use PVE::Tools;
use PVE::HA::Tools ':exit_codes';
use PVE::HA::NodeStatus;
+use PVE::HA::Usage::Basic;
+use PVE::HA::Usage::Static;
+
+## Variable Name & Abbreviations Convention
+#
+# The HA stack has some variables it uses frequently and thus abbreviates it such that it may be
+# confusing for new readers. Here's a short list of the most common used.
+#
+# NOTE: variables should be assumed to be read only if not otherwise stated, only use the specific
+# methods to re-compute/read/alter them.
+#
+# - $haenv -> HA environment, the main interface to the simulator/test/real world
+# - $sid -> Service ID, unique identifier for a service, `type:vmid` is common
+#
+# - $ms -> Master/Manager Status, contains runtime info from the current active manager
+# - $ns -> Node Status, hash holding online/offline status about all nodes
+#
+# - $ss -> Service Status, hash holding the current state (last LRM cmd result, failed starts
+# or migrates, maintenance fallback node, for *all* services ...
+# - $sd -> Service Data, the service status of a *single* service, iow. $ss->{$sid}
+#
+# - $sc -> Service Configuration, hash for all services including target state, group, ...
+# - $cd -> Configuration Data, the service config of a *single* service, iow. $sc->{$sid}
+#
+# Try to avoid adding new two letter (or similar over abbreviated) names, but also don't send
+# patches for changing above, as that set is mostly sensible and should be easy to remember once
+# spending a bit time in the HA code base.
sub new {
my ($this, $haenv) = @_;
my $class = ref($this) || $this;
- my $self = bless { haenv => $haenv }, $class;
+ my $self = bless { haenv => $haenv, crs => {} }, $class;
my $old_ms = $haenv->read_manager_status();
$self->{ms} = { master_node => $haenv->nodename() };
+ $self->update_crs_scheduler_mode(); # initial set, we update it once every loop
+
return $self;
}
+sub update_crs_scheduler_mode {
+ my ($self) = @_;
+
+ my $haenv = $self->{haenv};
+ my $dc_cfg = $haenv->get_datacenter_settings();
+
+ my $old_mode = $self->{crs}->{scheduler};
+ my $new_mode = $dc_cfg->{crs}->{ha} || 'basic';
+
+ if (!defined($old_mode)) {
+ $haenv->log('info', "using scheduler mode '$new_mode'") if $new_mode ne 'basic';
+ } elsif ($new_mode eq $old_mode) {
+ return; # nothing to do
+ } else {
+ $haenv->log('info', "switching scheduler mode from '$old_mode' to '$new_mode'");
+ }
+
+ $self->{crs}->{scheduler} = $new_mode;
+
+ return;
+}
+
sub cleanup {
my ($self) = @_;
my $group = {};
# add all online nodes to default group to allow try_next when no group set
- foreach my $node (keys %$online_node_usage) {
- $group->{nodes}->{$node} = 1;
- }
+ $group->{nodes}->{$_} = 1 for $online_node_usage->list_nodes();
# overwrite default if service is bound to a specific group
- $group = $groups->{ids}->{$service_conf->{group}} if $service_conf->{group} &&
- $groups->{ids}->{$service_conf->{group}};
+ if (my $group_id = $service_conf->{group}) {
+ $group = $groups->{ids}->{$group_id} if $groups->{ids}->{$group_id};
+ }
return $group;
}
if ($entry =~ m/^(\S+):(\d+)$/) {
($node, $pri) = ($1, $2);
}
- next if !defined($online_node_usage->{$node}); # offline
+ next if !$online_node_usage->contains_node($node); # offline
$pri_groups->{$pri}->{$node} = 1;
$group_members->{$node} = $pri;
}
# add non-group members to unrestricted groups (priority -1)
if (!$group->{restricted}) {
my $pri = -1;
- foreach my $node (keys %$online_node_usage) {
+ for my $node ($online_node_usage->list_nodes()) {
next if defined($group_members->{$node});
$pri_groups->{$pri}->{$node} = 1;
$group_members->{$node} = -1;
}
sub select_service_node {
- my ($groups, $online_node_usage, $service_conf, $current_node, $try_next, $tried_nodes) = @_;
+ my ($groups, $online_node_usage, $sid, $service_conf, $current_node, $try_next, $tried_nodes, $maintenance_fallback) = @_;
my $group = get_service_group($groups, $online_node_usage, $service_conf);
}
}
+ return $maintenance_fallback
+ if defined($maintenance_fallback) && $pri_groups->{$top_pri}->{$maintenance_fallback};
+
+ return $current_node if !$try_next && $pri_groups->{$top_pri}->{$current_node};
+
+ my $scores = $online_node_usage->score_nodes_to_start_service($sid, $current_node);
my @nodes = sort {
- $online_node_usage->{$a} <=> $online_node_usage->{$b} || $a cmp $b
+ $scores->{$a} <=> $scores->{$b} || $a cmp $b
} keys %{$pri_groups->{$top_pri}};
my $found;
my $node = $nodes[$i];
if ($node eq $current_node) {
$found = $i;
- last;
}
}
if ($try_next) {
-
if (defined($found) && ($found < (scalar(@nodes) - 1))) {
return $nodes[$found + 1];
} else {
return $nodes[0];
}
-
} else {
-
- return $nodes[$found] if defined($found);
-
return $nodes[0];
-
}
}
my $valid_service_states = {
stopped => 1,
request_stop => 1,
+ request_start => 1,
started => 1,
fence => 1,
+ recovery => 1,
migrate => 1,
relocate => 1,
freeze => 1,
error => 1,
};
+# FIXME with 'static' mode and thousands of services, the overhead can be noticable and the fact
+# that this function is called for each state change and upon recovery doesn't help.
sub recompute_online_node_usage {
my ($self) = @_;
- my $online_node_usage = {};
+ my $haenv = $self->{haenv};
my $online_nodes = $self->{ns}->list_online_nodes();
- foreach my $node (@$online_nodes) {
- $online_node_usage->{$node} = 0;
+ my $online_node_usage;
+
+ if (my $mode = $self->{crs}->{scheduler}) {
+ if ($mode eq 'static') {
+ $online_node_usage = eval {
+ my $scheduler = PVE::HA::Usage::Static->new($haenv);
+ $scheduler->add_node($_) for $online_nodes->@*;
+ return $scheduler;
+ };
+ $haenv->log('warning', "fallback to 'basic' scheduler mode, init for 'static' failed - $@")
+ if $@;
+ } elsif ($mode eq 'basic') {
+ # handled below in the general fall-back case
+ } else {
+ $haenv->log('warning', "got unknown scheduler mode '$mode', using 'basic'");
+ }
+ }
+
+ # fallback to the basic algorithm in any case
+ if (!$online_node_usage) {
+ $online_node_usage = PVE::HA::Usage::Basic->new($haenv);
+ $online_node_usage->add_node($_) for $online_nodes->@*;
}
- foreach my $sid (keys %{$self->{ss}}) {
+ foreach my $sid (sort keys %{$self->{ss}}) {
my $sd = $self->{ss}->{$sid};
my $state = $sd->{state};
- if (defined($online_node_usage->{$sd->{node}})) {
- if (($state eq 'started') || ($state eq 'request_stop') ||
- ($state eq 'fence') || ($state eq 'freeze') || ($state eq 'error')) {
- $online_node_usage->{$sd->{node}}++;
+ my $target = $sd->{target}; # optional
+ if ($online_node_usage->contains_node($sd->{node})) {
+ if (
+ $state eq 'started' || $state eq 'request_stop' || $state eq 'fence'
+ || $state eq 'freeze' || $state eq 'error' || $state eq 'recovery'
+ ) {
+ $online_node_usage->add_service_usage_to_node($sd->{node}, $sid, $sd->{node});
} elsif (($state eq 'migrate') || ($state eq 'relocate')) {
- $online_node_usage->{$sd->{target}}++;
- } elsif ($state eq 'stopped') {
+ my $source = $sd->{node};
+ # count it for both, source and target as load is put on both
+ $online_node_usage->add_service_usage_to_node($source, $sid, $source, $target);
+ $online_node_usage->add_service_usage_to_node($target, $sid, $source, $target);
+ } elsif ($state eq 'stopped' || $state eq 'request_start') {
# do nothing
} else {
- die "should not be reached";
+ die "should not be reached (sid = '$sid', state = '$state')";
+ }
+ } elsif (defined($target) && $online_node_usage->contains_node($target)) {
+ if ($state eq 'migrate' || $state eq 'relocate') {
+ # to correctly track maintenance modi and also consider the target as used for the
+ # case a node dies, as we cannot really know if the to-be-aborted incoming migration
+ # has already cleaned up all used resources
+ $online_node_usage->add_service_usage_to_node($target, $sid, $sd->{node}, $target);
}
}
}
my $old_state = $sd->{state};
my $old_node = $sd->{node};
my $old_failed_nodes = $sd->{failed_nodes};
+ my $old_maintenance_node = $sd->{maintenance_node};
die "no state change" if $old_state eq $new_state; # just to be sure
$sd->{state} = $new_state;
$sd->{node} = $old_node;
- $sd->{failed_nodes} = $old_failed_nodes;
+ $sd->{failed_nodes} = $old_failed_nodes if defined($old_failed_nodes);
+ $sd->{maintenance_node} = $old_maintenance_node if defined($old_maintenance_node);
my $text_state = '';
foreach my $k (sort keys %params) {
# should not happen
die "unknown resource type '$type'" if !$plugin;
- # locks may block recovery, cleanup those which are safe to remove after fencing
- my $removable_locks = ['backup', 'mounted'];
+ # locks may block recovery, cleanup those which are safe to remove after fencing,
+ # i.e., after the original node was reset and thus all it's state
+ my $removable_locks = [
+ 'backup',
+ 'mounted',
+ 'migrate',
+ 'clone',
+ 'rollback',
+ 'snapshot',
+ 'snapshot-delete',
+ 'suspending',
+ 'suspended',
+ ];
if (my $removed_lock = $plugin->remove_locks($haenv, $id, $removable_locks, $fenced_node)) {
$haenv->log('warning', "removed leftover lock '$removed_lock' from recovered " .
"service '$sid' to allow its start.");
}
};
-# after a node was fenced this recovers the service to a new node
-my $recover_fenced_service = sub {
- my ($self, $sid, $cd) = @_;
-
- my ($haenv, $ss) = ($self->{haenv}, $self->{ss});
-
- my $sd = $ss->{$sid};
-
- if ($sd->{state} ne 'fence') { # should not happen
- $haenv->log('err', "cannot recover service '$sid' from fencing," .
- " wrong state '$sd->{state}'");
- return;
- }
-
- my $fenced_node = $sd->{node}; # for logging purpose
-
- $self->recompute_online_node_usage(); # we want the most current node state
-
- my $recovery_node = select_service_node($self->{groups},
- $self->{online_node_usage},
- $cd, $sd->{node});
-
- if ($recovery_node) {
- $haenv->log('info', "recover service '$sid' from fenced node " .
- "'$fenced_node' to node '$recovery_node'");
-
- &$fence_recovery_cleanup($self, $sid, $fenced_node);
-
- $haenv->steal_service($sid, $sd->{node}, $recovery_node);
-
- # $sd *is normally read-only*, fencing is the exception
- $cd->{node} = $sd->{node} = $recovery_node;
- my $new_state = ($cd->{state} eq 'started') ? 'started' : 'request_stop';
- &$change_service_state($self, $sid, $new_state, node => $recovery_node);
- } else {
- # no possible node found, cannot recover
- $haenv->log('err', "recovering service '$sid' from fenced node " .
