openstack/os-brick
Code Issues Proposed changes
os-brick/os_brick/initiator/connectors/nvmeof.py
Rajat Dhasmana 8d919696a9 Fix multipathd dependency in NVMe-oF connections 
Previously, we required multipathd to be running
to perform multipathing in NVMe-oF connections
even though we were using native NVMe multipathing
(ANA).
This was because, traditionally we only supported
iSCSI and FC connections so we only needed to check
for ``multipathd`` running. The multipathing check
happened early on while fetching the connector
when we had no info what kind of connection will
be created.
With this patch, the logic to check for multipathing
is moved to when we are discovering the volume (after
mapping) and each connector provides it's own
implementation to report if it supports multipathing
or not (by default, the value is False).
Based on the values of ``multipath``, ``enforce_multipath``
and ``supports_multipath``, we decide if we want to
proceed with the connect request, log a warning or fail.

Following are some of the cases tested with iSCSI
and NVMe-oF connector:

Enable multipathing:
target_ip_address = <primary_ip>
target_secondary_ip_addresses = <secondary_ip>

Tested with:

1. iSCSI
1.1 enforce multipathing + multipathd running
a) use_multipath_for_image_xfer = True
b) enforce_multipath_for_image_xfer = True
c) multipathd running
Result: PASS

1.2 don't enforce multipathing + multipathd stopped
a) use_multipath_for_image_xfer = True
b) enforce_multipath_for_image_xfer = False
c) multipathd stopped
Result: Log Warning

1.3 enforce_multipathing + multipathd stopped
a) use_multipath_for_image_xfer = True
b) enforce_multipath_for_image_xfer = True
c) multipathd stopped
Result: Fails with BrickException

2. NVMe-TCP

NOTE: multipathd is stopped since it shouldn't affect nvme connections now

2.1 enforce multipathing + native multipathing
a) use_multipath_for_image_xfer = True
b) enforce_multipath_for_image_xfer = True
c) native multipathing = Y
Result: PASS

2.2 don't enforce multipathing + no native multipathing
a) use_multipath_for_image_xfer = True
b) enforce_multipath_for_image_xfer = False
c) native multipathing = N
Result: Log Warning

2.3 enforce_multipathing + no native multipathing
a) use_multipath_for_image_xfer = True
b) enforce_multipath_for_image_xfer = True
c) native multipathing = N
Result: Fails with BrickException

Depends-On: https://review.opendev.org/c/openstack/cinder/+/934011

Closes-Bug: #2085013
Change-Id: I61c7d7bc040633646411758c3245d8432910f766
2024-12-09 17:07:06 +05:30

