findbadpcu.py

   1 #!/usr/bin/python
   2
   3 import os
   4 import sys
   5 import string
   6 import time
   7 import socket
   8
   9
  10 import signal
  11
  12 #old_handler = signal.getsignal(signal.SIGCHLD)
  13
  14 #def sig_handler(signum, stack):
  15 #       """ Handle SIGCHLD signal """
  16 #       global old_handler
  17 #       if signum == signal.SIGCHLD:
  18 #               try:
  19 #                       os.wait()
  20 #               except:
  21 #                       pass
  22 #       if old_handler != signal.SIG_DFL:
  23 #               old_handler(signum, stack)
  24 #
  25 #orig_sig_handler = signal.signal(signal.SIGCHLD, sig_handler)
  26
  27 from config import config
  28 from optparse import OptionParser
  29 parser = OptionParser()
  30 parser.set_defaults(filename="",
  31                                         increment=False,
  32                                         dbname="findbadpcus",
  33                                         cachenodes=False,
  34                                         refresh=False,
  35                                         )
  36 parser.add_option("-f", "--nodelist", dest="filename", metavar="FILE",
  37                                         help="Provide the input file for the node list")
  38 parser.add_option("", "--cachenodes", action="store_true",
  39                                         help="Cache node lookup from PLC")
  40 parser.add_option("", "--dbname", dest="dbname", metavar="FILE",
  41                                         help="Specify the name of the database to which the information is saved")
  42 parser.add_option("", "--refresh", action="store_true", dest="refresh",
  43                                         help="Refresh the cached values")
  44 parser.add_option("-i", "--increment", action="store_true", dest="increment",
  45                                         help="Increment round number to force refresh or retry")
  46 config = config(parser)
  47 config.parse_args()
  48
  49 # QUERY all nodes.
  50 COMON_COTOPURL= "http://summer.cs.princeton.edu/status/tabulator.cgi?" + \
  51                                         "table=table_nodeview&" + \
  52                                     "dumpcols='name,resptime,sshstatus,uptime,lastcotop'&" + \
  53                                     "formatcsv"
  54                                     #"formatcsv&" + \
  55                                         #"select='lastcotop!=0'"
  56
  57 import threading
  58 plc_lock = threading.Lock()
  59 round = 1
  60 externalState = {'round': round, 'nodes': {'a': None}}
  61 errorState = {}
  62 count = 0
  63
  64 import reboot
  65 from reboot import pcu_name
  66
  67 import soltesz
  68 import plc
  69 import comon
  70 import threadpool
  71 import syncplcdb
  72
  73 def nmap_portstatus(status):
  74         ps = {}
  75         l_nmap = status.split()
  76         ports = l_nmap[4:]
  77
  78         continue_probe = False
  79         for port in ports:
  80                 results = port.split('/')
  81                 ps[results[0]] = results[1]
  82                 if results[1] == "open":
  83                         continue_probe = True
  84         return (ps, continue_probe)
  85
  86 def collectPingAndSSH(pcuname, cohash):
  87
  88         continue_probe = True
  89         errors = None
  90         values = {}
  91         ### GET PCU ######################
  92         try:
  93                 b_except = False
  94                 plc_lock.acquire()
  95
  96                 try:
  97                         l_pcu  = plc.GetPCUs({'pcu_id' : pcuname})
  98
  99                         if len(l_pcu) > 0:
 100                                 site_id = l_pcu[0]['site_id']
 101
 102                                 node_ids = l_pcu[0]['node_ids']
 103                                 l_node = plc.getNodes(node_ids, ['hostname', 'last_contact',
 104                                                                                                  'node_id', 'ports'])
 105                         if len(l_node) > 0:
 106                                 for node in l_node:
 107                                         values[node['hostname']] = node['ports'][0]
 108
 109                                 values['nodenames'] = [node['hostname'] for node in l_node]
 110                                 # NOTE: this is for a dry run later. It doesn't matter which node.
