xcp-xapi 1.3.2-5 / usr / lib / xcp / bin / perfmon

#!/usr/bin/env python
#
# perfmon - a daemon for monitoring performance of the host on which it is run 
# and of all the local VMs, and for generating events based on configurable 
# triggers
#
# Notes:
# ======
# The XAPI instance running on localhost monitors a number of variables
# for each VM running locally (i.e not on other pool members) and 
# for the host itself. Each variable is stored in 16 RRDs (Round Robin Databases).
#
#  Consolidation   Number of samples in RRD
#  function        5s/sample  1m/sample 1hr/sample 1day/sample
#  AVERAGE         120 (10m)  120 (2h)  ?          ?
#  MIN             120 (10m)  120 (2h)  ?          ?
#  MAX             120 (10m)  120 (2h)  ?          ?
#  LAST            120 (10m)  120 (2h)  ?          ?
#
# The "Consolidation function" tells how that RRD is built up from the 
# one with the next highest sample rate.  E.g. In the 1m/sample "AVERAGE" RRD
# each sample is the average of 12 from the 1s/sample "AVERAGE" RRD, whereas
# in the 1m/sample "MIN" RRD each sample is the minimum of 12 from the 1s/sample
# "AVERAGE" RRD.
#
# When XAPI is queried over http it selects the column (e.g. "1hr/sample")
# based on the "start" CGI param.  It will return the highest level of granularity
# available for the period requested.
#
# The "cf" CGI param specfies the row.  (All rows are returned if it's missing.)

import sys
import os
import getopt
import traceback
import XenAPI
import urllib
from xml import sax            # used to parse rrd_updates because this may be large and sax is more efficient
from xml.dom import minidom    # used to parse other-config:perfmon.  Efficiency is less important than reliability here
from xml.parsers.expat import ExpatError
import time
import re
import random
import syslog
import socket
import gc
import signal
import commands

def print_debug(string):
    if debug: 
        print >>sys.stderr, "DEBUG:", string
        syslog.syslog(syslog.LOG_USER | syslog.LOG_INFO, "PERFMON(DEBUG): %s" % string)

def log_err(string):
    print >>sys.stderr, string
    syslog.syslog(syslog.LOG_USER | syslog.LOG_ERR, "PERFMON: %s" % string)
    pass

def log_info(string):
    print >>sys.stderr, string
    syslog.syslog(syslog.LOG_INFO | syslog.LOG_INFO, "PERFMON: %s" % string)
    pass

def debug_mem():
    objCount = {}
    gc.collect()
    objList = gc.get_objects()
    for obj in objList:
        if getattr(obj, "__class__", None):
            name = obj.__class__.__name__
        else:
            name = type(obj)
        if objCount.has_key(name):
            objCount[name] += 1
        else:
            objCount[name] = 1
        
        output = []
        for name in objCount:
            output.append("%s :%s" % (name, objCount[name]))
    log_info("\n".join(output))

class PerfMonException(Exception):
    pass

class XmlConfigException(PerfMonException):
    pass

class UsageException(Exception):
    pass


# Start a session with the master of a pool.
# Note: when calling http://localhost/rrd_update we must pass the session
# ID as a param.  The host then uses this to verify our validity with 
# the master before responding.
# If the verification fails we should get a 401 response
class XapiSession(XenAPI.Session):
    """ Object that represents a XenAPI session with the pool master
    One of these is needed to refresh a VMMonitor or HOSTMonitor config, or 
    to refresh an RRDUpdates object
    """
    def __init__(self):
        XenAPI.Session.__init__(self, "http://_var_xapi_xapi", transport=XenAPI.UDSTransport())
        self.xenapi.login_with_password("", "")
    def __del__ (self):
        self.xenapi.session.logout()
    def id(self):
        return self._session

class ObjectReport:
    def __init__(self, objtype, uuid):
        self.objtype = objtype # a string like "vm", or "host" taken from an <entry> tag 
        self.uuid = uuid       # the object's uuid
        self.vars = {}         # maps rrd variable name to array of floats
    def get_uuid(self):
        return self.uuid
    def get_var_names(self):
        return self.vars.keys()
    def get_value(self, var_name, row):
        try:
            return (self.vars[var_name])[row]
        except:
            return 0.0
    def insert_value(self, var_name, index, value):
        if not self.vars.has_key(var_name):
            self.vars[var_name] = []
        self.vars[var_name].insert(index, value)

class RRDReport:
    "This is just a data structure passed that is completed by RRDContentHandler"
    def __init__(self):
        self.reset()
        
