#141 redo table open and close locking to avoid table opening pileup

storage/tokudb/ha_tokudb.cc

@@ -205,14 +205,39 @@ exit:
     return error;
 }
 
-/** @brief
-    Simple lock controls. The "share" it creates is a structure we will
-    pass to each tokudb handler. Do you have to have one of these? Well, you have
-    pieces that are used for locking, and they are needed to function.
-
-   MUST have tokudb_mutex locked on input
+static void free_key_and_col_info (KEY_AND_COL_INFO* kc_info) {
+    for (uint i = 0; i < MAX_KEY+1; i++) {
+        bitmap_free(&kc_info->key_filters[i]);
+    }
     
-*/
+    for (uint i = 0; i < MAX_KEY+1; i++) {
+        tokudb_my_free(kc_info->cp_info[i]);
+        kc_info->cp_info[i] = NULL; // 3144
+    }
+    
+    tokudb_my_free(kc_info->field_lengths);
+    tokudb_my_free(kc_info->length_bytes);
+    tokudb_my_free(kc_info->blob_fields);
+}
+
+void TOKUDB_SHARE::init(void) {
+    use_count = 0;
+    thr_lock_init(&lock);
+    tokudb_pthread_mutex_init(&mutex, MY_MUTEX_INIT_FAST);
+    my_rwlock_init(&num_DBs_lock, 0);
+    tokudb_pthread_cond_init(&m_openclose_cond, NULL);
+    m_state = CLOSED;
+}
+
+void TOKUDB_SHARE::destroy(void) {
+    assert(m_state == CLOSED);
+    thr_lock_delete(&lock);
+    tokudb_pthread_mutex_destroy(&mutex);
+    rwlock_destroy(&num_DBs_lock);
+    tokudb_pthread_cond_destroy(&m_openclose_cond);
+}
+
+// MUST have tokudb_mutex locked on input    
 static TOKUDB_SHARE *get_share(const char *table_name, TABLE_SHARE* table_share) {
     TOKUDB_SHARE *share = NULL;
     int error = 0;
@@ -234,7 +259,8 @@ static TOKUDB_SHARE *get_share(const char *table_name, TABLE_SHARE* table_share)
             );
         assert(share);
 
-        share->use_count = 0;
+        share->init();
+
         share->table_name_length = length;
         share->table_name = tmp_name;
         strmov(share->table_name, table_name);
@@ -244,54 +270,30 @@ static TOKUDB_SHARE *get_share(const char *table_name, TABLE_SHARE* table_share)
             goto exit;
         }
 
-        memset((void *) share->key_file, 0, sizeof(share->key_file));
-
         error = my_hash_insert(&tokudb_open_tables, (uchar *) share);
         if (error) {
+            free_key_and_col_info(&share->kc_info);
             goto exit;
         }
-        thr_lock_init(&share->lock);
-        pthread_mutex_init(&share->mutex, MY_MUTEX_INIT_FAST);
-        my_rwlock_init(&share->num_DBs_lock, 0);
     }
 
 exit:
     if (error) {
-        pthread_mutex_destroy(&share->mutex);
+        share->destroy();
         tokudb_my_free((uchar *) share);
         share = NULL;
     }
     return share;
 }
 
-
-static void free_key_and_col_info (KEY_AND_COL_INFO* kc_info) {
-    for (uint i = 0; i < MAX_KEY+1; i++) {
-        bitmap_free(&kc_info->key_filters[i]);
-    }
-    
-    for (uint i = 0; i < MAX_KEY+1; i++) {
-        tokudb_my_free(kc_info->cp_info[i]);
-        kc_info->cp_info[i] = NULL; // 3144
-    }
-    
-    tokudb_my_free(kc_info->field_lengths);
-    tokudb_my_free(kc_info->length_bytes);
-    tokudb_my_free(kc_info->blob_fields);
-}
-
-//
-// MUST have tokudb_mutex locked on input
-// bool mutex_is_locked specifies if share->mutex is locked
-//
-static int free_share(TOKUDB_SHARE * share, bool mutex_is_locked) {
+static int free_share(TOKUDB_SHARE * share) {
     int error, result = 0;
 
-    if (mutex_is_locked) {
-        pthread_mutex_unlock(&share->mutex);
-    }
+    tokudb_pthread_mutex_lock(&share->mutex);
+    DBUG_PRINT("info", ("share->use_count %u", share->use_count));
     if (!--share->use_count) {
-        DBUG_PRINT("info", ("share->use_count %u", share->use_count));
+        share->m_state = TOKUDB_SHARE::CLOSING;
+        tokudb_pthread_mutex_unlock(&share->mutex);
 
