Re: [RESEND][PATCH 0/3 BUG,RFC] release block-device-mapping buffer_heads which have the filesystem private data for avoiding oom-killer

[Toshiyuki Okajima <toshi.okajima@jp.fujitsu.com>]

Hi Andrew,
Thanks for your comments.

 > On Thu, 20 Nov 2008 09:27:11 +0900
 > Toshiyuki Okajima <toshi.okajima@jp.fujitsu.com> wrote:
<SNIP>
 >
 > I'm scratching my head trying to work out why we never encountered and
 > fixed this before.

 > Is it possible that you have a very large number of filesystems
 > mounted, and/or that they have large journals?

Yes, I think it happen more easily under those conditions.

Actually, I encountered this situation if conditions were:
- on the x86 architecture (The size of Normal zone is only 800MB
    even if the huge memory (more than 4GB) install.)
- reserving the big memory (more than 100MB) for the kdump kernel.
   (The memory obtains from Normal Zone.)
- mounting the large number of ext3 filesystems (more than 50).

And the following operations were done:
- many I/Os were issued to many filesystems sequentially and continuously.
(They made many journal_heads (and buffer_heads).
  => they were metadata.)
- issuing the I/Os to many filesystems were stopped.
(This caused many metadata to remain.)

By their operations, the number of remaining the journal_heads was
more than 100000 (They occupied 400MB (The same number of buffer_heads remained
and the block size was 4096B)). We cannot release those journal_heads because
checkpointing the transactions are not executed till some I/Os are issued to
the filesystems or the filesystems were unmounting.
And many other slab caches which couldn't be released occupied about 300MB.
Therefore about 800MB memory couldn't be released.
As a result, there was no room in the Normal zone.

I think you could not encounter it because you haven't done such the following:
- You reserve the big memory for the kdump kernel.
- You issue many I/Os to each ext3 filesystem sequentially and continuously,
  and then you never issue some I/Os to the filesystems at all afterwards.
  (Especially, you do the operations which causes many metadata to remain.
   Example: Delete many files which are huge.)

 > Would it not be more logical if the ->client_releasepage function
 > pointer were a member of the blockdev address_space_operations, rather
 > than some random field in the blockdev inode?  That arrangement might
 > well be reused in the future, when some other address_space needs to
 > talk to a different address_space to make a page reclaimable.

I think it logical to replace a default ->releasepage with a function pointer
which a client (FS) passed, but I don't think it logical to add a new member
function in address space in order to release a client page. Because new
function is called from ->releasepage, so I think this function pointer should
not be put in the same level as the releasepage of address space.

Though, it is difficult to replace ->releasepage member with a client function
because there is no exclusive operation while this function is calling.

So, I made this patch (without replacing ->releasepage).

How about my thought?

Best Regards,
Toshiyuki Okajima