summaryrefslogtreecommitdiffstats
path: root/include/linux
diff options
context:
space:
mode:
authorHans de Goede <hdegoede@redhat.com>2013-11-08 13:41:05 +0100
committerSarah Sharp <sarah.a.sharp@linux.intel.com>2014-03-04 15:38:15 -0800
commita82b76f7fa6154e8ab2d8071842a3e38b9c0d0ff (patch)
tree378682cb028e2f250b1bee815b9d429266d3eefc /include/linux
parent7a7b562d08ad6db98d6c8ec634620a11aaf8921a (diff)
downloadtalos-obmc-linux-a82b76f7fa6154e8ab2d8071842a3e38b9c0d0ff.tar.gz
talos-obmc-linux-a82b76f7fa6154e8ab2d8071842a3e38b9c0d0ff.zip
usb: Reset USB-3 devices on USB-3 link bounce
On disconnect USB3 protocol ports transit from U0 to SS.Inactive to Rx.Detect, on a recoverable error, the port stays in SS.Inactive and we recover from it by doing a warm-reset (through usb_device_reset if we have a udev for the port). If this really is a disconnect we may end up trying the warm-reset anyways, since khubd may run before the SS.Inactive to Rx.Detect transition, or it may get skipped if the transition to Rx.Detect happens before khubd gets run. With a loose connector, or in the case which actually led me to debugging this bad ACPI firmware toggling Vbus off and on in quick succession, the port may transition from Rx.Detect to U0 again before khubd gets run. In this case the device state is unknown really, but khubd happily goes into the resuscitate an existing device path, and the device driver never gets notified about the device state being messed up. If the above scenario happens with a streams using device, as soon as an urb is submitted to an endpoint with streams, the following appears in dmesg: ERROR Transfer event for disabled endpoint or incorrect stream ring @0000000036807420 00000000 00000000 04000000 04078000 Notice how the TRB address is all zeros. I've seen this both on Intel Pantherpoint and Nec xhci hosts. Luckily we can detect the U0 to SS.Inactive to Rx.Detect to U0 all having happened before khubd runs case since the C_LINK_STATE bit gets set in the portchange bits on the U0 -> SS.Inactive change. This bit will also be set on suspend / resume, but then it gets cleared by port_hub_init before khubd runs. So if the C_LINK_STATE bit is set and a warm-reset is not needed, iow the port is not still in SS.Inactive, and the port still has a connection, then the device needs to be reset to put it back in a known state. I've verified that doing the device reset also fixes the transfer event with all zeros address issue. Signed-off-by: Hans de Goede <hdegoede@redhat.com> Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Diffstat (limited to 'include/linux')
0 files changed, 0 insertions, 0 deletions
OpenPOWER on IntegriCloud