[LTS 9.2 RT] CVE-2023-4623, VULN-6713

[pvts-mat]

[LTS 9.2 RT]
CVE-2023-4623
VULN-6713

Problem

https://www.cve.org/CVERecord?id=CVE-2023-4623

A use-after-free vulnerability in the Linux kernel's net/sched: sch_hfsc (HFSC qdisc traffic control) component can be exploited to achieve local privilege escalation. If a class with a link-sharing curve (i.e. with the HFSC_FSC flag set) has a parent without a link-sharing curve, then init_vf() will call vttree_insert() on the parent, but vttree_remove() will be skipped in update_vf(). This leaves a dangling pointer that can cause a use-after-free.

Analysis and solution

A single commit was identified as a fix for this issue: b3d26c5702c7d6c45456326e56d2ccf3f103e60f net/sched: sch_hfsc: Ensure inner classes have fsc curve.

The solution consisted of rejecting the addition of a class with a link-sharing curve to the class without it (see Specific tests for details):

[root@ciqlts9 pvts]# (
    tc qdisc add dev lo root handle 1: hfsc default 10
    tc class add dev lo parent 1: classid 1:1 hfsc rt m2 100kbps
    tc class add dev lo parent 1:1 classid 1:10 hfsc ls m2 50kbps
    echo $?
)

Error: Invalid parent - parent class must have FSC.
2

The fix introduced a problem with existing network setup scripts for some users https://lore.kernel.org/all/[email protected]/:

I upgraded from 6.1.38 to 6.1.55 this morning and it broke my traffic shaping script, leaving me with a non-functional uplink on a remote router.

It was decided to fix the problem without breaking backwards compatibility https://lore.kernel.org/all/[email protected]/:

Setting 'rt' as a parent is incorrect and the man page is explicit about
it as it doesn't make sense 'qdisc wise'. Being able to set it has
always been wrong unfortunately…

Sure but unfortunately "we don't break backward compat" means
we can't really argue. It will take us more time to debate this
than to fix it (assuming we understand the initial problem).

Frankly one can even argue whether "exploitable by root / userns"
is more important than single user's init scripts breaking.
The "security" issues for root are dime a dozen.

The solution was to change the erroneous qdisc hierarchy to a correct one when the possible UAF condition was detected https://lore.kernel.org/all/[email protected]/:

As reported [1] disallowing 'rt' inner curves breaks already existing
scripts. Even though it doesn't make sense 'qdisc wise' users have been
relying on this behaviour since the qdisc inception.

We need users to update the scripts/applications to use 'sc' or 'ls'
as a inner curve instead of 'rt', but also avoid the UAF found by
Budimir, which was present since the qdisc inception.

The fix of the fix is given in the commit a13b67c9a015c4e21601ef9aa4ec9c5d972df1b4 net/sched: sch_hfsc: upgrade 'rt' to 'sc' when it becomes a inner curve

While the changes could be squashed into a single commit it was decided to retain the sequence of two commits for more straightforward LTS 9.2 RT - mainline patches correspondence.

The same solution was used already in other "version 9" branches centos9, ciqlts9_4, rocky9_4, rocky9_5, sig-cloud-9/5.14.0-427.37.1.el9_4, sig-cloud-9/5.14.0-427.40.1.el9_4, sig-cloud-9/5.14.0-427.42.1.el9_4, sig-cloud-9/5.14.0-503.19.1.el9_5, sig-cloud-9/5.14.0-503.22.1.el9_5:

git --no-pager log --decorate  --format="%h %cd %d %s" --date=short -n 2 9950d4d93

9950d4d93 2023-10-26  net/sched: sch_hfsc: upgrade 'rt' to 'sc' when it becomes a inner curve
ff4452fb0 2023-10-26  net/sched: sch_hfsc: Ensure inner classes have fsc curve

kABI check: omitted (unstable ABI of RT kernels)

Boot test: passed

Refer to Specific tests for implicit boot test passing.

Kselftests: passed relative

Methodology

A mix of kernel-selftests-internal and source-compiled tests were used:

• kernel-selftests-internal: bpf tests, except:
  • bpf:test_kmod.sh: takes very long time to finish and always fails anyway,
  • bpf:test_progs: unstable, can crash the machine,
  • bpf:test_progs-no_alu32: unstable, can crash the machine.
• source-compiled: all the rest.

