Applied innodb-5.1-ss1004

Fixes:
- Bug #15815: Very poor performance with multiple queries running concurrently
- Bug #22868: 'Thread thrashing' with > 50 concurrent conns under an upd-intensive workloadw
- Bug #24089: Race condition in fil_flush_file_spaces()


storage/innobase/btr/btr0btr.c:
  Applied innodb-5.1-ss1004
  
  Revision r934:
  btr_lift_page_up(): Set new page levels in all ancestor pages, not just
  in the father page.
  
  btr0btr.h: Add a BTR_MAX_LEVELS define.
storage/innobase/buf/buf0buf.c:
  Applied innodb-5.1-ss1004
  
  Revision r999:
  Reduce buffer pool mutex contention under >= 4 big concurrent
  CPU-bound SELECT queries.  (Bug #22868)
  
  Fix: replace the mutex by one mutex protecting the 'flush list'
  (and the free list) and several mutexes protecting portions of the
  buffer pool, where we keep several indivudual LRU lists of pages.
  
  This patch is from Sunny Bains and Heikki Tuuri.
storage/innobase/buf/buf0flu.c:
  Applied innodb-5.1-ss1004
  
  Revision r999:
  Reduce buffer pool mutex contention under >= 4 big concurrent
  CPU-bound SELECT queries.  (Bug #22868)
  
  Fix: replace the mutex by one mutex protecting the 'flush list'
  (and the free list) and several mutexes protecting portions of the
  buffer pool, where we keep several indivudual LRU lists of pages.
  
  This patch is from Sunny Bains and Heikki Tuuri.
storage/innobase/buf/buf0lru.c:
  Applied innodb-5.1-ss1004
  
  Revision r999:
  Reduce buffer pool mutex contention under >= 4 big concurrent
  CPU-bound SELECT queries.  (Bug #22868)
  
  Fix: replace the mutex by one mutex protecting the 'flush list'
  (and the free list) and several mutexes protecting portions of the
  buffer pool, where we keep several indivudual LRU lists of pages.
  
  This patch is from Sunny Bains and Heikki Tuuri.
  
  
  Revision r1004:
  UT_LIST_REMOVE(): Invalidate the node pointers #ifdef UNIV_DEBUG.
  
  buf_LRU_invalidate_tablespace(): Invoke UT_LIST_GET_PREV(LRU, block)
  before UT_LIST_REMOVE(LRU, buf_pool->LRU, block).
storage/innobase/dict/dict0crea.c:
  Applied innodb-5.1-ss1004
  
  Revision r973:
  Do not break the latching order in TRUNCATE TABLE.
  
  dict_truncate_index_tree(): Replace parameter rec_t* rec with
  btr_pcur_t* pcur.  Reposition pcur before calling btr_create().
  
  sync_thread_add_level(): Remove the relaxation of the assertion added in r968.
storage/innobase/fil/fil0fil.c:
  Applied innodb-5.1-ss1004
  
  Revision r1002:
  fil_flush_file_spaces(): Copy the system->unflushed_spaces list to an
  array while holding the mutex.  This removes the crash-triggering
  race condition that was introduced when fixing Bug 15653.  (Bug #24089)
storage/innobase/include/btr0btr.h:
  Applied innodb-5.1-ss1004
  
  Revision r934:
  btr_lift_page_up(): Set new page levels in all ancestor pages, not just
  in the father page.
  
  btr0btr.h: Add a BTR_MAX_LEVELS define.
storage/innobase/include/buf0buf.h:
  Applied innodb-5.1-ss1004
  
  Revision r999:
  Reduce buffer pool mutex contention under >= 4 big concurrent
  CPU-bound SELECT queries.  (Bug #22868)
  
  Fix: replace the mutex by one mutex protecting the 'flush list'
  (and the free list) and several mutexes protecting portions of the
  buffer pool, where we keep several indivudual LRU lists of pages.
  
  This patch is from Sunny Bains and Heikki Tuuri.
storage/innobase/include/buf0buf.ic:
  Applied innodb-5.1-ss1004
  
  Revision r999:
  Reduce buffer pool mutex contention under >= 4 big concurrent
  CPU-bound SELECT queries.  (Bug #22868)
  
  Fix: replace the mutex by one mutex protecting the 'flush list'
  (and the free list) and several mutexes protecting portions of the
  buffer pool, where we keep several indivudual LRU lists of pages.
  
