[PATCH net] net: af_key: fix use-after-free in dump_sp() on policy migrate

[Daniel Wade <danjwade95@gmail.com>]

pfkey_xfrm_policy2msg_size() derives an exported policy's message length
from xfrm_vec[].encap_family. dump_sp() derives it twice: once through
pfkey_xfrm_policy2msg_prep(), which allocates the skb, and again in
pfkey_xfrm_policy2msg(), which fills it in.

dump_sp() holds net->xfrm.xfrm_policy_lock rather than pol->lock, so
xfrm_policy_migrate() can rewrite encap_family between the two
derivations. The second length is then up to 24 bytes per template
larger than the first, and pfkey_xfrm_policy2msg() opens with
memset(hdr, 0, size), clearing the difference into the skb's shared
info. Four BEET templates put 80 bytes there, which reaches dataref:
alloc_skb() had set it to 1, the memset makes it 0, and the skb_clone()
in pfkey_broadcast_one() leaves two skbs sharing one head with a dataref
of 1. Freeing the original releases a head the queued clone still points
at, so recvmsg() copies freed memory to user space and consume_skb()
frees it a second time. The CAP_NET_ADMIN that pfkey_create() and the
xfrm netlink receive path check is namespace-relative, so an
unprivileged user can reach this. SLUB's double-free check fires within
a second of starting the reproducer; the oops kills a task, or panics
the machine when the corruption lands in softirq context.

Racing SADB_X_SPDDUMP against XFRM_MSG_MIGRATE as an ordinary user gave
4585 use-after-free reports and 765 double-frees in 300s on mainline
with KASAN and 8 CPUs. The reads reach user space (addresses trimmed):

  BUG: KASAN: slab-use-after-free in _copy_to_iter+0x1d8/0x1430
  Read of size 240 at addr ffff88810f98d140 by task pfkey_dataref_o/433
  CPU: 3 UID: 65534 PID: 433 Comm: pfkey_dataref_o Not tainted 7.4.0-rc2
  Call Trace:
   <TASK>
   dump_stack_lvl+0x66/0xa0
   print_report+0xd0/0x600
   kasan_report+0xc6/0x100
   kasan_check_range+0x105/0x1b0
   _copy_to_iter+0x1d8/0x1430
   __skb_datagram_iter+0xf8/0x820
   skb_copy_datagram_iter+0x6c/0x1a0
   pfkey_recvmsg+0x1ff/0x500
   sock_recvmsg+0x15e/0x1a0
   __sys_recvfrom+0x182/0x260
   do_syscall_64+0x10b/0x5f0
   entry_SYSCALL_64_after_hwframe+0x77/0x7f
  Allocated by task 433:
   kmalloc_reserve+0x108/0x2b0
   __alloc_skb+0x123/0x5b0
   dump_sp+0xdf/0x490
   xfrm_policy_walk+0x1d9/0x6c0
   pfkey_do_dump+0x61/0x420
   pfkey_recvmsg+0x358/0x500

Take xp->lock for reading across both calls, so both derivations see the
same encap_family. xfrm_policy_walk() already calls dump_sp() under
spin_lock_bh(), so the plain reader suffices, as in xfrm_policy_timer().
The other callers of this pair are already serialized against migrate by
net->xfrm.xfrm_cfg_mutex.

Fixes: 80c9abaabf42 ("[XFRM]: Extension for dynamic update of endpoint address(es)")
Cc: stable@vger.kernel.org
Signed-off-by: Daniel Wade <danjwade95@gmail.com>
---

 net/key/af_key.c |    9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/net/key/af_key.c b/net/key/af_key.c
index d215b33..ab4aaca 100644
--- a/net/key/af_key.c
+++ b/net/key/af_key.c
@@ -2712,11 +2712,18 @@ static int dump_sp(struct xfrm_policy *xp, int dir, int count, void *ptr)
 	if (!pfkey_can_dump(&pfk->sk))
 		return -ENOBUFS;
 
+	/* Both halves derive the size from xfrm_vec[].encap_family, so
+	 * hold xp->lock against a concurrent xfrm_policy_migrate().
+	 */
+	read_lock(&xp->lock);
 	out_skb = pfkey_xfrm_policy2msg_prep(xp);
-	if (IS_ERR(out_skb))
+	if (IS_ERR(out_skb)) {
+		read_unlock(&xp->lock);
 		return PTR_ERR(out_skb);
+	}
 
 	err = pfkey_xfrm_policy2msg(out_skb, xp, dir);
+	read_unlock(&xp->lock);
 	if (err < 0) {
 		kfree_skb(out_skb);
 		return err;