[PATCH 0/8] IB/rdmavt: Add functions for cq, qp, post send and recv

[PATCH 0/8] IB/rdmavt: Add functions for cq, qp, post send and recv

[Dennis Dalessandro]

This series continues adding support for verbs into rdmavt. This brings in
completion queue, queue pair, and post send/recv functionality.

It also introduces event tracing and moves some more of the hfi1/qib
registration functionality.

This patch set applies on top of the "Start to rely on rdmavt for qp support"
series previously submitted.

---

Dennis Dalessandro (8):
      IB/rdmavt: Add completion queue functions
      IB/rdmavt: Add post send to rdmavt
      IB/rdmavt: Add support for tracing events
      IB/rdmavt: Add modify qp
      IB/rdmavt: Add destroy qp verb
      IB/rdmavt: Add post receive to rdmavt
      IB/rdmavt: Add multicast functions
      IB/rdmavt: Add misc dev register functionality


 drivers/infiniband/sw/rdmavt/Makefile |    5 
 drivers/infiniband/sw/rdmavt/cq.c     |  445 +++++++++++++++++++
 drivers/infiniband/sw/rdmavt/cq.h     |    4 
 drivers/infiniband/sw/rdmavt/mcast.c  |  335 ++++++++++++++-
 drivers/infiniband/sw/rdmavt/mcast.h  |    2 
 drivers/infiniband/sw/rdmavt/qp.c     |  760 ++++++++++++++++++++++++++++++++-
 drivers/infiniband/sw/rdmavt/trace.c  |   49 ++
 drivers/infiniband/sw/rdmavt/trace.h  |  117 +++++
 drivers/infiniband/sw/rdmavt/vt.c     |   60 +++
 include/rdma/rdma_vt.h                |   66 +++
 include/rdma/rdmavt_cq.h              |   99 ++++
 include/rdma/rdmavt_mr.h              |    9 
 include/rdma/rdmavt_qp.h              |   66 +++
 13 files changed, 1974 insertions(+), 43 deletions(-)
 create mode 100644 drivers/infiniband/sw/rdmavt/trace.c
 create mode 100644 drivers/infiniband/sw/rdmavt/trace.h
 create mode 100644 include/rdma/rdmavt_cq.h

-- 
-Denny
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH 1/8] IB/rdmavt: Add completion queue functions

[Dennis Dalessandro]

Brings in completion queue functionality. A kthread worker is added to
the rvt_dev_info to serve as a worker for completion queues.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
Reviewed-by: Harish Chegondi <harish.chegondi-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
---
 drivers/infiniband/sw/rdmavt/cq.c |  445 +++++++++++++++++++++++++++++++++++++
 drivers/infiniband/sw/rdmavt/cq.h |    4 
 drivers/infiniband/sw/rdmavt/vt.c |   11 +
 include/rdma/rdma_vt.h            |    9 +
 include/rdma/rdmavt_cq.h          |   99 ++++++++
 5 files changed, 561 insertions(+), 7 deletions(-)
 create mode 100644 include/rdma/rdmavt_cq.h

diff --git a/drivers/infiniband/sw/rdmavt/cq.c b/drivers/infiniband/sw/rdmavt/cq.c
index 8d96194..7308a27 100644
--- a/drivers/infiniband/sw/rdmavt/cq.c
+++ b/drivers/infiniband/sw/rdmavt/cq.c
@@ -45,7 +45,126 @@
  *
  */
 
+#include <linux/slab.h>
+#include <linux/vmalloc.h>
+#include <linux/kthread.h>
 #include "cq.h"
+#include "vt.h"
+
+/**
+ * rvt_cq_enter - add a new entry to the completion queue
+ * @cq: completion queue
+ * @entry: work completion entry to add
+ * @sig: true if @entry is solicited
+ *
+ * This may be called with qp->s_lock held.
+ */
+void rvt_cq_enter(struct rvt_cq *cq, struct ib_wc *entry, bool solicited)
+{
+	struct rvt_cq_wc *wc;
+	unsigned long flags;
+	u32 head;
+	u32 next;
+
+	spin_lock_irqsave(&cq->lock, flags);
+
+	/*
+	 * Note that the head pointer might be writable by user processes.
+	 * Take care to verify it is a sane value.
+	 */
+	wc = cq->queue;
+	head = wc->head;
+	if (head >= (unsigned)cq->ibcq.cqe) {
+		head = cq->ibcq.cqe;
+		next = 0;
+	} else {
+		next = head + 1;
+	}
+
+	if (unlikely(next == wc->tail)) {
+		spin_unlock_irqrestore(&cq->lock, flags);
+		if (cq->ibcq.event_handler) {
+			struct ib_event ev;
+
+			ev.device = cq->ibcq.device;
+			ev.element.cq = &cq->ibcq;
+			ev.event = IB_EVENT_CQ_ERR;
+			cq->ibcq.event_handler(&ev, cq->ibcq.cq_context);
+		}
+		return;
+	}
+	if (cq->ip) {
+		wc->uqueue[head].wr_id = entry->wr_id;
+		wc->uqueue[head].status = entry->status;
+		wc->uqueue[head].opcode = entry->opcode;
+		wc->uqueue[head].vendor_err = entry->vendor_err;
+		wc->uqueue[head].byte_len = entry->byte_len;
+		wc->uqueue[head].ex.imm_data =
+			(__u32 __force)entry->ex.imm_data;
+		wc->uqueue[head].qp_num = entry->qp->qp_num;
+		wc->uqueue[head].src_qp = entry->src_qp;
+		wc->uqueue[head].wc_flags = entry->wc_flags;
+		wc->uqueue[head].pkey_index = entry->pkey_index;
+		wc->uqueue[head].slid = entry->slid;
+		wc->uqueue[head].sl = entry->sl;
+		wc->uqueue[head].dlid_path_bits = entry->dlid_path_bits;
+		wc->uqueue[head].port_num = entry->port_num;
+		/* Make sure entry is written before the head index. */
+		smp_wmb();
+	} else {
+		wc->kqueue[head] = *entry;
+	}
+	wc->head = next;
+
+	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.
+		 */
+		smp_read_barrier_depends(); /* see rvt_cq_exit */
+		worker = cq->rdi->worker;
+		if (likely(worker)) {
+			cq->notify = RVT_CQ_NONE;
+			cq->triggered++;
+			queue_kthread_work(worker, &cq->comptask);
+		}
+	}
+
+	spin_unlock_irqrestore(&cq->lock, flags);
+}
+EXPORT_SYMBOL(rvt_cq_enter);
+
+static void send_complete(struct kthread_work *work)
+{
+	struct rvt_cq *cq = container_of(work, struct rvt_cq, comptask);
+
+	/*
+	 * The completion handler will most likely rearm the notification
+	 * and poll for all pending entries.  If a new completion entry
+	 * is added while we are in this routine, queue_work()
+	 * won't call us again until we return so we check triggered to
+	 * see if we need to call the handler again.
+	 */
+	for (;;) {
+		u8 triggered = cq->triggered;
+
+		/*
+		 * IPoIB connected mode assumes the callback is from a
+		 * soft IRQ. We simulate this by blocking "bottom halves".
+		 * See the implementation for ipoib_cm_handle_tx_wc(),
+		 * netif_tx_lock_bh() and netif_tx_lock().
+		 */
+		local_bh_disable();
+		cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
+		local_bh_enable();
+
+		if (cq->triggered == triggered)
+			return;
+	}
+}
 
 /**
  * rvt_create_cq - create a completion queue
@@ -64,7 +183,103 @@ struct ib_cq *rvt_create_cq(struct ib_device *ibdev,
 			    struct ib_ucontext *context,
 			    struct ib_udata *udata)
 {
-	return ERR_PTR(-EOPNOTSUPP);
+	struct rvt_dev_info *rdi = ib_to_rvt(ibdev);
+	struct rvt_cq *cq;
+	struct rvt_cq_wc *wc;
+	struct ib_cq *ret;
+	u32 sz;
+	unsigned int entries = attr->cqe;
+
+	if (attr->flags)
+		return ERR_PTR(-EINVAL);
+
+	if (entries < 1 || entries > rdi->dparms.props.max_cqe)
+		return ERR_PTR(-EINVAL);
+
+	/* Allocate the completion queue structure. */
+	cq = kzalloc(sizeof(*cq), GFP_KERNEL);
+	if (!cq)
+		return ERR_PTR(-ENOMEM);
+
+	/*
+	 * Allocate the completion queue entries and head/tail pointers.
+	 * This is allocated separately so that it can be resized and
+	 * also mapped into user space.
+	 * We need to use vmalloc() in order to support mmap and large
+	 * numbers of entries.
+	 */
+	sz = sizeof(*wc);
+	if (udata && udata->outlen >= sizeof(__u64))
+		sz += sizeof(struct ib_uverbs_wc) * (entries + 1);
+	else
+		sz += sizeof(struct ib_wc) * (entries + 1);
+	wc = vmalloc_user(sz);
+	if (!wc) {
+		ret = ERR_PTR(-ENOMEM);
+		goto bail_cq;
+	}
+
+	/*
+	 * Return the address of the WC as the offset to mmap.
+	 * See rvt_mmap() for details.
+	 */
+	if (udata && udata->outlen >= sizeof(__u64)) {
+		int err;
+
+		cq->ip = rvt_create_mmap_info(rdi, sz, context, wc);
+		if (!cq->ip) {
+			ret = ERR_PTR(-ENOMEM);
+			goto bail_wc;
+		}
+
+		err = ib_copy_to_udata(udata, &cq->ip->offset,
+				       sizeof(cq->ip->offset));
+		if (err) {
+			ret = ERR_PTR(err);
+			goto bail_ip;
+		}
+	}
+
+	spin_lock(&rdi->n_cqs_lock);
+	if (rdi->n_cqs_allocated == rdi->dparms.props.max_cq) {
+		spin_unlock(&rdi->n_cqs_lock);
+		ret = ERR_PTR(-ENOMEM);
+		goto bail_ip;
+	}
+
+	rdi->n_cqs_allocated++;
+	spin_unlock(&rdi->n_cqs_lock);
+
+	if (cq->ip) {
+		spin_lock_irq(&rdi->pending_lock);
+		list_add(&cq->ip->pending_mmaps, &rdi->pending_mmaps);
+		spin_unlock_irq(&rdi->pending_lock);
+	}
+
+	/*
+	 * ib_create_cq() will initialize cq->ibcq except for cq->ibcq.cqe.
+	 * The number of entries should be >= the number requested or return
+	 * an error.
+	 */
+	cq->rdi = rdi;
+	cq->ibcq.cqe = entries;
+	cq->notify = RVT_CQ_NONE;
+	spin_lock_init(&cq->lock);
+	init_kthread_work(&cq->comptask, send_complete);
+	cq->queue = wc;
+
+	ret = &cq->ibcq;
+
+	goto done;
+
+bail_ip:
+	kfree(cq->ip);
+bail_wc:
+	vfree(wc);
+bail_cq:
+	kfree(cq);
+done:
+	return ret;
 }
 
 /**
@@ -77,12 +292,53 @@ struct ib_cq *rvt_create_cq(struct ib_device *ibdev,
  */
 int rvt_destroy_cq(struct ib_cq *ibcq)
 {
-	return -EOPNOTSUPP;
+	struct rvt_cq *cq = ibcq_to_rvtcq(ibcq);
+	struct rvt_dev_info *rdi = cq->rdi;
+
+	flush_kthread_work(&cq->comptask);
+	spin_lock(&rdi->n_cqs_lock);
+	rdi->n_cqs_allocated--;
+	spin_unlock(&rdi->n_cqs_lock);
+	if (cq->ip)
+		kref_put(&cq->ip->ref, rvt_release_mmap_info);
+	else
+		vfree(cq->queue);
+	kfree(cq);
+
+	return 0;
 }
 
+/**
+ * rvt_req_notify_cq - change the notification type for a completion queue
+ * @ibcq: the completion queue
+ * @notify_flags: the type of notification to request
+ *
+ * Returns 0 for success.
+ *
+ * This may be called from interrupt context.  Also called by
+ * ib_req_notify_cq() in the generic verbs code.
+ */
 int rvt_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags notify_flags)
 {
-	return -EOPNOTSUPP;
+	struct rvt_cq *cq = ibcq_to_rvtcq(ibcq);
+	unsigned long flags;
+	int ret = 0;
+
+	spin_lock_irqsave(&cq->lock, flags);
+	/*
+	 * Don't change IB_CQ_NEXT_COMP to IB_CQ_SOLICITED but allow
+	 * any other transitions (see C11-31 and C11-32 in ch. 11.4.2.2).
+	 */
+	if (cq->notify != IB_CQ_NEXT_COMP)
+		cq->notify = notify_flags & IB_CQ_SOLICITED_MASK;
+
+	if ((notify_flags & IB_CQ_REPORT_MISSED_EVENTS) &&
+	    cq->queue->head != cq->queue->tail)
+		ret = 1;
+
+	spin_unlock_irqrestore(&cq->lock, flags);
+
+	return ret;
 }
 
 /**
@@ -93,7 +349,107 @@ 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)
 {
-	return -EOPNOTSUPP;
+	struct rvt_cq *cq = ibcq_to_rvtcq(ibcq);
+	struct rvt_cq_wc *old_wc;
+	struct rvt_cq_wc *wc;
+	u32 head, tail, n;
+	int ret;
+	u32 sz;
+	struct rvt_dev_info *rdi = cq->rdi;
+
+	if (cqe < 1 || cqe > rdi->dparms.props.max_cqe)
+		return -EINVAL;
+
+	/*
+	 * Need to use vmalloc() if we want to support large #s of entries.
+	 */
+	sz = sizeof(*wc);
+	if (udata && udata->outlen >= sizeof(__u64))
+		sz += sizeof(struct ib_uverbs_wc) * (cqe + 1);
+	else
+		sz += sizeof(struct ib_wc) * (cqe + 1);
+	wc = vmalloc_user(sz);
+	if (!wc)
+		return -ENOMEM;
+
+	/* Check that we can write the offset to mmap. */
+	if (udata && udata->outlen >= sizeof(__u64)) {
+		__u64 offset = 0;
+
+		ret = ib_copy_to_udata(udata, &offset, sizeof(offset));
+		if (ret)
+			goto bail_free;
+	}
+
+	spin_lock_irq(&cq->lock);
+	/*
+	 * Make sure head and tail are sane since they
+	 * might be user writable.
+	 */
+	old_wc = cq->queue;
+	head = old_wc->head;
+	if (head > (u32)cq->ibcq.cqe)
+		head = (u32)cq->ibcq.cqe;
+	tail = old_wc->tail;
+	if (tail > (u32)cq->ibcq.cqe)
+		tail = (u32)cq->ibcq.cqe;
+	if (head < tail)
+		n = cq->ibcq.cqe + 1 + head - tail;
+	else
+		n = head - tail;
+	if (unlikely((u32)cqe < n)) {
+		ret = -EINVAL;
+		goto bail_unlock;
+	}
+	for (n = 0; tail != head; n++) {
+		if (cq->ip)
+			wc->uqueue[n] = old_wc->uqueue[tail];
+		else
+			wc->kqueue[n] = old_wc->kqueue[tail];
+		if (tail == (u32)cq->ibcq.cqe)
+			tail = 0;
+		else
+			tail++;
+	}
+	cq->ibcq.cqe = cqe;
+	wc->head = n;
+	wc->tail = 0;
+	cq->queue = wc;
+	spin_unlock_irq(&cq->lock);
+
+	vfree(old_wc);
+
+	if (cq->ip) {
+		struct rvt_mmap_info *ip = cq->ip;
+
+		rvt_update_mmap_info(rdi, ip, sz, wc);
+
+		/*
+		 * Return the offset to mmap.
+		 * See rvt_mmap() for details.
+		 */
+		if (udata && udata->outlen >= sizeof(__u64)) {
+			ret = ib_copy_to_udata(udata, &ip->offset,
+					       sizeof(ip->offset));
+			if (ret)
+				goto bail;
+		}
+
+		spin_lock_irq(&rdi->pending_lock);
+		if (list_empty(&ip->pending_mmaps))
+			list_add(&ip->pending_mmaps, &rdi->pending_mmaps);
+		spin_unlock_irq(&rdi->pending_lock);
+	}
+
+	return 0;
+
+bail_unlock:
+	spin_unlock_irq(&cq->lock);
+bail_free:
+	vfree(wc);
+bail:
+	return ret;
+
 }
 
