On Thu 16-04-20 14:08:59, Chengguang Xu wrote:
> ---- 在 星期四, 2020-04-16 03:19:50 Miklos Szeredi <miklos@xxxxxxxxxx> 撰写 ----
> > On Mon, Feb 10, 2020 at 4:10 AM Chengguang Xu <cgxu519@xxxxxxxxxxxx> wrote:
> > > +void ovl_evict_inode(struct inode *inode)
> > > +{
> > > + struct ovl_inode *oi = OVL_I(inode);
> > > + struct ovl_write_inode_work ovl_wiw;
> > > + DEFINE_WAIT_BIT(wait, &oi->flags, OVL_WRITE_INODE_PENDING);
> > > + wait_queue_head_t *wqh;
> > > +
> > > + if (ovl_inode_upper(inode)) {
> > > + if (current->flags & PF_MEMALLOC) {
> > > + spin_lock(&inode->i_lock);
> > > + ovl_set_flag(OVL_WRITE_INODE_PENDING, inode);
> > > + wqh = bit_waitqueue(&oi->flags,
> > > + OVL_WRITE_INODE_PENDING);
> > > + prepare_to_wait(wqh, &wait.wq_entry,
> > > + TASK_UNINTERRUPTIBLE);
> > > + spin_unlock(&inode->i_lock);
> > > +
> > > + ovl_wiw.inode = inode;
> > > + INIT_WORK(&ovl_wiw.work, ovl_write_inode_work_fn);
> > > + schedule_work(&ovl_wiw.work);
> > > +
> > > + schedule();
> > > + finish_wait(wqh, &wait.wq_entry);
> >
> > What is the reason to do this in another thread if this is a PF_MEMALLOC task?
>
> Some underlying filesystems(for example ext4) check the flag in
> ->write_inode() and treate it as an abnormal case.(warn and return)
>
> ext4_write_inode():
> if (WARN_ON_ONCE(current->flags & PF_MEMALLOC) ||
> sb_rdonly(inode->i_sb))
> return 0;
>
> overlayfs inodes are always keeping clean even after wring/modifying
> upperfile , so they are right target of kswapd but in the point of lower
> layer, ext4 just thinks kswapd is choosing a wrong dirty inode to reclam
> memory.
In ext4, it isn't a big problem if ext4_write_inode() is called from
kswapd. But if ext4_write_inode() is called from direct reclaim (which also
sets PF_MEMALLOC) we can deadlock because we may wait for transaction
commit and transaction commit may require locks (such as page lock or
waiting for page writeback to complete) which are held by the task
currently in direct reclaim. Your push to workqueue will silence the
warning but does not solve the possible deadlock.
I'm actually not sure why you need to writeback the upper inode when
reclaiming overlayfs inode. Why not leave it on flush worker on upper fs?
Honza
--
Jan Kara <jack@xxxxxxxx>
SUSE Labs, CR
