[PATCH v4 11/14] arm64/numa: support HAVE_MEMORYLESS_NODES

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

On 2016/6/7 22:01, Ganapatrao Kulkarni wrote:
> On Tue, Jun 7, 2016 at 6:27 PM, Leizhen (ThunderTown)
> <thunder.leizhen@huawei.com> wrote:
>>
>>
>> On 2016/6/7 16:31, Ganapatrao Kulkarni wrote:
>>> On Tue, Jun 7, 2016 at 1:38 PM, Zhen Lei <thunder.leizhen@huawei.com> wrote:
>>>> Some numa nodes may have no memory. For example:
>>>> 1. cpu0 on node0
>>>> 2. cpu1 on node1
>>>> 3. device0 access the momory from node0 and node1 take the same time.
>>>
>>> i am wondering, if access to both nodes is same, then why you need numa.
>>> the example you are quoting is against the basic principle of "numa"
>>> what is device0 here? cpu?
>> The device0 can also be a cpu. I drew a simple diagram:
>>
>>   cpu0     cpu1        cpu2/device0
>>     |        |              |
>>     |        |              |
>>    DDR0     DDR1    No DIMM slots or no DIMM plugged
>>  (node0)  (node1)         (node2)
>>
> 
> thanks for the clarification. your example is for 3 node system, where
> third node is memory less node.
> do you see any issue in supporting this topology with existing code?
If opened HAVE_MEMORYLESS_NODES, it will pick the nearest node for the cpus on
memoryless node.

For example, in include/linux/topology.h
#ifdef CONFIG_HAVE_MEMORYLESS_NODES
...
static inline int cpu_to_mem(int cpu)
{
	return per_cpu(_numa_mem_, cpu);
}
...
#else
...
static inline int cpu_to_mem(int cpu)
{
	return cpu_to_node(cpu);
}
...
#endif

> I think, this use case should be supported with present code.
> 
>>>>
>>>> So, we can not simply classify device0 to node0 or node1, but we can
>>>> define a node2 which distances to node0 and node1 are the same.
>>>>
>>>> Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
>>>> ---
>>>>  arch/arm64/Kconfig      |  4 ++++
>>>>  arch/arm64/kernel/smp.c |  1 +
>>>>  arch/arm64/mm/numa.c    | 43 +++++++++++++++++++++++++++++++++++++++++--
>>>>  3 files changed, 46 insertions(+), 2 deletions(-)
>>>>
>>>> diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
>>>> index 05c1bf1..5904a62 100644
>>>> --- a/arch/arm64/Kconfig
>>>> +++ b/arch/arm64/Kconfig
>>>> @@ -581,6 +581,10 @@ config NEED_PER_CPU_EMBED_FIRST_CHUNK
>>>>         def_bool y
>>>>         depends on NUMA
>>>>
>>>> +config HAVE_MEMORYLESS_NODES
>>>> +       def_bool y
>>>> +       depends on NUMA
>>>> +
>>>>  source kernel/Kconfig.preempt
>>>>  source kernel/Kconfig.hz
>>>>
>>>> diff --git a/arch/arm64/kernel/smp.c b/arch/arm64/kernel/smp.c
>>>> index d099306..9e15297 100644
>>>> --- a/arch/arm64/kernel/smp.c
>>>> +++ b/arch/arm64/kernel/smp.c
>>>> @@ -620,6 +620,7 @@ static void __init of_parse_and_init_cpus(void)
>>>>                         }
>>>>
>>>>                         bootcpu_valid = true;
>>>> +                       early_map_cpu_to_node(0, of_node_to_nid(dn));
>>>>
>>>>                         /*
>>>>                          * cpu_logical_map has already been
>>>> diff --git a/arch/arm64/mm/numa.c b/arch/arm64/mm/numa.c
>>>> index df5c842..d73b0a0 100644
>>>> --- a/arch/arm64/mm/numa.c
>>>> +++ b/arch/arm64/mm/numa.c
>>>> @@ -128,6 +128,14 @@ void __init early_map_cpu_to_node(unsigned int cpu, int nid)
>>>>                 nid = 0;
>>>>
>>>>         cpu_to_node_map[cpu] = nid;
>>>> +
>>>> +       /*
>>>> +        * We should set the numa node of cpu0 as soon as possible, because it
>>>> +        * has already been set up online before. cpu_to_node(0) will soon be
>>>> +        * called.
>>>> +        */
>>>> +       if (!cpu)
>>>> +               set_cpu_numa_node(cpu, nid);
>>>>  }
>>>>
>>>>  #ifdef CONFIG_HAVE_SETUP_PER_CPU_AREA
>>>> @@ -215,6 +223,35 @@ int __init numa_add_memblk(int nid, u64 start, u64 end)
>>>>         return ret;
>>>>  }
>>>>
>>>> +static u64 __init alloc_node_data_from_nearest_node(int nid, const size_t size)
>>>> +{
>>>> +       int i, best_nid, distance;
>>>> +       u64 pa;
>>>> +       DECLARE_BITMAP(nodes_map, MAX_NUMNODES);
>>>> +
>>>> +       bitmap_zero(nodes_map, MAX_NUMNODES);
>>>> +       bitmap_set(nodes_map, nid, 1);
>>>> +
>>>> +find_nearest_node:
>>>> +       best_nid = NUMA_NO_NODE;
>>>> +       distance = INT_MAX;
>>>> +
>>>> +       for_each_clear_bit(i, nodes_map, MAX_NUMNODES)
>>>> +               if (numa_distance[nid][i] < distance) {
>>>> +                       best_nid = i;
>>>> +                       distance = numa_distance[nid][i];
>>>> +               }
>>>> +
>>>> +       pa = memblock_alloc_nid(size, SMP_CACHE_BYTES, best_nid);
>>>> +       if (!pa) {
>>>> +               BUG_ON(best_nid == NUMA_NO_NODE);
>>>> +               bitmap_set(nodes_map, best_nid, 1);
>>>> +               goto find_nearest_node;
>>>> +       }
>>>> +
>>>> +       return pa;
>>>> +}
>>>> +
> 
> why do we need this function in arch specific code.
I also considered put these code(include HAVE_SETUP_PER_CPU_AREA) into drivers/of/of_numa.c,
but if I do that, it will make acpi numa dependent on of numa.

