[PATCH/RFC v2] serial: sh-sci: Fix fallback to PIO on DMA failure

[PATCH/RFC v2] serial: sh-sci: Fix fallback to PIO on DMA failure

[Geert Uytterhoeven]

When submitting a DMA request fails, the driver is supposed to fall back
to PIO.  However, this never really worked due to various reasons
(sh-sci driver issues and dmaengine framework limitations).

There are three places where DMA submission can fail, and the driver
should fall back to PIO:
  1. sci_dma_rx_complete(),
  2. sci_submit_rx(),
  3. work_fn_tx().

This RFC fixes fallback to PIO in the receive path (cases 1 and 2).
Fallback to PIO in the transmit path (case 3) already works fine.

Case 1: sci_dma_rx_complete()

  A. PIO cannot take over on SCIF if any DMA transactions are pending,
     hence they must be terminated first.

  B. The active cookies must be invalidated, else rx_timer_fn() may
     trigger a NULL pointer dereference.

  C. Restarting the port is not needed, as it is already running, but
     serial port interrupts must be directed back from the DMA engine to
     the CPU.

Case 2: sci_submit_rx()

  D. Some callers of sci_submit_rx() hold the port spinlock, others
     don't.  During fallback to PIO, the driver needs to obtain the port
     spinlock.  If the lock was already held, spinlock recursion is
     detected, causing a deadlock: BUG: spinlock recursion on CPU#0.

     Fix this by adding a flag parameter to sci_submit_rx() for the
     caller to indicate the port spinlock is already held, so spinlock
     recursion can be avoided.

     Move the spin_lock_irqsave() up, so all DMA disable steps are
     protected, which is safe as the recently introduced
     dmaengine_terminate_async() can be called in atomic context.

  E. When falling back to PIO, active_rx must be set to a different
     value than cookie_rx[i], else sci_dma_rx_find_active() will
     incorrectly find a match, leading to a NULL pointer dereference in
     rx_timer_fn() later.

  F. On (H)SCIF, sci_submit_rx() is called in the receive interrupt
     handler.  Hence if DMA submission fails, the interrupt handler
     should handle reception using PIO.

Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
---
Against tty/next + "serial: sh-sci: Fix receive on SCIFA/SCIFB variants
with DMA".

All testing was done on r8a7791/koelsch using SCIF1 on debug serial 1,
and SCIFA3 on EXIO-B, by introducing random failures in DMA submission
code.

Thanks for your comments!

v2:
  - Fix fallback in sci_dma_rx_complete(),
  - Fallback in the transmit path already works fine,
  - Widen audience, but keep RFC.
---
 drivers/tty/serial/sh-sci.c | 34 +++++++++++++++++++++++++---------
 1 file changed, 25 insertions(+), 9 deletions(-)

diff --git a/drivers/tty/serial/sh-sci.c b/drivers/tty/serial/sh-sci.c
index 3aad48e64b9b71ff..30b2728c20d6e982 100644
--- a/drivers/tty/serial/sh-sci.c
+++ b/drivers/tty/serial/sh-sci.c
@@ -1272,6 +1272,7 @@ static void sci_dma_rx_complete(void *arg)
 	struct dma_async_tx_descriptor *desc;
 	unsigned long flags;
 	int active, count = 0;
+	unsigned int i;
 
 	dev_dbg(port->dev, "%s(%d) active cookie %d\n", __func__, port->line,
 		s->active_rx);
@@ -1309,12 +1310,22 @@ static void sci_dma_rx_complete(void *arg)
 	return;
 
 fail:
+	dmaengine_terminate_async(chan);
 	spin_unlock_irqrestore(&port->lock, flags);
 	dev_warn(port->dev, "Failed submitting Rx DMA descriptor\n");
 	/* Switch to PIO */
 	spin_lock_irqsave(&port->lock, flags);
+	for (i = 0; i < 2; i++)
+		s->cookie_rx[i] = -EINVAL;
+	s->active_rx = 0;
 	s->chan_rx = NULL;
-	sci_start_rx(port);
+	/* Direct new serial port interrupts back to CPU */
+	scr = serial_port_in(port, SCSCR);
+	if (port->type == PORT_SCIFA || port->type == PORT_SCIFB) {
+		scr &= ~SCSCR_RDRQE;
+		enable_irq(s->irqs[SCIx_RXI_IRQ]);
+	}
+	serial_port_out(port, SCSCR, scr | SCSCR_RIE);
 	spin_unlock_irqrestore(&port->lock, flags);
 }
 
