bpf: Provide better register bounds after jmp32 instructions

With latest llvm (trunk https://github.com/llvm/llvm-project),
test_progs, which has +alu32 enabled, failed for strobemeta.o.
The verifier output looks like below with edit to replace large
decimal numbers with hex ones.
 193: (85) call bpf_probe_read_user_str#114
   R0=inv(id=0)
 194: (26) if w0 > 0x1 goto pc+4
   R0_w=inv(id=0,umax_value=0xffffffff00000001)
 195: (6b) *(u16 *)(r7 +80) = r0
 196: (bc) w6 = w0
   R6_w=inv(id=0,umax_value=0xffffffff,var_off=(0x0; 0xffffffff))
 197: (67) r6 <<= 32
   R6_w=inv(id=0,smax_value=0x7fffffff00000000,umax_value=0xffffffff00000000,
            var_off=(0x0; 0xffffffff00000000))
 198: (77) r6 >>= 32
   R6=inv(id=0,umax_value=0xffffffff,var_off=(0x0; 0xffffffff))
 ...
 201: (79) r8 = *(u64 *)(r10 -416)
   R8_w=map_value(id=0,off=40,ks=4,vs=13872,imm=0)
 202: (0f) r8 += r6
   R8_w=map_value(id=0,off=40,ks=4,vs=13872,umax_value=0xffffffff,var_off=(0x0; 0xffffffff))
 203: (07) r8 += 9696
   R8_w=map_value(id=0,off=9736,ks=4,vs=13872,umax_value=0xffffffff,var_off=(0x0; 0xffffffff))
 ...
 255: (bf) r1 = r8
   R1_w=map_value(id=0,off=9736,ks=4,vs=13872,umax_value=0xffffffff,var_off=(0x0; 0xffffffff))
 ...
 257: (85) call bpf_probe_read_user_str#114
 R1 unbounded memory access, make sure to bounds check any array access into a map

The value range for register r6 at insn 198 should be really just 0/1.
The umax_value=0xffffffff caused later verification failure.

After jmp instructions, the current verifier already tried to use just
obtained information to get better register range. The current mechanism is
for 64bit register only. This patch implemented to tighten the range
for 32bit sub-registers after jmp32 instructions.
With the patch, we have the below range ranges for the
above code sequence:
 193: (85) call bpf_probe_read_user_str#114
   R0=inv(id=0)
 194: (26) if w0 > 0x1 goto pc+4
   R0_w=inv(id=0,smax_value=0x7fffffff00000001,umax_value=0xffffffff00000001,
            var_off=(0x0; 0xffffffff00000001))
 195: (6b) *(u16 *)(r7 +80) = r0
 196: (bc) w6 = w0
   R6_w=inv(id=0,umax_value=0xffffffff,var_off=(0x0; 0x1))
 197: (67) r6 <<= 32
   R6_w=inv(id=0,umax_value=0x100000000,var_off=(0x0; 0x100000000))
 198: (77) r6 >>= 32
   R6=inv(id=0,umax_value=1,var_off=(0x0; 0x1))
 ...
 201: (79) r8 = *(u64 *)(r10 -416)
   R8_w=map_value(id=0,off=40,ks=4,vs=13872,imm=0)
 202: (0f) r8 += r6
   R8_w=map_value(id=0,off=40,ks=4,vs=13872,umax_value=1,var_off=(0x0; 0x1))
 203: (07) r8 += 9696
   R8_w=map_value(id=0,off=9736,ks=4,vs=13872,umax_value=1,var_off=(0x0; 0x1))
 ...
 255: (bf) r1 = r8
   R1_w=map_value(id=0,off=9736,ks=4,vs=13872,umax_value=1,var_off=(0x0; 0x1))
 ...
 257: (85) call bpf_probe_read_user_str#114
 ...

At insn 194, the register R0 has better var_off.mask and smax_value.
Especially, the var_off.mask ensures later lshift and rshift
maintains proper value range.
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Yonghong Song <yhs@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20191121170650.449030-1-yhs@fb.com

kernel/bpf/verifier.c

@@ -1007,6 +1007,17 @@ static void __reg_bound_offset(struct bpf_reg_state *reg)
 						 reg->umax_value));
 }
 
+static void __reg_bound_offset32(struct bpf_reg_state *reg)
+{
+	u64 mask = 0xffffFFFF;
+	struct tnum range = tnum_range(reg->umin_value & mask,
+				       reg->umax_value & mask);
+	struct tnum lo32 = tnum_cast(reg->var_off, 4);
+	struct tnum hi32 = tnum_lshift(tnum_rshift(reg->var_off, 32), 32);
+
+	reg->var_off = tnum_or(hi32, tnum_intersect(lo32, range));
+}
+
 /* Reset the min/max bounds of a register */
 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
 {
@@ -5589,6 +5600,10 @@ static void reg_set_min_max(struct bpf_reg_state *true_reg,
 	/* We might have learned some bits from the bounds. */
 	__reg_bound_offset(false_reg);
 	__reg_bound_offset(true_reg);
+	if (is_jmp32) {
+		__reg_bound_offset32(false_reg);
+		__reg_bound_offset32(true_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.
@@ -5698,6 +5713,10 @@ static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
 	/* We might have learned some bits from the bounds. */
 	__reg_bound_offset(false_reg);
 	__reg_bound_offset(true_reg);
+	if (is_jmp32) {
+		__reg_bound_offset32(false_reg);
+		__reg_bound_offset32(true_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.