B1;2802;0cOn Thu, 17 Mar 2016, Peter Zijlstra wrote:
> On Thu, Mar 17, 2016 at 11:21:24AM +0100, Thomas Gleixner wrote:
> > On Thu, 17 Mar 2016, Peter Zijlstra wrote:
> >
> > > Could you please try? I'm not sure how this would explain your loop
> > > device bug fail, but it certainly pointed towards broken.
> >
> > It definitely does not explain it. The wreckage that topo stuff causes is
> > that
> > it disables a cpu, but that really is not a reason for block/loop to
> > explode.
>
> Right. Sadly I could not reproduce that error on my machine. But we can
> at least start by fixing the 'obvious' problems and then maybe we get
> more clues ;-)
I'm able to reproduce by rejecting a cpu in that topology map function
forcefully.
That stuff explodes, because the block-mq code assumes that cpu_possible_mask
has no holes.
#define queue_for_each_ctx(q, ctx, i) \
for ((i) = 0; (i) < (q)->nr_queues && \
({ ctx = per_cpu_ptr((q)->queue_ctx, (i)); 1; }); (i)++)
is what makes that assumption about a consecutive possible mask.
The cure for now is the patch below on top of PeterZ's patch.
But we have to clarify and document whether holes in cpu_possible_mask are not
allowed at all or if code like the above is simply broken.
Thanks,
tglx
---
diff --git a/arch/x86/kernel/smpboot.c b/arch/x86/kernel/smpboot.c
index 643dbdccf4bc..f2ed8a01f870 100644
--- a/arch/x86/kernel/smpboot.c
+++ b/arch/x86/kernel/smpboot.c
@@ -345,7 +345,6 @@ static void __init smp_init_package_map(void)
continue;
pr_warn("CPU %u APICId %x disabled\n", cpu, apicid);
per_cpu(x86_bios_cpu_apicid, cpu) = BAD_APICID;
- set_cpu_possible(cpu, false);
set_cpu_present(cpu, false);
}
}
