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 import traceback
  12
  13 #old_handler = signal.getsignal(signal.SIGCHLD)
  14
  15 #def sig_handler(signum, stack):
  16 #       """ Handle SIGCHLD signal """
  17 #       global old_handler
  18 #       if signum == signal.SIGCHLD:
  19 #               try:
  20 #                       os.wait()
  21 #               except:
  22 #                       pass
  23 #       if old_handler != signal.SIG_DFL:
  24 #               old_handler(signum, stack)
  25 #
  26 #orig_sig_handler = signal.signal(signal.SIGCHLD, sig_handler)
  27
  28
  29 # QUERY all nodes.
  30 COMON_COTOPURL= "http://summer.cs.princeton.edu/status/tabulator.cgi?" + \
  31                                         "table=table_nodeview&" + \
  32                                     "dumpcols='name,resptime,sshstatus,uptime,lastcotop'&" + \
  33                                     "formatcsv"
  34                                     #"formatcsv&" + \
  35                                         #"select='lastcotop!=0'"
  36
  37 import threading
  38 plc_lock = threading.Lock()
  39 round = 1
  40 externalState = {'round': round, 'nodes': {'a': None}}
  41 errorState = {}
  42 count = 0
  43
  44 import reboot
  45 from reboot import pcu_name
  46
  47 import soltesz
  48 import plc
  49 import comon
  50 import threadpool
  51 import syncplcdb
  52
  53 def nmap_portstatus(status):
  54         ps = {}
  55         l_nmap = status.split()
  56         ports = l_nmap[4:]
  57
  58         continue_probe = False
  59         for port in ports:
  60                 results = port.split('/')
  61                 ps[results[0]] = results[1]
  62                 if results[1] == "open":
  63                         continue_probe = True
  64         return (ps, continue_probe)
  65
  66 def get_pcu(pcuname):
  67         plc_lock.acquire()
  68         try:
  69                 print "GetPCU from PLC %s" % pcuname
  70                 l_pcu  = plc.GetPCUs({'pcu_id' : pcuname})
  71                 print l_pcu
  72                 if len(l_pcu) > 0:
  73                         l_pcu = l_pcu[0]
  74         except:
  75                 try:
  76                         print "GetPCU from file %s" % pcuname
  77                         l_pcus = soltesz.dbLoad("pculist")
  78                         for i in l_pcus:
  79                                 if i['pcu_id'] == pcuname:
  80                                         l_pcu = i
  81                 except:
  82                         traceback.print_exc()
  83                         l_pcu = None
  84
  85         plc_lock.release()
  86         return l_pcu
  87
  88 def get_nodes(node_ids):
  89         plc_lock.acquire()
  90         l_node = []
  91         try:
  92                 l_node = plc.getNodes(node_ids, ['hostname', 'last_contact', 'node_id', 'ports'])
  93         except:
  94                 try:
  95                         plc_nodes = soltesz.dbLoad("l_plcnodes")
  96                         for n in plc_nodes:
  97                                 if n['node_id'] in node_ids:
  98                                         l_node.append(n)
  99                 except:
 100                         traceback.print_exc()
 101                         l_node = None
 102
 103         plc_lock.release()
 104         if l_node == []:
 105                 l_node = None
 106         return l_node
 107
 108
 109 def get_plc_pcu_values(pcuname):
 110         """
 111                 Try to contact PLC to get the PCU info.
 112                 If that fails, try a backup copy from the last run.
 113                 If that fails, return None
 114         """
 115         values = {}
 116
 117         l_pcu = get_pcu(pcuname)
 118
 119         if l_pcu is not None:
 120                 site_id = l_pcu['site_id']
 121                 node_ids = l_pcu['node_ids']
 122                 l_node = get_nodes(node_ids)
 123
 124                 if l_node is not None:
 125                         for node in l_node:
 126                                 values[node['hostname']] = node['ports'][0]
 127
 128                         values['nodenames'] = [node['hostname'] for node in l_node]
 129
 130                         # NOTE: this is for a dry run later. It doesn't matter which node.
