#!/bin/sh

# statd must be configured to use statd_callout, CTDB's binary
# counterpart to this script, as its availability call-out.
#
# Modern NFS utils versions use /etc/nfs.conf:
#
#   [statd]
#     name = mycluster
#     ha-callout = /usr/local/libexec/ctdb/statd_callout
#
# Older Linux versions may use something like the following...
#
# /etc/sysconfig/nfs (Red Hat) or /etc/default/nfs-common (Debian):
#   STATD_HOSTNAME="mycluster -H /usr/local/libexec/ctdb/statd_callout"
#
# If using Linux kernel NFS then the following should also be set in
# /etc/nfs.conf:
#
#   [sm-notify]
#      lift-grace = n
#
# See sm-notify(8) for details.  This doesn't matter when using
# NFS-Ganesha because sm-notify's attempt to lift grace will fail
# silently if /proc/fs/lockd/nlm_end_grace is not found.
#

if [ -z "$CTDB_BASE" ]; then
	export CTDB_BASE="/usr/local/etc/ctdb"
fi

. "${CTDB_BASE}/functions"

# Overwrite this so we get some logging
die()
{
	script_log "statd_callout_helper" "$@"
	exit 1
}

############################################################

ctdb_setup_state_dir "service" "nfs"

find_statd_sm_dir()
{
	if [ -n "$CTDB_TEST_MODE" ]; then
		_f="${CTDB_TEST_TMP_DIR}/sm"
		mkdir -p "$_f" "${_f}.bak"
		echo "$_f"
		return
	fi

	for _sm_dir in /var/lib/nfs/statd/sm /var/lib/nfs/sm; do
		if [ -d "$_sm_dir" ]; then
			echo "$_sm_dir"
			break
		fi
	done
}

# Ensure the state directory exists and can be written when called as
# a non-root user.  Assume the user to run as is the owner of the
# system statd sm directory, since both rpc.statd and sm-notify run as
# this directory's owner, so it can read and modify the directory.
create_add_del_client_dir()
{
	_dir="$1"

	if [ ! -d "$_dir" ]; then
		mkdir -p "$_dir" || die "Failed to create directory \"${_dir}\""
		ref=$(find_statd_sm_dir)
		[ -n "$ref" ] || die "Failed to find statd sm directory"
		chown --reference="$ref" "$_dir"
	fi
}

# script_state_dir set by ctdb_setup_state_dir()
# shellcheck disable=SC2154
statd_callout_state_dir="${script_state_dir}/statd_callout"

statd_callout_db="ctdb.tdb"
statd_callout_queue_dir="${statd_callout_state_dir}/queue"

############################################################

# Read pairs of:
#   server-IP client-IP
# from stdin and send associated SM_NOTIFY packets.
send_notifies()
{
	# State must monotonically increase, across the entire
	# cluster.  Use seconds since epoch and assume the time is in
	# sync across nodes.  Even numbers mean service is shut down,
	# odd numbers mean service is up.  However, sm-notify always
	# reads the state and converts it to odd (if necessary, by
	# adding 1 when it is even) because it only sends "up"
	# notifications.  Note that there is a 2038 issue here but we
	# will get to that later.
	_state=$(date '+%s')

	_helper="${CTDB_HELPER_BINDIR}/ctdb_smnotify_helper"

	_notify_dir="${statd_callout_state_dir}/sm-notify"
	mkdir -p "$_notify_dir"

	while read -r _sip _cip; do
		# Create a directory per server IP containing a file
		# for each client IP
		mkdir -p \
			"${_notify_dir}/${_sip}/sm" \
			"${_notify_dir}/${_sip}/sm.bak"

		_out="${_notify_dir}/${_sip}/sm/${_cip}"
		"$_helper" "monitor" "$_cip" "$_sip" >"$_out"
	done

	# Send notifications for server startup
	_ref=$(find_statd_sm_dir)
	for _sip_dir in "$_notify_dir"/*; do
		if [ "$_sip_dir" = "${_notify_dir}/*" ]; then
			break
		fi

		_sip="${_sip_dir##*/}" # basename

		# Write the state as a host order 32-bit integer.  See
		# note at top of function about state.
		_out="${_sip_dir}/state"
		"$_helper" "state" "$_state" >"$_out"

