[driver model] Rewrite system device API
System devices are special, and after two years of listening to Linus preach this, it finally sunk in enough to do something about. We don't need to regard them as real devices that reside on a peripheral bus and can be dynamically bound to drivers. If we discover, e.g. a CPU, we know by default that we have a driver for it, and we know damn well that we have a CPU. We still need to keep track of all the devices, and all the devices of a particular type. The kobject infrastructure allows us to do this, without the overhead of the regular model. A new subsystem is defined that registers as a child object of devices_subsys, giving us: /sys/devices/system/ struct sysdev_class { struct list_head drivers; /* Default operations for these types of devices */ int (*shutdown)(struct sys_device *); int (*suspend)(struct sys_device *, u32 state); int (*resume)(struct sys_device *); struct kset kset; }; Defines a type of system device. These are registered on startup, by e.g. drivers/base/cpu.c. The methods are default operations for devices of that type that may or may not be used. For things like the i8259 controller, these will be filled in, since it is registered by the same component that the device controls reside in. For things like CPUs, generic code will register the class, but other architecture-specific or otherwise configurable drivers may register auxillary drivers, that look like: struct sysdev_driver { struct list_head entry; int (*add)(struct sys_device *); int (*remove)(struct sys_device *); int (*shutdown)(struct sys_device *); int (*suspend)(struct sys_device *, u32 state); int (*resume)(struct sys_device *); }; Each auxillary driver gets called during each operation on a device of a particular class. Auxillary drivers may register with a NULL class parameter, in which case they will be added to a list of 'global drivers' that get called for each device of each class. Besides providing a decent of cleanup for system device drivers, this also allows: - Special handling of system devices during power transitions. We no longer have to worry about shutting down the PIC before we shut down any devices. We can shut down the system devices after we've shut down every other device. Ditto for suspend/resume cycles. Almost (if not) all PM actions for system devices happen with interrupts off, and require only one call, which makes that easier. But, we can also make sure we take care of these last during suspend and first during resume. - Easy expression of configurable device-specific interfaces. Namely cpufreq and mtrr. We don't have to worry about mispresentation in the driver model (like recent MTRR patches) or using a cumbersome interface ({device,class}_interface) that don't receive all the necessary calls. - Consolidation of userspace representation. No longer do we have /sys/devices/sys, /sys/bus/sys, and /sys/class/cpu, etc. We have only /sys/devices/system: # tree /sys/devices/system/ /sys/devices/system/ |-- cpu | `-- cpu0 |-- i8259 | `-- i82590 |-- lapic | `-- lapic0 |-- rtc | `-- rtc0 `-- timer `-- timer0 Each directory in 'system' is the class, and each directory under that is the instance of each device in that class.
Showing
Please register or sign in to comment