Re: [PATCH] mm: might_sleep warning

[Andrew Morton]

On Tue, 6 Mar 2018 16:04:06 -0500 Pavel Tatashin <pasha.tatashin@xxxxxxxxxx> wrote:

> > > > >       spin_lock(&deferred_zone_grow_lock);
> > > > > -     static_branch_disable(&deferred_pages);
> > > > > +     deferred_zone_grow = false;
> > > > >       spin_unlock(&deferred_zone_grow_lock);
> > > > > +     static_branch_disable(&deferred_pages);
> > > > >
> > > > >       /* There will be num_node_state(N_MEMORY) threads */
> > > > >       atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
> > > >
> > > > Kinda ugly, but I can see the logic behind the decisions.
> > > >
> > > > Can we instead turn deferred_zone_grow_lock into a mutex?
> >
> > (top-posting repaired.  Please don't top-post).
> >
> > > [CCed everyone]
> > >
> > > Hi Andrew,
> > >
> > > I afraid we cannot change this spinlock to mutex
> > > because deferred_grow_zone() might be called from an interrupt context if
> > > interrupt thread needs to allocate memory.
> > >
> >
> > OK.  But if deferred_grow_zone() can be called from interrupt then
> > page_alloc_init_late() should be using spin_lock_irq(), shouldn't it?
> > I'm surprised that lockdep didn't detect that.
> 
> No, page_alloc_init_late()  cannot be called from interrupt, it is
> called straight from:
> kernel_init_freeable(). But, I believe deferred_grow_zone(): can be called:
> 
> get_page_from_freelist()
>  _deferred_grow_zone()
>    deferred_grow_zone()

That's why page_alloc_init_late() needs spin_lock_irq().  If a CPU is
holding deferred_zone_grow_lock with enabled interrupts and an
interrupt comes in on that CPU and the CPU runs deferred_grow_zone() in
its interrupt handler, we deadlock.

lockdep knows about this bug and should have reported it.