN;
>+		if (data)
>+			rc = -ENOSYS;
>+	}
>+	return rc;
> }
> 
> static int xennet_set_tso(struct net_device *dev, u32 data)
> {
>+	int val, rc;
>+
> 	if (data) {
> 		struct netfront_info *np = netdev_priv(dev);
>-		int val;
>-
> 		if (xenbus_scanf(XBT_NIL, np->xbdev->otherend,
> 				 "feature-gso-tcpv4", "%d", &val) < 0)
> 			val = 0;
>-		if (!val)
>-			return -ENOSYS;
>-	}
>+	} else
>+		val = 0;
> 
>-	return ethtool_op_set_tso(dev, data);
>+	rc = ethtool_op_set_tso(dev, val);
>+	if (rc == 0 && !val && data)
>+		rc = -ENOSYS;
>+	return rc;
> }
> 
> static void xennet_set_features(struct net_device *dev)
>-- 
>1.7.6
>
>
>From 00fb21e9435005a01fdda65bd689e675d81523f6 Mon Sep 17 00:00:00 2001
>From: Paolo Bonzini <pbonzini@redhat.com>
>Date: Fri, 26 Aug 2011 14:03:42 +0200
>Subject: [RHEL6.3 PATCH 2/2] xen-netfront: do not temporarily disable s/g when negotiating features
>
>Following up on the previous patch, this one actually fixes the problem
>with jumbo frames after migration.  Now that xennet_set_{sg,tso} will
>force-disable unsupported offloading features, they can be called exactly
>once in xennet_set_features, without disabling s/g temporarily.
>
>xennet_create_dev will presume that all features are there, and
>if necessary, xennet_set_features will disable those that are absent.
>
>Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
>---
> drivers/net/xen-netfront.c |   29 ++++++++++++++++++-----------
> 1 files changed, 18 insertions(+), 11 deletions(-)
>
>diff --git a/drivers/net/xen-netfront.c b/drivers/net/xen-netfront.c
>index 15d0f79..b534ba6 100644
>--- a/drivers/net/xen-netfront.c
>+++ b/drivers/net/xen-netfront.c
>@@ -1179,7 +1179,9 @@ static struct net_device * __devinit xennet_create_dev(struct xenbus_device *dev
> 	netdev->netdev_ops	= &xennet_netdev_ops;
> 
> 	netif_napi_add(netdev, &np->napi, xennet_poll, 64);
>-	netdev->features        = NETIF_F_IP_CSUM;
>+
>+	/* Assume all features and let xennet_set_features fix up.  */
>+	netdev->features        = NETIF_F_IP_CSUM | NETIF_F_SG | NETIF_F_TSO;
> 
> 	SET_ETHTOOL_OPS(netdev, &xennet_ethtool_ops);
> 	SET_NETDEV_DEV(netdev, &dev->dev);
>@@ -1566,17 +1568,22 @@ static int xennet_set_tso(struct net_device *dev, u32 data)
> 
> static void xennet_set_features(struct net_device *dev)
> {
>-	/* Turn off all GSO bits except ROBUST. */
>-	dev->features &= ~NETIF_F_GSO_MASK;
>-	dev->features |= NETIF_F_GSO_ROBUST;
>-	xennet_set_sg(dev, 0);
>-
>-	/* We need checksum offload to enable scatter/gather and TSO. */
>-	if (!(dev->features & NETIF_F_IP_CSUM))
>-		return;
>+	/* Set ROBUST, turn off all other GSO bits except TSO. */
>+	dev->features =
>+		(dev->features & NETIF_F_TSO) |
>+		(dev->features & ~NETIF_F_GSO_MASK) |
>+		NETIF_F_GSO_ROBUST;
> 
>-	if (!xennet_set_sg(dev, 1))
>-		xennet_set_tso(dev, 1);
>+	/*
>+	 * We need checksum offload to enable scatter/gather, and
>+	 * scatter/gather to enable TSO.  Calling xennet_set_sg and
>+	 * xennet_set_tso ensures that Xenstore is probed for feature
>+	 * support in the backend.
>+	 */
>+	xennet_set_sg(dev, ((dev->features & (NETIF_F_IP_CSUM | NETIF_F_SG)) ==
>+			    (NETIF_F_IP_CSUM | NETIF_F_SG)));
>+	xennet_set_tso(dev, ((dev->features & (NETIF_F_SG | NETIF_F_TSO)) ==
>+			     (NETIF_F_SG | NETIF_F_TSO)));
> }
> 
> static int xennet_connect(struct net_device *dev)
>-- 
>1.7.6
>

Actions: View | Diff

Attachments on bug 733651: 520373