Hi Andrew.
Thank you for your useful comment.

Andrew Morton wrote:
 > (added linux-fsdevel)
 >
 > On Thu, 23 Oct 2008 17:41:01 +0900
 > Toshiyuki Okajima [off-list ref] wrote:
<SNIP>
 > > I managed to build a shrinker callback function for the journal_head slab cache.
 > > > This code size is less than before but the logic of it seems to be more complex
 > > >  than before.
 > > > However, I haven't got any troubles while I am testing some easy load operations
 > > > on the fixed kernel.
 > > > But I think a system may hang up if concurrently several journal_head shrinker
 > > > are executed.
 > > > So, I will retry to build more appropriate fix.
 >
 > yeah, that's not very pretty either, is it?
Yes...
I realized fixing only within buffer_head or jbd needs a more complex code.

 > > > Please give me comments if you have a nicer idea.
 > Stepping back a bit...
 >
 > The basic problem is, I believe, that some client of the blockdev
 > (ext3) is adding metadata to the blockdev's data structures
 > (buffer_heads) but we have no means by which the blockdev code can call
 > back into that client requesting that the metadata be released, yes?
Yes.

 > We can fix the problem which you've identified by adding a means for
 > the blockdev code (def_blk_aops.releasepage()) to call back into ext3,
 > yes?
Yes.

At first, I tried to fix by using only filesystem approach.
  - ver.1: fixing in buffer_head
  - ver.2: adding shrinker of journal_head (for releasing buffer_head)
But these approaches become complex code.
So, we should fix the essence of the problem.

 > If so, how do we do that?
 >
 > I seem to recall that there's code somewhere in the tree which does
 > things like taking a copy of bdev->address_space_operations and
 > reinstalling that, and overwriting selected fields, and then arranging
 > somehow for the old value to be reinstalled when the client releases
 > the blockdev.  That's plain nasty.
uh-huh.

I try to fix this problem again by using your approach,
blkdev_register_releasepage/blkdev_unregister_releasepage().

Best Regards,
Toshiyuki Okajima