@@ -1331,7 +1342,7 @@ static void sci_tx_dma_release(struct sci_port *s)
 	dma_release_channel(chan);
 }
 
-static void sci_submit_rx(struct sci_port *s)
+static int sci_submit_rx(struct sci_port *s, bool port_lock_held)
 {
 	struct dma_chan *chan = s->chan_rx;
 	struct uart_port *port = &s->port;
@@ -1359,19 +1370,22 @@ static void sci_submit_rx(struct sci_port *s)
 	s->active_rx = s->cookie_rx[0];
 
 	dma_async_issue_pending(chan);
-	return;
+	return 0;
 
 fail:
+	if (!port_lock_held)
+		spin_lock_irqsave(&port->lock, flags);
 	if (i)
 		dmaengine_terminate_async(chan);
 	for (i = 0; i < 2; i++)
 		s->cookie_rx[i] = -EINVAL;
-	s->active_rx = -EINVAL;
+	s->active_rx = 0;
 	/* Switch to PIO */
-	spin_lock_irqsave(&port->lock, flags);
 	s->chan_rx = NULL;
 	sci_start_rx(port);
-	spin_unlock_irqrestore(&port->lock, flags);
+	if (!port_lock_held)
+		spin_unlock_irqrestore(&port->lock, flags);
+	return -1;
 }
 
 static void work_fn_tx(struct work_struct *work)
@@ -1491,7 +1505,7 @@ static enum hrtimer_restart rx_timer_fn(struct hrtimer *t)
 	}
 
 	if (port->type == PORT_SCIFA || port->type == PORT_SCIFB)
-		sci_submit_rx(s);
+		sci_submit_rx(s, true);
 
 	/* Direct new serial port interrupts back to CPU */
 	scr = serial_port_in(port, SCSCR);
@@ -1617,7 +1631,7 @@ static void sci_request_dma(struct uart_port *port)
 		s->chan_rx_saved = s->chan_rx = chan;
 
 		if (port->type == PORT_SCIFA || port->type == PORT_SCIFB)
-			sci_submit_rx(s);
+			sci_submit_rx(s, false);
 	}
 }
 
@@ -1667,7 +1681,8 @@ static irqreturn_t sci_rx_interrupt(int irq, void *ptr)
 			scr |= SCSCR_RDRQE;
 		} else {
 			scr &= ~SCSCR_RIE;
-			sci_submit_rx(s);
+			if (sci_submit_rx(s, false))
+				goto handle_pio;
 		}
 		serial_port_out(port, SCSCR, scr);
 		/* Clear current interrupt */
@@ -1681,6 +1696,7 @@ static irqreturn_t sci_rx_interrupt(int irq, void *ptr)
 	}
 #endif
 
+handle_pio:
 	if (s->rx_trigger > 1 && s->rx_fifo_timeout > 0) {
 		if (!scif_rtrg_enabled(port))
 			scif_set_rtrg(port, s->rx_trigger);
-- 
2.17.1

Re: [PATCH/RFC v2] serial: sh-sci: Fix fallback to PIO on DMA failure

[Geert Uytterhoeven]

