[Devel] [PATCH rh7 v2] cgroup: rework reference acquisition for cgroup_find_inode

[Kirill Tkhai]

On 12.11.2020 18:08, Andrey Zhadchenko wrote:
> On Thu, 12 Nov 2020 17:29:21 +0300
> Kirill Tkhai <ktkhai at virtuozzo.com> wrote:
> 
> Hi, Kirill,
> 
>> Hi, Andrey,
>>
>> On 11.11.2020 10:32, Andrey Zhadchenko wrote:
>>> Use more generic igrab instead of atomic inc. Move cgroup_hash_del
>>> to eviction stage to avoid deadlock.
>>>
>>> Signed-off-by: Andrey Zhadchenko <andrey.zhadchenko at virtuozzo.com>
>>> ---
>>>
>>> v2: adjusted function call order in cgroup_evict_inode to match
>>> existing code
>>>
>>>  kernel/cgroup.c | 25 ++++++++-----------------
>>>  1 file changed, 8 insertions(+), 17 deletions(-)
>>>
>>> diff --git a/kernel/cgroup.c b/kernel/cgroup.c
>>> index 27d7a5e..8c2cef8 100644
>>> --- a/kernel/cgroup.c
>>> +++ b/kernel/cgroup.c
>>> @@ -1522,21 +1522,10 @@ static struct inode
>>> *cgroup_find_inode(unsigned long fh[2], char take_ref) struct inode
>>> *ret = NULL; 
>>>  	spin_lock(&cgroup_inode_table_lock);
>>> -	item = cgroup_find_item_no_lock(fh);
>>>  
>>> -	/*
>>> -	 * If we need to increase refcount, we should be aware of
>>> possible
>>> -	 * deadlock. Another thread may have started deleting this
>>> inode:
>>> -	 * iput->iput_final->cgroup_delete_inode->cgroup_hash_del
>>> -	 * If we just call igrab, it will try to take i_lock and
>>> this will
>>> -	 * result in deadlock, because deleting thread has already
>>> taken it
>>> -	 * and waits on cgroup_inode_table_lock to find inode in
>>> hashtable.
>>> -	 *
>>> -	 * If i_count is zero, someone is deleting it -> skip.
>>> -	 */
>>> -	if (take_ref && item)
>>> -		if (!atomic_inc_not_zero(&item->inode->i_count))
>>> -			item = NULL;
>>> +	item = cgroup_find_item_no_lock(fh);
>>> +	if (item && take_ref && !igrab(item->inode))
>>> +		item = NULL;  
>>
>> Here you call igrab() under cgroup_inode_table_lock, so the check for
>> (inode->i_state & I_FREEING|I_WILL_FREE) is done under that lock and
>> i_lock.
>>
>> But clear_inode() sets "inode->i_state = I_FREEING | I_CLEAR" without
>> any lock. This place does not look obvious. Isn't there some problem? 
>>
> 
> iput_final() sets inode to I_FREEING under i_lock, release it and calls
> evict(), which calls cgroup_evict_inode() via
> inode->i_sb->s_op->evict().

What is about inodes in LRU?

Say, iput_final() moves inode to LRU with zero counter. I_FREEING is not
set there. Later prune_icache_sb() operates with that inode with zero counter.

Doesn't dispose_list()->evict()->cgroup_evict_inode() race with cgroup_find_inode()?

> If I get it right, this seems to be fine. As soon as iput_final starts
> working igrab shouldn't be able to take ref on this inode.
> 
>>>  
>>>  	spin_unlock(&cgroup_inode_table_lock);
>>>  
>>> @@ -1634,15 +1623,17 @@ static const struct export_operations
>>> cgroup_export_ops = { .fh_to_dentry	= cgroup_fh_to_dentry,
>>>  };
>>>  
>>> -static int cgroup_delete_inode(struct inode *inode)
>>> +static void cgroup_evict_inode(struct inode *inode)
>>>  {
>>> +	truncate_inode_pages_final(&inode->i_data);
>>> +	clear_inode(inode);
>>>  	cgroup_hash_del(inode);
>>> -	return generic_delete_inode(inode);
>>>  }
>>>  
>>>  static const struct super_operations cgroup_ops = {
>>>  	.statfs = simple_statfs,
>>> -	.drop_inode = cgroup_delete_inode,
>>> +	.drop_inode = generic_delete_inode,
>>> +	.evict_inode = cgroup_evict_inode,
>>>  	.show_options = cgroup_show_options,
>>>  #ifdef CONFIG_VE
>>>  	.show_path = cgroup_show_path,
>>>   
>>
>