bpf: Fix insufficient bounds propagation from adjust_scalar_min_max_vals

Kuee reported a corner case where the tnum becomes constant after the call
to __reg_bound_offset(), but the register's bounds are not, that is, its
min bounds are still not equal to the register's max bounds.

This in turn allows to leak pointers through turning a pointer register as
is into an unknown scalar via adjust_ptr_min_max_vals().

Before:

func#0 @0
0: R1=ctx(off=0,imm=0,umax=0,var_off=(0x0; 0x0)) R10=fp(off=0,imm=0,umax=0,var_off=(0x0; 0x0))
0: (b7) r0 = 1 ; R0_w=scalar(imm=1,umin=1,umax=1,var_off=(0x1; 0x0))
1: (b7) r3 = 0 ; R3_w=scalar(imm=0,umax=0,var_off=(0x0; 0x0))
2: (87) r3 = -r3 ; R3_w=scalar()
3: (87) r3 = -r3 ; R3_w=scalar()
4: (47) r3 |= 32767 ; R3_w=scalar(smin=-9223372036854743041,umin=32767,var_off=(0x7fff; 0xffffffffffff8000),s32_min=-2147450881)
5: (75) if r3 s>= 0x0 goto pc+1 ; R3_w=scalar(umin=9223372036854808575,var_off=(0x8000000000007fff; 0x7fffffffffff8000),s32_min=-2147450881,u32_min=32767)
6: (95) exit

from 5 to 7: R0=scalar(imm=1,umin=1,umax=1,var_off=(0x1; 0x0)) R1=ctx(off=0,imm=0,umax=0,var_off=(0x0; 0x0)) R3=scalar(umin=32767,umax=9223372036854775807,var_off=(0x7fff; 0x7fffffffffff8000),s32_min=-2147450881) R10=fp(off=0,imm=0,umax=0,var_off=(0x0; 0x0))
7: (d5) if r3 s<= 0x8000 goto pc+1 ; R3=scalar(umin=32769,umax=9223372036854775807,var_off=(0x7fff; 0x7fffffffffff8000),s32_min=-2147450881,u32_min=32767)
8: (95) exit

from 7 to 9: R0=scalar(imm=1,umin=1,umax=1,var_off=(0x1; 0x0)) R1=ctx(off=0,imm=0,umax=0,var_off=(0x0; 0x0)) R3=scalar(umin=32767,umax=32768,var_off=(0x7fff; 0x8000)) R10=fp(off=0,imm=0,umax=0,var_off=(0x0; 0x0))
9: (07) r3 += -32767 ; R3_w=scalar(imm=0,umax=1,var_off=(0x0; 0x0)) <--- [*]
10: (95) exit

What can be seen here is that R3=scalar(umin=32767,umax=32768,var_off=(0x7fff;
0x8000)) after the operation R3 += -32767 results in a 'malformed' constant, that
is, R3_w=scalar(imm=0,umax=1,var_off=(0x0; 0x0)). Intersecting with var_off has
not been done at that point via __update_reg_bounds(), which would have improved
the umax to be equal to umin.

Refactor the tnum <> min/max bounds information flow into a reg_bounds_sync()
helper and use it consistently everywhere. After the fix, bounds have been
corrected to R3_w=scalar(imm=0,umax=0,var_off=(0x0; 0x0)) and thus the register
is regarded as a 'proper' constant scalar of 0.

After:

func#0 @0
0: R1=ctx(off=0,imm=0,umax=0,var_off=(0x0; 0x0)) R10=fp(off=0,imm=0,umax=0,var_off=(0x0; 0x0))
0: (b7) r0 = 1 ; R0_w=scalar(imm=1,umin=1,umax=1,var_off=(0x1; 0x0))
1: (b7) r3 = 0 ; R3_w=scalar(imm=0,umax=0,var_off=(0x0; 0x0))
2: (87) r3 = -r3 ; R3_w=scalar()
3: (87) r3 = -r3 ; R3_w=scalar()
4: (47) r3 |= 32767 ; R3_w=scalar(smin=-9223372036854743041,umin=32767,var_off=(0x7fff; 0xffffffffffff8000),s32_min=-2147450881)
5: (75) if r3 s>= 0x0 goto pc+1 ; R3_w=scalar(umin=9223372036854808575,var_off=(0x8000000000007fff; 0x7fffffffffff8000),s32_min=-2147450881,u32_min=32767)
6: (95) exit

from 5 to 7: R0=scalar(imm=1,umin=1,umax=1,var_off=(0x1; 0x0)) R1=ctx(off=0,imm=0,umax=0,var_off=(0x0; 0x0)) R3=scalar(umin=32767,umax=9223372036854775807,var_off=(0x7fff; 0x7fffffffffff8000),s32_min=-2147450881) R10=fp(off=0,imm=0,umax=0,var_off=(0x0; 0x0))
7: (d5) if r3 s<= 0x8000 goto pc+1 ; R3=scalar(umin=32769,umax=9223372036854775807,var_off=(0x7fff; 0x7fffffffffff8000),s32_min=-2147450881,u32_min=32767)
8: (95) exit

from 7 to 9: R0=scalar(imm=1,umin=1,umax=1,var_off=(0x1; 0x0)) R1=ctx(off=0,imm=0,umax=0,var_off=(0x0; 0x0)) R3=scalar(umin=32767,umax=32768,var_off=(0x7fff; 0x8000)) R10=fp(off=0,imm=0,umax=0,var_off=(0x0; 0x0))
9: (07) r3 += -32767 ; R3_w=scalar(imm=0,umax=0,var_off=(0x0; 0x0)) <--- [*]
10: (95) exit

