On 19/08/2020 03:30, Laurent Pinchart wrote: > Hi Hans, > > On Mon, Aug 17, 2020 at 01:00:49PM +0200, Hans Verkuil wrote: >> On 17/08/2020 01:51, Laurent Pinchart wrote: >>> On Sun, Aug 16, 2020 at 08:54:18AM -0700, Guenter Roeck wrote: >>>> On 8/16/20 5:18 AM, Laurent Pinchart wrote: >>>>> CC'ing Hans Verkuil and Sakari Ailus for the discussion about handling >>>>> file operations and disconnect in V4L2. >>>>> >>>>> On Sat, Aug 15, 2020 at 05:33:15PM -0700, Guenter Roeck wrote: >>>>>> + linux-uvc-devel@xxxxxxxxxxxxxxxxxxxxx >>>>>> + linux-media@xxxxxxxxxxxxxxx >>>>>> + laurent.pinchart@xxxxxxxxxxxxxxxx >>>>>> >>>>>> and changed subject >>>>>> >>>>>> On Fri, Aug 14, 2020 at 10:07:39PM -0400, Alan Stern wrote: >>>>>>> On Fri, Aug 14, 2020 at 04:07:03PM -0700, Guenter Roeck wrote: >>>>>>>> Hi all, >>>>>>>> >>>>>>>> over time, there have been a number of reports of crashes in usb_ifnum_to_if(), >>>>>>>> called from usb_hcd_alloc_bandwidth, which is in turn called from usb_set_interface(). >>>>>>>> Examples are [1] [2] [3]. A typical backtrace is: >>>>>>>> >>>>>>>> <3>[ 3489.445468] intel_sst_acpi 808622A8:00: sst: Busy wait failed, cant send this msg >>>>>>>> <6>[ 3490.507273] usb 1-4: USB disconnect, device number 3 >>>>>>>> <1>[ 3490.516670] BUG: unable to handle kernel NULL pointer dereference at 0000000000000000 >>>>>>>> <6>[ 3490.516680] PGD 0 P4D 0 >>>>>>>> <4>[ 3490.516687] Oops: 0000 [#1] PREEMPT SMP PTI >>>>>>>> <4>[ 3490.516693] CPU: 0 PID: 5633 Comm: V4L2CaptureThre Not tainted 4.19.113-08536-g5d29ca36db06 #1 >>>>>>>> <4>[ 3490.516696] Hardware name: GOOGLE Edgar, BIOS Google_Edgar.7287.167.156 03/25/2019 >>>>>>>> <4>[ 3490.516706] RIP: 0010:usb_ifnum_to_if+0x29/0x40 >>>>>>>> <4>[ 3490.516710] Code: ee 0f 1f 44 00 00 55 48 89 e5 48 8b 8f f8 03 00 00 48 85 c9 74 27 44 0f b6 41 04 4d 85 c0 74 1d 31 ff 48 8b 84 f9 98 00 00 00 <48> 8b 10 0f b6 52 02 39 f2 74 0a 48 ff c7 4c 39 c7 72 e5 31 c0 5d >>>>>>>> <4>[ 3490.516714] RSP: 0018:ffffa46f42a47a80 EFLAGS: 00010246 >>>>>>>> <4>[ 3490.516718] RAX: 0000000000000000 RBX: 0000000000000000 RCX: ffff904a396c9000 >>>>>>>> <4>[ 3490.516721] RDX: ffff904a39641320 RSI: 0000000000000001 RDI: 0000000000000000 >>>>>>>> <4>[ 3490.516724] RBP: ffffa46f42a47a80 R08: 0000000000000002 R09: 0000000000000000 >>>>>>>> <4>[ 3490.516727] R10: 0000000000009975 R11: 0000000000000009 R12: 0000000000000000 >>>>>>>> <4>[ 3490.516731] R13: ffff904a396b3800 R14: ffff904a39e88000 R15: 0000000000000000 >>>>>>>> <4>[ 3490.516735] FS: 00007f396448e700(0000) GS:ffff904a3ba00000(0000) knlGS:0000000000000000 >>>>>>>> <4>[ 3490.516738] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 >>>>>>>> <4>[ 3490.516742] CR2: 0000000000000000 CR3: 000000016cb46000 CR4: 00000000001006f0 >>>>>>>> <4>[ 3490.516745] Call Trace: >>>>>>>> <4>[ 3490.516756] usb_hcd_alloc_bandwidth+0x1ee/0x30f >>>>>>>> <4>[ 3490.516762] usb_set_interface+0x1a3/0x2b7 >>>>>>>> <4>[ 3490.516773] uvc_video_start_transfer+0x29b/0x4b8 [uvcvideo] >>>>>>>> <4>[ 3490.516781] uvc_video_start_streaming+0x91/0xdd [uvcvideo] >>>>>>>> <4>[ 3490.516787] uvc_start_streaming+0x28/0x5d [uvcvideo] >>>>>>>> <4>[ 3490.516795] vb2_start_streaming+0x61/0x143 [videobuf2_common] >>>>>>>> <4>[ 3490.516801] vb2_core_streamon+0xf7/0x10f [videobuf2_common] >>>>>>>> <4>[ 3490.516807] uvc_queue_streamon+0x2e/0x41 [uvcvideo] >>>>>>>> <4>[ 3490.516814] uvc_ioctl_streamon+0x42/0x5c [uvcvideo] >>>>>>>> <4>[ 3490.516820] __video_do_ioctl+0x33d/0x42a >>>>>>>> <4>[ 3490.516826] video_usercopy+0x34e/0x5ff >>>>>>>> <4>[ 