[CRIU] [PATCH v5] Add docker phaul driver

Hui Kang hkang.sunysb at gmail.com
Sat Oct 24 13:40:09 PDT 2015


See the instruction at test/docker/HOWTO

Changes since v4:
 - Remove all docker-specfic check in iter.py
 - dst node should create docker runtime meta state.json

Changes since v3:
 - Use sync_imgs_to_target to sync criu image for docker
 - Add can_pre_dump() to check if type supports pre_dump

TODO (suggestions from xemul, avagin, nikita):
    (1) Resolve the criu_conn error at the end of start_migration()
        for docker
    (2) Remove sleep, call wait for docker daemon
    (3) Remove get_driver_name; using some fake object, additional
        abstraction layers and such stuff
    (4) Wait for docker-py to integrate the C/R APIs

Signed-off-by: Hui Kang <hkang.sunysb at gmail.com>
---
 p.haul                   |   2 +-
 phaul/fs_haul_subtree.py |  29 ++++++---
 phaul/htype.py           |   1 +
 phaul/images.py          |   3 +-
 phaul/iters.py           |   7 ++-
 phaul/p_haul_docker.py   | 158 +++++++++++++++++++++++++++++++++++++++++++++++
 phaul/p_haul_pid.py      |   3 +
 phaul/p_haul_vz.py       |   3 +
 phaul/service.py         |   3 +-
 test/docker/HOWTO        |  82 ++++++++++++++++++++++++
 10 files changed, 277 insertions(+), 14 deletions(-)
 create mode 100644 phaul/p_haul_docker.py
 create mode 100644 test/docker/HOWTO

diff --git a/p.haul b/p.haul
index 11460b4..478e1c7 100755
--- a/p.haul
+++ b/p.haul
@@ -25,7 +25,7 @@ import phaul.htype
 
 parser = argparse.ArgumentParser("Process HAULer")
 parser.add_argument("type", choices=phaul.htype.get_haul_names(),
-	help="Type of hat to haul, e.g. vz or lxc")
+	help="Type of hat to haul, e.g. vz, lxc, or docker")
 parser.add_argument("id", help="ID of what to haul")
 parser.add_argument("--to", help="IP where to haul")
 parser.add_argument("--fdrpc", help="File descriptor of rpc socket", type=int, required=True)
diff --git a/phaul/fs_haul_subtree.py b/phaul/fs_haul_subtree.py
index a9bd559..e20ad1b 100644
--- a/phaul/fs_haul_subtree.py
+++ b/phaul/fs_haul_subtree.py
@@ -11,29 +11,38 @@ import logging
 rsync_log_file = "rsync.log"
 
 class p_haul_fs:
-	def __init__(self, subtree_path):
-		logging.info("Initialized subtree FS hauler (%s)", subtree_path)
-		self.__root = subtree_path
+	def __init__(self, subtree_paths):
+		self.__roots = []
+		for path in subtree_paths:
+			logging.info("Initialized subtree FS hauler (%s)", path)
+			self.__roots.append(path)
+
 		self.__thost = None
 
 	def set_options(self, opts):
 		self.__thost = opts["to"]
 
+	def set_target_host(self, thost):
+		self.__thost = thost
+
 	def set_work_dir(self, wdir):
 		self.__wdir = wdir
 
 	def __run_rsync(self):
 		logf = open(os.path.join(self.__wdir, rsync_log_file), "w+")
-		dst = "%s:%s" % (self.__thost, os.path.dirname(self.__root))
 
-		# First rsync might be very long. Wait for it not
-		# to produce big pause between the 1st pre-dump and
-		# .stop_migration
+		for dir_name in self.__roots:
+
+			dst = "%s:%s" % (self.__thost, os.path.dirname(dir_name))
+
+			# First rsync might be very long. Wait for it not
+			# to produce big pause between the 1st pre-dump and
+			# .stop_migration
 
