On Wed, 2020-08-19 at 10:01 +0800, Xiubo Li wrote: > On 2020/8/19 4:10, Patrick Donnelly wrote: > > On Tue, Aug 18, 2020 at 1:05 PM Jeff Layton <jlayton@xxxxxxxxxx> wrote: > > > Bear in mind that if the same file has been opened several times, then > > > you'll get an increment for each. > > Having an open file count (even for the same inode) and a count of > > inodes opened sounds useful to me. The latter would require some kind > > of refcounting for each inode? Maybe that's too expensive though. > > For the second, yes, we need one percpu refcount, which can be add in > ceph_get_fmode() when increasing any entry of the > ci->i_nr_by_mode[fmode] for the first time, to decrease it in > ceph_put_fmode() when the ci->i_nr_by_mode[fmode] is empty. IMO, it > should be okay and won't cost too much. > > Thanks > > BRs > Sure. To be clear, I'm not _really_ disputing the usefulness of these stats, but I think if we're going to measure stuff like this, the changelog needs to justify it in some way. We may find in 2-3 years that some of these are not as useful as we first thought, and if we don't have any of the original justification in the changelog for it, it's harder to determine whether removing them is ok. > > > Would it potentially be more useful to report the number of inodes that > > > have open file descriptions associated with them? It's hard for me to > > > know as I'm not clear on the intended use-case for this. > > Use-case is more information available via `fs top`. > > -- Jeff Layton <jlayton@xxxxxxxxxx>
