ctdb/config/statd-callout

   1 #!/bin/sh
   2
   3 # This must run as root as CTDB tool commands need to access CTDB socket
   4 [ "$(id -u)" -eq 0 ] || exec sudo "$0" "$@"
   5
   6 # statd must be configured to use this script as its high availability call-out.
   7 #
   8 # In most Linux versions this can be done using something like the following...
   9 #
  10 # /etc/sysconfig/nfs (Red Hat) or /etc/default/nfs-common (Debian):
  11 #   NFS_HOSTNAME=myhostname
  12 #   STATD_HOSTNAME="${NFS_HOSTNAME} -H /etc/ctdb/statd-callout"
  13 #
  14 # Newer Red Hat Linux variants instead use /etc/nfs.conf:
  15 #   [statd]
  16 #     name = myhostname
  17 #     ha-callout = /etc/ctdb/statd-callout
  18
  19 [ -n "$CTDB_BASE" ] || \
  20     CTDB_BASE=$(d=$(dirname "$0") ; cd -P "$d" ; echo "$PWD")
  21
  22 . "${CTDB_BASE}/functions"
  23
  24 # Overwrite this so we get some logging
  25 die ()
  26 {
  27     script_log "statd-callout" "$@"
  28     exit 1
  29 }
  30
  31 # Try different variables to find config file for NFS_HOSTNAME
  32 load_system_config "nfs" "nfs-common"
  33
  34 # If NFS_HOSTNAME not set then try to pull it out of /etc/nfs.conf
  35 if [ -z "$NFS_HOSTNAME" ] && type nfsconf >/dev/null 2>&1 ; then
  36         NFS_HOSTNAME=$(nfsconf --get statd name)
  37 fi
  38
  39 [ -n "$NFS_HOSTNAME" ] || \
  40     die "NFS_HOSTNAME is not configured. statd-callout failed"
  41
  42 ############################################################
  43
  44 ctdb_setup_state_dir "service" "nfs"
  45
  46 # script_state_dir set by ctdb_setup_state_dir()
  47 # shellcheck disable=SC2154
  48 d="${script_state_dir}/statd-callout"
  49
  50 mkdir -p "$d" || die "Failed to create directory \"${d}\""
  51 cd "$d" || die "Failed to change directory to \"${d}\""
  52
  53 pnn=$(ctdb_get_pnn)
  54
  55 ############################################################
  56
  57 send_notifies ()
  58 {
  59         _smnotify="${CTDB_HELPER_BINDIR}/smnotify"
  60
  61         # State must monotonically increase, across the entire
  62         # cluster.  Use seconds since epoch and hope the time is in
  63         # sync across nodes.  Even numbers mean service is shut down,
  64         # odd numbers mean service is started.
  65
  66         # Intentionally round to an even number
  67         # shellcheck disable=SC2017
  68         _state_even=$(( $(date '+%s') / 2 * 2))
  69
  70         _prev=""
  71         while read _sip _cip ; do
  72                 # NOTE: Consider optimising smnotify to read all the
  73                 # data from stdin and then run it in the background.
  74
  75                 # Reset stateval for each serverip
  76                 if [ "$_sip" != "$_prev" ] ; then
  77                         _stateval="$_state_even"
  78                 fi
  79
  80                 # Send notifies for server shutdown
  81                 "$_smnotify" --client="$_cip" --ip="$_sip" \
  82                              --server="$_sip" --stateval="$_stateval"
  83                 "$_smnotify" --client="$_cip" --ip="$_sip" \
  84                              --server="$NFS_HOSTNAME" --stateval="$_stateval"
  85
  86                 # Send notifies for server startup
  87                 _stateval=$((_stateval + 1))
  88                 "$_smnotify" --client="$_cip" --ip="$_sip" \
  89                              --server="$_sip" --stateval="$_stateval"
  90                 "$_smnotify" --client="$_cip" --ip="$_sip" \
  91                              --server="$NFS_HOSTNAME" --stateval="$_stateval"
  92         done
  93 }
  94
  95 delete_records ()
  96 {
  97         while read _sip _cip ; do
  98                 _key="statd-state@${_sip}@${_cip}"
  99                 echo "\"${_key}\" \"\""
 100         done | $CTDB ptrans "ctdb.tdb"
 101 }
 102
 103 ############################################################
 104
 105 case "$1" in
 106     # Keep a single file to keep track of the last "add-client" or
 107     # "del-client'.  These get pushed to ctdb.tdb during "update",
 108     # which will generally be run once each "monitor" cycle.  In this
 109     # way we avoid scalability problems with flood of persistent
 110     # transactions after a "notify" when all the clients re-take their
 111     # locks.
 112
 113     add-client)
 114         # statd does not tell us to which IP the client connected so
 115         # we must add it to all the IPs that we serve
 116         cip="$2"
 117         date=$(date '+%s')
 118         # x is intentionally ignored
 119         # shellcheck disable=SC2034
 120         $CTDB ip -X |
 121         tail -n +2 |
 122         while IFS="|" read x sip node x ; do
 123             [ "$node" = "$pnn" ] || continue # not us
 124             key="statd-state@${sip}@${cip}"
 125             echo "\"${key}\" \"${date}\"" >"$key"
 126         done
 127         ;;
 128
 129     del-client)
 130         # statd does not tell us from which IP the client disconnected
 131         # so we must add it to all the IPs that we serve
 132         cip="$2"
 133         # x is intentionally ignored
 134         # shellcheck disable=SC2034
 135         $CTDB ip -X |
 136         tail -n +2 |
 137         while IFS="|" read x sip node x ; do
 138             [ "$node" = "$pnn" ] || continue # not us
 139             key="statd-state@${sip}@${cip}"
 140             echo "\"${key}\" \"\"" >"$key"
 141         done
 142         ;;
 143
 144     update)
 145         files=$(echo statd-state@*)
 146         if [ "$files" = "statd-state@*" ] ; then
 147             # No files!
