Message-ID: <b14b296f-2e08-4edf-aeea-1c5b621e2d0c@google.com>
Date:   Fri, 6 Oct 2023 16:48:19 -0700
From:   Avichal Rakesh <arakesh@...gle.com>
To:     Michael Grzeschik <mgr@...gutronix.de>
Cc:     Laurent Pinchart <laurent.pinchart@...asonboard.com>,
        Daniel Scally <dan.scally@...asonboard.com>,
        Greg Kroah-Hartman <gregkh@...uxfoundation.org>,
        jchowdhary@...gle.com, etalvala@...gle.com,
        linux-usb@...r.kernel.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH v1 0/3] usb: gadget: uvc: stability fixes on STREAMOFF.



On 10/6/23 15:53, Michael Grzeschik wrote:
> On Fri, Oct 06, 2023 at 10:00:11AM -0700, Avichal Rakesh wrote:
>>
>>
>> On 10/5/23 15:05, Michael Grzeschik wrote:
>>> Hi Avichal,
>>>
>>> On Thu, Oct 05, 2023 at 11:30:32AM -0700, Avichal Rakesh wrote:
>>>> On 10/5/23 03:14, Michael Grzeschik wrote:
>>>>> On Thu, Oct 05, 2023 at 11:23:27AM +0300, Laurent Pinchart wrote:
>>>>>> On Tue, Oct 03, 2023 at 01:09:06PM +0200, Michael Grzeschik wrote:
>>>>>>> On Sat, Sep 30, 2023 at 11:48:18AM -0700, Avichal Rakesh wrote:
>>>>>>> > We have been seeing two main stability issues that uvc gadget driver
>>>>>>> > runs into when stopping streams:
>>>>>>> >  1. Attempting to queue usb_requests to a disabled usb_ep
>>>>>>> >  2. use-after-free issue for inflight usb_requests
>>>>>>> >
>>>>>>> > The three patches below fix the two issues above. Patch 1/3 fixes the
>>>>>>> > first issue, and Patch 2/3 and 3/3 fix the second issue.
>>>>>>> >
>>>>>>> > Avichal Rakesh (3):
>>>>>>> >   usb: gadget: uvc: prevent use of disabled endpoint
>>>>>>> >   usb: gadget: uvc: Allocate uvc_requests one at a time
>>>>>>> >   usb: gadget: uvc: Fix use-after-free for inflight usb_requests
>>>>>>> >
>>>>>>> > drivers/usb/gadget/function/f_uvc.c     |  11 +-
>>>>>>> > drivers/usb/gadget/function/f_uvc.h     |   2 +-
>>>>>>> > drivers/usb/gadget/function/uvc.h       |   6 +-
>>>>>>> > drivers/usb/gadget/function/uvc_v4l2.c  |  21 ++-
>>>>>>> > drivers/usb/gadget/function/uvc_video.c | 189 +++++++++++++++++-------
>>>>>>> > 5 files changed, 164 insertions(+), 65 deletions(-)
>>>>>>>
>>>>>>> These patches are not applying on gregkh/usb-testing since
>>>>>>> Greg did take my patches first. I have already rebased them.
>>>>>>
>>>>>> I think they got merged too soon :-( We could fix things on top, but
>>>>>> there's very little time to do so for v6.7.
>>>>>
>>>>> Agreed. I was jumping from one workaround to another one, since this
>>>>> is not easy to fix in a proper way. And still after this long discussion
>>>>> with Avichal I don't think we are there yet.
>>>>>
>>>>>
>>>>> So far the first two patches from Avichal look legit. But the overall
>>>>> Use-After-Free fix is yet to be done properly.
>>>>>
>>>>> The "abondoned" method he suggested is really bad to follow and will
>>>>> add too much complexity and will be hard to debug.
>>>>>
>>>>> IMHO it should be possible to introduce two cleanup pathes.
>>>>>
>>>>> One path would be in the uvc_cleanup_requests that will cleanup the
>>>>> requests that are actually not used in the controller and are registered
>>>>> in the req_free list.
>>>>>
>>>>> The second path would be the complete functions that are being run
>>>>> from the controller and will ensure that the cleanup will really free
>>>>> the requests from the controller after they were consumed.
>>>>>
>>>>> What do you think?
>>>>
>>>> I am not sure I follow. Patch 3/3 does exactly what you say here.
>>>
>>> Yes, it was just to summ up what the latest state of the idea was,
>>> so Laurent does not read the whole thread in detail. Sorry for not
>>> being clear enough about that.
>>
>> Whoops! Sorry about the misunderstanding!
>>
>>>
>>>> There are two cleanup paths:
>>>>  1. uvcg_video_disable cleans up only the requests in req_free, and
>>>>  2. complete handler cleans up the in-flight requests.
>>>>
>>>> The "abandoned" flag is simply to let the completion handler know
>>>> which requests to clean up and which ones to re-queue back to
>>>> the gadget driver.
>>>
>>> What I don't get is, why in the case of shutdown there needs to
>>> be something re-queued back to the gadget driver. There should not
>>> need to be any sort of barrier flag for the requests. Just the
>>> complete handler running past a barrier where it knows that the
>>> whole device is stopped. So every call on complete should then clean
>>> that exact request it is touching currently.
>>>
>>> I don't know where the extra complexity comes from.
>>
>> A lot of this complexity comes from assuming a back to back
>> STREAMOFF -> STREAMON sequence is possible where the gadget driver
>> doesn't have the time to clean up all in-flight usb_requests.
>> However, looking through the usb gadget APIs again, and it
>> looks like  usb_ep_disable enforces that all requests will
>> be sent back to the gadget driver before it returns.
> 
> Great!