On Tue, Oct 16, 2018 at 3:24 PM Geert Uytterhoeven
<geert+renesas@glider.be> wrote:
> When submitting a DMA request fails, the driver is supposed to fall back
> to PIO.  However, this never really worked due to various reasons
> (sh-sci driver issues and dmaengine framework limitations).
>
> There are three places where DMA submission can fail, and the driver
> should fall back to PIO:
>   1. sci_dma_rx_complete(),
>   2. sci_submit_rx(),
>   3. work_fn_tx().
>
> This RFC fixes fallback to PIO in the receive path (cases 1 and 2).
> Fallback to PIO in the transmit path (case 3) already works fine.
>
> Case 1: sci_dma_rx_complete()
>
>   A. PIO cannot take over on SCIF if any DMA transactions are pending,
>      hence they must be terminated first.
>
>   B. The active cookies must be invalidated, else rx_timer_fn() may
>      trigger a NULL pointer dereference.
>
>   C. Restarting the port is not needed, as it is already running, but
>      serial port interrupts must be directed back from the DMA engine to
>      the CPU.
>
> Case 2: sci_submit_rx()
>
>   D. Some callers of sci_submit_rx() hold the port spinlock, others
>      don't.  During fallback to PIO, the driver needs to obtain the port
>      spinlock.  If the lock was already held, spinlock recursion is
>      detected, causing a deadlock: BUG: spinlock recursion on CPU#0.
>
>      Fix this by adding a flag parameter to sci_submit_rx() for the
>      caller to indicate the port spinlock is already held, so spinlock
>      recursion can be avoided.
>
>      Move the spin_lock_irqsave() up, so all DMA disable steps are
>      protected, which is safe as the recently introduced
>      dmaengine_terminate_async() can be called in atomic context.
>
>   E. When falling back to PIO, active_rx must be set to a different
>      value than cookie_rx[i], else sci_dma_rx_find_active() will
>      incorrectly find a match, leading to a NULL pointer dereference in
>      rx_timer_fn() later.
>
>   F. On (H)SCIF, sci_submit_rx() is called in the receive interrupt
>      handler.  Hence if DMA submission fails, the interrupt handler
>      should handle reception using PIO.
>
> Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
> ---
> Against tty/next + "serial: sh-sci: Fix receive on SCIFA/SCIFB variants
> with DMA".
>
> All testing was done on r8a7791/koelsch using SCIF1 on debug serial 1,
> and SCIFA3 on EXIO-B, by introducing random failures in DMA submission
> code.
>
> Thanks for your comments!
>
> v2:
>   - Fix fallback in sci_dma_rx_complete(),
>   - Fallback in the transmit path already works fine,
>   - Widen audience, but keep RFC.
> ---
>  drivers/tty/serial/sh-sci.c | 34 +++++++++++++++++++++++++---------
>  1 file changed, 25 insertions(+), 9 deletions(-)
>
> diff --git a/drivers/tty/serial/sh-sci.c b/drivers/tty/serial/sh-sci.c
> index 3aad48e64b9b71ff..30b2728c20d6e982 100644
> --- a/drivers/tty/serial/sh-sci.c
> +++ b/drivers/tty/serial/sh-sci.c
> @@ -1272,6 +1272,7 @@ static void sci_dma_rx_complete(void *arg)
>         struct dma_async_tx_descriptor *desc;
>         unsigned long flags;
>         int active, count = 0;
> +       unsigned int i;

Oops, there's a "u16 scr;" missing here.

>
>         dev_dbg(port->dev, "%s(%d) active cookie %d\n", __func__, port->line,
>                 s->active_rx);
> @@ -1309,12 +1310,22 @@ static void sci_dma_rx_complete(void *arg)
>         return;
>
>  fail:
> +       dmaengine_terminate_async(chan);
>         spin_unlock_irqrestore(&port->lock, flags);
>         dev_warn(port->dev, "Failed submitting Rx DMA descriptor\n");
>         /* Switch to PIO */
>         spin_lock_irqsave(&port->lock, flags);
> +       for (i = 0; i < 2; i++)
> +               s->cookie_rx[i] = -EINVAL;
> +       s->active_rx = 0;
>         s->chan_rx = NULL;
> -       sci_start_rx(port);
> +       /* Direct new serial port interrupts back to CPU */
> +       scr = serial_port_in(port, SCSCR);
> +       if (port->type == PORT_SCIFA || port->type == PORT_SCIFB) {
> +               scr &= ~SCSCR_RDRQE;
> +               enable_irq(s->irqs[SCIx_RXI_IRQ]);
> +       }
> +       serial_port_out(port, SCSCR, scr | SCSCR_RIE);
>         spin_unlock_irqrestore(&port->lock, flags);
>  }
>

Gr{oetje,eeting}s,

                        Geert

-- 
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@linux-m68k.org

In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
                                -- Linus Torvalds