Add support for replacing/refreshing OSDs

Adds commands to both replace an OSD disk, and refresh (reimport) an
existing OSD disk on a new node. This handles the cases where an OSD
disk should be replaced (either due to upgrades or failures) or where a
node is rebuilt in-place and an existing OSD must be re-imported to it.

This should avoid the need to do a full remove/add sequence for either
case.

Also cleans up some aspects of OSD removal that are identical between
methods (e.g. using safe-to-destroy and sleeping after stopping) and
fixes a bug if an OSD does not truly exist when the daemon starts up.
This commit is contained in:
2022-05-06 15:31:58 -04:00
parent 8027a6efdc
commit 7a40c7a55b
7 changed files with 853 additions and 37 deletions

View File

@ -21,7 +21,6 @@
import time
import json
import psutil
import daemon_lib.common as common
@ -217,6 +216,10 @@ class CephOSDInstance(object):
retcode, stdout, stderr = common.run_os_command(
f"ceph-volume lvm list {find_device}"
)
osd_blockdev = None
osd_fsid = None
osd_clusterfsid = None
osd_device = None
for line in stdout.split("\n"):
if "block device" in line:
osd_blockdev = line.split()[-1]
@ -227,7 +230,7 @@ class CephOSDInstance(object):
if "devices" in line:
osd_device = line.split()[-1]
if not osd_fsid:
if not osd_blockdev or not osd_fsid or not osd_clusterfsid or not osd_device:
self.logger.out(
f"Failed to find updated OSD information via ceph-volume for {find_device}",
state="e",
@ -404,6 +407,7 @@ class CephOSDInstance(object):
print(stdout)
print(stderr)
raise Exception
time.sleep(0.5)
# 6. Verify it started
@ -447,6 +451,381 @@ class CephOSDInstance(object):
logger.out("Failed to create new OSD disk: {}".format(e), state="e")
return False
@staticmethod
def replace_osd(
zkhandler,
logger,
node,
osd_id,
old_device,
new_device,
weight,
ext_db_flag=False,
):
# Handle a detect device if that is passed
if match(r"detect:", new_device):
ddevice = get_detect_device(new_device)
if ddevice is None:
logger.out(
f"Failed to determine block device from detect string {new_device}",
state="e",
)
return False
else:
logger.out(
f"Determined block device {ddevice} from detect string {new_device}",
state="i",
)
new_device = ddevice
# We are ready to create a new OSD on this node
logger.out(
"Replacing OSD {} disk with block device {}".format(osd_id, new_device),
state="i",
)
try:
# Verify the OSD is present
retcode, stdout, stderr = common.run_os_command("ceph osd ls")
osd_list = stdout.split("\n")
if osd_id not in osd_list:
logger.out(
"Could not find OSD {} in the cluster".format(osd_id), state="e"
)
return True
# 1. Set the OSD down and out so it will flush
logger.out("Setting down OSD disk with ID {}".format(osd_id), state="i")
retcode, stdout, stderr = common.run_os_command(
"ceph osd down {}".format(osd_id)
)
if retcode:
print("ceph osd down")
print(stdout)
print(stderr)
raise Exception
logger.out("Setting out OSD disk with ID {}".format(osd_id), state="i")
retcode, stdout, stderr = common.run_os_command(
"ceph osd out {}".format(osd_id)
)
if retcode:
print("ceph osd out")
print(stdout)
print(stderr)
raise Exception
# 2. Wait for the OSD to be safe to remove (but don't wait for rebalancing to complete)
logger.out("Waiting for OSD {osd_id} to be safe to remove", state="i")
while True:
retcode, stdout, stderr = common.run_os_command(
f"ceph osd safe-to-destroy osd.{osd_id}"
)
if retcode in [0, 11]:
# Code 0 = success
# Code 11 = "Error EAGAIN: OSD(s) 5 have no reported stats, and not all PGs are active+clean; we cannot draw any conclusions." which means all PGs have been remappped but backfill is still occurring
break
else:
time.sleep(5)
