mgt/ibcm/ibcm_impl.c

	ibcm_impl.c revision 39b361b2ebefcef5612a54ae5cbd2179e19be296
/*
 * 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 2008 Sun Microsystems, Inc.  All rights reserved.
 * Use is subject to license terms.
 */

/*
 * ibcm_impl.c
 *
 * contains internal functions of IB CM module.
 *
 * TBD:
 * 1. HCA CATASTROPHIC/RECOVERED not handled yet
 */

#include <sys/ib/mgt/ibcm/ibcm_impl.h>
#include <sys/disp.h>


/* function prototypes */
static ibcm_status_t    ibcm_init(void);
static ibcm_status_t    ibcm_fini(void);

/* Routines to initialize and destory CM global locks and CVs */
static void     ibcm_init_locks(void);
static void     ibcm_fini_locks(void);

/* Routines that initialize/teardown CM's global hca structures */
static void     ibcm_init_hcas();
static ibcm_status_t    ibcm_fini_hcas();

static void     ibcm_init_classportinfo();
static void     ibcm_stop_timeout_thread();

/* Routines that handle HCA attach/detach asyncs */
static void     ibcm_hca_attach(ib_guid_t);
static ibcm_status_t    ibcm_hca_detach(ibcm_hca_info_t *);

/* Routines that initialize the HCA's port related fields */
static ibt_status_t ibcm_hca_init_port(ibcm_hca_info_t *hcap,
                uint8_t port_index);
static ibcm_status_t    ibcm_hca_fini_port(ibcm_hca_info_t *hcap,
                uint8_t port_index);

static void ibcm_rc_flow_control_init(void);
static void ibcm_rc_flow_control_fini(void);

/*
 * Routines that check if hca's avl trees and sidr lists are free of any
 * active client resources ie., RC or UD state structures in certain states
 */
static ibcm_status_t    ibcm_check_avl_clean(ibcm_hca_info_t *hcap);
static ibcm_status_t    ibcm_check_sidr_clean(ibcm_hca_info_t *hcap);

/* Add a new hca structure to CM's global hca list */
static ibcm_hca_info_t  *ibcm_add_hca_entry(ib_guid_t hcaguid, uint_t nports);

static void     ibcm_comm_est_handler(ibt_async_event_t *);
void            ibcm_async_handler(void *, ibt_hca_hdl_t,
                ibt_async_code_t, ibt_async_event_t *);

/* Global variables */
char            cmlog[] = "ibcm";   /* for debug log messages */
ibt_clnt_hdl_t      ibcm_ibt_handle;    /* IBT handle */
kmutex_t        ibcm_svc_info_lock; /* list lock */
kcondvar_t      ibcm_svc_info_cv;   /* cv for deregister */
kmutex_t        ibcm_recv_mutex;
avl_tree_t      ibcm_svc_avl_tree;
taskq_t         *ibcm_taskq = NULL;
int         taskq_dispatch_fail_cnt;

kmutex_t        ibcm_trace_mutex;   /* Trace mutex */
kmutex_t        ibcm_trace_print_mutex; /* Trace print mutex */
int         ibcm_conn_max_trcnt = IBCM_MAX_CONN_TRCNT;

int         ibcm_enable_trace = 2;  /* Trace level 4 by default */
int         ibcm_dtrace = 0; /* conditionally enable more dtrace */

_NOTE(MUTEX_PROTECTS_DATA(ibcm_svc_info_lock, ibcm_svc_info_s::{svc_bind_list
    svc_ref_cnt svc_to_delete}))

_NOTE(MUTEX_PROTECTS_DATA(ibcm_svc_info_lock, ibcm_svc_bind_s::{sbind_link}))

_NOTE(MUTEX_PROTECTS_DATA(ibcm_trace_mutex, ibcm_conn_trace_s))

_NOTE(DATA_READABLE_WITHOUT_LOCK(ibcm_conn_trace_s))

_NOTE(MUTEX_PROTECTS_DATA(ibcm_trace_print_mutex, ibcm_debug_buf))

_NOTE(DATA_READABLE_WITHOUT_LOCK(ibcm_debug_buf))

/*
 * Initial state is INIT. All hca dr's return success immediately in this
 * state, without adding or deleting any hca's to CM.
 */
ibcm_finit_state_t  ibcm_finit_state = IBCM_FINIT_INIT;

/* mutex and cv to manage hca's reference and resource count(s) */
kmutex_t        ibcm_global_hca_lock;
kcondvar_t      ibcm_global_hca_cv;

/* mutex and cv to sa session open */
kmutex_t        ibcm_sa_open_lock;
kcondvar_t      ibcm_sa_open_cv;
int         ibcm_sa_timeout_delay = 1;      /* in ticks */
_NOTE(MUTEX_PROTECTS_DATA(ibcm_sa_open_lock,
    ibcm_port_info_s::{port_ibmf_saa_hdl port_saa_open_in_progress}))

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

/* serialize sm notice callbacks */
kmutex_t        ibcm_sm_notice_serialize_lock;

_NOTE(LOCK_ORDER(ibcm_sm_notice_serialize_lock ibcm_global_hca_lock))

_NOTE(MUTEX_PROTECTS_DATA(ibcm_global_hca_lock, ibcm_hca_info_s::{hca_state
    hca_svc_cnt hca_acc_cnt hca_res_cnt hca_next}))

_NOTE(MUTEX_PROTECTS_DATA(ibcm_global_hca_lock,
    ibcm_port_info_s::{port_ibmf_hdl}))

_NOTE(MUTEX_PROTECTS_DATA(ibcm_sm_notice_serialize_lock,
    ibcm_port_info_s::{port_event_status}))

_NOTE(DATA_READABLE_WITHOUT_LOCK(ibcm_hca_info_s::{hca_state}))
_NOTE(DATA_READABLE_WITHOUT_LOCK(
    ibcm_hca_info_s::{hca_port_info.port_ibmf_hdl}))

/* mutex for CM's qp list management */
kmutex_t        ibcm_qp_list_lock;

_NOTE(MUTEX_PROTECTS_DATA(ibcm_qp_list_lock, ibcm_port_info_s::{port_qplist}))
_NOTE(MUTEX_PROTECTS_DATA(ibcm_qp_list_lock, ibcm_qp_list_s))
_NOTE(MUTEX_PROTECTS_DATA(ibcm_qp_list_lock, ibcm_qp_list_s))

kcondvar_t      ibcm_timeout_list_cv;
kcondvar_t      ibcm_timeout_thread_done_cv;
kt_did_t        ibcm_timeout_thread_did;
ibcm_state_data_t   *ibcm_timeout_list_hdr, *ibcm_timeout_list_tail;
ibcm_ud_state_data_t    *ibcm_ud_timeout_list_hdr, *ibcm_ud_timeout_list_tail;
kmutex_t        ibcm_timeout_list_lock;
uint8_t         ibcm_timeout_list_flags = 0;
pri_t           ibcm_timeout_thread_pri = MINCLSYSPRI;

_NOTE(MUTEX_PROTECTS_DATA(ibcm_timeout_list_lock,
    ibcm_state_data_s::timeout_next))

_NOTE(MUTEX_PROTECTS_DATA(ibcm_timeout_list_lock,
    ibcm_ud_state_data_s::ud_timeout_next))

/*
 * Flow control logic for open_rc_channel uses the following.
 */

struct ibcm_open_s {
    kmutex_t        mutex;
    kcondvar_t      cv;
    uint8_t         task_running;
    uint_t          queued;
    uint_t          exit_deferred;
    uint_t          in_progress;
    uint_t          in_progress_max;
    uint_t          sends;
    uint_t          sends_max;
    uint_t          sends_lowat;
    uint_t          sends_hiwat;
    ibcm_state_data_t   *tail;
    ibcm_state_data_t   head;
} ibcm_open;

static void ibcm_open_task(void *);

/*
 * Flow control logic for SA access and close_rc_channel calls follows.
 */

int ibcm_close_simul_max    = 12;
int ibcm_lapr_simul_max     = 12;
int ibcm_saa_simul_max      = 8;

typedef struct ibcm_flow1_s {
    struct ibcm_flow1_s *link;
    kcondvar_t      cv;
    uint8_t         waiters;    /* 1 to IBCM_FLOW_SIMUL_MAX */
} ibcm_flow1_t;

typedef struct ibcm_flow_s {
    ibcm_flow1_t        *list;
    uint_t          simul;  /* #requests currently outstanding */
    uint_t          simul_max;
    uint_t          waiters_per_chunk;
    uint_t          lowat;
    uint_t          lowat_default;
    /* statistics */
    uint_t          total;
} ibcm_flow_t;

ibcm_flow_t ibcm_saa_flow;
ibcm_flow_t ibcm_close_flow;
ibcm_flow_t ibcm_lapr_flow;

/* NONBLOCKING close requests are queued */
struct ibcm_close_s {
    kmutex_t        mutex;
    ibcm_state_data_t   *tail;
    ibcm_state_data_t   head;
} ibcm_close;

static ibt_clnt_modinfo_t ibcm_ibt_modinfo = {  /* Client's modinfop */
    IBTI_V2,
    IBT_CM,
    ibcm_async_handler,
    NULL,
    "IBCM"
};

/* IBCM's list of HCAs registered with it */
static ibcm_hca_info_t  *ibcm_hca_listp = NULL; /* CM's HCA list */

/* Array of CM state call table functions */
ibcm_state_handler_t    ibcm_sm_funcs_tbl[] = {
    ibcm_process_req_msg,
    ibcm_process_mra_msg,
    ibcm_process_rej_msg,
    ibcm_process_rep_msg,
    ibcm_process_rtu_msg,
    ibcm_process_dreq_msg,
    ibcm_process_drep_msg,
    ibcm_process_sidr_req_msg,
    ibcm_process_sidr_rep_msg,
    ibcm_process_lap_msg,
    ibcm_process_apr_msg
};

/* the following globals are CM tunables */
ibt_rnr_nak_time_t  ibcm_default_rnr_nak_time = IBT_RNR_NAK_655ms;

uint32_t    ibcm_max_retries = IBCM_MAX_RETRIES;
clock_t     ibcm_local_processing_time = IBCM_LOCAL_RESPONSE_TIME;
clock_t     ibcm_remote_response_time = IBCM_REMOTE_RESPONSE_TIME;
ib_time_t   ibcm_max_sidr_rep_proctime = IBCM_MAX_SIDR_PROCESS_TIME;
ib_time_t   ibcm_max_sidr_pktlife_time = IBCM_MAX_SIDR_PKT_LIFE_TIME;

ib_time_t   ibcm_max_sidr_rep_store_time = 18;
uint32_t    ibcm_wait_for_acc_cnt_timeout = 500000; /* 500 ms */
uint32_t    ibcm_wait_for_res_cnt_timeout = 500000; /* 500 ms */

ib_time_t   ibcm_max_ib_pkt_lt = IBCM_MAX_IB_PKT_LT;
ib_time_t   ibcm_max_ib_mad_pkt_lt = IBCM_MAX_IB_MAD_PKT_LT;

/*
 * This delay accounts for time involved in various activities as follows :
 *
 * IBMF delays for posting the MADs in non-blocking mode
 * IBMF delays for receiving the MADs and delivering to CM
 * CM delays in processing the MADs before invoking client handlers,
 * Any other delays associated with HCA driver in processing the MADs and
 *  other subsystems that CM may invoke (ex : SA, HCA driver)
 */
uint32_t    ibcm_sw_delay   = 1000; /* 1000us / 1ms */
uint32_t    ibcm_max_sa_retries = IBCM_MAX_SA_RETRIES + 1;

