bonding: do not enslave CAN devices
Subsystems (5)
PATCH 1/1 bonding: do not enslave CAN devices
Since commit 8df9ffb888c ("can: make use of preallocated can_ml_priv for per device struct can_dev_rcv_lists") the device specific CAN receive filter lists are stored in netdev_priv() and dev->ml_priv points to these filters.
In the bug report Syzkaller enslaved a vxcan1 CAN device and accessed the bonding device with a PF_CAN socket which lead to a crash due to an access of an unhandled bond_dev->ml_priv pointer.
Deny to enslave CAN devices by the bonding driver as the resulting bond_dev pretends to be a CAN device by copying dev->type without really being one.
device struct can_dev_rcv_lists")
--- drivers/net/bonding/bond_main.c | 12 ++++++++++++ 1 file changed, 12 insertions(+)
drivers/net/bonding/bond_main.c +12 -0
@@ -1475,6 +1475,18 @@ int bond_enslave(struct net_device *bond_dev, struct net_device *slave_dev,return -EPERM ; } +/* CAN network devices hold device specific filter lists in +* netdev_priv() where dev->ml_priv sets a reference to. +* As bonding assumes to have some ethernet-like device it doesn't +* take care about these CAN specific filter lists today. +* So we deny the enslaving of CAN interfaces here. +*/ +if (slave_dev->type ==ARPHRD_CAN ) { +NL_SET_ERR_MSG (extack,"CAN devices can not be enslaved" ); +slave_err (bond_dev, slave_dev,"no bonding on CAN devices\n" ); +return -EINVAL ; + } +/* set bonding device ether type by slave - bonding netdevices are * created with ether_setup, so when the slave type is not ARPHRD_ETHER * there is a need to override some of the type dependent attribs/funcs.
Signature · 1 line
2.20.1
Hello,
Quoted 10 lines · Oliver Hartkopp
Since commit
8df9ffb888c("can: make use of preallocatedcan_ml_privfor per device structcan_dev_rcv_lists") the device specific CAN receive filter lists are stored innetdev_priv()anddev->ml_privpoints to these filters.In the bug report Syzkaller enslaved a vxcan1 CAN device and accessed the bonding device with a
PF_CANsocket which lead to a crash due to an access of an unhandledbond_dev->ml_privpointer.Deny to enslave CAN devices by the bonding driver as the resulting
bond_devpretends to be a CAN device by copyingdev->typewithout really being one.
Does the team driver have the same problem?
Signature · 1 line
Sabrina
Quoted 5 lines · Sabrina Dubroca
Quoted 4 lines · Oliver Hartkopp
Since commit
8df9ffb888c("can: make use of preallocatedcan_ml_privfor per device structcan_dev_rcv_lists") the device specific CAN receive filter lists are stored innetdev_priv()anddev->ml_privpoints to these filters.
Quoted 8 lines
Quoted 6 lines
In the bug report Syzkaller enslaved a vxcan1 CAN device and accessed the bonding device with a
PF_CANsocket which lead to a crash due to an access of an unhandledbond_dev->ml_privpointer.Deny to enslave CAN devices by the bonding driver as the resulting
bond_devpretends to be a CAN device by copyingdev->typewithout really being one.Does the team driver have the same problem?
Good point!
From a first look into team_setup_by_port() in team.c I would say YES :-)
Thanks for watching out! I would suggest to wait for some more feedback and upstream of this fix.
Best regards,
Oliver
Any updates, reviews, acks on this?
As pointed out by Sabrina here https://marc.info/?l=linux-netdev&m=158039302905460&w=2
the issue is also relevant for the TEAM driver.
