Commit ec56dc0b authored by Michael S. Tsirkin's avatar Michael S. Tsirkin Committed by Roland Dreier

IPoIB/cm: Fix performance regression on Mellanox

commit 518b1646 ("IPoIB/cm: Fix SRQ WR leak") introduced a severe
performance regression on Mellanox cards, because keeping a QP in the
error state for extended periods of time moves hardware to the slow
path (until the QP is destroyed).  For example, MPI latency goes from
~3 usecs to ~7 usecs.

Fix this by posting a send WR on one of the QPs that are being
flushed, instead of using a separate drain QP that is kept in the
error state.

This fixes bug <https://bugs.openfabrics.org/show_bug.cgi?id=636>,
reported and bisected by Scott Weitzenkamp at Cisco and debugged by
Sasha Mikheev at Voltaire.
Signed-off-by: default avatarMichael S. Tsirkin <mst@dev.mellanox.co.il>
Signed-off-by: default avatarRoland Dreier <rolandd@cisco.com>
parent 8b7e1577
......@@ -156,7 +156,7 @@ struct ipoib_cm_data {
* - and then invoke a Destroy QP or Reset QP.
*
* We use the second option and wait for a completion on the
* rx_drain_qp before destroying QPs attached to our SRQ.
* same CQ before destroying QPs attached to our SRQ.
*/
enum ipoib_cm_state {
......@@ -199,7 +199,6 @@ struct ipoib_cm_dev_priv {
struct ib_srq *srq;
struct ipoib_cm_rx_buf *srq_ring;
struct ib_cm_id *id;
struct ib_qp *rx_drain_qp; /* generates WR described in 10.3.1 */
struct list_head passive_ids; /* state: LIVE */
struct list_head rx_error_list; /* state: ERROR */
struct list_head rx_flush_list; /* state: FLUSH, drain not started */
......
......@@ -69,8 +69,9 @@ static struct ib_qp_attr ipoib_cm_err_attr = {
#define IPOIB_CM_RX_DRAIN_WRID 0x7fffffff
static struct ib_recv_wr ipoib_cm_rx_drain_wr = {
.wr_id = IPOIB_CM_RX_DRAIN_WRID
static struct ib_send_wr ipoib_cm_rx_drain_wr = {
.wr_id = IPOIB_CM_RX_DRAIN_WRID,
.opcode = IB_WR_SEND,
};
static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
......@@ -163,16 +164,22 @@ static struct sk_buff *ipoib_cm_alloc_rx_skb(struct net_device *dev, int id, int
static void ipoib_cm_start_rx_drain(struct ipoib_dev_priv* priv)
{
struct ib_recv_wr *bad_wr;
struct ib_send_wr *bad_wr;
struct ipoib_cm_rx *p;
/* rx_drain_qp send queue depth is 1, so
/* We only reserved 1 extra slot in CQ for drain WRs, so
* make sure we have at most 1 outstanding WR. */
if (list_empty(&priv->cm.rx_flush_list) ||
!list_empty(&priv->cm.rx_drain_list))
return;
if (ib_post_recv(priv->cm.rx_drain_qp, &ipoib_cm_rx_drain_wr, &bad_wr))
ipoib_warn(priv, "failed to post rx_drain wr\n");
/*
* QPs on flush list are error state. This way, a "flush
* error" WC will be immediately generated for each WR we post.
*/
p = list_entry(priv->cm.rx_flush_list.next, typeof(*p), list);
if (ib_post_send(p->qp, &ipoib_cm_rx_drain_wr, &bad_wr))
ipoib_warn(priv, "failed to post drain wr\n");
list_splice_init(&priv->cm.rx_flush_list, &priv->cm.rx_drain_list);
}
......@@ -199,10 +206,10 @@ static struct ib_qp *ipoib_cm_create_rx_qp(struct net_device *dev,
struct ipoib_dev_priv *priv = netdev_priv(dev);
struct ib_qp_init_attr attr = {
.event_handler = ipoib_cm_rx_event_handler,
.send_cq = priv->cq, /* does not matter, we never send anything */
.send_cq = priv->cq, /* For drain WR */
.recv_cq = priv->cq,
.srq = priv->cm.srq,
.cap.max_send_wr = 1, /* FIXME: 0 Seems not to work */
.cap.max_send_wr = 1, /* For drain WR */
.cap.max_send_sge = 1, /* FIXME: 0 Seems not to work */
.sq_sig_type = IB_SIGNAL_ALL_WR,
.qp_type = IB_QPT_RC,
......@@ -242,6 +249,27 @@ static int ipoib_cm_modify_rx_qp(struct net_device *dev,
ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
return ret;
}
/*
* Current Mellanox HCA firmware won't generate completions
* with error for drain WRs unless the QP has been moved to
* RTS first. This work-around leaves a window where a QP has
* moved to error asynchronously, but this will eventually get
* fixed in firmware, so let's not error out if modify QP
* fails.
*/
qp_attr.qp_state = IB_QPS_RTS;
ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
if (ret) {
ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
return 0;
}
ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
if (ret) {
ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
return 0;
}
return 0;
}
......@@ -623,38 +651,11 @@ static void ipoib_cm_tx_completion(struct ib_cq *cq, void *tx_ptr)
int ipoib_cm_dev_open(struct net_device *dev)
{
struct ipoib_dev_priv *priv = netdev_priv(dev);
struct ib_qp_init_attr qp_init_attr = {
.send_cq = priv->cq, /* does not matter, we never send anything */
.recv_cq = priv->cq,
.cap.max_send_wr = 1, /* FIXME: 0 Seems not to work */
.cap.max_send_sge = 1, /* FIXME: 0 Seems not to work */
.cap.max_recv_wr = 1,
.cap.max_recv_sge = 1, /* FIXME: 0 Seems not to work */
.sq_sig_type = IB_SIGNAL_ALL_WR,
.qp_type = IB_QPT_UC,
};
int ret;
if (!IPOIB_CM_SUPPORTED(dev->dev_addr))
return 0;
priv->cm.rx_drain_qp = ib_create_qp(priv->pd, &qp_init_attr);
if (IS_ERR(priv->cm.rx_drain_qp)) {
printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name);
ret = PTR_ERR(priv->cm.rx_drain_qp);
return ret;
}
/*
* We put the QP in error state directly. This way, a "flush
* error" WC will be immediately generated for each WR we post.
*/
ret = ib_modify_qp(priv->cm.rx_drain_qp, &ipoib_cm_err_attr, IB_QP_STATE);
if (ret) {
ipoib_warn(priv, "failed to modify drain QP to error: %d\n", ret);
goto err_qp;
}
priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, dev);
if (IS_ERR(priv->cm.id)) {
printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name);
......@@ -676,8 +677,6 @@ int ipoib_cm_dev_open(struct net_device *dev)
ib_destroy_cm_id(priv->cm.id);
err_cm:
priv->cm.id = NULL;
err_qp:
ib_destroy_qp(priv->cm.rx_drain_qp);
return ret;
}
......@@ -740,7 +739,6 @@ void ipoib_cm_dev_stop(struct net_device *dev)
kfree(p);
}
ib_destroy_qp(priv->cm.rx_drain_qp);
cancel_delayed_work(&priv->cm.stale_task);
}
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment