Re: [RFC 1/2] deactive invalidated pages

[Andrew Morton <akpm@linux-foundation.org>]

On Tue, 23 Nov 2010 16:40:03 +0900 Minchan Kim <minchan.kim@gmail.com> wrote:

> Hi KOSAKI,
> 
> 2010/11/23 KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>:
> >> By Other approach, app developer uses POSIX_FADV_DONTNEED.
> >> But it has a problem. If kernel meets page is writing
> >> during invalidate_mapping_pages, it can't work.
> >> It is very hard for application programmer to use it.
> >> Because they always have to sync data before calling
> >> fadivse(..POSIX_FADV_DONTNEED) to make sure the pages could
> >> be discardable. At last, they can't use deferred write of kernel
> >> so that they could see performance loss.
> >> (http://insights.oetiker.ch/linux/fadvise.html)
> >
> > If rsync use the above url patch, we don't need your patch.
> > fdatasync() + POSIX_FADV_DONTNEED should work fine.
> 
> It works well. But it needs always fdatasync before calling fadvise.
> For small file, it hurt performance since we can't use the deferred write.

fdatasync() is (much) better than nothing, but a userspace application
which is carefully managing its IO scheduling should use
sync_file_range(SYNC_FILE_RANGE_WRITE) to push data at the disk and
should then run fadvise(DONTNEED) against the same data a few seconds
later, after the IO has completed.

That way, the application won't block against the write I/O at all,
unless of course someone else is thrashing the disk as well, etc.

If the app is doing a lot of file I/O (eg, rsync) then this shouldn't
be too hard to arrange.  Although the payback will be pretty small
unless the IO-intensive process is also compute-intensive at times. 
And such applications are a) fairly rare and b) poorly designed:
shouldn't be doing heavy IO and heavy compute in the same thread!