NBS-5637: Asynchronous disks allocation

cloud/blockstore/libs/storage/disk_registry/disk_registry_actor.h

@@ -83,6 +83,7 @@ class TDiskRegistryActor final
     // Pending requests
     TDeque<TPendingRequest> PendingRequests;
 
+    THashMap<TDiskId, TVector<TRequestInfoPtr>> PendingDiskAllocationRequests;
     THashMap<TDiskId, TVector<TRequestInfoPtr>> PendingDiskDeallocationRequests;
 
     bool BrokenDisksDestructionInProgress = false;
@@ -227,6 +228,20 @@ class TDiskRegistryActor final
         TDiskRegistryDatabase& db,
         TDiskRegistryStateSnapshot& args);
 
+    void AddPendingAllocation(
+        const NActors::TActorContext& ctx,
+        const TString& diskId,
+        TRequestInfoPtr requestInfoPtr);
+
+    void ReplyToPendingAllocations(
+        const NActors::TActorContext& ctx,
+        const TString& diskId);
+
+    void ReplyToPendingAllocations(
+        const NActors::TActorContext& ctx,
+        TVector<TRequestInfoPtr>& requestInfos,
+        NProto::TError error);
+
     void AddPendingDeallocation(
         const NActors::TActorContext& ctx,
         const TString& diskId,

cloud/blockstore/libs/storage/disk_registry/disk_registry_actor_allocate.cpp

@@ -145,6 +145,7 @@ void TDiskRegistryActor::CompleteAddDisk(
 {
     auto response = std::make_unique<TEvDiskRegistry::TEvAllocateDiskResponse>();
 
+    TVector<TDeviceId> dirtyDevices;
     if (HasError(args.Error)) {
         LOG_ERROR(ctx, TBlockStoreComponents::DISK_REGISTRY,
             "[%lu] AddDisk error: %s. DiskId=%s",
@@ -154,6 +155,14 @@ void TDiskRegistryActor::CompleteAddDisk(
 
         *response->Record.MutableError() = std::move(args.Error);
     } else {
+        // TODO: split this method, too complex
+        for (const auto& device: args.Devices) {
+            const bool dirty = State->IsDirtyDevice(device.GetDeviceUUID());
+            if (dirty) {
+                dirtyDevices.push_back(device.GetDeviceUUID());
+            }
+        }
+
         TStringBuilder devices;
         OutputDevices(args.Devices, devices);
 
@@ -240,12 +249,71 @@ void TDiskRegistryActor::CompleteAddDisk(
         response->Record.SetMuteIOErrors(args.MuteIOErrors);
     }
 
-    NCloud::Reply(ctx, *args.RequestInfo, std::move(response));
+    if (HasError(args.Error) || dirtyDevices.empty()) {
+        NCloud::Reply(ctx, *args.RequestInfo, std::move(response));
+    } else {
+        AddPendingAllocation(ctx, args.DiskId, args.RequestInfo);
+    }
 
     DestroyBrokenDisks(ctx);
+    SecureErase(ctx); // I think we need this only if allocating dirty devices
     NotifyUsers(ctx);
 }
 
+void TDiskRegistryActor::AddPendingAllocation(
+    const NActors::TActorContext& ctx,
+    const TString& diskId,
+    TRequestInfoPtr requestInfo)
+{
+    auto& requestInfos = PendingDiskAllocationRequests[diskId];
+
+    // TODO: GetMaxNonReplicatedDiskAllocationRequests
+    if (requestInfos.size() > Config->GetMaxNonReplicatedDiskDeallocationRequests()) {
+        LOG_WARN(ctx, TBlockStoreComponents::DISK_REGISTRY,
+            "Too many pending allocation requests (%lu) for disk %s. "
+            "Reject all requests.",
+            requestInfos.size(),
+            diskId.Quote().c_str());
+
+        ReplyToPendingAllocations(ctx, requestInfos, MakeError(E_REJECTED));
+    }
+
+    requestInfos.emplace_back(std::move(requestInfo));
+}
+
+void TDiskRegistryActor::ReplyToPendingAllocations(
+    const NActors::TActorContext& ctx,
+    TVector<TRequestInfoPtr>& requestInfos,
+    NProto::TError error)
+{
+    for (auto& requestInfo: requestInfos) {
+        NCloud::Reply(
+            ctx,
+            *requestInfo,
+            std::make_unique<TEvDiskRegistry::TEvAllocateDiskResponse>(error));
+    }
+    requestInfos.clear();
+}
+
+void TDiskRegistryActor::ReplyToPendingAllocations(
+    const NActors::TActorContext& ctx,
+    const TString& diskId)
+{
+    auto it = PendingDiskAllocationRequests.find(diskId);
+    if (it == PendingDiskAllocationRequests.end()) {
+        return;
+    }
+
+    LOG_INFO(ctx, TBlockStoreComponents::DISK_REGISTRY,
+        "Reply to pending allocation requests. DiskId=%s PendingRequests=%d",
+        diskId.Quote().c_str(),
+        static_cast<int>(it->second.size()));
+
+    ReplyToPendingAllocations(ctx, it->second, MakeError(S_OK));
+
+    PendingDiskAllocationRequests.erase(it);
+}
+
 ////////////////////////////////////////////////////////////////////////////////
 
