bpf: fix incorrect pruning decision when alignment must be tracked

Currently, when we enforce alignment tracking on direct packet access,
the verifier lets the following program pass despite doing a packet
write with unaligned access:

  0: (61) r2 = *(u32 *)(r1 +76)
  1: (61) r3 = *(u32 *)(r1 +80)
  2: (61) r7 = *(u32 *)(r1 +8)
  3: (bf) r0 = r2
  4: (07) r0 += 14
  5: (25) if r7 > 0x1 goto pc+4
   R0=pkt(id=0,off=14,r=0) R1=ctx R2=pkt(id=0,off=0,r=0)
   R3=pkt_end R7=inv,min_value=0,max_value=1 R10=fp
  6: (2d) if r0 > r3 goto pc+1
   R0=pkt(id=0,off=14,r=14) R1=ctx R2=pkt(id=0,off=0,r=14)
   R3=pkt_end R7=inv,min_value=0,max_value=1 R10=fp
  7: (63) *(u32 *)(r0 -4) = r0
  8: (b7) r0 = 0
  9: (95) exit

  from 6 to 8:
   R0=pkt(id=0,off=14,r=0) R1=ctx R2=pkt(id=0,off=0,r=0)
   R3=pkt_end R7=inv,min_value=0,max_value=1 R10=fp
  8: (b7) r0 = 0
  9: (95) exit

  from 5 to 10:
   R0=pkt(id=0,off=14,r=0) R1=ctx R2=pkt(id=0,off=0,r=0)
   R3=pkt_end R7=inv,min_value=2 R10=fp
  10: (07) r0 += 1
  11: (05) goto pc-6
  6: safe                           <----- here, wrongly found safe
  processed 15 insns

However, if we enforce a pruning mismatch by adding state into r8
which is then being mismatched in states_equal(), we find that for
the otherwise same program, the verifier detects a misaligned packet
access when actually walking that path:

  0: (61) r2 = *(u32 *)(r1 +76)
  1: (61) r3 = *(u32 *)(r1 +80)
  2: (61) r7 = *(u32 *)(r1 +8)
  3: (b7) r8 = 1
  4: (bf) r0 = r2
  5: (07) r0 += 14
  6: (25) if r7 > 0x1 goto pc+4
   R0=pkt(id=0,off=14,r=0) R1=ctx R2=pkt(id=0,off=0,r=0)
   R3=pkt_end R7=inv,min_value=0,max_value=1
   R8=imm1,min_value=1,max_value=1,min_align=1 R10=fp
  7: (2d) if r0 > r3 goto pc+1
   R0=pkt(id=0,off=14,r=14) R1=ctx R2=pkt(id=0,off=0,r=14)
   R3=pkt_end R7=inv,min_value=0,max_value=1
   R8=imm1,min_value=1,max_value=1,min_align=1 R10=fp
  8: (63) *(u32 *)(r0 -4) = r0
  9: (b7) r0 = 0
  10: (95) exit

  from 7 to 9:
   R0=pkt(id=0,off=14,r=0) R1=ctx R2=pkt(id=0,off=0,r=0)
   R3=pkt_end R7=inv,min_value=0,max_value=1
   R8=imm1,min_value=1,max_value=1,min_align=1 R10=fp
  9: (b7) r0 = 0
  10: (95) exit

  from 6 to 11:
   R0=pkt(id=0,off=14,r=0) R1=ctx R2=pkt(id=0,off=0,r=0)
   R3=pkt_end R7=inv,min_value=2
   R8=imm1,min_value=1,max_value=1,min_align=1 R10=fp
  11: (07) r0 += 1
  12: (b7) r8 = 0
  13: (05) goto pc-7                <----- mismatch due to r8
  7: (2d) if r0 > r3 goto pc+1
   R0=pkt(id=0,off=15,r=15) R1=ctx R2=pkt(id=0,off=0,r=15)
   R3=pkt_end R7=inv,min_value=2
   R8=imm0,min_value=0,max_value=0,min_align=2147483648 R10=fp
  8: (63) *(u32 *)(r0 -4) = r0
  misaligned packet access off 2+15+-4 size 4

The reason why we fail to see it in states_equal() is that the
third test in compare_ptrs_to_packet() ...

  if (old->off <= cur->off &&
      old->off >= old->range && cur->off >= cur->range)
          return true;

... will let the above pass. The situation we run into is that
old->off <= cur->off (14 <= 15), meaning that prior walked paths
went with smaller offset, which was later used in the packet
access after successful packet range check and found to be safe
already.

For example: Given is R0=pkt(id=0,off=0,r=0). Adding offset 14
as in above program to it, results in R0=pkt(id=0,off=14,r=0)
before the packet range test. Now, testing this against R3=pkt_end
with 'if r0 > r3 goto out' will transform R0 into R0=pkt(id=0,off=14,r=14)
for the case when we're within bounds. A write into the packet
at offset *(u32 *)(r0 -4), that is, 2 + 14 -4, is valid and
aligned (2 is for NET_IP_ALIGN). After processing this with
all fall-through paths, we later on check paths from branches.
When the above skb->mark test is true, then we jump near the
end of the program, perform r0 += 1, and jump back to the
'if r0 > r3 goto out' test we've visited earlier already. This
time, R0 is of type R0=pkt(id=0,off=15,r=0), and we'll prune
that part because this time we'll have a larger safe packet
range, and we already found that with off=14 all further insn
were already safe, so it's safe as well with a larger off.
However, the problem is that the subsequent write into the packet
with 2 + 15 -4 is then unaligned, and not caught by the alignment
tracking. Note that min_align, aux_off, and aux_off_align were
all 0 in this example.

Since we cannot tell at this time what kind of packet access was
performed in the prior walk and what minimal requirements it has
(we might do so in the future, but that requires more complexity),
fix it to disable this pruning case for strict alignment for now,
and let the verifier do check such paths instead. With that applied,
the test cases pass and reject the program due to misalignment.

Fixes: d1174416 ("bpf: Track alignment of register values in the verifier.")
Reference: http://patchwork.ozlabs.org/patch/761909/


Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>