rbd: avoid double free memory on error path in rbd_dev_create()

If rbd_dev_create() fails after assignment 'opts' to 'rbd_dev->opts',
double free of 'rbd_options' happens:
one is in rbd_dev_free() and another one is in do_rbd_add().

Found by Linux Verification Center (linuxtesting.org) with SVACE.

Fixes: 1643dfa4c2c8 ("rbd: introduce a per-device ordered workqueue")
Signed-off-by: Natalia Petrova <n.petrova@fintech.ru>
Signed-off-by: Alexey Khoroshilov <khoroshilov@ispras.ru>
---
 drivers/block/rbd.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

On Fri, Feb 3, 2023 at 3:15 PM Natalia Petrova <n.petrova@fintech.ru> wrote:
>
> If rbd_dev_create() fails after assignment 'opts' to 'rbd_dev->opts',
> double free of 'rbd_options' happens:
> one is in rbd_dev_free() and another one is in do_rbd_add().
>
> Found by Linux Verification Center (linuxtesting.org) with SVACE.
>
> Fixes: 1643dfa4c2c8 ("rbd: introduce a per-device ordered workqueue")
> Signed-off-by: Natalia Petrova <n.petrova@fintech.ru>
> Signed-off-by: Alexey Khoroshilov <khoroshilov@ispras.ru>
> ---
>  drivers/block/rbd.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/drivers/block/rbd.c b/drivers/block/rbd.c
> index 04453f4a319c..ab6bfc352cde 100644
> --- a/drivers/block/rbd.c
> +++ b/drivers/block/rbd.c
> @@ -5357,7 +5357,6 @@ static struct rbd_device *rbd_dev_create(struct rbd_client *rbdc,
>         if (!rbd_dev)
>                 return NULL;
>
> -       rbd_dev->opts = opts;
>
>         /* get an id and fill in device name */
>         rbd_dev->dev_id = ida_simple_get(&rbd_dev_id_ida, 0,
> @@ -5372,6 +5371,7 @@ static struct rbd_device *rbd_dev_create(struct rbd_client *rbdc,
>         if (!rbd_dev->task_wq)
>                 goto fail_dev_id;
>
> +       rbd_dev->opts = opts;
>         /* we have a ref from do_rbd_add() */
>         __module_get(THIS_MODULE);
>
> --
> 2.34.1
>

Hi Natalia,

It seems like a similar issue is affecting rbd_dev->rbd_client and
rbd_dev->spec.  Unlike rbd_dev->opts, they are ref-counted and I'm
guessing that the verification tool doesn't go that deep.

I'd prefer all three to be addressed in the same change, since it's the
same error path.  Would you be willing to look into that and post a new
revision or should I treat just this patch as a bug report?

Thanks,

                Ilya

Hi Ilya!
Thanks for your response! I don't quite understand your idea and suggestion. The patch is designed to avoid double free memory. I explored the code again and suppose there is another situation for rbd_dev->rbd_client and rbd_dev->spec. Free memory of these pointers is possible only once in rbd_dev_free() function. In do_rbd_add() deallocation memory is only for rbd_opts: drivers/block/rbd.c 7157.
Correct me if I'm wrong.

Thanks, 
Natalia

-----Original Message-----
From: Ilya Dryomov <idryomov@gmail.com> 
Sent: Monday, February 6, 2023 2:59 PM
To: Петрова Наталия Михайловна <n.petrova@fintech.ru>
Cc: Dongsheng Yang <dongsheng.yang@easystack.cn>; Jens Axboe <axboe@kernel.dk>; ceph-devel@vger.kernel.org; linux-block@vger.kernel.org; linux-kernel@vger.kernel.org; lvc-project@linuxtesting.org; Alexey Khoroshilov <khoroshilov@ispras.ru>
Subject: Re: [PATCH] rbd: avoid double free memory on error path in rbd_dev_create()

On Fri, Feb 3, 2023 at 3:15 PM Natalia Petrova <n.petrova@fintech.ru> wrote:
>
> If rbd_dev_create() fails after assignment 'opts' to 'rbd_dev->opts', 
> double free of 'rbd_options' happens:
> one is in rbd_dev_free() and another one is in do_rbd_add().
>
> Found by Linux Verification Center (linuxtesting.org) with SVACE.
>
> Fixes: 1643dfa4c2c8 ("rbd: introduce a per-device ordered workqueue")
> Signed-off-by: Natalia Petrova <n.petrova@fintech.ru>
> Signed-off-by: Alexey Khoroshilov <khoroshilov@ispras.ru>
> ---
>  drivers/block/rbd.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/drivers/block/rbd.c b/drivers/block/rbd.c index 
> 04453f4a319c..ab6bfc352cde 100644
> --- a/drivers/block/rbd.c
> +++ b/drivers/block/rbd.c
> @@ -5357,7 +5357,6 @@ static struct rbd_device *rbd_dev_create(struct rbd_client *rbdc,
>         if (!rbd_dev)
>                 return NULL;
>
> -       rbd_dev->opts = opts;
>
>         /* get an id and fill in device name */
>         rbd_dev->dev_id = ida_simple_get(&rbd_dev_id_ida, 0, @@ 
> -5372,6 +5371,7 @@ static struct rbd_device *rbd_dev_create(struct rbd_client *rbdc,
>         if (!rbd_dev->task_wq)
>                 goto fail_dev_id;
>
> +       rbd_dev->opts = opts;
>         /* we have a ref from do_rbd_add() */
>         __module_get(THIS_MODULE);
>
> --
> 2.34.1
>

