Messages in this thread | | | Subject | Re: INFO: possible circular locking dependency at cleanup_workqueue_thread | From | Peter Zijlstra <> | Date | Wed, 20 May 2009 15:44:52 +0200 |
| |
On Wed, 2009-05-20 at 15:18 +0200, Oleg Nesterov wrote: > On 05/20, Peter Zijlstra wrote: > > > > > > ======================================================= > > > > [ INFO: possible circular locking dependency detected ] > > > > 2.6.30-rc5-00097-gd665355 #59 > > > > ------------------------------------------------------- > > > > pm-suspend/12129 is trying to acquire lock: > > > > (events){+.+.+.}, at: [<ffffffff80259496>] cleanup_workqueue_thread+0x26/0xd0 > > > > > > > > but task is already holding lock: > > > > (cpu_add_remove_lock){+.+.+.}, at: [<ffffffff80246e57>] > > > > cpu_maps_update_begin+0x17/0x20 > > > > > > > > which lock already depends on the new lock. > > > > > > > > > > > > the existing dependency chain (in reverse order) is: > > > > > > > > -> #5 (cpu_add_remove_lock){+.+.+.}: > > > > [<ffffffff80271a64>] __lock_acquire+0xc64/0x10a0 > > > > [<ffffffff80271f38>] lock_acquire+0x98/0x140 > > > > [<ffffffff8054e78c>] __mutex_lock_common+0x4c/0x3b0 > > > > [<ffffffff8054ebf6>] mutex_lock_nested+0x46/0x60 > > > > [<ffffffff80246e57>] cpu_maps_update_begin+0x17/0x20 > > > > [<ffffffff80259c33>] __create_workqueue_key+0xc3/0x250 > > > > [<ffffffff80287b20>] stop_machine_create+0x40/0xb0 > > > > [<ffffffff8027a784>] sys_delete_module+0x84/0x270 > > > > [<ffffffff8020c15b>] system_call_fastpath+0x16/0x1b > > > > [<ffffffffffffffff>] 0xffffffffffffffff > > > > Oleg, why does __create_workqueue_key() require cpu_maps_update_begin()? > > Wouldn't get_online_cpus() be enough to freeze the online cpus? > > Yes, get_online_cpus() pins online CPUs. But CPU_POST_DEAD calls > cleanup_workqueue_thread() without cpu_hotplug.lock, this means > that create/destroy can race with cpu_down(). > > We can avoid cpu_add_remove_lock, but then we have to add another > lock to protect workqueues, cpu_populated_map, etc.
Joy..
> > Breaking the setup_lock -> cpu_add_remove_lock dependency seems > > sufficient. > > Hmm. What do you mean? Afaics setup_lock -> cpu_add_remove_lock > is not a problem?
>From what I could see that is the only dependency that makes cpu_add_remove_lock nest under "events" workqueue 'lock', which is what is generating the deadlock.
| |