Re: [RFC] memory controller : backgorund reclaim and avoid excessive locking [3/5] throttling

[Balbir Singh <balbir-23VcF4HTsmIX0ybBhKVfKdBPR1lH4CV8@public.gmane.org>]

KAMEZAWA Hiroyuki wrote:
> Throttle memory reclaim interface for cgroup.
> 
> This patch adds..
>  - a limit for simultaneous callers of try_to_free_mem_cgroup_pages().
>  - interface for that. memory.shrinkers.
> 
> There are some reasons.
>  - try_to_free... is very heavy and shoulnd't be called too much at once.
>  - When the number of callers of try_to_free.. is big, we'll reclaim
>    too much memory.
> 
> By this interface, a user can control the # of threads which can enter
> try_to_free...
> 

What happens to the other threads, do they sleep?

> Default is 10240 ...a enough big number for unlimited. Maybe this should
> be changed.
> 
> 
> Changes from previous one.
>   - Added an interface to control the limit.
>   - don't call wake_up at uncharge()...it seems hevay..
>     Instead of that, sleepers use schedule_timeout(HZ/100). (10ms)
> 
> Considerations:
>   - Should we add this 'throttle' to global_lru at first ?

Hmm.. interesting question. I think Rik is looking into some of these issues

>   - Do we need more knobs ?
>   - Should default value to be automtically estimated value ?
>     (I used '# of cpus/4' as default in previous version.)
> 
>     
> Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu-+CUm20s59erQFUHtdCDX3A@public.gmane.org>
> 
> 
>  mm/memcontrol.c |  235 ++++++++++++++++++++++++++++++--------------------------
>  1 files changed, 127 insertions(+), 108 deletions(-)
> 
> Index: linux-2.6.24-mm1/mm/memcontrol.c
> ===================================================================
> --- linux-2.6.24-mm1.orig/mm/memcontrol.c
> +++ linux-2.6.24-mm1/mm/memcontrol.c
> @@ -145,8 +145,17 @@ struct mem_cgroup {
>  		wait_queue_head_t	waitq;
>  		struct task_struct	*kthread;
>  	} daemon;
> +	/*
> +	 * throttling params for reclaim.
> +	 */
> +	struct {
> +		int		limit;

We might want to use something else instead of limit. How about
max_parallel_reclaimers?

> +		atomic_t	reclaimers;
> +		wait_queue_head_t waitq;
> +	} throttle;
>  };
> 
> +
>  /*
>   * We use the lower bit of the page->page_cgroup pointer as a bit spin
>   * lock. We need to ensure that page->page_cgroup is atleast two
> @@ -520,6 +529,27 @@ static inline void mem_cgroup_schedule_d
>  		wake_up_interruptible(&mem->daemon.waitq);
>  }
> 
> +static inline void mem_cgroup_wait_reclaim(struct mem_cgroup *mem)
> +{
> +	DEFINE_WAIT(wait);
> +	while (1) {
> +		prepare_to_wait(&mem->throttle.waitq, &wait,
> +						TASK_INTERRUPTIBLE);
> +		if (res_counter_check_under_limit(&mem->res)) {
> +			finish_wait(&mem->throttle.waitq, &wait);
> +			break;
> +		}
> +		/* expect some progress in... */
> +		schedule_timeout(HZ/50);

Can't we wait on someone to wake us up? Why do we need to do a schedule_timeout?
And why HZ/50 and not HZ/10? Too many timers distract the system from power
management :)

> +		finish_wait(&mem->throttle.waitq, &wait);
> +	}
> +}
> +
> +static inline int mem_cgroup_throttle_reclaim(struct mem_cgroup *mem)
> +{
> +	return atomic_add_unless(&mem->throttle.reclaimers, 1,
> +					mem->throttle.limit);
> +}
> 
>  unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
>  					struct list_head *dst,
> @@ -652,11 +682,22 @@ retry:
>  	 * the cgroup limit.
>  	 */
>  	while (res_counter_charge(&mem->res, PAGE_SIZE)) {
> +		int ret;
>  		if (!(gfp_mask & __GFP_WAIT))
>  			goto out;
> 
> -		if (try_to_free_mem_cgroup_pages(mem, gfp_mask))
> +		if (((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
> +		    || mem_cgroup_throttle_reclaim(mem)) {

I think we should split this check and add detailed comments. If we cannot
sleep, then we cannot be throttled, right?

> +			ret = try_to_free_mem_cgroup_pages(mem, gfp_mask);
> +			atomic_dec(&mem->throttle.reclaimers);
> +			if (waitqueue_active(&mem->throttle.waitq))
> +				wake_up_all(&mem->throttle.waitq);
> +			if (ret)
> +				continue;
> +		} else {
> +			mem_cgroup_wait_reclaim(mem);
>  			continue;
> +		}
> 
>  		/*
>   		 * try_to_free_mem_cgroup_pages() might not give us a full
> @@ -1054,6 +1095,19 @@ static ssize_t mem_force_empty_read(stru
>  	return -EINVAL;
>  }
> 
> +static int mem_throttle_write(struct cgroup *cont, struct cftype *cft, u64 val)
> +{
> +	struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
> +	int limit = (int)val;
> +	mem->throttle.limit = limit;
> +	return 0;
> +}
> +
> +static u64 mem_throttle_read(struct cgroup *cont, struct cftype *cft)
> +{
> +	struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
> +	return (u64)mem->throttle.limit;
> +}
> 
>  static const struct mem_cgroup_stat_desc {
>  	const char *msg;
> @@ -1146,6 +1200,11 @@ static struct cftype mem_cgroup_files[] 
>  		.read = mem_force_empty_read,
>  	},
>  	{
> +		.name = "shrinks",
> +		.write_uint = mem_throttle_write,
> +		.read_uint  = mem_throttle_read,
> +	},
> +	{
>  		.name = "stat",
>  		.open = mem_control_stat_open,
>  	},
> @@ -1215,6 +1274,11 @@ mem_cgroup_create(struct cgroup_subsys *
>  			goto free_out;
>  	init_waitqueue_head(&mem->daemon.waitq);
>  	mem->daemon.kthread = NULL;
> +
> +	init_waitqueue_head(&mem->throttle.waitq);
> +	mem->throttle.limit = 10240; /* maybe enough big for no throttle */

Why 10240? So, 10,000 threads can run in parallel, isn't that an overkill? How
about setting it to number of cpus/2 or something like that?

> +	atomic_set(&mem->throttle.reclaimers, 0);
> +
>  	return &mem->css;
>  free_out:
>  	for_each_node_state(node, N_POSSIBLE)
> 


-- 
	Warm Regards,
	Balbir Singh
	Linux Technology Center
	IBM, ISTL