I2C bus driver TIMEDOUT because of PM autosuspend
I am analysing a problem with I2C bus driver where the problem shows up as I2C bus completely blocked. The LX driver in question is /drivers/i2c/busses/i2c-xiic.c. Problem is difficult to reproduce, it happens very rarely. So far I saw that the main precondition is to have very heavy I2C traffic on bus. In my case this is achieved/reproduced via netdev driving SFP LEDs via /sys/class/leds/ (via gpio-pca953x). I generate traffic with iperf3. Network traffic is on 10Gbps EMAC. LX kernel is 4.14.0. What I saw from debugging this problem is that I2C bus get blocked when wait_event_timeout() completes because of timeout. The timeout handling in this driver is probably not robust enough (bus should not remain blocked), but at this moment this are just my speculations (don't know enough details). Looking the driver code and data on oscilloscope, I saw that SCL in single I2C data transfer sequence can be interrupted for very long delays, e.g up to hundredths of usec (SCL is 100kHz). I started to suspect that PM autosuspend delay could play some role here. There are only two delays in driver code, first in wait_event_timeout and second in set autosuspend delay. Case is a bit strange because in very busy I2C traffic, PM autosuspend should not be triggered at all. Additionally, if I lower PM timeout, e.g. from 1000 (default) to 100, I hit the problem sooner (waits for problem hit are in order of n*10minutes). It looks to me that PM autosupend is playing some role here. Power management options in my .config: # CONFIG_SUSPEND is not set # CONFIG_PM is not set CONFIG_ARCH_SUSPEND_POSSIBLE=y I intentionally did not put all detail descriptions of embedded system and test setup here (long list), because the main reason of this post is: The workaround that works for me/customer (at the moment) is to disable PM autosuspend in the driver code, either by incerementing PM delay from 1000 to 10000 or by disabling autosuspend (comment out call to pm_runtime_put_autosuspend() in xiic_xfer()). But, I would like to expose/discuss this issue (maintainer of the code, or others). The reason/source of the problem can be much more complex and in some other place. So my question is who should I contact, is this the M: in the MAINTAINERS list, the MODULE_AUTHOR, ...? How to proceed. WBR Primoz
On Fri, Nov 29, 2019 at 12:53 PM Primoz Beltram <primoz.beltram@gmail.com> wrote:
I am analysing a problem with I2C bus driver where the problem shows up as I2C bus completely blocked. The LX driver in question is /drivers/i2c/busses/i2c-xiic.c. Problem is difficult to reproduce, it happens very rarely. So far I saw that the main precondition is to have very heavy I2C traffic on bus. In my case this is achieved/reproduced via netdev driving SFP LEDs via /sys/class/leds/ (via gpio-pca953x). I generate traffic with iperf3. Network traffic is on 10Gbps EMAC. LX kernel is 4.14.0. What I saw from debugging this problem is that I2C bus get blocked when wait_event_timeout() completes because of timeout. The timeout handling in this driver is probably not robust enough (bus should not remain blocked), but at this moment this are just my speculations (don't know enough details).
Check with salea logic analyzer what happens to the i2c bus.
Looking the driver code and data on oscilloscope, I saw that SCL in single I2C data transfer sequence can be interrupted for very long delays, e.g up to hundredths of usec (SCL is 100kHz). I started to suspect that PM autosuspend delay could play some role here. There are only two delays in driver code, first in wait_event_timeout and second in set autosuspend delay. Case is a bit strange because in very busy I2C traffic, PM autosuspend should not be triggered at all. Additionally, if I lower PM timeout, e.g. from 1000 (default) to 100, I hit the problem sooner (waits for problem hit are in order of n*10minutes).
It looks to me that PM autosupend is playing some role here.
Power management options in my .config: # CONFIG_SUSPEND is not set # CONFIG_PM is not set CONFIG_ARCH_SUSPEND_POSSIBLE=y
I intentionally did not put all detail descriptions of embedded system and test setup here (long list), because the main reason of this post is:
The workaround that works for me/customer (at the moment) is to disable PM autosuspend in the driver code, either by incerementing PM delay from 1000 to 10000 or by disabling autosuspend (comment out call to pm_runtime_put_autosuspend() in xiic_xfer()).