 /**
@@ -109,5 +465,84 @@ 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)
 {
-	return -EOPNOTSUPP;
+	struct rvt_cq *cq = ibcq_to_rvtcq(ibcq);
+	struct rvt_cq_wc *wc;
+	unsigned long flags;
+	int npolled;
+	u32 tail;
+
+	/* The kernel can only poll a kernel completion queue */
+	if (cq->ip)
+		return -EINVAL;
+
+	spin_lock_irqsave(&cq->lock, flags);
+
+	wc = cq->queue;
+	tail = wc->tail;
+	if (tail > (u32)cq->ibcq.cqe)
+		tail = (u32)cq->ibcq.cqe;
+	for (npolled = 0; npolled < num_entries; ++npolled, ++entry) {
+		if (tail == wc->head)
+			break;
+		/* The kernel doesn't need a RMB since it has the lock. */
+		*entry = wc->kqueue[tail];
+		if (tail >= cq->ibcq.cqe)
+			tail = 0;
+		else
+			tail++;
+	}
+	wc->tail = tail;
+
+	spin_unlock_irqrestore(&cq->lock, flags);
+
+	return npolled;
+}
+
+int rvt_driver_cq_init(struct rvt_dev_info *rdi)
+{
+	int ret = 0;
+	int cpu;
+	struct task_struct *task;
+
+	if (rdi->flags & RVT_FLAG_CQ_INIT_DRIVER) {
+		rvt_pr_info(rdi, "Driver is doing CQ init.\n");
+		return 0;
+	}
+
+	if (rdi->worker)
+		return 0;
+	rdi->worker = kzalloc(sizeof(*rdi->worker), GFP_KERNEL);
+	if (!rdi->worker)
+		return -ENOMEM;
+	init_kthread_worker(rdi->worker);
+	task = kthread_create_on_node(
+		kthread_worker_fn,
+		rdi->worker,
+		rdi->dparms.node,
+		"%s", rdi->dparms.cq_name);
+	if (IS_ERR(task)) {
+		kfree(rdi->worker);
+		rdi->worker = NULL;
+		return PTR_ERR(task);
+	}
+
+	cpu = cpumask_first(cpumask_of_node(rdi->dparms.node));
+	kthread_bind(task, cpu);
+	wake_up_process(task);
+	return ret;
+}
+
+void rvt_cq_exit(struct rvt_dev_info *rdi)
+{
+	struct kthread_worker *worker;
+
+	worker = rdi->worker;
+	if (!worker)
+		return;
+	/* blocks future queuing from send_complete() */
+	rdi->worker = NULL;
+	smp_wmb(); /* See rdi_cq_enter */
+	flush_kthread_worker(worker);
+	kthread_stop(worker->task);
+	kfree(worker);
 }
diff --git a/drivers/infiniband/sw/rdmavt/cq.h b/drivers/infiniband/sw/rdmavt/cq.h
index 63a517d..3813d90 100644
--- a/drivers/infiniband/sw/rdmavt/cq.h
+++ b/drivers/infiniband/sw/rdmavt/cq.h
@@ -49,6 +49,7 @@
  */
 
 #include <rdma/rdma_vt.h>
+#include <rdma/rdmavt_cq.h>
 
 struct ib_cq *rvt_create_cq(struct ib_device *ibdev,
 			    const struct ib_cq_init_attr *attr,
@@ -58,5 +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);
 #endif          /* DEF_RVTCQ_H */
diff --git a/drivers/infiniband/sw/rdmavt/vt.c b/drivers/infiniband/sw/rdmavt/vt.c
index f2b6438..136cc21 100644
--- a/drivers/infiniband/sw/rdmavt/vt.c
+++ b/drivers/infiniband/sw/rdmavt/vt.c
@@ -324,6 +324,11 @@ int rvt_register_device(struct rvt_dev_info *rdi)
 	CHECK_DRIVER_OVERRIDE(rdi, mmap);
 
 	/* Completion queues */
+	ret = rvt_driver_cq_init(rdi);
+	if (ret) {
+		pr_err("Error in driver CQ init.\n");
+		goto bail_mr;
+	}
 	CHECK_DRIVER_OVERRIDE(rdi, create_cq);
 	CHECK_DRIVER_OVERRIDE(rdi, destroy_cq);
 	CHECK_DRIVER_OVERRIDE(rdi, poll_cq);
@@ -344,12 +349,15 @@ int rvt_register_device(struct rvt_dev_info *rdi)
 	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_mr;
+		goto bail_cq;
 	}
 
 	rvt_pr_info(rdi, "Registration with rdmavt done.\n");
 	return ret;
 
+bail_cq:
+	rvt_cq_exit(rdi);
+
 bail_mr:
 	rvt_mr_exit(rdi);
 
@@ -366,6 +374,7 @@ void rvt_unregister_device(struct rvt_dev_info *rdi)
 		return;
 
 	ib_unregister_device(&rdi->ibdev);
+	rvt_cq_exit(rdi);
 	rvt_mr_exit(rdi);
 	rvt_qp_exit(rdi);
 }
diff --git a/include/rdma/rdma_vt.h b/include/rdma/rdma_vt.h
index 725778a..f2d50b9 100644
--- a/include/rdma/rdma_vt.h
+++ b/include/rdma/rdma_vt.h
@@ -138,6 +138,8 @@ struct rvt_ibport {
 	/* TODO: Move sm_ah and smi_ah into here as well*/
 };
 
+#define RVT_CQN_MAX 16 /* maximum length of cq name */
+
 /*
  * Things that are driver specific, module parameters in hfi1 and qib
  */
@@ -190,6 +192,8 @@ struct rvt_driver_params {
 	int nports;
 	int npkeys;
 	u8 qos_shift;
+	char cq_name[RVT_CQN_MAX];
+	int node;
 };
 
 /* Protection domain */
@@ -280,6 +284,11 @@ struct rvt_dev_info {
 	spinlock_t mmap_offset_lock; /* protect mmap_offset */
 	u32 mmap_offset;
 	spinlock_t pending_lock; /* protect pending mmap list */
+
+	/* CQ */
+	struct kthread_worker *worker; /* per device cq worker */
+	u32 n_cqs_allocated;    /* number of CQs allocated for device */
+	spinlock_t n_cqs_lock; /* protect count of in use cqs */
 };
 
 static inline struct rvt_pd *ibpd_to_rvtpd(struct ib_pd *ibpd)
diff --git a/include/rdma/rdmavt_cq.h b/include/rdma/rdmavt_cq.h
new file mode 100644
index 0000000..51fd00b
--- /dev/null
+++ b/include/rdma/rdmavt_cq.h
@@ -0,0 +1,99 @@
+#ifndef DEF_RDMAVT_INCCQ_H
+#define DEF_RDMAVT_INCCQ_H
+
+/*
+ *
+ * This file is provided under a dual BSD/GPLv2 license.  When using or
+ * redistributing this file, you may do so under either license.
+ *
+ * GPL LICENSE SUMMARY
+ *
+ * Copyright(c) 2016 Intel Corporation.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of version 2 of the GNU General Public License as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * BSD LICENSE
+ *
+ * Copyright(c) 2015 Intel Corporation.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *
+ *  - Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ *  - Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in
+ *    the documentation and/or other materials provided with the
+ *    distribution.
+ *  - Neither the name of Intel Corporation nor the names of its
+ *    contributors may be used to endorse or promote products derived
+ *    from this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ */
+
+#include <linux/kthread.h>
+#include <rdma/ib_user_verbs.h>
+
+/*
+ * Define an ib_cq_notify value that is not valid so we know when CQ
+ * notifications are armed.
+ */
+#define RVT_CQ_NONE      (IB_CQ_NEXT_COMP + 1)
+
+/*
+ * This structure is used to contain the head pointer, tail pointer,
+ * and completion queue entries as a single memory allocation so
+ * it can be mmap'ed into user space.
+ */
+struct rvt_cq_wc {
+	u32 head;               /* index of next entry to fill */
+	u32 tail;               /* index of next ib_poll_cq() entry */
+	union {
+		/* these are actually size ibcq.cqe + 1 */
+		struct ib_uverbs_wc uqueue[0];
+		struct ib_wc kqueue[0];
+	};
+};
+
+/*
+ * The completion queue structure.
+ */
+struct rvt_cq {
+	struct ib_cq ibcq;
+	struct kthread_work comptask;
+	spinlock_t lock; /* protect changes in this struct */
+	u8 notify;
+	u8 triggered;
+	struct rvt_dev_info *rdi;
+	struct rvt_cq_wc *queue;
+	struct rvt_mmap_info *ip;
+};
+
+static inline struct rvt_cq *ibcq_to_rvtcq(struct ib_cq *ibcq)
+{
+	return container_of(ibcq, struct rvt_cq, ibcq);
+}
+
+void rvt_cq_enter(struct rvt_cq *cq, struct ib_wc *entry, bool solicited);
+
+#endif          /* DEF_RDMAVT_INCCQH */

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH 2/8] IB/rdmavt: Add post send to rdmavt

[Dennis Dalessandro]

Add in a post_send and post_one_send to rdmavt. The ULP will provide a WQE
to rdmavt which will then walk and queue each element. Rdmavt will then
queue the work to be done in the driver or kick the driver's progress
routine.

There needs to be a follow on patch which adds in another lock for the
head of the queue so that it can be added to and read from in parallel.
This will touch protocol handlers and require other changes in the
drivers. This will be done separately.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
Reviewed-by: Ira Weiny <ira.weiny-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
---
 drivers/infiniband/sw/rdmavt/qp.c |  183 +++++++++++++++++++++++++++++++++++--
 include/rdma/rdma_vt.h            |    7 +
 include/rdma/rdmavt_qp.h          |   26 +++++
 3 files changed, 204 insertions(+), 12 deletions(-)

diff --git a/drivers/infiniband/sw/rdmavt/qp.c b/drivers/infiniband/sw/rdmavt/qp.c
index ee19eae..5063e25 100644
--- a/drivers/infiniband/sw/rdmavt/qp.c
+++ b/drivers/infiniband/sw/rdmavt/qp.c
@@ -53,6 +53,27 @@
 #include "qp.h"
 #include "vt.h"
 
+/*
+ * Note that it is OK to post send work requests in the SQE and ERR
+ * states; rvt_do_send() will process them and generate error
+ * completions as per IB 1.2 C10-96.
+ */
+const int ib_rvt_state_ops[IB_QPS_ERR + 1] = {
+	[IB_QPS_RESET] = 0,
+	[IB_QPS_INIT] = RVT_POST_RECV_OK,
+	[IB_QPS_RTR] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK,
+	[IB_QPS_RTS] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK |
+	    RVT_POST_SEND_OK | RVT_PROCESS_SEND_OK |
+	    RVT_PROCESS_NEXT_SEND_OK,
+	[IB_QPS_SQD] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK |
+	    RVT_POST_SEND_OK | RVT_PROCESS_SEND_OK,
+	[IB_QPS_SQE] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK |
+	    RVT_POST_SEND_OK | RVT_FLUSH_SEND,
+	[IB_QPS_ERR] = RVT_POST_RECV_OK | RVT_FLUSH_RECV |
+	    RVT_POST_SEND_OK | RVT_FLUSH_SEND,
+};
+EXPORT_SYMBOL(ib_rvt_state_ops);
+
 static void get_map_page(struct rvt_qpn_table *qpt, struct rvt_qpn_map *map)
 {
 	unsigned long page = get_zeroed_page(GFP_KERNEL);
@@ -556,7 +577,7 @@ struct ib_qp *rvt_create_qp(struct ib_pd *ibpd,
 
 	/*
 	 * Return the address of the RWQ as the offset to mmap.
-	 * See hfi1_mmap() for details.
+	 * See rvt_mmap() for details.
 	 */
 	if (udata && udata->outlen >= sizeof(__u64)) {
 		if (!qp->r_rq.wq) {
@@ -720,6 +741,118 @@ int rvt_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
 }
 
