// SPDX-License-Identifier: GPL-2.0-or-later
/*
 * dlmmod.c
 *
 * standalone DLM module
 *
 * Copyright (C) 2004 Oracle.  All rights reserved.
 */


#include <linux/module.h>
#include <linux/fs.h>
#include <linux/types.h>
#include <linux/slab.h>
#include <linux/highmem.h>
#include <linux/init.h>
#include <linux/sysctl.h>
#include <linux/random.h>
#include <linux/blkdev.h>
#include <linux/socket.h>
#include <linux/inet.h>
#include <linux/spinlock.h>
#include <linux/delay.h>


#include "../cluster/heartbeat.h"
#include "../cluster/nodemanager.h"
#include "../cluster/tcp.h"

#include "dlmapi.h"
#include "dlmcommon.h"
#include "dlmdomain.h"
#include "dlmdebug.h"

#define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_MASTER)
#include "../cluster/masklog.h"

static void dlm_mle_node_down(struct dlm_ctxt *dlm,
                              struct dlm_master_list_entry *mle,
                              struct o2nm_node *node,
                              int idx);
static void dlm_mle_node_up(struct dlm_ctxt *dlm,
                            struct dlm_master_list_entry *mle,
                            struct o2nm_node *node,
                            int idx);

static void dlm_assert_master_worker(struct dlm_work_item *item, void *data);
static int dlm_do_assert_master(struct dlm_ctxt *dlm,
                                struct dlm_lock_resource *res,
                                void *nodemap, u32 flags);
static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data);

static inline int dlm_mle_equal(struct dlm_ctxt *dlm,
                                struct dlm_master_list_entry *mle,
                                const char *name,
                                unsigned int namelen)
{
        if (dlm != mle->dlm)
                return 0;

        if (namelen != mle->mnamelen ||
            memcmp(name, mle->mname, namelen) != 0)
                return 0;

        return 1;
}

static struct kmem_cache *dlm_lockres_cache;
static struct kmem_cache *dlm_lockname_cache;
static struct kmem_cache *dlm_mle_cache;

static void dlm_mle_release(struct kref *kref);
static void dlm_init_mle(struct dlm_master_list_entry *mle,
                        enum dlm_mle_type type,
                        struct dlm_ctxt *dlm,
                        struct dlm_lock_resource *res,
                        const char *name,
                        unsigned int namelen);
static void dlm_put_mle(struct dlm_master_list_entry *mle);
static void __dlm_put_mle(struct dlm_master_list_entry *mle);
static int dlm_find_mle(struct dlm_ctxt *dlm,
                        struct dlm_master_list_entry **mle,
                        char *name, unsigned int namelen);

static int dlm_do_master_request(struct dlm_lock_resource *res,
                                 struct dlm_master_list_entry *mle, int to);


static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
                                     struct dlm_lock_resource *res,
                                     struct dlm_master_list_entry *mle,
                                     int *blocked);
static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
                                    struct dlm_lock_resource *res,
                                    struct dlm_master_list_entry *mle,
                                    int blocked);
static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
                                 struct dlm_lock_resource *res,
                                 struct dlm_master_list_entry *mle,
                                 struct dlm_master_list_entry **oldmle,
                                 const char *name, unsigned int namelen,
                                 u8 new_master, u8 master);

static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
                                    struct dlm_lock_resource *res);
static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
                                      struct dlm_lock_resource *res);
static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
                                       struct dlm_lock_resource *res,
                                       u8 target);
static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
                                       struct dlm_lock_resource *res);


int dlm_is_host_down(int errno)
{
        switch (errno) {
                case -EBADF:
                case -ECONNREFUSED:
                case -ENOTCONN:
                case -ECONNRESET:
                case -EPIPE:
                case -EHOSTDOWN:
                case -EHOSTUNREACH:
                case -ETIMEDOUT:
                case -ECONNABORTED:
                case -ENETDOWN:
                case -ENETUNREACH:
                case -ENETRESET:
                case -ESHUTDOWN:
                case -ENOPROTOOPT:
                case -EINVAL:   /* if returned from our tcp code,
                                   this means there is no socket */
                        return 1;
        }
        return 0;
}


/*
 * MASTER LIST FUNCTIONS
 */


/*
 * regarding master list entries and heartbeat callbacks:
 *
 * in order to avoid sleeping and allocation that occurs in
 * heartbeat, master list entries are simply attached to the
 * dlm's established heartbeat callbacks.  the mle is attached
 * when it is created, and since the dlm->spinlock is held at
 * that time, any heartbeat event will be properly discovered
 * by the mle.  the mle needs to be detached from the
 * dlm->mle_hb_events list as soon as heartbeat events are no
 * longer useful to the mle, and before the mle is freed.
 *
 * as a general rule, heartbeat events are no longer needed by
 * the mle once an "answer" regarding the lock master has been
 * received.
 */
static inline void __dlm_mle_attach_hb_events(struct dlm_ctxt *dlm,
                                              struct dlm_master_list_entry *mle)
{
        assert_spin_locked(&dlm->spinlock);

        list_add_tail(&mle->hb_events, &dlm->mle_hb_events);
}


static inline void __dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
                                              struct dlm_master_list_entry *mle)
{
        if (!list_empty(&mle->hb_events))
                list_del_init(&mle->hb_events);
}


static inline void dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
                                            struct dlm_master_list_entry *mle)
{
        spin_lock(&dlm->spinlock);
        __dlm_mle_detach_hb_events(dlm, mle);
        spin_unlock(&dlm->spinlock);
}

static void dlm_get_mle_inuse(struct dlm_master_list_entry *mle)
{
        struct dlm_ctxt *dlm;
        dlm = mle->dlm;

        assert_spin_locked(&dlm->spinlock);
        assert_spin_locked(&dlm->master_lock);
        mle->inuse++;
        kref_get(&mle->mle_refs);
}

static void dlm_put_mle_inuse(struct dlm_master_list_entry *mle)
{
        struct dlm_ctxt *dlm;
        dlm = mle->dlm;

        spin_lock(&dlm->spinlock);
        spin_lock(&dlm->master_lock);
        mle->inuse--;
        __dlm_put_mle(mle);
        spin_unlock(&dlm->master_lock);
        spin_unlock(&dlm->spinlock);

}

/* remove from list and free */
static void __dlm_put_mle(struct dlm_master_list_entry *mle)
{
        struct dlm_ctxt *dlm;
        dlm = mle->dlm;

        assert_spin_locked(&dlm->spinlock);
        assert_spin_locked(&dlm->master_lock);
        if (!kref_read(&mle->mle_refs)) {
                /* this may or may not crash, but who cares.
                 * it's a BUG. */
                mlog(ML_ERROR, "bad mle: %p\n", mle);
                dlm_print_one_mle(mle);
                BUG();
        } else
                kref_put(&mle->mle_refs, dlm_mle_release);
}


/* must not have any spinlocks coming in */
static void dlm_put_mle(struct dlm_master_list_entry *mle)
{
        struct dlm_ctxt *dlm;
        dlm = mle->dlm;

        spin_lock(&dlm->spinlock);
        spin_lock(&dlm->master_lock);
        __dlm_put_mle(mle);
        spin_unlock(&dlm->master_lock);
        spin_unlock(&dlm->spinlock);
}

static inline void dlm_get_mle(struct dlm_master_list_entry *mle)
{
        kref_get(&mle->mle_refs);
}

static void dlm_init_mle(struct dlm_master_list_entry *mle,
                        enum dlm_mle_type type,
                        struct dlm_ctxt *dlm,
                        struct dlm_lock_resource *res,
                        const char *name,
                        unsigned int namelen)
{
        assert_spin_locked(&dlm->spinlock);

        mle->dlm = dlm;
        mle->type = type;
        INIT_HLIST_NODE(&mle->master_hash_node);
        INIT_LIST_HEAD(&mle->hb_events);
        memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
        spin_lock_init(&mle->spinlock);
        init_waitqueue_head(&mle->wq);
        atomic_set(&mle->woken, 0);
        kref_init(&mle->mle_refs);
        memset(mle->response_map, 0, sizeof(mle->response_map));
        mle->master = O2NM_MAX_NODES;
        mle->new_master = O2NM_MAX_NODES;
        mle->inuse = 0;

