Re: [RFC][PATCH 0/6] memcg updates (05/Nov)

[Balbir Singh <balbir@linux.vnet.ibm.com>]

KAMEZAWA Hiroyuki wrote:
> Weekly (RFC) update for memcg.
> 
> This set includes
> 
> 1. change force_empty to do move account rather than forget all

I would like this to be selectable, please. We don't want to break behaviour and
not everyone would like to pay the cost of movement.

> 2. swap cache handling
> 3. mem+swap controller kconfig
> 4. swap_cgroup for rememver swap account information
> 5. mem+swap controller core
> 6. synchronize memcg's LRU and global LRU.
> 
> "1" is already sent, "6" is a newcomer.
> I'd like to push out "2" or "2-5" in the next week (if no bugs.)
> 
> after 6, next candidates are
>   - dirty_ratio handler
>   - account move at task move.
> 
> Some more explanation about purpose of "6". (see details in patch itself)
> Now, one of complicated logic in memcg is LRU handling. Because the place of
> lru_head depends on page_cgroup->mem_cgroup pointer, we have to take
> lock as following even under zone->lru_lock.
> ==
>   pc = lookup_page_cgroup(page);
>   if (!trylock_page_cgroup(pc))
>   	return -EBUSY;
> 
>    if (PageCgroupUsed(pc)) {
> 	struct mem_cgroup_per_zone *mz = page_cgroup_zoneinfo(pc);
> 	spin_lock_irqsave(&mz->lru_lock, flags);
> 	....some operation on LRU.
> 	spin_unlock_irqrestore(&mz->lru_lock, flags);
>    }
>    unlock_page_cgroup(pc);
> ==
> Sigh..
> 
> After "6", page_cgroup's LRU management can be done independently to some extent.
> == as
>   (zone->lru_lock is held here)
>   pc = lookup_page_cgroup(page);
>   list operation on pc.
>   (unlock zone->lru_lock)
> ==
> Maybe good for maintainance and as a bonus, we can make use of isolate_lru_page() when
> doing some racy operation.
> 
> 	isolate_lru_page(page);
> 	pc = lookup_page_cgroup(page);
> 	do some jobs.
> 	putback_lru_page(page);
> 
> Maybe this will be a help to implement "account move at task move".

Sounds promising!

-- 
	Balbir