On Mon, Jan 19, 2026 at 12:03 PM Johan Hovold <[email protected]> wrote: > > On Tue, Dec 23, 2025 at 04:11:08PM +0100, Bartosz Golaszewski wrote: > > On Tue, Dec 23, 2025 at 3:24 PM Johan Hovold <[email protected]> wrote: > > > > > > On Tue, Dec 23, 2025 at 11:02:22AM +0100, Bartosz Golaszewski wrote: > > > > It's been another year of discussing the object life-time problems at > > > > conferences. I2C is one of the offenders and its problems are more > > > > complex than those of some other subsystems. It seems the revocable[1] > > > > API may make its way into the kernel this year but even with it in > > > > place, I2C won't be able to use it as there's currently nothing to > > > > *revoke*. The struct device is embedded within the i2c_adapter struct > > > > whose lifetime is tied to the provider device being bound to its driver. > > > > > > > > Fixing this won't be fast and easy but nothing's going to happen if we > > > > don't start chipping away at it. The ultimate goal in order to be able > > > > to use an SRCU-based solution (revocable or otherwise) is to convert the > > > > embedded struct device in struct i2c_adapter into an __rcu pointer that > > > > can be *revoked*. To that end we need to hide all dereferences of > > > > adap->dev in drivers. > > > > > > No, this is not the way to do it. You start with designing and showing > > > what the end result will look like *before* you start rewriting world > > > like you are doing here. > > > > The paragraph you're commenting under explains exactly what I propose > > to do: move struct device out of struct i2c_adapter and protect the > > pointer storing its address with SRCU. This is a well-known design > > that's being generalized to a common "revocable" API which will > > possibly be available upstream by the time we're ready to use it. > > Revocable, as presented in plumbers, is not going upstream. >
Oh really? :) https://lore.kernel.org/all/2026011607-canister-catalyst-9fdd@gregkh/ > > You know I can't possibly *show* the end result in a single series > > because - as the paragraph before explains - we need to first hide all > > direct dereferences of struct device in struct i2c_adapter behind > > dedicated interfaces so that we when do the conversion, it'll affect > > only a limited number of places. It can't realistically be done at > > once. > > You can post an RFC converting one driver with a proper description of > the problem you're trying to solve. > It's not a one-driver problem. It's a subsystem-wide problem requiring a subsystem-wide solution. Wolfram explained it really well in his summary, I'm not going to repeat it here. I also don't agree that i2c-specific helpers make code harder to read. Is device_set_node() harder to read than dev->fwnode = fwnode; dev->of_node = to_of_node(fwnode); ? Even if you answer yes - it at least helps hide the implementation details of the OF layer where fwnode-level is preferred. We do it all the time in the kernel. This kind of helpers allows easier transitions when some implementation detail needs to change - as is the case here. Bartosz
