On Mon 2020-01-06 17:56:54, Peter Zijlstra wrote:

On Mon, Jan 06, 2020 at 05:39:30PM +0100, Jann Horn wrote:

quoted

Hi!

I was chatting with kpsingh about BPF trampolines, and I noticed that
it looks like BPF trampolines (as of current bpf-next/master) seem to
be missing synchronization between trampoline code updates and
trampoline execution. Or maybe I'm missing something?

If I understand correctly, trampolines are executed directly from the
fentry placeholders at the start of arbitrary kernel functions, so
they can run without any locks held. So for example, if task A starts
executing a trampoline on entry to sys_open(), then gets preempted in
the middle of the trampoline, and then task B quickly calls
BPF_RAW_TRACEPOINT_OPEN twice, and then task A continues execution,
task A will end up executing the middle of newly-written machine code,
which can probably end up crashing the kernel somehow?

I think that at least to synchronize trampoline text freeing with
concurrent trampoline execution, it is necessary to do something
similar to what the livepatching code does with klp_check_stack(), and
then either use a callback from the scheduler to periodically re-check
tasks that were in the trampoline or let the trampoline tail-call into
a cleanup helper that is part of normal kernel text. And you'd
probably have to gate BPF trampolines on
CONFIG_HAVE_RELIABLE_STACKTRACE.

ftrace uses synchronize_rcu_tasks() to flip between trampolines iirc.

ftrace calls also schedule_on_each_cpu(ftrace_sync) to handle
situations where RCU is not watching, see rcu_is_watching().

The following is called in ftrace_shutdown():

	schedule_on_each_cpu(ftrace_sync);

	if (IS_ENABLED(CONFIG_PREEMPTION))
		synchronize_rcu_tasks();

	arch_ftrace_trampoline_free(ops);

Best Regards,
Petr