Merge tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/dledford/rdma

Pull rdma updates from Doug Ledford:
 "Initial roundup of 4.5 merge window patches

   - Remove usage of ib_query_device and instead store attributes in
     ib_device struct

   - Move iopoll out of block and into lib, rename to irqpoll, and use
     in several places in the rdma stack as our new completion queue
     polling library mechanism.  Update the other block drivers that
     already used iopoll to use the new mechanism too.

   - Replace the per-entry GID table locks with a single GID table lock

   - IPoIB multicast cleanup

   - Cleanups to the IB MR facility

   - Add support for 64bit extended IB counters

   - Fix for netlink oops while parsing RDMA nl messages

   - RoCEv2 support for the core IB code

   - mlx4 RoCEv2 support

   - mlx5 RoCEv2 support

   - Cross Channel support for mlx5

   - Timestamp support for mlx5

   - Atomic support for mlx5

   - Raw QP support for mlx5

   - MAINTAINERS update for mlx4/mlx5

   - Misc ocrdma, qib, nes, usNIC, cxgb3, cxgb4, mlx4, mlx5 updates

   - Add support for remote invalidate to the iSER driver (pushed
     through the RDMA tree due to dependencies, acknowledged by nab)

   - Update to NFSoRDMA (pushed through the RDMA tree due to
     dependencies, acknowledged by Bruce)"

* tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/dledford/rdma: (169 commits)
  IB/mlx5: Unify CQ create flags check
  IB/mlx5: Expose Raw Packet QP to user space consumers
  {IB, net}/mlx5: Move the modify QP operation table to mlx5_ib
  IB/mlx5: Support setting Ethernet priority for Raw Packet QPs
  IB/mlx5: Add Raw Packet QP query functionality
  IB/mlx5: Add create and destroy functionality for Raw Packet QP
  IB/mlx5: Refactor mlx5_ib_qp to accommodate other QP types
  IB/mlx5: Allocate a Transport Domain for each ucontext
  net/mlx5_core: Warn on unsupported events of QP/RQ/SQ
  net/mlx5_core: Add RQ and SQ event handling
  net/mlx5_core: Export transport objects
  IB/mlx5: Expose CQE version to user-space
  IB/mlx5: Add CQE version 1 support to user QPs and SRQs
  IB/mlx5: Fix data validation in mlx5_ib_alloc_ucontext
  IB/sa: Fix netlink local service GFP crash
  IB/srpt: Remove redundant wc array
  IB/qib: Improve ipoib UD performance
  IB/mlx4: Advertise RoCE v2 support
  IB/mlx4: Create and use another QP1 for RoCEv2
  IB/mlx4: Enable send of RoCE QP1 packets with IP/UDP headers
  ...

13 files changed, 744 insertions, 259 deletions

diff --git a/net/rds/ib.c b/net/rds/ib.c
index f222885ac0c7..9481d55ff6cb 100644
--- a/net/rds/ib.c
+++ b/net/rds/ib.c
@@ -122,44 +122,34 @@ void rds_ib_dev_put(struct rds_ib_device *rds_ibdev)
 static void rds_ib_add_one(struct ib_device *device)
 {
 	struct rds_ib_device *rds_ibdev;
-	struct ib_device_attr *dev_attr;
 
 	/* Only handle IB (no iWARP) devices */
 	if (device->node_type != RDMA_NODE_IB_CA)
 		return;
 
-	dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
-	if (!dev_attr)
-		return;
-
-	if (ib_query_device(device, dev_attr)) {
-		rdsdebug("Query device failed for %s\n", device->name);
-		goto free_attr;
-	}
-
 	rds_ibdev = kzalloc_node(sizeof(struct rds_ib_device), GFP_KERNEL,
 				 ibdev_to_node(device));
 	if (!rds_ibdev)
-		goto free_attr;
+		return;
 
 	spin_lock_init(&rds_ibdev->spinlock);
 	atomic_set(&rds_ibdev->refcount, 1);
 	INIT_WORK(&rds_ibdev->free_work, rds_ib_dev_free);
 
-	rds_ibdev->max_wrs = dev_attr->max_qp_wr;
-	rds_ibdev->max_sge = min(dev_attr->max_sge, RDS_IB_MAX_SGE);
+	rds_ibdev->max_wrs = device->attrs.max_qp_wr;
+	rds_ibdev->max_sge = min(device->attrs.max_sge, RDS_IB_MAX_SGE);
 
-	rds_ibdev->fmr_max_remaps = dev_attr->max_map_per_fmr?: 32;
-	rds_ibdev->max_1m_fmrs = dev_attr->max_mr ?
-		min_t(unsigned int, (dev_attr->max_mr / 2),
+	rds_ibdev->fmr_max_remaps = device->attrs.max_map_per_fmr?: 32;
+	rds_ibdev->max_1m_fmrs = device->attrs.max_mr ?
+		min_t(unsigned int, (device->attrs.max_mr / 2),
 		      rds_ib_fmr_1m_pool_size) : rds_ib_fmr_1m_pool_size;
 
-	rds_ibdev->max_8k_fmrs = dev_attr->max_mr ?
-		min_t(unsigned int, ((dev_attr->max_mr / 2) * RDS_MR_8K_SCALE),
+	rds_ibdev->max_8k_fmrs = device->attrs.max_mr ?
+		min_t(unsigned int, ((device->attrs.max_mr / 2) * RDS_MR_8K_SCALE),
 		      rds_ib_fmr_8k_pool_size) : rds_ib_fmr_8k_pool_size;
 
