Re: [PATCH] i2c-designware: fix RX FIFO overrun

[Mika Westerberg <mika.westerberg@linux.intel.com>]

On Mon, Apr 22, 2013 at 04:30:41AM -0700, Josef Ahmad wrote:
> > On Fri, Apr 19, 2013 at 07:05:30PM +0100, Josef Ahmad wrote:
> >> >From a969728248c3b439dc97a69e7dac133b5efa34e7 Mon Sep 17 00:00:00 2001
> >> From: Josef Ahmad <josef.ahmad@linux.intel.com>
> >> Date: Fri, 19 Apr 2013 17:28:10 +0100
> >> Subject: [PATCH] i2c-designware: fix RX FIFO overrun
> >>
> >> i2c_dw_xfer_msg() pushes a number of bytes to transmit/receive
> >> to/from the bus into the TX FIFO.
> >> For master-rx transactions, the maximum amount of data that can be
> >> received is calculated depending solely on TX and RX FIFO load.
> >>
> >> This is racy - TX FIFO may contain master-rx data yet to be
> >> processed, which will eventually land into the RX FIFO. This
> >> data is not taken into account and the function may request more
> >> data than the controller is actually capable of storing.
> >>
> >> This patch ensures the driver takes into account the outstanding
> >> master-rx data in TX FIFO to prevent RX FIFO overrun.
> >
> > Can you add something to the changelog to show what the error looks like
> > (a dump from dmesg for example)?
> >
> 
> The issue is, the data is silently corrupted and not notified to the I2C core
> driver.
> The master-rx transaction returns success and the RX buffer overflow is not
> reported by the driver, which will read dropped data.

OK.

> FWIW, I have a simple test application receiving well-known data from a slave
> in a single transaction, and comparing received to expected data.
> Here's the outcome of three runs:
> 
> [19/03/13 20:30:14] i2c-error   : rcv'd message != ref msg (first diff
> @byte 33)
> [19/03/13 20:30:43] i2c-error   : rcv'd message != ref msg (first diff
> @byte 108)
> [19/03/13 20:31:24] i2c-error   : rcv'd message != ref msg (first diff
> @byte 133)
> 
> >> Signed-off-by: Josef Ahmad <josef.ahmad@linux.intel.com>
> >> ---
> >>  drivers/i2c/busses/i2c-designware-core.c |   11 ++++++++++-
> >>  drivers/i2c/busses/i2c-designware-core.h |    2 ++
> >>  2 files changed, 12 insertions(+), 1 deletions(-)
> >>
> >> diff --git a/drivers/i2c/busses/i2c-designware-core.c
> >> b/drivers/i2c/busses/i2c-designware-core.c
> >> index 94fd818..8dbeef1 100644
> >> --- a/drivers/i2c/busses/i2c-designware-core.c
> >> +++ b/drivers/i2c/busses/i2c-designware-core.c
> >> @@ -426,8 +426,14 @@ i2c_dw_xfer_msg(struct dw_i2c_dev *dev)
> >>  				cmd |= BIT(9);
> >>
> >>  			if (msgs[dev->msg_write_idx].flags & I2C_M_RD) {
> >> +
> >> +				/* avoid rx buffer overrun */
> >> +				if (rx_limit - dev->rx_outstanding <= 0)
> >> +					break;
> >> +
> >>  				dw_writel(dev, cmd | 0x100, DW_IC_DATA_CMD);
> >>  				rx_limit--;
> >> +				dev->rx_outstanding++;
> >
> > Instead of adding a new variable, is there something preventing a use of
> > DW_IC_STATUS bits RFNE and TFNF?
> >
> 
> DW_IC_STATUS bits won't give information of the type of elements (read or
> write) that are in the fifos.
> What we need here is more specific information, i.e. how many RX elements are
> currently in TX fifo. The register set doesn't provide this information to my
> knowledge, so I had to work it out externally with a status variable.
> 
> Consider this example with 8-byte fifos (E=empty, R=read, W=write elements):
> 
> State of the fifos:
>         +-----------------+
> TX  ->  | E E E E E W W R |
>         +-----------------+
>         +-----------------+
> RX      | E E R R R R R R | <-
>         +-----------------+
> 
> Now, say the transaction requires to pump 2 additional R elements into TX
> fifo. We need to ensure that at this stage only 1 of the 2 R elements is
> actually put into TX fifo: this way we we won't saturate the RX fifo.
> Failing to do so exposes a race condition: if we don't read RX quickly
> enough,
> the R element + the new 2 R elements in the TX fifo will land into the RX,
> resulting in an element being dropped.

Thanks for the explanation. Makes sense to me now.

Feel free to add my

Acked-by: Mika Westerberg <mika.westerberg@linux.intel.com>