Re: [PATCH v3 0/2] quota: Add mountpath based quota support

From: Christian Brauner
Date: Wed May 12 2021 - 13:29:25 EST


On Wed, May 12, 2021 at 03:14:29PM +0200, Jan Kara wrote:
> On Wed 12-05-21 14:53:10, Christian Brauner wrote:
> > On Wed, May 12, 2021 at 01:01:49PM +0200, Jan Kara wrote:
> > > Added a few more CCs.
> > >
> > > On Tue 16-03-21 12:29:16, Jan Kara wrote:
> > > > On Thu 04-03-21 13:35:38, Sascha Hauer wrote:
> > > > > Current quotactl syscall uses a path to a block device to specify the
> > > > > filesystem to work on which makes it unsuitable for filesystems that
> > > > > do not have a block device. This series adds a new syscall quotactl_path()
> > > > > which replaces the path to the block device with a mountpath, but otherwise
> > > > > behaves like original quotactl.
> > > > >
> > > > > This is done to add quota support to UBIFS. UBIFS quota support has been
> > > > > posted several times with different approaches to put the mountpath into
> > > > > the existing quotactl() syscall until it has been suggested to make it a
> > > > > new syscall instead, so here it is.
> > > > >
> > > > > I'm not posting the full UBIFS quota series here as it remains unchanged
> > > > > and I'd like to get feedback to the new syscall first. For those interested
> > > > > the most recent series can be found here: https://lwn.net/Articles/810463/
> > > >
> > > > Thanks. I've merged the two patches into my tree and will push them to
> > > > Linus for the next merge window.
> > >
> > > So there are some people at LWN whining that quotactl_path() has no dirfd
> > > and flags arguments for specifying the target. Somewhat late in the game
> > > but since there's no major release with the syscall and no userspace using
> > > it, I think we could still change that. What do you think? What they
> > > suggest does make some sense. But then, rather then supporting API for
> > > million-and-one ways in which I may wish to lookup a fs object, won't it be
> > > better to just pass 'fd' in the new syscall (it may well be just O_PATH fd
> > > AFAICT) and be done with that?
> >
> > I think adding a dirfd argument makes a lot of sense (Unless there are
> > some restrictions around quotas I'm misunderstanding.).
> >
> > If I may: in general, I think we should aim to not add additional system
> > calls that operate on paths only. Purely path-based apis tend to be the
> > source of security issues especially when scoped lookups are really
> > important which given the ubiquity of sandboxing solutions nowadays is
> > quite often actually.
> > For example, when openat2() landed it gave such a boost in lookup
> > capabilities that I switched some libraries over to only ever do scoped
> > lookups, i.e. I decide on a starting point that gets opened path-based
> > and then explicitly express how I want that lookup to proceed ultimately
> > opening the final path component on which I want to perform operations.
> > Combined with the mount API almost everything can be done purely fd
> > based.
> >
> > In addition to that dirfd-scopable system calls allow for a much nicer
> > api experience when programming in userspace.
>
> OK, thanks for your insights. But when we add 'dirfd' I wonder whether we
> still need the 'path' component then. I mean you can always do fd =
> openat2(), quotactl_fd(fd, ...). After all ioctl() works exactly that way
> since the beginning. The only advantage of quotactl_xxx() taking path would
> be saving the open(2) call. That is somewhat convenient for simple cases
> (but also error prone in complex setups as you point out) and can be also
> sligthly faster (but quotactl is hardly a performance sensitive thing)...

That's a bit tricky indeed. It would feel consistent to add a path
argument as most of our fs apis seems to work that way even stuff like
fanotify_mark() but indeed a fd-only based api would be fine too. I
would try to follow recent additions/prior art here, I think.