Blob Blame History Raw
--- fence-agents-4.2.1/agents/gce/fence_gce.py	2022-04-28 15:58:42.949723547 +0200
+++ fence-agents-4.2.1/agents/gce/fence_gce.py2	2022-04-28 15:59:21.054915265 +0200
@@ -1,10 +1,22 @@
 #!@PYTHON@ -tt
 
+#
+# Requires the googleapiclient and oauth2client
+# RHEL 7.x: google-api-python-client==1.6.7 python-gflags==2.0 pyasn1==0.4.8 rsa==3.4.2 pysocks==1.7.1 httplib2==0.19.0
+# RHEL 8.x: pysocks==1.7.1 httplib2==0.19.0
+# SLES 12.x: python-google-api-python-client python-oauth2client python-oauth2client-gce pysocks==1.7.1 httplib2==0.19.0
+# SLES 15.x: python3-google-api-python-client python3-oauth2client pysocks==1.7.1 httplib2==0.19.0
+#
+
 import atexit
 import logging
+import json
+import re
 import os
+import socket
 import sys
 import time
+
 if sys.version_info >= (3, 0):
   # Python 3 imports.
   import urllib.parse as urlparse
@@ -15,7 +27,7 @@
   import urllib2 as urlrequest
 sys.path.append("@FENCEAGENTSLIBDIR@")
 
-from fencing import fail_usage, run_delay, all_opt, atexit_handler, check_input, process_input, show_docs, fence_action
+from fencing import fail_usage, run_delay, all_opt, atexit_handler, check_input, process_input, show_docs, fence_action, run_command
 try:
   sys.path.insert(0, '/usr/lib/fence-agents/bundled/google')
   import httplib2
@@ -30,6 +42,85 @@
 
 METADATA_SERVER = 'http://metadata.google.internal/computeMetadata/v1/'
 METADATA_HEADERS = {'Metadata-Flavor': 'Google'}
+INSTANCE_LINK = 'https://www.googleapis.com/compute/v1/projects/{}/zones/{}/instances/{}'
+
+def run_on_fail(options):
+	if "--runonfail" in options:
+		run_command(options, options["--runonfail"])
+
+def fail_fence_agent(options, message):
+	run_on_fail(options)
+	fail_usage(message)
+
+def raise_fence_agent(options, message):
+	run_on_fail(options)
+	raise Exception(message)
+
+#
+# Will use baremetalsolution setting or the environment variable
+# FENCE_GCE_URI_REPLACEMENTS to replace the uri for calls to *.googleapis.com.
+#
+def replace_api_uri(options, http_request):
+	uri_replacements = []
+	# put any env var replacements first, then baremetalsolution if in options
+	if "FENCE_GCE_URI_REPLACEMENTS" in os.environ:
+		logging.debug("FENCE_GCE_URI_REPLACEMENTS environment variable exists")
+		env_uri_replacements = os.environ["FENCE_GCE_URI_REPLACEMENTS"]
+		try:
+			uri_replacements_json = json.loads(env_uri_replacements)
+			if isinstance(uri_replacements_json, list):
+				uri_replacements = uri_replacements_json
+			else:
+				logging.warning("FENCE_GCE_URI_REPLACEMENTS exists, but is not a JSON List")
+		except ValueError as e:
+			logging.warning("FENCE_GCE_URI_REPLACEMENTS exists but is not valid JSON")
+	if "--baremetalsolution" in options:
+		uri_replacements.append(
+			{
+				"matchlength": 4,
+				"match": "https://compute.googleapis.com/compute/v1/projects/(.*)/zones/(.*)/instances/(.*)/reset(.*)",
+				"replace": "https://baremetalsolution.googleapis.com/v1/projects/\\1/locations/\\2/instances/\\3:resetInstance\\4"
+			})
+	for uri_replacement in uri_replacements:
+		# each uri_replacement should have matchlength, match, and replace
+		if "matchlength" not in uri_replacement or "match" not in uri_replacement or "replace" not in uri_replacement:
+			logging.warning("FENCE_GCE_URI_REPLACEMENTS missing matchlength, match, or replace in %s" % uri_replacement)
+			continue
+		match = re.match(uri_replacement["match"], http_request.uri)
+		if match is None or len(match.groups()) != uri_replacement["matchlength"]:
+			continue
+		replaced_uri = re.sub(uri_replacement["match"], uri_replacement["replace"], http_request.uri)
+		match = re.match("https:\/\/.*.googleapis.com", replaced_uri)
+		if match is None or match.start() != 0:
+			logging.warning("FENCE_GCE_URI_REPLACEMENTS replace is not "
+				"targeting googleapis.com, ignoring it: %s" % replaced_uri)
+			continue
+		logging.debug("Replacing googleapis uri %s with %s" % (http_request.uri, replaced_uri))
+		http_request.uri = replaced_uri
+		break
+	return http_request
+
+def retry_api_execute(options, http_request):
+	replaced_http_request = replace_api_uri(options, http_request)
+	retries = 3
+	if options.get("--retries"):
+		retries = int(options.get("--retries"))
+	retry_sleep = 5
+	if options.get("--retrysleep"):
+		retry_sleep = int(options.get("--retrysleep"))
+	retry = 0
+	current_err = None
+	while retry <= retries:
+		if retry > 0:
+			time.sleep(retry_sleep)
+		try:
+			return replaced_http_request.execute()
+		except Exception as err:
+			current_err = err
+			logging.warning("Could not execute api call to: %s, retry: %s, "
+				"err: %s" % (replaced_http_request.uri, retry, str(err)))
+		retry += 1
+	raise current_err
 
