bonding fixes
Subsystems (3)
cover bonding fixes
I've been investigating a bonding bug that results in skb_under_panic() on older kernels. The bug is still present but with less disastrous results. In the course of my investigation I found two regressions without actually looking very hard, which is somewhat worrying.
The fixes are:
1. Remove redundant VLAN tag insertion logic. Fixes a regression due to VLAN handling changes in 2.6.37, which corrupts *all* outgoing traffic on a VLAN over a bond when the slave does not implement VLAN tag insertion.
2. Change active slave quietly when bond is down. Fixes a crash introduced by 5a37e8ca8536c47871d46c82211f399adf06fd44 (bonding: rejoin multicast groups on VLANs) and also cleans up existing code.
3. Fix mangled NAs on slaves without VLAN tag insertion. This bug appears to have been present ever since gratuitous NAs were implemented in 2.6.29. It doesn't look possible to fix without relying on the recent VLAN handling changes. For stable/longterm and backports, I think the safest thing to do is not to send a gratuitous NA if it would end up mangled.
Ben.
Ben Hutchings (3):
bonding/vlan: Remove redundant VLAN tag insertion logic
bonding: Change active slave quietly when bond is down bonding/vlan: Fix mangled NAs on slaves without VLAN tag insertion
drivers/net/bonding/bond_ipv6.c | 7 +++++- drivers/net/bonding/bond_main.c | 42 +++++++++----------------------------- 2 files changed, 16 insertions(+), 33 deletions(-)
Signature · 7 lines
1.7.3.2 -- Ben Hutchings, Senior Software Engineer, Solarflare Communications Not speaking for my employer; that's the marketing department's job. They asked us to note that Solarflare product names are trademarked.
These should of course go to net-2.6, not net-next-2.6.
Ben.
Signature · 3 lines
Ben Hutchings, Senior Software Engineer, Solarflare Communications Not speaking for my employer; that's the marketing department's job. They asked us to note that Solarflare product names are trademarked.
David Miller
maintainerFrom: Ben Hutchings <[email protected]>
Date: Mon, 13 Dec 2010 18:18:03 +0000
Quoted 1 line
These should of course go to net-2.6, not net-next-2.6.
I'll apply these, thanks a lot Ben.
PATCH 1/3 bonding/vlan: Remove redundant VLAN tag insertion logic
In reply to Ben HutchingsA bond may have a mixture of slave devices with and without hardware VLAN tag insertion capability. Therefore it always claims this capability and performs software VLAN tag insertion if the slave does not.
Since commit 7b9c60903714bf0a19d746b228864bad3497284e, this has also been done by dev_hard_start_xmit(). The result is that VLAN- tagged skbs are now double-tagged when transmitted through slave devices without hardware VLAN tag insertion!
Remove the now-redundant logic from bond_dev_queue_xmit().
--- drivers/net/bonding/bond_main.c | 27 +-------------------------- 1 files changed, 1 insertions(+), 26 deletions(-)
drivers/net/bonding/bond_main.c +1 -26
@@ -418,36 +418,11 @@ struct vlan_entry *bond_next_vlan(struct bonding *bond, struct vlan_entry *curr)* @bond: bond device that got this skb for tx. * @skb: hw accel VLAN tagged skb to transmit * @slave_dev: slave that is supposed to xmit this skbuff -* -* When the bond gets an skb to transmit that is -* already hardware accelerated VLAN tagged, and it -* needs to relay this skb to a slave that is not -* hw accel capable, the skb needs to be "unaccelerated", -* i.e. strip the hwaccel tag and re-insert it as part -* of the payload. */ int bond_dev_queue_xmit (struct bonding *bond,struct sk_buff *skb,struct net_device *slave_dev) { -unsigned short uninitialized_var (vlan_id ); - -/* Test vlan_list not vlgrp to catch and handle 802.1p tags */ -if (!list_empty (&bond->vlan_list) && - !(slave_dev->features &NETIF_F_HW_VLAN_TX ) && -vlan_get_tag (skb, &vlan_id) ==0 ) { - skb->dev = slave_dev; - skb =vlan_put_tag (skb, vlan_id); -if (!skb) { -/* vlan_put_tag() frees the skb in case of error, -* so return success here so the calling functions -* won't attempt to free is again. -*/ -return 0 ; - } - }else { - skb->dev = slave_dev; - } - + skb->dev = slave_dev; skb->priority =1 ;#ifdef CONFIG_NET_POLL_CONTROLLER if (unlikely (bond->dev->priv_flags &IFF_IN_NETPOLL )) {
Signature · 8 lines
1.7.3.2 -- Ben Hutchings, Senior Software Engineer, Solarflare Communications Not speaking for my employer; that's the marketing department's job. They asked us to note that Solarflare product names are trademarked.
