[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: Mark Fasheh <mark.fasheh@oracle.com>
Signed-off-by: Kurt Hackel <kurt.hackel@oracle.com>

7 files changed, 2624 insertions, 0 deletions

diff --git a/fs/ocfs2/cluster/quorum.c b/fs/ocfs2/cluster/quorum.c
new file mode 100644
index 000000000000..7bba98fbfc15
--- /dev/null
+++ b/fs/ocfs2/cluster/quorum.c
@@ -0,0 +1,315 @@
+/* -*- 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();
+}
diff --git a/fs/ocfs2/cluster/quorum.h b/fs/ocfs2/cluster/quorum.h
new file mode 100644
index 000000000000..6649cc6f67c9
--- /dev/null
+++ b/fs/ocfs2/cluster/quorum.h
@@ -0,0 +1,36 @@
+/* -*- 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 */
diff --git a/fs/ocfs2/cluster/sys.c b/fs/ocfs2/cluster/sys.c
new file mode 100644
index 000000000000..f1e99461bb0d
--- /dev/null
+++ b/fs/ocfs2/cluster/sys.c
@@ -0,0 +1,124 @@
+/* -*- 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;
+}
diff --git a/fs/ocfs2/cluster/sys.h b/fs/ocfs2/cluster/sys.h
new file mode 100644
index 000000000000..d66b8ab0045e
--- /dev/null
+++ b/fs/ocfs2/cluster/sys.h
@@ -0,0 +1,33 @@
+/* -*- 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 */
diff --git a/fs/ocfs2/cluster/tcp.c b/fs/ocfs2/cluster/tcp.c
new file mode 100644
index 000000000000..35d92c01a972
--- /dev/null
+++ b/fs/ocfs2/cluster/tcp.c
@@ -0,0 +1,1829 @@
+/* -*- mode: c; c-basic-offset: 8; -*-
+ *
+ * vim: noexpandtab sw=8 ts=8 sts=0:
+ *
+ * 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.
+ *
+ * ----
+ *
+ * Callers for this were originally written against a very simple synchronus
+ * API.  This implementation reflects those simple callers.  Some day I'm sure
+ * we'll need to move to a more robust posting/callback mechanism.
+ *
+ * Transmit calls pass in kernel virtual addresses and block copying this into
+ * the socket's tx buffers via a usual blocking sendmsg.  They'll block waiting
+ * for a failed socket to timeout.  TX callers can also pass in a poniter to an
+ * 'int' which gets filled with an errno off the wire in response to the
+ * message they send.
+ *
+ * Handlers for unsolicited messages are registered.  Each socket has a page
+ * that incoming data is copied into.  First the header, then the data.
+ * Handlers are called from only one thread with a reference to this per-socket
+ * page.  This page is destroyed after the handler call, so it can't be
+ * referenced beyond the call.  Handlers may block but are discouraged from
+ * doing so.
+ *
+ * Any framing errors (bad magic, large payload lengths) close a connection.
+ *
+ * Our sock_container holds the state we associate with a socket.  It's current
+ * framing state is held there as well as the refcounting we do around when it
+ * is safe to tear down the socket.  The socket is only finally torn down from
+ * the container when the container loses all of its references -- so as long
+ * as you hold a ref on the container you can trust that the socket is valid
+ * for use with kernel socket APIs.
+ *
+ * Connections are initiated between a pair of nodes when the node with the
+ * higher node number gets a heartbeat callback which indicates that the lower
+ * numbered node has started heartbeating.  The lower numbered node is passive
+ * and only accepts the connection if the higher numbered node is heartbeating.
+ */
+
+#include <linux/kernel.h>
+#include <linux/jiffies.h>
+#include <linux/slab.h>
+#include <linux/idr.h>
+#include <linux/kref.h>
+#include <net/tcp.h>
+
+#include <asm/uaccess.h>
+
+#include "heartbeat.h"
+#include "tcp.h"
+#include "nodemanager.h"
+#define MLOG_MASK_PREFIX ML_TCP
+#include "masklog.h"
+#include "quorum.h"
+
+#include "tcp_internal.h"
+
+/* 
+ * The linux network stack isn't sparse endian clean.. It has macros like
+ * ntohs() which perform the endian checks and structs like sockaddr_in
+ * which aren't annotated.  So __force is found here to get the build
+ * clean.  When they emerge from the dark ages and annotate the code
+ * we can remove these.
