It saddened me that kmutex_t was 20 bytes on ARM, so I fixed it to be  
the optimal 8-bytes.  The restartable atomic sequence I added to  
support this can also be used for rwlock_t, eliminating the need to  
block interrupts to manipulate it; I'll make that change shortly.

Ben -- you need to update acorn26 to do the RAS check for _lock_cas()  
-- see the LOCK_CAS_CHECK macro in include/arm32/frame.h.

Uwe -- you might want to use this as a model for what to do for sh3.

Everyone else -- you might want to enable the LOCKDEBUG option in your  
kernels for a while.  Matt Thomas did so while testing these changes  
for me and found a latent bug in the Shark interrupt code.

Enjoy!

Begin forwarded message:

> From: Jason R Thorpe <thorpej@netbsd.org>
> Date: March 9, 2007 11:21:59 AM PST
> To: source-changes@NetBSD.org
> Subject: CVS commit: src/sys/arch
> Reply-To: thorpej@netbsd.org
>
>
> Module Name:    src
> Committed By:    thorpej
> Date:        Fri Mar  9 19:21:59 UTC 2007
>
> Modified Files:
>    src/sys/arch/arm/arm: arm_machdep.c
>    src/sys/arch/arm/arm32: genassym.cf irq_dispatch.S
>    src/sys/arch/arm/conf: files.arm
>    src/sys/arch/arm/include: mutex.h
>    src/sys/arch/arm/include/arm32: frame.h
>    src/sys/arch/arm/iomd: iomd_irq.S
>    src/sys/arch/arm/ofw: ofw_irq.S
>    src/sys/arch/arm/sa11x0: sa11x0_irq.S
>    src/sys/arch/arm/xscale: i80200_irq.S
>    src/sys/arch/shark/isa: isa_irq.S
>
> Log Message:
> Rewrite the ARM mutex implementation to be of the simple-mutex  
> variety.
> Because pre-v6 ARM lacks support for an atomic compare-and-swap, we
> implement _lock_cas() as a restartable atomic squence that is checked
> in the IRQ handler right before AST processing.  (This is safe  
> because,
> for all practical purposes, there are no SMP pre-v6 ARM systems.)
>
> This can serve as a model for other non-MP platforms that lack the
> necessary atomic operations for mutexes (SuperH, for example).
>
> Upshots of this change:
> - kmutex_t is now down to 8 bytes on ARM; about as good as we can get.
> - ARM2 systems don't have to trap and emulate SWP or SWPB for mutexes.
>
> The acorn26 port is not updated by this commit to do the  
> LOCK_CAS_CHECK.
> That is left as an exercise for the port maintainer.
>
> Reviewed and tested by Matt Thomas.
>
>
> To generate a diff of this commit:
> cvs rdiff -r1.13 -r1.14 src/sys/arch/arm/arm/arm_machdep.c
> cvs rdiff -r1.31 -r1.32 src/sys/arch/arm/arm32/genassym.cf
> cvs rdiff -r1.7 -r1.8 src/sys/arch/arm/arm32/irq_dispatch.S
> cvs rdiff -r1.82 -r1.83 src/sys/arch/arm/conf/files.arm
> cvs rdiff -r1.6 -r1.7 src/sys/arch/arm/include/mutex.h
> cvs rdiff -r1.14 -r1.15 src/sys/arch/arm/include/arm32/frame.h
> cvs rdiff -r1.5 -r1.6 src/sys/arch/arm/iomd/iomd_irq.S
> cvs rdiff -r1.6 -r1.7 src/sys/arch/arm/ofw/ofw_irq.S
> cvs rdiff -r1.9 -r1.10 src/sys/arch/arm/sa11x0/sa11x0_irq.S
> cvs rdiff -r1.13 -r1.14 src/sys/arch/arm/xscale/i80200_irq.S
> cvs rdiff -r1.6 -r1.7 src/sys/arch/shark/isa/isa_irq.S
>
> Please note that diffs are not public domain; they are subject to the
> copyright notices on the relevant files.

-- thorpej