Re: [PATCH] locks: export device name

From: Andrew Morton
Date: Thu Feb 16 2012 - 17:59:43 EST


On Thu, 16 Feb 2012 23:37:11 +0100
Davidlohr Bueso <dave@xxxxxxx> wrote:

> On Wed, 2012-02-15 at 12:39 -0800, Andrew Morton wrote:
> > On Wed, 15 Feb 2012 07:42:30 -0500
> > "J. Bruce Fields" <bfields@xxxxxxxxxxxx> wrote:
> >
> > > > > Perhaps safest would be to replace /proc/locks by another interface and
> > > > > deprecate this one.
> > > >
> > > > If exporting the name in the current /proc/locks file is out of the
> > > > question, then IMHO I don't think it would be worth adding a new
> > > > interface just for such a small change.
> > >
> > > OK.
> > >
> > > If you want to just change this over, I guess the thing to do would be
> > > to stick something in feature-removal-schedule.txt saying "we'll switch
> > > this in 2 years" (or however long you think before there are
> > > realistically no more lslk users left), then do it then.
> > >
> > > Switching to a new api would be better as we could warn users of the old
> > > api then. Maybe it'd be worth it if there was some other change we'd
> > > been wanting to make? Can't think of anything off the top of my head.
> > >
> > > We may be adding more lock types--will lslk and lslocks handle that
> > > gracefully?
> >
> > Adding a whole new interface is pretty attractive. It lets us get it
> > right this time. In particular, something which is extensible given
> > certain simple rules. As we've learned, the current /proc/locks didn't
> > get that right!
>
> Ok, however I'm a bit confused on what you mean by extensible; since
> what we decide to export to userspace is pretty much permanent, how can
> we change (extend) it later? We'd pretty much be running into
> the /proc/locks situation now.

Mainly by avoiding the use of implicit identification via fixed
positions. Look at /proc/stat and weep.

If we use a name:value format then we can add new fields later and
things work OK. We add stuff to /proc/meminfo and /proc/vmstat all the
time.

Removing things is of course much harder. The best fix is to avoid
adding things which we might ever have a reason for removing! If we
have a field which we simply can no longer support and which we think
we must retain for back-compat reasons then we just have to find some
way to emulate it. In extremis we could hardwire the value to "0" so
tools won't crash.

sysfs has a different convention: one-value-per-file. That means
there's no need for the name part of name:value. Extensibility means
"go add another sysfs file".

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/