diff mbox series

comedi: Fix potential memory leak in comedi_init()

Message ID 20221101032125.27337-1-shangxiaojing@huawei.com
State New
Headers
Series comedi: Fix potential memory leak in comedi_init() |

Commit Message

Shang XiaoJing Nov. 1, 2022, 3:21 a.m. UTC 
  comedi_init() will goto out_unregister_chrdev_region if cdev_add()
failed, which won't free the resource alloced in kobject_set_name().
Call kfree_const() to free the leaked name before goto
out_unregister_chrdev_region.

unreferenced object 0xffff8881000fa8c0 (size 8):
  comm "modprobe", pid 239, jiffies 4294905173 (age 51.308s)
  hex dump (first 8 bytes):
    63 6f 6d 65 64 69 00 ff                          comedi..
  backtrace:
    [<000000005f9878f7>] __kmalloc_node_track_caller+0x4c/0x1c0
    [<000000000fd70302>] kstrdup+0x3f/0x70
    [<000000009428bc33>] kstrdup_const+0x46/0x60
    [<00000000ed50d9de>] kvasprintf_const+0xdb/0xf0
    [<00000000b2766964>] kobject_set_name_vargs+0x3c/0xe0
    [<00000000f2424ef7>] kobject_set_name+0x62/0x90
    [<000000005d5a125b>] 0xffffffffa0013098
    [<00000000f331e663>] do_one_initcall+0x7a/0x380
    [<00000000aa7bac96>] do_init_module+0x5c/0x230
    [<000000005fd72335>] load_module+0x227d/0x2420
    [<00000000ad550cf1>] __do_sys_finit_module+0xd5/0x140
    [<00000000069a60c5>] do_syscall_64+0x3f/0x90
    [<00000000c5e0d521>] entry_SYSCALL_64_after_hwframe+0x63/0xcd

Fixes: ed9eccbe8970 ("Staging: add comedi core")
Signed-off-by: Shang XiaoJing <shangxiaojing@huawei.com>
---
 drivers/comedi/comedi_fops.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

Comments

Greg KH Nov. 1, 2022, 4:45 a.m. UTC | #1 
On Tue, Nov 01, 2022 at 11:21:25AM +0800, Shang XiaoJing wrote:
> comedi_init() will goto out_unregister_chrdev_region if cdev_add()
> failed, which won't free the resource alloced in kobject_set_name().
> Call kfree_const() to free the leaked name before goto
> out_unregister_chrdev_region.
> 
> unreferenced object 0xffff8881000fa8c0 (size 8):
>   comm "modprobe", pid 239, jiffies 4294905173 (age 51.308s)
>   hex dump (first 8 bytes):
>     63 6f 6d 65 64 69 00 ff                          comedi..
>   backtrace:
>     [<000000005f9878f7>] __kmalloc_node_track_caller+0x4c/0x1c0
>     [<000000000fd70302>] kstrdup+0x3f/0x70
>     [<000000009428bc33>] kstrdup_const+0x46/0x60
>     [<00000000ed50d9de>] kvasprintf_const+0xdb/0xf0
>     [<00000000b2766964>] kobject_set_name_vargs+0x3c/0xe0
>     [<00000000f2424ef7>] kobject_set_name+0x62/0x90
>     [<000000005d5a125b>] 0xffffffffa0013098
>     [<00000000f331e663>] do_one_initcall+0x7a/0x380
>     [<00000000aa7bac96>] do_init_module+0x5c/0x230
>     [<000000005fd72335>] load_module+0x227d/0x2420
>     [<00000000ad550cf1>] __do_sys_finit_module+0xd5/0x140
>     [<00000000069a60c5>] do_syscall_64+0x3f/0x90
>     [<00000000c5e0d521>] entry_SYSCALL_64_after_hwframe+0x63/0xcd
> 
> Fixes: ed9eccbe8970 ("Staging: add comedi core")
> Signed-off-by: Shang XiaoJing <shangxiaojing@huawei.com>
> ---
>  drivers/comedi/comedi_fops.c | 5 ++++-
>  1 file changed, 4 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/comedi/comedi_fops.c b/drivers/comedi/comedi_fops.c
> index e2114bcf815a..2c508c2cf6f6 100644
> --- a/drivers/comedi/comedi_fops.c
> +++ b/drivers/comedi/comedi_fops.c
> @@ -3379,8 +3379,11 @@ static int __init comedi_init(void)
>  
>  	retval = cdev_add(&comedi_cdev, MKDEV(COMEDI_MAJOR, 0),
>  			  COMEDI_NUM_MINORS);
> -	if (retval)
> +	if (retval) {
> +		kfree_const(comedi_cdev.kobj.name);
> +		comedi_cdev.kobj.name = NULL;
>  		goto out_unregister_chrdev_region;
> +	}

A driver should never have to poke around in the internals of a cdev
object like this.  Please fix the cdev core to not need this if
cdev_add() fails.

