Re: [PATCH] memcg: cgroup fix rmdir hang

[KAMEZAWA Hiroyuki]

On Fri, 26 Jun 2009 13:16:23 -0700
Paul Menage <menage@xxxxxxxxxx> wrote:

> Hi Kamezawa,
> 
> Sorry that I didn't get a chance to look at these patches before now.
> 
> On Thu, Jun 25, 2009 at 10:10 PM, KAMEZAWA
> Hiroyuki<kamezawa.hiroyu@xxxxxxxxxxxxxx> wrote:
> > I hope this will be a final bullet..
> > I myself think this one is enough simple and good.
> > I'm sorry that we need test again.
> >
> > ==
> > From: KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx>
> >
> > After commit: cgroup: fix frequent -EBUSY at rmdir
> > Â Â Â Â Â Â Âec64f51545fffbc4cb968f0cea56341a4b07e85a
> > cgroup's rmdir (especially against memcg) doesn't return -EBUSY
> > by temporal ref counts. That commit expects all refs after pre_destroy()
> > is temporary but...it wasn't. Then, rmdir can wait permanently.
> > This patch tries to fix that and change followings.
> >
> > Â- set CGRP_WAIT_ON_RMDIR flag before pre_destroy().
> > Â- clear CGRP_WAIT_ON_RMDIR flag when the subsys finds racy case.
> > Â if there are sleeping ones, wakes them up.
> > Â- rmdir() sleeps only when CGRP_WAIT_ON_RMDIR flag is set.
> >
> > Changelog v2->v3:
> > Â- removed retry_rmdir() callback.
> > Â- make use of CGRP_WAIT_ON_RMDIR flag more.
> >
> > Reported-by: Daisuke Nishimura <nishimura@xxxxxxxxxxxxxxxxx>
> > Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx>
> > ---
> > Âinclude/linux/cgroup.h | Â 13 +++++++++++++
> > Âkernel/cgroup.c    Â|  38 ++++++++++++++++++++++----------------
> > Âmm/memcontrol.c    Â|  25 +++++++++++++++++++++++--
> > Â3 files changed, 58 insertions(+), 18 deletions(-)
> >
> > Index: mmotm-2.6.31-Jun25/include/linux/cgroup.h
> > ===================================================================
> > --- mmotm-2.6.31-Jun25.orig/include/linux/cgroup.h
> > +++ mmotm-2.6.31-Jun25/include/linux/cgroup.h
> > @@ -366,6 +366,19 @@ int cgroup_task_count(const struct cgrou
> > Âint cgroup_is_descendant(const struct cgroup *cgrp, struct task_struct *task);
> >
> > Â/*
> > + * Allow to use CGRP_WAIT_ON_RMDIR flag to check race with rmdir() for subsys.
> > + * Subsys can call this function if it's necessary to call pre_destroy() again
> > + * because it adds not-temporary refs to css after or while pre_destroy().
> > + * The caller of this function should use css_tryget(), too.
> > + */
> > +void __cgroup_wakeup_rmdir_waiters(void);
> > +static inline void cgroup_wakeup_rmdir_waiters(struct cgroup *cgrp)
> > +{
> > + Â Â Â if (unlikely(test_and_clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags)))
> > + Â Â Â Â Â Â Â __cgroup_wakeup_rmdir_waiters();
> > +}
> > +
> > +/*
> > Â* Control Group subsystem type.
> > Â* See Documentation/cgroups/cgroups.txt for details
> > Â*/
> > Index: mmotm-2.6.31-Jun25/kernel/cgroup.c
> > ===================================================================
> > --- mmotm-2.6.31-Jun25.orig/kernel/cgroup.c
> > +++ mmotm-2.6.31-Jun25/kernel/cgroup.c
> > @@ -734,14 +734,13 @@ static void cgroup_d_remove_dir(struct d
> > Â* reference to css->refcnt. In general, this refcnt is expected to goes down
> > Â* to zero, soon.
> > Â*
> > - * CGRP_WAIT_ON_RMDIR flag is modified under cgroup's inode->i_mutex;
> > + * CGRP_WAIT_ON_RMDIR flag is set under cgroup's inode->i_mutex;
> > Â*/
> > ÂDECLARE_WAIT_QUEUE_HEAD(cgroup_rmdir_waitq);
> >
> > -static void cgroup_wakeup_rmdir_waiters(const struct cgroup *cgrp)
> > +void __cgroup_wakeup_rmdir_waiters(void)
> > Â{
> 
> Maybe we should name this wakeup_rmdir_waiter() to emphasise that fact
> that there will only be one waiter (the thread doing the rmdir).
> 
Hm, but, there is no guarantee that there will "an" waiter.


> > - Â Â Â if (unlikely(test_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags)))
> > - Â Â Â Â Â Â Â wake_up_all(&cgroup_rmdir_waitq);
> > + Â Â Â wake_up_all(&cgroup_rmdir_waitq);
> > Â}
> >
> > Âstatic int rebind_subsystems(struct cgroupfs_root *root,
> > @@ -2696,33 +2695,40 @@ again:
> > Â Â Â Âmutex_unlock(&cgroup_mutex);
> >
> > Â Â Â Â/*
> > + Â Â Â Â* css_put/get is provided for subsys to grab refcnt to css. In typical
> > + Â Â Â Â* case, subsystem has no reference after pre_destroy(). But, under
> > + Â Â Â Â* hierarchy management, some *temporal* refcnt can be hold.
> 
> This sentence needs improvement/clarification. (Yes, I know it's just
> copied from later in the file, but it wasn't clear there either :-) )
> 
ok, I'll modify here. How about this ?
==
In general, subsystem has no css->refcnt after pre_destroy(). But in racy cases,
subsystem may have to get css->refcnt after pre_destroy() and it makes rmdir
return with -EBUSY. But we don't like frequent -EBUSY. To avoid that, we use
waitqueue for cgroup's rmdir. CGROUP_WAIT_ON_RMDIR bit is for synchronizing
rmdir waitqueue and subsystem behavior. Please see css_get()/put()/tryget()
implementation, it allows subsystem to avoid unnecessary failure of rmdir().
If css_put()/get()/tryget() is not enough, cgroup_wakeup_rmdir_waiter() can be
used.
==


> 
> > + Â Â Â Â* To avoid returning -EBUSY to a user, waitqueue is used. If subsys
> > + Â Â Â Â* is really busy, it should return -EBUSY at pre_destroy(). wake_up
> > + Â Â Â Â* is called when css_put() is called and refcnt goes down to 0.
> > + Â Â Â Â* And this WAIT_ON_RMDIR flag is cleared when subsys detect a race
> > + Â Â Â Â* condition under pre_destroy()->rmdir.
> 
> What exactly do you mean by pre_destroy()->rmdir ?
> 
Ah, between pre_destroy() and "cgroup is removed" state. I'll remove this.
Maybe above comment is enough.


> > If flag is cleared, we need
> > + Â Â Â Â* to call pre_destroy(), again.
> > + Â Â Â Â*/
> > + Â Â Â set_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
> > +
> > + Â Â Â /*
> > Â Â Â Â * Call pre_destroy handlers of subsys. Notify subsystems
> > Â Â Â Â * that rmdir() request comes.
> > Â Â Â Â */
> > Â Â Â Âret = cgroup_call_pre_destroy(cgrp);
> > - Â Â Â if (ret)
> > + Â Â Â if (ret) {
> > + Â Â Â Â Â Â Â clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
> > Â Â Â Â Â Â Â Âreturn ret;
> > + Â Â Â }
> >
> > Â Â Â Âmutex_lock(&cgroup_mutex);
> > Â Â Â Âparent = cgrp->parent;
> > Â Â Â Âif (atomic_read(&cgrp->count) || !list_empty(&cgrp->children)) {
> > + Â Â Â Â Â Â Â clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
> > Â Â Â Â Â Â Â Âmutex_unlock(&cgroup_mutex);
> > Â Â Â Â Â Â Â Âreturn -EBUSY;
> > Â Â Â Â}
> > - Â Â Â /*
> > - Â Â Â Â* css_put/get is provided for subsys to grab refcnt to css. In typical
> > - Â Â Â Â* case, subsystem has no reference after pre_destroy(). But, under
> > - Â Â Â Â* hierarchy management, some *temporal* refcnt can be hold.
> > - Â Â Â Â* To avoid returning -EBUSY to a user, waitqueue is used. If subsys
> > - Â Â Â Â* is really busy, it should return -EBUSY at pre_destroy(). wake_up
> > - Â Â Â Â* is called when css_put() is called and refcnt goes down to 0.
> > - Â Â Â Â*/
> > - Â Â Â set_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
> > Â Â Â Âprepare_to_wait(&cgroup_rmdir_waitq, &wait, TASK_INTERRUPTIBLE);
> > -
> > Â Â Â Âif (!cgroup_clear_css_refs(cgrp)) {
> > Â Â Â Â Â Â Â Âmutex_unlock(&cgroup_mutex);
> > - Â Â Â Â Â Â Â schedule();
> > + Â Â Â Â Â Â Â if (test_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags))
> > + Â Â Â Â Â Â Â Â Â Â Â schedule();
> 
> The need for this test_bit() should be commented, I think - at first
> sight it doesn't seem necessary since if we've already been woken up,
> then the schedule() is logically a no-op anyway. We only need it
> because we are worried about the case where someone calls
> wakeup_rmdir_waiters() prior to the prepare_to_wait()
> 
> 
yes. will add comments.


