Re: [PATCH 1/1] sched/debug: fix dentry leak in update_sched_domain_debugfs

[Greg Kroah-Hartman <gregkh@linuxfoundation.org>]

On Fri, Sep 02, 2022 at 06:47:18PM +0800, Kuyo Chang wrote:
> On Fri, 2022-09-02 at 11:11 +0200, Greg Kroah-Hartman wrote:
> > On Fri, Sep 02, 2022 at 03:36:08PM +0800, Kuyo Chang wrote:
> > > On Fri, 2022-09-02 at 08:58 +0200, Greg Kroah-Hartman wrote:
> > > > On Fri, Sep 02, 2022 at 02:40:59PM +0800, Kuyo Chang wrote:
> > > > > On Fri, 2022-09-02 at 07:26 +0200, Greg Kroah-Hartman wrote:
> > > > > > On Fri, Sep 02, 2022 at 11:15:15AM +0800, Kuyo Chang wrote:
> > > > > > > From: kuyo chang <kuyo.chang@mediatek.com>
> > > > > > > 
> > > > > > > [Syndrome]
> > > > > > > Lowmemorykiller triggered while doing hotplug stress test
> > > > > > > as
> > > > > > > below
> > > > > > > cmd:
> > > > > > > echo [0/1] > /sys/devices/system/cpu/cpu${index}/online
> > > > > > > 
> > > > > > > Rootcause:
> > > > > > > Call trace of the slab owner & usage as below after hotplug
> > > > > > > stress
> > > > > > > test(4hr).
> > > > > > > There exists dentry leak at update_sched_domain_debugfs.
> > > > > > > 
> > > > > > > Total size : 322000KB
> > > > > > > <prep_new_page+44>:
> > > > > > > <get_page_from_freelist+672>:
> > > > > > > <__alloc_pages+304>:
> > > > > > > <allocate_slab+144>:
> > > > > > > <___slab_alloc+404>:
> > > > > > > <__slab_alloc+60>:
> > > > > > > <kmem_cache_alloc+1204>:
> > > > > > > <alloc_inode+100>:
> > > > > > > <new_inode+40>:
> > > > > > > <__debugfs_create_file+172>:
> > > > > > > <update_sched_domain_debugfs+824>:
> > > > > > > <partition_sched_domains_locked+1292>:
> > > > > > > <rebuild_sched_domains_locked+576>:
> > > > > > > <cpuset_hotplug_workfn+1052>:
> > > > > > > <process_one_work+584>:
> > > > > > > <worker_thread+1008>:
> > > > > > > 
> > > > > > > [Solution]
> > > > > > > Provided by Major Chen <major.chen@samsung.com> as below
> > > > > > > link.
> > > > > > > 
> > > > > 
> > > > > 
> > > 
> > > 
> https://lore.kernel.org/lkml/20220711030341epcms5p173848e98b13c09eb2fcdf2fd7287526a@epcms5p1/
> > > > > > > update_sched_domain_debugfs() uses debugfs_lookup() to find
> > > > > > > wanted
> > > > > > > dentry(which has
> > > > > > > been created by debugfs_create_dir() before), but not call
> > > > > > > dput()
> > > > > > > to return this dentry
> > > > > > > back. This result in dentry leak even debugfs_remove() is
> > > > > > > called.
> > > > > > > 
> > > > > > > [Test result]
> > > > > > > Using below commands to check inode_cache & dentry leak.
> > > > > > > cat /proc/slabinfo | grep -w inode_cache
> > > > > > > cat /proc/slabinfo | grep -w dentry
> > > > > > > 
> > > > > > > With the patch, the inode_cache & dentry stays consistent
> > > > > > > so the lowmemorykiller will not triggered anymore.
> > > > > > > 
> > > > > > > Fixes: 8a99b6833c88 ("sched: Move SCHED_DEBUG sysctl to
> > > > > > > debugfs")
> > > > > > > 
> > > > > > > Signed-off-by: Major Chen <major.chen@samsung.com>
> > > > > > > Signed-off-by: kuyo chang <kuyo.chang@mediatek.com>
> > > > > > > Tested-by: kuyo chang <kuyo.chang@mediatek.com>
> > > > > > > 
> > > > > > > ---
