dlmdomain.c

linux2.6.16版本

/* -*- mode: c; c-basic-offset: 8; -*-
 * vim: noexpandtab sw=8 ts=8 sts=0:
 *
 * dlmdomain.c
 *
 * defines domain join / leave apis
 *
 * Copyright (C) 2004 Oracle.  All rights reserved.
 *
 * This program is free software; you can redistribute it and/or
 * modify it under the terms of the GNU General Public
 * License as published by the Free Software Foundation; either
 * version 2 of the License, or (at your option) any later version.
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * General Public License for more details.
 *
 * You should have received a copy of the GNU General Public
 * License along with this program; if not, write to the
 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
 * Boston, MA 021110-1307, USA.
 *
 */

#include <linux/module.h>
#include <linux/types.h>
#include <linux/slab.h>
#include <linux/highmem.h>
#include <linux/utsname.h>
#include <linux/init.h>
#include <linux/spinlock.h>
#include <linux/delay.h>
#include <linux/err.h>

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

#include "dlmapi.h"
#include "dlmcommon.h"

#include "dlmdebug.h"
#include "dlmdomain.h"

#include "dlmver.h"

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

/*
 *
 * spinlock lock ordering: if multiple locks are needed, obey this ordering:
 *    dlm_domain_lock
 *    struct dlm_ctxt->spinlock
 *    struct dlm_lock_resource->spinlock
 *    struct dlm_ctxt->master_lock
 *    struct dlm_ctxt->ast_lock
 *    dlm_master_list_entry->spinlock
 *    dlm_lock->spinlock
 *
 */

spinlock_t dlm_domain_lock = SPIN_LOCK_UNLOCKED;
LIST_HEAD(dlm_domains);
static DECLARE_WAIT_QUEUE_HEAD(dlm_domain_events);

#define DLM_DOMAIN_BACKOFF_MS 200

static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data);
static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data);
static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data);
static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data);

static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm);

void __dlm_unhash_lockres(struct dlm_lock_resource *lockres)
{
	hlist_del_init(&lockres->hash_node);
	dlm_lockres_put(lockres);
}

void __dlm_insert_lockres(struct dlm_ctxt *dlm,
		       struct dlm_lock_resource *res)
{
	struct hlist_head *bucket;
	struct qstr *q;

	assert_spin_locked(&dlm->spinlock);

	q = &res->lockname;
	q->hash = full_name_hash(q->name, q->len);
	bucket = &(dlm->lockres_hash[q->hash % DLM_HASH_BUCKETS]);

	/* get a reference for our hashtable */
	dlm_lockres_get(res);

	hlist_add_head(&res->hash_node, bucket);
}

struct dlm_lock_resource * __dlm_lookup_lockres(struct dlm_ctxt *dlm,
					 const char *name,
					 unsigned int len)
{
	unsigned int hash;
	struct hlist_node *iter;
	struct dlm_lock_resource *tmpres=NULL;
	struct hlist_head *bucket;

	mlog_entry("%.*s\n", len, name);

	assert_spin_locked(&dlm->spinlock);

	hash = full_name_hash(name, len);

	bucket = &(dlm->lockres_hash[hash % DLM_HASH_BUCKETS]);

	/* check for pre-existing lock */
	hlist_for_each(iter, bucket) {
		tmpres = hlist_entry(iter, struct dlm_lock_resource, hash_node);
		if (tmpres->lockname.len == len &&
		    memcmp(tmpres->lockname.name, name, len) == 0) {
			dlm_lockres_get(tmpres);
			break;
		}

		tmpres = NULL;
	}
	return tmpres;
}

struct dlm_lock_resource * dlm_lookup_lockres(struct dlm_ctxt *dlm,
				    const char *name,
				    unsigned int len)
{
	struct dlm_lock_resource *res;

	spin_lock(&dlm->spinlock);
	res = __dlm_lookup_lockres(dlm, name, len);
	spin_unlock(&dlm->spinlock);
	return res;
}

static struct dlm_ctxt * __dlm_lookup_domain_full(const char *domain, int len)
{
	struct dlm_ctxt *tmp = NULL;
	struct list_head *iter;

	assert_spin_locked(&dlm_domain_lock);