/*  approx boot time */
uint32_t    ibcm_adj_btime = 4; /* 4 seconds */

/*
 * The information in ibcm_clpinfo is kept in wireformat and is setup at
 * init time, and used read-only after that
 */
ibcm_classportinfo_msg_t    ibcm_clpinfo;

char    *event_str[] = {
    "NEVER SEE THIS             ",
    "SESSION_ID                 ",
    "CHAN_HDL                   ",
    "LOCAL_COMID/HCA/PORT       ",
    "LOCAL_QPN                  ",
    "REMOTE_COMID/HCA           ",
    "REMOTE_QPN                 ",
    "BASE_TIME                  ",
    "INCOMING_REQ               ",
    "INCOMING_REP               ",
    "INCOMING_RTU               ",
    "INCOMING_COMEST            ",
    "INCOMING_MRA               ",
    "INCOMING_REJ               ",
    "INCOMING_LAP               ",
    "INCOMING_APR               ",
    "INCOMING_DREQ              ",
    "INCOMING_DREP              ",
    "OUTGOING_REQ               ",
    "OUTGOING_REP               ",
    "OUTGOING_RTU               ",
    "OUTGOING_LAP               ",
    "OUTGOING_APR               ",
    "OUTGOING_MRA               ",
    "OUTGOING_REJ               ",
    "OUTGOING_DREQ              ",
    "OUTGOING_DREP              ",
    "REQ_POST_COMPLETE          ",
    "REP_POST_COMPLETE          ",
    "RTU_POST_COMPLETE          ",
    "MRA_POST_COMPLETE          ",
    "REJ_POST_COMPLETE          ",
    "LAP_POST_COMPLETE          ",
    "APR_POST_COMPLETE          ",
    "DREQ_POST_COMPLETE         ",
    "DREP_POST_COMPLETE         ",
    "TIMEOUT_REP                ",
    "CALLED_REQ_RCVD_EVENT      ",
    "RET_REQ_RCVD_EVENT         ",
    "CALLED_REP_RCVD_EVENT      ",
    "RET_REP_RCVD_EVENT         ",
    "CALLED_CONN_EST_EVENT      ",
    "RET_CONN_EST_EVENT         ",
    "CALLED_CONN_FAIL_EVENT     ",
    "RET_CONN_FAIL_EVENT        ",
    "CALLED_CONN_CLOSE_EVENT    ",
    "RET_CONN_CLOSE_EVENT       ",
    "INIT_INIT                  ",
    "INIT_INIT_FAIL             ",
    "INIT_RTR                   ",
    "INIT_RTR_FAIL              ",
    "RTR_RTS                    ",
    "RTR_RTS_FAIL               ",
    "RTS_RTS                    ",
    "RTS_RTS_FAIL               ",
    "TO_ERROR                   ",
    "ERROR_FAIL                 ",
    "SET_ALT                    ",
    "SET_ALT_FAIL               ",
    "STALE_DETECT               ",
    "OUTGOING_REQ_RETRY         ",
    "OUTGOING_REP_RETRY         ",
    "OUTGOING_LAP_RETRY         ",
    "OUTGOING_MRA_RETRY         ",
    "OUTGOING_DREQ_RETRY        ",
    "NEVER SEE THIS             "
};

char    ibcm_debug_buf[IBCM_DEBUG_BUF_SIZE];

_NOTE(SCHEME_PROTECTS_DATA("used in a localized function consistently",
    ibcm_debug_buf))
_NOTE(READ_ONLY_DATA(ibcm_taskq))

_NOTE(MUTEX_PROTECTS_DATA(ibcm_timeout_list_lock, ibcm_timeout_list_flags))
_NOTE(MUTEX_PROTECTS_DATA(ibcm_timeout_list_lock, ibcm_timeout_list_hdr))
_NOTE(MUTEX_PROTECTS_DATA(ibcm_timeout_list_lock, ibcm_ud_timeout_list_hdr))

#ifdef DEBUG
int     ibcm_test_mode = 0; /* set to 1, if running tests */
#endif


/* Module Driver Info */
static struct modlmisc ibcm_modlmisc = {
    &mod_miscops,
    "IB Communication Manager"
};

/* Module Linkage */
static struct modlinkage ibcm_modlinkage = {
    MODREV_1,
    &ibcm_modlmisc,
    NULL
};


int
_init(void)
{
    int     rval;
    ibcm_status_t   status;

    status = ibcm_init();
    if (status != IBCM_SUCCESS) {
        IBTF_DPRINTF_L2(cmlog, "_init: ibcm failed %d", status);
        return (EINVAL);
    }

    rval = mod_install(&ibcm_modlinkage);
    if (rval != 0) {
        IBTF_DPRINTF_L2(cmlog, "_init: ibcm mod_install failed %d",
            rval);
        (void) ibcm_fini();
    }

    IBTF_DPRINTF_L5(cmlog, "_init: ibcm successful");
    return (rval);

}


int
_info(struct modinfo *modinfop)
{
    return (mod_info(&ibcm_modlinkage, modinfop));
}


int
_fini(void)
{
    int status;

    if (ibcm_fini() != IBCM_SUCCESS)
        return (EBUSY);

    if ((status = mod_remove(&ibcm_modlinkage)) != 0) {
        IBTF_DPRINTF_L2(cmlog, "_fini: ibcm mod_remove failed %d",
            status);
        return (status);
    }

    IBTF_DPRINTF_L5(cmlog, "_fini: ibcm successful");

    return (status);
}

/* Initializes all global mutex and CV in cm module */
static void
ibcm_init_locks()
{

    /* Verify CM MAD sizes */
#ifdef DEBUG

    if (ibcm_test_mode > 1) {

        IBTF_DPRINTF_L1(cmlog, "REQ MAD SIZE %d",
            sizeof (ibcm_req_msg_t));
        IBTF_DPRINTF_L1(cmlog, "REP MAD SIZE %d",
            sizeof (ibcm_rep_msg_t));
        IBTF_DPRINTF_L1(cmlog, "RTU MAD SIZE %d",
            sizeof (ibcm_rtu_msg_t));
        IBTF_DPRINTF_L1(cmlog, "MRA MAD SIZE %d",
            sizeof (ibcm_mra_msg_t));
        IBTF_DPRINTF_L1(cmlog, "REJ MAD SIZE %d",
            sizeof (ibcm_rej_msg_t));
        IBTF_DPRINTF_L1(cmlog, "LAP MAD SIZE %d",
            sizeof (ibcm_lap_msg_t));
        IBTF_DPRINTF_L1(cmlog, "APR MAD SIZE %d",
            sizeof (ibcm_apr_msg_t));
        IBTF_DPRINTF_L1(cmlog, "DREQ MAD SIZE %d",
            sizeof (ibcm_dreq_msg_t));
        IBTF_DPRINTF_L1(cmlog, "DREP MAD SIZE %d",
            sizeof (ibcm_drep_msg_t));
        IBTF_DPRINTF_L1(cmlog, "SIDR REQ MAD SIZE %d",
            sizeof (ibcm_sidr_req_msg_t));
        IBTF_DPRINTF_L1(cmlog, "SIDR REP MAD SIZE %d",
            sizeof (ibcm_sidr_rep_msg_t));
    }

#endif

    /* Create all global locks within cm module */
    mutex_init(&ibcm_svc_info_lock, NULL, MUTEX_DEFAULT, NULL);
    mutex_init(&ibcm_timeout_list_lock, NULL, MUTEX_DEFAULT, NULL);
    mutex_init(&ibcm_global_hca_lock, NULL, MUTEX_DEFAULT, NULL);
    mutex_init(&ibcm_sa_open_lock, NULL, MUTEX_DEFAULT, NULL);
    mutex_init(&ibcm_recv_mutex, NULL, MUTEX_DEFAULT, NULL);
    mutex_init(&ibcm_sm_notice_serialize_lock, NULL, MUTEX_DEFAULT, NULL);
    mutex_init(&ibcm_qp_list_lock, NULL, MUTEX_DEFAULT, NULL);
    mutex_init(&ibcm_trace_mutex, NULL, MUTEX_DEFAULT, NULL);
    mutex_init(&ibcm_trace_print_mutex, NULL, MUTEX_DEFAULT, NULL);
    cv_init(&ibcm_svc_info_cv, NULL, CV_DRIVER, NULL);
    cv_init(&ibcm_timeout_list_cv, NULL, CV_DRIVER, NULL);
    cv_init(&ibcm_timeout_thread_done_cv, NULL, CV_DRIVER, NULL);
    cv_init(&ibcm_global_hca_cv, NULL, CV_DRIVER, NULL);
    cv_init(&ibcm_sa_open_cv, NULL, CV_DRIVER, NULL);
    avl_create(&ibcm_svc_avl_tree, ibcm_svc_compare,
        sizeof (ibcm_svc_info_t),
        offsetof(struct ibcm_svc_info_s, svc_link));

    IBTF_DPRINTF_L5(cmlog, "ibcm_init_locks: done");
}

/* Destroys all global mutex and CV in cm module */
static void
ibcm_fini_locks()
{
    /* Destroy all global locks within cm module */
    mutex_destroy(&ibcm_svc_info_lock);
    mutex_destroy(&ibcm_timeout_list_lock);
    mutex_destroy(&ibcm_global_hca_lock);
    mutex_destroy(&ibcm_sa_open_lock);
    mutex_destroy(&ibcm_recv_mutex);
    mutex_destroy(&ibcm_sm_notice_serialize_lock);
    mutex_destroy(&ibcm_qp_list_lock);
    mutex_destroy(&ibcm_trace_mutex);
    mutex_destroy(&ibcm_trace_print_mutex);
    cv_destroy(&ibcm_svc_info_cv);
    cv_destroy(&ibcm_timeout_list_cv);
    cv_destroy(&ibcm_timeout_thread_done_cv);
    cv_destroy(&ibcm_global_hca_cv);
    cv_destroy(&ibcm_sa_open_cv);
    avl_destroy(&ibcm_svc_avl_tree);

    IBTF_DPRINTF_L5(cmlog, "ibcm_fini_locks: done");
}


/* Initialize CM's classport info */
static void
ibcm_init_classportinfo()
{
    _NOTE(NOW_INVISIBLE_TO_OTHER_THREADS(ibcm_clpinfo));

    ibcm_clpinfo.BaseVersion = IBCM_MAD_BASE_VERSION;
    ibcm_clpinfo.ClassVersion = IBCM_MAD_CLASS_VERSION;

    /* For now, CM supports same capabilities at all ports */
    ibcm_clpinfo.CapabilityMask =
        h2b16(IBCM_CPINFO_CAP_RC | IBCM_CPINFO_CAP_SIDR);

    /* Bits 0-7 are all 0 for Communication Mgmt Class */

    /* For now, CM has the same respvalue at all ports */
    ibcm_clpinfo.RespTimeValue_plus =
        h2b32(ibt_usec2ib(ibcm_local_processing_time) & 0x1f);

    /* For now, redirect fields are set to 0 */
    /* Trap fields are not applicable to CM, hence set to 0 */

    _NOTE(NOW_VISIBLE_TO_OTHER_THREADS(ibcm_clpinfo));
    IBTF_DPRINTF_L5(cmlog, "ibcm_init_classportinfo: done");
}

/*
 * ibcm_init():
 *  - call ibt_attach()
 *  - create AVL trees
 *  - Attach HCA handlers that are already present before
 *  CM got loaded.
 *
 * Arguments:   NONE
 *
 * Return values:
 *  IBCM_SUCCESS - success
 */
static ibcm_status_t
ibcm_init(void)
{
    ibt_status_t    status;
    kthread_t   *t;

    IBTF_DPRINTF_L3(cmlog, "ibcm_init:");

    ibcm_init_classportinfo();

    if (ibcm_init_ids() != IBCM_SUCCESS) {
        IBTF_DPRINTF_L1(cmlog, "ibcm_init: "
            "fatal error: vmem_create() failed");
        return (IBCM_FAILURE);
    }
    ibcm_init_locks();

    if (ibcm_ar_init() != IBCM_SUCCESS) {
        IBTF_DPRINTF_L1(cmlog, "ibcm_init: "
            "fatal error: ibcm_ar_init() failed");
        ibcm_fini_ids();
        ibcm_fini_locks();
        return (IBCM_FAILURE);
    }
    ibcm_rc_flow_control_init();

