On 2023/3/8 17:13, David Hildenbrand wrote: > On 08.03.23 10:03, Haifeng Xu wrote: >> >> >> On 2023/3/7 10:48, Matthew Wilcox wrote: >>> On Tue, Mar 07, 2023 at 10:36:55AM +0800, Haifeng Xu wrote: >>>> On 2023/3/6 21:49, David Hildenbrand wrote: >>>>> On 06.03.23 03:49, Haifeng Xu wrote: >>>>>> mem_cgroup_oom_synchronize() has checked whether current memcg_in_oom is >>>>>> set or not, so remove the check in handle_mm_fault(). >>>>> >>>>> "mem_cgroup_oom_synchronize() will returned immediately if memcg_in_oom is not set, so remove the check from handle_mm_fault()". >>>>> >>>>> However, that requires now always an indirect function call -- do we care about dropping that optimization? >>>>> >>>>> >>>> >>>> If memcg_in_oom is set, we will check it twice, one is from handle_mm_fault(), the other is from mem_cgroup_oom_synchronize(). That seems a bit redundant. >>>> >>>> if memcg_in_oom is not set, mem_cgroup_oom_synchronize() returns directly. Though it's an indirect function call, but the time spent can be negligible >>>> compare to the whole mm user falut preocess. And that won't cause stack overflow error. >>> >>> I suggest you measure it. >> >> test steps: >> 1) Run command: ./mmap_anon_test(global alloc, so the memcg_in_oom is not set) >> 2) Calculate the quotient of cost time and page-fault counts, run 10 rounds and average the results. >> >> The test result shows that whether using indirect function call or not, the time spent in user fault >> is almost the same, about 2.3ms. > > I guess most of the benchmark time is consumed by allocating fresh pages in your test (also, why exactly do you use MAP_SHARED?). > > Is 2.3ms the total time for writing to that 1GiB of memory or how did you derive that number? Posting both results would be cleaner (with more digits ;) ). > Hi Daivd, the details of test result were posted last week. Do you have any suggestions or more concerns about this change? Thanks.
