path: root/src/pybind/mgr/cephadm/serve.py

import ipaddress
import hashlib
import json
import logging
import uuid
import os
from collections import defaultdict
from typing import TYPE_CHECKING, Optional, List, cast, Dict, Any, Union, Tuple, Set, \
    DefaultDict, Callable

from ceph.deployment import inventory
from ceph.deployment.drive_group import DriveGroupSpec
from ceph.deployment.service_spec import (
    ArgumentList,
    ArgumentSpec,
    CustomContainerSpec,
    PlacementSpec,
    RGWSpec,
    ServiceSpec,
    IngressSpec,
)
from ceph.utils import datetime_now

import orchestrator
from orchestrator import OrchestratorError, set_exception_subject, OrchestratorEvent, \
    DaemonDescriptionStatus, daemon_type_to_service
from cephadm.services.cephadmservice import CephadmDaemonDeploySpec
from cephadm.schedule import HostAssignment
from cephadm.autotune import MemoryAutotuner
from cephadm.utils import forall_hosts, cephadmNoImage, is_repo_digest, \
    CephadmNoImage, CEPH_TYPES, ContainerInspectInfo, SpecialHostLabels
from mgr_module import MonCommandFailed
from mgr_util import format_bytes, verify_tls, get_cert_issuer_info, ServerConfigException

from . import utils
from . import exchange

if TYPE_CHECKING:
    from cephadm.module import CephadmOrchestrator

logger = logging.getLogger(__name__)

REQUIRES_POST_ACTIONS = ['grafana', 'iscsi', 'prometheus', 'alertmanager', 'rgw']


class CephadmServe:
    """
    This module contains functions that are executed in the
    serve() thread. Thus they don't block the CLI.

    Please see the `Note regarding network calls from CLI handlers`
    chapter in the cephadm developer guide.

    On the other hand, These function should *not* be called form
    CLI handlers, to avoid blocking the CLI
    """

    def __init__(self, mgr: "CephadmOrchestrator"):
        self.mgr: "CephadmOrchestrator" = mgr
        self.log = logger

    def serve(self) -> None:
        """
        The main loop of cephadm.

        A command handler will typically change the declarative state
        of cephadm. This loop will then attempt to apply this new state.
        """
        self.log.debug("serve starting")

        while self.mgr.run:
            self.log.debug("serve loop start")

            try:

                self.convert_tags_to_repo_digest()

                # refresh daemons
                self.log.debug('refreshing hosts and daemons')
                self._refresh_hosts_and_daemons()

                self._check_for_strays()

                self._update_paused_health()

                if self.mgr.need_connect_dashboard_rgw and self.mgr.config_dashboard:
                    self.mgr.need_connect_dashboard_rgw = False
                    if 'dashboard' in self.mgr.get('mgr_map')['modules']:
                        self.log.info('Checking dashboard <-> RGW credentials')
                        self.mgr.remote('dashboard', 'set_rgw_credentials')

                if not self.mgr.paused:
                    self._run_async_actions()

                    self.mgr.to_remove_osds.process_removal_queue()

                    self.mgr.migration.migrate()
                    if self.mgr.migration.is_migration_ongoing():
                        continue

                    if self._apply_all_services():
                        continue  # did something, refresh

                    self._check_daemons()

                    self._check_certificates()

                    self._purge_deleted_services()

                    self._check_for_moved_osds()

                    if self.mgr.agent_helpers._handle_use_agent_setting():
                        continue

                    if self.mgr.node_proxy_service.handle_hw_monitoring_setting():
                        continue

                    if self.mgr.upgrade.continue_upgrade():
                        continue

                    # refresh node-proxy cache
                    self.mgr.node_proxy_cache.load()

            except OrchestratorError as e:
                if e.event_subject:
                    self.mgr.events.from_orch_error(e)

            self.log.debug("serve loop sleep")
            self._serve_sleep()
            self.log.debug("serve loop wake")
        self.log.debug("serve exit")

    def _check_certificates(self) -> None:
        for d in self.mgr.cache.get_daemons_by_type('grafana'):
            cert = self.mgr.get_store(f'{d.hostname}/grafana_crt')
            key = self.mgr.get_store(f'{d.hostname}/grafana_key')
            if (not cert or not cert.strip()) and (not key or not key.strip()):
                # certificate/key are empty... nothing to check
                return

            try:
                get_cert_issuer_info(cert)
                verify_tls(cert, key)
                self.mgr.remove_health_warning('CEPHADM_CERT_ERROR')
            except ServerConfigException as e:
                err_msg = f"""
                Detected invalid grafana certificates. Please, use the following commands:

                  > ceph config-key set mgr/cephadm/{d.hostname}/grafana_crt -i <path-to-ctr-file>
                  > ceph config-key set mgr/cephadm/{d.hostname}/grafana_key -i <path-to-key-file>

                to set valid key and certificate or reset their value to an empty string
                in case you want cephadm to generate self-signed Grafana certificates.

                Once done, run the following command to reconfig the daemon:

                  > ceph orch daemon reconfig grafana.{d.hostname}

                """
                self.log.error(f'Detected invalid grafana certificate on host {d.hostname}: {e}')
                self.mgr.set_health_warning('CEPHADM_CERT_ERROR',
                                            f'Invalid grafana certificate on host {d.hostname}: {e}',
                                            1, [err_msg])
                break

    def _serve_sleep(self) -> None:
        sleep_interval = max(
            30,
            min(
                self.mgr.host_check_interval,
                self.mgr.facts_cache_timeout,
                self.mgr.daemon_cache_timeout,
                self.mgr.device_cache_timeout,
            )
        )
        self.log.debug('Sleeping for %d seconds', sleep_interval)
        self.mgr.event.wait(sleep_interval)
        self.mgr.event.clear()

    def _update_paused_health(self) -> None:
        self.log.debug('_update_paused_health')
        if self.mgr.paused:
            self.mgr.set_health_warning('CEPHADM_PAUSED', 'cephadm background work is paused', 1, [
                                        "'ceph orch resume' to resume"])
        else:
            self.mgr.remove_health_warning('CEPHADM_PAUSED')

    def _autotune_host_memory(self, host: str) -> None:
        total_mem = self.mgr.cache.get_facts(host).get('memory_total_kb', 0)
        if not total_mem:
            val = None
        else:
            total_mem *= 1024   # kb -> bytes
            total_mem *= self.mgr.autotune_memory_target_ratio
            a = MemoryAutotuner(
                daemons=self.mgr.cache.get_daemons_by_host(host),
                config_get=self.mgr.get_foreign_ceph_option,
                total_mem=total_mem,
            )
            val, osds = a.tune()
            any_changed = False
            for o in osds:
                if self.mgr.get_foreign_ceph_option(o, 'osd_memory_target') != val:
                    self.mgr.check_mon_command({
                        'prefix': 'config rm',
                        'who': o,
                        'name': 'osd_memory_target',
                    })
                    any_changed = True
        if val is not None:
            if any_changed:
                self.mgr.log.info(
                    f'Adjusting osd_memory_target on {host} to {format_bytes(val, 6)}'
                )
                ret, out, err = self.mgr.mon_command({
                    'prefix': 'config set',
                    'who': f'osd/host:{host.split(".")[0]}',
                    'name': 'osd_memory_target',
                    'value': str(val),
                })
                if ret:
                    self.log.warning(
                        f'Unable to set osd_memory_target on {host} to {val}: {err}'
                    )
        else:
            # if osd memory autotuning is off, we don't want to remove these config
            # options as users may be using them. Since there is no way to set autotuning
            # on/off at a host level, best we can do is check if it is globally on.
            if self.mgr.get_foreign_ceph_option('osd', 'osd_memory_target_autotune'):
                self.mgr.check_mon_command({
                    'prefix': 'config rm',
                    'who': f'osd/host:{host.split(".")[0]}',
                    'name': 'osd_memory_target',
                })
        self.mgr.cache.update_autotune(host)

    def _refresh_hosts_and_daemons(self) -> None:
        self.log.debug('_refresh_hosts_and_daemons')
        bad_hosts = []
        failures = []
        agents_down: List[str] = []

        @forall_hosts
        def refresh(host: str) -> None:

