Hi,
I am currently creating a "LOCK" prefix free and memory barrier free version
of atomic.h to fulfill my tracer (LTTng) needs, which is to atomically update
per-cpu data and have a minimal performance loss.
I just came across the MIPS atomic.h and system.h implementations in 2.6.18
which brings a question :
Why are the primitives in include/asm-mips/atomic.h using the "sync"
instruction even in the UP case ? system.h cmpxchg only uses the sync in the
SMP case.
Mathieu
OpenPGP public key: http://krystal.dyndns.org:8080/key/compudj.gpg
Key fingerprint: 8CD5 52C3 8E3C 4140 715F BA06 3F25 A8FE 3BAE 9A68
On Fri, Nov 10, 2006 at 01:40:49PM -0500, Mathieu Desnoyers wrote:
> I am currently creating a "LOCK" prefix free and memory barrier free version
> of atomic.h to fulfill my tracer (LTTng) needs, which is to atomically update
> per-cpu data and have a minimal performance loss.
>
> I just came across the MIPS atomic.h and system.h implementations in 2.6.18
> which brings a question :
>
> Why are the primitives in include/asm-mips/atomic.h using the "sync"
> instruction even in the UP case ? system.h cmpxchg only uses the sync in the
> SMP case.
Why are the standard atomic operations insufficient for your needs?
There is an enormous amout of subtilities in those atomic ops for some
architectures you probably do yourself a big favor by avoiding new
variants.
Ralf
* Ralf Baechle ([email protected]) wrote:
> On Fri, Nov 10, 2006 at 01:40:49PM -0500, Mathieu Desnoyers wrote:
>
> > I am currently creating a "LOCK" prefix free and memory barrier free version
> > of atomic.h to fulfill my tracer (LTTng) needs, which is to atomically update
> > per-cpu data and have a minimal performance loss.
> >
> > I just came across the MIPS atomic.h and system.h implementations in 2.6.18
> > which brings a question :
> >
> > Why are the primitives in include/asm-mips/atomic.h using the "sync"
> > instruction even in the UP case ? system.h cmpxchg only uses the sync in the
> > SMP case.
>
> Why are the standard atomic operations insufficient for your needs?
>
> There is an enormous amout of subtilities in those atomic ops for some
> architectures you probably do yourself a big favor by avoiding new
> variants.
>
Performance cost.
I add a memory barrier where needed when the data needs to appear to be written
sequentially from the other CPUs perspective.
Mathieu
OpenPGP public key: http://krystal.dyndns.org:8080/key/compudj.gpg
Key fingerprint: 8CD5 52C3 8E3C 4140 715F BA06 3F25 A8FE 3BAE 9A68