bpf, verifier: Improve precision of BPF_MUL
PATCH v2 1/1 bpf, verifier: Improve precision of BPF_MUL
This patch improves (or maintains) the precision of register value tracking in BPF_MUL across all possible inputs. It also simplifies scalar32_min_max_mul() and scalar_min_max_mul().
As it stands, BPF_MUL is composed of three functions:
case BPF_MUL:
tnum_mul();
scalar32_min_max_mul();
scalar_min_max_mul();
The current implementation of scalar_min_max_mul() restricts the u64 input ranges of dst_reg and src_reg to be within [0, U32_MAX]:
/* Both values are positive, so we can work with unsigned and
* copy the result to signed (unless it exceeds S64_MAX).
*/
if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
/* Potential overflow, we know nothing */
__mark_reg64_unbounded(dst_reg);
return;
}This restriction is done to avoid unsigned overflow, which could otherwise wrap the result around 0, and leave an unsound output where umin > umax. We also observe that limiting these u64 input ranges to [0, U32_MAX] leads to a loss of precision. Consider the case where the u64 bounds of dst_reg are [0, 2^34] and the u64 bounds of src_reg are [0, 2^2]. While the multiplication of these two bounds doesn't overflow and is sound [0, 2^36], the current scalar_min_max_mul() would set the entire register state to unbounded.
The key idea of our patch is that if there’s no possibility of overflow, we can multiply the unsigned bounds; otherwise, we set the 64-bit bounds to [0, U64_MAX], marking them as unbounded.
if (check_mul_overflow(*dst_umax, src_reg->umax_value, dst_umax) ||
(check_mul_overflow(*dst_umin, src_reg->umin_value, dst_umin))) {
/* Overflow possible, we know nothing */
dst_reg->umin_value = 0;
dst_reg->umax_value = U64_MAX;
}...
Now, to update the signed bounds based on the unsigned bounds, we need to ensure that the unsigned bounds don't cross the signed boundary (i.e., if ((s64)reg->umin_value <= (s64)reg->umax_value)). We observe that this is done anyway by __reg_deduce_bounds later, so we can just set signed bounds to unbounded [S64_MIN, S64_MAX]. Deferring the assignment of s64 bounds to reg_bounds_sync removes the current redundancy in scalar_min_max_mul(), which currently sets the s64 bounds based on the u64 bounds only in the case where umin <= umax <= 2^(63)-1.
Below, we provide an example BPF program (below) that exhibits the imprecision in the current BPF_MUL, where the outputs are all unbounded. In contrast, the updated BPF_MUL produces a bounded register state:
BPF_LD_IMM64(BPF_REG_1, 11),BPF_LD_IMM64(BPF_REG_2, 4503599627370624),BPF_ALU64_IMM(BPF_NEG, BPF_REG_2, 0),BPF_ALU64_IMM(BPF_NEG, BPF_REG_2, 0),BPF_ALU64_REG(BPF_AND, BPF_REG_1, BPF_REG_2),BPF_LD_IMM64(BPF_REG_3, 809591906117232263),BPF_ALU64_REG(BPF_MUL, BPF_REG_3, BPF_REG_1),BPF_MOV64_IMM(BPF_REG_0, 1),BPF_EXIT_INSN(),
Verifier log using the old BPF_MUL:
func#0 @0
0: R1=ctx() R10=fp0
0: (18) r1 = 0xb ; R1_w=11 2: (18) r2 = 0x10000000000080 ; R2_w=0x10000000000080 4: (87) r2 = -r2 ; R2_w=scalar() 5: (87) r2 = -r2 ; R2_w=scalar() 6: (5f) r1 &= r2 ; R1_w=scalar(smin=smin32=0,smax=umax=smax32=umax32=11,var_off=(0x0; 0xb)) R2_w=scalar() 7: (18) r3 = 0xb3c3f8c99262687 ; R3_w=0xb3c3f8c99262687 9: (2f) r3 *= r1 ; R1_w=scalar(smin=smin32=0,smax=umax=smax32=umax32=11,var_off=(0x0; 0xb)) R3_w=scalar()
...
