Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1534804 > unrolled thread

[PATCH] tty: serial: fsl_lpuart: fix del_timer_sync() vs timer routine deadlock

Back to article view | Back to linux.kernel

Contents

  [PATCH] tty: serial: fsl_lpuart: fix del_timer_sync() vs timer routine deadlock Nikita Yushchenko <nikita.yoush@cogentembedded.com> - 2016-12-02 11:30 +0100
    Re: [PATCH] tty: serial: fsl_lpuart: fix del_timer_sync() vs timer  routine deadlock Stefan Agner <stefan@agner.ch> - 2016-12-02 19:20 +0100
      Re: [PATCH] tty: serial: fsl_lpuart: fix del_timer_sync() vs timer  routine deadlock Nikita Yushchenko <nikita.yoush@cogentembedded.com> - 2016-12-02 22:30 +0100
        Re: [PATCH] tty: serial: fsl_lpuart: fix del_timer_sync() vs timer  routine deadlock Nikita Yushchenko <nikita.yoush@cogentembedded.com> - 2016-12-03 10:00 +0100

#1534804 — [PATCH] tty: serial: fsl_lpuart: fix del_timer_sync() vs timer routine deadlock

Problem found via lockdep:

- lpuart_set_termios() calls del_timer_sync(&sport->lpuart_timer) while
  holding sport->port.lock

- sport->lpuart_timer routine is lpuart_timer_func() that calls
  lpuart_copy_rx_to_tty() that acquires same lock.

To fix, move Rx DMA stopping out of lock, as it already is in other places
in the same file.

While at it, also make Rx DMA start/stop code to look the same is in
other places in the same file.

Signed-off-by: Nikita Yushchenko <nikita.yoush@cogentembedded.com>
---
 drivers/tty/serial/fsl_lpuart.c | 25 ++++++++++++++-----------
 1 file changed, 14 insertions(+), 11 deletions(-)

diff --git a/drivers/tty/serial/fsl_lpuart.c b/drivers/tty/serial/fsl_lpuart.c
index a1c6519837a4..81100c40f33b 100644
--- a/drivers/tty/serial/fsl_lpuart.c
+++ b/drivers/tty/serial/fsl_lpuart.c
@@ -1407,6 +1407,18 @@ lpuart_set_termios(struct uart_port *port, struct ktermios *termios,
 	/* ask the core to calculate the divisor */
 	baud = uart_get_baud_rate(port, termios, old, 50, port->uartclk / 16);
 
+	/*
+	 * Need to update the Ring buffer length according to the selected
+	 * baud rate and restart Rx DMA path.
+	 *
+	 * Since timer function acqures sport->port.lock, need to stop before
+	 * acquring same lock because otherwise del_timer_sync() can deadlock.
+	 */
+	if (sport->lpuart_dma_rx_use) {
+		del_timer_sync(&sport->lpuart_timer);
+		lpuart_dma_rx_free(&sport->port);
+	}
+
 	spin_lock_irqsave(&sport->port.lock, flags);
 
 	sport->port.read_status_mask = 0;
@@ -1456,17 +1468,8 @@ lpuart_set_termios(struct uart_port *port, struct ktermios *termios,
 	/* restore control register */
 	writeb(old_cr2, sport->port.membase + UARTCR2);
 
-	/*
-	 * If new baud rate is set, we will also need to update the Ring buffer
-	 * length according to the selected baud rate and restart Rx DMA path.
-	 */
-	if (old) {
-		if (sport->lpuart_dma_rx_use) {
-			del_timer_sync(&sport->lpuart_timer);
-			lpuart_dma_rx_free(&sport->port);
-		}
-
-		if (sport->dma_rx_chan && !lpuart_start_rx_dma(sport)) {
+	if (sport->lpuart_dma_rx_use) {
+		if (!lpuart_start_rx_dma(sport)) {
 			sport->lpuart_dma_rx_use = true;
 			rx_dma_timer_init(sport);
 		} else {
-- 
2.1.4

[toc] | [next] | [standalone]

#1535111 — Re: [PATCH] tty: serial: fsl_lpuart: fix del_timer_sync() vs timer routine deadlock

On 2016-12-02 02:28, Nikita Yushchenko wrote:
> Problem found via lockdep:
> 
> - lpuart_set_termios() calls del_timer_sync(&sport->lpuart_timer) while
>   holding sport->port.lock
> 
> - sport->lpuart_timer routine is lpuart_timer_func() that calls
>   lpuart_copy_rx_to_tty() that acquires same lock.
> 
> To fix, move Rx DMA stopping out of lock, as it already is in other places
> in the same file.
> 
> While at it, also make Rx DMA start/stop code to look the same is in
> other places in the same file.

Yeah I saw that too, never really came around to look closer into it.

Thanks for looking into it.

You removed the check whether there was an old configuration, I think
the idea of that was that we only resize DMA if it was configured
differently before... Not sure how startup/set_termios calls are
ordered. I guess in practice that shouldn't really make a difference
since lpuart_dma_rx_use can't be true until after startup?

One nit below.