# 3. Stop the OSD process
logger.out("Stopping OSD disk with ID {}".format(osd_id), state="i")
retcode, stdout, stderr = common.run_os_command(
"systemctl stop ceph-osd@{}".format(osd_id)
)
if retcode:
print("systemctl stop")
print(stdout)
print(stderr)
raise Exception
time.sleep(2)
# 4. Destroy the OSD
logger.out("Destroying OSD with ID {osd_id}", state="i")
retcode, stdout, stderr = common.run_os_command(
f"ceph osd destroy {osd_id} --yes-i-really-mean-it"
)
if retcode:
print("ceph osd destroy")
print(stdout)
print(stderr)
raise Exception
# 5. Adjust the weight
logger.out(
"Adjusting weight of OSD disk with ID {} in CRUSH map".format(osd_id),
state="i",
)
retcode, stdout, stderr = common.run_os_command(
"ceph osd crush reweight osd.{osdid} {weight}".format(
osdid=osd_id, weight=weight
)
)
if retcode:
print("ceph osd crush reweight")
print(stdout)
print(stderr)
raise Exception
# 6a. Zap the new disk to ensure it is ready to go
logger.out("Zapping disk {}".format(new_device), state="i")
retcode, stdout, stderr = common.run_os_command(
"ceph-volume lvm zap --destroy {}".format(new_device)
)
if retcode:
print("ceph-volume lvm zap")
print(stdout)
print(stderr)
raise Exception
dev_flags = "--data {}".format(new_device)
# 6b. Prepare the logical volume if ext_db_flag
if ext_db_flag:
db_device = "osd-db/osd-{}".format(osd_id)
dev_flags += " --block.db {}".format(db_device)
else:
db_device = ""
# 6c. Replace the OSD
logger.out(
"Preparing LVM for replaced OSD {} disk on {}".format(
osd_id, new_device
),
state="i",
)
retcode, stdout, stderr = common.run_os_command(
"ceph-volume lvm prepare --osd-id {osdid} --bluestore {devices}".format(
osdid=osd_id, devices=dev_flags
)
)
if retcode:
print("ceph-volume lvm prepare")
print(stdout)
print(stderr)
raise Exception
# 7a. Get OSD information
logger.out(
"Getting OSD information for ID {} on {}".format(osd_id, new_device),
state="i",
)
retcode, stdout, stderr = common.run_os_command(
"ceph-volume lvm list {device}".format(device=new_device)
)
for line in stdout.split("\n"):
if "block device" in line:
osd_blockdev = line.split()[-1]
if "osd fsid" in line:
osd_fsid = line.split()[-1]
if "cluster fsid" in line:
osd_clusterfsid = line.split()[-1]
if "devices" in line:
osd_device = line.split()[-1]
if not osd_fsid:
print("ceph-volume lvm list")
print("Could not find OSD information in data:")
print(stdout)
print(stderr)
raise Exception
# Split OSD blockdev into VG and LV components
# osd_blockdev = /dev/ceph-<uuid>/osd-block-<uuid>
_, _, osd_vg, osd_lv = osd_blockdev.split("/")
# Reset whatever we were given to Ceph's /dev/xdX naming
if new_device != osd_device:
new_device = osd_device
# 7b. Activate the OSD
logger.out("Activating new OSD disk with ID {}".format(osd_id), state="i")
retcode, stdout, stderr = common.run_os_command(
"ceph-volume lvm activate --bluestore {osdid} {osdfsid}".format(
osdid=osd_id, osdfsid=osd_fsid
)
)
if retcode:
print("ceph-volume lvm activate")
print(stdout)
print(stderr)
raise Exception
time.sleep(0.5)
# 8. Verify it started
retcode, stdout, stderr = common.run_os_command(
"systemctl status ceph-osd@{osdid}".format(osdid=osd_id)
)
if retcode:
print("systemctl status")
print(stdout)
print(stderr)
raise Exception
# 9. Update Zookeeper information
logger.out(
"Adding new OSD disk with ID {} to Zookeeper".format(osd_id), state="i"
)
zkhandler.write(
[
(("osd", osd_id), ""),
(("osd.node", osd_id), node),
(("osd.device", osd_id), new_device),
(("osd.db_device", osd_id), db_device),
(("osd.fsid", osd_id), ""),
(("osd.ofsid", osd_id), osd_fsid),
(("osd.cfsid", osd_id), osd_clusterfsid),
(("osd.lvm", osd_id), ""),
(("osd.vg", osd_id), osd_vg),
(("osd.lv", osd_id), osd_lv),
