On Tue, Jun 24 2008, Miklos Szeredi wrote:
> > > The 'confirm' operation was only used for splicing from page cache, to
> > > wait for read on a page to finish. But generic_file_splice_read()
> > > already blocks on readahead reads, so it seems logical to block on the
> > > rare and slow single page reads too.
> > >
> > > So wait for readpage to finish inside __generic_file_splice_read() and
> > > remove the 'confirm' method.
> > >
> > > This also fixes short return counts when the filesystem (e.g. fuse)
> > > invalidates the page between insertation and removal.
> >
> > One of the basic goals of splice is to allow the pipe buffer to only be
> > consisten when a consumer asks for it, otherwise the filling will always
> > be sync. There should be no blocking on reads in the splice-in path,
> > only on consumption for splice-out.
>
> What you are ignoring (and I've mentioned in the changelog) is that it
> is *already* sync. Look at the code: this starts I/O:
>
> page_cache_sync_readahead(mapping, &in->f_ra, in,
> index, req_pages - spd.nr_pages);
>
> And this waits for it to finish:
>
> if (!PageUptodate(page)) {
> ...
> lock_page(page);
>
> The only way it will be async, is if there's no readahead. But do we
> want to optmize that case?
It's an unfortunate side effect of the read-ahead, I'd much rather just
get rid of that. It _should_ behave like the non-ra case, when a page is
added it merely has IO started on it. So we want to have that be
something like
if (!PageUptodate(page) && !PageInFlight(page))
...
basically like PageWriteback(), but for read-in.
--
Jens Axboe
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
