drbd: new configuration parameter c-min-rate

We now track the data rate of locally submitted resync related requests,
and can thus detect non-resync activity on the lower level device.

If the current sync rate is above c-min-rate, and the lower level device
appears to be busy, we throttle the resyncer.

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
diff --git a/drivers/block/drbd/drbd_int.h b/drivers/block/drbd/drbd_int.h
index 0fce3f3..0fedcc0 100644
--- a/drivers/block/drbd/drbd_int.h
+++ b/drivers/block/drbd/drbd_int.h
@@ -1513,6 +1513,7 @@
 extern void resync_timer_fn(unsigned long data);
 
 /* drbd_receiver.c */
+extern int drbd_rs_should_slow_down(struct drbd_conf *mdev);
 extern int drbd_submit_ee(struct drbd_conf *mdev, struct drbd_epoch_entry *e,
 		const unsigned rw, const int fault_type);
 extern int drbd_release_ee(struct drbd_conf *mdev, struct list_head *list);
diff --git a/drivers/block/drbd/drbd_main.c b/drivers/block/drbd/drbd_main.c
index 1ff8418..db93eee 100644
--- a/drivers/block/drbd/drbd_main.c
+++ b/drivers/block/drbd/drbd_main.c
@@ -1098,6 +1098,8 @@
 		mdev->ov_left = mdev->rs_total
 			      - BM_SECT_TO_BIT(mdev->ov_position);
 		mdev->rs_start = now;
+		mdev->rs_last_events = 0;
+		mdev->rs_last_sect_ev = 0;
 		mdev->ov_last_oos_size = 0;
 		mdev->ov_last_oos_start = 0;
 
@@ -2706,7 +2708,8 @@
 		/* .c_plan_ahead = */	DRBD_C_PLAN_AHEAD_DEF,
 		/* .c_delay_target = */	DRBD_C_DELAY_TARGET_DEF,
 		/* .c_fill_target = */	DRBD_C_FILL_TARGET_DEF,
-		/* .c_max_rate = */	DRBD_C_MAX_RATE_DEF
+		/* .c_max_rate = */	DRBD_C_MAX_RATE_DEF,
+		/* .c_min_rate = */	DRBD_C_MIN_RATE_DEF
 	};
 
 	/* Have to use that way, because the layout differs between
@@ -2742,6 +2745,7 @@
 	atomic_set(&mdev->packet_seq, 0);
 	atomic_set(&mdev->pp_in_use, 0);
 	atomic_set(&mdev->rs_sect_in, 0);
+	atomic_set(&mdev->rs_sect_ev, 0);
 
 	mutex_init(&mdev->md_io_mutex);
 	mutex_init(&mdev->data.mutex);
@@ -2819,6 +2823,7 @@
 	mdev->rs_total     =
 	mdev->rs_failed    = 0;
 	mdev->rs_last_events = 0;
+	mdev->rs_last_sect_ev = 0;
 	for (i = 0; i < DRBD_SYNC_MARKS; i++) {
 		mdev->rs_mark_left[i] = 0;
 		mdev->rs_mark_time[i] = 0;
diff --git a/drivers/block/drbd/drbd_nl.c b/drivers/block/drbd/drbd_nl.c
index 295b8d5..6b35d41 100644
--- a/drivers/block/drbd/drbd_nl.c
+++ b/drivers/block/drbd/drbd_nl.c
@@ -1604,7 +1604,8 @@
 		sc.c_plan_ahead = DRBD_C_PLAN_AHEAD_DEF;
 		sc.c_delay_target = DRBD_C_DELAY_TARGET_DEF;
 		sc.c_fill_target = DRBD_C_FILL_TARGET_DEF;
-		sc.c_max_rate   = DRBD_C_MAX_RATE_DEF;
+		sc.c_max_rate = DRBD_C_MAX_RATE_DEF;
+		sc.c_min_rate = DRBD_C_MIN_RATE_DEF;
 	} else
 		memcpy(&sc, &mdev->sync_conf, sizeof(struct syncer_conf));
 
diff --git a/drivers/block/drbd/drbd_receiver.c b/drivers/block/drbd/drbd_receiver.c
index 346aed9..0d9967f 100644
--- a/drivers/block/drbd/drbd_receiver.c
+++ b/drivers/block/drbd/drbd_receiver.c
@@ -1561,6 +1561,7 @@
 	list_add(&e->w.list, &mdev->sync_ee);
 	spin_unlock_irq(&mdev->req_lock);
 
+	atomic_add(data_size >> 9, &mdev->rs_sect_ev);
 	if (drbd_submit_ee(mdev, e, WRITE, DRBD_FAULT_RS_WR) == 0)
 		return TRUE;
 
