From: Antoine Tenart <atenart@kernel.org> Date: 2021-11-29 19:27:47
When updating Rx and Tx queue kobjects, the queue count should always be
updated to match the queue kobjects count. This was not done in the net
device unregistration path and due to the Tx queue logic allowing
updates when unregistering (for qdisc cleanup) it was possible with
ethtool to manually add new queues after unregister, leading to NULL
pointer exceptions and UaFs, such as:
BUG: KASAN: use-after-free in kobject_get+0x14/0x90
Read of size 1 at addr ffff88801961248c by task ethtool/755
CPU: 0 PID: 755 Comm: ethtool Not tainted 5.15.0-rc6+ #778
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.14.0-4.fc34 04/014
Call Trace:
dump_stack_lvl+0x57/0x72
print_address_description.constprop.0+0x1f/0x140
kasan_report.cold+0x7f/0x11b
kobject_get+0x14/0x90
kobject_add_internal+0x3d1/0x450
kobject_init_and_add+0xba/0xf0
netdev_queue_update_kobjects+0xcf/0x200
netif_set_real_num_tx_queues+0xb4/0x310
veth_set_channels+0x1c3/0x550
ethnl_set_channels+0x524/0x610
Updating the queue counts in the unregistration path solve the above
issue, as the ethtool path updating the queue counts makes sanity checks
and a queue count of 0 should prevent any update.
Fixes: 5c56580b74e5 ("net: Adjust TX queue kobjects if number of queues changes during unregister")
Signed-off-by: Antoine Tenart <atenart@kernel.org>
---
Hello,
Following a previous thread[1] I had another look at this issue and now
have a better fix (this patch). In this previous thread we also
discussed preventing ethtool operations after unregister and adding a
warning in netdev_queue_update_kobjects; I'll send two patches for this
but targetting net-next.
Thanks!
Antoine
[1] https://lore.kernel.org/all/20211122162007.303623-1-atenart@kernel.org/T/
net/core/net-sysfs.c | 3 +++
1 file changed, 3 insertions(+)
From: Jakub Kicinski <kuba@kernel.org> Date: 2021-12-01 02:08:45
On Mon, 29 Nov 2021 16:45:20 +0100 Antoine Tenart wrote:
When updating Rx and Tx queue kobjects, the queue count should always be
updated to match the queue kobjects count. This was not done in the net
device unregistration path and due to the Tx queue logic allowing
updates when unregistering (for qdisc cleanup) it was possible with
ethtool to manually add new queues after unregister, leading to NULL
pointer exceptions and UaFs, such as:
BUG: KASAN: use-after-free in kobject_get+0x14/0x90
Read of size 1 at addr ffff88801961248c by task ethtool/755
CPU: 0 PID: 755 Comm: ethtool Not tainted 5.15.0-rc6+ #778
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.14.0-4.fc34 04/014
Call Trace:
dump_stack_lvl+0x57/0x72
print_address_description.constprop.0+0x1f/0x140
kasan_report.cold+0x7f/0x11b
kobject_get+0x14/0x90
kobject_add_internal+0x3d1/0x450
kobject_init_and_add+0xba/0xf0
netdev_queue_update_kobjects+0xcf/0x200
netif_set_real_num_tx_queues+0xb4/0x310
veth_set_channels+0x1c3/0x550
ethnl_set_channels+0x524/0x610
Updating the queue counts in the unregistration path solve the above
issue, as the ethtool path updating the queue counts makes sanity checks
and a queue count of 0 should prevent any update.
Would you mind pointing where in the code that happens? I can't seem
to find anything looking at real_num_.x_queues outside dev.c and
net-sysfs.c :S
Fixes: 5c56580b74e5 ("net: Adjust TX queue kobjects if number of queues changes during unregister")
Signed-off-by: Antoine Tenart <atenart@kernel.org>
---
Following a previous thread[1] I had another look at this issue and now
have a better fix (this patch). In this previous thread we also
discussed preventing ethtool operations after unregister and adding a
warning in netdev_queue_update_kobjects; I'll send two patches for this
but targetting net-next.
From: Antoine Tenart <atenart@kernel.org> Date: 2021-12-01 09:32:09
Quoting Jakub Kicinski (2021-12-01 03:08:39)
On Mon, 29 Nov 2021 16:45:20 +0100 Antoine Tenart wrote:
quoted
When updating Rx and Tx queue kobjects, the queue count should always be
updated to match the queue kobjects count. This was not done in the net
device unregistration path and due to the Tx queue logic allowing
updates when unregistering (for qdisc cleanup) it was possible with
ethtool to manually add new queues after unregister, leading to NULL
pointer exceptions and UaFs, such as:
BUG: KASAN: use-after-free in kobject_get+0x14/0x90
Read of size 1 at addr ffff88801961248c by task ethtool/755
CPU: 0 PID: 755 Comm: ethtool Not tainted 5.15.0-rc6+ #778
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.14.0-4.fc34 04/014
Call Trace:
dump_stack_lvl+0x57/0x72
print_address_description.constprop.0+0x1f/0x140
kasan_report.cold+0x7f/0x11b
kobject_get+0x14/0x90
kobject_add_internal+0x3d1/0x450
kobject_init_and_add+0xba/0xf0
netdev_queue_update_kobjects+0xcf/0x200
netif_set_real_num_tx_queues+0xb4/0x310
veth_set_channels+0x1c3/0x550
ethnl_set_channels+0x524/0x610
Updating the queue counts in the unregistration path solve the above
issue, as the ethtool path updating the queue counts makes sanity checks
and a queue count of 0 should prevent any update.
