Re: [PATCH] power, sched: stop updating inside arch_update_cpu_topology() when nothing to be update

[Srivatsa S. Bhat]

Hi Michael,

On 04/02/2014 08:59 AM, Michael wang wrote:
> During the testing, we encounter below WARN followed by Oops:
> 
> 	WARNING: at kernel/sched/core.c:6218
> 	...
> 	NIP [c000000000101660] .build_sched_domains+0x11d0/0x1200
> 	LR [c000000000101358] .build_sched_domains+0xec8/0x1200
> 	PACATMSCRATCH [800000000000f032]
> 	Call Trace:
> 	[c00000001b103850] [c000000000101358] .build_sched_domains+0xec8/0x1200
> 	[c00000001b1039a0] [c00000000010aad4] .partition_sched_domains+0x484/0x510
> 	[c00000001b103aa0] [c00000000016d0a8] .rebuild_sched_domains+0x68/0xa0
> 	[c00000001b103b30] [c00000000005cbf0] .topology_work_fn+0x10/0x30
> 	...
> 	Oops: Kernel access of bad area, sig: 11 [#1]
> 	...
> 	NIP [c00000000045c000] .__bitmap_weight+0x60/0xf0
> 	LR [c00000000010132c] .build_sched_domains+0xe9c/0x1200
> 	PACATMSCRATCH [8000000000029032]
> 	Call Trace:
> 	[c00000001b1037a0] [c000000000288ff4] .kmem_cache_alloc_node_trace+0x184/0x3a0
> 	[c00000001b103850] [c00000000010132c] .build_sched_domains+0xe9c/0x1200
> 	[c00000001b1039a0] [c00000000010aad4] .partition_sched_domains+0x484/0x510
> 	[c00000001b103aa0] [c00000000016d0a8] .rebuild_sched_domains+0x68/0xa0
> 	[c00000001b103b30] [c00000000005cbf0] .topology_work_fn+0x10/0x30
> 	...
> 
> This was caused by that 'sd->groups == NULL' after building groups, which
> was caused by the empty 'sd->span'.
> 
> The cpu's domain contain nothing because the cpu was assigned to wrong
> node inside arch_update_cpu_topology() by calling update_lookup_table()
> with the uninitialized param, in the case when there is nothing to be
> update.
> 

Can you reword the above paragraph to something like this:

The cpu's domain contained nothing because the cpu was assigned to a wrong
node, due to the following unfortunate sequence of events:

1. The hypervisor sent a topology update to the guest OS, to notify changes
   to the cpu-node mapping. However, the update was actually redundant - i.e.,
   the "new" mapping was exactly the same as the old one.

2. Due to this, the 'updated_cpus' mask turned out to be empty after exiting
   the 'for-loop' in arch_update_cpu_topology().

3. So we ended up calling stop-machine() with an empty cpumask list, which made
   stop-machine internally elect cpumask_first(cpu_online_mask), i.e., CPU0 as
   the cpu to run the payload (the update_cpu_topology() function).

4. This causes update_cpu_topology() to be run by CPU0. And since 'updates'
   is kzalloc()'ed inside arch_update_cpu_topology(), update_cpu_topology()
   finds update->cpu as well as update->new_nid to be 0. In other words, we
   end up assigning CPU0 (and eventually its siblings) to node 0, incorrectly.

This causes the sched-domain rebuild code to break and crash the system.


> Thus we should stop the updating in such cases, this patch will achieve
> this and fix the issue.
> 

We can reword this part as:

Fix this by skipping the topology update in cases where we find that
the topology has not actually changed in reality (ie., spurious updates).

> CC: Benjamin Herrenschmidt <benh@xxxxxxxxxxxxxxxxxxx>
> CC: Paul Mackerras <paulus@xxxxxxxxx>
> CC: Nathan Fontenot <nfont@xxxxxxxxxxxxxxxxxx>
> CC: Stephen Rothwell <sfr@xxxxxxxxxxxxxxxx>
> CC: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
> CC: Robert Jennings <rcj@xxxxxxxxxxxxxxxxxx>
> CC: Jesse Larrew <jlarrew@xxxxxxxxxxxxxxxxxx>
> CC: "Srivatsa S. Bhat" <srivatsa.bhat@xxxxxxxxxxxxxxxxxx>
> CC: Alistair Popple <alistair@xxxxxxxxxxxx>
> Signed-off-by: Michael Wang <wangyun@xxxxxxxxxxxxxxxxxx>
> ---
>  arch/powerpc/mm/numa.c |    9 +++++++++
>  1 file changed, 9 insertions(+)
> 
> diff --git a/arch/powerpc/mm/numa.c b/arch/powerpc/mm/numa.c
> index 30a42e2..6757690 100644
> --- a/arch/powerpc/mm/numa.c
> +++ b/arch/powerpc/mm/numa.c
> @@ -1591,6 +1591,14 @@ int arch_update_cpu_topology(void)
>  		cpu = cpu_last_thread_sibling(cpu);
>  	}
>  
> +	/*
> +	 * The 'cpu_associativity_changes_mask' could be cleared if
> +	 * all the cpus it indicates won't change their node, in
> +	 * which case the 'updated_cpus' will be empty.
> +	 */

How about rewording the comment like this:

In cases where we have nothing to update (because the updates list
is too short or because the new topology is same as the old one),
skip invoking update_cpu_topology() via stop-machine(). This is
necessary (and not just a fast-path optimization) because stop-machine
can end up electing a random CPU to run update_cpu_topology(), and
thus trick us into setting up incorrect cpu-node mappings (since
'updates' is kzalloc()'ed).

Regards,
Srivatsa S. Bhat

> +	if (!cpumask_weight(&updated_cpus))
> +		goto out;
> +
>  	stop_machine(update_cpu_topology, &updates[0], &updated_cpus);
>  
>  	/*
> @@ -1612,6 +1620,7 @@ int arch_update_cpu_topology(void)
>  		changed = 1;
>  	}
>  
> +out:
>  	kfree(updates);
>  	return changed;
>  }
> 

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