Stephen Hemminger [off-list ref] :

This driver was confused about how workqueue's work. It is save to
schedule multiple times, and the flag (bitfield) was unnecessary and racy.

static void rtl8139_start_thread(struct rtl8139_private *tp)
{
        tp->twistie = 0;
        if (tp->chipset == CH_8139_K)
                tp->twistie = 1;
        else if (tp->drv_flags & HAS_LNK_CHNG)
                return;
                ^^^^^^^

        tp->have_thread = 1;

The use of the thread was dependent on RTL8129_CAPS. Now it will be
scheduled in the tx_timeout handler for RTL8139_CAPS too.

Speaking of race, the bitfield ('have_thread') is initialized through
rtl8139_open
-> rtl8139_start_thread
Do you mean that the tx_timeout handler can kick before the bitfield
is set ?

The workqueue was never canceled either, so it is possible that the thread
could run after driver was removes [OOPS].

The driver uses flush_scheduled_work() when the pci device is removed.

Where would the workqueue be scheduled after / during the removal ?

-- 
Ueimor