RDS: Add iWARP support

Support for iWARP NICs is implemented as a separate
RDS transport from IB. The code, however, is very
similar to IB (it was forked, basically.) so let's keep
it in one changeset.

The reason for this duplicationis that despite its similarity
to IB, there are a number of places where it has different
semantics. iwarp zcopy support is still under development,
and giving it its own sandbox ensures that IB code isn't
disrupted while iwarp changes. Over time these transports
will re-converge.

Signed-off-by: Andy Grover <andy.grover@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

9 files changed, 4611 insertions, 0 deletions

diff --git a/net/rds/iw.c b/net/rds/iw.c
new file mode 100644
index 0000000..1b56905
--- /dev/null
+++ b/net/rds/iw.c
@@ -0,0 +1,333 @@
+/*
+ * Copyright (c) 2006 Oracle.  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
+ * OpenIB.org BSD 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.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/kernel.h>
+#include <linux/in.h>
+#include <linux/if.h>
+#include <linux/netdevice.h>
+#include <linux/inetdevice.h>
+#include <linux/if_arp.h>
+#include <linux/delay.h>
+
+#include "rds.h"
+#include "iw.h"
+
+unsigned int fastreg_pool_size = RDS_FASTREG_POOL_SIZE;
+unsigned int fastreg_message_size = RDS_FASTREG_SIZE + 1; /* +1 allows for unaligned MRs */
+
+module_param(fastreg_pool_size, int, 0444);
+MODULE_PARM_DESC(fastreg_pool_size, " Max number of fastreg MRs per device");
+module_param(fastreg_message_size, int, 0444);
+MODULE_PARM_DESC(fastreg_message_size, " Max size of a RDMA transfer (fastreg MRs)");
+
+struct list_head rds_iw_devices;
+
+DEFINE_SPINLOCK(iw_nodev_conns_lock);
+LIST_HEAD(iw_nodev_conns);
+
+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;
+
+	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->page_shift = max(PAGE_SHIFT, ffs(dev_attr->page_size_cap) - 1);
+
+	rds_iwdev->dev = device;
+	rds_iwdev->pd = ib_alloc_pd(device);
+	if (IS_ERR(rds_iwdev->pd))
+		goto free_dev;
+
+	if (!rds_iwdev->dma_local_lkey) {
+		if (device->node_type != RDMA_NODE_RNIC) {
+			rds_iwdev->mr = ib_get_dma_mr(rds_iwdev->pd,
+						IB_ACCESS_LOCAL_WRITE);
+		} else {
+			rds_iwdev->mr = ib_get_dma_mr(rds_iwdev->pd,
+						IB_ACCESS_REMOTE_READ |
+						IB_ACCESS_REMOTE_WRITE |
+						IB_ACCESS_LOCAL_WRITE);
+		}
+		if (IS_ERR(rds_iwdev->mr))
+			goto err_pd;
+	} else
+		rds_iwdev->mr = NULL;
+
+	rds_iwdev->mr_pool = rds_iw_create_mr_pool(rds_iwdev);
+	if (IS_ERR(rds_iwdev->mr_pool)) {
+		rds_iwdev->mr_pool = NULL;
+		goto err_mr;
+	}
+
+	INIT_LIST_HEAD(&rds_iwdev->cm_id_list);
+	INIT_LIST_HEAD(&rds_iwdev->conn_list);
+	list_add_tail(&rds_iwdev->list, &rds_iw_devices);
+
+	ib_set_client_data(device, &rds_iw_client, rds_iwdev);
+
+	goto free_attr;
+
+err_mr:
+	if (rds_iwdev->mr)
+		ib_dereg_mr(rds_iwdev->mr);
+err_pd:
+	ib_dealloc_pd(rds_iwdev->pd);
+free_dev:
+	kfree(rds_iwdev);
+free_attr:
+	kfree(dev_attr);
+}
+
+void rds_iw_remove_one(struct ib_device *device)
+{
+	struct rds_iw_device *rds_iwdev;
+	struct rds_iw_cm_id *i_cm_id, *next;
+
+	rds_iwdev = ib_get_client_data(device, &rds_iw_client);
+	if (!rds_iwdev)
+		return;
+
+	spin_lock_irq(&rds_iwdev->spinlock);
+	list_for_each_entry_safe(i_cm_id, next, &rds_iwdev->cm_id_list, list) {
+		list_del(&i_cm_id->list);
+		kfree(i_cm_id);
+	}
+	spin_unlock_irq(&rds_iwdev->spinlock);
+
+	rds_iw_remove_conns(rds_iwdev);
+
+	if (rds_iwdev->mr_pool)
+		rds_iw_destroy_mr_pool(rds_iwdev->mr_pool);
+
+	if (rds_iwdev->mr)
+		ib_dereg_mr(rds_iwdev->mr);
+
+	while (ib_dealloc_pd(rds_iwdev->pd)) {
+		rdsdebug("Failed to dealloc pd %p\n", rds_iwdev->pd);
+		msleep(1);
+	}
+
+	list_del(&rds_iwdev->list);
+	kfree(rds_iwdev);
+}
+
+struct ib_client rds_iw_client = {
+	.name   = "rds_iw",
+	.add    = rds_iw_add_one,
+	.remove = rds_iw_remove_one
+};
+
+static int rds_iw_conn_info_visitor(struct rds_connection *conn,
+				    void *buffer)
+{
+	struct rds_info_rdma_connection *iinfo = buffer;
+	struct rds_iw_connection *ic;
+
+	/* We will only ever look at IB transports */
+	if (conn->c_trans != &rds_iw_transport)
+		return 0;
+
+	iinfo->src_addr = conn->c_laddr;
+	iinfo->dst_addr = conn->c_faddr;
+
+	memset(&iinfo->src_gid, 0, sizeof(iinfo->src_gid));
+	memset(&iinfo->dst_gid, 0, sizeof(iinfo->dst_gid));
+	if (rds_conn_state(conn) == RDS_CONN_UP) {
+		struct rds_iw_device *rds_iwdev;
+		struct rdma_dev_addr *dev_addr;
+
+		ic = conn->c_transport_data;
+		dev_addr = &ic->i_cm_id->route.addr.dev_addr;
+
+		ib_addr_get_sgid(dev_addr, (union ib_gid *) &iinfo->src_gid);
+		ib_addr_get_dgid(dev_addr, (union ib_gid *) &iinfo->dst_gid);
+
+		rds_iwdev = ib_get_client_data(ic->i_cm_id->device, &rds_iw_client);
+		iinfo->max_send_wr = ic->i_send_ring.w_nr;
+		iinfo->max_recv_wr = ic->i_recv_ring.w_nr;
+		iinfo->max_send_sge = rds_iwdev->max_sge;
+		rds_iw_get_mr_info(rds_iwdev, iinfo);
+	}
+	return 1;
+}
+
+static void rds_iw_ic_info(struct socket *sock, unsigned int len,
+			   struct rds_info_iterator *iter,
+			   struct rds_info_lengths *lens)
+{
+	rds_for_each_conn_info(sock, len, iter, lens,
+				rds_iw_conn_info_visitor,
+				sizeof(struct rds_info_rdma_connection));
+}
+
+
+/*
+ * Early RDS/IB was built to only bind to an address if there is an IPoIB
+ * device with that address set.
+ *
+ * If it were me, I'd advocate for something more flexible.  Sending and
+ * receiving should be device-agnostic.  Transports would try and maintain
+ * connections between peers who have messages queued.  Userspace would be
+ * allowed to influence which paths have priority.  We could call userspace
+ * asserting this policy "routing".
+ */
+static int rds_iw_laddr_check(__be32 addr)
+{
+	int ret;
+	struct rdma_cm_id *cm_id;
+	struct sockaddr_in sin;
+
+	/* Create a CMA ID and try to bind it. This catches both
+	 * IB and iWARP capable NICs.
+	 */
+	cm_id = rdma_create_id(NULL, NULL, RDMA_PS_TCP);
+	if (!cm_id)
+		return -EADDRNOTAVAIL;
+
+	memset(&sin, 0, sizeof(sin));
+	sin.sin_family = AF_INET;
+	sin.sin_addr.s_addr = addr;
+
+	/* rdma_bind_addr will only succeed for IB & iWARP devices */
+	ret = rdma_bind_addr(cm_id, (struct sockaddr *)&sin);
+	/* due to this, we will claim to support IB devices unless we
+	   check node_type. */
+	if (ret || cm_id->device->node_type != RDMA_NODE_RNIC)
+		ret = -EADDRNOTAVAIL;
+
+	rdsdebug("addr %pI4 ret %d node type %d\n",
+		&addr, ret,
+		cm_id->device ? cm_id->device->node_type : -1);
+
+	rdma_destroy_id(cm_id);
+
+	return ret;
+}
+
+void rds_iw_exit(void)
+{
+	rds_info_deregister_func(RDS_INFO_IWARP_CONNECTIONS, rds_iw_ic_info);
+	rds_iw_remove_nodev_conns();
+	ib_unregister_client(&rds_iw_client);
+	rds_iw_sysctl_exit();
+	rds_iw_recv_exit();
+	rds_trans_unregister(&rds_iw_transport);
+}
+
+struct rds_transport rds_iw_transport = {
+	.laddr_check		= rds_iw_laddr_check,
+	.xmit_complete		= rds_iw_xmit_complete,
+	.xmit			= rds_iw_xmit,
+	.xmit_cong_map		= NULL,
+	.xmit_rdma		= rds_iw_xmit_rdma,
+	.recv			= rds_iw_recv,
+	.conn_alloc		= rds_iw_conn_alloc,
+	.conn_free		= rds_iw_conn_free,
+	.conn_connect		= rds_iw_conn_connect,
+	.conn_shutdown		= rds_iw_conn_shutdown,
+	.inc_copy_to_user	= rds_iw_inc_copy_to_user,
+	.inc_purge		= rds_iw_inc_purge,
+	.inc_free		= rds_iw_inc_free,
+	.cm_initiate_connect	= rds_iw_cm_initiate_connect,
+	.cm_handle_connect	= rds_iw_cm_handle_connect,
+	.cm_connect_complete	= rds_iw_cm_connect_complete,
+	.stats_info_copy	= rds_iw_stats_info_copy,
+	.exit			= rds_iw_exit,
+	.get_mr			= rds_iw_get_mr,
+	.sync_mr		= rds_iw_sync_mr,
+	.free_mr		= rds_iw_free_mr,
+	.flush_mrs		= rds_iw_flush_mrs,
+	.t_owner		= THIS_MODULE,
+	.t_name			= "iwarp",
+	.t_prefer_loopback	= 1,
+};
+
+int __init rds_iw_init(void)
+{
+	int ret;
+
+	INIT_LIST_HEAD(&rds_iw_devices);
+
+	ret = ib_register_client(&rds_iw_client);
+	if (ret)
+		goto out;
+
+	ret = rds_iw_sysctl_init();
+	if (ret)
+		goto out_ibreg;
+
+	ret = rds_iw_recv_init();
+	if (ret)
+		goto out_sysctl;
+
+	ret = rds_trans_register(&rds_iw_transport);
+	if (ret)
+		goto out_recv;
+
+	rds_info_register_func(RDS_INFO_IWARP_CONNECTIONS, rds_iw_ic_info);
+
+	goto out;
+
+out_recv:
+	rds_iw_recv_exit();
+out_sysctl:
+	rds_iw_sysctl_exit();
+out_ibreg:
+	ib_unregister_client(&rds_iw_client);
+out:
+	return ret;
+}
+
+MODULE_LICENSE("GPL");
+
diff --git a/net/rds/iw.h b/net/rds/iw.h
new file mode 100644
index 0000000..0ddda34
--- /dev/null
+++ b/net/rds/iw.h
@@ -0,0 +1,395 @@
+#ifndef _RDS_IW_H
+#define _RDS_IW_H
+
+#include <rdma/ib_verbs.h>
+#include <rdma/rdma_cm.h>
+#include "rds.h"
+#include "rdma_transport.h"
+
+#define RDS_FASTREG_SIZE		20
+#define RDS_FASTREG_POOL_SIZE		2048
+
+#define RDS_IW_MAX_SGE			8
+#define RDS_IW_RECV_SGE 		2
+
+#define RDS_IW_DEFAULT_RECV_WR		1024
+#define RDS_IW_DEFAULT_SEND_WR		256
+
+#define RDS_IW_SUPPORTED_PROTOCOLS	0x00000003	/* minor versions supported */
+
+extern struct list_head rds_iw_devices;
+
+/*
+ * IB posts RDS_FRAG_SIZE fragments of pages to the receive queues to
+ * try and minimize the amount of memory tied up both the device and
+ * socket receive queues.
+ */
+/* page offset of the final full frag that fits in the page */
+#define RDS_PAGE_LAST_OFF (((PAGE_SIZE  / RDS_FRAG_SIZE) - 1) * RDS_FRAG_SIZE)
+struct rds_page_frag {
+	struct list_head	f_item;
+	struct page		*f_page;
+	unsigned long		f_offset;
+	dma_addr_t 		f_mapped;
+};
+
+struct rds_iw_incoming {
+	struct list_head	ii_frags;
+	struct rds_incoming	ii_inc;
+};
+
+struct rds_iw_connect_private {
+	/* Add new fields at the end, and don't permute existing fields. */
+	__be32			dp_saddr;
+	__be32			dp_daddr;
+	u8			dp_protocol_major;
+	u8			dp_protocol_minor;
+	__be16			dp_protocol_minor_mask; /* bitmask */
+	__be32			dp_reserved1;
+	__be64			dp_ack_seq;
+	__be32			dp_credit;		/* non-zero enables flow ctl */
+};
+
+struct rds_iw_scatterlist {
+	struct scatterlist	*list;
+	unsigned int		len;
+	int			dma_len;
+	unsigned int		dma_npages;
+	unsigned int		bytes;
+};
+
+struct rds_iw_mapping {
+	spinlock_t		m_lock;	/* protect the mapping struct */
+	struct list_head	m_list;
+	struct rds_iw_mr	*m_mr;
+	uint32_t		m_rkey;
+	struct rds_iw_scatterlist m_sg;
+};
+
+struct rds_iw_send_work {
+	struct rds_message	*s_rm;
+
+	/* We should really put these into a union: */
+	struct rds_rdma_op	*s_op;
+	struct rds_iw_mapping	*s_mapping;
+	struct ib_mr		*s_mr;
+	struct ib_fast_reg_page_list *s_page_list;
+	unsigned char		s_remap_count;
+
+	struct ib_send_wr	s_wr;
+	struct ib_sge		s_sge[RDS_IW_MAX_SGE];
+	unsigned long		s_queued;
+};
+
+struct rds_iw_recv_work {
+	struct rds_iw_incoming 	*r_iwinc;
+	struct rds_page_frag	*r_frag;
+	struct ib_recv_wr	r_wr;
+	struct ib_sge		r_sge[2];
+};
+
+struct rds_iw_work_ring {
+	u32		w_nr;
+	u32		w_alloc_ptr;
+	u32		w_alloc_ctr;
+	u32		w_free_ptr;
+	atomic_t	w_free_ctr;
+};
+
+struct rds_iw_device;
+
+struct rds_iw_connection {
+
+	struct list_head	iw_node;
+	struct rds_iw_device 	*rds_iwdev;
+	struct rds_connection	*conn;
+
+	/* alphabet soup, IBTA style */
+	struct rdma_cm_id	*i_cm_id;
+	struct ib_pd		*i_pd;
+	struct ib_mr		*i_mr;
+	struct ib_cq		*i_send_cq;
+	struct ib_cq		*i_recv_cq;
+
+	/* tx */
+	struct rds_iw_work_ring	i_send_ring;
+	struct rds_message	*i_rm;
+	struct rds_header	*i_send_hdrs;
+	u64			i_send_hdrs_dma;
+	struct rds_iw_send_work *i_sends;
+
+	/* rx */
+	struct mutex		i_recv_mutex;
+	struct rds_iw_work_ring	i_recv_ring;
+	struct rds_iw_incoming	*i_iwinc;
+	u32			i_recv_data_rem;
+	struct rds_header	*i_recv_hdrs;
+	u64			i_recv_hdrs_dma;
+	struct rds_iw_recv_work *i_recvs;
+	struct rds_page_frag	i_frag;
+	u64			i_ack_recv;	/* last ACK received */
+
+	/* sending acks */
+	unsigned long		i_ack_flags;
+	u64			i_ack_next;	/* next ACK to send */
+	struct rds_header	*i_ack;
+	struct ib_send_wr	i_ack_wr;
+	struct ib_sge		i_ack_sge;
+	u64			i_ack_dma;
+	unsigned long		i_ack_queued;
+
+	/* Flow control related information
+	 *
+	 * Our algorithm uses a pair variables that we need to access
+	 * atomically - one for the send credits, and one posted
+	 * recv credits we need to transfer to remote.
+	 * Rather than protect them using a slow spinlock, we put both into
+	 * a single atomic_t and update it using cmpxchg
+	 */
+	atomic_t		i_credits;
+
+	/* Protocol version specific information */
+	unsigned int		i_flowctl:1;	/* enable/disable flow ctl */
+	unsigned int		i_dma_local_lkey:1;
+	unsigned int		i_fastreg_posted:1; /* fastreg posted on this connection */
+	/* Batched completions */
+	unsigned int		i_unsignaled_wrs;
+	long			i_unsignaled_bytes;
+};
+
+/* This assumes that atomic_t is at least 32 bits */
+#define IB_GET_SEND_CREDITS(v)	((v) & 0xffff)
+#define IB_GET_POST_CREDITS(v)	((v) >> 16)
+#define IB_SET_SEND_CREDITS(v)	((v) & 0xffff)
+#define IB_SET_POST_CREDITS(v)	((v) << 16)
+
+struct rds_iw_cm_id {
+	struct list_head	list;
+	struct rdma_cm_id	*cm_id;
+};
+
+struct rds_iw_device {
+	struct list_head	list;
+	struct list_head	cm_id_list;
+	struct list_head	conn_list;
+	struct ib_device	*dev;
+	struct ib_pd		*pd;
+	struct ib_mr		*mr;
+	struct rds_iw_mr_pool	*mr_pool;
+	int			page_shift;
+	int			max_sge;
+	unsigned int		max_wrs;
+	unsigned int		dma_local_lkey:1;
+	spinlock_t		spinlock;	/* protect the above */
+};
+
+/* bits for i_ack_flags */
+#define IB_ACK_IN_FLIGHT	0
+#define IB_ACK_REQUESTED	1
+
+/* Magic WR_ID for ACKs */
+#define RDS_IW_ACK_WR_ID	((u64)0xffffffffffffffffULL)
+#define RDS_IW_FAST_REG_WR_ID	((u64)0xefefefefefefefefULL)
+#define RDS_IW_LOCAL_INV_WR_ID	((u64)0xdfdfdfdfdfdfdfdfULL)
+
+struct rds_iw_statistics {
+	uint64_t	s_iw_connect_raced;
+	uint64_t	s_iw_listen_closed_stale;
+	uint64_t	s_iw_tx_cq_call;
+	uint64_t	s_iw_tx_cq_event;
+	uint64_t	s_iw_tx_ring_full;
+	uint64_t	s_iw_tx_throttle;
+	uint64_t	s_iw_tx_sg_mapping_failure;
+	uint64_t	s_iw_tx_stalled;
+	uint64_t	s_iw_tx_credit_updates;
+	uint64_t	s_iw_rx_cq_call;
+	uint64_t	s_iw_rx_cq_event;
+	uint64_t	s_iw_rx_ring_empty;
+	uint64_t	s_iw_rx_refill_from_cq;
+	uint64_t	s_iw_rx_refill_from_thread;
+	uint64_t	s_iw_rx_alloc_limit;
+	uint64_t	s_iw_rx_credit_updates;
+	uint64_t	s_iw_ack_sent;
+	uint64_t	s_iw_ack_send_failure;
+	uint64_t	s_iw_ack_send_delayed;
+	uint64_t	s_iw_ack_send_piggybacked;
+	uint64_t	s_iw_ack_received;
+	uint64_t	s_iw_rdma_mr_alloc;
+	uint64_t	s_iw_rdma_mr_free;
+	uint64_t	s_iw_rdma_mr_used;
+	uint64_t	s_iw_rdma_mr_pool_flush;
+	uint64_t	s_iw_rdma_mr_pool_wait;
+	uint64_t	s_iw_rdma_mr_pool_depleted;
+};
+
+extern struct workqueue_struct *rds_iw_wq;
+
+/*
+ * Fake ib_dma_sync_sg_for_{cpu,device} as long as ib_verbs.h
+ * doesn't define it.