 /**
+ * rvt_post_one_wr - post one RC, UC, or UD send work request
+ * @qp: the QP to post on
+ * @wr: the work request to send
+ */
+static int rvt_post_one_wr(struct rvt_qp *qp, struct ib_send_wr *wr)
+{
+	struct rvt_swqe *wqe;
+	u32 next;
+	int i;
+	int j;
+	int acc;
+	struct rvt_lkey_table *rkt;
+	struct rvt_pd *pd;
+	struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
+
+	/* IB spec says that num_sge == 0 is OK. */
+	if (unlikely(wr->num_sge > qp->s_max_sge))
+		return -EINVAL;
+
+	/*
+	 * Don't allow RDMA reads or atomic operations on UC or
+	 * undefined operations.
+	 * Make sure buffer is large enough to hold the result for atomics.
+	 */
+	if (qp->ibqp.qp_type == IB_QPT_UC) {
+		if ((unsigned)wr->opcode >= IB_WR_RDMA_READ)
+			return -EINVAL;
+	} else if (qp->ibqp.qp_type != IB_QPT_RC) {
+		/* Check IB_QPT_SMI, IB_QPT_GSI, IB_QPT_UD opcode */
+		if (wr->opcode != IB_WR_SEND &&
+		    wr->opcode != IB_WR_SEND_WITH_IMM)
+			return -EINVAL;
+		/* Check UD destination address PD */
+		if (qp->ibqp.pd != ud_wr(wr)->ah->pd)
+			return -EINVAL;
+	} else if ((unsigned)wr->opcode > IB_WR_ATOMIC_FETCH_AND_ADD) {
+		return -EINVAL;
+	} else if (wr->opcode >= IB_WR_ATOMIC_CMP_AND_SWP &&
+		   (wr->num_sge == 0 ||
+		    wr->sg_list[0].length < sizeof(u64) ||
+		    wr->sg_list[0].addr & (sizeof(u64) - 1))) {
+		return -EINVAL;
+	} else if (wr->opcode >= IB_WR_RDMA_READ && !qp->s_max_rd_atomic) {
+		return -EINVAL;
+	}
+
+	next = qp->s_head + 1;
+	if (next >= qp->s_size)
+		next = 0;
+	if (next == qp->s_last)
+		return -ENOMEM;
+
+	rkt = &rdi->lkey_table;
+	pd = ibpd_to_rvtpd(qp->ibqp.pd);
+	wqe = rvt_get_swqe_ptr(qp, qp->s_head);
+
+	if (qp->ibqp.qp_type != IB_QPT_UC &&
+	    qp->ibqp.qp_type != IB_QPT_RC)
+		memcpy(&wqe->ud_wr, ud_wr(wr), sizeof(wqe->ud_wr));
+	else if (wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM ||
+		 wr->opcode == IB_WR_RDMA_WRITE ||
+		 wr->opcode == IB_WR_RDMA_READ)
+		memcpy(&wqe->rdma_wr, rdma_wr(wr), sizeof(wqe->rdma_wr));
+	else if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
+		 wr->opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
+		memcpy(&wqe->atomic_wr, atomic_wr(wr), sizeof(wqe->atomic_wr));
+	else
+		memcpy(&wqe->wr, wr, sizeof(wqe->wr));
+
+	wqe->length = 0;
+	j = 0;
+	if (wr->num_sge) {
+		acc = wr->opcode >= IB_WR_RDMA_READ ?
+			IB_ACCESS_LOCAL_WRITE : 0;
+		for (i = 0; i < wr->num_sge; i++) {
+			u32 length = wr->sg_list[i].length;
+			int ok;
+
+			if (length == 0)
+				continue;
+			ok = rvt_lkey_ok(rkt, pd, &wqe->sg_list[j],
+					 &wr->sg_list[i], acc);
+			if (!ok)
+				goto bail_inval_free;
+			wqe->length += length;
+			j++;
+		}
+		wqe->wr.num_sge = j;
+	}
+	if (qp->ibqp.qp_type == IB_QPT_UC ||
+	    qp->ibqp.qp_type == IB_QPT_RC) {
+		if (wqe->length > 0x80000000U)
+			goto bail_inval_free;
+	} else {
+		atomic_inc(&ibah_to_rvtah(ud_wr(wr)->ah)->refcount);
+	}
+	wqe->ssn = qp->s_ssn++;
+	qp->s_head = next;
+
+	return 0;
+
+bail_inval_free:
+	/* release mr holds */
+	while (j) {
+		struct rvt_sge *sge = &wqe->sg_list[--j];
+
+		rvt_put_mr(sge->mr);
+	}
+	return -EINVAL;
+}
+
+/**
  * rvt_post_send - post a send on a QP
  * @ibqp: the QP to post the send on
  * @wr: the list of work requests to post
@@ -730,20 +863,46 @@ int rvt_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
 int rvt_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
 		  struct ib_send_wr **bad_wr)
 {
+	struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
+	struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
+	unsigned long flags = 0;
+	int call_send;
+	unsigned nreq = 0;
+	int err = 0;
+
+	spin_lock_irqsave(&qp->s_lock, flags);
+
 	/*
-	 * VT-DRIVER-API: do_send()
-	 * Driver needs to have a do_send() call which is a single entry point
-	 * to take an already formed packet and throw it out on the wire. Once
-	 * the packet is sent the driver needs to make an upcall to rvt so the
-	 * completion queue can be notified and/or any other outstanding
-	 * work/book keeping can be finished.
-	 *
-	 * Note that there should also be a way for rvt to protect itself
-	 * against hangs in the driver layer. If a send doesn't actually
-	 * complete in a timely manor rvt needs to return an error event.
+	 * Ensure QP state is such that we can send. If not bail out early,
+	 * there is no need to do this every time we post a send.
 	 */
+	if (unlikely(!(ib_rvt_state_ops[qp->state] & RVT_POST_SEND_OK))) {
+		spin_unlock_irqrestore(&qp->s_lock, flags);
+		return -EINVAL;
+	}
 
-	return -EOPNOTSUPP;
+	/*
+	 * If the send queue is empty, and we only have a single WR then just go
+	 * ahead and kick the send engine into gear. Otherwise we will always
+	 * just schedule the send to happen later.
+	 */
+	call_send = qp->s_head == ACCESS_ONCE(qp->s_last) && !wr->next;
+
+	for (; wr; wr = wr->next) {
+		err = rvt_post_one_wr(qp, wr);
+		if (unlikely(err)) {
+			*bad_wr = wr;
+			goto bail;
+		}
+		nreq++;
+	}
+bail:
+	if (nreq && !call_send)
+		rdi->driver_f.schedule_send(qp);
+	spin_unlock_irqrestore(&qp->s_lock, flags);
+	if (nreq && call_send)
+		rdi->driver_f.do_send(qp);
+	return err;
 }
 
 /**
diff --git a/include/rdma/rdma_vt.h b/include/rdma/rdma_vt.h
index f2d50b9..3d92239 100644
--- a/include/rdma/rdma_vt.h
+++ b/include/rdma/rdma_vt.h
@@ -230,6 +230,8 @@ struct rvt_driver_provided {
 	void * (*qp_priv_alloc)(struct rvt_dev_info *rdi, struct rvt_qp *qp);
 	void (*qp_priv_free)(struct rvt_dev_info *rdi, struct rvt_qp *qp);
 	void (*notify_qp_reset)(struct rvt_qp *qp);
+	void (*schedule_send)(struct rvt_qp *qp);
+	void (*do_send)(struct rvt_qp *qp);
 
 	/*--------------------*/
 	/* Optional functions */
@@ -311,6 +313,11 @@ static inline struct rvt_srq *ibsrq_to_rvtsrq(struct ib_srq *ibsrq)
 	return container_of(ibsrq, struct rvt_srq, ibsrq);
 }
 
+static inline struct rvt_qp *ibqp_to_rvtqp(struct ib_qp *ibqp)
+{
+	return container_of(ibqp, struct rvt_qp, ibqp);
+}
+
 static inline unsigned rvt_get_npkeys(struct rvt_dev_info *rdi)
 {
 	/*
diff --git a/include/rdma/rdmavt_qp.h b/include/rdma/rdmavt_qp.h
index bce0a03..3189f19 100644
--- a/include/rdma/rdmavt_qp.h
+++ b/include/rdma/rdmavt_qp.h
@@ -129,6 +129,17 @@
 /* Number of bits to pay attention to in the opcode for checking qp type */
 #define RVT_OPCODE_QP_MASK 0xE0
 
+/* Flags for checking QP state (see ib_rvt_state_ops[]) */
+#define RVT_POST_SEND_OK                0x01
+#define RVT_POST_RECV_OK                0x02
+#define RVT_PROCESS_RECV_OK             0x04
+#define RVT_PROCESS_SEND_OK             0x08
+#define RVT_PROCESS_NEXT_SEND_OK        0x10
+#define RVT_FLUSH_SEND			0x20
+#define RVT_FLUSH_RECV			0x40
+#define RVT_PROCESS_OR_FLUSH_SEND \
+	(RVT_PROCESS_SEND_OK | RVT_FLUSH_SEND)
+
 /*
  * Send work request queue entry.
  * The size of the sg_list is determined when the QP is created and stored
@@ -373,4 +384,19 @@ struct rvt_qp_ibdev {
 	struct rvt_qpn_table qpn_table;
 };
 
+/*
+ * Since struct rvt_swqe is not a fixed size, we can't simply index into
+ * struct hfi1_qp.s_wq.  This function does the array index computation.
+ */
+static inline struct rvt_swqe *rvt_get_swqe_ptr(struct rvt_qp *qp,
+						unsigned n)
+{
+	return (struct rvt_swqe *)((char *)qp->s_wq +
+				     (sizeof(struct rvt_swqe) +
+				      qp->s_max_sge *
+				      sizeof(struct rvt_sge)) * n);
+}
+
+extern const int  ib_rvt_state_ops[];
+
 #endif          /* DEF_RDMAVT_INCQP_H */

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH 3/8] IB/rdmavt: Add support for tracing events

[Dennis Dalessandro]

This patch adds support of tracing events using the kernels built-in event
tracing infrastructure. This can be extended to provide a wide range of
trace and debug capabilities which have a negligible impact on performance
when enabled. These should be preferred over the use of the rvt_pr*
functions.

Reviewed-by: Ira Weiny <ira.weiny-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
---
 drivers/infiniband/sw/rdmavt/Makefile |    5 ++
 drivers/infiniband/sw/rdmavt/trace.c  |   49 +++++++++++++++++++
 drivers/infiniband/sw/rdmavt/trace.h  |   86 +++++++++++++++++++++++++++++++++
 drivers/infiniband/sw/rdmavt/vt.c     |    5 ++
 4 files changed, 143 insertions(+), 2 deletions(-)
 create mode 100644 drivers/infiniband/sw/rdmavt/trace.c
 create mode 100644 drivers/infiniband/sw/rdmavt/trace.h

diff --git a/drivers/infiniband/sw/rdmavt/Makefile b/drivers/infiniband/sw/rdmavt/Makefile
index 00f0188..ccaa799 100644
--- a/drivers/infiniband/sw/rdmavt/Makefile
+++ b/drivers/infiniband/sw/rdmavt/Makefile
@@ -7,4 +7,7 @@
 #
 obj-$(CONFIG_INFINIBAND_RDMAVT) += rdmavt.o
 
-rdmavt-y := vt.o ah.o cq.o dma.o mad.o mcast.o mmap.o mr.o pd.o qp.o srq.o
+rdmavt-y := vt.o ah.o cq.o dma.o mad.o mcast.o mmap.o mr.o pd.o qp.o srq.o \
+	trace.o
+
+CFLAGS_trace.o = -I$(src)
diff --git a/drivers/infiniband/sw/rdmavt/trace.c b/drivers/infiniband/sw/rdmavt/trace.c
new file mode 100644
index 0000000..19afe39
--- /dev/null
+++ b/drivers/infiniband/sw/rdmavt/trace.c
@@ -0,0 +1,49 @@
+/*
+ * Copyright(c) 2015 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.
+ *
+ * GPL LICENSE SUMMARY
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of version 2 of the GNU General Public License as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * BSD LICENSE
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *
+ *  - Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ *  - Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in
+ *    the documentation and/or other materials provided with the
+ *    distribution.
+ *  - Neither the name of Intel Corporation nor the names of its
+ *    contributors may be used to endorse or promote products derived
+ *    from this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ */
+
+#define CREATE_TRACE_POINTS
+#include "trace.h"
diff --git a/drivers/infiniband/sw/rdmavt/trace.h b/drivers/infiniband/sw/rdmavt/trace.h
new file mode 100644
index 0000000..22e86ff
--- /dev/null
+++ b/drivers/infiniband/sw/rdmavt/trace.h
@@ -0,0 +1,86 @@
+/*
+ * Copyright(c) 2015 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.
+ *
+ * GPL LICENSE SUMMARY
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of version 2 of the GNU General Public License as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * BSD LICENSE
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *
+ *  - Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ *  - Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in
+ *    the documentation and/or other materials provided with the
+ *    distribution.
+ *  - Neither the name of Intel Corporation nor the names of its
+ *    contributors may be used to endorse or promote products derived
+ *    from this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ */
+
+#undef TRACE_SYSTEM_VAR
+#define TRACE_SYSTEM_VAR rdmavt
+
+#if !defined(__RDMAVT_TRACE_H) || defined(TRACE_HEADER_MULTI_READ)
+#define __RDMAVT_TRACE_H
+
+#include <linux/tracepoint.h>
+#include <linux/trace_seq.h>
+
+#include <rdma/rdma_vt.h>
+
+#define RDI_DEV_ENTRY(rdi)   __string(dev, rdi->driver_f.get_card_name(rdi))
+#define RDI_DEV_ASSIGN(rdi)  __assign_str(dev, rdi->driver_f.get_card_name(rdi))
+
+#undef TRACE_SYSTEM
+#define TRACE_SYSTEM rdmavt
+
+TRACE_EVENT(rvt_dbg,
+	TP_PROTO(struct rvt_dev_info *rdi,
+		 const char *msg),
+	TP_ARGS(rdi, msg),
+	TP_STRUCT__entry(
+		RDI_DEV_ENTRY(rdi)
+		__string(msg, msg)
+	),
+	TP_fast_assign(
+		RDI_DEV_ASSIGN(rdi);
+		__assign_str(msg, msg);
+	),
+	TP_printk("[%s]: %s", __get_str(dev), __get_str(msg))
+);
+
+#endif /* __RDMAVT_TRACE_H */
+
+#undef TRACE_INCLUDE_PATH
+#undef TRACE_INCLUDE_FILE
+#define TRACE_INCLUDE_PATH .
+#define TRACE_INCLUDE_FILE trace
+#include <trace/define_trace.h>
diff --git a/drivers/infiniband/sw/rdmavt/vt.c b/drivers/infiniband/sw/rdmavt/vt.c
index 136cc21..7f56a42 100644
--- a/drivers/infiniband/sw/rdmavt/vt.c
+++ b/drivers/infiniband/sw/rdmavt/vt.c
@@ -48,6 +48,7 @@
 #include <linux/module.h>
 #include <linux/kernel.h>
 #include "vt.h"
+#include "trace.h"
 
 MODULE_LICENSE("Dual BSD/GPL");
 MODULE_DESCRIPTION("RDMA Verbs Transport Library");
@@ -259,7 +260,8 @@ int rvt_register_device(struct rvt_dev_info *rdi)
 		return -EINVAL;
 	}
 
-	/* Once we get past here we can use the rvt_pr macros */
+	/* Once we get past here we can use rvt_pr macros and tracepoints */
+	trace_rvt_dbg(rdi, "Driver attempting registration");
 	rvt_mmap_init(rdi);
 
 	/* Dev Ops */
@@ -370,6 +372,7 @@ EXPORT_SYMBOL(rvt_register_device);
 
 void rvt_unregister_device(struct rvt_dev_info *rdi)
 {
+	trace_rvt_dbg(rdi, "Driver is unregistering.");
 	if (!rdi)
 		return;
 

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH 4/8] IB/rdmavt: Add modify qp

[Dennis Dalessandro]