Re: [PATCH net-2.6 1/3] bonding/vlan: Remove redundant VLAN tag insertion logic
Quoted 5 lines · Ben Hutchings
A bond may have a mixture of slave devices with and without hardware VLAN tag insertion capability. Therefore it always claims this capability and performs software VLAN tag insertion if the slave does not.
Quoted 8 lines
Since commit
7b9c60903714bf0a19d746b228864bad3497284e, this has also been done bydev_hard_start_xmit(). The result is that VLAN- tagged skbs are now double-tagged when transmitted through slave devices without hardware VLAN tag insertion!Remove the now-redundant logic from
bond_dev_queue_xmit().Signed-off-by Ben Hutchings <[email protected]>
Quoted 60 lines
---
drivers/net/bonding/bond_main.c| 27 +--------------------------1 files changed, 1 insertions(+), 26 deletions(-)drivers/net/bonding/bond_main.c +1
-26@@ -418,36 +418,11 @@ struct vlan_entry *bond_next_vlan(struct bonding *bond, struct vlan_entry *curr)* @bond: bond device that got this skb for tx. * @skb: hw accel VLAN tagged skb to transmit * @slave_dev: slave that is supposed to xmit this skbuff -* -* When the bond gets an skb to transmit that is -* already hardware accelerated VLAN tagged, and it -* needs to relay this skb to a slave that is not -* hw accel capable, the skb needs to be "unaccelerated", -* i.e. strip the hwaccel tag and re-insert it as part -* of the payload. */ int bond_dev_queue_xmit (struct bonding *bond,struct sk_buff *skb,struct net_device *slave_dev) { -unsigned short uninitialized_var (vlan_id ); - -/* Test vlan_list not vlgrp to catch and handle 802.1p tags */ -if (!list_empty (&bond->vlan_list) && - !(slave_dev->features &NETIF_F_HW_VLAN_TX ) && -vlan_get_tag (skb, &vlan_id) ==0 ) { - skb->dev = slave_dev; - skb =vlan_put_tag (skb, vlan_id); -if (!skb) { -/* vlan_put_tag() frees the skb in case of error, -* so return success here so the calling functions -* won't attempt to free is again. -*/ -return 0 ; - } - }else { - skb->dev = slave_dev; - } - + skb->dev = slave_dev; skb->priority =1 ;#ifdef CONFIG_NET_POLL_CONTROLLER if (unlikely (bond->dev->priv_flags &IFF_IN_NETPOLL )) {Signature · 13 lines
1.7.3.2 -- Ben Hutchings, Senior Software Engineer, Solarflare Communications Not speaking for my employer; that's the marketing department's job. They asked us to note that Solarflare product names are trademarked. -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
Jesse Gross
Reviewed-byRe: [PATCH net-2.6 1/3] bonding/vlan: Remove redundant VLAN tag insertion logic
Quoted 12 lines · Jay Vosburgh
Quoted 10 lines · Ben Hutchings
A bond may have a mixture of slave devices with and without hardware VLAN tag insertion capability. Therefore it always claims this capability and performs software VLAN tag insertion if the slave does not.
Since commit
7b9c60903714bf0a19d746b228864bad3497284e, this has also been done bydev_hard_start_xmit(). The result is that VLAN- tagged skbs are now double-tagged when transmitted through slave devices without hardware VLAN tag insertion!
Quoted 5 lines
Quoted 3 lines
Remove the now-redundant logic from
bond_dev_queue_xmit().Signed-off-by Ben Hutchings <[email protected]>Signed-off-by Jay Vosburgh <[email protected]>
Thanks for catching this. I had this change but it was grouped in with some other bonding/vlan simplifications that were dependent on the rest of the drivers being converted.