thanks,

greg k-h
Shang XiaoJing Nov. 1, 2022, 6:16 a.m. UTC | #2 
On 2022/11/1 12:45, Greg KH wrote:
> On Tue, Nov 01, 2022 at 11:21:25AM +0800, Shang XiaoJing wrote:
>> comedi_init() will goto out_unregister_chrdev_region if cdev_add()
>> failed, which won't free the resource alloced in kobject_set_name().
>> Call kfree_const() to free the leaked name before goto
>> out_unregister_chrdev_region.
>>
>> unreferenced object 0xffff8881000fa8c0 (size 8):
>>    comm "modprobe", pid 239, jiffies 4294905173 (age 51.308s)
>>    hex dump (first 8 bytes):
>>      63 6f 6d 65 64 69 00 ff                          comedi..
>>    backtrace:
>>      [<000000005f9878f7>] __kmalloc_node_track_caller+0x4c/0x1c0
>>      [<000000000fd70302>] kstrdup+0x3f/0x70
>>      [<000000009428bc33>] kstrdup_const+0x46/0x60
>>      [<00000000ed50d9de>] kvasprintf_const+0xdb/0xf0
>>      [<00000000b2766964>] kobject_set_name_vargs+0x3c/0xe0
>>      [<00000000f2424ef7>] kobject_set_name+0x62/0x90
>>      [<000000005d5a125b>] 0xffffffffa0013098
>>      [<00000000f331e663>] do_one_initcall+0x7a/0x380
>>      [<00000000aa7bac96>] do_init_module+0x5c/0x230
>>      [<000000005fd72335>] load_module+0x227d/0x2420
>>      [<00000000ad550cf1>] __do_sys_finit_module+0xd5/0x140
>>      [<00000000069a60c5>] do_syscall_64+0x3f/0x90
>>      [<00000000c5e0d521>] entry_SYSCALL_64_after_hwframe+0x63/0xcd
>>
>> Fixes: ed9eccbe8970 ("Staging: add comedi core")
>> Signed-off-by: Shang XiaoJing <shangxiaojing@huawei.com>
>> ---
>>   drivers/comedi/comedi_fops.c | 5 ++++-
>>   1 file changed, 4 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/comedi/comedi_fops.c b/drivers/comedi/comedi_fops.c
>> index e2114bcf815a..2c508c2cf6f6 100644
>> --- a/drivers/comedi/comedi_fops.c
>> +++ b/drivers/comedi/comedi_fops.c
>> @@ -3379,8 +3379,11 @@ static int __init comedi_init(void)
>>   
>>   	retval = cdev_add(&comedi_cdev, MKDEV(COMEDI_MAJOR, 0),
>>   			  COMEDI_NUM_MINORS);
>> -	if (retval)
>> +	if (retval) {
>> +		kfree_const(comedi_cdev.kobj.name);
>> +		comedi_cdev.kobj.name = NULL;
>>   		goto out_unregister_chrdev_region;
>> +	}
> 
> A driver should never have to poke around in the internals of a cdev
> object like this.  Please fix the cdev core to not need this if
> cdev_add() fails.
> 
Hi,

I had checked all 67 places that used cdev_add(), and only the following 
5 functions set the name before the cdev_add():

1. comedi_init(), will memleak and will be fixed by this patch.
2. hfi1_cdev_init(), won't memleak.
3. qib_cdev_init(), won't memleak.
4. uio_major_init(), won't memleak.
5. __register_chrdev(), won't memleak.

As the result, I just fix the code in comedi_init(). Should I still fix 
the cdev core code, which is free the name in cdev_add() fail path?

Thanks,
Ian Abbott Nov. 1, 2022, 11:40 a.m. UTC | #3 
On 01/11/2022 06:16, shangxiaojing wrote:
> 
> 
> On 2022/11/1 12:45, Greg KH wrote:
>> On Tue, Nov 01, 2022 at 11:21:25AM +0800, Shang XiaoJing wrote:
>>> comedi_init() will goto out_unregister_chrdev_region if cdev_add()
>>> failed, which won't free the resource alloced in kobject_set_name().
>>> Call kfree_const() to free the leaked name before goto
>>> out_unregister_chrdev_region.
>>>
>>> unreferenced object 0xffff8881000fa8c0 (size 8):
>>>    comm "modprobe", pid 239, jiffies 4294905173 (age 51.308s)
>>>    hex dump (first 8 bytes):
>>>      63 6f 6d 65 64 69 00 ff                          comedi..
>>>    backtrace:
>>>      [<000000005f9878f7>] __kmalloc_node_track_caller+0x4c/0x1c0
>>>      [<000000000fd70302>] kstrdup+0x3f/0x70
>>>      [<000000009428bc33>] kstrdup_const+0x46/0x60
>>>      [<00000000ed50d9de>] kvasprintf_const+0xdb/0xf0
>>>      [<00000000b2766964>] kobject_set_name_vargs+0x3c/0xe0
>>>      [<00000000f2424ef7>] kobject_set_name+0x62/0x90
>>>      [<000000005d5a125b>] 0xffffffffa0013098
>>>      [<00000000f331e663>] do_one_initcall+0x7a/0x380
>>>      [<00000000aa7bac96>] do_init_module+0x5c/0x230
>>>      [<000000005fd72335>] load_module+0x227d/0x2420
>>>      [<00000000ad550cf1>] __do_sys_finit_module+0xd5/0x140
>>>      [<00000000069a60c5>] do_syscall_64+0x3f/0x90
>>>      [<00000000c5e0d521>] entry_SYSCALL_64_after_hwframe+0x63/0xcd
>>>
>>> Fixes: ed9eccbe8970 ("Staging: add comedi core")
>>> Signed-off-by: Shang XiaoJing <shangxiaojing@huawei.com>
>>> ---
>>>   drivers/comedi/comedi_fops.c | 5 ++++-
>>>   1 file changed, 4 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/drivers/comedi/comedi_fops.c b/drivers/comedi/comedi_fops.c
>>> index e2114bcf815a..2c508c2cf6f6 100644
>>> --- a/drivers/comedi/comedi_fops.c
>>> +++ b/drivers/comedi/comedi_fops.c
>>> @@ -3379,8 +3379,11 @@ static int __init comedi_init(void)
>>>       retval = cdev_add(&comedi_cdev, MKDEV(COMEDI_MAJOR, 0),
>>>                 COMEDI_NUM_MINORS);
>>> -    if (retval)
>>> +    if (retval) {
>>> +        kfree_const(comedi_cdev.kobj.name);
>>> +        comedi_cdev.kobj.name = NULL;
>>>           goto out_unregister_chrdev_region;
>>> +    }
>>
>> A driver should never have to poke around in the internals of a cdev
>> object like this.  Please fix the cdev core to not need this if
>> cdev_add() fails.

Or at least there should be a function that can be called to undo the 
allocations of kobject_set_name().