But, I would like to expose/discuss this issue (maintainer of the code, or others). The reason/source of the problem can be much more complex and in some other place.
So my question is who should I contact, is this the M: in the MAINTAINERS list, the MODULE_AUTHOR, ...?
You can certainly add the author in loop but I am afraid you won't get any help as this would be specific to your board. So, best is to check soc vendor who has written your i2c bus driver or it could be a issue with your i2c client in that case show them your salea logic analyzer logs to see if they can figure out anything wrong.
How to proceed.
WBR Primoz
_______________________________________________ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org https://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies
On 3. 12. 19 02:30, anish singh wrote:
On Fri, Nov 29, 2019 at 12:53 PM Primoz Beltram <primoz.beltram@gmail.com> wrote:
I am analysing a problem with I2C bus driver where the problem shows up as I2C bus completely blocked. The LX driver in question is /drivers/i2c/busses/i2c-xiic.c. Problem is difficult to reproduce, it happens very rarely. So far I saw that the main precondition is to have very heavy I2C traffic on bus. In my case this is achieved/reproduced via netdev driving SFP LEDs via /sys/class/leds/ (via gpio-pca953x). I generate traffic with iperf3. Network traffic is on 10Gbps EMAC. LX kernel is 4.14.0. What I saw from debugging this problem is that I2C bus get blocked when wait_event_timeout() completes because of timeout. The timeout handling in this driver is probably not robust enough (bus should not remain blocked), but at this moment this are just my speculations (don't know enough details). Check with salea logic analyzer what happens to the i2c bus.
Looking the driver code and data on oscilloscope, I saw that SCL in single I2C data transfer sequence can be interrupted for very long delays, e.g up to hundredths of usec (SCL is 100kHz). I started to suspect that PM autosuspend delay could play some role here. There are only two delays in driver code, first in wait_event_timeout and second in set autosuspend delay. Case is a bit strange because in very busy I2C traffic, PM autosuspend should not be triggered at all. Additionally, if I lower PM timeout, e.g. from 1000 (default) to 100, I hit the problem sooner (waits for problem hit are in order of n*10minutes).
It looks to me that PM autosupend is playing some role here.
Power management options in my .config: # CONFIG_SUSPEND is not set # CONFIG_PM is not set CONFIG_ARCH_SUSPEND_POSSIBLE=y
I intentionally did not put all detail descriptions of embedded system and test setup here (long list), because the main reason of this post is:
The workaround that works for me/customer (at the moment) is to disable PM autosuspend in the driver code, either by incerementing PM delay from 1000 to 10000 or by disabling autosuspend (comment out call to pm_runtime_put_autosuspend() in xiic_xfer()).
But, I would like to expose/discuss this issue (maintainer of the code, or others). The reason/source of the problem can be much more complex and in some other place.
So my question is who should I contact, is this the M: in the MAINTAINERS list, the MODULE_AUTHOR, ...? You can certainly add the author in loop but I am afraid you won't get any help as this would be specific to your board. So, best is to check soc vendor who has written your i2c bus driver or it could be a issue with your i2c client in that case show them your salea logic analyzer logs to see if they can figure out anything wrong.
Thanks for reply and suggestions. My first suspicion was signal integrity on PCB, but if I add some debug prints in i2c-xiic driver (e.g. build with DEBUG define), the problem is no longer reproducible (not a single timeout completion in wait_event_timeout()). Signal integrity problem does not look credible to me. For my system I fixed the problem in i2c-xiic driver (in handIing timeout, not leave bus blocked). Found also a contact and fill report for SoC vendor. WBR Primoz
How to proceed.
WBR Primoz
_______________________________________________ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org https://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies
participants (3)
-
anish singh -
Primoz Beltram -
Primoz Beltram