2023-01-05 14:50:06

by Ricardo Ribalda

[permalink] [raw]
Subject: [PATCH v7] media: uvcvideo: Fix race condition with usb_kill_urb

usb_kill_urb warranties that all the handlers are finished when it
returns, but does not protect against threads that might be handling
asynchronously the urb.

For UVC, the function uvc_ctrl_status_event_async() takes care of
control changes asynchronously.

If the code is executed in the following order:

CPU 0 CPU 1
===== =====
uvc_status_complete()
uvc_status_stop()
uvc_ctrl_status_event_work()
uvc_status_start() -> FAIL

Then uvc_status_start will keep failing and this error will be shown:

<4>[ 5.540139] URB 0000000000000000 submitted while active
drivers/usb/core/urb.c:378 usb_submit_urb+0x4c3/0x528

Let's improve the current situation, by not re-submiting the urb if
we are stopping the status event. Also process the queued work
(if any) during stop.

CPU 0 CPU 1
===== =====
uvc_status_complete()
uvc_status_stop()
uvc_status_start()
uvc_ctrl_status_event_work() -> FAIL

Hopefully, with the usb layer protection this should be enough to cover
all the cases.

Cc: [email protected]
Fixes: e5225c820c05 ("media: uvcvideo: Send a control event when a Control Change interrupt arrives")
Reviewed-by: Yunke Cao <[email protected]>
Signed-off-by: Ricardo Ribalda <[email protected]>
---
uvc: Fix race condition on uvc

Make sure that all the async work is finished when we stop the status urb.

To: Hillf Danton <[email protected]>
To: Yunke Cao <[email protected]>
To: Sergey Senozhatsky <[email protected]>
To: Max Staudt <[email protected]>
To: Laurent Pinchart <[email protected]>
To: Mauro Carvalho Chehab <[email protected]>
Cc: [email protected]
Cc: [email protected]
---
Changes in v7:
- Use smp_store_release. (Thanks Hilf!)
- Rebase on top of uvc/next.
- Link to v6: https://lore.kernel.org/r/[email protected]

Changes in v6:
- Improve comments. (Thanks Laurent).
- Use true/false instead of 1/0 (Thanks Laurent).
- Link to v5: https://lore.kernel.org/r/[email protected]

Changes in v5:
- atomic_t do not impose barriers, use smp_mb() instead. (Thanks Laurent)
- Add an extra cancel_work_sync().
- Link to v4: https://lore.kernel.org/r/[email protected]

Changes in v4:
- Replace bool with atomic_t to avoid compiler reordering.
- First complete the async work and then kill the urb to avoid race (Thanks Laurent!)
- Link to v3: https://lore.kernel.org/r/[email protected]

Changes in v3:
- Remove the patch for dev->status, makes more sense in another series, and makes
the zero day less nervous.
- Update reviewed-by (thanks Yunke!).
- Link to v2: https://lore.kernel.org/r/[email protected]

Changes in v2:
- Add a patch for not kalloc dev->status
- Redo the logic mechanism, so it also works with suspend (Thanks Yunke!)
- Link to v1: https://lore.kernel.org/r/[email protected]
---
drivers/media/usb/uvc/uvc_ctrl.c | 5 +++++
drivers/media/usb/uvc/uvc_status.c | 33 +++++++++++++++++++++++++++++++++
drivers/media/usb/uvc/uvcvideo.h | 1 +
3 files changed, 39 insertions(+)

diff --git a/drivers/media/usb/uvc/uvc_ctrl.c b/drivers/media/usb/uvc/uvc_ctrl.c
index e07b56bbf853..30c417768376 100644
--- a/drivers/media/usb/uvc/uvc_ctrl.c
+++ b/drivers/media/usb/uvc/uvc_ctrl.c
@@ -6,6 +6,7 @@
* Laurent Pinchart ([email protected])
*/