        BUG_ON(mle->type != DLM_MLE_BLOCK &&
               mle->type != DLM_MLE_MASTER &&
               mle->type != DLM_MLE_MIGRATION);

        if (mle->type == DLM_MLE_MASTER) {
                BUG_ON(!res);
                mle->mleres = res;
                memcpy(mle->mname, res->lockname.name, res->lockname.len);
                mle->mnamelen = res->lockname.len;
                mle->mnamehash = res->lockname.hash;
        } else {
                BUG_ON(!name);
                mle->mleres = NULL;
                memcpy(mle->mname, name, namelen);
                mle->mnamelen = namelen;
                mle->mnamehash = dlm_lockid_hash(name, namelen);
        }

        atomic_inc(&dlm->mle_tot_count[mle->type]);
        atomic_inc(&dlm->mle_cur_count[mle->type]);

        /* copy off the node_map and register hb callbacks on our copy */
        memcpy(mle->node_map, dlm->domain_map, sizeof(mle->node_map));
        memcpy(mle->vote_map, dlm->domain_map, sizeof(mle->vote_map));
        clear_bit(dlm->node_num, mle->vote_map);
        clear_bit(dlm->node_num, mle->node_map);

        /* attach the mle to the domain node up/down events */
        __dlm_mle_attach_hb_events(dlm, mle);
}

void __dlm_unlink_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle)
{
        assert_spin_locked(&dlm->spinlock);
        assert_spin_locked(&dlm->master_lock);

        if (!hlist_unhashed(&mle->master_hash_node))
                hlist_del_init(&mle->master_hash_node);
}

void __dlm_insert_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle)
{
        struct hlist_head *bucket;

        assert_spin_locked(&dlm->master_lock);

        bucket = dlm_master_hash(dlm, mle->mnamehash);
        hlist_add_head(&mle->master_hash_node, bucket);
}

/* returns 1 if found, 0 if not */
static int dlm_find_mle(struct dlm_ctxt *dlm,
                        struct dlm_master_list_entry **mle,
                        char *name, unsigned int namelen)
{
        struct dlm_master_list_entry *tmpmle;
        struct hlist_head *bucket;
        unsigned int hash;

        assert_spin_locked(&dlm->master_lock);

        hash = dlm_lockid_hash(name, namelen);
        bucket = dlm_master_hash(dlm, hash);
        hlist_for_each_entry(tmpmle, bucket, master_hash_node) {
                if (!dlm_mle_equal(dlm, tmpmle, name, namelen))
                        continue;
                dlm_get_mle(tmpmle);
                *mle = tmpmle;
                return 1;
        }
        return 0;
}

void dlm_hb_event_notify_attached(struct dlm_ctxt *dlm, int idx, int node_up)
{
        struct dlm_master_list_entry *mle;

        assert_spin_locked(&dlm->spinlock);

        list_for_each_entry(mle, &dlm->mle_hb_events, hb_events) {
                if (node_up)
                        dlm_mle_node_up(dlm, mle, NULL, idx);
                else
                        dlm_mle_node_down(dlm, mle, NULL, idx);
        }
}

static void dlm_mle_node_down(struct dlm_ctxt *dlm,
                              struct dlm_master_list_entry *mle,
                              struct o2nm_node *node, int idx)
{
        spin_lock(&mle->spinlock);

        if (!test_bit(idx, mle->node_map))
                mlog(0, "node %u already removed from nodemap!\n", idx);
        else
                clear_bit(idx, mle->node_map);

        spin_unlock(&mle->spinlock);
}

static void dlm_mle_node_up(struct dlm_ctxt *dlm,
                            struct dlm_master_list_entry *mle,
                            struct o2nm_node *node, int idx)
{
        spin_lock(&mle->spinlock);

        if (test_bit(idx, mle->node_map))
                mlog(0, "node %u already in node map!\n", idx);
        else
                set_bit(idx, mle->node_map);

        spin_unlock(&mle->spinlock);
}


int dlm_init_mle_cache(void)
{
        dlm_mle_cache = kmem_cache_create("o2dlm_mle",
                                          sizeof(struct dlm_master_list_entry),
                                          0, SLAB_HWCACHE_ALIGN,
                                          NULL);
        if (dlm_mle_cache == NULL)
                return -ENOMEM;
        return 0;
}

void dlm_destroy_mle_cache(void)
{
        kmem_cache_destroy(dlm_mle_cache);
}

static void dlm_mle_release(struct kref *kref)
{
        struct dlm_master_list_entry *mle;
        struct dlm_ctxt *dlm;

        mle = container_of(kref, struct dlm_master_list_entry, mle_refs);
        dlm = mle->dlm;

        assert_spin_locked(&dlm->spinlock);
        assert_spin_locked(&dlm->master_lock);

        mlog(0, "Releasing mle for %.*s, type %d\n", mle->mnamelen, mle->mname,
             mle->type);

        /* remove from list if not already */
        __dlm_unlink_mle(dlm, mle);

        /* detach the mle from the domain node up/down events */
        __dlm_mle_detach_hb_events(dlm, mle);

        atomic_dec(&dlm->mle_cur_count[mle->type]);

        /* NOTE: kfree under spinlock here.
         * if this is bad, we can move this to a freelist. */
        kmem_cache_free(dlm_mle_cache, mle);
}


/*
 * LOCK RESOURCE FUNCTIONS
 */

int dlm_init_master_caches(void)
{
        dlm_lockres_cache = kmem_cache_create("o2dlm_lockres",
                                              sizeof(struct dlm_lock_resource),
                                              0, SLAB_HWCACHE_ALIGN, NULL);
        if (!dlm_lockres_cache)
                goto bail;

        dlm_lockname_cache = kmem_cache_create("o2dlm_lockname",
                                               DLM_LOCKID_NAME_MAX, 0,
                                               SLAB_HWCACHE_ALIGN, NULL);
        if (!dlm_lockname_cache)
                goto bail;

        return 0;
bail:
        dlm_destroy_master_caches();
        return -ENOMEM;
}

void dlm_destroy_master_caches(void)
{
        kmem_cache_destroy(dlm_lockname_cache);
        dlm_lockname_cache = NULL;

        kmem_cache_destroy(dlm_lockres_cache);
        dlm_lockres_cache = NULL;
}

static void dlm_lockres_release(struct kref *kref)
{
        struct dlm_lock_resource *res;
        struct dlm_ctxt *dlm;

        res = container_of(kref, struct dlm_lock_resource, refs);
        dlm = res->dlm;

        /* This should not happen -- all lockres' have a name
         * associated with them at init time. */
        BUG_ON(!res->lockname.name);

        mlog(0, "destroying lockres %.*s\n", res->lockname.len,
             res->lockname.name);

        atomic_dec(&dlm->res_cur_count);

        if (!hlist_unhashed(&res->hash_node) ||
            !list_empty(&res->granted) ||
            !list_empty(&res->converting) ||
            !list_empty(&res->blocked) ||
            !list_empty(&res->dirty) ||
            !list_empty(&res->recovering) ||
            !list_empty(&res->purge)) {
                mlog(ML_ERROR,
                     "Going to BUG for resource %.*s."
                     "  We're on a list! [%c%c%c%c%c%c%c]\n",
                     res->lockname.len, res->lockname.name,
                     !hlist_unhashed(&res->hash_node) ? 'H' : ' ',
                     !list_empty(&res->granted) ? 'G' : ' ',
                     !list_empty(&res->converting) ? 'C' : ' ',
                     !list_empty(&res->blocked) ? 'B' : ' ',
                     !list_empty(&res->dirty) ? 'D' : ' ',
                     !list_empty(&res->recovering) ? 'R' : ' ',
                     !list_empty(&res->purge) ? 'P' : ' ');

                dlm_print_one_lock_resource(res);
        }

        /* By the time we're ready to blow this guy away, we shouldn't
         * be on any lists. */
        BUG_ON(!hlist_unhashed(&res->hash_node));
        BUG_ON(!list_empty(&res->granted));
        BUG_ON(!list_empty(&res->converting));
        BUG_ON(!list_empty(&res->blocked));
        BUG_ON(!list_empty(&res->dirty));
        BUG_ON(!list_empty(&res->recovering));
        BUG_ON(!list_empty(&res->purge));

        kmem_cache_free(dlm_lockname_cache, (void *)res->lockname.name);

        kmem_cache_free(dlm_lockres_cache, res);
}

void dlm_lockres_put(struct dlm_lock_resource *res)
{
        kref_put(&res->refs, dlm_lockres_release);
}

static void dlm_init_lockres(struct dlm_ctxt *dlm,
                             struct dlm_lock_resource *res,
                             const char *name, unsigned int namelen)
{
        char *qname;

