Martin reported a verifier issue that hit the BUG_ON() for his
test case in the mark_reg_unknown_value() function:

  [  202.861380] kernel BUG at kernel/bpf/verifier.c:467!
  [...]
  [  203.291109] Call Trace:
  [  203.296501]  [<ffffffff811364d5>] mark_map_reg+0x45/0x50
  [  203.308225]  [<ffffffff81136558>] mark_map_regs+0x78/0x90
  [  203.320140]  [<ffffffff8113938d>] do_check+0x226d/0x2c90
  [  203.331865]  [<ffffffff8113a6ab>] bpf_check+0x48b/0x780
  [  203.343403]  [<ffffffff81134c8e>] bpf_prog_load+0x27e/0x440
  [  203.355705]  [<ffffffff8118a38f>] ? handle_mm_fault+0x11af/0x1230
  [  203.369158]  [<ffffffff812d8188>] ? security_capable+0x48/0x60
  [  203.382035]  [<ffffffff811351a4>] SyS_bpf+0x124/0x960
  [  203.393185]  [<ffffffff810515f6>] ? __do_page_fault+0x276/0x490
  [  203.406258]  [<ffffffff816db320>] entry_SYSCALL_64_fastpath+0x13/0x94

This issue got uncovered after the fix in a08dd0da ("bpf: fix
regression on verifier pruning wrt map lookups"). The reason why it
wasn't noticed before was, because as mentioned in a08dd0da,
mark_map_regs() was doing the id matching incorrectly based on the
uncached regs[regno].id. So, in the first loop, we walked all regs
and as soon as we found regno == i, then this reg's id was cleared
when calling mark_reg_unknown_value() thus that every subsequent
register was probed against id of 0 (which, in combination with the
PTR_TO_MAP_VALUE_OR_NULL type is an invalid condition that no other
register state can hold), and therefore wasn't type transitioned such
as in the spilled register case for the second loop.

Now since that got fixed, it turned out that 57a09bf0 ("bpf:
Detect identical PTR_TO_MAP_VALUE_OR_NULL registers") used
mark_reg_unknown_value() incorrectly for the spilled regs, and thus
hitting the BUG_ON() in some cases due to regno >= MAX_BPF_REG.

Although spilled regs have the same type as the non-spilled regs
for the verifier state, that is, struct bpf_reg_state, they are
semantically different from the non-spilled regs. In other words,
there can be up to 64 (MAX_BPF_STACK / BPF_REG_SIZE) spilled regs
in the stack, for example, register R<x> could have been spilled by
the program to stack location X, Y, Z, and in mark_map_regs() we
need to scan these stack slots of type STACK_SPILL for potential
registers that we have to transition from PTR_TO_MAP_VALUE_OR_NULL.
Therefore, depending on the location, the spilled_regs regno can
be a lot higher than just MAX_BPF_REG's value since we operate on
stack instead. The reset in mark_reg_unknown_value() itself is
just fine, only that the BUG_ON() was inappropriate for this. Fix
it by making a __mark_reg_unknown_value() version that can be
called from mark_map_reg() generically; we know for the non-spilled
case that the regno is always < MAX_BPF_REG anyway.

Fixes: 57a09bf0 ("bpf: Detect identical PTR_TO_MAP_VALUE_OR_NULL registers")
Reported-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>