RPCRDMA: rpc rdma verbs interface implementation

This implements the interface from rpcrdma to the RDMA verbs interface
supported by Infniband and iWARP.

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

1 files changed, 1621 insertions, 32 deletions

diff --git a/net/sunrpc/xprtrdma/verbs.c b/net/sunrpc/xprtrdma/verbs.c
index 0baf5338198..9ec8ca4f602 100644
--- a/net/sunrpc/xprtrdma/verbs.c
+++ b/net/sunrpc/xprtrdma/verbs.c
@@ -1,37 +1,1626 @@
 /*
- * Placeholders for subsequent patches
+ * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the BSD-type
+ * license below:
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *
+ *      Redistributions of source code must retain the above copyright
+ *      notice, this list of conditions and the following disclaimer.
+ *
+ *      Redistributions in binary form must reproduce the above
+ *      copyright notice, this list of conditions and the following
+ *      disclaimer in the documentation and/or other materials provided
+ *      with the distribution.
+ *
+ *      Neither the name of the Network Appliance, Inc. nor the names of
+ *      its contributors may be used to endorse or promote products
+ *      derived from this software without specific prior written
+ *      permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  */
 
+/*
+ * verbs.c
+ *
+ * Encapsulates the major functions managing:
+ *  o adapters
+ *  o endpoints
+ *  o connections
+ *  o buffer memory
+ */
+
+#include <linux/pci.h>	/* for Tavor hack below */
+
 #include "xprt_rdma.h"
 
-int rpcrdma_ia_open(struct rpcrdma_xprt *a, struct sockaddr *b, int c)
-{ return EINVAL; }
-void rpcrdma_ia_close(struct rpcrdma_ia *a) { }
-int rpcrdma_ep_create(struct rpcrdma_ep *a, struct rpcrdma_ia *b,
-struct rpcrdma_create_data_internal *c) { return EINVAL; }
-int rpcrdma_ep_destroy(struct rpcrdma_ep *a, struct rpcrdma_ia *b)
-{ return EINVAL; }
-int rpcrdma_ep_connect(struct rpcrdma_ep *a, struct rpcrdma_ia *b)
-{ return EINVAL; }
-int rpcrdma_ep_disconnect(struct rpcrdma_ep *a, struct rpcrdma_ia *b)
-{ return EINVAL; }
-int rpcrdma_ep_post(struct rpcrdma_ia *a, struct rpcrdma_ep *b,
-struct rpcrdma_req *c) { return EINVAL; }
-int rpcrdma_ep_post_recv(struct rpcrdma_ia *a, struct rpcrdma_ep *b,
-struct rpcrdma_rep *c) { return EINVAL; }
-int rpcrdma_buffer_create(struct rpcrdma_buffer *a, struct rpcrdma_ep *b,
-struct rpcrdma_ia *c, struct rpcrdma_create_data_internal *d) { return EINVAL; }
-void rpcrdma_buffer_destroy(struct rpcrdma_buffer *a) { }
-struct rpcrdma_req *rpcrdma_buffer_get(struct rpcrdma_buffer *a)
-{ return NULL; }
-void rpcrdma_buffer_put(struct rpcrdma_req *a) { }
-void rpcrdma_recv_buffer_get(struct rpcrdma_req *a) { }
-void rpcrdma_recv_buffer_put(struct rpcrdma_rep *a) { }
-int rpcrdma_register_internal(struct rpcrdma_ia *a, void *b, int c,
-struct ib_mr **d, struct ib_sge *e) { return EINVAL; }
-int rpcrdma_deregister_internal(struct rpcrdma_ia *a, struct ib_mr *b,
-struct ib_sge *c) { return EINVAL; }
-int rpcrdma_register_external(struct rpcrdma_mr_seg *a, int b, int c,
-struct rpcrdma_xprt *d) { return EINVAL; }
-int rpcrdma_deregister_external(struct rpcrdma_mr_seg *a,
-struct rpcrdma_xprt *b, void *c) { return EINVAL; }
+/*
+ * Globals/Macros
+ */
+
+#ifdef RPC_DEBUG
+# define RPCDBG_FACILITY	RPCDBG_TRANS
+#endif
+
+/*
+ * internal functions
+ */
+
+/*
+ * handle replies in tasklet context, using a single, global list
+ * rdma tasklet function -- just turn around and call the func
+ * for all replies on the list
+ */
+
+static DEFINE_SPINLOCK(rpcrdma_tk_lock_g);
+static LIST_HEAD(rpcrdma_tasklets_g);
+
+static void
+rpcrdma_run_tasklet(unsigned long data)
+{
+	struct rpcrdma_rep *rep;
+	void (*func)(struct rpcrdma_rep *);
+	unsigned long flags;
+
+	data = data;
+	spin_lock_irqsave(&rpcrdma_tk_lock_g, flags);
+	while (!list_empty(&rpcrdma_tasklets_g)) {
+		rep = list_entry(rpcrdma_tasklets_g.next,
+				 struct rpcrdma_rep, rr_list);
+		list_del(&rep->rr_list);
+		func = rep->rr_func;