1521 lines
59 KiB
Python
Raw Permalink Blame History

# Licensed under the Apache License, Version 2.0 (the "License"); you may
# not use this file except in compliance with the License. You may obtain
# a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
# WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
# License for the specific language governing permissions and limitations
# under the License.
# Look in the NVMeOFConnProps class docstring to see the format of the NVMe-oF
# connection properties
from __future__ import annotations
import errno
import functools
import glob
import json
import os.path
import time
from typing import (Callable, Optional, Sequence, Type, Union)
import uuid as uuid_lib
from oslo_concurrency import processutils as putils
from oslo_log import log as logging
from os_brick import exception
from os_brick.i18n import _
from os_brick.initiator.connectors import base
try:
from os_brick.initiator.connectors import nvmeof_agent
except ImportError:
nvmeof_agent = None
from os_brick.privileged import nvmeof as priv_nvmeof
from os_brick.privileged import rootwrap as priv_rootwrap
from os_brick import utils
DEV_SEARCH_PATH = '/dev/'
RAID_PATH = '/dev/md/'
NVME_CTRL_SYSFS_PATH = '/sys/class/nvme-fabrics/ctl/'
BLK_SYSFS_PATH = '/sys/class/block/'
DEVICE_SCAN_ATTEMPTS_DEFAULT = 5
LOG = logging.getLogger(__name__)
# #########################################################
# CONNECTION PROPERTIES KEYS start
# Only present in the old connection info format
OLD_NQN = 'nqn'
TRANSPORT = 'transport_type'
PORTAL = 'target_portal'
PORT = 'target_port'
# These were only present in the old connection info, but now we'll allow
# both in the old format, the new format, and as part of a volume_replicas
# element.
NGUID = 'volume_nguid'
NSID = 'ns_id'
HOST_NQN = 'host_nqn'
# Present in the new new connection info format
UUID = 'vol_uuid'
NQN = 'target_nqn'
PORTALS = 'portals'
ALIAS = 'alias'
REPLICAS = 'volume_replicas'
REPLICA_COUNT = 'replica_count'
# CONNECTION PROPERTIES KEYS end
# #########################################################
# #########################################################
# UTILITY METHODS start
def ctrl_property(prop_name: str, ctrl_name: str) -> Optional[str]:
"""Get a sysfs property of an nvme controller."""
return sysfs_property(prop_name, NVME_CTRL_SYSFS_PATH + ctrl_name)
def blk_property(prop_name: str, blk_name: str) -> Optional[str]:
"""Get a sysfs property of a block device."""
return sysfs_property(prop_name, BLK_SYSFS_PATH + blk_name)
def sysfs_property(prop_name: str, path_or_name: str) -> Optional[str]:
"""Get sysfs property by path returning None if not possible."""
filename = os.path.join(path_or_name, prop_name)
LOG.debug('Checking property at %s', filename)
try:
with open(filename, 'r') as f:
result = f.read().strip()
LOG.debug('Contents: %s', result)
return result
except (FileNotFoundError, IOError) as exc:
# May happens on race conditions with device removals
LOG.debug('Error reading file %s', exc)
return None
def nvme_basename(path: str) -> str:
"""Convert a sysfs control path into a namespace device.
We can have a basic namespace devices such as nvme0n10 which is already in
the desired form, but there's also channels when ANA is enabled on the
kernel which have the form nvme0c2n10 which need to be converted to
nvme0n10 to get the actual device.
"""
basename = os.path.basename(path)
if 'c' not in basename: # nvme0n10
return basename
# nvme0c1n10 ==> nvme0n10
ctrl, rest = basename.split('c', 1)
ns = rest[rest.index('n'):]
return ctrl + ns
# UTILITY METHODS end
# #########################################################
# #########################################################
# AUXILIARY CLASSES start
class Portal(object):
"""Representation of an NVMe-oF Portal with some related operations."""
LIVE = 'live'
MISSING = None # Unkown or not present in the system
CONNECTING = 'connecting'
# Default value of reconnect_delay in sysfs
DEFAULT_RECONNECT_DELAY = 10
controller: Optional[str] = None # Don't know controller name on start
def __str__(self) -> str:
return (f'Portal {self.transport} at {self.address}:{self.port} '
f'(ctrl: {self.controller})')
__repr__ = __str__
def __init__(self,
parent_target: 'Target',
address: str,
port: Union[str, int],
transport: str) -> None:
self.parent_target = parent_target
self.address = address
self.port = str(port)
# Convert the transport into our internal representation
if transport in ('RoCEv2', 'rdma'):
self.transport = 'rdma'
else:
self.transport = 'tcp'
@property
def is_live(self) -> bool:
"""Check if the portal is live.
Not being live can mean many things, such as being connecting because
the connection to the backend was lost, not knowing the controller name
because we haven't searched for it, or not being connected to the
backend.
"""
LOG.debug('Checking if %s is live', self)
return self.state == self.LIVE
@property
def state(self) -> Optional[str]:
"""Return the state if the controller is known, None otherwise."""
# Does not automatically search for the controller
if self.controller:
return ctrl_property('state', self.controller)
return None
@property
def reconnect_delay(self) -> int:
# 10 seconds is the default value of reconnect_delay
if self.controller:
res = ctrl_property('reconnect_delay', self.controller)
if res is not None:
return int(res)
return self.DEFAULT_RECONNECT_DELAY
def get_device(self) -> Optional[str]:
"""Get a device path using available volume identification markers.
Priority is given to the uuid, since that must be unique across ALL
devices, then nguid which is backend specific but a backend can reuse
in other connections after it has been disconnected, and then using
the namespace id (nsid) which changes based on the subsystem and the
moment the volume is connected.
If the target in the connection information didn't have any of those
identifiers, then let the parent Target instance try to figure out the
device based on the devices that existed when we started connecting and
the ones available now.
None is returned when a device cannot be found.
"""
target = self.parent_target
if target.uuid:
return self.get_device_by_property('uuid', target.uuid)
if target.nguid:
return self.get_device_by_property('nguid', target.nguid)
if target.ns_id:
return self.get_device_by_property('nsid', target.ns_id)
# Fallback to using the target to do the search
LOG.warning('Using unreliable mechanism to find device: '
'"devices_on_start"')
return target.get_device_path_by_initial_devices()
def get_all_namespaces_ctrl_paths(self) -> list[str]:
