IB/{hfi1, rdmavt, qib}: Implement CQ completion vector support

Currently the driver doesn't support completion vectors. These
are used to indicate which sets of CQs should be grouped together
into the same vector. A vector is a CQ processing thread that
runs on a specific CPU.

If an application has several CQs bound to different completion
vectors, and each completion vector runs on different CPUs, then
the completion queue workload is balanced. This helps scale as more
nodes are used.

Implement CQ completion vector support using a global workqueue
where a CQ entry is queued to the CPU corresponding to the CQ's
completion vector. Since the workqueue is global, it's guaranteed
to always be there when queueing CQ entries; Therefore, the RCU
locking for cq->rdi->worker in the hot path is superfluous.

Each completion vector is assigned to a different CPU. The number of
completion vectors available is computed by taking the number of
online, physical CPUs from the local NUMA node and subtracting the
CPUs used for kernel receive queues and the general interrupt.
Special use cases:

  * If there are no CPUs left for completion vectors, the same CPU
    for the general interrupt is used; Therefore, there would only
    be one completion vector available.

  * For multi-HFI systems, the number of completion vectors available
    for each device is the total number of completion vectors in
    the local NUMA node divided by the number of devices in the same
    NUMA node. If there's a division remainder, the first device to
    get initialized gets an extra completion vector.

Upon a CQ creation, an invalid completion vector could be specified.
Handle it as follows:

  * If the completion vector is less than 0, set it to 0.

  * Set the completion vector to the result of the passed completion
    vector moded with the number of device completion vectors
    available.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Sebastian Sanchez <sebastian.sanchez@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

4 files changed, 82 insertions, 75 deletions

diff --git a/drivers/infiniband/sw/rdmavt/cq.c b/drivers/infiniband/sw/rdmavt/cq.c
index 340c17aba3b0..4f1544ad4aff 100644
--- a/drivers/infiniband/sw/rdmavt/cq.c
+++ b/drivers/infiniband/sw/rdmavt/cq.c
@@ -1,5 +1,5 @@
 /*
- * Copyright(c) 2016 Intel Corporation.
+ * Copyright(c) 2016 - 2018 Intel Corporation.
  *
  * This file is provided under a dual BSD/GPLv2 license.  When using or
  * redistributing this file, you may do so under either license.
@@ -47,11 +47,12 @@
 
 #include <linux/slab.h>
 #include <linux/vmalloc.h>
-#include <linux/kthread.h>
 #include "cq.h"
 #include "vt.h"
 #include "trace.h"
 
+static struct workqueue_struct *comp_vector_wq;
+
 /**
  * rvt_cq_enter - add a new entry to the completion queue
  * @cq: completion queue
@@ -120,27 +121,21 @@ void rvt_cq_enter(struct rvt_cq *cq, struct ib_wc *entry, bool solicited)
 	if (cq->notify == IB_CQ_NEXT_COMP ||
 	    (cq->notify == IB_CQ_SOLICITED &&
 	     (solicited || entry->status != IB_WC_SUCCESS))) {
-		struct kthread_worker *worker;
-
 		/*
 		 * This will cause send_complete() to be called in
 		 * another thread.
 		 */
-		rcu_read_lock();
-		worker = rcu_dereference(cq->rdi->worker);
-		if (likely(worker)) {
-			cq->notify = RVT_CQ_NONE;
-			cq->triggered++;
-			kthread_queue_work(worker, &cq->comptask);
-		}
-		rcu_read_unlock();
+		cq->notify = RVT_CQ_NONE;
+		cq->triggered++;
+		queue_work_on(cq->comp_vector_cpu, comp_vector_wq,
+			      &cq->comptask);
 	}
 
 	spin_unlock_irqrestore(&cq->lock, flags);
 }
 EXPORT_SYMBOL(rvt_cq_enter);
 
-static void send_complete(struct kthread_work *work)
+static void send_complete(struct work_struct *work)
 {
 	struct rvt_cq *cq = container_of(work, struct rvt_cq, comptask);
 