    _NOTE(NOW_INVISIBLE_TO_OTHER_THREADS(ibcm_taskq))
    ibcm_taskq = system_taskq;
    _NOTE(NOW_VISIBLE_TO_OTHER_THREADS(ibcm_taskq))

    _NOTE(NOW_INVISIBLE_TO_OTHER_THREADS(ibcm_timeout_list_flags))
    _NOTE(NOW_INVISIBLE_TO_OTHER_THREADS(ibcm_timeout_thread_did))

    /* Start the timeout list processing thread */
    ibcm_timeout_list_flags = 0;
    t = thread_create(NULL, 0, ibcm_process_tlist, 0, 0, &p0, TS_RUN,
        ibcm_timeout_thread_pri);
    ibcm_timeout_thread_did = t->t_did;

    _NOTE(NOW_VISIBLE_TO_OTHER_THREADS(ibcm_timeout_list_flags))
    _NOTE(NOW_VISIBLE_TO_OTHER_THREADS(ibcm_timeout_thread_did))

    /*
     * NOTE : if ibt_attach is done after ibcm_init_hcas, then some
     * HCA DR events may be lost. CM could call re-init hca list
     * again, but it is more complicated. Some HCA's DR's lost may
     * be HCA detach, which makes hca list re-syncing and locking more
     * complex
     */
    status = ibt_attach(&ibcm_ibt_modinfo, NULL, NULL, &ibcm_ibt_handle);
    if (status != IBT_SUCCESS) {
        IBTF_DPRINTF_L2(cmlog, "ibcm_init(): ibt_attach failed %d",
            status);
        (void) ibcm_ar_fini();
        ibcm_stop_timeout_thread();
        ibcm_fini_ids();
        ibcm_fini_locks();
        ibcm_rc_flow_control_fini();
        return (IBCM_FAILURE);
    }

    /* Block all HCA attach/detach asyncs */
    mutex_enter(&ibcm_global_hca_lock);

    ibcm_init_hcas();
    ibcm_finit_state = IBCM_FINIT_IDLE;

    ibcm_path_cache_init();

    /* Unblock any waiting HCA DR asyncs in CM */
    mutex_exit(&ibcm_global_hca_lock);

    IBTF_DPRINTF_L4(cmlog, "ibcm_init: done");
    return (IBCM_SUCCESS);
}

/* Allocates and initializes the "per hca" global data in CM */
static void
ibcm_init_hcas()
{
    uint_t  num_hcas = 0;
    ib_guid_t *guid_array;
    int i;

    IBTF_DPRINTF_L5(cmlog, "ibcm_init_hcas:");

    /* Get the number of HCAs */
    num_hcas = ibt_get_hca_list(&guid_array);
    IBTF_DPRINTF_L4(cmlog, "ibcm_init_hcas: ibt_get_hca_list() "
        "returned %d hcas", num_hcas);

    ASSERT(MUTEX_HELD(&ibcm_global_hca_lock));

    for (i = 0; i < num_hcas; i++)
        ibcm_hca_attach(guid_array[i]);

    if (num_hcas)
        ibt_free_hca_list(guid_array, num_hcas);

    IBTF_DPRINTF_L5(cmlog, "ibcm_init_hcas: done");
}


/*
 * ibcm_fini():
 *  - Deregister w/ ibt
 *  - Cleanup IBCM HCA listp
 *  - Destroy mutexes
 *
 * Arguments:   NONE
 *
 * Return values:
 *  IBCM_SUCCESS - success
 */
static ibcm_status_t
ibcm_fini(void)
{
    ibt_status_t    status;

    IBTF_DPRINTF_L3(cmlog, "ibcm_fini:");

    /*
     * CM assumes that the all general clients got rid of all the
     * established connections and service registrations, completed all
     * pending SIDR operations before a call to ibcm_fini()
     */

    if (ibcm_ar_fini() != IBCM_SUCCESS) {
        IBTF_DPRINTF_L2(cmlog, "ibcm_fini: ibcm_ar_fini failed");
        return (IBCM_FAILURE);
    }

    /* cleanup the svcinfo list */
    mutex_enter(&ibcm_svc_info_lock);
    if (avl_first(&ibcm_svc_avl_tree) != NULL) {
        IBTF_DPRINTF_L2(cmlog, "ibcm_fini: "
            "ibcm_svc_avl_tree is not empty");
        mutex_exit(&ibcm_svc_info_lock);
        return (IBCM_FAILURE);
    }
    mutex_exit(&ibcm_svc_info_lock);

    /* disables any new hca attach/detaches */
    mutex_enter(&ibcm_global_hca_lock);

    ibcm_finit_state = IBCM_FINIT_BUSY;

    if (ibcm_fini_hcas() != IBCM_SUCCESS) {
        IBTF_DPRINTF_L2(cmlog, "ibcm_fini: "
            "some hca's still have client resources");

        /* First, re-initialize the hcas */
        ibcm_init_hcas();
        /* and then enable the HCA asyncs */
        ibcm_finit_state = IBCM_FINIT_IDLE;
        mutex_exit(&ibcm_global_hca_lock);
        if (ibcm_ar_init() != IBCM_SUCCESS) {
            IBTF_DPRINTF_L1(cmlog, "ibcm_fini:ibcm_ar_init failed");
        }
        return (IBCM_FAILURE);
    }

    _NOTE(NOW_INVISIBLE_TO_OTHER_THREADS(ibcm_timeout_list_hdr))
    _NOTE(NOW_INVISIBLE_TO_OTHER_THREADS(ibcm_ud_timeout_list_hdr))

    ASSERT(ibcm_timeout_list_hdr == NULL);
    ASSERT(ibcm_ud_timeout_list_hdr == NULL);

    _NOTE(NOW_VISIBLE_TO_OTHER_THREADS(ibcm_timeout_list_hdr))
    _NOTE(NOW_VISIBLE_TO_OTHER_THREADS(ibcm_ud_timeout_list_hdr))

    /* Release any pending asyncs on ibcm_global_hca_lock */
    ibcm_finit_state = IBCM_FINIT_SUCCESS;
    mutex_exit(&ibcm_global_hca_lock);

    ibcm_stop_timeout_thread();

    /*
     * Detach from IBTL. Waits until all pending asyncs are complete.
     * Above cv_broadcast wakes up any waiting hca attach/detach asyncs
     */
    status = ibt_detach(ibcm_ibt_handle);

    /* if detach fails, CM didn't free up some resources, so assert */
    if (status != IBT_SUCCESS)
        IBTF_DPRINTF_L1(cmlog, "ibcm_fini: ibt_detach failed %d",
            status);

    ibcm_rc_flow_control_fini();

    ibcm_path_cache_fini();

    ibcm_fini_ids();
    ibcm_fini_locks();
    IBTF_DPRINTF_L3(cmlog, "ibcm_fini: done");
    return (IBCM_SUCCESS);
}

/* This routine exit's the ibcm timeout thread  */
static void
ibcm_stop_timeout_thread()
{
    mutex_enter(&ibcm_timeout_list_lock);

    /* Stop the timeout list processing thread */
    ibcm_timeout_list_flags =
        ibcm_timeout_list_flags | IBCM_TIMEOUT_THREAD_EXIT;

    /* Wake up, if the timeout thread is on a cv_wait */
    cv_signal(&ibcm_timeout_list_cv);

    mutex_exit(&ibcm_timeout_list_lock);
    thread_join(ibcm_timeout_thread_did);

    IBTF_DPRINTF_L5(cmlog, "ibcm_stop_timeout_thread: done");
}


/* Attempts to release all the hca's associated with CM */
static ibcm_status_t
ibcm_fini_hcas()
{
    ibcm_hca_info_t *hcap, *next;

    IBTF_DPRINTF_L4(cmlog, "ibcm_fini_hcas:");

    ASSERT(MUTEX_HELD(&ibcm_global_hca_lock));

    hcap = ibcm_hca_listp;
    while (hcap != NULL) {
        next = hcap->hca_next;
        if (ibcm_hca_detach(hcap) != IBCM_SUCCESS) {
            ibcm_hca_listp = hcap;
            return (IBCM_FAILURE);
        }
        hcap = next;
    }

    IBTF_DPRINTF_L4(cmlog, "ibcm_fini_hcas: SUCCEEDED");
    return (IBCM_SUCCESS);
}


/*
 * ibcm_hca_attach():
 *  Called as an asynchronous event to notify CM of an attach of HCA.
 *  Here ibcm_hca_info_t is initialized and all fields are
 *  filled in along with SA Access handles and IBMA handles.
 *  Also called from ibcm_init to initialize ibcm_hca_info_t's for each
 *  hca's
 *
 * Arguments: (WILL CHANGE BASED ON ASYNC EVENT CODE)
 *  hca_guid    - HCA's guid
 *
 * Return values: NONE
 */
static void
ibcm_hca_attach(ib_guid_t hcaguid)
{
    int         i;
    ibt_status_t        status;
    uint_t          nports = 0;
    ibcm_hca_info_t     *hcap;
    ibt_hca_attr_t      hca_attrs;

    IBTF_DPRINTF_L3(cmlog, "ibcm_hca_attach: guid = 0x%llX", hcaguid);

    ASSERT(MUTEX_HELD(&ibcm_global_hca_lock));

    _NOTE(NOW_INVISIBLE_TO_OTHER_THREADS(*hcap))

    status = ibt_query_hca_byguid(hcaguid, &hca_attrs);
    if (status != IBT_SUCCESS) {
        IBTF_DPRINTF_L2(cmlog, "ibcm_hca_attach: "
            "ibt_query_hca_byguid failed = %d", status);
        return;
    }
    nports = hca_attrs.hca_nports;

    IBTF_DPRINTF_L4(cmlog, "ibcm_hca_attach: num ports = %x", nports);

    if ((hcap = ibcm_add_hca_entry(hcaguid, nports)) == NULL)
        return;

    hcap->hca_guid = hcaguid;   /* Set GUID */
    hcap->hca_num_ports = nports;   /* Set number of ports */

    if (ibcm_init_hca_ids(hcap) != IBCM_SUCCESS) {
        ibcm_delete_hca_entry(hcap);
        return;
    }

    /* Store the static hca attribute data */
    hcap->hca_caps = hca_attrs.hca_flags;
    hcap->hca_vendor_id = hca_attrs.hca_vendor_id;
    hcap->hca_device_id = hca_attrs.hca_device_id;
    hcap->hca_ack_delay = hca_attrs.hca_local_ack_delay;
    hcap->hca_max_rdma_in_qp = hca_attrs.hca_max_rdma_in_qp;
    hcap->hca_max_rdma_out_qp = hca_attrs.hca_max_rdma_out_qp;

    /* loop thru nports and initialize IBMF handles */
    for (i = 0; i < hcap->hca_num_ports; i++) {
        status = ibt_get_port_state_byguid(hcaguid, i + 1, NULL, NULL);
        if (status != IBT_SUCCESS) {
            IBTF_DPRINTF_L2(cmlog, "ibcm_hca_attach: "
                "port_num %d state DOWN", i + 1);
        }

        hcap->hca_port_info[i].port_hcap = hcap;
        hcap->hca_port_info[i].port_num = i+1;

        if (ibcm_hca_init_port(hcap, i) != IBT_SUCCESS)
            IBTF_DPRINTF_L2(cmlog, "ibcm_hca_attach: "
                "ibcm_hca_init_port failed %d port_num %d",
                status, i+1);
    }

    /* create the "active" CM AVL tree */
    avl_create(&hcap->hca_active_tree, ibcm_active_node_compare,
        sizeof (ibcm_state_data_t),
        offsetof(struct ibcm_state_data_s, avl_active_link));