Best,
Oliver
Quoted 44 lines · Oliver Hartkopp
Since commit
8df9ffb888c("can: make use of preallocatedcan_ml_privfor per device structcan_dev_rcv_lists") the device specific CAN receive filter lists are stored innetdev_priv()anddev->ml_privpoints to these filters.In the bug report Syzkaller enslaved a vxcan1 CAN device and accessed the bonding device with a
PF_CANsocket which lead to a crash due to an access of an unhandledbond_dev->ml_privpointer.Deny to enslave CAN devices by the bonding driver as the resulting
bond_devpretends to be a CAN device by copyingdev->typewithout really being one.Reported-by [email protected]Fixes 8df9ffb888c ("can: make use of preallocated can_ml_priv for perdevice struct
can_dev_rcv_lists")Cc linux-stable <[email protected]> # >= v5.4Signed-off-by Oliver Hartkopp <[email protected]>--- drivers/net/bonding/bond_main.c | 12 ++++++++++++ 1 file changed, 12 insertions(+)drivers/net/bonding/bond_main.c +12
-0@@ -1475,6 +1475,18 @@ int bond_enslave(struct net_device *bond_dev, struct net_device *slave_dev,return -EPERM ; } +/* CAN network devices hold device specific filter lists in +* netdev_priv() where dev->ml_priv sets a reference to. +* As bonding assumes to have some ethernet-like device it doesn't +* take care about these CAN specific filter lists today. +* So we deny the enslaving of CAN interfaces here. +*/ +if (slave_dev->type ==ARPHRD_CAN ) { +NL_SET_ERR_MSG (extack,"CAN devices can not be enslaved" ); +slave_err (bond_dev, slave_dev,"no bonding on CAN devices\n" ); +return -EINVAL ; + } +/* set bonding device ether type by slave - bonding netdevices are * created with ether_setup, so when the slave type is not ARPHRD_ETHER * there is a need to override some of the type dependent attribs/funcs.
Quoted 8 lines · Oliver Hartkopp
Since commit
8df9ffb888c("can: make use of preallocatedcan_ml_privfor per device structcan_dev_rcv_lists") the device specific CAN receive filter lists are stored innetdev_priv()anddev->ml_privpoints to these filters.In the bug report Syzkaller enslaved a vxcan1 CAN device and accessed the bonding device with a
PF_CANsocket which lead to a crash due to an access of an unhandledbond_dev->ml_privpointer.
Quoted 8 lines
Deny to enslave CAN devices by the bonding driver as the resulting
bond_devpretends to be a CAN device by copyingdev->typewithout really being one.Reported-by [email protected]Fixes 8df9ffb888c ("can: make use of preallocated can_ml_priv for perdevice struct
can_dev_rcv_lists")Cc linux-stable <[email protected]> # >= v5.4Signed-off-by Oliver Hartkopp <[email protected]>
What's the preferred to upstream this? I could take this via the linux-can tree.
regards,
Marc
Quoted 28 lines
--- drivers/net/bonding/bond_main.c | 12 ++++++++++++ 1 file changed, 12 insertions(+)drivers/net/bonding/bond_main.c +12
-0@@ -1475,6 +1475,18 @@ int bond_enslave(struct net_device *bond_dev, struct net_device *slave_dev,return -EPERM ; } +/* CAN network devices hold device specific filter lists in +* netdev_priv() where dev->ml_priv sets a reference to. +* As bonding assumes to have some ethernet-like device it doesn't +* take care about these CAN specific filter lists today. +* So we deny the enslaving of CAN interfaces here. +*/ +if (slave_dev->type ==ARPHRD_CAN ) { +NL_SET_ERR_MSG (extack,"CAN devices can not be enslaved" ); +slave_err (bond_dev, slave_dev,"no bonding on CAN devices\n" ); +return -EINVAL ; + } +/* set bonding device ether type by slave - bonding netdevices are * created with ether_setup, so when the slave type is not ARPHRD_ETHER * there is a need to override some of the type dependent attribs/funcs.
