On Wed, Jul 23, 2025 at 11:01 AM Thierry Reding
<thierry.reding@gmail.com> wrote:
>
> On Tue, Jul 22, 2025 at 04:08:09PM +0200, Greg Kroah-Hartman wrote:
> > On Tue, Jul 22, 2025 at 03:56:40PM +0200, Thierry Reding wrote:
> > > On Sat, Jul 19, 2025 at 08:52:41AM +0200, Greg Kroah-Hartman wrote:
> > > > On Fri, Jul 18, 2025 at 03:49:37PM +0200, Thierry Reding wrote:
> > > > > On Thu, Jul 17, 2025 at 02:11:41PM +0200, Greg Kroah-Hartman wrote:
> > > > > > On Thu, Jul 17, 2025 at 12:32:34PM +0200, Thierry Reding wrote:
> > > [...]
> > > > > struct syscore;
> > > > >
> > > > > struct syscore_ops {
> > > > > int (*suspend)(struct syscore *syscore);
> > > > > void (*resume)(struct syscore *syscore);
> > > > > void (*shutdown)(struct syscore *syscore);
> > > > > };
> > > > >
> > > > > struct syscore {
> > > > > const struct syscore_ops *ops;
> > > > > struct list_head node;
> > > > > };
> > > > >
> > > > > Is that what you had in mind?
> > > >
> > > > I missed the list_head, so yes, this would be better, but don't pass
> > > > back the syscore structure, how about just a void * instead, making the
> > > > whole container_of() stuff go away?
> > >
> > > Yeah, that's a possibility. I personally don't like passing the void *
> > > around because it's easier to make mistakes that way. I also find it
> > > unintuitive because it doesn't immediately show you what the functions
> > > expect.
> > >
> > > My understanding is that the container_of() should get optimized away
> > > most of the time, so there aren't any obvious downsides that I can see.
> >
> > container_of() is just pointer math, but a cast is even faster :)
> >
> > > But I don't feel very strongly, so if you have a strong preference for
> > > void pointers, I can do that.
> >
> > That's what you really want to have here, it's a syscore data type
> > thing, that the callback wants to reference. Just like a irqrequest_t
> > function passes back a void * that the handler "knows" how to deal with
> > properly.
>
> IRQ handlers are different, though, because you pass the void * data
> when you register the interrupt. That void * then gets stored and passed
> to the handler when the interrupt is processed.
>
> We'd have to change it to something like this:
>
> struct syscore_ops {
> /* parameters now changed to driver-specific data */
> int (*suspend)(void *data);
> void (*resume)(void *data);
> void (*shutdown)(void *data);
> };
>
> struct syscore {
> const struct syscore_ops *ops;
> struct list_head node;
> /* NEW driver-specific data */
> void *data;
> };
I like this more than the original, but I would do
struct syscore_ops_ops {
int (*suspend)(void *data);
void (*resume)(void *data);
void (*shutdown)(void *data);
};
struct syscore_ops {
struct list_head node;
const struct syscore_ops_ops *ops;
void *data;
};
and change register_syscore_ops() to take three arguments, the struct
syscore_ops pointer, the (constified) struct syscore_ops_ops one, and
the (void *) data one.
Note that it is not necessary to change the signature of
unregister_syscore_ops() in this case.
> It ends up increasing the syscore structure's size, about 33%, though
> given that there aren't a lot of these that's probably negligible.
That's not a problem IMV.
> What I think is a bit more unnatural about it in this case is that we
> embed the struct syscore into some driver-private data anyway so that
> it becomes per instance, and then we have a circular reference:
>
> foo->syscore.ops = &foo_syscore_ops;
> foo->syscore.data = foo;
That depends because "data" need not be "foo" in all cases, but also
see above. If the initialization of struct syscore_ops is all done by
register_syscore_ops(), it doesn't look circular any more.
> Which looks kind of weird. Alternatively I suppose we could completely
> rework it and make register_syscore_ops() allocate struct syscore, and
> hide the internals from drivers completely:
>
> err = register_syscore(&foo_syscore_ops, foo);
>
> With that it may be problematic that register_syscore() can now fail.
Yes, that might be a problem.