-	rds_ibdev->max_initiator_depth = dev_attr->max_qp_init_rd_atom;
-	rds_ibdev->max_responder_resources = dev_attr->max_qp_rd_atom;
+	rds_ibdev->max_initiator_depth = device->attrs.max_qp_init_rd_atom;
+	rds_ibdev->max_responder_resources = device->attrs.max_qp_rd_atom;
 
 	rds_ibdev->dev = device;
 	rds_ibdev->pd = ib_alloc_pd(device);
@@ -183,7 +173,7 @@ static void rds_ib_add_one(struct ib_device *device)
 	}
 
 	rdsdebug("RDS/IB: max_mr = %d, max_wrs = %d, max_sge = %d, fmr_max_remaps = %d, max_1m_fmrs = %d, max_8k_fmrs = %d\n",
-		 dev_attr->max_fmr, rds_ibdev->max_wrs, rds_ibdev->max_sge,
+		 device->attrs.max_fmr, rds_ibdev->max_wrs, rds_ibdev->max_sge,
 		 rds_ibdev->fmr_max_remaps, rds_ibdev->max_1m_fmrs,
 		 rds_ibdev->max_8k_fmrs);
 
@@ -202,8 +192,6 @@ static void rds_ib_add_one(struct ib_device *device)
 
 put_dev:
 	rds_ib_dev_put(rds_ibdev);
-free_attr:
-	kfree(dev_attr);
 }
 
 /*
diff --git a/net/rds/iw.c b/net/rds/iw.c
index 576f1825fc55..f4a9fff829e0 100644
--- a/net/rds/iw.c
+++ b/net/rds/iw.c
@@ -60,30 +60,20 @@ LIST_HEAD(iw_nodev_conns);
 static void rds_iw_add_one(struct ib_device *device)
 {
 	struct rds_iw_device *rds_iwdev;
-	struct ib_device_attr *dev_attr;
 
 	/* Only handle iwarp devices */
 	if (device->node_type != RDMA_NODE_RNIC)
 		return;
 
-	dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
-	if (!dev_attr)
-		return;
-
-	if (ib_query_device(device, dev_attr)) {
-		rdsdebug("Query device failed for %s\n", device->name);
-		goto free_attr;
-	}
-
 	rds_iwdev = kmalloc(sizeof *rds_iwdev, GFP_KERNEL);
 	if (!rds_iwdev)
-		goto free_attr;
+		return;
 
 	spin_lock_init(&rds_iwdev->spinlock);
 
-	rds_iwdev->dma_local_lkey = !!(dev_attr->device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY);
-	rds_iwdev->max_wrs = dev_attr->max_qp_wr;
-	rds_iwdev->max_sge = min(dev_attr->max_sge, RDS_IW_MAX_SGE);
+	rds_iwdev->dma_local_lkey = !!(device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY);
+	rds_iwdev->max_wrs = device->attrs.max_qp_wr;
+	rds_iwdev->max_sge = min(device->attrs.max_sge, RDS_IW_MAX_SGE);
 
 	rds_iwdev->dev = device;
 	rds_iwdev->pd = ib_alloc_pd(device);
@@ -111,8 +101,7 @@ static void rds_iw_add_one(struct ib_device *device)
 	list_add_tail(&rds_iwdev->list, &rds_iw_devices);
 
 	ib_set_client_data(device, &rds_iw_client, rds_iwdev);
-
-	goto free_attr;
+	return;
 
 err_mr:
 	if (rds_iwdev->mr)
@@ -121,8 +110,6 @@ err_pd:
 	ib_dealloc_pd(rds_iwdev->pd);
 free_dev:
 	kfree(rds_iwdev);
-free_attr:
-	kfree(dev_attr);
 }
 
 static void rds_iw_remove_one(struct ib_device *device, void *client_data)
diff --git a/net/sunrpc/xprt.c b/net/sunrpc/xprt.c
index 2e98f4a243e5..37edea6fa92d 100644
--- a/net/sunrpc/xprt.c
+++ b/net/sunrpc/xprt.c
@@ -1425,3 +1425,4 @@ void xprt_put(struct rpc_xprt *xprt)
 	if (atomic_dec_and_test(&xprt->count))
 		xprt_destroy(xprt);
 }
+EXPORT_SYMBOL_GPL(xprt_put);
diff --git a/net/sunrpc/xprtrdma/Makefile b/net/sunrpc/xprtrdma/Makefile
index 33f99d3004f2..dc9f3b513a05 100644
--- a/net/sunrpc/xprtrdma/Makefile
+++ b/net/sunrpc/xprtrdma/Makefile
@@ -2,7 +2,7 @@ obj-$(CONFIG_SUNRPC_XPRT_RDMA) += rpcrdma.o
 
 rpcrdma-y := transport.o rpc_rdma.o verbs.o \
 	fmr_ops.o frwr_ops.o physical_ops.o \
-	svc_rdma.o svc_rdma_transport.o \
+	svc_rdma.o svc_rdma_backchannel.o svc_rdma_transport.o \
 	svc_rdma_marshal.o svc_rdma_sendto.o svc_rdma_recvfrom.o \
 	module.o
 rpcrdma-$(CONFIG_SUNRPC_BACKCHANNEL) += backchannel.o
diff --git a/net/sunrpc/xprtrdma/frwr_ops.c b/net/sunrpc/xprtrdma/frwr_ops.c
index c6836844bd0e..e16567389e28 100644
--- a/net/sunrpc/xprtrdma/frwr_ops.c
+++ b/net/sunrpc/xprtrdma/frwr_ops.c
@@ -190,12 +190,11 @@ static int
 frwr_op_open(struct rpcrdma_ia *ia, struct rpcrdma_ep *ep,
 	     struct rpcrdma_create_data_internal *cdata)
 {
-	struct ib_device_attr *devattr = &ia->ri_devattr;
 	int depth, delta;
 
 	ia->ri_max_frmr_depth =
 			min_t(unsigned int, RPCRDMA_MAX_DATA_SEGS,
-			      devattr->max_fast_reg_page_list_len);
+			      ia->ri_device->attrs.max_fast_reg_page_list_len);
 	dprintk("RPC:       %s: device's max FR page list len = %u\n",
 		__func__, ia->ri_max_frmr_depth);
 
