Re: [RFC/PATCH] cgroup swap subsystem

[KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>]

On Wed, 05 Mar 2008 14:59:05 +0900
Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp> wrote:

>  #ifdef CONFIG_CGROUP_MEM_CONT
> +/*
> + * A page_cgroup page is associated with every page descriptor. The
> + * page_cgroup helps us identify information about the cgroup
> + */
> +struct page_cgroup {
> +	struct list_head lru;		/* per cgroup LRU list */
> +	struct page *page;
> +	struct mem_cgroup *mem_cgroup;
> +#ifdef CONFIG_CGROUP_SWAP_LIMIT
> +	struct mm_struct *pc_mm;
> +#endif
> +	atomic_t ref_cnt;		/* Helpful when pages move b/w  */
> +					/* mapped and cached states     */
> +	int	 flags;
> +};
>  
As first impression, I don't like to increase size of this...but have no alternative
idea.



>  static inline int page_cgroup_locked(struct page *page)
> @@ -664,6 +665,10 @@ retry:
>  	pc->flags = PAGE_CGROUP_FLAG_ACTIVE;
>  	if (ctype == MEM_CGROUP_CHARGE_TYPE_CACHE)
>  		pc->flags |= PAGE_CGROUP_FLAG_CACHE;
> +#ifdef CONFIG_CGROUP_SWAP_LIMIT
> +	atomic_inc(&mm->mm_count);
> +	pc->pc_mm = mm;
> +#endif
>  
Strongly Nack to this atomic_inc(). 
What happens when tmpfs pages goes to swap ?


>  	if (!page || page_cgroup_assign_new_page_cgroup(page, pc)) {
>  		/*
> @@ -673,6 +678,9 @@ retry:

> +int swap_cgroup_charge(struct page *page,
> +			struct swap_info_struct *si,
> +			unsigned long offset)
> +{
> +	int ret;
> +	struct page_cgroup *pc;
> +	struct mm_struct *mm;
> +	struct swap_cgroup *swap;
> +
> +	BUG_ON(!page);
> +
> +	/*
> +	 * Pages to be swapped out should have been charged by memory cgroup,
> +	 * but very rarely, pc would be NULL (pc is not reliable without lock,
> +	 * so I should fix here).
> +	 * In such cases, we charge the init_mm now.
> +	 */
> +	pc = page_get_page_cgroup(page);
> +	if (WARN_ON(!pc))
> +		mm = &init_mm;
> +	else
> +		mm = pc->pc_mm;
> +	BUG_ON(!mm);
> +
> +	rcu_read_lock();
> +	swap = rcu_dereference(mm->swap_cgroup);
> +	rcu_read_unlock();
> +	BUG_ON(!swap);
Is there no race ?

At first look, remembering mm struct is not very good.
Remembering swap controller itself is better.
If you go this direction, how about this way ?

==
enum {
#ifdef CONFIG_CGROUP_MEM_CONT
	MEMORY_RESOURCE_CONTROLLER,
#endif
#ifdef CONFIG_CGROUP_SWAP_CONT
	SWAP_CONTROLLER,
#endif
	NR_PAGE_CONTROLLER,
}

struct page_cgroup {
	......
	void*	controlls[NR_PAGE_CONTROLLER];
	....
};
==

Thanks,
-Kame





--
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>