[PATCH] net/sched: fix cls_flow SKUID/SKGID userns mask

[PATCH] net/sched: fix cls_flow SKUID/SKGID userns mask

[shaos]

Hello,

flow_change() checks FLOW_KEY_SKUID and FLOW_KEY_SKGID as bit values,
although they are enum indices and TCA_FLOW_KEYS is a bitmask.  The
current expression tests bits 0 through 3 instead of bits 14 and 15.

A process with CAP_NET_ADMIN in a non-initial user namespace can
therefore install SKUID/SKGID filters despite the user-namespace gate.
Those keys use initial-user-namespace credentials.  If another user's
locally generated traffic traverses the attacker's qdisc, the filter
can expose a property of that UID through class packet counters.  The
same typo also rejects some ordinary flow keys.  The demonstrated
cross-namespace inference requires such foreign traffic to reach the
attacker-owned qdisc; this does not disclose an arbitrary UID or grant
privileges.

HEAD commit:    c03c1a14507dfaf8b856c2608f9ac216021fc222
Git tree:       30fd2ef82912c26cf3a58ad0ba6afc375fa851ac
Source version: Linux 7.3
Compiler:       gcc 11.4.0, GNU ld 2.38

The console log, full kernel config, and reproducer are available from
one fixed-revision Gist:

Console output:
https://gist.githubusercontent.com/shaos26/a5571f1bfcd37c68c7270d296f1c12c0/raw/b3acb75e61d1cb703c7488ce0400651cd8c84b16/console-output.txt
Kernel config:
https://gist.githubusercontent.com/shaos26/a5571f1bfcd37c68c7270d296f1c12c0/raw/b3acb75e61d1cb703c7488ce0400651cd8c84b16/kernel-config.txt
PoC source:
https://gist.githubusercontent.com/shaos26/a5571f1bfcd37c68c7270d296f1c12c0/raw/b3acb75e61d1cb703c7488ce0400651cd8c84b16/poc.c

Clone this Gist and the exact Linux source revision, then build the
kernel with the attached config and run the PoC as PID 1 in an
initramfs:

    git clone \
      https://gist.github.com/shaos26/a5571f1bfcd37c68c7270d296f1c12c0.git \
      cls-flow-repro
    cd cls-flow-repro
    git clone \
      https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git \
      linux
    git -C linux checkout c03c1a14507dfaf8b856c2608f9ac216021fc222
    cp kernel-config.txt linux/.config
    make -C linux olddefconfig
    make -C linux -j"$(nproc)" bzImage
    gcc -static -O2 -Wall -Wextra -Werror -o init poc.c
    mkdir -p initramfs/dev initramfs/proc initramfs/sys
    cp init initramfs/init
    (cd initramfs && find . -print0 | \
     cpio --null -ov --format=newc | gzip -1 > ../flow-oracle.cpio.gz)
    timeout 180 qemu-system-x86_64 -m 768M -smp 2 \
      -kernel linux/arch/x86/boot/bzImage -initrd flow-oracle.cpio.gz \
      -append 'console=ttyS0 panic=-1 nokaslr rdinit=/init oracle_uid=424243' \
      -nographic -no-reboot > /tmp/flow-oracle-positive.log 2>&1

Change oracle_uid to 424242 for the even-UID control.  Run the same
command with the corrected-mask kernel to check the fixed behavior.

The vulnerable kernel accepted both restricted filters in two fresh
boots.  UID 424243 produced a 7/10 packet split between bands 1 and 2;
changing only the target to even UID 424242 put all 17 packets in band
1.  The corrected-mask kernel rejected SKUID and SKGID with -EOPNOTSUPP,
accepted the ordinary SRC key, and showed all 17 packets in band 2.
The fixed-kernel harness prints ORACLE_RESULT=FAIL intentionally,
because its predicate expects the vulnerable behavior.

The guard was added by commit a6c6796c7127
("userns: Convert cls_flow to work with user namespaces enabled").
Use shifted enum values when checking the TCA_FLOW_KEYS bitmask:

Fixes: a6c6796c7127 ("userns: Convert cls_flow to work with user namespaces enabled")
Signed-off-by: shaos <shawnq1@163.com>
---
 net/sched/cls_flow.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/net/sched/cls_flow.c b/net/sched/cls_flow.c
--- a/net/sched/cls_flow.c
+++ b/net/sched/cls_flow.c
@@ -433,7 +433,8 @@ static int flow_change(struct net *net, struct sk_buff *in_skb,
 		if (fls(keymask) - 1 > FLOW_KEY_MAX)
 			return -EOPNOTSUPP;
 
-		if ((keymask & (FLOW_KEY_SKUID|FLOW_KEY_SKGID)) &&
+		if ((keymask & ((1U << FLOW_KEY_SKUID) |
+				(1U << FLOW_KEY_SKGID))) &&
 		    sk_user_ns(NETLINK_CB(in_skb).sk) != &init_user_ns)
 			return -EOPNOTSUPP;
 	}
-- 
2.34.1