+		rep->rr_func = NULL;
+		spin_unlock_irqrestore(&rpcrdma_tk_lock_g, flags);
+
+		if (func)
+			func(rep);
+		else
+			rpcrdma_recv_buffer_put(rep);
+
+		spin_lock_irqsave(&rpcrdma_tk_lock_g, flags);
+	}
+	spin_unlock_irqrestore(&rpcrdma_tk_lock_g, flags);
+}
+
+static DECLARE_TASKLET(rpcrdma_tasklet_g, rpcrdma_run_tasklet, 0UL);
+
+static inline void
+rpcrdma_schedule_tasklet(struct rpcrdma_rep *rep)
+{
+	unsigned long flags;
+
+	spin_lock_irqsave(&rpcrdma_tk_lock_g, flags);
+	list_add_tail(&rep->rr_list, &rpcrdma_tasklets_g);
+	spin_unlock_irqrestore(&rpcrdma_tk_lock_g, flags);
+	tasklet_schedule(&rpcrdma_tasklet_g);
+}
+
+static void
+rpcrdma_qp_async_error_upcall(struct ib_event *event, void *context)
+{
+	struct rpcrdma_ep *ep = context;
+
+	dprintk("RPC:       %s: QP error %X on device %s ep %p\n",
+		__func__, event->event, event->device->name, context);
+	if (ep->rep_connected == 1) {
+		ep->rep_connected = -EIO;
+		ep->rep_func(ep);
+		wake_up_all(&ep->rep_connect_wait);
+	}
+}
+
+static void
+rpcrdma_cq_async_error_upcall(struct ib_event *event, void *context)
+{
+	struct rpcrdma_ep *ep = context;
+
+	dprintk("RPC:       %s: CQ error %X on device %s ep %p\n",
+		__func__, event->event, event->device->name, context);
+	if (ep->rep_connected == 1) {
+		ep->rep_connected = -EIO;
+		ep->rep_func(ep);
+		wake_up_all(&ep->rep_connect_wait);
+	}
+}
+
+static inline
+void rpcrdma_event_process(struct ib_wc *wc)
+{
+	struct rpcrdma_rep *rep =
+			(struct rpcrdma_rep *)(unsigned long) wc->wr_id;
+
+	dprintk("RPC:       %s: event rep %p status %X opcode %X length %u\n",
+		__func__, rep, wc->status, wc->opcode, wc->byte_len);
+
+	if (!rep) /* send or bind completion that we don't care about */
+		return;
+
+	if (IB_WC_SUCCESS != wc->status) {
+		dprintk("RPC:       %s: %s WC status %X, connection lost\n",
+			__func__, (wc->opcode & IB_WC_RECV) ? "recv" : "send",
+			 wc->status);
+		rep->rr_len = ~0U;
+		rpcrdma_schedule_tasklet(rep);
+		return;
+	}
+
+	switch (wc->opcode) {
+	case IB_WC_RECV:
+		rep->rr_len = wc->byte_len;
+		ib_dma_sync_single_for_cpu(
+			rdmab_to_ia(rep->rr_buffer)->ri_id->device,
+			rep->rr_iov.addr, rep->rr_len, DMA_FROM_DEVICE);
+		/* Keep (only) the most recent credits, after check validity */
+		if (rep->rr_len >= 16) {
+			struct rpcrdma_msg *p =
+					(struct rpcrdma_msg *) rep->rr_base;
+			unsigned int credits = ntohl(p->rm_credit);
+			if (credits == 0) {
+				dprintk("RPC:       %s: server"
+					" dropped credits to 0!\n", __func__);
+				/* don't deadlock */
+				credits = 1;
+			} else if (credits > rep->rr_buffer->rb_max_requests) {
+				dprintk("RPC:       %s: server"
+					" over-crediting: %d (%d)\n",
+					__func__, credits,
+					rep->rr_buffer->rb_max_requests);
+				credits = rep->rr_buffer->rb_max_requests;
+			}
+			atomic_set(&rep->rr_buffer->rb_credits, credits);
+		}
+		/* fall through */
+	case IB_WC_BIND_MW:
+		rpcrdma_schedule_tasklet(rep);
+		break;
+	default:
+		dprintk("RPC:       %s: unexpected WC event %X\n",
+			__func__, wc->opcode);
+		break;
+	}
+}
+
+static inline int
+rpcrdma_cq_poll(struct ib_cq *cq)
+{
+	struct ib_wc wc;
+	int rc;
+
+	for (;;) {
+		rc = ib_poll_cq(cq, 1, &wc);
+		if (rc < 0) {
+			dprintk("RPC:       %s: ib_poll_cq failed %i\n",
+				__func__, rc);
+			return rc;
+		}
+		if (rc == 0)
+			break;
+
+		rpcrdma_event_process(&wc);
+	}
+
+	return 0;
+}
+
+/*
+ * rpcrdma_cq_event_upcall
+ *
+ * This upcall handles recv, send, bind and unbind events.
+ * It is reentrant but processes single events in order to maintain
+ * ordering of receives to keep server credits.
+ *
+ * It is the responsibility of the scheduled tasklet to return
+ * recv buffers to the pool. NOTE: this affects synchronization of
+ * connection shutdown. That is, the structures required for
+ * the completion of the reply handler must remain intact until
+ * all memory has been reclaimed.
+ *
+ * Note that send events are suppressed and do not result in an upcall.
