Hi, back from a long sick leave
On 10.01.2019 11:39, Oliver Neukum wrote:
Hi,
looking at this code:
/* ring command ring doorbell to restart the command ring */
if ((xhci->cmd_ring->dequeue != xhci->cmd_ring->enqueue) &&
!(xhci->xhc_state & XHCI_STATE_DYING)) {
xhci->current_cmd = cur_cmd;
xhci_mod_cmd_timer(xhci, XHCI_CMD_DEFAULT_TIMEOUT);
xhci_ring_cmd_db(xhci);
}
Something looks fishy to me. The timer is started before the door bell
is rung. I cannot see how you can guarantee the timeout given that
order. Should it be reversed?
Could be reversed as we took spin_lock_irqsave() before this, so it shouldn't
cause any race issues with command completion canceling the timeout timer
before it is started.
Then again, the command timeout is 5 seconds, spin lock is taken with local
irqs disabled. The only possible problem case I can see is a 5 second SMI
between starting the timer and ringing the doorbell.
Sound more like a theoretical issues, or am I missing something?
Anyways, If you send a patch I can apply it
-Mathias
