| Message ID | 553F7B11.5000204@nokia.com |
|---|---|
| State | Superseded |
| Headers | show |
On Tue, Apr 28, 2015 at 02:20:33PM +0200, Alexander Sverdlin wrote: > Use recovery framework and implement bus recovery using "Bus Monitor" > register. Tests show that shortening SCL to GND results in "completion" > timeout with "BUSY" bit still set. So initiate recovery in case of generic > timeout and "Arbitration Lost" condition. So, how could you verify in this setup if the recovery actually worked? > + switch (idev->msg_err) { > + case -ETIMEDOUT: > + case -EAGAIN: > + i2c_recover_bus(&idev->adapter); > + } > + Doesn't look right. -EAGAIN means arbitration lost which is not a condition to reset the bus in a multi-master setup. Other than that, it looks like a nice example of using this framework. Just a few little helper functions and the core will do the rest.
Am 12.05.2015 um 21:12 schrieb Wolfram Sang: > Doesn't look right. -EAGAIN means arbitration lost which is not a > condition to reset the bus in a multi-master setup. IMHO in multi-master scenario bus recovery is never save. There is need of higher level synchronization to find out whether any master is active.
On Wed, May 13, 2015 at 08:19:59AM +0200, Michael Lawnick wrote: > Am 12.05.2015 um 21:12 schrieb Wolfram Sang: > >Doesn't look right. -EAGAIN means arbitration lost which is not a > >condition to reset the bus in a multi-master setup. > > IMHO in multi-master scenario bus recovery is never save. There is > need of higher level synchronization to find out whether any master > is active. Right. But you might be lucky and it's ok when done on error conditions like "sda is low since 300 ms". But on arbitration loss it's definitely wrong. Consider two masters where one resets the bus while the other does a transfer. That results in a arbitration loss for the latter. So the latter resets the bus and in the meantime the first one does a transfer now that the bus should be free. You see the loop? Best regards Uwe
Am 13.05.2015 um 08:55 schrieb ext Uwe Kleine-König: > On Wed, May 13, 2015 at 08:19:59AM +0200, Michael Lawnick wrote: >> Am 12.05.2015 um 21:12 schrieb Wolfram Sang: >>> Doesn't look right. -EAGAIN means arbitration lost which is not a >>> condition to reset the bus in a multi-master setup. >> >> IMHO in multi-master scenario bus recovery is never save. There is >> need of higher level synchronization to find out whether any master >> is active. > Right. But you might be lucky and it's ok when done on error conditions > like "sda is low since 300 ms". But on arbitration loss it's definitely > wrong. Consider two masters where one resets the bus while the other > does a transfer. That results in a arbitration loss for the latter. So > the latter resets the bus and in the meantime the first one does a transfer > now that the bus should be free. You see the loop? Of course. The question is: do/can controllers properly distinguish between arbitration loss and permanent SDL low? I always worked with single master assumption in last 10 years and /feel/ /to/ /remember/ to have observed situations were bus lock was signaled as arbitration loss. I'm simply in doubt that multi master can be handled at driver level at all. This might mean in consequence that unlocking the bus should not be done automatically, but by an extra I2C method triggered by application level.
Hello, On Wed, May 13, 2015 at 09:20:48AM +0200, Michael Lawnick wrote: > Am 13.05.2015 um 08:55 schrieb ext Uwe Kleine-König: > >On Wed, May 13, 2015 at 08:19:59AM +0200, Michael Lawnick wrote: > >>Am 12.05.2015 um 21:12 schrieb Wolfram Sang: > >>>Doesn't look right. -EAGAIN means arbitration lost which is not a > >>>condition to reset the bus in a multi-master setup. > >> > >>IMHO in multi-master scenario bus recovery is never save. There is > >>need of higher level synchronization to find out whether any master > >>is active. > >Right. But you might be lucky and it's ok when done on error conditions > >like "sda is low since 300 ms". But on arbitration loss it's definitely > >wrong. Consider two masters where one resets the bus while the other > >does a transfer. That results in a arbitration loss for the latter. So > >the latter resets the bus and in the meantime the first one does a transfer > >now that the bus should be free. You see the loop? > > Of course. > The question is: do/can controllers properly distinguish between > arbitration loss and permanent SDL low? I always worked with single > master assumption in last 10 years and /feel/ /to/ /remember/ to > have observed situations were bus lock was signaled as arbitration > loss. yeah, that might be true for some (or even most) controllers. Still from a high-level view "arbitration lost" means the controller started the transfer and in the middle observed a low SDA while it doesn't pull it low. If SDA is stuck low instead the controller cannot even put a start condition on the bus. > I'm simply in doubt that multi master can be handled at driver level > at all. This might mean in consequence that unlocking the bus should > not be done automatically, but by an extra I2C method triggered by > application level. At least it should not automatically triggered on error conditions that might result from another master on the bus. IMHO doing it once at probe time and then only when SDA is stuck low for say 1s should be fine. Best regards Uwe
Hi! On 12/05/15 21:12, ext Wolfram Sang wrote: >> Use recovery framework and implement bus recovery using "Bus Monitor" >> > register. Tests show that shortening SCL to GND results in "completion" >> > timeout with "BUSY" bit still set. So initiate recovery in case of generic >> > timeout and "Arbitration Lost" condition. > So, how could you verify in this setup if the recovery actually worked? Yes, it's tested with jumpers and oscilloscope on a custom Axxia AXM5508-based board. 9 CLK pulses are there and it's also jumping out of recovery if SDA is not low any more. >> > + switch (idev->msg_err) { >> > + case -ETIMEDOUT: >> > + case -EAGAIN: >> > + i2c_recover_bus(&idev->adapter); >> > + } >> > + > Doesn't look right. -EAGAIN means arbitration lost which is not a > condition to reset the bus in a multi-master setup. I'll resend with only -ETIMEDOUT triggering the recovery.
