i2c: ismt: kill transaction in hardware on timeout

On Intel Denverton SoC ismt controller may enter weird state when
transaction gets stuck. It times out in the driver, but unless
transaction is explicitly killed in the controller, it won't be able to
perform new transactions anymore.

The issue is extremely difficult to reproduce and may take weeks of non-
stop smbus traffic.

Numerous hours with logic analyzer didn't yield any useful results, it
looks like the controller stops toggling SCK line, i.e. the issue is
likely in the controller, since device doesn't do clock stretching, so
nothing is driving SCK except the host.

Explicitly kill transaction on timeout to recover the controller from
this state.

Signed-off-by: Vasily Khoruzhick <vasilykh@arista.com>
Signed-off-by: Andi Shyti <andi.shyti@kernel.org>

diff --git a/drivers/i2c/busses/i2c-ismt.c b/drivers/i2c/busses/i2c-ismt.c
index 655b5d851c48b04d6490d6b3e236f1dac23526c3..c93c02aa6ac8cce5ae6a3898a8487ed5d2b0a9c8 100644 (file)
--- a/drivers/i2c/busses/i2c-ismt.c
+++ b/drivers/i2c/busses/i2c-ismt.c
@@ -381,6 +381,15 @@ static int ismt_process_desc(const struct ismt_desc *desc,
        return -EIO;
 }
 
+/**
+ * ismt_kill_transaction() - kill current transaction
+ * @priv: iSMT private data
+ */
+static void ismt_kill_transaction(struct ismt_priv *priv)
+{
+       writel(ISMT_GCTRL_KILL, priv->smba + ISMT_GR_GCTRL);
+}
+
 /**
  * ismt_access() - process an SMBus command
  * @adap: the i2c host adapter
@@ -623,6 +632,7 @@ static int ismt_access(struct i2c_adapter *adap, u16 addr,
                dma_unmap_single(dev, dma_addr, dma_size, dma_direction);
 
        if (unlikely(!time_left)) {
+               ismt_kill_transaction(priv);
                ret = -ETIMEDOUT;
                goto out;
        }