-		ret = sp.call(["rsync", "-a", self.__root, dst],
+			ret = sp.call(["rsync", "-a", dir_name, dst],
 				stdout = logf, stderr = logf)
-		if ret != 0:
-			raise Exception("Rsync failed")
+			if ret != 0:
+				raise Exception("Rsync failed")
 
 	def start_migration(self):
 		logging.info("Starting FS migration")
diff --git a/phaul/htype.py b/phaul/htype.py
index f3a4774..bee49b4 100644
--- a/phaul/htype.py
+++ b/phaul/htype.py
@@ -11,6 +11,7 @@ __haul_modules = {
 	"vz": "p_haul_vz",
 	"pid": "p_haul_pid",
 	"lxc": "p_haul_lxc",
+	"docker": "p_haul_docker",
 }
 
 def __get(id):
diff --git a/phaul/images.py b/phaul/images.py
index 57dca59..e48f96f 100644
--- a/phaul/images.py
+++ b/phaul/images.py
@@ -145,7 +145,8 @@ class phaul_images:
 		tf = img_tar(sk, cdir)
 
 		logging.info("\tPack")
-		for img in filter(lambda x: x.endswith(".img"), os.listdir(cdir)):
+		# .json file is needed for docker, i.e., descriptor.json
+		for img in filter(lambda x: x.endswith(('.img', '.json')) , os.listdir(cdir)):
 			tf.add(img)
 
 		logging.info("\tAdd htype images")
diff --git a/phaul/iters.py b/phaul/iters.py
index 9be4e97..00b2a02 100644
--- a/phaul/iters.py
+++ b/phaul/iters.py
@@ -113,7 +113,7 @@ class phaul_iter_worker:
 		if self.pre_dump == PRE_DUMP_AUTO_DETECT:
 			# pre-dump auto-detection
 			try:
-				self.pre_dump = self.pre_dump_check()
+				self.pre_dump = (self.pre_dump_check() and self.htype.can_pre_dump())
 				logging.info("\t`- Auto %s" % (self.pre_dump and 'enabled' or 'disabled'))
 			except:
 				# The available criu seems to not
@@ -121,6 +121,11 @@ class phaul_iter_worker:
 				self.pre_dump = PRE_DUMP_DISABLE
 				logging.info("\t`- Auto detection not possible "
 						"- Disabled")
+		elif self.pre_dump == PRE_DUMP_DISABLE:
+			logging.info("\t`- Command-line disabled")
+		elif self.htype.can_pre_dump() == False:
+			self.pre_dump = False
+			logging.info("\t`- Type does not support pre-dump")
 		else:
 			logging.info("\t`- Command-line %s" % (self.pre_dump and 'enabled' or 'disabled'))
 
