RPCRDMA: rpc rdma transport switch

This implements the configuration and building of the core transport
switch implementation of the rpcrdma transport. Stubs are provided for
the rpcrdma protocol handling, and the infiniband/iwarp verbs interface.
These are provided in following patches.

Signed-off-by: Tom Talpey <talpey@netapp.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>

6 files changed, 1180 insertions, 0 deletions

diff --git a/net/sunrpc/Makefile b/net/sunrpc/Makefile
index 8ebfc4db7f51..5c69a725e530 100644
--- a/net/sunrpc/Makefile
+++ b/net/sunrpc/Makefile
@@ -5,6 +5,7 @@
 
 obj-$(CONFIG_SUNRPC) += sunrpc.o
 obj-$(CONFIG_SUNRPC_GSS) += auth_gss/
+obj-$(CONFIG_SUNRPC_XPRT_RDMA) += xprtrdma/
 
 sunrpc-y := clnt.o xprt.o socklib.o xprtsock.o sched.o \
 	    auth.o auth_null.o auth_unix.o \
diff --git a/net/sunrpc/xprtrdma/Makefile b/net/sunrpc/xprtrdma/Makefile
new file mode 100644
index 000000000000..264f0feeb513
--- /dev/null
+++ b/net/sunrpc/xprtrdma/Makefile
@@ -0,0 +1,3 @@
+obj-$(CONFIG_SUNRPC_XPRT_RDMA) += xprtrdma.o
+
+xprtrdma-y := transport.o rpc_rdma.o verbs.o
diff --git a/net/sunrpc/xprtrdma/rpc_rdma.c b/net/sunrpc/xprtrdma/rpc_rdma.c
new file mode 100644
index 000000000000..b0587f3a5d77
--- /dev/null
+++ b/net/sunrpc/xprtrdma/rpc_rdma.c
@@ -0,0 +1,9 @@
+/*
+ * Placeholders for subsequent patches
+ */
+
+#include "xprt_rdma.h"
+
+void rpcrdma_conn_func(struct rpcrdma_ep *a) { }
+void rpcrdma_reply_handler(struct rpcrdma_rep *a) { }
+int rpcrdma_marshal_req(struct rpc_rqst *a) { return EINVAL; }
diff --git a/net/sunrpc/xprtrdma/transport.c b/net/sunrpc/xprtrdma/transport.c
new file mode 100644
index 000000000000..dc55cc974c90
--- /dev/null
+++ b/net/sunrpc/xprtrdma/transport.c
@@ -0,0 +1,800 @@
+/*
+ * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the BSD-type
+ * license below:
+ *
+ * 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 the Network Appliance, Inc. 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.
+ */
+
+/*
+ * transport.c
+ *
+ * This file contains the top-level implementation of an RPC RDMA
+ * transport.
+ *
+ * Naming convention: functions beginning with xprt_ are part of the
+ * transport switch. All others are RPC RDMA internal.
+ */
+
+#include <linux/module.h>
+#include <linux/init.h>
+#include <linux/seq_file.h>
+
+#include "xprt_rdma.h"
+
+#ifdef RPC_DEBUG
+# define RPCDBG_FACILITY	RPCDBG_TRANS
+#endif
+
+MODULE_LICENSE("Dual BSD/GPL");
+
+MODULE_DESCRIPTION("RPC/RDMA Transport for Linux kernel NFS");
+MODULE_AUTHOR("Network Appliance, Inc.");
+
+/*
+ * tunables
+ */
+
+static unsigned int xprt_rdma_slot_table_entries = RPCRDMA_DEF_SLOT_TABLE;
+static unsigned int xprt_rdma_max_inline_read = RPCRDMA_DEF_INLINE;
+static unsigned int xprt_rdma_max_inline_write = RPCRDMA_DEF_INLINE;
+static unsigned int xprt_rdma_inline_write_padding;
+#if !RPCRDMA_PERSISTENT_REGISTRATION
+static unsigned int xprt_rdma_memreg_strategy = RPCRDMA_REGISTER; /* FMR? */
+#else
+static unsigned int xprt_rdma_memreg_strategy = RPCRDMA_ALLPHYSICAL;
+#endif
+
+#ifdef RPC_DEBUG
+
+static unsigned int min_slot_table_size = RPCRDMA_MIN_SLOT_TABLE;
+static unsigned int max_slot_table_size = RPCRDMA_MAX_SLOT_TABLE;
+static unsigned int zero;
+static unsigned int max_padding = PAGE_SIZE;
+static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS;
+static unsigned int max_memreg = RPCRDMA_LAST - 1;
+
+static struct ctl_table_header *sunrpc_table_header;
+
+static ctl_table xr_tunables_table[] = {
+	{
+		.ctl_name	= CTL_SLOTTABLE_RDMA,
+		.procname	= "rdma_slot_table_entries",
+		.data		= &xprt_rdma_slot_table_entries,
+		.maxlen		= sizeof(unsigned int),
+		.mode		= 0644,
+		.proc_handler	= &proc_dointvec_minmax,
+		.strategy	= &sysctl_intvec,
+		.extra1		= &min_slot_table_size,
+		.extra2		= &max_slot_table_size
+	},
+	{
+		.ctl_name	= CTL_RDMA_MAXINLINEREAD,