>>
> Hi,
> 
> I had checked all 67 places that used cdev_add(), and only the following 
> 5 functions set the name before the cdev_add():
> 
> 1. comedi_init(), will memleak and will be fixed by this patch.
> 2. hfi1_cdev_init(), won't memleak.
> 3. qib_cdev_init(), won't memleak.
> 4. uio_major_init(), won't memleak.
> 5. __register_chrdev(), won't memleak.
> 
> As the result, I just fix the code in comedi_init(). Should I still fix 
> the cdev core code, which is free the name in cdev_add() fail path?

I'm stuck trying to work out why hfi1_cdev_init() doesn't have the same 
problem when cdev_add() returns an error.
Shang XiaoJing Nov. 1, 2022, 11:57 a.m. UTC | #4 
On 2022/11/1 19:40, Ian Abbott wrote:
> On 01/11/2022 06:16, shangxiaojing wrote:
>>
>>
>> On 2022/11/1 12:45, Greg KH wrote:
>>> On Tue, Nov 01, 2022 at 11:21:25AM +0800, Shang XiaoJing wrote:
>>>> comedi_init() will goto out_unregister_chrdev_region if cdev_add()
>>>> failed, which won't free the resource alloced in kobject_set_name().
>>>> Call kfree_const() to free the leaked name before goto
>>>> out_unregister_chrdev_region.
>>>>
>>>> unreferenced object 0xffff8881000fa8c0 (size 8):
>>>>    comm "modprobe", pid 239, jiffies 4294905173 (age 51.308s)
>>>>    hex dump (first 8 bytes):
>>>>      63 6f 6d 65 64 69 00 ff                          comedi..
>>>>    backtrace:
>>>>      [<000000005f9878f7>] __kmalloc_node_track_caller+0x4c/0x1c0
>>>>      [<000000000fd70302>] kstrdup+0x3f/0x70
>>>>      [<000000009428bc33>] kstrdup_const+0x46/0x60
>>>>      [<00000000ed50d9de>] kvasprintf_const+0xdb/0xf0
>>>>      [<00000000b2766964>] kobject_set_name_vargs+0x3c/0xe0
>>>>      [<00000000f2424ef7>] kobject_set_name+0x62/0x90
>>>>      [<000000005d5a125b>] 0xffffffffa0013098
>>>>      [<00000000f331e663>] do_one_initcall+0x7a/0x380
>>>>      [<00000000aa7bac96>] do_init_module+0x5c/0x230
>>>>      [<000000005fd72335>] load_module+0x227d/0x2420
>>>>      [<00000000ad550cf1>] __do_sys_finit_module+0xd5/0x140
>>>>      [<00000000069a60c5>] do_syscall_64+0x3f/0x90
>>>>      [<00000000c5e0d521>] entry_SYSCALL_64_after_hwframe+0x63/0xcd
>>>>
>>>> Fixes: ed9eccbe8970 ("Staging: add comedi core")
>>>> Signed-off-by: Shang XiaoJing <shangxiaojing@huawei.com>
>>>> ---
>>>>   drivers/comedi/comedi_fops.c | 5 ++++-
>>>>   1 file changed, 4 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/drivers/comedi/comedi_fops.c 
>>>> b/drivers/comedi/comedi_fops.c
>>>> index e2114bcf815a..2c508c2cf6f6 100644
>>>> --- a/drivers/comedi/comedi_fops.c
>>>> +++ b/drivers/comedi/comedi_fops.c
>>>> @@ -3379,8 +3379,11 @@ static int __init comedi_init(void)
>>>>       retval = cdev_add(&comedi_cdev, MKDEV(COMEDI_MAJOR, 0),
>>>>                 COMEDI_NUM_MINORS);
>>>> -    if (retval)
>>>> +    if (retval) {
>>>> +        kfree_const(comedi_cdev.kobj.name);
>>>> +        comedi_cdev.kobj.name = NULL;
>>>>           goto out_unregister_chrdev_region;
>>>> +    }
>>>
>>> A driver should never have to poke around in the internals of a cdev
>>> object like this.  Please fix the cdev core to not need this if
>>> cdev_add() fails.
> 
> Or at least there should be a function that can be called to undo the 
> allocations of kobject_set_name().

It's ok to add a pair function, but maybe only one place where used 
cdev_add() need free name. Your mean all "kfree_const(name);" should be 
replaced to the new function?

> 
>>>
>> Hi,
>>
>> I had checked all 67 places that used cdev_add(), and only the 
>> following 5 functions set the name before the cdev_add():
>>
>> 1. comedi_init(), will memleak and will be fixed by this patch.
>> 2. hfi1_cdev_init(), won't memleak.
>> 3. qib_cdev_init(), won't memleak.
>> 4. uio_major_init(), won't memleak.
>> 5. __register_chrdev(), won't memleak.
>>
>> As the result, I just fix the code in comedi_init(). Should I still 
>> fix the cdev core code, which is free the name in cdev_add() fail path?
> 
> I'm stuck trying to work out why hfi1_cdev_init() doesn't have the same 
> problem when cdev_add() returns an error.
> 

Only user_add() calls the hfi1_cdev_init(), and will call user_remove() 
if hfi1_cdev_init() failed. In user_remove(), hfi1_cdev_cleanup() will 
call cdev_del().

