Re: [PATCH] uacce: unmap remaining mmapping from user space

[Xu Zaibo <xuzaibo@huawei.com>]

Hi,
On 2020/2/25 16:33, zhangfei wrote:
> Hi, Zaibo
>
> On 2020/2/24 下午3:17, Xu Zaibo wrote:
>>>   @@ -585,6 +595,13 @@ void uacce_remove(struct uacce_device *uacce)
>>>           cdev_device_del(uacce->cdev, &uacce->dev);
>>>       xa_erase(&uacce_xa, uacce->dev_id);
>>>       put_device(&uacce->dev);
>>> +
>>> +    /*
>>> +     * unmap remainning mapping from user space, preventing user still
>>> +     * access the mmaped area while parent device is already removed
>>> +     */
>>> +    if (uacce->inode)
>>> +        unmap_mapping_range(uacce->inode->i_mapping, 0, 0, 1);
>> Should we unmap them at the first of 'uacce_remove',  and before 
>> 'uacce_put_queue'?
>>
> We can do this,
> Though it does not matter, since user space can not interrupt kernel 
> function uacce_remove.
>
I think it matters :)

Image that the process holds the uacce queue is running(read and write 
the queue), then you do 'uacce_remove'.
The process is running(read and write the queue) well in the time 
between 'uacce_put_queue' and
'unmap_mapping_range', however, the queue with its DMA memory may be 
gotten and used by
other guys in this time, since you have released them in kernel. As a 
result, the running process will be a disaster.

cheers,
Zaibo

.


> Thanks
> .
>