diff --git a/drivers/i2c/busses/i2c-axxia.c b/drivers/i2c/busses/i2c-axxia.c index ad0ad15..070fd31 100644 --- a/drivers/i2c/busses/i2c-axxia.c +++ b/drivers/i2c/busses/i2c-axxia.c @@ -42,6 +42,10 @@ #define IBML_LOW_SEXT 0x18 #define TIMER_CLOCK_DIV 0x1c #define I2C_BUS_MONITOR 0x20 +#define BM_SDAC BIT(3) +#define BM_SCLC BIT(2) +#define BM_SDAS BIT(1) +#define BM_SCLS BIT(0) #define SOFT_RESET 0x24 #define MST_COMMAND 0x28 #define CMD_BUSY (1<<3) @@ -394,6 +398,12 @@ static int axxia_i2c_xfer_msg(struct axxia_i2c_dev *idev, struct i2c_msg *msg) if (time_left == 0) idev->msg_err = -ETIMEDOUT; + switch (idev->msg_err) { + case -ETIMEDOUT: + case -EAGAIN: + i2c_recover_bus(&idev->adapter); + } + if (unlikely(idev->msg_err) && idev->msg_err != -ENXIO) axxia_i2c_init(idev); @@ -437,6 +447,39 @@ axxia_i2c_xfer(struct i2c_adapter *adap, struct i2c_msg msgs[], int num) return ret ? : i; } +static int axxia_i2c_get_scl(struct i2c_adapter *adap) +{ + struct axxia_i2c_dev *idev = i2c_get_adapdata(adap); + + return !!(readl(idev->base + I2C_BUS_MONITOR) & BM_SCLS); +} + +static void axxia_i2c_set_scl(struct i2c_adapter *adap, int val) +{ + struct axxia_i2c_dev *idev = i2c_get_adapdata(adap); + u32 tmp; + + /* Preserve SDA Control */ + tmp = readl(idev->base + I2C_BUS_MONITOR) & BM_SDAC; + if (!val) + tmp |= BM_SCLC; + writel(tmp, idev->base + I2C_BUS_MONITOR); +} + +static int axxia_i2c_get_sda(struct i2c_adapter *adap) +{ + struct axxia_i2c_dev *idev = i2c_get_adapdata(adap); + + return !!(readl(idev->base + I2C_BUS_MONITOR) & BM_SDAS); +} + +static struct i2c_bus_recovery_info axxia_i2c_recovery_info = { + .recover_bus = i2c_generic_scl_recovery, + .get_scl = axxia_i2c_get_scl, + .set_scl = axxia_i2c_set_scl, + .get_sda = axxia_i2c_get_sda, +}; + static u32 axxia_i2c_func(struct i2c_adapter *adap) { u32 caps = (I2C_FUNC_I2C | I2C_FUNC_10BIT_ADDR | @@ -511,6 +554,7 @@ static int axxia_i2c_probe(struct platform_device *pdev) strlcpy(idev->adapter.name, pdev->name, sizeof(idev->adapter.name)); idev->adapter.owner = THIS_MODULE; idev->adapter.algo = &axxia_i2c_algo; + idev->adapter.bus_recovery_info = &axxia_i2c_recovery_info; idev->adapter.quirks = &axxia_i2c_quirks; idev->adapter.dev.parent = &pdev->dev; idev->adapter.dev.of_node = pdev->dev.of_node;
Use recovery framework and implement bus recovery using "Bus Monitor" register. Tests show that shortening SCL to GND results in "completion" timeout with "BUSY" bit still set. So initiate recovery in case of generic timeout and "Arbitration Lost" condition. Signed-off-by: Alexander Sverdlin <alexander.sverdlin@nokia.com> --- drivers/i2c/busses/i2c-axxia.c | 44 ++++++++++++++++++++++++++++++++++++++++ 1 files changed, 44 insertions(+), 0 deletions(-) -- To unsubscribe from this list: send the line "unsubscribe linux-i2c" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html