- "'$fenced_node' failed, no recovery node found");
- &$change_service_state($self, $sid, 'error');
- }
-};
-
# read LRM status for all nodes
sub read_lrm_status {
my ($self) = @_;
}
}
-
return ($results, $modes);
}
$haenv->log('info', "ignore crm command - service already on target node: $cmd");
} else {
$haenv->log('info', "got crm command: $cmd");
- $ss->{$sid}->{cmd} = [ $task, $node];
+ $ss->{$sid}->{cmd} = [ $task, $node ];
}
}
} else {
$haenv->log('err', "crm command error - no such service: $cmd");
}
+ } elsif ($cmd =~ m/^stop\s+(\S+)\s+(\S+)$/) {
+ my ($sid, $timeout) = ($1, $2);
+ if (my $sd = $ss->{$sid}) {
+ $haenv->log('info', "got crm command: $cmd");
+ $ss->{$sid}->{cmd} = [ 'stop', $timeout ];
+ } else {
+ $haenv->log('err', "crm command error - no such service: $cmd");
+ }
} else {
$haenv->log('err', "unable to parse crm command: $cmd");
}
my ($haenv, $ms, $ns, $ss) = ($self->{haenv}, $self->{ms}, $self->{ns}, $self->{ss});
- $ns->update($haenv->get_node_info());
+ my ($node_info) = $haenv->get_node_info();
+ my ($lrm_results, $lrm_modes) = $self->read_lrm_status();
+
+ $ns->update($node_info, $lrm_modes);
- if (!$ns->node_is_online($haenv->nodename())) {
+ if (!$ns->node_is_operational($haenv->nodename())) {
$haenv->log('info', "master seems offline");
return;
}
- my ($lrm_results, $lrm_modes) = $self->read_lrm_status();
+ $self->update_crs_scheduler_mode();
my $sc = $haenv->read_service_config();
$haenv->log('info', "adding new service '$sid' on node '$cd->{node}'");
# assume we are running to avoid relocate running service at add
- my $state = ($cd->{state} eq 'started') ? 'started' : 'request_stop';
- $ss->{$sid} = { state => $state, node => $cd->{node},
- uid => compute_new_uuid('started') };
+ my $state = ($cd->{state} eq 'started') ? 'request_start' : 'request_stop';
+ $ss->{$sid} = {
+ state => $state, node => $cd->{node}, uid => compute_new_uuid('started'),
+ };
}
# remove stale or ignored services from manager state
$self->next_state_started($sid, $cd, $sd, $lrm_res);
+ } elsif ($last_state eq 'request_start') {
+
+ $self->next_state_request_start($sid, $cd, $sd, $lrm_res);
+
} elsif ($last_state eq 'migrate' || $last_state eq 'relocate') {
$self->next_state_migrate_relocate($sid, $cd, $sd, $lrm_res);
# do nothing here - wait until fenced
+ } elsif ($last_state eq 'recovery') {
+
+ $self->next_state_recovery($sid, $cd, $sd, $lrm_res);
+
} elsif ($last_state eq 'request_stop') {
$self->next_state_request_stop($sid, $cd, $sd, $lrm_res);
} elsif ($last_state eq 'freeze') {
my $lrm_mode = $sd->{node} ? $lrm_modes->{$sd->{node}} : undef;
- # unfreeze
- my $state = ($cd->{state} eq 'started') ? 'started' : 'request_stop';
- &$change_service_state($self, $sid, $state)
- if $lrm_mode && $lrm_mode eq 'active';
+ if ($lrm_mode && $lrm_mode eq 'active') { # unfreeze if active again
+ my $state = ($cd->{state} eq 'started') ? 'started' : 'request_stop';
+ $change_service_state->($self, $sid, $state);
+ }
} elsif ($last_state eq 'error') {
my $lrm_mode = $sd->{node} ? $lrm_modes->{$sd->{node}} : undef;
if ($lrm_mode && $lrm_mode eq 'restart') {
- if (($sd->{state} eq 'started' || $sd->{state} eq 'stopped' ||
- $sd->{state} eq 'request_stop')) {
- &$change_service_state($self, $sid, 'freeze');
+ my $state = $sd->{state};
+ if ($state eq 'started' || $state eq 'stopped'|| $state eq 'request_stop') {
+ $change_service_state->($self, $sid, 'freeze');
}
}
# handle fencing
my $fenced_nodes = {};
foreach my $sid (sort keys %$ss) {
- my $sd = $ss->{$sid};
- next if $sd->{state} ne 'fence';
+ my ($service_state, $service_node) = $ss->{$sid}->@{'state', 'node'};
+ next if $service_state ne 'fence';
- if (!defined($fenced_nodes->{$sd->{node}})) {
- $fenced_nodes->{$sd->{node}} = $ns->fence_node($sd->{node}) || 0;
+ if (!defined($fenced_nodes->{$service_node})) {
+ $fenced_nodes->{$service_node} = $ns->fence_node($service_node) || 0;
}
- next if !$fenced_nodes->{$sd->{node}};
+ next if !$fenced_nodes->{$service_node};
# node fence was successful - recover service
- &$recover_fenced_service($self, $sid, $sc->{$sid});
+ $change_service_state->($self, $sid, 'recovery');
+ $repeat = 1; # for faster recovery execution
+ }
+
+ # Avoid that a node without services in 'fence' state (e.g., removed
+ # manually by admin) is stuck with the 'fence' node state.
