Messages in this thread |  | | | Date | Mon, 9 Nov 2009 07:57:35 -0800 (PST) | | From | Linus Torvalds <> | | Subject | Re: Help needed: Resume problems in 2.6.32-rc, perhaps related to preempt_count leakage in keventd |
On Mon, 9 Nov 2009, Thomas Gleixner wrote: > > I think it _IS_ releated because the worker_thread is CPU affine and > the debug_smp_processor_id() check does:
Hmm. We do know CPU affinity is destroyed by CPU hotplug. People have complained about that before (for user-space processes that get moved around due to hot-unplug/plug).
And the suspend/resume process does CPU hotplug to take down all but one CPU. The workqueues should act on those events already, but maybe there's a bug somewhere. None of that is new to 32-rc, though - is it?
And workqueue_cpu_callback() seems buggy. It loops over the 'workqueues' list with no protection. Yes, we do 'stop_machine' for CPU hotplug events, but only for the very internal one (CPU_DYING) will the CPU notifiers be called with the machine stopped).
Hmm. I don't see any changes in kernel/cpu.c or kernel/workqueue.c that look at all relevant. But scheduler changes could certainly matter.
Linus
|  |