> > > >> Looking into timestamps and transactions, DMA timeouts after start of DMA for I2C1 to touch during this transaction. > > > >> While it is waiting for I2C1 DMA transfer, lots of DVC > > > >> transactions happened thru DMA which are successful > > > >> > > > >> What is the I2C1 speed? > > > > > > > > 400KHz > > > > > > > >> Also incase if device is running slow for some reason, probably timeout was not enough as this patch series changes timeout with base 100mS + msg transfer time based on transfer size. > > > >> Can you give quick try with increased timeout incase if device is running slow? > > > >> > > > > > > > > Tried to increase the timeout to 1 second, doesn't help. > > > > > > > > What helped again is the I2C HW resetting after each transfer. Likely that means that HW isn't programmed correctly, please carefully check every bit. > > > > > > > > DMA-only + I2C HW reset: http://dpaste.com/26AQXFM.txt > > > > > > Seems I found what's the problem. Here is the fix, please include it to v14 if it is correct. > > > > > > diff --git a/drivers/i2c/busses/i2c-tegra.c > > > b/drivers/i2c/busses/i2c-tegra.c index a9391c3646b6..5ad54da70304 > > > 100644 > > > --- a/drivers/i2c/busses/i2c-tegra.c > > > +++ b/drivers/i2c/busses/i2c-tegra.c > > > @@ -912,7 +912,7 @@ static irqreturn_t tegra_i2c_isr(int irq, void *dev_id) static void tegra_i2c_config_fifo_trig(struct tegra_i2c_dev *i2c_dev, > > > size_t len) { > > > - u32 val, reg; > > > + u32 val = 0, reg; > > > u8 dma_burst = 0; > > > struct dma_slave_config slv_config = {0}; > > > struct dma_chan *chan; > > > @@ -922,7 +922,6 @@ static void tegra_i2c_config_fifo_trig(struct tegra_i2c_dev *i2c_dev, > > > reg = I2C_MST_FIFO_CONTROL; > > > else > > > reg = I2C_FIFO_CONTROL; > > > - val = i2c_readl(i2c_dev, reg); > > > > > > if (i2c_dev->is_curr_dma_xfer) { > > > if (len & 0xF) > > > > Thanks dmitry. Good catch. Didn't caught to my eyes. Yes FIFO_CONTROL > > is set based on read value without masking and oring causing back-back DMA transfers resulting in incorrect trig levels We can directly set trig levels. > > It might be safer to explicitly clear the trigger levels before we overwrite them. Though, admittedly, this register hasn't changed in years, so it's unlikely we'll ever see it change in a way that might result in breakage if we construct the register value from scratch. > > Thierry FIFO_CONTROL has trig and flush for TX & RX and since we do TX/RX and flush anyway prior to transfer, constructing from scratch should be fine as its same as doing mask-modify-write