  This patch is from Sunny Bains and Heikki Tuuri.
storage/innobase/include/dict0crea.h:
  Applied innodb-5.1-ss1004
  
  Revision r973:
  Do not break the latching order in TRUNCATE TABLE.
  
  dict_truncate_index_tree(): Replace parameter rec_t* rec with
  btr_pcur_t* pcur.  Reposition pcur before calling btr_create().
  
  sync_thread_add_level(): Remove the relaxation of the assertion added in r968.
storage/innobase/include/ut0lst.h:
  Applied innodb-5.1-ss1004
  
  Revision r1004:
  UT_LIST_REMOVE(): Invalidate the node pointers #ifdef UNIV_DEBUG.
  
  buf_LRU_invalidate_tablespace(): Invoke UT_LIST_GET_PREV(LRU, block)
  before UT_LIST_REMOVE(LRU, buf_pool->LRU, block).
storage/innobase/row/row0mysql.c:
  Applied innodb-5.1-ss1004
  
  Revision r973:
  Do not break the latching order in TRUNCATE TABLE.
  
  dict_truncate_index_tree(): Replace parameter rec_t* rec with
  btr_pcur_t* pcur.  Reposition pcur before calling btr_create().
  
  sync_thread_add_level(): Remove the relaxation of the assertion added in r968.

1 files changed, 63 insertions, 25 deletions

diff --git a/storage/innobase/buf/buf0flu.c b/storage/innobase/buf/buf0flu.c
index 8ef8c3b9358..49b81196a64 100644
--- a/storage/innobase/buf/buf0flu.c
+++ b/storage/innobase/buf/buf0flu.c
@@ -113,6 +113,7 @@ buf_flush_ready_for_replace(
 {
 #ifdef UNIV_SYNC_DEBUG
 	ut_ad(mutex_own(&(buf_pool->mutex)));
+	ut_ad(mutex_own(&block->mutex));
 #endif /* UNIV_SYNC_DEBUG */
 	if (block->state != BUF_BLOCK_FILE_PAGE) {
 		ut_print_timestamp(stderr);
@@ -148,6 +149,7 @@ buf_flush_ready_for_flush(
 {
 #ifdef UNIV_SYNC_DEBUG
 	ut_ad(mutex_own(&(buf_pool->mutex)));
+	ut_ad(mutex_own(&(block->mutex)));
 #endif /* UNIV_SYNC_DEBUG */
 	ut_a(block->state == BUF_BLOCK_FILE_PAGE);
 
@@ -559,8 +561,15 @@ buf_flush_try_page(
 
 	ut_a(!block || block->state == BUF_BLOCK_FILE_PAGE);
 
+	if (!block) {
+		mutex_exit(&(buf_pool->mutex));
+		return(0);
+	}
+
+	mutex_enter(&block->mutex);
+
 	if (flush_type == BUF_FLUSH_LIST
-	    && block && buf_flush_ready_for_flush(block, flush_type)) {
+	    && buf_flush_ready_for_flush(block, flush_type)) {
 
 		block->io_fix = BUF_IO_WRITE;
 
@@ -598,6 +607,7 @@ buf_flush_try_page(
 			locked = TRUE;
 		}
 
+		mutex_exit(&block->mutex);
 		mutex_exit(&(buf_pool->mutex));
 
 		if (!locked) {
@@ -618,7 +628,7 @@ buf_flush_try_page(
 
 		return(1);
 
-	} else if (flush_type == BUF_FLUSH_LRU && block
+	} else if (flush_type == BUF_FLUSH_LRU
 		   && buf_flush_ready_for_flush(block, flush_type)) {
 
 		/* VERY IMPORTANT:
@@ -659,13 +669,14 @@ buf_flush_try_page(
 		buf_pool mutex: this ensures that the latch is acquired
 		immediately. */
 
+		mutex_exit(&block->mutex);
 		mutex_exit(&(buf_pool->mutex));
 
 		buf_flush_write_block_low(block);
 
 		return(1);
 
-	} else if (flush_type == BUF_FLUSH_SINGLE_PAGE && block
+	} else if (flush_type == BUF_FLUSH_SINGLE_PAGE
 		   && buf_flush_ready_for_flush(block, flush_type)) {
 
 		block->io_fix = BUF_IO_WRITE;
@@ -692,6 +703,7 @@ buf_flush_try_page(
 
 		(buf_pool->n_flush[flush_type])++;
 