            # skip hosts that are in maintenance - they could be powered off
            if self.mgr.inventory._inventory[host].get("status", "").lower() == "maintenance":
                return

            if self.mgr.use_agent:
                if self.mgr.agent_helpers._check_agent(host):
                    agents_down.append(host)

            if self.mgr.cache.host_needs_check(host):
                r = self._check_host(host)
                if r is not None:
                    bad_hosts.append(r)

            if (
                not self.mgr.use_agent
                or self.mgr.cache.is_host_draining(host)
                or host in agents_down
            ):
                if self.mgr.cache.host_needs_daemon_refresh(host):
                    self.log.debug('refreshing %s daemons' % host)
                    r = self._refresh_host_daemons(host)
                    if r:
                        failures.append(r)

                if self.mgr.cache.host_needs_facts_refresh(host):
                    self.log.debug(('Refreshing %s facts' % host))
                    r = self._refresh_facts(host)
                    if r:
                        failures.append(r)

                if self.mgr.cache.host_needs_network_refresh(host):
                    self.log.debug(('Refreshing %s networks' % host))
                    r = self._refresh_host_networks(host)
                    if r:
                        failures.append(r)

                if self.mgr.cache.host_needs_device_refresh(host):
                    self.log.debug('refreshing %s devices' % host)
                    r = self._refresh_host_devices(host)
                    if r:
                        failures.append(r)
                self.mgr.cache.metadata_up_to_date[host] = True
            elif not self.mgr.cache.get_daemons_by_type('agent', host=host):
                if self.mgr.cache.host_needs_daemon_refresh(host):
                    self.log.debug('refreshing %s daemons' % host)
                    r = self._refresh_host_daemons(host)
                    if r:
                        failures.append(r)
                self.mgr.cache.metadata_up_to_date[host] = True

            if self.mgr.cache.host_needs_registry_login(host) and self.mgr.get_store('registry_credentials'):
                self.log.debug(f"Logging `{host}` into custom registry")
                with self.mgr.async_timeout_handler(host, 'cephadm registry-login'):
                    r = self.mgr.wait_async(self._registry_login(
                        host, json.loads(str(self.mgr.get_store('registry_credentials')))))
                if r:
                    bad_hosts.append(r)

            if self.mgr.cache.host_needs_osdspec_preview_refresh(host):
                self.log.debug(f"refreshing OSDSpec previews for {host}")
                r = self._refresh_host_osdspec_previews(host)
                if r:
                    failures.append(r)

            if (
                    self.mgr.cache.host_needs_autotune_memory(host)
                    and not self.mgr.inventory.has_label(host, SpecialHostLabels.NO_MEMORY_AUTOTUNE)
            ):
                self.log.debug(f"autotuning memory for {host}")
                self._autotune_host_memory(host)

        refresh(self.mgr.cache.get_hosts())

        self._write_all_client_files()

        self.mgr.agent_helpers._update_agent_down_healthcheck(agents_down)
        self.mgr.http_server.config_update()

        if self.mgr.config_checks_enabled:
            self.mgr.config_checker.load_network_config()
            self.mgr.config_checker.run_checks()

        for k in [
                'CEPHADM_HOST_CHECK_FAILED',
                'CEPHADM_REFRESH_FAILED',
        ]:
            self.mgr.remove_health_warning(k)
        if bad_hosts:
            self.mgr.set_health_warning(
                'CEPHADM_HOST_CHECK_FAILED', f'{len(bad_hosts)} hosts fail cephadm check', len(bad_hosts), bad_hosts)
        if failures:
            self.mgr.set_health_warning(
                'CEPHADM_REFRESH_FAILED', 'failed to probe daemons or devices', len(failures), failures)
        self.mgr.update_failed_daemon_health_check()

    def _check_host(self, host: str) -> Optional[str]:
        if host not in self.mgr.inventory:
            return None
        self.log.debug(' checking %s' % host)
        try:
            addr = self.mgr.inventory.get_addr(host) if host in self.mgr.inventory else host
            with self.mgr.async_timeout_handler(host, 'cephadm check-host'):
                out, err, code = self.mgr.wait_async(self._run_cephadm(
                    host, cephadmNoImage, 'check-host', [],
                    error_ok=True, no_fsid=True, log_output=self.mgr.log_refresh_metadata))
            self.mgr.cache.update_last_host_check(host)
            self.mgr.cache.save_host(host)
            if code:
                self.log.debug(' host %s (%s) failed check' % (host, addr))
                if self.mgr.warn_on_failed_host_check:
                    return 'host %s (%s) failed check: %s' % (host, addr, err)
            else:
                self.log.debug(' host %s (%s) ok' % (host, addr))
        except Exception as e:
            self.log.debug(' host %s (%s) failed check' % (host, addr))
            return 'host %s (%s) failed check: %s' % (host, addr, e)
        return None

    def _refresh_host_daemons(self, host: str) -> Optional[str]:
        try:
            with self.mgr.async_timeout_handler(host, 'cephadm ls'):
                ls = self.mgr.wait_async(self._run_cephadm_json(
                    host, 'mon', 'ls', [], no_fsid=True, log_output=self.mgr.log_refresh_metadata))
        except OrchestratorError as e:
            return str(e)
        self.mgr._process_ls_output(host, ls)
        return None

    def _refresh_facts(self, host: str) -> Optional[str]:
        try:
            with self.mgr.async_timeout_handler(host, 'cephadm gather-facts'):
                val = self.mgr.wait_async(self._run_cephadm_json(
                    host, cephadmNoImage, 'gather-facts', [],
                    no_fsid=True, log_output=self.mgr.log_refresh_metadata))
        except OrchestratorError as e:
            return str(e)

        self.mgr.cache.update_host_facts(host, val)

        return None

    def _refresh_host_devices(self, host: str) -> Optional[str]:
        with_lsm = self.mgr.device_enhanced_scan
        list_all = self.mgr.inventory_list_all
        inventory_args = ['--', 'inventory',
                          '--format=json-pretty',
                          '--filter-for-batch']
        if with_lsm:
            inventory_args.insert(-1, "--with-lsm")
        if list_all:
            inventory_args.insert(-1, "--list-all")

        try:
            try:
                with self.mgr.async_timeout_handler(host, 'cephadm ceph-volume -- inventory'):
                    devices = self.mgr.wait_async(self._run_cephadm_json(
                        host, 'osd', 'ceph-volume', inventory_args, log_output=self.mgr.log_refresh_metadata))
            except OrchestratorError as e:
                if 'unrecognized arguments: --filter-for-batch' in str(e):
                    rerun_args = inventory_args.copy()
                    rerun_args.remove('--filter-for-batch')
                    with self.mgr.async_timeout_handler(host, 'cephadm ceph-volume -- inventory'):
                        devices = self.mgr.wait_async(self._run_cephadm_json(
                            host, 'osd', 'ceph-volume', rerun_args, log_output=self.mgr.log_refresh_metadata))
                else:
                    raise

        except OrchestratorError as e:
            return str(e)

        self.log.debug('Refreshed host %s devices (%d)' % (
            host, len(devices)))
        ret = inventory.Devices.from_json(devices)
        self.mgr.cache.update_host_devices(host, ret.devices)
        self.update_osdspec_previews(host)
        self.mgr.cache.save_host(host)
        return None

    def _refresh_host_networks(self, host: str) -> Optional[str]:
        try:
            with self.mgr.async_timeout_handler(host, 'cephadm list-networks'):
                networks = self.mgr.wait_async(self._run_cephadm_json(
                    host, 'mon', 'list-networks', [], no_fsid=True, log_output=self.mgr.log_refresh_metadata))
        except OrchestratorError as e:
            return str(e)

        self.log.debug('Refreshed host %s networks (%s)' % (
            host, len(networks)))
        self.mgr.cache.update_host_networks(host, networks)
        self.mgr.cache.save_host(host)
        return None

    def _refresh_host_osdspec_previews(self, host: str) -> Optional[str]:
        self.update_osdspec_previews(host)
        self.mgr.cache.save_host(host)
        self.log.debug(f'Refreshed OSDSpec previews for host <{host}>')
        return None

    def update_osdspec_previews(self, search_host: str = '') -> None:
        # Set global 'pending' flag for host
        self.mgr.cache.loading_osdspec_preview.add(search_host)
        previews = []
        # query OSDSpecs for host <search host> and generate/get the preview
        # There can be multiple previews for one host due to multiple OSDSpecs.
        previews.extend(self.mgr.osd_service.get_previews(search_host))
        self.log.debug(f'Loading OSDSpec previews to HostCache for host <{search_host}>')
        self.mgr.cache.osdspec_previews[search_host] = previews
        # Unset global 'pending' flag for host
        self.mgr.cache.loading_osdspec_preview.remove(search_host)

    def _run_async_actions(self) -> None:
        while self.mgr.scheduled_async_actions:
            (self.mgr.scheduled_async_actions.pop(0))()

