Jann Horn reported following analysis that could potentially result
in a very hard to trigger (if not impossible) UAF race, to quote his
event timeline:

 - Set up a process with threads T1, T2 and T3
 - Let T1 set up a socket filter F1 that invokes another filter F2
   through a BPF map [tail call]
 - Let T1 trigger the socket filter via a unix domain socket write,
   don't wait for completion
 - Let T2 call PERF_EVENT_IOC_SET_BPF with F2, don't wait for completion
 - Now T2 should be behind bpf_prog_get(), but before bpf_prog_put()
 - Let T3 close the file descriptor for F2, dropping the reference
   count of F2 to 2
 - At this point, T1 should have looked up F2 from the map, but not
   finished executing it
 - Let T3 remove F2 from the BPF map, dropping the reference count of
   F2 to 1
 - Now T2 should call bpf_prog_put() (wrong BPF program type), dropping
   the reference count of F2 to 0 and scheduling bpf_prog_free_deferred()
   via schedule_work()
 - At this point, the BPF program could be freed
 - BPF execution is still running in a freed BPF program

While at PERF_EVENT_IOC_SET_BPF time it's only guaranteed that the perf
event fd we're doing the syscall on doesn't disappear from underneath us
for whole syscall time, it may not be the case for the bpf fd used as
an argument only after we did the put. It needs to be a valid fd pointing
to a BPF program at the time of the call to make the bpf_prog_get() and
while T2 gets preempted, F2 must have dropped reference to 1 on the other
CPU. The fput() from the close() in T3 should also add additionally delay
to the reference drop via exit_task_work() when bpf_prog_release() gets
called as well as scheduling bpf_prog_free_deferred().

That said, it makes nevertheless sense to move the BPF prog destruction
generally after RCU grace period to guarantee that such scenario above,
but also others as recently fixed in ceb56070

 ("bpf, perf: delay release
of BPF prog after grace period") with regards to tail calls won't happen.
Integrating bpf_prog_free_deferred() directly into the RCU callback is
not allowed since the invocation might happen from either softirq or
process context, so we're not permitted to block. Reviewing all bpf_prog_put()
invocations from eBPF side (note, cBPF -> eBPF progs don't use this for
their destruction) with call_rcu() look good to me.

Since we don't know whether at the time of attaching the program, we're
already part of a tail call map, we need to use RCU variant. However, due
to this, there won't be severely more stress on the RCU callback queue:
situations with above bpf_prog_get() and bpf_prog_put() combo in practice
normally won't lead to releases, but even if they would, enough effort/
cycles have to be put into loading a BPF program into the kernel already.

Reported-by: Jann Horn <jannh@google.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>