On Wed, Aug 03, 2016 at 05:17:09PM +0300, Nikolay Borisov wrote:
>
>
> On 08/03/2016 04:46 PM, Jeff Layton wrote:
> > On Wed, 2016-08-03 at 10:35 +0300, Nikolay Borisov wrote:
> >> On busy container servers reading /proc/locks shows all the locks
> >> created by all clients. This can cause large latency spikes. In my
> >> case I observed lsof taking up to 5-10 seconds while processing around
> >> 50k locks. Fix this by limiting the locks shown only to those created
> >> in the same pidns as the one the proc was mounted in. When reading
> >> /proc/locks from the init_pid_ns show everything.
> >>
> >>> Signed-off-by: Nikolay Borisov <kernel@xxxxxxxx>
> >> ---
> >> fs/locks.c | 6 ++++++
> >> 1 file changed, 6 insertions(+)
> >>
> >> diff --git a/fs/locks.c b/fs/locks.c
> >> index ee1b15f6fc13..751673d7f7fc 100644
> >> --- a/fs/locks.c
> >> +++ b/fs/locks.c
> >> @@ -2648,9 +2648,15 @@ static int locks_show(struct seq_file *f, void *v)
> >> {
> >>> struct locks_iterator *iter = f->private;
> >>> struct file_lock *fl, *bfl;
> >>> + struct pid_namespace *proc_pidns = file_inode(f->file)->i_sb->s_fs_info;
> >>> + struct pid_namespace *current_pidns = task_active_pid_ns(current);
> >>
> >>> fl = hlist_entry(v, struct file_lock, fl_link);
> >>
> >>>> + if ((current_pidns != &init_pid_ns) && fl->fl_nspid
> >
> > Ok, so when you read from a process that's in the init_pid_ns
> > namespace, then you'll get the whole pile of locks, even when reading
> > this from a filesystem that was mounted in a different pid_ns?
> >
> > That seems odd to me if so. Any reason not to just uniformly use the
> > proc_pidns here?
>
> [CCing some people from openvz/CRIU]
>
> My train of thought was "we should have means which would be the one
> universal truth about everything and this would be a process in the
> init_pid_ns".
OK, but why not make that means be "mount proc from the init_pid_ns and
read /proc/locks there". So just replace current_pidns with proc_pidns
in the above. I think that's all Jeff was suggesting.
--b.
> I don't have strong preference as long as I'm not breaking
> userspace. As I said before - I think the CRIU guys might be using that
> interface.
>
> >
> >>>> + && (proc_pidns != ns_of_pid(fl->fl_nspid)))
> >>> + return 0;
> >> +
> >>> lock_get_status(f, fl, iter->li_pos, "");
> >>
> >>> list_for_each_entry(bfl, &fl->fl_block, fl_block)
> >
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
