xref: /illumos-gate/usr/src/uts/common/io/ib/mgt/ibcm/ibcm_ti.c (revision fc8ae2ec)

/*
 * CDDL HEADER START
 *
 * The contents of this file are subject to the terms of the
 * Common Development and Distribution License (the "License").
 * You may not use this file except in compliance with the License.
 *
 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
 * or http://www.opensolaris.org/os/licensing.
 * See the License for the specific language governing permissions
 * and limitations under the License.
 *
 * When distributing Covered Code, include this CDDL HEADER in each
 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
 * If applicable, add the following below this CDDL HEADER, with the
 * fields enclosed by brackets "[]" replaced with your own identifying
 * information: Portions Copyright [yyyy] [name of copyright owner]
 *
 * CDDL HEADER END
 */
/*
 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
 */

#include <sys/ib/mgt/ibcm/ibcm_impl.h>
#include <sys/ib/ibtl/ibti.h>
#include <sys/ib/mgt/ibcm/ibcm_arp.h>

/*
 * ibcm_ti.c
 *	These routines implement the Communication Manager's interfaces to IBTL.
 */

/* CM rc recycle task args structure definition */
typedef struct ibcm_taskq_recycle_arg_s {
	ibt_channel_hdl_t	rc_chan;
	ibt_cep_flags_t		control;
	uint8_t			hca_port_num;
	ibt_recycle_handler_t	func;
	void			*arg;
} ibcm_taskq_recycle_arg_t;

_NOTE(READ_ONLY_DATA(ibcm_taskq_recycle_arg_s))

static ibt_status_t	ibcm_init_reply_addr(ibcm_hca_info_t *hcap,
    ibcm_mad_addr_t *reply_addr, ibt_chan_open_args_t *chan_args,
    ibt_chan_open_flags_t flags, ib_time_t *cm_pkt_lt, ib_lid_t prim_slid);
static void		ibcm_process_abort_via_taskq(void *args);
static ibt_status_t	ibcm_process_rc_recycle_ret(void *recycle_arg);
static ibt_status_t	ibcm_process_join_mcg(void *taskq_arg);
static void		ibcm_process_async_join_mcg(void *tq_arg);

ibt_status_t ibcm_get_node_rec(ibmf_saa_handle_t, sa_node_record_t *,
    uint64_t c_mask, void *, size_t *);

static ibt_status_t ibcm_close_rc_channel(ibt_channel_hdl_t channel,
    ibcm_state_data_t *statep, ibt_execution_mode_t mode);

/* Address Record management definitions */
#define	IBCM_DAPL_ATS_NAME	"DAPL Address Translation Service"
#define	IBCM_DAPL_ATS_SID	0x10000CE100415453ULL
#define	IBCM_DAPL_ATS_NBYTES	16
ibcm_svc_info_t *ibcm_ar_svcinfop;
ibcm_ar_t	*ibcm_ar_list;

/*
 * Tunable parameter to turnoff the overriding of pi_path_mtu value.
 *	1 	By default override the path record's pi_path_mtu value to
 *		IB_MTU_1K for all RC channels. This is done only for the
 *		channels established on Tavor HCA and the path's pi_path_mtu
 *		is greater than IB_MTU_1K.
 *	0	Do not override, use pi_path_mtu by default.
 */
int	ibcm_override_path_mtu = 1;

#ifdef DEBUG
static void	ibcm_print_reply_addr(ibt_channel_hdl_t channel,
		    ibcm_mad_addr_t *cm_reply_addr);
#endif

_NOTE(DATA_READABLE_WITHOUT_LOCK(ibcm_port_info_s::{port_ibmf_hdl}))

/* access is controlled between ibcm_sm.c and ibcm_ti.c by CVs */
_NOTE(SCHEME_PROTECTS_DATA("Serialized access by CV", {ibt_rc_returns_t
    ibt_ud_returns_t ibt_ap_returns_t ibt_ar_t}))

/*
 * Typically, clients initialize these args in one api call, and use in
 * another api
 */
_NOTE(SCHEME_PROTECTS_DATA("Expected usage of ibtl api by client",
    {ibt_path_info_s ibt_cep_path_s ibt_adds_vect_s ibt_mcg_info_s ib_gid_s
    ibt_ud_dest_attr_s ibt_ud_dest_s ibt_srv_data_s ibt_redirect_info_s}))

/*
 * ibt_open_rc_channel()
 *	ibt_open_rc_channel opens a communication channel on the specified
 *	channel to the specified service. For connection service type qp's
 *	the CM initiates the CEP to establish the connection and transitions
 *	the QP/EEC to the "Ready to send" State modifying the QP/EEC's
 *	attributes as necessary.
 *	The implementation of this function assumes that alt path is different
 *	from primary path. It is assumed that the Path functions ensure that.
 *
 * RETURN VALUES:
 *	IBT_SUCCESS	on success (or respective failure on error)
 */
ibt_status_t
ibt_open_rc_channel(ibt_channel_hdl_t channel, ibt_chan_open_flags_t flags,
    ibt_execution_mode_t mode, ibt_chan_open_args_t *chan_args,
    ibt_rc_returns_t *ret_args)
{
	/* all fields that are related to REQ MAD formation */

	ib_pkey_t		prim_pkey;
	ib_lid_t		primary_slid, alternate_slid;
	ib_qpn_t		local_qpn = 0;
	ib_guid_t		hca_guid;
	ib_qkey_t		local_qkey = 0;
	ib_eecn_t		local_eecn = 0;
	ib_eecn_t		remote_eecn = 0;
	boolean_t		primary_grh;
	boolean_t		alternate_grh = B_FALSE;
	ib_lid_t		base_lid;
	ib_com_id_t		local_comid;
	ibmf_msg_t		*ibmf_msg, *ibmf_msg_dreq;
	ibcm_req_msg_t		*req_msgp;

	uint8_t			rdma_in, rdma_out;
	uint8_t			cm_retries;
	uint64_t		local_cm_proc_time;	/* In usec */
	uint8_t			local_cm_resp_time;	/* IB time */
	uint64_t		remote_cm_resp_time;	/* In usec */
	uint32_t		starting_psn = 0;

	/* CM path related fields */
	ibmf_handle_t		ibmf_hdl;
	ibcm_qp_list_t		*cm_qp_entry;
	ibcm_mad_addr_t		cm_reply_addr;

	uint8_t			cm_pkt_lt;

	/* Local args for ibtl/internal CM functions called within */
	ibt_status_t		status;
	ibcm_status_t		lkup_status;
	ibt_qp_query_attr_t	qp_query_attr;

	/* Other misc local args */
	ibt_priv_data_len_t	len;
	ibcm_hca_info_t		*hcap;
	ibcm_state_data_t	*statep;
	uint8_t			port_no;

	IBTF_DPRINTF_L3(cmlog, "ibt_open_rc_channel(chan %p, %X, %x, %p, %p)",
	    channel, flags, mode, chan_args, ret_args);

	if (IBCM_INVALID_CHANNEL(channel)) {
		IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: invalid channel");
		return (IBT_CHAN_HDL_INVALID);
	}

	/* cm handler should always be specified */
	if (chan_args->oc_cm_handler == NULL) {
		IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: chan 0x%p "
		    "CM handler is not be specified", channel);
		return (IBT_INVALID_PARAM);
	}

	if (mode == IBT_NONBLOCKING) {
		if (ret_args != NULL) {
			IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: chan 0x%p"
			    " ret_args should be NULL when called in "
			    "non-blocking mode", channel);
			return (IBT_INVALID_PARAM);
		}
	} else if (mode == IBT_BLOCKING) {
		if (ret_args == NULL) {
			IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: chan 0x%p"
			    " ret_args should be Non-NULL when called in "
			    "blocking mode", channel);
			return (IBT_INVALID_PARAM);
		}
		if (ret_args->rc_priv_data_len > IBT_REP_PRIV_DATA_SZ) {
			IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: chan 0x%p"
			    " private data length is too large", channel);
			return (IBT_INVALID_PARAM);
		}
		if ((ret_args->rc_priv_data_len > 0) &&
		    (ret_args->rc_priv_data == NULL)) {
			IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: chan 0x%p"
			    " rc_priv_data_len > 0, but rc_priv_data NULL",
			    channel);
			return (IBT_INVALID_PARAM);
		}
	} else { /* any other mode is not valid for ibt_open_rc_channel */
		IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: chan 0x%p "
		    "invalid mode %x specified", channel, mode);
		return (IBT_INVALID_PARAM);
	}

	/*
	 * XXX: no support yet for ibt_chan_open_flags_t - IBT_OCHAN_DUP
	 */
	if (flags & IBT_OCHAN_DUP) {
		IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: chan 0x%p "
		    "Unsupported Flags specified: 0x%X", channel, flags);
		return (IBT_INVALID_PARAM);
	}

	if ((flags & IBT_OCHAN_REDIRECTED) &&
	    (flags & IBT_OCHAN_PORT_REDIRECTED)) {
		IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: chan 0x%p "
		    "Illegal to specify IBT_OCHAN_REDIRECTED and "
		    "IBT_OCHAN_PORT_REDIRECTED flags together", channel);
		return (IBT_INVALID_PARAM);
	}

	if (((flags & IBT_OCHAN_REDIRECTED) &&
	    (chan_args->oc_cm_redirect_info == NULL)) ||
	    ((flags & IBT_OCHAN_PORT_REDIRECTED) &&
	    (chan_args->oc_cm_cep_path == NULL))) {
		IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: chan 0x%p "
		    "Redirect flag specified, but respective arg is NULL",
		    channel);
		return (IBT_INVALID_PARAM);
	}

	if ((flags & IBT_OCHAN_REDIRECTED) &&
	    (chan_args->oc_cm_redirect_info->rdi_dlid == 0) &&
	    (chan_args->oc_cm_redirect_info->rdi_gid.gid_guid == 0)) {
		IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: chan 0x%p "
		    "Either rdi_dlid or rdi_gid must be specified for"
		    " IBT_OCHAN_REDIRECTED", channel);
		return (IBT_INVALID_PARAM);
	}

	/* primary dlid and hca_port_num should never be zero */
	port_no = IBCM_PRIM_CEP_PATH(chan_args).cep_hca_port_num;

	if ((IBCM_PRIM_ADDS_VECT(chan_args).av_dlid == 0) && (port_no == 0)) {
		IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: chan 0x%p "
		    "Primary Path's information is not valid", channel);
		return (IBT_INVALID_PARAM);
	}

	/* validate SID */
	if (chan_args->oc_path->pi_sid == 0) {
		IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: chan 0x%p "
		    "ERROR: Service ID in path information is 0", channel);
		return (IBT_INVALID_PARAM);
	}
	IBTF_DPRINTF_L3(cmlog, "ibt_open_rc_channel: chan 0x%p  SID %llX",
	    channel, chan_args->oc_path->pi_sid);

	/* validate rnr_retry_cnt (enum has more than 3 bits) */
	if ((uint_t)chan_args->oc_path_rnr_retry_cnt > IBT_RNR_INFINITE_RETRY) {
		IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: chan 0x%p "
		    "ERROR: oc_path_rnr_retry_cnt(%d) is out of range",
		    channel, chan_args->oc_path_rnr_retry_cnt);
		return (IBT_INVALID_PARAM);
	}

	/*
	 * Ensure that client is not re-using a QP that is still associated
	 * with a statep
	 */
	IBCM_GET_CHAN_PRIVATE(channel, statep);
	if (statep != NULL) {
		IBCM_RELEASE_CHAN_PRIVATE(channel);
		IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: chan 0x%p "
		    "Channel being re-used on active side", channel);
		return (IBT_CHAN_IN_USE);
	}

	/* Get GUID from Channel */
	hca_guid = ibt_channel_to_hca_guid(channel);

	/* validate QP's hca guid with that from primary path  */
	if (hca_guid != chan_args->oc_path->pi_hca_guid) {
		IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: chan 0x%p "
		    "GUID from Channel and primary path don't match", channel);
		IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: chan 0x%p "
		    "Channel GUID %llX primary path GUID %llX", channel,
		    hca_guid, chan_args->oc_path->pi_hca_guid);
		return (IBT_CHAN_HDL_INVALID);
	}

	IBTF_DPRINTF_L5(cmlog, "ibt_open_rc_channel: chan 0x%p "
	    "Local HCA GUID %llX", channel, hca_guid);

	status = ibt_query_qp(channel, &qp_query_attr);
	if (status != IBT_SUCCESS) {
		IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: chan 0x%p "
		    "ibt_query_qp failed %d", channel, status);
		return (status);
	}

	/* If client specified "no port change on QP" */
	if ((qp_query_attr.qp_info.qp_transport.rc.rc_path.cep_hca_port_num !=
	    port_no) && (flags & IBT_OCHAN_PORT_FIXED)) {
		IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: chan 0x%p "
		    "chan port %d and path port %d does not match", channel,
		    qp_query_attr.qp_info.qp_transport.rc.rc_path. \
		    cep_hca_port_num, port_no);
		return (IBT_INVALID_PARAM);
	}

	if (qp_query_attr.qp_info.qp_trans != IBT_RC_SRV) {
		IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: chan 0x%p "
		    "Invalid Channel type: Applicable only to RC Channel",
		    channel);
		return (IBT_CHAN_SRV_TYPE_INVALID);
	}

	/* Check if QP is in INIT state or not */
	if (qp_query_attr.qp_info.qp_state != IBT_STATE_INIT) {
		IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: chan 0x%p "
		    "QP is not in INIT state %x", channel,
		    qp_query_attr.qp_info.qp_state);
		return (IBT_CHAN_STATE_INVALID);
	}

	local_qpn = qp_query_attr.qp_qpn;

	IBTF_DPRINTF_L5(cmlog, "ibt_open_rc_channel: chan 0x%p Active QPN 0x%x",
	    channel, local_qpn);

#ifdef	NO_EEC_SUPPORT_YET

	if (flags & IBT_OCHAN_RDC_EXISTS) {
		ibt_eec_query_attr_t	eec_query_attr;

		local_qkey = qp_query_attr.qp_info.qp_transport.rd_qkey;

		IBTF_DPRINTF_L5(cmlog, "ibt_open_rc_channel: RD");

		status = ibt_query_eec(channel, &eec_query_attr);
		if (status != IBT_SUCCESS) {
			IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: chan 0x%p"
			    " ibt_query_eec failed %d", channel, status);
			return (status);
		}
		local_eecn = eec_query_attr.eec_eecn;
	}

#endif
	if (chan_args->oc_path->pi_prim_pkt_lt > ibcm_max_ib_pkt_lt) {
		IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: chan 0x%p "
		    "Huge PktLifeTime %d, Max is %d", channel,
		    chan_args->oc_path->pi_prim_pkt_lt, ibcm_max_ib_pkt_lt);
		return (IBT_PATH_PKT_LT_TOO_HIGH);
	}

	/* If no HCA found return failure */
	if ((hcap = ibcm_find_hca_entry(hca_guid)) == NULL) {
		IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: chan 0x%p "
		    "hcap is NULL. Probably hca is not in active state",
		    channel);
		return (IBT_CHAN_HDL_INVALID);
	}

	rdma_out = chan_args->oc_rdma_ra_out;
	rdma_in = chan_args->oc_rdma_ra_in;

	if ((rdma_in > hcap->hca_max_rdma_in_qp) ||
	    (rdma_out > hcap->hca_max_rdma_out_qp)) {
		IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: chan 0x%p "
		    "rdma in %d/out %d values exceed hca limits(%d/%d)",
		    channel, rdma_in, rdma_out, hcap->hca_max_rdma_in_qp,
		    hcap->hca_max_rdma_out_qp);
		ibcm_dec_hca_acc_cnt(hcap);
		return (IBT_INVALID_PARAM);
	}

	IBTF_DPRINTF_L5(cmlog, "ibt_open_rc_channel: chan 0x%p "
	    "rdma_in %d rdma_out %d", channel, rdma_in, rdma_out);

	status = ibt_get_port_state_byguid(hcap->hca_guid, port_no,
	    NULL, &base_lid);
	if (status != IBT_SUCCESS) {
		IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: chan 0x%p "
		    "primary port_num %d not active", channel, port_no);
		ibcm_dec_hca_acc_cnt(hcap);
		return (status);
	}

	/* Validate P_KEY Index */
	status = ibt_index2pkey_byguid(hcap->hca_guid, port_no,
	    IBCM_PRIM_CEP_PATH(chan_args).cep_pkey_ix, &prim_pkey);
	if (status != IBT_SUCCESS) {
		IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: chan 0x%p "
		    "Invalid Primary PKeyIx %x", channel,
		    IBCM_PRIM_CEP_PATH(chan_args).cep_pkey_ix);
		ibcm_dec_hca_acc_cnt(hcap);
		return (status);
	}

	IBTF_DPRINTF_L5(cmlog, "ibt_open_rc_channel: chan 0x%p "
	    "primary_port_num %d primary_pkey 0x%x", channel, port_no,
	    prim_pkey);

	if ((hcap->hca_port_info[port_no - 1].port_ibmf_hdl == NULL) &&
	    ((status = ibcm_hca_reinit_port(hcap, port_no - 1))
	    != IBT_SUCCESS)) {
		IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: chan 0x%p "
		    "ibmf reg or callback setup failed during re-initialize",
		    channel);
		ibcm_dec_hca_acc_cnt(hcap);
		return (status);
	}

	ibmf_hdl = hcap->hca_port_info[port_no - 1].port_ibmf_hdl;
	IBTF_DPRINTF_L5(cmlog, "ibt_open_rc_channel: chan 0x%p "
	    "primary ibmf_hdl = 0x%p", channel, ibmf_hdl);

	primary_slid = base_lid + IBCM_PRIM_ADDS_VECT(chan_args).av_src_path;