    /* create the "passive" CM AVL tree */
    avl_create(&hcap->hca_passive_tree, ibcm_passive_node_compare,
        sizeof (ibcm_state_data_t),
        offsetof(struct ibcm_state_data_s, avl_passive_link));

    /* create the "passive comid" CM AVL tree */
    avl_create(&hcap->hca_passive_comid_tree,
        ibcm_passive_comid_node_compare,
        sizeof (ibcm_state_data_t),
        offsetof(struct ibcm_state_data_s, avl_passive_comid_link));

    /*
     * Mark the state of the HCA to "attach" only at the end
     * Now CM starts accepting incoming MADs and client API calls
     */
    hcap->hca_state = IBCM_HCA_ACTIVE;

    _NOTE(NOW_VISIBLE_TO_OTHER_THREADS(*hcap))

    IBTF_DPRINTF_L3(cmlog, "ibcm_hca_attach: ATTACH Done");
}

/*
 * ibcm_hca_detach():
 *  Called as an asynchronous event to notify CM of a detach of HCA.
 *  Here ibcm_hca_info_t is freed up and all fields that
 *  were initialized earlier are cleaned up
 *
 * Arguments: (WILL CHANGE BASED ON ASYNC EVENT CODE)
 *  hca_guid    - HCA's guid
 *
 * Return values:
 *  IBCM_SUCCESS    - able to detach HCA
 *  IBCM_FAILURE    - failed to detach HCA
 */
static ibcm_status_t
ibcm_hca_detach(ibcm_hca_info_t *hcap)
{
    int     port_index, i;
    ibcm_status_t   status = IBCM_SUCCESS;
    clock_t     absolute_time;

    IBTF_DPRINTF_L3(cmlog, "ibcm_hca_detach: hcap = 0x%p guid = 0x%llX",
        hcap, hcap->hca_guid);

    ASSERT(MUTEX_HELD(&ibcm_global_hca_lock));

    /*
     * Declare hca is going away to all CM clients. Wait until the
     * access count becomes zero.
     */
    hcap->hca_state = IBCM_HCA_NOT_ACTIVE;

    /* wait on response CV to 500mS */
    absolute_time = ddi_get_lbolt() +
        drv_usectohz(ibcm_wait_for_acc_cnt_timeout);

    while (hcap->hca_acc_cnt > 0)
        if (cv_timedwait(&ibcm_global_hca_cv, &ibcm_global_hca_lock,
            absolute_time) == -1)
            break;

    if (hcap->hca_acc_cnt != 0) {
        /* We got a timeout */
#ifdef DEBUG
        if (ibcm_test_mode > 0)
            IBTF_DPRINTF_L1(cmlog, "ibcm_hca_detach: Unexpected "
                "abort due to timeout on acc_cnt %u",
                hcap->hca_acc_cnt);
        else
#endif
            IBTF_DPRINTF_L2(cmlog, "ibcm_hca_detach: Aborting due"
                " to timeout on acc_cnt %u", hcap->hca_acc_cnt);
        hcap->hca_state = IBCM_HCA_ACTIVE;
        return (IBCM_FAILURE);
    }

    /*
     * First make sure, there are no active users of ibma handles,
     * and then de-register handles.
     */

    /* make sure that there are no "Service"s registered w/ this HCA. */
    if (hcap->hca_svc_cnt != 0) {
        IBTF_DPRINTF_L2(cmlog, "ibcm_hca_detach: "
            "Active services still there %d", hcap->hca_svc_cnt);
        hcap->hca_state = IBCM_HCA_ACTIVE;
        return (IBCM_FAILURE);
    }

    if (ibcm_check_sidr_clean(hcap) != IBCM_SUCCESS) {
        IBTF_DPRINTF_L2(cmlog, "ibcm_hca_detach:"
            "There are active SIDR operations");
        hcap->hca_state = IBCM_HCA_ACTIVE;
        return (IBCM_FAILURE);
    }

    if (ibcm_check_avl_clean(hcap) != IBCM_SUCCESS) {
        IBTF_DPRINTF_L2(cmlog, "ibcm_hca_detach: "
            "There are active RC connections");
        hcap->hca_state = IBCM_HCA_ACTIVE;
        return (IBCM_FAILURE);
    }

    /*
     * Now, wait until all rc and sidr stateps go away
     * All these stateps must be short lived ones, waiting to be cleaned
     * up after some timeout value, based on the current state.
     */
    IBTF_DPRINTF_L5(cmlog, "ibcm_hca_detach:hca_guid = 0x%llX res_cnt = %d",
        hcap->hca_guid, hcap->hca_res_cnt);

    /* wait on response CV to 500mS */
    absolute_time = ddi_get_lbolt() +
        drv_usectohz(ibcm_wait_for_res_cnt_timeout);

    while (hcap->hca_res_cnt > 0)
        if (cv_timedwait(&ibcm_global_hca_cv, &ibcm_global_hca_lock,
            absolute_time) == -1)
            break;

    if (hcap->hca_res_cnt != 0) {
        /* We got a timeout waiting for hca_res_cnt to become 0 */
#ifdef DEBUG
        if (ibcm_test_mode > 0)
            IBTF_DPRINTF_L1(cmlog, "ibcm_hca_detach: Unexpected "
                "abort due to timeout on res_cnt %d",
                hcap->hca_res_cnt);
        else
#endif
            IBTF_DPRINTF_L2(cmlog, "ibcm_hca_detach: Aborting due"
                " to timeout on res_cnt %d", hcap->hca_res_cnt);
        hcap->hca_state = IBCM_HCA_ACTIVE;
        return (IBCM_FAILURE);
    }

    /* Re-assert the while loop step above */
    ASSERT(hcap->hca_sidr_list == NULL);
    avl_destroy(&hcap->hca_active_tree);
    avl_destroy(&hcap->hca_passive_tree);
    avl_destroy(&hcap->hca_passive_comid_tree);

    /*
     * Unregister all ports from IBMA
     * If there is a failure, re-initialize any free'd ibma handles. This
     * is required to receive the incoming mads
     */
    status = IBCM_SUCCESS;
    for (port_index = 0; port_index < hcap->hca_num_ports; port_index++) {
        if ((status = ibcm_hca_fini_port(hcap, port_index)) !=
            IBCM_SUCCESS) {
            IBTF_DPRINTF_L2(cmlog, "ibcm_hca_detach: "
                "Failed to free IBMA Handle for port_num %d",
                port_index + 1);
            break;
        }
    }

    /* If detach fails, re-initialize ibma handles for incoming mads */
    if (status != IBCM_SUCCESS)  {
        for (i = 0; i < port_index; i++) {
            if (ibcm_hca_init_port(hcap, i) != IBT_SUCCESS)
                IBTF_DPRINTF_L2(cmlog, "ibcm_hca_detach: "
                    "Failed to re-allocate IBMA Handles for"
                    " port_num %d", port_index + 1);
        }
        hcap->hca_state = IBCM_HCA_ACTIVE;
        return (IBCM_FAILURE);
    }

    ibcm_fini_hca_ids(hcap);
    ibcm_delete_hca_entry(hcap);

    IBTF_DPRINTF_L3(cmlog, "ibcm_hca_detach: DETACH succeeded");
    return (IBCM_SUCCESS);
}

/* Checks, if there are any active sidr state entries in the specified hca */
static ibcm_status_t
ibcm_check_sidr_clean(ibcm_hca_info_t *hcap)
{
    ibcm_ud_state_data_t    *usp;
    uint32_t        transient_cnt = 0;

    IBTF_DPRINTF_L5(cmlog, "ibcm_check_sidr_clean:");

    rw_enter(&hcap->hca_sidr_list_lock, RW_WRITER);
    usp = hcap->hca_sidr_list;  /* Point to the list */
    while (usp != NULL) {
        mutex_enter(&usp->ud_state_mutex);
        if ((usp->ud_state != IBCM_STATE_SIDR_REP_SENT) &&
            (usp->ud_state != IBCM_STATE_TIMED_OUT) &&
            (usp->ud_state != IBCM_STATE_DELETE)) {

            IBTF_DPRINTF_L3(cmlog, "ibcm_check_sidr_clean:"
                "usp = %p not in transient state = %d", usp,
                usp->ud_state);

            mutex_exit(&usp->ud_state_mutex);
            rw_exit(&hcap->hca_sidr_list_lock);
            return (IBCM_FAILURE);
        } else {
            mutex_exit(&usp->ud_state_mutex);
            ++transient_cnt;
        }

        usp = usp->ud_nextp;
    }
    rw_exit(&hcap->hca_sidr_list_lock);

    IBTF_DPRINTF_L4(cmlog, "ibcm_check_sidr_clean: transient_cnt %d",
        transient_cnt);

    return (IBCM_SUCCESS);
}

/* Checks, if there are any active rc state entries, in the specified hca */
static ibcm_status_t
ibcm_check_avl_clean(ibcm_hca_info_t *hcap)

{
    ibcm_state_data_t   *sp;
    avl_tree_t      *avl_tree;
    uint32_t        transient_cnt = 0;

    IBTF_DPRINTF_L5(cmlog, "ibcm_check_avl_clean:");
    /*
     * Both the trees ie., active and passive must reference to all
     * statep's, so let's use one
     */
    avl_tree = &hcap->hca_active_tree;

    rw_enter(&hcap->hca_state_rwlock, RW_WRITER);

    for (sp = avl_first(avl_tree); sp != NULL;
        sp = avl_walk(avl_tree, sp, AVL_AFTER)) {
        mutex_enter(&sp->state_mutex);
        if ((sp->state != IBCM_STATE_TIMEWAIT) &&
            (sp->state != IBCM_STATE_REJ_SENT) &&
            (sp->state != IBCM_STATE_DELETE)) {
            IBTF_DPRINTF_L3(cmlog, "ibcm_check_avl_clean: "
                "sp = %p not in transient state = %d", sp,
                sp->state);
            mutex_exit(&sp->state_mutex);
            rw_exit(&hcap->hca_state_rwlock);
            return (IBCM_FAILURE);
        } else {
            mutex_exit(&sp->state_mutex);
            ++transient_cnt;
        }
    }

    rw_exit(&hcap->hca_state_rwlock);

    IBTF_DPRINTF_L4(cmlog, "ibcm_check_avl_clean: transient_cnt %d",
        transient_cnt);

    return (IBCM_SUCCESS);
}

/* Adds a new entry into CM's global hca list, if hca_guid is not there yet */
static ibcm_hca_info_t *
ibcm_add_hca_entry(ib_guid_t hcaguid, uint_t nports)
{
    ibcm_hca_info_t *hcap;

    IBTF_DPRINTF_L5(cmlog, "ibcm_add_hca_entry: guid = 0x%llX",
        hcaguid);

    ASSERT(MUTEX_HELD(&ibcm_global_hca_lock));

    /*
     * Check if this hca_guid already in the list
     * If yes, then ignore this and return NULL
     */

    hcap = ibcm_hca_listp;

    /* search for this HCA */
    while (hcap != NULL) {
        if (hcap->hca_guid == hcaguid) {
            /* already exists */
            IBTF_DPRINTF_L2(cmlog, "ibcm_add_hca_entry: "
                "hcap %p guid 0x%llX, entry already exists !!",
                hcap, hcap->hca_guid);
            return (NULL);
        }
        hcap = hcap->hca_next;
    }

    /* Allocate storage for the new HCA entry found */
    hcap = kmem_zalloc(sizeof (ibcm_hca_info_t) +
        (nports - 1) * sizeof (ibcm_port_info_t), KM_SLEEP);

    /* initialize RW lock */
    rw_init(&hcap->hca_state_rwlock, NULL, RW_DRIVER, NULL);
    /* initialize SIDR list lock */
    rw_init(&hcap->hca_sidr_list_lock, NULL, RW_DRIVER, NULL);
    /* Insert "hcap" into the global HCA list maintained by CM */
    hcap->hca_next = ibcm_hca_listp;
    ibcm_hca_listp = hcap;