    def reset(self):
        self.columns     = 0      # num xapi vars in xml 
        self.rows        = 0      # num samples in xml
        self.start_time  = 0      # timestamp of 1st sample in xml
        self.end_time    = 0      # timestamp of last sample in xml
        self.step_time   = 0      # seconds between each pair of samples
        self.obj_reports  = {}     # maps uuids to ObjectReports, built from xml

class RRDColumn:
    "class used internally by RRDContentHandler"
    def __init__(self, paramname, obj_report):
        self.paramname = paramname
        self.obj_report = obj_report

class RRDContentHandler(sax.ContentHandler):
    """ Handles data in this format:
    <xport>
      <meta>
       <start>INTEGER</start>
       <step>INTEGER</step>
       <end>INTEGER</end>
       <rows>INTEGER</rows>
       <columns>INTEGER</columns>
       <legend>
        <entry>IGNOREME:(host|vm):UUID:PARAMNAME</entry>
        ... another COLUMNS-1 entries ...
       </legend>
      </meta>
      <data>
       <row>
        <t>INTEGER(END_TIME)</t>
        <v>FLOAT</v>
        ... another COLUMNS-1 values ...
       </row>
       ... another ROWS-2 rows
       <row>
        <t>INTEGER(START_TIME)</t>
        <v>FLOAT</v>
        ... another COLUMNS-1 values ...
       </row>
      </data>
    </xport>
    """
    def __init__(self, report):
        "report is saved and later updated by this object. report should contain defaults already"
        self.report = report

        self.in_start_tag = False
        self.in_step_tag = False
        self.in_end_tag = False
        self.in_rows_tag = False
        self.in_columns_tag = False
        self.in_entry_tag = False
        self.in_row_tag = False
        self.column_details = []
        self.row = 0
        
    def startElement(self, name, attrs):
        self.raw_text = ""
        if name == 'start':
            self.in_start_tag = True
        elif name == 'step':
            self.in_step_tag = True
        elif name == 'end':
            self.in_end_tag = True
        elif name == 'rows':
            self.in_rows_tag = True
        elif name == 'columns':
            self.in_columns_tag = True
        elif name == 'entry':
            self.in_entry_tag = True
        elif name == 'row':
            self.in_row_tag = True
            self.col = 0

        if self.in_row_tag:
            if name == 't':
                self.in_t_tag = True
            elif name == 'v':
                self.in_v_tag = True
            
    def characters(self, chars):
        if (self.in_start_tag or 
            self.in_step_tag or 
            self.in_end_tag or 
            self.in_rows_tag or 
            self.in_columns_tag or
            self.in_entry_tag or 
            #self.in_row_tag # ignore text under row tag, <row>s are just for holding <t> and <v> nodes
            self.in_t_tag or
            self.in_v_tag):
            self.raw_text += chars
            
    def endElement(self, name):
        if name == 'start':
            # This overwritten later if there are any rows
            self.report.start_time = int(self.raw_text)
            self.in_start_tag = False
        elif name == 'step':
            self.report.step_time = int(self.raw_text)
            self.in_step_tag = False
        elif name == 'end':
            # This overwritten later if there are any rows
            self.report.end_time = int(self.raw_text)
            self.in_end_tag = False
        elif name == 'rows':
            self.report.rows = int(self.raw_text)
            self.in_rows_tag = False
        elif name == 'columns':
            self.report.columns = int(self.raw_text)
            self.in_columns_tag = False
        elif name == 'entry':
            (_, objtype, uuid, paramname) = self.raw_text.split(':')
            # lookup the obj_report corresponding to this uuid, or create if it does not exist
            if not self.report.obj_reports.has_key(uuid):
                self.report.obj_reports[uuid] = ObjectReport(objtype, uuid)
            obj_report = self.report.obj_reports[uuid]

            # save the details of this column
            self.column_details.append(RRDColumn(paramname, obj_report))
            self.in_entry_tag = False
        elif name == 'row':
            self.in_row_tag = False
            self.row += 1
        elif name == 't':
            # Extract start and end time from row data as it's more reliable than the values in the meta data
            t = int(self.raw_text)
            # Last row corresponds to start time
            self.report.start_time = t
            if self.row == 0:
                # First row corresponds to end time
                self.report.end_time = t

            self.in_t_tag = False

        elif name == 'v':
            v = float(self.raw_text)

            # Find object report and paramname for this col
            col_details = self.column_details[self.col]
            obj_report = col_details.obj_report
            paramname = col_details.paramname
            