"""Return all nvme sysfs control paths for this portal.
The basename of the path can be single volume or a channel to an ANA
volume.
For example for the nvme1 controller we could return:
['/sys/class/nvme-fabrics/ctl/nvme1n1 ',
'/sys/class/nvme-fabrics/ctl/nvme0c1n1']
"""
if not self.controller:
return []
# Look under the controller, where we will have normal devices and ANA
# channel devices. For nvme1 we could find nvme1n1 or nvme0c1n1)
return glob.glob(f'{NVME_CTRL_SYSFS_PATH}{self.controller}/nvme*')
def get_device_by_property(self,
prop_name: str,
value: str) -> Optional[str]:
"""Look for a specific device (namespace) within a controller.
Use a specific property to identify the namespace within the
controller and returns the device path under /dev.
Returns None if device is not found.
"""
LOG.debug('Looking for device where %s=%s on controller %s',
prop_name, value, self.controller)
for path in self.get_all_namespaces_ctrl_paths():
prop_value = sysfs_property(prop_name, path)
if prop_value == value:
# Convert path to the namespace device name
result = DEV_SEARCH_PATH + nvme_basename(path)
LOG.debug('Device found at %s, using %s', path, result)
return result
LOG.debug('Block %s is not the one we are looking for (%s != %s)',
path, prop_value, value)
LOG.debug('No device Found on controller %s', self.controller)
return None
def can_disconnect(self) -> bool:
"""Check if this portal can be disconnected.
A portal can be disconnected if it is connected (has a controller name)
and the subsystem has no namespaces left or if it has only one and it
is from this target.
"""
if not self.controller:
LOG.debug('Portal %s is not present', self)
return False
ns_ctrl_paths = self.get_all_namespaces_ctrl_paths()
num_namespaces = len(ns_ctrl_paths)
# No namespaces => disconnect, >1 ns => can't disconnect
if num_namespaces != 1:
result = not bool(num_namespaces)
LOG.debug('There are %s namespaces on %s so we %s disconnect',
num_namespaces, self, 'can' if result else 'cannot')
return result
# With only 1 namespace, check if it belongs to the portal
# Get the device on this target's portal (may be None)
portal_dev = os.path.basename(self.get_device() or '')
result = portal_dev == nvme_basename(ns_ctrl_paths[0])
LOG.debug("The only namespace on portal %s is %s and %s this target's",
self, portal_dev, "matches" if result else "doesn't match")
return result
class Target(object):
"""Representation of an NVMe-oF Target and some related operations."""
# Only used if the target has no uuid, nguid, or ns_id information
devices_on_start = None
# Cache the device we find for cases where we do retries
_device = None
def __str__(self) -> str:
return f'Target {self.nqn} at {self.portals}'
__repr__ = __str__
@classmethod
def factory(cls: Type['Target'],
source_conn_props: 'NVMeOFConnProps',
target_nqn: str,
portals: list[str],
vol_uuid: Optional[str] = None,
volume_nguid: Optional[str] = None,
ns_id: Optional[str] = None,
host_nqn: Optional[str] = None,
find_controllers=False,
**ignore) -> 'Target':
"""Create an instance from the connection properties keys.
Extra keyword arguments are accepted (and ignored) for convenience, so
they don't need to be removed when calling the factory.
"""
target = cls(source_conn_props, target_nqn, portals, vol_uuid,
volume_nguid, ns_id, host_nqn, find_controllers)
return target
def __init__(self,
source_conn_props: 'NVMeOFConnProps',
nqn: str,
portals: list[str],
uuid: Optional[str] = None,
nguid: Optional[str] = None,
ns_id: Optional[str] = None,
host_nqn=None,
find_controllers=False) -> None:
"""Initialize instance.
Portals are converted from a list of length 3 tuple/list into a list of
Portal instances.
The find_controllers parameter controls the search of the controller
names in the system for each of the portals.
"""
self.source_conn_props = source_conn_props
self.nqn = nqn
self.portals = [Portal(self, *portal) for portal in portals]
self.uuid = uuid and str(uuid_lib.UUID(uuid))
self.nguid = nguid and str(uuid_lib.UUID(nguid))
self.ns_id = ns_id
self.host_nqn = host_nqn
if find_controllers:
self.set_portals_controllers()
# This only happens with some old connection properties format, where
# we may not have a way to identify the new volume and we'll have to
# try to guess it looking at existing volumes before the attach.
if not (uuid or nguid or ns_id):
self.devices_on_start = self._get_nvme_devices()
LOG.debug('Devices on start are: %s', self.devices_on_start)
@staticmethod
def _get_nvme_devices() -> list[str]:
"""Get all NVMe devices present in the system."""
pattern = '/dev/nvme*n*' # e.g. /dev/nvme10n10
return glob.glob(pattern)
@property
def live_portals(self) -> list[Portal]:
"""Get live portals.
Must have called set_portals_controllers first since portals without a
controller name will be skipped.
"""
return [p for p in self.portals if p.is_live]
@property
def present_portals(self) -> list[Portal]:
"""Get present portals.
Must have called set_portals_controllers first since portals without a
controller name will be skipped.
"""
return [p for p in self.portals if p.state is not None]
def set_portals_controllers(self) -> None:
"""Search and set controller names in the target's portals.
Compare the address, port, and transport protocol for each portal
against existing nvme subsystem controllers.
"""
if all(p.controller for p in self.portals): # all have been found
return
hostnqn: Optional[str] = self.host_nqn or utils.get_host_nqn()
# List of portal addresses and transports for this target
# Unlike "nvme list-subsys -o json" sysfs addr is separated by a comma
sysfs_portals: list[tuple[Optional[str],
Optional[str],
Optional[Union[str, utils.Anything]],
Optional[Union[str, utils.Anything]]]] = [
(p.address, p.port, p.transport, hostnqn)
for p in self.portals
]
known_names: list[str] = [p.controller for p in self.portals
if p.controller]
warned = False
LOG.debug('Search controllers for portals %s', sysfs_portals)
ctrl_paths = glob.glob(NVME_CTRL_SYSFS_PATH + 'nvme*')
for ctrl_path in ctrl_paths:
ctrl_name = os.path.basename(ctrl_path)
if ctrl_name in known_names:
continue
LOG.debug('Checking controller %s', ctrl_name)
nqn = sysfs_property('subsysnqn', ctrl_path)
if nqn != self.nqn:
LOG.debug("Skipping %s, doesn't match %s", nqn, self.nqn)
continue
# The right subsystem, but must also be the right portal
ctrl_transport = sysfs_property('transport', ctrl_path)
# Address in sysfs may contain src_addr in some systems. Parse and
# only use destination addr and port
address = sysfs_property('address', ctrl_path)
if not address:
LOG.error("Couldn't read address for %s", ctrl_path)
