On Wed, Feb 23, 2022 at 02:48:38PM +0100, Hans de Goede wrote:
> Lock the adapter while setting the suspended flag, to ensure that other
> locked code always sees the change immediately, rather then possibly using
> a stale value.
> 
> This involves splitting the suspend/resume callbacks into separate runtime
> and normal suspend/resume calls. This is necessary because i2c_dw_xfer()
> will get called by the i2c-core with the adapter locked and it in turn
> calls the runtime-resume callback through pm_runtime_get_sync().
> 
> So the runtime versions of the suspend/resume callbacks cannot take
> the adapter-lock. Note this patch simply makes the runtime suspend/resume
> callbacks not deal with the suspended flag at all. During runtime the
> pm_runtime_get_sync() from i2c_dw_xfer() will always ensure that the
> adapter is resumed when necessary.
> 
> The suspended flag check is only necessary to check proper suspend/resume
> ordering during normal suspend/resume which makes the pm_runtime_get_sync()
> call a no-op.
> 
> Signed-off-by: Hans de Goede <hdegoede@redhat.com>

Applied to for-next, thanks! We can still backport if we find out we
really need it...