+		mutex_exit(&block->mutex);
 		mutex_exit(&(buf_pool->mutex));
 
 		rw_lock_s_lock_gen(&(block->lock), BUF_IO_WRITE);
@@ -709,11 +721,12 @@ buf_flush_try_page(
 		buf_flush_write_block_low(block);
 
 		return(1);
-	} else {
-		mutex_exit(&(buf_pool->mutex));
-
-		return(0);
 	}
+
+	mutex_exit(&block->mutex);
+	mutex_exit(&(buf_pool->mutex));
+
+	return(0);
 }
 
 /***************************************************************
@@ -758,34 +771,48 @@ buf_flush_try_neighbors(
 		block = buf_page_hash_get(space, i);
 		ut_a(!block || block->state == BUF_BLOCK_FILE_PAGE);
 
-		if (block && flush_type == BUF_FLUSH_LRU && i != offset
-		    && !block->old) {
+		if (!block) {
+
+			continue;
+
+		} else if (flush_type == BUF_FLUSH_LRU && i != offset
+			   && !block->old) {
 
 			/* We avoid flushing 'non-old' blocks in an LRU flush,
 			because the flushed blocks are soon freed */
 
 			continue;
-		}
+		} else {
 
-		if (block && buf_flush_ready_for_flush(block, flush_type)
-		    && (i == offset || block->buf_fix_count == 0)) {
-			/* We only try to flush those neighbors != offset
-			where the buf fix count is zero, as we then know that
-			we probably can latch the page without a semaphore
-			wait. Semaphore waits are expensive because we must
-			flush the doublewrite buffer before we start
-			waiting. */
+			mutex_enter(&block->mutex);
 
-			mutex_exit(&(buf_pool->mutex));
+			if (buf_flush_ready_for_flush(block, flush_type)
+			    && (i == offset || block->buf_fix_count == 0)) {
+				/* We only try to flush those
+				neighbors != offset where the buf fix count is
+				zero, as we then know that we probably can
+				latch the page without a semaphore wait.
+				Semaphore waits are expensive because we must
+				flush the doublewrite buffer before we start
+				waiting. */
 
-			/* Note: as we release the buf_pool mutex above, in
-			buf_flush_try_page we cannot be sure the page is still
-			in a flushable state: therefore we check it again
-			inside that function. */
+				mutex_exit(&block->mutex);
 
-			count += buf_flush_try_page(space, i, flush_type);
+				mutex_exit(&(buf_pool->mutex));
 
-			mutex_enter(&(buf_pool->mutex));
+				/* Note: as we release the buf_pool mutex
+				above, in buf_flush_try_page we cannot be sure
+				the page is still in a flushable state:
+				therefore we check it again inside that
+				function. */
+
+				count += buf_flush_try_page(space, i,
+							    flush_type);
+
+				mutex_enter(&(buf_pool->mutex));
+			} else {
+				mutex_exit(&block->mutex);
+			}
 		}
 	}
 
@@ -879,12 +906,15 @@ buf_flush_batch(
 		while ((block != NULL) && !found) {
 			ut_a(block->state == BUF_BLOCK_FILE_PAGE);
 
+			mutex_enter(&block->mutex);
+
 			if (buf_flush_ready_for_flush(block, flush_type)) {
 
 				found = TRUE;
 				space = block->space;
 				offset = block->offset;
 
+				mutex_exit(&block->mutex);
 				mutex_exit(&(buf_pool->mutex));
 
 				old_page_count = page_count;
@@ -901,10 +931,14 @@ buf_flush_batch(
 
 			} else if (flush_type == BUF_FLUSH_LRU) {
 
+				mutex_exit(&block->mutex);
+
 				block = UT_LIST_GET_PREV(LRU, block);
 			} else {
 				ut_ad(flush_type == BUF_FLUSH_LIST);
 
+				mutex_exit(&block->mutex);
+
 				block = UT_LIST_GET_PREV(flush_list, block);
 			}
 		}
@@ -986,10 +1020,14 @@ buf_flush_LRU_recommendation(void)
 		   + BUF_FLUSH_EXTRA_MARGIN)
 	       && (distance < BUF_LRU_FREE_SEARCH_LEN)) {
 
+		mutex_enter(&block->mutex);
+
 		if (buf_flush_ready_for_replace(block)) {
 			n_replaceable++;
 		}
 
+		mutex_exit(&block->mutex);
+
 		distance++;
 
 		block = UT_LIST_GET_PREV(LRU, block);