@@ -222,8 +221,8 @@ frwr_op_open(struct rpcrdma_ia *ia, struct rpcrdma_ep *ep,
 	}
 
 	ep->rep_attr.cap.max_send_wr *= depth;
-	if (ep->rep_attr.cap.max_send_wr > devattr->max_qp_wr) {
-		cdata->max_requests = devattr->max_qp_wr / depth;
+	if (ep->rep_attr.cap.max_send_wr > ia->ri_device->attrs.max_qp_wr) {
+		cdata->max_requests = ia->ri_device->attrs.max_qp_wr / depth;
 		if (!cdata->max_requests)
 			return -EINVAL;
 		ep->rep_attr.cap.max_send_wr = cdata->max_requests *
diff --git a/net/sunrpc/xprtrdma/svc_rdma.c b/net/sunrpc/xprtrdma/svc_rdma.c
index 1b7051bdbdc8..c846ca9f1eba 100644
--- a/net/sunrpc/xprtrdma/svc_rdma.c
+++ b/net/sunrpc/xprtrdma/svc_rdma.c
@@ -55,6 +55,7 @@ unsigned int svcrdma_ord = RPCRDMA_ORD;
 static unsigned int min_ord = 1;
 static unsigned int max_ord = 4096;
 unsigned int svcrdma_max_requests = RPCRDMA_MAX_REQUESTS;
+unsigned int svcrdma_max_bc_requests = RPCRDMA_MAX_BC_REQUESTS;
 static unsigned int min_max_requests = 4;
 static unsigned int max_max_requests = 16384;
 unsigned int svcrdma_max_req_size = RPCRDMA_MAX_REQ_SIZE;
@@ -71,10 +72,6 @@ atomic_t rdma_stat_rq_prod;
 atomic_t rdma_stat_sq_poll;
 atomic_t rdma_stat_sq_prod;
 
-/* Temporary NFS request map and context caches */
-struct kmem_cache *svc_rdma_map_cachep;
-struct kmem_cache *svc_rdma_ctxt_cachep;
-
 struct workqueue_struct *svc_rdma_wq;
 
 /*
@@ -243,17 +240,16 @@ void svc_rdma_cleanup(void)
 	svc_unreg_xprt_class(&svc_rdma_bc_class);
 #endif
 	svc_unreg_xprt_class(&svc_rdma_class);
-	kmem_cache_destroy(svc_rdma_map_cachep);
-	kmem_cache_destroy(svc_rdma_ctxt_cachep);
 }
 
 int svc_rdma_init(void)
 {
 	dprintk("SVCRDMA Module Init, register RPC RDMA transport\n");
 	dprintk("\tsvcrdma_ord      : %d\n", svcrdma_ord);
-	dprintk("\tmax_requests     : %d\n", svcrdma_max_requests);
-	dprintk("\tsq_depth         : %d\n",
+	dprintk("\tmax_requests     : %u\n", svcrdma_max_requests);
+	dprintk("\tsq_depth         : %u\n",
 		svcrdma_max_requests * RPCRDMA_SQ_DEPTH_MULT);
+	dprintk("\tmax_bc_requests  : %u\n", svcrdma_max_bc_requests);
 	dprintk("\tmax_inline       : %d\n", svcrdma_max_req_size);
 
 	svc_rdma_wq = alloc_workqueue("svc_rdma", 0, 0);
@@ -264,39 +260,10 @@ int svc_rdma_init(void)
 		svcrdma_table_header =
 			register_sysctl_table(svcrdma_root_table);
 