		# The ownership of the directory and contents should
		# match the system's statd sm directory, so that
		# sm-notify drops privileges and switches to run as
		# the directory owner.
		chown -R --reference="$_ref" "$_sip_dir"
		timeout 10 sm-notify -d -f -m 0 -n -P "$_sip_dir" -v "$_sip"

		rm -rf "$_sip_dir"
	done
}

delete_records()
{
	while read -r _sip _cip; do
		_key="statd-state@${_sip}@${_cip}"
		echo "\"${_key}\" \"\""
	done | $CTDB ptrans "$statd_callout_db"
}

############################################################

# Keep a file per server-IP/client-IP pair, to keep track of the last
# "add-client" or "del-client'.  These get pushed to a database during
# "update", which will generally be run once each "monitor" cycle.  In
# this way we avoid scalability problems with flood of persistent
# transactions after a "notify" when all the clients re-take their
# locks.

startup()
{
	create_add_del_client_dir "$statd_callout_queue_dir"

	$CTDB attach "$statd_callout_db" persistent

	_default="${CTDB_SCRIPT_VARDIR}/statd_callout.conf"
	_config_file="${CTDB_STATD_CALLOUT_CONFIG_FILE:-"${_default}"}"
	cat >"$_config_file" <<EOF
persistent_db
${statd_callout_queue_dir}
${CTDB_MY_PUBLIC_IPS_CACHE}
EOF
}

############################################################

case "$1" in
startup)
	startup
	;;

update)
	cd "$statd_callout_queue_dir" ||
		die "Failed to change directory to \"${statd_callout_queue_dir}\""
	files=$(echo statd-state@*)
	if [ "$files" = "statd-state@*" ]; then
		# No files!
		exit 0
	fi
	sed_expr=$(awk '{
		ip = $1; gsub(/\./, "\\.", ip);
		printf "/statd-state@%s@/p\n", ip }' "$CTDB_MY_PUBLIC_IPS_CACHE")
	# Intentional multi-word expansion for multiple files
	# shellcheck disable=SC2086
	items=$(sed -n "$sed_expr" $files)
	if [ -n "$items" ]; then
		if echo "$items" | $CTDB ptrans "$statd_callout_db"; then
			# shellcheck disable=SC2086
			rm $files
		fi
	fi
	;;

notify)
	# we must restart the lockmanager (on all nodes) so that we get
	# a clusterwide grace period (so other clients don't take out
	# conflicting locks through other nodes before all locks have been
	# reclaimed)

	# we need these settings to make sure that no tcp connections survive
	# across a very fast failover/failback
	#echo 10 > /proc/sys/net/ipv4/tcp_fin_timeout
	#echo 0 > /proc/sys/net/ipv4/tcp_max_tw_buckets
	#echo 0 > /proc/sys/net/ipv4/tcp_max_orphans

	# Delete the notification list for statd, we don't want it to
	# ping any clients
	dir=$(find_statd_sm_dir)
	rm -f "${dir}/"* "${dir}.bak/"*

	# We must also let some time pass between stopping and
	# restarting the lock manager.  Otherwise there is a window
	# where the lock manager will respond "strangely" immediately
	# after restarting it, which causes clients to fail to reclaim
	# their locks.
	nfs_callout_init
	"$CTDB_NFS_CALLOUT" "stop" "nlockmgr" >/dev/null 2>&1
	sleep 2
	"$CTDB_NFS_CALLOUT" "start" "nlockmgr" >/dev/null 2>&1

	# Construct a sed expression to take catdb output and produce pairs of:
	#   server-IP client-IP
	# but only for the server-IPs that are hosted on this node.
	sed_expr=$(awk '{
		       ip = $1; gsub(/\./, "\\.", ip);
		       printf "s/^key.*=.*statd-state@\\(%s\\)@\\([^\"]*\\).*/\\1 \\2/p\n", ip }' \
			       "$CTDB_MY_PUBLIC_IPS_CACHE")

	statd_state=$($CTDB catdb "$statd_callout_db" |
		sed -n "$sed_expr" |
		sort)
	[ -n "$statd_state" ] || exit 0

	echo "$statd_state" | send_notifies
	echo "$statd_state" | delete_records

	# Remove any stale touch files (i.e. for IPs not currently
	# hosted on this node and created since the last "update").
	# There's nothing else we can do with them at this stage.
	pnn=$(ctdb_get_pnn)
	$CTDB ip all |
		tail -n +2 |
		awk -v pnn="$pnn" 'pnn != $2 { print $1 }' |
		while read -r sip; do
			rm -f "${statd_callout_queue_dir}/statd-state@${sip}@"*
		done
	;;
esac