Signature · 4 lines
Pengutronix e.K. | Marc Kleine-Budde | Embedded Linux | https://www.pengutronix.de | Vertretung West/Dortmund | Phone: +49-231-2826-924 | Amtsgericht Hildesheim, HRA 2686 | Fax: +49-5121-206917-5555 |
David Miller
maintainerFrom: Marc Kleine-Budde <[email protected]>
Date: Tue, 25 Feb 2020 21:32:41 +0100
Quoted 17 lines · Oliver Hartkopp
Quoted 8 lines
Since commit
8df9ffb888c("can: make use of preallocatedcan_ml_privfor per device structcan_dev_rcv_lists") the device specific CAN receive filter lists are stored innetdev_priv()anddev->ml_privpoints to these filters.In the bug report Syzkaller enslaved a vxcan1 CAN device and accessed the bonding device with a
PF_CANsocket which lead to a crash due to an access of an unhandledbond_dev->ml_privpointer.Quoted 8 lines
Deny to enslave CAN devices by the bonding driver as the resulting
bond_devpretends to be a CAN device by copyingdev->typewithout really being one.Reported-by [email protected]Fixes 8df9ffb888c ("can: make use of preallocated can_ml_priv for perdevice struct
can_dev_rcv_lists")Cc linux-stable <[email protected]> # >= v5.4Signed-off-by Oliver Hartkopp <[email protected]>
Quoted 4 lines
What's the preferred to upstream this? I could take this via the linux-can tree.
What I don't get is why the PF_CAN is blindly dereferencing a device assuming what is behind bond_dev->ml_priv.
If it assumes a device it access is CAN then it should check the device by comparing the netdev_ops or via some other means.
This restriction seems arbitrary.
Quoted 22 lines · David Miller
From: Marc Kleine-Budde <[email protected]>
Date: Tue, 25 Feb 2020 21:32:41 +0100Quoted 19 lines · Oliver Hartkopp
Quoted 8 lines
Since commit
8df9ffb888c("can: make use of preallocatedcan_ml_privfor per device structcan_dev_rcv_lists") the device specific CAN receive filter lists are stored innetdev_priv()anddev->ml_privpoints to these filters.In the bug report Syzkaller enslaved a vxcan1 CAN device and accessed the bonding device with a
PF_CANsocket which lead to a crash due to an access of an unhandledbond_dev->ml_privpointer.Quoted 8 lines
Deny to enslave CAN devices by the bonding driver as the resulting
bond_devpretends to be a CAN device by copyingdev->typewithout really being one.Reported-by [email protected]Fixes 8df9ffb888c ("can: make use of preallocated can_ml_priv for perdevice struct
can_dev_rcv_lists")Cc linux-stable <[email protected]> # >= v5.4Signed-off-by Oliver Hartkopp <[email protected]>
Quoted 8 lines
Quoted 2 lines
What's the preferred to upstream this? I could take this via the linux-can tree.
What I don't get is why the
PF_CANis blindly dereferencing a device assuming what is behindbond_dev->ml_priv.If it assumes a device it access is CAN then it should check the device by comparing the
netdev_opsor via some other means.
Yes we do.
Quoted 1 line
This restriction seems arbitrary.
Since commit 8df9ffb888c the data structures for the CAN filters have been moved from net/can/af_can.c into netdev->ml_priv.
PF_CAN only works with CAN interfaces and therefore always checks dev->type to be ARPHRD_CAN before accessing netdev->ml_priv.
Bonding and Team driver copy most of the device data structures to create bonding/team devices. They copy dev->type but *not* dev->ml_priv.
That leads to the problematic ml_priv access after passing the dev->type check ...
I don't know yet whether it makes sense to have CAN bonding/team devices. But if so we would need some more investigation. For now disabling CAN interfaces for bonding/team devices seems to be reasonable.
Regards,
Oliver
David Miller
maintainerFrom: Oliver Hartkopp <[email protected]>
Date: Mon, 2 Mar 2020 09:45:41 +0100
Quoted 4 lines
I don't know yet whether it makes sense to have CAN bonding/team devices. But if so we would need some more investigation. For now disabling CAN interfaces for bonding/team devices seems to be reasonable.
Every single interesting device that falls into a special use case like CAN is going to be tempted to add a similar check.
I don't want to set this precedence.