 void TDiskRegistryActor::HandleDeallocateDisk(
@@ -379,7 +447,7 @@ void TDiskRegistryActor::ReplyToPendingDeallocations(
     }
 
     LOG_INFO(ctx, TBlockStoreComponents::DISK_REGISTRY,
-        "Reply to pending deallocation requests. DiskId=%s PendingRquests=%d",
+        "Reply to pending deallocation requests. DiskId=%s PendingRequests=%d",
         diskId.Quote().c_str(),
         static_cast<int>(it->second.size()));
 

cloud/blockstore/libs/storage/disk_registry/disk_registry_actor_secure_erase.cpp

@@ -279,8 +279,8 @@ void TDiskRegistryActor::ExecuteCleanupDevices(
     TTxDiskRegistry::TCleanupDevices& args)
 {
     TDiskRegistryDatabase db(tx.DB);
-    args.SyncDeallocatedDisks =
-        State->MarkDevicesAsClean(ctx.Now(), db, args.Devices);
+    std::tie(args.SyncAllocatedDisks, args.SyncDeallocatedDisks) =
+        std::move(State->MarkDevicesAsClean(ctx.Now(), db, args.Devices));
 }
 
 void TDiskRegistryActor::CompleteCleanupDevices(
@@ -293,6 +293,10 @@ void TDiskRegistryActor::CompleteCleanupDevices(
     for (const auto& diskId: args.SyncDeallocatedDisks) {
         ReplyToPendingDeallocations(ctx, diskId);
     }
+
+    for (const auto& diskId: args.SyncAllocatedDisks) {
+        ReplyToPendingAllocations(ctx, diskId);
+    }
 }
 
 ////////////////////////////////////////////////////////////////////////////////

cloud/blockstore/libs/storage/disk_registry/disk_registry_state.cpp

@@ -1299,16 +1299,17 @@ NProto::TError TDiskRegistryState::ReplaceDeviceWithoutDiskStateUpdate(
         }
 
         if (!manual && !deviceReplacementId.empty()) {
-            auto cleaningDiskId =
+            auto [allocating, deallocating] =
                 PendingCleanup.FindDiskId(deviceReplacementId);
-            if (!cleaningDiskId.empty() && cleaningDiskId != diskId) {
+            Y_UNUSED(allocating); // TODO: probably should react to this
+            if (deallocating && *deallocating != diskId) {
                 return MakeError(
                     E_ARGUMENT,
                     TStringBuilder()
                         << "can't allocate specific device "
                         << deviceReplacementId.Quote() << " for disk " << diskId
                         << " since it is in pending cleanup for disk "
-                        << cleaningDiskId);
+                        << *deallocating);
             }
         }
 
@@ -1414,7 +1415,7 @@ NProto::TError TDiskRegistryState::ReplaceDeviceWithoutDiskStateUpdate(
         UpdatePlacementGroup(db, diskId, disk, "ReplaceDevice");
         UpdateAndReallocateDisk(db, diskId, disk);
 
-        error = PendingCleanup.Insert(diskId, deviceId);
+        error = PendingCleanup.Insert(diskId, deviceId); // TODO: allocating/deallocating?
         if (HasError(error)) {
             ReportDiskRegistryInsertToPendingCleanupFailed(
                 TStringBuilder() << "An error occurred while replacing device: "
@@ -2164,6 +2165,19 @@ NProto::TError TDiskRegistryState::AllocateDisk(
         ReallocateCheckpointByDisk(now, db, params.DiskId);
     }
 
+    // TODO: debug print that we put devices to pending cleanup
+    // TODO: check if allowed to be cleaned
+    // TODO: we should do this only for local disks
+    TVector<TString> dirtyDevices;
+    for (const auto& uuid: disk.Devices) {
+        const bool dirty = IsDirtyDevice(uuid);
+        if (dirty) {
+            dirtyDevices.push_back(uuid);
+        }
+    }
+    NProto::TError cleanupError = PendingCleanup.Insert(params.DiskId, std::move(dirtyDevices), /*allocation=*/true);
+    Y_UNUSED(cleanupError); // TODO: check for cleanupError
+
     return allocateError;
 }
 
