sun4v/io/dr_cpu.c

	dr_cpu.c revision 99c7e85592b5d673f1469b2767f876710004dc23
/*
 * 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.
 */

/*
 * sun4v CPU DR Module
 */

#include <sys/modctl.h>
#include <sys/processor.h>
#include <sys/cpuvar.h>
#include <sys/cpupart.h>
#include <sys/sunddi.h>
#include <sys/sunndi.h>
#include <sys/note.h>
#include <sys/sysevent/dr.h>
#include <sys/hypervisor_api.h>
#include <sys/mach_descrip.h>
#include <sys/mdesc.h>
#include <sys/ds.h>
#include <sys/drctl.h>
#include <sys/dr_util.h>
#include <sys/dr_cpu.h>
#include <sys/promif.h>
#include <sys/machsystm.h>


static struct modlmisc modlmisc = {
    &mod_miscops,
    "sun4v CPU DR"
};

static struct modlinkage modlinkage = {
    MODREV_1,
    (void *)&modlmisc,
    NULL
};

typedef int (*fn_t)(processorid_t, int *, boolean_t);

/*
 * Global DS Handle
 */
static ds_svc_hdl_t ds_handle;

/*
 * Supported DS Capability Versions
 */
static ds_ver_t     dr_cpu_vers[] = { { 1, 1 }, { 1, 0 } };
#define DR_CPU_NVERS    (sizeof (dr_cpu_vers) / sizeof (dr_cpu_vers[0]))

static ds_ver_t     version;

/*
 * DS Capability Description
 */
static ds_capability_t dr_cpu_cap = {
    DR_CPU_DS_ID,       /* svc_id */
    dr_cpu_vers,        /* vers */
    DR_CPU_NVERS        /* nvers */
};

#define DRCPU_VERS_EQ(_maj, _min) \
    ((version.major == (_maj)) && (version.minor == (_min)))

#define DRCPU_VERS_GTEQ(_maj, _min) \
    ((version.major > (_maj)) ||                    \
    ((version.major == (_maj)) && (version.minor >= (_min))))

/*
 * DS Callbacks
 */
static void dr_cpu_reg_handler(ds_cb_arg_t, ds_ver_t *, ds_svc_hdl_t);
static void dr_cpu_unreg_handler(ds_cb_arg_t arg);
static void dr_cpu_data_handler(ds_cb_arg_t arg, void *buf, size_t buflen);

/*
 * DS Client Ops Vector
 */
static ds_clnt_ops_t dr_cpu_ops = {
    dr_cpu_reg_handler, /* ds_reg_cb */
    dr_cpu_unreg_handler,   /* ds_unreg_cb */
    dr_cpu_data_handler,    /* ds_data_cb */
    NULL            /* cb_arg */
};

/*
 * Operation Results
 *
 * Used internally to gather results while an operation on a
 * list of CPUs is in progress. In particular, it is used to
 * keep track of which CPUs have already failed so that they are
 * not processed further, and the manner in which they failed.
 */
typedef struct {
    uint32_t    cpuid;
    uint32_t    result;
    uint32_t    status;
    char        *string;
} dr_cpu_res_t;

#define DR_CPU_MAX_ERR_LEN  64  /* maximum error string length */

/*
 * Internal Functions
 */
static int dr_cpu_init(void);
static int dr_cpu_fini(void);

static int dr_cpu_list_wrk(dr_cpu_hdr_t *, dr_cpu_hdr_t **, int *);
static int dr_cpu_list_status(dr_cpu_hdr_t *, dr_cpu_hdr_t **, int *);

static int dr_cpu_unconfigure(processorid_t, int *status, boolean_t force);
static int dr_cpu_configure(processorid_t, int *status, boolean_t force);
static int dr_cpu_status(processorid_t, int *status);

static void dr_cpu_check_cpus(dr_cpu_hdr_t *req, dr_cpu_res_t *res);
static void dr_cpu_check_psrset(uint32_t *cpuids, dr_cpu_res_t *res, int nres);
static int dr_cpu_check_bound_thr(cpu_t *cp, dr_cpu_res_t *res);

static dr_cpu_res_t *dr_cpu_res_array_init(dr_cpu_hdr_t *, drctl_rsrc_t *, int);
static void dr_cpu_res_array_fini(dr_cpu_res_t *res, int nres);
static size_t dr_cpu_pack_response(dr_cpu_hdr_t *req, dr_cpu_res_t *res,
    dr_cpu_hdr_t **respp);

static int dr_cpu_probe(processorid_t newcpuid);
static int dr_cpu_deprobe(processorid_t cpuid);

static dev_info_t *dr_cpu_find_node(processorid_t cpuid);
static mde_cookie_t dr_cpu_find_node_md(processorid_t, md_t *, mde_cookie_t *);

int
_init(void)
{
    int status;

    /* check that CPU DR is enabled */
    if (dr_is_disabled(DR_TYPE_CPU)) {
        cmn_err(CE_CONT, "!CPU DR is disabled\n");
        return (-1);
    }

    if ((status = dr_cpu_init()) != 0) {
        cmn_err(CE_NOTE, "CPU DR initialization failed");
        return (status);
    }

    if ((status = mod_install(&modlinkage)) != 0) {
        (void) dr_cpu_fini();
    }

    return (status);
}

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

int dr_cpu_allow_unload;

int
_fini(void)
{
    int status;

    if (dr_cpu_allow_unload == 0)
        return (EBUSY);

    if ((status = mod_remove(&modlinkage)) == 0) {
        (void) dr_cpu_fini();
    }

    return (status);
}

static int
dr_cpu_init(void)
{
    int rv;

    if ((rv = ds_cap_init(&dr_cpu_cap, &dr_cpu_ops)) != 0) {
        cmn_err(CE_NOTE, "ds_cap_init failed: %d", rv);
        return (-1);
    }

    return (0);
}

static int
dr_cpu_fini(void)
{
    int rv;

    if ((rv = ds_cap_fini(&dr_cpu_cap)) != 0) {
        cmn_err(CE_NOTE, "ds_cap_fini failed: %d", rv);
        return (-1);
    }

    return (0);
}

static void
dr_cpu_reg_handler(ds_cb_arg_t arg, ds_ver_t *ver, ds_svc_hdl_t hdl)
{
    DR_DBG_CPU("reg_handler: arg=0x%p, ver=%d.%d, hdl=0x%lx\n", arg,
        ver->major, ver->minor, hdl);

    version.major = ver->major;
    version.minor = ver->minor;
    ds_handle = hdl;
}

static void
dr_cpu_unreg_handler(ds_cb_arg_t arg)
{
    DR_DBG_CPU("unreg_handler: arg=0x%p\n", arg);

    ds_handle = DS_INVALID_HDL;
}

static void
dr_cpu_data_handler(ds_cb_arg_t arg, void *buf, size_t buflen)
{
    _NOTE(ARGUNUSED(arg))

    dr_cpu_hdr_t    *req = buf;
    dr_cpu_hdr_t    err_resp;
    dr_cpu_hdr_t    *resp = &err_resp;
    int     resp_len = 0;
    int     rv;

