Re: revert "fs/befs/linuxvfs.c: replace strncpy by strlcpy"

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

On Tue, Apr 28, 2015 at 02:48:45PM -0700, Linus Torvalds wrote:

> I suspect you could take that lib/strncpy_from_user.c and massage it
> reasonably trivially to be a good function.
> 
> That said, I can't think of a single strncpy (or strlcpy) in kernel
> space that is actually worth even optimizing for. They just don't tend
> to be in the critical path. So correctness is likely *much* more
> important than worrying about performance.

Indeed.  As it is, I suspect that strlcpy() use should be uniformly
discouraged; if nothing else, snprintf() gives the same semantics,
is less likely to cause confusion regardling the expected return
value and none of those paths are performance-critical.

strncpy() has another use, though, and it can't be replaced by strlcpy() -
see the commits that had started this thread.  IMO they (and anything
else of the same nature) really need to be reverted; using strlcpy() on
something that isn't guaranteed to be NUL-terminated is a serious bug.

And catching all such places is going to be quite a work - there are too
many strlcpy() callers out there.

Frankly, looking through call sites in fs...
* two callers in fs/9p - strlcpy() + sscanf(), both should've been
plain sscanf() (and the second should've been "HARDLINKCOUNT%u" instead
of "%13s %u" + comparison of string with "HARDLINKCOUNT" - sscanf() is
quite capable of matching explicit string literals)
* affs one: match_strdup + strlcpy + kfree.  Should just use match_strlcpy
instead (BTW, despite the name, it does *not* use strclpy() internally).
* afs: might be correct.
* two in befs: both broken.
* binfmt_misc: fishy; load_misc_binary() finds an object under rwlock, copies
one of its fields (->interpreter), drops rwlock and proceeds to do various
blocking operations (including open_exec()).  Using another field of the
same object (->interp_flags) all along.  If it gets freed and reused, well...
let's hope we won't fuck up too badly.  IMO we'd be better off if we added
a refcount to that sucker and used it to control the freeing.
* btrfs: undefined behaviour - potentially overlapping source and destination.
* another btrfs one:
                char b[BDEVNAME_SIZE];
                strlcpy(s->s_id, bdevname(bdev, b), sizeof(s->s_id));
complete garbage; might as well do bdevname(bdev, s->s_id) and be done with
that.

... and so on; this stuff is misused more often than not.