Re: [PATCH bpf-next 3/4] bpf: fix lockdep false positive in bpf_prog_register

[Peter Zijlstra]

On Tue, Jan 29, 2019 at 08:04:57PM -0800, Alexei Starovoitov wrote:
> Lockdep warns about false positive:

The report reads like:

        tracepoint_probe_register()
#0        mutex_lock(&tracepoint_mutex)
          tracepoint_add_func()
            static_key_slow_inc()
#1            cpus_read_lock();


        _cpu_up()
#1        cpus_write_lock();
          ...
          perf_event_init_cpu()
#2          mutex_lock(&pmus_lock);
#3          mutex_lock(&ctx->mutex);


        perf_ioctl()
#4        perf_event_ctx_lock();
          _perf_ioctl(IOC_QUERY_BPF)
            perf_event_query_prog_array()
#5            mutex_lock(&bpf_event_mutex);


        bpf_probe_register()
#5        mutex_lock(&bpf_event_mutex);
          __bpf_probe_register()
            tracepoint_probe_register()
#0            mutex_lock(&tracepoint_mutex);

Which to me reads like an entirely valid deadlock scenario.

And note that the first and last can be combined to give:

        bpf_probe_register()
#5        mutex_lock(&bpf_event_mutex);
          __bpf_probe_register()
            tracepoint_probe_register()
#0            mutex_lock(&tracepoint_mutex);
              tracepoint_add_func()
                static_key_slow_inc()
#1                cpus_read_lock();


Which generates a deadlock even without #0.

Why do you say this is not possible? All you need is 3 CPUs, one doing a
CPU online, one doing a perf ioctl() and one doing that
bpf_probe_register().