On Thu, Aug 08, 2019 at 01:34:08PM -0400, Alan Stern wrote:
> The syzbot fuzzer found a lockdep violation in the rio500 driver:
>
> ======================================================
> WARNING: possible circular locking dependency detected
> 5.3.0-rc2+ #23 Not tainted
> ------------------------------------------------------
> syz-executor.2/20386 is trying to acquire lock:
> 00000000772249c6 (rio500_mutex){+.+.}, at: open_rio+0x16/0xc0
> drivers/usb/misc/rio500.c:64
>
> but task is already holding lock:
> 00000000d3e8f4b9 (minor_rwsem){++++}, at: usb_open+0x23/0x270
> drivers/usb/core/file.c:39
>
> which lock already depends on the new lock.
>
> The problem is that the driver's open_rio() routine is called while
> the usbcore's minor_rwsem is locked for reading, and it acquires the
> rio500_mutex; whereas conversely, probe_rio() and disconnect_rio()
> first acquire the rio500_mutex and then call usb_register_dev() or
> usb_deregister_dev(), which lock minor_rwsem for writing.
>
> The correct ordering of acquisition should be: minor_rwsem first, then
> rio500_mutex (since the locking in open_rio() cannot be changed).
> Thus, the probe and disconnect routines should avoid holding
> rio500_mutex while doing their registration and deregistration.
>
> This patch adjusts the code in those two routines to do just that. It
> also relies on the fact that the probe and disconnect routines are
> protected by the device mutex, so the initial test of rio->present
> needs no extra locking.
>
> Reported-by: [email protected]
> Signed-off-by: Alan Stern <[email protected]>
> Fixes: d710734b0677 ("USB: rio500: simplify locking")
> CC: Oliver Neukum <[email protected]>
> CC: <[email protected]>
>
> ---
>
> This patch is different from the one I posted earlier. I realized that
> we don't want to register the device's char file until after the
> buffers have been allocated.
Should I revert Oliver's patch?
confused,
greg k-h
