On Wed, Apr 1, 2020 at 3:58 PM David Howells <dhowells@redhat.com> wrote:
>
> David Howells <dhowells@redhat.com> wrote:
>
> > > Attached patch applies against readfile patch.
> >
> > But doesn't actually do what Karel asked for.  show_mountinfo() itself does
> > not give you what Karel asked for.

Not sure what you mean.  I think it shows precisely the information
Karel asked for.

>  Plus there's more information you need to
> > add to it.

The mountinfo format is extensible (see
Documentation/filesystems/proc.txt) so for example adding the change
counters would be simple.

> And arguably, it's worse than just reading /proc/mounts.  If you get a
> notification that something changed (ie. you poll /proc/mounts or mount
> notifications gives you an overrun) you now have to read *every*
> /mountfs/*/info file.  That is way more expensive.

fsinfo(2) will never be substantially cheaper than reading and parsing
/mnt/MNT_ID/info.  In fact reading a large part of the mount table
using fsinfo(2) will be substantially slower than parsing
/proc/self/mountinfo (this doesn't actually do the parsing but that
would add a very small amount of overhead):

root@kvm:~# ./test-fsinfo-perf /tmp/a 30000
--- make mounts ---
--- test fsinfo by path ---
sum(mnt_id) = 960000
--- test fsinfo by mnt_id ---
sum(mnt_id) = 960000
--- test /proc/fdinfo ---
sum(mnt_id) = 960000
--- test mountfs ---
sum(mnt_id) = 960000
--- test mountinfo ---
sum(mnt_id) = 960000
For   30000 mounts, f=    154963us f2=    148337us p=   1803699us p2=
  257019us; m=     53996us; p=11.6*f p=12.2*f2 p=7.0*p2 p=33.4*m
--- umount ---

Yes, that's 33 times faster!

Thanks,
Miklos