| Message ID | 1569919869-3218-1-git-send-email-fabrice.gasnier@st.com |
|---|---|
| State | Accepted |
| Headers | show |
| Series | i2c: i2c-stm32f7: fix a race in slave mode with arbitration loss irq | expand |
Hi Reviewed-by: Pierre-Yves MORDRET <pierre-yves.mordret@st.com> Thx On 10/1/19 10:51 AM, Fabrice Gasnier wrote: > When in slave mode, an arbitration loss (ARLO) may be detected before the > slave had a chance to detect the stop condition (STOPF in ISR). > This is seen when two master + slave adapters switch their roles. It > provokes the i2c bus to be stuck, busy as SCL line is stretched. > - the I2C_SLAVE_STOP event is never generated due to STOPF flag is set but > don't generate an irq (race with ARLO irq, STOPIE is masked). STOPF flag > remains set until next master xfer (e.g. when STOPIE irq get unmasked). > In this case, completion is generated too early: immediately upon new > transfer request (then it doesn't send all data). > - Some data get stuck in TXDR register. As a consequence, the controller > stretches the SCL line: the bus gets busy until a future master transfer > triggers the bus busy / recovery mechanism (this can take time... and > may never happen at all) > > So choice is to let the STOPF being detected by the slave isr handler, > to properly handle this stop condition. E.g. don't mask IRQs in error > handler, when the slave is running. > > Fixes: 60d609f30de2 ("i2c: i2c-stm32f7: Add slave support") > > Signed-off-by: Fabrice Gasnier <fabrice.gasnier@st.com> > --- > drivers/i2c/busses/i2c-stm32f7.c | 17 ++++++++++------- > 1 file changed, 10 insertions(+), 7 deletions(-) > > diff --git a/drivers/i2c/busses/i2c-stm32f7.c b/drivers/i2c/busses/i2c-stm32f7.c > index 266d1c2..3a8ab0c 100644 > --- a/drivers/i2c/busses/i2c-stm32f7.c > +++ b/drivers/i2c/busses/i2c-stm32f7.c > @@ -1501,7 +1501,7 @@ static irqreturn_t stm32f7_i2c_isr_error(int irq, void *data) > void __iomem *base = i2c_dev->base; > struct device *dev = i2c_dev->dev; > struct stm32_i2c_dma *dma = i2c_dev->dma; > - u32 mask, status; > + u32 status; > > status = readl_relaxed(i2c_dev->base + STM32F7_I2C_ISR); > > @@ -1526,12 +1526,15 @@ static irqreturn_t stm32f7_i2c_isr_error(int irq, void *data) > f7_msg->result = -EINVAL; > } > > - /* Disable interrupts */ > - if (stm32f7_i2c_is_slave_registered(i2c_dev)) > - mask = STM32F7_I2C_XFER_IRQ_MASK; > - else > - mask = STM32F7_I2C_ALL_IRQ_MASK; > - stm32f7_i2c_disable_irq(i2c_dev, mask); > + if (!i2c_dev->slave_running) { > + u32 mask; > + /* Disable interrupts */ > + if (stm32f7_i2c_is_slave_registered(i2c_dev)) > + mask = STM32F7_I2C_XFER_IRQ_MASK; > + else > + mask = STM32F7_I2C_ALL_IRQ_MASK; > + stm32f7_i2c_disable_irq(i2c_dev, mask); > + } > > /* Disable dma */ > if (i2c_dev->use_dma) { >
On 10/1/19 10:51 AM, Fabrice Gasnier wrote: > When in slave mode, an arbitration loss (ARLO) may be detected before the > slave had a chance to detect the stop condition (STOPF in ISR). > This is seen when two master + slave adapters switch their roles. It > provokes the i2c bus to be stuck, busy as SCL line is stretched. > - the I2C_SLAVE_STOP event is never generated due to STOPF flag is set but > don't generate an irq (race with ARLO irq, STOPIE is masked). STOPF flag > remains set until next master xfer (e.g. when STOPIE irq get unmasked). > In this case, completion is generated too early: immediately upon new > transfer request (then it doesn't send all data). > - Some data get stuck in TXDR register. As a consequence, the controller > stretches the SCL line: the bus gets busy until a future master transfer > triggers the bus busy / recovery mechanism (this can take time... and > may never happen at all) > > So choice is to let the STOPF being detected by the slave isr handler, > to properly handle this stop condition. E.g. don't mask IRQs in error > handler, when the slave is running. > > Fixes: 60d609f30de2 ("i2c: i2c-stm32f7: Add slave support") > > Signed-off-by: Fabrice Gasnier <fabrice.gasnier@st.com> > --- Hi Wolfram, all, Gentle reminder on this patch. Thanks in advance ! Best Regards, Fabrice > drivers/i2c/busses/i2c-stm32f7.c | 17 ++++++++++------- > 1 file changed, 10 insertions(+), 7 deletions(-) > > diff --git a/drivers/i2c/busses/i2c-stm32f7.c b/drivers/i2c/busses/i2c-stm32f7.c > index 266d1c2..3a8ab0c 100644 > --- a/drivers/i2c/busses/i2c-stm32f7.c > +++ b/drivers/i2c/busses/i2c-stm32f7.c > @@ -1501,7 +1501,7 @@ static irqreturn_t stm32f7_i2c_isr_error(int