veth: don't free priv->status until dev->destructor (v2)
Subsystems (2)
PATCH 1/1 veth: don't free priv->status until dev->destructor (v2)
Based on the commit msg on ae0e8e82205c903978a79ebf5e31c670b61fa5b4, it looks like oopses were caused when people were reading the veth dev stats while the module was being unloaded, causing a deref of freed memory in veth_get_stats()? If so, I believe the following patch (still against mainline, so not on top of my previous patch or on top of a git-revert of ae0e8e82205c903978a79ebf5e31c670b61fa5b)) should prevent that. All the stats are gathered within one rcu cycle, while the device free hook first sets the device stats struct to NULL, waits an rcu cycle before freeing it.
I haven't been able to reproduce the original oops though (been trying to cat the stats sysfs files while rmmoding veth, to no avail, and haven't found an original bug report or testcase), so can't verify whether this patch prevents the original oops.
Does this look sufficient?
thanks,
-serge
From a8eb0950b47ff6c5dfe2debafbd203dcced75bd3 Mon Sep 17 00:00:00 2001
From: root <[email protected]>
Date: Wed, 24 Jun 2009 20:26:17 -0700
Subject: [PATCH 1/1] veth: don't free priv->status until dev->destructor (v2)
Since commit ae0e8e82205c903978a79ebf5e31c670b61fa5b4, priv->status has been freed at veth_close(). But that causes a NULL deref at veth_xmit. This patch moves priv->status free back to the device destructor. It also tries to prevent the original possible sysfs-induced oops. All the stats are now gathered within one rcu cycle, while the device free hook first sets the device stats struct to NULL, waits an rcu cycle before freeing
it.
Changelog:
June 26: try to fix the original oops.
--- drivers/net/veth.c | 22 ++++++++++++++++++---- 1 files changed, 18 insertions(+), 4 deletions(-)
drivers/net/veth.c +18 -4
@@ -208,7 +208,7 @@ rx_drop:static struct net_device_stats *veth_get_stats (struct net_device *dev) { -struct veth_priv *priv =netdev_priv (dev); +struct veth_priv *priv;struct net_device_stats *dev_stats = &dev->stats;unsigned int cpu;struct veth_net_stats *stats;
@@ -220,6 +220,8 @@ static struct net_device_stats *veth_get_stats(struct net_device *dev) dev_stats->tx_dropped =0 ; dev_stats->rx_dropped =0 ; +rcu_read_lock (); + priv =netdev_priv (dev);if (priv->stats)for_each_online_cpu (cpu) { stats =per_cpu_ptr (priv->stats, cpu);
@@ -231,6 +233,7 @@ static struct net_device_stats *veth_get_stats(struct net_device *dev) dev_stats->tx_dropped += stats->tx_dropped; dev_stats->rx_dropped += stats->rx_dropped; } +rcu_read_unlock ();return dev_stats; }
@@ -257,8 +260,6 @@ static int veth_close(struct net_device *dev)netif_carrier_off (dev);netif_carrier_off (priv->peer); -free_percpu (priv->stats); - priv->stats =NULL ;return 0 ; }
@@ -299,6 +300,19 @@ static const struct net_device_ops veth_netdev_ops = { .ndo_set_mac_address = eth_mac_addr, }; +static void veth_dev_free (struct net_device *dev) +{ +struct veth_priv *priv; +struct veth_net_stats *stats; + + priv =netdev_priv (dev); + stats = priv->stats; + priv->stats =NULL ; +synchronize_rcu (); +free_percpu (stats); +free_netdev (dev); +} +static void veth_setup (struct net_device *dev) {ether_setup (dev);
@@ -306,7 +320,7 @@ static void veth_setup(struct net_device *dev) dev->netdev_ops = &veth_netdev_ops; dev->ethtool_ops = &veth_ethtool_ops; dev->features |=NETIF_F_LLTX ; - dev->destructor = free_netdev; + dev->destructor = veth_dev_free; }/*
Signature · 1 line
1.6.2.3
David Miller
maintainerFrom: "Serge E. Hallyn" <[email protected]>
Date: Fri, 26 Jun 2009 11:24:18 -0500
Quoted 4 lines
I haven't been able to reproduce the original oops though (been trying to cat the stats sysfs files while rmmoding veth, to no avail, and haven't found an original bug report or testcase), so can't verify whether this patch prevents the original oops.
If you 'cat' it you're unlikely to trigger the oops.
You have to hold the sysfs files open, and then elsewhere do the rmmod, wait, and then continue with some access to those open sysfs file descriptors (f.e. do some reads).
I'd also need this patch to be against current sources as they'll never apply since I did the revert quite some time ago.
Thanks.
Quoted 10 lines · On Wed, 15 Jul 2009 08:50:12 -0700 (PDT) David Miller <[email protected]>…
From: "Serge E. Hallyn" <[email protected]>
Date: Fri, 26 Jun 2009 11:24:18 -0500Quoted 4 lines
I haven't been able to reproduce the original oops though (been trying to cat the stats sysfs files while rmmoding veth, to no avail, and haven't found an original bug report or testcase), so can't verify whether this patch prevents the original oops.
If you 'cat' it you're unlikely to trigger the oops.
Quoted 8 lines
You have to hold the sysfs files open, and then elsewhere do the rmmod, wait, and then continue with some access to those open sysfs file descriptors (f.e. do some reads).
I'd also need this patch to be against current sources as they'll never apply since I did the revert quite some time ago.
Thanks.
My usual way of doing this is:
# (sleep 30; cat /sys/class/net/ethX/statistics/tx_bytes) &
# rmmod the_buggy_driver
wait...