Check that the devices you get passed are actually CAN devices, it's easy, just compare the netdev_ops and make sure they equal the CAN ones.
Marc Kleine-Budde
maintainerQuoted 8 lines · David Miller
From: Oliver Hartkopp <[email protected]>
Date: Mon, 2 Mar 2020 09:45:41 +0100Quoted 4 lines
I don't know yet whether it makes sense to have CAN bonding/team devices. But if so we would need some more investigation. For now disabling CAN interfaces for bonding/team devices seems to be reasonable.
Quoted 8 lines
Every single interesting device that falls into a special use case like CAN is going to be tempted to add a similar check.
I don't want to set this precedence.
Check that the devices you get passed are actually CAN devices, it's easy, just compare the
netdev_opsand make sure they equal the CAN ones.
Sorry, I'm not really sure how to implement this check.
Should I maintain a list of all netdev_ops of all the CAN devices (= whitelist) and the compare against that list? Having a global list of pointers to network devices remind me of the old days of kernel-2.4.
regards,
Marc
Signature · 4 lines
Pengutronix e.K. | Marc Kleine-Budde | Embedded Linux | https://www.pengutronix.de | Vertretung West/Dortmund | Phone: +49-231-2826-924 | Amtsgericht Hildesheim, HRA 2686 | Fax: +49-5121-206917-5555 |
Quoted 19 lines · Marc Kleine-Budde
Quoted 8 lines · David Miller
From: Oliver Hartkopp <[email protected]>
Date: Mon, 2 Mar 2020 09:45:41 +0100Quoted 4 lines
I don't know yet whether it makes sense to have CAN bonding/team devices. But if so we would need some more investigation. For now disabling CAN interfaces for bonding/team devices seems to be reasonable.
Quoted 8 lines
Every single interesting device that falls into a special use case like CAN is going to be tempted to add a similar check.
I don't want to set this precedence.
Check that the devices you get passed are actually CAN devices, it's easy, just compare the
netdev_opsand make sure they equal the CAN ones.
Quoted 5 lines
Sorry, I'm not really sure how to implement this check.
Should I maintain a list of all
netdev_opsof all the CAN devices (= whitelist) and the compare against that list? Having a global list of pointers to network devices remind me of the old days of kernel-2.4.
I think Dave means something like this:
$ grep "netdev_ops == " drivers/net/*/*.c net/*/*.c drivers/net/hyperv/netvsc_drv.c: if (event_dev->netdev_ops == &device_ops) drivers/net/ppp/ppp_generic.c: if (dev->netdev_ops == &ppp_netdev_ops) net/dsa/slave.c: return dev->netdev_ops == &dsa_slave_netdev_ops; net/openvswitch/vport-internal_dev.c: return netdev->netdev_ops == &internal_dev_netdev_ops;
David Miller
maintainerFrom: Marc Kleine-Budde <[email protected]>
Date: Fri, 6 Mar 2020 15:12:48 +0100
Quoted 12 lines · David Miller
Quoted 11 lines
From: Oliver Hartkopp <[email protected]>
Date: Mon, 2 Mar 2020 09:45:41 +0100Quoted 4 lines
I don't know yet whether it makes sense to have CAN bonding/team devices. But if so we would need some more investigation. For now disabling CAN interfaces for bonding/team devices seems to be reasonable.
Every single interesting device that falls into a special use case like CAN is going to be tempted to add a similar check.
Quoted 7 lines
Quoted 5 lines
I don't want to set this precedence.
Check that the devices you get passed are actually CAN devices, it's easy, just compare the
netdev_opsand make sure they equal the CAN ones.Sorry, I'm not really sure how to implement this check.
Like this:
if (netdev->ops != &can_netdev_ops)
return;
Done.