    def _check_for_strays(self) -> None:
        self.log.debug('_check_for_strays')
        for k in ['CEPHADM_STRAY_HOST',
                  'CEPHADM_STRAY_DAEMON']:
            self.mgr.remove_health_warning(k)
        if self.mgr.warn_on_stray_hosts or self.mgr.warn_on_stray_daemons:
            ls = self.mgr.list_servers()
            self.log.debug(ls)
            managed = self.mgr.cache.get_daemon_names()
            host_detail = []     # type: List[str]
            host_num_daemons = 0
            daemon_detail = []  # type: List[str]
            for item in ls:
                host = item.get('hostname')
                assert isinstance(host, str)
                daemons = item.get('services')  # misnomer!
                assert isinstance(daemons, list)
                missing_names = []
                for s in daemons:
                    daemon_id = s.get('id')
                    assert daemon_id
                    name = '%s.%s' % (s.get('type'), daemon_id)
                    if s.get('type') in ['rbd-mirror', 'cephfs-mirror', 'rgw', 'rgw-nfs']:
                        metadata = self.mgr.get_metadata(
                            cast(str, s.get('type')), daemon_id, {})
                        assert metadata is not None
                        try:
                            if s.get('type') == 'rgw-nfs':
                                # https://tracker.ceph.com/issues/49573
                                name = metadata['id'][:-4]
                            else:
                                name = '%s.%s' % (s.get('type'), metadata['id'])
                        except (KeyError, TypeError):
                            self.log.debug(
                                "Failed to find daemon id for %s service %s" % (
                                    s.get('type'), s.get('id')
                                )
                            )
                    if s.get('type') == 'tcmu-runner':
                        # because we don't track tcmu-runner daemons in the host cache
                        # and don't have a way to check if the daemon is part of iscsi service
                        # we assume that all tcmu-runner daemons are managed by cephadm
                        managed.append(name)
                    if host not in self.mgr.inventory:
                        missing_names.append(name)
                        host_num_daemons += 1
                    if name not in managed:
                        daemon_detail.append(
                            'stray daemon %s on host %s not managed by cephadm' % (name, host))
                if missing_names:
                    host_detail.append(
                        'stray host %s has %d stray daemons: %s' % (
                            host, len(missing_names), missing_names))
            if self.mgr.warn_on_stray_hosts and host_detail:
                self.mgr.set_health_warning(
                    'CEPHADM_STRAY_HOST', f'{len(host_detail)} stray host(s) with {host_num_daemons} daemon(s) not managed by cephadm', len(host_detail), host_detail)
            if self.mgr.warn_on_stray_daemons and daemon_detail:
                self.mgr.set_health_warning(
                    'CEPHADM_STRAY_DAEMON', f'{len(daemon_detail)} stray daemon(s) not managed by cephadm', len(daemon_detail), daemon_detail)

    def _check_for_moved_osds(self) -> None:
        self.log.debug('_check_for_moved_osds')
        all_osds: DefaultDict[int, List[orchestrator.DaemonDescription]] = defaultdict(list)
        for dd in self.mgr.cache.get_daemons_by_type('osd'):
            assert dd.daemon_id
            all_osds[int(dd.daemon_id)].append(dd)
        for osd_id, dds in all_osds.items():
            if len(dds) <= 1:
                continue
            running = [dd for dd in dds if dd.status == DaemonDescriptionStatus.running]
            error = [dd for dd in dds if dd.status == DaemonDescriptionStatus.error]
            msg = f'Found duplicate OSDs: {", ".join(str(dd) for dd in dds)}'
            logger.info(msg)
            if len(running) != 1:
                continue
            osd = self.mgr.get_osd_by_id(osd_id)
            if not osd or not osd['up']:
                continue
            for e in error:
                assert e.hostname
                try:
                    self._remove_daemon(e.name(), e.hostname, no_post_remove=True)
                    self.mgr.events.for_daemon(
                        e.name(), 'INFO', f"Removed duplicated daemon on host '{e.hostname}'")
                except OrchestratorError as ex:
                    self.mgr.events.from_orch_error(ex)
                    logger.exception(f'failed to remove duplicated daemon {e}')

    def _apply_all_services(self) -> bool:
        self.log.debug('_apply_all_services')
        r = False
        specs = []  # type: List[ServiceSpec]
        # if metadata is not up to date, we still need to apply spec for agent
        # since the agent is the one who gather the metadata. If we don't we
        # end up stuck between wanting metadata to be up to date to apply specs
        # and needing to apply the agent spec to get up to date metadata
        if self.mgr.use_agent and not self.mgr.cache.all_host_metadata_up_to_date():
            self.log.info('Metadata not up to date on all hosts. Skipping non agent specs')
            try:
                specs.append(self.mgr.spec_store['agent'].spec)
            except Exception as e:
                self.log.debug(f'Failed to find agent spec: {e}')
                self.mgr.agent_helpers._apply_agent()
                return r
        else:
            _specs: List[ServiceSpec] = []
            for sn, spec in self.mgr.spec_store.active_specs.items():
                _specs.append(spec)
            # apply specs that don't use count first sice their placement is deterministic
            # and not dependant on other daemon's placements in any way
            specs = [s for s in _specs if not s.placement.count] + [s for s in _specs if s.placement.count]

        for name in ['CEPHADM_APPLY_SPEC_FAIL', 'CEPHADM_DAEMON_PLACE_FAIL']:
            self.mgr.remove_health_warning(name)
        self.mgr.apply_spec_fails = []
        for spec in specs:
            try:
                if self._apply_service(spec):
                    r = True
            except Exception as e:
                msg = f'Failed to apply {spec.service_name()} spec {spec}: {str(e)}'
                self.log.exception(msg)
                self.mgr.events.for_service(spec, 'ERROR', 'Failed to apply: ' + str(e))
                self.mgr.apply_spec_fails.append((spec.service_name(), str(e)))
                warnings = []
                for x in self.mgr.apply_spec_fails:
                    warnings.append(f'{x[0]}: {x[1]}')
                self.mgr.set_health_warning('CEPHADM_APPLY_SPEC_FAIL',
                                            f"Failed to apply {len(self.mgr.apply_spec_fails)} service(s): {','.join(x[0] for x in self.mgr.apply_spec_fails)}",
                                            len(self.mgr.apply_spec_fails),
                                            warnings)
        self.mgr.update_watched_hosts()
        self.mgr.tuned_profile_utils._write_all_tuned_profiles()
        return r

    def _apply_service_config(self, spec: ServiceSpec) -> None:
        if spec.config:
            section = utils.name_to_config_section(spec.service_name())
            for name in ['CEPHADM_INVALID_CONFIG_OPTION', 'CEPHADM_FAILED_SET_OPTION']:
                self.mgr.remove_health_warning(name)
            invalid_config_options = []
            options_failed_to_set = []
            for k, v in spec.config.items():
                try:
                    current = self.mgr.get_foreign_ceph_option(section, k)
                except KeyError:
                    msg = f'Ignoring invalid {spec.service_name()} config option {k}'
                    self.log.warning(msg)
                    self.mgr.events.for_service(
                        spec, OrchestratorEvent.ERROR, f'Invalid config option {k}'
                    )
                    invalid_config_options.append(msg)
                    continue
                if current != v:
                    self.log.debug(f'setting [{section}] {k} = {v}')
                    try:
                        self.mgr.check_mon_command({
                            'prefix': 'config set',
                            'name': k,
                            'value': str(v),
                            'who': section,
                        })
                    except MonCommandFailed as e:
                        msg = f'Failed to set {spec.service_name()} option {k}: {e}'
                        self.log.warning(msg)
                        options_failed_to_set.append(msg)

            if invalid_config_options:
                self.mgr.set_health_warning('CEPHADM_INVALID_CONFIG_OPTION', f'Ignoring {len(invalid_config_options)} invalid config option(s)', len(
                    invalid_config_options), invalid_config_options)
            if options_failed_to_set:
                self.mgr.set_health_warning('CEPHADM_FAILED_SET_OPTION', f'Failed to set {len(options_failed_to_set)} option(s)', len(
                    options_failed_to_set), options_failed_to_set)

    def _update_rgw_endpoints(self, rgw_spec: RGWSpec) -> None:

        if not rgw_spec.update_endpoints or rgw_spec.rgw_realm_token is None:
            return

        ep = []
        protocol = 'https' if rgw_spec.ssl else 'http'
        for s in self.mgr.cache.get_daemons_by_service(rgw_spec.service_name()):
            if s.ports:
                for p in s.ports:
                    ep.append(f'{protocol}://{s.hostname}:{p}')
        zone_update_cmd = {
            'prefix': 'rgw zone modify',
            'realm_name': rgw_spec.rgw_realm,
            'zonegroup_name': rgw_spec.rgw_zonegroup,
            'zone_name': rgw_spec.rgw_zone,
            'realm_token': rgw_spec.rgw_realm_token,
            'zone_endpoints': ep,
        }
        self.log.debug(f'rgw cmd: {zone_update_cmd}')
        rc, out, err = self.mgr.mon_command(zone_update_cmd)
        rgw_spec.update_endpoints = (rc != 0)  # keep trying on failure
        if rc != 0:
            self.log.error(f'Error when trying to update rgw zone: {err}')
            self.mgr.set_health_warning('CEPHADM_RGW', 'Cannot update rgw endpoints, error: {err}', 1,
                                        [f'Cannot update rgw endpoints for daemon {rgw_spec.service_name()}, error: {err}'])
        else:
            self.mgr.remove_health_warning('CEPHADM_RGW')