Would you mind pointing where in the code that happens? I can't seem
to find anything looking at real_num_.x_queues outside dev.c and
net-sysfs.c :S
I read the above commit message again; it's not well explained... Sorry
for that.
The above trace was triggered using veths and this patch would solve
this as veths do use real_num_x_queues to fill 'struct ethtool_channels'
in its get_channels ops[1] which is then used to avoid making channel
counts updates if it is 0[2].
In addition, keeping track of the queue counts in the unregistration
path do help other drivers as it will allow adding a warning in
netdev_queue_update_kobjects when adding queues after unregister
(without tracking the queue counts in the unregistration path we can't
detect illegal queue additions). We could also not only warn for illegal
uses of netdev_queue_update_kobjects but also return an error.
Another change that was discussed is to forbid ethtool ops after
unregister. This is good, but is outside of the queue code so it might
not solve all issues.
(I do have those two patches, warn + ethtool, in my local tree and plan
on targeting net-next).
As you can see I'm a bit puzzled at how to fix this in the best way
possible[3]. I think the combination of the three patches should be good
enough, with only one sent to net as it does fix veths which IMHO is
easier to trigger. WDYT?
Thanks!
Antoine
[1] https://elixir.bootlin.com/linux/latest/source/drivers/net/veth.c#L222
[2] https://elixir.bootlin.com/linux/latest/source/net/ethtool/channels.c#L175
[3] Because the queue code does rely on external states.
From: Jakub Kicinski <kuba@kernel.org> Date: 2021-12-01 15:05:28
On Wed, 01 Dec 2021 10:32:01 +0100 Antoine Tenart wrote:
quoted
Would you mind pointing where in the code that happens? I can't seem
to find anything looking at real_num_.x_queues outside dev.c and
net-sysfs.c :S
I read the above commit message again; it's not well explained... Sorry
for that.
The above trace was triggered using veths and this patch would solve
this as veths do use real_num_x_queues to fill 'struct ethtool_channels'
in its get_channels ops[1] which is then used to avoid making channel
counts updates if it is 0[2].
But when we are at line 175 in [2] we already updated the values from
the user space request at lines 144-151. This check validates the new
config so a transition from 0 -> n should not be prevented here AFAICT.
In addition, keeping track of the queue counts in the unregistration
path do help other drivers as it will allow adding a warning in
netdev_queue_update_kobjects when adding queues after unregister
(without tracking the queue counts in the unregistration path we can't
detect illegal queue additions). We could also not only warn for illegal
uses of netdev_queue_update_kobjects but also return an error.
Another change that was discussed is to forbid ethtool ops after
unregister. This is good, but is outside of the queue code so it might
not solve all issues.
(I do have those two patches, warn + ethtool, in my local tree and plan
on targeting net-next).
As you can see I'm a bit puzzled at how to fix this in the best way
possible[3]. I think the combination of the three patches should be good
enough, with only one sent to net as it does fix veths which IMHO is
easier to trigger. WDYT?
[1] https://elixir.bootlin.com/linux/latest/source/drivers/net/veth.c#L222
[2] https://elixir.bootlin.com/linux/latest/source/net/ethtool/channels.c#L175
[3] Because the queue code does rely on external states.
Any way of fixing this is fine. If you ask me personally I'd probably
go with the ethtool fix to net and the zeroing and warn to net-next.
Unless I'm misreading and this fix does work, in which case your plan
is good, too.
From: Antoine Tenart <atenart@kernel.org> Date: 2021-12-01 17:03:41
Quoting Jakub Kicinski (2021-12-01 16:05:13)
On Wed, 01 Dec 2021 10:32:01 +0100 Antoine Tenart wrote:
quoted
quoted
Would you mind pointing where in the code that happens? I can't seem
to find anything looking at real_num_.x_queues outside dev.c and
net-sysfs.c :S
The above trace was triggered using veths and this patch would solve
this as veths do use real_num_x_queues to fill 'struct ethtool_channels'
in its get_channels ops[1] which is then used to avoid making channel
counts updates if it is 0[2].
But when we are at line 175 in [2] we already updated the values from
the user space request at lines 144-151. This check validates the new
config so a transition from 0 -> n should not be prevented here AFAICT.
You're right. It worked in my testbed because I was not providing the
number of Rx channels with Ethtool, so channels.rx_counts wasn't updated
and still had the value veth provided. It "fixed" the issue for a
completely unrelated reason and obviously can't be a fix for the issue
described here. Thanks for having a second look at this, I completely
missed it...
Any way of fixing this is fine. If you ask me personally I'd probably
go with the ethtool fix to net and the zeroing and warn to net-next.
Unless I'm misreading and this fix does work, in which case your plan
is good, too.
I think you're right. Let's target net with the ethtool patch[1]. This
patch (still needed to keep track of the queue counts) and the warning
one can target net-next.
[1] And probably another one for the old ethtool ioctl interface too.
Thanks!
Antoine