            # Update object_report
            obj_report.insert_value(paramname, index=0, value=v) # use index=0 as this is the earliest sample so far
            
            # Update position in row
            self.col += 1
            
            self.in_t_tag = False
                
        

# An object of this class should persist the lifetime of the program
class RRDUpdates:
    """ Object used to get and parse the output the http://localhost/rrd_udpates?... 
    """
    def __init__(self):
        # params are what get passed to the CGI executable in the URL
        self.params = dict()
        self.params['start'] = int(time.time()) - interval # interval seconds ago
        self.params['host'] = 'true'   # include data for host (as well as for VMs)
        self.params['cf'] = 'AVERAGE'  # consolidation function, each sample averages 12 from the 5 second RRD
        self.params['interval'] = str(rrd_step)
        self.report = RRDReport()      # data structure updated by RRDContentHandler

    def __repr__(self):
        return '<RRDUpdates object: params=%s>' % str(self.params)
 
    def refresh(self, session, override_params = {}):
        "reread the rrd_updates over CGI and parse"
        params = override_params
        params['session_id'] = session.id()
        params.update(self.params)
        paramstr = "&".join(["%s=%s"  % (k,params[k]) for k in params])
        print_debug("Calling http://localhost/rrd_updates?%s" % paramstr)

        # this is better than urllib.urlopen() as it raises an Exception on http 401 'Unauthorised' error 
        # rather than drop into interactive mode
        sock = urllib.URLopener().open("http://localhost/rrd_updates?%s" % paramstr)
        xmlsource = sock.read()
        sock.close()

        # Use sax rather than minidom and save Vvvast amounts of time and memory.
        self.report.reset()
        sax.parseString(xmlsource, RRDContentHandler(self.report))

        # Update the time used on the next run
        self.params['start'] = self.report.end_time + 1 # avoid retrieving same data twice

        print_debug("Refreshed rrd_updates, start = %d, end = %d, rows = %d" % \
                        (self.report.start_time, self.report.end_time, self.report.rows))

    def get_num_rows(self):
        "Return the number of samples of each parameter"
        return self.report.rows

    def get_obj_report_by_uuid(self, uuid):
        "Return an ObjectReport for the object with this uuid"
        try:
            return self.report.obj_reports[uuid]
        except:
            return None
        
    def get_uuid_list_by_objtype(self, objtype):
        "Return a list of uuids corresonding to the objects of this type for which we have ObjectReports"
        return [ objrep.uuid 
                   for objrep in self.report.obj_reports.values() 
                     if objrep.objtype == objtype ]


# Consolidation functions:
supported_consolidation_functions = [ 'sum', 'average', 'max', 'get_percent_fs_usage' ]

def average(mylist):
    return sum(mylist)/float(len(mylist))

def get_percent_fs_usage(ignored):
    "Get the percent usage of the host filesystem.  Input list is ignored and should be empty"
    # this file is on the filesystem of interest in both OEM and Retail
    output = commands.getoutput('df /etc/passwd')
    output = ' '.join(output.splitlines()[1:]) # remove header line and rewrap on single line
    percentage = output.split()[4]
    # strip of % character and convert to float
    return float(percentage[0:-1])/100.0

class VariableConfig:
    """Object storing the configuration of a Variable
    
    Initialisation parameters:
             xmldoc = dom object representing the <variable> nodes in the ObjectMonitor config strings.
                      See VMMonitor.__doc__ and HOSTMonitor.__doc__
             alarm_create_callback = 
                      callback called by Variable.update() to create and send an alarm
             get_default_variable_config = 
                      a function that VariableConfig.__init__() uses to lookup default tag values
                      by variable name
    """
    def __init__(self, xmldoc, alarm_create_callback, get_default_variable_config):
        try: name = xmldoc.getElementsByTagName('name')[0].getAttribute('value')
        except IndexError: raise XmlConfigException, "variable missing 'name' tag" 
        def get_value(tag): 
            try:
                return xmldoc.getElementsByTagName(tag)[0].getAttribute('value')
            except:
                return get_default_variable_config(name, tag)
        rrd_regex = get_value('rrd_regex')
        consolidation_fn = get_value('consolidation_fn')
        alarm_trigger_level = get_value('alarm_trigger_level')
        alarm_trigger_period = get_value('alarm_trigger_period')
        alarm_auto_inhibit_period = get_value('alarm_auto_inhibit_period')
        alarm_trigger_sense = get_value('alarm_trigger_sense')
        alarm_priority = get_value('alarm_priority')

