Hi Philipp, On 30 June 2022 15:45 Philipp Zabel wrote: > On Do, 2022-06-30 at 13:43 +0000, Phil Edworthy wrote: > > On 29 June 2022 18:18 Geert Uytterhoeven wrote: > > > On Wed, Jun 29, 2022 at 6:21 PM Philipp Zabel wrote: > > > > On Di, 2022-06-28 at 20:45 +0100, Phil Edworthy wrote: > > > > > The Renesas RZ/V2M SoC (r9a09g011) has a new i2c controller. This > > > series > > > > > add the driver. One annoying problem is that the SoC uses a single > > > reset > > > > > line for two i2c controllers, and unfortunately one of the > controllers > > > > > is managed by some firmware, not by Linux. Therefore, the driver > just > > > > > deasserts the reset. > > > > > > > > This sounds scary. If the driver is never loaded, and the reset is > > > > never deasserted, what happens to the firmware trying to access the > > > > other i2c controller? Does it hang? Or write to the reset controller > > > > registers to deassert the reset? If so, is there any protection > against > > > > concurrent access from firmware and reset controller driver? > > Where a common reset is used by Linux and some firmware, I think we have > to > > ensure/assume that both only ever de-assert it. > > We also have to make sure that no read-modify-write cycles are required > to deassert the resets if we can't lock between firmware and kernel. > Otherwise concurrent access could cause a deassert to be reverted. Agreed > > In this particular SoC, the register used to assert/de-assert the reset > > has write enable bits in the upper half of the reg. There shouldn't be > any > > issues with both trying to de-assert the reset at the same time. > > Which reset driver is handling the reset for this i2c module? drivers/clk/renesas/rzg2l-cpg.c See rzg2l_cpg_assert() and rzg2l_cpg_deassert() Note this driver handles a few different SoCs, the SoC using this i2c driver is specified in drivers/clk/renesas/r9a09g011-cpg.c > > > In response to v1, I wrote > > > > > > > That is actually an integration issue, not an i2c controller issue. > > > > > > > > Perhaps we need a RESET_IS_CRITICAL flag, cfr. CLK_IS_CRITICAL, > > > > to be set by the reset provider? > > > > From what I understand, there are two main use cases for resets: > > 1. Often reset lines may be asserted at power on and so a driver needs > to > > de-assert them so that the module can be used. > > There are resets that are not initially asserted (among them the self- > deasserting resets) that are required to be asserted some time during > boot, to put some hardware into a well defined state. > I don't think those should be shared, but they sometimes are. > > > 2. A driver may need to reset the module for some reason. I have only > > seen this with watchdog timers with no way out. > > Grep for device_reset() or reset_control_reset() for some examples. > Also there are quite a few assert/udelay/deassert calls in drivers. Ok, though I'm not convinced that the driver specifying the reset period is the right way to support lots of different platforms. > Also many drivers assert the reset again during remove(). Whether that > is always necessary or useful, I can't say. Quite. It's difficult to know if the module requires a reset or that's just what the driver developer used. > It's sometimes nice during development, to be able to reload a kernel > module or rebind a driver to reset some locked up hardware. Ok, that's a good use case! > > So if a driver does not need to reset the module, shouldn't the driver > > only ever be de-asserting the reset line? > > I'm not sure the driver can always know this if it is used on different > platforms. > > > If so, it also doesn’t matter whether the reset is shared with other > > modules or not. > > If a driver needs to reset the module, then the reset cannot be shared > > with other modules used by firmware or Linux, or we cannot use any > > other modules that share the reset line. > > It can be shared for the special case of multiple modules requiring a > shared reset line to be asserted once, at some time before the modules > are used. The reset controller API supports this for the > reset_control_reset() call. In order for drivers to work on lots of platforms, should all drivers use devm_reset_control_get_shared() instead of devm_reset_control_get(), unless there is a need to reset the hardware at a specific time after boot (e.g. watchdog with no way out)? So where do we go with this for this i2c driver? Thanks Phil