+ */
+static inline void rds_iw_dma_sync_sg_for_cpu(struct ib_device *dev,
+		struct scatterlist *sg, unsigned int sg_dma_len, int direction)
+{
+	unsigned int i;
+
+	for (i = 0; i < sg_dma_len; ++i) {
+		ib_dma_sync_single_for_cpu(dev,
+				ib_sg_dma_address(dev, &sg[i]),
+				ib_sg_dma_len(dev, &sg[i]),
+				direction);
+	}
+}
+#define ib_dma_sync_sg_for_cpu	rds_iw_dma_sync_sg_for_cpu
+
+static inline void rds_iw_dma_sync_sg_for_device(struct ib_device *dev,
+		struct scatterlist *sg, unsigned int sg_dma_len, int direction)
+{
+	unsigned int i;
+
+	for (i = 0; i < sg_dma_len; ++i) {
+		ib_dma_sync_single_for_device(dev,
+				ib_sg_dma_address(dev, &sg[i]),
+				ib_sg_dma_len(dev, &sg[i]),
+				direction);
+	}
+}
+#define ib_dma_sync_sg_for_device	rds_iw_dma_sync_sg_for_device
+
+static inline u32 rds_iw_local_dma_lkey(struct rds_iw_connection *ic)
+{
+	return ic->i_dma_local_lkey ? ic->i_cm_id->device->local_dma_lkey : ic->i_mr->lkey;
+}
+
+/* ib.c */
+extern struct rds_transport rds_iw_transport;
+extern void rds_iw_add_one(struct ib_device *device);
+extern void rds_iw_remove_one(struct ib_device *device);
+extern struct ib_client rds_iw_client;
+
+extern unsigned int fastreg_pool_size;
+extern unsigned int fastreg_message_size;
+
+extern spinlock_t iw_nodev_conns_lock;
+extern struct list_head iw_nodev_conns;
+
+/* ib_cm.c */
+int rds_iw_conn_alloc(struct rds_connection *conn, gfp_t gfp);
+void rds_iw_conn_free(void *arg);
+int rds_iw_conn_connect(struct rds_connection *conn);
+void rds_iw_conn_shutdown(struct rds_connection *conn);
+void rds_iw_state_change(struct sock *sk);
+int __init rds_iw_listen_init(void);
+void rds_iw_listen_stop(void);
+void __rds_iw_conn_error(struct rds_connection *conn, const char *, ...);
+int rds_iw_cm_handle_connect(struct rdma_cm_id *cm_id,
+			     struct rdma_cm_event *event);
+int rds_iw_cm_initiate_connect(struct rdma_cm_id *cm_id);
+void rds_iw_cm_connect_complete(struct rds_connection *conn,
+				struct rdma_cm_event *event);
+
+
+#define rds_iw_conn_error(conn, fmt...) \
+	__rds_iw_conn_error(conn, KERN_WARNING "RDS/IW: " fmt)
+
+/* ib_rdma.c */
+int rds_iw_update_cm_id(struct rds_iw_device *rds_iwdev, struct rdma_cm_id *cm_id);
+int rds_iw_add_conn(struct rds_iw_device *rds_iwdev, struct rds_connection *conn);
+void rds_iw_remove_nodev_conns(void);
+void rds_iw_remove_conns(struct rds_iw_device *rds_iwdev);
+struct rds_iw_mr_pool *rds_iw_create_mr_pool(struct rds_iw_device *);
+void rds_iw_get_mr_info(struct rds_iw_device *rds_iwdev, struct rds_info_rdma_connection *iinfo);
+void rds_iw_destroy_mr_pool(struct rds_iw_mr_pool *);
+void *rds_iw_get_mr(struct scatterlist *sg, unsigned long nents,
+		    struct rds_sock *rs, u32 *key_ret);
+void rds_iw_sync_mr(void *trans_private, int dir);
+void rds_iw_free_mr(void *trans_private, int invalidate);
+void rds_iw_flush_mrs(void);
+void rds_iw_remove_cm_id(struct rds_iw_device *rds_iwdev, struct rdma_cm_id *cm_id);
+
+/* ib_recv.c */
+int __init rds_iw_recv_init(void);
+void rds_iw_recv_exit(void);
+int rds_iw_recv(struct rds_connection *conn);
+int rds_iw_recv_refill(struct rds_connection *conn, gfp_t kptr_gfp,
+		       gfp_t page_gfp, int prefill);
+void rds_iw_inc_purge(struct rds_incoming *inc);
+void rds_iw_inc_free(struct rds_incoming *inc);
+int rds_iw_inc_copy_to_user(struct rds_incoming *inc, struct iovec *iov,
+			     size_t size);
+void rds_iw_recv_cq_comp_handler(struct ib_cq *cq, void *context);
+void rds_iw_recv_init_ring(struct rds_iw_connection *ic);
+void rds_iw_recv_clear_ring(struct rds_iw_connection *ic);
+void rds_iw_recv_init_ack(struct rds_iw_connection *ic);
+void rds_iw_attempt_ack(struct rds_iw_connection *ic);
+void rds_iw_ack_send_complete(struct rds_iw_connection *ic);
+u64 rds_iw_piggyb_ack(struct rds_iw_connection *ic);
+
+/* ib_ring.c */
+void rds_iw_ring_init(struct rds_iw_work_ring *ring, u32 nr);
+void rds_iw_ring_resize(struct rds_iw_work_ring *ring, u32 nr);
+u32 rds_iw_ring_alloc(struct rds_iw_work_ring *ring, u32 val, u32 *pos);
+void rds_iw_ring_free(struct rds_iw_work_ring *ring, u32 val);
+void rds_iw_ring_unalloc(struct rds_iw_work_ring *ring, u32 val);
+int rds_iw_ring_empty(struct rds_iw_work_ring *ring);
+int rds_iw_ring_low(struct rds_iw_work_ring *ring);
+u32 rds_iw_ring_oldest(struct rds_iw_work_ring *ring);
+u32 rds_iw_ring_completed(struct rds_iw_work_ring *ring, u32 wr_id, u32 oldest);
+extern wait_queue_head_t rds_iw_ring_empty_wait;
+
+/* ib_send.c */
+void rds_iw_xmit_complete(struct rds_connection *conn);
+int rds_iw_xmit(struct rds_connection *conn, struct rds_message *rm,
+		unsigned int hdr_off, unsigned int sg, unsigned int off);
+void rds_iw_send_cq_comp_handler(struct ib_cq *cq, void *context);
+void rds_iw_send_init_ring(struct rds_iw_connection *ic);
+void rds_iw_send_clear_ring(struct rds_iw_connection *ic);
+int rds_iw_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op);
+void rds_iw_send_add_credits(struct rds_connection *conn, unsigned int credits);
+void rds_iw_advertise_credits(struct rds_connection *conn, unsigned int posted);
+int rds_iw_send_grab_credits(struct rds_iw_connection *ic, u32 wanted,
+			     u32 *adv_credits, int need_posted);
+
+/* ib_stats.c */
+DECLARE_PER_CPU(struct rds_iw_statistics, rds_iw_stats);
+#define rds_iw_stats_inc(member) rds_stats_inc_which(rds_iw_stats, member)
+unsigned int rds_iw_stats_info_copy(struct rds_info_iterator *iter,
+				    unsigned int avail);
+
+/* ib_sysctl.c */
+int __init rds_iw_sysctl_init(void);
+void rds_iw_sysctl_exit(void);
+extern unsigned long rds_iw_sysctl_max_send_wr;
+extern unsigned long rds_iw_sysctl_max_recv_wr;
+extern unsigned long rds_iw_sysctl_max_unsig_wrs;
+extern unsigned long rds_iw_sysctl_max_unsig_bytes;
+extern unsigned long rds_iw_sysctl_max_recv_allocation;
+extern unsigned int rds_iw_sysctl_flow_control;
+extern ctl_table rds_iw_sysctl_table[];
+
+/*
+ * Helper functions for getting/setting the header and data SGEs in
+ * RDS packets (not RDMA)
+ */
+static inline struct ib_sge *
+rds_iw_header_sge(struct rds_iw_connection *ic, struct ib_sge *sge)
+{
+	return &sge[0];
+}
+
+static inline struct ib_sge *
+rds_iw_data_sge(struct rds_iw_connection *ic, struct ib_sge *sge)
+{
+	return &sge[1];
+}
+
+static inline void rds_iw_set_64bit(u64 *ptr, u64 val)
+{
+#if BITS_PER_LONG == 64
+	*ptr = val;
+#else
+	set_64bit(ptr, val);
+#endif
+}
+
+#endif
diff --git a/net/rds/iw_cm.c b/net/rds/iw_cm.c
new file mode 100644
index 0000000..57ecb3d
--- /dev/null
+++ b/net/rds/iw_cm.c
@@ -0,0 +1,750 @@
+/*
+ * Copyright (c) 2006 Oracle.  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
+ * OpenIB.org BSD 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.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/kernel.h>
+#include <linux/in.h>
+#include <linux/vmalloc.h>
+
+#include "rds.h"
+#include "iw.h"
+
+/*
+ * Set the selected protocol version
+ */
+static void rds_iw_set_protocol(struct rds_connection *conn, unsigned int version)
+{
+	conn->c_version = version;
+}
+
+/*
+ * Set up flow control
+ */
+static void rds_iw_set_flow_control(struct rds_connection *conn, u32 credits)
+{
+	struct rds_iw_connection *ic = conn->c_transport_data;
+
+	if (rds_iw_sysctl_flow_control && credits != 0) {
+		/* We're doing flow control */
+		ic->i_flowctl = 1;
+		rds_iw_send_add_credits(conn, credits);
+	} else {
+		ic->i_flowctl = 0;
+	}
+}
+
+/*
+ * Connection established.
+ * We get here for both outgoing and incoming connection.
+ */
+void rds_iw_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_event *event)
+{
+	const struct rds_iw_connect_private *dp = NULL;
+	struct rds_iw_connection *ic = conn->c_transport_data;
+	struct rds_iw_device *rds_iwdev;
+	int err;
+
+	if (event->param.conn.private_data_len) {
+		dp = event->param.conn.private_data;
+
+		rds_iw_set_protocol(conn,
+				RDS_PROTOCOL(dp->dp_protocol_major,
+					dp->dp_protocol_minor));
+		rds_iw_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
+	}
+
+	/* update ib_device with this local ipaddr & conn */
+	rds_iwdev = ib_get_client_data(ic->i_cm_id->device, &rds_iw_client);
+	err = rds_iw_update_cm_id(rds_iwdev, ic->i_cm_id);
+	if (err)
+		printk(KERN_ERR "rds_iw_update_ipaddr failed (%d)\n", err);
+	err = rds_iw_add_conn(rds_iwdev, conn);
+	if (err)
+		printk(KERN_ERR "rds_iw_add_conn failed (%d)\n", err);
+
+	/* If the peer gave us the last packet it saw, process this as if
+	 * we had received a regular ACK. */
+	if (dp && dp->dp_ack_seq)
+		rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL);
+
+	printk(KERN_NOTICE "RDS/IW: connected to %pI4<->%pI4 version %u.%u%s\n",
+			&conn->c_laddr, &conn->c_faddr,
+			RDS_PROTOCOL_MAJOR(conn->c_version),
+			RDS_PROTOCOL_MINOR(conn->c_version),
+			ic->i_flowctl ? ", flow control" : "");
+
+	rds_connect_complete(conn);
+}
+
+static void rds_iw_cm_fill_conn_param(struct rds_connection *conn,
+			struct rdma_conn_param *conn_param,
+			struct rds_iw_connect_private *dp,
+			u32 protocol_version)
+{
+	struct rds_iw_connection *ic = conn->c_transport_data;
+
+	memset(conn_param, 0, sizeof(struct rdma_conn_param));
+	/* XXX tune these? */
+	conn_param->responder_resources = 1;
+	conn_param->initiator_depth = 1;
+
+	if (dp) {
+		memset(dp, 0, sizeof(*dp));
+		dp->dp_saddr = conn->c_laddr;
+		dp->dp_daddr = conn->c_faddr;
+		dp->dp_protocol_major = RDS_PROTOCOL_MAJOR(protocol_version);
+		dp->dp_protocol_minor = RDS_PROTOCOL_MINOR(protocol_version);
+		dp->dp_protocol_minor_mask = cpu_to_be16(RDS_IW_SUPPORTED_PROTOCOLS);
+		dp->dp_ack_seq = rds_iw_piggyb_ack(ic);
+
+		/* Advertise flow control */
+		if (ic->i_flowctl) {
+			unsigned int credits;
+
+			credits = IB_GET_POST_CREDITS(atomic_read(&ic->i_credits));
+			dp->dp_credit = cpu_to_be32(credits);
+			atomic_sub(IB_SET_POST_CREDITS(credits), &ic->i_credits);
+		}
+
+		conn_param->private_data = dp;
+		conn_param->private_data_len = sizeof(*dp);
+	}
+}
+
+static void rds_iw_cq_event_handler(struct ib_event *event, void *data)
+{
+	rdsdebug("event %u data %p\n", event->event, data);
+}
+
+static void rds_iw_qp_event_handler(struct ib_event *event, void *data)
+{
+	struct rds_connection *conn = data;
+	struct rds_iw_connection *ic = conn->c_transport_data;
+
+	rdsdebug("conn %p ic %p event %u\n", conn, ic, event->event);
+
+	switch (event->event) {
+	case IB_EVENT_COMM_EST:
+		rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST);
+		break;
+	case IB_EVENT_QP_REQ_ERR:
+	case IB_EVENT_QP_FATAL:
+	default:
+		rds_iw_conn_error(conn, "RDS/IW: Fatal QP Event %u - connection %pI4->%pI4...reconnecting\n",
+			event->event, &conn->c_laddr,
+			&conn->c_faddr);
+		break;
+	}
+}
+
+/*
+ * Create a QP
+ */
+static int rds_iw_init_qp_attrs(struct ib_qp_init_attr *attr,
+		struct rds_iw_device *rds_iwdev,
+		struct rds_iw_work_ring *send_ring,
+		void (*send_cq_handler)(struct ib_cq *, void *),
+		struct rds_iw_work_ring *recv_ring,
+		void (*recv_cq_handler)(struct ib_cq *, void *),
+		void *context)
+{
+	struct ib_device *dev = rds_iwdev->dev;
+	unsigned int send_size, recv_size;
+	int ret;
+
+	/* The offset of 1 is to accomodate the additional ACK WR. */
+	send_size = min_t(unsigned int, rds_iwdev->max_wrs, rds_iw_sysctl_max_send_wr + 1);
+	recv_size = min_t(unsigned int, rds_iwdev->max_wrs, rds_iw_sysctl_max_recv_wr + 1);
+	rds_iw_ring_resize(send_ring, send_size - 1);
+	rds_iw_ring_resize(recv_ring, recv_size - 1);
+
+	memset(attr, 0, sizeof(*attr));
+	attr->event_handler = rds_iw_qp_event_handler;
+	attr->qp_context = context;
+	attr->cap.max_send_wr = send_size;
+	attr->cap.max_recv_wr = recv_size;
+	attr->cap.max_send_sge = rds_iwdev->max_sge;
+	attr->cap.max_recv_sge = RDS_IW_RECV_SGE;
+	attr->sq_sig_type = IB_SIGNAL_REQ_WR;
+	attr->qp_type = IB_QPT_RC;
+
+	attr->send_cq = ib_create_cq(dev, send_cq_handler,
+				     rds_iw_cq_event_handler,
+				     context, send_size, 0);
+	if (IS_ERR(attr->send_cq)) {
+		ret = PTR_ERR(attr->send_cq);
+		attr->send_cq = NULL;
+		rdsdebug("ib_create_cq send failed: %d\n", ret);
+		goto out;
+	}
+
+	attr->recv_cq = ib_create_cq(dev, recv_cq_handler,
+				     rds_iw_cq_event_handler,
+				     context, recv_size, 0);
+	if (IS_ERR(attr->recv_cq)) {
+		ret = PTR_ERR(attr->recv_cq);
+		attr->recv_cq = NULL;
+		rdsdebug("ib_create_cq send failed: %d\n", ret);
+		goto out;
+	}
+
+	ret = ib_req_notify_cq(attr->send_cq, IB_CQ_NEXT_COMP);
+	if (ret) {
+		rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
+		goto out;
+	}
+
+	ret = ib_req_notify_cq(attr->recv_cq, IB_CQ_SOLICITED);
+	if (ret) {
+		rdsdebug("ib_req_notify_cq recv failed: %d\n", ret);
+		goto out;
+	}
+
+out:
+	if (ret) {
+		if (attr->send_cq)
+			ib_destroy_cq(attr->send_cq);
+		if (attr->recv_cq)
+			ib_destroy_cq(attr->recv_cq);
+	}
+	return ret;
+}
+
+/*
+ * This needs to be very careful to not leave IS_ERR pointers around for
+ * cleanup to trip over.
+ */
+static int rds_iw_setup_qp(struct rds_connection *conn)
+{
+	struct rds_iw_connection *ic = conn->c_transport_data;
+	struct ib_device *dev = ic->i_cm_id->device;
+	struct ib_qp_init_attr attr;
+	struct rds_iw_device *rds_iwdev;
+	int ret;
+
+	/* rds_iw_add_one creates a rds_iw_device object per IB device,
+	 * and allocates a protection domain, memory range and MR pool
+	 * for each.  If that fails for any reason, it will not register
+	 * the rds_iwdev at all.
+	 */
+	rds_iwdev = ib_get_client_data(dev, &rds_iw_client);
+	if (rds_iwdev == NULL) {
+		if (printk_ratelimit())
+			printk(KERN_NOTICE "RDS/IW: No client_data for device %s\n",
+					dev->name);
+		return -EOPNOTSUPP;
+	}
+
+	/* Protection domain and memory range */
+	ic->i_pd = rds_iwdev->pd;
+	ic->i_mr = rds_iwdev->mr;
+
+	ret = rds_iw_init_qp_attrs(&attr, rds_iwdev,
+			&ic->i_send_ring, rds_iw_send_cq_comp_handler,
+			&ic->i_recv_ring, rds_iw_recv_cq_comp_handler,
+			conn);
+	if (ret < 0)
+		goto out;
+
+	ic->i_send_cq = attr.send_cq;
+	ic->i_recv_cq = attr.recv_cq;
+
+	/*
+	 * XXX this can fail if max_*_wr is too large?  Are we supposed
+	 * to back off until we get a value that the hardware can support?
+	 */
+	ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr);
+	if (ret) {
+		rdsdebug("rdma_create_qp failed: %d\n", ret);
+		goto out;
+	}
+
+	ic->i_send_hdrs = ib_dma_alloc_coherent(dev,
+					   ic->i_send_ring.w_nr *
+						sizeof(struct rds_header),
+					   &ic->i_send_hdrs_dma, GFP_KERNEL);
+	if (ic->i_send_hdrs == NULL) {
+		ret = -ENOMEM;
+		rdsdebug("ib_dma_alloc_coherent send failed\n");
+		goto out;
+	}
+
+	ic->i_recv_hdrs = ib_dma_alloc_coherent(dev,
+					   ic->i_recv_ring.w_nr *
+						sizeof(struct rds_header),
+					   &ic->i_recv_hdrs_dma, GFP_KERNEL);
+	if (ic->i_recv_hdrs == NULL) {
+		ret = -ENOMEM;
+		rdsdebug("ib_dma_alloc_coherent recv failed\n");
+		goto out;
+	}
+
+	ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header),
+				       &ic->i_ack_dma, GFP_KERNEL);
+	if (ic->i_ack == NULL) {
+		ret = -ENOMEM;
+		rdsdebug("ib_dma_alloc_coherent ack failed\n");
+		goto out;
+	}
+
+	ic->i_sends = vmalloc(ic->i_send_ring.w_nr * sizeof(struct rds_iw_send_work));
+	if (ic->i_sends == NULL) {
+		ret = -ENOMEM;
+		rdsdebug("send allocation failed\n");
+		goto out;
+	}
+	rds_iw_send_init_ring(ic);
+
+	ic->i_recvs = vmalloc(ic->i_recv_ring.w_nr * sizeof(struct rds_iw_recv_work));
+	if (ic->i_recvs == NULL) {
+		ret = -ENOMEM;
+		rdsdebug("recv allocation failed\n");
+		goto out;
+	}
+
+	rds_iw_recv_init_ring(ic);
+	rds_iw_recv_init_ack(ic);
+
+	/* Post receive buffers - as a side effect, this will update
+	 * the posted credit count. */
+	rds_iw_recv_refill(conn, GFP_KERNEL, GFP_HIGHUSER, 1);
+
+	rdsdebug("conn %p pd %p mr %p cq %p %p\n", conn, ic->i_pd, ic->i_mr,
+		 ic->i_send_cq, ic->i_recv_cq);
+
+out:
+	return ret;
+}
+
+static u32 rds_iw_protocol_compatible(const struct rds_iw_connect_private *dp)
+{
+	u16 common;
+	u32 version = 0;
+
+	/* rdma_cm private data is odd - when there is any private data in the
+	 * request, we will be given a pretty large buffer without telling us the
+	 * original size. The only way to tell the difference is by looking at
+	 * the contents, which are initialized to zero.
+	 * If the protocol version fields aren't set, this is a connection attempt
+	 * from an older version. This could could be 3.0 or 2.0 - we can't tell.
+	 * We really should have changed this for OFED 1.3 :-( */
+	if (dp->dp_protocol_major == 0)
+		return RDS_PROTOCOL_3_0;
+
+	common = be16_to_cpu(dp->dp_protocol_minor_mask) & RDS_IW_SUPPORTED_PROTOCOLS;
+	if (dp->dp_protocol_major == 3 && common) {
+		version = RDS_PROTOCOL_3_0;
+		while ((common >>= 1) != 0)
+			version++;
+	} else if (printk_ratelimit()) {
+		printk(KERN_NOTICE "RDS: Connection from %pI4 using "
+			"incompatible protocol version %u.%u\n",
+			&dp->dp_saddr,
+			dp->dp_protocol_major,
+			dp->dp_protocol_minor);
+	}
+	return version;
+}
+
+int rds_iw_cm_handle_connect(struct rdma_cm_id *cm_id,
+				    struct rdma_cm_event *event)
+{
+	const struct rds_iw_connect_private *dp = event->param.conn.private_data;
+	struct rds_iw_connect_private dp_rep;
+	struct rds_connection *conn = NULL;
+	struct rds_iw_connection *ic = NULL;
+	struct rdma_conn_param conn_param;
+	struct rds_iw_device *rds_iwdev;
+	u32 version;
+	int err, destroy = 1;
+
+	/* Check whether the remote protocol version matches ours. */
+	version = rds_iw_protocol_compatible(dp);
+	if (!version)
+		goto out;
+
+	rdsdebug("saddr %pI4 daddr %pI4 RDSv%u.%u\n",
+		 &dp->dp_saddr, &dp->dp_daddr,
+		 RDS_PROTOCOL_MAJOR(version), RDS_PROTOCOL_MINOR(version));
+
+	conn = rds_conn_create(dp->dp_daddr, dp->dp_saddr, &rds_iw_transport,
+			       GFP_KERNEL);
+	if (IS_ERR(conn)) {
+		rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn));
+		conn = NULL;
+		goto out;
+	}
+
+	/*
+	 * The connection request may occur while the
+	 * previous connection exist, e.g. in case of failover.