+		.procname	= "rdma_max_inline_read",
+		.data		= &xprt_rdma_max_inline_read,
+		.maxlen		= sizeof(unsigned int),
+		.mode		= 0644,
+		.proc_handler	= &proc_dointvec,
+		.strategy	= &sysctl_intvec,
+	},
+	{
+		.ctl_name	= CTL_RDMA_MAXINLINEWRITE,
+		.procname	= "rdma_max_inline_write",
+		.data		= &xprt_rdma_max_inline_write,
+		.maxlen		= sizeof(unsigned int),
+		.mode		= 0644,
+		.proc_handler	= &proc_dointvec,
+		.strategy	= &sysctl_intvec,
+	},
+	{
+		.ctl_name	= CTL_RDMA_WRITEPADDING,
+		.procname	= "rdma_inline_write_padding",
+		.data		= &xprt_rdma_inline_write_padding,
+		.maxlen		= sizeof(unsigned int),
+		.mode		= 0644,
+		.proc_handler	= &proc_dointvec_minmax,
+		.strategy	= &sysctl_intvec,
+		.extra1		= &zero,
+		.extra2		= &max_padding,
+	},
+	{
+		.ctl_name	= CTL_RDMA_MEMREG,
+		.procname	= "rdma_memreg_strategy",
+		.data		= &xprt_rdma_memreg_strategy,
+		.maxlen		= sizeof(unsigned int),
+		.mode		= 0644,
+		.proc_handler	= &proc_dointvec_minmax,
+		.strategy	= &sysctl_intvec,
+		.extra1		= &min_memreg,
+		.extra2		= &max_memreg,
+	},
+	{
+		.ctl_name = 0,
+	},
+};
+
+static ctl_table sunrpc_table[] = {
+	{
+		.ctl_name	= CTL_SUNRPC,
+		.procname	= "sunrpc",
+		.mode		= 0555,
+		.child		= xr_tunables_table
+	},
+	{
+		.ctl_name = 0,
+	},
+};
+
+#endif
+
+static struct rpc_xprt_ops xprt_rdma_procs;	/* forward reference */
+
+static void
+xprt_rdma_format_addresses(struct rpc_xprt *xprt)
+{
+	struct sockaddr_in *addr = (struct sockaddr_in *)
+					&rpcx_to_rdmad(xprt).addr;
+	char *buf;
+
+	buf = kzalloc(20, GFP_KERNEL);
+	if (buf)
+		snprintf(buf, 20, NIPQUAD_FMT, NIPQUAD(addr->sin_addr.s_addr));
+	xprt->address_strings[RPC_DISPLAY_ADDR] = buf;
+
+	buf = kzalloc(8, GFP_KERNEL);
+	if (buf)
+		snprintf(buf, 8, "%u", ntohs(addr->sin_port));
+	xprt->address_strings[RPC_DISPLAY_PORT] = buf;
+
+	xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
+
+	buf = kzalloc(48, GFP_KERNEL);
+	if (buf)
+		snprintf(buf, 48, "addr="NIPQUAD_FMT" port=%u proto=%s",
+			NIPQUAD(addr->sin_addr.s_addr),
+			ntohs(addr->sin_port), "rdma");
+	xprt->address_strings[RPC_DISPLAY_ALL] = buf;
+
+	buf = kzalloc(10, GFP_KERNEL);
+	if (buf)
+		snprintf(buf, 10, "%02x%02x%02x%02x",
+			NIPQUAD(addr->sin_addr.s_addr));
+	xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = buf;
+
+	buf = kzalloc(8, GFP_KERNEL);
+	if (buf)
+		snprintf(buf, 8, "%4hx", ntohs(addr->sin_port));
+	xprt->address_strings[RPC_DISPLAY_HEX_PORT] = buf;
+
+	buf = kzalloc(30, GFP_KERNEL);
+	if (buf)
+		snprintf(buf, 30, NIPQUAD_FMT".%u.%u",
+			NIPQUAD(addr->sin_addr.s_addr),
+			ntohs(addr->sin_port) >> 8,
+			ntohs(addr->sin_port) & 0xff);
+	xprt->address_strings[RPC_DISPLAY_UNIVERSAL_ADDR] = buf;
+
+	/* netid */
+	xprt->address_strings[RPC_DISPLAY_NETID] = "rdma";
+}
+
+static void
+xprt_rdma_free_addresses(struct rpc_xprt *xprt)
+{
+	kfree(xprt->address_strings[RPC_DISPLAY_ADDR]);
+	kfree(xprt->address_strings[RPC_DISPLAY_PORT]);
+	kfree(xprt->address_strings[RPC_DISPLAY_ALL]);
+	kfree(xprt->address_strings[RPC_DISPLAY_HEX_ADDR]);
+	kfree(xprt->address_strings[RPC_DISPLAY_HEX_PORT]);
+	kfree(xprt->address_strings[RPC_DISPLAY_UNIVERSAL_ADDR]);
+}
+
+static void
+xprt_rdma_connect_worker(struct work_struct *work)
+{
+	struct rpcrdma_xprt *r_xprt =
+		container_of(work, struct rpcrdma_xprt, rdma_connect.work);
+	struct rpc_xprt *xprt = &r_xprt->xprt;
+	int rc = 0;
+
+	if (!xprt->shutdown) {
+		xprt_clear_connected(xprt);
+
+		dprintk("RPC:       %s: %sconnect\n", __func__,
+				r_xprt->rx_ep.rep_connected != 0 ? "re" : "");
+		rc = rpcrdma_ep_connect(&r_xprt->rx_ep, &r_xprt->rx_ia);
+		if (rc)
+			goto out;
+	}
+	goto out_clear;
+
+out:
+	xprt_wake_pending_tasks(xprt, rc);
+
+out_clear:
+	dprintk("RPC:       %s: exit\n", __func__);
+	xprt_clear_connecting(xprt);
+}
+
+/*
+ * xprt_rdma_destroy
+ *
+ * Destroy the xprt.