        /* If we memset here, we lose our reference to the kmalloc'd
         * res->lockname.name, so be sure to init every field
         * correctly! */

        qname = (char *) res->lockname.name;
        memcpy(qname, name, namelen);

        res->lockname.len = namelen;
        res->lockname.hash = dlm_lockid_hash(name, namelen);

        init_waitqueue_head(&res->wq);
        spin_lock_init(&res->spinlock);
        INIT_HLIST_NODE(&res->hash_node);
        INIT_LIST_HEAD(&res->granted);
        INIT_LIST_HEAD(&res->converting);
        INIT_LIST_HEAD(&res->blocked);
        INIT_LIST_HEAD(&res->dirty);
        INIT_LIST_HEAD(&res->recovering);
        INIT_LIST_HEAD(&res->purge);
        INIT_LIST_HEAD(&res->tracking);
        atomic_set(&res->asts_reserved, 0);
        res->migration_pending = 0;
        res->inflight_locks = 0;
        res->inflight_assert_workers = 0;

        res->dlm = dlm;

        kref_init(&res->refs);

        atomic_inc(&dlm->res_tot_count);
        atomic_inc(&dlm->res_cur_count);

        /* just for consistency */
        spin_lock(&res->spinlock);
        dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
        spin_unlock(&res->spinlock);

        res->state = DLM_LOCK_RES_IN_PROGRESS;

        res->last_used = 0;

        spin_lock(&dlm->track_lock);
        list_add_tail(&res->tracking, &dlm->tracking_list);
        spin_unlock(&dlm->track_lock);

        memset(res->lvb, 0, DLM_LVB_LEN);
        memset(res->refmap, 0, sizeof(res->refmap));
}

struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm,
                                   const char *name,
                                   unsigned int namelen)
{
        struct dlm_lock_resource *res = NULL;

        res = kmem_cache_zalloc(dlm_lockres_cache, GFP_NOFS);
        if (!res)
                goto error;

        res->lockname.name = kmem_cache_zalloc(dlm_lockname_cache, GFP_NOFS);
        if (!res->lockname.name)
                goto error;

        dlm_init_lockres(dlm, res, name, namelen);
        return res;

error:
        if (res)
                kmem_cache_free(dlm_lockres_cache, res);
        return NULL;
}

void dlm_lockres_set_refmap_bit(struct dlm_ctxt *dlm,
                                struct dlm_lock_resource *res, int bit)
{
        assert_spin_locked(&res->spinlock);

        mlog(0, "res %.*s, set node %u, %ps()\n", res->lockname.len,
             res->lockname.name, bit, __builtin_return_address(0));

        set_bit(bit, res->refmap);
}

void dlm_lockres_clear_refmap_bit(struct dlm_ctxt *dlm,
                                  struct dlm_lock_resource *res, int bit)
{
        assert_spin_locked(&res->spinlock);

        mlog(0, "res %.*s, clr node %u, %ps()\n", res->lockname.len,
             res->lockname.name, bit, __builtin_return_address(0));

        clear_bit(bit, res->refmap);
}

static void __dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
                                   struct dlm_lock_resource *res)
{
        res->inflight_locks++;

        mlog(0, "%s: res %.*s, inflight++: now %u, %ps()\n", dlm->name,
             res->lockname.len, res->lockname.name, res->inflight_locks,
             __builtin_return_address(0));
}

void dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
                                   struct dlm_lock_resource *res)
{
        assert_spin_locked(&res->spinlock);
        __dlm_lockres_grab_inflight_ref(dlm, res);
}

void dlm_lockres_drop_inflight_ref(struct dlm_ctxt *dlm,
                                   struct dlm_lock_resource *res)
{
        assert_spin_locked(&res->spinlock);

        BUG_ON(res->inflight_locks == 0);

        res->inflight_locks--;

        mlog(0, "%s: res %.*s, inflight--: now %u, %ps()\n", dlm->name,
             res->lockname.len, res->lockname.name, res->inflight_locks,
             __builtin_return_address(0));

        wake_up(&res->wq);
}

void __dlm_lockres_grab_inflight_worker(struct dlm_ctxt *dlm,
                struct dlm_lock_resource *res)
{
        assert_spin_locked(&res->spinlock);
        res->inflight_assert_workers++;
        mlog(0, "%s:%.*s: inflight assert worker++: now %u\n",
                        dlm->name, res->lockname.len, res->lockname.name,
                        res->inflight_assert_workers);
}

static void __dlm_lockres_drop_inflight_worker(struct dlm_ctxt *dlm,
                struct dlm_lock_resource *res)
{
        assert_spin_locked(&res->spinlock);
        BUG_ON(res->inflight_assert_workers == 0);
        res->inflight_assert_workers--;
        mlog(0, "%s:%.*s: inflight assert worker--: now %u\n",
                        dlm->name, res->lockname.len, res->lockname.name,
                        res->inflight_assert_workers);
}

static void dlm_lockres_drop_inflight_worker(struct dlm_ctxt *dlm,
                struct dlm_lock_resource *res)
{
        spin_lock(&res->spinlock);
        __dlm_lockres_drop_inflight_worker(dlm, res);
        spin_unlock(&res->spinlock);
}

/*
 * lookup a lock resource by name.
 * may already exist in the hashtable.
 * lockid is null terminated
 *
 * if not, allocate enough for the lockres and for
 * the temporary structure used in doing the mastering.
 *
 * also, do a lookup in the dlm->master_list to see
 * if another node has begun mastering the same lock.
 * if so, there should be a block entry in there
 * for this name, and we should *not* attempt to master
 * the lock here.   need to wait around for that node
 * to assert_master (or die).
 *
 */
struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm,
                                          const char *lockid,
                                          int namelen,
                                          int flags)
{
        struct dlm_lock_resource *tmpres=NULL, *res=NULL;
        struct dlm_master_list_entry *mle = NULL;
        struct dlm_master_list_entry *alloc_mle = NULL;
        int blocked = 0;
        int ret, nodenum;
        struct dlm_node_iter iter;
        unsigned int hash;
        int tries = 0;
        int bit, wait_on_recovery = 0;

        BUG_ON(!lockid);

        hash = dlm_lockid_hash(lockid, namelen);

        mlog(0, "get lockres %s (len %d)\n", lockid, namelen);

lookup:
        spin_lock(&dlm->spinlock);
        tmpres = __dlm_lookup_lockres_full(dlm, lockid, namelen, hash);
        if (tmpres) {
                spin_unlock(&dlm->spinlock);
                spin_lock(&tmpres->spinlock);

                /*
                 * Right after dlm spinlock was released, dlm_thread could have
                 * purged the lockres. Check if lockres got unhashed. If so
                 * start over.
                 */
                if (hlist_unhashed(&tmpres->hash_node)) {
                        spin_unlock(&tmpres->spinlock);
                        dlm_lockres_put(tmpres);
                        tmpres = NULL;
                        goto lookup;
                }

                /* Wait on the thread that is mastering the resource */
                if (tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
                        __dlm_wait_on_lockres(tmpres);
                        BUG_ON(tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN);
                        spin_unlock(&tmpres->spinlock);
                        dlm_lockres_put(tmpres);
                        tmpres = NULL;
                        goto lookup;
                }

                /* Wait on the resource purge to complete before continuing */
                if (tmpres->state & DLM_LOCK_RES_DROPPING_REF) {
                        BUG_ON(tmpres->owner == dlm->node_num);
                        __dlm_wait_on_lockres_flags(tmpres,
                                                    DLM_LOCK_RES_DROPPING_REF);
                        spin_unlock(&tmpres->spinlock);
                        dlm_lockres_put(tmpres);
                        tmpres = NULL;
                        goto lookup;
                }