Coverage (including tests skipped during execution)

bpf, breakpoints, capabilities, clone3, core, cpu-hotplug, cpufreq, drivers/dma-buf, drivers/net/bonding, drivers/net/team, efivarfs, filesystems, filesystems/binderfs, filesystems/epoll, firmware, fpu, ftrace, futex, gpio, intel_pstate, ipc, ir, kcmp, kvm, landlock, lib, livepatch, membarrier, memfd, memory-hotplug, mincore, mount, mqueue, nci, net, net/forwarding, net/mptcp, netfilter, nsfs, openat2, pid_namespace, pidfd, pstore, ptrace, rlimits, rseq, rtc, seccomp, sgx, sigaltstack, size, splice, static_keys, sync, syscall_user_dispatch, sysctl, tc-testing, tdx, timens, timers, tmpfs, tpm2, user, vDSO, vm, x86, zram

Reference ciqlts9_2-rt (4c5f656231e024966e5df6be5a66e3adf36379fb)

Three test runs were conducted on the reference kernel.
kselftests–mixed–ciqlts9_2-rt–run1.log
kselftests–mixed–ciqlts9_2-rt–run2.log
kselftests–mixed–ciqlts9_2-rt–run3.log

Patch ciqlts9_2-rt-CVE-2023-4623 (752717727a98bd7e29840fc4c65f8721aaae6a10)

A single test run was conducted on the patched kernel.
kselftests–mixed–ciqlts9_2-rt-CVE-2023-4623.log

Comparison

/home/pvts/gtd/projects/conclusive/ciq/rocky-patching/ktests.xsh  table --where "Summary = 'diff'"  kselftests*.log

Column    File
--------  -------------------------------------------------
Status0   kselftests--mixed--ciqlts9_2-rt--run1.log
Status1   kselftests--mixed--ciqlts9_2-rt--run2.log
Status2   kselftests--mixed--ciqlts9_2-rt--run3.log
Status3   kselftests--mixed--ciqlts9_2-rt-CVE-2023-4623.log

TestCase                  Status0  Status1  Status2  Status3  Summary
net/mptcp:mptcp_join.sh   fail     pass     pass     pass     diff
net:gro.sh                fail     pass     pass     pass     diff
net:txtimestamp.sh        pass     fail     fail     pass     diff
net:udpgso_bench.sh       pass     fail     pass     pass     diff
rtc:rtctest               fail     pass     fail     fail     diff
x86:syscall_numbering_64  fail     fail     pass     pass     diff

All of the tests with different results in the tested patch showed inconsistent behavior in the reference tests set itself. The net/mptcp:mptcp_join.sh test have shown inconsistent behavior before. Likely net:gro.sh, rtc:rtctest, net:txtimestamp.sh, although not yet on the ciqlts9_2-rt platform. Added to the list of flappy tests.

The net:udpgso_bench.sh did show inconsistent behavior before across different platforms but not on a single one. The one failing test seems to be associated with resources depletion

...
# udp gso zerocopy
# ./udpgso_bench_tx: sendmsg: No buffer space available
# udp gso timestamp
...

The x86:syscall_numbering_64 shows inconsistent behavior for the first time. All problems are associated with associated MODIFIED_BY_PTRACE errno, and returning -38 by the system call instead (ENOSYS 38 Function not implemented):