	IBTF_DPRINTF_L5(cmlog, "ibt_open_rc_channel: channel 0x%p "
	    "primary SLID = %x", channel, primary_slid);

	/* check first if alternate path exists or not as it is OPTIONAL */
	if (IBCM_ALT_CEP_PATH(chan_args).cep_hca_port_num != 0) {
		uint8_t	alt_port_no;

		alt_port_no = IBCM_ALT_CEP_PATH(chan_args).cep_hca_port_num;

		if (chan_args->oc_path->pi_alt_pkt_lt > ibcm_max_ib_pkt_lt) {
			IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: chan 0x%p "
			    "Huge Alt Pkt lt %d", channel,
			    chan_args->oc_path->pi_alt_pkt_lt);
			ibcm_dec_hca_acc_cnt(hcap);
			return (IBT_PATH_PKT_LT_TOO_HIGH);
		}

		if (port_no != alt_port_no) {

			status = ibt_get_port_state_byguid(hcap->hca_guid,
			    alt_port_no, NULL, &base_lid);
			if (status != IBT_SUCCESS) {

				IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: "
				    "chan 0x%p alt_port_num %d inactive %d",
				    channel, alt_port_no, status);
				ibcm_dec_hca_acc_cnt(hcap);
				return (status);
			}

		}
		alternate_slid =
		    base_lid + IBCM_ALT_ADDS_VECT(chan_args).av_src_path;

		IBTF_DPRINTF_L5(cmlog, "ibt_open_rc_channel: chan 0x%p "
		    "alternate SLID = %x", channel, alternate_slid);
	}

	/*
	 * only pkey needs to be zero'ed, because all other fields are set in
	 * in ibcm_init_reply_addr. But, let's bzero the complete struct for
	 * any future modifications.
	 */
	bzero(&cm_reply_addr, sizeof (cm_reply_addr));

	/* Initialize the MAD destination address in stored_reply_addr */
	if ((status = ibcm_init_reply_addr(hcap, &cm_reply_addr, chan_args,
	    flags, &cm_pkt_lt, primary_slid)) != IBT_SUCCESS) {

		IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: chan 0x%p "
		    "ibcm_init_reply_addr failed status %d ", channel, status);
		ibcm_dec_hca_acc_cnt(hcap);
		return (status);
	}


	/* Initialize the pkey for CM MAD communication */
	if (cm_reply_addr.rcvd_addr.ia_p_key == 0)
		cm_reply_addr.rcvd_addr.ia_p_key = prim_pkey;

#ifdef DEBUG
	ibcm_print_reply_addr(channel, &cm_reply_addr);
#endif

	/* Retrieve an ibmf qp for sending CM MADs */
	if ((cm_qp_entry = ibcm_find_qp(hcap, port_no,
	    cm_reply_addr.rcvd_addr.ia_p_key)) == NULL) {
		IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: chan 0x%p "
		    "unable to allocate ibmf qp for CM MADs", channel);
		ibcm_dec_hca_acc_cnt(hcap);
		return (IBT_INSUFF_RESOURCE);
	}


	if (ibcm_alloc_comid(hcap, &local_comid) != IBCM_SUCCESS) {
		ibcm_release_qp(cm_qp_entry);
		ibcm_dec_hca_acc_cnt(hcap);
		IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: chan 0x%p"
		    " Unable to allocate comid", channel);
		return (IBT_INSUFF_KERNEL_RESOURCE);
	}

	/* allocate an IBMF mad buffer (REQ) */
	if ((status = ibcm_alloc_out_msg(ibmf_hdl, &ibmf_msg,
	    MAD_METHOD_SEND)) != IBT_SUCCESS) {
		IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: "
		    "chan 0x%p ibcm_alloc_out_msg failed", channel);
		ibcm_release_qp(cm_qp_entry);
		ibcm_free_comid(hcap, local_comid);
		ibcm_dec_hca_acc_cnt(hcap);
		return (status);
	}

	/* allocate an IBMF mad buffer (DREQ) */
	if ((status = ibcm_alloc_out_msg(ibmf_hdl, &ibmf_msg_dreq,
	    MAD_METHOD_SEND)) != IBT_SUCCESS) {
		IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: "
		    "chan 0x%p ibcm_alloc_out_msg failed", channel);
		(void) ibcm_free_out_msg(ibmf_hdl, &ibmf_msg);
		ibcm_release_qp(cm_qp_entry);
		ibcm_free_comid(hcap, local_comid);
		ibcm_dec_hca_acc_cnt(hcap);
		return (status);
	}

	/* Init to Init, if QP's port does not match with path information */
	if (qp_query_attr.qp_info.qp_transport.rc.rc_path.cep_hca_port_num !=
	    IBCM_PRIM_CEP_PATH(chan_args).cep_hca_port_num) {

		ibt_qp_info_t		qp_info;
		ibt_cep_modify_flags_t	cep_flags;

		IBTF_DPRINTF_L5(cmlog, "ibt_open_rc_channel: "
		    "chan 0x%p chan port %d", channel,
		    qp_query_attr.qp_info.qp_transport.rc.rc_path.\
		    cep_hca_port_num);

		IBTF_DPRINTF_L5(cmlog, "ibt_open_rc_channel: "
		    "chan 0x%p path port %d", channel, port_no);

		bzero(&qp_info, sizeof (qp_info));
		/* For now, set it to RC type */

		qp_info.qp_trans = IBT_RC_SRV;
		qp_info.qp_state = IBT_STATE_INIT;
		qp_info.qp_transport.rc.rc_path.cep_hca_port_num = port_no;

		cep_flags = IBT_CEP_SET_STATE | IBT_CEP_SET_PORT;

		status = ibt_modify_qp(channel, cep_flags, &qp_info, NULL);

		if (status != IBT_SUCCESS) {
			IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: "
			    "chan 0x%p ibt_modify_qp() = %d", channel, status);
			ibcm_release_qp(cm_qp_entry);
			ibcm_free_comid(hcap, local_comid);
			ibcm_dec_hca_acc_cnt(hcap);
			(void) ibcm_free_out_msg(ibmf_hdl, &ibmf_msg);
			(void) ibcm_free_out_msg(ibmf_hdl, &ibmf_msg_dreq);
			return (status);
		} else
			IBTF_DPRINTF_L5(cmlog, "ibt_open_rc_channel: "
			    "chan 0x%p ibt_modify_qp() = %d", channel, status);
	}

	/* allocate ibcm_state_data_t before grabbing the WRITER lock */
	statep = kmem_zalloc(sizeof (ibcm_state_data_t), KM_SLEEP);
	rw_enter(&hcap->hca_state_rwlock, RW_WRITER);
	lkup_status = ibcm_lookup_msg(IBCM_OUTGOING_REQ, local_comid, 0, 0,
	    hcap, &statep);
	rw_exit(&hcap->hca_state_rwlock);

	/* CM should be seeing this for the first time */
	ASSERT(lkup_status == IBCM_LOOKUP_NEW);

	/* Increment the hca's resource count */
	ibcm_inc_hca_res_cnt(hcap);

	/* Once a resource created on hca, no need to hold the acc cnt */
	ibcm_dec_hca_acc_cnt(hcap);

	_NOTE(NOW_INVISIBLE_TO_OTHER_THREADS(*statep))

	statep->timerid = 0;
	statep->local_hca_guid = hca_guid;
	statep->local_qpn = local_qpn;
	statep->stored_reply_addr.cm_qp_entry = cm_qp_entry;
	statep->prim_port = IBCM_PRIM_CEP_PATH(chan_args).cep_hca_port_num;
	statep->alt_port = IBCM_ALT_CEP_PATH(chan_args).cep_hca_port_num;


	/* Save "statep" as channel's CM private data.  */
	statep->channel = channel;
	IBCM_SET_CHAN_PRIVATE(statep->channel, statep);

	statep->stored_msg = ibmf_msg;
	statep->dreq_msg = ibmf_msg_dreq;

	_NOTE(NOW_INVISIBLE_TO_OTHER_THREADS(*req_msgp))

	/* Start filling in the REQ MAD */
	req_msgp = (ibcm_req_msg_t *)IBCM_OUT_MSGP(statep->stored_msg);
	req_msgp->req_local_comm_id = h2b32(local_comid);
	req_msgp->req_svc_id = h2b64(chan_args->oc_path->pi_sid);
	req_msgp->req_local_ca_guid = h2b64(hca_guid);
	req_msgp->req_local_qkey = h2b32(local_qkey);	/* for EEC/RD */

	/* Bytes 32-35 are req_local_qpn and req_off_resp_resources */
	req_msgp->req_local_qpn_plus = h2b32(local_qpn << 8 | rdma_in);

	/* Bytes 36-39 are req_local_eec_no and req_off_initiator_depth */
	req_msgp->req_local_eec_no_plus = h2b32(local_eecn << 8 | rdma_out);

	if (flags & IBT_OCHAN_REMOTE_CM_TM)
		remote_cm_resp_time = chan_args->oc_remote_cm_time;
	else
		remote_cm_resp_time = ibcm_remote_response_time;

	/*
	 * Bytes 40-43 - remote_eecn, remote_cm_resp_time, tran_type,
	 * IBT_CM_FLOW_CONTROL is always set by default.
	 */
	req_msgp->req_remote_eecn_plus = h2b32(
	    remote_eecn << 8 | (ibt_usec2ib(remote_cm_resp_time) & 0x1f) << 3 |
	    IBT_RC_SRV << 1 | IBT_CM_FLOW_CONTROL);

	if (flags & IBT_OCHAN_LOCAL_CM_TM)
		local_cm_proc_time = chan_args->oc_local_cm_time;
	else
		local_cm_proc_time = ibcm_local_processing_time;

	local_cm_resp_time = ibt_usec2ib(local_cm_proc_time +
	    2 * ibt_ib2usec(chan_args->oc_path->pi_prim_pkt_lt) +
	    ibcm_sw_delay);

	/* save retry count */
	statep->cep_retry_cnt = chan_args->oc_path_retry_cnt;

	if (flags & IBT_OCHAN_STARTING_PSN)
		starting_psn = chan_args->oc_starting_psn;

	if (local_cm_resp_time > 0x1f)
		local_cm_resp_time = 0x1f;

	/* Bytes 44-47 are req_starting_psn, local_cm_resp_time and retry_cnt */
	req_msgp->req_starting_psn_plus = h2b32(starting_psn << 8 |
	    local_cm_resp_time << 3 | statep->cep_retry_cnt);

	IBTF_DPRINTF_L5(cmlog, "ibt_open_rc_channel: chan 0x%p "
	    "Prim Pkt lt (IB time) 0x%x", channel,
	    chan_args->oc_path->pi_prim_pkt_lt);

	IBTF_DPRINTF_L5(cmlog, "ibt_open_rc_channel: chan 0x%p "
	    "local_cm_proc_time(usec) %d ", channel, local_cm_proc_time);

	IBTF_DPRINTF_L5(cmlog, "ibt_open_rc_channel: chan 0x%p "
	    "local_cm_resp_time(ib_time) %d", channel, local_cm_resp_time);

	IBTF_DPRINTF_L5(cmlog, "ibt_open_rc_channel: chan 0x%p "
	    "remote_cm_resp_time (usec) %d", channel, remote_cm_resp_time);

	statep->starting_psn = starting_psn;

	/* Pkey - bytes 48-49 */
	req_msgp->req_part_key = h2b16(prim_pkey);

	if (flags & IBT_OCHAN_CM_RETRY)
		cm_retries = chan_args->oc_cm_retry_cnt;
	else
		cm_retries = ibcm_max_retries;

	statep->max_cm_retries = statep->remaining_retry_cnt = cm_retries;
	req_msgp->req_max_cm_retries_plus = statep->max_cm_retries << 4;

	/*
	 * Check whether SRQ is associated with this Channel, if yes, then
	 * set the SRQ Exists bit in the REQ.
	 */
	if (qp_query_attr.qp_srq != NULL) {
		req_msgp->req_max_cm_retries_plus |= (1 << 3);
	}

	/*
	 * By default on Tavor, we override the PathMTU to 1K.
	 * To turn this off, set ibcm_override_path_mtu = 0.
	 */
	if (ibcm_override_path_mtu && IBCM_IS_HCA_TAVOR(hcap) &&
	    (chan_args->oc_path->pi_path_mtu > IB_MTU_1K)) {
		req_msgp->req_mtu_plus = IB_MTU_1K << 4 |
		    chan_args->oc_path_rnr_retry_cnt;
		IBTF_DPRINTF_L3(cmlog, "ibt_open_rc_channel: chan 0x%p PathMTU"
		    " overridden to IB_MTU_1K(%d) from %d", channel, IB_MTU_1K,
		    chan_args->oc_path->pi_path_mtu);
	} else
		req_msgp->req_mtu_plus = chan_args->oc_path->pi_path_mtu << 4 |
		    chan_args->oc_path_rnr_retry_cnt;

	IBTF_DPRINTF_L5(cmlog, "ibt_open_rc_channel: chan 0x%p CM retry cnt %d"
	    " staring PSN %x", channel, cm_retries, starting_psn);


#ifdef	NO_EEC_SUPPORT_YET
	if (flags & IBT_OCHAN_RDC_EXISTS)
		req_msgp->req_mtu_plus |= 8;
#endif

	/* Initialize the "primary" port stuff next - bytes 52-95 */
	req_msgp->req_primary_l_port_lid = h2b16(primary_slid);
	req_msgp->req_primary_r_port_lid =
	    h2b16(IBCM_PRIM_ADDS_VECT(chan_args).av_dlid);
	req_msgp->req_primary_l_port_gid.gid_prefix =
	    h2b64(IBCM_PRIM_ADDS_VECT(chan_args).av_sgid.gid_prefix);
	req_msgp->req_primary_l_port_gid.gid_guid =
	    h2b64(IBCM_PRIM_ADDS_VECT(chan_args).av_sgid.gid_guid);
	req_msgp->req_primary_r_port_gid.gid_prefix =
	    h2b64(IBCM_PRIM_ADDS_VECT(chan_args).av_dgid.gid_prefix);
	req_msgp->req_primary_r_port_gid.gid_guid =
	    h2b64(IBCM_PRIM_ADDS_VECT(chan_args).av_dgid.gid_guid);
	primary_grh = IBCM_PRIM_ADDS_VECT(chan_args).av_send_grh;

	statep->remote_hca_guid = /* not correct, but helpful for debugging */
	    IBCM_PRIM_ADDS_VECT(chan_args).av_dgid.gid_guid;

	/* Bytes 88-91 - primary_flowlbl, and primary_srate */
	req_msgp->req_primary_flow_label_plus =
	    h2b32(((primary_grh == B_TRUE) ?
	    (IBCM_PRIM_ADDS_VECT(chan_args).av_flow << 12) : 0) |
	    IBCM_PRIM_ADDS_VECT(chan_args).av_srate);
	req_msgp->req_primary_traffic_class = (primary_grh == B_TRUE) ?
	    IBCM_PRIM_ADDS_VECT(chan_args).av_tclass : 0;
	req_msgp->req_primary_hop_limit = (primary_grh == B_TRUE) ?
	    IBCM_PRIM_ADDS_VECT(chan_args).av_hop : 1;
	req_msgp->req_primary_sl_plus =
	    IBCM_PRIM_ADDS_VECT(chan_args).av_srvl << 4 |
	    ((primary_grh == B_TRUE) ? 0 : 8);

	req_msgp->req_primary_localtime_plus =
	    ibt_usec2ib((2 * ibt_ib2usec(chan_args->oc_path->pi_prim_pkt_lt)) +
	    ibt_ib2usec(hcap->hca_ack_delay)) << 3;

	IBTF_DPRINTF_L2(cmlog, "ibt_open_rc_channel: chan %p statep %p",
	    channel, statep);
	IBTF_DPRINTF_L5(cmlog, "ibt_open_rc_channel: chan 0x%p "
	    "active hca_ack_delay (usec) %d", channel,
	    req_msgp->req_primary_localtime_plus);

	IBTF_DPRINTF_L5(cmlog, "ibt_open_rc_channel: chan 0x%p "
	    "Sent primary cep timeout (IB Time) %d", channel,
	    hcap->hca_ack_delay);

	IBTF_DPRINTF_L5(cmlog, "ibt_open_rc_channel: chan 0x%p prim_dlid %x ",
	    channel, IBCM_PRIM_ADDS_VECT(chan_args).av_dlid);

	IBTF_DPRINTF_L5(cmlog, "ibt_open_rc_channel: chan 0x%p "
	    "prim GID %llX:%llX", channel,
	    IBCM_PRIM_ADDS_VECT(chan_args).av_dgid.gid_prefix,
	    IBCM_PRIM_ADDS_VECT(chan_args).av_dgid.gid_guid);