+	 * But as connections may be initiated simultaneously
+	 * by both hosts, we have a random backoff mechanism -
+	 * see the comment above rds_queue_reconnect()
+	 */
+	mutex_lock(&conn->c_cm_lock);
+	if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
+		if (rds_conn_state(conn) == RDS_CONN_UP) {
+			rdsdebug("incoming connect while connecting\n");
+			rds_conn_drop(conn);
+			rds_iw_stats_inc(s_iw_listen_closed_stale);
+		} else
+		if (rds_conn_state(conn) == RDS_CONN_CONNECTING) {
+			/* Wait and see - our connect may still be succeeding */
+			rds_iw_stats_inc(s_iw_connect_raced);
+		}
+		mutex_unlock(&conn->c_cm_lock);
+		goto out;
+	}
+
+	ic = conn->c_transport_data;
+
+	rds_iw_set_protocol(conn, version);
+	rds_iw_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
+
+	/* If the peer gave us the last packet it saw, process this as if
+	 * we had received a regular ACK. */
+	if (dp->dp_ack_seq)
+		rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL);
+
+	BUG_ON(cm_id->context);
+	BUG_ON(ic->i_cm_id);
+
+	ic->i_cm_id = cm_id;
+	cm_id->context = conn;
+
+	rds_iwdev = ib_get_client_data(cm_id->device, &rds_iw_client);
+	ic->i_dma_local_lkey = rds_iwdev->dma_local_lkey;
+
+	/* We got halfway through setting up the ib_connection, if we
+	 * fail now, we have to take the long route out of this mess. */
+	destroy = 0;
+
+	err = rds_iw_setup_qp(conn);
+	if (err) {
+		rds_iw_conn_error(conn, "rds_iw_setup_qp failed (%d)\n", err);
+		goto out;
+	}
+
+	rds_iw_cm_fill_conn_param(conn, &conn_param, &dp_rep, version);
+
+	/* rdma_accept() calls rdma_reject() internally if it fails */
+	err = rdma_accept(cm_id, &conn_param);
+	mutex_unlock(&conn->c_cm_lock);
+	if (err) {
+		rds_iw_conn_error(conn, "rdma_accept failed (%d)\n", err);
+		goto out;
+	}
+
+	return 0;
+
+out:
+	rdma_reject(cm_id, NULL, 0);
+	return destroy;
+}
+
+
+int rds_iw_cm_initiate_connect(struct rdma_cm_id *cm_id)
+{
+	struct rds_connection *conn = cm_id->context;
+	struct rds_iw_connection *ic = conn->c_transport_data;
+	struct rdma_conn_param conn_param;
+	struct rds_iw_connect_private dp;
+	int ret;
+
+	/* If the peer doesn't do protocol negotiation, we must
+	 * default to RDSv3.0 */
+	rds_iw_set_protocol(conn, RDS_PROTOCOL_3_0);
+	ic->i_flowctl = rds_iw_sysctl_flow_control;	/* advertise flow control */
+
+	ret = rds_iw_setup_qp(conn);
+	if (ret) {
+		rds_iw_conn_error(conn, "rds_iw_setup_qp failed (%d)\n", ret);
+		goto out;
+	}
+
+	rds_iw_cm_fill_conn_param(conn, &conn_param, &dp, RDS_PROTOCOL_VERSION);
+
+	ret = rdma_connect(cm_id, &conn_param);
+	if (ret)
+		rds_iw_conn_error(conn, "rdma_connect failed (%d)\n", ret);
+
+out:
+	/* Beware - returning non-zero tells the rdma_cm to destroy
+	 * the cm_id. We should certainly not do it as long as we still
+	 * "own" the cm_id. */
+	if (ret) {
+		struct rds_iw_connection *ic = conn->c_transport_data;
+
+		if (ic->i_cm_id == cm_id)
+			ret = 0;
+	}
+	return ret;
+}
+
+int rds_iw_conn_connect(struct rds_connection *conn)
+{
+	struct rds_iw_connection *ic = conn->c_transport_data;
+	struct rds_iw_device *rds_iwdev;
+	struct sockaddr_in src, dest;
+	int ret;
+
+	/* XXX I wonder what affect the port space has */
+	/* delegate cm event handler to rdma_transport */
+	ic->i_cm_id = rdma_create_id(rds_rdma_cm_event_handler, conn,
+				     RDMA_PS_TCP);
+	if (IS_ERR(ic->i_cm_id)) {
+		ret = PTR_ERR(ic->i_cm_id);
+		ic->i_cm_id = NULL;
+		rdsdebug("rdma_create_id() failed: %d\n", ret);
+		goto out;
+	}
+
+	rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn);
+
+	src.sin_family = AF_INET;
+	src.sin_addr.s_addr = (__force u32)conn->c_laddr;
+	src.sin_port = (__force u16)htons(0);
+
+	/* First, bind to the local address and device. */
+	ret = rdma_bind_addr(ic->i_cm_id, (struct sockaddr *) &src);
+	if (ret) {
+		rdsdebug("rdma_bind_addr(%pI4) failed: %d\n",
+				&conn->c_laddr, ret);
+		rdma_destroy_id(ic->i_cm_id);
+		ic->i_cm_id = NULL;
+		goto out;
+	}
+
+	rds_iwdev = ib_get_client_data(ic->i_cm_id->device, &rds_iw_client);
+	ic->i_dma_local_lkey = rds_iwdev->dma_local_lkey;
+
+	dest.sin_family = AF_INET;
+	dest.sin_addr.s_addr = (__force u32)conn->c_faddr;
+	dest.sin_port = (__force u16)htons(RDS_PORT);
+
+	ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src,
+				(struct sockaddr *)&dest,
+				RDS_RDMA_RESOLVE_TIMEOUT_MS);
+	if (ret) {
+		rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id,
+			 ret);
+		rdma_destroy_id(ic->i_cm_id);
+		ic->i_cm_id = NULL;
+	}
+
+out:
+	return ret;
+}
+
+/*
+ * This is so careful about only cleaning up resources that were built up
+ * so that it can be called at any point during startup.  In fact it
+ * can be called multiple times for a given connection.
+ */
+void rds_iw_conn_shutdown(struct rds_connection *conn)
+{
+	struct rds_iw_connection *ic = conn->c_transport_data;
+	int err = 0;
+	struct ib_qp_attr qp_attr;
+
+	rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id,
+		 ic->i_pd, ic->i_send_cq, ic->i_recv_cq,
+		 ic->i_cm_id ? ic->i_cm_id->qp : NULL);
+
+	if (ic->i_cm_id) {
+		struct ib_device *dev = ic->i_cm_id->device;
+
+		rdsdebug("disconnecting cm %p\n", ic->i_cm_id);
+		err = rdma_disconnect(ic->i_cm_id);
+		if (err) {
+			/* Actually this may happen quite frequently, when
+			 * an outgoing connect raced with an incoming connect.
+			 */
+			rdsdebug("rds_iw_conn_shutdown: failed to disconnect,"
+				   " cm: %p err %d\n", ic->i_cm_id, err);
+		}
+
+		if (ic->i_cm_id->qp) {
+			qp_attr.qp_state = IB_QPS_ERR;
+			ib_modify_qp(ic->i_cm_id->qp, &qp_attr, IB_QP_STATE);
+		}
+
+		wait_event(rds_iw_ring_empty_wait,
+			rds_iw_ring_empty(&ic->i_send_ring) &&
+			rds_iw_ring_empty(&ic->i_recv_ring));
+
+		if (ic->i_send_hdrs)
+			ib_dma_free_coherent(dev,
+					   ic->i_send_ring.w_nr *
+						sizeof(struct rds_header),
+					   ic->i_send_hdrs,
+					   ic->i_send_hdrs_dma);
+
+		if (ic->i_recv_hdrs)
+			ib_dma_free_coherent(dev,
+					   ic->i_recv_ring.w_nr *
+						sizeof(struct rds_header),
+					   ic->i_recv_hdrs,
+					   ic->i_recv_hdrs_dma);
+
+		if (ic->i_ack)
+			ib_dma_free_coherent(dev, sizeof(struct rds_header),
+					     ic->i_ack, ic->i_ack_dma);
+
+		if (ic->i_sends)
+			rds_iw_send_clear_ring(ic);
+		if (ic->i_recvs)
+			rds_iw_recv_clear_ring(ic);
+
+		if (ic->i_cm_id->qp)
+			rdma_destroy_qp(ic->i_cm_id);
+		if (ic->i_send_cq)
+			ib_destroy_cq(ic->i_send_cq);
+		if (ic->i_recv_cq)
+			ib_destroy_cq(ic->i_recv_cq);
+
+		/*
+		 * If associated with an rds_iw_device:
+		 * 	Move connection back to the nodev list.
+		 * 	Remove cm_id from the device cm_id list.
+		 */
+		if (ic->rds_iwdev) {
+
+			spin_lock_irq(&ic->rds_iwdev->spinlock);
+			BUG_ON(list_empty(&ic->iw_node));
+			list_del(&ic->iw_node);
+			spin_unlock_irq(&ic->rds_iwdev->spinlock);
+
+			spin_lock_irq(&iw_nodev_conns_lock);
+			list_add_tail(&ic->iw_node, &iw_nodev_conns);
+			spin_unlock_irq(&iw_nodev_conns_lock);
+			rds_iw_remove_cm_id(ic->rds_iwdev, ic->i_cm_id);
+			ic->rds_iwdev = NULL;
+		}
+
+		rdma_destroy_id(ic->i_cm_id);
+
+		ic->i_cm_id = NULL;
+		ic->i_pd = NULL;
+		ic->i_mr = NULL;
+		ic->i_send_cq = NULL;
+		ic->i_recv_cq = NULL;
+		ic->i_send_hdrs = NULL;
+		ic->i_recv_hdrs = NULL;
+		ic->i_ack = NULL;
+	}
+	BUG_ON(ic->rds_iwdev);
+
+	/* Clear pending transmit */
+	if (ic->i_rm) {
+		rds_message_put(ic->i_rm);
+		ic->i_rm = NULL;
+	}
+
+	/* Clear the ACK state */
+	clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
+	rds_iw_set_64bit(&ic->i_ack_next, 0);
+	ic->i_ack_recv = 0;
+
+	/* Clear flow control state */
+	ic->i_flowctl = 0;
+	atomic_set(&ic->i_credits, 0);
+
+	rds_iw_ring_init(&ic->i_send_ring, rds_iw_sysctl_max_send_wr);
+	rds_iw_ring_init(&ic->i_recv_ring, rds_iw_sysctl_max_recv_wr);
+
+	if (ic->i_iwinc) {
+		rds_inc_put(&ic->i_iwinc->ii_inc);
+		ic->i_iwinc = NULL;
+	}
+
+	vfree(ic->i_sends);
+	ic->i_sends = NULL;
+	vfree(ic->i_recvs);
+	ic->i_recvs = NULL;
+	rdsdebug("shutdown complete\n");
+}
+
+int rds_iw_conn_alloc(struct rds_connection *conn, gfp_t gfp)
+{
+	struct rds_iw_connection *ic;
+	unsigned long flags;
+
+	/* XXX too lazy? */
+	ic = kzalloc(sizeof(struct rds_iw_connection), GFP_KERNEL);
+	if (ic == NULL)
+		return -ENOMEM;
+
+	INIT_LIST_HEAD(&ic->iw_node);
+	mutex_init(&ic->i_recv_mutex);
+
+	/*
+	 * rds_iw_conn_shutdown() waits for these to be emptied so they
+	 * must be initialized before it can be called.
+	 */
+	rds_iw_ring_init(&ic->i_send_ring, rds_iw_sysctl_max_send_wr);
+	rds_iw_ring_init(&ic->i_recv_ring, rds_iw_sysctl_max_recv_wr);
+
+	ic->conn = conn;
+	conn->c_transport_data = ic;
+
+	spin_lock_irqsave(&iw_nodev_conns_lock, flags);
+	list_add_tail(&ic->iw_node, &iw_nodev_conns);
+	spin_unlock_irqrestore(&iw_nodev_conns_lock, flags);
+
+
+	rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data);
+	return 0;
+}
+
+void rds_iw_conn_free(void *arg)
+{
+	struct rds_iw_connection *ic = arg;
+	rdsdebug("ic %p\n", ic);
+	list_del(&ic->iw_node);
+	kfree(ic);
+}
+
+/*
+ * An error occurred on the connection
+ */
+void
+__rds_iw_conn_error(struct rds_connection *conn, const char *fmt, ...)
+{
+	va_list ap;
+
+	rds_conn_drop(conn);
+
+	va_start(ap, fmt);
+	vprintk(fmt, ap);
+	va_end(ap);
+}
diff --git a/net/rds/iw_rdma.c b/net/rds/iw_rdma.c
new file mode 100644
index 0000000..1c02a8f
--- /dev/null
+++ b/net/rds/iw_rdma.c
@@ -0,0 +1,888 @@
+/*
+ * Copyright (c) 2006 Oracle.  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
+ * OpenIB.org BSD 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.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/kernel.h>
+
+#include "rds.h"
+#include "rdma.h"
+#include "iw.h"
+
+
+/*
+ * This is stored as mr->r_trans_private.
+ */
+struct rds_iw_mr {
+	struct rds_iw_device	*device;
+	struct rds_iw_mr_pool	*pool;
+	struct rdma_cm_id	*cm_id;
+
+	struct ib_mr	*mr;
+	struct ib_fast_reg_page_list *page_list;
+
+	struct rds_iw_mapping	mapping;
+	unsigned char		remap_count;
+};
+
+/*
+ * Our own little MR pool
+ */
+struct rds_iw_mr_pool {
+	struct rds_iw_device	*device;		/* back ptr to the device that owns us */
+
+	struct mutex		flush_lock;		/* serialize fmr invalidate */
+	struct work_struct	flush_worker;		/* flush worker */
+
+	spinlock_t		list_lock;		/* protect variables below */
+	atomic_t		item_count;		/* total # of MRs */
+	atomic_t		dirty_count;		/* # dirty of MRs */
+	struct list_head	dirty_list;		/* dirty mappings */
+	struct list_head	clean_list;		/* unused & unamapped MRs */
+	atomic_t		free_pinned;		/* memory pinned by free MRs */
+	unsigned long		max_message_size;	/* in pages */
+	unsigned long		max_items;
+	unsigned long		max_items_soft;
+	unsigned long		max_free_pinned;
+	int			max_pages;
+};
+
+static int rds_iw_flush_mr_pool(struct rds_iw_mr_pool *pool, int free_all);
+static void rds_iw_mr_pool_flush_worker(struct work_struct *work);
+static int rds_iw_init_fastreg(struct rds_iw_mr_pool *pool, struct rds_iw_mr *ibmr);
+static int rds_iw_map_fastreg(struct rds_iw_mr_pool *pool,
+			  struct rds_iw_mr *ibmr,
+			  struct scatterlist *sg, unsigned int nents);
+static void rds_iw_free_fastreg(struct rds_iw_mr_pool *pool, struct rds_iw_mr *ibmr);
+static unsigned int rds_iw_unmap_fastreg_list(struct rds_iw_mr_pool *pool,
+			struct list_head *unmap_list,
+			struct list_head *kill_list);
+static void rds_iw_destroy_fastreg(struct rds_iw_mr_pool *pool, struct rds_iw_mr *ibmr);
+
+static int rds_iw_get_device(struct rds_sock *rs, struct rds_iw_device **rds_iwdev, struct rdma_cm_id **cm_id)
+{
+	struct rds_iw_device *iwdev;
+	struct rds_iw_cm_id *i_cm_id;
+
+	*rds_iwdev = NULL;
+	*cm_id = NULL;
+
+	list_for_each_entry(iwdev, &rds_iw_devices, list) {
+		spin_lock_irq(&iwdev->spinlock);
+		list_for_each_entry(i_cm_id, &iwdev->cm_id_list, list) {
+			struct sockaddr_in *src_addr, *dst_addr;
+
+			src_addr = (struct sockaddr_in *)&i_cm_id->cm_id->route.addr.src_addr;
+			dst_addr = (struct sockaddr_in *)&i_cm_id->cm_id->route.addr.dst_addr;
+
+			rdsdebug("local ipaddr = %x port %d, "
+				 "remote ipaddr = %x port %d"
+				 "..looking for %x port %d, "
+				 "remote ipaddr = %x port %d\n",
+				src_addr->sin_addr.s_addr,
+				src_addr->sin_port,
+				dst_addr->sin_addr.s_addr,
+				dst_addr->sin_port,
+				rs->rs_bound_addr,
+				rs->rs_bound_port,
+				rs->rs_conn_addr,
+				rs->rs_conn_port);
+#ifdef WORKING_TUPLE_DETECTION
+			if (src_addr->sin_addr.s_addr == rs->rs_bound_addr &&
+			    src_addr->sin_port == rs->rs_bound_port &&
+			    dst_addr->sin_addr.s_addr == rs->rs_conn_addr &&
+			    dst_addr->sin_port == rs->rs_conn_port) {
+#else
+			/* FIXME - needs to compare the local and remote
+			 * ipaddr/port tuple, but the ipaddr is the only
+			 * available infomation in the rds_sock (as the rest are
+			 * zero'ed.  It doesn't appear to be properly populated
+			 * during connection setup...
