Re: percpu related boot crash on x86 (was: Linux 2.6.38-rc1)

[Peter Zijlstra <a.p.zijlstra@chello.nl>]

On Wed, 2011-01-19 at 14:56 +0200, Pekka Enberg wrote:
> 
> On Wed, Jan 19, 2011 at 2:48 PM, Peter Zijlstra
> <a.p.zijlstra@chello.nl> wrote:
> >  start_kernel()
> >   local_irq_disable()
> >   ...
> >   early_irq_init()
> >     alloc_desc()
> >       alloc_percpu()
> >         __alloc_percpu()
> >           pcpu_alloc()
> >             pcpu_extend_area_map()
> >               pcpu_mem_free()
> >                 vfree()
> >                   __vunmap()
> >                     remove_vm_area()
> >                       free_unmap_vmap_area()
> >                         vmap_debug_free_range()
> > #ifdef CONFIG_DEBUG_PAGEALLOC
> >                           flush_tlb_kernel_range()
> >                             flush_tlb_all()
> >                               on_each_cpu()
> >                                 smp_call_function()
> >                                   WARN_ON_ONCE(irqs_disabled()....);
> >
> >
> > Not quite sure that to do about that though..
> 
> Is vmalloc() and vfree() supposed to work with interrupts disabled? 

I'm not sure about vmalloc(), but vfree() has been known to not work
with IRQs disabled for a long long while.

Looking at the code neither can work, since vmalloc() can end up calling
vfree(), what's curious though is that all the BUG_ON()s in there are on
in_interrupt() instead of for irqs_disabled().

> I always thought they weren't which would mean something in
> pcpu_mem_alloc() needs changing... 

Is alloc_percpu() supposed to work from atomic context?