                /* Grab inflight ref to pin the resource */
                dlm_lockres_grab_inflight_ref(dlm, tmpres);

                spin_unlock(&tmpres->spinlock);
                if (res) {
                        spin_lock(&dlm->track_lock);
                        if (!list_empty(&res->tracking))
                                list_del_init(&res->tracking);
                        else
                                mlog(ML_ERROR, "Resource %.*s not "
                                                "on the Tracking list\n",
                                                res->lockname.len,
                                                res->lockname.name);
                        spin_unlock(&dlm->track_lock);
                        dlm_lockres_put(res);
                }
                res = tmpres;
                goto leave;
        }

        if (!res) {
                spin_unlock(&dlm->spinlock);
                mlog(0, "allocating a new resource\n");
                /* nothing found and we need to allocate one. */
                alloc_mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
                if (!alloc_mle)
                        goto leave;
                res = dlm_new_lockres(dlm, lockid, namelen);
                if (!res)
                        goto leave;
                goto lookup;
        }

        mlog(0, "no lockres found, allocated our own: %p\n", res);

        if (flags & LKM_LOCAL) {
                /* caller knows it's safe to assume it's not mastered elsewhere
                 * DONE!  return right away */
                spin_lock(&res->spinlock);
                dlm_change_lockres_owner(dlm, res, dlm->node_num);
                __dlm_insert_lockres(dlm, res);
                dlm_lockres_grab_inflight_ref(dlm, res);
                spin_unlock(&res->spinlock);
                spin_unlock(&dlm->spinlock);
                /* lockres still marked IN_PROGRESS */
                goto wake_waiters;
        }

        /* check master list to see if another node has started mastering it */
        spin_lock(&dlm->master_lock);

        /* if we found a block, wait for lock to be mastered by another node */
        blocked = dlm_find_mle(dlm, &mle, (char *)lockid, namelen);
        if (blocked) {
                int mig;
                if (mle->type == DLM_MLE_MASTER) {
                        mlog(ML_ERROR, "master entry for nonexistent lock!\n");
                        BUG();
                }
                mig = (mle->type == DLM_MLE_MIGRATION);
                /* if there is a migration in progress, let the migration
                 * finish before continuing.  we can wait for the absence
                 * of the MIGRATION mle: either the migrate finished or
                 * one of the nodes died and the mle was cleaned up.
                 * if there is a BLOCK here, but it already has a master
                 * set, we are too late.  the master does not have a ref
                 * for us in the refmap.  detach the mle and drop it.
                 * either way, go back to the top and start over. */
                if (mig || mle->master != O2NM_MAX_NODES) {
                        BUG_ON(mig && mle->master == dlm->node_num);
                        /* we arrived too late.  the master does not
                         * have a ref for us. retry. */
                        mlog(0, "%s:%.*s: late on %s\n",
                             dlm->name, namelen, lockid,
                             mig ?  "MIGRATION" : "BLOCK");
                        spin_unlock(&dlm->master_lock);
                        spin_unlock(&dlm->spinlock);

                        /* master is known, detach */
                        if (!mig)
                                dlm_mle_detach_hb_events(dlm, mle);
                        dlm_put_mle(mle);
                        mle = NULL;
                        /* this is lame, but we can't wait on either
                         * the mle or lockres waitqueue here */
                        if (mig)
                                msleep(100);
                        goto lookup;
                }
        } else {
                /* go ahead and try to master lock on this node */
                mle = alloc_mle;
                /* make sure this does not get freed below */
                alloc_mle = NULL;
                dlm_init_mle(mle, DLM_MLE_MASTER, dlm, res, NULL, 0);
                set_bit(dlm->node_num, mle->maybe_map);
                __dlm_insert_mle(dlm, mle);

                /* still holding the dlm spinlock, check the recovery map
                 * to see if there are any nodes that still need to be
                 * considered.  these will not appear in the mle nodemap
                 * but they might own this lockres.  wait on them. */
                bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
                if (bit < O2NM_MAX_NODES) {
                        mlog(0, "%s: res %.*s, At least one node (%d) "
                             "to recover before lock mastery can begin\n",
                             dlm->name, namelen, (char *)lockid, bit);
                        wait_on_recovery = 1;
                }
        }

        /* at this point there is either a DLM_MLE_BLOCK or a
         * DLM_MLE_MASTER on the master list, so it's safe to add the
         * lockres to the hashtable.  anyone who finds the lock will
         * still have to wait on the IN_PROGRESS. */

        /* finally add the lockres to its hash bucket */
        __dlm_insert_lockres(dlm, res);

        /* since this lockres is new it doesn't not require the spinlock */
        __dlm_lockres_grab_inflight_ref(dlm, res);

        /* get an extra ref on the mle in case this is a BLOCK
         * if so, the creator of the BLOCK may try to put the last
         * ref at this time in the assert master handler, so we
         * need an extra one to keep from a bad ptr deref. */
        dlm_get_mle_inuse(mle);
        spin_unlock(&dlm->master_lock);
        spin_unlock(&dlm->spinlock);

redo_request:
        while (wait_on_recovery) {
                /* any cluster changes that occurred after dropping the
                 * dlm spinlock would be detectable be a change on the mle,
                 * so we only need to clear out the recovery map once. */
                if (dlm_is_recovery_lock(lockid, namelen)) {
                        mlog(0, "%s: Recovery map is not empty, but must "
                             "master $RECOVERY lock now\n", dlm->name);
                        if (!dlm_pre_master_reco_lockres(dlm, res))
                                wait_on_recovery = 0;
                        else {
                                mlog(0, "%s: waiting 500ms for heartbeat state "
                                    "change\n", dlm->name);
                                msleep(500);
                        }
                        continue;
                }

                dlm_kick_recovery_thread(dlm);
                msleep(1000);
                dlm_wait_for_recovery(dlm);

                spin_lock(&dlm->spinlock);
                bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
                if (bit < O2NM_MAX_NODES) {
                        mlog(0, "%s: res %.*s, At least one node (%d) "
                             "to recover before lock mastery can begin\n",
                             dlm->name, namelen, (char *)lockid, bit);
                        wait_on_recovery = 1;
                } else
                        wait_on_recovery = 0;
                spin_unlock(&dlm->spinlock);

                if (wait_on_recovery)
                        dlm_wait_for_node_recovery(dlm, bit, 10000);
        }

        /* must wait for lock to be mastered elsewhere */
        if (blocked)
                goto wait;

        ret = -EINVAL;
        dlm_node_iter_init(mle->vote_map, &iter);
        while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
                ret = dlm_do_master_request(res, mle, nodenum);
                if (ret < 0)
                        mlog_errno(ret);
                if (mle->master != O2NM_MAX_NODES) {
                        /* found a master ! */
                        if (mle->master <= nodenum)
                                break;
                        /* if our master request has not reached the master
                         * yet, keep going until it does.  this is how the
                         * master will know that asserts are needed back to
                         * the lower nodes. */
                        mlog(0, "%s: res %.*s, Requests only up to %u but "
                             "master is %u, keep going\n", dlm->name, namelen,
                             lockid, nodenum, mle->master);
                }
        }

wait:
        /* keep going until the response map includes all nodes */
        ret = dlm_wait_for_lock_mastery(dlm, res, mle, &blocked);
        if (ret < 0) {
                wait_on_recovery = 1;
                mlog(0, "%s: res %.*s, Node map changed, redo the master "
                     "request now, blocked=%d\n", dlm->name, res->lockname.len,
                     res->lockname.name, blocked);
                if (++tries > 20) {
                        mlog(ML_ERROR, "%s: res %.*s, Spinning on "
                             "dlm_wait_for_lock_mastery, blocked = %d\n",
                             dlm->name, res->lockname.len,
                             res->lockname.name, blocked);
                        dlm_print_one_lock_resource(res);
                        dlm_print_one_mle(mle);
                        tries = 0;
                }
                goto redo_request;
        }

        mlog(0, "%s: res %.*s, Mastered by %u\n", dlm->name, res->lockname.len,
             res->lockname.name, res->owner);
        /* make sure we never continue without this */
        BUG_ON(res->owner == O2NM_MAX_NODES);

        /* master is known, detach if not already detached */
        dlm_mle_detach_hb_events(dlm, mle);
        dlm_put_mle(mle);
        /* put the extra ref */
        dlm_put_mle_inuse(mle);

wake_waiters:
        spin_lock(&res->spinlock);
        res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
        spin_unlock(&res->spinlock);
        wake_up(&res->wq);

leave:
        /* need to free the unused mle */
        if (alloc_mle)
                kmem_cache_free(dlm_mle_cache, alloc_mle);

        return res;
}


#define DLM_MASTERY_TIMEOUT_MS   5000

static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
                                     struct dlm_lock_resource *res,
                                     struct dlm_master_list_entry *mle,
                                     int *blocked)
{
        u8 m;
        int ret, bit;
        int map_changed, voting_done;
        int assert, sleep;

recheck:
        ret = 0;
        assert = 0;

