ehci_submit_async() doesn't really zero out the QH transfer overlay (as the EHCI specification suggests) which leads to the controller seeing the "token" field as the previous call has left it, i.e.: - if a timeout occured on the previous call (Active bit left as 1), controller incorrectly tries to complete a previous transaction on a newly programmed endpoint; - if a halt occured on the previous call (Halted bit set to 1), controller just ignores the newly programmed TD(s) and the function then keeps returning error ad infinitum.
This turned out to be caused by the wrong orger of the arguments to the memset() call in ehci_alloc(), so the allocated TDs weren't cleared either.
While at it, stop needlessly initializing the alternate next TD pointer in the QH transfer overlay...
Signed-off-by: Sergei Shtylyov sshtylyov@ru.mvista.com
--- Initialization of 'qh->qh_overlay.qt_next' turned out to be needed by the error cleanup code (seemingly useless, in its turn).
This is quite serious bug, so would be good to have the patch in v2010.06...
drivers/usb/host/ehci-hcd.c | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-)
Index: u-boot/drivers/usb/host/ehci-hcd.c =================================================================== --- u-boot.orig/drivers/usb/host/ehci-hcd.c +++ u-boot/drivers/usb/host/ehci-hcd.c @@ -275,7 +275,7 @@ static void *ehci_alloc(size_t sz, size_ return NULL; }
- memset(p, sz, 0); + memset(p, 0, sz); return p; }
@@ -350,7 +350,6 @@ ehci_submit_async(struct usb_device *dev (dev->parent->devnum << 16) | (0 << 8) | (0 << 0); qh->qh_endpt2 = cpu_to_hc32(endpt); qh->qh_overlay.qt_next = cpu_to_hc32(QT_NEXT_TERMINATE); - qh->qh_overlay.qt_altnext = cpu_to_hc32(QT_NEXT_TERMINATE);
td = NULL; tdp = &qh->qh_overlay.qt_next;