 
 def translate_status(instance_status):
@@ -43,86 +134,174 @@
 
 def get_nodes_list(conn, options):
 	result = {}
+	if "--zone" not in options:
+		fail_fence_agent(options, "Failed: get_nodes_list: Please specify the --zone in the command")
 	try:
-		instanceList = conn.instances().list(project=options["--project"], zone=options["--zone"]).execute()
-		for instance in instanceList["items"]:
-			result[instance["id"]] = (instance["name"], translate_status(instance["status"]))
+		for zone in options["--zone"].split(","):
+			instanceList = retry_api_execute(options, conn.instances().list(
+				project=options["--project"],
+				zone=zone))
+			for instance in instanceList["items"]:
+				result[instance["id"]] = (instance["name"], translate_status(instance["status"]))
 	except Exception as err:
-		fail_usage("Failed: get_nodes_list: {}".format(str(err)))
+		fail_fence_agent(options, "Failed: get_nodes_list: {}".format(str(err)))
 
 	return result
 
 
 def get_power_status(conn, options):
+	logging.debug("get_power_status")
+	# if this is bare metal we need to just send back the opposite of the
+	# requested action: if on send off, if off send on
+	if "--baremetalsolution" in options:
+		if options.get("--action") == "on":
+			return "off"
+		else:
+			return "on"
+	# If zone is not listed for an entry we attempt to get it automatically
+	instance = options["--plug"]
+	zone = get_zone(conn, options, instance) if "--plugzonemap" not in options else options["--plugzonemap"][instance]
+	instance_status = get_instance_power_status(conn, options, instance, zone)
+	# If any of the instances do not match the intended status we return the
+	# the opposite status so that the fence agent can change it.
+	if instance_status != options.get("--action"):
+		return instance_status
+
+	return options.get("--action")
+
+
+def get_instance_power_status(conn, options, instance, zone):
 	try:
-		instance = conn.instances().get(
-				project=options["--project"],
-				zone=options["--zone"],
-				instance=options["--plug"]).execute()
+		instance = retry_api_execute(
+				options,
+				conn.instances().get(project=options["--project"], zone=zone, instance=instance))
 		return translate_status(instance["status"])
 	except Exception as err:
-		fail_usage("Failed: get_power_status: {}".format(str(err)))
+		fail_fence_agent(options, "Failed: get_instance_power_status: {}".format(str(err)))
 
