On Fri 09-02-24 01:20:18, Kemeng Shi wrote:
> In kupdate writeback, only expired inode (have been dirty for longer than
> dirty_expire_interval) is supposed to be written back. However, kupdate
> writeback will writeback non-expired inode left in b_io or b_more_io from
> last wb_writeback. As a result, writeback will keep being triggered
> unexpected when we keep dirtying pages even dirty memory is under
> threshold and inode is not expired. To be more specific:
> Assume dirty background threshold is > 1G and dirty_expire_centisecs is
> > 60s. When we running fio -size=1G -invalidate=0 -ioengine=libaio
> --time_based -runtime=60... (keep dirtying), the writeback will keep
> being triggered as following:
> wb_workfn
> wb_do_writeback
> wb_check_background_flush
> /*
> * Wb dirty background threshold starts at 0 if device was idle and
> * grows up when bandwidth of wb is updated. So a background
> * writeback is triggered.
> */
> wb_over_bg_thresh
> /*
> * Dirtied inode will be written back and added to b_more_io list
> * after slice used up (because we keep dirtying the inode).
> */
> wb_writeback
>
> Writeback is triggered per dirty_writeback_centisecs as following:
> wb_workfn
> wb_do_writeback
> wb_check_old_data_flush
> /*
> * Write back inode left in b_io and b_more_io from last wb_writeback
> * even the inode is non-expired and it will be added to b_more_io
> * again as slice will be used up (because we keep dirtying the
> * inode)
> */
> wb_writeback
>
> Fix this by moving non-expired inode in io list from last wb_writeback to
> dirty list in kudpate writeback.
>
> Test as following:
> /* make it more easier to observe the issue */
> echo 300000 > /proc/sys/vm/dirty_expire_centisecs
> echo 100 > /proc/sys/vm/dirty_writeback_centisecs
> /* create a idle device */
> mkfs.ext4 -F /dev/vdb
> mount /dev/vdb /bdi1/
> /* run buffer write with fio */
> fio -name test -filename=/bdi1/file -size=800M -ioengine=libaio -bs=4K \
> -iodepth=1 -rw=write -direct=0 --time_based -runtime=60 -invalidate=0
>
> Result before fix (run three tests):
> 1360MB/s
> 1329MB/s
> 1455MB/s
>
> Result after fix (run three tests);
> 790MB/s
> 1820MB/s
> 1804MB/s
>
> Signed-off-by: Kemeng Shi <shikemeng@xxxxxxxxxxxxxxx>
OK, I don't find this a particularly troubling problem but I agree it might
be nice to fix. But filtering the lists in wb_writeback() like this seems
kind of wrong - the queueing is managed in queue_io() and I'd prefer to
keep it that way. What if we just modified requeue_inode() to not
requeue_io() inodes in case we are doing kupdate style writeback and inode
isn't expired?
Sure we will still possibly writeback unexpired inodes once before calling
redirty_tail_locked() on them but that shouldn't really be noticeable?
Honza
> ---
> fs/fs-writeback.c | 20 ++++++++++++++++++++
> 1 file changed, 20 insertions(+)
>
> diff --git a/fs/fs-writeback.c b/fs/fs-writeback.c
> index 5ab1aaf805f7..a9a918972719 100644
> --- a/fs/fs-writeback.c
> +++ b/fs/fs-writeback.c
> @@ -2046,6 +2046,23 @@ static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
> return nr_pages - work.nr_pages;
> }
>
> +static void filter_expired_io(struct bdi_writeback *wb)
> +{
> + struct inode *inode, *tmp;
> + unsigned long expired_jiffies = jiffies -
> + msecs_to_jiffies(dirty_expire_interval * 10);
> +
> + spin_lock(&wb->list_lock);
> + list_for_each_entry_safe(inode, tmp, &wb->b_io, i_io_list)
> + if (inode_dirtied_after(inode, expired_jiffies))
> + redirty_tail(inode, wb);
> +
> + list_for_each_entry_safe(inode, tmp, &wb->b_more_io, i_io_list)
> + if (inode_dirtied_after(inode, expired_jiffies))
> + redirty_tail(inode, wb);
> + spin_unlock(&wb->list_lock);
> +}
> +
> /*
> * Explicit flushing or periodic writeback of "old" data.
> *
> @@ -2070,6 +2087,9 @@ static long wb_writeback(struct bdi_writeback *wb,
> long progress;
> struct blk_plug plug;
>
> + if (work->for_kupdate)
> + filter_expired_io(wb);
> +
> blk_start_plug(&plug);
> for (;;) {
> /*
> --
> 2.30.0
>
--
Jan Kara <jack@xxxxxxxx>
SUSE Labs, CR