irq, void *data) > void __iomem *base = i2c_dev->base; > struct device *dev = i2c_dev->dev; > struct stm32_i2c_dma *dma = i2c_dev->dma; > - u32 mask, status; > + u32 status; > > status = readl_relaxed(i2c_dev->base + STM32F7_I2C_ISR); > > @@ -1526,12 +1526,15 @@ static irqreturn_t stm32f7_i2c_isr_error(int irq, void *data) > f7_msg->result = -EINVAL; > } > > - /* Disable interrupts */ > - if (stm32f7_i2c_is_slave_registered(i2c_dev)) > - mask = STM32F7_I2C_XFER_IRQ_MASK; > - else > - mask = STM32F7_I2C_ALL_IRQ_MASK; > - stm32f7_i2c_disable_irq(i2c_dev, mask); > + if (!i2c_dev->slave_running) { > + u32 mask; > + /* Disable interrupts */ > + if (stm32f7_i2c_is_slave_registered(i2c_dev)) > + mask = STM32F7_I2C_XFER_IRQ_MASK; > + else > + mask = STM32F7_I2C_ALL_IRQ_MASK; > + stm32f7_i2c_disable_irq(i2c_dev, mask); > + } > > /* Disable dma */ > if (i2c_dev->use_dma) { >
On Tue, Oct 01, 2019 at 10:51:09AM +0200, Fabrice Gasnier wrote: > When in slave mode, an arbitration loss (ARLO) may be detected before the > slave had a chance to detect the stop condition (STOPF in ISR). > This is seen when two master + slave adapters switch their roles. It > provokes the i2c bus to be stuck, busy as SCL line is stretched. > - the I2C_SLAVE_STOP event is never generated due to STOPF flag is set but > don't generate an irq (race with ARLO irq, STOPIE is masked). STOPF flag > remains set until next master xfer (e.g. when STOPIE irq get unmasked). > In this case, completion is generated too early: immediately upon new > transfer request (then it doesn't send all data). > - Some data get stuck in TXDR register. As a consequence, the controller > stretches the SCL line: the bus gets busy until a future master transfer > triggers the bus busy / recovery mechanism (this can take time... and > may never happen at all) > > So choice is to let the STOPF being detected by the slave isr handler, > to properly handle this stop condition. E.g. don't mask IRQs in error > handler, when the slave is running. > > Fixes: 60d609f30de2 ("i2c: i2c-stm32f7: Add slave support") > > Signed-off-by: Fabrice Gasnier <fabrice.gasnier@st.com> Applied to for-current, thanks!
diff --git a/drivers/i2c/busses/i2c-stm32f7.c b/drivers/i2c/busses/i2c-stm32f7.c index 266d1c2..3a8ab0c 100644 --- a/drivers/i2c/busses/i2c-stm32f7.c +++ b/drivers/i2c/busses/i2c-stm32f7.c @@ -1501,7 +1501,7 @@ static irqreturn_t stm32f7_i2c_isr_error(int irq, void *data) void __iomem *base = i2c_dev->base; struct device *dev = i2c_dev->dev; struct stm32_i2c_dma *dma = i2c_dev->dma; - u32 mask, status; + u32 status; status = readl_relaxed(i2c_dev->base + STM32F7_I2C_ISR); @@ -1526,12 +1526,15 @@ static irqreturn_t stm32f7_i2c_isr_error(int irq, void *data) f7_msg->result = -EINVAL; } - /* Disable interrupts */ - if (stm32f7_i2c_is_slave_registered(i2c_dev)) - mask = STM32F7_I2C_XFER_IRQ_MASK; - else - mask = STM32F7_I2C_ALL_IRQ_MASK; - stm32f7_i2c_disable_irq(i2c_dev, mask); + if (!i2c_dev->slave_running) { + u32 mask; + /* Disable interrupts */ + if (stm32f7_i2c_is_slave_registered(i2c_dev)) + mask = STM32F7_I2C_XFER_IRQ_MASK; + else + mask = STM32F7_I2C_ALL_IRQ_MASK; + stm32f7_i2c_disable_irq(i2c_dev, mask); + } /* Disable dma */ if (i2c_dev->use_dma) {
When in slave mode, an arbitration loss (ARLO) may be detected before the slave had a chance to detect the stop condition (STOPF in ISR). This is seen when two master + slave adapters switch their roles. It provokes the i2c bus to be stuck, busy as SCL line is stretched. - the I2C_SLAVE_STOP event is never generated due to STOPF flag is set but don't generate an irq (race with ARLO irq, STOPIE is masked). STOPF flag remains set until next master xfer (e.g. when STOPIE irq get unmasked). In this case, completion is generated too early: immediately upon new transfer request (then it doesn't send all data). - Some data get stuck in TXDR register. As a consequence, the controller stretches the SCL line: the bus gets busy until a future master transfer triggers the bus busy / recovery mechanism (this can take time... and may never happen at all) So choice is to let the STOPF being detected by the slave isr handler, to properly handle this stop condition. E.g. don't mask IRQs in error handler, when the slave is running. Fixes: 60d609f30de2 ("i2c: i2c-stm32f7: Add slave support") Signed-off-by: Fabrice Gasnier <fabrice.gasnier@st.com> --- drivers/i2c/busses/i2c-stm32f7.c | 17 ++++++++++------- 1 file changed, 10 insertions(+), 7 deletions(-)