[Bug] [Smart] terminal tty: ioctl passes raw user pointer into kernel dereference — arbitrary kernel read/write from unprivileged user process

Aberta
#11,937 2 comentários 0 reações 0 responsáveis Ver no GitHub

Mantenedores costumam responder em até 2 dias

Ninguém assumiu esta issue ainda.

Avaliação

Dificuldade
4/5
Tempo estimado
3-5 dias
Facilidade para iniciantes
48/100
Tipo de issue
Bug
Clareza
Claramente especificada
Status de atividade
Ativa
Stack de tecnologia
c, linux

Direção de pesquisa

Read components/lwp/lwp_syscall.c around sys_ioctl and components/lwp/terminal/freebsd/tty_compat.c at lwp_tty_ioctl_adapter(), then trace the winsize cases in tty.c and the ptmx default ioctl path. The issue provides a QEMU reproduction and identifies expected user pointer checks or copy in/out handling. Done when kernel addresses and NULL are rejected while valid user buffers still work across the forwarded ioctl commands.

Escrita pelo modelo de indexação a partir do texto da issue.

Descrição

bug

RT-Thread Version

master (v5.4.0-dev, verified unpatched 2026-10-28) and all releases since v5.1.0 (introduced by PR #8672, commit 83e95bdf, 2024-03-28). v5.0.x is unaffected (the terminal subsystem did not exist yet).

Affected area

RT-Thread Smart / terminal (tty) subsystem — components/lwp/terminal/ + components/drivers/serial/serial_tty.c + the sys_ioctl path.

Hardware-BSP vendor

Others (QEMU vexpress-a9; applies to every RT_USING_SMART board)

Architecture

ARM 32-bit Cortex-A verified (qemu-vexpress-a9); the defect is architecture-independent and also affects AArch64 RT-Smart builds.

Board and hardware details

QEMU vexpress-a9 (Cortex-A9) with the stock qemu-vexpress-a9 .config + RT-Thread CI rtsmart fragment (RT_USING_SMART=y → LWP_USING_TERMINAL default y). No extra hardware or mounts are needed: on any RT-Smart build every serial port is automatically registered as a tty device (/dev/ttyS<n>, dev_serial.c rt_hw_serial_register() → rt_hw_serial_register_tty()); the boot banner confirms Using /dev/ttyS0 as default console.

Develop Toolchain

arm-linux-musleabi gcc 7.3 (RT-Thread CI toolchain) for the kernel; freestanding user-space PoC using official syscalls only.

Describe the bug

An unprivileged RT-Smart user process can read and write arbitrary kernel memory (and kill the kernel with a single NULL pointer) using only official APIs: open("/dev/ttyS0") + ioctl(TIOCSWINSZ/TIOCGWINSZ).

Where this code comes from, and why it is safe in FreeBSD but not here

The terminal subsystem was ported from FreeBSD's tty framework into master by PR #8672 (commit 83e95bdf, 2024-03-28); the files still carry the original BSD-2-Clause headers ("Copyright (c) 2008 Ed Schouten [email protected]"). Comparing against FreeBSD sys/kern/tty.c (main branch), the relevant ioctl cases are verbatim identical, e.g.:

/* FreeBSD sys/kern/tty.c */
case TIOCGWINSZ:
    /* Obtain window size. */
    *(struct winsize*)data = tp->t_winsize;
    return (0);
case TIOCSWINSZ:
    /* Set window size. */
    tty_set_winsize(tp, data);
    return (0);

This code has no problem in FreeBSD: FreeBSD's sys_ioctl() (sys/kern/sys_generic.c) copyin()s the user argument into a kernel stack buffer before dispatching to the cdev driver, so data always points to kernel memory there — the dereference pattern is safe by construction upstream.

