On Sun, 24 Oct 2010 11:44:38 -0700
Greg Thelen <gthelen@google.com> wrote:

quoted text
> KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> writes:
> 
> > Fixed one here.
> > ==
> > From: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
> >
> > Now, at calculating dirty limit, vm_dirty_param() is called.
> > This function returns dirty-limit related parameters considering
> > memory cgroup settings.
> >
> > Now, assume that vm_dirty_bytes=100M (global dirty limit) and
> > memory cgroup has 1G of pages and 40 dirty_ratio, dirtyable memory is
> > 500MB.
> >
> > In this case, global_dirty_limits will consider dirty_limt as
> > 500 *0.4 = 200MB. This is bad...memory cgroup is not back door.
> >
> > This patch limits the return value of vm_dirty_param() considring
> > global settings.
> >
> > Changelog:
> >  - fixed an argument "mem" int to u64
> >  - fixed to use global available memory to cap memcg's value.
> >
> > Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
> > ---
> >  include/linux/memcontrol.h |    5 +++--
> >  mm/memcontrol.c            |   30 +++++++++++++++++++++++++++++-
> >  mm/page-writeback.c        |    3 ++-
> >  3 files changed, 34 insertions(+), 4 deletions(-)
> >
> > Index: dirty_limit_new/mm/memcontrol.c
> > ===================================================================
> > --- dirty_limit_new.orig/mm/memcontrol.c
> > +++ dirty_limit_new/mm/memcontrol.c
> > @@ -1171,9 +1171,11 @@ static void __mem_cgroup_dirty_param(str
> >   * can be moved after our access and writeback tends to take long time.  At
> >   * least, "memcg" will not be freed while holding rcu_read_lock().
> >   */
> > -void vm_dirty_param(struct vm_dirty_param *param)
> > +void vm_dirty_param(struct vm_dirty_param *param,
> > +	 u64 mem, u64 global)
> >  {
> >  	struct mem_cgroup *memcg;
> > +	u64 limit, bglimit;
> >  
> >  	if (mem_cgroup_disabled()) {
> >  		global_vm_dirty_param(param);
> > @@ -1183,6 +1185,32 @@ void vm_dirty_param(struct vm_dirty_para
> >  	rcu_read_lock();
> >  	memcg = mem_cgroup_from_task(current);
> >  	__mem_cgroup_dirty_param(param, memcg);
> > +	/*
> > +	 * A limitation under memory cgroup is under global vm, too.
> > +	 */
> > +	if (vm_dirty_ratio)
> > +		limit = global * vm_dirty_ratio / 100;
> > +	else
> > +		limit = vm_dirty_bytes;
> > +	if (param->dirty_ratio) {
> > +		param->dirty_bytes = mem * param->dirty_ratio / 100;
> > +		param->dirty_ratio = 0;
> > +	}
> > +	if (param->dirty_bytes > limit)
> > +		param->dirty_bytes = limit;
> > +
> > +	if (dirty_background_ratio)
> > +		bglimit = global * dirty_background_ratio / 100;
> > +	else
> > +		bglimit = dirty_background_bytes;
> > +
> > +	if (param->dirty_background_ratio) {
> > +		param->dirty_background_bytes =
> > +			mem * param->dirty_background_ratio / 100;
> > +		param->dirty_background_ratio = 0;
> > +	}
> > +	if (param->dirty_background_bytes > bglimit)
> > +		param->dirty_background_bytes = bglimit;
> >  	rcu_read_unlock();
> >  }
> >  
> > Index: dirty_limit_new/include/linux/memcontrol.h
> > ===================================================================
> > --- dirty_limit_new.orig/include/linux/memcontrol.h
> > +++ dirty_limit_new/include/linux/memcontrol.h
> > @@ -171,7 +171,7 @@ static inline void mem_cgroup_dec_page_s
> >  }
> >  
> >  bool mem_cgroup_has_dirty_limit(void);
> > -void vm_dirty_param(struct vm_dirty_param *param);
> > +void vm_dirty_param(struct vm_dirty_param *param, u64 mem, u64 global);
> >  s64 mem_cgroup_page_stat(enum mem_cgroup_nr_pages_item item);
> >  
> >  unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
> > @@ -360,7 +360,8 @@ static inline bool mem_cgroup_has_dirty_
> >  	return false;
> >  }
> >  
> > -static inline void vm_dirty_param(struct vm_dirty_param *param)
> > +static inline void vm_dirty_param(struct vm_dirty_param *param,
> > +		u64 mem, u64 global)
> >  {
> >  	global_vm_dirty_param(param);
> >  }
> > Index: dirty_limit_new/mm/page-writeback.c
> > ===================================================================
> > --- dirty_limit_new.orig/mm/page-writeback.c
> > +++ dirty_limit_new/mm/page-writeback.c
> > @@ -466,7 +466,8 @@ void global_dirty_limits(unsigned long *
> >  	struct task_struct *tsk;
> >  	struct vm_dirty_param dirty_param;
> >  
> > -	vm_dirty_param(&dirty_param);
> > +	vm_dirty_param(&dirty_param,
> > +		available_memory, global_dirtyable_memory());
> >  
> >  	if (dirty_param.dirty_bytes)
> >  		dirty = DIV_ROUND_UP(dirty_param.dirty_bytes, PAGE_SIZE);
> 
> I think there is a problem with the patch above.  In the patch
> vm_dirty_param() sets param->dirty_[background_]bytes to the smallest
> limits considering the memcg and global limits.  Assuming the current
> task is in a memcg, then the memcg dirty (not system-wide) usage is
> always compared to the selected limits (which may be per-memcg or
> system).  The problem is that if:
> a) per-memcg dirty limit is smaller than system then vm_dirty_param()
>    will select per-memcg dirty limit, and
> b) per-memcg dirty usage is well below memcg dirty limit, and
> b) system usage is at system limit
> Then the above patch will not trigger writeback.  Example with two
> memcg:
>          sys
>         B   C
>       
>       limit  usage
>   sys  10     10
>    B    7      6
>    C    5      4
> 
> If B wants to grow, the system will exceed system limit of 10 and should
> be throttled.  However, the smaller limit (7) will be selected and
> applied to memcg usage (6), which indicates no need to throttle, so the
> system could get as bad as:
> 
>       limit  usage
>   sys  10     12
>    B    7      7
>    C    5      5
> 
> In this case the system usage exceeds the system limit because each
> the per-memcg checks see no per-memcg problems.
> 
nice catch !

