On Fri, Nov 28, 2025 at 08:53:21PM -0400, Jason Gunthorpe wrote:

If the CM ID is destroyed while the CM event for multicast creating is still queued the cancel_work_sync() will prevent the work from running which also prevents destroying the ah_attr. This leaks a refcount and triggers a WARN:

GID entry ref leak for dev syz1 index 2 ref=573 WARNING: CPU: 1 PID: 655 at drivers/infiniband/core/cache.c:809 release_gid_table drivers/infiniband/core/cache.c:806 [inline] WARNING: CPU: 1 PID: 655 at drivers/infiniband/core/cache.c:809 gid_table_release_one+0x284/0x3cc drivers/infiniband/core/cache.c:886

Destroy the ah_attr after canceling the work, it is safe to call this twice.

Cc: stable@vger.kernel.org Fixes: fe454dc31e84 ("RDMA/ucma: Fix use-after-free bug in ucma_create_uevent") Reported-by: syzbot+b0da83a6c0e2e2bddbd4@syzkaller.appspotmail.com Closes: https://lore.kernel.org/all/68232e7b.050a0220.f2294.09f6.GAE@google.com Signed-off-by: Jason Gunthorpe jgg@nvidia.com

---

drivers/infiniband/core/cma.c | 2 ++ 1 file changed, 2 insertions(+)

Bah, a hunk got lost, it should be like this:

diff --git a/drivers/infiniband/core/cma.c b/drivers/infiniband/core/cma.c index 95e89f5c147c2c..f00f1d3fbd9c53 100644 --- a/drivers/infiniband/core/cma.c +++ b/drivers/infiniband/core/cma.c @@ -2009,6 +2009,7 @@ static void destroy_mc(struct rdma_id_private *id_priv, ib_sa_free_multicast(mc->sa_mc);

if (rdma_protocol_roce(id_priv->id.device, id_priv->id.port_num)) { + struct rdma_cm_event *event = &mc->iboe_join.event; struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; struct net_device *ndev = NULL; @@ -2031,6 +2032,8 @@ static void destroy_mc(struct rdma_id_private *id_priv, dev_put(ndev);

cancel_work_sync(&mc->iboe_join.work); + if (event->event == RDMA_CM_EVENT_MULTICAST_JOIN) + rdma_destroy_ah_attr(&event->param.ud.ah_attr); } kfree(mc); }