continue
ctrl_address = dict((x.split('=')
for x in address.split(',')))
ctrl_addr = ctrl_address['traddr']
ctrl_port = ctrl_address['trsvcid']
# hostnqn value not present in all OSs. Ignore when not present
ctrl_hostnqn = sysfs_property('hostnqn', ctrl_path) or utils.ANY
# Warn once per target for OS not presenting the hostnqn on sysfs
if ctrl_hostnqn is utils.ANY and not warned:
LOG.warning("OS doesn't present the host nqn information. "
"Controller may be incorrectly matched.")
warned = True
ctrl_portal = (ctrl_addr, ctrl_port, ctrl_transport, ctrl_hostnqn)
try:
index = sysfs_portals.index(ctrl_portal)
LOG.debug('Found a valid portal at %s', ctrl_portal)
# Update the portal with the found controller name
self.portals[index].controller = ctrl_name
known_names.append(ctrl_name) # One more controller found
except ValueError:
# If it's not one of our controllers ignore it
LOG.debug('Skipping %s, not part of portals %s',
ctrl_portal, sysfs_portals)
# short circuit if no more portals to find
if len(known_names) == len(sysfs_portals):
return
def get_devices(self, only_live=False, get_one=False) -> list[str]:
"""Return devices for this target
Optionally only return devices from portals that are live and also
optionally return on first device found.
Returns an empty list when not found.
"""
LOG.debug('Looking for volume at %s', self.nqn)
# Use a set because multiple portals can return the same device when
# using ANA (even if we are not intentionally doing multipathing)
result = set()
portals = self.live_portals if only_live else self.present_portals
for portal in portals:
device = portal.get_device()
if device:
result.add(device)
if get_one:
break
return list(result)
# NOTE: Don't change to a property, as it would hide VolumeDeviceNotFound
@utils.retry(exception.VolumeDeviceNotFound, retries=5)
def find_device(self) -> str:
"""Search for a device that is on a live subsystem
Must have called set_portals_controllers first since portals without a
controller name will be skipped.
Retries up to 5 times with exponential backoff to give time in case the
subsystem is currently reconnecting. Raises VolumeDeviceNotFound when
finally gives up trying.
"""
if not self._device:
devices = self.get_devices(only_live=True, get_one=True)
if not devices:
raise exception.VolumeDeviceNotFound(device=self.nqn)
self._device = devices[0]
return self._device
def get_device_path_by_initial_devices(self) -> Optional[str]:
"""Find target's device path from devices that were present before."""
ctrls = [p.controller for p in self.portals if p.controller]
def discard(devices):
"""Discard devices that don't belong to our controllers."""
if not devices:
return set()
return set(dev for dev in devices
if os.path.basename(dev).rsplit('n', 1)[0] in ctrls)
current_devices = self._get_nvme_devices()
LOG.debug('Initial devices: %s. Current devices %s. Controllers: %s',
self.devices_on_start, current_devices, ctrls)
devices = discard(current_devices) - discard(self.devices_on_start)
if not devices:
return None
if len(devices) == 1:
return devices.pop()
# With multiple devices they must all have the same uuid
if (len(devices) > 1 and
1 < len(set(blk_property('uuid', os.path.basename(d))
for d in devices))):
msg = _('Too many different volumes found for %s') % ctrls
LOG.error(msg)
return None
return devices.pop() # They are the same volume, return any of them
class NVMeOFConnProps(object):
"""Internal representation of the NVMe-oF connection properties
There is an old and a newer connection properties format, which result
in 2 variants for replicated connections and 2 for non replicated:
1- New format with multiples replicas information
2- New format with single replica information
3- New format with no replica information
4- Original format
Case #1 and #2 format:
{
'vol_uuid': <cinder_volume_id>,
'alias': <raid_alias>,
'volume_replicas': [ <target>, ... ],
'replica_count': len(volume_replicas),
}
Where:
cinder_volume_id ==> Cinder id, could be different from NVMe UUID.
with/without hyphens, uppper/lower cased.
target :== {
'target_nqn': <target_nqn>,
'vol_uuid': <nvme_volume_uuid>,
'portals': [ <portal>, ... ],
}
nvme_volume_uuid ==> NVMe UUID. Can be different than cinder's id.
With/without hyphens, uppper/lower cased
portal ::= tuple/list(
<target_portal>,
<target_port>,
<transport_type>
)
transport_type ::= 'RoCEv2' | <anything> # anything => tcp
Case #3 format:
<target> ==> As defined in case #1 & #2
Case #4 format:
{
'nqn': <nqn>,
'transport_type': <transport_type>,
'target_portal': <target_address>,
'target_port': <target_port>,
'volume_nguid': <volume_nguid>,
'ns_id': <target_namespace_id>,
'host_nqn': <connector_host_nqn>,
}
Where:
transport_type ::= 'rdma' | 'tcp'
volume_nguid ==> Optional, with/without hyphens, uppper/lower cased
target_namespace_id ==> Optional
connector_host_nqn> ==> Optional
This class unifies the representation of all these in the following
attributes:
replica_count: None for non replicated
alias: None for non replicated
cinder_volume_id: None for non replicated
is_replicated: True if replica count > 1, None if count = 1 else False
targets: List of Target instances
device_path: None if not present (it's set by Nova)
In this unification case#4 is treated as case#3 where the vol_uuid is None
and leaving all the additional information in the dictionary. This way non
replicated cases are always handled in the same way and we have a common
<target>" definition for all cases:
target :== {
'target_nqn': <target_nqn>,
'vol_uuid': <nvme_volume_uuid>,
'portals': [ <new_portal>, ... ],
'volume_nguid': <volume_nguid>,
'ns_id': <target_namespace_id>,
'host_nqn': <connector_host_nqn>,
}
new_portal ::= tuple/list(
<target_address>,
<target_port>,
<new_transport_type>
)
new_transport_type ::= 'rdma' | 'tcp'
Portals change the transport_type to the internal representation where:
'RoCEv2' ==> 'rdma'
<else> ==> 'tcp'
This means that the new connection format now accepts vol_uuid set to None,
and accepts ns_id, volume_nguid, and host_nqn parameters, as described in
the connect_volume docstring.
"""
RO = 'ro'
RW = 'rw'
replica_count = None
cinder_volume_id: Optional[str] = None
def __init__(self, conn_props: dict,
find_controllers: bool = False) -> None:
# Generic connection properties fields used by Nova
self.qos_specs = conn_props.get('qos_specs')
self.readonly = conn_props.get('access_mode', self.RW) == self.RO
self.encrypted = conn_props.get('encrypted', False)
self.cacheable = conn_props.get('cacheable', False)
self.discard = conn_props.get('discard', False)