+			 */
+			if (src_addr->sin_addr.s_addr == rs->rs_bound_addr) {
+#endif
+				spin_unlock_irq(&iwdev->spinlock);
+				*rds_iwdev = iwdev;
+				*cm_id = i_cm_id->cm_id;
+				return 0;
+			}
+		}
+		spin_unlock_irq(&iwdev->spinlock);
+	}
+
+	return 1;
+}
+
+static int rds_iw_add_cm_id(struct rds_iw_device *rds_iwdev, struct rdma_cm_id *cm_id)
+{
+	struct rds_iw_cm_id *i_cm_id;
+
+	i_cm_id = kmalloc(sizeof *i_cm_id, GFP_KERNEL);
+	if (!i_cm_id)
+		return -ENOMEM;
+
+	i_cm_id->cm_id = cm_id;
+
+	spin_lock_irq(&rds_iwdev->spinlock);
+	list_add_tail(&i_cm_id->list, &rds_iwdev->cm_id_list);
+	spin_unlock_irq(&rds_iwdev->spinlock);
+
+	return 0;
+}
+
+void rds_iw_remove_cm_id(struct rds_iw_device *rds_iwdev, struct rdma_cm_id *cm_id)
+{
+	struct rds_iw_cm_id *i_cm_id;
+
+	spin_lock_irq(&rds_iwdev->spinlock);
+	list_for_each_entry(i_cm_id, &rds_iwdev->cm_id_list, list) {
+		if (i_cm_id->cm_id == cm_id) {
+			list_del(&i_cm_id->list);
+			kfree(i_cm_id);
+			break;
+		}
+	}
+	spin_unlock_irq(&rds_iwdev->spinlock);
+}
+
+
+int rds_iw_update_cm_id(struct rds_iw_device *rds_iwdev, struct rdma_cm_id *cm_id)
+{
+	struct sockaddr_in *src_addr, *dst_addr;
+	struct rds_iw_device *rds_iwdev_old;
+	struct rds_sock rs;
+	struct rdma_cm_id *pcm_id;
+	int rc;
+
+	src_addr = (struct sockaddr_in *)&cm_id->route.addr.src_addr;
+	dst_addr = (struct sockaddr_in *)&cm_id->route.addr.dst_addr;
+
+	rs.rs_bound_addr = src_addr->sin_addr.s_addr;
+	rs.rs_bound_port = src_addr->sin_port;
+	rs.rs_conn_addr = dst_addr->sin_addr.s_addr;
+	rs.rs_conn_port = dst_addr->sin_port;
+
+	rc = rds_iw_get_device(&rs, &rds_iwdev_old, &pcm_id);
+	if (rc)
+		rds_iw_remove_cm_id(rds_iwdev, cm_id);
+
+	return rds_iw_add_cm_id(rds_iwdev, cm_id);
+}
+
+int rds_iw_add_conn(struct rds_iw_device *rds_iwdev, struct rds_connection *conn)
+{
+	struct rds_iw_connection *ic = conn->c_transport_data;
+
+	/* conn was previously on the nodev_conns_list */
+	spin_lock_irq(&iw_nodev_conns_lock);
+	BUG_ON(list_empty(&iw_nodev_conns));
+	BUG_ON(list_empty(&ic->iw_node));
+	list_del(&ic->iw_node);
+	spin_unlock_irq(&iw_nodev_conns_lock);
+
+	spin_lock_irq(&rds_iwdev->spinlock);
+	list_add_tail(&ic->iw_node, &rds_iwdev->conn_list);
+	spin_unlock_irq(&rds_iwdev->spinlock);
+
+	ic->rds_iwdev = rds_iwdev;
+
+	return 0;
+}
+
+void rds_iw_remove_nodev_conns(void)
+{
+	struct rds_iw_connection *ic, *_ic;
+	LIST_HEAD(tmp_list);
+
+	/* avoid calling conn_destroy with irqs off */
+	spin_lock_irq(&iw_nodev_conns_lock);
+	list_splice(&iw_nodev_conns, &tmp_list);
+	INIT_LIST_HEAD(&iw_nodev_conns);
+	spin_unlock_irq(&iw_nodev_conns_lock);
+
+	list_for_each_entry_safe(ic, _ic, &tmp_list, iw_node) {
+		if (ic->conn->c_passive)
+			rds_conn_destroy(ic->conn->c_passive);
+		rds_conn_destroy(ic->conn);
+	}
+}
+
+void rds_iw_remove_conns(struct rds_iw_device *rds_iwdev)
+{
+	struct rds_iw_connection *ic, *_ic;
+	LIST_HEAD(tmp_list);
+
+	/* avoid calling conn_destroy with irqs off */
+	spin_lock_irq(&rds_iwdev->spinlock);
+	list_splice(&rds_iwdev->conn_list, &tmp_list);
+	INIT_LIST_HEAD(&rds_iwdev->conn_list);
+	spin_unlock_irq(&rds_iwdev->spinlock);
+
+	list_for_each_entry_safe(ic, _ic, &tmp_list, iw_node) {
+		if (ic->conn->c_passive)
+			rds_conn_destroy(ic->conn->c_passive);
+		rds_conn_destroy(ic->conn);
+	}
+}
+
+static void rds_iw_set_scatterlist(struct rds_iw_scatterlist *sg,
+		struct scatterlist *list, unsigned int sg_len)
+{
+	sg->list = list;
+	sg->len = sg_len;
+	sg->dma_len = 0;
+	sg->dma_npages = 0;
+	sg->bytes = 0;
+}
+
+static u64 *rds_iw_map_scatterlist(struct rds_iw_device *rds_iwdev,
+			struct rds_iw_scatterlist *sg,
+			unsigned int dma_page_shift)
+{
+	struct ib_device *dev = rds_iwdev->dev;
+	u64 *dma_pages = NULL;
+	u64 dma_mask;
+	unsigned int dma_page_size;
+	int i, j, ret;
+
+	dma_page_size = 1 << dma_page_shift;
+	dma_mask = dma_page_size - 1;
+
+	WARN_ON(sg->dma_len);
+
+	sg->dma_len = ib_dma_map_sg(dev, sg->list, sg->len, DMA_BIDIRECTIONAL);
+	if (unlikely(!sg->dma_len)) {
+		printk(KERN_WARNING "RDS/IW: dma_map_sg failed!\n");
+		return ERR_PTR(-EBUSY);
+	}
+
+	sg->bytes = 0;
+	sg->dma_npages = 0;
+
+	ret = -EINVAL;
+	for (i = 0; i < sg->dma_len; ++i) {
+		unsigned int dma_len = ib_sg_dma_len(dev, &sg->list[i]);
+		u64 dma_addr = ib_sg_dma_address(dev, &sg->list[i]);
+		u64 end_addr;
+
+		sg->bytes += dma_len;
+
+		end_addr = dma_addr + dma_len;
+		if (dma_addr & dma_mask) {
+			if (i > 0)
+				goto out_unmap;
+			dma_addr &= ~dma_mask;
+		}
+		if (end_addr & dma_mask) {
+			if (i < sg->dma_len - 1)
+				goto out_unmap;
+			end_addr = (end_addr + dma_mask) & ~dma_mask;
+		}
+
+		sg->dma_npages += (end_addr - dma_addr) >> dma_page_shift;
+	}
+
+	/* Now gather the dma addrs into one list */
+	if (sg->dma_npages > fastreg_message_size)
+		goto out_unmap;
+
+	dma_pages = kmalloc(sizeof(u64) * sg->dma_npages, GFP_ATOMIC);
+	if (!dma_pages) {
+		ret = -ENOMEM;
+		goto out_unmap;
+	}
+
+	for (i = j = 0; i < sg->dma_len; ++i) {
+		unsigned int dma_len = ib_sg_dma_len(dev, &sg->list[i]);
+		u64 dma_addr = ib_sg_dma_address(dev, &sg->list[i]);
+		u64 end_addr;
+
+		end_addr = dma_addr + dma_len;
+		dma_addr &= ~dma_mask;
+		for (; dma_addr < end_addr; dma_addr += dma_page_size)
+			dma_pages[j++] = dma_addr;
+		BUG_ON(j > sg->dma_npages);
+	}
+
+	return dma_pages;
+
+out_unmap:
+	ib_dma_unmap_sg(rds_iwdev->dev, sg->list, sg->len, DMA_BIDIRECTIONAL);
+	sg->dma_len = 0;
+	kfree(dma_pages);
+	return ERR_PTR(ret);
+}
+
+
+struct rds_iw_mr_pool *rds_iw_create_mr_pool(struct rds_iw_device *rds_iwdev)
+{
+	struct rds_iw_mr_pool *pool;
+
+	pool = kzalloc(sizeof(*pool), GFP_KERNEL);
+	if (!pool) {
+		printk(KERN_WARNING "RDS/IW: rds_iw_create_mr_pool alloc error\n");
+		return ERR_PTR(-ENOMEM);
+	}
+
+	pool->device = rds_iwdev;
+	INIT_LIST_HEAD(&pool->dirty_list);
+	INIT_LIST_HEAD(&pool->clean_list);
+	mutex_init(&pool->flush_lock);
+	spin_lock_init(&pool->list_lock);
+	INIT_WORK(&pool->flush_worker, rds_iw_mr_pool_flush_worker);
+
+	pool->max_message_size = fastreg_message_size;
+	pool->max_items = fastreg_pool_size;
+	pool->max_free_pinned = pool->max_items * pool->max_message_size / 4;
+	pool->max_pages = fastreg_message_size;
+
+	/* We never allow more than max_items MRs to be allocated.
+	 * When we exceed more than max_items_soft, we start freeing
+	 * items more aggressively.
+	 * Make sure that max_items > max_items_soft > max_items / 2
+	 */
+	pool->max_items_soft = pool->max_items * 3 / 4;
+
+	return pool;
+}
+
+void rds_iw_get_mr_info(struct rds_iw_device *rds_iwdev, struct rds_info_rdma_connection *iinfo)
+{
+	struct rds_iw_mr_pool *pool = rds_iwdev->mr_pool;
+
+	iinfo->rdma_mr_max = pool->max_items;
+	iinfo->rdma_mr_size = pool->max_pages;
+}
+
+void rds_iw_destroy_mr_pool(struct rds_iw_mr_pool *pool)
+{
+	flush_workqueue(rds_wq);
+	rds_iw_flush_mr_pool(pool, 1);
+	BUG_ON(atomic_read(&pool->item_count));
+	BUG_ON(atomic_read(&pool->free_pinned));
+	kfree(pool);
+}
+
+static inline struct rds_iw_mr *rds_iw_reuse_fmr(struct rds_iw_mr_pool *pool)
+{
+	struct rds_iw_mr *ibmr = NULL;
+	unsigned long flags;
+
+	spin_lock_irqsave(&pool->list_lock, flags);
+	if (!list_empty(&pool->clean_list)) {
+		ibmr = list_entry(pool->clean_list.next, struct rds_iw_mr, mapping.m_list);
+		list_del_init(&ibmr->mapping.m_list);
+	}
+	spin_unlock_irqrestore(&pool->list_lock, flags);
+
+	return ibmr;
+}
+
+static struct rds_iw_mr *rds_iw_alloc_mr(struct rds_iw_device *rds_iwdev)
+{
+	struct rds_iw_mr_pool *pool = rds_iwdev->mr_pool;
+	struct rds_iw_mr *ibmr = NULL;
+	int err = 0, iter = 0;
+
+	while (1) {
+		ibmr = rds_iw_reuse_fmr(pool);
+		if (ibmr)
+			return ibmr;
+
+		/* No clean MRs - now we have the choice of either
+		 * allocating a fresh MR up to the limit imposed by the
+		 * driver, or flush any dirty unused MRs.
+		 * We try to avoid stalling in the send path if possible,
+		 * so we allocate as long as we're allowed to.
+		 *
+		 * We're fussy with enforcing the FMR limit, though. If the driver
+		 * tells us we can't use more than N fmrs, we shouldn't start
+		 * arguing with it */
+		if (atomic_inc_return(&pool->item_count) <= pool->max_items)
+			break;
+
+		atomic_dec(&pool->item_count);
+
+		if (++iter > 2) {
+			rds_iw_stats_inc(s_iw_rdma_mr_pool_depleted);
+			return ERR_PTR(-EAGAIN);
+		}
+
+		/* We do have some empty MRs. Flush them out. */
+		rds_iw_stats_inc(s_iw_rdma_mr_pool_wait);
+		rds_iw_flush_mr_pool(pool, 0);
+	}
+
+	ibmr = kzalloc(sizeof(*ibmr), GFP_KERNEL);
+	if (!ibmr) {
+		err = -ENOMEM;
+		goto out_no_cigar;
+	}
+
+	spin_lock_init(&ibmr->mapping.m_lock);
+	INIT_LIST_HEAD(&ibmr->mapping.m_list);
+	ibmr->mapping.m_mr = ibmr;
+
+	err = rds_iw_init_fastreg(pool, ibmr);
+	if (err)
+		goto out_no_cigar;
+
+	rds_iw_stats_inc(s_iw_rdma_mr_alloc);
+	return ibmr;
+
+out_no_cigar:
+	if (ibmr) {
+		rds_iw_destroy_fastreg(pool, ibmr);
+		kfree(ibmr);
+	}
+	atomic_dec(&pool->item_count);
+	return ERR_PTR(err);
+}
+
+void rds_iw_sync_mr(void *trans_private, int direction)
+{
+	struct rds_iw_mr *ibmr = trans_private;
+	struct rds_iw_device *rds_iwdev = ibmr->device;
+
+	switch (direction) {
+	case DMA_FROM_DEVICE:
+		ib_dma_sync_sg_for_cpu(rds_iwdev->dev, ibmr->mapping.m_sg.list,
+			ibmr->mapping.m_sg.dma_len, DMA_BIDIRECTIONAL);
+		break;
+	case DMA_TO_DEVICE:
+		ib_dma_sync_sg_for_device(rds_iwdev->dev, ibmr->mapping.m_sg.list,
+			ibmr->mapping.m_sg.dma_len, DMA_BIDIRECTIONAL);
+		break;
+	}
+}
+
+static inline unsigned int rds_iw_flush_goal(struct rds_iw_mr_pool *pool, int free_all)
+{
+	unsigned int item_count;
+
+	item_count = atomic_read(&pool->item_count);
+	if (free_all)
+		return item_count;
+
+	return 0;
+}
+
+/*
+ * Flush our pool of MRs.
+ * At a minimum, all currently unused MRs are unmapped.
+ * If the number of MRs allocated exceeds the limit, we also try
+ * to free as many MRs as needed to get back to this limit.
+ */
+static int rds_iw_flush_mr_pool(struct rds_iw_mr_pool *pool, int free_all)
+{
+	struct rds_iw_mr *ibmr, *next;
+	LIST_HEAD(unmap_list);
+	LIST_HEAD(kill_list);
+	unsigned long flags;
+	unsigned int nfreed = 0, ncleaned = 0, free_goal;
+	int ret = 0;
+
+	rds_iw_stats_inc(s_iw_rdma_mr_pool_flush);
+
+	mutex_lock(&pool->flush_lock);
+
+	spin_lock_irqsave(&pool->list_lock, flags);
+	/* Get the list of all mappings to be destroyed */
+	list_splice_init(&pool->dirty_list, &unmap_list);
+	if (free_all)
+		list_splice_init(&pool->clean_list, &kill_list);
+	spin_unlock_irqrestore(&pool->list_lock, flags);
+
+	free_goal = rds_iw_flush_goal(pool, free_all);
+
+	/* Batched invalidate of dirty MRs.
+	 * For FMR based MRs, the mappings on the unmap list are
+	 * actually members of an ibmr (ibmr->mapping). They either
+	 * migrate to the kill_list, or have been cleaned and should be
+	 * moved to the clean_list.
+	 * For fastregs, they will be dynamically allocated, and
+	 * will be destroyed by the unmap function.
+	 */
+	if (!list_empty(&unmap_list)) {
+		ncleaned = rds_iw_unmap_fastreg_list(pool, &unmap_list, &kill_list);
+		/* If we've been asked to destroy all MRs, move those
+		 * that were simply cleaned to the kill list */
+		if (free_all)
+			list_splice_init(&unmap_list, &kill_list);
+	}
+
+	/* Destroy any MRs that are past their best before date */
+	list_for_each_entry_safe(ibmr, next, &kill_list, mapping.m_list) {
+		rds_iw_stats_inc(s_iw_rdma_mr_free);
+		list_del(&ibmr->mapping.m_list);
+		rds_iw_destroy_fastreg(pool, ibmr);
+		kfree(ibmr);
+		nfreed++;
+	}
+
+	/* Anything that remains are laundered ibmrs, which we can add
+	 * back to the clean list. */
+	if (!list_empty(&unmap_list)) {
+		spin_lock_irqsave(&pool->list_lock, flags);
+		list_splice(&unmap_list, &pool->clean_list);
+		spin_unlock_irqrestore(&pool->list_lock, flags);
+	}
+
+	atomic_sub(ncleaned, &pool->dirty_count);
+	atomic_sub(nfreed, &pool->item_count);
+
+	mutex_unlock(&pool->flush_lock);
+	return ret;
+}
+
+static void rds_iw_mr_pool_flush_worker(struct work_struct *work)
+{
+	struct rds_iw_mr_pool *pool = container_of(work, struct rds_iw_mr_pool, flush_worker);
+
+	rds_iw_flush_mr_pool(pool, 0);
+}
+
+void rds_iw_free_mr(void *trans_private, int invalidate)
+{
+	struct rds_iw_mr *ibmr = trans_private;
+	struct rds_iw_mr_pool *pool = ibmr->device->mr_pool;
+
+	rdsdebug("RDS/IW: free_mr nents %u\n", ibmr->mapping.m_sg.len);
+	if (!pool)
+		return;
+
+	/* Return it to the pool's free list */
+	rds_iw_free_fastreg(pool, ibmr);
+
+	/* If we've pinned too many pages, request a flush */
+	if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned
+	 || atomic_read(&pool->dirty_count) >= pool->max_items / 10)
+		queue_work(rds_wq, &pool->flush_worker);
+
+	if (invalidate) {
+		if (likely(!in_interrupt())) {
+			rds_iw_flush_mr_pool(pool, 0);
+		} else {
+			/* We get here if the user created a MR marked
+			 * as use_once and invalidate at the same time. */
+			queue_work(rds_wq, &pool->flush_worker);
+		}
+	}
+}
+
+void rds_iw_flush_mrs(void)
+{
+	struct rds_iw_device *rds_iwdev;
+
+	list_for_each_entry(rds_iwdev, &rds_iw_devices, list) {
+		struct rds_iw_mr_pool *pool = rds_iwdev->mr_pool;
+
+		if (pool)
+			rds_iw_flush_mr_pool(pool, 0);
+	}
+}
+
+void *rds_iw_get_mr(struct scatterlist *sg, unsigned long nents,
+		    struct rds_sock *rs, u32 *key_ret)
+{
+	struct rds_iw_device *rds_iwdev;
+	struct rds_iw_mr *ibmr = NULL;
+	struct rdma_cm_id *cm_id;
+	int ret;
+
+	ret = rds_iw_get_device(rs, &rds_iwdev, &cm_id);
+	if (ret || !cm_id) {
+		ret = -ENODEV;
+		goto out;
+	}
+
+	if (!rds_iwdev->mr_pool) {
+		ret = -ENODEV;
+		goto out;
+	}
+
+	ibmr = rds_iw_alloc_mr(rds_iwdev);
+	if (IS_ERR(ibmr))
+		return ibmr;
+
+	ibmr->cm_id = cm_id;
+	ibmr->device = rds_iwdev;
+
+	ret = rds_iw_map_fastreg(rds_iwdev->mr_pool, ibmr, sg, nents);
+	if (ret == 0)
+		*key_ret = ibmr->mr->rkey;
+	else
+		printk(KERN_WARNING "RDS/IW: failed to map mr (errno=%d)\n", ret);
+
+out:
+	if (ret) {
+		if (ibmr)
+			rds_iw_free_mr(ibmr, 0);
+		ibmr = ERR_PTR(ret);
+	}
+	return ibmr;
+}
+
+/*
+ * iWARP fastreg handling
+ *
+ * The life cycle of a fastreg registration is a bit different from
+ * FMRs.
+ * The idea behind fastreg is to have one MR, to which we bind different
+ * mappings over time. To avoid stalling on the expensive map and invalidate
+ * operations, these operations are pipelined on the same send queue on
+ * which we want to send the message containing the r_key.
+ *
+ * This creates a bit of a problem for us, as we do not have the destination
+ * IP in GET_MR, so the connection must be setup prior to the GET_MR call for
+ * RDMA to be correctly setup.  If a fastreg request is present, rds_iw_xmit
+ * will try to queue a LOCAL_INV (if needed) and a FAST_REG_MR work request
+ * before queuing the SEND. When completions for these arrive, they are
+ * dispatched to the MR has a bit set showing that RDMa can be performed.
+ *
+ * There is another interesting aspect that's related to invalidation.
+ * The application can request that a mapping is invalidated in FREE_MR.
+ * The expectation there is that this invalidation step includes ALL
+ * PREVIOUSLY FREED MRs.
+ */
+static int rds_iw_init_fastreg(struct rds_iw_mr_pool *pool,
+				struct rds_iw_mr *ibmr)
+{
+	struct rds_iw_device *rds_iwdev = pool->device;
+	struct ib_fast_reg_page_list *page_list = NULL;
+	struct ib_mr *mr;
+	int err;
+
+	mr = ib_alloc_fast_reg_mr(rds_iwdev->pd, pool->max_message_size);
+	if (IS_ERR(mr)) {
+		err = PTR_ERR(mr);
+
+		printk(KERN_WARNING "RDS/IW: ib_alloc_fast_reg_mr failed (err=%d)\n", err);
+		return err;
+	}
+
+	/* FIXME - this is overkill, but mapping->m_sg.dma_len/mapping->m_sg.dma_npages
+	 * is not filled in.
+	 */
+	page_list = ib_alloc_fast_reg_page_list(rds_iwdev->dev, pool->max_message_size);
+	if (IS_ERR(page_list)) {
+		err = PTR_ERR(page_list);
+
+		printk(KERN_WARNING "RDS/IW: ib_alloc_fast_reg_page_list failed (err=%d)\n", err);
+		ib_dereg_mr(mr);
+		return err;
+	}
+
+	ibmr->page_list = page_list;
+	ibmr->mr = mr;
+	return 0;
+}
+
+static int rds_iw_rdma_build_fastreg(struct rds_iw_mapping *mapping)
+{
+	struct rds_iw_mr *ibmr = mapping->m_mr;
+	struct ib_send_wr f_wr, *failed_wr;
+	int ret;
+
+	/*
+	 * Perform a WR for the fast_reg_mr. Each individual page
+	 * in the sg list is added to the fast reg page list and placed
+	 * inside the fast_reg_mr WR.  The key used is a rolling 8bit
+	 * counter, which should guarantee uniqueness.
