Re: [PATCH] mm: page_alloc: don't allocate page from memoryless nodes

[David Hildenbrand <david@redhat.com>]

On 14.02.23 09:42, Vlastimil Babka wrote:
> On 2/13/23 12:00, Qi Zheng wrote:
>>
>>
>> On 2023/2/13 16:47, Vlastimil Babka wrote:
>>> On 2/12/23 12:03, Qi Zheng wrote:
>>>> In x86, numa_register_memblks() is only interested in
>>>> those nodes which have enough memory, so it skips over
>>>> all nodes with memory below NODE_MIN_SIZE (treated as
>>>> a memoryless node). Later on, we will initialize these
>>>> memoryless nodes (allocate pgdat in free_area_init()
>>>> and build zonelist etc), and will online these nodes
>>>> in init_cpu_to_node() and init_gi_nodes().
>>>>
>>>> After boot, these memoryless nodes are in N_ONLINE
>>>> state but not in N_MEMORY state. But we can still allocate
>>>> pages from these memoryless nodes.
>>>>
>>>> In SLUB, we only process nodes in the N_MEMORY state,
>>>> such as allocating their struct kmem_cache_node. So if
>>>> we allocate a page from the memoryless node above to
>>>> SLUB, the struct kmem_cache_node of the node corresponding
>>>> to this page is NULL, which will cause panic.
>>>>
>>>> For example, if we use qemu to start a two numa node kernel,
>>>> one of the nodes has 2M memory (less than NODE_MIN_SIZE),
>>>> and the other node has 2G, then we will encounter the
>>>> following panic:
>>>>
>>>> [    0.149844] BUG: kernel NULL pointer dereference, address: 0000000000000000
>>>> [    0.150783] #PF: supervisor write access in kernel mode
>>>> [    0.151488] #PF: error_code(0x0002) - not-present page
>>>> <...>
>>>> [    0.156056] RIP: 0010:_raw_spin_lock_irqsave+0x22/0x40
>>>> <...>
>>>> [    0.169781] Call Trace:
>>>> [    0.170159]  <TASK>
>>>> [    0.170448]  deactivate_slab+0x187/0x3c0
>>>> [    0.171031]  ? bootstrap+0x1b/0x10e
>>>> [    0.171559]  ? preempt_count_sub+0x9/0xa0
>>>> [    0.172145]  ? kmem_cache_alloc+0x12c/0x440
>>>> [    0.172735]  ? bootstrap+0x1b/0x10e
>>>> [    0.173236]  bootstrap+0x6b/0x10e
>>>> [    0.173720]  kmem_cache_init+0x10a/0x188
>>>> [    0.174240]  start_kernel+0x415/0x6ac
>>>> [    0.174738]  secondary_startup_64_no_verify+0xe0/0xeb
>>>> [    0.175417]  </TASK>
>>>> [    0.175713] Modules linked in:
>>>> [    0.176117] CR2: 0000000000000000
>>>>
>>>> In addition, we can also encountered this panic in the actual
>>>> production environment. We set up a 2c2g container with two
>>>> numa nodes, and then reserved 128M for kdump, and then we
>>>> can encountered the above panic in the kdump kernel.
>>>>
>>>> To fix it, we can filter memoryless nodes when allocating
>>>> pages.
>>>>
>>>> Signed-off-by: Qi Zheng <zhengqi.arch@bytedance.com>
>>>> Reported-by: Teng Hu <huteng.ht@bytedance.com>
>>>
>>> Well AFAIK the key mechanism to only allocate from "good" nodes is the
>>> zonelist, we shouldn't need to start putting extra checks like this. So it
>>> seems to me that the code building the zonelists should take the
>>> NODE_MIN_SIZE constraint in mind.
>>
>> Indeed. How about the following patch:
> 
> +Cc also David, forgot earlier.
> 
> Looks good to me, at least.
> 
>> @@ -6382,8 +6378,11 @@ int find_next_best_node(int node, nodemask_t
>> *used_node_mask)
>>           int min_val = INT_MAX;
>>           int best_node = NUMA_NO_NODE;
>>
>> -       /* Use the local node if we haven't already */
>> -       if (!node_isset(node, *used_node_mask)) {
>> +       /*
>> +        * Use the local node if we haven't already. But for memoryless
>> local
>> +        * node, we should skip it and fallback to other nodes.
>> +        */
>> +       if (!node_isset(node, *used_node_mask) && node_state(node,
>> N_MEMORY)) {
>>                   node_set(node, *used_node_mask);
>>                   return node;
>>           }
>>
>> For memoryless node, we skip it and fallback to other nodes when
>> build its zonelists.
>>
>> Say we have node0 and node1, and node0 is memoryless, then:
>>
>> [    0.102400] Fallback order for Node 0: 1
>> [    0.102931] Fallback order for Node 1: 1
>>
>> In this way, we will not allocate pages from memoryless node0.
>>

In offline_pages(), we'll first build_all_zonelists() to then 
node_states_clear_node()->node_clear_state(node, N_MEMORY);

So at least on the offlining path, we wouldn't detect it properly yet I 
assume, and build a zonelist that contains a now-memory-less node?

-- 
Thanks,

David / dhildenb