+ * Free all memory associated with the object, including its own.
+ * NOTE: none of the *destroy methods free memory for their top-level
+ * objects, even though they may have allocated it (they do free
+ * private memory). It's up to the caller to handle it. In this
+ * case (RDMA transport), all structure memory is inlined with the
+ * struct rpcrdma_xprt.
+ */
+static void
+xprt_rdma_destroy(struct rpc_xprt *xprt)
+{
+	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
+	int rc;
+
+	dprintk("RPC:       %s: called\n", __func__);
+
+	cancel_delayed_work(&r_xprt->rdma_connect);
+	flush_scheduled_work();
+
+	xprt_clear_connected(xprt);
+
+	rpcrdma_buffer_destroy(&r_xprt->rx_buf);
+	rc = rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia);
+	if (rc)
+		dprintk("RPC:       %s: rpcrdma_ep_destroy returned %i\n",
+			__func__, rc);
+	rpcrdma_ia_close(&r_xprt->rx_ia);
+
+	xprt_rdma_free_addresses(xprt);
+
+	kfree(xprt->slot);
+	xprt->slot = NULL;
+	kfree(xprt);
+
+	dprintk("RPC:       %s: returning\n", __func__);
+
+	module_put(THIS_MODULE);
+}
+
+/**
+ * xprt_setup_rdma - Set up transport to use RDMA
+ *
+ * @args: rpc transport arguments
+ */
+static struct rpc_xprt *
+xprt_setup_rdma(struct xprt_create *args)
+{
+	struct rpcrdma_create_data_internal cdata;
+	struct rpc_xprt *xprt;
+	struct rpcrdma_xprt *new_xprt;
+	struct rpcrdma_ep *new_ep;
+	struct sockaddr_in *sin;
+	int rc;
+
+	if (args->addrlen > sizeof(xprt->addr)) {
+		dprintk("RPC:       %s: address too large\n", __func__);
+		return ERR_PTR(-EBADF);
+	}
+
+	xprt = kzalloc(sizeof(struct rpcrdma_xprt), GFP_KERNEL);
+	if (xprt == NULL) {
+		dprintk("RPC:       %s: couldn't allocate rpcrdma_xprt\n",
+			__func__);
+		return ERR_PTR(-ENOMEM);
+	}
+
+	xprt->max_reqs = xprt_rdma_slot_table_entries;
+	xprt->slot = kcalloc(xprt->max_reqs,
+				sizeof(struct rpc_rqst), GFP_KERNEL);
+	if (xprt->slot == NULL) {
+		kfree(xprt);
+		dprintk("RPC:       %s: couldn't allocate %d slots\n",
+			__func__, xprt->max_reqs);
+		return ERR_PTR(-ENOMEM);
+	}
+
+	/* 60 second timeout, no retries */
+	xprt_set_timeout(&xprt->timeout, 0, 60UL * HZ);
+	xprt->bind_timeout = (60U * HZ);
+	xprt->connect_timeout = (60U * HZ);
+	xprt->reestablish_timeout = (5U * HZ);
+	xprt->idle_timeout = (5U * 60 * HZ);
+
+	xprt->resvport = 0;		/* privileged port not needed */
+	xprt->tsh_size = 0;		/* RPC-RDMA handles framing */
+	xprt->max_payload = RPCRDMA_MAX_DATA_SEGS * PAGE_SIZE;
+	xprt->ops = &xprt_rdma_procs;
+
+	/*
+	 * Set up RDMA-specific connect data.