+#include <asm/barrier.h>
#include <linux/bitops.h>
#include <linux/kernel.h>
#include <linux/list.h>
@@ -1509,6 +1510,10 @@ static void uvc_ctrl_status_event_work(struct work_struct *work)

uvc_ctrl_status_event(w->chain, w->ctrl, w->data);

+ /* The barrier is needed to synchronize with uvc_status_stop(). */
+ if (smp_load_acquire(&dev->flush_status))
+ return;
+
/* Resubmit the URB. */
w->urb->interval = dev->int_ep->desc.bInterval;
ret = usb_submit_urb(w->urb, GFP_KERNEL);
diff --git a/drivers/media/usb/uvc/uvc_status.c b/drivers/media/usb/uvc/uvc_status.c
index 602830a8023e..21e13b8441da 100644
--- a/drivers/media/usb/uvc/uvc_status.c
+++ b/drivers/media/usb/uvc/uvc_status.c
@@ -6,6 +6,7 @@
* Laurent Pinchart ([email protected])
*/

+#include <asm/barrier.h>
#include <linux/kernel.h>
#include <linux/input.h>
#include <linux/slab.h>
@@ -311,5 +312,37 @@ int uvc_status_start(struct uvc_device *dev, gfp_t flags)

void uvc_status_stop(struct uvc_device *dev)
{
+ struct uvc_ctrl_work *w = &dev->async_ctrl;
+
+ /*
+ * Prevent the asynchronous control handler from requeing the URB. The
+ * barrier is needed so the flush_status change is visible to other
+ * CPUs running the asynchronous handler before usb_kill_urb() is
+ * called below.
+ */
+ smp_store_release(&dev->flush_status, true);
+
+ /* If there is any status event on the queue, process it. */
+ if (cancel_work_sync(&w->work))
+ uvc_ctrl_status_event(w->chain, w->ctrl, w->data);
+
+ /* Kill the urb. */
usb_kill_urb(dev->int_urb);
+
+ /*
+ * The URB completion handler may have queued asynchronous work. This
+ * won't resubmit the URB as flush_status is set, but it needs to be
+ * cancelled before returning or it could then race with a future
+ * uvc_status_start() call.
+ */
+ if (cancel_work_sync(&w->work))
+ uvc_ctrl_status_event(w->chain, w->ctrl, w->data);
+
+ /*
+ * From this point, there are no events on the queue and the status URB
+ * is dead, this is, no events will be queued until uvc_status_start()
+ * is called. The barrier is needed to make sure that it is written to
+ * memory before uvc_status_start() is called again.
+ */
+ smp_store_release(&dev->flush_status, false);
}
diff --git a/drivers/media/usb/uvc/uvcvideo.h b/drivers/media/usb/uvc/uvcvideo.h
index ae0066eceffd..b2b277cccbdb 100644
--- a/drivers/media/usb/uvc/uvcvideo.h
+++ b/drivers/media/usb/uvc/uvcvideo.h
@@ -578,6 +578,7 @@ struct uvc_device {
struct usb_host_endpoint *int_ep;
struct urb *int_urb;
struct uvc_status *status;
+ bool flush_status;

struct input_dev *input;
char input_phys[64];

---
base-commit: fb1316b0ff3fc3cd98637040ee17ab7be753aac7
change-id: 20221212-uvc-race-09276ea68bf8

Best regards,
--
Ricardo Ribalda <[email protected]>


2023-01-07 00:59:48

by Laurent Pinchart

[permalink] [raw]
Subject: Re: [PATCH v7] media: uvcvideo: Fix race condition with usb_kill_urb

Hi Ricardo,

Thank you for the patch.

