[ovs-dev] [PATCH] ovs: do not allocate memory from offline numa node

[Jarno Rajahalme]

> On Oct 9, 2015, at 3:11 PM, Jesse Gross <jesse at nicira.com> wrote:
> 
> On Fri, Oct 9, 2015 at 8:54 AM, Jarno Rajahalme <jrajahalme at nicira.com <mailto:jrajahalme at nicira.com>> wrote:
>> 
>> On Oct 8, 2015, at 4:03 PM, Jesse Gross <jesse at nicira.com> wrote:
>> 
>> On Wed, Oct 7, 2015 at 10:47 AM, Jarno Rajahalme <jrajahalme at nicira.com>
>> wrote:
>> 
>> 
>> On Oct 6, 2015, at 6:01 PM, Jesse Gross <jesse at nicira.com> wrote:
>> 
>> On Mon, Oct 5, 2015 at 1:25 PM, Alexander Duyck
>> <alexander.duyck at gmail.com> wrote:
>> 
>> On 10/05/2015 06:59 AM, Vlastimil Babka wrote:
>> 
>> 
>> On 10/02/2015 12:18 PM, Konstantin Khlebnikov wrote:
>> 
>> 
>> When openvswitch tries allocate memory from offline numa node 0:
>> stats = kmem_cache_alloc_node(flow_stats_cache, GFP_KERNEL | __GFP_ZERO,
>> 0)
>> It catches VM_BUG_ON(nid < 0 || nid >= MAX_NUMNODES || !node_online(nid))
>> [ replaced with VM_WARN_ON(!node_online(nid)) recently ] in linux/gfp.h
>> This patch disables numa affinity in this case.
>> 
>> Signed-off-by: Konstantin Khlebnikov <khlebnikov at yandex-team.ru>
>> 
>> 
>> 
>> ...
>> 
>> diff --git a/net/openvswitch/flow_table.c b/net/openvswitch/flow_table.c
>> index f2ea83ba4763..c7f74aab34b9 100644
>> --- a/net/openvswitch/flow_table.c
>> +++ b/net/openvswitch/flow_table.c
>> @@ -93,7 +93,8 @@ struct sw_flow *ovs_flow_alloc(void)
>> 
>>    /* Initialize the default stat node. */
>>    stats = kmem_cache_alloc_node(flow_stats_cache,
>> -                      GFP_KERNEL | __GFP_ZERO, 0);
>> +                      GFP_KERNEL | __GFP_ZERO,
>> +                      node_online(0) ? 0 : NUMA_NO_NODE);
>> 
>> 
>> 
>> Stupid question: can node 0 become offline between this check, and the
>> VM_WARN_ON? :) BTW what kind of system has node 0 offline?
>> 
>> 
>> 
>> Another question to ask would be is it possible for node 0 to be online, but
>> be a memoryless node?
>> 
>> I would say you are better off just making this call kmem_cache_alloc.  I
>> don't see anything that indicates the memory has to come from node 0, so
>> adding the extra overhead doesn't provide any value.
>> 
>> 
>> I agree that this at least makes me wonder, though I actually have
>> concerns in the opposite direction - I see assumptions about this
>> being on node 0 in net/openvswitch/flow.c.
>> 
>> Jarno, since you original wrote this code, can you take a look to see
>> if everything still makes sense?
>> 
>> 
>> We keep the pre-allocated stats node at array index 0, which is initially
>> used by all CPUs, but if CPUs from multiple numa nodes start updating the
>> stats, we allocate additional stats nodes (up to one per numa node), and the
>> CPUs on node 0 keep using the preallocated entry. If stats cannot be
>> allocated from CPUs local node, then those CPUs keep using the entry at
>> index 0. Currently the code in net/openvswitch/flow.c will try to allocate
>> the local memory repeatedly, which may not be optimal when there is no
>> memory at the local node.
>> 
>> Allocating the memory for the index 0 from other than node 0, as discussed
>> here, just means that the CPUs on node 0 will keep on using non-local memory
>> for stats. In a scenario where there are CPUs on two nodes (0, 1), but only
>> the node 1 has memory, a shared flow entry will still end up having separate
>> memory allocated for both nodes, but both of the nodes would be at node 1.
>> However, there is still a high likelihood that the memory allocations would
>> not share a cache line, which should prevent the nodes from invalidating
>> each other’s caches. Based on this I do not see a problem relaxing the
>> memory allocation for the default stats node. If node 0 has memory, however,
>> it would be better to allocate the memory from node 0.
>> 
>> 
>> Thanks for going through all of that.
>> 
>> It seems like the question that is being raised is whether it actually
>> makes sense to try to get the initial memory on node 0, especially
>> since it seems to introduce some corner cases? Is there any reason why
>> the flow is more likely to hit node 0 than a randomly chosen one?
>> (Assuming that this is a multinode system, otherwise it's kind of a
>> moot point.) We could have a separate pointer to the default allocated
>> memory, so it wouldn't conflict with memory that was intentionally
>> allocated for node 0.
>> 
>> 
>> It would still be preferable to know from which node the default stats node
>> was allocated, and store it in the appropriate pointer in the array. We
>> could then add a new “default stats node index” that would be used to locate
>> the node in the array of pointers we already have. That way we would avoid
>> extra allocation and processing of the default stats node.
> 
> I agree, that sounds reasonable to me. Will you make that change?
> 
> Besides eliminating corner cases, it might help performance in some
> cases too by avoiding stressing memory bandwidth on node 0.

I’ll do this,

  Jarno