[PATCH bpf v4 0/2] Fix verifier id tracking of scalars on spill

[PATCH bpf v4 0/2] Fix verifier id tracking of scalars on spill

[Maxim Mikityanskiy]

From: Maxim Mikityanskiy <maxim@isovalent.com>

See the details in the commit message (TL/DR: under CAP_BPF, the
verifier can incorrectly conclude that a scalar is zero while in
fact it can be crafted to a predefined number.)

v1 and v2 were sent off-list.

v2 changes:

Added more tests, migrated them to inline asm, started using
bpf_get_prandom_u32, switched to a more bulletproof dead branch check
and modified the failing spill test scenarios so that an unauthorized
access attempt is performed in both branches.

v3 changes:

Dropped an improvement not necessary for the fix, changed the Fixes tag.

v4 changes:

Dropped supposedly redundant tests, kept the ones that result in
different verifier verdicts. Dropped the variable that is not yet
useful in this patch. Rephrased the commit message with Daniel's
suggestions.

Maxim Mikityanskiy (2):
  bpf: Fix verifier id tracking of scalars on spill
  selftests/bpf: Add test cases to assert proper ID tracking on spill

 kernel/bpf/verifier.c                         |  3 +
 .../selftests/bpf/progs/verifier_spill_fill.c | 79 +++++++++++++++++++
 2 files changed, 82 insertions(+)

-- 
2.40.1

[PATCH bpf v4 1/2] bpf: Fix verifier id tracking of scalars on spill

[Maxim Mikityanskiy]

From: Maxim Mikityanskiy <maxim@isovalent.com>

The following scenario describes a bug in the verifier where it
incorrectly concludes about equivalent scalar IDs which could lead to
verifier bypass in privileged mode:

1. Prepare a 32-bit rogue number.
2. Put the rogue number into the upper half of a 64-bit register, and
   roll a random (unknown to the verifier) bit in the lower half. The
   rest of the bits should be zero (although variations are possible).
3. Assign an ID to the register by MOVing it to another arbitrary
   register.
4. Perform a 32-bit spill of the register, then perform a 32-bit fill to
   another register. Due to a bug in the verifier, the ID will be
   preserved, although the new register will contain only the lower 32
   bits, i.e. all zeros except one random bit.

At this point there are two registers with different values but the same
ID, which means the integrity of the verifier state has been corrupted.

5. Compare the new 32-bit register with 0. In the branch where it's
   equal to 0, the verifier will believe that the original 64-bit
   register is also 0, because it has the same ID, but its actual value
   still contains the rogue number in the upper half.
   Some optimizations of the verifier prevent the actual bypass, so
   extra care is needed: the comparison must be between two registers,
   and both branches must be reachable (this is why one random bit is
   needed). Both branches are still suitable for the bypass.
6. Right shift the original register by 32 bits to pop the rogue number.
7. Use the rogue number as an offset with any pointer. The verifier will
   believe that the offset is 0, while in reality it's the given number.

The fix is similar to the 32-bit BPF_MOV handling in check_alu_op for
SCALAR_VALUE. If the spill is narrowing the actual register value, don't
keep the ID, make sure it's reset to 0.

Fixes: 354e8f1970f8 ("bpf: Support <8-byte scalar spill and refill")
Signed-off-by: Maxim Mikityanskiy <maxim@isovalent.com>
Acked-by: Yonghong Song <yhs@fb.com>
---
 kernel/bpf/verifier.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
index 5871aa78d01a..0dd8adc7a159 100644
--- a/kernel/bpf/verifier.c
+++ b/kernel/bpf/verifier.c
@@ -3868,6 +3868,9 @@ static int check_stack_write_fixed_off(struct bpf_verifier_env *env,
 				return err;
 		}
 		save_register_state(state, spi, reg, size);
+		/* Break the relation on a narrowing spill. */
+		if (fls64(reg->umax_value) > BITS_PER_BYTE * size)
+			state->stack[spi].spilled_ptr.id = 0;
 	} else if (!reg && !(off % BPF_REG_SIZE) && is_bpf_st_mem(insn) &&
 		   insn->imm != 0 && env->bpf_capable) {
 		struct bpf_reg_state fake_reg = {};
-- 
2.40.1

[PATCH bpf v4 2/2] selftests/bpf: Add test cases to assert proper ID tracking on spill

[Maxim Mikityanskiy]

From: Maxim Mikityanskiy <maxim@isovalent.com>

The previous commit fixed a verifier bypass by ensuring that ID is not
preserved on narrowing spills. Add the test cases to check the
problematic patterns.