        # Save xmldoc: we need this when creating the body of the alarms
        self.xmldoc = xmldoc

        self.name = name
        try:
            self.rrd_regex = re.compile(rrd_regex)
        except:
            raise XmlConfigException, "variable %s: regex %s does not compile" % (name, rrd_regex)

        if consolidation_fn not in supported_consolidation_functions:
            raise XmlConfigException, "variable %s: consolidation function %s not supported" \
                % (name, consolidation_fn)
        self.consolidation_fn = eval(consolidation_fn)

        try:
            self.alarm_trigger_period = int(alarm_trigger_period)
        except:
            raise XmlConfigException, "variable %s: alarm_trigger_level %s not an int" % \
                (name, alarm_trigger_period)

        try:
            self.alarm_auto_inhibit_period = int(alarm_auto_inhibit_period)
        except:
            raise XmlConfigException, "variable %s: alarm_auto_inhibit_period %s not an int" % \
                (name, alarm_auto_inhibit_period)
        try:
            trigger_level = float(alarm_trigger_level)
        except:
            raise XmlConfigException, "variable %s: alarm_trigger_level %s not a float" % \
                (name, alarm_trigger_level)

        self.alarm_priority = alarm_priority
            
        if alarm_trigger_sense == "high":
            self.test_level = lambda : (self.value > trigger_level)
        else:
            self.test_level = lambda : (self.value < trigger_level)
        self.alarm_create_callback = alarm_create_callback

def variable_configs_differ(vc1, vc2):
    "Say whether configuration of one variable differs from that of another"
    return vc1.xmldoc.toxml() != vc2.xmldoc.toxml()

class VariableState:
    """ Object storing the state of a Variable
    """
    def __init__(self):
        self.value = None
        self.timeof_last_alarm = time.time() - self.alarm_auto_inhibit_period
        self.trigger_down_counter = self.alarm_trigger_period

class Variable(VariableConfig, VariableState):
    """ Variable() is used by ObjectMonitor to create one Variable object for each
    variable specified in it's config string
    """
    def __init__(self, *args):
        VariableConfig.__init__(self, *args)
        VariableState.__init__(self)
        print_debug("Created Variable %s" % self.name)

    def __generate_alarm(self, session):
        """ Generate an alarm using callback provided by creator
        
        ... provided that one has not been generated in the last 
        self.alarm_auto_inhibit_period seconds
        """
        t = time.time()
        delta = t - self.timeof_last_alarm
        if delta < self.alarm_auto_inhibit_period:
            return # we are in the auto inhibit period - do nothing
        self.timeof_last_alarm = t
        message = "value: %f\nconfig:\n%s" % (self.value, self.xmldoc.toprettyxml())
        
        self.alarm_create_callback(self, session, message)

    def update(self, value, session):
        """Update the value of the variable using an RRDUpdates object
    
        Calls self.__generate_alarm() if level has been 'bad' for more than
        self.alarm_trigger_period seconds
        """
        self.value = value
        print_debug("Variable %s set to %f" % (self.name, value))
        if self.test_level():
            # level is bad
            self.trigger_down_counter -= rrd_step
            if self.trigger_down_counter <= 0:
                self.__generate_alarm(session)
                # reset trigger counter
                self.trigger_down_counter = self.alarm_trigger_period
        else:
            # level good - reset trigger counter
            self.trigger_down_counter = self.alarm_trigger_period
        

class ObjectMonitor:
    """Abstract class, used as base for VMMonitor and HOSTMonitor
    
    Public attributes are uuid, refresh_config()
    Inherited classes must implement a public attribute process_rrd_updates()
    """
    def __init__(self, uuid):
        self.uuid = uuid 
        self.xmlconfig = None
        # "variables" is the public attribute of interest
        self.variables = []
        self.refresh_config()
        
    def refresh_config(self):
        if self.__update_xmlconfig():
            # config has changed - reparse it
            try:
                self.__parse_xmlconfig()
            except XmlConfigException, e:
                errmsg = "\n".join([ str(x) for x in e.args ])
                log_err("%s %s config error: %s" % (self.monitortype, self.uuid, errmsg))
            except ExpatError, e:
                errmsg = "\n".join([ str(x) for x in e.args ])
                log_err("%s %s XML parse error: %s" % (self.monitortype, self.uuid, errmsg))

    def __update_xmlconfig(self):
        if not all_xmlconfigs.has_key(self.uuid):
            xmlconfig = None
        else:
            xmlconfig = all_xmlconfigs[self.uuid]
        
        if xmlconfig != self.xmlconfig:
            self.xmlconfig = xmlconfig
            return True
            
    def __parse_xmlconfig(self):
        if not self.xmlconfig:
            # Possible if this VM/host is not configured yet
            self.variables = []
            return
        xmldoc = minidom.parseString(self.xmlconfig)
        variable_nodes = xmldoc.getElementsByTagName('variable')
        variable_names = []
        
        for vn in variable_nodes:
            # create a variable using the config in vn
            var = Variable(vn, self.alarm_create, self.get_default_variable_config)