The port dropped that premise, and RT-Smart adds a boundary FreeBSD never had to defend inside this code:

  1. sys_ioctl() (components/lwp/lwp_syscall.c:817) forwards the raw user pointer for all non-flock commands — there is no copyin layer like FreeBSD's;
  2. the ported-in adapter lwp_tty_ioctl_adapter() (components/lwp/terminal/freebsd/tty_compat.c:508-521) — an RT-Thread-specific function that does not exist in FreeBSD — only copies the TCGETS/TCSETS family in/out (tty_compat.c:348-408) and forwards the raw pointer for TIOCSCTTY / TIOCNOTTY / TIOCGPGRP / TIOCSPGRP / TIOCGSID / TIOCGWINSZ / TIOCSWINSZ / FIONREAD (error = IOCTL(cmd, (rt_caddr_t)args, fflags, td);). The ptmx master side does the same in its default branch (tty_ptmx.c:200 → tty_pts.c ptsdev_ioctl, e.g. TIOCGETA writes sizeof(struct termios) bytes);
  3. in RT-Smart the kernel and the user process live in separate address spaces — so when a user process passes a kernel address as the ioctl argument, the FreeBSD-inherited dereference now runs in kernel mode on an address the caller was never allowed to touch:
/* components/lwp/terminal/freebsd/tty.c:1341 */
case TIOCGWINSZ:
    *(struct winsize *)data = tp->t_winsize;   /* 8-byte write to ANY kernel address */
    return 0;
/* tty.c:1343 → tty_set_winsize, tty.c:1095 */
case TIOCSWINSZ:
    tty_set_winsize(tp, data);                 /* memcmp + 8-byte read from ANY kernel address */

Notably, TIOCSPGRP in the same passthrough set does validate (lwp_in_user_space() + copy, tty.c:1291-1303 — an RT-Thread addition; FreeBSD has no such branch), which shows this exact hazard was recognized for one command, but the rest of the set was missed.

Exploitation primitives (user process passes a kernel address as arg):

  • Arbitrary kernel 8-byte READ: ioctl(fd, TIOCSWINSZ, KADDR) then ioctl(fd, TIOCGWINSZ, &buf)
  • Arbitrary kernel 8-byte WRITE with fully controlled value: ioctl(fd, TIOCSWINSZ, &pattern) then ioctl(fd, TIOCGWINSZ, KADDR)
  • ~60-byte partially controlled write on a ptmx master fd: ioctl(mfd, TIOCGETA, KADDR) (tty_pts.c:303)
  • Deterministic kernel panic: ioctl(fd, TIOCSWINSZ, NULL) (tty.c:1095 dereferences NULL)

Reproduction (full-chain verified on QEMU RT-Smart, kernel master@5d1fb66d, unprivileged user process as /sbin/init using only official syscalls):

  1. Control: with user-space buffers, TIOCSWINSZ(&w) / TIOCGWINSZ(&r) round-trips exactly (6f 00 de 00 33 33 44 44); the shell still answers version. Kernel alive.

  2. Arbitrary read: TIOCSWINSZ(0xc00b0388) (address of rt_tick_get in kernel .text) leaks 00 48 2d e9 04 b0 8d e2 — byte-identical to the ELF contents at that address; TIOCSWINSZ(0xc00f40fc) (_uart_device) leaks kernel pointers 0xc0114010 / 0xc00d93d8. Kernel stays alive.

  3. Arbitrary write: TIOCSWINSZ with pattern 42 41 44 43 46 45 48 47, then TIOCGWINSZ(0xc0102f58) (kernel .bss tty_nin). A GDB hardware watchpoint on 0xc0102f58 catches the store at tty.c:1341 (Old (tty_nin=0, tty_nout=0xB0) → New (tty_nin=0x43444142, …)); reading the address back through the read primitive returns the exact 8-byte pattern. Kernel stays alive.