Quoted 11 lines · Stephen Hemminger
Quoted 8 lines · On Wed, 15 Jul 2009 08:50:12 -0700 (PDT) David Miller <[email protected]>…
From: "Serge E. Hallyn" <[email protected]>
Date: Fri, 26 Jun 2009 11:24:18 -0500Quoted 4 lines
I haven't been able to reproduce the original oops though (been trying to cat the stats sysfs files while rmmoding veth, to no avail, and haven't found an original bug report or testcase), so can't verify whether this patch prevents the original oops.
Quoted 5 lines
Quoted 5 lines
If you 'cat' it you're unlikely to trigger the oops.
You have to hold the sysfs files open, and then elsewhere do the rmmod, wait, and then continue with some access to those open sysfs file descriptors (f.e. do some reads).
Yup, I was doing that too, but couldn't reproduce as yet.
Quoted 4 lines
Quoted 4 lines
I'd also need this patch to be against current sources as they'll never apply since I did the revert quite some time ago.
Thanks.
Ok, thanks - I'll generate a new patch against a fresh pull when I can confirm that it actually solves the problem.
Quoted 6 lines
My usual way of doing this is:
# (sleep 30; cat
/sys/class/net/ethX/statistics/tx_bytes) &# rmmod the_buggy_driverwait...
Can you oops the kernel this way on latest netns?
thanks,
-serge
drivers/net/veth.c static void veth_setup(struct net_device *dev) · line 324
Quoted 10 lines · On Fri, 26 Jun 2009 11:24:18 -0500 "Serge E. Hallyn" <[email protected]> wrote:
Based on the commit msg on
ae0e8e82205c903978a79ebf5e31c670b61fa5b4, it looksdrivers/net/veth.c +1
-1ether_setup (dev);@@ -306,7 +320,7 @@ static void veth_setup(struct net_device *dev) dev->netdev_ops = &veth_netdev_ops; dev->ethtool_ops = &veth_ethtool_ops; dev->features |=NETIF_F_LLTX ; - dev->destructor = free_netdev; + dev->destructor = veth_dev_free;
This is still going to oops if sysfs statistics referenced after module unload because module is unloaded (code is gone) and veth_dev_free no longer exists.
I'll respin the original patch (using free_netdev) and fix the statistics complaint.
Quoted 5 lines · Stephen Hemminger
Quoted 1 line · On Fri, 26 Jun 2009 11:24:18 -0500 "Serge E. Hallyn" <[email protected]> wrote:
Based on the commit msg on
ae0e8e82205c903978a79ebf5e31c670b61fa5b4, it looks
Quoted 12 lines
drivers/net/veth.c static void veth_setup(struct net_device *dev) · line 324
Quoted 8 lines
drivers/net/veth.c +1
-1ether_setup (dev);@@ -306,7 +320,7 @@ static void veth_setup(struct net_device *dev) dev->netdev_ops = &veth_netdev_ops; dev->ethtool_ops = &veth_ethtool_ops; dev->features |=NETIF_F_LLTX ; - dev->destructor = free_netdev; + dev->destructor = veth_dev_free;This is still going to oops if sysfs statistics referenced after module unload because module is unloaded (code is gone) and
veth_dev_freeno longer exists.
Has anyone actually seen that cause an oops?
The reason I am asking is that as I read the code we cannot have this problem. At worst the destructor callback is delayed until:
veth_exit
rtnl_link_unregister
rtnl_unlock
netdev_run_todo
dev->destructorSimilarly even if the sysfs filehandle is open we have called:
netdev_unregister_kobject
...
sysfs_addrm_finish
sysfs_deactivate
Which guarantees that sysfs_get_active_two will fail and all subsequent actions on that file will fail.
Eric
Quoted 39 lines · On Tue, 04 Aug 2009 23:40:47 -0700 [email protected] (Eric W. Biederman)…
Quoted 5 lines · Stephen Hemminger
Quoted 1 line · On Fri, 26 Jun 2009 11:24:18 -0500 "Serge E. Hallyn" <[email protected]> wrote:
Based on the commit msg on
ae0e8e82205c903978a79ebf5e31c670b61fa5b4, it looksQuoted 12 lines
drivers/net/veth.c static void veth_setup(struct net_device *dev) · line 324
Quoted 8 lines
drivers/net/veth.c +1
-1ether_setup (dev);@@ -306,7 +320,7 @@ static void veth_setup(struct net_device *dev) dev->netdev_ops = &veth_netdev_ops; dev->ethtool_ops = &veth_ethtool_ops; dev->features |=NETIF_F_LLTX ; - dev->destructor = free_netdev; + dev->destructor = veth_dev_free;This is still going to oops if sysfs statistics referenced after module unload because module is unloaded (code is gone) and
veth_dev_freeno longer exists.Has anyone actually seen that cause an oops?
The reason I am asking is that as I read the code we cannot have this problem. At worst the destructor callback is delayed until:
veth_exit
rtnl_link_unregisterrtnl_unlock netdev_run_todo dev->destructorSimilarly even if the sysfs filehandle is open we have called:
netdev_unregister_kobject
...
sysfs_addrm_finish
sysfs_deactivate
Quoted 4 lines
Which guarantees that
sysfs_get_active_twowill fail and all subsequent actions on that file will fail.Eric
Sysfs must be safer than it used to be.
Quoted 2 lines · Stephen Hemminger
Sysfs must be safer than it used to be.
Definitely. A lot of this dates to Tejun's cleanups which merged 2-3 years agos now. Just before I started working on sysfs.
Eric