Add modify qp and supporting functions.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
Reviewed-by: Ira Weiny <ira.weiny-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
---
 drivers/infiniband/sw/rdmavt/qp.c    |  503 +++++++++++++++++++++++++++++++++-
 drivers/infiniband/sw/rdmavt/trace.h |   31 ++
 include/rdma/rdma_vt.h               |   42 +++
 include/rdma/rdmavt_mr.h             |    9 +
 include/rdma/rdmavt_qp.h             |   20 +
 5 files changed, 596 insertions(+), 9 deletions(-)

diff --git a/drivers/infiniband/sw/rdmavt/qp.c b/drivers/infiniband/sw/rdmavt/qp.c
index 5063e25..b8c7c49 100644
--- a/drivers/infiniband/sw/rdmavt/qp.c
+++ b/drivers/infiniband/sw/rdmavt/qp.c
@@ -45,6 +45,7 @@
  *
  */
 
+#include <linux/hash.h>
 #include <linux/bitops.h>
 #include <linux/lockdep.h>
 #include <linux/vmalloc.h>
@@ -52,6 +53,7 @@
 #include <rdma/ib_verbs.h>
 #include "qp.h"
 #include "vt.h"
+#include "trace.h"
 
 /*
  * Note that it is OK to post send work requests in the SQE and ERR
@@ -377,19 +379,47 @@ static void free_qpn(struct rvt_qpn_table *qpt, u32 qpn)
  * reset_qp - initialize the QP state to the reset state
  * @qp: the QP to reset
  * @type: the QP type
+ * r and s lock are required to be held by the caller
  */
 void rvt_reset_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp,
 		  enum ib_qp_type type)
 {
-	qp->remote_qpn = 0;
-	qp->qkey = 0;
-	qp->qp_access_flags = 0;
+	if (qp->state != IB_QPS_RESET) {
+		qp->state = IB_QPS_RESET;
+
+		/* Let drivers flush their waitlist */
+		rdi->driver_f.flush_qp_waiters(qp);
+		qp->s_flags &= ~(RVT_S_TIMER | RVT_S_ANY_WAIT);
+		spin_unlock(&qp->s_lock);
+		spin_unlock_irq(&qp->r_lock);
+
+		/* Stop the send queue and the retry timer */
+		rdi->driver_f.stop_send_queue(qp);
+		del_timer_sync(&qp->s_timer);
+
+		/* Wait for things to stop */
+		rdi->driver_f.quiesce_qp(qp);
+
+		/* take qp out the hash and wait for it to be unused */
+		rvt_remove_qp(rdi, qp);
+		wait_event(qp->wait, !atomic_read(&qp->refcount));
+
+		/* grab the lock b/c it was locked at call time */
+		spin_lock_irq(&qp->r_lock);
+		spin_lock(&qp->s_lock);
+
+		rvt_clear_mr_refs(qp, 1);
+	}
 
 	/*
-	 * Let driver do anything it needs to for a new/reset qp
+	 * Let the driver do any tear down it needs to for a qp
+	 * that has been reset
 	 */
 	rdi->driver_f.notify_qp_reset(qp);
 
+	qp->remote_qpn = 0;
+	qp->qkey = 0;
+	qp->qp_access_flags = 0;
 	qp->s_flags &= RVT_S_SIGNAL_REQ_WR;
 	qp->s_hdrwords = 0;
 	qp->s_wqe = NULL;
@@ -672,6 +702,208 @@ bail_swq:
 	return ret;
 }
 
+void rvt_clear_mr_refs(struct rvt_qp *qp, int clr_sends)
+{
+	unsigned n;
+
+	if (test_and_clear_bit(RVT_R_REWIND_SGE, &qp->r_aflags))
+		rvt_put_ss(&qp->s_rdma_read_sge);
+
+	rvt_put_ss(&qp->r_sge);
+
+	if (clr_sends) {
+		while (qp->s_last != qp->s_head) {
+			struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, qp->s_last);
+			unsigned i;
+
+			for (i = 0; i < wqe->wr.num_sge; i++) {
+				struct rvt_sge *sge = &wqe->sg_list[i];
+
+				rvt_put_mr(sge->mr);
+			}
+			if (qp->ibqp.qp_type == IB_QPT_UD ||
+			    qp->ibqp.qp_type == IB_QPT_SMI ||
+			    qp->ibqp.qp_type == IB_QPT_GSI)
+				atomic_dec(&ibah_to_rvtah(
+						wqe->ud_wr.ah)->refcount);
+			if (++qp->s_last >= qp->s_size)
+				qp->s_last = 0;
+		}
+		if (qp->s_rdma_mr) {
+			rvt_put_mr(qp->s_rdma_mr);
+			qp->s_rdma_mr = NULL;
+		}
+	}
+
+	if (qp->ibqp.qp_type != IB_QPT_RC)
+		return;
+
+	for (n = 0; n < ARRAY_SIZE(qp->s_ack_queue); n++) {
+		struct rvt_ack_entry *e = &qp->s_ack_queue[n];
+
+		if (e->opcode == IB_OPCODE_RC_RDMA_READ_REQUEST &&
+		    e->rdma_sge.mr) {
+			rvt_put_mr(e->rdma_sge.mr);
+			e->rdma_sge.mr = NULL;
+		}
+	}
+}
+EXPORT_SYMBOL(rvt_clear_mr_refs);
+
+/**
+ * rvt_error_qp - put a QP into the error state
+ * @qp: the QP to put into the error state
+ * @err: the receive completion error to signal if a RWQE is active
+ *
+ * Flushes both send and receive work queues.
+ * Returns true if last WQE event should be generated.
+ * The QP r_lock and s_lock should be held and interrupts disabled.
+ * If we are already in error state, just return.
+ */
+int rvt_error_qp(struct rvt_qp *qp, enum ib_wc_status err)
+{
+	struct ib_wc wc;
+	int ret = 0;
+	struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
+
+	if (qp->state == IB_QPS_ERR || qp->state == IB_QPS_RESET)
+		goto bail;
+
+	qp->state = IB_QPS_ERR;
+
+	if (qp->s_flags & (RVT_S_TIMER | RVT_S_WAIT_RNR)) {
+		qp->s_flags &= ~(RVT_S_TIMER | RVT_S_WAIT_RNR);
+		del_timer(&qp->s_timer);
+	}
+
+	if (qp->s_flags & RVT_S_ANY_WAIT_SEND)
+		qp->s_flags &= ~RVT_S_ANY_WAIT_SEND;
+
+	rdi->driver_f.notify_error_qp(qp);
+
+	/* Schedule the sending tasklet to drain the send work queue. */
+	if (qp->s_last != qp->s_head)
+		rdi->driver_f.schedule_send(qp);
+
+	rvt_clear_mr_refs(qp, 0);
+
+	memset(&wc, 0, sizeof(wc));
+	wc.qp = &qp->ibqp;
+	wc.opcode = IB_WC_RECV;
+
+	if (test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags)) {
+		wc.wr_id = qp->r_wr_id;
+		wc.status = err;
+		rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1);
+	}
+	wc.status = IB_WC_WR_FLUSH_ERR;
+
+	if (qp->r_rq.wq) {
+		struct rvt_rwq *wq;
+		u32 head;
+		u32 tail;
+
+		spin_lock(&qp->r_rq.lock);
+
+		/* sanity check pointers before trusting them */
+		wq = qp->r_rq.wq;
+		head = wq->head;
+		if (head >= qp->r_rq.size)
+			head = 0;
+		tail = wq->tail;
+		if (tail >= qp->r_rq.size)
+			tail = 0;
+		while (tail != head) {
+			wc.wr_id = rvt_get_rwqe_ptr(&qp->r_rq, tail)->wr_id;
+			if (++tail >= qp->r_rq.size)
+				tail = 0;
+			rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1);
+		}
+		wq->tail = tail;
+
+		spin_unlock(&qp->r_rq.lock);
+	} else if (qp->ibqp.event_handler) {
+		ret = 1;
+	}
+
+bail:
+	return ret;
+}
+EXPORT_SYMBOL(rvt_error_qp);
+
+/*
+ * Put the QP into the hash table.
+ * The hash table holds a reference to the QP.
+ */
+static void rvt_insert_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp)
+{
+	struct rvt_ibport *rvp = rdi->ports[qp->port_num - 1];
+	unsigned long flags;
+
+	atomic_inc(&qp->refcount);
+	spin_lock_irqsave(&rdi->qp_dev->qpt_lock, flags);
+
+	if (qp->ibqp.qp_num <= 1) {
+		rcu_assign_pointer(rvp->qp[qp->ibqp.qp_num], qp);
+	} else {
+		u32 n = hash_32(qp->ibqp.qp_num, rdi->qp_dev->qp_table_bits);
+
+		qp->next = rdi->qp_dev->qp_table[n];
+		rcu_assign_pointer(rdi->qp_dev->qp_table[n], qp);
+		trace_rvt_qpinsert(qp, n);
+	}
+
+	spin_unlock_irqrestore(&rdi->qp_dev->qpt_lock, flags);
+}
+
+/*
+ * Remove the QP from the table so it can't be found asynchronously by
+ * the receive routine.
+ */
+void rvt_remove_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp)
+{
+	struct rvt_ibport *rvp = rdi->ports[qp->port_num - 1];
+	u32 n = hash_32(qp->ibqp.qp_num, rdi->qp_dev->qp_table_bits);
+	unsigned long flags;
+	int removed = 1;
+
+	spin_lock_irqsave(&rdi->qp_dev->qpt_lock, flags);
+
+	if (rcu_dereference_protected(rvp->qp[0],
+			lockdep_is_held(&rdi->qp_dev->qpt_lock)) == qp) {
+		RCU_INIT_POINTER(rvp->qp[0], NULL);
+	} else if (rcu_dereference_protected(rvp->qp[1],
+			lockdep_is_held(&rdi->qp_dev->qpt_lock)) == qp) {
+		RCU_INIT_POINTER(rvp->qp[1], NULL);
+	} else {
+		struct rvt_qp *q;
+		struct rvt_qp __rcu **qpp;
+
+		removed = 0;
+		qpp = &rdi->qp_dev->qp_table[n];
+		for (; (q = rcu_dereference_protected(*qpp,
+			lockdep_is_held(&rdi->qp_dev->qpt_lock))) != NULL;
+			qpp = &q->next) {
+			if (q == qp) {
+				RCU_INIT_POINTER(*qpp,
+				     rcu_dereference_protected(qp->next,
+				     lockdep_is_held(&rdi->qp_dev->qpt_lock)));
+				removed = 1;
+				trace_rvt_qpremove(qp, n);
+				break;
+			}
+		}
+	}
+
+	spin_unlock_irqrestore(&rdi->qp_dev->qpt_lock, flags);
+	if (removed) {
+		synchronize_rcu();
+		if (atomic_dec_and_test(&qp->refcount))
+			wake_up(&qp->wait);
+	}
+}
+EXPORT_SYMBOL(rvt_remove_qp);
+
 /**
  * qib_modify_qp - modify the attributes of a queue pair
  * @ibqp: the queue pair who's attributes we're modifying
@@ -684,13 +916,248 @@ bail_swq:
 int rvt_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 		  int attr_mask, struct ib_udata *udata)
 {
+	struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
+	struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
+	enum ib_qp_state cur_state, new_state;
+	struct ib_event ev;
+	int lastwqe = 0;
+	int mig = 0;
+	int pmtu = 0; /* for gcc warning only */
+	enum rdma_link_layer link;
+
+	link = rdma_port_get_link_layer(ibqp->device, qp->port_num);
+
+	spin_lock_irq(&qp->r_lock);
+	spin_lock(&qp->s_lock);
+
+	cur_state = attr_mask & IB_QP_CUR_STATE ?
+		attr->cur_qp_state : qp->state;
+	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
+
+	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
+				attr_mask, link))
+		goto inval;
+
+	if (attr_mask & IB_QP_AV) {
+		if (attr->ah_attr.dlid >= be16_to_cpu(IB_MULTICAST_LID_BASE))
+			goto inval;
+		if (rvt_check_ah(qp->ibqp.device, &attr->ah_attr))
+			goto inval;
+	}
+
+	if (attr_mask & IB_QP_ALT_PATH) {
+		if (attr->alt_ah_attr.dlid >=
+		    be16_to_cpu(IB_MULTICAST_LID_BASE))
+			goto inval;
+		if (rvt_check_ah(qp->ibqp.device, &attr->alt_ah_attr))
+			goto inval;
+		if (attr->alt_pkey_index >= rvt_get_npkeys(rdi))
+			goto inval;
+	}
+
+	if (attr_mask & IB_QP_PKEY_INDEX)
+		if (attr->pkey_index >= rvt_get_npkeys(rdi))
+			goto inval;
+
+	if (attr_mask & IB_QP_MIN_RNR_TIMER)
+		if (attr->min_rnr_timer > 31)
+			goto inval;
+
+	if (attr_mask & IB_QP_PORT)
+		if (qp->ibqp.qp_type == IB_QPT_SMI ||
+		    qp->ibqp.qp_type == IB_QPT_GSI ||
+		    attr->port_num == 0 ||
+		    attr->port_num > ibqp->device->phys_port_cnt)
+			goto inval;
+
+	if (attr_mask & IB_QP_DEST_QPN)
+		if (attr->dest_qp_num > RVT_QPN_MASK)
+			goto inval;
+
+	if (attr_mask & IB_QP_RETRY_CNT)
+		if (attr->retry_cnt > 7)
+			goto inval;
+
+	if (attr_mask & IB_QP_RNR_RETRY)
+		if (attr->rnr_retry > 7)
+			goto inval;
+
 	/*
-	 * VT-DRIVER-API: qp_mtu()
-	 * OPA devices have a per VL MTU the driver has a mapping of IB SL to SC
-	 * to VL and the mapping table of MTUs per VL. This is not something
-	 * that IB has and should not live in the rvt.
+	 * Don't allow invalid path_mtu values.  OK to set greater
+	 * than the active mtu (or even the max_cap, if we have tuned
+	 * that to a small mtu.  We'll set qp->path_mtu
+	 * to the lesser of requested attribute mtu and active,
+	 * for packetizing messages.
+	 * Note that the QP port has to be set in INIT and MTU in RTR.
 	 */
-	return -EOPNOTSUPP;
+	if (attr_mask & IB_QP_PATH_MTU) {
+		pmtu = rdi->driver_f.get_pmtu_from_attr(rdi, qp, attr);
+		if (pmtu < 0)
+			goto inval;
+	}
+
+	if (attr_mask & IB_QP_PATH_MIG_STATE) {
+		if (attr->path_mig_state == IB_MIG_REARM) {
+			if (qp->s_mig_state == IB_MIG_ARMED)
+				goto inval;
+			if (new_state != IB_QPS_RTS)
+				goto inval;
+		} else if (attr->path_mig_state == IB_MIG_MIGRATED) {
+			if (qp->s_mig_state == IB_MIG_REARM)
+				goto inval;
+			if (new_state != IB_QPS_RTS && new_state != IB_QPS_SQD)
+				goto inval;
+			if (qp->s_mig_state == IB_MIG_ARMED)
+				mig = 1;
+		} else {
+			goto inval;
+		}
+	}
+
+	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
+		if (attr->max_dest_rd_atomic > rdi->dparms.max_rdma_atomic)
+			goto inval;
+
+	switch (new_state) {
+	case IB_QPS_RESET:
+		if (qp->state != IB_QPS_RESET)
+			rvt_reset_qp(rdi, qp, ibqp->qp_type);
+		break;
+
+	case IB_QPS_RTR:
+		/* Allow event to re-trigger if QP set to RTR more than once */
+		qp->r_flags &= ~RVT_R_COMM_EST;
+		qp->state = new_state;
+		break;
+
+	case IB_QPS_SQD:
+		qp->s_draining = qp->s_last != qp->s_cur;
+		qp->state = new_state;
+		break;
+
+	case IB_QPS_SQE:
+		if (qp->ibqp.qp_type == IB_QPT_RC)
+			goto inval;
+		qp->state = new_state;
+		break;
+
+	case IB_QPS_ERR:
+		lastwqe = rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
+		break;
+
+	default:
+		qp->state = new_state;
+		break;
+	}
+
+	if (attr_mask & IB_QP_PKEY_INDEX)
+		qp->s_pkey_index = attr->pkey_index;
+
+	if (attr_mask & IB_QP_PORT)
+		qp->port_num = attr->port_num;
+
+	if (attr_mask & IB_QP_DEST_QPN)
+		qp->remote_qpn = attr->dest_qp_num;
+
+	if (attr_mask & IB_QP_SQ_PSN) {
+		qp->s_next_psn = attr->sq_psn & rdi->dparms.psn_modify_mask;
+		qp->s_psn = qp->s_next_psn;
+		qp->s_sending_psn = qp->s_next_psn;
+		qp->s_last_psn = qp->s_next_psn - 1;
+		qp->s_sending_hpsn = qp->s_last_psn;
+	}
+
+	if (attr_mask & IB_QP_RQ_PSN)
+		qp->r_psn = attr->rq_psn & rdi->dparms.psn_modify_mask;
+
+	if (attr_mask & IB_QP_ACCESS_FLAGS)
+		qp->qp_access_flags = attr->qp_access_flags;
+
+	if (attr_mask & IB_QP_AV) {
+		qp->remote_ah_attr = attr->ah_attr;
+		qp->s_srate = attr->ah_attr.static_rate;
+		qp->srate_mbps = ib_rate_to_mbps(qp->s_srate);
+	}
+
+	if (attr_mask & IB_QP_ALT_PATH) {
+		qp->alt_ah_attr = attr->alt_ah_attr;
+		qp->s_alt_pkey_index = attr->alt_pkey_index;
+	}
+
+	if (attr_mask & IB_QP_PATH_MIG_STATE) {
+		qp->s_mig_state = attr->path_mig_state;
+		if (mig) {
+			qp->remote_ah_attr = qp->alt_ah_attr;
+			qp->port_num = qp->alt_ah_attr.port_num;
+			qp->s_pkey_index = qp->s_alt_pkey_index;
+
+			/*
+			 * Ignored by drivers which do not support it. Not
+			 * really worth creating a call back into the driver
+			 * just to set a flag.
+			 */
+			qp->s_flags |= RVT_S_AHG_CLEAR;
+		}
+	}
+
+	if (attr_mask & IB_QP_PATH_MTU) {
+		qp->pmtu = rdi->driver_f.mtu_from_qp(rdi, qp, pmtu);
+		qp->path_mtu = rdi->driver_f.mtu_to_path_mtu(qp->pmtu);
+	}
+
+	if (attr_mask & IB_QP_RETRY_CNT) {
+		qp->s_retry_cnt = attr->retry_cnt;
+		qp->s_retry = attr->retry_cnt;
+	}
+
+	if (attr_mask & IB_QP_RNR_RETRY) {
+		qp->s_rnr_retry_cnt = attr->rnr_retry;
+		qp->s_rnr_retry = attr->rnr_retry;
+	}
+
+	if (attr_mask & IB_QP_MIN_RNR_TIMER)
+		qp->r_min_rnr_timer = attr->min_rnr_timer;
+
+	if (attr_mask & IB_QP_TIMEOUT) {
+		qp->timeout = attr->timeout;
+		qp->timeout_jiffies =
+			usecs_to_jiffies((4096UL * (1UL << qp->timeout)) /
+				1000UL);
+	}
+
+	if (attr_mask & IB_QP_QKEY)
+		qp->qkey = attr->qkey;
+
+	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
+		qp->r_max_rd_atomic = attr->max_dest_rd_atomic;
+
+	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC)
+		qp->s_max_rd_atomic = attr->max_rd_atomic;
+
+	spin_unlock(&qp->s_lock);
+	spin_unlock_irq(&qp->r_lock);
+
+	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
+		rvt_insert_qp(rdi, qp);
+
+	if (lastwqe) {
+		ev.device = qp->ibqp.device;
+		ev.element.qp = &qp->ibqp;
+		ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
+		qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
+	}
+	if (mig) {
+		ev.device = qp->ibqp.device;
+		ev.element.qp = &qp->ibqp;
+		ev.event = IB_EVENT_PATH_MIG;
+		qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
+	}
+	return 0;
+
+inval:
+	spin_unlock(&qp->s_lock);
+	spin_unlock_irq(&qp->r_lock);
+	return -EINVAL;
 }
 
