On Thu, Apr 10, 2003 at 10:15:03AM +0900, YAMAMOTO Takashi wrote:
> hi
> 
> > yea, this makes sense.  the more general description of the problem
> > is that nfs_strategy() should only use the commit mechanism when it's
> > being called from VOP_PUTPAGES() on an NFS vnode, but there are other
> > ways that nfs_strategy() can be called.  NFS swap files is one way we
> > can go wrong, but there are others, such as using an NFS file for
> > a vnd backing file.  one way to fix this more completely would be to
> > have the NFS VOP_PUTPAGES() path add an extra flag to its buffers
> > that nfs_strategy() can look at to decide if it should use stable writes.
> > I had written this up a while back but didn't commit it for some reason.
> > the diff is attached if you want to pursue it.
> 
> the patch seems to disable commit mechanism even in the case of
> synchronous B_PAGEIO. it's a overkill, isn't it?

I hadn't bothered to try to optimize synchronous i/o requests.
if you want to do so, that would be fine.

the change you checked in isn't quite correct, though.
if someone does a physio write (via vnd) to an NFS file from a mapping of
the same file at a different offset, you'd set PG_NEEDCOMMMIT on the
pages from the mapping instead of the pages for the range being modified
(which might not even exist on the client).

-Chuck