Re: [PATCH] aio: prevent the final fput() in the middle of vfs_poll() (Re: KASAN: use-after-free Read in unix_dgram_poll)

On Sun, Mar 03, 2019 at 02:23:33PM -0800, Linus Torvalds wrote:

OK, having dug through the archives, the reasons were not strong.
So that part is OK...

That reminds me - refcount_t here is rather ridiculous; what we
have is
	* everything other than aio_poll: ki_refcnt is 0 all along
	* aio_poll: originally 0, then set to 2, then two iocb_put()
are done (either both synchronous to aio_poll(), or one sync and one
async).

That's not a refcount at all.  It's a flag, set only for aio_poll ones.
And that test in iocb_put() is "if flag is set, clear it and bugger off".

What's worse, AFAICS the callers in aio_poll() are buggered (see below)

 static inline void aio_poll_complete(struct aio_kiocb *iocb, __poll_t mask)
 {
-	struct file *file = iocb->poll.file;
-
 	aio_complete(iocb, mangle_poll(mask), 0);
-	fput(file);
 }

No reasons to keep that function at all now...

-	if (unlikely(apt.error)) {
-		fput(req->file);
+	if (unlikely(apt.error))
 		return apt.error;
-	}
 
 	if (mask)
 		aio_poll_complete(aiocb, mask);

Looking at that thing...  How does it manage to avoid leaks
when we try to use it on e.g. /dev/tty, which has
        poll_wait(file, &tty->read_wait, wait);
        poll_wait(file, &tty->write_wait, wait);
in n_tty_poll()?

AFAICS, we'll succeed adding it to the first queue, then have
aio_poll_queue_proc() fail and set apt.error to -EINVAL.
Suppose we are looking for EPOLLIN and there's nothing ready
to read.  We'll go
        mask = vfs_poll(req->file, &apt.pt) & req->events;
mask is 0.
        if (unlikely(!req->head)) {
nope - it's &tty->read_wait, not NULL
                /* we did not manage to set up a waitqueue, done */
                goto out;
        }

        spin_lock_irq(&ctx->ctx_lock);
        spin_lock(&req->head->lock);
        if (req->woken) {
nope - no wakeups so far
                /* wake_up context handles the rest */
                mask = 0;
                apt.error = 0;
        } else if (mask || apt.error) {
apt.error is non-zero
                /* if we get an error or a mask we are done */
                WARN_ON_ONCE(list_empty(&req->wait.entry));
                list_del_init(&req->wait.entry);
OK, removed from queue
        } else {
                /* actually waiting for an event */
                list_add_tail(&aiocb->ki_list, &ctx->active_reqs);
                aiocb->ki_cancel = aio_poll_cancel;
        }
        spin_unlock(&req->head->lock);
        spin_unlock_irq(&ctx->ctx_lock);

out:
        if (unlikely(apt.error)) {
                fput(req->file);
                return apt.error;
... and we return -EINVAL to __io_submit_one(), where we hit
        /*
         * If ret is 0, we'd either done aio_complete() ourselves or have
         * arranged for that to be done asynchronously.  Anything non-zero
         * means that we need to destroy req ourselves.
         */
        if (ret)
                goto out_put_req;
        return 0;
out_put_req:
        if (req->ki_eventfd)
                eventfd_ctx_put(req->ki_eventfd);
        iocb_put(req);
out_put_reqs_available:
        put_reqs_available(ctx, 1);
        return ret;

and all knowledge of req is lost.  But we'd done only *one* iocb_put() in
that case, and ->ki_refcnt had been set to 2 by aio_poll().  How could it
avoid a leak?  The same goes for "->poll() returns something without
bothering to call poll_wait()" case, actually...

IOW, I would rather have aio_poll() (along with your fput-a-bit-later change)
do this -
out:
        if (mask && !apt.error)
                aio_complete(aiocb, mangle_poll(mask), 0);
        iocb_put(aiocb);
        return apt.error;

Comments?