3490.516831] ? video_ioctl2+0x16/0x16 >>>>>>>> <4>[ 3490.516837] v4l2_ioctl+0x46/0x53 >>>>>>>> <4>[ 3490.516843] do_vfs_ioctl+0x50a/0x76f >>>>>>>> <4>[ 3490.516848] ksys_ioctl+0x58/0x83 >>>>>>>> <4>[ 3490.516853] __x64_sys_ioctl+0x1a/0x1e >>>>>>>> <4>[ 3490.516858] do_syscall_64+0x54/0xde >>>>>>>> >>>>>>>> I have been able to reproduce the problem on a Chromebook by strategically placing >>>>>>>> msleep() calls into usb_set_interface() and usb_disable_device(). Ultimately, the >>>>>>>> problem boils down to lack of protection against device removal in usb_set_interface() >>>>>>>> [and/or possibly other callers of usb_ifnum_to_if()]. >>>>>>>> >>>>>>>> Sequence of events is roughly as follows: >>>>>>>> >>>>>>>> - usb_set_interface() is called and proceeds to some point, possibly to >>>>>>>> mutex_lock(hcd->bandwidth_mutex); >>>>>>>> - Device removal event is detected, and usb_disable_device() is called >>>>>>> >>>>>>> At this point all interface drivers get unbound (their disconnect >>>>>>> routines are called). >>>>>>> >>>>>>>> - usb_disable_device() starts removing actconfig data. It has removed >>>>>>>> and cleared dev->actconfig->interface[i], but not dev->actconfig >>>>>>>> - usb_set_interface() calls usb_hcd_alloc_bandwidth(), which calls >>>>>>>> usb_ifnum_to_if() >>>>>>>> - In usb_ifnum_to_if(), dev->actconfig is not NULL, but >>>>>>>> dev->actconfig->interface[i] is NULL >>>>>>>> - crash >>>>>>>> >>>>>>>> Question is what we can do about this. Checking if dev->state != USB_STATE_NOTATTACHED >>>>>>>> in usb_ifnum_to_if() might be a possible approach, but strictly speaking it would >>>>>>>> still be racy since there is still no lock against device removal. I have not tried >>>>>>>> calling usb_lock_device() in usb_set_interface() - would that possibly be an option ? >>>>>>> >>>>>>> As far as I know, protecting against these races is the responsibility >>>>>>> of the USB interface drivers. They must make sure that their disconnect >>>>>>> routines block until all outstanding calls to usb_set_interface return >>>>>>> (in fact, until all outstanding device accesses have finished). >>>>>>> >>>>>>> For instance, in the log extract you showed, it's obvious that the >>>>>>> uvc_start_streaming routine was running after the disconnect routine had >>>>>>> returned, which looks like a bug in itself: Once the disconnect routine >>>>>>> returns, the driver is not supposed to try to access the device at all >>>>>>> because some other driver may now be bound to it. >>>>>>> >>>>>>> We can't just call usb_lock_device from within usb_set_interface, >>>>>>> because usb_set_interface is often called with that lock already held. >>>>>>> >>>>>> I had a closer look into the uvcvideo driver and compared it to other usb >>>>>> drivers, including drivers in drivers/media/usb/ which connect to the video >>>>>> subsystem. >>>>>> >>>>>> The usbvideo driver lacks protection against calls to uvc_disconnect() while >>>>> >>>>> Are you confusing usbvideo and uvcvideo ? Both exist, and uvcvideo would >>>>> have been called usbvideo if the former hadn't already been in use. >>>> >>>> Yes, sorry :-(. I am not sure how s/uvc/usb/ happened. >>> >>> No worries. >>> >>>>>> calls into file operations are ongoing. This is pretty widespread, and not >>>>>> even limited to file operations (for example, there is a worker which is only >>>>>> canceled in uvc_delete, not in ucv_disconnect). The existing protection only >>>>>> ensures that no file operations are started after the call to ucv_disconnect, >>>>>> but