(
("osd.stats", osd_id),
'{"uuid": "|", "up": 0, "in": 0, "primary_affinity": "|", "utilization": "|", "var": "|", "pgs": "|", "kb": "|", "weight": "|", "reweight": "|", "node": "|", "used": "|", "avail": "|", "wr_ops": "|", "wr_data": "|", "rd_ops": "|", "rd_data": "|", "state": "|"}',
),
]
)
# Log it
logger.out(
"Replaced OSD {} disk with device {}".format(osd_id, new_device),
state="o",
)
return True
except Exception as e:
# Log it
logger.out("Failed to replace OSD {} disk: {}".format(osd_id, e), state="e")
return False
@staticmethod
def refresh_osd(zkhandler, logger, node, osd_id, device, ext_db_flag):
# Handle a detect device if that is passed
if match(r"detect:", device):
ddevice = get_detect_device(device)
if ddevice is None:
logger.out(
f"Failed to determine block device from detect string {device}",
state="e",
)
return False
else:
logger.out(
f"Determined block device {ddevice} from detect string {device}",
state="i",
)
device = ddevice
# We are ready to create a new OSD on this node
logger.out(
"Refreshing OSD {} disk on block device {}".format(osd_id, device),
state="i",
)
try:
# 1. Verify the OSD is present
retcode, stdout, stderr = common.run_os_command("ceph osd ls")
osd_list = stdout.split("\n")
if osd_id not in osd_list:
logger.out(
"Could not find OSD {} in the cluster".format(osd_id), state="e"
)
return True
dev_flags = "--data {}".format(device)
if ext_db_flag:
db_device = "osd-db/osd-{}".format(osd_id)
dev_flags += " --block.db {}".format(db_device)
else:
db_device = ""
# 2. Get OSD information
logger.out(
"Getting OSD information for ID {} on {}".format(osd_id, device),
state="i",
)
retcode, stdout, stderr = common.run_os_command(
"ceph-volume lvm list {device}".format(device=device)
)
for line in stdout.split("\n"):
if "block device" in line:
osd_blockdev = line.split()[-1]
if "osd fsid" in line:
osd_fsid = line.split()[-1]
if "cluster fsid" in line:
osd_clusterfsid = line.split()[-1]
if "devices" in line:
osd_device = line.split()[-1]
if not osd_fsid:
print("ceph-volume lvm list")
print("Could not find OSD information in data:")
print(stdout)
print(stderr)
raise Exception
# Split OSD blockdev into VG and LV components
# osd_blockdev = /dev/ceph-<uuid>/osd-block-<uuid>
_, _, osd_vg, osd_lv = osd_blockdev.split("/")
# Reset whatever we were given to Ceph's /dev/xdX naming
if device != osd_device:
device = osd_device
# 3. Activate the OSD
logger.out("Activating new OSD disk with ID {}".format(osd_id), state="i")
retcode, stdout, stderr = common.run_os_command(
"ceph-volume lvm activate --bluestore {osdid} {osdfsid}".format(
osdid=osd_id, osdfsid=osd_fsid
)
)
if retcode:
print("ceph-volume lvm activate")
print(stdout)
print(stderr)
raise Exception
time.sleep(0.5)
# 4. Verify it started
retcode, stdout, stderr = common.run_os_command(
"systemctl status ceph-osd@{osdid}".format(osdid=osd_id)
)
if retcode:
print("systemctl status")
print(stdout)
print(stderr)
raise Exception
# 5. Update Zookeeper information
logger.out(
"Adding new OSD disk with ID {} to Zookeeper".format(osd_id), state="i"
)
zkhandler.write(
[
(("osd", osd_id), ""),
(("osd.node", osd_id), node),
(("osd.device", osd_id), device),
(("osd.db_device", osd_id), db_device),
(("osd.fsid", osd_id), ""),
(("osd.ofsid", osd_id), osd_fsid),
(("osd.cfsid", osd_id), osd_clusterfsid),
(("osd.lvm", osd_id), ""),
(("osd.vg", osd_id), osd_vg),
(("osd.lv", osd_id), osd_lv),
(
("osd.stats", osd_id),
'{"uuid": "|", "up": 0, "in": 0, "primary_affinity": "|", "utilization": "|", "var": "|", "pgs": "|", "kb": "|", "weight": "|", "reweight": "|", "node": "|", "used": "|", "avail": "|", "wr_ops": "|", "wr_data": "|", "rd_ops": "|", "rd_data": "|", "state": "|"}',