-	/* Create the temporary map cache */
-	svc_rdma_map_cachep = kmem_cache_create("svc_rdma_map_cache",
-						sizeof(struct svc_rdma_req_map),
-						0,
-						SLAB_HWCACHE_ALIGN,
-						NULL);
-	if (!svc_rdma_map_cachep) {
-		printk(KERN_INFO "Could not allocate map cache.\n");
-		goto err0;
-	}
-
-	/* Create the temporary context cache */
-	svc_rdma_ctxt_cachep =
-		kmem_cache_create("svc_rdma_ctxt_cache",
-				  sizeof(struct svc_rdma_op_ctxt),
-				  0,
-				  SLAB_HWCACHE_ALIGN,
-				  NULL);
-	if (!svc_rdma_ctxt_cachep) {
-		printk(KERN_INFO "Could not allocate WR ctxt cache.\n");
-		goto err1;
-	}
-
 	/* Register RDMA with the SVC transport switch */
 	svc_reg_xprt_class(&svc_rdma_class);
 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
 	svc_reg_xprt_class(&svc_rdma_bc_class);
 #endif
 	return 0;
- err1:
-	kmem_cache_destroy(svc_rdma_map_cachep);
- err0:
-	unregister_sysctl_table(svcrdma_table_header);
-	destroy_workqueue(svc_rdma_wq);
-	return -ENOMEM;
 }
diff --git a/net/sunrpc/xprtrdma/svc_rdma_backchannel.c b/net/sunrpc/xprtrdma/svc_rdma_backchannel.c
new file mode 100644
index 000000000000..65a7c232a345
--- /dev/null
+++ b/net/sunrpc/xprtrdma/svc_rdma_backchannel.c
@@ -0,0 +1,371 @@
+/*
+ * Copyright (c) 2015 Oracle.  All rights reserved.
+ *
+ * Support for backward direction RPCs on RPC/RDMA (server-side).
+ */
+
+#include <linux/sunrpc/svc_rdma.h>
+#include "xprt_rdma.h"
+
+#define RPCDBG_FACILITY	RPCDBG_SVCXPRT
+
+#undef SVCRDMA_BACKCHANNEL_DEBUG
+
+int svc_rdma_handle_bc_reply(struct rpc_xprt *xprt, struct rpcrdma_msg *rmsgp,
+			     struct xdr_buf *rcvbuf)
+{
+	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
+	struct kvec *dst, *src = &rcvbuf->head[0];
+	struct rpc_rqst *req;
+	unsigned long cwnd;
+	u32 credits;
+	size_t len;
+	__be32 xid;
+	__be32 *p;
+	int ret;
+
+	p = (__be32 *)src->iov_base;
+	len = src->iov_len;
+	xid = rmsgp->rm_xid;
+
+#ifdef SVCRDMA_BACKCHANNEL_DEBUG
+	pr_info("%s: xid=%08x, length=%zu\n",
+		__func__, be32_to_cpu(xid), len);
+	pr_info("%s: RPC/RDMA: %*ph\n",
+		__func__, (int)RPCRDMA_HDRLEN_MIN, rmsgp);
+	pr_info("%s:      RPC: %*ph\n",
+		__func__, (int)len, p);
+#endif
+
+	ret = -EAGAIN;
+	if (src->iov_len < 24)
+		goto out_shortreply;
+
+	spin_lock_bh(&xprt->transport_lock);
+	req = xprt_lookup_rqst(xprt, xid);
+	if (!req)
+		goto out_notfound;
+
+	dst = &req->rq_private_buf.head[0];
+	memcpy(&req->rq_private_buf, &req->rq_rcv_buf, sizeof(struct xdr_buf));
+	if (dst->iov_len < len)
+		goto out_unlock;
+	memcpy(dst->iov_base, p, len);
+
+	credits = be32_to_cpu(rmsgp->rm_credit);
+	if (credits == 0)
+		credits = 1;	/* don't deadlock */
+	else if (credits > r_xprt->rx_buf.rb_bc_max_requests)
+		credits = r_xprt->rx_buf.rb_bc_max_requests;
+
+	cwnd = xprt->cwnd;
+	xprt->cwnd = credits << RPC_CWNDSHIFT;
+	if (xprt->cwnd > cwnd)
+		xprt_release_rqst_cong(req->rq_task);
+
+	ret = 0;
+	xprt_complete_rqst(req->rq_task, rcvbuf->len);
+	rcvbuf->len = 0;
+
+out_unlock:
+	spin_unlock_bh(&xprt->transport_lock);
+out:
+	return ret;
+
+out_shortreply:
+	dprintk("svcrdma: short bc reply: xprt=%p, len=%zu\n",
+		xprt, src->iov_len);
+	goto out;
+
+out_notfound:
+	dprintk("svcrdma: unrecognized bc reply: xprt=%p, xid=%08x\n",
+		xprt, be32_to_cpu(xid));
+
+	goto out_unlock;
+}
+
+/* Send a backwards direction RPC call.
+ *
+ * Caller holds the connection's mutex and has already marshaled
+ * the RPC/RDMA request.
+ *
+ * This is similar to svc_rdma_reply, but takes an rpc_rqst
+ * instead, does not support chunks, and avoids blocking memory
+ * allocation.
+ *
+ * XXX: There is still an opportunity to block in svc_rdma_send()
+ * if there are no SQ entries to post the Send. This may occur if
+ * the adapter has a small maximum SQ depth.
+ */
+static int svc_rdma_bc_sendto(struct svcxprt_rdma *rdma,
+			      struct rpc_rqst *rqst)
+{
+	struct xdr_buf *sndbuf = &rqst->rq_snd_buf;
+	struct svc_rdma_op_ctxt *ctxt;
+	struct svc_rdma_req_map *vec;
+	struct ib_send_wr send_wr;
+	int ret;
+
+	vec = svc_rdma_get_req_map(rdma);
+	ret = svc_rdma_map_xdr(rdma, sndbuf, vec);
+	if (ret)
+		goto out_err;
+
+	/* Post a recv buffer to handle the reply for this request. */
+	ret = svc_rdma_post_recv(rdma, GFP_NOIO);
+	if (ret) {
+		pr_err("svcrdma: Failed to post bc receive buffer, err=%d.\n",
+		       ret);
+		pr_err("svcrdma: closing transport %p.\n", rdma);
+		set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags);
+		ret = -ENOTCONN;
+		goto out_err;
+	}
+
+	ctxt = svc_rdma_get_context(rdma);
+	ctxt->pages[0] = virt_to_page(rqst->rq_buffer);
+	ctxt->count = 1;
+
+	ctxt->wr_op = IB_WR_SEND;
+	ctxt->direction = DMA_TO_DEVICE;
+	ctxt->sge[0].lkey = rdma->sc_pd->local_dma_lkey;
+	ctxt->sge[0].length = sndbuf->len;
+	ctxt->sge[0].addr =
+	    ib_dma_map_page(rdma->sc_cm_id->device, ctxt->pages[0], 0,
+			    sndbuf->len, DMA_TO_DEVICE);
+	if (ib_dma_mapping_error(rdma->sc_cm_id->device, ctxt->sge[0].addr)) {
+		ret = -EIO;
+		goto out_unmap;
+	}
+	atomic_inc(&rdma->sc_dma_used);
+
+	memset(&send_wr, 0, sizeof(send_wr));
+	send_wr.wr_id = (unsigned long)ctxt;
+	send_wr.sg_list = ctxt->sge;
+	send_wr.num_sge = 1;
+	send_wr.opcode = IB_WR_SEND;
+	send_wr.send_flags = IB_SEND_SIGNALED;
+
+	ret = svc_rdma_send(rdma, &send_wr);
+	if (ret) {
+		ret = -EIO;
+		goto out_unmap;
+	}
+
+out_err:
+	svc_rdma_put_req_map(rdma, vec);
+	dprintk("svcrdma: %s returns %d\n", __func__, ret);
+	return ret;
+
+out_unmap:
+	svc_rdma_unmap_dma(ctxt);
+	svc_rdma_put_context(ctxt, 1);
+	goto out_err;
+}
+
+/* Server-side transport endpoint wants a whole page for its send
+ * buffer. The client RPC code constructs the RPC header in this
+ * buffer before it invokes ->send_request.
+ *
+ * Returns NULL if there was a temporary allocation failure.
