Re: [PATCH] mm: fix page_mkclean_one (was: 2.6.19 file content corruption on ext3)

[Dave Kleikamp <shaggy@linux.vnet.ibm.com>]

On Wed, 2006-12-20 at 14:25 -0800, Linus Torvalds wrote:
> 
> On Wed, 20 Dec 2006, Dave Kleikamp wrote:
> >
> > This patch removes some questionable code that attempted to make a
> > no-longer-used page easier to reclaim.
> 
> If so, "cancel_dirty_page()" may actually be the right thing to use, but
> only if you can guarantee that the page isn't mapped anywhere (and from
> the name of the function I guess it's not something that you'll ever map?)

That's correct.  It can't be mapped.  It's a private mapping only used
for metadata.

I'm really not sure the code in question is having the intended effect.
Maybe one of the gurus on cc: can take a look at the code and tell me if
it's worth keeping.  I apologize in advance if it makes anyone lose
their lunch.

> So the JFS code _looks_ like you could just replace the
> 
> 	clear_page_dirty(page);
> 
> with
> 
> 	cancel_dirty_page(page, PAGE_CACHE_SIZE);
> 
> (where that second parameter is just used for statistics - it updates the
> "cancelled IO" byte-counts if CONFIG_TASK_IO_ACCOUNTING is set - so the
> number doesn't really matter, you could make it zero if you never want the
> thing to show up in the IO accounting).

I'm not sure whether zero or PAGE_CACHE_SIZE would be better.  The
situation is where some page of metadata is no longer used, say
shrinking a directory tree or truncating a file and throwing out the
extent tree.

Thanks,
Shaggy
-- 
David Kleikamp
IBM Linux Technology Center