PATCH 2/3 bonding: Change active slave quietly when bond is down
In reply to Ben Hutchingsbond_change_active_slave() may be called when a slave is added, even if the bond has not been brought up yet. It may then attempt to send packets, and further it may use mcast_work which is uninitialised before the bond is brought up. Add the necessary checks for netif_running(bond->dev).
--- drivers/net/bonding/bond_main.c | 15 +++++++++------ 1 files changed, 9 insertions(+), 6 deletions(-)
drivers/net/bonding/bond_main.c +9 -6
@@ -1178,11 +1178,13 @@ void bond_change_active_slave(struct bonding *bond, struct slave *new_active)bond_do_fail_over_mac (bond, new_active, old_active); - bond->send_grat_arp = bond->params.num_grat_arp; -bond_send_gratuitous_arp (bond); +if (netif_running (bond->dev)) { + bond->send_grat_arp = bond->params.num_grat_arp; +bond_send_gratuitous_arp (bond); - bond->send_unsol_na = bond->params.num_unsol_na; -bond_send_unsolicited_na (bond); + bond->send_unsol_na = bond->params.num_unsol_na; +bond_send_unsolicited_na (bond); + }write_unlock_bh (&bond->curr_slave_lock);read_unlock (&bond->lock);
@@ -1196,8 +1198,9 @@ void bond_change_active_slave(struct bonding *bond, struct slave *new_active)/* resend IGMP joins since active slave has changed or * all were sent on curr_active_slave */ -if ((USES_PRIMARY (bond->params.mode) && new_active) || - bond->params.mode ==BOND_MODE_ROUNDROBIN ) { +if (((USES_PRIMARY (bond->params.mode) && new_active) || + bond->params.mode ==BOND_MODE_ROUNDROBIN ) && +netif_running (bond->dev)) { bond->igmp_retrans = bond->params.resend_igmp;queue_delayed_work (bond->wq, &bond->mcast_work,0 ); }
Signature · 8 lines
1.7.3.2 -- Ben Hutchings, Senior Software Engineer, Solarflare Communications Not speaking for my employer; that's the marketing department's job. They asked us to note that Solarflare product names are trademarked.
Re: [PATCH net-2.6 2/3] bonding: Change active slave quietly when bond is down
Quoted 5 lines · Ben Hutchings
bond_change_active_slave()may be called when a slave is added, even if the bond has not been brought up yet. It may then attempt to send packets, and further it may usemcast_workwhich is uninitialised before the bond is brought up. Add the necessary checks fornetif_running(bond->dev).
I wondered if it would be better to instead arrange forbond_change_active_slave (and bond_select_active_slave, which calls it) to not be called when the bond is down, but that would require quite a bit of change to bond_enslave and bond_open.
-J
Quoted 53 lines
Signed-off-by Ben Hutchings <[email protected]>---
drivers/net/bonding/bond_main.c| 15 +++++++++------1 files changed, 9 insertions(+), 6 deletions(-)drivers/net/bonding/bond_main.c +9
-6@@ -1178,11 +1178,13 @@ void bond_change_active_slave(struct bonding *bond, struct slave *new_active)bond_do_fail_over_mac (bond, new_active, old_active); - bond->send_grat_arp = bond->params.num_grat_arp; -bond_send_gratuitous_arp (bond); +if (netif_running (bond->dev)) { + bond->send_grat_arp = bond->params.num_grat_arp; +bond_send_gratuitous_arp (bond); - bond->send_unsol_na = bond->params.num_unsol_na; -bond_send_unsolicited_na (bond); + bond->send_unsol_na = bond->params.num_unsol_na; +bond_send_unsolicited_na (bond); + }write_unlock_bh (&bond->curr_slave_lock);read_unlock (&bond->lock);@@ -1196,8 +1198,9 @@ void bond_change_active_slave(struct bonding *bond, struct slave *new_active)/* resend IGMP joins since active slave has changed or * all were sent on curr_active_slave */ -if ((USES_PRIMARY (bond->params.mode) && new_active) || - bond->params.mode ==BOND_MODE_ROUNDROBIN ) { +if (((USES_PRIMARY (bond->params.mode) && new_active) || + bond->params.mode ==BOND_MODE_ROUNDROBIN ) && +netif_running (bond->dev)) { bond->igmp_retrans = bond->params.resend_igmp;queue_delayed_work (bond->wq, &bond->mcast_work,0 ); }Signature · 13 lines
1.7.3.2 -- Ben Hutchings, Senior Software Engineer, Solarflare Communications Not speaking for my employer; that's the marketing department's job. They asked us to note that Solarflare product names are trademarked. -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [PATCH net-2.6 2/3] bonding: Change active slave quietly when bond is down
Quoted 8 lines · Jay Vosburgh
Quoted 5 lines · Ben Hutchings
bond_change_active_slave()may be called when a slave is added, even if the bond has not been brought up yet. It may then attempt to send packets, and further it may usemcast_workwhich is uninitialised before the bond is brought up. Add the necessary checks fornetif_running(bond->dev).
