RE: [PATCH v2] ceph: fix deadlock bugs by making iput() calls asynchronous

[Viacheslav Dubeyko <Slava.Dubeyko@ibm.com>]

On Wed, 2025-09-17 at 21:25 +0200, Max Kellermann wrote:
> On Wed, Sep 17, 2025 at 9:20 PM Viacheslav Dubeyko
> <Slava.Dubeyko@ibm.com> wrote:
> > > > > +     WARN_ON_ONCE(!queue_work(ceph_inode_to_fs_client(inode)->inode_wq,
> > > > > +                              &ceph_inode(inode)->i_work));
> > > > 
> > > > This function looks like ceph_queue_inode_work() [1]. Can we use
> > > > ceph_queue_inode_work()?
> > > 
> > > No, we can not, because that function adds an inode reference (instead
> > > of donating the existing reference) and there's no way we can safely
> > > get rid of it (even if we would accept paying the overhead of two
> > > extra atomic operations).
> > 
> > This function can call iput() too. Should we rework it, then? Also, as a result,
> > we will have two similar functions. And it could be confusing.
> 
> No. NOT calling iput() is the whole point of my patch. Did you read
> the patch description?

This function can call the iput:

void ceph_queue_inode_work(struct inode *inode, int work_bit)
{
	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
	struct ceph_client *cl = fsc->client;
	struct ceph_inode_info *ci = ceph_inode(inode);
	set_bit(work_bit, &ci->i_work_mask);

	ihold(inode);
	if (queue_work(fsc->inode_wq, &ci->i_work)) {
		doutc(cl, "%p %llx.%llx mask=%lx\n", inode,
		      ceph_vinop(inode), ci->i_work_mask);
	} else {
		doutc(cl, "%p %llx.%llx already queued, mask=%lx\n",
		      inode, ceph_vinop(inode), ci->i_work_mask);
		iput(inode); <-- we can call iput here.
	}
}

I am citing you: "NOT calling iput() is the whole point of my patch". This
function can call iput(). And this is my question: Should we rework it, then?

Thanks,
Slava.