+	 */
+
+	/* Put server RDMA address in local cdata */
+	memcpy(&cdata.addr, args->dstaddr, args->addrlen);
+
+	/* Ensure xprt->addr holds valid server TCP (not RDMA)
+	 * address, for any side protocols which peek at it */
+	xprt->prot = IPPROTO_TCP;
+	xprt->addrlen = args->addrlen;
+	memcpy(&xprt->addr, &cdata.addr, xprt->addrlen);
+
+	sin = (struct sockaddr_in *)&cdata.addr;
+	if (ntohs(sin->sin_port) != 0)
+		xprt_set_bound(xprt);
+
+	dprintk("RPC:       %s: %u.%u.%u.%u:%u\n", __func__,
+			NIPQUAD(sin->sin_addr.s_addr), ntohs(sin->sin_port));
+
+	/* Set max requests */
+	cdata.max_requests = xprt->max_reqs;
+
+	/* Set some length limits */
+	cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */
+	cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */
+
+	cdata.inline_wsize = xprt_rdma_max_inline_write;
+	if (cdata.inline_wsize > cdata.wsize)
+		cdata.inline_wsize = cdata.wsize;
+
+	cdata.inline_rsize = xprt_rdma_max_inline_read;
+	if (cdata.inline_rsize > cdata.rsize)
+		cdata.inline_rsize = cdata.rsize;
+
+	cdata.padding = xprt_rdma_inline_write_padding;
+
+	/*
+	 * Create new transport instance, which includes initialized
+	 *  o ia
+	 *  o endpoint
+	 *  o buffers
+	 */
+
+	new_xprt = rpcx_to_rdmax(xprt);
+
+	rc = rpcrdma_ia_open(new_xprt, (struct sockaddr *) &cdata.addr,
+				xprt_rdma_memreg_strategy);
+	if (rc)
+		goto out1;
+
+	/*
+	 * initialize and create ep
+	 */
+	new_xprt->rx_data = cdata;
+	new_ep = &new_xprt->rx_ep;
+	new_ep->rep_remote_addr = cdata.addr;
+
+	rc = rpcrdma_ep_create(&new_xprt->rx_ep,
+				&new_xprt->rx_ia, &new_xprt->rx_data);
+	if (rc)
+		goto out2;
+
+	/*
+	 * Allocate pre-registered send and receive buffers for headers and
+	 * any inline data. Also specify any padding which will be provided
+	 * from a preregistered zero buffer.
+	 */
+	rc = rpcrdma_buffer_create(&new_xprt->rx_buf, new_ep, &new_xprt->rx_ia,
+				&new_xprt->rx_data);
+	if (rc)
+		goto out3;
+
+	/*
+	 * Register a callback for connection events. This is necessary because
+	 * connection loss notification is async. We also catch connection loss
+	 * when reaping receives.
+	 */
+	INIT_DELAYED_WORK(&new_xprt->rdma_connect, xprt_rdma_connect_worker);
+	new_ep->rep_func = rpcrdma_conn_func;
+	new_ep->rep_xprt = xprt;
+
+	xprt_rdma_format_addresses(xprt);
+
+	if (!try_module_get(THIS_MODULE))
+		goto out4;
+
+	return xprt;
+
+out4:
+	xprt_rdma_free_addresses(xprt);
+	rc = -EINVAL;
+out3:
+	(void) rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia);
+out2:
+	rpcrdma_ia_close(&new_xprt->rx_ia);
+out1:
+	kfree(xprt->slot);
+	kfree(xprt);
+	return ERR_PTR(rc);
+}
+
+/*
+ * Close a connection, during shutdown or timeout/reconnect
+ */
+static void
+xprt_rdma_close(struct rpc_xprt *xprt)
+{
+	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
+
+	dprintk("RPC:       %s: closing\n", __func__);
+	xprt_disconnect(xprt);
+	(void) rpcrdma_ep_disconnect(&r_xprt->rx_ep, &r_xprt->rx_ia);
+}
+
+static void
+xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
+{
+	struct sockaddr_in *sap;
+
+	sap = (struct sockaddr_in *)&xprt->addr;
+	sap->sin_port = htons(port);
+	sap = (struct sockaddr_in *)&rpcx_to_rdmad(xprt).addr;
+	sap->sin_port = htons(port);
+	dprintk("RPC:       %s: %u\n", __func__, port);
+}
+
+static void
+xprt_rdma_connect(struct rpc_task *task)
+{
+	struct rpc_xprt *xprt = (struct rpc_xprt *)task->tk_xprt;
+	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
+
+	if (!xprt_test_and_set_connecting(xprt)) {
+		if (r_xprt->rx_ep.rep_connected != 0) {
+			/* Reconnect */
+			schedule_delayed_work(&r_xprt->rdma_connect,
+				xprt->reestablish_timeout);
+		} else {
+			schedule_delayed_work(&r_xprt->rdma_connect, 0);
+			if (!RPC_IS_ASYNC(task))
+				flush_scheduled_work();
+		}
+	}
+}
+
+static int
+xprt_rdma_reserve_xprt(struct rpc_task *task)
+{
+	struct rpc_xprt *xprt = task->tk_xprt;
+	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
+	int credits = atomic_read(&r_xprt->rx_buf.rb_credits);
+
+	/* == RPC_CWNDSCALE @ init, but *after* setup */
+	if (r_xprt->rx_buf.rb_cwndscale == 0UL) {
+		r_xprt->rx_buf.rb_cwndscale = xprt->cwnd;
+		dprintk("RPC:       %s: cwndscale %lu\n", __func__,
+			r_xprt->rx_buf.rb_cwndscale);
+		BUG_ON(r_xprt->rx_buf.rb_cwndscale <= 0);
+	}
+	xprt->cwnd = credits * r_xprt->rx_buf.rb_cwndscale;
+	return xprt_reserve_xprt_cong(task);
+}
+
+/*
+ * The RDMA allocate/free functions need the task structure as a place
+ * to hide the struct rpcrdma_req, which is necessary for the actual send/recv
+ * sequence. For this reason, the recv buffers are attached to send
+ * buffers for portions of the RPC. Note that the RPC layer allocates
+ * both send and receive buffers in the same call. We may register
+ * the receive buffer portion when using reply chunks.