+ */
+static void *
+xprt_rdma_bc_allocate(struct rpc_task *task, size_t size)
+{
+	struct rpc_rqst *rqst = task->tk_rqstp;
+	struct svc_xprt *sxprt = rqst->rq_xprt->bc_xprt;
+	struct svcxprt_rdma *rdma;
+	struct page *page;
+
+	rdma = container_of(sxprt, struct svcxprt_rdma, sc_xprt);
+
+	/* Prevent an infinite loop: try to make this case work */
+	if (size > PAGE_SIZE)
+		WARN_ONCE(1, "svcrdma: large bc buffer request (size %zu)\n",
+			  size);
+
+	page = alloc_page(RPCRDMA_DEF_GFP);
+	if (!page)
+		return NULL;
+
+	return page_address(page);
+}
+
+static void
+xprt_rdma_bc_free(void *buffer)
+{
+	/* No-op: ctxt and page have already been freed. */
+}
+
+static int
+rpcrdma_bc_send_request(struct svcxprt_rdma *rdma, struct rpc_rqst *rqst)
+{
+	struct rpc_xprt *xprt = rqst->rq_xprt;
+	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
+	struct rpcrdma_msg *headerp = (struct rpcrdma_msg *)rqst->rq_buffer;
+	int rc;
+
+	/* Space in the send buffer for an RPC/RDMA header is reserved
+	 * via xprt->tsh_size.
+	 */
+	headerp->rm_xid = rqst->rq_xid;
+	headerp->rm_vers = rpcrdma_version;
+	headerp->rm_credit = cpu_to_be32(r_xprt->rx_buf.rb_bc_max_requests);
+	headerp->rm_type = rdma_msg;
+	headerp->rm_body.rm_chunks[0] = xdr_zero;
+	headerp->rm_body.rm_chunks[1] = xdr_zero;
+	headerp->rm_body.rm_chunks[2] = xdr_zero;
+
+#ifdef SVCRDMA_BACKCHANNEL_DEBUG
+	pr_info("%s: %*ph\n", __func__, 64, rqst->rq_buffer);
+#endif
+
+	rc = svc_rdma_bc_sendto(rdma, rqst);
+	if (rc)
+		goto drop_connection;
+	return rc;
+
+drop_connection:
+	dprintk("svcrdma: failed to send bc call\n");
+	xprt_disconnect_done(xprt);
+	return -ENOTCONN;
+}
+
+/* Send an RPC call on the passive end of a transport
+ * connection.
+ */
+static int
+xprt_rdma_bc_send_request(struct rpc_task *task)
+{
+	struct rpc_rqst *rqst = task->tk_rqstp;
+	struct svc_xprt *sxprt = rqst->rq_xprt->bc_xprt;
+	struct svcxprt_rdma *rdma;
+	int ret;
+
+	dprintk("svcrdma: sending bc call with xid: %08x\n",
+		be32_to_cpu(rqst->rq_xid));
+
+	if (!mutex_trylock(&sxprt->xpt_mutex)) {
+		rpc_sleep_on(&sxprt->xpt_bc_pending, task, NULL);
+		if (!mutex_trylock(&sxprt->xpt_mutex))
+			return -EAGAIN;
+		rpc_wake_up_queued_task(&sxprt->xpt_bc_pending, task);
+	}
+
+	ret = -ENOTCONN;
+	rdma = container_of(sxprt, struct svcxprt_rdma, sc_xprt);
+	if (!test_bit(XPT_DEAD, &sxprt->xpt_flags))
+		ret = rpcrdma_bc_send_request(rdma, rqst);
+
+	mutex_unlock(&sxprt->xpt_mutex);
+
+	if (ret < 0)
+		return ret;
+	return 0;
+}
+
+static void
+xprt_rdma_bc_close(struct rpc_xprt *xprt)
+{
+	dprintk("svcrdma: %s: xprt %p\n", __func__, xprt);
+}
+
+static void
+xprt_rdma_bc_put(struct rpc_xprt *xprt)
+{
+	dprintk("svcrdma: %s: xprt %p\n", __func__, xprt);
+
+	xprt_free(xprt);
+	module_put(THIS_MODULE);
+}
+
+static struct rpc_xprt_ops xprt_rdma_bc_procs = {
+	.reserve_xprt		= xprt_reserve_xprt_cong,
+	.release_xprt		= xprt_release_xprt_cong,
+	.alloc_slot		= xprt_alloc_slot,
+	.release_request	= xprt_release_rqst_cong,
+	.buf_alloc		= xprt_rdma_bc_allocate,
+	.buf_free		= xprt_rdma_bc_free,
+	.send_request		= xprt_rdma_bc_send_request,
+	.set_retrans_timeout	= xprt_set_retrans_timeout_def,
+	.close			= xprt_rdma_bc_close,
+	.destroy		= xprt_rdma_bc_put,
+	.print_stats		= xprt_rdma_print_stats
+};
+
+static const struct rpc_timeout xprt_rdma_bc_timeout = {
+	.to_initval = 60 * HZ,
+	.to_maxval = 60 * HZ,
+};
+
+/* It shouldn't matter if the number of backchannel session slots
+ * doesn't match the number of RPC/RDMA credits. That just means
+ * one or the other will have extra slots that aren't used.
+ */
+static struct rpc_xprt *
+xprt_setup_rdma_bc(struct xprt_create *args)
+{
+	struct rpc_xprt *xprt;
+	struct rpcrdma_xprt *new_xprt;
+
+	if (args->addrlen > sizeof(xprt->addr)) {
+		dprintk("RPC:       %s: address too large\n", __func__);
+		return ERR_PTR(-EBADF);
+	}
+
+	xprt = xprt_alloc(args->net, sizeof(*new_xprt),
+			  RPCRDMA_MAX_BC_REQUESTS,
+			  RPCRDMA_MAX_BC_REQUESTS);
+	if (!xprt) {
+		dprintk("RPC:       %s: couldn't allocate rpc_xprt\n",
+			__func__);
+		return ERR_PTR(-ENOMEM);
+	}
+
+	xprt->timeout = &xprt_rdma_bc_timeout;
+	xprt_set_bound(xprt);
+	xprt_set_connected(xprt);
+	xprt->bind_timeout = RPCRDMA_BIND_TO;
+	xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
+	xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
+
+	xprt->prot = XPRT_TRANSPORT_BC_RDMA;
+	xprt->tsh_size = RPCRDMA_HDRLEN_MIN / sizeof(__be32);
+	xprt->ops = &xprt_rdma_bc_procs;
+
+	memcpy(&xprt->addr, args->dstaddr, args->addrlen);
+	xprt->addrlen = args->addrlen;
+	xprt_rdma_format_addresses(xprt, (struct sockaddr *)&xprt->addr);
+	xprt->resvport = 0;
+
+	xprt->max_payload = xprt_rdma_max_inline_read;
+
+	new_xprt = rpcx_to_rdmax(xprt);
+	new_xprt->rx_buf.rb_bc_max_requests = xprt->max_reqs;
+
+	xprt_get(xprt);
+	args->bc_xprt->xpt_bc_xprt = xprt;
+	xprt->bc_xprt = args->bc_xprt;
+
+	if (!try_module_get(THIS_MODULE))
+		goto out_fail;
+
+	/* Final put for backchannel xprt is in __svc_rdma_free */
+	xprt_get(xprt);
+	return xprt;
+
+out_fail:
+	xprt_rdma_free_addresses(xprt);
+	args->bc_xprt->xpt_bc_xprt = NULL;
+	xprt_put(xprt);
+	xprt_free(xprt);
+	return ERR_PTR(-EINVAL);
+}
+
+struct xprt_class xprt_rdma_bc = {
+	.list			= LIST_HEAD_INIT(xprt_rdma_bc.list),
+	.name			= "rdma backchannel",
+	.owner			= THIS_MODULE,
+	.ident			= XPRT_TRANSPORT_BC_RDMA,
+	.setup			= xprt_setup_rdma_bc,
+};
diff --git a/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c b/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c
index ff4f01e527ec..c8b8a8b4181e 100644
--- a/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c
+++ b/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c
@@ -144,6 +144,7 @@ int rdma_read_chunk_lcl(struct svcxprt_rdma *xprt,
 
