Currently, the destination register is marked as unknown for 32-bit
sub-register move (BPF_MOV | BPF_ALU) whenever the source register type is
SCALAR_VALUE.

This is too conservative that some valid cases will be rejected.
Especially, this may turn a constant scalar value into unknown value that
could break some assumptions of verifier.

For example, test_l4lb_noinline.c has the following C code:

    struct real_definition *dst

1:  if (!get_packet_dst(&dst, &pckt, vip_info, is_ipv6))
2:    return TC_ACT_SHOT;
3:
4:  if (dst->flags & F_IPV6) {

get_packet_dst is responsible for initializing "dst" into valid pointer and
return true (1), otherwise return false (0). The compiled instruction
sequence using alu32 will be:

  412: (54) (u32) r7 &= (u32) 1
  413: (bc) (u32) r0 = (u32) r7
  414: (95) exit

insn 413, a BPF_MOV | BPF_ALU, however will turn r0 into unknown value even
r7 contains SCALAR_VALUE 1.

This causes trouble when verifier is walking the code path that hasn't
initialized "dst" inside get_packet_dst, for which case 0 is returned and
we would then expect verifier concluding line 1 in the above C code pass
the "if" check, therefore would skip fall through path starting at line 4.
Now, because r0 returned from callee has became unknown value, so verifier
won't skip analyzing path starting at line 4 and "dst->flags" requires
dereferencing the pointer "dst" which actually hasn't be initialized for
this path.

This patch relaxed the code marking sub-register move destination. For a
SCALAR_VALUE, it is safe to just copy the value from source then truncate
it into 32-bit.

A unit test also included to demonstrate this issue. This test will fail
before this patch.

This relaxation could let verifier skipping more paths for conditional
comparison against immediate. It also let verifier recording a more
accurate/strict value for one register at one state, if this state end up
with going through exit without rejection and it is used for state
comparison later, then it is possible an inaccurate/permissive value is
better. So the real impact on verifier processed insn number is complex.
But in all, without this fix, valid program could be rejected.

>From real benchmarking on kernel selftests and Cilium bpf tests, there is
no impact on processed instruction number when tests ares compiled with
default compilation options. There is slightly improvements when they are
compiled with -mattr=+alu32 after this patch.

Also, test_xdp_noinline/-mattr=+alu32 now passed verification. It is
rejected before this fix.

Insn processed before/after this patch:

                        default     -mattr=+alu32

Kernel selftest

===
test_xdp.o              371/371      369/369
test_l4lb.o             6345/6345    5623/5623
test_xdp_noinline.o     2971/2971    rejected/2727
test_tcp_estates.o      429/429      430/430

Cilium bpf
===
bpf_lb-DLB_L3.o:        2085/2085     1685/1687
bpf_lb-DLB_L4.o:        2287/2287     1986/1982
bpf_lb-DUNKNOWN.o:      690/690       622/622
bpf_lxc.o:              95033/95033   N/A
bpf_netdev.o:           7245/7245     N/A
bpf_overlay.o:          2898/2898     3085/2947

NOTE:
  - bpf_lxc.o and bpf_netdev.o compiled by -mattr=+alu32 are rejected by
    verifier due to another issue inside verifier on supporting alu32
    binary.
  - Each cilium bpf program could generate several processed insn number,
    above number is sum of them.

v1->v2:
 - Restrict the change on SCALAR_VALUE.
 - Update benchmark numbers on Cilium bpf tests.
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>