Merge pull request #87 from skyzh/skyzh/zns-gc

add garbage collection for ZenFS

Author: Changlong Chen

db/db_impl.cc

@@ -68,6 +68,7 @@
 #include "table/block_based_table_factory.h"
 #include "table/merging_iterator.h"
 #include "table/two_level_iterator.h"
+#include "third-party/zenfs/fs/zbd_stat.h"
 #include "util/autovector.h"
 #include "util/build_version.h"
 #include "util/c_style_callback.h"
@@ -113,6 +114,8 @@
 
 namespace TERARKDB_NAMESPACE {
 
+std::vector<ZoneStat> GetStat(Env* env);
+
 const std::string kDefaultColumnFamilyName("default");
 const uint64_t kDumpStatsWaitMicroseconds = 10000;
 const std::string kPersistentStatsColumnFamilyName(
@@ -985,6 +988,163 @@ void DBImpl::ScheduleTtlGC() {
   log_buffer_debug.FlushBufferToLog();
 }
 
+#ifdef LIBZBD
+void DBImpl::ScheduleZNSGC() {
+  TEST_SYNC_POINT("DBImpl:ScheduleZNSGC");
+  uint64_t nowSeconds = env_->NowMicros() / 1000U / 1000U;
+  LogBuffer log_buffer_info(InfoLogLevel::INFO_LEVEL,
+                            immutable_db_options_.info_log.get());
+  LogBuffer log_buffer_debug(InfoLogLevel::DEBUG_LEVEL,
+                             immutable_db_options_.info_log.get());
+
+  chash_set<uint64_t> mark_for_gc;
+
+  if (initial_db_options_.zenfs_gc_ratio <= 0.0 ||
+      initial_db_options_.zenfs_gc_ratio >= 1.0) {
+    // GC is not enabled
+    return;
+  }
+
+  // Pick files for GC
+  auto stat = GetStat(env_);
+
+  uint64_t number;
+  FileType type;
+
+  // Merge db paths and column family paths together
+  chash_set<std::string> db_paths;
+
+  // Get column family paths
+  mutex_.Lock();
+  for (auto cfd : *versions_->GetColumnFamilySet()) {
+    for (const auto& path : cfd->ioptions()->db_paths) {
+      db_paths.emplace(path.path);
+    }
+  }
+  mutex_.Unlock();
+
+  // Get database paths
+  for (const auto& path : immutable_db_options_.db_paths) {
+    db_paths.emplace(path.path);
+  }
+
+  std::string strip_filename;
+
+  for (const auto& zone : stat) {
+    std::vector<uint64_t> sst_in_zone;
+    uint64_t written_data = zone.write_position - zone.start_position;
+    // zone is full
+    if (written_data == zone.total_capacity) {
+      uint64_t total_size = 0;
+      bool ignore_zone = false;
+      for (const auto& file : zone.files) {
+        strip_filename.clear();
+
+        for (const auto& path : db_paths) {
+          if (Slice(file.filename).starts_with(path)) {
+            strip_filename.assign(file.filename, path.length(),
+                                  file.filename.length() - path.length());
+            break;
+          }
+        }
+
+        if (strip_filename.empty()) {
+          // This file is not in DB folder.
+          ignore_zone = true;
+          break;
+        }
+
+        if (ParseFileName(strip_filename, &number, Slice(), &type)) {
+          // Is SST file, and is of current TerarkDB instance.
+          if (type == kTableFile) {
+            total_size += file.size_in_zone;
+            sst_in_zone.push_back(number);
+          } else {
+            // This zone contains file other than SSTs or files from other
+            // databases. We ignore the zone for now. When other files (like
+            // logs) have been deleted, we will come back and recycle this zone.
+            ignore_zone = true;
+            break;
+          }
+        } else {
+          // This file is not recognized by TerarkDB (or RocksDB). Even if we
+          // move the file, the zone may not be reset. Therefore, we simply
+          // ignore this zone.
+          ignore_zone = true;
+          break;
+        }
+      }
+
+      if (ignore_zone) {
+        continue;
+      }
+
+      // if data in zone <= (1 - ratio) * total_capacity, recycle the zone
+      if (total_size <=
+          (1.0 - initial_db_options_.zenfs_gc_ratio) * written_data) {
+        for (auto&& file_id : sst_in_zone) {
+          mark_for_gc.insert(file_id);
+        }
+      }
+    }
+  }
+
+  mutex_.Lock();
+  for (auto cfd : *versions_->GetColumnFamilySet()) {
+    uint64_t new_mark_count = 0;
+    uint64_t old_mark_count = 0;
+    uint64_t total_count = 0;
+    if (!cfd->initialized() || cfd->IsDropped()) {
+      continue;
+    }
+    VersionStorageInfo* vstorage = cfd->current()->storage_info();
+    // Level -1 contains SSTs inside lazy compaction SST index.
+    // By iterating level -1, we could collect that kind of garbage.
+    // But we still recommend using ZNS GC without lazy compaction
+    // enabled.
+    for (int l = -1; l < vstorage->num_non_empty_levels(); l++) {
+      for (auto meta : vstorage->LevelFiles(l)) {
+        if (meta->being_compacted) {
+          continue;
+        }
+        ++total_count;
+        old_mark_count += meta->marked_for_compaction;
+        TEST_SYNC_POINT("DBImpl:Exist-SST");
+        if (!meta->marked_for_compaction &&
+            mark_for_gc.count(meta->fd.GetNumber()) > 0) {
+          meta->marked_for_compaction = true;
+        }
+        if (meta->marked_for_compaction) {
+          new_mark_count++;
+          TEST_SYNC_POINT("DBImpl:ScheduleZNSGC-mark");
+        }
+      }
+    }
+    if (new_mark_count > old_mark_count) {
+      vstorage->ComputeCompactionScore(*cfd->ioptions(),
+                                       *cfd->GetLatestMutableCFOptions());
+      if (!cfd->queued_for_compaction()) {
+        AddToCompactionQueue(cfd);
+        unscheduled_compactions_++;
+      }
+    }
+    if (old_mark_count != 0 && new_mark_count != 0) {
+      ROCKS_LOG_BUFFER(&log_buffer_info,
+                       "[%s] ZNS GC: SSTs total marked = %" PRIu64
+                       ", new marked = %" PRIu64 ", file count: %" PRIu64,
+                       cfd->GetName().c_str(), old_mark_count, new_mark_count,
+                       total_count);
+    }
+  }
+  if (unscheduled_compactions_ > 0) {
+    MaybeScheduleFlushOrCompaction();
+  }
+  mutex_.Unlock();
+  log_buffer_info.FlushBufferToLog();
+  log_buffer_debug.FlushBufferToLog();
+}
+#endif
+
 void DBImpl::DumpStats() {
   TEST_SYNC_POINT("DBImpl::DumpStats:1");
 #ifndef ROCKSDB_LITE

