]> git.itanic.dy.fi Git - linux-stable/commitdiff
bpf: fix incorrect tracking of register size truncation
authorDaniel Borkmann <daniel@iogearbox.net>
Fri, 22 Dec 2017 15:23:06 +0000 (16:23 +0100)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Mon, 25 Dec 2017 13:26:32 +0000 (14:26 +0100)
From: Jann Horn <jannh@google.com>

[ Upstream commit 0c17d1d2c61936401f4702e1846e2c19b200f958 ]

Properly handle register truncation to a smaller size.

The old code 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 in that case apart from
the obvious [0, 0xffff'ffff].

Starting with v4.14, this is exploitable by unprivileged users as long as
the unprivileged_bpf_disabled sysctl isn't set.

Debian assigned CVE-2017-16996 for this issue.

v2:
 - flip the mask during arithmetic bounds calculation (Ben Hutchings)
v3:
 - add CVE number (Ben Hutchings)

Fixes: b03c9f9fdc37 ("bpf/verifier: track signed and unsigned min/max values")
Signed-off-by: Jann Horn <jannh@google.com>
Acked-by: Edward Cree <ecree@solarflare.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
kernel/bpf/verifier.c

index 2811ced3c2a991ec036d0abf2bcc5ecccb89acda..1e2afb264c31fe1972053b1f3cb99f71105ba4f9 100644 (file)
@@ -1068,6 +1068,29 @@ static int check_ptr_alignment(struct bpf_verifier_env *env,
        return check_generic_ptr_alignment(reg, pointer_desc, off, size, strict);
 }
 
+/* truncate register to smaller size (in bytes)
+ * must be called with size < BPF_REG_SIZE
+ */
+static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
+{
+       u64 mask;
+
+       /* clear high bits in bit representation */
+       reg->var_off = tnum_cast(reg->var_off, size);
+
+       /* fix arithmetic bounds */
+       mask = ((u64)1 << (size * 8)) - 1;
+       if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
+               reg->umin_value &= mask;
+               reg->umax_value &= mask;
+       } else {
+               reg->umin_value = 0;
+               reg->umax_value = mask;
+       }
+       reg->smin_value = reg->umin_value;
+       reg->smax_value = reg->umax_value;
+}
+
 /* check whether memory at (regno + off) is accessible for t = (read | write)
  * if t==write, value_regno is a register which value is stored into memory
  * if t==read, value_regno is a register which will receive the value from memory
@@ -1200,9 +1223,7 @@ static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regn
        if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
            state->regs[value_regno].type == SCALAR_VALUE) {
                /* b/h/w load zero-extends, mark upper bits as known 0 */
-               state->regs[value_regno].var_off = tnum_cast(
-                                       state->regs[value_regno].var_off, size);
-               __update_reg_bounds(&state->regs[value_regno]);
+               coerce_reg_to_size(&state->regs[value_regno], size);
        }
        return err;
 }
@@ -1742,14 +1763,6 @@ static int check_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
        return 0;
 }
 
-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);
-}
-
 static bool signed_add_overflows(s64 a, s64 b)
 {
        /* Do the add in u64, where overflow is well-defined */
@@ -1984,8 +1997,8 @@ static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
 
        if (BPF_CLASS(insn->code) != BPF_ALU64) {
                /* 32-bit ALU ops are (32,32)->64 */
-               coerce_reg_to_32(dst_reg);
-               coerce_reg_to_32(&src_reg);
+               coerce_reg_to_size(dst_reg, 4);
+               coerce_reg_to_size(&src_reg, 4);
        }
        smin_val = src_reg.smin_value;
        smax_val = src_reg.smax_value;
@@ -2364,10 +2377,7 @@ static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
                                        return -EACCES;
                                }
                                mark_reg_unknown(regs, insn->dst_reg);
-                               /* high 32 bits are known zero. */
-                               regs[insn->dst_reg].var_off = tnum_cast(
-                                               regs[insn->dst_reg].var_off, 4);
-                               __update_reg_bounds(&regs[insn->dst_reg]);
+                               coerce_reg_to_size(&regs[insn->dst_reg], 4);
                        }
                } else {
                        /* case: R = imm