 /**
@@ -918,3 +1385,21 @@ int rvt_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
 {
 	return -EOPNOTSUPP;
 }
+
+void rvt_free_qpn(struct rvt_qpn_table *qpt, u32 qpn)
+{
+	struct rvt_qpn_map *map;
+
+	map = qpt->map + qpn / RVT_BITS_PER_PAGE;
+	if (map->page)
+		clear_bit(qpn & RVT_BITS_PER_PAGE_MASK, map->page);
+}
+EXPORT_SYMBOL(rvt_free_qpn);
+
+void rvt_dec_qp_cnt(struct rvt_dev_info *rdi)
+{
+	spin_lock(&rdi->n_qps_lock);
+	rdi->n_qps_allocated--;
+	spin_unlock(&rdi->n_qps_lock);
+}
+EXPORT_SYMBOL(rvt_dec_qp_cnt);
diff --git a/drivers/infiniband/sw/rdmavt/trace.h b/drivers/infiniband/sw/rdmavt/trace.h
index 22e86ff..b269291 100644
--- a/drivers/infiniband/sw/rdmavt/trace.h
+++ b/drivers/infiniband/sw/rdmavt/trace.h
@@ -77,6 +77,37 @@ TRACE_EVENT(rvt_dbg,
 	TP_printk("[%s]: %s", __get_str(dev), __get_str(msg))
 );
 
+#undef TRACE_SYSTEM
+#define TRACE_SYSTEM rvt_qphash
+DECLARE_EVENT_CLASS(rvt_qphash_template,
+	TP_PROTO(struct rvt_qp *qp, u32 bucket),
+	TP_ARGS(qp, bucket),
+	TP_STRUCT__entry(
+		RDI_DEV_ENTRY(ib_to_rvt(qp->ibqp.device))
+		__field(u32, qpn)
+		__field(u32, bucket)
+	),
+	TP_fast_assign(
+		RDI_DEV_ASSIGN(ib_to_rvt(qp->ibqp.device))
+		__entry->qpn = qp->ibqp.qp_num;
+		__entry->bucket = bucket;
+	),
+	TP_printk(
+		"[%s] qpn 0x%x bucket %u",
+		__get_str(dev),
+		__entry->qpn,
+		__entry->bucket
+	)
+);
+
+DEFINE_EVENT(rvt_qphash_template, rvt_qpinsert,
+	TP_PROTO(struct rvt_qp *qp, u32 bucket),
+	TP_ARGS(qp, bucket));
+
+DEFINE_EVENT(rvt_qphash_template, rvt_qpremove,
+	TP_PROTO(struct rvt_qp *qp, u32 bucket),
+	TP_ARGS(qp, bucket));
+
 #endif /* __RDMAVT_TRACE_H */
 
 #undef TRACE_INCLUDE_PATH
diff --git a/include/rdma/rdma_vt.h b/include/rdma/rdma_vt.h
index 3d92239..fac3f9b 100644
--- a/include/rdma/rdma_vt.h
+++ b/include/rdma/rdma_vt.h
@@ -55,6 +55,7 @@
 
 #include <linux/spinlock.h>
 #include <linux/list.h>
+#include <linux/hash.h>
 #include <rdma/ib_verbs.h>
 #include <rdma/rdmavt_mr.h>
 #include <rdma/rdmavt_qp.h>
@@ -194,6 +195,10 @@ struct rvt_driver_params {
 	u8 qos_shift;
 	char cq_name[RVT_CQN_MAX];
 	int node;
+	int max_rdma_atomic;
+	int psn_mask;
+	int psn_shift;
+	int psn_modify_mask;
 };
 
 /* Protection domain */
@@ -232,6 +237,15 @@ struct rvt_driver_provided {
 	void (*notify_qp_reset)(struct rvt_qp *qp);
 	void (*schedule_send)(struct rvt_qp *qp);
 	void (*do_send)(struct rvt_qp *qp);
+	int (*get_pmtu_from_attr)(struct rvt_dev_info *rdi, struct rvt_qp *qp,
+				  struct ib_qp_attr *attr);
+	void (*flush_qp_waiters)(struct rvt_qp *qp);
+	void (*stop_send_queue)(struct rvt_qp *qp);
+	void (*quiesce_qp)(struct rvt_qp *qp);
+	void (*notify_error_qp)(struct rvt_qp *qp);
+	u32 (*mtu_from_qp)(struct rvt_dev_info *rdi, struct rvt_qp *qp,
+			   u32 pmtu);
+	int (*mtu_to_path_mtu)(u32 mtu);
 
 	/*--------------------*/
 	/* Optional functions */
@@ -339,6 +353,34 @@ static inline u16 rvt_get_pkey(struct rvt_dev_info *rdi,
 		return rdi->ports[port_index]->pkey_table[index];
 }
 
+/**
+ * rvt_lookup_qpn - return the QP with the given QPN
+ * @ibp: the ibport
+ * @qpn: the QP number to look up
+ *
+ * The caller must hold the rcu_read_lock(), and keep the lock until
+ * the returned qp is no longer in use.
+ */
+/* TODO: Remove this and put in rdmavt/qp.h when no longer needed by drivers */
+static inline struct rvt_qp *rvt_lookup_qpn(struct rvt_dev_info *rdi,
+					    struct rvt_ibport *rvp,
+					    u32 qpn) __must_hold(RCU)
+{
+	struct rvt_qp *qp = NULL;
+
+	if (unlikely(qpn <= 1)) {
+		qp = rcu_dereference(rvp->qp[qpn]);
+	} else {
+		u32 n = hash_32(qpn, rdi->qp_dev->qp_table_bits);
+
+		for (qp = rcu_dereference(rdi->qp_dev->qp_table[n]); qp;
+			qp = rcu_dereference(qp->next))
+			if (qp->ibqp.qp_num == qpn)
+				break;
+	}
+	return qp;
+}
+
 int rvt_register_device(struct rvt_dev_info *rvd);
 void rvt_unregister_device(struct rvt_dev_info *rvd);
 int rvt_check_ah(struct ib_device *ibdev, struct ib_ah_attr *ah_attr);
diff --git a/include/rdma/rdmavt_mr.h b/include/rdma/rdmavt_mr.h
index ea60476..4aa8171 100644
--- a/include/rdma/rdmavt_mr.h
+++ b/include/rdma/rdmavt_mr.h
@@ -127,4 +127,13 @@ static inline void rvt_get_mr(struct rvt_mregion *mr)
 	atomic_inc(&mr->refcount);
 }
 
+static inline void rvt_put_ss(struct rvt_sge_state *ss)
+{
+	while (ss->num_sge) {
+		rvt_put_mr(ss->sge.mr);
+		if (--ss->num_sge)
+			ss->sge = *ss->sg_list++;
+	}
+}
+
 #endif          /* DEF_RDMAVT_INCMRH */
diff --git a/include/rdma/rdmavt_qp.h b/include/rdma/rdmavt_qp.h
index 3189f19..e66bcc9 100644
--- a/include/rdma/rdmavt_qp.h
+++ b/include/rdma/rdmavt_qp.h
@@ -355,6 +355,7 @@ struct rvt_srq {
 #define RVT_QPNMAP_ENTRIES          (RVT_QPN_MAX / PAGE_SIZE / BITS_PER_BYTE)
 #define RVT_BITS_PER_PAGE           (PAGE_SIZE * BITS_PER_BYTE)
 #define RVT_BITS_PER_PAGE_MASK      (RVT_BITS_PER_PAGE - 1)
+#define RVT_QPN_MASK		    0xFFFFFF
 
 /*
  * QPN-map pages start out as NULL, they get allocated upon
@@ -397,6 +398,25 @@ static inline struct rvt_swqe *rvt_get_swqe_ptr(struct rvt_qp *qp,
 				      sizeof(struct rvt_sge)) * n);
 }
 
+/*
+ * Since struct rvt_rwqe is not a fixed size, we can't simply index into
+ * struct rvt_rwq.wq.  This function does the array index computation.
+ */
+static inline struct rvt_rwqe *rvt_get_rwqe_ptr(struct rvt_rq *rq, unsigned n)
+{
+	return (struct rvt_rwqe *)
+		((char *)rq->wq->wq +
+		 (sizeof(struct rvt_rwqe) +
+		  rq->max_sge * sizeof(struct ib_sge)) * n);
+}
+
 extern const int  ib_rvt_state_ops[];
 
+struct rvt_dev_info;
+void rvt_remove_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp);
+void rvt_clear_mr_refs(struct rvt_qp *qp, int clr_sends);
+int rvt_error_qp(struct rvt_qp *qp, enum ib_wc_status err);
+void rvt_free_qpn(struct rvt_qpn_table *qpt, u32 qpn);
+void rvt_dec_qp_cnt(struct rvt_dev_info *rdi);
+
 #endif          /* DEF_RDMAVT_INCQP_H */

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH 5/8] IB/rdmavt: Add destroy qp verb

[Dennis Dalessandro]

This patch adds in support the qp destroy verb call.