	/* Initialize the "alternate" port stuff - optional */
	if (chan_args->oc_path->pi_alt_cep_path.cep_hca_port_num != 0) {
		ib_gid_t	tmp_gid;

		req_msgp->req_alt_l_port_lid = h2b16(alternate_slid);
		req_msgp->req_alt_r_port_lid =
		    h2b16(IBCM_ALT_ADDS_VECT(chan_args).av_dlid);
		/*
		 * doing all this as req_alt_r/l_port_gid is at offset
		 * 100, 116 which is not divisible by 8
		 */

		tmp_gid.gid_prefix =
		    h2b64(IBCM_ALT_ADDS_VECT(chan_args).av_dgid.gid_prefix);
		tmp_gid.gid_guid =
		    h2b64(IBCM_ALT_ADDS_VECT(chan_args).av_dgid.gid_guid);
		bcopy(&tmp_gid, &req_msgp->req_alt_r_port_gid[0],
		    sizeof (ib_gid_t));
		tmp_gid.gid_prefix =
		    h2b64(IBCM_ALT_ADDS_VECT(chan_args).av_sgid.gid_prefix);
		tmp_gid.gid_guid =
		    h2b64(IBCM_ALT_ADDS_VECT(chan_args).av_sgid.gid_guid);

		bcopy(&tmp_gid, &req_msgp->req_alt_l_port_gid[0],
		    sizeof (ib_gid_t));
		alternate_grh = IBCM_ALT_ADDS_VECT(chan_args).av_send_grh;

		/* Bytes 132-135 - alternate_flow_label, and alternate srate */
		req_msgp->req_alt_flow_label_plus = h2b32(
		    (((alternate_grh == B_TRUE) ?
		    (IBCM_ALT_ADDS_VECT(chan_args).av_flow << 12) : 0) |
		    IBCM_ALT_ADDS_VECT(chan_args).av_srate));
		req_msgp->req_alt_traffic_class = (alternate_grh == B_TRUE) ?
		    IBCM_ALT_ADDS_VECT(chan_args).av_tclass : 0;
		req_msgp->req_alt_hop_limit = (alternate_grh == B_TRUE) ?
		    IBCM_ALT_ADDS_VECT(chan_args).av_hop : 1;
		req_msgp->req_alt_sl_plus =
		    IBCM_ALT_ADDS_VECT(chan_args).av_srvl << 4 |
		    ((alternate_grh == B_TRUE) ? 0 : 8);
		req_msgp->req_alt_localtime_plus = ibt_usec2ib((2 *
		    ibt_ib2usec(chan_args->oc_path->pi_alt_pkt_lt)) +
		    ibt_ib2usec(hcap->hca_ack_delay)) << 3;

		IBTF_DPRINTF_L5(cmlog, "ibt_open_rc_channel: chan 0x%p "
		    "alt_dlid %x ", channel,
		    IBCM_ALT_ADDS_VECT(chan_args).av_dlid);

		IBTF_DPRINTF_L5(cmlog, "ibt_open_rc_channel: chan 0x%p "
		    "alt GID %llX:%llX", channel,
		    IBCM_ALT_ADDS_VECT(chan_args).av_dgid.gid_prefix,
		    IBCM_ALT_ADDS_VECT(chan_args).av_dgid.gid_guid);
	}

	len = min(chan_args->oc_priv_data_len, IBT_REQ_PRIV_DATA_SZ);
	if ((len > 0) && chan_args->oc_priv_data)
		bcopy(chan_args->oc_priv_data, req_msgp->req_private_data, len);

	_NOTE(NOW_VISIBLE_TO_OTHER_THREADS(*req_msgp))

	/* return_data is filled up in the state machine code */
	if (ret_args != NULL) {
		statep->open_return_data = ret_args;
	}

	/* initialize some statep fields here */
	statep->mode = IBCM_ACTIVE_MODE;
	statep->hcap = hcap;

	statep->cm_handler = chan_args->oc_cm_handler;
	statep->state_cm_private = chan_args->oc_cm_clnt_private;

	statep->pkt_life_time =
	    ibt_ib2usec(chan_args->oc_path->pi_prim_pkt_lt);

	statep->timer_value = ibt_ib2usec(ibt_usec2ib(
	    2 * ibt_ib2usec(cm_pkt_lt) + remote_cm_resp_time));

	/* Initialize statep->stored_reply_addr */
	statep->stored_reply_addr.ibmf_hdl = ibmf_hdl;

	/* Initialize stored reply addr fields */
	statep->stored_reply_addr.grh_hdr = cm_reply_addr.grh_hdr;
	statep->stored_reply_addr.rcvd_addr = cm_reply_addr.rcvd_addr;
	statep->stored_reply_addr.grh_exists = cm_reply_addr.grh_exists;
	statep->stored_reply_addr.port_num = cm_reply_addr.port_num;

	/*
	 * The IPD on local/active side is calculated by path functions,
	 * hence available in the args of ibt_open_rc_channel
	 */
	statep->local_srate = IBCM_PRIM_ADDS_VECT(chan_args).av_srate;
	statep->local_alt_srate = IBCM_ALT_ADDS_VECT(chan_args).av_srate;

	/* Store the source path bits for primary and alt paths */
	statep->prim_src_path_bits = IBCM_PRIM_ADDS_VECT(chan_args).av_src_path;
	statep->alt_src_path_bits = IBCM_ALT_ADDS_VECT(chan_args).av_src_path;

	statep->open_flow = 1;
	statep->open_done = B_FALSE;
	statep->state = statep->timer_stored_state = IBCM_STATE_REQ_SENT;
	IBCM_REF_CNT_INCR(statep);	/* Decremented before return */
	IBCM_REF_CNT_INCR(statep);	/* Decremented after REQ is posted */
	statep->send_mad_flags |= IBCM_REQ_POST_BUSY;

	/*
	 * Skip moving channel to error state during close, for OFUV clients.
	 * OFUV clients transition the channel to error state by itself.
	 */
	if (flags & IBT_OCHAN_OFUV)
		statep->is_this_ofuv_chan = B_TRUE;

	IBCM_OUT_HDRP(statep->stored_msg)->AttributeID =
	    h2b16(IBCM_INCOMING_REQ + IBCM_ATTR_BASE_ID);

	IBCM_OUT_HDRP(statep->stored_msg)->TransactionID =
	    h2b64(ibcm_generate_tranid(IBCM_INCOMING_REQ, statep->local_comid,
	    0));

	_NOTE(NOW_VISIBLE_TO_OTHER_THREADS(*statep))

	ibtl_cm_chan_is_opening(channel);

	ibcm_open_enqueue(statep);

	mutex_enter(&statep->state_mutex);

	if (mode == IBT_BLOCKING) {

		/* wait for REQ/REP/RTU */
		while (statep->open_done != B_TRUE) {
			cv_wait(&statep->block_client_cv, &statep->state_mutex);
		}

		/*
		 * In the case that open_channel() fails because of a
		 * REJ or timeout, change retval to IBT_CM_FAILURE
		 */
		if (statep->open_return_data->rc_status != IBT_CM_SUCCESS) {
			status = IBT_CM_FAILURE;
			ibtl_cm_chan_open_is_aborted(channel);
		}

		IBTF_DPRINTF_L3(cmlog, "ibt_open_rc_channel: chan 0x%p "
		    "ret status %d cm status %d", channel, status,
		    statep->open_return_data->rc_status);
	}

	/* decrement the ref-count before leaving here */
	IBCM_REF_CNT_DECR(statep);

	mutex_exit(&statep->state_mutex);

	IBTF_DPRINTF_L4(cmlog, "ibt_open_rc_channel: chan 0x%p done", channel);
	return (status);
}

/*
 * ibcm_init_reply_addr:
 *
 * The brief description of functionality below.
 *
 * For IBT_OCHAN_PORT_REDIRECTED (ie., port redirected case):
 *	Build CM path from chan_args->oc_cm_cep_path
 *	Set CM pkt lt (ie.,life time) to chan_args->oc_cm_pkt_lt
 *
 * For IBT_OCHAN_REDIRECTED (ie., port and CM redirected case):
 *	If Redirect LID is specified,
 *		If Redirect GID is not specified or specified to be on the same
 *		    subnet, then
 *			Build CM path from chan_args->oc_cm_redirect_info
 *			Set CM pkt lt to subnet timeout
 *		Else (ie., GID specified, but on a different subnet)
 *			Do a path lookup to build CM Path and set CM pkt lt
 *
 */
static ibt_status_t
ibcm_init_reply_addr(ibcm_hca_info_t *hcap, ibcm_mad_addr_t *reply_addr,
    ibt_chan_open_args_t *chan_args, ibt_chan_open_flags_t flags,
    ib_time_t *cm_pkt_lt, ib_lid_t prim_slid)
{
	ibt_adds_vect_t	*cm_adds;
	ibt_path_info_t	path;
	boolean_t	cm_grh;
	ibt_status_t	status;

	IBTF_DPRINTF_L5(cmlog, "ibcm_init_reply_addr:");

	_NOTE(NOW_INVISIBLE_TO_OTHER_THREADS(*reply_addr))

	/*
	 * sending side CM lid/gid/port num are not based on any redirect
	 * params. These values are set to primary RC path lid/gid/port num.
	 * In the future, these values can be set based on framework policy
	 * decisions ensuring reachability.
	 */
	reply_addr->grh_hdr.ig_sender_gid =
	    IBCM_PRIM_ADDS_VECT(chan_args).av_sgid;
	reply_addr->rcvd_addr.ia_local_lid = prim_slid;
	reply_addr->port_num = IBCM_PRIM_CEP_PATH(chan_args).cep_hca_port_num;

	if (flags & IBT_OCHAN_PORT_REDIRECTED) {
		IBTF_DPRINTF_L4(cmlog, "ibcm_init_rely_addr: "
		    "IBT_OCHAN_PORT_REDIRECTED specified");

		status = ibt_index2pkey_byguid(hcap->hca_guid,
		    chan_args->oc_cm_cep_path->cep_hca_port_num,
		    chan_args->oc_cm_cep_path->cep_pkey_ix,
		    &reply_addr->rcvd_addr.ia_p_key);

		if (status != IBT_SUCCESS) {
			IBTF_DPRINTF_L2(cmlog, "ibcm_init_rely_addr: Invalid "
			    "CM PKeyIx %x port_num %x",
			    chan_args->oc_cm_cep_path->cep_pkey_ix,
			    chan_args->oc_cm_cep_path->cep_hca_port_num);
			return (status);
		}

		cm_adds = &(chan_args->oc_cm_cep_path->cep_adds_vect);
		IBTF_DPRINTF_L4(cmlog, "ibcm_init_rely_addr: dlid = %x",
		    cm_adds->av_dlid);

		reply_addr->rcvd_addr.ia_q_key = IB_GSI_QKEY;
		reply_addr->rcvd_addr.ia_remote_qno = 1;
		*cm_pkt_lt = chan_args->oc_cm_pkt_lt;

	} else if (flags & IBT_OCHAN_REDIRECTED) {
		ibt_redirect_info_t	*redirect_info;
		ibt_hca_portinfo_t	*port_infop;
		uint_t			psize, nports;

		IBTF_DPRINTF_L4(cmlog, "ibcm_init_rely_addr: "
		    "IBT_OCHAN_REDIRECTED specified");

		redirect_info = chan_args->oc_cm_redirect_info;

		if ((redirect_info->rdi_gid.gid_prefix == 0) ||
		    (redirect_info->rdi_gid.gid_guid == 0)) {
			IBTF_DPRINTF_L2(cmlog, "ibcm_init_reply_addr: "
			    "ERROR: Re-direct GID value NOT Provided.");
			return (IBT_INVALID_PARAM);
		}

		/* As per spec definition 1.1, it's always IB_GSI_QKEY */
		reply_addr->rcvd_addr.ia_q_key = redirect_info->rdi_qkey;
		reply_addr->rcvd_addr.ia_remote_qno = redirect_info->rdi_qpn;
		reply_addr->rcvd_addr.ia_p_key = redirect_info->rdi_pkey;

		/*
		 * if LID is non-zero in classportinfo then use classportinfo
		 * fields to form CM MAD destination address.
		 */
		if (redirect_info->rdi_dlid != 0) {
			status = ibtl_cm_query_hca_ports_byguid(hcap->hca_guid,
			    reply_addr->port_num, &port_infop, &nports, &psize);
			if ((status != IBT_SUCCESS) || (nports == 0)) {
				IBTF_DPRINTF_L2(cmlog, "ibcm_init_reply_addr: "
				    "Query Ports Failed: %d", status);
				return (status);
			} else if (port_infop->p_subnet_timeout >
			    ibcm_max_ib_pkt_lt) {
				IBTF_DPRINTF_L2(cmlog, "ibcm_init_reply_addr: "
				    "large subnet timeout %x port_no %x",
				    port_infop->p_subnet_timeout,
				    reply_addr->port_num);
				ibt_free_portinfo(port_infop, psize);
				return (IBT_PATH_PKT_LT_TOO_HIGH);
			} else {
				IBTF_DPRINTF_L3(cmlog, "ibcm_init_reply_addr: "
				    "subnet timeout %x port_no %x",
				    port_infop->p_subnet_timeout,
				    reply_addr->port_num);

				*cm_pkt_lt =
				    ibt_ib2usec(min(ibcm_max_ib_mad_pkt_lt,
				    port_infop->p_subnet_timeout));

				ibt_free_portinfo(port_infop, psize);
			}

			reply_addr->rcvd_addr.ia_remote_lid =
			    redirect_info->rdi_dlid;
			reply_addr->rcvd_addr.ia_service_level =
			    redirect_info->rdi_sl;
			reply_addr->grh_exists = B_TRUE;
			reply_addr->grh_hdr.ig_recver_gid =
			    redirect_info->rdi_gid;
			reply_addr->grh_hdr.ig_tclass =
			    redirect_info->rdi_tclass;
			reply_addr->grh_hdr.ig_flow_label =
			    redirect_info->rdi_flow;

			/* Classportinfo doesn't have hoplimit field */
			reply_addr->grh_hdr.ig_hop_limit = 1;
			return (IBT_SUCCESS);

		} else {
			ibt_path_attr_t	path_attr;
			ib_gid_t	path_dgid[1];

			/*
			 * If GID is specified, and LID is zero in classportinfo
			 * do a path lookup using specified GID, Pkey,
			 * in classportinfo
			 */

			bzero(&path_attr, sizeof (path_attr));

			path_attr.pa_dgids = &path_dgid[0];
			path_attr.pa_dgids[0] = redirect_info->rdi_gid;

			/*
			 * use reply_addr below, as sender_gid in reply_addr
			 * may have been set above based on some policy decision
			 * for originating end point for CM MADs above
			 */
			path_attr.pa_sgid = reply_addr->grh_hdr.ig_sender_gid;
			path_attr.pa_num_dgids = 1;
			path_attr.pa_pkey = redirect_info->rdi_pkey;

			if ((status = ibt_get_paths(ibcm_ibt_handle,
			    IBT_PATH_PKEY, &path_attr, 1, &path, NULL)) !=
			    IBT_SUCCESS)
				return (status);

			/* Initialize cm_adds */
			cm_adds = &path.pi_prim_cep_path.cep_adds_vect;
			*cm_pkt_lt = path.pi_prim_pkt_lt;
		}

	} else	{ /* cm_pkey initialized in ibt_open_rc_channel */
		reply_addr->rcvd_addr.ia_q_key = IB_GSI_QKEY;
		reply_addr->rcvd_addr.ia_remote_qno = 1;
		*cm_pkt_lt = chan_args->oc_path->pi_prim_pkt_lt;
		cm_adds = &(IBCM_PRIM_ADDS_VECT(chan_args));
	}


	cm_grh = cm_adds->av_send_grh;
	reply_addr->grh_exists = cm_grh;

	reply_addr->rcvd_addr.ia_remote_lid =
	    cm_adds->av_dlid;
	reply_addr->grh_hdr.ig_recver_gid =
	    cm_adds->av_dgid;
	reply_addr->grh_hdr.ig_flow_label =
	    cm_adds->av_flow & IB_GRH_FLOW_LABEL_MASK;
	reply_addr->grh_hdr.ig_tclass =
	    (cm_grh == B_TRUE) ? cm_adds->av_tclass : 0;
	reply_addr->grh_hdr.ig_hop_limit =
	    (cm_grh == B_TRUE) ? cm_adds->av_hop : 1;
	reply_addr->rcvd_addr.ia_service_level =
	    cm_adds->av_srvl;

	_NOTE(NOW_VISIBLE_TO_OTHER_THREADS(*reply_addr))

	return (IBT_SUCCESS);
}


/*
 * ibt_prime_close_rc_channel()
 *	It allocates resources required for close channel operation, so
 *	ibt_close_rc_channel can be called from interrupt routine.
 *
 * INPUTS:
 *	channel			The address of an ibt_channel_t struct that
 *				specifies the channel to open.
 *
 * RETURN VALUES:
 *	IBT_SUCCESS	on success(or respective failure on error)
 *
 * Clients are typically expected to call this function in established state
 */
ibt_status_t
ibt_prime_close_rc_channel(ibt_channel_hdl_t channel)
{
	ibcm_state_data_t	*statep;
	ibt_status_t		status = IBT_SUCCESS;

	IBTF_DPRINTF_L3(cmlog, "ibt_prime_close_rc_channel(%p)", channel);

	/* validate channel, first */
	if (IBCM_INVALID_CHANNEL(channel)) {
		IBTF_DPRINTF_L2(cmlog, "ibt_prime_close_rc_channel: chan 0x%p "
		    "invalid channel", channel);
		return (IBT_CHAN_HDL_INVALID);
	}

	if (ibtl_cm_get_chan_type(channel) != IBT_RC_SRV) {
		IBTF_DPRINTF_L2(cmlog, "ibt_prime_close_rc_channel: chan 0x%p "
		    "Invalid Channel type: Applicable only to RC Channel",
		    channel);
		return (IBT_CHAN_SRV_TYPE_INVALID);
	}

	/* get the statep */
	IBCM_GET_CHAN_PRIVATE(channel, statep);

