policy.py

   1 #
   2 # Copyright (c) 2004  The Trustees of Princeton University (Trustees).
   3 #
   4 # Faiyaz Ahmed <faiyaza@cs.princeton.edu>
   5 #
   6 # $Id: $
   7 #
   8 # Policy Engine.
   9
  10 #from monitor import *
  11 from threading import *
  12 import time
  13 import logging
  14 import mailer
  15 import emailTxt
  16 import pickle
  17 import Queue
  18 import plc
  19 import reboot
  20 import config
  21
  22 DAT="./monitor.dat"
  23
  24 logger = logging.getLogger("monitor")
  25
  26 # Time to enforce policy
  27 POLSLEEP = 7200
  28
  29 # Where to email the summary
  30 SUMTO = "pupadm@lists.planet-lab.org"
  31 TECHEMAIL="tech-%s@sites.planet-lab.org"
  32 PIEMAIL="pi-%s@sites.planet-lab.org"
  33 SLICEMAIL="%s@slices.planet-lab.org"
  34 PLCEMAIL="support@planet-lab.org"
  35
  36 #Thresholds (DAYS)
  37 SPERDAY = 86400
  38 PITHRESH = 1 * SPERDAY
  39 SLICETHRESH = 5 * SPERDAY
  40 # Days before attempting rins again
  41 RINSTHRESH = 5 * SPERDAY
  42
  43 # Minimum number of nodes up before squeezing
  44 MINUP = 2
  45
  46 # IF:
  47 #  no SSH, down.
  48 #  bad disk, down
  49 #  DNS, kinda down (sick)
  50 #  clock, kinda down (sick)
  51 #  Full disk, going to be down
  52
  53 # Actions:
  54 #  Email
  55 #  suspend slice creation
  56 #  kill slices
  57 class Policy(Thread):
  58         def __init__(self, comonthread, sickNoTicket, emailed):
  59                 self.cmn = comonthread
  60                 # host - > (time of email, type of email)
  61                 self.emailed = emailed
  62                 # all sick nodes w/o tickets
  63                 self.sickNoTicket = sickNoTicket
  64                 # Sitess we've Squeezed.
  65                 self.squeezed = {}
  66                 Thread.__init__(self)
  67
  68
  69         '''
  70         What to do when node is in dbg (as reported by CoMon).
  71         '''
  72         def __actOnDebug(self, node):
  73                 # Check to see if we've done this before
  74                 if (node in self.emailed.keys()):
  75                         if (self.emailed[node][0] == "dbg"):
  76                                 delta = time.time() - self.emailed[node][1]
  77                                 if (delta <= RINSTHRESH ):
  78                                         # Don't mess with node if under Thresh.
  79                                         # Return, move on.
  80                                         logger.info("POLICY:  %s in dbg, but acted on %s days ago" % (node, delta // SPERDAY))
  81                                         return
  82                         logger.info("POLICY:  Node in dbg - " + node)
  83                         plc.nodeBootState(node, "rins")
  84                         # If it has a PCU
  85                         return reboot.reboot(node)
  86
  87         '''
  88         What to do when node is in dbg (as reported by CoMon).
  89         '''
  90         def __actOnFilerw(self, node):
  91                 target = [PLCEMAIL]
  92                 logger.info("POLICY:  Emailing PLC for " + node)
  93                 tmp = emailTxt.mailtxt.filerw
  94                 sbj = tmp[0] % {'hostname': node}
  95                 msg = tmp[1] % {'hostname': node}
  96                 mailer.email(sbj, msg, target)
  97                 self.emailed[node] = ("filerw", time.time())
  98
  99
 100         '''
 101         Acts on sick nodes.
 102         '''
 103         def actOnSick(self):
 104                 # Get list of nodes in debug from PLC
 105                 #dbgNodes = NodesDebug()
 106                 global TECHEMAIL, PIEMAIL
 107                 # Grab a node from the queue (pushed by rt thread).
 108                 node = self.sickNoTicket.get(block = True)
 109                 # Get the login base
 110                 loginbase = plc.siteId(node)
 111
 112                 # Send appropriate message for node if in appropriate bucket.
 113                 # If we know where to send a message
 114                 if not loginbase:
 115                         logger.info("POLICY:  loginbase for %s not found" %node)
 116                 # And we didn't email already.
 117                 else:
 118                         # If first email, send to Tech
 119                         target = [TECHEMAIL % loginbase]
 120
 121                         # If disk is foobarred, PLC should check it.
 122                         if (node in self.cmn.filerw) and \
 123                         (node not in self.emailed.keys()):
 124                                 self.__actOnFilerw(node)
 125                                 return
 126
 127                         # If in dbg, set to rins, then reboot.  Inform PLC.
 128                         if (node in self.cmn.dbg):
 129                         # If reboot failure via PCU, POD and send email
 130                         # if contacted PCU, return
 131                                 if self.__actOnDebug(node):  return
 132
 133                         if (node in self.emailed.keys()) and \
 134                         (node not in self.cmn.filerw)    and \
 135                         (node not in self.cmn.clock_drift):
 136                                 # If we emailed before, how long ago?