+ */
+
+#define SC_NODEF_FMT "node %s (num %u) at %u.%u.%u.%u:%u"
+#define SC_NODEF_ARGS(sc) sc->sc_node->nd_name, sc->sc_node->nd_num,	\
+			  NIPQUAD(sc->sc_node->nd_ipv4_address),	\
+			  ntohs(sc->sc_node->nd_ipv4_port)
+
+/*
+ * In the following two log macros, the whitespace after the ',' just
+ * before ##args is intentional. Otherwise, gcc 2.95 will eat the
+ * previous token if args expands to nothing.
+ */
+#define msglog(hdr, fmt, args...) do {					\
+	typeof(hdr) __hdr = (hdr);					\
+	mlog(ML_MSG, "[mag %u len %u typ %u stat %d sys_stat %d "	\
+	     "key %08x num %u] " fmt,					\
+	     be16_to_cpu(__hdr->magic), be16_to_cpu(__hdr->data_len), 	\
+	     be16_to_cpu(__hdr->msg_type), be32_to_cpu(__hdr->status),	\
+	     be32_to_cpu(__hdr->sys_status), be32_to_cpu(__hdr->key),	\
+	     be32_to_cpu(__hdr->msg_num) ,  ##args);			\
+} while (0)
+
+#define sclog(sc, fmt, args...) do {					\
+	typeof(sc) __sc = (sc);						\
+	mlog(ML_SOCKET, "[sc %p refs %d sock %p node %u page %p "	\
+	     "pg_off %zu] " fmt, __sc,					\
+	     atomic_read(&__sc->sc_kref.refcount), __sc->sc_sock,	\
+	    __sc->sc_node->nd_num, __sc->sc_page, __sc->sc_page_off ,	\
+	    ##args);							\
+} while (0)
+
+static rwlock_t o2net_handler_lock = RW_LOCK_UNLOCKED;
+static struct rb_root o2net_handler_tree = RB_ROOT;
+
+static struct o2net_node o2net_nodes[O2NM_MAX_NODES];
+
+/* XXX someday we'll need better accounting */
+static struct socket *o2net_listen_sock = NULL;
+
+/*
+ * listen work is only queued by the listening socket callbacks on the
+ * o2net_wq.  teardown detaches the callbacks before destroying the workqueue.
+ * quorum work is queued as sock containers are shutdown.. stop_listening
+ * tears down all the node's sock containers, preventing future shutdowns
+ * and queued quroum work, before canceling delayed quorum work and
+ * destroying the work queue.
+ */
+static struct workqueue_struct *o2net_wq;
+static struct work_struct o2net_listen_work;
+
+static struct o2hb_callback_func o2net_hb_up, o2net_hb_down;
+#define O2NET_HB_PRI 0x1
+
+static struct o2net_handshake *o2net_hand;
+static struct o2net_msg *o2net_keep_req, *o2net_keep_resp;
+
+static int o2net_sys_err_translations[O2NET_ERR_MAX] =
+		{[O2NET_ERR_NONE]	= 0,
+		 [O2NET_ERR_NO_HNDLR]	= -ENOPROTOOPT,
+		 [O2NET_ERR_OVERFLOW]	= -EOVERFLOW,
+		 [O2NET_ERR_DIED]	= -EHOSTDOWN,};
+
+/* can't quite avoid *all* internal declarations :/ */
+static void o2net_sc_connect_completed(void *arg);
+static void o2net_rx_until_empty(void *arg);
+static void o2net_shutdown_sc(void *arg);
+static void o2net_listen_data_ready(struct sock *sk, int bytes);
+static void o2net_sc_send_keep_req(void *arg);
+static void o2net_idle_timer(unsigned long data);
+static void o2net_sc_postpone_idle(struct o2net_sock_container *sc);
+
+static inline int o2net_sys_err_to_errno(enum o2net_system_error err)
+{
+	int trans;
+	BUG_ON(err >= O2NET_ERR_MAX);
+	trans = o2net_sys_err_translations[err];
+
+	/* Just in case we mess up the translation table above */
+	BUG_ON(err != O2NET_ERR_NONE && trans == 0);
+	return trans;
+}
+
+static struct o2net_node * o2net_nn_from_num(u8 node_num)
+{
