Allow to reboot down nodes KIKIMR-8420

5 files changed, 55 insertions, 58 deletions

diff --git a/ydb/core/cms/cluster_info.cpp b/ydb/core/cms/cluster_info.cpp
index 0893fa2ebc..d14f6de637 100644
--- a/ydb/core/cms/cluster_info.cpp
+++ b/ydb/core/cms/cluster_info.cpp
@@ -646,6 +646,7 @@ void TClusterInfo::ApplyActionWithoutLog(const NKikimrCms::TAction &action)
     switch (action.GetType()) {
     case TAction::RESTART_SERVICES:
     case TAction::SHUTDOWN_HOST:
+    case TAction::REBOOT_HOST:
         if (auto nodes = NodePtrs(action.GetHost(), MakeServices(action))) {
             for (const auto node : nodes) {
                 for (auto &nodeGroup: node->NodeGroups)
@@ -696,6 +697,7 @@ TSet<TLockableItem *> TClusterInfo::FindLockedItems(const NKikimrCms::TAction &a
     switch (action.GetType()) {
     case TAction::RESTART_SERVICES:
     case TAction::SHUTDOWN_HOST:
+    case TAction::REBOOT_HOST:
         if (auto nodes = NodePtrs(action.GetHost(), MakeServices(action))) {
             for (const auto node : nodes) {
                 res.insert(node);
@@ -750,6 +752,7 @@ ui64 TClusterInfo::AddLocks(const TPermissionInfo &permission, const TActorConte
         if (item->State == DOWN
             && (permission.Action.GetType() == TAction::RESTART_SERVICES
                 || permission.Action.GetType() == TAction::SHUTDOWN_HOST
+                || permission.Action.GetType() == TAction::REBOOT_HOST
                 || permission.Action.GetType() == TAction::REPLACE_DEVICES)) {
             item->State = RESTART;
             lock = true;;
@@ -1001,6 +1004,7 @@ void TOperationLogManager::ApplyAction(const NKikimrCms::TAction &action,
     switch (action.GetType()) {
     case NKikimrCms::TAction::RESTART_SERVICES:
     case NKikimrCms::TAction::SHUTDOWN_HOST:
+    case NKikimrCms::TAction::REBOOT_HOST:
         if (auto nodes = clusterState->NodePtrs(action.GetHost(), MakeServices(action))) {
             for (const auto node : nodes) {
                 for (auto &nodeGroup: node->NodeGroups)
diff --git a/ydb/core/cms/cms.cpp b/ydb/core/cms/cms.cpp
index ffca58da2c..0834ef50c4 100644
--- a/ydb/core/cms/cms.cpp
+++ b/ydb/core/cms/cms.cpp
@@ -436,6 +436,7 @@ void TCms::AddPermissionExtensions(const TAction& action, TPermission& perm) con
     switch (action.GetType()) {
         case TAction::RESTART_SERVICES:
         case TAction::SHUTDOWN_HOST:
+        case TAction::REBOOT_HOST:
             AddHostExtensions(action.GetHost(), perm);
             break;
         default:
@@ -493,6 +494,7 @@ bool TCms::CheckAction(const TAction &action,
         case TAction::RESTART_SERVICES:
             return CheckActionRestartServices(action, opts, error, ctx);
         case TAction::SHUTDOWN_HOST:
+        case TAction::REBOOT_HOST:
             return CheckActionShutdownHost(action, opts, error, ctx);
         case TAction::REPLACE_DEVICES:
             return CheckActionReplaceDevices(action, opts.PermissionDuration, error);
@@ -581,8 +583,12 @@ bool TCms::CheckActionShutdownHost(const TAction &action,
                                    TErrorInfo &error,
                                    const TActorContext &ctx) const
 {
+    const bool forciblyAllow = action.GetType() == TAction::REBOOT_HOST;
     for (const auto node : ClusterInfo->HostNodes(action.GetHost())) {
         if (!CheckActionShutdownNode(action, opts, *node, error, ctx)) {
+            if (forciblyAllow && node->State == DOWN) {
+                continue;
+            }
             return false;
         }
     }
diff --git a/ydb/core/cms/cms_ut.cpp b/ydb/core/cms/cms_ut.cpp
index 3ff189cbd2..a82ea3f608 100644
--- a/ydb/core/cms/cms_ut.cpp
+++ b/ydb/core/cms/cms_ut.cpp
@@ -729,6 +729,19 @@ Y_UNIT_TEST_SUITE(TCmsTest) {
         env.CheckWalleCheckTask("task-2", TStatus::ALLOW, env.GetNodeId(1));
     }
 