    def _apply_service(self, spec: ServiceSpec) -> bool:
        """
        Schedule a service.  Deploy new daemons or remove old ones, depending
        on the target label and count specified in the placement.
        """
        self.mgr.migration.verify_no_migration()

        service_type = spec.service_type
        service_name = spec.service_name()
        if spec.unmanaged:
            self.log.debug('Skipping unmanaged service %s' % service_name)
            return False
        if spec.preview_only:
            self.log.debug('Skipping preview_only service %s' % service_name)
            return False
        self.log.debug('Applying service %s spec' % service_name)

        if service_type == 'agent':
            try:
                assert self.mgr.http_server.agent
                assert self.mgr.http_server.agent.ssl_certs.get_root_cert()
            except Exception:
                self.log.info(
                    'Delaying applying agent spec until cephadm endpoint root cert created')
                return False

        self._apply_service_config(spec)

        if service_type == 'osd':
            self.mgr.osd_service.create_from_spec(cast(DriveGroupSpec, spec))
            # TODO: return True would result in a busy loop
            # can't know if daemon count changed; create_from_spec doesn't
            # return a solid indication
            return False

        svc = self.mgr.cephadm_services[service_type]
        daemons = self.mgr.cache.get_daemons_by_service(service_name)
        related_service_daemons = self.mgr.cache.get_related_service_daemons(spec)

        public_networks: List[str] = []
        if service_type == 'mon':
            out = str(self.mgr.get_foreign_ceph_option('mon', 'public_network'))
            if '/' in out:
                public_networks = [x.strip() for x in out.split(',')]
                self.log.debug('mon public_network(s) is %s' % public_networks)

        def matches_public_network(host: str, sspec: ServiceSpec) -> bool:
            # make sure the host has at least one network that belongs to some configured public network(s)
            for pn in public_networks:
                public_network = ipaddress.ip_network(pn)
                for hn in self.mgr.cache.networks[host]:
                    host_network = ipaddress.ip_network(hn)
                    if host_network.overlaps(public_network):
                        return True

            host_networks = ','.join(self.mgr.cache.networks[host])
            pub_networks = ','.join(public_networks)
            self.log.info(
                f"Filtered out host {host}: does not belong to mon public_network(s): "
                f" {pub_networks}, host network(s): {host_networks}"
            )
            return False

        def has_interface_for_vip(host: str, sspec: ServiceSpec) -> bool:
            # make sure the host has an interface that can
            # actually accomodate the VIP
            if not sspec or sspec.service_type != 'ingress':
                return True
            ingress_spec = cast(IngressSpec, sspec)
            virtual_ips = []
            if ingress_spec.virtual_ip:
                virtual_ips.append(ingress_spec.virtual_ip)
            elif ingress_spec.virtual_ips_list:
                virtual_ips = ingress_spec.virtual_ips_list
            for vip in virtual_ips:
                found = False
                bare_ip = str(vip).split('/')[0]
                for subnet, ifaces in self.mgr.cache.networks.get(host, {}).items():
                    if ifaces and ipaddress.ip_address(bare_ip) in ipaddress.ip_network(subnet):
                        # found matching interface for this IP, move on
                        self.log.debug(
                            f'{bare_ip} is in {subnet} on {host} interface {list(ifaces.keys())[0]}'
                        )
                        found = True
                        break
                if not found:
                    self.log.info(
                        f"Filtered out host {host}: Host has no interface available for VIP: {vip}"
                    )
                    return False
            return True

        host_filters: Dict[str, Callable[[str, ServiceSpec], bool]] = {
            'mon': matches_public_network,
            'ingress': has_interface_for_vip
        }

        rank_map = None
        if svc.ranked():
            rank_map = self.mgr.spec_store[spec.service_name()].rank_map or {}
        ha = HostAssignment(
            spec=spec,
            hosts=self.mgr.cache.get_non_draining_hosts() if spec.service_name(
            ) == 'agent' else self.mgr.cache.get_schedulable_hosts(),
            unreachable_hosts=self.mgr.cache.get_unreachable_hosts(),
            draining_hosts=self.mgr.cache.get_draining_hosts(),
            daemons=daemons,
            related_service_daemons=related_service_daemons,
            networks=self.mgr.cache.networks,
            filter_new_host=host_filters.get(service_type, None),
            allow_colo=svc.allow_colo(),
            primary_daemon_type=svc.primary_daemon_type(spec),
            per_host_daemon_type=svc.per_host_daemon_type(spec),
            rank_map=rank_map,
        )

        try:
            all_slots, slots_to_add, daemons_to_remove = ha.place()
            daemons_to_remove = [d for d in daemons_to_remove if (d.hostname and self.mgr.inventory._inventory[d.hostname].get(
                'status', '').lower() not in ['maintenance', 'offline'] and d.hostname not in self.mgr.offline_hosts)]
            self.log.debug('Add %s, remove %s' % (slots_to_add, daemons_to_remove))
        except OrchestratorError as e:
            msg = f'Failed to apply {spec.service_name()} spec {spec}: {str(e)}'
            self.log.error(msg)
            self.mgr.events.for_service(spec, 'ERROR', 'Failed to apply: ' + str(e))
            self.mgr.apply_spec_fails.append((spec.service_name(), str(e)))
            warnings = []
            for x in self.mgr.apply_spec_fails:
                warnings.append(f'{x[0]}: {x[1]}')
            self.mgr.set_health_warning('CEPHADM_APPLY_SPEC_FAIL',
                                        f"Failed to apply {len(self.mgr.apply_spec_fails)} service(s): {','.join(x[0] for x in self.mgr.apply_spec_fails)}",
                                        len(self.mgr.apply_spec_fails),
                                        warnings)
            return False

        r = None

        # sanity check
        final_count = len(daemons) + len(slots_to_add) - len(daemons_to_remove)
        if service_type in ['mon', 'mgr'] and final_count < 1:
            self.log.debug('cannot scale mon|mgr below 1)')
            return False

        # progress
        progress_id = str(uuid.uuid4())
        delta: List[str] = []
        if slots_to_add:
            delta += [f'+{len(slots_to_add)}']
        if daemons_to_remove:
            delta += [f'-{len(daemons_to_remove)}']
        progress_title = f'Updating {spec.service_name()} deployment ({" ".join(delta)} -> {len(all_slots)})'
        progress_total = len(slots_to_add) + len(daemons_to_remove)
        progress_done = 0

        def update_progress() -> None:
            self.mgr.remote(
                'progress', 'update', progress_id,
                ev_msg=progress_title,
                ev_progress=(progress_done / progress_total),
                add_to_ceph_s=True,
            )

        if progress_total:
            update_progress()

        self.log.debug('Hosts that will receive new daemons: %s' % slots_to_add)
        self.log.debug('Daemons that will be removed: %s' % daemons_to_remove)

        hosts_altered: Set[str] = set()

        try:
            # assign names
            for i in range(len(slots_to_add)):
                slot = slots_to_add[i]
                slot = slot.assign_name(self.mgr.get_unique_name(
                    slot.daemon_type,
                    slot.hostname,
                    [d for d in daemons if d not in daemons_to_remove],
                    prefix=spec.service_id,
                    forcename=slot.name,
                    rank=slot.rank,
                    rank_generation=slot.rank_generation,
                ))
                slots_to_add[i] = slot
                if rank_map is not None:
                    assert slot.rank is not None
                    assert slot.rank_generation is not None
                    assert rank_map[slot.rank][slot.rank_generation] is None
                    rank_map[slot.rank][slot.rank_generation] = slot.name

            if rank_map:
                # record the rank_map before we make changes so that if we fail the
                # next mgr will clean up.
                self.mgr.spec_store.save_rank_map(spec.service_name(), rank_map)

                # remove daemons now, since we are going to fence them anyway
                for d in daemons_to_remove:
                    assert d.hostname is not None
                    self._remove_daemon(d.name(), d.hostname)
                daemons_to_remove = []

