Re: [PATCH 2/4] writeback: 3-queue based writeback schedule

[Fengguang Wu <wfg@mail.ustc.edu.cn>]

On Fri, Aug 10, 2007 at 02:34:14PM +0800, Fengguang Wu wrote:
> Properly manage the 3 queues of sb->s_dirty/s_io/s_more_io so that
> 	- time-ordering of dirtied_when can be easily maintained
> 	- writeback can continue from where previous run left out
> 
> The majority work has been done by Andrew Morton and Ken Chen,
> this patch just clarifies the roles of the 3 queues:
> - s_dirty   for io delay(up to dirty_expire_interval)
> - s_io      for io run(a full scan of s_io may involve multiple runs)
> - s_more_io for io continuation
> 
> The following paradigm shows the data flow.
> 
>                             requeue on new scan(empty s_io)
>                             +-----------------------------+
>                             |                             |
>  dirty           old        |                             |
>  inodes          enough     V                             |
>  ======> s_dirty ======> s_io                             |
>          ^                |     requeue io                |
>          |                +---------------------> s_more_io
>          |   hold back    |
>          +----------------+----------> disk write requests
> 
> sb->s_dirty: a FIFO queue
> - s_dirty hosts not-yet-expired(recently dirtied) dirty inodes
> - once expired, inodes will be moved out of s_dirty and *never put back*
>   (unless for some reason we have to hold on the inode for some time)
> 
> sb->s_io and sb->s_more_io: a cyclic queue scanned for io
> - on each run of generic_sync_sb_inodes(), some more s_dirty inodes may be
>   appended to s_io
> - on each full scan of s_io, all s_more_io inodes will be moved back to s_io
> - large files that cannot be synced in one run will be moved to s_more_io for
>   retry on next full scan

In fact s_more_io is no longer necessary. We end up with a priority
io-delaying queue s_dirty and a cyclic io-syncing queue s_io. They are
properly decoupled.  More flexible data structure can be used for
s_dirty, if we want to redirty an inode with arbitrary delays. Also
more priority queues can be introduced in addition to s_dirty. For
example, we can designate a queue s_dirty_atime for inodes dirtied
only by `atime', and sync them lazily.

> inode->dirtied_when
> - inode->dirtied_when is updated to the *current* jiffies on pushing into
>   s_dirty, and is never changed in other cases.
> - time-ordering thus can be simply ensured while moving inodes between lists,
>   since (time order == enqueue order)
> 
> Cc: Ken Chen <kenchen@google.com>
> Cc: Andrew Morton <akpm@linux-foundation.org>
> Signed-off-by: Fengguang Wu <wfg@mail.ustc.edu.cn>
> ---
>  fs/fs-writeback.c |  106 +++++++++++++++++++++-----------------------
>  1 file changed, 52 insertions(+), 54 deletions(-)