+	 */
+	ib_update_fast_reg_key(ibmr->mr, ibmr->remap_count++);
+	mapping->m_rkey = ibmr->mr->rkey;
+
+	memset(&f_wr, 0, sizeof(f_wr));
+	f_wr.wr_id = RDS_IW_FAST_REG_WR_ID;
+	f_wr.opcode = IB_WR_FAST_REG_MR;
+	f_wr.wr.fast_reg.length = mapping->m_sg.bytes;
+	f_wr.wr.fast_reg.rkey = mapping->m_rkey;
+	f_wr.wr.fast_reg.page_list = ibmr->page_list;
+	f_wr.wr.fast_reg.page_list_len = mapping->m_sg.dma_len;
+	f_wr.wr.fast_reg.page_shift = ibmr->device->page_shift;
+	f_wr.wr.fast_reg.access_flags = IB_ACCESS_LOCAL_WRITE |
+				IB_ACCESS_REMOTE_READ |
+				IB_ACCESS_REMOTE_WRITE;
+	f_wr.wr.fast_reg.iova_start = 0;
+	f_wr.send_flags = IB_SEND_SIGNALED;
+
+	failed_wr = &f_wr;
+	ret = ib_post_send(ibmr->cm_id->qp, &f_wr, &failed_wr);
+	BUG_ON(failed_wr != &f_wr);
+	if (ret && printk_ratelimit())
+		printk(KERN_WARNING "RDS/IW: %s:%d ib_post_send returned %d\n",
+			__func__, __LINE__, ret);
+	return ret;
+}
+
+static int rds_iw_rdma_fastreg_inv(struct rds_iw_mr *ibmr)
+{
+	struct ib_send_wr s_wr, *failed_wr;
+	int ret = 0;
+
+	if (!ibmr->cm_id->qp || !ibmr->mr)
+		goto out;
+
+	memset(&s_wr, 0, sizeof(s_wr));
+	s_wr.wr_id = RDS_IW_LOCAL_INV_WR_ID;
+	s_wr.opcode = IB_WR_LOCAL_INV;
+	s_wr.ex.invalidate_rkey = ibmr->mr->rkey;
+	s_wr.send_flags = IB_SEND_SIGNALED;
+
+	failed_wr = &s_wr;
+	ret = ib_post_send(ibmr->cm_id->qp, &s_wr, &failed_wr);
+	if (ret && printk_ratelimit()) {
+		printk(KERN_WARNING "RDS/IW: %s:%d ib_post_send returned %d\n",
+			__func__, __LINE__, ret);
+		goto out;
+	}
+out:
+	return ret;
+}
+
+static int rds_iw_map_fastreg(struct rds_iw_mr_pool *pool,
+			struct rds_iw_mr *ibmr,
+			struct scatterlist *sg,
+			unsigned int sg_len)
+{
+	struct rds_iw_device *rds_iwdev = pool->device;
+	struct rds_iw_mapping *mapping = &ibmr->mapping;
+	u64 *dma_pages;
+	int i, ret = 0;
+
+	rds_iw_set_scatterlist(&mapping->m_sg, sg, sg_len);
+
+	dma_pages = rds_iw_map_scatterlist(rds_iwdev,
+				&mapping->m_sg,
+				rds_iwdev->page_shift);
+	if (IS_ERR(dma_pages)) {
+		ret = PTR_ERR(dma_pages);
+		dma_pages = NULL;
+		goto out;
+	}
+
+	if (mapping->m_sg.dma_len > pool->max_message_size) {
+		ret = -EMSGSIZE;
+		goto out;
+	}
+
+	for (i = 0; i < mapping->m_sg.dma_npages; ++i)
+		ibmr->page_list->page_list[i] = dma_pages[i];
+
+	ret = rds_iw_rdma_build_fastreg(mapping);
+	if (ret)
+		goto out;
+
+	rds_iw_stats_inc(s_iw_rdma_mr_used);
+
+out:
+	kfree(dma_pages);
+
+	return ret;
+}
+
+/*
+ * "Free" a fastreg MR.
+ */
+static void rds_iw_free_fastreg(struct rds_iw_mr_pool *pool,
+		struct rds_iw_mr *ibmr)
+{
+	unsigned long flags;
+	int ret;
+
+	if (!ibmr->mapping.m_sg.dma_len)
+		return;
+
+	ret = rds_iw_rdma_fastreg_inv(ibmr);
+	if (ret)
+		return;
+
+	/* Try to post the LOCAL_INV WR to the queue. */
+	spin_lock_irqsave(&pool->list_lock, flags);
+
+	list_add_tail(&ibmr->mapping.m_list, &pool->dirty_list);
+	atomic_add(ibmr->mapping.m_sg.len, &pool->free_pinned);
+	atomic_inc(&pool->dirty_count);
+
+	spin_unlock_irqrestore(&pool->list_lock, flags);
+}
+
+static unsigned int rds_iw_unmap_fastreg_list(struct rds_iw_mr_pool *pool,
+				struct list_head *unmap_list,
+				struct list_head *kill_list)
+{
+	struct rds_iw_mapping *mapping, *next;
+	unsigned int ncleaned = 0;
+	LIST_HEAD(laundered);
+
+	/* Batched invalidation of fastreg MRs.
+	 * Why do we do it this way, even though we could pipeline unmap
+	 * and remap? The reason is the application semantics - when the
+	 * application requests an invalidation of MRs, it expects all
+	 * previously released R_Keys to become invalid.
+	 *
+	 * If we implement MR reuse naively, we risk memory corruption
+	 * (this has actually been observed). So the default behavior
+	 * requires that a MR goes through an explicit unmap operation before
+	 * we can reuse it again.
+	 *
+	 * We could probably improve on this a little, by allowing immediate
+	 * reuse of a MR on the same socket (eg you could add small
+	 * cache of unused MRs to strct rds_socket - GET_MR could grab one
+	 * of these without requiring an explicit invalidate).
+	 */
+	while (!list_empty(unmap_list)) {
+		unsigned long flags;
+
+		spin_lock_irqsave(&pool->list_lock, flags);
+		list_for_each_entry_safe(mapping, next, unmap_list, m_list) {
+			list_move(&mapping->m_list, &laundered);
+			ncleaned++;
+		}
+		spin_unlock_irqrestore(&pool->list_lock, flags);
+	}
+
+	/* Move all laundered mappings back to the unmap list.
+	 * We do not kill any WRs right now - it doesn't seem the
+	 * fastreg API has a max_remap limit. */
+	list_splice_init(&laundered, unmap_list);
+
+	return ncleaned;
+}
+
+static void rds_iw_destroy_fastreg(struct rds_iw_mr_pool *pool,
+		struct rds_iw_mr *ibmr)
+{
+	if (ibmr->page_list)
+		ib_free_fast_reg_page_list(ibmr->page_list);
+	if (ibmr->mr)
+		ib_dereg_mr(ibmr->mr);
+}
diff --git a/net/rds/iw_recv.c b/net/rds/iw_recv.c
new file mode 100644
index 0000000..a1931f0
--- /dev/null
+++ b/net/rds/iw_recv.c
@@ -0,0 +1,869 @@
+/*
+ * Copyright (c) 2006 Oracle.  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
+ * OpenIB.org BSD 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.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/kernel.h>
+#include <linux/pci.h>
+#include <linux/dma-mapping.h>
+#include <rdma/rdma_cm.h>
+
+#include "rds.h"
+#include "iw.h"
+
+static struct kmem_cache *rds_iw_incoming_slab;
+static struct kmem_cache *rds_iw_frag_slab;
+static atomic_t	rds_iw_allocation = ATOMIC_INIT(0);
+
+static void rds_iw_frag_drop_page(struct rds_page_frag *frag)
+{
+	rdsdebug("frag %p page %p\n", frag, frag->f_page);
+	__free_page(frag->f_page);
+	frag->f_page = NULL;
+}
+
+static void rds_iw_frag_free(struct rds_page_frag *frag)
+{
+	rdsdebug("frag %p page %p\n", frag, frag->f_page);
+	BUG_ON(frag->f_page != NULL);
+	kmem_cache_free(rds_iw_frag_slab, frag);
+}
+
+/*
+ * We map a page at a time.  Its fragments are posted in order.  This
+ * is called in fragment order as the fragments get send completion events.
+ * Only the last frag in the page performs the unmapping.
+ *
+ * It's OK for ring cleanup to call this in whatever order it likes because
+ * DMA is not in flight and so we can unmap while other ring entries still
+ * hold page references in their frags.
+ */
+static void rds_iw_recv_unmap_page(struct rds_iw_connection *ic,
+				   struct rds_iw_recv_work *recv)
+{
+	struct rds_page_frag *frag = recv->r_frag;
+
+	rdsdebug("recv %p frag %p page %p\n", recv, frag, frag->f_page);
+	if (frag->f_mapped)
+		ib_dma_unmap_page(ic->i_cm_id->device,
+			       frag->f_mapped,
+			       RDS_FRAG_SIZE, DMA_FROM_DEVICE);
+	frag->f_mapped = 0;
+}
+
+void rds_iw_recv_init_ring(struct rds_iw_connection *ic)
+{
+	struct rds_iw_recv_work *recv;
+	u32 i;
+
+	for (i = 0, recv = ic->i_recvs; i < ic->i_recv_ring.w_nr; i++, recv++) {
+		struct ib_sge *sge;
+
+		recv->r_iwinc = NULL;
+		recv->r_frag = NULL;
+
+		recv->r_wr.next = NULL;
+		recv->r_wr.wr_id = i;
+		recv->r_wr.sg_list = recv->r_sge;
+		recv->r_wr.num_sge = RDS_IW_RECV_SGE;
+
+		sge = rds_iw_data_sge(ic, recv->r_sge);
+		sge->addr = 0;
+		sge->length = RDS_FRAG_SIZE;
+		sge->lkey = 0;
+
+		sge = rds_iw_header_sge(ic, recv->r_sge);
+		sge->addr = ic->i_recv_hdrs_dma + (i * sizeof(struct rds_header));
+		sge->length = sizeof(struct rds_header);
+		sge->lkey = 0;
+	}
+}
+
+static void rds_iw_recv_clear_one(struct rds_iw_connection *ic,
+				  struct rds_iw_recv_work *recv)
+{
+	if (recv->r_iwinc) {
+		rds_inc_put(&recv->r_iwinc->ii_inc);
+		recv->r_iwinc = NULL;
+	}
+	if (recv->r_frag) {
+		rds_iw_recv_unmap_page(ic, recv);
+		if (recv->r_frag->f_page)
+			rds_iw_frag_drop_page(recv->r_frag);
+		rds_iw_frag_free(recv->r_frag);
+		recv->r_frag = NULL;
+	}
+}
+
+void rds_iw_recv_clear_ring(struct rds_iw_connection *ic)
+{
+	u32 i;
+
+	for (i = 0; i < ic->i_recv_ring.w_nr; i++)
+		rds_iw_recv_clear_one(ic, &ic->i_recvs[i]);
+
+	if (ic->i_frag.f_page)
+		rds_iw_frag_drop_page(&ic->i_frag);
+}
+
+static int rds_iw_recv_refill_one(struct rds_connection *conn,
+				  struct rds_iw_recv_work *recv,
+				  gfp_t kptr_gfp, gfp_t page_gfp)
+{
+	struct rds_iw_connection *ic = conn->c_transport_data;
+	dma_addr_t dma_addr;
+	struct ib_sge *sge;
+	int ret = -ENOMEM;
+
+	if (recv->r_iwinc == NULL) {
+		if (atomic_read(&rds_iw_allocation) >= rds_iw_sysctl_max_recv_allocation) {
+			rds_iw_stats_inc(s_iw_rx_alloc_limit);
+			goto out;
+		}
+		recv->r_iwinc = kmem_cache_alloc(rds_iw_incoming_slab,
+						 kptr_gfp);
+		if (recv->r_iwinc == NULL)
+			goto out;
+		atomic_inc(&rds_iw_allocation);
+		INIT_LIST_HEAD(&recv->r_iwinc->ii_frags);
+		rds_inc_init(&recv->r_iwinc->ii_inc, conn, conn->c_faddr);
+	}
+
+	if (recv->r_frag == NULL) {
+		recv->r_frag = kmem_cache_alloc(rds_iw_frag_slab, kptr_gfp);
+		if (recv->r_frag == NULL)
+			goto out;
+		INIT_LIST_HEAD(&recv->r_frag->f_item);
+		recv->r_frag->f_page = NULL;
+	}
+
+	if (ic->i_frag.f_page == NULL) {
+		ic->i_frag.f_page = alloc_page(page_gfp);
+		if (ic->i_frag.f_page == NULL)
+			goto out;
+		ic->i_frag.f_offset = 0;
+	}
+
+	dma_addr = ib_dma_map_page(ic->i_cm_id->device,
+				  ic->i_frag.f_page,
+				  ic->i_frag.f_offset,
+				  RDS_FRAG_SIZE,
+				  DMA_FROM_DEVICE);
+	if (ib_dma_mapping_error(ic->i_cm_id->device, dma_addr))
+		goto out;
+
+	/*
+	 * Once we get the RDS_PAGE_LAST_OFF frag then rds_iw_frag_unmap()
+	 * must be called on this recv.  This happens as completions hit
+	 * in order or on connection shutdown.
+	 */
+	recv->r_frag->f_page = ic->i_frag.f_page;
+	recv->r_frag->f_offset = ic->i_frag.f_offset;
+	recv->r_frag->f_mapped = dma_addr;
+
+	sge = rds_iw_data_sge(ic, recv->r_sge);
+	sge->addr = dma_addr;
+	sge->length = RDS_FRAG_SIZE;
+
+	sge = rds_iw_header_sge(ic, recv->r_sge);
+	sge->addr = ic->i_recv_hdrs_dma + (recv - ic->i_recvs) * sizeof(struct rds_header);
+	sge->length = sizeof(struct rds_header);
+
+	get_page(recv->r_frag->f_page);
+
+	if (ic->i_frag.f_offset < RDS_PAGE_LAST_OFF) {
+		ic->i_frag.f_offset += RDS_FRAG_SIZE;
+	} else {
+		put_page(ic->i_frag.f_page);
+		ic->i_frag.f_page = NULL;
+		ic->i_frag.f_offset = 0;
+	}
+
+	ret = 0;
+out:
+	return ret;
+}
+
+/*
+ * This tries to allocate and post unused work requests after making sure that
+ * they have all the allocations they need to queue received fragments into
+ * sockets.  The i_recv_mutex is held here so that ring_alloc and _unalloc
+ * pairs don't go unmatched.
+ *
+ * -1 is returned if posting fails due to temporary resource exhaustion.
+ */
+int rds_iw_recv_refill(struct rds_connection *conn, gfp_t kptr_gfp,
+		       gfp_t page_gfp, int prefill)
+{
+	struct rds_iw_connection *ic = conn->c_transport_data;
+	struct rds_iw_recv_work *recv;
+	struct ib_recv_wr *failed_wr;
+	unsigned int posted = 0;
+	int ret = 0;
+	u32 pos;
+
+	while ((prefill || rds_conn_up(conn))
+			&& rds_iw_ring_alloc(&ic->i_recv_ring, 1, &pos)) {
+		if (pos >= ic->i_recv_ring.w_nr) {
+			printk(KERN_NOTICE "Argh - ring alloc returned pos=%u\n",
+					pos);
+			ret = -EINVAL;
+			break;
+		}
+
+		recv = &ic->i_recvs[pos];
+		ret = rds_iw_recv_refill_one(conn, recv, kptr_gfp, page_gfp);
+		if (ret) {
+			ret = -1;
+			break;
+		}
+
+		/* XXX when can this fail? */
+		ret = ib_post_recv(ic->i_cm_id->qp, &recv->r_wr, &failed_wr);
+		rdsdebug("recv %p iwinc %p page %p addr %lu ret %d\n", recv,
+			 recv->r_iwinc, recv->r_frag->f_page,
+			 (long) recv->r_frag->f_mapped, ret);
+		if (ret) {
+			rds_iw_conn_error(conn, "recv post on "
+			       "%pI4 returned %d, disconnecting and "
+			       "reconnecting\n", &conn->c_faddr,
+			       ret);
+			ret = -1;
+			break;
+		}
+
+		posted++;
+	}
+
+	/* We're doing flow control - update the window. */
+	if (ic->i_flowctl && posted)
+		rds_iw_advertise_credits(conn, posted);
+
+	if (ret)
+		rds_iw_ring_unalloc(&ic->i_recv_ring, 1);
+	return ret;
+}
+
+void rds_iw_inc_purge(struct rds_incoming *inc)
+{
+	struct rds_iw_incoming *iwinc;
+	struct rds_page_frag *frag;
+	struct rds_page_frag *pos;
+
+	iwinc = container_of(inc, struct rds_iw_incoming, ii_inc);
+	rdsdebug("purging iwinc %p inc %p\n", iwinc, inc);
+
+	list_for_each_entry_safe(frag, pos, &iwinc->ii_frags, f_item) {
+		list_del_init(&frag->f_item);
+		rds_iw_frag_drop_page(frag);
+		rds_iw_frag_free(frag);
+	}
+}
+
+void rds_iw_inc_free(struct rds_incoming *inc)
+{
+	struct rds_iw_incoming *iwinc;
+
+	iwinc = container_of(inc, struct rds_iw_incoming, ii_inc);
+
+	rds_iw_inc_purge(inc);
+	rdsdebug("freeing iwinc %p inc %p\n", iwinc, inc);
+	BUG_ON(!list_empty(&iwinc->ii_frags));
+	kmem_cache_free(rds_iw_incoming_slab, iwinc);
+	atomic_dec(&rds_iw_allocation);
+	BUG_ON(atomic_read(&rds_iw_allocation) < 0);
+}
+
+int rds_iw_inc_copy_to_user(struct rds_incoming *inc, struct iovec *first_iov,
+			    size_t size)
+{
+	struct rds_iw_incoming *iwinc;
+	struct rds_page_frag *frag;
+	struct iovec *iov = first_iov;
+	unsigned long to_copy;
+	unsigned long frag_off = 0;
+	unsigned long iov_off = 0;
+	int copied = 0;
+	int ret;
+	u32 len;
+
+	iwinc = container_of(inc, struct rds_iw_incoming, ii_inc);
+	frag = list_entry(iwinc->ii_frags.next, struct rds_page_frag, f_item);
+	len = be32_to_cpu(inc->i_hdr.h_len);
+
+	while (copied < size && copied < len) {
+		if (frag_off == RDS_FRAG_SIZE) {
+			frag = list_entry(frag->f_item.next,
+					  struct rds_page_frag, f_item);
+			frag_off = 0;
+		}
+		while (iov_off == iov->iov_len) {
+			iov_off = 0;
+			iov++;
+		}
+
+		to_copy = min(iov->iov_len - iov_off, RDS_FRAG_SIZE - frag_off);
+		to_copy = min_t(size_t, to_copy, size - copied);
+		to_copy = min_t(unsigned long, to_copy, len - copied);
+
+		rdsdebug("%lu bytes to user [%p, %zu] + %lu from frag "
+			 "[%p, %lu] + %lu\n",
+			 to_copy, iov->iov_base, iov->iov_len, iov_off,
+			 frag->f_page, frag->f_offset, frag_off);
+
+		/* XXX needs + offset for multiple recvs per page */
+		ret = rds_page_copy_to_user(frag->f_page,
+					    frag->f_offset + frag_off,
+					    iov->iov_base + iov_off,
+					    to_copy);
+		if (ret) {
+			copied = ret;
+			break;
+		}
+
+		iov_off += to_copy;
+		frag_off += to_copy;
+		copied += to_copy;
+	}
+
+	return copied;
+}
+
+/* ic starts out kzalloc()ed */
+void rds_iw_recv_init_ack(struct rds_iw_connection *ic)
+{
+	struct ib_send_wr *wr = &ic->i_ack_wr;
+	struct ib_sge *sge = &ic->i_ack_sge;
+
+	sge->addr = ic->i_ack_dma;
+	sge->length = sizeof(struct rds_header);
+	sge->lkey = rds_iw_local_dma_lkey(ic);
+
+	wr->sg_list = sge;
+	wr->num_sge = 1;
+	wr->opcode = IB_WR_SEND;
+	wr->wr_id = RDS_IW_ACK_WR_ID;
+	wr->send_flags = IB_SEND_SIGNALED | IB_SEND_SOLICITED;
+}
+
+/*
+ * You'd think that with reliable IB connections you wouldn't need to ack
+ * messages that have been received.  The problem is that IB hardware generates
+ * an ack message before it has DMAed the message into memory.  This creates a
+ * potential message loss if the HCA is disabled for any reason between when it
+ * sends the ack and before the message is DMAed and processed.  This is only a
+ * potential issue if another HCA is available for fail-over.
+ *
+ * When the remote host receives our ack they'll free the sent message from
+ * their send queue.  To decrease the latency of this we always send an ack
+ * immediately after we've received messages.
+ *
+ * For simplicity, we only have one ack in flight at a time.  This puts
+ * pressure on senders to have deep enough send queues to absorb the latency of
+ * a single ack frame being in flight.  This might not be good enough.
+ *
+ * This is implemented by have a long-lived send_wr and sge which point to a
+ * statically allocated ack frame.  This ack wr does not fall under the ring
+ * accounting that the tx and rx wrs do.  The QP attribute specifically makes
+ * room for it beyond the ring size.  Send completion notices its special
+ * wr_id and avoids working with the ring in that case.
+ */
+static void rds_iw_set_ack(struct rds_iw_connection *ic, u64 seq,
+				int ack_required)
+{
+	rds_iw_set_64bit(&ic->i_ack_next, seq);
+	if (ack_required) {
+		smp_mb__before_clear_bit();
+		set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
+	}
+}
+
+static u64 rds_iw_get_ack(struct rds_iw_connection *ic)
+{
+	clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
+	smp_mb__after_clear_bit();
+
+	return ic->i_ack_next;
+}
+
+static void rds_iw_send_ack(struct rds_iw_connection *ic, unsigned int adv_credits)
+{
+	struct rds_header *hdr = ic->i_ack;
+	struct ib_send_wr *failed_wr;
+	u64 seq;
+	int ret;
+
+	seq = rds_iw_get_ack(ic);
+
+	rdsdebug("send_ack: ic %p ack %llu\n", ic, (unsigned long long) seq);
+	rds_message_populate_header(hdr, 0, 0, 0);
+	hdr->h_ack = cpu_to_be64(seq);
+	hdr->h_credit = adv_credits;
+	rds_message_make_checksum(hdr);
+	ic->i_ack_queued = jiffies;
+
+	ret = ib_post_send(ic->i_cm_id->qp, &ic->i_ack_wr, &failed_wr);
+	if (unlikely(ret)) {
+		/* Failed to send. Release the WR, and
+		 * force another ACK.
+		 */
+		clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
+		set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
+
+		rds_iw_stats_inc(s_iw_ack_send_failure);
+		/* Need to finesse this later. */
+		BUG();
+	} else
+		rds_iw_stats_inc(s_iw_ack_sent);
+}
+
+/*
+ * There are 3 ways of getting acknowledgements to the peer:
+ *  1.	We call rds_iw_attempt_ack from the recv completion handler
+ *	to send an ACK-only frame.
+ *	However, there can be only one such frame in the send queue
+ *	at any time, so we may have to postpone it.
