ring-buffer: Fix polling on trace_pipe
authorMartin Lau <kafai@fb.com>
Tue, 10 Jun 2014 06:06:42 +0000 (23:06 -0700)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Mon, 28 Jul 2014 15:00:06 +0000 (08:00 -0700)
commit 97b8ee845393701edc06e27ccec2876ff9596019 upstream.

ring_buffer_poll_wait() should always put the poll_table to its wait_queue
even there is immediate data available.  Otherwise, the following epoll and
read sequence will eventually hang forever:

1. Put some data to make the trace_pipe ring_buffer read ready first
2. epoll_ctl(efd, EPOLL_CTL_ADD, trace_pipe_fd, ee)
3. epoll_wait()
4. read(trace_pipe_fd) till EAGAIN
5. Add some more data to the trace_pipe ring_buffer
6. epoll_wait() -> this epoll_wait() will block forever

~ During the epoll_ctl(efd, EPOLL_CTL_ADD,...) call in step 2,
  ring_buffer_poll_wait() returns immediately without adding poll_table,
  which has poll_table->_qproc pointing to ep_poll_callback(), to its
  wait_queue.
~ During the epoll_wait() call in step 3 and step 6,
  ring_buffer_poll_wait() cannot add ep_poll_callback() to its wait_queue
  because the poll_table->_qproc is NULL and it is how epoll works.
~ When there is new data available in step 6, ring_buffer does not know
  it has to call ep_poll_callback() because it is not in its wait queue.
  Hence, block forever.

Other poll implementation seems to call poll_wait() unconditionally as the very
first thing to do.  For example, tcp_poll() in tcp.c.

Link: http://lkml.kernel.org/p/20140610060637.GA14045@devbig242.prn2.facebook.com
Fixes: 2a2cc8f7c4d0 "ftrace: allow the event pipe to be polled"
Reviewed-by: Chris Mason <clm@fb.com>
Signed-off-by: Martin Lau <kafai@fb.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
kernel/trace/ring_buffer.c

index 8e94c1102636670dabf637e5aff3e2e65fa661bc..4063d5fe5e44bf1a36f51da6f1f76773164aaafc 100644 (file)
@@ -616,10 +616,6 @@ int ring_buffer_poll_wait(struct ring_buffer *buffer, int cpu,
        struct ring_buffer_per_cpu *cpu_buffer;
        struct rb_irq_work *work;
 
-       if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
-           (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
-               return POLLIN | POLLRDNORM;
-
        if (cpu == RING_BUFFER_ALL_CPUS)
                work = &buffer->irq_work;
        else {