On 3/18/2023 6:24 AM, Daniel Borkmann wrote:
On 3/14/23 9:34 PM, Xu Kuohai wrote:
From: Xu Kuohai <xukuohai@xxxxxxxxxx>
After commit 3f50f132d840 ("bpf: Verifier, do explicit ALU32 bounds tracking"),
the following bpf prog is rejected:
0: (61) r2 = *(u32 *)(r1 +0) ; R2_w=pkt(off=0,r=0,imm=0)
1: (61) r3 = *(u32 *)(r1 +4) ; R3_w=pkt_end(off=0,imm=0)
2: (bf) r1 = r2
3: (07) r1 += 1
4: (2d) if r1 > r3 goto pc+8
5: (71) r1 = *(u8 *)(r2 +0) ; R1_w=scalar(umax=255,var_off=(0x0; 0xff))
6: (18) r0 = 0x7fffffffffffff10
8: (0f) r1 += r0 ; R1_w=scalar(umin=0x7fffffffffffff10,umax=0x800000000000000f)
9: (18) r0 = 0x8000000000000000
11: (07) r0 += 1
12: (ad) if r0 < r1 goto pc-2
13: (b7) r0 = 0
14: (95) exit
And the verifier log says:
[...]
from 12 to 11: R0_w=-9223372036854775794 R1=scalar(umin=9223372036854775823,umax=9223372036854775823,var_off=(0x8000000000000000; 0xffffffff))
11: (07) r0 += 1 ; R0_w=-9223372036854775793
12: (ad) if r0 < r1 goto pc-2 ; R0_w=-9223372036854775793 R1=scalar(umin=9223372036854775823,umax=9223372036854775823,var_off=(0x8000000000000000; 0xffffffff))
13: safe
from 12 to 11: R0_w=-9223372036854775793 R1=scalar(umin=9223372036854775824,umax=9223372036854775823,var_off=(0x8000000000000000; 0xffffffff))
11: (07) r0 += 1 ; R0_w=-9223372036854775792
12: (ad) if r0 < r1 goto pc-2 ; R0_w=-9223372036854775792 R1=scalar(umin=9223372036854775824,umax=9223372036854775823,var_off=(0x8000000000000000; 0xffffffff))
13: safe
[...]
What can be seen here is that r1->umin grows blindly and becomes bigger
than r1->umax. The reason is because the loop does not terminate, when
r0 increases to r1->umax_value, the following code in reg_set_min_max()
sets r1->umin_value to r1->umax_value + 1 blindly:
case BPF_JGT:
{
if (is_jmp32) {
[...]
} else {
u64 false_umax = opcode == BPF_JGT ? val : val - 1;
u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
false_reg->umax_value = min(false_reg->umax_value, false_umax);
true_reg->umin_value = max(true_reg->umin_value, true_umin);
}
break;
}
Why the loop does not terminate is because tnum_is_const(src_reg->var_off)
always returns false, causing is_branch_taken() to be skipped:
if (src_reg->type == SCALAR_VALUE &&
!is_jmp32 && tnum_is_const(src_reg->var_off)) {
pred = is_branch_taken(dst_reg, // could not reach here
src_reg->var_off.value,
opcode,
is_jmp32);
}
Why tnum_is_const(src_reg->var_off) always returns false is because
r1->umin_value starts increasing from 0x7fffffffffffff10, always bigger
than U32_MAX, causing the __reg_combine_64_into_32() to mark the lower
32 bits unbounded, i.e. not a constant.
To fix it:
1. avoid increasing reg lower bound to a value bigger than the upper bound,
or decreasing reg upper bound to a value smaller than the lower bound.
2. set 32-bit min/max values to the lower 32 bits of the 64-bit min/max values
when the 64-bit min/max values are equal.
Should both these be separate patches, meaning are both of them strictly
required as one logical entity or not? From your description it's not really
clear wrt reg_{inc,dec}_{u32,u64}_{min,max} and if this is mainly defensive
or required.
It's defensive, not required. It was added to address "umin > umax" concerns
received in v1. Maybe I misunderstood something.
Also, while you describe to some degree how we get here, there is no analysis
on why your proposed changes are safe. If you want to make the verifier less
conservative to start accepting such progs, can you then elaborate on the latter?