+ for my $node (sort grep { !defined($fenced_nodes->{$_}) } keys $ns->{status}->%*) {
+ next if $ns->get_node_state($node) ne 'fence';
+
+ $haenv->log('notice', "node '$node' in fence state but no services to-fence! admin interference?!");
+ $repeat = 1 if $ns->fence_node($node);
}
last if !$repeat;
}
}
-
sub next_state_stopped {
my ($self, $sid, $cd, $sd, $lrm_res) = @_;
}
if ($sd->{cmd}) {
- my ($cmd, $target) = @{$sd->{cmd}};
- delete $sd->{cmd};
+ my $cmd = shift @{$sd->{cmd}};
if ($cmd eq 'migrate' || $cmd eq 'relocate') {
+ my $target = shift @{$sd->{cmd}};
if (!$ns->node_is_online($target)) {
$haenv->log('err', "ignore service '$sid' $cmd request - node '$target' not online");
} elsif ($sd->{node} eq $target) {
$haenv->log('info', "ignore service '$sid' $cmd request - service already on node '$target'");
} else {
- &$change_service_state($self, $sid, $cmd, node => $sd->{node},
- target => $target);
+ &$change_service_state($self, $sid, $cmd, node => $sd->{node}, target => $target);
return;
}
+ } elsif ($cmd eq 'stop') {
+ $haenv->log('info', "ignore service '$sid' $cmd request - service already stopped");
} else {
$haenv->log('err', "unknown command '$cmd' for service '$sid'");
}
+ delete $sd->{cmd};
}
if ($cd->{state} eq 'disabled') {
return;
}
- if ($ns->node_is_offline_delayed($sd->{node})) {
+ if ($ns->node_is_offline_delayed($sd->{node}) && $ns->get_node_state($sd->{node}) ne 'maintenance') {
&$change_service_state($self, $sid, 'fence');
return;
}
}
if ($cd->{state} eq 'started') {
- # simply mark it started, if it's on the wrong node
- # next_state_started will fix that for us
- &$change_service_state($self, $sid, 'started', node => $sd->{node});
+ # simply mark it started, if it's on the wrong node next_state_started will fix that for us
+ $change_service_state->($self, $sid, 'request_start', node => $sd->{node});
return;
}
$haenv->log('err', "service '$sid' - unknown state '$cd->{state}' in service configuration");
}
+sub next_state_request_start {
+ my ($self, $sid, $cd, $sd, $lrm_res) = @_;
+
+ $change_service_state->($self, $sid, 'started', node => $sd->{node});
+}
+
sub record_service_failed_on_node {
my ($self, $sid, $node) = @_;
if ($ns->node_is_offline_delayed($sd->{node})) {
&$change_service_state($self, $sid, 'fence');
}
- return;
+ if ($ns->get_node_state($sd->{node}) ne 'maintenance') {
+ return;
+ } else {
+ # save current node as fallback for when it comes out of
+ # maintenance
+ $sd->{maintenance_node} = $sd->{node};
+ }
}
if ($cd->{state} eq 'disabled' || $cd->{state} eq 'stopped') {
if ($cd->{state} eq 'started') {
if ($sd->{cmd}) {
- my ($cmd, $target) = @{$sd->{cmd}};
- delete $sd->{cmd};
+ my $cmd = shift @{$sd->{cmd}};
if ($cmd eq 'migrate' || $cmd eq 'relocate') {
+ my $target = shift @{$sd->{cmd}};
if (!$ns->node_is_online($target)) {
$haenv->log('err', "ignore service '$sid' $cmd request - node '$target' not online");
} elsif ($sd->{node} eq $target) {
$haenv->log('info', "$cmd service '$sid' to node '$target'");
&$change_service_state($self, $sid, $cmd, node => $sd->{node}, target => $target);
}
+ } elsif ($cmd eq 'stop') {
+ my $timeout = shift @{$sd->{cmd}};
+ if ($timeout == 0) {
+ $haenv->log('info', "request immediate service hard-stop for service '$sid'");
+ } else {
+ $haenv->log('info', "request graceful stop with timeout '$timeout' for service '$sid'");
+ }
+ &$change_service_state($self, $sid, 'request_stop', timeout => $timeout);