         //
         // number of open DB's may not be equal to number of keys we have because add_index
@@ -316,13 +318,25 @@ static int free_share(TOKUDB_SHARE * share, bool mutex_is_locked) {
 
         error = tokudb::close_status(&share->status_block);
         assert(error == 0);
-        
-        my_hash_delete(&tokudb_open_tables, (uchar *) share);
-        thr_lock_delete(&share->lock);
-        pthread_mutex_destroy(&share->mutex);
-        rwlock_destroy(&share->num_DBs_lock);
 
-        tokudb_my_free((uchar *) share);
+        tokudb_pthread_mutex_lock(&tokudb_mutex);
+        tokudb_pthread_mutex_lock(&share->mutex);
+        share->m_state = TOKUDB_SHARE::CLOSED;
+        if (share->use_count > 0) {
+            tokudb_pthread_cond_broadcast(&share->m_openclose_cond);
+            tokudb_pthread_mutex_unlock(&share->mutex);
+            tokudb_pthread_mutex_unlock(&tokudb_mutex);
+        } else {
+            my_hash_delete(&tokudb_open_tables, (uchar *) share);
+            
+            tokudb_pthread_mutex_unlock(&share->mutex);
+            tokudb_pthread_mutex_unlock(&tokudb_mutex);
+
+            share->destroy();
+            tokudb_my_free((uchar *) share);
+        }
+    } else {
+        tokudb_pthread_mutex_unlock(&share->mutex);
     }
 
     return result;
@@ -588,7 +602,7 @@ smart_dbt_callback_ir_rowread(DBT const *key, DBT  const *row, void *context) {
 // Returns:
 //      The value of the auto increment column in record
 //
-ulonglong retrieve_auto_increment(uint16 type, uint32 offset,const uchar *record)
+static ulonglong retrieve_auto_increment(uint16 type, uint32 offset,const uchar *record)
 {
     const uchar *key;     /* Key */
     ulonglong   unsigned_autoinc = 0;  /* Unsigned auto-increment */
@@ -1710,8 +1724,6 @@ exit:
     return error;
 }
 
-
-
 //
 // Creates and opens a handle to a table which already exists in a tokudb
 // database.
@@ -1788,36 +1800,49 @@ int ha_tokudb::open(const char *name, int mode, uint test_if_locked) {
         goto exit;
     }
 
-    /* Init shared structure */
-    pthread_mutex_lock(&tokudb_mutex);
+    // lookup or create share
+    tokudb_pthread_mutex_lock(&tokudb_mutex);
     share = get_share(name, table_share);
     assert(share);
 
     thr_lock_data_init(&share->lock, &lock, NULL);
 
-    /* Fill in shared structure, if needed */
-    pthread_mutex_lock(&share->mutex);
-    if (!share->use_count++) {
-        ret_val = initialize_share(
-            name,
-            mode
-            );
-        if (ret_val) {
-            free_share(share, true);
-            pthread_mutex_unlock(&tokudb_mutex);
-            goto exit;
+    tokudb_pthread_mutex_lock(&share->mutex);
+    tokudb_pthread_mutex_unlock(&tokudb_mutex);
+    share->use_count++;
+    while (share->m_state == TOKUDB_SHARE::OPENING || share->m_state == TOKUDB_SHARE::CLOSING) {
+        tokudb_pthread_cond_wait(&share->m_openclose_cond, &share->mutex);
+    }
+    if (share->m_state == TOKUDB_SHARE::CLOSED) {
+        share->m_state = TOKUDB_SHARE::OPENING;
+        tokudb_pthread_mutex_unlock(&share->mutex);
+
+        ret_val = initialize_share(name, mode);
+
+        tokudb_pthread_mutex_lock(&share->mutex);
+        if (ret_val == 0) {
+            share->m_state = TOKUDB_SHARE::OPENED;
+        } else {
+            share->m_state = TOKUDB_SHARE::ERROR;
+            share->m_error = ret_val;
         }
+        tokudb_pthread_cond_broadcast(&share->m_openclose_cond);
+    }
+    if (share->m_state == TOKUDB_SHARE::ERROR) {
+        ret_val = share->m_error;
+        tokudb_pthread_mutex_unlock(&share->mutex);
+        free_share(share);
+        goto exit;
+    } else {
+        assert(share->m_state == TOKUDB_SHARE::OPENED);
+        tokudb_pthread_mutex_unlock(&share->mutex);
     }
-    pthread_mutex_unlock(&share->mutex);
-    pthread_mutex_unlock(&tokudb_mutex);
 
