Re: [PATCH 5/6] writeback: sync expired inodes first in background writeback

[Wu Fengguang <fengguang.wu@intel.com>]

> One of the many requirements for writeback is that if userspace is
> continually dirtying pages in a particular file, that shouldn't cause
> the kupdate function to concentrate on that file's newly-dirtied pages,
> neglecting pages from other files which were less-recently dirtied. 
> (and dirty nodes, etc).

Sadly I do find the old pages that the flusher never get a chance to
catch and write them out.

In the below case, if the task dirties pages fast enough at the end of
file, writeback_index will never get a chance to wrap back. There may
be various variations of this case.

file head
[          ***                        ==>***************]==>
           old pages          writeback_index            fresh dirties

Ironically the current kernel relies on pageout() to catch these
old pages, which is not only inefficient, but also not reliable.
If a full LRU walk takes an hour, the old pages may stay dirtied
for an hour.

We may have to do (conditional) tagged ->writepages to safeguard users
from losing data he'd expect to be written hours ago.

Thanks,
Fengguang