On Thu, May 11, 2023 at 10:15:53AM +0800, Zhangfei Gao wrote: > The inode can be different in a container, for example, a docker and host > both open the same uacce parent device, which uses the same uacce struct > but different inode, so uacce->inode is not enough. > > What's worse, when docker stops, the inode will be destroyed as well, > causing use-after-free in uacce_remove. > > So use q->filep->f_mapping to replace uacce->inode->i_mapping. > @@ -574,12 +574,6 @@ void uacce_remove(struct uacce_device *uacce) > > if (!uacce) > return; > - /* > - * unmap remaining 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); > > /* > * uacce_fops_open() may be running concurrently, even after we remove > @@ -589,6 +583,8 @@ void uacce_remove(struct uacce_device *uacce) > mutex_lock(&uacce->mutex); > /* ensure no open queue remains */ > list_for_each_entry_safe(q, next_q, &uacce->queues, list) { > + struct file *filep = q->private_data; > + > /* > * Taking q->mutex ensures that fops do not use the defunct > * uacce->ops after the queue is disabled. > @@ -597,6 +593,12 @@ void uacce_remove(struct uacce_device *uacce) > uacce_put_queue(q); > mutex_unlock(&q->mutex); > uacce_unbind_queue(q); > + > + /* > + * unmap remaining mapping from user space, preventing user still > + * access the mmaped area while parent device is already removed > + */ > + unmap_mapping_range(filep->f_mapping, 0, 0, 1); IDGI. Going through uacce_queue instead of uacce_device is fine, but why bother with file *or* inode? Just store a reference to struct address_space in your uacce_queue and be done with that... Another problem in that driver is uacce_vma_close(); this if (vma->vm_pgoff < UACCE_MAX_REGION) qfr = q->qfrs[vma->vm_pgoff]; kfree(qfr); can't be right - you have q->qfrs left pointing to freed object. If nothing else, subsequent mmap() will fail with -EEXIST, won't it?