self.enforce_multipath = conn_props.get('enforce_multipath', False)
# old connection properties format
if REPLICAS not in conn_props and NQN not in conn_props:
LOG.debug('Converting old connection info to new format')
conn_props[UUID] = None
conn_props[NQN] = conn_props.pop(OLD_NQN)
conn_props[PORTALS] = [(conn_props.pop(PORTAL),
conn_props.pop(PORT),
conn_props.pop(TRANSPORT))]
# Leave other fields as they are: volume_nguid, ns_id, host_nqn
# NVMe-oF specific fields below
self.alias = conn_props.get(ALIAS)
if REPLICAS in conn_props:
self.replica_count = (conn_props[REPLICA_COUNT] or
len(conn_props[REPLICAS]))
self.is_replicated = True if self.replica_count > 1 else None
targets = conn_props[REPLICAS]
self.cinder_volume_id = str(uuid_lib.UUID(conn_props[UUID]))
else:
self.is_replicated = False
targets = [conn_props]
self.targets = [Target.factory(source_conn_props=self,
find_controllers=find_controllers,
**target) for target in targets]
# Below fields may have been added by nova
self.device_path = conn_props.get('device_path')
def get_devices(self, only_live: bool = False) -> list[str]:
"""Get all device paths present in the system for all targets."""
result = []
for target in self.targets:
result.extend(target.get_devices(only_live))
return result
@classmethod
def from_dictionary_parameter(cls: Type['NVMeOFConnProps'],
func: Callable) -> Callable:
"""Decorator to convert connection properties dictionary.
It converts the connection properties into a NVMeOFConnProps instance
and finds the controller names for all portals present in the system.
"""
@functools.wraps(func)
def wrapper(self, connection_properties, *args, **kwargs):
conn_props = cls(connection_properties, find_controllers=True)
return func(self, conn_props, *args, **kwargs)
return wrapper
# AUXILIARY CLASSES end
# #########################################################
# #########################################################
# CONNECTOR CLASS start
class NVMeOFConnector(base.BaseLinuxConnector):
"""Connector class to attach/detach NVMe-oF volumes."""
# Use a class attribute since a restart is needed to change it on the host
native_multipath_supported = None
# Time we think is more than reasonable to establish an NVMe-oF connection
TIME_TO_CONNECT = 10
def __init__(self,
root_helper: str,
driver: Optional[base.host_driver.HostDriver] = None,
use_multipath: bool = False,
device_scan_attempts: int = DEVICE_SCAN_ATTEMPTS_DEFAULT,
*args, **kwargs) -> None:
super(NVMeOFConnector, self).__init__(
root_helper,
driver,
device_scan_attemps=device_scan_attempts,
*args, **kwargs)
self.use_multipath = use_multipath
self._set_native_multipath_supported()
if self.use_multipath and not self.native_multipath_supported:
LOG.warning('native multipath is not enabled')
@staticmethod
def get_search_path() -> str:
"""Necessary implementation for an os-brick connector."""
return DEV_SEARCH_PATH
def get_volume_paths(
self,
connection_properties: NVMeOFConnProps,
device_info: Optional[dict[str, str]] = None) -> list[str]:
"""Return paths where the volume is present."""
# Priority is on the value returned by connect_volume method
if device_info and device_info.get('path'):
return [device_info['path']]
# Nova may add the path on the connection properties as a device_path
device_path = connection_properties.device_path
if device_path:
return [device_path]
# If we don't get the info on the connection properties it could be
# problematic because we could have multiple devices and not know which
# one we used.
LOG.warning('We are being called without the path information!')
# TODO: For raids it would be good to ensure they are actually there
if connection_properties.is_replicated:
if connection_properties.alias is None:
raise exception.BrickException('Alias missing in connection '
'info')
return [RAID_PATH + connection_properties.alias]
# TODO: Return live devices first?
devices = connection_properties.get_devices()
# If we are not sure if it's replicated or not, find out
if connection_properties.is_replicated is None:
if any(self._is_raid_device(dev) for dev in devices):
if connection_properties.alias is None:
raise exception.BrickException('Alias missing in '
'connection info')
return [RAID_PATH + connection_properties.alias]
return devices
# ####### Connector Properties methods ########
@classmethod
def nvme_present(cls: type) -> bool:
"""Check if the nvme CLI is present."""
try:
priv_rootwrap.custom_execute('nvme', 'version')
return True
except Exception as exc:
if isinstance(exc, OSError) and exc.errno == errno.ENOENT:
LOG.debug('nvme not present on system')
else:
LOG.warning('Unknown error when checking presence of nvme: %s',
exc)
return False
@classmethod
def get_connector_properties(cls, root_helper, *args, **kwargs) -> dict:
"""The NVMe-oF connector properties (initiator uuid and nqn.)"""
execute = kwargs.get('execute') or priv_rootwrap.execute
nvmf = NVMeOFConnector(root_helper=root_helper, execute=execute)
ret = {}
nqn = None
hostid = None
uuid = nvmf._get_host_uuid()
suuid = priv_nvmeof.get_system_uuid()
if cls.nvme_present():
nqn = utils.get_host_nqn(suuid)
# Ensure /etc/nvme/hostid exists and defaults to the system uuid,
# or a random value.
hostid = utils.get_nvme_host_id(suuid)
if hostid:
ret['nvme_hostid'] = hostid
if uuid:
ret['uuid'] = uuid
if suuid:
ret['system uuid'] = suuid # compatibility
if nqn:
ret['nqn'] = nqn
ret['nvme_native_multipath'] = cls._set_native_multipath_supported()
return ret
def _get_host_uuid(self) -> Optional[str]:
"""Get the UUID of the first mounted filesystem."""
cmd = ('findmnt', '-v', '/', '-n', '-o', 'SOURCE')
try:
lines, err = self._execute(
*cmd, run_as_root=True, root_helper=self._root_helper)
source = lines.split('\n')[0]
# In a container this could be 'overlay', which causes the blkid
# command to fail.
if source == "overlay":
return None
blkid_cmd = (
'blkid', source, '-s', 'UUID', '-o', 'value')
lines, _err = self._execute(
*blkid_cmd, run_as_root=True, root_helper=self._root_helper)
return lines.split('\n')[0]
except putils.ProcessExecutionError as e:
LOG.warning(
"Process execution error in _get_host_uuid: %s", e)
return None
@classmethod
def _set_native_multipath_supported(cls):
if cls.native_multipath_supported is None:
cls.native_multipath_supported = \
cls._is_native_multipath_supported()
return cls.native_multipath_supported
@staticmethod
def _is_native_multipath_supported():
try:
with open('/sys/module/nvme_core/parameters/multipath', 'rt') as f:
return f.read().strip() == 'Y'
except Exception:
LOG.warning("Could not find nvme_core/parameters/multipath")
return False