    IBTF_DPRINTF_L5(cmlog, "ibcm_add_hca_entry: done hcap = 0x%p", hcap);

    return (hcap);

}

/* deletes the given ibcm_hca_info_t from CM's global hca list */
void
ibcm_delete_hca_entry(ibcm_hca_info_t *hcap)
{
    ibcm_hca_info_t *headp, *prevp = NULL;

    /* ibcm_hca_global_lock is held */
    IBTF_DPRINTF_L5(cmlog, "ibcm_delete_hca_entry: guid = 0x%llX "
        "hcap = 0x%p", hcap->hca_guid, hcap);

    ASSERT(MUTEX_HELD(&ibcm_global_hca_lock));

    headp = ibcm_hca_listp;
    while (headp != NULL) {
        if (headp == hcap) {
            IBTF_DPRINTF_L3(cmlog, "ibcm_delete_hca_entry: "
                "deleting hcap %p hcaguid %llX", hcap,
                hcap->hca_guid);
            if (prevp) {
                prevp->hca_next = headp->hca_next;
            } else {
                prevp = headp->hca_next;
                ibcm_hca_listp = prevp;
            }
            rw_destroy(&hcap->hca_state_rwlock);
            rw_destroy(&hcap->hca_sidr_list_lock);
            kmem_free(hcap, sizeof (ibcm_hca_info_t) +
                (hcap->hca_num_ports - 1) *
                sizeof (ibcm_port_info_t));
            return;
        }

        prevp = headp;
        headp = headp->hca_next;
    }
}

/*
 * ibcm_find_hca_entry:
 *  Given a HCA's GUID find out ibcm_hca_info_t entry for that HCA
 *  This entry can be then used to access AVL tree/SIDR list etc.
 *  If entry exists and in HCA ATTACH state, then hca's ref cnt is
 *  incremented and entry returned. Else NULL returned.
 *
 *  All functions that use ibcm_find_hca_entry and get a non-NULL
 *  return values must call ibcm_dec_hca_acc_cnt to decrement the
 *  respective hca ref cnt. There shouldn't be any usage of
 *  ibcm_hca_info_t * returned from ibcm_find_hca_entry,
 *  after decrementing the hca_acc_cnt
 *
 * INPUTS:
 *  hca_guid    - HCA's guid
 *
 * RETURN VALUE:
 *  hcap        - if a match is found, else NULL
 */
ibcm_hca_info_t *
ibcm_find_hca_entry(ib_guid_t hca_guid)
{
    ibcm_hca_info_t *hcap;

    IBTF_DPRINTF_L5(cmlog, "ibcm_find_hca_entry: guid = 0x%llX", hca_guid);

    mutex_enter(&ibcm_global_hca_lock);

    hcap = ibcm_hca_listp;
    /* search for this HCA */
    while (hcap != NULL) {
        if (hcap->hca_guid == hca_guid)
            break;
        hcap = hcap->hca_next;
    }

    /* if no hcap for the hca_guid, return NULL */
    if (hcap == NULL) {
        mutex_exit(&ibcm_global_hca_lock);
        return (NULL);
    }

    /* return hcap, only if it valid to use */
    if (hcap->hca_state == IBCM_HCA_ACTIVE) {
        ++(hcap->hca_acc_cnt);

        IBTF_DPRINTF_L5(cmlog, "ibcm_find_hca_entry: "
            "found hcap = 0x%p hca_acc_cnt %u", hcap,
            hcap->hca_acc_cnt);

        mutex_exit(&ibcm_global_hca_lock);
        return (hcap);
    } else {
        mutex_exit(&ibcm_global_hca_lock);

        IBTF_DPRINTF_L2(cmlog, "ibcm_find_hca_entry: "
            "found hcap = 0x%p not in active state", hcap);
        return (NULL);
    }
}

/*
 * Searches for ibcm_hca_info_t entry based on hca_guid, but doesn't increment
 * the hca's reference count. This function is used, where the calling context
 * is attempting to delete hcap itself and hence acc_cnt cannot be incremented
 * OR assumes that valid hcap must be available in ibcm's global hca list.
 */
ibcm_hca_info_t *
ibcm_find_hcap_entry(ib_guid_t hca_guid)
{
    ibcm_hca_info_t *hcap;

    IBTF_DPRINTF_L5(cmlog, "ibcm_find_hcap_entry: guid = 0x%llX", hca_guid);

    ASSERT(MUTEX_HELD(&ibcm_global_hca_lock));

    hcap = ibcm_hca_listp;
    /* search for this HCA */
    while (hcap != NULL) {
        if (hcap->hca_guid == hca_guid)
            break;
        hcap = hcap->hca_next;
    }

    if (hcap == NULL)
        IBTF_DPRINTF_L2(cmlog, "ibcm_find_hcap_entry: No hcap found for"
            " hca_guid 0x%llX", hca_guid);
    else
        IBTF_DPRINTF_L5(cmlog, "ibcm_find_hcap_entry: hcap found for"
            " hca_guid 0x%llX", hca_guid);

    return (hcap);
}

/* increment the hca's temporary reference count */
ibcm_status_t
ibcm_inc_hca_acc_cnt(ibcm_hca_info_t *hcap)
{
    mutex_enter(&ibcm_global_hca_lock);
    if (hcap->hca_state == IBCM_HCA_ACTIVE) {
        ++(hcap->hca_acc_cnt);
        IBTF_DPRINTF_L5(cmlog, "ibcm_inc_hca_acc_cnt: "
            "hcap = 0x%p  acc_cnt = %d ", hcap, hcap->hca_acc_cnt);
        mutex_exit(&ibcm_global_hca_lock);
        return (IBCM_SUCCESS);
    } else {
        IBTF_DPRINTF_L2(cmlog, "ibcm_inc_hca_acc_cnt: "
            "hcap INACTIVE 0x%p  acc_cnt = %d ", hcap,
            hcap->hca_acc_cnt);
        mutex_exit(&ibcm_global_hca_lock);
        return (IBCM_FAILURE);
    }
}

/* decrement the hca's ref count, and wake up any waiting threads */
void
ibcm_dec_hca_acc_cnt(ibcm_hca_info_t *hcap)
{
    mutex_enter(&ibcm_global_hca_lock);
    ASSERT(hcap->hca_acc_cnt > 0);
    --(hcap->hca_acc_cnt);
    IBTF_DPRINTF_L5(cmlog, "ibcm_dec_hca_acc_cnt: hcap = 0x%p "
        "acc_cnt = %d", hcap, hcap->hca_acc_cnt);
    if ((hcap->hca_state == IBCM_HCA_NOT_ACTIVE) &&
        (hcap->hca_acc_cnt == 0)) {
        IBTF_DPRINTF_L3(cmlog, "ibcm_dec_hca_acc_cnt: "
            "cv_broadcast for hcap = 0x%p", hcap);
        cv_broadcast(&ibcm_global_hca_cv);
    }
    mutex_exit(&ibcm_global_hca_lock);
}

/* increment the hca's resource count */
void
ibcm_inc_hca_res_cnt(ibcm_hca_info_t *hcap)

{
    mutex_enter(&ibcm_global_hca_lock);
    ++(hcap->hca_res_cnt);
    IBTF_DPRINTF_L5(cmlog, "ibcm_inc_hca_res_cnt: hcap = 0x%p "
        "ref_cnt = %d", hcap, hcap->hca_res_cnt);
    mutex_exit(&ibcm_global_hca_lock);
}

/* decrement the hca's resource count, and wake up any waiting threads */
void
ibcm_dec_hca_res_cnt(ibcm_hca_info_t *hcap)
{
    mutex_enter(&ibcm_global_hca_lock);
    ASSERT(hcap->hca_res_cnt > 0);
    --(hcap->hca_res_cnt);
    IBTF_DPRINTF_L5(cmlog, "ibcm_dec_hca_res_cnt: hcap = 0x%p "
        "ref_cnt = %d", hcap, hcap->hca_res_cnt);
    if ((hcap->hca_state == IBCM_HCA_NOT_ACTIVE) &&
        (hcap->hca_res_cnt == 0)) {
        IBTF_DPRINTF_L3(cmlog, "ibcm_dec_hca_res_cnt: "
            "cv_broadcast for hcap = 0x%p", hcap);
        cv_broadcast(&ibcm_global_hca_cv);
    }
    mutex_exit(&ibcm_global_hca_lock);
}

/* increment the hca's service count */
void
ibcm_inc_hca_svc_cnt(ibcm_hca_info_t *hcap)

{
    mutex_enter(&ibcm_global_hca_lock);
    ++(hcap->hca_svc_cnt);
    IBTF_DPRINTF_L5(cmlog, "ibcm_inc_hca_svc_cnt: hcap = 0x%p "
        "svc_cnt = %d", hcap, hcap->hca_svc_cnt);
    mutex_exit(&ibcm_global_hca_lock);
}

/* decrement the hca's service count */
void
ibcm_dec_hca_svc_cnt(ibcm_hca_info_t *hcap)
{
    mutex_enter(&ibcm_global_hca_lock);
    ASSERT(hcap->hca_svc_cnt > 0);
    --(hcap->hca_svc_cnt);
    IBTF_DPRINTF_L5(cmlog, "ibcm_dec_hca_svc_cnt: hcap = 0x%p "
        "svc_cnt = %d", hcap, hcap->hca_svc_cnt);
    mutex_exit(&ibcm_global_hca_lock);
}

/*
 * The following code manages three classes of requests that CM makes to
 * the fabric.  Those three classes are SA_ACCESS, REQ/REP/RTU, and DREQ/DREP.
 * The main issue is that the fabric can become very busy, and the CM
 * protocols rely on responses being made based on a predefined timeout
 * value.  By managing how many simultaneous sessions are allowed, there
 * is observed extremely high reliability of CM protocol succeeding when
 * it should.
 *
 * SA_ACCESS and DREQ/DREP are managed at the thread level, whereby the
 * thread blocks until there are less than some number of threads doing
 * similar requests.
 *
 * REQ/REP/RTU requests beyond a given limit are added to a list,
 * allowing the thread to return immediately to its caller in the
 * case where the "mode" is IBT_NONBLOCKING.  This is the mode used
 * by uDAPL and seems to be an important feature/behavior.
 */

static int
ibcm_ok_to_start(struct ibcm_open_s *openp)
{
    return (openp->sends < openp->sends_hiwat &&
        openp->in_progress < openp->in_progress_max);
}

void
ibcm_open_done(ibcm_state_data_t *statep)
{
    int run;
    ibcm_state_data_t **linkp, *tmp;

    ASSERT(MUTEX_HELD(&statep->state_mutex));
    if (statep->open_flow == 1) {
        statep->open_flow = 0;
        mutex_enter(&ibcm_open.mutex);
        if (statep->open_link == NULL) {
            ibcm_open.in_progress--;
            run = ibcm_ok_to_start(&ibcm_open);
        } else {
            ibcm_open.queued--;
            linkp = &ibcm_open.head.open_link;
            while (*linkp != statep)
                linkp = &((*linkp)->open_link);
            *linkp = statep->open_link;
            statep->open_link = NULL;
            /*
             * If we remove what tail pointed to, we need
             * to reassign tail (it is never NULL).
             * tail points to head for the empty list.
             */
            if (ibcm_open.tail == statep) {
                tmp = &ibcm_open.head;
                while (tmp->open_link != &ibcm_open.head)
                    tmp = tmp->open_link;
                ibcm_open.tail = tmp;
            }
            run = 0;
        }
        mutex_exit(&ibcm_open.mutex);
        if (run)
            ibcm_run_tlist_thread();
    }
}

