From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1755029AbZDUNva (ORCPT ); Tue, 21 Apr 2009 09:51:30 -0400 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S1752306AbZDUNvU (ORCPT ); Tue, 21 Apr 2009 09:51:20 -0400 Received: from cantor2.suse.de ([195.135.220.15]:38625 "EHLO mx2.suse.de" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751932AbZDUNvS (ORCPT ); Tue, 21 Apr 2009 09:51:18 -0400 Date: Tue, 21 Apr 2009 15:51:16 +0200 (CEST) From: Jiri Kosina X-X-Sender: jkosina@wotan.suse.de To: Johannes Berg Cc: Dmitry Torokhov , linux-input , linux-kernel , "Rafael J. Wysocki" , Oleg Nesterov , Oliver Neukum Subject: Re: lockdep report at resume In-Reply-To: <1238578840.5970.178.camel@johannes.local> Message-ID: References: <1234022517.4175.107.camel@johannes.local> <1237536764.5100.124.camel@johannes.local> <1237800572.19647.97.camel@johannes.local> <1238488692.5970.80.camel@johannes.local> <1238578840.5970.178.camel@johannes.local> User-Agent: Alpine 2.00 (LNX 1167 2008-08-23) MIME-Version: 1.0 Content-Type: TEXT/PLAIN; charset=US-ASCII Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org [ thanks for the analysis Johannes ... involving also Oliver here; really seems like a false positive to me ] On Wed, 1 Apr 2009, Johannes Berg wrote: > On Tue, 2009-03-31 at 10:40 +0200, Jiri Kosina wrote: > > > > > Could you please send me your config? > > > Sure, attached. I haven't yet tried to reproduce on .29 though, which > > > this config is for (but I haven't changed it since, only taken it > > > forward). > > I've now gotten it again on 2.6.29-wl-20327-g8f2487d-dirty. > > I've analysed a bit more. > > Let's start from the bottom: > > -> #0 (&dev->mutex){--..}: > [] check_prev_add+0x57/0x770 > [] validate_chain+0x5f6/0x6b0 > [] __lock_acquire+0x43f/0xa10 > [] lock_acquire+0x91/0xc0 > [] mutex_lock_nested+0xfc/0x390 > [] input_disconnect_device+0x31/0xf0 > [] input_unregister_device+0x1a/0x110 > [] bcm5974_disconnect+0x29/0x90 [bcm5974] > [] usb_unbind_interface+0x6d/0x180 [usbcore] > [] __device_release_driver+0x81/0xc0 > [] device_release_driver+0x30/0x50 > [] usb_driver_release_interface+0xc8/0xf0 [usbcore] > [] usb_forced_unbind_intf+0x39/0x90 [usbcore] > [] usb_reset_device+0xd5/0x220 [usbcore] > [] hid_reset+0x18a/0x280 [usbhid] > [] run_workqueue+0x10d/0x250 > > > Here we have hid_reset being called off schedule_work. It eventually > calls into bcm5974 which will, from its usb_driver disconnect call, call > input_unregister_device(), which acquires &dev->mutex. > > > -> #1 (polldev_mutex){--..}: > [] check_prev_add+0x3b7/0x770 > [] validate_chain+0x5f6/0x6b0 > [] __lock_acquire+0x43f/0xa10 > [] lock_acquire+0x91/0xc0 > [] mutex_lock_interruptible_nested+0xec/0x430 > [] input_open_polled_device+0x21/0xd0 [input_polldev] > [] input_open_device+0x98/0xc0 > [] evdev_open+0x1c8/0x1f0 [evdev] > [] input_open_file+0x10f/0x200 > [] chrdev_open+0x147/0x220 > [] __dentry_open+0x11b/0x350 > [] nameidata_to_filp+0x57/0x70 > [] do_filp_open+0x1fe/0x970 > [] do_sys_open+0x80/0x110 > [] sys_open+0x20/0x30 > > This is another code path -- evdev triggered here. Any input polldev > will acquire polldev_mutex within its struct input_dev->open() callback, > and thus create a dependency of &dev->mutex on polldev_mutex because > input_open_device() is called with &dev->mutex held. > > -> #2 (cpu_add_remove_lock){--..