I have some discussion below and hope it makes some sense.
Fixes: 3f50f132d840 ("bpf: Verifier, do explicit ALU32 bounds tracking")
Signed-off-by: Xu Kuohai <xukuohai@xxxxxxxxxx>
---
kernel/bpf/verifier.c | 143 +++++++++++++++++++++++++++---------------
1 file changed, 93 insertions(+), 50 deletions(-)
diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
index 2bbd89279070..b775b50353d6 100644
--- a/kernel/bpf/verifier.c
+++ b/kernel/bpf/verifier.c
@@ -2223,14 +2223,21 @@ static bool __reg64_bound_u32(u64 a)
static void __reg_combine_64_into_32(struct bpf_reg_state *reg)
{
+ s64 smin = reg->smin_value;
+ s64 smax = reg->smax_value;
+ u64 umin = reg->umin_value;
+ u64 umax = reg->umax_value;
+
__mark_reg32_unbounded(reg);
- if (__reg64_bound_s32(reg->smin_value) && __reg64_bound_s32(reg->smax_value)) {
- reg->s32_min_value = (s32)reg->smin_value;
- reg->s32_max_value = (s32)reg->smax_value;
+ if ((__reg64_bound_s32(smin) && __reg64_bound_s32(smax)) ||
+ smin == smax) {
Did you look into debugging reg_bounds_sync()? Assumption for constant
register is register_is_const() and it's explicitly looking at var_off.
I checked this function and saw that the low 32 bits and high 32 bits of
var_off are updated separately, and the new low/high 32 bits are the
intersection of the old low/high 32 bits and the new [u32_min, u32_max]
or [umin, umax] range. So as the range [u32_min, u32_max] or [umin, umax]
converges, the low/high 32 bits converge as well.
This converge process is reasonable. The lower 32 bits of var_off do not
converge to constant is because the range [u32_min, u32_max] does not
converge to constant.
IIUC, what _reg_combine_64_into_32 does before calling reg_bounds_sync is
find the value range of y when the 64-bit variable x changes from smin to
smax for function y = (s32)x, and the value range of y when x changes from
umin to umax for function y = (u32)x.
So when the value range of x is a constant, the value range of y is naturally
also a constant, that is, it's reasonable to infer u32_min == u32_max from
the 64-bit umin == umax.
In fact, the range of y could be found from function graph directly.
1. function graph for y = (u32)x:
^ y = (u32)x
|
U32_MAX + _ _
| _/ | _| |
| _| | _| |
| _| | _| |
| _| | _| |
| _| | _| |
| _| | _| |
| _| | _| |
| _| | _| |
| | | | |
0 +-------------------+-------------------+----> x/u64
|<---U32_MAX + 1--->|<---U32_MAX + 1--->|
1A) if umin == umax, then u32_min = (u32)umin, u32_max = (u32)umax:
^ y = (u32)x
|
U32_MAX + _ _
| _/ | _| |
| _| | _| |
| _| | _| |
u32_min == u32_max | _ _ _ _ _| | _| |
| _| | _| |
| _| | | _| |
| _| | | _| |
| _| | | _| |
| | | | | |
0 +---------+---------+-------------------+----> x/u64
umin == umax
|<---U32_MAX + 1--->|<---U32_MAX + 1--->|
1B) if umax - umin <= U32_MAX and (u32)umin < (u32)umax, then
u32_min = (u32)umin, u32_max = (u32)umax:
^ y = (u32)x
|
U32_MAX + _ _
u32_max = (u32)umax | _ _ _ _ _ _ _ _/ | _| |
| _| | _| |
| _| | | _| |
u32_min = (u32)umin | _ _ _ _ _| | | _| |
| _| | | _| |
| _| | | | _| |
| _| | | | _| |
| _| | | | _| |
| | | | | | |
0 +---------+-----+---+-------------------+----> x/u64
