tty: fix stall caused by missing memory barrier in drivers/tty/n_tty.c
authorKosuke Tatsukawa <tatsu@ab.jp.nec.com>
Fri, 2 Oct 2015 08:27:05 +0000 (08:27 +0000)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Sun, 4 Oct 2015 18:03:40 +0000 (19:03 +0100)
commite81107d4c6bd098878af9796b24edc8d4a9524fd
treed3f13777691b356ce1eaf73c92a98f7d5961fbb3
parent8f1bd8f2ad2358d6a88c115481ff3e69817d1bde
tty: fix stall caused by missing memory barrier in drivers/tty/n_tty.c

My colleague ran into a program stall on a x86_64 server, where
n_tty_read() was waiting for data even if there was data in the buffer
in the pty.  kernel stack for the stuck process looks like below.
 #0 [ffff88303d107b58] __schedule at ffffffff815c4b20
 #1 [ffff88303d107bd0] schedule at ffffffff815c513e
 #2 [ffff88303d107bf0] schedule_timeout at ffffffff815c7818
 #3 [ffff88303d107ca0] wait_woken at ffffffff81096bd2
 #4 [ffff88303d107ce0] n_tty_read at ffffffff8136fa23
 #5 [ffff88303d107dd0] tty_read at ffffffff81368013
 #6 [ffff88303d107e20] __vfs_read at ffffffff811a3704
 #7 [ffff88303d107ec0] vfs_read at ffffffff811a3a57
 #8 [ffff88303d107f00] sys_read at ffffffff811a4306
 #9 [ffff88303d107f50] entry_SYSCALL_64_fastpath at ffffffff815c86d7

There seems to be two problems causing this issue.

First, in drivers/tty/n_tty.c, __receive_buf() stores the data and
updates ldata->commit_head using smp_store_release() and then checks
the wait queue using waitqueue_active().  However, since there is no
memory barrier, __receive_buf() could return without calling
wake_up_interactive_poll(), and at the same time, n_tty_read() could
start to wait in wait_woken() as in the following chart.

        __receive_buf()                         n_tty_read()
------------------------------------------------------------------------
if (waitqueue_active(&tty->read_wait))
/* Memory operations issued after the
   RELEASE may be completed before the
   RELEASE operation has completed */
                                        add_wait_queue(&tty->read_wait, &wait);
                                        ...
                                        if (!input_available_p(tty, 0)) {
smp_store_release(&ldata->commit_head,
                  ldata->read_head);
                                        ...
                                        timeout = wait_woken(&wait,
                                          TASK_INTERRUPTIBLE, timeout);
------------------------------------------------------------------------

The second problem is that n_tty_read() also lacks a memory barrier
call and could also cause __receive_buf() to return without calling
wake_up_interactive_poll(), and n_tty_read() to wait in wait_woken()
as in the chart below.

        __receive_buf()                         n_tty_read()
------------------------------------------------------------------------
                                        spin_lock_irqsave(&q->lock, flags);
                                        /* from add_wait_queue() */
                                        ...
                                        if (!input_available_p(tty, 0)) {
                                        /* Memory operations issued after the
                                           RELEASE may be completed before the
                                           RELEASE operation has completed */
smp_store_release(&ldata->commit_head,
                  ldata->read_head);
if (waitqueue_active(&tty->read_wait))
                                        __add_wait_queue(q, wait);
                                        spin_unlock_irqrestore(&q->lock,flags);
                                        /* from add_wait_queue() */
                                        ...
                                        timeout = wait_woken(&wait,
                                          TASK_INTERRUPTIBLE, timeout);
------------------------------------------------------------------------

There are also other places in drivers/tty/n_tty.c which have similar
calls to waitqueue_active(), so instead of adding many memory barrier
calls, this patch simply removes the call to waitqueue_active(),
leaving just wake_up*() behind.

This fixes both problems because, even though the memory access before
or after the spinlocks in both wake_up*() and add_wait_queue() can
sneak into the critical section, it cannot go past it and the critical
section assures that they will be serialized (please see "INTER-CPU
ACQUIRING BARRIER EFFECTS" in Documentation/memory-barriers.txt for a
better explanation).  Moreover, the resulting code is much simpler.

Latency measurement using a ping-pong test over a pty doesn't show any
visible performance drop.

Signed-off-by: Kosuke Tatsukawa <tatsu@ab.jp.nec.com>
Cc: stable@vger.kernel.org
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
drivers/tty/n_tty.c