[PATCH] powerpc/numa: fix hot-added CPU on memory-less node

[Laurent Vivier]

On 15/11/2018 10:19, Satheesh Rajendran wrote:
> On Wed, Nov 14, 2018 at 06:03:19PM +0100, Laurent Vivier wrote:
>> Trying to hotplug a CPU on an empty NUMA node (without
>> memory or CPU) crashes the kernel when the CPU is onlined.
>>
>> During the onlining process, the kernel calls start_secondary()
>> that ends by calling
>> set_numa_mem(local_memory_node(numa_cpu_lookup_table[cpu]))
>> that relies on NODE_DATA(nid)->node_zonelists and in our case
>> NODE_DATA(nid) is NULL.
>>
>> To fix that, add the same checking as we already have in
>> find_and_online_cpu_nid(): if NODE_DATA() is NULL, use
>> the first online node.
>>
>> Bug: https://github.com/linuxppc/linux/issues/184
>> Fixes: ea05ba7c559c8e5a5946c3a94a2a266e9a6680a6
>>        (powerpc/numa: Ensure nodes initialized for hotplug)
>> Signed-off-by: Laurent Vivier <lvivier at redhat.com>
>> ---
>>  arch/powerpc/mm/numa.c | 9 +++++++++
>>  1 file changed, 9 insertions(+)
> 
> This patch causes regression for cold plug numa case(Case 1) and 
> hotplug case + reboot(Case 2) with adding all vcpus into node 0.
> 
> 
> Env: HW: Power8 Host.
> Kernel: 4.20-rc2 + this patch
> 
> Case 1:
> 1. boot a guest with 8 vcpus(all available), spreadout in 4 numa nodes.
> <vcpu placement='static'>8</vcpu>
> ...
>    <numa>
>       <cell id='0' cpus='0-1' memory='4194304' unit='KiB'/>
>       <cell id='1' cpus='2-3' memory='4194304' unit='KiB'/>
>       <cell id='2' cpus='4-5' memory='0' unit='KiB'/>
>       <cell id='3' cpus='6-7' memory='0' unit='KiB'/>
>     </numa>
> 
> 2. Check lscpu --- all vcpus are added to node0 --> NOK
> 
> # lscpu
> Architecture:        ppc64le
> Byte Order:          Little Endian
> CPU(s):              8
> On-line CPU(s) list: 0-7
> Thread(s) per core:  1
> Core(s) per socket:  8
> Socket(s):           1
> NUMA node(s):        4
> Model:               2.1 (pvr 004b 0201)
> Model name:          POWER8 (architected), altivec supported
> Hypervisor vendor:   KVM
> Virtualization type: para
> L1d cache:           64K
> L1i cache:           32K
> NUMA node0 CPU(s):   0-7
> NUMA node1 CPU(s):   
> NUMA node2 CPU(s):   
> NUMA node3 CPU(s): 
> 
> without this patch it was working fine.
> # lscpu
> Architecture:        ppc64le
> Byte Order:          Little Endian
> CPU(s):              8
> On-line CPU(s) list: 0-7
> Thread(s) per core:  1
> Core(s) per socket:  8
> Socket(s):           1
> NUMA node(s):        4
> Model:               2.1 (pvr 004b 0201)
> Model name:          POWER8 (architected), altivec supported
> Hypervisor vendor:   KVM
> Virtualization type: para
> L1d cache:           64K
> L1i cache:           32K
> NUMA node0 CPU(s):   0,1
> NUMA node1 CPU(s):   2,3
> NUMA node2 CPU(s):   4,5
> NUMA node3 CPU(s):   6,7
> 
> 
> Case 2:
> 1. boot a guest with 8 vcpus(2 available, 6 possible), spreadout in 4 numa nodes.
> <vcpu placement='static' current='2'>8</vcpu>
> ...
>    <numa>
>       <cell id='0' cpus='0-1' memory='0' unit='KiB'/>
>       <cell id='1' cpus='2-3' memory='4194304' unit='KiB'/>
>       <cell id='2' cpus='4-5' memory='0' unit='KiB'/>
>       <cell id='3' cpus='6-7' memory='0' unit='KiB'/>
>     </numa>
> 
> 2. Hotplug all vcpus
> # lscpu
> Architecture:        ppc64le
> Byte Order:          Little Endian
> CPU(s):              8
> On-line CPU(s) list: 0-7
> Thread(s) per core:  1
> Core(s) per socket:  8
> Socket(s):           1
> NUMA node(s):        2
> Model:               2.1 (pvr 004b 0201)
> Model name:          POWER8 (architected), altivec supported
> Hypervisor vendor:   KVM
> Virtualization type: para
> L1d cache:           64K
> L1i cache:           32K
> NUMA node0 CPU(s):   0,1,4-7
> NUMA node1 CPU(s):   2,3
> 
> 
> 3. reboot the guest
> # lscpu
> Architecture:        ppc64le
> Byte Order:          Little Endian
> CPU(s):              8
> On-line CPU(s) list: 0-7
> Thread(s) per core:  1
> Core(s) per socket:  8
> Socket(s):           1
> NUMA node(s):        4
> Model:               2.1 (pvr 004b 0201)
> Model name:          POWER8 (architected), altivec supported
> Hypervisor vendor:   KVM
> Virtualization type: para
> L1d cache:           64K
> L1i cache:           32K
> NUMA node0 CPU(s):   0-7
> NUMA node1 CPU(s):
> NUMA node2 CPU(s):
> NUMA node3 CPU(s):
> 
> 
> Without this patch, Case 2 crashes the guest during hotplug, i.e
> issue reported in https://github.com/linuxppc/linux/issues/184
> 
> Regards,
> -Satheesh.
> 
>>
>> diff --git a/arch/powerpc/mm/numa.c b/arch/powerpc/mm/numa.c
>> index 3a048e98a132..1b2d25a3c984 100644
>> --- a/arch/powerpc/mm/numa.c
>> +++ b/arch/powerpc/mm/numa.c
>> @@ -483,6 +483,15 @@ static int numa_setup_cpu(unsigned long lcpu)
>>  	if (nid < 0 || !node_possible(nid))
>>  		nid = first_online_node;
>>
>> +	if (NODE_DATA(nid) == NULL) {
>> +		/*
>> +		 * Default to using the nearest node that has memory installed.
>> +		 * Otherwise, it would be necessary to patch the kernel MM code
>> +		 * to deal with more memoryless-node error conditions.
>> +		 */
>> +		nid = first_online_node;
>> +	}
>> +
>>  	map_cpu_to_node(lcpu, nid);
>>  	of_node_put(cpu);
>>  out:
>> -- 
>> 2.17.2
>>
> 

I have worked a while on this problem, and I don't see any easy fix for
that. It seems kernel is not ready to online a memory-less/cpu-less node
when someone hotplug a CPU in it. I think we have to fix several areas
to be able to do that.

Perhaps someone from IBM could have a better view on what we need?

Michael? Nathan?

Thanks,
Laurent