Fixes: b03c9f9fdc37 ("bpf/verifier: track signed and unsigned min/max values")
Reported-by: Kuee K1r0a <[e-mail]>
Signed-off-by: Daniel Borkmann <[e-mail]>
Signed-off-by: Andrii Nakryiko <[e-mail]>
Acked-by: John Fastabend <[e-mail]>
Link: https://lore.kernel.org/bpf/[e-mail]

kernel/bpf/verifier.c

···
 	reg->var_off = tnum_or(tnum_clear_subreg(var64_off), var32_off);
 }
 
+static void reg_bounds_sync(struct bpf_reg_state *reg)
+{
+	/* We might have learned new bounds from the var_off. */
+	__update_reg_bounds(reg);
+	/* We might have learned something about the sign bit. */
+	__reg_deduce_bounds(reg);
+	/* We might have learned some bits from the bounds. */
+	__reg_bound_offset(reg);
+	/* Intersecting with the old var_off might have improved our bounds
+	 * slightly, e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
+	 * then new var_off is (0; 0x7f...fc) which improves our umax.
+	 */
+	__update_reg_bounds(reg);
+}
+
 static bool __reg32_bound_s64(s32 a)
 {
 	return a >= 0 && a <= S32_MAX;
···
 		 * so they do not impact tnum bounds calculation.
 		 */
 		__mark_reg64_unbounded(reg);
-		__update_reg_bounds(reg);
 	}
-
-	/* Intersecting with the old var_off might have improved our bounds
-	 * slightly.  e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
-	 * then new var_off is (0; 0x7f...fc) which improves our umax.
-	 */
-	__reg_deduce_bounds(reg);
-	__reg_bound_offset(reg);
-	__update_reg_bounds(reg);
+	reg_bounds_sync(reg);
 }
 
 static bool __reg64_bound_s32(s64 a)
···
 static void __reg_combine_64_into_32(struct bpf_reg_state *reg)
 {
 	__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;
···
 		reg->u32_min_value = (u32)reg->umin_value;
 		reg->u32_max_value = (u32)reg->umax_value;
 	}
-
-	/* Intersecting with the old var_off might have improved our bounds
-	 * slightly.  e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
-	 * then new var_off is (0; 0x7f...fc) which improves our umax.
-	 */
-	__reg_deduce_bounds(reg);
-	__reg_bound_offset(reg);
-	__update_reg_bounds(reg);
+	reg_bounds_sync(reg);
 }
 
 /* Mark a register as having a completely unknown (scalar) value. */
···
 	ret_reg->s32_max_value = meta->msize_max_value;
 	ret_reg->smin_value = -MAX_ERRNO;
 	ret_reg->s32_min_value = -MAX_ERRNO;
-	__reg_deduce_bounds(ret_reg);
-	__reg_bound_offset(ret_reg);
-	__update_reg_bounds(ret_reg);
+	reg_bounds_sync(ret_reg);
 }
 
 static int
···
 
 	if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
 		return -EINVAL;
-
-	__update_reg_bounds(dst_reg);
-	__reg_deduce_bounds(dst_reg);
-	__reg_bound_offset(dst_reg);
-
+	reg_bounds_sync(dst_reg);
 	if (sanitize_check_bounds(env, insn, dst_reg) < 0)
 		return -EACCES;
 	if (sanitize_needed(opcode)) {
···
 	/* ALU32 ops are zero extended into 64bit register */
 	if (alu32)
 		zext_32_to_64(dst_reg);
-
-	__update_reg_bounds(dst_reg);
-	__reg_deduce_bounds(dst_reg);
-	__reg_bound_offset(dst_reg);
+	reg_bounds_sync(dst_reg);
 	return 0;
 }
 
···
 							 insn->dst_reg);
 				}
 				zext_32_to_64(dst_reg);
-
-				__update_reg_bounds(dst_reg);
-				__reg_deduce_bounds(dst_reg);
-				__reg_bound_offset(dst_reg);
+				reg_bounds_sync(dst_reg);
 			}
 		} else {
 			/* case: R = imm
···
 							dst_reg->smax_value);
 	src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
 							     dst_reg->var_off);
-	/* We might have learned new bounds from the var_off. */
-	__update_reg_bounds(src_reg);
-	__update_reg_bounds(dst_reg);
-	/* We might have learned something about the sign bit. */
-	__reg_deduce_bounds(src_reg);
-	__reg_deduce_bounds(dst_reg);
-	/* We might have learned some bits from the bounds. */
-	__reg_bound_offset(src_reg);
-	__reg_bound_offset(dst_reg);
-	/* Intersecting with the old var_off might have improved our bounds
-	 * slightly.  e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
-	 * then new var_off is (0; 0x7f...fc) which improves our umax.
-	 */
-	__update_reg_bounds(src_reg);
-	__update_reg_bounds(dst_reg);
+	reg_bounds_sync(src_reg);
+	reg_bounds_sync(dst_reg);
 }
 
 static void reg_combine_min_max(struct bpf_reg_state *true_src,