recover.c

linux 内核源代码

/******************************************************************************
*******************************************************************************
**
**  Copyright (C) Sistina Software, Inc.  1997-2003  All rights reserved.
**  Copyright (C) 2004-2005 Red Hat, Inc.  All rights reserved.
**
**  This copyrighted material is made available to anyone wishing to use,
**  modify, copy, or redistribute it subject to the terms and conditions
**  of the GNU General Public License v.2.
**
*******************************************************************************
******************************************************************************/

#include "dlm_internal.h"
#include "lockspace.h"
#include "dir.h"
#include "config.h"
#include "ast.h"
#include "memory.h"
#include "rcom.h"
#include "lock.h"
#include "lowcomms.h"
#include "member.h"
#include "recover.h"


/*
 * Recovery waiting routines: these functions wait for a particular reply from
 * a remote node, or for the remote node to report a certain status.  They need
 * to abort if the lockspace is stopped indicating a node has failed (perhaps
 * the one being waited for).
 */

/*
 * Wait until given function returns non-zero or lockspace is stopped
 * (LS_RECOVERY_STOP set due to failure of a node in ls_nodes).  When another
 * function thinks it could have completed the waited-on task, they should wake
 * up ls_wait_general to get an immediate response rather than waiting for the
 * timer to detect the result.  A timer wakes us up periodically while waiting
 * to see if we should abort due to a node failure.  This should only be called
 * by the dlm_recoverd thread.
 */

static void dlm_wait_timer_fn(unsigned long data)
{
	struct dlm_ls *ls = (struct dlm_ls *) data;
	mod_timer(&ls->ls_timer, jiffies + (dlm_config.ci_recover_timer * HZ));
	wake_up(&ls->ls_wait_general);
}

int dlm_wait_function(struct dlm_ls *ls, int (*testfn) (struct dlm_ls *ls))
{
	int error = 0;

	init_timer(&ls->ls_timer);
	ls->ls_timer.function = dlm_wait_timer_fn;
	ls->ls_timer.data = (long) ls;
	ls->ls_timer.expires = jiffies + (dlm_config.ci_recover_timer * HZ);
	add_timer(&ls->ls_timer);

	wait_event(ls->ls_wait_general, testfn(ls) || dlm_recovery_stopped(ls));
	del_timer_sync(&ls->ls_timer);

	if (dlm_recovery_stopped(ls)) {
		log_debug(ls, "dlm_wait_function aborted");
		error = -EINTR;
	}
	return error;
}

/*
 * An efficient way for all nodes to wait for all others to have a certain
 * status.  The node with the lowest nodeid polls all the others for their
 * status (wait_status_all) and all the others poll the node with the low id
 * for its accumulated result (wait_status_low).  When all nodes have set
 * status flag X, then status flag X_ALL will be set on the low nodeid.
 */

uint32_t dlm_recover_status(struct dlm_ls *ls)
{
	uint32_t status;
	spin_lock(&ls->ls_recover_lock);
	status = ls->ls_recover_status;
	spin_unlock(&ls->ls_recover_lock);
	return status;
}

void dlm_set_recover_status(struct dlm_ls *ls, uint32_t status)
{
	spin_lock(&ls->ls_recover_lock);
	ls->ls_recover_status |= status;
	spin_unlock(&ls->ls_recover_lock);
}

static int wait_status_all(struct dlm_ls *ls, uint32_t wait_status)
{
	struct dlm_rcom *rc = (struct dlm_rcom *) ls->ls_recover_buf;
	struct dlm_member *memb;
	int error = 0, delay;

	list_for_each_entry(memb, &ls->ls_nodes, list) {
		delay = 0;
		for (;;) {
			if (dlm_recovery_stopped(ls)) {
				error = -EINTR;
				goto out;
			}

			error = dlm_rcom_status(ls, memb->nodeid);
			if (error)
				goto out;

			if (rc->rc_result & wait_status)
				break;
			if (delay < 1000)
				delay += 20;
			msleep(delay);
		}
	}
 out:
	return error;
}

