| Message ID | 20190731224955.10908-1-nikolay@cumulusnetworks.com |
|---|---|
| State | Superseded |
| Delegated to: | David Miller |
| Headers | show |
| Series | [net,v3] net: bridge: move vlan init/deinit to NETDEV_REGISTER/UNREGISTER | expand |
> -int br_vlan_init(struct net_bridge *br) > +static int br_vlan_init(struct net_bridge *br) > { > struct net_bridge_vlan_group *vg; > int ret = -ENOMEM; > @@ -1083,6 +1085,8 @@ int br_vlan_init(struct net_bridge *br) > return ret; > > err_vlan_add: > + RCU_INIT_POINTER(br->vlgrp, NULL); > + synchronize_rcu(); Calling sychronize_rcu is expensive. And the callback for notifier is always called with rtnl_head. Why not just keep the pointer initialization back in the code where bridge is created, it was safe there.
On 8/1/19 1:53 AM, Stephen Hemminger wrote: > >> -int br_vlan_init(struct net_bridge *br) >> +static int br_vlan_init(struct net_bridge *br) >> { >> struct net_bridge_vlan_group *vg; >> int ret = -ENOMEM; >> @@ -1083,6 +1085,8 @@ int br_vlan_init(struct net_bridge *br) >> return ret; >> >> err_vlan_add: >> + RCU_INIT_POINTER(br->vlgrp, NULL); >> + synchronize_rcu(); > > Calling sychronize_rcu is expensive. And the callback for > notifier is always called with rtnl_head. > > Why not just keep the pointer initialization back in the > code where bridge is created, it was safe there. > Because now the device registered and we've published the group, right now it is not an issue but if we expose an rcu helper we'll have to fix this because it'd become a bug. I'd prefer to have the error path correct and future-proof it, since it's an error path we're not concerned with speed, but rather correctness. Also these are rarely exercised so the bug might remain for a very long time.
On 8/1/19 2:32 AM, Nikolay Aleksandrov wrote: > On 8/1/19 1:53 AM, Stephen Hemminger wrote: >> >>> -int br_vlan_init(struct net_bridge *br) >>> +static int br_vlan_init(struct net_bridge *br) >>> { >>> struct net_bridge_vlan_group *vg; >>> int ret = -ENOMEM; >>> @@ -1083,6 +1085,8 @@ int br_vlan_init(struct net_bridge *br) >>> return ret; >>> >>> err_vlan_add: >>> + RCU_INIT_POINTER(br->vlgrp, NULL); >>> + synchronize_rcu(); >> >> Calling sychronize_rcu is expensive. And the callback for >> notifier is always called with rtnl_head. >> >> Why not just keep the pointer initialization back in the >> code where bridge is created, it was safe there. >> > > Because now the device registered and we've published the group, right now > it is not an issue but if we expose an rcu helper we'll have to fix this > because it'd become a bug. > I'd prefer to have the error path correct and future-proof it, since it's > an error path we're not concerned with speed, but rather correctness. Also > these are rarely exercised so the bug might remain for a very long time. > > About why move it - I've explained in the commit message, it might be safe but it has presented a lot of bugs, we'll have to separate it in two parts one that initializes the vlan group and second one which adds the fdb, then we'll have to split the flush/delete in two different places to cleanup. This way we have only a single exit point that cleans up and it works for all cases. The synchronize there wouldn't be called under normal circumstances.
On 8/1/19 1:49 AM, Nikolay Aleksandrov wrote: [snip] > [0] https://bugzilla.kernel.org/show_bug.cgi?id=204389 > > Reported-by: michael-dev <michael-dev@fami-braun.de> > Fixes: 5be5a2df40f0 ("bridge: Add filtering support for default_pvid") > Signed-off-by: Nikolay Aleksandrov <nikolay@cumulusnetworks.com> > --- > I tried a few different approaches to resolve this but they were all > unsuitable for some kernels, this approach can go to stables easily > and IMO is the way this had to be done from the start. Alternatively > we could move only the br_vlan_add and pair it with a br_vlan_del of > default_pvid on the same events, but I don't think it hurts to move > the whole init/deinit there as it'd help older stable releases as well. > > I also tested the br_vlan_init error handling after the move by always > returning errors from all over it. Since errors at NETDEV_REGISTER cause > NETDEV_UNREGISTER we can deinit vlans properly for all cases regardless > why it happened (e.g. device destruction or init error). > > net/bridge/br.c | 5 ++++- > net/bridge/br_device.c | 10 ---------- > net/bridge/br_private.h | 20 +++++--------------- > net/bridge/br_vlan.c | 25 ++++++++++++++++++------- > 4 files changed, 27 insertions(+), 33 deletions(-) > Self-NAK, after thinking more about how to best handle this and running more tests I believe it'll be better to go with the alternative I suggested above - to move out only the default pvid add out of br_vlan_init to NETDEV_REGSITER and pair it with br_vlan_delete in NETDEV_UNREGISTER. That way we'll split the init/deinit in 2 steps, but we'll keep the current order and will reduce the churn for this fix, functionally it should be equivalent as that is the problematic part of the init which has to be done after the netdev has been registered. I'll spin v4 tomorrow after running more tests with it.