# ####### Connect Volume methods ########
def supports_multipath(self):
return self.native_multipath_supported
@utils.trace
@utils.connect_volume_prepare_result
@base.synchronized('connect_volume', external=True)
@NVMeOFConnProps.from_dictionary_parameter
def connect_volume(
self, connection_properties: NVMeOFConnProps) -> dict[str, str]:
"""Attach and discover the volume."""
self.check_multipath(
{'enforce_multipath': getattr(
connection_properties, 'enforce_multipath', False)})
try:
if connection_properties.is_replicated is False:
LOG.debug('Starting non replicated connection')
path = self._connect_target(connection_properties.targets[0])
else: # If we know it's replicated or we don't yet know
LOG.debug('Starting replicated connection')
path = self._connect_volume_replicated(connection_properties)
except Exception:
self._try_disconnect_all(connection_properties)
raise
return {'type': 'block', 'path': path}
def _do_multipath(self):
return self.use_multipath and self.native_multipath_supported
@utils.retry(exception.VolumeDeviceNotFound, interval=2)
def _connect_target(self, target: Target) -> str:
"""Attach a specific target to present a volume on the system
If we are already connected to any of the portals (and it's live) we
send a rescan (because the backend may not support AER messages),
otherwise we iterate through the portals trying to do an nvme-of
connection.
This method assumes that the controllers for the portals have already
been set. For example using the from_dictionary_parameter decorator
in the NVMeOFConnProps class.
Returns the path of the connected device.
"""
connected = False
missing_portals = []
reconnecting_portals = []
for portal in target.portals:
state = portal.state # store it so we read only once from sysfs
# Rescan live controllers in case backend doesn't support AER
if state == portal.LIVE:
connected = True
self.rescan(portal.controller) # type: ignore
# Remember portals that are not present in the system
elif state == portal.MISSING:
missing_portals.append(portal)
elif state == portal.CONNECTING:
LOG.debug('%s is reconnecting', portal)
reconnecting_portals.append(portal)
# Ignore reconnecting/dead portals
else:
LOG.debug('%s exists but is %s', portal, state)
# If no live portals exist or if we want to use multipath
do_multipath = self._do_multipath()
if do_multipath or not connected:
for portal in missing_portals:
cmd = ['connect', '-a', portal.address, '-s', portal.port,
'-t', portal.transport, '-n', target.nqn, '-Q', '128',
'-l', '-1']
if target.host_nqn:
cmd.extend(['-q', target.host_nqn])
try:
self.run_nvme_cli(cmd)
connected = True
except putils.ProcessExecutionError as exc:
# In some nvme versions code 70 means target is already
# connected, but in newer versions code is EALREADY.
# Those should only happen if there is a race condition
# because something is incorrectly configured (n-cpu and
# c-vol running on same node with different lock paths) or
# an admin is touching things manually. Not passing these
# exit codes in check_exit_code parameter to _execute so we
# can log it. nvme cli v2 returns 1, so we parse the
# message. Some nvme cli versions return errors in stdout,
# so we look in stderr and stdout.
if not (exc.exit_code in (70, errno.EALREADY) or
(exc.exit_code == 1 and
'already connected' in exc.stderr + exc.stdout)):
LOG.error('Could not connect to %s: exit_code: %s, '
'stdout: "%s", stderr: "%s",', portal,
exc.exit_code, exc.stdout, exc.stderr)
continue
LOG.warning('Race condition with some other application '
'when connecting to %s, please check your '
'system configuration.', portal)
state = portal.state
if state == portal.LIVE:
connected = True
elif state == portal.CONNECTING:
reconnecting_portals.append(portal)
else:
LOG.error('Ignoring %s due to unknown state (%s)',
portal, state)
if not do_multipath:
break # We are connected
if not connected and reconnecting_portals:
delay = self.TIME_TO_CONNECT + max(p.reconnect_delay
for p in reconnecting_portals)
LOG.debug('Waiting %s seconds for some nvme controllers to '
'reconnect', delay)
timeout = time.time() + delay
while time.time() < timeout:
time.sleep(1)
if any(p.is_live for p in reconnecting_portals):
LOG.debug('Reconnected')
connected = True
break
LOG.debug('No controller reconnected')
if not connected:
raise exception.VolumeDeviceNotFound(device=target.nqn)
# Ensure controller names of new connections are set
target.set_portals_controllers()
dev_path = target.find_device()
return dev_path
@utils.trace
def _connect_volume_replicated(
self, connection_properties: NVMeOFConnProps) -> str:
"""Connect to a replicated volume and prepare the RAID
Connection properties must contain all the necessary replica
information, even if there is only 1 replica.
Returns the /dev/md path
Raises VolumeDeviceNotFound when cannot present the device in the
system.
"""
host_device_paths = []
if not connection_properties.alias:
raise exception.BrickException('Alias missing in connection info')
for replica in connection_properties.targets:
try:
rep_host_device_path = self._connect_target(replica)
host_device_paths.append(rep_host_device_path)
except Exception as ex:
LOG.error("_connect_target: %s", ex)
if not host_device_paths:
raise exception.VolumeDeviceNotFound(
device=connection_properties.targets)
if connection_properties.is_replicated:
device_path = self._handle_replicated_volume(
host_device_paths, connection_properties)
else:
device_path = self._handle_single_replica(
host_device_paths, connection_properties.alias)
if nvmeof_agent:
nvmeof_agent.NVMeOFAgent.ensure_running(self)
return device_path
def _handle_replicated_volume(self,
host_device_paths: list[str],
conn_props: NVMeOFConnProps) -> str:
"""Assemble the raid from found devices."""
path_in_raid = False
for dev_path in host_device_paths:
path_in_raid = self._is_device_in_raid(dev_path)
if path_in_raid:
break
device_path = RAID_PATH + conn_props.alias # type: ignore
if path_in_raid:
self.stop_and_assemble_raid(host_device_paths, device_path, False)
else:
paths_found = len(host_device_paths)
if conn_props.replica_count > paths_found: # type: ignore
LOG.error(
'Cannot create MD as %s out of %s legs were found.',
paths_found, conn_props.replica_count)
raise exception.VolumeDeviceNotFound(device=conn_props.alias)
self.create_raid(host_device_paths, '1',
conn_props.alias, # type: ignore
conn_props.alias, # type: ignore
False)
return device_path
def _handle_single_replica(self,
host_device_paths: list[str],
volume_alias: str) -> str:
"""Assemble the raid from a single device."""
if self._is_raid_device(host_device_paths[0]):
md_path = RAID_PATH + volume_alias