+    Y_UNIT_TEST(WalleRebootDownNode)
+    {
+        TCmsTestEnv env(8);
+
+        // alllow
+        env.CheckWalleCreateTask("task-1", "reboot", false, TStatus::ALLOW, env.GetNodeId(0));
+        // disallow (up)
+        env.CheckWalleCreateTask("task-2", "reboot", false, TStatus::DISALLOW_TEMP, env.GetNodeId(1));
+        // allow (down)
+        TFakeNodeWhiteboardService::Info[env.GetNodeId(1)].Connected = false;
+        env.CheckWalleCheckTask("task-2", TStatus::ALLOW, env.GetNodeId(1));
+    }
+
     Y_UNIT_TEST(Notifications)
     {
         TCmsTestEnv env(8);
diff --git a/ydb/core/cms/walle_create_task_adapter.cpp b/ydb/core/cms/walle_create_task_adapter.cpp
index c1a2866e37..5c7ce0393b 100644
--- a/ydb/core/cms/walle_create_task_adapter.cpp
+++ b/ydb/core/cms/walle_create_task_adapter.cpp
@@ -5,6 +5,8 @@
 #include <library/cpp/actors/core/actor_bootstrapped.h>
 #include <library/cpp/actors/core/hfunc.h>
 
+#include <optional>
+
 namespace NKikimr::NCms {
 
 using namespace NKikimrCms;
@@ -27,19 +29,7 @@ public:
         LOG_INFO(ctx, NKikimrServices::CMS, "Processing Wall-E request: %s",
                   rec.ShortDebugString().data());
 
-        if (rec.GetAction() != "reboot"
-            && rec.GetAction() != "power-off"
-            && rec.GetAction() != "change-disk"
-            && rec.GetAction() != "change-memory"
-            && rec.GetAction() != "profile"
-            && rec.GetAction() != "redeploy"
-            && rec.GetAction() != "prepare"
-            && rec.GetAction() != "repair-link"
-            && rec.GetAction() != "repair-bmc"
-            && rec.GetAction() != "repair-overheat"
-            && rec.GetAction() != "repair-capping"
-            && rec.GetAction() != "deactivate"
-            && rec.GetAction() != "temporary-unreachable") {
+        if (!Actions.contains(rec.GetAction())) {
             ReplyWithErrorAndDie(TStatus::WRONG_REQUEST, "Unsupported action", ctx);
             return;
         }
@@ -132,9 +122,10 @@ private:
         request->Record.SetSchedule(true);
         request->Record.SetDryRun(task.GetDryRun());
 
