curvefs/metaserver: fix trash bugs

curvefs/src/metaserver/copyset/copyset_node.cpp

@@ -170,7 +170,7 @@ bool CopysetNode::Start() {
         LOG(ERROR) << "Fail to init raft node, copyset: " << name_;
         return false;
     }
-
+    metaStore_->LoadDeletedInodes();
     LOG(INFO) << "Run copyset success, copyset: " << name_;
     return true;
 }

curvefs/src/metaserver/inode_manager.cpp

@@ -391,11 +391,6 @@ MetaStatusCode InodeManager::UpdateInode(const UpdateInodeRequest& request,
         }
     }
 
-    if (needAddTrash) {
-        trash_->Add(old.inodeid(), old.dtime());
-        --(*type2InodeNum_)[old.type()];
-    }
-
     const S3ChunkInfoMap &map2add = request.s3chunkinfoadd();
     const S3ChunkInfoList *list2add;
     VLOG(9) << "UpdateInode inode " << old.inodeid() << " map2add size "
@@ -446,6 +441,13 @@ MetaStatusCode InodeManager::UpdateInode(const UpdateInodeRequest& request,
             return MetaStatusCode::STORAGE_INTERNAL_ERROR;
         }
     }
+
+    if (needAddTrash) {
+        inodeStorage_->UpdateDeletingKey(old, logIndex);
+        trash_->Add(old.inodeid(), old.dtime(), false);
+        --(*type2InodeNum_)[old.type()];
+    }
+
     VLOG(9) << "UpdateInode success, " << request.ShortDebugString();
     return MetaStatusCode::OK;
 }

curvefs/src/metaserver/inode_storage.cpp

@@ -67,6 +67,7 @@ InodeStorage::InodeStorage(std::shared_ptr<KVStorage> kvStorage,
                            uint64_t nInode)
     : kvStorage_(std::move(kvStorage)),
       table4Inode_(nameGenerator->GetInodeTableName()),
+      table4DelInode_(nameGenerator->GetDelInodeTableName()),
       table4S3ChunkInfo_(nameGenerator->GetS3ChunkInfoTableName()),
       table4VolumeExtent_(nameGenerator->GetVolumeExtentTableName()),
       table4InodeAuxInfo_(nameGenerator->GetInodeAuxInfoTableName()),
@@ -185,9 +186,74 @@ MetaStatusCode InodeStorage::Insert(const Inode& inode, int64_t logIndex) {
     return MetaStatusCode::STORAGE_INTERNAL_ERROR;
 }
 
