IB/qib: Reduce sdma_lock contention

Profiling has shown that sdma_lock is proving a bottleneck for
performance. The situations include:
 - RDMA reads when krcvqs > 1
 - post sends from multiple threads

For RDMA read the current global qib_wq mechanism runs on all CPUs
and contends for the sdma_lock when multiple RMDA read requests are
fielded on differenct CPUs. For post sends, the direct call to
qib_do_send() from multiple threads causes the contention.

Since the sdma mechanism is per port, this fix converts the existing
workqueue to a per port single thread workqueue to reduce the lock
contention in the RDMA read case, and for any other case where the QP
is scheduled via the workqueue mechanism from more than 1 CPU.

For the post send case, This patch modifies the post send code to test
for a non empty sdma engine.  If the sdma is not idle the (now single
thread) workqueue will be used to trigger the send engine instead of
the direct call to qib_do_send().

Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>

1 files changed, 49 insertions, 2 deletions

diff --git a/drivers/infiniband/hw/qib/qib_init.c b/drivers/infiniband/hw/qib/qib_init.c
index dc14e100a7f1..306e65e99e99 100644
--- a/drivers/infiniband/hw/qib/qib_init.c
+++ b/drivers/infiniband/hw/qib/qib_init.c
@@ -1,6 +1,6 @@
 /*
- * Copyright (c) 2006, 2007, 2008, 2009, 2010 QLogic Corporation.
- * All rights reserved.
+ * Copyright (c) 2012 Intel Corporation.  All rights reserved.
+ * Copyright (c) 2006 - 2012 QLogic Corporation. All rights reserved.
  * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
  *
  * This software is available to you under a choice of one of two
@@ -210,6 +210,8 @@ void qib_init_pportdata(struct qib_pportdata *ppd, struct qib_devdata *dd,
 	init_timer(&ppd->symerr_clear_timer);
 	ppd->symerr_clear_timer.function = qib_clear_symerror_on_linkup;
 	ppd->symerr_clear_timer.data = (unsigned long)ppd;
+
+	ppd->qib_wq = NULL;
 }
 
 static int init_pioavailregs(struct qib_devdata *dd)
@@ -483,6 +485,42 @@ static void init_piobuf_state(struct qib_devdata *dd)
 }
 
 /**
+ * qib_create_workqueues - create per port workqueues
+ * @dd: the qlogic_ib device
+ */
+static int qib_create_workqueues(struct qib_devdata *dd)
+{
+	int pidx;
+	struct qib_pportdata *ppd;
+
+	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
+		ppd = dd->pport + pidx;
+		if (!ppd->qib_wq) {
+			char wq_name[8]; /* 3 + 2 + 1 + 1 + 1 */
+			snprintf(wq_name, sizeof(wq_name), "qib%d_%d",
+				dd->unit, pidx);
+			ppd->qib_wq =
+				create_singlethread_workqueue(wq_name);
+			if (!ppd->qib_wq)
+				goto wq_error;
+		}
+	}
+	return 0;
+wq_error:
+	pr_err(
+	 QIB_DRV_NAME ": create_singlethread_workqueue failed for port %d\n",
+	 pidx + 1);
+	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
+		ppd = dd->pport + pidx;
+		if (ppd->qib_wq) {
+			destroy_workqueue(ppd->qib_wq);
+			ppd->qib_wq = NULL;
+		}
+	}
+	return -ENOMEM;
+}
+
+/**
  * qib_init - do the actual initialization sequence on the chip
  * @dd: the qlogic_ib device
  * @reinit: reinitializing, so don't allocate new memory
@@ -764,6 +802,11 @@ static void qib_shutdown_device(struct qib_devdata *dd)
 		 * We can't count on interrupts since we are stopping.
 		 */
 		dd->f_quiet_serdes(ppd);
+
+		if (ppd->qib_wq) {
+			destroy_workqueue(ppd->qib_wq);
+			ppd->qib_wq = NULL;
+		}
 	}
 
 	qib_update_eeprom_log(dd);
@@ -1249,6 +1292,10 @@ static int __devinit qib_init_one(struct pci_dev *pdev,
 	if (ret)
 		goto bail; /* error already printed */
 
+	ret = qib_create_workqueues(dd);
+	if (ret)
+		goto bail;
+
 	/* do the generic initialization */
 	initfail = qib_init(dd, 0);