 137                                 delta = time.time() - self.emailed[node][1]
 138                                 if delta < SPERDAY:
 139                                         logger.info("POLICY:  already acted on %s today." % node)
 140                                         return
 141
 142                                 logger.info("POLICY:  acted %s on %s days ago" % (node,
 143                                 delta // SPERDAY))
 144
 145                                 # If more than PI thresh, but less than slicethresh
 146                                 if (delta >= PITHRESH) and (delta < SLICETHRESH):
 147                                         target.append(PIEMAIL % loginbase)
 148                                         #remove slice creation if enough nodes arent up
 149                                         if not self.enoughUp(loginbase):
 150                                                 logger.info("POLICY:  Removing slice creation from %s" % loginbase)
 151                                                 plc.removeSliceCreation(node)
 152                                                 self.squeezed[loginbase] = (time.time(), "creation")
 153                                 # If more than PI thresh and slicethresh
 154                                 if (delta >= PITHRESH) and (delta > SLICETHRESH):
 155                                         # Email slices at site.
 156                                         slices = plc.slices(loginbase)
 157                                         if len(slices) >= 1:
 158                                                 for slice in slices:
 159                                                         target.append(SLICEMAIL % slice)
 160                                                 if not self.enoughUp(loginbase):
 161                                                         plc.suspendSlices(node)
 162                                                         self.squeezed[loginbase] = (time.time(),
 163                                                                  "freeze")
 164
 165                         # Find the bucket the node is in and send appropriate email
 166                         # to approriate list of people.
 167                         for bkt in self.cmn.comonbkts.keys():
 168                                 if (node in getattr(self.cmn, bkt)):
 169                                         # Send predefined message for that bucket.
 170                                         logger.info("POLICY: Emailing (%s) %s - %s"\
 171                                                 %(bkt, node, target))
 172                                         tmp = getattr(emailTxt.mailtxt, bkt)
 173                                         sbj = tmp[0] % {'hostname': node}
 174                                         msg = tmp[1] % {'hostname': node}
 175                                         mailer.email(sbj, msg, target)
 176                                         self.emailed[node] = (bkt , time.time())
 177                                         return
 178
 179
 180         '''
 181         Prints, logs, and emails status of up nodes, down nodes, and buckets.
 182         '''
 183         def status(self):
 184                 sub = "Monitor Summary"
 185                 msg = "\nThe following nodes were acted upon:  \n\n"
 186                 for (node, (type, date)) in self.emailed.items():
 187                         # Print only things acted on today.
 188                         if (time.gmtime(time.time())[2] == time.gmtime(date)[2]):
 189                                 msg +="%s\t(%s)\t%s\n" %(node, type, time.ctime(date))
 190                 msg +="\n\nThe following sites have been 'squeezed':\n\n"
 191                 for (loginbase, (date, type)) in self.squeezed.items():
 192                         # Print only things acted on today.
 193                         if (time.gmtime(time.time())[2] == time.gmtime(date)[2]):
 194                                 msg +="%s\t(%s)\t%s\n" %(loginbase, type, time.ctime(date))
 195                 mailer.email(sub, msg, [SUMTO])
 196                 logger.info(msg)
 197                 return
 198
 199         '''
 200         Store/Load state of emails.  When, where, what.
 201         '''
 202         def emailedStore(self, action):
 203                 try:
 204                         if action == "LOAD":
 205                                 f = open(DAT, "r+")
 206                                 logger.info("POLICY:  Found and reading " + DAT)
 207                                 self.emailed.update(pickle.load(f))
 208                         if action == "WRITE":
 209                                 f = open(DAT, "w")
 210                                 #logger.debug("Writing " + DAT)
 211                                 pickle.dump(self.emailed, f)
 212                         f.close()
 213                 except Exception, err:
 214                         logger.info("POLICY:  Problem with DAT, %s" %err)
 215
 216         '''
 217         Returns True if more than MINUP nodes are up at a site.
 218         '''
 219         def enoughUp(self, loginbase):
 220                 allsitenodes = plc.getSiteNodes(loginbase)
 221                 if len(allsitenodes) == 0:
 222                         logger.info("Node not in db")
 223                         return
 224
 225                 numnodes = len(allsitenodes)
 226                 sicknodes = []
 227                 # Get all sick nodes from comon
 228                 for bucket in self.cmn.comonbkts.keys():
 229                         for host in getattr(self.cmn, bucket):
 230                                 sicknodes.append(host)
 231                 # Diff.
 232                 for node in allsitenodes:
 233                         if node in sicknodes:
 234                                 numnodes -= 1
 235
 236                 if numnodes < MINUP:
 237                         logger.info(\
 238 "POLICY:  site with %s has nodes %s up." %(loginbase, numnodes))
 239                         return False
 240                 else:
 241                         return True
 242
 243
 244
 245
 246         def run(self):
 247                 while 1:
 248                         self.actOnSick()
 249                         self.emailedStore("WRITE")
 250
 251
 252 def main():
 253         logger.setLevel(logging.DEBUG)
 254         ch = logging.StreamHandler()
 255         ch.setLevel(logging.DEBUG)
 256         formatter = logging.Formatter('%(message)s')
 257         ch.setFormatter(formatter)
 258         logger.addHandler(ch)
 259
 260         #print NodesDebug()
 261         #tmp = Queue.Queue()
 262         #a = Policy(None, tmp)
 263         #a.emailedStore("LOAD")
 264         #print a.emailed
 265
 266         print plc.slices(plc.siteId("alice.cs.princeton.edu"))
 267         os._exit(0)
 268 if __name__ == '__main__':
 269         import os
 270         import plc
 271         try:
 272                 main()
 273         except KeyboardInterrupt:
 274                 print "Killed.  Exitting."
 275                 logger.info('Monitor Killed')
 276                 os._exit(0)