Quoted 4 lines
I wondered if it would be better to instead arrange for
bond_change_active_slave(andbond_select_active_slave, which calls it) to not be called when the bond is down, but that would require quite a bit of change tobond_enslaveandbond_open.
[...]
Yes, it did seem to me that it would be better to defer all these actions until bond_open, but I thought that would be too risky at this stage in the release cycle. Please do clean this up further in net-next-2.6.
Ben.
Signature · 3 lines
Ben Hutchings, Senior Software Engineer, Solarflare Communications Not speaking for my employer; that's the marketing department's job. They asked us to note that Solarflare product names are trademarked.
Re: [PATCH net-2.6 2/3] bonding: Change active slave quietly when bond is down
In reply to Ben HutchingsQuoted 47 lines · Ben Hutchings
bond_change_active_slave()may be called when a slave is added, even if the bond has not been brought up yet. It may then attempt to send packets, and further it may usemcast_workwhich is uninitialised before the bond is brought up. Add the necessary checks fornetif_running(bond->dev).Signed-off-by Ben Hutchings <[email protected]>--- drivers/net/bonding/bond_main.c | 15 +++++++++------ 1 files changed, 9 insertions(+), 6 deletions(-)drivers/net/bonding/bond_main.c +9
-6@@ -1178,11 +1178,13 @@ void bond_change_active_slave(struct bonding *bond, struct slave *new_active)bond_do_fail_over_mac (bond, new_active, old_active); - bond->send_grat_arp = bond->params.num_grat_arp; -bond_send_gratuitous_arp (bond); +if (netif_running (bond->dev)) { + bond->send_grat_arp = bond->params.num_grat_arp; +bond_send_gratuitous_arp (bond); - bond->send_unsol_na = bond->params.num_unsol_na; -bond_send_unsolicited_na (bond); + bond->send_unsol_na = bond->params.num_unsol_na; +bond_send_unsolicited_na (bond); + }write_unlock_bh (&bond->curr_slave_lock);read_unlock (&bond->lock);@@ -1196,8 +1198,9 @@ void bond_change_active_slave(struct bonding *bond, struct slave *new_active)/* resend IGMP joins since active slave has changed or * all were sent on curr_active_slave */ -if ((USES_PRIMARY (bond->params.mode) && new_active) || - bond->params.mode ==BOND_MODE_ROUNDROBIN ) { +if (((USES_PRIMARY (bond->params.mode) && new_active) || + bond->params.mode ==BOND_MODE_ROUNDROBIN ) && +netif_running (bond->dev)) { bond->igmp_retrans = bond->params.resend_igmp;queue_delayed_work (bond->wq, &bond->mcast_work,0 ); }Signature · 1 line
1.7.3.2
I had a script randomly adding/removing slaves and setting the bonding device up and down to test that commit, so I'm surprised that it didn't catch this combination before.
Thanks,
Signature · 1 line
Flavio
PATCH 3/3 bonding/vlan: Fix mangled NAs on slaves without VLAN tag insertion
In reply to Ben Hutchingsbond_na_send() attempts to insert a VLAN tag in between building and sending packets of the respective formats. If the slave does not implement hardware VLAN tag insertion then vlan_put_tag() will mangle the network-layer header because the Ethernet header is not present at this point (unlike in bond_arp_send()).
Fix this by adding the tag out-of-line and relying on dev_hard_start_xmit() to insert it inline if necessary.