--
Stefan

> 
> Signed-off-by: Nikita Yushchenko <nikita.yoush@cogentembedded.com>
> ---
>  drivers/tty/serial/fsl_lpuart.c | 25 ++++++++++++++-----------
>  1 file changed, 14 insertions(+), 11 deletions(-)
> 
> diff --git a/drivers/tty/serial/fsl_lpuart.c b/drivers/tty/serial/fsl_lpuart.c
> index a1c6519837a4..81100c40f33b 100644
> --- a/drivers/tty/serial/fsl_lpuart.c
> +++ b/drivers/tty/serial/fsl_lpuart.c
> @@ -1407,6 +1407,18 @@ lpuart_set_termios(struct uart_port *port,
> struct ktermios *termios,
>  	/* ask the core to calculate the divisor */
>  	baud = uart_get_baud_rate(port, termios, old, 50, port->uartclk / 16);
>  
> +	/*
> +	 * Need to update the Ring buffer length according to the selected
> +	 * baud rate and restart Rx DMA path.
> +	 *
> +	 * Since timer function acqures sport->port.lock, need to stop before
> +	 * acquring same lock because otherwise del_timer_sync() can deadlock.
> +	 */
> +	if (sport->lpuart_dma_rx_use) {
> +		del_timer_sync(&sport->lpuart_timer);
> +		lpuart_dma_rx_free(&sport->port);
> +	}
> +
>  	spin_lock_irqsave(&sport->port.lock, flags);
>  
>  	sport->port.read_status_mask = 0;
> @@ -1456,17 +1468,8 @@ lpuart_set_termios(struct uart_port *port,
> struct ktermios *termios,
>  	/* restore control register */
>  	writeb(old_cr2, sport->port.membase + UARTCR2);
>  
> -	/*
> -	 * If new baud rate is set, we will also need to update the Ring buffer
> -	 * length according to the selected baud rate and restart Rx DMA path.
> -	 */
> -	if (old) {
> -		if (sport->lpuart_dma_rx_use) {
> -			del_timer_sync(&sport->lpuart_timer);
> -			lpuart_dma_rx_free(&sport->port);
> -		}
> -
> -		if (sport->dma_rx_chan && !lpuart_start_rx_dma(sport)) {
> +	if (sport->lpuart_dma_rx_use) {
> +		if (!lpuart_start_rx_dma(sport)) {
>  			sport->lpuart_dma_rx_use = true;

No need to set to true here, it is guaranteed to be true at this point.

>  			rx_dma_timer_init(sport);
>  		} else {

[toc] | [prev] | [next] | [standalone]

#1535234 — Re: [PATCH] tty: serial: fsl_lpuart: fix del_timer_sync() vs timer routine deadlock

>> Problem found via lockdep:
>>
>> - lpuart_set_termios() calls del_timer_sync(&sport->lpuart_timer) while
>>   holding sport->port.lock
>>
>> - sport->lpuart_timer routine is lpuart_timer_func() that calls
>>   lpuart_copy_rx_to_tty() that acquires same lock.
>>
>> To fix, move Rx DMA stopping out of lock, as it already is in other places
>> in the same file.
>>
>> While at it, also make Rx DMA start/stop code to look the same is in
>> other places in the same file.
> 
> Yeah I saw that too, never really came around to look closer into it.
> 
> Thanks for looking into it.
> 
> You removed the check whether there was an old configuration, I think
> the idea of that was that we only resize DMA if it was configured
> differently before...

Per my code reading, checking for sport->lpuart_dma_rx_use should be
enough, this flag will be set only if DMA was previously enabled,

>> +	if (sport->lpuart_dma_rx_use) {
>> +		if (!lpuart_start_rx_dma(sport)) {
>>  			sport->lpuart_dma_rx_use = true;
> 
> No need to set to true here, it is guaranteed to be true at this point.

I've seen this...  However elsewhere in this file (namely in
lpuart_resume(), in very similar situation, code is exactly the same,
i.e. it sets sport->lpuart_dma_rx_use in both clauses. I thought it
could be for a reason (i.e. for readability).

Nikita

[toc] | [prev] | [next] | [standalone]

#1535349 — Re: [PATCH] tty: serial: fsl_lpuart: fix del_timer_sync() vs timer routine deadlock

03.12.2016 10:06, Bhuvanchandra DV пишет:
> On 12/03/2016 02:58 AM, Nikita Yushchenko wrote:
> 
>>>> Problem found via lockdep:
>>>>
>>>> - lpuart_set_termios() calls del_timer_sync(&sport->lpuart_timer) while
>>>>    holding sport->port.lock
>>>>
>>>> - sport->lpuart_timer routine is lpuart_timer_func() that calls
>>>>    lpuart_copy_rx_to_tty() that acquires same lock.
>>>>
>>>> To fix, move Rx DMA stopping out of lock, as it already is in other
>>>> places
>>>> in the same file.
>>>>
>>>> While at it, also make Rx DMA start/stop code to look the same is in
>>>> other places in the same file.
>>> Yeah I saw that too, never really came around to look closer into it.
>>>
>>> Thanks for looking into it.
>>>
>>> You removed the check whether there was an old configuration, I think
>>> the idea of that was that we only resize DMA if it was configured
>>> differently before...
>> Per my code reading, checking for sport->lpuart_dma_rx_use should be
>> enough, this flag will be set only if DMA was previously enabled,
> 
> The check is to make sure the reconfiguration of DMA is done only when
> the baudrate is altered.

Then, ok to use

if (old && sport->lpuart_dma_rx_use) {...}

in both places?

[toc] | [prev] | [standalone]

Back to top | Article view | linux.kernel

csiph-web