static int wait_status_low(struct dlm_ls *ls, uint32_t wait_status)
{
	struct dlm_rcom *rc = (struct dlm_rcom *) ls->ls_recover_buf;
	int error = 0, delay = 0, nodeid = ls->ls_low_nodeid;

	for (;;) {
		if (dlm_recovery_stopped(ls)) {
			error = -EINTR;
			goto out;
		}

		error = dlm_rcom_status(ls, nodeid);
		if (error)
			break;

		if (rc->rc_result & wait_status)
			break;
		if (delay < 1000)
			delay += 20;
		msleep(delay);
	}
 out:
	return error;
}

static int wait_status(struct dlm_ls *ls, uint32_t status)
{
	uint32_t status_all = status << 1;
	int error;

	if (ls->ls_low_nodeid == dlm_our_nodeid()) {
		error = wait_status_all(ls, status);
		if (!error)
			dlm_set_recover_status(ls, status_all);
	} else
		error = wait_status_low(ls, status_all);

	return error;
}

int dlm_recover_members_wait(struct dlm_ls *ls)
{
	return wait_status(ls, DLM_RS_NODES);
}

int dlm_recover_directory_wait(struct dlm_ls *ls)
{
	return wait_status(ls, DLM_RS_DIR);
}

int dlm_recover_locks_wait(struct dlm_ls *ls)
{
	return wait_status(ls, DLM_RS_LOCKS);
}

int dlm_recover_done_wait(struct dlm_ls *ls)
{
	return wait_status(ls, DLM_RS_DONE);
}

/*
 * The recover_list contains all the rsb's for which we've requested the new
 * master nodeid.  As replies are returned from the resource directories the
 * rsb's are removed from the list.  When the list is empty we're done.
 *
 * The recover_list is later similarly used for all rsb's for which we've sent
 * new lkb's and need to receive new corresponding lkid's.
 *
 * We use the address of the rsb struct as a simple local identifier for the
 * rsb so we can match an rcom reply with the rsb it was sent for.
 */

static int recover_list_empty(struct dlm_ls *ls)
{
	int empty;

	spin_lock(&ls->ls_recover_list_lock);
	empty = list_empty(&ls->ls_recover_list);
	spin_unlock(&ls->ls_recover_list_lock);

	return empty;
}

static void recover_list_add(struct dlm_rsb *r)
{
	struct dlm_ls *ls = r->res_ls;

	spin_lock(&ls->ls_recover_list_lock);
	if (list_empty(&r->res_recover_list)) {
		list_add_tail(&r->res_recover_list, &ls->ls_recover_list);
		ls->ls_recover_list_count++;
		dlm_hold_rsb(r);
	}
	spin_unlock(&ls->ls_recover_list_lock);
}

static void recover_list_del(struct dlm_rsb *r)
{
	struct dlm_ls *ls = r->res_ls;

	spin_lock(&ls->ls_recover_list_lock);
	list_del_init(&r->res_recover_list);
	ls->ls_recover_list_count--;
	spin_unlock(&ls->ls_recover_list_lock);

	dlm_put_rsb(r);
}

static struct dlm_rsb *recover_list_find(struct dlm_ls *ls, uint64_t id)
{
	struct dlm_rsb *r = NULL;

	spin_lock(&ls->ls_recover_list_lock);

	list_for_each_entry(r, &ls->ls_recover_list, res_recover_list) {
		if (id == (unsigned long) r)
			goto out;
	}
	r = NULL;
 out:
	spin_unlock(&ls->ls_recover_list_lock);
	return r;
}

static void recover_list_clear(struct dlm_ls *ls)
{
	struct dlm_rsb *r, *s;

	spin_lock(&ls->ls_recover_list_lock);
	list_for_each_entry_safe(r, s, &ls->ls_recover_list, res_recover_list) {
		list_del_init(&r->res_recover_list);
		r->res_recover_locks_count = 0;
		dlm_put_rsb(r);
		ls->ls_recover_list_count--;
	}

	if (ls->ls_recover_list_count != 0) {
		log_error(ls, "warning: recover_list_count %d",
			  ls->ls_recover_list_count);
		ls->ls_recover_list_count = 0;
	}
	spin_unlock(&ls->ls_recover_list_lock);
}