+ */
+static void *
+xprt_rdma_allocate(struct rpc_task *task, size_t size)
+{
+	struct rpc_xprt *xprt = task->tk_xprt;
+	struct rpcrdma_req *req, *nreq;
+
+	req = rpcrdma_buffer_get(&rpcx_to_rdmax(xprt)->rx_buf);
+	BUG_ON(NULL == req);
+
+	if (size > req->rl_size) {
+		dprintk("RPC:       %s: size %zd too large for buffer[%zd]: "
+			"prog %d vers %d proc %d\n",
+			__func__, size, req->rl_size,
+			task->tk_client->cl_prog, task->tk_client->cl_vers,
+			task->tk_msg.rpc_proc->p_proc);
+		/*
+		 * Outgoing length shortage. Our inline write max must have
+		 * been configured to perform direct i/o.
+		 *
+		 * This is therefore a large metadata operation, and the
+		 * allocate call was made on the maximum possible message,
+		 * e.g. containing long filename(s) or symlink data. In
+		 * fact, while these metadata operations *might* carry
+		 * large outgoing payloads, they rarely *do*. However, we
+		 * have to commit to the request here, so reallocate and
+		 * register it now. The data path will never require this
+		 * reallocation.
+		 *
+		 * If the allocation or registration fails, the RPC framework
+		 * will (doggedly) retry.
+		 */
+		if (rpcx_to_rdmax(xprt)->rx_ia.ri_memreg_strategy ==
+				RPCRDMA_BOUNCEBUFFERS) {
+			/* forced to "pure inline" */
+			dprintk("RPC:       %s: too much data (%zd) for inline "
+					"(r/w max %d/%d)\n", __func__, size,
+					rpcx_to_rdmad(xprt).inline_rsize,
+					rpcx_to_rdmad(xprt).inline_wsize);
+			size = req->rl_size;
+			rpc_exit(task, -EIO);		/* fail the operation */
+			rpcx_to_rdmax(xprt)->rx_stats.failed_marshal_count++;
+			goto out;
+		}
+		if (task->tk_flags & RPC_TASK_SWAPPER)
+			nreq = kmalloc(sizeof *req + size, GFP_ATOMIC);
+		else
+			nreq = kmalloc(sizeof *req + size, GFP_NOFS);
+		if (nreq == NULL)
+			goto outfail;
+
+		if (rpcrdma_register_internal(&rpcx_to_rdmax(xprt)->rx_ia,
+				nreq->rl_base, size + sizeof(struct rpcrdma_req)
+				- offsetof(struct rpcrdma_req, rl_base),
+				&nreq->rl_handle, &nreq->rl_iov)) {
+			kfree(nreq);
+			goto outfail;
+		}
+		rpcx_to_rdmax(xprt)->rx_stats.hardway_register_count += size;
+		nreq->rl_size = size;
+		nreq->rl_niovs = 0;
+		nreq->rl_nchunks = 0;
+		nreq->rl_buffer = (struct rpcrdma_buffer *)req;
+		nreq->rl_reply = req->rl_reply;
+		memcpy(nreq->rl_segments,
+			req->rl_segments, sizeof nreq->rl_segments);
+		/* flag the swap with an unused field */
+		nreq->rl_iov.length = 0;
+		req->rl_reply = NULL;
+		req = nreq;
+	}
+	dprintk("RPC:       %s: size %zd, request 0x%p\n", __func__, size, req);
+out:
+	return req->rl_xdr_buf;
+
+outfail:
+	rpcrdma_buffer_put(req);
+	rpcx_to_rdmax(xprt)->rx_stats.failed_marshal_count++;
+	return NULL;
+}
+
+/*
+ * This function returns all RDMA resources to the pool.
+ */
+static void
+xprt_rdma_free(void *buffer)
+{
+	struct rpcrdma_req *req;
+	struct rpcrdma_xprt *r_xprt;
+	struct rpcrdma_rep *rep;
+	int i;
+
+	if (buffer == NULL)
+		return;
+
+	req = container_of(buffer, struct rpcrdma_req, rl_xdr_buf[0]);
+	r_xprt = container_of(req->rl_buffer, struct rpcrdma_xprt, rx_buf);
+	rep = req->rl_reply;
+
+	dprintk("RPC:       %s: called on 0x%p%s\n",
+		__func__, rep, (rep && rep->rr_func) ? " (with waiter)" : "");
+
+	/*
+	 * Finish the deregistration. When using mw bind, this was
+	 * begun in rpcrdma_reply_handler(). In all other modes, we
+	 * do it here, in thread context. The process is considered
+	 * complete when the rr_func vector becomes NULL - this
+	 * was put in place during rpcrdma_reply_handler() - the wait
+	 * call below will not block if the dereg is "done". If
+	 * interrupted, our framework will clean up.