+	BUG_ON(node_num >= ARRAY_SIZE(o2net_nodes));
+	return &o2net_nodes[node_num];
+}
+
+static u8 o2net_num_from_nn(struct o2net_node *nn)
+{
+	BUG_ON(nn == NULL);
+	return nn - o2net_nodes;
+}
+
+/* ------------------------------------------------------------ */
+
+static int o2net_prep_nsw(struct o2net_node *nn, struct o2net_status_wait *nsw)
+{
+	int ret = 0;
+
+	do {
+		if (!idr_pre_get(&nn->nn_status_idr, GFP_ATOMIC)) {
+			ret = -EAGAIN;
+			break;
+		}
+		spin_lock(&nn->nn_lock);
+		ret = idr_get_new(&nn->nn_status_idr, nsw, &nsw->ns_id);
+		if (ret == 0)
+			list_add_tail(&nsw->ns_node_item,
+				      &nn->nn_status_list);
+		spin_unlock(&nn->nn_lock);
+	} while (ret == -EAGAIN);
+
+	if (ret == 0)  {
+		init_waitqueue_head(&nsw->ns_wq);
+		nsw->ns_sys_status = O2NET_ERR_NONE;
+		nsw->ns_status = 0;
+	}
+
+	return ret;
+}
+
+static void o2net_complete_nsw_locked(struct o2net_node *nn,
+				      struct o2net_status_wait *nsw,
+				      enum o2net_system_error sys_status,
+				      s32 status)
+{
+	assert_spin_locked(&nn->nn_lock);
+
+	if (!list_empty(&nsw->ns_node_item)) {
+		list_del_init(&nsw->ns_node_item);
+		nsw->ns_sys_status = sys_status;
+		nsw->ns_status = status;
+		idr_remove(&nn->nn_status_idr, nsw->ns_id);
+		wake_up(&nsw->ns_wq);
+	}
+}
+
+static void o2net_complete_nsw(struct o2net_node *nn,
+			       struct o2net_status_wait *nsw,
+			       u64 id, enum o2net_system_error sys_status,
+			       s32 status)
+{
+	spin_lock(&nn->nn_lock);
+	if (nsw == NULL) {
+		if (id > INT_MAX)
+			goto out;
+
+		nsw = idr_find(&nn->nn_status_idr, id);
+		if (nsw == NULL)
+			goto out;
+	}
+
+	o2net_complete_nsw_locked(nn, nsw, sys_status, status);
+
+out:
+	spin_unlock(&nn->nn_lock);
+	return;
+}
+
+static void o2net_complete_nodes_nsw(struct o2net_node *nn)
+{
+	struct list_head *iter, *tmp;
+	unsigned int num_kills = 0;
+	struct o2net_status_wait *nsw;
+
+	assert_spin_locked(&nn->nn_lock);
+
+	list_for_each_safe(iter, tmp, &nn->nn_status_list) {
+		nsw = list_entry(iter, struct o2net_status_wait, ns_node_item);
+		o2net_complete_nsw_locked(nn, nsw, O2NET_ERR_DIED, 0);
+		num_kills++;
+	}
+
+	mlog(0, "completed %d messages for node %u\n", num_kills,
+	     o2net_num_from_nn(nn));
+}
+
+static int o2net_nsw_completed(struct o2net_node *nn,
+			       struct o2net_status_wait *nsw)
+{
+	int completed;
+	spin_lock(&nn->nn_lock);
+	completed = list_empty(&nsw->ns_node_item);
+	spin_unlock(&nn->nn_lock);
+	return completed;
+}
+
+/* ------------------------------------------------------------ */
+
+static void sc_kref_release(struct kref *kref)
+{
+	struct o2net_sock_container *sc = container_of(kref,
+					struct o2net_sock_container, sc_kref);
+	sclog(sc, "releasing\n");
+
+	if (sc->sc_sock) {
+		sock_release(sc->sc_sock);
+		sc->sc_sock = NULL;
+	}
+
+	o2nm_node_put(sc->sc_node);
+	sc->sc_node = NULL;
+
+	kfree(sc);
+}
+
+static void sc_put(struct o2net_sock_container *sc)
+{
+	sclog(sc, "put\n");
+	kref_put(&sc->sc_kref, sc_kref_release);
+}
+static void sc_get(struct o2net_sock_container *sc)
+{
+	sclog(sc, "get\n");
+	kref_get(&sc->sc_kref);
+}
+static struct o2net_sock_container *sc_alloc(struct o2nm_node *node)