Verifier using the new updated BPF_MUL (more precise bounds at label 9)
func#0 @0
0: R1=ctx() R10=fp0
0: (18) r1 = 0xb ; R1_w=11 2: (18) r2 = 0x10000000000080 ; R2_w=0x10000000000080 4: (87) r2 = -r2 ; R2_w=scalar() 5: (87) r2 = -r2 ; R2_w=scalar() 6: (5f) r1 &= r2 ; R1_w=scalar(smin=smin32=0,smax=umax=smax32=umax32=11,var_off=(0x0; 0xb)) R2_w=scalar() 7: (18) r3 = 0xb3c3f8c99262687 ; R3_w=0xb3c3f8c99262687 9: (2f) r3 *= r1 ; R1_w=scalar(smin=smin32=0,smax=umax=smax32=umax32=11,var_off=(0x0; 0xb)) R3_w=scalar(smin=0,smax=umax=0x7b96bb0a94a3a7cd,var_off=(0x0; 0x7fffffffffffffff))
...
Finally, we proved the soundness of the new scalar_min_max_mul() and scalar32_min_max_mul() functions. Typically, multiplication operations are expensive to check with bitvector-based solvers. We were able to prove the soundness of these functions using Non-Linear Integer Arithmetic (NIA) theory. Additionally, using Agni [2,3], we obtained the encodings for scalar32_min_max_mul() and scalar_min_max_mul() in bitvector theory, and were able to prove their soundness using 16-bit bitvectors (instead of 64-bit bitvectors that the functions actually use).
In conclusion, with this patch,
1. We were able to show that we can improve the overall precision of BPF_MUL. We proved (using an SMT solver) that this new version of BPF_MUL is at least as precise as the current version for all inputs.
2. We are able to prove the soundness of the new scalar_min_max_mul() and scalar32_min_max_mul(). By leveraging the existing proof of tnum_mul [1], we can say that the composition of these three functions within BPF_MUL is sound.
[1] https://ieeexplore.ieee.org/abstract/document/9741267
[2] https://link.springer.com/chapter/10.1007/978-3-031-37709-9_12
[3] https://people.cs.rutgers.edu/~sn349/papers/sas24-preprint.pdf
--- kernel/bpf/verifier.c | 72 +++++++++++++++---------------------------- 1 file changed, 24 insertions(+), 48 deletions(-)
kernel/bpf/verifier.c +24 -48
@@ -13827,65 +13827,41 @@ static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,static void scalar32_min_max_mul (struct bpf_reg_state *dst_reg,struct bpf_reg_state *src_reg) { -s32 smin_val = src_reg->s32_min_value; -u32 umin_val = src_reg->u32_min_value; -u32 umax_val = src_reg->u32_max_value; +u32 *dst_umin = &dst_reg->u32_min_value; +u32 *dst_umax = &dst_reg->u32_max_value; -if (smin_val <0 || dst_reg->s32_min_value <0 ) { -/* Ain't nobody got time to multiply that sign */ -__mark_reg32_unbounded (dst_reg); -return ; - } -/* Both values are positive, so we can work with unsigned and -* copy the result to signed (unless it exceeds S32_MAX). -*/ -if (umax_val >U16_MAX || dst_reg->u32_max_value >U16_MAX ) { -/* Potential overflow, we know nothing */ -__mark_reg32_unbounded (dst_reg); -return ; - } - dst_reg->u32_min_value *= umin_val; - dst_reg->u32_max_value *= umax_val; -if (dst_reg->u32_max_value >S32_MAX ) { +if (check_mul_overflow (*dst_umax, src_reg->u32_max_value, dst_umax) || +check_mul_overflow (*dst_umin, src_reg->u32_min_value, dst_umin)) {/* Overflow possible, we know nothing */ - dst_reg->s32_min_value =S32_MIN ; - dst_reg->s32_max_value =S32_MAX ; - }else { - dst_reg->s32_min_value = dst_reg->u32_min_value; - dst_reg->s32_max_value = dst_reg->u32_max_value; + dst_reg->u32_min_value =0 ; + dst_reg->u32_max_value =U32_MAX ; } + +/* Set signed bounds to unbounded and improve precision in +* reg_bounds_sync() +*/ + dst_reg->s32_min_value =S32_MIN ; + dst_reg->s32_max_value =S32_MAX ; }static void scalar_min_max_mul (struct bpf_reg_state *dst_reg,struct bpf_reg_state *src_reg) { -s64 smin_val = src_reg->smin_value; -u64 umin_val = src_reg->umin_value; -u64 umax_val = src_reg->umax_value; +u64 *dst_umin = &dst_reg->umin_value; +u64 *dst_umax = &dst_reg->umax_value; -if (smin_val <0 || dst_reg->smin_value <0 ) { -/* Ain't nobody got time to multiply that sign */ -__mark_reg64_unbounded (dst_reg); -return ; - } -/* Both values are positive, so we can work with unsigned and -* copy the result to signed (unless it exceeds S64_MAX). -*/ -if (umax_val >U32_MAX || dst_reg->umax_value >U32_MAX ) { -/* Potential overflow, we know nothing */ -__mark_reg64_unbounded (dst_reg); -return ; - } - dst_reg->umin_value *= umin_val; - dst_reg->umax_value *= umax_val; -if (dst_reg->umax_value >S64_MAX ) { +if (check_mul_overflow (*dst_umax, src_reg->umax_value, dst_umax) || +check_mul_overflow (*dst_umin, src_reg->umin_value, dst_umin)) {/* Overflow possible, we know nothing */ - dst_reg->smin_value =S64_MIN ; - dst_reg->smax_value =S64_MAX ; - }else { - dst_reg->smin_value = dst_reg->umin_value; - dst_reg->smax_value = dst_reg->umax_value; + dst_reg->umin_value =0 ; + dst_reg->umax_value =U64_MAX ; } + +/* Set signed bounds to unbounded and improve precision in +* reg_bounds_sync() +*/ + dst_reg->smin_value =S64_MIN ; + dst_reg->smax_value =S64_MAX ; } static void scalar32_min_max_and(struct bpf_reg_state *dst_reg,
Signature · 1 line
2.25.1
Eduard Zingerman
maintainerOn Wed, 2024-11-27 at 02:41 -0500, Matan Shachnai wrote:
[...]
Quoted 10 lines
In conclusion, with this patch,
1. We were able to show that we can improve the overall precision of
BPF_MUL. We proved (using an SMT solver) that this new version ofBPF_MULis at least as precise as the current version for all inputs.2. We are able to prove the soundness of the new
scalar_min_max_mul()andscalar32_min_max_mul(). By leveraging the existing proof oftnum_mul[1], we can say that the composition of these three functions withinBPF_MULis sound.
Hi Matan,
I think this is a nice simplification of the existing code. Could you please also add a few canary tests in the tools/testing/selftests/bpf/progs/verifier_bounds.c ? (e.g. simple case plus possible edge cases). Something like:
SEC("tc")
__success __log_level(2)
__msg("r6 *= r7 {{.*}}; R6_w=some-range-here")
__naked void mult_mixed_sign(void)
{
asm volatile (
"call %[bpf_get_prandom_u32];"
"r6 = r0;"
"call %[bpf_get_prandom_u32];"
"r7 = r0;"
"r6 &= 0xf;"
"r6 -= 1000000000;"
"r7 &= 0xf;"
"r7 -= 2000000000;"
"r6 *= r7;"
"exit"
:
: __imm(bpf_get_prandom_u32),
__imm(bpf_skb_store_bytes)
: __clobber_all);
}We usually do this as a separate patch in a patch-set.
Also, it looks like this has limited applicability in practice, because small negative values denote huge unsigned values, hence overflow check kicks in for such values.
