On Sun, Mar 03, 2019 at 04:03:13PM +0100, Wolfram Sang wrote:
> From: Hiromitsu Yamasaki <hiromitsu.yamasaki.ym@xxxxxxxxxxx>
>
> This patch fixes the problem that an interrupt may set up a new I2C
> message and the DMA callback overwrites this setup.
>
> By disabling the DMA Enable Register(ICDMAER), rcar_i2c_dma_unmap()
> enables interrupts for register settings (such as Master Control
> Register(ICMCR)) and advances the I2C transfer sequence.
>
> If an interrupt occurs immediately after ICDMAER is disabled, the
> callback handler later continues and overwrites the previous settings
> from the interrupt. So, disable ICDMAER at the end of the callback to
> ensure other interrupts are masked until then.
>
> Note that this driver needs to work lock-free because there are IP cores
> with a HW race condition which prevent us from using a spinlock in the
> interrupt handler.
>
> Reproduction test:
> 1. Add udelay(1) after disabling ICDMAER. (It is expected to
> generate an interrupt of rcar_i2c_irq())
>
> void rcar_i2c_dma_unmap(struct rcar_i2c_priv *priv)
> {
> ...
> rcar_i2c_write(priv, ICDMAER, 0);
> udelay(1); <-- Add this line
> ...
> priv->dma_direction = DMA_NONE;
> }
>
> 2. Execute DMA transfer.
> Performs DMA transfer of read or write. In the sample, write was used.
>
> $ i2cset -y -f 4 0x6a 0x01 0x01 0x02 0x03 0x04 0x05 0x06 0x07 0x08 i
>
> 3. A log message of BUG_ON() will be displayed.
>
> Signed-off-by: Hiromitsu Yamasaki <hiromitsu.yamasaki.ym@xxxxxxxxxxx>
> Signed-off-by: Wolfram Sang <wsa+renesas@xxxxxxxxxxxxxxxxxxxx>
Updated the test case, added a note about all this to the comment and
applied to for-current, thanks!
Attachment:
signature.asc
Description: PGP signature