+ *  2.	When another (data) packet is transmitted while there's
+ *	an ACK in the queue, we piggyback the ACK sequence number
+ *	on the data packet.
+ *  3.	If the ACK WR is done sending, we get called from the
+ *	send queue completion handler, and check whether there's
+ *	another ACK pending (postponed because the WR was on the
+ *	queue). If so, we transmit it.
+ *
+ * We maintain 2 variables:
+ *  -	i_ack_flags, which keeps track of whether the ACK WR
+ *	is currently in the send queue or not (IB_ACK_IN_FLIGHT)
+ *  -	i_ack_next, which is the last sequence number we received
+ *
+ * Potentially, send queue and receive queue handlers can run concurrently.
+ *
+ * Reconnecting complicates this picture just slightly. When we
+ * reconnect, we may be seeing duplicate packets. The peer
+ * is retransmitting them, because it hasn't seen an ACK for
+ * them. It is important that we ACK these.
+ *
+ * ACK mitigation adds a header flag "ACK_REQUIRED"; any packet with
+ * this flag set *MUST* be acknowledged immediately.
+ */
+
+/*
+ * When we get here, we're called from the recv queue handler.
+ * Check whether we ought to transmit an ACK.
+ */
+void rds_iw_attempt_ack(struct rds_iw_connection *ic)
+{
+	unsigned int adv_credits;
+
+	if (!test_bit(IB_ACK_REQUESTED, &ic->i_ack_flags))
+		return;
+
+	if (test_and_set_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags)) {
+		rds_iw_stats_inc(s_iw_ack_send_delayed);
+		return;
+	}
+
+	/* Can we get a send credit? */
+	if (!rds_iw_send_grab_credits(ic, 1, &adv_credits, 0)) {
+		rds_iw_stats_inc(s_iw_tx_throttle);
+		clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
+		return;
+	}
+
+	clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
+	rds_iw_send_ack(ic, adv_credits);
+}
+
+/*
+ * We get here from the send completion handler, when the
+ * adapter tells us the ACK frame was sent.
+ */
+void rds_iw_ack_send_complete(struct rds_iw_connection *ic)
+{
+	clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
+	rds_iw_attempt_ack(ic);
+}
+
+/*
+ * This is called by the regular xmit code when it wants to piggyback
+ * an ACK on an outgoing frame.
+ */
+u64 rds_iw_piggyb_ack(struct rds_iw_connection *ic)
+{
+	if (test_and_clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags))
+		rds_iw_stats_inc(s_iw_ack_send_piggybacked);
+	return rds_iw_get_ack(ic);
+}
+
+/*
+ * It's kind of lame that we're copying from the posted receive pages into
+ * long-lived bitmaps.  We could have posted the bitmaps and rdma written into
+ * them.  But receiving new congestion bitmaps should be a *rare* event, so
+ * hopefully we won't need to invest that complexity in making it more
+ * efficient.  By copying we can share a simpler core with TCP which has to
+ * copy.
+ */
+static void rds_iw_cong_recv(struct rds_connection *conn,
+			      struct rds_iw_incoming *iwinc)
+{
+	struct rds_cong_map *map;
+	unsigned int map_off;
+	unsigned int map_page;
+	struct rds_page_frag *frag;
+	unsigned long frag_off;
+	unsigned long to_copy;
+	unsigned long copied;
+	uint64_t uncongested = 0;
+	void *addr;
+
+	/* catch completely corrupt packets */
+	if (be32_to_cpu(iwinc->ii_inc.i_hdr.h_len) != RDS_CONG_MAP_BYTES)
+		return;
+
+	map = conn->c_fcong;
+	map_page = 0;
+	map_off = 0;
+
+	frag = list_entry(iwinc->ii_frags.next, struct rds_page_frag, f_item);
+	frag_off = 0;
+
+	copied = 0;
+
+	while (copied < RDS_CONG_MAP_BYTES) {
+		uint64_t *src, *dst;
+		unsigned int k;
+
+		to_copy = min(RDS_FRAG_SIZE - frag_off, PAGE_SIZE - map_off);
+		BUG_ON(to_copy & 7); /* Must be 64bit aligned. */
+
+		addr = kmap_atomic(frag->f_page, KM_SOFTIRQ0);
+
+		src = addr + frag_off;
+		dst = (void *)map->m_page_addrs[map_page] + map_off;
+		for (k = 0; k < to_copy; k += 8) {
+			/* Record ports that became uncongested, ie
+			 * bits that changed from 0 to 1. */
+			uncongested |= ~(*src) & *dst;
+			*dst++ = *src++;
+		}
+		kunmap_atomic(addr, KM_SOFTIRQ0);
+
+		copied += to_copy;
+
+		map_off += to_copy;
+		if (map_off == PAGE_SIZE) {
+			map_off = 0;
+			map_page++;
+		}
+
+		frag_off += to_copy;
+		if (frag_off == RDS_FRAG_SIZE) {
+			frag = list_entry(frag->f_item.next,
+					  struct rds_page_frag, f_item);
+			frag_off = 0;
+		}
+	}
+
+	/* the congestion map is in little endian order */
+	uncongested = le64_to_cpu(uncongested);
+
+	rds_cong_map_updated(map, uncongested);
+}
+
+/*
+ * Rings are posted with all the allocations they'll need to queue the
+ * incoming message to the receiving socket so this can't fail.
+ * All fragments start with a header, so we can make sure we're not receiving
+ * garbage, and we can tell a small 8 byte fragment from an ACK frame.
+ */
+struct rds_iw_ack_state {
+	u64		ack_next;
+	u64		ack_recv;
+	unsigned int	ack_required:1;
+	unsigned int	ack_next_valid:1;
+	unsigned int	ack_recv_valid:1;
+};
+
+static void rds_iw_process_recv(struct rds_connection *conn,
+				struct rds_iw_recv_work *recv, u32 byte_len,
+				struct rds_iw_ack_state *state)
+{
+	struct rds_iw_connection *ic = conn->c_transport_data;
+	struct rds_iw_incoming *iwinc = ic->i_iwinc;
+	struct rds_header *ihdr, *hdr;
+
+	/* XXX shut down the connection if port 0,0 are seen? */
+
+	rdsdebug("ic %p iwinc %p recv %p byte len %u\n", ic, iwinc, recv,
+		 byte_len);
+
+	if (byte_len < sizeof(struct rds_header)) {
+		rds_iw_conn_error(conn, "incoming message "
+		       "from %pI4 didn't inclue a "
+		       "header, disconnecting and "
+		       "reconnecting\n",
+		       &conn->c_faddr);
+		return;
+	}
+	byte_len -= sizeof(struct rds_header);
+
+	ihdr = &ic->i_recv_hdrs[recv - ic->i_recvs];
+
+	/* Validate the checksum. */
+	if (!rds_message_verify_checksum(ihdr)) {
+		rds_iw_conn_error(conn, "incoming message "
+		       "from %pI4 has corrupted header - "
+		       "forcing a reconnect\n",
+		       &conn->c_faddr);
+		rds_stats_inc(s_recv_drop_bad_checksum);
+		return;
+	}
+
+	/* Process the ACK sequence which comes with every packet */
+	state->ack_recv = be64_to_cpu(ihdr->h_ack);
+	state->ack_recv_valid = 1;
+
+	/* Process the credits update if there was one */
+	if (ihdr->h_credit)
+		rds_iw_send_add_credits(conn, ihdr->h_credit);
+
+	if (ihdr->h_sport == 0 && ihdr->h_dport == 0 && byte_len == 0) {
+		/* This is an ACK-only packet. The fact that it gets
+		 * special treatment here is that historically, ACKs
+		 * were rather special beasts.
+		 */
+		rds_iw_stats_inc(s_iw_ack_received);
+
+		/*
+		 * Usually the frags make their way on to incs and are then freed as
+		 * the inc is freed.  We don't go that route, so we have to drop the
+		 * page ref ourselves.  We can't just leave the page on the recv
+		 * because that confuses the dma mapping of pages and each recv's use
+		 * of a partial page.  We can leave the frag, though, it will be
+		 * reused.
+		 *
+		 * FIXME: Fold this into the code path below.
+		 */
+		rds_iw_frag_drop_page(recv->r_frag);
+		return;
+	}
+
+	/*
+	 * If we don't already have an inc on the connection then this
+	 * fragment has a header and starts a message.. copy its header
+	 * into the inc and save the inc so we can hang upcoming fragments
+	 * off its list.
+	 */
+	if (iwinc == NULL) {
+		iwinc = recv->r_iwinc;
+		recv->r_iwinc = NULL;
+		ic->i_iwinc = iwinc;
+
+		hdr = &iwinc->ii_inc.i_hdr;
+		memcpy(hdr, ihdr, sizeof(*hdr));
+		ic->i_recv_data_rem = be32_to_cpu(hdr->h_len);
+
+		rdsdebug("ic %p iwinc %p rem %u flag 0x%x\n", ic, iwinc,
+			 ic->i_recv_data_rem, hdr->h_flags);
+	} else {
+		hdr = &iwinc->ii_inc.i_hdr;
+		/* We can't just use memcmp here; fragments of a
+		 * single message may carry different ACKs */
+		if (hdr->h_sequence != ihdr->h_sequence
+		 || hdr->h_len != ihdr->h_len
+		 || hdr->h_sport != ihdr->h_sport
+		 || hdr->h_dport != ihdr->h_dport) {
+			rds_iw_conn_error(conn,
+				"fragment header mismatch; forcing reconnect\n");
+			return;
+		}
+	}
+
+	list_add_tail(&recv->r_frag->f_item, &iwinc->ii_frags);
+	recv->r_frag = NULL;
+
+	if (ic->i_recv_data_rem > RDS_FRAG_SIZE)
+		ic->i_recv_data_rem -= RDS_FRAG_SIZE;
+	else {
+		ic->i_recv_data_rem = 0;
+		ic->i_iwinc = NULL;
+
+		if (iwinc->ii_inc.i_hdr.h_flags == RDS_FLAG_CONG_BITMAP)
+			rds_iw_cong_recv(conn, iwinc);
+		else {
+			rds_recv_incoming(conn, conn->c_faddr, conn->c_laddr,
+					  &iwinc->ii_inc, GFP_ATOMIC,
+					  KM_SOFTIRQ0);
+			state->ack_next = be64_to_cpu(hdr->h_sequence);
+			state->ack_next_valid = 1;
+		}
+
+		/* Evaluate the ACK_REQUIRED flag *after* we received
+		 * the complete frame, and after bumping the next_rx
+		 * sequence. */
+		if (hdr->h_flags & RDS_FLAG_ACK_REQUIRED) {
+			rds_stats_inc(s_recv_ack_required);
+			state->ack_required = 1;
+		}
+
+		rds_inc_put(&iwinc->ii_inc);
+	}
+}
+
+/*
+ * Plucking the oldest entry from the ring can be done concurrently with
+ * the thread refilling the ring.  Each ring operation is protected by
+ * spinlocks and the transient state of refilling doesn't change the
+ * recording of which entry is oldest.
+ *
+ * This relies on IB only calling one cq comp_handler for each cq so that
+ * there will only be one caller of rds_recv_incoming() per RDS connection.
+ */
+void rds_iw_recv_cq_comp_handler(struct ib_cq *cq, void *context)
+{
+	struct rds_connection *conn = context;
+	struct rds_iw_connection *ic = conn->c_transport_data;
+	struct ib_wc wc;
+	struct rds_iw_ack_state state = { 0, };
+	struct rds_iw_recv_work *recv;
+
+	rdsdebug("conn %p cq %p\n", conn, cq);
+
+	rds_iw_stats_inc(s_iw_rx_cq_call);
+
+	ib_req_notify_cq(cq, IB_CQ_SOLICITED);
+
+	while (ib_poll_cq(cq, 1, &wc) > 0) {
+		rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
+			 (unsigned long long)wc.wr_id, wc.status, wc.byte_len,
+			 be32_to_cpu(wc.ex.imm_data));
+		rds_iw_stats_inc(s_iw_rx_cq_event);
+
+		recv = &ic->i_recvs[rds_iw_ring_oldest(&ic->i_recv_ring)];
+
+		rds_iw_recv_unmap_page(ic, recv);
+
+		/*
+		 * Also process recvs in connecting state because it is possible
+		 * to get a recv completion _before_ the rdmacm ESTABLISHED
+		 * event is processed.
+		 */
+		if (rds_conn_up(conn) || rds_conn_connecting(conn)) {
+			/* We expect errors as the qp is drained during shutdown */
+			if (wc.status == IB_WC_SUCCESS) {
+				rds_iw_process_recv(conn, recv, wc.byte_len, &state);
+			} else {
+				rds_iw_conn_error(conn, "recv completion on "
+				       "%pI4 had status %u, disconnecting and "
+				       "reconnecting\n", &conn->c_faddr,
+				       wc.status);
+			}
+		}
+
+		rds_iw_ring_free(&ic->i_recv_ring, 1);
+	}
+
+	if (state.ack_next_valid)
+		rds_iw_set_ack(ic, state.ack_next, state.ack_required);
+	if (state.ack_recv_valid && state.ack_recv > ic->i_ack_recv) {
+		rds_send_drop_acked(conn, state.ack_recv, NULL);
+		ic->i_ack_recv = state.ack_recv;
+	}
+	if (rds_conn_up(conn))
+		rds_iw_attempt_ack(ic);
+
+	/* If we ever end up with a really empty receive ring, we're
+	 * in deep trouble, as the sender will definitely see RNR
+	 * timeouts. */
+	if (rds_iw_ring_empty(&ic->i_recv_ring))
+		rds_iw_stats_inc(s_iw_rx_ring_empty);
+
+	/*
+	 * If the ring is running low, then schedule the thread to refill.
+	 */
+	if (rds_iw_ring_low(&ic->i_recv_ring))
+		queue_delayed_work(rds_wq, &conn->c_recv_w, 0);
+}
+
+int rds_iw_recv(struct rds_connection *conn)
+{
+	struct rds_iw_connection *ic = conn->c_transport_data;
+	int ret = 0;
+
+	rdsdebug("conn %p\n", conn);
+
+	/*
+	 * If we get a temporary posting failure in this context then
+	 * we're really low and we want the caller to back off for a bit.
+	 */
+	mutex_lock(&ic->i_recv_mutex);
+	if (rds_iw_recv_refill(conn, GFP_KERNEL, GFP_HIGHUSER, 0))
+		ret = -ENOMEM;
+	else
+		rds_iw_stats_inc(s_iw_rx_refill_from_thread);
+	mutex_unlock(&ic->i_recv_mutex);
+
+	if (rds_conn_up(conn))
+		rds_iw_attempt_ack(ic);
+
+	return ret;
+}
+
+int __init rds_iw_recv_init(void)
+{
+	struct sysinfo si;
+	int ret = -ENOMEM;
+
+	/* Default to 30% of all available RAM for recv memory */
+	si_meminfo(&si);
+	rds_iw_sysctl_max_recv_allocation = si.totalram / 3 * PAGE_SIZE / RDS_FRAG_SIZE;
+
+	rds_iw_incoming_slab = kmem_cache_create("rds_iw_incoming",
+					sizeof(struct rds_iw_incoming),
+					0, 0, NULL);
+	if (rds_iw_incoming_slab == NULL)
+		goto out;
+
+	rds_iw_frag_slab = kmem_cache_create("rds_iw_frag",
+					sizeof(struct rds_page_frag),
+					0, 0, NULL);
+	if (rds_iw_frag_slab == NULL)
+		kmem_cache_destroy(rds_iw_incoming_slab);
+	else
+		ret = 0;
+out:
+	return ret;
+}
+
+void rds_iw_recv_exit(void)
+{
+	kmem_cache_destroy(rds_iw_incoming_slab);
+	kmem_cache_destroy(rds_iw_frag_slab);
+}
diff --git a/net/rds/iw_ring.c b/net/rds/iw_ring.c
new file mode 100644
index 0000000..d422d4b
--- /dev/null
+++ b/net/rds/iw_ring.c
@@ -0,0 +1,169 @@
+/*
+ * Copyright (c) 2006 Oracle.  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
+ * OpenIB.org BSD 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.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/kernel.h>
+
+#include "rds.h"
+#include "iw.h"
+
+/*
+ * Locking for IB rings.
+ * We assume that allocation is always protected by a mutex
+ * in the caller (this is a valid assumption for the current
+ * implementation).
+ *
+ * Freeing always happens in an interrupt, and hence only
+ * races with allocations, but not with other free()s.
+ *
+ * The interaction between allocation and freeing is that
+ * the alloc code has to determine the number of free entries.
+ * To this end, we maintain two counters; an allocation counter
+ * and a free counter. Both are allowed to run freely, and wrap
+ * around.
+ * The number of used entries is always (alloc_ctr - free_ctr) % NR.
+ *
+ * The current implementation makes free_ctr atomic. When the
+ * caller finds an allocation fails, it should set an "alloc fail"
+ * bit and retry the allocation. The "alloc fail" bit essentially tells
+ * the CQ completion handlers to wake it up after freeing some
+ * more entries.
+ */
+
+/*
+ * This only happens on shutdown.
+ */
+DECLARE_WAIT_QUEUE_HEAD(rds_iw_ring_empty_wait);
+
+void rds_iw_ring_init(struct rds_iw_work_ring *ring, u32 nr)
+{
+	memset(ring, 0, sizeof(*ring));
+	ring->w_nr = nr;
+	rdsdebug("ring %p nr %u\n", ring, ring->w_nr);
+}
+
+static inline u32 __rds_iw_ring_used(struct rds_iw_work_ring *ring)
+{
+	u32 diff;
+
+	/* This assumes that atomic_t has at least as many bits as u32 */
+	diff = ring->w_alloc_ctr - (u32) atomic_read(&ring->w_free_ctr);
+	BUG_ON(diff > ring->w_nr);
+
+	return diff;
+}
+
+void rds_iw_ring_resize(struct rds_iw_work_ring *ring, u32 nr)
+{
+	/* We only ever get called from the connection setup code,
+	 * prior to creating the QP. */
+	BUG_ON(__rds_iw_ring_used(ring));
+	ring->w_nr = nr;
+}
+
+static int __rds_iw_ring_empty(struct rds_iw_work_ring *ring)
+{
+	return __rds_iw_ring_used(ring) == 0;
+}
+
+u32 rds_iw_ring_alloc(struct rds_iw_work_ring *ring, u32 val, u32 *pos)
+{
+	u32 ret = 0, avail;
+
+	avail = ring->w_nr - __rds_iw_ring_used(ring);
+
+	rdsdebug("ring %p val %u next %u free %u\n", ring, val,
+		 ring->w_alloc_ptr, avail);
+
+	if (val && avail) {
+		ret = min(val, avail);
+		*pos = ring->w_alloc_ptr;
+
+		ring->w_alloc_ptr = (ring->w_alloc_ptr + ret) % ring->w_nr;
+		ring->w_alloc_ctr += ret;
+	}
+
+	return ret;
+}
+
+void rds_iw_ring_free(struct rds_iw_work_ring *ring, u32 val)
+{
+	ring->w_free_ptr = (ring->w_free_ptr + val) % ring->w_nr;
+	atomic_add(val, &ring->w_free_ctr);
+
+	if (__rds_iw_ring_empty(ring) &&
+	    waitqueue_active(&rds_iw_ring_empty_wait))
+		wake_up(&rds_iw_ring_empty_wait);
+}
+
+void rds_iw_ring_unalloc(struct rds_iw_work_ring *ring, u32 val)
+{
+	ring->w_alloc_ptr = (ring->w_alloc_ptr - val) % ring->w_nr;
+	ring->w_alloc_ctr -= val;
+}
+
+int rds_iw_ring_empty(struct rds_iw_work_ring *ring)
+{
+	return __rds_iw_ring_empty(ring);
+}
+
+int rds_iw_ring_low(struct rds_iw_work_ring *ring)
+{
+	return __rds_iw_ring_used(ring) <= (ring->w_nr >> 2);
+}
+
+
+/*
+ * returns the oldest alloced ring entry.  This will be the next one
+ * freed.  This can't be called if there are none allocated.
+ */
+u32 rds_iw_ring_oldest(struct rds_iw_work_ring *ring)
+{
+	return ring->w_free_ptr;
+}
+
+/*
+ * returns the number of completed work requests.