    /*
     * Sanity check the message
     */
    if (buflen < sizeof (dr_cpu_hdr_t)) {
        DR_DBG_CPU("incoming message short: expected at least %ld "
            "bytes, received %ld\n", sizeof (dr_cpu_hdr_t), buflen);
        goto done;
    }

    if (req == NULL) {
        DR_DBG_CPU("empty message: expected at least %ld bytes\n",
            sizeof (dr_cpu_hdr_t));
        goto done;
    }

    DR_DBG_CPU("incoming request:\n");
    DR_DBG_DUMP_MSG(buf, buflen);

    if (req->num_records > NCPU) {
        DR_DBG_CPU("CPU list too long: %d when %d is the maximum\n",
            req->num_records, NCPU);
        goto done;
    }

    if (req->num_records == 0) {
        DR_DBG_CPU("No CPU specified for operation\n");
        goto done;
    }

    /*
     * Process the command
     */
    switch (req->msg_type) {
    case DR_CPU_CONFIGURE:
    case DR_CPU_UNCONFIGURE:
    case DR_CPU_FORCE_UNCONFIG:
        if ((rv = dr_cpu_list_wrk(req, &resp, &resp_len)) != 0) {
            DR_DBG_CPU("%s%s failed (%d)\n",
                (req->msg_type == DR_CPU_CONFIGURE) ?
                "CPU configure" : "CPU unconfigure",
                (req->msg_type == DR_CPU_FORCE_UNCONFIG) ?
                " (forced)" : "", rv);
        }
        break;

    case DR_CPU_STATUS:
        if ((rv = dr_cpu_list_status(req, &resp, &resp_len)) != 0)
            DR_DBG_CPU("CPU status failed (%d)\n", rv);
        break;

    default:
        cmn_err(CE_NOTE, "unsupported DR operation (%d)",
            req->msg_type);
        break;
    }

done:
    /* check if an error occurred */
    if (resp == &err_resp) {
        resp->req_num = (req) ? req->req_num : 0;
        resp->msg_type = DR_CPU_ERROR;
        resp->num_records = 0;
        resp_len = sizeof (dr_cpu_hdr_t);
    }

    DR_DBG_CPU("outgoing response:\n");
    DR_DBG_DUMP_MSG(resp, resp_len);

    /* send back the response */
    if (ds_cap_send(ds_handle, resp, resp_len) != 0) {
        DR_DBG_CPU("ds_send failed\n");
    }

    /* free any allocated memory */
    if (DRCPU_VERS_GTEQ(1, 1) || (resp != &err_resp)) {
        DR_DBG_KMEM("%s: free addr %p size %d\n",
            __func__, (void *)resp, resp_len);
        kmem_free(resp, resp_len);
    }
}

/*
 * Create a response message which consists of a header followed
 * by the error string passed in.
 */
static size_t
dr_cpu_err_resp(dr_cpu_hdr_t *req, dr_cpu_hdr_t **respp, char *msg)
{
    size_t size;
    dr_cpu_hdr_t *resp;

    ASSERT((msg != NULL) && (strlen(msg) > 0));

    size = sizeof (*req) + strlen(msg) + 1;
    resp = kmem_alloc(size, KM_SLEEP);
    DR_DBG_KMEM("%s: alloc addr %p size %ld\n",
        __func__, (void *)resp, size);

    resp->req_num = req->req_num;
    resp->msg_type = DR_CPU_ERROR;
    resp->num_records = 0;

    (void) strcpy((char *)(resp) + sizeof (*resp), msg);

    *respp = resp;

    return (size);
}

/*
 * Common routine to config or unconfig multiple cpus.  The unconfig
 * case checks with the OS to see if the removal of cpus will be
 * permitted, but can be overridden by the "force" version of the
 * command.  Otherwise, the logic for both cases is identical.
 *
 * Note: Do not modify result buffer or length on error.
 */
static int
dr_cpu_list_wrk(dr_cpu_hdr_t *req, dr_cpu_hdr_t **resp, int *resp_len)
{
    int     rv;
    int     idx;
    int     count;
    fn_t        dr_fn;
    int     se_hint;
    boolean_t   force = B_FALSE;
    uint32_t    *req_cpus;
    dr_cpu_res_t    *res;
    int     drctl_cmd;
    int     drctl_flags = 0;
    drctl_rsrc_t    *drctl_req;
    size_t      drctl_req_len;
    drctl_resp_t    *drctl_resp;
    drctl_rsrc_t    *drctl_rsrc;
    size_t      drctl_resp_len = 0;
    drctl_cookie_t  drctl_res_ck;

    ASSERT((req != NULL) && (req->num_records != 0));

    count = req->num_records;

    /*
     * Extract all information that is specific
     * to the various types of operations.
     */
    switch (req->msg_type) {
    case DR_CPU_CONFIGURE:
        dr_fn = dr_cpu_configure;
        drctl_cmd = DRCTL_CPU_CONFIG_REQUEST;
        se_hint = SE_HINT_INSERT;
        break;
    case DR_CPU_FORCE_UNCONFIG:
        drctl_flags = DRCTL_FLAG_FORCE;
        force = B_TRUE;
        _NOTE(FALLTHROUGH)
    case DR_CPU_UNCONFIGURE:
        dr_fn = dr_cpu_unconfigure;
        drctl_cmd = DRCTL_CPU_UNCONFIG_REQUEST;
        se_hint = SE_HINT_REMOVE;
        break;
    default:
        /* Programming error if we reach this. */
        cmn_err(CE_NOTE,
            "%s: bad msg_type %d\n", __func__, req->msg_type);
        ASSERT(0);
        return (-1);
    }

    /* the incoming array of cpuids to operate on */
    req_cpus = DR_CPU_CMD_CPUIDS(req);

