imporve jbd2 fsync batching

index 8b119e1..3169db9 100644
@@ -332,6 +332,8 @@ void jbd2_journal_commit_transaction(journal_t *journal)
 	int flags;
 	int err;
 	unsigned long long blocknr;
+	ktime_t start_time;
+	u64 commit_time;
 	char *tagp = NULL;
 	journal_header_t *header;
 	journal_block_tag_t *tag = NULL;

@@ -458,6 +460,7 @@ void jbd2_journal_commit_transaction(journal_t *journal)
 	commit_transaction->t_state = T_FLUSH;
 	journal->j_committing_transaction = commit_transaction;
 	journal->j_running_transaction = NULL;
+	start_time = ktime_get();
 	commit_transaction->t_log_start = journal->j_head;
 	wake_up(&journal->j_wait_transaction_locked);
 	spin_unlock(&journal->j_state_lock);

@@ -972,6 +975,17 @@ restart_loop:
 	J_ASSERT(commit_transaction == journal->j_committing_transaction);
 	journal->j_commit_sequence = commit_transaction->t_tid;
 	journal->j_committing_transaction = NULL;
+	commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
+
+	/*
+	 * weight the commit time higher than the average time so we don't
+	 * react too strongly to vast changes in the commit time
+	 */
+	if (likely(journal->j_average_commit_time))
+		journal->j_average_commit_time = (commit_time*3 +
+				journal->j_average_commit_time) / 4;
+	else
+		journal->j_average_commit_time = commit_time;
 	spin_unlock(&journal->j_state_lock);
 
 	if (commit_transaction->t_checkpoint_list == NULL &&

index 39b7805..ef705f9 100644
@@ -25,6 +25,7 @@
 #include <linux/timer.h>
 #include <linux/mm.h>
 #include <linux/highmem.h>
+#include <linux/hrtimer.h>
 
 static void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh);
 

@@ -48,6 +49,7 @@ jbd2_get_transaction(journal_t *journal, transaction_t *transaction)
 {
 	transaction->t_journal = journal;
 	transaction->t_state = T_RUNNING;
+	transaction->t_start_time = ktime_get();
 	transaction->t_tid = journal->j_transaction_sequence++;
 	transaction->t_expires = jiffies + journal->j_commit_interval;
 	spin_lock_init(&transaction->t_handle_lock);

@@ -1193,7 +1195,7 @@ int jbd2_journal_stop(handle_t *handle)
 {
 	transaction_t *transaction = handle->h_transaction;
 	journal_t *journal = transaction->t_journal;
-	int old_handle_count, err;
+	int err;
 	pid_t pid;
 
 	J_ASSERT(journal_current_handle() == handle);

@@ -1222,6 +1224,17 @@ int jbd2_journal_stop(handle_t *handle)
 	 * on IO anyway.  Speeds up many-threaded, many-dir operations
 	 * by 30x or more...
 	 *
+	 * We try and optimize the sleep time against what the underlying disk
+	 * can do, instead of having a static sleep time.  This is usefull for
+	 * the case where our storage is so fast that it is more optimal to go
+	 * ahead and force a flush and wait for the transaction to be committed
+	 * than it is to wait for an arbitrary amount of time for new writers to
+	 * join the transaction.  We acheive this by measuring how long it takes
+	 * to commit a transaction, and compare it with how long this
+	 * transaction has been running, and if run time < commit time then we
+	 * sleep for the delta and commit.  This greatly helps super fast disks
+	 * that would see slowdowns as more threads started doing fsyncs.
+	 *
 	 * But don't do this if this process was the most recent one to
 	 * perform a synchronous write.  We do this to detect the case where a
 	 * single process is doing a stream of sync writes.  No point in waiting

@@ -1229,11 +1242,26 @@ int jbd2_journal_stop(handle_t *handle)
 	 */
 	pid = current->pid;
 	if (handle->h_sync && journal->j_last_sync_writer != pid) {
+		u64 commit_time, trans_time;
+
 		journal->j_last_sync_writer = pid;
-		do {
-			old_handle_count = transaction->t_handle_count;
-			schedule_timeout_uninterruptible(1);
-		} while (old_handle_count != transaction->t_handle_count);
+
+		spin_lock(&journal->j_state_lock);
+		commit_time = journal->j_average_commit_time;
+		spin_unlock(&journal->j_state_lock);
+
+		trans_time = ktime_to_ns(ktime_sub(ktime_get(),
+						   transaction->t_start_time));
+
+		commit_time = min_t(u64, commit_time,
+				    1000*jiffies_to_usecs(1));
+
+		if (trans_time < commit_time) {
+			ktime_t expires = ktime_add_ns(ktime_get(),
+						       commit_time);
+			set_current_state(TASK_UNINTERRUPTIBLE);
+			schedule_hrtimeout(&expires, HRTIMER_MODE_ABS);
+		}
 	}
 
 	current->journal_info = NULL;

index c7d106e..b8b8744 100644
@@ -637,6 +637,11 @@ struct transaction_s
 	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;

@@ -938,8 +943,18 @@ struct journal_s
 	struct buffer_head	**j_wbuf;
 	int			j_wbufsize;
 
+	/*
+	 * this is the pid of hte 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 disk. [j_state_lock]
+	 */
+	u64			j_average_commit_time;
+
 	/* This function is called when a transaction is closed */
 	void			(*j_commit_callback)(journal_t *,
 						     transaction_t *);

index 8370ffd..6244c5d 100644
@@ -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];

@@ -806,6 +807,12 @@ static inline int ext4_valid_inum(struct super_block *sb, unsigned long ino)
 #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
  */
 #define EXT4_NAME_LEN 255

index 2f3b8b1..f5b5d56 100644
@@ -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) */

96 lines not on this page · raw message

index e70d657..dfacd9e 100644
@@ -961,6 +961,8 @@ 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;

index 0e1fc34..e7de7ac 100644
@@ -1255,9 +1255,11 @@ 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));