block: revert part of 18ce3751ccd488c78d3827e9f6bf54e6322676fb

commit 78f707bfc723552e8309b7c38a8d0cc51012e813 upstream.

The above commit added WRITE_SYNC and switched various places to using
that for committing writes that will be waited upon immediately after
submission. However, this causes a performance regression with AS and CFQ
for ext3 at least, since sync_dirty_buffer() will submit some writes with
WRITE_SYNC while ext3 has sumitted others dependent writes without the sync
flag set. This causes excessive anticipation/idling in the IO scheduler
because sync and async writes get interleaved, causing a big performance
regression for the below test case (which is meant to simulate sqlite
like behaviour).

---- test case ----

int main(int argc, char **argv)
{

	int fdes, i;
	FILE *fp;
	struct timeval start;
	struct timeval end;
	struct timeval res;

	gettimeofday(&start, NULL);
	for (i=0; i<ROWS; i++) {
		fp = fopen("test_file", "a");
		fprintf(fp, "Some Text Data\n");
		fdes = fileno(fp);
		fsync(fdes);
		fclose(fp);
	}
	gettimeofday(&end, NULL);

	timersub(&end, &start, &res);
	fprintf(stdout, "time to write %d lines is %ld(msec)\n", ROWS,
			(res.tv_sec*1000000 + res.tv_usec)/1000);

	return 0;
}

-------------------

Thanks to Sean.White@APCC.com for tracking down this performance
regression and providing a test case.

Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>

diff --git a/fs/buffer.c b/fs/buffer.c
index a542f97..a5d806d 100644 (file)
--- a/fs/buffer.c
+++ b/fs/buffer.c
@@ -3042,7 +3042,7 @@ int sync_dirty_buffer(struct buffer_head *bh)
        if (test_clear_buffer_dirty(bh)) {
                get_bh(bh);
                bh->b_end_io = end_buffer_write_sync;
-               ret = submit_bh(WRITE_SYNC, bh);
+               ret = submit_bh(WRITE, bh);
                wait_on_buffer(bh);
                if (buffer_eopnotsupp(bh)) {
                        clear_buffer_eopnotsupp(bh);