Thanks,
Ian Abbott Nov. 1, 2022, 3:59 p.m. UTC | #5 
On 01/11/2022 11:57, shangxiaojing wrote:
> 
> 
> On 2022/11/1 19:40, Ian Abbott wrote:
>> On 01/11/2022 06:16, shangxiaojing wrote:
>>>
>>>
>>> On 2022/11/1 12:45, Greg KH wrote:
>>>> On Tue, Nov 01, 2022 at 11:21:25AM +0800, Shang XiaoJing wrote:
>>>>> comedi_init() will goto out_unregister_chrdev_region if cdev_add()
>>>>> failed, which won't free the resource alloced in kobject_set_name().
>>>>> Call kfree_const() to free the leaked name before goto
>>>>> out_unregister_chrdev_region.
>>>>>
>>>>> unreferenced object 0xffff8881000fa8c0 (size 8):
>>>>>    comm "modprobe", pid 239, jiffies 4294905173 (age 51.308s)
>>>>>    hex dump (first 8 bytes):
>>>>>      63 6f 6d 65 64 69 00 ff                          comedi..
>>>>>    backtrace:
>>>>>      [<000000005f9878f7>] __kmalloc_node_track_caller+0x4c/0x1c0
>>>>>      [<000000000fd70302>] kstrdup+0x3f/0x70
>>>>>      [<000000009428bc33>] kstrdup_const+0x46/0x60
>>>>>      [<00000000ed50d9de>] kvasprintf_const+0xdb/0xf0
>>>>>      [<00000000b2766964>] kobject_set_name_vargs+0x3c/0xe0
>>>>>      [<00000000f2424ef7>] kobject_set_name+0x62/0x90
>>>>>      [<000000005d5a125b>] 0xffffffffa0013098
>>>>>      [<00000000f331e663>] do_one_initcall+0x7a/0x380
>>>>>      [<00000000aa7bac96>] do_init_module+0x5c/0x230
>>>>>      [<000000005fd72335>] load_module+0x227d/0x2420
>>>>>      [<00000000ad550cf1>] __do_sys_finit_module+0xd5/0x140
>>>>>      [<00000000069a60c5>] do_syscall_64+0x3f/0x90
>>>>>      [<00000000c5e0d521>] entry_SYSCALL_64_after_hwframe+0x63/0xcd
>>>>>
>>>>> Fixes: ed9eccbe8970 ("Staging: add comedi core")
>>>>> Signed-off-by: Shang XiaoJing <shangxiaojing@huawei.com>
>>>>> ---
>>>>>   drivers/comedi/comedi_fops.c | 5 ++++-
>>>>>   1 file changed, 4 insertions(+), 1 deletion(-)
>>>>>
>>>>> diff --git a/drivers/comedi/comedi_fops.c 
>>>>> b/drivers/comedi/comedi_fops.c
>>>>> index e2114bcf815a..2c508c2cf6f6 100644
>>>>> --- a/drivers/comedi/comedi_fops.c
>>>>> +++ b/drivers/comedi/comedi_fops.c
>>>>> @@ -3379,8 +3379,11 @@ static int __init comedi_init(void)
>>>>>       retval = cdev_add(&comedi_cdev, MKDEV(COMEDI_MAJOR, 0),
>>>>>                 COMEDI_NUM_MINORS);
>>>>> -    if (retval)
>>>>> +    if (retval) {
>>>>> +        kfree_const(comedi_cdev.kobj.name);
>>>>> +        comedi_cdev.kobj.name = NULL;
>>>>>           goto out_unregister_chrdev_region;
>>>>> +    }
>>>>
>>>> A driver should never have to poke around in the internals of a cdev
>>>> object like this.  Please fix the cdev core to not need this if
>>>> cdev_add() fails.
>>
>> Or at least there should be a function that can be called to undo the 
>> allocations of kobject_set_name().
> 
> It's ok to add a pair function, but maybe only one place where used 
> cdev_add() need free name. Your mean all "kfree_const(name);" should be 
> replaced to the new function?
> 
>>
>>>>
>>> Hi,
>>>
>>> I had checked all 67 places that used cdev_add(), and only the 
>>> following 5 functions set the name before the cdev_add():
>>>
>>> 1. comedi_init(), will memleak and will be fixed by this patch.
>>> 2. hfi1_cdev_init(), won't memleak.
>>> 3. qib_cdev_init(), won't memleak.
>>> 4. uio_major_init(), won't memleak.
>>> 5. __register_chrdev(), won't memleak.
>>>
>>> As the result, I just fix the code in comedi_init(). Should I still 
>>> fix the cdev core code, which is free the name in cdev_add() fail path?
>>
>> I'm stuck trying to work out why hfi1_cdev_init() doesn't have the 
>> same problem when cdev_add() returns an error.
>>
> 
> Only user_add() calls the hfi1_cdev_init(), and will call user_remove() 
> if hfi1_cdev_init() failed. In user_remove(), hfi1_cdev_cleanup() will 
> call cdev_del().