> dont you think common code will take care of this? when you define
> HAVE_MEMORYLESS_NODES

I have searched CONFIG_HAVE_MEMORYLESS_NODES in *.c, but did not find the relevant content.
So maybe other ARCHs also missed this.

> 
>>>>  /**
>>>>   * Initialize NODE_DATA for a node on the local memory
>>>>   */
>>>> @@ -228,7 +265,9 @@ static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
>>>>         pr_info("Initmem setup node %d [mem %#010Lx-%#010Lx]\n",
>>>>                 nid, start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
>>>>
>>>> -       nd_pa = memblock_alloc_try_nid(nd_size, SMP_CACHE_BYTES, nid);
> 
> this function try to allocate from a nid, if fails, it allocates from
> node 0(local node).
> this is ok for memory less node i guess.
Yes, the function is OK, but the performance is not.

Suppose there are 3 nodes:
1. cpu0 on node0, cpu1 on node1, cpu2 on node2.
2. cpu2 access the memory on node1 take 1us, but access the memory on node1 take 5us.
   That is, distance[2,1] is shorter than distance[2,0].
3. And node2 is a memoryless node.

So if NODE_DATA(2) allocated from node0, it will take more time than allocted from node1 at run time.
Because NODE_DATA will be accessed at run time.

> 
>>>> +       nd_pa = memblock_alloc_nid(nd_size, SMP_CACHE_BYTES, nid);
>>>> +       if (!nd_pa)
>>>> +               nd_pa = alloc_node_data_from_nearest_node(nid, nd_size);
>>>>         nd = __va(nd_pa);
>>>>
>>>>         /* report and initialize */
>>>> @@ -238,7 +277,7 @@ static void __init setup_node_data(int nid, u64 start_pfn, u64 end_pfn)
>>>>         if (tnid != nid)
>>>>                 pr_info("    NODE_DATA(%d) on node %d\n", nid, tnid);
>>>>
>>>> -       node_data[nid] = nd;
>>>> +       NODE_DATA(nid) = nd;
>>>>         memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
>>>>         NODE_DATA(nid)->node_id = nid;
>>>>         NODE_DATA(nid)->node_start_pfn = start_pfn;
>>>> --
>>>> 2.5.0
>>>>
>>>>
>>> Ganapat
>>>>
>>>> _______________________________________________
>>>> linux-arm-kernel mailing list
>>>> linux-arm-kernel at lists.infradead.org
>>>> http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
>>>
>>> .
>>>
>>
> 
> .
>