slapos node boot: Start computer partitiong without connecting to master

See merge request nexedi/slapos.core!431

slapos/cli/boot.py

@@ -31,6 +31,7 @@ from __future__ import print_function
 
 import subprocess
 from six.moves.urllib.parse import urlparse
+from six.moves import xmlrpc_client as xmlrpclib
 from time import sleep
 import glob
 import os
@@ -41,6 +42,11 @@ from slapos.cli.command import check_root_user
 from slapos.cli.entry import SlapOSApp
 from slapos.cli.config import ConfigCommand
 from slapos.format import isGlobalScopeAddress
+from slapos.grid.slapgrid import (COMPUTER_PARTITION_REQUESTED_STATE_FILENAME,
+                                  COMPUTER_PARTITION_STARTED_STATE)
+from slapos.grid.svcbackend import  (_getSupervisordSocketPath,
+                                     getSupervisorRPC,
+                                     launchSupervisord)
 from slapos.util import string_to_boolean
 import argparse
 import logging
@@ -59,6 +65,48 @@ def _removeTimestamp(instancehome, partition_base_name):
        logger.info("Removing %s", timestamp_path)
        os.remove(timestamp_path)
 
+def _startComputerPartition(partition_id, supervisord_socket):
+    """
+    With supervisord, start the instance that was deployed
+    """
+    try:
+      with getSupervisorRPC(supervisord_socket) as supervisor:
+        supervisor.startProcessGroup(partition_id, False)
+    except xmlrpclib.Fault as exc:
+      if exc.faultString.startswith('BAD_NAME:'):
+        logger.info("Nothing to start on %s...", partition_id)
+      else:
+        raise
+    else:
+      logger.info("Requested start of %s...", partition_id)
+
+def _startComputerPartitionList(instance_root, partition_base_name):
+  """
+  Start services for partition which has requested state to 'started'
+  """
+  partition_glob_path = os.path.join(
+        instance_root,
+        "%s*" % partition_base_name)
+  launchSupervisord(instance_root=instance_root, logger=logger)
+  for partition_path in glob.glob(partition_glob_path):
+    partition_state_path = os.path.join(
+        partition_path,
+        COMPUTER_PARTITION_REQUESTED_STATE_FILENAME
+    )
+    supervisord_socket_path = _getSupervisordSocketPath(
+      instance_root,
+      logger
+    )
+    if os.path.exists(partition_state_path):
+      partition_state = ""
+      with open(partition_state_path) as f:
+        partition_state = f.read()
+      if partition_state == COMPUTER_PARTITION_STARTED_STATE:
+        # Call start for this computer partition
+        _startComputerPartition(
+          os.path.basename(partition_path.rstrip('/')),
+          supervisord_socket_path
+        )
 
 def _runBang(app):
     """
@@ -76,7 +124,9 @@ def _runFormat(app):
     Launch slapos node format.
     """
     logger.info("[BOOT] Invoking slapos node format...")
-    result = app.run(['node', 'format', '--now', '--verbose'])
+    # '--local' parameter is to prevent node format command to post data to
+    # master, so this command can work without internet and setup partitions IP.
+    result = app.run(['node', 'format', '--now', '--local', '--verbose'])
     if result == 1:
       return 0
     return 1
@@ -196,6 +246,15 @@ class BootCommand(ConfigCommand):
         if ipv6_interface is not None:
             _waitIpv6Ready(ipv6_interface)
 
+        app = SlapOSApp()
+        # Make sure slapos node format returns ok
+        while not _runFormat(app):
+            logger.error("[BOOT] Fail to format, try again in 15 seconds...")
+            sleep(15)
+
+        # Start computer partition services
+        _startComputerPartitionList(instance_root, partition_base_name)
+
         # Check that node can ping master
         if valid_ipv4(master_hostname):
             _test_ping(master_hostname)
@@ -205,12 +264,6 @@ class BootCommand(ConfigCommand):
             # hostname
             _ping_hostname(master_hostname)
 
-        app = SlapOSApp()
-        # Make sure slapos node format returns ok
-        while not _runFormat(app):
-            logger.error("[BOOT] Fail to format, try again in 15 seconds...")
-            sleep(15)
-
         # Make sure slapos node bang returns ok
         while not _runBang(app):
             logger.error("[BOOT] Fail to bang, try again in 15 seconds...")