 		head->arg.pages[pg_no] = rqstp->rq_arg.pages[pg_no];
 		head->arg.page_len += len;
+
 		head->arg.len += len;
 		if (!pg_off)
 			head->count++;
@@ -160,8 +161,7 @@ int rdma_read_chunk_lcl(struct svcxprt_rdma *xprt,
 			goto err;
 		atomic_inc(&xprt->sc_dma_used);
 
-		/* The lkey here is either a local dma lkey or a dma_mr lkey */
-		ctxt->sge[pno].lkey = xprt->sc_dma_lkey;
+		ctxt->sge[pno].lkey = xprt->sc_pd->local_dma_lkey;
 		ctxt->sge[pno].length = len;
 		ctxt->count++;
 
@@ -567,6 +567,38 @@ static int rdma_read_complete(struct svc_rqst *rqstp,
 	return ret;
 }
 
+/* By convention, backchannel calls arrive via rdma_msg type
+ * messages, and never populate the chunk lists. This makes
+ * the RPC/RDMA header small and fixed in size, so it is
+ * straightforward to check the RPC header's direction field.
+ */
+static bool
+svc_rdma_is_backchannel_reply(struct svc_xprt *xprt, struct rpcrdma_msg *rmsgp)
+{
+	__be32 *p = (__be32 *)rmsgp;
+
+	if (!xprt->xpt_bc_xprt)
+		return false;
+
+	if (rmsgp->rm_type != rdma_msg)
+		return false;
+	if (rmsgp->rm_body.rm_chunks[0] != xdr_zero)
+		return false;
+	if (rmsgp->rm_body.rm_chunks[1] != xdr_zero)
+		return false;
+	if (rmsgp->rm_body.rm_chunks[2] != xdr_zero)
+		return false;
+
+	/* sanity */
+	if (p[7] != rmsgp->rm_xid)
+		return false;
+	/* call direction */
+	if (p[8] == cpu_to_be32(RPC_CALL))
+		return false;
+
+	return true;
+}
+
 /*
  * Set up the rqstp thread context to point to the RQ buffer. If
  * necessary, pull additional data from the client with an RDMA_READ
@@ -632,6 +664,15 @@ int svc_rdma_recvfrom(struct svc_rqst *rqstp)
 		goto close_out;
 	}
 