self.stop_and_assemble_raid(host_device_paths, md_path, False)
return md_path
return host_device_paths[0]
# ####### Disconnect methods ########
@utils.trace
@base.synchronized('connect_volume', external=True)
@utils.connect_volume_undo_prepare_result(unlink_after=True)
def disconnect_volume(self,
connection_properties: dict,
device_info: dict[str, str],
force: bool = False,
ignore_errors: bool = False) -> None:
"""Flush the volume.
Disconnect of volumes happens on storage system side. Here we could
remove connections to subsystems if no volumes are left. But new
volumes can pop up asynchronously in the meantime. So the only thing
left is flushing or disassembly of a correspondng RAID device.
:param connection_properties: The dictionary that describes all
of the target volume attributes as
described in connect_volume but also with
the "device_path" key containing the path
to the volume that was connected (this is
added by Nova).
:type connection_properties: dict
:param device_info: historical difference, but same as connection_props
:type device_info: dict
"""
# NOTE: Cannot use NVMeOFConnProps's decorator to create instance from
# conn props because "connect_volume_undo_prepare_result" must be the
# first decorator and it expects a dictionary.
conn_props = NVMeOFConnProps(connection_properties)
try:
device_path = self.get_volume_paths(conn_props, device_info)[0]
except IndexError:
LOG.warning(
"Cannot find the device for %s, assuming it's not there.",
conn_props.cinder_volume_id or conn_props.targets[0].nqn)
return
exc = exception.ExceptionChainer()
if not os.path.exists(device_path):
LOG.warning("Trying to disconnect device %(device_path)s, but "
"it is not connected. Skipping.",
{'device_path': device_path})
return
# We assume that raid devices are flushed when ending the raid
if device_path.startswith(RAID_PATH):
with exc.context(force, 'Failed to end raid %s', device_path):
self.end_raid(device_path)
else:
with exc.context(force, 'Failed to flush %s', device_path):
self._linuxscsi.flush_device_io(device_path)
self._try_disconnect_all(conn_props, exc)
if exc:
LOG.warning('There were errors removing %s', device_path)
if not ignore_errors:
raise exc
def _try_disconnect_all(
self,
conn_props: NVMeOFConnProps,
exc: Optional[exception.ExceptionChainer] = None) -> None:
"""Disconnect all subsystems that are not being used.
Only sees if it has to disconnect this connection properties portals,
leaves other alone.
Since this is unrelated to the flushing of the devices failures will be
logged but they won't be raised.
"""
if exc is None:
exc = exception.ExceptionChainer()
for target in conn_props.targets:
# Associate each portal with its controller name
target.set_portals_controllers()
for portal in target.portals:
# Ignore exceptions to disconnect as many as possible.
with exc.context(True, 'Failed to disconnect %s', portal):
self._try_disconnect(portal)
def _try_disconnect(self, portal: Portal) -> None:
"""Disconnect a specific subsystem if it's safe.
Only disconnect if it has no namespaces left or has only one left and
it is from this connection.
"""
LOG.debug('Checking if %s can be disconnected', portal)
if portal.can_disconnect():
self._execute('nvme', 'disconnect',
'-d', '/dev/' + portal.controller, # type: ignore
root_helper=self._root_helper, run_as_root=True)
# ####### Extend methods ########
@staticmethod
def _get_sizes_from_lba(ns_data: dict) -> tuple[Optional[int],
Optional[int]]:
"""Return size in bytes and the nsze of the volume from NVMe NS data.
nsze is the namespace size that defines the total size of the namespace
in logical blocks (LBA 0 through n-1), as per NVMe-oF specs.
Returns a tuple of nsze and size
"""
try:
lbads = ns_data['lbafs'][0]['ds']
# Don't know what to do with more than 1 LBA and as per NVMe specs
# if LBADS < 9 then LBA is not supported
if len(ns_data['lbafs']) != 1 or lbads < 9:
LOG.warning("Cannot calculate new size with LBAs")
return None, None
nsze = ns_data['nsze']
new_size = nsze * (1 << lbads)
except Exception:
return None, None
LOG.debug('New volume size is %s and nsze is %s', new_size, nsze)
return nsze, new_size
@utils.trace
@base.synchronized('extend_volume', external=True)
@utils.connect_volume_undo_prepare_result
def extend_volume(self, connection_properties: dict[str, str]) -> int:
"""Update an attached volume to reflect the current size after extend
The only way to reflect the new size of an NVMe-oF volume on the host
is a rescan, which rescans the whole subsystem. This is a problem on
attach_volume and detach_volume, but not here, since we will have at
least the namespace we are operating on in the subsystem.
The tricky part is knowing when a rescan has already been completed and
the volume size on sysfs is final. The rescan may already have
happened before this method is called due to an AER message or we may
need to trigger it here.
Scans can be triggered manually with 'nvme ns-rescan' or writing 1 in
configf's rescan file, or they can be triggered indirectly when calling
the 'nvme list', 'nvme id-ns', or even using the 'nvme admin-passthru'
command.
Even after getting the new size with any of the NVMe commands above, we
still need to wait until this is reflected on the host device, because
we cannot return to the caller until the new size is in effect.
If we don't see the new size taking effect on the system after 5
seconds, or if we cannot get the new size with nvme, then we rescan in
the latter and in both cases we blindly wait 5 seconds and return
whatever size is present.
For replicated volumes, the RAID needs to be extended.
"""
# NOTE: Cannot use NVMeOFConnProps's decorator to create instance from
# conn props because "connect_volume_undo_prepare_result" must be the
# first decorator and it expects a dictionary.
conn_props = NVMeOFConnProps(connection_properties)
try:
device_path = self.get_volume_paths(conn_props)[0]
except IndexError:
raise exception.VolumeDeviceNotFound()
# Replicated needs to grow the raid, even if there's only 1 device
if device_path.startswith(RAID_PATH):
# NOTE: May not work without backend AER support and may have races
self.run_mdadm(('mdadm', '--grow', '--size', 'max', device_path))
else:
dev_name = os.path.basename(device_path)
ctrl_name = dev_name.rsplit('n', 1)[0]
nsze: Optional[Union[str, int]] = None
try:
# With many devices, id-ns command generates less data
out, err = self._execute('nvme', 'id-ns', '-ojson',
device_path, run_as_root=True,
root_helper=self._root_helper)
ns_data = json.loads(out)
nsze, new_size = self._get_sizes_from_lba(ns_data)
except Exception as exc:
