cvs commit: src/sys/conf NOTES files options src/sys/kern vfs_bio.c src/sys/ufs/ffs ffs_rawread.c ffs_vnops.c

[Eivind Eklund]

On Thu, Mar 27, 2003 at 09:39:18AM -0800, Alfred Perlstein wrote:
> * Eivind Eklund <eivind at FreeBSD.org> [030327 03:16] wrote:
> > On Wed, Mar 26, 2003 at 03:40:42PM -0800, Tor Egge wrote:
> > > tegge       2003/03/26 15:40:42 PST
> > > 
> > >   FreeBSD src repository
> > > 
> > >   Modified files:
> > >     sys/conf             NOTES files options 
> > >     sys/kern             vfs_bio.c 
> > >     sys/ufs/ffs          ffs_vnops.c 
> > >   Added files:
> > >     sys/ufs/ffs          ffs_rawread.c 
> > >   Log:
> > >   Add support for reading directly from file to userland buffer when the
> > >   O_DIRECT descriptor status flag is set and both offset and length is a
> > >   multiple of the physical media sector size.
> > 
> > Why only in the O_DIRECT case?  O_DIRECT is normally used to avoid caching
> > effects *in order to avoid side effects of caching*, at the cost of speed.
> > This patch seems to do the opposite - speed up the codepath significantly.
> > 
> > Not that it isn't nice - it would just be even nicer to have available some
> > way to just say "max speed, please".
> 
> Actually this is what O_DIRECT was intended to be as soon as someone
> got around to it.  I initially thought it would be dillon, but tegge
> was a pleasant suprise.  :)

I am not protesting that this is appropriate for O_DIRECT - I think it
is a brilliant improvement for O_DIRECT :-)  However, if the intent of
O_DIRECT is to avoid caching in order to provide maximum speed, then the
documentation is really off target.  Either the docs or the code or both
seems to be in need of updating.

Eivind.