+ */
+
+u32 rds_iw_ring_completed(struct rds_iw_work_ring *ring, u32 wr_id, u32 oldest)
+{
+	u32 ret;
+
+	if (oldest <= (unsigned long long)wr_id)
+		ret = (unsigned long long)wr_id - oldest + 1;
+	else
+		ret = ring->w_nr - oldest + (unsigned long long)wr_id + 1;
+
+	rdsdebug("ring %p ret %u wr_id %u oldest %u\n", ring, ret,
+		 wr_id, oldest);
+	return ret;
+}
diff --git a/net/rds/iw_send.c b/net/rds/iw_send.c
new file mode 100644
index 0000000..22dd38f
--- /dev/null
+++ b/net/rds/iw_send.c
@@ -0,0 +1,975 @@
+/*
+ * Copyright (c) 2006 Oracle.  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
+ * OpenIB.org BSD 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.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/kernel.h>
+#include <linux/in.h>
+#include <linux/device.h>
+#include <linux/dmapool.h>
+
+#include "rds.h"
+#include "rdma.h"
+#include "iw.h"
+
+static void rds_iw_send_rdma_complete(struct rds_message *rm,
+				      int wc_status)
+{
+	int notify_status;
+
+	switch (wc_status) {
+	case IB_WC_WR_FLUSH_ERR:
+		return;
+
+	case IB_WC_SUCCESS:
+		notify_status = RDS_RDMA_SUCCESS;
+		break;
+
+	case IB_WC_REM_ACCESS_ERR:
+		notify_status = RDS_RDMA_REMOTE_ERROR;
+		break;
+
+	default:
+		notify_status = RDS_RDMA_OTHER_ERROR;
+		break;
+	}
+	rds_rdma_send_complete(rm, notify_status);
+}
+
+static void rds_iw_send_unmap_rdma(struct rds_iw_connection *ic,
+				   struct rds_rdma_op *op)
+{
+	if (op->r_mapped) {
+		ib_dma_unmap_sg(ic->i_cm_id->device,
+			op->r_sg, op->r_nents,
+			op->r_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
+		op->r_mapped = 0;
+	}
+}
+
+static void rds_iw_send_unmap_rm(struct rds_iw_connection *ic,
+			  struct rds_iw_send_work *send,
+			  int wc_status)
+{
+	struct rds_message *rm = send->s_rm;
+
+	rdsdebug("ic %p send %p rm %p\n", ic, send, rm);
+
+	ib_dma_unmap_sg(ic->i_cm_id->device,
+		     rm->m_sg, rm->m_nents,
+		     DMA_TO_DEVICE);
+
+	if (rm->m_rdma_op != NULL) {
+		rds_iw_send_unmap_rdma(ic, rm->m_rdma_op);
+
+		/* If the user asked for a completion notification on this
+		 * message, we can implement three different semantics:
+		 *  1.	Notify when we received the ACK on the RDS message
+		 *	that was queued with the RDMA. This provides reliable
+		 *	notification of RDMA status at the expense of a one-way
+		 *	packet delay.
+		 *  2.	Notify when the IB stack gives us the completion event for
+		 *	the RDMA operation.
+		 *  3.	Notify when the IB stack gives us the completion event for
+		 *	the accompanying RDS messages.
+		 * Here, we implement approach #3. To implement approach #2,
+		 * call rds_rdma_send_complete from the cq_handler. To implement #1,
+		 * don't call rds_rdma_send_complete at all, and fall back to the notify
+		 * handling in the ACK processing code.
+		 *
+		 * Note: There's no need to explicitly sync any RDMA buffers using
+		 * ib_dma_sync_sg_for_cpu - the completion for the RDMA
+		 * operation itself unmapped the RDMA buffers, which takes care
+		 * of synching.
+		 */
+		rds_iw_send_rdma_complete(rm, wc_status);
+
+		if (rm->m_rdma_op->r_write)
+			rds_stats_add(s_send_rdma_bytes, rm->m_rdma_op->r_bytes);
+		else
+			rds_stats_add(s_recv_rdma_bytes, rm->m_rdma_op->r_bytes);
+	}
+
+	/* If anyone waited for this message to get flushed out, wake
+	 * them up now */
+	rds_message_unmapped(rm);
+
+	rds_message_put(rm);
+	send->s_rm = NULL;
+}
+
+void rds_iw_send_init_ring(struct rds_iw_connection *ic)
+{
+	struct rds_iw_send_work *send;
+	u32 i;
+
+	for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
+		struct ib_sge *sge;
+
+		send->s_rm = NULL;
+		send->s_op = NULL;
+		send->s_mapping = NULL;
+
+		send->s_wr.next = NULL;
+		send->s_wr.wr_id = i;
+		send->s_wr.sg_list = send->s_sge;
+		send->s_wr.num_sge = 1;
+		send->s_wr.opcode = IB_WR_SEND;
+		send->s_wr.send_flags = 0;
+		send->s_wr.ex.imm_data = 0;
+
+		sge = rds_iw_data_sge(ic, send->s_sge);
+		sge->lkey = 0;
+
+		sge = rds_iw_header_sge(ic, send->s_sge);
+		sge->addr = ic->i_send_hdrs_dma + (i * sizeof(struct rds_header));
+		sge->length = sizeof(struct rds_header);
+		sge->lkey = 0;
+
+		send->s_mr = ib_alloc_fast_reg_mr(ic->i_pd, fastreg_message_size);
+		if (IS_ERR(send->s_mr)) {
+			printk(KERN_WARNING "RDS/IW: ib_alloc_fast_reg_mr failed\n");
+			break;
+		}
+
+		send->s_page_list = ib_alloc_fast_reg_page_list(
+			ic->i_cm_id->device, fastreg_message_size);
+		if (IS_ERR(send->s_page_list)) {
+			printk(KERN_WARNING "RDS/IW: ib_alloc_fast_reg_page_list failed\n");
+			break;
+		}
+	}
+}
+
+void rds_iw_send_clear_ring(struct rds_iw_connection *ic)
+{
+	struct rds_iw_send_work *send;
+	u32 i;
+
+	for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
+		BUG_ON(!send->s_mr);
+		ib_dereg_mr(send->s_mr);
+		BUG_ON(!send->s_page_list);
+		ib_free_fast_reg_page_list(send->s_page_list);
+		if (send->s_wr.opcode == 0xdead)
+			continue;
+		if (send->s_rm)
+			rds_iw_send_unmap_rm(ic, send, IB_WC_WR_FLUSH_ERR);
+		if (send->s_op)
+			rds_iw_send_unmap_rdma(ic, send->s_op);
+	}
+}
+
+/*
+ * The _oldest/_free ring operations here race cleanly with the alloc/unalloc
+ * operations performed in the send path.  As the sender allocs and potentially
+ * unallocs the next free entry in the ring it doesn't alter which is
+ * the next to be freed, which is what this is concerned with.
+ */
+void rds_iw_send_cq_comp_handler(struct ib_cq *cq, void *context)
+{
+	struct rds_connection *conn = context;
+	struct rds_iw_connection *ic = conn->c_transport_data;
+	struct ib_wc wc;
+	struct rds_iw_send_work *send;
+	u32 completed;
+	u32 oldest;
+	u32 i;
+	int ret;
+
+	rdsdebug("cq %p conn %p\n", cq, conn);
+	rds_iw_stats_inc(s_iw_tx_cq_call);
+	ret = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
+	if (ret)
+		rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
+
+	while (ib_poll_cq(cq, 1, &wc) > 0) {
+		rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
+			 (unsigned long long)wc.wr_id, wc.status, wc.byte_len,
+			 be32_to_cpu(wc.ex.imm_data));
+		rds_iw_stats_inc(s_iw_tx_cq_event);
+
+		if (wc.status != IB_WC_SUCCESS) {
+			printk(KERN_ERR "WC Error:  status = %d opcode = %d\n", wc.status, wc.opcode);
+			break;
+		}
+
+		if (wc.opcode == IB_WC_LOCAL_INV && wc.wr_id == RDS_IW_LOCAL_INV_WR_ID) {
+			ic->i_fastreg_posted = 0;
+			continue;
+		}
+
+		if (wc.opcode == IB_WC_FAST_REG_MR && wc.wr_id == RDS_IW_FAST_REG_WR_ID) {
+			ic->i_fastreg_posted = 1;
+			continue;
+		}
+
+		if (wc.wr_id == RDS_IW_ACK_WR_ID) {
+			if (ic->i_ack_queued + HZ/2 < jiffies)
+				rds_iw_stats_inc(s_iw_tx_stalled);
+			rds_iw_ack_send_complete(ic);
+			continue;
+		}
+
+		oldest = rds_iw_ring_oldest(&ic->i_send_ring);
+
+		completed = rds_iw_ring_completed(&ic->i_send_ring, wc.wr_id, oldest);
+
+		for (i = 0; i < completed; i++) {
+			send = &ic->i_sends[oldest];
+
+			/* In the error case, wc.opcode sometimes contains garbage */
+			switch (send->s_wr.opcode) {
+			case IB_WR_SEND:
+				if (send->s_rm)
+					rds_iw_send_unmap_rm(ic, send, wc.status);
+				break;
+			case IB_WR_FAST_REG_MR:
+			case IB_WR_RDMA_WRITE:
+			case IB_WR_RDMA_READ:
+			case IB_WR_RDMA_READ_WITH_INV:
+				/* Nothing to be done - the SG list will be unmapped
+				 * when the SEND completes. */
+				break;
+			default:
+				if (printk_ratelimit())
+					printk(KERN_NOTICE
+						"RDS/IW: %s: unexpected opcode 0x%x in WR!\n",
+						__func__, send->s_wr.opcode);
+				break;
+			}
+
+			send->s_wr.opcode = 0xdead;
+			send->s_wr.num_sge = 1;
+			if (send->s_queued + HZ/2 < jiffies)
+				rds_iw_stats_inc(s_iw_tx_stalled);
+
+			/* If a RDMA operation produced an error, signal this right
+			 * away. If we don't, the subsequent SEND that goes with this
+			 * RDMA will be canceled with ERR_WFLUSH, and the application
+			 * never learn that the RDMA failed. */
+			if (unlikely(wc.status == IB_WC_REM_ACCESS_ERR && send->s_op)) {
+				struct rds_message *rm;
+
+				rm = rds_send_get_message(conn, send->s_op);
+				if (rm)
+					rds_iw_send_rdma_complete(rm, wc.status);
+			}
+
+			oldest = (oldest + 1) % ic->i_send_ring.w_nr;
+		}
+
+		rds_iw_ring_free(&ic->i_send_ring, completed);
+
+		if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags)
+		 || test_bit(0, &conn->c_map_queued))
+			queue_delayed_work(rds_wq, &conn->c_send_w, 0);
+
+		/* We expect errors as the qp is drained during shutdown */
+		if (wc.status != IB_WC_SUCCESS && rds_conn_up(conn)) {
+			rds_iw_conn_error(conn,
+				"send completion on %pI4 "
+				"had status %u, disconnecting and reconnecting\n",
+				&conn->c_faddr, wc.status);
+		}
+	}
+}
+
+/*
+ * This is the main function for allocating credits when sending
+ * messages.
+ *
+ * Conceptually, we have two counters:
+ *  -	send credits: this tells us how many WRs we're allowed
+ *	to submit without overruning the reciever's queue. For
+ *	each SEND WR we post, we decrement this by one.
+ *
+ *  -	posted credits: this tells us how many WRs we recently
+ *	posted to the receive queue. This value is transferred
+ *	to the peer as a "credit update" in a RDS header field.
+ *	Every time we transmit credits to the peer, we subtract
+ *	the amount of transferred credits from this counter.
+ *
+ * It is essential that we avoid situations where both sides have
+ * exhausted their send credits, and are unable to send new credits
+ * to the peer. We achieve this by requiring that we send at least
+ * one credit update to the peer before exhausting our credits.
+ * When new credits arrive, we subtract one credit that is withheld
+ * until we've posted new buffers and are ready to transmit these
+ * credits (see rds_iw_send_add_credits below).
+ *
+ * The RDS send code is essentially single-threaded; rds_send_xmit
+ * grabs c_send_lock to ensure exclusive access to the send ring.
+ * However, the ACK sending code is independent and can race with
+ * message SENDs.
+ *
+ * In the send path, we need to update the counters for send credits
+ * and the counter of posted buffers atomically - when we use the
+ * last available credit, we cannot allow another thread to race us
+ * and grab the posted credits counter.  Hence, we have to use a
+ * spinlock to protect the credit counter, or use atomics.
+ *
+ * Spinlocks shared between the send and the receive path are bad,
+ * because they create unnecessary delays. An early implementation
+ * using a spinlock showed a 5% degradation in throughput at some
+ * loads.
+ *
+ * This implementation avoids spinlocks completely, putting both
+ * counters into a single atomic, and updating that atomic using
+ * atomic_add (in the receive path, when receiving fresh credits),
+ * and using atomic_cmpxchg when updating the two counters.
+ */
+int rds_iw_send_grab_credits(struct rds_iw_connection *ic,
+			     u32 wanted, u32 *adv_credits, int need_posted)
+{
+	unsigned int avail, posted, got = 0, advertise;
+	long oldval, newval;
+
+	*adv_credits = 0;
+	if (!ic->i_flowctl)
+		return wanted;
+
+try_again:
+	advertise = 0;
+	oldval = newval = atomic_read(&ic->i_credits);
+	posted = IB_GET_POST_CREDITS(oldval);
+	avail = IB_GET_SEND_CREDITS(oldval);
+
+	rdsdebug("rds_iw_send_grab_credits(%u): credits=%u posted=%u\n",
+			wanted, avail, posted);
+
+	/* The last credit must be used to send a credit update. */
+	if (avail && !posted)
+		avail--;
+
+	if (avail < wanted) {
+		struct rds_connection *conn = ic->i_cm_id->context;
+
+		/* Oops, there aren't that many credits left! */
+		set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
+		got = avail;
+	} else {
+		/* Sometimes you get what you want, lalala. */
+		got = wanted;
+	}
+	newval -= IB_SET_SEND_CREDITS(got);
+
+	/*
+	 * If need_posted is non-zero, then the caller wants
+	 * the posted regardless of whether any send credits are
+	 * available.
+	 */
+	if (posted && (got || need_posted)) {
+		advertise = min_t(unsigned int, posted, RDS_MAX_ADV_CREDIT);
+		newval -= IB_SET_POST_CREDITS(advertise);
+	}
+
+	/* Finally bill everything */
+	if (atomic_cmpxchg(&ic->i_credits, oldval, newval) != oldval)
+		goto try_again;
+
+	*adv_credits = advertise;
+	return got;
+}
+
+void rds_iw_send_add_credits(struct rds_connection *conn, unsigned int credits)
+{
+	struct rds_iw_connection *ic = conn->c_transport_data;
+
+	if (credits == 0)
+		return;
+
+	rdsdebug("rds_iw_send_add_credits(%u): current=%u%s\n",
+			credits,
+			IB_GET_SEND_CREDITS(atomic_read(&ic->i_credits)),
+			test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ? ", ll_send_full" : "");
+
+	atomic_add(IB_SET_SEND_CREDITS(credits), &ic->i_credits);
+	if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags))
+		queue_delayed_work(rds_wq, &conn->c_send_w, 0);
+
+	WARN_ON(IB_GET_SEND_CREDITS(credits) >= 16384);
+
+	rds_iw_stats_inc(s_iw_rx_credit_updates);
+}
+
+void rds_iw_advertise_credits(struct rds_connection *conn, unsigned int posted)
+{
+	struct rds_iw_connection *ic = conn->c_transport_data;
+
+	if (posted == 0)
+		return;
+
+	atomic_add(IB_SET_POST_CREDITS(posted), &ic->i_credits);
+
+	/* Decide whether to send an update to the peer now.
+	 * If we would send a credit update for every single buffer we
+	 * post, we would end up with an ACK storm (ACK arrives,
+	 * consumes buffer, we refill the ring, send ACK to remote
+	 * advertising the newly posted buffer... ad inf)
+	 *
+	 * Performance pretty much depends on how often we send
+	 * credit updates - too frequent updates mean lots of ACKs.
+	 * Too infrequent updates, and the peer will run out of
+	 * credits and has to throttle.
+	 * For the time being, 16 seems to be a good compromise.
+	 */
+	if (IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)) >= 16)
+		set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
+}
+
+static inline void
+rds_iw_xmit_populate_wr(struct rds_iw_connection *ic,
+		struct rds_iw_send_work *send, unsigned int pos,
+		unsigned long buffer, unsigned int length,
+		int send_flags)
+{
+	struct ib_sge *sge;
+
+	WARN_ON(pos != send - ic->i_sends);
+
+	send->s_wr.send_flags = send_flags;
+	send->s_wr.opcode = IB_WR_SEND;
+	send->s_wr.num_sge = 2;
+	send->s_wr.next = NULL;
+	send->s_queued = jiffies;
+	send->s_op = NULL;
+
+	if (length != 0) {
+		sge = rds_iw_data_sge(ic, send->s_sge);
+		sge->addr = buffer;
+		sge->length = length;
+		sge->lkey = rds_iw_local_dma_lkey(ic);
+
+		sge = rds_iw_header_sge(ic, send->s_sge);
+	} else {
+		/* We're sending a packet with no payload. There is only
+		 * one SGE */
+		send->s_wr.num_sge = 1;
+		sge = &send->s_sge[0];
+	}
+
+	sge->addr = ic->i_send_hdrs_dma + (pos * sizeof(struct rds_header));
+	sge->length = sizeof(struct rds_header);
+	sge->lkey = rds_iw_local_dma_lkey(ic);
+}
+
+/*
+ * This can be called multiple times for a given message.  The first time
+ * we see a message we map its scatterlist into the IB device so that
+ * we can provide that mapped address to the IB scatter gather entries
+ * in the IB work requests.  We translate the scatterlist into a series
+ * of work requests that fragment the message.  These work requests complete
+ * in order so we pass ownership of the message to the completion handler
+ * once we send the final fragment.
+ *
+ * The RDS core uses the c_send_lock to only enter this function once
+ * per connection.  This makes sure that the tx ring alloc/unalloc pairs
+ * don't get out of sync and confuse the ring.
+ */
+int rds_iw_xmit(struct rds_connection *conn, struct rds_message *rm,
+		unsigned int hdr_off, unsigned int sg, unsigned int off)
+{
+	struct rds_iw_connection *ic = conn->c_transport_data;
+	struct ib_device *dev = ic->i_cm_id->device;
+	struct rds_iw_send_work *send = NULL;
+	struct rds_iw_send_work *first;
+	struct rds_iw_send_work *prev;
+	struct ib_send_wr *failed_wr;
+	struct scatterlist *scat;
+	u32 pos;
+	u32 i;
+	u32 work_alloc;
+	u32 credit_alloc;
+	u32 posted;
+	u32 adv_credits = 0;
+	int send_flags = 0;
+	int sent;
+	int ret;
+	int flow_controlled = 0;
+
+	BUG_ON(off % RDS_FRAG_SIZE);
+	BUG_ON(hdr_off != 0 && hdr_off != sizeof(struct rds_header));
+
+	/* Fastreg support */
+	if (rds_rdma_cookie_key(rm->m_rdma_cookie)
+	 && !ic->i_fastreg_posted) {
+		ret = -EAGAIN;
+		goto out;
+	}
+
+	/* FIXME we may overallocate here */
+	if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0)
+		i = 1;
+	else
+		i = ceil(be32_to_cpu(rm->m_inc.i_hdr.h_len), RDS_FRAG_SIZE);
+
+	work_alloc = rds_iw_ring_alloc(&ic->i_send_ring, i, &pos);
+	if (work_alloc == 0) {
+		set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
+		rds_iw_stats_inc(s_iw_tx_ring_full);
+		ret = -ENOMEM;
+		goto out;
+	}
+
+	credit_alloc = work_alloc;
+	if (ic->i_flowctl) {
+		credit_alloc = rds_iw_send_grab_credits(ic, work_alloc, &posted, 0);
+		adv_credits += posted;
+		if (credit_alloc < work_alloc) {
+			rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc - credit_alloc);
+			work_alloc = credit_alloc;
+			flow_controlled++;
+		}
+		if (work_alloc == 0) {
+			rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
+			rds_iw_stats_inc(s_iw_tx_throttle);
+			ret = -ENOMEM;
+			goto out;
+		}
+	}
+
+	/* map the message the first time we see it */
+	if (ic->i_rm == NULL) {
+		/*
+		printk(KERN_NOTICE "rds_iw_xmit prep msg dport=%u flags=0x%x len=%d\n",
+				be16_to_cpu(rm->m_inc.i_hdr.h_dport),
+				rm->m_inc.i_hdr.h_flags,
+				be32_to_cpu(rm->m_inc.i_hdr.h_len));
+		   */
+		if (rm->m_nents) {
+			rm->m_count = ib_dma_map_sg(dev,
+					 rm->m_sg, rm->m_nents, DMA_TO_DEVICE);
+			rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->m_count);
+			if (rm->m_count == 0) {
+				rds_iw_stats_inc(s_iw_tx_sg_mapping_failure);
+				rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
+				ret = -ENOMEM; /* XXX ? */
+				goto out;
+			}
+		} else {
+			rm->m_count = 0;
+		}
+
+		ic->i_unsignaled_wrs = rds_iw_sysctl_max_unsig_wrs;
+		ic->i_unsignaled_bytes = rds_iw_sysctl_max_unsig_bytes;
+		rds_message_addref(rm);
+		ic->i_rm = rm;
+
+		/* Finalize the header */
+		if (test_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags))
+			rm->m_inc.i_hdr.h_flags |= RDS_FLAG_ACK_REQUIRED;
+		if (test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))
+			rm->m_inc.i_hdr.h_flags |= RDS_FLAG_RETRANSMITTED;
+
+		/* If it has a RDMA op, tell the peer we did it. This is
+		 * used by the peer to release use-once RDMA MRs. */
+		if (rm->m_rdma_op) {
+			struct rds_ext_header_rdma ext_hdr;
+
+			ext_hdr.h_rdma_rkey = cpu_to_be32(rm->m_rdma_op->r_key);
+			rds_message_add_extension(&rm->m_inc.i_hdr,
+					RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr));
+		}
+		if (rm->m_rdma_cookie) {
+			rds_message_add_rdma_dest_extension(&rm->m_inc.i_hdr,
+					rds_rdma_cookie_key(rm->m_rdma_cookie),
+					rds_rdma_cookie_offset(rm->m_rdma_cookie));
+		}
+
+		/* Note - rds_iw_piggyb_ack clears the ACK_REQUIRED bit, so
+		 * we should not do this unless we have a chance of at least
+		 * sticking the header into the send ring. Which is why we
+		 * should call rds_iw_ring_alloc first. */
+		rm->m_inc.i_hdr.h_ack = cpu_to_be64(rds_iw_piggyb_ack(ic));
+		rds_message_make_checksum(&rm->m_inc.i_hdr);
+
+		/*
+		 * Update adv_credits since we reset the ACK_REQUIRED bit.
