ext4: add fsync batch tuning knobs

Add new mount options, min_batch_time and max_batch_time, which
controls how long the jbd2 layer should wait for additional filesystem
operations to get batched with a synchronous write transaction.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>

Documentation/filesystems/ext4.txt

@@ -283,6 +283,35 @@ delalloc	(*)	Deferring block allocation until write-out time.
 nodelalloc		Disable delayed allocation. Blocks are allocation
 			when data is copied from user to page cache.
 
+max_batch_time=usec	Maximum amount of time ext4 should wait for
+			additional filesystem operations to be batch
+			together with a synchronous write operation.
+			Since a synchronous write operation is going to
+			force a commit and then a wait for the I/O
+			complete, it doesn't cost much, and can be a
+			huge throughput win, we wait for a small amount
+			of time to see if any other transactions can
+			piggyback on the synchronous write.   The
+			algorithm used is designed to automatically tune
+			for the speed of the disk, by measuring the
+			amount of time (on average) that it takes to
+			finish committing a transaction.  Call this time
+			the "commit time".  If the time that the
+			transactoin has been running is less than the
+			commit time, ext4 will try sleeping for the
+			commit time to see if other operations will join
+			the transaction.   The commit time is capped by
+			the max_batch_time, which defaults to 15000us
+			(15ms).   This optimization can be turned off
+			entirely by setting max_batch_time to 0.
+
+min_batch_time=usec	This parameter sets the commit time (as
+			described above) to be at least min_batch_time.
+			It defaults to zero microseconds.  Increasing
+			this parameter may improve the throughput of
+			multi-threaded, synchronous workloads on very
+			fast disks, at the cost of increasing latency.
+
 Data Mode
 =========
 There are 3 different data modes:

fs/ext4/ext4.h

@@ -328,6 +328,7 @@ struct ext4_mount_options {
 	uid_t s_resuid;
 	gid_t s_resgid;
 	unsigned long s_commit_interval;
+	u32 s_min_batch_time, s_max_batch_time;
 #ifdef CONFIG_QUOTA
 	int s_jquota_fmt;
 	char *s_qf_names[MAXQUOTAS];
@@ -805,6 +806,12 @@ static inline int ext4_valid_inum(struct super_block *sb, unsigned long ino)
 #define EXT4_DEFM_JMODE_ORDERED	0x0040
 #define EXT4_DEFM_JMODE_WBACK	0x0060
 
+/*
+ * Default journal batch times
+ */
+#define EXT4_DEF_MIN_BATCH_TIME	0
+#define EXT4_DEF_MAX_BATCH_TIME	15000 /* 15ms */
+
 /*
  * Structure of a directory entry
  */

fs/ext4/ext4_sb.h

@@ -74,6 +74,8 @@ struct ext4_sb_info {
 	struct journal_s *s_journal;
 	struct list_head s_orphan;
 	unsigned long s_commit_interval;
+	u32 s_max_batch_time;
+	u32 s_min_batch_time;
 	struct block_device *journal_bdev;
 #ifdef CONFIG_JBD2_DEBUG
 	struct timer_list turn_ro_timer;	/* For turning read-only (crash simulation) */

fs/ext4/super.c

@@ -705,10 +705,19 @@ static int ext4_show_options(struct seq_file *seq, struct vfsmount *vfs)
 #endif
 	if (!test_opt(sb, RESERVATION))
 		seq_puts(seq, ",noreservation");
-	if (sbi->s_commit_interval) {
+	if (sbi->s_commit_interval != JBD2_DEFAULT_MAX_COMMIT_AGE*HZ) {
 		seq_printf(seq, ",commit=%u",
 			   (unsigned) (sbi->s_commit_interval / HZ));
 	}
+	if (sbi->s_min_batch_time != EXT4_DEF_MIN_BATCH_TIME) {
+		seq_printf(seq, ",min_batch_time=%u",
+			   (unsigned) sbi->s_min_batch_time);
+	}
+	if (sbi->s_max_batch_time != EXT4_DEF_MAX_BATCH_TIME) {
+		seq_printf(seq, ",max_batch_time=%u",
+			   (unsigned) sbi->s_min_batch_time);
+	}
+
 	/*
 	 * We're changing the default of barrier mount option, so
 	 * let's always display its mount state so it's clear what its
@@ -874,7 +883,8 @@ enum {
 	Opt_nouid32, Opt_debug, Opt_oldalloc, Opt_orlov,
 	Opt_user_xattr, Opt_nouser_xattr, Opt_acl, Opt_noacl,
 	Opt_reservation, Opt_noreservation, Opt_noload, Opt_nobh, Opt_bh,
-	Opt_commit, Opt_journal_update, Opt_journal_inum, Opt_journal_dev,
+	Opt_commit, Opt_min_batch_time, Opt_max_batch_time,
+	Opt_journal_update, Opt_journal_inum, Opt_journal_dev,
 	Opt_journal_checksum, Opt_journal_async_commit,
 	Opt_abort, Opt_data_journal, Opt_data_ordered, Opt_data_writeback,
 	Opt_data_err_abort, Opt_data_err_ignore,
@@ -913,6 +923,8 @@ static const match_table_t tokens = {
 	{Opt_nobh, "nobh"},
 	{Opt_bh, "bh"},
 	{Opt_commit, "commit=%u"},
+	{Opt_min_batch_time, "min_batch_time=%u"},
+	{Opt_max_batch_time, "max_batch_time=%u"},
 	{Opt_journal_update, "journal=update"},
 	{Opt_journal_inum, "journal=%u"},
 	{Opt_journal_dev, "journal_dev=%u"},
@@ -1131,6 +1143,22 @@ static int parse_options(char *options, struct super_block *sb,
 				option = JBD2_DEFAULT_MAX_COMMIT_AGE;
 			sbi->s_commit_interval = HZ * option;
 			break;
+		case Opt_max_batch_time:
+			if (match_int(&args[0], &option))
+				return 0;
+			if (option < 0)
+				return 0;
+			if (option == 0)
+				option = EXT4_DEF_MAX_BATCH_TIME;
+			sbi->s_max_batch_time = option;
+			break;
+		case Opt_min_batch_time:
+			if (match_int(&args[0], &option))
+				return 0;
+			if (option < 0)
+				return 0;
+			sbi->s_min_batch_time = option;
+			break;
 		case Opt_data_journal:
 			data_opt = EXT4_MOUNT_JOURNAL_DATA;
 			goto datacheck;
@@ -1979,6 +2007,9 @@ static int ext4_fill_super(struct super_block *sb, void *data, int silent)
 