/* dtrace */
void
ibcm_open_wait(hrtime_t delta)
{
    if (delta > 1000000)
        IBTF_DPRINTF_L2(cmlog, "ibcm_open_wait: flow more %lld", delta);
}

void
ibcm_open_start(ibcm_state_data_t *statep)
{
    ibcm_insert_trace(statep, IBCM_TRACE_OUTGOING_REQ);

    mutex_enter(&statep->state_mutex);
    ibcm_open_wait(gethrtime() - statep->post_time);
    mutex_exit(&statep->state_mutex);

    ibcm_post_rc_mad(statep, statep->stored_msg, ibcm_post_req_complete,
        statep);

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

void
ibcm_open_enqueue(ibcm_state_data_t *statep)
{
    int run;

    mutex_enter(&statep->state_mutex);
    statep->post_time = gethrtime();
    mutex_exit(&statep->state_mutex);
    mutex_enter(&ibcm_open.mutex);
    if (ibcm_open.queued == 0 && ibcm_ok_to_start(&ibcm_open)) {
        ibcm_open.in_progress++;
        mutex_exit(&ibcm_open.mutex);
        ibcm_open_start(statep);
    } else {
        ibcm_open.queued++;
        statep->open_link = &ibcm_open.head;
        ibcm_open.tail->open_link = statep;
        ibcm_open.tail = statep;
        run = ibcm_ok_to_start(&ibcm_open);
        mutex_exit(&ibcm_open.mutex);
        if (run)
            ibcm_run_tlist_thread();
    }
}

ibcm_state_data_t *
ibcm_open_dequeue(void)
{
    ibcm_state_data_t *statep;

    ASSERT(MUTEX_HELD(&ibcm_open.mutex));
    ibcm_open.queued--;
    ibcm_open.in_progress++;
    statep = ibcm_open.head.open_link;
    ibcm_open.head.open_link = statep->open_link;
    statep->open_link = NULL;
    /*
     * If we remove what tail pointed to, we need
     * to reassign tail (it is never NULL).
     * tail points to head for the empty list.
     */
    if (ibcm_open.tail == statep)
        ibcm_open.tail = &ibcm_open.head;
    return (statep);
}

void
ibcm_check_for_opens(void)
{
    ibcm_state_data_t   *statep;

    mutex_enter(&ibcm_open.mutex);

    while (ibcm_open.queued > 0) {
        if (ibcm_ok_to_start(&ibcm_open)) {
            statep = ibcm_open_dequeue();
            mutex_exit(&ibcm_open.mutex);

            ibcm_open_start(statep);

            mutex_enter(&ibcm_open.mutex);
        } else {
            break;
        }
    }
    mutex_exit(&ibcm_open.mutex);
}


static void
ibcm_flow_init(ibcm_flow_t *flow, uint_t simul_max)
{
    flow->list          = NULL;
    flow->simul         = 0;
    flow->waiters_per_chunk     = 4;
    flow->simul_max         = simul_max;
    flow->lowat         = simul_max - flow->waiters_per_chunk;
    flow->lowat_default     = flow->lowat;
    /* stats */
    flow->total         = 0;
}

static void
ibcm_rc_flow_control_init(void)
{
    mutex_init(&ibcm_open.mutex, NULL, MUTEX_DEFAULT, NULL);
    mutex_enter(&ibcm_open.mutex);
    ibcm_flow_init(&ibcm_close_flow, ibcm_close_simul_max);
    ibcm_flow_init(&ibcm_lapr_flow, ibcm_lapr_simul_max);
    ibcm_flow_init(&ibcm_saa_flow, ibcm_saa_simul_max);

    ibcm_open.queued        = 0;
    ibcm_open.exit_deferred     = 0;
    ibcm_open.in_progress       = 0;
    ibcm_open.in_progress_max   = 16;
    ibcm_open.sends         = 0;
    ibcm_open.sends_max         = 0;
    ibcm_open.sends_lowat       = 8;
    ibcm_open.sends_hiwat       = 16;
    ibcm_open.tail          = &ibcm_open.head;
    ibcm_open.head.open_link    = NULL;
    mutex_exit(&ibcm_open.mutex);

    mutex_init(&ibcm_close.mutex, NULL, MUTEX_DEFAULT, NULL);
    mutex_enter(&ibcm_close.mutex);
    ibcm_close.tail         = &ibcm_close.head;
    ibcm_close.head.close_link  = NULL;
    mutex_exit(&ibcm_close.mutex);
}

static void
ibcm_rc_flow_control_fini(void)
{
    mutex_destroy(&ibcm_open.mutex);
    mutex_destroy(&ibcm_close.mutex);
}

static ibcm_flow1_t *
ibcm_flow_find(ibcm_flow_t *flow)
{
    ibcm_flow1_t *flow1;
    ibcm_flow1_t *f;

    f = flow->list;
    if (f) {    /* most likely code path */
        while (f->link != NULL)
            f = f->link;
        if (f->waiters < flow->waiters_per_chunk)
            return (f);
    }

    /* There was no flow1 list element ready for another waiter */
    mutex_exit(&ibcm_open.mutex);
    flow1 = kmem_alloc(sizeof (*flow1), KM_SLEEP);
    mutex_enter(&ibcm_open.mutex);

    f = flow->list;
    if (f) {
        while (f->link != NULL)
            f = f->link;
        if (f->waiters < flow->waiters_per_chunk) {
            kmem_free(flow1, sizeof (*flow1));
            return (f);
        }
        f->link = flow1;
    } else {
        flow->list = flow1;
    }
    cv_init(&flow1->cv, NULL, CV_DRIVER, NULL);
    flow1->waiters = 0;
    flow1->link = NULL;
    return (flow1);
}

static void
ibcm_flow_enter(ibcm_flow_t *flow)
{
    mutex_enter(&ibcm_open.mutex);
    if (flow->list == NULL && flow->simul < flow->simul_max) {
        flow->simul++;
        flow->total++;
        mutex_exit(&ibcm_open.mutex);
    } else {
        ibcm_flow1_t *flow1;

        flow1 = ibcm_flow_find(flow);
        flow1->waiters++;
        cv_wait(&flow1->cv, &ibcm_open.mutex);
        if (--flow1->waiters == 0) {
            cv_destroy(&flow1->cv);
            mutex_exit(&ibcm_open.mutex);
            kmem_free(flow1, sizeof (*flow1));
        } else
            mutex_exit(&ibcm_open.mutex);
    }
}

static void
ibcm_flow_exit(ibcm_flow_t *flow)
{
    mutex_enter(&ibcm_open.mutex);
    if (--flow->simul < flow->lowat) {
        if (flow->lowat < flow->lowat_default)
            flow->lowat++;
        if (flow->list) {
            ibcm_flow1_t *flow1;

            flow1 = flow->list;
            flow->list = flow1->link;   /* unlink */
            flow1->link = NULL;     /* be clean */
            flow->total += flow1->waiters;
            flow->simul += flow1->waiters;
            cv_broadcast(&flow1->cv);
        }
    }
    mutex_exit(&ibcm_open.mutex);
}

void
ibcm_flow_inc(void)
{
    mutex_enter(&ibcm_open.mutex);
    if (++ibcm_open.sends > ibcm_open.sends_max) {
        ibcm_open.sends_max = ibcm_open.sends;
        IBTF_DPRINTF_L2(cmlog, "ibcm_flow_inc: sends max = %d",
            ibcm_open.sends_max);
    }
    mutex_exit(&ibcm_open.mutex);
}

static void
ibcm_check_send_cmpltn_time(hrtime_t delta, char *event_msg)
{
    if (delta > 4000000LL) {
        IBTF_DPRINTF_L2(cmlog, "ibcm_check_send_cmpltn_time: "
            "%s: %lldns", event_msg, delta);
    }
}

void
ibcm_flow_dec(hrtime_t time, char *mad_type)
{
    int flow_exit = 0;
    int run = 0;

    if (ibcm_dtrace)
        ibcm_check_send_cmpltn_time(gethrtime() - time, mad_type);
    mutex_enter(&ibcm_open.mutex);
    ibcm_open.sends--;
    if (ibcm_open.sends < ibcm_open.sends_lowat) {
        run = ibcm_ok_to_start(&ibcm_open);
        if (ibcm_open.exit_deferred) {
            ibcm_open.exit_deferred--;
            flow_exit = 1;
        }
    }
    mutex_exit(&ibcm_open.mutex);
    if (flow_exit)
        ibcm_flow_exit(&ibcm_close_flow);
    if (run)
        ibcm_run_tlist_thread();
}

void
ibcm_close_enqueue(ibcm_state_data_t *statep)
{
    mutex_enter(&ibcm_close.mutex);
    statep->close_link = NULL;
    ibcm_close.tail->close_link = statep;
    ibcm_close.tail = statep;
    mutex_exit(&ibcm_close.mutex);
    ibcm_run_tlist_thread();
}

void
ibcm_check_for_async_close()
{
    ibcm_state_data_t   *statep;

    mutex_enter(&ibcm_close.mutex);

    while (ibcm_close.head.close_link) {
        statep = ibcm_close.head.close_link;
        ibcm_close.head.close_link = statep->close_link;
        statep->close_link = NULL;
        if (ibcm_close.tail == statep)
            ibcm_close.tail = &ibcm_close.head;
        mutex_exit(&ibcm_close.mutex);
        ibcm_close_start(statep);
        mutex_enter(&ibcm_close.mutex);
    }
    mutex_exit(&ibcm_close.mutex);
}

void
ibcm_close_enter(void)
{
    ibcm_flow_enter(&ibcm_close_flow);
}

void
ibcm_close_exit(void)
{
    int flow_exit;

    mutex_enter(&ibcm_open.mutex);
    if (ibcm_open.sends < ibcm_open.sends_lowat ||
        ibcm_open.exit_deferred >= 4)
        flow_exit = 1;
    else {
        flow_exit = 0;
        ibcm_open.exit_deferred++;
    }
    mutex_exit(&ibcm_open.mutex);
    if (flow_exit)
        ibcm_flow_exit(&ibcm_close_flow);
}

/*
 * This function needs to be called twice to finish our flow
 * control accounting when closing down a connection.  One
 * call has send_done set to 1, while the other has it set to 0.
 * Because of retries, this could get called more than once
 * with either 0 or 1, but additional calls have no effect.
 */
void
ibcm_close_done(ibcm_state_data_t *statep, int send_done)
{
    int flow_exit;

    ASSERT(MUTEX_HELD(&statep->state_mutex));
    if (statep->close_flow == 1) {
        if (send_done)
            statep->close_flow = 3;
        else
            statep->close_flow = 2;
    } else if ((send_done && statep->close_flow == 2) ||
        (!send_done && statep->close_flow == 3)) {
        statep->close_flow = 0;
        mutex_enter(&ibcm_open.mutex);
        if (ibcm_open.sends < ibcm_open.sends_lowat ||
            ibcm_open.exit_deferred >= 4)
            flow_exit = 1;
        else {
            flow_exit = 0;
            ibcm_open.exit_deferred++;
        }
        mutex_exit(&ibcm_open.mutex);
        if (flow_exit)
            ibcm_flow_exit(&ibcm_close_flow);
    }
}

void
ibcm_lapr_enter(void)
{
    ibcm_flow_enter(&ibcm_lapr_flow);
}

void
ibcm_lapr_exit(void)
{
    ibcm_flow_exit(&ibcm_lapr_flow);
}

void
ibcm_sa_access_enter()
{
    ibcm_flow_enter(&ibcm_saa_flow);
}

void
ibcm_sa_access_exit()
{
    ibcm_flow_exit(&ibcm_saa_flow);
}

static void
ibcm_sm_notice_handler(ibmf_saa_handle_t saa_handle,
    ibmf_saa_subnet_event_t saa_event_code,
    ibmf_saa_event_details_t *saa_event_details,
    void *callback_arg)
{
    ibcm_port_info_t    *portp = (ibcm_port_info_t *)callback_arg;
    ibt_subnet_event_code_t code;
    ibt_subnet_event_t  event;
    uint8_t         event_status;

