On 14.05.2018 15:21, Laxman Dewangan wrote: > > > On Monday 14 May 2018 05:29 PM, Thierry Reding wrote: >> * PGP Signed by an unknown key >> >> On Mon, May 14, 2018 at 12:13:47AM +0300, Dmitry Osipenko wrote: >>> Nothing prevents I2C clients to access I2C while Tegra's driver is being >>> suspended, this results in -EBUSY error returned to the clients and that >>> may have unfortunate consequences. In particular this causes problems >>> for the TPS6586x MFD driver which emits hundreds of "failed to read >>> interrupt status" error messages on resume from suspend. This happens if >>> TPS6586X is used to wake system from suspend by the expired RTC alarm >>> timer because TPS6586X is an I2C device driver and its IRQ handler reads >>> the status register while Tegra's I2C driver is suspended, i.e. just after >>> kernel enabled IRQ's during of resume-from-suspend process. >>> >>> Note that the removed tegra_i2c_resume() invoked tegra_i2c_init() which >>> performs HW reset. That seems was also not entirely correct because moving >>> tegra_i2c_resume to an earlier stage of resume-from-suspend process causes >>> I2C transfer to fail in the case of TPS6586X. It is fine to remove the >>> HW-reinitialization for now because it should be only needed in a case of >>> using lowest power-mode during suspend, which upstream kernel doesn't >>> support. >>> >>> Signed-off-by: Dmitry Osipenko <digetx@xxxxxxxxx> >>> Cc: <stable@xxxxxxxxxxxxxxx> >>> --- >>> drivers/i2c/busses/i2c-tegra.c | 33 --------------------------------- >>> 1 file changed, 33 deletions(-) >> Shardar, Laxman, any thoughts on this? The is_suspended thing looks to >> me like a workaround of some sort that may not be needed if clients have >> proper suspend/resume implementations. Even without suspend/resume >> support in client drivers, the driver core should resume devices in the >> right order (I2C adapter before any of the clients), so I don't see any >> cases where the is_suspended logic would be useful. >> > > Our I2C driver is based on the interrupt. So we have converted the > suspend/resume to suspend_noirq and reseume_noirq so that we will not allow the > transfer when system interrupt disabled in downstream. > SET_NOIRQ_SYSTEM_SLEEP_PM_OPS(tegra_i2c_suspend, tegra_i2c_resume) > > In shutdown path, where interrupt disabled and still need i2c, we use the > bit-bang method via GPIO for i2c transfer. In the current upstream kernel suspend/resume can't be simply moved to the 'noirq' stage because resume invokes tegra_i2c_init() which uses runtime PM and that doesn't work with the IRQ's being disabled. But things do not work even with the tegra_i2c_init() changed to work with the disabled IRQ's, like I wrote above the I2C transfer fails (due to timeout) and a "fix" for that failure was to remove reset_control_assert/deassert from the tegra_i2c_init(). So I'd go for a complete suspend/resume removal for now as it is causes problem.