slapos/cli/format.py

@@ -83,6 +83,12 @@ class FormatCommand(ConfigCommand):
                         help='Launch slapformat without delay'
                              ' (default: %(default)s)')
 
+        ap.add_argument('--local',
+                        default=False, # can have a default as it is not in .cfg
+                        action="store_true",
+                        help='Keep format data locally, do not post xml to master'
+                             ' (default: %(default)s)')
+
         ap.add_argument('-n', '--dry_run',
                         default=False, # can have a default as it is not in .cfg
                         action="store_true",

slapos/format.py

@@ -1408,6 +1408,7 @@ def do_format(conf):
     computer.dump(path_to_xml=conf.computer_xml,
                   path_to_json=conf.computer_json,
                   logger=conf.logger)
+  if not conf.local:
     conf.logger.info('Posting information to %r' % conf.master_url)
     computer.send(conf)
   conf.logger.info('slapos successfully prepared the computer.')

slapos/grid/slapgrid.py

@@ -90,6 +90,7 @@ SLAPGRID_PROMISE_FAIL = 2
 PROMISE_TIMEOUT = 20
 
 COMPUTER_PARTITION_TIMESTAMP_FILENAME = '.timestamp'
+COMPUTER_PARTITION_REQUESTED_STATE_FILENAME = '.requested_state'
 COMPUTER_PARTITION_LATEST_BANG_TIMESTAMP_FILENAME = '.slapos_latest_bang_timestamp'
 COMPUTER_PARTITION_INSTALL_ERROR_FILENAME = '.slapgrid-%s-error.log'
 COMPUTER_PARTITION_WAIT_LIST_FILENAME = '.slapos-report-wait-service-list'
@@ -1125,6 +1126,10 @@ stderr_logfile_backups=1
         instance_path,
         COMPUTER_PARTITION_TIMESTAMP_FILENAME
     )
+    partition_state_path = os.path.join(
+        instance_path,
+        COMPUTER_PARTITION_REQUESTED_STATE_FILENAME
+    )
     parameter_dict = computer_partition.getInstanceParameterDict()
     timestamp = parameter_dict.get('timestamp')
 
@@ -1225,6 +1230,7 @@ stderr_logfile_backups=1
 
               return
         os.remove(timestamp_path)
+        os.remove(partition_state_path)
 
     # Include Partition Logging
     log_folder_path = "%s/.slapgrid/log" % instance_path
@@ -1339,6 +1345,8 @@ stderr_logfile_backups=1
     if timestamp:
       with open(timestamp_path, 'w') as f:
         f.write(str(timestamp))
+      with open(partition_state_path, 'w') as f:
+        f.write(str(computer_partition_state))
 
   def FilterComputerPartitionList(self, computer_partition_list):
     """

slapos/tests/test_cli.py

@@ -415,8 +415,13 @@ class TestCliBoot(CliMixin):
     os.mkdir(os.path.join(instance_root, partition_base_name + '1'))
     timestamp = os.path.join(
         instance_root, partition_base_name + '1', '.timestamp')
+    requested_state_path = os.path.join(instance_root,
+                                        partition_base_name + '1',
+                                        '.requested_state')
     with open(timestamp, 'w') as f:
       f.write("1578552471")
+    with open(requested_state_path, 'w') as f:
+      f.write("started")
 
     # make a config file using this instance root
     with tempfile.NamedTemporaryFile(mode='w') as slapos_conf:
@@ -441,17 +446,21 @@ class TestCliBoot(CliMixin):
           patch(
               'slapos.cli.boot.netifaces.ifaddresses',
               return_value={socket.AF_INET6: ({'addr': '2000::1'},),},) as ifaddresses,\
+          patch('slapos.cli.boot._startComputerPartition', return_value=None) as start_partition,\
+          patch('slapos.cli.boot.launchSupervisord', return_value=None),\
           patch('slapos.cli.boot._ping_hostname', return_value=1) as _ping_hostname:
         app.run(('node', 'boot'))
 
       # boot command runs as root
       check_root_user.assert_called_once()
+      # Computer partition was started during boot
+      start_partition.assert_called_once()
       # it waits for interface to have an IPv6 address
       ifaddresses.assert_called_once_with('interface_name_from_config')
       # then ping master hostname to wait for connectivity
       _ping_hostname.assert_called_once_with('slap.vifib.com')
       # then format and bang
-      SlapOSApp().run.assert_any_call(['node', 'format', '--now', '--verbose'])
+      SlapOSApp().run.assert_any_call(['node', 'format', '--now', '--local', '--verbose'])
       SlapOSApp().run.assert_any_call(['node', 'bang', '-m', 'Reboot'])
 