 148             exit 0
 149         fi
 150         # Filter out lines for any IP addresses that are not currently
 151         # hosted public IP addresses.
 152         ctdb_ips=$($CTDB ip | tail -n +2)
 153         sed_expr=$(echo "$ctdb_ips" |
 154             awk -v pnn="$pnn" 'pnn == $2 {
 155                 ip = $1; gsub(/\./, "\\.", ip);
 156                 printf "/statd-state@%s@/p\n", ip }')
 157         # Intentional multi-word expansion for multiple files
 158         # shellcheck disable=SC2086
 159         items=$(sed -n "$sed_expr" $files)
 160         if [ -n "$items" ] ; then
 161                 if echo "$items"  | $CTDB ptrans "ctdb.tdb" ; then
 162                         # shellcheck disable=SC2086
 163                         rm $files
 164                 fi
 165         fi
 166         ;;
 167
 168     notify)
 169         # we must restart the lockmanager (on all nodes) so that we get
 170         # a clusterwide grace period (so other clients don't take out
 171         # conflicting locks through other nodes before all locks have been
 172         # reclaimed)
 173
 174         # we need these settings to make sure that no tcp connections survive
 175         # across a very fast failover/failback
 176         #echo 10 > /proc/sys/net/ipv4/tcp_fin_timeout
 177         #echo 0 > /proc/sys/net/ipv4/tcp_max_tw_buckets
 178         #echo 0 > /proc/sys/net/ipv4/tcp_max_orphans
 179
 180         # Delete the notification list for statd, we don't want it to
 181         # ping any clients
 182         rm -f /var/lib/nfs/statd/sm/*
 183         rm -f /var/lib/nfs/statd/sm.bak/*
 184
 185         # We must also let some time pass between stopping and
 186         # restarting the lock manager.  Otherwise there is a window
 187         # where the lock manager will respond "strangely" immediately
 188         # after restarting it, which causes clients to fail to reclaim
 189         # their locks.
 190         nfs_callout_init
 191         "$CTDB_NFS_CALLOUT" "stop" "nlockmgr" >/dev/null 2>&1
 192         sleep 2
 193         "$CTDB_NFS_CALLOUT" "start" "nlockmgr" >/dev/null 2>&1
 194
 195         # we now need to send out additional statd notifications to ensure
 196         # that clients understand that the lockmanager has restarted.
 197         # we have three cases:
 198         # 1, clients that ignore the ip address the stat notification came from
 199         #    and ONLY care about the 'name' in the notify packet.
 200         #    these clients ONLY work with lock failover IFF that name
 201         #    can be resolved into an ipaddress that matches the one used
 202         #    to mount the share.  (==linux clients)
 203         #    This is handled when starting lockmanager above,  but those
 204         #    packets are sent from the "wrong" ip address, something linux
 205         #    clients are ok with, buth other clients will barf at.
 206         # 2, Some clients only accept statd packets IFF they come from the
 207         #    'correct' ip address.
 208         # 2a,Send out the notification using the 'correct' ip address and also
 209         #    specify the 'correct' hostname in the statd packet.
 210         #    Some clients require both the correct source address and also the
 211         #    correct name. (these clients also ONLY work if the ip addresses
 212         #    used to map the share can be resolved into the name returned in
 213         #    the notify packet.)
 214         # 2b,Other clients require that the source ip address of the notify
 215         #    packet matches the ip address used to take out the lock.
 216         #    I.e. that the correct source address is used.
 217         #    These clients also require that the statd notify packet contains
 218         #    the name as the ip address used when the lock was taken out.
 219         #
 220         # Both 2a and 2b are commonly used in lockmanagers since they maximize
 221         # probability that the client will accept the statd notify packet and
 222         # not just ignore it.
 223         # For all IPs we serve, collect info and push to the config database
 224
 225         # Construct a sed expression to take catdb output and produce pairs of:
 226         #   server-IP client-IP
 227         # but only for the server-IPs that are hosted on this node.
 228         ctdb_all_ips=$($CTDB ip all | tail -n +2)
 229         sed_expr=$(echo "$ctdb_all_ips" |
 230             awk -v pnn="$pnn" 'pnn == $2 {
 231                 ip = $1; gsub(/\./, "\\.", ip);
 232                 printf "s/^key.*=.*statd-state@\\(%s\\)@\\([^\"]*\\).*/\\1 \\2/p\n", ip }')
 233
 234         statd_state=$($CTDB catdb ctdb.tdb | sed -n "$sed_expr" | sort)
 235         [ -n "$statd_state" ] || exit 0
 236
 237         echo "$statd_state" | send_notifies
 238         echo "$statd_state" | delete_records
 239
 240         # Remove any stale touch files (i.e. for IPs not currently
 241         # hosted on this node and created since the last "update").
 242         # There's nothing else we can do with them at this stage.
 243         echo "$ctdb_all_ips" |
 244             awk -v pnn="$pnn" 'pnn != $2 { print $1 }' |
 245             while read sip ; do
 246                 rm -f "statd-state@${sip}@"*
 247             done
 248         ;;
 249 esac