 
-def wait_for_operation(conn, project, zone, operation):
+def check_for_existing_operation(conn, options, instance, zone, operation_type):
+	logging.debug("check_for_existing_operation")
+	if "--baremetalsolution" in options:
+		# There is no API for checking in progress operations
+		return False
+
+	project = options["--project"]
+	target_link = INSTANCE_LINK.format(project, zone, instance)
+	query_filter = '(targetLink = "{}") AND (operationType = "{}") AND (status = "RUNNING")'.format(target_link, operation_type)
+	result = retry_api_execute(
+			options,
+			conn.zoneOperations().list(project=project, zone=zone, filter=query_filter, maxResults=1))
+
+	if "items" in result and result["items"]:
+		logging.info("Existing %s operation found", operation_type)
+		return result["items"][0]
+
+
+def wait_for_operation(conn, options, zone, operation):
+	if 'name' not in operation:
+		logging.warning('Cannot wait for operation to complete, the'
+		' requested operation will continue asynchronously')
+		return False
+
+	wait_time = 0
+	project = options["--project"]
 	while True:
-		result = conn.zoneOperations().get(
+		result = retry_api_execute(options, conn.zoneOperations().get(
 			project=project,
 			zone=zone,
-			operation=operation['name']).execute()
+			operation=operation['name']))
 		if result['status'] == 'DONE':
 			if 'error' in result:
-				raise Exception(result['error'])
-			return
+				raise_fence_agent(options, result['error'])
+			return True
+
+		if "--errortimeout" in options and wait_time > int(options["--errortimeout"]):
+			raise_fence_agent(options, "Operation did not complete before the timeout.")
+
+		if "--warntimeout" in options and wait_time > int(options["--warntimeout"]):
+			logging.warning("Operation did not complete before the timeout.")
+			if "--runonwarn" in options:
+				run_command(options, options["--runonwarn"])
+			return False
+
+		wait_time = wait_time + 1
 		time.sleep(1)
 
 
 def set_power_status(conn, options):
+	logging.debug("set_power_status")
+	instance = options["--plug"]
+	# If zone is not listed for an entry we attempt to get it automatically
+	zone = get_zone(conn, options, instance) if "--plugzonemap" not in options else options["--plugzonemap"][instance]
+	set_instance_power_status(conn, options, instance, zone, options["--action"])
+
+
+def set_instance_power_status(conn, options, instance, zone, action):
+	logging.info("Setting power status of %s in zone %s", instance, zone)
+	project = options["--project"]
+
 	try:
-		if options["--action"] == "off":
-			logging.info("Issuing poweroff of %s in zone %s" % (options["--plug"], options["--zone"]))
-			operation = conn.instances().stop(
-					project=options["--project"],
-					zone=options["--zone"],
-					instance=options["--plug"]).execute()
-			wait_for_operation(conn, options["--project"], options["--zone"], operation)
-			logging.info("Poweroff of %s in zone %s complete" % (options["--plug"], options["--zone"]))
-		elif options["--action"] == "on":
-			logging.info("Issuing poweron of %s in zone %s" % (options["--plug"], options["--zone"]))
-			operation = conn.instances().start(
-					project=options["--project"],
-					zone=options["--zone"],
-					instance=options["--plug"]).execute()
-			wait_for_operation(conn, options["--project"], options["--zone"], operation)
-			logging.info("Poweron of %s in zone %s complete" % (options["--plug"], options["--zone"]))
+		if action == "off":
+			logging.info("Issuing poweroff of %s in zone %s", instance, zone)
+			operation = check_for_existing_operation(conn, options, instance, zone, "stop")
+			if operation and "--earlyexit" in options:
+				return
+			if not operation:
+				operation = retry_api_execute(
+						options,
+						conn.instances().stop(project=project, zone=zone, instance=instance))
+			logging.info("Poweroff command completed, waiting for the operation to complete")
+			if wait_for_operation(conn, options, zone, operation):
+				logging.info("Poweroff of %s in zone %s complete", instance, zone)
+		elif action == "on":
+			logging.info("Issuing poweron of %s in zone %s", instance, zone)
+			operation = check_for_existing_operation(conn, options, instance, zone, "start")
+			if operation and "--earlyexit" in options:
+				return
+			if not operation:
+				operation = retry_api_execute(
+						options,
+						conn.instances().start(project=project, zone=zone, instance=instance))
+			if wait_for_operation(conn, options, zone, operation):
+				logging.info("Poweron of %s in zone %s complete", instance, zone)
 	except Exception as err:
-		fail_usage("Failed: set_power_status: {}".format(str(err)))
-
+		fail_fence_agent(options, "Failed: set_instance_power_status: {}".format(str(err)))
 