	/*
	 * This can happen, if the statep is already gone by a DREQ from
	 * the remote side
	 */

	if (statep == NULL) {
		IBTF_DPRINTF_L2(cmlog, "ibt_prime_close_rc_channel: chan 0x%p "
		    "statep NULL", channel);
		return (IBT_SUCCESS);
	}

	mutex_enter(&statep->state_mutex);
	IBCM_RELEASE_CHAN_PRIVATE(channel);
	if (statep->state != IBCM_STATE_ESTABLISHED) {
		mutex_exit(&statep->state_mutex);
		return (IBT_CHAN_STATE_INVALID);
	}
	IBCM_REF_CNT_INCR(statep);
	IBTF_DPRINTF_L4(cmlog, "ibt_prime_close_rc_channel: chan 0x%p statep %p"
	    " state %x", channel, statep, statep->state);
	mutex_exit(&statep->state_mutex);

	/* clients could pre-allocate dreq mad, even before connection est */
	if (statep->dreq_msg == NULL)
		status = ibcm_alloc_out_msg(statep->stored_reply_addr.ibmf_hdl,
		    &statep->dreq_msg, MAD_METHOD_SEND);

	mutex_enter(&statep->state_mutex);
	IBCM_REF_CNT_DECR(statep);
	mutex_exit(&statep->state_mutex);

	if (status != IBT_SUCCESS) {
		IBTF_DPRINTF_L2(cmlog, "ibt_prime_close_rc_channel: chan 0x%p "
		    "ibcm_alloc_out_msg failed ", channel);
		return (status);
	}

	/* If this message isn't seen then ibt_prime_close_rc_channel failed */
	IBTF_DPRINTF_L5(cmlog, "ibt_prime_close_rc_channel: chan 0x%p done",
	    channel);

	return (IBT_SUCCESS);
}

/*
 * ibt_close_rc_channel()
 *	It closes an established channel.
 *
 * RETURN VALUES:
 *	IBT_SUCCESS	on success(or respective failure on error)
 */
ibt_status_t
ibt_close_rc_channel(ibt_channel_hdl_t channel, ibt_execution_mode_t mode,
    void *priv_data, ibt_priv_data_len_t priv_data_len, uint8_t *ret_status,
    void *ret_priv_data, ibt_priv_data_len_t *ret_priv_data_len_p)
{
	ibcm_state_data_t	*statep;

	IBTF_DPRINTF_L3(cmlog, "ibt_close_rc_channel(%p, %x, %p, %d, %p)",
	    channel, mode, priv_data, priv_data_len,
	    (ret_priv_data_len_p == NULL) ? 0 : *ret_priv_data_len_p);

	/* validate channel, first */
	if (IBCM_INVALID_CHANNEL(channel)) {
		IBTF_DPRINTF_L2(cmlog, "ibt_close_rc_channel: chan 0x%p "
		    "invalid channel", channel);
		return (IBT_CHAN_HDL_INVALID);
	}

	if (ibtl_cm_get_chan_type(channel) != IBT_RC_SRV) {
		IBTF_DPRINTF_L2(cmlog, "ibt_close_rc_channel: chan 0x%p "
		    "Invalid Channel type: Applicable only to RC Channel",
		    channel);
		return (IBT_CHAN_SRV_TYPE_INVALID);
	}

	if (mode == IBT_BLOCKING) {
		/* valid only for BLOCKING MODE */
		if ((ret_priv_data_len_p != NULL) &&
		    (*ret_priv_data_len_p > IBT_DREP_PRIV_DATA_SZ)) {
			IBTF_DPRINTF_L2(cmlog, "ibt_close_rc_channel: chan 0x%p"
			    " private data len %d is too large", channel,
			    *ret_priv_data_len_p);
			return (IBT_INVALID_PARAM);
		}
	} else if ((mode != IBT_NONBLOCKING) && (mode != IBT_NOCALLBACKS)) {
		IBTF_DPRINTF_L2(cmlog, "ibt_close_rc_channel: chan 0x%p "
		    "invalid mode %x specified", channel, mode);
		return (IBT_INVALID_PARAM);
	}

	if (ibtl_cm_is_chan_closing(channel) ||
	    ibtl_cm_is_chan_closed(channel)) {
		if (ret_status)
			*ret_status = IBT_CM_CLOSED_ALREADY;

		/* No private data to return to the client */
		if (ret_priv_data_len_p != NULL)
			*ret_priv_data_len_p = 0;

		if ((mode == IBT_BLOCKING) ||
		    (mode == IBT_NOCALLBACKS)) {
			IBCM_GET_CHAN_PRIVATE(channel, statep);
			if (statep == NULL)
				return (IBT_SUCCESS);
			mutex_enter(&statep->state_mutex);
			IBCM_RELEASE_CHAN_PRIVATE(channel);
			IBCM_REF_CNT_INCR(statep);
			while (statep->close_done != B_TRUE)
				cv_wait(&statep->block_client_cv,
				    &statep->state_mutex);
			IBCM_REF_CNT_DECR(statep);
			mutex_exit(&statep->state_mutex);
		}

		IBTF_DPRINTF_L3(cmlog, "ibt_close_rc_channel: chan 0x%p "
		    "already marked for closing", channel);

		return (IBT_SUCCESS);
	}

	/* get the statep */
	IBCM_GET_CHAN_PRIVATE(channel, statep);
	if (statep == NULL) {
		IBTF_DPRINTF_L2(cmlog, "ibt_close_rc_channel: chan 0x%p "
		    "statep NULL", channel);
		return (IBT_CHAN_STATE_INVALID);
	}

	mutex_enter(&statep->state_mutex);

	if (statep->dreq_msg == NULL) {
		IBTF_DPRINTF_L2(cmlog, "ibt_close_rc_channel: chan 0x%p "
		    "Fatal Error: dreq_msg is NULL", channel);
		IBCM_RELEASE_CHAN_PRIVATE(channel);
		mutex_exit(&statep->state_mutex);
		return (IBT_CHAN_STATE_INVALID);
	}

	if ((ret_priv_data == NULL) || (ret_priv_data_len_p == NULL)) {
		statep->close_ret_priv_data = NULL;
		statep->close_ret_priv_data_len = NULL;
	} else {
		statep->close_ret_priv_data = ret_priv_data;
		statep->close_ret_priv_data_len = ret_priv_data_len_p;
	}

	priv_data_len = min(priv_data_len, IBT_DREQ_PRIV_DATA_SZ);
	if ((priv_data != NULL) && (priv_data_len > 0)) {
		bcopy(priv_data, ((ibcm_dreq_msg_t *)
		    IBCM_OUT_MSGP(statep->dreq_msg))->dreq_private_data,
		    priv_data_len);
	}
	statep->close_ret_status = ret_status;

	IBCM_RELEASE_CHAN_PRIVATE(channel);
	IBCM_REF_CNT_INCR(statep);

	if (mode != IBT_NONBLOCKING) {
		return (ibcm_close_rc_channel(channel, statep, mode));
	}

	/* IBT_NONBLOCKING */
	ibcm_close_enqueue(statep);
	mutex_exit(&statep->state_mutex);

	return (IBT_SUCCESS);
}

void
ibcm_close_start(ibcm_state_data_t *statep)
{
	mutex_enter(&statep->state_mutex);
	(void) ibcm_close_rc_channel(statep->channel, statep, IBT_NONBLOCKING);
}

static
ibt_status_t
ibcm_close_rc_channel(ibt_channel_hdl_t channel, ibcm_state_data_t *statep,
    ibt_execution_mode_t mode)
{
	ibcm_hca_info_t		*hcap;

	_NOTE(LOCK_RELEASED_AS_SIDE_EFFECT(&statep->state_mutex));
	ASSERT(MUTEX_HELD(&statep->state_mutex));

	IBTF_DPRINTF_L3(cmlog, "ibcm_close_rc_channel: chan 0x%p statep %p",
	    channel, statep);

	hcap = statep->hcap;

	/* HCA must have been in active state. If not, it's a client bug */
	if (!IBCM_ACCESS_HCA_OK(hcap)) {
		IBTF_DPRINTF_L2(cmlog, "ibcm_close_rc_channel: chan 0x%p "
		    "hcap 0x%p not active", channel, hcap);
		IBCM_REF_CNT_DECR(statep);
		mutex_exit(&statep->state_mutex);
		return (IBT_CHAN_HDL_INVALID);
	}

	if (statep->state == IBCM_STATE_TRANSIENT_ESTABLISHED) {
		while (statep->cep_in_rts == IBCM_BLOCK)
			cv_wait(&statep->block_mad_cv, &statep->state_mutex);
	}

	/* Do TRANSIENT_DREQ check after TRANSIENT_ESTABLISHED check */
	while (statep->state == IBCM_STATE_TRANSIENT_DREQ_SENT)
		cv_wait(&statep->block_mad_cv, &statep->state_mutex);

	IBTF_DPRINTF_L4(cmlog, "ibcm_close_rc_channel: chan 0x%p "
	    "connection state is %x", channel, statep->state);

	/* If state is in pre-established states, abort the connection est */
	if (statep->state != IBCM_STATE_ESTABLISHED) {
		statep->cm_retries++;	/* ensure connection trace is dumped */

		/* No DREP private data possible */
		if (statep->close_ret_priv_data_len != NULL)
			*statep->close_ret_priv_data_len = 0;

		/*
		 * If waiting for a response mad, then cancel the timer,
		 * and delete the connection
		 */
		if (statep->state == IBCM_STATE_REQ_SENT ||
		    statep->state == IBCM_STATE_REP_SENT ||
		    statep->state == IBCM_STATE_REP_WAIT ||
		    statep->state == IBCM_STATE_MRA_REP_RCVD) {
			timeout_id_t		timer_val = statep->timerid;
			ibcm_conn_state_t	old_state;

			IBTF_DPRINTF_L4(cmlog, "ibcm_close_rc_channel: "
			    "chan 0x%p connection aborted in state %x", channel,
			    statep->state);

			old_state = statep->state;
			statep->state = IBCM_STATE_DELETE;

			if (mode == IBT_NONBLOCKING) {
				if (taskq_dispatch(ibcm_taskq,
				    ibcm_process_abort_via_taskq, statep,
				    TQ_NOSLEEP) == TASKQID_INVALID) {

					IBCM_REF_CNT_DECR(statep);
					statep->state = old_state;
					mutex_exit(&statep->state_mutex);
					return (IBT_INSUFF_KERNEL_RESOURCE);
				}	/* if taskq_dispatch succeeds */
				/* Cancel the timer */
				statep->timerid = 0;
				mutex_exit(&statep->state_mutex);
			} else {
				/* Cancel the timer */
				statep->timerid = 0;
				mutex_exit(&statep->state_mutex);
				(void) taskq_dispatch(ibcm_taskq,
				    ibcm_process_abort_via_taskq, statep,
				    TQ_SLEEP);
			}

			/* cancel the currently running timer */
			if (timer_val != 0)
				(void) untimeout(timer_val);

			/* wait until cm handler returns for BLOCKING cases */
			mutex_enter(&statep->state_mutex);
			if ((mode == IBT_BLOCKING) ||
			    (mode == IBT_NOCALLBACKS)) {
				while (statep->close_done != B_TRUE)
					cv_wait(&statep->block_client_cv,
					    &statep->state_mutex);
			}

			if (statep->close_ret_status)
				*statep->close_ret_status = IBT_CM_CLOSED_ABORT;
			mutex_exit(&statep->state_mutex);

			/*
			 * It would ideal to post a REJ MAD, but that would
			 * be non-conformance to spec. Hence, delete the state
			 * data. Assuming that happens quickly, any retransmits
			 * from the remote are replied by CM with reject
			 * reason " no valid com id". That would stop remote
			 * sending any more MADs.
			 */
			ibcm_delete_state_data(statep);
			return (IBT_SUCCESS);

		/* if CM busy in cm handler, wait until cm handler returns */
		} else if (statep->state == IBCM_STATE_REQ_RCVD ||
		    statep->state == IBCM_STATE_REP_RCVD ||
		    statep->state == IBCM_STATE_MRA_SENT ||
		    statep->state == IBCM_STATE_MRA_REP_SENT) {

			/* take control of statep */
			statep->abort_flag |= IBCM_ABORT_CLIENT;

			IBTF_DPRINTF_L4(cmlog, "ibcm_close_rc_channel: "
			    "chan 0x%p connection aborted in state = %x",
			    channel, statep->state);

			/*
			 * wait until state machine modifies qp state to error,
			 * including disassociating statep and QP
			 */
			if ((mode == IBT_BLOCKING) || (mode == IBT_NOCALLBACKS))
				while (statep->close_done != B_TRUE)
					cv_wait(&statep->block_client_cv,
					    &statep->state_mutex);

			/* a sanity setting */
			if (mode == IBT_NOCALLBACKS)
				statep->cm_handler = NULL;
			IBCM_REF_CNT_DECR(statep);

			/*
			 * In rare situations, connection attempt could be
			 * terminated for some other reason, before abort is
			 * processed, but CM still returns ret_status as abort
			 */
			if (statep->close_ret_status)
				*statep->close_ret_status = IBT_CM_CLOSED_ABORT;
			mutex_exit(&statep->state_mutex);

			/*
			 * REJ MAD is posted by the CM state machine for this
			 * case, hence state structure is deleted in the
			 * state machine processing.
			 */
			return (IBT_SUCCESS);

		} else if ((statep->state == IBCM_STATE_TIMEWAIT) ||
		    (statep->state == IBCM_STATE_DELETE)) {

			/* State already in timewait, so no return priv data */
			IBCM_REF_CNT_DECR(statep);

			/* The teardown has already been done */
			if (statep->close_ret_status)
				*statep->close_ret_status =
				    IBT_CM_CLOSED_ALREADY;
			mutex_exit(&statep->state_mutex);

			return (IBT_SUCCESS);

		} else if ((statep->state == IBCM_STATE_DREQ_RCVD) ||
		    (statep->state == IBCM_STATE_DREQ_SENT) ||
		    (statep->state == IBCM_STATE_DREP_RCVD) ||
		    ((statep->state == IBCM_STATE_TIMED_OUT) &&
		    (statep->timedout_state == IBCM_STATE_DREQ_SENT))) {

			/*
			 * Either the remote or local client has already
			 * initiated the teardown.  IBCM_STATE_DREP_RCVD is
			 * possible, if CM initiated teardown without client's
			 * knowledge, for stale handling, etc.,
			 */
			if (mode == IBT_NOCALLBACKS) {
				if (statep->close_nocb_state == IBCM_UNBLOCK) {
					statep->close_nocb_state = IBCM_FAIL;
					/* enable free qp after return */
					ibtl_cm_chan_is_closing(
					    statep->channel);
				} else while (statep->close_nocb_state ==
				    IBCM_BLOCK)
					cv_wait(&statep->block_client_cv,
					    &statep->state_mutex);
				statep->cm_handler = NULL; /* sanity setting */
				if (statep->close_ret_status)
					*statep->close_ret_status =
					    IBT_CM_CLOSED_ALREADY;
			} else if (mode == IBT_BLOCKING) {
				/* wait until state is moved to timewait */
				while (statep->close_done != B_TRUE)
					cv_wait(&statep->block_client_cv,
					    &statep->state_mutex);
			}

			IBCM_REF_CNT_DECR(statep);
			mutex_exit(&statep->state_mutex);

			/* ret_status is set in state machine code */
			return (IBT_SUCCESS);

		} else if (statep->state == IBCM_STATE_TIMED_OUT) {

			if ((mode == IBT_BLOCKING) ||
			    (mode == IBT_NOCALLBACKS)) {

				/*
				 * wait until cm handler invocation and
				 * disassociation between statep and channel
				 * is complete
				 */
				while (statep->close_done != B_TRUE)
					cv_wait(&statep->block_client_cv,
					    &statep->state_mutex);
			}

			if (statep->close_ret_status)
				*statep->close_ret_status = IBT_CM_CLOSED_ABORT;
			IBCM_REF_CNT_DECR(statep);
			mutex_exit(&statep->state_mutex);

			return (IBT_SUCCESS);
		} else {
			IBCM_REF_CNT_DECR(statep);
			mutex_exit(&statep->state_mutex);

			return (IBT_CM_FAILURE);
		}
	}

	ASSERT(statep->close_nocb_state != IBCM_BLOCK);

	if (mode == IBT_NOCALLBACKS) {
		statep->close_nocb_state = IBCM_FAIL;
		statep->cm_handler = NULL;
		ibtl_cm_chan_is_closing(statep->channel);
		IBTF_DPRINTF_L4(cmlog, "ibcm_close_rc_channel: "
		    "NOCALLBACKS on in statep = %p", statep);
	}

	if (statep->state != IBCM_STATE_ESTABLISHED) {
		goto lost_race;
	}

	/*
	 * Cancel/wait for any pending ibt_set_alt_path, and
	 * release state mutex
	 */
	ibcm_sync_lapr_idle(statep);

	ibcm_close_enter();

	mutex_enter(&statep->state_mutex);
	if (statep->state != IBCM_STATE_ESTABLISHED) {
		ibcm_close_exit();
		goto lost_race;
	}

	statep->state = IBCM_STATE_TRANSIENT_DREQ_SENT;
	statep->timerid = 0;
	statep->close_done = B_FALSE;
	statep->close_flow = 1;
	mutex_exit(&statep->state_mutex);

	ibcm_post_dreq_mad(statep);

	mutex_enter(&statep->state_mutex);

lost_race:
	if (mode == IBT_BLOCKING) {

		/* wait for DREP */
		while (statep->close_done != B_TRUE)
			cv_wait(&statep->block_client_cv,
			    &statep->state_mutex);