            # Update list of variable names
            if var.name not in variable_names:
                variable_names.append(var.name)

            # build list of variables already present with same name
            vars_with_same_name = [ v for v in self.variables if v.name == var.name ]
            count = 0
            append_var = True
            for v in vars_with_same_name:
                # this list should be 0 or 1 long!
                if count > 0: 
                    log_err("programmer error: found duplicate variable %s (uuid %s)" % (var.name, self.uuid))
                    self.variables.remove(v)
                    continue
                count += 1

                # only replace variable in self.variables if its config has changed.
                # This way we don't reset its state
                if variable_configs_differ(var, v):
                    self.variables.remove(v)
                else:
                    append_var = False
            
            if append_var:
                print_debug("Appending %s to list of variables for UUID=%s" % (var.name, self.uuid))
                self.variables.append(var)

        # Now delete any old variables that do not appear in the new variable_nodes
        variables_to_remove = [ v for v in self.variables if v.name not in variable_names ]
        for v in variables_to_remove:
            print_debug("Deleting %s from list of variables for UUID=%s" % (v.name, self.uuid))
            self.variables.remove(v)
        

    def process_rrd_updates(self, rrd_updates, session):
        print_debug("%sMonitor processing rrd_updates for %s" % (self.monitortype, self.uuid))
        obj_report = rrd_updates.get_obj_report_by_uuid(self.uuid)
        num_rows   = rrd_updates.get_num_rows()
        if not obj_report:
            return
        params_in_obj_report = obj_report.get_var_names()
        for var in self.variables:
            # find the subset of the params returned for this object that we need to consolidate into var
            params_to_consolidate = filter(var.rrd_regex.search, params_in_obj_report)
            for row in range(num_rows):
                # Get the values to consolidate
                values_to_consolidate = map(lambda param: obj_report.get_value(param, row), params_to_consolidate)
                # Consolidate them
                value = var.consolidation_fn(values_to_consolidate)
                # Pass result on to the variable object - this may result in an alarm being generated
                var.update(value, session)

    def alarm_create(self, var, session, message):
        "Callback used by Variable var to actually send an alarm"
        print_debug("Creating an alarm for %s %s, message: %s" % (self.monitortype, self.uuid, message))
        session.xenapi.message.create("ALARM", var.alarm_priority, self.monitortype, self.uuid, message)
            
class VMMonitor(ObjectMonitor):
    """Object that maintains state of one VM
    
    Configured by writing an xml string into an other-config key, e.g.
    xe vm-param-set uuid=$vmuuid other-config:perfmon=\
       '<config><variable><name value="cpu_usage"/><alarm_trigger_level value="0.5"/></variable></config>'

    Notes:
     - Multiple <variable> nodes allowed
     - full list of child nodes is
       * name: what to call the variable (no default)
       * alarm_priority: the priority of the messages generated (default '5')
       * alarm_trigger_level: level of value that triggers an alarm (no default)
       * alarm_trigger_sense: 'high' if alarm_trigger_level is a max, otherwise 'low'. (default 'high')
       * alarm_trigger_period: num seconds of 'bad' values before an alarm is sent (default '60')
       * alarm_auto_inhibit_period: num seconds this alarm disabled after an alarm is sent (default '3600')
       * consolidation_fn: how to combine variables from rrd_updates into one value
         (default is 'average' for 'cpu_usage' & 'sum' for everything else)
       * rrd_regex matches the names of variables from (xe vm-data-sources-list uuid=$vmuuid) used to compute value
         (only has defaults for "cpu_usage", "network_usage", and "disk_usage")
    """
    def __init__(self, *args):
        self.monitortype = "VM"
        ObjectMonitor.__init__(self, *args)
        print_debug("Created VMMonitor with uuid %s" % self.uuid)