# [FAIL]              x32 syscall 0:-1 returned -38, but it should have returned MODIFIED_BY_PTRACE
# [FAIL]              x32 syscall 1:-1 returned -38, but it should have returned MODIFIED_BY_PTRACE
# [FAIL]              x32 syscall -1:-1 returned -38, but it should have returned MODIFIED_BY_PTRACE
# [FAIL]              x32 syscall 1073741823:-1 returned -38, but it should have returned MODIFIED_BY_PTRACE
# [FAIL]              x32 syscall 1073741824:-1 returned -38, but it should have returned MODIFIED_BY_PTRACE
# [FAIL]              x32 syscall 1073741823:-1 returned -38, but it should have returned MODIFIED_BY_PTRACE
# [FAIL]              x32 syscall -1073741824:-1 returned -38, but it should have returned MODIFIED_BY_PTRACE
# [FAIL]              x32 syscall 2147483647:-1 returned -38, but it should have returned MODIFIED_BY_PTRACE
# [FAIL]              x32 syscall -2147483648:-1 returned -38, but it should have returned MODIFIED_BY_PTRACE
# [FAIL]              x32 syscall -2147483647:-1 returned -38, but it should have returned MODIFIED_BY_PTRACE
# [FAIL]              x32 syscall 0:-1 returned -38, but it should have returned MODIFIED_BY_PTRACE
# [FAIL]              x32 syscall 1:-1 returned -38, but it should have returned MODIFIED_BY_PTRACE
# [FAIL]              x32 syscall -1:-1 returned -38, but it should have returned MODIFIED_BY_PTRACE
# [FAIL]              x32 syscall 1073741823:-1 returned -38, but it should have returned MODIFIED_BY_PTRACE
# [FAIL]              x32 syscall 1073741824:-1 returned -38, but it should have returned MODIFIED_BY_PTRACE
# [FAIL]              x32 syscall 1073741823:-1 returned -38, but it should have returned MODIFIED_BY_PTRACE
# [FAIL]              x32 syscall -1073741824:-1 returned -38, but it should have returned MODIFIED_BY_PTRACE
# [FAIL]              x32 syscall 2147483647:-1 returned -38, but it should have returned MODIFIED_BY_PTRACE
# [FAIL]              x32 syscall -2147483648:-1 returned -38, but it should have returned MODIFIED_BY_PTRACE
# [FAIL]              x32 syscall -2147483647:-1 returned -38, but it should have returned MODIFIED_BY_PTRACE
# [FAIL]              x32 syscall 0:-1 returned -38, but it should have returned MODIFIED_BY_PTRACE
# [FAIL]              x32 syscall 1:-1 returned -38, but it should have returned MODIFIED_BY_PTRACE
# [FAIL]              x32 syscall -1:-1 returned -38, but it should have returned MODIFIED_BY_PTRACE
# [FAIL]              x32 syscall 1073741823:-1 returned -38, but it should have returned MODIFIED_BY_PTRACE
# [FAIL]              x32 syscall 1073741824:-1 returned -38, but it should have returned MODIFIED_BY_PTRACE
# [FAIL]              x32 syscall 1073741823:-1 returned -38, but it should have returned MODIFIED_BY_PTRACE
# [FAIL]              x32 syscall -1073741824:-1 returned -38, but it should have returned MODIFIED_BY_PTRACE
# [FAIL]              x32 syscall 2147483647:-1 returned -38, but it should have returned MODIFIED_BY_PTRACE
# [FAIL]              x32 syscall -2147483648:-1 returned -38, but it should have returned MODIFIED_BY_PTRACE
# [FAIL]              x32 syscall -2147483647:-1 returned -38, but it should have returned MODIFIED_BY_PTRACE
# [FAIL]  A total of 30 system calls had incorrect behavior

The details of this failure were not investigated further

Specific tests: passed

The potential UAF condition was found to be reproducible with the following tc commands sequence:

tc qdisc add dev lo root handle 1: hfsc default 10
# Inner hfsc class constructed with 'rt' - realtime service
tc class add dev lo parent 1: classid 1:1 hfsc rt m2 100kbps
# Leaf hfsc class constructed with 'ls' - linksharing service
tc class add dev lo parent 1:1 classid 1:10 hfsc ls m2 50kbps

The "100kbps", "50kbps" parts are arbitrary. What's important is the use of rt for the inner class and ls for the leaf class. While the exact UAF was not obtained the commands helped confirm the efficacy of the patch.

Reference

The incorrect qdisc hierarchy can be created without any guardrails.

[root@ciqlts9 pvts]# (
    tc qdisc add dev lo root handle 1: hfsc default 10
    tc class add dev lo parent 1: classid 1:1 hfsc rt m2 100kbps
    tc class add dev lo parent 1:1 classid 1:10 hfsc ls m2 50kbps
    echo $?
)
0
[root@ciqlts9 pvts]# tc -g class show dev lo
tc -g class show dev lo
+---(1:) hfsc 
     +---(1:1) hfsc rt m1 0bit d 0us m2 800Kbit 
          +---(1:10) hfsc ls m1 0bit d 0us m2 400Kbit 

Full logs:
fix-replicate–ciqlts9_2-rt.log

Patch

Creating the incorrect qdisc hierarchy raises a warning, but succeeds. Notice the type of inner class being sc instead of rt as shown by tc -g class show dev lo command.

[root@ciqlts9 pvts]# (
    tc qdisc add dev lo root handle 1: hfsc default 10
    tc class add dev lo parent 1: classid 1:1 hfsc rt m2 100kbps
    tc class add dev lo parent 1:1 classid 1:10 hfsc ls m2 50kbps
    echo $?
)
Warning: Forced curve change on parent 'rt' to 'sc'.
0
[root@ciqlts9 pvts]# tc -g class show dev lo
tc -g class show dev lo
+---(1:) hfsc 
     +---(1:1) hfsc sc m1 0bit d 0us m2 800Kbit 
          +---(1:10) hfsc ls m1 0bit d 0us m2 400Kbit 

Full logs:
fix-replicate–ciqlts9_2-rt-CVE-2023-4623.log

[bmastbergen]

🥌

[gvrose8192]

LGTM - Thanks!

[gvrose8192]

Passes checks and is reviewed and approved.