jbd: improve fsync batching

There is a flaw with the way jbd handles fsync batching.  If we fsync() a
file and we were not the last person to run fsync() on this fs then we
automatically sleep for 1 jiffie in order to wait for new writers to join
into the transaction before forcing the commit.  The problem with this is
that with really fast storage (ie a Clariion) the time it takes to commit
a transaction to disk is way faster than 1 jiffie in most cases, so
sleeping means waiting longer with nothing to do than if we just committed
the transaction and kept going.  Ric Wheeler noticed this when using
fs_mark with more than 1 thread, the throughput would plummet as he added
more threads.

This patch attempts to fix this problem by recording the average time in
nanoseconds that it takes to commit a transaction to disk, and what time
we started the transaction.  If we run an fsync() and we have been running
for less time than it takes to commit the transaction to disk, we sleep
for the delta amount of time and then commit to disk.  We acheive
sub-jiffie sleeping using schedule_hrtimeout.  This means that the wait
time is auto-tuned to the speed of the underlying disk, instead of having
this static timeout.  I weighted the average according to somebody's
comments (Andreas Dilger I think) in order to help normalize random
outliers where we take way longer or way less time to commit than the
average.  I also have a min() check in there to make sure we don't sleep
longer than a jiffie in case our storage is super slow, this was requested
by Andrew.

I unfortunately do not have access to a Clariion, so I had to use a
ramdisk to represent a super fast array.  I tested with a SATA drive with
barrier=1 to make sure there was no regression with local disks, I tested
with a 4 way multipathed Apple Xserve RAID array and of course the
ramdisk.  I ran the following command

fs_mark -d /mnt/ext3-test -s 4096 -n 2000 -D 64 -t $i

where $i was 2, 4, 8, 16 and 32.  I mkfs'ed the fs each time.  Here are my
results

type	threads		with patch	without patch
sata	2		24.6		26.3
sata	4		49.2		48.1
sata	8		70.1		67.0
sata	16		104.0		94.1
sata	32		153.6		142.7

xserve	2		246.4		222.0
xserve	4		480.0		440.8
xserve	8		829.5		730.8
xserve	16		1172.7		1026.9
xserve	32		1816.3		1650.5

ramdisk	2		2538.3		1745.6
ramdisk	4		2942.3		661.9
ramdisk	8		2882.5		999.8
ramdisk	16		2738.7		1801.9
ramdisk	32		2541.9		2394.0

Signed-off-by: Josef Bacik <jbacik@redhat.com>
Cc: Andreas Dilger <adilger@sun.com>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Ric Wheeler <rwheeler@redhat.com>
Cc: <linux-ext4@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
diff --git a/include/linux/jbd.h b/include/linux/jbd.h
index 346e2b8..6384b19 100644
--- a/include/linux/jbd.h
+++ b/include/linux/jbd.h
@@ -543,6 +543,11 @@
 	unsigned long		t_expires;
 
 	/*
+	 * When this transaction started, in nanoseconds [no locking]
+	 */
+	ktime_t			t_start_time;
+
+	/*
 	 * How many handles used this transaction? [t_handle_lock]
 	 */
 	int t_handle_count;
@@ -798,9 +803,19 @@
 	struct buffer_head	**j_wbuf;
 	int			j_wbufsize;
 
+	/*
+	 * this is the pid of the last person to run a synchronous operation
+	 * through the journal.
+	 */
 	pid_t			j_last_sync_writer;
 
 	/*
+	 * the average amount of time in nanoseconds it takes to commit a
+	 * transaction to the disk.  [j_state_lock]
+	 */
+	u64			j_average_commit_time;
+
+	/*
 	 * An opaque pointer to fs-private information.  ext3 puts its
 	 * superblock pointer here
 	 */