 	sbi->s_resuid = le16_to_cpu(es->s_def_resuid);
 	sbi->s_resgid = le16_to_cpu(es->s_def_resgid);
+	sbi->s_commit_interval = JBD2_DEFAULT_MAX_COMMIT_AGE * HZ;
+	sbi->s_min_batch_time = EXT4_DEF_MIN_BATCH_TIME;
+	sbi->s_max_batch_time = EXT4_DEF_MAX_BATCH_TIME;
 
 	set_opt(sbi->s_mount_opt, RESERVATION);
 	set_opt(sbi->s_mount_opt, BARRIER);
@@ -2524,11 +2555,9 @@ static void ext4_init_journal_params(struct super_block *sb, journal_t *journal)
 {
 	struct ext4_sb_info *sbi = EXT4_SB(sb);
 
-	if (sbi->s_commit_interval)
-		journal->j_commit_interval = sbi->s_commit_interval;
-	/* We could also set up an ext4-specific default for the commit
-	 * interval here, but for now we'll just fall back to the jbd
-	 * default. */
+	journal->j_commit_interval = sbi->s_commit_interval;
+	journal->j_min_batch_time = sbi->s_min_batch_time;
+	journal->j_max_batch_time = sbi->s_max_batch_time;
 
 	spin_lock(&journal->j_state_lock);
 	if (test_opt(sb, BARRIER))
@@ -3042,6 +3071,8 @@ static int ext4_remount(struct super_block *sb, int *flags, char *data)
 	old_opts.s_resuid = sbi->s_resuid;
 	old_opts.s_resgid = sbi->s_resgid;
 	old_opts.s_commit_interval = sbi->s_commit_interval;
+	old_opts.s_min_batch_time = sbi->s_min_batch_time;
+	old_opts.s_max_batch_time = sbi->s_max_batch_time;
 #ifdef CONFIG_QUOTA
 	old_opts.s_jquota_fmt = sbi->s_jquota_fmt;
 	for (i = 0; i < MAXQUOTAS; i++)
@@ -3178,6 +3209,8 @@ static int ext4_remount(struct super_block *sb, int *flags, char *data)
 	sbi->s_resuid = old_opts.s_resuid;
 	sbi->s_resgid = old_opts.s_resgid;
 	sbi->s_commit_interval = old_opts.s_commit_interval;
+	sbi->s_min_batch_time = old_opts.s_min_batch_time;
+	sbi->s_max_batch_time = old_opts.s_max_batch_time;
 #ifdef CONFIG_QUOTA
 	sbi->s_jquota_fmt = old_opts.s_jquota_fmt;
 	for (i = 0; i < MAXQUOTAS; i++) {

fs/jbd2/journal.c

@@ -964,6 +964,8 @@ static journal_t * journal_init_common (void)
 	spin_lock_init(&journal->j_state_lock);
 
 	journal->j_commit_interval = (HZ * JBD2_DEFAULT_MAX_COMMIT_AGE);
+	journal->j_min_batch_time = 0;
+	journal->j_max_batch_time = 15000; /* 15ms */
 
 	/* The journal is marked for error until we succeed with recovery! */
 	journal->j_flags = JBD2_ABORT;

fs/jbd2/transaction.c

@@ -1255,8 +1255,10 @@ int jbd2_journal_stop(handle_t *handle)
 		trans_time = ktime_to_ns(ktime_sub(ktime_get(),
 						   transaction->t_start_time));
 
+		commit_time = max_t(u64, commit_time,
+				    1000*journal->j_min_batch_time);
 		commit_time = min_t(u64, commit_time,
-				    1000*jiffies_to_usecs(1));
+				    1000*journal->j_max_batch_time);
 
 		if (trans_time < commit_time) {
 			ktime_t expires = ktime_add_ns(ktime_get(),

include/linux/jbd2.h

@@ -956,6 +956,14 @@ struct journal_s
 	 */
 	u64			j_average_commit_time;
 
+	/*
+	 * minimum and maximum times that we should wait for
+	 * additional filesystem operations to get batched into a
+	 * synchronous handle in microseconds
+	 */
+	u32			j_min_batch_time;
+	u32			j_max_batch_time;
+
 	/* This function is called when a transaction is closed */
 	void			(*j_commit_callback)(journal_t *,
 						     transaction_t *);