Re: [RFC][EXPERIMENTAL][PATCH 0/8] memcg: migrate charge at task move

[Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>]

On Thu, 17 Sep 2009 11:23:04 +0900, Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp> wrote:
> Hi.
> 
> These are patches for migrating memcg's charge at task move.
> 
> I know we should fix res_counter's scalability problem first,
> but this feature is also important for me, so I tried making patches
> and they seem to work basically.
> I post them(based on mmotm-2009-09-14-01-57) before going further to get some feedbacks.
> 
> 
> Basic design:
> - Add flag file "memory.migrate_charge" to determine whether charges should be
>   migrated or not. Each bit of "memory.migrate_charge" has meaning(indicate the
>   type of pages the charges of which should be migrated).
> - At can_attach(), isolate pages of the task, call __mem_cgroup_try_charge,
>   and move them to a private list.
> - Call mem_cgroup_move_account() at attach() about all pages on the private list
>   after necessary checks under page_cgroup lock, and put back them to LRU.
> - Cancel charges about all pages remains on the private list on failure or at the end
>   of charge migration, and put back them to LRU.
> 
> 
> I think this design is simple but it has a problem when mounted on the same hierarchy
> with cpuset. This design isolate pages of the task at can_attach(), but attach() of cpuset
> also tries to isolate pages of the task and migrate them if cpuset.memory_migrate is set.
> As a result, pages cannot be memory migrated by cpuset if we set memory.migrate_charge.
> But I think this problem can be handled by user-space to some extent: for example,
> move the task back and move it again with memory.migrate_charge unset.
> So I went to this direction as a first step(I'm considering how to avoid this issue).
> 
> 
As I said in another mail, I'll change the design.

- At can_attach(), parse the page table of the task, do try_charge, and remember the
  count of try_charge(no page isolation).
- At attach, parse the page table again and do move_account.

So the problem above with cpuset disappears.


Thanks,
Daisuke Nishimura.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>