Re: [PATCH v1 net] af_packet: Fix fortified memcpy() without flex array.

[Kees Cook <keescook@chromium.org>]

On Mon, Oct 09, 2023 at 10:12:28AM -0700, Kuniyuki Iwashima wrote:
> From: Kees Cook <keescook@chromium.org>
> Date: Mon, 9 Oct 2023 09:01:34 -0700
> > On Mon, Oct 09, 2023 at 08:31:52AM -0700, Kuniyuki Iwashima wrote:
> > > Sergei Trofimovich reported a regression [0] caused by commit a0ade8404c3b
> > > ("af_packet: Fix warning of fortified memcpy() in packet_getname().").
> > > 
> > > It introduced a flex array sll_addr_flex in struct sockaddr_ll as a
> > > union-ed member with sll_addr to work around the fortified memcpy() check.
> > > 
> > > However, a userspace program uses a struct that has struct sockaddr_ll in
> > > the middle, where a flex array is illegal to exist.
> > > 
> > >   include/linux/if_packet.h:24:17: error: flexible array member 'sockaddr_ll::<unnamed union>::<unnamed struct>::sll_addr_flex' not at end of 'struct packet_info_t'
> > >      24 |                 __DECLARE_FLEX_ARRAY(unsigned char, sll_addr_flex);
> > >         |                 ^~~~~~~~~~~~~~~~~~~~
> > > To fix the regression, let's go back to the first attempt [1] telling
> > > memcpy() the actual size of the array.
> > > 
> > > Reported-by: Sergei Trofimovich <slyich@gmail.com>
> > > Closes: https://github.com/NixOS/nixpkgs/pull/252587#issuecomment-1741733002 [0]
> > 
> > Eww. That's a buggy definition -- it could get overflowed.
> 
> Only if they pass sizeof(struct sockaddr_storage) to getsockname().
> 
> 
> > 
> > But okay, we don't break userspace.
> > 
> > > Link: https://lore.kernel.org/netdev/20230720004410.87588-3-kuniyu@amazon.com/ [1]
> > > Fixes: a0ade8404c3b ("af_packet: Fix warning of fortified memcpy() in packet_getname().")
> > > Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
> > > ---
> > >  include/uapi/linux/if_packet.h | 6 +-----
> > >  net/packet/af_packet.c         | 7 ++++++-
> > >  2 files changed, 7 insertions(+), 6 deletions(-)
> > > 
> > > diff --git a/include/uapi/linux/if_packet.h b/include/uapi/linux/if_packet.h
> > > index 4d0ad22f83b5..9efc42382fdb 100644
> > > --- a/include/uapi/linux/if_packet.h
> > > +++ b/include/uapi/linux/if_packet.h
> > > @@ -18,11 +18,7 @@ struct sockaddr_ll {
> > >  	unsigned short	sll_hatype;
> > >  	unsigned char	sll_pkttype;
> > >  	unsigned char	sll_halen;
> > > -	union {
> > > -		unsigned char	sll_addr[8];
> > > -		/* Actual length is in sll_halen. */
> > > -		__DECLARE_FLEX_ARRAY(unsigned char, sll_addr_flex);
> > > -	};
> > > +	unsigned char	sll_addr[8];
> > >  };
> > 
> > Yup, we need to do at least this.
> > 
> > >  
> > >  /* Packet types */
> > > diff --git a/net/packet/af_packet.c b/net/packet/af_packet.c
> > > index 8f97648d652f..a84e00b5904b 100644
> > > --- a/net/packet/af_packet.c
> > > +++ b/net/packet/af_packet.c
> > > @@ -3607,7 +3607,12 @@ static int packet_getname(struct socket *sock, struct sockaddr *uaddr,
> > >  	if (dev) {
> > >  		sll->sll_hatype = dev->type;
> > >  		sll->sll_halen = dev->addr_len;
> > > -		memcpy(sll->sll_addr_flex, dev->dev_addr, dev->addr_len);
> > > +
> > > +		/* Let __fortify_memcpy_chk() know the actual buffer size. */
> > > +		memcpy(((struct sockaddr_storage *)sll)->__data +
> > > +		       offsetof(struct sockaddr_ll, sll_addr) -
> > > +		       offsetofend(struct sockaddr_ll, sll_family),
> > > +		       dev->dev_addr, dev->addr_len);
> > >  	} else {
> > >  		sll->sll_hatype = 0;	/* Bad: we have no ARPHRD_UNSPEC */
> > >  		sll->sll_halen = 0;
> > 
> > I still think this is a mistake. We're papering over so many lies to the
> > compiler. :P If "uaddr" is actually "struct sockaddr_storage", then we
> > should update the callers...
> 
> We could update all callers to pass sockaddr_storage but it seems too much
> for net.git.. :/  I think the conversion should be done later for net-next.
> 
>   $ grep -rn -E "\.getname.*?=" | cut -f 2 -d"=" | sort | uniq | wc -l
>   40

Agreed -- it's something to do going forward. Your fix is likely the
right approach to take to undo the UAPI change.

-Kees

-- 
Kees Cook