+ */
+static void
+rpcrdma_cq_event_upcall(struct ib_cq *cq, void *context)
+{
+	int rc;
+
+	rc = rpcrdma_cq_poll(cq);
+	if (rc)
+		return;
+
+	rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
+	if (rc) {
+		dprintk("RPC:       %s: ib_req_notify_cq failed %i\n",
+			__func__, rc);
+		return;
+	}
+
+	rpcrdma_cq_poll(cq);
+}
+
+#ifdef RPC_DEBUG
+static const char * const conn[] = {
+	"address resolved",
+	"address error",
+	"route resolved",
+	"route error",
+	"connect request",
+	"connect response",
+	"connect error",
+	"unreachable",
+	"rejected",
+	"established",
+	"disconnected",
+	"device removal"
+};
+#endif
+
+static int
+rpcrdma_conn_upcall(struct rdma_cm_id *id, struct rdma_cm_event *event)
+{
+	struct rpcrdma_xprt *xprt = id->context;
+	struct rpcrdma_ia *ia = &xprt->rx_ia;
+	struct rpcrdma_ep *ep = &xprt->rx_ep;
+	struct sockaddr_in *addr = (struct sockaddr_in *) &ep->rep_remote_addr;
+	struct ib_qp_attr attr;
+	struct ib_qp_init_attr iattr;
+	int connstate = 0;
+
+	switch (event->event) {
+	case RDMA_CM_EVENT_ADDR_RESOLVED:
+	case RDMA_CM_EVENT_ROUTE_RESOLVED:
+		complete(&ia->ri_done);
+		break;
+	case RDMA_CM_EVENT_ADDR_ERROR:
+		ia->ri_async_rc = -EHOSTUNREACH;
+		dprintk("RPC:       %s: CM address resolution error, ep 0x%p\n",
+			__func__, ep);
+		complete(&ia->ri_done);
+		break;
+	case RDMA_CM_EVENT_ROUTE_ERROR:
+		ia->ri_async_rc = -ENETUNREACH;
+		dprintk("RPC:       %s: CM route resolution error, ep 0x%p\n",
+			__func__, ep);
+		complete(&ia->ri_done);
+		break;
+	case RDMA_CM_EVENT_ESTABLISHED:
+		connstate = 1;
+		ib_query_qp(ia->ri_id->qp, &attr,
+			IB_QP_MAX_QP_RD_ATOMIC | IB_QP_MAX_DEST_RD_ATOMIC,
+			&iattr);
+		dprintk("RPC:       %s: %d responder resources"
+			" (%d initiator)\n",
+			__func__, attr.max_dest_rd_atomic, attr.max_rd_atomic);
+		goto connected;
+	case RDMA_CM_EVENT_CONNECT_ERROR:
+		connstate = -ENOTCONN;
+		goto connected;
+	case RDMA_CM_EVENT_UNREACHABLE:
+		connstate = -ENETDOWN;
+		goto connected;
+	case RDMA_CM_EVENT_REJECTED:
+		connstate = -ECONNREFUSED;
+		goto connected;
+	case RDMA_CM_EVENT_DISCONNECTED:
+		connstate = -ECONNABORTED;
+		goto connected;
+	case RDMA_CM_EVENT_DEVICE_REMOVAL:
+		connstate = -ENODEV;
+connected:
+		dprintk("RPC:       %s: %s: %u.%u.%u.%u:%u"
+			" (ep 0x%p event 0x%x)\n",
+			__func__,
+			(event->event <= 11) ? conn[event->event] :
+						"unknown connection error",
+			NIPQUAD(addr->sin_addr.s_addr),
+			ntohs(addr->sin_port),
+			ep, event->event);
+		atomic_set(&rpcx_to_rdmax(ep->rep_xprt)->rx_buf.rb_credits, 1);
+		dprintk("RPC:       %s: %sconnected\n",
+					__func__, connstate > 0 ? "" : "dis");
+		ep->rep_connected = connstate;
+		ep->rep_func(ep);
+		wake_up_all(&ep->rep_connect_wait);
+		break;
+	default:
+		ia->ri_async_rc = -EINVAL;
+		dprintk("RPC:       %s: unexpected CM event %X\n",
+			__func__, event->event);
+		complete(&ia->ri_done);
+		break;
+	}
+
+	return 0;
+}
+
+static struct rdma_cm_id *
+rpcrdma_create_id(struct rpcrdma_xprt *xprt,
+			struct rpcrdma_ia *ia, struct sockaddr *addr)
+{
+	struct rdma_cm_id *id;
+	int rc;
+
+	id = rdma_create_id(rpcrdma_conn_upcall, xprt, RDMA_PS_TCP);
+	if (IS_ERR(id)) {
+		rc = PTR_ERR(id);
+		dprintk("RPC:       %s: rdma_create_id() failed %i\n",
+			__func__, rc);
+		return id;
+	}
+
+	ia->ri_async_rc = 0;
+	rc = rdma_resolve_addr(id, NULL, addr, RDMA_RESOLVE_TIMEOUT);
+	if (rc) {
+		dprintk("RPC:       %s: rdma_resolve_addr() failed %i\n",
+			__func__, rc);
+		goto out;
+	}
+	wait_for_completion(&ia->ri_done);
+	rc = ia->ri_async_rc;
+	if (rc)
+		goto out;
+
+	ia->ri_async_rc = 0;
+	rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
+	if (rc) {
+		dprintk("RPC:       %s: rdma_resolve_route() failed %i\n",
+			__func__, rc);
+		goto out;
+	}
+	wait_for_completion(&ia->ri_done);
+	rc = ia->ri_async_rc;
+	if (rc)
+		goto out;
+
+	return id;
+
+out:
+	rdma_destroy_id(id);
+	return ERR_PTR(rc);
+}
+
+/*
+ * Drain any cq, prior to teardown.
