usb: core: hub: hub_port_init lock controller instead of bus
authorChris Bainbridge <chris.bainbridge@gmail.com>
Mon, 25 Apr 2016 12:48:38 +0000 (13:48 +0100)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Thu, 28 Apr 2016 19:40:46 +0000 (12:40 -0700)
The XHCI controller presents two USB buses to the system - one for USB2
and one for USB3. The hub init code (hub_port_init) is reentrant but
only locks one bus per thread, leading to a race condition failure when
two threads attempt to simultaneously initialise a USB2 and USB3 device:

[    8.034843] xhci_hcd 0000:00:14.0: Timeout while waiting for setup device command
[   13.183701] usb 3-3: device descriptor read/all, error -110

On a test system this failure occurred on 6% of all boots.

The call traces at the point of failure are:

Call Trace:
 [<ffffffff81b9bab7>] schedule+0x37/0x90
 [<ffffffff817da7cd>] usb_kill_urb+0x8d/0xd0
 [<ffffffff8111e5e0>] ? wake_up_atomic_t+0x30/0x30
 [<ffffffff817dafbe>] usb_start_wait_urb+0xbe/0x150
 [<ffffffff817db10c>] usb_control_msg+0xbc/0xf0
 [<ffffffff817d07de>] hub_port_init+0x51e/0xb70
 [<ffffffff817d4697>] hub_event+0x817/0x1570
 [<ffffffff810f3e6f>] process_one_work+0x1ff/0x620
 [<ffffffff810f3dcf>] ? process_one_work+0x15f/0x620
 [<ffffffff810f4684>] worker_thread+0x64/0x4b0
 [<ffffffff810f4620>] ? rescuer_thread+0x390/0x390
 [<ffffffff810fa7f5>] kthread+0x105/0x120
 [<ffffffff810fa6f0>] ? kthread_create_on_node+0x200/0x200
 [<ffffffff81ba183f>] ret_from_fork+0x3f/0x70
 [<ffffffff810fa6f0>] ? kthread_create_on_node+0x200/0x200

Call Trace:
 [<ffffffff817fd36d>] xhci_setup_device+0x53d/0xa40
 [<ffffffff817fd87e>] xhci_address_device+0xe/0x10
 [<ffffffff817d047f>] hub_port_init+0x1bf/0xb70
 [<ffffffff811247ed>] ? trace_hardirqs_on+0xd/0x10
 [<ffffffff817d4697>] hub_event+0x817/0x1570
 [<ffffffff810f3e6f>] process_one_work+0x1ff/0x620
 [<ffffffff810f3dcf>] ? process_one_work+0x15f/0x620
 [<ffffffff810f4684>] worker_thread+0x64/0x4b0
 [<ffffffff810f4620>] ? rescuer_thread+0x390/0x390
 [<ffffffff810fa7f5>] kthread+0x105/0x120
 [<ffffffff810fa6f0>] ? kthread_create_on_node+0x200/0x200
 [<ffffffff81ba183f>] ret_from_fork+0x3f/0x70
 [<ffffffff810fa6f0>] ? kthread_create_on_node+0x200/0x200

Which results from the two call chains:

hub_port_init
 usb_get_device_descriptor
  usb_get_descriptor
   usb_control_msg
    usb_internal_control_msg
     usb_start_wait_urb
      usb_submit_urb / wait_for_completion_timeout / usb_kill_urb

hub_port_init
 hub_set_address
  xhci_address_device
   xhci_setup_device

Mathias Nyman explains the current behaviour violates the XHCI spec:

 hub_port_reset() will end up moving the corresponding xhci device slot
 to default state.

 As hub_port_reset() is called several times in hub_port_init() it
 sounds reasonable that we could end up with two threads having their
 xhci device slots in default state at the same time, which according to
 xhci 4.5.3 specs still is a big no no:

 "Note: Software shall not transition more than one Device Slot to the
  Default State at a time"

 So both threads fail at their next task after this.
 One fails to read the descriptor, and the other fails addressing the
 device.

Fix this in hub_port_init by locking the USB controller (instead of an
individual bus) to prevent simultaneous initialisation of both buses.

Fixes: 638139eb95d2 ("usb: hub: allow to process more usb hub events in parallel")
Link: https://lkml.org/lkml/2016/2/8/312
Link: https://lkml.org/lkml/2016/2/4/748
Signed-off-by: Chris Bainbridge <chris.bainbridge@gmail.com>
Cc: stable <stable@vger.kernel.org>
Acked-by: Mathias Nyman <mathias.nyman@linux.intel.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
drivers/usb/core/hcd.c
drivers/usb/core/hub.c
include/linux/usb.h
include/linux/usb/hcd.h

index 2ca2cef7f6811abd03bf036770667e31091c53df..980fc5774151baf47ce9598a480d30d91ff98464 100644 (file)
@@ -994,7 +994,7 @@ static void usb_bus_init (struct usb_bus *bus)
        bus->bandwidth_allocated = 0;
        bus->bandwidth_int_reqs  = 0;
        bus->bandwidth_isoc_reqs = 0;
-       mutex_init(&bus->usb_address0_mutex);
+       mutex_init(&bus->devnum_next_mutex);
 }
 