Marc Kleine-Budde
maintainerQuoted 21 lines · David Miller
From: Marc Kleine-Budde <[email protected]>
Date: Fri, 6 Mar 2020 15:12:48 +0100Quoted 18 lines · David Miller
Quoted 16 lines
From: Oliver Hartkopp <[email protected]>
Date: Mon, 2 Mar 2020 09:45:41 +0100Quoted 4 lines
I don't know yet whether it makes sense to have CAN bonding/team devices. But if so we would need some more investigation. For now disabling CAN interfaces for bonding/team devices seems to be reasonable.
Every single interesting device that falls into a special use case like CAN is going to be tempted to add a similar check.
I don't want to set this precedence.
Check that the devices you get passed are actually CAN devices, it's easy, just compare the
netdev_opsand make sure they equal the CAN ones.
Quoted 6 lines
Quoted 1 line
Sorry, I'm not really sure how to implement this check.
Like this:
if (
netdev->ops!=&can_netdev_ops)return;
There is no single can_netdev_ops. The netdev_ops are per CAN-network driver. But the ml_priv is used in the generic CAN code.
regards,
Marc
Signature · 4 lines
Pengutronix e.K. | Marc Kleine-Budde | Embedded Linux | https://www.pengutronix.de | Vertretung West/Dortmund | Phone: +49-231-2826-924 | Amtsgericht Hildesheim, HRA 2686 | Fax: +49-5121-206917-5555 |
Quoted 24 lines · Marc Kleine-Budde
Quoted 23 lines · David Miller
From: Marc Kleine-Budde <[email protected]>
Date: Fri, 6 Mar 2020 15:12:48 +0100Quoted 19 lines · David Miller
Quoted 8 lines
From: Oliver Hartkopp <[email protected]>
Date: Mon, 2 Mar 2020 09:45:41 +0100Quoted 4 lines
I don't know yet whether it makes sense to have CAN bonding/team devices. But if so we would need some more investigation. For now disabling CAN interfaces for bonding/team devices seems to be reasonable.
Quoted 8 lines
Every single interesting device that falls into a special use case like CAN is going to be tempted to add a similar check.
I don't want to set this precedence.
Check that the devices you get passed are actually CAN devices, it's easy, just compare the
netdev_opsand make sure they equal the CAN ones.Sorry, I'm not really sure how to implement this check.
Quoted 7 lines
Quoted 4 lines
Like this:
if (
netdev->ops!=&can_netdev_ops)return;There is no single
can_netdev_ops. Thenetdev_opsare per CAN-network driver. But theml_privis used in the generic CAN code.
ping,
are there any other ways or ideas how to solve this issue?
Regards,
Oleksij
Signature · 4 lines
Pengutronix e.K. | | Steuerwalder Str. 21 | http://www.pengutronix.de/ | 31137 Hildesheim, Germany | Phone: +49-5121-206917-0 | Amtsgericht Hildesheim, HRA 2686 | Fax: +49-5121-206917-5555 |
+ Sabrina Dubroca (takes care of team driver which has the same issue)
Quoted 8 lines · Oleksij Rempel
Quoted 7 lines · Marc Kleine-Budde
Quoted 4 lines · David Miller
Like this:
if (
netdev->ops!=&can_netdev_ops)return;
Quoted 6 lines
Quoted 2 lines
There is no single
can_netdev_ops. Thenetdev_opsare per CAN-network driver. But theml_privis used in the generic CAN code.ping,
are there any other ways or ideas how to solve this issue?
Well, IMO the patch from
https://marc.info/?l=linux-can&m=158039108004683
is still valid.
Although the attribution that commit 8df9ffb888c ("can: make use of preallocated can_ml_priv for per device struct can_dev_rcv_lists") introduced the problem could be removed.
Even before this commit dev->ml_priv of CAN interfaces has been used to store the filter lists. And either the bonding and the team driver do not take care of ml_priv.
They pretend to be CAN interfaces by faking dev->type to be ARPHRD_CAN - but they are not. When we dereference dev->ml_priv in (badly) faked CAN devices we run into the reported issue.
So the approach is to tell bonding and team driver to keep the fingers away from CAN interfaces like we do with some ARPHRD_INFINIBAND setups in bond_enslave() too.
Regards,
Oliver