     ref_length = share->ref_length;     // If second open
     
     if (tokudb_debug & TOKUDB_DEBUG_OPEN) {
-        pthread_mutex_lock(&share->mutex);
         TOKUDB_TRACE("tokudbopen:%p:share=%p:file=%p:table=%p:table->s=%p:%d\n", 
                      this, share, share->file, table, table->s, share->use_count);
-        pthread_mutex_unlock(&share->mutex);
     }
 
     key_read = false;
@@ -2128,9 +2153,7 @@ int ha_tokudb::__close() {
     rec_update_buff = NULL;
     alloc_ptr = NULL;
     ha_tokudb::reset();
-    pthread_mutex_lock(&tokudb_mutex);
-    int retval = free_share(share, false);
-    pthread_mutex_unlock(&tokudb_mutex);
+    int retval = free_share(share);
     TOKUDB_DBUG_RETURN(retval);
 }
 
@@ -2926,7 +2949,7 @@ DBT *ha_tokudb::pack_ext_key(
 //
 void ha_tokudb::init_hidden_prim_key_info() {
     TOKUDB_DBUG_ENTER("ha_tokudb::init_prim_key_info");
-    pthread_mutex_lock(&share->mutex);
+    tokudb_pthread_mutex_lock(&share->mutex);
     if (!(share->status & STATUS_PRIMARY_KEY_INIT)) {
         int error = 0;
         THD* thd = ha_thd();
@@ -2966,7 +2989,7 @@ void ha_tokudb::init_hidden_prim_key_info() {
         }
         share->status |= STATUS_PRIMARY_KEY_INIT;
     }
-    pthread_mutex_unlock(&share->mutex);
+    tokudb_pthread_mutex_unlock(&share->mutex);
     DBUG_VOID_RETURN;
 }
 
@@ -3246,9 +3269,9 @@ void ha_tokudb::start_bulk_insert(ha_rows rows) {
             }
         }
     exit_try_table_lock:
-        pthread_mutex_lock(&share->mutex);
+        tokudb_pthread_mutex_lock(&share->mutex);
         share->try_table_lock = false;
-        pthread_mutex_unlock(&share->mutex);
+        tokudb_pthread_mutex_unlock(&share->mutex);
     }
     DBUG_VOID_RETURN;
 }
@@ -3265,9 +3288,9 @@ int ha_tokudb::end_bulk_insert(bool abort) {
     tokudb_trx_data* trx = (tokudb_trx_data *) thd_data_get(thd, tokudb_hton->slot);
     bool using_loader = (loader != NULL);
     if (ai_metadata_update_required) {
-        pthread_mutex_lock(&share->mutex);
+        tokudb_pthread_mutex_lock(&share->mutex);
         error = update_max_auto_inc(share->status_block, share->last_auto_increment);
-        pthread_mutex_unlock(&share->mutex);
+        tokudb_pthread_mutex_unlock(&share->mutex);
         if (error) { goto cleanup; }
     }
     delay_updating_ai_metadata = false;
@@ -3891,7 +3914,7 @@ int ha_tokudb::write_row(uchar * record) {
     // of the auto inc field.
     //
     if (share->has_auto_inc && record == table->record[0]) {
-        pthread_mutex_lock(&share->mutex);
+        tokudb_pthread_mutex_lock(&share->mutex);
         ulonglong curr_auto_inc = retrieve_auto_increment(
             table->field[share->ai_field_index]->key_type(), 
             field_offset(table->field[share->ai_field_index], table),
@@ -3906,7 +3929,7 @@ int ha_tokudb::write_row(uchar * record) {
                 update_max_auto_inc(share->status_block, share->last_auto_increment);
             }
         }
-        pthread_mutex_unlock(&share->mutex);
+        tokudb_pthread_mutex_unlock(&share->mutex);
     }
 
