On Sunday 13 January 2008 18:08:57 Alan Stern wrote: > On Sun, 13 Jan 2008, Rafael J. Wysocki wrote: > > > On Sunday, 13 of January 2008, Michael Buesch wrote: > > > On Sunday 13 January 2008 00:08:29 Rafael J. Wysocki wrote: > > > > There is a problem with b43_suspend() that it (indirectly) causes > > > > b43_leds_exit() to be called, which attempts to unregister the leds device > > > > objects, which is forbidden (ie. you can't unregister and/or register devices > > > > during a suspend or resume). > > > > > > Why? > > > > Well, the unregistering itself is not really harmful, provided that the device > > is not registered back during the subsequent resume. > > > > The PM core uses a list of active devices that are added to the list in > > device_add(). The ordering of this list is important, because it is expected > > to reflect the order in which the devices are to be suspended. > > > > This list is manipulated during suspend/resume and devices are moved from it > > and back to it, so unregistering devices during a suspend and registering them > > during the subsequent resume generally changes its ordering and may lead to > > problems during the next suspend/resume cycle. > > > > This is also undesirable if we're going to stop using the freezer for > > suspend/resume at one point in the future. > > > > I'm sure Alan can add some more details. > > Indeed. A system suspend is a delicate operation, and the PM core > needs to access all the devices in the system. To have devices being > registered and unregistered at the same time would cause a lot of > confusion. In self defense, the PM core starts out by acquiring all > the device semaphores before calling the suspend routines. This means > that if a suspend routine tries to unregister anything, it will > deadlock at the point where the driver core tries to acquire the device > semaphore prior to invoking the driver's remove method. > > Besides, if you're going to register the device right back again during > the subsequent resume, then why go to the trouble of unregistering it > during suspend? Why not just leave it registered the whole time and > avoid all the complication (and excess meaningless uevents)? Well, because not doing it complicates code :) Currently suspend/resume calls the same code as init/exit. The b43 init/exit code is really complicated, compared to other drivers, due to dozens of hardware versions. So I just avoided having yet other codepaths for suspend/resume. But I will add a flag to the device structure that's used to avoid unregistering stuff. -- Greetings Michael. _______________________________________________ linux-pm mailing list linux-pm@xxxxxxxxxxxxxxxxxxxxxxxxxx https://lists.linux-foundation.org/mailman/listinfo/linux-pm
