On Mon, Feb 10, 2020 at 4:11 AM Chengguang Xu <cgxu519@xxxxxxxxxxxx> wrote:
> static int ovl_mmap(struct file *file, struct vm_area_struct *vma)
> {
> struct file *realfile = file->private_data;
> const struct cred *old_cred;
> + struct inode *inode = file->f_inode;
> + struct ovl_copy_up_work ovl_cuw;
> + DEFINE_WAIT_BIT(wait, &ovl_cuw.flags, OVL_COPY_UP_PENDING);
> + wait_queue_head_t *wqh;
> int ret;
>
> + if (vma->vm_flags & MAP_SHARED &&
> + ovl_copy_up_shared(file_inode(file)->i_sb)) {
> + ovl_cuw.err = 0;
> + ovl_cuw.flags = 0;
> + ovl_cuw.dentry = file_dentry(file);
> + set_bit(OVL_COPY_UP_PENDING, &ovl_cuw.flags);
> +
> + wqh = bit_waitqueue(&ovl_cuw.flags, OVL_COPY_UP_PENDING);
> + prepare_to_wait(wqh, &wait.wq_entry, TASK_UNINTERRUPTIBLE);
> +
> + INIT_WORK(&ovl_cuw.work, ovl_copy_up_work_fn);
> + schedule_work(&ovl_cuw.work);
> +
> + schedule();
> + finish_wait(wqh, &wait.wq_entry);
This just hides the bad lock dependency, it does not remove it.
The solution we've come up with is arguably more complex, but it does
fix this properly: make overlay use its own address space operations
in case of a shared map.
Amir, I lost track, do you remember what's the status of this work?
Thanks,
Miklos