diff --git a/net/bridge/br.c b/net/bridge/br.c index d164f63a4345..8a8f9e5f264f 100644 --- a/net/bridge/br.c +++ b/net/bridge/br.c @@ -37,12 +37,15 @@ static int br_device_event(struct notifier_block *unused, unsigned long event, v int err; if (dev->priv_flags & IFF_EBRIDGE) { + err = br_vlan_bridge_event(dev, event, ptr); + if (err) + return notifier_from_errno(err); + if (event == NETDEV_REGISTER) { /* register of bridge completed, add sysfs entries */ br_sysfs_addbr(dev); return NOTIFY_DONE; } - br_vlan_bridge_event(dev, event, ptr); } /* not a port of a bridge */ diff --git a/net/bridge/br_device.c b/net/bridge/br_device.c index 681b72862c16..b3e3de2ecf95 100644 --- a/net/bridge/br_device.c +++ b/net/bridge/br_device.c @@ -135,18 +135,9 @@ static int br_dev_init(struct net_device *dev) return err; } - err = br_vlan_init(br); - if (err) { - free_percpu(br->stats); - br_mdb_hash_fini(br); - br_fdb_hash_fini(br); - return err; - } - err = br_multicast_init_stats(br); if (err) { free_percpu(br->stats); - br_vlan_flush(br); br_mdb_hash_fini(br); br_fdb_hash_fini(br); } @@ -161,7 +152,6 @@ static void br_dev_uninit(struct net_device *dev) br_multicast_dev_del(br); br_multicast_uninit_stats(br); - br_vlan_flush(br); br_mdb_hash_fini(br); br_fdb_hash_fini(br); free_percpu(br->stats); diff --git a/net/bridge/br_private.h b/net/bridge/br_private.h index e8cf03b43b7d..782f18c6b2a9 100644 --- a/net/bridge/br_private.h +++ b/net/bridge/br_private.h @@ -872,7 +872,6 @@ struct sk_buff *br_handle_vlan(struct net_bridge *br, int br_vlan_add(struct net_bridge *br, u16 vid, u16 flags, bool *changed, struct netlink_ext_ack *extack); int br_vlan_delete(struct net_bridge *br, u16 vid); -void br_vlan_flush(struct net_bridge *br); struct net_bridge_vlan *br_vlan_find(struct net_bridge_vlan_group *vg, u16 vid); void br_recalculate_fwd_mask(struct net_bridge *br); int __br_vlan_filter_toggle(struct net_bridge *br, unsigned long val); @@ -881,7 +880,6 @@ int __br_vlan_set_proto(struct net_bridge *br, __be16 proto); int br_vlan_set_proto(struct net_bridge *br, unsigned long val); int br_vlan_set_stats(struct net_bridge *br, unsigned long val); int br_vlan_set_stats_per_port(struct net_bridge *br, unsigned long val); -int br_vlan_init(struct net_bridge *br); int br_vlan_set_default_pvid(struct net_bridge *br, unsigned long val); int __br_vlan_set_default_pvid(struct net_bridge *br, u16 pvid, struct netlink_ext_ack *extack); @@ -894,8 +892,8 @@ int nbp_get_num_vlan_infos(struct net_bridge_port *p, u32 filter_mask); void br_vlan_get_stats(const struct net_bridge_vlan *v, struct br_vlan_stats *stats); void br_vlan_port_event(struct net_bridge_port *p, unsigned long event); -void br_vlan_bridge_event(struct net_device *dev, unsigned long event, - void *ptr); +int br_vlan_bridge_event(struct net_device *dev, unsigned long event, + void *ptr); static inline struct net_bridge_vlan_group *br_vlan_group( const struct net_bridge *br) @@ -988,19 +986,10 @@ static inline int br_vlan_delete(struct net_bridge *br, u16 vid) return -EOPNOTSUPP; } -static inline void br_vlan_flush(struct net_bridge *br) -{ -} - static inline void br_recalculate_fwd_mask(struct net_bridge *br) { } -static inline int br_vlan_init(struct net_bridge *br) -{ - return 0; -} - static inline int nbp_vlan_add(struct net_bridge_port *port, u16 vid, u16 flags, bool *changed, struct netlink_ext_ack *extack) { @@ -1085,9 +1074,10 @@ static inline void br_vlan_port_event(struct net_bridge_port *p, { } -static inline void br_vlan_bridge_event(struct net_device *dev, - unsigned long event, void *ptr) +static inline int br_vlan_bridge_event(struct net_device *dev, + unsigned long event, void *ptr) { + return 0; } #endif diff --git a/net/bridge/br_vlan.c b/net/bridge/br_vlan.c index a544e161c7fa..6a17408b4eb8 100644 --- a/net/bridge/br_vlan.c +++ b/net/bridge/br_vlan.c @@ -709,7 +709,7 @@ int br_vlan_delete(struct net_bridge *br, u16 vid) return __vlan_del(v); } -void br_vlan_flush(struct net_bridge *br) +static void br_vlan_flush(struct net_bridge *br) { struct net_bridge_vlan_group *vg; @@ -721,6 +721,8 @@ void br_vlan_flush(struct net_bridge *br) br_fdb_delete_by_port(br, NULL, 0, 1); vg = br_vlan_group(br); + if (!vg) + return; __vlan_flush(vg); RCU_INIT_POINTER(br->vlgrp, NULL); synchronize_rcu(); @@ -1054,7 +1056,7 @@ int