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

[Trond Myklebust]

On Wed, 2006-11-15 at 17:07 +0000, David Howells wrote:
> Trond Myklebust <trond.myklebust@xxxxxxxxxx> wrote:
> 
> > > 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.
> 
> Okay... invalidate_complete_page2() passes __GFP_WAIT through.  I'll make
> nfs_fscache_release_page() check for that, and if it's set, it'll wait for
> FS-Cache to finish with the page before returning true.
> 
> This sounds like invalidate_inode_pages2() is doing the wrong thing.  After
> all, releasepage() _is_ conditional.  It sounds like it should be calling
> invalidatepage() instead.  Either that or NFS should be calling something
> else entirely.

No! invalidate_inode_pages2() is supposed to be non-destructive w.r.t.
dirty pages. That is why it calls try_to_release_page(). If you want a
destructive truncate, then you should be calling truncate_inode_pages().

> > > 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.
> 
> So I can assume that, as we're only caching read-only files, I don't need to
> invoke FS-Cache here.

I would assume so.

> > > > >  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.
> 
> Sorry, I meant can I make use of the nfs_page struct that was handed to
> nfs_readpage_release() by queuing it for an auxiliary thread to call
> nfs_readpage_to_fscache() on?  I didn't intend to call nfs_readpage_release()
> in another thread.

If you do, then you will either have to get rid of the call to
nfs_clear_request(), or track the struct page yourself.

At this point, I think that removing the call to nfs_clear_request is
safe: there should be nothing that checks page_count() in the mm layer
that we can race with AFAICS.

Cheers,
  Trond

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/