]> err.no Git - linux-2.6/commitdiff
IB/cm: Include HCA ACK delay in local ACK timeout
authorSean Hefty <sean.hefty@intel.com>
Mon, 18 Jun 2007 18:09:37 +0000 (11:09 -0700)
committerRoland Dreier <rolandd@cisco.com>
Wed, 11 Jul 2007 04:50:05 +0000 (21:50 -0700)
The IB CM should include the HCA ACK delay when calculating the local
ACK timeout value to use for RC QPs.  If the HCA ACK delay is large
enough relative to the packet life time, then if it is not taken into
account, the calculated timeout value ends up being too small, which
can result in "retry exceeded" errors.

Signed-off-by: Sean Hefty <sean.hefty@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
drivers/infiniband/core/cm.c
drivers/infiniband/core/cma.c
drivers/infiniband/core/ucm.c
drivers/infiniband/ulp/ipoib/ipoib_cm.c
include/rdma/ib_cm.h

index 16181d6558548837b22decaa9484d9c033527458..c7007c45889ab10469198c0174a2fb0f3d1b3b43 100644 (file)
@@ -87,6 +87,7 @@ struct cm_port {
 struct cm_device {
        struct list_head list;
        struct ib_device *device;
+       u8 ack_delay;
        struct cm_port port[0];
 };
 
@@ -95,7 +96,7 @@ struct cm_av {
        union ib_gid dgid;
        struct ib_ah_attr ah_attr;
        u16 pkey_index;
-       u8 packet_life_time;
+       u8 timeout;
 };
 
 struct cm_work {
@@ -154,6 +155,7 @@ struct cm_id_private {
        u8 retry_count;
        u8 rnr_retry_count;
        u8 service_timeout;
+       u8 target_ack_delay;
 
        struct list_head work_list;
        atomic_t work_count;
@@ -293,7 +295,7 @@ static int cm_init_av_by_path(struct ib_sa_path_rec *path, struct cm_av *av)
        av->port = port;
        ib_init_ah_from_path(cm_dev->device, port->port_num, path,
                             &av->ah_attr);
-       av->packet_life_time = path->packet_life_time;
+       av->timeout = path->packet_life_time + 1;
        return 0;
 }
 
@@ -643,6 +645,25 @@ static inline int cm_convert_to_ms(int iba_time)
        return 1 << max(iba_time - 8, 0);
 }
 
+/*
+ * calculate: 4.096x2^ack_timeout = 4.096x2^ack_delay + 2x4.096x2^life_time
+ * Because of how ack_timeout is stored, adding one doubles the timeout.
+ * To avoid large timeouts, select the max(ack_delay, life_time + 1), and
+ * increment it (round up) only if the other is within 50%.
+ */
+static u8 cm_ack_timeout(u8 ca_ack_delay, u8 packet_life_time)
+{
+       int ack_timeout = packet_life_time + 1;
+
+       if (ack_timeout >= ca_ack_delay)
+               ack_timeout += (ca_ack_delay >= (ack_timeout - 1));
+       else
+               ack_timeout = ca_ack_delay +
+                             (ack_timeout >= (ca_ack_delay - 1));
+
+       return min(31, ack_timeout);
+}
+
 static void cm_cleanup_timewait(struct cm_timewait_info *timewait_info)
 {
        if (timewait_info->inserted_remote_id) {
@@ -686,7 +707,7 @@ static void cm_enter_timewait(struct cm_id_private *cm_id_priv)
         * timewait before notifying the user that we've exited timewait.
         */
        cm_id_priv->id.state = IB_CM_TIMEWAIT;
-       wait_time = cm_convert_to_ms(cm_id_priv->av.packet_life_time + 1);
+       wait_time = cm_convert_to_ms(cm_id_priv->av.timeout);
        queue_delayed_work(cm.wq, &cm_id_priv->timewait_info->work.work,
                           msecs_to_jiffies(wait_time));
        cm_id_priv->timewait_info = NULL;
@@ -908,7 +929,8 @@ static void cm_format_req(struct cm_req_msg *req_msg,
        cm_req_set_primary_sl(req_msg, param->primary_path->sl);
        cm_req_set_primary_subnet_local(req_msg, 1); /* local only... */
        cm_req_set_primary_local_ack_timeout(req_msg,
-               min(31, param->primary_path->packet_life_time + 1));
+               cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
+                              param->primary_path->packet_life_time));
 