diff --git a/phaul/p_haul_docker.py b/phaul/p_haul_docker.py
new file mode 100644
index 0000000..d2bb1e3
--- /dev/null
+++ b/phaul/p_haul_docker.py
@@ -0,0 +1,158 @@
+#
+# Docker container hauler
+#
+
+import os
+import logging
+import time
+import fs_haul_subtree
+import pycriu.rpc
+import json
+from pprint import pprint
+import subprocess as sp
+from subprocess import PIPE
+
+# TODO use docker-py
+# import docker
+
+# Some constants for docker
+name = "docker"
+docker_exec = "/usr/bin/docker-1.9.0-dev"
+docker_dir = "/var/lib/docker/"
+docker_run_meta_dir = "/var/run/docker/execdriver/native"
+
+class p_haul_type:
+	def __init__(self, ctid):
+
+		# TODO ctid must > 3 digit; with docker-py, we can also resolve
+		#	  container name
+		if len(ctid) < 3:
+			raise Exception("Docker container ID must be > 3 digits")
+
+		self._ctid = ctid
+		self._ct_rootfs = ""
+
+	def get_driver_name(self):
+		return name
+
+	def init_src(self):
+		self.full_ctid = self.get_full_ctid()
+		self.__load_ct_config(docker_dir)
+
+	def init_dst(self):
+		pass
+
+	def adjust_criu_req(self, req):
+		"""Add module-specific options to criu request"""
+		pass
+
+	def root_task_pid(self):
+		# Do we need this for Docker?
+		return self.full_ctid
+
+	def __load_ct_config(self, path):
+
+		# Each docker container has 3 directories that need to be
+		# migrated: (1) root filesystem, (2) container configuration,
+		# (3) runtime meta state
+		self._ct_rootfs = os.path.join(docker_dir, "aufs/mnt", self.full_ctid)
+		self._ct_config_dir = os.path.join(docker_dir, "containers", self.full_ctid)
+		self._ct_run_meta_dir = os.path.join(docker_run_meta_dir, self.full_ctid)
+		logging.info("Container rootfs: %s", self._ct_rootfs)
+		logging.info("Container config: %s", self._ct_config_dir)
+		logging.info("Container meta: %s", self._ct_run_meta_dir)
+
+	def set_options(self, opts):
+		pass
+
+	# Remove any specific FS setup
+	def umount(self):
+		pass
+
+	def get_fs(self, fs_sk=None):
+		# use rsync for rootfs and configuration directories
+		return fs_haul_subtree.p_haul_fs([self._ct_rootfs, self._ct_config_dir])
+
+	def get_fs_receiver(self, fs_sk=None):
+		return None
+
+	def get_full_ctid(self):
+		dir_name_list = os.listdir(os.path.join(docker_dir, "containers"))
+
+		full_id = ""
+		for name in dir_name_list:
+			name = name.rsplit("/")
+			if (name[0].find(self._ctid) == 0):
+				full_id = name[0]
+				break
+
+		if full_id != "":
+			return full_id
+		else:
+			raise Exception("Can not find container fs")
+
+	def final_dump(self, pid, img, ccon, fs):
+		logging.info("Dump docker container %s", pid)
+
+		# TODO: docker API does not have checkpoint right now
+		# cli.checkpoint() so we have to use the command line
+		# cli = docker.Client(base_url='unix://var/run/docker.sock')
+		# output = cli.info()
+		# call docker API
+
+		logf = open("/tmp/docker_checkpoint.log", "w+")
+		image_path_opt = "--image-dir=" + img.image_dir()
+		ret = sp.call([docker_exec, "checkpoint", image_path_opt, self._ctid],
+			stdout = logf, stderr = logf)
+		if ret != 0:
+			raise Exception("docker checkpoint failed")
+	#
+	# Meta-images for docker -- /var/run/docker
+	#
+	def get_meta_images(self, path):
+		# Send the meta state file with criu images
+		return [(os.path.join(self._ct_run_meta_dir, "state.json"), "state.json")]
+
+	def put_meta_images(self, dir):
+		# Create docker runtime meta dir on dst side
+		with open(os.path.join(dir, "state.json")) as data_file:
+			data = json.load(data_file)
+		self.full_ctid=data["id"]
+
+		self.__load_ct_config(docker_dir)
+		os.makedirs(self._ct_run_meta_dir)
+		pd = sp.Popen(["cp", os.path.join(dir, "state.json"), self._ct_run_meta_dir], stdout = PIPE)
+		status = pd.wait()
+
+	def kill_last_docker_daemon(self):
+		p = sp.Popen(['pgrep', '-l' , docker_exec], stdout=sp.PIPE)
+		out, err = p.communicate()
+
+		for line in out.splitlines():
+			line = bytes.decode(line)
+			pid = int(line.split(None, 1)[0])
+			os.kill(pid, signal.SIGKILL)
+
+	def final_restore(self, img, criu):
+		logf = open("/tmp/docker_restore.log", "w+")
+
+		# Kill any previous docker daemon in order to reload the
+		# status of the migrated container
+		self.kill_last_docker_daemon()
+
+		# start docker daemon in background
+		daemon = sp.Popen([docker_exec, "daemon", "-s", "aufs"],
+				 stdout = logf, stderr = logf)
+		# daemon.wait() TODO: docker daemon not return
+		time.sleep(2)
+
+		image_path_opt = "--image-dir=" + img.image_dir()
+		ret = sp.call([docker_exec, "restore", image_path_opt, self._ctid],
+						stdout = logf, stderr = logf)
+		if ret != 0:
+			raise Exception("docker restore failed")
+
+	def can_pre_dump(self):
+		# XXX: Do not do predump for docker right now. Add page-server
+		#	to docker C/R API, then we can enable the pre-dump
+		return False
diff --git a/phaul/p_haul_pid.py b/phaul/p_haul_pid.py
index e354dff..43af654 100644
--- a/phaul/p_haul_pid.py
+++ b/phaul/p_haul_pid.py
@@ -94,3 +94,6 @@ class p_haul_type:
 	# Get list of veth pairs if any
 	def veths(self):
 		return []
+
+	def can_pre_dump(self):
+		return True
diff --git a/phaul/p_haul_vz.py b/phaul/p_haul_vz.py
index 903006d..e346bdf 100644
--- a/phaul/p_haul_vz.py
+++ b/phaul/p_haul_vz.py
@@ -244,6 +244,9 @@ class p_haul_type:
 		return True
 