E.g. no range inferred for [-10,5] * [-20,-5]:
0: (85) call bpf_get_prandom_u32#7 ; R0_w=scalar() 1: (bf) r6 = r0 ; R0_w=scalar(id=1) R6_w=scalar(id=1) 2: (85) call bpf_get_prandom_u32#7 ; R0_w=scalar() 3: (bf) r7 = r0 ; R0_w=scalar(id=2) R7_w=scalar(id=2) 4: (57) r6 &= 15 ; R6_w=scalar(smin=smin32=0,smax=umax=smax32=umax32=15,var_off=(0x0; 0xf)) 5: (17) r6 -= 10 ; R6_w=scalar(smin=smin32=-10,smax=smax32=5) 6: (57) r7 &= 15 ; R7_w=scalar(smin=smin32=0,smax=umax=smax32=umax32=15,var_off=(0x0; 0xf)) 7: (17) r7 -= 20 ; R7_w=scalar(smin=smin32=-20,smax=smax32=-5,umin=0xffffffffffffffec,umax=0xfffffffffffffffb,umin32=0xffffffec,umax32=0xfffffffb,var_off=(0xffffffffffffffe0; 0x1f)) 8: (2f) r6 *= r7 ; R6_w=scalar() R7_w=scalar(smin=smin32=-20,smax=smax32=-5,umin=0xffffffffffffffec,umax=0xfffffffffffffffb,umin32=0xffffffec,umax32=0xfffffffb,var_off=(0xffffffffffffffe0; 0x1f))
9: (95) exit
Compared to:
0: R1=ctx() R10=fp0
; asm volatile ( @ verifier_bounds.c:1208
0: (85) call bpf_get_prandom_u32#7 ; R0_w=scalar() 1: (bf) r6 = r0 ; R0_w=scalar(id=1) R6_w=scalar(id=1) 2: (85) call bpf_get_prandom_u32#7 ; R0_w=scalar() 3: (bf) r7 = r0 ; R0_w=scalar(id=2) R7_w=scalar(id=2) 4: (57) r6 &= 15 ; R6_w=scalar(smin=smin32=0,smax=umax=smax32=umax32=15,var_off=(0x0; 0xf)) 5: (17) r6 -= 1000000000 ; R6_w=scalar(smin=0xffffffffc4653600,smax=0xffffffffc465360f,umin=0xffffffffc4653600,umax=0xffffffffc465360f,smin32=umin32=0xc4653600,smax32=umax32=0xc465360f,var_off=(0xffffffffc4653600; 0xf)) 6: (57) r7 &= 15 ; R7_w=scalar(smin=smin32=0,smax=umax=smax32=umax32=15,var_off=(0x0; 0xf)) 7: (17) r7 -= 2000000000 ; R7_w=scalar(smin=0xffffffff88ca6c00,smax=0xffffffff88ca6c0f,umin=0xffffffff88ca6c00,umax=0xffffffff88ca6c0f,smin32=umin32=0x88ca6c00,smax32=umax32=0x88ca6c0f,var_off=(0xffffffff88ca6c00; 0xf)) 8: (2f) r6 *= r7 ; R6_w=scalar(smax=0x7ffffffffffffeff,umax=0xfffffffffffffeff,smax32=0x7ffffeff,umax32=0xfffffeff,var_off=(0x0; 0xfffffffffffffeff)) R7_w=scalar(smin=0xffffffff88ca6c00,smax=0xffffffff88ca6c0f,umin=0xffffffff88ca6c00,umax=0xffffffff88ca6c0f,smin32=umin32=0x88ca6c00,smax32=umax32=0x88ca6c0f,var_off=(0xffffffff88ca6c00; 0xf))
9: (95) exit
Is it possible to do check_mul_overflow() for signed bounds and rely on reg_bounds_sync() for unsigned?
[...]
Quoted 16 lines · Eduard Zingerman
On Wed, 2024-11-27 at 02:41 -0500, Matan Shachnai wrote:
[...]
Quoted 10 lines
In conclusion, with this patch,
1. We were able to show that we can improve the overall precision of
BPF_MUL. We proved (using an SMT solver) that this new version ofBPF_MULis at least as precise as the current version for all inputs.2. We are able to prove the soundness of the new
scalar_min_max_mul()andscalar32_min_max_mul(). By leveraging the existing proof oftnum_mul[1], we can say that the composition of these three functions withinBPF_MULis sound.
Quoted 6 lines
Hi Matan,
I think this is a nice simplification of the existing code. Could you please also add a few canary tests in the
tools/testing/selftests/bpf/progs/verifier_bounds.c? (e.g. simple case plus possible edge cases).
Thanks for your feedback, Eduard! We'll be happy to add test-cases to exercise BPF_MUL.
Quoted 55 lines
Something like:
SEC("tc") __success __log_level(2) __msg("r6 *= r7 {{.*}}; R6_w=some-range-here") __naked void mult_mixed_sign(void) { asm volatile ( "call %[bpf_get_prandom_u32];" "r6 = r0;" "call %[bpf_get_prandom_u32];" "r7 = r0;" "r6 &= 0xf;" "r6 -= 1000000000;" "r7 &= 0xf;" "r7 -= 2000000000;" "r6 *= r7;" "exit" : : __imm(bpf_get_prandom_u32), __imm(bpf_skb_store_bytes) : __clobber_all); }We usually do this as a separate patch in a patch-set.
Also, it looks like this has limited applicability in practice, because small negative values denote huge unsigned values, hence overflow check kicks in for such values.
