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)
commitfeb26ac31a2a5cb88d86680d9a94916a6343e9e6
treeaa6b222fdf730e07e6e7f0f4d77684d859213ed7
parentedc8c54bf454e6ba4f4631db1603f2297e4d4b67
usb: core: hub: hub_port_init lock controller instead of bus

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