    IBTF_DPRINTF_L3(cmlog, "ibcm_sm_notice_handler: saa_hdl %p, code = %d",
        saa_handle, saa_event_code);

    mutex_enter(&ibcm_sm_notice_serialize_lock);

    switch (saa_event_code) {
    case IBMF_SAA_EVENT_MCG_CREATED:
        code = IBT_SM_EVENT_MCG_CREATED;
        break;
    case IBMF_SAA_EVENT_MCG_DELETED:
        code = IBT_SM_EVENT_MCG_DELETED;
        break;
    case IBMF_SAA_EVENT_GID_AVAILABLE:
        code = IBT_SM_EVENT_GID_AVAIL;
        ibcm_path_cache_purge();
        break;
    case IBMF_SAA_EVENT_GID_UNAVAILABLE:
        code = IBT_SM_EVENT_GID_UNAVAIL;
        ibcm_path_cache_purge();
        break;
    case IBMF_SAA_EVENT_SUBSCRIBER_STATUS_CHG:
        event_status =
            saa_event_details->ie_producer_event_status_mask &
            IBMF_SAA_EVENT_STATUS_MASK_PRODUCER_SM;
        if (event_status == (portp->port_event_status &
            IBMF_SAA_EVENT_STATUS_MASK_PRODUCER_SM)) {
            mutex_exit(&ibcm_sm_notice_serialize_lock);
            return; /* no change */
        }
        portp->port_event_status = event_status;
        if (event_status == IBMF_SAA_EVENT_STATUS_MASK_PRODUCER_SM)
            code = IBT_SM_EVENT_AVAILABLE;
        else
            code = IBT_SM_EVENT_UNAVAILABLE;
        break;
    default:
        mutex_exit(&ibcm_sm_notice_serialize_lock);
        return;
    }

    mutex_enter(&ibcm_global_hca_lock);

    /* don't send the event if we're tearing down */
    if (!IBCM_ACCESS_HCA_OK(portp->port_hcap)) {
        mutex_exit(&ibcm_global_hca_lock);
        mutex_exit(&ibcm_sm_notice_serialize_lock);
        return;
    }

    ++(portp->port_hcap->hca_acc_cnt);
    mutex_exit(&ibcm_global_hca_lock);

    event.sm_notice_gid = saa_event_details->ie_gid;
    ibtl_cm_sm_notice_handler(portp->port_sgid0, code, &event);

    mutex_exit(&ibcm_sm_notice_serialize_lock);

    ibcm_dec_hca_acc_cnt(portp->port_hcap);
}

void
ibt_register_subnet_notices(ibt_clnt_hdl_t ibt_hdl,
    ibt_sm_notice_handler_t sm_notice_handler, void *private)
{
    ibcm_port_info_t    *portp;
    ibcm_hca_info_t     *hcap;
    uint8_t         port;
    int         num_failed_sgids;
    ibtl_cm_sm_init_fail_t  *ifail;
    ib_gid_t        *sgidp;

    IBTF_DPRINTF_L3(cmlog, "ibt_register_subnet_notices: ibt_hdl = %p",
        ibt_hdl);

    mutex_enter(&ibcm_sm_notice_serialize_lock);

    ibtl_cm_set_sm_notice_handler(ibt_hdl, sm_notice_handler, private);
    if (sm_notice_handler == NULL) {
        mutex_exit(&ibcm_sm_notice_serialize_lock);
        return;
    }

    /* for each port, if service is not available, make a call */
    mutex_enter(&ibcm_global_hca_lock);
    num_failed_sgids = 0;
    hcap = ibcm_hca_listp;
    while (hcap != NULL) {
        portp = hcap->hca_port_info;
        for (port = 0; port < hcap->hca_num_ports; port++) {
            if (!(portp->port_event_status &
                IBMF_SAA_EVENT_STATUS_MASK_PRODUCER_SM))
                num_failed_sgids++;
            portp++;
        }
        hcap = hcap->hca_next;
    }
    if (num_failed_sgids != 0) {
        ifail = kmem_alloc(sizeof (*ifail) +
            (num_failed_sgids - 1) * sizeof (ib_gid_t), KM_SLEEP);
        ifail->smf_num_sgids = num_failed_sgids;
        ifail->smf_ibt_hdl = ibt_hdl;
        sgidp = &ifail->smf_sgid[0];
        hcap = ibcm_hca_listp;
        while (hcap != NULL) {
            portp = hcap->hca_port_info;
            for (port = 0; port < hcap->hca_num_ports; port++) {
                if (!(portp->port_event_status &
                    IBMF_SAA_EVENT_STATUS_MASK_PRODUCER_SM))
                    *sgidp++ = portp->port_sgid0;
                portp++;
            }
            hcap = hcap->hca_next;
        }
    }
    mutex_exit(&ibcm_global_hca_lock);

    if (num_failed_sgids != 0) {
        ibtl_cm_sm_notice_init_failure(ifail);
        kmem_free(ifail, sizeof (*ifail) +
            (num_failed_sgids - 1) * sizeof (ib_gid_t));
    }
    mutex_exit(&ibcm_sm_notice_serialize_lock);
}

/* The following is run from a taskq because we've seen the stack overflow. */
static void
ibcm_init_saa(void *arg)
{
    ibcm_port_info_t        *portp = (ibcm_port_info_t *)arg;
    int             status;
    ib_guid_t           port_guid;
    ibmf_saa_subnet_event_args_t    event_args;

    port_guid = portp->port_sgid0.gid_guid;

    IBTF_DPRINTF_L3(cmlog, "ibcm_init_saa: port guid %llX", port_guid);

    event_args.is_event_callback_arg = portp;
    event_args.is_event_callback = ibcm_sm_notice_handler;

    if ((status = ibmf_sa_session_open(port_guid, 0, &event_args,
        IBMF_VERSION, 0, &portp->port_ibmf_saa_hdl)) != IBMF_SUCCESS) {
        IBTF_DPRINTF_L2(cmlog, "ibcm_init_saa: "
            "ibmf_sa_session_open failed for port guid %llX "
            "status = %d", port_guid, status);
    } else {
        IBTF_DPRINTF_L2(cmlog, "ibcm_init_saa: "
            "registered sa_hdl 0x%p for port guid %llX",
            portp->port_ibmf_saa_hdl, port_guid);
    }

    mutex_enter(&ibcm_sa_open_lock);
    portp->port_saa_open_in_progress = 0;
    cv_broadcast(&ibcm_sa_open_cv);
    mutex_exit(&ibcm_sa_open_lock);
}

void
ibcm_init_saa_handle(ibcm_hca_info_t *hcap, uint8_t port)
{
    ibmf_saa_handle_t   saa_handle;
    uint8_t         port_index = port - 1;
    ibcm_port_info_t    *portp = &hcap->hca_port_info[port_index];
    ibt_status_t        ibt_status;

    if (port_index >= hcap->hca_num_ports)
        return;

    mutex_enter(&ibcm_sa_open_lock);
    if (portp->port_saa_open_in_progress) {
        mutex_exit(&ibcm_sa_open_lock);
        return;
    }

    saa_handle = portp->port_ibmf_saa_hdl;
    if (saa_handle != NULL) {
        mutex_exit(&ibcm_sa_open_lock);
        return;
    }

    portp->port_saa_open_in_progress = 1;
    mutex_exit(&ibcm_sa_open_lock);

    _NOTE(NOW_INVISIBLE_TO_OTHER_THREADS(portp->port_event_status))

    /* The assumption is that we're getting event notifications */
    portp->port_event_status = IBMF_SAA_EVENT_STATUS_MASK_PRODUCER_SM;

    _NOTE(NOW_VISIBLE_TO_OTHER_THREADS(portp->port_event_status))

    ibt_status = ibt_get_port_state_byguid(portp->port_hcap->hca_guid,
        portp->port_num, &portp->port_sgid0, NULL);
    if (ibt_status != IBT_SUCCESS) {
        IBTF_DPRINTF_L2(cmlog, "ibcm_init_saa_handle: "
            "ibt_get_port_state_byguid failed for guid %llX "
            "with status %d", portp->port_hcap->hca_guid, ibt_status);
        mutex_enter(&ibcm_sa_open_lock);
        portp->port_saa_open_in_progress = 0;
        cv_broadcast(&ibcm_sa_open_cv);
        mutex_exit(&ibcm_sa_open_lock);
        return;
    }
    /* if the port is UP, try sa_session_open */
    (void) taskq_dispatch(ibcm_taskq, ibcm_init_saa, portp, TQ_SLEEP);
}


ibmf_saa_handle_t
ibcm_get_saa_handle(ibcm_hca_info_t *hcap, uint8_t port)
{
    ibmf_saa_handle_t   saa_handle;
    uint8_t         port_index = port - 1;
    ibcm_port_info_t    *portp = &hcap->hca_port_info[port_index];
    ibt_status_t        ibt_status;

    if (port_index >= hcap->hca_num_ports)
        return (NULL);

    mutex_enter(&ibcm_sa_open_lock);
    while (portp->port_saa_open_in_progress) {
        cv_wait(&ibcm_sa_open_cv, &ibcm_sa_open_lock);
    }

    saa_handle = portp->port_ibmf_saa_hdl;
    if (saa_handle != NULL) {
        mutex_exit(&ibcm_sa_open_lock);
        return (saa_handle);
    }

    portp->port_saa_open_in_progress = 1;
    mutex_exit(&ibcm_sa_open_lock);

    ibt_status = ibt_get_port_state_byguid(portp->port_hcap->hca_guid,
        portp->port_num, &portp->port_sgid0, NULL);
    if (ibt_status != IBT_SUCCESS) {
        IBTF_DPRINTF_L2(cmlog, "ibcm_get_saa_handle: "
            "ibt_get_port_state_byguid failed for guid %llX "
            "with status %d", portp->port_hcap->hca_guid, ibt_status);
        mutex_enter(&ibcm_sa_open_lock);
        portp->port_saa_open_in_progress = 0;
        cv_broadcast(&ibcm_sa_open_cv);
        mutex_exit(&ibcm_sa_open_lock);
        return (NULL);
    }
    /* if the port is UP, try sa_session_open */
    (void) taskq_dispatch(ibcm_taskq, ibcm_init_saa, portp, TQ_SLEEP);

    mutex_enter(&ibcm_sa_open_lock);
    while (portp->port_saa_open_in_progress) {
        cv_wait(&ibcm_sa_open_cv, &ibcm_sa_open_lock);
    }
    saa_handle = portp->port_ibmf_saa_hdl;
    mutex_exit(&ibcm_sa_open_lock);
    return (saa_handle);
}


/*
 * ibcm_hca_init_port():
 *  - Register port with IBMA
 *
 * Arguments:
 *  hcap        - HCA's guid
 *  port_index  - port number minus 1
 *
 * Return values:
 *  IBCM_SUCCESS - success
 */
ibt_status_t
ibcm_hca_init_port(ibcm_hca_info_t *hcap, uint8_t port_index)
{
    int         status;
    ibmf_register_info_t    *ibmf_reg;

    IBTF_DPRINTF_L4(cmlog, "ibcm_hca_init_port: hcap = 0x%p port_num %d",
        hcap, port_index + 1);

    ASSERT(MUTEX_HELD(&ibcm_global_hca_lock));

    _NOTE(NOW_INVISIBLE_TO_OTHER_THREADS(hcap->hca_port_info))

    if (hcap->hca_port_info[port_index].port_ibmf_hdl == NULL) {
        /* Register with IBMF */
        ibmf_reg = &hcap->hca_port_info[port_index].port_ibmf_reg;
        ibmf_reg->ir_ci_guid = hcap->hca_guid;
        ibmf_reg->ir_port_num = port_index + 1;
        ibmf_reg->ir_client_class = COMM_MGT_MANAGER_AGENT;