    /* allocate drctl request msg based on incoming resource count */
    drctl_req_len = sizeof (drctl_rsrc_t) * count;
    drctl_req = kmem_zalloc(drctl_req_len, KM_SLEEP);
    DR_DBG_KMEM("%s: alloc addr %p size %ld\n",
        __func__, (void *)drctl_req, drctl_req_len);

    /* copy the cpuids for the drctl call from the incoming request msg */
    for (idx = 0; idx < count; idx++)
        drctl_req[idx].res_cpu_id = req_cpus[idx];

    rv = drctl_config_init(drctl_cmd, drctl_flags, drctl_req,
        count, &drctl_resp, &drctl_resp_len, &drctl_res_ck);

    ASSERT((drctl_resp != NULL) && (drctl_resp_len != 0));

    if (rv != 0) {
        DR_DBG_CPU("%s: drctl_config_init "
            "returned: %d\n", __func__, rv);

        if (DRCPU_VERS_EQ(1, 0)) {
            rv = -1;
        } else {
            ASSERT(DRCPU_VERS_GTEQ(1, 1));
            ASSERT(drctl_resp->resp_type == DRCTL_RESP_ERR);

            *resp_len = dr_cpu_err_resp(req,
                resp, drctl_resp->resp_err_msg);
        }

        DR_DBG_KMEM("%s: free addr %p size %ld\n",
            __func__, (void *)drctl_resp, drctl_resp_len);
        kmem_free(drctl_resp, drctl_resp_len);
        DR_DBG_KMEM("%s: free addr %p size %ld\n",
            __func__, (void *)drctl_req, drctl_req_len);
        kmem_free(drctl_req, drctl_req_len);

        return (rv);
    }

    ASSERT(drctl_resp->resp_type == DRCTL_RESP_OK);

    drctl_rsrc = drctl_resp->resp_resources;

    /* create the result scratch array */
    res = dr_cpu_res_array_init(req, drctl_rsrc, count);

    /*
     * For unconfigure, check if there are any conditions
     * that will cause the operation to fail. These are
     * performed before the actual unconfigure attempt so
     * that a meaningful error message can be generated.
     */
    if (req->msg_type != DR_CPU_CONFIGURE)
        dr_cpu_check_cpus(req, res);

    /* perform the specified operation on each of the CPUs */
    for (idx = 0; idx < count; idx++) {
        int result;
        int status;

        /*
         * If no action will be taken against the current
         * CPU, update the drctl resource information to
         * ensure that it gets recovered properly during
         * the drctl fini() call.
         */
        if (res[idx].result != DR_CPU_RES_OK) {
            drctl_req[idx].status = DRCTL_STATUS_CONFIG_FAILURE;
            continue;
        }

        /* call the function to perform the actual operation */
        result = (*dr_fn)(req_cpus[idx], &status, force);

        /* save off results of the operation */
        res[idx].result = result;
        res[idx].status = status;

        /* save result for drctl fini() reusing init() msg memory */
        drctl_req[idx].status = (result != DR_CPU_RES_OK) ?
            DRCTL_STATUS_CONFIG_FAILURE : DRCTL_STATUS_CONFIG_SUCCESS;

        DR_DBG_CPU("%s: cpuid %d status %d result %d off '%s'\n",
            __func__, req_cpus[idx], drctl_req[idx].status, result,
            (res[idx].string) ? res[idx].string : "");
    }

    if ((rv = drctl_config_fini(&drctl_res_ck, drctl_req, count)) != 0)
        DR_DBG_CPU("%s: drctl_config_fini "
            "returned: %d\n", __func__, rv);

    /*
     * Operation completed without any fatal errors.
     * Pack the response for transmission.
     */
    *resp_len = dr_cpu_pack_response(req, res, resp);

    /* notify interested parties about the operation */
    dr_generate_event(DR_TYPE_CPU, se_hint);

    /*
     * Deallocate any scratch memory.
     */
    DR_DBG_KMEM("%s: free addr %p size %ld\n",
        __func__, (void *)drctl_resp, drctl_resp_len);
    kmem_free(drctl_resp, drctl_resp_len);
    DR_DBG_KMEM("%s: free addr %p size %ld\n",
        __func__, (void *)drctl_req, drctl_req_len);
    kmem_free(drctl_req, drctl_req_len);

    dr_cpu_res_array_fini(res, count);

    return (0);
}

/*
 * Allocate and initialize a result array based on the initial
 * drctl operation. A valid result array is always returned.
 */
static dr_cpu_res_t *
dr_cpu_res_array_init(dr_cpu_hdr_t *req, drctl_rsrc_t *rsrc, int nrsrc)
{
    int     idx;
    dr_cpu_res_t    *res;
    char        *err_str;
    size_t      err_len;

    /* allocate zero filled buffer to initialize fields */
    res = kmem_zalloc(nrsrc * sizeof (dr_cpu_res_t), KM_SLEEP);
    DR_DBG_KMEM("%s: alloc addr %p size %ld\n",
        __func__, (void *)res, nrsrc * sizeof (dr_cpu_res_t));

    /*
     * Fill in the result information for each resource.
     */
    for (idx = 0; idx < nrsrc; idx++) {
        res[idx].cpuid = rsrc[idx].res_cpu_id;
        res[idx].result = DR_CPU_RES_OK;

        if (rsrc[idx].status == DRCTL_STATUS_ALLOW)
            continue;

        /*
         * Update the state information for this CPU.
         */
        res[idx].result = DR_CPU_RES_BLOCKED;
        res[idx].status = (req->msg_type == DR_CPU_CONFIGURE) ?
            DR_CPU_STAT_UNCONFIGURED : DR_CPU_STAT_CONFIGURED;

        /*
         * If an error string exists, copy it out of the
         * message buffer. This eliminates any dependency
         * on the memory allocated for the message buffer
         * itself.
         */
        if (rsrc[idx].offset != NULL) {
            err_str = (char *)rsrc + rsrc[idx].offset;
            err_len = strlen(err_str) + 1;

            res[idx].string = kmem_alloc(err_len, KM_SLEEP);
            DR_DBG_KMEM("%s: alloc addr %p size %ld\n",
                __func__, (void *)(res[idx].string), err_len);
            bcopy(err_str, res[idx].string, err_len);
        }
    }

    return (res);
}

static void
dr_cpu_res_array_fini(dr_cpu_res_t *res, int nres)
{
    int idx;
    size_t  str_len;

    for (idx = 0; idx < nres; idx++) {
        /* deallocate the error string if present */
        if (res[idx].string) {
            str_len = strlen(res[idx].string) + 1;
            DR_DBG_KMEM("%s: free addr %p size %ld\n",
                __func__, (void *)(res[idx].string), str_len);
            kmem_free(res[idx].string, str_len);
        }
    }