		IBTF_DPRINTF_L4(cmlog, "ibcm_close_rc_channel: chan 0x%p "
		    "done blocking", channel);
	}

	IBCM_REF_CNT_DECR(statep);
	mutex_exit(&statep->state_mutex);

	/* If this message isn't seen then ibt_close_rc_channel failed */
	IBTF_DPRINTF_L5(cmlog, "ibcm_close_rc_channel: chan 0x%p done",
	    channel);

	return (IBT_SUCCESS);
}

ibt_status_t
ibt_recycle_rc(ibt_channel_hdl_t rc_chan, ibt_cep_flags_t control,
    uint8_t hca_port_num, ibt_recycle_handler_t func, void *arg)
{
	ibcm_state_data_t		*statep;
	ibcm_taskq_recycle_arg_t	*ibcm_tq_recycle_arg;
	ibt_qp_query_attr_t		qp_attr;
	ibt_status_t			retval;

	IBTF_DPRINTF_L3(cmlog, "ibt_recycle_rc (%p, 0x%X, %d, %p, %p)", rc_chan,
	    control, hca_port_num, func, arg);

	if (IBCM_INVALID_CHANNEL(rc_chan)) {
		IBTF_DPRINTF_L2(cmlog, "ibt_recycle_rc: invalid channel");
		return (IBT_CHAN_HDL_INVALID);
	}

	/* check qp state */
	retval = ibt_query_qp(rc_chan, &qp_attr);

	if (retval != IBT_SUCCESS)
		return (retval);

	if (qp_attr.qp_info.qp_trans != IBT_RC_SRV)
		return (IBT_CHAN_SRV_TYPE_INVALID);

	if (qp_attr.qp_info.qp_state != IBT_STATE_ERROR)
		return (IBT_CHAN_STATE_INVALID);

	_NOTE(NOW_INVISIBLE_TO_OTHER_THREADS(*ibcm_tq_recycle_arg))

	ibcm_tq_recycle_arg = kmem_alloc(sizeof (ibcm_taskq_recycle_arg_t),
	    KM_SLEEP);

	ibcm_tq_recycle_arg->rc_chan		= rc_chan;
	ibcm_tq_recycle_arg->control		= control;
	ibcm_tq_recycle_arg->hca_port_num	= hca_port_num;
	ibcm_tq_recycle_arg->func		= func;
	ibcm_tq_recycle_arg->arg		= arg;

	_NOTE(NOW_VISIBLE_TO_OTHER_THREADS(*ibcm_tq_recycle_arg))

	IBCM_GET_CHAN_PRIVATE(rc_chan, statep);

	/*
	 * If non-blocking ie., func specified and channel has not yet completed
	 * the timewait, then schedule the work for later
	 */
	if ((func != NULL) && (statep != NULL)) {
		IBCM_RELEASE_CHAN_PRIVATE(rc_chan);
		_NOTE(NOW_INVISIBLE_TO_OTHER_THREADS(statep->recycle_arg))
		statep->recycle_arg = ibcm_tq_recycle_arg;
		_NOTE(NOW_VISIBLE_TO_OTHER_THREADS(statep->recycle_arg))
		return (IBT_SUCCESS);
	}

	/*
	 * if blocking ie., func specified, and channel has not yet completed
	 * the timewait, then block until the channel completes the timewait
	 */
	if (statep != NULL)
		IBCM_RELEASE_CHAN_PRIVATE(rc_chan);
	IBCM_WAIT_CHAN_PRIVATE(rc_chan);

	if (func) {	/* NON BLOCKING case. Taskq for QP state change */
		(void) taskq_dispatch(ibcm_taskq, ibcm_process_rc_recycle,
		    ibcm_tq_recycle_arg, TQ_SLEEP);
		return (IBT_SUCCESS);
	} else	/* BLOCKING case */
		return (ibcm_process_rc_recycle_ret(ibcm_tq_recycle_arg));
}

void
ibcm_process_rc_recycle(void *recycle_arg)
{
	(void) ibcm_process_rc_recycle_ret(recycle_arg);
}

static ibt_status_t
ibcm_process_rc_recycle_ret(void *recycle_arg)
{
	ibt_qp_info_t			qp_info;
	ibt_status_t			ibt_status = IBT_SUCCESS;
	ibt_cep_modify_flags_t		cep_flags;
	ibt_qp_query_attr_t		qp_attr;
	ibcm_taskq_recycle_arg_t	*ibcm_tq_recycle_arg =
	    (ibcm_taskq_recycle_arg_t *)recycle_arg;

	/* QP must have been in error state */
	ibt_status = ibt_query_qp(ibcm_tq_recycle_arg->rc_chan, &qp_attr);
	if (ibt_status != IBT_SUCCESS)
		IBTF_DPRINTF_L2(cmlog, "ibcm_process_rc_recycle_ret: "
		    "chanp %p ibt_query_qp() = %d",
		    ibcm_tq_recycle_arg->rc_chan, ibt_status);
	else {
		/* perform the QP state change from ERROR to RESET */
		bzero(&qp_info, sizeof (qp_info));

		qp_info.qp_trans = IBT_RC_SRV;
		qp_info.qp_state = IBT_STATE_RESET;

		/* Call modify_qp to move to RESET state */
		ibt_status = ibt_modify_qp(ibcm_tq_recycle_arg->rc_chan,
		    IBT_CEP_SET_STATE, &qp_info, NULL);

		if (ibt_status != IBT_SUCCESS)
			IBTF_DPRINTF_L2(cmlog, "ibcm_process_rc_recycle_ret: "
			    "chanp %p ibt_modify_qp() = %d for ERROR to RESET",
			    ibcm_tq_recycle_arg->rc_chan, ibt_status);
	}

	if (ibt_status == IBT_SUCCESS) {

		qp_info.qp_state = IBT_STATE_INIT;

		/* set flags for all mandatory args from RESET to INIT */
		cep_flags = IBT_CEP_SET_STATE | IBT_CEP_SET_PORT;
		cep_flags |= IBT_CEP_SET_RDMA_R | IBT_CEP_SET_RDMA_W;
		cep_flags |= IBT_CEP_SET_ATOMIC;

		qp_info.qp_transport.rc.rc_path.cep_hca_port_num =
		    ibcm_tq_recycle_arg->hca_port_num;
		qp_info.qp_flags |=
		    ibcm_tq_recycle_arg->control & IBT_CEP_RDMA_RD;
		qp_info.qp_flags |=
		    ibcm_tq_recycle_arg->control & IBT_CEP_RDMA_WR;
		qp_info.qp_flags |=
		    ibcm_tq_recycle_arg->control & IBT_CEP_ATOMIC;

		/* Always use the existing pkey */
		qp_info.qp_transport.rc.rc_path.cep_pkey_ix =
		    qp_attr. qp_info.qp_transport.rc.rc_path.cep_pkey_ix;

		/* Call modify_qp to move to INIT state */
		ibt_status = ibt_modify_qp(ibcm_tq_recycle_arg->rc_chan,
		    cep_flags, &qp_info, NULL);

		if (ibt_status != IBT_SUCCESS)
			IBTF_DPRINTF_L2(cmlog, "ibcm_process_rc_recycle_ret: "
			    "chanp %p ibt_modify_qp() = %d for RESET to INIT",
			    ibcm_tq_recycle_arg->rc_chan, ibt_status);
	}

	/* Change the QP CM state to indicate QP being re-used */
	if (ibt_status == IBT_SUCCESS)
		ibtl_cm_chan_is_reused(ibcm_tq_recycle_arg->rc_chan);

	/* Call func, if defined */
	if (ibcm_tq_recycle_arg->func)
		(*(ibcm_tq_recycle_arg->func))(ibt_status,
		    ibcm_tq_recycle_arg->arg);

	kmem_free(ibcm_tq_recycle_arg, sizeof (ibcm_taskq_recycle_arg_t));

	return (ibt_status);
}

static void
ibcm_process_abort_via_taskq(void *args)
{
	ibcm_state_data_t	*statep = (ibcm_state_data_t *)args;

	ibcm_process_abort(statep);
	mutex_enter(&statep->state_mutex);
	IBCM_REF_CNT_DECR(statep);
	mutex_exit(&statep->state_mutex);
}

/*
 * Local UD CM Handler's private data, used during ibt_request_ud_dest() in
 * Non-Blocking mode operations.
 */
typedef struct ibcm_local_handler_s {
	ibt_cm_ud_handler_t	actual_cm_handler;
	void			*actual_cm_private;
	ibt_ud_dest_t		*dest_hdl;
} ibcm_local_handler_t;

_NOTE(READ_ONLY_DATA(ibcm_local_handler_s))

/*
 * Local UD CM Handler, used when ibt_alloc_ud_dest() is issued in
 * NON-Blocking mode.
 *
 * Out here, we update the UD Destination handle with
 * the obtained DQPN and QKey (from SIDR REP) and invokes actual client
 * handler that was specified by the client.
 */
static ibt_cm_status_t
ibcm_local_cm_handler(void *priv, ibt_cm_ud_event_t *event,
    ibt_cm_ud_return_args_t *ret_args, void *priv_data, ibt_priv_data_len_t len)
{
	ibcm_local_handler_t	*handler_priv = (ibcm_local_handler_t *)priv;

	IBTF_DPRINTF_L4(cmlog, "ibcm_local_cm_handler: event %d",
	    event->cm_type);

	ASSERT(handler_priv != NULL);

	switch (event->cm_type) {
	case IBT_CM_UD_EVENT_SIDR_REP:
		/* Update QPN & QKey from event into destination handle. */
		if (handler_priv->dest_hdl != NULL) {
			handler_priv->dest_hdl->ud_dst_qpn =
			    event->cm_event.sidr_rep.srep_remote_qpn;
			handler_priv->dest_hdl->ud_qkey =
			    event->cm_event.sidr_rep.srep_remote_qkey;
		}

		/* Invoke the client handler - inform only, so ignore retval */
		(void) handler_priv->actual_cm_handler(
		    handler_priv->actual_cm_private, event, ret_args, priv_data,
		    len);

		/* Free memory allocated for local handler's private data. */
		if (handler_priv != NULL)
			kmem_free(handler_priv, sizeof (*handler_priv));

		break;
	default:
		IBTF_DPRINTF_L2(cmlog, "ibcm_local_cm_handler: ERROR");
		break;
	}

	return (IBT_CM_ACCEPT);
}


/* Validate the input UD destination attributes.  */
static ibt_status_t
ibcm_validate_dqpn_data(ibt_ud_dest_attr_t *attr, ibt_execution_mode_t mode,
    ibt_ud_returns_t *ret_args)
{
	/* cm handler must always be specified */
	if (mode == IBT_NONBLOCKING && attr->ud_cm_handler == NULL) {
		IBTF_DPRINTF_L2(cmlog, "ibcm_validate_dqpn_data: "
		    "CM handler is not specified ");
		return (IBT_INVALID_PARAM);
	}

	if (mode == IBT_NONBLOCKING) {
		if (ret_args != NULL) {
			IBTF_DPRINTF_L2(cmlog, "ibcm_validate_dqpn_data: "
			    "ret_args should be NULL when called in "
			    "non-blocking mode");
			return (IBT_INVALID_PARAM);
		}
	} else if (mode == IBT_BLOCKING) {
		if (ret_args == NULL) {
			IBTF_DPRINTF_L2(cmlog, "ibcm_validate_dqpn_data: "
			    "ret_args should be Non-NULL when called in "
			    "blocking mode");
			return (IBT_INVALID_PARAM);
		}
	} else {
		IBTF_DPRINTF_L2(cmlog, "ibcm_validate_dqpn_data: "
		    "invalid mode %x specified ", mode);
		return (IBT_INVALID_PARAM);
	}

	if (attr->ud_sid == 0) {
		IBTF_DPRINTF_L2(cmlog, "ibcm_validate_dqpn_data: "
		    "ServiceID must be specified. ");
		return (IBT_INVALID_PARAM);
	}

	if (attr->ud_addr == NULL) {
		IBTF_DPRINTF_L2(cmlog, "ibcm_validate_dqpn_data: "
		    "Address Info NULL");
		return (IBT_INVALID_PARAM);
	}

	/* Validate SGID */
	if ((attr->ud_addr->av_sgid.gid_prefix == 0) ||
	    (attr->ud_addr->av_sgid.gid_guid == 0)) {
		IBTF_DPRINTF_L2(cmlog, "ibcm_validate_dqpn_data: Invalid SGID");
		return (IBT_INVALID_PARAM);
	}
	IBTF_DPRINTF_L3(cmlog, "ibcm_validate_dqpn_data: SGID<%llX:%llX>",
	    attr->ud_addr->av_sgid.gid_prefix,
	    attr->ud_addr->av_sgid.gid_guid);

	/* Validate DGID */
	if ((attr->ud_addr->av_dgid.gid_prefix == 0) ||
	    (attr->ud_addr->av_dgid.gid_guid == 0)) {
		IBTF_DPRINTF_L2(cmlog, "ibcm_validate_dqpn_data: Invalid DGID");
		return (IBT_INVALID_PARAM);
	}
	IBTF_DPRINTF_L3(cmlog, "ibcm_validate_dqpn_data: DGID<%llX:%llX>",
	    attr->ud_addr->av_dgid.gid_prefix,
	    attr->ud_addr->av_dgid.gid_guid);

	return (IBT_SUCCESS);
}


/* Perform SIDR to retrieve DQPN and QKey.  */
static ibt_status_t
ibcm_ud_get_dqpn(ibt_ud_dest_attr_t *attr, ibt_execution_mode_t mode,
    ibt_ud_returns_t *ret_args)
{
	ibt_status_t		retval;
	ib_pkey_t		ud_pkey;
	ibmf_handle_t		ibmf_hdl;
	ibmf_msg_t		*ibmf_msg;
	ibcm_hca_info_t		*hcap;
	ibcm_sidr_req_msg_t	*sidr_req_msgp;
	ibcm_ud_state_data_t	*ud_statep;
	ibtl_cm_hca_port_t	port;
	ibcm_sidr_srch_t	sidr_entry;
	ibcm_qp_list_t		*cm_qp_entry;

	/* Retrieve HCA GUID value from the available SGID info. */
	retval = ibtl_cm_get_hca_port(attr->ud_addr->av_sgid, 0, &port);
	if ((retval != IBT_SUCCESS) || (port.hp_port == 0)) {
		IBTF_DPRINTF_L2(cmlog, "ibcm_ud_get_dqpn: "
		    "ibtl_cm_get_hca_port failed: %d", retval);
		return (retval);
	}

	IBTF_DPRINTF_L4(cmlog, "ibcm_ud_get_dqpn: "
	    "HCA GUID:%llX, port_num:%d", port.hp_hca_guid, port.hp_port);

	/* Lookup the HCA info for this GUID */
	if ((hcap = ibcm_find_hca_entry(port.hp_hca_guid)) == NULL) {
		IBTF_DPRINTF_L2(cmlog, "ibcm_ud_get_dqpn: hcap is NULL");
		return (IBT_HCA_INVALID);
	}

	/* Return failure if the HCA device or Port is not operational */

	if ((retval = ibt_get_port_state_byguid(port.hp_hca_guid, port.hp_port,
	    NULL, NULL)) != IBT_SUCCESS) {
		/* Device Port is not in good state, don't use it. */
		IBTF_DPRINTF_L2(cmlog, "ibcm_ud_get_dqpn: Invalid "
		    "port specified or port not active");
		ibcm_dec_hca_acc_cnt(hcap);
		return (retval);
	}

	retval = ibt_index2pkey_byguid(port.hp_hca_guid, port.hp_port,
	    attr->ud_pkey_ix, &ud_pkey);
	if (retval != IBT_SUCCESS) {
		IBTF_DPRINTF_L2(cmlog, "ibcm_ud_get_dqpn: "
		    "Failed to convert index2pkey: %d", retval);
		ibcm_dec_hca_acc_cnt(hcap);
		return (retval);
	}

	_NOTE(NOW_INVISIBLE_TO_OTHER_THREADS(sidr_entry))

	/* Allocate a new request id */
	if (ibcm_alloc_reqid(hcap, &sidr_entry.srch_req_id) == IBCM_FAILURE) {
		IBTF_DPRINTF_L2(cmlog, "ibcm_ud_get_dqpn: "
		    "no req id available");
		ibcm_dec_hca_acc_cnt(hcap);
		return (IBT_INSUFF_KERNEL_RESOURCE);
	}

	if ((hcap->hca_port_info[port.hp_port - 1].port_ibmf_hdl == NULL) &&
	    ((retval = ibcm_hca_reinit_port(hcap, port.hp_port - 1))
	    != IBT_SUCCESS)) {
		IBTF_DPRINTF_L2(cmlog, "ibcm_ud_get_dqpn: "
		    "ibmf reg or callback setup failed during re-initialize");
		return (retval);
	}

	ibmf_hdl = hcap->hca_port_info[port.hp_port - 1].port_ibmf_hdl;

