Re: [PATCH v3 3/4] mm/page_owner: Print memcg information

[Michal Hocko <mhocko@suse.com>]

On Tue 01-02-22 12:04:37, Waiman Long wrote:
> On 2/1/22 05:54, Michal Hocko wrote:
> > On Mon 31-01-22 14:23:07, Waiman Long wrote:
> > > It was found that a number of offlined memcgs were not freed because
> > > they were pinned by some charged pages that were present. Even "echo
> > > 1 > /proc/sys/vm/drop_caches" wasn't able to free those pages. These
> > > offlined but not freed memcgs tend to increase in number over time with
> > > the side effect that percpu memory consumption as shown in /proc/meminfo
> > > also increases over time.
> > > 
> > > In order to find out more information about those pages that pin
> > > offlined memcgs, the page_owner feature is extended to print memory
> > > cgroup information especially whether the cgroup is offlined or not.
> > > 
> > > Signed-off-by: Waiman Long <longman@redhat.com>
> > > Acked-by: David Rientjes <rientjes@google.com>
> > > ---
> > >   mm/page_owner.c | 39 +++++++++++++++++++++++++++++++++++++++
> > >   1 file changed, 39 insertions(+)
> > > 
> > > diff --git a/mm/page_owner.c b/mm/page_owner.c
> > > index 28dac73e0542..a471c74c7fe0 100644
> > > --- a/mm/page_owner.c
> > > +++ b/mm/page_owner.c
> > > @@ -10,6 +10,7 @@
> > >   #include <linux/migrate.h>
> > >   #include <linux/stackdepot.h>
> > >   #include <linux/seq_file.h>
> > > +#include <linux/memcontrol.h>
> > >   #include <linux/sched/clock.h>
> > >   #include "internal.h"
> > > @@ -325,6 +326,42 @@ void pagetypeinfo_showmixedcount_print(struct seq_file *m,
> > >   	seq_putc(m, '\n');
> > >   }
> > > +#ifdef CONFIG_MEMCG
> > > +/*
> > > + * Looking for memcg information and print it out
> > > + */
> > > +static inline void print_page_owner_memcg(char *kbuf, size_t count, int *pret,
> > > +					  struct page *page)
> > > +{
> > > +	unsigned long memcg_data = READ_ONCE(page->memcg_data);
> > > +	struct mem_cgroup *memcg;
> > > +	bool onlined;
> > > +	char name[80];
> > > +
> > > +	if (!memcg_data)
> > > +		return;
> > > +
> > > +	if (memcg_data & MEMCG_DATA_OBJCGS)
> > > +		*pret += scnprintf(kbuf + *pret, count - *pret,
> > > +				"Slab cache page\n");
> > > +
> > > +	memcg = page_memcg_check(page);
> > > +	if (!memcg)
> > > +		return;
> > > +
> > > +	onlined = (memcg->css.flags & CSS_ONLINE);
> > > +	cgroup_name(memcg->css.cgroup, name, sizeof(name));
> > > +	*pret += scnprintf(kbuf + *pret, count - *pret,
> > > +			"Charged %sto %smemcg %s\n",
> > > +			PageMemcgKmem(page) ? "(via objcg) " : "",
> > > +			onlined ? "" : "offlined ",
> > > +			name);
> > I have asked in the previous version already but what makes the memcg
> > stable (why it cannot go away and be reallocated for something else)
> > while you are trying to get its name?
> 
> The memcg is not going away as long as the page isn't freed unless if it is
> indirectly connected via objcg. Of course, there can be a race between the
> page is going to be freed while the page_owner information is being
> displayed.

Right. And that means that cgtoup_name can go off the rail and wander
through memory correct?

> One solution is to add a simple bit lock to each of the
> page_owner structure and acquire the lock when it is being written to or
> read from.

I do not really see how a bit lock could prevent memcg from going away.
On the other hand I think RCU read lock should be sufficient to keep the
memcg from going away completely.

> Anyway a lot of these debugging aids or tools don't eliminate all
> the race conditions that affect the accuracy of the displayed information. I
> can add a patch to eliminate this direct memcg race if you think this is
> necessary.

I do not mind inaccurate information. That is natural but reading
through a freed memory can be really harmfull. So this really need to be
sorted out.
-- 
Michal Hocko
SUSE Labs