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