    /* deallocate the result array itself */
    DR_DBG_KMEM("%s: free addr %p size %ld\n",
        __func__, (void *)res, sizeof (dr_cpu_res_t) * nres);
    kmem_free(res, sizeof (dr_cpu_res_t) * nres);
}

/*
 * Allocate and pack a response message for transmission based
 * on the specified result array. A valid response message and
 * valid size information is always returned.
 */
static size_t
dr_cpu_pack_response(dr_cpu_hdr_t *req, dr_cpu_res_t *res, dr_cpu_hdr_t **respp)
{
    int     idx;
    dr_cpu_hdr_t    *resp;
    dr_cpu_stat_t   *resp_stat;
    size_t      resp_len;
    uint32_t    curr_off;
    caddr_t     curr_str;
    size_t      str_len;
    size_t      stat_len;
    int     nstat = req->num_records;

    /*
     * Calculate the size of the response message
     * and allocate an appropriately sized buffer.
     */
    resp_len = 0;

    /* add the header size */
    resp_len += sizeof (dr_cpu_hdr_t);

    /* add the stat array size */
    stat_len = sizeof (dr_cpu_stat_t) * nstat;
    resp_len += stat_len;

    /* add the size of any error strings */
    for (idx = 0; idx < nstat; idx++) {
        if (res[idx].string != NULL) {
            resp_len += strlen(res[idx].string) + 1;
        }
    }

    /* allocate the message buffer */
    resp = kmem_zalloc(resp_len, KM_SLEEP);
    DR_DBG_KMEM("%s: alloc addr %p size %ld\n",
        __func__, (void *)resp, resp_len);

    /*
     * Fill in the header information.
     */
    resp->req_num = req->req_num;
    resp->msg_type = DR_CPU_OK;
    resp->num_records = nstat;

    /*
     * Fill in the stat information.
     */
    resp_stat = DR_CPU_RESP_STATS(resp);

    /* string offsets start immediately after stat array */
    curr_off = sizeof (dr_cpu_hdr_t) + stat_len;
    curr_str = (char *)resp_stat + stat_len;

    for (idx = 0; idx < nstat; idx++) {
        resp_stat[idx].cpuid = res[idx].cpuid;
        resp_stat[idx].result = res[idx].result;
        resp_stat[idx].status = res[idx].status;

        if (res[idx].string != NULL) {
            /* copy over the error string */
            str_len = strlen(res[idx].string) + 1;
            bcopy(res[idx].string, curr_str, str_len);
            resp_stat[idx].string_off = curr_off;

            curr_off += str_len;
            curr_str += str_len;
        }
    }

    /* buffer should be exactly filled */
    ASSERT(curr_off == resp_len);

    *respp = resp;
    return (resp_len);
}

/*
 * Check for conditions that will prevent a CPU from being offlined.
 * This provides the opportunity to generate useful information to
 * help diagnose the failure rather than letting the offline attempt
 * fail in a more generic way.
 */
static void
dr_cpu_check_cpus(dr_cpu_hdr_t *req, dr_cpu_res_t *res)
{
    int     idx;
    cpu_t       *cp;
    uint32_t    *cpuids;

    ASSERT((req->msg_type == DR_CPU_UNCONFIGURE) ||
        (req->msg_type == DR_CPU_FORCE_UNCONFIG));

    DR_DBG_CPU("dr_cpu_check_cpus...\n");

    /* array of cpuids start just after the header */
    cpuids = DR_CPU_CMD_CPUIDS(req);

    mutex_enter(&cpu_lock);

    /*
     * Always check processor set membership first. The
     * last CPU in a processor set will fail to offline
     * even if the operation if forced, so any failures
     * should always be reported.
     */
    dr_cpu_check_psrset(cpuids, res, req->num_records);

    /* process each cpu that is part of the request */
    for (idx = 0; idx < req->num_records; idx++) {

        /* nothing to check if the CPU has already failed */
        if (res[idx].result != DR_CPU_RES_OK)
            continue;

        if ((cp = cpu_get(cpuids[idx])) == NULL)
            continue;

        /*
         * Only check if there are bound threads if the
         * operation is not a forced unconfigure. In a
         * forced request, threads are automatically
         * unbound before they are offlined.
         */
        if (req->msg_type == DR_CPU_UNCONFIGURE) {
            /*
             * The return value is only interesting if other
             * checks are added to this loop and a decision
             * is needed on whether to continue checking.
             */
            (void) dr_cpu_check_bound_thr(cp, &res[idx]);
        }
    }

    mutex_exit(&cpu_lock);
}

/*
 * Examine the processor set configuration for the specified
 * CPUs and see if the unconfigure operation would result in
 * trying to remove the last CPU in any processor set.
 */
static void
dr_cpu_check_psrset(uint32_t *cpuids, dr_cpu_res_t *res, int nres)
{
    int     cpu_idx;
    int     set_idx;
    cpu_t       *cp;
    cpupart_t   *cpp;
    char        err_str[DR_CPU_MAX_ERR_LEN];
    size_t      err_len;
    struct {
        cpupart_t   *cpp;
        int     ncpus;
    } *psrset;

    ASSERT(MUTEX_HELD(&cpu_lock));

    /*
     * Allocate a scratch array to count the CPUs in
     * the various processor sets. A CPU always belongs
     * to exactly one processor set, so by definition,
     * the scratch array never needs to be larger than
     * the number of CPUs.
     */
    psrset = kmem_zalloc(sizeof (*psrset) * nres, KM_SLEEP);
    DR_DBG_KMEM("%s: alloc addr %p size %ld\n",
        __func__, (void *)psrset, sizeof (*psrset) * nres);

    for (cpu_idx = 0; cpu_idx < nres; cpu_idx++) {

        /* skip any CPUs that have already failed */
        if (res[cpu_idx].result != DR_CPU_RES_OK)
            continue;

        if ((cp = cpu_get(cpuids[cpu_idx])) == NULL)
            continue;

        cpp = cp->cpu_part;

        /* lookup the set this CPU belongs to */
        for (set_idx = 0; set_idx < nres; set_idx++) {

            /* matching set found */
            if (cpp == psrset[set_idx].cpp)
                break;

