[PATCH] tty: serial: fsl_lpuart: fix race on RX DMA shutdown

[PATCH] tty: serial: fsl_lpuart: fix race on RX DMA shutdown

[A. Sverdlin]

From: Alexander Sverdlin <alexander.sverdlin@siemens.com>

From time to time DMA completion can come in the middle of DMA shutdown:

<process ctx>:				<IRQ>:
lpuart32_shutdown()
  lpuart_dma_shutdown()
    del_timer_sync()
					lpuart_dma_rx_complete()
					  lpuart_copy_rx_to_tty()
					    mod_timer()
    lpuart_dma_rx_free()

When the timer fires a bit later, sport->dma_rx_desc is NULL:

Unable to handle kernel NULL pointer dereference at virtual address 0000000000000004
pc : lpuart_copy_rx_to_tty+0xcc/0x5bc
lr : lpuart_timer_func+0x1c/0x2c
Call trace:
 lpuart_copy_rx_to_tty
 lpuart_timer_func
 call_timer_fn
 __run_timers.part.0
 run_timer_softirq
 __do_softirq
 __irq_exit_rcu
 irq_exit
 handle_domain_irq
 gic_handle_irq
 call_on_irq_stack
 do_interrupt_handler
 ...

To fix this fold del_timer_sync() into lpuart_dma_rx_free() after
dmaengine_terminate_sync() to make sure timer will not be re-started in
lpuart_copy_rx_to_tty() <= lpuart_dma_rx_complete().

Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com>
---
 drivers/tty/serial/fsl_lpuart.c | 11 +++--------
 1 file changed, 3 insertions(+), 8 deletions(-)

diff --git a/drivers/tty/serial/fsl_lpuart.c b/drivers/tty/serial/fsl_lpuart.c
index e945f41b93d43..47c267ee22e04 100644
--- a/drivers/tty/serial/fsl_lpuart.c
+++ b/drivers/tty/serial/fsl_lpuart.c
@@ -1354,6 +1354,7 @@ static void lpuart_dma_rx_free(struct uart_port *port)
 	struct dma_chan *chan = sport->dma_rx_chan;
 
 	dmaengine_terminate_sync(chan);
+	del_timer_sync(&sport->lpuart_timer);
 	dma_unmap_sg(chan->device->dev, &sport->rx_sgl, 1, DMA_FROM_DEVICE);
 	kfree(sport->rx_ring.buf);
 	sport->rx_ring.tail = 0;
@@ -1813,7 +1814,6 @@ static int lpuart32_startup(struct uart_port *port)
 static void lpuart_dma_shutdown(struct lpuart_port *sport)
 {
 	if (sport->lpuart_dma_rx_use) {
-		del_timer_sync(&sport->lpuart_timer);
 		lpuart_dma_rx_free(&sport->port);
 		sport->lpuart_dma_rx_use = false;
 	}
@@ -1973,10 +1973,8 @@ lpuart_set_termios(struct uart_port *port, struct ktermios *termios,
 	 * Since timer function acqures sport->port.lock, need to stop before
 	 * acquring same lock because otherwise del_timer_sync() can deadlock.
 	 */
-	if (old && sport->lpuart_dma_rx_use) {
-		del_timer_sync(&sport->lpuart_timer);
+	if (old && sport->lpuart_dma_rx_use)
 		lpuart_dma_rx_free(&sport->port);
-	}
 
 	spin_lock_irqsave(&sport->port.lock, flags);
 
@@ -2210,10 +2208,8 @@ lpuart32_set_termios(struct uart_port *port, struct ktermios *termios,
 	 * Since timer function acqures sport->port.lock, need to stop before
 	 * acquring same lock because otherwise del_timer_sync() can deadlock.
 	 */
-	if (old && sport->lpuart_dma_rx_use) {
-		del_timer_sync(&sport->lpuart_timer);
+	if (old && sport->lpuart_dma_rx_use)
 		lpuart_dma_rx_free(&sport->port);
-	}
 
 	spin_lock_irqsave(&sport->port.lock, flags);
 
@@ -3014,7 +3010,6 @@ static int lpuart_suspend(struct device *dev)
 			 * cannot resume as expected, hence gracefully release the
 			 * Rx DMA path before suspend and start Rx DMA path on resume.
 			 */
-			del_timer_sync(&sport->lpuart_timer);
 			lpuart_dma_rx_free(&sport->port);
 
 			/* Disable Rx DMA to use UART port as wakeup source */
-- 
2.34.1

Re: [PATCH] tty: serial: fsl_lpuart: fix race on RX DMA shutdown

[Jiri Slaby]

On 06. 03. 23, 10:00, A. Sverdlin wrote:
> From: Alexander Sverdlin <alexander.sverdlin@siemens.com>
> 
>  From time to time DMA completion can come in the middle of DMA shutdown:
> 
> <process ctx>:				<IRQ>:
> lpuart32_shutdown()
>    lpuart_dma_shutdown()
>      del_timer_sync()
> 					lpuart_dma_rx_complete()
> 					  lpuart_copy_rx_to_tty()
> 					    mod_timer()
>      lpuart_dma_rx_free()
> 
> When the timer fires a bit later, sport->dma_rx_desc is NULL:
> 
> Unable to handle kernel NULL pointer dereference at virtual address 0000000000000004
> pc : lpuart_copy_rx_to_tty+0xcc/0x5bc
> lr : lpuart_timer_func+0x1c/0x2c
> Call trace:
>   lpuart_copy_rx_to_tty
>   lpuart_timer_func
>   call_timer_fn
>   __run_timers.part.0
>   run_timer_softirq
>   __do_softirq
>   __irq_exit_rcu
>   irq_exit
>   handle_domain_irq
>   gic_handle_irq
>   call_on_irq_stack
>   do_interrupt_handler
>   ...
> 
> To fix this fold del_timer_sync() into lpuart_dma_rx_free() after
> dmaengine_terminate_sync() to make sure timer will not be re-started in
> lpuart_copy_rx_to_tty() <= lpuart_dma_rx_complete().
> 
> Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com>

This should have some Fixes: tag, I believe.

