Commit 98211489 authored by Zach Brown's avatar Zach Brown Committed by Joel Becker
Browse files

[PATCH] OCFS2: The Second Oracle Cluster Filesystem



Node messaging via tcp. Used by the dlm and the file system for point
to point communication between nodes.
Signed-off-by: default avatarMark Fasheh <mark.fasheh@oracle.com>
Signed-off-by: default avatarKurt Hackel <kurt.hackel@oracle.com>
parent a7f6a5fb
/* -*- mode: c; c-basic-offset: 8; -*-
*
* vim: noexpandtab sw=8 ts=8 sts=0:
*
* Copyright (C) 2005 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.
*/
/* This quorum hack is only here until we transition to some more rational
* approach that is driven from userspace. Honest. No foolin'.
*
* Imagine two nodes lose network connectivity to each other but they're still
* up and operating in every other way. Presumably a network timeout indicates
* that a node is broken and should be recovered. They can't both recover each
* other and both carry on without serialising their access to the file system.
* They need to decide who is authoritative. Now extend that problem to
* arbitrary groups of nodes losing connectivity between each other.
*
* So we declare that a node which has given up on connecting to a majority
* of nodes who are still heartbeating will fence itself.
*
* There are huge opportunities for races here. After we give up on a node's
* connection we need to wait long enough to give heartbeat an opportunity
* to declare the node as truly dead. We also need to be careful with the
* race between when we see a node start heartbeating and when we connect
* to it.
*
* So nodes that are in this transtion put a hold on the quorum decision
* with a counter. As they fall out of this transition they drop the count
* and if they're the last, they fire off the decision.
*/
#include <linux/kernel.h>
#include <linux/slab.h>
#include <linux/workqueue.h>
#include "heartbeat.h"
#include "nodemanager.h"
#define MLOG_MASK_PREFIX ML_QUORUM
#include "masklog.h"
#include "quorum.h"
static struct o2quo_state {
spinlock_t qs_lock;
struct work_struct qs_work;
int qs_pending;
int qs_heartbeating;
unsigned long qs_hb_bm[BITS_TO_LONGS(O2NM_MAX_NODES)];
int qs_connected;
unsigned long qs_conn_bm[BITS_TO_LONGS(O2NM_MAX_NODES)];
int qs_holds;
unsigned long qs_hold_bm[BITS_TO_LONGS(O2NM_MAX_NODES)];
} o2quo_state;
/* this is horribly heavy-handed. It should instead flip the file
* system RO and call some userspace script. */
static void o2quo_fence_self(void)
{
/* panic spins with interrupts enabled. with preempt
* threads can still schedule, etc, etc */
o2hb_stop_all_regions();
panic("ocfs2 is very sorry to be fencing this system by panicing\n");
}
/* Indicate that a timeout occured on a hearbeat region write. The
* other nodes in the cluster may consider us dead at that time so we
* want to "fence" ourselves so that we don't scribble on the disk
* after they think they've recovered us. This can't solve all
* problems related to writeout after recovery but this hack can at
* least close some of those gaps. When we have real fencing, this can
* go away as our node would be fenced externally before other nodes
* begin recovery. */
void o2quo_disk_timeout(void)
{
o2quo_fence_self();
}
static void o2quo_make_decision(void *arg)
{
int quorum;
int lowest_hb, lowest_reachable = 0, fence = 0;
struct o2quo_state *qs = &o2quo_state;
spin_lock(&qs->qs_lock);
lowest_hb = find_first_bit(qs->qs_hb_bm, O2NM_MAX_NODES);
if (lowest_hb != O2NM_MAX_NODES)
lowest_reachable = test_bit(lowest_hb, qs->qs_conn_bm);
mlog(0, "heartbeating: %d, connected: %d, "
"lowest: %d (%sreachable)\n", qs->qs_heartbeating,
qs->qs_connected, lowest_hb, lowest_reachable ? "" : "un");
if (!test_bit(o2nm_this_node(), qs->qs_hb_bm) ||
qs->qs_heartbeating == 1)
goto out;
if (qs->qs_heartbeating & 1) {
/* the odd numbered cluster case is straight forward --
* if we can't talk to the majority we're hosed */
