Re: [PATCH] mm, numa: boot cpu should bound to the node0 when node_off enable

["Leizhen (ThunderTown)" <thunder.leizhen@huawei.com>]

On 2016/8/22 22:28, Catalin Marinas wrote:
> On Sat, Aug 20, 2016 at 05:38:59PM +0800, zhong jiang wrote:
>> On 2016/8/19 12:11, Ganapatrao Kulkarni wrote:
>>> On Fri, Aug 19, 2016 at 9:30 AM, Ganapatrao Kulkarni
>>> <gpkulkarni@gmail.com> wrote:
>>>> On Fri, Aug 19, 2016 at 7:28 AM, zhong jiang <zhongjiang@huawei.com> wrote:
>>>>> On 2016/8/19 1:45, Ganapatrao Kulkarni wrote:
>>>>>> On Thu, Aug 18, 2016 at 9:34 PM, Catalin Marinas
>>>>>> <catalin.marinas@arm.com> wrote:
>>>>>>> On Thu, Aug 18, 2016 at 09:09:26PM +0800, zhongjiang wrote:
>>>>>>>> At present, boot cpu will bound to a node from device tree when node_off enable.
>>>>>>>> if the node is not initialization, it will lead to a following problem.
> [...]
>>>>>>>> --- a/arch/arm64/mm/numa.c
>>>>>>>> +++ b/arch/arm64/mm/numa.c
>>>>>>>> @@ -119,7 +119,7 @@ void numa_store_cpu_info(unsigned int cpu)
>>>>>>>>  void __init early_map_cpu_to_node(unsigned int cpu, int nid)
>>>>>>>>  {
>>>>>>>>       /* fallback to node 0 */
>>>>>>>> -     if (nid < 0 || nid >= MAX_NUMNODES)
>>>>>>>> +     if (nid < 0 || nid >= MAX_NUMNODES || numa_off)
>>>>>>
>>>>>> i  did not understood how this line change fixes the issue that you
>>>>>> have mentioned (i too not understood fully the issue description)
>>>>>> this array used while mapping node id when secondary cores comes up
>>>>>> when numa_off is set the cpu_to_node_map[cpu] is not used and set to
>>>>>> node0 always( refer function numa_store_cpu_info)..
>>>>>> please provide more details to understand the issue you are facing.
>>>>>> /*
>>>>>>  *  Set the cpu to node and mem mapping
>>>>>>  */
>>>>>> void numa_store_cpu_info(unsigned int cpu)
>>>>>> {
>>>>>>         map_cpu_to_node(cpu, numa_off ? 0 : cpu_to_node_map[cpu]);
>>>>>> }
>>>>>
>>>>> The issue comes up when we test the kdump. it will leads to kernel crash.
>>>>> when I debug the issue, I find boot cpu actually bound to the node1. while
>>>>> node1 is not real existence when numa_off enable.
>>>>
>>>> boot cpu is default mapped to node0
>>>> are you running with any other patches?
He applied my patches, which I mentioned these days.

I chated with ZhongJiang, this problem is only exist for my patches, and no matter
whether use kdump or not. Mainline doesn't have this problem.

The details of this problem is(suppose numa_off is true), according to the code execution sequence :

1. setup_arch-->bootmem_init-->arm64_numa_init
When numa_off is true, all memory blocks will add into node 0.

2. setup_arch-->of_smp_init_cpus
I added early_map_cpu_to_node for boot cpu, so that the nid of cpu0 will change to the value read from dt node.
With ZhongJiang's patch, it will correct the nid of cpu0 to zero when numa_off is true.

3. build_all_zonelists
Because numa is off, so that only the control block of node 0 had been initialized. So cpu0 with non-zero nid will lead the kernel crash.

4. kernel_init_freeable-->smp_prepare_cpus-->smp_store_cpu_info
Set the nid of cpu0 to zero, but it's too late.

5. secondary_start_kernel-->smp_store_cpu_info
Set the nid of other cpus to zero.

I will update my patch series and resend it again.

Best regards,
     Town·Thunder
     (My Chinese name Zhen Lei direct translation into English)

>>>
>>> if you added any patch to change this code
>>>   /* init boot processor */
>>>         cpu_to_node_map[0] = 0;
>>>         map_cpu_to_node(0, 0);
>>>
>>> then adding code to take-care numa_off here might solve your issue.
>>
>>  but in of_smp_init_cpus, boot cpu will call early_map_cpu_to_node[] to get
>>  the relation node. and the node is from devicetree.
>>
>>  you points to the code will be covered with another node. therefore, it is
>>  possible that cpu_to_node[cpu] will leads to the incorrect results. therefore,
>>  The crash will come up.
> 
> I think I get Ganapat's point. The cpu_to_node_map[0] may be incorrectly
> set by early_map_cpu_to_node() when called from smp_init_cpus() ->
> of_parse_and_init_cpus(). However, the cpu_to_node_map[] array is *only*
> read by numa_store_cpu_info(). This latter function calls
> map_cpu_to_node() and, if numa_off, will only ever pass 0 as the nid.
> 
> Given that the cpu_to_node_map[] array is static, I don't see how any
> non-zero value could leak outside the arch/arm64/mm/numa.c file.
> 
> So please give more details of any additional patches you have on top of
> mainline or whether you reproduced this issue with the vanilla kernel
> (since you mentioned kdump, that's not in mainline yet).
>