No it won't.  hfi1_cdev_cleanup() only calls cdev_del() if device is 
non-null, but device will be null if the call to cdev_add() failed in 
hfi1_cdev_init().
Ian Abbott Nov. 1, 2022, 4:41 p.m. UTC | #6 
On 01/11/2022 11:40, Ian Abbott wrote:
> On 01/11/2022 06:16, shangxiaojing wrote:
>>
>>
>> On 2022/11/1 12:45, Greg KH wrote:
>>> On Tue, Nov 01, 2022 at 11:21:25AM +0800, Shang XiaoJing wrote:
>>>> comedi_init() will goto out_unregister_chrdev_region if cdev_add()
>>>> failed, which won't free the resource alloced in kobject_set_name().
>>>> Call kfree_const() to free the leaked name before goto
>>>> out_unregister_chrdev_region.
>>>>
>>>> unreferenced object 0xffff8881000fa8c0 (size 8):
>>>>    comm "modprobe", pid 239, jiffies 4294905173 (age 51.308s)
>>>>    hex dump (first 8 bytes):
>>>>      63 6f 6d 65 64 69 00 ff                          comedi..
>>>>    backtrace:
>>>>      [<000000005f9878f7>] __kmalloc_node_track_caller+0x4c/0x1c0
>>>>      [<000000000fd70302>] kstrdup+0x3f/0x70
>>>>      [<000000009428bc33>] kstrdup_const+0x46/0x60
>>>>      [<00000000ed50d9de>] kvasprintf_const+0xdb/0xf0
>>>>      [<00000000b2766964>] kobject_set_name_vargs+0x3c/0xe0
>>>>      [<00000000f2424ef7>] kobject_set_name+0x62/0x90
>>>>      [<000000005d5a125b>] 0xffffffffa0013098
>>>>      [<00000000f331e663>] do_one_initcall+0x7a/0x380
>>>>      [<00000000aa7bac96>] do_init_module+0x5c/0x230
>>>>      [<000000005fd72335>] load_module+0x227d/0x2420
>>>>      [<00000000ad550cf1>] __do_sys_finit_module+0xd5/0x140
>>>>      [<00000000069a60c5>] do_syscall_64+0x3f/0x90
>>>>      [<00000000c5e0d521>] entry_SYSCALL_64_after_hwframe+0x63/0xcd
>>>>
>>>> Fixes: ed9eccbe8970 ("Staging: add comedi core")
>>>> Signed-off-by: Shang XiaoJing <shangxiaojing@huawei.com>
>>>> ---
>>>>   drivers/comedi/comedi_fops.c | 5 ++++-
>>>>   1 file changed, 4 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/drivers/comedi/comedi_fops.c 
>>>> b/drivers/comedi/comedi_fops.c
>>>> index e2114bcf815a..2c508c2cf6f6 100644
>>>> --- a/drivers/comedi/comedi_fops.c
>>>> +++ b/drivers/comedi/comedi_fops.c
>>>> @@ -3379,8 +3379,11 @@ static int __init comedi_init(void)
>>>>       retval = cdev_add(&comedi_cdev, MKDEV(COMEDI_MAJOR, 0),
>>>>                 COMEDI_NUM_MINORS);
>>>> -    if (retval)
>>>> +    if (retval) {
>>>> +        kfree_const(comedi_cdev.kobj.name);
>>>> +        comedi_cdev.kobj.name = NULL;
>>>>           goto out_unregister_chrdev_region;
>>>> +    }
>>>
>>> A driver should never have to poke around in the internals of a cdev
>>> object like this.  Please fix the cdev core to not need this if
>>> cdev_add() fails.
> 
> Or at least there should be a function that can be called to undo the 
> allocations of kobject_set_name().

Looking at it a bit more, cdev_init() calls kobject_init(), and 
kobject_init() requires a matching call to kobject_put().  Nothing is 
calling kobject_put() in this situation.  Calling cdev_del() will call 
kobject_put(), so is that the correct way to clean up after cdev_init() 
if the call to cdev_add() fails?
Shang XiaoJing Nov. 2, 2022, 2:41 a.m. UTC | #7 
On 2022/11/1 23:59, Ian Abbott wrote:
> On 01/11/2022 11:57, shangxiaojing wrote:
>>
>>
>> On 2022/11/1 19:40, Ian Abbott wrote:
>>> On 01/11/2022 06:16, shangxiaojing wrote:
>>>>
>>>>
>>>> On 2022/11/1 12:45, Greg KH wrote:
>>>>> On Tue, Nov 01, 2022 at 11:21:25AM +0800, Shang XiaoJing wrote:
>>>>>> comedi_init() will goto out_unregister_chrdev_region if cdev_add()
>>>>>> failed, which won't free the resource alloced in kobject_set_name().
>>>>>> Call kfree_const() to free the leaked name before goto
>>>>>> out_unregister_chrdev_region.
>>>>>>
>>>>>> unreferenced object 0xffff8881000fa8c0 (size 8):
>>>>>>    comm "modprobe", pid 239, jiffies 4294905173 (age 51.308s)
>>>>>>    hex dump (first 8 bytes):
>>>>>>      63 6f 6d 65 64 69 00 ff                          comedi..
>>>>>>    backtrace:
>>>>>>      [<000000005f9878f7>] __kmalloc_node_track_caller+0x4c/0x1c0
>>>>>>      [<000000000fd70302>] kstrdup+0x3f/0x70
>>>>>>      [<000000009428bc33>] kstrdup_const+0x46/0x60
>>>>>>      [<00000000ed50d9de>] kvasprintf_const+0xdb/0xf0
>>>>>>      [<00000000b2766964>] kobject_set_name_vargs+0x3c/0xe0
>>>>>>      [<00000000f2424ef7>] kobject_set_name+0x62/0x90
>>>>>>      [<000000005d5a125b>] 0xffffffffa0013098
>>>>>>      [<00000000f331e663>] do_one_initcall+0x7a/0x380
>>>>>>      [<00000000aa7bac96>] do_init_module+0x5c/0x230
>>>>>>      [<000000005fd72335>] load_module+0x227d/0x2420
>>>>>>      [<00000000ad550cf1>] __do_sys_finit_module+0xd5/0x140
>>>>>>      [<00000000069a60c5>] do_syscall_64+0x3f/0x90
>>>>>>      [<00000000c5e0d521>] entry_SYSCALL_64_after_hwframe+0x63/0xcd
>>>>>>
>>>>>> Fixes: ed9eccbe8970 ("Staging: add comedi core")
>>>>>> Signed-off-by: Shang XiaoJing <shangxiaojing@huawei.com>
>>>>>> ---
>>>>>>   drivers/comedi/comedi_fops.c | 5 ++++-
>>>>>>   1 file changed, 4 insertions(+), 1 deletion(-)
>>>>>>
>>>>>> diff --git a/drivers/comedi/comedi_fops.c 
>>>>>> b/drivers/comedi/comedi_fops.c
>>>>>> index e2114bcf815a..2c508c2cf6f6 100644
>>>>>> --- a/drivers/comedi/comedi_fops.c
>>>>>> +++ b/drivers/comedi/comedi_fops.c
>>>>>> @@ -3379,8 +3379,11 @@ static int __init comedi_init(void)
>>>>>>       retval = cdev_add(&comedi_cdev, MKDEV(COMEDI_MAJOR, 0),
>>>>>>                 COMEDI_NUM_MINORS);
>>>>>> -    if (retval)
>>>>>> +    if (retval) {
>>>>>> +        kfree_const(comedi_cdev.kobj.name);
>>>>>> +        comedi_cdev.kobj.name = NULL;
>>>>>>           goto out_unregister_chrdev_region;
>>>>>> +    }
>>>>>
>>>>> A driver should never have to poke around in the internals of a cdev
>>>>> object like this.  Please fix the cdev core to not need this if
>>>>> cdev_add() fails.
>>>
>>> Or at least there should be a function that can be called to undo the 
>>> allocations of kobject_set_name().
>>
>> It's ok to add a pair function, but maybe only one place where used 
>> cdev_add() need free name. Your mean all "kfree_const(name);" should 
>> be replaced to the new function?
>>
>>>
>>>>>
>>>> Hi,
>>>>
>>>> I had checked all 67 places that used cdev_add(), and only the 
>>>> following 5 functions set the name before the cdev_add():
>>>>
>>>> 1. comedi_init(), will memleak and will be fixed by this patch.
>>>> 2. hfi1_cdev_init(), won't memleak.
>>>> 3. qib_cdev_init(), won't memleak.
>>>> 4. uio_major_init(), won't memleak.
>>>> 5. __register_chrdev(), won't memleak.
>>>>
>>>> As the result, I just fix the code in comedi_init(). Should I still 
>>>> fix the cdev core code, which is free the name in cdev_add() fail path?
>>>
>>> I'm stuck trying to work out why hfi1_cdev_init() doesn't have the 
>>> same problem when cdev_add() returns an error.
>>>
>>
>> Only user_add() calls the hfi1_cdev_init(), and will call 
>> user_remove() if hfi1_cdev_init() failed. In user_remove(), 
>> hfi1_cdev_cleanup() will call cdev_del().
> 
> No it won't.  hfi1_cdev_cleanup() only calls cdev_del() if device is 
> non-null, but device will be null if the call to cdev_add() failed in 
> hfi1_cdev_init().
> 

