Re: [patch 0/9] mutex subsystem, -V4

[Steven Rostedt <rostedt@goodmis.org>]

On Thu, 2005-12-22 at 10:34 -0500, Nicolas Pitre wrote:

> > Actually just havign asm/mutex.h implement the faspath per-arch and get
> > rid of all the oddball atomic.h additions would be even better.  While
> > this means we need per-arch code it also means the code is a lot easier
> > understandable, and we don't add odd public APIs.
> 
> I'm with Christoph here.  Please preserve my 
> arch_mutex_fast_lock/arch_mutex_fast_unlock helpers.  I did it that way 
> because the most important thing they bring is flexibility where it is 
> needed i.e. in architecture specific implementations.  And done that way 
> the architecture specific part is well abstracted with the minimum 
> semantics allowing flexibility in the implementation.
> 
> I insist on that because, even if ARM currently relies on the atomic 
> swap behavior, on ARMv6 at least this can be improved even further, but 
> a special implementation which is neither a fully qualified atomic 
> decrement nor an atomic swap is needed.  That's why I insist that you 
> should keep my arch_mutex_fast_lock and friends (rename them if you 
> wish) and remove __ARCH_WANT_XCHG_BASED_ATOMICS entirely.
> 

Not sure how well this is accepted, but would it be acceptable to have
the mutex_lock and friends covered with the (weak) attribute?

ie.

void fastcall __sched __attribute__((weak)) mutex_lock(struct mutex *lock)

Then let the archs override them if they wish?

You would just need to make an extra slow path mutex visible to the archs.

-- Steve