+ *
+ * TODO:
+ * - Now that the driver was significantly simplified, I need to rework
+ * the locking. I'm sure we don't need _2_ spinlocks, and we probably
+ * can avoid taking most of them for so long period of time (and schedule
+ * instead). The main issues at this point are caused by the netdev layer
+ * though:
+ *
+ * gem_change_mtu() and gem_set_multicast() are called with a read_lock()
+ * help by net/core/dev.c, thus they can't schedule. That means they can't
+ * call netif_poll_disable() neither, thus force gem_poll() to keep a spinlock
+ * where it could have been dropped. change_mtu especially would love also to
+ * be able to msleep instead of horrid locked delays when resetting the HW,
+ * but that read_lock() makes it impossible, unless I defer it's action to
+ * the reset task, which means it'll be asynchronous (won't take effect until
+ * the system schedules a bit).
+ *
+ * Also, it would probably be possible to also remove most of the long-life
+ * locking in open/resume code path (gem_reinit_chip) by beeing more careful
+ * about when we can start taking interrupts or get xmit() called...