+	if (svc_rdma_is_backchannel_reply(xprt, rmsgp)) {
+		ret = svc_rdma_handle_bc_reply(xprt->xpt_bc_xprt, rmsgp,
+					       &rqstp->rq_arg);
+		svc_rdma_put_context(ctxt, 0);
+		if (ret)
+			goto repost;
+		return ret;
+	}
+
 	/* Read read-list data. */
 	ret = rdma_read_chunks(rdma_xprt, rmsgp, rqstp, ctxt);
 	if (ret > 0) {
@@ -668,4 +709,15 @@ int svc_rdma_recvfrom(struct svc_rqst *rqstp)
 	set_bit(XPT_CLOSE, &xprt->xpt_flags);
 defer:
 	return 0;
+
+repost:
+	ret = svc_rdma_post_recv(rdma_xprt, GFP_KERNEL);
+	if (ret) {
+		pr_err("svcrdma: could not post a receive buffer, err=%d.\n",
+		       ret);
+		pr_err("svcrdma: closing transport %p.\n", rdma_xprt);
+		set_bit(XPT_CLOSE, &rdma_xprt->sc_xprt.xpt_flags);
+		ret = -ENOTCONN;
+	}
+	return ret;
 }
diff --git a/net/sunrpc/xprtrdma/svc_rdma_sendto.c b/net/sunrpc/xprtrdma/svc_rdma_sendto.c
index 969a1ab75fc3..df57f3ce6cd2 100644
--- a/net/sunrpc/xprtrdma/svc_rdma_sendto.c
+++ b/net/sunrpc/xprtrdma/svc_rdma_sendto.c
@@ -50,9 +50,9 @@
 
 #define RPCDBG_FACILITY	RPCDBG_SVCXPRT
 
-static int map_xdr(struct svcxprt_rdma *xprt,
-		   struct xdr_buf *xdr,
-		   struct svc_rdma_req_map *vec)
+int svc_rdma_map_xdr(struct svcxprt_rdma *xprt,
+		     struct xdr_buf *xdr,
+		     struct svc_rdma_req_map *vec)
 {
 	int sge_no;
 	u32 sge_bytes;
@@ -62,7 +62,7 @@ static int map_xdr(struct svcxprt_rdma *xprt,
 
 	if (xdr->len !=
 	    (xdr->head[0].iov_len + xdr->page_len + xdr->tail[0].iov_len)) {
-		pr_err("svcrdma: map_xdr: XDR buffer length error\n");
+		pr_err("svcrdma: %s: XDR buffer length error\n", __func__);
 		return -EIO;
 	}
 
@@ -97,9 +97,9 @@ static int map_xdr(struct svcxprt_rdma *xprt,
 		sge_no++;
 	}
 
-	dprintk("svcrdma: map_xdr: sge_no %d page_no %d "
+	dprintk("svcrdma: %s: sge_no %d page_no %d "
 		"page_base %u page_len %u head_len %zu tail_len %zu\n",
-		sge_no, page_no, xdr->page_base, xdr->page_len,
+		__func__, sge_no, page_no, xdr->page_base, xdr->page_len,
 		xdr->head[0].iov_len, xdr->tail[0].iov_len);
 
 	vec->count = sge_no;
@@ -265,7 +265,7 @@ static int send_write(struct svcxprt_rdma *xprt, struct svc_rqst *rqstp,
 					 sge[sge_no].addr))
 			goto err;
 		atomic_inc(&xprt->sc_dma_used);
-		sge[sge_no].lkey = xprt->sc_dma_lkey;
+		sge[sge_no].lkey = xprt->sc_pd->local_dma_lkey;
 		ctxt->count++;
 		sge_off = 0;
 		sge_no++;
@@ -465,7 +465,7 @@ static int send_reply(struct svcxprt_rdma *rdma,
 	int ret;
 
 	/* Post a recv buffer to handle another request. */
-	ret = svc_rdma_post_recv(rdma);
+	ret = svc_rdma_post_recv(rdma, GFP_KERNEL);
 	if (ret) {
 		printk(KERN_INFO
 		       "svcrdma: could not post a receive buffer, err=%d."
@@ -480,7 +480,7 @@ static int send_reply(struct svcxprt_rdma *rdma,
 	ctxt->count = 1;
 
 	/* Prepare the SGE for the RPCRDMA Header */
-	ctxt->sge[0].lkey = rdma->sc_dma_lkey;
+	ctxt->sge[0].lkey = rdma->sc_pd->local_dma_lkey;
 	ctxt->sge[0].length = svc_rdma_xdr_get_reply_hdr_len(rdma_resp);
 	ctxt->sge[0].addr =
 	    ib_dma_map_page(rdma->sc_cm_id->device, page, 0,
@@ -504,7 +504,7 @@ static int send_reply(struct svcxprt_rdma *rdma,
 					 ctxt->sge[sge_no].addr))
 			goto err;
 		atomic_inc(&rdma->sc_dma_used);
-		ctxt->sge[sge_no].lkey = rdma->sc_dma_lkey;
+		ctxt->sge[sge_no].lkey = rdma->sc_pd->local_dma_lkey;
 		ctxt->sge[sge_no].length = sge_bytes;
 	}
 	if (byte_count != 0) {
@@ -591,14 +591,17 @@ int svc_rdma_sendto(struct svc_rqst *rqstp)
 	/* Build an req vec for the XDR */
 	ctxt = svc_rdma_get_context(rdma);
 	ctxt->direction = DMA_TO_DEVICE;
-	vec = svc_rdma_get_req_map();
-	ret = map_xdr(rdma, &rqstp->rq_res, vec);
+	vec = svc_rdma_get_req_map(rdma);
+	ret = svc_rdma_map_xdr(rdma, &rqstp->rq_res, vec);
 	if (ret)
 		goto err0;
 	inline_bytes = rqstp->rq_res.len;
 
