bumping mount path lengths in struct statfs

[Kostik Belousov]

On Wed, Jun 22, 2011 at 01:13:28PM +0300, Gleb Kurtsou wrote:
> On (21/06/2011 22:21), Kostik Belousov wrote:
> > On Tue, Jun 21, 2011 at 09:16:18AM -0600, Will Andrews wrote:
> > > Hi,
> > > 
> > > struct statfs contains the following:
> > > 
> > >  90         char      f_mntfromname[MNAMELEN];  /* mounted filesystem */
> > >  91         char      f_mntonname[MNAMELEN];    /* directory on which mounted */
> > > 
> > > Where MNAMELEN is, currently, 88.  These limit the length of the path
> > > that a filesystem can be mounted to.  This is enforced by
> > > kern/vfs_mount.c:vfs_donmount().  This limit seems archaic, especially
> > > given use cases like virtualization (large filesystem structures to
> > > support underlying VMs), builds (which often make extensive use of
> > > chroot with nullfs/NFS), ZFS, snapshots, etc.  Does anyone object to
> > > bumping MNAMELEN to 1024 (PATH_MAX/MAXPATHLEN)?  Or some other
> > > reasonably large value?
> > 
> > There is nothing inherently wrong with bumping the length. But the
> > work required is probably more then you estimated. The cause is the
> > ABI breakage. For sure, you will need to provide the shims for compat
> > syscalls. Unfortunately, this is not enough.
> > 
> > Even quick look over our tree shows that struct statfs is used in API by
> > several base libraries. Look e.g. at the getmntinfo(3). Libc would need
> > shims too, at least.
> > 
> > You will need to do the ABI analisys of the whole system, provide the
> > shims for the symbol-versioned libraries, and bump so version for
> > unversioned.
> 
> I could do it as part of my 64-bit ino_t GSoC project. So should I
> change MNAMELEN to 1024? What about MFSNAMELEN, it's now 16.
> I think removing or increasing size of f_charspare field might also be a
> good idea.

I think it may be good to have this change done. I do think that this is 
of much lower importance then 64bit ino_t. Also, if done, I prefer to have
struct statfs change be separate from ino_t change. That said, statfs
change should be much smaller then ino_t, thus easier to review, allowing
it to be committed before ino_t.

Increasing f_fstypenamehave no obvious benefits, do we expect to have filesystem
types with the names longer then 16 bytes ?

I do not think that having much spare fields in struct statfs is definitely
good. The structure will become quite large as is after MNAMELEN bump, and
the balance should be provided between application memory usage and future
extensibility.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 196 bytes
Desc: not available
Url : http://lists.freebsd.org/pipermail/freebsd-fs/attachments/20110622/9caf0d4e/attachment.pgp