+		 */
+		rds_iw_send_grab_credits(ic, 0, &posted, 1);
+		adv_credits += posted;
+		BUG_ON(adv_credits > 255);
+	} else if (ic->i_rm != rm)
+		BUG();
+
+	send = &ic->i_sends[pos];
+	first = send;
+	prev = NULL;
+	scat = &rm->m_sg[sg];
+	sent = 0;
+	i = 0;
+
+	/* Sometimes you want to put a fence between an RDMA
+	 * READ and the following SEND.
+	 * We could either do this all the time
+	 * or when requested by the user. Right now, we let
+	 * the application choose.
+	 */
+	if (rm->m_rdma_op && rm->m_rdma_op->r_fence)
+		send_flags = IB_SEND_FENCE;
+
+	/*
+	 * We could be copying the header into the unused tail of the page.
+	 * That would need to be changed in the future when those pages might
+	 * be mapped userspace pages or page cache pages.  So instead we always
+	 * use a second sge and our long-lived ring of mapped headers.  We send
+	 * the header after the data so that the data payload can be aligned on
+	 * the receiver.
+	 */
+
+	/* handle a 0-len message */
+	if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0) {
+		rds_iw_xmit_populate_wr(ic, send, pos, 0, 0, send_flags);
+		goto add_header;
+	}
+
+	/* if there's data reference it with a chain of work reqs */
+	for (; i < work_alloc && scat != &rm->m_sg[rm->m_count]; i++) {
+		unsigned int len;
+
+		send = &ic->i_sends[pos];
+
+		len = min(RDS_FRAG_SIZE, ib_sg_dma_len(dev, scat) - off);
+		rds_iw_xmit_populate_wr(ic, send, pos,
+				ib_sg_dma_address(dev, scat) + off, len,
+				send_flags);
+
+		/*
+		 * We want to delay signaling completions just enough to get
+		 * the batching benefits but not so much that we create dead time
+		 * on the wire.
+		 */
+		if (ic->i_unsignaled_wrs-- == 0) {
+			ic->i_unsignaled_wrs = rds_iw_sysctl_max_unsig_wrs;
+			send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
+		}
+
+		ic->i_unsignaled_bytes -= len;
+		if (ic->i_unsignaled_bytes <= 0) {
+			ic->i_unsignaled_bytes = rds_iw_sysctl_max_unsig_bytes;
+			send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
+		}
+
+		/*
+		 * Always signal the last one if we're stopping due to flow control.
+		 */
+		if (flow_controlled && i == (work_alloc-1))
+			send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
+
+		rdsdebug("send %p wr %p num_sge %u next %p\n", send,
+			 &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
+
+		sent += len;
+		off += len;
+		if (off == ib_sg_dma_len(dev, scat)) {
+			scat++;
+			off = 0;
+		}
+
+add_header:
+		/* Tack on the header after the data. The header SGE should already
+		 * have been set up to point to the right header buffer. */
+		memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header));
+
+		if (0) {
+			struct rds_header *hdr = &ic->i_send_hdrs[pos];
+
+			printk(KERN_NOTICE "send WR dport=%u flags=0x%x len=%d\n",
+				be16_to_cpu(hdr->h_dport),
+				hdr->h_flags,
+				be32_to_cpu(hdr->h_len));
+		}
+		if (adv_credits) {
+			struct rds_header *hdr = &ic->i_send_hdrs[pos];
+
+			/* add credit and redo the header checksum */
+			hdr->h_credit = adv_credits;
+			rds_message_make_checksum(hdr);
+			adv_credits = 0;
+			rds_iw_stats_inc(s_iw_tx_credit_updates);
+		}
+
+		if (prev)
+			prev->s_wr.next = &send->s_wr;
+		prev = send;
+
+		pos = (pos + 1) % ic->i_send_ring.w_nr;
+	}
+
+	/* Account the RDS header in the number of bytes we sent, but just once.
+	 * The caller has no concept of fragmentation. */
+	if (hdr_off == 0)
+		sent += sizeof(struct rds_header);
+
+	/* if we finished the message then send completion owns it */
+	if (scat == &rm->m_sg[rm->m_count]) {
+		prev->s_rm = ic->i_rm;
+		prev->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
+		ic->i_rm = NULL;
+	}
+
+	if (i < work_alloc) {
+		rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc - i);
+		work_alloc = i;
+	}
+	if (ic->i_flowctl && i < credit_alloc)
+		rds_iw_send_add_credits(conn, credit_alloc - i);
+
+	/* XXX need to worry about failed_wr and partial sends. */
+	failed_wr = &first->s_wr;
+	ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
+	rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
+		 first, &first->s_wr, ret, failed_wr);
+	BUG_ON(failed_wr != &first->s_wr);
+	if (ret) {
+		printk(KERN_WARNING "RDS/IW: ib_post_send to %pI4 "
+		       "returned %d\n", &conn->c_faddr, ret);
+		rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
+		if (prev->s_rm) {
+			ic->i_rm = prev->s_rm;
+			prev->s_rm = NULL;
+		}
+		goto out;
+	}
+
+	ret = sent;
+out:
+	BUG_ON(adv_credits);
+	return ret;
+}
+
+static void rds_iw_build_send_fastreg(struct rds_iw_device *rds_iwdev, struct rds_iw_connection *ic, struct rds_iw_send_work *send, int nent, int len, u64 sg_addr)
+{
+	BUG_ON(nent > send->s_page_list->max_page_list_len);
+	/*
+	 * Perform a WR for the fast_reg_mr. Each individual page
+	 * in the sg list is added to the fast reg page list and placed
+	 * inside the fast_reg_mr WR.
+	 */
+	send->s_wr.opcode = IB_WR_FAST_REG_MR;
+	send->s_wr.wr.fast_reg.length = len;
+	send->s_wr.wr.fast_reg.rkey = send->s_mr->rkey;
+	send->s_wr.wr.fast_reg.page_list = send->s_page_list;
+	send->s_wr.wr.fast_reg.page_list_len = nent;
+	send->s_wr.wr.fast_reg.page_shift = rds_iwdev->page_shift;
+	send->s_wr.wr.fast_reg.access_flags = IB_ACCESS_REMOTE_WRITE;
+	send->s_wr.wr.fast_reg.iova_start = sg_addr;
+
+	ib_update_fast_reg_key(send->s_mr, send->s_remap_count++);
+}
+
+int rds_iw_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
+{
+	struct rds_iw_connection *ic = conn->c_transport_data;
+	struct rds_iw_send_work *send = NULL;
+	struct rds_iw_send_work *first;
+	struct rds_iw_send_work *prev;
+	struct ib_send_wr *failed_wr;
+	struct rds_iw_device *rds_iwdev;
+	struct scatterlist *scat;
+	unsigned long len;
+	u64 remote_addr = op->r_remote_addr;
+	u32 pos, fr_pos;
+	u32 work_alloc;
+	u32 i;
+	u32 j;
+	int sent;
+	int ret;
+	int num_sge;
+
+	rds_iwdev = ib_get_client_data(ic->i_cm_id->device, &rds_iw_client);
+
+	/* map the message the first time we see it */
+	if (!op->r_mapped) {
+		op->r_count = ib_dma_map_sg(ic->i_cm_id->device,
+					op->r_sg, op->r_nents, (op->r_write) ?
+					DMA_TO_DEVICE : DMA_FROM_DEVICE);
+		rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->r_count);
+		if (op->r_count == 0) {
+			rds_iw_stats_inc(s_iw_tx_sg_mapping_failure);
+			ret = -ENOMEM; /* XXX ? */
+			goto out;
+		}
+
+		op->r_mapped = 1;
+	}
+
+	if (!op->r_write) {
+		/* Alloc space on the send queue for the fastreg */
+		work_alloc = rds_iw_ring_alloc(&ic->i_send_ring, 1, &fr_pos);
+		if (work_alloc != 1) {
+			rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
+			rds_iw_stats_inc(s_iw_tx_ring_full);
+			ret = -ENOMEM;
+			goto out;
+		}
+	}
+
+	/*
+	 * Instead of knowing how to return a partial rdma read/write we insist that there
+	 * be enough work requests to send the entire message.
+	 */
+	i = ceil(op->r_count, rds_iwdev->max_sge);
+
+	work_alloc = rds_iw_ring_alloc(&ic->i_send_ring, i, &pos);
+	if (work_alloc != i) {
+		rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
+		rds_iw_stats_inc(s_iw_tx_ring_full);
+		ret = -ENOMEM;
+		goto out;
+	}
+
+	send = &ic->i_sends[pos];
+	if (!op->r_write) {
+		first = prev = &ic->i_sends[fr_pos];
+	} else {
+		first = send;
+		prev = NULL;
+	}
+	scat = &op->r_sg[0];
+	sent = 0;
+	num_sge = op->r_count;
+
+	for (i = 0; i < work_alloc && scat != &op->r_sg[op->r_count]; i++) {
+		send->s_wr.send_flags = 0;
+		send->s_queued = jiffies;
+
+		/*
+		 * We want to delay signaling completions just enough to get
+		 * the batching benefits but not so much that we create dead time on the wire.
+		 */
+		if (ic->i_unsignaled_wrs-- == 0) {
+			ic->i_unsignaled_wrs = rds_iw_sysctl_max_unsig_wrs;
+			send->s_wr.send_flags = IB_SEND_SIGNALED;
+		}
+
+		/* To avoid the need to have the plumbing to invalidate the fastreg_mr used
+		 * for local access after RDS is finished with it, using
+		 * IB_WR_RDMA_READ_WITH_INV will invalidate it after the read has completed.
+		 */
+		if (op->r_write)
+			send->s_wr.opcode = IB_WR_RDMA_WRITE;
+		else
+			send->s_wr.opcode = IB_WR_RDMA_READ_WITH_INV;
+
+		send->s_wr.wr.rdma.remote_addr = remote_addr;
+		send->s_wr.wr.rdma.rkey = op->r_key;
+		send->s_op = op;
+
+		if (num_sge > rds_iwdev->max_sge) {
+			send->s_wr.num_sge = rds_iwdev->max_sge;
+			num_sge -= rds_iwdev->max_sge;
+		} else
+			send->s_wr.num_sge = num_sge;
+
+		send->s_wr.next = NULL;
+
+		if (prev)
+			prev->s_wr.next = &send->s_wr;
+
+		for (j = 0; j < send->s_wr.num_sge && scat != &op->r_sg[op->r_count]; j++) {
+			len = ib_sg_dma_len(ic->i_cm_id->device, scat);
+
+			if (send->s_wr.opcode == IB_WR_RDMA_READ_WITH_INV)
+				send->s_page_list->page_list[j] = ib_sg_dma_address(ic->i_cm_id->device, scat);
+			else {
+				send->s_sge[j].addr = ib_sg_dma_address(ic->i_cm_id->device, scat);
+				send->s_sge[j].length = len;
+				send->s_sge[j].lkey = rds_iw_local_dma_lkey(ic);
+			}
+
+			sent += len;
+			rdsdebug("ic %p sent %d remote_addr %llu\n", ic, sent, remote_addr);
+			remote_addr += len;
+
+			scat++;
+		}
+
+		if (send->s_wr.opcode == IB_WR_RDMA_READ_WITH_INV) {
+			send->s_wr.num_sge = 1;
+			send->s_sge[0].addr = conn->c_xmit_rm->m_rs->rs_user_addr;
+			send->s_sge[0].length = conn->c_xmit_rm->m_rs->rs_user_bytes;
+			send->s_sge[0].lkey = ic->i_sends[fr_pos].s_mr->lkey;
+		}
+
+		rdsdebug("send %p wr %p num_sge %u next %p\n", send,
+			&send->s_wr, send->s_wr.num_sge, send->s_wr.next);
+
+		prev = send;
+		if (++send == &ic->i_sends[ic->i_send_ring.w_nr])
+			send = ic->i_sends;
+	}
+
+	/* if we finished the message then send completion owns it */
+	if (scat == &op->r_sg[op->r_count])
+		first->s_wr.send_flags = IB_SEND_SIGNALED;
+
+	if (i < work_alloc) {
+		rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc - i);
+		work_alloc = i;
+	}
+
+	/* On iWARP, local memory access by a remote system (ie, RDMA Read) is not
+	 * recommended.  Putting the lkey on the wire is a security hole, as it can
+	 * allow for memory access to all of memory on the remote system.  Some
+	 * adapters do not allow using the lkey for this at all.  To bypass this use a
+	 * fastreg_mr (or possibly a dma_mr)
+	 */
+	if (!op->r_write) {
+		rds_iw_build_send_fastreg(rds_iwdev, ic, &ic->i_sends[fr_pos],
+			op->r_count, sent, conn->c_xmit_rm->m_rs->rs_user_addr);
+		work_alloc++;
+	}
+
+	failed_wr = &first->s_wr;
+	ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
+	rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
+		 first, &first->s_wr, ret, failed_wr);
+	BUG_ON(failed_wr != &first->s_wr);
+	if (ret) {
+		printk(KERN_WARNING "RDS/IW: rdma ib_post_send to %pI4 "
+		       "returned %d\n", &conn->c_faddr, ret);
+		rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
+		goto out;
+	}
+
+out:
+	return ret;
+}
+
+void rds_iw_xmit_complete(struct rds_connection *conn)
+{
+	struct rds_iw_connection *ic = conn->c_transport_data;
+
+	/* We may have a pending ACK or window update we were unable
+	 * to send previously (due to flow control). Try again. */
+	rds_iw_attempt_ack(ic);
+}
diff --git a/net/rds/iw_stats.c b/net/rds/iw_stats.c
new file mode 100644
index 0000000..ccc7e8f
--- /dev/null
+++ b/net/rds/iw_stats.c
@@ -0,0 +1,95 @@
+/*
+ * Copyright (c) 2006 Oracle.  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
+ * OpenIB.org BSD 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.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/percpu.h>
+#include <linux/seq_file.h>
+#include <linux/proc_fs.h>
+
+#include "rds.h"
+#include "iw.h"
+
+DEFINE_PER_CPU(struct rds_iw_statistics, rds_iw_stats) ____cacheline_aligned;
+
+static char *rds_iw_stat_names[] = {
+	"iw_connect_raced",
+	"iw_listen_closed_stale",
+	"iw_tx_cq_call",
+	"iw_tx_cq_event",
+	"iw_tx_ring_full",
+	"iw_tx_throttle",
+	"iw_tx_sg_mapping_failure",
+	"iw_tx_stalled",
+	"iw_tx_credit_updates",
+	"iw_rx_cq_call",
+	"iw_rx_cq_event",
+	"iw_rx_ring_empty",
+	"iw_rx_refill_from_cq",
+	"iw_rx_refill_from_thread",
+	"iw_rx_alloc_limit",
+	"iw_rx_credit_updates",
+	"iw_ack_sent",
+	"iw_ack_send_failure",
+	"iw_ack_send_delayed",
+	"iw_ack_send_piggybacked",
+	"iw_ack_received",
+	"iw_rdma_mr_alloc",
+	"iw_rdma_mr_free",
+	"iw_rdma_mr_used",
+	"iw_rdma_mr_pool_flush",
+	"iw_rdma_mr_pool_wait",
+	"iw_rdma_mr_pool_depleted",
+};
+
+unsigned int rds_iw_stats_info_copy(struct rds_info_iterator *iter,
+				    unsigned int avail)
+{
+	struct rds_iw_statistics stats = {0, };
+	uint64_t *src;
+	uint64_t *sum;
+	size_t i;
+	int cpu;
+
+	if (avail < ARRAY_SIZE(rds_iw_stat_names))
+		goto out;
+
+	for_each_online_cpu(cpu) {
+		src = (uint64_t *)&(per_cpu(rds_iw_stats, cpu));
+		sum = (uint64_t *)&stats;
+		for (i = 0; i < sizeof(stats) / sizeof(uint64_t); i++)
+			*(sum++) += *(src++);
+	}
+
+	rds_stats_info_copy(iter, (uint64_t *)&stats, rds_iw_stat_names,
+			    ARRAY_SIZE(rds_iw_stat_names));
+out:
+	return ARRAY_SIZE(rds_iw_stat_names);
+}
diff --git a/net/rds/iw_sysctl.c b/net/rds/iw_sysctl.c
new file mode 100644
index 0000000..9590678
--- /dev/null
+++ b/net/rds/iw_sysctl.c
@@ -0,0 +1,137 @@
+/*
+ * Copyright (c) 2006 Oracle.  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
+ * OpenIB.org BSD 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.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+#include <linux/kernel.h>
+#include <linux/sysctl.h>
+#include <linux/proc_fs.h>
+
+#include "iw.h"
+
+static struct ctl_table_header *rds_iw_sysctl_hdr;
+
+unsigned long rds_iw_sysctl_max_send_wr = RDS_IW_DEFAULT_SEND_WR;
+unsigned long rds_iw_sysctl_max_recv_wr = RDS_IW_DEFAULT_RECV_WR;
+unsigned long rds_iw_sysctl_max_recv_allocation = (128 * 1024 * 1024) / RDS_FRAG_SIZE;
+static unsigned long rds_iw_sysctl_max_wr_min = 1;
+/* hardware will fail CQ creation long before this */
+static unsigned long rds_iw_sysctl_max_wr_max = (u32)~0;
+
+unsigned long rds_iw_sysctl_max_unsig_wrs = 16;
+static unsigned long rds_iw_sysctl_max_unsig_wr_min = 1;
+static unsigned long rds_iw_sysctl_max_unsig_wr_max = 64;
+
+unsigned long rds_iw_sysctl_max_unsig_bytes = (16 << 20);
+static unsigned long rds_iw_sysctl_max_unsig_bytes_min = 1;
+static unsigned long rds_iw_sysctl_max_unsig_bytes_max = ~0UL;
+
+unsigned int rds_iw_sysctl_flow_control = 1;
+
+ctl_table rds_iw_sysctl_table[] = {
+	{
+		.ctl_name       = CTL_UNNUMBERED,
+		.procname       = "max_send_wr",
+		.data		= &rds_iw_sysctl_max_send_wr,
+		.maxlen         = sizeof(unsigned long),
+		.mode           = 0644,
+		.proc_handler   = &proc_doulongvec_minmax,
+		.extra1		= &rds_iw_sysctl_max_wr_min,
+		.extra2		= &rds_iw_sysctl_max_wr_max,
+	},
+	{
+		.ctl_name       = CTL_UNNUMBERED,
+		.procname       = "max_recv_wr",
+		.data		= &rds_iw_sysctl_max_recv_wr,
+		.maxlen         = sizeof(unsigned long),
+		.mode           = 0644,
+		.proc_handler   = &proc_doulongvec_minmax,
+		.extra1		= &rds_iw_sysctl_max_wr_min,
+		.extra2		= &rds_iw_sysctl_max_wr_max,
+	},
+	{
+		.ctl_name       = CTL_UNNUMBERED,
+		.procname       = "max_unsignaled_wr",
+		.data		= &rds_iw_sysctl_max_unsig_wrs,
+		.maxlen         = sizeof(unsigned long),
+		.mode           = 0644,
+		.proc_handler   = &proc_doulongvec_minmax,
+		.extra1		= &rds_iw_sysctl_max_unsig_wr_min,
+		.extra2		= &rds_iw_sysctl_max_unsig_wr_max,
+	},
+	{
+		.ctl_name       = CTL_UNNUMBERED,
+		.procname       = "max_unsignaled_bytes",
+		.data		= &rds_iw_sysctl_max_unsig_bytes,
+		.maxlen         = sizeof(unsigned long),
+		.mode           = 0644,
+		.proc_handler   = &proc_doulongvec_minmax,
+		.extra1		= &rds_iw_sysctl_max_unsig_bytes_min,
+		.extra2		= &rds_iw_sysctl_max_unsig_bytes_max,
+	},
+	{
+		.ctl_name       = CTL_UNNUMBERED,
+		.procname       = "max_recv_allocation",
+		.data		= &rds_iw_sysctl_max_recv_allocation,
+		.maxlen         = sizeof(unsigned long),
+		.mode           = 0644,
+		.proc_handler   = &proc_doulongvec_minmax,
+	},
+	{
+		.ctl_name	= CTL_UNNUMBERED,
+		.procname	= "flow_control",
+		.data		= &rds_iw_sysctl_flow_control,
+		.maxlen		= sizeof(rds_iw_sysctl_flow_control),
+		.mode		= 0644,
+		.proc_handler	= &proc_dointvec,
+	},
+	{ .ctl_name = 0}
+};
+
+static struct ctl_path rds_iw_sysctl_path[] = {
+	{ .procname = "net", .ctl_name = CTL_NET, },
+	{ .procname = "rds", .ctl_name = CTL_UNNUMBERED, },
+	{ .procname = "iw", .ctl_name = CTL_UNNUMBERED, },
+	{ }
+};
+
+void rds_iw_sysctl_exit(void)
+{
+	if (rds_iw_sysctl_hdr)
+		unregister_sysctl_table(rds_iw_sysctl_hdr);
+}
+
+int __init rds_iw_sysctl_init(void)
+{
+	rds_iw_sysctl_hdr = register_sysctl_paths(rds_iw_sysctl_path, rds_iw_sysctl_table);
+	if (rds_iw_sysctl_hdr == NULL)
+		return -ENOMEM;
+	return 0;
+}