Re: hci_ldsic nested locking problem

[Felipe Balbi]

On Thu, Mar 20, 2014 at 01:25:28PM -0500, Felipe Balbi wrote:
> On Thu, Mar 20, 2014 at 02:21:17PM -0400, Peter Hurley wrote:
> > On 03/20/2014 02:11 PM, Felipe Balbi wrote:
> > >On Thu, Mar 20, 2014 at 01:31:40PM -0400, Peter Hurley wrote:
> > >>[ +cc Huang Shijie ]
> > >>
> > >>On 03/20/2014 01:16 PM, Felipe Balbi wrote:
> > >>>On Thu, Mar 20, 2014 at 04:42:16PM +0000, Alan Cox wrote:
> > >>>>On Thu, 2014-03-20 at 11:34 -0500, Felipe Balbi wrote:
> > >>>>>Hi,
> > >>>>>
> > >>>>>when 8250 driver calls uart_write_wakeup(), the tty port lock is already
> > >>>>>taken. hci_ldisc.c's implementation of ->write_wakeup() calls
> > >>>>>tty->ops->write() to actually send the characters, but that call will
> > >>>>>try to acquire the same port lock again.
> > >>>>>
> > >>>>>Looking at other line disciplines that looks like a bug in hci_ldisc.c.
> > >>>>>Am I correct to assume that ->write_wakeup() is supposed to *just*
> > >>>>>wakeup the bottom half so we handle ->write() in another context ?
> > >>>>>
> > >>>>>Is it legal to call tty->ops->write() from within ->write_wakeup() ?
> > >>>>
> > >>>>It isn't because you might send all the bytes and go
> > >>>>
> > >>>>	write
> > >>>>		write_wakeup
> > >>>>			write
> > >>>>				write wakeup
> > >>>>					...
> > >>>>
> > >>>>and recurse
> > >>>
> > >>>cool, so there really is a bug in hci_ldisc. Marcel, any tips on how do
> > >>>you want this to be sorted out ?
> > >>
> > >>hci_uart_tx_wakeup() should perform the I/O as work.
> > >>FWIW, this was reported by Huang Shijie back on Dec 6.
> > >>
> > >>I'd fix it but I have no way to test it.
> > >
> > >here's a build-tested only patch which is waiting for testing from other
> > >colleagues who've got a platform to reproduce the problem:
> > 
> > Where's the cancel_work_sync() on teardown?
> 
> here, as a patch too this time:
> 
> From 3ee6b74833f154df64a6164476b854846206a3f2 Mon Sep 17 00:00:00 2001
> From: Felipe Balbi <balbi@xxxxxx>
> Date: Thu, 20 Mar 2014 13:20:10 -0500
> Subject: [PATCH] bluetooth: hci_ldisc: fix deadlock condition
> 
> LDISCs shouldn't call tty->ops->write() from within
> ->write_wakeup().
> 
> ->write_wakeup() is called with port lock taken and
> IRQs disabled, tty->ops->write() will try to acquire
> the same port lock and we will deadlock.
> 
> Signed-off-by: Felipe Balbi <balbi@xxxxxx>
> ---
>  drivers/bluetooth/hci_ldisc.c | 20 +++++++++++++++-----
>  drivers/bluetooth/hci_uart.h  |  1 +
>  2 files changed, 16 insertions(+), 5 deletions(-)
> 
> diff --git a/drivers/bluetooth/hci_ldisc.c b/drivers/bluetooth/hci_ldisc.c
> index 6e06f6f..ecdd765 100644
> --- a/drivers/bluetooth/hci_ldisc.c
> +++ b/drivers/bluetooth/hci_ldisc.c
> @@ -118,10 +118,6 @@ static inline struct sk_buff *hci_uart_dequeue(struct hci_uart *hu)
>  
>  int hci_uart_tx_wakeup(struct hci_uart *hu)
>  {
> -	struct tty_struct *tty = hu->tty;
> -	struct hci_dev *hdev = hu->hdev;
> -	struct sk_buff *skb;
> -
>  	if (test_and_set_bit(HCI_UART_SENDING, &hu->tx_state)) {
>  		set_bit(HCI_UART_TX_WAKEUP, &hu->tx_state);
>  		return 0;
> @@ -129,6 +125,18 @@ int hci_uart_tx_wakeup(struct hci_uart *hu)
>  
>  	BT_DBG("");
>  
> +	schedule_work(&hu->write_work);
> +
> +	return 0;
> +}
> +
> +static void hci_uart_write_work(struct work_struct *work)
> +{
> +	struct hci_uart *hu = container_of(work, struct hci_uart, init_ready);
> +	struct tty_struct *tty = hu->tty;
> +	struct hci_dev *hdev = hu->hdev;
> +	struct sk_buff *skb;
> +
>  restart:
>  	clear_bit(HCI_UART_TX_WAKEUP, &hu->tx_state);
>  
> @@ -153,7 +161,6 @@ restart:
>  		goto restart;
>  
>  	clear_bit(HCI_UART_SENDING, &hu->tx_state);
> -	return 0;
>  }
>  
>  static void hci_uart_init_work(struct work_struct *work)
> @@ -281,6 +288,7 @@ static int hci_uart_tty_open(struct tty_struct *tty)
>  	tty->receive_room = 65536;
>  
>  	INIT_WORK(&hu->init_ready, hci_uart_init_work);
> +	INIT_WORK(&hu->write_work, hci_uart_write_work);
>  
>  	spin_lock_init(&hu->rx_lock);
>  
> @@ -318,6 +326,8 @@ static void hci_uart_tty_close(struct tty_struct *tty)
>  	if (hdev)
>  		hci_uart_close(hdev);
>  
> +	cancel_work_sync(&hy->write_work);

forgot to commit, darn it

-- 
balbi

Attachment: signature.asc
Description: Digital signature