 def power_cycle(conn, options):
+	logging.debug("power_cycle")
+	instance = options["--plug"]
+	# If zone is not listed for an entry we attempt to get it automatically
+	zone = get_zone(conn, options, instance) if "--plugzonemap" not in options else options["--plugzonemap"][instance]
+	return power_cycle_instance(conn, options, instance, zone)
+
+
+def power_cycle_instance(conn, options, instance, zone):
+	logging.info("Issuing reset of %s in zone %s", instance, zone)
+	project = options["--project"]
+
 	try:
-		logging.info('Issuing reset of %s in zone %s' % (options["--plug"], options["--zone"]))
-		operation = conn.instances().reset(
-				project=options["--project"],
-				zone=options["--zone"],
-				instance=options["--plug"]).execute()
-		wait_for_operation(conn, options["--project"], options["--zone"], operation)
-		logging.info('Reset of %s in zone %s complete' % (options["--plug"], options["--zone"]))
+		operation = check_for_existing_operation(conn, options, instance, zone, "reset")
+		if operation and "--earlyexit" in options:
+			return True
+		if not operation:
+			operation = retry_api_execute(
+					options,
+					conn.instances().reset(project=project, zone=zone, instance=instance))
+		logging.info("Reset command sent, waiting for the operation to complete")
+		if wait_for_operation(conn, options, zone, operation):
+			logging.info("Reset of %s in zone %s complete", instance, zone)
 		return True
 	except Exception as err:
-		logging.error("Failed: power_cycle: {}".format(str(err)))
-		return False
-
-
-def get_instance(conn, project, zone, instance):
-	request = conn.instances().get(
-			project=project, zone=zone, instance=instance)
-	return request.execute()
+		logging.exception("Failed: power_cycle")
+		raise err
 
 
-def get_zone(conn, project, instance):
+def get_zone(conn, options, instance):
+	logging.debug("get_zone");
+	project = options['--project']
 	fl = 'name="%s"' % instance
-	request = conn.instances().aggregatedList(project=project, filter=fl)
+	request = replace_api_uri(options, conn.instances().aggregatedList(project=project, filter=fl))
 	while request is not None:
 		response = request.execute()
 		zones = response.get('items', {})
@@ -130,9 +309,9 @@
 			for inst in zone.get('instances', []):
 				if inst['name'] == instance:
 					return inst['zone'].split("/")[-1]
-		request = conn.instances().aggregatedList_next(
-				previous_request=request, previous_response=response)
-	raise Exception("Unable to find instance %s" % (instance))
+		request = replace_api_uri(options, conn.instances().aggregatedList_next(
+				previous_request=request, previous_response=response))
+	raise_fence_agent(options, "Unable to find instance %s" % (instance))
 