	/* find the ibmf QP to post the SIDR REQ */
	if ((cm_qp_entry = ibcm_find_qp(hcap, port.hp_port, ud_pkey)) ==
	    NULL) {
		IBTF_DPRINTF_L2(cmlog, "ibcm_ud_get_dqpn: IBMF QP allocation"
		    " failed");
		ibcm_dec_hca_acc_cnt(hcap);
		return (IBT_INSUFF_RESOURCE);
	}

	if ((retval = ibcm_alloc_out_msg(ibmf_hdl, &ibmf_msg, MAD_METHOD_SEND))
	    != IBT_SUCCESS) {
		IBTF_DPRINTF_L2(cmlog, "ibcm_ud_get_dqpn: IBMF MSG allocation"
		    " failed");
		ibcm_release_qp(cm_qp_entry);
		ibcm_dec_hca_acc_cnt(hcap);
		return (retval);
	}

	sidr_entry.srch_lid = port.hp_base_lid;
	sidr_entry.srch_gid = attr->ud_addr->av_sgid;
	sidr_entry.srch_grh_exists = attr->ud_addr->av_send_grh;
	sidr_entry.srch_mode = IBCM_ACTIVE_MODE;

	_NOTE(NOW_VISIBLE_TO_OTHER_THREADS(sidr_entry))

	/* do various allocations needed here */
	rw_enter(&hcap->hca_sidr_list_lock, RW_WRITER);

	(void) ibcm_find_sidr_entry(&sidr_entry, hcap, &ud_statep,
	    IBCM_FLAG_ADD);
	rw_exit(&hcap->hca_sidr_list_lock);

	/* Increment hca's resource count */
	ibcm_inc_hca_res_cnt(hcap);

	/* After a resource created on hca, no need to hold the acc cnt */
	ibcm_dec_hca_acc_cnt(hcap);

	_NOTE(NOW_INVISIBLE_TO_OTHER_THREADS(*ud_statep))

	/* Initialize some ud_statep fields */
	ud_statep->ud_stored_msg = ibmf_msg;
	ud_statep->ud_svc_id = attr->ud_sid;
	ud_statep->ud_pkt_life_time =
	    ibt_ib2usec(attr->ud_pkt_lt);
	ud_statep->ud_stored_reply_addr.cm_qp_entry = cm_qp_entry;

	/* set remaining retry cnt */
	ud_statep->ud_remaining_retry_cnt = ud_statep->ud_max_cm_retries;

	/*
	 * Get UD handler and corresponding args which is pass it back
	 * as first argument for the handler.
	 */
	ud_statep->ud_state_cm_private = attr->ud_cm_private;

	if (mode == IBT_BLOCKING)
		ud_statep->ud_return_data = ret_args;
	else
		ud_statep->ud_cm_handler = attr->ud_cm_handler;

	/* Initialize the fields of ud_statep->ud_stored_reply_addr */
	ud_statep->ud_stored_reply_addr.grh_exists = attr->ud_addr->av_send_grh;
	ud_statep->ud_stored_reply_addr.ibmf_hdl = ibmf_hdl;
	ud_statep->ud_stored_reply_addr.grh_hdr.ig_hop_limit =
	    attr->ud_addr->av_hop;
	ud_statep->ud_stored_reply_addr.grh_hdr.ig_sender_gid =
	    attr->ud_addr->av_sgid;
	ud_statep->ud_stored_reply_addr.grh_hdr.ig_recver_gid =
	    attr->ud_addr->av_dgid;
	ud_statep->ud_stored_reply_addr.grh_hdr.ig_tclass =
	    attr->ud_addr->av_tclass;
	ud_statep->ud_stored_reply_addr.grh_hdr.ig_flow_label =
	    attr->ud_addr->av_flow & IB_GRH_FLOW_LABEL_MASK;

	/* needs to be derived based on the base LID and path bits */
	ud_statep->ud_stored_reply_addr.rcvd_addr.ia_local_lid =
	    port.hp_base_lid;
	ud_statep->ud_stored_reply_addr.rcvd_addr.ia_remote_lid =
	    attr->ud_addr->av_dlid;
	ud_statep->ud_stored_reply_addr.rcvd_addr.ia_p_key = ud_pkey;
	ud_statep->ud_stored_reply_addr.rcvd_addr.ia_q_key = IB_GSI_QKEY;
	ud_statep->ud_stored_reply_addr.rcvd_addr.ia_service_level =
	    attr->ud_addr->av_srvl;

	/*
	 * This may be enchanced later, to use a remote qno based on past
	 * redirect rej mad responses. This would be the place to specify
	 * appropriate remote qno
	 */
	ud_statep->ud_stored_reply_addr.rcvd_addr.ia_remote_qno = 1;

	_NOTE(NOW_INVISIBLE_TO_OTHER_THREADS(*sidr_req_msgp))

	/* Initialize the SIDR REQ message fields */
	sidr_req_msgp =
	    (ibcm_sidr_req_msg_t *)IBCM_OUT_MSGP(ud_statep->ud_stored_msg);

	sidr_req_msgp->sidr_req_request_id = h2b32(ud_statep->ud_req_id);
	sidr_req_msgp->sidr_req_service_id = h2b64(attr->ud_sid);
	sidr_req_msgp->sidr_req_pkey = h2b16(ud_pkey);
	IBCM_OUT_HDRP(ud_statep->ud_stored_msg)->AttributeID =
	    h2b16(IBCM_INCOMING_SIDR_REQ + IBCM_ATTR_BASE_ID);

	if ((attr->ud_priv_data != NULL) && (attr->ud_priv_data_len > 0)) {
		bcopy(attr->ud_priv_data, sidr_req_msgp->sidr_req_private_data,
		    min(attr->ud_priv_data_len, IBT_SIDR_REQ_PRIV_DATA_SZ));
	}

	/* Send out the SIDR REQ message */
	ud_statep->ud_state = IBCM_STATE_SIDR_REQ_SENT;
	ud_statep->ud_timer_stored_state = IBCM_STATE_SIDR_REQ_SENT;
	IBCM_UD_REF_CNT_INCR(ud_statep); /* for non-blocking SIDR REQ post */
	ud_statep->ud_timer_value = ibt_ib2usec(ibcm_max_sidr_rep_proctime) +
	    (ud_statep->ud_pkt_life_time * 2);

	IBCM_OUT_HDRP(ud_statep->ud_stored_msg)->TransactionID =
	    h2b64(ibcm_generate_tranid(IBCM_INCOMING_SIDR_REQ,
	    ud_statep->ud_req_id, 0));

	IBTF_DPRINTF_L4(cmlog, "ibcm_ud_get_dqpn: timer_value in HZ = %x",
	    ud_statep->ud_timer_value);

	_NOTE(NOW_VISIBLE_TO_OTHER_THREADS(*ud_statep))
	_NOTE(NOW_VISIBLE_TO_OTHER_THREADS(*sidr_req_msgp))

	ibcm_post_ud_mad(ud_statep, ud_statep->ud_stored_msg,
	    ibcm_post_sidr_req_complete, ud_statep);

	mutex_enter(&ud_statep->ud_state_mutex);

	/* Wait for SIDR_REP */
	if (mode == IBT_BLOCKING) {
		IBTF_DPRINTF_L4(cmlog, "ibcm_ud_get_dqpn: blocking");

		while (ud_statep->ud_blocking_done != B_TRUE) {
			cv_wait(&ud_statep->ud_block_client_cv,
			    &ud_statep->ud_state_mutex);
		}

		IBTF_DPRINTF_L4(cmlog, "ibcm_ud_get_dqpn: finished blocking");

		if (ret_args->ud_status == IBT_CM_SREP_QPN_VALID) {
			IBTF_DPRINTF_L4(cmlog, "ibcm_ud_get_dqpn: DQPN = %x, "
			    "status = %x, QKey = %x", ret_args->ud_dqpn,
			    ret_args->ud_status, ret_args->ud_qkey);

		} else {
			IBTF_DPRINTF_L4(cmlog, "ibcm_ud_get_dqpn: Status<%x>",
			    ret_args->ud_status);
			retval = IBT_CM_FAILURE;
		}
	}

	IBCM_UD_REF_CNT_DECR(ud_statep);
	mutex_exit(&ud_statep->ud_state_mutex);

	IBTF_DPRINTF_L4(cmlog, "ibcm_ud_get_dqpn: done");

	return (retval);
}


/*
 * Function:
 *	ibt_request_ud_dest
 * Input:
 *	ud_dest		A previously allocated UD destination handle.
 *	mode		This function can execute in blocking or non blocking
 *			modes.
 *	attr		UD destination attributes to be modified.
 * Output:
 *	ud_ret_args	If the function is called in blocking mode, ud_ret_args
 *			should be a pointer to an ibt_ud_returns_t struct.
 * Returns:
 *	IBT_SUCCESS
 * Description:
 *	Modify a previously allocated UD destination handle based on the
 *	results of doing the SIDR protocol.
 */
ibt_status_t
ibt_request_ud_dest(ibt_ud_dest_hdl_t ud_dest, ibt_execution_mode_t mode,
    ibt_ud_dest_attr_t *attr, ibt_ud_returns_t *ud_ret_args)
{
	ibt_status_t		retval;
	ibt_ud_dest_t		*ud_destp;
	ibcm_local_handler_t	*local_handler_priv = NULL;

	IBTF_DPRINTF_L3(cmlog, "ibt_request_ud_dest(%p, %x, %p, %p)",
	    ud_dest, mode, attr, ud_ret_args);

	retval = ibcm_validate_dqpn_data(attr, mode, ud_ret_args);
	if (retval != IBT_SUCCESS) {
		return (retval);
	}

	ud_destp = ud_dest;

	/* Allocate an Address handle. */
	retval = ibt_modify_ah(ud_destp->ud_dest_hca, ud_destp->ud_ah,
	    attr->ud_addr);
	if (retval != IBT_SUCCESS) {
		IBTF_DPRINTF_L2(cmlog, "ibt_request_ud_dest: "
		    "Address Handle Modification failed: %d", retval);
		return (retval);
	}

	if (mode == IBT_NONBLOCKING) {
		/*
		 * In NON-BLOCKING mode, and we need to update the destination
		 * handle with the DQPN and QKey that are obtained from
		 * SIDR REP, hook-up our own handler, so that we can catch
		 * the event, and we ourselves call the actual client's
		 * ud_cm_handler, in our handler.
		 */

		/* Allocate memory for local handler's private data. */
		local_handler_priv =
		    kmem_alloc(sizeof (*local_handler_priv), KM_SLEEP);

		_NOTE(NOW_INVISIBLE_TO_OTHER_THREADS(*local_handler_priv))

		local_handler_priv->actual_cm_handler = attr->ud_cm_handler;
		local_handler_priv->actual_cm_private = attr->ud_cm_private;
		local_handler_priv->dest_hdl = ud_destp;

		_NOTE(NOW_VISIBLE_TO_OTHER_THREADS(*local_handler_priv))

		attr->ud_cm_handler = ibcm_local_cm_handler;
		attr->ud_cm_private = local_handler_priv;
	}

	/* In order to get DQPN and Destination QKey, perform SIDR */
	retval = ibcm_ud_get_dqpn(attr, mode, ud_ret_args);
	if (retval != IBT_SUCCESS) {
		IBTF_DPRINTF_L2(cmlog, "ibt_request_ud_dest: "
		    "Failed to get DQPN: %d", retval);

		/* Free memory allocated for local handler's private data. */
		if (local_handler_priv != NULL)
			kmem_free(local_handler_priv,
			    sizeof (*local_handler_priv));
		return (retval);
	}

	/*
	 * Fill in the dqpn and dqkey as obtained from ud_ret_args,
	 * values will be valid only on BLOCKING mode.
	 */
	if (mode == IBT_BLOCKING) {
		ud_destp->ud_dst_qpn = ud_ret_args->ud_dqpn;
		ud_destp->ud_qkey = ud_ret_args->ud_qkey;
	}

	return (retval);
}

/*
 * Function:
 *	ibt_ud_get_dqpn
 * Input:
 *	attr		A pointer to an ibt_ud_dest_attr_t struct that are
 *			required for SIDR REQ message. Not specified attributes
 *			should be set to "NULL" or "0".
 *			ud_sid, ud_addr and ud_pkt_lt must be specified.
 *	mode		This function can execute in blocking or non blocking
 *			modes.
 * Output:
 *	returns		If the function is called in blocking mode, returns
 *			should be a pointer to an ibt_ud_returns_t struct.
 * Return:
 *	IBT_SUCCESS	on success or respective failure on error.
 * Description:
 *	Finds the destination QPN at the specified destination that the
 *	specified service can be reached on. The IBTF CM initiates the
 *	service ID resolution protocol (SIDR) to determine a destination QPN.
 *
 * NOTE: SIDR_REQ is initiated from active side.
 */
ibt_status_t
ibt_ud_get_dqpn(ibt_ud_dest_attr_t *attr, ibt_execution_mode_t mode,
    ibt_ud_returns_t *returns)
{
	ibt_status_t		retval;

	IBTF_DPRINTF_L3(cmlog, "ibt_ud_get_dqpn(%p, %x, %p)",
	    attr, mode, returns);

	retval = ibcm_validate_dqpn_data(attr, mode, returns);
	if (retval != IBT_SUCCESS) {
		return (retval);
	}

	return (ibcm_ud_get_dqpn(attr, mode, returns));
}


/*
 * ibt_cm_delay:
 *	A client CM handler function can call this function
 *	to extend its response time to a CM event.
 * INPUTS:
 *	flags		Indicates what CM message processing is being delayed
 *			by the CM handler, valid values are:
 *				IBT_CM_DELAY_REQ
 *				IBT_CM_DELAY_REP
 *				IBT_CM_DELAY_LAP
 *	cm_session_id	The session ID that was passed to client srv_handler
 *			by the CM
 *	service_time	The extended service time
 *	priv_data	Vendor specific data to be sent in the CM generated
 *			MRA message. Should be NULL if not specified.
 *	len		The number of bytes of data specified by priv_data.
 *
 * RETURN VALUES:
 *	IBT_SUCCESS	on success (or respective failure on error)
 */
ibt_status_t
ibt_cm_delay(ibt_cmdelay_flags_t flags, void *cm_session_id,
    clock_t service_time, void *priv_data, ibt_priv_data_len_t len)
{
	uint8_t			msg_typ = 0;
	ibcm_mra_msg_t		*mra_msgp;
	ibcm_state_data_t	*statep;
	ibt_status_t		status;

	IBTF_DPRINTF_L3(cmlog, "ibt_cm_delay(0x%x, %p, 0x%x)",
	    flags, cm_session_id, service_time);

	/*
	 * Make sure channel is associated with a statep
	 */
	statep = (ibcm_state_data_t *)cm_session_id;

	if (statep == NULL) {
		IBTF_DPRINTF_L2(cmlog, "ibt_cm_delay: statep NULL");
		return (IBT_INVALID_PARAM);
	}

	IBTF_DPRINTF_L4(cmlog, "ibt_cm_delay: statep %p", statep);

	/* Allocate an ibmf msg for mra, if not allocated yet */
	if (statep->mra_msg == NULL) {
		if ((status = ibcm_alloc_out_msg(
		    statep->stored_reply_addr.ibmf_hdl, &statep->mra_msg,
		    MAD_METHOD_SEND)) != IBT_SUCCESS) {
			IBTF_DPRINTF_L2(cmlog, "ibt_cm_delay: chan 0x%p"
			    "IBMF MSG allocation failed", statep->channel);
			return (status);
		}
	}

	_NOTE(NOW_INVISIBLE_TO_OTHER_THREADS(*mra_msgp))

	mra_msgp = (ibcm_mra_msg_t *)IBCM_OUT_MSGP(statep->mra_msg);
	mra_msgp->mra_local_comm_id = h2b32(statep->local_comid);
	mra_msgp->mra_remote_comm_id = h2b32(statep->remote_comid);

	/* fill in rest of MRA's fields - Message MRAed and Service Timeout */
	if (flags == IBT_CM_DELAY_REQ) {
		msg_typ = IBT_CM_MRA_TYPE_REQ;
	} else if (flags == IBT_CM_DELAY_REP) {
		msg_typ = IBT_CM_MRA_TYPE_REP;
	} else if (flags == IBT_CM_DELAY_LAP) {
		msg_typ = IBT_CM_MRA_TYPE_LAP;
	}

	mra_msgp->mra_message_type_plus = msg_typ << 6;
	mra_msgp->mra_service_timeout_plus = ibt_usec2ib(service_time) << 3;

	len = min(len, IBT_MRA_PRIV_DATA_SZ);
	if (priv_data && (len > 0))
		bcopy(priv_data, mra_msgp->mra_private_data, len);

	IBCM_OUT_HDRP(statep->mra_msg)->AttributeID =
	    h2b16(IBCM_INCOMING_MRA + IBCM_ATTR_BASE_ID);

	_NOTE(NOW_VISIBLE_TO_OTHER_THREADS(*mra_msgp))

	mutex_enter(&statep->state_mutex);

	if ((statep->mode == IBCM_ACTIVE_MODE) &&
	    (statep->state == IBCM_STATE_REP_RCVD)) {
		statep->state = IBCM_STATE_MRA_REP_SENT;
	} else if (statep->mode == IBCM_PASSIVE_MODE) {
		if (statep->state == IBCM_STATE_REQ_RCVD) {
			statep->state = IBCM_STATE_MRA_SENT;
		} else if (statep->ap_state == IBCM_AP_STATE_LAP_RCVD) {
			statep->ap_state = IBCM_AP_STATE_MRA_LAP_RCVD;
		} else {
			IBTF_DPRINTF_L2(cmlog, "ibt_cm_delay: invalid state "
			    "/ap_state/mode %x, %x, %x", statep->state,
			    statep->ap_state, statep->mode);
			mutex_exit(&statep->state_mutex);
			return (IBT_CHAN_STATE_INVALID);
		}
	} else {
		IBTF_DPRINTF_L2(cmlog, "ibt_cm_delay: invalid state "
		    "/ap_state/mode %x, %x, %x", statep->state,
		    statep->ap_state, statep->mode);
		mutex_exit(&statep->state_mutex);

		return (IBT_CHAN_STATE_INVALID);
	}
	/* service time is usecs, stale_clock is nsecs */
	statep->stale_clock = gethrtime() +
	    (hrtime_t)ibt_ib2usec(ibt_usec2ib(service_time)) * (1000 *
	    statep->max_cm_retries);

	statep->send_mad_flags |= IBCM_MRA_POST_BUSY;
	IBCM_REF_CNT_INCR(statep);	/* for ibcm_post_mra_complete */
	mutex_exit(&statep->state_mutex);

