On 10/7/20 9:48 AM, Shuah Khan wrote:
On 10/7/20 8:28 AM, Andrey Konovalov wrote:
On Wed, Oct 7, 2020 at 3:56 PM Shuah Khan <skhan@xxxxxxxxxxxxxxxxxxx>
wrote:
[snip]
Hi Andrey,
I have been unable to reproduce the problem with the reproducer
so far. You mentioned it happens quite often.
- matched config with yours
- load vhci_hcd module and run the reproducer
Hm, if you matched the config, then the module should be built-in?
Right. I did notice that your config has built-in. This shouldn't
matter, I have a kernel built with it static. I will try it to
see if it makes a difference.
I do see the messages during shutdown - stop threads etc.
What am I missing?
This appears to be a race that requires precise timings. I failed to
reproduce it with the C reproducer, but I managed to reproduce it with
the syzkaller repro program:
https://syzkaller.appspot.com/x/repro.syz?x=16cbaa7d900000
To do that you need to build syzkaller, and copy ./bin/syz-execprog
and ./bin/syz-executor into your testing environment, and then do:
./syz-execprog -sandbox=none -repeat=0 -procs=6 ./repro.prog
Andrey,
I am unable to reproduce the problem. I even tweaked the reproducer
to launch 10 procs and 100 threads.
Can you test the following patch with your setup
Here it is - also attached.
=======================================================================
diff --git a/drivers/usb/usbip/vhci_hcd.c b/drivers/usb/usbip/vhci_hcd.c
index 66cde5e5f796..3fdcb93b667d 100644
--- a/drivers/usb/usbip/vhci_hcd.c
+++ b/drivers/usb/usbip/vhci_hcd.c
@@ -997,14 +997,12 @@ static void vhci_device_unlink_cleanup(struct
vhci_device *vdev)
spin_unlock_irqrestore(&vhci->lock, flags);
}
-/*
- * The important thing is that only one context begins cleanup.
- * This is why error handling and cleanup become simple.
- * We do not want to consider race condition as possible.
- */
static void vhci_shutdown_connection(struct usbip_device *ud)
{
struct vhci_device *vdev = container_of(ud, struct vhci_device, ud);
+ struct vhci_hcd *vhci_hcd = vdev_to_vhci_hcd(vdev);
+ struct vhci *vhci = vhci_hcd->vhci;
+ unsigned long flags;
/* need this? see stub_dev.c */
if (ud->tcp_socket) {
@@ -1012,6 +1010,10 @@ static void vhci_shutdown_connection(struct
usbip_device *ud)
kernel_sock_shutdown(ud->tcp_socket, SHUT_RDWR);
}
+ /* avoid races during shutdown */
+ spin_lock_irqsave(&vhci->lock, flags);
+ spin_lock(&vdev->priv_lock);
+
/* kill threads related to this sdev */
if (vdev->ud.tcp_rx) {
kthread_stop_put(vdev->ud.tcp_rx);
@@ -1031,6 +1033,10 @@ static void vhci_shutdown_connection(struct
usbip_device *ud)
}
pr_info("release socket\n");
+ /* unlock - vhci_device_unlink_cleanup() holds the locks */
+ spin_unlock(&vdev->priv_lock);
+ spin_unlock_irqrestore(&vhci->lock, flags);
+
vhci_device_unlink_cleanup(vdev);
/*
=======================================================================
thanks,
-- Shuah
diff --git a/drivers/usb/usbip/vhci_hcd.c b/drivers/usb/usbip/vhci_hcd.c
index 66cde5e5f796..3fdcb93b667d 100644
--- a/drivers/usb/usbip/vhci_hcd.c
+++ b/drivers/usb/usbip/vhci_hcd.c
@@ -997,14 +997,12 @@ static void vhci_device_unlink_cleanup(struct vhci_device *vdev)
spin_unlock_irqrestore(&vhci->lock, flags);
}
-/*
- * The important thing is that only one context begins cleanup.
- * This is why error handling and cleanup become simple.
- * We do not want to consider race condition as possible.
- */
static void vhci_shutdown_connection(struct usbip_device *ud)
{
struct vhci_device *vdev = container_of(ud, struct vhci_device, ud);
+ struct vhci_hcd *vhci_hcd = vdev_to_vhci_hcd(vdev);
+ struct vhci *vhci = vhci_hcd->vhci;
+ unsigned long flags;
/* need this? see stub_dev.c */
if (ud->tcp_socket) {
@@ -1012,6 +1010,10 @@ static void vhci_shutdown_connection(struct usbip_device *ud)
kernel_sock_shutdown(ud->tcp_socket, SHUT_RDWR);
}
+ /* avoid races during shutdown */
+ spin_lock_irqsave(&vhci->lock, flags);
+ spin_lock(&vdev->priv_lock);
+
/* kill threads related to this sdev */
if (vdev->ud.tcp_rx) {
kthread_stop_put(vdev->ud.tcp_rx);
@@ -1031,6 +1033,10 @@ static void vhci_shutdown_connection(struct usbip_device *ud)
}
pr_info("release socket\n");
+ /* unlock - vhci_device_unlink_cleanup() holds the locks */
+ spin_unlock(&vdev->priv_lock);
+ spin_unlock_irqrestore(&vhci->lock, flags);
+
vhci_device_unlink_cleanup(vdev);
/*