            /* set not found, start a new entry */
            if (psrset[set_idx].cpp == NULL) {
                psrset[set_idx].cpp = cpp;
                psrset[set_idx].ncpus = cpp->cp_ncpus;
                break;
            }
        }

        ASSERT(set_idx != nres);

        /*
         * Remove the current CPU from the set total but only
         * generate an error for the last CPU. The correct CPU
         * will get the error because the unconfigure attempts
         * will occur in the same order in which the CPUs are
         * examined in this loop.
         */
        if (--psrset[set_idx].ncpus == 0) {
            /*
             * Fill in the various pieces of information
             * to report that the operation will fail.
             */
            res[cpu_idx].result = DR_CPU_RES_BLOCKED;
            res[cpu_idx].status = DR_CPU_STAT_CONFIGURED;

            (void) snprintf(err_str, DR_CPU_MAX_ERR_LEN,
                "last online cpu in processor set %d", cpp->cp_id);

            err_len = strlen(err_str) + 1;

            res[cpu_idx].string = kmem_alloc(err_len, KM_SLEEP);
            DR_DBG_KMEM("%s: alloc addr %p size %ld\n",
                __func__, (void *)(res[cpu_idx].string), err_len);
            bcopy(err_str, res[cpu_idx].string, err_len);

            DR_DBG_CPU("cpu %d: %s\n", cpuids[cpu_idx], err_str);
        }
    }

    DR_DBG_KMEM("%s: free addr %p size %ld\n",
        __func__, (void *)psrset, sizeof (*psrset) * nres);
    kmem_free(psrset, sizeof (*psrset) * nres);
}

/*
 * Check if any threads are bound to the specified CPU. If the
 * condition is true, DR_CPU_RES_BLOCKED is returned and an error
 * string is generated and placed in the specified result structure.
 * Otherwise, DR_CPU_RES_OK is returned.
 */
static int
dr_cpu_check_bound_thr(cpu_t *cp, dr_cpu_res_t *res)
{
    int     nbound;
    proc_t      *pp;
    kthread_t   *tp;
    char        err_str[DR_CPU_MAX_ERR_LEN];
    size_t      err_len;

    /*
     * Error string allocation makes an assumption
     * that no blocking condition has been identified.
     */
    ASSERT(res->result == DR_CPU_RES_OK);
    ASSERT(res->string == NULL);

    ASSERT(MUTEX_HELD(&cpu_lock));

    mutex_enter(&pidlock);

    nbound = 0;

    /*
     * Walk the active processes, checking if each
     * thread belonging to the process is bound.
     */
    for (pp = practive; (pp != NULL) && (nbound <= 1); pp = pp->p_next) {
        mutex_enter(&pp->p_lock);

        tp = pp->p_tlist;

        if ((tp == NULL) || (pp->p_flag & SSYS)) {
            mutex_exit(&pp->p_lock);
            continue;
        }

        do {
            if (tp->t_bind_cpu != cp->cpu_id)
                continue;

            /*
             * Update the running total of bound
             * threads. Continue the search until
             * it can be determined if more than
             * one thread is bound to the CPU.
             */
            if (++nbound > 1)
                break;

        } while ((tp = tp->t_forw) != pp->p_tlist);

        mutex_exit(&pp->p_lock);
    }

    mutex_exit(&pidlock);

    if (nbound) {
        /*
         * Threads are bound to the CPU. Fill in
         * various pieces of information to report
         * that the operation will fail.
         */
        res->result = DR_CPU_RES_BLOCKED;
        res->status = DR_CPU_STAT_CONFIGURED;

        (void) snprintf(err_str, DR_CPU_MAX_ERR_LEN, "cpu has bound "
            "thread%s", (nbound > 1) ? "s" : "");

        err_len = strlen(err_str) + 1;

        res->string = kmem_alloc(err_len, KM_SLEEP);
        DR_DBG_KMEM("%s: alloc addr %p size %ld\n",
            __func__, (void *)(res->string), err_len);
        bcopy(err_str, res->string, err_len);

        DR_DBG_CPU("cpu %d: %s\n", cp->cpu_id, err_str);
    }

    return (res->result);
}

/*
 * Do not modify result buffer or length on error.
 */
static int
dr_cpu_list_status(dr_cpu_hdr_t *req, dr_cpu_hdr_t **resp, int *resp_len)
{
    int     idx;
    int     result;
    int     status;
    int     rlen;
    uint32_t    *cpuids;
    dr_cpu_hdr_t    *rp;
    dr_cpu_stat_t   *stat;
    md_t        *mdp = NULL;
    int     num_nodes;
    int     listsz;
    mde_cookie_t    *listp = NULL;
    mde_cookie_t    cpunode;
    boolean_t   walk_md = B_FALSE;

    /* the incoming array of cpuids to configure */
    cpuids = DR_CPU_CMD_CPUIDS(req);

    /* allocate a response message */
    rlen = sizeof (dr_cpu_hdr_t);
    rlen += req->num_records * sizeof (dr_cpu_stat_t);
    rp = kmem_zalloc(rlen, KM_SLEEP);
    DR_DBG_KMEM("%s: alloc addr %p size %d\n", __func__, (void *)rp, rlen);

    /* fill in the known data */
    rp->req_num = req->req_num;
    rp->msg_type = DR_CPU_STATUS;
    rp->num_records = req->num_records;

    /* stat array for the response */
    stat = DR_CPU_RESP_STATS(rp);

    /* get the status for each of the CPUs */
    for (idx = 0; idx < req->num_records; idx++) {

        result = dr_cpu_status(cpuids[idx], &status);

        if (result == DR_CPU_RES_FAILURE)
            walk_md = B_TRUE;

        /* save off results of the status */
        stat[idx].cpuid = cpuids[idx];
        stat[idx].result = result;
        stat[idx].status = status;
    }

    if (walk_md == B_FALSE)
        goto done;

    /*
     * At least one of the cpus did not have a CPU
     * structure. So, consult the MD to determine if
     * they are present.
     */

    if ((mdp = md_get_handle()) == NULL) {
        DR_DBG_CPU("unable to initialize MD\n");
        goto done;
    }

    num_nodes = md_node_count(mdp);
    ASSERT(num_nodes > 0);

    listsz = num_nodes * sizeof (mde_cookie_t);
    listp = kmem_zalloc(listsz, KM_SLEEP);
    DR_DBG_KMEM("%s: alloc addr %p size %d\n",
        __func__, (void *)listp, listsz);

    for (idx = 0; idx < req->num_records; idx++) {

        if (stat[idx].result != DR_CPU_RES_FAILURE)
            continue;