       # timestamp files have been removed
@@ -473,6 +482,7 @@ class TestCliBoot(CliMixin):
         patch('slapos.cli.boot.netifaces.ifaddresses',
               side_effect=[net1, net2, net3]),\
         patch('slapos.cli.boot._ping_hostname', return_value=0),\
+        patch('slapos.cli.boot._startComputerPartitionList', return_value=None) as start_partition,\
         patch('slapos.cli.format.check_root_user', return_value=True),\
         patch('slapos.cli.format.logging.FileHandler', return_value=logging.NullHandler()),\
         patch('slapos.cli.bang.check_root_user', return_value=True),\
@@ -482,6 +492,7 @@ class TestCliBoot(CliMixin):
       app.run(('node', 'boot'))
 
     check_root_user.assert_called_once()
+    start_partition.assert_called_once()
 
     self.assertEqual(do_format.call_count, 3)
     self.assertEqual(do_bang.call_count, 3)

slapos/tests/test_slapgrid.py

@@ -1401,7 +1401,7 @@ class TestSlapgridCPPartitionProcessing(MasterMixin, unittest.TestCase):
       self.assertInstanceDirectoryListEqual(['0'])
       partition = os.path.join(self.instance_root, '0')
       six.assertCountEqual(self, os.listdir(partition),
-                            ['.slapgrid', '.timestamp', 'buildout.cfg', 'software_release', 'worked', '.slapos-retention-lock-delay'])
+                            ['.slapgrid', '.timestamp', '.requested_state', 'buildout.cfg', 'software_release', 'worked', '.slapos-retention-lock-delay'])
       six.assertCountEqual(self, os.listdir(self.software_root), [instance.software.software_hash])
       timestamp_path = os.path.join(instance.partition_path, '.timestamp')
       self.setSlapgrid()
@@ -1422,7 +1422,7 @@ class TestSlapgridCPPartitionProcessing(MasterMixin, unittest.TestCase):
       self.assertInstanceDirectoryListEqual(['0'])
       partition = os.path.join(self.instance_root, '0')
       six.assertCountEqual(self, os.listdir(partition),
-                            ['.slapgrid', '.timestamp', 'buildout.cfg',
+                            ['.slapgrid', '.timestamp', '.requested_state', 'buildout.cfg',
                              'software_release', 'worked', '.slapos-retention-lock-delay'])
       six.assertCountEqual(self, os.listdir(self.software_root), [instance.software.software_hash])
 
@@ -1445,7 +1445,7 @@ class TestSlapgridCPPartitionProcessing(MasterMixin, unittest.TestCase):
       self.assertInstanceDirectoryListEqual(['0'])
       partition = os.path.join(self.instance_root, '0')
       six.assertCountEqual(self, os.listdir(partition),
-                            ['.slapgrid', '.timestamp', 'buildout.cfg', 'software_release', 'worked', '.slapos-retention-lock-delay'])
+                            ['.slapgrid', '.timestamp', '.requested_state', 'buildout.cfg', 'software_release', 'worked', '.slapos-retention-lock-delay'])
       six.assertCountEqual(self, os.listdir(self.software_root), [instance.software.software_hash])
       instance.timestamp = str(int(timestamp) - 1)
       self.assertEqual(self.launchSlapgrid(), slapgrid.SLAPGRID_SUCCESS)
@@ -1463,7 +1463,7 @@ class TestSlapgridCPPartitionProcessing(MasterMixin, unittest.TestCase):
       self.assertInstanceDirectoryListEqual(['0'])
       partition = os.path.join(self.instance_root, '0')
       six.assertCountEqual(self, os.listdir(partition),
-                            ['.slapgrid', '.timestamp', 'buildout.cfg', 'software_release', 'worked', '.slapos-retention-lock-delay'])
+                            ['.slapgrid', '.timestamp', '.requested_state', 'buildout.cfg', 'software_release', 'worked', '.slapos-retention-lock-delay'])
       six.assertCountEqual(self, os.listdir(self.software_root), [instance.software.software_hash])
       instance.timestamp = str(int(timestamp) + 1)
       self.assertEqual(self.launchSlapgrid(), slapgrid.SLAPGRID_SUCCESS)
@@ -1491,7 +1491,7 @@ class TestSlapgridCPPartitionProcessing(MasterMixin, unittest.TestCase):
       self.assertInstanceDirectoryListEqual(['0'])
       partition = os.path.join(self.instance_root, '0')
       six.assertCountEqual(self, os.listdir(partition),
-                            ['.slapgrid', '.timestamp', 'buildout.cfg', 'software_release', 'worked', '.slapos-retention-lock-delay'])
+                            ['.slapgrid', '.timestamp', '.requested_state', 'buildout.cfg', 'software_release', 'worked', '.slapos-retention-lock-delay'])
       six.assertCountEqual(self, os.listdir(self.software_root),
                             [instance.software.software_hash])
       instance.timestamp = None
@@ -1523,7 +1523,7 @@ class TestSlapgridCPPartitionProcessing(MasterMixin, unittest.TestCase):
       self.launchSlapgrid()
       partition = os.path.join(self.instance_root, '0')
       six.assertCountEqual(self, os.listdir(partition),
-                            ['.slapgrid', '.timestamp', 'buildout.cfg',
+                            ['.slapgrid', '.timestamp', '.requested_state', 'buildout.cfg',
                              'software_release', 'worked', '.slapos-retention-lock-delay'])
 