E.g. no range inferred for [-10,5] * [-20,-5]:0: (85) call bpf_get_prandom_u32#7 ; R0_w=scalar() 1: (bf) r6 = r0 ; R0_w=scalar(id=1) R6_w=scalar(id=1) 2: (85) call bpf_get_prandom_u32#7 ; R0_w=scalar() 3: (bf) r7 = r0 ; R0_w=scalar(id=2) R7_w=scalar(id=2) 4: (57) r6 &= 15 ; R6_w=scalar(smin=smin32=0,smax=umax=smax32=umax32=15,var_off=(0x0; 0xf)) 5: (17) r6 -= 10 ; R6_w=scalar(smin=smin32=-10,smax=smax32=5) 6: (57) r7 &= 15 ; R7_w=scalar(smin=smin32=0,smax=umax=smax32=umax32=15,var_off=(0x0; 0xf)) 7: (17) r7 -= 20 ; R7_w=scalar(smin=smin32=-20,smax=smax32=-5,umin=0xffffffffffffffec,umax=0xfffffffffffffffb,umin32=0xffffffec,umax32=0xfffffffb,var_off=(0xffffffffffffffe0; 0x1f)) 8: (2f) r6 *= r7 ; R6_w=scalar() R7_w=scalar(smin=smin32=-20,smax=smax32=-5,umin=0xffffffffffffffec,umax=0xfffffffffffffffb,umin32=0xffffffec,umax32=0xfffffffb,var_off=(0xffffffffffffffe0; 0x1f))9: (95) exit
Compared to:
0: R1=ctx() R10=fp0
; asm volatile ( @ verifier_bounds.c:12080: (85) call bpf_get_prandom_u32#7 ; R0_w=scalar() 1: (bf) r6 = r0 ; R0_w=scalar(id=1) R6_w=scalar(id=1) 2: (85) call bpf_get_prandom_u32#7 ; R0_w=scalar() 3: (bf) r7 = r0 ; R0_w=scalar(id=2) R7_w=scalar(id=2) 4: (57) r6 &= 15 ; R6_w=scalar(smin=smin32=0,smax=umax=smax32=umax32=15,var_off=(0x0; 0xf)) 5: (17) r6 -= 1000000000 ; R6_w=scalar(smin=0xffffffffc4653600,smax=0xffffffffc465360f,umin=0xffffffffc4653600,umax=0xffffffffc465360f,smin32=umin32=0xc4653600,smax32=umax32=0xc465360f,var_off=(0xffffffffc4653600; 0xf)) 6: (57) r7 &= 15 ; R7_w=scalar(smin=smin32=0,smax=umax=smax32=umax32=15,var_off=(0x0; 0xf)) 7: (17) r7 -= 2000000000 ; R7_w=scalar(smin=0xffffffff88ca6c00,smax=0xffffffff88ca6c0f,umin=0xffffffff88ca6c00,umax=0xffffffff88ca6c0f,smin32=umin32=0x88ca6c00,smax32=umax32=0x88ca6c0f,var_off=(0xffffffff88ca6c00; 0xf)) 8: (2f) r6 *= r7 ; R6_w=scalar(smax=0x7ffffffffffffeff,umax=0xfffffffffffffeff,smax32=0x7ffffeff,umax32=0xfffffeff,var_off=(0x0; 0xfffffffffffffeff)) R7_w=scalar(smin=0xffffffff88ca6c00,smax=0xffffffff88ca6c0f,umin=0xffffffff88ca6c00,umax=0xffffffff88ca6c0f,smin32=umin32=0x88ca6c00,smax32=umax32=0x88ca6c0f,var_off=(0xffffffff88ca6c00; 0xf))
Quoted 5 lines
9: (95) exit
Is it possible to do
check_mul_overflow()for signed bounds and rely onreg_bounds_sync()for unsigned?
The patch in its current form (and the existing BPF_MUL version in the verifier) doesn't handle negative values well, as the example you gave here illustrates. The initial goal of this patch was to improve precision of unsigned multiplication. However, there is a canonical way to perform signed multiplication which is sound and is able to handle negative values. Specifically, signed multiplication can be performed soundly by [min(a, b, c, d), max(a, b, c, d)], where a, b, c, d correspond to the four products obtained by multiplying all the bounds (these products are checked for overflows). For better precision, we propose having both unsigned multiplication as well as signed multiplication. The resulting bounds can then be refined in reg_bounds_sync().
We will update our patch with both signed and unsigned multiplication, add test-cases, and send it all as a patch-set soon.
Best,
Matan
Quoted 2 lines
[...]