+	 */
+	for (i = 0; req->rl_nchunks;) {
+		--req->rl_nchunks;
+		i += rpcrdma_deregister_external(
+			&req->rl_segments[i], r_xprt, NULL);
+	}
+
+	if (rep && wait_event_interruptible(rep->rr_unbind, !rep->rr_func)) {
+		rep->rr_func = NULL;	/* abandon the callback */
+		req->rl_reply = NULL;
+	}
+
+	if (req->rl_iov.length == 0) {	/* see allocate above */
+		struct rpcrdma_req *oreq = (struct rpcrdma_req *)req->rl_buffer;
+		oreq->rl_reply = req->rl_reply;
+		(void) rpcrdma_deregister_internal(&r_xprt->rx_ia,
+						   req->rl_handle,
+						   &req->rl_iov);
+		kfree(req);
+		req = oreq;
+	}
+
+	/* Put back request+reply buffers */
+	rpcrdma_buffer_put(req);
+}
+
+/*
+ * send_request invokes the meat of RPC RDMA. It must do the following:
+ *  1.  Marshal the RPC request into an RPC RDMA request, which means
+ *	putting a header in front of data, and creating IOVs for RDMA
+ *	from those in the request.
+ *  2.  In marshaling, detect opportunities for RDMA, and use them.
+ *  3.  Post a recv message to set up asynch completion, then send
+ *	the request (rpcrdma_ep_post).
+ *  4.  No partial sends are possible in the RPC-RDMA protocol (as in UDP).
+ */
+
+static int
+xprt_rdma_send_request(struct rpc_task *task)
+{
+	struct rpc_rqst *rqst = task->tk_rqstp;
+	struct rpc_xprt *xprt = task->tk_xprt;
+	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
+	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
+
+	/* marshal the send itself */
+	if (req->rl_niovs == 0 && rpcrdma_marshal_req(rqst) != 0) {
+		r_xprt->rx_stats.failed_marshal_count++;
+		dprintk("RPC:       %s: rpcrdma_marshal_req failed\n",
+			__func__);
+		return -EIO;
+	}
+
+	if (req->rl_reply == NULL) 		/* e.g. reconnection */
+		rpcrdma_recv_buffer_get(req);
+
+	if (req->rl_reply) {
+		req->rl_reply->rr_func = rpcrdma_reply_handler;
+		/* this need only be done once, but... */
+		req->rl_reply->rr_xprt = xprt;
+	}
+
+	if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req)) {
+		xprt_disconnect(xprt);
+		return -ENOTCONN;	/* implies disconnect */
+	}
+
+	rqst->rq_bytes_sent = 0;
+	return 0;
+}
+
+static void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
+{
+	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
+	long idle_time = 0;
+
+	if (xprt_connected(xprt))
+		idle_time = (long)(jiffies - xprt->last_used) / HZ;
+
+	seq_printf(seq,
+	  "\txprt:\trdma %u %lu %lu %lu %ld %lu %lu %lu %Lu %Lu "
+	  "%lu %lu %lu %Lu %Lu %Lu %Lu %lu %lu %lu\n",
+
+	   0,	/* need a local port? */
+	   xprt->stat.bind_count,
+	   xprt->stat.connect_count,
+	   xprt->stat.connect_time,
+	   idle_time,
+	   xprt->stat.sends,
+	   xprt->stat.recvs,
+	   xprt->stat.bad_xids,
+	   xprt->stat.req_u,
+	   xprt->stat.bklog_u,
+
+	   r_xprt->rx_stats.read_chunk_count,
+	   r_xprt->rx_stats.write_chunk_count,
+	   r_xprt->rx_stats.reply_chunk_count,
+	   r_xprt->rx_stats.total_rdma_request,
+	   r_xprt->rx_stats.total_rdma_reply,
+	   r_xprt->rx_stats.pullup_copy_count,
+	   r_xprt->rx_stats.fixup_copy_count,
+	   r_xprt->rx_stats.hardway_register_count,
+	   r_xprt->rx_stats.failed_marshal_count,
+	   r_xprt->rx_stats.bad_reply_count);
+}
+
+/*
+ * Plumbing for rpc transport switch and kernel module
+ */
+
+static struct rpc_xprt_ops xprt_rdma_procs = {
+	.reserve_xprt		= xprt_rdma_reserve_xprt,
+	.release_xprt		= xprt_release_xprt_cong, /* sunrpc/xprt.c */
+	.release_request	= xprt_release_rqst_cong,       /* ditto */
+	.set_retrans_timeout	= xprt_set_retrans_timeout_def, /* ditto */
+	.rpcbind		= rpcb_getport_async,	/* sunrpc/rpcb_clnt.c */
+	.set_port		= xprt_rdma_set_port,
+	.connect		= xprt_rdma_connect,
+	.buf_alloc		= xprt_rdma_allocate,
+	.buf_free		= xprt_rdma_free,
+	.send_request		= xprt_rdma_send_request,
+	.close			= xprt_rdma_close,
+	.destroy		= xprt_rdma_destroy,