),
]
)
# Log it
logger.out("Refreshed OSD {} disk on {}".format(osd_id, device), state="o")
return True
except Exception as e:
# Log it
logger.out("Failed to refresh OSD {} disk: {}".format(osd_id, e), state="e")
return False
@staticmethod
def remove_osd(zkhandler, logger, osd_id, osd_obj, force_flag):
logger.out("Removing OSD disk {}".format(osd_id), state="i")
@ -490,29 +869,16 @@ class CephOSDInstance(object):
else:
raise Exception
# 2. Wait for the OSD to flush
logger.out("Flushing OSD disk with ID {}".format(osd_id), state="i")
osd_string = str()
# 2. Wait for the OSD to be safe to remove (but don't wait for rebalancing to complete)
logger.out("Waiting for OSD {osd_id} to be safe to remove", state="i")
while True:
try:
retcode, stdout, stderr = common.run_os_command(
"ceph pg dump osds --format json"
)
dump_string = json.loads(stdout)
for osd in dump_string:
if str(osd["osd"]) == osd_id:
osd_string = osd
num_pgs = osd_string["num_pgs"]
if num_pgs > 0:
time.sleep(5)
else:
if force_flag:
logger.out("Ignoring error due to force flag", state="i")
else:
raise Exception
except Exception:
retcode, stdout, stderr = common.run_os_command(
f"ceph osd safe-to-destroy osd.{osd_id}"
)
if int(retcode) in [0, 11]:
break
else:
time.sleep(5)
# 3. Stop the OSD process and wait for it to be terminated
logger.out("Stopping OSD disk with ID {}".format(osd_id), state="i")
@ -527,19 +893,7 @@ class CephOSDInstance(object):
logger.out("Ignoring error due to force flag", state="i")
else:
raise Exception
# FIXME: There has to be a better way to do this /shrug
while True:
is_osd_up = False
# Find if there is a process named ceph-osd with arg '--id {id}'
for p in psutil.process_iter(attrs=["name", "cmdline"]):
if "ceph-osd" == p.info["name"] and "--id {}".format(
osd_id
) in " ".join(p.info["cmdline"]):
is_osd_up = True
# If there isn't, continue
if not is_osd_up:
break
time.sleep(2)
# 4. Determine the block devices
osd_vg = zkhandler.read(("osd.vg", osd_id))
@ -912,6 +1266,59 @@ def ceph_command(zkhandler, logger, this_node, data, d_osd):
# Wait 1 seconds before we free the lock, to ensure the client hits the lock
time.sleep(1)
# Replacing an OSD
if command == "osd_replace":
node, osd_id, old_device, new_device, weight, ext_db_flag = args.split(",")
ext_db_flag = bool(strtobool(ext_db_flag))
if node == this_node.name:
# Lock the command queue
zk_lock = zkhandler.writelock("base.cmd.ceph")
with zk_lock:
# Add the OSD
result = CephOSDInstance.replace_osd(
zkhandler,
logger,
node,
osd_id,
old_device,
new_device,
weight,
ext_db_flag,
)
# Command succeeded
if result:
# Update the command queue
zkhandler.write([("base.cmd.ceph", "success-{}".format(data))])
# Command failed
else:
# Update the command queue
zkhandler.write([("base.cmd.ceph", "failure-{}".format(data))])
# Wait 1 seconds before we free the lock, to ensure the client hits the lock
time.sleep(1)
# Refreshing an OSD
if command == "osd_refresh":
node, osd_id, device, ext_db_flag = args.split(",")
ext_db_flag = bool(strtobool(ext_db_flag))
if node == this_node.name:
# Lock the command queue
zk_lock = zkhandler.writelock("base.cmd.ceph")
with zk_lock:
# Add the OSD
result = CephOSDInstance.refresh_osd(
zkhandler, logger, node, osd_id, device, ext_db_flag
)
# Command succeeded
if result:
# Update the command queue
zkhandler.write([("base.cmd.ceph", "success-{}".format(data))])
# Command failed
else:
# Update the command queue
zkhandler.write([("base.cmd.ceph", "failure-{}".format(data))])
# Wait 1 seconds before we free the lock, to ensure the client hits the lock
time.sleep(1)
# Removing an OSD
elif command == "osd_remove":
osd_id, force = args.split(",")