        /* check if another node has already become the owner */
        spin_lock(&res->spinlock);
        if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
                mlog(0, "%s:%.*s: owner is suddenly %u\n", dlm->name,
                     res->lockname.len, res->lockname.name, res->owner);
                spin_unlock(&res->spinlock);
                /* this will cause the master to re-assert across
                 * the whole cluster, freeing up mles */
                if (res->owner != dlm->node_num) {
                        ret = dlm_do_master_request(res, mle, res->owner);
                        if (ret < 0) {
                                /* give recovery a chance to run */
                                mlog(ML_ERROR, "link to %u went down?: %d\n", res->owner, ret);
                                msleep(500);
                                goto recheck;
                        }
                }
                ret = 0;
                goto leave;
        }
        spin_unlock(&res->spinlock);

        spin_lock(&mle->spinlock);
        m = mle->master;
        map_changed = (memcmp(mle->vote_map, mle->node_map,
                              sizeof(mle->vote_map)) != 0);
        voting_done = (memcmp(mle->vote_map, mle->response_map,
                             sizeof(mle->vote_map)) == 0);

        /* restart if we hit any errors */
        if (map_changed) {
                int b;
                mlog(0, "%s: %.*s: node map changed, restarting\n",
                     dlm->name, res->lockname.len, res->lockname.name);
                ret = dlm_restart_lock_mastery(dlm, res, mle, *blocked);
                b = (mle->type == DLM_MLE_BLOCK);
                if ((*blocked && !b) || (!*blocked && b)) {
                        mlog(0, "%s:%.*s: status change: old=%d new=%d\n",
                             dlm->name, res->lockname.len, res->lockname.name,
                             *blocked, b);
                        *blocked = b;
                }
                spin_unlock(&mle->spinlock);
                if (ret < 0) {
                        mlog_errno(ret);
                        goto leave;
                }
                mlog(0, "%s:%.*s: restart lock mastery succeeded, "
                     "rechecking now\n", dlm->name, res->lockname.len,
                     res->lockname.name);
                goto recheck;
        } else {
                if (!voting_done) {
                        mlog(0, "map not changed and voting not done "
                             "for %s:%.*s\n", dlm->name, res->lockname.len,
                             res->lockname.name);
                }
        }

        if (m != O2NM_MAX_NODES) {
                /* another node has done an assert!
                 * all done! */
                sleep = 0;
        } else {
                sleep = 1;
                /* have all nodes responded? */
                if (voting_done && !*blocked) {
                        bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
                        if (dlm->node_num <= bit) {
                                /* my node number is lowest.
                                 * now tell other nodes that I am
                                 * mastering this. */
                                mle->master = dlm->node_num;
                                /* ref was grabbed in get_lock_resource
                                 * will be dropped in dlmlock_master */
                                assert = 1;
                                sleep = 0;
                        }
                        /* if voting is done, but we have not received
                         * an assert master yet, we must sleep */
                }
        }

        spin_unlock(&mle->spinlock);

        /* sleep if we haven't finished voting yet */
        if (sleep) {
                unsigned long timeo = msecs_to_jiffies(DLM_MASTERY_TIMEOUT_MS);
                atomic_set(&mle->woken, 0);
                (void)wait_event_timeout(mle->wq,
                                         (atomic_read(&mle->woken) == 1),
                                         timeo);
                if (res->owner == O2NM_MAX_NODES) {
                        mlog(0, "%s:%.*s: waiting again\n", dlm->name,
                             res->lockname.len, res->lockname.name);
                        goto recheck;
                }
                mlog(0, "done waiting, master is %u\n", res->owner);
                ret = 0;
                goto leave;
        }

        ret = 0;   /* done */
        if (assert) {
                m = dlm->node_num;
                mlog(0, "about to master %.*s here, this=%u\n",
                     res->lockname.len, res->lockname.name, m);
                ret = dlm_do_assert_master(dlm, res, mle->vote_map, 0);
                if (ret) {
                        /* This is a failure in the network path,
                         * not in the response to the assert_master
                         * (any nonzero response is a BUG on this node).
                         * Most likely a socket just got disconnected
                         * due to node death. */
                        mlog_errno(ret);
                }
                /* no longer need to restart lock mastery.
                 * all living nodes have been contacted. */
                ret = 0;
        }

        /* set the lockres owner */
        spin_lock(&res->spinlock);
        /* mastery reference obtained either during
         * assert_master_handler or in get_lock_resource */
        dlm_change_lockres_owner(dlm, res, m);
        spin_unlock(&res->spinlock);

leave:
        return ret;
}

struct dlm_bitmap_diff_iter
{
        int curnode;
        unsigned long *orig_bm;
        unsigned long *cur_bm;
        unsigned long diff_bm[BITS_TO_LONGS(O2NM_MAX_NODES)];
};

enum dlm_node_state_change
{
        NODE_DOWN = -1,
        NODE_NO_CHANGE = 0,
        NODE_UP
};

static void dlm_bitmap_diff_iter_init(struct dlm_bitmap_diff_iter *iter,
                                      unsigned long *orig_bm,
                                      unsigned long *cur_bm)
{
        unsigned long p1, p2;
        int i;

        iter->curnode = -1;
        iter->orig_bm = orig_bm;
        iter->cur_bm = cur_bm;

        for (i = 0; i < BITS_TO_LONGS(O2NM_MAX_NODES); i++) {
                p1 = *(iter->orig_bm + i);
                p2 = *(iter->cur_bm + i);
                iter->diff_bm[i] = (p1 & ~p2) | (p2 & ~p1);
        }
}

static int dlm_bitmap_diff_iter_next(struct dlm_bitmap_diff_iter *iter,
                                     enum dlm_node_state_change *state)
{
        int bit;

        if (iter->curnode >= O2NM_MAX_NODES)
                return -ENOENT;

        bit = find_next_bit(iter->diff_bm, O2NM_MAX_NODES,
                            iter->curnode+1);
        if (bit >= O2NM_MAX_NODES) {
                iter->curnode = O2NM_MAX_NODES;
                return -ENOENT;
        }

        /* if it was there in the original then this node died */
        if (test_bit(bit, iter->orig_bm))
                *state = NODE_DOWN;
        else
                *state = NODE_UP;

        iter->curnode = bit;
        return bit;
}


static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
                                    struct dlm_lock_resource *res,
                                    struct dlm_master_list_entry *mle,
                                    int blocked)
{
        struct dlm_bitmap_diff_iter bdi;
        enum dlm_node_state_change sc;
        int node;
        int ret = 0;

        mlog(0, "something happened such that the "
             "master process may need to be restarted!\n");

        assert_spin_locked(&mle->spinlock);

        dlm_bitmap_diff_iter_init(&bdi, mle->vote_map, mle->node_map);
        node = dlm_bitmap_diff_iter_next(&bdi, &sc);
        while (node >= 0) {
                if (sc == NODE_UP) {
                        /* a node came up.  clear any old vote from
                         * the response map and set it in the vote map
                         * then restart the mastery. */
                        mlog(ML_NOTICE, "node %d up while restarting\n", node);

                        /* redo the master request, but only for the new node */
                        mlog(0, "sending request to new node\n");
                        clear_bit(node, mle->response_map);
                        set_bit(node, mle->vote_map);
                } else {
                        mlog(ML_ERROR, "node down! %d\n", node);
                        if (blocked) {
                                int lowest = find_next_bit(mle->maybe_map,
                                                       O2NM_MAX_NODES, 0);

                                /* act like it was never there */
                                clear_bit(node, mle->maybe_map);

                                if (node == lowest) {
                                        mlog(0, "expected master %u died"
                                            " while this node was blocked "
                                            "waiting on it!\n", node);
                                        lowest = find_next_bit(mle->maybe_map,
                                                        O2NM_MAX_NODES,
                                                        lowest+1);
                                        if (lowest < O2NM_MAX_NODES) {
                                                mlog(0, "%s:%.*s:still "
                                                     "blocked. waiting on %u "
                                                     "now\n", dlm->name,
                                                     res->lockname.len,
                                                     res->lockname.name,
                                                     lowest);
                                        } else {
                                                /* mle is an MLE_BLOCK, but
                                                 * there is now nothing left to
                                                 * block on.  we need to return
                                                 * all the way back out and try
                                                 * again with an MLE_MASTER.
                                                 * dlm_do_local_recovery_cleanup
                                                 * has already run, so the mle
                                                 * refcount is ok */
                                                mlog(0, "%s:%.*s: no "
                                                     "longer blocking. try to "
                                                     "master this here\n",
                                                     dlm->name,
                                                     res->lockname.len,
                                                     res->lockname.name);
                                                mle->type = DLM_MLE_MASTER;
                                                mle->mleres = res;
                                        }
                                }
                        }