Reviewed-by: Ira Weiny <ira.weiny-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
---
 drivers/infiniband/sw/rdmavt/qp.c |   29 +++++++++++++++++++++++------
 1 files changed, 23 insertions(+), 6 deletions(-)

diff --git a/drivers/infiniband/sw/rdmavt/qp.c b/drivers/infiniband/sw/rdmavt/qp.c
index b8c7c49..6aa7d6f 100644
--- a/drivers/infiniband/sw/rdmavt/qp.c
+++ b/drivers/infiniband/sw/rdmavt/qp.c
@@ -1171,13 +1171,30 @@ inval:
  */
 int rvt_destroy_qp(struct ib_qp *ibqp)
 {
-	/*
-	 * VT-DRIVER-API: qp_flush()
-	 * Driver provies a mechanism to flush and wait for that flush to
-	 * finish.
-	 */
+	struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
+	struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
 
-	return -EOPNOTSUPP;
+	spin_lock_irq(&qp->r_lock);
+	spin_lock(&qp->s_lock);
+	rvt_reset_qp(rdi, qp, ibqp->qp_type);
+	spin_unlock(&qp->s_lock);
+	spin_unlock_irq(&qp->r_lock);
+
+	/* qpn is now available for use again */
+	rvt_free_qpn(&rdi->qp_dev->qpn_table, qp->ibqp.qp_num);
+
+	spin_lock(&rdi->n_qps_lock);
+	rdi->n_qps_allocated--;
+	spin_unlock(&rdi->n_qps_lock);
+
+	if (qp->ip)
+		kref_put(&qp->ip->ref, rvt_release_mmap_info);
+	else
+		vfree(qp->r_rq.wq);
+	vfree(qp->s_wq);
+	rdi->driver_f.qp_priv_free(rdi, qp);
+	kfree(qp);
+	return 0;
 }
 
 int rvt_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH 6/8] IB/rdmavt: Add post receive to rdmavt

[Dennis Dalessandro]

This patch adds the simple post receive verbs call to rdmavt. The actual
interrupt handling and packet processing is still done in the low level
driver.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
Reviewed-by: Harish Chegondi <harish.chegondi-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
---
 drivers/infiniband/sw/rdmavt/qp.c |   47 +++++++++++++++++++++++++++++++------
 1 files changed, 40 insertions(+), 7 deletions(-)

diff --git a/drivers/infiniband/sw/rdmavt/qp.c b/drivers/infiniband/sw/rdmavt/qp.c
index 6aa7d6f..0b606b1 100644
--- a/drivers/infiniband/sw/rdmavt/qp.c
+++ b/drivers/infiniband/sw/rdmavt/qp.c
@@ -1214,14 +1214,47 @@ int rvt_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 int rvt_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
 		  struct ib_recv_wr **bad_wr)
 {
-	/*
-	 * When a packet arrives the driver needs to call up to rvt to process
-	 * the packet. The UD, RC, UC processing will be done in rvt, however
-	 * the driver should be able to override this if it so choses. Perhaps a
-	 * set of function pointers set up at registration time.
-	 */
+	struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
+	struct rvt_rwq *wq = qp->r_rq.wq;
+	unsigned long flags;
 
-	return -EOPNOTSUPP;
+	/* Check that state is OK to post receive. */
+	if (!(ib_rvt_state_ops[qp->state] & RVT_POST_RECV_OK) || !wq) {
+		*bad_wr = wr;
+		return -EINVAL;
+	}
+
+	for (; wr; wr = wr->next) {
+		struct rvt_rwqe *wqe;
+		u32 next;
+		int i;
+
+		if ((unsigned)wr->num_sge > qp->r_rq.max_sge) {
+			*bad_wr = wr;
+			return -EINVAL;
+		}
+
+		spin_lock_irqsave(&qp->r_rq.lock, flags);
+		next = wq->head + 1;
+		if (next >= qp->r_rq.size)
+			next = 0;
+		if (next == wq->tail) {
+			spin_unlock_irqrestore(&qp->r_rq.lock, flags);
+			*bad_wr = wr;
+			return -ENOMEM;
+		}
+
+		wqe = rvt_get_rwqe_ptr(&qp->r_rq, wq->head);
+		wqe->wr_id = wr->wr_id;
+		wqe->num_sge = wr->num_sge;
+		for (i = 0; i < wr->num_sge; i++)
+			wqe->sg_list[i] = wr->sg_list[i];
+		/* Make sure queue entry is written before the head index. */
+		smp_wmb();
+		wq->head = next;
+		spin_unlock_irqrestore(&qp->r_rq.lock, flags);
+	}
+	return 0;
 }
 
 /**

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH 7/8] IB/rdmavt: Add multicast functions

[Dennis Dalessandro]

This patch adds in the multicast add and remove functions as well as the
ancillary infrastructure needed.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
---
 drivers/infiniband/sw/rdmavt/mcast.c |  335 ++++++++++++++++++++++++++++++++++
 drivers/infiniband/sw/rdmavt/mcast.h |    2 
 drivers/infiniband/sw/rdmavt/qp.c    |    2 
 drivers/infiniband/sw/rdmavt/vt.c    |    1 
 include/rdma/rdma_vt.h               |    8 +
 include/rdma/rdmavt_qp.h             |   22 ++
 6 files changed, 367 insertions(+), 3 deletions(-)

diff --git a/drivers/infiniband/sw/rdmavt/mcast.c b/drivers/infiniband/sw/rdmavt/mcast.c
index 5a78dc7..528c1ca 100644
--- a/drivers/infiniband/sw/rdmavt/mcast.c
+++ b/drivers/infiniband/sw/rdmavt/mcast.c
@@ -45,14 +45,345 @@
  *
  */
 
+#include <linux/slab.h>
+#include <linux/sched.h>
+#include <linux/rculist.h>
+#include <rdma/rdma_vt.h>
+#include <rdma/rdmavt_qp.h>
+
 #include "mcast.h"
 
+void rvt_driver_mcast_init(struct rvt_dev_info *rdi)
+{
+	/*
+	 * Anything that needs setup for multicast on a per driver or per rdi
+	 * basis should be done in here.
+	 */
+	spin_lock_init(&rdi->n_mcast_grps_lock);
+}
+
+/**
+ * mcast_qp_alloc - alloc a struct to link a QP to mcast GID struct
+ * @qp: the QP to link
+ */
+static struct rvt_mcast_qp *rvt_mcast_qp_alloc(struct rvt_qp *qp)
+{
+	struct rvt_mcast_qp *mqp;
+
+	mqp = kmalloc(sizeof(*mqp), GFP_KERNEL);
+	if (!mqp)
+		goto bail;
+
+	mqp->qp = qp;
+	atomic_inc(&qp->refcount);
+
+bail:
+	return mqp;
+}
+
+static void rvt_mcast_qp_free(struct rvt_mcast_qp *mqp)
+{
+	struct rvt_qp *qp = mqp->qp;
+
+	/* Notify hfi1_destroy_qp() if it is waiting. */
+	if (atomic_dec_and_test(&qp->refcount))
+		wake_up(&qp->wait);
+
+	kfree(mqp);
+}
+
+/**
+ * mcast_alloc - allocate the multicast GID structure
+ * @mgid: the multicast GID
+ *
+ * A list of QPs will be attached to this structure.
+ */
+static struct rvt_mcast *rvt_mcast_alloc(union ib_gid *mgid)
+{
+	struct rvt_mcast *mcast;
+
+	mcast = kzalloc(sizeof(*mcast), GFP_KERNEL);
+	if (!mcast)
+		goto bail;
+
+	mcast->mgid = *mgid;
+	INIT_LIST_HEAD(&mcast->qp_list);
+	init_waitqueue_head(&mcast->wait);
+	atomic_set(&mcast->refcount, 0);
+
+bail:
+	return mcast;
+}
+
+static void rvt_mcast_free(struct rvt_mcast *mcast)
+{
+	struct rvt_mcast_qp *p, *tmp;
+
+	list_for_each_entry_safe(p, tmp, &mcast->qp_list, list)
+		rvt_mcast_qp_free(p);
+
+	kfree(mcast);
+}
+
+/**
+ * rvt_mcast_find - search the global table for the given multicast GID
+ * @ibp: the IB port structure
+ * @mgid: the multicast GID to search for
+ *
+ * Returns NULL if not found.
+ *
+ * The caller is responsible for decrementing the reference count if found.
+ */
+struct rvt_mcast *rvt_mcast_find(struct rvt_ibport *ibp, union ib_gid *mgid)
+{
+	struct rb_node *n;
+	unsigned long flags;
+	struct rvt_mcast *found = NULL;
+
+	spin_lock_irqsave(&ibp->lock, flags);
+	n = ibp->mcast_tree.rb_node;
+	while (n) {
+		int ret;
+		struct rvt_mcast *mcast;
+
+		mcast = rb_entry(n, struct rvt_mcast, rb_node);
+
+		ret = memcmp(mgid->raw, mcast->mgid.raw,
+			     sizeof(union ib_gid));
+		if (ret < 0) {
+			n = n->rb_left;
+		} else if (ret > 0) {
+			n = n->rb_right;
+		} else {
+			atomic_inc(&mcast->refcount);
+			found = mcast;
+			break;
+		}
+	}
+	spin_unlock_irqrestore(&ibp->lock, flags);
+	return found;
+}
+EXPORT_SYMBOL(rvt_mcast_find);
+
+/**
+ * mcast_add - insert mcast GID into table and attach QP struct
+ * @mcast: the mcast GID table
+ * @mqp: the QP to attach
+ *
+ * Return zero if both were added.  Return EEXIST if the GID was already in
+ * the table but the QP was added.  Return ESRCH if the QP was already
+ * attached and neither structure was added.
+ */
+static int rvt_mcast_add(struct rvt_dev_info *rdi, struct rvt_ibport *ibp,
+			 struct rvt_mcast *mcast, struct rvt_mcast_qp *mqp)
+{
+	struct rb_node **n = &ibp->mcast_tree.rb_node;
+	struct rb_node *pn = NULL;
+	int ret;
+
+	spin_lock_irq(&ibp->lock);
+
+	while (*n) {
+		struct rvt_mcast *tmcast;
+		struct rvt_mcast_qp *p;
+
+		pn = *n;
+		tmcast = rb_entry(pn, struct rvt_mcast, rb_node);
+
+		ret = memcmp(mcast->mgid.raw, tmcast->mgid.raw,
+			     sizeof(union ib_gid));
+		if (ret < 0) {
+			n = &pn->rb_left;
+			continue;
+		}
+		if (ret > 0) {
+			n = &pn->rb_right;
+			continue;
+		}
+
+		/* Search the QP list to see if this is already there. */
+		list_for_each_entry_rcu(p, &tmcast->qp_list, list) {
+			if (p->qp == mqp->qp) {
+				ret = ESRCH;
+				goto bail;
+			}
+		}
+		if (tmcast->n_attached ==
+		    rdi->dparms.props.max_mcast_qp_attach) {
+			ret = ENOMEM;
+			goto bail;
+		}
+
+		tmcast->n_attached++;
+
+		list_add_tail_rcu(&mqp->list, &tmcast->qp_list);
+		ret = EEXIST;
+		goto bail;
+	}
+
+	spin_lock(&rdi->n_mcast_grps_lock);
+	if (rdi->n_mcast_grps_allocated == rdi->dparms.props.max_mcast_grp) {
+		spin_unlock(&rdi->n_mcast_grps_lock);
+		ret = ENOMEM;
+		goto bail;
+	}
+
+	rdi->n_mcast_grps_allocated++;
+	spin_unlock(&rdi->n_mcast_grps_lock);
+
+	mcast->n_attached++;
+
+	list_add_tail_rcu(&mqp->list, &mcast->qp_list);
+
+	atomic_inc(&mcast->refcount);
+	rb_link_node(&mcast->rb_node, pn, n);
+	rb_insert_color(&mcast->rb_node, &ibp->mcast_tree);
+
+	ret = 0;
+
+bail:
+	spin_unlock_irq(&ibp->lock);
+
+	return ret;
+}
+
 int rvt_attach_mcast(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
 {
-	return -EOPNOTSUPP;
+	struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
+	struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
+	struct rvt_ibport *ibp = rdi->ports[qp->port_num - 1];
+	struct rvt_mcast *mcast;
+	struct rvt_mcast_qp *mqp;
+	int ret = -ENOMEM;
+
+	if (ibqp->qp_num <= 1 || qp->state == IB_QPS_RESET)
+		return -EINVAL;
+
+	/*
+	 * Allocate data structures since its better to do this outside of
+	 * spin locks and it will most likely be needed.
+	 */
+	mcast = rvt_mcast_alloc(gid);
+	if (!mcast)
+		return -ENOMEM;
+
+	mqp = rvt_mcast_qp_alloc(qp);
+	if (!mqp)
+		goto bail_mcast;
+
+	switch (rvt_mcast_add(rdi, ibp, mcast, mqp)) {
+	case ESRCH:
+		/* Neither was used: OK to attach the same QP twice. */
+		ret = 0;
+		goto bail_mqp;
+	case EEXIST: /* The mcast wasn't used */
+		ret = 0;
+		goto bail_mcast;
+	case ENOMEM:
+		/* Exceeded the maximum number of mcast groups. */
+		ret = -ENOMEM;
+		goto bail_mqp;
+	default:
+		break;
+	}
+
+	return 0;
+
+bail_mqp:
+	rvt_mcast_qp_free(mqp);
+
+bail_mcast:
+	rvt_mcast_free(mcast);
+
+	return ret;
 }
 
 int rvt_detach_mcast(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
 {
-	return -EOPNOTSUPP;
+	struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
+	struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
+	struct rvt_ibport *ibp = rdi->ports[qp->port_num - 1];
+	struct rvt_mcast *mcast = NULL;
+	struct rvt_mcast_qp *p, *tmp, *delp = NULL;
+	struct rb_node *n;
+	int last = 0;
+	int ret = 0;
+
+	if (ibqp->qp_num <= 1 || qp->state == IB_QPS_RESET)
+		return -EINVAL;
+
+	spin_lock_irq(&ibp->lock);
+
+	/* Find the GID in the mcast table. */
+	n = ibp->mcast_tree.rb_node;
+	while (1) {
+		if (!n) {
+			spin_unlock_irq(&ibp->lock);
+			return -EINVAL;
+		}
+
+		mcast = rb_entry(n, struct rvt_mcast, rb_node);
+		ret = memcmp(gid->raw, mcast->mgid.raw,
+			     sizeof(union ib_gid));
+		if (ret < 0)
+			n = n->rb_left;
+		else if (ret > 0)
+			n = n->rb_right;
+		else
+			break;
+	}
+
+	/* Search the QP list. */
+	list_for_each_entry_safe(p, tmp, &mcast->qp_list, list) {
+		if (p->qp != qp)
+			continue;
+		/*
+		 * We found it, so remove it, but don't poison the forward
+		 * link until we are sure there are no list walkers.
+		 */
+		list_del_rcu(&p->list);
+		mcast->n_attached--;
+		delp = p;
+
+		/* If this was the last attached QP, remove the GID too. */
+		if (list_empty(&mcast->qp_list)) {
+			rb_erase(&mcast->rb_node, &ibp->mcast_tree);
+			last = 1;
+		}
+		break;
+	}
+
+	spin_unlock_irq(&ibp->lock);
+	/* QP not attached */
+	if (!delp)
+		return -EINVAL;
+
+	/*
+	 * Wait for any list walkers to finish before freeing the
+	 * list element.
+	 */
+	wait_event(mcast->wait, atomic_read(&mcast->refcount) <= 1);
+	rvt_mcast_qp_free(delp);
+
+	if (last) {
+		atomic_dec(&mcast->refcount);
+		wait_event(mcast->wait, !atomic_read(&mcast->refcount));
+		rvt_mcast_free(mcast);
+		spin_lock_irq(&rdi->n_mcast_grps_lock);
+		rdi->n_mcast_grps_allocated--;
+		spin_unlock_irq(&rdi->n_mcast_grps_lock);
+	}
+
+	return 0;
+}
+
+int rvt_mcast_tree_empty(struct rvt_dev_info *rdi)
+{
+	int i;
+	int in_use = 0;
+
+	for (i = 0; i < rdi->dparms.nports; i++)
+		if (rdi->ports[i]->mcast_tree.rb_node)
+			in_use++;
+	return in_use;
 }
diff --git a/drivers/infiniband/sw/rdmavt/mcast.h b/drivers/infiniband/sw/rdmavt/mcast.h
index 21647c3..cd15a98 100644
--- a/drivers/infiniband/sw/rdmavt/mcast.h
+++ b/drivers/infiniband/sw/rdmavt/mcast.h
@@ -50,7 +50,9 @@
 