@@ -2851,7 +2865,7 @@ NProto::TError TDiskRegistryState::DeallocateDisk(
         JoinSeq(", ", devicesAllowedToBeCleaned).c_str());
 
     auto error =
-        PendingCleanup.Insert(diskId, std::move(devicesAllowedToBeCleaned));
+        PendingCleanup.Insert(diskId, std::move(devicesAllowedToBeCleaned), /*allocation=*/false);
     if (HasError(error)) {
         ReportDiskRegistryInsertToPendingCleanupFailed(
             TStringBuilder() << "An error occurred while deallocating disk: "
@@ -2998,6 +3012,11 @@ auto TDiskRegistryState::DeallocateSimpleDisk(
         db.UpdateDirtyDevice(uuid, diskId);
     }
 
+    auto agents = FindDiskDevicesAgents(disk);
+    for (const auto* agent: agents) {
+        DeviceList.UpdateDevices(*agent, DevicePoolConfigs);
+    }
+
     DeleteAllDeviceMigrations(diskId);
     DeleteDisk(db, diskId);
 
@@ -3855,16 +3874,20 @@ bool TDiskRegistryState::MarkDeviceAsDirty(
     return true;
 }
 
-TDiskRegistryState::TDiskId TDiskRegistryState::MarkDeviceAsClean(
+TDiskRegistryState::TOpt2Disk TDiskRegistryState::MarkDeviceAsClean(
     TInstant now,
     TDiskRegistryDatabase& db,
     const TDeviceId& uuid)
 {
-    auto ret = MarkDevicesAsClean(now, db, TVector<TDeviceId>{uuid});
-    return ret.empty() ? "" : ret[0];
+    auto [alloc, dealloc] = MarkDevicesAsClean(now, db, TVector<TDeviceId>{uuid});
+    return {
+        alloc.empty() ? std::nullopt : std::make_optional(std::move(alloc[0])),
+        dealloc.empty() ? std::nullopt
+                        : std::make_optional(std::move(dealloc[0]))};
 }
 
-TVector<TDiskRegistryState::TDiskId> TDiskRegistryState::MarkDevicesAsClean(
+std::pair<TDiskRegistryState::TAllocatedDisksList, TDiskRegistryState::TDellocatedDisksList>
+TDiskRegistryState::MarkDevicesAsClean(
     TInstant now,
     TDiskRegistryDatabase& db,
     const TVector<TDeviceId>& uuids)
@@ -3878,14 +3901,19 @@ TVector<TDiskRegistryState::TDiskId> TDiskRegistryState::MarkDevicesAsClean(
         }
     }
 
-    TVector<TDiskId> ret;
+    TAllocatedDisksList allocatedDisks;
+    TDellocatedDisksList dellocatedDisks;
     for (const auto& uuid: TryUpdateDevices(now, db, uuids)) {
-        if (auto diskId = PendingCleanup.EraseDevice(uuid); !diskId.empty()) {
-            ret.push_back(std::move(diskId));
+        auto [allocatedDisk, deallocatedDisk] = PendingCleanup.EraseDevice(uuid);
+        if (allocatedDisk) {
+            allocatedDisks.push_back(std::move(*allocatedDisk));
+        }
+        if (deallocatedDisk) {
+            dellocatedDisks.push_back(std::move(*deallocatedDisk));
         }
     }
 
-    return ret;
+    return {std::move(allocatedDisks), std::move(dellocatedDisks)};
 }
 
 bool TDiskRegistryState::TryUpdateDevice(
@@ -4811,7 +4839,7 @@ void TDiskRegistryState::RemoveFinishedMigrations(
 
             DeviceList.ReleaseDevice(m.DeviceId);
             db.UpdateDirtyDevice(m.DeviceId, diskId);
-            auto error = PendingCleanup.Insert(diskId, m.DeviceId);
+            auto error = PendingCleanup.Insert(diskId, m.DeviceId); // TODO: allocating/deallocating?
             if (HasError(error)) {
                 ReportDiskRegistryInsertToPendingCleanupFailed(
                     TStringBuilder()
@@ -5072,8 +5100,9 @@ bool TDiskRegistryState::HasDependentSsdDisks(
             continue;
         }
 
+        auto [allocating, deallocating] = PendingCleanup.FindDiskId(d.GetDeviceUUID());
         if (d.GetPoolKind() == NProto::DEVICE_POOL_KIND_LOCAL &&
-            PendingCleanup.FindDiskId(d.GetDeviceUUID()))
+            (allocating || deallocating))
         {
             return true;
         }
@@ -5598,6 +5627,24 @@ auto TDiskRegistryState::FindDeviceLocation(const TDeviceId& deviceId) const
     return const_cast<TDiskRegistryState*>(this)->FindDeviceLocation(deviceId);
 }
 
