On Thu, 2005-03-10 at 09:43 +1100, Nigel Cunningham wrote:
> > > > > <heresy> I wonder if using sysfs is even the best method for doing
> > > > > run-time PM. It will force your imaginary nice userspace interface to
> > > > > include code to scan the whole directory tree looking for files of each
> > > > > kind, perhaps sorting and collating and so on. Maybe a DBus interface
> > > > > would be better? </heresy>
> > > >
> > > > DBus is too hard to use from shell, and I do not think whole directory
> > > > scan is so hard to do.
> > >
> > > Perhaps provide both? Of course I don't know what the implications
> >
> > No...
> >
> > I do not believe sysfs is ever going to be efficiency problem. If your
> > machine is so big that your /sys is big, then your machine is fast
> > enough to handle it.
>
> It is possible, I agree. But how slow would it be?
>
> I guess the real show stopper would be the fact that you'd need to have
> your userspace daemon constantly scanning these sysfs files for events.
> That or you'd need to provide one sysfs entry that simplified things by
> duplicating all events through one file. If we go that far, it's not
> much further to a netlink socket.
>
Isn't this exactly what the kobject user events are for ? There is one
here
KOBJ_CHANGE = (__force kobject_action_t) 0x03, /* a sysfs attribute file has changed */
to be used when a sysfs attribute has changed. Userspace would listen
for this; it's broadcast on the netlink socket and thus pretty easy to
be picked up. Events can be emitted from the kernel using these
functions
int kobject_uevent(struct kobject *kobj,
enum kobject_action action,
struct attribute *attr);
int kobject_uevent_atomic(struct kobject *kobj,
enum kobject_action action,
struct attribute *attr);
defined in include/linux/kobject_uevent.h. You may also define your own
signals; see the header file; for instance we now have messages for
mount/umount so userspace don't have to use the dreadful D_NOTIFY on
e.g. /etc/mtab or /proc/mounts.
I also think there is some general confusion about using D-BUS in kernel
space in this thread - the whole reason for doing the uevents stuff,
IIRC, was that they are easy to be picked up by the D-BUS system message
bus and pushed out to the rest of user space. IIRC there were never any
talk about having user space using D-BUS to invoke methods in the kernel
- it was strictly one-way.
It seems to me that sysfs + uevents for change notification is the
appropriate interface for user space.
Cheers,
David