+MetaStatusCode InodeStorage::UpdateDeletingKey(
+  const Inode& inode, int64_t logIndex) {
+    WriteLockGuard lg(rwLock_);
+    Key4Inode key(inode.fsid(), inode.inodeid());
+    std::string skey = conv_.SerializeToString(key);
+    VLOG(9) << "update deleting key, " << inode.inodeid();
+    const char* step = "Begin transaction";
+    std::shared_ptr<storage::StorageTransaction> txn;
+    txn = kvStorage_->BeginTransaction();
+    if (txn == nullptr) {
+        LOG(ERROR) << "Begin transaction failed";
+        return MetaStatusCode::STORAGE_INTERNAL_ERROR;
+    }
+    auto rc = txn->HSet(table4DelInode_, skey , inode);
+    step = "insert inode ";
+     if (rc.ok()) {
+        rc = SetAppliedIndex(txn.get(), logIndex);
+        step = "Insert applied index to transaction";
+    }
+    if (rc.ok()) {
+        rc = txn->Commit();
+        step = "commit";
+    }
+    if (rc.ok()) {
+        VLOG(9) << "set deleting key ok";
+        return MetaStatusCode::OK;
+    }
+    LOG(ERROR) << step << "failed, status = " << rc.ToString();
+    if (txn != nullptr && !txn->Rollback().ok()) {
+        LOG(ERROR) << "Rollback delete inode transaction failed, status = "
+                   << rc.ToString();
+    }
+    return MetaStatusCode::STORAGE_INTERNAL_ERROR;
+}
+
+MetaStatusCode InodeStorage::ClearDelKey(const Key4Inode& key) {
+    WriteLockGuard lg(rwLock_);
+    std::string skey = conv_.SerializeToString(key);
+    VLOG(9) << "clear deleting key start, " << skey;
+    std::shared_ptr<storage::StorageTransaction> txn = nullptr;
+    const char* step = "Begin transaction";
+    txn = kvStorage_->BeginTransaction();
+    if (txn == nullptr) {
+        LOG(ERROR) << "Begin transaction failed";
+        return MetaStatusCode::STORAGE_INTERNAL_ERROR;
+    }
+    step = "Delete inode from transaction";
+    auto s = txn->HDel(table4DelInode_, skey);
+    if (s.ok()) {
+        step = "Delete inode";
+        s = txn->Commit();
+    }
+    if (s.ok()) {
+        VLOG(9) << "clear deleting key ok, " << skey;
+        return MetaStatusCode::OK;
+    }
+    LOG(ERROR) << step << " failed, status = " << s.ToString();
+    if (!txn->Rollback().ok()) {
+        LOG(ERROR) << "Rollback delete inode transaction failed, status = "
+                   << s.ToString();
+    }
+    return MetaStatusCode::STORAGE_INTERNAL_ERROR;
+}
+
 MetaStatusCode InodeStorage::Get(const Key4Inode& key, Inode* inode) {
     ReadLockGuard lg(rwLock_);
     std::string skey = conv_.SerializeToString(key);
+
     Status s = kvStorage_->HGet(table4Inode_, skey, inode);
     if (s.ok()) {
         return MetaStatusCode::OK;
@@ -471,7 +537,6 @@ MetaStatusCode InodeStorage::Clear() {
     // because if we fail stop, we will replay
     // raft logs and clear it again
     WriteLockGuard lg(rwLock_);
-
     Status s = kvStorage_->HClear(table4Inode_);
     if (!s.ok()) {
         LOG(ERROR) << "InodeStorage clear inode table failed, status = "
@@ -492,7 +557,6 @@ MetaStatusCode InodeStorage::Clear() {
             << s.ToString();
         return MetaStatusCode::STORAGE_INTERNAL_ERROR;
     }
-
     s = kvStorage_->HClear(table4InodeAuxInfo_);
     if (!s.ok()) {
         LOG(ERROR)

curvefs/src/metaserver/inode_storage.h

@@ -76,6 +76,16 @@ class InodeStorage {
      */
     MetaStatusCode Insert(const Inode& inode, int64_t logIndex);
 
+    /**
+     * @brief update deleting inode key in storage
+     * @param[in] inode: the inode want to update
+     * @param[in] logIndex: the index of raft log
+     * @return
+     */
+    MetaStatusCode UpdateDeletingKey(const Inode& inode, int64_t logIndex);
+
+    MetaStatusCode ClearDelKey(const Key4Inode& key);
+
     /**
      * @brief get inode from storage
      * @param[in] key: the key of inode want to get
@@ -237,6 +247,7 @@ class InodeStorage {
     RWLock rwLock_;
     std::shared_ptr<KVStorage> kvStorage_;
     std::string table4Inode_;
+    std::string table4DelInode_;
     std::string table4S3ChunkInfo_;
     std::string table4VolumeExtent_;
     std::string table4InodeAuxInfo_;

curvefs/src/metaserver/metastore.cpp

@@ -25,6 +25,7 @@
 #include <glog/logging.h>
 #include <sys/types.h>
 
+#include <bitset>
 #include <memory>
 #include <thread>  // NOLINT
 #include <unordered_map>
@@ -60,6 +61,7 @@ using KVStorage = ::curvefs::metaserver::storage::KVStorage;
 using Key4S3ChunkInfoList = ::curvefs::metaserver::storage::Key4S3ChunkInfoList;
 
 using ::curvefs::metaserver::storage::MemoryStorage;
+using ::curvefs::metaserver::storage::NameGenerator;
 using ::curvefs::metaserver::storage::RocksDBStorage;
 using ::curvefs::metaserver::storage::StorageOptions;
 
@@ -147,6 +149,7 @@ bool MetaStoreImpl::Load(const std::string &pathname) {
     }
 
     startCompacts();
+
     return true;
 }
 
@@ -866,5 +869,27 @@ bool MetaStoreImpl::InitStorage() {
     return kvStorage_->Open();
 }
 