that is insufficient. >>>>>> >>>>>> Other drivers do have that protection and make sure that no usb operations >>>>>> can happen after the disconnect call. >>>>>> >>>>>> The only remedy I can see is to rework the usbvideo driver and add the >>>>>> necessary protections. At first glance, it looks like this may be a >>>>>> substantial amount of work. I'd sign up for that, but before I start, >>>>>> I would like to get input from the usbvideo community. Is such an effort >>>>>> already going on ? If yes, how can I help ? If not, is the problem >>>>>> understood and accepted ? Are there any ideas on how to solve it ? >>>>> >>>>> This is something that has been discussed before, and needs to be solved >>>>> in the V4L2 framework itself, not in individual drivers. Not only would >>>>> this avoid rolling out the same code manually everywhere (in different >>>>> incorrect ways, as races are difficult to solve and implementations are >>>>> more often wrong than right), but it will also avoid similar issues for >>>>> non-USB devices. >>>> >>>> You mean code that ensures that no user-space v4l2 operation is in progress >>>> after video_device_unregister / v4l2_device_unregister return ? I agree, >>>> that would simplify the necessary changes on the uvc side. >>> >>> I was thinking about adding a new function to be called from the >>> disconnect handler to implement the wait on end of userspace access, but >>> video_device_unregister() seems an even better idea. >>> v4l2_device_unregister() is probably not very useful as v4l2_device >>> isn't exposed to userspace, only video_device is (and v4l2_subdev and >>> media_device, but that's a different story, although probably still an >>> issue for the latter in the UVC driver). >> >> Actually, all that is needed is to take the ioctl serialization lock in the disconnect >> function. > > It's not just ioctls though, the other file operations also need to be > handled (read, write, mmap). Correct. And AFAIK all vb2-based drivers do take that serialization lock in the file ops. > >> See last paragraph in 1.4.1 here: >> >> https://hverkuil.home.xs4all.nl/spec/driver-api/v4l2-dev.html >> >> Since uvc uses its own lock, you need to take that one. > > Drivers that use their own lock do so to avoid serializing all ioctls. Let's agree to disagree :-) In my experience it is just too hard to keep track of locking and with very little advantages. You are the only developer that I know of that insists on doing your own locking. Luckily you are very good at your job, but everyone else uses the v4l2/vb2 core locking. > This means that different ioctls may be covered by different locks > (possibly with part of some ioctls running without locking). I don't > think we can just dismiss the issue saying those drivers need to > implement the disconnection manually. It would be much better to > integrate handling of userspace access with video_device_unregister() > like proposed above, as that will work for all drivers in a transparent > way. It would also be fairly simple to implement in the V4L2 core. I'm not really sure what you want. Should video_unregister_device() take the core lock (i.e. vdev->lock)? Regards, Hans > >>> We also have a v4l2_device_disconnect() function which is supposed to >>> handle hot-pluggable device disconnection, but it's fairly useless (I'd >>> even say harmful as it gives the illusion that hotplugging is correctly >>> handled, while in reality the media subsystem is plagged by hot-unplug >>> issues :-S). >> >> The