 	/* Create the RDMA response header */
-	res_page = alloc_page(GFP_KERNEL | __GFP_NOFAIL);
+	ret = -ENOMEM;
+	res_page = alloc_page(GFP_KERNEL);
+	if (!res_page)
+		goto err0;
 	rdma_resp = page_address(res_page);
 	reply_ary = svc_rdma_get_reply_array(rdma_argp);
 	if (reply_ary)
@@ -630,14 +633,14 @@ int svc_rdma_sendto(struct svc_rqst *rqstp)
 
 	ret = send_reply(rdma, rqstp, res_page, rdma_resp, ctxt, vec,
 			 inline_bytes);
-	svc_rdma_put_req_map(vec);
+	svc_rdma_put_req_map(rdma, vec);
 	dprintk("svcrdma: send_reply returns %d\n", ret);
 	return ret;
 
  err1:
 	put_page(res_page);
  err0:
-	svc_rdma_put_req_map(vec);
+	svc_rdma_put_req_map(rdma, vec);
 	svc_rdma_put_context(ctxt, 0);
 	return ret;
 }
diff --git a/net/sunrpc/xprtrdma/svc_rdma_transport.c b/net/sunrpc/xprtrdma/svc_rdma_transport.c
index b348b4adef29..5763825d09bf 100644
--- a/net/sunrpc/xprtrdma/svc_rdma_transport.c
+++ b/net/sunrpc/xprtrdma/svc_rdma_transport.c
@@ -153,18 +153,76 @@ static void svc_rdma_bc_free(struct svc_xprt *xprt)
 }
 #endif	/* CONFIG_SUNRPC_BACKCHANNEL */
 
-struct svc_rdma_op_ctxt *svc_rdma_get_context(struct svcxprt_rdma *xprt)
+static struct svc_rdma_op_ctxt *alloc_ctxt(struct svcxprt_rdma *xprt,
+					   gfp_t flags)
 {
 	struct svc_rdma_op_ctxt *ctxt;
 
-	ctxt = kmem_cache_alloc(svc_rdma_ctxt_cachep,
-				GFP_KERNEL | __GFP_NOFAIL);
-	ctxt->xprt = xprt;
-	INIT_LIST_HEAD(&ctxt->dto_q);
+	ctxt = kmalloc(sizeof(*ctxt), flags);
+	if (ctxt) {
+		ctxt->xprt = xprt;
+		INIT_LIST_HEAD(&ctxt->free);
+		INIT_LIST_HEAD(&ctxt->dto_q);
+	}
+	return ctxt;
+}
+
+static bool svc_rdma_prealloc_ctxts(struct svcxprt_rdma *xprt)
+{
+	unsigned int i;
+
+	/* Each RPC/RDMA credit can consume a number of send
+	 * and receive WQEs. One ctxt is allocated for each.
+	 */
+	i = xprt->sc_sq_depth + xprt->sc_rq_depth;
+
+	while (i--) {
+		struct svc_rdma_op_ctxt *ctxt;
+
+		ctxt = alloc_ctxt(xprt, GFP_KERNEL);
+		if (!ctxt) {
+			dprintk("svcrdma: No memory for RDMA ctxt\n");
+			return false;
+		}
+		list_add(&ctxt->free, &xprt->sc_ctxts);
+	}
+	return true;
+}
+
+struct svc_rdma_op_ctxt *svc_rdma_get_context(struct svcxprt_rdma *xprt)
+{
+	struct svc_rdma_op_ctxt *ctxt = NULL;
+
+	spin_lock_bh(&xprt->sc_ctxt_lock);
+	xprt->sc_ctxt_used++;
+	if (list_empty(&xprt->sc_ctxts))
+		goto out_empty;
+
+	ctxt = list_first_entry(&xprt->sc_ctxts,
+				struct svc_rdma_op_ctxt, free);
+	list_del_init(&ctxt->free);
+	spin_unlock_bh(&xprt->sc_ctxt_lock);
+
+out:
 	ctxt->count = 0;
 	ctxt->frmr = NULL;
-	atomic_inc(&xprt->sc_ctxt_used);
 	return ctxt;
+
+out_empty:
+	/* Either pre-allocation missed the mark, or send
+	 * queue accounting is broken.
+	 */
+	spin_unlock_bh(&xprt->sc_ctxt_lock);
+
+	ctxt = alloc_ctxt(xprt, GFP_NOIO);
+	if (ctxt)
+		goto out;
+
+	spin_lock_bh(&xprt->sc_ctxt_lock);
+	xprt->sc_ctxt_used--;
+	spin_unlock_bh(&xprt->sc_ctxt_lock);
+	WARN_ONCE(1, "svcrdma: empty RDMA ctxt list?\n");
+	return NULL;
 }
 
 void svc_rdma_unmap_dma(struct svc_rdma_op_ctxt *ctxt)
@@ -174,11 +232,11 @@ void svc_rdma_unmap_dma(struct svc_rdma_op_ctxt *ctxt)
 	for (i = 0; i < ctxt->count && ctxt->sge[i].length; i++) {
 		/*
 		 * Unmap the DMA addr in the SGE if the lkey matches
-		 * the sc_dma_lkey, otherwise, ignore it since it is
+		 * the local_dma_lkey, otherwise, ignore it since it is
 		 * an FRMR lkey and will be unmapped later when the
 		 * last WR that us