+{
+	struct o2net_sock_container *sc, *ret = NULL;
+	struct page *page = NULL;
+
+	page = alloc_page(GFP_NOFS);
+	sc = kcalloc(1, sizeof(*sc), GFP_NOFS);
+	if (sc == NULL || page == NULL)
+		goto out;
+
+	kref_init(&sc->sc_kref);
+	o2nm_node_get(node);
+	sc->sc_node = node;
+
+	INIT_WORK(&sc->sc_connect_work, o2net_sc_connect_completed, sc);
+	INIT_WORK(&sc->sc_rx_work, o2net_rx_until_empty, sc);
+	INIT_WORK(&sc->sc_shutdown_work, o2net_shutdown_sc, sc);
+	INIT_WORK(&sc->sc_keepalive_work, o2net_sc_send_keep_req, sc);
+
+	init_timer(&sc->sc_idle_timeout);
+	sc->sc_idle_timeout.function = o2net_idle_timer;
+	sc->sc_idle_timeout.data = (unsigned long)sc;
+
+	sclog(sc, "alloced\n");
+
+	ret = sc;
+	sc->sc_page = page;
+	sc = NULL;
+	page = NULL;
+
+out:
+	if (page)
+		__free_page(page);
+	kfree(sc);
+
+	return ret;
+}
+
+/* ------------------------------------------------------------ */
+
+static void o2net_sc_queue_work(struct o2net_sock_container *sc,
+				struct work_struct *work)
+{
+	sc_get(sc);
+	if (!queue_work(o2net_wq, work))
+		sc_put(sc);
+}
+static void o2net_sc_queue_delayed_work(struct o2net_sock_container *sc,
+					struct work_struct *work,
+					int delay)
+{
+	sc_get(sc);
+	if (!queue_delayed_work(o2net_wq, work, delay))
+		sc_put(sc);
+}
+static void o2net_sc_cancel_delayed_work(struct o2net_sock_container *sc,
+					 struct work_struct *work)
+{
+	if (cancel_delayed_work(work))
+		sc_put(sc);
+}
+
+static void o2net_set_nn_state(struct o2net_node *nn,
+			       struct o2net_sock_container *sc,
+			       unsigned valid, int err)
+{
+	int was_valid = nn->nn_sc_valid;
+	int was_err = nn->nn_persistent_error;
+	struct o2net_sock_container *old_sc = nn->nn_sc;
+
+	assert_spin_locked(&nn->nn_lock);
+
+	/* the node num comparison and single connect/accept path should stop
+	 * an non-null sc from being overwritten with another */
+	BUG_ON(sc && nn->nn_sc && nn->nn_sc != sc);
+	mlog_bug_on_msg(err && valid, "err %d valid %u\n", err, valid);
+	mlog_bug_on_msg(valid && !sc, "valid %u sc %p\n", valid, sc);
+
+	/* we won't reconnect after our valid conn goes away for
+	 * this hb iteration.. here so it shows up in the logs */
+	if (was_valid && !valid && err == 0)
+		err = -ENOTCONN;
+
+	mlog(ML_CONN, "node %u sc: %p -> %p, valid %u -> %u, err %d -> %d\n",
+	     o2net_num_from_nn(nn), nn->nn_sc, sc, nn->nn_sc_valid, valid,
+	     nn->nn_persistent_error, err);
+
+	nn->nn_sc = sc;
+	nn->nn_sc_valid = valid ? 1 : 0;
+	nn->nn_persistent_error = err;
+
+	/* mirrors o2net_tx_can_proceed() */
+	if (nn->nn_persistent_error || nn->nn_sc_valid)
+		wake_up(&nn->nn_sc_wq);
+
+	if (!was_err && nn->nn_persistent_error) {
+		o2quo_conn_err(o2net_num_from_nn(nn));
+		queue_delayed_work(o2net_wq, &nn->nn_still_up,
+				   msecs_to_jiffies(O2NET_QUORUM_DELAY_MS));
+	}
+
+	if (was_valid && !valid) {
+		mlog(ML_NOTICE, "no longer connected to " SC_NODEF_FMT "\n",
+		     SC_NODEF_ARGS(old_sc));
+		o2net_complete_nodes_nsw(nn);
+	}
+
+	if (!was_valid && valid) {
+		o2quo_conn_up(o2net_num_from_nn(nn));
+		/* this is a bit of a hack.  we only try reconnecting
+		 * when heartbeating starts until we get a connection.
