lkml.org 
[lkml]   [2009]   [Aug]   [1]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: Bug in kernel 2.6.31, Slow wb_kupdate writeout
On Thu, Jul 30, 2009 at 03:34:12PM -0700, Martin Bligh wrote:
> > The test case above on a 4G machine is only generating 1G of dirty data.
> > I ran the same test case on the 16G, resulting in only background
> > writeout. The relevant bit here being that the background writeout
> > finished quickly, writing at disk speed.
> >
> > I re-ran the same test, but using 300 100MB files instead. While the
> > dd's are running, we are going at ~80MB/sec (this is disk speed, it's an
> > x25-m). When the dd's are done, it continues doing 80MB/sec for 10
> > seconds or so. Then the remainder (about 2G) is written in bursts at
> > disk speeds, but with some time in between.
>
> OK, I think the test case is sensitive to how many files you have - if
> we punt them to the back of the list, and yet we still have 299 other
> ones, it may well be able to keep the disk spinning despite the bug
> I outlined.Try using 30 1GB files?
>
> Though it doesn't seem to happen with just one dd streamer, and
> I don't see why the bug doesn't trigger in that case either.

I guess the bug is not related to number dd streamers, but whether
there is a stream of newly dirtied inodes (atime dirtiness would be
enough). Because wb_kupdate() itself won't give up on congestion, but
redirty_tail() would refresh the inode dirty time if there are newly
dirtied inodes in front. And we cannot claim it to be a bug of the
list based redirty_tail(), since we call it with the belief that the
inode is somehow blocked. In this manner redirty_tail() can refresh
the inode dirty time (and therefore delay its writeback for up to 30s)
at will.

> I believe the bugfix is correct independent of any bdi changes?

Agreed.

Thanks,
Fengguang



\
 
 \ /
  Last update: 2009-08-01 06:05    [W:0.893 / U:0.768 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site