@@ -2017,17 +2018,66 @@
 	return FALSE;
 }
 
+/* We may throttle resync, if the lower device seems to be busy,
+ * and current sync rate is above c_min_rate.
+ *
+ * To decide whether or not the lower device is busy, we use a scheme similar
+ * to MD RAID is_mddev_idle(): if the partition stats reveal "significant"
+ * (more than 64 sectors) of activity we cannot account for with our own resync
+ * activity, it obviously is "busy".
+ *
+ * The current sync rate used here uses only the most recent two step marks,
+ * to have a short time average so we can react faster.
+ */
+int drbd_rs_should_slow_down(struct drbd_conf *mdev)
+{
+	struct gendisk *disk = mdev->ldev->backing_bdev->bd_contains->bd_disk;
+	unsigned long db, dt, dbdt;
+	int curr_events;
+	int throttle = 0;
+
+	/* feature disabled? */
+	if (mdev->sync_conf.c_min_rate == 0)
+		return 0;
+
+	curr_events = (int)part_stat_read(&disk->part0, sectors[0]) +
+		      (int)part_stat_read(&disk->part0, sectors[1]) -
+			atomic_read(&mdev->rs_sect_ev);
+	if (!mdev->rs_last_events || curr_events - mdev->rs_last_events > 64) {
+		unsigned long rs_left;
+		int i;
+
+		mdev->rs_last_events = curr_events;
+
+		/* sync speed average over the last 2*DRBD_SYNC_MARK_STEP,
+		 * approx. */
+		i = (mdev->rs_last_mark + DRBD_SYNC_MARKS-2) % DRBD_SYNC_MARKS;
+		rs_left = drbd_bm_total_weight(mdev) - mdev->rs_failed;
+
+		dt = ((long)jiffies - (long)mdev->rs_mark_time[i]) / HZ;
+		if (!dt)
+			dt++;
+		db = mdev->rs_mark_left[i] - rs_left;
+		dbdt = Bit2KB(db/dt);
+
+		if (dbdt > mdev->sync_conf.c_min_rate)
+			throttle = 1;
+	}
+	return throttle;
+}
+
+
 static int receive_DataRequest(struct drbd_conf *mdev, struct p_header *h)
 {
 	sector_t sector;
 	const sector_t capacity = drbd_get_capacity(mdev->this_bdev);
 	struct drbd_epoch_entry *e;
 	struct digest_info *di = NULL;
+	struct p_block_req *p = (struct p_block_req *)h;
+	const int brps = sizeof(*p)-sizeof(*h);
 	int size, digest_size;
 	unsigned int fault_type;
-	struct p_block_req *p =
-		(struct p_block_req *)h;
-	const int brps = sizeof(*p)-sizeof(*h);
+
 
 	if (drbd_recv(mdev, h->payload, brps) != brps)
 		return FALSE;
@@ -2099,8 +2149,9 @@
 		} else if (h->command == P_OV_REPLY) {
 			e->w.cb = w_e_end_ov_reply;
 			dec_rs_pending(mdev);
-			/* drbd_rs_begin_io done when we sent this request */
-			goto submit;
+			/* drbd_rs_begin_io done when we sent this request,
+			 * but accounting still needs to be done. */
+			goto submit_for_resync;
 		}
 		break;
 
@@ -2128,9 +2179,36 @@
 		goto out_free_e;
 	}
 
+	/* Throttle, drbd_rs_begin_io and submit should become asynchronous
+	 * wrt the receiver, but it is not as straightforward as it may seem.
+	 * Various places in the resync start and stop logic assume resync
+	 * requests are processed in order, requeuing this on the worker thread
+	 * introduces a bunch of new code for synchronization between threads.
+	 *
+	 * Unlimited throttling before drbd_rs_begin_io may stall the resync
+	 * "forever", throttling after drbd_rs_begin_io will lock that extent
+	 * for application writes for the same time.  For now, just throttle
+	 * here, where the rest of the code expects the receiver to sleep for
+	 * a while, anyways.
+	 */
+
+	/* Throttle before drbd_rs_begin_io, as that locks out application IO;
+	 * this defers syncer requests for some time, before letting at least
+	 * on request through.  The resync controller on the receiving side
+	 * will adapt to the incoming rate accordingly.
+	 *
+	 * We cannot throttle here if remote is Primary/SyncTarget:
+	 * we would also throttle its application reads.
+	 * In that case, throttling is done on the SyncTarget only.
+	 */
+	if (mdev->state.peer != R_PRIMARY && drbd_rs_should_slow_down(mdev))
+		msleep(100);
 	if (drbd_rs_begin_io(mdev, e->sector))
 		goto out_free_e;
 