    def get_default_variable_config(self, variable_name, config_tag):
        "This allows user to not specify full set of tags for each variable in xml config"
        if config_tag == 'consolidation_fn':
            if variable_name == "cpu_usage": return 'average'
            elif variable_name == "fs_usage": return 'get_percent_fs_usage'
            else: return 'sum'
        elif config_tag == 'rrd_regex':
            if variable_name == "cpu_usage": return "cpu[0-9]+"
            elif variable_name == "network_usage": return "vif_[0-9]+_[rt]x"
            elif variable_name == "disk_usage": return "vbd_(xvd|hd)[a-z]+_(read|write)"
            elif variable_name == "fs_usage": return "__DUMMY__"      # match nothing
            else: raise XmlConfigException, "variable %s: no default rrd_regex - please specify one" % variable_name
        elif config_tag == 'alarm_trigger_period':      return '60'   # 1 minute
        elif config_tag == 'alarm_auto_inhibit_period': 
            if variable_name == "fs_usage": return '604800'           # 1 week
            else: return '3600'                                       # 1 hour
        elif config_tag == 'alarm_trigger_level': 
            if variable_name == "fs_usage": return '0.9'               # trigger when 90% full
            else: raise XmlConfigException, "variable %s: no default alarm_trigger_level - please specify one" % variable_name
        elif config_tag == 'alarm_trigger_sense':       return 'high' # trigger if *above*
        elif config_tag == 'alarm_priority':             return '5'    # the minimum priority required for mail-alarm to send
        else: raise XmlConfigException, "variable %s: no default available for tag %s" % (variable_name, config_tag)

class HOSTMonitor(ObjectMonitor):
    """Object that maintains state of one Host
    
    Configured by writing an xml string into an other-config key, e.g.
    xe host-param-set uuid=$hostuuid other-config:perfmon=\
       '<config><variable><name value="cpu_usage"/><alarm_trigger_level value="0.5"/></variable></config>'

    Notes:
     - Multiple <variable> nodes allowed
     - full list of child nodes is
       * name: what to call the variable (no default)
       * alarm_priority: the priority of the messages generated (default '5')
       * alarm_trigger_level: level of value that triggers an alarm (no default)
       * alarm_trigger_sense: 'high' if alarm_trigger_level is a max, otherwise 'low'. (default 'high')
       * alarm_trigger_period: num seconds of 'bad' values before an alarm is sent (default '60')
       * alarm_auto_inhibit_period: num seconds this alarm disabled after an alarm is sent (default '3600')
       * consolidation_fn: how to combine variables from rrd_updates into one value
         (default is 'average' for 'cpu_usage' & 'sum' for everything else)
       * rrd_regex matches the names of variables from (xe host-data-sources-list uuid=$hostuuid) used to compute value
         (only has defaults for "cpu_usage" and "network_usage")
    """
    def __init__(self, *args):
        self.monitortype = "HOST"
        ObjectMonitor.__init__(self, *args)
        print_debug("Created HOSTMonitor with uuid %s" % self.uuid)

    def get_default_variable_config(self, variable_name, config_tag):
        "This allows user to not specify full set of tags for each variable in xml config"
        if config_tag == 'consolidation_fn':
            if variable_name == "cpu_usage": return 'average'
            else: return 'sum'
        elif config_tag == 'rrd_regex':
            if variable_name == "cpu_usage": return "cpu[0-9]+"
            elif variable_name == "network_usage": return "pif_eth[0-9]+_[rt]x"
            else: raise XmlConfigException, "variable %s: no default rrd_regex - please specify one" % variable_name
        elif config_tag == 'alarm_trigger_period':      return '60'   # 1 minute
        elif config_tag == 'alarm_auto_inhibit_period': return '3600' # 1 hour
        elif config_tag == 'alarm_trigger_sense':       return 'high' # trigger if *above*
        elif config_tag == 'alarm_priority':             return '5'    # the minimum priority required for mail-alarm to send
        else: raise XmlConfigException, "variable %s: no default available for tag %s" % (variable_name, config_tag)

all_xmlconfigs = {}
def update_all_xmlconfigs(session):
    """Update all_xmlconfigs, a global dictionary that maps any uuid
    (host or VM) to the xml config string in other-config:perfmon keys"""
    global all_xmlconfigs
    
    all_host_recs = session.xenapi.host.get_all_records()
    all_vm_recs   = session.xenapi.VM.get_all_records()
    
    # build dictionary mapping uuids to other_configs
    all_otherconfigs = {}
    all_otherconfigs.update([
            (all_host_recs[ref]['uuid'], all_host_recs[ref]['other_config'])
            for ref in all_host_recs.keys()
            ])
    all_otherconfigs.update([
            (all_vm_recs[ref]['uuid'], all_vm_recs[ref]['other_config'])
            for ref in all_vm_recs.keys()
            ])
    