-        TAction action;
-        if (task.GetAction() == "prepare"
-            || task.GetAction() == "deactivate") {
+        auto it = Actions.find(task.GetAction());
+        Y_ABORT_UNLESS(it != Actions.end());
+
+        if (!it->second) {
             TAutoPtr<TEvCms::TEvWalleCreateTaskResponse> resp = new TEvCms::TEvWalleCreateTaskResponse;
             resp->Record.SetTaskId(task.GetTaskId());
             resp->Record.MutableHosts()->CopyFrom(task.GetHosts());
@@ -142,52 +133,17 @@ private:
             ReplyAndDie(resp.Release(), ctx);
             return;
         } else {
-            // We always use infinite duration.
-            // Wall-E MUST delete processed tasks.
-            if (task.GetAction() == "reboot") {
-                action.SetType(TAction::SHUTDOWN_HOST);
-                action.SetDuration(TDuration::Max().GetValue());
-            } else if (task.GetAction() == "power-off") {
-                action.SetType(TAction::SHUTDOWN_HOST);
-                action.SetDuration(TDuration::Max().GetValue());
-            } else if (task.GetAction() == "change-disk") {
-                action.SetType(TAction::REPLACE_DEVICES);
-                action.SetDuration(TDuration::Max().GetValue());
-            } else if (task.GetAction() == "change-memory") {
-                action.SetType(TAction::SHUTDOWN_HOST);
-                action.SetDuration(TDuration::Max().GetValue());
-            } else if (task.GetAction() == "profile") {
-                action.SetType(TAction::SHUTDOWN_HOST);
-                action.SetDuration(TDuration::Max().GetValue());
-            } else if (task.GetAction() == "redeploy") {
-                action.SetType(TAction::SHUTDOWN_HOST);
-                action.SetDuration(TDuration::Max().GetValue());
-            } else if (task.GetAction() == "repair-link") {
-                action.SetType(TAction::SHUTDOWN_HOST);
-                action.SetDuration(TDuration::Max().GetValue());
-            } else if (task.GetAction() == "repair-bmc") {
-                action.SetType(TAction::SHUTDOWN_HOST);
-                action.SetDuration(TDuration::Max().GetValue());
-            } else if (task.GetAction() == "repair-overheat") {
-                action.SetType(TAction::SHUTDOWN_HOST);
-                action.SetDuration(TDuration::Max().GetValue());
-            } else if (task.GetAction() == "repair-capping") {
-                action.SetType(TAction::SHUTDOWN_HOST);
-                action.SetDuration(TDuration::Max().GetValue());
-            } else if (task.GetAction() == "temporary-unreachable") {
-                action.SetType(TAction::SHUTDOWN_HOST);
-                action.SetDuration(TDuration::Max().GetValue());
-            } else
-                Y_ABORT("Unknown action");
-
             for (auto &host : task.GetHosts()) {
-                auto &hostAction = *request->Record.AddActions();
-                hostAction.CopyFrom(action);
-                hostAction.SetHost(host);
+                auto &action = *request->Record.AddActions();
+                action.SetHost(host);
+                action.SetType(*it->second);
+                // We always use infinite duration.
+                // Wall-E MUST delete processed tasks.
+                action.SetDuration(TDuration::Max().GetValue());
                 if (action.GetType() == TAction::REPLACE_DEVICES) {
                     for (const auto node : cluster->HostNodes(host)) {
                         for (auto &pdiskId : node->PDisks)
-                            *hostAction.AddDevices() = cluster->PDisk(pdiskId).GetDeviceName();
+                            *action.AddDevices() = cluster->PDisk(pdiskId).GetDeviceName();
                     }
                 }
             }
@@ -204,11 +160,28 @@ private:
         ReplyAndDie(Response, ctx);
     }
 
+    static const THashMap<TString, std::optional<TAction::EType>> Actions;
     TEvCms::TEvWalleCreateTaskRequest::TPtr RequestEvent;
     TAutoPtr<TEvCms::TEvWalleCreateTaskResponse> Response;
     TActorId Cms;
 };
 
+const THashMap<TString, std::optional<TAction::EType>> TWalleCreateTaskAdapter::Actions = {
+    {"reboot", TAction::REBOOT_HOST},
+    {"power-off", TAction::SHUTDOWN_HOST},
+    {"change-disk", TAction::REPLACE_DEVICES},
+    {"change-memory", TAction::SHUTDOWN_HOST},
+    {"profile", TAction::SHUTDOWN_HOST},
+    {"redeploy", TAction::SHUTDOWN_HOST},
+    {"repair-link", TAction::SHUTDOWN_HOST},
+    {"repair-bmc", TAction::SHUTDOWN_HOST},
+    {"repair-overheat", TAction::SHUTDOWN_HOST},
+    {"repair-capping", TAction::SHUTDOWN_HOST},
+    {"temporary-unreachable", TAction::SHUTDOWN_HOST},
+    {"prepare", std::nullopt},
+    {"deactivate", std::nullopt},
+};
+
 IActor *CreateWalleAdapter(TEvCms::TEvWalleCreateTaskRequest::TPtr &ev, TActorId cms) {
     return new TWalleCreateTaskAdapter(ev, cms);
 }
diff --git a/ydb/core/protos/cms.proto b/ydb/core/protos/cms.proto
index 6658e0611f..86ee47ce66 100644
--- a/ydb/core/protos/cms.proto
+++ b/ydb/core/protos/cms.proto
@@ -106,6 +106,7 @@ message TAction {
     ADD_DEVICES = 7;
     REPLACE_DEVICES = 8;
     REMOVE_DEVICES = 9;
+    REBOOT_HOST = 10; // Same as SHUTDOWN_HOST, but forcibly allowed if host is down
   }
 
   optional EType     Type = 1;