> ---
>   drivers/tty/serial/fsl_lpuart.c | 11 +++--------
>   1 file changed, 3 insertions(+), 8 deletions(-)
> 
> diff --git a/drivers/tty/serial/fsl_lpuart.c b/drivers/tty/serial/fsl_lpuart.c
> index e945f41b93d43..47c267ee22e04 100644
> --- a/drivers/tty/serial/fsl_lpuart.c
> +++ b/drivers/tty/serial/fsl_lpuart.c
> @@ -1354,6 +1354,7 @@ static void lpuart_dma_rx_free(struct uart_port *port)
>   	struct dma_chan *chan = sport->dma_rx_chan;
>   
>   	dmaengine_terminate_sync(chan);

Maybe I'm missing something (I haven't looked into the code), but what 
happens if the timer ticks here? Won't the dmaengine be restarted there?

> +	del_timer_sync(&sport->lpuart_timer);
>   	dma_unmap_sg(chan->device->dev, &sport->rx_sgl, 1, DMA_FROM_DEVICE);
>   	kfree(sport->rx_ring.buf);
>   	sport->rx_ring.tail = 0;


-- 
js
suse labs

Re: [PATCH] tty: serial: fsl_lpuart: fix race on RX DMA shutdown

[Sverdlin, Alexander]

Hello Jiri,

On Mon, 2023-03-06 at 10:43 +0100, Jiri Slaby wrote:
> On 06. 03. 23, 10:00, A. Sverdlin wrote:
> > From: Alexander Sverdlin <alexander.sverdlin@siemens.com>
> > 
> >  From time to time DMA completion can come in the middle of DMA
> > shutdown:
> > 
> > <process ctx>:                          <IRQ>:
> > lpuart32_shutdown()
> >    lpuart_dma_shutdown()
> >      del_timer_sync()
> >                                         lpuart_dma_rx_complete()
> >                                           lpuart_copy_rx_to_tty()
> >                                             mod_timer()
> >      lpuart_dma_rx_free()
> > 
> > When the timer fires a bit later, sport->dma_rx_desc is NULL:
> > 
> > Unable to handle kernel NULL pointer dereference at virtual address
> > 0000000000000004
> > pc : lpuart_copy_rx_to_tty+0xcc/0x5bc
> > lr : lpuart_timer_func+0x1c/0x2c
> > Call trace:
> >   lpuart_copy_rx_to_tty
> >   lpuart_timer_func
> >   call_timer_fn
> >   __run_timers.part.0
> >   run_timer_softirq
> >   __do_softirq
> >   __irq_exit_rcu
> >   irq_exit
> >   handle_domain_irq
> >   gic_handle_irq
> >   call_on_irq_stack
> >   do_interrupt_handler
> >   ...
> > 
> > To fix this fold del_timer_sync() into lpuart_dma_rx_free() after
> > dmaengine_terminate_sync() to make sure timer will not be re-
> > started in
> > lpuart_copy_rx_to_tty() <= lpuart_dma_rx_complete().
> > 
> > Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com>
> 
> This should have some Fixes: tag, I believe.
> 

sure, you are right, I'll use
Fixes: 4a8588a1cf86 ("serial: fsl_lpuart: delete timer on shutdown")

on respin.

> > ---
> >   drivers/tty/serial/fsl_lpuart.c | 11 +++--------
> >   1 file changed, 3 insertions(+), 8 deletions(-)
> > 
> > diff --git a/drivers/tty/serial/fsl_lpuart.c
> > b/drivers/tty/serial/fsl_lpuart.c
> > index e945f41b93d43..47c267ee22e04 100644
> > --- a/drivers/tty/serial/fsl_lpuart.c
> > +++ b/drivers/tty/serial/fsl_lpuart.c
> > @@ -1354,6 +1354,7 @@ static void lpuart_dma_rx_free(struct
> > uart_port *port)
> >         struct dma_chan *chan = sport->dma_rx_chan;
> >   
> >         dmaengine_terminate_sync(chan);
> 
> Maybe I'm missing something (I haven't looked into the code), but
> what 
> happens if the timer ticks here? Won't the dmaengine be restarted
> there?
> 

It's cyclic (used dmaengine_prep_dma_cyclic()), I see no restart in
lpuart_copy_rx_to_tty(), but maybe someone could double-check.

> > +       del_timer_sync(&sport->lpuart_timer);
> >         dma_unmap_sg(chan->device->dev, &sport->rx_sgl, 1,
> > DMA_FROM_DEVICE);
> >         kfree(sport->rx_ring.buf);
> >         sport->rx_ring.tail = 0;
> 
> 

-- 
Alexander Sverdlin
Siemens AG
www.siemens.com