quoted text
> To solve this I propose we create a new structure to aggregate both
> dirty limit and usage data:
> 	struct dirty_limits {
> 	       unsigned long dirty_thresh;
> 	       unsigned long background_thresh;
> 	       unsigned long nr_reclaimable;
> 	       unsigned long nr_writeback;
> 	};
> 
> global_dirty_limits() would then query both the global and memcg limits
> and dirty usage of one that is closest to its limit.  This change makes
> global_dirty_limits() look like:
> 
> void global_dirty_limits(struct dirty_limits *limits)
> {
> 	unsigned long background;
> 	unsigned long dirty;
> 	unsigned long nr_reclaimable;
> 	unsigned long nr_writeback;
> 	unsigned long available_memory = determine_dirtyable_memory();
> 	struct task_struct *tsk;
> 
> 	if (vm_dirty_bytes)
> 		dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
> 	else
> 		dirty = (vm_dirty_ratio * available_memory) / 100;
> 
> 	if (dirty_background_bytes)
> 		background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
> 	else
> 		background = (dirty_background_ratio * available_memory) / 100;
> 
> 	nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
> 				global_page_state(NR_UNSTABLE_NFS);
> 	nr_writeback = global_page_state(NR_WRITEBACK);
> 
> 	if (mem_cgroup_dirty_limits(available_memory, limits) &&
> 	    dirty_available(limits->dirty_thresh, limits->nr_reclaimable,
> 			    limits->nr_writeback) <
> 	    dirty_available(dirty, nr_reclaimable, nr_writeback)) {
> 		dirty = min(dirty, limits->dirty_thresh);
> 		background = min(background, limits->background_thresh);
> 	} else {
> 		limits->nr_reclaimable = nr_reclaimable;
> 		limits->nr_writeback = nr_writeback;
> 	}
> 
> 	if (background >= dirty)
> 		background = dirty / 2;
> 	tsk = current;
> 	if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
> 		background += background / 4;
> 		dirty += dirty / 4;
> 	}
> 	limits->background_thresh = background;
> 	limits->dirty_thresh = dirty;
> }
> 
> Because this approach considered both memcg and system limits, the
> problem described above is avoided.
> 
> I have this change integrated into the memcg dirty limit series (-v3 was
> the last post; v4 is almost ready with this change).  I will post -v4
> with this approach is there is no strong objection.
> 
I have one concern now.

I've noticed that global_dirty_limits() is called bdi_debug_stats_show(),
so the output of "cat /sys/kernel/debug/bdi/.../stats" changes depending
on the memcg where the command is executed.

Can you take it into account too ?

Thanks,
Daisuke Nishimura.
--
unsubscribe notice
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/