                # fence them
                svc.fence_old_ranks(spec, rank_map, len(all_slots))

            # create daemons
            daemon_place_fails = []
            for slot in slots_to_add:
                # first remove daemon with conflicting port or name?
                if slot.ports or slot.name in [d.name() for d in daemons_to_remove]:
                    for d in daemons_to_remove:
                        if (
                            d.hostname != slot.hostname
                            or not (set(d.ports or []) & set(slot.ports))
                            or (d.ip and slot.ip and d.ip != slot.ip)
                            and d.name() != slot.name
                        ):
                            continue
                        if d.name() != slot.name:
                            self.log.info(
                                f'Removing {d.name()} before deploying to {slot} to avoid a port or conflict'
                            )
                        # NOTE: we don't check ok-to-stop here to avoid starvation if
                        # there is only 1 gateway.
                        self._remove_daemon(d.name(), d.hostname)
                        daemons_to_remove.remove(d)
                        progress_done += 1
                        hosts_altered.add(d.hostname)
                        break

                # do not attempt to deploy node-proxy agent when oob details are not provided.
                if slot.daemon_type == 'node-proxy' and slot.hostname not in self.mgr.node_proxy_cache.oob.keys():
                    self.log.debug(
                        f'Not deploying node-proxy agent on {slot.hostname} as oob details are not present.'
                    )
                    continue

                # deploy new daemon
                daemon_id = slot.name

                daemon_spec = svc.make_daemon_spec(
                    slot.hostname, daemon_id, slot.network, spec,
                    daemon_type=slot.daemon_type,
                    ports=slot.ports,
                    ip=slot.ip,
                    rank=slot.rank,
                    rank_generation=slot.rank_generation,
                )
                self.log.debug('Placing %s.%s on host %s' % (
                    slot.daemon_type, daemon_id, slot.hostname))

                try:
                    daemon_spec = svc.prepare_create(daemon_spec)
                    with self.mgr.async_timeout_handler(slot.hostname, f'cephadm deploy ({daemon_spec.daemon_type} type dameon)'):
                        self.mgr.wait_async(self._create_daemon(daemon_spec))
                    r = True
                    progress_done += 1
                    update_progress()
                    hosts_altered.add(daemon_spec.host)
                    self.mgr.spec_store.mark_needs_configuration(spec.service_name())
                except (RuntimeError, OrchestratorError) as e:
                    msg = (f"Failed while placing {slot.daemon_type}.{daemon_id} "
                           f"on {slot.hostname}: {e}")
                    self.mgr.events.for_service(spec, 'ERROR', msg)
                    self.mgr.log.error(msg)
                    daemon_place_fails.append(msg)
                    # only return "no change" if no one else has already succeeded.
                    # later successes will also change to True
                    if r is None:
                        r = False
                    progress_done += 1
                    update_progress()
                    continue

                # add to daemon list so next name(s) will also be unique
                sd = orchestrator.DaemonDescription(
                    hostname=slot.hostname,
                    daemon_type=slot.daemon_type,
                    daemon_id=daemon_id,
                    service_name=spec.service_name()
                )
                daemons.append(sd)
                self.mgr.cache.append_tmp_daemon(slot.hostname, sd)

            if daemon_place_fails:
                self.mgr.set_health_warning('CEPHADM_DAEMON_PLACE_FAIL', f'Failed to place {len(daemon_place_fails)} daemon(s)', len(
                    daemon_place_fails), daemon_place_fails)

            if service_type == 'mgr':
                active_mgr = svc.get_active_daemon(self.mgr.cache.get_daemons_by_type('mgr'))
                if active_mgr.daemon_id in [d.daemon_id for d in daemons_to_remove]:
                    # We can't just remove the active mgr like any other daemon.
                    # Need to fail over later so it can be removed on next pass.
                    # This can be accomplished by scheduling a restart of the active mgr.
                    self.mgr._schedule_daemon_action(active_mgr.name(), 'restart')

            if service_type == 'rgw':
                self._update_rgw_endpoints(cast(RGWSpec, spec))

            # remove any?
            def _ok_to_stop(remove_daemons: List[orchestrator.DaemonDescription]) -> bool:
                daemon_ids = [d.daemon_id for d in remove_daemons]
                assert None not in daemon_ids
                # setting force flag retains previous behavior
                r = svc.ok_to_stop(cast(List[str], daemon_ids), force=True)
                return not r.retval

            while daemons_to_remove and not _ok_to_stop(daemons_to_remove):
                # let's find a subset that is ok-to-stop
                non_error_daemon_index = -1
                # prioritize removing daemons in error state
                for i, dmon in enumerate(daemons_to_remove):
                    if dmon.status != DaemonDescriptionStatus.error:
                        non_error_daemon_index = i
                        break
                if non_error_daemon_index != -1:
                    daemons_to_remove.pop(non_error_daemon_index)
                else:
                    # all daemons in list are in error state
                    # we should be able to remove all of them
                    break
            for d in daemons_to_remove:
                r = True
                assert d.hostname is not None
                self._remove_daemon(d.name(), d.hostname)

                progress_done += 1
                update_progress()
                hosts_altered.add(d.hostname)
                self.mgr.spec_store.mark_needs_configuration(spec.service_name())

            self.mgr.remote('progress', 'complete', progress_id)
        except Exception as e:
            self.mgr.remote('progress', 'fail', progress_id, str(e))
            raise
        finally:
            if self.mgr.spec_store.needs_configuration(spec.service_name()):
                svc.config(spec)
                self.mgr.spec_store.mark_configured(spec.service_name())
            if self.mgr.use_agent:
                # can only send ack to agents if we know for sure port they bound to
                hosts_altered = set([h for h in hosts_altered if (h in self.mgr.agent_cache.agent_ports and not self.mgr.cache.is_host_draining(h))])
                self.mgr.agent_helpers._request_agent_acks(hosts_altered, increment=True)

        if r is None:
            r = False
        return r

    def _check_daemons(self) -> None:
        self.log.debug('_check_daemons')
        daemons = self.mgr.cache.get_daemons()
        daemons_post: Dict[str, List[orchestrator.DaemonDescription]] = defaultdict(list)
        for dd in daemons:
            # orphan?
            spec = self.mgr.spec_store.active_specs.get(dd.service_name(), None)
            assert dd.hostname is not None
            assert dd.daemon_type is not None
            assert dd.daemon_id is not None

            # any action we can try will fail for a daemon on an offline host,
            # including removing the daemon
            if dd.hostname in self.mgr.offline_hosts:
                continue

            if not spec and dd.daemon_type not in ['mon', 'mgr', 'osd']:
                # (mon and mgr specs should always exist; osds aren't matched
                # to a service spec)
                self.log.info('Removing orphan daemon %s...' % dd.name())
                self._remove_daemon(dd.name(), dd.hostname)

            # ignore unmanaged services
            if spec and spec.unmanaged:
                continue

            # ignore daemons for deleted services
            if dd.service_name() in self.mgr.spec_store.spec_deleted:
                continue

            if dd.daemon_type == 'agent':
                try:
                    self.mgr.agent_helpers._check_agent(dd.hostname)
                except Exception as e:
                    self.log.debug(
                        f'Agent {dd.name()} could not be checked in _check_daemons: {e}')
                continue

            # These daemon types require additional configs after creation
            if dd.daemon_type in REQUIRES_POST_ACTIONS:
                daemons_post[dd.daemon_type].append(dd)

            if self.mgr.cephadm_services[daemon_type_to_service(dd.daemon_type)].get_active_daemon(
               self.mgr.cache.get_daemons_by_service(dd.service_name())).daemon_id == dd.daemon_id:
                dd.is_active = True
            else:
                dd.is_active = False

            deps = self.mgr._calc_daemon_deps(spec, dd.daemon_type, dd.daemon_id)
            last_deps, last_config = self.mgr.cache.get_daemon_last_config_deps(
                dd.hostname, dd.name())
            if last_deps is None:
                last_deps = []
            action = self.mgr.cache.get_scheduled_daemon_action(dd.hostname, dd.name())
            if not last_config:
                self.log.info('Reconfiguring %s (unknown last config time)...' % (
                    dd.name()))
                action = 'reconfig'
            elif last_deps != deps:
                self.log.debug(f'{dd.name()} deps {last_deps} -> {deps}')
                self.log.info(f'Reconfiguring {dd.name()} (dependencies changed)...')
                action = 'reconfig'
                # we need only redeploy if secure_monitoring_stack value has changed:
                if dd.daemon_type in ['prometheus', 'node-exporter', 'alertmanager']:
                    diff = list(set(last_deps) - set(deps))
                    if any('secure_monitoring_stack' in e for e in diff):
                        action = 'redeploy'
                elif dd.daemon_type == 'jaeger-agent':
                    # changes to jaeger-agent deps affect the way the unit.run for
                    # the daemon is written, which we rewrite on redeploy, but not
                    # on reconfig.
                    action = 'redeploy'