On Thu, Jan 05, 2023 at 03:31:29PM +0100, Ricardo Ribalda wrote:
> usb_kill_urb warranties that all the handlers are finished when it
> returns, but does not protect against threads that might be handling
> asynchronously the urb.
>
> For UVC, the function uvc_ctrl_status_event_async() takes care of
> control changes asynchronously.
>
> If the code is executed in the following order:
>
> CPU 0 CPU 1
> ===== =====
> uvc_status_complete()
> uvc_status_stop()
> uvc_ctrl_status_event_work()
> uvc_status_start() -> FAIL
>
> Then uvc_status_start will keep failing and this error will be shown:
>
> <4>[ 5.540139] URB 0000000000000000 submitted while active
> drivers/usb/core/urb.c:378 usb_submit_urb+0x4c3/0x528
>
> Let's improve the current situation, by not re-submiting the urb if
> we are stopping the status event. Also process the queued work
> (if any) during stop.
>
> CPU 0 CPU 1
> ===== =====
> uvc_status_complete()
> uvc_status_stop()
> uvc_status_start()
> uvc_ctrl_status_event_work() -> FAIL
>
> Hopefully, with the usb layer protection this should be enough to cover
> all the cases.

For some reason the word "hopefully" in a bug fix doesn't make me very
hopeful ;-)

> Cc: [email protected]
> Fixes: e5225c820c05 ("media: uvcvideo: Send a control event when a Control Change interrupt arrives")
> Reviewed-by: Yunke Cao <[email protected]>
> Signed-off-by: Ricardo Ribalda <[email protected]>
> ---
> uvc: Fix race condition on uvc
>
> Make sure that all the async work is finished when we stop the status urb.
>
> To: Hillf Danton <[email protected]>
> To: Yunke Cao <[email protected]>
> To: Sergey Senozhatsky <[email protected]>
> To: Max Staudt <[email protected]>
> To: Laurent Pinchart <[email protected]>
> To: Mauro Carvalho Chehab <[email protected]>
> Cc: [email protected]
> Cc: [email protected]
> ---
> Changes in v7:
> - Use smp_store_release. (Thanks Hilf!)
> - Rebase on top of uvc/next.
> - Link to v6: https://lore.kernel.org/r/[email protected]
>
> Changes in v6:
> - Improve comments. (Thanks Laurent).
> - Use true/false instead of 1/0 (Thanks Laurent).
> - Link to v5: https://lore.kernel.org/r/[email protected]
>
> Changes in v5:
> - atomic_t do not impose barriers, use smp_mb() instead. (Thanks Laurent)
> - Add an extra cancel_work_sync().
> - Link to v4: https://lore.kernel.org/r/[email protected]
>
> Changes in v4:
> - Replace bool with atomic_t to avoid compiler reordering.
> - First complete the async work and then kill the urb to avoid race (Thanks Laurent!)
> - Link to v3: https://lore.kernel.org/r/[email protected]
>
> Changes in v3:
> - Remove the patch for dev->status, makes more sense in another series, and makes
> the zero day less nervous.
> - Update reviewed-by (thanks Yunke!).
> - Link to v2: https://lore.kernel.org/r/[email protected]
>
> Changes in v2:
> - Add a patch for not kalloc dev->status
> - Redo the logic mechanism, so it also works with suspend (Thanks Yunke!)
> - Link to v1: https://lore.kernel.org/r/[email protected]
> ---
> drivers/media/usb/uvc/uvc_ctrl.c | 5 +++++
> drivers/media/usb/uvc/uvc_status.c | 33 +++++++++++++++++++++++++++++++++
> drivers/media/usb/uvc/uvcvideo.h | 1 +
> 3 files changed, 39 insertions(+)
>
> diff --git a/drivers/media/usb/uvc/uvc_ctrl.c b/drivers/media/usb/uvc/uvc_ctrl.c
> index e07b56bbf853..30c417768376 100644
> --- a/drivers/media/usb/uvc/uvc_ctrl.c
> +++ b/drivers/media/usb/uvc/uvc_ctrl.c
> @@ -6,6 +6,7 @@
> * Laurent Pinchart ([email protected])
> */
>
> +#include <asm/barrier.h>
> #include <linux/bitops.h>
> #include <linux/kernel.h>
> #include <linux/list.h>
> @@ -1509,6 +1510,10 @@ static void uvc_ctrl_status_event_work(struct work_struct *work)
>
> uvc_ctrl_status_event(w->chain, w->ctrl, w->data);
>
> + /* The barrier is needed to synchronize with uvc_status_stop(). */
> + if (smp_load_acquire(&dev->flush_status))
> + return;
> +
> /* Resubmit the URB. */
> w->urb->interval = dev->int_ep->desc.bInterval;
> ret = usb_submit_urb(w->urb, GFP_KERNEL);
> diff --git a/drivers/media/usb/uvc/uvc_status.c b/drivers/media/usb/uvc/uvc_status.c
> index 602830a8023e..21e13b8441da 100644
> --- a/drivers/media/usb/uvc/uvc_status.c
> +++ b/drivers/media/usb/uvc/uvc_status.c
> @@ -6,6 +6,7 @@
> * Laurent Pinchart ([email protected])
> */
>
> +#include <asm/barrier.h>
> #include <linux/kernel.h>
> #include <linux/input.h>
> #include <linux/slab.h>
> @@ -311,5 +312,37 @@ int uvc_status_start(struct uvc_device *dev, gfp_t flags)
>
> void uvc_status_stop(struct uvc_device *dev)
> {
> + struct uvc_ctrl_work *w = &dev->async_ctrl;
> +
> + /*
> + * Prevent the asynchronous control handler from requeing the URB. The
> + * barrier is needed so the flush_status change is visible to other
> + * CPUs running the asynchronous handler before usb_kill_urb() is
> + * called below.
> + */
> + smp_store_release(&dev->flush_status, true);
> +
> + /* If there is any status event on the queue, process it. */

