Re: [PATCH v2] mm: shrinker: avoid memleak in alloc_shrinker_info

[Qi Zheng <zhengqi.arch@bytedance.com>]

On 2024/10/16 22:22, Muchun Song wrote:
> 
> 
>> On Oct 16, 2024, at 20:13, Vlastimil Babka <vbabka@suse.cz> wrote:
>>
>> On 10/14/24 11:20, Muchun Song wrote:
>>>
>>>
>>>>> On Oct 14, 2024, at 17:04, chenridong <chenridong@huawei.com> wrote:
>>>>
>>>>
>>>>
>>>> On 2024/10/14 16:43, Muchun Song wrote:
>>>>>> On Oct 14, 2024, at 16:13, Anshuman Khandual <anshuman.khandual@arm.com> wrote:
>>>>>>
>>>>>>
>>>>>>
>>>>>> On 10/14/24 08:53, Chen Ridong wrote:
>>>>>>> From: Chen Ridong <chenridong@huawei.com>
>>>>>>>
>>>>>>> A memleak was found as bellow:
>>>>>>>
>>>>>>> unreferenced object 0xffff8881010d2a80 (size 32):
>>>>>>> comm "mkdir", pid 1559, jiffies 4294932666
>>>>>>> hex dump (first 32 bytes):
>>>>>>>    00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00  ................
>>>>>>>    40 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00  @...............
>>>>>>> backtrace (crc 2e7ef6fa):
>>>>>>>    [<ffffffff81372754>] __kmalloc_node_noprof+0x394/0x470
>>>>>>>    [<ffffffff813024ab>] alloc_shrinker_info+0x7b/0x1a0
>>>>>>>    [<ffffffff813b526a>] mem_cgroup_css_online+0x11a/0x3b0
>>>>>>>    [<ffffffff81198dd9>] online_css+0x29/0xa0
>>>>>>>    [<ffffffff811a243d>] cgroup_apply_control_enable+0x20d/0x360
>>>>>>>    [<ffffffff811a5728>] cgroup_mkdir+0x168/0x5f0
>>>>>>>    [<ffffffff8148543e>] kernfs_iop_mkdir+0x5e/0x90
>>>>>>>    [<ffffffff813dbb24>] vfs_mkdir+0x144/0x220
>>>>>>>    [<ffffffff813e1c97>] do_mkdirat+0x87/0x130
>>>>>>>    [<ffffffff813e1de9>] __x64_sys_mkdir+0x49/0x70
>>>>>>>    [<ffffffff81f8c928>] do_syscall_64+0x68/0x140
>>>>>>>    [<ffffffff8200012f>] entry_SYSCALL_64_after_hwframe+0x76/0x7e
>>>>>>>
>>>>>>> In the alloc_shrinker_info function, when shrinker_unit_alloc return
>>>>>>> err, the info won't be freed. Just fix it.
>>>>>>>
>>>>>>> Fixes: 307bececcd12 ("mm: shrinker: add a secondary array for shrinker_info::{map, nr_deferred}")
>>>>>>> Signed-off-by: Chen Ridong <chenridong@huawei.com>
>>>>>>> ---
>>>>>>> mm/shrinker.c | 1 +
>>>>>>> 1 file changed, 1 insertion(+)
>>>>>>>
>>>>>>> diff --git a/mm/shrinker.c b/mm/shrinker.c
>>>>>>> index dc5d2a6fcfc4..92270413190d 100644
>>>>>>> --- a/mm/shrinker.c
>>>>>>> +++ b/mm/shrinker.c
>>>>>>> @@ -97,6 +97,7 @@ int alloc_shrinker_info(struct mem_cgroup *memcg)
>>>>>>>
>>>>>>> err:
>>>>>>> mutex_unlock(&shrinker_mutex);
>>>>>>> + kvfree(info);
>>>>>>> free_shrinker_info(memcg);
>>>>>>> return -ENOMEM;
>>>>>>> }
>>>>>>
>>>>>> There are two scenarios when "goto err:" gets called
>>>>>>
>>>>>> - When shrinker_info allocations fails, no kvfree() is required
>>>>>> - but after this change kvfree() would be called even
>>>>>> when the allocation had failed originally, which does
>>>>>>    not sound right
>>>>> Yes. In this case, @info is NULL and kvfree could handle NULL.
>>>>> It seems strange but the final behaviour correct.
>>>>>>
>>>>>> - shrinker_unit_alloc() fails, kvfree() is actually required
>>>>>>
>>>>>> I guess kvfree() should be called just after shrinker_unit_alloc()
>>>>>> fails but before calling into "goto err".
>>>>> We could do it like this, which avoids ambiguity (if someone ignores
>>>>> that kvfree could handle NULL). Something like:
>>>>> --- a/mm/shrinker.c
>>>>> +++ b/mm/shrinker.c
>>>>> @@ -88,13 +88,14 @@ int alloc_shrinker_info(struct mem_cgroup *memcg)
>>>>>                         goto err;
>>>>>                 info->map_nr_max = shrinker_nr_max;
>>>>>                 if (shrinker_unit_alloc(info, NULL, nid))
>>>>> -                       goto err;
>>>>> +                       goto free;
>>>>>                 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_info, info);
>>>>>         }
>>>>>         mutex_unlock(&shrinker_mutex);
>>>>>         return ret;
>>>>> -
>>>>> +free:
>>>>> +       kvfree(info);
>>>>> err:
>>>>>         mutex_unlock(&shrinker_mutex);
>>>>>         free_shrinker_info(memcg);
>>>>> Thanks.
>>>>>>
>>>>>> But curious, should not both kvzalloc_node()/kvfree() be avoided
>>>>>> while inside mutex lock to avoid possible lockdep issues ?
>>>> How about:
>>>>
>>>> diff --git a/mm/shrinker.c b/mm/shrinker.c
>>>> index dc5d2a6fcfc4..7baee7f00497 100644
>>>> --- a/mm/shrinker.c
>>>> +++ b/mm/shrinker.c
>>>> @@ -87,9 +87,9 @@ int alloc_shrinker_info(struct mem_cgroup *memcg)
>>>>                 if (!info)
>>>>                         goto err;
>>>>                 info->map_nr_max = shrinker_nr_max;
>>>> +               rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_info, info);
>>>>                 if (shrinker_unit_alloc(info, NULL, nid))
>>>>                         goto err;
>>>> -               rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_info, info);
>>>>         }
>>>>         mutex_unlock(&shrinker_mutex);
>>>
>>> No. We should make sure the @info is fully initialized before others
>>> could see it. That's why rcu_assign_pointer is used here.
>>
>> If the info is immediately visible, is the failure cleanup
>> free_shrinker_info() safe? It uses kvfree(info) and not kvfree_rcu(), and
>> shrinker_unit_free() is also doing kfree().
> 
> Qi told me that the @info will not visible immediately yesterday.
> So non-rcu-based kvfree is safe. Even if this fix could

Yes, alloc_shrinker_info() is only called by mem_cgroup_css_online(). At
this time, the memcg is not online yet, so it is not visible to 
shrink_slab(). And free_shrinker_info() is also called by 
mem_cgroup_css_free(),
the memcg has already been offline. The shrinker_unit_free() is
also called by expand_one_shrinker_info(), but the shrinker_info 'new'
is also not visible at that time. So non-rcu-based kvfree is safe.

> work properly, it’s a bit strange for me to use
> rcu_assign_pointer to assign the @info without full initialization to it.

Agree.

> 
> Muchun,
> Thanks.
> 
>>
>>>>
>>>> I think this is concise.
>>>>
>>>> Best regards,
>>>> Ridong
>>>
>>>
>>