@@ -192,6 +187,7 @@ struct ib_cq *rvt_create_cq(struct ib_device *ibdev,
 	struct ib_cq *ret;
 	u32 sz;
 	unsigned int entries = attr->cqe;
+	int comp_vector = attr->comp_vector;
 
 	if (attr->flags)
 		return ERR_PTR(-EINVAL);
@@ -199,6 +195,11 @@ struct ib_cq *rvt_create_cq(struct ib_device *ibdev,
 	if (entries < 1 || entries > rdi->dparms.props.max_cqe)
 		return ERR_PTR(-EINVAL);
 
+	if (comp_vector < 0)
+		comp_vector = 0;
+
+	comp_vector = comp_vector % rdi->ibdev.num_comp_vectors;
+
 	/* Allocate the completion queue structure. */
 	cq = kzalloc_node(sizeof(*cq), GFP_KERNEL, rdi->dparms.node);
 	if (!cq)
@@ -267,14 +268,22 @@ struct ib_cq *rvt_create_cq(struct ib_device *ibdev,
 	 * an error.
 	 */
 	cq->rdi = rdi;
+	if (rdi->driver_f.comp_vect_cpu_lookup)
+		cq->comp_vector_cpu =
+			rdi->driver_f.comp_vect_cpu_lookup(rdi, comp_vector);
+	else
+		cq->comp_vector_cpu =
+			cpumask_first(cpumask_of_node(rdi->dparms.node));
+
 	cq->ibcq.cqe = entries;
 	cq->notify = RVT_CQ_NONE;
 	spin_lock_init(&cq->lock);
-	kthread_init_work(&cq->comptask, send_complete);
+	INIT_WORK(&cq->comptask, send_complete);
 	cq->queue = wc;
 
 	ret = &cq->ibcq;
 
+	trace_rvt_create_cq(cq, attr);
 	goto done;
 
 bail_ip:
@@ -300,7 +309,7 @@ int rvt_destroy_cq(struct ib_cq *ibcq)
 	struct rvt_cq *cq = ibcq_to_rvtcq(ibcq);
 	struct rvt_dev_info *rdi = cq->rdi;
 
-	kthread_flush_work(&cq->comptask);
+	flush_work(&cq->comptask);
 	spin_lock_irq(&rdi->n_cqs_lock);
 	rdi->n_cqs_allocated--;
 	spin_unlock_irq(&rdi->n_cqs_lock);
@@ -510,24 +519,13 @@ int rvt_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *entry)
  *
  * Return: 0 on success
  */
-int rvt_driver_cq_init(struct rvt_dev_info *rdi)
+int rvt_driver_cq_init(void)
 {
-	int cpu;
-	struct kthread_worker *worker;
-
-	if (rcu_access_pointer(rdi->worker))
-		return 0;
-
-	spin_lock_init(&rdi->n_cqs_lock);
-
-	cpu = cpumask_first(cpumask_of_node(rdi->dparms.node));
-	worker = kthread_create_worker_on_cpu(cpu, 0,
-					      "%s", rdi->dparms.cq_name);
-	if (IS_ERR(worker))
-		return PTR_ERR(worker);
+	comp_vector_wq = alloc_workqueue("%s", WQ_HIGHPRI | WQ_CPU_INTENSIVE,
+					 0, "rdmavt_cq");
+	if (!comp_vector_wq)
+		return -ENOMEM;
 
-	set_user_nice(worker->task, MIN_NICE);
-	RCU_INIT_POINTER(rdi->worker, worker);
 	return 0;
 }
 
