bpf: Fix incorrect verifier simulation around jmp32's jeq/jne

commit a12ca62 upstream.

Kuee reported a quirk in the jmp32's jeq/jne simulation, namely that the
register value does not match expectations for the fall-through path. For
example:

Before fix:

  0: R1=ctx(off=0,imm=0) R10=fp0
  0: (b7) r2 = 0                        ; R2_w=P0
  1: (b7) r6 = 563                      ; R6_w=P563
  2: (87) r2 = -r2                      ; R2_w=Pscalar()
  3: (87) r2 = -r2                      ; R2_w=Pscalar()
  4: (4c) w2 |= w6                      ; R2_w=Pscalar(umin=563,umax=4294967295,var_off=(0x233; 0xfffffdcc),s32_min=-2147483085) R6_w=P563
  5: (56) if w2 != 0x8 goto pc+1        ; R2_w=P571  <--- [*]
  6: (95) exit
  R0 !read_ok

After fix:

  0: R1=ctx(off=0,imm=0) R10=fp0
  0: (b7) r2 = 0                        ; R2_w=P0
  1: (b7) r6 = 563                      ; R6_w=P563
  2: (87) r2 = -r2                      ; R2_w=Pscalar()
  3: (87) r2 = -r2                      ; R2_w=Pscalar()
  4: (4c) w2 |= w6                      ; R2_w=Pscalar(umin=563,umax=4294967295,var_off=(0x233; 0xfffffdcc),s32_min=-2147483085) R6_w=P563
  5: (56) if w2 != 0x8 goto pc+1        ; R2_w=P8  <--- [*]
  6: (95) exit
  R0 !read_ok

As can be seen on line 5 for the branch fall-through path in R2 [*] is that
given condition w2 != 0x8 is false, verifier should conclude that r2 = 8 as
upper 32 bit are known to be zero. However, verifier incorrectly concludes
that r2 = 571 which is far off.

The problem is it only marks false{true}_reg as known in the switch for JE/NE
case, but at the end of the function, it uses {false,true}_{64,32}off to
update {false,true}_reg->var_off and they still hold the prior value of
{false,true}_reg->var_off before it got marked as known. The subsequent
__reg_combine_32_into_64() then propagates this old var_off and derives new
bounds. The information between min/max bounds on {false,true}_reg from
setting the register to known const combined with the {false,true}_reg->var_off
based on the old information then derives wrong register data.

Fix it by detangling the BPF_JEQ/BPF_JNE cases and updating relevant
{false,true}_{64,32}off tnums along with the register marking to known
constant.

Fixes: 3f50f13 ("bpf: Verifier, do explicit ALU32 bounds tracking")
Reported-by: Kuee K1r0a <liulin063@gmail.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Acked-by: John Fastabend <john.fastabend@gmail.com>
Link: https://lore.kernel.org/bpf/20220701124727.11153-1-daniel@iogearbox.net
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

kernel/bpf/verifier.c

@@ -7512,26 +7512,33 @@ static void reg_set_min_max(struct bpf_reg_state *true_reg,
 		return;
 
 	switch (opcode) {
+	/* JEQ/JNE comparison doesn't change the register equivalence.
+	 *
+	 * r1 = r2;
+	 * if (r1 == 42) goto label;
+	 * ...
+	 * label: // here both r1 and r2 are known to be 42.
+	 *
+	 * Hence when marking register as known preserve it's ID.
+	 */
 	case BPF_JEQ:
+		if (is_jmp32) {
+			__mark_reg32_known(true_reg, val32);
+			true_32off = tnum_subreg(true_reg->var_off);
+		} else {
+			___mark_reg_known(true_reg, val);
+			true_64off = true_reg->var_off;
+		}
+		break;
 	case BPF_JNE:
-	{
-		struct bpf_reg_state *reg =
-			opcode == BPF_JEQ ? true_reg : false_reg;
-
-		/* JEQ/JNE comparison doesn't change the register equivalence.
-		 * r1 = r2;
-		 * if (r1 == 42) goto label;
-		 * ...
-		 * label: // here both r1 and r2 are known to be 42.
-		 *
-		 * Hence when marking register as known preserve it's ID.
-		 */
-		if (is_jmp32)
-			__mark_reg32_known(reg, val32);
-		else
-			___mark_reg_known(reg, val);
+		if (is_jmp32) {
+			__mark_reg32_known(false_reg, val32);
+			false_32off = tnum_subreg(false_reg->var_off);
+		} else {
+			___mark_reg_known(false_reg, val);
+			false_64off = false_reg->var_off;
+		}
 		break;
-	}
 	case BPF_JSET:
 		if (is_jmp32) {
 			false_32off = tnum_and(false_32off, tnum_const(~val32));