On Thursday 17 December 2009 20:11:55 Sven Eckelmann wrote:
forw_bcast_list_lock is spin_locked in both process and softirq context. SoftIRQ calls the spinlock with disabled IRQ and normal process context with enabled IRQs.
When process context is inside an spin_locked area protected by forw_bcast_list_lock and gets interrupted by an IRQ, it could happen that something tries to lock forw_bcast_list_lock again in SoftIRQ context. It cannot proceed further since the lock is already taken somewhere else, but no reschedule will happen inside the SoftIRQ context. This leads to an complete kernel hang without any chance of resurrection.
All functions called in process context must disable IRQs when they try to get get that lock to to prevent any reschedule due to IRQs.
Thanks - nice catch (applied in rev 1504)!
Regards, Marek