@@ -535,23 +533,8 @@ int rvt_driver_cq_init(struct rvt_dev_info *rdi)
  * rvt_cq_exit - tear down cq reources
  * @rdi: rvt dev structure
  */
-void rvt_cq_exit(struct rvt_dev_info *rdi)
+void rvt_cq_exit(void)
 {
-	struct kthread_worker *worker;
-
-	if (!rcu_access_pointer(rdi->worker))
-		return;
-
-	spin_lock(&rdi->n_cqs_lock);
-	worker = rcu_dereference_protected(rdi->worker,
-					   lockdep_is_held(&rdi->n_cqs_lock));
-	if (!worker) {
-		spin_unlock(&rdi->n_cqs_lock);
-		return;
-	}
-	RCU_INIT_POINTER(rdi->worker, NULL);
-	spin_unlock(&rdi->n_cqs_lock);
-	synchronize_rcu();
-
-	kthread_destroy_worker(worker);
+	destroy_workqueue(comp_vector_wq);
+	comp_vector_wq = NULL;
 }
diff --git a/drivers/infiniband/sw/rdmavt/cq.h b/drivers/infiniband/sw/rdmavt/cq.h
index 6182c29eff66..72184b1c176b 100644
--- a/drivers/infiniband/sw/rdmavt/cq.h
+++ b/drivers/infiniband/sw/rdmavt/cq.h
@@ -2,7 +2,7 @@
 #define DEF_RVTCQ_H
 
 /*
- * Copyright(c) 2016 Intel Corporation.
+ * Copyright(c) 2016 - 2018 Intel Corporation.
  *
  * This file is provided under a dual BSD/GPLv2 license.  When using or
  * redistributing this file, you may do so under either license.
@@ -59,6 +59,6 @@ int rvt_destroy_cq(struct ib_cq *ibcq);
 int rvt_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags notify_flags);
 int rvt_resize_cq(struct ib_cq *ibcq, int cqe, struct ib_udata *udata);
 int rvt_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *entry);
-int rvt_driver_cq_init(struct rvt_dev_info *rdi);
-void rvt_cq_exit(struct rvt_dev_info *rdi);
+int rvt_driver_cq_init(void);
+void rvt_cq_exit(void);
 #endif          /* DEF_RVTCQ_H */
diff --git a/drivers/infiniband/sw/rdmavt/trace_cq.h b/drivers/infiniband/sw/rdmavt/trace_cq.h
index a315850aa9bb..df8e1adbef9d 100644
--- a/drivers/infiniband/sw/rdmavt/trace_cq.h
+++ b/drivers/infiniband/sw/rdmavt/trace_cq.h
@@ -1,5 +1,5 @@
 /*
- * Copyright(c) 2016 Intel Corporation.
+ * Copyright(c) 2016 - 2018 Intel Corporation.
  *
  * This file is provided under a dual BSD/GPLv2 license.  When using or
  * redistributing this file, you may do so under either license.
@@ -71,6 +71,39 @@ __print_symbolic(opcode,                                      \
 	wc_opcode_name(RECV),                                 \
 	wc_opcode_name(RECV_RDMA_WITH_IMM))
 
+#define CQ_ATTR_PRINT \
+"[%s] user cq %s cqe %u comp_vector %d comp_vector_cpu %d flags %x"
+
+DECLARE_EVENT_CLASS(rvt_cq_template,
+		    TP_PROTO(struct rvt_cq *cq,
+			     const struct ib_cq_init_attr *attr),
+		    TP_ARGS(cq, attr),
+		    TP_STRUCT__entry(RDI_DEV_ENTRY(cq->rdi)
+				     __field(struct rvt_mmap_info *, ip)
+				     __field(unsigned int, cqe)
+				     __field(int, comp_vector)
+				     __field(int, comp_vector_cpu)
+				     __field(u32, flags)
+				     ),
+		    TP_fast_assign(RDI_DEV_ASSIGN(cq->rdi)
+				   __entry->ip = cq->ip;
+				   __entry->cqe = attr->cqe;
+				   __entry->comp_vector = attr->comp_vector;
+				   __entry->comp_vector_cpu =
+							cq->comp_vector_cpu;
+				   __entry->flags = attr->flags;
+				   ),
+		    TP_printk(CQ_ATTR_PRINT, __get_str(dev),
+			      __entry->ip ? "true" : "false", __entry->cqe,
+			      __entry->comp_vector, __entry->comp_vector_cpu,
+			      __entry->flags
+			      )
+);
+
+DEFINE_EVENT(rvt_cq_template, rvt_create_cq,
+	     TP_PROTO(struct rvt_cq *cq, const struct ib_cq_init_attr *attr),
+	     TP_ARGS(cq, attr));
+
 #define CQ_PRN \
 "[%s] idx %u wr_id %llx status %u opcode %u,%s length %u qpn %x"
 
