anolis: futex: Allow disbling per numa node futex from kernel cmdline · Pull Request !5514 · anolis/cloud-kernel

bugzilla: https://bugzilla.openanolis.cn/show_bug.cgi?id=21547

Currently, the Linux kernel allocates memory for the futex hash table
without considering the NUMA node. This can result in futex hash table
accesses crossing NUMA nodes. If the memory latency between NUMA nodes
is sufficiently large, the futex access latency may also increase.

To address this, we introduce a per-NUMA-node futex hash table. To
maintain user-space compatibility, the NUMA node index for the futex
hash table is selected when the process first accesses the futex and is
used for the entire process lifecycle.

This solution targets scenarios where users bind processes/workloads to
specific NUMA nodes. In such cases, cross-NUMA-node futex hash table
accesses are guaranteed to be avoided. If no NUMA node binding is
present, cross-NUMA-node accesses may still occur. However, this will
not be worse than the vanilla kernel.

We collected results from the perf bench futex test suite and included
them in the commit message for patch [4/5].

In addition to perf bench testing, we performed the following
validations:
  - Kernel self-tests (futex): No new regressions were introduced.
  - Schbench: No deadlocks and no futex waiters were not woken.
  - Stress-ng mutex/futex tests: No deadlocks and no mutex/futexwaiters
                                 were not woken.
  - PTG QA test suite: No functional regressions were in introduced.

类型	指派人员	状态
审查	hsiangkao josephhz JeffleXu zhongjiang winterddd xueshuai 赵鑫鹏陶义 vxuyu dtc baolin.wang health_jay	已完成 (0/0人)
测试	小龙已测试通过	已完成 (1/1人)

GVPanolis/cloud-kernel

搜索帮助

GVP anolis/cloud-kernel