Right, device won't be inited, so there has the same potential memleak.

Thanks,
Shang XiaoJing Nov. 2, 2022, 2:51 a.m. UTC | #8 
On 2022/11/2 0:41, Ian Abbott wrote:
> On 01/11/2022 11:40, Ian Abbott wrote:
>> On 01/11/2022 06:16, shangxiaojing wrote:
>>>
>>>
>>> On 2022/11/1 12:45, Greg KH wrote:
>>>> On Tue, Nov 01, 2022 at 11:21:25AM +0800, Shang XiaoJing wrote:
>>>>> comedi_init() will goto out_unregister_chrdev_region if cdev_add()
>>>>> failed, which won't free the resource alloced in kobject_set_name().
>>>>> Call kfree_const() to free the leaked name before goto
>>>>> out_unregister_chrdev_region.
>>>>>
>>>>> unreferenced object 0xffff8881000fa8c0 (size 8):
>>>>>    comm "modprobe", pid 239, jiffies 4294905173 (age 51.308s)
>>>>>    hex dump (first 8 bytes):
>>>>>      63 6f 6d 65 64 69 00 ff                          comedi..
>>>>>    backtrace:
>>>>>      [<000000005f9878f7>] __kmalloc_node_track_caller+0x4c/0x1c0
>>>>>      [<000000000fd70302>] kstrdup+0x3f/0x70
>>>>>      [<000000009428bc33>] kstrdup_const+0x46/0x60
>>>>>      [<00000000ed50d9de>] kvasprintf_const+0xdb/0xf0
>>>>>      [<00000000b2766964>] kobject_set_name_vargs+0x3c/0xe0
>>>>>      [<00000000f2424ef7>] kobject_set_name+0x62/0x90
>>>>>      [<000000005d5a125b>] 0xffffffffa0013098
>>>>>      [<00000000f331e663>] do_one_initcall+0x7a/0x380
>>>>>      [<00000000aa7bac96>] do_init_module+0x5c/0x230
>>>>>      [<000000005fd72335>] load_module+0x227d/0x2420
>>>>>      [<00000000ad550cf1>] __do_sys_finit_module+0xd5/0x140
>>>>>      [<00000000069a60c5>] do_syscall_64+0x3f/0x90
>>>>>      [<00000000c5e0d521>] entry_SYSCALL_64_after_hwframe+0x63/0xcd
>>>>>
>>>>> Fixes: ed9eccbe8970 ("Staging: add comedi core")
>>>>> Signed-off-by: Shang XiaoJing <shangxiaojing@huawei.com>
>>>>> ---
>>>>>   drivers/comedi/comedi_fops.c | 5 ++++-
>>>>>   1 file changed, 4 insertions(+), 1 deletion(-)
>>>>>
>>>>> diff --git a/drivers/comedi/comedi_fops.c 
>>>>> b/drivers/comedi/comedi_fops.c
>>>>> index e2114bcf815a..2c508c2cf6f6 100644
>>>>> --- a/drivers/comedi/comedi_fops.c
>>>>> +++ b/drivers/comedi/comedi_fops.c
>>>>> @@ -3379,8 +3379,11 @@ static int __init comedi_init(void)
>>>>>       retval = cdev_add(&comedi_cdev, MKDEV(COMEDI_MAJOR, 0),
>>>>>                 COMEDI_NUM_MINORS);
>>>>> -    if (retval)
>>>>> +    if (retval) {
>>>>> +        kfree_const(comedi_cdev.kobj.name);
>>>>> +        comedi_cdev.kobj.name = NULL;
>>>>>           goto out_unregister_chrdev_region;
>>>>> +    }
>>>>
>>>> A driver should never have to poke around in the internals of a cdev
>>>> object like this.  Please fix the cdev core to not need this if
>>>> cdev_add() fails.
>>
>> Or at least there should be a function that can be called to undo the 
>> allocations of kobject_set_name().
> 
> Looking at it a bit more, cdev_init() calls kobject_init(), and 
> kobject_init() requires a matching call to kobject_put().  Nothing is 
> calling kobject_put() in this situation.  Calling cdev_del() will call 
> kobject_put(), so is that the correct way to clean up after cdev_init() 
> if the call to cdev_add() fails?
> 

Some cdev call cdev_del() when cdev_add() failed (like init_dvbdev()), 
and at that time the cdev_unmap() in cdev_del() won't do anything due to 
the failure of cdev_add(), which is calling the kobject_put() when 
cdev_add() failed. But I'm not sure which one is better.

