Re: [PATCH v2] fs: Set file_handle::handle_bytes before referencing file_handle::f_handle

[Kees Cook]

On Fri, Apr 05, 2024 at 12:21:57PM +0200, Jan Kara wrote:
> On Thu 04-04-24 14:12:15, Kees Cook wrote:
> > Since __counted_by(handle_bytes) was added to struct file_handle, we need
> > to explicitly set it in the one place it wasn't yet happening prior to
> > accessing the flex array "f_handle". For robustness also check for a
> > negative value for handle_bytes, which is possible for an "int", but
> > nothing appears to set.
> > 
> > Fixes: 1b43c4629756 ("fs: Annotate struct file_handle with __counted_by() and use struct_size()")
> > Signed-off-by: Kees Cook <keescook@xxxxxxxxxxxx>
> > ---
> > Cc: Jan Kara <jack@xxxxxxx>
> > Cc: Chuck Lever <chuck.lever@xxxxxxxxxx>
> > Cc: Gustavo A. R. Silva <gustavoars@xxxxxxxxxx>
> > Cc: Christian Brauner <brauner@xxxxxxxxxx>
> > Cc: Alexander Viro <viro@xxxxxxxxxxxxxxxxxx>
> > Cc: Jeff Layton <jlayton@xxxxxxxxxx>
> > Cc: Amir Goldstein <amir73il@xxxxxxxxx>
> > Cc: linux-fsdevel@xxxxxxxxxxxxxxx
> > Cc: linux-nfs@xxxxxxxxxxxxxxx
> > Cc: linux-hardening@xxxxxxxxxxxxxxx
> >  v2: more bounds checking, add comments, dropped reviews since logic changed
> >  v1: https://lore.kernel.org/all/20240403215358.work.365-kees@xxxxxxxxxx/
> > ---
> >  fs/fhandle.c | 11 +++++++++--
> >  1 file changed, 9 insertions(+), 2 deletions(-)
> > 
> > diff --git a/fs/fhandle.c b/fs/fhandle.c
> > index 8a7f86c2139a..854f866eaad2 100644
> > --- a/fs/fhandle.c
> > +++ b/fs/fhandle.c
> > @@ -40,6 +40,11 @@ static long do_sys_name_to_handle(const struct path *path,
> >  			 GFP_KERNEL);
> >  	if (!handle)
> >  		return -ENOMEM;
> > +	/*
> > +	 * Since handle->f_handle is about to be written, make sure the
> > +	 * associated __counted_by(handle_bytes) variable is correct.
> > +	 */
> > +	handle->handle_bytes = f_handle.handle_bytes;
> >  
> >  	/* convert handle size to multiple of sizeof(u32) */
> >  	handle_dwords = f_handle.handle_bytes >> 2;
> > @@ -51,8 +56,8 @@ static long do_sys_name_to_handle(const struct path *path,
> >  	handle->handle_type = retval;
> >  	/* convert handle size to bytes */
> >  	handle_bytes = handle_dwords * sizeof(u32);
> > -	handle->handle_bytes = handle_bytes;
> > -	if ((handle->handle_bytes > f_handle.handle_bytes) ||
> > +	/* check if handle_bytes would have exceeded the allocation */
> > +	if ((handle_bytes < 0) || (handle_bytes > f_handle.handle_bytes) ||
> 
> This is broken. Let me explain: Userspace passes in struct file_handle
> (ufh) and says how many bytes it has reserved for the variable length
> contents in ufh->handle_bytes. We call exportfs_encode_fh() to create
> the file handle. If it fits into the provided space, the function returns
> in handle_dwords how many uints it has actually stored. If the handle
> didn't fit, handle_dword contains number of uints we'd need in the variable
> length part to be able to fit the handle in.
> 
> Now your patch destroys this behavior by storing 0 to handle_bytes in case
> the handle didn't fit *before* the returned value is actually stored to a
> struct copied to userspace.

Ah yes, I see now how they're used for separate things. I will respin
again and possibly fix up the "int" to "u32" as suggested by willy.

-- 
Kees Cook