Re: [Linux Kernel Bug] UBSAN: array-index-out-of-bounds in rds_cmsg_recv

From: Zhu Yanjun
Date: Sat Jan 27 2024 - 22:53:38 EST



在 2024/1/27 8:00, Allison Henderson 写道:
On Mon, 2024-01-22 at 16:49 +0800, Zhu Yanjun wrote:
在 2024/1/22 13:48, Randy Dunlap 写道:
Hi,


On 1/21/24 00:34, Zhu Yanjun wrote:
在 2024/1/19 22:29, Chenyuan Yang 写道:
Dear Linux Kernel Developers for Network RDS,

We encountered "UBSAN: array-index-out-of-bounds in
rds_cmsg_recv"
when testing the RDS with our generated specifications. The C
reproduce program and logs for this crash are attached.

This crash happens when RDS receives messages by using
`rds_cmsg_recv`, which reads the `j+1` index of the array
`inc->i_rx_lat_trace`
(
https://urldefense.com/v3/__https://elixir.bootlin.com/linux/v6.
7/source/net/rds/recv.c*L585__;Iw!!ACWV5N9M2RV99hQ!J8QGG3fi_O0g
6p3oOboqNj5BuTcMuLuF-7-
SATmNj8EFTKyC68co6cnoG6LQzY1lJ9M_XA6voErOfj-qXTq3BSnW21Tk$ ).
The length of `inc->i_rx_lat_trace` array is 4 (defined by
`RDS_RX_MAX_TRACES`,
https://urldefense.com/v3/__https://elixir.bootlin.com/linux/v6.7/source/net/rds/rds.h*L289__;Iw!!ACWV5N9M2RV99hQ!J8QGG3fi_O0g6p3oOboqNj5BuTcMuLuF-7-SATmNj8EFTKyC68co6cnoG6LQzY1lJ9M_XA6voErOfj-qXTq3BYX3yVFo$
 ) while
`j` is the value stored in another array `rs->rs_rx_trace`
(
https://urldefense.com/v3/__https://elixir.bootlin.com/linux/v6.
7/source/net/rds/recv.c*L583__;Iw!!ACWV5N9M2RV99hQ!J8QGG3fi_O0g
6p3oOboqNj5BuTcMuLuF-7-
SATmNj8EFTKyC68co6cnoG6LQzY1lJ9M_XA6voErOfj-qXTq3BVTaaNkx$ ),
which is sent from others and could be arbitrary value.
I recommend to use the latest rds to make tests. The rds in linux
kernel upstream is too old. The rds in oracle linux is newer.
Why is the upstream kernel lagging behind?  Is the RDS maintainer
going
to submit patches to update mainline?
When I was in Oracle and worked with RDS, I have planned to upgrade
kernel rds to the latest. But after I submitted several patch series,
Oracle Developing Center of China was shutdown. I can not finish the
plan. But the UEK kernel in Oracle linux has the latest RDS.

If you want to make tests with rds, I recommend to use UEK kernel in
Oracle Linux.

Or you can install UEK kernel in RedHat. IMO, this UEK kernel can
also
work in RedHat Linux.

Zhu Yanjun
The challenge with updateing rds in upstream is that the uek rds
diverged from upstream a long time ago. So most of the uek patches
wont apply very well with a pretty big revert to bring it back to the
point of divergence. It not entirly clear how much rds is used outside
of oracle linux, but we are looking at how we might go about updating
at least the rds_tcp module, as we think this area would have less

From my perspective, a lot of people are more interested in rds_rdma module.

Exactly the gap between linux upstream and UEK is very big. But based on the rds features,

we can backport these features to linux upstream.

Zhu Yanjun

patching conflicts, and may be of more interest to community folks.
This is still very much a work in progress though, and still undergoing
a lot of investigation, so Zhu is likley correct in that for now it's
probably best to simply use a uek kernel if you are just wanting to
develop test cases.

Zhu, I was unaware that an effort had been submitted, but I am still
very much learning rds. If you want to point me to your set, I would
be happy to study it even if it was submitted a long time ago. Thanks!

Allison

Thanks.

Zhu Yanjun

This crash might be exploited to read the value out-of-bound
from the
array by setting arbitrary values for the array `rs-
rs_rx_trace`.
If you have any questions or require more information, please
feel
free to contact us.

Best,
Chenyuan