            elif spec is not None and hasattr(spec, 'extra_container_args') and dd.extra_container_args != spec.extra_container_args:
                self.log.debug(
                    f'{dd.name()} container cli args {dd.extra_container_args} -> {spec.extra_container_args}')
                self.log.info(f'Redeploying {dd.name()}, (container cli args changed) . . .')
                dd.extra_container_args = spec.extra_container_args
                action = 'redeploy'
            elif spec is not None and hasattr(spec, 'extra_entrypoint_args') and dd.extra_entrypoint_args != spec.extra_entrypoint_args:
                self.log.info(f'Redeploying {dd.name()}, (entrypoint args changed) . . .')
                self.log.debug(
                    f'{dd.name()} daemon entrypoint args {dd.extra_entrypoint_args} -> {spec.extra_entrypoint_args}')
                dd.extra_entrypoint_args = spec.extra_entrypoint_args
                action = 'redeploy'
            elif self.mgr.last_monmap and \
                    self.mgr.last_monmap > last_config and \
                    dd.daemon_type in CEPH_TYPES:
                self.log.info('Reconfiguring %s (monmap changed)...' % dd.name())
                action = 'reconfig'
            elif self.mgr.extra_ceph_conf_is_newer(last_config) and \
                    dd.daemon_type in CEPH_TYPES:
                self.log.info('Reconfiguring %s (extra config changed)...' % dd.name())
                action = 'reconfig'
            if action:
                if self.mgr.cache.get_scheduled_daemon_action(dd.hostname, dd.name()) == 'redeploy' \
                        and action == 'reconfig':
                    action = 'redeploy'
                try:
                    daemon_spec = CephadmDaemonDeploySpec.from_daemon_description(dd)
                    self.mgr._daemon_action(daemon_spec, action=action)
                    if self.mgr.cache.rm_scheduled_daemon_action(dd.hostname, dd.name()):
                        self.mgr.cache.save_host(dd.hostname)
                except OrchestratorError as e:
                    self.log.exception(e)
                    self.mgr.events.from_orch_error(e)
                    if dd.daemon_type in daemons_post:
                        del daemons_post[dd.daemon_type]
                    # continue...
                except Exception as e:
                    self.log.exception(e)
                    self.mgr.events.for_daemon_from_exception(dd.name(), e)
                    if dd.daemon_type in daemons_post:
                        del daemons_post[dd.daemon_type]
                    # continue...

        # do daemon post actions
        for daemon_type, daemon_descs in daemons_post.items():
            run_post = False
            for d in daemon_descs:
                if d.name() in self.mgr.requires_post_actions:
                    self.mgr.requires_post_actions.remove(d.name())
                    run_post = True
            if run_post:
                self.mgr._get_cephadm_service(daemon_type_to_service(
                    daemon_type)).daemon_check_post(daemon_descs)

    def _purge_deleted_services(self) -> None:
        self.log.debug('_purge_deleted_services')
        existing_services = self.mgr.spec_store.all_specs.items()
        for service_name, spec in list(existing_services):
            if service_name not in self.mgr.spec_store.spec_deleted:
                continue
            if self.mgr.cache.get_daemons_by_service(service_name):
                continue
            if spec.service_type in ['mon', 'mgr']:
                continue

            logger.info(f'Purge service {service_name}')

            self.mgr.cephadm_services[spec.service_type].purge(service_name)
            self.mgr.spec_store.finally_rm(service_name)

    def convert_tags_to_repo_digest(self) -> None:
        if not self.mgr.use_repo_digest:
            return
        settings = self.mgr.upgrade.get_distinct_container_image_settings()
        digests: Dict[str, ContainerInspectInfo] = {}
        for container_image_ref in set(settings.values()):
            if not is_repo_digest(container_image_ref):
                with self.mgr.async_timeout_handler(cmd=f'cephadm inspect-image (image {container_image_ref})'):
                    image_info = self.mgr.wait_async(
                        self._get_container_image_info(container_image_ref))
                if image_info.repo_digests:
                    # FIXME: we assume the first digest here is the best
                    assert is_repo_digest(image_info.repo_digests[0]), image_info
                digests[container_image_ref] = image_info

        for entity, container_image_ref in settings.items():
            if not is_repo_digest(container_image_ref):
                image_info = digests[container_image_ref]
                if image_info.repo_digests:
                    # FIXME: we assume the first digest here is the best
                    self.mgr.set_container_image(entity, image_info.repo_digests[0])

    def _calc_client_files(self) -> Dict[str, Dict[str, Tuple[int, int, int, bytes, str]]]:
        # host -> path -> (mode, uid, gid, content, digest)
        client_files: Dict[str, Dict[str, Tuple[int, int, int, bytes, str]]] = {}

        # ceph.conf
        config = self.mgr.get_minimal_ceph_conf().encode('utf-8')
        config_digest = ''.join('%02x' % c for c in hashlib.sha256(config).digest())
        cluster_cfg_dir = f'/var/lib/ceph/{self.mgr._cluster_fsid}/config'

        if self.mgr.manage_etc_ceph_ceph_conf:
            try:
                pspec = PlacementSpec.from_string(self.mgr.manage_etc_ceph_ceph_conf_hosts)
                ha = HostAssignment(
                    spec=ServiceSpec('mon', placement=pspec),
                    hosts=self.mgr.cache.get_conf_keyring_available_hosts(),
                    unreachable_hosts=self.mgr.cache.get_unreachable_hosts(),
                    draining_hosts=self.mgr.cache.get_conf_keyring_draining_hosts(),
                    daemons=[],
                    networks=self.mgr.cache.networks,
                )
                all_slots, _, _ = ha.place()
                for host in {s.hostname for s in all_slots}:
                    if host not in client_files:
                        client_files[host] = {}
                    ceph_conf = (0o644, 0, 0, bytes(config), str(config_digest))
                    client_files[host]['/etc/ceph/ceph.conf'] = ceph_conf
                    client_files[host][f'{cluster_cfg_dir}/ceph.conf'] = ceph_conf
            except Exception as e:
                self.mgr.log.warning(
                    f'unable to calc conf hosts: {self.mgr.manage_etc_ceph_ceph_conf_hosts}: {e}')

        # client keyrings
        for ks in self.mgr.keys.keys.values():
            try:
                ret, keyring, err = self.mgr.mon_command({
                    'prefix': 'auth get',
                    'entity': ks.entity,
                })
                if ret:
                    self.log.warning(f'unable to fetch keyring for {ks.entity}')
                    continue
                digest = ''.join('%02x' % c for c in hashlib.sha256(
                    keyring.encode('utf-8')).digest())
                ha = HostAssignment(
                    spec=ServiceSpec('mon', placement=ks.placement),
                    hosts=self.mgr.cache.get_conf_keyring_available_hosts(),
                    unreachable_hosts=self.mgr.cache.get_unreachable_hosts(),
                    draining_hosts=self.mgr.cache.get_conf_keyring_draining_hosts(),
                    daemons=[],
                    networks=self.mgr.cache.networks,
                )
                all_slots, _, _ = ha.place()
                for host in {s.hostname for s in all_slots}:
                    if host not in client_files:
                        client_files[host] = {}
                    ceph_conf = (0o644, 0, 0, bytes(config), str(config_digest))
                    client_files[host]['/etc/ceph/ceph.conf'] = ceph_conf
                    client_files[host][f'{cluster_cfg_dir}/ceph.conf'] = ceph_conf
                    ceph_admin_key = (ks.mode, ks.uid, ks.gid, keyring.encode('utf-8'), digest)
                    client_files[host][ks.path] = ceph_admin_key
                    client_files[host][f'{cluster_cfg_dir}/{os.path.basename(ks.path)}'] = ceph_admin_key
            except Exception as e:
                self.log.warning(
                    f'unable to calc client keyring {ks.entity} placement {ks.placement}: {e}')
        return client_files

    def _write_all_client_files(self) -> None:
        if self.mgr.manage_etc_ceph_ceph_conf or self.mgr.keys.keys:
            client_files = self._calc_client_files()
        else:
            client_files = {}

        @forall_hosts
        def _write_files(host: str) -> None:
            self._write_client_files(client_files, host)