     //
@@ -4084,7 +4107,7 @@ int ha_tokudb::update_row(const uchar * old_row, uchar * new_row) {
     // of the auto inc field.
     //
     if (share->has_auto_inc && new_row == table->record[0]) {
-        pthread_mutex_lock(&share->mutex);
+        tokudb_pthread_mutex_lock(&share->mutex);
         ulonglong curr_auto_inc = retrieve_auto_increment(
             table->field[share->ai_field_index]->key_type(), 
             field_offset(table->field[share->ai_field_index], table),
@@ -4096,7 +4119,7 @@ int ha_tokudb::update_row(const uchar * old_row, uchar * new_row) {
                 share->last_auto_increment = curr_auto_inc;
             }
         }
-        pthread_mutex_unlock(&share->mutex);
+        tokudb_pthread_mutex_unlock(&share->mutex);
     }
 
     //
@@ -5377,6 +5400,8 @@ int ha_tokudb::index_prev(uchar * buf) {
     TOKUDB_DBUG_RETURN(error);
 }
 
+volatile int tokudb_index_first_wait = 0;
+
 //
 // Reads the first row from the active index (cursor) into buf, and advances cursor
 // Parameters:
@@ -5398,6 +5423,8 @@ int ha_tokudb::index_first(uchar * buf) {
 
     ha_statistic_increment(&SSV::ha_read_first_count);
 
+    while (tokudb_index_first_wait) sleep(1);
+
     info.ha = this;
     info.buf = buf;
     info.keynr = tokudb_active_index;
@@ -6125,7 +6152,7 @@ int ha_tokudb::external_lock(THD * thd, int lock_type) {
         transaction = trx->sub_sp_level;
     }
     else {
-        pthread_mutex_lock(&share->mutex);
+        tokudb_pthread_mutex_lock(&share->mutex);
         // hate dealing with comparison of signed vs unsigned, so doing this
         if (deleted_rows > added_rows && share->rows < (deleted_rows - added_rows)) {
             share->rows = 0;
@@ -6133,7 +6160,7 @@ int ha_tokudb::external_lock(THD * thd, int lock_type) {
         else {
             share->rows += (added_rows - deleted_rows);
         }
-        pthread_mutex_unlock(&share->mutex);
+        tokudb_pthread_mutex_unlock(&share->mutex);
         added_rows = 0;
         deleted_rows = 0;
         share->rows_from_locked_table = 0;
@@ -7443,7 +7470,7 @@ void ha_tokudb::get_auto_increment(ulonglong offset, ulonglong increment, ulongl
     ulonglong nr;
     bool over;
 
-    pthread_mutex_lock(&share->mutex);
+    tokudb_pthread_mutex_lock(&share->mutex);
 
     if (share->auto_inc_create_value > share->last_auto_increment) {
         nr = share->auto_inc_create_value;
@@ -7472,7 +7499,7 @@ void ha_tokudb::get_auto_increment(ulonglong offset, ulonglong increment, ulongl
     }
     *first_value = nr;
     *nb_reserved_values = nb_desired_values;
-    pthread_mutex_unlock(&share->mutex);
+    tokudb_pthread_mutex_unlock(&share->mutex);
     DBUG_VOID_RETURN;
 }
 
@@ -7838,18 +7865,18 @@ int ha_tokudb::tokudb_add_index(
     // We have an accurate row count, might as well update share->rows
     //
     if(!creating_hot_index) {
-        pthread_mutex_lock(&share->mutex);
+        tokudb_pthread_mutex_lock(&share->mutex);
         share->rows = num_processed;
-        pthread_mutex_unlock(&share->mutex);
+        tokudb_pthread_mutex_unlock(&share->mutex);
     }
     //
     // now write stuff to status.tokudb
     //
-    pthread_mutex_lock(&share->mutex);
+    tokudb_pthread_mutex_lock(&share->mutex);
     for (uint i = 0; i < num_of_keys; i++) {
         write_key_name_to_status(share->status_block, key_info[i].name, txn);
     }
-    pthread_mutex_unlock(&share->mutex);
+    tokudb_pthread_mutex_unlock(&share->mutex);
     
     error = 0;
 cleanup:

storage/tokudb/ha_tokudb.h

@@ -126,6 +126,10 @@ typedef struct hot_optimize_context {
 // and auto increment information.
 //
 class TOKUDB_SHARE {
+public:
+    void init(void);
+    void destroy(void);
+
 public:
     char *table_name;
     uint table_name_length, use_count;
@@ -184,6 +188,10 @@ public:
     bool replace_into_fast;
     rw_lock_t num_DBs_lock;
     uint32_t num_DBs;
+
+    pthread_cond_t m_openclose_cond;
+    enum { CLOSED, OPENING, OPENED, CLOSING, ERROR } m_state;
+    int m_error;
 };
 