+auto TDiskRegistryState::FindDiskDevicesAgents(const TDiskState& disk) const
+    -> std::set<const NProto::TAgentConfig*>
+{
+    std::set<TString> diskAgents;
+    for (const auto& uuid: disk.Devices) {
+        // TODO: handle when not found
+        diskAgents.insert(DeviceList.FindAgentId(uuid));
+    }
+
+    std::set<const NProto::TAgentConfig*> agents;
+    for (const auto& agentId: diskAgents) {
+        // TODO: handle when not found
+        agents.insert(AgentList.FindAgent(agentId));
+    }
+
+    return std::move(agents);
+}
+
 auto TDiskRegistryState::FindDeviceLocation(const TDeviceId& deviceId)
     -> std::pair<NProto::TAgentConfig*, NProto::TDeviceConfig*>
 {
@@ -6354,7 +6401,9 @@ NProto::TDiskRegistryStateBackup TDiskRegistryState::BackupState() const
     transform(GetDirtyDevices(), backup.MutableDirtyDevices(), [this] (auto& x) {
         NProto::TDiskRegistryStateBackup::TDirtyDevice dd;
         dd.SetId(x.GetDeviceUUID());
-        dd.SetDiskId(PendingCleanup.FindDiskId(x.GetDeviceUUID()));
+        auto [allocating, deallocating] = PendingCleanup.FindDiskId(x.GetDeviceUUID()); // TODO: need to backup
+        Y_UNUSED(allocating);
+        dd.SetDiskId(*deallocating);
 
         return dd;
     });

cloud/blockstore/libs/storage/disk_registry/disk_registry_state.h

@@ -281,6 +281,15 @@ class TDiskRegistryState
     using TCheckpoints = THashMap<TCheckpointId, TCheckpointInfo>;
     using TPlacementGroups = THashMap<TString, TPlacementGroupInfo>;
 
+    using TAllocatingDiskId = TDiskRegistryState::TDiskId;
+    using TDeallocatingDiskId = TDiskRegistryState::TDiskId;
+    using TAllocatedDisksList = TVector<TAllocatingDiskId>;
+    using TDellocatedDisksList = TVector<TDeallocatingDiskId>;
+
+    template<typename T, typename U>
+    using TOpt2 = std::pair<std::optional<T>, std::optional<U>>;
+    using TOpt2Disk = TOpt2<TAllocatingDiskId, TDeallocatingDiskId>;
+
 private:
     TLog Log;
 
@@ -503,16 +512,17 @@ class TDiskRegistryState
 
     /// Mark selected device as clean and remove it
     /// from lists of suspended/dirty/pending cleanup devices
-    /// @return disk id where selected device was allocated
-    TDiskId MarkDeviceAsClean(
+    /// @return allocated/deallocated disk id of where selected device was allocated/deallocated
+    TOpt2Disk MarkDeviceAsClean(
         TInstant now,
         TDiskRegistryDatabase& db,
         const TDeviceId& uuid);
 
     /// Mark selected devices as clean and remove them
     /// from lists of suspended/dirty/pending cleanup devices
-    /// @return vector of disk ids where selected devices were allocated
-    TVector<TDiskId> MarkDevicesAsClean(
+    /// @return vector of allocated/deallocated disk ids where selected devices were allocated/deallocated
+    std::pair<TAllocatedDisksList, TDellocatedDisksList>
+    MarkDevicesAsClean(
         TInstant now,
         TDiskRegistryDatabase& db,
         const TVector<TDeviceId>& uuids);
@@ -933,6 +943,9 @@ class TDiskRegistryState
     auto FindDeviceLocation(const TDeviceId& uuid) const
         -> std::pair<const NProto::TAgentConfig*, const NProto::TDeviceConfig*>;
 
+    auto FindDiskDevicesAgents(const TDiskState& disk) const
+        -> std::set<const NProto::TAgentConfig*>;
+
     TVector<NProto::TDeviceConfig> FindDevices(
         const TString& agentId,
         const TString& path) const;

cloud/blockstore/libs/storage/disk_registry/disk_registry_tx.h

@@ -305,6 +305,7 @@ struct TTxDiskRegistry
         const TRequestInfoPtr RequestInfo;
         const TVector<TString> Devices;
 
+        TVector<TString> SyncAllocatedDisks;
         TVector<TString> SyncDeallocatedDisks;
 
         explicit TCleanupDevices(
@@ -316,6 +317,7 @@ struct TTxDiskRegistry
 
         void Clear()
         {
+            SyncAllocatedDisks.clear();
             SyncDeallocatedDisks.clear();
         }
     };