 #include <rdma/rdma_vt.h>
 
+void rvt_driver_mcast_init(struct rvt_dev_info *rdi);
 int rvt_attach_mcast(struct ib_qp *ibqp, union ib_gid *gid, u16 lid);
 int rvt_detach_mcast(struct ib_qp *ibqp, union ib_gid *gid, u16 lid);
+int rvt_mcast_tree_empty(struct rvt_dev_info *rdi);
 
 #endif          /* DEF_RVTMCAST_H */
diff --git a/drivers/infiniband/sw/rdmavt/qp.c b/drivers/infiniband/sw/rdmavt/qp.c
index 0b606b1..6b76cf2 100644
--- a/drivers/infiniband/sw/rdmavt/qp.c
+++ b/drivers/infiniband/sw/rdmavt/qp.c
@@ -238,6 +238,8 @@ static unsigned rvt_free_all_qps(struct rvt_dev_info *rdi)
 	if (rdi->driver_f.free_all_qps)
 		qp_inuse = rdi->driver_f.free_all_qps(rdi);
 
+	qp_inuse += rvt_mcast_tree_empty(rdi);
+
 	if (!rdi->qp_dev)
 		return qp_inuse;
 
diff --git a/drivers/infiniband/sw/rdmavt/vt.c b/drivers/infiniband/sw/rdmavt/vt.c
index 7f56a42..5a094eb 100644
--- a/drivers/infiniband/sw/rdmavt/vt.c
+++ b/drivers/infiniband/sw/rdmavt/vt.c
@@ -305,6 +305,7 @@ int rvt_register_device(struct rvt_dev_info *rdi)
 	CHECK_DRIVER_OVERRIDE(rdi, query_srq);
 
 	/* Multicast */
+	rvt_driver_mcast_init(rdi);
 	CHECK_DRIVER_OVERRIDE(rdi, attach_mcast);
 	CHECK_DRIVER_OVERRIDE(rdi, detach_mcast);
 
diff --git a/include/rdma/rdma_vt.h b/include/rdma/rdma_vt.h
index fac3f9b..1ad902d 100644
--- a/include/rdma/rdma_vt.h
+++ b/include/rdma/rdma_vt.h
@@ -305,6 +305,11 @@ struct rvt_dev_info {
 	struct kthread_worker *worker; /* per device cq worker */
 	u32 n_cqs_allocated;    /* number of CQs allocated for device */
 	spinlock_t n_cqs_lock; /* protect count of in use cqs */
+
+	/* Multicast */
+	u32 n_mcast_grps_allocated; /* number of mcast groups allocated */
+	spinlock_t n_mcast_grps_lock;
+
 };
 
 static inline struct rvt_pd *ibpd_to_rvtpd(struct ib_pd *ibpd)
@@ -398,8 +403,11 @@ struct rvt_mmap_info *rvt_create_mmap_info(struct rvt_dev_info *rdi,
 					   void *obj);
 void rvt_update_mmap_info(struct rvt_dev_info *rdi, struct rvt_mmap_info *ip,
 			  u32 size, void *obj);
+int rvt_reg_mr(struct rvt_qp *qp, struct ib_reg_wr *wr);
+struct rvt_mcast *rvt_mcast_find(struct rvt_ibport *ibp, union ib_gid *mgid);
 
 /* Temporary export */
 void rvt_reset_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp,
 		  enum ib_qp_type type);
+
 #endif          /* DEF_RDMA_VT_H */
diff --git a/include/rdma/rdmavt_qp.h b/include/rdma/rdmavt_qp.h
index e66bcc9..a97b95b 100644
--- a/include/rdma/rdmavt_qp.h
+++ b/include/rdma/rdmavt_qp.h
@@ -50,6 +50,7 @@
 
 #include <rdma/rdma_vt.h>
 #include <rdma/ib_pack.h>
+#include <rdma/ib_verbs.h>
 /*
  * Atomic bit definitions for r_aflags.
  */
@@ -386,8 +387,27 @@ struct rvt_qp_ibdev {
 };
 
 /*
+ * There is one struct rvt_mcast for each multicast GID.
+ * All attached QPs are then stored as a list of
+ * struct rvt_mcast_qp.
+ */
+struct rvt_mcast_qp {
+	struct list_head list;
+	struct rvt_qp *qp;
+};
+
+struct rvt_mcast {
+	struct rb_node rb_node;
+	union ib_gid mgid;
+	struct list_head qp_list;
+	wait_queue_head_t wait;
+	atomic_t refcount;
+	int n_attached;
+};
+
+/*
  * Since struct rvt_swqe is not a fixed size, we can't simply index into
- * struct hfi1_qp.s_wq.  This function does the array index computation.
+ * struct rvt_qp.s_wq.  This function does the array index computation.
  */
 static inline struct rvt_swqe *rvt_get_swqe_ptr(struct rvt_qp *qp,
 						unsigned n)

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH 8/8] IB/rdmavt: Add misc dev register functionality

[Dennis Dalessandro]

There are a number of minor things that should be set by rdmavt rather
than by the drivers. Now that rdmavt has solidified in its design we can
go ahead and clean up this stuff.

Reviewed-by: Ira Weiny <ira.weiny-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
---
 drivers/infiniband/sw/rdmavt/vt.c |   43 +++++++++++++++++++++++++++++++++++++
 1 files changed, 43 insertions(+), 0 deletions(-)

diff --git a/drivers/infiniband/sw/rdmavt/vt.c b/drivers/infiniband/sw/rdmavt/vt.c
index 5a094eb..cf7cac6 100644
--- a/drivers/infiniband/sw/rdmavt/vt.c
+++ b/drivers/infiniband/sw/rdmavt/vt.c
@@ -50,6 +50,8 @@
 #include "vt.h"
 #include "trace.h"
 
+#define RVT_UVERBS_ABI_VERSION 2
+
 MODULE_LICENSE("Dual BSD/GPL");
 MODULE_DESCRIPTION("RDMA Verbs Transport Library");
 
@@ -348,6 +350,47 @@ int rvt_register_device(struct rvt_dev_info *rdi)
 	spin_lock_init(&rdi->n_pds_lock);
 	rdi->n_pds_allocated = 0;
 
+	/*
+	 * There are some things which could be set by underlying drivers but
+	 * really should be up to rdmavt to set. For instance drivers can't know
+	 * exactly which functions rdmavt supports, nor do they know the ABI
+	 * version, so we do all of this sort of stuff here.
+	 */
+	rdi->ibdev.uverbs_abi_ver = RVT_UVERBS_ABI_VERSION;
+	rdi->ibdev.uverbs_cmd_mask =
+		(1ull << IB_USER_VERBS_CMD_GET_CONTEXT)         |
+		(1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)        |
+		(1ull << IB_USER_VERBS_CMD_QUERY_PORT)          |
+		(1ull << IB_USER_VERBS_CMD_ALLOC_PD)            |
+		(1ull << IB_USER_VERBS_CMD_DEALLOC_PD)          |
+		(1ull << IB_USER_VERBS_CMD_CREATE_AH)           |
+		(1ull << IB_USER_VERBS_CMD_MODIFY_AH)           |
+		(1ull << IB_USER_VERBS_CMD_QUERY_AH)            |
+		(1ull << IB_USER_VERBS_CMD_DESTROY_AH)          |
+		(1ull << IB_USER_VERBS_CMD_REG_MR)              |
+		(1ull << IB_USER_VERBS_CMD_DEREG_MR)            |
+		(1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
+		(1ull << IB_USER_VERBS_CMD_CREATE_CQ)           |
+		(1ull << IB_USER_VERBS_CMD_RESIZE_CQ)           |
+		(1ull << IB_USER_VERBS_CMD_DESTROY_CQ)          |
+		(1ull << IB_USER_VERBS_CMD_POLL_CQ)             |
+		(1ull << IB_USER_VERBS_CMD_REQ_NOTIFY_CQ)       |
+		(1ull << IB_USER_VERBS_CMD_CREATE_QP)           |
+		(1ull << IB_USER_VERBS_CMD_QUERY_QP)            |
+		(1ull << IB_USER_VERBS_CMD_MODIFY_QP)           |
+		(1ull << IB_USER_VERBS_CMD_DESTROY_QP)          |
+		(1ull << IB_USER_VERBS_CMD_POST_SEND)           |
+		(1ull << IB_USER_VERBS_CMD_POST_RECV)           |
+		(1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)        |
+		(1ull << IB_USER_VERBS_CMD_DETACH_MCAST)        |
+		(1ull << IB_USER_VERBS_CMD_CREATE_SRQ)          |
+		(1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)          |
+		(1ull << IB_USER_VERBS_CMD_QUERY_SRQ)           |
+		(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;
+
 	/* We are now good to announce we exist */
 	ret =  ib_register_device(&rdi->ibdev, rdi->driver_f.port_callback);
 	if (ret) {

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH 2/8] IB/rdmavt: Add post send to rdmavt

[Moni Shoua]

