verbs.c

/*
 * 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 <linux/slab.h>

#include "xprt_rdma.h"

/*
 * 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_mw *frmr;
	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: WC opcode %d status %X, connection lost\n",
			__func__, wc->opcode, wc->status);
		rep->rr_len = ~0U;
		if (wc->opcode != IB_WC_FAST_REG_MR && wc->opcode != IB_WC_LOCAL_INV)
			rpcrdma_schedule_tasklet(rep);
		return;
	}

	switch (wc->opcode) {
	case IB_WC_FAST_REG_MR:
		frmr = (struct rpcrdma_mw *)(unsigned long)wc->wr_id;
		frmr->r.frmr.state = FRMR_IS_VALID;
		break;
	case IB_WC_LOCAL_INV:
		frmr = (struct rpcrdma_mw *)(unsigned long)wc->wr_id;
		frmr->r.frmr.state = FRMR_IS_INVALID;
		break;
	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;
#ifdef RPC_DEBUG
	struct sockaddr_in *addr = (struct sockaddr_in *) &ep->rep_remote_addr;
#endif
	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:
		ia->ri_async_rc = 0;
		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: %pI4:%u (ep 0x%p event 0x%x)\n",
			__func__,
			(event->event <= 11) ? conn[event->event] :
						"unknown connection error",
			&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:
		dprintk("RPC:       %s: unexpected CM event %d\n",
			__func__, event->event);
		break;
	}

#ifdef RPC_DEBUG
	if (connstate == 1) {
		int ird = attr.max_dest_rd_atomic;
		int tird = ep->rep_remote_cma.responder_resources;
		printk(KERN_INFO "rpcrdma: connection to %pI4:%u "
			"on %s, memreg %d slots %d ird %d%s\n",
			&addr->sin_addr.s_addr,
			ntohs(addr->sin_port),
			ia->ri_id->device->name,
			ia->ri_memreg_strategy,
			xprt->rx_buf.rb_max_requests,
			ird, ird < 4 && ird < tird / 2 ? " (low!)" : "");
	} else if (connstate < 0) {
		printk(KERN_INFO "rpcrdma: connection to %pI4:%u closed (%d)\n",
			&addr->sin_addr.s_addr,
			ntohs(addr->sin_port),
			connstate);
	}
#endif

	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;

	init_completion(&ia->ri_done);

	id = rdma_create_id(rpcrdma_conn_upcall, xprt, RDMA_PS_TCP, IB_QPT_RC);
	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 = -ETIMEDOUT;
	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_interruptible_timeout(&ia->ri_done,
				msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1);
	rc = ia->ri_async_rc;
	if (rc)
		goto out;

	ia->ri_async_rc = -ETIMEDOUT;
	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_interruptible_timeout(&ia->ri_done,
				msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1);
	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, mem_priv;
	struct ib_device_attr devattr;
	struct rpcrdma_ia *ia = &xprt->rx_ia;

	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;
	}

	/*
	 * Query the device to determine if the requested memory
	 * registration strategy is supported. If it isn't, set the
	 * strategy to a globally supported model.
	 */
	rc = ib_query_device(ia->ri_id->device, &devattr);
	if (rc) {
		dprintk("RPC:       %s: ib_query_device failed %d\n",
			__func__, rc);
		goto out2;
	}

	if (devattr.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY) {
		ia->ri_have_dma_lkey = 1;
		ia->ri_dma_lkey = ia->ri_id->device->local_dma_lkey;
	}

	switch (memreg) {
	case RPCRDMA_MEMWINDOWS:
	case RPCRDMA_MEMWINDOWS_ASYNC:
		if (!(devattr.device_cap_flags & IB_DEVICE_MEM_WINDOW)) {
			dprintk("RPC:       %s: MEMWINDOWS registration "
				"specified but not supported by adapter, "
				"using slower RPCRDMA_REGISTER\n",
				__func__);
			memreg = RPCRDMA_REGISTER;
		}
		break;
	case RPCRDMA_MTHCAFMR:
		if (!ia->ri_id->device->alloc_fmr) {
#if RPCRDMA_PERSISTENT_REGISTRATION
			dprintk("RPC:       %s: MTHCAFMR registration "
				"specified but not supported by adapter, "
				"using riskier RPCRDMA_ALLPHYSICAL\n",
				__func__);
			memreg = RPCRDMA_ALLPHYSICAL;
#else
			dprintk("RPC:       %s: MTHCAFMR registration "
				"specified but not supported by adapter, "
				"using slower RPCRDMA_REGISTER\n",
				__func__);
			memreg = RPCRDMA_REGISTER;
#endif
		}
		break;
	case RPCRDMA_FRMR:
		/* Requires both frmr reg and local dma lkey */
		if ((devattr.device_cap_flags &
		     (IB_DEVICE_MEM_MGT_EXTENSIONS|IB_DEVICE_LOCAL_DMA_LKEY)) !=
		    (IB_DEVICE_MEM_MGT_EXTENSIONS|IB_DEVICE_LOCAL_DMA_LKEY)) {
#if RPCRDMA_PERSISTENT_REGISTRATION
			dprintk("RPC:       %s: FRMR registration "
				"specified but not supported by adapter, "
				"using riskier RPCRDMA_ALLPHYSICAL\n",
				__func__);
			memreg = RPCRDMA_ALLPHYSICAL;
#else
			dprintk("RPC:       %s: FRMR registration "
				"specified but not supported by adapter, "
				"using slower RPCRDMA_REGISTER\n",
				__func__);
			memreg = RPCRDMA_REGISTER;
#endif
		}
		break;
	}

