Re: [PATCH] fix remap of shared read only mappings

[Daniel Phillips <phillips@arcor.de>]

On Friday 05 September 2003 03:07, Alan Cox wrote:
> On Gwe, 2003-09-05 at 01:49, Linus Torvalds wrote:
> > What really matters is that mmap() under Linux is 100% coherent, as far
> > as the hardware just allows. We haven't taken the easy way out. We
> > shouldn't start now.
>
> NFS ?

NFS doesn't attempt to implement local Posix semantics, but a DFS should.  
Anyway, Linus has ruled we're being held to the higher standard of "Linux 
semantics".

> The problem with OpenGFS is that it is a network file system so
> implementing "perfect" shared mmap semantics might actually reduce it
> from handy to useless. Right now the worst we have to do is mark pages
> uncached in some weird shared map cases, with pages being bounced across
> firewire its a bit different.

Sistina has been doing it the "perfect" way for some time now, and it's worked 
out OK.  This relies on writes being rare.

Things will definitely slow to a crawl if several nodes keep writing little 
bits into the same mmap, but that's a case of "doctor it hurts" I think.  If 
there's a common application that does this, I'd appreciate knowing about it.

Regards,

Daniel