Thanks,
Greg KH Nov. 2, 2022, 2:59 a.m. UTC | #9 
On Wed, Nov 02, 2022 at 10:51:35AM +0800, shangxiaojing wrote:
> 
> 
> On 2022/11/2 0:41, Ian Abbott wrote:
> > On 01/11/2022 11:40, Ian Abbott wrote:
> > > On 01/11/2022 06:16, shangxiaojing wrote:
> > > > 
> > > > 
> > > > On 2022/11/1 12:45, Greg KH wrote:
> > > > > On Tue, Nov 01, 2022 at 11:21:25AM +0800, Shang XiaoJing wrote:
> > > > > > comedi_init() will goto out_unregister_chrdev_region if cdev_add()
> > > > > > failed, which won't free the resource alloced in kobject_set_name().
> > > > > > Call kfree_const() to free the leaked name before goto
> > > > > > out_unregister_chrdev_region.
> > > > > > 
> > > > > > unreferenced object 0xffff8881000fa8c0 (size 8):
> > > > > >    comm "modprobe", pid 239, jiffies 4294905173 (age 51.308s)
> > > > > >    hex dump (first 8 bytes):
> > > > > >      63 6f 6d 65 64 69 00 ff                          comedi..
> > > > > >    backtrace:
> > > > > >      [<000000005f9878f7>] __kmalloc_node_track_caller+0x4c/0x1c0
> > > > > >      [<000000000fd70302>] kstrdup+0x3f/0x70
> > > > > >      [<000000009428bc33>] kstrdup_const+0x46/0x60
> > > > > >      [<00000000ed50d9de>] kvasprintf_const+0xdb/0xf0
> > > > > >      [<00000000b2766964>] kobject_set_name_vargs+0x3c/0xe0
> > > > > >      [<00000000f2424ef7>] kobject_set_name+0x62/0x90
> > > > > >      [<000000005d5a125b>] 0xffffffffa0013098
> > > > > >      [<00000000f331e663>] do_one_initcall+0x7a/0x380
> > > > > >      [<00000000aa7bac96>] do_init_module+0x5c/0x230
> > > > > >      [<000000005fd72335>] load_module+0x227d/0x2420
> > > > > >      [<00000000ad550cf1>] __do_sys_finit_module+0xd5/0x140
> > > > > >      [<00000000069a60c5>] do_syscall_64+0x3f/0x90
> > > > > >      [<00000000c5e0d521>] entry_SYSCALL_64_after_hwframe+0x63/0xcd
> > > > > > 
> > > > > > Fixes: ed9eccbe8970 ("Staging: add comedi core")
> > > > > > Signed-off-by: Shang XiaoJing <shangxiaojing@huawei.com>
> > > > > > ---
> > > > > >   drivers/comedi/comedi_fops.c | 5 ++++-
> > > > > >   1 file changed, 4 insertions(+), 1 deletion(-)
> > > > > > 
> > > > > > diff --git a/drivers/comedi/comedi_fops.c
> > > > > > b/drivers/comedi/comedi_fops.c
> > > > > > index e2114bcf815a..2c508c2cf6f6 100644
> > > > > > --- a/drivers/comedi/comedi_fops.c
> > > > > > +++ b/drivers/comedi/comedi_fops.c
> > > > > > @@ -3379,8 +3379,11 @@ static int __init comedi_init(void)
> > > > > >       retval = cdev_add(&comedi_cdev, MKDEV(COMEDI_MAJOR, 0),
> > > > > >                 COMEDI_NUM_MINORS);
> > > > > > -    if (retval)
> > > > > > +    if (retval) {
> > > > > > +        kfree_const(comedi_cdev.kobj.name);
> > > > > > +        comedi_cdev.kobj.name = NULL;
> > > > > >           goto out_unregister_chrdev_region;
> > > > > > +    }
> > > > > 
> > > > > A driver should never have to poke around in the internals of a cdev
> > > > > object like this.  Please fix the cdev core to not need this if
> > > > > cdev_add() fails.
> > > 
> > > Or at least there should be a function that can be called to undo
> > > the allocations of kobject_set_name().
> > 
> > Looking at it a bit more, cdev_init() calls kobject_init(), and
> > kobject_init() requires a matching call to kobject_put().  Nothing is
> > calling kobject_put() in this situation.  Calling cdev_del() will call
> > kobject_put(), so is that the correct way to clean up after cdev_init()
> > if the call to cdev_add() fails?
> > 
> 
> Some cdev call cdev_del() when cdev_add() failed (like init_dvbdev()), and
> at that time the cdev_unmap() in cdev_del() won't do anything due to the
> failure of cdev_add(), which is calling the kobject_put() when cdev_add()
> failed. But I'm not sure which one is better.

My point is that no one calling the cdev_add() call should have to do
this type of "cleanup"  The cdev code should do it automatically, we
don't want to have to audit every user of cdev_add() today, and in the
future for forever, to ensure they got it right.