    # rebuild dictionary mapping uuids to xmlconfigs
    all_xmlconfigs.clear()
    all_xmlconfigs.update([
            (uuid, other_config['perfmon'])
            for (uuid, other_config) in all_otherconfigs.items()
            if other_config.has_key('perfmon')
            ])

# 5 minute default interval
interval = 300
interval_percent_dither = 5
rrd_step = 60
debug = False

# rate to call update_all_xmlconfigs()
config_update_period = 1800

cmdsockname = "\0perfmon"  # an af_unix socket name (the "\0" stops socket.bind() creating a fs node)
cmdmaxlen = 256

def main():
    global interval
    global interval_percent_dither
    global rrd_step
    global debug
    global config_update_period
    maxruns=None
    try:
        argv = sys.argv[1:]
        opts, args = getopt.getopt(argv, "i:n:ds:c:D:",
            ["interval=", "numloops=","debug","rrdstep=","config_update_period=","interval_percent_dither="])
    except getopt.GetoptError:
        raise UsageException
    
    configfname = None
    for opt, arg in opts:
        if opt == '-i' or opt == '--interval':
            interval = int(arg)
        elif opt == '-n' or opt == '--numloops':
            maxruns = int(arg)
        elif opt == '-d' or opt == '--debug':
            debug = True
        elif opt == '-s' or opt == '--rrdstep':
            rrd_step = int(arg)
            if rrd_step != 5 and rrd_step != 60:
                raise UsageException
        elif opt == '-c' or opt == '--config_update_period':
            config_update_period = int(arg)
        elif opt == '-D' or opt == '--interval_percent_dither':
            interval_percent_dither = int(arg)
        else:
            raise UsageException

    # open the cmd socket (over which we listen for commands such as "refresh")
    cmdsock = socket.socket(socket.AF_UNIX, socket.SOCK_DGRAM)
    cmdsock.bind(cmdsockname)


    # The dither on each loop (prevents stampede on master)
    rand = random.Random().uniform
    dither = (interval * interval_percent_dither)/100.0

    # Create a XAPI session on first run
    restart_session = True

    # Create a client for getting the rrd_updates over HTTP
    rrd_updates = RRDUpdates() 

    # Work out when next to update all the xmlconfigs for all the 
    # hosts and all the VMs.  This causes a lot of data to be retrieved
    # from the master, so we only do it once every config_update_period
    # and we cache the results
    next_config_update = time.time()
    
    # monitors for vms running on this host.
    # This dictionary uses uuids to lookup each monitor object
    vm_mon_lookup = {}
    
    # The monitor for the host
    host_mon = None
    
    runs = 0
    while True:
        print_debug("Run: %d" % runs)

        # Get new updates - and catch any http errors
        try:
            # if session has failed on last run we need to restart it
            if restart_session:
                session = XapiSession()
                restart_session = False

            rrd_updates.refresh(session)
        
            # Should we update all_xmlconfigs
            if time.time() >= next_config_update:
                print_debug("Updating all_xmlconfigs")
                # yes - update all the xml configs: this generates a few LARGE xapi messages from the master
                update_all_xmlconfigs(session)
                
                # Set time when to do this next
                next_config_update = time.time() + config_update_period

            # List of VMs present in rrd_updates
            vm_uuid_list = rrd_updates.get_uuid_list_by_objtype('vm')

            # Remove any monitors for VMs no longer listed in rrd_updates page
            for uuid in vm_mon_lookup.keys():
                if uuid not in vm_uuid_list:
                    vm_mon_lookup.pop(uuid)
                
            # Create monitors for VMs that have just appeared in rrd_updates page
            for uuid in vm_uuid_list:
                if uuid not in vm_mon_lookup.keys():
                    vm_mon_lookup[uuid] = VMMonitor(uuid)
                else:
                    # check if the config has changed, e.g. by XenCenter
                    vm_mon_lookup[uuid].refresh_config()
            
            # Remove monitor for the host if it's no longer listed in rrd_updates page
            # Create monitor for the host if it has just appeared in rrd_updates page
            try:
                host_uuid = rrd_updates.get_uuid_list_by_objtype('host')[0] # should only ever be one of these
            except:
                # list may be empty!
                host_uuid = None

            if not host_uuid:
                host_mon = None
            elif not host_mon:
                host_mon = HOSTMonitor(host_uuid)
            elif host_mon.uuid != host_uuid:
                raise PerfMonException, "host uuid in rrd_updates changed (old: %s, new %s)" % \
                    (host_mon.uuid, host_uuid)
            else:
                # check if the config has changed, e.g. by XenCenter
                host_mon.refresh_config()
            
