Re: PATCH: (as177) Add class_device_unregister_wait() andplatform_device_unregister_wait() to the driver model core

[Rusty Russell]

On Sun, 25 Jan 2004 11:02:58 -0800 (PST)
Linus Torvalds <torvalds@xxxxxxxx> wrote:

> On Sun, 25 Jan 2004, Alan Stern wrote:
> > 
> > Is there some reason why modules don't work like this?
> 
> There's a few:
> 
>  - pain. pain. pain.
> 
>  - doing proper refcounting of modules is _really_ really hard. The reason 
>    is that proper refcounting is a "local" issue: you reference count a
>    single data structure. It's basically impossible to make a "global" 
>    reference count without jumping through hoops.
> 
>  - lack of testing. Unloading a module happens once in a blue moon, if 
>    even then.

And modules do work like you proposed, if you use "rmmod --wait".

Doing proper refcounting is actually fairly easy: every function pointer
has an associated reference count (or pointer to the module containing
the refcount).

But how much pain are you prepared to put up with to have a pseudo-pagable
kernel?

> (As an example of "too painful, too slow", think of something like a 
> packet filter module. You'd literally have to increment the count in every 
> part that gets a packet, and decrement the count at every point where it 
> lets the packet go.  And since it would have to be SMP-safe, it would have 
> to be a locked cycle, or we'd have to have per-CPU counters - at which 
> point you now also have to worry about things like preemption and 
> sleeping, which just means that it would be a _lot_ of very fragile code).

Actually, this is already handled.  The module reference counts are per-cpu
and don't contain any barriers.  We go to an *awful* lot of pain on remove
to synchronize, but as Linus says, it's not the normal case.

Since we hit the (atomic_t) ref to the devices on every packet, bumping
the refcount on the module is lost in the noise.

But Dave doesn't want to do it: it makes the code uglier and painful.

Cheers,
Rusty.
-- 
   there are those who do and those who hang on and you don't see too
   many doers quoting their contemporaries.  -- Larry McVoy
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/