 /*-------------------------------------------------------------------------*/
@@ -2521,6 +2521,14 @@ struct usb_hcd *usb_create_shared_hcd(const struct hc_driver *driver,
                return NULL;
        }
        if (primary_hcd == NULL) {
+               hcd->address0_mutex = kmalloc(sizeof(*hcd->address0_mutex),
+                               GFP_KERNEL);
+               if (!hcd->address0_mutex) {
+                       kfree(hcd);
+                       dev_dbg(dev, "hcd address0 mutex alloc failed\n");
+                       return NULL;
+               }
+               mutex_init(hcd->address0_mutex);
                hcd->bandwidth_mutex = kmalloc(sizeof(*hcd->bandwidth_mutex),
                                GFP_KERNEL);
                if (!hcd->bandwidth_mutex) {
@@ -2532,6 +2540,7 @@ struct usb_hcd *usb_create_shared_hcd(const struct hc_driver *driver,
                dev_set_drvdata(dev, hcd);
        } else {
                mutex_lock(&usb_port_peer_mutex);
+               hcd->address0_mutex = primary_hcd->address0_mutex;
                hcd->bandwidth_mutex = primary_hcd->bandwidth_mutex;
                hcd->primary_hcd = primary_hcd;
                primary_hcd->primary_hcd = primary_hcd;
@@ -2598,8 +2607,10 @@ static void hcd_release(struct kref *kref)
        struct usb_hcd *hcd = container_of (kref, struct usb_hcd, kref);
 
        mutex_lock(&usb_port_peer_mutex);
-       if (usb_hcd_is_primary_hcd(hcd))
+       if (usb_hcd_is_primary_hcd(hcd)) {
+               kfree(hcd->address0_mutex);
                kfree(hcd->bandwidth_mutex);
+       }
        if (hcd->shared_hcd) {
                struct usb_hcd *peer = hcd->shared_hcd;
 
index c2270d8fac127d4fa7a8f962291164931256a7f2..bee13517676f9b2f00e3cd4e3fd4185f5aaa13c9 100644 (file)
@@ -2082,7 +2082,7 @@ static void choose_devnum(struct usb_device *udev)
        struct usb_bus  *bus = udev->bus;
 
        /* be safe when more hub events are proceed in parallel */
-       mutex_lock(&bus->usb_address0_mutex);
+       mutex_lock(&bus->devnum_next_mutex);
        if (udev->wusb) {
                devnum = udev->portnum + 1;
                BUG_ON(test_bit(devnum, bus->devmap.devicemap));
@@ -2100,7 +2100,7 @@ static void choose_devnum(struct usb_device *udev)
                set_bit(devnum, bus->devmap.devicemap);
                udev->devnum = devnum;
        }
-       mutex_unlock(&bus->usb_address0_mutex);
+       mutex_unlock(&bus->devnum_next_mutex);
 }
 
 static void release_devnum(struct usb_device *udev)
@@ -4366,7 +4366,7 @@ hub_port_init(struct usb_hub *hub, struct usb_device *udev, int port1,
        if (oldspeed == USB_SPEED_LOW)
                delay = HUB_LONG_RESET_TIME;
 
-       mutex_lock(&hdev->bus->usb_address0_mutex);
+       mutex_lock(hcd->address0_mutex);
 
        /* Reset the device; full speed may morph to high speed */
        /* FIXME a USB 2.0 device may morph into SuperSpeed on reset. */
@@ -4652,7 +4652,7 @@ fail:
                hub_port_disable(hub, port1, 0);
                update_devnum(udev, devnum);    /* for disconnect processing */
        }
-       mutex_unlock(&hdev->bus->usb_address0_mutex);
+       mutex_unlock(hcd->address0_mutex);
        return retval;
 }
 
index 7824f4557d508d0adb7be676b6aa04cb1ab4a4d6..01b6c61cf9bb7e0959a7fc9ab50f49ca2fc89440 100644 (file)
@@ -374,13 +374,12 @@ struct usb_bus {
 
        int devnum_next;                /* Next open device number in
                                         * round-robin allocation */
+       struct mutex devnum_next_mutex; /* devnum_next mutex */
 
        struct usb_devmap devmap;       /* device address allocation map */
        struct usb_device *root_hub;    /* Root hub */
        struct usb_bus *hs_companion;   /* Companion EHCI bus, if any */
 
-       struct mutex usb_address0_mutex; /* unaddressed device mutex */
-
        int bandwidth_allocated;        /* on this bus: how much of the time
                                         * reserved for periodic (intr/iso)
                                         * requests is used, on average?
index b98f831dcda32f7eb27b5df41242613f44e08474..66fc13705ab7a45af7211a967b0b377a14127cd6 100644 (file)
@@ -181,6 +181,7 @@ struct usb_hcd {
         * bandwidth_mutex should be dropped after a successful control message
         * to the device, or resetting the bandwidth after a failed attempt.
         */
+       struct mutex            *address0_mutex;
        struct mutex            *bandwidth_mutex;
        struct usb_hcd          *shared_hcd;
        struct usb_hcd          *primary_hcd;