Hi, On 10/21/2022 12:22 PM, Alexei Starovoitov wrote: > On Thu, Oct 20, 2022 at 7:26 PM Hou Tao <houtao@xxxxxxxxxxxxxxx> wrote: >> How about reject the NUMA node setting for non-preallocated hash table in >> hashtab.c ? > It's easy to ask the question, but please answer it yourself. > Analyze the code and describe what you think is happening now > and what should or should not be the behavior. > . I found it is a bad idea to reject the numa node setting for non-preallocated hash table. The reason is that now hash buckets are still allocated according to the numa node setting. If the numa node setting is rejected, the use case below won't work normally, so I will keep it as-is: 1. a non-preallocated hash table is created on numa node 2 (buckets are allocated from node 2) 2. bpf program is running on numa node 2 (elements are also allocated from node 2) 3. all used memories are allocated from node 2.
