Re: usb: use-after-free write in usb_hcd_link_urb_to_ep

[Dmitry Vyukov <dvyukov@google.com>]

On Thu, Mar 23, 2017 at 4:22 PM, Dmitry Vyukov <dvyukov@google.com> wrote:
>> On Thu, 23 Mar 2017, Dmitry Vyukov wrote:
>>
>>> > Putting these together:
>>> >
>>> >         The memory was allocated in usb_internal_control_msg() line 93.
>>> >         The later events occurred within the call in line 100 to
>>> >         usb_start_wait_urb().
>>> >
>>> >         The invalid access occurred within usb_start_wait_urb() line 56.
>>> >
>>> >         The memory was deallocated within usb_start_wait_urb() line 78.
>>> >
>>> > Since these routines don't involve any loops or backward jumps, this
>>> > says that the invalid access occurred before the memory was
>>> > deallocated!  So why is it reported as a problem?
>>>
>>>
>>> My first guess would be that pid 3348 did 2 calls to open and the urb
>>> was somehow referenced across these calls. Is it possible?
>>
>> I don't think so.  The URB gets allocated and deallocated separately
>> for each call.  You can see this very plainly by reading the source
>> code for usb_internal_control_msg() and usb_start_wait_urb().
>>
>> It's possible that the same memory location was allocated and
>> deallocated for two different calls at different times.  That wouldn't
>> fool syzkaller, would it?
>
>
> Generally it does not fool KASAN because of heap memory quarantine.
> I will take a closer look tomorrow.
> Thanks for looking into this.


The bug looks real to me and it can be easily reproduced by executing:

mmap(&(0x7f0000000000/0xfff000)=nil, (0xfff000), 0x3, 0x32,
0xffffffffffffffff, 0x0)
syz_open_dev$usb(&(0x7f0000001000-0x15)="2f6465762f6275732f7573622f3030232f30302300",
0x1ff, 0x200)
syz_open_dev$usb(&(0x7f0000002000-0x15)="2f6465762f6275732f7573622f3030232f30302300",
0x3f, 0x0)

and failing 7-th malloc in the first one, this one:

 kzalloc include/linux/slab.h:663 [inline]
 rh_call_control drivers/usb/core/hcd.c:522 [inline]

>From the report:

(from the failure stack)
kmalloc in rh_call_control fails, and it straight returns -ENOMEM
without calling  usb_hcd_unlink_urb_from_ep(hcd, urb) leaving urb
linked into hcd (and that's probably the root problem)
consequently:
rh_urb_enqueue return -ENOMEM
usb_hcd_submit_urb does INIT_LIST_HEAD(&urb->urb_list), but the urb is
still linked, so we got corrupted list
(from the free stack)
eventually usb_start_wait_urb frees the still linked urb by calling usb_free_urb
(form the use-after-free stack)
the subsequent open links a new urb into the corrupted list in
usb_hcd_link_urb_to_ep
bang!