+void MetaStoreImpl::LoadDeletedInodes() {
+    VLOG(3) << "LoadDeletedInodes start";
+    std::list<std::string> items;
+    kvStorage_->LoadDeletedInodes(&items);
+    VLOG(3) << "build trash items size: " << items.size();
+    std::vector<std::string> names;
+    for (auto iter : items) {
+        curve::common::SplitString(iter , ":", &names);
+        char* tmp = const_cast<char*>(names[2].c_str());
+        uint32_t partitionId = *reinterpret_cast<uint32_t*>(tmp);
+        VLOG(3) << "partitionId is: " << partitionId;
+        std::shared_ptr<Trash> trash =
+          TrashManager::GetInstance().GetTrash(partitionId);
+        if (nullptr == trash) {
+            LOG(INFO) << "trash is null: " << partitionId;
+            continue;
+        }
+        trash->Add(std::stoull(names[names.size() - 1 ]), 0, true);
+    }
+    VLOG(3) << "LoadDeletedInodes over.";
+}
+
 }  // namespace metaserver
 }  // namespace curvefs

curvefs/src/metaserver/metastore.h

@@ -117,6 +117,7 @@ class MetaStore {
     virtual bool SaveData(const std::string& dir,
                           std::vector<std::string>* files) = 0;
     virtual bool Clear() = 0;
+    virtual void LoadDeletedInodes() {}
     virtual bool Destroy() = 0;
     virtual MetaStatusCode CreatePartition(
         const CreatePartitionRequest* request,
@@ -223,6 +224,7 @@ class MetaStoreImpl : public MetaStore {
                   std::vector<std::string>* files) override;
     bool Clear() override;
     bool Destroy() override;
+    void LoadDeletedInodes() override;
 
     MetaStatusCode CreatePartition(const CreatePartitionRequest* request,
                                    CreatePartitionResponse* response,

curvefs/src/metaserver/storage/converter.cpp

@@ -23,6 +23,7 @@
 #include <inttypes.h>
 #include <glog/logging.h>
 
+#include <bitset>
 #include <cstring>
 #include <string>
 #include <vector>
@@ -54,6 +55,7 @@ static bool CompareType(const std::string& str, KEY_TYPE keyType) {
 
 NameGenerator::NameGenerator(uint32_t partitionId)
     : tableName4Inode_(Format(kTypeInode, partitionId)),
+      tableName4DelInode_(FormatDel(kTypeInode, partitionId)),
       tableName4DeallocatableIndoe_(
           Format(kTypeDeallocatableInode, partitionId)),
       tableName4DeallocatableBlockGroup_(
@@ -71,6 +73,10 @@ std::string NameGenerator::GetInodeTableName() const {
     return tableName4Inode_;
 }
 
+std::string NameGenerator::GetDelInodeTableName() const {
+    return tableName4DelInode_;
+}
+
 std::string NameGenerator::GetDeallocatableInodeTableName() const {
     return tableName4DeallocatableIndoe_;
 }
@@ -123,6 +129,13 @@ std::string NameGenerator::Format(KEY_TYPE type, uint32_t partitionId) {
     return absl::StrCat(type, kDelimiter, absl::string_view(buf, sizeof(buf)));
 }
 
+std::string NameGenerator::FormatDel(KEY_TYPE type, uint32_t partitionId) {
+    char buf[sizeof(partitionId)];
+    std::memcpy(buf, reinterpret_cast<char*>(&partitionId), sizeof(buf));
+    return absl::StrCat(DELETEPREFIXINKV, type,
+      kDelimiter, absl::string_view(buf, sizeof(buf)));
+}
+
 Key4Inode::Key4Inode() : fsId(0), inodeId(0) {}
 
 Key4Inode::Key4Inode(uint32_t fsId, uint64_t inodeId)
@@ -146,6 +159,11 @@ bool Key4Inode::ParseFromString(const std::string& value) {
            StringToUl(items[1], &fsId) && StringToUll(items[2], &inodeId);
 }
 
+void Key4Inode::GetPrefix(bool ordered, std::string* value) {
+    *value = absl::StrCat(std::to_string(
+      ordered), ":" , DELETEPREFIXINKV);
+}
+
 std::string Prefix4AllInode::SerializeToString() const {
     return absl::StrCat(keyType_, ":");
 }

curvefs/src/metaserver/storage/converter.h

@@ -34,9 +34,10 @@ namespace curvefs {
 namespace metaserver {
 
 class MetaStoreFStream;
-
 namespace storage {
 
+#define DELETEPREFIXINKV "deleting:"
+
 enum KEY_TYPE : unsigned char {
     kTypeInode = 1,
     kTypeS3ChunkInfo = 2,
@@ -61,6 +62,8 @@ class NameGenerator {
 
     std::string GetInodeTableName() const;
 