> > > > > > >  kernel/sched/debug.c | 7 +++++--
> > > > > > >  1 file changed, 5 insertions(+), 2 deletions(-)
> > > > > > > 
> > > > > > > diff --git a/kernel/sched/debug.c b/kernel/sched/debug.c
> > > > > > > index bb3d63bdf4ae..4ffea2dc01da 100644
> > > > > > > --- a/kernel/sched/debug.c
> > > > > > > +++ b/kernel/sched/debug.c
> > > > > > > @@ -412,11 +412,14 @@ void
> > > > > > > update_sched_domain_debugfs(void)
> > > > > > >  
> > > > > > >  	for_each_cpu(cpu, sd_sysctl_cpus) {
> > > > > > >  		struct sched_domain *sd;
> > > > > > > -		struct dentry *d_cpu;
> > > > > > > +		struct dentry *d_cpu, *d_lookup;
> > > > > > >  		char buf[32];
> > > > > > >  
> > > > > > >  		snprintf(buf, sizeof(buf), "cpu%d", cpu);
> > > > > > > -		debugfs_remove(debugfs_lookup(buf, sd_dentry));
> > > > > > > +		d_lookup = debugfs_lookup(buf, sd_dentry);
> > > > > > > +		debugfs_remove(d_lookup);
> > > > > > > +		if (!IS_ERR_OR_NULL(d_lookup))
> > > > > > > +			dput(d_lookup);
> > > > > > 
> > > > > > That's odd, and means that something else is removing this
> > > > > > file
> > > > > > right
> > > > > > after we looked it up?  Is there a missing lock here that
> > > > > > should
> > > > > > be
> > > > > > used
> > > > > > instead?
> > > > > > 
> > > > > > thanks,
> > > > > > 
> > > > > > greg k-h
> > > > > 
> > > > > 
> > > > > While doing cpu hotlug, the cpu_active_mask is changed, 
> > > > > so it need to update_sched_domain_debugfs.
> > > > > 
> > > > > The original design is to recreate sd_dentry, so it doing
> > > > > debugfs_remove and then debugfs_create_dir.
> > > > > However, by debugfs_lookup function usage.
> > > > > The returned dentry must be passed to dput() when it is no
> > > > > longer
> > > > > needed to avoid dentry leak.
> > > > 
> > > > Eeeek, nice find!  I've been adding this pattern:
> > > > 	debugfs_remove(debugfs_lookup(...));
> > > > all over the place, and as you point out, that's wrong!
> > > > 
> > > > It's as if I didn't even read the documentation I wrote.
> > > > 
> > > > {sigh}
> > > > 
> > > > Ok, as this is going to be a very common pattern, how about we
> > > > create:
> > > > 	debugfs_lookup_and_remove()
> > > > function that does the above logic all in one place and then we
> > > > don't
> > > > have to put that logic everywhere in the kernel.  My goal is for
> > > > users
> > > > of debugfs to never have to worry about anything about 'struct
> > > > dentry'
> > > > at all, and I really failed that goal here in a major way.
> > > > 
> > > > I can work on that this afternoon after I get some other things
> > > > done,
> > > > unless you want to do it now?
> > > > 
> > > > Again, very nice find, thank you for this.
> > > > 
> > > 
> > > Thanks for your kindly support !
> > > Please help to add debugfs_lookup_and_remove() and then we can use
> > > this api to fix this denrty leak issue.
> > 
> > Ok, can you test the patch below with your reproducer to see if it is
> > fixed or not?  If this resolves the issue, I'll split it up into the
> > proper set of patches and sweep the whole tree as every user of
> > debugfs_lookup() that I've found so far got it wrong :(