quorum = (qs->qs_heartbeating + 1)/2;
if (qs->qs_connected < quorum) {
mlog(ML_ERROR, "fencing this node because it is "
"only connected to %u nodes and %u is needed "
"to make a quorum out of %u heartbeating nodes\n",
qs->qs_connected, quorum,
qs->qs_heartbeating);
fence = 1;
}
} else {
/* the even numbered cluster adds the possibility of each half
* of the cluster being able to talk amongst themselves.. in
* that case we're hosed if we can't talk to the group that has
* the lowest numbered node */
quorum = qs->qs_heartbeating / 2;
if (qs->qs_connected < quorum) {
mlog(ML_ERROR, "fencing this node because it is "
"only connected to %u nodes and %u is needed "
"to make a quorum out of %u heartbeating nodes\n",
qs->qs_connected, quorum,
qs->qs_heartbeating);
fence = 1;
}
else if ((qs->qs_connected == quorum) &&
!lowest_reachable) {
mlog(ML_ERROR, "fencing this node because it is "
"connected to a half-quorum of %u out of %u "
"nodes which doesn't include the lowest active "
"node %u\n", quorum, qs->qs_heartbeating,
lowest_hb);
fence = 1;
}
}
out:
spin_unlock(&qs->qs_lock);
if (fence)
o2quo_fence_self();
}
static void o2quo_set_hold(struct o2quo_state *qs, u8 node)
{
assert_spin_locked(&qs->qs_lock);
if (!test_and_set_bit(node, qs->qs_hold_bm)) {
qs->qs_holds++;
mlog_bug_on_msg(qs->qs_holds == O2NM_MAX_NODES,
"node %u\n", node);
mlog(0, "node %u, %d total\n", node, qs->qs_holds);
}
}
static void o2quo_clear_hold(struct o2quo_state *qs, u8 node)
{
assert_spin_locked(&qs->qs_lock);
if (test_and_clear_bit(node, qs->qs_hold_bm)) {
mlog(0, "node %u, %d total\n", node, qs->qs_holds - 1);
if (--qs->qs_holds == 0) {
if (qs->qs_pending) {
qs->qs_pending = 0;
schedule_work(&qs->qs_work);
}
}
mlog_bug_on_msg(qs->qs_holds < 0, "node %u, holds %d\n",
node, qs->qs_holds);
}
}
/* as a node comes up we delay the quorum decision until we know the fate of
* the connection. the hold will be droped in conn_up or hb_down. it might be
* perpetuated by con_err until hb_down. if we already have a conn, we might
* be dropping a hold that conn_up got. */
void o2quo_hb_up(u8 node)
{
struct o2quo_state *qs = &o2quo_state;
spin_lock(&qs->qs_lock);
qs->qs_heartbeating++;
mlog_bug_on_msg(qs->qs_heartbeating == O2NM_MAX_NODES,
"node %u\n", node);
mlog_bug_on_msg(test_bit(node, qs->qs_hb_bm), "node %u\n", node);
set_bit(node, qs->qs_hb_bm);
mlog(0, "node %u, %d total\n", node, qs->qs_heartbeating);
if (!test_bit(node, qs->qs_conn_bm))
o2quo_set_hold(qs, node);
else
o2quo_clear_hold(qs, node);
spin_unlock(&qs->qs_lock);
}
/* hb going down releases any holds we might have had due to this node from
* conn_up, conn_err, or hb_up */
void o2quo_hb_down(u8 node)
{
struct o2quo_state *qs = &o2quo_state;
spin_lock(&qs->qs_lock);
qs->qs_heartbeating--;
mlog_bug_on_msg(qs->qs_heartbeating < 0,
"node %u, %d heartbeating\n",
node, qs->qs_heartbeating);
mlog_bug_on_msg(!test_bit(node, qs->qs_hb_bm), "node %u\n", node);
clear_bit(node, qs->qs_hb_bm);
mlog(0, "node %u, %d total\n", node, qs->qs_heartbeating);
o2quo_clear_hold(qs, node);
spin_unlock(&qs->qs_lock);
}
/* this tells us that we've decided that the node is still heartbeating
* even though we've lost it's conn. it must only be called after conn_err
* and indicates that we must now make a quorum decision in the future,
* though we might be doing so after waiting for holds to drain. Here
* we'll be dropping the hold from conn_err. */
void o2quo_hb_still_up(u8 node)
{
struct o2quo_state *qs = &o2quo_state;
spin_lock(&qs->qs_lock);
mlog(0, "node %u\n", node);
qs->qs_pending = 1;
o2quo_clear_hold(qs, node);
spin_unlock(&qs->qs_lock);
}
/* This is analagous to hb_up. as a node's connection comes up we delay the
* quorum decision until we see it heartbeating. the hold will be droped in
* hb_up or hb_down. it might be perpetuated by con_err until hb_down. if
* it's already heartbeating we we might be dropping a hold that conn_up got.
