@@ -2456,8 +2456,10 @@ static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
if (IS_ERR(id))
return PTR_ERR(id);
+ mutex_lock(&id_priv->qp_mutex);
id->tos = id_priv->tos;
id->tos_set = id_priv->tos_set;
+ mutex_unlock(&id_priv->qp_mutex);
id->afonly = id_priv->afonly;
id_priv->cm_id.iw = id;
@@ -2518,8 +2520,10 @@ static int cma_listen_on_dev(struct rdma_id_private *id_priv,
cma_id_get(id_priv);
dev_id_priv->internal_id = 1;
dev_id_priv->afonly = id_priv->afonly;
+ mutex_lock(&id_priv->qp_mutex);
dev_id_priv->tos_set = id_priv->tos_set;
dev_id_priv->tos = id_priv->tos;
+ mutex_unlock(&id_priv->qp_mutex);
ret = rdma_listen(&dev_id_priv->id, id_priv->backlog);
if (ret)
@@ -2566,8 +2570,10 @@ void rdma_set_service_type(struct rdma_cm_id *id, int tos)
struct rdma_id_private *id_priv;
id_priv = container_of(id, struct rdma_id_private, id);
+ mutex_lock(&id_priv->qp_mutex);
id_priv->tos = (u8) tos;
id_priv->tos_set = true;
+ mutex_unlock(&id_priv->qp_mutex);
}
EXPORT_SYMBOL(rdma_set_service_type);
@@ -2594,8 +2600,10 @@ int rdma_set_ack_timeout(struct rdma_cm_id *id, u8 timeout)
return -EINVAL;
id_priv = container_of(id, struct rdma_id_private, id);
+ mutex_lock(&id_priv->qp_mutex);
id_priv->timeout = timeout;
id_priv->timeout_set = true;
+ mutex_unlock(&id_priv->qp_mutex);
return 0;
}
@@ -2631,8 +2639,10 @@ int rdma_set_min_rnr_timer(struct rdma_cm_id *id, u8 min_rnr_timer)
return -EINVAL;
id_priv = container_of(id, struct rdma_id_private, id);
+ mutex_lock(&id_priv->qp_mutex);
id_priv->min_rnr_timer = min_rnr_timer;
id_priv->min_rnr_timer_set = true;
+ mutex_unlock(&id_priv->qp_mutex);
return 0;
}
@@ -3018,8 +3028,11 @@ static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
u8 default_roce_tos = id_priv->cma_dev->default_roce_tos[id_priv->id.port_num -
rdma_start_port(id_priv->cma_dev->device)];
- u8 tos = id_priv->tos_set ? id_priv->tos : default_roce_tos;
+ u8 tos;
+ mutex_lock(&id_priv->qp_mutex);
+ tos = id_priv->tos_set ? id_priv->tos : default_roce_tos;
+ mutex_unlock(&id_priv->qp_mutex);
work = kzalloc(sizeof *work, GFP_KERNEL);
if (!work)
@@ -3066,8 +3079,10 @@ static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
* PacketLifeTime = local ACK timeout/2
* as a reasonable approximation for RoCE networks.
*/
+ mutex_lock(&id_priv->qp_mutex);
route->path_rec->packet_life_time = id_priv->timeout_set ?
id_priv->timeout - 1 : CMA_IBOE_PACKET_LIFETIME;
+ mutex_unlock(&id_priv->qp_mutex);
if (!route->path_rec->mtu) {
ret = -EINVAL;
@@ -4091,8 +4106,11 @@ static int cma_connect_iw(struct rdma_id_private *id_priv,
if (IS_ERR(cm_id))
return PTR_ERR(cm_id);
+ mutex_lock(&id_priv->qp_mutex);
cm_id->tos = id_priv->tos;
cm_id->tos_set = id_priv->tos_set;
+ mutex_unlock(&id_priv->qp_mutex);
+
id_priv->cm_id.iw = cm_id;
memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
The struct rdma_id_private contains three bit-fields, tos_set, timeout_set, and min_rnr_timer_set. These are set by accessor functions without any synchronization. If two or all accessor functions are invoked in close proximity in time, there will be Read-Modify-Write from several contexts to the same variable, and the result will be intermittent. Fixed by protecting the bit-fields by the qp_mutex in the accessor functions. The consumer of timeout_set and min_rnr_timer_set is in rdma_init_qp_attr(), which is called with qp_mutex held for connected QPs. Explicit locking is added for the consumers of tos and tos_set. This commit depends on ("RDMA/cma: Remove unnecessary INIT->INIT transition"), since the call to rdma_init_qp_attr() from cma_init_conn_qp() does not hold the qp_mutex. Fixes: 2c1619edef61 ("IB/cma: Define option to set ack timeout and pack tos_set") Fixes: 3aeffc46afde ("IB/cma: Introduce rdma_set_min_rnr_timer()") Signed-off-by: Håkon Bugge <haakon.bugge@oracle.com> --- v1 -> v2 * Added mutex exclusion in cma_resolve_iboe_route() protecting timeout/timeout_set --- drivers/infiniband/core/cma.c | 20 +++++++++++++++++++- 1 file changed, 19 insertions(+), 1 deletion(-)