Created attachment 1349 [details] perf record & mem use Description of problem: Environment information： CPU: Hygon 7360 * 2 Mem: 16G * 8 Network adapter：MLX IB 200GE (driver version 24.10-1.1.4) Compared to the 4.19 kernel version under the same use case, the memory bandwidth on the NUMA node where the network interface card resides is nearly saturated by the operating system during data reception（by TCP protocol）. In the 4.19 kernel version, both memory bandwidth and CPU utilization are reduced by 30%. Captured via perf, the time slice consumed by the function “copy_user_generic_string” increased by 30% in the 5.10 kernel. Version-Release number of selected component (if applicable): Sugon os kernel：5.10.134-17.1.3 (base on Anolis 8.9, and has picked this issue:https://gitee.com/anolis/cloud-kernel/pulls/3496) How reproducible: Steps to Reproduce: 1、When sending TCP requests of 4MB in size from multiple clients to a server with a Hygon CPU (3rd generation), if the server memory of the NUMA node where the network interface card (NIC) resides is 16GB, the network bandwidth of a single server must be driven to 2.5 GB/s . 2、Under the same operation and environment, compared to the 4.19 kernel. 3、Monitor the memory bandwidth and CPU utilization on the NUMA node where the network interface card resides. Actual results: In the 4.19 kernel version, both memory bandwidth and CPU utilization are reduced by 30%. Expected results: The memory bandwidth and CPU utilization of the 4.19 kernel are on par with those of the 5.10 kernel. Additional info: