bpf: Fix a verifier issue when assigning 32bit reg states to 64bit ones

With the latest trunk llvm (llvm 11), I hit a verifier issue for
test_prog subtest test_verif_scale1.

The following simplified example illustrate the issue:
    w9 = 0  /* R9_w=inv0 */
    r8 = *(u32 *)(r1 + 80)  /* __sk_buff->data_end */
    r7 = *(u32 *)(r1 + 76)  /* __sk_buff->data */
    ......
    w2 = w9 /* R2_w=inv0 */
    r6 = r7 /* R6_w=pkt(id=0,off=0,r=0,imm=0) */
    r6 += r2 /* R6_w=inv(id=0) */
    r3 = r6 /* R3_w=inv(id=0) */
    r3 += 14 /* R3_w=inv(id=0) */
    if r3 > r8 goto end
    r5 = *(u32 *)(r6 + 0) /* R6_w=inv(id=0) */
       <== error here: R6 invalid mem access 'inv'
    ...
  end:

In real test_verif_scale1 code, "w9 = 0" and "w2 = w9" are in
different basic blocks.

In the above, after "r6 += r2", r6 becomes a scalar, which eventually
caused the memory access error. The correct register state should be
a pkt pointer.

The inprecise register state starts at "w2 = w9".
The 32bit register w9 is 0, in __reg_assign_32_into_64(),
the 64bit reg->smax_value is assigned to be U32_MAX.
The 64bit reg->smin_value is 0 and the 64bit register
itself remains constant based on reg->var_off.

In adjust_ptr_min_max_vals(), the verifier checks for a known constant,
smin_val must be equal to smax_val. Since they are not equal,
the verifier decides r6 is a unknown scalar, which caused later failure.

The llvm10 does not have this issue as it generates different code:
    w9 = 0  /* R9_w=inv0 */
    r8 = *(u32 *)(r1 + 80)  /* __sk_buff->data_end */
    r7 = *(u32 *)(r1 + 76)  /* __sk_buff->data */
    ......
    r6 = r7 /* R6_w=pkt(id=0,off=0,r=0,imm=0) */
    r6 += r9 /* R6_w=pkt(id=0,off=0,r=0,imm=0) */
    r3 = r6 /* R3_w=pkt(id=0,off=0,r=0,imm=0) */
    r3 += 14 /* R3_w=pkt(id=0,off=14,r=0,imm=0) */
    if r3 > r8 goto end
    ...

To fix the above issue, we can include zero in the test condition for
assigning the s32_max_value and s32_min_value to their 64-bit equivalents
smax_value and smin_value.

Further, fix the condition to avoid doing zero extension bounds checks
when s32_min_value <= 0. This could allow for the case where bounds
32-bit bounds (-1,1) get incorrectly translated to (0,1) 64-bit bounds.
When in-fact the -1 min value needs to force U32_MAX bound.

Fixes: 3f50f132 ("bpf: Verifier, do explicit ALU32 bounds tracking")
Signed-off-by: John Fastabend <john.fastabend@gmail.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/bpf/159077331983.6014.5758956193749002737.stgit@john-Precision-5820-Tower

kernel/bpf/verifier.c

@@ -1168,14 +1168,14 @@ static void __reg_assign_32_into_64(struct bpf_reg_state *reg)
 	 * but must be positive otherwise set to worse case bounds
 	 * and refine later from tnum.
 	 */
-	if (reg->s32_min_value > 0)
-		reg->smin_value = reg->s32_min_value;
-	else
-		reg->smin_value = 0;
-	if (reg->s32_max_value > 0)
+	if (reg->s32_min_value >= 0 && reg->s32_max_value >= 0)
 		reg->smax_value = reg->s32_max_value;
 	else
 		reg->smax_value = U32_MAX;
+	if (reg->s32_min_value >= 0)
+		reg->smin_value = reg->s32_min_value;
+	else
+		reg->smin_value = 0;
 }
 
 static void __reg_combine_32_into_64(struct bpf_reg_state *reg)