Re: sunxigpio vs. LOCKDEBUG (PR 54664)

[Jared McNeill <jmcneill%invisible.ca@localhost>]

Yeah don't commit that, that's pretty gross as-is:
 1) Silent failure in the kmem_alloc failure path.
 2) No comment describing _why_ it's doing this bad thing.

How about adding a struct fdtbus_regulator to struct 
fdtbus_regulator_controller? Then container_of() magic to derive the 
controller from the regulator handle in subsequent fdtbus_regulator_* 
calls, and we can avoid the allocation completely.

Take care,
Jared

On Thu, 5 Aug 2021, Tobias Nygren wrote:

> Hi,
>
> sunxi boards can't boot LOCKDEBUG kernels. As I described
> in the PR:
>
> > The problem is that sunxi_pinctrl_set_config() calls
> > fdtbus_regulator_acquire() with spin mutex
> > sunxi_gpio_softc.sc_lock held, but fdtbus_regulator_acquire()
> > calls kmem_alloc with flag KM_SLEEP. That's not permitted.
>
> Refactoring sunxi_gpio to not hold it's sc lock seems
> not straightforward. Any objections to this patch?
> I'm not a fan of using KM_NOSLEEP, but don't have any
> better ideas ...
>
> --- sys/dev/fdt/fdt_regulator.c	27 May 2019 23:18:33 -0000	1.8
> +++ sys/dev/fdt/fdt_regulator.c	5 Aug 2021 18:29:22 -0000
> @@ -99,13 +99,16 @@ fdtbus_regulator_acquire(int phandle, co
> 		return NULL;
> 	}
>
> +	reg = kmem_alloc(sizeof(*reg), KM_NOSLEEP);
> +	if (!reg) {
> +		return NULL;
> +	}
> 	error = rc->rc_funcs->acquire(rc->rc_dev);
> 	if (error) {
> 		aprint_error_dev(rc->rc_dev, "failed to acquire regulator: %d\n", error);
> +		kmem_free(reg, sizeof(*reg));
> 		return NULL;
> 	}
> -
> -	reg = kmem_alloc(sizeof(*reg), KM_SLEEP);
> 	reg->reg_rc = rc;
>
> 	return reg;