On Wed, Sep 09, 2020 at 01:39:33AM +0300, Dmitry Osipenko wrote:
> The tegra_i2c_flush_fifos() shouldn't sleep in atomic transfer and jiffies
> are not updating if interrupts are disabled. Let's switch to use iopoll
> API helpers for register-polling. The iopoll API provides helpers for both
> atomic and non-atomic cases.
> 
> Note that this patch doesn't fix any known problem because normally FIFO
> is flushed at the time of starting a new transfer.
> 
> Reviewed-by: Michał Mirosław <mirq-linux@rere.qmqm.pl>
> Signed-off-by: Dmitry Osipenko <digetx@gmail.com>
> ---
>  drivers/i2c/busses/i2c-tegra.c | 25 ++++++++++++++++---------
>  1 file changed, 16 insertions(+), 9 deletions(-)
> 
> diff --git a/drivers/i2c/busses/i2c-tegra.c b/drivers/i2c/busses/i2c-tegra.c
> index 00d3e4d7a01e..ab88cdd70376 100644
> --- a/drivers/i2c/busses/i2c-tegra.c
> +++ b/drivers/i2c/busses/i2c-tegra.c
> @@ -470,9 +470,9 @@ static int tegra_i2c_init_dma(struct tegra_i2c_dev *i2c_dev)
>  
>  static int tegra_i2c_flush_fifos(struct tegra_i2c_dev *i2c_dev)
>  {
> -	unsigned long timeout = jiffies + HZ;
> -	unsigned int offset;
> -	u32 mask, val;
> +	u32 mask, val, offset, reg_offset;

Is there are reason why we need reg_offset? Seems to me like we could
simplify this, see below.

> +	void __iomem *addr;
> +	int err;
>  
>  	if (i2c_dev->hw->has_mst_fifo) {
>  		mask = I2C_MST_FIFO_CONTROL_TX_FLUSH |
> @@ -488,12 +488,19 @@ static int tegra_i2c_flush_fifos(struct tegra_i2c_dev *i2c_dev)
>  	val |= mask;
>  	i2c_writel(i2c_dev, val, offset);
>  
> -	while (i2c_readl(i2c_dev, offset) & mask) {
> -		if (time_after(jiffies, timeout)) {
> -			dev_warn(i2c_dev->dev, "timeout waiting for fifo flush\n");
> -			return -ETIMEDOUT;
> -		}
> -		usleep_range(1000, 2000);
> +	reg_offset = tegra_i2c_reg_addr(i2c_dev, offset);
> +	addr = i2c_dev->base + reg_offset;

Why not just:

	offset = tegra_i2c_reg_offset(i2c_dev, offset);
	addr = i2c_dev->base + offset;

or even just:

	addr = i2c_dev->base + tegra_i2c_reg_offset(i2c_dev, offset);

? That makes the patch much smaller because you don't have to rewrite
the whole variable declaration block and just add the "addr" and "err"
variables while removing "timeout".

Thierry