        /* check the MD for the current cpuid */
        cpunode = dr_cpu_find_node_md(stat[idx].cpuid, mdp, listp);

        stat[idx].result = DR_CPU_RES_OK;

        if (cpunode == MDE_INVAL_ELEM_COOKIE) {
            stat[idx].status = DR_CPU_STAT_NOT_PRESENT;
        } else {
            stat[idx].status = DR_CPU_STAT_UNCONFIGURED;
        }
    }

    DR_DBG_KMEM("%s: free addr %p size %d\n",
        __func__, (void *)listp, listsz);
    kmem_free(listp, listsz);

    (void) md_fini_handle(mdp);

done:
    *resp = rp;
    *resp_len = rlen;

    return (0);
}

static int
dr_cpu_configure(processorid_t cpuid, int *status, boolean_t force)
{
     _NOTE(ARGUNUSED(force))
    struct cpu  *cp;
    int     rv = 0;

    DR_DBG_CPU("dr_cpu_configure...\n");

    /*
     * Build device tree node for the CPU
     */
    if ((rv = dr_cpu_probe(cpuid)) != 0) {
        DR_DBG_CPU("failed to probe CPU %d (%d)\n", cpuid, rv);
        if (rv == EINVAL) {
            *status = DR_CPU_STAT_NOT_PRESENT;
            return (DR_CPU_RES_NOT_IN_MD);
        }
        *status = DR_CPU_STAT_UNCONFIGURED;
        return (DR_CPU_RES_FAILURE);
    }

    mutex_enter(&cpu_lock);

    /*
     * Configure the CPU
     */
    if ((cp = cpu_get(cpuid)) == NULL) {

        if ((rv = cpu_configure(cpuid)) != 0) {
            DR_DBG_CPU("failed to configure CPU %d (%d)\n",
                cpuid, rv);
            rv = DR_CPU_RES_FAILURE;
            *status = DR_CPU_STAT_UNCONFIGURED;
            goto done;
        }

        DR_DBG_CPU("CPU %d configured\n", cpuid);

        /* CPU struct should exist now */
        cp = cpu_get(cpuid);
    }

    ASSERT(cp);

    /*
     * Power on the CPU. In sun4v, this brings the stopped
     * CPU into the guest from the Hypervisor.
     */
    if (cpu_is_poweredoff(cp)) {

        if ((rv = cpu_poweron(cp)) != 0) {
            DR_DBG_CPU("failed to power on CPU %d (%d)\n",
                cpuid, rv);
            rv = DR_CPU_RES_FAILURE;
            *status = DR_CPU_STAT_UNCONFIGURED;
            goto done;
        }

        DR_DBG_CPU("CPU %d powered on\n", cpuid);
    }

    /*
     * Online the CPU
     */
    if (cpu_is_offline(cp)) {

        if ((rv = cpu_online(cp)) != 0) {
            DR_DBG_CPU("failed to online CPU %d (%d)\n",
                cpuid, rv);
            rv = DR_CPU_RES_FAILURE;
            /* offline is still configured */
            *status = DR_CPU_STAT_CONFIGURED;
            goto done;
        }

        DR_DBG_CPU("CPU %d online\n", cpuid);
    }

    rv = DR_CPU_RES_OK;
    *status = DR_CPU_STAT_CONFIGURED;

done:
    mutex_exit(&cpu_lock);

    return (rv);
}

static int
dr_cpu_unconfigure(processorid_t cpuid, int *status, boolean_t force)
{
    struct cpu  *cp;
    int     rv = 0;
    int     cpu_flags;

    DR_DBG_CPU("dr_cpu_unconfigure%s...\n", (force) ? " (force)" : "");

    mutex_enter(&cpu_lock);

    cp = cpu_get(cpuid);

    if (cp == NULL) {

        /*
         * The OS CPU structures are already torn down,
         * Attempt to deprobe the CPU to make sure the
         * device tree is up to date.
         */
        if (dr_cpu_deprobe(cpuid) != 0) {
            DR_DBG_CPU("failed to deprobe CPU %d\n", cpuid);
            rv = DR_CPU_RES_FAILURE;
            *status = DR_CPU_STAT_UNCONFIGURED;
            goto done;
        }

        goto done;
    }

    ASSERT(cp->cpu_id == cpuid);

    /*
     * Offline the CPU
     */
    if (cpu_is_active(cp)) {

        /* set the force flag correctly */
        cpu_flags = (force) ? CPU_FORCED : 0;

        if ((rv = cpu_offline(cp, cpu_flags)) != 0) {
            DR_DBG_CPU("failed to offline CPU %d (%d)\n",
                cpuid, rv);

            rv = DR_CPU_RES_FAILURE;
            *status = DR_CPU_STAT_CONFIGURED;
            goto done;
        }

        DR_DBG_CPU("CPU %d offline\n", cpuid);
    }

    /*
     * Power off the CPU. In sun4v, this puts the running
     * CPU into the stopped state in the Hypervisor.
     */
    if (!cpu_is_poweredoff(cp)) {

        if ((rv = cpu_poweroff(cp)) != 0) {
            DR_DBG_CPU("failed to power off CPU %d (%d)\n",
                cpuid, rv);
            rv = DR_CPU_RES_FAILURE;
            *status = DR_CPU_STAT_CONFIGURED;
            goto done;
        }

        DR_DBG_CPU("CPU %d powered off\n", cpuid);
    }

    /*
     * Unconfigure the CPU
     */
    if ((rv = cpu_unconfigure(cpuid)) != 0) {
        DR_DBG_CPU("failed to unconfigure CPU %d (%d)\n", cpuid, rv);
        rv = DR_CPU_RES_FAILURE;
        *status = DR_CPU_STAT_UNCONFIGURED;
        goto done;
    }

    DR_DBG_CPU("CPU %d unconfigured\n", cpuid);

    /*
     * Tear down device tree.
     */
    if ((rv = dr_cpu_deprobe(cpuid)) != 0) {
        DR_DBG_CPU("failed to deprobe CPU %d (%d)\n", cpuid, rv);
        rv = DR_CPU_RES_FAILURE;
        *status = DR_CPU_STAT_UNCONFIGURED;
        goto done;
    }

    rv = DR_CPU_RES_OK;
    *status = DR_CPU_STAT_UNCONFIGURED;

done:
    mutex_exit(&cpu_lock);

    return (rv);
}

/*
 * Determine the state of a CPU. If the CPU structure is not present,
 * it does not attempt to determine whether or not the CPU is in the
 * MD. It is more efficient to do this at the higher level for all
 * CPUs since it may not even be necessary to search the MD if all
 * the CPUs are accounted for. Returns DR_CPU_RES_OK if the CPU
 * structure is present, and DR_CPU_RES_FAILURE otherwise as a signal
 * that an MD walk is necessary.
 */
static int
dr_cpu_status(processorid_t cpuid, int *status)
{
    int     rv;
    struct cpu  *cp;