                        /* now blank out everything, as if we had never
                         * contacted anyone */
                        memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
                        memset(mle->response_map, 0, sizeof(mle->response_map));
                        /* reset the vote_map to the current node_map */
                        memcpy(mle->vote_map, mle->node_map,
                               sizeof(mle->node_map));
                        /* put myself into the maybe map */
                        if (mle->type != DLM_MLE_BLOCK)
                                set_bit(dlm->node_num, mle->maybe_map);
                }
                ret = -EAGAIN;
                node = dlm_bitmap_diff_iter_next(&bdi, &sc);
        }
        return ret;
}


/*
 * DLM_MASTER_REQUEST_MSG
 *
 * returns: 0 on success,
 *          -errno on a network error
 *
 * on error, the caller should assume the target node is "dead"
 *
 */

static int dlm_do_master_request(struct dlm_lock_resource *res,
                                 struct dlm_master_list_entry *mle, int to)
{
        struct dlm_ctxt *dlm = mle->dlm;
        struct dlm_master_request request;
        int ret, response=0, resend;

        memset(&request, 0, sizeof(request));
        request.node_idx = dlm->node_num;

        BUG_ON(mle->type == DLM_MLE_MIGRATION);

        request.namelen = (u8)mle->mnamelen;
        memcpy(request.name, mle->mname, request.namelen);

again:
        ret = o2net_send_message(DLM_MASTER_REQUEST_MSG, dlm->key, &request,
                                 sizeof(request), to, &response);
        if (ret < 0)  {
                if (ret == -ESRCH) {
                        /* should never happen */
                        mlog(ML_ERROR, "TCP stack not ready!\n");
                        BUG();
                } else if (ret == -EINVAL) {
                        mlog(ML_ERROR, "bad args passed to o2net!\n");
                        BUG();
                } else if (ret == -ENOMEM) {
                        mlog(ML_ERROR, "out of memory while trying to send "
                             "network message!  retrying\n");
                        /* this is totally crude */
                        msleep(50);
                        goto again;
                } else if (!dlm_is_host_down(ret)) {
                        /* not a network error. bad. */
                        mlog_errno(ret);
                        mlog(ML_ERROR, "unhandled error!");
                        BUG();
                }
                /* all other errors should be network errors,
                 * and likely indicate node death */
                mlog(ML_ERROR, "link to %d went down!\n", to);
                goto out;
        }

        ret = 0;
        resend = 0;
        spin_lock(&mle->spinlock);
        switch (response) {
                case DLM_MASTER_RESP_YES:
                        set_bit(to, mle->response_map);
                        mlog(0, "node %u is the master, response=YES\n", to);
                        mlog(0, "%s:%.*s: master node %u now knows I have a "
                             "reference\n", dlm->name, res->lockname.len,
                             res->lockname.name, to);
                        mle->master = to;
                        break;
                case DLM_MASTER_RESP_NO:
                        mlog(0, "node %u not master, response=NO\n", to);
                        set_bit(to, mle->response_map);
                        break;
                case DLM_MASTER_RESP_MAYBE:
                        mlog(0, "node %u not master, response=MAYBE\n", to);
                        set_bit(to, mle->response_map);
                        set_bit(to, mle->maybe_map);
                        break;
                case DLM_MASTER_RESP_ERROR:
                        mlog(0, "node %u hit an error, resending\n", to);
                        resend = 1;
                        response = 0;
                        break;
                default:
                        mlog(ML_ERROR, "bad response! %u\n", response);
                        BUG();
        }
        spin_unlock(&mle->spinlock);
        if (resend) {
                /* this is also totally crude */
                msleep(50);
                goto again;
        }

out:
        return ret;
}

/*
 * locks that can be taken here:
 * dlm->spinlock
 * res->spinlock
 * mle->spinlock
 * dlm->master_list
 *
 * if possible, TRIM THIS DOWN!!!
 */
int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data,
                               void **ret_data)
{
        u8 response = DLM_MASTER_RESP_MAYBE;
        struct dlm_ctxt *dlm = data;
        struct dlm_lock_resource *res = NULL;
        struct dlm_master_request *request = (struct dlm_master_request *) msg->buf;
        struct dlm_master_list_entry *mle = NULL, *tmpmle = NULL;
        char *name;
        unsigned int namelen, hash;
        int found, ret;
        int set_maybe;
        int dispatch_assert = 0;
        int dispatched = 0;

        if (!dlm_grab(dlm))
                return DLM_MASTER_RESP_NO;

        if (!dlm_domain_fully_joined(dlm)) {
                response = DLM_MASTER_RESP_NO;
                goto send_response;
        }

        name = request->name;
        namelen = request->namelen;
        hash = dlm_lockid_hash(name, namelen);

        if (namelen > DLM_LOCKID_NAME_MAX) {
                response = DLM_IVBUFLEN;
                goto send_response;
        }

way_up_top:
        spin_lock(&dlm->spinlock);
        res = __dlm_lookup_lockres(dlm, name, namelen, hash);
        if (res) {
                spin_unlock(&dlm->spinlock);

                /* take care of the easy cases up front */
                spin_lock(&res->spinlock);

                /*
                 * Right after dlm spinlock was released, dlm_thread could have
                 * purged the lockres. Check if lockres got unhashed. If so
                 * start over.
                 */
                if (hlist_unhashed(&res->hash_node)) {
                        spin_unlock(&res->spinlock);
                        dlm_lockres_put(res);
                        goto way_up_top;
                }

                if (res->state & (DLM_LOCK_RES_RECOVERING|
                                  DLM_LOCK_RES_MIGRATING)) {
                        spin_unlock(&res->spinlock);
                        mlog(0, "returning DLM_MASTER_RESP_ERROR since res is "
                             "being recovered/migrated\n");
                        response = DLM_MASTER_RESP_ERROR;
                        if (mle)
                                kmem_cache_free(dlm_mle_cache, mle);
                        goto send_response;
                }

                if (res->owner == dlm->node_num) {
                        dlm_lockres_set_refmap_bit(dlm, res, request->node_idx);
                        spin_unlock(&res->spinlock);
                        response = DLM_MASTER_RESP_YES;
                        if (mle)
                                kmem_cache_free(dlm_mle_cache, mle);

                        /* this node is the owner.
                         * there is some extra work that needs to
                         * happen now.  the requesting node has
                         * caused all nodes up to this one to
                         * create mles.  this node now needs to
                         * go back and clean those up. */
                        dispatch_assert = 1;
                        goto send_response;
                } else if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
                        spin_unlock(&res->spinlock);
                        // mlog(0, "node %u is the master\n", res->owner);
                        response = DLM_MASTER_RESP_NO;
                        if (mle)
                                kmem_cache_free(dlm_mle_cache, mle);
                        goto send_response;
                }

                /* ok, there is no owner.  either this node is
                 * being blocked, or it is actively trying to
                 * master this lock. */
                if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
                        mlog(ML_ERROR, "lock with no owner should be "
                             "in-progress!\n");
                        BUG();
                }

                // mlog(0, "lockres is in progress...\n");
                spin_lock(&dlm->master_lock);
                found = dlm_find_mle(dlm, &tmpmle, name, namelen);
                if (!found) {
                        mlog(ML_ERROR, "no mle found for this lock!\n");
                        BUG();
                }
                set_maybe = 1;
                spin_lock(&tmpmle->spinlock);
                if (tmpmle->type == DLM_MLE_BLOCK) {
                        // mlog(0, "this node is waiting for "
                        // "lockres to be mastered\n");
                        response = DLM_MASTER_RESP_NO;
                } else if (tmpmle->type == DLM_MLE_MIGRATION) {
                        mlog(0, "node %u is master, but trying to migrate to "
                             "node %u.\n", tmpmle->master, tmpmle->new_master);
                        if (tmpmle->master == dlm->node_num) {
                                mlog(ML_ERROR, "no owner on lockres, but this "
                                     "node is trying to migrate it to %u?!\n",
                                     tmpmle->new_master);
                                BUG();
                        } else {
                                /* the real master can respond on its own */
                                response = DLM_MASTER_RESP_NO;
                        }
                } else if (tmpmle->master != DLM_LOCK_RES_OWNER_UNKNOWN) {
                        set_maybe = 0;
                        if (tmpmle->master == dlm->node_num) {
                                response = DLM_MASTER_RESP_YES;
                                /* this node will be the owner.
                                 * go back and clean the mles on any
                                 * other nodes */
                                dispatch_assert = 1;
                                dlm_lockres_set_refmap_bit(dlm, res,
                                                           request->node_idx);
                        } else
                                response = DLM_MASTER_RESP_NO;
                } else {
                        // mlog(0, "this node is attempting to "
                        // "master lockres\n");
                        response = DLM_MASTER_RESP_MAYBE;
                }
                if (set_maybe)
                        set_bit(request->node_idx, tmpmle->maybe_map);
                spin_unlock(&tmpmle->spinlock);

                spin_unlock(&dlm->master_lock);
                spin_unlock(&res->spinlock);