+submit_for_resync:
+	atomic_add(size >> 9, &mdev->rs_sect_ev);
+
 submit:
 	inc_unacked(mdev);
 	spin_lock_irq(&mdev->req_lock);
diff --git a/drivers/block/drbd/drbd_worker.c b/drivers/block/drbd/drbd_worker.c
index f5d779b..99c937a 100644
--- a/drivers/block/drbd/drbd_worker.c
+++ b/drivers/block/drbd/drbd_worker.c
@@ -215,10 +215,8 @@
  */
 void drbd_endio_pri(struct bio *bio, int error)
 {
-	unsigned long flags;
 	struct drbd_request *req = bio->bi_private;
 	struct drbd_conf *mdev = req->mdev;
-	struct bio_and_error m;
 	enum drbd_req_event what;
 	int uptodate = bio_flagged(bio, BIO_UPTODATE);
 
@@ -244,12 +242,7 @@
 	bio_put(req->private_bio);
 	req->private_bio = ERR_PTR(error);
 
-	spin_lock_irqsave(&mdev->req_lock, flags);
-	__req_mod(req, what, &m);
-	spin_unlock_irqrestore(&mdev->req_lock, flags);
-
-	if (m.bio)
-		complete_master_bio(mdev, &m);
+	req_mod(req, what);
 }
 
 int w_read_retry_remote(struct drbd_conf *mdev, struct drbd_work *w, int cancel)
@@ -376,6 +369,9 @@
 	if (!get_ldev(mdev))
 		return -EIO;
 
+	if (drbd_rs_should_slow_down(mdev))
+		goto defer;
+
 	/* GFP_TRY, because if there is no memory available right now, this may
 	 * be rescheduled for later. It is "only" background resync, after all. */
 	e = drbd_alloc_ee(mdev, DRBD_MAGIC+0xbeef, sector, size, GFP_TRY);
@@ -387,6 +383,7 @@
 	list_add(&e->w.list, &mdev->read_ee);
 	spin_unlock_irq(&mdev->req_lock);
 
+	atomic_add(size >> 9, &mdev->rs_sect_ev);
 	if (drbd_submit_ee(mdev, e, READ, DRBD_FAULT_RS_RD) == 0)
 		return 0;
 
@@ -512,8 +509,9 @@
 	sector_t sector;
 	const sector_t capacity = drbd_get_capacity(mdev->this_bdev);
 	int max_segment_size;
-	int number, i, rollback_i, size, pe, mx;
+	int number, rollback_i, size, pe, mx;
 	int align, queued, sndbuf;
+	int i = 0;
 
 	if (unlikely(cancel))
 		return 1;
@@ -549,7 +547,14 @@
 		mdev->c_sync_rate = mdev->sync_conf.rate;
 		number = SLEEP_TIME * mdev->c_sync_rate  / ((BM_BLOCK_SIZE / 1024) * HZ);
 	}
-	pe = atomic_read(&mdev->rs_pending_cnt);
+
+	/* Throttle resync on lower level disk activity, which may also be
+	 * caused by application IO on Primary/SyncTarget.
+	 * Keep this after the call to drbd_rs_controller, as that assumes
+	 * to be called as precisely as possible every SLEEP_TIME,
+	 * and would be confused otherwise. */
+	if (drbd_rs_should_slow_down(mdev))
+		goto requeue;
 
 	mutex_lock(&mdev->data.mutex);
 	if (mdev->data.socket)
@@ -563,6 +568,7 @@
 		mx = number;
 
 	/* Limit the number of pending RS requests to no more than the peer's receive buffer */
+	pe = atomic_read(&mdev->rs_pending_cnt);
 	if ((pe + number) > mx) {
 		number = mx - pe;
 	}
@@ -1492,6 +1498,8 @@
 		mdev->rs_failed    = 0;
 		mdev->rs_paused    = 0;
 		mdev->rs_same_csum = 0;
+		mdev->rs_last_events = 0;
+		mdev->rs_last_sect_ev = 0;
 		mdev->rs_total     = tw;
 		mdev->rs_start     = now;
 		for (i = 0; i < DRBD_SYNC_MARKS; i++) {
@@ -1516,6 +1524,7 @@
 		}
 
 		atomic_set(&mdev->rs_sect_in, 0);
+		atomic_set(&mdev->rs_sect_ev, 0);
 		mdev->rs_in_flight = 0;
 		mdev->rs_planed = 0;
 		spin_lock(&mdev->peer_seq_lock);