	/* tmp->name here is always NULL terminated,
	 * but domain may not be! */
	list_for_each(iter, &dlm_domains) {
		tmp = list_entry (iter, struct dlm_ctxt, list);
		if (strlen(tmp->name) == len &&
		    memcmp(tmp->name, domain, len)==0)
			break;
		tmp = NULL;
	}

	return tmp;
}

/* For null terminated domain strings ONLY */
static struct dlm_ctxt * __dlm_lookup_domain(const char *domain)
{
	assert_spin_locked(&dlm_domain_lock);

	return __dlm_lookup_domain_full(domain, strlen(domain));
}


/* returns true on one of two conditions:
 * 1) the domain does not exist
 * 2) the domain exists and it's state is "joined" */
static int dlm_wait_on_domain_helper(const char *domain)
{
	int ret = 0;
	struct dlm_ctxt *tmp = NULL;

	spin_lock(&dlm_domain_lock);

	tmp = __dlm_lookup_domain(domain);
	if (!tmp)
		ret = 1;
	else if (tmp->dlm_state == DLM_CTXT_JOINED)
		ret = 1;

	spin_unlock(&dlm_domain_lock);
	return ret;
}

static void dlm_free_ctxt_mem(struct dlm_ctxt *dlm)
{
	if (dlm->lockres_hash)
		free_page((unsigned long) dlm->lockres_hash);

	if (dlm->name)
		kfree(dlm->name);

	kfree(dlm);
}

/* A little strange - this function will be called while holding
 * dlm_domain_lock and is expected to be holding it on the way out. We
 * will however drop and reacquire it multiple times */
static void dlm_ctxt_release(struct kref *kref)
{
	struct dlm_ctxt *dlm;

	dlm = container_of(kref, struct dlm_ctxt, dlm_refs);

	BUG_ON(dlm->num_joins);
	BUG_ON(dlm->dlm_state == DLM_CTXT_JOINED);

	/* we may still be in the list if we hit an error during join. */
	list_del_init(&dlm->list);

	spin_unlock(&dlm_domain_lock);

	mlog(0, "freeing memory from domain %s\n", dlm->name);

	wake_up(&dlm_domain_events);

	dlm_free_ctxt_mem(dlm);

	spin_lock(&dlm_domain_lock);
}

void dlm_put(struct dlm_ctxt *dlm)
{
	spin_lock(&dlm_domain_lock);
	kref_put(&dlm->dlm_refs, dlm_ctxt_release);
	spin_unlock(&dlm_domain_lock);
}

static void __dlm_get(struct dlm_ctxt *dlm)
{
	kref_get(&dlm->dlm_refs);
}

/* given a questionable reference to a dlm object, gets a reference if
 * it can find it in the list, otherwise returns NULL in which case
 * you shouldn't trust your pointer. */
struct dlm_ctxt *dlm_grab(struct dlm_ctxt *dlm)
{
	struct list_head *iter;
	struct dlm_ctxt *target = NULL;

	spin_lock(&dlm_domain_lock);

	list_for_each(iter, &dlm_domains) {
		target = list_entry (iter, struct dlm_ctxt, list);

		if (target == dlm) {
			__dlm_get(target);
			break;
		}

		target = NULL;
	}

	spin_unlock(&dlm_domain_lock);

	return target;
}

int dlm_domain_fully_joined(struct dlm_ctxt *dlm)
{
	int ret;

	spin_lock(&dlm_domain_lock);
	ret = (dlm->dlm_state == DLM_CTXT_JOINED) ||
		(dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN);
	spin_unlock(&dlm_domain_lock);

	return ret;
}

static void dlm_complete_dlm_shutdown(struct dlm_ctxt *dlm)
{
	dlm_unregister_domain_handlers(dlm);
	dlm_complete_thread(dlm);
	dlm_complete_recovery_thread(dlm);

	/* We've left the domain. Now we can take ourselves out of the
	 * list and allow the kref stuff to help us free the
	 * memory. */
	spin_lock(&dlm_domain_lock);
	list_del_init(&dlm->list);
	spin_unlock(&dlm_domain_lock);

	/* Wake up anyone waiting for us to remove this domain */
	wake_up(&dlm_domain_events);
}

static void dlm_migrate_all_locks(struct dlm_ctxt *dlm)
{
	int i;
	struct dlm_lock_resource *res;

	mlog(0, "Migrating locks from domain %s\n", dlm->name);
restart:
	spin_lock(&dlm->spinlock);
	for (i = 0; i < DLM_HASH_BUCKETS; i++) {
		while (!hlist_empty(&dlm->lockres_hash[i])) {
			res = hlist_entry(dlm->lockres_hash[i].first,
					  struct dlm_lock_resource, hash_node);
			/* need reference when manually grabbing lockres */
			dlm_lockres_get(res);
			/* this should unhash the lockres
			 * and exit with dlm->spinlock */
			mlog(0, "purging res=%p\n", res);
			if (dlm_lockres_is_dirty(dlm, res)) {
				/* HACK!  this should absolutely go.
				 * need to figure out why some empty
				 * lockreses are still marked dirty */
				mlog(ML_ERROR, "lockres %.*s dirty!\n",
				     res->lockname.len, res->lockname.name);

				spin_unlock(&dlm->spinlock);
				dlm_kick_thread(dlm, res);
				wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res));
				dlm_lockres_put(res);
				goto restart;
			}
			dlm_purge_lockres(dlm, res);
			dlm_lockres_put(res);
		}
	}
	spin_unlock(&dlm->spinlock);

