Re: [RFC] killing boilerplate checks in ->link/->mkdir/->rename

[Al Viro <viro@ZenIV.linux.org.uk>]

On Thu, Feb 02, 2012 at 03:46:06PM -0800, Linus Torvalds wrote:
> On Thu, Feb 2, 2012 at 1:24 PM, Al Viro <viro@zeniv.linux.org.uk> wrote:
> >
> > Comments? ?Boilerplate removal follows (22 files changed, 45 insertions(+),
> > 120 deletions(-)), but it's *not* for immediate merge; it's really completely
> > untested.
> 
> Looks ok to me. Historically, the more things we can check at the VFS
> layer, the better.

After looking a bit more: nlink_t is a f*cking mess.  Almost any code
using that type kernel-side is broken.  Crap galore:
	* sometimes it's 32 bits, sometimes 16, sometimes 64.  Essentially
at random.
	* almost all have it unsigned, except for sparc32, where it's
signed short [inherited from v7 via SunOS?  BTW, in v6 it used to be even
funnier - char, which is where ridiculous LINK_MAX == 127 comes from]

IOW, nlink_t is an attractive nuisance - it's nearly impossible to use in
a portable way and we are lucky that almost nobody tries to.  Exceptions:
ocfs2_rename() does
        nlink_t old_dir_nlink = old_dir->i_nlink;
	...
followed later by comparison with old_dir->i_nlink.  And no, it's not to
handle truncation - it's "what if i_nlink changed while ocfs2_rename()
had been grabbing the cluster lock" kind of thing.  OCFS2 can have up
to 2^32 links to file, so truncation is really possible...  AFAICS,
that one is a genuine bug - this nlink_t should be u32...
Another one is proc_dir_entry ->nlink and it would cause Bad Things(tm)
on architecture with 16bit nlink_t if we could end up with 65534
subdirectories in some procfs dir.  Might be possible, might be not -
doing that under /proc/sys is definitely possible, but that won't be
enough; needs to be proc_dir_entry-backed directory.  Again, solution
is to use explicit u32 anyway.

	* compat_nlink_t is even funnier - it's signed in *two* cases; sparc
and ppc.  No, nlink_t on ppc32 is unsigned.  Not that anyone cared, really,
since the _only_ use of that type is in struct compat_stat.  For exactly
one field.  Only used as left-hand side of assignment, which is actually
broken since unlike cp_new_stat(), cp_compat_stat() does *not* check if the
value fits into st_nlink.  Bug, needs to be fixed.  Incidentally, just what
should we do on sparc32 if we run into a file with 4G-10 links?  -EOVERFLOW
or silently put 65536-10 in st_nlink and be done with that?  Note that
filesystems allowing that many links *do* exist...

	* when does jfs dtInsert() return -EMLINK?  Can it ever get triggered?
	* WTF is XFS doing with these checks?  Note that we have them
done _twice_ on all paths - explictly from xfs_create(), xfs_link(),
xfs_rename() and then from xfs_bumplink() called by exactly the same
set of functions.

	* what's up with btrfs_insert_inode_ref()?  I've tried to trace
the codepaths around there, but... Incidentally, when could fixup_low_keys()
return non-zero?  I don't see any candidates for that in there...  Chris?

	* ubifs, hfsplus, jffs2 - definitely broken if you create enough
links.  i_nlink wraparound to zero, confused inode eviction logics.