On Wed 02-01-13 11:44:21, Michal Hocko wrote:
> On Wed 26-12-12 01:26:07, Sha Zhengju wrote:
> > From: Sha Zhengju <[email protected]>
> >
> > This patch adds memcg routines to count dirty pages, which allows memory controller
> > to maintain an accurate view of the amount of its dirty memory and can provide some
> > info for users while cgroup's direct reclaim is working.
>
> I guess you meant targeted resp. (hard/soft) limit reclaim here,
> right? It is true that this is direct reclaim but it is not clear to me
> why the usefulnes should be limitted to the reclaim for users. I would
> understand this if the users was in fact in-kernel users.
>
> [...]
> > To prevent AB/BA deadlock mentioned by Greg Thelen in previous version
> > (https://lkml.org/lkml/2012/7/30/227), we adjust the lock order:
> > ->private_lock --> mapping->tree_lock --> memcg->move_lock.
> > So we need to make mapping->tree_lock ahead of TestSetPageDirty in __set_page_dirty()
> > and __set_page_dirty_nobuffers(). But in order to avoiding useless spinlock contention,
> > a prepare PageDirty() checking is added.
>
> But there is another AA deadlock here I believe.
> page_remove_rmap
> mem_cgroup_begin_update_page_stat <<< 1
> set_page_dirty
> __set_page_dirty_buffers
> __set_page_dirty
> mem_cgroup_begin_update_page_stat <<< 2
> move_lock_mem_cgroup
> spin_lock_irqsave(&memcg->move_lock, *flags);
JFYI since abf09bed (s390/mm: implement software dirty bits) this is no
longer possible. I haven't checked wheter there are other cases like
this one and it should be better if mem_cgroup_begin_update_page_stat
was recursive safe if that can be done without too many hacks.
I will have a look at this (hopefully) sometimes next week.
[...]
--
Michal Hocko
SUSE Labs
On Fri, May 3, 2013 at 5:11 PM, Michal Hocko <[email protected]> wrote:
> On Wed 02-01-13 11:44:21, Michal Hocko wrote:
>> On Wed 26-12-12 01:26:07, Sha Zhengju wrote:
>> > From: Sha Zhengju <[email protected]>
>> >
>> > This patch adds memcg routines to count dirty pages, which allows memory controller
>> > to maintain an accurate view of the amount of its dirty memory and can provide some
>> > info for users while cgroup's direct reclaim is working.
>>
>> I guess you meant targeted resp. (hard/soft) limit reclaim here,
>> right? It is true that this is direct reclaim but it is not clear to me
>> why the usefulnes should be limitted to the reclaim for users. I would
>> understand this if the users was in fact in-kernel users.
>>
>> [...]
>> > To prevent AB/BA deadlock mentioned by Greg Thelen in previous version
>> > (https://lkml.org/lkml/2012/7/30/227), we adjust the lock order:
>> > ->private_lock --> mapping->tree_lock --> memcg->move_lock.
>> > So we need to make mapping->tree_lock ahead of TestSetPageDirty in __set_page_dirty()
>> > and __set_page_dirty_nobuffers(). But in order to avoiding useless spinlock contention,
>> > a prepare PageDirty() checking is added.
>>
>> But there is another AA deadlock here I believe.
>> page_remove_rmap
>> mem_cgroup_begin_update_page_stat <<< 1
>> set_page_dirty
>> __set_page_dirty_buffers
>> __set_page_dirty
>> mem_cgroup_begin_update_page_stat <<< 2
>> move_lock_mem_cgroup
>> spin_lock_irqsave(&memcg->move_lock, *flags);
>
> JFYI since abf09bed (s390/mm: implement software dirty bits) this is no
> longer possible. I haven't checked wheter there are other cases like
> this one and it should be better if mem_cgroup_begin_update_page_stat
> was recursive safe if that can be done without too many hacks.
> I will have a look at this (hopefully) sometimes next week.
>
Hi Michal,
I'm sorry for not being able to return to this problem immediately after LSF/MM.
That is good news. IIRC, it's the only place we have encountered
recursive problem in accounting memcg dirty pages. But I'll try to
revive my previous work of simplifying
mem_cgroup_begin_update_page_stat() lock.
I'll back to it in next few days.
--
Thanks,
Sha