implement bit_arr to speed up searching the first free slot in a pline

Author: trdtnguyen

storage/innobase/buf/buf0flu.cc

@@ -3654,6 +3654,28 @@ buf_flush_request_force(
 	os_event_set(buf_flush_event);
 }
 #if defined (UNIV_PMEMOBJ_PART_PL)
+void
+pm_ppl_buf_flush_recv_note_modification(
+	PMEMobjpool*		pop,
+	PMEM_PAGE_PART_LOG*	ppl,
+	buf_block_t*    block,
+	lsn_t       start_lsn,
+	lsn_t       end_lsn) 
+{
+	buf_page_mutex_enter(block);
+
+	block->page.newest_modification = end_lsn;
+	if (!block->page.oldest_modification) {
+		buf_pool_t*	buf_pool = buf_pool_from_block(block);
+
+		buf_flush_insert_sorted_into_flush_list(
+			buf_pool, block, start_lsn);
+	} else {
+		ut_ad(block->page.oldest_modification <= start_lsn);
+	}
+
+	buf_page_mutex_exit(block);
+}
 /*
  *Called by pm_ppl_checkpoint()
  * */

storage/innobase/include/my_pmemobj.h

@@ -571,12 +571,17 @@ struct plog_hash_t {
  * A hashed line has array of log blocks share the same hashed value and log buffer
  * */
 struct __pmem_page_log_hashed_line {
+	/*general lock protect data in PART 1 and PART 3*/
 	PMEMrwlock		lock;
 
+	/*PART 1, writing log rec to logbuf*/
 	int hashed_id;
 	TOID(PMEM_PAGE_LOG_BUF)	logbuf;	//pointer to current log buffer (head)
 	TOID(PMEM_PAGE_LOG_BUF)	tail_logbuf;	//pointer to the oldest log buffer (tail)
 
+	os_event_t		log_flush_event;
+	bool			is_flushing;
+
 	uint64_t		diskaddr; //log file offset, update when flush log, reset when purging file
 	uint64_t		write_diskaddr; //diskaddr that log recs are durable write write_diskaddr < diskaddr
 
@@ -589,7 +594,9 @@ struct __pmem_page_log_hashed_line {
 	uint64_t		recv_diskaddr; //the diskaddr begin the recover, min of log blocks diskaddr
 	uint64_t		recv_off; //the offset begin the recover
 	uint64_t		recv_lsn; //min lsn of block's beginLSN in this line
-	/*end test */
+	
+	/*PART 2, metadata: log_block array, hashtable, std::map*/
+	PMEMrwlock		meta_lock;
 
 	TOID_ARRAY(TOID(PMEM_PAGE_LOG_BLOCK)) arr;
 	uint64_t n_blocks; //the current non-free blocks
@@ -602,6 +609,7 @@ struct __pmem_page_log_hashed_line {
 
 	std::map<uint64_t, uint32_t>* offset_map;
 
+	/*PART 3: recovery*/
 	//Alternative to recv_sys_t in InnoDB, allocate in DRAM when recovery
 	PMEM_RECV_LINE* recv_line;
 	bool			is_redoing;
@@ -1118,6 +1126,11 @@ void
 pm_ppl_init_in_mem(
 		PMEMobjpool*		pop,
 		PMEM_PAGE_PART_LOG*		ppl);
+
+void 
+pm_ppl_free_in_mem(
+		PMEMobjpool*		pop,
+		PMEM_PAGE_PART_LOG*		ppl);
 void
 pm_page_part_log_hash_create(
 		PMEMobjpool*		pop,
@@ -1541,6 +1554,14 @@ pm_ppl_recv_apply_hashed_line(
 	PMEM_PAGE_LOG_HASHED_LINE* pline,
 	ibool	allow_ibuf);
 
+void
+pm_ppl_buf_flush_recv_note_modification(
+	PMEMobjpool*		pop,
+	PMEM_PAGE_PART_LOG*	ppl,
+	buf_block_t*    block,
+	lsn_t       start_lsn,
+	lsn_t       end_lsn);
+
 void
 pm_ppl_check_input_rec(
 		byte*		    ptr,
@@ -1575,18 +1596,21 @@ pm_ppl_remove_fil_spaces();
 //////////// BIT ARRAY /////////
 void
 pm_bit_set(
+		PMEM_PAGE_LOG_HASHED_LINE* pline,
 		long long*	arr,
 		size_t		block_size,
 		uint64_t	bit_i);
 
 void
 pm_bit_clear(
+		PMEM_PAGE_LOG_HASHED_LINE* pline,
 		long long*	arr,
 		size_t		block_size,
 		uint64_t	bit_i);
 
 int32_t 
 pm_search_first_free_slot(
+		PMEM_PAGE_LOG_HASHED_LINE* pline,
 		long long*	bit_arr,
 		uint16_t	n_bit_blocks,
 		uint16_t	block_size);

storage/innobase/log/log0recv.cc

@@ -82,7 +82,11 @@ this must be less than UNIV_PAGE_SIZE as it is stored in the buffer pool */
 #define RECV_DATA_BLOCK_SIZE	(MEM_MAX_ALLOC_IN_BUF - sizeof(recv_data_t))
 
 /** Read-ahead area in applying log records to file pages */
+#if defined (UNIV_PMEMOBJ_PART_PL)
+#define RECV_READ_AHEAD_AREA	16
+#else //original
 #define RECV_READ_AHEAD_AREA	32
+#endif //UNIV_PMEMOBJ_PART_PL
 
 /** The recovery system */
 recv_sys_t*	recv_sys = NULL;
@@ -5950,10 +5954,13 @@ pm_ppl_recv_recover_page_func(
 	if (modification_to_page) {
 		ut_a(block);
 
-		log_flush_order_mutex_enter();
-		//buf_flush_recv_note_modification(block, start_lsn, end_lsn);
-		buf_flush_recv_note_modification(block, start_lsn, start_lsn);
-		log_flush_order_mutex_exit();
+		//log_flush_order_mutex_enter();
+		//buf_flush_recv_note_modification(block, start_lsn, start_lsn);
+		//log_flush_order_mutex_exit();
+
+		/*the flush_order_mutex may has high contention, try to simulate buf_flush_recv_note_modification()*/
+		pm_ppl_buf_flush_recv_note_modification(
+				pop, ppl, block, start_lsn, start_lsn);
 	}
 
 	/* Make sure that committing mtr does not change the modification lsn values of page */
@@ -6309,9 +6316,7 @@ pm_ppl_recv_apply_hashed_log_recs(
 		recv_line->is_ibuf_avail = allow_ibuf;
         //Asign pline to a redoer thread
         redoer->hashed_line_arr[i] = pline;
-
     }
-	
 
     /*trigger REDOer threads phase 2. 
 	 * Call pm_ppl_recv_apply_hashed_line() */
@@ -6575,16 +6580,7 @@ pm_ppl_recv_apply_hashed_line(
 					mtr_commit(&mtr);
 				} else {
 					/* page is not cached, fetch it from disk and apply is done in IO thread -> pm_ppl_recv_recover_page_func */
-
-					/*read-ahead approach*/
 					pm_ppl_recv_read_in_area (pop, ppl, recv_line, page_id);
-					/*single read approach*/
-
-				//	recv_line->n_read_reqs++;
-
-				//	bool read_ok = buf_read_page(page_id, page_size);
-
-				//	assert(read_ok);
 				}
 			}
 			cnt++;