* */
void o2quo_conn_up(u8 node)
{
struct o2quo_state *qs = &o2quo_state;
spin_lock(&qs->qs_lock);
qs->qs_connected++;
mlog_bug_on_msg(qs->qs_connected == O2NM_MAX_NODES,
"node %u\n", node);
mlog_bug_on_msg(test_bit(node, qs->qs_conn_bm), "node %u\n", node);
set_bit(node, qs->qs_conn_bm);
mlog(0, "node %u, %d total\n", node, qs->qs_connected);
if (!test_bit(node, qs->qs_hb_bm))
o2quo_set_hold(qs, node);
else
o2quo_clear_hold(qs, node);
spin_unlock(&qs->qs_lock);
}
/* we've decided that we won't ever be connecting to the node again. if it's
* still heartbeating we grab a hold that will delay decisions until either the
* node stops heartbeating from hb_down or the caller decides that the node is
* still up and calls still_up */
void o2quo_conn_err(u8 node)
{
struct o2quo_state *qs = &o2quo_state;
spin_lock(&qs->qs_lock);
if (test_bit(node, qs->qs_conn_bm)) {
qs->qs_connected--;
mlog_bug_on_msg(qs->qs_connected < 0,
"node %u, connected %d\n",
node, qs->qs_connected);
clear_bit(node, qs->qs_conn_bm);
}
mlog(0, "node %u, %d total\n", node, qs->qs_connected);
if (test_bit(node, qs->qs_hb_bm))
o2quo_set_hold(qs, node);
spin_unlock(&qs->qs_lock);
}
void o2quo_init(void)
{
struct o2quo_state *qs = &o2quo_state;
spin_lock_init(&qs->qs_lock);
INIT_WORK(&qs->qs_work, o2quo_make_decision, NULL);
}
void o2quo_exit(void)
{
flush_scheduled_work();
}
/* -*- mode: c; c-basic-offset: 8; -*-
* vim: noexpandtab sw=8 ts=8 sts=0:
*
* Copyright (C) 2005 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.
*
*/
#ifndef O2CLUSTER_QUORUM_H
#define O2CLUSTER_QUORUM_H
void o2quo_init(void);
void o2quo_exit(void);
void o2quo_hb_up(u8 node);
void o2quo_hb_down(u8 node);
void o2quo_hb_still_up(u8 node);
void o2quo_conn_up(u8 node);
void o2quo_conn_err(u8 node);
void o2quo_disk_timeout(void);
#endif /* O2CLUSTER_QUORUM_H */
/* -*- mode: c; c-basic-offset: 8; -*-
* vim: noexpandtab sw=8 ts=8 sts=0:
*
* sys.c
*
* OCFS2 cluster sysfs interface
*
* Copyright (C) 2005 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,
* version 2 of the License.
*
* 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/kernel.h>
#include <linux/module.h>
#include <linux/kobject.h>
#include <linux/sysfs.h>
#include "ocfs2_nodemanager.h"
#include "masklog.h"
#include "sys.h"
struct o2cb_attribute {
struct attribute attr;
ssize_t (*show)(char *buf);
ssize_t (*store)(const char *buf, size_t count);
};
#define O2CB_ATTR(_name, _mode, _show, _store) \
struct o2cb_attribute o2cb_attr_##_name = __ATTR(_name, _mode, _show, _store)
#define to_o2cb_subsys(k) container_of(to_kset(k), struct subsystem, kset)
#define to_o2cb_attr(_attr) container_of(_attr, struct o2cb_attribute, attr)
static ssize_t o2cb_interface_revision_show(char *buf)
{
return snprintf(buf, PAGE_SIZE, "%u\n", O2NM_API_VERSION);
}
O2CB_ATTR(interface_revision, S_IFREG | S_IRUGO, o2cb_interface_revision_show, NULL);
static struct attribute *o2cb_attrs[] = {
&o2cb_attr_interface_revision.attr,
NULL,
};
static ssize_t
o2cb_show(struct kobject * kobj, struct attribute * attr, char * buffer);
static ssize_t
o2cb_store(struct kobject * kobj, struct attribute * attr,
const char * buffer, size_t count);
static struct sysfs_ops o2cb_sysfs_ops = {
.show = o2cb_show,
.store = o2cb_store,
};
static struct kobj_type o2cb_subsys_type = {
.default_attrs = o2cb_attrs,
.sysfs_ops = &o2cb_sysfs_ops,
};
/* gives us o2cb_subsys */
decl_subsys(o2cb, NULL, NULL);
static ssize_t
o2cb_show(struct kobject * kobj, struct attribute * attr, char * buffer)
{
