When ctx access is used, the kernel often needs to expand/rewrite
instructions, so after that patching, branch offsets have to be
adjusted for both forward and backward jumps in the new eBPF program,
but for backward jumps it fails to account the delta. Meaning, for
example, if the expansion happens exactly on the insn that sits at
the jump target, it doesn't fix up the back jump offset.

Analysis on what the check in adjust_branches() is currently doing:

  /* adjust offset of jmps if necessary */
  if (i < pos && i + insn->off + 1 > pos)
    insn->off += delta;
  else if (i > pos && i + insn->off + 1 < pos)
    insn->off -= delta;

First condition (forward jumps):

  Before:                         After:

  insns[0]                        insns[0]
  insns[1] <--- i/insn            insns[1] <--- i/insn
  insns[2] <--- pos               insns[P] <--- pos
  insns[3]                        insns[P]  `------| delta
  insns[4] <--- target_X          insns[P]   `-----|
  insns[5]                        insns[3]
                                  insns[4] <--- target_X
                                  insns[5]

First case is if we cross pos-boundary and the jump instruction was
before pos. This is handeled correctly. I.e. if i == pos, then this
would mean our jump that we currently check was the patchlet itself
that we just injected. Since such patchlets are self-contained and
have no awareness of any insns before or after the patched one, the
delta is correctly not adjusted. Also, for the second condition in
case of i + insn->off + 1 == pos, means we jump to that newly patched
instruction, so no offset adjustment are needed. That part is correct.

Second condition (backward jumps):

  Before:                         After:

  insns[0]                        insns[0]
  insns[1] <--- target_X          insns[1] <--- target_X
  insns[2] <--- pos <-- target_Y  insns[P] <--- pos <-- target_Y
  insns[3]                        insns[P]  `------| delta
  insns[4] <--- i/insn            insns[P]   `-----|
  insns[5]                        insns[3]
                                  insns[4] <--- i/insn
                                  insns[5]

Second interesting case is where we cross pos-boundary and the jump
instruction was after pos. Backward jump with i == pos would be
impossible and pose a bug somewhere in the patchlet, so the first
condition checking i > pos is okay only by itself. However, i +
insn->off + 1 < pos does not always work as intended to trigger the
adjustment. It works when jump targets would be far off where the
delta wouldn't matter. But, for example, where the fixed insn->off
before pointed to pos (target_Y), it now points to pos + delta, so
that additional room needs to be taken into account for the check.
This means that i) both tests here need to be adjusted into pos + delta,
and ii) for the second condition, the test needs to be <= as pos
itself can be a target in the backjump, too.

Fixes: 9bac3d6d ("bpf: allow extended BPF programs access skb fields")
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>