Hi Tejun, 在 2019/5/10 0:48, Tejun Heo 写道:
Hi Tejun, On Thu, May 09, 2019 at 04:03:53PM +0800, zhangliguang wrote:There might have tons of files queued in the writeback, awaiting for writing back. Unfortunately, the writeback's cgroup has been dead. In this case, we reassociate the inode with another writeback cgroup, but we possibly can't because the writeback associated with the dead cgroup is the only valid one. In this case, the new writeback is allocated, initialized and associated with the inode. It causes unnecessary high system load and latency. This fixes the issue by enforce moving the inode to root cgroup when the previous binding cgroup becomes dead. With it, no more unnecessary writebacks are created, populated and the system load decreased by about 6x in the online service we encounted: Without the patch: about 30% system load With the patch: about 5% system loadCan you please describe the scenario with more details? I'm having a bit of hard time understanding the amount of cpu cycles being consumed. Thanks.
Our search line reported a problem, when containerA was removed, containerB and containerC's system load were up to 30%. We record the trace with 'perf record cycles:k -g -a', found that wb_init was the hotspot function. Function call: generic_file_direct_write filemap_write_and_wait_range __filemap_fdatawrite_range wbc_attach_fdatawrite_inode inode_attach_wb __inode_attach_wb wb_get_create wbc_attach_and_unlock_inode if (unlikely(wb_dying(wbc->wb))) inode_switch_wbs wb_get_create ; Search bdi->cgwb_tree from memcg_css->id ; OR cgwb_create kmalloc wb_init // hot spot ; Insert to bdi->cgwb_tree, mmecg_css->id as key We discussed it through, base on the analysis: When we running into the issue, there is cgroups are being deleted. The inodes (files) that were associated with these cgroups have to switch into another newly created writeback. We think there are huge amount of inodes in the writeback list that time. So we don't think there is anything abnormal. However, one thing we possibly can do: enforce these inodes to BDI embedded wirteback and we needn't create huge amount of writebacks in that case, to avoid the high system load phenomenon. We expect correct wb (best candidate) is picked up in next round. Thanks, Liguang