	/*
	 * 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.
	 */
	switch (memreg) {
	case RPCRDMA_BOUNCEBUFFERS:
	case RPCRDMA_REGISTER:
	case RPCRDMA_FRMR:
		break;
#if RPCRDMA_PERSISTENT_REGISTRATION
	case RPCRDMA_ALLPHYSICAL:
		mem_priv = IB_ACCESS_LOCAL_WRITE |
				IB_ACCESS_REMOTE_WRITE |
				IB_ACCESS_REMOTE_READ;
		goto register_setup;
#endif
	case RPCRDMA_MEMWINDOWS_ASYNC:
	case RPCRDMA_MEMWINDOWS:
		mem_priv = IB_ACCESS_LOCAL_WRITE |
				IB_ACCESS_MW_BIND;
		goto register_setup;
	case RPCRDMA_MTHCAFMR:
		if (ia->ri_have_dma_lkey)
			break;
		mem_priv = IB_ACCESS_LOCAL_WRITE;
	register_setup:
		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;
		}
		break;
	default:
		printk(KERN_ERR "%s: invalid memory registration mode %d\n",
				__func__, memreg);
		rc = -EINVAL;
		goto out2;
	}
	dprintk("RPC:       %s: memory registration strategy is %d\n",
		__func__, memreg);

	/* Else will do memory reg/dereg for each chunk */
	ia->ri_memreg_strategy = memreg;

	return 0;
out2:
	rdma_destroy_id(ia->ri_id);
	ia->ri_id = NULL;
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)) {
		if (ia->ri_id->qp)
			rdma_destroy_qp(ia->ri_id);
		rdma_destroy_id(ia->ri_id);
		ia->ri_id = NULL;
	}
	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);
	}
}

/*
 * 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, err;

	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_FRMR:
		/* Add room for frmr register and invalidate WRs.
		 * 1. FRMR reg WR for head
		 * 2. FRMR invalidate WR for head
		 * 3. FRMR reg WR for pagelist
		 * 4. FRMR invalidate WR for pagelist
		 * 5. FRMR reg WR for tail
		 * 6. FRMR invalidate WR for tail
		 * 7. The RDMA_SEND WR
		 */
		ep->rep_attr.cap.max_send_wr *= 7;
		if (ep->rep_attr.cap.max_send_wr > devattr.max_qp_wr) {
			cdata->max_requests = devattr.max_qp_wr / 7;
			if (!cdata->max_requests)
				return -EINVAL;
			ep->rep_attr.cap.max_send_wr = cdata->max_requests * 7;
		}
		break;
	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 */
	ep->rep_remote_cma.initiator_depth = 0;
	if (ia->ri_memreg_strategy == RPCRDMA_BOUNCEBUFFERS)
		ep->rep_remote_cma.responder_resources = 0;
	else if (devattr.max_qp_rd_atom > 32)	/* arbitrary but <= 255 */
		ep->rep_remote_cma.responder_resources = 32;
	else
		ep->rep_remote_cma.responder_resources = devattr.max_qp_rd_atom;

	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:
	err = ib_destroy_cq(ep->rep_cq);
	if (err)
		dprintk("RPC:       %s: ib_destroy_cq returned %i\n",
			__func__, err);
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);
		rdma_destroy_qp(ia->ri_id);
		ia->ri_id->qp = 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;
	}

	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;

	if (ep->rep_connected != 0) {
		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_qp(ia->ri_id);
		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);
	}
}

	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;
	}

	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. */
		if (retry_count++ <= RDMA_CONNECT_RETRY_MAX + 1 &&
		    (ep->rep_remote_cma.responder_resources == 0 ||
		     ep->rep_remote_cma.initiator_depth !=
				ep->rep_remote_cma.responder_resources)) {
			if (ep->rep_remote_cma.responder_resources == 0)
				ep->rep_remote_cma.responder_resources = 1;
			ep->rep_remote_cma.initiator_depth =
				ep->rep_remote_cma.responder_resources;
			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;
	struct rpcrdma_mw *r;

	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, fmr's or frmr'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_FRMR:
		len += buf->rb_max_requests * RPCRDMA_MAX_SEGS *
				sizeof(struct rpcrdma_mw);
		break;
	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: