lkml.org 
[lkml]   [2019]   [Apr]   [23]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v2] Input: uinput: Avoid Object-Already-Free with a global lock
On Tue, Apr 23, 2019 at 12:51:13PM +0530, Mukesh Ojha wrote:
> On 4/23/2019 8:58 AM, dmitry.torokhov@gmail.com wrote:
> > On Fri, Apr 19, 2019 at 02:13:48PM +0530, Mukesh Ojha wrote:
> > > On 4/19/2019 12:41 PM, dmitry.torokhov@gmail.com wrote:
> > > > On Fri, Apr 19, 2019 at 12:17:44PM +0530, Mukesh Ojha wrote:
> > > > > On 4/18/2019 11:55 AM, Mukesh Ojha wrote:
> > > > > > On 4/18/2019 7:13 AM, dmitry.torokhov@gmail.com wrote:
> > > > > > > > On 4/10/2019 1:29 PM, Mukesh Ojha wrote:
> > > > > > > > > uinput_destroy_device() gets called from two places. In one place,
> > > > > > > > > uinput_ioctl_handler() where it is protected under a lock
> > > > > > > > > udev->mutex but there is no protection on udev device from freeing
> > > > > > > > > inside uinput_release().
> > > > > > > uinput_release() should be called when last file handle to the uinput
> > > > > > > instance is being dropped, so there should be no other users and thus we
> > > > > > > can't be racing with anyone.
> > > > > > Lets say an example where i am creating input device quite frequently
> > > > > >
> > > > > > [   97.836603] input: syz0 as /devices/virtual/input/input262
> > > > > > [   97.845589] input: syz0 as /devices/virtual/input/input261
> > > > > > [   97.849415] input: syz0 as /devices/virtual/input/input263
> > > > > > [   97.856479] input: syz0 as /devices/virtual/input/input264
> > > > > > [   97.936128] input: syz0 as /devices/virtual/input/input265
> > > > > >
> > > > > > e.g input265
> > > > > >
> > > > > > while input265 gets created [1] and handlers are getting registered on
> > > > > > that device*fput* gets called on
> > > > > > that device as user space got to know that input265 is created and its
> > > > > > reference is still 1(rare but possible).
> > > > Are you saying that there are 2 threads sharing the same file
> > > > descriptor, one issuing the registration ioctl while the other closing
> > > > the same fd?
> > > Dmitry,
> > >
> > > I don't have a the exact look inside the app here, but this looks like the
> > > same as it is able to do
> > > fput on the uinput device.
> > >
> > > FYI
> > > Syskaller app is running in userspace (which is for syscall fuzzing) on
> > > kernel which is enabled
> > > with various config fault injection, FAULT_INJECTION,FAIL_SLAB,
> > > FAIL_PAGEALLOC, KASAN etc.
> > Mukesh,
> >
> > We need to understand exactly the failure mode. I do not think that
> > introducing another mutex into uinput actually fixes the issue, as we do
> > not order mutex acquisition, so I think it is still possible for the
> > release function to acquire the mutex and run first, and then ioctl
> > would run with freed object.
>
> I have taken care this case from ioctl and release point of view.
>
> Even if the release gets called first it will make the
> file->private_data=NULL.
> and further call to ioctl will not be a problem as the check is already
> there.

Al, do we have any protections in VFS layer from userspace hanging onto
a file descriptor and calling ioctl() on it even as another thread
calls close() on the same fd?

Should the issue be solved by individual drivers, or more careful
accounting for currently running operations is needed at VFS layer?

Thanks.

--
Dmitry

\
 
 \ /
  Last update: 2019-04-23 10:50    [W:0.087 / U:1.140 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site