LOG.warning('Failed to get id-ns %s', exc)
# Assume that nvme command failed, so didn't scan
self.rescan(ctrl_name)
if nsze: # Wait for the system to reflect the new size
nsze = str(nsze) # To compare with contents of sysfs
for x in range(10): # Wait 5 secs for size to appear in sysfs
current_nsze = blk_property('size', dev_name)
if current_nsze == nsze:
return new_size # type: ignore
LOG.debug('Sysfs size is still %s', current_nsze)
time.sleep(0.5)
LOG.warning('Timeout waiting for sysfs to reflect the right '
'volume size.')
# Last resort when id-ns failed or system didn't reflect new size
LOG.info('Wait 5 seconds and return whatever size is present')
time.sleep(5)
size = utils.get_device_size(self, device_path)
if size is None:
raise exception.BrickException(
'get_device_size returned non-numeric size')
return size
# ####### RAID methods ########
def run_mdadm(self,
cmd: Sequence[str],
raise_exception: bool = False) -> Optional[str]:
cmd_output = None
try:
lines, err = self._execute(
*cmd, run_as_root=True, root_helper=self._root_helper)
for line in lines.split('\n'):
cmd_output = line
break
except putils.ProcessExecutionError as ex:
LOG.warning("[!] Could not run mdadm: %s", str(ex))
if raise_exception:
raise ex
return cmd_output
def _is_device_in_raid(self, device_path: str) -> bool:
cmd = ['mdadm', '--examine', device_path]
raid_expected = device_path + ':'
try:
lines, err = self._execute(
*cmd, run_as_root=True, root_helper=self._root_helper)
for line in lines.split('\n'):
if line == raid_expected:
return True
else:
return False
except putils.ProcessExecutionError:
pass
return False
@staticmethod
def ks_readlink(dest: str) -> str:
try:
return os.readlink(dest)
except Exception:
return ''
@staticmethod
def get_md_name(device_name: str) -> Optional[str]:
try:
with open('/proc/mdstat', 'r') as f:
lines = [line.split(' ')[0]
for line in f
if device_name in line]
if lines:
return lines[0]
except Exception as exc:
LOG.debug("[!] Could not find md name for %s in mdstat: %s",
device_name, exc)
return None
def stop_and_assemble_raid(self,
drives: list[str],
md_path: str,
read_only: bool) -> None:
md_name = None
i = 0
assembled = False
link = ''
while i < 5 and not assembled:
for drive in drives:
device_name = drive[5:]
md_name = self.get_md_name(device_name)
link = NVMeOFConnector.ks_readlink(md_path)
if link != '':
link = os.path.basename(link)
if md_name and md_name == link:
return
LOG.debug(
"sleeping 1 sec -allow auto assemble link = %(link)s "
"md path = %(md_path)s",
{'link': link, 'md_path': md_path})
time.sleep(1)
if md_name and md_name != link:
self.stop_raid(md_name)
try:
assembled = self.assemble_raid(drives, md_path, read_only)
except Exception:
i += 1
def assemble_raid(self,
drives: list[str],
md_path: str,
read_only: bool) -> bool:
cmd = ['mdadm', '--assemble', '--run', md_path]
if read_only:
cmd.append('-o')
for i in range(len(drives)):
cmd.append(drives[i])
try:
self.run_mdadm(cmd, True)
except putils.ProcessExecutionError as ex:
LOG.warning("[!] Could not _assemble_raid: %s", str(ex))
raise ex
return True
def create_raid(self,
drives: list[str],
raid_type: str,
device_name: str,
name: str,
read_only: bool) -> None:
cmd = ['mdadm']
num_drives = len(drives)
cmd.append('-C')
if read_only:
cmd.append('-o')
cmd.append(device_name)
cmd.append('-R')
if name:
cmd.append('-N')
cmd.append(name)
cmd.append('--level')
cmd.append(raid_type)
cmd.append('--raid-devices=' + str(num_drives))
cmd.append('--bitmap=internal')
cmd.append('--homehost=any')
cmd.append('--failfast')
cmd.append('--assume-clean')
for i in range(len(drives)):
cmd.append(drives[i])
LOG.debug('[!] cmd = %s', cmd)
self.run_mdadm(cmd)
# sometimes under load, md is not created right away so we wait
for i in range(60):
try:
is_exist = os.path.exists(RAID_PATH + name)
LOG.debug("[!] md is_exist = %s", is_exist)
if is_exist:
return
time.sleep(1)
except Exception:
LOG.debug('[!] Exception_wait_raid!')
msg = _("md: /dev/md/%s not found.") % name
LOG.error(msg)
raise exception.NotFound(message=msg)
def end_raid(self, device_path: str) -> None:
raid_exists = self.is_raid_exists(device_path)
if raid_exists:
for i in range(10):
try:
cmd_out = self.stop_raid(device_path, True)
if not cmd_out:
break
except Exception:
time.sleep(1)
try:
is_exist = os.path.exists(device_path)
LOG.debug("[!] is_exist = %s", is_exist)
if is_exist:
self.remove_raid(device_path)
os.remove(device_path)
except Exception:
LOG.debug('[!] Exception_stop_raid!')
def stop_raid(self,
md_path: str,
raise_exception: bool = False) -> Optional[str]:
cmd = ['mdadm', '--stop', md_path]
LOG.debug("[!] cmd = %s", cmd)
cmd_out = self.run_mdadm(cmd, raise_exception)
return cmd_out
def is_raid_exists(self, device_path: str) -> bool:
cmd = ['mdadm', '--detail', device_path]
LOG.debug("[!] cmd = %s", cmd)
raid_expected = device_path + ':'
try:
lines, err = self._execute(
*cmd, run_as_root=True, root_helper=self._root_helper)
for line in lines.split('\n'):
LOG.debug("[!] line = %s", line)
if line == raid_expected:
return True
else:
return False
except putils.ProcessExecutionError:
pass
return False
def remove_raid(self, device_path: str) -> None:
cmd = ['mdadm', '--remove', device_path]
LOG.debug("[!] cmd = %s", cmd)
self.run_mdadm(cmd)
def _is_raid_device(self, device: str) -> bool:
return self._get_fs_type(device) == 'linux_raid_member'
def _get_fs_type(self, device_path: str) -> Optional[str]:
cmd = ['blkid', device_path, '-s', 'TYPE', '-o', 'value']
LOG.debug("[!] cmd = %s", cmd)
fs_type = None
# We don't care about errors, on error lines will be '' so it's ok
lines, err = self._execute(
*cmd, run_as_root=True, root_helper=self._root_helper,
check_exit_code=False)
fs_type = lines.split('\n')[0]
return fs_type or None
# ####### NVMe methods ########
def run_nvme_cli(self,
nvme_command: Sequence[str],
**kwargs) -> tuple[str, str]:
"""Run an nvme cli command and return stdout and stderr output."""
(out, err) = self._execute('nvme', *nvme_command, run_as_root=True,
root_helper=self._root_helper,
check_exit_code=True)
msg = ("nvme %(nvme_command)s: stdout=%(out)s stderr=%(err)s" %
{'nvme_command': nvme_command, 'out': out, 'err': err})
LOG.debug("[!] %s", msg)
return out, err
def rescan(self, controller_name: str) -> None:
"""Rescan an nvme controller."""
nvme_command = ('ns-rescan', DEV_SEARCH_PATH + controller_name)
try:
self.run_nvme_cli(nvme_command)
except Exception as e:
raise exception.CommandExecutionFailed(e, cmd=nvme_command)
# CONNECTOR CLASS end
# #########################################################
View Git Blame Copy Permalink