RDMA/cma: Ensure rdma_addr_cancel() happens before issuing more requests

The FSM can run in a circle allowing rdma_resolve_ip() to be called twice
on the same id_priv. While this cannot happen without going through the
work, it violates the invariant that the same address resolution
background request cannot be active twice.

       CPU 1                                  CPU 2

rdma_resolve_addr():
  RDMA_CM_IDLE -> RDMA_CM_ADDR_QUERY
  rdma_resolve_ip(addr_handler)  #1

			 process_one_req(): for #1
                          addr_handler():
                            RDMA_CM_ADDR_QUERY -> RDMA_CM_ADDR_BOUND
                            mutex_unlock(&id_priv->handler_mutex);
                            [.. handler still running ..]

rdma_resolve_addr():
  RDMA_CM_ADDR_BOUND -> RDMA_CM_ADDR_QUERY
  rdma_resolve_ip(addr_handler)
    !! two requests are now on the req_list

rdma_destroy_id():
 destroy_id_handler_unlock():
  _destroy_id():
   cma_cancel_operation():
    rdma_addr_cancel()

                          // process_one_req() self removes it
		          spin_lock_bh(&lock);
                           cancel_delayed_work(&req->work);
	                   if (!list_empty(&req->list)) == true

      ! rdma_addr_cancel() returns after process_on_req #1 is done

   kfree(id_priv)

			 process_one_req(): for #2
                          addr_handler():
	                    mutex_lock(&id_priv->handler_mutex);
                            !! Use after free on id_priv

rdma_addr_cancel() expects there to be one req on the list and only
cancels the first one. The self-removal behavior of the work only happens
after the handler has returned. This yields a situations where the
req_list can have two reqs for the same "handle" but rdma_addr_cancel()
only cancels the first one.

The second req remains active beyond rdma_destroy_id() and will
use-after-free id_priv once it inevitably triggers.

Fix this by remembering if the id_priv has called rdma_resolve_ip() and
always cancel before calling it again. This ensures the req_list never
gets more than one item in it and doesn't cost anything in the normal flow
that never uses this strange error path.

Link: https://lore.kernel.org/r/0-v1-3bc675b8006d+22-syz_cancel_uaf_jgg@nvidia.com
Cc: stable@vger.kernel.org
Fixes: e51060f08a61 ("IB: IP address based RDMA connection manager")
Reported-by: syzbot+dc3dfba010d7671e05f5@syzkaller.appspotmail.com
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>

diff --git a/drivers/infiniband/core/cma.c b/drivers/infiniband/core/cma.c
index 8862b0e572f0fe6ecf329e6e03c66c46a868eba1..704ce595542c5735e2390ebedc2cf2697333b8e5 100644 (file)
--- a/drivers/infiniband/core/cma.c
+++ b/drivers/infiniband/core/cma.c
@@ -1783,6 +1783,14 @@ static void cma_cancel_operation(struct rdma_id_private *id_priv,
 {
        switch (state) {
        case RDMA_CM_ADDR_QUERY:
+               /*
+                * We can avoid doing the rdma_addr_cancel() based on state,
+                * only RDMA_CM_ADDR_QUERY has a work that could still execute.
+                * Notice that the addr_handler work could still be exiting
+                * outside this state, however due to the interaction with the
+                * handler_mutex the work is guaranteed not to touch id_priv
+                * during exit.
+                */
                rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
                break;
        case RDMA_CM_ROUTE_QUERY:
@@ -3425,6 +3433,21 @@ int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
                if (dst_addr->sa_family == AF_IB) {
                        ret = cma_resolve_ib_addr(id_priv);
                } else {
+                       /*
+                        * The FSM can return back to RDMA_CM_ADDR_BOUND after
+                        * rdma_resolve_ip() is called, eg through the error
+                        * path in addr_handler(). If this happens the existing
+                        * request must be canceled before issuing a new one.
+                        * Since canceling a request is a bit slow and this
+                        * oddball path is rare, keep track once a request has
+                        * been issued. The track turns out to be a permanent
+                        * state since this is the only cancel as it is
+                        * immediately before rdma_resolve_ip().
+                        */
+                       if (id_priv->used_resolve_ip)
+                               rdma_addr_cancel(&id->route.addr.dev_addr);
+                       else
+                               id_priv->used_resolve_ip = 1;
                        ret = rdma_resolve_ip(cma_src_addr(id_priv), dst_addr,
                                              &id->route.addr.dev_addr,
                                              timeout_ms, addr_handler,

diff --git a/drivers/infiniband/core/cma_priv.h b/drivers/infiniband/core/cma_priv.h
index 5c463da9984536c6795ee7905fb08848235b4cb9..f92f101ea9818f8c1ecf11375353103dce234858 100644 (file)
--- a/drivers/infiniband/core/cma_priv.h
+++ b/drivers/infiniband/core/cma_priv.h
@@ -91,6 +91,7 @@ struct rdma_id_private {
        u8                      afonly;
        u8                      timeout;
        u8                      min_rnr_timer;
+       u8 used_resolve_ip;
        enum ib_gid_type        gid_type;
 
        /*