lkml.org 
[lkml]   [2010]   [Jul]   [15]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 4/6] writeback: dont redirty tail an inode with dirty pages
On Tue, Jul 13, 2010 at 06:13:17AM +0800, Andrew Morton wrote:
> On Mon, 12 Jul 2010 23:31:27 +0800
> Wu Fengguang <fengguang.wu@intel.com> wrote:
>
> > > > + } else if (inode->i_state & I_DIRTY) {
> > > > + /*
> > > > + * At least XFS will redirty the inode during the
> > > > + * writeback (delalloc) and on io completion (isize).
> > > > + */
> > > > + redirty_tail(inode);
> > >
> > > I'd drop the mention of XFS here - any filesystem that does delayed
> > > allocation or unwritten extent conversion after Io completion will
> > > cause this. Perhaps make the comment:
> > >
> > > /*
> > > * Filesystems can dirty the inode during writeback
> > > * operations, such as delayed allocation during submission
> > > * or metadata updates after data IO completion.
> > > */
> >
> > Thanks, comments updated accordingly.
> >
> > ---
> > writeback: don't redirty tail an inode with dirty pages
> >
> > This avoids delaying writeback for an expired (XFS) inode with lots of
> > dirty pages, but no active dirtier at the moment. Previously we only do
> > that for the kupdate case.
> >
>
> You didn't actually explain the _reason_ for making this change.
> Please always do that.

OK. It's actually extending commit b3af9468ae from the kupdate-only case to
both kupdate and !kupdate cases.

The commit documented the reason:

Debug traces show that in per-bdi writeback, the inode under writeback
almost always get redirtied by a busy dirtier. We used to call
redirty_tail() in this case, which could delay inode for up to 30s.

This is unacceptable because it now happens so frequently for plain cp/dd,
that the accumulated delays could make writeback of big files very slow.

So let's distinguish between data redirty and metadata only redirty.
The first one is caused by a busy dirtier, while the latter one could
happen in XFS, NFS, etc. when they are doing delalloc or updating isize.

Commit b3af9468ae only does that for kupdate case because requeue_io() was
only called in the kupdate case. Now we are merging the kupdate and !kupdate
cases in patch 6/6 (why not?), so is this patch.

> The patch is... surprisingly complicated, although the end result
> looks OK. This is not aided by the partial duplication between
> mapping_tagged(PAGECACHE_TAG_DIRTY) and I_DIRTY_PAGES. I don't think
> we can easily remove I_DIRTY_PAGES because it's used for the
> did-someone-just-dirty-a-page test here.

I double checked I_DIRTY_PAGES. The main difference to PAGECACHE_TAG_DIRTY is:
I_DIRTY_PAGES (at the line removed by this patch) means there are _new_ pages
get dirtied during writeback, while PAGECACHE_TAG_DIRTY means there are dirty
pages. In this sense, if the I_DIRTY_PAGES handling is the same as
PAGECACHE_TAG_DIRTY, the code can be merged into PAGECACHE_TAG_DIRTY, as this
patch does.

The other minor differences are

- in *_set_page_dirty*(), PAGECACHE_TAG_DIRTY is set racelessly, while
I_DIRTY_PAGES might be set on the inode for a page just truncated.
The difference has no real impact on this patch (it's actually
slightly better now).

- afs_fsync() always set I_DIRTY_PAGES after calling afs_writepages().
The call was there in the first day (introduce by David Howells).
What was the intention, hmm..?

> This code is way too complex and fragile and I fear that anything we do
> to it will break something :(

Agreed. Let's try to simplify it :)

Thanks,
Fengguang


\
 
 \ /
  Last update: 2010-07-15 17:37    [from the cache]
©2003-2011 Jasper Spaans