os0file.c Always do fsync after a file write to reduce possibility of a partially written page in an OS crash

buf0rea.c	Fixed a recovery hang associated with ibuf bitmap pages


innobase/buf/buf0rea.c:
  Fixed a recovery hang associated with ibuf bitmap pages
innobase/os/os0file.c:
  Always do fsync after a file write to reduce possibility of a partially written page in an OS crash

2 files changed, 33 insertions, 5 deletions

diff --git a/innobase/buf/buf0rea.c b/innobase/buf/buf0rea.c
index 13e9ed0476b..644dd226a0e 100644
--- a/innobase/buf/buf0rea.c
+++ b/innobase/buf/buf0rea.c
@@ -73,11 +73,13 @@ buf_read_page_low(
 		sync = TRUE;
 	}
 #endif
-	if (trx_sys_hdr_page(space, offset)) {
+	if (ibuf_bitmap_page(offset) || trx_sys_hdr_page(space, offset)) {
 
 		/* Trx sys header is so low in the latching order that we play
 		safe and do not leave the i/o-completion to an asynchronous
-		i/o-thread: */
+		i/o-thread. Ibuf bitmap pages must always be read with
+                syncronous i/o, to make sure they do not get involved in
+                thread deadlocks. */
 		
 		sync = TRUE;
 	}
diff --git a/innobase/os/os0file.c b/innobase/os/os0file.c
index 804a63507ce..c2cedba5137 100644
--- a/innobase/os/os0file.c
+++ b/innobase/os/os0file.c
@@ -559,10 +559,18 @@ os_file_pwrite(
 	ulint		n,	/* in: number of bytes to write */	
 	ulint		offset)	/* in: offset where to write */
 {
+	ssize_t	ret;
+
 #ifdef HAVE_PWRITE
-	return(pwrite(file, buf, n, (off_t) offset));
+	ret = pwrite(file, buf, n, (off_t) offset);
+
+	/* Always do fsync to reduce the probability that when the OS crashes,
+	a database page is only partially physically written to disk. */
+
+	ut_a(TRUE == os_file_flush(file));
+
+        return(ret);
 #else
-	ssize_t	ret;
 	ulint	i;
 
 	/* Protect the seek / write operation with a mutex */
@@ -580,6 +588,11 @@ os_file_pwrite(
 	
 	ret = write(file, buf, n);
 
+	/* Always do fsync to reduce the probability that when the OS crashes,
+	a database page is only partially physically written to disk. */
+
+	ut_a(TRUE == os_file_flush(file));
+
 	os_mutex_exit(os_file_seek_mutexes[i]);
 
 	return(ret);
@@ -733,9 +746,14 @@ try_again:
 	} 
 
 	ret = WriteFile(file, buf, n, &len, NULL);
+	
+	/* Always do fsync to reduce the probability that when the OS crashes,
+	a database page is only partially physically written to disk. */
+
+	ut_a(TRUE == os_file_flush(file));
 
 	os_mutex_exit(os_file_seek_mutexes[i]);
-	
+
 	if (ret && len == n) {
 		return(TRUE);
 	}
@@ -1497,6 +1515,10 @@ os_aio_windows_handle(
 
 	if (ret && len == slot->len) {
 		ret_val = TRUE;
+
+		if (slot->type == OS_FILE_WRITE) {
+		         ut_a(TRUE == os_file_flush(slot->file));
+		}
 	} else {
 		err = GetLastError();
 		ut_error;
@@ -1578,6 +1600,10 @@ os_aio_posix_handle(
 	*message1 = slot->message1;
 	*message2 = slot->message2;
 
+	if (slot->type == OS_FILE_WRITE) {
+		ut_a(TRUE == os_file_flush(slot->file));
+	}
+
 	os_mutex_exit(array->mutex);
 
 	os_aio_array_free_slot(array, slot);