+		 * if that connection then dies we don't try reconnecting.
+		 * the only way to start connecting again is to down
+		 * heartbeat and bring it back up. */
+		cancel_delayed_work(&nn->nn_connect_expired);
+		mlog(ML_NOTICE, "%s " SC_NODEF_FMT "\n", 
+		     o2nm_this_node() > sc->sc_node->nd_num ?
+		     	"connected to" : "accepted connection from",
+		     SC_NODEF_ARGS(sc));
+	}
+
+	/* trigger the connecting worker func as long as we're not valid,
+	 * it will back off if it shouldn't connect.  This can be called
+	 * from node config teardown and so needs to be careful about
+	 * the work queue actually being up. */
+	if (!valid && o2net_wq) {
+		unsigned long delay;
+		/* delay if we're withing a RECONNECT_DELAY of the
+		 * last attempt */
+		delay = (nn->nn_last_connect_attempt +
+			 msecs_to_jiffies(O2NET_RECONNECT_DELAY_MS))
+			- jiffies;
+		if (delay > msecs_to_jiffies(O2NET_RECONNECT_DELAY_MS))
+			delay = 0;
+		mlog(ML_CONN, "queueing conn attempt in %lu jiffies\n", delay);
+		queue_delayed_work(o2net_wq, &nn->nn_connect_work, delay);
+	}
+
+	/* keep track of the nn's sc ref for the caller */
+	if ((old_sc == NULL) && sc)
+		sc_get(sc);
+	if (old_sc && (old_sc != sc)) {
+		o2net_sc_queue_work(old_sc, &old_sc->sc_shutdown_work);
+		sc_put(old_sc);
+	}
+}
+
+/* see o2net_register_callbacks() */
+static void o2net_data_ready(struct sock *sk, int bytes)
+{
+	void (*ready)(struct sock *sk, int bytes);
+
+	read_lock(&sk->sk_callback_lock);
+	if (sk->sk_user_data) {
+		struct o2net_sock_container *sc = sk->sk_user_data;
+		sclog(sc, "data_ready hit\n");
+		do_gettimeofday(&sc->sc_tv_data_ready);
+		o2net_sc_queue_work(sc, &sc->sc_rx_work);
+		ready = sc->sc_data_ready;
+	} else {
+		ready = sk->sk_data_ready;
+	}
+	read_unlock(&sk->sk_callback_lock);
+
+	ready(sk, bytes);
+}
+
+/* see o2net_register_callbacks() */
+static void o2net_state_change(struct sock *sk)
+{
+	void (*state_change)(struct sock *sk);
+	struct o2net_sock_container *sc;
+
+	read_lock(&sk->sk_callback_lock);
+	sc = sk->sk_user_data;
+	if (sc == NULL) {
+		state_change = sk->sk_state_change;
+		goto out;
+	}
+
+	sclog(sc, "state_change to %d\n", sk->sk_state);
+
+	state_change = sc->sc_state_change;
+
+	switch(sk->sk_state) {
+		/* ignore connecting sockets as they make progress */
+		case TCP_SYN_SENT:
+		case TCP_SYN_RECV:
+			break;
+		case TCP_ESTABLISHED:
+			o2net_sc_queue_work(sc, &sc->sc_connect_work);
+			break;
+		default:
+			o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
+			break;
+	}
+out:
+	read_unlock(&sk->sk_callback_lock);
+	state_change(sk);
+}
+
+/*
+ * we register callbacks so we can queue work on events before calling
+ * the original callbacks.  our callbacks our careful to test user_data
+ * to discover when they've reaced with o2net_unregister_callbacks().
+ */
+static void o2net_register_callbacks(struct sock *sk,
+				     struct o2net_so