On 10/31/2013 11:16 PM, Domenico Andreoli wrote: > On Thu, Oct 31, 2013 at 03:49:18PM -0600, Stephen Warren wrote: >> On 10/31/2013 12:27 AM, Domenico Andreoli wrote: >>> Hi, >>> >>> I've been looking for a solution to my bcm4760 watchdog based restart >>> hook when I noticed that the kernel reboot/shutdown mechanism is having >>> a few unaddressed issues. >>> >>> Those I identified are: >>> >>> 1) context pointer often needed by the reset hook >>> (currently local static data is used for this pourpose) >>> 2) unclear ownership/policy in case of multiple reset hooks >>> (currently almost nobody seems to care much) >> >> I'm not sure how this patchset solves (2); even with the new API, it's >> still the case that whichever code calls set_machine_reset() last wins, >> just like before where whichever code wrote to pm_power_off won. I'm not >> sure what this series attempts to solve. > > That's right, the last wins. But the previous has a chance to know. > > I only supposed there is somebody in charge of selecting the best handler > for the machine. Don't know how fancy this decision is but at least for > the vexpress there is also a sysfs way to configure different reset methods > from user-space. For PMICs that provide power off, we've been adding a property to DT to indicate whether the PMIC is *the* system power off controller or not. If the property is present, the PMIC registers itself in the poweroff hook. If not, it doesn't. So, there really isn't an algorithm for selecting the power off mechanism, but rather we designate one mechanism ahead of time, and that's the only one that's relevant. We could probably do the same for reset mechanisms. I guess the vexpress situation is actually the same; there's a single concept of a custom vexpress reset, it's just that sysfs is used to select exactly what that does? > So cleaning up things after the handler is replaced seemed a sensible > thing to do. Can't we avoid replacing handlers, but only registering a single handler? > Another "problem" this patch would solve is the registration of the > reset handler in a architecture independent way. Now an otherwise platform > generic gpio HW reset driver would need to do different things on different > architectures. OK, if there are architecture differences in how the hooks are registered, that seems like a good thing to fix.
