Michael Buesch <m...@bu3sch.de> wrote on 31/10/2009 21:35:31:
>
> On Saturday 31 October 2009 21:14:07 Joakim Tjernlund wrote:
> > Michael Buesch <m...@bu3sch.de> wrote on 31/10/2009 17:42:54:
> > >
> > > On Saturday 31 October 2009 14:26:48 Joakim Tjernlund wrote:
> > > > >
> > > > > > On Friday 30 October 2009 16:08:55 Alessandro Rubini wrote:
> > > > > > > > asm("eieio; sync");
> > > > > > >
> > > > > > > Hmm...
> > > > > > > : : : "memory"
> > > > > > >
> > > > > > > And, doesn't ";" start a comment in assembly? (no, not on powerpc
> > > > > > it seems)
> > > > > >
> > > > > > Yes, I think the barrier is wrong.
> > > > > > Please try with
> > > > > >
> > > > > > #define mb() __asm__ __volatile__("eieio\n sync\n" : : :
> > > > > > "memory")
> > > > >
> > > > > That definition worked great. I must have missed the : : : "memory"
> > > > > bit when
> > > > > I was digging through code.
> > > > >
> > > > > Thanks, that gives me about a 2x speedup over the msync() calls.
> > > >
> > > > Exactly when should you use the barrier? At every access,
> > > > every read or when changing from write to read?
> > >
> > > Well, it depends on the device you are accessing. I'll give you a small
> pseudo example.
> > >
> > > mmio[0] = address;
> > > mmio[1] = data;
> > > mb();
> > > mmio[3] |= 0x01; /* This triggers an operation -> address=data */
> > > /* probably also need an mb() here, if the following code
> > > * depends on the operation to be triggered. */
> >
> > So anything that depends on the previous accesses needs a mb()
>
> No, not really. I would always put an mb() where the comment is.
> Imagine you have two instances of the above (probably in a loop):
>
> mmio[0] = address;
> mmio[1] = data;
> mb();
> mmio[3] |= 0x01; /* This triggers an operation -> address=data */
> mb();
>
> mmio[0] = address;
> mmio[1] = data;
> mb();
> mmio[3] |= 0x01; /* This triggers an operation -> address=data */
> mb();
>
> > hmm, the mmio[0] and mmio[1] are written in order I hope?
>
> We do not care in this example, as the write to [3] does trigger
> the device operation. We do only care that [0] and [1] are set
> when [3] is written. We do not care in what order [0] and [1] are written.
In this example yes, I was wondering in general.
>
> This is just an artificial example, but I think you get my point.
> My point just being is that an mb() does only belong where it matters.
> And what matters is defined by your device specifications.
So what does guarded memory mapping on ppc mean really? If I need
that much mb(), guarded does not seem to do much.
Jocke
_______________________________________________
Linuxppc-dev mailing list
Linuxppc-dev@lists.ozlabs.org
https://lists.ozlabs.org/listinfo/linuxppc-dev
