在 2020/9/21 下午10:50, Tejun Heo 写道: > Hello, > > On Tue, Sep 15, 2020 at 11:53:49PM +0800, Chengming Zhou wrote: >> In the cgroup v1, we can use netlink interface to get cgroupstats for >> a cgroup. But it has been excluded from cgroup v2 interface intentionally >> due to the duplication and inconsistencies with other statistics. >> To make container monitor tool like "cadvisor" continue to work, we add >> these cgroupstats numbers to the cgroup.stat file, and change the >> admin-guide doc accordingly. > So, we can't add O(nr_threads) operations to cgroup.stat reads. There are > two ways forward that I can see. > > * Investigate how these counters are being used. If it's used for congestion > detection, pressure metrics are likely better indicators to use anyway. If > the usage frequency is low enough, maybe querying from userspace should > work? > > * If the need for these per-cgroup task state counters is really > justifiable, the counters should be maintained from scheduling event > directly and summed up using rstat like other statistics. > > Thanks. > Well, I see. These counters are being used for load monitor and debug, pressure metrics is good for congestion detection, but more details needed for debug when problem happened. And I noticed PSI has maintained these taskstats from scheduler event already. I think we can just export these counters gracefully. Thanks.
