On Wed 11-09-13 11:45:03, Bernd Schubert wrote:

On 09/10/2013 09:40 PM, Jan Kara wrote:

quoted

Linux Kernel Performance project guys have reported that commit 4e7ea81db5
introduces a performance regression for the following fio workload:
[global]
direct=0
ioengine=mmap
size=1500M
bs=4k
pre_read=1
numjobs=1
overwrite=1
loops=5
runtime=300
group_reporting
invalidate=0
directory=/mnt/
file_service_type=random:36
file_service_type=random:36

[job0]
startdelay=0
rw=randrw
filename=data0/f1:data0/f2

[job1]
startdelay=0
rw=randrw
filename=data0/f2:data0/f1
...

[job7]
startdelay=0
rw=randrw
filename=data0/f2:data0/f1

The culprit of the problem is that after the commit ext4_writepages()
are more aggressive in writing back pages. Thus we have less consecutive
dirty pages resulting in more seeking.

This increased aggressivity is caused by a bug in the condition
terminating ext4_writepages(). We start writing from the beginning of
the file even if we should have terminated ext4_writepages() because
wbc->nr_to_write <= 0.

After fixing the condition the throughput of the fio workload is about 20%
better than before writeback reorganization.

Reported-by: "Yan, Zheng" <redacted>
Signed-off-by: Jan Kara <jack@suse.cz>
---
 fs/ext4/inode.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
index c79fd7d..7914c05 100644
--- a/fs/ext4/inode.c
+++ b/fs/ext4/inode.c

@@ -2563,7 +2563,7 @@ retry:
 			break;
 	}
 	blk_finish_plug(&plug);
-	if (!ret && !cycled) {
+	if (!ret && !cycled && wbc->nr_to_write > 0) {
 		cycled = 1;
 		mpd.last_page = writeback_index - 1;
 		mpd.first_page = 0;

Interesting, doesn't that mean generic_writepages (sub-sequent
write_cache_pages() ) and all other file systems implementing their
own ->writepages()  should be updated?

  No. write_cache_pages() has the condition like:
if (!cycled && !done) {

  and 'done' is set when wbc->nr_to_write drops to zero. So that function
is OK. We cannot use 'done' in ext4_writepages() because the functions are
structured a bit differently and 'done' gets set also when reach end of
file.

								Honza
-- 
Jan Kara [off-list ref]
SUSE Labs, CR