        /*
         * register with management framework
         */
        status = ibmf_register(ibmf_reg, IBMF_VERSION,
            IBMF_REG_FLAG_NO_OFFLOAD, NULL, NULL,
            &(hcap->hca_port_info[port_index].port_ibmf_hdl),
            &(hcap->hca_port_info[port_index].port_ibmf_caps));

        if (status != IBMF_SUCCESS) {
            IBTF_DPRINTF_L2(cmlog, "ibcm_hca_init_port: "
                "ibmf_register failed for port_num %x, "
                "status = %x", port_index + 1, status);
            return (ibcm_ibmf_analyze_error(status));
        }

        hcap->hca_port_info[port_index].port_qp1.qp_cm =
            IBMF_QP_HANDLE_DEFAULT;
        hcap->hca_port_info[port_index].port_qp1.qp_port =
            &(hcap->hca_port_info[port_index]);

        /*
         * Register the read callback with IBMF.
         * Since we just did an ibmf_register, handle is
         * valid and ibcm_recv_cb() is valid so we can
         * safely assert for success of ibmf_setup_recv_cb()
         *
         * Depending on the "state" of the HCA,
         * CM may drop incoming packets
         */
        status = ibmf_setup_async_cb(
            hcap->hca_port_info[port_index].port_ibmf_hdl,
            IBMF_QP_HANDLE_DEFAULT, ibcm_recv_cb,
            &(hcap->hca_port_info[port_index].port_qp1), 0);
        ASSERT(status == IBMF_SUCCESS);

        IBTF_DPRINTF_L5(cmlog, "ibcm_hca_init_port: "
            "IBMF hdl[%x] = 0x%p", port_index,
            hcap->hca_port_info[port_index].port_ibmf_hdl);

        /* Attempt to get the saa_handle for this port */
        ibcm_init_saa_handle(hcap, port_index + 1);
    }

    return (IBT_SUCCESS);
}

/*
 * useful, to re attempt to initialize port ibma handles from elsewhere in
 * cm code
 */
ibt_status_t
ibcm_hca_reinit_port(ibcm_hca_info_t *hcap, uint8_t port_index)
{
    ibt_status_t    status;

    IBTF_DPRINTF_L5(cmlog, "ibcm_hca_reinit_port: hcap 0x%p port_num %d",
        hcap, port_index + 1);

    mutex_enter(&ibcm_global_hca_lock);
    status = ibcm_hca_init_port(hcap, port_index);
    mutex_exit(&ibcm_global_hca_lock);
    return (status);
}


/*
 * ibcm_hca_fini_port():
 *  - Deregister port with IBMA
 *
 * Arguments:
 *  hcap        - HCA's guid
 *  port_index  - port number minus 1
 *
 * Return values:
 *  IBCM_SUCCESS - success
 */
static ibcm_status_t
ibcm_hca_fini_port(ibcm_hca_info_t *hcap, uint8_t port_index)
{
    int         ibmf_status;
    ibcm_status_t       ibcm_status;

    IBTF_DPRINTF_L4(cmlog, "ibcm_hca_fini_port: hcap = 0x%p port_num %d ",
        hcap, port_index + 1);

    ASSERT(MUTEX_HELD(&ibcm_global_hca_lock));

    if (hcap->hca_port_info[port_index].port_ibmf_saa_hdl != NULL) {
        IBTF_DPRINTF_L5(cmlog, "ibcm_hca_fini_port: "
            "ibmf_sa_session_close IBMF SAA hdl %p",
            hcap->hca_port_info[port_index].port_ibmf_saa_hdl);

        ibmf_status = ibmf_sa_session_close(
            &hcap->hca_port_info[port_index].port_ibmf_saa_hdl, 0);
        if (ibmf_status != IBMF_SUCCESS) {
            IBTF_DPRINTF_L2(cmlog, "ibcm_hca_fini_port: "
                "ibmf_sa_session_close of port %d returned %x",
                port_index + 1, ibmf_status);
            return (IBCM_FAILURE);
        }
    }

    if (hcap->hca_port_info[port_index].port_ibmf_hdl != NULL) {
        IBTF_DPRINTF_L5(cmlog, "ibcm_hca_fini_port: "
            "ibmf_unregister IBMF Hdl %p",
            hcap->hca_port_info[port_index].port_ibmf_hdl);

        /* clean-up all the ibmf qp's allocated on this port */
        ibcm_status = ibcm_free_allqps(hcap, port_index + 1);

        if (ibcm_status != IBCM_SUCCESS) {

            IBTF_DPRINTF_L2(cmlog, "ibcm_hca_fini_port: "
                "ibcm_free_allqps failed for port_num %d",
                port_index + 1);
            return (IBCM_FAILURE);
        }

        /* Tear down the receive callback */
        ibmf_status = ibmf_tear_down_async_cb(
            hcap->hca_port_info[port_index].port_ibmf_hdl,
            IBMF_QP_HANDLE_DEFAULT, 0);

        if (ibmf_status != IBMF_SUCCESS) {
            IBTF_DPRINTF_L2(cmlog, "ibcm_hca_fini_port: "
                "ibmf_tear_down_async_cb failed %d port_num %d",
                ibmf_status, port_index + 1);
            return (IBCM_FAILURE);
        }

        /* Now, unregister with IBMF */
        ibmf_status = ibmf_unregister(
            &hcap->hca_port_info[port_index].port_ibmf_hdl, 0);
        IBTF_DPRINTF_L4(cmlog, "ibcm_hca_fini_port: "
            "ibmf_unregister of port_num %x returned %x",
            port_index + 1, ibmf_status);

        if (ibmf_status == IBMF_SUCCESS)
            hcap->hca_port_info[port_index].port_ibmf_hdl = NULL;
        else {
            IBTF_DPRINTF_L2(cmlog, "ibcm_hca_fini_port: "
                "ibmf_unregister failed %d port_num %d",
                ibmf_status, port_index + 1);
            return (IBCM_FAILURE);
        }
    }
    return (IBCM_SUCCESS);
}

/*
 * ibcm_comm_est_handler():
 *  Check if the given channel is in ESTABLISHED state or not
 *
 * Arguments:
 *  eventp  - A pointer to an ibt_async_event_t struct
 *
 * Return values: NONE
 */
static void
ibcm_comm_est_handler(ibt_async_event_t *eventp)
{
    ibcm_state_data_t   *statep;

    IBTF_DPRINTF_L4(cmlog, "ibcm_comm_est_handler:");

    /* Both QP and EEC handles can't be NULL */
    if (eventp->ev_chan_hdl == NULL) {
        IBTF_DPRINTF_L2(cmlog, "ibcm_comm_est_handler: "
            "both QP and EEC handles are NULL");
        return;
    }

    /* get the "statep" from qp/eec handles */
    IBCM_GET_CHAN_PRIVATE(eventp->ev_chan_hdl, statep);
    if (statep == NULL) {
        IBTF_DPRINTF_L2(cmlog, "ibcm_comm_est_handler: statep is NULL");
        return;
    }

    mutex_enter(&statep->state_mutex);

    IBCM_RELEASE_CHAN_PRIVATE(eventp->ev_chan_hdl);

    IBTF_DPRINTF_L4(cmlog, "ibcm_comm_est_handler: statep = %p", statep);

    IBCM_REF_CNT_INCR(statep);

    if ((statep->state == IBCM_STATE_REP_SENT) ||
        (statep->state == IBCM_STATE_MRA_REP_RCVD)) {
        timeout_id_t    timer_val = statep->timerid;

        statep->state = IBCM_STATE_TRANSIENT_ESTABLISHED;

        if (timer_val) {
            statep->timerid = 0;
            mutex_exit(&statep->state_mutex);
            (void) untimeout(timer_val);
        } else
            mutex_exit(&statep->state_mutex);

        /* CM doesn't have RTU message here */
        ibcm_cep_state_rtu(statep, NULL);

    } else {
        if (statep->state == IBCM_STATE_ESTABLISHED ||
            statep->state == IBCM_STATE_TRANSIENT_ESTABLISHED) {
            IBTF_DPRINTF_L4(cmlog, "ibcm_comm_est_handler: "
                "Channel already in ESTABLISHED state");
        } else {
            /* An unexpected behavior from remote */
            IBTF_DPRINTF_L2(cmlog, "ibcm_comm_est_handler: "
                "Unexpected in state = %d", statep->state);
        }
        mutex_exit(&statep->state_mutex);

        ibcm_insert_trace(statep, IBCM_TRACE_INCOMING_COMEST);
    }

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


/*
 * ibcm_async_handler():
 *  CM's Async Handler
 *  (Handles ATTACH, DETACH, COM_EST events)
 *
 * Arguments:
 *  eventp  - A pointer to an ibt_async_event_t struct
 *
 * Return values: None
 *
 * NOTE : CM assumes that all HCA DR events are delivered sequentially
 * i.e., until ibcm_async_handler  completes for a given HCA DR, framework
 * shall not invoke ibcm_async_handler with another DR event for the same
 * HCA
 */
/* ARGSUSED */
void
ibcm_async_handler(void *clnt_hdl, ibt_hca_hdl_t hca_hdl,
    ibt_async_code_t code, ibt_async_event_t *eventp)
{
    ibcm_hca_info_t     *hcap;
    ibcm_port_up_t      *pup;

    IBTF_DPRINTF_L3(cmlog, "ibcm_async_handler: "
        "clnt_hdl = %p, code = 0x%x, eventp = 0x%p",
        clnt_hdl, code, eventp);

    mutex_enter(&ibcm_global_hca_lock);

    /* If fini is going to complete successfully, then return */
    if (ibcm_finit_state != IBCM_FINIT_IDLE) {

        /*
         * This finit state implies one of the following:
         * Init either didn't start or didn't complete OR
         * Fini is about to return SUCCESS and release the global lock.
         * In all these cases, it is safe to ignore the async.
         */

        IBTF_DPRINTF_L2(cmlog, "ibcm_async_handler: ignoring event %x, "
            "as either init didn't complete or fini about to succeed",
            code);
        mutex_exit(&ibcm_global_hca_lock);
        return;
    }

    switch (code) {
    case IBT_EVENT_PORT_UP:
        mutex_exit(&ibcm_global_hca_lock);
        pup = kmem_alloc(sizeof (ibcm_port_up_t), KM_SLEEP);
        _NOTE(NOW_INVISIBLE_TO_OTHER_THREADS(*pup))
        pup->pup_hca_guid = eventp->ev_hca_guid;
        pup->pup_port = eventp->ev_port;
        _NOTE(NOW_VISIBLE_TO_OTHER_THREADS(*pup))
        (void) taskq_dispatch(ibcm_taskq,
            ibcm_service_record_rewrite_task, pup, TQ_SLEEP);
        ibcm_path_cache_purge();
        return;

    case IBT_HCA_ATTACH_EVENT:

        /* eventp->ev_hcaguid is the HCA GUID of interest */
        ibcm_hca_attach(eventp->ev_hca_guid);
        break;

    case IBT_HCA_DETACH_EVENT:

        /* eventp->ev_hca_guid is the HCA GUID of interest */
        if ((hcap = ibcm_find_hcap_entry(eventp->ev_hca_guid)) ==
            NULL) {
            IBTF_DPRINTF_L2(cmlog, "ibcm_async_handler:"
                " hca %llX doesn't exist", eventp->ev_hca_guid);
            break;
        }

        (void) ibcm_hca_detach(hcap);
        break;

    case IBT_EVENT_COM_EST_QP:
        /* eventp->ev_qp_hdl is the ibt_qp_hdl_t of interest */
    case IBT_EVENT_COM_EST_EEC:
        /* eventp->ev_eec_hdl is the ibt_eec_hdl_t of interest */
        ibcm_comm_est_handler(eventp);
        break;
    default:
        break;
    }

    /* Unblock, any blocked fini/init operations */
    mutex_exit(&ibcm_global_hca_lock);
}