        _write_files(self.mgr.cache.get_hosts())

    def _write_client_files(self,
                            client_files: Dict[str, Dict[str, Tuple[int, int, int, bytes, str]]],
                            host: str) -> None:
        updated_files = False
        if self.mgr.cache.is_host_unreachable(host):
            return
        old_files = self.mgr.cache.get_host_client_files(host).copy()
        for path, m in client_files.get(host, {}).items():
            mode, uid, gid, content, digest = m
            if path in old_files:
                match = old_files[path] == (digest, mode, uid, gid)
                del old_files[path]
                if match:
                    continue
            self.log.info(f'Updating {host}:{path}')
            self.mgr.ssh.write_remote_file(host, path, content, mode, uid, gid)
            self.mgr.cache.update_client_file(host, path, digest, mode, uid, gid)
            updated_files = True
        for path in old_files.keys():
            if path == '/etc/ceph/ceph.conf':
                continue
            self.log.info(f'Removing {host}:{path}')
            cmd = ['rm', '-f', path]
            self.mgr.ssh.check_execute_command(host, cmd)
            updated_files = True
            self.mgr.cache.removed_client_file(host, path)
        if updated_files:
            self.mgr.cache.save_host(host)

    async def _create_daemon(self,
                             daemon_spec: CephadmDaemonDeploySpec,
                             reconfig: bool = False,
                             osd_uuid_map: Optional[Dict[str, Any]] = None,
                             ) -> str:

        daemon_params: Dict[str, Any] = {}
        with set_exception_subject('service', orchestrator.DaemonDescription(
                daemon_type=daemon_spec.daemon_type,
                daemon_id=daemon_spec.daemon_id,
                hostname=daemon_spec.host,
        ).service_id(), overwrite=True):

            try:
                image = ''
                start_time = datetime_now()
                ports: List[int] = daemon_spec.ports if daemon_spec.ports else []
                port_ips: Dict[str, str] = daemon_spec.port_ips if daemon_spec.port_ips else {}

                if daemon_spec.daemon_type == 'container':
                    spec = cast(CustomContainerSpec,
                                self.mgr.spec_store[daemon_spec.service_name].spec)
                    image = spec.image
                    if spec.ports:
                        ports.extend(spec.ports)

                # TCP port to open in the host firewall
                if len(ports) > 0:
                    daemon_params['tcp_ports'] = list(ports)

                if port_ips:
                    daemon_params['port_ips'] = port_ips

                # osd deployments needs an --osd-uuid arg
                if daemon_spec.daemon_type == 'osd':
                    if not osd_uuid_map:
                        osd_uuid_map = self.mgr.get_osd_uuid_map()
                    osd_uuid = osd_uuid_map.get(daemon_spec.daemon_id)
                    if not osd_uuid:
                        raise OrchestratorError('osd.%s not in osdmap' % daemon_spec.daemon_id)
                    daemon_params['osd_fsid'] = osd_uuid

                if reconfig:
                    daemon_params['reconfig'] = True
                if self.mgr.allow_ptrace:
                    daemon_params['allow_ptrace'] = True

                daemon_spec, extra_container_args, extra_entrypoint_args = self._setup_extra_deployment_args(daemon_spec, daemon_params)

                if daemon_spec.service_name in self.mgr.spec_store:
                    configs = self.mgr.spec_store[daemon_spec.service_name].spec.custom_configs
                    if configs is not None:
                        daemon_spec.final_config.update(
                            {'custom_config_files': [c.to_json() for c in configs]})

                if self.mgr.cache.host_needs_registry_login(daemon_spec.host) and self.mgr.registry_url:
                    await self._registry_login(daemon_spec.host, json.loads(str(self.mgr.get_store('registry_credentials'))))

                self.log.info('%s daemon %s on %s' % (
                    'Reconfiguring' if reconfig else 'Deploying',
                    daemon_spec.name(), daemon_spec.host))

                out, err, code = await self._run_cephadm(
                    daemon_spec.host,
                    daemon_spec.name(),
                    ['_orch', 'deploy'],
                    [],
                    stdin=exchange.Deploy(
                        fsid=self.mgr._cluster_fsid,
                        name=daemon_spec.name(),
                        image=image,
                        params=daemon_params,
                        meta=exchange.DeployMeta(
                            service_name=daemon_spec.service_name,
                            ports=daemon_spec.ports,
                            ip=daemon_spec.ip,
                            deployed_by=self.mgr.get_active_mgr_digests(),
                            rank=daemon_spec.rank,
                            rank_generation=daemon_spec.rank_generation,
                            extra_container_args=ArgumentSpec.map_json(
                                extra_container_args,
                            ),
                            extra_entrypoint_args=ArgumentSpec.map_json(
                                extra_entrypoint_args,
                            ),
                        ),
                        config_blobs=daemon_spec.final_config,
                    ).dump_json_str(),
                )

                if daemon_spec.daemon_type == 'agent':
                    self.mgr.agent_cache.agent_timestamp[daemon_spec.host] = datetime_now()
                    self.mgr.agent_cache.agent_counter[daemon_spec.host] = 1

                # refresh daemon state?  (ceph daemon reconfig does not need it)
                if not reconfig or daemon_spec.daemon_type not in CEPH_TYPES:
                    if not code and daemon_spec.host in self.mgr.cache.daemons:
                        # prime cached service state with what we (should have)
                        # just created
                        sd = daemon_spec.to_daemon_description(
                            DaemonDescriptionStatus.starting, 'starting')
                        self.mgr.cache.add_daemon(daemon_spec.host, sd)
                        if daemon_spec.daemon_type in REQUIRES_POST_ACTIONS:
                            self.mgr.requires_post_actions.add(daemon_spec.name())
                    self.mgr.cache.invalidate_host_daemons(daemon_spec.host)

                if daemon_spec.daemon_type != 'agent':
                    self.mgr.cache.update_daemon_config_deps(
                        daemon_spec.host, daemon_spec.name(), daemon_spec.deps, start_time)
                    self.mgr.cache.save_host(daemon_spec.host)
                else:
                    self.mgr.agent_cache.update_agent_config_deps(
                        daemon_spec.host, daemon_spec.deps, start_time)
                    self.mgr.agent_cache.save_agent(daemon_spec.host)
                msg = "{} {} on host '{}'".format(
                    'Reconfigured' if reconfig else 'Deployed', daemon_spec.name(), daemon_spec.host)
                if not code:
                    self.mgr.events.for_daemon(daemon_spec.name(), OrchestratorEvent.INFO, msg)
                else:
                    what = 'reconfigure' if reconfig else 'deploy'
                    self.mgr.events.for_daemon(
                        daemon_spec.name(), OrchestratorEvent.ERROR, f'Failed to {what}: {err}')
                return msg
            except OrchestratorError:
                redeploy = daemon_spec.name() in self.mgr.cache.get_daemon_names()
                if not reconfig and not redeploy:
                    # we have to clean up the daemon. E.g. keyrings.
                    servict_type = daemon_type_to_service(daemon_spec.daemon_type)
                    dd = daemon_spec.to_daemon_description(DaemonDescriptionStatus.error, 'failed')
                    self.mgr.cephadm_services[servict_type].post_remove(dd, is_failed_deploy=True)
                raise

    def _setup_extra_deployment_args(
        self,
        daemon_spec: CephadmDaemonDeploySpec,
        params: Dict[str, Any],
    ) -> Tuple[CephadmDaemonDeploySpec, Optional[ArgumentList], Optional[ArgumentList]]:
        # this function is for handling any potential user specified
        # (in the service spec) extra runtime or entrypoint args for a daemon
        # we are going to deploy. Effectively just adds a set of extra args to
        # pass to the cephadm binary to indicate the daemon being deployed
        # needs extra runtime/entrypoint args. Returns the modified daemon spec
        # as well as what args were added (as those are included in unit.meta file)
        def _to_args(lst: ArgumentList) -> List[str]:
            out: List[str] = []
            for argspec in lst:
                out.extend(argspec.to_args())
            return out

        try:
            eca = daemon_spec.extra_container_args
            if eca:
                params['extra_container_args'] = _to_args(eca)
        except AttributeError:
            eca = None
        try:
            eea = daemon_spec.extra_entrypoint_args
            if eea:
                params['extra_entrypoint_args'] = _to_args(eea)
        except AttributeError:
            eea = None
        return daemon_spec, eca, eea

    def _remove_daemon(self, name: str, host: str, no_post_remove: bool = False) -> str:
        """
        Remove a daemon
        """
        (daemon_type, daemon_id) = name.split('.', 1)
        daemon = orchestrator.DaemonDescription(
            daemon_type=daemon_type,
            daemon_id=daemon_id,
            hostname=host)

        with set_exception_subject('service', daemon.service_id(), overwrite=True):

            self.mgr.cephadm_services[daemon_type_to_service(daemon_type)].pre_remove(daemon)
            # NOTE: we are passing the 'force' flag here, which means
            # we can delete a mon instances data.
            dd = self.mgr.cache.get_daemon(daemon.daemon_name)
            if dd.ports:
                args = ['--name', name, '--force', '--tcp-ports', ' '.join(map(str, dd.ports))]
            else:
                args = ['--name', name, '--force']