}: > [] check_prev_add+0x3b7/0x770 > [] validate_chain+0x5f6/0x6b0 > [] __lock_acquire+0x43f/0xa10 > [] lock_acquire+0x91/0xc0 > [] mutex_lock_nested+0xfc/0x390 > [] cpu_maps_update_begin+0x17/0x20 > [] destroy_workqueue+0x38/0xb0 > [] input_close_polled_device+0x45/0x60 [input_polldev] > [] input_close_device+0x5c/0x90 > [] evdev_release+0xa9/0xd0 [evdev] > [] __fput+0xd5/0x1e0 > [] fput+0x25/0x30 > [] filp_close+0x58/0x90 > [] sys_close+0xbe/0x120 > [] system_call_fastpath+0x16/0x1b > [] 0xffffffffffffffff > > This is cute. > So input-polldev uses its own workqueue, and it's singlethread. But > destroy_workqueue must stop CPU hotplug anyway, calls > cpu_map_update_begin() which locks cpu_add_remove_lock. > > -> #3 (events){--..}: > [] check_prev_add+0x3b7/0x770 > [] validate_chain+0x5f6/0x6b0 > [] __lock_acquire+0x43f/0xa10 > [] lock_acquire+0x91/0xc0 > [] cleanup_workqueue_thread+0x42/0x90 > [] workqueue_cpu_callback+0x9d/0x132 > [] notifier_call_chain+0x65/0xa0 > [] raw_notifier_call_chain+0x16/0x20 > [] _cpu_down+0x1db/0x350 > [] disable_nonboot_cpus+0xe5/0x170 > [] hibernation_snapshot+0x135/0x170 > [] snapshot_ioctl+0x425/0x620 > [] vfs_ioctl+0x36/0xb0 > [] do_vfs_ioctl+0x89/0x350 > [] sys_ioctl+0x4f/0x80 > [] system_call_fastpath+0x16/0x1b > [] 0xffffffffffffffff > > Here we have hibernation, which needs to call disable_nonboot_cpus. This > takes down all CPUs, and causes the workqueue code, now running off the > workqueue_cpu_callback, to call cleanup_workqueue_thread(), which > "acquires" the workqueue. I suspect this will also happen if you go into > sysfs and disable a CPU manually, which may help you reproduce this. > > disable_nonboot_cpus calls cpu_map_update_begin() to avoid other things > interfering, and thus creates the dependency of the workqueue on that. > > > -> #4 (&usbhid->reset_work){--..}: > [] check_prev_add+0x3b7/0x770 > [] validate_chain+0x5f6/0x6b0 > [] __lock_acquire+0x43f/0xa10 > [] lock_acquire+0x91/0xc0 > [] run_workqueue+0x107/0x250 > [] worker_thread+0xaf/0x130 > [] kthread+0x4e/0x90 > [] child_rip+0xa/0x20 > [] 0xffffffffffffffff > > Now, of course, usbhid->reset_work runs off the schedule_work workqueue, > which was stopped during hibernation, so it depends on that workqueue. > > > Finally, we're back at the top, with input_disconnect_device() acquiring > &dev->mutex. > > > > Now, how can a deadlock happen? > > I think it cannot -- unless you have a polled USB device. The two > "&dev->mutex" instances here are from difference devices, but lockdep > cannot tell them apart, and if you have a polled USB device then the > same can happen. > > Assume you had a polled USB driver using input_unregister_polled_device, > and thus input_unregister_device, in its usb_driver disconnect call. In > that case you could potentially trigger the deadlock when you manage to > get that usb device reset very very close before calling > disable_nonboot_cpus, so close that the usb reset_work is still > scheduled or something like that... > > > I don't really see a good way to solve it -- but I hope the analysis > helps some -- also adding lots of people to CC. > > johannes >