Delaying page writeouts

Delaying page writeouts

[Martin Jambor]

Hi,

our filesystem does not write individual pages at a time but rather a
bunch of them together. Because of that our writepage(s) handlers need
to queue pages until there is enough of them.  In case of sync
operations, the queue may be flushed to disk at any time but doing so
for every inode would be terribly ineffective.

The solution I am thinking of at the moment is following:

1. Put the page into my queue of requests.
2. Get an extra reference on every page passed to writepage with
page_cache_get().
3. Call set_page_writeback(), unlock_page() and end_page_writeback().
4. if there is enough pages, build necessary metadata structures and
start BIOing the
    whole thing to disk.
5. Once BIOs return call page_cache_put() to release the pages.

In case of sync and fsync, the queue would be flushed after all
relevant inodes have written their dirty data into it.

The question is: Do you think this is viable?

Thank you for any comment on this,

Martin Jambor