Re: sleeping functions called in invalid context during resume

[Andrew Morton <akpm@osdl.org>]

On Thu, 16 Nov 2006 21:21:58 -0800
Stephen Hemminger <shemminger@osdl.org> wrote:

> 
> > Stephen Hemminger <shemminger@osdl.org> wrote:
> > 
> > > Lots of sleeping while atomic warnings on 2.6.19-rc5
> > > During resume I see the following:
> > > 
> > > 
> > > platform floppy.0: EARLY resume
> > > APIC error on CPU0: 00(00)
> > > PM: Finishing wakeup.
> > > BUG: sleeping function called from invalid context at drivers/base/power/resume.c:99
> > > in_atomic():1, irqs_disabled():0
> > > 
> > > Call Trace:  
> > >  [<ffffffff80266117>] show_trace+0x34/0x47
> > >  [<ffffffff8026613c>] dump_stack+0x12/0x17
> > >  [<ffffffff803734e5>] device_resume+0x19/0x51
> > >  [<ffffffff80292157>] enter_state+0x19b/0x1b5
> > >  [<ffffffff802921cf>] state_store+0x5e/0x79
> > >  [<ffffffff802cc157>] sysfs_write_file+0xc5/0xf8
> > >  [<ffffffff80215059>] vfs_write+0xce/0x174
> > >  [<ffffffff802159a5>] sys_write+0x45/0x6e
> > >  [<ffffffff802593de>] system_call+0x7e/0x83  
> > > DWARF2 unwinder stuck at system_call+0x7e/0x83
> > > 
> 
> Ingo, the later version of your lockdep patch (with the x86_64 fix), worked.
> There is nothing locked during these errors.
> 
> The problem was the APIC error is leaving preempt-disabled.
> 
> I have no idea what causes:
> 
> APIC error on CPU0: 00(00)
> 
> Is it an ACPI problem?

Strange.  x86_64 has that stray exit_idle() in smp_error_interrupt() but
afaict it won't cause this to happen.

What's that idle_notifier doing in x86_64 anyway?  It appears to have no
users.  If there _is_ a user, and if its IDLE_END handler is altering the
preempt-count then perhaps there's your explanation.

But it all appears to be dead code to me.