        if (param->alternate_path) {
                req_msg->alt_local_lid = param->alternate_path->slid;
@@ -923,7 +945,8 @@ static void cm_format_req(struct cm_req_msg *req_msg,
                cm_req_set_alt_sl(req_msg, param->alternate_path->sl);
                cm_req_set_alt_subnet_local(req_msg, 1); /* local only... */
                cm_req_set_alt_local_ack_timeout(req_msg,
-                       min(31, param->alternate_path->packet_life_time + 1));
+                       cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
+                                      param->alternate_path->packet_life_time));
        }
 
        if (param->private_data && param->private_data_len)
@@ -1433,7 +1456,8 @@ static void cm_format_rep(struct cm_rep_msg *rep_msg,
        cm_rep_set_starting_psn(rep_msg, cpu_to_be32(param->starting_psn));
        rep_msg->resp_resources = param->responder_resources;
        rep_msg->initiator_depth = param->initiator_depth;
-       cm_rep_set_target_ack_delay(rep_msg, param->target_ack_delay);
+       cm_rep_set_target_ack_delay(rep_msg,
+                                   cm_id_priv->av.port->cm_dev->ack_delay);
        cm_rep_set_failover(rep_msg, param->failover_accepted);
        cm_rep_set_flow_ctrl(rep_msg, param->flow_control);
        cm_rep_set_rnr_retry_count(rep_msg, param->rnr_retry_count);
@@ -1680,6 +1704,13 @@ static int cm_rep_handler(struct cm_work *work)
        cm_id_priv->responder_resources = rep_msg->initiator_depth;
        cm_id_priv->sq_psn = cm_rep_get_starting_psn(rep_msg);
        cm_id_priv->rnr_retry_count = cm_rep_get_rnr_retry_count(rep_msg);
+       cm_id_priv->target_ack_delay = cm_rep_get_target_ack_delay(rep_msg);
+       cm_id_priv->av.timeout =
+                       cm_ack_timeout(cm_id_priv->target_ack_delay,
+                                      cm_id_priv->av.timeout - 1);
+       cm_id_priv->alt_av.timeout =
+                       cm_ack_timeout(cm_id_priv->target_ack_delay,
+                                      cm_id_priv->alt_av.timeout - 1);
 
        /* todo: handle peer_to_peer */
 
@@ -2291,7 +2322,7 @@ static int cm_mra_handler(struct cm_work *work)
        work->cm_event.param.mra_rcvd.service_timeout =
                                        cm_mra_get_service_timeout(mra_msg);
        timeout = cm_convert_to_ms(cm_mra_get_service_timeout(mra_msg)) +
-                 cm_convert_to_ms(cm_id_priv->av.packet_life_time);
+                 cm_convert_to_ms(cm_id_priv->av.timeout);
 
        spin_lock_irq(&cm_id_priv->lock);
        switch (cm_id_priv->id.state) {
@@ -2363,7 +2394,8 @@ static void cm_format_lap(struct cm_lap_msg *lap_msg,
        cm_lap_set_sl(lap_msg, alternate_path->sl);
        cm_lap_set_subnet_local(lap_msg, 1); /* local only... */
        cm_lap_set_local_ack_timeout(lap_msg,
-               min(31, alternate_path->packet_life_time + 1));
+               cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
+                              alternate_path->packet_life_time));
 
        if (private_data && private_data_len)
                memcpy(lap_msg->private_data, private_data, private_data_len);
@@ -2394,6 +2426,9 @@ int ib_send_cm_lap(struct ib_cm_id *cm_id,
        ret = cm_init_av_by_path(alternate_path, &cm_id_priv->alt_av);
        if (ret)
                goto out;
+       cm_id_priv->alt_av.timeout =
+                       cm_ack_timeout(cm_id_priv->target_ack_delay,
+                                      cm_id_priv->alt_av.timeout - 1);
 