 typedef struct st_filter_key_part_info {
@@ -443,10 +451,7 @@ private:
     int write_auto_inc_create(DB* db, ulonglong val, DB_TXN* txn);
     void init_auto_increment();
     bool can_replace_into_be_fast(TABLE_SHARE* table_share, KEY_AND_COL_INFO* kc_info, uint pk);
-    int initialize_share(
-        const char* name,
-        int mode
-        );
+    int initialize_share(const char* name, int mode);
 
     void set_query_columns(uint keynr);
     int prelock_range (const key_range *start_key, const key_range *end_key);
@@ -599,10 +604,10 @@ public:
     int get_status(DB_TXN* trans);
     void init_hidden_prim_key_info();
     inline void get_auto_primary_key(uchar * to) {
-        pthread_mutex_lock(&share->mutex);
+        tokudb_pthread_mutex_lock(&share->mutex);
         share->auto_ident++;
         hpk_num_to_char(to, share->auto_ident);
-        pthread_mutex_unlock(&share->mutex);
+        tokudb_pthread_mutex_unlock(&share->mutex);
     }
     virtual void get_auto_increment(ulonglong offset, ulonglong increment, ulonglong nb_desired_values, ulonglong * first_value, ulonglong * nb_reserved_values);
     bool is_optimize_blocking();

storage/tokudb/hatoku_defines.h

@@ -443,4 +443,44 @@ static inline void* tokudb_my_multi_malloc(myf myFlags, ...) {
   return start;
 }
 
+static inline void tokudb_pthread_mutex_init(pthread_mutex_t *mutex, const pthread_mutexattr_t *attr) {
+    int r = pthread_mutex_init(mutex, attr);
+    assert(r == 0);
+}
+
+static inline void tokudb_pthread_mutex_destroy(pthread_mutex_t *mutex) {
+    int r = pthread_mutex_destroy(mutex);
+    assert(r == 0);
+}
+
+static inline void tokudb_pthread_mutex_lock(pthread_mutex_t *mutex) {
+    int r = pthread_mutex_lock(mutex);
+    assert(r == 0);
+}
+
+static inline void tokudb_pthread_mutex_unlock(pthread_mutex_t *mutex) {
+    int r = pthread_mutex_unlock(mutex);
+    assert(r == 0);
+}
+
+static inline void tokudb_pthread_cond_init(pthread_cond_t *cond, const pthread_condattr_t *attr) {
+    int r = pthread_cond_init(cond, attr);
+    assert(r == 0);
+}
+
+static inline void tokudb_pthread_cond_destroy(pthread_cond_t *cond) {
+    int r = pthread_cond_destroy(cond);
+    assert(r == 0);
+}
+
+static inline void tokudb_pthread_cond_wait(pthread_cond_t *cond, pthread_mutex_t *mutex) {
+    int r = pthread_cond_wait(cond, mutex);
+    assert(r == 0);
+}
+
+static inline void tokudb_pthread_cond_broadcast(pthread_cond_t *cond) {
+    int r = pthread_cond_broadcast(cond);
+    assert(r == 0);
+}
+
 #endif

storage/tokudb/hatoku_hton.cc

@@ -311,7 +311,7 @@ static int tokudb_init_func(void *p) {
 
     tokudb_hton = (handlerton *) p;
 
-    pthread_mutex_init(&tokudb_mutex, MY_MUTEX_INIT_FAST);
+    tokudb_pthread_mutex_init(&tokudb_mutex, MY_MUTEX_INIT_FAST);
     (void) my_hash_init(&tokudb_open_tables, table_alias_charset, 32, 0, 0, (my_hash_get_key) tokudb_get_key, 0, 0);
 
     tokudb_hton->state = SHOW_OPTION_YES;
@@ -529,7 +529,7 @@ static int tokudb_done_func(void *p) {
     tokudb_my_free(toku_global_status_rows);
     toku_global_status_rows = NULL;
     my_hash_free(&tokudb_open_tables);
-    pthread_mutex_destroy(&tokudb_mutex);
+    tokudb_pthread_mutex_destroy(&tokudb_mutex);
 #if defined(_WIN64)
     toku_ydb_destroy();
 #endif