+ $haenv->update_service_config($sid, {'state' => 'stopped'});
} else {
$haenv->log('err', "unknown command '$cmd' for service '$sid'");
}
+
+ delete $sd->{cmd};
+
} else {
my $try_next = 0;
# store flag to indicate successful start - only valid while state == 'started'
$sd->{running} = 1;
- } elsif ($ec == ERROR) {
+ } elsif ($ec == ERROR || $ec == EWRONG_NODE) {
delete $sd->{running};
}
}
- my $node = select_service_node($self->{groups}, $self->{online_node_usage},
- $cd, $sd->{node}, $try_next, $sd->{failed_nodes});
+ my $node = select_service_node(
+ $self->{groups},
+ $self->{online_node_usage},
+ $sid,
+ $cd,
+ $sd->{node},
+ $try_next,
+ $sd->{failed_nodes},
+ $sd->{maintenance_node},
+ );
if ($node && ($sd->{node} ne $node)) {
+ $self->{online_node_usage}->add_service_usage_to_node($node, $sid, $sd->{node});
+
+ if (defined(my $fallback = $sd->{maintenance_node})) {
+ if ($node eq $fallback) {
+ $haenv->log('info', "moving service '$sid' back to '$fallback', node came back from maintenance.");
+ delete $sd->{maintenance_node};
+ } elsif ($sd->{node} ne $fallback) {
+ $haenv->log('info', "dropping maintenance fallback node '$fallback' for '$sid'");
+ delete $sd->{maintenance_node};
+ }
+ }
+
if ($cd->{type} eq 'vm') {
$haenv->log('info', "migrate service '$sid' to node '$node' (running)");
&$change_service_state($self, $sid, 'migrate', node => $sd->{node}, target => $node);
}
+# after a node was fenced this recovers the service to a new node
+sub next_state_recovery {
+ my ($self, $sid, $cd, $sd, $lrm_res) = @_;
+
+ my ($haenv, $ss) = ($self->{haenv}, $self->{ss});
+ my $ns = $self->{ns};
+ my $ms = $self->{ms};
+
+ if ($sd->{state} ne 'recovery') { # should not happen
+ $haenv->log('err', "cannot recover service '$sid' from fencing, wrong state '$sd->{state}'");
+ return;
+ }
+
+ my $fenced_node = $sd->{node}; # for logging purpose
+
+ $self->recompute_online_node_usage(); # we want the most current node state
+
+ my $recovery_node = select_service_node(
+ $self->{groups},
+ $self->{online_node_usage},
+ $sid,
+ $cd,
+ $sd->{node},
+ );
+
+ if ($recovery_node) {
+ my $msg = "recover service '$sid' from fenced node '$fenced_node' to node '$recovery_node'";
+ if ($recovery_node eq $fenced_node) {
+ # can happen if restriced groups and the node came up again OK
+ $msg = "recover service '$sid' to previous failed and fenced node '$fenced_node' again";
+ }
+ $haenv->log('info', "$msg");
+
+ $fence_recovery_cleanup->($self, $sid, $fenced_node);
+
+ $haenv->steal_service($sid, $sd->{node}, $recovery_node);
+ $self->{online_node_usage}->add_service_usage_to_node($recovery_node, $sid, $recovery_node);
+
+ # NOTE: $sd *is normally read-only*, fencing is the exception
+ $cd->{node} = $sd->{node} = $recovery_node;
+ my $new_state = ($cd->{state} eq 'started') ? 'started' : 'request_stop';
+ $change_service_state->($self, $sid, $new_state, node => $recovery_node);
+ } else {
+ # no possible node found, cannot recover - but retry later, as we always try to make it available
+ $haenv->log('err', "recovering service '$sid' from fenced node '$fenced_node' failed, no recovery node found");
+
+ if ($cd->{state} eq 'disabled') {
+ # allow getting a service out of recovery manually if an admin disables it.
+ delete $sd->{failed_nodes}; # clean up on recovery to stopped
+ $change_service_state->($self, $sid, 'stopped'); # must NOT go through request_stop
+ return;
+ }
+ }
+}
+
1;
projects / pve-ha-manager.git / blobdiff