 131                         values['node_id'] = l_node[0]['node_id']
 132
 133                 values.update(l_pcu)
 134         else:
 135                 values = None
 136
 137         return values
 138
 139 def get_plc_site_values(site_id):
 140         ### GET PLC SITE ######################
 141         plc_lock.acquire()
 142         values = {}
 143         d_site = None
 144
 145         try:
 146                 d_site = plc.getSites({'site_id': site_id}, ['max_slices', 'slice_ids', 'node_ids', 'login_base'])
 147                 if len(d_site) > 0:
 148                         d_site = d_site[0]
 149         except:
 150                 try:
 151                         plc_sites = soltesz.dbLoad("l_plcsites")
 152                         for site in plc_sites:
 153                                 if site['site_id'] == site_id:
 154                                         d_site = site
 155                                         break
 156                 except:
 157                         traceback.print_exc()
 158                         values = None
 159
 160         plc_lock.release()
 161
 162         if d_site is not None:
 163                 max_slices = d_site['max_slices']
 164                 num_slices = len(d_site['slice_ids'])
 165                 num_nodes = len(d_site['node_ids'])
 166                 loginbase = d_site['login_base']
 167                 values['plcsite'] = {'num_nodes' : num_nodes,
 168                                                         'max_slices' : max_slices,
 169                                                         'num_slices' : num_slices,
 170                                                         'login_base' : loginbase,
 171                                                         'status'     : 'SUCCESS'}
 172         else:
 173                 values = None
 174
 175
 176         return values
 177
 178
 179 def collectPingAndSSH(pcuname, cohash):
 180
 181         continue_probe = True
 182         errors = None
 183         values = {}
 184         ### GET PCU ######################
 185         try:
 186                 b_except = False
 187                 try:
 188                         v = get_plc_pcu_values(pcuname)
 189                         if v is not None:
 190                                 values.update(v)
 191                         else:
 192                                 continue_probe = False
 193                 except:
 194                         b_except = True
 195                         traceback.print_exc()
 196                         continue_probe = False
 197
 198                 if b_except or not continue_probe: return (None, None, None)
 199
 200                 if values['hostname'] is not None:
 201                         values['hostname'] = values['hostname'].strip()
 202
 203                 if values['ip'] is not None:
 204                         values['ip'] = values['ip'].strip()
 205
 206                 #### COMPLETE ENTRY   #######################
 207
 208                 values['complete_entry'] = []
 209                 #if values['protocol'] is None or values['protocol'] is "":
 210                 #       values['complete_entry'] += ["protocol"]
 211                 if values['model'] is None or values['model'] is "":
 212                         values['complete_entry'] += ["model"]
 213                         # Cannot continue due to this condition
 214                         continue_probe = False
 215
 216                 if values['password'] is None or values['password'] is "":
 217                         values['complete_entry'] += ["password"]
 218                         # Cannot continue due to this condition
 219                         continue_probe = False
 220
 221                 if len(values['complete_entry']) > 0:
 222                         continue_probe = False
 223
 224                 if values['hostname'] is None or values['hostname'] is "":
 225                         values['complete_entry'] += ["hostname"]
 226                 if values['ip'] is None or values['ip'] is "":
 227                         values['complete_entry'] += ["ip"]
 228
 229                 # If there are no nodes associated with this PCU, then we cannot continue.
