Potentially, but it depends on the implementation of copyin()/copyout() 
for that particular operating system.  

Machines which support two or more address spaces at the same time, 
such as SPARC, use use memcpy with a few hooks to handle trap recovery.  

But PowerPC only supports a single address space, so the copyin()/copyout() 
implementations are a bit more compilcated.  Either the page needs to be 
mapped into the kernel, which is usually an expensive operation, or it 
needs to use some machine-dependent magic to allow access to user 
addresses.  

I don't know how AIX does this.

Eduardo

On Tue, Aug 19, 2003 at 06:35:57PM +0530, Kamal R Prasad wrote:
> Hello,
> 
> If I understand correctly, if the userland page is pinned -then 
> memcpy(userspace, kernelspace) is as expensive as within a single context. 
> Pl. correct me if Im wrong.
> 
> thanks
> -kamal
> 
> 
> 
> 
> 
> 
> Eduardo Horvath <eeh@netbsd.org>
> 08/18/2003 11:55 PM
>  
>         To:     Kamal R Prasad/India/IBM@IBMIN
>         cc:     der Mouse <mouse@Rodents.Montreal.QC.CA>, 
> tech-kern@netbsd.org
>         Subject:        Re: Measuring memory to memory copy operations (+ 
> interrupts, DMA)
> 
>  
> 
> On Mon, Aug 18, 2003 at 04:01:07PM +0530, Kamal R Prasad wrote:
> > > I am modifying [the clock interrupt handler] and want to add a couple
> > > of mem-to-mem copys [copy from userspace to kernel space] in the
> > > handler.
> > 
> > > The code is on AIX -but I expect similar performance on most versions
> > > of UNIX.
> > 
> > >Well, this is a NetBSD list.  If you want AIX-specific advice, this is
> > >a wrong place to ask; my comments will apply to NetBSD.
> > 
> > Im aware of this.
> > 
> > >This is a rather dangerous thing to do.  Unless you have arranged for
> > >something to be present in every process at some userspace address
> > >that's fixed (or otherwise easily computable), there's no safe way to
> > >do this.  Even if you have, I'm not entirely convinced it's enough, as
> > >I'm not sure there's always a userland in the MMU (for example, if no
> > >process has been run since the last process-exit, I'm not sure what's
> > >left in the MMU - it's probably port-specific).
> > 
> > http://netbsd.gw.com/cgi-bin/man-cgi?profil+2+NetBSD-current
> > 
> > this function and the code for updating timer ticks is inside the 
> kernel.
> > 
> > > is it affordable to copy in about 100 bytes twice [and copy out a few
> > > bytes] within the routine if it is being called every one second?
> > 
> > >"Probably."  This depends on your CPU speed, how expensive copyin() is
> > >on the port you're using, and how much performance degradation you're
> > [snip]
> > 
> > lets assume a 1 GHZ cpu speed and the port is obviously powerpc.
> > I am looking for info about overhead associated with copying from one 
> > context to another.
> 
> copyin()/copyout() can result in pages being retrieved from swap.  If
> this operation is done in the clock interrupt handler you can end up
> with a stopped clock for the duration of the disk I/O or worse, a hard
> hang as the storage driver waits for an interrupt that never arrives
> because it's currently masked.
> 
> Eduardo
> 
>