/*
* Cancel any pending asynchronous work. If any status event was queued,
* process it synchronously.
*/

> + if (cancel_work_sync(&w->work))
> + uvc_ctrl_status_event(w->chain, w->ctrl, w->data);
> +
> + /* Kill the urb. */
> usb_kill_urb(dev->int_urb);
> +
> + /*
> + * The URB completion handler may have queued asynchronous work. This
> + * won't resubmit the URB as flush_status is set, but it needs to be
> + * cancelled before returning or it could then race with a future
> + * uvc_status_start() call.
> + */
> + if (cancel_work_sync(&w->work))
> + uvc_ctrl_status_event(w->chain, w->ctrl, w->data);
> +
> + /*
> + * From this point, there are no events on the queue and the status URB
> + * is dead, this is, no events will be queued until uvc_status_start()
> + * is called. The barrier is needed to make sure that it is written to
> + * memory before uvc_status_start() is called again.

With data races, the concept of "written to memory" doesn't make much
sense anymore.

* From this point, there are no events on the queue and the status URB
* is dead. No events will be queued until uvc_status_start() is called.
* The barrier is needed to make sure that flush_status is visible to
* uvc_ctrl_status_event_work() when uvc_status_start() will be called
* again.

I'll update the comments locally.

Reviewed-by: Laurent Pinchart <[email protected]>

> + */
> + smp_store_release(&dev->flush_status, false);
> }
> diff --git a/drivers/media/usb/uvc/uvcvideo.h b/drivers/media/usb/uvc/uvcvideo.h
> index ae0066eceffd..b2b277cccbdb 100644
> --- a/drivers/media/usb/uvc/uvcvideo.h
> +++ b/drivers/media/usb/uvc/uvcvideo.h
> @@ -578,6 +578,7 @@ struct uvc_device {
> struct usb_host_endpoint *int_ep;
> struct urb *int_urb;
> struct uvc_status *status;
> + bool flush_status;
>
> struct input_dev *input;
> char input_phys[64];
>
> ---
> base-commit: fb1316b0ff3fc3cd98637040ee17ab7be753aac7
> change-id: 20221212-uvc-race-09276ea68bf8

--
Regards,

Laurent Pinchart