v4l2_device_disconnect() is there to remove a v4l2_dev reference to >> the device that is about to be removed when the disconnect() exists. >> Otherwise v4l2_dev->dev would point to a missing device. >> >> However, I wonder if it is still needed: commit 236c5441d703 from 2011 added >> code to take a reference to v4l2_dev->dev in v4l2_device_register(). This >> should prevent the device from disappearing until v4l2_device_unregister() is >> called. I suspect that v4l2_device_disconnect() can be removed completely, and >> instead v4l2_device_unregister() just calls put_device(v4l2_dev->dev). >> >> I don't like v4l2_device_disconnect() either, so if this works, then that would >> be a nice simplification. >> >>>> I actually came from the other side - I assumed that there is a reason >>>> that is not already the case, and that the problem therefore has to be >>>> resolved on the driver side. >>>> >>>> So I guess the next question is: Is this already being addressed on the >>>> v4l2 side ? >>> >>> I'm not aware of anyone working on this. >>> >>>>> It shouldn't take more than two flags (to track user-space operations in >>>>> progress and disconnection), a spinlock and a wait queue entry. I'm not >>>>> sure if someone has already given it a try, and don't recall why this >>>>> hasn't been done yet, as it should be fairly straightforward. >>>>> >>>>> On the UVC side, the work queue probably has to be flushed in >>>>> uvc_disconnect(). I'd keep the destroy call in uvc_delete() though. >>>>> Please make sure to look for potential race conditions between the URB >>>>> completion handler and the .disconnect() handler (they shouldn't be any, >>>>> but I haven't checked lately myself). >>>> >>>> My current solution for this problem is to call uvc_ctrl_cleanup_device() >>>> from uvc_disconnect(), after uvc_unregister_video(). >>> >>> I'd rather avoid that, as the cleanup functions in the UVC driver are >>> generally meant to free memory when the last user disappears. While no >>> new userspace operation will be started after disconnection once the >>> above fix will be in place, there's one operation we can't avoid: the >>> file release. This will access some of the memory allocated by the >>> driver, and while the current implementation probably doesn't access in >>> .release() any memory freed by uvc_ctrl_cleanup_device(), I think it's a >>> good practice to only shut down the userspace API in .disconnect(), and >>> free memory when the last reference is released. >>> >>>> An alternative might >>>> be to add a uvc_ctrl_stop_device() function which would just cancel the >>>> worker. >>> >>> I think that would be best. Should stream->async_wq (in uvc_video.c) be >>> similarly flushed ? The driver does so in stream->async_wq(), called >>> from uvc_video_stop_transfer(), itself called from >>> uvc_video_stop_streaming() (among other places, that are either error >>> paths or system suspend handling). The call stack goes to >>> uvc_stop_streaming(), and, through the videobuf2 helpers, to >>> vb2_queue_release() called by uvc_queue_release() itself called by >>> uvc_v4l2_release() (in the non-disconnect case, >>> uvc_video_stop_streaming() will be called through videobuf2 by >>> uvc_queue_streamoff(), in response to a VIDIOC_STREAMOFF ioctl). We thus >>> flush the workqueue too late, and also access the device in >>> uvc_video_stop_streaming() long after .disconnect() returns. >>> >>> I think uvc_video_stop_streaming() could be called in uvc_disconnect() >>> after uvc_unregister_video(). >