+ */
+static void
+rpcrdma_clean_cq(struct ib_cq *cq)
+{
+	struct ib_wc wc;
+	int count = 0;
+
+	while (1 == ib_poll_cq(cq, 1, &wc))
+		++count;
+
+	if (count)
+		dprintk("RPC:       %s: flushed %d events (last 0x%x)\n",
+			__func__, count, wc.opcode);
+}
+
+/*
+ * Exported functions.
+ */
+
+/*
+ * Open and initialize an Interface Adapter.
+ *  o initializes fields of struct rpcrdma_ia, including
+ *    interface and provider attributes and protection zone.
+ */
+int
+rpcrdma_ia_open(struct rpcrdma_xprt *xprt, struct sockaddr *addr, int memreg)
+{
+	int rc;
+	struct rpcrdma_ia *ia = &xprt->rx_ia;
+
+	init_completion(&ia->ri_done);
+
+	ia->ri_id = rpcrdma_create_id(xprt, ia, addr);
+	if (IS_ERR(ia->ri_id)) {
+		rc = PTR_ERR(ia->ri_id);
+		goto out1;
+	}
+
+	ia->ri_pd = ib_alloc_pd(ia->ri_id->device);
+	if (IS_ERR(ia->ri_pd)) {
+		rc = PTR_ERR(ia->ri_pd);
+		dprintk("RPC:       %s: ib_alloc_pd() failed %i\n",
+			__func__, rc);
+		goto out2;
+	}
+
+	/*
+	 * Optionally obtain an underlying physical identity mapping in
+	 * order to do a memory window-based bind. This base registration
+	 * is protected from remote access - that is enabled only by binding
+	 * for the specific bytes targeted during each RPC operation, and
+	 * revoked after the corresponding completion similar to a storage
+	 * adapter.
+	 */
+	if (memreg > RPCRDMA_REGISTER) {
+		int mem_priv = IB_ACCESS_LOCAL_WRITE;
+		switch (memreg) {
+#if RPCRDMA_PERSISTENT_REGISTRATION
+		case RPCRDMA_ALLPHYSICAL:
+			mem_priv |= IB_ACCESS_REMOTE_WRITE;
+			mem_priv |= IB_ACCESS_REMOTE_READ;
+			break;
+#endif
+		case RPCRDMA_MEMWINDOWS_ASYNC:
+		case RPCRDMA_MEMWINDOWS:
+			mem_priv |= IB_ACCESS_MW_BIND;
+			break;
+		default:
+			break;
+		}
+		ia->ri_bind_mem = ib_get_dma_mr(ia->ri_pd, mem_priv);
+		if (IS_ERR(ia->ri_bind_mem)) {
+			printk(KERN_ALERT "%s: ib_get_dma_mr for "
+				"phys register failed with %lX\n\t"
+				"Will continue with degraded performance\n",
+				__func__, PTR_ERR(ia->ri_bind_mem));
+			memreg = RPCRDMA_REGISTER;
+			ia->ri_bind_mem = NULL;
+		}
+	}
+
+	/* Else will do memory reg/dereg for each chunk */
+	ia->ri_memreg_strategy = memreg;
+
+	return 0;
+out2:
+	rdma_destroy_id(ia->ri_id);
+out1:
+	return rc;
+}
+
+/*
+ * Clean up/close an IA.
+ *   o if event handles and PD have been initialized, free them.
+ *   o close the IA
+ */
+void
+rpcrdma_ia_close(struct rpcrdma_ia *ia)
+{
+	int rc;
+
+	dprintk("RPC:       %s: entering\n", __func__);
+	if (ia->ri_bind_mem != NULL) {
+		rc = ib_dereg_mr(ia->ri_bind_mem);
+		dprintk("RPC:       %s: ib_dereg_mr returned %i\n",
+			__func__, rc);
+	}
+	if (ia->ri_id != NULL && !IS_ERR(ia->ri_id) && ia->ri_id->qp)
+		rdma_destroy_qp(ia->ri_id);
+	if (ia->ri_pd != NULL && !IS_ERR(ia->ri_pd)) {
+		rc = ib_dealloc_pd(ia->ri_pd);
+		dprintk("RPC:       %s: ib_dealloc_pd returned %i\n",
+			__func__, rc);
+	}
+	if (ia->ri_id != NULL && !IS_ERR(ia->ri_id))
+		rdma_destroy_id(ia->ri_id);
+}
+
+/*
+ * Create unconnected endpoint.