+	.print_stats		= xprt_rdma_print_stats
+};
+
+static struct xprt_class xprt_rdma = {
+	.list			= LIST_HEAD_INIT(xprt_rdma.list),
+	.name			= "rdma",
+	.owner			= THIS_MODULE,
+	.ident			= XPRT_TRANSPORT_RDMA,
+	.setup			= xprt_setup_rdma,
+};
+
+static void __exit xprt_rdma_cleanup(void)
+{
+	int rc;
+
+	dprintk("RPCRDMA Module Removed, deregister RPC RDMA transport\n");
+#ifdef RPC_DEBUG
+	if (sunrpc_table_header) {
+		unregister_sysctl_table(sunrpc_table_header);
+		sunrpc_table_header = NULL;
+	}
+#endif
+	rc = xprt_unregister_transport(&xprt_rdma);
+	if (rc)
+		dprintk("RPC:       %s: xprt_unregister returned %i\n",
+			__func__, rc);
+}
+
+static int __init xprt_rdma_init(void)
+{
+	int rc;
+
+	rc = xprt_register_transport(&xprt_rdma);
+
+	if (rc)
+		return rc;
+
+	dprintk(KERN_INFO "RPCRDMA Module Init, register RPC RDMA transport\n");
+
+	dprintk(KERN_INFO "Defaults:\n");
+	dprintk(KERN_INFO "\tSlots %d\n"
+		"\tMaxInlineRead %d\n\tMaxInlineWrite %d\n",
+		xprt_rdma_slot_table_entries,
+		xprt_rdma_max_inline_read, xprt_rdma_max_inline_write);
+	dprintk(KERN_INFO "\tPadding %d\n\tMemreg %d\n",
+		xprt_rdma_inline_write_padding, xprt_rdma_memreg_strategy);
+
+#ifdef RPC_DEBUG
+	if (!sunrpc_table_header)
+		sunrpc_table_header = register_sysctl_table(sunrpc_table);
+#endif
+	return 0;
+}
+
+module_init(xprt_rdma_init);
+module_exit(xprt_rdma_cleanup);
diff --git a/net/sunrpc/xprtrdma/verbs.c b/net/sunrpc/xprtrdma/verbs.c
new file mode 100644
index 000000000000..0baf53381987
--- /dev/null
+++ b/net/sunrpc/xprtrdma/verbs.c
@@ -0,0 +1,37 @@
+/*
+ * Placeholders for subsequent patches
+ */
+
+#include "xprt_rdma.h"
+
+int rpcrdma_ia_open(struct rpcrdma_xprt *a, struct sockaddr *b, int c)
+{ return EINVAL; }
+void rpcrdma_ia_close(struct rpcrdma_ia *a) { }
+int rpcrdma_ep_create(struct rpcrdma_ep *a, struct rpcrdma_ia *b,
+struct rpcrdma_create_data_internal *c) { return EINVAL; }
+int rpcrdma_ep_destroy(struct rpcrdma_ep *a, struct rpcrdma_ia *b)
+{ return EINVAL; }
+int rpcrdma_ep_connect(struct rpcrdma_ep *a, struct rpcrdma_ia *b)
+{ return EINVAL; }
+int rpcrdma_ep_disconnect(struct rpcrdma_ep *a, struct rpcrdma_ia *b)
+{ return EINVAL; }
+int rpcrdma_ep_post(struct rpcrdma_ia *a, struct rpcrdma_ep *b,
+struct rpcrdma_req *c) { return EINVAL; }
+int rpcrdma_ep_post_recv(struct rpcrdma_ia *a, struct rpcrdma_ep *b,
+struct rpcrdma_rep *c) { return EINVAL; }
+int rpcrdma_buffer_create(struct rpcrdma_buffer *a, struct rpcrdma_ep *b,
+struct rpcrdma_ia *c, struct rpcrdma_create_data_internal *d) { return EINVAL; }
+void rpcrdma_buffer_destroy(struct rpcrdma_buffer *a) { }
+struct rpcrdma_req *rpcrdma_buffer_get(struct rpcrdma_buffer *a)
+{ return NULL; }
+void rpcrdma_buffer_put(struct rpcrdma_req *a) { }
+void rpcrdma_recv_buffer_get(struct rpcrdma_req *a) { }
+void rpcrdma_recv_buffer_put(struct rpcrdma_rep *a) { }
+int rpcrdma_register_internal(struct rpcrdma_ia *a, void *b, int c,
+struct ib_mr **d, struct ib_sge *e) { return EINVAL; }
+int rpcrdma_deregister_internal(struct rpcrdma_ia *a, struct ib_mr *b,
+struct ib_sge *c) { return EINVAL; }
+int rpcrdma_register_external(struct rpcrdma_mr_seg *a, int b, int c,
+struct rpcrdma_xprt *d) { return EINVAL; }
+int rpcrdma_deregister_external(struct rpcrdma_mr_seg *a,
+struct rpcrdma_xprt *b, void *c) { return EINVAL; }
diff --git a/net/sunrpc/xprtrdma/xprt_rdma.h b/net/sunrpc/xprtrdma/xprt_rdma.h
new file mode 100644
index 000000000000..2427822f8bd4
--- /dev/null
+++ b/net/sunrpc/xprtrdma/xprt_rdma.h
@@ -0,0 +1,330 @@
+/*
+ * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the BSD-type
+ * license below:
+ *
+ * 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 the Network Appliance, Inc. 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.