--- drivers/net/bonding/bond_ipv6.c | 7 ++++++- 1 files changed, 6 insertions(+), 1 deletions(-)
drivers/net/bonding/bond_ipv6.c +6 -1
@@ -88,7 +88,12 @@ static void bond_na_send(struct net_device *slave_dev, }if (vlan_id) { - skb =vlan_put_tag (skb, vlan_id); +/* The Ethernet header is not present yet, so it is +* too early to insert a VLAN tag. Force use of an +* out-of-line tag here and let dev_hard_start_xmit() +* insert it if the slave hardware can't. +*/ + skb =__vlan_hwaccel_put_tag (skb, vlan_id);if (!skb) {pr_err ("failed to insert VLAN tag\n" );return ;
Signature · 7 lines
1.7.3.2 -- Ben Hutchings, Senior Software Engineer, Solarflare Communications Not speaking for my employer; that's the marketing department's job. They asked us to note that Solarflare product names are trademarked.
Re: [PATCH net-2.6 3/3] bonding/vlan: Fix mangled NAs on slaves without VLAN tag insertion
Quoted 10 lines · Ben Hutchings
bond_na_send()attempts to insert a VLAN tag in between building and sending packets of the respective formats. If the slave does not implement hardware VLAN tag insertion thenvlan_put_tag()will mangle the network-layer header because the Ethernet header is not present at this point (unlike inbond_arp_send()).Fix this by adding the tag out-of-line and relying on
dev_hard_start_xmit()to insert it inline if necessary.Signed-off-by Ben Hutchings <[email protected]>
Quoted 31 lines
---
drivers/net/bonding/bond_ipv6.c| 7 ++++++-1 files changed, 6 insertions(+), 1 deletions(-)drivers/net/bonding/bond_ipv6.c +6
-1@@ -88,7 +88,12 @@ static void bond_na_send(struct net_device *slave_dev, }if (vlan_id) { - skb =vlan_put_tag (skb, vlan_id); +/* The Ethernet header is not present yet, so it is +* too early to insert a VLAN tag. Force use of an +* out-of-line tag here and let dev_hard_start_xmit() +* insert it if the slave hardware can't. +*/ + skb =__vlan_hwaccel_put_tag (skb, vlan_id);if (!skb) {pr_err ("failed to insert VLAN tag\n" );return ;Signature · 7 lines
1.7.3.2 -- Ben Hutchings, Senior Software Engineer, Solarflare Communications Not speaking for my employer; that's the marketing department's job. They asked us to note that Solarflare product names are trademarked.
Re: [PATCH net-2.6 3/3] bonding/vlan: Fix mangled NAs on slaves without VLAN tag insertion
In reply to Ben HutchingsQuoted 22 lines · Ben Hutchings
bond_na_send()attempts to insert a VLAN tag in between building and sending packets of the respective formats. If the slave does not implement hardware VLAN tag insertion thenvlan_put_tag()will mangle the network-layer header because the Ethernet header is not present at this point (unlike inbond_arp_send()).Fix this by adding the tag out-of-line and relying on
dev_hard_start_xmit()to insert it inline if necessary.Signed-off-by Ben Hutchings <[email protected]>---
drivers/net/bonding/bond_ipv6.c | 7 ++++++-1 files changed, 6 insertions(+), 1 deletions(-)drivers/net/bonding/bond_ipv6.c +0
-0@@ -88,7 +88,12 @@ static void bond_na_send(struct net_device *slave_dev, } if (vlan_id) {
drivers/net/bonding/bond_ipv6.c static void bond_na_send(struct net_device *slave_dev, · line 99
Quoted 10 lines
drivers/net/bonding/bond_ipv6.c +6
-1- skb = vlan_put_tag(skb, vlan_id); + /* The Ethernet header is not present yet, so it is + * too early to insert a VLAN tag. Force use of an + * out-of-line tag here and let dev_hard_start_xmit() + * insert it if the slave hardware can't. + */ + skb = __vlan_hwaccel_put_tag(skb, vlan_id); if (!skb) { pr_err("failed to insert VLAN tag\n"); return;
You could drop the check for !skb, since __vlan_hwaccel_put_tag() can never fail.
Actually, my plan is to remove all current users of vlan_put_tag() and replace them with __vlan_hwaccel_put_tag (and then rename that to vlan_put_tag()). There's really no reason why the upper layers should have to care what the underlying device is capable of, now that dev_hard_start_xmit() can handle it.