diff --git a/drivers/infiniband/sw/rdmavt/vt.c b/drivers/infiniband/sw/rdmavt/vt.c
index 434199d0bc96..17e4abc067af 100644
--- a/drivers/infiniband/sw/rdmavt/vt.c
+++ b/drivers/infiniband/sw/rdmavt/vt.c
@@ -1,5 +1,5 @@
 /*
- * Copyright(c) 2016 Intel Corporation.
+ * Copyright(c) 2016 - 2018 Intel Corporation.
  *
  * This file is provided under a dual BSD/GPLv2 license.  When using or
  * redistributing this file, you may do so under either license.
@@ -49,6 +49,7 @@
 #include <linux/kernel.h>
 #include <linux/dma-mapping.h>
 #include "vt.h"
+#include "cq.h"
 #include "trace.h"
 
 #define RVT_UVERBS_ABI_VERSION 2
@@ -58,21 +59,18 @@ MODULE_DESCRIPTION("RDMA Verbs Transport Library");
 
 static int rvt_init(void)
 {
-	/*
-	 * rdmavt does not need to do anything special when it starts up. All it
-	 * needs to do is sit and wait until a driver attempts registration.
-	 */
-	return 0;
+	int ret = rvt_driver_cq_init();
+
+	if (ret)
+		pr_err("Error in driver CQ init.\n");
+
+	return ret;
 }
 module_init(rvt_init);
 
 static void rvt_cleanup(void)
 {
-	/*
-	 * Nothing to do at exit time either. The module won't be able to be
-	 * removed until all drivers are gone which means all the dev structs
-	 * are gone so there is really nothing to do.
-	 */
+	rvt_cq_exit();
 }
 module_exit(rvt_cleanup);
 
@@ -777,11 +775,7 @@ int rvt_register_device(struct rvt_dev_info *rdi, u32 driver_id)
 	}
 
 	/* Completion queues */
-	ret = rvt_driver_cq_init(rdi);
-	if (ret) {
-		pr_err("Error in driver CQ init.\n");
-		goto bail_mr;
-	}
+	spin_lock_init(&rdi->n_cqs_lock);
 
 	/* DMA Operations */
 	rdi->ibdev.dev.dma_ops = rdi->ibdev.dev.dma_ops ? : &dma_virt_ops;
@@ -829,14 +823,15 @@ int rvt_register_device(struct rvt_dev_info *rdi, u32 driver_id)
 		(1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)         |
 		(1ull << IB_USER_VERBS_CMD_POST_SRQ_RECV);
 	rdi->ibdev.node_type = RDMA_NODE_IB_CA;
-	rdi->ibdev.num_comp_vectors = 1;
+	if (!rdi->ibdev.num_comp_vectors)
+		rdi->ibdev.num_comp_vectors = 1;
 
 	rdi->ibdev.driver_id = driver_id;
 	/* We are now good to announce we exist */
 	ret =  ib_register_device(&rdi->ibdev, rdi->driver_f.port_callback);
 	if (ret) {
 		rvt_pr_err(rdi, "Failed to register driver with ib core.\n");
-		goto bail_cq;
+		goto bail_mr;
 	}
 
 	rvt_create_mad_agents(rdi);
@@ -844,9 +839,6 @@ int rvt_register_device(struct rvt_dev_info *rdi, u32 driver_id)
 	rvt_pr_info(rdi, "Registration with rdmavt done.\n");
 	return ret;
 
-bail_cq:
-	rvt_cq_exit(rdi);
-
 bail_mr:
 	rvt_mr_exit(rdi);
 
@@ -870,7 +862,6 @@ void rvt_unregister_device(struct rvt_dev_info *rdi)
 	rvt_free_mad_agents(rdi);
 
 	ib_unregister_device(&rdi->ibdev);
-	rvt_cq_exit(rdi);
 	rvt_mr_exit(rdi);
 	rvt_qp_exit(rdi);
 }