 
+	def can_pre_dump(self):
+		return True
+
 def parse_vz_config(body):
 	"""Parse shell-like virtuozzo config file"""
 
diff --git a/phaul/service.py b/phaul/service.py
index a03b43e..507d171 100644
--- a/phaul/service.py
+++ b/phaul/service.py
@@ -72,7 +72,8 @@ class phaul_service:
 	def rpc_start_iter(self):
 		self.dump_iter_index += 1
 		self.img.new_image_dir()
-		self.start_page_server()
+		if self.htype.get_driver_name() != "docker" :
+			self.start_page_server()
 
 	def rpc_end_iter(self):
 		pass
diff --git a/test/docker/HOWTO b/test/docker/HOWTO
new file mode 100644
index 0000000..3255360
--- /dev/null
+++ b/test/docker/HOWTO
@@ -0,0 +1,82 @@
+This HOWTO describes how to _non_ live-migrate a docker container from one
+docker host to another.
+
+** This is an experimental implementation of docker migration, which may affect
+your running containers.
+
+0. Install CRIU, p.haul, docker on both nodes
+
+   Besides the packages that are needed to compile and run CRIU and p.haul,
+   the specific docker binary that supports checkpoint/restore should be used.
+
+   Refer to step 0 in test/mtouch/HOWTO about the pacekages for CRIU and p.haul.
+
+   The docker version that supports checkpoint and restore can be obtained by
+
+   # git clone https://github.com/boucher/docker.git
+   # cd docker.git
+   # git checkout cr-combined
+
+   On both nodes, compile and store the the docker binary as
+
+   /usr/bin/docker-1.9.0-dev
+
+   Note that the path above is for now hard-coded in p_haul_docker.py
+
+1. Prepare criu and p.haul on both nodes (adapted from test/mtouch/HOWTO)
+
+   a) CRIU
+
+   * Clone CRIU repository from git://github.com/xemul/criu
+     and compile it with 'make'
+
+   * Make _local_ directory for service ($csdir)
+   * Start CRIU service by root user
+     Note that this step is mandatory. Although runC will start criu in swrk
+     mode, p.haul needs to connect to criu service for validating CPU.
+
+   # criu service -o $csdir/log -v4 --daemon
+
+   b) On destination node start the p.haul service
+
+   [dst]# ./p.haul-wrap service
+   Starting p.haul rpyc service
+
+   It will go daemon and will print logs on the terminals
+
+3. Run the test container on source node
+
+   a) Start the docker daemon
+
+   # /usr/bin/docker-1.9.0-dev daemon -s aufs
+
+   b) Start the container
+
+   # /usr/bin/docker-1.9.0-dev run -d busybox:latest /bin/sh -c 'i=0; while true; do echo $i >> /foo; i=$(expr $i + 1); sleep 1; done'
+
+   This command will return the container's ID, e.g.,  d78.
+   (borrowed from https://criu.org/Docker)
+
+4. Migrate container from source node
+
+   [src]# ./p.haul-wrap client to docker [container ID, e.g., d78]
+
+   to is the ip address of the dst node
+
+   For example:
+
+   [src]# ./p.haul-wrap client 192.168.11.106 docker d78
+   192.168.11.106 is the destination node IP and d78 is the container ID
+
+
+   When the command returns, on the destination node run
+
+   [dst]# /usr/bin/docker-1.9.0-dev ps
+   [dst]# /usr/bin/docker-1.9.0-dev exec d78 cat /foo
+
+   to verify the counter is continuously being incremented.
+
+Known limitations.
+
+1. No support from docker python binding
+2. Docker daemon has to be restarted on the destination node
-- 
1.9.1



More information about the CRIU mailing list