[PATCH] RDMA/rxe: destroy the mcg when rxe_mcast_add() fails in rxe_get_mcg()

[Michael Bommarito]

rxe_get_mcg() inserts the new mcg into rxe->mcg_tree and takes the tree
reference before calling rxe_mcast_add() outside mcg_lock. On failure
the error path frees the mcg with a bare kfree() without erasing the
tree node or dropping the tree reference, so the freed mcg stays linked
in mcg_tree and the next __rxe_lookup_mcg() on the same mgid uses it
after free. rxe_mcast_add() fails reachably from an unprivileged caller:
-ENODEV when the backing netdev is removed, or a propagated dev_mc_add()
error.

Tear the mcg down with __rxe_destroy_mcg() on the failure path, as
rxe_attach_mcast() already does.

Reproduced under KASAN on QEMU by forcing the rxe_mcast_add() failure;
the use-after-free in __rxe_lookup_mcg() is gone after this change.

Fixes: a926a903b7dc ("RDMA/rxe: Do not call  dev_mc_add/del() under a spinlock")
Cc: stable@xxxxxxxxxxxxxxx # v5.18+
Assisted-by: Claude:claude-opus-4-8
Signed-off-by: Michael Bommarito <michael.bommarito@xxxxxxxxx>
---
Reproduction (v7.1-rc4, x86_64 QEMU/KVM, KASAN, Soft-RoCE):

Forcing rxe_mcast_add() to return -ENODEV, an unprivileged ATTACH_MCAST
on a UD QP leaves the freed mcg linked in mcg_tree. On the stock kernel
the next lookup reports

  BUG: KASAN: slab-use-after-free in __rxe_lookup_mcg

and the subsequent rb_erase() panics. Patched, the forced failure
returns cleanly. Control: with injection disabled, re-attach and detach
of the same MGID and a two-QP join/leave are KASAN-clean on both trees.

tools/testing/selftests/rdma has no rxe_mcast coverage; harness off-list
on request.

 drivers/infiniband/sw/rxe/rxe_mcast.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

diff --git a/drivers/infiniband/sw/rxe/rxe_mcast.c b/drivers/infiniband/sw/rxe/rxe_mcast.c
index 5cad720..7f148d4 100644
--- a/drivers/infiniband/sw/rxe/rxe_mcast.c
+++ b/drivers/infiniband/sw/rxe/rxe_mcast.c
@@ -196,6 +196,8 @@ static void __rxe_init_mcg(struct rxe_dev *rxe, union ib_gid *mgid,
 	__rxe_insert_mcg(mcg);
 }
 
+static void __rxe_destroy_mcg(struct rxe_mcg *mcg);
+
 /**
  * rxe_get_mcg - lookup or allocate a mcg
  * @rxe: rxe device object
@@ -247,7 +249,13 @@ static struct rxe_mcg *rxe_get_mcg(struct rxe_dev *rxe, union ib_gid *mgid)
 	if (!err)
 		return mcg;
 
-	kfree(mcg);
+	/* mcg was made visible in mcg_tree; unwind the insert before freeing. */
+	spin_lock_bh(&rxe->mcg_lock);
+	__rxe_destroy_mcg(mcg);
+	spin_unlock_bh(&rxe->mcg_lock);
+	kref_put(&mcg->ref_cnt, rxe_cleanup_mcg);
+	return ERR_PTR(err);
+
 err_dec:
 	atomic_dec(&rxe->mcg_num);
 	return ERR_PTR(err);
base-commit: 5200f5f493f79f14bbdc349e402a40dfb32f23c8
-- 
2.53.0