Hi Grygorii, Doug and Heiko, Thanks for your replies. I will do 2 steps: 1. Add "suspended" flag in suspend_noirq()/resume_noirq() callback to prevent new i2c started, and use i2c_lock_adapter() to wait for current i2c transfer finished. 2. IRQF_NO_SUSPEND added could make i2c work well during the time between suspend_device_irqs() and i2c_suspend_noirq() callback. In the other side, it is the the time between resume_device_irqs() and i2c_resume_noirq() callback. If any i2c client try to access I2C after suspend_noirq() or before resume_noirq() callback, print the warning, and they should fix it, not to start i2c access and the moment. ? 2016/12/8 0:27, Grygorii Strashko ??: > > > On 12/06/2016 09:37 PM, David.Wu wrote: >> Hi Doug, >> >> ? 2016/12/7 0:31, Doug Anderson ??: >>> Hi, >>> >>> On Tue, Dec 6, 2016 at 12:12 AM, David.Wu <david.wu at rock-chips.com> >>> wrote: >>>> Hi Heiko, >>>> >>>> ? 2016/12/5 18:54, Heiko Stuebner ??: >>>>> >>>>> Hi David, >>>>> >>>>> Am Montag, 5. Dezember 2016, 16:02:59 CET schrieb David Wu: >>>>>> >>>>>> During suspend there may still be some i2c access happening. >>>>>> And if we don't keep i2c irq ON, there may be i2c access timeout if >>>>>> i2c is in irq mode of operation. >>>>> >>>>> >>>>> can you describe the issue you're trying to fix a bit more please? >>>> >>>> >>>> Sometimes we could see the i2c timeout errors during suspend/resume, >>>> which >>>> makes the duration of suspend/resume too longer. >>>> >>>> [ 484.171541] CPU4: Booted secondary processor [410fd082] >>>> [ 485.172777] rk3x-i2c ff3c0000.i2c: timeout, ipd: 0x10, state: 1 >>>> [ 486.172760] rk3x-i2c ff3c0000.i2c: timeout, ipd: 0x10, state: 1 >>>> [ 487.172759] rk3x-i2c ff3c0000.i2c: timeout, ipd: 0x10, state: 1 >>>> [ 487.172840] cpu cpu4: _set_opp_voltage: failed to set voltage (800000 >>>> 800000 800000 mV): -110 >>>> [ 487.172874] cpu cpu4: failed to set volt 800000 >>>> >>>>> >>>>> I.e. I'd think the i2c-core does suspend i2c-client devices first, >>>>> so that >>>>> these should be able to finish up their ongoing transfers and not start >>>>> any >>>>> new ones instead? >>>>> >>>>> Your irq can still happen slightly after the system started going to >>>>> actually >>>>> sleep, so to me it looks like you just widened the window where irqs >>>>> can >>>>> be >>>>> handled. Especially as your irq could also just simply stem from the >>>>> start >>>>> state, so you cannot even be sure if your transaction actually is >>>>> finished. >>>> >>>> >>>> Okay, you are right. I want to give it a double insurance at first, >>>> but it >>>> may hide the unhappend issue. >>>> >>>>> >>>>> So to me it looks like the i2c-connected device driver should be fixed >>>>> instead? >>>> >>>> >>>> I tell them to fix it in rk808 driver. >>> >>> To me it seems like perhaps cpufreq should not be changing frequencies >>> until it is resumed properly. Presumably if all the ordering is done >>> right then cpufreq should be resumed _after_ the i2c regulator so you >>> should be OK. ...or am I somehow confused about that? >> >> yes?the cpufreq and regulator should start i2c job after they resume >> properly. >> >>> >>> Also note that previous i2c busses I worked with simply returned -EIO >>> in the case where they were called when suspended. See >>> "i2c-exynos5.c" and "i2c-s3c2410.c". >> >> In "i2c-exynos5.c", it seems that using the "i2c->suspended" to protect >> i2c transfer works most of the time. Of course it could prevent the next >> new i2c transfer to start. But in one case, if the current i2c job was >> not finished until the i2c irq was disabled by system suspend, the i2c >> timeout error would also happen, as the current i2c job may have a large >> data to transfer and it lasts from a long time. > > And this means you have bug in some of I2C client drivers which do not stop > their activities during suspend properly (most usual case - driver uses work > and this work still active during suspend and can run on one CPU while suspend > runs on another). > > At the moment .suspend_noirq() callback is called there should be no active > I2C transactions in general. > >> >> So is it necessary to add a mutex lock to wait the current job to be >> finished before the "i2c->suspended" is changed in i2c_suspend_noirq()? >> > > You need to catch and fix all driver who will try to access I2C after your > I2C bus driver passes suspend_noirq stage. Smth, like [1], uses i2c_lock_adapter(). > > > [1] https://git.ti.com/android-sdk/kernel-omap/commit/125ef8f7016e7b205886f93862288a45a312b1d8 >
