On Tue, 2011-11-22 at 08:48 +0100, Eric Dumazet wrote: > Le lundi 21 novembre 2011 à 21:18 -0600, Christoph Lameter a écrit : > > > Hmmm... That means that c->page points to page not frozen. Per cpu > > partial pages are frozen until they are reused or until the partial list > > is flushed. > > > > Does this ever happen on x86 or only on other platforms? In put_cpu_partial() the > > this_cpu_cmpxchg really needs really to be irq safe. this_cpu_cmpxchg is > > only preempt safe. > > > > Index: linux-2.6/mm/slub.c > > =================================================================== > > --- linux-2.6.orig/mm/slub.c 2011-11-21 21:15:41.575673204 -0600 > > +++ linux-2.6/mm/slub.c 2011-11-21 21:16:33.442336849 -0600 > > @@ -1969,7 +1969,7 @@ > > page->pobjects = pobjects; > > page->next = oldpage; > > > > - } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage); > > + } while (irqsafe_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage); > > stat(s, CPU_PARTIAL_FREE); > > return pobjects; > > } > > > > For x86, I wonder if our !X86_FEATURE_CX16 support is correct on SMP > machines. > > this_cpu_cmpxchg16b_emu() claims to be IRQ safe, but may be buggy... > > Could we have somewhere a NMI handler calling kmalloc() ? Christian and I are on ppc, which uses the generic implementation of this_cpu_cmpxchg() which is not irq safe. So the above patch is needed regardless. Christian, can you try it see if that helps in your case ? Cheers, Ben. -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/ Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>