	mlog(0, "DONE Migrating locks from domain %s\n", dlm->name);
}

static int dlm_no_joining_node(struct dlm_ctxt *dlm)
{
	int ret;

	spin_lock(&dlm->spinlock);
	ret = dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN;
	spin_unlock(&dlm->spinlock);

	return ret;
}

static void dlm_mark_domain_leaving(struct dlm_ctxt *dlm)
{
	/* Yikes, a double spinlock! I need domain_lock for the dlm
	 * state and the dlm spinlock for join state... Sorry! */
again:
	spin_lock(&dlm_domain_lock);
	spin_lock(&dlm->spinlock);

	if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
		mlog(0, "Node %d is joining, we wait on it.\n",
			  dlm->joining_node);
		spin_unlock(&dlm->spinlock);
		spin_unlock(&dlm_domain_lock);

		wait_event(dlm->dlm_join_events, dlm_no_joining_node(dlm));
		goto again;
	}

	dlm->dlm_state = DLM_CTXT_LEAVING;
	spin_unlock(&dlm->spinlock);
	spin_unlock(&dlm_domain_lock);
}

static void __dlm_print_nodes(struct dlm_ctxt *dlm)
{
	int node = -1;

	assert_spin_locked(&dlm->spinlock);

	mlog(ML_NOTICE, "Nodes in my domain (\"%s\"):\n", dlm->name);

	while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
				     node + 1)) < O2NM_MAX_NODES) {
		mlog(ML_NOTICE, " node %d\n", node);
	}
}

static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data)
{
	struct dlm_ctxt *dlm = data;
	unsigned int node;
	struct dlm_exit_domain *exit_msg = (struct dlm_exit_domain *) msg->buf;

	mlog_entry("%p %u %p", msg, len, data);

	if (!dlm_grab(dlm))
		return 0;

	node = exit_msg->node_idx;

	mlog(0, "Node %u leaves domain %s\n", node, dlm->name);

	spin_lock(&dlm->spinlock);
	clear_bit(node, dlm->domain_map);
	__dlm_print_nodes(dlm);

	/* notify anything attached to the heartbeat events */
	dlm_hb_event_notify_attached(dlm, node, 0);

	spin_unlock(&dlm->spinlock);

	dlm_put(dlm);

	return 0;
}

static int dlm_send_one_domain_exit(struct dlm_ctxt *dlm,
				    unsigned int node)
{
	int status;
	struct dlm_exit_domain leave_msg;

	mlog(0, "Asking node %u if we can leave the domain %s me = %u\n",
		  node, dlm->name, dlm->node_num);

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

	status = o2net_send_message(DLM_EXIT_DOMAIN_MSG, dlm->key,
				    &leave_msg, sizeof(leave_msg), node,
				    NULL);

	mlog(0, "status return %d from o2net_send_message\n", status);

	return status;
}


static void dlm_leave_domain(struct dlm_ctxt *dlm)
{
	int node, clear_node, status;

	/* At this point we've migrated away all our locks and won't
	 * accept mastership of new ones. The dlm is responsible for
	 * almost nothing now. We make sure not to confuse any joining
	 * nodes and then commence shutdown procedure. */

	spin_lock(&dlm->spinlock);
	/* Clear ourselves from the domain map */
	clear_bit(dlm->node_num, dlm->domain_map);
	while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
				     0)) < O2NM_MAX_NODES) {
		/* Drop the dlm spinlock. This is safe wrt the domain_map.
		 * -nodes cannot be added now as the
		 *   query_join_handlers knows to respond with OK_NO_MAP
		 * -we catch the right network errors if a node is
		 *   removed from the map while we're sending him the
		 *   exit message. */
		spin_unlock(&dlm->spinlock);

		clear_node = 1;

		status = dlm_send_one_domain_exit(dlm, node);
		if (status < 0 &&
		    status != -ENOPROTOOPT &&
		    status != -ENOTCONN) {
			mlog(ML_NOTICE, "Error %d sending domain exit message "
			     "to node %d\n", status, node);

			/* Not sure what to do here but lets sleep for
			 * a bit in case this was a transient
			 * error... */
			msleep(DLM_DOMAIN_BACKOFF_MS);
			clear_node = 0;
		}

		spin_lock(&dlm->spinlock);
		/* If we're not clearing the node bit then we intend
		 * to loop back around to try again. */
		if (clear_node)
			clear_bit(node, dlm->domain_map);
	}
	spin_unlock(&dlm->spinlock);
}

int dlm_joined(struct dlm_ctxt *dlm)
{
	int ret = 0;

	spin_lock(&dlm_domain_lock);

	if (dlm->dlm_state == DLM_CTXT_JOINED)
		ret = 1;

	spin_unlock(&dlm_domain_lock);

	return ret;
}