umin umax
|<---U32_MAX + 1--->|<---U32_MAX + 1--->|
1C) if umax - umin <= U32_MAX and (u32)umin > (u32)umax, then
u32_min = U32_MIN, u32_max = U32_MAX:
^ y = (u32)x
|
u32_max = 32_MAX + _ _
| _/ | _| |
| _| | _| |
| _| | _| |
(u32)umin | _ _ _ _ _| | _| |
| _| | _| |
| _| | | _| |
(u32)umax | _ _| _ |_ _ _ _ _|_ _ _| |
| _| | | _| | |
| | | | | | |
u32_min = U32_MIN 0 +---------+---------+-----+-------------+----> x/u64
umin umax
|<---U32_MAX + 1--->|<---U32_MAX + 1--->|
1D) if umax - umin > U32_MAX, then u32_min = U32_MIN, u32_max = U32_MAX:
^ y = (u32)x
|
u32_max = 32_MAX + _ _
| _/ | _| |
(u32)umax | _ _ _ _ _ _ _|_ _|_ _ _ _ _ _ _ _| |
| _| | _| |
(u32)umin | _ _ _ _ _| | _| | |
| _| | _| | |
| _| | | _| | |
| _| | | _| | |
| _| | | _| | |
| | | | | | |
u32_min = U32_MIN 0 +---------+---------+-------------+-----+----> x/u64
umin umax
|<---U32_MAX + 1--->|<---U32_MAX + 1--->|
2. function graph for y = (s32)x:
^ y = (s32)x
|
S32_MAX | _ _ _ _ _ _
| _| | _| |
| _| | _| |
| _| | _| |
| _| | _| |
--+---------0----------+---------+----------+---> x/s64
| _| | _|
| _| | | _|
| _| | | _|
| _| | | _|
|_| _ _ _ | S32_MIN |_|
|
|
|<--- U32_MAX + 1--->|<--- U32_MAX + 1--->|
2A) if smin == smax, then s32_min = (s32)smin, s32_max = (s32)smax:
^ y = (s32)x
|
S32_MAX | _ _ _ _ _ _
| _| | _| |
| _| | _| |
s32_min == s32_max|__ _| | _| |
| _| | | _| |
--+---------0----+-----+---------+----------+---> x/s64
| _| smin | _|
| _| | smax | _|
| _| | | _|
| _| | | _|
|_| _ _ _ | S32_MIN |_|
|
|
|<--- U32_MAX + 1--->|<--- U32_MAX + 1--->|
2B) if smax - smin <= U32_MAX and (s32)smin < (s32)smax, then
s32_min = (u32)smin, s32_max = (s32)smax:
^ y = (s32)x
|
S32_MAX | _ _ _ _ _ _
s32_max | _ _ _ _| | _| |
| _| | _| |
| _| | | _| |
| _| | | _| |
--+---+-----0------+---+---------+----------+---> x/s64
| smin _| smax | _|
| | _| | | _|
| |_|_ _| s32_min | _|
| _| | | _|
|_| _ _ _ | S32_MIN |_|
|
|
|<--- U32_MAX + 1--->|<--- U32_MAX + 1--->|
2C) if smax - smin <= U32_MAX and (s32)smin > (s32)smax, then
s32_min = S32_MIN, s32_max = S32_MAX:
^ y = (s32)x
|
s32_max = S32_MAX | _ _ _ _ _ _
| _| | _| |
(s32)smin |_ _ _| | _| |
| _| | _| |
| _| | | smax _| |
--+---------0----+-----+-----+---+----------+---> x/s64
| smin | | _|
(s32)smax | _ _ _ _ _|_ _ _|_|
| | _|
| | _|
s32_min = S32_MIN | _ _ _ _ _|_|
|
|
|<--- U32_MAX + 1--->|<--- U32_MAX + 1--->|
2D) if smax - smin > U32_MAX, then s32_min = S32_MIN, s32_max = S32_MAX:
^ y = (s32)x
|
s32_max = S32_MAX | _ _ _ _ _ _
(s32)smax | _ _ _ _|_|_ _ _ _ _ _ _ _ _| |
| _| | _| |
(s32)smin | _ _| | _| | |
| _| | _| | |
--+---------0--+-------+---------+------+---+---> x/s64
| smin | _| smax
| | _|
| | _|
| | _|
s32_min = S32_MIN | _ _ _ _ _|_|
|
|
|<--- U32_MAX + 1--->|<--- U32_MAX + 1--->|
+ reg->s32_min_value = (s32)smin;
+ reg->s32_max_value = (s32)smax;
}
- if (__reg64_bound_u32(reg->umin_value) && __reg64_bound_u32(reg->umax_value)) {
- reg->u32_min_value = (u32)reg->umin_value;
- reg->u32_max_value = (u32)reg->umax_value;
+ if ((__reg64_bound_u32(umin) && __reg64_bound_u32(umax)) ||
+ umin == umax) {