 
 def get_metadata(metadata_key, params=None, timeout=None):
@@ -149,6 +328,7 @@
 	Raises:
 		urlerror.HTTPError: raises when the GET request fails.
 	"""
+	logging.debug("get_metadata");
 	timeout = timeout or 60
 	metadata_url = os.path.join(METADATA_SERVER, metadata_key)
 	params = urlparse.urlencode(params or {})
@@ -178,12 +358,50 @@
 	all_opt["stackdriver-logging"] = {
 		"getopt" : "",
 		"longopt" : "stackdriver-logging",
-		"help" : "--stackdriver-logging		Enable Logging to Stackdriver. Using stackdriver logging requires additional libraries (google-cloud-logging).",
-		"shortdesc" : "Stackdriver-logging support. Requires additional libraries (google-cloud-logging).",
-		"longdesc" : "If enabled IP failover logs will be posted to stackdriver logging. Using stackdriver logging requires additional libraries (google-cloud-logging).",
+		"help" : "--stackdriver-logging          Enable Logging to Stackdriver",
+		"shortdesc" : "Stackdriver-logging support.",
+		"longdesc" : "If enabled IP failover logs will be posted to stackdriver logging.",
 		"required" : "0",
 		"order" : 4
 	}
+	all_opt["baremetalsolution"] = {
+		"getopt" : "",
+		"longopt" : "baremetalsolution",
+		"help" : "--baremetalsolution            Enable on bare metal",
+		"shortdesc" : "If enabled this is a bare metal offering from google.",
+		"required" : "0",
+		"order" : 5
+	}
+	all_opt["apitimeout"] = {
+		"getopt" : ":",
+		"type" : "second",
+		"longopt" : "apitimeout",
+		"help" : "--apitimeout=[seconds]         Timeout to use for API calls",
+		"shortdesc" : "Timeout in seconds to use for API calls, default is 60.",
+		"required" : "0",
+		"default" : 60,
+		"order" : 6
+	}
+	all_opt["retries"] = {
+		"getopt" : ":",
+		"type" : "integer",
+		"longopt" : "retries",
+		"help" : "--retries=[retries]            Number of retries on failure for API calls",
+		"shortdesc" : "Number of retries on failure for API calls, default is 3.",
+		"required" : "0",
+		"default" : 3,
+		"order" : 7
+	}
+	all_opt["retrysleep"] = {
+		"getopt" : ":",
+		"type" : "second",
+		"longopt" : "retrysleep",
+		"help" : "--retrysleep=[seconds]         Time to sleep between API retries",
+		"shortdesc" : "Time to sleep in seconds between API retries, default is 5.",
+		"required" : "0",
+		"default" : 5,
+		"order" : 8
+	}
 	all_opt["serviceaccount"] = {
 		"getopt" : ":",
 		"longopt" : "serviceaccount",
@@ -192,13 +410,21 @@
 		"required" : "0",
 		"order" : 9
 	}
+	all_opt["plugzonemap"] = {
+		"getopt" : ":",
+		"longopt" : "plugzonemap",
+		"help" : "--plugzonemap=[plugzonemap]    Comma separated zone map when fencing multiple plugs",
+		"shortdesc" : "Comma separated zone map when fencing multiple plugs.",
+		"required" : "0",
+		"order" : 10
+	}
 	all_opt["proxyhost"] = {
 		"getopt" : ":",
 		"longopt" : "proxyhost",
 		"help" : "--proxyhost=[proxy_host]       The proxy host to use, if one is needed to access the internet (Example: 10.122.0.33)",
 		"shortdesc" : "If a proxy is used for internet access, the proxy host should be specified.",
 		"required" : "0",
-		"order" : 10
+		"order" : 11
 	}
 	all_opt["proxyport"] = {
 		"getopt" : ":",
@@ -207,7 +433,49 @@
 		"help" : "--proxyport=[proxy_port]       The proxy port to use, if one is needed to access the internet (Example: 3127)",
 		"shortdesc" : "If a proxy is used for internet access, the proxy port should be specified.",
 		"required" : "0",
-		"order" : 11
+		"order" : 12
+	}
+	all_opt["earlyexit"] = {
+		"getopt" : "",
+		"longopt" : "earlyexit",
+		"help" : "--earlyexit                    Return early if reset is already in progress",
+		"shortdesc" : "If an existing reset operation is detected, the fence agent will return before the operation completes with a 0 return code.",
+		"required" : "0",
+		"order" : 13
+	}
+	all_opt["warntimeout"] = {
+		"getopt" : ":",
+		"type" : "second",
+		"longopt" : "warntimeout",
+		"help" : "--warntimeout=[warn_timeout]   Timeout seconds before logging a warning and returning a 0 status code",
+		"shortdesc" : "If the operation is not completed within the timeout, the cluster operations are allowed to continue.",
+		"required" : "0",
+		"order" : 14
+	}
+	all_opt["errortimeout"] = {
+		"getopt" : ":",
+		"type" : "second",
+		"longopt" : "errortimeout",
+		"help" : "--errortimeout=[error_timeout] Timeout seconds before failing and returning a non-zero status code",
+		"shortdesc" : "If the operation is not completed within the timeout, cluster is notified of the operation failure.",
+		"required" : "0",
+		"order" : 15
+	}
+	all_opt["runonwarn"] = {
+		"getopt" : ":",
+		"longopt" : "runonwarn",
+		"help" : "--runonwarn=[run_on_warn]      If a timeout occurs and warning is generated, run the supplied command",
+		"shortdesc" : "If a timeout would occur while running the agent, then the supplied command is run.",
+		"required" : "0",
+		"order" : 16
+	}
+	all_opt["runonfail"] = {
+		"getopt" : ":",
+		"longopt" : "runonfail",
+		"help" : "--runonfail=[run_on_fail]      If a failure occurs, run the supplied command",
+		"shortdesc" : "If a failure would occur while running the agent, then the supplied command is run.",
+		"required" : "0",
+		"order" : 17
 	}
 
