Received-SPF: pass (google.com: domain of linux-bluetooth-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
MIME-Version: 1.0
References: <20210717000731.3836303-1-luiz.dentz@gmail.com> <7a0405c7-63ba-3e5c-7b22-503d7b326a11@i-love.sakura.ne.jp>
In-Reply-To: <7a0405c7-63ba-3e5c-7b22-503d7b326a11@i-love.sakura.ne.jp>
From:   Luiz Augusto von Dentz <luiz.dentz@gmail.com>
Date:   Sat, 17 Jul 2021 22:16:29 -0700
Message-ID: <CABBYNZLdYXHxD47xO+TX6U=MK1RtvuAAHLYvS1vTsuk4nFQfrg@mail.gmail.com>
Subject: Re: [RFC] Bluetooth: hci_sock: Fix calling lock_sock when handling HCI_DEV_UNREG
To:     Tetsuo Handa <penguin-kernel@i-love.sakura.ne.jp>
Cc:     "linux-bluetooth@vger.kernel.org" <linux-bluetooth@vger.kernel.org>,
        LinMa <linma@zju.edu.cn>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

Hi Tetsuo,

On Sat, Jul 17, 2021 at 7:22 PM Tetsuo Handa
<penguin-kernel@i-love.sakura.ne.jp> wrote:
>
> Luiz Augusto von Dentz wrote:
> > This removes the reference of hci_dev to hci_pinfo since the reference
> > cannot prevent hdev to be freed hci_pinfo only keeps the index so in
> > case the device is unregistered and freed hci_dev_get will return NULL
> > thus prevent UAF.
>
> I'm not convinced that this change is safe.
>
> vhci_release() (which will call hci_unregister_dev(hdev)) is called when
> refcount to /dev/vchi dropped to 0. That is, vhci_release() might be called
> while e.g. hci_sock_bound_ioctl() is in progress.
>
> Since hci_unregister_dev(hdev) calls list_del(&hdev->list) with hci_dev_list_lock
> held for write, hci_dev_get(hci_pi(sk)->dev) which scans hci_dev_list with
> hci_dev_list_lock held for read will start returning NULL. But I think that
> this change introduces two race windows.
>
> Since hci_unregister_dev(hdev) then calls hci_sock_dev_event(hdev, HCI_DEV_UNREG)
> and finally calls ida_simple_remove(&hci_index_ida, id), subsequent
> hci_register_dev(struct hci_dev *hdev) call can re-assign the id which
> hci_pi(sk)->dev is tracking, by calling ida_simple_get() and finally calling
> list_add(&hdev->list, &hci_dev_list) with hci_dev_list_lock held for write.

We can perform a pointer comparison if that makes you happy. Anyway I
don't know if you guys have realized already but this is probably
impossible to reproduce with real hardware since the
registration/unregistration comes for other buses no memfault would
hold the thread that long for unplugging and replugging a device on
the bus, vhci is usually only used for emulation/testing/CI, not sure
who got the idea that finding vulnerabilities on our emulator would be
a great feat.

> Therefore, I think that first race window is that
>
> +       /* Commands may use copy_from_user which is unsafe while holding hdev as
> +        * it could be unregistered in the meantime.
> +        */
> +       hci_dev_put(hdev);
> +       hdev = NULL;
>
> causes hci_sock_bound_ioctl() to check flags on an intended hdev and
> e.g. hci_sock_reject_list_add() to operate on an unintended hdev.
>
> Also, even if hci_sock_bound_ioctl() and hci_sock_reject_list_add() reached
> the same hdev, I think that there still is second race window that
>
>   hci_unregister_dev() {                       hci_sock_reject_list_add() {
>                                                  hdev = hci_dev_get(hci_pi(sk)->dev);
>     write_lock(&hci_dev_list_lock);
>     list_del(&hdev->list);
>     write_unlock(&hci_dev_list_lock);
>
>     hci_sock_dev_event(hdev, HCI_DEV_UNREG);
>
>     hci_dev_lock(hdev);
>     hci_bdaddr_list_clear(&hdev->reject_list);
>     hci_dev_unlock(hdev);
>                                                  hci_dev_lock(hdev);
>                                                  err = hci_bdaddr_list_add(&hdev->reject_list, &bdaddr, BDADDR_BREDR); // <= Adding after clear all; at least memory leak.
>                                                  hci_dev_unlock(hdev);
>                                                  hci_dev_put(hdev);
>   }
>
> . That is, an attempt to replace pointer reference with index number is racy.
>
> After all, hci_sock_dev_event(hdev, HCI_DEV_UNREG) is responsible for
> waiting for already started e.g. hci_sock_reject_list_add().

Both blocks do require hci_dev_lock, so if the second block had
acquired the lock isn't it obvious that the first won't execute until
hci_dev_unlock is complete? Anyway after all these discussion Im even
more convinced that the real problem lies in hci_dev_get/hold, after
all references are usually used to prevent the objects to be freed but
in this case it doesn't and no locking will gonna fix that.

-- 
Luiz Augusto von Dentz