struct o2cb_attribute *o2cb_attr = to_o2cb_attr(attr);
struct subsystem *sbs = to_o2cb_subsys(kobj);
BUG_ON(sbs != &o2cb_subsys);
if (o2cb_attr->show)
return o2cb_attr->show(buffer);
return -EIO;
}
static ssize_t
o2cb_store(struct kobject * kobj, struct attribute * attr,
const char * buffer, size_t count)
{
struct o2cb_attribute *o2cb_attr = to_o2cb_attr(attr);
struct subsystem *sbs = to_o2cb_subsys(kobj);
BUG_ON(sbs != &o2cb_subsys);
if (o2cb_attr->store)
return o2cb_attr->store(buffer, count);
return -EIO;
}
void o2cb_sys_shutdown(void)
{
mlog_sys_shutdown();
subsystem_unregister(&o2cb_subsys);
}
int o2cb_sys_init(void)
{
int ret;
o2cb_subsys.kset.kobj.ktype = &o2cb_subsys_type;
ret = subsystem_register(&o2cb_subsys);
if (ret)
return ret;
ret = mlog_sys_init(&o2cb_subsys);
if (ret)
subsystem_unregister(&o2cb_subsys);
return ret;
}
/* -*- mode: c; c-basic-offset: 8; -*-
* vim: noexpandtab sw=8 ts=8 sts=0:
*
* sys.h
*
* Function prototypes for o2cb sysfs interface
*
* Copyright (C) 2005 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,
* version 2 of the License.
*
* 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.
*
*/
#ifndef O2CLUSTER_SYS_H
#define O2CLUSTER_SYS_H
void o2cb_sys_shutdown(void);
int o2cb_sys_init(void);
#endif /* O2CLUSTER_SYS_H */
This diff is collapsed.
/* -*- mode: c; c-basic-offset: 8; -*-
* vim: noexpandtab sw=8 ts=8 sts=0:
*
* tcp.h
*
* Function prototypes
*
* 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.
*
*/
#ifndef O2CLUSTER_TCP_H
#define O2CLUSTER_TCP_H
#include <linux/socket.h>
#ifdef __KERNEL__
#include <net/sock.h>
#include <linux/tcp.h>
#else
#include <sys/socket.h>
#endif
#include <linux/inet.h>
#include <linux/in.h>
struct o2net_msg
{
__be16 magic;
__be16 data_len;
__be16 msg_type;
__be16 pad1;
__be32 sys_status;
__be32 status;
__be32 key;
__be32 msg_num;
__u8 buf[0];
};
typedef int (o2net_msg_handler_func)(struct o2net_msg *msg, u32 len, void *data);
#define O2NET_MAX_PAYLOAD_BYTES (4096 - sizeof(struct o2net_msg))
/* TODO: figure this out.... */
static inline int o2net_link_down(int err, struct socket *sock)
{
if (sock) {
if (sock->sk->sk_state != TCP_ESTABLISHED &&
sock->sk->sk_state != TCP_CLOSE_WAIT)
return 1;
}
if (err >= 0)
return 0;
switch (err) {
/* ????????????????????????? */
case -ERESTARTSYS:
case -EBADF:
/* When the server has died, an ICMP port unreachable
* message prompts ECONNREFUSED. */
case -ECONNREFUSED:
case -ENOTCONN:
case -ECONNRESET:
case -EPIPE:
return 1;
}
return 0;
}
enum {
O2NET_DRIVER_UNINITED,
O2NET_DRIVER_READY,
};
int o2net_init_tcp_sock(struct inode *inode);
int o2net_send_message(u32 msg_type, u32 key, void *data, u32 len,
u8 target_node, int *status);
int o2net_send_message_vec(u32 msg_type, u32 key, struct kvec *vec,
size_t veclen, u8 target_node, int *status);
int o2net_broadcast_message(u32 msg_type, u32 key, void *data, u32 len,
struct inode *group);
int o2net_register_handler(u32 msg_type, u32 key, u32 max_len,
o2net_msg_handler_func *func, void *data,
struct list_head *unreg_list);
void o2net_unregister_handler_list(struct list_head *list);
struct o2nm_node;
int o2net_register_hb_callbacks(void);
void o2net_unregister_hb_callbacks(void);
int o2net_start_listening(struct o2nm_node *node);
void o2net_stop_listening(struct o2nm_node *node);
void o2net_disconnect_node(struct o2nm_node *node);
int o2net_init(void);
void o2net_exit(void);
int o2net_proc_init(struct proc_dir_entry *parent);
void o2net_proc_exit(struct proc_dir_entry *parent);
#endif /* O2CLUSTER_TCP_H */