br_vlan_set_default_pvid(struct net_bridge *br, unsigned long val) return err; } -int br_vlan_init(struct net_bridge *br) +static int br_vlan_init(struct net_bridge *br) { struct net_bridge_vlan_group *vg; int ret = -ENOMEM; @@ -1083,6 +1085,8 @@ int br_vlan_init(struct net_bridge *br) return ret; err_vlan_add: + RCU_INIT_POINTER(br->vlgrp, NULL); + synchronize_rcu(); vlan_tunnel_deinit(vg); err_tunnel_init: rhashtable_destroy(&vg->vlan_hash); @@ -1469,13 +1473,19 @@ static void nbp_vlan_set_vlan_dev_state(struct net_bridge_port *p, u16 vid) } /* Must be protected by RTNL. */ -void br_vlan_bridge_event(struct net_device *dev, unsigned long event, - void *ptr) +int br_vlan_bridge_event(struct net_device *dev, unsigned long event, void *ptr) { struct netdev_notifier_changeupper_info *info; - struct net_bridge *br; + struct net_bridge *br = netdev_priv(dev); + int ret = 0; switch (event) { + case NETDEV_REGISTER: + ret = br_vlan_init(br); + break; + case NETDEV_UNREGISTER: + br_vlan_flush(br); + break; case NETDEV_CHANGEUPPER: info = ptr; br_vlan_upper_change(dev, info->upper_dev, info->linking); @@ -1483,12 +1493,13 @@ void br_vlan_bridge_event(struct net_device *dev, unsigned long event, case NETDEV_CHANGE: case NETDEV_UP: - br = netdev_priv(dev); if (!br_opt_get(br, BROPT_VLAN_BRIDGE_BINDING)) - return; + break; br_vlan_link_state_change(dev, br); break; } + + return ret; } /* Must be protected by RTNL. */
Most of the bridge device's vlan init bugs come from the fact that it's done in the wrong place, way too early in ndo_init() before the device is even assigned an ifindex. That makes error handling harder, especially for older kernels which don't have bridge ndo_uninit callback. It also introduces another bug when the bridge's dev_addr is added as fdb in the the initial default pvid on vlan initialization, the fdb notification has ifindex/NDA_MASTER both equal to 0 (see example below) which really makes no sense for user-space[0]. Usually user-space software would ignore such entries, but they are actually valid and will eventually have all necessary attributes. I chose to change the order because this can be backported to all kernels even pre-ndo_uninit ones without many changes and it keeps init/deinit symmetric. As a bonus this allows us to keep the vlan init/deinit entirely in br_vlan.c and remove those exports. It makes much more sense to send a notification *after* the device has registered and has a proper ifindex allocated rather than before when there's a chance that the registration might still fail. For the demonstration below a small change to iproute2 for printing all fdb notifications is added, because it contained a workaround not to show entries with ifindex == 0. Command executed while monitoring: $ ip l add br0 type bridge Before (both ifindex and master == 0): $ bridge monitor fdb 36:7e:8a:b3:56:ba dev * vlan 1 master * permanent After (proper br0 ifindex): $ bridge monitor fdb e6:2a:ae:7a:b7:48 dev br0 vlan 1 master br0 permanent v3: send the correct v2 patch with all changes (stub should return 0) v2: on error in br_vlan_init set br->vlgrp to NULL and return 0 in the br_vlan_bridge_event stub when bridge vlans are disabled [0] https://bugzilla.kernel.org/show_bug.cgi?id=204389 Reported-by: michael-dev <michael-dev@fami-braun.de> Fixes: 5be5a2df40f0 ("bridge: Add filtering support for default_pvid") Signed-off-by: Nikolay Aleksandrov <nikolay@cumulusnetworks.com> --- I tried a few different approaches to resolve this but they were all unsuitable for some kernels, this approach can go to stables easily and IMO is the way this had to be done from the start. Alternatively we could move only the br_vlan_add and pair it with a br_vlan_del of default_pvid on the same events, but I don't think it hurts to move the whole init/deinit there as it'd help older stable releases as well. I also tested the br_vlan_init error handling after the move by always returning errors from all over it. Since errors at NETDEV_REGISTER cause NETDEV_UNREGISTER we can deinit vlans properly for all cases regardless why it happened (e.g. device destruction or init error). net/bridge/br.c | 5 ++++- net/bridge/br_device.c | 10 ---------- net/bridge/br_private.h | 20 +++++--------------- net/bridge/br_vlan.c | 25 ++++++++++++++++++------- 4 files changed, 27 insertions(+), 33 deletions(-)