    Backtrace for the write (user-supplied kernel pointer flowing unvalidated through every layer):

    sys_ioctl (lwp_syscall.c:817, data=0xc0102f58 <tty_nin>)
    → ioctl (dfs_posix.c:737) → fcntl → dfs_file_ioctl (dfs_file.c:1658)
    → dfs_devfs_ioctl (devfs.c:234, args=0xc0102f58)
    → tty_fops_ioctl (tty_device.c:114, arg=0xc0102f58)
    → lwp_tty_ioctl_adapter (tty_compat.c:520, args=0xc0102f58)
    → ttydev_ioctl (tty.c:650) → tty_ioctl (tty.c:1395)
    → tty_generic_ioctl (tty.c:1341)  ← *(struct winsize *)0xc0102f58 = tp->t_winsize
    
  4. Availability: TIOCGWINSZ(0xc00b0388) overwrites the first instructions of rt_tick_get (kernel .text is writable in this configuration) → repeated data aborts → kernel dies. Also TIOCSWINSZ(NULL) → data abort at 0x0 → kernel dies.

Expected behavior

The tty layer must never dereference a user-supplied raw pointer: FreeBSD guarantees this by copying ioctl args in/out at the syscall boundary. In RT-Thread, a kernel address or NULL passed as arg must be rejected with EFAULT, and user addresses must go through lwp_get_from_user()/lwp_put_to_user().

Actual behavior

Arbitrary kernel read/write plus trivial kernel panic, reachable by any unprivileged RT-Smart user process through official POSIX APIs on a default image (no mounts, no privileges, no races).

Suggested fix

In lwp_tty_ioctl_adapter() (and the ptm_fops_ioctl() default branch), copy in/out the payloads of every forwarded command into kernel stack buffers, mirroring FreeBSD sys_ioctl — one change closes all sibling sinks at once (TIOCSWINSZ/TIOCGWINSZ/TIOCGPGRP/TIOCGSID/FIONREAD on the slave side; TIOCGETA/TIOCGPTN/TIOCGPGRP/TIOCGSID/FIONREAD/TIOCSIG/TIOCPKT on the master side). Alternatively, reject lwp_in_user_space(data) inside tty_generic_ioctl()/ptsdev_ioctl().

Other additional context

  • CWE-787 / CWE-125 (out-of-bounds write/read via unchecked pointer trust); CWE-476 for the NULL variant.
  • CVSS 3.1: AV:L/AC:L/PR:L/UI:N/S:C/C:H/I:H/A:H = 8.8 (scope-change reading: the attacked authority is the unprivileged user process, the impacted resource is the kernel; the S:U reading gives 7.8). All three impact axes are demonstrated with serial logs and a GDB watchpoint, not inferred.
  • Attack preconditions, stated explicitly: requires an RT-Smart build (RT_USING_SMART=y; LWP_USING_TERMINAL is default-y and serial ttys are auto-registered, so default Smart images ship the vulnerable devices); the attacker is any user-mode process; the /dev/ttyS<n> entry needs no mount (a /dev/pts/* pty entry also exists when ptyfs is mounted).
  • Not a duplicate: no existing issue/PR touches the terminal subsystem (searched winsize / TIOCGWINSZ / terminal ioctl / lwp tty / ptmx); none of the known CVEs cover it; verified still unpatched on origin/master as of 2026-10-28 (last touch was #8996, 2024-05-28).
  • The bug is in the RT-Thread porting/adaptation layer (the missing copyin premise + the RT-Thread-specific passthrough adapter), not in the FreeBSD-derived logic itself — the FreeBSD code only becomes unsafe because its precondition disappeared. Full repro artifacts (4-mode PoC, QEMU session harness, GDB watchpoint script, complete serial logs) available on request.
Linguagem predominante
C
Estrelas
12.4k
Forks
5.5k
Merge médio
3d 7h
PRs com merge (30d)
79

Preparar o ambiente

Abrir no Codespaces

Inicia o contêiner de desenvolvimento do projeto no navegador, com a sua própria conta do GitHub.

Primeiros passos

  1. Leia a issue inteira e depois o guia de contribuição do projeto.
  2. Comente na issue dizendo que vai assumir — evita que duas pessoas façam o mesmo trabalho.
  3. Faça um fork do repositório e trabalhe em uma branch.
  4. Abra um pull request que referencie o número da issue.

Mais de RT-Thread/rt-thread

Todas as issues de RT-Thread/rt-thread

Issues semelhantes

Mais issues de C

Receba novas issues na sua caixa de entrada

Um resumo curto de issues do GitHub para quem está começando.