            self.log.info('Removing daemon %s from %s -- ports %s' % (name, host, dd.ports))
            with self.mgr.async_timeout_handler(host, f'cephadm rm-daemon (daemon {name})'):
                out, err, code = self.mgr.wait_async(self._run_cephadm(
                    host, name, 'rm-daemon', args))
            if not code:
                # remove item from cache
                self.mgr.cache.rm_daemon(host, name)
            self.mgr.cache.invalidate_host_daemons(host)

            if not no_post_remove:
                if daemon_type not in ['iscsi']:
                    self.mgr.cephadm_services[daemon_type_to_service(
                        daemon_type)].post_remove(daemon, is_failed_deploy=False)
                else:
                    self.mgr.scheduled_async_actions.append(lambda: self.mgr.cephadm_services[daemon_type_to_service(
                                                            daemon_type)].post_remove(daemon, is_failed_deploy=False))
                    self.mgr._kick_serve_loop()

            return "Removed {} from host '{}'".format(name, host)

    async def _run_cephadm_json(self,
                                host: str,
                                entity: Union[CephadmNoImage, str],
                                command: str,
                                args: List[str],
                                no_fsid: Optional[bool] = False,
                                error_ok: Optional[bool] = False,
                                image: Optional[str] = "",
                                log_output: Optional[bool] = True,
                                ) -> Any:
        try:
            out, err, code = await self._run_cephadm(
                host, entity, command, args, no_fsid=no_fsid, error_ok=error_ok,
                image=image, log_output=log_output)
            if code:
                raise OrchestratorError(f'host {host} `cephadm {command}` returned {code}: {err}')
        except Exception as e:
            raise OrchestratorError(f'host {host} `cephadm {command}` failed: {e}')
        try:
            return json.loads(''.join(out))
        except (ValueError, KeyError):
            msg = f'host {host} `cephadm {command}` failed: Cannot decode JSON'
            self.log.exception(f'{msg}: {"".join(out)}')
            raise OrchestratorError(msg)

    async def _run_cephadm(self,
                           host: str,
                           entity: Union[CephadmNoImage, str],
                           command: Union[str, List[str]],
                           args: List[str],
                           addr: Optional[str] = "",
                           stdin: Optional[str] = "",
                           no_fsid: Optional[bool] = False,
                           error_ok: Optional[bool] = False,
                           image: Optional[str] = "",
                           env_vars: Optional[List[str]] = None,
                           log_output: Optional[bool] = True,
                           timeout: Optional[int] = None,  # timeout in seconds
                           ) -> Tuple[List[str], List[str], int]:
        """
        Run cephadm on the remote host with the given command + args

        Important: You probably don't want to run _run_cephadm from CLI handlers

        :env_vars: in format -> [KEY=VALUE, ..]
        """

        await self.mgr.ssh._remote_connection(host, addr)

        self.log.debug(f"_run_cephadm : command = {command}")
        self.log.debug(f"_run_cephadm : args = {args}")

        bypass_image = ('agent')

        assert image or entity
        # Skip the image check for daemons deployed that are not ceph containers
        if not str(entity).startswith(bypass_image):
            if not image and entity is not cephadmNoImage:
                image = self.mgr._get_container_image(entity)

        final_args = []

        # global args
        if env_vars:
            for env_var_pair in env_vars:
                final_args.extend(['--env', env_var_pair])

        if image:
            final_args.extend(['--image', image])

        if not self.mgr.container_init:
            final_args += ['--no-container-init']

        if not self.mgr.cgroups_split:
            final_args += ['--no-cgroups-split']

        if not timeout:
            # default global timeout if no timeout was passed
            timeout = self.mgr.default_cephadm_command_timeout
            # put a lower bound of 60 seconds in case users
            # accidentally set it to something unreasonable.
            # For example if they though it was in minutes
            # rather than seconds
            if timeout < 60:
                self.log.info(f'Found default timeout set to {timeout}. Instead trying minimum of 60.')
                timeout = 60
            # subtract a small amount to give this timeout
            # in the binary a chance to actually happen over
            # the asyncio based timeout in the mgr module
            timeout -= 5
        final_args += ['--timeout', str(timeout)]

        # subcommand
        if isinstance(command, list):
            final_args.extend([str(v) for v in command])
        else:
            final_args.append(command)

        # subcommand args
        if not no_fsid:
            final_args += ['--fsid', self.mgr._cluster_fsid]

        final_args += args

        # exec
        self.log.debug('args: %s' % (' '.join(final_args)))
        if self.mgr.mode == 'root':
            # agent has cephadm binary as an extra file which is
            # therefore passed over stdin. Even for debug logs it's too much
            if stdin and 'agent' not in str(entity):
                self.log.debug('stdin: %s' % stdin)

            cmd = ['which', 'python3']
            python = await self.mgr.ssh._check_execute_command(host, cmd, addr=addr)
            cmd = [python, self.mgr.cephadm_binary_path] + final_args

            try:
                out, err, code = await self.mgr.ssh._execute_command(
                    host, cmd, stdin=stdin, addr=addr)
                if code == 2:
                    ls_cmd = ['ls', self.mgr.cephadm_binary_path]
                    out_ls, err_ls, code_ls = await self.mgr.ssh._execute_command(host, ls_cmd, addr=addr,
                                                                                  log_command=log_output)
                    if code_ls == 2:
                        await self._deploy_cephadm_binary(host, addr)
                        out, err, code = await self.mgr.ssh._execute_command(
                            host, cmd, stdin=stdin, addr=addr)
                        # if there is an agent on this host, make sure it is using the most recent
                        # version of cephadm binary
                        if host in self.mgr.inventory:
                            for agent in self.mgr.cache.get_daemons_by_type('agent', host):
                                self.mgr._schedule_daemon_action(agent.name(), 'redeploy')

            except Exception as e:
                await self.mgr.ssh._reset_con(host)
                if error_ok:
                    return [], [str(e)], 1
                raise

        elif self.mgr.mode == 'cephadm-package':
            try:
                cmd = ['/usr/bin/cephadm'] + final_args
                out, err, code = await self.mgr.ssh._execute_command(
                    host, cmd, stdin=stdin, addr=addr)
            except Exception as e:
                await self.mgr.ssh._reset_con(host)
                if error_ok:
                    return [], [str(e)], 1
                raise
        else:
            assert False, 'unsupported mode'

        if log_output:
            self.log.debug(f'code: {code}')
            if out:
                self.log.debug(f'out: {out}')
            if err:
                self.log.debug(f'err: {err}')
        if code and not error_ok:
            raise OrchestratorError(
                f'cephadm exited with an error code: {code}, stderr: {err}')
        return [out], [err], code

    async def _get_container_image_info(self, image_name: str) -> ContainerInspectInfo:
        # pick a random host...
        host = None
        for host_name in self.mgr.inventory.keys():
            host = host_name
            break
        if not host:
            raise OrchestratorError('no hosts defined')
        if self.mgr.cache.host_needs_registry_login(host) and self.mgr.registry_url:
            await self._registry_login(host, json.loads(str(self.mgr.get_store('registry_credentials'))))

        j = None
        try:
            j = await self._run_cephadm_json(host, '', 'inspect-image', [],
                                             image=image_name, no_fsid=True,
                                             error_ok=True)
        except OrchestratorError:
            pass

        if not j:
            pullargs: List[str] = []
            if self.mgr.registry_insecure:
                pullargs.append("--insecure")

            j = await self._run_cephadm_json(host, '', 'pull', pullargs,
                                             image=image_name, no_fsid=True)
        r = ContainerInspectInfo(
            j['image_id'],
            j.get('ceph_version'),
            j.get('repo_digests')
        )
        self.log.debug(f'image {image_name} -> {r}')
        return r

    # function responsible for logging single host into custom registry
    async def _registry_login(self, host: str, registry_json: Dict[str, str]) -> Optional[str]:
        self.log.debug(
            f"Attempting to log host {host} into custom registry @ {registry_json['url']}")
        # want to pass info over stdin rather than through normal list of args
        out, err, code = await self._run_cephadm(
            host, 'mon', 'registry-login',
            ['--registry-json', '-'], stdin=json.dumps(registry_json), error_ok=True)
        if code:
            return f"Host {host} failed to login to {registry_json['url']} as {registry_json['username']} with given password"
        return None

    async def _deploy_cephadm_binary(self, host: str, addr: Optional[str] = None) -> None:
        # Use tee (from coreutils) to create a copy of cephadm on the target machine
        self.log.info(f"Deploying cephadm binary to {host}")
        await self.mgr.ssh._write_remote_file(host, self.mgr.cephadm_binary_path,
                                              self.mgr._cephadm, addr=addr)