Uhh...apologies, I will have to take this back. I've been
trying to use uvc->state as the condition for when completion
handler should clean up usb_requests, and I cannot figure 
out a way to do so cleanly.

The fundamental problem with using uvc->state is that it is
not protected by any locks. So there is no real way to
assert that its value has not changed between reading
uvc->state and acting on it.

Naively we can write something like the following in the 
completion handler:

void uvc_video_complete(...) {
    if (uvc->state != UVC_EVENT_STREAMING) {
        usb_ep_free_request(....);
    } else { 
        // handle usb_request normally
    }
}

But without any locks, there are no guarantees that 
uvc->state didn't mutate immediately after the if 
condition was checked, and the complete handler is 
handling a request that it should've freed instead
or vice-versa. This argument would hold for any logic 
we guard with uvc->state, making uvc->state effectively 
useless as a check for freeing memory.

We can work around it by either
1. Locking uvc->state with some driver level lock
   to ensure that we can trust the value of uvc->state
   at least for a little while, or
2. Using some other barrier condition that is protected by 
   another lock

If we go with (1), we'd have to add a lock around every
and every write to uvc->state, which isn't terrible, but
would require more testing to ensure that it doesn't
create any new deadlocks.

For (2), with the realization that usb_ep_disable flushes 
all requests, we can add a barrier in uvc_video, protected by 
req_lock. That should simplify the logic a little bit and 
will hopefully be easier to reason about.

I could of course be missing a simpler solution here,
and am happy to be wrong. So please let me know if you 
have any other ideas on how to guarantee such a check. 


> 
>> So you're right:
>> With Patch 1/3 in place, I think we can just guard on uvc->state
>> alone, because control requests are blocked until usb_ep_disable
>> is finished anyway. I'll upload v4 with the "is_abandoned"
>> flag removed and the checks simplified once I've verified the
>> fix locally.
>>
>> That should also remove any bookkeeping issues that may have
>> triggered the stack below.
> 
> I am currious if we should handle -ECONNRESET and -ESHUTDOWN in more
> detail in the completion handler and make sure that the request will not
> be added into the req_free list from there.
> 
> Will review your v4 then.

Appreciate the reviews, thank you!

> 
>>>> The other "complications" are around making sure we can trust
>>>> the values in an inherently racey situation. The reasoning
>>>> can admittedly be difficult to follow at a glance, which incidentally
>>>> is why I went with a simple to prove timed wait in the past
>>>> (https://lore.kernel.org/20230912041910.726442-3-arakesh@google.com).
>>>>
>>>> I am not suggesting we go back to a timed wait, but please do look
>>>> at the patch and let me know which parts don't make sense, or are
>>>> difficult to understand. We can add more documentation about our
>>>> assumptions there, or if you have a way to do this that you
>>>> think is simpler to reason about, then please let me know and I'll
>>>> be more than happy to use that!
>>>
>>> I really try to spin my head around the idea of the is_abondoned flag
>>> you are using. Unfortunatly for now I am out to debug the issues I see
>>> with your series.
>>>
>>> So I did try these patches you send. Yes the deadlock error is gone with
>>> v3. But the linked list is still running into cases where
>>> dwc3_gadget_giveback(complete) is touching requests that are already
>>> freed.
>>>
>>> [   61.408715] ------------[ cut here ]------------
>>> [   61.413897] kernel BUG at lib/list_debug.c:56!
>>> ...
>>> [   61.590762] Call trace:
>>> [   61.596890]  __list_del_entry_valid+0xb8/0xe8
>>> [   61.603408]  dwc3_gadget_giveback+0x3c/0x1b0
>>> [   61.607594]  dwc3_remove_requests.part.0+0xcc/0x100
>>> [   61.612948]  __dwc3_gadget_ep_disable+0xbc/0x1b8
>>> [   61.621019]  dwc3_gadget_ep_disable+0x48/0x100
>>> [   61.627925]  usb_ep_disable+0x3c/0x138
>>> [   61.638230]  uvc_function_setup_continue+0x3c/0x60
>>> [   61.645040]  uvc_v4l2_streamoff+0x5c/0x80
>>> [   61.659812]  v4l_streamoff+0x40/0x60
>>> [   61.668950]  __video_do_ioctl+0x344/0x420
>>> [   61.679548]  video_usercopy+0x1d0/0x788
>>> [   61.685677]  video_ioctl2+0x40/0x70
>>> [   61.697439]  v4l2_ioctl+0x68/0xa0
>>> [   61.709200]  __arm64_sys_ioctl+0x304/0xda0
>>> [   61.720768]  invoke_syscall.constprop.0+0x70/0x130

Just to confirm: this stack was with all 3 patches applied? 
Want to make sure this won't happen with v4.



Regards,
Avi.

Powered by blists - more mailing lists