On 4/12/07, Roman Zippel <zippel@linux-m68k.org> wrote:
quoted text
> Hi,
>
> On Thu, 12 Apr 2007, Christoph Hellwig wrote:
>
> > On Wed, Apr 11, 2007 at 07:49:38PM -0700, Nate Diller wrote:
> > > read_mapping_page_async() is going away, so convert its only user to
> > > read_mapping_page().  This change has not been benchmarked, however, in
> > > order to get real parallelism this wants something completely different,
> > > like __do_page_cache_readahead(), which is not currently exported.
> >
> > Why is read_mapping_page_async going away?  This probably needs a lot more
> > testing, and I'd be much happier if you split it out of the series and
> > sent it separately at the end.
>
> That function wasn't fully async anyway, as it would often sleep in
> lock_page(). AFAICT only in the special case of a partial written page
> would this function return a not yet uptodate page.

yes, exactly, the structure of read_cache_page() and friends is
totally not appropriate for doing async I/O to more than one page at a
time, and the whole point of the special treatment in cramfs was to
read 4 pages at once rather than synchronously reading each of the 4
seperately.  read_cache_page_async() is totally wrong for that use,
its purpose would be to get a reference to a single page that is
likely to be in cache already without having to take the page_lock.
Turns out nobody needs to do that, so there's no point in keeping it
around.

If the performance gain of reading all 4 pages at once would be worth
the effort, this code should be using __do_page_cache_readahead().
That function allocates all the pages first, then reads them in
asynchronously as a group.  It is currently not exported.

NATE
-
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html