Re: [PATCH v2] tty: fix stall caused by missing memory barrier in drivers/tty/n_tty.c

[Peter Hurley <peter@hurleysoftware.com>]

On 10/02/2015 04:27 AM, Kosuke Tatsukawa wrote:
> 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.

Reviewed-by: Peter Hurley <peter@hurleysoftware.com>