	IBCM_OUT_HDRP(statep->mra_msg)->TransactionID =
	    IBCM_OUT_HDRP(statep->stored_msg)->TransactionID;

	/* post the MRA mad in blocking mode, as no timers involved */
	ibcm_post_rc_mad(statep, statep->mra_msg, ibcm_post_mra_complete,
	    statep);
	ibcm_insert_trace(statep, IBCM_TRACE_OUTGOING_MRA);
	/* If this message isn't seen then ibt_cm_delay failed */
	IBTF_DPRINTF_L3(cmlog, "ibt_cm_delay: done !!");

	return (IBT_SUCCESS);
}


/*
 * ibt_register_service()
 *	Register a service with the IBCM
 *
 * INPUTS:
 *	ibt_hdl		The IBT client handle returned to the client
 *			on an ibt_attach() call.
 *
 *	srv		The address of a ibt_srv_desc_t that describes
 *			the service, containing the following:
 *
 *		sd_ud_handler	The Service CM UD event Handler.
 *		sd_handler	The Service CM RC/UC/RD event Handler.
 *		sd_flags	Service flags (peer-to-peer, or not).
 *
 *	sid		This tells CM if the service is local (sid is 0) or
 *			wellknown (sid is the starting service id of the range).
 *
 *	num_sids	The number of contiguous service-ids to reserve.
 *
 *	srv_hdl		The address of a service identification handle, used
 *			to deregister a service, and to bind GIDs to.
 *
 *	ret_sid		The address to store the Service ID return value.
 *			If num_sids > 1, ret_sid is the first Service ID
 *			in the range.
 *
 * ibt_register_service() returns:
 *	IBT_SUCCESS		- added a service successfully.
 *	IBT_INVALID_PARAM	- invalid input parameter.
 *	IBT_CM_FAILURE		- failed to add the service.
 *	IBT_CM_SERVICE_EXISTS	- service already exists.
 *	IBT_INSUFF_KERNEL_RESOURCE - ran out of local service ids (should
 *				     never happen).
 */
ibt_status_t
ibt_register_service(ibt_clnt_hdl_t ibt_hdl, ibt_srv_desc_t *srv,
    ib_svc_id_t sid, int num_sids, ibt_srv_hdl_t *srv_hdl, ib_svc_id_t *ret_sid)
{
	ibcm_svc_info_t		*svcinfop;

	IBTF_DPRINTF_L2(cmlog, "ibt_register_service(%p (%s), %p, 0x%llX, %d)",
	    ibt_hdl, ibtl_cm_get_clnt_name(ibt_hdl), srv, (longlong_t)sid,
	    num_sids);

	_NOTE(NOW_INVISIBLE_TO_OTHER_THREADS(*svcinfop))

	*srv_hdl = NULL;

	if (num_sids <= 0) {
		IBTF_DPRINTF_L2(cmlog, "ibt_register_service: "
		    "Invalid number of service-ids specified (%d)", num_sids);
		return (IBT_INVALID_PARAM);
	}

	if (sid == 0) {
		if (ret_sid == NULL)
			return (IBT_INVALID_PARAM);
		sid = ibcm_alloc_local_sids(num_sids);
		if (sid == 0)
			return (IBT_INSUFF_KERNEL_RESOURCE);

	/* Make sure that the ServiceId specified is not of LOCAL AGN type. */
	} else if ((sid & IB_SID_AGN_MASK) == IB_SID_AGN_LOCAL) {
		IBTF_DPRINTF_L2(cmlog, "ibt_register_service: "
		    "Invalid non-LOCAL SID specified: 0x%llX",
		    (longlong_t)sid);
		return (IBT_INVALID_PARAM);
	}

	svcinfop = ibcm_create_svc_entry(sid, num_sids);

	if (svcinfop == NULL) {
		IBTF_DPRINTF_L2(cmlog, "ibt_register_service: "
		    "Service-ID 0x%llx already registered", (longlong_t)sid);
		return (IBT_CM_SERVICE_EXISTS);
	}

	/*
	 * 'sid' and 'num_sids' are filled in ibcm_create_svc_entry()
	 */
	svcinfop->svc_flags = srv->sd_flags;
	svcinfop->svc_rc_handler = srv->sd_handler;
	svcinfop->svc_ud_handler = srv->sd_ud_handler;

	if (ret_sid != NULL)
		*ret_sid = sid;

	*srv_hdl = svcinfop;

	ibtl_cm_change_service_cnt(ibt_hdl, num_sids);

	/* If this message isn't seen, then ibt_register_service failed. */
	IBTF_DPRINTF_L2(cmlog, "ibt_register_service: done (%p, %llX)",
	    svcinfop, sid);

	_NOTE(NOW_VISIBLE_TO_OTHER_THREADS(*svcinfop))

	return (IBT_SUCCESS);
}


static ibt_status_t
ibcm_write_service_record(ibmf_saa_handle_t saa_handle,
    sa_service_record_t *srv_recp, ibmf_saa_access_type_t saa_type)
{
	int	rval;
	int	retry;

	ibcm_sa_access_enter();
	for (retry = 0; retry < ibcm_max_sa_retries; retry++) {
		rval = ibmf_saa_update_service_record(
		    saa_handle, srv_recp, saa_type, 0);
		if (rval != IBMF_TRANS_TIMEOUT) {
			break;
		}
		IBTF_DPRINTF_L2(cmlog, "ibcm_write_service_record: "
		    "ibmf_saa_update_service_record timed out"
		    " SID = %llX, rval = %d, saa_type = %d",
		    (longlong_t)srv_recp->ServiceID, rval, saa_type);
		delay(ibcm_sa_timeout_delay);
	}
	ibcm_sa_access_exit();

	if (rval != IBMF_SUCCESS) {
		IBTF_DPRINTF_L2(cmlog, "ibcm_write_service_record: "
		    "ibmf_saa_update_service_record() : Failed - %d", rval);
		return (ibcm_ibmf_analyze_error(rval));
	} else
		return (IBT_SUCCESS);
}


static void
ibcm_rem_stale_srec(ibmf_saa_handle_t saa_handle, sa_service_record_t *srec)
{
	ibt_status_t		retval;
	uint_t			num_found;
	size_t			length;
	sa_service_record_t	*srv_resp;
	void			*results_p;
	uint_t			i;
	uint64_t		component_mask;
	ibmf_saa_access_args_t	access_args;

	component_mask =
	    SA_SR_COMPMASK_PKEY | SA_SR_COMPMASK_NAME | SA_SR_COMPMASK_GID;

	/* Call in SA Access retrieve routine to get Service Records. */
	access_args.sq_attr_id = SA_SERVICERECORD_ATTRID;
	access_args.sq_access_type = IBMF_SAA_RETRIEVE;
	access_args.sq_component_mask = component_mask;
	access_args.sq_template = srec;
	access_args.sq_template_length = sizeof (sa_service_record_t);
	access_args.sq_callback = NULL;
	access_args.sq_callback_arg = NULL;

	retval = ibcm_contact_sa_access(saa_handle, &access_args, &length,
	    &results_p);
	if (retval != IBT_SUCCESS) {
		IBTF_DPRINTF_L2(cmlog, "ibcm_rem_stale_srec: "
		    "SA Access Failure");
		return;
	}

	num_found = length / sizeof (sa_service_record_t);

	if (num_found)
		IBTF_DPRINTF_L3(cmlog, "ibcm_rem_stale_srec: "
		    "Found %d matching Service Records.", num_found);

	/* Validate the returned number of records. */
	if ((results_p != NULL) && (num_found > 0)) {

		/* Remove all the records. */
		for (i = 0; i < num_found; i++) {

			srv_resp = (sa_service_record_t *)
			    ((uchar_t *)results_p +
			    i * sizeof (sa_service_record_t));

			/*
			 * Found some matching records, but check out whether
			 * this Record is really stale or just happens to match
			 * the current session records. If yes, don't remove it.
			 */
			mutex_enter(&ibcm_svc_info_lock);
			if (ibcm_find_svc_entry(srv_resp->ServiceID) != NULL) {
				/* This record is NOT STALE. */
				mutex_exit(&ibcm_svc_info_lock);
				IBTF_DPRINTF_L3(cmlog, "ibcm_rem_stale_srec: "
				    "This is not Stale, it's an active record");
				continue;
			}
			mutex_exit(&ibcm_svc_info_lock);

			IBTF_DPRINTF_L2(cmlog, "ibcm_rem_stale_srec: "
			    "Removing Stale Rec: %s, %llX",
			    srv_resp->ServiceName, srv_resp->ServiceID);

			IBCM_DUMP_SERVICE_REC(srv_resp);

			/*
			 * Remove the Service Record Entry from SA.
			 *
			 * Get ServiceID info from Response Buf, other
			 * attributes are already filled-in.
			 */

			 _NOTE(NOW_INVISIBLE_TO_OTHER_THREADS(srec->ServiceID))

			srec->ServiceID = srv_resp->ServiceID;

			 _NOTE(NOW_VISIBLE_TO_OTHER_THREADS(srec->ServiceID))

			(void) ibcm_write_service_record(saa_handle, srec,
			    IBMF_SAA_DELETE);
		}

		/* Deallocate the memory for results_p. */
		kmem_free(results_p, length);
	}
}



/*
 * ibt_bind_service()
 *	Register a service with the IBCM
 *
 * INPUTS:
 *	srv_hdl		The service id handle returned to the client
 *			on an ibt_service_register() call.
 *
 *	gid		The GID to which to bind the service.
 *
 *	srv_bind	The address of a ibt_srv_bind_t that describes
 *			the service record.  This should be NULL if there
 *			is to be no service record.  This contains:
 *
 *		sb_lease	Lease period
 *		sb_pkey		Partition
 *		sb_name		pointer to ASCII string Service Name,
 *				NULL terminated.
 *		sb_key[]	Key to secure the service record.
 *		sb_data		Service Data structure (64-byte)
 *
 *	cm_private	First argument of Service handler.
 *
 *	sb_hdl_p	The address of a service bind handle, used
 *			to undo the service binding.
 *
 * ibt_bind_service() returns:
 *	IBT_SUCCESS		- added a service successfully.
 *	IBT_INVALID_PARAM	- invalid input parameter.
 *	IBT_CM_FAILURE		- failed to add the service.
 *	IBT_CM_SERVICE_EXISTS	- service already exists.
 */
ibt_status_t
ibt_bind_service(ibt_srv_hdl_t srv_hdl, ib_gid_t gid, ibt_srv_bind_t *srv_bind,
    void *cm_private, ibt_sbind_hdl_t *sb_hdl_p)
{
	ibt_status_t		status;
	ibtl_cm_hca_port_t	port;
	ibcm_svc_bind_t		*sbindp, *sbp;
	ibcm_hca_info_t		*hcap;
	ib_svc_id_t		sid, start_sid, end_sid;
	ibmf_saa_handle_t	saa_handle;
	sa_service_record_t	srv_rec;
	uint16_t		pkey_ix;

	if (sb_hdl_p != NULL)
		*sb_hdl_p = NULL;	/* return value for error cases */

	IBTF_DPRINTF_L2(cmlog, "ibt_bind_service: srv_hdl %p, gid (%llX:%llX)",
	    srv_hdl, (longlong_t)gid.gid_prefix, (longlong_t)gid.gid_guid);

	_NOTE(NOW_INVISIBLE_TO_OTHER_THREADS(*sbindp))

	/* Call ibtl_cm_get_hca_port to get the port number and the HCA GUID. */
	if ((status = ibtl_cm_get_hca_port(gid, 0, &port)) != IBT_SUCCESS) {
		IBTF_DPRINTF_L2(cmlog, "ibt_bind_service: "
		    "ibtl_cm_get_hca_port failed: %d", status);
		return (status);
	}
	IBTF_DPRINTF_L4(cmlog, "ibt_bind_service: Port:%d HCA GUID:%llX",
	    port.hp_port, port.hp_hca_guid);

	hcap = ibcm_find_hca_entry(port.hp_hca_guid);
	if (hcap == NULL) {
		IBTF_DPRINTF_L2(cmlog, "ibt_bind_service: NO HCA found");
		return (IBT_HCA_BUSY_DETACHING);
	}
	IBTF_DPRINTF_L4(cmlog, "ibt_bind_service: hcap = %p", hcap);

	if (srv_bind != NULL) {
		saa_handle = ibcm_get_saa_handle(hcap, port.hp_port);
		if (saa_handle == NULL) {
			IBTF_DPRINTF_L2(cmlog, "ibt_bind_service: "
			    "saa_handle is NULL");
			ibcm_dec_hca_acc_cnt(hcap);
			return (IBT_HCA_PORT_NOT_ACTIVE);
		}
		if (srv_bind->sb_pkey == 0) {
			IBTF_DPRINTF_L2(cmlog, "ibt_bind_service: "
			    "P_Key must not be 0");
			ibcm_dec_hca_acc_cnt(hcap);
			return (IBT_INVALID_PARAM);
		}
		if (strlen(srv_bind->sb_name) >= IB_SVC_NAME_LEN) {
			IBTF_DPRINTF_L2(cmlog, "ibt_bind_service: "
			    "Service Name is too long");
			ibcm_dec_hca_acc_cnt(hcap);
			return (IBT_INVALID_PARAM);
		} else
			IBTF_DPRINTF_L3(cmlog, "ibt_bind_service: "
			    "Service Name='%s'", srv_bind->sb_name);
		status = ibt_pkey2index_byguid(port.hp_hca_guid,
		    port.hp_port, srv_bind->sb_pkey, &pkey_ix);
		if (status != IBT_SUCCESS) {
			IBTF_DPRINTF_L2(cmlog, "ibt_bind_service: "
			    "P_Key 0x%x not found in P_Key_Table",
			    srv_bind->sb_pkey);
			ibcm_dec_hca_acc_cnt(hcap);
			return (status);
		}
	}

	/* assume success - allocate before locking */
	sbindp = kmem_zalloc(sizeof (*sbindp), KM_SLEEP);
	sbindp->sbind_cm_private = cm_private;
	sbindp->sbind_gid = gid;
	sbindp->sbind_hcaguid = port.hp_hca_guid;
	sbindp->sbind_port = port.hp_port;

	mutex_enter(&ibcm_svc_info_lock);

	sbp = srv_hdl->svc_bind_list;
	while (sbp != NULL) {
		if (sbp->sbind_gid.gid_guid == gid.gid_guid &&
		    sbp->sbind_gid.gid_prefix == gid.gid_prefix) {
			if (srv_bind == NULL ||
			    srv_bind->sb_pkey == sbp->sbind_pkey) {
				IBTF_DPRINTF_L2(cmlog, "ibt_bind_service: "
				    "failed: GID %llX:%llX and PKEY %x is "
				    "already bound", gid.gid_prefix,
				    gid.gid_guid, sbp->sbind_pkey);
				mutex_exit(&ibcm_svc_info_lock);
				ibcm_dec_hca_acc_cnt(hcap);
				kmem_free(sbindp, sizeof (*sbindp));
				return (IBT_CM_SERVICE_EXISTS);
			}
		}
		sbp = sbp->sbind_link;
	}
	/* no entry found */

	sbindp->sbind_link = srv_hdl->svc_bind_list;
	srv_hdl->svc_bind_list = sbindp;

	mutex_exit(&ibcm_svc_info_lock);

	if (srv_bind != NULL) {
		bzero(&srv_rec, sizeof (srv_rec));

		srv_rec.ServiceLease =
		    sbindp->sbind_lease = srv_bind->sb_lease;
		srv_rec.ServiceP_Key =
		    sbindp->sbind_pkey = srv_bind->sb_pkey;
		srv_rec.ServiceKey_hi =
		    sbindp->sbind_key[0] = srv_bind->sb_key[0];
		srv_rec.ServiceKey_lo =
		    sbindp->sbind_key[1] = srv_bind->sb_key[1];
		(void) strcpy(sbindp->sbind_name, srv_bind->sb_name);
		(void) strcpy((char *)srv_rec.ServiceName, srv_bind->sb_name);
		srv_rec.ServiceGID = gid;

		/*
		 * Find out whether we have any stale Local Service records
		 * matching the current attributes.  If yes, we shall try to
		 * remove them from SA using the current request's ServiceKey.
		 *
		 * We will perform this operation only for Local Services, as
		 * it is handled by SA automatically for WellKnown Services.
		 *
		 * Ofcourse, clients can specify NOT to do this clean-up by
		 * setting IBT_SBIND_NO_CLEANUP flag (srv_bind->sb_flag).
		 */
		if ((srv_hdl->svc_id & IB_SID_AGN_LOCAL) &&
		    (!(srv_bind->sb_flag & IBT_SBIND_NO_CLEANUP))) {
			ibcm_rem_stale_srec(saa_handle, &srv_rec);
		}

