Currently, when having map file descriptors pointing to program arrays,
there's still the issue that we unconditionally flush program array
contents via bpf_fd_array_map_clear() in bpf_map_release(). This happens
when such a file descriptor is released and is independent of the map's
refcount.

Having this flush independent of the refcount is for a reason: there
can be arbitrary complex dependency chains among tail calls, also circular
ones (direct or indirect, nesting limit determined during runtime), and
we need to make sure that the map drops all references to eBPF programs
it holds, so that the map's refcount can eventually drop to zero and
initiate its freeing. Btw, a walk of the whole dependency graph would
not be possible for various reasons, one being complexity and another
one inconsistency, i.e. new programs can be added to parts of the graph
at any time, so there's no guaranteed consistent state for the time of
such a walk.

Now, the program array pinning itself works, but the issue is that each
derived file descriptor on close would nevertheless call unconditionally
into bpf_fd_array_map_clear(). Instead, keep track of users and postpone
this flush until the last reference to a user is dropped. As this only
concerns a subset of references (f.e. a prog array could hold a program
that itself has reference on the prog array holding it, etc), we need to
track them separately.

Short analysis on the refcounting: on map creation time usercnt will be
one, so there's no change in behaviour for bpf_map_release(), if unpinned.
If we already fail in map_create(), we are immediately freed, and no
file descriptor has been made public yet. In bpf_obj_pin_user(), we need
to probe for a possible map in bpf_fd_probe_obj() already with a usercnt
reference, so before we drop the reference on the fd with fdput().
Therefore, if actual pinning fails, we need to drop that reference again
in bpf_any_put(), otherwise we keep holding it. When last reference
drops on the inode, the bpf_any_put() in bpf_evict_inode() will take
care of dropping the usercnt again. In the bpf_obj_get_user() case, the
bpf_any_get() will grab a reference on the usercnt, still at a time when
we have the reference on the path. Should we later on fail to grab a new
file descriptor, bpf_any_put() will drop it, otherwise we hold it until
bpf_map_release() time.

Joint work with Alexei.

Fixes: b2197755

 ("bpf: add support for persistent maps/progs")
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>