On Thu, Nov 30, 2017 at 05:18:33AM -0800, Christoph Hellwig wrote:

On Thu, Nov 30, 2017 at 02:07:19AM +0000, Al Viro wrote:

quoted

Incidentally, grepping for sys_close() shows another piece of fun in
net/netfilter/xt_bpf.c.  Folks, ONCE DESCRIPTOR IS INSTALLED, THAT'S
IT; THERE'S NO REMOVING IT ON FAILURE EXITS.  sys_close() should
never, ever be used that way.  Sigh...

Would be great do unexport the thing.  Except that we also have
binfmt_misc (which looks legit) and autofs4, which on crack decided
that close() isn't a fun syscall, they'd much rather have an ioctl
that does exactly the same..

Yes, since binfmt_misc one is guaranteed that its descriptor table is
not shared - all callchains go through do_execveat_common(), where we'd
use unshare_files().  autofs one is... not in good taste, but still
safe; there the descriptor is preexisting and it's essentially a weird
way of spelling close(2).  References from syscall tables are, of course,
OK.  init/*.c uses are done pretty much from userland - they could have
been straight syscalls, if not for the lack of klibc in kernel tree.
Everything else, though...

IMO we need a whack-a-mole list somewhere; "new callers of sys_close()
anywhere outside of init/* and syscall tables" definitely should be
on it...