        ret = cm_alloc_msg(cm_id_priv, &msg);
        if (ret)
@@ -3248,8 +3283,7 @@ static int cm_init_qp_rtr_attr(struct cm_id_private *cm_id_priv,
                        *qp_attr_mask |= IB_QP_ALT_PATH;
                        qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
                        qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
-                       qp_attr->alt_timeout =
-                                       cm_id_priv->alt_av.packet_life_time + 1;
+                       qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
                        qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
                }
                ret = 0;
@@ -3287,8 +3321,7 @@ static int cm_init_qp_rts_attr(struct cm_id_private *cm_id_priv,
                                *qp_attr_mask |= IB_QP_TIMEOUT | IB_QP_RETRY_CNT |
                                                 IB_QP_RNR_RETRY |
                                                 IB_QP_MAX_QP_RD_ATOMIC;
-                               qp_attr->timeout =
-                                       cm_id_priv->av.packet_life_time + 1;
+                               qp_attr->timeout = cm_id_priv->av.timeout;
                                qp_attr->retry_cnt = cm_id_priv->retry_count;
                                qp_attr->rnr_retry = cm_id_priv->rnr_retry_count;
                                qp_attr->max_rd_atomic =
@@ -3302,8 +3335,7 @@ static int cm_init_qp_rts_attr(struct cm_id_private *cm_id_priv,
                        *qp_attr_mask = IB_QP_ALT_PATH | IB_QP_PATH_MIG_STATE;
                        qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
                        qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
-                       qp_attr->alt_timeout =
-                               cm_id_priv->alt_av.packet_life_time + 1;
+                       qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
                        qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
                        qp_attr->path_mig_state = IB_MIG_REARM;
                }
@@ -3343,6 +3375,16 @@ int ib_cm_init_qp_attr(struct ib_cm_id *cm_id,
 }
 EXPORT_SYMBOL(ib_cm_init_qp_attr);
 
+void cm_get_ack_delay(struct cm_device *cm_dev)
+{
+       struct ib_device_attr attr;
+
+       if (ib_query_device(cm_dev->device, &attr))
+               cm_dev->ack_delay = 0; /* acks will rely on packet life time */
+       else
+               cm_dev->ack_delay = attr.local_ca_ack_delay;
+}
+
 static void cm_add_one(struct ib_device *device)
 {
        struct cm_device *cm_dev;
@@ -3367,6 +3409,7 @@ static void cm_add_one(struct ib_device *device)
                return;
 
        cm_dev->device = device;
+       cm_get_ack_delay(cm_dev);
 
        set_bit(IB_MGMT_METHOD_SEND, reg_req.method_mask);
        for (i = 1; i <= device->phys_port_cnt; i++) {
index 32a0e66d2a235215bece711619ae36003660142d..23af7a032a035fbb1c082ca69fdc69053d026515 100644 (file)
@@ -2326,7 +2326,6 @@ static int cma_accept_ib(struct rdma_id_private *id_priv,
        rep.private_data_len = conn_param->private_data_len;
        rep.responder_resources = conn_param->responder_resources;
        rep.initiator_depth = conn_param->initiator_depth;
-       rep.target_ack_delay = CMA_CM_RESPONSE_TIMEOUT;
        rep.failover_accepted = 0;
        rep.flow_control = conn_param->flow_control;
        rep.rnr_retry_count = conn_param->rnr_retry_count;
index 2586a3ee8eba6ae35c77a09460d3b9ea18e954fa..424983f5b1ee249f5685907d6f695c03c20d4c4a 100644 (file)
@@ -823,7 +823,6 @@ static ssize_t ib_ucm_send_rep(struct ib_ucm_file *file,
        param.private_data_len    = cmd.len;
        param.responder_resources = cmd.responder_resources;
        param.initiator_depth     = cmd.initiator_depth;
-       param.target_ack_delay    = cmd.target_ack_delay;
        param.failover_accepted   = cmd.failover_accepted;
        param.flow_control        = cmd.flow_control;
        param.rnr_retry_count     = cmd.rnr_retry_count;
index 6764d216c887927d11ee34e06630945bff3504ff..08b4676a38201ad94fc5d165ce2a27ed139a36a1 100644 (file)
@@ -281,7 +281,6 @@ static int ipoib_cm_send_rep(struct net_device *dev, struct ib_cm_id *cm_id,
        rep.private_data_len = sizeof data;
        rep.flow_control = 0;
        rep.rnr_retry_count = req->rnr_retry_count;
-       rep.target_ack_delay = 20; /* FIXME */
        rep.srq = 1;
        rep.qp_num = qp->qp_num;
        rep.starting_psn = psn;
index 5c070176d9ab30bb1bdc8b5ae7524322070b70fe..12243e80c70615676b224310ed90d1785128662b 100644 (file)
@@ -385,7 +385,6 @@ struct ib_cm_rep_param {
        u8              private_data_len;
        u8              responder_resources;
        u8              initiator_depth;
-       u8              target_ack_delay;
        u8              failover_accepted;
        u8              flow_control;
        u8              rnr_retry_count;