Re: 2.6.18 Headers - Long

[David Miller <davem@davemloft.net>]

From: Kyle Moffett <mrmacman_g4@mac.com>
Date: Sun, 16 Jul 2006 08:34:53 -0400

> Both of which may be easily cut and pasted into your GPL programs  
> with little or no effort (Hint: I do this all the time).  Those are  
> so stable you don't even have to maintain it!  IMHO, what you really  
> want, though, is for GCC to export a library of ASM intrinsics (like  
> memory barriers, atomic ops, etc), that are available on your current  
> architecture.  If there is no __gcc_atomic_inc then it wouldn't  
> #define it and you can just go back to pthread_mutex_lock/unlock for  
> protecting an atomic variable.  Such a library layer certainly  
> doesn't belong in the kernel, although if GCC got such a library  
> right the kernel might start to use it (although only the most recent  
> GCC would support it so it wouldn't be very useful).

I agree with your assertions that the stuff in asm/atomic.h should be
steered away from, since they generally are not expected to work in
userspace.  In fact, if you try to use the 32-bit sparc or parisc
ones, it simply won't link because the implementations are external
and use a hash table spinlock scheme which is in the kernel image.

However, the pthread based locking is bad if you want to recover
from arbitrary signals correctly.  I ran into this problem while
trying to get Linux/Sparc Mono working well.

When people want atomic cmpxchg or atomic increment/decrement, they
want a true atomic.  This means either finish the whole thing, or
leave no trace of the atomic having started in the first place.
Spinlock based atomic implementations, which is what a lot of
platforms lacking true atomic use, absolutely cannot reasonably
provide this semantic.