Re: [PATCH v2 0/3] ceph: don't NULL terminate virtual xattr values

From: Jeff Layton
Date: Thu Jun 20 2019 - 07:46:25 EST


On Thu, 2019-06-20 at 13:24 +0300, Andy Shevchenko wrote:
> On Wed, Jun 19, 2019 at 12:45:25PM -0400, Jeff Layton wrote:
> > v2: drop bogus EXPORT_SYMBOL of static function
> >
> > The only real difference between this set and the one I sent originally
> > is the removal of a spurious EXPORT_SYMBOL in the snprintf patch.
> >
> > I'm mostly sending this with a wider cc list in an effort to get a
> > review from the maintainers of the printf code. Basically ceph needs a
> > snprintf variant that does not NULL terminate in order to handle its
> > virtual xattrs.
> >
> > Joe Perches had expressed some concerns about stack usage in vsnprintf
> > with this, but I'm not sure I really understand the basis of that
> > concern. If it is problematic, then I could use suggestions as to how
> > best to fix that up.
>
> It might be problematic, since vsnprintf() can be called recursively.
>

So the concern is that we'd have extra call/ret activity in the stack?
That seems like a lot of hand-wringing over very little, but ok if so.

> > ----------------------------8<-----------------------------
> >
> > kcephfs has several "virtual" xattrs that return strings that are
> > currently populated using snprintf(), which always NULL terminates the
> > string.
> >
> > This leads to the string being truncated when we use a buffer length
> > acquired by calling getxattr with a 0 size first. The last character
> > of the string ends up being clobbered by the termination.
>
> So, then don't use snprintf() for this, simple memcpy() designed for that kind
> of things.
>

memcpy from what? For many of these xattrs, we need to format integer
data into strings. I could roll my own routine to do this formatting,
but that's sort of what sprintf and its variants are for and I'd rather
not reimplement all of it from scratch.

> > The convention with xattrs is to not store the termination with string
> > data, given that we have the length. This is how setfattr/getfattr
> > operate.
>
> Fine.
>
> > This patch makes ceph's virtual xattrs not include NULL termination
> > when formatting their values. In order to handle this, a new
> > snprintf_noterm function is added, and ceph is changed over to use
> > this to populate the xattr value buffer.
>
> In terms of vsnprintf(), and actually compiler point of view, it's not a string
> anymore, it's a text-based data.
>
> Personally, I don't see an advantage of a deep intrusion into vsnprintf().
> The wrapper can be made to achieve this w/o touching the generic code. Thus,
> you can quickly and cleanly fix the issue, while discussing this with wider
> audience.
>

Sorry, if I'm being dense but I'm not sure I follow here.

Are you suggesting I should just copy/paste most of vsnprintf into a new
function that just leaves off the termination at the end, and leave the
original alone? That seems like a bit of a waste, but if that's the
consensus then ok.

> > Finally, we fix ceph to
> > return -ERANGE properly when the string didn't fit in the buffer.
> >
> > Jeff Layton (3):
> > lib/vsprintf: add snprintf_noterm
> > ceph: don't NULL terminate virtual xattr strings
> > ceph: return -ERANGE if virtual xattr value didn't fit in buffer
> >
> > fs/ceph/xattr.c | 49 +++++++-------
> > include/linux/kernel.h | 2 +
> > lib/vsprintf.c | 144 ++++++++++++++++++++++++++++-------------
> > 3 files changed, 129 insertions(+), 66 deletions(-)
> >
> > --
> > 2.21.0
> >

--
Jeff Layton <jlayton@xxxxxxxxxx>