+ */
+int
+rpcrdma_ep_create(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia,
+				struct rpcrdma_create_data_internal *cdata)
+{
+	struct ib_device_attr devattr;
+	int rc;
+
+	rc = ib_query_device(ia->ri_id->device, &devattr);
+	if (rc) {
+		dprintk("RPC:       %s: ib_query_device failed %d\n",
+			__func__, rc);
+		return rc;
+	}
+
+	/* check provider's send/recv wr limits */
+	if (cdata->max_requests > devattr.max_qp_wr)
+		cdata->max_requests = devattr.max_qp_wr;
+
+	ep->rep_attr.event_handler = rpcrdma_qp_async_error_upcall;
+	ep->rep_attr.qp_context = ep;
+	/* send_cq and recv_cq initialized below */
+	ep->rep_attr.srq = NULL;
+	ep->rep_attr.cap.max_send_wr = cdata->max_requests;
+	switch (ia->ri_memreg_strategy) {
+	case RPCRDMA_MEMWINDOWS_ASYNC:
+	case RPCRDMA_MEMWINDOWS:
+		/* Add room for mw_binds+unbinds - overkill! */
+		ep->rep_attr.cap.max_send_wr++;
+		ep->rep_attr.cap.max_send_wr *= (2 * RPCRDMA_MAX_SEGS);
+		if (ep->rep_attr.cap.max_send_wr > devattr.max_qp_wr)
+			return -EINVAL;
+		break;
+	default:
+		break;
+	}
+	ep->rep_attr.cap.max_recv_wr = cdata->max_requests;
+	ep->rep_attr.cap.max_send_sge = (cdata->padding ? 4 : 2);
+	ep->rep_attr.cap.max_recv_sge = 1;
+	ep->rep_attr.cap.max_inline_data = 0;
+	ep->rep_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
+	ep->rep_attr.qp_type = IB_QPT_RC;
+	ep->rep_attr.port_num = ~0;
+
+	dprintk("RPC:       %s: requested max: dtos: send %d recv %d; "
+		"iovs: send %d recv %d\n",
+		__func__,
+		ep->rep_attr.cap.max_send_wr,
+		ep->rep_attr.cap.max_recv_wr,
+		ep->rep_attr.cap.max_send_sge,
+		ep->rep_attr.cap.max_recv_sge);
+
+	/* set trigger for requesting send completion */
+	ep->rep_cqinit = ep->rep_attr.cap.max_send_wr/2 /*  - 1*/;
+	switch (ia->ri_memreg_strategy) {
+	case RPCRDMA_MEMWINDOWS_ASYNC:
+	case RPCRDMA_MEMWINDOWS:
+		ep->rep_cqinit -= RPCRDMA_MAX_SEGS;
+		break;
+	default:
+		break;
+	}
+	if (ep->rep_cqinit <= 2)
+		ep->rep_cqinit = 0;
+	INIT_CQCOUNT(ep);
+	ep->rep_ia = ia;
+	init_waitqueue_head(&ep->rep_connect_wait);
+
+	/*
+	 * Create a single cq for receive dto and mw_bind (only ever
+	 * care about unbind, really). Send completions are suppressed.
+	 * Use single threaded tasklet upcalls to maintain ordering.
+	 */
+	ep->rep_cq = ib_create_cq(ia->ri_id->device, rpcrdma_cq_event_upcall,
+				  rpcrdma_cq_async_error_upcall, NULL,
+				  ep->rep_attr.cap.max_recv_wr +
+				  ep->rep_attr.cap.max_send_wr + 1, 0);
+	if (IS_ERR(ep->rep_cq)) {
+		rc = PTR_ERR(ep->rep_cq);
+		dprintk("RPC:       %s: ib_create_cq failed: %i\n",
+			__func__, rc);
+		goto out1;
+	}
+
+	rc = ib_req_notify_cq(ep->rep_cq, IB_CQ_NEXT_COMP);
+	if (rc) {
+		dprintk("RPC:       %s: ib_req_notify_cq failed: %i\n",
+			__func__, rc);
+		goto out2;
+	}
+
+	ep->rep_attr.send_cq = ep->rep_cq;
+	ep->rep_attr.recv_cq = ep->rep_cq;
+
+	/* Initialize cma parameters */
+
+	/* RPC/RDMA does not use private data */
+	ep->rep_remote_cma.private_data = NULL;
+	ep->rep_remote_cma.private_data_len = 0;
+
+	/* Client offers RDMA Read but does not initiate */
+	switch (ia->ri_memreg_strategy) {
+	case RPCRDMA_BOUNCEBUFFERS:
+		ep->rep_remote_cma.responder_resources = 0;
+		break;
+	case RPCRDMA_MTHCAFMR:
+	case RPCRDMA_REGISTER:
+		ep->rep_remote_cma.responder_resources = cdata->max_requests *
+				(RPCRDMA_MAX_DATA_SEGS / 8);
+		break;
+	case RPCRDMA_MEMWINDOWS:
+	case RPCRDMA_MEMWINDOWS_ASYNC:
+#if RPCRDMA_PERSISTENT_REGISTRATION
+	case RPCRDMA_ALLPHYSICAL:
+#endif
+		ep->rep_remote_cma.responder_resources = cdata->max_requests *
+				(RPCRDMA_MAX_DATA_SEGS / 2);
+		break;
+	default:
+		break;
+	}
+	if (ep->rep_remote_cma.responder_resources > devattr.max_qp_rd_atom)
+		ep->rep_remote_cma.responder_resources = devattr.max_qp_rd_atom;
+	ep->rep_remote_cma.initiator_depth = 0;
+
+	ep->rep_remote_cma.retry_count = 7;
+	ep->rep_remote_cma.flow_control = 0;
+	ep->rep_remote_cma.rnr_retry_count = 0;
+
+	return 0;
+
+out2:
+	if (ib_destroy_cq(ep->rep_cq))
+		;
+out1:
+	return rc;
+}
+
+/*
+ * rpcrdma_ep_destroy
+ *
+ * Disconnect and destroy endpoint. After this, the only
+ * valid operations on the ep are to free it (if dynamically
+ * allocated) or re-create it.
+ *
+ * The caller's error handling must be sure to not leak the endpoint
+ * if this function fails.