db/db_impl.h

@@ -410,11 +410,11 @@ class DBImpl : public DB {
   // Implemented in db_impl_debug.cc
 
   // Compact any files in the named level that overlap [*begin, *end]
-  Status TEST_CompactRange(int level, const Slice* begin, const Slice* end,
-                           ColumnFamilyHandle* column_family = nullptr,
-                           SeparationType separation_type =
-                               kCompactionTransToSeparate,
-                           bool disallow_trivial_move = false);
+  Status TEST_CompactRange(
+      int level, const Slice* begin, const Slice* end,
+      ColumnFamilyHandle* column_family = nullptr,
+      SeparationType separation_type = kCompactionTransToSeparate,
+      bool disallow_trivial_move = false);
 
   void TEST_SwitchWAL();
 
@@ -806,6 +806,11 @@ class DBImpl : public DB {
 
   void ScheduleTtlGC();
 
+#ifdef LIBZBD
+  // schedule GC by polling ZNS zone status
+  void ScheduleZNSGC();
+#endif
+
  protected:
   Env* const env_;
   const std::string dbname_;

db/db_impl_files.cc

@@ -532,6 +532,22 @@ void DBImpl::PurgeObsoleteFiles(JobContext& state, bool schedule_only) {
               to_delete;
     }
 
+    // TODO: Workaround for ZNS and Windows.
+    // In TerarkDB, when a WAL is not needed, it is first deleted, and then
+    // closed. This means that the underlying FS must support deferred delete.
+    // In this case, we delete the writer before issuing delete to FS.
+    if (type == kLogFile) {
+      auto it =
+          std::find_if(state.logs_to_free.begin(), state.logs_to_free.end(),
+                       [number](log::Writer* writer) {
+                         return writer->get_log_number() == number;
+                       });
+      if (it != state.logs_to_free.end()) {
+        delete *it;
+        *it = nullptr;
+      }
+    }
+
 #ifndef ROCKSDB_LITE
     if (type == kLogFile && (immutable_db_options_.wal_ttl_seconds > 0 ||
                              immutable_db_options_.wal_size_limit_mb > 0)) {

db/periodic_work_scheduler.cc

@@ -45,6 +45,13 @@ void PeriodicWorkScheduler::Register(DBImpl* dbi,
              initial_delay.fetch_add(1) % kDefaultScheduleGCTTLPeriodSec *
                  kMicrosInSecond,
              kDefaultScheduleGCTTLPeriodSec * kMicrosInSecond);
+#ifdef LIBZBD
+  timer->Add([dbi]() { dbi->ScheduleZNSGC(); },
+             GetTaskName(dbi, "schedule_gc_zns"),
+             initial_delay.fetch_add(1) % kDefaultScheduleZNSTTLPeriodSec *
+                 kMicrosInSecond,
+             kDefaultScheduleZNSTTLPeriodSec * kMicrosInSecond);
+#endif
 }
 