+ reg->u32_min_value = (u32)umin;
+ reg->u32_max_value = (u32)umax;
}
reg_bounds_sync(reg);
}
@@ -12828,6 +12835,62 @@ static int is_pkt_ptr_branch_taken(struct bpf_reg_state *dst_reg,
return -1;
}
+static void reg_inc_u32_min(struct bpf_reg_state *reg, u32 val)
+{
+ reg->u32_min_value = max(reg->u32_min_value, val);
+ if (reg->u32_min_value > reg->u32_max_value)
+ reg->u32_min_value = reg->u32_max_value;
+}
+
+static void reg_dec_u32_max(struct bpf_reg_state *reg, u32 val)
+{
+ reg->u32_max_value = min(reg->u32_max_value, val);
+ if (reg->u32_max_value < reg->u32_min_value)
+ reg->u32_max_value = reg->u32_min_value;
+}
+
+static void reg_inc_s32_min(struct bpf_reg_state *reg, s32 val)
+{
+ reg->s32_min_value = max(reg->s32_min_value, val);
+ if (reg->s32_min_value > reg->s32_max_value)
+ reg->s32_min_value = reg->s32_max_value;
+}
+
+static void reg_dec_s32_max(struct bpf_reg_state *reg, s32 val)
+{
+ reg->s32_max_value = min(reg->s32_max_value, val);
+ if (reg->s32_max_value < reg->s32_min_value)
+ reg->s32_max_value = reg->s32_min_value;
+}
+
+static void reg_inc_u64_min(struct bpf_reg_state *reg, u64 val)
+{
+ reg->umin_value = max(reg->umin_value, val);
+ if (reg->umin_value > reg->umax_value)
+ reg->umin_value = reg->umax_value;
+}
+
+static void reg_dec_u64_max(struct bpf_reg_state *reg, u64 val)
+{
+ reg->umax_value = min(reg->umax_value, val);
+ if (reg->umax_value < reg->umin_value)
+ reg->umax_value = reg->umin_value;
+}
+
+static void reg_inc_s64_min(struct bpf_reg_state *reg, s64 val)
+{
+ reg->smin_value = max(reg->smin_value, val);
+ if (reg->smin_value > reg->smax_value)
+ reg->smin_value = reg->smax_value;
+}
+
+static void reg_dec_s64_max(struct bpf_reg_state *reg, s64 val)
+{
+ reg->smax_value = min(reg->smax_value, val);
+ if (reg->smax_value < reg->smin_value)
+ reg->smax_value = reg->smin_value;
+}
All this feels more like a workaround and papering over the issue.
Agree, making umin > umax disappear from error log does not fix the
issue, and may make debugging more difficult.
/* Adjusts the register min/max values in the case that the dst_reg is the
* variable register that we are working on, and src_reg is a constant or we're
* simply doing a BPF_K check.
@@ -12898,76 +12961,56 @@ static void reg_set_min_max(struct bpf_reg_state *true_reg,
case BPF_JGE:
case BPF_JGT:
{
- if (is_jmp32) {
- u32 false_umax = opcode == BPF_JGT ? val32 : val32 - 1;
- u32 true_umin = opcode == BPF_JGT ? val32 + 1 : val32;
+ bool neq = (opcode == BPF_JGT);
- false_reg->u32_max_value = min(false_reg->u32_max_value,
- false_umax);
- true_reg->u32_min_value = max(true_reg->u32_min_value,
- true_umin);
+ if (is_jmp32) {
+ reg_dec_u32_max(false_reg, neq ? val32 : val32 - 1);
+ reg_inc_u32_min(true_reg, neq ? val32 + 1 : val32);
} else {
- u64 false_umax = opcode == BPF_JGT ? val : val - 1;
- u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
-
- false_reg->umax_value = min(false_reg->umax_value, false_umax);
- true_reg->umin_value = max(true_reg->umin_value, true_umin);
+ reg_dec_u64_max(false_reg, neq ? val : val - 1);
+ reg_inc_u64_min(true_reg, neq ? val + 1 : val);
.