                /* keep the mle attached to heartbeat events */
                dlm_put_mle(tmpmle);
                if (mle)
                        kmem_cache_free(dlm_mle_cache, mle);
                goto send_response;
        }

        /*
         * lockres doesn't exist on this node
         * if there is an MLE_BLOCK, return NO
         * if there is an MLE_MASTER, return MAYBE
         * otherwise, add an MLE_BLOCK, return NO
         */
        spin_lock(&dlm->master_lock);
        found = dlm_find_mle(dlm, &tmpmle, name, namelen);
        if (!found) {
                /* this lockid has never been seen on this node yet */
                // mlog(0, "no mle found\n");
                if (!mle) {
                        spin_unlock(&dlm->master_lock);
                        spin_unlock(&dlm->spinlock);

                        mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
                        if (!mle) {
                                response = DLM_MASTER_RESP_ERROR;
                                mlog_errno(-ENOMEM);
                                goto send_response;
                        }
                        goto way_up_top;
                }

                // mlog(0, "this is second time thru, already allocated, "
                // "add the block.\n");
                dlm_init_mle(mle, DLM_MLE_BLOCK, dlm, NULL, name, namelen);
                set_bit(request->node_idx, mle->maybe_map);
                __dlm_insert_mle(dlm, mle);
                response = DLM_MASTER_RESP_NO;
        } else {
                spin_lock(&tmpmle->spinlock);
                if (tmpmle->master == dlm->node_num) {
                        mlog(ML_ERROR, "no lockres, but an mle with this node as master!\n");
                        BUG();
                }
                if (tmpmle->type == DLM_MLE_BLOCK)
                        response = DLM_MASTER_RESP_NO;
                else if (tmpmle->type == DLM_MLE_MIGRATION) {
                        mlog(0, "migration mle was found (%u->%u)\n",
                             tmpmle->master, tmpmle->new_master);
                        /* real master can respond on its own */
                        response = DLM_MASTER_RESP_NO;
                } else
                        response = DLM_MASTER_RESP_MAYBE;
                set_bit(request->node_idx, tmpmle->maybe_map);
                spin_unlock(&tmpmle->spinlock);
        }
        spin_unlock(&dlm->master_lock);
        spin_unlock(&dlm->spinlock);

        if (found) {
                /* keep the mle attached to heartbeat events */
                dlm_put_mle(tmpmle);
        }
send_response:
        /*
         * __dlm_lookup_lockres() grabbed a reference to this lockres.
         * The reference is released by dlm_assert_master_worker() under
         * the call to dlm_dispatch_assert_master().  If
         * dlm_assert_master_worker() isn't called, we drop it here.
         */
        if (dispatch_assert) {
                mlog(0, "%u is the owner of %.*s, cleaning everyone else\n",
                             dlm->node_num, res->lockname.len, res->lockname.name);
                spin_lock(&res->spinlock);
                ret = dlm_dispatch_assert_master(dlm, res, 0, request->node_idx,
                                                 DLM_ASSERT_MASTER_MLE_CLEANUP);
                if (ret < 0) {
                        mlog(ML_ERROR, "failed to dispatch assert master work\n");
                        response = DLM_MASTER_RESP_ERROR;
                        spin_unlock(&res->spinlock);
                        dlm_lockres_put(res);
                } else {
                        dispatched = 1;
                        __dlm_lockres_grab_inflight_worker(dlm, res);
                        spin_unlock(&res->spinlock);
                }
        } else {
                if (res)
                        dlm_lockres_put(res);
        }

        if (!dispatched)
                dlm_put(dlm);
        return response;
}

/*
 * DLM_ASSERT_MASTER_MSG
 */


/*
 * NOTE: this can be used for debugging
 * can periodically run all locks owned by this node
 * and re-assert across the cluster...
 */
static int dlm_do_assert_master(struct dlm_ctxt *dlm,
                                struct dlm_lock_resource *res,
                                void *nodemap, u32 flags)
{
        struct dlm_assert_master assert;
        int to, tmpret;
        struct dlm_node_iter iter;
        int ret = 0;
        int reassert;
        const char *lockname = res->lockname.name;
        unsigned int namelen = res->lockname.len;

        BUG_ON(namelen > O2NM_MAX_NAME_LEN);

        spin_lock(&res->spinlock);
        res->state |= DLM_LOCK_RES_SETREF_INPROG;
        spin_unlock(&res->spinlock);

again:
        reassert = 0;

        /* note that if this nodemap is empty, it returns 0 */
        dlm_node_iter_init(nodemap, &iter);
        while ((to = dlm_node_iter_next(&iter)) >= 0) {
                int r = 0;
                struct dlm_master_list_entry *mle = NULL;

                mlog(0, "sending assert master to %d (%.*s)\n", to,
                     namelen, lockname);
                memset(&assert, 0, sizeof(assert));
                assert.node_idx = dlm->node_num;
                assert.namelen = namelen;
                memcpy(assert.name, lockname, namelen);
                assert.flags = cpu_to_be32(flags);

                tmpret = o2net_send_message(DLM_ASSERT_MASTER_MSG, dlm->key,
                                            &assert, sizeof(assert), to, &r);
                if (tmpret < 0) {
                        mlog(ML_ERROR, "Error %d when sending message %u (key "
                             "0x%x) to node %u\n", tmpret,
                             DLM_ASSERT_MASTER_MSG, dlm->key, to);
                        if (!dlm_is_host_down(tmpret)) {
                                mlog(ML_ERROR, "unhandled error=%d!\n", tmpret);
                                BUG();
                        }
                        /* a node died.  finish out the rest of the nodes. */
                        mlog(0, "link to %d went down!\n", to);
                        /* any nonzero status return will do */
                        ret = tmpret;
                        r = 0;
                } else if (r < 0) {
                        /* ok, something horribly messed.  kill thyself. */
                        mlog(ML_ERROR,"during assert master of %.*s to %u, "
                             "got %d.\n", namelen, lockname, to, r);
                        spin_lock(&dlm->spinlock);
                        spin_lock(&dlm->master_lock);
                        if (dlm_find_mle(dlm, &mle, (char *)lockname,
                                         namelen)) {
                                dlm_print_one_mle(mle);
                                __dlm_put_mle(mle);
                        }
                        spin_unlock(&dlm->master_lock);
                        spin_unlock(&dlm->spinlock);
                        BUG();
                }

                if (r & DLM_ASSERT_RESPONSE_REASSERT &&
                    !(r & DLM_ASSERT_RESPONSE_MASTERY_REF)) {
                                mlog(ML_ERROR, "%.*s: very strange, "
                                     "master MLE but no lockres on %u\n",
                                     namelen, lockname, to);
                }

                if (r & DLM_ASSERT_RESPONSE_REASSERT) {
                        mlog(0, "%.*s: node %u create mles on other "
                             "nodes and requests a re-assert\n",
                             namelen, lockname, to);
                        reassert = 1;
                }
                if (r & DLM_ASSERT_RESPONSE_MASTERY_REF) {
                        mlog(0, "%.*s: node %u has a reference to this "
                             "lockres, set the bit in the refmap\n",
                             namelen, lockname, to);
                        spin_lock(&res->spinlock);
                        dlm_lockres_set_refmap_bit(dlm, res, to);
                        spin_unlock(&res->spinlock);
                }
        }

        if (reassert)
                goto again;

        spin_lock(&res->spinlock);
        res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
        spin_unlock(&res->spinlock);
        wake_up(&res->wq);

        return ret;
}

/*
 * locks that can be taken here:
 * dlm->spinlock
 * res->spinlock
 * mle->spinlock
 * dlm->master_list
 *
 * if possible, TRIM THIS DOWN!!!
 */
int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data,
                              void **ret_data)
{
        struct dlm_ctxt *dlm = data;
        struct dlm_master_list_entry *mle = NULL;
        struct dlm_assert_master *assert = (struct dlm_assert_master *)msg->buf;
        struct dlm_lock_resource *res = NULL;
        char *name;
        unsigned int namelen, hash;
        u32 flags;
        int master_request = 0, have_lockres_ref = 0;
        int ret = 0;

        if (!dlm_grab(dlm))
                return 0;

        name = assert->name;
        namelen = assert->namelen;
        hash = dlm_lockid_hash(name, namelen);
        flags = be32_to_cpu(assert->flags);

        if (namelen > DLM_LOCKID_NAME_MAX) {
                mlog(ML_ERROR, "Invalid name length!");
                goto done;
        }

        spin_lock(&dlm->spinlock);

        if (flags)
                mlog(0, "assert_master with flags: %u\n", flags);

