On Fri, Aug 29, 2025 at 11:28:55AM +0000, Dmitry Torokhov wrote: > On Thu, Aug 28, 2025 at 08:35:56AM +0000, Tzung-Bi Shih wrote: > > A race is observed when cros_ec_lpc and cros-ec-keyb are all built as > > modules. cros_ec_lpc is cros-ec-keyb's parent. However, they can be > > probed at the same time. > > > > Example: > > > > + -----------------------------------------------------------------+ > > | Some init process (e.g. udevd) | deferred_probe_work_func worker | > > + -----------------------------------------------------------------+ > > | Probe cros-ec-keyb. | | > > | - Decide to defer[1]. | | > > | | A device bound to a driver[2]. | > > | Probe cros_ec_lpc. | | > > | - Init the struct[3]. | | > > | | Retry cros-ec-keyb from the | > > | | deferred list[4]. | > > | | - Won't defer again as [3]. | > > | | - Access uninitialized data in | > > | | the struct. | > > | - Register the device. | | > > + -----------------------------------------------------------------+ > > > > [1] https://elixir.bootlin.com/linux/v6.16/source/drivers/input/keyboard/cros_ec_keyb.c#L707 > > [2] https://elixir.bootlin.com/linux/v6.16/source/drivers/base/dd.c#L405 > > [3] https://elixir.bootlin.com/linux/v6.16/source/drivers/platform/chrome/cros_ec_lpc.c#L644 > > [4] https://elixir.bootlin.com/linux/v6.16/source/drivers/base/dd.c#L418 > > > > Note that the device link[5] can't help as in the observed environment, > > the devices are already added via device_add()[6]. > > > > [5] https://www.kernel.org/doc/html/latest/driver-api/device_link.html#usage > > [6] https://elixir.bootlin.com/linux/v6.16/source/drivers/acpi/acpi_platform.c#L177 > > > > The series fixes the issue by ensuring the struct is ready for accessing > > before continuing to probe cros-ec-keyb. > > Why is the keyboard platform device instantiated before the transport > (cros_ec_lpc) is done initializing? I think this is the root of the > issue... I may misunderstand but it seems to me: - The ACPI bus enumerated and instantiated the platform devices[6] first. - The keyboard platform device was probed when `cros_ec_keyb_driver` registered. It deferred as its parent's drvdata was NULL[1]. - The transport platform device was probed when `cros_ec_lpc_driver` registered. It set the drvdata[3]. - The keyboard platform device was probed again from retrying the deferred list, by another thread `deferred_probe_work_func`. The parent's drvdata wasn't NULL and cros_ec_register() for the transport device weren't finished. The race happened.