Hi Natalia,

It seems like a similar issue is affecting rbd_dev->rbd_client and rbd_dev->spec.  Unlike rbd_dev->opts, they are ref-counted and I'm guessing that the verification tool doesn't go that deep.

I'd prefer all three to be addressed in the same change, since it's the same error path.  Would you be willing to look into that and post a new revision or should I treat just this patch as a bug report?

Thanks,

                Ilya

On 06/02/2023 23:15, Петрова Наталия Михайловна wrote:
> Hi Ilya!
> Thanks for your response! I don't quite understand your idea and suggestion. The patch is designed to avoid double free memory. I explored the code again and suppose there is another situation for rbd_dev->rbd_client and rbd_dev->spec. Free memory of these pointers is possible only once in rbd_dev_free() function. In do_rbd_add() deallocation memory is only for rbd_opts: drivers/block/rbd.c 7157.

Hi Петрова,

If the rbd_dev_create() fails, for spec it will be freed in 
rbd_dev_create()->rbd_spec_put() first and then in do_rbd_add() it will 
call rbd_spec_put() again.

It won't trigger double free but this should generate a warning when the 
refcount underflow, because the refcount_dec_and_test() will warn and 
then return false when underflow happens.

The same for rbd_client.

Thanks,

- Xiubo

> Correct me if I'm wrong.
>
> Thanks,
> Natalia
>
> -----Original Message-----
> From: Ilya Dryomov <idryomov@gmail.com>
> Sent: Monday, February 6, 2023 2:59 PM
> To: Петрова Наталия Михайловна <n.petrova@fintech.ru>
> Cc: Dongsheng Yang <dongsheng.yang@easystack.cn>; Jens Axboe <axboe@kernel.dk>; ceph-devel@vger.kernel.org; linux-block@vger.kernel.org; linux-kernel@vger.kernel.org; lvc-project@linuxtesting.org; Alexey Khoroshilov <khoroshilov@ispras.ru>
> Subject: Re: [PATCH] rbd: avoid double free memory on error path in rbd_dev_create()
>
> On Fri, Feb 3, 2023 at 3:15 PM Natalia Petrova <n.petrova@fintech.ru> wrote:
>> If rbd_dev_create() fails after assignment 'opts' to 'rbd_dev->opts',
>> double free of 'rbd_options' happens:
>> one is in rbd_dev_free() and another one is in do_rbd_add().
>>
>> Found by Linux Verification Center (linuxtesting.org) with SVACE.
>>
>> Fixes: 1643dfa4c2c8 ("rbd: introduce a per-device ordered workqueue")
>> Signed-off-by: Natalia Petrova <n.petrova@fintech.ru>
>> Signed-off-by: Alexey Khoroshilov <khoroshilov@ispras.ru>
>> ---
>>   drivers/block/rbd.c | 2 +-
>>   1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/drivers/block/rbd.c b/drivers/block/rbd.c index
>> 04453f4a319c..ab6bfc352cde 100644
>> --- a/drivers/block/rbd.c
>> +++ b/drivers/block/rbd.c
>> @@ -5357,7 +5357,6 @@ static struct rbd_device *rbd_dev_create(struct rbd_client *rbdc,
>>          if (!rbd_dev)
>>                  return NULL;
>>
>> -       rbd_dev->opts = opts;
>>
>>          /* get an id and fill in device name */
>>          rbd_dev->dev_id = ida_simple_get(&rbd_dev_id_ida, 0, @@
>> -5372,6 +5371,7 @@ static struct rbd_device *rbd_dev_create(struct rbd_client *rbdc,
>>          if (!rbd_dev->task_wq)
>>                  goto fail_dev_id;
>>
>> +       rbd_dev->opts = opts;
>>          /* we have a ref from do_rbd_add() */
>>          __module_get(THIS_MODULE);
>>
>> --
>> 2.34.1
>>
> Hi Natalia,
>
> It seems like a similar issue is affecting rbd_dev->rbd_client and rbd_dev->spec.  Unlike rbd_dev->opts, they are ref-counted and I'm guessing that the verification tool doesn't go that deep.
>
> I'd prefer all three to be addressed in the same change, since it's the same error path.  Would you be willing to look into that and post a new revision or should I treat just this patch as a bug report?
>
> Thanks,
>
>                  Ilya