/* Master recovery: find new master node for rsb's that were
   mastered on nodes that have been removed.

   dlm_recover_masters
   recover_master
   dlm_send_rcom_lookup            ->  receive_rcom_lookup
                                       dlm_dir_lookup
   receive_rcom_lookup_reply       <-
   dlm_recover_master_reply
   set_new_master
   set_master_lkbs
   set_lock_master
*/

/*
 * Set the lock master for all LKBs in a lock queue
 * If we are the new master of the rsb, we may have received new
 * MSTCPY locks from other nodes already which we need to ignore
 * when setting the new nodeid.
 */

static void set_lock_master(struct list_head *queue, int nodeid)
{
	struct dlm_lkb *lkb;

	list_for_each_entry(lkb, queue, lkb_statequeue)
		if (!(lkb->lkb_flags & DLM_IFL_MSTCPY))
			lkb->lkb_nodeid = nodeid;
}

static void set_master_lkbs(struct dlm_rsb *r)
{
	set_lock_master(&r->res_grantqueue, r->res_nodeid);
	set_lock_master(&r->res_convertqueue, r->res_nodeid);
	set_lock_master(&r->res_waitqueue, r->res_nodeid);
}

/*
 * Propogate the new master nodeid to locks
 * The NEW_MASTER flag tells dlm_recover_locks() which rsb's to consider.
 * The NEW_MASTER2 flag tells recover_lvb() and set_locks_purged() which
 * rsb's to consider.
 */

static void set_new_master(struct dlm_rsb *r, int nodeid)
{
	lock_rsb(r);
	r->res_nodeid = nodeid;
	set_master_lkbs(r);
	rsb_set_flag(r, RSB_NEW_MASTER);
	rsb_set_flag(r, RSB_NEW_MASTER2);
	unlock_rsb(r);
}

/*
 * We do async lookups on rsb's that need new masters.  The rsb's
 * waiting for a lookup reply are kept on the recover_list.
 */

static int recover_master(struct dlm_rsb *r)
{
	struct dlm_ls *ls = r->res_ls;
	int error, dir_nodeid, ret_nodeid, our_nodeid = dlm_our_nodeid();

	dir_nodeid = dlm_dir_nodeid(r);

	if (dir_nodeid == our_nodeid) {
		error = dlm_dir_lookup(ls, our_nodeid, r->res_name,
				       r->res_length, &ret_nodeid);
		if (error)
			log_error(ls, "recover dir lookup error %d", error);

		if (ret_nodeid == our_nodeid)
			ret_nodeid = 0;
		set_new_master(r, ret_nodeid);
	} else {
		recover_list_add(r);
		error = dlm_send_rcom_lookup(r, dir_nodeid);
	}

	return error;
}

/*
 * When not using a directory, most resource names will hash to a new static
 * master nodeid and the resource will need to be remastered.
 */

static int recover_master_static(struct dlm_rsb *r)
{
	int master = dlm_dir_nodeid(r);

	if (master == dlm_our_nodeid())
		master = 0;

	if (r->res_nodeid != master) {
		if (is_master(r))
			dlm_purge_mstcpy_locks(r);
		set_new_master(r, master);
		return 1;
	}
	return 0;
}

/*
 * Go through local root resources and for each rsb which has a master which
 * has departed, get the new master nodeid from the directory.  The dir will
 * assign mastery to the first node to look up the new master.  That means
 * we'll discover in this lookup if we're the new master of any rsb's.
 *
 * We fire off all the dir lookup requests individually and asynchronously to
 * the correct dir node.
 */

int dlm_recover_masters(struct dlm_ls *ls)
{
	struct dlm_rsb *r;