Let's fix it up in the cdev code itself please.  The kobject in a cdev
is a very odd thing, it's not the "normal" type of kobject that you
expect, so the cdev code should handle all of that on its own and that
should not "leak" into any caller.

thanks,

greg k-h
Shang XiaoJing Nov. 2, 2022, 6:06 a.m. UTC | #10 
On 2022/11/2 10:59, Greg KH wrote:
> On Wed, Nov 02, 2022 at 10:51:35AM +0800, shangxiaojing wrote:
>>
>>
>> On 2022/11/2 0:41, Ian Abbott wrote:
>>> On 01/11/2022 11:40, Ian Abbott wrote:
>>>> On 01/11/2022 06:16, shangxiaojing wrote:
>>>>>
>>>>>
>>>>> On 2022/11/1 12:45, Greg KH wrote:
>>>>>> On Tue, Nov 01, 2022 at 11:21:25AM +0800, Shang XiaoJing wrote:
>>>>>>> comedi_init() will goto out_unregister_chrdev_region if cdev_add()
>>>>>>> failed, which won't free the resource alloced in kobject_set_name().
>>>>>>> Call kfree_const() to free the leaked name before goto
>>>>>>> out_unregister_chrdev_region.
>>>>>>>
>>>>>>> unreferenced object 0xffff8881000fa8c0 (size 8):
>>>>>>>     comm "modprobe", pid 239, jiffies 4294905173 (age 51.308s)
>>>>>>>     hex dump (first 8 bytes):
>>>>>>>       63 6f 6d 65 64 69 00 ff                          comedi..
>>>>>>>     backtrace:
>>>>>>>       [<000000005f9878f7>] __kmalloc_node_track_caller+0x4c/0x1c0
>>>>>>>       [<000000000fd70302>] kstrdup+0x3f/0x70
>>>>>>>       [<000000009428bc33>] kstrdup_const+0x46/0x60
>>>>>>>       [<00000000ed50d9de>] kvasprintf_const+0xdb/0xf0
>>>>>>>       [<00000000b2766964>] kobject_set_name_vargs+0x3c/0xe0
>>>>>>>       [<00000000f2424ef7>] kobject_set_name+0x62/0x90
>>>>>>>       [<000000005d5a125b>] 0xffffffffa0013098
>>>>>>>       [<00000000f331e663>] do_one_initcall+0x7a/0x380
>>>>>>>       [<00000000aa7bac96>] do_init_module+0x5c/0x230
>>>>>>>       [<000000005fd72335>] load_module+0x227d/0x2420
>>>>>>>       [<00000000ad550cf1>] __do_sys_finit_module+0xd5/0x140
>>>>>>>       [<00000000069a60c5>] do_syscall_64+0x3f/0x90
>>>>>>>       [<00000000c5e0d521>] entry_SYSCALL_64_after_hwframe+0x63/0xcd
>>>>>>>
>>>>>>> Fixes: ed9eccbe8970 ("Staging: add comedi core")
>>>>>>> Signed-off-by: Shang XiaoJing <shangxiaojing@huawei.com>
>>>>>>> ---
>>>>>>>    drivers/comedi/comedi_fops.c | 5 ++++-
>>>>>>>    1 file changed, 4 insertions(+), 1 deletion(-)
>>>>>>>
>>>>>>> diff --git a/drivers/comedi/comedi_fops.c
>>>>>>> b/drivers/comedi/comedi_fops.c
>>>>>>> index e2114bcf815a..2c508c2cf6f6 100644
>>>>>>> --- a/drivers/comedi/comedi_fops.c
>>>>>>> +++ b/drivers/comedi/comedi_fops.c
>>>>>>> @@ -3379,8 +3379,11 @@ static int __init comedi_init(void)
>>>>>>>        retval = cdev_add(&comedi_cdev, MKDEV(COMEDI_MAJOR, 0),
>>>>>>>                  COMEDI_NUM_MINORS);
>>>>>>> -    if (retval)
>>>>>>> +    if (retval) {
>>>>>>> +        kfree_const(comedi_cdev.kobj.name);
>>>>>>> +        comedi_cdev.kobj.name = NULL;
>>>>>>>            goto out_unregister_chrdev_region;
>>>>>>> +    }
>>>>>>
>>>>>> A driver should never have to poke around in the internals of a cdev
>>>>>> object like this.  Please fix the cdev core to not need this if
>>>>>> cdev_add() fails.
>>>>
>>>> Or at least there should be a function that can be called to undo
>>>> the allocations of kobject_set_name().
>>>
>>> Looking at it a bit more, cdev_init() calls kobject_init(), and
>>> kobject_init() requires a matching call to kobject_put().  Nothing is
>>> calling kobject_put() in this situation.  Calling cdev_del() will call
>>> kobject_put(), so is that the correct way to clean up after cdev_init()
>>> if the call to cdev_add() fails?
>>>
>>
>> Some cdev call cdev_del() when cdev_add() failed (like init_dvbdev()), and
>> at that time the cdev_unmap() in cdev_del() won't do anything due to the
>> failure of cdev_add(), which is calling the kobject_put() when cdev_add()
>> failed. But I'm not sure which one is better.
> 
> My point is that no one calling the cdev_add() call should have to do
> this type of "cleanup"  The cdev code should do it automatically, we
> don't want to have to audit every user of cdev_add() today, and in the
> future for forever, to ensure they got it right.
> 
> Let's fix it up in the cdev code itself please.  The kobject in a cdev
> is a very odd thing, it's not the "normal" type of kobject that you
> expect, so the cdev code should handle all of that on its own and that
> should not "leak" into any caller.
> 

ok, I'll fix in cdev code and send v2.

Thanks,
diff mbox series

Patch

diff --git a/drivers/comedi/comedi_fops.c b/drivers/comedi/comedi_fops.c
index e2114bcf815a..2c508c2cf6f6 100644
--- a/drivers/comedi/comedi_fops.c
+++ b/drivers/comedi/comedi_fops.c
@@ -3379,8 +3379,11 @@  static int __init comedi_init(void)
 
 	retval = cdev_add(&comedi_cdev, MKDEV(COMEDI_MAJOR, 0),
 			  COMEDI_NUM_MINORS);
-	if (retval)
+	if (retval) {
+		kfree_const(comedi_cdev.kobj.name);
+		comedi_cdev.kobj.name = NULL;
 		goto out_unregister_chrdev_region;
+	}
 
 	comedi_class = class_create(THIS_MODULE, "comedi");
 	if (IS_ERR(comedi_class)) {