        /* find the MLE */
        spin_lock(&dlm->master_lock);
        if (!dlm_find_mle(dlm, &mle, name, namelen)) {
                /* not an error, could be master just re-asserting */
                mlog(0, "just got an assert_master from %u, but no "
                     "MLE for it! (%.*s)\n", assert->node_idx,
                     namelen, name);
        } else {
                int bit = find_next_bit (mle->maybe_map, O2NM_MAX_NODES, 0);
                if (bit >= O2NM_MAX_NODES) {
                        /* not necessarily an error, though less likely.
                         * could be master just re-asserting. */
                        mlog(0, "no bits set in the maybe_map, but %u "
                             "is asserting! (%.*s)\n", assert->node_idx,
                             namelen, name);
                } else if (bit != assert->node_idx) {
                        if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
                                mlog(0, "master %u was found, %u should "
                                     "back off\n", assert->node_idx, bit);
                        } else {
                                /* with the fix for bug 569, a higher node
                                 * number winning the mastery will respond
                                 * YES to mastery requests, but this node
                                 * had no way of knowing.  let it pass. */
                                mlog(0, "%u is the lowest node, "
                                     "%u is asserting. (%.*s)  %u must "
                                     "have begun after %u won.\n", bit,
                                     assert->node_idx, namelen, name, bit,
                                     assert->node_idx);
                        }
                }
                if (mle->type == DLM_MLE_MIGRATION) {
                        if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
                                mlog(0, "%s:%.*s: got cleanup assert"
                                     " from %u for migration\n",
                                     dlm->name, namelen, name,
                                     assert->node_idx);
                        } else if (!(flags & DLM_ASSERT_MASTER_FINISH_MIGRATION)) {
                                mlog(0, "%s:%.*s: got unrelated assert"
                                     " from %u for migration, ignoring\n",
                                     dlm->name, namelen, name,
                                     assert->node_idx);
                                __dlm_put_mle(mle);
                                spin_unlock(&dlm->master_lock);
                                spin_unlock(&dlm->spinlock);
                                goto done;
                        }
                }
        }
        spin_unlock(&dlm->master_lock);

        /* ok everything checks out with the MLE
         * now check to see if there is a lockres */
        res = __dlm_lookup_lockres(dlm, name, namelen, hash);
        if (res) {
                spin_lock(&res->spinlock);
                if (res->state & DLM_LOCK_RES_RECOVERING)  {
                        mlog(ML_ERROR, "%u asserting but %.*s is "
                             "RECOVERING!\n", assert->node_idx, namelen, name);
                        goto kill;
                }
                if (!mle) {
                        if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN &&
                            res->owner != assert->node_idx) {
                                mlog(ML_ERROR, "DIE! Mastery assert from %u, "
                                     "but current owner is %u! (%.*s)\n",
                                     assert->node_idx, res->owner, namelen,
                                     name);
                                __dlm_print_one_lock_resource(res);
                                BUG();
                        }
                } else if (mle->type != DLM_MLE_MIGRATION) {
                        if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
                                /* owner is just re-asserting */
                                if (res->owner == assert->node_idx) {
                                        mlog(0, "owner %u re-asserting on "
                                             "lock %.*s\n", assert->node_idx,
                                             namelen, name);
                                        goto ok;
                                }
                                mlog(ML_ERROR, "got assert_master from "
                                     "node %u, but %u is the owner! "
                                     "(%.*s)\n", assert->node_idx,
                                     res->owner, namelen, name);
                                goto kill;
                        }
                        if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
                                mlog(ML_ERROR, "got assert from %u, but lock "
                                     "with no owner should be "
                                     "in-progress! (%.*s)\n",
                                     assert->node_idx,
                                     namelen, name);
                                goto kill;
                        }
                } else /* mle->type == DLM_MLE_MIGRATION */ {
                        /* should only be getting an assert from new master */
                        if (assert->node_idx != mle->new_master) {
                                mlog(ML_ERROR, "got assert from %u, but "
                                     "new master is %u, and old master "
                                     "was %u (%.*s)\n",
                                     assert->node_idx, mle->new_master,
                                     mle->master, namelen, name);
                                goto kill;
                        }

                }
ok:
                spin_unlock(&res->spinlock);
        }

        // mlog(0, "woo!  got an assert_master from node %u!\n",
        //           assert->node_idx);
        if (mle) {
                int extra_ref = 0;
                int nn = -1;
                int rr, err = 0;

                spin_lock(&mle->spinlock);
                if (mle->type == DLM_MLE_BLOCK || mle->type == DLM_MLE_MIGRATION)
                        extra_ref = 1;
                else {
                        /* MASTER mle: if any bits set in the response map
                         * then the calling node needs to re-assert to clear
                         * up nodes that this node contacted */
                        while ((nn = find_next_bit (mle->response_map, O2NM_MAX_NODES,
                                                    nn+1)) < O2NM_MAX_NODES) {
                                if (nn != dlm->node_num && nn != assert->node_idx) {
                                        master_request = 1;
                                        break;
                                }
                        }
                }
                mle->master = assert->node_idx;
                atomic_set(&mle->woken, 1);
                wake_up(&mle->wq);
                spin_unlock(&mle->spinlock);

                if (res) {
                        int wake = 0;
                        spin_lock(&res->spinlock);
                        if (mle->type == DLM_MLE_MIGRATION) {
                                mlog(0, "finishing off migration of lockres %.*s, "
                                        "from %u to %u\n",
                                        res->lockname.len, res->lockname.name,
                                        dlm->node_num, mle->new_master);
                                res->state &= ~DLM_LOCK_RES_MIGRATING;
                                wake = 1;
                                dlm_change_lockres_owner(dlm, res, mle->new_master);
                                BUG_ON(res->state & DLM_LOCK_RES_DIRTY);
                        } else {
                                dlm_change_lockres_owner(dlm, res, mle->master);
                        }
                        spin_unlock(&res->spinlock);
                        have_lockres_ref = 1;
                        if (wake)
                                wake_up(&res->wq);
                }

                /* master is known, detach if not already detached.
                 * ensures that only one assert_master call will happen
                 * on this mle. */
                spin_lock(&dlm->master_lock);

                rr = kref_read(&mle->mle_refs);
                if (mle->inuse > 0) {
                        if (extra_ref && rr < 3)
                                err = 1;
                        else if (!extra_ref && rr < 2)
                                err = 1;
                } else {
                        if (extra_ref && rr < 2)
                                err = 1;
                        else if (!extra_ref && rr < 1)
                                err = 1;
                }
                if (err) {
                        mlog(ML_ERROR, "%s:%.*s: got assert master from %u "
                             "that will mess up this node, refs=%d, extra=%d, "
                             "inuse=%d\n", dlm->name, namelen, name,
                             assert->node_idx, rr, extra_ref, mle->inuse);
                        dlm_print_one_mle(mle);
                }
                __dlm_unlink_mle(dlm, mle);
                __dlm_mle_detach_hb_events(dlm, mle);
                __dlm_put_mle(mle);
                if (extra_ref) {
                        /* the assert master message now balances the extra
                         * ref given by the master / migration request message.
                         * if this is the last put, it will be removed
                         * from the list. */
                        __dlm_put_mle(mle);
                }
                spin_unlock(&dlm->master_lock);
        } else if (res) {
                if (res->owner != assert->node_idx) {
                        mlog(0, "assert_master from %u, but current "
                             "owner is %u (%.*s), no mle\n", assert->node_idx,
                             res->owner, namelen, name);
                }
        }
        spin_unlock(&dlm->spinlock);

done:
        ret = 0;
        if (res) {
                spin_lock(&res->spinlock);
                res->state |= DLM_LOCK_RES_SETREF_INPROG;
                spin_unlock(&res->spinlock);
                *ret_data = (void *)res;
        }
        dlm_put(dlm);