Re: [PATCH 05/19] NFS: Use local caching

[Trond Myklebust <trond.myklebust@fys.uio.no>]

On Wed, 2006-11-15 at 16:00 +0000, David Howells wrote:
> Trond Myklebust <trond.myklebust@fys.uio.no> wrote:
> 
> > Why is fscache being given a vote on whether or not the NFS page can be
> > removed from the mapping? If the file has changed on the server, so that
> > we have to invalidate the mapping, then I don't care about the fact that
> > fscache is busy: the page has to go.
> 
> This is releasepage() not invalidatepage().  It is conditional.

...and invalidate_complete_page2() calls try_to_release_page() which
again calls releasepage(). Success or failure of the latter should
therefore not depend on the internal fscache state.

> > > @@ -942,11 +954,13 @@ static int nfs_update_inode(struct inode
> > > ...
> > > +			nfs_fscache_set_size(inode);
> > 
> > Doesn't nfs_fscache_set_size try to grab rw_semaphores? This function is
> > _always_ called with the inode->i_lock spinlock held.
> 
> Hmmm...  I wonder if I need to do this in nfs_update_inode() at all.  Won't the
> pages and the cache object attached to an inode be discarded anyway if the file
> attributes returned by the server change?

In the case of a read-only file, yes. That is not true of a read/write
file.

> > >  static void nfs_readpage_release(struct nfs_page *req)
> > >  {
> > > +	struct inode *d_inode = req->wb_context->dentry->d_inode;
> > > +
> > > +	if (PageUptodate(req->wb_page))
> > > +		nfs_readpage_to_fscache(d_inode, req->wb_page, 0);
> > > +
> > 
> > Will usually be called from an rpciod context. Should therefore not be
> > grabbing semaphores, doing memory allocation etc.
> 
> Is it possible to make an NFS kernel thread that can have completed nfs_page
> structs queued for writing to the cache?

Why should we add extra context switches for the non-fscache case? Just
move the call to nfs_readpage_to_fscache into its own kernel thread.

Trond