		/* Handle endianess for service data. */
		ibcm_swizzle_from_srv(&srv_bind->sb_data, sbindp->sbind_data);

		bcopy(sbindp->sbind_data, srv_rec.ServiceData, IB_SVC_DATA_LEN);

		/* insert srv record into the SA */
		start_sid = srv_hdl->svc_id;
		end_sid = start_sid + srv_hdl->svc_num_sids - 1;
		for (sid = start_sid; sid <= end_sid; sid++) {

			srv_rec.ServiceID = sid;

			IBCM_DUMP_SERVICE_REC(&srv_rec);

			IBTF_DPRINTF_L4(cmlog, "ibt_bind_service: "
			    "ibmf_saa_write_service_record, SvcId = %llX",
			    (longlong_t)sid);

			status = ibcm_write_service_record(saa_handle, &srv_rec,
			    IBMF_SAA_UPDATE);
			if (status != IBT_SUCCESS) {
				IBTF_DPRINTF_L2(cmlog, "ibt_bind_service:"
				    " ibcm_write_service_record fails %d, "
				    "sid %llX", status, (longlong_t)sid);

				if (sid != start_sid) {
					/*
					 * Bind failed while bind SID other than
					 * first in the sid_range.  So we need
					 * to unbind those, which are passed.
					 *
					 * Need to increment svc count to
					 * compensate for ibt_unbind_service().
					 */
					ibcm_inc_hca_svc_cnt(hcap);
					ibcm_dec_hca_acc_cnt(hcap);

					(void) ibt_unbind_service(srv_hdl,
					    sbindp);
				} else {
					ibcm_svc_bind_t		**sbpp;

					/*
					 * Bind failed for the first SID or the
					 * only SID in question, then no need
					 * to unbind, just free memory and
					 * return error.
					 */
					mutex_enter(&ibcm_svc_info_lock);

					sbpp = &srv_hdl->svc_bind_list;
					sbp = *sbpp;
					while (sbp != NULL) {
						if (sbp == sbindp) {
							*sbpp = sbp->sbind_link;
							break;
						}
						sbpp = &sbp->sbind_link;
						sbp = *sbpp;
					}
					mutex_exit(&ibcm_svc_info_lock);
					ibcm_dec_hca_acc_cnt(hcap);

					kmem_free(sbindp, sizeof (*sbindp));
				}
				return (status);
			}
		}
	}
	ibcm_inc_hca_svc_cnt(hcap);
	ibcm_dec_hca_acc_cnt(hcap);

	/* If this message isn't seen then ibt_bind_service failed */
	IBTF_DPRINTF_L2(cmlog, "ibt_bind_service: DONE (%p, %llX:%llX)",
	    srv_hdl, gid.gid_prefix, gid.gid_guid);

	if (sb_hdl_p != NULL)
		*sb_hdl_p = sbindp;

	_NOTE(NOW_VISIBLE_TO_OTHER_THREADS(*sbindp))

	return (IBT_SUCCESS);
}

ibt_status_t
ibt_unbind_service(ibt_srv_hdl_t srv_hdl, ibt_sbind_hdl_t sbindp)
{
	ib_svc_id_t	sid, end_sid;
	ibt_status_t	rval;
	ibcm_hca_info_t	*hcap;
	ibcm_svc_bind_t	*sbp, **sbpp;

	IBTF_DPRINTF_L2(cmlog, "ibt_unbind_service(%p, %p)",
	    srv_hdl, sbindp);

	hcap = ibcm_find_hca_entry(sbindp->sbind_hcaguid);

	/* If there is a service on hca, respective hcap cannot go away */
	ASSERT(hcap != NULL);

	mutex_enter(&ibcm_svc_info_lock);

	sbpp = &srv_hdl->svc_bind_list;
	sbp = *sbpp;
	while (sbp != NULL) {
		if (sbp == sbindp) {
			*sbpp = sbp->sbind_link;
			break;
		}
		sbpp = &sbp->sbind_link;
		sbp = *sbpp;
	}
	sid = srv_hdl->svc_id;
	end_sid = srv_hdl->svc_id + srv_hdl->svc_num_sids - 1;
	if (sbp != NULL)
		while (sbp->sbind_rewrite_state == IBCM_REWRITE_BUSY)
			cv_wait(&ibcm_svc_info_cv, &ibcm_svc_info_lock);
	mutex_exit(&ibcm_svc_info_lock);

	if (sbp == NULL) {
		IBTF_DPRINTF_L2(cmlog, "ibt_unbind_service: "
		    "service binding not found: srv_hdl %p, srv_bind %p",
		    srv_hdl, sbindp);
		ibcm_dec_hca_acc_cnt(hcap);
		return (IBT_INVALID_PARAM);
	}

	if (sbindp->sbind_pkey != 0) {	/* Are there service records? */
		ibtl_cm_hca_port_t	port;
		sa_service_record_t	srv_rec;
		ibmf_saa_handle_t	saa_handle;
		ibt_status_t		status;

		/* get the default SGID of the port */
		if ((status = ibtl_cm_get_hca_port(sbindp->sbind_gid, 0, &port))
		    != IBT_SUCCESS) {
			IBTF_DPRINTF_L2(cmlog, "ibt_unbind_service: "
			    "ibtl_cm_get_hca_port failed: %d", status);
			/* we're done, but there may be stale service records */
			goto done;
		}

		saa_handle = ibcm_get_saa_handle(hcap, port.hp_port);
		if (saa_handle == NULL) {
			IBTF_DPRINTF_L2(cmlog, "ibt_unbind_service: "
			    "saa_handle is NULL");
			/* we're done, but there may be stale service records */
			goto done;
		}

		/* Fill in fields of srv_rec */
		bzero(&srv_rec, sizeof (srv_rec));

		srv_rec.ServiceP_Key = sbindp->sbind_pkey;
		srv_rec.ServiceKey_hi = sbindp->sbind_key[0];
		srv_rec.ServiceKey_lo = sbindp->sbind_key[1];
		srv_rec.ServiceGID = sbindp->sbind_gid;
		(void) strcpy((char *)srv_rec.ServiceName, sbindp->sbind_name);

		while (sid <= end_sid) {

			srv_rec.ServiceID = sid;
			IBCM_DUMP_SERVICE_REC(&srv_rec);

			rval = ibcm_write_service_record(saa_handle, &srv_rec,
			    IBMF_SAA_DELETE);

			IBTF_DPRINTF_L4(cmlog, "ibt_unbind_service: "
			    "ibcm_write_service_record rval = %d, SID %llx",
			    rval, sid);
			if (rval != IBT_SUCCESS) {
				/* this is not considered a reason to fail */
				IBTF_DPRINTF_L2(cmlog, "ibt_unbind_service: "
				    "ibcm_write_service_record fails %d, "
				    "sid %llx", rval, sid);
			}
			sid++;
		}
	}
done:
	ibcm_dec_hca_svc_cnt(hcap);
	ibcm_dec_hca_acc_cnt(hcap);
	kmem_free(sbindp, sizeof (*sbindp));

	/* If this message isn't seen then ibt_unbind_service failed */
	IBTF_DPRINTF_L2(cmlog, "ibt_unbind_service: done !!");

	return (IBT_SUCCESS);
}

/*
 * Simply pull off each binding from the list and unbind it.
 * If any of the unbind calls fail, we fail.
 */
ibt_status_t
ibt_unbind_all_services(ibt_srv_hdl_t srv_hdl)
{
	ibt_status_t	status;
	ibcm_svc_bind_t	*sbp;

	mutex_enter(&ibcm_svc_info_lock);
	sbp = NULL;

	/* this compare keeps the loop from being infinite */
	while (sbp != srv_hdl->svc_bind_list) {
		sbp = srv_hdl->svc_bind_list;
		mutex_exit(&ibcm_svc_info_lock);
		status = ibt_unbind_service(srv_hdl, sbp);
		if (status != IBT_SUCCESS)
			return (status);
		mutex_enter(&ibcm_svc_info_lock);
		if (srv_hdl->svc_bind_list == NULL)
			break;
	}
	mutex_exit(&ibcm_svc_info_lock);
	return (IBT_SUCCESS);
}

/*
 * ibt_deregister_service()
 *	Deregister a service with the IBCM
 *
 * INPUTS:
 *	ibt_hdl		The IBT client handle returned to the client
 *			on an ibt_attach() call.
 *
 *	srv_hdl		The address of a service identification handle, used
 *			to de-register a service.
 * RETURN VALUES:
 *	IBT_SUCCESS	on success (or respective failure on error)
 */
ibt_status_t
ibt_deregister_service(ibt_clnt_hdl_t ibt_hdl, ibt_srv_hdl_t srv_hdl)
{
	ibcm_svc_info_t		*svcp;
	ibcm_svc_lookup_t	svc;

	IBTF_DPRINTF_L2(cmlog, "ibt_deregister_service(%p (%s), %p)",
	    ibt_hdl, ibtl_cm_get_clnt_name(ibt_hdl), srv_hdl);

	mutex_enter(&ibcm_svc_info_lock);

	if (srv_hdl->svc_bind_list != NULL) {
		IBTF_DPRINTF_L2(cmlog, "ibt_deregister_service:"
		    " srv_hdl %p still has bindings", srv_hdl);
		mutex_exit(&ibcm_svc_info_lock);
		return (IBT_CM_SERVICE_BUSY);
	}
	svc.sid = srv_hdl->svc_id;
	svc.num_sids = 1;
	IBTF_DPRINTF_L3(cmlog, "ibt_deregister_service: SID 0x%llX, numsids %d",
	    srv_hdl->svc_id, srv_hdl->svc_num_sids);

#ifdef __lock_lint
	ibcm_svc_compare(NULL, NULL);
#endif
	svcp = avl_find(&ibcm_svc_avl_tree, &svc, NULL);
	if (svcp != srv_hdl) {
		mutex_exit(&ibcm_svc_info_lock);
		IBTF_DPRINTF_L2(cmlog, "ibt_deregister_service(): "
		    "srv_hdl %p not found", srv_hdl);
		return (IBT_INVALID_PARAM);
	}
	avl_remove(&ibcm_svc_avl_tree, svcp);

	/* wait for active REQ/SREQ handling to be done */
	svcp->svc_to_delete = 1;
	while (svcp->svc_ref_cnt != 0)
		cv_wait(&ibcm_svc_info_cv, &ibcm_svc_info_lock);

	mutex_exit(&ibcm_svc_info_lock);

	if ((srv_hdl->svc_id & IB_SID_AGN_MASK) == IB_SID_AGN_LOCAL)
		ibcm_free_local_sids(srv_hdl->svc_id, srv_hdl->svc_num_sids);

	ibtl_cm_change_service_cnt(ibt_hdl, -srv_hdl->svc_num_sids);
	kmem_free(srv_hdl, sizeof (*srv_hdl));

	/* If this message isn't seen then ibt_deregister_service failed */
	IBTF_DPRINTF_L2(cmlog, "ibt_deregister_service: done !!");

	return (IBT_SUCCESS);
}

ibcm_status_t
ibcm_ar_init(void)
{
	ib_svc_id_t	sid = IBCM_DAPL_ATS_SID;
	ibcm_svc_info_t *tmp_svcp;

	IBTF_DPRINTF_L3(cmlog, "ibcm_ar_init()");

	/* remove this special SID from the pool of available SIDs */
	if ((tmp_svcp = ibcm_create_svc_entry(sid, 1)) == NULL) {
		IBTF_DPRINTF_L3(cmlog, "ibcm_ar_init: "
		    "DAPL ATS SID 0x%llx already registered", (longlong_t)sid);
		return (IBCM_FAILURE);
	}
	mutex_enter(&ibcm_svc_info_lock);
	ibcm_ar_svcinfop = tmp_svcp;
	ibcm_ar_list = NULL;	/* no address records registered yet */
	mutex_exit(&ibcm_svc_info_lock);
	return (IBCM_SUCCESS);
}

ibcm_status_t
ibcm_ar_fini(void)
{
	ibcm_ar_t	*ar_list;
	ibcm_svc_info_t	*tmp_svcp;

	mutex_enter(&ibcm_svc_info_lock);
	ar_list = ibcm_ar_list;

	if (ar_list == NULL &&
	    avl_numnodes(&ibcm_svc_avl_tree) == 1 &&
	    avl_first(&ibcm_svc_avl_tree) == ibcm_ar_svcinfop) {
		avl_remove(&ibcm_svc_avl_tree, ibcm_ar_svcinfop);
		tmp_svcp = ibcm_ar_svcinfop;
		mutex_exit(&ibcm_svc_info_lock);
		kmem_free(tmp_svcp, sizeof (*ibcm_ar_svcinfop));
		return (IBCM_SUCCESS);
	}
	mutex_exit(&ibcm_svc_info_lock);
	return (IBCM_FAILURE);
}


/*
 * Return to the caller:
 *	IBT_SUCCESS		Found a perfect match.
 *				*arpp is set to the record.
 *	IBT_INCONSISTENT_AR	Found a record that's inconsistent.
 *	IBT_AR_NOT_REGISTERED	Found no record with same GID/pkey and
 *				found no record with same data.
 */
static ibt_status_t
ibcm_search_ar(ibt_ar_t *arp, ibcm_ar_t **arpp)
{
	ibcm_ar_t	*tmp;
	int		i;

	ASSERT(MUTEX_HELD(&ibcm_svc_info_lock));
	tmp = ibcm_ar_list;
	while (tmp != NULL) {
		if (tmp->ar.ar_gid.gid_prefix == arp->ar_gid.gid_prefix &&
		    tmp->ar.ar_gid.gid_guid == arp->ar_gid.gid_guid &&
		    tmp->ar.ar_pkey == arp->ar_pkey) {
			for (i = 0; i < IBCM_DAPL_ATS_NBYTES; i++)
				if (tmp->ar.ar_data[i] != arp->ar_data[i])
					return (IBT_INCONSISTENT_AR);
			*arpp = tmp;
			return (IBT_SUCCESS);
		} else {
			/* if all the data bytes match, we have inconsistency */
			for (i = 0; i < IBCM_DAPL_ATS_NBYTES; i++)
				if (tmp->ar.ar_data[i] != arp->ar_data[i])
					break;
			if (i == IBCM_DAPL_ATS_NBYTES)
				return (IBT_INCONSISTENT_AR);
			/* try next address record */
		}
		tmp = tmp->ar_link;
	}
	return (IBT_AR_NOT_REGISTERED);
}

ibt_status_t
ibt_register_ar(ibt_clnt_hdl_t ibt_hdl, ibt_ar_t *arp)
{
	ibcm_ar_t		*found;
	ibcm_ar_t		*tmp;
	ibt_status_t		status;
	ibt_status_t		s1, s2;
	char			*s;
	ibcm_ar_ref_t		*hdlp;
	ibcm_ar_t		*new;
	ibcm_ar_t		**linkp;
	ibtl_cm_hca_port_t	cm_port;
	uint16_t		pkey_ix;
	ibcm_hca_info_t		*hcap;
	ibmf_saa_handle_t	saa_handle;
	sa_service_record_t	*srv_recp;
	uint64_t		gid_ored;

	IBTF_DPRINTF_L3(cmlog, "ibt_register_ar: PKey 0x%X GID %llX:%llX",
	    arp->ar_pkey, (longlong_t)arp->ar_gid.gid_prefix,
	    (longlong_t)arp->ar_gid.gid_guid);

	/*
	 * If P_Key is 0, but GID is not, this query is invalid.
	 * If GID is 0, but P_Key is not, this query is invalid.
	 */
	gid_ored = arp->ar_gid.gid_guid | arp->ar_gid.gid_prefix;
	if ((arp->ar_pkey == 0 && gid_ored != 0ULL) ||
	    (arp->ar_pkey != 0 && gid_ored == 0ULL)) {
		IBTF_DPRINTF_L2(cmlog, "ibt_register_ar: "
		    "GID/P_Key is not valid");
		return (IBT_INVALID_PARAM);
	}

	/* assume success, so these might be needed */
	hdlp = kmem_alloc(sizeof (*hdlp), KM_SLEEP);
	new = kmem_zalloc(sizeof (*new), KM_SLEEP);

	mutex_enter(&ibcm_svc_info_lock);
	/* search for existing GID/pkey (there can be at most 1) */
	status = ibcm_search_ar(arp, &found);
	if (status == IBT_INCONSISTENT_AR) {
		mutex_exit(&ibcm_svc_info_lock);
		kmem_free(new, sizeof (*new));
		kmem_free(hdlp, sizeof (*hdlp));
		IBTF_DPRINTF_L2(cmlog, "ibt_register_ar: "
		    "address record is inconsistent with a known one");
		return (IBT_INCONSISTENT_AR);
	} else if (status == IBT_SUCCESS) {
		if (found->ar_flags == IBCM_AR_INITING) {
			found->ar_waiters++;
			cv_wait(&found->ar_cv, &ibcm_svc_info_lock);
			found->ar_waiters--;
		}
		if (found->ar_flags == IBCM_AR_FAILED) {
			if (found->ar_waiters == 0) {
				cv_destroy(&found->ar_cv);
				kmem_free(found, sizeof (*found));
			}
			mutex_exit(&ibcm_svc_info_lock);
			kmem_free(new, sizeof (*new));
			kmem_free(hdlp, sizeof (*hdlp));
			return (ibt_get_module_failure(IBT_FAILURE_IBCM, 0));
		}
		hdlp->ar_ibt_hdl = ibt_hdl;
		hdlp->ar_ref_link = found->ar_ibt_hdl_list;
		found->ar_ibt_hdl_list =