+ */
+int
+rpcrdma_ep_destroy(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
+{
+	int rc;
+
+	dprintk("RPC:       %s: entering, connected is %d\n",
+		__func__, ep->rep_connected);
+
+	if (ia->ri_id->qp) {
+		rc = rpcrdma_ep_disconnect(ep, ia);
+		if (rc)
+			dprintk("RPC:       %s: rpcrdma_ep_disconnect"
+				" returned %i\n", __func__, rc);
+	}
+
+	ep->rep_func = NULL;
+
+	/* padding - could be done in rpcrdma_buffer_destroy... */
+	if (ep->rep_pad_mr) {
+		rpcrdma_deregister_internal(ia, ep->rep_pad_mr, &ep->rep_pad);
+		ep->rep_pad_mr = NULL;
+	}
+
+	if (ia->ri_id->qp) {
+		rdma_destroy_qp(ia->ri_id);
+		ia->ri_id->qp = NULL;
+	}
+
+	rpcrdma_clean_cq(ep->rep_cq);
+	rc = ib_destroy_cq(ep->rep_cq);
+	if (rc)
+		dprintk("RPC:       %s: ib_destroy_cq returned %i\n",
+			__func__, rc);
+
+	return rc;
+}
+
+/*
+ * Connect unconnected endpoint.
+ */
+int
+rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
+{
+	struct rdma_cm_id *id;
+	int rc = 0;
+	int retry_count = 0;
+	int reconnect = (ep->rep_connected != 0);
+
+	if (reconnect) {
+		struct rpcrdma_xprt *xprt;
+retry:
+		rc = rpcrdma_ep_disconnect(ep, ia);
+		if (rc && rc != -ENOTCONN)
+			dprintk("RPC:       %s: rpcrdma_ep_disconnect"
+				" status %i\n", __func__, rc);
+		rpcrdma_clean_cq(ep->rep_cq);
+
+		xprt = container_of(ia, struct rpcrdma_xprt, rx_ia);
+		id = rpcrdma_create_id(xprt, ia,
+				(struct sockaddr *)&xprt->rx_data.addr);
+		if (IS_ERR(id)) {
+			rc = PTR_ERR(id);
+			goto out;
+		}
+		/* TEMP TEMP TEMP - fail if new device:
+		 * Deregister/remarshal *all* requests!
+		 * Close and recreate adapter, pd, etc!
+		 * Re-determine all attributes still sane!
+		 * More stuff I haven't thought of!
+		 * Rrrgh!
+		 */
+		if (ia->ri_id->device != id->device) {
+			printk("RPC:       %s: can't reconnect on "
+				"different device!\n", __func__);
+			rdma_destroy_id(id);
+			rc = -ENETDOWN;
+			goto out;
+		}
+		/* END TEMP */
+		rdma_destroy_id(ia->ri_id);
+		ia->ri_id = id;
+	}
+
+	rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
+	if (rc) {
+		dprintk("RPC:       %s: rdma_create_qp failed %i\n",
+			__func__, rc);
+		goto out;
+	}
+
+/* XXX Tavor device performs badly with 2K MTU! */
+if (strnicmp(ia->ri_id->device->dma_device->bus->name, "pci", 3) == 0) {
+	struct pci_dev *pcid = to_pci_dev(ia->ri_id->device->dma_device);
+	if (pcid->device == PCI_DEVICE_ID_MELLANOX_TAVOR &&
+	    (pcid->vendor == PCI_VENDOR_ID_MELLANOX ||
+	     pcid->vendor == PCI_VENDOR_ID_TOPSPIN)) {
+		struct ib_qp_attr attr = {
+			.path_mtu = IB_MTU_1024
+		};
+		rc = ib_modify_qp(ia->ri_id->qp, &attr, IB_QP_PATH_MTU);
+	}
+}
+
+	/* Theoretically a client initiator_depth > 0 is not needed,
+	 * but many peers fail to complete the connection unless they
+	 * == responder_resources! */
+	if (ep->rep_remote_cma.initiator_depth !=
+				ep->rep_remote_cma.responder_resources)
+		ep->rep_remote_cma.initiator_depth =
+			ep->rep_remote_cma.responder_resources;
+
+	ep->rep_connected = 0;
+
+	rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma);
+	if (rc) {
+		dprintk("RPC:       %s: rdma_connect() failed with %i\n",
+				__func__, rc);
+		goto out;
+	}
+
+	if (reconnect)
+		return 0;
+
+	wait_event_interruptible(ep->rep_connect_wait, ep->rep_connected != 0);
+
+	/*
+	 * Check state. A non-peer reject indicates no listener
+	 * (ECONNREFUSED), which may be a transient state. All
+	 * others indicate a transport condition which has already
+	 * undergone a best-effort.
+	 */
+	if (ep->rep_connected == -ECONNREFUSED
+	    && ++retry_count <= RDMA_CONNECT_RETRY_MAX) {
+		dprintk("RPC:       %s: non-peer_reject, retry\n", __func__);
+		goto retry;
+	}
+	if (ep->rep_connected <= 0) {
+		/* Sometimes, the only way to reliably connect to remote
+		 * CMs is to use same nonzero values for ORD and IRD. */
+		ep->rep_remote_cma.initiator_depth =
+					ep->rep_remote_cma.responder_resources;
+		if (ep->rep_remote_cma.initiator_depth == 0)
+			++ep->rep_remote_cma.initiator_depth;
+		if (ep->rep_remote_cma.responder_resources == 0)
+			++ep->rep_remote_cma.responder_resources;
+		if (retry_count++ == 0)
+			goto retry;
+		rc = ep->rep_connected;
+	} else {
+		dprintk("RPC:       %s: connected\n", __func__);
+	}
+
+out:
+	if (rc)
+		ep->rep_connected = rc;
+	return rc;
+}
+
+/*
+ * rpcrdma_ep_disconnect
+ *
+ * This is separate from destroy to facilitate the ability
+ * to reconnect without recreating the endpoint.
+ *
+ * This call is not reentrant, and must not be made in parallel
+ * on the same endpoint.
+ */
+int
+rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
+{
+	int rc;
+
+	rpcrdma_clean_cq(ep->rep_cq);
+	rc = rdma_disconnect(ia->ri_id);
+	if (!rc) {
+		/* returns without wait if not connected */
+		wait_event_interruptible(ep->rep_connect_wait,
+							ep->rep_connected != 1);
+		dprintk("RPC:       %s: after wait, %sconnected\n", __func__,
+			(ep->rep_connected == 1) ? "still " : "dis");
+	} else {
+		dprintk("RPC:       %s: rdma_disconnect %i\n", __func__, rc);
+		ep->rep_connected = rc;
+	}
+	return rc;
+}
+
+/*
+ * Initialize buffer memory
+ */
+int
+rpcrdma_buffer_create(struct rpcrdma_buffer *buf, struct rpcrdma_ep *ep,
+	struct rpcrdma_ia *ia, struct rpcrdma_create_data_internal *cdata)
+{
+	char *p;
+	size_t len;
+	int i, rc;
+
+	buf->rb_max_requests = cdata->max_requests;
+	spin_lock_init(&buf->rb_lock);
+	atomic_set(&buf->rb_credits, 1);
+
+	/* Need to allocate:
+	 *   1.  arrays for send and recv pointers
+	 *   2.  arrays of struct rpcrdma_req to fill in pointers
+	 *   3.  array of struct rpcrdma_rep for replies
+	 *   4.  padding, if any
+	 *   5.  mw's, if any
+	 * Send/recv buffers in req/rep need to be registered
+	 */
+
+	len = buf->rb_max_requests *
+		(sizeof(struct rpcrdma_req *) + sizeof(struct rpcrdma_rep *));
+	len += cdata->padding;
+	switch (ia->ri_memreg_strategy) {
+	case RPCRDMA_MTHCAFMR:
+		/* TBD we are perhaps overallocating here */
+		len += (buf->rb_max_requests + 1) * RPCRDMA_MAX_SEGS *
+				sizeof(struct rpcrdma_mw);
+		break;
+	case RPCRDMA_MEMWINDOWS_ASYNC:
+	case RPCRDMA_MEMWINDOWS:
+		len += (buf->rb_max_requests + 1) * RPCRDMA_MAX_SEGS *
+				sizeof(struct rpcrdma_mw);
+		break;
+	default:
+		break;
+	}
+
+	/* allocate 1, 4 and 5 in one shot */
+	p = kzalloc(len, GFP_KERNEL);
+	if (p == NULL) {
+		dprintk("RPC:       %s: req_t/rep_t/pad kzalloc(%zd) failed\n",
+			__func__, len);
+		rc = -ENOMEM;
+		goto out;
+	}
+	buf->rb_pool = p;	/* for freeing it later */
+
+	buf->rb_send_bufs = (struct rpcrdma_req **) p;
+	p = (char *) &buf->rb_send_bufs[buf->rb_max_requests];
+	buf->rb_recv_bufs = (struct rpcrdma_rep **) p;
+	p = (char *) &buf->rb_recv_bufs[buf->rb_max_requests];
+
+	/*
+	 * Register the zeroed pad buffer, if any.
+	 */
+	if (cdata->padding) {
+		rc = rpcrdma_register_internal(ia, p, cdata->padding,
+					    &ep->rep_pad_mr, &ep->rep_pad);
+		if (rc)
+			goto out;
+	}
+	p += cdata->padding;
+
+	/*
+	 * Allocate the fmr's, or mw's for mw_bind chunk registration.
+	 * We "cycle" the mw's in order to minimize rkey reuse,
+	 * and also reduce unbind-to-bind collision.
+	 */
+	INIT_LIST_HEAD(&buf->rb_mws);
+	switch (ia->ri_memreg_strategy) {
+	case RPCRDMA_MTHCAFMR:
+		{
+		struct rpcrdma_mw *r = (struct rpcrdma_mw *)p;
+		struct ib_fmr_attr fa = {
+			RPCRDMA_MAX_DATA_SEGS, 1, PAGE_SHIFT
+		};
+		/* TBD we are perhaps overallocating here */
+		for (i = (buf->rb_max_requests+1) * RPCRDMA_MAX_SEGS; i; i--) {
+			r->r.fmr = ib_alloc_fmr(ia->ri_pd,
+				IB_ACCESS_REMOTE_WRITE | IB_ACCESS_REMOTE_READ,
+				&fa);
+			if (IS_ERR(r->r.fmr)) {
+				rc = PTR_ERR(r->r.fmr);
+				dprintk("RPC:       %s: ib_alloc_fmr"
+					" failed %i\n", __func__, rc);
+				goto out;
+			}
+			list_add(&r->mw_list, &buf->rb_mws);
+			++r;
+		}
+		}
+		break;
+	case RPCRDMA_MEMWINDOWS_ASYNC:
+	case RPCRDMA_MEMWINDOWS:
+		{
+		struct rpcrdma_mw *r = (struct rpcrdma_mw *)p;
+		/* Allocate one extra request's worth, for full cycling */
+		for (i = (buf->rb_max_requests+1) * RPCRDMA_MAX_SEGS; i; i--) {
+			r->r.mw = ib_alloc_mw(ia->ri_pd);
+			if (IS_ERR(r->r.mw)) {
+				rc = PTR_ERR(r->r.mw);
+				dprintk("RPC:       %s: ib_alloc_mw"
+					" failed %i\n", __func__, rc);
+				goto out;
+			}
+			list_add(&r->mw_list, &buf->rb_mws);
+			++r;
+		}
+		}
+		break;
+	default:
+		break;
+	}
+
+	/*
+	 * Allocate/init the request/reply buffers. Doing this
+	 * using kmalloc for now -- one for each buf.
+	 */
+	for (i = 0; i < buf->rb_max_requests; i++) {
+		struct rpcrdma_req *req;
+		struct rpcrdma_rep *rep;
+
+		len = cdata->inline_wsize + sizeof(struct rpcrdma_req);
+		/* RPC layer requests *double* size + 1K RPC_SLACK_SPACE! */
+		/* Typical ~2400b, so rounding up saves work later */
+		if (len < 4096)
+			len = 4096;
+		req = kmalloc(len, GFP_KERNEL);
+		if (req == NULL) {
+			dprintk("RPC:       %s: request buffer %d alloc"
+				" failed\n", __func__, i);
+			rc = -ENOMEM;
+			goto out;
+		}
+		memset(req, 0, sizeof(struct rpcrdma_req));
+		buf->rb_send_bufs[i] = req;
+		buf->rb_send_bufs[i]->rl_buffer = buf;
+
+		rc = rpcrdma_register_internal(ia, req->rl_base,
+				len - offsetof(struct rpcrdma_req, rl_base),
+				&buf->rb_send_bufs[i]->rl_handle,
+				&buf->rb_send_bufs[i]->rl_iov);
+		if (rc)
+			goto out;
+
+		buf->rb_send_bufs[i]->rl_size = len-sizeof(struct rpcrdma_req);
+
+		len = cdata->inline_rsize + sizeof(struct rpcrdma_rep);
+		rep = kmalloc(len, GFP_KERNEL);
+		if (rep == NULL) {
+			dprintk("RPC:       %s: reply buffer %d alloc failed\n",
+				__func__, i);
+			rc = -ENOMEM;
+			goto out;
+		}
+		memset(rep, 0, sizeof(struct rpcrdma_rep));
+		buf->rb_recv_bufs[i] = rep;
+		buf->rb_recv_bufs[i]->rr_buffer = buf;
+		init_waitqueue_head(&rep->rr_unbind);
+
+		rc = rpcrdma_register_internal(ia, rep->rr_base,
+				len - offsetof(struct rpcrdma_rep, rr_base),
+				&buf->rb_recv_bufs[i]->rr_handle,
+				&buf->rb_recv_bufs[i]->rr_iov);
+		if (rc)
+			goto out;
+
+	}
+	dprintk("RPC:       %s: max_requests %d\n",
+		__func__, buf->rb_max_requests);
+	/* done */
+	return 0;
+out:
+	rpcrdma_buffer_destroy(buf);
+	return rc;
+}
+
+/*
+ * Unregister and destroy buffer memory. Need to deal with
+ * partial initialization, so it's callable from failed create.
+ * Must be called before destroying endpoint, as registrations
+ * reference it.
+ */
+void
+rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
+{
+	int rc, i;
+	struct rpcrdma_ia *ia = rdmab_to_ia(buf);
+
+	/* clean up in reverse order from create
+	 *   1.  recv mr memory (mr free, then kfree)
+	 *   1a. bind mw memory
+	 *   2.  send mr memory (mr free, then kfree)
+	 *   3.  padding (if any) [moved to rpcrdma_ep_destroy]
+	 *   4.  arrays
+	 */
+	dprintk("RPC:       %s: entering\n", __func__);
+
+	for (i = 0; i < buf->rb_max_requests; i++) {
+		if (buf->rb_recv_bufs && buf->rb_recv_bufs[i]) {
+			rpcrdma_deregister_internal(ia,
+					buf->rb_recv_bufs[i]->rr_handle,
+					&buf->rb_recv_bufs[i]->rr_iov);
+			kfree(buf->rb_recv_bufs[i]);
+		}
+		if (buf->rb_send_bufs && buf->rb_send_bufs[i]) {
+			while (!list_empty(&buf->rb_mws)) {
+				struct rpcrdma_mw *r;
+				r = list_entry(buf->rb_mws.next,
+					struct rpcrdma_mw, mw_list);
+				list_del(&r->mw_list);
+				switch (ia->ri_memreg_strategy) {
+				case RPCRDMA_MTHCAFMR:
+					rc = ib_dealloc_fmr(r->r.fmr);
+					if (rc)
+						dprintk("RPC:       %s:"
+							" ib_dealloc_fmr"
+							" failed %i\n",
+							__func__, rc);
+					break;
+				case RPCRDMA_MEMWINDOWS_ASYNC:
+				case RPCRDMA_MEMWINDOWS:
+					rc = ib_dealloc_mw(r->r.mw);
+					if (rc)
+						dprintk("RPC:       %s:"
+							" ib_dealloc_mw"
+							" failed %i\n",
+							__func__, rc);
+					break;
+				default:
+					break;
+				}
+			}
+			rpcrdma_deregister_internal(ia,
+					buf->rb_send_bufs[i]->rl_handle,
+					&buf->rb_send_bufs[i]->rl_iov);
+			kfree(buf->rb_send_bufs[i]);
+		}
+	}
+
+	kfree(buf->rb_pool);
+}
+
+/*
+ * Get a set of request/reply buffers.
+ *
+ * Reply buffer (if needed) is attached to send buffer upon return.
+ * Rule:
+ *    rb_send_index and rb_recv_index MUST always be pointing to the
+ *    *next* available buff