 230                 if len(values['node_ids']) == 0:
 231                         continue_probe = False
 232                         values['complete_entry'] += ['NoNodeIds']
 233
 234                 #### DNS and IP MATCH #######################
 235                 if values['hostname'] is not None and values['hostname'] is not "" and \
 236                    values['ip'] is not None and values['ip'] is not "":
 237                         #print "Calling socket.gethostbyname(%s)" % values['hostname']
 238                         try:
 239                                 ipaddr = socket.gethostbyname(values['hostname'])
 240                                 if ipaddr == values['ip']:
 241                                         values['dnsmatch'] = "DNS-OK"
 242                                 else:
 243                                         values['dnsmatch'] = "DNS-MISMATCH"
 244                                         continue_probe = False
 245
 246                         except Exception, err:
 247                                 values['dnsmatch'] = "DNS-NOENTRY"
 248                                 values['hostname'] = values['ip']
 249                                 #print err
 250                 else:
 251                         if values['ip'] is not None and values['ip'] is not "":
 252                                 values['dnsmatch'] = "NOHOSTNAME"
 253                                 values['hostname'] = values['ip']
 254                         else:
 255                                 values['dnsmatch'] = "NO-DNS-OR-IP"
 256                                 values['hostname'] = "No_entry_in_DB"
 257                                 continue_probe = False
 258
 259                 #### RUN NMAP ###############################
 260                 if continue_probe:
 261                         nmap = soltesz.CMD()
 262                         (oval,eval) = nmap.run_noexcept("nmap -oG - -P0 -p22,23,80,443,5869,9100,16992 %s | grep Host:" % pcu_name(values))
 263                         # NOTE: an empty / error value for oval, will still work.
 264                         (values['portstatus'], continue_probe) = nmap_portstatus(oval)
 265                 else:
 266                         values['portstatus'] = None
 267
 268
 269                 ######  DRY RUN  ############################
 270                 if 'node_ids' in values and len(values['node_ids']) > 0:
 271                         rb_ret = reboot.reboot_test(values['nodenames'][0], values, continue_probe, 1, True)
 272                 else:
 273                         rb_ret = "Not_Run" # No nodes to test"
 274
 275                 values['reboot'] = rb_ret
 276
 277                 ### GET PLC SITE ######################
 278                 v = get_plc_site_values(values['site_id'])
 279                 if v is not None:
 280                         values.update(v)
 281                 else:
 282                         values['plcsite'] = {'status' : "GS_FAILED"}
 283
 284         except:
 285                 print "____________________________________"
 286                 print values
 287                 errors = values
 288                 print "____________________________________"
 289                 errors['traceback'] = traceback.format_exc()
 290                 print errors['traceback']
 291
 292         values['checked'] = time.time()
 293         return (pcuname, values, errors)
 294
 295 def recordPingAndSSH(request, result):
 296         global errorState
 297         global externalState
 298         global count
 299         (nodename, values, errors) = result
 300
 301         if values is not None:
 302                 global_round = externalState['round']
 303                 pcu_id = "id_%s" % nodename
 304                 externalState['nodes'][pcu_id]['values'] = values
 305                 externalState['nodes'][pcu_id]['round'] = global_round
 306
 307                 count += 1
 308                 print "%d %s %s" % (count, nodename, externalState['nodes'][pcu_id]['values'])
 309                 soltesz.dbDump(config.dbname, externalState)
 310
 311         if errors is not None:
 312                 pcu_id = "id_%s" % nodename
 313                 errorState[pcu_id] = errors
 314                 soltesz.dbDump("findbadpcu_errors", errorState)
 315
 316 # this will be called when an exception occurs within a thread
 317 def handle_exception(request, result):
 318         print "Exception occured in request %s" % request.requestID
 319         for i in result:
 320                 print "Result: %s" % i
 321
 322
 323 def checkAndRecordState(l_pcus, cohash):
 324         global externalState
 325         global count
 326         global_round = externalState['round']
 327
 328         tp = threadpool.ThreadPool(20)
 329
 330         # CREATE all the work requests
 331         for pcuname in l_pcus:
 332                 pcu_id = "id_%s" % pcuname
 333                 if pcuname not in externalState['nodes']:
 334                         #print type(externalState['nodes'])
 335
 336                         externalState['nodes'][pcu_id] = {'round': 0, 'values': []}
 337
 338                 node_round   = externalState['nodes'][pcu_id]['round']
 339                 if node_round < global_round:
 340                         # recreate node stats when refreshed
 341                         #print "%s" % nodename
 342                         req = threadpool.WorkRequest(collectPingAndSSH, [pcuname, cohash], {},
 343                                                                                  None, recordPingAndSSH, handle_exception)
 344                         tp.putRequest(req)
 345                 else:
 346                         # We just skip it, since it's "up to date"
 347                         count += 1
 348                         print "%d %s %s" % (count, pcu_id, externalState['nodes'][pcu_id]['values'])
 349                         pass
 350
 351         # WAIT while all the work requests are processed.
