Re: [PATCH] nilfs2: fix memory leak in nilfs_sysfs_create_device_group

[Pavel Skripkin <paskripkin-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org>]

On 9/6/21 19:47, Pavel Skripkin wrote:
> On 9/6/21 18:56, Pavel Skripkin wrote:
>> On 9/6/21 08:43, Dongliang Mu wrote:
>>> On Mon, Sep 6, 2021 at 12:13 PM Dongliang Mu <mudongliangabcd-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org> wrote:
>>>>
>>>> The commit 8fd0c1b0647a ("nilfs2: fix memory leak in
>>>> nilfs_sysfs_delete_device_group") adds a kobject_put to free the leaking
>>>> object name. However, it is incomplete to only add kobject_put in the
>>>> nilfs_sysfs_delete_device_group. The function
>>>> nilfs_sysfs_create_device_group also needs the kobject_put to
>>>> free the object name in the error handling part.
>>>>
>>>> Fix this by adding kobject_put in the error handling code of
>>>> nilfs_sysfs_create_device_group.
>>> 
>>> Even after I add this patch, my local syzkaller still reports this
>>> memory leak one hour later. Therefore, there are some other paths or
>>> magics which can trigger the memory leak. I need to dig deeper.
>>> 
>>> Pavel, do you have any idea about this crash report?
>>> 
>>> BUG: memory leak
>>> unreferenced object 0xffff88804a1a8a60 (size 32):
>>>    comm "syz-executor", pid 14551, jiffies 4294960586 (age 14.780s)
>>>    hex dump (first 32 bytes):
>>>      6c 6f 6f 70 35 00 00 00 00 00 00 00 00 00 00 00  loop5...........
>>>      00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00  ................
>>>    backtrace:
>>>      [<ffffffff814750c6>] kstrdup+0x36/0x70
>>>      [<ffffffff81475153>] kstrdup_const+0x53/0x80
>>>      [<ffffffff822773a2>] kvasprintf_const+0xc2/0x110
>>>      [<ffffffff82337c5b>] kobject_set_name_vargs+0x3b/0xe0
>>>      [<ffffffff823385ed>] kobject_init_and_add+0x6d/0xc0
>>>      [<ffffffff81d2bd08>] nilfs_sysfs_create_device_group+0x98/0x3a0
>>>      [<ffffffff81d14fc4>] init_nilfs+0x424/0x580
>>>      [<ffffffff81d02962>] nilfs_mount+0x532/0x8c0
>>>      [<ffffffff815c754b>] legacy_get_tree+0x2b/0x90
>>>      [<ffffffff81565158>] vfs_get_tree+0x28/0x100
>>>      [<ffffffff815a3a82>] path_mount+0xb92/0xfe0
>>>      [<ffffffff815a3f71>] do_mount+0xa1/0xc0
>>>      [<ffffffff815a4584>] __x64_sys_mount+0xf4/0x160
>>>      [<ffffffff8433fd35>] do_syscall_64+0x35/0xb0
>>>      [<ffffffff84400068>] entry_SYSCALL_64_after_hwframe+0x44/0xae
>>> 
>>> 
>> 
>> Hi, Dongliang!
>> 
>> 
>> This report says nothing to me... It shows, that there is missing
>> kobject_put() somewhere. I think, we need a reproducer for this leak,
>> otherwise only code review can help :(
>> 
>> 
>> 
> 
> Hm, I guess, I see... We should call kobject_put() in case of
> kobject_init_and_add() failure:
> 
> lib/kobject.c:459
> 
>    * If this function returns an error, kobject_put() must be called to
>    * properly clean up the memory associated with the object.  This is the
> 
> 
> so I suggest:
> 
> diff --git a/fs/nilfs2/sysfs.c b/fs/nilfs2/sysfs.c
> index 68e8d61e28dd..e8717f4ba2a1 100644
> --- a/fs/nilfs2/sysfs.c
> +++ b/fs/nilfs2/sysfs.c
> @@ -1026,6 +1026,7 @@ int nilfs_sysfs_create_device_group(struct
> super_block *sb) 	kobject_del(&nilfs->ns_dev_kobj);
> 
>    free_dev_subgroups:
> +	kobject_put(&nilfs->ns_dev_kobj);
>    	kfree(nilfs->ns_dev_subgroups);
> 
>    failed_create_device_group:
> 
> 
> This patch should fix 2 memory leaks :)
> 
> 

Also, I think, we should add kobject_put() call to

1. nilfs_sysfs_create_##name##_group
2. nilfs_sysfs_create_snapshot_group


To prevent other random kobject memory leaks in nilfs...


With regards,
Pavel Skripkin