 
@@ -215,7 +483,9 @@
 	conn = None
 
 	device_opt = ["port", "no_password", "zone", "project", "stackdriver-logging",
-		"method", "serviceaccount", "proxyhost", "proxyport"]
+		"method", "baremetalsolution", "apitimeout", "retries", "retrysleep",
+		"serviceaccount", "plugzonemap", "proxyhost", "proxyport", "earlyexit",
+		"warntimeout", "errortimeout", "runonwarn", "runonfail"]
 
 	atexit.register(atexit_handler)
 
@@ -259,6 +529,11 @@
 			logging.error('Couldn\'t import google.cloud.logging, '
 				'disabling Stackdriver-logging support')
 
+  # if apitimeout is defined we set the socket timeout, if not we keep the
+  # socket default which is 60s
+	if options.get("--apitimeout"):
+		socket.setdefaulttimeout(options["--apitimeout"])
+
 	# Prepare cli
 	try:
 		serviceaccount = options.get("--serviceaccount")
@@ -291,20 +566,39 @@
 			conn = googleapiclient.discovery.build(
 				'compute', 'v1', credentials=credentials, cache_discovery=False)
 	except Exception as err:
-		fail_usage("Failed: Create GCE compute v1 connection: {}".format(str(err)))
+		fail_fence_agent(options, "Failed: Create GCE compute v1 connection: {}".format(str(err)))
 
 	# Get project and zone
 	if not options.get("--project"):
 		try:
 			options["--project"] = get_metadata('project/project-id')
 		except Exception as err:
-			fail_usage("Failed retrieving GCE project. Please provide --project option: {}".format(str(err)))
+			fail_fence_agent(options, "Failed retrieving GCE project. Please provide --project option: {}".format(str(err)))
 
-	if not options.get("--zone"):
-		try:
-			options["--zone"] = get_zone(conn, options['--project'], options['--plug'])
-		except Exception as err:
-			fail_usage("Failed retrieving GCE zone. Please provide --zone option: {}".format(str(err)))
+	if "--baremetalsolution" in options:
+		options["--zone"] = "none"
+
+	# Populates zone automatically if missing from the command
+	zones = [] if not "--zone" in options else options["--zone"].split(",")
+	options["--plugzonemap"] = {}
+	if "--plug" in options:
+		for i, instance in enumerate(options["--plug"].split(",")):
+			if len(zones) == 1:
+				# If only one zone is specified, use it across all plugs
+				options["--plugzonemap"][instance] = zones[0]
+				continue
+
+			if len(zones) - 1 >= i:
+				# If we have enough zones specified with the --zone flag use the zone at
+				# the same index as the plug
+				options["--plugzonemap"][instance] = zones[i]
+				continue
+
+			try:
+				# In this case we do not have a zone specified so we attempt to detect it
+				options["--plugzonemap"][instance] = get_zone(conn, options, instance)
+			except Exception as err:
+				fail_fence_agent(options, "Failed retrieving GCE zone. Please provide --zone option: {}".format(str(err)))
 
 	# Operate the fencing device
 	result = fence_action(conn, options, set_power_status, get_power_status, get_nodes_list, power_cycle)