[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <95d5b806-d912-4a63-add6-ac115e8f181d@huawei.com>
Date: Mon, 14 Oct 2024 17:38:07 +0800
From: chenridong <chenridong@...wei.com>
To: Muchun Song <muchun.song@...ux.dev>
CC: Anshuman Khandual <anshuman.khandual@....com>, Chen Ridong
<chenridong@...weicloud.com>, <akpm@...ux-foundation.org>,
<david@...morbit.com>, <zhengqi.arch@...edance.com>,
<roman.gushchin@...ux.dev>, <linux-mm@...ck.org>,
<linux-kernel@...r.kernel.org>, <wangweiyang2@...wei.com>
Subject: Re: [PATCH v2] mm: shrinker: avoid memleak in alloc_shrinker_info
On 2024/10/14 17:20, Muchun Song wrote:
>
>
>> On Oct 14, 2024, at 17:04, chenridong <chenridong@...wei.com> wrote:
>>
>>
>>
>> On 2024/10/14 16:43, Muchun Song wrote:
>>>> On Oct 14, 2024, at 16:13, Anshuman Khandual <anshuman.khandual@....com> wrote:
>>>>
>>>>
>>>>
>>>> On 10/14/24 08:53, Chen Ridong wrote:
>>>>> From: Chen Ridong <chenridong@...wei.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@...wei.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.
>
Thank you, it seems that 'goto free' is a better choice.
Will update.
Thanks,
Ridong
>>
>> I think this is concise.
>>
>> Best regards,
>> Ridong
>
Powered by blists - more mailing lists