Re: 2.6.22-rc3-mm1

[Christoph Lameter <clameter@sgi.com>]

On Thu, 31 May 2007, Andrew Morton wrote:

> > l *0xc0181288
> > 0xc0181288 is in add_partial (/home/devel/linux-mm/mm/slub.c:1193).
> > 1188    }
> > 1189
> > 1190    static void add_partial(struct kmem_cache_node *n, struct page *page)
> > 1191    {
> > 1192            spin_lock(&n->list_lock);
> > 1193            n->nr_partial++;
> > 1194            list_add(&page->lru, &n->partial);
> > 1195            spin_unlock(&n->list_lock);
> > 1196    }
> > 1197

add_partial runs with interrupts disabled. The interrupts are disabled 
when we enter SLUB via allocator or free functions.

> > [ 4972.087650] {in-hardirq-W} state was registered at:
> > [ 4972.092656]   [<c0143bfe>] mark_lock+0x82/0x557
> > [ 4972.097323]   [<c0144c5f>] __lock_acquire+0x476/0xd36
> > [ 4972.102562]   [<c01455bd>] lock_acquire+0x9e/0xb8
> > [ 4972.107342]   [<c0348993>] _spin_lock+0x38/0x62
> > [ 4972.111993]   [<c0181d29>] deactivate_slab+0xb9/0x179
> > [ 4972.117300]   [<c0181e56>] flush_slab+0x6d/0x72
> > [ 4972.122063]   [<c0181e8c>] __flush_cpu_slab+0x31/0x36
> > [ 4972.127335]   [<c0181ea5>] flush_cpu_slab+0x14/0x17
> > [ 4972.132401]   [<c0113d6f>] smp_call_function_interrupt+0x3a/0x56
> > [ 4972.138607]   [<c0104c73>] call_function_interrupt+0x33/0x38
> > [ 4972.144503]   [<c0102b59>] default_idle+0x50/0x69
> > [ 4972.149421]   [<c01023eb>] cpu_idle+0xb3/0xf8
> > [ 4972.153889]   [<c03454fa>] rest_init+0x56/0x58
> > [ 4972.158402]   [<c04f39c7>] start_kernel+0x351/0x359
> > [ 4972.163450]   [<ffffffff>] 0xffffffff
> > [ 4972.167221] irq event stamp: 2451
> 
> Yep, that's a bug in slub.  We take that lock in the IPI handler.  If a CPU
> is currently holding that lock and then takes the IPI and enters
> add_partial(), it'll deadlock.

A cpu cannot enter an IPI handler while interrupts are disabled. That 
needs to be always the case when the list_lock is held.

> Perhaps a suitable fix would be local_irq_disable() in flush_slab().

As far as I can tell: Interrupts are always disabled when flush_slab is 
run. 

Sometimes we use spin_lock_irqsave for the list_lock and at other times 
spin_lock if interrupts are already disabled. Is that the problem?

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/