lkml.org 
[lkml]   [2009]   [Apr]   [6]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [RFC][PATCH 4/9] soft limit queue and priority
On Mon, 6 Apr 2009 16:35:34 +0530
Balbir Singh <balbir@linux.vnet.ibm.com> wrote:

> * KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> [2009-04-03 17:12:48]:
>
> > From: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
> >
> > Softlimitq. for memcg.
> >
> > Implements an array of queue to list memcgs, array index is determined by
> > the amount of memory usage excess the soft limit.
> >
> > While Balbir's one uses RB-tree and my old one used a per-zone queue
> > (with round-robin), this is one of mixture of them.
> > (I'd like to use rotation of queue in later patches)
> >
> > Priority is determined by following.
> > Assume unit = total pages/1024. (the code uses different value)
> > if excess is...
> > < unit, priority = 0,
> > < unit*2, priority = 1,
> > < unit*2*2, priority = 2,
> > ...
> > < unit*2^9, priority = 9,
> > < unit*2^10, priority = 10, (> 50% to total mem)
> >
> > This patch just includes queue management part and not includes
> > selection logic from queue. Some trick will be used for selecting victims at
> > soft limit in efficient way.
> >
> > And this equips 2 queues, for anon and file. Inset/Delete of both list is
> > done at once but scan will be independent. (These 2 queues are used later.)
> >
> > Major difference from Balbir's one other than RB-tree is bahavior under
> > hierarchy. This one adds all children to queue by checking hierarchical
> > priority. This is for helping per-zone usage check on victim-selection logic.
> >
> > Changelog: v1->v2
> > - fixed comments.
> > - change base size to exponent.
> > - some micro optimization to reduce code size.
> > - considering memory hotplug, it's not good to record a value calculated
> > from totalram_pages at boot and using it later is bad manner. Fixed it.
> > - removed soft_limit_lock (spinlock)
> > - added soft_limit_update counter for avoiding mulptiple update at once.
> >
> >
> > Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
> > ---
> > mm/memcontrol.c | 118 +++++++++++++++++++++++++++++++++++++++++++++++++++++++-
> > 1 file changed, 117 insertions(+), 1 deletion(-)
> >
> > Index: softlimit-test2/mm/memcontrol.c
> > ===================================================================
> > --- softlimit-test2.orig/mm/memcontrol.c
> > +++ softlimit-test2/mm/memcontrol.c
> > @@ -192,7 +192,14 @@ struct mem_cgroup {
> > atomic_t refcnt;
> >
> > unsigned int swappiness;
> > -
> > + /*
> > + * For soft limit.
> > + */
> > + int soft_limit_priority;
> > + struct list_head soft_limit_list[2];
> > +#define SL_ANON (0)
> > +#define SL_FILE (1)
>
> Comments for the #define please.
>
Sure.

> > + atomic_t soft_limit_update;
> > /*
> > * statistics. This must be placed at the end of memcg.
> > */
> > @@ -938,11 +945,115 @@ static bool mem_cgroup_soft_limit_check(
> > return ret;
> > }
> >
> > +/*
> > + * Assume "base_amount", and excess = usage - soft limit.
> > + *
> > + * 0...... if excess < base_amount
> > + * 1...... if excess < base_amount * 2
> > + * 2...... if excess < base_amount * 2^2
> > + * 3.......if excess < base_amount * 2^3
> > + * ....
> > + * 9.......if excess < base_amount * 2^9
> > + * 10 .....if excess < base_amount * 2^10
> > + *
> > + * base_amount is detemined from total pages in the system.
> > + */
> > +
> > +#define SLQ_MAXPRIO (11)
> > +static struct {
> > + spinlock_t lock;
> > + struct list_head queue[SLQ_MAXPRIO][2]; /* 0:anon 1:file */
> > +} softlimitq;
> > +
> > +#define SLQ_PRIO_FACTOR (1024) /* 2^10 */
> > +
> > +static int __calc_soft_limit_prio(unsigned long excess)
> > +{
> > + unsigned long factor = totalram_pages /SLQ_PRIO_FACTOR;
>
> I would prefer to use global_lru_pages()
>
Hmm, ok.

Thanks,
-Kame



\
 
 \ /
  Last update: 2009-04-07 01:59    [from the cache]
©2003-2011 Jasper Spaans