Signed-off-by: Maxim Mikityanskiy <maxim@isovalent.com>
---
 .../selftests/bpf/progs/verifier_spill_fill.c | 79 +++++++++++++++++++
 1 file changed, 79 insertions(+)

diff --git a/tools/testing/selftests/bpf/progs/verifier_spill_fill.c b/tools/testing/selftests/bpf/progs/verifier_spill_fill.c
index 136e5530b72c..6115520154e3 100644
--- a/tools/testing/selftests/bpf/progs/verifier_spill_fill.c
+++ b/tools/testing/selftests/bpf/progs/verifier_spill_fill.c
@@ -371,4 +371,83 @@ __naked void and_then_at_fp_8(void)
 "	::: __clobber_all);
 }
 
+SEC("xdp")
+__description("32-bit spill of 64-bit reg should clear ID")
+__failure __msg("math between ctx pointer and 4294967295 is not allowed")
+__naked void spill_32bit_of_64bit_fail(void)
+{
+	asm volatile ("					\
+	r6 = r1;					\
+	/* Roll one bit to force the verifier to track both branches. */\
+	call %[bpf_get_prandom_u32];			\
+	r0 &= 0x8;					\
+	/* Put a large number into r1. */		\
+	r1 = 0xffffffff;				\
+	r1 <<= 32;					\
+	r1 += r0;					\
+	/* Assign an ID to r1. */			\
+	r2 = r1;					\
+	/* 32-bit spill r1 to stack - should clear the ID! */\
+	*(u32*)(r10 - 8) = r1;				\
+	/* 32-bit fill r2 from stack. */		\
+	r2 = *(u32*)(r10 - 8);				\
+	/* Compare r2 with another register to trigger find_equal_scalars.\
+	 * Having one random bit is important here, otherwise the verifier cuts\
+	 * the corners. If the ID was mistakenly preserved on spill, this would\
+	 * cause the verifier to think that r1 is also equal to zero in one of\
+	 * the branches, and equal to eight on the other branch.\
+	 */						\
+	r3 = 0;						\
+	if r2 != r3 goto l0_%=;				\
+l0_%=:	r1 >>= 32;					\
+	/* At this point, if the verifier thinks that r1 is 0, an out-of-bounds\
+	 * read will happen, because it actually contains 0xffffffff.\
+	 */						\
+	r6 += r1;					\
+	r0 = *(u32*)(r6 + 0);				\
+	exit;						\
+"	:
+	: __imm(bpf_get_prandom_u32)
+	: __clobber_all);
+}
+
+SEC("xdp")
+__description("16-bit spill of 32-bit reg should clear ID")
+__failure __msg("dereference of modified ctx ptr R6 off=65535 disallowed")
+__naked void spill_16bit_of_32bit_fail(void)
+{
+	asm volatile ("					\
+	r6 = r1;					\
+	/* Roll one bit to force the verifier to track both branches. */\
+	call %[bpf_get_prandom_u32];			\
+	r0 &= 0x8;					\
+	/* Put a large number into r1. */		\
+	w1 = 0xffff0000;				\
+	r1 += r0;					\
+	/* Assign an ID to r1. */			\
+	r2 = r1;					\
+	/* 16-bit spill r1 to stack - should clear the ID! */\
+	*(u16*)(r10 - 8) = r1;				\
+	/* 16-bit fill r2 from stack. */		\
+	r2 = *(u16*)(r10 - 8);				\
+	/* Compare r2 with another register to trigger find_equal_scalars.\
+	 * Having one random bit is important here, otherwise the verifier cuts\
+	 * the corners. If the ID was mistakenly preserved on spill, this would\
+	 * cause the verifier to think that r1 is also equal to zero in one of\
+	 * the branches, and equal to eight on the other branch.\
+	 */						\
+	r3 = 0;						\
+	if r2 != r3 goto l0_%=;				\
+l0_%=:	r1 >>= 16;					\
+	/* At this point, if the verifier thinks that r1 is 0, an out-of-bounds\
+	 * read will happen, because it actually contains 0xffff.\
+	 */						\
+	r6 += r1;					\
+	r0 = *(u32*)(r6 + 0);				\
+	exit;						\
+"	:
+	: __imm(bpf_get_prandom_u32)
+	: __clobber_all);
+}
+
 char _license[] SEC("license") = "GPL";
-- 
2.40.1

Re: [PATCH bpf v4 2/2] selftests/bpf: Add test cases to assert proper ID tracking on spill

[Yonghong Song]

On 6/7/23 5:39 AM, Maxim Mikityanskiy wrote:
> From: Maxim Mikityanskiy <maxim@isovalent.com>
> 
> The previous commit fixed a verifier bypass by ensuring that ID is not
> preserved on narrowing spills. Add the test cases to check the
> problematic patterns.
> 
> Signed-off-by: Maxim Mikityanskiy <maxim@isovalent.com>

Acked-by: Yonghong Song <yhs@fb.com>

Re: [PATCH bpf v4 0/2] Fix verifier id tracking of scalars on spill

[Daniel Borkmann]

On 6/7/23 2:39 PM, Maxim Mikityanskiy wrote:
[...]
> v4 changes:
> 
> Dropped supposedly redundant tests, kept the ones that result in
> different verifier verdicts. Dropped the variable that is not yet
> useful in this patch. Rephrased the commit message with Daniel's
> suggestions.

Andrii mentioned he did some veristat measurements and they looked good
to him. Looks like patchbot didn't reply, I've pushed it to bpf, thanks!