       time.sleep(2)
@@ -1533,7 +1533,7 @@ class TestSlapgridCPPartitionProcessing(MasterMixin, unittest.TestCase):
 
       self.launchSlapgrid()
       six.assertCountEqual(self, os.listdir(partition),
-                            ['.slapgrid', '.timestamp', 'buildout.cfg',
+                            ['.slapgrid', '.timestamp', '.requested_state', 'buildout.cfg',
                              'software_release', 'worked', '.slapos-retention-lock-delay'])
 
   def test_one_partition_periodicity_from_file_does_not_disturb_others(self):
@@ -1710,6 +1710,43 @@ class TestSlapgridCPPartitionProcessing(MasterMixin, unittest.TestCase):
         self.launchSlapgrid()
         self.assertEqual(mock_method.call_count, 2)
 
+  def test_partition_requested_state_created(self):
+    computer = self.getTestComputerClass()(self.software_root, self.instance_root)
+    with httmock.HTTMock(computer.request_handler):
+      instance = computer.instance_list[0]
+      timestamp = str(int(time.time()))
+      instance.timestamp = timestamp
+
+      self.assertEqual(self.grid.processComputerPartitionList(), slapgrid.SLAPGRID_SUCCESS)
+      self.assertInstanceDirectoryListEqual(['0'])
+      partition = os.path.join(self.instance_root, '0')
+      six.assertCountEqual(self, os.listdir(partition),
+                            ['.slapgrid', '.timestamp', '.requested_state', 'buildout.cfg',
+                             'software_release', 'worked', '.slapos-retention-lock-delay'])
+      six.assertCountEqual(self, os.listdir(self.software_root), [instance.software.software_hash])
+      requested_state_path = os.path.join(instance.partition_path, '.requested_state')
+
+      with open(requested_state_path) as f:
+        self.assertEqual(f.read(), slapgrid.COMPUTER_PARTITION_STOPPED_STATE)
+      self.assertEqual(instance.sequence,
+                       ['/stoppedComputerPartition'])
+
+  def test_partition_requested_state_not_created_if_failed(self):
+    computer = self.getTestComputerClass()(self.software_root, self.instance_root)
+    with httmock.HTTMock(computer.request_handler):
+      instance = computer.instance_list[0]
+      timestamp = str(int(time.time()))
+      instance.timestamp = timestamp
+
+      instance.software.setBuildout("""#!/bin/sh
+exit 3""")
+      self.assertEqual(self.grid.processComputerPartitionList(), slapgrid.SLAPGRID_FAIL)
+      self.assertInstanceDirectoryListEqual(['0'])
+      self.assertEqual(instance.sequence,
+                       ['/softwareInstanceError'])
+      requested_state_path = os.path.join(instance.partition_path, '.requested_state')
+      self.assertFalse(os.path.exists(requested_state_path))
+
   def test_one_partition_buildout_fail_does_not_disturb_others(self):
     """
     1. We set up two instance one using a corrupted buildout