> > 
> > thanks again!
> > 
> > greg k-h
> > 
> > diff --git a/fs/debugfs/inode.c b/fs/debugfs/inode.c
> > index 3dcf0b8b4e93..87ccd6280a9b 100644
> > --- a/fs/debugfs/inode.c
> > +++ b/fs/debugfs/inode.c
> > @@ -744,6 +744,28 @@ void debugfs_remove(struct dentry *dentry)
> >  }
> >  EXPORT_SYMBOL_GPL(debugfs_remove);
> >  
> > +/**
> > + * debugfs_lookup_and_remove - lookup a directory or file and
> > recursively remove it
> > + * @name: a pointer to a string containing the name of the item to
> > look up.
> > + * @parent: a pointer to the parent dentry of the item.
> > + *
> > + * This is the equlivant of doing something like
> > + * debugfs_remove(debugfs_lookup(..)) but with the proper reference
> > counting
> > + * handled for the directory being looked up.
> > + */
> > +void debugfs_lookup_and_remove(const char *name, struct dentry
> > *parent)
> > +{
> > +	struct dentry *dentry;
> > +
> > +	dentry = debugfs_lookup(name, parent);
> > +	if (IS_ERR_OR_NULL(dentry))
> > +		return;
> > +
> > +	debugfs_remove(dentry);
> > +	dput(dentry);
> > +}
> > +EXPORT_SYMBOL_GPL(debugfs_lookup_and_remove);
> > +
> >  /**
> >   * debugfs_rename - rename a file/directory in the debugfs
> > filesystem
> >   * @old_dir: a pointer to the parent dentry for the renamed object.
> > This
> > diff --git a/include/linux/debugfs.h b/include/linux/debugfs.h
> > index c869f1e73d75..f60674692d36 100644
> > --- a/include/linux/debugfs.h
> > +++ b/include/linux/debugfs.h
> > @@ -91,6 +91,8 @@ struct dentry *debugfs_create_automount(const char
> > *name,
> >  void debugfs_remove(struct dentry *dentry);
> >  #define debugfs_remove_recursive debugfs_remove
> >  
> > +void debugfs_lookup_and_remove(const char *name, struct dentry
> > *parent);
> > +
> >  const struct file_operations *debugfs_real_fops(const struct file
> > *filp);
> >  
> >  int debugfs_file_get(struct dentry *dentry);
> > @@ -225,6 +227,10 @@ static inline void debugfs_remove(struct dentry
> > *dentry)
> >  static inline void debugfs_remove_recursive(struct dentry *dentry)
> >  { }
> >  
> > +static inline void debugfs_lookup_and_remove(const char *name,
> > +					     struct dentry *parent)
> > +{ }
> > +
> >  const struct file_operations *debugfs_real_fops(const struct file
> > *filp);
> >  
> >  static inline int debugfs_file_get(struct dentry *dentry)
> > diff --git a/kernel/sched/debug.c b/kernel/sched/debug.c
> > index bb3d63bdf4ae..667876da8382 100644
> > --- a/kernel/sched/debug.c
> > +++ b/kernel/sched/debug.c
> > @@ -416,7 +416,7 @@ void update_sched_domain_debugfs(void)
> >  		char buf[32];
> >  
> >  		snprintf(buf, sizeof(buf), "cpu%d", cpu);
> > -		debugfs_remove(debugfs_lookup(buf, sd_dentry));
> > +		debugfs_lookup_and_remove(buf, sd_dentry);
> >  		d_cpu = debugfs_create_dir(buf, sd_dentry);
> >  
> >  		i = 0;
> 
> The lowmemorykiller triggered issue while doing hotplug stress is fixed
> after applying your patch.
> Thanks for your kindly support !

Wonderful!  I added your Tested-by: to this in a 2 part patch series
that is now at:
	https://lore.kernel.org/r/20220902123107.109274-1-gregkh@linuxfoundation.org

thanks,

greg k-h