lkml.org 
[lkml]   [2009]   [Jan]   [15]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
SubjectRe: [RFC][PATCH] memcg: get/put parents at create/free
On Thu, 15 Jan 2009 16:54:53 +0900, KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> wrote:
> On Thu, 15 Jan 2009 16:45:37 +0900
> Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp> wrote:
>
> > On Thu, 15 Jan 2009 13:38:14 +0900, Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp> wrote:
> > > On Thu, 15 Jan 2009 11:14:20 +0900, Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp> wrote:
> > > > > > > To handle the problem "parent may be obsolete",
> > > > > > >
> > > > > > > call mem_cgroup_get(parent) at create()
> > > > > > > call mem_cgroup_put(parent) at freeing memcg.
> > > > > > > (regardless of use_hierarchy.)
> > > > > > >
> > > > > > > is clearer way to go, I think.
> > > > > > >
> > > > > > > I wonder whether there is mis-accounting problem or not..
> > > > > > >
> > > hmm, after more consideration, although this patch can prevent the BUG,
> > > it can leak memsw accounting of parents because memsw of parents, which
> > > have been incremented by charge, does not decremented.
> > >
> > > I'll try pet/put parent approach..
> > > Or any other good ideas ?
> > >
> > I attach a tryial patch.
> >
> > It has been working fine so far(for about 1 hour).
> >
> > Thanks,
> > Daisuke Nishimura.
> > ===
> > From: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
> >
> > mem_cgroup_get ensures that the memcg that has been got can be accessed
> > even after the directory has been removed, but it doesn't ensure that parents
> > of it can be accessed: parents might have been freed already by rmdir.
> >
> > This causes a bug in case of use_hierarchy==1, because res_counter_uncharge
> > climb up the tree.
> >
> > This patch tries to fix this probrem by getting parents at create, and
> > putting them at freeing.
> >
> > Signed-off-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
> > ---
> > mm/memcontrol.c | 33 ++++++++++++++++++++++++++++++++-
> > 1 files changed, 32 insertions(+), 1 deletions(-)
> >
> > diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> > index fb62b43..b4aed07 100644
> > --- a/mm/memcontrol.c
> > +++ b/mm/memcontrol.c
> > @@ -202,6 +202,8 @@ pcg_default_flags[NR_CHARGE_TYPE] = {
> >
> > static void mem_cgroup_get(struct mem_cgroup *mem);
> > static void mem_cgroup_put(struct mem_cgroup *mem);
> > +static void mem_cgroup_get_parents(struct mem_cgroup *mem);
> > +static void mem_cgroup_put_parents(struct mem_cgroup *mem);
> >
> > static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
> > struct page_cgroup *pc,
> > @@ -2185,10 +2187,38 @@ static void mem_cgroup_get(struct mem_cgroup *mem)
> >
> > static void mem_cgroup_put(struct mem_cgroup *mem)
> > {
> > - if (atomic_dec_and_test(&mem->refcnt))
> > + if (atomic_dec_and_test(&mem->refcnt)) {
> > + mem_cgroup_put_parents(mem);
> > __mem_cgroup_free(mem);
> > + }
> > +}
> > +
> > +static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
> > +{
> > + if (!mem->res.parent)
> > + return NULL;
> > + return mem_cgroup_from_res_counter(mem->res.parent, res);
> > +}
> > +
> > +static void mem_cgroup_get_parents(struct mem_cgroup *mem)
> > +{
> > + struct mem_cgroup *parent = parent_mem_cgroup(mem);
> > +
> > + while (parent) {
> > + mem_cgroup_get(parent);
> > + parent = parent_mem_cgroup(parent);
> > + }
> > }
> >
>
> does we have to add refcnt to all ancestors ?
>
Ah, no need.

Just ensureing "a parent memcg is not freed while it has child memcg unfreed"
would be enough, because a parent is a child of parent of the parent.

This is the updated version.

Thanks,
Daisuke Nishimura.
===
From: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>

mem_cgroup_get ensures that the memcg that has been got can be accessed
even after the directory has been removed, but it doesn't ensure that parents
of it can be accessed: parents might have been freed already by rmdir.

This causes a bug in case of use_hierarchy==1, because res_counter_uncharge
climb up the tree.

This patch tries to fix this probrem by getting the parent at create, and
putting it at freeing.

Signed-off-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
---
mm/memcontrol.c | 29 ++++++++++++++++++++++++++++-
1 files changed, 28 insertions(+), 1 deletions(-)

diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index fb62b43..a80ba68 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -202,6 +202,8 @@ pcg_default_flags[NR_CHARGE_TYPE] = {

static void mem_cgroup_get(struct mem_cgroup *mem);
static void mem_cgroup_put(struct mem_cgroup *mem);
+static void mem_cgroup_get_parent(struct mem_cgroup *mem);
+static void mem_cgroup_put_parent(struct mem_cgroup *mem);

static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
struct page_cgroup *pc,
@@ -2185,10 +2187,34 @@ static void mem_cgroup_get(struct mem_cgroup *mem)

static void mem_cgroup_put(struct mem_cgroup *mem)
{
- if (atomic_dec_and_test(&mem->refcnt))
+ if (atomic_dec_and_test(&mem->refcnt)) {
+ mem_cgroup_put_parent(mem);
__mem_cgroup_free(mem);
+ }
+}
+
+static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
+{
+ if (!mem->res.parent)
+ return NULL;
+ return mem_cgroup_from_res_counter(mem->res.parent, res);
+}
+
+static void mem_cgroup_get_parent(struct mem_cgroup *mem)
+{
+ struct mem_cgroup *parent = parent_mem_cgroup(mem);
+
+ if (parent)
+ mem_cgroup_get(parent);
}

+static void mem_cgroup_put_parent(struct mem_cgroup *mem)
+{
+ struct mem_cgroup *parent = parent_mem_cgroup(mem);
+
+ if (parent)
+ mem_cgroup_put(parent);
+}

#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
static void __init enable_swap_cgroup(void)
@@ -2237,6 +2263,7 @@ mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
if (parent)
mem->swappiness = get_swappiness(parent);
atomic_set(&mem->refcnt, 1);
+ mem_cgroup_get_parent(mem);
return &mem->css;
free_out:
__mem_cgroup_free(mem);

\
 
 \ /
  Last update: 2009-01-15 09:21    [W:1.685 / U:0.036 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site