We can potentially run into a couple of issues with the XDP
bpf_redirect_map() helper. The ri->map in the per CPU storage
can become stale in several ways, mostly due to misuse, where
we can then trigger a use after free on the map:

i) prog A is calling bpf_redirect_map(), returning XDP_REDIRECT
and running on a driver not supporting XDP_REDIRECT yet. The
ri->map on that CPU becomes stale when the XDP program is unloaded
on the driver, and a prog B loaded on a different driver which
supports XDP_REDIRECT return code. prog B would have to omit
calling to bpf_redirect_map() and just return XDP_REDIRECT, which
would then access the freed map in xdp_do_redirect() since not
cleared for that CPU.

ii) prog A is calling bpf_redirect_map(), returning a code other
than XDP_REDIRECT. prog A is then detached, which triggers release
of the map. prog B is attached which, similarly as in i), would
just return XDP_REDIRECT without having called bpf_redirect_map()
and thus be accessing the freed map in xdp_do_redirect() since
not cleared for that CPU.

iii) prog A is attached to generic XDP, calling the bpf_redirect_map()
helper and returning XDP_REDIRECT. xdp_do_generic_redirect() is
currently not handling ri->map (will be fixed by Jesper), so it's
not being reset. Later loading a e.g. native prog B which would,
say, call bpf_xdp_redirect() and then returns XDP_REDIRECT would
find in xdp_do_redirect() that a map was set and uses that causing
use after free on map access.

Fix thus needs to avoid accessing stale ri->map pointers, naive
way would be to call a BPF function from drivers that just resets
it to NULL for all XDP return codes but XDP_REDIRECT and including
XDP_REDIRECT for drivers not supporting it yet (and let ri->map
being handled in xdp_do_generic_redirect()). There is a less
intrusive way w/o letting drivers call a reset for each BPF run.

The verifier knows we're calling into bpf_xdp_redirect_map()
helper, so it can do a small insn rewrite transparent to the prog
itself in the sense that it fills R4 with a pointer to the own
bpf_prog. We have that pointer at verification time anyway and
R4 is allowed to be used as per calling convention we scratch
R0 to R5 anyway, so they become inaccessible and program cannot
read them prior to a write. Then, the helper would store the prog
pointer in the current CPUs struct redirect_info. Later in
xdp_do_*_redirect() we check whether the redirect_info's prog
pointer is the same as passed xdp_prog pointer, and if that's
the case then all good, since the prog holds a ref on the map
anyway, so it is always valid at that point in time and must
have a reference count of at least 1. If in the unlikely case
they are not equal, it means we got a stale pointer, so we clear
and bail out right there. Also do reset map and the owning prog
in bpf_xdp_redirect(), so that bpf_xdp_redirect_map() and
bpf_xdp_redirect() won't get mixed up, only the last call should
take precedence. A tc bpf_redirect() doesn't use map anywhere
yet, so no need to clear it there since never accessed in that
layer.

Note that in case the prog is released, and thus the map as
well we're still under RCU read critical section at that time
and have preemption disabled as well. Once we commit with the
__dev_map_insert_ctx() from xdp_do_redirect_map() and set the
map to ri->map_to_flush, we still wait for a xdp_do_flush_map()
to finish in devmap dismantle time once flush_needed bit is set,
so that is fine.

Fixes: 97f91a7c

 ("bpf: add bpf_redirect_map helper routine")
Reported-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: John Fastabend <john.fastabend@gmail.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>