+    std::string GetDelInodeTableName() const;
+
     std::string GetDeallocatableInodeTableName() const;
 
     std::string GetS3ChunkInfoTableName() const;
@@ -87,9 +90,12 @@ class NameGenerator {
 
  private:
     std::string Format(KEY_TYPE type, uint32_t partitionId);
+    std::string FormatDel(KEY_TYPE type, uint32_t partitionId);
+
 
  private:
     std::string tableName4Inode_;
+    std::string tableName4DelInode_;
     std::string tableName4DeallocatableIndoe_;
     std::string tableName4DeallocatableBlockGroup_;
     std::string tableName4S3ChunkInfo_;
@@ -108,6 +114,7 @@ class StorageKey {
 
     virtual std::string SerializeToString() const = 0;
     virtual bool ParseFromString(const std::string& value) = 0;
+    virtual void GetPrefix(bool ordered, std::string* value) {}
 };
 
 /* rules for key serialization:
@@ -142,6 +149,8 @@ class Key4Inode : public StorageKey {
 
     bool ParseFromString(const std::string& value) override;
 
+    void GetPrefix(bool ordered, std::string* value) override;
+
  public:
     static const KEY_TYPE keyType_ = kTypeInode;
 

curvefs/src/metaserver/storage/rocksdb_storage.cpp

@@ -26,6 +26,7 @@
 #include <iostream>
 #include <unordered_map>
 
+#include "src/common/string_util.h"
 #include "src/common/timeutility.h"
 #include "curvefs/src/metaserver/storage/utils.h"
 #include "curvefs/src/metaserver/storage/storage.h"
@@ -209,6 +210,7 @@ Status RocksDBStorage::Get(const std::string& name,
     ROCKSDB_NAMESPACE::Status s;
     std::string svalue;
     std::string ikey = ToInternalKey(name, key, ordered);
+    VLOG(9) << "Get  key: " << ikey << ", " << options_.dataDir;
     auto handle = GetColumnFamilyHandle(ordered);
     {
         RocksDBPerfGuard guard(OP_GET);
@@ -218,6 +220,7 @@ Status RocksDBStorage::Get(const std::string& name,
     if (s.ok() && !value->ParseFromString(svalue)) {
         return Status::ParsedFailed();
     }
+
     return ToStorageStatus(s);
 }
 
@@ -234,6 +237,7 @@ Status RocksDBStorage::Set(const std::string& name,
 
     auto handle = GetColumnFamilyHandle(ordered);
     std::string ikey = ToInternalKey(name, key, ordered);
+    VLOG(9) << "set  key: " << ikey << ", " << options_.dataDir;
     RocksDBPerfGuard guard(OP_PUT);
     ROCKSDB_NAMESPACE::Status s = InTransaction_ ?
         txn_->Put(handle, ikey, svalue) :
@@ -250,6 +254,7 @@ Status RocksDBStorage::Del(const std::string& name,
 
     std::string ikey = ToInternalKey(name, key, ordered);
     auto handle = GetColumnFamilyHandle(ordered);
+    VLOG(9) << "del key: " << ikey << ", " << options_.dataDir;
     RocksDBPerfGuard guard(OP_DELETE);
     ROCKSDB_NAMESPACE::Status s = InTransaction_ ?
         txn_->Delete(handle, ikey) :
@@ -505,6 +510,27 @@ bool RocksDBStorage::Recover(const std::string& dir) {
     return true;
 }
 
+void  RocksDBStorage::LoadDeletedInodes(std::list<std::string>* item) {
+    std::string sprefix;
+    Key4Inode key;
+    key.GetPrefix(0, &sprefix);
+
+    LOG(INFO) << "LoadAll storage from: "
+      << options_.dataDir << ", " << sprefix;
+
+    int counts = 0;
+    rocksdb::Iterator* it = db_->NewIterator(rocksdb::ReadOptions());
+    for (it->Seek(sprefix); it->Valid() &&
+      it->key().starts_with(sprefix); it->Next()) {
+        std::string key = it->key().ToString();
+        item->push_back(key);
+        VLOG(3) << "load key: " << key << ", " << key.size();
+        counts++;
+     }
+    LOG(INFO) << "LoadAll storage from. size: " << item->size()
+              << ", " << counts << ", " << options_.dataDir;
+}
+
 }  // namespace storage
 }  // namespace metaserver
 }  // namespace curvefs