 111                                 values['node_id'] = l_node[0]['node_id']
 112
 113                         if len(l_pcu) > 0:
 114                                 values.update(l_pcu[0])
 115                         else:
 116                                 continue_probe = False
 117
 118                 except:
 119                         b_except = True
 120                         import traceback
 121                         traceback.print_exc()
 122
 123                         continue_probe = False
 124
 125                 plc_lock.release()
 126                 if b_except: return (None, None)
 127
 128                 if values['hostname'] is not None:
 129                         values['hostname'] = values['hostname'].strip()
 130
 131                 if values['ip'] is not None:
 132                         values['ip'] = values['ip'].strip()
 133
 134                 #### COMPLETE ENTRY   #######################
 135
 136                 values['complete_entry'] = []
 137                 #if values['protocol'] is None or values['protocol'] is "":
 138                 #       values['complete_entry'] += ["protocol"]
 139                 if values['model'] is None or values['model'] is "":
 140                         values['complete_entry'] += ["model"]
 141                         # Cannot continue due to this condition
 142                         continue_probe = False
 143
 144                 if values['password'] is None or values['password'] is "":
 145                         values['complete_entry'] += ["password"]
 146                         # Cannot continue due to this condition
 147                         continue_probe = False
 148
 149                 if len(values['complete_entry']) > 0:
 150                         continue_probe = False
 151
 152                 if values['hostname'] is None or values['hostname'] is "":
 153                         values['complete_entry'] += ["hostname"]
 154                 if values['ip'] is None or values['ip'] is "":
 155                         values['complete_entry'] += ["ip"]
 156
 157                 # If there are no nodes associated with this PCU, then we cannot continue.
 158                 if len(values['node_ids']) == 0:
 159                         continue_probe = False
 160                         values['complete_entry'] += ['NoNodeIds']
 161
 162                 #### DNS and IP MATCH #######################
 163                 if values['hostname'] is not None and values['hostname'] is not "" and \
 164                    values['ip'] is not None and values['ip'] is not "":
 165                         #print "Calling socket.gethostbyname(%s)" % values['hostname']
 166                         try:
 167                                 ipaddr = socket.gethostbyname(values['hostname'])
 168                                 if ipaddr == values['ip']:
 169                                         values['dnsmatch'] = "DNS-OK"
 170                                 else:
 171                                         values['dnsmatch'] = "DNS-MISMATCH"
 172                                         continue_probe = False
 173
 174                         except Exception, err:
 175                                 values['dnsmatch'] = "DNS-NOENTRY"
 176                                 values['hostname'] = values['ip']
 177                                 #print err
 178                 else:
 179                         if values['ip'] is not None and values['ip'] is not "":
 180                                 values['dnsmatch'] = "NOHOSTNAME"
 181                                 values['hostname'] = values['ip']
 182                         else:
 183                                 values['dnsmatch'] = "NO-DNS-OR-IP"
 184                                 values['hostname'] = "No_entry_in_DB"
 185                                 continue_probe = False
 186
 187                 #### RUN NMAP ###############################
 188                 if continue_probe:
 189                         nmap = soltesz.CMD()
 190                         (oval,eval) = nmap.run_noexcept("nmap -oG - -P0 -p22,23,80,443,5869,16992 %s | grep Host:" % pcu_name(values))
 191                         # NOTE: an empty / error value for oval, will still work.
 192                         (values['portstatus'], continue_probe) = nmap_portstatus(oval)
 193                 else:
 194                         values['portstatus'] = None
 195
 196
 197                 ######  DRY RUN  ############################
 198                 if 'node_ids' in values and len(values['node_ids']) > 0:
 199                         rb_ret = reboot.reboot_test(values['nodenames'][0], values, continue_probe, 1, True)
 200                 else:
 201                         rb_ret = "Not_Run" # No nodes to test"
 202
 203                 values['reboot'] = rb_ret
 204
 205                 ### GET PLC SITE ######################
 206                 b_except = False
 207                 plc_lock.acquire()
 208
 209                 try:
 210                         d_site = plc.getSites({'site_id': site_id},
 211                                                                 ['max_slices', 'slice_ids', 'node_ids', 'login_base'])
 212                 except:
 213                         b_except = True
 214                         import traceback
 215                         traceback.print_exc()
 216
 217                 plc_lock.release()
 218                 if b_except: return (None, None)
 219
 220                 if d_site and len(d_site) > 0:
 221                         max_slices = d_site[0]['max_slices']
 222                         num_slices = len(d_site[0]['slice_ids'])
 223                         num_nodes = len(d_site[0]['node_ids'])
 224                         loginbase = d_site[0]['login_base']
 225                         values['plcsite'] = {'num_nodes' : num_nodes,