    DR_DBG_CPU("dr_cpu_status...\n");

    mutex_enter(&cpu_lock);

    if ((cp = cpu_get(cpuid)) == NULL) {
        /* need to check if cpu is in the MD */
        rv = DR_CPU_RES_FAILURE;
        goto done;
    }

    if (cpu_is_poweredoff(cp)) {
        /*
         * The CPU is powered off, so it is considered
         * unconfigured from the service entity point of
         * view. The CPU is not available to the system
         * and intervention by the service entity would
         * be required to change that.
         */
        *status = DR_CPU_STAT_UNCONFIGURED;
    } else {
        /*
         * The CPU is powered on, so it is considered
         * configured from the service entity point of
         * view. It is available for use by the system
         * and service entities are not concerned about
         * the operational status (offline, online, etc.)
         * of the CPU in terms of DR.
         */
        *status = DR_CPU_STAT_CONFIGURED;
    }

    rv = DR_CPU_RES_OK;

done:
    mutex_exit(&cpu_lock);

    return (rv);
}

typedef struct {
    md_t        *mdp;
    mde_cookie_t    cpunode;
    dev_info_t  *dip;
} cb_arg_t;

#define STR_ARR_LEN 5

static int
new_cpu_node(dev_info_t *new_node, void *arg, uint_t flags)
{
    _NOTE(ARGUNUSED(flags))

    char        *compat;
    uint64_t    freq;
    uint64_t    cpuid = 0;
    int     regbuf[4];
    int     len = 0;
    cb_arg_t    *cba;
    char        *str_arr[STR_ARR_LEN];
    char        *curr;
    int     idx = 0;

    DR_DBG_CPU("new_cpu_node...\n");

    cba = (cb_arg_t *)arg;

    /*
     * Add 'name' property
     */
    if (ndi_prop_update_string(DDI_DEV_T_NONE, new_node,
        "name", "cpu") != DDI_SUCCESS) {
        DR_DBG_CPU("new_cpu_node: failed to create 'name' property\n");
        return (DDI_WALK_ERROR);
    }

    /*
     * Add 'compatible' property
     */
    if (md_get_prop_data(cba->mdp, cba->cpunode, "compatible",
        (uint8_t **)(&compat), &len)) {
        DR_DBG_CPU("new_cpu_node: failed to read 'compatible' property "
            "from MD\n");
        return (DDI_WALK_ERROR);
    }

    DR_DBG_CPU("'compatible' len is %d\n", len);

    /* parse the MD string array */
    curr = compat;
    while (curr < (compat + len)) {

        DR_DBG_CPU("adding '%s' to 'compatible' property\n", curr);

        str_arr[idx++] = curr;
        curr += strlen(curr) + 1;

        if (idx == STR_ARR_LEN) {
            DR_DBG_CPU("exceeded str_arr len (%d)\n", STR_ARR_LEN);
            break;
        }
    }

    if (ndi_prop_update_string_array(DDI_DEV_T_NONE, new_node,
        "compatible", str_arr, idx) != DDI_SUCCESS) {
        DR_DBG_CPU("new_cpu_node: failed to create 'compatible' "
            "property\n");
        return (DDI_WALK_ERROR);
    }

    /*
     * Add 'device_type' property
     */
    if (ndi_prop_update_string(DDI_DEV_T_NONE, new_node,
        "device_type", "cpu") != DDI_SUCCESS) {
        DR_DBG_CPU("new_cpu_node: failed to create 'device_type' "
            "property\n");
        return (DDI_WALK_ERROR);
    }

    /*
     * Add 'clock-frequency' property
     */
    if (md_get_prop_val(cba->mdp, cba->cpunode, "clock-frequency", &freq)) {
        DR_DBG_CPU("new_cpu_node: failed to read 'clock-frequency' "
            "property from MD\n");
        return (DDI_WALK_ERROR);
    }

    if (ndi_prop_update_int(DDI_DEV_T_NONE, new_node,
        "clock-frequency", freq) != DDI_SUCCESS) {
        DR_DBG_CPU("new_cpu_node: failed to create 'clock-frequency' "
            "property\n");
        return (DDI_WALK_ERROR);
    }

    /*
     * Add 'reg' (cpuid) property
     */
    if (md_get_prop_val(cba->mdp, cba->cpunode, "id", &cpuid)) {
        DR_DBG_CPU("new_cpu_node: failed to read 'id' property "
            "from MD\n");
        return (DDI_WALK_ERROR);
    }

    DR_DBG_CPU("new cpuid=0x%lx\n", cpuid);

    bzero(regbuf, 4 * sizeof (int));
    regbuf[0] = 0xc0000000 | cpuid;

    if (ndi_prop_update_int_array(DDI_DEV_T_NONE, new_node,
        "reg", regbuf, 4) != DDI_SUCCESS) {
        DR_DBG_CPU("new_cpu_node: failed to create 'reg' property\n");
        return (DDI_WALK_ERROR);
    }

    cba->dip = new_node;

    return (DDI_WALK_TERMINATE);
}

static int
dr_cpu_probe(processorid_t cpuid)
{
    dev_info_t  *pdip;
    dev_info_t  *dip;
    devi_branch_t   br;
    md_t        *mdp = NULL;
    int     num_nodes;
    int     rv = 0;
    int     listsz;
    mde_cookie_t    *listp = NULL;
    cb_arg_t    cba;
    mde_cookie_t    cpunode;

    if ((dip = dr_cpu_find_node(cpuid)) != NULL) {
        /* nothing to do */
        e_ddi_branch_rele(dip);
        return (0);
    }

    if ((mdp = md_get_handle()) == NULL) {
        DR_DBG_CPU("unable to initialize machine description\n");
        return (-1);
    }

    num_nodes = md_node_count(mdp);
    ASSERT(num_nodes > 0);

    listsz = num_nodes * sizeof (mde_cookie_t);
    listp = kmem_zalloc(listsz, KM_SLEEP);
    DR_DBG_KMEM("%s: alloc addr %p size %d\n",
        __func__, (void *)listp, listsz);

    cpunode = dr_cpu_find_node_md(cpuid, mdp, listp);

    if (cpunode == MDE_INVAL_ELEM_COOKIE) {
        rv = EINVAL;
        goto done;
    }