            # Go through each vm_mon and update it using the rrd_udpates - this may generate alarms
            for vm_mon in vm_mon_lookup.values():
                vm_mon.process_rrd_updates(rrd_updates, session)

            # Ditto for the host_mon
            if host_mon:
                host_mon.process_rrd_updates(rrd_updates, session)

        except IOError, e:
            if e.args[0] == 'http error' and e.args[1] in (401, 500):
                # Error getting rrd_updates: 401=Unauthorised, 500=Internal - start new session
                pass 
            elif e.args[0] == 'socket error':
                # This happens if we send messages or read other-config:perfmon after xapi is restarted
                pass
            else:
                # Don't know why we got this error - crash, die and look at logs later
                raise

            log_err("caught IOError: (%s) - restarting XAPI session" % " ".join([str(x) for x in e.args]))
            restart_session = True

        except socket.error, e:
            if e.args[0] == 111:
                # "Connection refused" - this happens when we try to restart session and *that* fails
                pass

            log_err("caught socket.error: (%s) - restarting XAPI session" % " ".join([str(x) for x in e.args]))
            restart_session = True

        runs += 1
        if maxruns is not None and runs >= maxruns:
            break

        # Force collection of cyclically referenced objects cos we don't
        # trust GC to do it on its own
        gc.collect()

        # Sleep for interval +/- dither, exiting early if we recv a cmd
        timeout = rand(interval - dither, interval + dither)
        cmdsock.settimeout(timeout)
        try:
            cmd = cmdsock.recv(cmdmaxlen)
        except socket.timeout:
            pass
        else:
            if cmd == "refresh":
                # This forces a re-read of all the configs on the next loop
                next_config_update = time.time()
            elif cmd == "debug_mem":
                debug_mem()
            else:
                log_err("received unhandled command %s" % cmd)

        # continue to next run

    return 0

def sigterm_handler(sig, stack_frame):
    log_err("Caught signal %d - exiting" % sig)
    sys.exit(1)

pidfile = "/var/run/perfmon.pid"

if __name__ == "__main__":
    
    # setup signal handler to print out notice when killed
    signal.signal(signal.SIGTERM, sigterm_handler)
    
    if '--daemon' in sys.argv[1:]:
        sys.argv.remove('--daemon')
        if os.fork() != 0:
            sys.exit(0)
        os.setsid()
        sys.stdout=open("/dev/null", 'w')
        sys.stdin=open("/dev/null", 'r')
        sys.stderr=sys.stdout
        
    # Exit if perfmon already running
    if os.path.exists(pidfile):
        pid = open(pidfile).read()
        if os.path.exists("/proc/%s" % pid):
            log_err("perfmon already running - exiting")
            sys.exit(3)

    try:
        # Write out pidfile 
        fd = open(pidfile,"w")
        fd.write("%d" % os.getpid())
        fd.close()

        # run the main loop
        rc = main()

    except UsageException, e:
        # Print the usage
        log_err("usage: %s [-i <interval> -n <loops> -d -s <rrd_step> -c <config_update_period> -D <interval_percent_dither>] \\\n" \
                    "\t[--interval=<interval> --numloops=<loops> --debug \\\n" \
                    "\t --rrdstep=<rrd_step> --daemon]\n" \
                    "\t --config_update_period=<config_update_period>\n" \
                    "\t --interval_percent_dither=<interval_percent_dither>\n" \
                    "  interval:\tseconds between reads of http://localhost/rrd_updates?...\n" \
                    "  loops:\tnumber of times to run before exiting\n" \
                    "  rrd_step:\tseconds between samples provided by rrd_updates.  Valid values are 5 or 60\n" \
                    "  config_update_period:\tseconds between getting updates of all VM/host records from master\n" \
                    "  interval_percent_dither:\tmax percent dither in each loop - prevents stampede on master\n" \
                    % (sys.argv[0]))
        rc = 1

    except SystemExit:
        # we caught a signal which we have already logged
        pass

    except Exception, e:
        ex = sys.exc_info()
        err = traceback.format_exception(*ex)
        errmsg = "\n".join([ str(x) for x in e.args ])

        # print the exception args nicely
        log_err(errmsg)

        # now log the traceback to syslog
        for exline in err:
            log_err(exline)
        rc = 2

    # remove pidfile and exit
    os.unlink(pidfile)
    sys.exit(rc)