 226                                                                 'max_slices' : max_slices,
 227                                                                 'num_slices' : num_slices,
 228                                                                 'login_base' : loginbase,
 229                                                                 'status'     : 'SUCCESS'}
 230                 else:
 231                         values['plcsite'] = {'status' : "GS_FAILED"}
 232         except:
 233                 print "____________________________________"
 234                 print values
 235                 errors = values
 236                 print "____________________________________"
 237                 import traceback
 238                 errors['traceback'] = traceback.format_exc()
 239                 print errors['traceback']
 240
 241         values['checked'] = time.time()
 242         return (pcuname, values, errors)
 243
 244 def recordPingAndSSH(request, result):
 245         global errorState
 246         global externalState
 247         global count
 248         (nodename, values, errors) = result
 249
 250         if values is not None:
 251                 global_round = externalState['round']
 252                 pcu_id = "id_%s" % nodename
 253                 externalState['nodes'][pcu_id]['values'] = values
 254                 externalState['nodes'][pcu_id]['round'] = global_round
 255
 256                 count += 1
 257                 print "%d %s %s" % (count, nodename, externalState['nodes'][pcu_id]['values'])
 258                 soltesz.dbDump(config.dbname, externalState)
 259
 260         if errors is not None:
 261                 pcu_id = "id_%s" % nodename
 262                 errorState[pcu_id] = errors
 263                 soltesz.dbDump("findbadpcu_errors", errorState)
 264
 265 # this will be called when an exception occurs within a thread
 266 def handle_exception(request, result):
 267         print "Exception occured in request %s" % request.requestID
 268         for i in result:
 269                 print "Result: %s" % i
 270
 271
 272 def checkAndRecordState(l_pcus, cohash):
 273         global externalState
 274         global count
 275         global_round = externalState['round']
 276
 277         tp = threadpool.ThreadPool(20)
 278
 279         # CREATE all the work requests
 280         for pcuname in l_pcus:
 281                 pcu_id = "id_%s" % pcuname
 282                 if pcuname not in externalState['nodes']:
 283                         #print type(externalState['nodes'])
 284
 285                         externalState['nodes'][pcu_id] = {'round': 0, 'values': []}
 286
 287                 node_round   = externalState['nodes'][pcu_id]['round']
 288                 if node_round < global_round:
 289                         # recreate node stats when refreshed
 290                         #print "%s" % nodename
 291                         req = threadpool.WorkRequest(collectPingAndSSH, [pcuname, cohash], {},
 292                                                                                  None, recordPingAndSSH, handle_exception)
 293                         tp.putRequest(req)
 294                 else:
 295                         # We just skip it, since it's "up to date"
 296                         count += 1
 297                         print "%d %s %s" % (count, pcu_id, externalState['nodes'][pcu_id]['values'])
 298                         pass
 299
 300         # WAIT while all the work requests are processed.
 301         while 1:
 302                 try:
 303                         time.sleep(1)
 304                         tp.poll()
 305                 except KeyboardInterrupt:
 306                         print "Interrupted!"
 307                         break
 308                 except threadpool.NoResultsPending:
 309                         print "All results collected."
 310                         break
 311
 312
 313
 314 def main():
 315         global externalState
 316
 317         externalState = soltesz.if_cached_else(1, config.dbname, lambda : externalState)
 318         cohash = {}
 319
 320         if config.increment:
 321                 # update global round number to force refreshes across all nodes
 322                 externalState['round'] += 1
 323
 324         if config.filename == "":
 325                 print "Calling API GetPCUs() : refresh(%s)" % config.refresh
 326                 l_pcus = soltesz.if_cached_else_refresh(1,
 327                                                                 config.refresh, "pculist", lambda : plc.GetPCUs())
 328                 l_pcus  = [pcu['pcu_id'] for pcu in l_pcus]
 329         else:
 330                 l_pcus = config.getListFromFile(config.filename)
 331                 l_pcus = [int(pcu) for pcu in l_pcus]
 332
 333         checkAndRecordState(l_pcus, cohash)
 334
 335         return 0
 336
 337 import logging
 338 logger = logging.getLogger("monitor")
 339 logger.setLevel(logging.DEBUG)
 340 fh = logging.FileHandler("monitor.log", mode = 'a')
 341 fh.setLevel(logging.DEBUG)
 342 formatter = logging.Formatter('%(asctime)s %(levelname)s %(message)s')
 343 fh.setFormatter(formatter)
 344 logger.addHandler(fh)
 345
 346
 347 if __name__ == '__main__':
 348         try:
 349                 # NOTE: evidently, there is a bizarre interaction between iLO and ssh
 350                 # when LANG is set... Do not know why.  Unsetting LANG, fixes the problem.
 351                 if 'LANG' in os.environ:
 352                         del os.environ['LANG']
 353                 main()
 354                 time.sleep(1)
 355         except Exception, err:
 356                 print "Exception: %s" % err
 357                 print "Saving data... exitting."
 358                 soltesz.dbDump(config.dbname, externalState)
 359                 sys.exit(0)