 352         while 1:
 353                 try:
 354                         time.sleep(1)
 355                         tp.poll()
 356                 except KeyboardInterrupt:
 357                         print "Interrupted!"
 358                         break
 359                 except threadpool.NoResultsPending:
 360                         print "All results collected."
 361                         break
 362
 363
 364
 365 def main():
 366         global externalState
 367
 368         l_pcus = soltesz.if_cached_else_refresh(1, config.refresh, "pculist", lambda : plc.GetPCUs())
 369         externalState = soltesz.if_cached_else(1, config.dbname, lambda : externalState)
 370         cohash = {}
 371
 372         if config.increment:
 373                 # update global round number to force refreshes across all nodes
 374                 externalState['round'] += 1
 375
 376         if config.filename == None and config.pcuid == None:
 377                 print "Calling API GetPCUs() : refresh(%s)" % config.refresh
 378                 l_pcus  = [pcu['pcu_id'] for pcu in l_pcus]
 379         elif config.filename is not None:
 380                 l_pcus = config.getListFromFile(config.filename)
 381                 l_pcus = [int(pcu) for pcu in l_pcus]
 382         elif config.pcuid is not None:
 383                 l_pcus = [ config.pcuid ]
 384                 l_pcus = [int(pcu) for pcu in l_pcus]
 385
 386         checkAndRecordState(l_pcus, cohash)
 387
 388         return 0
 389
 390
 391 if __name__ == '__main__':
 392         import logging
 393         logger = logging.getLogger("monitor")
 394         logger.setLevel(logging.DEBUG)
 395         fh = logging.FileHandler("monitor.log", mode = 'a')
 396         fh.setLevel(logging.DEBUG)
 397         formatter = logging.Formatter('%(asctime)s %(levelname)s %(message)s')
 398         fh.setFormatter(formatter)
 399         logger.addHandler(fh)
 400         from config import config
 401         from optparse import OptionParser
 402         parser = OptionParser()
 403         parser.set_defaults(filename=None,
 404                                                 increment=False,
 405                                                 pcuid=None,
 406                                                 dbname="findbadpcus",
 407                                                 cachenodes=False,
 408                                                 refresh=False,
 409                                                 )
 410         parser.add_option("-f", "--nodelist", dest="filename", metavar="FILE",
 411                                                 help="Provide the input file for the node list")
 412         parser.add_option("", "--pcuid", dest="pcuid", metavar="id",
 413                                                 help="Provide the id for a single pcu")
 414         parser.add_option("", "--cachenodes", action="store_true",
 415                                                 help="Cache node lookup from PLC")
 416         parser.add_option("", "--dbname", dest="dbname", metavar="FILE",
 417                                                 help="Specify the name of the database to which the information is saved")
 418         parser.add_option("", "--refresh", action="store_true", dest="refresh",
 419                                                 help="Refresh the cached values")
 420         parser.add_option("-i", "--increment", action="store_true", dest="increment",
 421                                                 help="Increment round number to force refresh or retry")
 422         config = config(parser)
 423         config.parse_args()
 424         try:
 425                 # NOTE: evidently, there is a bizarre interaction between iLO and ssh
 426                 # when LANG is set... Do not know why.  Unsetting LANG, fixes the problem.
 427                 if 'LANG' in os.environ:
 428                         del os.environ['LANG']
 429                 main()
 430                 time.sleep(1)
 431         except Exception, err:
 432                 traceback.print_exc()
 433                 print "Exception: %s" % err
 434                 print "Saving data... exitting."
 435                 soltesz.dbDump(config.dbname, externalState)
 436                 sys.exit(0)