    /* pass in MD cookie for CPU */
    cba.mdp = mdp;
    cba.cpunode = cpunode;

    br.arg = (void *)&cba;
    br.type = DEVI_BRANCH_SID;
    br.create.sid_branch_create = new_cpu_node;
    br.devi_branch_callback = NULL;
    pdip = ddi_root_node();

    if ((rv = e_ddi_branch_create(pdip, &br, NULL, 0))) {
        DR_DBG_CPU("e_ddi_branch_create failed: %d\n", rv);
        rv = -1;
        goto done;
    }

    DR_DBG_CPU("CPU %d probed\n", cpuid);

    rv = 0;

done:
    if (listp) {
        DR_DBG_KMEM("%s: free addr %p size %d\n",
            __func__, (void *)listp, listsz);
        kmem_free(listp, listsz);
    }

    if (mdp)
        (void) md_fini_handle(mdp);

    return (rv);
}

static int
dr_cpu_deprobe(processorid_t cpuid)
{
    dev_info_t  *fdip = NULL;
    dev_info_t  *dip;

    if ((dip = dr_cpu_find_node(cpuid)) == NULL) {
        DR_DBG_CPU("cpuid %d already deprobed\n", cpuid);
        return (0);
    }

    ASSERT(e_ddi_branch_held(dip));

    if (e_ddi_branch_destroy(dip, &fdip, 0)) {
        char *path = kmem_alloc(MAXPATHLEN, KM_SLEEP);

        DR_DBG_KMEM("%s: alloc addr %p size %d\n",
            __func__, (void *)path, MAXPATHLEN);
        /*
         * If non-NULL, fdip is held and must be released.
         */
        if (fdip != NULL) {
            (void) ddi_pathname(fdip, path);
            ddi_release_devi(fdip);
        } else {
            (void) ddi_pathname(dip, path);
        }
        cmn_err(CE_NOTE, "node removal failed: %s (%p)",
            path, (fdip) ? (void *)fdip : (void *)dip);

        DR_DBG_KMEM("%s: free addr %p size %d\n",
            __func__, (void *)path, MAXPATHLEN);
        kmem_free(path, MAXPATHLEN);

        return (-1);
    }

    DR_DBG_CPU("CPU %d deprobed\n", cpuid);

    return (0);
}

typedef struct {
    processorid_t   cpuid;
    dev_info_t  *dip;
} dr_search_arg_t;

static int
dr_cpu_check_node(dev_info_t *dip, void *arg)
{
    char        *name;
    processorid_t   cpuid;
    dr_search_arg_t *sarg = (dr_search_arg_t *)arg;

    if (dip == ddi_root_node()) {
        return (DDI_WALK_CONTINUE);
    }

    name = ddi_node_name(dip);

    if (strcmp(name, "cpu") != 0) {
        return (DDI_WALK_PRUNECHILD);
    }

    cpuid = ddi_prop_get_int(DDI_DEV_T_ANY, dip, DDI_PROP_DONTPASS,
        "reg", -1);

    cpuid = PROM_CFGHDL_TO_CPUID(cpuid);

    DR_DBG_CPU("found cpuid=0x%x, looking for 0x%x\n", cpuid, sarg->cpuid);

    if (cpuid == sarg->cpuid) {
        DR_DBG_CPU("matching node\n");

        /* matching node must be returned held */
        if (!e_ddi_branch_held(dip))
            e_ddi_branch_hold(dip);

        sarg->dip = dip;
        return (DDI_WALK_TERMINATE);
    }

    return (DDI_WALK_CONTINUE);
}

/*
 * Walk the device tree to find the dip corresponding to the cpuid
 * passed in. If present, the dip is returned held. The caller must
 * release the hold on the dip once it is no longer required. If no
 * matching node if found, NULL is returned.
 */
static dev_info_t *
dr_cpu_find_node(processorid_t cpuid)
{
    dr_search_arg_t arg;

    DR_DBG_CPU("dr_cpu_find_node...\n");

    arg.cpuid = cpuid;
    arg.dip = NULL;

    ddi_walk_devs(ddi_root_node(), dr_cpu_check_node, &arg);

    ASSERT((arg.dip == NULL) || (e_ddi_branch_held(arg.dip)));

    return ((arg.dip) ? arg.dip : NULL);
}

/*
 * Look up a particular cpuid in the MD. Returns the mde_cookie_t
 * representing that CPU if present, and MDE_INVAL_ELEM_COOKIE
 * otherwise. It is assumed the scratch array has already been
 * allocated so that it can accommodate the worst case scenario,
 * every node in the MD.
 */
static mde_cookie_t
dr_cpu_find_node_md(processorid_t cpuid, md_t *mdp, mde_cookie_t *listp)
{
    int     idx;
    int     nnodes;
    mde_cookie_t    rootnode;
    uint64_t    cpuid_prop;
    mde_cookie_t    result = MDE_INVAL_ELEM_COOKIE;

    rootnode = md_root_node(mdp);
    ASSERT(rootnode != MDE_INVAL_ELEM_COOKIE);

    /*
     * Scan the DAG for all the CPU nodes
     */
    nnodes = md_scan_dag(mdp, rootnode, md_find_name(mdp, "cpu"),
        md_find_name(mdp, "fwd"), listp);

    if (nnodes < 0) {
        DR_DBG_CPU("Scan for CPUs failed\n");
        return (result);
    }

    DR_DBG_CPU("dr_cpu_find_node_md: found %d CPUs in the MD\n", nnodes);

    /*
     * Find the CPU of interest
     */
    for (idx = 0; idx < nnodes; idx++) {

        if (md_get_prop_val(mdp, listp[idx], "id", &cpuid_prop)) {
            DR_DBG_CPU("Missing 'id' property for CPU node %d\n",
                idx);
            break;
        }

        if (cpuid_prop == cpuid) {
            /* found a match */
            DR_DBG_CPU("dr_cpu_find_node_md: found CPU %d "
                "in MD\n", cpuid);
            result = listp[idx];
            break;
        }
    }

    if (result == MDE_INVAL_ELEM_COOKIE) {
        DR_DBG_CPU("CPU %d not in MD\n", cpuid);
    }

    return (result);
}