 void PeriodicWorkScheduler::Unregister(DBImpl* dbi) {
@@ -53,6 +60,9 @@ void PeriodicWorkScheduler::Unregister(DBImpl* dbi) {
   timer->Cancel(GetTaskName(dbi, "pst_st"));
   timer->Cancel(GetTaskName(dbi, "flush_info_log"));
   timer->Cancel(GetTaskName(dbi, "schedule_gc_ttl"));
+#ifdef LIBZBD
+  timer->Cancel(GetTaskName(dbi, "schedule_gc_zns"));
+#endif
   if (!timer->HasPendingTask()) {
     timer->Shutdown();
   }

db/periodic_work_scheduler.h

@@ -41,6 +41,7 @@ class PeriodicWorkScheduler {
   // log.
   static const uint64_t kDefaultFlushInfoLogPeriodSec = 10;
   static const uint64_t kDefaultScheduleGCTTLPeriodSec = 10;
+  static const uint64_t kDefaultScheduleZNSTTLPeriodSec = 10;
 
  protected:
   std::unique_ptr<Timer> timer;

env/env_zenfs.cc

@@ -4,6 +4,7 @@
 
 #ifdef LIBZBD
 #include "third-party/zenfs/fs/fs_zenfs.h"
+#include "third-party/zenfs/fs/zbd_stat.h"
 #include "third-party/zenfs/fs/zbd_zenfs.h"
 
 namespace TERARKDB_NAMESPACE {
@@ -166,9 +167,7 @@ class ZenfsDirectory : public Directory {
   explicit ZenfsDirectory(std::unique_ptr<FSDirectory>&& target)
       : target_(std::move(target)) {}
 
-  Status Fsync() override {
-    return target_->Fsync(IOOptions(), nullptr);
-  }
+  Status Fsync() override { return target_->Fsync(IOOptions(), nullptr); }
   size_t GetUniqueId(char* id, size_t max_size) const override {
     return target_->GetUniqueId(id, max_size);
   }
@@ -469,14 +468,20 @@ class ZenfsEnv : public EnvWrapper {
     target_->SanitizeEnvOptions(env_opts);
   }
 
-  Status GetZbdDiskSpaceInfo(uint64_t &total_size, uint64_t &avail_size, uint64_t &used_size) {
+  Status GetZbdDiskSpaceInfo(uint64_t& total_size, uint64_t& avail_size,
+                             uint64_t& used_size) {
     auto zbd = dynamic_cast<ZenFS*>(fs_)->GetZonedBlockDevice();
     used_size = zbd->GetUsedSpace() + zbd->GetReclaimableSpace();
     avail_size = zbd->GetFreeSpace();
     total_size = used_size + avail_size;
     return Status::OK();
   }
 
+  std::vector<ZoneStat> GetStat() {
+    auto zen_fs = dynamic_cast<ZenFS*>(fs_);
+    return zen_fs->GetStat();
+  }
+
  private:
   Env* target_;
   FileSystem* fs_;
@@ -490,8 +495,16 @@ Status NewZenfsEnv(Env** zenfs_env, const std::string& zdb_path) {
   return s;
 }
 
-Status GetZbdDiskSpaceInfo(Env* env, uint64_t &total_size, uint64_t &avail_size, uint64_t &used_size) {
-  return dynamic_cast<ZenfsEnv*>(env)->GetZbdDiskSpaceInfo(total_size, avail_size, used_size);
+Status GetZbdDiskSpaceInfo(Env* env, uint64_t& total_size, uint64_t& avail_size,
+                           uint64_t& used_size) {
+  return dynamic_cast<ZenfsEnv*>(env)->GetZbdDiskSpaceInfo(
+      total_size, avail_size, used_size);
+}
+
+std::vector<ZoneStat> GetStat(Env* env) {
+  auto zen_env = dynamic_cast<ZenfsEnv*>(env);
+  if (!zen_env) return {};
+  return zen_env->GetStat();
 }
 
 }  // namespace TERARKDB_NAMESPACE
@@ -505,11 +518,13 @@ Status NewZenfsEnv(Env** zenfs_env, const std::string& zdb_path) {
   return Status::NotSupported("ZenFSEnv is not implemented.");
 }
 
-Status GetZbdDiskSpaceInfo(Env* env, uint64_t &total_size, uint64_t &avail_size, uint64_t &used_size) {
+Status GetZbdDiskSpaceInfo(Env* env, uint64_t& total_size, uint64_t& avail_size,
+                           uint64_t& used_size) {
   return Status::NotSupported("GetZbdDiskSpaceInfo is not implemented.");
 }
 
+std::vector<ZoneStat> GetStat(Env* env) { return {}; }
+
 }  // namespace TERARKDB_NAMESPACE
 
 #endif
-