Bug with BPF_ALU64 | BPF_END?

Is BPF_END supposed to only be used with BPF_ALU, never with BPF_ALU64?
In kernel/bpf/core.c:___bpf_prog_run(), there are only jump table targets
 for the BPF_ALU case, not for the BPF_ALU64 case (opcodes 0xd7 and 0xdf).
But the verifier doesn't enforce this; by crafting a program that uses
 these opcodes I can get a WARN when they're run (without JIT; it looks
 like the x86 JIT, at least, won't like it either).
Proposed patch below the cut; build-tested only.

-Ed
---

[PATCH net] bpf/verifier: reject BPF_ALU64|BPF_END

Neither ___bpf_prog_run nor the JITs accept it.

Fixes: 17a5267067f3 ("bpf: verifier (add verifier core)")
Signed-off-by: Edward Cree <redacted>
---
 kernel/bpf/verifier.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
index 477b693..799b245 100644
--- a/kernel/bpf/verifier.c
+++ b/kernel/bpf/verifier.c

@@ -2292,7 +2292,8 @@ static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
 			}
 		} else {
 			if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
-			    (insn->imm != 16 && insn->imm != 32 && insn->imm != 64)) {
+			    (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
+			    BPF_CLASS(insn->code) == BPF_ALU64) {
 				verbose("BPF_END uses reserved fields\n");
 				return -EINVAL;
 			}

From: Edward Cree <redacted>
Date: Thu, 14 Sep 2017 18:53:17 +0100

Is BPF_END supposed to only be used with BPF_ALU, never with BPF_ALU64?
In kernel/bpf/core.c:___bpf_prog_run(), there are only jump table targets
 for the BPF_ALU case, not for the BPF_ALU64 case (opcodes 0xd7 and 0xdf).
But the verifier doesn't enforce this; by crafting a program that uses
 these opcodes I can get a WARN when they're run (without JIT; it looks
 like the x86 JIT, at least, won't like it either).
Proposed patch below the cut; build-tested only.

Good catch.

A really neat test would be a program that uploads random BPF programs
into the kernel, in a syzkaller'ish way.  It might have triggered this
(eventually).

On Thu, Sep 14, 2017 at 11:14 AM, David Miller [off-list ref] wrote:

From: Edward Cree <redacted>
Date: Thu, 14 Sep 2017 18:53:17 +0100

quoted

Is BPF_END supposed to only be used with BPF_ALU, never with BPF_ALU64?

Yes, only BPF_ALU. The below is LLVM bpf swap insn encoding:

...
// bswap16, bswap32, bswap64
class BSWAP ...
...
  let op = 0xd;     // BPF_END
  let BPFSrc = 1;   // BPF_TO_BE (TODO: use BPF_TO_LE for big-endian target)
  let BPFClass = 4; // BPF_ALU
...

quoted

In kernel/bpf/core.c:___bpf_prog_run(), there are only jump table targets
 for the BPF_ALU case, not for the BPF_ALU64 case (opcodes 0xd7 and 0xdf).
But the verifier doesn't enforce this; by crafting a program that uses
 these opcodes I can get a WARN when they're run (without JIT; it looks
 like the x86 JIT, at least, won't like it either).
Proposed patch below the cut; build-tested only.

Good catch.

A really neat test would be a program that uploads random BPF programs
into the kernel, in a syzkaller'ish way.  It might have triggered this
(eventually).

On 09/14/2017 07:53 PM, Edward Cree wrote:

Is BPF_END supposed to only be used with BPF_ALU, never with BPF_ALU64?
In kernel/bpf/core.c:___bpf_prog_run(), there are only jump table targets
  for the BPF_ALU case, not for the BPF_ALU64 case (opcodes 0xd7 and 0xdf).
But the verifier doesn't enforce this; by crafting a program that uses
  these opcodes I can get a WARN when they're run (without JIT; it looks
  like the x86 JIT, at least, won't like it either).
Proposed patch below the cut; build-tested only.

-Ed
---

[PATCH net] bpf/verifier: reject BPF_ALU64|BPF_END

Neither ___bpf_prog_run nor the JITs accept it.

Fixes: 17a5267067f3 ("bpf: verifier (add verifier core)")
Signed-off-by: Edward Cree <redacted>

Good catch! Can you submit this as an official patch for -net together
with a test case for tools/testing/selftests/bpf/test_verifier.c?

Thanks!

Acked-by: Daniel Borkmann <daniel@iogearbox.net>