linux-kernel.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
From: Al Viro <viro@zeniv.linux.org.uk>
To: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Eric Dumazet <eric.dumazet@gmail.com>,
	David Miller <davem@davemloft.net>,
	Jason Baron <jbaron@akamai.com>,
	kgraul@linux.ibm.com, ktkhai@virtuozzo.com,
	kyeongdon.kim@lge.com,
	Linux List Kernel Mailing <linux-kernel@vger.kernel.org>,
	Netdev <netdev@vger.kernel.org>,
	pabeni@redhat.com, syzkaller-bugs@googlegroups.com,
	xiyou.wangcong@gmail.com, Christoph Hellwig <hch@lst.de>
Subject: Re: [PATCH] aio: prevent the final fput() in the middle of vfs_poll() (Re: KASAN: use-after-free Read in unix_dgram_poll)
Date: Mon, 4 Mar 2019 02:36:18 +0000	[thread overview]
Message-ID: <20190304023618.GS2217@ZenIV.linux.org.uk> (raw)
In-Reply-To: <CAHk-=widuz9QBY-Xhpc1YHf=EV1uRGyOv-v-BNSa1zW-nCCC6Q@mail.gmail.com>

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...

> @@ -1060,6 +1071,8 @@ static inline void iocb_put(struct aio_kiocb *iocb)
>  {
>  	if (refcount_read(&iocb->ki_refcnt) == 0 ||
>  	    refcount_dec_and_test(&iocb->ki_refcnt)) {
> +		if (iocb->ki_filp)
> +			fput(iocb->ki_filp);
>  		percpu_ref_put(&iocb->ki_ctx->reqs);
>  		kmem_cache_free(kiocb_cachep, iocb);
>  	}

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?

  reply	other threads:[~2019-03-04  2:36 UTC|newest]

Thread overview: 53+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2019-03-03 10:22 KASAN: use-after-free Read in unix_dgram_poll syzbot
2019-03-03 13:55 ` Al Viro
2019-03-03 15:18   ` [PATCH] aio: prevent the final fput() in the middle of vfs_poll() (Re: KASAN: use-after-free Read in unix_dgram_poll) Al Viro
2019-03-03 18:37     ` Eric Dumazet
2019-03-03 19:44     ` Linus Torvalds
2019-03-03 20:13       ` Linus Torvalds
2019-03-03 20:30       ` Al Viro
2019-03-03 22:23         ` Linus Torvalds
2019-03-04  2:36           ` Al Viro [this message]
2019-03-04 21:22             ` Linus Torvalds
2019-03-07  0:03               ` [PATCH 1/8] aio: make sure file is pinned Al Viro
2019-03-07  0:03                 ` [PATCH 2/8] aio_poll_wake(): don't set ->woken if we ignore the wakeup Al Viro
2019-03-07  2:18                   ` Al Viro
2019-03-08 11:16                     ` zhengbin (A)
2019-03-07  0:03                 ` [PATCH 3/8] aio_poll(): sanitize the logics after vfs_poll(), get rid of leak on error Al Viro
2019-03-07  2:11                   ` zhengbin (A)
2019-03-07  0:03                 ` [PATCH 4/8] aio_poll(): get rid of weird refcounting Al Viro
2019-03-07  0:03                 ` [PATCH 5/8] make aio_read()/aio_write() return int Al Viro
2019-03-07  0:03                 ` [PATCH 6/8] move dropping ->ki_eventfd into iocb_put() Al Viro
2019-03-07  0:03                 ` [PATCH 7/8] deal with get_reqs_available() in aio_get_req() itself Al Viro
2019-03-07  0:03                 ` [PATCH 8/8] aio: move sanity checks and request allocation to io_submit_one() Al Viro
2019-03-07  0:23                 ` [PATCH 1/8] aio: make sure file is pinned Linus Torvalds
2019-03-07  0:41                   ` Al Viro
2019-03-07  0:48                     ` Al Viro
2019-03-07  1:20                       ` Al Viro
2019-03-07  1:30                         ` Linus Torvalds
2019-03-08  3:36                           ` Al Viro
2019-03-08 15:50                             ` Christoph Hellwig
2019-03-10  7:06                             ` Al Viro
2019-03-10  7:08                               ` [PATCH 1/8] pin iocb through aio Al Viro
2019-03-10  7:08                                 ` [PATCH 2/8] keep io_event in aio_kiocb Al Viro
2019-03-11 19:43                                   ` Christoph Hellwig
2019-03-11 21:17                                     ` Al Viro
2019-03-10  7:08                                 ` [PATCH 3/8] aio: store event at final iocb_put() Al Viro
2019-03-11 19:44                                   ` Christoph Hellwig
2019-03-11 21:13                                     ` Al Viro
2019-03-11 22:52                                       ` Al Viro
2019-03-10  7:08                                 ` [PATCH 4/8] Fix aio_poll() races Al Viro
2019-03-11 19:58                                   ` Christoph Hellwig
2019-03-11 21:06                                     ` Al Viro
2019-03-12 19:18                                       ` Christoph Hellwig
2019-03-10  7:08                                 ` [PATCH 5/8] make aio_read()/aio_write() return int Al Viro
2019-03-11 19:44                                   ` Christoph Hellwig
2019-03-10  7:08                                 ` [PATCH 6/8] move dropping ->ki_eventfd into iocb_destroy() Al Viro
2019-03-11 19:46                                   ` Christoph Hellwig
2019-03-10  7:08                                 ` [PATCH 7/8] deal with get_reqs_available() in aio_get_req() itself Al Viro
2019-03-11 19:46                                   ` Christoph Hellwig
2019-03-10  7:08                                 ` [PATCH 8/8] aio: move sanity checks and request allocation to io_submit_one() Al Viro
2019-03-11 19:48                                   ` Christoph Hellwig
2019-03-11 21:12                                     ` Al Viro
2019-03-11 19:41                                 ` [PATCH 1/8] pin iocb through aio Christoph Hellwig
2019-03-11 19:41                               ` [PATCH 1/8] aio: make sure file is pinned Christoph Hellwig
2019-03-04  7:53     ` [PATCH] aio: prevent the final fput() in the middle of vfs_poll() (Re: KASAN: use-after-free Read in unix_dgram_poll) Dmitry Vyukov

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20190304023618.GS2217@ZenIV.linux.org.uk \
    --to=viro@zeniv.linux.org.uk \
    --cc=davem@davemloft.net \
    --cc=eric.dumazet@gmail.com \
    --cc=hch@lst.de \
    --cc=jbaron@akamai.com \
    --cc=kgraul@linux.ibm.com \
    --cc=ktkhai@virtuozzo.com \
    --cc=kyeongdon.kim@lge.com \
    --cc=linux-kernel@vger.kernel.org \
    --cc=netdev@vger.kernel.org \
    --cc=pabeni@redhat.com \
    --cc=syzkaller-bugs@googlegroups.com \
    --cc=torvalds@linux-foundation.org \
    --cc=xiyou.wangcong@gmail.com \
    --subject='Re: [PATCH] aio: prevent the final fput() in the middle of vfs_poll() (Re: KASAN: use-after-free Read in unix_dgram_poll)' \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).