Issue#45 Replace trx_sys_t::rw_trx_set with LF_HASH

trx_sys_t::rw_trx_set is implemented as std::set, which does a few quite
expensive operations under trx_sys_t::mutex protection: e.g. malloc/free
when adding/removing elements. Traversing b-tree is not that cheap either.

This has negative scalability impact, which is especially visible when running
oltp_update_index.lua benchmark on a ramdisk.

To reduce trx_sys_t::mutex contention std::set is replaced with LF_HASH. None
of LF_HASH operations require trx_sys_t::mutex (nor any other global mutex)
protection.

storage/innobase/clone/clone0copy.cc

@@ -327,14 +327,14 @@ int Clone_Snapshot::update_binlog_position() {
 }
 
 int Clone_Snapshot::wait_trx_end(THD *thd, trx_id_t trx_id) {
-  auto trx = trx_rw_is_active(trx_id, nullptr, false);
+  auto trx = trx_sys->rw_trx_hash.find(trx_id, false);
   if (trx == nullptr) {
     return (0);
   }
 
   auto wait_cond = [&](bool alert, bool &result) {
     /* Check if transaction is still active. */
-    auto trx = trx_rw_is_active(trx_id, nullptr, false);
+    auto trx = trx_sys->rw_trx_hash.find(trx_id, false);
     if (trx == nullptr) {
       result = false;
       return (0);

storage/innobase/handler/ha_innodb.cc

@@ -652,6 +652,7 @@ static PSI_mutex_info all_innodb_mutexes[] = {
 #endif /* UNIV_DEBUG */
     PSI_MUTEX_KEY(rw_lock_list_mutex, 0, 0, PSI_DOCUMENT_ME),
     PSI_MUTEX_KEY(rw_lock_mutex, 0, 0, PSI_DOCUMENT_ME),
+    PSI_MUTEX_KEY(rw_trx_hash_element_mutex, 0, 0, PSI_DOCUMENT_ME),
     PSI_MUTEX_KEY(srv_innodb_monitor_mutex, 0, 0, PSI_DOCUMENT_ME),
     PSI_MUTEX_KEY(srv_misc_tmpfile_mutex, 0, 0, PSI_DOCUMENT_ME),
     PSI_MUTEX_KEY(srv_monitor_file_mutex, 0, 0, PSI_DOCUMENT_ME),
@@ -19322,6 +19323,22 @@ static bool innodb_buffer_pool_size_validate(THD *thd,
   return true;
 }
 
+/**
+  Gets current trx.
+
+  This function may be called during InnoDB initialisation, when
+  innodb_hton_ptr->slot is not yet set to meaningful value.
+*/
+trx_t *current_trx() {
+  THD *thd = current_thd;
+  if (likely(thd != nullptr) && innodb_hton_ptr->slot != HA_SLOT_UNDEF) {
+    trx_t *&trx = thd_to_trx(thd);
+    return (trx);
+  } else {
+    return (nullptr);
+  }
+}
+
 /** Update the system variable innodb_buffer_pool_size using the "saved"
 value. This function is registered as a callback with MySQL.
 @param[in]	thd	thread handle

storage/innobase/include/row0vers.h

@@ -48,6 +48,7 @@ class ReadView;
 
 /** Finds out if an active transaction has inserted or modified a secondary
  index record.
+ @param[in,out] caller_trx  trx of current thread
  @param[in]   rec       record in a secondary index
  @param[in]   index     the secondary index
  @param[in]   offsets   rec_get_offsets(rec, index)
@@ -56,8 +57,8 @@ class ReadView;
  negatives. The caller must confirm all positive results by checking if the trx
  is still active.
 */
-trx_t *row_vers_impl_x_locked(const rec_t *rec, const dict_index_t *index,
-                              const ulint *offsets);
+trx_t *row_vers_impl_x_locked(trx_t *caller_trx, const rec_t *rec,
+                              const dict_index_t *index, const ulint *offsets);
 
 /** Finds out if we must preserve a delete marked earlier version of a clustered
  index record, because it is >= the purge view.
@@ -119,6 +120,7 @@ dberr_t row_vers_build_for_consistent_read(
 /** Constructs the last committed version of a clustered index record,
  which should be seen by a semi-consistent read. */
 void row_vers_build_for_semi_consistent_read(
+    trx_t* caller_trx,        /*!< in/out: trx of current thread */
     const rec_t *rec,         /*!< in: record in a clustered index; the
                               caller must have a latch on the page; this
                               latch locks the top of the stack of versions

storage/innobase/include/sync0sync.h

@@ -175,6 +175,7 @@ extern mysql_pfs_key_t clone_task_mutex_key;
 extern mysql_pfs_key_t clone_snapshot_mutex_key;
 extern mysql_pfs_key_t parallel_read_mutex_key;
 extern mysql_pfs_key_t dblwr_mutex_key;
+extern mysql_pfs_key_t rw_trx_hash_element_mutex_key;
 #endif /* UNIV_PFS_MUTEX */
 
 #ifdef UNIV_PFS_RWLOCK

storage/innobase/include/sync0types.h

@@ -275,6 +275,7 @@ enum latch_level_t {
   SYNC_THREADS,
   SYNC_TRX,
   SYNC_POOL,
+  SYNC_RW_TRX_HASH_ELEMENT,
   SYNC_POOL_MANAGER,
   SYNC_TRX_SYS,
   SYNC_LOCK_SYS_SHARDED,
@@ -466,6 +467,7 @@ enum latch_id_t {
   LATCH_ID_DBLR,
   LATCH_ID_REDO_LOG_ARCHIVE_ADMIN_MUTEX,
   LATCH_ID_REDO_LOG_ARCHIVE_QUEUE_MUTEX,
+  LATCH_ID_RW_TRX_HASH_ELEMENT,
   LATCH_ID_TEST_MUTEX,
   LATCH_ID_MAX = LATCH_ID_TEST_MUTEX
 };