On Mon, Aug 12, 2019 at 04:58:12PM +0530, Krishna Yarlagadda wrote:
> From: Ahung Cheng <ahcheng@nvidia.com>
> 
> This avoids two race conditions from the UART shutdown sequence both
> leading to 'Machine check error in AXI2APB' and kernel oops.
> 
> One was that the clock was disabled before the DMA was terminated making
> it possible for the DMA callbacks to be called after the clock was
> disabled. These callbacks could write to the UART registers causing
> timeout.
> 
> The second was that the clock was disabled before the UART was
> completely flagged as closed. This is done after the shutdown is called
> and a new write could be started after the clock was disabled.
> tegra_uart_start_pio_tx could be called causing timeout.
> 
> Given that the baud rate is reset at the end of shutdown sequence, this
> fix is to examine the baud rate to avoid register access from both race
> conditions.
> 
> Besides, terminate the DMA before disabling the clock.
> 
> Signed-off-by: Ahung Cheng <ahcheng@nvidia.com>
> Signed-off-by: Shardar Mohammed <smohammed@nvidia.com>
> Signed-off-by: Krishna Yarlagadda <kyarlagadda@nvidia.com>
> ---
>  drivers/tty/serial/serial-tegra.c | 17 +++++++++++------
>  1 file changed, 11 insertions(+), 6 deletions(-)
> 
> diff --git a/drivers/tty/serial/serial-tegra.c b/drivers/tty/serial/serial-tegra.c
> index 93d299e..d908465 100644
> --- a/drivers/tty/serial/serial-tegra.c
> +++ b/drivers/tty/serial/serial-tegra.c
> @@ -126,6 +126,8 @@ struct tegra_uart_port {
>  
>  static void tegra_uart_start_next_tx(struct tegra_uart_port *tup);
>  static int tegra_uart_start_rx_dma(struct tegra_uart_port *tup);
> +static void tegra_uart_dma_channel_free(struct tegra_uart_port *tup,
> +					bool dma_to_memory);
>  
>  static inline unsigned long tegra_uart_read(struct tegra_uart_port *tup,
>  		unsigned long reg)
> @@ -458,6 +460,9 @@ static void tegra_uart_start_next_tx(struct tegra_uart_port *tup)
>  	unsigned long count;
>  	struct circ_buf *xmit = &tup->uport.state->xmit;
>  
> +	if (WARN_ON(!tup->current_baud))
> +		return;

Are the race conditions that you are describing something which can be
triggered by the user? If so, it's not a good idea to use a WARN_ON,
because that could lead to some userspace spamming the log with these,
potentially on purpose.

Thierry

> +
>  	tail = (unsigned long)&xmit->buf[xmit->tail];
>  	count = CIRC_CNT_TO_END(xmit->head, xmit->tail, UART_XMIT_SIZE);
>  	if (!count)
> @@ -829,6 +834,12 @@ static void tegra_uart_hw_deinit(struct tegra_uart_port *tup)
>  	tup->current_baud = 0;
>  	spin_unlock_irqrestore(&tup->uport.lock, flags);
>  
> +	tup->rx_in_progress = 0;
> +	tup->tx_in_progress = 0;
> +
> +	tegra_uart_dma_channel_free(tup, true);
> +	tegra_uart_dma_channel_free(tup, false);
> +
>  	clk_disable_unprepare(tup->uart_clk);
>  }
>  
> @@ -1066,12 +1077,6 @@ static void tegra_uart_shutdown(struct uart_port *u)
>  	struct tegra_uart_port *tup = to_tegra_uport(u);
>  
>  	tegra_uart_hw_deinit(tup);
> -
> -	tup->rx_in_progress = 0;
> -	tup->tx_in_progress = 0;
> -
> -	tegra_uart_dma_channel_free(tup, true);
> -	tegra_uart_dma_channel_free(tup, false);
>  	free_irq(u->irq, tup);
>  }
>  
> -- 
> 2.7.4
>