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

[Chen Ridong <chenridong@huaweicloud.com>]

On 2024/10/17 1:31, Roman Gushchin wrote:
> On Wed, Oct 16, 2024 at 07:02:23PM +0200, Vlastimil Babka wrote:
>> On 10/16/24 16:08, Muchun Song wrote:
>>>
>>>
>>>> On Oct 16, 2024, at 19:43, Vlastimil Babka <vbabka@suse.cz> wrote:
>>>> On 10/16/24 04:21, Muchun Song wrote:
>>>>>
>>>>>
>>>>>> On Oct 16, 2024, at 09:25, chenridong <chenridong@huawei.com> wrote:
>>>>>> On 2024/10/15 14:55, Anshuman Khandual wrote:
>>>>>>> On 10/14/24 16:59, Kirill A. Shutemov wrote:
>>>>>>>> On Mon, Oct 14, 2024 at 03:23:36AM +0000, 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;
>>>>>>>>> }
>>>>>>>> NAK. If in the future there going to one more error case after
>>>>>>>> rcu_assign_pointer() we will end up with double free.
>>>>>>>> This should be safer:
>>>>>>>> diff --git a/mm/shrinker.c b/mm/shrinker.c
>>>>>>>> index dc5d2a6fcfc4..763fd556bc7d 100644
>>>>>>>> --- a/mm/shrinker.c
>>>>>>>> +++ b/mm/shrinker.c
>>>>>>>> @@ -87,8 +87,10 @@ int alloc_shrinker_info(struct mem_cgroup *memcg)
>>>>>>>>  if (!info)
>>>>>>>>  goto err;
>>>>>>>>  info->map_nr_max = shrinker_nr_max;
>>>>>>>> - if (shrinker_unit_alloc(info, NULL, nid))
>>>>>>>> + if (shrinker_unit_alloc(info, NULL, nid)) {
>>>>>>>> + kvfree(info);
>>>>>>>>  goto err;
>>>>>>>> + }
>>>>>>>>  rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_info, info);
>>>>>>>>  }
>>>>>>>>  mutex_unlock(&shrinker_mutex);
>>>>>>> Agreed, this is what I mentioned earlier as well.
>>>>>>> ------------------------------------------------------------------
>>>>>>> I guess kvfree() should be called just after shrinker_unit_alloc()
>>>>>>> fails but before calling into "goto err"
>>>>>>> ------------------------------------------------------------------
>>>>>> After discussion, it seems that v1 is acceptable.
>>>>>> Hi, Muchun, do you have any other opinions?
>>>>>
>>>>> I insist on my opinion, not mixing two different approaches
>>>>> to do release resources.
>>>>
>>>> So instead we mix the cleanup of the whole function with the cleanup of what
>>>> is effectively a per-iteration temporary variable?
>>>>
>>>> The fact there was already a confusion in this thread about whether it's
>>>> safe and relies on kvfree(NULL) to be a no-op, should be a hint.
>>>
>>> Yes. I think someone is confused about my opinion.
>>> I don’t care about whether we should apply this hit.
>>> If we think the hint is tricky, we could add another
>>> label to fix it like I suggested previously. Because
>>> we already use goto-based approaches to
>>> cleanup the resources, why not keeping
>>> consistent?
>>
>> I think we're rather pragmatic than striving to be consistent for the sake
>> of consistency. goto is not the nicest thing in the world, but we (unlike
>> other projects) use it where it makes sense to avoid if/else nesting
>> explosion. Here for the info it's not the most pragmatic option.
>>
>>> It will be easier for us to add a new
>>> "if" statement and handle the failure case in the future.
>>
>> Let's not overengineer things for hypothetical future.
>>
>>> For example, if we use his v1 proposal, we should do
>>> the cleanups again for info. But for goto-based
>>> version, we just add another label to do the
>>> cleanups and go to the new label for failure case. goto-based fix is what I insisted on. I copied my previous suggested fix here to clarify my opinion.
>>
>> Again, info is a loop-iteration-local variable, v1 fix making it truly local
>> is the way to go. If there are further cleanups added in the loop itself in
>> the future, they could hopefully keep being local to the loop as well.
>> Cleanup of info outside the loop iteration is breaking its real scope.
> 
> +1 to that.
> 
> I don't think it's such a big deal and both versions are ok, but I strongly
> prefer the original version (without introduction of a new label).
> 
> Please, feel free to use
> Acked-by: Roman Gushchin <roman.gushchin@linux.dev>
> with the original version.
> 
> Thanks!

I agree with Roman.
Hello, Andrew and Dave, Do you have any opinions?

The original version:
https://lore.kernel.org/linux-kernel/4184c61f-80f7-4adc-8929-c29f959cb8df@huawei.com/

Best regards,
Ridong