> > ===================================================================
> > --- mmotm-2.6.31-Jun25.orig/mm/memcontrol.c
> > +++ mmotm-2.6.31-Jun25/mm/memcontrol.c
> > @@ -1234,6 +1234,12 @@ static int mem_cgroup_move_account(struc
> > Â Â Â Âret = 0;
> > Âout:
> > Â Â Â Âunlock_page_cgroup(pc);
> > + Â Â Â /*
> > + Â Â Â Â* We charges against "to" which may not have any tasks. Then, "to"
> > + Â Â Â Â* can be under rmdir(). But in current implementation, caller of
> > + Â Â Â Â* this function is just force_empty() and it's garanteed that
> > + Â Â Â Â* "to" is never removed. So, we don't check rmdir status here.
> > + Â Â Â Â*/
> > Â Â Â Âreturn ret;
> > Â}
> >
> > @@ -1455,6 +1461,7 @@ __mem_cgroup_commit_charge_swapin(struct
> > Â Â Â Â Â Â Â Âreturn;
> > Â Â Â Âif (!ptr)
> > Â Â Â Â Â Â Â Âreturn;
> > + Â Â Â css_get(&ptr->css);
> > Â Â Â Âpc = lookup_page_cgroup(page);
> > Â Â Â Âmem_cgroup_lru_del_before_commit_swapcache(page);
> > Â Â Â Â__mem_cgroup_commit_charge(ptr, pc, ctype);
> > @@ -1484,7 +1491,13 @@ __mem_cgroup_commit_charge_swapin(struct
> > Â Â Â Â Â Â Â Â}
> > Â Â Â Â Â Â Â Ârcu_read_unlock();
> > Â Â Â Â}
> > - Â Â Â /* add this page(page_cgroup) to the LRU we want. */
> > + Â Â Â /*
> > + Â Â Â Â* At swapin, we may charge account against cgroup which has no tasks.
> > + Â Â Â Â* So, rmdir()->pre_destroy() can be called while we do this charge.
> > + Â Â Â Â* In that case, we need to call pre_destroy() again. check it here.
> > + Â Â Â Â*/
> > + Â Â Â cgroup_wakeup_rmdir_waiters(ptr->css.cgroup);
> > + Â Â Â css_put(&ptr->css);
> >
> > Â}
> >
> > @@ -1691,7 +1704,7 @@ void mem_cgroup_end_migration(struct mem
> >
> > Â Â Â Âif (!mem)
> > Â Â Â Â Â Â Â Âreturn;
> > -
> > + Â Â Â css_get(&mem->css);
> > Â Â Â Â/* at migration success, oldpage->mapping is NULL. */
> > Â Â Â Âif (oldpage->mapping) {
> > Â Â Â Â Â Â Â Âtarget = oldpage;
> > @@ -1731,6 +1744,14 @@ void mem_cgroup_end_migration(struct mem
> > Â Â Â Â */
> > Â Â Â Âif (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED)
> > Â Â Â Â Â Â Â Âmem_cgroup_uncharge_page(target);
> > + Â Â Â /*
> > + Â Â Â Â* At migration, we may charge account against cgroup which has no tasks
> > + Â Â Â Â* So, rmdir()->pre_destroy() can be called while we do this charge.
> > + Â Â Â Â* In that case, we need to call pre_destroy() again. check it here.
> > + Â Â Â Â*/
> > + Â Â Â cgroup_wakeup_rmdir_waiters(mem->css.cgroup);
> > + Â Â Â css_put(&mem->css);
> 
> Having to do an extra get/put purely in order to seems unfortunate -
> is that purely to force the cgroup_clear_css_refs() to fail?
> 
yes.

> Maybe we could wrap the get and the wakeup/put inside functions named
> "cgroup_exclude_rmdir()" "cgroup_release_rmdir()" so that the mm
> cgroup code is more self-explanatory.
> 
Ah, it sounds nice idea. I'll prepare that as [2/2] patch.

Thanks,
-Kame


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

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