+ */
+
+#ifndef _LINUX_SUNRPC_XPRT_RDMA_H
+#define _LINUX_SUNRPC_XPRT_RDMA_H
+
+#include <linux/wait.h> 		/* wait_queue_head_t, etc */
+#include <linux/spinlock.h> 		/* spinlock_t, etc */
+#include <asm/atomic.h>			/* atomic_t, etc */
+
+#include <rdma/rdma_cm.h>		/* RDMA connection api */
+#include <rdma/ib_verbs.h>		/* RDMA verbs api */
+
+#include <linux/sunrpc/clnt.h> 		/* rpc_xprt */
+#include <linux/sunrpc/rpc_rdma.h> 	/* RPC/RDMA protocol */
+#include <linux/sunrpc/xprtrdma.h> 	/* xprt parameters */
+
+/*
+ * Interface Adapter -- one per transport instance
+ */
+struct rpcrdma_ia {
+	struct rdma_cm_id 	*ri_id;
+	struct ib_pd		*ri_pd;
+	struct ib_mr		*ri_bind_mem;
+	struct completion	ri_done;
+	int			ri_async_rc;
+	enum rpcrdma_memreg	ri_memreg_strategy;
+};
+
+/*
+ * RDMA Endpoint -- one per transport instance
+ */
+
+struct rpcrdma_ep {
+	atomic_t		rep_cqcount;
+	int			rep_cqinit;
+	int			rep_connected;
+	struct rpcrdma_ia	*rep_ia;
+	struct ib_cq		*rep_cq;
+	struct ib_qp_init_attr	rep_attr;
+	wait_queue_head_t 	rep_connect_wait;
+	struct ib_sge		rep_pad;	/* holds zeroed pad */
+	struct ib_mr		*rep_pad_mr;	/* holds zeroed pad */
+	void			(*rep_func)(struct rpcrdma_ep *);
+	struct rpc_xprt		*rep_xprt;	/* for rep_func */
+	struct rdma_conn_param	rep_remote_cma;
+	struct sockaddr_storage	rep_remote_addr;
+};
+
+#define INIT_CQCOUNT(ep) atomic_set(&(ep)->rep_cqcount, (ep)->rep_cqinit)
+#define DECR_CQCOUNT(ep) atomic_sub_return(1, &(ep)->rep_cqcount)
+
+/*
+ * struct rpcrdma_rep -- this structure encapsulates state required to recv
+ * and complete a reply, asychronously. It needs several pieces of
+ * state:
+ *   o recv buffer (posted to provider)
+ *   o ib_sge (also donated to provider)
+ *   o status of reply (length, success or not)
+ *   o bookkeeping state to get run by tasklet (list, etc)
+ *
+ * These are allocated during initialization, per-transport instance;
+ * however, the tasklet execution list itself is global, as it should
+ * always be pretty short.
+ *
+ * N of these are associated with a transport instance, and stored in
+ * struct rpcrdma_buffer. N is the max number of outstanding requests.
+ */
+
+/* temporary static scatter/gather max */
+#define RPCRDMA_MAX_DATA_SEGS	(8)	/* max scatter/gather */
+#define RPCRDMA_MAX_SEGS 	(RPCRDMA_MAX_DATA_SEGS + 2) /* head+tail = 2 */
+#define MAX_RPCRDMAHDR	(\
+	/* max supported RPC/RDMA header */ \
+	sizeof(struct rpcrdma_msg) + (2 * sizeof(u32)) + \
+	(sizeof(struct rpcrdma_read_chunk) * RPCRDMA_MAX_SEGS) + sizeof(u32))
+
+struct rpcrdma_buffer;
+
+struct rpcrdma_rep {
+	unsigned int	rr_len;		/* actual received reply length */
+	struct rpcrdma_buffer *rr_buffer; /* home base for this structure */
+	struct rpc_xprt	*rr_xprt;	/* needed for request/reply matching */
+	void (*rr_func)(struct rpcrdma_rep *);/* called by tasklet in softint */
+	struct list_head rr_list;	/* tasklet list */
+	wait_queue_head_t rr_unbind;	/* optional unbind wait */
+	struct ib_sge	rr_iov;		/* for posting */
+	struct ib_mr	*rr_handle;	/* handle for mem in rr_iov */
+	char	rr_base[MAX_RPCRDMAHDR]; /* minimal inline receive buffer */
+};
+
+/*
+ * struct rpcrdma_req -- structure central to the request/reply sequence.
+ *
+ * N of these are associated with a transport instance, and stored in
+ * struct rpcrdma_buffer. N is the max number of outstanding requests.
+ *
+ * It includes pre-registered buffer memory for send AND recv.
+ * The recv buffer, however, is not owned by this structure, and
+ * is "donated" to the hardware when a recv is posted. When a
+ * reply is handled, the recv buffer used is given back to the
+ * struct rpcrdma_req associated with the request.
+ *
+ * In addition to the basic memory, this structure includes an arra