On Sat, Jan 9, 2016 at 5:00 PM, Dennis Dalessandro
<dennis.dalessandro-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org> wrote:
> Add in a post_send and post_one_send to rdmavt. The ULP will provide a WQE
> to rdmavt which will then walk and queue each element. Rdmavt will then
> queue the work to be done in the driver or kick the driver's progress
> routine.
>
> There needs to be a follow on patch which adds in another lock for the
> head of the queue so that it can be added to and read from in parallel.
> This will touch protocol handlers and require other changes in the
> drivers. This will be done separately.
>
> Reviewed-by: Mike Marciniszyn <mike.marciniszyn-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
> Reviewed-by: Ira Weiny <ira.weiny-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
> Signed-off-by: Dennis Dalessandro <dennis.dalessandro-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
> ---
>  drivers/infiniband/sw/rdmavt/qp.c |  183 +++++++++++++++++++++++++++++++++++--
>  include/rdma/rdma_vt.h            |    7 +
>  include/rdma/rdmavt_qp.h          |   26 +++++
>  3 files changed, 204 insertions(+), 12 deletions(-)
>
> diff --git a/drivers/infiniband/sw/rdmavt/qp.c b/drivers/infiniband/sw/rdmavt/qp.c
> index ee19eae..5063e25 100644
> --- a/drivers/infiniband/sw/rdmavt/qp.c
> +++ b/drivers/infiniband/sw/rdmavt/qp.c
> @@ -53,6 +53,27 @@
>  #include "qp.h"
>  #include "vt.h"
>
> +/*
> + * Note that it is OK to post send work requests in the SQE and ERR
> + * states; rvt_do_send() will process them and generate error
> + * completions as per IB 1.2 C10-96.
> + */
> +const int ib_rvt_state_ops[IB_QPS_ERR + 1] = {
> +       [IB_QPS_RESET] = 0,
> +       [IB_QPS_INIT] = RVT_POST_RECV_OK,
> +       [IB_QPS_RTR] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK,
> +       [IB_QPS_RTS] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK |
> +           RVT_POST_SEND_OK | RVT_PROCESS_SEND_OK |
> +           RVT_PROCESS_NEXT_SEND_OK,
> +       [IB_QPS_SQD] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK |
> +           RVT_POST_SEND_OK | RVT_PROCESS_SEND_OK,
> +       [IB_QPS_SQE] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK |
> +           RVT_POST_SEND_OK | RVT_FLUSH_SEND,
> +       [IB_QPS_ERR] = RVT_POST_RECV_OK | RVT_FLUSH_RECV |
> +           RVT_POST_SEND_OK | RVT_FLUSH_SEND,
> +};
> +EXPORT_SYMBOL(ib_rvt_state_ops);
> +
>  static void get_map_page(struct rvt_qpn_table *qpt, struct rvt_qpn_map *map)
>  {
>         unsigned long page = get_zeroed_page(GFP_KERNEL);
> @@ -556,7 +577,7 @@ struct ib_qp *rvt_create_qp(struct ib_pd *ibpd,
>
>         /*
>          * Return the address of the RWQ as the offset to mmap.
> -        * See hfi1_mmap() for details.
> +        * See rvt_mmap() for details.
>          */
>         if (udata && udata->outlen >= sizeof(__u64)) {
>                 if (!qp->r_rq.wq) {
> @@ -720,6 +741,118 @@ int rvt_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
>  }
>
>  /**
> + * rvt_post_one_wr - post one RC, UC, or UD send work request
> + * @qp: the QP to post on
> + * @wr: the work request to send
> + */
> +static int rvt_post_one_wr(struct rvt_qp *qp, struct ib_send_wr *wr)
> +{
> +       struct rvt_swqe *wqe;
> +       u32 next;
> +       int i;
> +       int j;
> +       int acc;
> +       struct rvt_lkey_table *rkt;
> +       struct rvt_pd *pd;
> +       struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
> +
> +       /* IB spec says that num_sge == 0 is OK. */
> +       if (unlikely(wr->num_sge > qp->s_max_sge))
> +               return -EINVAL;
> +
> +       /*
> +        * Don't allow RDMA reads or atomic operations on UC or
> +        * undefined operations.
> +        * Make sure buffer is large enough to hold the result for atomics.
> +        */
> +       if (qp->ibqp.qp_type == IB_QPT_UC) {
> +               if ((unsigned)wr->opcode >= IB_WR_RDMA_READ)
> +                       return -EINVAL;
> +       } else if (qp->ibqp.qp_type != IB_QPT_RC) {
> +               /* Check IB_QPT_SMI, IB_QPT_GSI, IB_QPT_UD opcode */
> +               if (wr->opcode != IB_WR_SEND &&
> +                   wr->opcode != IB_WR_SEND_WITH_IMM)
> +                       return -EINVAL;
> +               /* Check UD destination address PD */
> +               if (qp->ibqp.pd != ud_wr(wr)->ah->pd)
> +                       return -EINVAL;
> +       } else if ((unsigned)wr->opcode > IB_WR_ATOMIC_FETCH_AND_ADD) {
> +               return -EINVAL;
> +       } else if (wr->opcode >= IB_WR_ATOMIC_CMP_AND_SWP &&
> +                  (wr->num_sge == 0 ||
> +                   wr->sg_list[0].length < sizeof(u64) ||
> +                   wr->sg_list[0].addr & (sizeof(u64) - 1))) {
> +               return -EINVAL;
> +       } else if (wr->opcode >= IB_WR_RDMA_READ && !qp->s_max_rd_atomic) {
> +               return -EINVAL;
> +       }
> +
> +       next = qp->s_head + 1;
> +       if (next >= qp->s_size)
> +               next = 0;
> +       if (next == qp->s_last)
> +               return -ENOMEM;
> +
> +       rkt = &rdi->lkey_table;
> +       pd = ibpd_to_rvtpd(qp->ibqp.pd);
> +       wqe = rvt_get_swqe_ptr(qp, qp->s_head);
> +
> +       if (qp->ibqp.qp_type != IB_QPT_UC &&
> +           qp->ibqp.qp_type != IB_QPT_RC)
> +               memcpy(&wqe->ud_wr, ud_wr(wr), sizeof(wqe->ud_wr));
> +       else if (wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM ||
> +                wr->opcode == IB_WR_RDMA_WRITE ||
> +                wr->opcode == IB_WR_RDMA_READ)
> +               memcpy(&wqe->rdma_wr, rdma_wr(wr), sizeof(wqe->rdma_wr));
> +       else if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
> +                wr->opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
> +               memcpy(&wqe->atomic_wr, atomic_wr(wr), sizeof(wqe->atomic_wr));
> +       else
> +               memcpy(&wqe->wr, wr, sizeof(wqe->wr));
> +
> +       wqe->length = 0;
> +       j = 0;
> +       if (wr->num_sge) {
> +               acc = wr->opcode >= IB_WR_RDMA_READ ?
> +                       IB_ACCESS_LOCAL_WRITE : 0;
> +               for (i = 0; i < wr->num_sge; i++) {
> +                       u32 length = wr->sg_list[i].length;
> +                       int ok;
> +
> +                       if (length == 0)
> +                               continue;
> +                       ok = rvt_lkey_ok(rkt, pd, &wqe->sg_list[j],
> +                                        &wr->sg_list[i], acc);
> +                       if (!ok)
> +                               goto bail_inval_free;
> +                       wqe->length += length;
> +                       j++;
> +               }
> +               wqe->wr.num_sge = j;
> +       }
> +       if (qp->ibqp.qp_type == IB_QPT_UC ||
> +           qp->ibqp.qp_type == IB_QPT_RC) {
> +               if (wqe->length > 0x80000000U)
> +                       goto bail_inval_free;
> +       } else {
> +               atomic_inc(&ibah_to_rvtah(ud_wr(wr)->ah)->refcount);
> +       }
> +       wqe->ssn = qp->s_ssn++;
> +       qp->s_head = next;
> +
> +       return 0;
> +
> +bail_inval_free:
> +       /* release mr holds */
> +       while (j) {
> +               struct rvt_sge *sge = &wqe->sg_list[--j];
> +
> +               rvt_put_mr(sge->mr);
> +       }
> +       return -EINVAL;
> +}
> +
> +/**
>   * rvt_post_send - post a send on a QP
>   * @ibqp: the QP to post the send on
>   * @wr: the list of work requests to post
> @@ -730,20 +863,46 @@ int rvt_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
>  int rvt_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
>                   struct ib_send_wr **bad_wr)
>  {
> +       struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
> +       struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
> +       unsigned long flags = 0;
> +       int call_send;
> +       unsigned nreq = 0;
> +       int err = 0;
> +
> +       spin_lock_irqsave(&qp->s_lock, flags);
> +
>         /*
> -        * VT-DRIVER-API: do_send()
> -        * Driver needs to have a do_send() call which is a single entry point
> -        * to take an already formed packet and throw it out on the wire. Once
> -        * the packet is sent the driver needs to make an upcall to rvt so the
> -        * completion queue can be notified and/or any other outstanding
> -        * work/book keeping can be finished.
Why was this design for do_send() removed?
This requires that each low level drivers do the packet formatting by
themselves --> code duplication.
> -        *
> -        * Note that there should also be a way for rvt to protect itself
> -        * against hangs in the driver layer. If a send doesn't actually
> -        * complete in a timely manor rvt needs to return an error event.
> +        * Ensure QP state is such that we can send. If not bail out early,
> +        * there is no need to do this every time we post a send.
>          */
> +       if (unlikely(!(ib_rvt_state_ops[qp->state] & RVT_POST_SEND_OK))) {
> +               spin_unlock_irqrestore(&qp->s_lock, flags);
> +               return -EINVAL;
> +       }
>
> -       return -EOPNOTSUPP;
> +       /*
> +        * If the send queue is empty, and we only have a single WR then just go
> +        * ahead and kick the send engine into gear. Otherwise we will always
> +        * just schedule the send to happen later.
> +        */
> +       call_send = qp->s_head == ACCESS_ONCE(qp->s_last) && !wr->next;
> +
> +       for (; wr; wr = wr->next) {
> +               err = rvt_post_one_wr(qp, wr);
> +               if (unlikely(err)) {
> +                       *bad_wr = wr;
> +                       goto bail;
> +               }
> +               nreq++;
> +       }
> +bail:
> +       if (nreq && !call_send)
> +               rdi->driver_f.schedule_send(qp);
> +       spin_unlock_irqrestore(&qp->s_lock, flags);
> +       if (nreq && call_send)
> +               rdi->driver_f.do_send(qp);
> +       return err;
>  }
Wouldn't it be better if scheduling for sending packets be done in RVT?
>
>  /**
> diff --git a/include/rdma/rdma_vt.h b/include/rdma/rdma_vt.h
> index f2d50b9..3d92239 100644
> --- a/include/rdma/rdma_vt.h
> +++ b/include/rdma/rdma_vt.h
> @@ -230,6 +230,8 @@ struct rvt_driver_provided {
>         void * (*qp_priv_alloc)(struct rvt_dev_info *rdi, struct rvt_qp *qp);
>         void (*qp_priv_free)(struct rvt_dev_info *rdi, struct rvt_qp *qp);
>         void (*notify_qp_reset)(struct rvt_qp *qp);
> +       void (*schedule_send)(struct rvt_qp *qp);
> +       void (*do_send)(struct rvt_qp *qp);
A proper explanation about the meaning for each driver provided
function is required.
This is true for the 2 new added function and for any other that was
already there
>
>         /*--------------------*/
>         /* Optional functions */
> @@ -311,6 +313,11 @@ static inline struct rvt_srq *ibsrq_to_rvtsrq(struct ib_srq *ibsrq)
>         return container_of(ibsrq, struct rvt_srq, ibsrq);
>  }
>
> +static inline struct rvt_qp *ibqp_to_rvtqp(struct ib_qp *ibqp)
> +{
> +       return container_of(ibqp, struct rvt_qp, ibqp);
> +}
> +
>  static inline unsigned rvt_get_npkeys(struct rvt_dev_info *rdi)
>  {
>         /*
> diff --git a/include/rdma/rdmavt_qp.h b/include/rdma/rdmavt_qp.h
> index bce0a03..3189f19 100644
> --- a/include/rdma/rdmavt_qp.h
> +++ b/include/rdma/rdmavt_qp.h
> @@ -129,6 +129,17 @@
>  /* Number of bits to pay attention to in the opcode for checking qp type */
>  #define RVT_OPCODE_QP_MASK 0xE0
>
> +/* Flags for checking QP state (see ib_rvt_state_ops[]) */
> +#define RVT_POST_SEND_OK                0x01
> +#define RVT_POST_RECV_OK                0x02
> +#define RVT_PROCESS_RECV_OK             0x04
> +#define RVT_PROCESS_SEND_OK             0x08
> +#define RVT_PROCESS_NEXT_SEND_OK        0x10
> +#define RVT_FLUSH_SEND                 0x20
> +#define RVT_FLUSH_RECV                 0x40
> +#define RVT_PROCESS_OR_FLUSH_SEND \
> +       (RVT_PROCESS_SEND_OK | RVT_FLUSH_SEND)
> +
>  /*
>   * Send work request queue entry.
>   * The size of the sg_list is determined when the QP is created and stored
> @@ -373,4 +384,19 @@ struct rvt_qp_ibdev {
>         struct rvt_qpn_table qpn_table;
>  };
>
> +/*
> + * Since struct rvt_swqe is not a fixed size, we can't simply index into
> + * struct hfi1_qp.s_wq.  This function does the array index computation.
> + */
> +static inline struct rvt_swqe *rvt_get_swqe_ptr(struct rvt_qp *qp,
> +                                               unsigned n)
> +{
> +       return (struct rvt_swqe *)((char *)qp->s_wq +
> +                                    (sizeof(struct rvt_swqe) +
> +                                     qp->s_max_sge *
> +                                     sizeof(struct rvt_sge)) * n);
> +}
> +
> +extern const int  ib_rvt_state_ops[];
> +
>  #endif          /* DEF_RDMAVT_INCQP_H */
>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
> the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH 0/8] IB/rdmavt: Add functions for cq, qp, post send and recv

[Moni Shoua]

On Sat, Jan 9, 2016 at 4:59 PM, Dennis Dalessandro
<dennis.dalessandro-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org> wrote:
> This series continues adding support for verbs into rdmavt. This brings in
> completion queue, queue pair, and post send/recv functionality.
>
> It also introduces event tracing and moves some more of the hfi1/qib
> registration functionality.
>
> This patch set applies on top of the "Start to rely on rdmavt for qp support"
> series previously submitted.
>
> ---
>
> Dennis Dalessandro (8):
>       IB/rdmavt: Add completion queue functions
>       IB/rdmavt: Add post send to rdmavt
>       IB/rdmavt: Add support for tracing events
>       IB/rdmavt: Add modify qp
>       IB/rdmavt: Add destroy qp verb
>       IB/rdmavt: Add post receive to rdmavt
>       IB/rdmavt: Add multicast functions
>       IB/rdmavt: Add misc dev register functionality

I didn't see anywhere an API for RVT that handles received packets
Did I miss it?
If not, what is expected from the lower driver to do when a packet is received?
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH 2/8] IB/rdmavt: Add post send to rdmavt

[Dennis Dalessandro]

On Tue, Jan 19, 2016 at 06:23:14PM +0200, Moni Shoua wrote:
>> @@ -730,20 +863,46 @@ int rvt_post_recv(struct ib_qp *ibqp, struct 
>> ib_recv_wr *wr,
>>  int rvt_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
>>                   struct ib_send_wr **bad_wr)
>>  {
>> +       struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
>> +       struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
>> +       unsigned long flags = 0;
>> +       int call_send;
>> +       unsigned nreq = 0;
>> +       int err = 0;
>> +
>> +       spin_lock_irqsave(&qp->s_lock, flags);
>> +
>>         /*
>> -        * VT-DRIVER-API: do_send()
>> -        * Driver needs to have a do_send() call which is a single entry point
>> -        * to take an already formed packet and throw it out on the wire. Once
>> -        * the packet is sent the driver needs to make an upcall to rvt so the
>> -        * completion queue can be notified and/or any other outstanding
>> -        * work/book keeping can be finished.
>Why was this design for do_send() removed?
>This requires that each low level drivers do the packet formatting by
>themselves --> code duplication.

The packet building is still being done in the drivers right now. While the 
code is similar between the existing drivers there are hardware specific 
aspects which are critical to performance which will need untangled very 
carefully.  We do plan to move this stuff into rdmavt as well. However, I 
think it makes sense to let this initial cut of rdmavt get some soak time 
given the scope of the changes. So this would be a next step in the 
iterative development.

>> +bail:
>> +       if (nreq && !call_send)
>> +               rdi->driver_f.schedule_send(qp);
>> +       spin_unlock_irqrestore(&qp->s_lock, flags);
>> +       if (nreq && call_send)
>> +               rdi->driver_f.do_send(qp);
>> +       return err;
>>  }
>Wouldn't it be better if scheduling for sending packets be done in RVT?

The progress mechanism is probably a device specific thing. One 
implementation may use work queues, another, something else. For this reason 
we leave it down in the driver. Rdmavt has the two driver calls shown here, 
one to tell the driver to schedule a packet for sending based on whatever 
it's policy is. The other is to kick the driver to do a send right away, for 
performance reasons.

>>  /**
>> diff --git a/include/rdma/rdma_vt.h b/include/rdma/rdma_vt.h
>> index f2d50b9..3d92239 100644
>> --- a/include/rdma/rdma_vt.h
>> +++ b/include/rdma/rdma_vt.h
>> @@ -230,6 +230,8 @@ struct rvt_driver_provided {
>>         void * (*qp_priv_alloc)(struct rvt_dev_info *rdi, struct rvt_qp *qp);
>>         void (*qp_priv_free)(struct rvt_dev_info *rdi, struct rvt_qp *qp);
>>         void (*notify_qp_reset)(struct rvt_qp *qp);
>> +       void (*schedule_send)(struct rvt_qp *qp);
>> +       void (*do_send)(struct rvt_qp *qp);
>A proper explanation about the meaning for each driver provided
>function is required.
>This is true for the 2 new added function and for any other that was
>already there

Agree. I'll perhaps add another patch which cleans up these and other 
comments. Make sure it all makes sense as a number of the original comments 
were put in place to get the ball rolling and stimulate discussion.

-Denny
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH 0/8] IB/rdmavt: Add functions for cq, qp, post send and recv

[Dennis Dalessandro]

On Tue, Jan 19, 2016 at 06:44:11PM +0200, Moni Shoua wrote:
>On Sat, Jan 9, 2016 at 4:59 PM, Dennis Dalessandro
><dennis.dalessandro-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org> wrote:
>> This series continues adding support for verbs into rdmavt. This brings in
>> completion queue, queue pair, and post send/recv functionality.
>>
>> It also introduces event tracing and moves some more of the hfi1/qib
>> registration functionality.
>>
>> This patch set applies on top of the "Start to rely on rdmavt for qp support"
>> series previously submitted.
>>
>> ---
>>
>> Dennis Dalessandro (8):
>>       IB/rdmavt: Add completion queue functions
>>       IB/rdmavt: Add post send to rdmavt
>>       IB/rdmavt: Add support for tracing events
>>       IB/rdmavt: Add modify qp
>>       IB/rdmavt: Add destroy qp verb
>>       IB/rdmavt: Add post receive to rdmavt
>>       IB/rdmavt: Add multicast functions
>>       IB/rdmavt: Add misc dev register functionality
>
>I didn't see anywhere an API for RVT that handles received packets
>Did I miss it?
>If not, what is expected from the lower driver to do when a packet is received?

Similar to my response for the send side, this is the first step. When a 
packet is received it is handled by the driver. As to the details of just 
what moves into rdmavt and what stays in the driver is a matter of debate, 
and a decision which will have to take into account performance aspects. A 
perfect design from a software engineering standpoint may not always be 
possible, these are high performance devices and we can't sacrifice that 
performance.

At any rate, I'd like to focus on the send side next with the recv to 
follow.

-Denny
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html